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;
1736 #endif /* CONFIG_NUMA_BALANCING */
1739 #ifdef CONFIG_NUMA_BALANCING
1740 #ifdef CONFIG_SCHED_DEBUG
1741 void set_numabalancing_state(bool enabled)
1744 sched_feat_set("NUMA");
1746 sched_feat_set("NO_NUMA");
1749 __read_mostly bool numabalancing_enabled;
1751 void set_numabalancing_state(bool enabled)
1753 numabalancing_enabled = enabled;
1755 #endif /* CONFIG_SCHED_DEBUG */
1756 #endif /* CONFIG_NUMA_BALANCING */
1759 * fork()/clone()-time setup:
1761 void sched_fork(struct task_struct *p)
1763 unsigned long flags;
1764 int cpu = get_cpu();
1768 * We mark the process as running here. This guarantees that
1769 * nobody will actually run it, and a signal or other external
1770 * event cannot wake it up and insert it on the runqueue either.
1772 p->state = TASK_RUNNING;
1775 * Make sure we do not leak PI boosting priority to the child.
1777 p->prio = current->normal_prio;
1780 * Revert to default priority/policy on fork if requested.
1782 if (unlikely(p->sched_reset_on_fork)) {
1783 if (task_has_rt_policy(p)) {
1784 p->policy = SCHED_NORMAL;
1785 p->static_prio = NICE_TO_PRIO(0);
1787 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1788 p->static_prio = NICE_TO_PRIO(0);
1790 p->prio = p->normal_prio = __normal_prio(p);
1794 * We don't need the reset flag anymore after the fork. It has
1795 * fulfilled its duty:
1797 p->sched_reset_on_fork = 0;
1800 if (!rt_prio(p->prio))
1801 p->sched_class = &fair_sched_class;
1803 if (p->sched_class->task_fork)
1804 p->sched_class->task_fork(p);
1807 * The child is not yet in the pid-hash so no cgroup attach races,
1808 * and the cgroup is pinned to this child due to cgroup_fork()
1809 * is ran before sched_fork().
1811 * Silence PROVE_RCU.
1813 raw_spin_lock_irqsave(&p->pi_lock, flags);
1814 set_task_cpu(p, cpu);
1815 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1817 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1818 if (likely(sched_info_on()))
1819 memset(&p->sched_info, 0, sizeof(p->sched_info));
1821 #if defined(CONFIG_SMP)
1824 init_task_preempt_count(p);
1826 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1833 * wake_up_new_task - wake up a newly created task for the first time.
1835 * This function will do some initial scheduler statistics housekeeping
1836 * that must be done for every newly created context, then puts the task
1837 * on the runqueue and wakes it.
1839 void wake_up_new_task(struct task_struct *p)
1841 unsigned long flags;
1844 raw_spin_lock_irqsave(&p->pi_lock, flags);
1847 * Fork balancing, do it here and not earlier because:
1848 * - cpus_allowed can change in the fork path
1849 * - any previously selected cpu might disappear through hotplug
1851 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1854 /* Initialize new task's runnable average */
1855 init_task_runnable_average(p);
1856 rq = __task_rq_lock(p);
1857 activate_task(rq, p, 0);
1859 trace_sched_wakeup_new(p, true);
1860 check_preempt_curr(rq, p, WF_FORK);
1862 if (p->sched_class->task_woken)
1863 p->sched_class->task_woken(rq, p);
1865 task_rq_unlock(rq, p, &flags);
1868 #ifdef CONFIG_PREEMPT_NOTIFIERS
1871 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1872 * @notifier: notifier struct to register
1874 void preempt_notifier_register(struct preempt_notifier *notifier)
1876 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1878 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1881 * preempt_notifier_unregister - no longer interested in preemption notifications
1882 * @notifier: notifier struct to unregister
1884 * This is safe to call from within a preemption notifier.
1886 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1888 hlist_del(¬ifier->link);
1890 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1892 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1894 struct preempt_notifier *notifier;
1896 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1897 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1901 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1902 struct task_struct *next)
1904 struct preempt_notifier *notifier;
1906 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1907 notifier->ops->sched_out(notifier, next);
1910 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1912 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1917 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1918 struct task_struct *next)
1922 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1925 * prepare_task_switch - prepare to switch tasks
1926 * @rq: the runqueue preparing to switch
1927 * @prev: the current task that is being switched out
1928 * @next: the task we are going to switch to.
1930 * This is called with the rq lock held and interrupts off. It must
1931 * be paired with a subsequent finish_task_switch after the context
1934 * prepare_task_switch sets up locking and calls architecture specific
1938 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1939 struct task_struct *next)
1941 trace_sched_switch(prev, next);
1942 sched_info_switch(rq, prev, next);
1943 perf_event_task_sched_out(prev, next);
1944 fire_sched_out_preempt_notifiers(prev, next);
1945 prepare_lock_switch(rq, next);
1946 prepare_arch_switch(next);
1950 * finish_task_switch - clean up after a task-switch
1951 * @rq: runqueue associated with task-switch
1952 * @prev: the thread we just switched away from.
1954 * finish_task_switch must be called after the context switch, paired
1955 * with a prepare_task_switch call before the context switch.
1956 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1957 * and do any other architecture-specific cleanup actions.
1959 * Note that we may have delayed dropping an mm in context_switch(). If
1960 * so, we finish that here outside of the runqueue lock. (Doing it
1961 * with the lock held can cause deadlocks; see schedule() for
1964 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1965 __releases(rq->lock)
1967 struct mm_struct *mm = rq->prev_mm;
1973 * A task struct has one reference for the use as "current".
1974 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1975 * schedule one last time. The schedule call will never return, and
1976 * the scheduled task must drop that reference.
1977 * The test for TASK_DEAD must occur while the runqueue locks are
1978 * still held, otherwise prev could be scheduled on another cpu, die
1979 * there before we look at prev->state, and then the reference would
1981 * Manfred Spraul <manfred@colorfullife.com>
1983 prev_state = prev->state;
1984 vtime_task_switch(prev);
1985 finish_arch_switch(prev);
1986 perf_event_task_sched_in(prev, current);
1987 finish_lock_switch(rq, prev);
1988 finish_arch_post_lock_switch();
1990 fire_sched_in_preempt_notifiers(current);
1993 if (unlikely(prev_state == TASK_DEAD)) {
1994 task_numa_free(prev);
1997 * Remove function-return probe instances associated with this
1998 * task and put them back on the free list.
2000 kprobe_flush_task(prev);
2001 put_task_struct(prev);
2004 tick_nohz_task_switch(current);
2009 /* assumes rq->lock is held */
2010 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2012 if (prev->sched_class->pre_schedule)
2013 prev->sched_class->pre_schedule(rq, prev);
2016 /* rq->lock is NOT held, but preemption is disabled */
2017 static inline void post_schedule(struct rq *rq)
2019 if (rq->post_schedule) {
2020 unsigned long flags;
2022 raw_spin_lock_irqsave(&rq->lock, flags);
2023 if (rq->curr->sched_class->post_schedule)
2024 rq->curr->sched_class->post_schedule(rq);
2025 raw_spin_unlock_irqrestore(&rq->lock, flags);
2027 rq->post_schedule = 0;
2033 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2037 static inline void post_schedule(struct rq *rq)
2044 * schedule_tail - first thing a freshly forked thread must call.
2045 * @prev: the thread we just switched away from.
2047 asmlinkage void schedule_tail(struct task_struct *prev)
2048 __releases(rq->lock)
2050 struct rq *rq = this_rq();
2052 finish_task_switch(rq, prev);
2055 * FIXME: do we need to worry about rq being invalidated by the
2060 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2061 /* In this case, finish_task_switch does not reenable preemption */
2064 if (current->set_child_tid)
2065 put_user(task_pid_vnr(current), current->set_child_tid);
2069 * context_switch - switch to the new MM and the new
2070 * thread's register state.
2073 context_switch(struct rq *rq, struct task_struct *prev,
2074 struct task_struct *next)
2076 struct mm_struct *mm, *oldmm;
2078 prepare_task_switch(rq, prev, next);
2081 oldmm = prev->active_mm;
2083 * For paravirt, this is coupled with an exit in switch_to to
2084 * combine the page table reload and the switch backend into
2087 arch_start_context_switch(prev);
2090 next->active_mm = oldmm;
2091 atomic_inc(&oldmm->mm_count);
2092 enter_lazy_tlb(oldmm, next);
2094 switch_mm(oldmm, mm, next);
2097 prev->active_mm = NULL;
2098 rq->prev_mm = oldmm;
2101 * Since the runqueue lock will be released by the next
2102 * task (which is an invalid locking op but in the case
2103 * of the scheduler it's an obvious special-case), so we
2104 * do an early lockdep release here:
2106 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2107 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2110 context_tracking_task_switch(prev, next);
2111 /* Here we just switch the register state and the stack. */
2112 switch_to(prev, next, prev);
2116 * this_rq must be evaluated again because prev may have moved
2117 * CPUs since it called schedule(), thus the 'rq' on its stack
2118 * frame will be invalid.
2120 finish_task_switch(this_rq(), prev);
2124 * nr_running and nr_context_switches:
2126 * externally visible scheduler statistics: current number of runnable
2127 * threads, total number of context switches performed since bootup.
2129 unsigned long nr_running(void)
2131 unsigned long i, sum = 0;
2133 for_each_online_cpu(i)
2134 sum += cpu_rq(i)->nr_running;
2139 unsigned long long nr_context_switches(void)
2142 unsigned long long sum = 0;
2144 for_each_possible_cpu(i)
2145 sum += cpu_rq(i)->nr_switches;
2150 unsigned long nr_iowait(void)
2152 unsigned long i, sum = 0;
2154 for_each_possible_cpu(i)
2155 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2160 unsigned long nr_iowait_cpu(int cpu)
2162 struct rq *this = cpu_rq(cpu);
2163 return atomic_read(&this->nr_iowait);
2169 * sched_exec - execve() is a valuable balancing opportunity, because at
2170 * this point the task has the smallest effective memory and cache footprint.
2172 void sched_exec(void)
2174 struct task_struct *p = current;
2175 unsigned long flags;
2178 raw_spin_lock_irqsave(&p->pi_lock, flags);
2179 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2180 if (dest_cpu == smp_processor_id())
2183 if (likely(cpu_active(dest_cpu))) {
2184 struct migration_arg arg = { p, dest_cpu };
2186 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2187 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2191 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2196 DEFINE_PER_CPU(struct kernel_stat, kstat);
2197 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2199 EXPORT_PER_CPU_SYMBOL(kstat);
2200 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2203 * Return any ns on the sched_clock that have not yet been accounted in
2204 * @p in case that task is currently running.
2206 * Called with task_rq_lock() held on @rq.
2208 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2212 if (task_current(rq, p)) {
2213 update_rq_clock(rq);
2214 ns = rq_clock_task(rq) - p->se.exec_start;
2222 unsigned long long task_delta_exec(struct task_struct *p)
2224 unsigned long flags;
2228 rq = task_rq_lock(p, &flags);
2229 ns = do_task_delta_exec(p, rq);
2230 task_rq_unlock(rq, p, &flags);
2236 * Return accounted runtime for the task.
2237 * In case the task is currently running, return the runtime plus current's
2238 * pending runtime that have not been accounted yet.
2240 unsigned long long task_sched_runtime(struct task_struct *p)
2242 unsigned long flags;
2246 rq = task_rq_lock(p, &flags);
2247 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2248 task_rq_unlock(rq, p, &flags);
2254 * This function gets called by the timer code, with HZ frequency.
2255 * We call it with interrupts disabled.
2257 void scheduler_tick(void)
2259 int cpu = smp_processor_id();
2260 struct rq *rq = cpu_rq(cpu);
2261 struct task_struct *curr = rq->curr;
2265 raw_spin_lock(&rq->lock);
2266 update_rq_clock(rq);
2267 curr->sched_class->task_tick(rq, curr, 0);
2268 update_cpu_load_active(rq);
2269 raw_spin_unlock(&rq->lock);
2271 perf_event_task_tick();
2274 rq->idle_balance = idle_cpu(cpu);
2275 trigger_load_balance(rq, cpu);
2277 rq_last_tick_reset(rq);
2280 #ifdef CONFIG_NO_HZ_FULL
2282 * scheduler_tick_max_deferment
2284 * Keep at least one tick per second when a single
2285 * active task is running because the scheduler doesn't
2286 * yet completely support full dynticks environment.
2288 * This makes sure that uptime, CFS vruntime, load
2289 * balancing, etc... continue to move forward, even
2290 * with a very low granularity.
2292 * Return: Maximum deferment in nanoseconds.
2294 u64 scheduler_tick_max_deferment(void)
2296 struct rq *rq = this_rq();
2297 unsigned long next, now = ACCESS_ONCE(jiffies);
2299 next = rq->last_sched_tick + HZ;
2301 if (time_before_eq(next, now))
2304 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2308 notrace unsigned long get_parent_ip(unsigned long addr)
2310 if (in_lock_functions(addr)) {
2311 addr = CALLER_ADDR2;
2312 if (in_lock_functions(addr))
2313 addr = CALLER_ADDR3;
2318 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2319 defined(CONFIG_PREEMPT_TRACER))
2321 void __kprobes preempt_count_add(int val)
2323 #ifdef CONFIG_DEBUG_PREEMPT
2327 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2330 __preempt_count_add(val);
2331 #ifdef CONFIG_DEBUG_PREEMPT
2333 * Spinlock count overflowing soon?
2335 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2338 if (preempt_count() == val)
2339 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2341 EXPORT_SYMBOL(preempt_count_add);
2343 void __kprobes preempt_count_sub(int val)
2345 #ifdef CONFIG_DEBUG_PREEMPT
2349 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2352 * Is the spinlock portion underflowing?
2354 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2355 !(preempt_count() & PREEMPT_MASK)))
2359 if (preempt_count() == val)
2360 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2361 __preempt_count_sub(val);
2363 EXPORT_SYMBOL(preempt_count_sub);
2368 * Print scheduling while atomic bug:
2370 static noinline void __schedule_bug(struct task_struct *prev)
2372 if (oops_in_progress)
2375 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2376 prev->comm, prev->pid, preempt_count());
2378 debug_show_held_locks(prev);
2380 if (irqs_disabled())
2381 print_irqtrace_events(prev);
2383 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2387 * Various schedule()-time debugging checks and statistics:
2389 static inline void schedule_debug(struct task_struct *prev)
2392 * Test if we are atomic. Since do_exit() needs to call into
2393 * schedule() atomically, we ignore that path for now.
2394 * Otherwise, whine if we are scheduling when we should not be.
2396 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2397 __schedule_bug(prev);
2400 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2402 schedstat_inc(this_rq(), sched_count);
2405 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2407 if (prev->on_rq || rq->skip_clock_update < 0)
2408 update_rq_clock(rq);
2409 prev->sched_class->put_prev_task(rq, prev);
2413 * Pick up the highest-prio task:
2415 static inline struct task_struct *
2416 pick_next_task(struct rq *rq)
2418 const struct sched_class *class;
2419 struct task_struct *p;
2422 * Optimization: we know that if all tasks are in
2423 * the fair class we can call that function directly:
2425 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2426 p = fair_sched_class.pick_next_task(rq);
2431 for_each_class(class) {
2432 p = class->pick_next_task(rq);
2437 BUG(); /* the idle class will always have a runnable task */
2441 * __schedule() is the main scheduler function.
2443 * The main means of driving the scheduler and thus entering this function are:
2445 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2447 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2448 * paths. For example, see arch/x86/entry_64.S.
2450 * To drive preemption between tasks, the scheduler sets the flag in timer
2451 * interrupt handler scheduler_tick().
2453 * 3. Wakeups don't really cause entry into schedule(). They add a
2454 * task to the run-queue and that's it.
