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 static void __migrate_swap_task(struct task_struct *p, int cpu)
1019 struct rq *src_rq, *dst_rq;
1021 src_rq = task_rq(p);
1022 dst_rq = cpu_rq(cpu);
1024 deactivate_task(src_rq, p, 0);
1025 set_task_cpu(p, cpu);
1026 activate_task(dst_rq, p, 0);
1027 check_preempt_curr(dst_rq, p, 0);
1030 * Task isn't running anymore; make it appear like we migrated
1031 * it before it went to sleep. This means on wakeup we make the
1032 * previous cpu our targer instead of where it really is.
1038 struct migration_swap_arg {
1039 struct task_struct *src_task, *dst_task;
1040 int src_cpu, dst_cpu;
1043 static int migrate_swap_stop(void *data)
1045 struct migration_swap_arg *arg = data;
1046 struct rq *src_rq, *dst_rq;
1049 src_rq = cpu_rq(arg->src_cpu);
1050 dst_rq = cpu_rq(arg->dst_cpu);
1052 double_rq_lock(src_rq, dst_rq);
1053 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1056 if (task_cpu(arg->src_task) != arg->src_cpu)
1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1065 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1071 double_rq_unlock(src_rq, dst_rq);
1077 * Cross migrate two tasks
1079 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1081 struct migration_swap_arg arg;
1086 arg = (struct migration_swap_arg){
1088 .src_cpu = task_cpu(cur),
1090 .dst_cpu = task_cpu(p),
1093 if (arg.src_cpu == arg.dst_cpu)
1096 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1099 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1102 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1105 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1112 struct migration_arg {
1113 struct task_struct *task;
1117 static int migration_cpu_stop(void *data);
1120 * wait_task_inactive - wait for a thread to unschedule.
1122 * If @match_state is nonzero, it's the @p->state value just checked and
1123 * not expected to change. If it changes, i.e. @p might have woken up,
1124 * then return zero. When we succeed in waiting for @p to be off its CPU,
1125 * we return a positive number (its total switch count). If a second call
1126 * a short while later returns the same number, the caller can be sure that
1127 * @p has remained unscheduled the whole time.
1129 * The caller must ensure that the task *will* unschedule sometime soon,
1130 * else this function might spin for a *long* time. This function can't
1131 * be called with interrupts off, or it may introduce deadlock with
1132 * smp_call_function() if an IPI is sent by the same process we are
1133 * waiting to become inactive.
1135 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1137 unsigned long flags;
1144 * We do the initial early heuristics without holding
1145 * any task-queue locks at all. We'll only try to get
1146 * the runqueue lock when things look like they will
1152 * If the task is actively running on another CPU
1153 * still, just relax and busy-wait without holding
1156 * NOTE! Since we don't hold any locks, it's not
1157 * even sure that "rq" stays as the right runqueue!
1158 * But we don't care, since "task_running()" will
1159 * return false if the runqueue has changed and p
1160 * is actually now running somewhere else!
1162 while (task_running(rq, p)) {
1163 if (match_state && unlikely(p->state != match_state))
1169 * Ok, time to look more closely! We need the rq
1170 * lock now, to be *sure*. If we're wrong, we'll
1171 * just go back and repeat.
1173 rq = task_rq_lock(p, &flags);
1174 trace_sched_wait_task(p);
1175 running = task_running(rq, p);
1178 if (!match_state || p->state == match_state)
1179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1180 task_rq_unlock(rq, p, &flags);
1183 * If it changed from the expected state, bail out now.
1185 if (unlikely(!ncsw))
1189 * Was it really running after all now that we
1190 * checked with the proper locks actually held?
1192 * Oops. Go back and try again..
1194 if (unlikely(running)) {
1200 * It's not enough that it's not actively running,
1201 * it must be off the runqueue _entirely_, and not
1204 * So if it was still runnable (but just not actively
1205 * running right now), it's preempted, and we should
1206 * yield - it could be a while.
1208 if (unlikely(on_rq)) {
1209 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1211 set_current_state(TASK_UNINTERRUPTIBLE);
1212 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217 * Ahh, all good. It wasn't running, and it wasn't
1218 * runnable, which means that it will never become
1219 * running in the future either. We're all done!
1228 * kick_process - kick a running thread to enter/exit the kernel
1229 * @p: the to-be-kicked thread
1231 * Cause a process which is running on another CPU to enter
1232 * kernel-mode, without any delay. (to get signals handled.)
1234 * NOTE: this function doesn't have to take the runqueue lock,
1235 * because all it wants to ensure is that the remote task enters
1236 * the kernel. If the IPI races and the task has been migrated
1237 * to another CPU then no harm is done and the purpose has been
1240 void kick_process(struct task_struct *p)
1246 if ((cpu != smp_processor_id()) && task_curr(p))
1247 smp_send_reschedule(cpu);
1250 EXPORT_SYMBOL_GPL(kick_process);
1251 #endif /* CONFIG_SMP */
1255 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1257 static int select_fallback_rq(int cpu, struct task_struct *p)
1259 int nid = cpu_to_node(cpu);
1260 const struct cpumask *nodemask = NULL;
1261 enum { cpuset, possible, fail } state = cpuset;
1265 * If the node that the cpu is on has been offlined, cpu_to_node()
1266 * will return -1. There is no cpu on the node, and we should
1267 * select the cpu on the other node.
1270 nodemask = cpumask_of_node(nid);
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu, nodemask) {
1274 if (!cpu_online(dest_cpu))
1276 if (!cpu_active(dest_cpu))
1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1284 /* Any allowed, online CPU? */
1285 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1286 if (!cpu_online(dest_cpu))
1288 if (!cpu_active(dest_cpu))
1295 /* No more Mr. Nice Guy. */
1296 cpuset_cpus_allowed_fallback(p);
1301 do_set_cpus_allowed(p, cpu_possible_mask);
1312 if (state != cpuset) {
1314 * Don't tell them about moving exiting tasks or
1315 * kernel threads (both mm NULL), since they never
1318 if (p->mm && printk_ratelimit()) {
1319 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1320 task_pid_nr(p), p->comm, cpu);
1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1331 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1333 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1336 * In order not to call set_task_cpu() on a blocking task we need
1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1340 * Since this is common to all placement strategies, this lives here.
1342 * [ this allows ->select_task() to simply return task_cpu(p) and
1343 * not worry about this generic constraint ]
1345 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1347 cpu = select_fallback_rq(task_cpu(p), p);
1352 static void update_avg(u64 *avg, u64 sample)
1354 s64 diff = sample - *avg;
1360 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1362 #ifdef CONFIG_SCHEDSTATS
1363 struct rq *rq = this_rq();
1366 int this_cpu = smp_processor_id();
1368 if (cpu == this_cpu) {
1369 schedstat_inc(rq, ttwu_local);
1370 schedstat_inc(p, se.statistics.nr_wakeups_local);
1372 struct sched_domain *sd;
1374 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1376 for_each_domain(this_cpu, sd) {
1377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1378 schedstat_inc(sd, ttwu_wake_remote);
1385 if (wake_flags & WF_MIGRATED)
1386 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1388 #endif /* CONFIG_SMP */
1390 schedstat_inc(rq, ttwu_count);
1391 schedstat_inc(p, se.statistics.nr_wakeups);
1393 if (wake_flags & WF_SYNC)
1394 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1396 #endif /* CONFIG_SCHEDSTATS */
1399 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1401 activate_task(rq, p, en_flags);
1404 /* if a worker is waking up, notify workqueue */
1405 if (p->flags & PF_WQ_WORKER)
1406 wq_worker_waking_up(p, cpu_of(rq));
1410 * Mark the task runnable and perform wakeup-preemption.
1413 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1415 check_preempt_curr(rq, p, wake_flags);
1416 trace_sched_wakeup(p, true);
1418 p->state = TASK_RUNNING;
1420 if (p->sched_class->task_woken)
1421 p->sched_class->task_woken(rq, p);
1423 if (rq->idle_stamp) {
1424 u64 delta = rq_clock(rq) - rq->idle_stamp;
1425 u64 max = 2*rq->max_idle_balance_cost;
1427 update_avg(&rq->avg_idle, delta);
1429 if (rq->avg_idle > max)
1438 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1441 if (p->sched_contributes_to_load)
1442 rq->nr_uninterruptible--;
1445 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1446 ttwu_do_wakeup(rq, p, wake_flags);
1450 * Called in case the task @p isn't fully descheduled from its runqueue,
1451 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1452 * since all we need to do is flip p->state to TASK_RUNNING, since
1453 * the task is still ->on_rq.
1455 static int ttwu_remote(struct task_struct *p, int wake_flags)
1460 rq = __task_rq_lock(p);
1462 /* check_preempt_curr() may use rq clock */
1463 update_rq_clock(rq);
1464 ttwu_do_wakeup(rq, p, wake_flags);
1467 __task_rq_unlock(rq);
1473 static void sched_ttwu_pending(void)
1475 struct rq *rq = this_rq();
1476 struct llist_node *llist = llist_del_all(&rq->wake_list);
1477 struct task_struct *p;
1479 raw_spin_lock(&rq->lock);
1482 p = llist_entry(llist, struct task_struct, wake_entry);
1483 llist = llist_next(llist);
1484 ttwu_do_activate(rq, p, 0);
1487 raw_spin_unlock(&rq->lock);
1490 void scheduler_ipi(void)
1493 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1494 * TIF_NEED_RESCHED remotely (for the first time) will also send
1497 if (tif_need_resched())
1498 set_preempt_need_resched();
1500 if (llist_empty(&this_rq()->wake_list)
1501 && !tick_nohz_full_cpu(smp_processor_id())
1502 && !got_nohz_idle_kick())
1506 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1507 * traditionally all their work was done from the interrupt return
1508 * path. Now that we actually do some work, we need to make sure
1511 * Some archs already do call them, luckily irq_enter/exit nest
1514 * Arguably we should visit all archs and update all handlers,
1515 * however a fair share of IPIs are still resched only so this would
1516 * somewhat pessimize the simple resched case.
1519 tick_nohz_full_check();
1520 sched_ttwu_pending();
1523 * Check if someone kicked us for doing the nohz idle load balance.
1525 if (unlikely(got_nohz_idle_kick())) {
1526 this_rq()->idle_balance = 1;
1527 raise_softirq_irqoff(SCHED_SOFTIRQ);
1532 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1534 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1535 smp_send_reschedule(cpu);
1538 bool cpus_share_cache(int this_cpu, int that_cpu)
1540 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1542 #endif /* CONFIG_SMP */
1544 static void ttwu_queue(struct task_struct *p, int cpu)
1546 struct rq *rq = cpu_rq(cpu);
1548 #if defined(CONFIG_SMP)
1549 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1550 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1551 ttwu_queue_remote(p, cpu);
1556 raw_spin_lock(&rq->lock);
1557 ttwu_do_activate(rq, p, 0);
1558 raw_spin_unlock(&rq->lock);
1562 * try_to_wake_up - wake up a thread
1563 * @p: the thread to be awakened
1564 * @state: the mask of task states that can be woken
1565 * @wake_flags: wake modifier flags (WF_*)
1567 * Put it on the run-queue if it's not already there. The "current"
1568 * thread is always on the run-queue (except when the actual
1569 * re-schedule is in progress), and as such you're allowed to do
1570 * the simpler "current->state = TASK_RUNNING" to mark yourself
1571 * runnable without the overhead of this.
1573 * Return: %true if @p was woken up, %false if it was already running.
1574 * or @state didn't match @p's state.
1577 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1579 unsigned long flags;
1580 int cpu, success = 0;
1583 * If we are going to wake up a thread waiting for CONDITION we
1584 * need to ensure that CONDITION=1 done by the caller can not be
1585 * reordered with p->state check below. This pairs with mb() in
1586 * set_current_state() the waiting thread does.
1588 smp_mb__before_spinlock();
1589 raw_spin_lock_irqsave(&p->pi_lock, flags);
1590 if (!(p->state & state))
1593 success = 1; /* we're going to change ->state */
1596 if (p->on_rq && ttwu_remote(p, wake_flags))
1601 * If the owning (remote) cpu is still in the middle of schedule() with
1602 * this task as prev, wait until its done referencing the task.
1607 * Pairs with the smp_wmb() in finish_lock_switch().
1611 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1612 p->state = TASK_WAKING;
1614 if (p->sched_class->task_waking)
1615 p->sched_class->task_waking(p);
1617 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1618 if (task_cpu(p) != cpu) {
1619 wake_flags |= WF_MIGRATED;
1620 set_task_cpu(p, cpu);
1622 #endif /* CONFIG_SMP */
1626 ttwu_stat(p, cpu, wake_flags);
1628 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1634 * try_to_wake_up_local - try to wake up a local task with rq lock held
1635 * @p: the thread to be awakened
1637 * Put @p on the run-queue if it's not already there. The caller must
1638 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1641 static void try_to_wake_up_local(struct task_struct *p)
1643 struct rq *rq = task_rq(p);
1645 if (WARN_ON_ONCE(rq != this_rq()) ||
1646 WARN_ON_ONCE(p == current))
1649 lockdep_assert_held(&rq->lock);
1651 if (!raw_spin_trylock(&p->pi_lock)) {
1652 raw_spin_unlock(&rq->lock);
1653 raw_spin_lock(&p->pi_lock);
1654 raw_spin_lock(&rq->lock);
1657 if (!(p->state & TASK_NORMAL))
1661 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1663 ttwu_do_wakeup(rq, p, 0);
1664 ttwu_stat(p, smp_processor_id(), 0);
1666 raw_spin_unlock(&p->pi_lock);
1670 * wake_up_process - Wake up a specific process
1671 * @p: The process to be woken up.
1673 * Attempt to wake up the nominated process and move it to the set of runnable
1676 * Return: 1 if the process was woken up, 0 if it was already running.
1678 * It may be assumed that this function implies a write memory barrier before
1679 * changing the task state if and only if any tasks are woken up.
1681 int wake_up_process(struct task_struct *p)
1683 WARN_ON(task_is_stopped_or_traced(p));
1684 return try_to_wake_up(p, TASK_NORMAL, 0);
1686 EXPORT_SYMBOL(wake_up_process);
1688 int wake_up_state(struct task_struct *p, unsigned int state)
1690 return try_to_wake_up(p, state, 0);
1694 * Perform scheduler related setup for a newly forked process p.
1695 * p is forked by current.
1697 * __sched_fork() is basic setup used by init_idle() too:
1699 static void __sched_fork(struct task_struct *p)
1704 p->se.exec_start = 0;
1705 p->se.sum_exec_runtime = 0;
1706 p->se.prev_sum_exec_runtime = 0;
1707 p->se.nr_migrations = 0;
1709 INIT_LIST_HEAD(&p->se.group_node);
1711 #ifdef CONFIG_SCHEDSTATS
1712 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1715 INIT_LIST_HEAD(&p->rt.run_list);
1717 #ifdef CONFIG_PREEMPT_NOTIFIERS
1718 INIT_HLIST_HEAD(&p->preempt_notifiers);
1721 #ifdef CONFIG_NUMA_BALANCING
1722 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1723 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1724 p->mm->numa_next_reset = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1725 p->mm->numa_scan_seq = 0;
1728 p->node_stamp = 0ULL;
1729 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1730 p->numa_migrate_seq = 1;
1731 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1732 p->numa_preferred_nid = -1;
1733 p->numa_work.next = &p->numa_work;
1734 p->numa_faults = NULL;
1735 p->numa_faults_buffer = NULL;
1737 INIT_LIST_HEAD(&p->numa_entry);
1738 p->numa_group = NULL;
1739 #endif /* CONFIG_NUMA_BALANCING */
1742 #ifdef CONFIG_NUMA_BALANCING
1743 #ifdef CONFIG_SCHED_DEBUG
1744 void set_numabalancing_state(bool enabled)
1747 sched_feat_set("NUMA");
1749 sched_feat_set("NO_NUMA");
1752 __read_mostly bool numabalancing_enabled;
1754 void set_numabalancing_state(bool enabled)
1756 numabalancing_enabled = enabled;
1758 #endif /* CONFIG_SCHED_DEBUG */
1759 #endif /* CONFIG_NUMA_BALANCING */
1762 * fork()/clone()-time setup:
1764 void sched_fork(struct task_struct *p)
1766 unsigned long flags;
1767 int cpu = get_cpu();
1771 * We mark the process as running here. This guarantees that
1772 * nobody will actually run it, and a signal or other external
1773 * event cannot wake it up and insert it on the runqueue either.
1775 p->state = TASK_RUNNING;
1778 * Make sure we do not leak PI boosting priority to the child.
1780 p->prio = current->normal_prio;
1783 * Revert to default priority/policy on fork if requested.
1785 if (unlikely(p->sched_reset_on_fork)) {
1786 if (task_has_rt_policy(p)) {
1787 p->policy = SCHED_NORMAL;
1788 p->static_prio = NICE_TO_PRIO(0);
1790 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1791 p->static_prio = NICE_TO_PRIO(0);
1793 p->prio = p->normal_prio = __normal_prio(p);
1797 * We don't need the reset flag anymore after the fork. It has
1798 * fulfilled its duty:
1800 p->sched_reset_on_fork = 0;
1803 if (!rt_prio(p->prio))
1804 p->sched_class = &fair_sched_class;
1806 if (p->sched_class->task_fork)
1807 p->sched_class->task_fork(p);
1810 * The child is not yet in the pid-hash so no cgroup attach races,
1811 * and the cgroup is pinned to this child due to cgroup_fork()
1812 * is ran before sched_fork().
1814 * Silence PROVE_RCU.
