4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
375 static void hrtick_clear(struct rq *rq)
377 if (hrtimer_active(&rq->hrtick_timer))
378 hrtimer_cancel(&rq->hrtick_timer);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart hrtick(struct hrtimer *timer)
387 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
391 raw_spin_lock(&rq->lock);
393 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
394 raw_spin_unlock(&rq->lock);
396 return HRTIMER_NORESTART;
401 static int __hrtick_restart(struct rq *rq)
403 struct hrtimer *timer = &rq->hrtick_timer;
404 ktime_t time = hrtimer_get_softexpires(timer);
406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg)
416 raw_spin_lock(&rq->lock);
417 __hrtick_restart(rq);
418 rq->hrtick_csd_pending = 0;
419 raw_spin_unlock(&rq->lock);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq *rq, u64 delay)
429 struct hrtimer *timer = &rq->hrtick_timer;
430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
432 hrtimer_set_expires(timer, time);
434 if (rq == this_rq()) {
435 __hrtick_restart(rq);
436 } else if (!rq->hrtick_csd_pending) {
437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
438 rq->hrtick_csd_pending = 1;
443 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
445 int cpu = (int)(long)hcpu;
448 case CPU_UP_CANCELED:
449 case CPU_UP_CANCELED_FROZEN:
450 case CPU_DOWN_PREPARE:
451 case CPU_DOWN_PREPARE_FROZEN:
453 case CPU_DEAD_FROZEN:
454 hrtick_clear(cpu_rq(cpu));
461 static __init void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq *rq, u64 delay)
473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
474 HRTIMER_MODE_REL_PINNED, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq *rq)
485 rq->hrtick_csd_pending = 0;
487 rq->hrtick_csd.flags = 0;
488 rq->hrtick_csd.func = __hrtick_start;
489 rq->hrtick_csd.info = rq;
492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
493 rq->hrtick_timer.function = hrtick;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq *rq)
500 static inline void init_rq_hrtick(struct rq *rq)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
517 void resched_task(struct task_struct *p)
521 assert_raw_spin_locked(&task_rq(p)->lock);
523 if (test_tsk_need_resched(p))
526 set_tsk_need_resched(p);
529 if (cpu == smp_processor_id())
532 /* NEED_RESCHED must be visible before we test polling */
534 if (!tsk_is_polling(p))
535 smp_send_reschedule(cpu);
538 void resched_cpu(int cpu)
540 struct rq *rq = cpu_rq(cpu);
543 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
545 resched_task(cpu_curr(cpu));
546 raw_spin_unlock_irqrestore(&rq->lock, flags);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu = smp_processor_id();
562 struct sched_domain *sd;
565 for_each_domain(cpu, sd) {
566 for_each_cpu(i, sched_domain_span(sd)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu)
589 struct rq *rq = cpu_rq(cpu);
591 if (cpu == smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq->curr != rq->idle)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq->idle);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq->idle))
614 smp_send_reschedule(cpu);
617 static bool wake_up_full_nohz_cpu(int cpu)
619 if (tick_nohz_full_cpu(cpu)) {
620 if (cpu != smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu);
629 void wake_up_nohz_cpu(int cpu)
631 if (!wake_up_full_nohz_cpu(cpu))
632 wake_up_idle_cpu(cpu);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu = smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
642 if (idle_cpu(cpu) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq->nr_running > 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq *rq)
682 s64 period = sched_avg_period();
684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq->age_stamp));
691 rq->age_stamp += period;
696 #else /* !CONFIG_SMP */
697 void resched_task(struct task_struct *p)
699 assert_raw_spin_locked(&task_rq(p)->lock);
700 set_tsk_need_resched(p);
702 #endif /* CONFIG_SMP */
704 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
707 * Iterate task_group tree rooted at *from, calling @down when first entering a
708 * node and @up when leaving it for the final time.
710 * Caller must hold rcu_lock or sufficient equivalent.
712 int walk_tg_tree_from(struct task_group *from,
713 tg_visitor down, tg_visitor up, void *data)
715 struct task_group *parent, *child;
721 ret = (*down)(parent, data);
724 list_for_each_entry_rcu(child, &parent->children, siblings) {
731 ret = (*up)(parent, data);
732 if (ret || parent == from)
736 parent = parent->parent;
743 int tg_nop(struct task_group *tg, void *data)
749 static void set_load_weight(struct task_struct *p)
751 int prio = p->static_prio - MAX_RT_PRIO;
752 struct load_weight *load = &p->se.load;
755 * SCHED_IDLE tasks get minimal weight:
757 if (p->policy == SCHED_IDLE) {
758 load->weight = scale_load(WEIGHT_IDLEPRIO);
759 load->inv_weight = WMULT_IDLEPRIO;
763 load->weight = scale_load(prio_to_weight[prio]);
764 load->inv_weight = prio_to_wmult[prio];
767 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
770 sched_info_queued(p);
771 p->sched_class->enqueue_task(rq, p, flags);
774 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
777 sched_info_dequeued(p);
778 p->sched_class->dequeue_task(rq, p, flags);
781 void activate_task(struct rq *rq, struct task_struct *p, int flags)
783 if (task_contributes_to_load(p))
784 rq->nr_uninterruptible--;
786 enqueue_task(rq, p, flags);
789 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
791 if (task_contributes_to_load(p))
792 rq->nr_uninterruptible++;
794 dequeue_task(rq, p, flags);
797 static void update_rq_clock_task(struct rq *rq, s64 delta)
800 * In theory, the compile should just see 0 here, and optimize out the call
801 * to sched_rt_avg_update. But I don't trust it...
803 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
804 s64 steal = 0, irq_delta = 0;
806 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
807 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
810 * Since irq_time is only updated on {soft,}irq_exit, we might run into
811 * this case when a previous update_rq_clock() happened inside a
814 * When this happens, we stop ->clock_task and only update the
815 * prev_irq_time stamp to account for the part that fit, so that a next
816 * update will consume the rest. This ensures ->clock_task is
819 * It does however cause some slight miss-attribution of {soft,}irq
820 * time, a more accurate solution would be to update the irq_time using
821 * the current rq->clock timestamp, except that would require using
824 if (irq_delta > delta)
827 rq->prev_irq_time += irq_delta;
830 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
831 if (static_key_false((¶virt_steal_rq_enabled))) {
834 steal = paravirt_steal_clock(cpu_of(rq));
835 steal -= rq->prev_steal_time_rq;
837 if (unlikely(steal > delta))
840 st = steal_ticks(steal);
841 steal = st * TICK_NSEC;
843 rq->prev_steal_time_rq += steal;
849 rq->clock_task += delta;
851 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
852 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
853 sched_rt_avg_update(rq, irq_delta + steal);
857 void sched_set_stop_task(int cpu, struct task_struct *stop)
859 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
860 struct task_struct *old_stop = cpu_rq(cpu)->stop;
864 * Make it appear like a SCHED_FIFO task, its something
865 * userspace knows about and won't get confused about.
867 * Also, it will make PI more or less work without too
868 * much confusion -- but then, stop work should not
869 * rely on PI working anyway.
871 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
873 stop->sched_class = &stop_sched_class;
876 cpu_rq(cpu)->stop = stop;
880 * Reset it back to a normal scheduling class so that
881 * it can die in pieces.
883 old_stop->sched_class = &rt_sched_class;
888 * __normal_prio - return the priority that is based on the static prio
890 static inline int __normal_prio(struct task_struct *p)
892 return p->static_prio;
896 * Calculate the expected normal priority: i.e. priority
897 * without taking RT-inheritance into account. Might be
898 * boosted by interactivity modifiers. Changes upon fork,
899 * setprio syscalls, and whenever the interactivity
900 * estimator recalculates.
902 static inline int normal_prio(struct task_struct *p)
906 if (task_has_rt_policy(p))
907 prio = MAX_RT_PRIO-1 - p->rt_priority;
909 prio = __normal_prio(p);
914 * Calculate the current priority, i.e. the priority
915 * taken into account by the scheduler. This value might
916 * be boosted by RT tasks, or might be boosted by
917 * interactivity modifiers. Will be RT if the task got
918 * RT-boosted. If not then it returns p->normal_prio.
920 static int effective_prio(struct task_struct *p)
922 p->normal_prio = normal_prio(p);
924 * If we are RT tasks or we were boosted to RT priority,
925 * keep the priority unchanged. Otherwise, update priority
926 * to the normal priority:
928 if (!rt_prio(p->prio))
929 return p->normal_prio;
934 * task_curr - is this task currently executing on a CPU?
935 * @p: the task in question.
937 inline int task_curr(const struct task_struct *p)
939 return cpu_curr(task_cpu(p)) == p;
942 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
943 const struct sched_class *prev_class,
946 if (prev_class != p->sched_class) {
947 if (prev_class->switched_from)
948 prev_class->switched_from(rq, p);
949 p->sched_class->switched_to(rq, p);
950 } else if (oldprio != p->prio)
951 p->sched_class->prio_changed(rq, p, oldprio);
954 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
956 const struct sched_class *class;
958 if (p->sched_class == rq->curr->sched_class) {
959 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
961 for_each_class(class) {
962 if (class == rq->curr->sched_class)
964 if (class == p->sched_class) {
965 resched_task(rq->curr);
972 * A queue event has occurred, and we're going to schedule. In
973 * this case, we can save a useless back to back clock update.
975 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
976 rq->skip_clock_update = 1;
979 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
981 void register_task_migration_notifier(struct notifier_block *n)
983 atomic_notifier_chain_register(&task_migration_notifier, n);
987 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
989 #ifdef CONFIG_SCHED_DEBUG
991 * We should never call set_task_cpu() on a blocked task,
992 * ttwu() will sort out the placement.
994 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
995 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
997 #ifdef CONFIG_LOCKDEP
999 * The caller should hold either p->pi_lock or rq->lock, when changing
1000 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1002 * sched_move_task() holds both and thus holding either pins the cgroup,
1005 * Furthermore, all task_rq users should acquire both locks, see
1008 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1009 lockdep_is_held(&task_rq(p)->lock)));
1013 trace_sched_migrate_task(p, new_cpu);
1015 if (task_cpu(p) != new_cpu) {
1016 struct task_migration_notifier tmn;
1018 if (p->sched_class->migrate_task_rq)
1019 p->sched_class->migrate_task_rq(p, new_cpu);
1020 p->se.nr_migrations++;
1021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1024 tmn.from_cpu = task_cpu(p);
1025 tmn.to_cpu = new_cpu;
1027 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1030 __set_task_cpu(p, new_cpu);
1033 struct migration_arg {
1034 struct task_struct *task;
1038 static int migration_cpu_stop(void *data);
1041 * wait_task_inactive - wait for a thread to unschedule.
1043 * If @match_state is nonzero, it's the @p->state value just checked and
1044 * not expected to change. If it changes, i.e. @p might have woken up,
1045 * then return zero. When we succeed in waiting for @p to be off its CPU,
1046 * we return a positive number (its total switch count). If a second call
1047 * a short while later returns the same number, the caller can be sure that
1048 * @p has remained unscheduled the whole time.
1050 * The caller must ensure that the task *will* unschedule sometime soon,
1051 * else this function might spin for a *long* time. This function can't
1052 * be called with interrupts off, or it may introduce deadlock with
1053 * smp_call_function() if an IPI is sent by the same process we are
1054 * waiting to become inactive.
1056 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1058 unsigned long flags;
1065 * We do the initial early heuristics without holding
1066 * any task-queue locks at all. We'll only try to get
1067 * the runqueue lock when things look like they will
1073 * If the task is actively running on another CPU
1074 * still, just relax and busy-wait without holding
1077 * NOTE! Since we don't hold any locks, it's not
1078 * even sure that "rq" stays as the right runqueue!
1079 * But we don't care, since "task_running()" will
1080 * return false if the runqueue has changed and p
1081 * is actually now running somewhere else!
1083 while (task_running(rq, p)) {
1084 if (match_state && unlikely(p->state != match_state))
1090 * Ok, time to look more closely! We need the rq
1091 * lock now, to be *sure*. If we're wrong, we'll
1092 * just go back and repeat.
1094 rq = task_rq_lock(p, &flags);
1095 trace_sched_wait_task(p);
1096 running = task_running(rq, p);
1099 if (!match_state || p->state == match_state)
1100 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1101 task_rq_unlock(rq, p, &flags);
1104 * If it changed from the expected state, bail out now.
1106 if (unlikely(!ncsw))
1110 * Was it really running after all now that we
1111 * checked with the proper locks actually held?
1113 * Oops. Go back and try again..
1115 if (unlikely(running)) {
1121 * It's not enough that it's not actively running,
1122 * it must be off the runqueue _entirely_, and not
1125 * So if it was still runnable (but just not actively
1126 * running right now), it's preempted, and we should
1127 * yield - it could be a while.
1129 if (unlikely(on_rq)) {
1130 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1132 set_current_state(TASK_UNINTERRUPTIBLE);
1133 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1138 * Ahh, all good. It wasn't running, and it wasn't
1139 * runnable, which means that it will never become
1140 * running in the future either. We're all done!
1149 * kick_process - kick a running thread to enter/exit the kernel
1150 * @p: the to-be-kicked thread
1152 * Cause a process which is running on another CPU to enter
1153 * kernel-mode, without any delay. (to get signals handled.)
1155 * NOTE: this function doesn't have to take the runqueue lock,
1156 * because all it wants to ensure is that the remote task enters
1157 * the kernel. If the IPI races and the task has been migrated
1158 * to another CPU then no harm is done and the purpose has been
1161 void kick_process(struct task_struct *p)
1167 if ((cpu != smp_processor_id()) && task_curr(p))
1168 smp_send_reschedule(cpu);
1171 EXPORT_SYMBOL_GPL(kick_process);
1172 #endif /* CONFIG_SMP */
1176 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1178 static int select_fallback_rq(int cpu, struct task_struct *p)
1180 int nid = cpu_to_node(cpu);
1181 const struct cpumask *nodemask = NULL;
1182 enum { cpuset, possible, fail } state = cpuset;
1186 * If the node that the cpu is on has been offlined, cpu_to_node()
1187 * will return -1. There is no cpu on the node, and we should
1188 * select the cpu on the other node.
1191 nodemask = cpumask_of_node(nid);
1193 /* Look for allowed, online CPU in same node. */
1194 for_each_cpu(dest_cpu, nodemask) {
1195 if (!cpu_online(dest_cpu))
1197 if (!cpu_active(dest_cpu))
1199 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1205 /* Any allowed, online CPU? */
1206 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1207 if (!cpu_online(dest_cpu))
1209 if (!cpu_active(dest_cpu))
1216 /* No more Mr. Nice Guy. */
1217 cpuset_cpus_allowed_fallback(p);
1222 do_set_cpus_allowed(p, cpu_possible_mask);
1233 if (state != cpuset) {
1235 * Don't tell them about moving exiting tasks or
1236 * kernel threads (both mm NULL), since they never
1239 if (p->mm && printk_ratelimit()) {
1240 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1241 task_pid_nr(p), p->comm, cpu);
1249 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1252 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1254 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1257 * In order not to call set_task_cpu() on a blocking task we need
1258 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1261 * Since this is common to all placement strategies, this lives here.
1263 * [ this allows ->select_task() to simply return task_cpu(p) and
1264 * not worry about this generic constraint ]
1266 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1268 cpu = select_fallback_rq(task_cpu(p), p);
1273 static void update_avg(u64 *avg, u64 sample)
1275 s64 diff = sample - *avg;
1281 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1283 #ifdef CONFIG_SCHEDSTATS
1284 struct rq *rq = this_rq();
1287 int this_cpu = smp_processor_id();
1289 if (cpu == this_cpu) {
1290 schedstat_inc(rq, ttwu_local);
1291 schedstat_inc(p, se.statistics.nr_wakeups_local);
1293 struct sched_domain *sd;
1295 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1297 for_each_domain(this_cpu, sd) {
1298 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1299 schedstat_inc(sd, ttwu_wake_remote);
1306 if (wake_flags & WF_MIGRATED)
1307 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1309 #endif /* CONFIG_SMP */
1311 schedstat_inc(rq, ttwu_count);
1312 schedstat_inc(p, se.statistics.nr_wakeups);
1314 if (wake_flags & WF_SYNC)
1315 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1317 #endif /* CONFIG_SCHEDSTATS */
1320 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1322 activate_task(rq, p, en_flags);
1325 /* if a worker is waking up, notify workqueue */
1326 if (p->flags & PF_WQ_WORKER)
1327 wq_worker_waking_up(p, cpu_of(rq));
1331 * Mark the task runnable and perform wakeup-preemption.
1334 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1336 check_preempt_curr(rq, p, wake_flags);
1337 trace_sched_wakeup(p, true);
1339 p->state = TASK_RUNNING;
1341 if (p->sched_class->task_woken)
1342 p->sched_class->task_woken(rq, p);
1344 if (rq->idle_stamp) {
1345 u64 delta = rq_clock(rq) - rq->idle_stamp;
1346 u64 max = 2*sysctl_sched_migration_cost;
1351 update_avg(&rq->avg_idle, delta);
1358 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1361 if (p->sched_contributes_to_load)
1362 rq->nr_uninterruptible--;
1365 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1366 ttwu_do_wakeup(rq, p, wake_flags);
1370 * Called in case the task @p isn't fully descheduled from its runqueue,
1371 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1372 * since all we need to do is flip p->state to TASK_RUNNING, since
1373 * the task is still ->on_rq.
1375 static int ttwu_remote(struct task_struct *p, int wake_flags)
1380 rq = __task_rq_lock(p);
1382 /* check_preempt_curr() may use rq clock */
1383 update_rq_clock(rq);
1384 ttwu_do_wakeup(rq, p, wake_flags);
1387 __task_rq_unlock(rq);
1393 static void sched_ttwu_pending(void)
1395 struct rq *rq = this_rq();
1396 struct llist_node *llist = llist_del_all(&rq->wake_list);
1397 struct task_struct *p;
1399 raw_spin_lock(&rq->lock);
1402 p = llist_entry(llist, struct task_struct, wake_entry);
1403 llist = llist_next(llist);
1404 ttwu_do_activate(rq, p, 0);
1407 raw_spin_unlock(&rq->lock);
1410 void scheduler_ipi(void)
1412 if (llist_empty(&this_rq()->wake_list)
1413 && !tick_nohz_full_cpu(smp_processor_id())
1414 && !got_nohz_idle_kick())
1418 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1419 * traditionally all their work was done from the interrupt return
1420 * path. Now that we actually do some work, we need to make sure
1423 * Some archs already do call them, luckily irq_enter/exit nest
1426 * Arguably we should visit all archs and update all handlers,
1427 * however a fair share of IPIs are still resched only so this would
1428 * somewhat pessimize the simple resched case.