2456 * Now, if the new task added to the run-queue preempts the current
2457 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2458 * called on the nearest possible occasion:
2460 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2462 * - in syscall or exception context, at the next outmost
2463 * preempt_enable(). (this might be as soon as the wake_up()'s
2466 * - in IRQ context, return from interrupt-handler to
2467 * preemptible context
2469 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2472 * - cond_resched() call
2473 * - explicit schedule() call
2474 * - return from syscall or exception to user-space
2475 * - return from interrupt-handler to user-space
2477 static void __sched __schedule(void)
2479 struct task_struct *prev, *next;
2480 unsigned long *switch_count;
2486 cpu = smp_processor_id();
2488 rcu_note_context_switch(cpu);
2491 schedule_debug(prev);
2493 if (sched_feat(HRTICK))
2497 * Make sure that signal_pending_state()->signal_pending() below
2498 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2499 * done by the caller to avoid the race with signal_wake_up().
2501 smp_mb__before_spinlock();
2502 raw_spin_lock_irq(&rq->lock);
2504 switch_count = &prev->nivcsw;
2505 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2506 if (unlikely(signal_pending_state(prev->state, prev))) {
2507 prev->state = TASK_RUNNING;
2509 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2513 * If a worker went to sleep, notify and ask workqueue
2514 * whether it wants to wake up a task to maintain
2517 if (prev->flags & PF_WQ_WORKER) {
2518 struct task_struct *to_wakeup;
2520 to_wakeup = wq_worker_sleeping(prev, cpu);
2522 try_to_wake_up_local(to_wakeup);
2525 switch_count = &prev->nvcsw;
2528 pre_schedule(rq, prev);
2530 if (unlikely(!rq->nr_running))
2531 idle_balance(cpu, rq);
2533 put_prev_task(rq, prev);
2534 next = pick_next_task(rq);
2535 clear_tsk_need_resched(prev);
2536 clear_preempt_need_resched();
2537 rq->skip_clock_update = 0;
2539 if (likely(prev != next)) {
2544 context_switch(rq, prev, next); /* unlocks the rq */
2546 * The context switch have flipped the stack from under us
2547 * and restored the local variables which were saved when
2548 * this task called schedule() in the past. prev == current
2549 * is still correct, but it can be moved to another cpu/rq.
2551 cpu = smp_processor_id();
2554 raw_spin_unlock_irq(&rq->lock);
2558 sched_preempt_enable_no_resched();
2563 static inline void sched_submit_work(struct task_struct *tsk)
2565 if (!tsk->state || tsk_is_pi_blocked(tsk))
2568 * If we are going to sleep and we have plugged IO queued,
2569 * make sure to submit it to avoid deadlocks.
2571 if (blk_needs_flush_plug(tsk))
2572 blk_schedule_flush_plug(tsk);
2575 asmlinkage void __sched schedule(void)
2577 struct task_struct *tsk = current;
2579 sched_submit_work(tsk);
2582 EXPORT_SYMBOL(schedule);
2584 #ifdef CONFIG_CONTEXT_TRACKING
2585 asmlinkage void __sched schedule_user(void)
2588 * If we come here after a random call to set_need_resched(),
2589 * or we have been woken up remotely but the IPI has not yet arrived,
2590 * we haven't yet exited the RCU idle mode. Do it here manually until
2591 * we find a better solution.
2600 * schedule_preempt_disabled - called with preemption disabled
2602 * Returns with preemption disabled. Note: preempt_count must be 1
2604 void __sched schedule_preempt_disabled(void)
2606 sched_preempt_enable_no_resched();
2611 #ifdef CONFIG_PREEMPT
2613 * this is the entry point to schedule() from in-kernel preemption
2614 * off of preempt_enable. Kernel preemptions off return from interrupt
2615 * occur there and call schedule directly.
2617 asmlinkage void __sched notrace preempt_schedule(void)
2620 * If there is a non-zero preempt_count or interrupts are disabled,
2621 * we do not want to preempt the current task. Just return..
2623 if (likely(!preemptible()))
2627 __preempt_count_add(PREEMPT_ACTIVE);
2629 __preempt_count_sub(PREEMPT_ACTIVE);
2632 * Check again in case we missed a preemption opportunity
2633 * between schedule and now.
2636 } while (need_resched());
2638 EXPORT_SYMBOL(preempt_schedule);
2641 * this is the entry point to schedule() from kernel preemption
2642 * off of irq context.
2643 * Note, that this is called and return with irqs disabled. This will
2644 * protect us against recursive calling from irq.
2646 asmlinkage void __sched preempt_schedule_irq(void)
2648 enum ctx_state prev_state;
2650 /* Catch callers which need to be fixed */
2651 BUG_ON(preempt_count() || !irqs_disabled());
2653 prev_state = exception_enter();
2656 __preempt_count_add(PREEMPT_ACTIVE);
2659 local_irq_disable();
2660 __preempt_count_sub(PREEMPT_ACTIVE);
2663 * Check again in case we missed a preemption opportunity
2664 * between schedule and now.
2667 } while (need_resched());
2669 exception_exit(prev_state);
2672 #endif /* CONFIG_PREEMPT */
2674 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2677 return try_to_wake_up(curr->private, mode, wake_flags);
2679 EXPORT_SYMBOL(default_wake_function);
2682 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2683 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2684 * number) then we wake all the non-exclusive tasks and one exclusive task.
2686 * There are circumstances in which we can try to wake a task which has already
2687 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2688 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2690 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2691 int nr_exclusive, int wake_flags, void *key)
2693 wait_queue_t *curr, *next;
2695 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2696 unsigned flags = curr->flags;
2698 if (curr->func(curr, mode, wake_flags, key) &&
2699 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2705 * __wake_up - wake up threads blocked on a waitqueue.
2707 * @mode: which threads
2708 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2709 * @key: is directly passed to the wakeup function
2711 * It may be assumed that this function implies a write memory barrier before
2712 * changing the task state if and only if any tasks are woken up.
2714 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2715 int nr_exclusive, void *key)
2717 unsigned long flags;
2719 spin_lock_irqsave(&q->lock, flags);
2720 __wake_up_common(q, mode, nr_exclusive, 0, key);
2721 spin_unlock_irqrestore(&q->lock, flags);
2723 EXPORT_SYMBOL(__wake_up);
2726 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2728 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2730 __wake_up_common(q, mode, nr, 0, NULL);
2732 EXPORT_SYMBOL_GPL(__wake_up_locked);
2734 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2736 __wake_up_common(q, mode, 1, 0, key);
2738 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2741 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2743 * @mode: which threads
2744 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2745 * @key: opaque value to be passed to wakeup targets
2747 * The sync wakeup differs that the waker knows that it will schedule
2748 * away soon, so while the target thread will be woken up, it will not
2749 * be migrated to another CPU - ie. the two threads are 'synchronized'
2750 * with each other. This can prevent needless bouncing between CPUs.
2752 * On UP it can prevent extra preemption.
2754 * It may be assumed that this function implies a write memory barrier before
2755 * changing the task state if and only if any tasks are woken up.
2757 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2758 int nr_exclusive, void *key)
2760 unsigned long flags;
2761 int wake_flags = WF_SYNC;
2766 if (unlikely(nr_exclusive != 1))
2769 spin_lock_irqsave(&q->lock, flags);
2770 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2771 spin_unlock_irqrestore(&q->lock, flags);
2773 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2776 * __wake_up_sync - see __wake_up_sync_key()
2778 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2780 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2782 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2785 * complete: - signals a single thread waiting on this completion
2786 * @x: holds the state of this particular completion
2788 * This will wake up a single thread waiting on this completion. Threads will be
2789 * awakened in the same order in which they were queued.
2791 * See also complete_all(), wait_for_completion() and related routines.
2793 * It may be assumed that this function implies a write memory barrier before
2794 * changing the task state if and only if any tasks are woken up.
2796 void complete(struct completion *x)
2798 unsigned long flags;
2800 spin_lock_irqsave(&x->wait.lock, flags);
2802 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2803 spin_unlock_irqrestore(&x->wait.lock, flags);
2805 EXPORT_SYMBOL(complete);
2808 * complete_all: - signals all threads waiting on this completion
2809 * @x: holds the state of this particular completion
2811 * This will wake up all threads waiting on this particular completion event.
2813 * It may be assumed that this function implies a write memory barrier before
2814 * changing the task state if and only if any tasks are woken up.
2816 void complete_all(struct completion *x)
2818 unsigned long flags;
2820 spin_lock_irqsave(&x->wait.lock, flags);
2821 x->done += UINT_MAX/2;
2822 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2823 spin_unlock_irqrestore(&x->wait.lock, flags);
2825 EXPORT_SYMBOL(complete_all);
2827 static inline long __sched
2828 do_wait_for_common(struct completion *x,
2829 long (*action)(long), long timeout, int state)
2832 DECLARE_WAITQUEUE(wait, current);
2834 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2836 if (signal_pending_state(state, current)) {
2837 timeout = -ERESTARTSYS;
2840 __set_current_state(state);
2841 spin_unlock_irq(&x->wait.lock);
2842 timeout = action(timeout);
2843 spin_lock_irq(&x->wait.lock);
2844 } while (!x->done && timeout);
2845 __remove_wait_queue(&x->wait, &wait);
2850 return timeout ?: 1;
2853 static inline long __sched
2854 __wait_for_common(struct completion *x,
2855 long (*action)(long), long timeout, int state)
2859 spin_lock_irq(&x->wait.lock);
2860 timeout = do_wait_for_common(x, action, timeout, state);
2861 spin_unlock_irq(&x->wait.lock);
2866 wait_for_common(struct completion *x, long timeout, int state)
2868 return __wait_for_common(x, schedule_timeout, timeout, state);
2872 wait_for_common_io(struct completion *x, long timeout, int state)
2874 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2878 * wait_for_completion: - waits for completion of a task
2879 * @x: holds the state of this particular completion
2881 * This waits to be signaled for completion of a specific task. It is NOT
2882 * interruptible and there is no timeout.
2884 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2885 * and interrupt capability. Also see complete().
2887 void __sched wait_for_completion(struct completion *x)
2889 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2891 EXPORT_SYMBOL(wait_for_completion);
2894 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2895 * @x: holds the state of this particular completion
2896 * @timeout: timeout value in jiffies
2898 * This waits for either a completion of a specific task to be signaled or for a
2899 * specified timeout to expire. The timeout is in jiffies. It is not
2902 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2903 * till timeout) if completed.
2905 unsigned long __sched
2906 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2908 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2910 EXPORT_SYMBOL(wait_for_completion_timeout);
2913 * wait_for_completion_io: - waits for completion of a task
2914 * @x: holds the state of this particular completion
2916 * This waits to be signaled for completion of a specific task. It is NOT
2917 * interruptible and there is no timeout. The caller is accounted as waiting
2920 void __sched wait_for_completion_io(struct completion *x)
2922 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2924 EXPORT_SYMBOL(wait_for_completion_io);
2927 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2928 * @x: holds the state of this particular completion
2929 * @timeout: timeout value in jiffies
2931 * This waits for either a completion of a specific task to be signaled or for a
2932 * specified timeout to expire. The timeout is in jiffies. It is not
2933 * interruptible. The caller is accounted as waiting for IO.
2935 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2936 * till timeout) if completed.
2938 unsigned long __sched
2939 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2941 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2943 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2946 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2947 * @x: holds the state of this particular completion
2949 * This waits for completion of a specific task to be signaled. It is
2952 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2954 int __sched wait_for_completion_interruptible(struct completion *x)
2956 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2957 if (t == -ERESTARTSYS)
2961 EXPORT_SYMBOL(wait_for_completion_interruptible);
2964 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2965 * @x: holds the state of this particular completion
2966 * @timeout: timeout value in jiffies
2968 * This waits for either a completion of a specific task to be signaled or for a
2969 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2971 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2972 * or number of jiffies left till timeout) if completed.
2975 wait_for_completion_interruptible_timeout(struct completion *x,
2976 unsigned long timeout)
2978 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2980 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2983 * wait_for_completion_killable: - waits for completion of a task (killable)
2984 * @x: holds the state of this particular completion
2986 * This waits to be signaled for completion of a specific task. It can be
2987 * interrupted by a kill signal.
2989 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2991 int __sched wait_for_completion_killable(struct completion *x)
2993 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2994 if (t == -ERESTARTSYS)
2998 EXPORT_SYMBOL(wait_for_completion_killable);
3001 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3002 * @x: holds the state of this particular completion
3003 * @timeout: timeout value in jiffies
3005 * This waits for either a completion of a specific task to be
3006 * signaled or for a specified timeout to expire. It can be
3007 * interrupted by a kill signal. The timeout is in jiffies.
3009 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
3010 * or number of jiffies left till timeout) if completed.
3013 wait_for_completion_killable_timeout(struct completion *x,
3014 unsigned long timeout)
3016 return wait_for_common(x, timeout, TASK_KILLABLE);
3018 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3021 * try_wait_for_completion - try to decrement a completion without blocking
3022 * @x: completion structure
3024 * Return: 0 if a decrement cannot be done without blocking
3025 * 1 if a decrement succeeded.
3027 * If a completion is being used as a counting completion,
3028 * attempt to decrement the counter without blocking. This
3029 * enables us to avoid waiting if the resource the completion
3030 * is protecting is not available.
3032 bool try_wait_for_completion(struct completion *x)
3034 unsigned long flags;
3037 spin_lock_irqsave(&x->wait.lock, flags);
3042 spin_unlock_irqrestore(&x->wait.lock, flags);
3045 EXPORT_SYMBOL(try_wait_for_completion);
3048 * completion_done - Test to see if a completion has any waiters
3049 * @x: completion structure
3051 * Return: 0 if there are waiters (wait_for_completion() in progress)
3052 * 1 if there are no waiters.
3055 bool completion_done(struct completion *x)
3057 unsigned long flags;
3060 spin_lock_irqsave(&x->wait.lock, flags);
3063 spin_unlock_irqrestore(&x->wait.lock, flags);
3066 EXPORT_SYMBOL(completion_done);
3069 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3071 unsigned long flags;
3074 init_waitqueue_entry(&wait, current);
3076 __set_current_state(state);
3078 spin_lock_irqsave(&q->lock, flags);
3079 __add_wait_queue(q, &wait);
3080 spin_unlock(&q->lock);
3081 timeout = schedule_timeout(timeout);
3082 spin_lock_irq(&q->lock);
3083 __remove_wait_queue(q, &wait);
3084 spin_unlock_irqrestore(&q->lock, flags);
3089 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3091 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3093 EXPORT_SYMBOL(interruptible_sleep_on);
3096 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3098 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3100 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3102 void __sched sleep_on(wait_queue_head_t *q)
3104 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3106 EXPORT_SYMBOL(sleep_on);
3108 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3110 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3112 EXPORT_SYMBOL(sleep_on_timeout);
3114 #ifdef CONFIG_RT_MUTEXES
3117 * rt_mutex_setprio - set the current priority of a task
3119 * @prio: prio value (kernel-internal form)
3121 * This function changes the 'effective' priority of a task. It does
3122 * not touch ->normal_prio like __setscheduler().
3124 * Used by the rt_mutex code to implement priority inheritance logic.
3126 void rt_mutex_setprio(struct task_struct *p, int prio)
3128 int oldprio, on_rq, running;
3130 const struct sched_class *prev_class;
3132 BUG_ON(prio < 0 || prio > MAX_PRIO);
3134 rq = __task_rq_lock(p);
3137 * Idle task boosting is a nono in general. There is one
3138 * exception, when PREEMPT_RT and NOHZ is active:
3140 * The idle task calls get_next_timer_interrupt() and holds
3141 * the timer wheel base->lock on the CPU and another CPU wants
3142 * to access the timer (probably to cancel it). We can safely
3143 * ignore the boosting request, as the idle CPU runs this code
3144 * with interrupts disabled and will complete the lock
3145 * protected section without being interrupted. So there is no
3146 * real need to boost.
