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.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 assert_raw_spin_locked(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p))
533 smp_send_reschedule(cpu);
536 void resched_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
541 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 resched_task(cpu_curr(cpu));
544 raw_spin_unlock_irqrestore(&rq->lock, flags);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu = smp_processor_id();
560 struct sched_domain *sd;
563 for_each_domain(cpu, sd) {
564 for_each_cpu(i, sched_domain_span(sd)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 static void wake_up_idle_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
589 if (cpu == smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq->curr != rq->idle)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq->idle);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq->idle))
612 smp_send_reschedule(cpu);
615 static bool wake_up_full_nohz_cpu(int cpu)
617 if (tick_nohz_full_cpu(cpu)) {
618 if (cpu != smp_processor_id() ||
619 tick_nohz_tick_stopped())
620 smp_send_reschedule(cpu);
627 void wake_up_nohz_cpu(int cpu)
629 if (!wake_up_full_nohz_cpu(cpu))
630 wake_up_idle_cpu(cpu);
633 static inline bool got_nohz_idle_kick(void)
635 int cpu = smp_processor_id();
637 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
640 if (idle_cpu(cpu) && !need_resched())
644 * We can't run Idle Load Balance on this CPU for this time so we
645 * cancel it and clear NOHZ_BALANCE_KICK
647 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
651 #else /* CONFIG_NO_HZ_COMMON */
653 static inline bool got_nohz_idle_kick(void)
658 #endif /* CONFIG_NO_HZ_COMMON */
660 #ifdef CONFIG_NO_HZ_FULL
661 bool sched_can_stop_tick(void)
667 /* Make sure rq->nr_running update is visible after the IPI */
670 /* More than one running task need preemption */
671 if (rq->nr_running > 1)
676 #endif /* CONFIG_NO_HZ_FULL */
678 void sched_avg_update(struct rq *rq)
680 s64 period = sched_avg_period();
682 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
684 * Inline assembly required to prevent the compiler
685 * optimising this loop into a divmod call.
686 * See __iter_div_u64_rem() for another example of this.
688 asm("" : "+rm" (rq->age_stamp));
689 rq->age_stamp += period;
694 #else /* !CONFIG_SMP */
695 void resched_task(struct task_struct *p)
697 assert_raw_spin_locked(&task_rq(p)->lock);
698 set_tsk_need_resched(p);
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
713 struct task_group *parent, *child;
719 ret = (*down)(parent, data);
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
734 parent = parent->parent;
741 int tg_nop(struct task_group *tg, void *data)
747 static void set_load_weight(struct task_struct *p)
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p->policy == SCHED_IDLE) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
761 load->weight = scale_load(prio_to_weight[prio]);
762 load->inv_weight = prio_to_wmult[prio];
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
768 sched_info_queued(p);
769 p->sched_class->enqueue_task(rq, p, flags);
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
775 sched_info_dequeued(p);
776 p->sched_class->dequeue_task(rq, p, flags);
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible--;
784 enqueue_task(rq, p, flags);
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
789 if (task_contributes_to_load(p))
790 rq->nr_uninterruptible++;
792 dequeue_task(rq, p, flags);
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal = 0, irq_delta = 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta > delta)
825 rq->prev_irq_time += irq_delta;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled))) {
832 steal = paravirt_steal_clock(cpu_of(rq));
833 steal -= rq->prev_steal_time_rq;
835 if (unlikely(steal > delta))
838 st = steal_ticks(steal);
839 steal = st * TICK_NSEC;
841 rq->prev_steal_time_rq += steal;
847 rq->clock_task += delta;
849 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
850 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
851 sched_rt_avg_update(rq, irq_delta + steal);
855 void sched_set_stop_task(int cpu, struct task_struct *stop)
857 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
858 struct task_struct *old_stop = cpu_rq(cpu)->stop;
862 * Make it appear like a SCHED_FIFO task, its something
863 * userspace knows about and won't get confused about.
865 * Also, it will make PI more or less work without too
866 * much confusion -- but then, stop work should not
867 * rely on PI working anyway.
869 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
871 stop->sched_class = &stop_sched_class;
874 cpu_rq(cpu)->stop = stop;
878 * Reset it back to a normal scheduling class so that
879 * it can die in pieces.
881 old_stop->sched_class = &rt_sched_class;
886 * __normal_prio - return the priority that is based on the static prio
888 static inline int __normal_prio(struct task_struct *p)
890 return p->static_prio;
894 * Calculate the expected normal priority: i.e. priority
895 * without taking RT-inheritance into account. Might be
896 * boosted by interactivity modifiers. Changes upon fork,
897 * setprio syscalls, and whenever the interactivity
898 * estimator recalculates.
900 static inline int normal_prio(struct task_struct *p)
904 if (task_has_rt_policy(p))
905 prio = MAX_RT_PRIO-1 - p->rt_priority;
907 prio = __normal_prio(p);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct *p)
920 p->normal_prio = normal_prio(p);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p->prio))
927 return p->normal_prio;
932 * task_curr - is this task currently executing on a CPU?
933 * @p: the task in question.
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;
977 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
979 void register_task_migration_notifier(struct notifier_block *n)
981 atomic_notifier_chain_register(&task_migration_notifier, n);
985 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
987 #ifdef CONFIG_SCHED_DEBUG
989 * We should never call set_task_cpu() on a blocked task,
990 * ttwu() will sort out the placement.
992 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
993 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
995 #ifdef CONFIG_LOCKDEP
997 * The caller should hold either p->pi_lock or rq->lock, when changing
998 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1000 * sched_move_task() holds both and thus holding either pins the cgroup,
1003 * Furthermore, all task_rq users should acquire both locks, see
1006 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1007 lockdep_is_held(&task_rq(p)->lock)));
1011 trace_sched_migrate_task(p, new_cpu);
1013 if (task_cpu(p) != new_cpu) {
1014 struct task_migration_notifier tmn;
1016 if (p->sched_class->migrate_task_rq)
1017 p->sched_class->migrate_task_rq(p, new_cpu);
1018 p->se.nr_migrations++;
1019 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1022 tmn.from_cpu = task_cpu(p);
1023 tmn.to_cpu = new_cpu;
1025 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1028 __set_task_cpu(p, new_cpu);
1031 struct migration_arg {
1032 struct task_struct *task;
1036 static int migration_cpu_stop(void *data);
1039 * wait_task_inactive - wait for a thread to unschedule.
1041 * If @match_state is nonzero, it's the @p->state value just checked and
1042 * not expected to change. If it changes, i.e. @p might have woken up,
1043 * then return zero. When we succeed in waiting for @p to be off its CPU,
1044 * we return a positive number (its total switch count). If a second call
1045 * a short while later returns the same number, the caller can be sure that
1046 * @p has remained unscheduled the whole time.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1056 unsigned long flags;
1063 * We do the initial early heuristics without holding
1064 * any task-queue locks at all. We'll only try to get
1065 * the runqueue lock when things look like they will
1071 * If the task is actively running on another CPU
1072 * still, just relax and busy-wait without holding
1075 * NOTE! Since we don't hold any locks, it's not
1076 * even sure that "rq" stays as the right runqueue!
1077 * But we don't care, since "task_running()" will
1078 * return false if the runqueue has changed and p
1079 * is actually now running somewhere else!
1081 while (task_running(rq, p)) {
1082 if (match_state && unlikely(p->state != match_state))
1088 * Ok, time to look more closely! We need the rq
1089 * lock now, to be *sure*. If we're wrong, we'll
1090 * just go back and repeat.
1092 rq = task_rq_lock(p, &flags);
1093 trace_sched_wait_task(p);
1094 running = task_running(rq, p);
1097 if (!match_state || p->state == match_state)
1098 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1099 task_rq_unlock(rq, p, &flags);
1102 * If it changed from the expected state, bail out now.
1104 if (unlikely(!ncsw))
1108 * Was it really running after all now that we
1109 * checked with the proper locks actually held?
1111 * Oops. Go back and try again..
1113 if (unlikely(running)) {
1119 * It's not enough that it's not actively running,
1120 * it must be off the runqueue _entirely_, and not
1123 * So if it was still runnable (but just not actively
1124 * running right now), it's preempted, and we should
1125 * yield - it could be a while.
1127 if (unlikely(on_rq)) {
1128 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1130 set_current_state(TASK_UNINTERRUPTIBLE);
1131 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1136 * Ahh, all good. It wasn't running, and it wasn't
1137 * runnable, which means that it will never become
1138 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesn't have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct *p)
1165 if ((cpu != smp_processor_id()) && task_curr(p))
1166 smp_send_reschedule(cpu);
1169 EXPORT_SYMBOL_GPL(kick_process);
1170 #endif /* CONFIG_SMP */
1174 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1176 static int select_fallback_rq(int cpu, struct task_struct *p)
1178 int nid = cpu_to_node(cpu);
1179 const struct cpumask *nodemask = NULL;
1180 enum { cpuset, possible, fail } state = cpuset;
1184 * If the node that the cpu is on has been offlined, cpu_to_node()
1185 * will return -1. There is no cpu on the node, and we should
1186 * select the cpu on the other node.
1189 nodemask = cpumask_of_node(nid);
1191 /* Look for allowed, online CPU in same node. */
1192 for_each_cpu(dest_cpu, nodemask) {
1193 if (!cpu_online(dest_cpu))
1195 if (!cpu_active(dest_cpu))
1197 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1203 /* Any allowed, online CPU? */
1204 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1205 if (!cpu_online(dest_cpu))
1207 if (!cpu_active(dest_cpu))
1214 /* No more Mr. Nice Guy. */
1215 cpuset_cpus_allowed_fallback(p);
1220 do_set_cpus_allowed(p, cpu_possible_mask);
1231 if (state != cpuset) {
1233 * Don't tell them about moving exiting tasks or
1234 * kernel threads (both mm NULL), since they never
1237 if (p->mm && printk_ratelimit()) {
1238 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1239 task_pid_nr(p), p->comm, cpu);
1247 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1250 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1252 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1255 * In order not to call set_task_cpu() on a blocking task we need
1256 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1259 * Since this is common to all placement strategies, this lives here.
1261 * [ this allows ->select_task() to simply return task_cpu(p) and
1262 * not worry about this generic constraint ]
1264 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1266 cpu = select_fallback_rq(task_cpu(p), p);
1271 static void update_avg(u64 *avg, u64 sample)
1273 s64 diff = sample - *avg;
1279 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1281 #ifdef CONFIG_SCHEDSTATS
1282 struct rq *rq = this_rq();
1285 int this_cpu = smp_processor_id();
1287 if (cpu == this_cpu) {
1288 schedstat_inc(rq, ttwu_local);
1289 schedstat_inc(p, se.statistics.nr_wakeups_local);
1291 struct sched_domain *sd;
1293 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1295 for_each_domain(this_cpu, sd) {
1296 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1297 schedstat_inc(sd, ttwu_wake_remote);
1304 if (wake_flags & WF_MIGRATED)
1305 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1307 #endif /* CONFIG_SMP */
1309 schedstat_inc(rq, ttwu_count);
1310 schedstat_inc(p, se.statistics.nr_wakeups);
1312 if (wake_flags & WF_SYNC)
1313 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1315 #endif /* CONFIG_SCHEDSTATS */
1318 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1320 activate_task(rq, p, en_flags);
1323 /* if a worker is waking up, notify workqueue */
1324 if (p->flags & PF_WQ_WORKER)
1325 wq_worker_waking_up(p, cpu_of(rq));
1329 * Mark the task runnable and perform wakeup-preemption.
1332 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1334 check_preempt_curr(rq, p, wake_flags);
1335 trace_sched_wakeup(p, true);
1337 p->state = TASK_RUNNING;
1339 if (p->sched_class->task_woken)
1340 p->sched_class->task_woken(rq, p);
1342 if (rq->idle_stamp) {
1343 u64 delta = rq_clock(rq) - rq->idle_stamp;
1344 u64 max = 2*sysctl_sched_migration_cost;
1349 update_avg(&rq->avg_idle, delta);
1356 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1359 if (p->sched_contributes_to_load)
1360 rq->nr_uninterruptible--;
1363 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1364 ttwu_do_wakeup(rq, p, wake_flags);
1368 * Called in case the task @p isn't fully descheduled from its runqueue,
1369 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1370 * since all we need to do is flip p->state to TASK_RUNNING, since
1371 * the task is still ->on_rq.
1373 static int ttwu_remote(struct task_struct *p, int wake_flags)
1378 rq = __task_rq_lock(p);
1380 /* check_preempt_curr() may use rq clock */
1381 update_rq_clock(rq);
1382 ttwu_do_wakeup(rq, p, wake_flags);
1385 __task_rq_unlock(rq);
1391 static void sched_ttwu_pending(void)
1393 struct rq *rq = this_rq();
1394 struct llist_node *llist = llist_del_all(&rq->wake_list);
1395 struct task_struct *p;
1397 raw_spin_lock(&rq->lock);
1400 p = llist_entry(llist, struct task_struct, wake_entry);
1401 llist = llist_next(llist);
1402 ttwu_do_activate(rq, p, 0);
1405 raw_spin_unlock(&rq->lock);
1408 void scheduler_ipi(void)
1410 if (llist_empty(&this_rq()->wake_list)
1411 && !tick_nohz_full_cpu(smp_processor_id())
1412 && !got_nohz_idle_kick())
1416 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1417 * traditionally all their work was done from the interrupt return
1418 * path. Now that we actually do some work, we need to make sure
1421 * Some archs already do call them, luckily irq_enter/exit nest
1424 * Arguably we should visit all archs and update all handlers,
1425 * however a fair share of IPIs are still resched only so this would
1426 * somewhat pessimize the simple resched case.
1429 tick_nohz_full_check();
1430 sched_ttwu_pending();
1433 * Check if someone kicked us for doing the nohz idle load balance.
1435 if (unlikely(got_nohz_idle_kick())) {
1436 this_rq()->idle_balance = 1;
1437 raise_softirq_irqoff(SCHED_SOFTIRQ);
1442 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1444 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1445 smp_send_reschedule(cpu);
1448 bool cpus_share_cache(int this_cpu, int that_cpu)
1450 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1452 #endif /* CONFIG_SMP */
1454 static void ttwu_queue(struct task_struct *p, int cpu)
1456 struct rq *rq = cpu_rq(cpu);
1458 #if defined(CONFIG_SMP)
1459 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1460 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1461 ttwu_queue_remote(p, cpu);
1466 raw_spin_lock(&rq->lock);
1467 ttwu_do_activate(rq, p, 0);
1468 raw_spin_unlock(&rq->lock);
1472 * try_to_wake_up - wake up a thread
1473 * @p: the thread to be awakened
1474 * @state: the mask of task states that can be woken
1475 * @wake_flags: wake modifier flags (WF_*)
1477 * Put it on the run-queue if it's not already there. The "current"
1478 * thread is always on the run-queue (except when the actual
1479 * re-schedule is in progress), and as such you're allowed to do
1480 * the simpler "current->state = TASK_RUNNING" to mark yourself
1481 * runnable without the overhead of this.
1483 * Returns %true if @p was woken up, %false if it was already running
1484 * or @state didn't match @p's state.
1487 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1489 unsigned long flags;
1490 int cpu, success = 0;
1493 raw_spin_lock_irqsave(&p->pi_lock, flags);
1494 if (!(p->state & state))
1497 success = 1; /* we're going to change ->state */
1500 if (p->on_rq && ttwu_remote(p, wake_flags))
1505 * If the owning (remote) cpu is still in the middle of schedule() with
1506 * this task as prev, wait until its done referencing the task.
1511 * Pairs with the smp_wmb() in finish_lock_switch().
1515 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1516 p->state = TASK_WAKING;
1518 if (p->sched_class->task_waking)
1519 p->sched_class->task_waking(p);
1521 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1522 if (task_cpu(p) != cpu) {
1523 wake_flags |= WF_MIGRATED;
1524 set_task_cpu(p, cpu);
1526 #endif /* CONFIG_SMP */
1530 ttwu_stat(p, cpu, wake_flags);
1532 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1538 * try_to_wake_up_local - try to wake up a local task with rq lock held
1539 * @p: the thread to be awakened
1541 * Put @p on the run-queue if it's not already there. The caller must
1542 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1545 static void try_to_wake_up_local(struct task_struct *p)
1547 struct rq *rq = task_rq(p);
1549 if (WARN_ON_ONCE(rq != this_rq()) ||
1550 WARN_ON_ONCE(p == current))
1553 lockdep_assert_held(&rq->lock);
1555 if (!raw_spin_trylock(&p->pi_lock)) {
1556 raw_spin_unlock(&rq->lock);
1557 raw_spin_lock(&p->pi_lock);
1558 raw_spin_lock(&rq->lock);
1561 if (!(p->state & TASK_NORMAL))
1565 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1567 ttwu_do_wakeup(rq, p, 0);
1568 ttwu_stat(p, smp_processor_id(), 0);
1570 raw_spin_unlock(&p->pi_lock);
1574 * wake_up_process - Wake up a specific process
1575 * @p: The process to be woken up.
1577 * Attempt to wake up the nominated process and move it to the set of runnable
1578 * processes. Returns 1 if the process was woken up, 0 if it was already
1581 * It may be assumed that this function implies a write memory barrier before
1582 * changing the task state if and only if any tasks are woken up.
1584 int wake_up_process(struct task_struct *p)
1586 WARN_ON(task_is_stopped_or_traced(p));
1587 return try_to_wake_up(p, TASK_NORMAL, 0);
1589 EXPORT_SYMBOL(wake_up_process);
1591 int wake_up_state(struct task_struct *p, unsigned int state)
1593 return try_to_wake_up(p, state, 0);
1597 * Perform scheduler related setup for a newly forked process p.
1598 * p is forked by current.