1431 tick_nohz_full_check();
1432 sched_ttwu_pending();
1435 * Check if someone kicked us for doing the nohz idle load balance.
1437 if (unlikely(got_nohz_idle_kick())) {
1438 this_rq()->idle_balance = 1;
1439 raise_softirq_irqoff(SCHED_SOFTIRQ);
1444 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1446 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1447 smp_send_reschedule(cpu);
1450 bool cpus_share_cache(int this_cpu, int that_cpu)
1452 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1454 #endif /* CONFIG_SMP */
1456 static void ttwu_queue(struct task_struct *p, int cpu)
1458 struct rq *rq = cpu_rq(cpu);
1460 #if defined(CONFIG_SMP)
1461 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1462 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1463 ttwu_queue_remote(p, cpu);
1468 raw_spin_lock(&rq->lock);
1469 ttwu_do_activate(rq, p, 0);
1470 raw_spin_unlock(&rq->lock);
1474 * try_to_wake_up - wake up a thread
1475 * @p: the thread to be awakened
1476 * @state: the mask of task states that can be woken
1477 * @wake_flags: wake modifier flags (WF_*)
1479 * Put it on the run-queue if it's not already there. The "current"
1480 * thread is always on the run-queue (except when the actual
1481 * re-schedule is in progress), and as such you're allowed to do
1482 * the simpler "current->state = TASK_RUNNING" to mark yourself
1483 * runnable without the overhead of this.
1485 * Returns %true if @p was woken up, %false if it was already running
1486 * or @state didn't match @p's state.
1489 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1491 unsigned long flags;
1492 int cpu, success = 0;
1495 raw_spin_lock_irqsave(&p->pi_lock, flags);
1496 if (!(p->state & state))
1499 success = 1; /* we're going to change ->state */
1502 if (p->on_rq && ttwu_remote(p, wake_flags))
1507 * If the owning (remote) cpu is still in the middle of schedule() with
1508 * this task as prev, wait until its done referencing the task.
1513 * Pairs with the smp_wmb() in finish_lock_switch().
1517 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1518 p->state = TASK_WAKING;
1520 if (p->sched_class->task_waking)
1521 p->sched_class->task_waking(p);
1523 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1524 if (task_cpu(p) != cpu) {
1525 wake_flags |= WF_MIGRATED;
1526 set_task_cpu(p, cpu);
1528 #endif /* CONFIG_SMP */
1532 ttwu_stat(p, cpu, wake_flags);
1534 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1540 * try_to_wake_up_local - try to wake up a local task with rq lock held
1541 * @p: the thread to be awakened
1543 * Put @p on the run-queue if it's not already there. The caller must
1544 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1547 static void try_to_wake_up_local(struct task_struct *p)
1549 struct rq *rq = task_rq(p);
1551 if (WARN_ON_ONCE(rq != this_rq()) ||
1552 WARN_ON_ONCE(p == current))
1555 lockdep_assert_held(&rq->lock);
1557 if (!raw_spin_trylock(&p->pi_lock)) {
1558 raw_spin_unlock(&rq->lock);
1559 raw_spin_lock(&p->pi_lock);
1560 raw_spin_lock(&rq->lock);
1563 if (!(p->state & TASK_NORMAL))
1567 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1569 ttwu_do_wakeup(rq, p, 0);
1570 ttwu_stat(p, smp_processor_id(), 0);
1572 raw_spin_unlock(&p->pi_lock);
1576 * wake_up_process - Wake up a specific process
1577 * @p: The process to be woken up.
1579 * Attempt to wake up the nominated process and move it to the set of runnable
1580 * processes. Returns 1 if the process was woken up, 0 if it was already
1583 * It may be assumed that this function implies a write memory barrier before
1584 * changing the task state if and only if any tasks are woken up.
1586 int wake_up_process(struct task_struct *p)
1588 WARN_ON(task_is_stopped_or_traced(p));
1589 return try_to_wake_up(p, TASK_NORMAL, 0);
1591 EXPORT_SYMBOL(wake_up_process);
1593 int wake_up_state(struct task_struct *p, unsigned int state)
1595 return try_to_wake_up(p, state, 0);
1599 * Perform scheduler related setup for a newly forked process p.
1600 * p is forked by current.
1602 * __sched_fork() is basic setup used by init_idle() too:
1604 static void __sched_fork(struct task_struct *p)
1609 p->se.exec_start = 0;
1610 p->se.sum_exec_runtime = 0;
1611 p->se.prev_sum_exec_runtime = 0;
1612 p->se.nr_migrations = 0;
1614 INIT_LIST_HEAD(&p->se.group_node);
1616 #ifdef CONFIG_SCHEDSTATS
1617 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1620 INIT_LIST_HEAD(&p->rt.run_list);
1622 #ifdef CONFIG_PREEMPT_NOTIFIERS
1623 INIT_HLIST_HEAD(&p->preempt_notifiers);
1626 #ifdef CONFIG_NUMA_BALANCING
1627 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1628 p->mm->numa_next_scan = jiffies;
1629 p->mm->numa_next_reset = jiffies;
1630 p->mm->numa_scan_seq = 0;
1633 p->node_stamp = 0ULL;
1634 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1635 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1636 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1637 p->numa_work.next = &p->numa_work;
1638 #endif /* CONFIG_NUMA_BALANCING */
1641 #ifdef CONFIG_NUMA_BALANCING
1642 #ifdef CONFIG_SCHED_DEBUG
1643 void set_numabalancing_state(bool enabled)
1646 sched_feat_set("NUMA");
1648 sched_feat_set("NO_NUMA");
1651 __read_mostly bool numabalancing_enabled;
1653 void set_numabalancing_state(bool enabled)
1655 numabalancing_enabled = enabled;
1657 #endif /* CONFIG_SCHED_DEBUG */
1658 #endif /* CONFIG_NUMA_BALANCING */
1661 * fork()/clone()-time setup:
1663 void sched_fork(struct task_struct *p)
1665 unsigned long flags;
1666 int cpu = get_cpu();
1670 * We mark the process as running here. This guarantees that
1671 * nobody will actually run it, and a signal or other external
1672 * event cannot wake it up and insert it on the runqueue either.
1674 p->state = TASK_RUNNING;
1677 * Make sure we do not leak PI boosting priority to the child.
1679 p->prio = current->normal_prio;
1682 * Revert to default priority/policy on fork if requested.
1684 if (unlikely(p->sched_reset_on_fork)) {
1685 if (task_has_rt_policy(p)) {
1686 p->policy = SCHED_NORMAL;
1687 p->static_prio = NICE_TO_PRIO(0);
1689 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1690 p->static_prio = NICE_TO_PRIO(0);
1692 p->prio = p->normal_prio = __normal_prio(p);
1696 * We don't need the reset flag anymore after the fork. It has
1697 * fulfilled its duty:
1699 p->sched_reset_on_fork = 0;
1702 if (!rt_prio(p->prio))
1703 p->sched_class = &fair_sched_class;
1705 if (p->sched_class->task_fork)
1706 p->sched_class->task_fork(p);
1709 * The child is not yet in the pid-hash so no cgroup attach races,
1710 * and the cgroup is pinned to this child due to cgroup_fork()
1711 * is ran before sched_fork().
1713 * Silence PROVE_RCU.
1715 raw_spin_lock_irqsave(&p->pi_lock, flags);
1716 set_task_cpu(p, cpu);
1717 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1719 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1720 if (likely(sched_info_on()))
1721 memset(&p->sched_info, 0, sizeof(p->sched_info));
1723 #if defined(CONFIG_SMP)
1726 #ifdef CONFIG_PREEMPT_COUNT
1727 /* Want to start with kernel preemption disabled. */
1728 task_thread_info(p)->preempt_count = 1;
1731 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1738 * wake_up_new_task - wake up a newly created task for the first time.
1740 * This function will do some initial scheduler statistics housekeeping
1741 * that must be done for every newly created context, then puts the task
1742 * on the runqueue and wakes it.
1744 void wake_up_new_task(struct task_struct *p)
1746 unsigned long flags;
1749 raw_spin_lock_irqsave(&p->pi_lock, flags);
1752 * Fork balancing, do it here and not earlier because:
1753 * - cpus_allowed can change in the fork path
1754 * - any previously selected cpu might disappear through hotplug
1756 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1759 /* Initialize new task's runnable average */
1760 init_task_runnable_average(p);
1761 rq = __task_rq_lock(p);
1762 activate_task(rq, p, 0);
1764 trace_sched_wakeup_new(p, true);
1765 check_preempt_curr(rq, p, WF_FORK);
1767 if (p->sched_class->task_woken)
1768 p->sched_class->task_woken(rq, p);
1770 task_rq_unlock(rq, p, &flags);
1773 #ifdef CONFIG_PREEMPT_NOTIFIERS
1776 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1777 * @notifier: notifier struct to register
1779 void preempt_notifier_register(struct preempt_notifier *notifier)
1781 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1783 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1786 * preempt_notifier_unregister - no longer interested in preemption notifications
1787 * @notifier: notifier struct to unregister
1789 * This is safe to call from within a preemption notifier.
1791 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1793 hlist_del(¬ifier->link);
1795 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1797 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1799 struct preempt_notifier *notifier;
1801 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1802 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1806 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1807 struct task_struct *next)
1809 struct preempt_notifier *notifier;
1811 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1812 notifier->ops->sched_out(notifier, next);
1815 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1817 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1822 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1823 struct task_struct *next)
1827 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1830 * prepare_task_switch - prepare to switch tasks
1831 * @rq: the runqueue preparing to switch
1832 * @prev: the current task that is being switched out
1833 * @next: the task we are going to switch to.
1835 * This is called with the rq lock held and interrupts off. It must
1836 * be paired with a subsequent finish_task_switch after the context
1839 * prepare_task_switch sets up locking and calls architecture specific
1843 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1844 struct task_struct *next)
1846 trace_sched_switch(prev, next);
1847 sched_info_switch(prev, next);
1848 perf_event_task_sched_out(prev, next);
1849 fire_sched_out_preempt_notifiers(prev, next);
1850 prepare_lock_switch(rq, next);
1851 prepare_arch_switch(next);
1855 * finish_task_switch - clean up after a task-switch
1856 * @rq: runqueue associated with task-switch
1857 * @prev: the thread we just switched away from.
1859 * finish_task_switch must be called after the context switch, paired
1860 * with a prepare_task_switch call before the context switch.
1861 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1862 * and do any other architecture-specific cleanup actions.
1864 * Note that we may have delayed dropping an mm in context_switch(). If
1865 * so, we finish that here outside of the runqueue lock. (Doing it
1866 * with the lock held can cause deadlocks; see schedule() for
1869 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1870 __releases(rq->lock)
1872 struct mm_struct *mm = rq->prev_mm;
1878 * A task struct has one reference for the use as "current".
1879 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1880 * schedule one last time. The schedule call will never return, and
1881 * the scheduled task must drop that reference.
1882 * The test for TASK_DEAD must occur while the runqueue locks are
1883 * still held, otherwise prev could be scheduled on another cpu, die
1884 * there before we look at prev->state, and then the reference would
1886 * Manfred Spraul <manfred@colorfullife.com>
1888 prev_state = prev->state;
1889 vtime_task_switch(prev);
1890 finish_arch_switch(prev);
1891 perf_event_task_sched_in(prev, current);
1892 finish_lock_switch(rq, prev);
1893 finish_arch_post_lock_switch();
1895 fire_sched_in_preempt_notifiers(current);
1898 if (unlikely(prev_state == TASK_DEAD)) {
1900 * Remove function-return probe instances associated with this
1901 * task and put them back on the free list.
1903 kprobe_flush_task(prev);
1904 put_task_struct(prev);
1907 tick_nohz_task_switch(current);
1912 /* assumes rq->lock is held */
1913 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1915 if (prev->sched_class->pre_schedule)
1916 prev->sched_class->pre_schedule(rq, prev);
1919 /* rq->lock is NOT held, but preemption is disabled */
1920 static inline void post_schedule(struct rq *rq)
1922 if (rq->post_schedule) {
1923 unsigned long flags;
1925 raw_spin_lock_irqsave(&rq->lock, flags);
1926 if (rq->curr->sched_class->post_schedule)
1927 rq->curr->sched_class->post_schedule(rq);
1928 raw_spin_unlock_irqrestore(&rq->lock, flags);
1930 rq->post_schedule = 0;
1936 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1940 static inline void post_schedule(struct rq *rq)
1947 * schedule_tail - first thing a freshly forked thread must call.
1948 * @prev: the thread we just switched away from.
1950 asmlinkage void schedule_tail(struct task_struct *prev)
1951 __releases(rq->lock)
1953 struct rq *rq = this_rq();
1955 finish_task_switch(rq, prev);
1958 * FIXME: do we need to worry about rq being invalidated by the
1963 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1964 /* In this case, finish_task_switch does not reenable preemption */
1967 if (current->set_child_tid)
1968 put_user(task_pid_vnr(current), current->set_child_tid);
1972 * context_switch - switch to the new MM and the new
1973 * thread's register state.
1976 context_switch(struct rq *rq, struct task_struct *prev,
1977 struct task_struct *next)
1979 struct mm_struct *mm, *oldmm;
1981 prepare_task_switch(rq, prev, next);
1984 oldmm = prev->active_mm;
1986 * For paravirt, this is coupled with an exit in switch_to to
1987 * combine the page table reload and the switch backend into
1990 arch_start_context_switch(prev);
1993 next->active_mm = oldmm;
1994 atomic_inc(&oldmm->mm_count);
1995 enter_lazy_tlb(oldmm, next);
1997 switch_mm(oldmm, mm, next);
2000 prev->active_mm = NULL;
2001 rq->prev_mm = oldmm;
2004 * Since the runqueue lock will be released by the next
2005 * task (which is an invalid locking op but in the case
2006 * of the scheduler it's an obvious special-case), so we
2007 * do an early lockdep release here:
2009 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2010 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2013 context_tracking_task_switch(prev, next);
2014 /* Here we just switch the register state and the stack. */
2015 switch_to(prev, next, prev);
2019 * this_rq must be evaluated again because prev may have moved
2020 * CPUs since it called schedule(), thus the 'rq' on its stack
2021 * frame will be invalid.
2023 finish_task_switch(this_rq(), prev);
2027 * nr_running and nr_context_switches:
2029 * externally visible scheduler statistics: current number of runnable
2030 * threads, total number of context switches performed since bootup.
2032 unsigned long nr_running(void)
2034 unsigned long i, sum = 0;
2036 for_each_online_cpu(i)
2037 sum += cpu_rq(i)->nr_running;
2042 unsigned long long nr_context_switches(void)
2045 unsigned long long sum = 0;
2047 for_each_possible_cpu(i)
2048 sum += cpu_rq(i)->nr_switches;
2053 unsigned long nr_iowait(void)
2055 unsigned long i, sum = 0;
2057 for_each_possible_cpu(i)
2058 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2063 unsigned long nr_iowait_cpu(int cpu)
2065 struct rq *this = cpu_rq(cpu);
2066 return atomic_read(&this->nr_iowait);
2072 * sched_exec - execve() is a valuable balancing opportunity, because at
2073 * this point the task has the smallest effective memory and cache footprint.
2075 void sched_exec(void)
2077 struct task_struct *p = current;
2078 unsigned long flags;
2081 raw_spin_lock_irqsave(&p->pi_lock, flags);
2082 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2083 if (dest_cpu == smp_processor_id())
2086 if (likely(cpu_active(dest_cpu))) {
2087 struct migration_arg arg = { p, dest_cpu };
2089 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2090 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2094 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2099 DEFINE_PER_CPU(struct kernel_stat, kstat);
2100 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2102 EXPORT_PER_CPU_SYMBOL(kstat);
2103 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2106 * Return any ns on the sched_clock that have not yet been accounted in
2107 * @p in case that task is currently running.
2109 * Called with task_rq_lock() held on @rq.
2111 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2115 if (task_current(rq, p)) {
2116 update_rq_clock(rq);
2117 ns = rq_clock_task(rq) - p->se.exec_start;
2125 unsigned long long task_delta_exec(struct task_struct *p)
2127 unsigned long flags;
2131 rq = task_rq_lock(p, &flags);
2132 ns = do_task_delta_exec(p, rq);
2133 task_rq_unlock(rq, p, &flags);
2139 * Return accounted runtime for the task.
2140 * In case the task is currently running, return the runtime plus current's
2141 * pending runtime that have not been accounted yet.
2143 unsigned long long task_sched_runtime(struct task_struct *p)
2145 unsigned long flags;
2149 rq = task_rq_lock(p, &flags);
2150 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2151 task_rq_unlock(rq, p, &flags);
2157 * This function gets called by the timer code, with HZ frequency.
2158 * We call it with interrupts disabled.
2160 void scheduler_tick(void)
2162 int cpu = smp_processor_id();
2163 struct rq *rq = cpu_rq(cpu);
2164 struct task_struct *curr = rq->curr;
2168 raw_spin_lock(&rq->lock);
2169 update_rq_clock(rq);
2170 curr->sched_class->task_tick(rq, curr, 0);
2171 update_cpu_load_active(rq);
2172 raw_spin_unlock(&rq->lock);
2174 perf_event_task_tick();
2177 rq->idle_balance = idle_cpu(cpu);
2178 trigger_load_balance(rq, cpu);
2180 rq_last_tick_reset(rq);
2183 #ifdef CONFIG_NO_HZ_FULL
2185 * scheduler_tick_max_deferment
2187 * Keep at least one tick per second when a single
2188 * active task is running because the scheduler doesn't
2189 * yet completely support full dynticks environment.
2191 * This makes sure that uptime, CFS vruntime, load
2192 * balancing, etc... continue to move forward, even
2193 * with a very low granularity.
2195 u64 scheduler_tick_max_deferment(void)
2197 struct rq *rq = this_rq();
2198 unsigned long next, now = ACCESS_ONCE(jiffies);
2200 next = rq->last_sched_tick + HZ;
2202 if (time_before_eq(next, now))
2205 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2209 notrace unsigned long get_parent_ip(unsigned long addr)
2211 if (in_lock_functions(addr)) {
2212 addr = CALLER_ADDR2;
2213 if (in_lock_functions(addr))
2214 addr = CALLER_ADDR3;
2219 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2220 defined(CONFIG_PREEMPT_TRACER))
2222 void __kprobes add_preempt_count(int val)
2224 #ifdef CONFIG_DEBUG_PREEMPT
2228 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2231 preempt_count() += val;
2232 #ifdef CONFIG_DEBUG_PREEMPT
2234 * Spinlock count overflowing soon?