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1817 set_task_cpu(p, cpu);
1818 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1820 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1821 if (likely(sched_info_on()))
1822 memset(&p->sched_info, 0, sizeof(p->sched_info));
1824 #if defined(CONFIG_SMP)
1827 init_task_preempt_count(p);
1829 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1836 * wake_up_new_task - wake up a newly created task for the first time.
1838 * This function will do some initial scheduler statistics housekeeping
1839 * that must be done for every newly created context, then puts the task
1840 * on the runqueue and wakes it.
1842 void wake_up_new_task(struct task_struct *p)
1844 unsigned long flags;
1847 raw_spin_lock_irqsave(&p->pi_lock, flags);
1850 * Fork balancing, do it here and not earlier because:
1851 * - cpus_allowed can change in the fork path
1852 * - any previously selected cpu might disappear through hotplug
1854 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1857 /* Initialize new task's runnable average */
1858 init_task_runnable_average(p);
1859 rq = __task_rq_lock(p);
1860 activate_task(rq, p, 0);
1862 trace_sched_wakeup_new(p, true);
1863 check_preempt_curr(rq, p, WF_FORK);
1865 if (p->sched_class->task_woken)
1866 p->sched_class->task_woken(rq, p);
1868 task_rq_unlock(rq, p, &flags);
1871 #ifdef CONFIG_PREEMPT_NOTIFIERS
1874 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1875 * @notifier: notifier struct to register
1877 void preempt_notifier_register(struct preempt_notifier *notifier)
1879 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1881 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1884 * preempt_notifier_unregister - no longer interested in preemption notifications
1885 * @notifier: notifier struct to unregister
1887 * This is safe to call from within a preemption notifier.
1889 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1891 hlist_del(¬ifier->link);
1893 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1895 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1897 struct preempt_notifier *notifier;
1899 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1900 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1904 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1905 struct task_struct *next)
1907 struct preempt_notifier *notifier;
1909 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1910 notifier->ops->sched_out(notifier, next);
1913 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1915 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1920 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1921 struct task_struct *next)
1925 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1928 * prepare_task_switch - prepare to switch tasks
1929 * @rq: the runqueue preparing to switch
1930 * @prev: the current task that is being switched out
1931 * @next: the task we are going to switch to.
1933 * This is called with the rq lock held and interrupts off. It must
1934 * be paired with a subsequent finish_task_switch after the context
1937 * prepare_task_switch sets up locking and calls architecture specific
1941 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1942 struct task_struct *next)
1944 trace_sched_switch(prev, next);
1945 sched_info_switch(rq, prev, next);
1946 perf_event_task_sched_out(prev, next);
1947 fire_sched_out_preempt_notifiers(prev, next);
1948 prepare_lock_switch(rq, next);
1949 prepare_arch_switch(next);
1953 * finish_task_switch - clean up after a task-switch
1954 * @rq: runqueue associated with task-switch
1955 * @prev: the thread we just switched away from.
1957 * finish_task_switch must be called after the context switch, paired
1958 * with a prepare_task_switch call before the context switch.
1959 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1960 * and do any other architecture-specific cleanup actions.
1962 * Note that we may have delayed dropping an mm in context_switch(). If
1963 * so, we finish that here outside of the runqueue lock. (Doing it
1964 * with the lock held can cause deadlocks; see schedule() for
1967 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1968 __releases(rq->lock)
1970 struct mm_struct *mm = rq->prev_mm;
1976 * A task struct has one reference for the use as "current".
1977 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1978 * schedule one last time. The schedule call will never return, and
1979 * the scheduled task must drop that reference.
1980 * The test for TASK_DEAD must occur while the runqueue locks are
1981 * still held, otherwise prev could be scheduled on another cpu, die
1982 * there before we look at prev->state, and then the reference would
1984 * Manfred Spraul <manfred@colorfullife.com>
1986 prev_state = prev->state;
1987 vtime_task_switch(prev);
1988 finish_arch_switch(prev);
1989 perf_event_task_sched_in(prev, current);
1990 finish_lock_switch(rq, prev);
1991 finish_arch_post_lock_switch();
1993 fire_sched_in_preempt_notifiers(current);
1996 if (unlikely(prev_state == TASK_DEAD)) {
1997 task_numa_free(prev);
2000 * Remove function-return probe instances associated with this
2001 * task and put them back on the free list.
2003 kprobe_flush_task(prev);
2004 put_task_struct(prev);
2007 tick_nohz_task_switch(current);
2012 /* assumes rq->lock is held */
2013 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2015 if (prev->sched_class->pre_schedule)
2016 prev->sched_class->pre_schedule(rq, prev);
2019 /* rq->lock is NOT held, but preemption is disabled */
2020 static inline void post_schedule(struct rq *rq)
2022 if (rq->post_schedule) {
2023 unsigned long flags;
2025 raw_spin_lock_irqsave(&rq->lock, flags);
2026 if (rq->curr->sched_class->post_schedule)
2027 rq->curr->sched_class->post_schedule(rq);
2028 raw_spin_unlock_irqrestore(&rq->lock, flags);
2030 rq->post_schedule = 0;
2036 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2040 static inline void post_schedule(struct rq *rq)
2047 * schedule_tail - first thing a freshly forked thread must call.
2048 * @prev: the thread we just switched away from.
2050 asmlinkage void schedule_tail(struct task_struct *prev)
2051 __releases(rq->lock)
2053 struct rq *rq = this_rq();
2055 finish_task_switch(rq, prev);
2058 * FIXME: do we need to worry about rq being invalidated by the
2063 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2064 /* In this case, finish_task_switch does not reenable preemption */
2067 if (current->set_child_tid)
2068 put_user(task_pid_vnr(current), current->set_child_tid);
2072 * context_switch - switch to the new MM and the new
2073 * thread's register state.
2076 context_switch(struct rq *rq, struct task_struct *prev,
2077 struct task_struct *next)
2079 struct mm_struct *mm, *oldmm;
2081 prepare_task_switch(rq, prev, next);
2084 oldmm = prev->active_mm;
2086 * For paravirt, this is coupled with an exit in switch_to to
2087 * combine the page table reload and the switch backend into
2090 arch_start_context_switch(prev);
2093 next->active_mm = oldmm;
2094 atomic_inc(&oldmm->mm_count);
2095 enter_lazy_tlb(oldmm, next);
2097 switch_mm(oldmm, mm, next);
2100 prev->active_mm = NULL;
2101 rq->prev_mm = oldmm;
2104 * Since the runqueue lock will be released by the next
2105 * task (which is an invalid locking op but in the case
2106 * of the scheduler it's an obvious special-case), so we
2107 * do an early lockdep release here:
2109 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2110 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2113 context_tracking_task_switch(prev, next);
2114 /* Here we just switch the register state and the stack. */
2115 switch_to(prev, next, prev);
2119 * this_rq must be evaluated again because prev may have moved
2120 * CPUs since it called schedule(), thus the 'rq' on its stack
2121 * frame will be invalid.
2123 finish_task_switch(this_rq(), prev);
2127 * nr_running and nr_context_switches:
2129 * externally visible scheduler statistics: current number of runnable
2130 * threads, total number of context switches performed since bootup.
2132 unsigned long nr_running(void)
2134 unsigned long i, sum = 0;
2136 for_each_online_cpu(i)
2137 sum += cpu_rq(i)->nr_running;
2142 unsigned long long nr_context_switches(void)
2145 unsigned long long sum = 0;
2147 for_each_possible_cpu(i)
2148 sum += cpu_rq(i)->nr_switches;
2153 unsigned long nr_iowait(void)
2155 unsigned long i, sum = 0;
2157 for_each_possible_cpu(i)
2158 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2163 unsigned long nr_iowait_cpu(int cpu)
2165 struct rq *this = cpu_rq(cpu);
2166 return atomic_read(&this->nr_iowait);
2172 * sched_exec - execve() is a valuable balancing opportunity, because at
2173 * this point the task has the smallest effective memory and cache footprint.
2175 void sched_exec(void)
2177 struct task_struct *p = current;
2178 unsigned long flags;
2181 raw_spin_lock_irqsave(&p->pi_lock, flags);
2182 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2183 if (dest_cpu == smp_processor_id())
2186 if (likely(cpu_active(dest_cpu))) {
2187 struct migration_arg arg = { p, dest_cpu };
2189 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2190 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2194 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2199 DEFINE_PER_CPU(struct kernel_stat, kstat);
2200 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2202 EXPORT_PER_CPU_SYMBOL(kstat);
2203 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2206 * Return any ns on the sched_clock that have not yet been accounted in
2207 * @p in case that task is currently running.
2209 * Called with task_rq_lock() held on @rq.
2211 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2215 if (task_current(rq, p)) {
2216 update_rq_clock(rq);
2217 ns = rq_clock_task(rq) - p->se.exec_start;
2225 unsigned long long task_delta_exec(struct task_struct *p)
2227 unsigned long flags;
2231 rq = task_rq_lock(p, &flags);
2232 ns = do_task_delta_exec(p, rq);
2233 task_rq_unlock(rq, p, &flags);
2239 * Return accounted runtime for the task.
2240 * In case the task is currently running, return the runtime plus current's
2241 * pending runtime that have not been accounted yet.
2243 unsigned long long task_sched_runtime(struct task_struct *p)
2245 unsigned long flags;
2249 rq = task_rq_lock(p, &flags);
2250 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2251 task_rq_unlock(rq, p, &flags);
2257 * This function gets called by the timer code, with HZ frequency.
2258 * We call it with interrupts disabled.
2260 void scheduler_tick(void)
2262 int cpu = smp_processor_id();
2263 struct rq *rq = cpu_rq(cpu);
2264 struct task_struct *curr = rq->curr;
2268 raw_spin_lock(&rq->lock);
2269 update_rq_clock(rq);
2270 curr->sched_class->task_tick(rq, curr, 0);
2271 update_cpu_load_active(rq);
2272 raw_spin_unlock(&rq->lock);
2274 perf_event_task_tick();
2277 rq->idle_balance = idle_cpu(cpu);
2278 trigger_load_balance(rq, cpu);
2280 rq_last_tick_reset(rq);
2283 #ifdef CONFIG_NO_HZ_FULL
2285 * scheduler_tick_max_deferment
2287 * Keep at least one tick per second when a single
2288 * active task is running because the scheduler doesn't
2289 * yet completely support full dynticks environment.
2291 * This makes sure that uptime, CFS vruntime, load
2292 * balancing, etc... continue to move forward, even
2293 * with a very low granularity.
2295 * Return: Maximum deferment in nanoseconds.
2297 u64 scheduler_tick_max_deferment(void)
2299 struct rq *rq = this_rq();
2300 unsigned long next, now = ACCESS_ONCE(jiffies);
2302 next = rq->last_sched_tick + HZ;
2304 if (time_before_eq(next, now))
2307 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2311 notrace unsigned long get_parent_ip(unsigned long addr)
2313 if (in_lock_functions(addr)) {
2314 addr = CALLER_ADDR2;
2315 if (in_lock_functions(addr))
2316 addr = CALLER_ADDR3;
2321 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2322 defined(CONFIG_PREEMPT_TRACER))
2324 void __kprobes preempt_count_add(int val)
2326 #ifdef CONFIG_DEBUG_PREEMPT
2330 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2333 __preempt_count_add(val);
2334 #ifdef CONFIG_DEBUG_PREEMPT
2336 * Spinlock count overflowing soon?
2338 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2341 if (preempt_count() == val)
2342 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2344 EXPORT_SYMBOL(preempt_count_add);
2346 void __kprobes preempt_count_sub(int val)
2348 #ifdef CONFIG_DEBUG_PREEMPT
2352 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2355 * Is the spinlock portion underflowing?
2357 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2358 !(preempt_count() & PREEMPT_MASK)))
2362 if (preempt_count() == val)
2363 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2364 __preempt_count_sub(val);
2366 EXPORT_SYMBOL(preempt_count_sub);
2371 * Print scheduling while atomic bug:
2373 static noinline void __schedule_bug(struct task_struct *prev)
2375 if (oops_in_progress)
2378 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2379 prev->comm, prev->pid, preempt_count());
2381 debug_show_held_locks(prev);
2383 if (irqs_disabled())
2384 print_irqtrace_events(prev);
2386 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2390 * Various schedule()-time debugging checks and statistics:
2392 static inline void schedule_debug(struct task_struct *prev)
2395 * Test if we are atomic. Since do_exit() needs to call into
2396 * schedule() atomically, we ignore that path for now.
2397 * Otherwise, whine if we are scheduling when we should not be.
2399 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2400 __schedule_bug(prev);
2403 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2405 schedstat_inc(this_rq(), sched_count);
2408 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2410 if (prev->on_rq || rq->skip_clock_update < 0)
2411 update_rq_clock(rq);
2412 prev->sched_class->put_prev_task(rq, prev);
2416 * Pick up the highest-prio task:
2418 static inline struct task_struct *
2419 pick_next_task(struct rq *rq)
2421 const struct sched_class *class;
2422 struct task_struct *p;
2425 * Optimization: we know that if all tasks are in
2426 * the fair class we can call that function directly:
2428 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2429 p = fair_sched_class.pick_next_task(rq);
2434 for_each_class(class) {
2435 p = class->pick_next_task(rq);
2440 BUG(); /* the idle class will always have a runnable task */
2444 * __schedule() is the main scheduler function.
2446 * The main means of driving the scheduler and thus entering this function are:
2448 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2450 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2451 * paths. For example, see arch/x86/entry_64.S.
2453 * To drive preemption between tasks, the scheduler sets the flag in timer
2454 * interrupt handler scheduler_tick().
2456 * 3. Wakeups don't really cause entry into schedule(). They add a
2457 * task to the run-queue and that's it.
2459 * Now, if the new task added to the run-queue preempts the current
2460 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2461 * called on the nearest possible occasion:
2463 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2465 * - in syscall or exception context, at the next outmost
2466 * preempt_enable(). (this might be as soon as the wake_up()'s
2469 * - in IRQ context, return from interrupt-handler to
2470 * preemptible context
2472 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2475 * - cond_resched() call
2476 * - explicit schedule() call
2477 * - return from syscall or exception to user-space
2478 * - return from interrupt-handler to user-space
2480 static void __sched __schedule(void)
2482 struct task_struct *prev, *next;
2483 unsigned long *switch_count;
2489 cpu = smp_processor_id();
2491 rcu_note_context_switch(cpu);
2494 schedule_debug(prev);
2496 if (sched_feat(HRTICK))
2500 * Make sure that signal_pending_state()->signal_pending() below
2501 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2502 * done by the caller to avoid the race with signal_wake_up().
2504 smp_mb__before_spinlock();
2505 raw_spin_lock_irq(&rq->lock);
2507 switch_count = &prev->nivcsw;
2508 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2509 if (unlikely(signal_pending_state(prev->state, prev))) {
2510 prev->state = TASK_RUNNING;
2512 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2516 * If a worker went to sleep, notify and ask workqueue
2517 * whether it wants to wake up a task to maintain
2520 if (prev->flags & PF_WQ_WORKER) {
2521 struct task_struct *to_wakeup;
2523 to_wakeup = wq_worker_sleeping(prev, cpu);
2525 try_to_wake_up_local(to_wakeup);
2528 switch_count = &prev->nvcsw;
2531 pre_schedule(rq, prev);
2533 if (unlikely(!rq->nr_running))
2534 idle_balance(cpu, rq);
2536 put_prev_task(rq, prev);
2537 next = pick_next_task(rq);
2538 clear_tsk_need_resched(prev);
2539 clear_preempt_need_resched();
2540 rq->skip_clock_update = 0;
2542 if (likely(prev != next)) {
2547 context_switch(rq, prev, next); /* unlocks the rq */
2549 * The context switch have flipped the stack from under us
2550 * and restored the local variables which were saved when
2551 * this task called schedule() in the past. prev == current
2552 * is still correct, but it can be moved to another cpu/rq.
2554 cpu = smp_processor_id();
2557 raw_spin_unlock_irq(&rq->lock);
2561 sched_preempt_enable_no_resched();
2566 static inline void sched_submit_work(struct task_struct *tsk)
2568 if (!tsk->state || tsk_is_pi_blocked(tsk))
2571 * If we are going to sleep and we have plugged IO queued,
2572 * make sure to submit it to avoid deadlocks.
2574 if (blk_needs_flush_plug(tsk))
2575 blk_schedule_flush_plug(tsk);
2578 asmlinkage void __sched schedule(void)
2580 struct task_struct *tsk = current;
2582 sched_submit_work(tsk);
2585 EXPORT_SYMBOL(schedule);
2587 #ifdef CONFIG_CONTEXT_TRACKING
2588 asmlinkage void __sched schedule_user(void)
2591 * If we come here after a random call to set_need_resched(),
2592 * or we have been woken up remotely but the IPI has not yet arrived,
2593 * we haven't yet exited the RCU idle mode. Do it here manually until
2594 * we find a better solution.
2603 * schedule_preempt_disabled - called with preemption disabled
2605 * Returns with preemption disabled. Note: preempt_count must be 1
2607 void __sched schedule_preempt_disabled(void)
2609 sched_preempt_enable_no_resched();
2614 #ifdef CONFIG_PREEMPT
2616 * this is the entry point to schedule() from in-kernel preemption
2617 * off of preempt_enable. Kernel preemptions off return from interrupt
2618 * occur there and call schedule directly.
2620 asmlinkage void __sched notrace preempt_schedule(void)
2623 * If there is a non-zero preempt_count or interrupts are disabled,
2624 * we do not want to preempt the current task. Just return..
2626 if (likely(!preemptible()))
2630 __preempt_count_add(PREEMPT_ACTIVE);
2632 __preempt_count_sub(PREEMPT_ACTIVE);
2635 * Check again in case we missed a preemption opportunity
2636 * between schedule and now.
2639 } while (need_resched());
2641 EXPORT_SYMBOL(preempt_schedule);
2644 * this is the entry point to schedule() from kernel preemption
2645 * off of irq context.