1600 * __sched_fork() is basic setup used by init_idle() too:
1602 static void __sched_fork(struct task_struct *p)
1607 p->se.exec_start = 0;
1608 p->se.sum_exec_runtime = 0;
1609 p->se.prev_sum_exec_runtime = 0;
1610 p->se.nr_migrations = 0;
1612 INIT_LIST_HEAD(&p->se.group_node);
1615 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1616 * removed when useful for applications beyond shares distribution (e.g.
1619 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1620 p->se.avg.runnable_avg_period = 0;
1621 p->se.avg.runnable_avg_sum = 0;
1623 #ifdef CONFIG_SCHEDSTATS
1624 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1627 INIT_LIST_HEAD(&p->rt.run_list);
1629 #ifdef CONFIG_PREEMPT_NOTIFIERS
1630 INIT_HLIST_HEAD(&p->preempt_notifiers);
1633 #ifdef CONFIG_NUMA_BALANCING
1634 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1635 p->mm->numa_next_scan = jiffies;
1636 p->mm->numa_next_reset = jiffies;
1637 p->mm->numa_scan_seq = 0;
1640 p->node_stamp = 0ULL;
1641 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1642 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1643 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1644 p->numa_work.next = &p->numa_work;
1645 #endif /* CONFIG_NUMA_BALANCING */
1648 #ifdef CONFIG_NUMA_BALANCING
1649 #ifdef CONFIG_SCHED_DEBUG
1650 void set_numabalancing_state(bool enabled)
1653 sched_feat_set("NUMA");
1655 sched_feat_set("NO_NUMA");
1658 __read_mostly bool numabalancing_enabled;
1660 void set_numabalancing_state(bool enabled)
1662 numabalancing_enabled = enabled;
1664 #endif /* CONFIG_SCHED_DEBUG */
1665 #endif /* CONFIG_NUMA_BALANCING */
1668 * fork()/clone()-time setup:
1670 void sched_fork(struct task_struct *p)
1672 unsigned long flags;
1673 int cpu = get_cpu();
1677 * We mark the process as running here. This guarantees that
1678 * nobody will actually run it, and a signal or other external
1679 * event cannot wake it up and insert it on the runqueue either.
1681 p->state = TASK_RUNNING;
1684 * Make sure we do not leak PI boosting priority to the child.
1686 p->prio = current->normal_prio;
1689 * Revert to default priority/policy on fork if requested.
1691 if (unlikely(p->sched_reset_on_fork)) {
1692 if (task_has_rt_policy(p)) {
1693 p->policy = SCHED_NORMAL;
1694 p->static_prio = NICE_TO_PRIO(0);
1696 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1697 p->static_prio = NICE_TO_PRIO(0);
1699 p->prio = p->normal_prio = __normal_prio(p);
1703 * We don't need the reset flag anymore after the fork. It has
1704 * fulfilled its duty:
1706 p->sched_reset_on_fork = 0;
1709 if (!rt_prio(p->prio))
1710 p->sched_class = &fair_sched_class;
1712 if (p->sched_class->task_fork)
1713 p->sched_class->task_fork(p);
1716 * The child is not yet in the pid-hash so no cgroup attach races,
1717 * and the cgroup is pinned to this child due to cgroup_fork()
1718 * is ran before sched_fork().
1720 * Silence PROVE_RCU.
1722 raw_spin_lock_irqsave(&p->pi_lock, flags);
1723 set_task_cpu(p, cpu);
1724 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1726 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1727 if (likely(sched_info_on()))
1728 memset(&p->sched_info, 0, sizeof(p->sched_info));
1730 #if defined(CONFIG_SMP)
1733 #ifdef CONFIG_PREEMPT_COUNT
1734 /* Want to start with kernel preemption disabled. */
1735 task_thread_info(p)->preempt_count = 1;
1738 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1745 * wake_up_new_task - wake up a newly created task for the first time.
1747 * This function will do some initial scheduler statistics housekeeping
1748 * that must be done for every newly created context, then puts the task
1749 * on the runqueue and wakes it.
1751 void wake_up_new_task(struct task_struct *p)
1753 unsigned long flags;
1756 raw_spin_lock_irqsave(&p->pi_lock, flags);
1759 * Fork balancing, do it here and not earlier because:
1760 * - cpus_allowed can change in the fork path
1761 * - any previously selected cpu might disappear through hotplug
1763 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1766 rq = __task_rq_lock(p);
1767 activate_task(rq, p, 0);
1769 trace_sched_wakeup_new(p, true);
1770 check_preempt_curr(rq, p, WF_FORK);
1772 if (p->sched_class->task_woken)
1773 p->sched_class->task_woken(rq, p);
1775 task_rq_unlock(rq, p, &flags);
1778 #ifdef CONFIG_PREEMPT_NOTIFIERS
1781 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1782 * @notifier: notifier struct to register
1784 void preempt_notifier_register(struct preempt_notifier *notifier)
1786 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1788 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1791 * preempt_notifier_unregister - no longer interested in preemption notifications
1792 * @notifier: notifier struct to unregister
1794 * This is safe to call from within a preemption notifier.
1796 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1798 hlist_del(¬ifier->link);
1800 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1802 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1804 struct preempt_notifier *notifier;
1806 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1807 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1811 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1812 struct task_struct *next)
1814 struct preempt_notifier *notifier;
1816 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1817 notifier->ops->sched_out(notifier, next);
1820 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1822 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1827 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1828 struct task_struct *next)
1832 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1835 * prepare_task_switch - prepare to switch tasks
1836 * @rq: the runqueue preparing to switch
1837 * @prev: the current task that is being switched out
1838 * @next: the task we are going to switch to.
1840 * This is called with the rq lock held and interrupts off. It must
1841 * be paired with a subsequent finish_task_switch after the context
1844 * prepare_task_switch sets up locking and calls architecture specific
1848 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1849 struct task_struct *next)
1851 trace_sched_switch(prev, next);
1852 sched_info_switch(prev, next);
1853 perf_event_task_sched_out(prev, next);
1854 fire_sched_out_preempt_notifiers(prev, next);
1855 prepare_lock_switch(rq, next);
1856 prepare_arch_switch(next);
1860 * finish_task_switch - clean up after a task-switch
1861 * @rq: runqueue associated with task-switch
1862 * @prev: the thread we just switched away from.
1864 * finish_task_switch must be called after the context switch, paired
1865 * with a prepare_task_switch call before the context switch.
1866 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1867 * and do any other architecture-specific cleanup actions.
1869 * Note that we may have delayed dropping an mm in context_switch(). If
1870 * so, we finish that here outside of the runqueue lock. (Doing it
1871 * with the lock held can cause deadlocks; see schedule() for
1874 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1875 __releases(rq->lock)
1877 struct mm_struct *mm = rq->prev_mm;
1883 * A task struct has one reference for the use as "current".
1884 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1885 * schedule one last time. The schedule call will never return, and
1886 * the scheduled task must drop that reference.
1887 * The test for TASK_DEAD must occur while the runqueue locks are
1888 * still held, otherwise prev could be scheduled on another cpu, die
1889 * there before we look at prev->state, and then the reference would
1891 * Manfred Spraul <manfred@colorfullife.com>
1893 prev_state = prev->state;
1894 vtime_task_switch(prev);
1895 finish_arch_switch(prev);
1896 perf_event_task_sched_in(prev, current);
1897 finish_lock_switch(rq, prev);
1898 finish_arch_post_lock_switch();
1900 fire_sched_in_preempt_notifiers(current);
1903 if (unlikely(prev_state == TASK_DEAD)) {
1905 * Remove function-return probe instances associated with this
1906 * task and put them back on the free list.
1908 kprobe_flush_task(prev);
1909 put_task_struct(prev);
1912 tick_nohz_task_switch(current);
1917 /* assumes rq->lock is held */
1918 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1920 if (prev->sched_class->pre_schedule)
1921 prev->sched_class->pre_schedule(rq, prev);
1924 /* rq->lock is NOT held, but preemption is disabled */
1925 static inline void post_schedule(struct rq *rq)
1927 if (rq->post_schedule) {
1928 unsigned long flags;
1930 raw_spin_lock_irqsave(&rq->lock, flags);
1931 if (rq->curr->sched_class->post_schedule)
1932 rq->curr->sched_class->post_schedule(rq);
1933 raw_spin_unlock_irqrestore(&rq->lock, flags);
1935 rq->post_schedule = 0;
1941 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1945 static inline void post_schedule(struct rq *rq)
1952 * schedule_tail - first thing a freshly forked thread must call.
1953 * @prev: the thread we just switched away from.
1955 asmlinkage void schedule_tail(struct task_struct *prev)
1956 __releases(rq->lock)
1958 struct rq *rq = this_rq();
1960 finish_task_switch(rq, prev);
1963 * FIXME: do we need to worry about rq being invalidated by the
1968 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1969 /* In this case, finish_task_switch does not reenable preemption */
1972 if (current->set_child_tid)
1973 put_user(task_pid_vnr(current), current->set_child_tid);
1977 * context_switch - switch to the new MM and the new
1978 * thread's register state.
1981 context_switch(struct rq *rq, struct task_struct *prev,
1982 struct task_struct *next)
1984 struct mm_struct *mm, *oldmm;
1986 prepare_task_switch(rq, prev, next);
1989 oldmm = prev->active_mm;
1991 * For paravirt, this is coupled with an exit in switch_to to
1992 * combine the page table reload and the switch backend into
1995 arch_start_context_switch(prev);
1998 next->active_mm = oldmm;
1999 atomic_inc(&oldmm->mm_count);
2000 enter_lazy_tlb(oldmm, next);
2002 switch_mm(oldmm, mm, next);
2005 prev->active_mm = NULL;
2006 rq->prev_mm = oldmm;
2009 * Since the runqueue lock will be released by the next
2010 * task (which is an invalid locking op but in the case
2011 * of the scheduler it's an obvious special-case), so we
2012 * do an early lockdep release here:
2014 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2015 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2018 context_tracking_task_switch(prev, next);
2019 /* Here we just switch the register state and the stack. */
2020 switch_to(prev, next, prev);
2024 * this_rq must be evaluated again because prev may have moved
2025 * CPUs since it called schedule(), thus the 'rq' on its stack
2026 * frame will be invalid.
2028 finish_task_switch(this_rq(), prev);
2032 * nr_running and nr_context_switches:
2034 * externally visible scheduler statistics: current number of runnable
2035 * threads, total number of context switches performed since bootup.
2037 unsigned long nr_running(void)
2039 unsigned long i, sum = 0;
2041 for_each_online_cpu(i)
2042 sum += cpu_rq(i)->nr_running;
2047 unsigned long long nr_context_switches(void)
2050 unsigned long long sum = 0;
2052 for_each_possible_cpu(i)
2053 sum += cpu_rq(i)->nr_switches;
2058 unsigned long nr_iowait(void)
2060 unsigned long i, sum = 0;
2062 for_each_possible_cpu(i)
2063 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2068 unsigned long nr_iowait_cpu(int cpu)
2070 struct rq *this = cpu_rq(cpu);
2071 return atomic_read(&this->nr_iowait);
2077 * sched_exec - execve() is a valuable balancing opportunity, because at
2078 * this point the task has the smallest effective memory and cache footprint.
2080 void sched_exec(void)
2082 struct task_struct *p = current;
2083 unsigned long flags;
2086 raw_spin_lock_irqsave(&p->pi_lock, flags);
2087 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2088 if (dest_cpu == smp_processor_id())
2091 if (likely(cpu_active(dest_cpu))) {
2092 struct migration_arg arg = { p, dest_cpu };
2094 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2095 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2099 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2104 DEFINE_PER_CPU(struct kernel_stat, kstat);
2105 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2107 EXPORT_PER_CPU_SYMBOL(kstat);
2108 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2111 * Return any ns on the sched_clock that have not yet been accounted in
2112 * @p in case that task is currently running.
2114 * Called with task_rq_lock() held on @rq.
2116 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2120 if (task_current(rq, p)) {
2121 update_rq_clock(rq);
2122 ns = rq_clock_task(rq) - p->se.exec_start;
2130 unsigned long long task_delta_exec(struct task_struct *p)
2132 unsigned long flags;
2136 rq = task_rq_lock(p, &flags);
2137 ns = do_task_delta_exec(p, rq);
2138 task_rq_unlock(rq, p, &flags);
2144 * Return accounted runtime for the task.
2145 * In case the task is currently running, return the runtime plus current's
2146 * pending runtime that have not been accounted yet.
2148 unsigned long long task_sched_runtime(struct task_struct *p)
2150 unsigned long flags;
2154 rq = task_rq_lock(p, &flags);
2155 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2156 task_rq_unlock(rq, p, &flags);
2162 * This function gets called by the timer code, with HZ frequency.
2163 * We call it with interrupts disabled.
2165 void scheduler_tick(void)
2167 int cpu = smp_processor_id();
2168 struct rq *rq = cpu_rq(cpu);
2169 struct task_struct *curr = rq->curr;
2173 raw_spin_lock(&rq->lock);
2174 update_rq_clock(rq);
2175 update_cpu_load_active(rq);
2176 curr->sched_class->task_tick(rq, curr, 0);
2177 raw_spin_unlock(&rq->lock);
2179 perf_event_task_tick();
2182 rq->idle_balance = idle_cpu(cpu);
2183 trigger_load_balance(rq, cpu);
2185 rq_last_tick_reset(rq);
2188 #ifdef CONFIG_NO_HZ_FULL
2190 * scheduler_tick_max_deferment
2192 * Keep at least one tick per second when a single
2193 * active task is running because the scheduler doesn't
2194 * yet completely support full dynticks environment.
2196 * This makes sure that uptime, CFS vruntime, load
2197 * balancing, etc... continue to move forward, even
2198 * with a very low granularity.
2200 u64 scheduler_tick_max_deferment(void)
2202 struct rq *rq = this_rq();
2203 unsigned long next, now = ACCESS_ONCE(jiffies);
2205 next = rq->last_sched_tick + HZ;
2207 if (time_before_eq(next, now))
2210 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2214 notrace unsigned long get_parent_ip(unsigned long addr)
2216 if (in_lock_functions(addr)) {
2217 addr = CALLER_ADDR2;
2218 if (in_lock_functions(addr))
2219 addr = CALLER_ADDR3;
2224 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2225 defined(CONFIG_PREEMPT_TRACER))
2227 void __kprobes add_preempt_count(int val)
2229 #ifdef CONFIG_DEBUG_PREEMPT
2233 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2236 preempt_count() += val;
2237 #ifdef CONFIG_DEBUG_PREEMPT
2239 * Spinlock count overflowing soon?
2241 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2244 if (preempt_count() == val)
2245 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2247 EXPORT_SYMBOL(add_preempt_count);
2249 void __kprobes sub_preempt_count(int val)
2251 #ifdef CONFIG_DEBUG_PREEMPT
2255 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2258 * Is the spinlock portion underflowing?
2260 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2261 !(preempt_count() & PREEMPT_MASK)))
2265 if (preempt_count() == val)
2266 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2267 preempt_count() -= val;
2269 EXPORT_SYMBOL(sub_preempt_count);
2274 * Print scheduling while atomic bug:
2276 static noinline void __schedule_bug(struct task_struct *prev)
2278 if (oops_in_progress)
2281 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2282 prev->comm, prev->pid, preempt_count());
2284 debug_show_held_locks(prev);
2286 if (irqs_disabled())
2287 print_irqtrace_events(prev);
2289 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2293 * Various schedule()-time debugging checks and statistics:
2295 static inline void schedule_debug(struct task_struct *prev)
2298 * Test if we are atomic. Since do_exit() needs to call into
2299 * schedule() atomically, we ignore that path for now.
2300 * Otherwise, whine if we are scheduling when we should not be.
2302 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2303 __schedule_bug(prev);
2306 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2308 schedstat_inc(this_rq(), sched_count);
2311 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2313 if (prev->on_rq || rq->skip_clock_update < 0)
2314 update_rq_clock(rq);
2315 prev->sched_class->put_prev_task(rq, prev);
2319 * Pick up the highest-prio task:
2321 static inline struct task_struct *
2322 pick_next_task(struct rq *rq)
2324 const struct sched_class *class;
2325 struct task_struct *p;
2328 * Optimization: we know that if all tasks are in
2329 * the fair class we can call that function directly:
2331 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2332 p = fair_sched_class.pick_next_task(rq);
2337 for_each_class(class) {
2338 p = class->pick_next_task(rq);
2343 BUG(); /* the idle class will always have a runnable task */
2347 * __schedule() is the main scheduler function.
2349 * The main means of driving the scheduler and thus entering this function are:
2351 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2353 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2354 * paths. For example, see arch/x86/entry_64.S.
2356 * To drive preemption between tasks, the scheduler sets the flag in timer
2357 * interrupt handler scheduler_tick().
2359 * 3. Wakeups don't really cause entry into schedule(). They add a
2360 * task to the run-queue and that's it.
2362 * Now, if the new task added to the run-queue preempts the current
2363 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2364 * called on the nearest possible occasion:
2366 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2368 * - in syscall or exception context, at the next outmost
2369 * preempt_enable(). (this might be as soon as the wake_up()'s
2372 * - in IRQ context, return from interrupt-handler to
2373 * preemptible context
2375 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2378 * - cond_resched() call
2379 * - explicit schedule() call
2380 * - return from syscall or exception to user-space
2381 * - return from interrupt-handler to user-space
2383 static void __sched __schedule(void)
2385 struct task_struct *prev, *next;
2386 unsigned long *switch_count;
2392 cpu = smp_processor_id();
2394 rcu_note_context_switch(cpu);
2397 schedule_debug(prev);
2399 if (sched_feat(HRTICK))
2402 raw_spin_lock_irq(&rq->lock);
2404 switch_count = &prev->nivcsw;
2405 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2406 if (unlikely(signal_pending_state(prev->state, prev))) {
2407 prev->state = TASK_RUNNING;
2409 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2413 * If a worker went to sleep, notify and ask workqueue
2414 * whether it wants to wake up a task to maintain
2417 if (prev->flags & PF_WQ_WORKER) {
2418 struct task_struct *to_wakeup;
2420 to_wakeup = wq_worker_sleeping(prev, cpu);
2422 try_to_wake_up_local(to_wakeup);
2425 switch_count = &prev->nvcsw;
2428 pre_schedule(rq, prev);
2430 if (unlikely(!rq->nr_running))
2431 idle_balance(cpu, rq);
2433 put_prev_task(rq, prev);
2434 next = pick_next_task(rq);
2435 clear_tsk_need_resched(prev);
2436 rq->skip_clock_update = 0;
2438 if (likely(prev != next)) {
2443 context_switch(rq, prev, next); /* unlocks the rq */
2445 * The context switch have flipped the stack from under us
2446 * and restored the local variables which were saved when
2447 * this task called schedule() in the past. prev == current
2448 * is still correct, but it can be moved to another cpu/rq.