2236 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2239 if (preempt_count() == val)
2240 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2242 EXPORT_SYMBOL(add_preempt_count);
2244 void __kprobes sub_preempt_count(int val)
2246 #ifdef CONFIG_DEBUG_PREEMPT
2250 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2253 * Is the spinlock portion underflowing?
2255 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2256 !(preempt_count() & PREEMPT_MASK)))
2260 if (preempt_count() == val)
2261 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2262 preempt_count() -= val;
2264 EXPORT_SYMBOL(sub_preempt_count);
2269 * Print scheduling while atomic bug:
2271 static noinline void __schedule_bug(struct task_struct *prev)
2273 if (oops_in_progress)
2276 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2277 prev->comm, prev->pid, preempt_count());
2279 debug_show_held_locks(prev);
2281 if (irqs_disabled())
2282 print_irqtrace_events(prev);
2284 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2288 * Various schedule()-time debugging checks and statistics:
2290 static inline void schedule_debug(struct task_struct *prev)
2293 * Test if we are atomic. Since do_exit() needs to call into
2294 * schedule() atomically, we ignore that path for now.
2295 * Otherwise, whine if we are scheduling when we should not be.
2297 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2298 __schedule_bug(prev);
2301 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2303 schedstat_inc(this_rq(), sched_count);
2306 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2308 if (prev->on_rq || rq->skip_clock_update < 0)
2309 update_rq_clock(rq);
2310 prev->sched_class->put_prev_task(rq, prev);
2314 * Pick up the highest-prio task:
2316 static inline struct task_struct *
2317 pick_next_task(struct rq *rq)
2319 const struct sched_class *class;
2320 struct task_struct *p;
2323 * Optimization: we know that if all tasks are in
2324 * the fair class we can call that function directly:
2326 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2327 p = fair_sched_class.pick_next_task(rq);
2332 for_each_class(class) {
2333 p = class->pick_next_task(rq);
2338 BUG(); /* the idle class will always have a runnable task */
2342 * __schedule() is the main scheduler function.
2344 * The main means of driving the scheduler and thus entering this function are:
2346 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2348 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2349 * paths. For example, see arch/x86/entry_64.S.
2351 * To drive preemption between tasks, the scheduler sets the flag in timer
2352 * interrupt handler scheduler_tick().
2354 * 3. Wakeups don't really cause entry into schedule(). They add a
2355 * task to the run-queue and that's it.
2357 * Now, if the new task added to the run-queue preempts the current
2358 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2359 * called on the nearest possible occasion:
2361 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2363 * - in syscall or exception context, at the next outmost
2364 * preempt_enable(). (this might be as soon as the wake_up()'s
2367 * - in IRQ context, return from interrupt-handler to
2368 * preemptible context
2370 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2373 * - cond_resched() call
2374 * - explicit schedule() call
2375 * - return from syscall or exception to user-space
2376 * - return from interrupt-handler to user-space
2378 static void __sched __schedule(void)
2380 struct task_struct *prev, *next;
2381 unsigned long *switch_count;
2387 cpu = smp_processor_id();
2389 rcu_note_context_switch(cpu);
2392 schedule_debug(prev);
2394 if (sched_feat(HRTICK))
2397 raw_spin_lock_irq(&rq->lock);
2399 switch_count = &prev->nivcsw;
2400 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2401 if (unlikely(signal_pending_state(prev->state, prev))) {
2402 prev->state = TASK_RUNNING;
2404 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2408 * If a worker went to sleep, notify and ask workqueue
2409 * whether it wants to wake up a task to maintain
2412 if (prev->flags & PF_WQ_WORKER) {
2413 struct task_struct *to_wakeup;
2415 to_wakeup = wq_worker_sleeping(prev, cpu);
2417 try_to_wake_up_local(to_wakeup);
2420 switch_count = &prev->nvcsw;
2423 pre_schedule(rq, prev);
2425 if (unlikely(!rq->nr_running))
2426 idle_balance(cpu, rq);
2428 put_prev_task(rq, prev);
2429 next = pick_next_task(rq);
2430 clear_tsk_need_resched(prev);
2431 rq->skip_clock_update = 0;
2433 if (likely(prev != next)) {
2438 context_switch(rq, prev, next); /* unlocks the rq */
2440 * The context switch have flipped the stack from under us
2441 * and restored the local variables which were saved when
2442 * this task called schedule() in the past. prev == current
2443 * is still correct, but it can be moved to another cpu/rq.
2445 cpu = smp_processor_id();
2448 raw_spin_unlock_irq(&rq->lock);
2452 sched_preempt_enable_no_resched();
2457 static inline void sched_submit_work(struct task_struct *tsk)
2459 if (!tsk->state || tsk_is_pi_blocked(tsk))
2462 * If we are going to sleep and we have plugged IO queued,
2463 * make sure to submit it to avoid deadlocks.
2465 if (blk_needs_flush_plug(tsk))
2466 blk_schedule_flush_plug(tsk);
2469 asmlinkage void __sched schedule(void)
2471 struct task_struct *tsk = current;
2473 sched_submit_work(tsk);
2476 EXPORT_SYMBOL(schedule);
2478 #ifdef CONFIG_CONTEXT_TRACKING
2479 asmlinkage void __sched schedule_user(void)
2482 * If we come here after a random call to set_need_resched(),
2483 * or we have been woken up remotely but the IPI has not yet arrived,
2484 * we haven't yet exited the RCU idle mode. Do it here manually until
2485 * we find a better solution.
2494 * schedule_preempt_disabled - called with preemption disabled
2496 * Returns with preemption disabled. Note: preempt_count must be 1
2498 void __sched schedule_preempt_disabled(void)
2500 sched_preempt_enable_no_resched();
2505 #ifdef CONFIG_PREEMPT
2507 * this is the entry point to schedule() from in-kernel preemption
2508 * off of preempt_enable. Kernel preemptions off return from interrupt
2509 * occur there and call schedule directly.
2511 asmlinkage void __sched notrace preempt_schedule(void)
2513 struct thread_info *ti = current_thread_info();
2516 * If there is a non-zero preempt_count or interrupts are disabled,
2517 * we do not want to preempt the current task. Just return..
2519 if (likely(ti->preempt_count || irqs_disabled()))
2523 add_preempt_count_notrace(PREEMPT_ACTIVE);
2525 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2528 * Check again in case we missed a preemption opportunity
2529 * between schedule and now.
2532 } while (need_resched());
2534 EXPORT_SYMBOL(preempt_schedule);
2537 * this is the entry point to schedule() from kernel preemption
2538 * off of irq context.
2539 * Note, that this is called and return with irqs disabled. This will
2540 * protect us against recursive calling from irq.
2542 asmlinkage void __sched preempt_schedule_irq(void)
2544 struct thread_info *ti = current_thread_info();
2545 enum ctx_state prev_state;
2547 /* Catch callers which need to be fixed */
2548 BUG_ON(ti->preempt_count || !irqs_disabled());
2550 prev_state = exception_enter();
2553 add_preempt_count(PREEMPT_ACTIVE);
2556 local_irq_disable();
2557 sub_preempt_count(PREEMPT_ACTIVE);
2560 * Check again in case we missed a preemption opportunity
2561 * between schedule and now.
2564 } while (need_resched());
2566 exception_exit(prev_state);
2569 #endif /* CONFIG_PREEMPT */
2571 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2574 return try_to_wake_up(curr->private, mode, wake_flags);
2576 EXPORT_SYMBOL(default_wake_function);
2579 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2580 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2581 * number) then we wake all the non-exclusive tasks and one exclusive task.
2583 * There are circumstances in which we can try to wake a task which has already
2584 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2585 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2587 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2588 int nr_exclusive, int wake_flags, void *key)
2590 wait_queue_t *curr, *next;
2592 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2593 unsigned flags = curr->flags;
2595 if (curr->func(curr, mode, wake_flags, key) &&
2596 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2602 * __wake_up - wake up threads blocked on a waitqueue.
2604 * @mode: which threads
2605 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2606 * @key: is directly passed to the wakeup function
2608 * It may be assumed that this function implies a write memory barrier before
2609 * changing the task state if and only if any tasks are woken up.
2611 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2612 int nr_exclusive, void *key)
2614 unsigned long flags;
2616 spin_lock_irqsave(&q->lock, flags);
2617 __wake_up_common(q, mode, nr_exclusive, 0, key);
2618 spin_unlock_irqrestore(&q->lock, flags);
2620 EXPORT_SYMBOL(__wake_up);
2623 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2625 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2627 __wake_up_common(q, mode, nr, 0, NULL);
2629 EXPORT_SYMBOL_GPL(__wake_up_locked);
2631 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2633 __wake_up_common(q, mode, 1, 0, key);
2635 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2638 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2640 * @mode: which threads
2641 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2642 * @key: opaque value to be passed to wakeup targets
2644 * The sync wakeup differs that the waker knows that it will schedule
2645 * away soon, so while the target thread will be woken up, it will not
2646 * be migrated to another CPU - ie. the two threads are 'synchronized'
2647 * with each other. This can prevent needless bouncing between CPUs.
2649 * On UP it can prevent extra preemption.
2651 * It may be assumed that this function implies a write memory barrier before
2652 * changing the task state if and only if any tasks are woken up.
2654 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2655 int nr_exclusive, void *key)
2657 unsigned long flags;
2658 int wake_flags = WF_SYNC;
2663 if (unlikely(!nr_exclusive))
2666 spin_lock_irqsave(&q->lock, flags);
2667 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2668 spin_unlock_irqrestore(&q->lock, flags);
2670 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2673 * __wake_up_sync - see __wake_up_sync_key()
2675 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2677 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2679 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2682 * complete: - signals a single thread waiting on this completion
2683 * @x: holds the state of this particular completion
2685 * This will wake up a single thread waiting on this completion. Threads will be
2686 * awakened in the same order in which they were queued.
2688 * See also complete_all(), wait_for_completion() and related routines.
2690 * It may be assumed that this function implies a write memory barrier before
2691 * changing the task state if and only if any tasks are woken up.
2693 void complete(struct completion *x)
2695 unsigned long flags;
2697 spin_lock_irqsave(&x->wait.lock, flags);
2699 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2700 spin_unlock_irqrestore(&x->wait.lock, flags);
2702 EXPORT_SYMBOL(complete);
2705 * complete_all: - signals all threads waiting on this completion
2706 * @x: holds the state of this particular completion
2708 * This will wake up all threads waiting on this particular completion event.
2710 * It may be assumed that this function implies a write memory barrier before
2711 * changing the task state if and only if any tasks are woken up.
2713 void complete_all(struct completion *x)
2715 unsigned long flags;
2717 spin_lock_irqsave(&x->wait.lock, flags);
2718 x->done += UINT_MAX/2;
2719 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2720 spin_unlock_irqrestore(&x->wait.lock, flags);
2722 EXPORT_SYMBOL(complete_all);
2724 static inline long __sched
2725 do_wait_for_common(struct completion *x,
2726 long (*action)(long), long timeout, int state)
2729 DECLARE_WAITQUEUE(wait, current);
2731 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2733 if (signal_pending_state(state, current)) {
2734 timeout = -ERESTARTSYS;
2737 __set_current_state(state);
2738 spin_unlock_irq(&x->wait.lock);
2739 timeout = action(timeout);
2740 spin_lock_irq(&x->wait.lock);
2741 } while (!x->done && timeout);
2742 __remove_wait_queue(&x->wait, &wait);
2747 return timeout ?: 1;
2750 static inline long __sched
2751 __wait_for_common(struct completion *x,
2752 long (*action)(long), long timeout, int state)
2756 spin_lock_irq(&x->wait.lock);
2757 timeout = do_wait_for_common(x, action, timeout, state);
2758 spin_unlock_irq(&x->wait.lock);
2763 wait_for_common(struct completion *x, long timeout, int state)
2765 return __wait_for_common(x, schedule_timeout, timeout, state);
2769 wait_for_common_io(struct completion *x, long timeout, int state)
2771 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2775 * wait_for_completion: - waits for completion of a task
2776 * @x: holds the state of this particular completion
2778 * This waits to be signaled for completion of a specific task. It is NOT
2779 * interruptible and there is no timeout.
2781 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2782 * and interrupt capability. Also see complete().
2784 void __sched wait_for_completion(struct completion *x)
2786 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2788 EXPORT_SYMBOL(wait_for_completion);
2791 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2792 * @x: holds the state of this particular completion
2793 * @timeout: timeout value in jiffies
2795 * This waits for either a completion of a specific task to be signaled or for a
2796 * specified timeout to expire. The timeout is in jiffies. It is not
2799 * The return value is 0 if timed out, and positive (at least 1, or number of
2800 * jiffies left till timeout) if completed.
2802 unsigned long __sched
2803 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2805 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2807 EXPORT_SYMBOL(wait_for_completion_timeout);
2810 * wait_for_completion_io: - waits for completion of a task
2811 * @x: holds the state of this particular completion
2813 * This waits to be signaled for completion of a specific task. It is NOT
2814 * interruptible and there is no timeout. The caller is accounted as waiting
2817 void __sched wait_for_completion_io(struct completion *x)
2819 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2821 EXPORT_SYMBOL(wait_for_completion_io);
2824 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2825 * @x: holds the state of this particular completion
2826 * @timeout: timeout value in jiffies
2828 * This waits for either a completion of a specific task to be signaled or for a
2829 * specified timeout to expire. The timeout is in jiffies. It is not
2830 * interruptible. The caller is accounted as waiting for IO.
2832 * The return value is 0 if timed out, and positive (at least 1, or number of
2833 * jiffies left till timeout) if completed.
2835 unsigned long __sched
2836 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2838 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2840 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2843 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2844 * @x: holds the state of this particular completion
2846 * This waits for completion of a specific task to be signaled. It is
2849 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2851 int __sched wait_for_completion_interruptible(struct completion *x)
2853 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2854 if (t == -ERESTARTSYS)
2858 EXPORT_SYMBOL(wait_for_completion_interruptible);
2861 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2862 * @x: holds the state of this particular completion
2863 * @timeout: timeout value in jiffies
2865 * This waits for either a completion of a specific task to be signaled or for a
2866 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2868 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2869 * positive (at least 1, or number of jiffies left till timeout) if completed.
2872 wait_for_completion_interruptible_timeout(struct completion *x,
2873 unsigned long timeout)
2875 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2877 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2880 * wait_for_completion_killable: - waits for completion of a task (killable)
2881 * @x: holds the state of this particular completion
2883 * This waits to be signaled for completion of a specific task. It can be
2884 * interrupted by a kill signal.
2886 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2888 int __sched wait_for_completion_killable(struct completion *x)
2890 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2891 if (t == -ERESTARTSYS)
2895 EXPORT_SYMBOL(wait_for_completion_killable);
2898 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2899 * @x: holds the state of this particular completion
2900 * @timeout: timeout value in jiffies
2902 * This waits for either a completion of a specific task to be
2903 * signaled or for a specified timeout to expire. It can be
2904 * interrupted by a kill signal. The timeout is in jiffies.
2906 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2907 * positive (at least 1, or number of jiffies left till timeout) if completed.
2910 wait_for_completion_killable_timeout(struct completion *x,
2911 unsigned long timeout)
2913 return wait_for_common(x, timeout, TASK_KILLABLE);
2915 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2918 * try_wait_for_completion - try to decrement a completion without blocking
2919 * @x: completion structure
2921 * Returns: 0 if a decrement cannot be done without blocking
2922 * 1 if a decrement succeeded.
2924 * If a completion is being used as a counting completion,
2925 * attempt to decrement the counter without blocking. This
2926 * enables us to avoid waiting if the resource the completion
2927 * is protecting is not available.
2929 bool try_wait_for_completion(struct completion *x)
2931 unsigned long flags;
2934 spin_lock_irqsave(&x->wait.lock, flags);
2939 spin_unlock_irqrestore(&x->wait.lock, flags);
2942 EXPORT_SYMBOL(try_wait_for_completion);
2945 * completion_done - Test to see if a completion has any waiters
2946 * @x: completion structure
2948 * Returns: 0 if there are waiters (wait_for_completion() in progress)
2949 * 1 if there are no waiters.
2952 bool completion_done(struct completion *x)
2954 unsigned long flags;
2957 spin_lock_irqsave(&x->wait.lock, flags);
2960 spin_unlock_irqrestore(&x->wait.lock, flags);
2963 EXPORT_SYMBOL(completion_done);
2966 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2968 unsigned long flags;
2971 init_waitqueue_entry(&wait, current);
2973 __set_current_state(state);
2975 spin_lock_irqsave(&q->lock, flags);
2976 __add_wait_queue(q, &wait);
2977 spin_unlock(&q->lock);
2978 timeout = schedule_timeout(timeout);
2979 spin_lock_irq(&q->lock);
2980 __remove_wait_queue(q, &wait);
2981 spin_unlock_irqrestore(&q->lock, flags);
2986 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2988 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2990 EXPORT_SYMBOL(interruptible_sleep_on);
2993 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2995 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2997 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2999 void __sched sleep_on(wait_queue_head_t *q)
3001 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3003 EXPORT_SYMBOL(sleep_on);
3005 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3007 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3009 EXPORT_SYMBOL(sleep_on_timeout);
3011 #ifdef CONFIG_RT_MUTEXES
3014 * rt_mutex_setprio - set the current priority of a task
3016 * @prio: prio value (kernel-internal form)
3018 * This function changes the 'effective' priority of a task. It does
3019 * not touch ->normal_prio like __setscheduler().
3021 * Used by the rt_mutex code to implement priority inheritance logic.
3023 void rt_mutex_setprio(struct task_struct *p, int prio)
3025 int oldprio, on_rq, running;
3027 const struct sched_class *prev_class;
3029 BUG_ON(prio < 0 || prio > MAX_PRIO);
3031 rq = __task_rq_lock(p);
3034 * Idle task boosting is a nono in general. There is one
3035 * exception, when PREEMPT_RT and NOHZ is active:
3037 * The idle task calls get_next_timer_interrupt() and holds
3038 * the timer wheel base->lock on the CPU and another CPU wants
3039 * to access the timer (probably to cancel it). We can safely
3040 * ignore the boosting request, as the idle CPU runs this code
3041 * with interrupts disabled and will complete the lock
3042 * protected section without being interrupted. So there is no
3043 * real need to boost.
