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_raw_lock(&arg->src_task->pi_lock,
1053 &arg->dst_task->pi_lock);
1054 double_rq_lock(src_rq, dst_rq);
1055 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1058 if (task_cpu(arg->src_task) != arg->src_cpu)
1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1067 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1068 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1073 double_rq_unlock(src_rq, dst_rq);
1074 raw_spin_unlock(&arg->dst_task->pi_lock);
1075 raw_spin_unlock(&arg->src_task->pi_lock);
1081 * Cross migrate two tasks
1083 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1085 struct migration_swap_arg arg;
1088 arg = (struct migration_swap_arg){
1090 .src_cpu = task_cpu(cur),
1092 .dst_cpu = task_cpu(p),
1095 if (arg.src_cpu == arg.dst_cpu)
1099 * These three tests are all lockless; this is OK since all of them
1100 * will be re-checked with proper locks held further down the line.
1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1111 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1117 struct migration_arg {
1118 struct task_struct *task;
1122 static int migration_cpu_stop(void *data);
1125 * wait_task_inactive - wait for a thread to unschedule.
1127 * If @match_state is nonzero, it's the @p->state value just checked and
1128 * not expected to change. If it changes, i.e. @p might have woken up,
1129 * then return zero. When we succeed in waiting for @p to be off its CPU,
1130 * we return a positive number (its total switch count). If a second call
1131 * a short while later returns the same number, the caller can be sure that
1132 * @p has remained unscheduled the whole time.
1134 * The caller must ensure that the task *will* unschedule sometime soon,
1135 * else this function might spin for a *long* time. This function can't
1136 * be called with interrupts off, or it may introduce deadlock with
1137 * smp_call_function() if an IPI is sent by the same process we are
1138 * waiting to become inactive.
1140 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1142 unsigned long flags;
1149 * We do the initial early heuristics without holding
1150 * any task-queue locks at all. We'll only try to get
1151 * the runqueue lock when things look like they will
1157 * If the task is actively running on another CPU
1158 * still, just relax and busy-wait without holding
1161 * NOTE! Since we don't hold any locks, it's not
1162 * even sure that "rq" stays as the right runqueue!
1163 * But we don't care, since "task_running()" will
1164 * return false if the runqueue has changed and p
1165 * is actually now running somewhere else!
1167 while (task_running(rq, p)) {
1168 if (match_state && unlikely(p->state != match_state))
1174 * Ok, time to look more closely! We need the rq
1175 * lock now, to be *sure*. If we're wrong, we'll
1176 * just go back and repeat.
1178 rq = task_rq_lock(p, &flags);
1179 trace_sched_wait_task(p);
1180 running = task_running(rq, p);
1183 if (!match_state || p->state == match_state)
1184 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1185 task_rq_unlock(rq, p, &flags);
1188 * If it changed from the expected state, bail out now.
1190 if (unlikely(!ncsw))
1194 * Was it really running after all now that we
1195 * checked with the proper locks actually held?
1197 * Oops. Go back and try again..
1199 if (unlikely(running)) {
1205 * It's not enough that it's not actively running,
1206 * it must be off the runqueue _entirely_, and not
1209 * So if it was still runnable (but just not actively
1210 * running right now), it's preempted, and we should
1211 * yield - it could be a while.
1213 if (unlikely(on_rq)) {
1214 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1216 set_current_state(TASK_UNINTERRUPTIBLE);
1217 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1222 * Ahh, all good. It wasn't running, and it wasn't
1223 * runnable, which means that it will never become
1224 * running in the future either. We're all done!
1233 * kick_process - kick a running thread to enter/exit the kernel
1234 * @p: the to-be-kicked thread
1236 * Cause a process which is running on another CPU to enter
1237 * kernel-mode, without any delay. (to get signals handled.)
1239 * NOTE: this function doesn't have to take the runqueue lock,
1240 * because all it wants to ensure is that the remote task enters
1241 * the kernel. If the IPI races and the task has been migrated
1242 * to another CPU then no harm is done and the purpose has been
1245 void kick_process(struct task_struct *p)
1251 if ((cpu != smp_processor_id()) && task_curr(p))
1252 smp_send_reschedule(cpu);
1255 EXPORT_SYMBOL_GPL(kick_process);
1256 #endif /* CONFIG_SMP */
1260 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1262 static int select_fallback_rq(int cpu, struct task_struct *p)
1264 int nid = cpu_to_node(cpu);
1265 const struct cpumask *nodemask = NULL;
1266 enum { cpuset, possible, fail } state = cpuset;
1270 * If the node that the cpu is on has been offlined, cpu_to_node()
1271 * will return -1. There is no cpu on the node, and we should
1272 * select the cpu on the other node.
1275 nodemask = cpumask_of_node(nid);
1277 /* Look for allowed, online CPU in same node. */
1278 for_each_cpu(dest_cpu, nodemask) {
1279 if (!cpu_online(dest_cpu))
1281 if (!cpu_active(dest_cpu))
1283 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1289 /* Any allowed, online CPU? */
1290 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1291 if (!cpu_online(dest_cpu))
1293 if (!cpu_active(dest_cpu))
1300 /* No more Mr. Nice Guy. */
1301 cpuset_cpus_allowed_fallback(p);
1306 do_set_cpus_allowed(p, cpu_possible_mask);
1317 if (state != cpuset) {
1319 * Don't tell them about moving exiting tasks or
1320 * kernel threads (both mm NULL), since they never
1323 if (p->mm && printk_ratelimit()) {
1324 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1325 task_pid_nr(p), p->comm, cpu);
1333 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1336 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1338 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1341 * In order not to call set_task_cpu() on a blocking task we need
1342 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1345 * Since this is common to all placement strategies, this lives here.
1347 * [ this allows ->select_task() to simply return task_cpu(p) and
1348 * not worry about this generic constraint ]
1350 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1352 cpu = select_fallback_rq(task_cpu(p), p);
1357 static void update_avg(u64 *avg, u64 sample)
1359 s64 diff = sample - *avg;
1365 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1367 #ifdef CONFIG_SCHEDSTATS
1368 struct rq *rq = this_rq();
1371 int this_cpu = smp_processor_id();
1373 if (cpu == this_cpu) {
1374 schedstat_inc(rq, ttwu_local);
1375 schedstat_inc(p, se.statistics.nr_wakeups_local);
1377 struct sched_domain *sd;
1379 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1381 for_each_domain(this_cpu, sd) {
1382 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1383 schedstat_inc(sd, ttwu_wake_remote);
1390 if (wake_flags & WF_MIGRATED)
1391 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1393 #endif /* CONFIG_SMP */
1395 schedstat_inc(rq, ttwu_count);
1396 schedstat_inc(p, se.statistics.nr_wakeups);
1398 if (wake_flags & WF_SYNC)
1399 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1401 #endif /* CONFIG_SCHEDSTATS */
1404 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1406 activate_task(rq, p, en_flags);
1409 /* if a worker is waking up, notify workqueue */
1410 if (p->flags & PF_WQ_WORKER)
1411 wq_worker_waking_up(p, cpu_of(rq));
1415 * Mark the task runnable and perform wakeup-preemption.
1418 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1420 check_preempt_curr(rq, p, wake_flags);
1421 trace_sched_wakeup(p, true);
1423 p->state = TASK_RUNNING;
1425 if (p->sched_class->task_woken)
1426 p->sched_class->task_woken(rq, p);
1428 if (rq->idle_stamp) {
1429 u64 delta = rq_clock(rq) - rq->idle_stamp;
1430 u64 max = 2*rq->max_idle_balance_cost;
1432 update_avg(&rq->avg_idle, delta);
1434 if (rq->avg_idle > max)
1443 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1446 if (p->sched_contributes_to_load)
1447 rq->nr_uninterruptible--;
1450 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1451 ttwu_do_wakeup(rq, p, wake_flags);
1455 * Called in case the task @p isn't fully descheduled from its runqueue,
1456 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1457 * since all we need to do is flip p->state to TASK_RUNNING, since
1458 * the task is still ->on_rq.
1460 static int ttwu_remote(struct task_struct *p, int wake_flags)
1465 rq = __task_rq_lock(p);
1467 /* check_preempt_curr() may use rq clock */
1468 update_rq_clock(rq);
1469 ttwu_do_wakeup(rq, p, wake_flags);
1472 __task_rq_unlock(rq);
1478 static void sched_ttwu_pending(void)
1480 struct rq *rq = this_rq();
1481 struct llist_node *llist = llist_del_all(&rq->wake_list);
1482 struct task_struct *p;
1484 raw_spin_lock(&rq->lock);
1487 p = llist_entry(llist, struct task_struct, wake_entry);
1488 llist = llist_next(llist);
1489 ttwu_do_activate(rq, p, 0);
1492 raw_spin_unlock(&rq->lock);
1495 void scheduler_ipi(void)
1498 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1499 * TIF_NEED_RESCHED remotely (for the first time) will also send
1502 if (tif_need_resched())
1503 set_preempt_need_resched();
1505 if (llist_empty(&this_rq()->wake_list)
1506 && !tick_nohz_full_cpu(smp_processor_id())
1507 && !got_nohz_idle_kick())
1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1512 * traditionally all their work was done from the interrupt return
1513 * path. Now that we actually do some work, we need to make sure
1516 * Some archs already do call them, luckily irq_enter/exit nest
1519 * Arguably we should visit all archs and update all handlers,
1520 * however a fair share of IPIs are still resched only so this would
1521 * somewhat pessimize the simple resched case.
1524 tick_nohz_full_check();
1525 sched_ttwu_pending();
1528 * Check if someone kicked us for doing the nohz idle load balance.
1530 if (unlikely(got_nohz_idle_kick())) {
1531 this_rq()->idle_balance = 1;
1532 raise_softirq_irqoff(SCHED_SOFTIRQ);
1537 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1540 smp_send_reschedule(cpu);
1543 bool cpus_share_cache(int this_cpu, int that_cpu)
1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1547 #endif /* CONFIG_SMP */
1549 static void ttwu_queue(struct task_struct *p, int cpu)
1551 struct rq *rq = cpu_rq(cpu);
1553 #if defined(CONFIG_SMP)
1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1556 ttwu_queue_remote(p, cpu);
1561 raw_spin_lock(&rq->lock);
1562 ttwu_do_activate(rq, p, 0);
1563 raw_spin_unlock(&rq->lock);
1567 * try_to_wake_up - wake up a thread
1568 * @p: the thread to be awakened
1569 * @state: the mask of task states that can be woken
1570 * @wake_flags: wake modifier flags (WF_*)
1572 * Put it on the run-queue if it's not already there. The "current"
1573 * thread is always on the run-queue (except when the actual
1574 * re-schedule is in progress), and as such you're allowed to do
1575 * the simpler "current->state = TASK_RUNNING" to mark yourself
1576 * runnable without the overhead of this.
1578 * Return: %true if @p was woken up, %false if it was already running.
1579 * or @state didn't match @p's state.
1582 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1584 unsigned long flags;
1585 int cpu, success = 0;
1588 * If we are going to wake up a thread waiting for CONDITION we
1589 * need to ensure that CONDITION=1 done by the caller can not be
1590 * reordered with p->state check below. This pairs with mb() in
1591 * set_current_state() the waiting thread does.
1593 smp_mb__before_spinlock();
1594 raw_spin_lock_irqsave(&p->pi_lock, flags);
1595 if (!(p->state & state))
1598 success = 1; /* we're going to change ->state */
1601 if (p->on_rq && ttwu_remote(p, wake_flags))
1606 * If the owning (remote) cpu is still in the middle of schedule() with
1607 * this task as prev, wait until its done referencing the task.
1612 * Pairs with the smp_wmb() in finish_lock_switch().
1616 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1617 p->state = TASK_WAKING;
1619 if (p->sched_class->task_waking)
1620 p->sched_class->task_waking(p);
1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1623 if (task_cpu(p) != cpu) {
1624 wake_flags |= WF_MIGRATED;
1625 set_task_cpu(p, cpu);
1627 #endif /* CONFIG_SMP */
1631 ttwu_stat(p, cpu, wake_flags);
1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1639 * try_to_wake_up_local - try to wake up a local task with rq lock held
1640 * @p: the thread to be awakened
1642 * Put @p on the run-queue if it's not already there. The caller must
1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1646 static void try_to_wake_up_local(struct task_struct *p)
1648 struct rq *rq = task_rq(p);
1650 if (WARN_ON_ONCE(rq != this_rq()) ||
1651 WARN_ON_ONCE(p == current))
1654 lockdep_assert_held(&rq->lock);
1656 if (!raw_spin_trylock(&p->pi_lock)) {
1657 raw_spin_unlock(&rq->lock);
1658 raw_spin_lock(&p->pi_lock);
1659 raw_spin_lock(&rq->lock);
1662 if (!(p->state & TASK_NORMAL))
1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1668 ttwu_do_wakeup(rq, p, 0);
1669 ttwu_stat(p, smp_processor_id(), 0);
1671 raw_spin_unlock(&p->pi_lock);
1675 * wake_up_process - Wake up a specific process
1676 * @p: The process to be woken up.
1678 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * Return: 1 if the process was woken up, 0 if it was already running.
1683 * It may be assumed that this function implies a write memory barrier before
1684 * changing the task state if and only if any tasks are woken up.
1686 int wake_up_process(struct task_struct *p)
1688 WARN_ON(task_is_stopped_or_traced(p));
1689 return try_to_wake_up(p, TASK_NORMAL, 0);
1691 EXPORT_SYMBOL(wake_up_process);
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1695 return try_to_wake_up(p, state, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1720 INIT_LIST_HEAD(&p->rt.run_list);
1722 #ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1726 #ifdef CONFIG_NUMA_BALANCING
1727 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1728 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1729 p->mm->numa_scan_seq = 0;
1732 if (clone_flags & CLONE_VM)
1733 p->numa_preferred_nid = current->numa_preferred_nid;
1735 p->numa_preferred_nid = -1;
1737 p->node_stamp = 0ULL;
1738 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1739 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1740 p->numa_work.next = &p->numa_work;
1741 p->numa_faults = NULL;
1742 p->numa_faults_buffer = NULL;
1744 INIT_LIST_HEAD(&p->numa_entry);
1745 p->numa_group = NULL;
1746 #endif /* CONFIG_NUMA_BALANCING */
1749 #ifdef CONFIG_NUMA_BALANCING
1750 #ifdef CONFIG_SCHED_DEBUG
1751 void set_numabalancing_state(bool enabled)
1754 sched_feat_set("NUMA");
1756 sched_feat_set("NO_NUMA");
1759 __read_mostly bool numabalancing_enabled;
1761 void set_numabalancing_state(bool enabled)
1763 numabalancing_enabled = enabled;
1765 #endif /* CONFIG_SCHED_DEBUG */
1766 #endif /* CONFIG_NUMA_BALANCING */
1769 * fork()/clone()-time setup:
1771 void sched_fork(unsigned long clone_flags, struct task_struct *p)
1773 unsigned long flags;
1774 int cpu = get_cpu();
1776 __sched_fork(clone_flags, p);
1778 * We mark the process as running here. This guarantees that
1779 * nobody will actually run it, and a signal or other external
1780 * event cannot wake it up and insert it on the runqueue either.
1782 p->state = TASK_RUNNING;
1785 * Make sure we do not leak PI boosting priority to the child.
1787 p->prio = current->normal_prio;
1790 * Revert to default priority/policy on fork if requested.
1792 if (unlikely(p->sched_reset_on_fork)) {
1793 if (task_has_rt_policy(p)) {
1794 p->policy = SCHED_NORMAL;
1795 p->static_prio = NICE_TO_PRIO(0);
1797 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1798 p->static_prio = NICE_TO_PRIO(0);
1800 p->prio = p->normal_prio = __normal_prio(p);
1804 * We don't need the reset flag anymore after the fork. It has
1805 * fulfilled its duty:
1807 p->sched_reset_on_fork = 0;
1810 if (!rt_prio(p->prio))
1811 p->sched_class = &fair_sched_class;
1813 if (p->sched_class->task_fork)
1814 p->sched_class->task_fork(p);
1817 * The child is not yet in the pid-hash so no cgroup attach races,
1818 * and the cgroup is pinned to this child due to cgroup_fork()
1819 * is ran before sched_fork().
1821 * Silence PROVE_RCU.
1823 raw_spin_lock_irqsave(&p->pi_lock, flags);
1824 set_task_cpu(p, cpu);
1825 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1827 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1828 if (likely(sched_info_on()))
1829 memset(&p->sched_info, 0, sizeof(p->sched_info));
1831 #if defined(CONFIG_SMP)
1834 init_task_preempt_count(p);
1836 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1843 * wake_up_new_task - wake up a newly created task for the first time.
1845 * This function will do some initial scheduler statistics housekeeping
1846 * that must be done for every newly created context, then puts the task
1847 * on the runqueue and wakes it.
1849 void wake_up_new_task(struct task_struct *p)
1851 unsigned long flags;
1854 raw_spin_lock_irqsave(&p->pi_lock, flags);
1857 * Fork balancing, do it here and not earlier because:
1858 * - cpus_allowed can change in the fork path
1859 * - any previously selected cpu might disappear through hotplug
1861 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1864 /* Initialize new task's runnable average */
1865 init_task_runnable_average(p);
1866 rq = __task_rq_lock(p);
1867 activate_task(rq, p, 0);
1869 trace_sched_wakeup_new(p, true);
1870 check_preempt_curr(rq, p, WF_FORK);
1872 if (p->sched_class->task_woken)
1873 p->sched_class->task_woken(rq, p);
1875 task_rq_unlock(rq, p, &flags);
1878 #ifdef CONFIG_PREEMPT_NOTIFIERS
1881 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1882 * @notifier: notifier struct to register
1884 void preempt_notifier_register(struct preempt_notifier *notifier)
1886 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1888 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1891 * preempt_notifier_unregister - no longer interested in preemption notifications
1892 * @notifier: notifier struct to unregister
1894 * This is safe to call from within a preemption notifier.