2450 cpu = smp_processor_id();
2453 raw_spin_unlock_irq(&rq->lock);
2457 sched_preempt_enable_no_resched();
2462 static inline void sched_submit_work(struct task_struct *tsk)
2464 if (!tsk->state || tsk_is_pi_blocked(tsk))
2467 * If we are going to sleep and we have plugged IO queued,
2468 * make sure to submit it to avoid deadlocks.
2470 if (blk_needs_flush_plug(tsk))
2471 blk_schedule_flush_plug(tsk);
2474 asmlinkage void __sched schedule(void)
2476 struct task_struct *tsk = current;
2478 sched_submit_work(tsk);
2481 EXPORT_SYMBOL(schedule);
2483 #ifdef CONFIG_CONTEXT_TRACKING
2484 asmlinkage void __sched schedule_user(void)
2487 * If we come here after a random call to set_need_resched(),
2488 * or we have been woken up remotely but the IPI has not yet arrived,
2489 * we haven't yet exited the RCU idle mode. Do it here manually until
2490 * we find a better solution.
2499 * schedule_preempt_disabled - called with preemption disabled
2501 * Returns with preemption disabled. Note: preempt_count must be 1
2503 void __sched schedule_preempt_disabled(void)
2505 sched_preempt_enable_no_resched();
2510 #ifdef CONFIG_PREEMPT
2512 * this is the entry point to schedule() from in-kernel preemption
2513 * off of preempt_enable. Kernel preemptions off return from interrupt
2514 * occur there and call schedule directly.
2516 asmlinkage void __sched notrace preempt_schedule(void)
2518 struct thread_info *ti = current_thread_info();
2521 * If there is a non-zero preempt_count or interrupts are disabled,
2522 * we do not want to preempt the current task. Just return..
2524 if (likely(ti->preempt_count || irqs_disabled()))
2528 add_preempt_count_notrace(PREEMPT_ACTIVE);
2530 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2533 * Check again in case we missed a preemption opportunity
2534 * between schedule and now.
2537 } while (need_resched());
2539 EXPORT_SYMBOL(preempt_schedule);
2542 * this is the entry point to schedule() from kernel preemption
2543 * off of irq context.
2544 * Note, that this is called and return with irqs disabled. This will
2545 * protect us against recursive calling from irq.
2547 asmlinkage void __sched preempt_schedule_irq(void)
2549 struct thread_info *ti = current_thread_info();
2550 enum ctx_state prev_state;
2552 /* Catch callers which need to be fixed */
2553 BUG_ON(ti->preempt_count || !irqs_disabled());
2555 prev_state = exception_enter();
2558 add_preempt_count(PREEMPT_ACTIVE);
2561 local_irq_disable();
2562 sub_preempt_count(PREEMPT_ACTIVE);
2565 * Check again in case we missed a preemption opportunity
2566 * between schedule and now.
2569 } while (need_resched());
2571 exception_exit(prev_state);
2574 #endif /* CONFIG_PREEMPT */
2576 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2579 return try_to_wake_up(curr->private, mode, wake_flags);
2581 EXPORT_SYMBOL(default_wake_function);
2584 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2585 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2586 * number) then we wake all the non-exclusive tasks and one exclusive task.
2588 * There are circumstances in which we can try to wake a task which has already
2589 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2590 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2592 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2593 int nr_exclusive, int wake_flags, void *key)
2595 wait_queue_t *curr, *next;
2597 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2598 unsigned flags = curr->flags;
2600 if (curr->func(curr, mode, wake_flags, key) &&
2601 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2607 * __wake_up - wake up threads blocked on a waitqueue.
2609 * @mode: which threads
2610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2611 * @key: is directly passed to the wakeup function
2613 * It may be assumed that this function implies a write memory barrier before
2614 * changing the task state if and only if any tasks are woken up.
2616 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2617 int nr_exclusive, void *key)
2619 unsigned long flags;
2621 spin_lock_irqsave(&q->lock, flags);
2622 __wake_up_common(q, mode, nr_exclusive, 0, key);
2623 spin_unlock_irqrestore(&q->lock, flags);
2625 EXPORT_SYMBOL(__wake_up);
2628 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2630 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2632 __wake_up_common(q, mode, nr, 0, NULL);
2634 EXPORT_SYMBOL_GPL(__wake_up_locked);
2636 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2638 __wake_up_common(q, mode, 1, 0, key);
2640 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2643 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2645 * @mode: which threads
2646 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2647 * @key: opaque value to be passed to wakeup targets
2649 * The sync wakeup differs that the waker knows that it will schedule
2650 * away soon, so while the target thread will be woken up, it will not
2651 * be migrated to another CPU - ie. the two threads are 'synchronized'
2652 * with each other. This can prevent needless bouncing between CPUs.
2654 * On UP it can prevent extra preemption.
2656 * It may be assumed that this function implies a write memory barrier before
2657 * changing the task state if and only if any tasks are woken up.
2659 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2660 int nr_exclusive, void *key)
2662 unsigned long flags;
2663 int wake_flags = WF_SYNC;
2668 if (unlikely(!nr_exclusive))
2671 spin_lock_irqsave(&q->lock, flags);
2672 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2673 spin_unlock_irqrestore(&q->lock, flags);
2675 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2678 * __wake_up_sync - see __wake_up_sync_key()
2680 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2682 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2684 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2687 * complete: - signals a single thread waiting on this completion
2688 * @x: holds the state of this particular completion
2690 * This will wake up a single thread waiting on this completion. Threads will be
2691 * awakened in the same order in which they were queued.
2693 * See also complete_all(), wait_for_completion() and related routines.
2695 * It may be assumed that this function implies a write memory barrier before
2696 * changing the task state if and only if any tasks are woken up.
2698 void complete(struct completion *x)
2700 unsigned long flags;
2702 spin_lock_irqsave(&x->wait.lock, flags);
2704 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2705 spin_unlock_irqrestore(&x->wait.lock, flags);
2707 EXPORT_SYMBOL(complete);
2710 * complete_all: - signals all threads waiting on this completion
2711 * @x: holds the state of this particular completion
2713 * This will wake up all threads waiting on this particular completion event.
2715 * It may be assumed that this function implies a write memory barrier before
2716 * changing the task state if and only if any tasks are woken up.
2718 void complete_all(struct completion *x)
2720 unsigned long flags;
2722 spin_lock_irqsave(&x->wait.lock, flags);
2723 x->done += UINT_MAX/2;
2724 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2725 spin_unlock_irqrestore(&x->wait.lock, flags);
2727 EXPORT_SYMBOL(complete_all);
2729 static inline long __sched
2730 do_wait_for_common(struct completion *x,
2731 long (*action)(long), long timeout, int state)
2734 DECLARE_WAITQUEUE(wait, current);
2736 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2738 if (signal_pending_state(state, current)) {
2739 timeout = -ERESTARTSYS;
2742 __set_current_state(state);
2743 spin_unlock_irq(&x->wait.lock);
2744 timeout = action(timeout);
2745 spin_lock_irq(&x->wait.lock);
2746 } while (!x->done && timeout);
2747 __remove_wait_queue(&x->wait, &wait);
2752 return timeout ?: 1;
2755 static inline long __sched
2756 __wait_for_common(struct completion *x,
2757 long (*action)(long), long timeout, int state)
2761 spin_lock_irq(&x->wait.lock);
2762 timeout = do_wait_for_common(x, action, timeout, state);
2763 spin_unlock_irq(&x->wait.lock);
2768 wait_for_common(struct completion *x, long timeout, int state)
2770 return __wait_for_common(x, schedule_timeout, timeout, state);
2774 wait_for_common_io(struct completion *x, long timeout, int state)
2776 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2780 * wait_for_completion: - waits for completion of a task
2781 * @x: holds the state of this particular completion
2783 * This waits to be signaled for completion of a specific task. It is NOT
2784 * interruptible and there is no timeout.
2786 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2787 * and interrupt capability. Also see complete().
2789 void __sched wait_for_completion(struct completion *x)
2791 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2793 EXPORT_SYMBOL(wait_for_completion);
2796 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2797 * @x: holds the state of this particular completion
2798 * @timeout: timeout value in jiffies
2800 * This waits for either a completion of a specific task to be signaled or for a
2801 * specified timeout to expire. The timeout is in jiffies. It is not
2804 * The return value is 0 if timed out, and positive (at least 1, or number of
2805 * jiffies left till timeout) if completed.
2807 unsigned long __sched
2808 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2810 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2812 EXPORT_SYMBOL(wait_for_completion_timeout);
2815 * wait_for_completion_io: - waits for completion of a task
2816 * @x: holds the state of this particular completion
2818 * This waits to be signaled for completion of a specific task. It is NOT
2819 * interruptible and there is no timeout. The caller is accounted as waiting
2822 void __sched wait_for_completion_io(struct completion *x)
2824 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2826 EXPORT_SYMBOL(wait_for_completion_io);
2829 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2830 * @x: holds the state of this particular completion
2831 * @timeout: timeout value in jiffies
2833 * This waits for either a completion of a specific task to be signaled or for a
2834 * specified timeout to expire. The timeout is in jiffies. It is not
2835 * interruptible. The caller is accounted as waiting for IO.
2837 * The return value is 0 if timed out, and positive (at least 1, or number of
2838 * jiffies left till timeout) if completed.
2840 unsigned long __sched
2841 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2843 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2845 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2848 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2849 * @x: holds the state of this particular completion
2851 * This waits for completion of a specific task to be signaled. It is
2854 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2856 int __sched wait_for_completion_interruptible(struct completion *x)
2858 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2859 if (t == -ERESTARTSYS)
2863 EXPORT_SYMBOL(wait_for_completion_interruptible);
2866 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2867 * @x: holds the state of this particular completion
2868 * @timeout: timeout value in jiffies
2870 * This waits for either a completion of a specific task to be signaled or for a
2871 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2873 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2874 * positive (at least 1, or number of jiffies left till timeout) if completed.
2877 wait_for_completion_interruptible_timeout(struct completion *x,
2878 unsigned long timeout)
2880 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2882 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2885 * wait_for_completion_killable: - waits for completion of a task (killable)
2886 * @x: holds the state of this particular completion
2888 * This waits to be signaled for completion of a specific task. It can be
2889 * interrupted by a kill signal.
2891 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2893 int __sched wait_for_completion_killable(struct completion *x)
2895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2896 if (t == -ERESTARTSYS)
2900 EXPORT_SYMBOL(wait_for_completion_killable);
2903 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2904 * @x: holds the state of this particular completion
2905 * @timeout: timeout value in jiffies
2907 * This waits for either a completion of a specific task to be
2908 * signaled or for a specified timeout to expire. It can be
2909 * interrupted by a kill signal. The timeout is in jiffies.
2911 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2912 * positive (at least 1, or number of jiffies left till timeout) if completed.
2915 wait_for_completion_killable_timeout(struct completion *x,
2916 unsigned long timeout)
2918 return wait_for_common(x, timeout, TASK_KILLABLE);
2920 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2923 * try_wait_for_completion - try to decrement a completion without blocking
2924 * @x: completion structure
2926 * Returns: 0 if a decrement cannot be done without blocking
2927 * 1 if a decrement succeeded.
2929 * If a completion is being used as a counting completion,
2930 * attempt to decrement the counter without blocking. This
2931 * enables us to avoid waiting if the resource the completion
2932 * is protecting is not available.
2934 bool try_wait_for_completion(struct completion *x)
2936 unsigned long flags;
2939 spin_lock_irqsave(&x->wait.lock, flags);
2944 spin_unlock_irqrestore(&x->wait.lock, flags);
2947 EXPORT_SYMBOL(try_wait_for_completion);
2950 * completion_done - Test to see if a completion has any waiters
2951 * @x: completion structure
2953 * Returns: 0 if there are waiters (wait_for_completion() in progress)
2954 * 1 if there are no waiters.
2957 bool completion_done(struct completion *x)
2959 unsigned long flags;
2962 spin_lock_irqsave(&x->wait.lock, flags);
2965 spin_unlock_irqrestore(&x->wait.lock, flags);
2968 EXPORT_SYMBOL(completion_done);
2971 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2973 unsigned long flags;
2976 init_waitqueue_entry(&wait, current);
2978 __set_current_state(state);
2980 spin_lock_irqsave(&q->lock, flags);
2981 __add_wait_queue(q, &wait);
2982 spin_unlock(&q->lock);
2983 timeout = schedule_timeout(timeout);
2984 spin_lock_irq(&q->lock);
2985 __remove_wait_queue(q, &wait);
2986 spin_unlock_irqrestore(&q->lock, flags);
2991 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2993 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2995 EXPORT_SYMBOL(interruptible_sleep_on);
2998 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3000 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3002 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3004 void __sched sleep_on(wait_queue_head_t *q)
3006 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3008 EXPORT_SYMBOL(sleep_on);
3010 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3012 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3014 EXPORT_SYMBOL(sleep_on_timeout);
3016 #ifdef CONFIG_RT_MUTEXES
3019 * rt_mutex_setprio - set the current priority of a task
3021 * @prio: prio value (kernel-internal form)
3023 * This function changes the 'effective' priority of a task. It does
3024 * not touch ->normal_prio like __setscheduler().
3026 * Used by the rt_mutex code to implement priority inheritance logic.
3028 void rt_mutex_setprio(struct task_struct *p, int prio)
3030 int oldprio, on_rq, running;
3032 const struct sched_class *prev_class;
3034 BUG_ON(prio < 0 || prio > MAX_PRIO);
3036 rq = __task_rq_lock(p);
3039 * Idle task boosting is a nono in general. There is one
3040 * exception, when PREEMPT_RT and NOHZ is active:
3042 * The idle task calls get_next_timer_interrupt() and holds
3043 * the timer wheel base->lock on the CPU and another CPU wants
3044 * to access the timer (probably to cancel it). We can safely
3045 * ignore the boosting request, as the idle CPU runs this code
3046 * with interrupts disabled and will complete the lock
3047 * protected section without being interrupted. So there is no
3048 * real need to boost.
3050 if (unlikely(p == rq->idle)) {
3051 WARN_ON(p != rq->curr);
3052 WARN_ON(p->pi_blocked_on);
3056 trace_sched_pi_setprio(p, prio);
3058 prev_class = p->sched_class;
3060 running = task_current(rq, p);
3062 dequeue_task(rq, p, 0);
3064 p->sched_class->put_prev_task(rq, p);
3067 p->sched_class = &rt_sched_class;
3069 p->sched_class = &fair_sched_class;
3074 p->sched_class->set_curr_task(rq);
3076 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3078 check_class_changed(rq, p, prev_class, oldprio);
3080 __task_rq_unlock(rq);
3083 void set_user_nice(struct task_struct *p, long nice)
3085 int old_prio, delta, on_rq;
3086 unsigned long flags;
3089 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3092 * We have to be careful, if called from sys_setpriority(),
3093 * the task might be in the middle of scheduling on another CPU.
3095 rq = task_rq_lock(p, &flags);
3097 * The RT priorities are set via sched_setscheduler(), but we still
3098 * allow the 'normal' nice value to be set - but as expected
3099 * it wont have any effect on scheduling until the task is
3100 * SCHED_FIFO/SCHED_RR:
3102 if (task_has_rt_policy(p)) {
3103 p->static_prio = NICE_TO_PRIO(nice);
3108 dequeue_task(rq, p, 0);
3110 p->static_prio = NICE_TO_PRIO(nice);
3113 p->prio = effective_prio(p);
3114 delta = p->prio - old_prio;
3117 enqueue_task(rq, p, 0);
3119 * If the task increased its priority or is running and
3120 * lowered its priority, then reschedule its CPU:
3122 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3123 resched_task(rq->curr);
3126 task_rq_unlock(rq, p, &flags);
3128 EXPORT_SYMBOL(set_user_nice);
3131 * can_nice - check if a task can reduce its nice value
3135 int can_nice(const struct task_struct *p, const int nice)
3137 /* convert nice value [19,-20] to rlimit style value [1,40] */
3138 int nice_rlim = 20 - nice;
3140 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3141 capable(CAP_SYS_NICE));
3144 #ifdef __ARCH_WANT_SYS_NICE
3147 * sys_nice - change the priority of the current process.
3148 * @increment: priority increment
3150 * sys_setpriority is a more generic, but much slower function that
3151 * does similar things.
3153 SYSCALL_DEFINE1(nice, int, increment)
3158 * Setpriority might change our priority at the same moment.
3159 * We don't have to worry. Conceptually one call occurs first
3160 * and we have a single winner.
3162 if (increment < -40)
3167 nice = TASK_NICE(current) + increment;
3173 if (increment < 0 && !can_nice(current, nice))
3176 retval = security_task_setnice(current, nice);
3180 set_user_nice(current, nice);
3187 * task_prio - return the priority value of a given task.
3188 * @p: the task in question.
3190 * This is the priority value as seen by users in /proc.
3191 * RT tasks are offset by -200. Normal tasks are centered
3192 * around 0, value goes from -16 to +15.
3194 int task_prio(const struct task_struct *p)
3196 return p->prio - MAX_RT_PRIO;
3200 * task_nice - return the nice value of a given task.
