4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
375 static void hrtick_clear(struct rq *rq)
377 if (hrtimer_active(&rq->hrtick_timer))
378 hrtimer_cancel(&rq->hrtick_timer);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart hrtick(struct hrtimer *timer)
387 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
391 raw_spin_lock(&rq->lock);
393 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
394 raw_spin_unlock(&rq->lock);
396 return HRTIMER_NORESTART;
401 static int __hrtick_restart(struct rq *rq)
403 struct hrtimer *timer = &rq->hrtick_timer;
404 ktime_t time = hrtimer_get_softexpires(timer);
406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg)
416 raw_spin_lock(&rq->lock);
417 __hrtick_restart(rq);
418 rq->hrtick_csd_pending = 0;
419 raw_spin_unlock(&rq->lock);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq *rq, u64 delay)
429 struct hrtimer *timer = &rq->hrtick_timer;
430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
432 hrtimer_set_expires(timer, time);
434 if (rq == this_rq()) {
435 __hrtick_restart(rq);
436 } else if (!rq->hrtick_csd_pending) {
437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
438 rq->hrtick_csd_pending = 1;
443 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
445 int cpu = (int)(long)hcpu;
448 case CPU_UP_CANCELED:
449 case CPU_UP_CANCELED_FROZEN:
450 case CPU_DOWN_PREPARE:
451 case CPU_DOWN_PREPARE_FROZEN:
453 case CPU_DEAD_FROZEN:
454 hrtick_clear(cpu_rq(cpu));
461 static __init void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq *rq, u64 delay)
473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
474 HRTIMER_MODE_REL_PINNED, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq *rq)
485 rq->hrtick_csd_pending = 0;
487 rq->hrtick_csd.flags = 0;
488 rq->hrtick_csd.func = __hrtick_start;
489 rq->hrtick_csd.info = rq;
492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
493 rq->hrtick_timer.function = hrtick;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq *rq)
500 static inline void init_rq_hrtick(struct rq *rq)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
516 void resched_task(struct task_struct *p)
520 lockdep_assert_held(&task_rq(p)->lock);
522 if (test_tsk_need_resched(p))
525 set_tsk_need_resched(p);
528 if (cpu == smp_processor_id()) {
529 set_preempt_need_resched();
533 /* NEED_RESCHED must be visible before we test polling */
535 if (!tsk_is_polling(p))
536 smp_send_reschedule(cpu);
539 void resched_cpu(int cpu)
541 struct rq *rq = cpu_rq(cpu);
544 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
546 resched_task(cpu_curr(cpu));
547 raw_spin_unlock_irqrestore(&rq->lock, flags);
551 #ifdef CONFIG_NO_HZ_COMMON
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu = smp_processor_id();
564 struct sched_domain *sd;
567 for_each_domain(cpu, sd) {
568 for_each_cpu(i, sched_domain_span(sd)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 static void wake_up_idle_cpu(int cpu)
591 struct rq *rq = cpu_rq(cpu);
593 if (cpu == smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq->curr != rq->idle)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq->idle);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq->idle))
616 smp_send_reschedule(cpu);
619 static bool wake_up_full_nohz_cpu(int cpu)
621 if (tick_nohz_full_cpu(cpu)) {
622 if (cpu != smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 smp_send_reschedule(cpu);
631 void wake_up_nohz_cpu(int cpu)
633 if (!wake_up_full_nohz_cpu(cpu))
634 wake_up_idle_cpu(cpu);
637 static inline bool got_nohz_idle_kick(void)
639 int cpu = smp_processor_id();
641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644 if (idle_cpu(cpu) && !need_resched())
648 * We can't run Idle Load Balance on this CPU for this time so we
649 * cancel it and clear NOHZ_BALANCE_KICK
651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
655 #else /* CONFIG_NO_HZ_COMMON */
657 static inline bool got_nohz_idle_kick(void)
662 #endif /* CONFIG_NO_HZ_COMMON */
664 #ifdef CONFIG_NO_HZ_FULL
665 bool sched_can_stop_tick(void)
671 /* Make sure rq->nr_running update is visible after the IPI */
674 /* More than one running task need preemption */
675 if (rq->nr_running > 1)
680 #endif /* CONFIG_NO_HZ_FULL */
682 void sched_avg_update(struct rq *rq)
684 s64 period = sched_avg_period();
686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
688 * Inline assembly required to prevent the compiler
689 * optimising this loop into a divmod call.
690 * See __iter_div_u64_rem() for another example of this.
692 asm("" : "+rm" (rq->age_stamp));
693 rq->age_stamp += period;
698 #endif /* CONFIG_SMP */
700 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
703 * Iterate task_group tree rooted at *from, calling @down when first entering a
704 * node and @up when leaving it for the final time.
706 * Caller must hold rcu_lock or sufficient equivalent.
708 int walk_tg_tree_from(struct task_group *from,
709 tg_visitor down, tg_visitor up, void *data)
711 struct task_group *parent, *child;
717 ret = (*down)(parent, data);
720 list_for_each_entry_rcu(child, &parent->children, siblings) {
727 ret = (*up)(parent, data);
728 if (ret || parent == from)
732 parent = parent->parent;
739 int tg_nop(struct task_group *tg, void *data)
745 static void set_load_weight(struct task_struct *p)
747 int prio = p->static_prio - MAX_RT_PRIO;
748 struct load_weight *load = &p->se.load;
751 * SCHED_IDLE tasks get minimal weight:
753 if (p->policy == SCHED_IDLE) {
754 load->weight = scale_load(WEIGHT_IDLEPRIO);
755 load->inv_weight = WMULT_IDLEPRIO;
759 load->weight = scale_load(prio_to_weight[prio]);
760 load->inv_weight = prio_to_wmult[prio];
763 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 sched_info_queued(rq, p);
767 p->sched_class->enqueue_task(rq, p, flags);
770 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 sched_info_dequeued(rq, p);
774 p->sched_class->dequeue_task(rq, p, flags);
777 void activate_task(struct rq *rq, struct task_struct *p, int flags)
779 if (task_contributes_to_load(p))
780 rq->nr_uninterruptible--;
782 enqueue_task(rq, p, flags);
785 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
787 if (task_contributes_to_load(p))
788 rq->nr_uninterruptible++;
790 dequeue_task(rq, p, flags);
793 static void update_rq_clock_task(struct rq *rq, s64 delta)
796 * In theory, the compile should just see 0 here, and optimize out the call
797 * to sched_rt_avg_update. But I don't trust it...
799 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
800 s64 steal = 0, irq_delta = 0;
802 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806 * Since irq_time is only updated on {soft,}irq_exit, we might run into
807 * this case when a previous update_rq_clock() happened inside a
810 * When this happens, we stop ->clock_task and only update the
811 * prev_irq_time stamp to account for the part that fit, so that a next
812 * update will consume the rest. This ensures ->clock_task is
815 * It does however cause some slight miss-attribution of {soft,}irq
816 * time, a more accurate solution would be to update the irq_time using
817 * the current rq->clock timestamp, except that would require using
820 if (irq_delta > delta)
823 rq->prev_irq_time += irq_delta;
826 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
827 if (static_key_false((¶virt_steal_rq_enabled))) {
830 steal = paravirt_steal_clock(cpu_of(rq));
831 steal -= rq->prev_steal_time_rq;
833 if (unlikely(steal > delta))
836 st = steal_ticks(steal);
837 steal = st * TICK_NSEC;
839 rq->prev_steal_time_rq += steal;
845 rq->clock_task += delta;
847 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
849 sched_rt_avg_update(rq, irq_delta + steal);
853 void sched_set_stop_task(int cpu, struct task_struct *stop)
855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
856 struct task_struct *old_stop = cpu_rq(cpu)->stop;
860 * Make it appear like a SCHED_FIFO task, its something
861 * userspace knows about and won't get confused about.
863 * Also, it will make PI more or less work without too
864 * much confusion -- but then, stop work should not
865 * rely on PI working anyway.
867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
869 stop->sched_class = &stop_sched_class;
872 cpu_rq(cpu)->stop = stop;
876 * Reset it back to a normal scheduling class so that
877 * it can die in pieces.
879 old_stop->sched_class = &rt_sched_class;
884 * __normal_prio - return the priority that is based on the static prio
886 static inline int __normal_prio(struct task_struct *p)
888 return p->static_prio;
892 * Calculate the expected normal priority: i.e. priority
893 * without taking RT-inheritance into account. Might be
894 * boosted by interactivity modifiers. Changes upon fork,
895 * setprio syscalls, and whenever the interactivity
896 * estimator recalculates.
898 static inline int normal_prio(struct task_struct *p)
902 if (task_has_rt_policy(p))
903 prio = MAX_RT_PRIO-1 - p->rt_priority;
905 prio = __normal_prio(p);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct *p)
918 p->normal_prio = normal_prio(p);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p->prio))
925 return p->normal_prio;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio)
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
986 !(task_preempt_count(p) & PREEMPT_ACTIVE));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1000 lockdep_is_held(&task_rq(p)->lock)));
1004 trace_sched_migrate_task(p, new_cpu);
1006 if (task_cpu(p) != new_cpu) {
1007 if (p->sched_class->migrate_task_rq)
1008 p->sched_class->migrate_task_rq(p, new_cpu);
1009 p->se.nr_migrations++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1013 __set_task_cpu(p, new_cpu);
1016 static void __migrate_swap_task(struct task_struct *p, int cpu)
1019 struct rq *src_rq, *dst_rq;
1021 src_rq = task_rq(p);
1022 dst_rq = cpu_rq(cpu);
1024 deactivate_task(src_rq, p, 0);
1025 set_task_cpu(p, cpu);
1026 activate_task(dst_rq, p, 0);
1027 check_preempt_curr(dst_rq, p, 0);
1030 * Task isn't running anymore; make it appear like we migrated
1031 * it before it went to sleep. This means on wakeup we make the
1032 * previous cpu our targer instead of where it really is.
1038 struct migration_swap_arg {
1039 struct task_struct *src_task, *dst_task;
1040 int src_cpu, dst_cpu;
1043 static int migrate_swap_stop(void *data)
1045 struct migration_swap_arg *arg = data;
1046 struct rq *src_rq, *dst_rq;
1049 src_rq = cpu_rq(arg->src_cpu);
1050 dst_rq = cpu_rq(arg->dst_cpu);
1052 double_rq_lock(src_rq, dst_rq);
1053 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1056 if (task_cpu(arg->src_task) != arg->src_cpu)
1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1065 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1071 double_rq_unlock(src_rq, dst_rq);
1077 * Cross migrate two tasks
1079 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1081 struct migration_swap_arg arg;
1086 arg = (struct migration_swap_arg){
1088 .src_cpu = task_cpu(cur),
1090 .dst_cpu = task_cpu(p),
1093 if (arg.src_cpu == arg.dst_cpu)
1096 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1099 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1102 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1105 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1112 struct migration_arg {
1113 struct task_struct *task;
1117 static int migration_cpu_stop(void *data);
1120 * wait_task_inactive - wait for a thread to unschedule.
1122 * If @match_state is nonzero, it's the @p->state value just checked and
1123 * not expected to change. If it changes, i.e. @p might have woken up,
1124 * then return zero. When we succeed in waiting for @p to be off its CPU,
1125 * we return a positive number (its total switch count). If a second call
1126 * a short while later returns the same number, the caller can be sure that
1127 * @p has remained unscheduled the whole time.
1129 * The caller must ensure that the task *will* unschedule sometime soon,
1130 * else this function might spin for a *long* time. This function can't
1131 * be called with interrupts off, or it may introduce deadlock with
1132 * smp_call_function() if an IPI is sent by the same process we are
1133 * waiting to become inactive.
1135 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1137 unsigned long flags;
1144 * We do the initial early heuristics without holding
1145 * any task-queue locks at all. We'll only try to get
1146 * the runqueue lock when things look like they will
1152 * If the task is actively running on another CPU
1153 * still, just relax and busy-wait without holding
1156 * NOTE! Since we don't hold any locks, it's not
1157 * even sure that "rq" stays as the right runqueue!
1158 * But we don't care, since "task_running()" will
1159 * return false if the runqueue has changed and p
1160 * is actually now running somewhere else!
1162 while (task_running(rq, p)) {
1163 if (match_state && unlikely(p->state != match_state))
1169 * Ok, time to look more closely! We need the rq
1170 * lock now, to be *sure*. If we're wrong, we'll
1171 * just go back and repeat.
1173 rq = task_rq_lock(p, &flags);
1174 trace_sched_wait_task(p);
1175 running = task_running(rq, p);
1178 if (!match_state || p->state == match_state)
1179 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1180 task_rq_unlock(rq, p, &flags);
1183 * If it changed from the expected state, bail out now.
1185 if (unlikely(!ncsw))
1189 * Was it really running after all now that we
1190 * checked with the proper locks actually held?
1192 * Oops. Go back and try again..
1194 if (unlikely(running)) {
1200 * It's not enough that it's not actively running,
1201 * it must be off the runqueue _entirely_, and not
1204 * So if it was still runnable (but just not actively
1205 * running right now), it's preempted, and we should
1206 * yield - it could be a while.
1208 if (unlikely(on_rq)) {
1209 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1211 set_current_state(TASK_UNINTERRUPTIBLE);
1212 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1217 * Ahh, all good. It wasn't running, and it wasn't
1218 * runnable, which means that it will never become
1219 * running in the future either. We're all done!
1228 * kick_process - kick a running thread to enter/exit the kernel
1229 * @p: the to-be-kicked thread
1231 * Cause a process which is running on another CPU to enter
1232 * kernel-mode, without any delay. (to get signals handled.)
1234 * NOTE: this function doesn't have to take the runqueue lock,
1235 * because all it wants to ensure is that the remote task enters
1236 * the kernel. If the IPI races and the task has been migrated
1237 * to another CPU then no harm is done and the purpose has been
1240 void kick_process(struct task_struct *p)
1246 if ((cpu != smp_processor_id()) && task_curr(p))
1247 smp_send_reschedule(cpu);
1250 EXPORT_SYMBOL_GPL(kick_process);
1251 #endif /* CONFIG_SMP */
1255 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1257 static int select_fallback_rq(int cpu, struct task_struct *p)
1259 int nid = cpu_to_node(cpu);
1260 const struct cpumask *nodemask = NULL;
1261 enum { cpuset, possible, fail } state = cpuset;
1265 * If the node that the cpu is on has been offlined, cpu_to_node()
1266 * will return -1. There is no cpu on the node, and we should
1267 * select the cpu on the other node.
1270 nodemask = cpumask_of_node(nid);
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu, nodemask) {
1274 if (!cpu_online(dest_cpu))
1276 if (!cpu_active(dest_cpu))
1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1284 /* Any allowed, online CPU? */
1285 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1286 if (!cpu_online(dest_cpu))
1288 if (!cpu_active(dest_cpu))
1295 /* No more Mr. Nice Guy. */
1296 cpuset_cpus_allowed_fallback(p);
1301 do_set_cpus_allowed(p, cpu_possible_mask);
1312 if (state != cpuset) {
1314 * Don't tell them about moving exiting tasks or
1315 * kernel threads (both mm NULL), since they never
1318 if (p->mm && printk_ratelimit()) {
1319 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1320 task_pid_nr(p), p->comm, cpu);
1328 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1331 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1333 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1336 * In order not to call set_task_cpu() on a blocking task we need
1337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1340 * Since this is common to all placement strategies, this lives here.
1342 * [ this allows ->select_task() to simply return task_cpu(p) and
1343 * not worry about this generic constraint ]
1345 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1347 cpu = select_fallback_rq(task_cpu(p), p);
1352 static void update_avg(u64 *avg, u64 sample)
1354 s64 diff = sample - *avg;
1360 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1362 #ifdef CONFIG_SCHEDSTATS
1363 struct rq *rq = this_rq();
1366 int this_cpu = smp_processor_id();
1368 if (cpu == this_cpu) {
1369 schedstat_inc(rq, ttwu_local);
1370 schedstat_inc(p, se.statistics.nr_wakeups_local);
1372 struct sched_domain *sd;
1374 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1376 for_each_domain(this_cpu, sd) {
1377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1378 schedstat_inc(sd, ttwu_wake_remote);
1385 if (wake_flags & WF_MIGRATED)
1386 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1388 #endif /* CONFIG_SMP */
1390 schedstat_inc(rq, ttwu_count);
1391 schedstat_inc(p, se.statistics.nr_wakeups);
1393 if (wake_flags & WF_SYNC)
1394 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1396 #endif /* CONFIG_SCHEDSTATS */
1399 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1401 activate_task(rq, p, en_flags);
1404 /* if a worker is waking up, notify workqueue */
1405 if (p->flags & PF_WQ_WORKER)
1406 wq_worker_waking_up(p, cpu_of(rq));
1410 * Mark the task runnable and perform wakeup-preemption.
1413 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1415 check_preempt_curr(rq, p, wake_flags);
1416 trace_sched_wakeup(p, true);
1418 p->state = TASK_RUNNING;
1420 if (p->sched_class->task_woken)
1421 p->sched_class->task_woken(rq, p);
1423 if (rq->idle_stamp) {
1424 u64 delta = rq_clock(rq) - rq->idle_stamp;
1425 u64 max = 2*rq->max_idle_balance_cost;
1427 update_avg(&rq->avg_idle, delta);
1429 if (rq->avg_idle > max)
1438 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1441 if (p->sched_contributes_to_load)
1442 rq->nr_uninterruptible--;
1445 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1446 ttwu_do_wakeup(rq, p, wake_flags);
1450 * Called in case the task @p isn't fully descheduled from its runqueue,
1451 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1452 * since all we need to do is flip p->state to TASK_RUNNING, since
1453 * the task is still ->on_rq.
1455 static int ttwu_remote(struct task_struct *p, int wake_flags)
1460 rq = __task_rq_lock(p);
1462 /* check_preempt_curr() may use rq clock */
1463 update_rq_clock(rq);
1464 ttwu_do_wakeup(rq, p, wake_flags);
1467 __task_rq_unlock(rq);
1473 static void sched_ttwu_pending(void)
1475 struct rq *rq = this_rq();
1476 struct llist_node *llist = llist_del_all(&rq->wake_list);
1477 struct task_struct *p;
1479 raw_spin_lock(&rq->lock);
1482 p = llist_entry(llist, struct task_struct, wake_entry);
1483 llist = llist_next(llist);
1484 ttwu_do_activate(rq, p, 0);
1487 raw_spin_unlock(&rq->lock);
1490 void scheduler_ipi(void)
1493 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1494 * TIF_NEED_RESCHED remotely (for the first time) will also send
1497 if (tif_need_resched())
1498 set_preempt_need_resched();
1500 if (llist_empty(&this_rq()->wake_list)
1501 && !tick_nohz_full_cpu(smp_processor_id())
1502 && !got_nohz_idle_kick())
1506 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1507 * traditionally all their work was done from the interrupt return
1508 * path. Now that we actually do some work, we need to make sure
1511 * Some archs already do call them, luckily irq_enter/exit nest
1514 * Arguably we should visit all archs and update all handlers,
1515 * however a fair share of IPIs are still resched only so this would
1516 * somewhat pessimize the simple resched case.
