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 * If we are going to wake up a thread waiting for CONDITION we
1496 * need to ensure that CONDITION=1 done by the caller can not be
1497 * reordered with p->state check below. This pairs with mb() in
1498 * set_current_state() the waiting thread does.
1500 smp_mb__before_spinlock();
1501 raw_spin_lock_irqsave(&p->pi_lock, flags);
1502 if (!(p->state & state))
1505 success = 1; /* we're going to change ->state */
1508 if (p->on_rq && ttwu_remote(p, wake_flags))
1513 * If the owning (remote) cpu is still in the middle of schedule() with
1514 * this task as prev, wait until its done referencing the task.
1519 * Pairs with the smp_wmb() in finish_lock_switch().
1523 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1524 p->state = TASK_WAKING;
1526 if (p->sched_class->task_waking)
1527 p->sched_class->task_waking(p);
1529 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1530 if (task_cpu(p) != cpu) {
1531 wake_flags |= WF_MIGRATED;
1532 set_task_cpu(p, cpu);
1534 #endif /* CONFIG_SMP */
1538 ttwu_stat(p, cpu, wake_flags);
1540 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1546 * try_to_wake_up_local - try to wake up a local task with rq lock held
1547 * @p: the thread to be awakened
1549 * Put @p on the run-queue if it's not already there. The caller must
1550 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1553 static void try_to_wake_up_local(struct task_struct *p)
1555 struct rq *rq = task_rq(p);
1557 if (WARN_ON_ONCE(rq != this_rq()) ||
1558 WARN_ON_ONCE(p == current))
1561 lockdep_assert_held(&rq->lock);
1563 if (!raw_spin_trylock(&p->pi_lock)) {
1564 raw_spin_unlock(&rq->lock);
1565 raw_spin_lock(&p->pi_lock);
1566 raw_spin_lock(&rq->lock);
1569 if (!(p->state & TASK_NORMAL))
1573 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1575 ttwu_do_wakeup(rq, p, 0);
1576 ttwu_stat(p, smp_processor_id(), 0);
1578 raw_spin_unlock(&p->pi_lock);
1582 * wake_up_process - Wake up a specific process
1583 * @p: The process to be woken up.
1585 * Attempt to wake up the nominated process and move it to the set of runnable
1586 * processes. Returns 1 if the process was woken up, 0 if it was already
1589 * It may be assumed that this function implies a write memory barrier before
1590 * changing the task state if and only if any tasks are woken up.
1592 int wake_up_process(struct task_struct *p)
1594 WARN_ON(task_is_stopped_or_traced(p));
1595 return try_to_wake_up(p, TASK_NORMAL, 0);
1597 EXPORT_SYMBOL(wake_up_process);
1599 int wake_up_state(struct task_struct *p, unsigned int state)
1601 return try_to_wake_up(p, state, 0);
1605 * Perform scheduler related setup for a newly forked process p.
1606 * p is forked by current.
1608 * __sched_fork() is basic setup used by init_idle() too:
1610 static void __sched_fork(struct task_struct *p)
1615 p->se.exec_start = 0;
1616 p->se.sum_exec_runtime = 0;
1617 p->se.prev_sum_exec_runtime = 0;
1618 p->se.nr_migrations = 0;
1620 INIT_LIST_HEAD(&p->se.group_node);
1622 #ifdef CONFIG_SCHEDSTATS
1623 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1626 INIT_LIST_HEAD(&p->rt.run_list);
1628 #ifdef CONFIG_PREEMPT_NOTIFIERS
1629 INIT_HLIST_HEAD(&p->preempt_notifiers);
1632 #ifdef CONFIG_NUMA_BALANCING
1633 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1634 p->mm->numa_next_scan = jiffies;
1635 p->mm->numa_next_reset = jiffies;
1636 p->mm->numa_scan_seq = 0;
1639 p->node_stamp = 0ULL;
1640 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1641 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1642 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1643 p->numa_work.next = &p->numa_work;
1644 #endif /* CONFIG_NUMA_BALANCING */
1647 #ifdef CONFIG_NUMA_BALANCING
1648 #ifdef CONFIG_SCHED_DEBUG
1649 void set_numabalancing_state(bool enabled)
1652 sched_feat_set("NUMA");
1654 sched_feat_set("NO_NUMA");
1657 __read_mostly bool numabalancing_enabled;
1659 void set_numabalancing_state(bool enabled)
1661 numabalancing_enabled = enabled;
1663 #endif /* CONFIG_SCHED_DEBUG */
1664 #endif /* CONFIG_NUMA_BALANCING */
1667 * fork()/clone()-time setup:
1669 void sched_fork(struct task_struct *p)
1671 unsigned long flags;
1672 int cpu = get_cpu();
1676 * We mark the process as running here. This guarantees that
1677 * nobody will actually run it, and a signal or other external
1678 * event cannot wake it up and insert it on the runqueue either.
1680 p->state = TASK_RUNNING;
1683 * Make sure we do not leak PI boosting priority to the child.
1685 p->prio = current->normal_prio;
1688 * Revert to default priority/policy on fork if requested.
1690 if (unlikely(p->sched_reset_on_fork)) {
1691 if (task_has_rt_policy(p)) {
1692 p->policy = SCHED_NORMAL;
1693 p->static_prio = NICE_TO_PRIO(0);
1695 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1696 p->static_prio = NICE_TO_PRIO(0);
1698 p->prio = p->normal_prio = __normal_prio(p);
1702 * We don't need the reset flag anymore after the fork. It has
1703 * fulfilled its duty:
1705 p->sched_reset_on_fork = 0;
1708 if (!rt_prio(p->prio))
1709 p->sched_class = &fair_sched_class;
1711 if (p->sched_class->task_fork)
1712 p->sched_class->task_fork(p);
1715 * The child is not yet in the pid-hash so no cgroup attach races,
1716 * and the cgroup is pinned to this child due to cgroup_fork()
1717 * is ran before sched_fork().
1719 * Silence PROVE_RCU.
1721 raw_spin_lock_irqsave(&p->pi_lock, flags);
1722 set_task_cpu(p, cpu);
1723 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1725 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1726 if (likely(sched_info_on()))
1727 memset(&p->sched_info, 0, sizeof(p->sched_info));
1729 #if defined(CONFIG_SMP)
1732 #ifdef CONFIG_PREEMPT_COUNT
1733 /* Want to start with kernel preemption disabled. */
1734 task_thread_info(p)->preempt_count = 1;
1737 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1744 * wake_up_new_task - wake up a newly created task for the first time.
1746 * This function will do some initial scheduler statistics housekeeping
1747 * that must be done for every newly created context, then puts the task
1748 * on the runqueue and wakes it.
1750 void wake_up_new_task(struct task_struct *p)
1752 unsigned long flags;
1755 raw_spin_lock_irqsave(&p->pi_lock, flags);
1758 * Fork balancing, do it here and not earlier because:
1759 * - cpus_allowed can change in the fork path
1760 * - any previously selected cpu might disappear through hotplug
1762 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1765 /* Initialize new task's runnable average */
1766 init_task_runnable_average(p);
1767 rq = __task_rq_lock(p);
1768 activate_task(rq, p, 0);
1770 trace_sched_wakeup_new(p, true);
1771 check_preempt_curr(rq, p, WF_FORK);
1773 if (p->sched_class->task_woken)
1774 p->sched_class->task_woken(rq, p);
1776 task_rq_unlock(rq, p, &flags);
1779 #ifdef CONFIG_PREEMPT_NOTIFIERS
1782 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1783 * @notifier: notifier struct to register
1785 void preempt_notifier_register(struct preempt_notifier *notifier)
1787 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1789 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1792 * preempt_notifier_unregister - no longer interested in preemption notifications
1793 * @notifier: notifier struct to unregister
1795 * This is safe to call from within a preemption notifier.
1797 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1799 hlist_del(¬ifier->link);
1801 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1803 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1805 struct preempt_notifier *notifier;
1807 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1808 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1812 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1813 struct task_struct *next)
1815 struct preempt_notifier *notifier;
1817 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1818 notifier->ops->sched_out(notifier, next);
1821 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1823 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1828 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1829 struct task_struct *next)
1833 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1836 * prepare_task_switch - prepare to switch tasks
1837 * @rq: the runqueue preparing to switch
1838 * @prev: the current task that is being switched out
1839 * @next: the task we are going to switch to.
1841 * This is called with the rq lock held and interrupts off. It must
1842 * be paired with a subsequent finish_task_switch after the context
1845 * prepare_task_switch sets up locking and calls architecture specific
1849 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1850 struct task_struct *next)
1852 trace_sched_switch(prev, next);
1853 sched_info_switch(prev, next);
1854 perf_event_task_sched_out(prev, next);
1855 fire_sched_out_preempt_notifiers(prev, next);
1856 prepare_lock_switch(rq, next);
1857 prepare_arch_switch(next);
1861 * finish_task_switch - clean up after a task-switch
1862 * @rq: runqueue associated with task-switch
1863 * @prev: the thread we just switched away from.
1865 * finish_task_switch must be called after the context switch, paired
1866 * with a prepare_task_switch call before the context switch.
1867 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1868 * and do any other architecture-specific cleanup actions.
1870 * Note that we may have delayed dropping an mm in context_switch(). If
1871 * so, we finish that here outside of the runqueue lock. (Doing it
1872 * with the lock held can cause deadlocks; see schedule() for
1875 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1876 __releases(rq->lock)
1878 struct mm_struct *mm = rq->prev_mm;
1884 * A task struct has one reference for the use as "current".
1885 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1886 * schedule one last time. The schedule call will never return, and
1887 * the scheduled task must drop that reference.