3148 if (unlikely(p == rq->idle)) {
3149 WARN_ON(p != rq->curr);
3150 WARN_ON(p->pi_blocked_on);
3154 trace_sched_pi_setprio(p, prio);
3156 prev_class = p->sched_class;
3158 running = task_current(rq, p);
3160 dequeue_task(rq, p, 0);
3162 p->sched_class->put_prev_task(rq, p);
3165 p->sched_class = &rt_sched_class;
3167 p->sched_class = &fair_sched_class;
3172 p->sched_class->set_curr_task(rq);
3174 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3176 check_class_changed(rq, p, prev_class, oldprio);
3178 __task_rq_unlock(rq);
3181 void set_user_nice(struct task_struct *p, long nice)
3183 int old_prio, delta, on_rq;
3184 unsigned long flags;
3187 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3190 * We have to be careful, if called from sys_setpriority(),
3191 * the task might be in the middle of scheduling on another CPU.
3193 rq = task_rq_lock(p, &flags);
3195 * The RT priorities are set via sched_setscheduler(), but we still
3196 * allow the 'normal' nice value to be set - but as expected
3197 * it wont have any effect on scheduling until the task is
3198 * SCHED_FIFO/SCHED_RR:
3200 if (task_has_rt_policy(p)) {
3201 p->static_prio = NICE_TO_PRIO(nice);
3206 dequeue_task(rq, p, 0);
3208 p->static_prio = NICE_TO_PRIO(nice);
3211 p->prio = effective_prio(p);
3212 delta = p->prio - old_prio;
3215 enqueue_task(rq, p, 0);
3217 * If the task increased its priority or is running and
3218 * lowered its priority, then reschedule its CPU:
3220 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3221 resched_task(rq->curr);
3224 task_rq_unlock(rq, p, &flags);
3226 EXPORT_SYMBOL(set_user_nice);
3229 * can_nice - check if a task can reduce its nice value
3233 int can_nice(const struct task_struct *p, const int nice)
3235 /* convert nice value [19,-20] to rlimit style value [1,40] */
3236 int nice_rlim = 20 - nice;
3238 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3239 capable(CAP_SYS_NICE));
3242 #ifdef __ARCH_WANT_SYS_NICE
3245 * sys_nice - change the priority of the current process.
3246 * @increment: priority increment
3248 * sys_setpriority is a more generic, but much slower function that
3249 * does similar things.
3251 SYSCALL_DEFINE1(nice, int, increment)
3256 * Setpriority might change our priority at the same moment.
3257 * We don't have to worry. Conceptually one call occurs first
3258 * and we have a single winner.
3260 if (increment < -40)
3265 nice = TASK_NICE(current) + increment;
3271 if (increment < 0 && !can_nice(current, nice))
3274 retval = security_task_setnice(current, nice);
3278 set_user_nice(current, nice);
3285 * task_prio - return the priority value of a given task.
3286 * @p: the task in question.
3288 * Return: The priority value as seen by users in /proc.
3289 * RT tasks are offset by -200. Normal tasks are centered
3290 * around 0, value goes from -16 to +15.
3292 int task_prio(const struct task_struct *p)
3294 return p->prio - MAX_RT_PRIO;
3298 * task_nice - return the nice value of a given task.
3299 * @p: the task in question.
3301 * Return: The nice value [ -20 ... 0 ... 19 ].
3303 int task_nice(const struct task_struct *p)
3305 return TASK_NICE(p);
3307 EXPORT_SYMBOL(task_nice);
3310 * idle_cpu - is a given cpu idle currently?
3311 * @cpu: the processor in question.
3313 * Return: 1 if the CPU is currently idle. 0 otherwise.
3315 int idle_cpu(int cpu)
3317 struct rq *rq = cpu_rq(cpu);
3319 if (rq->curr != rq->idle)
3326 if (!llist_empty(&rq->wake_list))
3334 * idle_task - return the idle task for a given cpu.
3335 * @cpu: the processor in question.
3337 * Return: The idle task for the cpu @cpu.
3339 struct task_struct *idle_task(int cpu)
3341 return cpu_rq(cpu)->idle;
3345 * find_process_by_pid - find a process with a matching PID value.
3346 * @pid: the pid in question.
3348 * The task of @pid, if found. %NULL otherwise.
3350 static struct task_struct *find_process_by_pid(pid_t pid)
3352 return pid ? find_task_by_vpid(pid) : current;
3355 /* Actually do priority change: must hold rq lock. */
3357 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3360 p->rt_priority = prio;
3361 p->normal_prio = normal_prio(p);
3362 /* we are holding p->pi_lock already */
3363 p->prio = rt_mutex_getprio(p);
3364 if (rt_prio(p->prio))
3365 p->sched_class = &rt_sched_class;
3367 p->sched_class = &fair_sched_class;
3372 * check the target process has a UID that matches the current process's
3374 static bool check_same_owner(struct task_struct *p)
3376 const struct cred *cred = current_cred(), *pcred;
3380 pcred = __task_cred(p);
3381 match = (uid_eq(cred->euid, pcred->euid) ||
3382 uid_eq(cred->euid, pcred->uid));
3387 static int __sched_setscheduler(struct task_struct *p, int policy,
3388 const struct sched_param *param, bool user)
3390 int retval, oldprio, oldpolicy = -1, on_rq, running;
3391 unsigned long flags;
3392 const struct sched_class *prev_class;
3396 /* may grab non-irq protected spin_locks */
3397 BUG_ON(in_interrupt());
3399 /* double check policy once rq lock held */
3401 reset_on_fork = p->sched_reset_on_fork;
3402 policy = oldpolicy = p->policy;
3404 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3405 policy &= ~SCHED_RESET_ON_FORK;
3407 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3408 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3409 policy != SCHED_IDLE)
3414 * Valid priorities for SCHED_FIFO and SCHED_RR are
3415 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3416 * SCHED_BATCH and SCHED_IDLE is 0.
3418 if (param->sched_priority < 0 ||
3419 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3420 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3422 if (rt_policy(policy) != (param->sched_priority != 0))
3426 * Allow unprivileged RT tasks to decrease priority:
3428 if (user && !capable(CAP_SYS_NICE)) {
3429 if (rt_policy(policy)) {
3430 unsigned long rlim_rtprio =
3431 task_rlimit(p, RLIMIT_RTPRIO);
3433 /* can't set/change the rt policy */
3434 if (policy != p->policy && !rlim_rtprio)
3437 /* can't increase priority */
3438 if (param->sched_priority > p->rt_priority &&
3439 param->sched_priority > rlim_rtprio)
3444 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3445 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3447 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3448 if (!can_nice(p, TASK_NICE(p)))
3452 /* can't change other user's priorities */
3453 if (!check_same_owner(p))
3456 /* Normal users shall not reset the sched_reset_on_fork flag */
3457 if (p->sched_reset_on_fork && !reset_on_fork)
3462 retval = security_task_setscheduler(p);
3468 * make sure no PI-waiters arrive (or leave) while we are
3469 * changing the priority of the task:
3471 * To be able to change p->policy safely, the appropriate
3472 * runqueue lock must be held.
3474 rq = task_rq_lock(p, &flags);
3477 * Changing the policy of the stop threads its a very bad idea
3479 if (p == rq->stop) {
3480 task_rq_unlock(rq, p, &flags);
3485 * If not changing anything there's no need to proceed further:
3487 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3488 param->sched_priority == p->rt_priority))) {
3489 task_rq_unlock(rq, p, &flags);
3493 #ifdef CONFIG_RT_GROUP_SCHED
3496 * Do not allow realtime tasks into groups that have no runtime
3499 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3500 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3501 !task_group_is_autogroup(task_group(p))) {
3502 task_rq_unlock(rq, p, &flags);
3508 /* recheck policy now with rq lock held */
3509 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3510 policy = oldpolicy = -1;
3511 task_rq_unlock(rq, p, &flags);
3515 running = task_current(rq, p);
3517 dequeue_task(rq, p, 0);
3519 p->sched_class->put_prev_task(rq, p);
3521 p->sched_reset_on_fork = reset_on_fork;
3524 prev_class = p->sched_class;
3525 __setscheduler(rq, p, policy, param->sched_priority);
3528 p->sched_class->set_curr_task(rq);
3530 enqueue_task(rq, p, 0);
3532 check_class_changed(rq, p, prev_class, oldprio);
3533 task_rq_unlock(rq, p, &flags);
3535 rt_mutex_adjust_pi(p);
3541 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3542 * @p: the task in question.
3543 * @policy: new policy.
3544 * @param: structure containing the new RT priority.
3546 * Return: 0 on success. An error code otherwise.
3548 * NOTE that the task may be already dead.
3550 int sched_setscheduler(struct task_struct *p, int policy,
3551 const struct sched_param *param)
3553 return __sched_setscheduler(p, policy, param, true);
3555 EXPORT_SYMBOL_GPL(sched_setscheduler);
3558 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3559 * @p: the task in question.
3560 * @policy: new policy.
3561 * @param: structure containing the new RT priority.
3563 * Just like sched_setscheduler, only don't bother checking if the
3564 * current context has permission. For example, this is needed in
3565 * stop_machine(): we create temporary high priority worker threads,
3566 * but our caller might not have that capability.
3568 * Return: 0 on success. An error code otherwise.
3570 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3571 const struct sched_param *param)
3573 return __sched_setscheduler(p, policy, param, false);
3577 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3579 struct sched_param lparam;
3580 struct task_struct *p;
3583 if (!param || pid < 0)
3585 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3590 p = find_process_by_pid(pid);
3592 retval = sched_setscheduler(p, policy, &lparam);
3599 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3600 * @pid: the pid in question.
3601 * @policy: new policy.
3602 * @param: structure containing the new RT priority.
3604 * Return: 0 on success. An error code otherwise.
3606 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3607 struct sched_param __user *, param)
3609 /* negative values for policy are not valid */
3613 return do_sched_setscheduler(pid, policy, param);
3617 * sys_sched_setparam - set/change the RT priority of a thread
3618 * @pid: the pid in question.
3619 * @param: structure containing the new RT priority.
3621 * Return: 0 on success. An error code otherwise.
3623 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3625 return do_sched_setscheduler(pid, -1, param);
3629 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3630 * @pid: the pid in question.
3632 * Return: On success, the policy of the thread. Otherwise, a negative error
3635 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3637 struct task_struct *p;
3645 p = find_process_by_pid(pid);
3647 retval = security_task_getscheduler(p);
3650 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3657 * sys_sched_getparam - get the RT priority of a thread
3658 * @pid: the pid in question.
3659 * @param: structure containing the RT priority.
3661 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3664 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3666 struct sched_param lp;
3667 struct task_struct *p;
3670 if (!param || pid < 0)
3674 p = find_process_by_pid(pid);
3679 retval = security_task_getscheduler(p);
3683 lp.sched_priority = p->rt_priority;
3687 * This one might sleep, we cannot do it with a spinlock held ...
3689 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3698 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3700 cpumask_var_t cpus_allowed, new_mask;
3701 struct task_struct *p;
3707 p = find_process_by_pid(pid);
3714 /* Prevent p going away */
3718 if (p->flags & PF_NO_SETAFFINITY) {
3722 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3726 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3728 goto out_free_cpus_allowed;
3731 if (!check_same_owner(p)) {
3733 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3740 retval = security_task_setscheduler(p);
3744 cpuset_cpus_allowed(p, cpus_allowed);
3745 cpumask_and(new_mask, in_mask, cpus_allowed);
3747 retval = set_cpus_allowed_ptr(p, new_mask);
3750 cpuset_cpus_allowed(p, cpus_allowed);
3751 if (!cpumask_subset(new_mask, cpus_allowed)) {
3753 * We must have raced with a concurrent cpuset
3754 * update. Just reset the cpus_allowed to the
3755 * cpuset's cpus_allowed
3757 cpumask_copy(new_mask, cpus_allowed);
3762 free_cpumask_var(new_mask);
3763 out_free_cpus_allowed:
3764 free_cpumask_var(cpus_allowed);
3771 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3772 struct cpumask *new_mask)
3774 if (len < cpumask_size())
3775 cpumask_clear(new_mask);
3776 else if (len > cpumask_size())
3777 len = cpumask_size();
3779 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3783 * sys_sched_setaffinity - set the cpu affinity of a process
3784 * @pid: pid of the process
3785 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3786 * @user_mask_ptr: user-space pointer to the new cpu mask
3788 * Return: 0 on success. An error code otherwise.
3790 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3791 unsigned long __user *, user_mask_ptr)
3793 cpumask_var_t new_mask;
3796 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3799 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3801 retval = sched_setaffinity(pid, new_mask);
3802 free_cpumask_var(new_mask);
3806 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3808 struct task_struct *p;
3809 unsigned long flags;
3816 p = find_process_by_pid(pid);
3820 retval = security_task_getscheduler(p);
3824 raw_spin_lock_irqsave(&p->pi_lock, flags);
3825 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3826 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3836 * sys_sched_getaffinity - get the cpu affinity of a process
3837 * @pid: pid of the process
3838 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3839 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3841 * Return: 0 on success. An error code otherwise.
3843 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3844 unsigned long __user *, user_mask_ptr)
3849 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3851 if (len & (sizeof(unsigned long)-1))
3854 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3857 ret = sched_getaffinity(pid, mask);
3859 size_t retlen = min_t(size_t, len, cpumask_size());
3861 if (copy_to_user(user_mask_ptr, mask, retlen))
3866 free_cpumask_var(mask);
3872 * sys_sched_yield - yield the current processor to other threads.
3874 * This function yields the current CPU to other tasks. If there are no
3875 * other threads running on this CPU then this function will return.
3879 SYSCALL_DEFINE0(sched_yield)
3881 struct rq *rq = this_rq_lock();
3883 schedstat_inc(rq, yld_count);
3884 current->sched_class->yield_task(rq);
3887 * Since we are going to call schedule() anyway, there's
3888 * no need to preempt or enable interrupts:
3890 __release(rq->lock);
3891 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3892 do_raw_spin_unlock(&rq->lock);
3893 sched_preempt_enable_no_resched();
3900 static void __cond_resched(void)
3902 __preempt_count_add(PREEMPT_ACTIVE);
3904 __preempt_count_sub(PREEMPT_ACTIVE);
3907 int __sched _cond_resched(void)
3909 if (should_resched()) {
3915 EXPORT_SYMBOL(_cond_resched);
3918 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3919 * call schedule, and on return reacquire the lock.
3921 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3922 * operations here to prevent schedule() from being called twice (once via
3923 * spin_unlock(), once by hand).
3925 int __cond_resched_lock(spinlock_t *lock)
3927 int resched = should_resched();
3930 lockdep_assert_held(lock);
3932 if (spin_needbreak(lock) || resched) {
3943 EXPORT_SYMBOL(__cond_resched_lock);
3945 int __sched __cond_resched_softirq(void)
3947 BUG_ON(!in_softirq());
3949 if (should_resched()) {
3957 EXPORT_SYMBOL(__cond_resched_softirq);
3960 * yield - yield the current processor to other threads.
3962 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3964 * The scheduler is at all times free to pick the calling task as the most
3965 * eligible task to run, if removing the yield() call from your code breaks
3966 * it, its already broken.
3968 * Typical broken usage is:
3973 * where one assumes that yield() will let 'the other' process run that will
3974 * make event true. If the current task is a SCHED_FIFO task that will never
3975 * happen. Never use yield() as a progress guarantee!!
3977 * If you want to use yield() to wait for something, use wait_event().
3978 * If you want to use yield() to be 'nice' for others, use cond_resched().
3979 * If you still want to use yield(), do not!
3981 void __sched yield(void)
3983 set_current_state(TASK_RUNNING);
3986 EXPORT_SYMBOL(yield);
3989 * yield_to - yield the current processor to another thread in
3990 * your thread group, or accelerate that thread toward the
3991 * processor it's on.