1519 tick_nohz_full_check();
1520 sched_ttwu_pending();
1523 * Check if someone kicked us for doing the nohz idle load balance.
1525 if (unlikely(got_nohz_idle_kick())) {
1526 this_rq()->idle_balance = 1;
1527 raise_softirq_irqoff(SCHED_SOFTIRQ);
1532 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1534 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1535 smp_send_reschedule(cpu);
1538 bool cpus_share_cache(int this_cpu, int that_cpu)
1540 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1542 #endif /* CONFIG_SMP */
1544 static void ttwu_queue(struct task_struct *p, int cpu)
1546 struct rq *rq = cpu_rq(cpu);
1548 #if defined(CONFIG_SMP)
1549 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1550 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1551 ttwu_queue_remote(p, cpu);
1556 raw_spin_lock(&rq->lock);
1557 ttwu_do_activate(rq, p, 0);
1558 raw_spin_unlock(&rq->lock);
1562 * try_to_wake_up - wake up a thread
1563 * @p: the thread to be awakened
1564 * @state: the mask of task states that can be woken
1565 * @wake_flags: wake modifier flags (WF_*)
1567 * Put it on the run-queue if it's not already there. The "current"
1568 * thread is always on the run-queue (except when the actual
1569 * re-schedule is in progress), and as such you're allowed to do
1570 * the simpler "current->state = TASK_RUNNING" to mark yourself
1571 * runnable without the overhead of this.
1573 * Return: %true if @p was woken up, %false if it was already running.
1574 * or @state didn't match @p's state.
1577 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1579 unsigned long flags;
1580 int cpu, success = 0;
1583 * If we are going to wake up a thread waiting for CONDITION we
1584 * need to ensure that CONDITION=1 done by the caller can not be
1585 * reordered with p->state check below. This pairs with mb() in
1586 * set_current_state() the waiting thread does.
1588 smp_mb__before_spinlock();
1589 raw_spin_lock_irqsave(&p->pi_lock, flags);
1590 if (!(p->state & state))
1593 success = 1; /* we're going to change ->state */
1596 if (p->on_rq && ttwu_remote(p, wake_flags))
1601 * If the owning (remote) cpu is still in the middle of schedule() with
1602 * this task as prev, wait until its done referencing the task.
1607 * Pairs with the smp_wmb() in finish_lock_switch().
1611 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1612 p->state = TASK_WAKING;
1614 if (p->sched_class->task_waking)
1615 p->sched_class->task_waking(p);
1617 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1618 if (task_cpu(p) != cpu) {
1619 wake_flags |= WF_MIGRATED;
1620 set_task_cpu(p, cpu);
1622 #endif /* CONFIG_SMP */
1626 ttwu_stat(p, cpu, wake_flags);
1628 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1634 * try_to_wake_up_local - try to wake up a local task with rq lock held
1635 * @p: the thread to be awakened
1637 * Put @p on the run-queue if it's not already there. The caller must
1638 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1641 static void try_to_wake_up_local(struct task_struct *p)
1643 struct rq *rq = task_rq(p);
1645 if (WARN_ON_ONCE(rq != this_rq()) ||
1646 WARN_ON_ONCE(p == current))
1649 lockdep_assert_held(&rq->lock);
1651 if (!raw_spin_trylock(&p->pi_lock)) {
1652 raw_spin_unlock(&rq->lock);
1653 raw_spin_lock(&p->pi_lock);
1654 raw_spin_lock(&rq->lock);
1657 if (!(p->state & TASK_NORMAL))
1661 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1663 ttwu_do_wakeup(rq, p, 0);
1664 ttwu_stat(p, smp_processor_id(), 0);
1666 raw_spin_unlock(&p->pi_lock);
1670 * wake_up_process - Wake up a specific process
1671 * @p: The process to be woken up.
1673 * Attempt to wake up the nominated process and move it to the set of runnable
1676 * Return: 1 if the process was woken up, 0 if it was already running.
1678 * It may be assumed that this function implies a write memory barrier before
1679 * changing the task state if and only if any tasks are woken up.
1681 int wake_up_process(struct task_struct *p)
1683 WARN_ON(task_is_stopped_or_traced(p));
1684 return try_to_wake_up(p, TASK_NORMAL, 0);
1686 EXPORT_SYMBOL(wake_up_process);
1688 int wake_up_state(struct task_struct *p, unsigned int state)
1690 return try_to_wake_up(p, state, 0);
1694 * Perform scheduler related setup for a newly forked process p.
1695 * p is forked by current.
1697 * __sched_fork() is basic setup used by init_idle() too:
1699 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1704 p->se.exec_start = 0;
1705 p->se.sum_exec_runtime = 0;
1706 p->se.prev_sum_exec_runtime = 0;
1707 p->se.nr_migrations = 0;
1709 INIT_LIST_HEAD(&p->se.group_node);
1711 #ifdef CONFIG_SCHEDSTATS
1712 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1715 INIT_LIST_HEAD(&p->rt.run_list);
1717 #ifdef CONFIG_PREEMPT_NOTIFIERS
1718 INIT_HLIST_HEAD(&p->preempt_notifiers);
1721 #ifdef CONFIG_NUMA_BALANCING
1722 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1723 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1724 p->mm->numa_next_reset = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_period_reset);
1725 p->mm->numa_scan_seq = 0;
1728 if (clone_flags & CLONE_VM)
1729 p->numa_preferred_nid = current->numa_preferred_nid;
1731 p->numa_preferred_nid = -1;
1733 p->node_stamp = 0ULL;
1734 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1735 p->numa_migrate_seq = 1;
1736 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1737 p->numa_work.next = &p->numa_work;
1738 p->numa_faults = NULL;
1739 p->numa_faults_buffer = NULL;
1741 INIT_LIST_HEAD(&p->numa_entry);
1742 p->numa_group = NULL;
1743 #endif /* CONFIG_NUMA_BALANCING */
1746 #ifdef CONFIG_NUMA_BALANCING
1747 #ifdef CONFIG_SCHED_DEBUG
1748 void set_numabalancing_state(bool enabled)
1751 sched_feat_set("NUMA");
1753 sched_feat_set("NO_NUMA");
1756 __read_mostly bool numabalancing_enabled;
1758 void set_numabalancing_state(bool enabled)
1760 numabalancing_enabled = enabled;
1762 #endif /* CONFIG_SCHED_DEBUG */
1763 #endif /* CONFIG_NUMA_BALANCING */
1766 * fork()/clone()-time setup:
1768 void sched_fork(unsigned long clone_flags, struct task_struct *p)
1770 unsigned long flags;
1771 int cpu = get_cpu();
1773 __sched_fork(clone_flags, p);
1775 * We mark the process as running here. This guarantees that
1776 * nobody will actually run it, and a signal or other external
1777 * event cannot wake it up and insert it on the runqueue either.
1779 p->state = TASK_RUNNING;
1782 * Make sure we do not leak PI boosting priority to the child.
1784 p->prio = current->normal_prio;
1787 * Revert to default priority/policy on fork if requested.
1789 if (unlikely(p->sched_reset_on_fork)) {
1790 if (task_has_rt_policy(p)) {
1791 p->policy = SCHED_NORMAL;
1792 p->static_prio = NICE_TO_PRIO(0);
1794 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1795 p->static_prio = NICE_TO_PRIO(0);
1797 p->prio = p->normal_prio = __normal_prio(p);
1801 * We don't need the reset flag anymore after the fork. It has
1802 * fulfilled its duty:
1804 p->sched_reset_on_fork = 0;
1807 if (!rt_prio(p->prio))
1808 p->sched_class = &fair_sched_class;
1810 if (p->sched_class->task_fork)
1811 p->sched_class->task_fork(p);
1814 * The child is not yet in the pid-hash so no cgroup attach races,
1815 * and the cgroup is pinned to this child due to cgroup_fork()
1816 * is ran before sched_fork().
1818 * Silence PROVE_RCU.
1820 raw_spin_lock_irqsave(&p->pi_lock, flags);
1821 set_task_cpu(p, cpu);
1822 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1824 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1825 if (likely(sched_info_on()))
1826 memset(&p->sched_info, 0, sizeof(p->sched_info));
1828 #if defined(CONFIG_SMP)
1831 init_task_preempt_count(p);
1833 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1840 * wake_up_new_task - wake up a newly created task for the first time.
1842 * This function will do some initial scheduler statistics housekeeping
1843 * that must be done for every newly created context, then puts the task
1844 * on the runqueue and wakes it.
1846 void wake_up_new_task(struct task_struct *p)
1848 unsigned long flags;
1851 raw_spin_lock_irqsave(&p->pi_lock, flags);
1854 * Fork balancing, do it here and not earlier because:
1855 * - cpus_allowed can change in the fork path
1856 * - any previously selected cpu might disappear through hotplug
1858 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1861 /* Initialize new task's runnable average */
1862 init_task_runnable_average(p);
1863 rq = __task_rq_lock(p);
1864 activate_task(rq, p, 0);
1866 trace_sched_wakeup_new(p, true);
1867 check_preempt_curr(rq, p, WF_FORK);
1869 if (p->sched_class->task_woken)
1870 p->sched_class->task_woken(rq, p);
1872 task_rq_unlock(rq, p, &flags);
1875 #ifdef CONFIG_PREEMPT_NOTIFIERS
1878 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1879 * @notifier: notifier struct to register
1881 void preempt_notifier_register(struct preempt_notifier *notifier)
1883 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1885 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1888 * preempt_notifier_unregister - no longer interested in preemption notifications
1889 * @notifier: notifier struct to unregister
1891 * This is safe to call from within a preemption notifier.
1893 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1895 hlist_del(¬ifier->link);
1897 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1899 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1901 struct preempt_notifier *notifier;
1903 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1904 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1908 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1909 struct task_struct *next)
1911 struct preempt_notifier *notifier;
1913 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1914 notifier->ops->sched_out(notifier, next);
1917 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1919 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1924 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1925 struct task_struct *next)
1929 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1932 * prepare_task_switch - prepare to switch tasks
1933 * @rq: the runqueue preparing to switch
1934 * @prev: the current task that is being switched out
1935 * @next: the task we are going to switch to.
1937 * This is called with the rq lock held and interrupts off. It must
1938 * be paired with a subsequent finish_task_switch after the context
1941 * prepare_task_switch sets up locking and calls architecture specific
1945 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1946 struct task_struct *next)
1948 trace_sched_switch(prev, next);
1949 sched_info_switch(rq, prev, next);
1950 perf_event_task_sched_out(prev, next);
1951 fire_sched_out_preempt_notifiers(prev, next);
1952 prepare_lock_switch(rq, next);
1953 prepare_arch_switch(next);
1957 * finish_task_switch - clean up after a task-switch
1958 * @rq: runqueue associated with task-switch
1959 * @prev: the thread we just switched away from.
1961 * finish_task_switch must be called after the context switch, paired
1962 * with a prepare_task_switch call before the context switch.
1963 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1964 * and do any other architecture-specific cleanup actions.
1966 * Note that we may have delayed dropping an mm in context_switch(). If
1967 * so, we finish that here outside of the runqueue lock. (Doing it
1968 * with the lock held can cause deadlocks; see schedule() for
1971 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1972 __releases(rq->lock)
1974 struct mm_struct *mm = rq->prev_mm;
1980 * A task struct has one reference for the use as "current".
1981 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1982 * schedule one last time. The schedule call will never return, and
1983 * the scheduled task must drop that reference.
1984 * The test for TASK_DEAD must occur while the runqueue locks are
1985 * still held, otherwise prev could be scheduled on another cpu, die
1986 * there before we look at prev->state, and then the reference would
1988 * Manfred Spraul <manfred@colorfullife.com>
1990 prev_state = prev->state;
1991 vtime_task_switch(prev);
1992 finish_arch_switch(prev);
1993 perf_event_task_sched_in(prev, current);
1994 finish_lock_switch(rq, prev);
1995 finish_arch_post_lock_switch();
1997 fire_sched_in_preempt_notifiers(current);
2000 if (unlikely(prev_state == TASK_DEAD)) {
2001 task_numa_free(prev);
2004 * Remove function-return probe instances associated with this
2005 * task and put them back on the free list.
2007 kprobe_flush_task(prev);
2008 put_task_struct(prev);
2011 tick_nohz_task_switch(current);
2016 /* assumes rq->lock is held */
2017 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2019 if (prev->sched_class->pre_schedule)
2020 prev->sched_class->pre_schedule(rq, prev);
2023 /* rq->lock is NOT held, but preemption is disabled */
2024 static inline void post_schedule(struct rq *rq)
2026 if (rq->post_schedule) {
2027 unsigned long flags;
2029 raw_spin_lock_irqsave(&rq->lock, flags);
2030 if (rq->curr->sched_class->post_schedule)
2031 rq->curr->sched_class->post_schedule(rq);
2032 raw_spin_unlock_irqrestore(&rq->lock, flags);
2034 rq->post_schedule = 0;
2040 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2044 static inline void post_schedule(struct rq *rq)
2051 * schedule_tail - first thing a freshly forked thread must call.
2052 * @prev: the thread we just switched away from.
2054 asmlinkage void schedule_tail(struct task_struct *prev)
2055 __releases(rq->lock)
2057 struct rq *rq = this_rq();
2059 finish_task_switch(rq, prev);
2062 * FIXME: do we need to worry about rq being invalidated by the
2067 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2068 /* In this case, finish_task_switch does not reenable preemption */
2071 if (current->set_child_tid)
2072 put_user(task_pid_vnr(current), current->set_child_tid);
2076 * context_switch - switch to the new MM and the new
2077 * thread's register state.
2080 context_switch(struct rq *rq, struct task_struct *prev,
2081 struct task_struct *next)
2083 struct mm_struct *mm, *oldmm;
2085 prepare_task_switch(rq, prev, next);
2088 oldmm = prev->active_mm;
2090 * For paravirt, this is coupled with an exit in switch_to to
2091 * combine the page table reload and the switch backend into
2094 arch_start_context_switch(prev);
2097 next->active_mm = oldmm;
2098 atomic_inc(&oldmm->mm_count);
2099 enter_lazy_tlb(oldmm, next);
2101 switch_mm(oldmm, mm, next);
2104 prev->active_mm = NULL;
2105 rq->prev_mm = oldmm;
2108 * Since the runqueue lock will be released by the next
2109 * task (which is an invalid locking op but in the case
2110 * of the scheduler it's an obvious special-case), so we
2111 * do an early lockdep release here:
2113 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2114 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2117 context_tracking_task_switch(prev, next);
2118 /* Here we just switch the register state and the stack. */
2119 switch_to(prev, next, prev);
2123 * this_rq must be evaluated again because prev may have moved
2124 * CPUs since it called schedule(), thus the 'rq' on its stack
2125 * frame will be invalid.
2127 finish_task_switch(this_rq(), prev);
2131 * nr_running and nr_context_switches:
2133 * externally visible scheduler statistics: current number of runnable
2134 * threads, total number of context switches performed since bootup.
2136 unsigned long nr_running(void)
2138 unsigned long i, sum = 0;
2140 for_each_online_cpu(i)
2141 sum += cpu_rq(i)->nr_running;
2146 unsigned long long nr_context_switches(void)
2149 unsigned long long sum = 0;
2151 for_each_possible_cpu(i)
2152 sum += cpu_rq(i)->nr_switches;
2157 unsigned long nr_iowait(void)
2159 unsigned long i, sum = 0;
2161 for_each_possible_cpu(i)
2162 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2167 unsigned long nr_iowait_cpu(int cpu)
2169 struct rq *this = cpu_rq(cpu);
2170 return atomic_read(&this->nr_iowait);
2176 * sched_exec - execve() is a valuable balancing opportunity, because at
2177 * this point the task has the smallest effective memory and cache footprint.
2179 void sched_exec(void)
2181 struct task_struct *p = current;
2182 unsigned long flags;
2185 raw_spin_lock_irqsave(&p->pi_lock, flags);
2186 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2187 if (dest_cpu == smp_processor_id())
2190 if (likely(cpu_active(dest_cpu))) {
2191 struct migration_arg arg = { p, dest_cpu };
2193 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2194 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2198 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2203 DEFINE_PER_CPU(struct kernel_stat, kstat);
2204 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2206 EXPORT_PER_CPU_SYMBOL(kstat);
2207 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2210 * Return any ns on the sched_clock that have not yet been accounted in
2211 * @p in case that task is currently running.
2213 * Called with task_rq_lock() held on @rq.
2215 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2219 if (task_current(rq, p)) {
2220 update_rq_clock(rq);
2221 ns = rq_clock_task(rq) - p->se.exec_start;
2229 unsigned long long task_delta_exec(struct task_struct *p)
2231 unsigned long flags;
2235 rq = task_rq_lock(p, &flags);
2236 ns = do_task_delta_exec(p, rq);
2237 task_rq_unlock(rq, p, &flags);
2243 * Return accounted runtime for the task.
2244 * In case the task is currently running, return the runtime plus current's
2245 * pending runtime that have not been accounted yet.
2247 unsigned long long task_sched_runtime(struct task_struct *p)
2249 unsigned long flags;
2253 rq = task_rq_lock(p, &flags);
2254 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2255 task_rq_unlock(rq, p, &flags);
2261 * This function gets called by the timer code, with HZ frequency.
2262 * We call it with interrupts disabled.
2264 void scheduler_tick(void)
2266 int cpu = smp_processor_id();
2267 struct rq *rq = cpu_rq(cpu);
2268 struct task_struct *curr = rq->curr;
2272 raw_spin_lock(&rq->lock);
2273 update_rq_clock(rq);
2274 curr->sched_class->task_tick(rq, curr, 0);
2275 update_cpu_load_active(rq);
2276 raw_spin_unlock(&rq->lock);
2278 perf_event_task_tick();
2281 rq->idle_balance = idle_cpu(cpu);
2282 trigger_load_balance(rq, cpu);
2284 rq_last_tick_reset(rq);
2287 #ifdef CONFIG_NO_HZ_FULL
2289 * scheduler_tick_max_deferment
2291 * Keep at least one tick per second when a single
2292 * active task is running because the scheduler doesn't
2293 * yet completely support full dynticks environment.
2295 * This makes sure that uptime, CFS vruntime, load
2296 * balancing, etc... continue to move forward, even
2297 * with a very low granularity.
2299 * Return: Maximum deferment in nanoseconds.
2301 u64 scheduler_tick_max_deferment(void)
2303 struct rq *rq = this_rq();
2304 unsigned long next, now = ACCESS_ONCE(jiffies);
2306 next = rq->last_sched_tick + HZ;
2308 if (time_before_eq(next, now))
2311 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2315 notrace unsigned long get_parent_ip(unsigned long addr)
2317 if (in_lock_functions(addr)) {
2318 addr = CALLER_ADDR2;
2319 if (in_lock_functions(addr))
2320 addr = CALLER_ADDR3;
2325 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2326 defined(CONFIG_PREEMPT_TRACER))
2328 void __kprobes preempt_count_add(int val)
2330 #ifdef CONFIG_DEBUG_PREEMPT
2334 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2337 __preempt_count_add(val);
2338 #ifdef CONFIG_DEBUG_PREEMPT
2340 * Spinlock count overflowing soon?