1888 * The test for TASK_DEAD must occur while the runqueue locks are
1889 * still held, otherwise prev could be scheduled on another cpu, die
1890 * there before we look at prev->state, and then the reference would
1892 * Manfred Spraul <manfred@colorfullife.com>
1894 prev_state = prev->state;
1895 vtime_task_switch(prev);
1896 finish_arch_switch(prev);
1897 perf_event_task_sched_in(prev, current);
1898 finish_lock_switch(rq, prev);
1899 finish_arch_post_lock_switch();
1901 fire_sched_in_preempt_notifiers(current);
1904 if (unlikely(prev_state == TASK_DEAD)) {
1906 * Remove function-return probe instances associated with this
1907 * task and put them back on the free list.
1909 kprobe_flush_task(prev);
1910 put_task_struct(prev);
1913 tick_nohz_task_switch(current);
1918 /* assumes rq->lock is held */
1919 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1921 if (prev->sched_class->pre_schedule)
1922 prev->sched_class->pre_schedule(rq, prev);
1925 /* rq->lock is NOT held, but preemption is disabled */
1926 static inline void post_schedule(struct rq *rq)
1928 if (rq->post_schedule) {
1929 unsigned long flags;
1931 raw_spin_lock_irqsave(&rq->lock, flags);
1932 if (rq->curr->sched_class->post_schedule)
1933 rq->curr->sched_class->post_schedule(rq);
1934 raw_spin_unlock_irqrestore(&rq->lock, flags);
1936 rq->post_schedule = 0;
1942 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1946 static inline void post_schedule(struct rq *rq)
1953 * schedule_tail - first thing a freshly forked thread must call.
1954 * @prev: the thread we just switched away from.
1956 asmlinkage void schedule_tail(struct task_struct *prev)
1957 __releases(rq->lock)
1959 struct rq *rq = this_rq();
1961 finish_task_switch(rq, prev);
1964 * FIXME: do we need to worry about rq being invalidated by the
1969 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1970 /* In this case, finish_task_switch does not reenable preemption */
1973 if (current->set_child_tid)
1974 put_user(task_pid_vnr(current), current->set_child_tid);
1978 * context_switch - switch to the new MM and the new
1979 * thread's register state.
1982 context_switch(struct rq *rq, struct task_struct *prev,
1983 struct task_struct *next)
1985 struct mm_struct *mm, *oldmm;
1987 prepare_task_switch(rq, prev, next);
1990 oldmm = prev->active_mm;
1992 * For paravirt, this is coupled with an exit in switch_to to
1993 * combine the page table reload and the switch backend into
1996 arch_start_context_switch(prev);
1999 next->active_mm = oldmm;
2000 atomic_inc(&oldmm->mm_count);
2001 enter_lazy_tlb(oldmm, next);
2003 switch_mm(oldmm, mm, next);
2006 prev->active_mm = NULL;
2007 rq->prev_mm = oldmm;
2010 * Since the runqueue lock will be released by the next
2011 * task (which is an invalid locking op but in the case
2012 * of the scheduler it's an obvious special-case), so we
2013 * do an early lockdep release here:
2015 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2016 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2019 context_tracking_task_switch(prev, next);
2020 /* Here we just switch the register state and the stack. */
2021 switch_to(prev, next, prev);
2025 * this_rq must be evaluated again because prev may have moved
2026 * CPUs since it called schedule(), thus the 'rq' on its stack
2027 * frame will be invalid.
2029 finish_task_switch(this_rq(), prev);
2033 * nr_running and nr_context_switches:
2035 * externally visible scheduler statistics: current number of runnable
2036 * threads, total number of context switches performed since bootup.
2038 unsigned long nr_running(void)
2040 unsigned long i, sum = 0;
2042 for_each_online_cpu(i)
2043 sum += cpu_rq(i)->nr_running;
2048 unsigned long long nr_context_switches(void)
2051 unsigned long long sum = 0;
2053 for_each_possible_cpu(i)
2054 sum += cpu_rq(i)->nr_switches;
2059 unsigned long nr_iowait(void)
2061 unsigned long i, sum = 0;
2063 for_each_possible_cpu(i)
2064 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2069 unsigned long nr_iowait_cpu(int cpu)
2071 struct rq *this = cpu_rq(cpu);
2072 return atomic_read(&this->nr_iowait);
2078 * sched_exec - execve() is a valuable balancing opportunity, because at
2079 * this point the task has the smallest effective memory and cache footprint.
2081 void sched_exec(void)
2083 struct task_struct *p = current;
2084 unsigned long flags;
2087 raw_spin_lock_irqsave(&p->pi_lock, flags);
2088 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2089 if (dest_cpu == smp_processor_id())
2092 if (likely(cpu_active(dest_cpu))) {
2093 struct migration_arg arg = { p, dest_cpu };
2095 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2096 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2100 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2105 DEFINE_PER_CPU(struct kernel_stat, kstat);
2106 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2108 EXPORT_PER_CPU_SYMBOL(kstat);
2109 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2112 * Return any ns on the sched_clock that have not yet been accounted in
2113 * @p in case that task is currently running.
2115 * Called with task_rq_lock() held on @rq.
2117 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2121 if (task_current(rq, p)) {
2122 update_rq_clock(rq);
2123 ns = rq_clock_task(rq) - p->se.exec_start;
2131 unsigned long long task_delta_exec(struct task_struct *p)
2133 unsigned long flags;
2137 rq = task_rq_lock(p, &flags);
2138 ns = do_task_delta_exec(p, rq);
2139 task_rq_unlock(rq, p, &flags);
2145 * Return accounted runtime for the task.
2146 * In case the task is currently running, return the runtime plus current's
2147 * pending runtime that have not been accounted yet.
2149 unsigned long long task_sched_runtime(struct task_struct *p)
2151 unsigned long flags;
2155 rq = task_rq_lock(p, &flags);
2156 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2157 task_rq_unlock(rq, p, &flags);
2163 * This function gets called by the timer code, with HZ frequency.
2164 * We call it with interrupts disabled.
2166 void scheduler_tick(void)
2168 int cpu = smp_processor_id();
2169 struct rq *rq = cpu_rq(cpu);
2170 struct task_struct *curr = rq->curr;
2174 raw_spin_lock(&rq->lock);
2175 update_rq_clock(rq);
2176 curr->sched_class->task_tick(rq, curr, 0);
2177 update_cpu_load_active(rq);
2178 raw_spin_unlock(&rq->lock);
2180 perf_event_task_tick();
2183 rq->idle_balance = idle_cpu(cpu);
2184 trigger_load_balance(rq, cpu);
2186 rq_last_tick_reset(rq);
2189 #ifdef CONFIG_NO_HZ_FULL
2191 * scheduler_tick_max_deferment
2193 * Keep at least one tick per second when a single
2194 * active task is running because the scheduler doesn't
2195 * yet completely support full dynticks environment.
2197 * This makes sure that uptime, CFS vruntime, load
2198 * balancing, etc... continue to move forward, even
2199 * with a very low granularity.
2201 u64 scheduler_tick_max_deferment(void)
2203 struct rq *rq = this_rq();
2204 unsigned long next, now = ACCESS_ONCE(jiffies);
2206 next = rq->last_sched_tick + HZ;
2208 if (time_before_eq(next, now))
2211 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2215 notrace unsigned long get_parent_ip(unsigned long addr)
2217 if (in_lock_functions(addr)) {
2218 addr = CALLER_ADDR2;
2219 if (in_lock_functions(addr))
2220 addr = CALLER_ADDR3;
2225 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2226 defined(CONFIG_PREEMPT_TRACER))
2228 void __kprobes add_preempt_count(int val)
2230 #ifdef CONFIG_DEBUG_PREEMPT
2234 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2237 preempt_count() += val;
2238 #ifdef CONFIG_DEBUG_PREEMPT
2240 * Spinlock count overflowing soon?
2242 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2245 if (preempt_count() == val)
2246 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2248 EXPORT_SYMBOL(add_preempt_count);
2250 void __kprobes sub_preempt_count(int val)
2252 #ifdef CONFIG_DEBUG_PREEMPT
2256 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2259 * Is the spinlock portion underflowing?
2261 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2262 !(preempt_count() & PREEMPT_MASK)))
2266 if (preempt_count() == val)
2267 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2268 preempt_count() -= val;
2270 EXPORT_SYMBOL(sub_preempt_count);
2275 * Print scheduling while atomic bug:
2277 static noinline void __schedule_bug(struct task_struct *prev)
2279 if (oops_in_progress)
2282 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2283 prev->comm, prev->pid, preempt_count());
2285 debug_show_held_locks(prev);
2287 if (irqs_disabled())
2288 print_irqtrace_events(prev);
2290 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2294 * Various schedule()-time debugging checks and statistics:
2296 static inline void schedule_debug(struct task_struct *prev)
2299 * Test if we are atomic. Since do_exit() needs to call into
2300 * schedule() atomically, we ignore that path for now.
2301 * Otherwise, whine if we are scheduling when we should not be.
2303 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2304 __schedule_bug(prev);
2307 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2309 schedstat_inc(this_rq(), sched_count);
2312 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2314 if (prev->on_rq || rq->skip_clock_update < 0)
2315 update_rq_clock(rq);
2316 prev->sched_class->put_prev_task(rq, prev);
2320 * Pick up the highest-prio task:
2322 static inline struct task_struct *
2323 pick_next_task(struct rq *rq)
2325 const struct sched_class *class;
2326 struct task_struct *p;
2329 * Optimization: we know that if all tasks are in
2330 * the fair class we can call that function directly:
2332 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2333 p = fair_sched_class.pick_next_task(rq);
2338 for_each_class(class) {
2339 p = class->pick_next_task(rq);
2344 BUG(); /* the idle class will always have a runnable task */
2348 * __schedule() is the main scheduler function.
2350 * The main means of driving the scheduler and thus entering this function are:
2352 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2354 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2355 * paths. For example, see arch/x86/entry_64.S.