3201 * @p: the task in question.
3203 int task_nice(const struct task_struct *p)
3205 return TASK_NICE(p);
3207 EXPORT_SYMBOL(task_nice);
3210 * idle_cpu - is a given cpu idle currently?
3211 * @cpu: the processor in question.
3213 int idle_cpu(int cpu)
3215 struct rq *rq = cpu_rq(cpu);
3217 if (rq->curr != rq->idle)
3224 if (!llist_empty(&rq->wake_list))
3232 * idle_task - return the idle task for a given cpu.
3233 * @cpu: the processor in question.
3235 struct task_struct *idle_task(int cpu)
3237 return cpu_rq(cpu)->idle;
3241 * find_process_by_pid - find a process with a matching PID value.
3242 * @pid: the pid in question.
3244 static struct task_struct *find_process_by_pid(pid_t pid)
3246 return pid ? find_task_by_vpid(pid) : current;
3249 /* Actually do priority change: must hold rq lock. */
3251 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3254 p->rt_priority = prio;
3255 p->normal_prio = normal_prio(p);
3256 /* we are holding p->pi_lock already */
3257 p->prio = rt_mutex_getprio(p);
3258 if (rt_prio(p->prio))
3259 p->sched_class = &rt_sched_class;
3261 p->sched_class = &fair_sched_class;
3266 * check the target process has a UID that matches the current process's
3268 static bool check_same_owner(struct task_struct *p)
3270 const struct cred *cred = current_cred(), *pcred;
3274 pcred = __task_cred(p);
3275 match = (uid_eq(cred->euid, pcred->euid) ||
3276 uid_eq(cred->euid, pcred->uid));
3281 static int __sched_setscheduler(struct task_struct *p, int policy,
3282 const struct sched_param *param, bool user)
3284 int retval, oldprio, oldpolicy = -1, on_rq, running;
3285 unsigned long flags;
3286 const struct sched_class *prev_class;
3290 /* may grab non-irq protected spin_locks */
3291 BUG_ON(in_interrupt());
3293 /* double check policy once rq lock held */
3295 reset_on_fork = p->sched_reset_on_fork;
3296 policy = oldpolicy = p->policy;
3298 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3299 policy &= ~SCHED_RESET_ON_FORK;
3301 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3302 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3303 policy != SCHED_IDLE)
3308 * Valid priorities for SCHED_FIFO and SCHED_RR are
3309 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3310 * SCHED_BATCH and SCHED_IDLE is 0.
3312 if (param->sched_priority < 0 ||
3313 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3314 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3316 if (rt_policy(policy) != (param->sched_priority != 0))
3320 * Allow unprivileged RT tasks to decrease priority:
3322 if (user && !capable(CAP_SYS_NICE)) {
3323 if (rt_policy(policy)) {
3324 unsigned long rlim_rtprio =
3325 task_rlimit(p, RLIMIT_RTPRIO);
3327 /* can't set/change the rt policy */
3328 if (policy != p->policy && !rlim_rtprio)
3331 /* can't increase priority */
3332 if (param->sched_priority > p->rt_priority &&
3333 param->sched_priority > rlim_rtprio)
3338 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3339 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3341 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3342 if (!can_nice(p, TASK_NICE(p)))
3346 /* can't change other user's priorities */
3347 if (!check_same_owner(p))
3350 /* Normal users shall not reset the sched_reset_on_fork flag */
3351 if (p->sched_reset_on_fork && !reset_on_fork)
3356 retval = security_task_setscheduler(p);
3362 * make sure no PI-waiters arrive (or leave) while we are
3363 * changing the priority of the task:
3365 * To be able to change p->policy safely, the appropriate
3366 * runqueue lock must be held.
3368 rq = task_rq_lock(p, &flags);
3371 * Changing the policy of the stop threads its a very bad idea
3373 if (p == rq->stop) {
3374 task_rq_unlock(rq, p, &flags);
3379 * If not changing anything there's no need to proceed further:
3381 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3382 param->sched_priority == p->rt_priority))) {
3383 task_rq_unlock(rq, p, &flags);
3387 #ifdef CONFIG_RT_GROUP_SCHED
3390 * Do not allow realtime tasks into groups that have no runtime
3393 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3394 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3395 !task_group_is_autogroup(task_group(p))) {
3396 task_rq_unlock(rq, p, &flags);
3402 /* recheck policy now with rq lock held */
3403 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3404 policy = oldpolicy = -1;
3405 task_rq_unlock(rq, p, &flags);
3409 running = task_current(rq, p);
3411 dequeue_task(rq, p, 0);
3413 p->sched_class->put_prev_task(rq, p);
3415 p->sched_reset_on_fork = reset_on_fork;
3418 prev_class = p->sched_class;
3419 __setscheduler(rq, p, policy, param->sched_priority);
3422 p->sched_class->set_curr_task(rq);
3424 enqueue_task(rq, p, 0);
3426 check_class_changed(rq, p, prev_class, oldprio);
3427 task_rq_unlock(rq, p, &flags);
3429 rt_mutex_adjust_pi(p);
3435 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3436 * @p: the task in question.
3437 * @policy: new policy.
3438 * @param: structure containing the new RT priority.
3440 * NOTE that the task may be already dead.
3442 int sched_setscheduler(struct task_struct *p, int policy,
3443 const struct sched_param *param)
3445 return __sched_setscheduler(p, policy, param, true);
3447 EXPORT_SYMBOL_GPL(sched_setscheduler);
3450 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3451 * @p: the task in question.
3452 * @policy: new policy.
3453 * @param: structure containing the new RT priority.
3455 * Just like sched_setscheduler, only don't bother checking if the
3456 * current context has permission. For example, this is needed in
3457 * stop_machine(): we create temporary high priority worker threads,
3458 * but our caller might not have that capability.
3460 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3461 const struct sched_param *param)
3463 return __sched_setscheduler(p, policy, param, false);
3467 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3469 struct sched_param lparam;
3470 struct task_struct *p;
3473 if (!param || pid < 0)
3475 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3480 p = find_process_by_pid(pid);
3482 retval = sched_setscheduler(p, policy, &lparam);
3489 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3490 * @pid: the pid in question.
3491 * @policy: new policy.
3492 * @param: structure containing the new RT priority.
3494 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3495 struct sched_param __user *, param)
3497 /* negative values for policy are not valid */
3501 return do_sched_setscheduler(pid, policy, param);
3505 * sys_sched_setparam - set/change the RT priority of a thread
3506 * @pid: the pid in question.
3507 * @param: structure containing the new RT priority.
3509 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3511 return do_sched_setscheduler(pid, -1, param);
3515 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3516 * @pid: the pid in question.
3518 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3520 struct task_struct *p;
3528 p = find_process_by_pid(pid);
3530 retval = security_task_getscheduler(p);
3533 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3540 * sys_sched_getparam - get the RT priority of a thread
3541 * @pid: the pid in question.
3542 * @param: structure containing the RT priority.
3544 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3546 struct sched_param lp;
3547 struct task_struct *p;
3550 if (!param || pid < 0)
3554 p = find_process_by_pid(pid);
3559 retval = security_task_getscheduler(p);
3563 lp.sched_priority = p->rt_priority;
3567 * This one might sleep, we cannot do it with a spinlock held ...
3569 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3578 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3580 cpumask_var_t cpus_allowed, new_mask;
3581 struct task_struct *p;
3587 p = find_process_by_pid(pid);
3594 /* Prevent p going away */
3598 if (p->flags & PF_NO_SETAFFINITY) {
3602 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3606 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3608 goto out_free_cpus_allowed;
3611 if (!check_same_owner(p)) {
3613 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3620 retval = security_task_setscheduler(p);
3624 cpuset_cpus_allowed(p, cpus_allowed);
3625 cpumask_and(new_mask, in_mask, cpus_allowed);
3627 retval = set_cpus_allowed_ptr(p, new_mask);
3630 cpuset_cpus_allowed(p, cpus_allowed);
3631 if (!cpumask_subset(new_mask, cpus_allowed)) {
3633 * We must have raced with a concurrent cpuset
3634 * update. Just reset the cpus_allowed to the
3635 * cpuset's cpus_allowed
3637 cpumask_copy(new_mask, cpus_allowed);
3642 free_cpumask_var(new_mask);
3643 out_free_cpus_allowed:
3644 free_cpumask_var(cpus_allowed);
3651 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3652 struct cpumask *new_mask)
3654 if (len < cpumask_size())
3655 cpumask_clear(new_mask);
3656 else if (len > cpumask_size())
3657 len = cpumask_size();
3659 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3663 * sys_sched_setaffinity - set the cpu affinity of a process
3664 * @pid: pid of the process
3665 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3666 * @user_mask_ptr: user-space pointer to the new cpu mask
3668 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3669 unsigned long __user *, user_mask_ptr)
3671 cpumask_var_t new_mask;
3674 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3677 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3679 retval = sched_setaffinity(pid, new_mask);
3680 free_cpumask_var(new_mask);
3684 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3686 struct task_struct *p;
3687 unsigned long flags;
3694 p = find_process_by_pid(pid);
3698 retval = security_task_getscheduler(p);
3702 raw_spin_lock_irqsave(&p->pi_lock, flags);
3703 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3704 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3714 * sys_sched_getaffinity - get the cpu affinity of a process
3715 * @pid: pid of the process
3716 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3717 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3719 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3720 unsigned long __user *, user_mask_ptr)
3725 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3727 if (len & (sizeof(unsigned long)-1))
3730 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3733 ret = sched_getaffinity(pid, mask);
3735 size_t retlen = min_t(size_t, len, cpumask_size());
3737 if (copy_to_user(user_mask_ptr, mask, retlen))
3742 free_cpumask_var(mask);
3748 * sys_sched_yield - yield the current processor to other threads.
3750 * This function yields the current CPU to other tasks. If there are no
3751 * other threads running on this CPU then this function will return.
3753 SYSCALL_DEFINE0(sched_yield)
3755 struct rq *rq = this_rq_lock();
3757 schedstat_inc(rq, yld_count);
3758 current->sched_class->yield_task(rq);
3761 * Since we are going to call schedule() anyway, there's
3762 * no need to preempt or enable interrupts:
3764 __release(rq->lock);
3765 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3766 do_raw_spin_unlock(&rq->lock);
3767 sched_preempt_enable_no_resched();
3774 static inline int should_resched(void)
3776 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
3779 static void __cond_resched(void)
3781 add_preempt_count(PREEMPT_ACTIVE);
3783 sub_preempt_count(PREEMPT_ACTIVE);
3786 int __sched _cond_resched(void)
3788 if (should_resched()) {
3794 EXPORT_SYMBOL(_cond_resched);
3797 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3798 * call schedule, and on return reacquire the lock.
3800 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3801 * operations here to prevent schedule() from being called twice (once via
3802 * spin_unlock(), once by hand).
3804 int __cond_resched_lock(spinlock_t *lock)
3806 int resched = should_resched();
3809 lockdep_assert_held(lock);
3811 if (spin_needbreak(lock) || resched) {
3822 EXPORT_SYMBOL(__cond_resched_lock);
3824 int __sched __cond_resched_softirq(void)
3826 BUG_ON(!in_softirq());
3828 if (should_resched()) {
3836 EXPORT_SYMBOL(__cond_resched_softirq);
3839 * yield - yield the current processor to other threads.
3841 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3843 * The scheduler is at all times free to pick the calling task as the most
3844 * eligible task to run, if removing the yield() call from your code breaks
3845 * it, its already broken.
3847 * Typical broken usage is:
3852 * where one assumes that yield() will let 'the other' process run that will
3853 * make event true. If the current task is a SCHED_FIFO task that will never
3854 * happen. Never use yield() as a progress guarantee!!
3856 * If you want to use yield() to wait for something, use wait_event().
3857 * If you want to use yield() to be 'nice' for others, use cond_resched().
3858 * If you still want to use yield(), do not!
3860 void __sched yield(void)
3862 set_current_state(TASK_RUNNING);
3865 EXPORT_SYMBOL(yield);
3868 * yield_to - yield the current processor to another thread in
3869 * your thread group, or accelerate that thread toward the
3870 * processor it's on.
3872 * @preempt: whether task preemption is allowed or not
3874 * It's the caller's job to ensure that the target task struct
3875 * can't go away on us before we can do any checks.
3878 * true (>0) if we indeed boosted the target task.
3879 * false (0) if we failed to boost the target.
3880 * -ESRCH if there's no task to yield to.
3882 bool __sched yield_to(struct task_struct *p, bool preempt)
3884 struct task_struct *curr = current;
3885 struct rq *rq, *p_rq;
3886 unsigned long flags;
3889 local_irq_save(flags);
3895 * If we're the only runnable task on the rq and target rq also
3896 * has only one task, there's absolutely no point in yielding.
3898 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3903 double_rq_lock(rq, p_rq);
3904 while (task_rq(p) != p_rq) {
3905 double_rq_unlock(rq, p_rq);
3909 if (!curr->sched_class->yield_to_task)
3912 if (curr->sched_class != p->sched_class)
3915 if (task_running(p_rq, p) || p->state)
3918 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3920 schedstat_inc(rq, yld_count);
3922 * Make p's CPU reschedule; pick_next_entity takes care of
3925 if (preempt && rq != p_rq)
3926 resched_task(p_rq->curr);
3930 double_rq_unlock(rq, p_rq);
3932 local_irq_restore(flags);
3939 EXPORT_SYMBOL_GPL(yield_to);
3942 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3943 * that process accounting knows that this is a task in IO wait state.
3945 void __sched io_schedule(void)
3947 struct rq *rq = raw_rq();
3949 delayacct_blkio_start();
3950 atomic_inc(&rq->nr_iowait);
3951 blk_flush_plug(current);
3952 current->in_iowait = 1;
3954 current->in_iowait = 0;
3955 atomic_dec(&rq->nr_iowait);
3956 delayacct_blkio_end();
3958 EXPORT_SYMBOL(io_schedule);
3960 long __sched io_schedule_timeout(long timeout)
3962 struct rq *rq = raw_rq();
3965 delayacct_blkio_start();
3966 atomic_inc(&rq->nr_iowait);
3967 blk_flush_plug(current);
3968 current->in_iowait = 1;
3969 ret = schedule_timeout(timeout);
3970 current->in_iowait = 0;
3971 atomic_dec(&rq->nr_iowait);
3972 delayacct_blkio_end();
3977 * sys_sched_get_priority_max - return maximum RT priority.
3978 * @policy: scheduling class.
3980 * this syscall returns the maximum rt_priority that can be used
3981 * by a given scheduling class.
3983 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3990 ret = MAX_USER_RT_PRIO-1;
4002 * sys_sched_get_priority_min - return minimum RT priority.
4003 * @policy: scheduling class.
4005 * this syscall returns the minimum rt_priority that can be used
4006 * by a given scheduling class.
4008 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4026 * sys_sched_rr_get_interval - return the default timeslice of a process.
4027 * @pid: pid of the process.
4028 * @interval: userspace pointer to the timeslice value.
4030 * this syscall writes the default timeslice value of a given process
4031 * into the user-space timespec buffer. A value of '0' means infinity.
4033 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4034 struct timespec __user *, interval)
4036 struct task_struct *p;
4037 unsigned int time_slice;
4038 unsigned long flags;
4048 p = find_process_by_pid(pid);
4052 retval = security_task_getscheduler(p);
4056 rq = task_rq_lock(p, &flags);
4057 time_slice = p->sched_class->get_rr_interval(rq, p);
4058 task_rq_unlock(rq, p, &flags);
4061 jiffies_to_timespec(time_slice, &t);
4062 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4070 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4072 void sched_show_task(struct task_struct *p)
4074 unsigned long free = 0;
4078 state = p->state ? __ffs(p->state) + 1 : 0;
4079 printk(KERN_INFO "%-15.15s %c", p->comm,
4080 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4081 #if BITS_PER_LONG == 32
4082 if (state == TASK_RUNNING)
4083 printk(KERN_CONT " running ");
4085 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4087 if (state == TASK_RUNNING)
4088 printk(KERN_CONT " running task ");
4090 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4092 #ifdef CONFIG_DEBUG_STACK_USAGE
4093 free = stack_not_used(p);
4096 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4098 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4099 task_pid_nr(p), ppid,
4100 (unsigned long)task_thread_info(p)->flags);
4102 print_worker_info(KERN_INFO, p);
4103 show_stack(p, NULL);
4106 void show_state_filter(unsigned long state_filter)
4108 struct task_struct *g, *p;
4110 #if BITS_PER_LONG == 32
4112 " task PC stack pid father\n");
4115 " task PC stack pid father\n");
4118 do_each_thread(g, p) {
4120 * reset the NMI-timeout, listing all files on a slow
4121 * console might take a lot of time:
4123 touch_nmi_watchdog();
4124 if (!state_filter || (p->state & state_filter))
4126 } while_each_thread(g, p);
4128 touch_all_softlockup_watchdogs();
4130 #ifdef CONFIG_SCHED_DEBUG
4131 sysrq_sched_debug_show();
4135 * Only show locks if all tasks are dumped:
4138 debug_show_all_locks();
4141 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4143 idle->sched_class = &idle_sched_class;
4147 * init_idle - set up an idle thread for a given CPU
4148 * @idle: task in question
4149 * @cpu: cpu the idle task belongs to
4151 * NOTE: this function does not set the idle thread's NEED_RESCHED
4152 * flag, to make booting more robust.
4154 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4156 struct rq *rq = cpu_rq(cpu);
4157 unsigned long flags;
4159 raw_spin_lock_irqsave(&rq->lock, flags);
4162 idle->state = TASK_RUNNING;
4163 idle->se.exec_start = sched_clock();
4165 do_set_cpus_allowed(idle, cpumask_of(cpu));
4167 * We're having a chicken and egg problem, even though we are
4168 * holding rq->lock, the cpu isn't yet set to this cpu so the
4169 * lockdep check in task_group() will fail.