2342 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2345 if (preempt_count() == val)
2346 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2348 EXPORT_SYMBOL(preempt_count_add);
2350 void __kprobes preempt_count_sub(int val)
2352 #ifdef CONFIG_DEBUG_PREEMPT
2356 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2359 * Is the spinlock portion underflowing?
2361 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2362 !(preempt_count() & PREEMPT_MASK)))
2366 if (preempt_count() == val)
2367 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2368 __preempt_count_sub(val);
2370 EXPORT_SYMBOL(preempt_count_sub);
2375 * Print scheduling while atomic bug:
2377 static noinline void __schedule_bug(struct task_struct *prev)
2379 if (oops_in_progress)
2382 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2383 prev->comm, prev->pid, preempt_count());
2385 debug_show_held_locks(prev);
2387 if (irqs_disabled())
2388 print_irqtrace_events(prev);
2390 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2394 * Various schedule()-time debugging checks and statistics:
2396 static inline void schedule_debug(struct task_struct *prev)
2399 * Test if we are atomic. Since do_exit() needs to call into
2400 * schedule() atomically, we ignore that path for now.
2401 * Otherwise, whine if we are scheduling when we should not be.
2403 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2404 __schedule_bug(prev);
2407 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2409 schedstat_inc(this_rq(), sched_count);
2412 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2414 if (prev->on_rq || rq->skip_clock_update < 0)
2415 update_rq_clock(rq);
2416 prev->sched_class->put_prev_task(rq, prev);
2420 * Pick up the highest-prio task:
2422 static inline struct task_struct *
2423 pick_next_task(struct rq *rq)
2425 const struct sched_class *class;
2426 struct task_struct *p;
2429 * Optimization: we know that if all tasks are in
2430 * the fair class we can call that function directly:
2432 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2433 p = fair_sched_class.pick_next_task(rq);
2438 for_each_class(class) {
2439 p = class->pick_next_task(rq);
2444 BUG(); /* the idle class will always have a runnable task */
2448 * __schedule() is the main scheduler function.
2450 * The main means of driving the scheduler and thus entering this function are:
2452 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2454 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2455 * paths. For example, see arch/x86/entry_64.S.
2457 * To drive preemption between tasks, the scheduler sets the flag in timer
2458 * interrupt handler scheduler_tick().
2460 * 3. Wakeups don't really cause entry into schedule(). They add a
2461 * task to the run-queue and that's it.
2463 * Now, if the new task added to the run-queue preempts the current
2464 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2465 * called on the nearest possible occasion:
2467 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2469 * - in syscall or exception context, at the next outmost
2470 * preempt_enable(). (this might be as soon as the wake_up()'s
2473 * - in IRQ context, return from interrupt-handler to
2474 * preemptible context
2476 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2479 * - cond_resched() call
2480 * - explicit schedule() call
2481 * - return from syscall or exception to user-space
2482 * - return from interrupt-handler to user-space
2484 static void __sched __schedule(void)
2486 struct task_struct *prev, *next;
2487 unsigned long *switch_count;
2493 cpu = smp_processor_id();
2495 rcu_note_context_switch(cpu);
2498 schedule_debug(prev);
2500 if (sched_feat(HRTICK))
2504 * Make sure that signal_pending_state()->signal_pending() below
2505 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2506 * done by the caller to avoid the race with signal_wake_up().
2508 smp_mb__before_spinlock();
2509 raw_spin_lock_irq(&rq->lock);
2511 switch_count = &prev->nivcsw;
2512 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2513 if (unlikely(signal_pending_state(prev->state, prev))) {
2514 prev->state = TASK_RUNNING;
2516 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2520 * If a worker went to sleep, notify and ask workqueue
2521 * whether it wants to wake up a task to maintain
2524 if (prev->flags & PF_WQ_WORKER) {
2525 struct task_struct *to_wakeup;
2527 to_wakeup = wq_worker_sleeping(prev, cpu);
2529 try_to_wake_up_local(to_wakeup);
2532 switch_count = &prev->nvcsw;
2535 pre_schedule(rq, prev);
2537 if (unlikely(!rq->nr_running))
2538 idle_balance(cpu, rq);
2540 put_prev_task(rq, prev);
2541 next = pick_next_task(rq);
2542 clear_tsk_need_resched(prev);
2543 clear_preempt_need_resched();
2544 rq->skip_clock_update = 0;
2546 if (likely(prev != next)) {
2551 context_switch(rq, prev, next); /* unlocks the rq */
2553 * The context switch have flipped the stack from under us
2554 * and restored the local variables which were saved when
2555 * this task called schedule() in the past. prev == current
2556 * is still correct, but it can be moved to another cpu/rq.
2558 cpu = smp_processor_id();
2561 raw_spin_unlock_irq(&rq->lock);
2565 sched_preempt_enable_no_resched();
2570 static inline void sched_submit_work(struct task_struct *tsk)
2572 if (!tsk->state || tsk_is_pi_blocked(tsk))
2575 * If we are going to sleep and we have plugged IO queued,
2576 * make sure to submit it to avoid deadlocks.
2578 if (blk_needs_flush_plug(tsk))
2579 blk_schedule_flush_plug(tsk);
2582 asmlinkage void __sched schedule(void)
2584 struct task_struct *tsk = current;
2586 sched_submit_work(tsk);
2589 EXPORT_SYMBOL(schedule);
2591 #ifdef CONFIG_CONTEXT_TRACKING
2592 asmlinkage void __sched schedule_user(void)
2595 * If we come here after a random call to set_need_resched(),
2596 * or we have been woken up remotely but the IPI has not yet arrived,
2597 * we haven't yet exited the RCU idle mode. Do it here manually until
2598 * we find a better solution.
2607 * schedule_preempt_disabled - called with preemption disabled
2609 * Returns with preemption disabled. Note: preempt_count must be 1
2611 void __sched schedule_preempt_disabled(void)
2613 sched_preempt_enable_no_resched();
2618 #ifdef CONFIG_PREEMPT
2620 * this is the entry point to schedule() from in-kernel preemption
2621 * off of preempt_enable. Kernel preemptions off return from interrupt
2622 * occur there and call schedule directly.
2624 asmlinkage void __sched notrace preempt_schedule(void)
2627 * If there is a non-zero preempt_count or interrupts are disabled,
2628 * we do not want to preempt the current task. Just return..
2630 if (likely(!preemptible()))
2634 __preempt_count_add(PREEMPT_ACTIVE);
2636 __preempt_count_sub(PREEMPT_ACTIVE);
2639 * Check again in case we missed a preemption opportunity
2640 * between schedule and now.
2643 } while (need_resched());
2645 EXPORT_SYMBOL(preempt_schedule);
2648 * this is the entry point to schedule() from kernel preemption
2649 * off of irq context.
2650 * Note, that this is called and return with irqs disabled. This will
2651 * protect us against recursive calling from irq.
2653 asmlinkage void __sched preempt_schedule_irq(void)
2655 enum ctx_state prev_state;
2657 /* Catch callers which need to be fixed */
2658 BUG_ON(preempt_count() || !irqs_disabled());
2660 prev_state = exception_enter();
2663 __preempt_count_add(PREEMPT_ACTIVE);
2666 local_irq_disable();
2667 __preempt_count_sub(PREEMPT_ACTIVE);
2670 * Check again in case we missed a preemption opportunity
2671 * between schedule and now.
2674 } while (need_resched());
2676 exception_exit(prev_state);
2679 #endif /* CONFIG_PREEMPT */
2681 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2684 return try_to_wake_up(curr->private, mode, wake_flags);
2686 EXPORT_SYMBOL(default_wake_function);
2689 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2690 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2691 * number) then we wake all the non-exclusive tasks and one exclusive task.
2693 * There are circumstances in which we can try to wake a task which has already
2694 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2695 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2697 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2698 int nr_exclusive, int wake_flags, void *key)
2700 wait_queue_t *curr, *next;
2702 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2703 unsigned flags = curr->flags;
2705 if (curr->func(curr, mode, wake_flags, key) &&
2706 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2712 * __wake_up - wake up threads blocked on a waitqueue.
2714 * @mode: which threads
2715 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2716 * @key: is directly passed to the wakeup function
2718 * It may be assumed that this function implies a write memory barrier before
2719 * changing the task state if and only if any tasks are woken up.
2721 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2722 int nr_exclusive, void *key)
2724 unsigned long flags;
2726 spin_lock_irqsave(&q->lock, flags);
2727 __wake_up_common(q, mode, nr_exclusive, 0, key);
2728 spin_unlock_irqrestore(&q->lock, flags);
2730 EXPORT_SYMBOL(__wake_up);
2733 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2735 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2737 __wake_up_common(q, mode, nr, 0, NULL);
2739 EXPORT_SYMBOL_GPL(__wake_up_locked);
2741 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2743 __wake_up_common(q, mode, 1, 0, key);
2745 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2748 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2750 * @mode: which threads
2751 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2752 * @key: opaque value to be passed to wakeup targets
2754 * The sync wakeup differs that the waker knows that it will schedule
2755 * away soon, so while the target thread will be woken up, it will not
2756 * be migrated to another CPU - ie. the two threads are 'synchronized'
2757 * with each other. This can prevent needless bouncing between CPUs.
2759 * On UP it can prevent extra preemption.
2761 * It may be assumed that this function implies a write memory barrier before
2762 * changing the task state if and only if any tasks are woken up.
2764 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2765 int nr_exclusive, void *key)
2767 unsigned long flags;
2768 int wake_flags = WF_SYNC;
2773 if (unlikely(nr_exclusive != 1))
2776 spin_lock_irqsave(&q->lock, flags);
2777 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2778 spin_unlock_irqrestore(&q->lock, flags);
2780 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2783 * __wake_up_sync - see __wake_up_sync_key()
2785 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2787 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2789 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2792 * complete: - signals a single thread waiting on this completion
2793 * @x: holds the state of this particular completion
2795 * This will wake up a single thread waiting on this completion. Threads will be
2796 * awakened in the same order in which they were queued.
2798 * See also complete_all(), wait_for_completion() and related routines.
2800 * It may be assumed that this function implies a write memory barrier before
2801 * changing the task state if and only if any tasks are woken up.
2803 void complete(struct completion *x)
2805 unsigned long flags;
2807 spin_lock_irqsave(&x->wait.lock, flags);
2809 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2810 spin_unlock_irqrestore(&x->wait.lock, flags);
2812 EXPORT_SYMBOL(complete);
2815 * complete_all: - signals all threads waiting on this completion
2816 * @x: holds the state of this particular completion
2818 * This will wake up all threads waiting on this particular completion event.
2820 * It may be assumed that this function implies a write memory barrier before
2821 * changing the task state if and only if any tasks are woken up.
2823 void complete_all(struct completion *x)
2825 unsigned long flags;
2827 spin_lock_irqsave(&x->wait.lock, flags);
2828 x->done += UINT_MAX/2;
2829 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2830 spin_unlock_irqrestore(&x->wait.lock, flags);
2832 EXPORT_SYMBOL(complete_all);
2834 static inline long __sched
2835 do_wait_for_common(struct completion *x,
2836 long (*action)(long), long timeout, int state)
2839 DECLARE_WAITQUEUE(wait, current);
2841 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2843 if (signal_pending_state(state, current)) {
2844 timeout = -ERESTARTSYS;
2847 __set_current_state(state);
2848 spin_unlock_irq(&x->wait.lock);
2849 timeout = action(timeout);
2850 spin_lock_irq(&x->wait.lock);
2851 } while (!x->done && timeout);
2852 __remove_wait_queue(&x->wait, &wait);
2857 return timeout ?: 1;
2860 static inline long __sched
2861 __wait_for_common(struct completion *x,
2862 long (*action)(long), long timeout, int state)
2866 spin_lock_irq(&x->wait.lock);
2867 timeout = do_wait_for_common(x, action, timeout, state);
2868 spin_unlock_irq(&x->wait.lock);
2873 wait_for_common(struct completion *x, long timeout, int state)
2875 return __wait_for_common(x, schedule_timeout, timeout, state);
2879 wait_for_common_io(struct completion *x, long timeout, int state)
2881 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2885 * wait_for_completion: - waits for completion of a task
2886 * @x: holds the state of this particular completion
2888 * This waits to be signaled for completion of a specific task. It is NOT
2889 * interruptible and there is no timeout.
2891 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2892 * and interrupt capability. Also see complete().
2894 void __sched wait_for_completion(struct completion *x)
2896 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2898 EXPORT_SYMBOL(wait_for_completion);
2901 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2902 * @x: holds the state of this particular completion
2903 * @timeout: timeout value in jiffies
2905 * This waits for either a completion of a specific task to be signaled or for a
2906 * specified timeout to expire. The timeout is in jiffies. It is not
2909 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2910 * till timeout) if completed.
2912 unsigned long __sched
2913 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2915 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2917 EXPORT_SYMBOL(wait_for_completion_timeout);
2920 * wait_for_completion_io: - waits for completion of a task
2921 * @x: holds the state of this particular completion
2923 * This waits to be signaled for completion of a specific task. It is NOT
2924 * interruptible and there is no timeout. The caller is accounted as waiting
2927 void __sched wait_for_completion_io(struct completion *x)
2929 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2931 EXPORT_SYMBOL(wait_for_completion_io);
2934 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2935 * @x: holds the state of this particular completion
2936 * @timeout: timeout value in jiffies
2938 * This waits for either a completion of a specific task to be signaled or for a
2939 * specified timeout to expire. The timeout is in jiffies. It is not
2940 * interruptible. The caller is accounted as waiting for IO.
2942 * Return: 0 if timed out, and positive (at least 1, or number of jiffies left
2943 * till timeout) if completed.
2945 unsigned long __sched
2946 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2948 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2950 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2953 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2954 * @x: holds the state of this particular completion
2956 * This waits for completion of a specific task to be signaled. It is
2959 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2961 int __sched wait_for_completion_interruptible(struct completion *x)
2963 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2964 if (t == -ERESTARTSYS)
2968 EXPORT_SYMBOL(wait_for_completion_interruptible);
2971 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2972 * @x: holds the state of this particular completion
2973 * @timeout: timeout value in jiffies
2975 * This waits for either a completion of a specific task to be signaled or for a
2976 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2978 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
2979 * or number of jiffies left till timeout) if completed.
2982 wait_for_completion_interruptible_timeout(struct completion *x,
2983 unsigned long timeout)
2985 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2987 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2990 * wait_for_completion_killable: - waits for completion of a task (killable)
2991 * @x: holds the state of this particular completion
2993 * This waits to be signaled for completion of a specific task. It can be
2994 * interrupted by a kill signal.
2996 * Return: -ERESTARTSYS if interrupted, 0 if completed.
2998 int __sched wait_for_completion_killable(struct completion *x)
3000 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3001 if (t == -ERESTARTSYS)
3005 EXPORT_SYMBOL(wait_for_completion_killable);
3008 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3009 * @x: holds the state of this particular completion
3010 * @timeout: timeout value in jiffies
3012 * This waits for either a completion of a specific task to be
3013 * signaled or for a specified timeout to expire. It can be
3014 * interrupted by a kill signal. The timeout is in jiffies.
3016 * Return: -ERESTARTSYS if interrupted, 0 if timed out, positive (at least 1,
3017 * or number of jiffies left till timeout) if completed.
3020 wait_for_completion_killable_timeout(struct completion *x,
3021 unsigned long timeout)
3023 return wait_for_common(x, timeout, TASK_KILLABLE);
3025 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3028 * try_wait_for_completion - try to decrement a completion without blocking
3029 * @x: completion structure
3031 * Return: 0 if a decrement cannot be done without blocking
3032 * 1 if a decrement succeeded.
3034 * If a completion is being used as a counting completion,
3035 * attempt to decrement the counter without blocking. This
3036 * enables us to avoid waiting if the resource the completion
3037 * is protecting is not available.
3039 bool try_wait_for_completion(struct completion *x)
3041 unsigned long flags;
3044 spin_lock_irqsave(&x->wait.lock, flags);
3049 spin_unlock_irqrestore(&x->wait.lock, flags);
3052 EXPORT_SYMBOL(try_wait_for_completion);
3055 * completion_done - Test to see if a completion has any waiters
3056 * @x: completion structure
3058 * Return: 0 if there are waiters (wait_for_completion() in progress)
3059 * 1 if there are no waiters.
3062 bool completion_done(struct completion *x)
3064 unsigned long flags;
3067 spin_lock_irqsave(&x->wait.lock, flags);
3070 spin_unlock_irqrestore(&x->wait.lock, flags);
3073 EXPORT_SYMBOL(completion_done);
3076 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3078 unsigned long flags;
3081 init_waitqueue_entry(&wait, current);
3083 __set_current_state(state);
3085 spin_lock_irqsave(&q->lock, flags);
3086 __add_wait_queue(q, &wait);
3087 spin_unlock(&q->lock);
3088 timeout = schedule_timeout(timeout);
3089 spin_lock_irq(&q->lock);
3090 __remove_wait_queue(q, &wait);
3091 spin_unlock_irqrestore(&q->lock, flags);
3096 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3098 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3100 EXPORT_SYMBOL(interruptible_sleep_on);
3103 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3105 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3107 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3109 void __sched sleep_on(wait_queue_head_t *q)
3111 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3113 EXPORT_SYMBOL(sleep_on);
3115 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3117 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3119 EXPORT_SYMBOL(sleep_on_timeout);
3121 #ifdef CONFIG_RT_MUTEXES
3124 * rt_mutex_setprio - set the current priority of a task
3126 * @prio: prio value (kernel-internal form)
3128 * This function changes the 'effective' priority of a task. It does
3129 * not touch ->normal_prio like __setscheduler().
3131 * Used by the rt_mutex code to implement priority inheritance logic.
3133 void rt_mutex_setprio(struct task_struct *p, int prio)
3135 int oldprio, on_rq, running;
3137 const struct sched_class *prev_class;
3139 BUG_ON(prio < 0 || prio > MAX_PRIO);
3141 rq = __task_rq_lock(p);
3144 * Idle task boosting is a nono in general. There is one
3145 * exception, when PREEMPT_RT and NOHZ is active:
3147 * The idle task calls get_next_timer_interrupt() and holds
3148 * the timer wheel base->lock on the CPU and another CPU wants
3149 * to access the timer (probably to cancel it). We can safely
3150 * ignore the boosting request, as the idle CPU runs this code
3151 * with interrupts disabled and will complete the lock
3152 * protected section without being interrupted. So there is no
3153 * real need to boost.