3045 if (unlikely(p == rq->idle)) {
3046 WARN_ON(p != rq->curr);
3047 WARN_ON(p->pi_blocked_on);
3051 trace_sched_pi_setprio(p, prio);
3053 prev_class = p->sched_class;
3055 running = task_current(rq, p);
3057 dequeue_task(rq, p, 0);
3059 p->sched_class->put_prev_task(rq, p);
3062 p->sched_class = &rt_sched_class;
3064 p->sched_class = &fair_sched_class;
3069 p->sched_class->set_curr_task(rq);
3071 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3073 check_class_changed(rq, p, prev_class, oldprio);
3075 __task_rq_unlock(rq);
3078 void set_user_nice(struct task_struct *p, long nice)
3080 int old_prio, delta, on_rq;
3081 unsigned long flags;
3084 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3087 * We have to be careful, if called from sys_setpriority(),
3088 * the task might be in the middle of scheduling on another CPU.
3090 rq = task_rq_lock(p, &flags);
3092 * The RT priorities are set via sched_setscheduler(), but we still
3093 * allow the 'normal' nice value to be set - but as expected
3094 * it wont have any effect on scheduling until the task is
3095 * SCHED_FIFO/SCHED_RR:
3097 if (task_has_rt_policy(p)) {
3098 p->static_prio = NICE_TO_PRIO(nice);
3103 dequeue_task(rq, p, 0);
3105 p->static_prio = NICE_TO_PRIO(nice);
3108 p->prio = effective_prio(p);
3109 delta = p->prio - old_prio;
3112 enqueue_task(rq, p, 0);
3114 * If the task increased its priority or is running and
3115 * lowered its priority, then reschedule its CPU:
3117 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3118 resched_task(rq->curr);
3121 task_rq_unlock(rq, p, &flags);
3123 EXPORT_SYMBOL(set_user_nice);
3126 * can_nice - check if a task can reduce its nice value
3130 int can_nice(const struct task_struct *p, const int nice)
3132 /* convert nice value [19,-20] to rlimit style value [1,40] */
3133 int nice_rlim = 20 - nice;
3135 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3136 capable(CAP_SYS_NICE));
3139 #ifdef __ARCH_WANT_SYS_NICE
3142 * sys_nice - change the priority of the current process.
3143 * @increment: priority increment
3145 * sys_setpriority is a more generic, but much slower function that
3146 * does similar things.
3148 SYSCALL_DEFINE1(nice, int, increment)
3153 * Setpriority might change our priority at the same moment.
3154 * We don't have to worry. Conceptually one call occurs first
3155 * and we have a single winner.
3157 if (increment < -40)
3162 nice = TASK_NICE(current) + increment;
3168 if (increment < 0 && !can_nice(current, nice))
3171 retval = security_task_setnice(current, nice);
3175 set_user_nice(current, nice);
3182 * task_prio - return the priority value of a given task.
3183 * @p: the task in question.
3185 * This is the priority value as seen by users in /proc.
3186 * RT tasks are offset by -200. Normal tasks are centered
3187 * around 0, value goes from -16 to +15.
3189 int task_prio(const struct task_struct *p)
3191 return p->prio - MAX_RT_PRIO;
3195 * task_nice - return the nice value of a given task.
3196 * @p: the task in question.
3198 int task_nice(const struct task_struct *p)
3200 return TASK_NICE(p);
3202 EXPORT_SYMBOL(task_nice);
3205 * idle_cpu - is a given cpu idle currently?
3206 * @cpu: the processor in question.
3208 int idle_cpu(int cpu)
3210 struct rq *rq = cpu_rq(cpu);
3212 if (rq->curr != rq->idle)
3219 if (!llist_empty(&rq->wake_list))
3227 * idle_task - return the idle task for a given cpu.
3228 * @cpu: the processor in question.
3230 struct task_struct *idle_task(int cpu)
3232 return cpu_rq(cpu)->idle;
3236 * find_process_by_pid - find a process with a matching PID value.
3237 * @pid: the pid in question.
3239 static struct task_struct *find_process_by_pid(pid_t pid)
3241 return pid ? find_task_by_vpid(pid) : current;
3244 /* Actually do priority change: must hold rq lock. */
3246 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3249 p->rt_priority = prio;
3250 p->normal_prio = normal_prio(p);
3251 /* we are holding p->pi_lock already */
3252 p->prio = rt_mutex_getprio(p);
3253 if (rt_prio(p->prio))
3254 p->sched_class = &rt_sched_class;
3256 p->sched_class = &fair_sched_class;
3261 * check the target process has a UID that matches the current process's
3263 static bool check_same_owner(struct task_struct *p)
3265 const struct cred *cred = current_cred(), *pcred;
3269 pcred = __task_cred(p);
3270 match = (uid_eq(cred->euid, pcred->euid) ||
3271 uid_eq(cred->euid, pcred->uid));
3276 static int __sched_setscheduler(struct task_struct *p, int policy,
3277 const struct sched_param *param, bool user)
3279 int retval, oldprio, oldpolicy = -1, on_rq, running;
3280 unsigned long flags;
3281 const struct sched_class *prev_class;
3285 /* may grab non-irq protected spin_locks */
3286 BUG_ON(in_interrupt());
3288 /* double check policy once rq lock held */
3290 reset_on_fork = p->sched_reset_on_fork;
3291 policy = oldpolicy = p->policy;
3293 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3294 policy &= ~SCHED_RESET_ON_FORK;
3296 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3297 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3298 policy != SCHED_IDLE)
3303 * Valid priorities for SCHED_FIFO and SCHED_RR are
3304 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3305 * SCHED_BATCH and SCHED_IDLE is 0.
3307 if (param->sched_priority < 0 ||
3308 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3309 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3311 if (rt_policy(policy) != (param->sched_priority != 0))
3315 * Allow unprivileged RT tasks to decrease priority:
3317 if (user && !capable(CAP_SYS_NICE)) {
3318 if (rt_policy(policy)) {
3319 unsigned long rlim_rtprio =
3320 task_rlimit(p, RLIMIT_RTPRIO);
3322 /* can't set/change the rt policy */
3323 if (policy != p->policy && !rlim_rtprio)
3326 /* can't increase priority */
3327 if (param->sched_priority > p->rt_priority &&
3328 param->sched_priority > rlim_rtprio)
3333 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3334 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3336 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3337 if (!can_nice(p, TASK_NICE(p)))
3341 /* can't change other user's priorities */
3342 if (!check_same_owner(p))
3345 /* Normal users shall not reset the sched_reset_on_fork flag */
3346 if (p->sched_reset_on_fork && !reset_on_fork)
3351 retval = security_task_setscheduler(p);
3357 * make sure no PI-waiters arrive (or leave) while we are
3358 * changing the priority of the task:
3360 * To be able to change p->policy safely, the appropriate
3361 * runqueue lock must be held.
3363 rq = task_rq_lock(p, &flags);
3366 * Changing the policy of the stop threads its a very bad idea
3368 if (p == rq->stop) {
3369 task_rq_unlock(rq, p, &flags);
3374 * If not changing anything there's no need to proceed further:
3376 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3377 param->sched_priority == p->rt_priority))) {
3378 task_rq_unlock(rq, p, &flags);
3382 #ifdef CONFIG_RT_GROUP_SCHED
3385 * Do not allow realtime tasks into groups that have no runtime
3388 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3389 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3390 !task_group_is_autogroup(task_group(p))) {
3391 task_rq_unlock(rq, p, &flags);
3397 /* recheck policy now with rq lock held */
3398 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3399 policy = oldpolicy = -1;
3400 task_rq_unlock(rq, p, &flags);
3404 running = task_current(rq, p);
3406 dequeue_task(rq, p, 0);
3408 p->sched_class->put_prev_task(rq, p);
3410 p->sched_reset_on_fork = reset_on_fork;
3413 prev_class = p->sched_class;
3414 __setscheduler(rq, p, policy, param->sched_priority);
3417 p->sched_class->set_curr_task(rq);
3419 enqueue_task(rq, p, 0);
3421 check_class_changed(rq, p, prev_class, oldprio);
3422 task_rq_unlock(rq, p, &flags);
3424 rt_mutex_adjust_pi(p);
3430 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3431 * @p: the task in question.
3432 * @policy: new policy.
3433 * @param: structure containing the new RT priority.
3435 * NOTE that the task may be already dead.
3437 int sched_setscheduler(struct task_struct *p, int policy,
3438 const struct sched_param *param)
3440 return __sched_setscheduler(p, policy, param, true);
3442 EXPORT_SYMBOL_GPL(sched_setscheduler);
3445 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3446 * @p: the task in question.
3447 * @policy: new policy.
3448 * @param: structure containing the new RT priority.
3450 * Just like sched_setscheduler, only don't bother checking if the
3451 * current context has permission. For example, this is needed in
3452 * stop_machine(): we create temporary high priority worker threads,
3453 * but our caller might not have that capability.
3455 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3456 const struct sched_param *param)
3458 return __sched_setscheduler(p, policy, param, false);
3462 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3464 struct sched_param lparam;
3465 struct task_struct *p;
3468 if (!param || pid < 0)
3470 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3475 p = find_process_by_pid(pid);
3477 retval = sched_setscheduler(p, policy, &lparam);
3484 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3485 * @pid: the pid in question.
3486 * @policy: new policy.
3487 * @param: structure containing the new RT priority.
3489 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3490 struct sched_param __user *, param)
3492 /* negative values for policy are not valid */
3496 return do_sched_setscheduler(pid, policy, param);
3500 * sys_sched_setparam - set/change the RT priority of a thread
3501 * @pid: the pid in question.
3502 * @param: structure containing the new RT priority.
3504 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3506 return do_sched_setscheduler(pid, -1, param);
3510 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3511 * @pid: the pid in question.
3513 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3515 struct task_struct *p;
3523 p = find_process_by_pid(pid);
3525 retval = security_task_getscheduler(p);
3528 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3535 * sys_sched_getparam - get the RT priority of a thread
3536 * @pid: the pid in question.
3537 * @param: structure containing the RT priority.
3539 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3541 struct sched_param lp;
3542 struct task_struct *p;
3545 if (!param || pid < 0)
3549 p = find_process_by_pid(pid);
3554 retval = security_task_getscheduler(p);
3558 lp.sched_priority = p->rt_priority;
3562 * This one might sleep, we cannot do it with a spinlock held ...
3564 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3573 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3575 cpumask_var_t cpus_allowed, new_mask;
3576 struct task_struct *p;
3582 p = find_process_by_pid(pid);
3589 /* Prevent p going away */
3593 if (p->flags & PF_NO_SETAFFINITY) {
3597 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3601 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3603 goto out_free_cpus_allowed;
3606 if (!check_same_owner(p)) {
3608 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3615 retval = security_task_setscheduler(p);
3619 cpuset_cpus_allowed(p, cpus_allowed);
3620 cpumask_and(new_mask, in_mask, cpus_allowed);
3622 retval = set_cpus_allowed_ptr(p, new_mask);
3625 cpuset_cpus_allowed(p, cpus_allowed);
3626 if (!cpumask_subset(new_mask, cpus_allowed)) {
3628 * We must have raced with a concurrent cpuset
3629 * update. Just reset the cpus_allowed to the
3630 * cpuset's cpus_allowed
3632 cpumask_copy(new_mask, cpus_allowed);
3637 free_cpumask_var(new_mask);
3638 out_free_cpus_allowed:
3639 free_cpumask_var(cpus_allowed);
3646 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3647 struct cpumask *new_mask)
3649 if (len < cpumask_size())
3650 cpumask_clear(new_mask);
3651 else if (len > cpumask_size())
3652 len = cpumask_size();
3654 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3658 * sys_sched_setaffinity - set the cpu affinity of a process
3659 * @pid: pid of the process
3660 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3661 * @user_mask_ptr: user-space pointer to the new cpu mask
3663 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3664 unsigned long __user *, user_mask_ptr)
3666 cpumask_var_t new_mask;
3669 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3672 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3674 retval = sched_setaffinity(pid, new_mask);
3675 free_cpumask_var(new_mask);
3679 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3681 struct task_struct *p;
3682 unsigned long flags;
3689 p = find_process_by_pid(pid);
3693 retval = security_task_getscheduler(p);
3697 raw_spin_lock_irqsave(&p->pi_lock, flags);
3698 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3699 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3709 * sys_sched_getaffinity - get the cpu affinity of a process
3710 * @pid: pid of the process
3711 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3712 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3714 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3715 unsigned long __user *, user_mask_ptr)
3720 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3722 if (len & (sizeof(unsigned long)-1))
3725 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3728 ret = sched_getaffinity(pid, mask);
3730 size_t retlen = min_t(size_t, len, cpumask_size());
3732 if (copy_to_user(user_mask_ptr, mask, retlen))
3737 free_cpumask_var(mask);
3743 * sys_sched_yield - yield the current processor to other threads.
3745 * This function yields the current CPU to other tasks. If there are no
3746 * other threads running on this CPU then this function will return.
3748 SYSCALL_DEFINE0(sched_yield)
3750 struct rq *rq = this_rq_lock();
3752 schedstat_inc(rq, yld_count);
3753 current->sched_class->yield_task(rq);
3756 * Since we are going to call schedule() anyway, there's
3757 * no need to preempt or enable interrupts:
3759 __release(rq->lock);
3760 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3761 do_raw_spin_unlock(&rq->lock);
3762 sched_preempt_enable_no_resched();
3769 static inline int should_resched(void)
3771 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
3774 static void __cond_resched(void)
3776 add_preempt_count(PREEMPT_ACTIVE);
3778 sub_preempt_count(PREEMPT_ACTIVE);
3781 int __sched _cond_resched(void)
3783 if (should_resched()) {
3789 EXPORT_SYMBOL(_cond_resched);
3792 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3793 * call schedule, and on return reacquire the lock.
3795 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3796 * operations here to prevent schedule() from being called twice (once via
3797 * spin_unlock(), once by hand).
3799 int __cond_resched_lock(spinlock_t *lock)
3801 int resched = should_resched();
3804 lockdep_assert_held(lock);
3806 if (spin_needbreak(lock) || resched) {
3817 EXPORT_SYMBOL(__cond_resched_lock);
3819 int __sched __cond_resched_softirq(void)
3821 BUG_ON(!in_softirq());
3823 if (should_resched()) {
3831 EXPORT_SYMBOL(__cond_resched_softirq);
3834 * yield - yield the current processor to other threads.
3836 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3838 * The scheduler is at all times free to pick the calling task as the most
3839 * eligible task to run, if removing the yield() call from your code breaks
3840 * it, its already broken.
3842 * Typical broken usage is:
3847 * where one assumes that yield() will let 'the other' process run that will
3848 * make event true. If the current task is a SCHED_FIFO task that will never
3849 * happen. Never use yield() as a progress guarantee!!
3851 * If you want to use yield() to wait for something, use wait_event().
3852 * If you want to use yield() to be 'nice' for others, use cond_resched().
3853 * If you still want to use yield(), do not!
3855 void __sched yield(void)
3857 set_current_state(TASK_RUNNING);
3860 EXPORT_SYMBOL(yield);
3863 * yield_to - yield the current processor to another thread in
3864 * your thread group, or accelerate that thread toward the
3865 * processor it's on.
3867 * @preempt: whether task preemption is allowed or not
3869 * It's the caller's job to ensure that the target task struct
3870 * can't go away on us before we can do any checks.
3873 * true (>0) if we indeed boosted the target task.
3874 * false (0) if we failed to boost the target.
3875 * -ESRCH if there's no task to yield to.
3877 bool __sched yield_to(struct task_struct *p, bool preempt)
3879 struct task_struct *curr = current;
3880 struct rq *rq, *p_rq;
3881 unsigned long flags;
3884 local_irq_save(flags);
3890 * If we're the only runnable task on the rq and target rq also
3891 * has only one task, there's absolutely no point in yielding.
3893 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3898 double_rq_lock(rq, p_rq);
3899 while (task_rq(p) != p_rq) {
3900 double_rq_unlock(rq, p_rq);
3904 if (!curr->sched_class->yield_to_task)
3907 if (curr->sched_class != p->sched_class)
3910 if (task_running(p_rq, p) || p->state)
3913 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3915 schedstat_inc(rq, yld_count);
3917 * Make p's CPU reschedule; pick_next_entity takes care of
3920 if (preempt && rq != p_rq)
3921 resched_task(p_rq->curr);
3925 double_rq_unlock(rq, p_rq);
3927 local_irq_restore(flags);
3934 EXPORT_SYMBOL_GPL(yield_to);
3937 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3938 * that process accounting knows that this is a task in IO wait state.
3940 void __sched io_schedule(void)
3942 struct rq *rq = raw_rq();
3944 delayacct_blkio_start();
3945 atomic_inc(&rq->nr_iowait);
3946 blk_flush_plug(current);
3947 current->in_iowait = 1;
3949 current->in_iowait = 0;
3950 atomic_dec(&rq->nr_iowait);
3951 delayacct_blkio_end();
3953 EXPORT_SYMBOL(io_schedule);
3955 long __sched io_schedule_timeout(long timeout)
3957 struct rq *rq = raw_rq();
3960 delayacct_blkio_start();
3961 atomic_inc(&rq->nr_iowait);
3962 blk_flush_plug(current);
3963 current->in_iowait = 1;
3964 ret = schedule_timeout(timeout);
3965 current->in_iowait = 0;
3966 atomic_dec(&rq->nr_iowait);
3967 delayacct_blkio_end();
3972 * sys_sched_get_priority_max - return maximum RT priority.
3973 * @policy: scheduling class.
3975 * this syscall returns the maximum rt_priority that can be used
3976 * by a given scheduling class.
3978 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3985 ret = MAX_USER_RT_PRIO-1;
3997 * sys_sched_get_priority_min - return minimum RT priority.
3998 * @policy: scheduling class.
4000 * this syscall returns the minimum rt_priority that can be used
4001 * by a given scheduling class.
4003 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4021 * sys_sched_rr_get_interval - return the default timeslice of a process.
4022 * @pid: pid of the process.
4023 * @interval: userspace pointer to the timeslice value.
4025 * this syscall writes the default timeslice value of a given process
4026 * into the user-space timespec buffer. A value of '0' means infinity.