1896 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1898 hlist_del(¬ifier->link);
1900 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1902 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1904 struct preempt_notifier *notifier;
1906 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1907 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1911 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1912 struct task_struct *next)
1914 struct preempt_notifier *notifier;
1916 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1917 notifier->ops->sched_out(notifier, next);
1920 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1922 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1927 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1928 struct task_struct *next)
1932 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1935 * prepare_task_switch - prepare to switch tasks
1936 * @rq: the runqueue preparing to switch
1937 * @prev: the current task that is being switched out
1938 * @next: the task we are going to switch to.
1940 * This is called with the rq lock held and interrupts off. It must
1941 * be paired with a subsequent finish_task_switch after the context
1944 * prepare_task_switch sets up locking and calls architecture specific
1948 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1949 struct task_struct *next)
1951 trace_sched_switch(prev, next);
1952 sched_info_switch(rq, prev, next);
1953 perf_event_task_sched_out(prev, next);
1954 fire_sched_out_preempt_notifiers(prev, next);
1955 prepare_lock_switch(rq, next);
1956 prepare_arch_switch(next);
1960 * finish_task_switch - clean up after a task-switch
1961 * @rq: runqueue associated with task-switch
1962 * @prev: the thread we just switched away from.
1964 * finish_task_switch must be called after the context switch, paired
1965 * with a prepare_task_switch call before the context switch.
1966 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1967 * and do any other architecture-specific cleanup actions.
1969 * Note that we may have delayed dropping an mm in context_switch(). If
1970 * so, we finish that here outside of the runqueue lock. (Doing it
1971 * with the lock held can cause deadlocks; see schedule() for
1974 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1975 __releases(rq->lock)
1977 struct mm_struct *mm = rq->prev_mm;
1983 * A task struct has one reference for the use as "current".
1984 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1985 * schedule one last time. The schedule call will never return, and
1986 * the scheduled task must drop that reference.
1987 * The test for TASK_DEAD must occur while the runqueue locks are
1988 * still held, otherwise prev could be scheduled on another cpu, die
1989 * there before we look at prev->state, and then the reference would
1991 * Manfred Spraul <manfred@colorfullife.com>
1993 prev_state = prev->state;
1994 vtime_task_switch(prev);
1995 finish_arch_switch(prev);
1996 perf_event_task_sched_in(prev, current);
1997 finish_lock_switch(rq, prev);
1998 finish_arch_post_lock_switch();
2000 fire_sched_in_preempt_notifiers(current);
2003 if (unlikely(prev_state == TASK_DEAD)) {
2004 task_numa_free(prev);
2007 * Remove function-return probe instances associated with this
2008 * task and put them back on the free list.
2010 kprobe_flush_task(prev);
2011 put_task_struct(prev);
2014 tick_nohz_task_switch(current);
2019 /* assumes rq->lock is held */
2020 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2022 if (prev->sched_class->pre_schedule)
2023 prev->sched_class->pre_schedule(rq, prev);
2026 /* rq->lock is NOT held, but preemption is disabled */
2027 static inline void post_schedule(struct rq *rq)
2029 if (rq->post_schedule) {
2030 unsigned long flags;
2032 raw_spin_lock_irqsave(&rq->lock, flags);
2033 if (rq->curr->sched_class->post_schedule)
2034 rq->curr->sched_class->post_schedule(rq);
2035 raw_spin_unlock_irqrestore(&rq->lock, flags);
2037 rq->post_schedule = 0;
2043 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2047 static inline void post_schedule(struct rq *rq)
2054 * schedule_tail - first thing a freshly forked thread must call.
2055 * @prev: the thread we just switched away from.
2057 asmlinkage void schedule_tail(struct task_struct *prev)
2058 __releases(rq->lock)
2060 struct rq *rq = this_rq();
2062 finish_task_switch(rq, prev);
2065 * FIXME: do we need to worry about rq being invalidated by the
2070 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2071 /* In this case, finish_task_switch does not reenable preemption */
2074 if (current->set_child_tid)
2075 put_user(task_pid_vnr(current), current->set_child_tid);
2079 * context_switch - switch to the new MM and the new
2080 * thread's register state.
2083 context_switch(struct rq *rq, struct task_struct *prev,
2084 struct task_struct *next)
2086 struct mm_struct *mm, *oldmm;
2088 prepare_task_switch(rq, prev, next);
2091 oldmm = prev->active_mm;
2093 * For paravirt, this is coupled with an exit in switch_to to
2094 * combine the page table reload and the switch backend into
2097 arch_start_context_switch(prev);
2100 next->active_mm = oldmm;
2101 atomic_inc(&oldmm->mm_count);
2102 enter_lazy_tlb(oldmm, next);
2104 switch_mm(oldmm, mm, next);
2107 prev->active_mm = NULL;
2108 rq->prev_mm = oldmm;
2111 * Since the runqueue lock will be released by the next
2112 * task (which is an invalid locking op but in the case
2113 * of the scheduler it's an obvious special-case), so we
2114 * do an early lockdep release here:
2116 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2117 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2120 context_tracking_task_switch(prev, next);
2121 /* Here we just switch the register state and the stack. */
2122 switch_to(prev, next, prev);
2126 * this_rq must be evaluated again because prev may have moved
2127 * CPUs since it called schedule(), thus the 'rq' on its stack
2128 * frame will be invalid.
2130 finish_task_switch(this_rq(), prev);
2134 * nr_running and nr_context_switches:
2136 * externally visible scheduler statistics: current number of runnable
2137 * threads, total number of context switches performed since bootup.
2139 unsigned long nr_running(void)
2141 unsigned long i, sum = 0;
2143 for_each_online_cpu(i)
2144 sum += cpu_rq(i)->nr_running;
2149 unsigned long long nr_context_switches(void)
2152 unsigned long long sum = 0;
2154 for_each_possible_cpu(i)
2155 sum += cpu_rq(i)->nr_switches;
2160 unsigned long nr_iowait(void)
2162 unsigned long i, sum = 0;
2164 for_each_possible_cpu(i)
2165 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2170 unsigned long nr_iowait_cpu(int cpu)
2172 struct rq *this = cpu_rq(cpu);
2173 return atomic_read(&this->nr_iowait);
2179 * sched_exec - execve() is a valuable balancing opportunity, because at
2180 * this point the task has the smallest effective memory and cache footprint.
2182 void sched_exec(void)
2184 struct task_struct *p = current;
2185 unsigned long flags;
2188 raw_spin_lock_irqsave(&p->pi_lock, flags);
2189 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2190 if (dest_cpu == smp_processor_id())
2193 if (likely(cpu_active(dest_cpu))) {
2194 struct migration_arg arg = { p, dest_cpu };
2196 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2197 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2201 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2206 DEFINE_PER_CPU(struct kernel_stat, kstat);
2207 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2209 EXPORT_PER_CPU_SYMBOL(kstat);
2210 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2213 * Return any ns on the sched_clock that have not yet been accounted in
2214 * @p in case that task is currently running.
2216 * Called with task_rq_lock() held on @rq.
2218 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2222 if (task_current(rq, p)) {
2223 update_rq_clock(rq);
2224 ns = rq_clock_task(rq) - p->se.exec_start;
2232 unsigned long long task_delta_exec(struct task_struct *p)
2234 unsigned long flags;
2238 rq = task_rq_lock(p, &flags);
2239 ns = do_task_delta_exec(p, rq);
2240 task_rq_unlock(rq, p, &flags);
2246 * Return accounted runtime for the task.
2247 * In case the task is currently running, return the runtime plus current's
2248 * pending runtime that have not been accounted yet.
2250 unsigned long long task_sched_runtime(struct task_struct *p)
2252 unsigned long flags;
2256 rq = task_rq_lock(p, &flags);
2257 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2258 task_rq_unlock(rq, p, &flags);
2264 * This function gets called by the timer code, with HZ frequency.
2265 * We call it with interrupts disabled.
2267 void scheduler_tick(void)
2269 int cpu = smp_processor_id();
2270 struct rq *rq = cpu_rq(cpu);
2271 struct task_struct *curr = rq->curr;
2275 raw_spin_lock(&rq->lock);
2276 update_rq_clock(rq);
2277 curr->sched_class->task_tick(rq, curr, 0);
2278 update_cpu_load_active(rq);
2279 raw_spin_unlock(&rq->lock);
2281 perf_event_task_tick();
2284 rq->idle_balance = idle_cpu(cpu);
2285 trigger_load_balance(rq, cpu);
2287 rq_last_tick_reset(rq);
2290 #ifdef CONFIG_NO_HZ_FULL
2292 * scheduler_tick_max_deferment
2294 * Keep at least one tick per second when a single
2295 * active task is running because the scheduler doesn't
2296 * yet completely support full dynticks environment.
2298 * This makes sure that uptime, CFS vruntime, load
2299 * balancing, etc... continue to move forward, even
2300 * with a very low granularity.
2302 * Return: Maximum deferment in nanoseconds.
2304 u64 scheduler_tick_max_deferment(void)
2306 struct rq *rq = this_rq();
2307 unsigned long next, now = ACCESS_ONCE(jiffies);
2309 next = rq->last_sched_tick + HZ;
2311 if (time_before_eq(next, now))
2314 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2318 notrace unsigned long get_parent_ip(unsigned long addr)
2320 if (in_lock_functions(addr)) {
2321 addr = CALLER_ADDR2;
2322 if (in_lock_functions(addr))
2323 addr = CALLER_ADDR3;
2328 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2329 defined(CONFIG_PREEMPT_TRACER))
2331 void __kprobes preempt_count_add(int val)
2333 #ifdef CONFIG_DEBUG_PREEMPT
2337 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2340 __preempt_count_add(val);
2341 #ifdef CONFIG_DEBUG_PREEMPT
2343 * Spinlock count overflowing soon?
2345 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2348 if (preempt_count() == val)
2349 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2351 EXPORT_SYMBOL(preempt_count_add);
2353 void __kprobes preempt_count_sub(int val)
2355 #ifdef CONFIG_DEBUG_PREEMPT
2359 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2362 * Is the spinlock portion underflowing?
2364 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2365 !(preempt_count() & PREEMPT_MASK)))
2369 if (preempt_count() == val)
2370 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2371 __preempt_count_sub(val);
2373 EXPORT_SYMBOL(preempt_count_sub);
2378 * Print scheduling while atomic bug:
2380 static noinline void __schedule_bug(struct task_struct *prev)
2382 if (oops_in_progress)
2385 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2386 prev->comm, prev->pid, preempt_count());
2388 debug_show_held_locks(prev);
2390 if (irqs_disabled())
2391 print_irqtrace_events(prev);
2393 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2397 * Various schedule()-time debugging checks and statistics:
2399 static inline void schedule_debug(struct task_struct *prev)
2402 * Test if we are atomic. Since do_exit() needs to call into
2403 * schedule() atomically, we ignore that path for now.
2404 * Otherwise, whine if we are scheduling when we should not be.
2406 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2407 __schedule_bug(prev);
2410 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2412 schedstat_inc(this_rq(), sched_count);
2415 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2417 if (prev->on_rq || rq->skip_clock_update < 0)
2418 update_rq_clock(rq);
2419 prev->sched_class->put_prev_task(rq, prev);
2423 * Pick up the highest-prio task:
2425 static inline struct task_struct *
2426 pick_next_task(struct rq *rq)
2428 const struct sched_class *class;
2429 struct task_struct *p;
2432 * Optimization: we know that if all tasks are in
2433 * the fair class we can call that function directly:
2435 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2436 p = fair_sched_class.pick_next_task(rq);
2441 for_each_class(class) {
2442 p = class->pick_next_task(rq);
2447 BUG(); /* the idle class will always have a runnable task */
2451 * __schedule() is the main scheduler function.
2453 * The main means of driving the scheduler and thus entering this function are:
2455 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2457 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2458 * paths. For example, see arch/x86/entry_64.S.
2460 * To drive preemption between tasks, the scheduler sets the flag in timer
2461 * interrupt handler scheduler_tick().
2463 * 3. Wakeups don't really cause entry into schedule(). They add a
2464 * task to the run-queue and that's it.
2466 * Now, if the new task added to the run-queue preempts the current
2467 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2468 * called on the nearest possible occasion:
2470 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2472 * - in syscall or exception context, at the next outmost
2473 * preempt_enable(). (this might be as soon as the wake_up()'s
2476 * - in IRQ context, return from interrupt-handler to
2477 * preemptible context
2479 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2482 * - cond_resched() call
2483 * - explicit schedule() call
2484 * - return from syscall or exception to user-space
2485 * - return from interrupt-handler to user-space
2487 static void __sched __schedule(void)
2489 struct task_struct *prev, *next;
2490 unsigned long *switch_count;
2496 cpu = smp_processor_id();
2498 rcu_note_context_switch(cpu);
2501 schedule_debug(prev);
2503 if (sched_feat(HRTICK))
2507 * Make sure that signal_pending_state()->signal_pending() below
2508 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2509 * done by the caller to avoid the race with signal_wake_up().
2511 smp_mb__before_spinlock();
2512 raw_spin_lock_irq(&rq->lock);
2514 switch_count = &prev->nivcsw;
2515 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2516 if (unlikely(signal_pending_state(prev->state, prev))) {
2517 prev->state = TASK_RUNNING;
2519 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2523 * If a worker went to sleep, notify and ask workqueue
2524 * whether it wants to wake up a task to maintain
2527 if (prev->flags & PF_WQ_WORKER) {
2528 struct task_struct *to_wakeup;
2530 to_wakeup = wq_worker_sleeping(prev, cpu);
2532 try_to_wake_up_local(to_wakeup);
2535 switch_count = &prev->nvcsw;
2538 pre_schedule(rq, prev);
2540 if (unlikely(!rq->nr_running))
2541 idle_balance(cpu, rq);
2543 put_prev_task(rq, prev);
2544 next = pick_next_task(rq);
2545 clear_tsk_need_resched(prev);
2546 clear_preempt_need_resched();
2547 rq->skip_clock_update = 0;
2549 if (likely(prev != next)) {
2554 context_switch(rq, prev, next); /* unlocks the rq */
2556 * The context switch have flipped the stack from under us
2557 * and restored the local variables which were saved when
2558 * this task called schedule() in the past. prev == current
2559 * is still correct, but it can be moved to another cpu/rq.
2561 cpu = smp_processor_id();
2564 raw_spin_unlock_irq(&rq->lock);
2568 sched_preempt_enable_no_resched();
2573 static inline void sched_submit_work(struct task_struct *tsk)
2575 if (!tsk->state || tsk_is_pi_blocked(tsk))
2578 * If we are going to sleep and we have plugged IO queued,
2579 * make sure to submit it to avoid deadlocks.
2581 if (blk_needs_flush_plug(tsk))
2582 blk_schedule_flush_plug(tsk);
2585 asmlinkage void __sched schedule(void)
2587 struct task_struct *tsk = current;
2589 sched_submit_work(tsk);
2592 EXPORT_SYMBOL(schedule);
2594 #ifdef CONFIG_CONTEXT_TRACKING
2595 asmlinkage void __sched schedule_user(void)
2598 * If we come here after a random call to set_need_resched(),
2599 * or we have been woken up remotely but the IPI has not yet arrived,
2600 * we haven't yet exited the RCU idle mode. Do it here manually until
2601 * we find a better solution.
2610 * schedule_preempt_disabled - called with preemption disabled
2612 * Returns with preemption disabled. Note: preempt_count must be 1
2614 void __sched schedule_preempt_disabled(void)
2616 sched_preempt_enable_no_resched();
2621 #ifdef CONFIG_PREEMPT
2623 * this is the entry point to schedule() from in-kernel preemption
2624 * off of preempt_enable. Kernel preemptions off return from interrupt
2625 * occur there and call schedule directly.
2627 asmlinkage void __sched notrace preempt_schedule(void)
2630 * If there is a non-zero preempt_count or interrupts are disabled,
2631 * we do not want to preempt the current task. Just return..
2633 if (likely(!preemptible()))
2637 __preempt_count_add(PREEMPT_ACTIVE);
2639 __preempt_count_sub(PREEMPT_ACTIVE);
2642 * Check again in case we missed a preemption opportunity
2643 * between schedule and now.
2646 } while (need_resched());
2648 EXPORT_SYMBOL(preempt_schedule);
2651 * this is the entry point to schedule() from kernel preemption
2652 * off of irq context.
2653 * Note, that this is called and return with irqs disabled. This will
2654 * protect us against recursive calling from irq.
2656 asmlinkage void __sched preempt_schedule_irq(void)
2658 enum ctx_state prev_state;
2660 /* Catch callers which need to be fixed */
2661 BUG_ON(preempt_count() || !irqs_disabled());
2663 prev_state = exception_enter();
2666 __preempt_count_add(PREEMPT_ACTIVE);
2669 local_irq_disable();
2670 __preempt_count_sub(PREEMPT_ACTIVE);
2673 * Check again in case we missed a preemption opportunity
2674 * between schedule and now.