2646 * Note, that this is called and return with irqs disabled. This will
2647 * protect us against recursive calling from irq.
2649 asmlinkage void __sched preempt_schedule_irq(void)
2651 enum ctx_state prev_state;
2653 /* Catch callers which need to be fixed */
2654 BUG_ON(preempt_count() || !irqs_disabled());
2656 prev_state = exception_enter();
2659 __preempt_count_add(PREEMPT_ACTIVE);
2662 local_irq_disable();
2663 __preempt_count_sub(PREEMPT_ACTIVE);
2666 * Check again in case we missed a preemption opportunity
2667 * between schedule and now.
2670 } while (need_resched());
2672 exception_exit(prev_state);
2675 #endif /* CONFIG_PREEMPT */
2677 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2680 return try_to_wake_up(curr->private, mode, wake_flags);
2682 EXPORT_SYMBOL(default_wake_function);
2685 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2686 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2687 * number) then we wake all the non-exclusive tasks and one exclusive task.
2689 * There are circumstances in which we can try to wake a task which has already
2690 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2691 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2693 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2694 int nr_exclusive, int wake_flags, void *key)
2696 wait_queue_t *curr, *next;
2698 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2699 unsigned flags = curr->flags;
2701 if (curr->func(curr, mode, wake_flags, key) &&
2702 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2708 * __wake_up - wake up threads blocked on a waitqueue.
2710 * @mode: which threads
2711 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2712 * @key: is directly passed to the wakeup function
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 __wake_up(wait_queue_head_t *q, unsigned int mode,
2718 int nr_exclusive, void *key)
2720 unsigned long flags;
2722 spin_lock_irqsave(&q->lock, flags);
2723 __wake_up_common(q, mode, nr_exclusive, 0, key);
2724 spin_unlock_irqrestore(&q->lock, flags);
2726 EXPORT_SYMBOL(__wake_up);
2729 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2731 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2733 __wake_up_common(q, mode, nr, 0, NULL);
2735 EXPORT_SYMBOL_GPL(__wake_up_locked);
2737 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2739 __wake_up_common(q, mode, 1, 0, key);
2741 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2744 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2746 * @mode: which threads
2747 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2748 * @key: opaque value to be passed to wakeup targets
2750 * The sync wakeup differs that the waker knows that it will schedule
2751 * away soon, so while the target thread will be woken up, it will not
2752 * be migrated to another CPU - ie. the two threads are 'synchronized'
2753 * with each other. This can prevent needless bouncing between CPUs.
2755 * On UP it can prevent extra preemption.
2757 * It may be assumed that this function implies a write memory barrier before
2758 * changing the task state if and only if any tasks are woken up.
2760 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2761 int nr_exclusive, void *key)
2763 unsigned long flags;
2764 int wake_flags = WF_SYNC;
2769 if (unlikely(nr_exclusive != 1))
2772 spin_lock_irqsave(&q->lock, flags);
2773 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2774 spin_unlock_irqrestore(&q->lock, flags);
2776 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2779 * __wake_up_sync - see __wake_up_sync_key()
2781 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2783 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2785 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2788 * complete: - signals a single thread waiting on this completion
2789 * @x: holds the state of this particular completion
2791 * This will wake up a single thread waiting on this completion. Threads will be
2792 * awakened in the same order in which they were queued.
2794 * See also complete_all(), wait_for_completion() and related routines.
2796 * It may be assumed that this function implies a write memory barrier before
2797 * changing the task state if and only if any tasks are woken up.
2799 void complete(struct completion *x)
2801 unsigned long flags;
2803 spin_lock_irqsave(&x->wait.lock, flags);
2805 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2806 spin_unlock_irqrestore(&x->wait.lock, flags);
2808 EXPORT_SYMBOL(complete);
2811 * complete_all: - signals all threads waiting on this completion
2812 * @x: holds the state of this particular completion
2814 * This will wake up all threads waiting on this particular completion event.
2816 * It may be assumed that this function implies a write memory barrier before
2817 * changing the task state if and only if any tasks are woken up.
2819 void complete_all(struct completion *x)
2821 unsigned long flags;
2823 spin_lock_irqsave(&x->wait.lock, flags);
2824 x->done += UINT_MAX/2;
2825 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2826 spin_unlock_irqrestore(&x->wait.lock, flags);
2828 EXPORT_SYMBOL(complete_all);
2830 static inline long __sched
2831 do_wait_for_common(struct completion *x,
2832 long (*action)(long), long timeout, int state)
2835 DECLARE_WAITQUEUE(wait, current);
2837 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2839 if (signal_pending_state(state, current)) {
2840 timeout = -ERESTARTSYS;
2843 __set_current_state(state);
2844 spin_unlock_irq(&x->wait.lock);
2845 timeout = action(timeout);
2846 spin_lock_irq(&x->wait.lock);
2847 } while (!x->done && timeout);
2848 __remove_wait_queue(&x->wait, &wait);
2853 return timeout ?: 1;
2856 static inline long __sched
2857 __wait_for_common(struct completion *x,
2858 long (*action)(long), long timeout, int state)
2862 spin_lock_irq(&x->wait.lock);
2863 timeout = do_wait_for_common(x, action, timeout, state);
2864 spin_unlock_irq(&x->wait.lock);
2869 wait_for_common(struct completion *x, long timeout, int state)
2871 return __wait_for_common(x, schedule_timeout, timeout, state);
2875 wait_for_common_io(struct completion *x, long timeout, int state)
2877 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2881 * wait_for_completion: - waits for completion of a task
2882 * @x: holds the state of this particular completion
2884 * This waits to be signaled for completion of a specific task. It is NOT
2885 * interruptible and there is no timeout.
2887 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2888 * and interrupt capability. Also see complete().
2890 void __sched wait_for_completion(struct completion *x)
2892 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2894 EXPORT_SYMBOL(wait_for_completion);
2897 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2898 * @x: holds the state of this particular completion
2899 * @timeout: timeout value in jiffies
2901 * This waits for either a completion of a specific task to be signaled or for a
2902 * specified timeout to expire. The timeout is in jiffies. It is not
2905 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2906 * till timeout) if completed.
2908 unsigned long __sched
2909 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2911 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2913 EXPORT_SYMBOL(wait_for_completion_timeout);
2916 * wait_for_completion_io: - waits for completion of a task
2917 * @x: holds the state of this particular completion
2919 * This waits to be signaled for completion of a specific task. It is NOT
2920 * interruptible and there is no timeout. The caller is accounted as waiting
2923 void __sched wait_for_completion_io(struct completion *x)
2925 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2927 EXPORT_SYMBOL(wait_for_completion_io);
2930 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2931 * @x: holds the state of this particular completion
2932 * @timeout: timeout value in jiffies
2934 * This waits for either a completion of a specific task to be signaled or for a
2935 * specified timeout to expire. The timeout is in jiffies. It is not
2936 * interruptible. The caller is accounted as waiting for IO.
2938 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2939 * till timeout) if completed.
2941 unsigned long __sched
2942 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2944 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2946 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2949 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2950 * @x: holds the state of this particular completion
2952 * This waits for completion of a specific task to be signaled. It is
2955 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2957 int __sched wait_for_completion_interruptible(struct completion *x)
2959 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2960 if (t == -ERESTARTSYS)
2964 EXPORT_SYMBOL(wait_for_completion_interruptible);
2967 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2968 * @x: holds the state of this particular completion
2969 * @timeout: timeout value in jiffies
2971 * This waits for either a completion of a specific task to be signaled or for a
2972 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2974 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2975 * or number of jiffies left till timeout) if completed.
2978 wait_for_completion_interruptible_timeout(struct completion *x,
2979 unsigned long timeout)
2981 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2983 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2986 * wait_for_completion_killable: - waits for completion of a task (killable)
2987 * @x: holds the state of this particular completion
2989 * This waits to be signaled for completion of a specific task. It can be
2990 * interrupted by a kill signal.
2992 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2994 int __sched wait_for_completion_killable(struct completion *x)
2996 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2997 if (t == -ERESTARTSYS)
3001 EXPORT_SYMBOL(wait_for_completion_killable);
3004 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3005 * @x: holds the state of this particular completion
3006 * @timeout: timeout value in jiffies
3008 * This waits for either a completion of a specific task to be
3009 * signaled or for a specified timeout to expire. It can be
3010 * interrupted by a kill signal. The timeout is in jiffies.
3012 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
3013 * or number of jiffies left till timeout) if completed.
3016 wait_for_completion_killable_timeout(struct completion *x,
3017 unsigned long timeout)
3019 return wait_for_common(x, timeout, TASK_KILLABLE);
3021 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3024 * try_wait_for_completion - try to decrement a completion without blocking
3025 * @x: completion structure
3027 * Return: 0 if a decrement cannot be done without blocking
3028 * 1 if a decrement succeeded.
3030 * If a completion is being used as a counting completion,
3031 * attempt to decrement the counter without blocking. This
3032 * enables us to avoid waiting if the resource the completion
3033 * is protecting is not available.
3035 bool try_wait_for_completion(struct completion *x)
3037 unsigned long flags;
3040 spin_lock_irqsave(&x->wait.lock, flags);
3045 spin_unlock_irqrestore(&x->wait.lock, flags);
3048 EXPORT_SYMBOL(try_wait_for_completion);
3051 * completion_done - Test to see if a completion has any waiters
3052 * @x: completion structure
3054 * Return: 0 if there are waiters (wait_for_completion() in progress)
3055 * 1 if there are no waiters.
3058 bool completion_done(struct completion *x)
3060 unsigned long flags;
3063 spin_lock_irqsave(&x->wait.lock, flags);
3066 spin_unlock_irqrestore(&x->wait.lock, flags);
3069 EXPORT_SYMBOL(completion_done);
3072 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3074 unsigned long flags;
3077 init_waitqueue_entry(&wait, current);
3079 __set_current_state(state);
3081 spin_lock_irqsave(&q->lock, flags);
3082 __add_wait_queue(q, &wait);
3083 spin_unlock(&q->lock);
3084 timeout = schedule_timeout(timeout);
3085 spin_lock_irq(&q->lock);
3086 __remove_wait_queue(q, &wait);
3087 spin_unlock_irqrestore(&q->lock, flags);
3092 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3094 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3096 EXPORT_SYMBOL(interruptible_sleep_on);
3099 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3101 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3103 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3105 void __sched sleep_on(wait_queue_head_t *q)
3107 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3109 EXPORT_SYMBOL(sleep_on);
3111 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3113 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3115 EXPORT_SYMBOL(sleep_on_timeout);
3117 #ifdef CONFIG_RT_MUTEXES
3120 * rt_mutex_setprio - set the current priority of a task
3122 * @prio: prio value (kernel-internal form)
3124 * This function changes the 'effective' priority of a task. It does
3125 * not touch ->normal_prio like __setscheduler().
3127 * Used by the rt_mutex code to implement priority inheritance logic.
3129 void rt_mutex_setprio(struct task_struct *p, int prio)
3131 int oldprio, on_rq, running;
3133 const struct sched_class *prev_class;
3135 BUG_ON(prio < 0 || prio > MAX_PRIO);
3137 rq = __task_rq_lock(p);
3140 * Idle task boosting is a nono in general. There is one
3141 * exception, when PREEMPT_RT and NOHZ is active:
3143 * The idle task calls get_next_timer_interrupt() and holds
3144 * the timer wheel base->lock on the CPU and another CPU wants
3145 * to access the timer (probably to cancel it). We can safely
3146 * ignore the boosting request, as the idle CPU runs this code
3147 * with interrupts disabled and will complete the lock
3148 * protected section without being interrupted. So there is no
3149 * real need to boost.
3151 if (unlikely(p == rq->idle)) {
3152 WARN_ON(p != rq->curr);
3153 WARN_ON(p->pi_blocked_on);
3157 trace_sched_pi_setprio(p, prio);
3159 prev_class = p->sched_class;
3161 running = task_current(rq, p);
3163 dequeue_task(rq, p, 0);
3165 p->sched_class->put_prev_task(rq, p);
3168 p->sched_class = &rt_sched_class;
3170 p->sched_class = &fair_sched_class;
3175 p->sched_class->set_curr_task(rq);
3177 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3179 check_class_changed(rq, p, prev_class, oldprio);
3181 __task_rq_unlock(rq);
3184 void set_user_nice(struct task_struct *p, long nice)
3186 int old_prio, delta, on_rq;
3187 unsigned long flags;
3190 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3193 * We have to be careful, if called from sys_setpriority(),
3194 * the task might be in the middle of scheduling on another CPU.
3196 rq = task_rq_lock(p, &flags);
3198 * The RT priorities are set via sched_setscheduler(), but we still
3199 * allow the 'normal' nice value to be set - but as expected
3200 * it wont have any effect on scheduling until the task is
3201 * SCHED_FIFO/SCHED_RR:
3203 if (task_has_rt_policy(p)) {
3204 p->static_prio = NICE_TO_PRIO(nice);
3209 dequeue_task(rq, p, 0);
3211 p->static_prio = NICE_TO_PRIO(nice);
3214 p->prio = effective_prio(p);
3215 delta = p->prio - old_prio;
3218 enqueue_task(rq, p, 0);
3220 * If the task increased its priority or is running and
3221 * lowered its priority, then reschedule its CPU:
3223 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3224 resched_task(rq->curr);
3227 task_rq_unlock(rq, p, &flags);
3229 EXPORT_SYMBOL(set_user_nice);
3232 * can_nice - check if a task can reduce its nice value
3236 int can_nice(const struct task_struct *p, const int nice)
3238 /* convert nice value [19,-20] to rlimit style value [1,40] */
3239 int nice_rlim = 20 - nice;
3241 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3242 capable(CAP_SYS_NICE));
3245 #ifdef __ARCH_WANT_SYS_NICE
3248 * sys_nice - change the priority of the current process.
3249 * @increment: priority increment
3251 * sys_setpriority is a more generic, but much slower function that
3252 * does similar things.
3254 SYSCALL_DEFINE1(nice, int, increment)
3259 * Setpriority might change our priority at the same moment.
3260 * We don't have to worry. Conceptually one call occurs first
3261 * and we have a single winner.
3263 if (increment < -40)
3268 nice = TASK_NICE(current) + increment;
3274 if (increment < 0 && !can_nice(current, nice))
3277 retval = security_task_setnice(current, nice);
3281 set_user_nice(current, nice);
3288 * task_prio - return the priority value of a given task.
3289 * @p: the task in question.
3291 * Return: The priority value as seen by users in /proc.
3292 * RT tasks are offset by -200. Normal tasks are centered
3293 * around 0, value goes from -16 to +15.
3295 int task_prio(const struct task_struct *p)
3297 return p->prio - MAX_RT_PRIO;
3301 * task_nice - return the nice value of a given task.
3302 * @p: the task in question.
3304 * Return: The nice value [ -20 ... 0 ... 19 ].
3306 int task_nice(const struct task_struct *p)
3308 return TASK_NICE(p);
3310 EXPORT_SYMBOL(task_nice);
3313 * idle_cpu - is a given cpu idle currently?
3314 * @cpu: the processor in question.
3316 * Return: 1 if the CPU is currently idle. 0 otherwise.
3318 int idle_cpu(int cpu)
3320 struct rq *rq = cpu_rq(cpu);
3322 if (rq->curr != rq->idle)
3329 if (!llist_empty(&rq->wake_list))
3337 * idle_task - return the idle task for a given cpu.
3338 * @cpu: the processor in question.
3340 * Return: The idle task for the cpu @cpu.
3342 struct task_struct *idle_task(int cpu)
3344 return cpu_rq(cpu)->idle;
3348 * find_process_by_pid - find a process with a matching PID value.
3349 * @pid: the pid in question.
3351 * The task of @pid, if found. %NULL otherwise.
3353 static struct task_struct *find_process_by_pid(pid_t pid)
3355 return pid ? find_task_by_vpid(pid) : current;
3358 /* Actually do priority change: must hold rq lock. */
3360 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3363 p->rt_priority = prio;
3364 p->normal_prio = normal_prio(p);
3365 /* we are holding p->pi_lock already */
3366 p->prio = rt_mutex_getprio(p);
3367 if (rt_prio(p->prio))
3368 p->sched_class = &rt_sched_class;
3370 p->sched_class = &fair_sched_class;
3375 * check the target process has a UID that matches the current process's
3377 static bool check_same_owner(struct task_struct *p)
3379 const struct cred *cred = current_cred(), *pcred;
3383 pcred = __task_cred(p);
3384 match = (uid_eq(cred->euid, pcred->euid) ||
3385 uid_eq(cred->euid, pcred->uid));
3390 static int __sched_setscheduler(struct task_struct *p, int policy,
3391 const struct sched_param *param, bool user)
3393 int retval, oldprio, oldpolicy = -1, on_rq, running;
3394 unsigned long flags;
3395 const struct sched_class *prev_class;
3399 /* may grab non-irq protected spin_locks */
3400 BUG_ON(in_interrupt());
3402 /* double check policy once rq lock held */
3404 reset_on_fork = p->sched_reset_on_fork;
3405 policy = oldpolicy = p->policy;
3407 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3408 policy &= ~SCHED_RESET_ON_FORK;
3410 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3411 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3412 policy != SCHED_IDLE)
3417 * Valid priorities for SCHED_FIFO and SCHED_RR are
3418 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3419 * SCHED_BATCH and SCHED_IDLE is 0.