4028 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4029 struct timespec __user *, interval)
4031 struct task_struct *p;
4032 unsigned int time_slice;
4033 unsigned long flags;
4043 p = find_process_by_pid(pid);
4047 retval = security_task_getscheduler(p);
4051 rq = task_rq_lock(p, &flags);
4052 time_slice = p->sched_class->get_rr_interval(rq, p);
4053 task_rq_unlock(rq, p, &flags);
4056 jiffies_to_timespec(time_slice, &t);
4057 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4065 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4067 void sched_show_task(struct task_struct *p)
4069 unsigned long free = 0;
4073 state = p->state ? __ffs(p->state) + 1 : 0;
4074 printk(KERN_INFO "%-15.15s %c", p->comm,
4075 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4076 #if BITS_PER_LONG == 32
4077 if (state == TASK_RUNNING)
4078 printk(KERN_CONT " running ");
4080 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4082 if (state == TASK_RUNNING)
4083 printk(KERN_CONT " running task ");
4085 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4087 #ifdef CONFIG_DEBUG_STACK_USAGE
4088 free = stack_not_used(p);
4091 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4093 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4094 task_pid_nr(p), ppid,
4095 (unsigned long)task_thread_info(p)->flags);
4097 print_worker_info(KERN_INFO, p);
4098 show_stack(p, NULL);
4101 void show_state_filter(unsigned long state_filter)
4103 struct task_struct *g, *p;
4105 #if BITS_PER_LONG == 32
4107 " task PC stack pid father\n");
4110 " task PC stack pid father\n");
4113 do_each_thread(g, p) {
4115 * reset the NMI-timeout, listing all files on a slow
4116 * console might take a lot of time:
4118 touch_nmi_watchdog();
4119 if (!state_filter || (p->state & state_filter))
4121 } while_each_thread(g, p);
4123 touch_all_softlockup_watchdogs();
4125 #ifdef CONFIG_SCHED_DEBUG
4126 sysrq_sched_debug_show();
4130 * Only show locks if all tasks are dumped:
4133 debug_show_all_locks();
4136 void init_idle_bootup_task(struct task_struct *idle)
4138 idle->sched_class = &idle_sched_class;
4142 * init_idle - set up an idle thread for a given CPU
4143 * @idle: task in question
4144 * @cpu: cpu the idle task belongs to
4146 * NOTE: this function does not set the idle thread's NEED_RESCHED
4147 * flag, to make booting more robust.
4149 void init_idle(struct task_struct *idle, int cpu)
4151 struct rq *rq = cpu_rq(cpu);
4152 unsigned long flags;
4154 raw_spin_lock_irqsave(&rq->lock, flags);
4157 idle->state = TASK_RUNNING;
4158 idle->se.exec_start = sched_clock();
4160 do_set_cpus_allowed(idle, cpumask_of(cpu));
4162 * We're having a chicken and egg problem, even though we are
4163 * holding rq->lock, the cpu isn't yet set to this cpu so the
4164 * lockdep check in task_group() will fail.
4166 * Similar case to sched_fork(). / Alternatively we could
4167 * use task_rq_lock() here and obtain the other rq->lock.
4172 __set_task_cpu(idle, cpu);
4175 rq->curr = rq->idle = idle;
4176 #if defined(CONFIG_SMP)
4179 raw_spin_unlock_irqrestore(&rq->lock, flags);
4181 /* Set the preempt count _outside_ the spinlocks! */
4182 task_thread_info(idle)->preempt_count = 0;
4185 * The idle tasks have their own, simple scheduling class:
4187 idle->sched_class = &idle_sched_class;
4188 ftrace_graph_init_idle_task(idle, cpu);
4189 vtime_init_idle(idle, cpu);
4190 #if defined(CONFIG_SMP)
4191 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4196 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4198 if (p->sched_class && p->sched_class->set_cpus_allowed)
4199 p->sched_class->set_cpus_allowed(p, new_mask);
4201 cpumask_copy(&p->cpus_allowed, new_mask);
4202 p->nr_cpus_allowed = cpumask_weight(new_mask);
4206 * This is how migration works:
4208 * 1) we invoke migration_cpu_stop() on the target CPU using
4210 * 2) stopper starts to run (implicitly forcing the migrated thread
4212 * 3) it checks whether the migrated task is still in the wrong runqueue.
4213 * 4) if it's in the wrong runqueue then the migration thread removes
4214 * it and puts it into the right queue.
4215 * 5) stopper completes and stop_one_cpu() returns and the migration
4220 * Change a given task's CPU affinity. Migrate the thread to a
4221 * proper CPU and schedule it away if the CPU it's executing on
4222 * is removed from the allowed bitmask.
4224 * NOTE: the caller must have a valid reference to the task, the
4225 * task must not exit() & deallocate itself prematurely. The
4226 * call is not atomic; no spinlocks may be held.
4228 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4230 unsigned long flags;
4232 unsigned int dest_cpu;
4235 rq = task_rq_lock(p, &flags);
4237 if (cpumask_equal(&p->cpus_allowed, new_mask))
4240 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4245 do_set_cpus_allowed(p, new_mask);
4247 /* Can the task run on the task's current CPU? If so, we're done */
4248 if (cpumask_test_cpu(task_cpu(p), new_mask))
4251 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4253 struct migration_arg arg = { p, dest_cpu };
4254 /* Need help from migration thread: drop lock and wait. */
4255 task_rq_unlock(rq, p, &flags);
4256 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4257 tlb_migrate_finish(p->mm);
4261 task_rq_unlock(rq, p, &flags);
4265 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4268 * Move (not current) task off this cpu, onto dest cpu. We're doing
4269 * this because either it can't run here any more (set_cpus_allowed()
4270 * away from this CPU, or CPU going down), or because we're
4271 * attempting to rebalance this task on exec (sched_exec).
4273 * So we race with normal scheduler movements, but that's OK, as long
4274 * as the task is no longer on this CPU.
4276 * Returns non-zero if task was successfully migrated.
4278 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4280 struct rq *rq_dest, *rq_src;
4283 if (unlikely(!cpu_active(dest_cpu)))
4286 rq_src = cpu_rq(src_cpu);
4287 rq_dest = cpu_rq(dest_cpu);
4289 raw_spin_lock(&p->pi_lock);
4290 double_rq_lock(rq_src, rq_dest);
4291 /* Already moved. */
4292 if (task_cpu(p) != src_cpu)
4294 /* Affinity changed (again). */
4295 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4299 * If we're not on a rq, the next wake-up will ensure we're
4303 dequeue_task(rq_src, p, 0);
4304 set_task_cpu(p, dest_cpu);
4305 enqueue_task(rq_dest, p, 0);
4306 check_preempt_curr(rq_dest, p, 0);
4311 double_rq_unlock(rq_src, rq_dest);
4312 raw_spin_unlock(&p->pi_lock);
4317 * migration_cpu_stop - this will be executed by a highprio stopper thread
4318 * and performs thread migration by bumping thread off CPU then
4319 * 'pushing' onto another runqueue.
4321 static int migration_cpu_stop(void *data)
4323 struct migration_arg *arg = data;
4326 * The original target cpu might have gone down and we might
4327 * be on another cpu but it doesn't matter.
4329 local_irq_disable();
4330 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4335 #ifdef CONFIG_HOTPLUG_CPU
4338 * Ensures that the idle task is using init_mm right before its cpu goes
4341 void idle_task_exit(void)
4343 struct mm_struct *mm = current->active_mm;
4345 BUG_ON(cpu_online(smp_processor_id()));
4348 switch_mm(mm, &init_mm, current);
4353 * Since this CPU is going 'away' for a while, fold any nr_active delta
4354 * we might have. Assumes we're called after migrate_tasks() so that the
4355 * nr_active count is stable.
4357 * Also see the comment "Global load-average calculations".
4359 static void calc_load_migrate(struct rq *rq)
4361 long delta = calc_load_fold_active(rq);
4363 atomic_long_add(delta, &calc_load_tasks);
4367 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4368 * try_to_wake_up()->select_task_rq().
4370 * Called with rq->lock held even though we'er in stop_machine() and
4371 * there's no concurrency possible, we hold the required locks anyway
4372 * because of lock validation efforts.
4374 static void migrate_tasks(unsigned int dead_cpu)
4376 struct rq *rq = cpu_rq(dead_cpu);
4377 struct task_struct *next, *stop = rq->stop;
4381 * Fudge the rq selection such that the below task selection loop
4382 * doesn't get stuck on the currently eligible stop task.
4384 * We're currently inside stop_machine() and the rq is either stuck
4385 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4386 * either way we should never end up calling schedule() until we're
4392 * put_prev_task() and pick_next_task() sched
4393 * class method both need to have an up-to-date
4394 * value of rq->clock[_task]
4396 update_rq_clock(rq);
4400 * There's this thread running, bail when that's the only
4403 if (rq->nr_running == 1)
4406 next = pick_next_task(rq);
4408 next->sched_class->put_prev_task(rq, next);
4410 /* Find suitable destination for @next, with force if needed. */
4411 dest_cpu = select_fallback_rq(dead_cpu, next);
4412 raw_spin_unlock(&rq->lock);
4414 __migrate_task(next, dead_cpu, dest_cpu);
4416 raw_spin_lock(&rq->lock);
4422 #endif /* CONFIG_HOTPLUG_CPU */
4424 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4426 static struct ctl_table sd_ctl_dir[] = {
4428 .procname = "sched_domain",
4434 static struct ctl_table sd_ctl_root[] = {
4436 .procname = "kernel",
4438 .child = sd_ctl_dir,
4443 static struct ctl_table *sd_alloc_ctl_entry(int n)
4445 struct ctl_table *entry =
4446 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4451 static void sd_free_ctl_entry(struct ctl_table **tablep)
4453 struct ctl_table *entry;
4456 * In the intermediate directories, both the child directory and
4457 * procname are dynamically allocated and could fail but the mode
4458 * will always be set. In the lowest directory the names are
4459 * static strings and all have proc handlers.
4461 for (entry = *tablep; entry->mode; entry++) {
4463 sd_free_ctl_entry(&entry->child);
4464 if (entry->proc_handler == NULL)
4465 kfree(entry->procname);
4472 static int min_load_idx = 0;
4473 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4476 set_table_entry(struct ctl_table *entry,
4477 const char *procname, void *data, int maxlen,
4478 umode_t mode, proc_handler *proc_handler,
4481 entry->procname = procname;
4483 entry->maxlen = maxlen;
4485 entry->proc_handler = proc_handler;
4488 entry->extra1 = &min_load_idx;
4489 entry->extra2 = &max_load_idx;
4493 static struct ctl_table *
4494 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4496 struct ctl_table *table = sd_alloc_ctl_entry(13);
4501 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4502 sizeof(long), 0644, proc_doulongvec_minmax, false);
4503 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4504 sizeof(long), 0644, proc_doulongvec_minmax, false);
4505 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4506 sizeof(int), 0644, proc_dointvec_minmax, true);
4507 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4508 sizeof(int), 0644, proc_dointvec_minmax, true);
4509 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4510 sizeof(int), 0644, proc_dointvec_minmax, true);
4511 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4512 sizeof(int), 0644, proc_dointvec_minmax, true);
4513 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4514 sizeof(int), 0644, proc_dointvec_minmax, true);
4515 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4516 sizeof(int), 0644, proc_dointvec_minmax, false);
4517 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4518 sizeof(int), 0644, proc_dointvec_minmax, false);
4519 set_table_entry(&table[9], "cache_nice_tries",
4520 &sd->cache_nice_tries,
4521 sizeof(int), 0644, proc_dointvec_minmax, false);
4522 set_table_entry(&table[10], "flags", &sd->flags,
4523 sizeof(int), 0644, proc_dointvec_minmax, false);
4524 set_table_entry(&table[11], "name", sd->name,
4525 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4526 /* &table[12] is terminator */
4531 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4533 struct ctl_table *entry, *table;
4534 struct sched_domain *sd;
4535 int domain_num = 0, i;
4538 for_each_domain(cpu, sd)
4540 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4545 for_each_domain(cpu, sd) {
4546 snprintf(buf, 32, "domain%d", i);
4547 entry->procname = kstrdup(buf, GFP_KERNEL);
4549 entry->child = sd_alloc_ctl_domain_table(sd);
4556 static struct ctl_table_header *sd_sysctl_header;
4557 static void register_sched_domain_sysctl(void)
4559 int i, cpu_num = num_possible_cpus();
4560 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4563 WARN_ON(sd_ctl_dir[0].child);
4564 sd_ctl_dir[0].child = entry;
4569 for_each_possible_cpu(i) {
4570 snprintf(buf, 32, "cpu%d", i);
4571 entry->procname = kstrdup(buf, GFP_KERNEL);
4573 entry->child = sd_alloc_ctl_cpu_table(i);
4577 WARN_ON(sd_sysctl_header);
4578 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4581 /* may be called multiple times per register */
4582 static void unregister_sched_domain_sysctl(void)
4584 if (sd_sysctl_header)
4585 unregister_sysctl_table(sd_sysctl_header);
4586 sd_sysctl_header = NULL;
4587 if (sd_ctl_dir[0].child)
4588 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4591 static void register_sched_domain_sysctl(void)
4594 static void unregister_sched_domain_sysctl(void)
4599 static void set_rq_online(struct rq *rq)
4602 const struct sched_class *class;
4604 cpumask_set_cpu(rq->cpu, rq->rd->online);
4607 for_each_class(class) {
4608 if (class->rq_online)
4609 class->rq_online(rq);
4614 static void set_rq_offline(struct rq *rq)
4617 const struct sched_class *class;
4619 for_each_class(class) {
4620 if (class->rq_offline)
4621 class->rq_offline(rq);
4624 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4630 * migration_call - callback that gets triggered when a CPU is added.
4631 * Here we can start up the necessary migration thread for the new CPU.
4634 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4636 int cpu = (long)hcpu;
4637 unsigned long flags;
4638 struct rq *rq = cpu_rq(cpu);
4640 switch (action & ~CPU_TASKS_FROZEN) {
4642 case CPU_UP_PREPARE:
4643 rq->calc_load_update = calc_load_update;
4647 /* Update our root-domain */
4648 raw_spin_lock_irqsave(&rq->lock, flags);
4650 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4654 raw_spin_unlock_irqrestore(&rq->lock, flags);
4657 #ifdef CONFIG_HOTPLUG_CPU
4659 sched_ttwu_pending();
4660 /* Update our root-domain */
4661 raw_spin_lock_irqsave(&rq->lock, flags);
4663 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4667 BUG_ON(rq->nr_running != 1); /* the migration thread */
4668 raw_spin_unlock_irqrestore(&rq->lock, flags);
4672 calc_load_migrate(rq);
4677 update_max_interval();
4683 * Register at high priority so that task migration (migrate_all_tasks)
4684 * happens before everything else. This has to be lower priority than
4685 * the notifier in the perf_event subsystem, though.
4687 static struct notifier_block migration_notifier = {
4688 .notifier_call = migration_call,
4689 .priority = CPU_PRI_MIGRATION,
4692 static int sched_cpu_active(struct notifier_block *nfb,
4693 unsigned long action, void *hcpu)
4695 switch (action & ~CPU_TASKS_FROZEN) {
4697 case CPU_DOWN_FAILED:
4698 set_cpu_active((long)hcpu, true);
4705 static int sched_cpu_inactive(struct notifier_block *nfb,
4706 unsigned long action, void *hcpu)
4708 switch (action & ~CPU_TASKS_FROZEN) {
4709 case CPU_DOWN_PREPARE:
4710 set_cpu_active((long)hcpu, false);
4717 static int __init migration_init(void)
4719 void *cpu = (void *)(long)smp_processor_id();
4722 /* Initialize migration for the boot CPU */
4723 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4724 BUG_ON(err == NOTIFY_BAD);
4725 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4726 register_cpu_notifier(&migration_notifier);
4728 /* Register cpu active notifiers */
4729 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4730 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4734 early_initcall(migration_init);
4739 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4741 #ifdef CONFIG_SCHED_DEBUG
4743 static __read_mostly int sched_debug_enabled;
4745 static int __init sched_debug_setup(char *str)
4747 sched_debug_enabled = 1;
4751 early_param("sched_debug", sched_debug_setup);
4753 static inline bool sched_debug(void)
4755 return sched_debug_enabled;
4758 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4759 struct cpumask *groupmask)
4761 struct sched_group *group = sd->groups;
4764 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4765 cpumask_clear(groupmask);
4767 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4769 if (!(sd->flags & SD_LOAD_BALANCE)) {
4770 printk("does not load-balance\n");
4772 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4777 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4779 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4780 printk(KERN_ERR "ERROR: domain->span does not contain "
4783 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4784 printk(KERN_ERR "ERROR: domain->groups does not contain"
4788 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4792 printk(KERN_ERR "ERROR: group is NULL\n");
4797 * Even though we initialize ->power to something semi-sane,
4798 * we leave power_orig unset. This allows us to detect if
4799 * domain iteration is still funny without causing /0 traps.
4801 if (!group->sgp->power_orig) {
4802 printk(KERN_CONT "\n");
4803 printk(KERN_ERR "ERROR: domain->cpu_power not "
4808 if (!cpumask_weight(sched_group_cpus(group))) {
4809 printk(KERN_CONT "\n");
4810 printk(KERN_ERR "ERROR: empty group\n");
4814 if (!(sd->flags & SD_OVERLAP) &&
4815 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4816 printk(KERN_CONT "\n");
4817 printk(KERN_ERR "ERROR: repeated CPUs\n");
4821 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4823 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4825 printk(KERN_CONT " %s", str);
4826 if (group->sgp->power != SCHED_POWER_SCALE) {
4827 printk(KERN_CONT " (cpu_power = %d)",
4831 group = group->next;
4832 } while (group != sd->groups);
4833 printk(KERN_CONT "\n");
4835 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4836 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4839 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4840 printk(KERN_ERR "ERROR: parent span is not a superset "
4841 "of domain->span\n");
4845 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4849 if (!sched_debug_enabled)
4853 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4857 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4860 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4868 #else /* !CONFIG_SCHED_DEBUG */
4869 # define sched_domain_debug(sd, cpu) do { } while (0)
4870 static inline bool sched_debug(void)
4874 #endif /* CONFIG_SCHED_DEBUG */
4876 static int sd_degenerate(struct sched_domain *sd)
4878 if (cpumask_weight(sched_domain_span(sd)) == 1)
4881 /* Following flags need at least 2 groups */
4882 if (sd->flags & (SD_LOAD_BALANCE |
4883 SD_BALANCE_NEWIDLE |
4887 SD_SHARE_PKG_RESOURCES)) {
4888 if (sd->groups != sd->groups->next)
4892 /* Following flags don't use groups */
4893 if (sd->flags & (SD_WAKE_AFFINE))
4900 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4902 unsigned long cflags = sd->flags, pflags = parent->flags;
4904 if (sd_degenerate(parent))
4907 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4910 /* Flags needing groups don't count if only 1 group in parent */
4911 if (parent->groups == parent->groups->next) {
4912 pflags &= ~(SD_LOAD_BALANCE |
4913 SD_BALANCE_NEWIDLE |
4917 SD_SHARE_PKG_RESOURCES);
4918 if (nr_node_ids == 1)
4919 pflags &= ~SD_SERIALIZE;
4921 if (~cflags & pflags)
4927 static void free_rootdomain(struct rcu_head *rcu)
4929 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4931 cpupri_cleanup(&rd->cpupri);
4932 free_cpumask_var(rd->rto_mask);
4933 free_cpumask_var(rd->online);
4934 free_cpumask_var(rd->span);
4938 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4940 struct root_domain *old_rd = NULL;
4941 unsigned long flags;
4943 raw_spin_lock_irqsave(&rq->lock, flags);
4948 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4951 cpumask_clear_cpu(rq->cpu, old_rd->span);
4954 * If we dont want to free the old_rt yet then
4955 * set old_rd to NULL to skip the freeing later
4958 if (!atomic_dec_and_test(&old_rd->refcount))
4962 atomic_inc(&rd->refcount);
4965 cpumask_set_cpu(rq->cpu, rd->span);
4966 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4969 raw_spin_unlock_irqrestore(&rq->lock, flags);
4972 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4975 static int init_rootdomain(struct root_domain *rd)
4977 memset(rd, 0, sizeof(*rd));
4979 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4981 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4983 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4986 if (cpupri_init(&rd->cpupri) != 0)
4991 free_cpumask_var(rd->rto_mask);
4993 free_cpumask_var(rd->online);
4995 free_cpumask_var(rd->span);
5001 * By default the system creates a single root-domain with all cpus as
5002 * members (mimicking the global state we have today).