4171 * Similar case to sched_fork(). / Alternatively we could
4172 * use task_rq_lock() here and obtain the other rq->lock.
4177 __set_task_cpu(idle, cpu);
4180 rq->curr = rq->idle = idle;
4181 #if defined(CONFIG_SMP)
4184 raw_spin_unlock_irqrestore(&rq->lock, flags);
4186 /* Set the preempt count _outside_ the spinlocks! */
4187 task_thread_info(idle)->preempt_count = 0;
4190 * The idle tasks have their own, simple scheduling class:
4192 idle->sched_class = &idle_sched_class;
4193 ftrace_graph_init_idle_task(idle, cpu);
4194 vtime_init_idle(idle);
4195 #if defined(CONFIG_SMP)
4196 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4201 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4203 if (p->sched_class && p->sched_class->set_cpus_allowed)
4204 p->sched_class->set_cpus_allowed(p, new_mask);
4206 cpumask_copy(&p->cpus_allowed, new_mask);
4207 p->nr_cpus_allowed = cpumask_weight(new_mask);
4211 * This is how migration works:
4213 * 1) we invoke migration_cpu_stop() on the target CPU using
4215 * 2) stopper starts to run (implicitly forcing the migrated thread
4217 * 3) it checks whether the migrated task is still in the wrong runqueue.
4218 * 4) if it's in the wrong runqueue then the migration thread removes
4219 * it and puts it into the right queue.
4220 * 5) stopper completes and stop_one_cpu() returns and the migration
4225 * Change a given task's CPU affinity. Migrate the thread to a
4226 * proper CPU and schedule it away if the CPU it's executing on
4227 * is removed from the allowed bitmask.
4229 * NOTE: the caller must have a valid reference to the task, the
4230 * task must not exit() & deallocate itself prematurely. The
4231 * call is not atomic; no spinlocks may be held.
4233 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4235 unsigned long flags;
4237 unsigned int dest_cpu;
4240 rq = task_rq_lock(p, &flags);
4242 if (cpumask_equal(&p->cpus_allowed, new_mask))
4245 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4250 do_set_cpus_allowed(p, new_mask);
4252 /* Can the task run on the task's current CPU? If so, we're done */
4253 if (cpumask_test_cpu(task_cpu(p), new_mask))
4256 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4258 struct migration_arg arg = { p, dest_cpu };
4259 /* Need help from migration thread: drop lock and wait. */
4260 task_rq_unlock(rq, p, &flags);
4261 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4262 tlb_migrate_finish(p->mm);
4266 task_rq_unlock(rq, p, &flags);
4270 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4273 * Move (not current) task off this cpu, onto dest cpu. We're doing
4274 * this because either it can't run here any more (set_cpus_allowed()
4275 * away from this CPU, or CPU going down), or because we're
4276 * attempting to rebalance this task on exec (sched_exec).
4278 * So we race with normal scheduler movements, but that's OK, as long
4279 * as the task is no longer on this CPU.
4281 * Returns non-zero if task was successfully migrated.
4283 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4285 struct rq *rq_dest, *rq_src;
4288 if (unlikely(!cpu_active(dest_cpu)))
4291 rq_src = cpu_rq(src_cpu);
4292 rq_dest = cpu_rq(dest_cpu);
4294 raw_spin_lock(&p->pi_lock);
4295 double_rq_lock(rq_src, rq_dest);
4296 /* Already moved. */
4297 if (task_cpu(p) != src_cpu)
4299 /* Affinity changed (again). */
4300 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4304 * If we're not on a rq, the next wake-up will ensure we're
4308 dequeue_task(rq_src, p, 0);
4309 set_task_cpu(p, dest_cpu);
4310 enqueue_task(rq_dest, p, 0);
4311 check_preempt_curr(rq_dest, p, 0);
4316 double_rq_unlock(rq_src, rq_dest);
4317 raw_spin_unlock(&p->pi_lock);
4322 * migration_cpu_stop - this will be executed by a highprio stopper thread
4323 * and performs thread migration by bumping thread off CPU then
4324 * 'pushing' onto another runqueue.
4326 static int migration_cpu_stop(void *data)
4328 struct migration_arg *arg = data;
4331 * The original target cpu might have gone down and we might
4332 * be on another cpu but it doesn't matter.
4334 local_irq_disable();
4335 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4340 #ifdef CONFIG_HOTPLUG_CPU
4343 * Ensures that the idle task is using init_mm right before its cpu goes
4346 void idle_task_exit(void)
4348 struct mm_struct *mm = current->active_mm;
4350 BUG_ON(cpu_online(smp_processor_id()));
4353 switch_mm(mm, &init_mm, current);
4358 * Since this CPU is going 'away' for a while, fold any nr_active delta
4359 * we might have. Assumes we're called after migrate_tasks() so that the
4360 * nr_active count is stable.
4362 * Also see the comment "Global load-average calculations".
4364 static void calc_load_migrate(struct rq *rq)
4366 long delta = calc_load_fold_active(rq);
4368 atomic_long_add(delta, &calc_load_tasks);
4372 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4373 * try_to_wake_up()->select_task_rq().
4375 * Called with rq->lock held even though we'er in stop_machine() and
4376 * there's no concurrency possible, we hold the required locks anyway
4377 * because of lock validation efforts.
4379 static void migrate_tasks(unsigned int dead_cpu)
4381 struct rq *rq = cpu_rq(dead_cpu);
4382 struct task_struct *next, *stop = rq->stop;
4386 * Fudge the rq selection such that the below task selection loop
4387 * doesn't get stuck on the currently eligible stop task.
4389 * We're currently inside stop_machine() and the rq is either stuck
4390 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4391 * either way we should never end up calling schedule() until we're
4397 * put_prev_task() and pick_next_task() sched
4398 * class method both need to have an up-to-date
4399 * value of rq->clock[_task]
4401 update_rq_clock(rq);
4405 * There's this thread running, bail when that's the only
4408 if (rq->nr_running == 1)
4411 next = pick_next_task(rq);
4413 next->sched_class->put_prev_task(rq, next);
4415 /* Find suitable destination for @next, with force if needed. */
4416 dest_cpu = select_fallback_rq(dead_cpu, next);
4417 raw_spin_unlock(&rq->lock);
4419 __migrate_task(next, dead_cpu, dest_cpu);
4421 raw_spin_lock(&rq->lock);
4427 #endif /* CONFIG_HOTPLUG_CPU */
4429 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4431 static struct ctl_table sd_ctl_dir[] = {
4433 .procname = "sched_domain",
4439 static struct ctl_table sd_ctl_root[] = {
4441 .procname = "kernel",
4443 .child = sd_ctl_dir,
4448 static struct ctl_table *sd_alloc_ctl_entry(int n)
4450 struct ctl_table *entry =
4451 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4456 static void sd_free_ctl_entry(struct ctl_table **tablep)
4458 struct ctl_table *entry;
4461 * In the intermediate directories, both the child directory and
4462 * procname are dynamically allocated and could fail but the mode
4463 * will always be set. In the lowest directory the names are
4464 * static strings and all have proc handlers.
4466 for (entry = *tablep; entry->mode; entry++) {
4468 sd_free_ctl_entry(&entry->child);
4469 if (entry->proc_handler == NULL)
4470 kfree(entry->procname);
4477 static int min_load_idx = 0;
4478 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4481 set_table_entry(struct ctl_table *entry,
4482 const char *procname, void *data, int maxlen,
4483 umode_t mode, proc_handler *proc_handler,
4486 entry->procname = procname;
4488 entry->maxlen = maxlen;
4490 entry->proc_handler = proc_handler;
4493 entry->extra1 = &min_load_idx;
4494 entry->extra2 = &max_load_idx;
4498 static struct ctl_table *
4499 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4501 struct ctl_table *table = sd_alloc_ctl_entry(13);
4506 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4507 sizeof(long), 0644, proc_doulongvec_minmax, false);
4508 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4509 sizeof(long), 0644, proc_doulongvec_minmax, false);
4510 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4511 sizeof(int), 0644, proc_dointvec_minmax, true);
4512 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4513 sizeof(int), 0644, proc_dointvec_minmax, true);
4514 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4515 sizeof(int), 0644, proc_dointvec_minmax, true);
4516 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4517 sizeof(int), 0644, proc_dointvec_minmax, true);
4518 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4519 sizeof(int), 0644, proc_dointvec_minmax, true);
4520 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4521 sizeof(int), 0644, proc_dointvec_minmax, false);
4522 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4523 sizeof(int), 0644, proc_dointvec_minmax, false);
4524 set_table_entry(&table[9], "cache_nice_tries",
4525 &sd->cache_nice_tries,
4526 sizeof(int), 0644, proc_dointvec_minmax, false);
4527 set_table_entry(&table[10], "flags", &sd->flags,
4528 sizeof(int), 0644, proc_dointvec_minmax, false);
4529 set_table_entry(&table[11], "name", sd->name,
4530 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4531 /* &table[12] is terminator */
4536 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4538 struct ctl_table *entry, *table;
4539 struct sched_domain *sd;
4540 int domain_num = 0, i;
4543 for_each_domain(cpu, sd)
4545 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4550 for_each_domain(cpu, sd) {
4551 snprintf(buf, 32, "domain%d", i);
4552 entry->procname = kstrdup(buf, GFP_KERNEL);
4554 entry->child = sd_alloc_ctl_domain_table(sd);
4561 static struct ctl_table_header *sd_sysctl_header;
4562 static void register_sched_domain_sysctl(void)
4564 int i, cpu_num = num_possible_cpus();
4565 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4568 WARN_ON(sd_ctl_dir[0].child);
4569 sd_ctl_dir[0].child = entry;
4574 for_each_possible_cpu(i) {
4575 snprintf(buf, 32, "cpu%d", i);
4576 entry->procname = kstrdup(buf, GFP_KERNEL);
4578 entry->child = sd_alloc_ctl_cpu_table(i);
4582 WARN_ON(sd_sysctl_header);
4583 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4586 /* may be called multiple times per register */
4587 static void unregister_sched_domain_sysctl(void)
4589 if (sd_sysctl_header)
4590 unregister_sysctl_table(sd_sysctl_header);
4591 sd_sysctl_header = NULL;
4592 if (sd_ctl_dir[0].child)
4593 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4596 static void register_sched_domain_sysctl(void)
4599 static void unregister_sched_domain_sysctl(void)
4604 static void set_rq_online(struct rq *rq)
4607 const struct sched_class *class;
4609 cpumask_set_cpu(rq->cpu, rq->rd->online);
4612 for_each_class(class) {
4613 if (class->rq_online)
4614 class->rq_online(rq);
4619 static void set_rq_offline(struct rq *rq)
4622 const struct sched_class *class;
4624 for_each_class(class) {
4625 if (class->rq_offline)
4626 class->rq_offline(rq);
4629 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4635 * migration_call - callback that gets triggered when a CPU is added.
4636 * Here we can start up the necessary migration thread for the new CPU.
4638 static int __cpuinit
4639 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4641 int cpu = (long)hcpu;
4642 unsigned long flags;
4643 struct rq *rq = cpu_rq(cpu);
4645 switch (action & ~CPU_TASKS_FROZEN) {
4647 case CPU_UP_PREPARE:
4648 rq->calc_load_update = calc_load_update;
4652 /* Update our root-domain */
4653 raw_spin_lock_irqsave(&rq->lock, flags);
4655 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4659 raw_spin_unlock_irqrestore(&rq->lock, flags);
4662 #ifdef CONFIG_HOTPLUG_CPU
4664 sched_ttwu_pending();
4665 /* Update our root-domain */
4666 raw_spin_lock_irqsave(&rq->lock, flags);
4668 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4672 BUG_ON(rq->nr_running != 1); /* the migration thread */
4673 raw_spin_unlock_irqrestore(&rq->lock, flags);
4677 calc_load_migrate(rq);
4682 update_max_interval();
4688 * Register at high priority so that task migration (migrate_all_tasks)
4689 * happens before everything else. This has to be lower priority than
4690 * the notifier in the perf_event subsystem, though.
4692 static struct notifier_block __cpuinitdata migration_notifier = {
4693 .notifier_call = migration_call,
4694 .priority = CPU_PRI_MIGRATION,
4697 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
4698 unsigned long action, void *hcpu)
4700 switch (action & ~CPU_TASKS_FROZEN) {
4702 case CPU_DOWN_FAILED:
4703 set_cpu_active((long)hcpu, true);
4710 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
4711 unsigned long action, void *hcpu)
4713 switch (action & ~CPU_TASKS_FROZEN) {
4714 case CPU_DOWN_PREPARE:
4715 set_cpu_active((long)hcpu, false);
4722 static int __init migration_init(void)
4724 void *cpu = (void *)(long)smp_processor_id();
4727 /* Initialize migration for the boot CPU */
4728 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4729 BUG_ON(err == NOTIFY_BAD);
4730 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4731 register_cpu_notifier(&migration_notifier);
4733 /* Register cpu active notifiers */
4734 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4735 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4739 early_initcall(migration_init);
4744 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4746 #ifdef CONFIG_SCHED_DEBUG
4748 static __read_mostly int sched_debug_enabled;
4750 static int __init sched_debug_setup(char *str)
4752 sched_debug_enabled = 1;
4756 early_param("sched_debug", sched_debug_setup);
4758 static inline bool sched_debug(void)
4760 return sched_debug_enabled;
4763 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4764 struct cpumask *groupmask)
4766 struct sched_group *group = sd->groups;
4769 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4770 cpumask_clear(groupmask);
4772 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4774 if (!(sd->flags & SD_LOAD_BALANCE)) {
4775 printk("does not load-balance\n");
4777 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4782 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4784 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4785 printk(KERN_ERR "ERROR: domain->span does not contain "
4788 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4789 printk(KERN_ERR "ERROR: domain->groups does not contain"
4793 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4797 printk(KERN_ERR "ERROR: group is NULL\n");
4802 * Even though we initialize ->power to something semi-sane,
4803 * we leave power_orig unset. This allows us to detect if
4804 * domain iteration is still funny without causing /0 traps.
4806 if (!group->sgp->power_orig) {
4807 printk(KERN_CONT "\n");
4808 printk(KERN_ERR "ERROR: domain->cpu_power not "
4813 if (!cpumask_weight(sched_group_cpus(group))) {
4814 printk(KERN_CONT "\n");
4815 printk(KERN_ERR "ERROR: empty group\n");
4819 if (!(sd->flags & SD_OVERLAP) &&
4820 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4821 printk(KERN_CONT "\n");
4822 printk(KERN_ERR "ERROR: repeated CPUs\n");
4826 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4828 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4830 printk(KERN_CONT " %s", str);
4831 if (group->sgp->power != SCHED_POWER_SCALE) {
4832 printk(KERN_CONT " (cpu_power = %d)",
4836 group = group->next;
4837 } while (group != sd->groups);
4838 printk(KERN_CONT "\n");
4840 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4841 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4844 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4845 printk(KERN_ERR "ERROR: parent span is not a superset "
4846 "of domain->span\n");
4850 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4854 if (!sched_debug_enabled)
4858 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4862 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4865 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4873 #else /* !CONFIG_SCHED_DEBUG */
4874 # define sched_domain_debug(sd, cpu) do { } while (0)
4875 static inline bool sched_debug(void)
4879 #endif /* CONFIG_SCHED_DEBUG */
4881 static int sd_degenerate(struct sched_domain *sd)
4883 if (cpumask_weight(sched_domain_span(sd)) == 1)
4886 /* Following flags need at least 2 groups */
4887 if (sd->flags & (SD_LOAD_BALANCE |
4888 SD_BALANCE_NEWIDLE |
4892 SD_SHARE_PKG_RESOURCES)) {
4893 if (sd->groups != sd->groups->next)
4897 /* Following flags don't use groups */
4898 if (sd->flags & (SD_WAKE_AFFINE))
4905 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4907 unsigned long cflags = sd->flags, pflags = parent->flags;
4909 if (sd_degenerate(parent))
4912 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4915 /* Flags needing groups don't count if only 1 group in parent */
4916 if (parent->groups == parent->groups->next) {
4917 pflags &= ~(SD_LOAD_BALANCE |
4918 SD_BALANCE_NEWIDLE |
4922 SD_SHARE_PKG_RESOURCES);
4923 if (nr_node_ids == 1)
4924 pflags &= ~SD_SERIALIZE;
4926 if (~cflags & pflags)
4932 static void free_rootdomain(struct rcu_head *rcu)
4934 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4936 cpupri_cleanup(&rd->cpupri);
4937 free_cpumask_var(rd->rto_mask);
4938 free_cpumask_var(rd->online);
4939 free_cpumask_var(rd->span);
4943 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4945 struct root_domain *old_rd = NULL;
4946 unsigned long flags;
4948 raw_spin_lock_irqsave(&rq->lock, flags);
4953 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4956 cpumask_clear_cpu(rq->cpu, old_rd->span);
4959 * If we dont want to free the old_rt yet then
4960 * set old_rd to NULL to skip the freeing later
4963 if (!atomic_dec_and_test(&old_rd->refcount))
4967 atomic_inc(&rd->refcount);
4970 cpumask_set_cpu(rq->cpu, rd->span);
4971 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4974 raw_spin_unlock_irqrestore(&rq->lock, flags);
4977 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4980 static int init_rootdomain(struct root_domain *rd)
4982 memset(rd, 0, sizeof(*rd));
4984 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4986 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4988 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4991 if (cpupri_init(&rd->cpupri) != 0)
4996 free_cpumask_var(rd->rto_mask);
4998 free_cpumask_var(rd->online);
5000 free_cpumask_var(rd->span);
5006 * By default the system creates a single root-domain with all cpus as
5007 * members (mimicking the global state we have today).