2677 } while (need_resched());
2679 exception_exit(prev_state);
2682 #endif /* CONFIG_PREEMPT */
2684 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2687 return try_to_wake_up(curr->private, mode, wake_flags);
2689 EXPORT_SYMBOL(default_wake_function);
2692 * complete: - signals a single thread waiting on this completion
2693 * @x: holds the state of this particular completion
2695 * This will wake up a single thread waiting on this completion. Threads will be
2696 * awakened in the same order in which they were queued.
2698 * See also complete_all(), wait_for_completion() and related routines.
2700 * It may be assumed that this function implies a write memory barrier before
2701 * changing the task state if and only if any tasks are woken up.
2703 void complete(struct completion *x)
2705 unsigned long flags;
2707 spin_lock_irqsave(&x->wait.lock, flags);
2709 __wake_up_locked(&x->wait, TASK_NORMAL, 1);
2710 spin_unlock_irqrestore(&x->wait.lock, flags);
2712 EXPORT_SYMBOL(complete);
2715 * complete_all: - signals all threads waiting on this completion
2716 * @x: holds the state of this particular completion
2718 * This will wake up all threads waiting on this particular completion event.
2720 * It may be assumed that this function implies a write memory barrier before
2721 * changing the task state if and only if any tasks are woken up.
2723 void complete_all(struct completion *x)
2725 unsigned long flags;
2727 spin_lock_irqsave(&x->wait.lock, flags);
2728 x->done += UINT_MAX/2;
2729 __wake_up_locked(&x->wait, TASK_NORMAL, 0);
2730 spin_unlock_irqrestore(&x->wait.lock, flags);
2732 EXPORT_SYMBOL(complete_all);
2734 static inline long __sched
2735 do_wait_for_common(struct completion *x,
2736 long (*action)(long), long timeout, int state)
2739 DECLARE_WAITQUEUE(wait, current);
2741 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2743 if (signal_pending_state(state, current)) {
2744 timeout = -ERESTARTSYS;
2747 __set_current_state(state);
2748 spin_unlock_irq(&x->wait.lock);
2749 timeout = action(timeout);
2750 spin_lock_irq(&x->wait.lock);
2751 } while (!x->done && timeout);
2752 __remove_wait_queue(&x->wait, &wait);
2757 return timeout ?: 1;
2760 static inline long __sched
2761 __wait_for_common(struct completion *x,
2762 long (*action)(long), long timeout, int state)
2766 spin_lock_irq(&x->wait.lock);
2767 timeout = do_wait_for_common(x, action, timeout, state);
2768 spin_unlock_irq(&x->wait.lock);
2773 wait_for_common(struct completion *x, long timeout, int state)
2775 return __wait_for_common(x, schedule_timeout, timeout, state);
2779 wait_for_common_io(struct completion *x, long timeout, int state)
2781 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2785 * wait_for_completion: - waits for completion of a task
2786 * @x: holds the state of this particular completion
2788 * This waits to be signaled for completion of a specific task. It is NOT
2789 * interruptible and there is no timeout.
2791 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2792 * and interrupt capability. Also see complete().
2794 void __sched wait_for_completion(struct completion *x)
2796 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2798 EXPORT_SYMBOL(wait_for_completion);
2801 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2802 * @x: holds the state of this particular completion
2803 * @timeout: timeout value in jiffies
2805 * This waits for either a completion of a specific task to be signaled or for a
2806 * specified timeout to expire. The timeout is in jiffies. It is not
2809 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2810 * till timeout) if completed.
2812 unsigned long __sched
2813 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2815 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2817 EXPORT_SYMBOL(wait_for_completion_timeout);
2820 * wait_for_completion_io: - waits for completion of a task
2821 * @x: holds the state of this particular completion
2823 * This waits to be signaled for completion of a specific task. It is NOT
2824 * interruptible and there is no timeout. The caller is accounted as waiting
2827 void __sched wait_for_completion_io(struct completion *x)
2829 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2831 EXPORT_SYMBOL(wait_for_completion_io);
2834 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2835 * @x: holds the state of this particular completion
2836 * @timeout: timeout value in jiffies
2838 * This waits for either a completion of a specific task to be signaled or for a
2839 * specified timeout to expire. The timeout is in jiffies. It is not
2840 * interruptible. The caller is accounted as waiting for IO.
2842 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2843 * till timeout) if completed.
2845 unsigned long __sched
2846 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2848 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2850 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2853 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2854 * @x: holds the state of this particular completion
2856 * This waits for completion of a specific task to be signaled. It is
2859 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2861 int __sched wait_for_completion_interruptible(struct completion *x)
2863 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2864 if (t == -ERESTARTSYS)
2868 EXPORT_SYMBOL(wait_for_completion_interruptible);
2871 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2872 * @x: holds the state of this particular completion
2873 * @timeout: timeout value in jiffies
2875 * This waits for either a completion of a specific task to be signaled or for a
2876 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2878 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2879 * or number of jiffies left till timeout) if completed.
2882 wait_for_completion_interruptible_timeout(struct completion *x,
2883 unsigned long timeout)
2885 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2887 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2890 * wait_for_completion_killable: - waits for completion of a task (killable)
2891 * @x: holds the state of this particular completion
2893 * This waits to be signaled for completion of a specific task. It can be
2894 * interrupted by a kill signal.
2896 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2898 int __sched wait_for_completion_killable(struct completion *x)
2900 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2901 if (t == -ERESTARTSYS)
2905 EXPORT_SYMBOL(wait_for_completion_killable);
2908 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2909 * @x: holds the state of this particular completion
2910 * @timeout: timeout value in jiffies
2912 * This waits for either a completion of a specific task to be
2913 * signaled or for a specified timeout to expire. It can be
2914 * interrupted by a kill signal. The timeout is in jiffies.
2916 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2917 * or number of jiffies left till timeout) if completed.
2920 wait_for_completion_killable_timeout(struct completion *x,
2921 unsigned long timeout)
2923 return wait_for_common(x, timeout, TASK_KILLABLE);
2925 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2928 * try_wait_for_completion - try to decrement a completion without blocking
2929 * @x: completion structure
2931 * Return: 0 if a decrement cannot be done without blocking
2932 * 1 if a decrement succeeded.
2934 * If a completion is being used as a counting completion,
2935 * attempt to decrement the counter without blocking. This
2936 * enables us to avoid waiting if the resource the completion
2937 * is protecting is not available.
2939 bool try_wait_for_completion(struct completion *x)
2941 unsigned long flags;
2944 spin_lock_irqsave(&x->wait.lock, flags);
2949 spin_unlock_irqrestore(&x->wait.lock, flags);
2952 EXPORT_SYMBOL(try_wait_for_completion);
2955 * completion_done - Test to see if a completion has any waiters
2956 * @x: completion structure
2958 * Return: 0 if there are waiters (wait_for_completion() in progress)
2959 * 1 if there are no waiters.
2962 bool completion_done(struct completion *x)
2964 unsigned long flags;
2967 spin_lock_irqsave(&x->wait.lock, flags);
2970 spin_unlock_irqrestore(&x->wait.lock, flags);
2973 EXPORT_SYMBOL(completion_done);
2976 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2978 unsigned long flags;
2981 init_waitqueue_entry(&wait, current);
2983 __set_current_state(state);
2985 spin_lock_irqsave(&q->lock, flags);
2986 __add_wait_queue(q, &wait);
2987 spin_unlock(&q->lock);
2988 timeout = schedule_timeout(timeout);
2989 spin_lock_irq(&q->lock);
2990 __remove_wait_queue(q, &wait);
2991 spin_unlock_irqrestore(&q->lock, flags);
2996 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2998 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3000 EXPORT_SYMBOL(interruptible_sleep_on);
3003 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3005 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3007 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3009 void __sched sleep_on(wait_queue_head_t *q)
3011 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3013 EXPORT_SYMBOL(sleep_on);
3015 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3017 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3019 EXPORT_SYMBOL(sleep_on_timeout);
3021 #ifdef CONFIG_RT_MUTEXES
3024 * rt_mutex_setprio - set the current priority of a task
3026 * @prio: prio value (kernel-internal form)
3028 * This function changes the 'effective' priority of a task. It does
3029 * not touch ->normal_prio like __setscheduler().
3031 * Used by the rt_mutex code to implement priority inheritance logic.
3033 void rt_mutex_setprio(struct task_struct *p, int prio)
3035 int oldprio, on_rq, running;
3037 const struct sched_class *prev_class;
3039 BUG_ON(prio < 0 || prio > MAX_PRIO);
3041 rq = __task_rq_lock(p);
3044 * Idle task boosting is a nono in general. There is one
3045 * exception, when PREEMPT_RT and NOHZ is active:
3047 * The idle task calls get_next_timer_interrupt() and holds
3048 * the timer wheel base->lock on the CPU and another CPU wants
3049 * to access the timer (probably to cancel it). We can safely
3050 * ignore the boosting request, as the idle CPU runs this code
3051 * with interrupts disabled and will complete the lock
3052 * protected section without being interrupted. So there is no
3053 * real need to boost.
3055 if (unlikely(p == rq->idle)) {
3056 WARN_ON(p != rq->curr);
3057 WARN_ON(p->pi_blocked_on);
3061 trace_sched_pi_setprio(p, prio);
3063 prev_class = p->sched_class;
3065 running = task_current(rq, p);
3067 dequeue_task(rq, p, 0);
3069 p->sched_class->put_prev_task(rq, p);
3072 p->sched_class = &rt_sched_class;
3074 p->sched_class = &fair_sched_class;
3079 p->sched_class->set_curr_task(rq);
3081 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3083 check_class_changed(rq, p, prev_class, oldprio);
3085 __task_rq_unlock(rq);
3088 void set_user_nice(struct task_struct *p, long nice)
3090 int old_prio, delta, on_rq;
3091 unsigned long flags;
3094 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3097 * We have to be careful, if called from sys_setpriority(),
3098 * the task might be in the middle of scheduling on another CPU.
3100 rq = task_rq_lock(p, &flags);
3102 * The RT priorities are set via sched_setscheduler(), but we still
3103 * allow the 'normal' nice value to be set - but as expected
3104 * it wont have any effect on scheduling until the task is
3105 * SCHED_FIFO/SCHED_RR:
3107 if (task_has_rt_policy(p)) {
3108 p->static_prio = NICE_TO_PRIO(nice);
3113 dequeue_task(rq, p, 0);
3115 p->static_prio = NICE_TO_PRIO(nice);
3118 p->prio = effective_prio(p);
3119 delta = p->prio - old_prio;
3122 enqueue_task(rq, p, 0);
3124 * If the task increased its priority or is running and
3125 * lowered its priority, then reschedule its CPU:
3127 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3128 resched_task(rq->curr);
3131 task_rq_unlock(rq, p, &flags);
3133 EXPORT_SYMBOL(set_user_nice);
3136 * can_nice - check if a task can reduce its nice value
3140 int can_nice(const struct task_struct *p, const int nice)
3142 /* convert nice value [19,-20] to rlimit style value [1,40] */
3143 int nice_rlim = 20 - nice;
3145 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3146 capable(CAP_SYS_NICE));
3149 #ifdef __ARCH_WANT_SYS_NICE
3152 * sys_nice - change the priority of the current process.
3153 * @increment: priority increment
3155 * sys_setpriority is a more generic, but much slower function that
3156 * does similar things.
3158 SYSCALL_DEFINE1(nice, int, increment)
3163 * Setpriority might change our priority at the same moment.
3164 * We don't have to worry. Conceptually one call occurs first
3165 * and we have a single winner.
3167 if (increment < -40)
3172 nice = TASK_NICE(current) + increment;
3178 if (increment < 0 && !can_nice(current, nice))
3181 retval = security_task_setnice(current, nice);
3185 set_user_nice(current, nice);
3192 * task_prio - return the priority value of a given task.
3193 * @p: the task in question.
3195 * Return: The priority value as seen by users in /proc.
3196 * RT tasks are offset by -200. Normal tasks are centered
3197 * around 0, value goes from -16 to +15.
3199 int task_prio(const struct task_struct *p)
3201 return p->prio - MAX_RT_PRIO;
3205 * task_nice - return the nice value of a given task.
3206 * @p: the task in question.
3208 * Return: The nice value [ -20 ... 0 ... 19 ].
3210 int task_nice(const struct task_struct *p)
3212 return TASK_NICE(p);
3214 EXPORT_SYMBOL(task_nice);
3217 * idle_cpu - is a given cpu idle currently?
3218 * @cpu: the processor in question.
3220 * Return: 1 if the CPU is currently idle. 0 otherwise.
3222 int idle_cpu(int cpu)
3224 struct rq *rq = cpu_rq(cpu);
3226 if (rq->curr != rq->idle)
3233 if (!llist_empty(&rq->wake_list))
3241 * idle_task - return the idle task for a given cpu.
3242 * @cpu: the processor in question.
3244 * Return: The idle task for the cpu @cpu.
3246 struct task_struct *idle_task(int cpu)
3248 return cpu_rq(cpu)->idle;
3252 * find_process_by_pid - find a process with a matching PID value.
3253 * @pid: the pid in question.
3255 * The task of @pid, if found. %NULL otherwise.
3257 static struct task_struct *find_process_by_pid(pid_t pid)
3259 return pid ? find_task_by_vpid(pid) : current;
3262 /* Actually do priority change: must hold rq lock. */
3264 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3267 p->rt_priority = prio;
3268 p->normal_prio = normal_prio(p);
3269 /* we are holding p->pi_lock already */
3270 p->prio = rt_mutex_getprio(p);
3271 if (rt_prio(p->prio))
3272 p->sched_class = &rt_sched_class;
3274 p->sched_class = &fair_sched_class;
3279 * check the target process has a UID that matches the current process's
3281 static bool check_same_owner(struct task_struct *p)
3283 const struct cred *cred = current_cred(), *pcred;
3287 pcred = __task_cred(p);
3288 match = (uid_eq(cred->euid, pcred->euid) ||
3289 uid_eq(cred->euid, pcred->uid));
3294 static int __sched_setscheduler(struct task_struct *p, int policy,
3295 const struct sched_param *param, bool user)
3297 int retval, oldprio, oldpolicy = -1, on_rq, running;
3298 unsigned long flags;
3299 const struct sched_class *prev_class;
3303 /* may grab non-irq protected spin_locks */
3304 BUG_ON(in_interrupt());
3306 /* double check policy once rq lock held */
3308 reset_on_fork = p->sched_reset_on_fork;
3309 policy = oldpolicy = p->policy;
3311 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3312 policy &= ~SCHED_RESET_ON_FORK;
3314 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3315 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3316 policy != SCHED_IDLE)
3321 * Valid priorities for SCHED_FIFO and SCHED_RR are
3322 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3323 * SCHED_BATCH and SCHED_IDLE is 0.
3325 if (param->sched_priority < 0 ||
3326 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3327 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3329 if (rt_policy(policy) != (param->sched_priority != 0))
3333 * Allow unprivileged RT tasks to decrease priority:
3335 if (user && !capable(CAP_SYS_NICE)) {
3336 if (rt_policy(policy)) {
3337 unsigned long rlim_rtprio =
3338 task_rlimit(p, RLIMIT_RTPRIO);
3340 /* can't set/change the rt policy */
3341 if (policy != p->policy && !rlim_rtprio)
3344 /* can't increase priority */
3345 if (param->sched_priority > p->rt_priority &&
3346 param->sched_priority > rlim_rtprio)
3351 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3352 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3354 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3355 if (!can_nice(p, TASK_NICE(p)))
3359 /* can't change other user's priorities */
3360 if (!check_same_owner(p))
3363 /* Normal users shall not reset the sched_reset_on_fork flag */
3364 if (p->sched_reset_on_fork && !reset_on_fork)
3369 retval = security_task_setscheduler(p);
3375 * make sure no PI-waiters arrive (or leave) while we are
3376 * changing the priority of the task:
3378 * To be able to change p->policy safely, the appropriate
3379 * runqueue lock must be held.
3381 rq = task_rq_lock(p, &flags);
3384 * Changing the policy of the stop threads its a very bad idea
3386 if (p == rq->stop) {
3387 task_rq_unlock(rq, p, &flags);
3392 * If not changing anything there's no need to proceed further:
3394 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3395 param->sched_priority == p->rt_priority))) {
3396 task_rq_unlock(rq, p, &flags);
3400 #ifdef CONFIG_RT_GROUP_SCHED
3403 * Do not allow realtime tasks into groups that have no runtime
3406 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3407 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3408 !task_group_is_autogroup(task_group(p))) {
3409 task_rq_unlock(rq, p, &flags);
3415 /* recheck policy now with rq lock held */
3416 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3417 policy = oldpolicy = -1;
3418 task_rq_unlock(rq, p, &flags);
3422 running = task_current(rq, p);
3424 dequeue_task(rq, p, 0);
3426 p->sched_class->put_prev_task(rq, p);
3428 p->sched_reset_on_fork = reset_on_fork;
3431 prev_class = p->sched_class;
3432 __setscheduler(rq, p, policy, param->sched_priority);
3435 p->sched_class->set_curr_task(rq);
3437 enqueue_task(rq, p, 0);
3439 check_class_changed(rq, p, prev_class, oldprio);
3440 task_rq_unlock(rq, p, &flags);
3442 rt_mutex_adjust_pi(p);
3448 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3449 * @p: the task in question.
3450 * @policy: new policy.
3451 * @param: structure containing the new RT priority.
3453 * Return: 0 on success. An error code otherwise.
3455 * NOTE that the task may be already dead.