3155 if (unlikely(p == rq->idle)) {
3156 WARN_ON(p != rq->curr);
3157 WARN_ON(p->pi_blocked_on);
3161 trace_sched_pi_setprio(p, prio);
3163 prev_class = p->sched_class;
3165 running = task_current(rq, p);
3167 dequeue_task(rq, p, 0);
3169 p->sched_class->put_prev_task(rq, p);
3172 p->sched_class = &rt_sched_class;
3174 p->sched_class = &fair_sched_class;
3179 p->sched_class->set_curr_task(rq);
3181 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3183 check_class_changed(rq, p, prev_class, oldprio);
3185 __task_rq_unlock(rq);
3188 void set_user_nice(struct task_struct *p, long nice)
3190 int old_prio, delta, on_rq;
3191 unsigned long flags;
3194 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3197 * We have to be careful, if called from sys_setpriority(),
3198 * the task might be in the middle of scheduling on another CPU.
3200 rq = task_rq_lock(p, &flags);
3202 * The RT priorities are set via sched_setscheduler(), but we still
3203 * allow the 'normal' nice value to be set - but as expected
3204 * it wont have any effect on scheduling until the task is
3205 * SCHED_FIFO/SCHED_RR:
3207 if (task_has_rt_policy(p)) {
3208 p->static_prio = NICE_TO_PRIO(nice);
3213 dequeue_task(rq, p, 0);
3215 p->static_prio = NICE_TO_PRIO(nice);
3218 p->prio = effective_prio(p);
3219 delta = p->prio - old_prio;
3222 enqueue_task(rq, p, 0);
3224 * If the task increased its priority or is running and
3225 * lowered its priority, then reschedule its CPU:
3227 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3228 resched_task(rq->curr);
3231 task_rq_unlock(rq, p, &flags);
3233 EXPORT_SYMBOL(set_user_nice);
3236 * can_nice - check if a task can reduce its nice value
3240 int can_nice(const struct task_struct *p, const int nice)
3242 /* convert nice value [19,-20] to rlimit style value [1,40] */
3243 int nice_rlim = 20 - nice;
3245 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3246 capable(CAP_SYS_NICE));
3249 #ifdef __ARCH_WANT_SYS_NICE
3252 * sys_nice - change the priority of the current process.
3253 * @increment: priority increment
3255 * sys_setpriority is a more generic, but much slower function that
3256 * does similar things.
3258 SYSCALL_DEFINE1(nice, int, increment)
3263 * Setpriority might change our priority at the same moment.
3264 * We don't have to worry. Conceptually one call occurs first
3265 * and we have a single winner.
3267 if (increment < -40)
3272 nice = TASK_NICE(current) + increment;
3278 if (increment < 0 && !can_nice(current, nice))
3281 retval = security_task_setnice(current, nice);
3285 set_user_nice(current, nice);
3292 * task_prio - return the priority value of a given task.
3293 * @p: the task in question.
3295 * Return: The priority value as seen by users in /proc.
3296 * RT tasks are offset by -200. Normal tasks are centered
3297 * around 0, value goes from -16 to +15.
3299 int task_prio(const struct task_struct *p)
3301 return p->prio - MAX_RT_PRIO;
3305 * task_nice - return the nice value of a given task.
3306 * @p: the task in question.
3308 * Return: The nice value [ -20 ... 0 ... 19 ].
3310 int task_nice(const struct task_struct *p)
3312 return TASK_NICE(p);
3314 EXPORT_SYMBOL(task_nice);
3317 * idle_cpu - is a given cpu idle currently?
3318 * @cpu: the processor in question.
3320 * Return: 1 if the CPU is currently idle. 0 otherwise.
3322 int idle_cpu(int cpu)
3324 struct rq *rq = cpu_rq(cpu);
3326 if (rq->curr != rq->idle)
3333 if (!llist_empty(&rq->wake_list))
3341 * idle_task - return the idle task for a given cpu.
3342 * @cpu: the processor in question.
3344 * Return: The idle task for the cpu @cpu.
3346 struct task_struct *idle_task(int cpu)
3348 return cpu_rq(cpu)->idle;
3352 * find_process_by_pid - find a process with a matching PID value.
3353 * @pid: the pid in question.
3355 * The task of @pid, if found. %NULL otherwise.
3357 static struct task_struct *find_process_by_pid(pid_t pid)
3359 return pid ? find_task_by_vpid(pid) : current;
3362 /* Actually do priority change: must hold rq lock. */
3364 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3367 p->rt_priority = prio;
3368 p->normal_prio = normal_prio(p);
3369 /* we are holding p->pi_lock already */
3370 p->prio = rt_mutex_getprio(p);
3371 if (rt_prio(p->prio))
3372 p->sched_class = &rt_sched_class;
3374 p->sched_class = &fair_sched_class;
3379 * check the target process has a UID that matches the current process's
3381 static bool check_same_owner(struct task_struct *p)
3383 const struct cred *cred = current_cred(), *pcred;
3387 pcred = __task_cred(p);
3388 match = (uid_eq(cred->euid, pcred->euid) ||
3389 uid_eq(cred->euid, pcred->uid));
3394 static int __sched_setscheduler(struct task_struct *p, int policy,
3395 const struct sched_param *param, bool user)
3397 int retval, oldprio, oldpolicy = -1, on_rq, running;
3398 unsigned long flags;
3399 const struct sched_class *prev_class;
3403 /* may grab non-irq protected spin_locks */
3404 BUG_ON(in_interrupt());
3406 /* double check policy once rq lock held */
3408 reset_on_fork = p->sched_reset_on_fork;
3409 policy = oldpolicy = p->policy;
3411 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3412 policy &= ~SCHED_RESET_ON_FORK;
3414 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3415 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3416 policy != SCHED_IDLE)
3421 * Valid priorities for SCHED_FIFO and SCHED_RR are
3422 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3423 * SCHED_BATCH and SCHED_IDLE is 0.
3425 if (param->sched_priority < 0 ||
3426 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3427 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3429 if (rt_policy(policy) != (param->sched_priority != 0))
3433 * Allow unprivileged RT tasks to decrease priority:
3435 if (user && !capable(CAP_SYS_NICE)) {
3436 if (rt_policy(policy)) {
3437 unsigned long rlim_rtprio =
3438 task_rlimit(p, RLIMIT_RTPRIO);
3440 /* can't set/change the rt policy */
3441 if (policy != p->policy && !rlim_rtprio)
3444 /* can't increase priority */
3445 if (param->sched_priority > p->rt_priority &&
3446 param->sched_priority > rlim_rtprio)
3451 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3452 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3454 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3455 if (!can_nice(p, TASK_NICE(p)))
3459 /* can't change other user's priorities */
3460 if (!check_same_owner(p))
3463 /* Normal users shall not reset the sched_reset_on_fork flag */
3464 if (p->sched_reset_on_fork && !reset_on_fork)
3469 retval = security_task_setscheduler(p);
3475 * make sure no PI-waiters arrive (or leave) while we are
3476 * changing the priority of the task:
3478 * To be able to change p->policy safely, the appropriate
3479 * runqueue lock must be held.
3481 rq = task_rq_lock(p, &flags);
3484 * Changing the policy of the stop threads its a very bad idea
3486 if (p == rq->stop) {
3487 task_rq_unlock(rq, p, &flags);
3492 * If not changing anything there's no need to proceed further:
3494 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3495 param->sched_priority == p->rt_priority))) {
3496 task_rq_unlock(rq, p, &flags);
3500 #ifdef CONFIG_RT_GROUP_SCHED
3503 * Do not allow realtime tasks into groups that have no runtime
3506 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3507 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3508 !task_group_is_autogroup(task_group(p))) {
3509 task_rq_unlock(rq, p, &flags);
3515 /* recheck policy now with rq lock held */
3516 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3517 policy = oldpolicy = -1;
3518 task_rq_unlock(rq, p, &flags);
3522 running = task_current(rq, p);
3524 dequeue_task(rq, p, 0);
3526 p->sched_class->put_prev_task(rq, p);
3528 p->sched_reset_on_fork = reset_on_fork;
3531 prev_class = p->sched_class;
3532 __setscheduler(rq, p, policy, param->sched_priority);
3535 p->sched_class->set_curr_task(rq);
3537 enqueue_task(rq, p, 0);
3539 check_class_changed(rq, p, prev_class, oldprio);
3540 task_rq_unlock(rq, p, &flags);
3542 rt_mutex_adjust_pi(p);
3548 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3549 * @p: the task in question.
3550 * @policy: new policy.
3551 * @param: structure containing the new RT priority.
3553 * Return: 0 on success. An error code otherwise.
3555 * NOTE that the task may be already dead.
3557 int sched_setscheduler(struct task_struct *p, int policy,
3558 const struct sched_param *param)
3560 return __sched_setscheduler(p, policy, param, true);
3562 EXPORT_SYMBOL_GPL(sched_setscheduler);
3565 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3566 * @p: the task in question.
3567 * @policy: new policy.
3568 * @param: structure containing the new RT priority.
3570 * Just like sched_setscheduler, only don't bother checking if the
3571 * current context has permission. For example, this is needed in
3572 * stop_machine(): we create temporary high priority worker threads,
3573 * but our caller might not have that capability.
3575 * Return: 0 on success. An error code otherwise.
3577 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3578 const struct sched_param *param)
3580 return __sched_setscheduler(p, policy, param, false);
3584 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3586 struct sched_param lparam;
3587 struct task_struct *p;
3590 if (!param || pid < 0)
3592 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3597 p = find_process_by_pid(pid);
3599 retval = sched_setscheduler(p, policy, &lparam);
3606 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3607 * @pid: the pid in question.
3608 * @policy: new policy.
3609 * @param: structure containing the new RT priority.
3611 * Return: 0 on success. An error code otherwise.
3613 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3614 struct sched_param __user *, param)
3616 /* negative values for policy are not valid */
3620 return do_sched_setscheduler(pid, policy, param);
3624 * sys_sched_setparam - set/change the RT priority of a thread
3625 * @pid: the pid in question.
3626 * @param: structure containing the new RT priority.
3628 * Return: 0 on success. An error code otherwise.
3630 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3632 return do_sched_setscheduler(pid, -1, param);
3636 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3637 * @pid: the pid in question.
3639 * Return: On success, the policy of the thread. Otherwise, a negative error
3642 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3644 struct task_struct *p;
3652 p = find_process_by_pid(pid);
3654 retval = security_task_getscheduler(p);
3657 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3664 * sys_sched_getparam - get the RT priority of a thread
3665 * @pid: the pid in question.
3666 * @param: structure containing the RT priority.
3668 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3671 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3673 struct sched_param lp;
3674 struct task_struct *p;
3677 if (!param || pid < 0)
3681 p = find_process_by_pid(pid);
3686 retval = security_task_getscheduler(p);
3690 lp.sched_priority = p->rt_priority;
3694 * This one might sleep, we cannot do it with a spinlock held ...
3696 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3705 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3707 cpumask_var_t cpus_allowed, new_mask;
3708 struct task_struct *p;
3714 p = find_process_by_pid(pid);
3721 /* Prevent p going away */
3725 if (p->flags & PF_NO_SETAFFINITY) {
3729 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3733 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3735 goto out_free_cpus_allowed;
3738 if (!check_same_owner(p)) {
3740 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3747 retval = security_task_setscheduler(p);
3751 cpuset_cpus_allowed(p, cpus_allowed);
3752 cpumask_and(new_mask, in_mask, cpus_allowed);
3754 retval = set_cpus_allowed_ptr(p, new_mask);
3757 cpuset_cpus_allowed(p, cpus_allowed);
3758 if (!cpumask_subset(new_mask, cpus_allowed)) {
3760 * We must have raced with a concurrent cpuset
3761 * update. Just reset the cpus_allowed to the
3762 * cpuset's cpus_allowed
3764 cpumask_copy(new_mask, cpus_allowed);
3769 free_cpumask_var(new_mask);
3770 out_free_cpus_allowed:
3771 free_cpumask_var(cpus_allowed);
3778 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3779 struct cpumask *new_mask)
3781 if (len < cpumask_size())
3782 cpumask_clear(new_mask);
3783 else if (len > cpumask_size())
3784 len = cpumask_size();
3786 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3790 * sys_sched_setaffinity - set the cpu affinity of a process
3791 * @pid: pid of the process
3792 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3793 * @user_mask_ptr: user-space pointer to the new cpu mask
3795 * Return: 0 on success. An error code otherwise.
3797 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3798 unsigned long __user *, user_mask_ptr)
3800 cpumask_var_t new_mask;
3803 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3806 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3808 retval = sched_setaffinity(pid, new_mask);
3809 free_cpumask_var(new_mask);
3813 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3815 struct task_struct *p;
3816 unsigned long flags;
3823 p = find_process_by_pid(pid);
3827 retval = security_task_getscheduler(p);
3831 raw_spin_lock_irqsave(&p->pi_lock, flags);
3832 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3833 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3843 * sys_sched_getaffinity - get the cpu affinity of a process
3844 * @pid: pid of the process
3845 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3846 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3848 * Return: 0 on success. An error code otherwise.
3850 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3851 unsigned long __user *, user_mask_ptr)
3856 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3858 if (len & (sizeof(unsigned long)-1))
3861 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3864 ret = sched_getaffinity(pid, mask);
3866 size_t retlen = min_t(size_t, len, cpumask_size());
3868 if (copy_to_user(user_mask_ptr, mask, retlen))
3873 free_cpumask_var(mask);
3879 * sys_sched_yield - yield the current processor to other threads.
3881 * This function yields the current CPU to other tasks. If there are no
3882 * other threads running on this CPU then this function will return.
3886 SYSCALL_DEFINE0(sched_yield)
3888 struct rq *rq = this_rq_lock();
3890 schedstat_inc(rq, yld_count);
3891 current->sched_class->yield_task(rq);
3894 * Since we are going to call schedule() anyway, there's
3895 * no need to preempt or enable interrupts:
3897 __release(rq->lock);
3898 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3899 do_raw_spin_unlock(&rq->lock);
3900 sched_preempt_enable_no_resched();
3907 static void __cond_resched(void)
3909 __preempt_count_add(PREEMPT_ACTIVE);
3911 __preempt_count_sub(PREEMPT_ACTIVE);
3914 int __sched _cond_resched(void)
3916 if (should_resched()) {
3922 EXPORT_SYMBOL(_cond_resched);
3925 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3926 * call schedule, and on return reacquire the lock.
3928 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3929 * operations here to prevent schedule() from being called twice (once via
3930 * spin_unlock(), once by hand).
3932 int __cond_resched_lock(spinlock_t *lock)
3934 int resched = should_resched();
3937 lockdep_assert_held(lock);
3939 if (spin_needbreak(lock) || resched) {
3950 EXPORT_SYMBOL(__cond_resched_lock);
3952 int __sched __cond_resched_softirq(void)
3954 BUG_ON(!in_softirq());
3956 if (should_resched()) {
3964 EXPORT_SYMBOL(__cond_resched_softirq);
3967 * yield - yield the current processor to other threads.
3969 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3971 * The scheduler is at all times free to pick the calling task as the most
3972 * eligible task to run, if removing the yield() call from your code breaks
3973 * it, its already broken.
3975 * Typical broken usage is:
3980 * where one assumes that yield() will let 'the other' process run that will
3981 * make event true. If the current task is a SCHED_FIFO task that will never
3982 * happen. Never use yield() as a progress guarantee!!
3984 * If you want to use yield() to wait for something, use wait_event().
3985 * If you want to use yield() to be 'nice' for others, use cond_resched().
3986 * If you still want to use yield(), do not!
3988 void __sched yield(void)
3990 set_current_state(TASK_RUNNING);
3993 EXPORT_SYMBOL(yield);
3996 * yield_to - yield the current processor to another thread in
3997 * your thread group, or accelerate that thread toward the
3998 * processor it's on.
4000 * @preempt: whether task preemption is allowed or not
4002 * It's the caller's job to ensure that the target task struct
4003 * can't go away on us before we can do any checks.
4006 * true (>0) if we indeed boosted the target task.
4007 * false (0) if we failed to boost the target.
4008 * -ESRCH if there's no task to yield to.
4010 bool __sched yield_to(struct task_struct *p, bool preempt)
4012 struct task_struct *curr = current;
4013 struct rq *rq, *p_rq;
4014 unsigned long flags;
4017 local_irq_save(flags);
4023 * If we're the only runnable task on the rq and target rq also
4024 * has only one task, there's absolutely no point in yielding.
4026 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4031 double_rq_lock(rq, p_rq);
4032 while (task_rq(p) != p_rq) {
4033 double_rq_unlock(rq, p_rq);
4037 if (!curr->sched_class->yield_to_task)
4040 if (curr->sched_class != p->sched_class)
4043 if (task_running(p_rq, p) || p->state)
4046 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4048 schedstat_inc(rq, yld_count);
4050 * Make p's CPU reschedule; pick_next_entity takes care of
4053 if (preempt && rq != p_rq)
4054 resched_task(p_rq->curr);
4058 double_rq_unlock(rq, p_rq);
4060 local_irq_restore(flags);
4067 EXPORT_SYMBOL_GPL(yield_to);
4070 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4071 * that process accounting knows that this is a task in IO wait state.
4073 void __sched io_schedule(void)
4075 struct rq *rq = raw_rq();
4077 delayacct_blkio_start();
4078 atomic_inc(&rq->nr_iowait);
4079 blk_flush_plug(current);
4080 current->in_iowait = 1;
4082 current->in_iowait = 0;
4083 atomic_dec(&rq->nr_iowait);
4084 delayacct_blkio_end();
4086 EXPORT_SYMBOL(io_schedule);
4088 long __sched io_schedule_timeout(long timeout)
4090 struct rq *rq = raw_rq();
4093 delayacct_blkio_start();
4094 atomic_inc(&rq->nr_iowait);
4095 blk_flush_plug(current);
4096 current->in_iowait = 1;
4097 ret = schedule_timeout(timeout);
4098 current->in_iowait = 0;
4099 atomic_dec(&rq->nr_iowait);
4100 delayacct_blkio_end();
4105 * sys_sched_get_priority_max - return maximum RT priority.
4106 * @policy: scheduling class.
4108 * Return: On success, this syscall returns the maximum
4109 * rt_priority that can be used by a given scheduling class.
4110 * On failure, a negative error code is returned.
4112 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4119 ret = MAX_USER_RT_PRIO-1;
4131 * sys_sched_get_priority_min - return minimum RT priority.
4132 * @policy: scheduling class.
4134 * Return: On success, this syscall returns the minimum
4135 * rt_priority that can be used by a given scheduling class.
4136 * On failure, a negative error code is returned.
4138 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4156 * sys_sched_rr_get_interval - return the default timeslice of a process.
4157 * @pid: pid of the process.
4158 * @interval: userspace pointer to the timeslice value.
4160 * this syscall writes the default timeslice value of a given process
4161 * into the user-space timespec buffer. A value of '0' means infinity.