3993 * @preempt: whether task preemption is allowed or not
3995 * It's the caller's job to ensure that the target task struct
3996 * can't go away on us before we can do any checks.
3999 * true (>0) if we indeed boosted the target task.
4000 * false (0) if we failed to boost the target.
4001 * -ESRCH if there's no task to yield to.
4003 bool __sched yield_to(struct task_struct *p, bool preempt)
4005 struct task_struct *curr = current;
4006 struct rq *rq, *p_rq;
4007 unsigned long flags;
4010 local_irq_save(flags);
4016 * If we're the only runnable task on the rq and target rq also
4017 * has only one task, there's absolutely no point in yielding.
4019 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4024 double_rq_lock(rq, p_rq);
4025 while (task_rq(p) != p_rq) {
4026 double_rq_unlock(rq, p_rq);
4030 if (!curr->sched_class->yield_to_task)
4033 if (curr->sched_class != p->sched_class)
4036 if (task_running(p_rq, p) || p->state)
4039 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4041 schedstat_inc(rq, yld_count);
4043 * Make p's CPU reschedule; pick_next_entity takes care of
4046 if (preempt && rq != p_rq)
4047 resched_task(p_rq->curr);
4051 double_rq_unlock(rq, p_rq);
4053 local_irq_restore(flags);
4060 EXPORT_SYMBOL_GPL(yield_to);
4063 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4064 * that process accounting knows that this is a task in IO wait state.
4066 void __sched io_schedule(void)
4068 struct rq *rq = raw_rq();
4070 delayacct_blkio_start();
4071 atomic_inc(&rq->nr_iowait);
4072 blk_flush_plug(current);
4073 current->in_iowait = 1;
4075 current->in_iowait = 0;
4076 atomic_dec(&rq->nr_iowait);
4077 delayacct_blkio_end();
4079 EXPORT_SYMBOL(io_schedule);
4081 long __sched io_schedule_timeout(long timeout)
4083 struct rq *rq = raw_rq();
4086 delayacct_blkio_start();
4087 atomic_inc(&rq->nr_iowait);
4088 blk_flush_plug(current);
4089 current->in_iowait = 1;
4090 ret = schedule_timeout(timeout);
4091 current->in_iowait = 0;
4092 atomic_dec(&rq->nr_iowait);
4093 delayacct_blkio_end();
4098 * sys_sched_get_priority_max - return maximum RT priority.
4099 * @policy: scheduling class.
4101 * Return: On success, this syscall returns the maximum
4102 * rt_priority that can be used by a given scheduling class.
4103 * On failure, a negative error code is returned.
4105 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4112 ret = MAX_USER_RT_PRIO-1;
4124 * sys_sched_get_priority_min - return minimum RT priority.
4125 * @policy: scheduling class.
4127 * Return: On success, this syscall returns the minimum
4128 * rt_priority that can be used by a given scheduling class.
4129 * On failure, a negative error code is returned.
4131 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4149 * sys_sched_rr_get_interval - return the default timeslice of a process.
4150 * @pid: pid of the process.
4151 * @interval: userspace pointer to the timeslice value.
4153 * this syscall writes the default timeslice value of a given process
4154 * into the user-space timespec buffer. A value of '0' means infinity.
4156 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4159 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4160 struct timespec __user *, interval)
4162 struct task_struct *p;
4163 unsigned int time_slice;
4164 unsigned long flags;
4174 p = find_process_by_pid(pid);
4178 retval = security_task_getscheduler(p);
4182 rq = task_rq_lock(p, &flags);
4183 time_slice = p->sched_class->get_rr_interval(rq, p);
4184 task_rq_unlock(rq, p, &flags);
4187 jiffies_to_timespec(time_slice, &t);
4188 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4196 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4198 void sched_show_task(struct task_struct *p)
4200 unsigned long free = 0;
4204 state = p->state ? __ffs(p->state) + 1 : 0;
4205 printk(KERN_INFO "%-15.15s %c", p->comm,
4206 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4207 #if BITS_PER_LONG == 32
4208 if (state == TASK_RUNNING)
4209 printk(KERN_CONT " running ");
4211 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4213 if (state == TASK_RUNNING)
4214 printk(KERN_CONT " running task ");
4216 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4218 #ifdef CONFIG_DEBUG_STACK_USAGE
4219 free = stack_not_used(p);
4222 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4224 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4225 task_pid_nr(p), ppid,
4226 (unsigned long)task_thread_info(p)->flags);
4228 print_worker_info(KERN_INFO, p);
4229 show_stack(p, NULL);
4232 void show_state_filter(unsigned long state_filter)
4234 struct task_struct *g, *p;
4236 #if BITS_PER_LONG == 32
4238 " task PC stack pid father\n");
4241 " task PC stack pid father\n");
4244 do_each_thread(g, p) {
4246 * reset the NMI-timeout, listing all files on a slow
4247 * console might take a lot of time:
4249 touch_nmi_watchdog();
4250 if (!state_filter || (p->state & state_filter))
4252 } while_each_thread(g, p);
4254 touch_all_softlockup_watchdogs();
4256 #ifdef CONFIG_SCHED_DEBUG
4257 sysrq_sched_debug_show();
4261 * Only show locks if all tasks are dumped:
4264 debug_show_all_locks();
4267 void init_idle_bootup_task(struct task_struct *idle)
4269 idle->sched_class = &idle_sched_class;
4273 * init_idle - set up an idle thread for a given CPU
4274 * @idle: task in question
4275 * @cpu: cpu the idle task belongs to
4277 * NOTE: this function does not set the idle thread's NEED_RESCHED
4278 * flag, to make booting more robust.
4280 void init_idle(struct task_struct *idle, int cpu)
4282 struct rq *rq = cpu_rq(cpu);
4283 unsigned long flags;
4285 raw_spin_lock_irqsave(&rq->lock, flags);
4288 idle->state = TASK_RUNNING;
4289 idle->se.exec_start = sched_clock();
4291 do_set_cpus_allowed(idle, cpumask_of(cpu));
4293 * We're having a chicken and egg problem, even though we are
4294 * holding rq->lock, the cpu isn't yet set to this cpu so the
4295 * lockdep check in task_group() will fail.
4297 * Similar case to sched_fork(). / Alternatively we could
4298 * use task_rq_lock() here and obtain the other rq->lock.
4303 __set_task_cpu(idle, cpu);
4306 rq->curr = rq->idle = idle;
4307 #if defined(CONFIG_SMP)
4310 raw_spin_unlock_irqrestore(&rq->lock, flags);
4312 /* Set the preempt count _outside_ the spinlocks! */
4313 init_idle_preempt_count(idle, cpu);
4316 * The idle tasks have their own, simple scheduling class:
4318 idle->sched_class = &idle_sched_class;
4319 ftrace_graph_init_idle_task(idle, cpu);
4320 vtime_init_idle(idle, cpu);
4321 #if defined(CONFIG_SMP)
4322 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4327 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4329 if (p->sched_class && p->sched_class->set_cpus_allowed)
4330 p->sched_class->set_cpus_allowed(p, new_mask);
4332 cpumask_copy(&p->cpus_allowed, new_mask);
4333 p->nr_cpus_allowed = cpumask_weight(new_mask);
4337 * This is how migration works:
4339 * 1) we invoke migration_cpu_stop() on the target CPU using
4341 * 2) stopper starts to run (implicitly forcing the migrated thread
4343 * 3) it checks whether the migrated task is still in the wrong runqueue.
4344 * 4) if it's in the wrong runqueue then the migration thread removes
4345 * it and puts it into the right queue.
4346 * 5) stopper completes and stop_one_cpu() returns and the migration
4351 * Change a given task's CPU affinity. Migrate the thread to a
4352 * proper CPU and schedule it away if the CPU it's executing on
4353 * is removed from the allowed bitmask.
4355 * NOTE: the caller must have a valid reference to the task, the
4356 * task must not exit() & deallocate itself prematurely. The
4357 * call is not atomic; no spinlocks may be held.
4359 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4361 unsigned long flags;
4363 unsigned int dest_cpu;
4366 rq = task_rq_lock(p, &flags);
4368 if (cpumask_equal(&p->cpus_allowed, new_mask))
4371 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4376 do_set_cpus_allowed(p, new_mask);
4378 /* Can the task run on the task's current CPU? If so, we're done */
4379 if (cpumask_test_cpu(task_cpu(p), new_mask))
4382 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4384 struct migration_arg arg = { p, dest_cpu };
4385 /* Need help from migration thread: drop lock and wait. */
4386 task_rq_unlock(rq, p, &flags);
4387 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4388 tlb_migrate_finish(p->mm);
4392 task_rq_unlock(rq, p, &flags);
4396 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4399 * Move (not current) task off this cpu, onto dest cpu. We're doing
4400 * this because either it can't run here any more (set_cpus_allowed()
4401 * away from this CPU, or CPU going down), or because we're
4402 * attempting to rebalance this task on exec (sched_exec).
4404 * So we race with normal scheduler movements, but that's OK, as long
4405 * as the task is no longer on this CPU.
4407 * Returns non-zero if task was successfully migrated.
4409 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4411 struct rq *rq_dest, *rq_src;
4414 if (unlikely(!cpu_active(dest_cpu)))
4417 rq_src = cpu_rq(src_cpu);
4418 rq_dest = cpu_rq(dest_cpu);
4420 raw_spin_lock(&p->pi_lock);
4421 double_rq_lock(rq_src, rq_dest);
4422 /* Already moved. */
4423 if (task_cpu(p) != src_cpu)
4425 /* Affinity changed (again). */
4426 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4430 * If we're not on a rq, the next wake-up will ensure we're
4434 dequeue_task(rq_src, p, 0);
4435 set_task_cpu(p, dest_cpu);
4436 enqueue_task(rq_dest, p, 0);
4437 check_preempt_curr(rq_dest, p, 0);
4442 double_rq_unlock(rq_src, rq_dest);
4443 raw_spin_unlock(&p->pi_lock);
4447 #ifdef CONFIG_NUMA_BALANCING
4448 /* Migrate current task p to target_cpu */
4449 int migrate_task_to(struct task_struct *p, int target_cpu)
4451 struct migration_arg arg = { p, target_cpu };
4452 int curr_cpu = task_cpu(p);
4454 if (curr_cpu == target_cpu)
4457 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4460 /* TODO: This is not properly updating schedstats */
4462 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4467 * migration_cpu_stop - this will be executed by a highprio stopper thread
4468 * and performs thread migration by bumping thread off CPU then
4469 * 'pushing' onto another runqueue.
4471 static int migration_cpu_stop(void *data)
4473 struct migration_arg *arg = data;
4476 * The original target cpu might have gone down and we might
4477 * be on another cpu but it doesn't matter.
4479 local_irq_disable();
4480 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4485 #ifdef CONFIG_HOTPLUG_CPU
4488 * Ensures that the idle task is using init_mm right before its cpu goes
4491 void idle_task_exit(void)
4493 struct mm_struct *mm = current->active_mm;
4495 BUG_ON(cpu_online(smp_processor_id()));
4498 switch_mm(mm, &init_mm, current);
4503 * Since this CPU is going 'away' for a while, fold any nr_active delta
4504 * we might have. Assumes we're called after migrate_tasks() so that the
4505 * nr_active count is stable.
4507 * Also see the comment "Global load-average calculations".
4509 static void calc_load_migrate(struct rq *rq)
4511 long delta = calc_load_fold_active(rq);
4513 atomic_long_add(delta, &calc_load_tasks);
4517 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4518 * try_to_wake_up()->select_task_rq().
4520 * Called with rq->lock held even though we'er in stop_machine() and
4521 * there's no concurrency possible, we hold the required locks anyway
4522 * because of lock validation efforts.
4524 static void migrate_tasks(unsigned int dead_cpu)
4526 struct rq *rq = cpu_rq(dead_cpu);
4527 struct task_struct *next, *stop = rq->stop;
4531 * Fudge the rq selection such that the below task selection loop
4532 * doesn't get stuck on the currently eligible stop task.
4534 * We're currently inside stop_machine() and the rq is either stuck
4535 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4536 * either way we should never end up calling schedule() until we're
4542 * put_prev_task() and pick_next_task() sched
4543 * class method both need to have an up-to-date
4544 * value of rq->clock[_task]
4546 update_rq_clock(rq);
4550 * There's this thread running, bail when that's the only
4553 if (rq->nr_running == 1)
4556 next = pick_next_task(rq);
4558 next->sched_class->put_prev_task(rq, next);
4560 /* Find suitable destination for @next, with force if needed. */
4561 dest_cpu = select_fallback_rq(dead_cpu, next);
4562 raw_spin_unlock(&rq->lock);
4564 __migrate_task(next, dead_cpu, dest_cpu);
4566 raw_spin_lock(&rq->lock);
4572 #endif /* CONFIG_HOTPLUG_CPU */
4574 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4576 static struct ctl_table sd_ctl_dir[] = {
4578 .procname = "sched_domain",
4584 static struct ctl_table sd_ctl_root[] = {
4586 .procname = "kernel",
4588 .child = sd_ctl_dir,
4593 static struct ctl_table *sd_alloc_ctl_entry(int n)
4595 struct ctl_table *entry =
4596 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4601 static void sd_free_ctl_entry(struct ctl_table **tablep)
4603 struct ctl_table *entry;
4606 * In the intermediate directories, both the child directory and
4607 * procname are dynamically allocated and could fail but the mode
4608 * will always be set. In the lowest directory the names are
4609 * static strings and all have proc handlers.
4611 for (entry = *tablep; entry->mode; entry++) {
4613 sd_free_ctl_entry(&entry->child);
4614 if (entry->proc_handler == NULL)
4615 kfree(entry->procname);
4622 static int min_load_idx = 0;
4623 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4626 set_table_entry(struct ctl_table *entry,
4627 const char *procname, void *data, int maxlen,
4628 umode_t mode, proc_handler *proc_handler,
4631 entry->procname = procname;
4633 entry->maxlen = maxlen;
4635 entry->proc_handler = proc_handler;
4638 entry->extra1 = &min_load_idx;
4639 entry->extra2 = &max_load_idx;
4643 static struct ctl_table *
4644 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4646 struct ctl_table *table = sd_alloc_ctl_entry(13);
4651 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4652 sizeof(long), 0644, proc_doulongvec_minmax, false);
4653 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4654 sizeof(long), 0644, proc_doulongvec_minmax, false);
4655 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4656 sizeof(int), 0644, proc_dointvec_minmax, true);
4657 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4658 sizeof(int), 0644, proc_dointvec_minmax, true);
4659 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4660 sizeof(int), 0644, proc_dointvec_minmax, true);
4661 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4662 sizeof(int), 0644, proc_dointvec_minmax, true);
4663 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4664 sizeof(int), 0644, proc_dointvec_minmax, true);
4665 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4666 sizeof(int), 0644, proc_dointvec_minmax, false);
4667 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4668 sizeof(int), 0644, proc_dointvec_minmax, false);
4669 set_table_entry(&table[9], "cache_nice_tries",
4670 &sd->cache_nice_tries,
4671 sizeof(int), 0644, proc_dointvec_minmax, false);
4672 set_table_entry(&table[10], "flags", &sd->flags,
4673 sizeof(int), 0644, proc_dointvec_minmax, false);
4674 set_table_entry(&table[11], "name", sd->name,
4675 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4676 /* &table[12] is terminator */
4681 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4683 struct ctl_table *entry, *table;
4684 struct sched_domain *sd;
4685 int domain_num = 0, i;
4688 for_each_domain(cpu, sd)
4690 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4695 for_each_domain(cpu, sd) {
4696 snprintf(buf, 32, "domain%d", i);
4697 entry->procname = kstrdup(buf, GFP_KERNEL);
4699 entry->child = sd_alloc_ctl_domain_table(sd);
4706 static struct ctl_table_header *sd_sysctl_header;
4707 static void register_sched_domain_sysctl(void)
4709 int i, cpu_num = num_possible_cpus();
4710 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4713 WARN_ON(sd_ctl_dir[0].child);
4714 sd_ctl_dir[0].child = entry;
4719 for_each_possible_cpu(i) {
4720 snprintf(buf, 32, "cpu%d", i);
4721 entry->procname = kstrdup(buf, GFP_KERNEL);
4723 entry->child = sd_alloc_ctl_cpu_table(i);
4727 WARN_ON(sd_sysctl_header);
4728 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4731 /* may be called multiple times per register */
4732 static void unregister_sched_domain_sysctl(void)
4734 if (sd_sysctl_header)
4735 unregister_sysctl_table(sd_sysctl_header);
4736 sd_sysctl_header = NULL;
4737 if (sd_ctl_dir[0].child)
4738 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4741 static void register_sched_domain_sysctl(void)
4744 static void unregister_sched_domain_sysctl(void)
4749 static void set_rq_online(struct rq *rq)
4752 const struct sched_class *class;
4754 cpumask_set_cpu(rq->cpu, rq->rd->online);
4757 for_each_class(class) {
4758 if (class->rq_online)
4759 class->rq_online(rq);
4764 static void set_rq_offline(struct rq *rq)
4767 const struct sched_class *class;
4769 for_each_class(class) {
4770 if (class->rq_offline)
4771 class->rq_offline(rq);
4774 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4780 * migration_call - callback that gets triggered when a CPU is added.