5004 struct root_domain def_root_domain;
5006 static void init_defrootdomain(void)
5008 init_rootdomain(&def_root_domain);
5010 atomic_set(&def_root_domain.refcount, 1);
5013 static struct root_domain *alloc_rootdomain(void)
5015 struct root_domain *rd;
5017 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5021 if (init_rootdomain(rd) != 0) {
5029 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5031 struct sched_group *tmp, *first;
5040 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5045 } while (sg != first);
5048 static void free_sched_domain(struct rcu_head *rcu)
5050 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5053 * If its an overlapping domain it has private groups, iterate and
5056 if (sd->flags & SD_OVERLAP) {
5057 free_sched_groups(sd->groups, 1);
5058 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5059 kfree(sd->groups->sgp);
5065 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5067 call_rcu(&sd->rcu, free_sched_domain);
5070 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5072 for (; sd; sd = sd->parent)
5073 destroy_sched_domain(sd, cpu);
5077 * Keep a special pointer to the highest sched_domain that has
5078 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5079 * allows us to avoid some pointer chasing select_idle_sibling().
5081 * Also keep a unique ID per domain (we use the first cpu number in
5082 * the cpumask of the domain), this allows us to quickly tell if
5083 * two cpus are in the same cache domain, see cpus_share_cache().
5085 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5086 DEFINE_PER_CPU(int, sd_llc_id);
5088 static void update_top_cache_domain(int cpu)
5090 struct sched_domain *sd;
5093 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5095 id = cpumask_first(sched_domain_span(sd));
5097 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5098 per_cpu(sd_llc_id, cpu) = id;
5102 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5103 * hold the hotplug lock.
5106 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5108 struct rq *rq = cpu_rq(cpu);
5109 struct sched_domain *tmp;
5111 /* Remove the sched domains which do not contribute to scheduling. */
5112 for (tmp = sd; tmp; ) {
5113 struct sched_domain *parent = tmp->parent;
5117 if (sd_parent_degenerate(tmp, parent)) {
5118 tmp->parent = parent->parent;
5120 parent->parent->child = tmp;
5121 destroy_sched_domain(parent, cpu);
5126 if (sd && sd_degenerate(sd)) {
5129 destroy_sched_domain(tmp, cpu);
5134 sched_domain_debug(sd, cpu);
5136 rq_attach_root(rq, rd);
5138 rcu_assign_pointer(rq->sd, sd);
5139 destroy_sched_domains(tmp, cpu);
5141 update_top_cache_domain(cpu);
5144 /* cpus with isolated domains */
5145 static cpumask_var_t cpu_isolated_map;
5147 /* Setup the mask of cpus configured for isolated domains */
5148 static int __init isolated_cpu_setup(char *str)
5150 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5151 cpulist_parse(str, cpu_isolated_map);
5155 __setup("isolcpus=", isolated_cpu_setup);
5157 static const struct cpumask *cpu_cpu_mask(int cpu)
5159 return cpumask_of_node(cpu_to_node(cpu));
5163 struct sched_domain **__percpu sd;
5164 struct sched_group **__percpu sg;
5165 struct sched_group_power **__percpu sgp;
5169 struct sched_domain ** __percpu sd;
5170 struct root_domain *rd;
5180 struct sched_domain_topology_level;
5182 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5183 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5185 #define SDTL_OVERLAP 0x01
5187 struct sched_domain_topology_level {
5188 sched_domain_init_f init;
5189 sched_domain_mask_f mask;
5192 struct sd_data data;
5196 * Build an iteration mask that can exclude certain CPUs from the upwards
5199 * Asymmetric node setups can result in situations where the domain tree is of
5200 * unequal depth, make sure to skip domains that already cover the entire
5203 * In that case build_sched_domains() will have terminated the iteration early
5204 * and our sibling sd spans will be empty. Domains should always include the
5205 * cpu they're built on, so check that.
5208 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5210 const struct cpumask *span = sched_domain_span(sd);
5211 struct sd_data *sdd = sd->private;
5212 struct sched_domain *sibling;
5215 for_each_cpu(i, span) {
5216 sibling = *per_cpu_ptr(sdd->sd, i);
5217 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5220 cpumask_set_cpu(i, sched_group_mask(sg));
5225 * Return the canonical balance cpu for this group, this is the first cpu
5226 * of this group that's also in the iteration mask.
5228 int group_balance_cpu(struct sched_group *sg)
5230 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5234 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5236 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5237 const struct cpumask *span = sched_domain_span(sd);
5238 struct cpumask *covered = sched_domains_tmpmask;
5239 struct sd_data *sdd = sd->private;
5240 struct sched_domain *child;
5243 cpumask_clear(covered);
5245 for_each_cpu(i, span) {
5246 struct cpumask *sg_span;
5248 if (cpumask_test_cpu(i, covered))
5251 child = *per_cpu_ptr(sdd->sd, i);
5253 /* See the comment near build_group_mask(). */
5254 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5257 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5258 GFP_KERNEL, cpu_to_node(cpu));
5263 sg_span = sched_group_cpus(sg);
5265 child = child->child;
5266 cpumask_copy(sg_span, sched_domain_span(child));
5268 cpumask_set_cpu(i, sg_span);
5270 cpumask_or(covered, covered, sg_span);
5272 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5273 if (atomic_inc_return(&sg->sgp->ref) == 1)
5274 build_group_mask(sd, sg);
5277 * Initialize sgp->power such that even if we mess up the
5278 * domains and no possible iteration will get us here, we won't
5281 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5284 * Make sure the first group of this domain contains the
5285 * canonical balance cpu. Otherwise the sched_domain iteration
5286 * breaks. See update_sg_lb_stats().
5288 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5289 group_balance_cpu(sg) == cpu)
5299 sd->groups = groups;
5304 free_sched_groups(first, 0);
5309 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5311 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5312 struct sched_domain *child = sd->child;
5315 cpu = cpumask_first(sched_domain_span(child));
5318 *sg = *per_cpu_ptr(sdd->sg, cpu);
5319 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5320 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5327 * build_sched_groups will build a circular linked list of the groups
5328 * covered by the given span, and will set each group's ->cpumask correctly,
5329 * and ->cpu_power to 0.
5331 * Assumes the sched_domain tree is fully constructed
5334 build_sched_groups(struct sched_domain *sd, int cpu)
5336 struct sched_group *first = NULL, *last = NULL;
5337 struct sd_data *sdd = sd->private;
5338 const struct cpumask *span = sched_domain_span(sd);
5339 struct cpumask *covered;
5342 get_group(cpu, sdd, &sd->groups);
5343 atomic_inc(&sd->groups->ref);
5345 if (cpu != cpumask_first(span))
5348 lockdep_assert_held(&sched_domains_mutex);
5349 covered = sched_domains_tmpmask;
5351 cpumask_clear(covered);
5353 for_each_cpu(i, span) {
5354 struct sched_group *sg;
5357 if (cpumask_test_cpu(i, covered))
5360 group = get_group(i, sdd, &sg);
5361 cpumask_clear(sched_group_cpus(sg));
5363 cpumask_setall(sched_group_mask(sg));
5365 for_each_cpu(j, span) {
5366 if (get_group(j, sdd, NULL) != group)
5369 cpumask_set_cpu(j, covered);
5370 cpumask_set_cpu(j, sched_group_cpus(sg));
5385 * Initialize sched groups cpu_power.
5387 * cpu_power indicates the capacity of sched group, which is used while
5388 * distributing the load between different sched groups in a sched domain.
5389 * Typically cpu_power for all the groups in a sched domain will be same unless
5390 * there are asymmetries in the topology. If there are asymmetries, group
5391 * having more cpu_power will pickup more load compared to the group having
5394 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5396 struct sched_group *sg = sd->groups;
5401 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5403 } while (sg != sd->groups);
5405 if (cpu != group_balance_cpu(sg))
5408 update_group_power(sd, cpu);
5409 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5412 int __weak arch_sd_sibling_asym_packing(void)
5414 return 0*SD_ASYM_PACKING;
5418 * Initializers for schedule domains
5419 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5422 #ifdef CONFIG_SCHED_DEBUG
5423 # define SD_INIT_NAME(sd, type) sd->name = #type
5425 # define SD_INIT_NAME(sd, type) do { } while (0)
5428 #define SD_INIT_FUNC(type) \
5429 static noinline struct sched_domain * \
5430 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5432 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5433 *sd = SD_##type##_INIT; \
5434 SD_INIT_NAME(sd, type); \
5435 sd->private = &tl->data; \
5440 #ifdef CONFIG_SCHED_SMT
5441 SD_INIT_FUNC(SIBLING)
5443 #ifdef CONFIG_SCHED_MC
5446 #ifdef CONFIG_SCHED_BOOK
5450 static int default_relax_domain_level = -1;
5451 int sched_domain_level_max;
5453 static int __init setup_relax_domain_level(char *str)
5455 if (kstrtoint(str, 0, &default_relax_domain_level))
5456 pr_warn("Unable to set relax_domain_level\n");
5460 __setup("relax_domain_level=", setup_relax_domain_level);
5462 static void set_domain_attribute(struct sched_domain *sd,
5463 struct sched_domain_attr *attr)
5467 if (!attr || attr->relax_domain_level < 0) {
5468 if (default_relax_domain_level < 0)
5471 request = default_relax_domain_level;
5473 request = attr->relax_domain_level;
5474 if (request < sd->level) {
5475 /* turn off idle balance on this domain */
5476 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5478 /* turn on idle balance on this domain */
5479 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5483 static void __sdt_free(const struct cpumask *cpu_map);
5484 static int __sdt_alloc(const struct cpumask *cpu_map);
5486 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5487 const struct cpumask *cpu_map)
5491 if (!atomic_read(&d->rd->refcount))
5492 free_rootdomain(&d->rd->rcu); /* fall through */
5494 free_percpu(d->sd); /* fall through */
5496 __sdt_free(cpu_map); /* fall through */
5502 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5503 const struct cpumask *cpu_map)
5505 memset(d, 0, sizeof(*d));
5507 if (__sdt_alloc(cpu_map))
5508 return sa_sd_storage;
5509 d->sd = alloc_percpu(struct sched_domain *);
5511 return sa_sd_storage;
5512 d->rd = alloc_rootdomain();
5515 return sa_rootdomain;
5519 * NULL the sd_data elements we've used to build the sched_domain and
5520 * sched_group structure so that the subsequent __free_domain_allocs()
5521 * will not free the data we're using.
5523 static void claim_allocations(int cpu, struct sched_domain *sd)
5525 struct sd_data *sdd = sd->private;
5527 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5528 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5530 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5531 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5533 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5534 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5537 #ifdef CONFIG_SCHED_SMT
5538 static const struct cpumask *cpu_smt_mask(int cpu)
5540 return topology_thread_cpumask(cpu);
5545 * Topology list, bottom-up.
5547 static struct sched_domain_topology_level default_topology[] = {
5548 #ifdef CONFIG_SCHED_SMT
5549 { sd_init_SIBLING, cpu_smt_mask, },
5551 #ifdef CONFIG_SCHED_MC
5552 { sd_init_MC, cpu_coregroup_mask, },
5554 #ifdef CONFIG_SCHED_BOOK
5555 { sd_init_BOOK, cpu_book_mask, },
5557 { sd_init_CPU, cpu_cpu_mask, },
5561 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5563 #define for_each_sd_topology(tl) \
5564 for (tl = sched_domain_topology; tl->init; tl++)
5568 static int sched_domains_numa_levels;
5569 static int *sched_domains_numa_distance;
5570 static struct cpumask ***sched_domains_numa_masks;
5571 static int sched_domains_curr_level;
5573 static inline int sd_local_flags(int level)
5575 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5578 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5581 static struct sched_domain *
5582 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5584 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5585 int level = tl->numa_level;
5586 int sd_weight = cpumask_weight(
5587 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5589 *sd = (struct sched_domain){
5590 .min_interval = sd_weight,
5591 .max_interval = 2*sd_weight,
5593 .imbalance_pct = 125,
5594 .cache_nice_tries = 2,
5601 .flags = 1*SD_LOAD_BALANCE
5602 | 1*SD_BALANCE_NEWIDLE
5607 | 0*SD_SHARE_CPUPOWER
5608 | 0*SD_SHARE_PKG_RESOURCES
5610 | 0*SD_PREFER_SIBLING
5611 | sd_local_flags(level)
5613 .last_balance = jiffies,
5614 .balance_interval = sd_weight,
5616 SD_INIT_NAME(sd, NUMA);
5617 sd->private = &tl->data;
5620 * Ugly hack to pass state to sd_numa_mask()...
5622 sched_domains_curr_level = tl->numa_level;
5627 static const struct cpumask *sd_numa_mask(int cpu)
5629 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5632 static void sched_numa_warn(const char *str)
5634 static int done = false;
5642 printk(KERN_WARNING "ERROR: %s\n\n", str);
5644 for (i = 0; i < nr_node_ids; i++) {
5645 printk(KERN_WARNING " ");
5646 for (j = 0; j < nr_node_ids; j++)
5647 printk(KERN_CONT "%02d ", node_distance(i,j));
5648 printk(KERN_CONT "\n");
5650 printk(KERN_WARNING "\n");
5653 static bool find_numa_distance(int distance)
5657 if (distance == node_distance(0, 0))
5660 for (i = 0; i < sched_domains_numa_levels; i++) {
5661 if (sched_domains_numa_distance[i] == distance)
5668 static void sched_init_numa(void)
5670 int next_distance, curr_distance = node_distance(0, 0);
5671 struct sched_domain_topology_level *tl;
5675 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5676 if (!sched_domains_numa_distance)
5680 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5681 * unique distances in the node_distance() table.
5683 * Assumes node_distance(0,j) includes all distances in
5684 * node_distance(i,j) in order to avoid cubic time.
5686 next_distance = curr_distance;
5687 for (i = 0; i < nr_node_ids; i++) {
5688 for (j = 0; j < nr_node_ids; j++) {
5689 for (k = 0; k < nr_node_ids; k++) {
5690 int distance = node_distance(i, k);
5692 if (distance > curr_distance &&
5693 (distance < next_distance ||
5694 next_distance == curr_distance))
5695 next_distance = distance;
5698 * While not a strong assumption it would be nice to know
5699 * about cases where if node A is connected to B, B is not
5700 * equally connected to A.
5702 if (sched_debug() && node_distance(k, i) != distance)
5703 sched_numa_warn("Node-distance not symmetric");
5705 if (sched_debug() && i && !find_numa_distance(distance))
5706 sched_numa_warn("Node-0 not representative");
5708 if (next_distance != curr_distance) {
5709 sched_domains_numa_distance[level++] = next_distance;
5710 sched_domains_numa_levels = level;
5711 curr_distance = next_distance;
5716 * In case of sched_debug() we verify the above assumption.
5722 * 'level' contains the number of unique distances, excluding the
5723 * identity distance node_distance(i,i).
5725 * The sched_domains_numa_distance[] array includes the actual distance
5730 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5731 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5732 * the array will contain less then 'level' members. This could be
5733 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5734 * in other functions.
5736 * We reset it to 'level' at the end of this function.
5738 sched_domains_numa_levels = 0;
5740 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5741 if (!sched_domains_numa_masks)
5745 * Now for each level, construct a mask per node which contains all
5746 * cpus of nodes that are that many hops away from us.
5748 for (i = 0; i < level; i++) {
5749 sched_domains_numa_masks[i] =
5750 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5751 if (!sched_domains_numa_masks[i])
5754 for (j = 0; j < nr_node_ids; j++) {
5755 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5759 sched_domains_numa_masks[i][j] = mask;
5761 for (k = 0; k < nr_node_ids; k++) {
5762 if (node_distance(j, k) > sched_domains_numa_distance[i])
5765 cpumask_or(mask, mask, cpumask_of_node(k));
5770 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5771 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5776 * Copy the default topology bits..
5778 for (i = 0; default_topology[i].init; i++)
5779 tl[i] = default_topology[i];
5782 * .. and append 'j' levels of NUMA goodness.