5009 struct root_domain def_root_domain;
5011 static void init_defrootdomain(void)
5013 init_rootdomain(&def_root_domain);
5015 atomic_set(&def_root_domain.refcount, 1);
5018 static struct root_domain *alloc_rootdomain(void)
5020 struct root_domain *rd;
5022 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5026 if (init_rootdomain(rd) != 0) {
5034 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5036 struct sched_group *tmp, *first;
5045 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5050 } while (sg != first);
5053 static void free_sched_domain(struct rcu_head *rcu)
5055 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5058 * If its an overlapping domain it has private groups, iterate and
5061 if (sd->flags & SD_OVERLAP) {
5062 free_sched_groups(sd->groups, 1);
5063 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5064 kfree(sd->groups->sgp);
5070 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5072 call_rcu(&sd->rcu, free_sched_domain);
5075 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5077 for (; sd; sd = sd->parent)
5078 destroy_sched_domain(sd, cpu);
5082 * Keep a special pointer to the highest sched_domain that has
5083 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5084 * allows us to avoid some pointer chasing select_idle_sibling().
5086 * Also keep a unique ID per domain (we use the first cpu number in
5087 * the cpumask of the domain), this allows us to quickly tell if
5088 * two cpus are in the same cache domain, see cpus_share_cache().
5090 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5091 DEFINE_PER_CPU(int, sd_llc_id);
5093 static void update_top_cache_domain(int cpu)
5095 struct sched_domain *sd;
5098 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5100 id = cpumask_first(sched_domain_span(sd));
5102 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5103 per_cpu(sd_llc_id, cpu) = id;
5107 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5108 * hold the hotplug lock.
5111 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5113 struct rq *rq = cpu_rq(cpu);
5114 struct sched_domain *tmp;
5116 /* Remove the sched domains which do not contribute to scheduling. */
5117 for (tmp = sd; tmp; ) {
5118 struct sched_domain *parent = tmp->parent;
5122 if (sd_parent_degenerate(tmp, parent)) {
5123 tmp->parent = parent->parent;
5125 parent->parent->child = tmp;
5126 destroy_sched_domain(parent, cpu);
5131 if (sd && sd_degenerate(sd)) {
5134 destroy_sched_domain(tmp, cpu);
5139 sched_domain_debug(sd, cpu);
5141 rq_attach_root(rq, rd);
5143 rcu_assign_pointer(rq->sd, sd);
5144 destroy_sched_domains(tmp, cpu);
5146 update_top_cache_domain(cpu);
5149 /* cpus with isolated domains */
5150 static cpumask_var_t cpu_isolated_map;
5152 /* Setup the mask of cpus configured for isolated domains */
5153 static int __init isolated_cpu_setup(char *str)
5155 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5156 cpulist_parse(str, cpu_isolated_map);
5160 __setup("isolcpus=", isolated_cpu_setup);
5162 static const struct cpumask *cpu_cpu_mask(int cpu)
5164 return cpumask_of_node(cpu_to_node(cpu));
5168 struct sched_domain **__percpu sd;
5169 struct sched_group **__percpu sg;
5170 struct sched_group_power **__percpu sgp;
5174 struct sched_domain ** __percpu sd;
5175 struct root_domain *rd;
5185 struct sched_domain_topology_level;
5187 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5188 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5190 #define SDTL_OVERLAP 0x01
5192 struct sched_domain_topology_level {
5193 sched_domain_init_f init;
5194 sched_domain_mask_f mask;
5197 struct sd_data data;
5201 * Build an iteration mask that can exclude certain CPUs from the upwards
5204 * Asymmetric node setups can result in situations where the domain tree is of
5205 * unequal depth, make sure to skip domains that already cover the entire
5208 * In that case build_sched_domains() will have terminated the iteration early
5209 * and our sibling sd spans will be empty. Domains should always include the
5210 * cpu they're built on, so check that.
5213 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5215 const struct cpumask *span = sched_domain_span(sd);
5216 struct sd_data *sdd = sd->private;
5217 struct sched_domain *sibling;
5220 for_each_cpu(i, span) {
5221 sibling = *per_cpu_ptr(sdd->sd, i);
5222 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5225 cpumask_set_cpu(i, sched_group_mask(sg));
5230 * Return the canonical balance cpu for this group, this is the first cpu
5231 * of this group that's also in the iteration mask.
5233 int group_balance_cpu(struct sched_group *sg)
5235 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5239 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5241 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5242 const struct cpumask *span = sched_domain_span(sd);
5243 struct cpumask *covered = sched_domains_tmpmask;
5244 struct sd_data *sdd = sd->private;
5245 struct sched_domain *child;
5248 cpumask_clear(covered);
5250 for_each_cpu(i, span) {
5251 struct cpumask *sg_span;
5253 if (cpumask_test_cpu(i, covered))
5256 child = *per_cpu_ptr(sdd->sd, i);
5258 /* See the comment near build_group_mask(). */
5259 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5262 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5263 GFP_KERNEL, cpu_to_node(cpu));
5268 sg_span = sched_group_cpus(sg);
5270 child = child->child;
5271 cpumask_copy(sg_span, sched_domain_span(child));
5273 cpumask_set_cpu(i, sg_span);
5275 cpumask_or(covered, covered, sg_span);
5277 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5278 if (atomic_inc_return(&sg->sgp->ref) == 1)
5279 build_group_mask(sd, sg);
5282 * Initialize sgp->power such that even if we mess up the
5283 * domains and no possible iteration will get us here, we won't
5286 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5289 * Make sure the first group of this domain contains the
5290 * canonical balance cpu. Otherwise the sched_domain iteration
5291 * breaks. See update_sg_lb_stats().
5293 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5294 group_balance_cpu(sg) == cpu)
5304 sd->groups = groups;
5309 free_sched_groups(first, 0);
5314 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5316 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5317 struct sched_domain *child = sd->child;
5320 cpu = cpumask_first(sched_domain_span(child));
5323 *sg = *per_cpu_ptr(sdd->sg, cpu);
5324 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5325 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5332 * build_sched_groups will build a circular linked list of the groups
5333 * covered by the given span, and will set each group's ->cpumask correctly,
5334 * and ->cpu_power to 0.
5336 * Assumes the sched_domain tree is fully constructed
5339 build_sched_groups(struct sched_domain *sd, int cpu)
5341 struct sched_group *first = NULL, *last = NULL;
5342 struct sd_data *sdd = sd->private;
5343 const struct cpumask *span = sched_domain_span(sd);
5344 struct cpumask *covered;
5347 get_group(cpu, sdd, &sd->groups);
5348 atomic_inc(&sd->groups->ref);
5350 if (cpu != cpumask_first(sched_domain_span(sd)))
5353 lockdep_assert_held(&sched_domains_mutex);
5354 covered = sched_domains_tmpmask;
5356 cpumask_clear(covered);
5358 for_each_cpu(i, span) {
5359 struct sched_group *sg;
5360 int group = get_group(i, sdd, &sg);
5363 if (cpumask_test_cpu(i, covered))
5366 cpumask_clear(sched_group_cpus(sg));
5368 cpumask_setall(sched_group_mask(sg));
5370 for_each_cpu(j, span) {
5371 if (get_group(j, sdd, NULL) != group)
5374 cpumask_set_cpu(j, covered);
5375 cpumask_set_cpu(j, sched_group_cpus(sg));
5390 * Initialize sched groups cpu_power.
5392 * cpu_power indicates the capacity of sched group, which is used while
5393 * distributing the load between different sched groups in a sched domain.
5394 * Typically cpu_power for all the groups in a sched domain will be same unless
5395 * there are asymmetries in the topology. If there are asymmetries, group
5396 * having more cpu_power will pickup more load compared to the group having
5399 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5401 struct sched_group *sg = sd->groups;
5403 WARN_ON(!sd || !sg);
5406 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5408 } while (sg != sd->groups);
5410 if (cpu != group_balance_cpu(sg))
5413 update_group_power(sd, cpu);
5414 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5417 int __weak arch_sd_sibling_asym_packing(void)
5419 return 0*SD_ASYM_PACKING;
5423 * Initializers for schedule domains
5424 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5427 #ifdef CONFIG_SCHED_DEBUG
5428 # define SD_INIT_NAME(sd, type) sd->name = #type
5430 # define SD_INIT_NAME(sd, type) do { } while (0)
5433 #define SD_INIT_FUNC(type) \
5434 static noinline struct sched_domain * \
5435 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5437 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5438 *sd = SD_##type##_INIT; \
5439 SD_INIT_NAME(sd, type); \
5440 sd->private = &tl->data; \
5445 #ifdef CONFIG_SCHED_SMT
5446 SD_INIT_FUNC(SIBLING)
5448 #ifdef CONFIG_SCHED_MC
5451 #ifdef CONFIG_SCHED_BOOK
5455 static int default_relax_domain_level = -1;
5456 int sched_domain_level_max;
5458 static int __init setup_relax_domain_level(char *str)
5460 if (kstrtoint(str, 0, &default_relax_domain_level))
5461 pr_warn("Unable to set relax_domain_level\n");
5465 __setup("relax_domain_level=", setup_relax_domain_level);
5467 static void set_domain_attribute(struct sched_domain *sd,
5468 struct sched_domain_attr *attr)
5472 if (!attr || attr->relax_domain_level < 0) {
5473 if (default_relax_domain_level < 0)
5476 request = default_relax_domain_level;
5478 request = attr->relax_domain_level;
5479 if (request < sd->level) {
5480 /* turn off idle balance on this domain */
5481 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5483 /* turn on idle balance on this domain */
5484 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5488 static void __sdt_free(const struct cpumask *cpu_map);
5489 static int __sdt_alloc(const struct cpumask *cpu_map);
5491 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5492 const struct cpumask *cpu_map)
5496 if (!atomic_read(&d->rd->refcount))
5497 free_rootdomain(&d->rd->rcu); /* fall through */
5499 free_percpu(d->sd); /* fall through */
5501 __sdt_free(cpu_map); /* fall through */
5507 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5508 const struct cpumask *cpu_map)
5510 memset(d, 0, sizeof(*d));
5512 if (__sdt_alloc(cpu_map))
5513 return sa_sd_storage;
5514 d->sd = alloc_percpu(struct sched_domain *);
5516 return sa_sd_storage;
5517 d->rd = alloc_rootdomain();
5520 return sa_rootdomain;
5524 * NULL the sd_data elements we've used to build the sched_domain and
5525 * sched_group structure so that the subsequent __free_domain_allocs()
5526 * will not free the data we're using.
5528 static void claim_allocations(int cpu, struct sched_domain *sd)
5530 struct sd_data *sdd = sd->private;
5532 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5533 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5535 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5536 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5538 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5539 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5542 #ifdef CONFIG_SCHED_SMT
5543 static const struct cpumask *cpu_smt_mask(int cpu)
5545 return topology_thread_cpumask(cpu);
5550 * Topology list, bottom-up.
5552 static struct sched_domain_topology_level default_topology[] = {
5553 #ifdef CONFIG_SCHED_SMT
5554 { sd_init_SIBLING, cpu_smt_mask, },
5556 #ifdef CONFIG_SCHED_MC
5557 { sd_init_MC, cpu_coregroup_mask, },
5559 #ifdef CONFIG_SCHED_BOOK
5560 { sd_init_BOOK, cpu_book_mask, },
5562 { sd_init_CPU, cpu_cpu_mask, },
5566 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5570 static int sched_domains_numa_levels;
5571 static int *sched_domains_numa_distance;
5572 static struct cpumask ***sched_domains_numa_masks;
5573 static int sched_domains_curr_level;
5575 static inline int sd_local_flags(int level)
5577 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5580 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5583 static struct sched_domain *
5584 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5586 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5587 int level = tl->numa_level;
5588 int sd_weight = cpumask_weight(
5589 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5591 *sd = (struct sched_domain){
5592 .min_interval = sd_weight,
5593 .max_interval = 2*sd_weight,
5595 .imbalance_pct = 125,
5596 .cache_nice_tries = 2,
5603 .flags = 1*SD_LOAD_BALANCE
5604 | 1*SD_BALANCE_NEWIDLE
5609 | 0*SD_SHARE_CPUPOWER
5610 | 0*SD_SHARE_PKG_RESOURCES
5612 | 0*SD_PREFER_SIBLING
5613 | sd_local_flags(level)
5615 .last_balance = jiffies,
5616 .balance_interval = sd_weight,
5618 SD_INIT_NAME(sd, NUMA);
5619 sd->private = &tl->data;
5622 * Ugly hack to pass state to sd_numa_mask()...
5624 sched_domains_curr_level = tl->numa_level;
5629 static const struct cpumask *sd_numa_mask(int cpu)
5631 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5634 static void sched_numa_warn(const char *str)
5636 static int done = false;
5644 printk(KERN_WARNING "ERROR: %s\n\n", str);
5646 for (i = 0; i < nr_node_ids; i++) {
5647 printk(KERN_WARNING " ");
5648 for (j = 0; j < nr_node_ids; j++)
5649 printk(KERN_CONT "%02d ", node_distance(i,j));
5650 printk(KERN_CONT "\n");
5652 printk(KERN_WARNING "\n");
5655 static bool find_numa_distance(int distance)
5659 if (distance == node_distance(0, 0))
5662 for (i = 0; i < sched_domains_numa_levels; i++) {
5663 if (sched_domains_numa_distance[i] == distance)
5670 static void sched_init_numa(void)
5672 int next_distance, curr_distance = node_distance(0, 0);
5673 struct sched_domain_topology_level *tl;
5677 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5678 if (!sched_domains_numa_distance)
5682 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5683 * unique distances in the node_distance() table.
5685 * Assumes node_distance(0,j) includes all distances in
5686 * node_distance(i,j) in order to avoid cubic time.
5688 next_distance = curr_distance;
5689 for (i = 0; i < nr_node_ids; i++) {
5690 for (j = 0; j < nr_node_ids; j++) {
5691 for (k = 0; k < nr_node_ids; k++) {
5692 int distance = node_distance(i, k);
5694 if (distance > curr_distance &&
5695 (distance < next_distance ||
5696 next_distance == curr_distance))
5697 next_distance = distance;
5700 * While not a strong assumption it would be nice to know
5701 * about cases where if node A is connected to B, B is not
5702 * equally connected to A.
5704 if (sched_debug() && node_distance(k, i) != distance)
5705 sched_numa_warn("Node-distance not symmetric");
5707 if (sched_debug() && i && !find_numa_distance(distance))
5708 sched_numa_warn("Node-0 not representative");
5710 if (next_distance != curr_distance) {
5711 sched_domains_numa_distance[level++] = next_distance;
5712 sched_domains_numa_levels = level;
5713 curr_distance = next_distance;
5718 * In case of sched_debug() we verify the above assumption.
5724 * 'level' contains the number of unique distances, excluding the
5725 * identity distance node_distance(i,i).
5727 * The sched_domains_numa_distance[] array includes the actual distance
5732 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5733 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5734 * the array will contain less then 'level' members. This could be
5735 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5736 * in other functions.
5738 * We reset it to 'level' at the end of this function.
5740 sched_domains_numa_levels = 0;
5742 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5743 if (!sched_domains_numa_masks)
5747 * Now for each level, construct a mask per node which contains all
5748 * cpus of nodes that are that many hops away from us.
5750 for (i = 0; i < level; i++) {
5751 sched_domains_numa_masks[i] =
5752 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5753 if (!sched_domains_numa_masks[i])
5756 for (j = 0; j < nr_node_ids; j++) {
5757 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5761 sched_domains_numa_masks[i][j] = mask;
5763 for (k = 0; k < nr_node_ids; k++) {
5764 if (node_distance(j, k) > sched_domains_numa_distance[i])
5767 cpumask_or(mask, mask, cpumask_of_node(k));
5772 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5773 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5778 * Copy the default topology bits..
5780 for (i = 0; default_topology[i].init; i++)
5781 tl[i] = default_topology[i];
5784 * .. and append 'j' levels of NUMA goodness.