4163 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4166 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4167 struct timespec __user *, interval)
4169 struct task_struct *p;
4170 unsigned int time_slice;
4171 unsigned long flags;
4181 p = find_process_by_pid(pid);
4185 retval = security_task_getscheduler(p);
4189 rq = task_rq_lock(p, &flags);
4190 time_slice = p->sched_class->get_rr_interval(rq, p);
4191 task_rq_unlock(rq, p, &flags);
4194 jiffies_to_timespec(time_slice, &t);
4195 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4203 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4205 void sched_show_task(struct task_struct *p)
4207 unsigned long free = 0;
4211 state = p->state ? __ffs(p->state) + 1 : 0;
4212 printk(KERN_INFO "%-15.15s %c", p->comm,
4213 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4214 #if BITS_PER_LONG == 32
4215 if (state == TASK_RUNNING)
4216 printk(KERN_CONT " running ");
4218 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4220 if (state == TASK_RUNNING)
4221 printk(KERN_CONT " running task ");
4223 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4225 #ifdef CONFIG_DEBUG_STACK_USAGE
4226 free = stack_not_used(p);
4229 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4231 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4232 task_pid_nr(p), ppid,
4233 (unsigned long)task_thread_info(p)->flags);
4235 print_worker_info(KERN_INFO, p);
4236 show_stack(p, NULL);
4239 void show_state_filter(unsigned long state_filter)
4241 struct task_struct *g, *p;
4243 #if BITS_PER_LONG == 32
4245 " task PC stack pid father\n");
4248 " task PC stack pid father\n");
4251 do_each_thread(g, p) {
4253 * reset the NMI-timeout, listing all files on a slow
4254 * console might take a lot of time:
4256 touch_nmi_watchdog();
4257 if (!state_filter || (p->state & state_filter))
4259 } while_each_thread(g, p);
4261 touch_all_softlockup_watchdogs();
4263 #ifdef CONFIG_SCHED_DEBUG
4264 sysrq_sched_debug_show();
4268 * Only show locks if all tasks are dumped:
4271 debug_show_all_locks();
4274 void init_idle_bootup_task(struct task_struct *idle)
4276 idle->sched_class = &idle_sched_class;
4280 * init_idle - set up an idle thread for a given CPU
4281 * @idle: task in question
4282 * @cpu: cpu the idle task belongs to
4284 * NOTE: this function does not set the idle thread's NEED_RESCHED
4285 * flag, to make booting more robust.
4287 void init_idle(struct task_struct *idle, int cpu)
4289 struct rq *rq = cpu_rq(cpu);
4290 unsigned long flags;
4292 raw_spin_lock_irqsave(&rq->lock, flags);
4294 __sched_fork(0, idle);
4295 idle->state = TASK_RUNNING;
4296 idle->se.exec_start = sched_clock();
4298 do_set_cpus_allowed(idle, cpumask_of(cpu));
4300 * We're having a chicken and egg problem, even though we are
4301 * holding rq->lock, the cpu isn't yet set to this cpu so the
4302 * lockdep check in task_group() will fail.
4304 * Similar case to sched_fork(). / Alternatively we could
4305 * use task_rq_lock() here and obtain the other rq->lock.
4310 __set_task_cpu(idle, cpu);
4313 rq->curr = rq->idle = idle;
4314 #if defined(CONFIG_SMP)
4317 raw_spin_unlock_irqrestore(&rq->lock, flags);
4319 /* Set the preempt count _outside_ the spinlocks! */
4320 init_idle_preempt_count(idle, cpu);
4323 * The idle tasks have their own, simple scheduling class:
4325 idle->sched_class = &idle_sched_class;
4326 ftrace_graph_init_idle_task(idle, cpu);
4327 vtime_init_idle(idle, cpu);
4328 #if defined(CONFIG_SMP)
4329 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4334 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4336 if (p->sched_class && p->sched_class->set_cpus_allowed)
4337 p->sched_class->set_cpus_allowed(p, new_mask);
4339 cpumask_copy(&p->cpus_allowed, new_mask);
4340 p->nr_cpus_allowed = cpumask_weight(new_mask);
4344 * This is how migration works:
4346 * 1) we invoke migration_cpu_stop() on the target CPU using
4348 * 2) stopper starts to run (implicitly forcing the migrated thread
4350 * 3) it checks whether the migrated task is still in the wrong runqueue.
4351 * 4) if it's in the wrong runqueue then the migration thread removes
4352 * it and puts it into the right queue.
4353 * 5) stopper completes and stop_one_cpu() returns and the migration
4358 * Change a given task's CPU affinity. Migrate the thread to a
4359 * proper CPU and schedule it away if the CPU it's executing on
4360 * is removed from the allowed bitmask.
4362 * NOTE: the caller must have a valid reference to the task, the
4363 * task must not exit() & deallocate itself prematurely. The
4364 * call is not atomic; no spinlocks may be held.
4366 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4368 unsigned long flags;
4370 unsigned int dest_cpu;
4373 rq = task_rq_lock(p, &flags);
4375 if (cpumask_equal(&p->cpus_allowed, new_mask))
4378 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4383 do_set_cpus_allowed(p, new_mask);
4385 /* Can the task run on the task's current CPU? If so, we're done */
4386 if (cpumask_test_cpu(task_cpu(p), new_mask))
4389 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4391 struct migration_arg arg = { p, dest_cpu };
4392 /* Need help from migration thread: drop lock and wait. */
4393 task_rq_unlock(rq, p, &flags);
4394 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4395 tlb_migrate_finish(p->mm);
4399 task_rq_unlock(rq, p, &flags);
4403 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4406 * Move (not current) task off this cpu, onto dest cpu. We're doing
4407 * this because either it can't run here any more (set_cpus_allowed()
4408 * away from this CPU, or CPU going down), or because we're
4409 * attempting to rebalance this task on exec (sched_exec).
4411 * So we race with normal scheduler movements, but that's OK, as long
4412 * as the task is no longer on this CPU.
4414 * Returns non-zero if task was successfully migrated.
4416 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4418 struct rq *rq_dest, *rq_src;
4421 if (unlikely(!cpu_active(dest_cpu)))
4424 rq_src = cpu_rq(src_cpu);
4425 rq_dest = cpu_rq(dest_cpu);
4427 raw_spin_lock(&p->pi_lock);
4428 double_rq_lock(rq_src, rq_dest);
4429 /* Already moved. */
4430 if (task_cpu(p) != src_cpu)
4432 /* Affinity changed (again). */
4433 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4437 * If we're not on a rq, the next wake-up will ensure we're
4441 dequeue_task(rq_src, p, 0);
4442 set_task_cpu(p, dest_cpu);
4443 enqueue_task(rq_dest, p, 0);
4444 check_preempt_curr(rq_dest, p, 0);
4449 double_rq_unlock(rq_src, rq_dest);
4450 raw_spin_unlock(&p->pi_lock);
4454 #ifdef CONFIG_NUMA_BALANCING
4455 /* Migrate current task p to target_cpu */
4456 int migrate_task_to(struct task_struct *p, int target_cpu)
4458 struct migration_arg arg = { p, target_cpu };
4459 int curr_cpu = task_cpu(p);
4461 if (curr_cpu == target_cpu)
4464 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4467 /* TODO: This is not properly updating schedstats */
4469 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4473 * Requeue a task on a given node and accurately track the number of NUMA
4474 * tasks on the runqueues
4476 void sched_setnuma(struct task_struct *p, int nid)
4479 unsigned long flags;
4480 bool on_rq, running;
4482 rq = task_rq_lock(p, &flags);
4484 running = task_current(rq, p);
4487 dequeue_task(rq, p, 0);
4489 p->sched_class->put_prev_task(rq, p);
4491 p->numa_preferred_nid = nid;
4492 p->numa_migrate_seq = 1;
4495 p->sched_class->set_curr_task(rq);
4497 enqueue_task(rq, p, 0);
4498 task_rq_unlock(rq, p, &flags);
4503 * migration_cpu_stop - this will be executed by a highprio stopper thread
4504 * and performs thread migration by bumping thread off CPU then
4505 * 'pushing' onto another runqueue.
4507 static int migration_cpu_stop(void *data)
4509 struct migration_arg *arg = data;
4512 * The original target cpu might have gone down and we might
4513 * be on another cpu but it doesn't matter.
4515 local_irq_disable();
4516 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4521 #ifdef CONFIG_HOTPLUG_CPU
4524 * Ensures that the idle task is using init_mm right before its cpu goes
4527 void idle_task_exit(void)
4529 struct mm_struct *mm = current->active_mm;
4531 BUG_ON(cpu_online(smp_processor_id()));
4534 switch_mm(mm, &init_mm, current);
4539 * Since this CPU is going 'away' for a while, fold any nr_active delta
4540 * we might have. Assumes we're called after migrate_tasks() so that the
4541 * nr_active count is stable.
4543 * Also see the comment "Global load-average calculations".
4545 static void calc_load_migrate(struct rq *rq)
4547 long delta = calc_load_fold_active(rq);
4549 atomic_long_add(delta, &calc_load_tasks);
4553 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4554 * try_to_wake_up()->select_task_rq().
4556 * Called with rq->lock held even though we'er in stop_machine() and
4557 * there's no concurrency possible, we hold the required locks anyway
4558 * because of lock validation efforts.
4560 static void migrate_tasks(unsigned int dead_cpu)
4562 struct rq *rq = cpu_rq(dead_cpu);
4563 struct task_struct *next, *stop = rq->stop;
4567 * Fudge the rq selection such that the below task selection loop
4568 * doesn't get stuck on the currently eligible stop task.
4570 * We're currently inside stop_machine() and the rq is either stuck
4571 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4572 * either way we should never end up calling schedule() until we're
4578 * put_prev_task() and pick_next_task() sched
4579 * class method both need to have an up-to-date
4580 * value of rq->clock[_task]
4582 update_rq_clock(rq);
4586 * There's this thread running, bail when that's the only
4589 if (rq->nr_running == 1)
4592 next = pick_next_task(rq);
4594 next->sched_class->put_prev_task(rq, next);
4596 /* Find suitable destination for @next, with force if needed. */
4597 dest_cpu = select_fallback_rq(dead_cpu, next);
4598 raw_spin_unlock(&rq->lock);
4600 __migrate_task(next, dead_cpu, dest_cpu);
4602 raw_spin_lock(&rq->lock);
4608 #endif /* CONFIG_HOTPLUG_CPU */
4610 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4612 static struct ctl_table sd_ctl_dir[] = {
4614 .procname = "sched_domain",
4620 static struct ctl_table sd_ctl_root[] = {
4622 .procname = "kernel",
4624 .child = sd_ctl_dir,
4629 static struct ctl_table *sd_alloc_ctl_entry(int n)
4631 struct ctl_table *entry =
4632 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4637 static void sd_free_ctl_entry(struct ctl_table **tablep)
4639 struct ctl_table *entry;
4642 * In the intermediate directories, both the child directory and
4643 * procname are dynamically allocated and could fail but the mode
4644 * will always be set. In the lowest directory the names are
4645 * static strings and all have proc handlers.
4647 for (entry = *tablep; entry->mode; entry++) {
4649 sd_free_ctl_entry(&entry->child);
4650 if (entry->proc_handler == NULL)
4651 kfree(entry->procname);
4658 static int min_load_idx = 0;
4659 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4662 set_table_entry(struct ctl_table *entry,
4663 const char *procname, void *data, int maxlen,
4664 umode_t mode, proc_handler *proc_handler,
4667 entry->procname = procname;
4669 entry->maxlen = maxlen;
4671 entry->proc_handler = proc_handler;
4674 entry->extra1 = &min_load_idx;
4675 entry->extra2 = &max_load_idx;
4679 static struct ctl_table *
4680 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4682 struct ctl_table *table = sd_alloc_ctl_entry(13);
4687 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4688 sizeof(long), 0644, proc_doulongvec_minmax, false);
4689 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4690 sizeof(long), 0644, proc_doulongvec_minmax, false);
4691 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4692 sizeof(int), 0644, proc_dointvec_minmax, true);
4693 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4694 sizeof(int), 0644, proc_dointvec_minmax, true);
4695 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4696 sizeof(int), 0644, proc_dointvec_minmax, true);
4697 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4698 sizeof(int), 0644, proc_dointvec_minmax, true);
4699 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4700 sizeof(int), 0644, proc_dointvec_minmax, true);
4701 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4702 sizeof(int), 0644, proc_dointvec_minmax, false);
4703 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4704 sizeof(int), 0644, proc_dointvec_minmax, false);
4705 set_table_entry(&table[9], "cache_nice_tries",
4706 &sd->cache_nice_tries,
4707 sizeof(int), 0644, proc_dointvec_minmax, false);
4708 set_table_entry(&table[10], "flags", &sd->flags,
4709 sizeof(int), 0644, proc_dointvec_minmax, false);
4710 set_table_entry(&table[11], "name", sd->name,
4711 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4712 /* &table[12] is terminator */
4717 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4719 struct ctl_table *entry, *table;
4720 struct sched_domain *sd;
4721 int domain_num = 0, i;
4724 for_each_domain(cpu, sd)
4726 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4731 for_each_domain(cpu, sd) {
4732 snprintf(buf, 32, "domain%d", i);
4733 entry->procname = kstrdup(buf, GFP_KERNEL);
4735 entry->child = sd_alloc_ctl_domain_table(sd);
4742 static struct ctl_table_header *sd_sysctl_header;
4743 static void register_sched_domain_sysctl(void)
4745 int i, cpu_num = num_possible_cpus();
4746 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4749 WARN_ON(sd_ctl_dir[0].child);
4750 sd_ctl_dir[0].child = entry;
4755 for_each_possible_cpu(i) {
4756 snprintf(buf, 32, "cpu%d", i);
4757 entry->procname = kstrdup(buf, GFP_KERNEL);
4759 entry->child = sd_alloc_ctl_cpu_table(i);
4763 WARN_ON(sd_sysctl_header);
4764 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4767 /* may be called multiple times per register */
4768 static void unregister_sched_domain_sysctl(void)
4770 if (sd_sysctl_header)
4771 unregister_sysctl_table(sd_sysctl_header);
4772 sd_sysctl_header = NULL;
4773 if (sd_ctl_dir[0].child)
4774 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4777 static void register_sched_domain_sysctl(void)
4780 static void unregister_sched_domain_sysctl(void)
4785 static void set_rq_online(struct rq *rq)
4788 const struct sched_class *class;
4790 cpumask_set_cpu(rq->cpu, rq->rd->online);
4793 for_each_class(class) {
4794 if (class->rq_online)
4795 class->rq_online(rq);
4800 static void set_rq_offline(struct rq *rq)
4803 const struct sched_class *class;
4805 for_each_class(class) {
4806 if (class->rq_offline)
4807 class->rq_offline(rq);
4810 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4816 * migration_call - callback that gets triggered when a CPU is added.
4817 * Here we can start up the necessary migration thread for the new CPU.
4820 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4822 int cpu = (long)hcpu;
4823 unsigned long flags;
4824 struct rq *rq = cpu_rq(cpu);
4826 switch (action & ~CPU_TASKS_FROZEN) {
4828 case CPU_UP_PREPARE:
4829 rq->calc_load_update = calc_load_update;
4833 /* Update our root-domain */
4834 raw_spin_lock_irqsave(&rq->lock, flags);
4836 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4840 raw_spin_unlock_irqrestore(&rq->lock, flags);
4843 #ifdef CONFIG_HOTPLUG_CPU
4845 sched_ttwu_pending();
4846 /* Update our root-domain */
4847 raw_spin_lock_irqsave(&rq->lock, flags);
4849 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4853 BUG_ON(rq->nr_running != 1); /* the migration thread */
4854 raw_spin_unlock_irqrestore(&rq->lock, flags);
4858 calc_load_migrate(rq);
4863 update_max_interval();
4869 * Register at high priority so that task migration (migrate_all_tasks)
4870 * happens before everything else. This has to be lower priority than
4871 * the notifier in the perf_event subsystem, though.
4873 static struct notifier_block migration_notifier = {
4874 .notifier_call = migration_call,
4875 .priority = CPU_PRI_MIGRATION,
4878 static int sched_cpu_active(struct notifier_block *nfb,
4879 unsigned long action, void *hcpu)
4881 switch (action & ~CPU_TASKS_FROZEN) {
4883 case CPU_DOWN_FAILED:
4884 set_cpu_active((long)hcpu, true);
4891 static int sched_cpu_inactive(struct notifier_block *nfb,
4892 unsigned long action, void *hcpu)
4894 switch (action & ~CPU_TASKS_FROZEN) {
4895 case CPU_DOWN_PREPARE:
4896 set_cpu_active((long)hcpu, false);
4903 static int __init migration_init(void)
4905 void *cpu = (void *)(long)smp_processor_id();
4908 /* Initialize migration for the boot CPU */
4909 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4910 BUG_ON(err == NOTIFY_BAD);
4911 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4912 register_cpu_notifier(&migration_notifier);
4914 /* Register cpu active notifiers */
4915 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4916 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4920 early_initcall(migration_init);
4925 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4927 #ifdef CONFIG_SCHED_DEBUG
4929 static __read_mostly int sched_debug_enabled;
4931 static int __init sched_debug_setup(char *str)
4933 sched_debug_enabled = 1;
4937 early_param("sched_debug", sched_debug_setup);
4939 static inline bool sched_debug(void)
4941 return sched_debug_enabled;
4944 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4945 struct cpumask *groupmask)
4947 struct sched_group *group = sd->groups;
4950 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4951 cpumask_clear(groupmask);
4953 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4955 if (!(sd->flags & SD_LOAD_BALANCE)) {
4956 printk("does not load-balance\n");
4958 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4963 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4965 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4966 printk(KERN_ERR "ERROR: domain->span does not contain "
4969 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4970 printk(KERN_ERR "ERROR: domain->groups does not contain"
4974 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4978 printk(KERN_ERR "ERROR: group is NULL\n");
4983 * Even though we initialize ->power to something semi-sane,
4984 * we leave power_orig unset. This allows us to detect if
4985 * domain iteration is still funny without causing /0 traps.