3421 if (param->sched_priority < 0 ||
3422 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3423 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3425 if (rt_policy(policy) != (param->sched_priority != 0))
3429 * Allow unprivileged RT tasks to decrease priority:
3431 if (user && !capable(CAP_SYS_NICE)) {
3432 if (rt_policy(policy)) {
3433 unsigned long rlim_rtprio =
3434 task_rlimit(p, RLIMIT_RTPRIO);
3436 /* can't set/change the rt policy */
3437 if (policy != p->policy && !rlim_rtprio)
3440 /* can't increase priority */
3441 if (param->sched_priority > p->rt_priority &&
3442 param->sched_priority > rlim_rtprio)
3447 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3448 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3450 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3451 if (!can_nice(p, TASK_NICE(p)))
3455 /* can't change other user's priorities */
3456 if (!check_same_owner(p))
3459 /* Normal users shall not reset the sched_reset_on_fork flag */
3460 if (p->sched_reset_on_fork && !reset_on_fork)
3465 retval = security_task_setscheduler(p);
3471 * make sure no PI-waiters arrive (or leave) while we are
3472 * changing the priority of the task:
3474 * To be able to change p->policy safely, the appropriate
3475 * runqueue lock must be held.
3477 rq = task_rq_lock(p, &flags);
3480 * Changing the policy of the stop threads its a very bad idea
3482 if (p == rq->stop) {
3483 task_rq_unlock(rq, p, &flags);
3488 * If not changing anything there's no need to proceed further:
3490 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3491 param->sched_priority == p->rt_priority))) {
3492 task_rq_unlock(rq, p, &flags);
3496 #ifdef CONFIG_RT_GROUP_SCHED
3499 * Do not allow realtime tasks into groups that have no runtime
3502 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3503 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3504 !task_group_is_autogroup(task_group(p))) {
3505 task_rq_unlock(rq, p, &flags);
3511 /* recheck policy now with rq lock held */
3512 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3513 policy = oldpolicy = -1;
3514 task_rq_unlock(rq, p, &flags);
3518 running = task_current(rq, p);
3520 dequeue_task(rq, p, 0);
3522 p->sched_class->put_prev_task(rq, p);
3524 p->sched_reset_on_fork = reset_on_fork;
3527 prev_class = p->sched_class;
3528 __setscheduler(rq, p, policy, param->sched_priority);
3531 p->sched_class->set_curr_task(rq);
3533 enqueue_task(rq, p, 0);
3535 check_class_changed(rq, p, prev_class, oldprio);
3536 task_rq_unlock(rq, p, &flags);
3538 rt_mutex_adjust_pi(p);
3544 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3545 * @p: the task in question.
3546 * @policy: new policy.
3547 * @param: structure containing the new RT priority.
3549 * Return: 0 on success. An error code otherwise.
3551 * NOTE that the task may be already dead.
3553 int sched_setscheduler(struct task_struct *p, int policy,
3554 const struct sched_param *param)
3556 return __sched_setscheduler(p, policy, param, true);
3558 EXPORT_SYMBOL_GPL(sched_setscheduler);
3561 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3562 * @p: the task in question.
3563 * @policy: new policy.
3564 * @param: structure containing the new RT priority.
3566 * Just like sched_setscheduler, only don't bother checking if the
3567 * current context has permission. For example, this is needed in
3568 * stop_machine(): we create temporary high priority worker threads,
3569 * but our caller might not have that capability.
3571 * Return: 0 on success. An error code otherwise.
3573 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3574 const struct sched_param *param)
3576 return __sched_setscheduler(p, policy, param, false);
3580 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3582 struct sched_param lparam;
3583 struct task_struct *p;
3586 if (!param || pid < 0)
3588 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3593 p = find_process_by_pid(pid);
3595 retval = sched_setscheduler(p, policy, &lparam);
3602 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3603 * @pid: the pid in question.
3604 * @policy: new policy.
3605 * @param: structure containing the new RT priority.
3607 * Return: 0 on success. An error code otherwise.
3609 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3610 struct sched_param __user *, param)
3612 /* negative values for policy are not valid */
3616 return do_sched_setscheduler(pid, policy, param);
3620 * sys_sched_setparam - set/change the RT priority of a thread
3621 * @pid: the pid in question.
3622 * @param: structure containing the new RT priority.
3624 * Return: 0 on success. An error code otherwise.
3626 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3628 return do_sched_setscheduler(pid, -1, param);
3632 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3633 * @pid: the pid in question.
3635 * Return: On success, the policy of the thread. Otherwise, a negative error
3638 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3640 struct task_struct *p;
3648 p = find_process_by_pid(pid);
3650 retval = security_task_getscheduler(p);
3653 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3660 * sys_sched_getparam - get the RT priority of a thread
3661 * @pid: the pid in question.
3662 * @param: structure containing the RT priority.
3664 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3667 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3669 struct sched_param lp;
3670 struct task_struct *p;
3673 if (!param || pid < 0)
3677 p = find_process_by_pid(pid);
3682 retval = security_task_getscheduler(p);
3686 lp.sched_priority = p->rt_priority;
3690 * This one might sleep, we cannot do it with a spinlock held ...
3692 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3701 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3703 cpumask_var_t cpus_allowed, new_mask;
3704 struct task_struct *p;
3710 p = find_process_by_pid(pid);
3717 /* Prevent p going away */
3721 if (p->flags & PF_NO_SETAFFINITY) {
3725 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3729 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3731 goto out_free_cpus_allowed;
3734 if (!check_same_owner(p)) {
3736 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3743 retval = security_task_setscheduler(p);
3747 cpuset_cpus_allowed(p, cpus_allowed);
3748 cpumask_and(new_mask, in_mask, cpus_allowed);
3750 retval = set_cpus_allowed_ptr(p, new_mask);
3753 cpuset_cpus_allowed(p, cpus_allowed);
3754 if (!cpumask_subset(new_mask, cpus_allowed)) {
3756 * We must have raced with a concurrent cpuset
3757 * update. Just reset the cpus_allowed to the
3758 * cpuset's cpus_allowed
3760 cpumask_copy(new_mask, cpus_allowed);
3765 free_cpumask_var(new_mask);
3766 out_free_cpus_allowed:
3767 free_cpumask_var(cpus_allowed);
3774 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3775 struct cpumask *new_mask)
3777 if (len < cpumask_size())
3778 cpumask_clear(new_mask);
3779 else if (len > cpumask_size())
3780 len = cpumask_size();
3782 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3786 * sys_sched_setaffinity - set the cpu affinity of a process
3787 * @pid: pid of the process
3788 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3789 * @user_mask_ptr: user-space pointer to the new cpu mask
3791 * Return: 0 on success. An error code otherwise.
3793 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3794 unsigned long __user *, user_mask_ptr)
3796 cpumask_var_t new_mask;
3799 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3802 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3804 retval = sched_setaffinity(pid, new_mask);
3805 free_cpumask_var(new_mask);
3809 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3811 struct task_struct *p;
3812 unsigned long flags;
3819 p = find_process_by_pid(pid);
3823 retval = security_task_getscheduler(p);
3827 raw_spin_lock_irqsave(&p->pi_lock, flags);
3828 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3829 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3839 * sys_sched_getaffinity - get the cpu affinity of a process
3840 * @pid: pid of the process
3841 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3842 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3844 * Return: 0 on success. An error code otherwise.
3846 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3847 unsigned long __user *, user_mask_ptr)
3852 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3854 if (len & (sizeof(unsigned long)-1))
3857 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3860 ret = sched_getaffinity(pid, mask);
3862 size_t retlen = min_t(size_t, len, cpumask_size());
3864 if (copy_to_user(user_mask_ptr, mask, retlen))
3869 free_cpumask_var(mask);
3875 * sys_sched_yield - yield the current processor to other threads.
3877 * This function yields the current CPU to other tasks. If there are no
3878 * other threads running on this CPU then this function will return.
3882 SYSCALL_DEFINE0(sched_yield)
3884 struct rq *rq = this_rq_lock();
3886 schedstat_inc(rq, yld_count);
3887 current->sched_class->yield_task(rq);
3890 * Since we are going to call schedule() anyway, there's
3891 * no need to preempt or enable interrupts:
3893 __release(rq->lock);
3894 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3895 do_raw_spin_unlock(&rq->lock);
3896 sched_preempt_enable_no_resched();
3903 static void __cond_resched(void)
3905 __preempt_count_add(PREEMPT_ACTIVE);
3907 __preempt_count_sub(PREEMPT_ACTIVE);
3910 int __sched _cond_resched(void)
3912 if (should_resched()) {
3918 EXPORT_SYMBOL(_cond_resched);
3921 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3922 * call schedule, and on return reacquire the lock.
3924 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3925 * operations here to prevent schedule() from being called twice (once via
3926 * spin_unlock(), once by hand).
3928 int __cond_resched_lock(spinlock_t *lock)
3930 int resched = should_resched();
3933 lockdep_assert_held(lock);
3935 if (spin_needbreak(lock) || resched) {
3946 EXPORT_SYMBOL(__cond_resched_lock);
3948 int __sched __cond_resched_softirq(void)
3950 BUG_ON(!in_softirq());
3952 if (should_resched()) {
3960 EXPORT_SYMBOL(__cond_resched_softirq);
3963 * yield - yield the current processor to other threads.
3965 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3967 * The scheduler is at all times free to pick the calling task as the most
3968 * eligible task to run, if removing the yield() call from your code breaks
3969 * it, its already broken.
3971 * Typical broken usage is:
3976 * where one assumes that yield() will let 'the other' process run that will
3977 * make event true. If the current task is a SCHED_FIFO task that will never
3978 * happen. Never use yield() as a progress guarantee!!
3980 * If you want to use yield() to wait for something, use wait_event().
3981 * If you want to use yield() to be 'nice' for others, use cond_resched().
3982 * If you still want to use yield(), do not!
3984 void __sched yield(void)
3986 set_current_state(TASK_RUNNING);
3989 EXPORT_SYMBOL(yield);
3992 * yield_to - yield the current processor to another thread in
3993 * your thread group, or accelerate that thread toward the
3994 * processor it's on.
3996 * @preempt: whether task preemption is allowed or not
3998 * It's the caller's job to ensure that the target task struct
3999 * can't go away on us before we can do any checks.
4002 * true (>0) if we indeed boosted the target task.
4003 * false (0) if we failed to boost the target.
4004 * -ESRCH if there's no task to yield to.
4006 bool __sched yield_to(struct task_struct *p, bool preempt)
4008 struct task_struct *curr = current;
4009 struct rq *rq, *p_rq;
4010 unsigned long flags;
4013 local_irq_save(flags);
4019 * If we're the only runnable task on the rq and target rq also
4020 * has only one task, there's absolutely no point in yielding.
4022 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4027 double_rq_lock(rq, p_rq);
4028 while (task_rq(p) != p_rq) {
4029 double_rq_unlock(rq, p_rq);
4033 if (!curr->sched_class->yield_to_task)
4036 if (curr->sched_class != p->sched_class)
4039 if (task_running(p_rq, p) || p->state)
4042 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4044 schedstat_inc(rq, yld_count);
4046 * Make p's CPU reschedule; pick_next_entity takes care of
4049 if (preempt && rq != p_rq)
4050 resched_task(p_rq->curr);
4054 double_rq_unlock(rq, p_rq);
4056 local_irq_restore(flags);
4063 EXPORT_SYMBOL_GPL(yield_to);
4066 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4067 * that process accounting knows that this is a task in IO wait state.
4069 void __sched io_schedule(void)
4071 struct rq *rq = raw_rq();
4073 delayacct_blkio_start();
4074 atomic_inc(&rq->nr_iowait);
4075 blk_flush_plug(current);
4076 current->in_iowait = 1;
4078 current->in_iowait = 0;
4079 atomic_dec(&rq->nr_iowait);
4080 delayacct_blkio_end();
4082 EXPORT_SYMBOL(io_schedule);
4084 long __sched io_schedule_timeout(long timeout)
4086 struct rq *rq = raw_rq();
4089 delayacct_blkio_start();
4090 atomic_inc(&rq->nr_iowait);
4091 blk_flush_plug(current);
4092 current->in_iowait = 1;
4093 ret = schedule_timeout(timeout);
4094 current->in_iowait = 0;
4095 atomic_dec(&rq->nr_iowait);
4096 delayacct_blkio_end();
4101 * sys_sched_get_priority_max - return maximum RT priority.
4102 * @policy: scheduling class.
4104 * Return: On success, this syscall returns the maximum
4105 * rt_priority that can be used by a given scheduling class.
4106 * On failure, a negative error code is returned.
4108 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4115 ret = MAX_USER_RT_PRIO-1;
4127 * sys_sched_get_priority_min - return minimum RT priority.
4128 * @policy: scheduling class.
4130 * Return: On success, this syscall returns the minimum
4131 * rt_priority that can be used by a given scheduling class.
4132 * On failure, a negative error code is returned.
4134 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4152 * sys_sched_rr_get_interval - return the default timeslice of a process.
4153 * @pid: pid of the process.
4154 * @interval: userspace pointer to the timeslice value.
4156 * this syscall writes the default timeslice value of a given process
4157 * into the user-space timespec buffer. A value of '0' means infinity.
4159 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4162 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4163 struct timespec __user *, interval)
4165 struct task_struct *p;
4166 unsigned int time_slice;
4167 unsigned long flags;
4177 p = find_process_by_pid(pid);
4181 retval = security_task_getscheduler(p);
4185 rq = task_rq_lock(p, &flags);
4186 time_slice = p->sched_class->get_rr_interval(rq, p);
4187 task_rq_unlock(rq, p, &flags);
4190 jiffies_to_timespec(time_slice, &t);
4191 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4199 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4201 void sched_show_task(struct task_struct *p)
4203 unsigned long free = 0;
4207 state = p->state ? __ffs(p->state) + 1 : 0;
4208 printk(KERN_INFO "%-15.15s %c", p->comm,
4209 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4210 #if BITS_PER_LONG == 32
4211 if (state == TASK_RUNNING)
4212 printk(KERN_CONT " running ");
4214 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4216 if (state == TASK_RUNNING)
4217 printk(KERN_CONT " running task ");
4219 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4221 #ifdef CONFIG_DEBUG_STACK_USAGE
4222 free = stack_not_used(p);
4225 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4227 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4228 task_pid_nr(p), ppid,
4229 (unsigned long)task_thread_info(p)->flags);
4231 print_worker_info(KERN_INFO, p);
4232 show_stack(p, NULL);
4235 void show_state_filter(unsigned long state_filter)
4237 struct task_struct *g, *p;
4239 #if BITS_PER_LONG == 32
4241 " task PC stack pid father\n");
4244 " task PC stack pid father\n");
4247 do_each_thread(g, p) {
4249 * reset the NMI-timeout, listing all files on a slow
4250 * console might take a lot of time:
4252 touch_nmi_watchdog();
4253 if (!state_filter || (p->state & state_filter))
4255 } while_each_thread(g, p);
4257 touch_all_softlockup_watchdogs();
4259 #ifdef CONFIG_SCHED_DEBUG
4260 sysrq_sched_debug_show();
4264 * Only show locks if all tasks are dumped:
4267 debug_show_all_locks();
4270 void init_idle_bootup_task(struct task_struct *idle)
4272 idle->sched_class = &idle_sched_class;
4276 * init_idle - set up an idle thread for a given CPU
4277 * @idle: task in question
4278 * @cpu: cpu the idle task belongs to
4280 * NOTE: this function does not set the idle thread's NEED_RESCHED
4281 * flag, to make booting more robust.
4283 void init_idle(struct task_struct *idle, int cpu)
4285 struct rq *rq = cpu_rq(cpu);
4286 unsigned long flags;
4288 raw_spin_lock_irqsave(&rq->lock, flags);
4291 idle->state = TASK_RUNNING;
4292 idle->se.exec_start = sched_clock();
4294 do_set_cpus_allowed(idle, cpumask_of(cpu));
4296 * We're having a chicken and egg problem, even though we are
4297 * holding rq->lock, the cpu isn't yet set to this cpu so the
4298 * lockdep check in task_group() will fail.
4300 * Similar case to sched_fork(). / Alternatively we could
4301 * use task_rq_lock() here and obtain the other rq->lock.
4306 __set_task_cpu(idle, cpu);
4309 rq->curr = rq->idle = idle;
4310 #if defined(CONFIG_SMP)
4313 raw_spin_unlock_irqrestore(&rq->lock, flags);
4315 /* Set the preempt count _outside_ the spinlocks! */
4316 init_idle_preempt_count(idle, cpu);
4319 * The idle tasks have their own, simple scheduling class:
4321 idle->sched_class = &idle_sched_class;
4322 ftrace_graph_init_idle_task(idle, cpu);
4323 vtime_init_idle(idle, cpu);
4324 #if defined(CONFIG_SMP)
4325 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4330 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4332 if (p->sched_class && p->sched_class->set_cpus_allowed)
4333 p->sched_class->set_cpus_allowed(p, new_mask);
4335 cpumask_copy(&p->cpus_allowed, new_mask);
4336 p->nr_cpus_allowed = cpumask_weight(new_mask);
4340 * This is how migration works:
4342 * 1) we invoke migration_cpu_stop() on the target CPU using
4344 * 2) stopper starts to run (implicitly forcing the migrated thread
4346 * 3) it checks whether the migrated task is still in the wrong runqueue.
4347 * 4) if it's in the wrong runqueue then the migration thread removes
4348 * it and puts it into the right queue.
4349 * 5) stopper completes and stop_one_cpu() returns and the migration
4354 * Change a given task's CPU affinity. Migrate the thread to a
4355 * proper CPU and schedule it away if the CPU it's executing on
4356 * is removed from the allowed bitmask.
4358 * NOTE: the caller must have a valid reference to the task, the
4359 * task must not exit() & deallocate itself prematurely. The
4360 * call is not atomic; no spinlocks may be held.