5786 for (j = 0; j < level; i++, j++) {
5787 tl[i] = (struct sched_domain_topology_level){
5788 .init = sd_numa_init,
5789 .mask = sd_numa_mask,
5790 .flags = SDTL_OVERLAP,
5795 sched_domain_topology = tl;
5797 sched_domains_numa_levels = level;
5800 static void sched_domains_numa_masks_set(int cpu)
5803 int node = cpu_to_node(cpu);
5805 for (i = 0; i < sched_domains_numa_levels; i++) {
5806 for (j = 0; j < nr_node_ids; j++) {
5807 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5808 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5813 static void sched_domains_numa_masks_clear(int cpu)
5816 for (i = 0; i < sched_domains_numa_levels; i++) {
5817 for (j = 0; j < nr_node_ids; j++)
5818 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5823 * Update sched_domains_numa_masks[level][node] array when new cpus
5826 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5827 unsigned long action,
5830 int cpu = (long)hcpu;
5832 switch (action & ~CPU_TASKS_FROZEN) {
5834 sched_domains_numa_masks_set(cpu);
5838 sched_domains_numa_masks_clear(cpu);
5848 static inline void sched_init_numa(void)
5852 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5853 unsigned long action,
5858 #endif /* CONFIG_NUMA */
5860 static int __sdt_alloc(const struct cpumask *cpu_map)
5862 struct sched_domain_topology_level *tl;
5865 for (tl = sched_domain_topology; tl->init; tl++) {
5866 struct sd_data *sdd = &tl->data;
5868 sdd->sd = alloc_percpu(struct sched_domain *);
5872 sdd->sg = alloc_percpu(struct sched_group *);
5876 sdd->sgp = alloc_percpu(struct sched_group_power *);
5880 for_each_cpu(j, cpu_map) {
5881 struct sched_domain *sd;
5882 struct sched_group *sg;
5883 struct sched_group_power *sgp;
5885 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5886 GFP_KERNEL, cpu_to_node(j));
5890 *per_cpu_ptr(sdd->sd, j) = sd;
5892 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5893 GFP_KERNEL, cpu_to_node(j));
5899 *per_cpu_ptr(sdd->sg, j) = sg;
5901 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5902 GFP_KERNEL, cpu_to_node(j));
5906 *per_cpu_ptr(sdd->sgp, j) = sgp;
5913 static void __sdt_free(const struct cpumask *cpu_map)
5915 struct sched_domain_topology_level *tl;
5918 for (tl = sched_domain_topology; tl->init; tl++) {
5919 struct sd_data *sdd = &tl->data;
5921 for_each_cpu(j, cpu_map) {
5922 struct sched_domain *sd;
5925 sd = *per_cpu_ptr(sdd->sd, j);
5926 if (sd && (sd->flags & SD_OVERLAP))
5927 free_sched_groups(sd->groups, 0);
5928 kfree(*per_cpu_ptr(sdd->sd, j));
5932 kfree(*per_cpu_ptr(sdd->sg, j));
5934 kfree(*per_cpu_ptr(sdd->sgp, j));
5936 free_percpu(sdd->sd);
5938 free_percpu(sdd->sg);
5940 free_percpu(sdd->sgp);
5945 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5946 struct s_data *d, const struct cpumask *cpu_map,
5947 struct sched_domain_attr *attr, struct sched_domain *child,
5950 struct sched_domain *sd = tl->init(tl, cpu);
5954 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5956 sd->level = child->level + 1;
5957 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5961 set_domain_attribute(sd, attr);
5967 * Build sched domains for a given set of cpus and attach the sched domains
5968 * to the individual cpus
5970 static int build_sched_domains(const struct cpumask *cpu_map,
5971 struct sched_domain_attr *attr)
5973 enum s_alloc alloc_state = sa_none;
5974 struct sched_domain *sd;
5976 int i, ret = -ENOMEM;
5978 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5979 if (alloc_state != sa_rootdomain)
5982 /* Set up domains for cpus specified by the cpu_map. */
5983 for_each_cpu(i, cpu_map) {
5984 struct sched_domain_topology_level *tl;
5987 for (tl = sched_domain_topology; tl->init; tl++) {
5988 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
5989 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
5990 sd->flags |= SD_OVERLAP;
5991 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
5998 *per_cpu_ptr(d.sd, i) = sd;
6001 /* Build the groups for the domains */
6002 for_each_cpu(i, cpu_map) {
6003 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6004 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6005 if (sd->flags & SD_OVERLAP) {
6006 if (build_overlap_sched_groups(sd, i))
6009 if (build_sched_groups(sd, i))
6015 /* Calculate CPU power for physical packages and nodes */
6016 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6017 if (!cpumask_test_cpu(i, cpu_map))
6020 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6021 claim_allocations(i, sd);
6022 init_sched_groups_power(i, sd);
6026 /* Attach the domains */
6028 for_each_cpu(i, cpu_map) {
6029 sd = *per_cpu_ptr(d.sd, i);
6030 cpu_attach_domain(sd, d.rd, i);
6036 __free_domain_allocs(&d, alloc_state, cpu_map);
6040 static cpumask_var_t *doms_cur; /* current sched domains */
6041 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6042 static struct sched_domain_attr *dattr_cur;
6043 /* attribues of custom domains in 'doms_cur' */
6046 * Special case: If a kmalloc of a doms_cur partition (array of
6047 * cpumask) fails, then fallback to a single sched domain,
6048 * as determined by the single cpumask fallback_doms.
6050 static cpumask_var_t fallback_doms;
6053 * arch_update_cpu_topology lets virtualized architectures update the
6054 * cpu core maps. It is supposed to return 1 if the topology changed
6055 * or 0 if it stayed the same.
6057 int __attribute__((weak)) arch_update_cpu_topology(void)
6062 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6065 cpumask_var_t *doms;
6067 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6070 for (i = 0; i < ndoms; i++) {
6071 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6072 free_sched_domains(doms, i);
6079 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6082 for (i = 0; i < ndoms; i++)
6083 free_cpumask_var(doms[i]);
6088 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6089 * For now this just excludes isolated cpus, but could be used to
6090 * exclude other special cases in the future.
6092 static int init_sched_domains(const struct cpumask *cpu_map)
6096 arch_update_cpu_topology();
6098 doms_cur = alloc_sched_domains(ndoms_cur);
6100 doms_cur = &fallback_doms;
6101 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6102 err = build_sched_domains(doms_cur[0], NULL);
6103 register_sched_domain_sysctl();
6109 * Detach sched domains from a group of cpus specified in cpu_map
6110 * These cpus will now be attached to the NULL domain
6112 static void detach_destroy_domains(const struct cpumask *cpu_map)
6117 for_each_cpu(i, cpu_map)
6118 cpu_attach_domain(NULL, &def_root_domain, i);
6122 /* handle null as "default" */
6123 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6124 struct sched_domain_attr *new, int idx_new)
6126 struct sched_domain_attr tmp;
6133 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6134 new ? (new + idx_new) : &tmp,
6135 sizeof(struct sched_domain_attr));
6139 * Partition sched domains as specified by the 'ndoms_new'
6140 * cpumasks in the array doms_new[] of cpumasks. This compares
6141 * doms_new[] to the current sched domain partitioning, doms_cur[].
6142 * It destroys each deleted domain and builds each new domain.
6144 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6145 * The masks don't intersect (don't overlap.) We should setup one
6146 * sched domain for each mask. CPUs not in any of the cpumasks will
6147 * not be load balanced. If the same cpumask appears both in the
6148 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6151 * The passed in 'doms_new' should be allocated using
6152 * alloc_sched_domains. This routine takes ownership of it and will
6153 * free_sched_domains it when done with it. If the caller failed the
6154 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6155 * and partition_sched_domains() will fallback to the single partition
6156 * 'fallback_doms', it also forces the domains to be rebuilt.
6158 * If doms_new == NULL it will be replaced with cpu_online_mask.
6159 * ndoms_new == 0 is a special case for destroying existing domains,
6160 * and it will not create the default domain.
6162 * Call with hotplug lock held
6164 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6165 struct sched_domain_attr *dattr_new)
6170 mutex_lock(&sched_domains_mutex);
6172 /* always unregister in case we don't destroy any domains */
6173 unregister_sched_domain_sysctl();
6175 /* Let architecture update cpu core mappings. */
6176 new_topology = arch_update_cpu_topology();
6178 n = doms_new ? ndoms_new : 0;
6180 /* Destroy deleted domains */
6181 for (i = 0; i < ndoms_cur; i++) {
6182 for (j = 0; j < n && !new_topology; j++) {
6183 if (cpumask_equal(doms_cur[i], doms_new[j])
6184 && dattrs_equal(dattr_cur, i, dattr_new, j))
6187 /* no match - a current sched domain not in new doms_new[] */
6188 detach_destroy_domains(doms_cur[i]);
6193 if (doms_new == NULL) {
6195 doms_new = &fallback_doms;
6196 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6197 WARN_ON_ONCE(dattr_new);
6200 /* Build new domains */
6201 for (i = 0; i < ndoms_new; i++) {
6202 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6203 if (cpumask_equal(doms_new[i], doms_cur[j])
6204 && dattrs_equal(dattr_new, i, dattr_cur, j))
6207 /* no match - add a new doms_new */
6208 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6213 /* Remember the new sched domains */
6214 if (doms_cur != &fallback_doms)
6215 free_sched_domains(doms_cur, ndoms_cur);
6216 kfree(dattr_cur); /* kfree(NULL) is safe */
6217 doms_cur = doms_new;
6218 dattr_cur = dattr_new;
6219 ndoms_cur = ndoms_new;
6221 register_sched_domain_sysctl();
6223 mutex_unlock(&sched_domains_mutex);
6226 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6229 * Update cpusets according to cpu_active mask. If cpusets are
6230 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6231 * around partition_sched_domains().
6233 * If we come here as part of a suspend/resume, don't touch cpusets because we
6234 * want to restore it back to its original state upon resume anyway.
6236 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6240 case CPU_ONLINE_FROZEN:
6241 case CPU_DOWN_FAILED_FROZEN:
6244 * num_cpus_frozen tracks how many CPUs are involved in suspend
6245 * resume sequence. As long as this is not the last online
6246 * operation in the resume sequence, just build a single sched
6247 * domain, ignoring cpusets.
6250 if (likely(num_cpus_frozen)) {
6251 partition_sched_domains(1, NULL, NULL);
6256 * This is the last CPU online operation. So fall through and
6257 * restore the original sched domains by considering the
6258 * cpuset configurations.
6262 case CPU_DOWN_FAILED:
6263 cpuset_update_active_cpus(true);
6271 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6275 case CPU_DOWN_PREPARE:
6276 cpuset_update_active_cpus(false);
6278 case CPU_DOWN_PREPARE_FROZEN:
6280 partition_sched_domains(1, NULL, NULL);
6288 void __init sched_init_smp(void)
6290 cpumask_var_t non_isolated_cpus;
6292 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6293 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6298 mutex_lock(&sched_domains_mutex);
6299 init_sched_domains(cpu_active_mask);
6300 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6301 if (cpumask_empty(non_isolated_cpus))
6302 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6303 mutex_unlock(&sched_domains_mutex);
6306 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6307 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6308 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6312 /* Move init over to a non-isolated CPU */
6313 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6315 sched_init_granularity();
6316 free_cpumask_var(non_isolated_cpus);
6318 init_sched_rt_class();
6321 void __init sched_init_smp(void)
6323 sched_init_granularity();
6325 #endif /* CONFIG_SMP */
6327 const_debug unsigned int sysctl_timer_migration = 1;
6329 int in_sched_functions(unsigned long addr)
6331 return in_lock_functions(addr) ||
6332 (addr >= (unsigned long)__sched_text_start
6333 && addr < (unsigned long)__sched_text_end);
6336 #ifdef CONFIG_CGROUP_SCHED
6338 * Default task group.
6339 * Every task in system belongs to this group at bootup.
6341 struct task_group root_task_group;
6342 LIST_HEAD(task_groups);
6345 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6347 void __init sched_init(void)
6350 unsigned long alloc_size = 0, ptr;
6352 #ifdef CONFIG_FAIR_GROUP_SCHED
6353 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6355 #ifdef CONFIG_RT_GROUP_SCHED
6356 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6358 #ifdef CONFIG_CPUMASK_OFFSTACK
6359 alloc_size += num_possible_cpus() * cpumask_size();
6362 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6364 #ifdef CONFIG_FAIR_GROUP_SCHED
6365 root_task_group.se = (struct sched_entity **)ptr;
6366 ptr += nr_cpu_ids * sizeof(void **);
6368 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6369 ptr += nr_cpu_ids * sizeof(void **);
6371 #endif /* CONFIG_FAIR_GROUP_SCHED */
6372 #ifdef CONFIG_RT_GROUP_SCHED
6373 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6374 ptr += nr_cpu_ids * sizeof(void **);
6376 root_task_group.rt_rq = (struct rt_rq **)ptr;
6377 ptr += nr_cpu_ids * sizeof(void **);
6379 #endif /* CONFIG_RT_GROUP_SCHED */
6380 #ifdef CONFIG_CPUMASK_OFFSTACK
6381 for_each_possible_cpu(i) {
6382 per_cpu(load_balance_mask, i) = (void *)ptr;
6383 ptr += cpumask_size();
6385 #endif /* CONFIG_CPUMASK_OFFSTACK */
6389 init_defrootdomain();
6392 init_rt_bandwidth(&def_rt_bandwidth,
6393 global_rt_period(), global_rt_runtime());
6395 #ifdef CONFIG_RT_GROUP_SCHED
6396 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6397 global_rt_period(), global_rt_runtime());
6398 #endif /* CONFIG_RT_GROUP_SCHED */
6400 #ifdef CONFIG_CGROUP_SCHED
6401 list_add(&root_task_group.list, &task_groups);
6402 INIT_LIST_HEAD(&root_task_group.children);
6403 INIT_LIST_HEAD(&root_task_group.siblings);
6404 autogroup_init(&init_task);
6406 #endif /* CONFIG_CGROUP_SCHED */
6408 for_each_possible_cpu(i) {
6412 raw_spin_lock_init(&rq->lock);
6414 rq->calc_load_active = 0;
6415 rq->calc_load_update = jiffies + LOAD_FREQ;
6416 init_cfs_rq(&rq->cfs);
6417 init_rt_rq(&rq->rt, rq);
6418 #ifdef CONFIG_FAIR_GROUP_SCHED
6419 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6420 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6422 * How much cpu bandwidth does root_task_group get?
6424 * In case of task-groups formed thr' the cgroup filesystem, it
6425 * gets 100% of the cpu resources in the system. This overall
6426 * system cpu resource is divided among the tasks of
6427 * root_task_group and its child task-groups in a fair manner,
6428 * based on each entity's (task or task-group's) weight
6429 * (se->load.weight).
6431 * In other words, if root_task_group has 10 tasks of weight
6432 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6433 * then A0's share of the cpu resource is:
6435 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6437 * We achieve this by letting root_task_group's tasks sit
6438 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6440 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6441 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6442 #endif /* CONFIG_FAIR_GROUP_SCHED */
6444 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6445 #ifdef CONFIG_RT_GROUP_SCHED
6446 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6447 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6450 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6451 rq->cpu_load[j] = 0;
6453 rq->last_load_update_tick = jiffies;
6458 rq->cpu_power = SCHED_POWER_SCALE;
6459 rq->post_schedule = 0;
6460 rq->active_balance = 0;
6461 rq->next_balance = jiffies;
6466 rq->avg_idle = 2*sysctl_sched_migration_cost;
6468 INIT_LIST_HEAD(&rq->cfs_tasks);
6470 rq_attach_root(rq, &def_root_domain);
6471 #ifdef CONFIG_NO_HZ_COMMON
6474 #ifdef CONFIG_NO_HZ_FULL
6475 rq->last_sched_tick = 0;
6479 atomic_set(&rq->nr_iowait, 0);
6482 set_load_weight(&init_task);
6484 #ifdef CONFIG_PREEMPT_NOTIFIERS
6485 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6488 #ifdef CONFIG_RT_MUTEXES
6489 plist_head_init(&init_task.pi_waiters);
6493 * The boot idle thread does lazy MMU switching as well:
6495 atomic_inc(&init_mm.mm_count);
6496 enter_lazy_tlb(&init_mm, current);
6499 * Make us the idle thread. Technically, schedule() should not be
6500 * called from this thread, however somewhere below it might be,
6501 * but because we are the idle thread, we just pick up running again
6502 * when this runqueue becomes "idle".
6504 init_idle(current, smp_processor_id());
6506 calc_load_update = jiffies + LOAD_FREQ;
6509 * During early bootup we pretend to be a normal task:
6511 current->sched_class = &fair_sched_class;
6514 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6515 /* May be allocated at isolcpus cmdline parse time */
6516 if (cpu_isolated_map == NULL)
6517 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6518 idle_thread_set_boot_cpu();
6520 init_sched_fair_class();
6522 scheduler_running = 1;
6525 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6526 static inline int preempt_count_equals(int preempt_offset)
6528 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6530 return (nested == preempt_offset);
6533 void __might_sleep(const char *file, int line, int preempt_offset)
6535 static unsigned long prev_jiffy; /* ratelimiting */
6537 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6538 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6539 system_state != SYSTEM_RUNNING || oops_in_progress)
6541 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6543 prev_jiffy = jiffies;
6546 "BUG: sleeping function called from invalid context at %s:%d\n",
6549 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6550 in_atomic(), irqs_disabled(),
6551 current->pid, current->comm);
6553 debug_show_held_locks(current);
6554 if (irqs_disabled())
6555 print_irqtrace_events(current);
6558 EXPORT_SYMBOL(__might_sleep);
6561 #ifdef CONFIG_MAGIC_SYSRQ
6562 static void normalize_task(struct rq *rq, struct task_struct *p)
6564 const struct sched_class *prev_class = p->sched_class;
6565 int old_prio = p->prio;
6570 dequeue_task(rq, p, 0);
6571 __setscheduler(rq, p, SCHED_NORMAL, 0);
6573 enqueue_task(rq, p, 0);
6574 resched_task(rq->curr);
6577 check_class_changed(rq, p, prev_class, old_prio);
6580 void normalize_rt_tasks(void)
6582 struct task_struct *g, *p;
6583 unsigned long flags;
6586 read_lock_irqsave(&tasklist_lock, flags);
6587 do_each_thread(g, p) {
6589 * Only normalize user tasks:
6594 p->se.exec_start = 0;
6595 #ifdef CONFIG_SCHEDSTATS
6596 p->se.statistics.wait_start = 0;
6597 p->se.statistics.sleep_start = 0;
6598 p->se.statistics.block_start = 0;
6603 * Renice negative nice level userspace
6606 if (TASK_NICE(p) < 0 && p->mm)
6607 set_user_nice(p, 0);
6611 raw_spin_lock(&p->pi_lock);
6612 rq = __task_rq_lock(p);
6614 normalize_task(rq, p);
6616 __task_rq_unlock(rq);
6617 raw_spin_unlock(&p->pi_lock);
6618 } while_each_thread(g, p);
6620 read_unlock_irqrestore(&tasklist_lock, flags);
6623 #endif /* CONFIG_MAGIC_SYSRQ */
6625 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6627 * These functions are only useful for the IA64 MCA handling, or kdb.
6629 * They can only be called when the whole system has been
6630 * stopped - every CPU needs to be quiescent, and no scheduling
6631 * activity can take place. Using them for anything else would
6632 * be a serious bug, and as a result, they aren't even visible
6633 * under any other configuration.
6637 * curr_task - return the current task for a given cpu.
6638 * @cpu: the processor in question.
6640 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6642 struct task_struct *curr_task(int cpu)
6644 return cpu_curr(cpu);
6647 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6651 * set_curr_task - set the current task for a given cpu.
6652 * @cpu: the processor in question.
6653 * @p: the task pointer to set.
6655 * Description: This function must only be used when non-maskable interrupts
6656 * are serviced on a separate stack. It allows the architecture to switch the
6657 * notion of the current task on a cpu in a non-blocking manner. This function
6658 * must be called with all CPU's synchronized, and interrupts disabled, the
6659 * and caller must save the original value of the current task (see
6660 * curr_task() above) and restore that value before reenabling interrupts and
6661 * re-starting the system.