3457 int sched_setscheduler(struct task_struct *p, int policy,
3458 const struct sched_param *param)
3460 return __sched_setscheduler(p, policy, param, true);
3462 EXPORT_SYMBOL_GPL(sched_setscheduler);
3465 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3466 * @p: the task in question.
3467 * @policy: new policy.
3468 * @param: structure containing the new RT priority.
3470 * Just like sched_setscheduler, only don't bother checking if the
3471 * current context has permission. For example, this is needed in
3472 * stop_machine(): we create temporary high priority worker threads,
3473 * but our caller might not have that capability.
3475 * Return: 0 on success. An error code otherwise.
3477 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3478 const struct sched_param *param)
3480 return __sched_setscheduler(p, policy, param, false);
3484 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3486 struct sched_param lparam;
3487 struct task_struct *p;
3490 if (!param || pid < 0)
3492 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3497 p = find_process_by_pid(pid);
3499 retval = sched_setscheduler(p, policy, &lparam);
3506 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3507 * @pid: the pid in question.
3508 * @policy: new policy.
3509 * @param: structure containing the new RT priority.
3511 * Return: 0 on success. An error code otherwise.
3513 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3514 struct sched_param __user *, param)
3516 /* negative values for policy are not valid */
3520 return do_sched_setscheduler(pid, policy, param);
3524 * sys_sched_setparam - set/change the RT priority of a thread
3525 * @pid: the pid in question.
3526 * @param: structure containing the new RT priority.
3528 * Return: 0 on success. An error code otherwise.
3530 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3532 return do_sched_setscheduler(pid, -1, param);
3536 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3537 * @pid: the pid in question.
3539 * Return: On success, the policy of the thread. Otherwise, a negative error
3542 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3544 struct task_struct *p;
3552 p = find_process_by_pid(pid);
3554 retval = security_task_getscheduler(p);
3557 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3564 * sys_sched_getparam - get the RT priority of a thread
3565 * @pid: the pid in question.
3566 * @param: structure containing the RT priority.
3568 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3571 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3573 struct sched_param lp;
3574 struct task_struct *p;
3577 if (!param || pid < 0)
3581 p = find_process_by_pid(pid);
3586 retval = security_task_getscheduler(p);
3590 lp.sched_priority = p->rt_priority;
3594 * This one might sleep, we cannot do it with a spinlock held ...
3596 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3605 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3607 cpumask_var_t cpus_allowed, new_mask;
3608 struct task_struct *p;
3613 p = find_process_by_pid(pid);
3619 /* Prevent p going away */
3623 if (p->flags & PF_NO_SETAFFINITY) {
3627 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3631 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3633 goto out_free_cpus_allowed;
3636 if (!check_same_owner(p)) {
3638 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3645 retval = security_task_setscheduler(p);
3649 cpuset_cpus_allowed(p, cpus_allowed);
3650 cpumask_and(new_mask, in_mask, cpus_allowed);
3652 retval = set_cpus_allowed_ptr(p, new_mask);
3655 cpuset_cpus_allowed(p, cpus_allowed);
3656 if (!cpumask_subset(new_mask, cpus_allowed)) {
3658 * We must have raced with a concurrent cpuset
3659 * update. Just reset the cpus_allowed to the
3660 * cpuset's cpus_allowed
3662 cpumask_copy(new_mask, cpus_allowed);
3667 free_cpumask_var(new_mask);
3668 out_free_cpus_allowed:
3669 free_cpumask_var(cpus_allowed);
3675 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3676 struct cpumask *new_mask)
3678 if (len < cpumask_size())
3679 cpumask_clear(new_mask);
3680 else if (len > cpumask_size())
3681 len = cpumask_size();
3683 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3687 * sys_sched_setaffinity - set the cpu affinity of a process
3688 * @pid: pid of the process
3689 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3690 * @user_mask_ptr: user-space pointer to the new cpu mask
3692 * Return: 0 on success. An error code otherwise.
3694 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3695 unsigned long __user *, user_mask_ptr)
3697 cpumask_var_t new_mask;
3700 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3703 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3705 retval = sched_setaffinity(pid, new_mask);
3706 free_cpumask_var(new_mask);
3710 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3712 struct task_struct *p;
3713 unsigned long flags;
3719 p = find_process_by_pid(pid);
3723 retval = security_task_getscheduler(p);
3727 raw_spin_lock_irqsave(&p->pi_lock, flags);
3728 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3729 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3738 * sys_sched_getaffinity - get the cpu affinity of a process
3739 * @pid: pid of the process
3740 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3741 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3743 * Return: 0 on success. An error code otherwise.
3745 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3746 unsigned long __user *, user_mask_ptr)
3751 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3753 if (len & (sizeof(unsigned long)-1))
3756 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3759 ret = sched_getaffinity(pid, mask);
3761 size_t retlen = min_t(size_t, len, cpumask_size());
3763 if (copy_to_user(user_mask_ptr, mask, retlen))
3768 free_cpumask_var(mask);
3774 * sys_sched_yield - yield the current processor to other threads.
3776 * This function yields the current CPU to other tasks. If there are no
3777 * other threads running on this CPU then this function will return.
3781 SYSCALL_DEFINE0(sched_yield)
3783 struct rq *rq = this_rq_lock();
3785 schedstat_inc(rq, yld_count);
3786 current->sched_class->yield_task(rq);
3789 * Since we are going to call schedule() anyway, there's
3790 * no need to preempt or enable interrupts:
3792 __release(rq->lock);
3793 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3794 do_raw_spin_unlock(&rq->lock);
3795 sched_preempt_enable_no_resched();
3802 static void __cond_resched(void)
3804 __preempt_count_add(PREEMPT_ACTIVE);
3806 __preempt_count_sub(PREEMPT_ACTIVE);
3809 int __sched _cond_resched(void)
3811 if (should_resched()) {
3817 EXPORT_SYMBOL(_cond_resched);
3820 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3821 * call schedule, and on return reacquire the lock.
3823 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3824 * operations here to prevent schedule() from being called twice (once via
3825 * spin_unlock(), once by hand).
3827 int __cond_resched_lock(spinlock_t *lock)
3829 int resched = should_resched();
3832 lockdep_assert_held(lock);
3834 if (spin_needbreak(lock) || resched) {
3845 EXPORT_SYMBOL(__cond_resched_lock);
3847 int __sched __cond_resched_softirq(void)
3849 BUG_ON(!in_softirq());
3851 if (should_resched()) {
3859 EXPORT_SYMBOL(__cond_resched_softirq);
3862 * yield - yield the current processor to other threads.
3864 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3866 * The scheduler is at all times free to pick the calling task as the most
3867 * eligible task to run, if removing the yield() call from your code breaks
3868 * it, its already broken.
3870 * Typical broken usage is:
3875 * where one assumes that yield() will let 'the other' process run that will
3876 * make event true. If the current task is a SCHED_FIFO task that will never
3877 * happen. Never use yield() as a progress guarantee!!
3879 * If you want to use yield() to wait for something, use wait_event().
3880 * If you want to use yield() to be 'nice' for others, use cond_resched().
3881 * If you still want to use yield(), do not!
3883 void __sched yield(void)
3885 set_current_state(TASK_RUNNING);
3888 EXPORT_SYMBOL(yield);
3891 * yield_to - yield the current processor to another thread in
3892 * your thread group, or accelerate that thread toward the
3893 * processor it's on.
3895 * @preempt: whether task preemption is allowed or not
3897 * It's the caller's job to ensure that the target task struct
3898 * can't go away on us before we can do any checks.
3901 * true (>0) if we indeed boosted the target task.
3902 * false (0) if we failed to boost the target.
3903 * -ESRCH if there's no task to yield to.
3905 bool __sched yield_to(struct task_struct *p, bool preempt)
3907 struct task_struct *curr = current;
3908 struct rq *rq, *p_rq;
3909 unsigned long flags;
3912 local_irq_save(flags);
3918 * If we're the only runnable task on the rq and target rq also
3919 * has only one task, there's absolutely no point in yielding.
3921 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3926 double_rq_lock(rq, p_rq);
3927 while (task_rq(p) != p_rq) {
3928 double_rq_unlock(rq, p_rq);
3932 if (!curr->sched_class->yield_to_task)
3935 if (curr->sched_class != p->sched_class)
3938 if (task_running(p_rq, p) || p->state)
3941 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3943 schedstat_inc(rq, yld_count);
3945 * Make p's CPU reschedule; pick_next_entity takes care of
3948 if (preempt && rq != p_rq)
3949 resched_task(p_rq->curr);
3953 double_rq_unlock(rq, p_rq);
3955 local_irq_restore(flags);
3962 EXPORT_SYMBOL_GPL(yield_to);
3965 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3966 * that process accounting knows that this is a task in IO wait state.
3968 void __sched io_schedule(void)
3970 struct rq *rq = raw_rq();
3972 delayacct_blkio_start();
3973 atomic_inc(&rq->nr_iowait);
3974 blk_flush_plug(current);
3975 current->in_iowait = 1;
3977 current->in_iowait = 0;
3978 atomic_dec(&rq->nr_iowait);
3979 delayacct_blkio_end();
3981 EXPORT_SYMBOL(io_schedule);
3983 long __sched io_schedule_timeout(long timeout)
3985 struct rq *rq = raw_rq();
3988 delayacct_blkio_start();
3989 atomic_inc(&rq->nr_iowait);
3990 blk_flush_plug(current);
3991 current->in_iowait = 1;
3992 ret = schedule_timeout(timeout);
3993 current->in_iowait = 0;
3994 atomic_dec(&rq->nr_iowait);
3995 delayacct_blkio_end();
4000 * sys_sched_get_priority_max - return maximum RT priority.
4001 * @policy: scheduling class.
4003 * Return: On success, this syscall returns the maximum
4004 * rt_priority that can be used by a given scheduling class.
4005 * On failure, a negative error code is returned.
4007 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4014 ret = MAX_USER_RT_PRIO-1;
4026 * sys_sched_get_priority_min - return minimum RT priority.
4027 * @policy: scheduling class.
4029 * Return: On success, this syscall returns the minimum
4030 * rt_priority that can be used by a given scheduling class.
4031 * On failure, a negative error code is returned.
4033 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4051 * sys_sched_rr_get_interval - return the default timeslice of a process.
4052 * @pid: pid of the process.
4053 * @interval: userspace pointer to the timeslice value.
4055 * this syscall writes the default timeslice value of a given process
4056 * into the user-space timespec buffer. A value of '0' means infinity.
4058 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4061 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4062 struct timespec __user *, interval)
4064 struct task_struct *p;
4065 unsigned int time_slice;
4066 unsigned long flags;
4076 p = find_process_by_pid(pid);
4080 retval = security_task_getscheduler(p);
4084 rq = task_rq_lock(p, &flags);
4085 time_slice = p->sched_class->get_rr_interval(rq, p);
4086 task_rq_unlock(rq, p, &flags);
4089 jiffies_to_timespec(time_slice, &t);
4090 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4098 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4100 void sched_show_task(struct task_struct *p)
4102 unsigned long free = 0;
4106 state = p->state ? __ffs(p->state) + 1 : 0;
4107 printk(KERN_INFO "%-15.15s %c", p->comm,
4108 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4109 #if BITS_PER_LONG == 32
4110 if (state == TASK_RUNNING)
4111 printk(KERN_CONT " running ");
4113 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4115 if (state == TASK_RUNNING)
4116 printk(KERN_CONT " running task ");
4118 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4120 #ifdef CONFIG_DEBUG_STACK_USAGE
4121 free = stack_not_used(p);
4124 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4126 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4127 task_pid_nr(p), ppid,
4128 (unsigned long)task_thread_info(p)->flags);
4130 print_worker_info(KERN_INFO, p);
4131 show_stack(p, NULL);
4134 void show_state_filter(unsigned long state_filter)
4136 struct task_struct *g, *p;
4138 #if BITS_PER_LONG == 32
4140 " task PC stack pid father\n");
4143 " task PC stack pid father\n");
4146 do_each_thread(g, p) {
4148 * reset the NMI-timeout, listing all files on a slow
4149 * console might take a lot of time:
4151 touch_nmi_watchdog();
4152 if (!state_filter || (p->state & state_filter))
4154 } while_each_thread(g, p);
4156 touch_all_softlockup_watchdogs();
4158 #ifdef CONFIG_SCHED_DEBUG
4159 sysrq_sched_debug_show();
4163 * Only show locks if all tasks are dumped:
4166 debug_show_all_locks();
4169 void init_idle_bootup_task(struct task_struct *idle)
4171 idle->sched_class = &idle_sched_class;
4175 * init_idle - set up an idle thread for a given CPU
4176 * @idle: task in question
4177 * @cpu: cpu the idle task belongs to
4179 * NOTE: this function does not set the idle thread's NEED_RESCHED
4180 * flag, to make booting more robust.
4182 void init_idle(struct task_struct *idle, int cpu)
4184 struct rq *rq = cpu_rq(cpu);
4185 unsigned long flags;
4187 raw_spin_lock_irqsave(&rq->lock, flags);
4189 __sched_fork(0, idle);
4190 idle->state = TASK_RUNNING;
4191 idle->se.exec_start = sched_clock();
4193 do_set_cpus_allowed(idle, cpumask_of(cpu));
4195 * We're having a chicken and egg problem, even though we are
4196 * holding rq->lock, the cpu isn't yet set to this cpu so the
4197 * lockdep check in task_group() will fail.
4199 * Similar case to sched_fork(). / Alternatively we could
4200 * use task_rq_lock() here and obtain the other rq->lock.
4205 __set_task_cpu(idle, cpu);
4208 rq->curr = rq->idle = idle;
4209 #if defined(CONFIG_SMP)
4212 raw_spin_unlock_irqrestore(&rq->lock, flags);
4214 /* Set the preempt count _outside_ the spinlocks! */
4215 init_idle_preempt_count(idle, cpu);
4218 * The idle tasks have their own, simple scheduling class:
4220 idle->sched_class = &idle_sched_class;
4221 ftrace_graph_init_idle_task(idle, cpu);
4222 vtime_init_idle(idle, cpu);
4223 #if defined(CONFIG_SMP)
4224 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4229 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4231 if (p->sched_class && p->sched_class->set_cpus_allowed)
4232 p->sched_class->set_cpus_allowed(p, new_mask);
4234 cpumask_copy(&p->cpus_allowed, new_mask);
4235 p->nr_cpus_allowed = cpumask_weight(new_mask);
4239 * This is how migration works:
4241 * 1) we invoke migration_cpu_stop() on the target CPU using
4243 * 2) stopper starts to run (implicitly forcing the migrated thread
4245 * 3) it checks whether the migrated task is still in the wrong runqueue.
4246 * 4) if it's in the wrong runqueue then the migration thread removes
4247 * it and puts it into the right queue.
4248 * 5) stopper completes and stop_one_cpu() returns and the migration
4253 * Change a given task's CPU affinity. Migrate the thread to a
4254 * proper CPU and schedule it away if the CPU it's executing on
4255 * is removed from the allowed bitmask.
4257 * NOTE: the caller must have a valid reference to the task, the
4258 * task must not exit() & deallocate itself prematurely. The
4259 * call is not atomic; no spinlocks may be held.
4261 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4263 unsigned long flags;
4265 unsigned int dest_cpu;
4268 rq = task_rq_lock(p, &flags);
4270 if (cpumask_equal(&p->cpus_allowed, new_mask))
4273 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4278 do_set_cpus_allowed(p, new_mask);
4280 /* Can the task run on the task's current CPU? If so, we're done */
4281 if (cpumask_test_cpu(task_cpu(p), new_mask))
4284 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4286 struct migration_arg arg = { p, dest_cpu };
4287 /* Need help from migration thread: drop lock and wait. */
4288 task_rq_unlock(rq, p, &flags);
4289 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4290 tlb_migrate_finish(p->mm);
4294 task_rq_unlock(rq, p, &flags);
4298 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4301 * Move (not current) task off this cpu, onto dest cpu. We're doing
4302 * this because either it can't run here any more (set_cpus_allowed()
4303 * away from this CPU, or CPU going down), or because we're
4304 * attempting to rebalance this task on exec (sched_exec).
4306 * So we race with normal scheduler movements, but that's OK, as long
4307 * as the task is no longer on this CPU.
4309 * Returns non-zero if task was successfully migrated.
4311 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4313 struct rq *rq_dest, *rq_src;
4316 if (unlikely(!cpu_active(dest_cpu)))
4319 rq_src = cpu_rq(src_cpu);
4320 rq_dest = cpu_rq(dest_cpu);
4322 raw_spin_lock(&p->pi_lock);
4323 double_rq_lock(rq_src, rq_dest);
4324 /* Already moved. */
4325 if (task_cpu(p) != src_cpu)
4327 /* Affinity changed (again). */
4328 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4332 * If we're not on a rq, the next wake-up will ensure we're
4336 dequeue_task(rq_src, p, 0);
4337 set_task_cpu(p, dest_cpu);
4338 enqueue_task(rq_dest, p, 0);
4339 check_preempt_curr(rq_dest, p, 0);
4344 double_rq_unlock(rq_src, rq_dest);
4345 raw_spin_unlock(&p->pi_lock);
4349 #ifdef CONFIG_NUMA_BALANCING
4350 /* Migrate current task p to target_cpu */
4351 int migrate_task_to(struct task_struct *p, int target_cpu)
4353 struct migration_arg arg = { p, target_cpu };
4354 int curr_cpu = task_cpu(p);
4356 if (curr_cpu == target_cpu)
4359 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4362 /* TODO: This is not properly updating schedstats */
4364 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4368 * Requeue a task on a given node and accurately track the number of NUMA
4369 * tasks on the runqueues
4371 void sched_setnuma(struct task_struct *p, int nid)
4374 unsigned long flags;
4375 bool on_rq, running;
4377 rq = task_rq_lock(p, &flags);
4379 running = task_current(rq, p);
4382 dequeue_task(rq, p, 0);
4384 p->sched_class->put_prev_task(rq, p);
4386 p->numa_preferred_nid = nid;
4389 p->sched_class->set_curr_task(rq);
4391 enqueue_task(rq, p, 0);
4392 task_rq_unlock(rq, p, &flags);
4397 * migration_cpu_stop - this will be executed by a highprio stopper thread
4398 * and performs thread migration by bumping thread off CPU then
4399 * 'pushing' onto another runqueue.