4987 if (!group->sgp->power_orig) {
4988 printk(KERN_CONT "\n");
4989 printk(KERN_ERR "ERROR: domain->cpu_power not "
4994 if (!cpumask_weight(sched_group_cpus(group))) {
4995 printk(KERN_CONT "\n");
4996 printk(KERN_ERR "ERROR: empty group\n");
5000 if (!(sd->flags & SD_OVERLAP) &&
5001 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5002 printk(KERN_CONT "\n");
5003 printk(KERN_ERR "ERROR: repeated CPUs\n");
5007 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5009 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5011 printk(KERN_CONT " %s", str);
5012 if (group->sgp->power != SCHED_POWER_SCALE) {
5013 printk(KERN_CONT " (cpu_power = %d)",
5017 group = group->next;
5018 } while (group != sd->groups);
5019 printk(KERN_CONT "\n");
5021 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5022 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5025 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5026 printk(KERN_ERR "ERROR: parent span is not a superset "
5027 "of domain->span\n");
5031 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5035 if (!sched_debug_enabled)
5039 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5043 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5046 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5054 #else /* !CONFIG_SCHED_DEBUG */
5055 # define sched_domain_debug(sd, cpu) do { } while (0)
5056 static inline bool sched_debug(void)
5060 #endif /* CONFIG_SCHED_DEBUG */
5062 static int sd_degenerate(struct sched_domain *sd)
5064 if (cpumask_weight(sched_domain_span(sd)) == 1)
5067 /* Following flags need at least 2 groups */
5068 if (sd->flags & (SD_LOAD_BALANCE |
5069 SD_BALANCE_NEWIDLE |
5073 SD_SHARE_PKG_RESOURCES)) {
5074 if (sd->groups != sd->groups->next)
5078 /* Following flags don't use groups */
5079 if (sd->flags & (SD_WAKE_AFFINE))
5086 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5088 unsigned long cflags = sd->flags, pflags = parent->flags;
5090 if (sd_degenerate(parent))
5093 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5096 /* Flags needing groups don't count if only 1 group in parent */
5097 if (parent->groups == parent->groups->next) {
5098 pflags &= ~(SD_LOAD_BALANCE |
5099 SD_BALANCE_NEWIDLE |
5103 SD_SHARE_PKG_RESOURCES |
5105 if (nr_node_ids == 1)
5106 pflags &= ~SD_SERIALIZE;
5108 if (~cflags & pflags)
5114 static void free_rootdomain(struct rcu_head *rcu)
5116 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5118 cpupri_cleanup(&rd->cpupri);
5119 free_cpumask_var(rd->rto_mask);
5120 free_cpumask_var(rd->online);
5121 free_cpumask_var(rd->span);
5125 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5127 struct root_domain *old_rd = NULL;
5128 unsigned long flags;
5130 raw_spin_lock_irqsave(&rq->lock, flags);
5135 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5138 cpumask_clear_cpu(rq->cpu, old_rd->span);
5141 * If we dont want to free the old_rt yet then
5142 * set old_rd to NULL to skip the freeing later
5145 if (!atomic_dec_and_test(&old_rd->refcount))
5149 atomic_inc(&rd->refcount);
5152 cpumask_set_cpu(rq->cpu, rd->span);
5153 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5156 raw_spin_unlock_irqrestore(&rq->lock, flags);
5159 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5162 static int init_rootdomain(struct root_domain *rd)
5164 memset(rd, 0, sizeof(*rd));
5166 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5168 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5170 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5173 if (cpupri_init(&rd->cpupri) != 0)
5178 free_cpumask_var(rd->rto_mask);
5180 free_cpumask_var(rd->online);
5182 free_cpumask_var(rd->span);
5188 * By default the system creates a single root-domain with all cpus as
5189 * members (mimicking the global state we have today).
5191 struct root_domain def_root_domain;
5193 static void init_defrootdomain(void)
5195 init_rootdomain(&def_root_domain);
5197 atomic_set(&def_root_domain.refcount, 1);
5200 static struct root_domain *alloc_rootdomain(void)
5202 struct root_domain *rd;
5204 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5208 if (init_rootdomain(rd) != 0) {
5216 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5218 struct sched_group *tmp, *first;
5227 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5232 } while (sg != first);
5235 static void free_sched_domain(struct rcu_head *rcu)
5237 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5240 * If its an overlapping domain it has private groups, iterate and
5243 if (sd->flags & SD_OVERLAP) {
5244 free_sched_groups(sd->groups, 1);
5245 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5246 kfree(sd->groups->sgp);
5252 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5254 call_rcu(&sd->rcu, free_sched_domain);
5257 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5259 for (; sd; sd = sd->parent)
5260 destroy_sched_domain(sd, cpu);
5264 * Keep a special pointer to the highest sched_domain that has
5265 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5266 * allows us to avoid some pointer chasing select_idle_sibling().
5268 * Also keep a unique ID per domain (we use the first cpu number in
5269 * the cpumask of the domain), this allows us to quickly tell if
5270 * two cpus are in the same cache domain, see cpus_share_cache().
5272 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5273 DEFINE_PER_CPU(int, sd_llc_size);
5274 DEFINE_PER_CPU(int, sd_llc_id);
5275 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5277 static void update_top_cache_domain(int cpu)
5279 struct sched_domain *sd;
5283 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5285 id = cpumask_first(sched_domain_span(sd));
5286 size = cpumask_weight(sched_domain_span(sd));
5289 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5290 per_cpu(sd_llc_size, cpu) = size;
5291 per_cpu(sd_llc_id, cpu) = id;
5293 sd = lowest_flag_domain(cpu, SD_NUMA);
5294 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5298 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5299 * hold the hotplug lock.
5302 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5304 struct rq *rq = cpu_rq(cpu);
5305 struct sched_domain *tmp;
5307 /* Remove the sched domains which do not contribute to scheduling. */
5308 for (tmp = sd; tmp; ) {
5309 struct sched_domain *parent = tmp->parent;
5313 if (sd_parent_degenerate(tmp, parent)) {
5314 tmp->parent = parent->parent;
5316 parent->parent->child = tmp;
5318 * Transfer SD_PREFER_SIBLING down in case of a
5319 * degenerate parent; the spans match for this
5320 * so the property transfers.
5322 if (parent->flags & SD_PREFER_SIBLING)
5323 tmp->flags |= SD_PREFER_SIBLING;
5324 destroy_sched_domain(parent, cpu);
5329 if (sd && sd_degenerate(sd)) {
5332 destroy_sched_domain(tmp, cpu);
5337 sched_domain_debug(sd, cpu);
5339 rq_attach_root(rq, rd);
5341 rcu_assign_pointer(rq->sd, sd);
5342 destroy_sched_domains(tmp, cpu);
5344 update_top_cache_domain(cpu);
5347 /* cpus with isolated domains */
5348 static cpumask_var_t cpu_isolated_map;
5350 /* Setup the mask of cpus configured for isolated domains */
5351 static int __init isolated_cpu_setup(char *str)
5353 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5354 cpulist_parse(str, cpu_isolated_map);
5358 __setup("isolcpus=", isolated_cpu_setup);
5360 static const struct cpumask *cpu_cpu_mask(int cpu)
5362 return cpumask_of_node(cpu_to_node(cpu));
5366 struct sched_domain **__percpu sd;
5367 struct sched_group **__percpu sg;
5368 struct sched_group_power **__percpu sgp;
5372 struct sched_domain ** __percpu sd;
5373 struct root_domain *rd;
5383 struct sched_domain_topology_level;
5385 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5386 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5388 #define SDTL_OVERLAP 0x01
5390 struct sched_domain_topology_level {
5391 sched_domain_init_f init;
5392 sched_domain_mask_f mask;
5395 struct sd_data data;
5399 * Build an iteration mask that can exclude certain CPUs from the upwards
5402 * Asymmetric node setups can result in situations where the domain tree is of
5403 * unequal depth, make sure to skip domains that already cover the entire
5406 * In that case build_sched_domains() will have terminated the iteration early
5407 * and our sibling sd spans will be empty. Domains should always include the
5408 * cpu they're built on, so check that.
5411 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5413 const struct cpumask *span = sched_domain_span(sd);
5414 struct sd_data *sdd = sd->private;
5415 struct sched_domain *sibling;
5418 for_each_cpu(i, span) {
5419 sibling = *per_cpu_ptr(sdd->sd, i);
5420 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5423 cpumask_set_cpu(i, sched_group_mask(sg));
5428 * Return the canonical balance cpu for this group, this is the first cpu
5429 * of this group that's also in the iteration mask.
5431 int group_balance_cpu(struct sched_group *sg)
5433 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5437 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5439 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5440 const struct cpumask *span = sched_domain_span(sd);
5441 struct cpumask *covered = sched_domains_tmpmask;
5442 struct sd_data *sdd = sd->private;
5443 struct sched_domain *child;
5446 cpumask_clear(covered);
5448 for_each_cpu(i, span) {
5449 struct cpumask *sg_span;
5451 if (cpumask_test_cpu(i, covered))
5454 child = *per_cpu_ptr(sdd->sd, i);
5456 /* See the comment near build_group_mask(). */
5457 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5460 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5461 GFP_KERNEL, cpu_to_node(cpu));
5466 sg_span = sched_group_cpus(sg);
5468 child = child->child;
5469 cpumask_copy(sg_span, sched_domain_span(child));
5471 cpumask_set_cpu(i, sg_span);
5473 cpumask_or(covered, covered, sg_span);
5475 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5476 if (atomic_inc_return(&sg->sgp->ref) == 1)
5477 build_group_mask(sd, sg);
5480 * Initialize sgp->power such that even if we mess up the
5481 * domains and no possible iteration will get us here, we won't
5484 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5487 * Make sure the first group of this domain contains the
5488 * canonical balance cpu. Otherwise the sched_domain iteration
5489 * breaks. See update_sg_lb_stats().
5491 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5492 group_balance_cpu(sg) == cpu)
5502 sd->groups = groups;
5507 free_sched_groups(first, 0);
5512 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5514 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5515 struct sched_domain *child = sd->child;
5518 cpu = cpumask_first(sched_domain_span(child));
5521 *sg = *per_cpu_ptr(sdd->sg, cpu);
5522 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5523 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5530 * build_sched_groups will build a circular linked list of the groups
5531 * covered by the given span, and will set each group's ->cpumask correctly,
5532 * and ->cpu_power to 0.
5534 * Assumes the sched_domain tree is fully constructed
5537 build_sched_groups(struct sched_domain *sd, int cpu)
5539 struct sched_group *first = NULL, *last = NULL;
5540 struct sd_data *sdd = sd->private;
5541 const struct cpumask *span = sched_domain_span(sd);
5542 struct cpumask *covered;
5545 get_group(cpu, sdd, &sd->groups);
5546 atomic_inc(&sd->groups->ref);
5548 if (cpu != cpumask_first(span))
5551 lockdep_assert_held(&sched_domains_mutex);
5552 covered = sched_domains_tmpmask;
5554 cpumask_clear(covered);
5556 for_each_cpu(i, span) {
5557 struct sched_group *sg;
5560 if (cpumask_test_cpu(i, covered))
5563 group = get_group(i, sdd, &sg);
5564 cpumask_clear(sched_group_cpus(sg));
5566 cpumask_setall(sched_group_mask(sg));
5568 for_each_cpu(j, span) {
5569 if (get_group(j, sdd, NULL) != group)
5572 cpumask_set_cpu(j, covered);
5573 cpumask_set_cpu(j, sched_group_cpus(sg));
5588 * Initialize sched groups cpu_power.
5590 * cpu_power indicates the capacity of sched group, which is used while
5591 * distributing the load between different sched groups in a sched domain.
5592 * Typically cpu_power for all the groups in a sched domain will be same unless
5593 * there are asymmetries in the topology. If there are asymmetries, group
5594 * having more cpu_power will pickup more load compared to the group having
5597 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5599 struct sched_group *sg = sd->groups;
5604 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5606 } while (sg != sd->groups);
5608 if (cpu != group_balance_cpu(sg))
5611 update_group_power(sd, cpu);
5612 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5615 int __weak arch_sd_sibling_asym_packing(void)
5617 return 0*SD_ASYM_PACKING;
5621 * Initializers for schedule domains
5622 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5625 #ifdef CONFIG_SCHED_DEBUG
5626 # define SD_INIT_NAME(sd, type) sd->name = #type
5628 # define SD_INIT_NAME(sd, type) do { } while (0)
5631 #define SD_INIT_FUNC(type) \
5632 static noinline struct sched_domain * \
5633 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5635 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5636 *sd = SD_##type##_INIT; \
5637 SD_INIT_NAME(sd, type); \
5638 sd->private = &tl->data; \
5643 #ifdef CONFIG_SCHED_SMT
5644 SD_INIT_FUNC(SIBLING)
5646 #ifdef CONFIG_SCHED_MC
5649 #ifdef CONFIG_SCHED_BOOK
5653 static int default_relax_domain_level = -1;
5654 int sched_domain_level_max;
5656 static int __init setup_relax_domain_level(char *str)
5658 if (kstrtoint(str, 0, &default_relax_domain_level))
5659 pr_warn("Unable to set relax_domain_level\n");
5663 __setup("relax_domain_level=", setup_relax_domain_level);
5665 static void set_domain_attribute(struct sched_domain *sd,
5666 struct sched_domain_attr *attr)
5670 if (!attr || attr->relax_domain_level < 0) {
5671 if (default_relax_domain_level < 0)
5674 request = default_relax_domain_level;
5676 request = attr->relax_domain_level;
5677 if (request < sd->level) {
5678 /* turn off idle balance on this domain */
5679 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5681 /* turn on idle balance on this domain */
5682 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5686 static void __sdt_free(const struct cpumask *cpu_map);
5687 static int __sdt_alloc(const struct cpumask *cpu_map);
5689 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5690 const struct cpumask *cpu_map)
5694 if (!atomic_read(&d->rd->refcount))
5695 free_rootdomain(&d->rd->rcu); /* fall through */
5697 free_percpu(d->sd); /* fall through */
5699 __sdt_free(cpu_map); /* fall through */
5705 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5706 const struct cpumask *cpu_map)
5708 memset(d, 0, sizeof(*d));
5710 if (__sdt_alloc(cpu_map))
5711 return sa_sd_storage;
5712 d->sd = alloc_percpu(struct sched_domain *);
5714 return sa_sd_storage;
5715 d->rd = alloc_rootdomain();
5718 return sa_rootdomain;
5722 * NULL the sd_data elements we've used to build the sched_domain and
5723 * sched_group structure so that the subsequent __free_domain_allocs()
5724 * will not free the data we're using.
5726 static void claim_allocations(int cpu, struct sched_domain *sd)
5728 struct sd_data *sdd = sd->private;
5730 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5731 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5733 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5734 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5736 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5737 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5740 #ifdef CONFIG_SCHED_SMT
5741 static const struct cpumask *cpu_smt_mask(int cpu)
5743 return topology_thread_cpumask(cpu);
5748 * Topology list, bottom-up.
5750 static struct sched_domain_topology_level default_topology[] = {
5751 #ifdef CONFIG_SCHED_SMT
5752 { sd_init_SIBLING, cpu_smt_mask, },
5754 #ifdef CONFIG_SCHED_MC
5755 { sd_init_MC, cpu_coregroup_mask, },
5757 #ifdef CONFIG_SCHED_BOOK
5758 { sd_init_BOOK, cpu_book_mask, },
5760 { sd_init_CPU, cpu_cpu_mask, },
5764 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5766 #define for_each_sd_topology(tl) \
5767 for (tl = sched_domain_topology; tl->init; tl++)
5771 static int sched_domains_numa_levels;
5772 static int *sched_domains_numa_distance;
5773 static struct cpumask ***sched_domains_numa_masks;
5774 static int sched_domains_curr_level;
5776 static inline int sd_local_flags(int level)
5778 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5781 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5784 static struct sched_domain *
5785 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5787 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5788 int level = tl->numa_level;
5789 int sd_weight = cpumask_weight(
5790 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5792 *sd = (struct sched_domain){
5793 .min_interval = sd_weight,
5794 .max_interval = 2*sd_weight,
5796 .imbalance_pct = 125,
5797 .cache_nice_tries = 2,
5804 .flags = 1*SD_LOAD_BALANCE
5805 | 1*SD_BALANCE_NEWIDLE
5810 | 0*SD_SHARE_CPUPOWER
5811 | 0*SD_SHARE_PKG_RESOURCES
5813 | 0*SD_PREFER_SIBLING
5815 | sd_local_flags(level)
5817 .last_balance = jiffies,
5818 .balance_interval = sd_weight,
5820 SD_INIT_NAME(sd, NUMA);
5821 sd->private = &tl->data;
5824 * Ugly hack to pass state to sd_numa_mask()...
5826 sched_domains_curr_level = tl->numa_level;
5831 static const struct cpumask *sd_numa_mask(int cpu)
5833 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5836 static void sched_numa_warn(const char *str)
5838 static int done = false;
5846 printk(KERN_WARNING "ERROR: %s\n\n", str);
5848 for (i = 0; i < nr_node_ids; i++) {
5849 printk(KERN_WARNING " ");
5850 for (j = 0; j < nr_node_ids; j++)
5851 printk(KERN_CONT "%02d ", node_distance(i,j));
5852 printk(KERN_CONT "\n");
5854 printk(KERN_WARNING "\n");
5857 static bool find_numa_distance(int distance)
5861 if (distance == node_distance(0, 0))
5864 for (i = 0; i < sched_domains_numa_levels; i++) {
5865 if (sched_domains_numa_distance[i] == distance)
5872 static void sched_init_numa(void)
5874 int next_distance, curr_distance = node_distance(0, 0);
5875 struct sched_domain_topology_level *tl;
5879 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5880 if (!sched_domains_numa_distance)
5884 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5885 * unique distances in the node_distance() table.
5887 * Assumes node_distance(0,j) includes all distances in
5888 * node_distance(i,j) in order to avoid cubic time.
5890 next_distance = curr_distance;
5891 for (i = 0; i < nr_node_ids; i++) {
5892 for (j = 0; j < nr_node_ids; j++) {
5893 for (k = 0; k < nr_node_ids; k++) {
5894 int distance = node_distance(i, k);
5896 if (distance > curr_distance &&
5897 (distance < next_distance ||
5898 next_distance == curr_distance))
5899 next_distance = distance;
5902 * While not a strong assumption it would be nice to know
5903 * about cases where if node A is connected to B, B is not
5904 * equally connected to A.
5906 if (sched_debug() && node_distance(k, i) != distance)
5907 sched_numa_warn("Node-distance not symmetric");
5909 if (sched_debug() && i && !find_numa_distance(distance))
5910 sched_numa_warn("Node-0 not representative");
5912 if (next_distance != curr_distance) {
5913 sched_domains_numa_distance[level++] = next_distance;
5914 sched_domains_numa_levels = level;
5915 curr_distance = next_distance;
5920 * In case of sched_debug() we verify the above assumption.
5926 * 'level' contains the number of unique distances, excluding the
5927 * identity distance node_distance(i,i).
5929 * The sched_domains_numa_distance[] array includes the actual distance
5934 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5935 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5936 * the array will contain less then 'level' members. This could be
5937 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5938 * in other functions.
5940 * We reset it to 'level' at the end of this function.
5942 sched_domains_numa_levels = 0;
5944 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5945 if (!sched_domains_numa_masks)
5949 * Now for each level, construct a mask per node which contains all
5950 * cpus of nodes that are that many hops away from us.
5952 for (i = 0; i < level; i++) {
5953 sched_domains_numa_masks[i] =
5954 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5955 if (!sched_domains_numa_masks[i])
5958 for (j = 0; j < nr_node_ids; j++) {
5959 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5963 sched_domains_numa_masks[i][j] = mask;
5965 for (k = 0; k < nr_node_ids; k++) {
5966 if (node_distance(j, k) > sched_domains_numa_distance[i])
5969 cpumask_or(mask, mask, cpumask_of_node(k));
5974 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5975 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5980 * Copy the default topology bits..
5982 for (i = 0; default_topology[i].init; i++)
5983 tl[i] = default_topology[i];
5986 * .. and append 'j' levels of NUMA goodness.