2357 * To drive preemption between tasks, the scheduler sets the flag in timer
2358 * interrupt handler scheduler_tick().
2360 * 3. Wakeups don't really cause entry into schedule(). They add a
2361 * task to the run-queue and that's it.
2363 * Now, if the new task added to the run-queue preempts the current
2364 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2365 * called on the nearest possible occasion:
2367 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2369 * - in syscall or exception context, at the next outmost
2370 * preempt_enable(). (this might be as soon as the wake_up()'s
2373 * - in IRQ context, return from interrupt-handler to
2374 * preemptible context
2376 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2379 * - cond_resched() call
2380 * - explicit schedule() call
2381 * - return from syscall or exception to user-space
2382 * - return from interrupt-handler to user-space
2384 static void __sched __schedule(void)
2386 struct task_struct *prev, *next;
2387 unsigned long *switch_count;
2393 cpu = smp_processor_id();
2395 rcu_note_context_switch(cpu);
2398 schedule_debug(prev);
2400 if (sched_feat(HRTICK))
2404 * Make sure that signal_pending_state()->signal_pending() below
2405 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2406 * done by the caller to avoid the race with signal_wake_up().
2408 smp_mb__before_spinlock();
2409 raw_spin_lock_irq(&rq->lock);
2411 switch_count = &prev->nivcsw;
2412 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2413 if (unlikely(signal_pending_state(prev->state, prev))) {
2414 prev->state = TASK_RUNNING;
2416 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2420 * If a worker went to sleep, notify and ask workqueue
2421 * whether it wants to wake up a task to maintain
2424 if (prev->flags & PF_WQ_WORKER) {
2425 struct task_struct *to_wakeup;
2427 to_wakeup = wq_worker_sleeping(prev, cpu);
2429 try_to_wake_up_local(to_wakeup);
2432 switch_count = &prev->nvcsw;
2435 pre_schedule(rq, prev);
2437 if (unlikely(!rq->nr_running))
2438 idle_balance(cpu, rq);
2440 put_prev_task(rq, prev);
2441 next = pick_next_task(rq);
2442 clear_tsk_need_resched(prev);
2443 rq->skip_clock_update = 0;
2445 if (likely(prev != next)) {
2450 context_switch(rq, prev, next); /* unlocks the rq */
2452 * The context switch have flipped the stack from under us
2453 * and restored the local variables which were saved when
2454 * this task called schedule() in the past. prev == current
2455 * is still correct, but it can be moved to another cpu/rq.
2457 cpu = smp_processor_id();
2460 raw_spin_unlock_irq(&rq->lock);
2464 sched_preempt_enable_no_resched();
2469 static inline void sched_submit_work(struct task_struct *tsk)
2471 if (!tsk->state || tsk_is_pi_blocked(tsk))
2474 * If we are going to sleep and we have plugged IO queued,
2475 * make sure to submit it to avoid deadlocks.
2477 if (blk_needs_flush_plug(tsk))
2478 blk_schedule_flush_plug(tsk);
2481 asmlinkage void __sched schedule(void)
2483 struct task_struct *tsk = current;
2485 sched_submit_work(tsk);
2488 EXPORT_SYMBOL(schedule);
2490 #ifdef CONFIG_CONTEXT_TRACKING
2491 asmlinkage void __sched schedule_user(void)
2494 * If we come here after a random call to set_need_resched(),
2495 * or we have been woken up remotely but the IPI has not yet arrived,
2496 * we haven't yet exited the RCU idle mode. Do it here manually until
2497 * we find a better solution.
2506 * schedule_preempt_disabled - called with preemption disabled
2508 * Returns with preemption disabled. Note: preempt_count must be 1
2510 void __sched schedule_preempt_disabled(void)
2512 sched_preempt_enable_no_resched();
2517 #ifdef CONFIG_PREEMPT
2519 * this is the entry point to schedule() from in-kernel preemption
2520 * off of preempt_enable. Kernel preemptions off return from interrupt
2521 * occur there and call schedule directly.
2523 asmlinkage void __sched notrace preempt_schedule(void)
2525 struct thread_info *ti = current_thread_info();
2528 * If there is a non-zero preempt_count or interrupts are disabled,
2529 * we do not want to preempt the current task. Just return..
2531 if (likely(ti->preempt_count || irqs_disabled()))
2535 add_preempt_count_notrace(PREEMPT_ACTIVE);
2537 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2540 * Check again in case we missed a preemption opportunity
2541 * between schedule and now.
2544 } while (need_resched());
2546 EXPORT_SYMBOL(preempt_schedule);
2549 * this is the entry point to schedule() from kernel preemption
2550 * off of irq context.
2551 * Note, that this is called and return with irqs disabled. This will
2552 * protect us against recursive calling from irq.
2554 asmlinkage void __sched preempt_schedule_irq(void)
2556 struct thread_info *ti = current_thread_info();
2557 enum ctx_state prev_state;
2559 /* Catch callers which need to be fixed */
2560 BUG_ON(ti->preempt_count || !irqs_disabled());
2562 prev_state = exception_enter();
2565 add_preempt_count(PREEMPT_ACTIVE);
2568 local_irq_disable();
2569 sub_preempt_count(PREEMPT_ACTIVE);
2572 * Check again in case we missed a preemption opportunity
2573 * between schedule and now.
2576 } while (need_resched());
2578 exception_exit(prev_state);
2581 #endif /* CONFIG_PREEMPT */
2583 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2586 return try_to_wake_up(curr->private, mode, wake_flags);
2588 EXPORT_SYMBOL(default_wake_function);
2591 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2592 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2593 * number) then we wake all the non-exclusive tasks and one exclusive task.
2595 * There are circumstances in which we can try to wake a task which has already
2596 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2597 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2599 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2600 int nr_exclusive, int wake_flags, void *key)
2602 wait_queue_t *curr, *next;
2604 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2605 unsigned flags = curr->flags;
2607 if (curr->func(curr, mode, wake_flags, key) &&
2608 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2614 * __wake_up - wake up threads blocked on a waitqueue.
2616 * @mode: which threads
2617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2618 * @key: is directly passed to the wakeup function
2620 * It may be assumed that this function implies a write memory barrier before
2621 * changing the task state if and only if any tasks are woken up.
2623 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2624 int nr_exclusive, void *key)
2626 unsigned long flags;
2628 spin_lock_irqsave(&q->lock, flags);
2629 __wake_up_common(q, mode, nr_exclusive, 0, key);
2630 spin_unlock_irqrestore(&q->lock, flags);
2632 EXPORT_SYMBOL(__wake_up);
2635 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2637 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2639 __wake_up_common(q, mode, nr, 0, NULL);
2641 EXPORT_SYMBOL_GPL(__wake_up_locked);
2643 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2645 __wake_up_common(q, mode, 1, 0, key);
2647 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2650 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2652 * @mode: which threads
2653 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2654 * @key: opaque value to be passed to wakeup targets
2656 * The sync wakeup differs that the waker knows that it will schedule
2657 * away soon, so while the target thread will be woken up, it will not
2658 * be migrated to another CPU - ie. the two threads are 'synchronized'
2659 * with each other. This can prevent needless bouncing between CPUs.
2661 * On UP it can prevent extra preemption.
2663 * It may be assumed that this function implies a write memory barrier before
2664 * changing the task state if and only if any tasks are woken up.
2666 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2667 int nr_exclusive, void *key)
2669 unsigned long flags;
2670 int wake_flags = WF_SYNC;
2675 if (unlikely(!nr_exclusive))
2678 spin_lock_irqsave(&q->lock, flags);
2679 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2680 spin_unlock_irqrestore(&q->lock, flags);
2682 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2685 * __wake_up_sync - see __wake_up_sync_key()
2687 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2689 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2691 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2694 * complete: - signals a single thread waiting on this completion
2695 * @x: holds the state of this particular completion
2697 * This will wake up a single thread waiting on this completion. Threads will be
2698 * awakened in the same order in which they were queued.
2700 * See also complete_all(), wait_for_completion() and related routines.
2702 * It may be assumed that this function implies a write memory barrier before
2703 * changing the task state if and only if any tasks are woken up.
2705 void complete(struct completion *x)
2707 unsigned long flags;
2709 spin_lock_irqsave(&x->wait.lock, flags);
2711 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2712 spin_unlock_irqrestore(&x->wait.lock, flags);
2714 EXPORT_SYMBOL(complete);
2717 * complete_all: - signals all threads waiting on this completion
2718 * @x: holds the state of this particular completion
2720 * This will wake up all threads waiting on this particular completion event.
2722 * It may be assumed that this function implies a write memory barrier before
2723 * changing the task state if and only if any tasks are woken up.
2725 void complete_all(struct completion *x)
2727 unsigned long flags;
2729 spin_lock_irqsave(&x->wait.lock, flags);
2730 x->done += UINT_MAX/2;
2731 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2732 spin_unlock_irqrestore(&x->wait.lock, flags);
2734 EXPORT_SYMBOL(complete_all);
2736 static inline long __sched
2737 do_wait_for_common(struct completion *x,
2738 long (*action)(long), long timeout, int state)
2741 DECLARE_WAITQUEUE(wait, current);
2743 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2745 if (signal_pending_state(state, current)) {
2746 timeout = -ERESTARTSYS;
2749 __set_current_state(state);
2750 spin_unlock_irq(&x->wait.lock);
2751 timeout = action(timeout);
2752 spin_lock_irq(&x->wait.lock);
2753 } while (!x->done && timeout);
2754 __remove_wait_queue(&x->wait, &wait);
2759 return timeout ?: 1;
2762 static inline long __sched
2763 __wait_for_common(struct completion *x,
2764 long (*action)(long), long timeout, int state)
2768 spin_lock_irq(&x->wait.lock);
2769 timeout = do_wait_for_common(x, action, timeout, state);
2770 spin_unlock_irq(&x->wait.lock);
2775 wait_for_common(struct completion *x, long timeout, int state)
2777 return __wait_for_common(x, schedule_timeout, timeout, state);
2781 wait_for_common_io(struct completion *x, long timeout, int state)
2783 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2787 * wait_for_completion: - waits for completion of a task
2788 * @x: holds the state of this particular completion
2790 * This waits to be signaled for completion of a specific task. It is NOT
2791 * interruptible and there is no timeout.