4401 static int migration_cpu_stop(void *data)
4403 struct migration_arg *arg = data;
4406 * The original target cpu might have gone down and we might
4407 * be on another cpu but it doesn't matter.
4409 local_irq_disable();
4410 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4415 #ifdef CONFIG_HOTPLUG_CPU
4418 * Ensures that the idle task is using init_mm right before its cpu goes
4421 void idle_task_exit(void)
4423 struct mm_struct *mm = current->active_mm;
4425 BUG_ON(cpu_online(smp_processor_id()));
4428 switch_mm(mm, &init_mm, current);
4433 * Since this CPU is going 'away' for a while, fold any nr_active delta
4434 * we might have. Assumes we're called after migrate_tasks() so that the
4435 * nr_active count is stable.
4437 * Also see the comment "Global load-average calculations".
4439 static void calc_load_migrate(struct rq *rq)
4441 long delta = calc_load_fold_active(rq);
4443 atomic_long_add(delta, &calc_load_tasks);
4447 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4448 * try_to_wake_up()->select_task_rq().
4450 * Called with rq->lock held even though we'er in stop_machine() and
4451 * there's no concurrency possible, we hold the required locks anyway
4452 * because of lock validation efforts.
4454 static void migrate_tasks(unsigned int dead_cpu)
4456 struct rq *rq = cpu_rq(dead_cpu);
4457 struct task_struct *next, *stop = rq->stop;
4461 * Fudge the rq selection such that the below task selection loop
4462 * doesn't get stuck on the currently eligible stop task.
4464 * We're currently inside stop_machine() and the rq is either stuck
4465 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4466 * either way we should never end up calling schedule() until we're
4472 * put_prev_task() and pick_next_task() sched
4473 * class method both need to have an up-to-date
4474 * value of rq->clock[_task]
4476 update_rq_clock(rq);
4480 * There's this thread running, bail when that's the only
4483 if (rq->nr_running == 1)
4486 next = pick_next_task(rq);
4488 next->sched_class->put_prev_task(rq, next);
4490 /* Find suitable destination for @next, with force if needed. */
4491 dest_cpu = select_fallback_rq(dead_cpu, next);
4492 raw_spin_unlock(&rq->lock);
4494 __migrate_task(next, dead_cpu, dest_cpu);
4496 raw_spin_lock(&rq->lock);
4502 #endif /* CONFIG_HOTPLUG_CPU */
4504 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4506 static struct ctl_table sd_ctl_dir[] = {
4508 .procname = "sched_domain",
4514 static struct ctl_table sd_ctl_root[] = {
4516 .procname = "kernel",
4518 .child = sd_ctl_dir,
4523 static struct ctl_table *sd_alloc_ctl_entry(int n)
4525 struct ctl_table *entry =
4526 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4531 static void sd_free_ctl_entry(struct ctl_table **tablep)
4533 struct ctl_table *entry;
4536 * In the intermediate directories, both the child directory and
4537 * procname are dynamically allocated and could fail but the mode
4538 * will always be set. In the lowest directory the names are
4539 * static strings and all have proc handlers.
4541 for (entry = *tablep; entry->mode; entry++) {
4543 sd_free_ctl_entry(&entry->child);
4544 if (entry->proc_handler == NULL)
4545 kfree(entry->procname);
4552 static int min_load_idx = 0;
4553 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4556 set_table_entry(struct ctl_table *entry,
4557 const char *procname, void *data, int maxlen,
4558 umode_t mode, proc_handler *proc_handler,
4561 entry->procname = procname;
4563 entry->maxlen = maxlen;
4565 entry->proc_handler = proc_handler;
4568 entry->extra1 = &min_load_idx;
4569 entry->extra2 = &max_load_idx;
4573 static struct ctl_table *
4574 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4576 struct ctl_table *table = sd_alloc_ctl_entry(13);
4581 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4582 sizeof(long), 0644, proc_doulongvec_minmax, false);
4583 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4584 sizeof(long), 0644, proc_doulongvec_minmax, false);
4585 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4586 sizeof(int), 0644, proc_dointvec_minmax, true);
4587 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4588 sizeof(int), 0644, proc_dointvec_minmax, true);
4589 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4590 sizeof(int), 0644, proc_dointvec_minmax, true);
4591 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4592 sizeof(int), 0644, proc_dointvec_minmax, true);
4593 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4594 sizeof(int), 0644, proc_dointvec_minmax, true);
4595 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4596 sizeof(int), 0644, proc_dointvec_minmax, false);
4597 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4598 sizeof(int), 0644, proc_dointvec_minmax, false);
4599 set_table_entry(&table[9], "cache_nice_tries",
4600 &sd->cache_nice_tries,
4601 sizeof(int), 0644, proc_dointvec_minmax, false);
4602 set_table_entry(&table[10], "flags", &sd->flags,
4603 sizeof(int), 0644, proc_dointvec_minmax, false);
4604 set_table_entry(&table[11], "name", sd->name,
4605 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4606 /* &table[12] is terminator */
4611 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4613 struct ctl_table *entry, *table;
4614 struct sched_domain *sd;
4615 int domain_num = 0, i;
4618 for_each_domain(cpu, sd)
4620 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4625 for_each_domain(cpu, sd) {
4626 snprintf(buf, 32, "domain%d", i);
4627 entry->procname = kstrdup(buf, GFP_KERNEL);
4629 entry->child = sd_alloc_ctl_domain_table(sd);
4636 static struct ctl_table_header *sd_sysctl_header;
4637 static void register_sched_domain_sysctl(void)
4639 int i, cpu_num = num_possible_cpus();
4640 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4643 WARN_ON(sd_ctl_dir[0].child);
4644 sd_ctl_dir[0].child = entry;
4649 for_each_possible_cpu(i) {
4650 snprintf(buf, 32, "cpu%d", i);
4651 entry->procname = kstrdup(buf, GFP_KERNEL);
4653 entry->child = sd_alloc_ctl_cpu_table(i);
4657 WARN_ON(sd_sysctl_header);
4658 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4661 /* may be called multiple times per register */
4662 static void unregister_sched_domain_sysctl(void)
4664 if (sd_sysctl_header)
4665 unregister_sysctl_table(sd_sysctl_header);
4666 sd_sysctl_header = NULL;
4667 if (sd_ctl_dir[0].child)
4668 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4671 static void register_sched_domain_sysctl(void)
4674 static void unregister_sched_domain_sysctl(void)
4679 static void set_rq_online(struct rq *rq)
4682 const struct sched_class *class;
4684 cpumask_set_cpu(rq->cpu, rq->rd->online);
4687 for_each_class(class) {
4688 if (class->rq_online)
4689 class->rq_online(rq);
4694 static void set_rq_offline(struct rq *rq)
4697 const struct sched_class *class;
4699 for_each_class(class) {
4700 if (class->rq_offline)
4701 class->rq_offline(rq);
4704 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4710 * migration_call - callback that gets triggered when a CPU is added.
4711 * Here we can start up the necessary migration thread for the new CPU.
4714 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4716 int cpu = (long)hcpu;
4717 unsigned long flags;
4718 struct rq *rq = cpu_rq(cpu);
4720 switch (action & ~CPU_TASKS_FROZEN) {
4722 case CPU_UP_PREPARE:
4723 rq->calc_load_update = calc_load_update;
4727 /* Update our root-domain */
4728 raw_spin_lock_irqsave(&rq->lock, flags);
4730 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4734 raw_spin_unlock_irqrestore(&rq->lock, flags);
4737 #ifdef CONFIG_HOTPLUG_CPU
4739 sched_ttwu_pending();
4740 /* Update our root-domain */
4741 raw_spin_lock_irqsave(&rq->lock, flags);
4743 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4747 BUG_ON(rq->nr_running != 1); /* the migration thread */
4748 raw_spin_unlock_irqrestore(&rq->lock, flags);
4752 calc_load_migrate(rq);
4757 update_max_interval();
4763 * Register at high priority so that task migration (migrate_all_tasks)
4764 * happens before everything else. This has to be lower priority than
4765 * the notifier in the perf_event subsystem, though.
4767 static struct notifier_block migration_notifier = {
4768 .notifier_call = migration_call,
4769 .priority = CPU_PRI_MIGRATION,
4772 static int sched_cpu_active(struct notifier_block *nfb,
4773 unsigned long action, void *hcpu)
4775 switch (action & ~CPU_TASKS_FROZEN) {
4777 case CPU_DOWN_FAILED:
4778 set_cpu_active((long)hcpu, true);
4785 static int sched_cpu_inactive(struct notifier_block *nfb,
4786 unsigned long action, void *hcpu)
4788 switch (action & ~CPU_TASKS_FROZEN) {
4789 case CPU_DOWN_PREPARE:
4790 set_cpu_active((long)hcpu, false);
4797 static int __init migration_init(void)
4799 void *cpu = (void *)(long)smp_processor_id();
4802 /* Initialize migration for the boot CPU */
4803 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4804 BUG_ON(err == NOTIFY_BAD);
4805 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4806 register_cpu_notifier(&migration_notifier);
4808 /* Register cpu active notifiers */
4809 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4810 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4814 early_initcall(migration_init);
4819 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4821 #ifdef CONFIG_SCHED_DEBUG
4823 static __read_mostly int sched_debug_enabled;
4825 static int __init sched_debug_setup(char *str)
4827 sched_debug_enabled = 1;
4831 early_param("sched_debug", sched_debug_setup);
4833 static inline bool sched_debug(void)
4835 return sched_debug_enabled;
4838 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4839 struct cpumask *groupmask)
4841 struct sched_group *group = sd->groups;
4844 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4845 cpumask_clear(groupmask);
4847 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4849 if (!(sd->flags & SD_LOAD_BALANCE)) {
4850 printk("does not load-balance\n");
4852 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4857 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4859 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4860 printk(KERN_ERR "ERROR: domain->span does not contain "
4863 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4864 printk(KERN_ERR "ERROR: domain->groups does not contain"
4868 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4872 printk(KERN_ERR "ERROR: group is NULL\n");
4877 * Even though we initialize ->power to something semi-sane,
4878 * we leave power_orig unset. This allows us to detect if
4879 * domain iteration is still funny without causing /0 traps.
4881 if (!group->sgp->power_orig) {
4882 printk(KERN_CONT "\n");
4883 printk(KERN_ERR "ERROR: domain->cpu_power not "
4888 if (!cpumask_weight(sched_group_cpus(group))) {
4889 printk(KERN_CONT "\n");
4890 printk(KERN_ERR "ERROR: empty group\n");
4894 if (!(sd->flags & SD_OVERLAP) &&
4895 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4896 printk(KERN_CONT "\n");
4897 printk(KERN_ERR "ERROR: repeated CPUs\n");
4901 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4903 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4905 printk(KERN_CONT " %s", str);
4906 if (group->sgp->power != SCHED_POWER_SCALE) {
4907 printk(KERN_CONT " (cpu_power = %d)",
4911 group = group->next;
4912 } while (group != sd->groups);
4913 printk(KERN_CONT "\n");
4915 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4916 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4919 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4920 printk(KERN_ERR "ERROR: parent span is not a superset "
4921 "of domain->span\n");
4925 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4929 if (!sched_debug_enabled)
4933 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4937 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4940 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4948 #else /* !CONFIG_SCHED_DEBUG */
4949 # define sched_domain_debug(sd, cpu) do { } while (0)
4950 static inline bool sched_debug(void)
4954 #endif /* CONFIG_SCHED_DEBUG */
4956 static int sd_degenerate(struct sched_domain *sd)
4958 if (cpumask_weight(sched_domain_span(sd)) == 1)
4961 /* Following flags need at least 2 groups */
4962 if (sd->flags & (SD_LOAD_BALANCE |
4963 SD_BALANCE_NEWIDLE |
4967 SD_SHARE_PKG_RESOURCES)) {
4968 if (sd->groups != sd->groups->next)
4972 /* Following flags don't use groups */
4973 if (sd->flags & (SD_WAKE_AFFINE))
4980 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4982 unsigned long cflags = sd->flags, pflags = parent->flags;
4984 if (sd_degenerate(parent))
4987 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4990 /* Flags needing groups don't count if only 1 group in parent */
4991 if (parent->groups == parent->groups->next) {
4992 pflags &= ~(SD_LOAD_BALANCE |
4993 SD_BALANCE_NEWIDLE |
4997 SD_SHARE_PKG_RESOURCES |
4999 if (nr_node_ids == 1)
5000 pflags &= ~SD_SERIALIZE;
5002 if (~cflags & pflags)
5008 static void free_rootdomain(struct rcu_head *rcu)
5010 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5012 cpupri_cleanup(&rd->cpupri);
5013 free_cpumask_var(rd->rto_mask);
5014 free_cpumask_var(rd->online);
5015 free_cpumask_var(rd->span);
5019 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5021 struct root_domain *old_rd = NULL;
5022 unsigned long flags;
5024 raw_spin_lock_irqsave(&rq->lock, flags);
5029 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5032 cpumask_clear_cpu(rq->cpu, old_rd->span);
5035 * If we dont want to free the old_rt yet then
5036 * set old_rd to NULL to skip the freeing later
5039 if (!atomic_dec_and_test(&old_rd->refcount))
5043 atomic_inc(&rd->refcount);
5046 cpumask_set_cpu(rq->cpu, rd->span);
5047 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5050 raw_spin_unlock_irqrestore(&rq->lock, flags);
5053 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5056 static int init_rootdomain(struct root_domain *rd)
5058 memset(rd, 0, sizeof(*rd));
5060 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5062 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5064 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5067 if (cpupri_init(&rd->cpupri) != 0)
5072 free_cpumask_var(rd->rto_mask);
5074 free_cpumask_var(rd->online);
5076 free_cpumask_var(rd->span);
5082 * By default the system creates a single root-domain with all cpus as
5083 * members (mimicking the global state we have today).
5085 struct root_domain def_root_domain;
5087 static void init_defrootdomain(void)
5089 init_rootdomain(&def_root_domain);
5091 atomic_set(&def_root_domain.refcount, 1);
5094 static struct root_domain *alloc_rootdomain(void)
5096 struct root_domain *rd;
5098 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5102 if (init_rootdomain(rd) != 0) {
5110 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5112 struct sched_group *tmp, *first;
5121 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5126 } while (sg != first);
5129 static void free_sched_domain(struct rcu_head *rcu)
5131 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5134 * If its an overlapping domain it has private groups, iterate and
5137 if (sd->flags & SD_OVERLAP) {
5138 free_sched_groups(sd->groups, 1);
5139 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5140 kfree(sd->groups->sgp);
5146 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5148 call_rcu(&sd->rcu, free_sched_domain);
5151 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5153 for (; sd; sd = sd->parent)
5154 destroy_sched_domain(sd, cpu);
5158 * Keep a special pointer to the highest sched_domain that has
5159 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5160 * allows us to avoid some pointer chasing select_idle_sibling().
5162 * Also keep a unique ID per domain (we use the first cpu number in
5163 * the cpumask of the domain), this allows us to quickly tell if
5164 * two cpus are in the same cache domain, see cpus_share_cache().
5166 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5167 DEFINE_PER_CPU(int, sd_llc_size);
5168 DEFINE_PER_CPU(int, sd_llc_id);
5169 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5171 static void update_top_cache_domain(int cpu)
5173 struct sched_domain *sd;
5177 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5179 id = cpumask_first(sched_domain_span(sd));
5180 size = cpumask_weight(sched_domain_span(sd));
5183 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5184 per_cpu(sd_llc_size, cpu) = size;
5185 per_cpu(sd_llc_id, cpu) = id;
5187 sd = lowest_flag_domain(cpu, SD_NUMA);
5188 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5192 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5193 * hold the hotplug lock.
5196 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5198 struct rq *rq = cpu_rq(cpu);
5199 struct sched_domain *tmp;
5201 /* Remove the sched domains which do not contribute to scheduling. */
5202 for (tmp = sd; tmp; ) {
5203 struct sched_domain *parent = tmp->parent;
5207 if (sd_parent_degenerate(tmp, parent)) {
5208 tmp->parent = parent->parent;
5210 parent->parent->child = tmp;
5212 * Transfer SD_PREFER_SIBLING down in case of a
5213 * degenerate parent; the spans match for this
5214 * so the property transfers.