4362 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4364 unsigned long flags;
4366 unsigned int dest_cpu;
4369 rq = task_rq_lock(p, &flags);
4371 if (cpumask_equal(&p->cpus_allowed, new_mask))
4374 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4379 do_set_cpus_allowed(p, new_mask);
4381 /* Can the task run on the task's current CPU? If so, we're done */
4382 if (cpumask_test_cpu(task_cpu(p), new_mask))
4385 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4387 struct migration_arg arg = { p, dest_cpu };
4388 /* Need help from migration thread: drop lock and wait. */
4389 task_rq_unlock(rq, p, &flags);
4390 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4391 tlb_migrate_finish(p->mm);
4395 task_rq_unlock(rq, p, &flags);
4399 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4402 * Move (not current) task off this cpu, onto dest cpu. We're doing
4403 * this because either it can't run here any more (set_cpus_allowed()
4404 * away from this CPU, or CPU going down), or because we're
4405 * attempting to rebalance this task on exec (sched_exec).
4407 * So we race with normal scheduler movements, but that's OK, as long
4408 * as the task is no longer on this CPU.
4410 * Returns non-zero if task was successfully migrated.
4412 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4414 struct rq *rq_dest, *rq_src;
4417 if (unlikely(!cpu_active(dest_cpu)))
4420 rq_src = cpu_rq(src_cpu);
4421 rq_dest = cpu_rq(dest_cpu);
4423 raw_spin_lock(&p->pi_lock);
4424 double_rq_lock(rq_src, rq_dest);
4425 /* Already moved. */
4426 if (task_cpu(p) != src_cpu)
4428 /* Affinity changed (again). */
4429 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4433 * If we're not on a rq, the next wake-up will ensure we're
4437 dequeue_task(rq_src, p, 0);
4438 set_task_cpu(p, dest_cpu);
4439 enqueue_task(rq_dest, p, 0);
4440 check_preempt_curr(rq_dest, p, 0);
4445 double_rq_unlock(rq_src, rq_dest);
4446 raw_spin_unlock(&p->pi_lock);
4450 #ifdef CONFIG_NUMA_BALANCING
4451 /* Migrate current task p to target_cpu */
4452 int migrate_task_to(struct task_struct *p, int target_cpu)
4454 struct migration_arg arg = { p, target_cpu };
4455 int curr_cpu = task_cpu(p);
4457 if (curr_cpu == target_cpu)
4460 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4463 /* TODO: This is not properly updating schedstats */
4465 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4470 * migration_cpu_stop - this will be executed by a highprio stopper thread
4471 * and performs thread migration by bumping thread off CPU then
4472 * 'pushing' onto another runqueue.
4474 static int migration_cpu_stop(void *data)
4476 struct migration_arg *arg = data;
4479 * The original target cpu might have gone down and we might
4480 * be on another cpu but it doesn't matter.
4482 local_irq_disable();
4483 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4488 #ifdef CONFIG_HOTPLUG_CPU
4491 * Ensures that the idle task is using init_mm right before its cpu goes
4494 void idle_task_exit(void)
4496 struct mm_struct *mm = current->active_mm;
4498 BUG_ON(cpu_online(smp_processor_id()));
4501 switch_mm(mm, &init_mm, current);
4506 * Since this CPU is going 'away' for a while, fold any nr_active delta
4507 * we might have. Assumes we're called after migrate_tasks() so that the
4508 * nr_active count is stable.
4510 * Also see the comment "Global load-average calculations".
4512 static void calc_load_migrate(struct rq *rq)
4514 long delta = calc_load_fold_active(rq);
4516 atomic_long_add(delta, &calc_load_tasks);
4520 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4521 * try_to_wake_up()->select_task_rq().
4523 * Called with rq->lock held even though we'er in stop_machine() and
4524 * there's no concurrency possible, we hold the required locks anyway
4525 * because of lock validation efforts.
4527 static void migrate_tasks(unsigned int dead_cpu)
4529 struct rq *rq = cpu_rq(dead_cpu);
4530 struct task_struct *next, *stop = rq->stop;
4534 * Fudge the rq selection such that the below task selection loop
4535 * doesn't get stuck on the currently eligible stop task.
4537 * We're currently inside stop_machine() and the rq is either stuck
4538 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4539 * either way we should never end up calling schedule() until we're
4545 * put_prev_task() and pick_next_task() sched
4546 * class method both need to have an up-to-date
4547 * value of rq->clock[_task]
4549 update_rq_clock(rq);
4553 * There's this thread running, bail when that's the only
4556 if (rq->nr_running == 1)
4559 next = pick_next_task(rq);
4561 next->sched_class->put_prev_task(rq, next);
4563 /* Find suitable destination for @next, with force if needed. */
4564 dest_cpu = select_fallback_rq(dead_cpu, next);
4565 raw_spin_unlock(&rq->lock);
4567 __migrate_task(next, dead_cpu, dest_cpu);
4569 raw_spin_lock(&rq->lock);
4575 #endif /* CONFIG_HOTPLUG_CPU */
4577 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4579 static struct ctl_table sd_ctl_dir[] = {
4581 .procname = "sched_domain",
4587 static struct ctl_table sd_ctl_root[] = {
4589 .procname = "kernel",
4591 .child = sd_ctl_dir,
4596 static struct ctl_table *sd_alloc_ctl_entry(int n)
4598 struct ctl_table *entry =
4599 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4604 static void sd_free_ctl_entry(struct ctl_table **tablep)
4606 struct ctl_table *entry;
4609 * In the intermediate directories, both the child directory and
4610 * procname are dynamically allocated and could fail but the mode
4611 * will always be set. In the lowest directory the names are
4612 * static strings and all have proc handlers.
4614 for (entry = *tablep; entry->mode; entry++) {
4616 sd_free_ctl_entry(&entry->child);
4617 if (entry->proc_handler == NULL)
4618 kfree(entry->procname);
4625 static int min_load_idx = 0;
4626 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4629 set_table_entry(struct ctl_table *entry,
4630 const char *procname, void *data, int maxlen,
4631 umode_t mode, proc_handler *proc_handler,
4634 entry->procname = procname;
4636 entry->maxlen = maxlen;
4638 entry->proc_handler = proc_handler;
4641 entry->extra1 = &min_load_idx;
4642 entry->extra2 = &max_load_idx;
4646 static struct ctl_table *
4647 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4649 struct ctl_table *table = sd_alloc_ctl_entry(13);
4654 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4655 sizeof(long), 0644, proc_doulongvec_minmax, false);
4656 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4657 sizeof(long), 0644, proc_doulongvec_minmax, false);
4658 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4659 sizeof(int), 0644, proc_dointvec_minmax, true);
4660 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4661 sizeof(int), 0644, proc_dointvec_minmax, true);
4662 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4663 sizeof(int), 0644, proc_dointvec_minmax, true);
4664 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4665 sizeof(int), 0644, proc_dointvec_minmax, true);
4666 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4667 sizeof(int), 0644, proc_dointvec_minmax, true);
4668 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4669 sizeof(int), 0644, proc_dointvec_minmax, false);
4670 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4671 sizeof(int), 0644, proc_dointvec_minmax, false);
4672 set_table_entry(&table[9], "cache_nice_tries",
4673 &sd->cache_nice_tries,
4674 sizeof(int), 0644, proc_dointvec_minmax, false);
4675 set_table_entry(&table[10], "flags", &sd->flags,
4676 sizeof(int), 0644, proc_dointvec_minmax, false);
4677 set_table_entry(&table[11], "name", sd->name,
4678 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4679 /* &table[12] is terminator */
4684 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4686 struct ctl_table *entry, *table;
4687 struct sched_domain *sd;
4688 int domain_num = 0, i;
4691 for_each_domain(cpu, sd)
4693 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4698 for_each_domain(cpu, sd) {
4699 snprintf(buf, 32, "domain%d", i);
4700 entry->procname = kstrdup(buf, GFP_KERNEL);
4702 entry->child = sd_alloc_ctl_domain_table(sd);
4709 static struct ctl_table_header *sd_sysctl_header;
4710 static void register_sched_domain_sysctl(void)
4712 int i, cpu_num = num_possible_cpus();
4713 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4716 WARN_ON(sd_ctl_dir[0].child);
4717 sd_ctl_dir[0].child = entry;
4722 for_each_possible_cpu(i) {
4723 snprintf(buf, 32, "cpu%d", i);
4724 entry->procname = kstrdup(buf, GFP_KERNEL);
4726 entry->child = sd_alloc_ctl_cpu_table(i);
4730 WARN_ON(sd_sysctl_header);
4731 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4734 /* may be called multiple times per register */
4735 static void unregister_sched_domain_sysctl(void)
4737 if (sd_sysctl_header)
4738 unregister_sysctl_table(sd_sysctl_header);
4739 sd_sysctl_header = NULL;
4740 if (sd_ctl_dir[0].child)
4741 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4744 static void register_sched_domain_sysctl(void)
4747 static void unregister_sched_domain_sysctl(void)
4752 static void set_rq_online(struct rq *rq)
4755 const struct sched_class *class;
4757 cpumask_set_cpu(rq->cpu, rq->rd->online);
4760 for_each_class(class) {
4761 if (class->rq_online)
4762 class->rq_online(rq);
4767 static void set_rq_offline(struct rq *rq)
4770 const struct sched_class *class;
4772 for_each_class(class) {
4773 if (class->rq_offline)
4774 class->rq_offline(rq);
4777 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4783 * migration_call - callback that gets triggered when a CPU is added.
4784 * Here we can start up the necessary migration thread for the new CPU.
4787 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4789 int cpu = (long)hcpu;
4790 unsigned long flags;
4791 struct rq *rq = cpu_rq(cpu);
4793 switch (action & ~CPU_TASKS_FROZEN) {
4795 case CPU_UP_PREPARE:
4796 rq->calc_load_update = calc_load_update;
4800 /* Update our root-domain */
4801 raw_spin_lock_irqsave(&rq->lock, flags);
4803 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4807 raw_spin_unlock_irqrestore(&rq->lock, flags);
4810 #ifdef CONFIG_HOTPLUG_CPU
4812 sched_ttwu_pending();
4813 /* Update our root-domain */
4814 raw_spin_lock_irqsave(&rq->lock, flags);
4816 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4820 BUG_ON(rq->nr_running != 1); /* the migration thread */
4821 raw_spin_unlock_irqrestore(&rq->lock, flags);
4825 calc_load_migrate(rq);
4830 update_max_interval();
4836 * Register at high priority so that task migration (migrate_all_tasks)
4837 * happens before everything else. This has to be lower priority than
4838 * the notifier in the perf_event subsystem, though.
4840 static struct notifier_block migration_notifier = {
4841 .notifier_call = migration_call,
4842 .priority = CPU_PRI_MIGRATION,
4845 static int sched_cpu_active(struct notifier_block *nfb,
4846 unsigned long action, void *hcpu)
4848 switch (action & ~CPU_TASKS_FROZEN) {
4850 case CPU_DOWN_FAILED:
4851 set_cpu_active((long)hcpu, true);
4858 static int sched_cpu_inactive(struct notifier_block *nfb,
4859 unsigned long action, void *hcpu)
4861 switch (action & ~CPU_TASKS_FROZEN) {
4862 case CPU_DOWN_PREPARE:
4863 set_cpu_active((long)hcpu, false);
4870 static int __init migration_init(void)
4872 void *cpu = (void *)(long)smp_processor_id();
4875 /* Initialize migration for the boot CPU */
4876 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4877 BUG_ON(err == NOTIFY_BAD);
4878 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4879 register_cpu_notifier(&migration_notifier);
4881 /* Register cpu active notifiers */
4882 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4883 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4887 early_initcall(migration_init);
4892 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4894 #ifdef CONFIG_SCHED_DEBUG
4896 static __read_mostly int sched_debug_enabled;
4898 static int __init sched_debug_setup(char *str)
4900 sched_debug_enabled = 1;
4904 early_param("sched_debug", sched_debug_setup);
4906 static inline bool sched_debug(void)
4908 return sched_debug_enabled;
4911 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4912 struct cpumask *groupmask)
4914 struct sched_group *group = sd->groups;
4917 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4918 cpumask_clear(groupmask);
4920 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4922 if (!(sd->flags & SD_LOAD_BALANCE)) {
4923 printk("does not load-balance\n");
4925 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4930 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4932 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4933 printk(KERN_ERR "ERROR: domain->span does not contain "
4936 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4937 printk(KERN_ERR "ERROR: domain->groups does not contain"
4941 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4945 printk(KERN_ERR "ERROR: group is NULL\n");
4950 * Even though we initialize ->power to something semi-sane,
4951 * we leave power_orig unset. This allows us to detect if
4952 * domain iteration is still funny without causing /0 traps.
4954 if (!group->sgp->power_orig) {
4955 printk(KERN_CONT "\n");
4956 printk(KERN_ERR "ERROR: domain->cpu_power not "
4961 if (!cpumask_weight(sched_group_cpus(group))) {
4962 printk(KERN_CONT "\n");
4963 printk(KERN_ERR "ERROR: empty group\n");
4967 if (!(sd->flags & SD_OVERLAP) &&
4968 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4969 printk(KERN_CONT "\n");
4970 printk(KERN_ERR "ERROR: repeated CPUs\n");
4974 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4976 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4978 printk(KERN_CONT " %s", str);
4979 if (group->sgp->power != SCHED_POWER_SCALE) {
4980 printk(KERN_CONT " (cpu_power = %d)",
4984 group = group->next;
4985 } while (group != sd->groups);
4986 printk(KERN_CONT "\n");
4988 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4989 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4992 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4993 printk(KERN_ERR "ERROR: parent span is not a superset "
4994 "of domain->span\n");
4998 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5002 if (!sched_debug_enabled)
5006 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5010 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5013 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5021 #else /* !CONFIG_SCHED_DEBUG */
5022 # define sched_domain_debug(sd, cpu) do { } while (0)
5023 static inline bool sched_debug(void)
5027 #endif /* CONFIG_SCHED_DEBUG */
5029 static int sd_degenerate(struct sched_domain *sd)
5031 if (cpumask_weight(sched_domain_span(sd)) == 1)
5034 /* Following flags need at least 2 groups */
5035 if (sd->flags & (SD_LOAD_BALANCE |
5036 SD_BALANCE_NEWIDLE |
5040 SD_SHARE_PKG_RESOURCES)) {
5041 if (sd->groups != sd->groups->next)
5045 /* Following flags don't use groups */
5046 if (sd->flags & (SD_WAKE_AFFINE))
5053 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5055 unsigned long cflags = sd->flags, pflags = parent->flags;
5057 if (sd_degenerate(parent))
5060 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5063 /* Flags needing groups don't count if only 1 group in parent */
5064 if (parent->groups == parent->groups->next) {
5065 pflags &= ~(SD_LOAD_BALANCE |
5066 SD_BALANCE_NEWIDLE |
5070 SD_SHARE_PKG_RESOURCES |
5072 if (nr_node_ids == 1)
5073 pflags &= ~SD_SERIALIZE;
5075 if (~cflags & pflags)
5081 static void free_rootdomain(struct rcu_head *rcu)
5083 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5085 cpupri_cleanup(&rd->cpupri);
5086 free_cpumask_var(rd->rto_mask);
5087 free_cpumask_var(rd->online);
5088 free_cpumask_var(rd->span);
5092 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5094 struct root_domain *old_rd = NULL;
5095 unsigned long flags;
5097 raw_spin_lock_irqsave(&rq->lock, flags);
5102 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5105 cpumask_clear_cpu(rq->cpu, old_rd->span);
5108 * If we dont want to free the old_rt yet then
5109 * set old_rd to NULL to skip the freeing later
5112 if (!atomic_dec_and_test(&old_rd->refcount))
5116 atomic_inc(&rd->refcount);
5119 cpumask_set_cpu(rq->cpu, rd->span);
5120 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5123 raw_spin_unlock_irqrestore(&rq->lock, flags);
5126 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5129 static int init_rootdomain(struct root_domain *rd)
5131 memset(rd, 0, sizeof(*rd));
5133 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5135 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5137 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5140 if (cpupri_init(&rd->cpupri) != 0)
5145 free_cpumask_var(rd->rto_mask);
5147 free_cpumask_var(rd->online);
5149 free_cpumask_var(rd->span);
5155 * By default the system creates a single root-domain with all cpus as
5156 * members (mimicking the global state we have today).
5158 struct root_domain def_root_domain;
5160 static void init_defrootdomain(void)
5162 init_rootdomain(&def_root_domain);
5164 atomic_set(&def_root_domain.refcount, 1);
5167 static struct root_domain *alloc_rootdomain(void)
5169 struct root_domain *rd;
5171 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5175 if (init_rootdomain(rd) != 0) {
5183 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5185 struct sched_group *tmp, *first;
5194 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5199 } while (sg != first);
5202 static void free_sched_domain(struct rcu_head *rcu)
5204 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5207 * If its an overlapping domain it has private groups, iterate and
5210 if (sd->flags & SD_OVERLAP) {
5211 free_sched_groups(sd->groups, 1);
5212 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5213 kfree(sd->groups->sgp);
5219 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5221 call_rcu(&sd->rcu, free_sched_domain);
5224 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5226 for (; sd; sd = sd->parent)
5227 destroy_sched_domain(sd, cpu);
5231 * Keep a special pointer to the highest sched_domain that has
5232 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5233 * allows us to avoid some pointer chasing select_idle_sibling().
5235 * Also keep a unique ID per domain (we use the first cpu number in
5236 * the cpumask of the domain), this allows us to quickly tell if
5237 * two cpus are in the same cache domain, see cpus_share_cache().