2793 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2794 * and interrupt capability. Also see complete().
2796 void __sched wait_for_completion(struct completion *x)
2798 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2800 EXPORT_SYMBOL(wait_for_completion);
2803 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2804 * @x: holds the state of this particular completion
2805 * @timeout: timeout value in jiffies
2807 * This waits for either a completion of a specific task to be signaled or for a
2808 * specified timeout to expire. The timeout is in jiffies. It is not
2811 * The return value is 0 if timed out, and positive (at least 1, or number of
2812 * jiffies left till timeout) if completed.
2814 unsigned long __sched
2815 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2817 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2819 EXPORT_SYMBOL(wait_for_completion_timeout);
2822 * wait_for_completion_io: - waits for completion of a task
2823 * @x: holds the state of this particular completion
2825 * This waits to be signaled for completion of a specific task. It is NOT
2826 * interruptible and there is no timeout. The caller is accounted as waiting
2829 void __sched wait_for_completion_io(struct completion *x)
2831 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2833 EXPORT_SYMBOL(wait_for_completion_io);
2836 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2837 * @x: holds the state of this particular completion
2838 * @timeout: timeout value in jiffies
2840 * This waits for either a completion of a specific task to be signaled or for a
2841 * specified timeout to expire. The timeout is in jiffies. It is not
2842 * interruptible. The caller is accounted as waiting for IO.
2844 * The return value is 0 if timed out, and positive (at least 1, or number of
2845 * jiffies left till timeout) if completed.
2847 unsigned long __sched
2848 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2850 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2852 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2855 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2856 * @x: holds the state of this particular completion
2858 * This waits for completion of a specific task to be signaled. It is
2861 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2863 int __sched wait_for_completion_interruptible(struct completion *x)
2865 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2866 if (t == -ERESTARTSYS)
2870 EXPORT_SYMBOL(wait_for_completion_interruptible);
2873 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2874 * @x: holds the state of this particular completion
2875 * @timeout: timeout value in jiffies
2877 * This waits for either a completion of a specific task to be signaled or for a
2878 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2880 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2881 * positive (at least 1, or number of jiffies left till timeout) if completed.
2884 wait_for_completion_interruptible_timeout(struct completion *x,
2885 unsigned long timeout)
2887 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2889 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2892 * wait_for_completion_killable: - waits for completion of a task (killable)
2893 * @x: holds the state of this particular completion
2895 * This waits to be signaled for completion of a specific task. It can be
2896 * interrupted by a kill signal.
2898 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2900 int __sched wait_for_completion_killable(struct completion *x)
2902 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2903 if (t == -ERESTARTSYS)
2907 EXPORT_SYMBOL(wait_for_completion_killable);
2910 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2911 * @x: holds the state of this particular completion
2912 * @timeout: timeout value in jiffies
2914 * This waits for either a completion of a specific task to be
2915 * signaled or for a specified timeout to expire. It can be
2916 * interrupted by a kill signal. The timeout is in jiffies.
2918 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2919 * positive (at least 1, or number of jiffies left till timeout) if completed.
2922 wait_for_completion_killable_timeout(struct completion *x,
2923 unsigned long timeout)
2925 return wait_for_common(x, timeout, TASK_KILLABLE);
2927 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2930 * try_wait_for_completion - try to decrement a completion without blocking
2931 * @x: completion structure
2933 * Returns: 0 if a decrement cannot be done without blocking
2934 * 1 if a decrement succeeded.
2936 * If a completion is being used as a counting completion,
2937 * attempt to decrement the counter without blocking. This
2938 * enables us to avoid waiting if the resource the completion
2939 * is protecting is not available.
2941 bool try_wait_for_completion(struct completion *x)
2943 unsigned long flags;
2946 spin_lock_irqsave(&x->wait.lock, flags);
2951 spin_unlock_irqrestore(&x->wait.lock, flags);
2954 EXPORT_SYMBOL(try_wait_for_completion);
2957 * completion_done - Test to see if a completion has any waiters
2958 * @x: completion structure
2960 * Returns: 0 if there are waiters (wait_for_completion() in progress)
2961 * 1 if there are no waiters.
2964 bool completion_done(struct completion *x)
2966 unsigned long flags;
2969 spin_lock_irqsave(&x->wait.lock, flags);
2972 spin_unlock_irqrestore(&x->wait.lock, flags);
2975 EXPORT_SYMBOL(completion_done);
2978 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2980 unsigned long flags;
2983 init_waitqueue_entry(&wait, current);
2985 __set_current_state(state);
2987 spin_lock_irqsave(&q->lock, flags);
2988 __add_wait_queue(q, &wait);
2989 spin_unlock(&q->lock);
2990 timeout = schedule_timeout(timeout);
2991 spin_lock_irq(&q->lock);
2992 __remove_wait_queue(q, &wait);
2993 spin_unlock_irqrestore(&q->lock, flags);
2998 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3000 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3002 EXPORT_SYMBOL(interruptible_sleep_on);
3005 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3007 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3009 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3011 void __sched sleep_on(wait_queue_head_t *q)
3013 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3015 EXPORT_SYMBOL(sleep_on);
3017 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3019 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3021 EXPORT_SYMBOL(sleep_on_timeout);
3023 #ifdef CONFIG_RT_MUTEXES
3026 * rt_mutex_setprio - set the current priority of a task
3028 * @prio: prio value (kernel-internal form)
3030 * This function changes the 'effective' priority of a task. It does
3031 * not touch ->normal_prio like __setscheduler().
3033 * Used by the rt_mutex code to implement priority inheritance logic.
3035 void rt_mutex_setprio(struct task_struct *p, int prio)
3037 int oldprio, on_rq, running;
3039 const struct sched_class *prev_class;
3041 BUG_ON(prio < 0 || prio > MAX_PRIO);
3043 rq = __task_rq_lock(p);
3046 * Idle task boosting is a nono in general. There is one
3047 * exception, when PREEMPT_RT and NOHZ is active:
3049 * The idle task calls get_next_timer_interrupt() and holds
3050 * the timer wheel base->lock on the CPU and another CPU wants
3051 * to access the timer (probably to cancel it). We can safely
3052 * ignore the boosting request, as the idle CPU runs this code
3053 * with interrupts disabled and will complete the lock
3054 * protected section without being interrupted. So there is no
3055 * real need to boost.
3057 if (unlikely(p == rq->idle)) {
3058 WARN_ON(p != rq->curr);
3059 WARN_ON(p->pi_blocked_on);
3063 trace_sched_pi_setprio(p, prio);
3065 prev_class = p->sched_class;
3067 running = task_current(rq, p);
3069 dequeue_task(rq, p, 0);
3071 p->sched_class->put_prev_task(rq, p);
3074 p->sched_class = &rt_sched_class;
3076 p->sched_class = &fair_sched_class;
3081 p->sched_class->set_curr_task(rq);
3083 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3085 check_class_changed(rq, p, prev_class, oldprio);
3087 __task_rq_unlock(rq);
3090 void set_user_nice(struct task_struct *p, long nice)
3092 int old_prio, delta, on_rq;
3093 unsigned long flags;
3096 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3099 * We have to be careful, if called from sys_setpriority(),
3100 * the task might be in the middle of scheduling on another CPU.
3102 rq = task_rq_lock(p, &flags);
3104 * The RT priorities are set via sched_setscheduler(), but we still
3105 * allow the 'normal' nice value to be set - but as expected
3106 * it wont have any effect on scheduling until the task is
3107 * SCHED_FIFO/SCHED_RR:
3109 if (task_has_rt_policy(p)) {
3110 p->static_prio = NICE_TO_PRIO(nice);
3115 dequeue_task(rq, p, 0);
3117 p->static_prio = NICE_TO_PRIO(nice);
3120 p->prio = effective_prio(p);
3121 delta = p->prio - old_prio;
3124 enqueue_task(rq, p, 0);
3126 * If the task increased its priority or is running and
3127 * lowered its priority, then reschedule its CPU:
3129 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3130 resched_task(rq->curr);
3133 task_rq_unlock(rq, p, &flags);
3135 EXPORT_SYMBOL(set_user_nice);
3138 * can_nice - check if a task can reduce its nice value
3142 int can_nice(const struct task_struct *p, const int nice)
3144 /* convert nice value [19,-20] to rlimit style value [1,40] */
3145 int nice_rlim = 20 - nice;
3147 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3148 capable(CAP_SYS_NICE));
3151 #ifdef __ARCH_WANT_SYS_NICE
3154 * sys_nice - change the priority of the current process.
3155 * @increment: priority increment
3157 * sys_setpriority is a more generic, but much slower function that
3158 * does similar things.
3160 SYSCALL_DEFINE1(nice, int, increment)
3165 * Setpriority might change our priority at the same moment.
3166 * We don't have to worry. Conceptually one call occurs first
3167 * and we have a single winner.
3169 if (increment < -40)
3174 nice = TASK_NICE(current) + increment;
3180 if (increment < 0 && !can_nice(current, nice))
3183 retval = security_task_setnice(current, nice);
3187 set_user_nice(current, nice);
3194 * task_prio - return the priority value of a given task.
3195 * @p: the task in question.
3197 * This is the priority value as seen by users in /proc.
3198 * RT tasks are offset by -200. Normal tasks are centered
3199 * around 0, value goes from -16 to +15.
3201 int task_prio(const struct task_struct *p)
3203 return p->prio - MAX_RT_PRIO;
3207 * task_nice - return the nice value of a given task.