4781 * Here we can start up the necessary migration thread for the new CPU.
4784 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4786 int cpu = (long)hcpu;
4787 unsigned long flags;
4788 struct rq *rq = cpu_rq(cpu);
4790 switch (action & ~CPU_TASKS_FROZEN) {
4792 case CPU_UP_PREPARE:
4793 rq->calc_load_update = calc_load_update;
4797 /* Update our root-domain */
4798 raw_spin_lock_irqsave(&rq->lock, flags);
4800 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4804 raw_spin_unlock_irqrestore(&rq->lock, flags);
4807 #ifdef CONFIG_HOTPLUG_CPU
4809 sched_ttwu_pending();
4810 /* Update our root-domain */
4811 raw_spin_lock_irqsave(&rq->lock, flags);
4813 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4817 BUG_ON(rq->nr_running != 1); /* the migration thread */
4818 raw_spin_unlock_irqrestore(&rq->lock, flags);
4822 calc_load_migrate(rq);
4827 update_max_interval();
4833 * Register at high priority so that task migration (migrate_all_tasks)
4834 * happens before everything else. This has to be lower priority than
4835 * the notifier in the perf_event subsystem, though.
4837 static struct notifier_block migration_notifier = {
4838 .notifier_call = migration_call,
4839 .priority = CPU_PRI_MIGRATION,
4842 static int sched_cpu_active(struct notifier_block *nfb,
4843 unsigned long action, void *hcpu)
4845 switch (action & ~CPU_TASKS_FROZEN) {
4847 case CPU_DOWN_FAILED:
4848 set_cpu_active((long)hcpu, true);
4855 static int sched_cpu_inactive(struct notifier_block *nfb,
4856 unsigned long action, void *hcpu)
4858 switch (action & ~CPU_TASKS_FROZEN) {
4859 case CPU_DOWN_PREPARE:
4860 set_cpu_active((long)hcpu, false);
4867 static int __init migration_init(void)
4869 void *cpu = (void *)(long)smp_processor_id();
4872 /* Initialize migration for the boot CPU */
4873 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4874 BUG_ON(err == NOTIFY_BAD);
4875 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4876 register_cpu_notifier(&migration_notifier);
4878 /* Register cpu active notifiers */
4879 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4880 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4884 early_initcall(migration_init);
4889 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4891 #ifdef CONFIG_SCHED_DEBUG
4893 static __read_mostly int sched_debug_enabled;
4895 static int __init sched_debug_setup(char *str)
4897 sched_debug_enabled = 1;
4901 early_param("sched_debug", sched_debug_setup);
4903 static inline bool sched_debug(void)
4905 return sched_debug_enabled;
4908 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4909 struct cpumask *groupmask)
4911 struct sched_group *group = sd->groups;
4914 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4915 cpumask_clear(groupmask);
4917 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4919 if (!(sd->flags & SD_LOAD_BALANCE)) {
4920 printk("does not load-balance\n");
4922 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4927 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4929 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4930 printk(KERN_ERR "ERROR: domain->span does not contain "
4933 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4934 printk(KERN_ERR "ERROR: domain->groups does not contain"
4938 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4942 printk(KERN_ERR "ERROR: group is NULL\n");
4947 * Even though we initialize ->power to something semi-sane,
4948 * we leave power_orig unset. This allows us to detect if
4949 * domain iteration is still funny without causing /0 traps.
4951 if (!group->sgp->power_orig) {
4952 printk(KERN_CONT "\n");
4953 printk(KERN_ERR "ERROR: domain->cpu_power not "
4958 if (!cpumask_weight(sched_group_cpus(group))) {
4959 printk(KERN_CONT "\n");
4960 printk(KERN_ERR "ERROR: empty group\n");
4964 if (!(sd->flags & SD_OVERLAP) &&
4965 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4966 printk(KERN_CONT "\n");
4967 printk(KERN_ERR "ERROR: repeated CPUs\n");
4971 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4973 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4975 printk(KERN_CONT " %s", str);
4976 if (group->sgp->power != SCHED_POWER_SCALE) {
4977 printk(KERN_CONT " (cpu_power = %d)",
4981 group = group->next;
4982 } while (group != sd->groups);
4983 printk(KERN_CONT "\n");
4985 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4986 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4989 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4990 printk(KERN_ERR "ERROR: parent span is not a superset "
4991 "of domain->span\n");
4995 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4999 if (!sched_debug_enabled)
5003 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5007 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5010 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5018 #else /* !CONFIG_SCHED_DEBUG */
5019 # define sched_domain_debug(sd, cpu) do { } while (0)
5020 static inline bool sched_debug(void)
5024 #endif /* CONFIG_SCHED_DEBUG */
5026 static int sd_degenerate(struct sched_domain *sd)
5028 if (cpumask_weight(sched_domain_span(sd)) == 1)
5031 /* Following flags need at least 2 groups */
5032 if (sd->flags & (SD_LOAD_BALANCE |
5033 SD_BALANCE_NEWIDLE |
5037 SD_SHARE_PKG_RESOURCES)) {
5038 if (sd->groups != sd->groups->next)
5042 /* Following flags don't use groups */
5043 if (sd->flags & (SD_WAKE_AFFINE))
5050 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5052 unsigned long cflags = sd->flags, pflags = parent->flags;
5054 if (sd_degenerate(parent))
5057 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5060 /* Flags needing groups don't count if only 1 group in parent */
5061 if (parent->groups == parent->groups->next) {
5062 pflags &= ~(SD_LOAD_BALANCE |
5063 SD_BALANCE_NEWIDLE |
5067 SD_SHARE_PKG_RESOURCES |
5069 if (nr_node_ids == 1)
5070 pflags &= ~SD_SERIALIZE;
5072 if (~cflags & pflags)
5078 static void free_rootdomain(struct rcu_head *rcu)
5080 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5082 cpupri_cleanup(&rd->cpupri);
5083 free_cpumask_var(rd->rto_mask);
5084 free_cpumask_var(rd->online);
5085 free_cpumask_var(rd->span);
5089 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5091 struct root_domain *old_rd = NULL;
5092 unsigned long flags;
5094 raw_spin_lock_irqsave(&rq->lock, flags);
5099 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5102 cpumask_clear_cpu(rq->cpu, old_rd->span);
5105 * If we dont want to free the old_rt yet then
5106 * set old_rd to NULL to skip the freeing later
5109 if (!atomic_dec_and_test(&old_rd->refcount))
5113 atomic_inc(&rd->refcount);
5116 cpumask_set_cpu(rq->cpu, rd->span);
5117 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5120 raw_spin_unlock_irqrestore(&rq->lock, flags);
5123 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5126 static int init_rootdomain(struct root_domain *rd)
5128 memset(rd, 0, sizeof(*rd));
5130 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5132 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5134 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5137 if (cpupri_init(&rd->cpupri) != 0)
5142 free_cpumask_var(rd->rto_mask);
5144 free_cpumask_var(rd->online);
5146 free_cpumask_var(rd->span);
5152 * By default the system creates a single root-domain with all cpus as
5153 * members (mimicking the global state we have today).
5155 struct root_domain def_root_domain;
5157 static void init_defrootdomain(void)
5159 init_rootdomain(&def_root_domain);
5161 atomic_set(&def_root_domain.refcount, 1);
5164 static struct root_domain *alloc_rootdomain(void)
5166 struct root_domain *rd;
5168 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5172 if (init_rootdomain(rd) != 0) {
5180 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5182 struct sched_group *tmp, *first;
5191 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5196 } while (sg != first);
5199 static void free_sched_domain(struct rcu_head *rcu)
5201 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5204 * If its an overlapping domain it has private groups, iterate and
5207 if (sd->flags & SD_OVERLAP) {
5208 free_sched_groups(sd->groups, 1);
5209 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5210 kfree(sd->groups->sgp);
5216 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5218 call_rcu(&sd->rcu, free_sched_domain);
5221 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5223 for (; sd; sd = sd->parent)
5224 destroy_sched_domain(sd, cpu);
5228 * Keep a special pointer to the highest sched_domain that has
5229 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5230 * allows us to avoid some pointer chasing select_idle_sibling().
5232 * Also keep a unique ID per domain (we use the first cpu number in
5233 * the cpumask of the domain), this allows us to quickly tell if
5234 * two cpus are in the same cache domain, see cpus_share_cache().
5236 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5237 DEFINE_PER_CPU(int, sd_llc_size);
5238 DEFINE_PER_CPU(int, sd_llc_id);
5239 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5241 static void update_top_cache_domain(int cpu)
5243 struct sched_domain *sd;
5247 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5249 id = cpumask_first(sched_domain_span(sd));
5250 size = cpumask_weight(sched_domain_span(sd));
5253 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5254 per_cpu(sd_llc_size, cpu) = size;
5255 per_cpu(sd_llc_id, cpu) = id;
5257 sd = lowest_flag_domain(cpu, SD_NUMA);
5258 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5262 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5263 * hold the hotplug lock.
5266 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5268 struct rq *rq = cpu_rq(cpu);
5269 struct sched_domain *tmp;
5271 /* Remove the sched domains which do not contribute to scheduling. */
5272 for (tmp = sd; tmp; ) {
5273 struct sched_domain *parent = tmp->parent;
5277 if (sd_parent_degenerate(tmp, parent)) {
5278 tmp->parent = parent->parent;
5280 parent->parent->child = tmp;
5282 * Transfer SD_PREFER_SIBLING down in case of a
5283 * degenerate parent; the spans match for this
5284 * so the property transfers.
5286 if (parent->flags & SD_PREFER_SIBLING)
5287 tmp->flags |= SD_PREFER_SIBLING;
5288 destroy_sched_domain(parent, cpu);
5293 if (sd && sd_degenerate(sd)) {
5296 destroy_sched_domain(tmp, cpu);
5301 sched_domain_debug(sd, cpu);
5303 rq_attach_root(rq, rd);
5305 rcu_assign_pointer(rq->sd, sd);
5306 destroy_sched_domains(tmp, cpu);
5308 update_top_cache_domain(cpu);
5311 /* cpus with isolated domains */
5312 static cpumask_var_t cpu_isolated_map;
5314 /* Setup the mask of cpus configured for isolated domains */
5315 static int __init isolated_cpu_setup(char *str)
5317 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5318 cpulist_parse(str, cpu_isolated_map);
5322 __setup("isolcpus=", isolated_cpu_setup);
5324 static const struct cpumask *cpu_cpu_mask(int cpu)
5326 return cpumask_of_node(cpu_to_node(cpu));
5330 struct sched_domain **__percpu sd;
5331 struct sched_group **__percpu sg;
5332 struct sched_group_power **__percpu sgp;
5336 struct sched_domain ** __percpu sd;
5337 struct root_domain *rd;
5347 struct sched_domain_topology_level;
5349 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5350 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5352 #define SDTL_OVERLAP 0x01
5354 struct sched_domain_topology_level {
5355 sched_domain_init_f init;
5356 sched_domain_mask_f mask;
5359 struct sd_data data;
5363 * Build an iteration mask that can exclude certain CPUs from the upwards
5366 * Asymmetric node setups can result in situations where the domain tree is of
5367 * unequal depth, make sure to skip domains that already cover the entire
5370 * In that case build_sched_domains() will have terminated the iteration early
5371 * and our sibling sd spans will be empty. Domains should always include the
5372 * cpu they're built on, so check that.
5375 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5377 const struct cpumask *span = sched_domain_span(sd);
5378 struct sd_data *sdd = sd->private;
5379 struct sched_domain *sibling;
5382 for_each_cpu(i, span) {
5383 sibling = *per_cpu_ptr(sdd->sd, i);
5384 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5387 cpumask_set_cpu(i, sched_group_mask(sg));
5392 * Return the canonical balance cpu for this group, this is the first cpu
5393 * of this group that's also in the iteration mask.
5395 int group_balance_cpu(struct sched_group *sg)
5397 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5401 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5403 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5404 const struct cpumask *span = sched_domain_span(sd);
5405 struct cpumask *covered = sched_domains_tmpmask;
5406 struct sd_data *sdd = sd->private;
5407 struct sched_domain *child;
5410 cpumask_clear(covered);
5412 for_each_cpu(i, span) {
5413 struct cpumask *sg_span;
5415 if (cpumask_test_cpu(i, covered))
5418 child = *per_cpu_ptr(sdd->sd, i);
5420 /* See the comment near build_group_mask(). */
5421 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5424 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5425 GFP_KERNEL, cpu_to_node(cpu));
5430 sg_span = sched_group_cpus(sg);
5432 child = child->child;
5433 cpumask_copy(sg_span, sched_domain_span(child));
5435 cpumask_set_cpu(i, sg_span);
5437 cpumask_or(covered, covered, sg_span);
5439 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5440 if (atomic_inc_return(&sg->sgp->ref) == 1)
5441 build_group_mask(sd, sg);
5444 * Initialize sgp->power such that even if we mess up the
5445 * domains and no possible iteration will get us here, we won't
5448 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5451 * Make sure the first group of this domain contains the
5452 * canonical balance cpu. Otherwise the sched_domain iteration
5453 * breaks. See update_sg_lb_stats().
5455 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5456 group_balance_cpu(sg) == cpu)
5466 sd->groups = groups;
5471 free_sched_groups(first, 0);
5476 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5478 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5479 struct sched_domain *child = sd->child;
5482 cpu = cpumask_first(sched_domain_span(child));
5485 *sg = *per_cpu_ptr(sdd->sg, cpu);
5486 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5487 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5494 * build_sched_groups will build a circular linked list of the groups
5495 * covered by the given span, and will set each group's ->cpumask correctly,
5496 * and ->cpu_power to 0.
5498 * Assumes the sched_domain tree is fully constructed
5501 build_sched_groups(struct sched_domain *sd, int cpu)
5503 struct sched_group *first = NULL, *last = NULL;
5504 struct sd_data *sdd = sd->private;
5505 const struct cpumask *span = sched_domain_span(sd);
5506 struct cpumask *covered;
5509 get_group(cpu, sdd, &sd->groups);
5510 atomic_inc(&sd->groups->ref);
5512 if (cpu != cpumask_first(span))
5515 lockdep_assert_held(&sched_domains_mutex);
5516 covered = sched_domains_tmpmask;
5518 cpumask_clear(covered);
5520 for_each_cpu(i, span) {
5521 struct sched_group *sg;
5524 if (cpumask_test_cpu(i, covered))
5527 group = get_group(i, sdd, &sg);
5528 cpumask_clear(sched_group_cpus(sg));
5530 cpumask_setall(sched_group_mask(sg));
5532 for_each_cpu(j, span) {
5533 if (get_group(j, sdd, NULL) != group)
5536 cpumask_set_cpu(j, covered);
5537 cpumask_set_cpu(j, sched_group_cpus(sg));
5552 * Initialize sched groups cpu_power.