5784 for (j = 0; j < level; i++, j++) {
5785 tl[i] = (struct sched_domain_topology_level){
5786 .init = sd_numa_init,
5787 .mask = sd_numa_mask,
5788 .flags = SDTL_OVERLAP,
5793 sched_domain_topology = tl;
5795 sched_domains_numa_levels = level;
5798 static void sched_domains_numa_masks_set(int cpu)
5801 int node = cpu_to_node(cpu);
5803 for (i = 0; i < sched_domains_numa_levels; i++) {
5804 for (j = 0; j < nr_node_ids; j++) {
5805 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5806 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5811 static void sched_domains_numa_masks_clear(int cpu)
5814 for (i = 0; i < sched_domains_numa_levels; i++) {
5815 for (j = 0; j < nr_node_ids; j++)
5816 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5821 * Update sched_domains_numa_masks[level][node] array when new cpus
5824 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5825 unsigned long action,
5828 int cpu = (long)hcpu;
5830 switch (action & ~CPU_TASKS_FROZEN) {
5832 sched_domains_numa_masks_set(cpu);
5836 sched_domains_numa_masks_clear(cpu);
5846 static inline void sched_init_numa(void)
5850 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5851 unsigned long action,
5856 #endif /* CONFIG_NUMA */
5858 static int __sdt_alloc(const struct cpumask *cpu_map)
5860 struct sched_domain_topology_level *tl;
5863 for_each_sd_topology(tl) {
5864 struct sd_data *sdd = &tl->data;
5866 sdd->sd = alloc_percpu(struct sched_domain *);
5870 sdd->sg = alloc_percpu(struct sched_group *);
5874 sdd->sgp = alloc_percpu(struct sched_group_power *);
5878 for_each_cpu(j, cpu_map) {
5879 struct sched_domain *sd;
5880 struct sched_group *sg;
5881 struct sched_group_power *sgp;
5883 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5884 GFP_KERNEL, cpu_to_node(j));
5888 *per_cpu_ptr(sdd->sd, j) = sd;
5890 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5891 GFP_KERNEL, cpu_to_node(j));
5897 *per_cpu_ptr(sdd->sg, j) = sg;
5899 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5900 GFP_KERNEL, cpu_to_node(j));
5904 *per_cpu_ptr(sdd->sgp, j) = sgp;
5911 static void __sdt_free(const struct cpumask *cpu_map)
5913 struct sched_domain_topology_level *tl;
5916 for_each_sd_topology(tl) {
5917 struct sd_data *sdd = &tl->data;
5919 for_each_cpu(j, cpu_map) {
5920 struct sched_domain *sd;
5923 sd = *per_cpu_ptr(sdd->sd, j);
5924 if (sd && (sd->flags & SD_OVERLAP))
5925 free_sched_groups(sd->groups, 0);
5926 kfree(*per_cpu_ptr(sdd->sd, j));
5930 kfree(*per_cpu_ptr(sdd->sg, j));
5932 kfree(*per_cpu_ptr(sdd->sgp, j));
5934 free_percpu(sdd->sd);
5936 free_percpu(sdd->sg);
5938 free_percpu(sdd->sgp);
5943 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5944 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5945 struct sched_domain *child, int cpu)
5947 struct sched_domain *sd = tl->init(tl, cpu);
5951 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5953 sd->level = child->level + 1;
5954 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5958 set_domain_attribute(sd, attr);
5964 * Build sched domains for a given set of cpus and attach the sched domains
5965 * to the individual cpus
5967 static int build_sched_domains(const struct cpumask *cpu_map,
5968 struct sched_domain_attr *attr)
5970 enum s_alloc alloc_state;
5971 struct sched_domain *sd;
5973 int i, ret = -ENOMEM;
5975 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5976 if (alloc_state != sa_rootdomain)
5979 /* Set up domains for cpus specified by the cpu_map. */
5980 for_each_cpu(i, cpu_map) {
5981 struct sched_domain_topology_level *tl;
5984 for_each_sd_topology(tl) {
5985 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
5986 if (tl == sched_domain_topology)
5987 *per_cpu_ptr(d.sd, i) = sd;
5988 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
5989 sd->flags |= SD_OVERLAP;
5990 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
5995 /* Build the groups for the domains */
5996 for_each_cpu(i, cpu_map) {
5997 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5998 sd->span_weight = cpumask_weight(sched_domain_span(sd));
5999 if (sd->flags & SD_OVERLAP) {
6000 if (build_overlap_sched_groups(sd, i))
6003 if (build_sched_groups(sd, i))
6009 /* Calculate CPU power for physical packages and nodes */
6010 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6011 if (!cpumask_test_cpu(i, cpu_map))
6014 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6015 claim_allocations(i, sd);
6016 init_sched_groups_power(i, sd);
6020 /* Attach the domains */
6022 for_each_cpu(i, cpu_map) {
6023 sd = *per_cpu_ptr(d.sd, i);
6024 cpu_attach_domain(sd, d.rd, i);
6030 __free_domain_allocs(&d, alloc_state, cpu_map);
6034 static cpumask_var_t *doms_cur; /* current sched domains */
6035 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6036 static struct sched_domain_attr *dattr_cur;
6037 /* attribues of custom domains in 'doms_cur' */
6040 * Special case: If a kmalloc of a doms_cur partition (array of
6041 * cpumask) fails, then fallback to a single sched domain,
6042 * as determined by the single cpumask fallback_doms.
6044 static cpumask_var_t fallback_doms;
6047 * arch_update_cpu_topology lets virtualized architectures update the
6048 * cpu core maps. It is supposed to return 1 if the topology changed
6049 * or 0 if it stayed the same.
6051 int __attribute__((weak)) arch_update_cpu_topology(void)
6056 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6059 cpumask_var_t *doms;
6061 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6064 for (i = 0; i < ndoms; i++) {
6065 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6066 free_sched_domains(doms, i);
6073 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6076 for (i = 0; i < ndoms; i++)
6077 free_cpumask_var(doms[i]);
6082 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6083 * For now this just excludes isolated cpus, but could be used to
6084 * exclude other special cases in the future.
6086 static int init_sched_domains(const struct cpumask *cpu_map)
6090 arch_update_cpu_topology();
6092 doms_cur = alloc_sched_domains(ndoms_cur);
6094 doms_cur = &fallback_doms;
6095 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6096 err = build_sched_domains(doms_cur[0], NULL);
6097 register_sched_domain_sysctl();
6103 * Detach sched domains from a group of cpus specified in cpu_map
6104 * These cpus will now be attached to the NULL domain
6106 static void detach_destroy_domains(const struct cpumask *cpu_map)
6111 for_each_cpu(i, cpu_map)
6112 cpu_attach_domain(NULL, &def_root_domain, i);
6116 /* handle null as "default" */
6117 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6118 struct sched_domain_attr *new, int idx_new)
6120 struct sched_domain_attr tmp;
6127 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6128 new ? (new + idx_new) : &tmp,
6129 sizeof(struct sched_domain_attr));
6133 * Partition sched domains as specified by the 'ndoms_new'
6134 * cpumasks in the array doms_new[] of cpumasks. This compares
6135 * doms_new[] to the current sched domain partitioning, doms_cur[].
6136 * It destroys each deleted domain and builds each new domain.
6138 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6139 * The masks don't intersect (don't overlap.) We should setup one
6140 * sched domain for each mask. CPUs not in any of the cpumasks will
6141 * not be load balanced. If the same cpumask appears both in the
6142 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6145 * The passed in 'doms_new' should be allocated using
6146 * alloc_sched_domains. This routine takes ownership of it and will
6147 * free_sched_domains it when done with it. If the caller failed the
6148 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6149 * and partition_sched_domains() will fallback to the single partition
6150 * 'fallback_doms', it also forces the domains to be rebuilt.
6152 * If doms_new == NULL it will be replaced with cpu_online_mask.
6153 * ndoms_new == 0 is a special case for destroying existing domains,
6154 * and it will not create the default domain.
6156 * Call with hotplug lock held
6158 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6159 struct sched_domain_attr *dattr_new)
6164 mutex_lock(&sched_domains_mutex);
6166 /* always unregister in case we don't destroy any domains */
6167 unregister_sched_domain_sysctl();
6169 /* Let architecture update cpu core mappings. */
6170 new_topology = arch_update_cpu_topology();
6172 n = doms_new ? ndoms_new : 0;
6174 /* Destroy deleted domains */
6175 for (i = 0; i < ndoms_cur; i++) {
6176 for (j = 0; j < n && !new_topology; j++) {
6177 if (cpumask_equal(doms_cur[i], doms_new[j])
6178 && dattrs_equal(dattr_cur, i, dattr_new, j))
6181 /* no match - a current sched domain not in new doms_new[] */
6182 detach_destroy_domains(doms_cur[i]);
6187 if (doms_new == NULL) {
6189 doms_new = &fallback_doms;
6190 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6191 WARN_ON_ONCE(dattr_new);
6194 /* Build new domains */
6195 for (i = 0; i < ndoms_new; i++) {
6196 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6197 if (cpumask_equal(doms_new[i], doms_cur[j])
6198 && dattrs_equal(dattr_new, i, dattr_cur, j))
6201 /* no match - add a new doms_new */
6202 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6207 /* Remember the new sched domains */
6208 if (doms_cur != &fallback_doms)
6209 free_sched_domains(doms_cur, ndoms_cur);
6210 kfree(dattr_cur); /* kfree(NULL) is safe */
6211 doms_cur = doms_new;
6212 dattr_cur = dattr_new;
6213 ndoms_cur = ndoms_new;
6215 register_sched_domain_sysctl();
6217 mutex_unlock(&sched_domains_mutex);
6220 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6223 * Update cpusets according to cpu_active mask. If cpusets are
6224 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6225 * around partition_sched_domains().
6227 * If we come here as part of a suspend/resume, don't touch cpusets because we
6228 * want to restore it back to its original state upon resume anyway.
6230 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6234 case CPU_ONLINE_FROZEN:
6235 case CPU_DOWN_FAILED_FROZEN:
6238 * num_cpus_frozen tracks how many CPUs are involved in suspend
6239 * resume sequence. As long as this is not the last online
6240 * operation in the resume sequence, just build a single sched
6241 * domain, ignoring cpusets.
6244 if (likely(num_cpus_frozen)) {
6245 partition_sched_domains(1, NULL, NULL);
6250 * This is the last CPU online operation. So fall through and
6251 * restore the original sched domains by considering the
6252 * cpuset configurations.
6256 case CPU_DOWN_FAILED:
6257 cpuset_update_active_cpus(true);
6265 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6269 case CPU_DOWN_PREPARE:
6270 cpuset_update_active_cpus(false);
6272 case CPU_DOWN_PREPARE_FROZEN:
6274 partition_sched_domains(1, NULL, NULL);
6282 void __init sched_init_smp(void)
6284 cpumask_var_t non_isolated_cpus;
6286 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6287 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6292 mutex_lock(&sched_domains_mutex);
6293 init_sched_domains(cpu_active_mask);
6294 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6295 if (cpumask_empty(non_isolated_cpus))
6296 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6297 mutex_unlock(&sched_domains_mutex);
6300 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6301 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6302 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6306 /* Move init over to a non-isolated CPU */
6307 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6309 sched_init_granularity();
6310 free_cpumask_var(non_isolated_cpus);
6312 init_sched_rt_class();
6315 void __init sched_init_smp(void)
6317 sched_init_granularity();
6319 #endif /* CONFIG_SMP */
6321 const_debug unsigned int sysctl_timer_migration = 1;
6323 int in_sched_functions(unsigned long addr)
6325 return in_lock_functions(addr) ||
6326 (addr >= (unsigned long)__sched_text_start
6327 && addr < (unsigned long)__sched_text_end);
6330 #ifdef CONFIG_CGROUP_SCHED
6332 * Default task group.
6333 * Every task in system belongs to this group at bootup.
6335 struct task_group root_task_group;
6336 LIST_HEAD(task_groups);
6339 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6341 void __init sched_init(void)
6344 unsigned long alloc_size = 0, ptr;
6346 #ifdef CONFIG_FAIR_GROUP_SCHED
6347 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6349 #ifdef CONFIG_RT_GROUP_SCHED
6350 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6352 #ifdef CONFIG_CPUMASK_OFFSTACK
6353 alloc_size += num_possible_cpus() * cpumask_size();
6356 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6358 #ifdef CONFIG_FAIR_GROUP_SCHED
6359 root_task_group.se = (struct sched_entity **)ptr;
6360 ptr += nr_cpu_ids * sizeof(void **);
6362 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6363 ptr += nr_cpu_ids * sizeof(void **);
6365 #endif /* CONFIG_FAIR_GROUP_SCHED */
6366 #ifdef CONFIG_RT_GROUP_SCHED
6367 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6368 ptr += nr_cpu_ids * sizeof(void **);
6370 root_task_group.rt_rq = (struct rt_rq **)ptr;
6371 ptr += nr_cpu_ids * sizeof(void **);
6373 #endif /* CONFIG_RT_GROUP_SCHED */
6374 #ifdef CONFIG_CPUMASK_OFFSTACK
6375 for_each_possible_cpu(i) {
6376 per_cpu(load_balance_mask, i) = (void *)ptr;
6377 ptr += cpumask_size();
6379 #endif /* CONFIG_CPUMASK_OFFSTACK */
6383 init_defrootdomain();
6386 init_rt_bandwidth(&def_rt_bandwidth,
6387 global_rt_period(), global_rt_runtime());
6389 #ifdef CONFIG_RT_GROUP_SCHED
6390 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6391 global_rt_period(), global_rt_runtime());
6392 #endif /* CONFIG_RT_GROUP_SCHED */
6394 #ifdef CONFIG_CGROUP_SCHED
6395 list_add(&root_task_group.list, &task_groups);
6396 INIT_LIST_HEAD(&root_task_group.children);
6397 INIT_LIST_HEAD(&root_task_group.siblings);
6398 autogroup_init(&init_task);
6400 #endif /* CONFIG_CGROUP_SCHED */
6402 for_each_possible_cpu(i) {
6406 raw_spin_lock_init(&rq->lock);
6408 rq->calc_load_active = 0;
6409 rq->calc_load_update = jiffies + LOAD_FREQ;
6410 init_cfs_rq(&rq->cfs);
6411 init_rt_rq(&rq->rt, rq);
6412 #ifdef CONFIG_FAIR_GROUP_SCHED
6413 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6414 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6416 * How much cpu bandwidth does root_task_group get?
6418 * In case of task-groups formed thr' the cgroup filesystem, it
6419 * gets 100% of the cpu resources in the system. This overall
6420 * system cpu resource is divided among the tasks of
6421 * root_task_group and its child task-groups in a fair manner,
6422 * based on each entity's (task or task-group's) weight
6423 * (se->load.weight).
6425 * In other words, if root_task_group has 10 tasks of weight
6426 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6427 * then A0's share of the cpu resource is:
6429 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6431 * We achieve this by letting root_task_group's tasks sit
6432 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6434 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6435 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6436 #endif /* CONFIG_FAIR_GROUP_SCHED */
6438 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6439 #ifdef CONFIG_RT_GROUP_SCHED
6440 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6441 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6444 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6445 rq->cpu_load[j] = 0;
6447 rq->last_load_update_tick = jiffies;
6452 rq->cpu_power = SCHED_POWER_SCALE;
6453 rq->post_schedule = 0;
6454 rq->active_balance = 0;
6455 rq->next_balance = jiffies;
6460 rq->avg_idle = 2*sysctl_sched_migration_cost;
6462 INIT_LIST_HEAD(&rq->cfs_tasks);
6464 rq_attach_root(rq, &def_root_domain);
6465 #ifdef CONFIG_NO_HZ_COMMON
6468 #ifdef CONFIG_NO_HZ_FULL
6469 rq->last_sched_tick = 0;
6473 atomic_set(&rq->nr_iowait, 0);
6476 set_load_weight(&init_task);
6478 #ifdef CONFIG_PREEMPT_NOTIFIERS
6479 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6482 #ifdef CONFIG_RT_MUTEXES
6483 plist_head_init(&init_task.pi_waiters);
6487 * The boot idle thread does lazy MMU switching as well:
6489 atomic_inc(&init_mm.mm_count);
6490 enter_lazy_tlb(&init_mm, current);
6493 * Make us the idle thread. Technically, schedule() should not be
6494 * called from this thread, however somewhere below it might be,
6495 * but because we are the idle thread, we just pick up running again
6496 * when this runqueue becomes "idle".
6498 init_idle(current, smp_processor_id());
6500 calc_load_update = jiffies + LOAD_FREQ;
6503 * During early bootup we pretend to be a normal task:
6505 current->sched_class = &fair_sched_class;
6508 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6509 /* May be allocated at isolcpus cmdline parse time */
6510 if (cpu_isolated_map == NULL)
6511 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6512 idle_thread_set_boot_cpu();
6514 init_sched_fair_class();
6516 scheduler_running = 1;
6519 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6520 static inline int preempt_count_equals(int preempt_offset)
6522 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6524 return (nested == preempt_offset);
6527 void __might_sleep(const char *file, int line, int preempt_offset)
6529 static unsigned long prev_jiffy; /* ratelimiting */
6531 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6532 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6533 system_state != SYSTEM_RUNNING || oops_in_progress)
6535 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6537 prev_jiffy = jiffies;
6540 "BUG: sleeping function called from invalid context at %s:%d\n",
6543 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6544 in_atomic(), irqs_disabled(),
6545 current->pid, current->comm);
6547 debug_show_held_locks(current);
6548 if (irqs_disabled())
6549 print_irqtrace_events(current);
6552 EXPORT_SYMBOL(__might_sleep);
6555 #ifdef CONFIG_MAGIC_SYSRQ
6556 static void normalize_task(struct rq *rq, struct task_struct *p)
6558 const struct sched_class *prev_class = p->sched_class;
6559 int old_prio = p->prio;
6564 dequeue_task(rq, p, 0);
6565 __setscheduler(rq, p, SCHED_NORMAL, 0);
6567 enqueue_task(rq, p, 0);
6568 resched_task(rq->curr);
6571 check_class_changed(rq, p, prev_class, old_prio);
6574 void normalize_rt_tasks(void)
6576 struct task_struct *g, *p;
6577 unsigned long flags;
6580 read_lock_irqsave(&tasklist_lock, flags);
6581 do_each_thread(g, p) {
6583 * Only normalize user tasks:
6588 p->se.exec_start = 0;
6589 #ifdef CONFIG_SCHEDSTATS
6590 p->se.statistics.wait_start = 0;
6591 p->se.statistics.sleep_start = 0;
6592 p->se.statistics.block_start = 0;
6597 * Renice negative nice level userspace
6600 if (TASK_NICE(p) < 0 && p->mm)
6601 set_user_nice(p, 0);
6605 raw_spin_lock(&p->pi_lock);
6606 rq = __task_rq_lock(p);
6608 normalize_task(rq, p);
6610 __task_rq_unlock(rq);
6611 raw_spin_unlock(&p->pi_lock);
6612 } while_each_thread(g, p);
6614 read_unlock_irqrestore(&tasklist_lock, flags);
6617 #endif /* CONFIG_MAGIC_SYSRQ */
6619 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6621 * These functions are only useful for the IA64 MCA handling, or kdb.
6623 * They can only be called when the whole system has been
6624 * stopped - every CPU needs to be quiescent, and no scheduling
6625 * activity can take place. Using them for anything else would
6626 * be a serious bug, and as a result, they aren't even visible
6627 * under any other configuration.