3208 * @p: the task in question.
3210 int task_nice(const struct task_struct *p)
3212 return TASK_NICE(p);
3214 EXPORT_SYMBOL(task_nice);
3217 * idle_cpu - is a given cpu idle currently?
3218 * @cpu: the processor in question.
3220 int idle_cpu(int cpu)
3222 struct rq *rq = cpu_rq(cpu);
3224 if (rq->curr != rq->idle)
3231 if (!llist_empty(&rq->wake_list))
3239 * idle_task - return the idle task for a given cpu.
3240 * @cpu: the processor in question.
3242 struct task_struct *idle_task(int cpu)
3244 return cpu_rq(cpu)->idle;
3248 * find_process_by_pid - find a process with a matching PID value.
3249 * @pid: the pid in question.
3251 static struct task_struct *find_process_by_pid(pid_t pid)
3253 return pid ? find_task_by_vpid(pid) : current;
3256 /* Actually do priority change: must hold rq lock. */
3258 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3261 p->rt_priority = prio;
3262 p->normal_prio = normal_prio(p);
3263 /* we are holding p->pi_lock already */
3264 p->prio = rt_mutex_getprio(p);
3265 if (rt_prio(p->prio))
3266 p->sched_class = &rt_sched_class;
3268 p->sched_class = &fair_sched_class;
3273 * check the target process has a UID that matches the current process's
3275 static bool check_same_owner(struct task_struct *p)
3277 const struct cred *cred = current_cred(), *pcred;
3281 pcred = __task_cred(p);
3282 match = (uid_eq(cred->euid, pcred->euid) ||
3283 uid_eq(cred->euid, pcred->uid));
3288 static int __sched_setscheduler(struct task_struct *p, int policy,
3289 const struct sched_param *param, bool user)
3291 int retval, oldprio, oldpolicy = -1, on_rq, running;
3292 unsigned long flags;
3293 const struct sched_class *prev_class;
3297 /* may grab non-irq protected spin_locks */
3298 BUG_ON(in_interrupt());
3300 /* double check policy once rq lock held */
3302 reset_on_fork = p->sched_reset_on_fork;
3303 policy = oldpolicy = p->policy;
3305 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3306 policy &= ~SCHED_RESET_ON_FORK;
3308 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3309 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3310 policy != SCHED_IDLE)
3315 * Valid priorities for SCHED_FIFO and SCHED_RR are
3316 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3317 * SCHED_BATCH and SCHED_IDLE is 0.
3319 if (param->sched_priority < 0 ||
3320 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3321 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3323 if (rt_policy(policy) != (param->sched_priority != 0))
3327 * Allow unprivileged RT tasks to decrease priority:
3329 if (user && !capable(CAP_SYS_NICE)) {
3330 if (rt_policy(policy)) {
3331 unsigned long rlim_rtprio =
3332 task_rlimit(p, RLIMIT_RTPRIO);
3334 /* can't set/change the rt policy */
3335 if (policy != p->policy && !rlim_rtprio)
3338 /* can't increase priority */
3339 if (param->sched_priority > p->rt_priority &&
3340 param->sched_priority > rlim_rtprio)
3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3349 if (!can_nice(p, TASK_NICE(p)))
3353 /* can't change other user's priorities */
3354 if (!check_same_owner(p))
3357 /* Normal users shall not reset the sched_reset_on_fork flag */
3358 if (p->sched_reset_on_fork && !reset_on_fork)
3363 retval = security_task_setscheduler(p);
3369 * make sure no PI-waiters arrive (or leave) while we are
3370 * changing the priority of the task:
3372 * To be able to change p->policy safely, the appropriate
3373 * runqueue lock must be held.
3375 rq = task_rq_lock(p, &flags);
3378 * Changing the policy of the stop threads its a very bad idea
3380 if (p == rq->stop) {
3381 task_rq_unlock(rq, p, &flags);
3386 * If not changing anything there's no need to proceed further:
3388 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3389 param->sched_priority == p->rt_priority))) {
3390 task_rq_unlock(rq, p, &flags);
3394 #ifdef CONFIG_RT_GROUP_SCHED
3397 * Do not allow realtime tasks into groups that have no runtime
3400 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3401 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3402 !task_group_is_autogroup(task_group(p))) {
3403 task_rq_unlock(rq, p, &flags);
3409 /* recheck policy now with rq lock held */
3410 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3411 policy = oldpolicy = -1;
3412 task_rq_unlock(rq, p, &flags);
3416 running = task_current(rq, p);
3418 dequeue_task(rq, p, 0);
3420 p->sched_class->put_prev_task(rq, p);
3422 p->sched_reset_on_fork = reset_on_fork;
3425 prev_class = p->sched_class;
3426 __setscheduler(rq, p, policy, param->sched_priority);
3429 p->sched_class->set_curr_task(rq);
3431 enqueue_task(rq, p, 0);
3433 check_class_changed(rq, p, prev_class, oldprio);
3434 task_rq_unlock(rq, p, &flags);
3436 rt_mutex_adjust_pi(p);
3442 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3443 * @p: the task in question.
3444 * @policy: new policy.
3445 * @param: structure containing the new RT priority.
3447 * NOTE that the task may be already dead.
3449 int sched_setscheduler(struct task_struct *p, int policy,
3450 const struct sched_param *param)
3452 return __sched_setscheduler(p, policy, param, true);
3454 EXPORT_SYMBOL_GPL(sched_setscheduler);
3457 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3458 * @p: the task in question.
3459 * @policy: new policy.
3460 * @param: structure containing the new RT priority.
3462 * Just like sched_setscheduler, only don't bother checking if the
3463 * current context has permission. For example, this is needed in
3464 * stop_machine(): we create temporary high priority worker threads,
3465 * but our caller might not have that capability.
3467 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3468 const struct sched_param *param)
3470 return __sched_setscheduler(p, policy, param, false);
3474 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3476 struct sched_param lparam;
3477 struct task_struct *p;
3480 if (!param || pid < 0)
3482 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3487 p = find_process_by_pid(pid);
3489 retval = sched_setscheduler(p, policy, &lparam);
3496 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3497 * @pid: the pid in question.
3498 * @policy: new policy.
3499 * @param: structure containing the new RT priority.
3501 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3502 struct sched_param __user *, param)
3504 /* negative values for policy are not valid */
3508 return do_sched_setscheduler(pid, policy, param);
3512 * sys_sched_setparam - set/change the RT priority of a thread
3513 * @pid: the pid in question.
3514 * @param: structure containing the new RT priority.
3516 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3518 return do_sched_setscheduler(pid, -1, param);
3522 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3523 * @pid: the pid in question.
3525 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3527 struct task_struct *p;
3535 p = find_process_by_pid(pid);
3537 retval = security_task_getscheduler(p);
3540 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3547 * sys_sched_getparam - get the RT priority of a thread
3548 * @pid: the pid in question.
3549 * @param: structure containing the RT priority.
3551 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3553 struct sched_param lp;
3554 struct task_struct *p;
3557 if (!param || pid < 0)
3561 p = find_process_by_pid(pid);
3566 retval = security_task_getscheduler(p);
3570 lp.sched_priority = p->rt_priority;
3574 * This one might sleep, we cannot do it with a spinlock held ...
3576 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3585 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3587 cpumask_var_t cpus_allowed, new_mask;
3588 struct task_struct *p;
3594 p = find_process_by_pid(pid);
3601 /* Prevent p going away */
3605 if (p->flags & PF_NO_SETAFFINITY) {
3609 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3613 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3615 goto out_free_cpus_allowed;
3618 if (!check_same_owner(p)) {
3620 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3627 retval = security_task_setscheduler(p);
3631 cpuset_cpus_allowed(p, cpus_allowed);
3632 cpumask_and(new_mask, in_mask, cpus_allowed);
3634 retval = set_cpus_allowed_ptr(p, new_mask);
3637 cpuset_cpus_allowed(p, cpus_allowed);
3638 if (!cpumask_subset(new_mask, cpus_allowed)) {
3640 * We must have raced with a concurrent cpuset
3641 * update. Just reset the cpus_allowed to the
3642 * cpuset's cpus_allowed
3644 cpumask_copy(new_mask, cpus_allowed);
3649 free_cpumask_var(new_mask);
3650 out_free_cpus_allowed:
3651 free_cpumask_var(cpus_allowed);
3658 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3659 struct cpumask *new_mask)
3661 if (len < cpumask_size())
3662 cpumask_clear(new_mask);
3663 else if (len > cpumask_size())
3664 len = cpumask_size();
3666 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3670 * sys_sched_setaffinity - set the cpu affinity of a process
3671 * @pid: pid of the process
3672 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3673 * @user_mask_ptr: user-space pointer to the new cpu mask
3675 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3676 unsigned long __user *, user_mask_ptr)
3678 cpumask_var_t new_mask;
3681 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3684 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3686 retval = sched_setaffinity(pid, new_mask);
3687 free_cpumask_var(new_mask);
3691 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3693 struct task_struct *p;
3694 unsigned long flags;
3701 p = find_process_by_pid(pid);
3705 retval = security_task_getscheduler(p);
3709 raw_spin_lock_irqsave(&p->pi_lock, flags);
3710 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3711 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3721 * sys_sched_getaffinity - get the cpu affinity of a process
3722 * @pid: pid of the process
3723 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3724 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3726 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3727 unsigned long __user *, user_mask_ptr)
3732 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3734 if (len & (sizeof(unsigned long)-1))
3737 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3740 ret = sched_getaffinity(pid, mask);
3742 size_t retlen = min_t(size_t, len, cpumask_size());
3744 if (copy_to_user(user_mask_ptr, mask, retlen))
3749 free_cpumask_var(mask);
3755 * sys_sched_yield - yield the current processor to other threads.