5554 * cpu_power indicates the capacity of sched group, which is used while
5555 * distributing the load between different sched groups in a sched domain.
5556 * Typically cpu_power for all the groups in a sched domain will be same unless
5557 * there are asymmetries in the topology. If there are asymmetries, group
5558 * having more cpu_power will pickup more load compared to the group having
5561 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5563 struct sched_group *sg = sd->groups;
5568 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5570 } while (sg != sd->groups);
5572 if (cpu != group_balance_cpu(sg))
5575 update_group_power(sd, cpu);
5576 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5579 int __weak arch_sd_sibling_asym_packing(void)
5581 return 0*SD_ASYM_PACKING;
5585 * Initializers for schedule domains
5586 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5589 #ifdef CONFIG_SCHED_DEBUG
5590 # define SD_INIT_NAME(sd, type) sd->name = #type
5592 # define SD_INIT_NAME(sd, type) do { } while (0)
5595 #define SD_INIT_FUNC(type) \
5596 static noinline struct sched_domain * \
5597 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5599 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5600 *sd = SD_##type##_INIT; \
5601 SD_INIT_NAME(sd, type); \
5602 sd->private = &tl->data; \
5607 #ifdef CONFIG_SCHED_SMT
5608 SD_INIT_FUNC(SIBLING)
5610 #ifdef CONFIG_SCHED_MC
5613 #ifdef CONFIG_SCHED_BOOK
5617 static int default_relax_domain_level = -1;
5618 int sched_domain_level_max;
5620 static int __init setup_relax_domain_level(char *str)
5622 if (kstrtoint(str, 0, &default_relax_domain_level))
5623 pr_warn("Unable to set relax_domain_level\n");
5627 __setup("relax_domain_level=", setup_relax_domain_level);
5629 static void set_domain_attribute(struct sched_domain *sd,
5630 struct sched_domain_attr *attr)
5634 if (!attr || attr->relax_domain_level < 0) {
5635 if (default_relax_domain_level < 0)
5638 request = default_relax_domain_level;
5640 request = attr->relax_domain_level;
5641 if (request < sd->level) {
5642 /* turn off idle balance on this domain */
5643 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5645 /* turn on idle balance on this domain */
5646 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5650 static void __sdt_free(const struct cpumask *cpu_map);
5651 static int __sdt_alloc(const struct cpumask *cpu_map);
5653 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5654 const struct cpumask *cpu_map)
5658 if (!atomic_read(&d->rd->refcount))
5659 free_rootdomain(&d->rd->rcu); /* fall through */
5661 free_percpu(d->sd); /* fall through */
5663 __sdt_free(cpu_map); /* fall through */
5669 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5670 const struct cpumask *cpu_map)
5672 memset(d, 0, sizeof(*d));
5674 if (__sdt_alloc(cpu_map))
5675 return sa_sd_storage;
5676 d->sd = alloc_percpu(struct sched_domain *);
5678 return sa_sd_storage;
5679 d->rd = alloc_rootdomain();
5682 return sa_rootdomain;
5686 * NULL the sd_data elements we've used to build the sched_domain and
5687 * sched_group structure so that the subsequent __free_domain_allocs()
5688 * will not free the data we're using.
5690 static void claim_allocations(int cpu, struct sched_domain *sd)
5692 struct sd_data *sdd = sd->private;
5694 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5695 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5697 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5698 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5700 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5701 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5704 #ifdef CONFIG_SCHED_SMT
5705 static const struct cpumask *cpu_smt_mask(int cpu)
5707 return topology_thread_cpumask(cpu);
5712 * Topology list, bottom-up.
5714 static struct sched_domain_topology_level default_topology[] = {
5715 #ifdef CONFIG_SCHED_SMT
5716 { sd_init_SIBLING, cpu_smt_mask, },
5718 #ifdef CONFIG_SCHED_MC
5719 { sd_init_MC, cpu_coregroup_mask, },
5721 #ifdef CONFIG_SCHED_BOOK
5722 { sd_init_BOOK, cpu_book_mask, },
5724 { sd_init_CPU, cpu_cpu_mask, },
5728 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5730 #define for_each_sd_topology(tl) \
5731 for (tl = sched_domain_topology; tl->init; tl++)
5735 static int sched_domains_numa_levels;
5736 static int *sched_domains_numa_distance;
5737 static struct cpumask ***sched_domains_numa_masks;
5738 static int sched_domains_curr_level;
5740 static inline int sd_local_flags(int level)
5742 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5745 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5748 static struct sched_domain *
5749 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5751 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5752 int level = tl->numa_level;
5753 int sd_weight = cpumask_weight(
5754 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5756 *sd = (struct sched_domain){
5757 .min_interval = sd_weight,
5758 .max_interval = 2*sd_weight,
5760 .imbalance_pct = 125,
5761 .cache_nice_tries = 2,
5768 .flags = 1*SD_LOAD_BALANCE
5769 | 1*SD_BALANCE_NEWIDLE
5774 | 0*SD_SHARE_CPUPOWER
5775 | 0*SD_SHARE_PKG_RESOURCES
5777 | 0*SD_PREFER_SIBLING
5779 | sd_local_flags(level)
5781 .last_balance = jiffies,
5782 .balance_interval = sd_weight,
5784 SD_INIT_NAME(sd, NUMA);
5785 sd->private = &tl->data;
5788 * Ugly hack to pass state to sd_numa_mask()...
5790 sched_domains_curr_level = tl->numa_level;
5795 static const struct cpumask *sd_numa_mask(int cpu)
5797 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5800 static void sched_numa_warn(const char *str)
5802 static int done = false;
5810 printk(KERN_WARNING "ERROR: %s\n\n", str);
5812 for (i = 0; i < nr_node_ids; i++) {
5813 printk(KERN_WARNING " ");
5814 for (j = 0; j < nr_node_ids; j++)
5815 printk(KERN_CONT "%02d ", node_distance(i,j));
5816 printk(KERN_CONT "\n");
5818 printk(KERN_WARNING "\n");
5821 static bool find_numa_distance(int distance)
5825 if (distance == node_distance(0, 0))
5828 for (i = 0; i < sched_domains_numa_levels; i++) {
5829 if (sched_domains_numa_distance[i] == distance)
5836 static void sched_init_numa(void)
5838 int next_distance, curr_distance = node_distance(0, 0);
5839 struct sched_domain_topology_level *tl;
5843 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5844 if (!sched_domains_numa_distance)
5848 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5849 * unique distances in the node_distance() table.
5851 * Assumes node_distance(0,j) includes all distances in
5852 * node_distance(i,j) in order to avoid cubic time.
5854 next_distance = curr_distance;
5855 for (i = 0; i < nr_node_ids; i++) {
5856 for (j = 0; j < nr_node_ids; j++) {
5857 for (k = 0; k < nr_node_ids; k++) {
5858 int distance = node_distance(i, k);
5860 if (distance > curr_distance &&
5861 (distance < next_distance ||
5862 next_distance == curr_distance))
5863 next_distance = distance;
5866 * While not a strong assumption it would be nice to know
5867 * about cases where if node A is connected to B, B is not
5868 * equally connected to A.
5870 if (sched_debug() && node_distance(k, i) != distance)
5871 sched_numa_warn("Node-distance not symmetric");
5873 if (sched_debug() && i && !find_numa_distance(distance))
5874 sched_numa_warn("Node-0 not representative");
5876 if (next_distance != curr_distance) {
5877 sched_domains_numa_distance[level++] = next_distance;
5878 sched_domains_numa_levels = level;
5879 curr_distance = next_distance;
5884 * In case of sched_debug() we verify the above assumption.
5890 * 'level' contains the number of unique distances, excluding the
5891 * identity distance node_distance(i,i).
5893 * The sched_domains_numa_distance[] array includes the actual distance
5898 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5899 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5900 * the array will contain less then 'level' members. This could be
5901 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5902 * in other functions.
5904 * We reset it to 'level' at the end of this function.
5906 sched_domains_numa_levels = 0;
5908 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5909 if (!sched_domains_numa_masks)
5913 * Now for each level, construct a mask per node which contains all
5914 * cpus of nodes that are that many hops away from us.
5916 for (i = 0; i < level; i++) {
5917 sched_domains_numa_masks[i] =
5918 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5919 if (!sched_domains_numa_masks[i])
5922 for (j = 0; j < nr_node_ids; j++) {
5923 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5927 sched_domains_numa_masks[i][j] = mask;
5929 for (k = 0; k < nr_node_ids; k++) {
5930 if (node_distance(j, k) > sched_domains_numa_distance[i])
5933 cpumask_or(mask, mask, cpumask_of_node(k));
5938 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5939 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5944 * Copy the default topology bits..
5946 for (i = 0; default_topology[i].init; i++)
5947 tl[i] = default_topology[i];
5950 * .. and append 'j' levels of NUMA goodness.
5952 for (j = 0; j < level; i++, j++) {
5953 tl[i] = (struct sched_domain_topology_level){
5954 .init = sd_numa_init,
5955 .mask = sd_numa_mask,
5956 .flags = SDTL_OVERLAP,
5961 sched_domain_topology = tl;
5963 sched_domains_numa_levels = level;
5966 static void sched_domains_numa_masks_set(int cpu)
5969 int node = cpu_to_node(cpu);
5971 for (i = 0; i < sched_domains_numa_levels; i++) {
5972 for (j = 0; j < nr_node_ids; j++) {
5973 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5974 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5979 static void sched_domains_numa_masks_clear(int cpu)
5982 for (i = 0; i < sched_domains_numa_levels; i++) {
5983 for (j = 0; j < nr_node_ids; j++)
5984 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5989 * Update sched_domains_numa_masks[level][node] array when new cpus
5992 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5993 unsigned long action,
5996 int cpu = (long)hcpu;
5998 switch (action & ~CPU_TASKS_FROZEN) {
6000 sched_domains_numa_masks_set(cpu);
6004 sched_domains_numa_masks_clear(cpu);
6014 static inline void sched_init_numa(void)
6018 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6019 unsigned long action,
6024 #endif /* CONFIG_NUMA */
6026 static int __sdt_alloc(const struct cpumask *cpu_map)
6028 struct sched_domain_topology_level *tl;
6031 for_each_sd_topology(tl) {
6032 struct sd_data *sdd = &tl->data;
6034 sdd->sd = alloc_percpu(struct sched_domain *);
6038 sdd->sg = alloc_percpu(struct sched_group *);
6042 sdd->sgp = alloc_percpu(struct sched_group_power *);
6046 for_each_cpu(j, cpu_map) {
6047 struct sched_domain *sd;
6048 struct sched_group *sg;
6049 struct sched_group_power *sgp;
6051 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6052 GFP_KERNEL, cpu_to_node(j));
6056 *per_cpu_ptr(sdd->sd, j) = sd;
6058 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6059 GFP_KERNEL, cpu_to_node(j));
6065 *per_cpu_ptr(sdd->sg, j) = sg;
6067 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6068 GFP_KERNEL, cpu_to_node(j));
6072 *per_cpu_ptr(sdd->sgp, j) = sgp;
6079 static void __sdt_free(const struct cpumask *cpu_map)
6081 struct sched_domain_topology_level *tl;
6084 for_each_sd_topology(tl) {
6085 struct sd_data *sdd = &tl->data;
6087 for_each_cpu(j, cpu_map) {
6088 struct sched_domain *sd;
6091 sd = *per_cpu_ptr(sdd->sd, j);
6092 if (sd && (sd->flags & SD_OVERLAP))
6093 free_sched_groups(sd->groups, 0);
6094 kfree(*per_cpu_ptr(sdd->sd, j));
6098 kfree(*per_cpu_ptr(sdd->sg, j));
6100 kfree(*per_cpu_ptr(sdd->sgp, j));
6102 free_percpu(sdd->sd);
6104 free_percpu(sdd->sg);
6106 free_percpu(sdd->sgp);
6111 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6112 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6113 struct sched_domain *child, int cpu)
6115 struct sched_domain *sd = tl->init(tl, cpu);
6119 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6121 sd->level = child->level + 1;
6122 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6126 set_domain_attribute(sd, attr);
6132 * Build sched domains for a given set of cpus and attach the sched domains
6133 * to the individual cpus
6135 static int build_sched_domains(const struct cpumask *cpu_map,
6136 struct sched_domain_attr *attr)
6138 enum s_alloc alloc_state;
6139 struct sched_domain *sd;
6141 int i, ret = -ENOMEM;
6143 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6144 if (alloc_state != sa_rootdomain)
6147 /* Set up domains for cpus specified by the cpu_map. */
6148 for_each_cpu(i, cpu_map) {
6149 struct sched_domain_topology_level *tl;
6152 for_each_sd_topology(tl) {
6153 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6154 if (tl == sched_domain_topology)
6155 *per_cpu_ptr(d.sd, i) = sd;
6156 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6157 sd->flags |= SD_OVERLAP;
6158 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6163 /* Build the groups for the domains */
6164 for_each_cpu(i, cpu_map) {
6165 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6166 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6167 if (sd->flags & SD_OVERLAP) {
6168 if (build_overlap_sched_groups(sd, i))
6171 if (build_sched_groups(sd, i))
6177 /* Calculate CPU power for physical packages and nodes */
6178 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6179 if (!cpumask_test_cpu(i, cpu_map))
6182 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6183 claim_allocations(i, sd);
6184 init_sched_groups_power(i, sd);
6188 /* Attach the domains */
6190 for_each_cpu(i, cpu_map) {
6191 sd = *per_cpu_ptr(d.sd, i);
6192 cpu_attach_domain(sd, d.rd, i);
6198 __free_domain_allocs(&d, alloc_state, cpu_map);
6202 static cpumask_var_t *doms_cur; /* current sched domains */
6203 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6204 static struct sched_domain_attr *dattr_cur;
6205 /* attribues of custom domains in 'doms_cur' */
6208 * Special case: If a kmalloc of a doms_cur partition (array of
6209 * cpumask) fails, then fallback to a single sched domain,
6210 * as determined by the single cpumask fallback_doms.
6212 static cpumask_var_t fallback_doms;
6215 * arch_update_cpu_topology lets virtualized architectures update the
6216 * cpu core maps. It is supposed to return 1 if the topology changed
6217 * or 0 if it stayed the same.
6219 int __attribute__((weak)) arch_update_cpu_topology(void)
6224 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6227 cpumask_var_t *doms;
6229 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6232 for (i = 0; i < ndoms; i++) {
6233 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6234 free_sched_domains(doms, i);
6241 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6244 for (i = 0; i < ndoms; i++)
6245 free_cpumask_var(doms[i]);
6250 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6251 * For now this just excludes isolated cpus, but could be used to
6252 * exclude other special cases in the future.
6254 static int init_sched_domains(const struct cpumask *cpu_map)
6258 arch_update_cpu_topology();
6260 doms_cur = alloc_sched_domains(ndoms_cur);
6262 doms_cur = &fallback_doms;
6263 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6264 err = build_sched_domains(doms_cur[0], NULL);
6265 register_sched_domain_sysctl();
6271 * Detach sched domains from a group of cpus specified in cpu_map
6272 * These cpus will now be attached to the NULL domain
6274 static void detach_destroy_domains(const struct cpumask *cpu_map)
6279 for_each_cpu(i, cpu_map)
6280 cpu_attach_domain(NULL, &def_root_domain, i);
6284 /* handle null as "default" */
6285 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6286 struct sched_domain_attr *new, int idx_new)
6288 struct sched_domain_attr tmp;
6295 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6296 new ? (new + idx_new) : &tmp,
6297 sizeof(struct sched_domain_attr));
6301 * Partition sched domains as specified by the 'ndoms_new'
6302 * cpumasks in the array doms_new[] of cpumasks. This compares
6303 * doms_new[] to the current sched domain partitioning, doms_cur[].