6631 * curr_task - return the current task for a given cpu.
6632 * @cpu: the processor in question.
6634 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6636 struct task_struct *curr_task(int cpu)
6638 return cpu_curr(cpu);
6641 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6645 * set_curr_task - set the current task for a given cpu.
6646 * @cpu: the processor in question.
6647 * @p: the task pointer to set.
6649 * Description: This function must only be used when non-maskable interrupts
6650 * are serviced on a separate stack. It allows the architecture to switch the
6651 * notion of the current task on a cpu in a non-blocking manner. This function
6652 * must be called with all CPU's synchronized, and interrupts disabled, the
6653 * and caller must save the original value of the current task (see
6654 * curr_task() above) and restore that value before reenabling interrupts and
6655 * re-starting the system.
6657 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6659 void set_curr_task(int cpu, struct task_struct *p)
6666 #ifdef CONFIG_CGROUP_SCHED
6667 /* task_group_lock serializes the addition/removal of task groups */
6668 static DEFINE_SPINLOCK(task_group_lock);
6670 static void free_sched_group(struct task_group *tg)
6672 free_fair_sched_group(tg);
6673 free_rt_sched_group(tg);
6678 /* allocate runqueue etc for a new task group */
6679 struct task_group *sched_create_group(struct task_group *parent)
6681 struct task_group *tg;
6683 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6685 return ERR_PTR(-ENOMEM);
6687 if (!alloc_fair_sched_group(tg, parent))
6690 if (!alloc_rt_sched_group(tg, parent))
6696 free_sched_group(tg);
6697 return ERR_PTR(-ENOMEM);
6700 void sched_online_group(struct task_group *tg, struct task_group *parent)
6702 unsigned long flags;
6704 spin_lock_irqsave(&task_group_lock, flags);
6705 list_add_rcu(&tg->list, &task_groups);
6707 WARN_ON(!parent); /* root should already exist */
6709 tg->parent = parent;
6710 INIT_LIST_HEAD(&tg->children);
6711 list_add_rcu(&tg->siblings, &parent->children);
6712 spin_unlock_irqrestore(&task_group_lock, flags);
6715 /* rcu callback to free various structures associated with a task group */
6716 static void free_sched_group_rcu(struct rcu_head *rhp)
6718 /* now it should be safe to free those cfs_rqs */
6719 free_sched_group(container_of(rhp, struct task_group, rcu));
6722 /* Destroy runqueue etc associated with a task group */
6723 void sched_destroy_group(struct task_group *tg)
6725 /* wait for possible concurrent references to cfs_rqs complete */
6726 call_rcu(&tg->rcu, free_sched_group_rcu);
6729 void sched_offline_group(struct task_group *tg)
6731 unsigned long flags;
6734 /* end participation in shares distribution */
6735 for_each_possible_cpu(i)
6736 unregister_fair_sched_group(tg, i);
6738 spin_lock_irqsave(&task_group_lock, flags);
6739 list_del_rcu(&tg->list);
6740 list_del_rcu(&tg->siblings);
6741 spin_unlock_irqrestore(&task_group_lock, flags);
6744 /* change task's runqueue when it moves between groups.
6745 * The caller of this function should have put the task in its new group
6746 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6747 * reflect its new group.
6749 void sched_move_task(struct task_struct *tsk)
6751 struct task_group *tg;
6753 unsigned long flags;
6756 rq = task_rq_lock(tsk, &flags);
6758 running = task_current(rq, tsk);
6762 dequeue_task(rq, tsk, 0);
6763 if (unlikely(running))
6764 tsk->sched_class->put_prev_task(rq, tsk);
6766 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
6767 lockdep_is_held(&tsk->sighand->siglock)),
6768 struct task_group, css);
6769 tg = autogroup_task_group(tsk, tg);
6770 tsk->sched_task_group = tg;
6772 #ifdef CONFIG_FAIR_GROUP_SCHED
6773 if (tsk->sched_class->task_move_group)
6774 tsk->sched_class->task_move_group(tsk, on_rq);
6777 set_task_rq(tsk, task_cpu(tsk));
6779 if (unlikely(running))
6780 tsk->sched_class->set_curr_task(rq);
6782 enqueue_task(rq, tsk, 0);
6784 task_rq_unlock(rq, tsk, &flags);
6786 #endif /* CONFIG_CGROUP_SCHED */
6788 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6789 static unsigned long to_ratio(u64 period, u64 runtime)
6791 if (runtime == RUNTIME_INF)
6794 return div64_u64(runtime << 20, period);
6798 #ifdef CONFIG_RT_GROUP_SCHED
6800 * Ensure that the real time constraints are schedulable.
6802 static DEFINE_MUTEX(rt_constraints_mutex);
6804 /* Must be called with tasklist_lock held */
6805 static inline int tg_has_rt_tasks(struct task_group *tg)
6807 struct task_struct *g, *p;
6809 do_each_thread(g, p) {
6810 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6812 } while_each_thread(g, p);
6817 struct rt_schedulable_data {
6818 struct task_group *tg;
6823 static int tg_rt_schedulable(struct task_group *tg, void *data)
6825 struct rt_schedulable_data *d = data;
6826 struct task_group *child;
6827 unsigned long total, sum = 0;
6828 u64 period, runtime;
6830 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6831 runtime = tg->rt_bandwidth.rt_runtime;
6834 period = d->rt_period;
6835 runtime = d->rt_runtime;
6839 * Cannot have more runtime than the period.
6841 if (runtime > period && runtime != RUNTIME_INF)
6845 * Ensure we don't starve existing RT tasks.
6847 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6850 total = to_ratio(period, runtime);
6853 * Nobody can have more than the global setting allows.
6855 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6859 * The sum of our children's runtime should not exceed our own.
6861 list_for_each_entry_rcu(child, &tg->children, siblings) {
6862 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6863 runtime = child->rt_bandwidth.rt_runtime;
6865 if (child == d->tg) {
6866 period = d->rt_period;
6867 runtime = d->rt_runtime;
6870 sum += to_ratio(period, runtime);
6879 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6883 struct rt_schedulable_data data = {
6885 .rt_period = period,
6886 .rt_runtime = runtime,
6890 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6896 static int tg_set_rt_bandwidth(struct task_group *tg,
6897 u64 rt_period, u64 rt_runtime)
6901 mutex_lock(&rt_constraints_mutex);
6902 read_lock(&tasklist_lock);
6903 err = __rt_schedulable(tg, rt_period, rt_runtime);
6907 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6908 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6909 tg->rt_bandwidth.rt_runtime = rt_runtime;
6911 for_each_possible_cpu(i) {
6912 struct rt_rq *rt_rq = tg->rt_rq[i];
6914 raw_spin_lock(&rt_rq->rt_runtime_lock);
6915 rt_rq->rt_runtime = rt_runtime;
6916 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6918 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6920 read_unlock(&tasklist_lock);
6921 mutex_unlock(&rt_constraints_mutex);
6926 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6928 u64 rt_runtime, rt_period;
6930 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6931 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6932 if (rt_runtime_us < 0)
6933 rt_runtime = RUNTIME_INF;
6935 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6938 static long sched_group_rt_runtime(struct task_group *tg)
6942 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6945 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6946 do_div(rt_runtime_us, NSEC_PER_USEC);
6947 return rt_runtime_us;
6950 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6952 u64 rt_runtime, rt_period;
6954 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6955 rt_runtime = tg->rt_bandwidth.rt_runtime;
6960 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6963 static long sched_group_rt_period(struct task_group *tg)
6967 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6968 do_div(rt_period_us, NSEC_PER_USEC);
6969 return rt_period_us;
6972 static int sched_rt_global_constraints(void)
6974 u64 runtime, period;
6977 if (sysctl_sched_rt_period <= 0)
6980 runtime = global_rt_runtime();
6981 period = global_rt_period();
6984 * Sanity check on the sysctl variables.
6986 if (runtime > period && runtime != RUNTIME_INF)
6989 mutex_lock(&rt_constraints_mutex);
6990 read_lock(&tasklist_lock);
6991 ret = __rt_schedulable(NULL, 0, 0);
6992 read_unlock(&tasklist_lock);
6993 mutex_unlock(&rt_constraints_mutex);
6998 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7000 /* Don't accept realtime tasks when there is no way for them to run */
7001 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7007 #else /* !CONFIG_RT_GROUP_SCHED */
7008 static int sched_rt_global_constraints(void)
7010 unsigned long flags;
7013 if (sysctl_sched_rt_period <= 0)
7017 * There's always some RT tasks in the root group
7018 * -- migration, kstopmachine etc..
7020 if (sysctl_sched_rt_runtime == 0)
7023 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7024 for_each_possible_cpu(i) {
7025 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7027 raw_spin_lock(&rt_rq->rt_runtime_lock);
7028 rt_rq->rt_runtime = global_rt_runtime();
7029 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7031 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7035 #endif /* CONFIG_RT_GROUP_SCHED */
7037 int sched_rr_handler(struct ctl_table *table, int write,
7038 void __user *buffer, size_t *lenp,
7042 static DEFINE_MUTEX(mutex);
7045 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7046 /* make sure that internally we keep jiffies */
7047 /* also, writing zero resets timeslice to default */
7048 if (!ret && write) {
7049 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7050 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7052 mutex_unlock(&mutex);
7056 int sched_rt_handler(struct ctl_table *table, int write,
7057 void __user *buffer, size_t *lenp,
7061 int old_period, old_runtime;
7062 static DEFINE_MUTEX(mutex);
7065 old_period = sysctl_sched_rt_period;
7066 old_runtime = sysctl_sched_rt_runtime;
7068 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7070 if (!ret && write) {
7071 ret = sched_rt_global_constraints();
7073 sysctl_sched_rt_period = old_period;
7074 sysctl_sched_rt_runtime = old_runtime;
7076 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7077 def_rt_bandwidth.rt_period =
7078 ns_to_ktime(global_rt_period());
7081 mutex_unlock(&mutex);
7086 #ifdef CONFIG_CGROUP_SCHED
7088 /* return corresponding task_group object of a cgroup */
7089 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7091 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7092 struct task_group, css);
7095 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7097 struct task_group *tg, *parent;
7099 if (!cgrp->parent) {
7100 /* This is early initialization for the top cgroup */
7101 return &root_task_group.css;
7104 parent = cgroup_tg(cgrp->parent);
7105 tg = sched_create_group(parent);
7107 return ERR_PTR(-ENOMEM);
7112 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7114 struct task_group *tg = cgroup_tg(cgrp);
7115 struct task_group *parent;
7120 parent = cgroup_tg(cgrp->parent);
7121 sched_online_group(tg, parent);
7125 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7127 struct task_group *tg = cgroup_tg(cgrp);
7129 sched_destroy_group(tg);
7132 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7134 struct task_group *tg = cgroup_tg(cgrp);
7136 sched_offline_group(tg);
7139 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7140 struct cgroup_taskset *tset)
7142 struct task_struct *task;
7144 cgroup_taskset_for_each(task, cgrp, tset) {
7145 #ifdef CONFIG_RT_GROUP_SCHED
7146 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7149 /* We don't support RT-tasks being in separate groups */
7150 if (task->sched_class != &fair_sched_class)
7157 static void cpu_cgroup_attach(struct cgroup *cgrp,
7158 struct cgroup_taskset *tset)
7160 struct task_struct *task;
7162 cgroup_taskset_for_each(task, cgrp, tset)
7163 sched_move_task(task);
7167 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7168 struct task_struct *task)
7171 * cgroup_exit() is called in the copy_process() failure path.
7172 * Ignore this case since the task hasn't ran yet, this avoids
7173 * trying to poke a half freed task state from generic code.
7175 if (!(task->flags & PF_EXITING))
7178 sched_move_task(task);
7181 #ifdef CONFIG_FAIR_GROUP_SCHED
7182 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7185 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7188 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7190 struct task_group *tg = cgroup_tg(cgrp);
7192 return (u64) scale_load_down(tg->shares);
7195 #ifdef CONFIG_CFS_BANDWIDTH
7196 static DEFINE_MUTEX(cfs_constraints_mutex);
7198 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7199 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7201 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7203 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7205 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7206 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7208 if (tg == &root_task_group)
7212 * Ensure we have at some amount of bandwidth every period. This is
7213 * to prevent reaching a state of large arrears when throttled via
7214 * entity_tick() resulting in prolonged exit starvation.
7216 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7220 * Likewise, bound things on the otherside by preventing insane quota
7221 * periods. This also allows us to normalize in computing quota
7224 if (period > max_cfs_quota_period)
7227 mutex_lock(&cfs_constraints_mutex);
7228 ret = __cfs_schedulable(tg, period, quota);
7232 runtime_enabled = quota != RUNTIME_INF;
7233 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7234 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7235 raw_spin_lock_irq(&cfs_b->lock);
7236 cfs_b->period = ns_to_ktime(period);
7237 cfs_b->quota = quota;
7239 __refill_cfs_bandwidth_runtime(cfs_b);
7240 /* restart the period timer (if active) to handle new period expiry */
7241 if (runtime_enabled && cfs_b->timer_active) {
7242 /* force a reprogram */
7243 cfs_b->timer_active = 0;
7244 __start_cfs_bandwidth(cfs_b);
7246 raw_spin_unlock_irq(&cfs_b->lock);
7248 for_each_possible_cpu(i) {
7249 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7250 struct rq *rq = cfs_rq->rq;
7252 raw_spin_lock_irq(&rq->lock);
7253 cfs_rq->runtime_enabled = runtime_enabled;
7254 cfs_rq->runtime_remaining = 0;
7256 if (cfs_rq->throttled)
7257 unthrottle_cfs_rq(cfs_rq);
7258 raw_spin_unlock_irq(&rq->lock);
7261 mutex_unlock(&cfs_constraints_mutex);
7266 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7270 period = ktime_to_ns(tg->cfs_bandwidth.period);
7271 if (cfs_quota_us < 0)
7272 quota = RUNTIME_INF;
7274 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7276 return tg_set_cfs_bandwidth(tg, period, quota);
7279 long tg_get_cfs_quota(struct task_group *tg)
7283 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7286 quota_us = tg->cfs_bandwidth.quota;
7287 do_div(quota_us, NSEC_PER_USEC);
7292 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7296 period = (u64)cfs_period_us * NSEC_PER_USEC;
7297 quota = tg->cfs_bandwidth.quota;
7299 return tg_set_cfs_bandwidth(tg, period, quota);
7302 long tg_get_cfs_period(struct task_group *tg)
7306 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7307 do_div(cfs_period_us, NSEC_PER_USEC);
7309 return cfs_period_us;
7312 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7314 return tg_get_cfs_quota(cgroup_tg(cgrp));
7317 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7320 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7323 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7325 return tg_get_cfs_period(cgroup_tg(cgrp));
7328 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7331 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7334 struct cfs_schedulable_data {
7335 struct task_group *tg;
7340 * normalize group quota/period to be quota/max_period
7341 * note: units are usecs
7343 static u64 normalize_cfs_quota(struct task_group *tg,
7344 struct cfs_schedulable_data *d)
7352 period = tg_get_cfs_period(tg);
7353 quota = tg_get_cfs_quota(tg);
7356 /* note: these should typically be equivalent */
7357 if (quota == RUNTIME_INF || quota == -1)
7360 return to_ratio(period, quota);
7363 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7365 struct cfs_schedulable_data *d = data;
7366 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7367 s64 quota = 0, parent_quota = -1;
7370 quota = RUNTIME_INF;
7372 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7374 quota = normalize_cfs_quota(tg, d);
7375 parent_quota = parent_b->hierarchal_quota;
7378 * ensure max(child_quota) <= parent_quota, inherit when no
7381 if (quota == RUNTIME_INF)
7382 quota = parent_quota;
7383 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7386 cfs_b->hierarchal_quota = quota;
7391 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7394 struct cfs_schedulable_data data = {
7400 if (quota != RUNTIME_INF) {
7401 do_div(data.period, NSEC_PER_USEC);
7402 do_div(data.quota, NSEC_PER_USEC);
7406 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7412 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7413 struct cgroup_map_cb *cb)
7415 struct task_group *tg = cgroup_tg(cgrp);
7416 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7418 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7419 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7420 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7424 #endif /* CONFIG_CFS_BANDWIDTH */
7425 #endif /* CONFIG_FAIR_GROUP_SCHED */
7427 #ifdef CONFIG_RT_GROUP_SCHED
7428 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7431 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7434 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7436 return sched_group_rt_runtime(cgroup_tg(cgrp));
7439 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7442 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7445 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7447 return sched_group_rt_period(cgroup_tg(cgrp));
7449 #endif /* CONFIG_RT_GROUP_SCHED */
7451 static struct cftype cpu_files[] = {
7452 #ifdef CONFIG_FAIR_GROUP_SCHED
7455 .read_u64 = cpu_shares_read_u64,
7456 .write_u64 = cpu_shares_write_u64,
7459 #ifdef CONFIG_CFS_BANDWIDTH
7461 .name = "cfs_quota_us",
7462 .read_s64 = cpu_cfs_quota_read_s64,
7463 .write_s64 = cpu_cfs_quota_write_s64,
7466 .name = "cfs_period_us",
7467 .read_u64 = cpu_cfs_period_read_u64,
7468 .write_u64 = cpu_cfs_period_write_u64,
7472 .read_map = cpu_stats_show,
7475 #ifdef CONFIG_RT_GROUP_SCHED
7477 .name = "rt_runtime_us",
7478 .read_s64 = cpu_rt_runtime_read,
7479 .write_s64 = cpu_rt_runtime_write,
7482 .name = "rt_period_us",
7483 .read_u64 = cpu_rt_period_read_uint,
7484 .write_u64 = cpu_rt_period_write_uint,
7490 struct cgroup_subsys cpu_cgroup_subsys = {
7492 .css_alloc = cpu_cgroup_css_alloc,
7493 .css_free = cpu_cgroup_css_free,
7494 .css_online = cpu_cgroup_css_online,
7495 .css_offline = cpu_cgroup_css_offline,
7496 .can_attach = cpu_cgroup_can_attach,
7497 .attach = cpu_cgroup_attach,
7498 .exit = cpu_cgroup_exit,
7499 .subsys_id = cpu_cgroup_subsys_id,
7500 .base_cftypes = cpu_files,
7504 #endif /* CONFIG_CGROUP_SCHED */
7506 void dump_cpu_task(int cpu)
7508 pr_info("Task dump for CPU %d:\n", cpu);
7509 sched_show_task(cpu_curr(cpu));