3757 * This function yields the current CPU to other tasks. If there are no
3758 * other threads running on this CPU then this function will return.
3760 SYSCALL_DEFINE0(sched_yield)
3762 struct rq *rq = this_rq_lock();
3764 schedstat_inc(rq, yld_count);
3765 current->sched_class->yield_task(rq);
3768 * Since we are going to call schedule() anyway, there's
3769 * no need to preempt or enable interrupts:
3771 __release(rq->lock);
3772 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3773 do_raw_spin_unlock(&rq->lock);
3774 sched_preempt_enable_no_resched();
3781 static inline int should_resched(void)
3783 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
3786 static void __cond_resched(void)
3788 add_preempt_count(PREEMPT_ACTIVE);
3790 sub_preempt_count(PREEMPT_ACTIVE);
3793 int __sched _cond_resched(void)
3795 if (should_resched()) {
3801 EXPORT_SYMBOL(_cond_resched);
3804 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3805 * call schedule, and on return reacquire the lock.
3807 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3808 * operations here to prevent schedule() from being called twice (once via
3809 * spin_unlock(), once by hand).
3811 int __cond_resched_lock(spinlock_t *lock)
3813 int resched = should_resched();
3816 lockdep_assert_held(lock);
3818 if (spin_needbreak(lock) || resched) {
3829 EXPORT_SYMBOL(__cond_resched_lock);
3831 int __sched __cond_resched_softirq(void)
3833 BUG_ON(!in_softirq());
3835 if (should_resched()) {
3843 EXPORT_SYMBOL(__cond_resched_softirq);
3846 * yield - yield the current processor to other threads.
3848 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3850 * The scheduler is at all times free to pick the calling task as the most
3851 * eligible task to run, if removing the yield() call from your code breaks
3852 * it, its already broken.
3854 * Typical broken usage is:
3859 * where one assumes that yield() will let 'the other' process run that will
3860 * make event true. If the current task is a SCHED_FIFO task that will never
3861 * happen. Never use yield() as a progress guarantee!!
3863 * If you want to use yield() to wait for something, use wait_event().
3864 * If you want to use yield() to be 'nice' for others, use cond_resched().
3865 * If you still want to use yield(), do not!
3867 void __sched yield(void)
3869 set_current_state(TASK_RUNNING);
3872 EXPORT_SYMBOL(yield);
3875 * yield_to - yield the current processor to another thread in
3876 * your thread group, or accelerate that thread toward the
3877 * processor it's on.
3879 * @preempt: whether task preemption is allowed or not
3881 * It's the caller's job to ensure that the target task struct
3882 * can't go away on us before we can do any checks.
3885 * true (>0) if we indeed boosted the target task.
3886 * false (0) if we failed to boost the target.
3887 * -ESRCH if there's no task to yield to.
3889 bool __sched yield_to(struct task_struct *p, bool preempt)
3891 struct task_struct *curr = current;
3892 struct rq *rq, *p_rq;
3893 unsigned long flags;
3896 local_irq_save(flags);
3902 * If we're the only runnable task on the rq and target rq also
3903 * has only one task, there's absolutely no point in yielding.
3905 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3910 double_rq_lock(rq, p_rq);
3911 while (task_rq(p) != p_rq) {
3912 double_rq_unlock(rq, p_rq);
3916 if (!curr->sched_class->yield_to_task)
3919 if (curr->sched_class != p->sched_class)
3922 if (task_running(p_rq, p) || p->state)
3925 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3927 schedstat_inc(rq, yld_count);
3929 * Make p's CPU reschedule; pick_next_entity takes care of
3932 if (preempt && rq != p_rq)
3933 resched_task(p_rq->curr);
3937 double_rq_unlock(rq, p_rq);
3939 local_irq_restore(flags);
3946 EXPORT_SYMBOL_GPL(yield_to);
3949 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3950 * that process accounting knows that this is a task in IO wait state.
3952 void __sched io_schedule(void)
3954 struct rq *rq = raw_rq();
3956 delayacct_blkio_start();
3957 atomic_inc(&rq->nr_iowait);
3958 blk_flush_plug(current);
3959 current->in_iowait = 1;
3961 current->in_iowait = 0;
3962 atomic_dec(&rq->nr_iowait);
3963 delayacct_blkio_end();
3965 EXPORT_SYMBOL(io_schedule);
3967 long __sched io_schedule_timeout(long timeout)
3969 struct rq *rq = raw_rq();
3972 delayacct_blkio_start();
3973 atomic_inc(&rq->nr_iowait);
3974 blk_flush_plug(current);
3975 current->in_iowait = 1;
3976 ret = schedule_timeout(timeout);
3977 current->in_iowait = 0;
3978 atomic_dec(&rq->nr_iowait);
3979 delayacct_blkio_end();
3984 * sys_sched_get_priority_max - return maximum RT priority.
3985 * @policy: scheduling class.
3987 * this syscall returns the maximum rt_priority that can be used
3988 * by a given scheduling class.
3990 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3997 ret = MAX_USER_RT_PRIO-1;
4009 * sys_sched_get_priority_min - return minimum RT priority.
4010 * @policy: scheduling class.
4012 * this syscall returns the minimum rt_priority that can be used
4013 * by a given scheduling class.
4015 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4033 * sys_sched_rr_get_interval - return the default timeslice of a process.
4034 * @pid: pid of the process.
4035 * @interval: userspace pointer to the timeslice value.
4037 * this syscall writes the default timeslice value of a given process
4038 * into the user-space timespec buffer. A value of '0' means infinity.
4040 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4041 struct timespec __user *, interval)
4043 struct task_struct *p;
4044 unsigned int time_slice;
4045 unsigned long flags;
4055 p = find_process_by_pid(pid);
4059 retval = security_task_getscheduler(p);
4063 rq = task_rq_lock(p, &flags);
4064 time_slice = p->sched_class->get_rr_interval(rq, p);
4065 task_rq_unlock(rq, p, &flags);
4068 jiffies_to_timespec(time_slice, &t);
4069 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4077 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4079 void sched_show_task(struct task_struct *p)
4081 unsigned long free = 0;
4085 state = p->state ? __ffs(p->state) + 1 : 0;
4086 printk(KERN_INFO "%-15.15s %c", p->comm,
4087 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4088 #if BITS_PER_LONG == 32
4089 if (state == TASK_RUNNING)
4090 printk(KERN_CONT " running ");
4092 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4094 if (state == TASK_RUNNING)
4095 printk(KERN_CONT " running task ");
4097 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4099 #ifdef CONFIG_DEBUG_STACK_USAGE
4100 free = stack_not_used(p);
4103 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4105 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4106 task_pid_nr(p), ppid,
4107 (unsigned long)task_thread_info(p)->flags);
4109 print_worker_info(KERN_INFO, p);
4110 show_stack(p, NULL);
4113 void show_state_filter(unsigned long state_filter)
4115 struct task_struct *g, *p;
4117 #if BITS_PER_LONG == 32
4119 " task PC stack pid father\n");
4122 " task PC stack pid father\n");
4125 do_each_thread(g, p) {
4127 * reset the NMI-timeout, listing all files on a slow
4128 * console might take a lot of time:
4130 touch_nmi_watchdog();
4131 if (!state_filter || (p->state & state_filter))
4133 } while_each_thread(g, p);
4135 touch_all_softlockup_watchdogs();
4137 #ifdef CONFIG_SCHED_DEBUG
4138 sysrq_sched_debug_show();
4142 * Only show locks if all tasks are dumped:
4145 debug_show_all_locks();
4148 void init_idle_bootup_task(struct task_struct *idle)
4150 idle->sched_class = &idle_sched_class;
4154 * init_idle - set up an idle thread for a given CPU
4155 * @idle: task in question
4156 * @cpu: cpu the idle task belongs to
4158 * NOTE: this function does not set the idle thread's NEED_RESCHED
4159 * flag, to make booting more robust.
4161 void init_idle(struct task_struct *idle, int cpu)
4163 struct rq *rq = cpu_rq(cpu);
4164 unsigned long flags;
4166 raw_spin_lock_irqsave(&rq->lock, flags);
4169 idle->state = TASK_RUNNING;
4170 idle->se.exec_start = sched_clock();
4172 do_set_cpus_allowed(idle, cpumask_of(cpu));
4174 * We're having a chicken and egg problem, even though we are
4175 * holding rq->lock, the cpu isn't yet set to this cpu so the
4176 * lockdep check in task_group() will fail.
4178 * Similar case to sched_fork(). / Alternatively we could
4179 * use task_rq_lock() here and obtain the other rq->lock.
4184 __set_task_cpu(idle, cpu);
4187 rq->curr = rq->idle = idle;
4188 #if defined(CONFIG_SMP)
4191 raw_spin_unlock_irqrestore(&rq->lock, flags);
4193 /* Set the preempt count _outside_ the spinlocks! */
4194 task_thread_info(idle)->preempt_count = 0;
4197 * The idle tasks have their own, simple scheduling class:
4199 idle->sched_class = &idle_sched_class;
4200 ftrace_graph_init_idle_task(idle, cpu);
4201 vtime_init_idle(idle, cpu);
4202 #if defined(CONFIG_SMP)
4203 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4208 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4210 if (p->sched_class && p->sched_class->set_cpus_allowed)
4211 p->sched_class->set_cpus_allowed(p, new_mask);
4213 cpumask_copy(&p->cpus_allowed, new_mask);
4214 p->nr_cpus_allowed = cpumask_weight(new_mask);
4218 * This is how migration works:
4220 * 1) we invoke migration_cpu_stop() on the target CPU using
4222 * 2) stopper starts to run (implicitly forcing the migrated thread
4224 * 3) it checks whether the migrated task is still in the wrong runqueue.
4225 * 4) if it's in the wrong runqueue then the migration thread removes
4226 * it and puts it into the right queue.