5988 for (j = 0; j < level; i++, j++) {
5989 tl[i] = (struct sched_domain_topology_level){
5990 .init = sd_numa_init,
5991 .mask = sd_numa_mask,
5992 .flags = SDTL_OVERLAP,
5997 sched_domain_topology = tl;
5999 sched_domains_numa_levels = level;
6002 static void sched_domains_numa_masks_set(int cpu)
6005 int node = cpu_to_node(cpu);
6007 for (i = 0; i < sched_domains_numa_levels; i++) {
6008 for (j = 0; j < nr_node_ids; j++) {
6009 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6010 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6015 static void sched_domains_numa_masks_clear(int cpu)
6018 for (i = 0; i < sched_domains_numa_levels; i++) {
6019 for (j = 0; j < nr_node_ids; j++)
6020 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6025 * Update sched_domains_numa_masks[level][node] array when new cpus
6028 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6029 unsigned long action,
6032 int cpu = (long)hcpu;
6034 switch (action & ~CPU_TASKS_FROZEN) {
6036 sched_domains_numa_masks_set(cpu);
6040 sched_domains_numa_masks_clear(cpu);
6050 static inline void sched_init_numa(void)
6054 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6055 unsigned long action,
6060 #endif /* CONFIG_NUMA */
6062 static int __sdt_alloc(const struct cpumask *cpu_map)
6064 struct sched_domain_topology_level *tl;
6067 for_each_sd_topology(tl) {
6068 struct sd_data *sdd = &tl->data;
6070 sdd->sd = alloc_percpu(struct sched_domain *);
6074 sdd->sg = alloc_percpu(struct sched_group *);
6078 sdd->sgp = alloc_percpu(struct sched_group_power *);
6082 for_each_cpu(j, cpu_map) {
6083 struct sched_domain *sd;
6084 struct sched_group *sg;
6085 struct sched_group_power *sgp;
6087 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6088 GFP_KERNEL, cpu_to_node(j));
6092 *per_cpu_ptr(sdd->sd, j) = sd;
6094 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6095 GFP_KERNEL, cpu_to_node(j));
6101 *per_cpu_ptr(sdd->sg, j) = sg;
6103 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6104 GFP_KERNEL, cpu_to_node(j));
6108 *per_cpu_ptr(sdd->sgp, j) = sgp;
6115 static void __sdt_free(const struct cpumask *cpu_map)
6117 struct sched_domain_topology_level *tl;
6120 for_each_sd_topology(tl) {
6121 struct sd_data *sdd = &tl->data;
6123 for_each_cpu(j, cpu_map) {
6124 struct sched_domain *sd;
6127 sd = *per_cpu_ptr(sdd->sd, j);
6128 if (sd && (sd->flags & SD_OVERLAP))
6129 free_sched_groups(sd->groups, 0);
6130 kfree(*per_cpu_ptr(sdd->sd, j));
6134 kfree(*per_cpu_ptr(sdd->sg, j));
6136 kfree(*per_cpu_ptr(sdd->sgp, j));
6138 free_percpu(sdd->sd);
6140 free_percpu(sdd->sg);
6142 free_percpu(sdd->sgp);
6147 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6148 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6149 struct sched_domain *child, int cpu)
6151 struct sched_domain *sd = tl->init(tl, cpu);
6155 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6157 sd->level = child->level + 1;
6158 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6162 set_domain_attribute(sd, attr);
6168 * Build sched domains for a given set of cpus and attach the sched domains
6169 * to the individual cpus
6171 static int build_sched_domains(const struct cpumask *cpu_map,
6172 struct sched_domain_attr *attr)
6174 enum s_alloc alloc_state;
6175 struct sched_domain *sd;
6177 int i, ret = -ENOMEM;
6179 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6180 if (alloc_state != sa_rootdomain)
6183 /* Set up domains for cpus specified by the cpu_map. */
6184 for_each_cpu(i, cpu_map) {
6185 struct sched_domain_topology_level *tl;
6188 for_each_sd_topology(tl) {
6189 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6190 if (tl == sched_domain_topology)
6191 *per_cpu_ptr(d.sd, i) = sd;
6192 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6193 sd->flags |= SD_OVERLAP;
6194 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6199 /* Build the groups for the domains */
6200 for_each_cpu(i, cpu_map) {
6201 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6202 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6203 if (sd->flags & SD_OVERLAP) {
6204 if (build_overlap_sched_groups(sd, i))
6207 if (build_sched_groups(sd, i))
6213 /* Calculate CPU power for physical packages and nodes */
6214 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6215 if (!cpumask_test_cpu(i, cpu_map))
6218 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6219 claim_allocations(i, sd);
6220 init_sched_groups_power(i, sd);
6224 /* Attach the domains */
6226 for_each_cpu(i, cpu_map) {
6227 sd = *per_cpu_ptr(d.sd, i);
6228 cpu_attach_domain(sd, d.rd, i);
6234 __free_domain_allocs(&d, alloc_state, cpu_map);
6238 static cpumask_var_t *doms_cur; /* current sched domains */
6239 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6240 static struct sched_domain_attr *dattr_cur;
6241 /* attribues of custom domains in 'doms_cur' */
6244 * Special case: If a kmalloc of a doms_cur partition (array of
6245 * cpumask) fails, then fallback to a single sched domain,
6246 * as determined by the single cpumask fallback_doms.
6248 static cpumask_var_t fallback_doms;
6251 * arch_update_cpu_topology lets virtualized architectures update the
6252 * cpu core maps. It is supposed to return 1 if the topology changed
6253 * or 0 if it stayed the same.
6255 int __attribute__((weak)) arch_update_cpu_topology(void)
6260 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6263 cpumask_var_t *doms;
6265 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6268 for (i = 0; i < ndoms; i++) {
6269 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6270 free_sched_domains(doms, i);
6277 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6280 for (i = 0; i < ndoms; i++)
6281 free_cpumask_var(doms[i]);
6286 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6287 * For now this just excludes isolated cpus, but could be used to
6288 * exclude other special cases in the future.
6290 static int init_sched_domains(const struct cpumask *cpu_map)
6294 arch_update_cpu_topology();
6296 doms_cur = alloc_sched_domains(ndoms_cur);
6298 doms_cur = &fallback_doms;
6299 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6300 err = build_sched_domains(doms_cur[0], NULL);
6301 register_sched_domain_sysctl();
6307 * Detach sched domains from a group of cpus specified in cpu_map
6308 * These cpus will now be attached to the NULL domain
6310 static void detach_destroy_domains(const struct cpumask *cpu_map)
6315 for_each_cpu(i, cpu_map)
6316 cpu_attach_domain(NULL, &def_root_domain, i);
6320 /* handle null as "default" */
6321 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6322 struct sched_domain_attr *new, int idx_new)
6324 struct sched_domain_attr tmp;
6331 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6332 new ? (new + idx_new) : &tmp,
6333 sizeof(struct sched_domain_attr));
6337 * Partition sched domains as specified by the 'ndoms_new'
6338 * cpumasks in the array doms_new[] of cpumasks. This compares
6339 * doms_new[] to the current sched domain partitioning, doms_cur[].
6340 * It destroys each deleted domain and builds each new domain.
6342 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6343 * The masks don't intersect (don't overlap.) We should setup one
6344 * sched domain for each mask. CPUs not in any of the cpumasks will
6345 * not be load balanced. If the same cpumask appears both in the
6346 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6349 * The passed in 'doms_new' should be allocated using
6350 * alloc_sched_domains. This routine takes ownership of it and will
6351 * free_sched_domains it when done with it. If the caller failed the
6352 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6353 * and partition_sched_domains() will fallback to the single partition
6354 * 'fallback_doms', it also forces the domains to be rebuilt.
6356 * If doms_new == NULL it will be replaced with cpu_online_mask.
6357 * ndoms_new == 0 is a special case for destroying existing domains,
6358 * and it will not create the default domain.
6360 * Call with hotplug lock held
6362 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6363 struct sched_domain_attr *dattr_new)
6368 mutex_lock(&sched_domains_mutex);
6370 /* always unregister in case we don't destroy any domains */
6371 unregister_sched_domain_sysctl();
6373 /* Let architecture update cpu core mappings. */
6374 new_topology = arch_update_cpu_topology();
6376 n = doms_new ? ndoms_new : 0;
6378 /* Destroy deleted domains */
6379 for (i = 0; i < ndoms_cur; i++) {
6380 for (j = 0; j < n && !new_topology; j++) {
6381 if (cpumask_equal(doms_cur[i], doms_new[j])
6382 && dattrs_equal(dattr_cur, i, dattr_new, j))
6385 /* no match - a current sched domain not in new doms_new[] */
6386 detach_destroy_domains(doms_cur[i]);
6392 if (doms_new == NULL) {
6394 doms_new = &fallback_doms;
6395 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6396 WARN_ON_ONCE(dattr_new);
6399 /* Build new domains */
6400 for (i = 0; i < ndoms_new; i++) {
6401 for (j = 0; j < n && !new_topology; j++) {
6402 if (cpumask_equal(doms_new[i], doms_cur[j])
6403 && dattrs_equal(dattr_new, i, dattr_cur, j))
6406 /* no match - add a new doms_new */
6407 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6412 /* Remember the new sched domains */
6413 if (doms_cur != &fallback_doms)
6414 free_sched_domains(doms_cur, ndoms_cur);
6415 kfree(dattr_cur); /* kfree(NULL) is safe */
6416 doms_cur = doms_new;
6417 dattr_cur = dattr_new;
6418 ndoms_cur = ndoms_new;
6420 register_sched_domain_sysctl();
6422 mutex_unlock(&sched_domains_mutex);
6425 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6428 * Update cpusets according to cpu_active mask. If cpusets are
6429 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6430 * around partition_sched_domains().
6432 * If we come here as part of a suspend/resume, don't touch cpusets because we
6433 * want to restore it back to its original state upon resume anyway.
6435 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6439 case CPU_ONLINE_FROZEN:
6440 case CPU_DOWN_FAILED_FROZEN:
6443 * num_cpus_frozen tracks how many CPUs are involved in suspend
6444 * resume sequence. As long as this is not the last online
6445 * operation in the resume sequence, just build a single sched
6446 * domain, ignoring cpusets.
6449 if (likely(num_cpus_frozen)) {
6450 partition_sched_domains(1, NULL, NULL);
6455 * This is the last CPU online operation. So fall through and
6456 * restore the original sched domains by considering the
6457 * cpuset configurations.
6461 case CPU_DOWN_FAILED:
6462 cpuset_update_active_cpus(true);
6470 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6474 case CPU_DOWN_PREPARE:
6475 cpuset_update_active_cpus(false);
6477 case CPU_DOWN_PREPARE_FROZEN:
6479 partition_sched_domains(1, NULL, NULL);
6487 void __init sched_init_smp(void)
6489 cpumask_var_t non_isolated_cpus;
6491 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6492 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6497 mutex_lock(&sched_domains_mutex);
6498 init_sched_domains(cpu_active_mask);
6499 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6500 if (cpumask_empty(non_isolated_cpus))
6501 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6502 mutex_unlock(&sched_domains_mutex);
6505 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6506 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6507 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6511 /* Move init over to a non-isolated CPU */
6512 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6514 sched_init_granularity();
6515 free_cpumask_var(non_isolated_cpus);
6517 init_sched_rt_class();
6520 void __init sched_init_smp(void)
6522 sched_init_granularity();
6524 #endif /* CONFIG_SMP */
6526 const_debug unsigned int sysctl_timer_migration = 1;
6528 int in_sched_functions(unsigned long addr)
6530 return in_lock_functions(addr) ||
6531 (addr >= (unsigned long)__sched_text_start
6532 && addr < (unsigned long)__sched_text_end);
6535 #ifdef CONFIG_CGROUP_SCHED
6537 * Default task group.
6538 * Every task in system belongs to this group at bootup.
6540 struct task_group root_task_group;
6541 LIST_HEAD(task_groups);
6544 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6546 void __init sched_init(void)
6549 unsigned long alloc_size = 0, ptr;
6551 #ifdef CONFIG_FAIR_GROUP_SCHED
6552 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6554 #ifdef CONFIG_RT_GROUP_SCHED
6555 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6557 #ifdef CONFIG_CPUMASK_OFFSTACK
6558 alloc_size += num_possible_cpus() * cpumask_size();
6561 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6563 #ifdef CONFIG_FAIR_GROUP_SCHED
6564 root_task_group.se = (struct sched_entity **)ptr;
6565 ptr += nr_cpu_ids * sizeof(void **);
6567 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6568 ptr += nr_cpu_ids * sizeof(void **);
6570 #endif /* CONFIG_FAIR_GROUP_SCHED */
6571 #ifdef CONFIG_RT_GROUP_SCHED
6572 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6573 ptr += nr_cpu_ids * sizeof(void **);
6575 root_task_group.rt_rq = (struct rt_rq **)ptr;
6576 ptr += nr_cpu_ids * sizeof(void **);
6578 #endif /* CONFIG_RT_GROUP_SCHED */
6579 #ifdef CONFIG_CPUMASK_OFFSTACK
6580 for_each_possible_cpu(i) {
6581 per_cpu(load_balance_mask, i) = (void *)ptr;
6582 ptr += cpumask_size();
6584 #endif /* CONFIG_CPUMASK_OFFSTACK */
6588 init_defrootdomain();
6591 init_rt_bandwidth(&def_rt_bandwidth,
6592 global_rt_period(), global_rt_runtime());
6594 #ifdef CONFIG_RT_GROUP_SCHED
6595 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6596 global_rt_period(), global_rt_runtime());
6597 #endif /* CONFIG_RT_GROUP_SCHED */
6599 #ifdef CONFIG_CGROUP_SCHED
6600 list_add(&root_task_group.list, &task_groups);
6601 INIT_LIST_HEAD(&root_task_group.children);
6602 INIT_LIST_HEAD(&root_task_group.siblings);
6603 autogroup_init(&init_task);
6605 #endif /* CONFIG_CGROUP_SCHED */
6607 for_each_possible_cpu(i) {
6611 raw_spin_lock_init(&rq->lock);
6613 rq->calc_load_active = 0;
6614 rq->calc_load_update = jiffies + LOAD_FREQ;
6615 init_cfs_rq(&rq->cfs);
6616 init_rt_rq(&rq->rt, rq);
6617 #ifdef CONFIG_FAIR_GROUP_SCHED
6618 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6619 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6621 * How much cpu bandwidth does root_task_group get?
6623 * In case of task-groups formed thr' the cgroup filesystem, it
6624 * gets 100% of the cpu resources in the system. This overall
6625 * system cpu resource is divided among the tasks of
6626 * root_task_group and its child task-groups in a fair manner,
6627 * based on each entity's (task or task-group's) weight
6628 * (se->load.weight).
6630 * In other words, if root_task_group has 10 tasks of weight
6631 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6632 * then A0's share of the cpu resource is:
6634 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6636 * We achieve this by letting root_task_group's tasks sit
6637 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6639 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6640 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6641 #endif /* CONFIG_FAIR_GROUP_SCHED */
6643 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6644 #ifdef CONFIG_RT_GROUP_SCHED
6645 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6646 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6649 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6650 rq->cpu_load[j] = 0;
6652 rq->last_load_update_tick = jiffies;
6657 rq->cpu_power = SCHED_POWER_SCALE;
6658 rq->post_schedule = 0;
6659 rq->active_balance = 0;
6660 rq->next_balance = jiffies;
6665 rq->avg_idle = 2*sysctl_sched_migration_cost;
6666 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6668 INIT_LIST_HEAD(&rq->cfs_tasks);
6670 rq_attach_root(rq, &def_root_domain);
6671 #ifdef CONFIG_NO_HZ_COMMON
6674 #ifdef CONFIG_NO_HZ_FULL
6675 rq->last_sched_tick = 0;
6679 atomic_set(&rq->nr_iowait, 0);
6682 set_load_weight(&init_task);
6684 #ifdef CONFIG_PREEMPT_NOTIFIERS
6685 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6688 #ifdef CONFIG_RT_MUTEXES
6689 plist_head_init(&init_task.pi_waiters);
6693 * The boot idle thread does lazy MMU switching as well:
6695 atomic_inc(&init_mm.mm_count);
6696 enter_lazy_tlb(&init_mm, current);
6699 * Make us the idle thread. Technically, schedule() should not be
6700 * called from this thread, however somewhere below it might be,
6701 * but because we are the idle thread, we just pick up running again
6702 * when this runqueue becomes "idle".
6704 init_idle(current, smp_processor_id());
6706 calc_load_update = jiffies + LOAD_FREQ;
6709 * During early bootup we pretend to be a normal task:
6711 current->sched_class = &fair_sched_class;
6714 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6715 /* May be allocated at isolcpus cmdline parse time */
6716 if (cpu_isolated_map == NULL)
6717 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6718 idle_thread_set_boot_cpu();
6720 init_sched_fair_class();
6722 scheduler_running = 1;
6725 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6726 static inline int preempt_count_equals(int preempt_offset)
6728 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6730 return (nested == preempt_offset);
6733 void __might_sleep(const char *file, int line, int preempt_offset)
6735 static unsigned long prev_jiffy; /* ratelimiting */
6737 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6738 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6739 system_state != SYSTEM_RUNNING || oops_in_progress)
6741 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6743 prev_jiffy = jiffies;
6746 "BUG: sleeping function called from invalid context at %s:%d\n",
6749 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6750 in_atomic(), irqs_disabled(),
6751 current->pid, current->comm);
6753 debug_show_held_locks(current);
6754 if (irqs_disabled())
6755 print_irqtrace_events(current);
6758 EXPORT_SYMBOL(__might_sleep);
6761 #ifdef CONFIG_MAGIC_SYSRQ
6762 static void normalize_task(struct rq *rq, struct task_struct *p)
6764 const struct sched_class *prev_class = p->sched_class;
6765 int old_prio = p->prio;
6770 dequeue_task(rq, p, 0);
6771 __setscheduler(rq, p, SCHED_NORMAL, 0);
6773 enqueue_task(rq, p, 0);
6774 resched_task(rq->curr);
6777 check_class_changed(rq, p, prev_class, old_prio);
6780 void normalize_rt_tasks(void)
6782 struct task_struct *g, *p;
6783 unsigned long flags;
6786 read_lock_irqsave(&tasklist_lock, flags);
6787 do_each_thread(g, p) {
6789 * Only normalize user tasks:
6794 p->se.exec_start = 0;
6795 #ifdef CONFIG_SCHEDSTATS
6796 p->se.statistics.wait_start = 0;
6797 p->se.statistics.sleep_start = 0;
6798 p->se.statistics.block_start = 0;
6803 * Renice negative nice level userspace
6806 if (TASK_NICE(p) < 0 && p->mm)
6807 set_user_nice(p, 0);
6811 raw_spin_lock(&p->pi_lock);
6812 rq = __task_rq_lock(p);
6814 normalize_task(rq, p);
6816 __task_rq_unlock(rq);
6817 raw_spin_unlock(&p->pi_lock);
6818 } while_each_thread(g, p);
6820 read_unlock_irqrestore(&tasklist_lock, flags);
6823 #endif /* CONFIG_MAGIC_SYSRQ */
6825 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6827 * These functions are only useful for the IA64 MCA handling, or kdb.