5216 if (parent->flags & SD_PREFER_SIBLING)
5217 tmp->flags |= SD_PREFER_SIBLING;
5218 destroy_sched_domain(parent, cpu);
5223 if (sd && sd_degenerate(sd)) {
5226 destroy_sched_domain(tmp, cpu);
5231 sched_domain_debug(sd, cpu);
5233 rq_attach_root(rq, rd);
5235 rcu_assign_pointer(rq->sd, sd);
5236 destroy_sched_domains(tmp, cpu);
5238 update_top_cache_domain(cpu);
5241 /* cpus with isolated domains */
5242 static cpumask_var_t cpu_isolated_map;
5244 /* Setup the mask of cpus configured for isolated domains */
5245 static int __init isolated_cpu_setup(char *str)
5247 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5248 cpulist_parse(str, cpu_isolated_map);
5252 __setup("isolcpus=", isolated_cpu_setup);
5254 static const struct cpumask *cpu_cpu_mask(int cpu)
5256 return cpumask_of_node(cpu_to_node(cpu));
5260 struct sched_domain **__percpu sd;
5261 struct sched_group **__percpu sg;
5262 struct sched_group_power **__percpu sgp;
5266 struct sched_domain ** __percpu sd;
5267 struct root_domain *rd;
5277 struct sched_domain_topology_level;
5279 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5280 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5282 #define SDTL_OVERLAP 0x01
5284 struct sched_domain_topology_level {
5285 sched_domain_init_f init;
5286 sched_domain_mask_f mask;
5289 struct sd_data data;
5293 * Build an iteration mask that can exclude certain CPUs from the upwards
5296 * Asymmetric node setups can result in situations where the domain tree is of
5297 * unequal depth, make sure to skip domains that already cover the entire
5300 * In that case build_sched_domains() will have terminated the iteration early
5301 * and our sibling sd spans will be empty. Domains should always include the
5302 * cpu they're built on, so check that.
5305 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5307 const struct cpumask *span = sched_domain_span(sd);
5308 struct sd_data *sdd = sd->private;
5309 struct sched_domain *sibling;
5312 for_each_cpu(i, span) {
5313 sibling = *per_cpu_ptr(sdd->sd, i);
5314 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5317 cpumask_set_cpu(i, sched_group_mask(sg));
5322 * Return the canonical balance cpu for this group, this is the first cpu
5323 * of this group that's also in the iteration mask.
5325 int group_balance_cpu(struct sched_group *sg)
5327 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5331 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5333 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5334 const struct cpumask *span = sched_domain_span(sd);
5335 struct cpumask *covered = sched_domains_tmpmask;
5336 struct sd_data *sdd = sd->private;
5337 struct sched_domain *child;
5340 cpumask_clear(covered);
5342 for_each_cpu(i, span) {
5343 struct cpumask *sg_span;
5345 if (cpumask_test_cpu(i, covered))
5348 child = *per_cpu_ptr(sdd->sd, i);
5350 /* See the comment near build_group_mask(). */
5351 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5354 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5355 GFP_KERNEL, cpu_to_node(cpu));
5360 sg_span = sched_group_cpus(sg);
5362 child = child->child;
5363 cpumask_copy(sg_span, sched_domain_span(child));
5365 cpumask_set_cpu(i, sg_span);
5367 cpumask_or(covered, covered, sg_span);
5369 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5370 if (atomic_inc_return(&sg->sgp->ref) == 1)
5371 build_group_mask(sd, sg);
5374 * Initialize sgp->power such that even if we mess up the
5375 * domains and no possible iteration will get us here, we won't
5378 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5381 * Make sure the first group of this domain contains the
5382 * canonical balance cpu. Otherwise the sched_domain iteration
5383 * breaks. See update_sg_lb_stats().
5385 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5386 group_balance_cpu(sg) == cpu)
5396 sd->groups = groups;
5401 free_sched_groups(first, 0);
5406 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5408 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5409 struct sched_domain *child = sd->child;
5412 cpu = cpumask_first(sched_domain_span(child));
5415 *sg = *per_cpu_ptr(sdd->sg, cpu);
5416 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5417 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5424 * build_sched_groups will build a circular linked list of the groups
5425 * covered by the given span, and will set each group's ->cpumask correctly,
5426 * and ->cpu_power to 0.
5428 * Assumes the sched_domain tree is fully constructed
5431 build_sched_groups(struct sched_domain *sd, int cpu)
5433 struct sched_group *first = NULL, *last = NULL;
5434 struct sd_data *sdd = sd->private;
5435 const struct cpumask *span = sched_domain_span(sd);
5436 struct cpumask *covered;
5439 get_group(cpu, sdd, &sd->groups);
5440 atomic_inc(&sd->groups->ref);
5442 if (cpu != cpumask_first(span))
5445 lockdep_assert_held(&sched_domains_mutex);
5446 covered = sched_domains_tmpmask;
5448 cpumask_clear(covered);
5450 for_each_cpu(i, span) {
5451 struct sched_group *sg;
5454 if (cpumask_test_cpu(i, covered))
5457 group = get_group(i, sdd, &sg);
5458 cpumask_clear(sched_group_cpus(sg));
5460 cpumask_setall(sched_group_mask(sg));
5462 for_each_cpu(j, span) {
5463 if (get_group(j, sdd, NULL) != group)
5466 cpumask_set_cpu(j, covered);
5467 cpumask_set_cpu(j, sched_group_cpus(sg));
5482 * Initialize sched groups cpu_power.
5484 * cpu_power indicates the capacity of sched group, which is used while
5485 * distributing the load between different sched groups in a sched domain.
5486 * Typically cpu_power for all the groups in a sched domain will be same unless
5487 * there are asymmetries in the topology. If there are asymmetries, group
5488 * having more cpu_power will pickup more load compared to the group having
5491 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5493 struct sched_group *sg = sd->groups;
5498 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5500 } while (sg != sd->groups);
5502 if (cpu != group_balance_cpu(sg))
5505 update_group_power(sd, cpu);
5506 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5509 int __weak arch_sd_sibling_asym_packing(void)
5511 return 0*SD_ASYM_PACKING;
5515 * Initializers for schedule domains
5516 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5519 #ifdef CONFIG_SCHED_DEBUG
5520 # define SD_INIT_NAME(sd, type) sd->name = #type
5522 # define SD_INIT_NAME(sd, type) do { } while (0)
5525 #define SD_INIT_FUNC(type) \
5526 static noinline struct sched_domain * \
5527 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5529 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5530 *sd = SD_##type##_INIT; \
5531 SD_INIT_NAME(sd, type); \
5532 sd->private = &tl->data; \
5537 #ifdef CONFIG_SCHED_SMT
5538 SD_INIT_FUNC(SIBLING)
5540 #ifdef CONFIG_SCHED_MC
5543 #ifdef CONFIG_SCHED_BOOK
5547 static int default_relax_domain_level = -1;
5548 int sched_domain_level_max;
5550 static int __init setup_relax_domain_level(char *str)
5552 if (kstrtoint(str, 0, &default_relax_domain_level))
5553 pr_warn("Unable to set relax_domain_level\n");
5557 __setup("relax_domain_level=", setup_relax_domain_level);
5559 static void set_domain_attribute(struct sched_domain *sd,
5560 struct sched_domain_attr *attr)
5564 if (!attr || attr->relax_domain_level < 0) {
5565 if (default_relax_domain_level < 0)
5568 request = default_relax_domain_level;
5570 request = attr->relax_domain_level;
5571 if (request < sd->level) {
5572 /* turn off idle balance on this domain */
5573 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5575 /* turn on idle balance on this domain */
5576 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5580 static void __sdt_free(const struct cpumask *cpu_map);
5581 static int __sdt_alloc(const struct cpumask *cpu_map);
5583 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5584 const struct cpumask *cpu_map)
5588 if (!atomic_read(&d->rd->refcount))
5589 free_rootdomain(&d->rd->rcu); /* fall through */
5591 free_percpu(d->sd); /* fall through */
5593 __sdt_free(cpu_map); /* fall through */
5599 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5600 const struct cpumask *cpu_map)
5602 memset(d, 0, sizeof(*d));
5604 if (__sdt_alloc(cpu_map))
5605 return sa_sd_storage;
5606 d->sd = alloc_percpu(struct sched_domain *);
5608 return sa_sd_storage;
5609 d->rd = alloc_rootdomain();
5612 return sa_rootdomain;
5616 * NULL the sd_data elements we've used to build the sched_domain and
5617 * sched_group structure so that the subsequent __free_domain_allocs()
5618 * will not free the data we're using.
5620 static void claim_allocations(int cpu, struct sched_domain *sd)
5622 struct sd_data *sdd = sd->private;
5624 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5625 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5627 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5628 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5630 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5631 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5634 #ifdef CONFIG_SCHED_SMT
5635 static const struct cpumask *cpu_smt_mask(int cpu)
5637 return topology_thread_cpumask(cpu);
5642 * Topology list, bottom-up.
5644 static struct sched_domain_topology_level default_topology[] = {
5645 #ifdef CONFIG_SCHED_SMT
5646 { sd_init_SIBLING, cpu_smt_mask, },
5648 #ifdef CONFIG_SCHED_MC
5649 { sd_init_MC, cpu_coregroup_mask, },
5651 #ifdef CONFIG_SCHED_BOOK
5652 { sd_init_BOOK, cpu_book_mask, },
5654 { sd_init_CPU, cpu_cpu_mask, },
5658 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5660 #define for_each_sd_topology(tl) \
5661 for (tl = sched_domain_topology; tl->init; tl++)
5665 static int sched_domains_numa_levels;
5666 static int *sched_domains_numa_distance;
5667 static struct cpumask ***sched_domains_numa_masks;
5668 static int sched_domains_curr_level;
5670 static inline int sd_local_flags(int level)
5672 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5675 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5678 static struct sched_domain *
5679 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5681 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5682 int level = tl->numa_level;
5683 int sd_weight = cpumask_weight(
5684 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5686 *sd = (struct sched_domain){
5687 .min_interval = sd_weight,
5688 .max_interval = 2*sd_weight,
5690 .imbalance_pct = 125,
5691 .cache_nice_tries = 2,
5698 .flags = 1*SD_LOAD_BALANCE
5699 | 1*SD_BALANCE_NEWIDLE
5704 | 0*SD_SHARE_CPUPOWER
5705 | 0*SD_SHARE_PKG_RESOURCES
5707 | 0*SD_PREFER_SIBLING
5709 | sd_local_flags(level)
5711 .last_balance = jiffies,
5712 .balance_interval = sd_weight,
5714 SD_INIT_NAME(sd, NUMA);
5715 sd->private = &tl->data;
5718 * Ugly hack to pass state to sd_numa_mask()...
5720 sched_domains_curr_level = tl->numa_level;
5725 static const struct cpumask *sd_numa_mask(int cpu)
5727 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5730 static void sched_numa_warn(const char *str)
5732 static int done = false;
5740 printk(KERN_WARNING "ERROR: %s\n\n", str);
5742 for (i = 0; i < nr_node_ids; i++) {
5743 printk(KERN_WARNING " ");
5744 for (j = 0; j < nr_node_ids; j++)
5745 printk(KERN_CONT "%02d ", node_distance(i,j));
5746 printk(KERN_CONT "\n");
5748 printk(KERN_WARNING "\n");
5751 static bool find_numa_distance(int distance)
5755 if (distance == node_distance(0, 0))
5758 for (i = 0; i < sched_domains_numa_levels; i++) {
5759 if (sched_domains_numa_distance[i] == distance)
5766 static void sched_init_numa(void)
5768 int next_distance, curr_distance = node_distance(0, 0);
5769 struct sched_domain_topology_level *tl;
5773 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5774 if (!sched_domains_numa_distance)
5778 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5779 * unique distances in the node_distance() table.
5781 * Assumes node_distance(0,j) includes all distances in
5782 * node_distance(i,j) in order to avoid cubic time.
5784 next_distance = curr_distance;
5785 for (i = 0; i < nr_node_ids; i++) {
5786 for (j = 0; j < nr_node_ids; j++) {
5787 for (k = 0; k < nr_node_ids; k++) {
5788 int distance = node_distance(i, k);
5790 if (distance > curr_distance &&
5791 (distance < next_distance ||
5792 next_distance == curr_distance))
5793 next_distance = distance;
5796 * While not a strong assumption it would be nice to know
5797 * about cases where if node A is connected to B, B is not
5798 * equally connected to A.
5800 if (sched_debug() && node_distance(k, i) != distance)
5801 sched_numa_warn("Node-distance not symmetric");
5803 if (sched_debug() && i && !find_numa_distance(distance))
5804 sched_numa_warn("Node-0 not representative");
5806 if (next_distance != curr_distance) {
5807 sched_domains_numa_distance[level++] = next_distance;
5808 sched_domains_numa_levels = level;
5809 curr_distance = next_distance;
5814 * In case of sched_debug() we verify the above assumption.
5820 * 'level' contains the number of unique distances, excluding the
5821 * identity distance node_distance(i,i).
5823 * The sched_domains_numa_distance[] array includes the actual distance
5828 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5829 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5830 * the array will contain less then 'level' members. This could be
5831 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5832 * in other functions.
5834 * We reset it to 'level' at the end of this function.
5836 sched_domains_numa_levels = 0;
5838 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5839 if (!sched_domains_numa_masks)
5843 * Now for each level, construct a mask per node which contains all
5844 * cpus of nodes that are that many hops away from us.
5846 for (i = 0; i < level; i++) {
5847 sched_domains_numa_masks[i] =
5848 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5849 if (!sched_domains_numa_masks[i])
5852 for (j = 0; j < nr_node_ids; j++) {
5853 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5857 sched_domains_numa_masks[i][j] = mask;
5859 for (k = 0; k < nr_node_ids; k++) {
5860 if (node_distance(j, k) > sched_domains_numa_distance[i])
5863 cpumask_or(mask, mask, cpumask_of_node(k));
5868 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5869 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5874 * Copy the default topology bits..
5876 for (i = 0; default_topology[i].init; i++)
5877 tl[i] = default_topology[i];
5880 * .. and append 'j' levels of NUMA goodness.
5882 for (j = 0; j < level; i++, j++) {
5883 tl[i] = (struct sched_domain_topology_level){
5884 .init = sd_numa_init,
5885 .mask = sd_numa_mask,
5886 .flags = SDTL_OVERLAP,
5891 sched_domain_topology = tl;
5893 sched_domains_numa_levels = level;
5896 static void sched_domains_numa_masks_set(int cpu)
5899 int node = cpu_to_node(cpu);
5901 for (i = 0; i < sched_domains_numa_levels; i++) {
5902 for (j = 0; j < nr_node_ids; j++) {
5903 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5904 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5909 static void sched_domains_numa_masks_clear(int cpu)
5912 for (i = 0; i < sched_domains_numa_levels; i++) {
5913 for (j = 0; j < nr_node_ids; j++)
5914 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5919 * Update sched_domains_numa_masks[level][node] array when new cpus
5922 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5923 unsigned long action,
5926 int cpu = (long)hcpu;
5928 switch (action & ~CPU_TASKS_FROZEN) {
5930 sched_domains_numa_masks_set(cpu);
5934 sched_domains_numa_masks_clear(cpu);
5944 static inline void sched_init_numa(void)
5948 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5949 unsigned long action,
5954 #endif /* CONFIG_NUMA */
5956 static int __sdt_alloc(const struct cpumask *cpu_map)
5958 struct sched_domain_topology_level *tl;
5961 for_each_sd_topology(tl) {
5962 struct sd_data *sdd = &tl->data;
5964 sdd->sd = alloc_percpu(struct sched_domain *);
5968 sdd->sg = alloc_percpu(struct sched_group *);
5972 sdd->sgp = alloc_percpu(struct sched_group_power *);
5976 for_each_cpu(j, cpu_map) {
5977 struct sched_domain *sd;
5978 struct sched_group *sg;
5979 struct sched_group_power *sgp;
5981 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5982 GFP_KERNEL, cpu_to_node(j));
5986 *per_cpu_ptr(sdd->sd, j) = sd;
5988 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5989 GFP_KERNEL, cpu_to_node(j));
5995 *per_cpu_ptr(sdd->sg, j) = sg;
5997 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5998 GFP_KERNEL, cpu_to_node(j));
6002 *per_cpu_ptr(sdd->sgp, j) = sgp;
6009 static void __sdt_free(const struct cpumask *cpu_map)
6011 struct sched_domain_topology_level *tl;
6014 for_each_sd_topology(tl) {
6015 struct sd_data *sdd = &tl->data;
6017 for_each_cpu(j, cpu_map) {
6018 struct sched_domain *sd;
6021 sd = *per_cpu_ptr(sdd->sd, j);
6022 if (sd && (sd->flags & SD_OVERLAP))
6023 free_sched_groups(sd->groups, 0);
6024 kfree(*per_cpu_ptr(sdd->sd, j));
6028 kfree(*per_cpu_ptr(sdd->sg, j));
6030 kfree(*per_cpu_ptr(sdd->sgp, j));
6032 free_percpu(sdd->sd);
6034 free_percpu(sdd->sg);
6036 free_percpu(sdd->sgp);
6041 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6042 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6043 struct sched_domain *child, int cpu)
6045 struct sched_domain *sd = tl->init(tl, cpu);
6049 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6051 sd->level = child->level + 1;
6052 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6056 set_domain_attribute(sd, attr);
6062 * Build sched domains for a given set of cpus and attach the sched domains
6063 * to the individual cpus
6065 static int build_sched_domains(const struct cpumask *cpu_map,
6066 struct sched_domain_attr *attr)
6068 enum s_alloc alloc_state;
6069 struct sched_domain *sd;
6071 int i, ret = -ENOMEM;
6073 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6074 if (alloc_state != sa_rootdomain)
6077 /* Set up domains for cpus specified by the cpu_map. */
6078 for_each_cpu(i, cpu_map) {
6079 struct sched_domain_topology_level *tl;
6082 for_each_sd_topology(tl) {
6083 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6084 if (tl == sched_domain_topology)
6085 *per_cpu_ptr(d.sd, i) = sd;
6086 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6087 sd->flags |= SD_OVERLAP;
6088 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6093 /* Build the groups for the domains */
6094 for_each_cpu(i, cpu_map) {
6095 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6096 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6097 if (sd->flags & SD_OVERLAP) {
6098 if (build_overlap_sched_groups(sd, i))
6101 if (build_sched_groups(sd, i))
6107 /* Calculate CPU power for physical packages and nodes */
6108 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6109 if (!cpumask_test_cpu(i, cpu_map))
6112 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6113 claim_allocations(i, sd);
6114 init_sched_groups_power(i, sd);
6118 /* Attach the domains */
6120 for_each_cpu(i, cpu_map) {
6121 sd = *per_cpu_ptr(d.sd, i);
6122 cpu_attach_domain(sd, d.rd, i);
6128 __free_domain_allocs(&d, alloc_state, cpu_map);
6132 static cpumask_var_t *doms_cur; /* current sched domains */
6133 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6134 static struct sched_domain_attr *dattr_cur;
6135 /* attribues of custom domains in 'doms_cur' */
6138 * Special case: If a kmalloc of a doms_cur partition (array of
6139 * cpumask) fails, then fallback to a single sched domain,
6140 * as determined by the single cpumask fallback_doms.