4227 * 5) stopper completes and stop_one_cpu() returns and the migration
4232 * Change a given task's CPU affinity. Migrate the thread to a
4233 * proper CPU and schedule it away if the CPU it's executing on
4234 * is removed from the allowed bitmask.
4236 * NOTE: the caller must have a valid reference to the task, the
4237 * task must not exit() & deallocate itself prematurely. The
4238 * call is not atomic; no spinlocks may be held.
4240 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4242 unsigned long flags;
4244 unsigned int dest_cpu;
4247 rq = task_rq_lock(p, &flags);
4249 if (cpumask_equal(&p->cpus_allowed, new_mask))
4252 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4257 do_set_cpus_allowed(p, new_mask);
4259 /* Can the task run on the task's current CPU? If so, we're done */
4260 if (cpumask_test_cpu(task_cpu(p), new_mask))
4263 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4265 struct migration_arg arg = { p, dest_cpu };
4266 /* Need help from migration thread: drop lock and wait. */
4267 task_rq_unlock(rq, p, &flags);
4268 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4269 tlb_migrate_finish(p->mm);
4273 task_rq_unlock(rq, p, &flags);
4277 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4280 * Move (not current) task off this cpu, onto dest cpu. We're doing
4281 * this because either it can't run here any more (set_cpus_allowed()
4282 * away from this CPU, or CPU going down), or because we're
4283 * attempting to rebalance this task on exec (sched_exec).
4285 * So we race with normal scheduler movements, but that's OK, as long
4286 * as the task is no longer on this CPU.
4288 * Returns non-zero if task was successfully migrated.
4290 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4292 struct rq *rq_dest, *rq_src;
4295 if (unlikely(!cpu_active(dest_cpu)))
4298 rq_src = cpu_rq(src_cpu);
4299 rq_dest = cpu_rq(dest_cpu);
4301 raw_spin_lock(&p->pi_lock);
4302 double_rq_lock(rq_src, rq_dest);
4303 /* Already moved. */
4304 if (task_cpu(p) != src_cpu)
4306 /* Affinity changed (again). */
4307 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4311 * If we're not on a rq, the next wake-up will ensure we're
4315 dequeue_task(rq_src, p, 0);
4316 set_task_cpu(p, dest_cpu);
4317 enqueue_task(rq_dest, p, 0);
4318 check_preempt_curr(rq_dest, p, 0);
4323 double_rq_unlock(rq_src, rq_dest);
4324 raw_spin_unlock(&p->pi_lock);
4329 * migration_cpu_stop - this will be executed by a highprio stopper thread
4330 * and performs thread migration by bumping thread off CPU then
4331 * 'pushing' onto another runqueue.
4333 static int migration_cpu_stop(void *data)
4335 struct migration_arg *arg = data;
4338 * The original target cpu might have gone down and we might
4339 * be on another cpu but it doesn't matter.
4341 local_irq_disable();
4342 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4347 #ifdef CONFIG_HOTPLUG_CPU
4350 * Ensures that the idle task is using init_mm right before its cpu goes
4353 void idle_task_exit(void)
4355 struct mm_struct *mm = current->active_mm;
4357 BUG_ON(cpu_online(smp_processor_id()));
4360 switch_mm(mm, &init_mm, current);
4365 * Since this CPU is going 'away' for a while, fold any nr_active delta
4366 * we might have. Assumes we're called after migrate_tasks() so that the
4367 * nr_active count is stable.
4369 * Also see the comment "Global load-average calculations".
4371 static void calc_load_migrate(struct rq *rq)
4373 long delta = calc_load_fold_active(rq);
4375 atomic_long_add(delta, &calc_load_tasks);
4379 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4380 * try_to_wake_up()->select_task_rq().
4382 * Called with rq->lock held even though we'er in stop_machine() and
4383 * there's no concurrency possible, we hold the required locks anyway
4384 * because of lock validation efforts.
4386 static void migrate_tasks(unsigned int dead_cpu)
4388 struct rq *rq = cpu_rq(dead_cpu);
4389 struct task_struct *next, *stop = rq->stop;
4393 * Fudge the rq selection such that the below task selection loop
4394 * doesn't get stuck on the currently eligible stop task.
4396 * We're currently inside stop_machine() and the rq is either stuck
4397 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4398 * either way we should never end up calling schedule() until we're
4404 * put_prev_task() and pick_next_task() sched
4405 * class method both need to have an up-to-date
4406 * value of rq->clock[_task]
4408 update_rq_clock(rq);
4412 * There's this thread running, bail when that's the only
4415 if (rq->nr_running == 1)
4418 next = pick_next_task(rq);
4420 next->sched_class->put_prev_task(rq, next);
4422 /* Find suitable destination for @next, with force if needed. */
4423 dest_cpu = select_fallback_rq(dead_cpu, next);
4424 raw_spin_unlock(&rq->lock);
4426 __migrate_task(next, dead_cpu, dest_cpu);
4428 raw_spin_lock(&rq->lock);
4434 #endif /* CONFIG_HOTPLUG_CPU */
4436 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4438 static struct ctl_table sd_ctl_dir[] = {
4440 .procname = "sched_domain",
4446 static struct ctl_table sd_ctl_root[] = {
4448 .procname = "kernel",
4450 .child = sd_ctl_dir,
4455 static struct ctl_table *sd_alloc_ctl_entry(int n)
4457 struct ctl_table *entry =
4458 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4463 static void sd_free_ctl_entry(struct ctl_table **tablep)
4465 struct ctl_table *entry;
4468 * In the intermediate directories, both the child directory and
4469 * procname are dynamically allocated and could fail but the mode
4470 * will always be set. In the lowest directory the names are
4471 * static strings and all have proc handlers.
4473 for (entry = *tablep; entry->mode; entry++) {
4475 sd_free_ctl_entry(&entry->child);
4476 if (entry->proc_handler == NULL)
4477 kfree(entry->procname);
4484 static int min_load_idx = 0;
4485 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4488 set_table_entry(struct ctl_table *entry,
4489 const char *procname, void *data, int maxlen,
4490 umode_t mode, proc_handler *proc_handler,
4493 entry->procname = procname;
4495 entry->maxlen = maxlen;
4497 entry->proc_handler = proc_handler;
4500 entry->extra1 = &min_load_idx;
4501 entry->extra2 = &max_load_idx;
4505 static struct ctl_table *
4506 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4508 struct ctl_table *table = sd_alloc_ctl_entry(13);
4513 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4514 sizeof(long), 0644, proc_doulongvec_minmax, false);
4515 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4516 sizeof(long), 0644, proc_doulongvec_minmax, false);
4517 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4518 sizeof(int), 0644, proc_dointvec_minmax, true);
4519 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4520 sizeof(int), 0644, proc_dointvec_minmax, true);
4521 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4522 sizeof(int), 0644, proc_dointvec_minmax, true);
4523 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4524 sizeof(int), 0644, proc_dointvec_minmax, true);
4525 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4526 sizeof(int), 0644, proc_dointvec_minmax, true);
4527 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4528 sizeof(int), 0644, proc_dointvec_minmax, false);
4529 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4530 sizeof(int), 0644, proc_dointvec_minmax, false);
4531 set_table_entry(&table[9], "cache_nice_tries",
4532 &sd->cache_nice_tries,
4533 sizeof(int), 0644, proc_dointvec_minmax, false);
4534 set_table_entry(&table[10], "flags", &sd->flags,
4535 sizeof(int), 0644, proc_dointvec_minmax, false);
4536 set_table_entry(&table[11], "name", sd->name,
4537 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4538 /* &table[12] is terminator */
4543 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4545 struct ctl_table *entry, *table;
4546 struct sched_domain *sd;
4547 int domain_num = 0, i;
4550 for_each_domain(cpu, sd)
4552 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4557 for_each_domain(cpu, sd) {
4558 snprintf(buf, 32, "domain%d", i);
4559 entry->procname = kstrdup(buf, GFP_KERNEL);
4561 entry->child = sd_alloc_ctl_domain_table(sd);
4568 static struct ctl_table_header *sd_sysctl_header;
4569 static void register_sched_domain_sysctl(void)
4571 int i, cpu_num = num_possible_cpus();
4572 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4575 WARN_ON(sd_ctl_dir[0].child);
4576 sd_ctl_dir[0].child = entry;
4581 for_each_possible_cpu(i) {
4582 snprintf(buf, 32, "cpu%d", i);
4583 entry->procname = kstrdup(buf, GFP_KERNEL);
4585 entry->child = sd_alloc_ctl_cpu_table(i);
4589 WARN_ON(sd_sysctl_header);
4590 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4593 /* may be called multiple times per register */
4594 static void unregister_sched_domain_sysctl(void)
4596 if (sd_sysctl_header)
4597 unregister_sysctl_table(sd_sysctl_header);
4598 sd_sysctl_header = NULL;
4599 if (sd_ctl_dir[0].child)
4600 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4603 static void register_sched_domain_sysctl(void)
4606 static void unregister_sched_domain_sysctl(void)
4611 static void set_rq_online(struct rq *rq)
4614 const struct sched_class *class;
4616 cpumask_set_cpu(rq->cpu, rq->rd->online);
4619 for_each_class(class) {
4620 if (class->rq_online)
4621 class->rq_online(rq);
4626 static void set_rq_offline(struct rq *rq)
4629 const struct sched_class *class;
4631 for_each_class(class) {
4632 if (class->rq_offline)
4633 class->rq_offline(rq);
4636 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4642 * migration_call - callback that gets triggered when a CPU is added.
4643 * Here we can start up the necessary migration thread for the new CPU.