5239 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5240 DEFINE_PER_CPU(int, sd_llc_size);
5241 DEFINE_PER_CPU(int, sd_llc_id);
5242 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5244 static void update_top_cache_domain(int cpu)
5246 struct sched_domain *sd;
5250 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5252 id = cpumask_first(sched_domain_span(sd));
5253 size = cpumask_weight(sched_domain_span(sd));
5256 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5257 per_cpu(sd_llc_size, cpu) = size;
5258 per_cpu(sd_llc_id, cpu) = id;
5260 sd = lowest_flag_domain(cpu, SD_NUMA);
5261 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5265 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5266 * hold the hotplug lock.
5269 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5271 struct rq *rq = cpu_rq(cpu);
5272 struct sched_domain *tmp;
5274 /* Remove the sched domains which do not contribute to scheduling. */
5275 for (tmp = sd; tmp; ) {
5276 struct sched_domain *parent = tmp->parent;
5280 if (sd_parent_degenerate(tmp, parent)) {
5281 tmp->parent = parent->parent;
5283 parent->parent->child = tmp;
5285 * Transfer SD_PREFER_SIBLING down in case of a
5286 * degenerate parent; the spans match for this
5287 * so the property transfers.
5289 if (parent->flags & SD_PREFER_SIBLING)
5290 tmp->flags |= SD_PREFER_SIBLING;
5291 destroy_sched_domain(parent, cpu);
5296 if (sd && sd_degenerate(sd)) {
5299 destroy_sched_domain(tmp, cpu);
5304 sched_domain_debug(sd, cpu);
5306 rq_attach_root(rq, rd);
5308 rcu_assign_pointer(rq->sd, sd);
5309 destroy_sched_domains(tmp, cpu);
5311 update_top_cache_domain(cpu);
5314 /* cpus with isolated domains */
5315 static cpumask_var_t cpu_isolated_map;
5317 /* Setup the mask of cpus configured for isolated domains */
5318 static int __init isolated_cpu_setup(char *str)
5320 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5321 cpulist_parse(str, cpu_isolated_map);
5325 __setup("isolcpus=", isolated_cpu_setup);
5327 static const struct cpumask *cpu_cpu_mask(int cpu)
5329 return cpumask_of_node(cpu_to_node(cpu));
5333 struct sched_domain **__percpu sd;
5334 struct sched_group **__percpu sg;
5335 struct sched_group_power **__percpu sgp;
5339 struct sched_domain ** __percpu sd;
5340 struct root_domain *rd;
5350 struct sched_domain_topology_level;
5352 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5353 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5355 #define SDTL_OVERLAP 0x01
5357 struct sched_domain_topology_level {
5358 sched_domain_init_f init;
5359 sched_domain_mask_f mask;
5362 struct sd_data data;
5366 * Build an iteration mask that can exclude certain CPUs from the upwards
5369 * Asymmetric node setups can result in situations where the domain tree is of
5370 * unequal depth, make sure to skip domains that already cover the entire
5373 * In that case build_sched_domains() will have terminated the iteration early
5374 * and our sibling sd spans will be empty. Domains should always include the
5375 * cpu they're built on, so check that.
5378 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5380 const struct cpumask *span = sched_domain_span(sd);
5381 struct sd_data *sdd = sd->private;
5382 struct sched_domain *sibling;
5385 for_each_cpu(i, span) {
5386 sibling = *per_cpu_ptr(sdd->sd, i);
5387 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5390 cpumask_set_cpu(i, sched_group_mask(sg));
5395 * Return the canonical balance cpu for this group, this is the first cpu
5396 * of this group that's also in the iteration mask.
5398 int group_balance_cpu(struct sched_group *sg)
5400 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5404 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5406 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5407 const struct cpumask *span = sched_domain_span(sd);
5408 struct cpumask *covered = sched_domains_tmpmask;
5409 struct sd_data *sdd = sd->private;
5410 struct sched_domain *child;
5413 cpumask_clear(covered);
5415 for_each_cpu(i, span) {
5416 struct cpumask *sg_span;
5418 if (cpumask_test_cpu(i, covered))
5421 child = *per_cpu_ptr(sdd->sd, i);
5423 /* See the comment near build_group_mask(). */
5424 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5427 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5428 GFP_KERNEL, cpu_to_node(cpu));
5433 sg_span = sched_group_cpus(sg);
5435 child = child->child;
5436 cpumask_copy(sg_span, sched_domain_span(child));
5438 cpumask_set_cpu(i, sg_span);
5440 cpumask_or(covered, covered, sg_span);
5442 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5443 if (atomic_inc_return(&sg->sgp->ref) == 1)
5444 build_group_mask(sd, sg);
5447 * Initialize sgp->power such that even if we mess up the
5448 * domains and no possible iteration will get us here, we won't
5451 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5454 * Make sure the first group of this domain contains the
5455 * canonical balance cpu. Otherwise the sched_domain iteration
5456 * breaks. See update_sg_lb_stats().
5458 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5459 group_balance_cpu(sg) == cpu)
5469 sd->groups = groups;
5474 free_sched_groups(first, 0);
5479 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5481 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5482 struct sched_domain *child = sd->child;
5485 cpu = cpumask_first(sched_domain_span(child));
5488 *sg = *per_cpu_ptr(sdd->sg, cpu);
5489 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5490 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5497 * build_sched_groups will build a circular linked list of the groups
5498 * covered by the given span, and will set each group's ->cpumask correctly,
5499 * and ->cpu_power to 0.
5501 * Assumes the sched_domain tree is fully constructed
5504 build_sched_groups(struct sched_domain *sd, int cpu)
5506 struct sched_group *first = NULL, *last = NULL;
5507 struct sd_data *sdd = sd->private;
5508 const struct cpumask *span = sched_domain_span(sd);
5509 struct cpumask *covered;
5512 get_group(cpu, sdd, &sd->groups);
5513 atomic_inc(&sd->groups->ref);
5515 if (cpu != cpumask_first(span))
5518 lockdep_assert_held(&sched_domains_mutex);
5519 covered = sched_domains_tmpmask;
5521 cpumask_clear(covered);
5523 for_each_cpu(i, span) {
5524 struct sched_group *sg;
5527 if (cpumask_test_cpu(i, covered))
5530 group = get_group(i, sdd, &sg);
5531 cpumask_clear(sched_group_cpus(sg));
5533 cpumask_setall(sched_group_mask(sg));
5535 for_each_cpu(j, span) {
5536 if (get_group(j, sdd, NULL) != group)
5539 cpumask_set_cpu(j, covered);
5540 cpumask_set_cpu(j, sched_group_cpus(sg));
5555 * Initialize sched groups cpu_power.
5557 * cpu_power indicates the capacity of sched group, which is used while
5558 * distributing the load between different sched groups in a sched domain.
5559 * Typically cpu_power for all the groups in a sched domain will be same unless
5560 * there are asymmetries in the topology. If there are asymmetries, group
5561 * having more cpu_power will pickup more load compared to the group having
5564 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5566 struct sched_group *sg = sd->groups;
5571 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5573 } while (sg != sd->groups);
5575 if (cpu != group_balance_cpu(sg))
5578 update_group_power(sd, cpu);
5579 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5582 int __weak arch_sd_sibling_asym_packing(void)
5584 return 0*SD_ASYM_PACKING;
5588 * Initializers for schedule domains
5589 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5592 #ifdef CONFIG_SCHED_DEBUG
5593 # define SD_INIT_NAME(sd, type) sd->name = #type
5595 # define SD_INIT_NAME(sd, type) do { } while (0)
5598 #define SD_INIT_FUNC(type) \
5599 static noinline struct sched_domain * \
5600 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5602 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5603 *sd = SD_##type##_INIT; \
5604 SD_INIT_NAME(sd, type); \
5605 sd->private = &tl->data; \
5610 #ifdef CONFIG_SCHED_SMT
5611 SD_INIT_FUNC(SIBLING)
5613 #ifdef CONFIG_SCHED_MC
5616 #ifdef CONFIG_SCHED_BOOK
5620 static int default_relax_domain_level = -1;
5621 int sched_domain_level_max;
5623 static int __init setup_relax_domain_level(char *str)
5625 if (kstrtoint(str, 0, &default_relax_domain_level))
5626 pr_warn("Unable to set relax_domain_level\n");
5630 __setup("relax_domain_level=", setup_relax_domain_level);
5632 static void set_domain_attribute(struct sched_domain *sd,
5633 struct sched_domain_attr *attr)
5637 if (!attr || attr->relax_domain_level < 0) {
5638 if (default_relax_domain_level < 0)
5641 request = default_relax_domain_level;
5643 request = attr->relax_domain_level;
5644 if (request < sd->level) {
5645 /* turn off idle balance on this domain */
5646 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5648 /* turn on idle balance on this domain */
5649 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5653 static void __sdt_free(const struct cpumask *cpu_map);
5654 static int __sdt_alloc(const struct cpumask *cpu_map);
5656 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5657 const struct cpumask *cpu_map)
5661 if (!atomic_read(&d->rd->refcount))
5662 free_rootdomain(&d->rd->rcu); /* fall through */
5664 free_percpu(d->sd); /* fall through */
5666 __sdt_free(cpu_map); /* fall through */
5672 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5673 const struct cpumask *cpu_map)
5675 memset(d, 0, sizeof(*d));
5677 if (__sdt_alloc(cpu_map))
5678 return sa_sd_storage;
5679 d->sd = alloc_percpu(struct sched_domain *);
5681 return sa_sd_storage;
5682 d->rd = alloc_rootdomain();
5685 return sa_rootdomain;
5689 * NULL the sd_data elements we've used to build the sched_domain and
5690 * sched_group structure so that the subsequent __free_domain_allocs()
5691 * will not free the data we're using.
5693 static void claim_allocations(int cpu, struct sched_domain *sd)
5695 struct sd_data *sdd = sd->private;
5697 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5698 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5700 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5701 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5703 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5704 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5707 #ifdef CONFIG_SCHED_SMT
5708 static const struct cpumask *cpu_smt_mask(int cpu)
5710 return topology_thread_cpumask(cpu);
5715 * Topology list, bottom-up.
5717 static struct sched_domain_topology_level default_topology[] = {
5718 #ifdef CONFIG_SCHED_SMT
5719 { sd_init_SIBLING, cpu_smt_mask, },
5721 #ifdef CONFIG_SCHED_MC
5722 { sd_init_MC, cpu_coregroup_mask, },
5724 #ifdef CONFIG_SCHED_BOOK
5725 { sd_init_BOOK, cpu_book_mask, },
5727 { sd_init_CPU, cpu_cpu_mask, },
5731 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5733 #define for_each_sd_topology(tl) \
5734 for (tl = sched_domain_topology; tl->init; tl++)
5738 static int sched_domains_numa_levels;
5739 static int *sched_domains_numa_distance;
5740 static struct cpumask ***sched_domains_numa_masks;
5741 static int sched_domains_curr_level;
5743 static inline int sd_local_flags(int level)
5745 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5748 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5751 static struct sched_domain *
5752 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5754 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5755 int level = tl->numa_level;
5756 int sd_weight = cpumask_weight(
5757 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5759 *sd = (struct sched_domain){
5760 .min_interval = sd_weight,
5761 .max_interval = 2*sd_weight,
5763 .imbalance_pct = 125,
5764 .cache_nice_tries = 2,
5771 .flags = 1*SD_LOAD_BALANCE
5772 | 1*SD_BALANCE_NEWIDLE
5777 | 0*SD_SHARE_CPUPOWER
5778 | 0*SD_SHARE_PKG_RESOURCES
5780 | 0*SD_PREFER_SIBLING
5782 | sd_local_flags(level)
5784 .last_balance = jiffies,
5785 .balance_interval = sd_weight,
5787 SD_INIT_NAME(sd, NUMA);
5788 sd->private = &tl->data;
5791 * Ugly hack to pass state to sd_numa_mask()...
5793 sched_domains_curr_level = tl->numa_level;
5798 static const struct cpumask *sd_numa_mask(int cpu)
5800 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5803 static void sched_numa_warn(const char *str)
5805 static int done = false;
5813 printk(KERN_WARNING "ERROR: %s\n\n", str);
5815 for (i = 0; i < nr_node_ids; i++) {
5816 printk(KERN_WARNING " ");
5817 for (j = 0; j < nr_node_ids; j++)
5818 printk(KERN_CONT "%02d ", node_distance(i,j));
5819 printk(KERN_CONT "\n");
5821 printk(KERN_WARNING "\n");
5824 static bool find_numa_distance(int distance)
5828 if (distance == node_distance(0, 0))
5831 for (i = 0; i < sched_domains_numa_levels; i++) {
5832 if (sched_domains_numa_distance[i] == distance)
5839 static void sched_init_numa(void)
5841 int next_distance, curr_distance = node_distance(0, 0);
5842 struct sched_domain_topology_level *tl;
5846 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5847 if (!sched_domains_numa_distance)
5851 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5852 * unique distances in the node_distance() table.
5854 * Assumes node_distance(0,j) includes all distances in
5855 * node_distance(i,j) in order to avoid cubic time.
5857 next_distance = curr_distance;
5858 for (i = 0; i < nr_node_ids; i++) {
5859 for (j = 0; j < nr_node_ids; j++) {
5860 for (k = 0; k < nr_node_ids; k++) {
5861 int distance = node_distance(i, k);
5863 if (distance > curr_distance &&
5864 (distance < next_distance ||
5865 next_distance == curr_distance))
5866 next_distance = distance;
5869 * While not a strong assumption it would be nice to know
5870 * about cases where if node A is connected to B, B is not
5871 * equally connected to A.
5873 if (sched_debug() && node_distance(k, i) != distance)
5874 sched_numa_warn("Node-distance not symmetric");
5876 if (sched_debug() && i && !find_numa_distance(distance))
5877 sched_numa_warn("Node-0 not representative");
5879 if (next_distance != curr_distance) {
5880 sched_domains_numa_distance[level++] = next_distance;
5881 sched_domains_numa_levels = level;
5882 curr_distance = next_distance;
5887 * In case of sched_debug() we verify the above assumption.
5893 * 'level' contains the number of unique distances, excluding the
5894 * identity distance node_distance(i,i).
5896 * The sched_domains_numa_distance[] array includes the actual distance
5901 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5902 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5903 * the array will contain less then 'level' members. This could be
5904 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5905 * in other functions.
5907 * We reset it to 'level' at the end of this function.
5909 sched_domains_numa_levels = 0;
5911 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5912 if (!sched_domains_numa_masks)
5916 * Now for each level, construct a mask per node which contains all
5917 * cpus of nodes that are that many hops away from us.
5919 for (i = 0; i < level; i++) {
5920 sched_domains_numa_masks[i] =
5921 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5922 if (!sched_domains_numa_masks[i])
5925 for (j = 0; j < nr_node_ids; j++) {
5926 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5930 sched_domains_numa_masks[i][j] = mask;
5932 for (k = 0; k < nr_node_ids; k++) {
5933 if (node_distance(j, k) > sched_domains_numa_distance[i])
5936 cpumask_or(mask, mask, cpumask_of_node(k));
5941 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5942 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5947 * Copy the default topology bits..
5949 for (i = 0; default_topology[i].init; i++)
5950 tl[i] = default_topology[i];
5953 * .. and append 'j' levels of NUMA goodness.
5955 for (j = 0; j < level; i++, j++) {
5956 tl[i] = (struct sched_domain_topology_level){
5957 .init = sd_numa_init,
5958 .mask = sd_numa_mask,
5959 .flags = SDTL_OVERLAP,
5964 sched_domain_topology = tl;
5966 sched_domains_numa_levels = level;
5969 static void sched_domains_numa_masks_set(int cpu)
5972 int node = cpu_to_node(cpu);
5974 for (i = 0; i < sched_domains_numa_levels; i++) {
5975 for (j = 0; j < nr_node_ids; j++) {
5976 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5977 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5982 static void sched_domains_numa_masks_clear(int cpu)
5985 for (i = 0; i < sched_domains_numa_levels; i++) {
5986 for (j = 0; j < nr_node_ids; j++)
5987 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5992 * Update sched_domains_numa_masks[level][node] array when new cpus
5995 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5996 unsigned long action,
5999 int cpu = (long)hcpu;
6001 switch (action & ~CPU_TASKS_FROZEN) {
6003 sched_domains_numa_masks_set(cpu);
6007 sched_domains_numa_masks_clear(cpu);
6017 static inline void sched_init_numa(void)
6021 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6022 unsigned long action,
6027 #endif /* CONFIG_NUMA */
6029 static int __sdt_alloc(const struct cpumask *cpu_map)
6031 struct sched_domain_topology_level *tl;
6034 for_each_sd_topology(tl) {
6035 struct sd_data *sdd = &tl->data;
6037 sdd->sd = alloc_percpu(struct sched_domain *);
6041 sdd->sg = alloc_percpu(struct sched_group *);
6045 sdd->sgp = alloc_percpu(struct sched_group_power *);
6049 for_each_cpu(j, cpu_map) {
6050 struct sched_domain *sd;
6051 struct sched_group *sg;
6052 struct sched_group_power *sgp;
6054 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6055 GFP_KERNEL, cpu_to_node(j));
6059 *per_cpu_ptr(sdd->sd, j) = sd;
6061 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6062 GFP_KERNEL, cpu_to_node(j));
6068 *per_cpu_ptr(sdd->sg, j) = sg;
6070 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6071 GFP_KERNEL, cpu_to_node(j));
6075 *per_cpu_ptr(sdd->sgp, j) = sgp;
6082 static void __sdt_free(const struct cpumask *cpu_map)
6084 struct sched_domain_topology_level *tl;
6087 for_each_sd_topology(tl) {
6088 struct sd_data *sdd = &tl->data;
6090 for_each_cpu(j, cpu_map) {
6091 struct sched_domain *sd;
6094 sd = *per_cpu_ptr(sdd->sd, j);
6095 if (sd && (sd->flags & SD_OVERLAP))
6096 free_sched_groups(sd->groups, 0);
6097 kfree(*per_cpu_ptr(sdd->sd, j));
6101 kfree(*per_cpu_ptr(sdd->sg, j));
6103 kfree(*per_cpu_ptr(sdd->sgp, j));
6105 free_percpu(sdd->sd);
6107 free_percpu(sdd->sg);
6109 free_percpu(sdd->sgp);
6114 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6115 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6116 struct sched_domain *child, int cpu)
6118 struct sched_domain *sd = tl->init(tl, cpu);
6122 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6124 sd->level = child->level + 1;
6125 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6129 set_domain_attribute(sd, attr);
6135 * Build sched domains for a given set of cpus and attach the sched domains
6136 * to the individual cpus
6138 static int build_sched_domains(const struct cpumask *cpu_map,
6139 struct sched_domain_attr *attr)
6141 enum s_alloc alloc_state;
6142 struct sched_domain *sd;
6144 int i, ret = -ENOMEM;
6146 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6147 if (alloc_state != sa_rootdomain)
6150 /* Set up domains for cpus specified by the cpu_map. */
6151 for_each_cpu(i, cpu_map) {
6152 struct sched_domain_topology_level *tl;
6155 for_each_sd_topology(tl) {
6156 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6157 if (tl == sched_domain_topology)
6158 *per_cpu_ptr(d.sd, i) = sd;
6159 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6160 sd->flags |= SD_OVERLAP;
6161 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6166 /* Build the groups for the domains */
6167 for_each_cpu(i, cpu_map) {
6168 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6169 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6170 if (sd->flags & SD_OVERLAP) {
6171 if (build_overlap_sched_groups(sd, i))
6174 if (build_sched_groups(sd, i))
6180 /* Calculate CPU power for physical packages and nodes */
6181 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6182 if (!cpumask_test_cpu(i, cpu_map))
6185 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6186 claim_allocations(i, sd);
6187 init_sched_groups_power(i, sd);
6191 /* Attach the domains */
6193 for_each_cpu(i, cpu_map) {
6194 sd = *per_cpu_ptr(d.sd, i);
6195 cpu_attach_domain(sd, d.rd, i);
6201 __free_domain_allocs(&d, alloc_state, cpu_map);
6205 static cpumask_var_t *doms_cur; /* current sched domains */
6206 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6207 static struct sched_domain_attr *dattr_cur;
6208 /* attribues of custom domains in 'doms_cur' */
6211 * Special case: If a kmalloc of a doms_cur partition (array of
6212 * cpumask) fails, then fallback to a single sched domain,
6213 * as determined by the single cpumask fallback_doms.