6142 static cpumask_var_t fallback_doms;
6145 * arch_update_cpu_topology lets virtualized architectures update the
6146 * cpu core maps. It is supposed to return 1 if the topology changed
6147 * or 0 if it stayed the same.
6149 int __attribute__((weak)) arch_update_cpu_topology(void)
6154 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6157 cpumask_var_t *doms;
6159 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6162 for (i = 0; i < ndoms; i++) {
6163 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6164 free_sched_domains(doms, i);
6171 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6174 for (i = 0; i < ndoms; i++)
6175 free_cpumask_var(doms[i]);
6180 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6181 * For now this just excludes isolated cpus, but could be used to
6182 * exclude other special cases in the future.
6184 static int init_sched_domains(const struct cpumask *cpu_map)
6188 arch_update_cpu_topology();
6190 doms_cur = alloc_sched_domains(ndoms_cur);
6192 doms_cur = &fallback_doms;
6193 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6194 err = build_sched_domains(doms_cur[0], NULL);
6195 register_sched_domain_sysctl();
6201 * Detach sched domains from a group of cpus specified in cpu_map
6202 * These cpus will now be attached to the NULL domain
6204 static void detach_destroy_domains(const struct cpumask *cpu_map)
6209 for_each_cpu(i, cpu_map)
6210 cpu_attach_domain(NULL, &def_root_domain, i);
6214 /* handle null as "default" */
6215 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6216 struct sched_domain_attr *new, int idx_new)
6218 struct sched_domain_attr tmp;
6225 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6226 new ? (new + idx_new) : &tmp,
6227 sizeof(struct sched_domain_attr));
6231 * Partition sched domains as specified by the 'ndoms_new'
6232 * cpumasks in the array doms_new[] of cpumasks. This compares
6233 * doms_new[] to the current sched domain partitioning, doms_cur[].
6234 * It destroys each deleted domain and builds each new domain.
6236 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6237 * The masks don't intersect (don't overlap.) We should setup one
6238 * sched domain for each mask. CPUs not in any of the cpumasks will
6239 * not be load balanced. If the same cpumask appears both in the
6240 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6243 * The passed in 'doms_new' should be allocated using
6244 * alloc_sched_domains. This routine takes ownership of it and will
6245 * free_sched_domains it when done with it. If the caller failed the
6246 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6247 * and partition_sched_domains() will fallback to the single partition
6248 * 'fallback_doms', it also forces the domains to be rebuilt.
6250 * If doms_new == NULL it will be replaced with cpu_online_mask.
6251 * ndoms_new == 0 is a special case for destroying existing domains,
6252 * and it will not create the default domain.
6254 * Call with hotplug lock held
6256 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6257 struct sched_domain_attr *dattr_new)
6262 mutex_lock(&sched_domains_mutex);
6264 /* always unregister in case we don't destroy any domains */
6265 unregister_sched_domain_sysctl();
6267 /* Let architecture update cpu core mappings. */
6268 new_topology = arch_update_cpu_topology();
6270 n = doms_new ? ndoms_new : 0;
6272 /* Destroy deleted domains */
6273 for (i = 0; i < ndoms_cur; i++) {
6274 for (j = 0; j < n && !new_topology; j++) {
6275 if (cpumask_equal(doms_cur[i], doms_new[j])
6276 && dattrs_equal(dattr_cur, i, dattr_new, j))
6279 /* no match - a current sched domain not in new doms_new[] */
6280 detach_destroy_domains(doms_cur[i]);
6286 if (doms_new == NULL) {
6288 doms_new = &fallback_doms;
6289 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6290 WARN_ON_ONCE(dattr_new);
6293 /* Build new domains */
6294 for (i = 0; i < ndoms_new; i++) {
6295 for (j = 0; j < n && !new_topology; j++) {
6296 if (cpumask_equal(doms_new[i], doms_cur[j])
6297 && dattrs_equal(dattr_new, i, dattr_cur, j))
6300 /* no match - add a new doms_new */
6301 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6306 /* Remember the new sched domains */
6307 if (doms_cur != &fallback_doms)
6308 free_sched_domains(doms_cur, ndoms_cur);
6309 kfree(dattr_cur); /* kfree(NULL) is safe */
6310 doms_cur = doms_new;
6311 dattr_cur = dattr_new;
6312 ndoms_cur = ndoms_new;
6314 register_sched_domain_sysctl();
6316 mutex_unlock(&sched_domains_mutex);
6319 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6322 * Update cpusets according to cpu_active mask. If cpusets are
6323 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6324 * around partition_sched_domains().
6326 * If we come here as part of a suspend/resume, don't touch cpusets because we
6327 * want to restore it back to its original state upon resume anyway.
6329 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6333 case CPU_ONLINE_FROZEN:
6334 case CPU_DOWN_FAILED_FROZEN:
6337 * num_cpus_frozen tracks how many CPUs are involved in suspend
6338 * resume sequence. As long as this is not the last online
6339 * operation in the resume sequence, just build a single sched
6340 * domain, ignoring cpusets.
6343 if (likely(num_cpus_frozen)) {
6344 partition_sched_domains(1, NULL, NULL);
6349 * This is the last CPU online operation. So fall through and
6350 * restore the original sched domains by considering the
6351 * cpuset configurations.
6355 case CPU_DOWN_FAILED:
6356 cpuset_update_active_cpus(true);
6364 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6368 case CPU_DOWN_PREPARE:
6369 cpuset_update_active_cpus(false);
6371 case CPU_DOWN_PREPARE_FROZEN:
6373 partition_sched_domains(1, NULL, NULL);
6381 void __init sched_init_smp(void)
6383 cpumask_var_t non_isolated_cpus;
6385 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6386 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6391 * There's no userspace yet to cause hotplug operations; hence all the
6392 * cpu masks are stable and all blatant races in the below code cannot
6395 mutex_lock(&sched_domains_mutex);
6396 init_sched_domains(cpu_active_mask);
6397 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6398 if (cpumask_empty(non_isolated_cpus))
6399 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6400 mutex_unlock(&sched_domains_mutex);
6402 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6403 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6404 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6408 /* Move init over to a non-isolated CPU */
6409 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6411 sched_init_granularity();
6412 free_cpumask_var(non_isolated_cpus);
6414 init_sched_rt_class();
6417 void __init sched_init_smp(void)
6419 sched_init_granularity();
6421 #endif /* CONFIG_SMP */
6423 const_debug unsigned int sysctl_timer_migration = 1;
6425 int in_sched_functions(unsigned long addr)
6427 return in_lock_functions(addr) ||
6428 (addr >= (unsigned long)__sched_text_start
6429 && addr < (unsigned long)__sched_text_end);
6432 #ifdef CONFIG_CGROUP_SCHED
6434 * Default task group.
6435 * Every task in system belongs to this group at bootup.
6437 struct task_group root_task_group;
6438 LIST_HEAD(task_groups);
6441 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6443 void __init sched_init(void)
6446 unsigned long alloc_size = 0, ptr;
6448 #ifdef CONFIG_FAIR_GROUP_SCHED
6449 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6451 #ifdef CONFIG_RT_GROUP_SCHED
6452 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6454 #ifdef CONFIG_CPUMASK_OFFSTACK
6455 alloc_size += num_possible_cpus() * cpumask_size();
6458 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6460 #ifdef CONFIG_FAIR_GROUP_SCHED
6461 root_task_group.se = (struct sched_entity **)ptr;
6462 ptr += nr_cpu_ids * sizeof(void **);
6464 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6465 ptr += nr_cpu_ids * sizeof(void **);
6467 #endif /* CONFIG_FAIR_GROUP_SCHED */
6468 #ifdef CONFIG_RT_GROUP_SCHED
6469 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6470 ptr += nr_cpu_ids * sizeof(void **);
6472 root_task_group.rt_rq = (struct rt_rq **)ptr;
6473 ptr += nr_cpu_ids * sizeof(void **);
6475 #endif /* CONFIG_RT_GROUP_SCHED */
6476 #ifdef CONFIG_CPUMASK_OFFSTACK
6477 for_each_possible_cpu(i) {
6478 per_cpu(load_balance_mask, i) = (void *)ptr;
6479 ptr += cpumask_size();
6481 #endif /* CONFIG_CPUMASK_OFFSTACK */
6485 init_defrootdomain();
6488 init_rt_bandwidth(&def_rt_bandwidth,
6489 global_rt_period(), global_rt_runtime());
6491 #ifdef CONFIG_RT_GROUP_SCHED
6492 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6493 global_rt_period(), global_rt_runtime());
6494 #endif /* CONFIG_RT_GROUP_SCHED */
6496 #ifdef CONFIG_CGROUP_SCHED
6497 list_add(&root_task_group.list, &task_groups);
6498 INIT_LIST_HEAD(&root_task_group.children);
6499 INIT_LIST_HEAD(&root_task_group.siblings);
6500 autogroup_init(&init_task);
6502 #endif /* CONFIG_CGROUP_SCHED */
6504 for_each_possible_cpu(i) {
6508 raw_spin_lock_init(&rq->lock);
6510 rq->calc_load_active = 0;
6511 rq->calc_load_update = jiffies + LOAD_FREQ;
6512 init_cfs_rq(&rq->cfs);
6513 init_rt_rq(&rq->rt, rq);
6514 #ifdef CONFIG_FAIR_GROUP_SCHED
6515 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6516 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6518 * How much cpu bandwidth does root_task_group get?
6520 * In case of task-groups formed thr' the cgroup filesystem, it
6521 * gets 100% of the cpu resources in the system. This overall
6522 * system cpu resource is divided among the tasks of
6523 * root_task_group and its child task-groups in a fair manner,
6524 * based on each entity's (task or task-group's) weight
6525 * (se->load.weight).
6527 * In other words, if root_task_group has 10 tasks of weight
6528 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6529 * then A0's share of the cpu resource is:
6531 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6533 * We achieve this by letting root_task_group's tasks sit
6534 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6536 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6537 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6538 #endif /* CONFIG_FAIR_GROUP_SCHED */
6540 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6541 #ifdef CONFIG_RT_GROUP_SCHED
6542 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6543 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6546 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6547 rq->cpu_load[j] = 0;
6549 rq->last_load_update_tick = jiffies;
6554 rq->cpu_power = SCHED_POWER_SCALE;
6555 rq->post_schedule = 0;
6556 rq->active_balance = 0;
6557 rq->next_balance = jiffies;
6562 rq->avg_idle = 2*sysctl_sched_migration_cost;
6563 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6565 INIT_LIST_HEAD(&rq->cfs_tasks);
6567 rq_attach_root(rq, &def_root_domain);
6568 #ifdef CONFIG_NO_HZ_COMMON
6571 #ifdef CONFIG_NO_HZ_FULL
6572 rq->last_sched_tick = 0;
6576 atomic_set(&rq->nr_iowait, 0);
6579 set_load_weight(&init_task);
6581 #ifdef CONFIG_PREEMPT_NOTIFIERS
6582 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6585 #ifdef CONFIG_RT_MUTEXES
6586 plist_head_init(&init_task.pi_waiters);
6590 * The boot idle thread does lazy MMU switching as well:
6592 atomic_inc(&init_mm.mm_count);
6593 enter_lazy_tlb(&init_mm, current);
6596 * Make us the idle thread. Technically, schedule() should not be
6597 * called from this thread, however somewhere below it might be,
6598 * but because we are the idle thread, we just pick up running again
6599 * when this runqueue becomes "idle".
6601 init_idle(current, smp_processor_id());
6603 calc_load_update = jiffies + LOAD_FREQ;
6606 * During early bootup we pretend to be a normal task:
6608 current->sched_class = &fair_sched_class;
6611 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6612 /* May be allocated at isolcpus cmdline parse time */
6613 if (cpu_isolated_map == NULL)
6614 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6615 idle_thread_set_boot_cpu();
6617 init_sched_fair_class();
6619 scheduler_running = 1;
6622 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6623 static inline int preempt_count_equals(int preempt_offset)
6625 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6627 return (nested == preempt_offset);
6630 void __might_sleep(const char *file, int line, int preempt_offset)
6632 static unsigned long prev_jiffy; /* ratelimiting */
6634 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6635 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6636 system_state != SYSTEM_RUNNING || oops_in_progress)
6638 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6640 prev_jiffy = jiffies;
6643 "BUG: sleeping function called from invalid context at %s:%d\n",
6646 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6647 in_atomic(), irqs_disabled(),
6648 current->pid, current->comm);
6650 debug_show_held_locks(current);
6651 if (irqs_disabled())
6652 print_irqtrace_events(current);
6655 EXPORT_SYMBOL(__might_sleep);
6658 #ifdef CONFIG_MAGIC_SYSRQ
6659 static void normalize_task(struct rq *rq, struct task_struct *p)
6661 const struct sched_class *prev_class = p->sched_class;
6662 int old_prio = p->prio;
6667 dequeue_task(rq, p, 0);
6668 __setscheduler(rq, p, SCHED_NORMAL, 0);
6670 enqueue_task(rq, p, 0);
6671 resched_task(rq->curr);
6674 check_class_changed(rq, p, prev_class, old_prio);
6677 void normalize_rt_tasks(void)
6679 struct task_struct *g, *p;
6680 unsigned long flags;
6683 read_lock_irqsave(&tasklist_lock, flags);
6684 do_each_thread(g, p) {
6686 * Only normalize user tasks:
6691 p->se.exec_start = 0;
6692 #ifdef CONFIG_SCHEDSTATS
6693 p->se.statistics.wait_start = 0;
6694 p->se.statistics.sleep_start = 0;
6695 p->se.statistics.block_start = 0;
6700 * Renice negative nice level userspace
6703 if (TASK_NICE(p) < 0 && p->mm)
6704 set_user_nice(p, 0);
6708 raw_spin_lock(&p->pi_lock);
6709 rq = __task_rq_lock(p);
6711 normalize_task(rq, p);
6713 __task_rq_unlock(rq);
6714 raw_spin_unlock(&p->pi_lock);
6715 } while_each_thread(g, p);
6717 read_unlock_irqrestore(&tasklist_lock, flags);
6720 #endif /* CONFIG_MAGIC_SYSRQ */
6722 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6724 * These functions are only useful for the IA64 MCA handling, or kdb.
6726 * They can only be called when the whole system has been
6727 * stopped - every CPU needs to be quiescent, and no scheduling
6728 * activity can take place. Using them for anything else would
6729 * be a serious bug, and as a result, they aren't even visible
6730 * under any other configuration.
6734 * curr_task - return the current task for a given cpu.
6735 * @cpu: the processor in question.
6737 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6739 * Return: The current task for @cpu.
6741 struct task_struct *curr_task(int cpu)
6743 return cpu_curr(cpu);
6746 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6750 * set_curr_task - set the current task for a given cpu.
6751 * @cpu: the processor in question.
6752 * @p: the task pointer to set.
6754 * Description: This function must only be used when non-maskable interrupts
6755 * are serviced on a separate stack. It allows the architecture to switch the
6756 * notion of the current task on a cpu in a non-blocking manner. This function
6757 * must be called with all CPU's synchronized, and interrupts disabled, the
6758 * and caller must save the original value of the current task (see
6759 * curr_task() above) and restore that value before reenabling interrupts and
6760 * re-starting the system.