6663 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6665 void set_curr_task(int cpu, struct task_struct *p)
6672 #ifdef CONFIG_CGROUP_SCHED
6673 /* task_group_lock serializes the addition/removal of task groups */
6674 static DEFINE_SPINLOCK(task_group_lock);
6676 static void free_sched_group(struct task_group *tg)
6678 free_fair_sched_group(tg);
6679 free_rt_sched_group(tg);
6684 /* allocate runqueue etc for a new task group */
6685 struct task_group *sched_create_group(struct task_group *parent)
6687 struct task_group *tg;
6689 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6691 return ERR_PTR(-ENOMEM);
6693 if (!alloc_fair_sched_group(tg, parent))
6696 if (!alloc_rt_sched_group(tg, parent))
6702 free_sched_group(tg);
6703 return ERR_PTR(-ENOMEM);
6706 void sched_online_group(struct task_group *tg, struct task_group *parent)
6708 unsigned long flags;
6710 spin_lock_irqsave(&task_group_lock, flags);
6711 list_add_rcu(&tg->list, &task_groups);
6713 WARN_ON(!parent); /* root should already exist */
6715 tg->parent = parent;
6716 INIT_LIST_HEAD(&tg->children);
6717 list_add_rcu(&tg->siblings, &parent->children);
6718 spin_unlock_irqrestore(&task_group_lock, flags);
6721 /* rcu callback to free various structures associated with a task group */
6722 static void free_sched_group_rcu(struct rcu_head *rhp)
6724 /* now it should be safe to free those cfs_rqs */
6725 free_sched_group(container_of(rhp, struct task_group, rcu));
6728 /* Destroy runqueue etc associated with a task group */
6729 void sched_destroy_group(struct task_group *tg)
6731 /* wait for possible concurrent references to cfs_rqs complete */
6732 call_rcu(&tg->rcu, free_sched_group_rcu);
6735 void sched_offline_group(struct task_group *tg)
6737 unsigned long flags;
6740 /* end participation in shares distribution */
6741 for_each_possible_cpu(i)
6742 unregister_fair_sched_group(tg, i);
6744 spin_lock_irqsave(&task_group_lock, flags);
6745 list_del_rcu(&tg->list);
6746 list_del_rcu(&tg->siblings);
6747 spin_unlock_irqrestore(&task_group_lock, flags);
6750 /* change task's runqueue when it moves between groups.
6751 * The caller of this function should have put the task in its new group
6752 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6753 * reflect its new group.
6755 void sched_move_task(struct task_struct *tsk)
6757 struct task_group *tg;
6759 unsigned long flags;
6762 rq = task_rq_lock(tsk, &flags);
6764 running = task_current(rq, tsk);
6768 dequeue_task(rq, tsk, 0);
6769 if (unlikely(running))
6770 tsk->sched_class->put_prev_task(rq, tsk);
6772 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
6773 lockdep_is_held(&tsk->sighand->siglock)),
6774 struct task_group, css);
6775 tg = autogroup_task_group(tsk, tg);
6776 tsk->sched_task_group = tg;
6778 #ifdef CONFIG_FAIR_GROUP_SCHED
6779 if (tsk->sched_class->task_move_group)
6780 tsk->sched_class->task_move_group(tsk, on_rq);
6783 set_task_rq(tsk, task_cpu(tsk));
6785 if (unlikely(running))
6786 tsk->sched_class->set_curr_task(rq);
6788 enqueue_task(rq, tsk, 0);
6790 task_rq_unlock(rq, tsk, &flags);
6792 #endif /* CONFIG_CGROUP_SCHED */
6794 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6795 static unsigned long to_ratio(u64 period, u64 runtime)
6797 if (runtime == RUNTIME_INF)
6800 return div64_u64(runtime << 20, period);
6804 #ifdef CONFIG_RT_GROUP_SCHED
6806 * Ensure that the real time constraints are schedulable.
6808 static DEFINE_MUTEX(rt_constraints_mutex);
6810 /* Must be called with tasklist_lock held */
6811 static inline int tg_has_rt_tasks(struct task_group *tg)
6813 struct task_struct *g, *p;
6815 do_each_thread(g, p) {
6816 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6818 } while_each_thread(g, p);
6823 struct rt_schedulable_data {
6824 struct task_group *tg;
6829 static int tg_rt_schedulable(struct task_group *tg, void *data)
6831 struct rt_schedulable_data *d = data;
6832 struct task_group *child;
6833 unsigned long total, sum = 0;
6834 u64 period, runtime;
6836 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6837 runtime = tg->rt_bandwidth.rt_runtime;
6840 period = d->rt_period;
6841 runtime = d->rt_runtime;
6845 * Cannot have more runtime than the period.
6847 if (runtime > period && runtime != RUNTIME_INF)
6851 * Ensure we don't starve existing RT tasks.
6853 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6856 total = to_ratio(period, runtime);
6859 * Nobody can have more than the global setting allows.
6861 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6865 * The sum of our children's runtime should not exceed our own.
6867 list_for_each_entry_rcu(child, &tg->children, siblings) {
6868 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6869 runtime = child->rt_bandwidth.rt_runtime;
6871 if (child == d->tg) {
6872 period = d->rt_period;
6873 runtime = d->rt_runtime;
6876 sum += to_ratio(period, runtime);
6885 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6889 struct rt_schedulable_data data = {
6891 .rt_period = period,
6892 .rt_runtime = runtime,
6896 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6902 static int tg_set_rt_bandwidth(struct task_group *tg,
6903 u64 rt_period, u64 rt_runtime)
6907 mutex_lock(&rt_constraints_mutex);
6908 read_lock(&tasklist_lock);
6909 err = __rt_schedulable(tg, rt_period, rt_runtime);
6913 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6914 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6915 tg->rt_bandwidth.rt_runtime = rt_runtime;
6917 for_each_possible_cpu(i) {
6918 struct rt_rq *rt_rq = tg->rt_rq[i];
6920 raw_spin_lock(&rt_rq->rt_runtime_lock);
6921 rt_rq->rt_runtime = rt_runtime;
6922 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6924 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6926 read_unlock(&tasklist_lock);
6927 mutex_unlock(&rt_constraints_mutex);
6932 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6934 u64 rt_runtime, rt_period;
6936 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6937 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6938 if (rt_runtime_us < 0)
6939 rt_runtime = RUNTIME_INF;
6941 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6944 static long sched_group_rt_runtime(struct task_group *tg)
6948 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6951 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6952 do_div(rt_runtime_us, NSEC_PER_USEC);
6953 return rt_runtime_us;
6956 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6958 u64 rt_runtime, rt_period;
6960 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6961 rt_runtime = tg->rt_bandwidth.rt_runtime;
6966 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6969 static long sched_group_rt_period(struct task_group *tg)
6973 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6974 do_div(rt_period_us, NSEC_PER_USEC);
6975 return rt_period_us;
6978 static int sched_rt_global_constraints(void)
6980 u64 runtime, period;
6983 if (sysctl_sched_rt_period <= 0)
6986 runtime = global_rt_runtime();
6987 period = global_rt_period();
6990 * Sanity check on the sysctl variables.
6992 if (runtime > period && runtime != RUNTIME_INF)
6995 mutex_lock(&rt_constraints_mutex);
6996 read_lock(&tasklist_lock);
6997 ret = __rt_schedulable(NULL, 0, 0);
6998 read_unlock(&tasklist_lock);
6999 mutex_unlock(&rt_constraints_mutex);
7004 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7006 /* Don't accept realtime tasks when there is no way for them to run */
7007 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7013 #else /* !CONFIG_RT_GROUP_SCHED */
7014 static int sched_rt_global_constraints(void)
7016 unsigned long flags;
7019 if (sysctl_sched_rt_period <= 0)
7023 * There's always some RT tasks in the root group
7024 * -- migration, kstopmachine etc..
7026 if (sysctl_sched_rt_runtime == 0)
7029 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7030 for_each_possible_cpu(i) {
7031 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7033 raw_spin_lock(&rt_rq->rt_runtime_lock);
7034 rt_rq->rt_runtime = global_rt_runtime();
7035 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7037 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7041 #endif /* CONFIG_RT_GROUP_SCHED */
7043 int sched_rr_handler(struct ctl_table *table, int write,
7044 void __user *buffer, size_t *lenp,
7048 static DEFINE_MUTEX(mutex);
7051 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7052 /* make sure that internally we keep jiffies */
7053 /* also, writing zero resets timeslice to default */
7054 if (!ret && write) {
7055 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7056 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7058 mutex_unlock(&mutex);
7062 int sched_rt_handler(struct ctl_table *table, int write,
7063 void __user *buffer, size_t *lenp,
7067 int old_period, old_runtime;
7068 static DEFINE_MUTEX(mutex);
7071 old_period = sysctl_sched_rt_period;
7072 old_runtime = sysctl_sched_rt_runtime;
7074 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7076 if (!ret && write) {
7077 ret = sched_rt_global_constraints();
7079 sysctl_sched_rt_period = old_period;
7080 sysctl_sched_rt_runtime = old_runtime;
7082 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7083 def_rt_bandwidth.rt_period =
7084 ns_to_ktime(global_rt_period());
7087 mutex_unlock(&mutex);
7092 #ifdef CONFIG_CGROUP_SCHED
7094 /* return corresponding task_group object of a cgroup */
7095 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7097 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7098 struct task_group, css);
7101 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7103 struct task_group *tg, *parent;
7105 if (!cgrp->parent) {
7106 /* This is early initialization for the top cgroup */
7107 return &root_task_group.css;
7110 parent = cgroup_tg(cgrp->parent);
7111 tg = sched_create_group(parent);
7113 return ERR_PTR(-ENOMEM);
7118 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7120 struct task_group *tg = cgroup_tg(cgrp);
7121 struct task_group *parent;
7126 parent = cgroup_tg(cgrp->parent);
7127 sched_online_group(tg, parent);
7131 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7133 struct task_group *tg = cgroup_tg(cgrp);
7135 sched_destroy_group(tg);
7138 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7140 struct task_group *tg = cgroup_tg(cgrp);
7142 sched_offline_group(tg);
7145 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7146 struct cgroup_taskset *tset)
7148 struct task_struct *task;
7150 cgroup_taskset_for_each(task, cgrp, tset) {
7151 #ifdef CONFIG_RT_GROUP_SCHED
7152 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7155 /* We don't support RT-tasks being in separate groups */
7156 if (task->sched_class != &fair_sched_class)
7163 static void cpu_cgroup_attach(struct cgroup *cgrp,
7164 struct cgroup_taskset *tset)
7166 struct task_struct *task;
7168 cgroup_taskset_for_each(task, cgrp, tset)
7169 sched_move_task(task);
7173 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7174 struct task_struct *task)
7177 * cgroup_exit() is called in the copy_process() failure path.
7178 * Ignore this case since the task hasn't ran yet, this avoids
7179 * trying to poke a half freed task state from generic code.
7181 if (!(task->flags & PF_EXITING))
7184 sched_move_task(task);
7187 #ifdef CONFIG_FAIR_GROUP_SCHED
7188 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7191 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7194 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7196 struct task_group *tg = cgroup_tg(cgrp);
7198 return (u64) scale_load_down(tg->shares);
7201 #ifdef CONFIG_CFS_BANDWIDTH
7202 static DEFINE_MUTEX(cfs_constraints_mutex);
7204 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7205 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7207 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7209 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7211 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7212 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7214 if (tg == &root_task_group)
7218 * Ensure we have at some amount of bandwidth every period. This is
7219 * to prevent reaching a state of large arrears when throttled via
7220 * entity_tick() resulting in prolonged exit starvation.
7222 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7226 * Likewise, bound things on the otherside by preventing insane quota
7227 * periods. This also allows us to normalize in computing quota
7230 if (period > max_cfs_quota_period)
7233 mutex_lock(&cfs_constraints_mutex);
7234 ret = __cfs_schedulable(tg, period, quota);
7238 runtime_enabled = quota != RUNTIME_INF;
7239 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7240 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7241 raw_spin_lock_irq(&cfs_b->lock);
7242 cfs_b->period = ns_to_ktime(period);
7243 cfs_b->quota = quota;
7245 __refill_cfs_bandwidth_runtime(cfs_b);
7246 /* restart the period timer (if active) to handle new period expiry */
7247 if (runtime_enabled && cfs_b->timer_active) {
7248 /* force a reprogram */
7249 cfs_b->timer_active = 0;
7250 __start_cfs_bandwidth(cfs_b);
7252 raw_spin_unlock_irq(&cfs_b->lock);
7254 for_each_possible_cpu(i) {
7255 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7256 struct rq *rq = cfs_rq->rq;
7258 raw_spin_lock_irq(&rq->lock);
7259 cfs_rq->runtime_enabled = runtime_enabled;
7260 cfs_rq->runtime_remaining = 0;
7262 if (cfs_rq->throttled)
7263 unthrottle_cfs_rq(cfs_rq);
7264 raw_spin_unlock_irq(&rq->lock);
7267 mutex_unlock(&cfs_constraints_mutex);
7272 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7276 period = ktime_to_ns(tg->cfs_bandwidth.period);
7277 if (cfs_quota_us < 0)
7278 quota = RUNTIME_INF;
7280 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7282 return tg_set_cfs_bandwidth(tg, period, quota);
7285 long tg_get_cfs_quota(struct task_group *tg)
7289 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7292 quota_us = tg->cfs_bandwidth.quota;
7293 do_div(quota_us, NSEC_PER_USEC);
7298 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7302 period = (u64)cfs_period_us * NSEC_PER_USEC;
7303 quota = tg->cfs_bandwidth.quota;
7305 return tg_set_cfs_bandwidth(tg, period, quota);
7308 long tg_get_cfs_period(struct task_group *tg)
7312 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7313 do_div(cfs_period_us, NSEC_PER_USEC);
7315 return cfs_period_us;
7318 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7320 return tg_get_cfs_quota(cgroup_tg(cgrp));
7323 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7326 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7329 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7331 return tg_get_cfs_period(cgroup_tg(cgrp));
7334 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7337 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7340 struct cfs_schedulable_data {
7341 struct task_group *tg;
7346 * normalize group quota/period to be quota/max_period
7347 * note: units are usecs
7349 static u64 normalize_cfs_quota(struct task_group *tg,
7350 struct cfs_schedulable_data *d)
7358 period = tg_get_cfs_period(tg);
7359 quota = tg_get_cfs_quota(tg);
7362 /* note: these should typically be equivalent */
7363 if (quota == RUNTIME_INF || quota == -1)
7366 return to_ratio(period, quota);
7369 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7371 struct cfs_schedulable_data *d = data;
7372 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7373 s64 quota = 0, parent_quota = -1;
7376 quota = RUNTIME_INF;
7378 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7380 quota = normalize_cfs_quota(tg, d);
7381 parent_quota = parent_b->hierarchal_quota;
7384 * ensure max(child_quota) <= parent_quota, inherit when no
7387 if (quota == RUNTIME_INF)
7388 quota = parent_quota;
7389 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7392 cfs_b->hierarchal_quota = quota;
7397 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7400 struct cfs_schedulable_data data = {
7406 if (quota != RUNTIME_INF) {
7407 do_div(data.period, NSEC_PER_USEC);
7408 do_div(data.quota, NSEC_PER_USEC);
7412 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7418 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7419 struct cgroup_map_cb *cb)
7421 struct task_group *tg = cgroup_tg(cgrp);
7422 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7424 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7425 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7426 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7430 #endif /* CONFIG_CFS_BANDWIDTH */
7431 #endif /* CONFIG_FAIR_GROUP_SCHED */
7433 #ifdef CONFIG_RT_GROUP_SCHED
7434 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7437 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7440 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7442 return sched_group_rt_runtime(cgroup_tg(cgrp));
7445 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7448 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7451 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7453 return sched_group_rt_period(cgroup_tg(cgrp));
7455 #endif /* CONFIG_RT_GROUP_SCHED */
7457 static struct cftype cpu_files[] = {
7458 #ifdef CONFIG_FAIR_GROUP_SCHED
7461 .read_u64 = cpu_shares_read_u64,
7462 .write_u64 = cpu_shares_write_u64,
7465 #ifdef CONFIG_CFS_BANDWIDTH
7467 .name = "cfs_quota_us",
7468 .read_s64 = cpu_cfs_quota_read_s64,
7469 .write_s64 = cpu_cfs_quota_write_s64,
7472 .name = "cfs_period_us",
7473 .read_u64 = cpu_cfs_period_read_u64,
7474 .write_u64 = cpu_cfs_period_write_u64,
7478 .read_map = cpu_stats_show,
7481 #ifdef CONFIG_RT_GROUP_SCHED
7483 .name = "rt_runtime_us",
7484 .read_s64 = cpu_rt_runtime_read,
7485 .write_s64 = cpu_rt_runtime_write,
7488 .name = "rt_period_us",
7489 .read_u64 = cpu_rt_period_read_uint,
7490 .write_u64 = cpu_rt_period_write_uint,
7496 struct cgroup_subsys cpu_cgroup_subsys = {
7498 .css_alloc = cpu_cgroup_css_alloc,
7499 .css_free = cpu_cgroup_css_free,
7500 .css_online = cpu_cgroup_css_online,
7501 .css_offline = cpu_cgroup_css_offline,
7502 .can_attach = cpu_cgroup_can_attach,
7503 .attach = cpu_cgroup_attach,
7504 .exit = cpu_cgroup_exit,
7505 .subsys_id = cpu_cgroup_subsys_id,
7506 .base_cftypes = cpu_files,
7510 #endif /* CONFIG_CGROUP_SCHED */
7512 void dump_cpu_task(int cpu)
7514 pr_info("Task dump for CPU %d:\n", cpu);
7515 sched_show_task(cpu_curr(cpu));