6215 static cpumask_var_t fallback_doms;
6218 * arch_update_cpu_topology lets virtualized architectures update the
6219 * cpu core maps. It is supposed to return 1 if the topology changed
6220 * or 0 if it stayed the same.
6222 int __attribute__((weak)) arch_update_cpu_topology(void)
6227 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6230 cpumask_var_t *doms;
6232 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6235 for (i = 0; i < ndoms; i++) {
6236 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6237 free_sched_domains(doms, i);
6244 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6247 for (i = 0; i < ndoms; i++)
6248 free_cpumask_var(doms[i]);
6253 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6254 * For now this just excludes isolated cpus, but could be used to
6255 * exclude other special cases in the future.
6257 static int init_sched_domains(const struct cpumask *cpu_map)
6261 arch_update_cpu_topology();
6263 doms_cur = alloc_sched_domains(ndoms_cur);
6265 doms_cur = &fallback_doms;
6266 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6267 err = build_sched_domains(doms_cur[0], NULL);
6268 register_sched_domain_sysctl();
6274 * Detach sched domains from a group of cpus specified in cpu_map
6275 * These cpus will now be attached to the NULL domain
6277 static void detach_destroy_domains(const struct cpumask *cpu_map)
6282 for_each_cpu(i, cpu_map)
6283 cpu_attach_domain(NULL, &def_root_domain, i);
6287 /* handle null as "default" */
6288 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6289 struct sched_domain_attr *new, int idx_new)
6291 struct sched_domain_attr tmp;
6298 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6299 new ? (new + idx_new) : &tmp,
6300 sizeof(struct sched_domain_attr));
6304 * Partition sched domains as specified by the 'ndoms_new'
6305 * cpumasks in the array doms_new[] of cpumasks. This compares
6306 * doms_new[] to the current sched domain partitioning, doms_cur[].
6307 * It destroys each deleted domain and builds each new domain.
6309 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6310 * The masks don't intersect (don't overlap.) We should setup one
6311 * sched domain for each mask. CPUs not in any of the cpumasks will
6312 * not be load balanced. If the same cpumask appears both in the
6313 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6316 * The passed in 'doms_new' should be allocated using
6317 * alloc_sched_domains. This routine takes ownership of it and will
6318 * free_sched_domains it when done with it. If the caller failed the
6319 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6320 * and partition_sched_domains() will fallback to the single partition
6321 * 'fallback_doms', it also forces the domains to be rebuilt.
6323 * If doms_new == NULL it will be replaced with cpu_online_mask.
6324 * ndoms_new == 0 is a special case for destroying existing domains,
6325 * and it will not create the default domain.
6327 * Call with hotplug lock held
6329 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6330 struct sched_domain_attr *dattr_new)
6335 mutex_lock(&sched_domains_mutex);
6337 /* always unregister in case we don't destroy any domains */
6338 unregister_sched_domain_sysctl();
6340 /* Let architecture update cpu core mappings. */
6341 new_topology = arch_update_cpu_topology();
6343 n = doms_new ? ndoms_new : 0;
6345 /* Destroy deleted domains */
6346 for (i = 0; i < ndoms_cur; i++) {
6347 for (j = 0; j < n && !new_topology; j++) {
6348 if (cpumask_equal(doms_cur[i], doms_new[j])
6349 && dattrs_equal(dattr_cur, i, dattr_new, j))
6352 /* no match - a current sched domain not in new doms_new[] */
6353 detach_destroy_domains(doms_cur[i]);
6359 if (doms_new == NULL) {
6361 doms_new = &fallback_doms;
6362 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6363 WARN_ON_ONCE(dattr_new);
6366 /* Build new domains */
6367 for (i = 0; i < ndoms_new; i++) {
6368 for (j = 0; j < n && !new_topology; j++) {
6369 if (cpumask_equal(doms_new[i], doms_cur[j])
6370 && dattrs_equal(dattr_new, i, dattr_cur, j))
6373 /* no match - add a new doms_new */
6374 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6379 /* Remember the new sched domains */
6380 if (doms_cur != &fallback_doms)
6381 free_sched_domains(doms_cur, ndoms_cur);
6382 kfree(dattr_cur); /* kfree(NULL) is safe */
6383 doms_cur = doms_new;
6384 dattr_cur = dattr_new;
6385 ndoms_cur = ndoms_new;
6387 register_sched_domain_sysctl();
6389 mutex_unlock(&sched_domains_mutex);
6392 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6395 * Update cpusets according to cpu_active mask. If cpusets are
6396 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6397 * around partition_sched_domains().
6399 * If we come here as part of a suspend/resume, don't touch cpusets because we
6400 * want to restore it back to its original state upon resume anyway.
6402 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6406 case CPU_ONLINE_FROZEN:
6407 case CPU_DOWN_FAILED_FROZEN:
6410 * num_cpus_frozen tracks how many CPUs are involved in suspend
6411 * resume sequence. As long as this is not the last online
6412 * operation in the resume sequence, just build a single sched
6413 * domain, ignoring cpusets.
6416 if (likely(num_cpus_frozen)) {
6417 partition_sched_domains(1, NULL, NULL);
6422 * This is the last CPU online operation. So fall through and
6423 * restore the original sched domains by considering the
6424 * cpuset configurations.
6428 case CPU_DOWN_FAILED:
6429 cpuset_update_active_cpus(true);
6437 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6441 case CPU_DOWN_PREPARE:
6442 cpuset_update_active_cpus(false);
6444 case CPU_DOWN_PREPARE_FROZEN:
6446 partition_sched_domains(1, NULL, NULL);
6454 void __init sched_init_smp(void)
6456 cpumask_var_t non_isolated_cpus;
6458 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6459 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6464 mutex_lock(&sched_domains_mutex);
6465 init_sched_domains(cpu_active_mask);
6466 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6467 if (cpumask_empty(non_isolated_cpus))
6468 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6469 mutex_unlock(&sched_domains_mutex);
6472 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6473 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6474 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6478 /* Move init over to a non-isolated CPU */
6479 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6481 sched_init_granularity();
6482 free_cpumask_var(non_isolated_cpus);
6484 init_sched_rt_class();
6487 void __init sched_init_smp(void)
6489 sched_init_granularity();
6491 #endif /* CONFIG_SMP */
6493 const_debug unsigned int sysctl_timer_migration = 1;
6495 int in_sched_functions(unsigned long addr)
6497 return in_lock_functions(addr) ||
6498 (addr >= (unsigned long)__sched_text_start
6499 && addr < (unsigned long)__sched_text_end);
6502 #ifdef CONFIG_CGROUP_SCHED
6504 * Default task group.
6505 * Every task in system belongs to this group at bootup.
6507 struct task_group root_task_group;
6508 LIST_HEAD(task_groups);
6511 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6513 void __init sched_init(void)
6516 unsigned long alloc_size = 0, ptr;
6518 #ifdef CONFIG_FAIR_GROUP_SCHED
6519 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6521 #ifdef CONFIG_RT_GROUP_SCHED
6522 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6524 #ifdef CONFIG_CPUMASK_OFFSTACK
6525 alloc_size += num_possible_cpus() * cpumask_size();
6528 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6530 #ifdef CONFIG_FAIR_GROUP_SCHED
6531 root_task_group.se = (struct sched_entity **)ptr;
6532 ptr += nr_cpu_ids * sizeof(void **);
6534 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6535 ptr += nr_cpu_ids * sizeof(void **);
6537 #endif /* CONFIG_FAIR_GROUP_SCHED */
6538 #ifdef CONFIG_RT_GROUP_SCHED
6539 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6540 ptr += nr_cpu_ids * sizeof(void **);
6542 root_task_group.rt_rq = (struct rt_rq **)ptr;
6543 ptr += nr_cpu_ids * sizeof(void **);
6545 #endif /* CONFIG_RT_GROUP_SCHED */
6546 #ifdef CONFIG_CPUMASK_OFFSTACK
6547 for_each_possible_cpu(i) {
6548 per_cpu(load_balance_mask, i) = (void *)ptr;
6549 ptr += cpumask_size();
6551 #endif /* CONFIG_CPUMASK_OFFSTACK */
6555 init_defrootdomain();
6558 init_rt_bandwidth(&def_rt_bandwidth,
6559 global_rt_period(), global_rt_runtime());
6561 #ifdef CONFIG_RT_GROUP_SCHED
6562 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6563 global_rt_period(), global_rt_runtime());
6564 #endif /* CONFIG_RT_GROUP_SCHED */
6566 #ifdef CONFIG_CGROUP_SCHED
6567 list_add(&root_task_group.list, &task_groups);
6568 INIT_LIST_HEAD(&root_task_group.children);
6569 INIT_LIST_HEAD(&root_task_group.siblings);
6570 autogroup_init(&init_task);
6572 #endif /* CONFIG_CGROUP_SCHED */
6574 for_each_possible_cpu(i) {
6578 raw_spin_lock_init(&rq->lock);
6580 rq->calc_load_active = 0;
6581 rq->calc_load_update = jiffies + LOAD_FREQ;
6582 init_cfs_rq(&rq->cfs);
6583 init_rt_rq(&rq->rt, rq);
6584 #ifdef CONFIG_FAIR_GROUP_SCHED
6585 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6586 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6588 * How much cpu bandwidth does root_task_group get?
6590 * In case of task-groups formed thr' the cgroup filesystem, it
6591 * gets 100% of the cpu resources in the system. This overall
6592 * system cpu resource is divided among the tasks of
6593 * root_task_group and its child task-groups in a fair manner,
6594 * based on each entity's (task or task-group's) weight
6595 * (se->load.weight).
6597 * In other words, if root_task_group has 10 tasks of weight
6598 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6599 * then A0's share of the cpu resource is:
6601 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6603 * We achieve this by letting root_task_group's tasks sit
6604 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6606 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6607 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6608 #endif /* CONFIG_FAIR_GROUP_SCHED */
6610 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6611 #ifdef CONFIG_RT_GROUP_SCHED
6612 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6613 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6616 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6617 rq->cpu_load[j] = 0;
6619 rq->last_load_update_tick = jiffies;
6624 rq->cpu_power = SCHED_POWER_SCALE;
6625 rq->post_schedule = 0;
6626 rq->active_balance = 0;
6627 rq->next_balance = jiffies;
6632 rq->avg_idle = 2*sysctl_sched_migration_cost;
6633 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6635 INIT_LIST_HEAD(&rq->cfs_tasks);
6637 rq_attach_root(rq, &def_root_domain);
6638 #ifdef CONFIG_NO_HZ_COMMON
6641 #ifdef CONFIG_NO_HZ_FULL
6642 rq->last_sched_tick = 0;
6646 atomic_set(&rq->nr_iowait, 0);
6649 set_load_weight(&init_task);
6651 #ifdef CONFIG_PREEMPT_NOTIFIERS
6652 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6655 #ifdef CONFIG_RT_MUTEXES
6656 plist_head_init(&init_task.pi_waiters);
6660 * The boot idle thread does lazy MMU switching as well:
6662 atomic_inc(&init_mm.mm_count);
6663 enter_lazy_tlb(&init_mm, current);
6666 * Make us the idle thread. Technically, schedule() should not be
6667 * called from this thread, however somewhere below it might be,
6668 * but because we are the idle thread, we just pick up running again
6669 * when this runqueue becomes "idle".
6671 init_idle(current, smp_processor_id());
6673 calc_load_update = jiffies + LOAD_FREQ;
6676 * During early bootup we pretend to be a normal task:
6678 current->sched_class = &fair_sched_class;
6681 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6682 /* May be allocated at isolcpus cmdline parse time */
6683 if (cpu_isolated_map == NULL)
6684 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6685 idle_thread_set_boot_cpu();
6687 init_sched_fair_class();
6689 scheduler_running = 1;
6692 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6693 static inline int preempt_count_equals(int preempt_offset)
6695 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6697 return (nested == preempt_offset);
6700 void __might_sleep(const char *file, int line, int preempt_offset)
6702 static unsigned long prev_jiffy; /* ratelimiting */
6704 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6705 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6706 system_state != SYSTEM_RUNNING || oops_in_progress)
6708 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6710 prev_jiffy = jiffies;
6713 "BUG: sleeping function called from invalid context at %s:%d\n",
6716 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6717 in_atomic(), irqs_disabled(),
6718 current->pid, current->comm);
6720 debug_show_held_locks(current);
6721 if (irqs_disabled())
6722 print_irqtrace_events(current);
6725 EXPORT_SYMBOL(__might_sleep);
6728 #ifdef CONFIG_MAGIC_SYSRQ
6729 static void normalize_task(struct rq *rq, struct task_struct *p)
6731 const struct sched_class *prev_class = p->sched_class;
6732 int old_prio = p->prio;
6737 dequeue_task(rq, p, 0);
6738 __setscheduler(rq, p, SCHED_NORMAL, 0);
6740 enqueue_task(rq, p, 0);
6741 resched_task(rq->curr);
6744 check_class_changed(rq, p, prev_class, old_prio);
6747 void normalize_rt_tasks(void)
6749 struct task_struct *g, *p;
6750 unsigned long flags;
6753 read_lock_irqsave(&tasklist_lock, flags);
6754 do_each_thread(g, p) {
6756 * Only normalize user tasks:
6761 p->se.exec_start = 0;
6762 #ifdef CONFIG_SCHEDSTATS
6763 p->se.statistics.wait_start = 0;
6764 p->se.statistics.sleep_start = 0;
6765 p->se.statistics.block_start = 0;
6770 * Renice negative nice level userspace
6773 if (TASK_NICE(p) < 0 && p->mm)
6774 set_user_nice(p, 0);
6778 raw_spin_lock(&p->pi_lock);
6779 rq = __task_rq_lock(p);
6781 normalize_task(rq, p);
6783 __task_rq_unlock(rq);
6784 raw_spin_unlock(&p->pi_lock);
6785 } while_each_thread(g, p);
6787 read_unlock_irqrestore(&tasklist_lock, flags);
6790 #endif /* CONFIG_MAGIC_SYSRQ */
6792 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6794 * These functions are only useful for the IA64 MCA handling, or kdb.
6796 * They can only be called when the whole system has been
6797 * stopped - every CPU needs to be quiescent, and no scheduling
6798 * activity can take place. Using them for anything else would
6799 * be a serious bug, and as a result, they aren't even visible
6800 * under any other configuration.
6804 * curr_task - return the current task for a given cpu.
6805 * @cpu: the processor in question.
6807 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6809 * Return: The current task for @cpu.
6811 struct task_struct *curr_task(int cpu)
6813 return cpu_curr(cpu);
6816 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6820 * set_curr_task - set the current task for a given cpu.
6821 * @cpu: the processor in question.
6822 * @p: the task pointer to set.
6824 * Description: This function must only be used when non-maskable interrupts
6825 * are serviced on a separate stack. It allows the architecture to switch the
6826 * notion of the current task on a cpu in a non-blocking manner. This function
6827 * must be called with all CPU's synchronized, and interrupts disabled, the
6828 * and caller must save the original value of the current task (see
6829 * curr_task() above) and restore that value before reenabling interrupts and
6830 * re-starting the system.