6304 * It destroys each deleted domain and builds each new domain.
6306 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6307 * The masks don't intersect (don't overlap.) We should setup one
6308 * sched domain for each mask. CPUs not in any of the cpumasks will
6309 * not be load balanced. If the same cpumask appears both in the
6310 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6313 * The passed in 'doms_new' should be allocated using
6314 * alloc_sched_domains. This routine takes ownership of it and will
6315 * free_sched_domains it when done with it. If the caller failed the
6316 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6317 * and partition_sched_domains() will fallback to the single partition
6318 * 'fallback_doms', it also forces the domains to be rebuilt.
6320 * If doms_new == NULL it will be replaced with cpu_online_mask.
6321 * ndoms_new == 0 is a special case for destroying existing domains,
6322 * and it will not create the default domain.
6324 * Call with hotplug lock held
6326 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6327 struct sched_domain_attr *dattr_new)
6332 mutex_lock(&sched_domains_mutex);
6334 /* always unregister in case we don't destroy any domains */
6335 unregister_sched_domain_sysctl();
6337 /* Let architecture update cpu core mappings. */
6338 new_topology = arch_update_cpu_topology();
6340 n = doms_new ? ndoms_new : 0;
6342 /* Destroy deleted domains */
6343 for (i = 0; i < ndoms_cur; i++) {
6344 for (j = 0; j < n && !new_topology; j++) {
6345 if (cpumask_equal(doms_cur[i], doms_new[j])
6346 && dattrs_equal(dattr_cur, i, dattr_new, j))
6349 /* no match - a current sched domain not in new doms_new[] */
6350 detach_destroy_domains(doms_cur[i]);
6356 if (doms_new == NULL) {
6358 doms_new = &fallback_doms;
6359 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6360 WARN_ON_ONCE(dattr_new);
6363 /* Build new domains */
6364 for (i = 0; i < ndoms_new; i++) {
6365 for (j = 0; j < n && !new_topology; j++) {
6366 if (cpumask_equal(doms_new[i], doms_cur[j])
6367 && dattrs_equal(dattr_new, i, dattr_cur, j))
6370 /* no match - add a new doms_new */
6371 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6376 /* Remember the new sched domains */
6377 if (doms_cur != &fallback_doms)
6378 free_sched_domains(doms_cur, ndoms_cur);
6379 kfree(dattr_cur); /* kfree(NULL) is safe */
6380 doms_cur = doms_new;
6381 dattr_cur = dattr_new;
6382 ndoms_cur = ndoms_new;
6384 register_sched_domain_sysctl();
6386 mutex_unlock(&sched_domains_mutex);
6389 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6392 * Update cpusets according to cpu_active mask. If cpusets are
6393 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6394 * around partition_sched_domains().
6396 * If we come here as part of a suspend/resume, don't touch cpusets because we
6397 * want to restore it back to its original state upon resume anyway.
6399 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6403 case CPU_ONLINE_FROZEN:
6404 case CPU_DOWN_FAILED_FROZEN:
6407 * num_cpus_frozen tracks how many CPUs are involved in suspend
6408 * resume sequence. As long as this is not the last online
6409 * operation in the resume sequence, just build a single sched
6410 * domain, ignoring cpusets.
6413 if (likely(num_cpus_frozen)) {
6414 partition_sched_domains(1, NULL, NULL);
6419 * This is the last CPU online operation. So fall through and
6420 * restore the original sched domains by considering the
6421 * cpuset configurations.
6425 case CPU_DOWN_FAILED:
6426 cpuset_update_active_cpus(true);
6434 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6438 case CPU_DOWN_PREPARE:
6439 cpuset_update_active_cpus(false);
6441 case CPU_DOWN_PREPARE_FROZEN:
6443 partition_sched_domains(1, NULL, NULL);
6451 void __init sched_init_smp(void)
6453 cpumask_var_t non_isolated_cpus;
6455 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6456 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6461 mutex_lock(&sched_domains_mutex);
6462 init_sched_domains(cpu_active_mask);
6463 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6464 if (cpumask_empty(non_isolated_cpus))
6465 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6466 mutex_unlock(&sched_domains_mutex);
6469 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6470 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6471 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6475 /* Move init over to a non-isolated CPU */
6476 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6478 sched_init_granularity();
6479 free_cpumask_var(non_isolated_cpus);
6481 init_sched_rt_class();
6484 void __init sched_init_smp(void)
6486 sched_init_granularity();
6488 #endif /* CONFIG_SMP */
6490 const_debug unsigned int sysctl_timer_migration = 1;
6492 int in_sched_functions(unsigned long addr)
6494 return in_lock_functions(addr) ||
6495 (addr >= (unsigned long)__sched_text_start
6496 && addr < (unsigned long)__sched_text_end);
6499 #ifdef CONFIG_CGROUP_SCHED
6501 * Default task group.
6502 * Every task in system belongs to this group at bootup.
6504 struct task_group root_task_group;
6505 LIST_HEAD(task_groups);
6508 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6510 void __init sched_init(void)
6513 unsigned long alloc_size = 0, ptr;
6515 #ifdef CONFIG_FAIR_GROUP_SCHED
6516 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6518 #ifdef CONFIG_RT_GROUP_SCHED
6519 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6521 #ifdef CONFIG_CPUMASK_OFFSTACK
6522 alloc_size += num_possible_cpus() * cpumask_size();
6525 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6527 #ifdef CONFIG_FAIR_GROUP_SCHED
6528 root_task_group.se = (struct sched_entity **)ptr;
6529 ptr += nr_cpu_ids * sizeof(void **);
6531 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6532 ptr += nr_cpu_ids * sizeof(void **);
6534 #endif /* CONFIG_FAIR_GROUP_SCHED */
6535 #ifdef CONFIG_RT_GROUP_SCHED
6536 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6537 ptr += nr_cpu_ids * sizeof(void **);
6539 root_task_group.rt_rq = (struct rt_rq **)ptr;
6540 ptr += nr_cpu_ids * sizeof(void **);
6542 #endif /* CONFIG_RT_GROUP_SCHED */
6543 #ifdef CONFIG_CPUMASK_OFFSTACK
6544 for_each_possible_cpu(i) {
6545 per_cpu(load_balance_mask, i) = (void *)ptr;
6546 ptr += cpumask_size();
6548 #endif /* CONFIG_CPUMASK_OFFSTACK */
6552 init_defrootdomain();
6555 init_rt_bandwidth(&def_rt_bandwidth,
6556 global_rt_period(), global_rt_runtime());
6558 #ifdef CONFIG_RT_GROUP_SCHED
6559 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6560 global_rt_period(), global_rt_runtime());
6561 #endif /* CONFIG_RT_GROUP_SCHED */
6563 #ifdef CONFIG_CGROUP_SCHED
6564 list_add(&root_task_group.list, &task_groups);
6565 INIT_LIST_HEAD(&root_task_group.children);
6566 INIT_LIST_HEAD(&root_task_group.siblings);
6567 autogroup_init(&init_task);
6569 #endif /* CONFIG_CGROUP_SCHED */
6571 for_each_possible_cpu(i) {
6575 raw_spin_lock_init(&rq->lock);
6577 rq->calc_load_active = 0;
6578 rq->calc_load_update = jiffies + LOAD_FREQ;
6579 init_cfs_rq(&rq->cfs);
6580 init_rt_rq(&rq->rt, rq);
6581 #ifdef CONFIG_FAIR_GROUP_SCHED
6582 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6583 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6585 * How much cpu bandwidth does root_task_group get?
6587 * In case of task-groups formed thr' the cgroup filesystem, it
6588 * gets 100% of the cpu resources in the system. This overall
6589 * system cpu resource is divided among the tasks of
6590 * root_task_group and its child task-groups in a fair manner,
6591 * based on each entity's (task or task-group's) weight
6592 * (se->load.weight).
6594 * In other words, if root_task_group has 10 tasks of weight
6595 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6596 * then A0's share of the cpu resource is:
6598 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6600 * We achieve this by letting root_task_group's tasks sit
6601 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6603 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6604 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6605 #endif /* CONFIG_FAIR_GROUP_SCHED */
6607 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6608 #ifdef CONFIG_RT_GROUP_SCHED
6609 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6610 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6613 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6614 rq->cpu_load[j] = 0;
6616 rq->last_load_update_tick = jiffies;
6621 rq->cpu_power = SCHED_POWER_SCALE;
6622 rq->post_schedule = 0;
6623 rq->active_balance = 0;
6624 rq->next_balance = jiffies;
6629 rq->avg_idle = 2*sysctl_sched_migration_cost;
6630 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6632 INIT_LIST_HEAD(&rq->cfs_tasks);
6634 rq_attach_root(rq, &def_root_domain);
6635 #ifdef CONFIG_NO_HZ_COMMON
6638 #ifdef CONFIG_NO_HZ_FULL
6639 rq->last_sched_tick = 0;
6643 atomic_set(&rq->nr_iowait, 0);
6646 set_load_weight(&init_task);
6648 #ifdef CONFIG_PREEMPT_NOTIFIERS
6649 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6652 #ifdef CONFIG_RT_MUTEXES
6653 plist_head_init(&init_task.pi_waiters);
6657 * The boot idle thread does lazy MMU switching as well:
6659 atomic_inc(&init_mm.mm_count);
6660 enter_lazy_tlb(&init_mm, current);
6663 * Make us the idle thread. Technically, schedule() should not be
6664 * called from this thread, however somewhere below it might be,
6665 * but because we are the idle thread, we just pick up running again
6666 * when this runqueue becomes "idle".
6668 init_idle(current, smp_processor_id());
6670 calc_load_update = jiffies + LOAD_FREQ;
6673 * During early bootup we pretend to be a normal task:
6675 current->sched_class = &fair_sched_class;
6678 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6679 /* May be allocated at isolcpus cmdline parse time */
6680 if (cpu_isolated_map == NULL)
6681 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6682 idle_thread_set_boot_cpu();
6684 init_sched_fair_class();
6686 scheduler_running = 1;
6689 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6690 static inline int preempt_count_equals(int preempt_offset)
6692 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6694 return (nested == preempt_offset);
6697 void __might_sleep(const char *file, int line, int preempt_offset)
6699 static unsigned long prev_jiffy; /* ratelimiting */
6701 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6702 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6703 system_state != SYSTEM_RUNNING || oops_in_progress)
6705 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6707 prev_jiffy = jiffies;
6710 "BUG: sleeping function called from invalid context at %s:%d\n",
6713 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6714 in_atomic(), irqs_disabled(),
6715 current->pid, current->comm);
6717 debug_show_held_locks(current);
6718 if (irqs_disabled())
6719 print_irqtrace_events(current);
6722 EXPORT_SYMBOL(__might_sleep);
6725 #ifdef CONFIG_MAGIC_SYSRQ
6726 static void normalize_task(struct rq *rq, struct task_struct *p)
6728 const struct sched_class *prev_class = p->sched_class;
6729 int old_prio = p->prio;
6734 dequeue_task(rq, p, 0);
6735 __setscheduler(rq, p, SCHED_NORMAL, 0);
6737 enqueue_task(rq, p, 0);
6738 resched_task(rq->curr);
6741 check_class_changed(rq, p, prev_class, old_prio);
6744 void normalize_rt_tasks(void)
6746 struct task_struct *g, *p;
6747 unsigned long flags;
6750 read_lock_irqsave(&tasklist_lock, flags);
6751 do_each_thread(g, p) {
6753 * Only normalize user tasks:
6758 p->se.exec_start = 0;
6759 #ifdef CONFIG_SCHEDSTATS
6760 p->se.statistics.wait_start = 0;
6761 p->se.statistics.sleep_start = 0;
6762 p->se.statistics.block_start = 0;
6767 * Renice negative nice level userspace
6770 if (TASK_NICE(p) < 0 && p->mm)
6771 set_user_nice(p, 0);
6775 raw_spin_lock(&p->pi_lock);
6776 rq = __task_rq_lock(p);
6778 normalize_task(rq, p);
6780 __task_rq_unlock(rq);
6781 raw_spin_unlock(&p->pi_lock);
6782 } while_each_thread(g, p);
6784 read_unlock_irqrestore(&tasklist_lock, flags);
6787 #endif /* CONFIG_MAGIC_SYSRQ */
6789 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6791 * These functions are only useful for the IA64 MCA handling, or kdb.
6793 * They can only be called when the whole system has been
6794 * stopped - every CPU needs to be quiescent, and no scheduling
6795 * activity can take place. Using them for anything else would
6796 * be a serious bug, and as a result, they aren't even visible
6797 * under any other configuration.
6801 * curr_task - return the current task for a given cpu.
6802 * @cpu: the processor in question.
6804 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6806 * Return: The current task for @cpu.
6808 struct task_struct *curr_task(int cpu)
6810 return cpu_curr(cpu);
6813 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6817 * set_curr_task - set the current task for a given cpu.
6818 * @cpu: the processor in question.
6819 * @p: the task pointer to set.
6821 * Description: This function must only be used when non-maskable interrupts
6822 * are serviced on a separate stack. It allows the architecture to switch the
6823 * notion of the current task on a cpu in a non-blocking manner. This function
6824 * must be called with all CPU's synchronized, and interrupts disabled, the
6825 * and caller must save the original value of the current task (see
6826 * curr_task() above) and restore that value before reenabling interrupts and
6827 * re-starting the system.
6829 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6831 void set_curr_task(int cpu, struct task_struct *p)
6838 #ifdef CONFIG_CGROUP_SCHED
6839 /* task_group_lock serializes the addition/removal of task groups */
6840 static DEFINE_SPINLOCK(task_group_lock);
6842 static void free_sched_group(struct task_group *tg)
6844 free_fair_sched_group(tg);
6845 free_rt_sched_group(tg);
6850 /* allocate runqueue etc for a new task group */
6851 struct task_group *sched_create_group(struct task_group *parent)
6853 struct task_group *tg;
6855 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6857 return ERR_PTR(-ENOMEM);
6859 if (!alloc_fair_sched_group(tg, parent))
6862 if (!alloc_rt_sched_group(tg, parent))
6868 free_sched_group(tg);
6869 return ERR_PTR(-ENOMEM);
6872 void sched_online_group(struct task_group *tg, struct task_group *parent)
6874 unsigned long flags;
6876 spin_lock_irqsave(&task_group_lock, flags);
6877 list_add_rcu(&tg->list, &task_groups);
6879 WARN_ON(!parent); /* root should already exist */
6881 tg->parent = parent;
6882 INIT_LIST_HEAD(&tg->children);
6883 list_add_rcu(&tg->siblings, &parent->children);
6884 spin_unlock_irqrestore(&task_group_lock, flags);
6887 /* rcu callback to free various structures associated with a task group */
6888 static void free_sched_group_rcu(struct rcu_head *rhp)
6890 /* now it should be safe to free those cfs_rqs */
6891 free_sched_group(container_of(rhp, struct task_group, rcu));
6894 /* Destroy runqueue etc associated with a task group */
6895 void sched_destroy_group(struct task_group *tg)
6897 /* wait for possible concurrent references to cfs_rqs complete */
6898 call_rcu(&tg->rcu, free_sched_group_rcu);
6901 void sched_offline_group(struct task_group *tg)
6903 unsigned long flags;
6906 /* end participation in shares distribution */
6907 for_each_possible_cpu(i)
6908 unregister_fair_sched_group(tg, i);
6910 spin_lock_irqsave(&task_group_lock, flags);
6911 list_del_rcu(&tg->list);
6912 list_del_rcu(&tg->siblings);
6913 spin_unlock_irqrestore(&task_group_lock, flags);
6916 /* change task's runqueue when it moves between groups.