6762 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6764 void set_curr_task(int cpu, struct task_struct *p)
6771 #ifdef CONFIG_CGROUP_SCHED
6772 /* task_group_lock serializes the addition/removal of task groups */
6773 static DEFINE_SPINLOCK(task_group_lock);
6775 static void free_sched_group(struct task_group *tg)
6777 free_fair_sched_group(tg);
6778 free_rt_sched_group(tg);
6783 /* allocate runqueue etc for a new task group */
6784 struct task_group *sched_create_group(struct task_group *parent)
6786 struct task_group *tg;
6788 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6790 return ERR_PTR(-ENOMEM);
6792 if (!alloc_fair_sched_group(tg, parent))
6795 if (!alloc_rt_sched_group(tg, parent))
6801 free_sched_group(tg);
6802 return ERR_PTR(-ENOMEM);
6805 void sched_online_group(struct task_group *tg, struct task_group *parent)
6807 unsigned long flags;
6809 spin_lock_irqsave(&task_group_lock, flags);
6810 list_add_rcu(&tg->list, &task_groups);
6812 WARN_ON(!parent); /* root should already exist */
6814 tg->parent = parent;
6815 INIT_LIST_HEAD(&tg->children);
6816 list_add_rcu(&tg->siblings, &parent->children);
6817 spin_unlock_irqrestore(&task_group_lock, flags);
6820 /* rcu callback to free various structures associated with a task group */
6821 static void free_sched_group_rcu(struct rcu_head *rhp)
6823 /* now it should be safe to free those cfs_rqs */
6824 free_sched_group(container_of(rhp, struct task_group, rcu));
6827 /* Destroy runqueue etc associated with a task group */
6828 void sched_destroy_group(struct task_group *tg)
6830 /* wait for possible concurrent references to cfs_rqs complete */
6831 call_rcu(&tg->rcu, free_sched_group_rcu);
6834 void sched_offline_group(struct task_group *tg)
6836 unsigned long flags;
6839 /* end participation in shares distribution */
6840 for_each_possible_cpu(i)
6841 unregister_fair_sched_group(tg, i);
6843 spin_lock_irqsave(&task_group_lock, flags);
6844 list_del_rcu(&tg->list);
6845 list_del_rcu(&tg->siblings);
6846 spin_unlock_irqrestore(&task_group_lock, flags);
6849 /* change task's runqueue when it moves between groups.
6850 * The caller of this function should have put the task in its new group
6851 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6852 * reflect its new group.
6854 void sched_move_task(struct task_struct *tsk)
6856 struct task_group *tg;
6858 unsigned long flags;
6861 rq = task_rq_lock(tsk, &flags);
6863 running = task_current(rq, tsk);
6867 dequeue_task(rq, tsk, 0);
6868 if (unlikely(running))
6869 tsk->sched_class->put_prev_task(rq, tsk);
6871 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6872 lockdep_is_held(&tsk->sighand->siglock)),
6873 struct task_group, css);
6874 tg = autogroup_task_group(tsk, tg);
6875 tsk->sched_task_group = tg;
6877 #ifdef CONFIG_FAIR_GROUP_SCHED
6878 if (tsk->sched_class->task_move_group)
6879 tsk->sched_class->task_move_group(tsk, on_rq);
6882 set_task_rq(tsk, task_cpu(tsk));
6884 if (unlikely(running))
6885 tsk->sched_class->set_curr_task(rq);
6887 enqueue_task(rq, tsk, 0);
6889 task_rq_unlock(rq, tsk, &flags);
6891 #endif /* CONFIG_CGROUP_SCHED */
6893 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6894 static unsigned long to_ratio(u64 period, u64 runtime)
6896 if (runtime == RUNTIME_INF)
6899 return div64_u64(runtime << 20, period);
6903 #ifdef CONFIG_RT_GROUP_SCHED
6905 * Ensure that the real time constraints are schedulable.
6907 static DEFINE_MUTEX(rt_constraints_mutex);
6909 /* Must be called with tasklist_lock held */
6910 static inline int tg_has_rt_tasks(struct task_group *tg)
6912 struct task_struct *g, *p;
6914 do_each_thread(g, p) {
6915 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6917 } while_each_thread(g, p);
6922 struct rt_schedulable_data {
6923 struct task_group *tg;
6928 static int tg_rt_schedulable(struct task_group *tg, void *data)
6930 struct rt_schedulable_data *d = data;
6931 struct task_group *child;
6932 unsigned long total, sum = 0;
6933 u64 period, runtime;
6935 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6936 runtime = tg->rt_bandwidth.rt_runtime;
6939 period = d->rt_period;
6940 runtime = d->rt_runtime;
6944 * Cannot have more runtime than the period.
6946 if (runtime > period && runtime != RUNTIME_INF)
6950 * Ensure we don't starve existing RT tasks.
6952 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6955 total = to_ratio(period, runtime);
6958 * Nobody can have more than the global setting allows.
6960 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6964 * The sum of our children's runtime should not exceed our own.
6966 list_for_each_entry_rcu(child, &tg->children, siblings) {
6967 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6968 runtime = child->rt_bandwidth.rt_runtime;
6970 if (child == d->tg) {
6971 period = d->rt_period;
6972 runtime = d->rt_runtime;
6975 sum += to_ratio(period, runtime);
6984 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6988 struct rt_schedulable_data data = {
6990 .rt_period = period,
6991 .rt_runtime = runtime,
6995 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7001 static int tg_set_rt_bandwidth(struct task_group *tg,
7002 u64 rt_period, u64 rt_runtime)
7006 mutex_lock(&rt_constraints_mutex);
7007 read_lock(&tasklist_lock);
7008 err = __rt_schedulable(tg, rt_period, rt_runtime);
7012 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7013 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7014 tg->rt_bandwidth.rt_runtime = rt_runtime;
7016 for_each_possible_cpu(i) {
7017 struct rt_rq *rt_rq = tg->rt_rq[i];
7019 raw_spin_lock(&rt_rq->rt_runtime_lock);
7020 rt_rq->rt_runtime = rt_runtime;
7021 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7023 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7025 read_unlock(&tasklist_lock);
7026 mutex_unlock(&rt_constraints_mutex);
7031 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7033 u64 rt_runtime, rt_period;
7035 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7036 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7037 if (rt_runtime_us < 0)
7038 rt_runtime = RUNTIME_INF;
7040 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7043 static long sched_group_rt_runtime(struct task_group *tg)
7047 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7050 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7051 do_div(rt_runtime_us, NSEC_PER_USEC);
7052 return rt_runtime_us;
7055 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7057 u64 rt_runtime, rt_period;
7059 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7060 rt_runtime = tg->rt_bandwidth.rt_runtime;
7065 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7068 static long sched_group_rt_period(struct task_group *tg)
7072 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7073 do_div(rt_period_us, NSEC_PER_USEC);
7074 return rt_period_us;
7077 static int sched_rt_global_constraints(void)
7079 u64 runtime, period;
7082 if (sysctl_sched_rt_period <= 0)
7085 runtime = global_rt_runtime();
7086 period = global_rt_period();
7089 * Sanity check on the sysctl variables.
7091 if (runtime > period && runtime != RUNTIME_INF)
7094 mutex_lock(&rt_constraints_mutex);
7095 read_lock(&tasklist_lock);
7096 ret = __rt_schedulable(NULL, 0, 0);
7097 read_unlock(&tasklist_lock);
7098 mutex_unlock(&rt_constraints_mutex);
7103 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7105 /* Don't accept realtime tasks when there is no way for them to run */
7106 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7112 #else /* !CONFIG_RT_GROUP_SCHED */
7113 static int sched_rt_global_constraints(void)
7115 unsigned long flags;
7118 if (sysctl_sched_rt_period <= 0)
7122 * There's always some RT tasks in the root group
7123 * -- migration, kstopmachine etc..
7125 if (sysctl_sched_rt_runtime == 0)
7128 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7129 for_each_possible_cpu(i) {
7130 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7132 raw_spin_lock(&rt_rq->rt_runtime_lock);
7133 rt_rq->rt_runtime = global_rt_runtime();
7134 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7136 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7140 #endif /* CONFIG_RT_GROUP_SCHED */
7142 int sched_rr_handler(struct ctl_table *table, int write,
7143 void __user *buffer, size_t *lenp,
7147 static DEFINE_MUTEX(mutex);
7150 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7151 /* make sure that internally we keep jiffies */
7152 /* also, writing zero resets timeslice to default */
7153 if (!ret && write) {
7154 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7155 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7157 mutex_unlock(&mutex);
7161 int sched_rt_handler(struct ctl_table *table, int write,
7162 void __user *buffer, size_t *lenp,
7166 int old_period, old_runtime;
7167 static DEFINE_MUTEX(mutex);
7170 old_period = sysctl_sched_rt_period;
7171 old_runtime = sysctl_sched_rt_runtime;
7173 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7175 if (!ret && write) {
7176 ret = sched_rt_global_constraints();
7178 sysctl_sched_rt_period = old_period;
7179 sysctl_sched_rt_runtime = old_runtime;
7181 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7182 def_rt_bandwidth.rt_period =
7183 ns_to_ktime(global_rt_period());
7186 mutex_unlock(&mutex);
7191 #ifdef CONFIG_CGROUP_SCHED
7193 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7195 return css ? container_of(css, struct task_group, css) : NULL;
7198 static struct cgroup_subsys_state *
7199 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7201 struct task_group *parent = css_tg(parent_css);
7202 struct task_group *tg;
7205 /* This is early initialization for the top cgroup */
7206 return &root_task_group.css;
7209 tg = sched_create_group(parent);
7211 return ERR_PTR(-ENOMEM);
7216 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7218 struct task_group *tg = css_tg(css);
7219 struct task_group *parent = css_tg(css_parent(css));
7222 sched_online_group(tg, parent);
7226 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7228 struct task_group *tg = css_tg(css);
7230 sched_destroy_group(tg);
7233 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7235 struct task_group *tg = css_tg(css);
7237 sched_offline_group(tg);
7240 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7241 struct cgroup_taskset *tset)
7243 struct task_struct *task;
7245 cgroup_taskset_for_each(task, css, tset) {
7246 #ifdef CONFIG_RT_GROUP_SCHED
7247 if (!sched_rt_can_attach(css_tg(css), task))
7250 /* We don't support RT-tasks being in separate groups */
7251 if (task->sched_class != &fair_sched_class)
7258 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7259 struct cgroup_taskset *tset)
7261 struct task_struct *task;
7263 cgroup_taskset_for_each(task, css, tset)
7264 sched_move_task(task);
7267 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7268 struct cgroup_subsys_state *old_css,
7269 struct task_struct *task)
7272 * cgroup_exit() is called in the copy_process() failure path.
7273 * Ignore this case since the task hasn't ran yet, this avoids
7274 * trying to poke a half freed task state from generic code.
7276 if (!(task->flags & PF_EXITING))
7279 sched_move_task(task);
7282 #ifdef CONFIG_FAIR_GROUP_SCHED
7283 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7284 struct cftype *cftype, u64 shareval)
7286 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7289 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7292 struct task_group *tg = css_tg(css);
7294 return (u64) scale_load_down(tg->shares);
7297 #ifdef CONFIG_CFS_BANDWIDTH
7298 static DEFINE_MUTEX(cfs_constraints_mutex);
7300 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7301 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7303 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7305 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7307 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7308 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7310 if (tg == &root_task_group)
7314 * Ensure we have at some amount of bandwidth every period. This is
7315 * to prevent reaching a state of large arrears when throttled via
7316 * entity_tick() resulting in prolonged exit starvation.
7318 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7322 * Likewise, bound things on the otherside by preventing insane quota
7323 * periods. This also allows us to normalize in computing quota
7326 if (period > max_cfs_quota_period)
7329 mutex_lock(&cfs_constraints_mutex);
7330 ret = __cfs_schedulable(tg, period, quota);
7334 runtime_enabled = quota != RUNTIME_INF;
7335 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7337 * If we need to toggle cfs_bandwidth_used, off->on must occur
7338 * before making related changes, and on->off must occur afterwards
7340 if (runtime_enabled && !runtime_was_enabled)
7341 cfs_bandwidth_usage_inc();
7342 raw_spin_lock_irq(&cfs_b->lock);
7343 cfs_b->period = ns_to_ktime(period);
7344 cfs_b->quota = quota;
7346 __refill_cfs_bandwidth_runtime(cfs_b);
7347 /* restart the period timer (if active) to handle new period expiry */
7348 if (runtime_enabled && cfs_b->timer_active) {
7349 /* force a reprogram */
7350 cfs_b->timer_active = 0;
7351 __start_cfs_bandwidth(cfs_b);
7353 raw_spin_unlock_irq(&cfs_b->lock);
7355 for_each_possible_cpu(i) {
7356 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7357 struct rq *rq = cfs_rq->rq;
7359 raw_spin_lock_irq(&rq->lock);
7360 cfs_rq->runtime_enabled = runtime_enabled;
7361 cfs_rq->runtime_remaining = 0;
7363 if (cfs_rq->throttled)
7364 unthrottle_cfs_rq(cfs_rq);
7365 raw_spin_unlock_irq(&rq->lock);
7367 if (runtime_was_enabled && !runtime_enabled)
7368 cfs_bandwidth_usage_dec();
7370 mutex_unlock(&cfs_constraints_mutex);
7375 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7379 period = ktime_to_ns(tg->cfs_bandwidth.period);
7380 if (cfs_quota_us < 0)
7381 quota = RUNTIME_INF;
7383 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7385 return tg_set_cfs_bandwidth(tg, period, quota);
7388 long tg_get_cfs_quota(struct task_group *tg)
7392 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7395 quota_us = tg->cfs_bandwidth.quota;
7396 do_div(quota_us, NSEC_PER_USEC);
7401 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7405 period = (u64)cfs_period_us * NSEC_PER_USEC;
7406 quota = tg->cfs_bandwidth.quota;
7408 return tg_set_cfs_bandwidth(tg, period, quota);
7411 long tg_get_cfs_period(struct task_group *tg)
7415 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7416 do_div(cfs_period_us, NSEC_PER_USEC);
7418 return cfs_period_us;
7421 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7424 return tg_get_cfs_quota(css_tg(css));
7427 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7428 struct cftype *cftype, s64 cfs_quota_us)
7430 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7433 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7436 return tg_get_cfs_period(css_tg(css));
7439 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7440 struct cftype *cftype, u64 cfs_period_us)
7442 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7445 struct cfs_schedulable_data {
7446 struct task_group *tg;
7451 * normalize group quota/period to be quota/max_period
7452 * note: units are usecs
7454 static u64 normalize_cfs_quota(struct task_group *tg,
7455 struct cfs_schedulable_data *d)
7463 period = tg_get_cfs_period(tg);
7464 quota = tg_get_cfs_quota(tg);
7467 /* note: these should typically be equivalent */
7468 if (quota == RUNTIME_INF || quota == -1)
7471 return to_ratio(period, quota);
7474 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7476 struct cfs_schedulable_data *d = data;
7477 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7478 s64 quota = 0, parent_quota = -1;
7481 quota = RUNTIME_INF;
7483 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7485 quota = normalize_cfs_quota(tg, d);
7486 parent_quota = parent_b->hierarchal_quota;
7489 * ensure max(child_quota) <= parent_quota, inherit when no
7492 if (quota == RUNTIME_INF)
7493 quota = parent_quota;
7494 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7497 cfs_b->hierarchal_quota = quota;
7502 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7505 struct cfs_schedulable_data data = {
7511 if (quota != RUNTIME_INF) {
7512 do_div(data.period, NSEC_PER_USEC);
7513 do_div(data.quota, NSEC_PER_USEC);
7517 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7523 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7524 struct cgroup_map_cb *cb)
7526 struct task_group *tg = css_tg(css);
7527 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7529 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7530 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7531 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7535 #endif /* CONFIG_CFS_BANDWIDTH */
7536 #endif /* CONFIG_FAIR_GROUP_SCHED */
7538 #ifdef CONFIG_RT_GROUP_SCHED
7539 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7540 struct cftype *cft, s64 val)
7542 return sched_group_set_rt_runtime(css_tg(css), val);
7545 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7548 return sched_group_rt_runtime(css_tg(css));
7551 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7552 struct cftype *cftype, u64 rt_period_us)
7554 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7557 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7560 return sched_group_rt_period(css_tg(css));
7562 #endif /* CONFIG_RT_GROUP_SCHED */
7564 static struct cftype cpu_files[] = {
7565 #ifdef CONFIG_FAIR_GROUP_SCHED
7568 .read_u64 = cpu_shares_read_u64,
7569 .write_u64 = cpu_shares_write_u64,
7572 #ifdef CONFIG_CFS_BANDWIDTH
7574 .name = "cfs_quota_us",
7575 .read_s64 = cpu_cfs_quota_read_s64,
7576 .write_s64 = cpu_cfs_quota_write_s64,
7579 .name = "cfs_period_us",
7580 .read_u64 = cpu_cfs_period_read_u64,
7581 .write_u64 = cpu_cfs_period_write_u64,
7585 .read_map = cpu_stats_show,
7588 #ifdef CONFIG_RT_GROUP_SCHED
7590 .name = "rt_runtime_us",
7591 .read_s64 = cpu_rt_runtime_read,
7592 .write_s64 = cpu_rt_runtime_write,
7595 .name = "rt_period_us",
7596 .read_u64 = cpu_rt_period_read_uint,
7597 .write_u64 = cpu_rt_period_write_uint,
7603 struct cgroup_subsys cpu_cgroup_subsys = {
7605 .css_alloc = cpu_cgroup_css_alloc,
7606 .css_free = cpu_cgroup_css_free,
7607 .css_online = cpu_cgroup_css_online,
7608 .css_offline = cpu_cgroup_css_offline,
7609 .can_attach = cpu_cgroup_can_attach,
7610 .attach = cpu_cgroup_attach,
7611 .exit = cpu_cgroup_exit,
7612 .subsys_id = cpu_cgroup_subsys_id,
7613 .base_cftypes = cpu_files,
7617 #endif /* CONFIG_CGROUP_SCHED */
7619 void dump_cpu_task(int cpu)
7621 pr_info("Task dump for CPU %d:\n", cpu);
7622 sched_show_task(cpu_curr(cpu));