6832 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6834 void set_curr_task(int cpu, struct task_struct *p)
6841 #ifdef CONFIG_CGROUP_SCHED
6842 /* task_group_lock serializes the addition/removal of task groups */
6843 static DEFINE_SPINLOCK(task_group_lock);
6845 static void free_sched_group(struct task_group *tg)
6847 free_fair_sched_group(tg);
6848 free_rt_sched_group(tg);
6853 /* allocate runqueue etc for a new task group */
6854 struct task_group *sched_create_group(struct task_group *parent)
6856 struct task_group *tg;
6858 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6860 return ERR_PTR(-ENOMEM);
6862 if (!alloc_fair_sched_group(tg, parent))
6865 if (!alloc_rt_sched_group(tg, parent))
6871 free_sched_group(tg);
6872 return ERR_PTR(-ENOMEM);
6875 void sched_online_group(struct task_group *tg, struct task_group *parent)
6877 unsigned long flags;
6879 spin_lock_irqsave(&task_group_lock, flags);
6880 list_add_rcu(&tg->list, &task_groups);
6882 WARN_ON(!parent); /* root should already exist */
6884 tg->parent = parent;
6885 INIT_LIST_HEAD(&tg->children);
6886 list_add_rcu(&tg->siblings, &parent->children);
6887 spin_unlock_irqrestore(&task_group_lock, flags);
6890 /* rcu callback to free various structures associated with a task group */
6891 static void free_sched_group_rcu(struct rcu_head *rhp)
6893 /* now it should be safe to free those cfs_rqs */
6894 free_sched_group(container_of(rhp, struct task_group, rcu));
6897 /* Destroy runqueue etc associated with a task group */
6898 void sched_destroy_group(struct task_group *tg)
6900 /* wait for possible concurrent references to cfs_rqs complete */
6901 call_rcu(&tg->rcu, free_sched_group_rcu);
6904 void sched_offline_group(struct task_group *tg)
6906 unsigned long flags;
6909 /* end participation in shares distribution */
6910 for_each_possible_cpu(i)
6911 unregister_fair_sched_group(tg, i);
6913 spin_lock_irqsave(&task_group_lock, flags);
6914 list_del_rcu(&tg->list);
6915 list_del_rcu(&tg->siblings);
6916 spin_unlock_irqrestore(&task_group_lock, flags);
6919 /* change task's runqueue when it moves between groups.
6920 * The caller of this function should have put the task in its new group
6921 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6922 * reflect its new group.
6924 void sched_move_task(struct task_struct *tsk)
6926 struct task_group *tg;
6928 unsigned long flags;
6931 rq = task_rq_lock(tsk, &flags);
6933 running = task_current(rq, tsk);
6937 dequeue_task(rq, tsk, 0);
6938 if (unlikely(running))
6939 tsk->sched_class->put_prev_task(rq, tsk);
6941 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6942 lockdep_is_held(&tsk->sighand->siglock)),
6943 struct task_group, css);
6944 tg = autogroup_task_group(tsk, tg);
6945 tsk->sched_task_group = tg;
6947 #ifdef CONFIG_FAIR_GROUP_SCHED
6948 if (tsk->sched_class->task_move_group)
6949 tsk->sched_class->task_move_group(tsk, on_rq);
6952 set_task_rq(tsk, task_cpu(tsk));
6954 if (unlikely(running))
6955 tsk->sched_class->set_curr_task(rq);
6957 enqueue_task(rq, tsk, 0);
6959 task_rq_unlock(rq, tsk, &flags);
6961 #endif /* CONFIG_CGROUP_SCHED */
6963 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6964 static unsigned long to_ratio(u64 period, u64 runtime)
6966 if (runtime == RUNTIME_INF)
6969 return div64_u64(runtime << 20, period);
6973 #ifdef CONFIG_RT_GROUP_SCHED
6975 * Ensure that the real time constraints are schedulable.
6977 static DEFINE_MUTEX(rt_constraints_mutex);
6979 /* Must be called with tasklist_lock held */
6980 static inline int tg_has_rt_tasks(struct task_group *tg)
6982 struct task_struct *g, *p;
6984 do_each_thread(g, p) {
6985 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6987 } while_each_thread(g, p);
6992 struct rt_schedulable_data {
6993 struct task_group *tg;
6998 static int tg_rt_schedulable(struct task_group *tg, void *data)
7000 struct rt_schedulable_data *d = data;
7001 struct task_group *child;
7002 unsigned long total, sum = 0;
7003 u64 period, runtime;
7005 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7006 runtime = tg->rt_bandwidth.rt_runtime;
7009 period = d->rt_period;
7010 runtime = d->rt_runtime;
7014 * Cannot have more runtime than the period.
7016 if (runtime > period && runtime != RUNTIME_INF)
7020 * Ensure we don't starve existing RT tasks.
7022 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7025 total = to_ratio(period, runtime);
7028 * Nobody can have more than the global setting allows.
7030 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7034 * The sum of our children's runtime should not exceed our own.
7036 list_for_each_entry_rcu(child, &tg->children, siblings) {
7037 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7038 runtime = child->rt_bandwidth.rt_runtime;
7040 if (child == d->tg) {
7041 period = d->rt_period;
7042 runtime = d->rt_runtime;
7045 sum += to_ratio(period, runtime);
7054 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7058 struct rt_schedulable_data data = {
7060 .rt_period = period,
7061 .rt_runtime = runtime,
7065 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7071 static int tg_set_rt_bandwidth(struct task_group *tg,
7072 u64 rt_period, u64 rt_runtime)
7076 mutex_lock(&rt_constraints_mutex);
7077 read_lock(&tasklist_lock);
7078 err = __rt_schedulable(tg, rt_period, rt_runtime);
7082 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7083 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7084 tg->rt_bandwidth.rt_runtime = rt_runtime;
7086 for_each_possible_cpu(i) {
7087 struct rt_rq *rt_rq = tg->rt_rq[i];
7089 raw_spin_lock(&rt_rq->rt_runtime_lock);
7090 rt_rq->rt_runtime = rt_runtime;
7091 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7093 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7095 read_unlock(&tasklist_lock);
7096 mutex_unlock(&rt_constraints_mutex);
7101 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7103 u64 rt_runtime, rt_period;
7105 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7106 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7107 if (rt_runtime_us < 0)
7108 rt_runtime = RUNTIME_INF;
7110 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7113 static long sched_group_rt_runtime(struct task_group *tg)
7117 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7120 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7121 do_div(rt_runtime_us, NSEC_PER_USEC);
7122 return rt_runtime_us;
7125 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7127 u64 rt_runtime, rt_period;
7129 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7130 rt_runtime = tg->rt_bandwidth.rt_runtime;
7135 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7138 static long sched_group_rt_period(struct task_group *tg)
7142 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7143 do_div(rt_period_us, NSEC_PER_USEC);
7144 return rt_period_us;
7147 static int sched_rt_global_constraints(void)
7149 u64 runtime, period;
7152 if (sysctl_sched_rt_period <= 0)
7155 runtime = global_rt_runtime();
7156 period = global_rt_period();
7159 * Sanity check on the sysctl variables.
7161 if (runtime > period && runtime != RUNTIME_INF)
7164 mutex_lock(&rt_constraints_mutex);
7165 read_lock(&tasklist_lock);
7166 ret = __rt_schedulable(NULL, 0, 0);
7167 read_unlock(&tasklist_lock);
7168 mutex_unlock(&rt_constraints_mutex);
7173 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7175 /* Don't accept realtime tasks when there is no way for them to run */
7176 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7182 #else /* !CONFIG_RT_GROUP_SCHED */
7183 static int sched_rt_global_constraints(void)
7185 unsigned long flags;
7188 if (sysctl_sched_rt_period <= 0)
7192 * There's always some RT tasks in the root group
7193 * -- migration, kstopmachine etc..
7195 if (sysctl_sched_rt_runtime == 0)
7198 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7199 for_each_possible_cpu(i) {
7200 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7202 raw_spin_lock(&rt_rq->rt_runtime_lock);
7203 rt_rq->rt_runtime = global_rt_runtime();
7204 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7206 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7210 #endif /* CONFIG_RT_GROUP_SCHED */
7212 int sched_rr_handler(struct ctl_table *table, int write,
7213 void __user *buffer, size_t *lenp,
7217 static DEFINE_MUTEX(mutex);
7220 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7221 /* make sure that internally we keep jiffies */
7222 /* also, writing zero resets timeslice to default */
7223 if (!ret && write) {
7224 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7225 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7227 mutex_unlock(&mutex);
7231 int sched_rt_handler(struct ctl_table *table, int write,
7232 void __user *buffer, size_t *lenp,
7236 int old_period, old_runtime;
7237 static DEFINE_MUTEX(mutex);
7240 old_period = sysctl_sched_rt_period;
7241 old_runtime = sysctl_sched_rt_runtime;
7243 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7245 if (!ret && write) {
7246 ret = sched_rt_global_constraints();
7248 sysctl_sched_rt_period = old_period;
7249 sysctl_sched_rt_runtime = old_runtime;
7251 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7252 def_rt_bandwidth.rt_period =
7253 ns_to_ktime(global_rt_period());
7256 mutex_unlock(&mutex);
7261 #ifdef CONFIG_CGROUP_SCHED
7263 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7265 return css ? container_of(css, struct task_group, css) : NULL;
7268 static struct cgroup_subsys_state *
7269 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7271 struct task_group *parent = css_tg(parent_css);
7272 struct task_group *tg;
7275 /* This is early initialization for the top cgroup */
7276 return &root_task_group.css;
7279 tg = sched_create_group(parent);
7281 return ERR_PTR(-ENOMEM);
7286 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7288 struct task_group *tg = css_tg(css);
7289 struct task_group *parent = css_tg(css_parent(css));
7292 sched_online_group(tg, parent);
7296 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7298 struct task_group *tg = css_tg(css);
7300 sched_destroy_group(tg);
7303 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7305 struct task_group *tg = css_tg(css);
7307 sched_offline_group(tg);
7310 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7311 struct cgroup_taskset *tset)
7313 struct task_struct *task;
7315 cgroup_taskset_for_each(task, css, tset) {
7316 #ifdef CONFIG_RT_GROUP_SCHED
7317 if (!sched_rt_can_attach(css_tg(css), task))
7320 /* We don't support RT-tasks being in separate groups */
7321 if (task->sched_class != &fair_sched_class)
7328 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7329 struct cgroup_taskset *tset)
7331 struct task_struct *task;
7333 cgroup_taskset_for_each(task, css, tset)
7334 sched_move_task(task);
7337 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7338 struct cgroup_subsys_state *old_css,
7339 struct task_struct *task)
7342 * cgroup_exit() is called in the copy_process() failure path.
7343 * Ignore this case since the task hasn't ran yet, this avoids
7344 * trying to poke a half freed task state from generic code.
7346 if (!(task->flags & PF_EXITING))
7349 sched_move_task(task);
7352 #ifdef CONFIG_FAIR_GROUP_SCHED
7353 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7354 struct cftype *cftype, u64 shareval)
7356 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7359 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7362 struct task_group *tg = css_tg(css);
7364 return (u64) scale_load_down(tg->shares);
7367 #ifdef CONFIG_CFS_BANDWIDTH
7368 static DEFINE_MUTEX(cfs_constraints_mutex);
7370 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7371 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7373 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7375 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7377 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7378 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7380 if (tg == &root_task_group)
7384 * Ensure we have at some amount of bandwidth every period. This is
7385 * to prevent reaching a state of large arrears when throttled via
7386 * entity_tick() resulting in prolonged exit starvation.
7388 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7392 * Likewise, bound things on the otherside by preventing insane quota
7393 * periods. This also allows us to normalize in computing quota
7396 if (period > max_cfs_quota_period)
7399 mutex_lock(&cfs_constraints_mutex);
7400 ret = __cfs_schedulable(tg, period, quota);
7404 runtime_enabled = quota != RUNTIME_INF;
7405 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7406 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7407 raw_spin_lock_irq(&cfs_b->lock);
7408 cfs_b->period = ns_to_ktime(period);
7409 cfs_b->quota = quota;
7411 __refill_cfs_bandwidth_runtime(cfs_b);
7412 /* restart the period timer (if active) to handle new period expiry */
7413 if (runtime_enabled && cfs_b->timer_active) {
7414 /* force a reprogram */
7415 cfs_b->timer_active = 0;
7416 __start_cfs_bandwidth(cfs_b);
7418 raw_spin_unlock_irq(&cfs_b->lock);
7420 for_each_possible_cpu(i) {
7421 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7422 struct rq *rq = cfs_rq->rq;
7424 raw_spin_lock_irq(&rq->lock);
7425 cfs_rq->runtime_enabled = runtime_enabled;
7426 cfs_rq->runtime_remaining = 0;
7428 if (cfs_rq->throttled)
7429 unthrottle_cfs_rq(cfs_rq);
7430 raw_spin_unlock_irq(&rq->lock);
7433 mutex_unlock(&cfs_constraints_mutex);
7438 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7442 period = ktime_to_ns(tg->cfs_bandwidth.period);
7443 if (cfs_quota_us < 0)
7444 quota = RUNTIME_INF;
7446 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7448 return tg_set_cfs_bandwidth(tg, period, quota);
7451 long tg_get_cfs_quota(struct task_group *tg)
7455 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7458 quota_us = tg->cfs_bandwidth.quota;
7459 do_div(quota_us, NSEC_PER_USEC);
7464 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7468 period = (u64)cfs_period_us * NSEC_PER_USEC;
7469 quota = tg->cfs_bandwidth.quota;
7471 return tg_set_cfs_bandwidth(tg, period, quota);
7474 long tg_get_cfs_period(struct task_group *tg)
7478 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7479 do_div(cfs_period_us, NSEC_PER_USEC);
7481 return cfs_period_us;
7484 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7487 return tg_get_cfs_quota(css_tg(css));
7490 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7491 struct cftype *cftype, s64 cfs_quota_us)
7493 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7496 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7499 return tg_get_cfs_period(css_tg(css));
7502 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7503 struct cftype *cftype, u64 cfs_period_us)
7505 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7508 struct cfs_schedulable_data {
7509 struct task_group *tg;
7514 * normalize group quota/period to be quota/max_period
7515 * note: units are usecs
7517 static u64 normalize_cfs_quota(struct task_group *tg,
7518 struct cfs_schedulable_data *d)
7526 period = tg_get_cfs_period(tg);
7527 quota = tg_get_cfs_quota(tg);
7530 /* note: these should typically be equivalent */
7531 if (quota == RUNTIME_INF || quota == -1)
7534 return to_ratio(period, quota);
7537 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7539 struct cfs_schedulable_data *d = data;
7540 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7541 s64 quota = 0, parent_quota = -1;
7544 quota = RUNTIME_INF;
7546 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7548 quota = normalize_cfs_quota(tg, d);
7549 parent_quota = parent_b->hierarchal_quota;
7552 * ensure max(child_quota) <= parent_quota, inherit when no
7555 if (quota == RUNTIME_INF)
7556 quota = parent_quota;
7557 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7560 cfs_b->hierarchal_quota = quota;
7565 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7568 struct cfs_schedulable_data data = {
7574 if (quota != RUNTIME_INF) {
7575 do_div(data.period, NSEC_PER_USEC);
7576 do_div(data.quota, NSEC_PER_USEC);
7580 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7586 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7587 struct cgroup_map_cb *cb)
7589 struct task_group *tg = css_tg(css);
7590 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7592 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7593 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7594 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7598 #endif /* CONFIG_CFS_BANDWIDTH */
7599 #endif /* CONFIG_FAIR_GROUP_SCHED */
7601 #ifdef CONFIG_RT_GROUP_SCHED
7602 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7603 struct cftype *cft, s64 val)
7605 return sched_group_set_rt_runtime(css_tg(css), val);
7608 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7611 return sched_group_rt_runtime(css_tg(css));
7614 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7615 struct cftype *cftype, u64 rt_period_us)
7617 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7620 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7623 return sched_group_rt_period(css_tg(css));
7625 #endif /* CONFIG_RT_GROUP_SCHED */
7627 static struct cftype cpu_files[] = {
7628 #ifdef CONFIG_FAIR_GROUP_SCHED
7631 .read_u64 = cpu_shares_read_u64,
7632 .write_u64 = cpu_shares_write_u64,
7635 #ifdef CONFIG_CFS_BANDWIDTH
7637 .name = "cfs_quota_us",
7638 .read_s64 = cpu_cfs_quota_read_s64,
7639 .write_s64 = cpu_cfs_quota_write_s64,
7642 .name = "cfs_period_us",
7643 .read_u64 = cpu_cfs_period_read_u64,
7644 .write_u64 = cpu_cfs_period_write_u64,
7648 .read_map = cpu_stats_show,
7651 #ifdef CONFIG_RT_GROUP_SCHED
7653 .name = "rt_runtime_us",
7654 .read_s64 = cpu_rt_runtime_read,
7655 .write_s64 = cpu_rt_runtime_write,
7658 .name = "rt_period_us",
7659 .read_u64 = cpu_rt_period_read_uint,
7660 .write_u64 = cpu_rt_period_write_uint,
7666 struct cgroup_subsys cpu_cgroup_subsys = {
7668 .css_alloc = cpu_cgroup_css_alloc,
7669 .css_free = cpu_cgroup_css_free,
7670 .css_online = cpu_cgroup_css_online,
7671 .css_offline = cpu_cgroup_css_offline,
7672 .can_attach = cpu_cgroup_can_attach,
7673 .attach = cpu_cgroup_attach,
7674 .exit = cpu_cgroup_exit,
7675 .subsys_id = cpu_cgroup_subsys_id,
7676 .base_cftypes = cpu_files,
7680 #endif /* CONFIG_CGROUP_SCHED */
7682 void dump_cpu_task(int cpu)
7684 pr_info("Task dump for CPU %d:\n", cpu);
7685 sched_show_task(cpu_curr(cpu));