6917 * The caller of this function should have put the task in its new group
6918 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6919 * reflect its new group.
6921 void sched_move_task(struct task_struct *tsk)
6923 struct task_group *tg;
6925 unsigned long flags;
6928 rq = task_rq_lock(tsk, &flags);
6930 running = task_current(rq, tsk);
6934 dequeue_task(rq, tsk, 0);
6935 if (unlikely(running))
6936 tsk->sched_class->put_prev_task(rq, tsk);
6938 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6939 lockdep_is_held(&tsk->sighand->siglock)),
6940 struct task_group, css);
6941 tg = autogroup_task_group(tsk, tg);
6942 tsk->sched_task_group = tg;
6944 #ifdef CONFIG_FAIR_GROUP_SCHED
6945 if (tsk->sched_class->task_move_group)
6946 tsk->sched_class->task_move_group(tsk, on_rq);
6949 set_task_rq(tsk, task_cpu(tsk));
6951 if (unlikely(running))
6952 tsk->sched_class->set_curr_task(rq);
6954 enqueue_task(rq, tsk, 0);
6956 task_rq_unlock(rq, tsk, &flags);
6958 #endif /* CONFIG_CGROUP_SCHED */
6960 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6961 static unsigned long to_ratio(u64 period, u64 runtime)
6963 if (runtime == RUNTIME_INF)
6966 return div64_u64(runtime << 20, period);
6970 #ifdef CONFIG_RT_GROUP_SCHED
6972 * Ensure that the real time constraints are schedulable.
6974 static DEFINE_MUTEX(rt_constraints_mutex);
6976 /* Must be called with tasklist_lock held */
6977 static inline int tg_has_rt_tasks(struct task_group *tg)
6979 struct task_struct *g, *p;
6981 do_each_thread(g, p) {
6982 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6984 } while_each_thread(g, p);
6989 struct rt_schedulable_data {
6990 struct task_group *tg;
6995 static int tg_rt_schedulable(struct task_group *tg, void *data)
6997 struct rt_schedulable_data *d = data;
6998 struct task_group *child;
6999 unsigned long total, sum = 0;
7000 u64 period, runtime;
7002 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7003 runtime = tg->rt_bandwidth.rt_runtime;
7006 period = d->rt_period;
7007 runtime = d->rt_runtime;
7011 * Cannot have more runtime than the period.
7013 if (runtime > period && runtime != RUNTIME_INF)
7017 * Ensure we don't starve existing RT tasks.
7019 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7022 total = to_ratio(period, runtime);
7025 * Nobody can have more than the global setting allows.
7027 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7031 * The sum of our children's runtime should not exceed our own.
7033 list_for_each_entry_rcu(child, &tg->children, siblings) {
7034 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7035 runtime = child->rt_bandwidth.rt_runtime;
7037 if (child == d->tg) {
7038 period = d->rt_period;
7039 runtime = d->rt_runtime;
7042 sum += to_ratio(period, runtime);
7051 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7055 struct rt_schedulable_data data = {
7057 .rt_period = period,
7058 .rt_runtime = runtime,
7062 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7068 static int tg_set_rt_bandwidth(struct task_group *tg,
7069 u64 rt_period, u64 rt_runtime)
7073 mutex_lock(&rt_constraints_mutex);
7074 read_lock(&tasklist_lock);
7075 err = __rt_schedulable(tg, rt_period, rt_runtime);
7079 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7080 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7081 tg->rt_bandwidth.rt_runtime = rt_runtime;
7083 for_each_possible_cpu(i) {
7084 struct rt_rq *rt_rq = tg->rt_rq[i];
7086 raw_spin_lock(&rt_rq->rt_runtime_lock);
7087 rt_rq->rt_runtime = rt_runtime;
7088 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7090 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7092 read_unlock(&tasklist_lock);
7093 mutex_unlock(&rt_constraints_mutex);
7098 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7100 u64 rt_runtime, rt_period;
7102 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7103 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7104 if (rt_runtime_us < 0)
7105 rt_runtime = RUNTIME_INF;
7107 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7110 static long sched_group_rt_runtime(struct task_group *tg)
7114 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7117 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7118 do_div(rt_runtime_us, NSEC_PER_USEC);
7119 return rt_runtime_us;
7122 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7124 u64 rt_runtime, rt_period;
7126 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7127 rt_runtime = tg->rt_bandwidth.rt_runtime;
7132 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7135 static long sched_group_rt_period(struct task_group *tg)
7139 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7140 do_div(rt_period_us, NSEC_PER_USEC);
7141 return rt_period_us;
7144 static int sched_rt_global_constraints(void)
7146 u64 runtime, period;
7149 if (sysctl_sched_rt_period <= 0)
7152 runtime = global_rt_runtime();
7153 period = global_rt_period();
7156 * Sanity check on the sysctl variables.
7158 if (runtime > period && runtime != RUNTIME_INF)
7161 mutex_lock(&rt_constraints_mutex);
7162 read_lock(&tasklist_lock);
7163 ret = __rt_schedulable(NULL, 0, 0);
7164 read_unlock(&tasklist_lock);
7165 mutex_unlock(&rt_constraints_mutex);
7170 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7172 /* Don't accept realtime tasks when there is no way for them to run */
7173 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7179 #else /* !CONFIG_RT_GROUP_SCHED */
7180 static int sched_rt_global_constraints(void)
7182 unsigned long flags;
7185 if (sysctl_sched_rt_period <= 0)
7189 * There's always some RT tasks in the root group
7190 * -- migration, kstopmachine etc..
7192 if (sysctl_sched_rt_runtime == 0)
7195 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7196 for_each_possible_cpu(i) {
7197 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7199 raw_spin_lock(&rt_rq->rt_runtime_lock);
7200 rt_rq->rt_runtime = global_rt_runtime();
7201 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7203 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7207 #endif /* CONFIG_RT_GROUP_SCHED */
7209 int sched_rr_handler(struct ctl_table *table, int write,
7210 void __user *buffer, size_t *lenp,
7214 static DEFINE_MUTEX(mutex);
7217 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7218 /* make sure that internally we keep jiffies */
7219 /* also, writing zero resets timeslice to default */
7220 if (!ret && write) {
7221 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7222 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7224 mutex_unlock(&mutex);
7228 int sched_rt_handler(struct ctl_table *table, int write,
7229 void __user *buffer, size_t *lenp,
7233 int old_period, old_runtime;
7234 static DEFINE_MUTEX(mutex);
7237 old_period = sysctl_sched_rt_period;
7238 old_runtime = sysctl_sched_rt_runtime;
7240 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7242 if (!ret && write) {
7243 ret = sched_rt_global_constraints();
7245 sysctl_sched_rt_period = old_period;
7246 sysctl_sched_rt_runtime = old_runtime;
7248 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7249 def_rt_bandwidth.rt_period =
7250 ns_to_ktime(global_rt_period());
7253 mutex_unlock(&mutex);
7258 #ifdef CONFIG_CGROUP_SCHED
7260 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7262 return css ? container_of(css, struct task_group, css) : NULL;
7265 static struct cgroup_subsys_state *
7266 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7268 struct task_group *parent = css_tg(parent_css);
7269 struct task_group *tg;
7272 /* This is early initialization for the top cgroup */
7273 return &root_task_group.css;
7276 tg = sched_create_group(parent);
7278 return ERR_PTR(-ENOMEM);
7283 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7285 struct task_group *tg = css_tg(css);
7286 struct task_group *parent = css_tg(css_parent(css));
7289 sched_online_group(tg, parent);
7293 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7295 struct task_group *tg = css_tg(css);
7297 sched_destroy_group(tg);
7300 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7302 struct task_group *tg = css_tg(css);
7304 sched_offline_group(tg);
7307 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7308 struct cgroup_taskset *tset)
7310 struct task_struct *task;
7312 cgroup_taskset_for_each(task, css, tset) {
7313 #ifdef CONFIG_RT_GROUP_SCHED
7314 if (!sched_rt_can_attach(css_tg(css), task))
7317 /* We don't support RT-tasks being in separate groups */
7318 if (task->sched_class != &fair_sched_class)
7325 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7326 struct cgroup_taskset *tset)
7328 struct task_struct *task;
7330 cgroup_taskset_for_each(task, css, tset)
7331 sched_move_task(task);
7334 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7335 struct cgroup_subsys_state *old_css,
7336 struct task_struct *task)
7339 * cgroup_exit() is called in the copy_process() failure path.
7340 * Ignore this case since the task hasn't ran yet, this avoids
7341 * trying to poke a half freed task state from generic code.
7343 if (!(task->flags & PF_EXITING))
7346 sched_move_task(task);
7349 #ifdef CONFIG_FAIR_GROUP_SCHED
7350 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7351 struct cftype *cftype, u64 shareval)
7353 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7356 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7359 struct task_group *tg = css_tg(css);
7361 return (u64) scale_load_down(tg->shares);
7364 #ifdef CONFIG_CFS_BANDWIDTH
7365 static DEFINE_MUTEX(cfs_constraints_mutex);
7367 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7368 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7370 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7372 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7374 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7375 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7377 if (tg == &root_task_group)
7381 * Ensure we have at some amount of bandwidth every period. This is
7382 * to prevent reaching a state of large arrears when throttled via
7383 * entity_tick() resulting in prolonged exit starvation.
7385 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7389 * Likewise, bound things on the otherside by preventing insane quota
7390 * periods. This also allows us to normalize in computing quota
7393 if (period > max_cfs_quota_period)
7396 mutex_lock(&cfs_constraints_mutex);
7397 ret = __cfs_schedulable(tg, period, quota);
7401 runtime_enabled = quota != RUNTIME_INF;
7402 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7403 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7404 raw_spin_lock_irq(&cfs_b->lock);
7405 cfs_b->period = ns_to_ktime(period);
7406 cfs_b->quota = quota;
7408 __refill_cfs_bandwidth_runtime(cfs_b);
7409 /* restart the period timer (if active) to handle new period expiry */
7410 if (runtime_enabled && cfs_b->timer_active) {
7411 /* force a reprogram */
7412 cfs_b->timer_active = 0;
7413 __start_cfs_bandwidth(cfs_b);
7415 raw_spin_unlock_irq(&cfs_b->lock);
7417 for_each_possible_cpu(i) {
7418 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7419 struct rq *rq = cfs_rq->rq;
7421 raw_spin_lock_irq(&rq->lock);
7422 cfs_rq->runtime_enabled = runtime_enabled;
7423 cfs_rq->runtime_remaining = 0;
7425 if (cfs_rq->throttled)
7426 unthrottle_cfs_rq(cfs_rq);
7427 raw_spin_unlock_irq(&rq->lock);
7430 mutex_unlock(&cfs_constraints_mutex);
7435 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7439 period = ktime_to_ns(tg->cfs_bandwidth.period);
7440 if (cfs_quota_us < 0)
7441 quota = RUNTIME_INF;
7443 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7445 return tg_set_cfs_bandwidth(tg, period, quota);
7448 long tg_get_cfs_quota(struct task_group *tg)
7452 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7455 quota_us = tg->cfs_bandwidth.quota;
7456 do_div(quota_us, NSEC_PER_USEC);
7461 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7465 period = (u64)cfs_period_us * NSEC_PER_USEC;
7466 quota = tg->cfs_bandwidth.quota;
7468 return tg_set_cfs_bandwidth(tg, period, quota);
7471 long tg_get_cfs_period(struct task_group *tg)
7475 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7476 do_div(cfs_period_us, NSEC_PER_USEC);
7478 return cfs_period_us;
7481 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7484 return tg_get_cfs_quota(css_tg(css));
7487 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7488 struct cftype *cftype, s64 cfs_quota_us)
7490 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7493 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7496 return tg_get_cfs_period(css_tg(css));
7499 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7500 struct cftype *cftype, u64 cfs_period_us)
7502 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7505 struct cfs_schedulable_data {
7506 struct task_group *tg;
7511 * normalize group quota/period to be quota/max_period
7512 * note: units are usecs
7514 static u64 normalize_cfs_quota(struct task_group *tg,
7515 struct cfs_schedulable_data *d)
7523 period = tg_get_cfs_period(tg);
7524 quota = tg_get_cfs_quota(tg);
7527 /* note: these should typically be equivalent */
7528 if (quota == RUNTIME_INF || quota == -1)
7531 return to_ratio(period, quota);
7534 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7536 struct cfs_schedulable_data *d = data;
7537 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7538 s64 quota = 0, parent_quota = -1;
7541 quota = RUNTIME_INF;
7543 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7545 quota = normalize_cfs_quota(tg, d);
7546 parent_quota = parent_b->hierarchal_quota;
7549 * ensure max(child_quota) <= parent_quota, inherit when no
7552 if (quota == RUNTIME_INF)
7553 quota = parent_quota;
7554 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7557 cfs_b->hierarchal_quota = quota;
7562 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7565 struct cfs_schedulable_data data = {
7571 if (quota != RUNTIME_INF) {
7572 do_div(data.period, NSEC_PER_USEC);
7573 do_div(data.quota, NSEC_PER_USEC);
7577 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7583 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7584 struct cgroup_map_cb *cb)
7586 struct task_group *tg = css_tg(css);
7587 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7589 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7590 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7591 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7595 #endif /* CONFIG_CFS_BANDWIDTH */
7596 #endif /* CONFIG_FAIR_GROUP_SCHED */
7598 #ifdef CONFIG_RT_GROUP_SCHED
7599 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7600 struct cftype *cft, s64 val)
7602 return sched_group_set_rt_runtime(css_tg(css), val);
7605 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7608 return sched_group_rt_runtime(css_tg(css));
7611 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7612 struct cftype *cftype, u64 rt_period_us)
7614 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7617 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7620 return sched_group_rt_period(css_tg(css));
7622 #endif /* CONFIG_RT_GROUP_SCHED */
7624 static struct cftype cpu_files[] = {
7625 #ifdef CONFIG_FAIR_GROUP_SCHED
7628 .read_u64 = cpu_shares_read_u64,
7629 .write_u64 = cpu_shares_write_u64,
7632 #ifdef CONFIG_CFS_BANDWIDTH
7634 .name = "cfs_quota_us",
7635 .read_s64 = cpu_cfs_quota_read_s64,
7636 .write_s64 = cpu_cfs_quota_write_s64,
7639 .name = "cfs_period_us",
7640 .read_u64 = cpu_cfs_period_read_u64,
7641 .write_u64 = cpu_cfs_period_write_u64,
7645 .read_map = cpu_stats_show,
7648 #ifdef CONFIG_RT_GROUP_SCHED
7650 .name = "rt_runtime_us",
7651 .read_s64 = cpu_rt_runtime_read,
7652 .write_s64 = cpu_rt_runtime_write,
7655 .name = "rt_period_us",
7656 .read_u64 = cpu_rt_period_read_uint,
7657 .write_u64 = cpu_rt_period_write_uint,
7663 struct cgroup_subsys cpu_cgroup_subsys = {
7665 .css_alloc = cpu_cgroup_css_alloc,
7666 .css_free = cpu_cgroup_css_free,
7667 .css_online = cpu_cgroup_css_online,
7668 .css_offline = cpu_cgroup_css_offline,
7669 .can_attach = cpu_cgroup_can_attach,
7670 .attach = cpu_cgroup_attach,
7671 .exit = cpu_cgroup_exit,
7672 .subsys_id = cpu_cgroup_subsys_id,
7673 .base_cftypes = cpu_files,
7677 #endif /* CONFIG_CGROUP_SCHED */
7679 void dump_cpu_task(int cpu)
7681 pr_info("Task dump for CPU %d:\n", cpu);
7682 sched_show_task(cpu_curr(cpu));