6829 * They can only be called when the whole system has been
6830 * stopped - every CPU needs to be quiescent, and no scheduling
6831 * activity can take place. Using them for anything else would
6832 * be a serious bug, and as a result, they aren't even visible
6833 * under any other configuration.
6837 * curr_task - return the current task for a given cpu.
6838 * @cpu: the processor in question.
6840 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6842 * Return: The current task for @cpu.
6844 struct task_struct *curr_task(int cpu)
6846 return cpu_curr(cpu);
6849 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6853 * set_curr_task - set the current task for a given cpu.
6854 * @cpu: the processor in question.
6855 * @p: the task pointer to set.
6857 * Description: This function must only be used when non-maskable interrupts
6858 * are serviced on a separate stack. It allows the architecture to switch the
6859 * notion of the current task on a cpu in a non-blocking manner. This function
6860 * must be called with all CPU's synchronized, and interrupts disabled, the
6861 * and caller must save the original value of the current task (see
6862 * curr_task() above) and restore that value before reenabling interrupts and
6863 * re-starting the system.
6865 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6867 void set_curr_task(int cpu, struct task_struct *p)
6874 #ifdef CONFIG_CGROUP_SCHED
6875 /* task_group_lock serializes the addition/removal of task groups */
6876 static DEFINE_SPINLOCK(task_group_lock);
6878 static void free_sched_group(struct task_group *tg)
6880 free_fair_sched_group(tg);
6881 free_rt_sched_group(tg);
6886 /* allocate runqueue etc for a new task group */
6887 struct task_group *sched_create_group(struct task_group *parent)
6889 struct task_group *tg;
6891 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6893 return ERR_PTR(-ENOMEM);
6895 if (!alloc_fair_sched_group(tg, parent))
6898 if (!alloc_rt_sched_group(tg, parent))
6904 free_sched_group(tg);
6905 return ERR_PTR(-ENOMEM);
6908 void sched_online_group(struct task_group *tg, struct task_group *parent)
6910 unsigned long flags;
6912 spin_lock_irqsave(&task_group_lock, flags);
6913 list_add_rcu(&tg->list, &task_groups);
6915 WARN_ON(!parent); /* root should already exist */
6917 tg->parent = parent;
6918 INIT_LIST_HEAD(&tg->children);
6919 list_add_rcu(&tg->siblings, &parent->children);
6920 spin_unlock_irqrestore(&task_group_lock, flags);
6923 /* rcu callback to free various structures associated with a task group */
6924 static void free_sched_group_rcu(struct rcu_head *rhp)
6926 /* now it should be safe to free those cfs_rqs */
6927 free_sched_group(container_of(rhp, struct task_group, rcu));
6930 /* Destroy runqueue etc associated with a task group */
6931 void sched_destroy_group(struct task_group *tg)
6933 /* wait for possible concurrent references to cfs_rqs complete */
6934 call_rcu(&tg->rcu, free_sched_group_rcu);
6937 void sched_offline_group(struct task_group *tg)
6939 unsigned long flags;
6942 /* end participation in shares distribution */
6943 for_each_possible_cpu(i)
6944 unregister_fair_sched_group(tg, i);
6946 spin_lock_irqsave(&task_group_lock, flags);
6947 list_del_rcu(&tg->list);
6948 list_del_rcu(&tg->siblings);
6949 spin_unlock_irqrestore(&task_group_lock, flags);
6952 /* change task's runqueue when it moves between groups.
6953 * The caller of this function should have put the task in its new group
6954 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6955 * reflect its new group.
6957 void sched_move_task(struct task_struct *tsk)
6959 struct task_group *tg;
6961 unsigned long flags;
6964 rq = task_rq_lock(tsk, &flags);
6966 running = task_current(rq, tsk);
6970 dequeue_task(rq, tsk, 0);
6971 if (unlikely(running))
6972 tsk->sched_class->put_prev_task(rq, tsk);
6974 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6975 lockdep_is_held(&tsk->sighand->siglock)),
6976 struct task_group, css);
6977 tg = autogroup_task_group(tsk, tg);
6978 tsk->sched_task_group = tg;
6980 #ifdef CONFIG_FAIR_GROUP_SCHED
6981 if (tsk->sched_class->task_move_group)
6982 tsk->sched_class->task_move_group(tsk, on_rq);
6985 set_task_rq(tsk, task_cpu(tsk));
6987 if (unlikely(running))
6988 tsk->sched_class->set_curr_task(rq);
6990 enqueue_task(rq, tsk, 0);
6992 task_rq_unlock(rq, tsk, &flags);
6994 #endif /* CONFIG_CGROUP_SCHED */
6996 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6997 static unsigned long to_ratio(u64 period, u64 runtime)
6999 if (runtime == RUNTIME_INF)
7002 return div64_u64(runtime << 20, period);
7006 #ifdef CONFIG_RT_GROUP_SCHED
7008 * Ensure that the real time constraints are schedulable.
7010 static DEFINE_MUTEX(rt_constraints_mutex);
7012 /* Must be called with tasklist_lock held */
7013 static inline int tg_has_rt_tasks(struct task_group *tg)
7015 struct task_struct *g, *p;
7017 do_each_thread(g, p) {
7018 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7020 } while_each_thread(g, p);
7025 struct rt_schedulable_data {
7026 struct task_group *tg;
7031 static int tg_rt_schedulable(struct task_group *tg, void *data)
7033 struct rt_schedulable_data *d = data;
7034 struct task_group *child;
7035 unsigned long total, sum = 0;
7036 u64 period, runtime;
7038 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7039 runtime = tg->rt_bandwidth.rt_runtime;
7042 period = d->rt_period;
7043 runtime = d->rt_runtime;
7047 * Cannot have more runtime than the period.
7049 if (runtime > period && runtime != RUNTIME_INF)
7053 * Ensure we don't starve existing RT tasks.
7055 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7058 total = to_ratio(period, runtime);
7061 * Nobody can have more than the global setting allows.
7063 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7067 * The sum of our children's runtime should not exceed our own.
7069 list_for_each_entry_rcu(child, &tg->children, siblings) {
7070 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7071 runtime = child->rt_bandwidth.rt_runtime;
7073 if (child == d->tg) {
7074 period = d->rt_period;
7075 runtime = d->rt_runtime;
7078 sum += to_ratio(period, runtime);
7087 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7091 struct rt_schedulable_data data = {
7093 .rt_period = period,
7094 .rt_runtime = runtime,
7098 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7104 static int tg_set_rt_bandwidth(struct task_group *tg,
7105 u64 rt_period, u64 rt_runtime)
7109 mutex_lock(&rt_constraints_mutex);
7110 read_lock(&tasklist_lock);
7111 err = __rt_schedulable(tg, rt_period, rt_runtime);
7115 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7116 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7117 tg->rt_bandwidth.rt_runtime = rt_runtime;
7119 for_each_possible_cpu(i) {
7120 struct rt_rq *rt_rq = tg->rt_rq[i];
7122 raw_spin_lock(&rt_rq->rt_runtime_lock);
7123 rt_rq->rt_runtime = rt_runtime;
7124 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7126 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7128 read_unlock(&tasklist_lock);
7129 mutex_unlock(&rt_constraints_mutex);
7134 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7136 u64 rt_runtime, rt_period;
7138 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7139 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7140 if (rt_runtime_us < 0)
7141 rt_runtime = RUNTIME_INF;
7143 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7146 static long sched_group_rt_runtime(struct task_group *tg)
7150 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7153 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7154 do_div(rt_runtime_us, NSEC_PER_USEC);
7155 return rt_runtime_us;
7158 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7160 u64 rt_runtime, rt_period;
7162 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7163 rt_runtime = tg->rt_bandwidth.rt_runtime;
7168 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7171 static long sched_group_rt_period(struct task_group *tg)
7175 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7176 do_div(rt_period_us, NSEC_PER_USEC);
7177 return rt_period_us;
7180 static int sched_rt_global_constraints(void)
7182 u64 runtime, period;
7185 if (sysctl_sched_rt_period <= 0)
7188 runtime = global_rt_runtime();
7189 period = global_rt_period();
7192 * Sanity check on the sysctl variables.
7194 if (runtime > period && runtime != RUNTIME_INF)
7197 mutex_lock(&rt_constraints_mutex);
7198 read_lock(&tasklist_lock);
7199 ret = __rt_schedulable(NULL, 0, 0);
7200 read_unlock(&tasklist_lock);
7201 mutex_unlock(&rt_constraints_mutex);
7206 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7208 /* Don't accept realtime tasks when there is no way for them to run */
7209 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7215 #else /* !CONFIG_RT_GROUP_SCHED */
7216 static int sched_rt_global_constraints(void)
7218 unsigned long flags;
7221 if (sysctl_sched_rt_period <= 0)
7225 * There's always some RT tasks in the root group
7226 * -- migration, kstopmachine etc..
7228 if (sysctl_sched_rt_runtime == 0)
7231 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7232 for_each_possible_cpu(i) {
7233 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7235 raw_spin_lock(&rt_rq->rt_runtime_lock);
7236 rt_rq->rt_runtime = global_rt_runtime();
7237 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7239 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7243 #endif /* CONFIG_RT_GROUP_SCHED */
7245 int sched_rr_handler(struct ctl_table *table, int write,
7246 void __user *buffer, size_t *lenp,
7250 static DEFINE_MUTEX(mutex);
7253 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7254 /* make sure that internally we keep jiffies */
7255 /* also, writing zero resets timeslice to default */
7256 if (!ret && write) {
7257 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7258 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7260 mutex_unlock(&mutex);
7264 int sched_rt_handler(struct ctl_table *table, int write,
7265 void __user *buffer, size_t *lenp,
7269 int old_period, old_runtime;
7270 static DEFINE_MUTEX(mutex);
7273 old_period = sysctl_sched_rt_period;
7274 old_runtime = sysctl_sched_rt_runtime;
7276 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7278 if (!ret && write) {
7279 ret = sched_rt_global_constraints();
7281 sysctl_sched_rt_period = old_period;
7282 sysctl_sched_rt_runtime = old_runtime;
7284 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7285 def_rt_bandwidth.rt_period =
7286 ns_to_ktime(global_rt_period());
7289 mutex_unlock(&mutex);
7294 #ifdef CONFIG_CGROUP_SCHED
7296 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7298 return css ? container_of(css, struct task_group, css) : NULL;
7301 static struct cgroup_subsys_state *
7302 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7304 struct task_group *parent = css_tg(parent_css);
7305 struct task_group *tg;
7308 /* This is early initialization for the top cgroup */
7309 return &root_task_group.css;
7312 tg = sched_create_group(parent);
7314 return ERR_PTR(-ENOMEM);
7319 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7321 struct task_group *tg = css_tg(css);
7322 struct task_group *parent = css_tg(css_parent(css));
7325 sched_online_group(tg, parent);
7329 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7331 struct task_group *tg = css_tg(css);
7333 sched_destroy_group(tg);
7336 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7338 struct task_group *tg = css_tg(css);
7340 sched_offline_group(tg);
7343 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7344 struct cgroup_taskset *tset)
7346 struct task_struct *task;
7348 cgroup_taskset_for_each(task, css, tset) {
7349 #ifdef CONFIG_RT_GROUP_SCHED
7350 if (!sched_rt_can_attach(css_tg(css), task))
7353 /* We don't support RT-tasks being in separate groups */
7354 if (task->sched_class != &fair_sched_class)
7361 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7362 struct cgroup_taskset *tset)
7364 struct task_struct *task;
7366 cgroup_taskset_for_each(task, css, tset)
7367 sched_move_task(task);
7370 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7371 struct cgroup_subsys_state *old_css,
7372 struct task_struct *task)
7375 * cgroup_exit() is called in the copy_process() failure path.
7376 * Ignore this case since the task hasn't ran yet, this avoids
7377 * trying to poke a half freed task state from generic code.
7379 if (!(task->flags & PF_EXITING))
7382 sched_move_task(task);
7385 #ifdef CONFIG_FAIR_GROUP_SCHED
7386 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7387 struct cftype *cftype, u64 shareval)
7389 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7392 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7395 struct task_group *tg = css_tg(css);
7397 return (u64) scale_load_down(tg->shares);
7400 #ifdef CONFIG_CFS_BANDWIDTH
7401 static DEFINE_MUTEX(cfs_constraints_mutex);
7403 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7404 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7406 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7408 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7410 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7411 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7413 if (tg == &root_task_group)
7417 * Ensure we have at some amount of bandwidth every period. This is
7418 * to prevent reaching a state of large arrears when throttled via
7419 * entity_tick() resulting in prolonged exit starvation.
7421 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7425 * Likewise, bound things on the otherside by preventing insane quota
7426 * periods. This also allows us to normalize in computing quota
7429 if (period > max_cfs_quota_period)
7432 mutex_lock(&cfs_constraints_mutex);
7433 ret = __cfs_schedulable(tg, period, quota);
7437 runtime_enabled = quota != RUNTIME_INF;
7438 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7439 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7440 raw_spin_lock_irq(&cfs_b->lock);
7441 cfs_b->period = ns_to_ktime(period);
7442 cfs_b->quota = quota;
7444 __refill_cfs_bandwidth_runtime(cfs_b);
7445 /* restart the period timer (if active) to handle new period expiry */
7446 if (runtime_enabled && cfs_b->timer_active) {
7447 /* force a reprogram */
7448 cfs_b->timer_active = 0;
7449 __start_cfs_bandwidth(cfs_b);
7451 raw_spin_unlock_irq(&cfs_b->lock);
7453 for_each_possible_cpu(i) {
7454 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7455 struct rq *rq = cfs_rq->rq;
7457 raw_spin_lock_irq(&rq->lock);
7458 cfs_rq->runtime_enabled = runtime_enabled;
7459 cfs_rq->runtime_remaining = 0;
7461 if (cfs_rq->throttled)
7462 unthrottle_cfs_rq(cfs_rq);
7463 raw_spin_unlock_irq(&rq->lock);
7466 mutex_unlock(&cfs_constraints_mutex);
7471 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7475 period = ktime_to_ns(tg->cfs_bandwidth.period);
7476 if (cfs_quota_us < 0)
7477 quota = RUNTIME_INF;
7479 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7481 return tg_set_cfs_bandwidth(tg, period, quota);
7484 long tg_get_cfs_quota(struct task_group *tg)
7488 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7491 quota_us = tg->cfs_bandwidth.quota;
7492 do_div(quota_us, NSEC_PER_USEC);
7497 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7501 period = (u64)cfs_period_us * NSEC_PER_USEC;
7502 quota = tg->cfs_bandwidth.quota;
7504 return tg_set_cfs_bandwidth(tg, period, quota);
7507 long tg_get_cfs_period(struct task_group *tg)
7511 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7512 do_div(cfs_period_us, NSEC_PER_USEC);
7514 return cfs_period_us;
7517 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7520 return tg_get_cfs_quota(css_tg(css));
7523 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7524 struct cftype *cftype, s64 cfs_quota_us)
7526 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7529 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7532 return tg_get_cfs_period(css_tg(css));
7535 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7536 struct cftype *cftype, u64 cfs_period_us)
7538 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7541 struct cfs_schedulable_data {
7542 struct task_group *tg;
7547 * normalize group quota/period to be quota/max_period
7548 * note: units are usecs
7550 static u64 normalize_cfs_quota(struct task_group *tg,
7551 struct cfs_schedulable_data *d)
7559 period = tg_get_cfs_period(tg);
7560 quota = tg_get_cfs_quota(tg);
7563 /* note: these should typically be equivalent */
7564 if (quota == RUNTIME_INF || quota == -1)
7567 return to_ratio(period, quota);
7570 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7572 struct cfs_schedulable_data *d = data;
7573 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7574 s64 quota = 0, parent_quota = -1;
7577 quota = RUNTIME_INF;
7579 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7581 quota = normalize_cfs_quota(tg, d);
7582 parent_quota = parent_b->hierarchal_quota;
7585 * ensure max(child_quota) <= parent_quota, inherit when no
7588 if (quota == RUNTIME_INF)
7589 quota = parent_quota;
7590 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7593 cfs_b->hierarchal_quota = quota;
7598 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7601 struct cfs_schedulable_data data = {
7607 if (quota != RUNTIME_INF) {
7608 do_div(data.period, NSEC_PER_USEC);
7609 do_div(data.quota, NSEC_PER_USEC);
7613 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7619 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7620 struct cgroup_map_cb *cb)
7622 struct task_group *tg = css_tg(css);
7623 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7625 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7626 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7627 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7631 #endif /* CONFIG_CFS_BANDWIDTH */
7632 #endif /* CONFIG_FAIR_GROUP_SCHED */
7634 #ifdef CONFIG_RT_GROUP_SCHED
7635 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7636 struct cftype *cft, s64 val)
7638 return sched_group_set_rt_runtime(css_tg(css), val);
7641 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7644 return sched_group_rt_runtime(css_tg(css));
7647 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7648 struct cftype *cftype, u64 rt_period_us)
7650 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7653 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7656 return sched_group_rt_period(css_tg(css));
7658 #endif /* CONFIG_RT_GROUP_SCHED */
7660 static struct cftype cpu_files[] = {
7661 #ifdef CONFIG_FAIR_GROUP_SCHED
7664 .read_u64 = cpu_shares_read_u64,
7665 .write_u64 = cpu_shares_write_u64,
7668 #ifdef CONFIG_CFS_BANDWIDTH
7670 .name = "cfs_quota_us",
7671 .read_s64 = cpu_cfs_quota_read_s64,
7672 .write_s64 = cpu_cfs_quota_write_s64,
7675 .name = "cfs_period_us",
7676 .read_u64 = cpu_cfs_period_read_u64,
7677 .write_u64 = cpu_cfs_period_write_u64,
7681 .read_map = cpu_stats_show,
7684 #ifdef CONFIG_RT_GROUP_SCHED
7686 .name = "rt_runtime_us",
7687 .read_s64 = cpu_rt_runtime_read,
7688 .write_s64 = cpu_rt_runtime_write,
7691 .name = "rt_period_us",
7692 .read_u64 = cpu_rt_period_read_uint,
7693 .write_u64 = cpu_rt_period_write_uint,
7699 struct cgroup_subsys cpu_cgroup_subsys = {
7701 .css_alloc = cpu_cgroup_css_alloc,
7702 .css_free = cpu_cgroup_css_free,
7703 .css_online = cpu_cgroup_css_online,
7704 .css_offline = cpu_cgroup_css_offline,
7705 .can_attach = cpu_cgroup_can_attach,
7706 .attach = cpu_cgroup_attach,
7707 .exit = cpu_cgroup_exit,
7708 .subsys_id = cpu_cgroup_subsys_id,
7709 .base_cftypes = cpu_files,
7713 #endif /* CONFIG_CGROUP_SCHED */
7715 void dump_cpu_task(int cpu)
7717 pr_info("Task dump for CPU %d:\n", cpu);
7718 sched_show_task(cpu_curr(cpu));