4646 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4648 int cpu = (long)hcpu;
4649 unsigned long flags;
4650 struct rq *rq = cpu_rq(cpu);
4652 switch (action & ~CPU_TASKS_FROZEN) {
4654 case CPU_UP_PREPARE:
4655 rq->calc_load_update = calc_load_update;
4659 /* Update our root-domain */
4660 raw_spin_lock_irqsave(&rq->lock, flags);
4662 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4666 raw_spin_unlock_irqrestore(&rq->lock, flags);
4669 #ifdef CONFIG_HOTPLUG_CPU
4671 sched_ttwu_pending();
4672 /* Update our root-domain */
4673 raw_spin_lock_irqsave(&rq->lock, flags);
4675 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4679 BUG_ON(rq->nr_running != 1); /* the migration thread */
4680 raw_spin_unlock_irqrestore(&rq->lock, flags);
4684 calc_load_migrate(rq);
4689 update_max_interval();
4695 * Register at high priority so that task migration (migrate_all_tasks)
4696 * happens before everything else. This has to be lower priority than
4697 * the notifier in the perf_event subsystem, though.
4699 static struct notifier_block migration_notifier = {
4700 .notifier_call = migration_call,
4701 .priority = CPU_PRI_MIGRATION,
4704 static int sched_cpu_active(struct notifier_block *nfb,
4705 unsigned long action, void *hcpu)
4707 switch (action & ~CPU_TASKS_FROZEN) {
4709 case CPU_DOWN_FAILED:
4710 set_cpu_active((long)hcpu, true);
4717 static int sched_cpu_inactive(struct notifier_block *nfb,
4718 unsigned long action, void *hcpu)
4720 switch (action & ~CPU_TASKS_FROZEN) {
4721 case CPU_DOWN_PREPARE:
4722 set_cpu_active((long)hcpu, false);
4729 static int __init migration_init(void)
4731 void *cpu = (void *)(long)smp_processor_id();
4734 /* Initialize migration for the boot CPU */
4735 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4736 BUG_ON(err == NOTIFY_BAD);
4737 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4738 register_cpu_notifier(&migration_notifier);
4740 /* Register cpu active notifiers */
4741 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4742 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4746 early_initcall(migration_init);
4751 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4753 #ifdef CONFIG_SCHED_DEBUG
4755 static __read_mostly int sched_debug_enabled;
4757 static int __init sched_debug_setup(char *str)
4759 sched_debug_enabled = 1;
4763 early_param("sched_debug", sched_debug_setup);
4765 static inline bool sched_debug(void)
4767 return sched_debug_enabled;
4770 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4771 struct cpumask *groupmask)
4773 struct sched_group *group = sd->groups;
4776 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4777 cpumask_clear(groupmask);
4779 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4781 if (!(sd->flags & SD_LOAD_BALANCE)) {
4782 printk("does not load-balance\n");
4784 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4789 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4791 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4792 printk(KERN_ERR "ERROR: domain->span does not contain "
4795 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4796 printk(KERN_ERR "ERROR: domain->groups does not contain"
4800 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4804 printk(KERN_ERR "ERROR: group is NULL\n");
4809 * Even though we initialize ->power to something semi-sane,
4810 * we leave power_orig unset. This allows us to detect if
4811 * domain iteration is still funny without causing /0 traps.
4813 if (!group->sgp->power_orig) {
4814 printk(KERN_CONT "\n");
4815 printk(KERN_ERR "ERROR: domain->cpu_power not "
4820 if (!cpumask_weight(sched_group_cpus(group))) {
4821 printk(KERN_CONT "\n");
4822 printk(KERN_ERR "ERROR: empty group\n");
4826 if (!(sd->flags & SD_OVERLAP) &&
4827 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4828 printk(KERN_CONT "\n");
4829 printk(KERN_ERR "ERROR: repeated CPUs\n");
4833 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4835 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4837 printk(KERN_CONT " %s", str);
4838 if (group->sgp->power != SCHED_POWER_SCALE) {
4839 printk(KERN_CONT " (cpu_power = %d)",
4843 group = group->next;
4844 } while (group != sd->groups);
4845 printk(KERN_CONT "\n");
4847 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4848 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4851 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4852 printk(KERN_ERR "ERROR: parent span is not a superset "
4853 "of domain->span\n");
4857 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4861 if (!sched_debug_enabled)
4865 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4869 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4872 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4880 #else /* !CONFIG_SCHED_DEBUG */
4881 # define sched_domain_debug(sd, cpu) do { } while (0)
4882 static inline bool sched_debug(void)
4886 #endif /* CONFIG_SCHED_DEBUG */
4888 static int sd_degenerate(struct sched_domain *sd)
4890 if (cpumask_weight(sched_domain_span(sd)) == 1)
4893 /* Following flags need at least 2 groups */
4894 if (sd->flags & (SD_LOAD_BALANCE |
4895 SD_BALANCE_NEWIDLE |
4899 SD_SHARE_PKG_RESOURCES)) {
4900 if (sd->groups != sd->groups->next)
4904 /* Following flags don't use groups */
4905 if (sd->flags & (SD_WAKE_AFFINE))
4912 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4914 unsigned long cflags = sd->flags, pflags = parent->flags;
4916 if (sd_degenerate(parent))
4919 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4922 /* Flags needing groups don't count if only 1 group in parent */
4923 if (parent->groups == parent->groups->next) {
4924 pflags &= ~(SD_LOAD_BALANCE |
4925 SD_BALANCE_NEWIDLE |
4929 SD_SHARE_PKG_RESOURCES);
4930 if (nr_node_ids == 1)
4931 pflags &= ~SD_SERIALIZE;
4933 if (~cflags & pflags)
4939 static void free_rootdomain(struct rcu_head *rcu)
4941 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4943 cpupri_cleanup(&rd->cpupri);
4944 free_cpumask_var(rd->rto_mask);
4945 free_cpumask_var(rd->online);
4946 free_cpumask_var(rd->span);
4950 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4952 struct root_domain *old_rd = NULL;
4953 unsigned long flags;
4955 raw_spin_lock_irqsave(&rq->lock, flags);
4960 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4963 cpumask_clear_cpu(rq->cpu, old_rd->span);
4966 * If we dont want to free the old_rt yet then
4967 * set old_rd to NULL to skip the freeing later
4970 if (!atomic_dec_and_test(&old_rd->refcount))
4974 atomic_inc(&rd->refcount);
4977 cpumask_set_cpu(rq->cpu, rd->span);
4978 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4981 raw_spin_unlock_irqrestore(&rq->lock, flags);
4984 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4987 static int init_rootdomain(struct root_domain *rd)
4989 memset(rd, 0, sizeof(*rd));
4991 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4993 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4995 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4998 if (cpupri_init(&rd->cpupri) != 0)
5003 free_cpumask_var(rd->rto_mask);
5005 free_cpumask_var(rd->online);
5007 free_cpumask_var(rd->span);
5013 * By default the system creates a single root-domain with all cpus as
5014 * members (mimicking the global state we have today).
5016 struct root_domain def_root_domain;
5018 static void init_defrootdomain(void)
5020 init_rootdomain(&def_root_domain);
5022 atomic_set(&def_root_domain.refcount, 1);
5025 static struct root_domain *alloc_rootdomain(void)
5027 struct root_domain *rd;
5029 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5033 if (init_rootdomain(rd) != 0) {
5041 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5043 struct sched_group *tmp, *first;
5052 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5057 } while (sg != first);
5060 static void free_sched_domain(struct rcu_head *rcu)
5062 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5065 * If its an overlapping domain it has private groups, iterate and
5068 if (sd->flags & SD_OVERLAP) {
5069 free_sched_groups(sd->groups, 1);
5070 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5071 kfree(sd->groups->sgp);
5077 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5079 call_rcu(&sd->rcu, free_sched_domain);
5082 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5084 for (; sd; sd = sd->parent)
5085 destroy_sched_domain(sd, cpu);
5089 * Keep a special pointer to the highest sched_domain that has
5090 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5091 * allows us to avoid some pointer chasing select_idle_sibling().
5093 * Also keep a unique ID per domain (we use the first cpu number in
5094 * the cpumask of the domain), this allows us to quickly tell if
5095 * two cpus are in the same cache domain, see cpus_share_cache().
5097 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5098 DEFINE_PER_CPU(int, sd_llc_id);
5100 static void update_top_cache_domain(int cpu)
5102 struct sched_domain *sd;
5105 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5107 id = cpumask_first(sched_domain_span(sd));
5109 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5110 per_cpu(sd_llc_id, cpu) = id;
5114 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5115 * hold the hotplug lock.
5118 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5120 struct rq *rq = cpu_rq(cpu);
5121 struct sched_domain *tmp;
5123 /* Remove the sched domains which do not contribute to scheduling. */
5124 for (tmp = sd; tmp; ) {
5125 struct sched_domain *parent = tmp->parent;
5129 if (sd_parent_degenerate(tmp, parent)) {
5130 tmp->parent = parent->parent;
5132 parent->parent->child = tmp;
5133 destroy_sched_domain(parent, cpu);
5138 if (sd && sd_degenerate(sd)) {
5141 destroy_sched_domain(tmp, cpu);
5146 sched_domain_debug(sd, cpu);
5148 rq_attach_root(rq, rd);
5150 rcu_assign_pointer(rq->sd, sd);
5151 destroy_sched_domains(tmp, cpu);
5153 update_top_cache_domain(cpu);
5156 /* cpus with isolated domains */
5157 static cpumask_var_t cpu_isolated_map;
5159 /* Setup the mask of cpus configured for isolated domains */
5160 static int __init isolated_cpu_setup(char *str)
5162 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5163 cpulist_parse(str, cpu_isolated_map);
5167 __setup("isolcpus=", isolated_cpu_setup);
5169 static const struct cpumask *cpu_cpu_mask(int cpu)
5171 return cpumask_of_node(cpu_to_node(cpu));
5175 struct sched_domain **__percpu sd;
5176 struct sched_group **__percpu sg;
5177 struct sched_group_power **__percpu sgp;
5181 struct sched_domain ** __percpu sd;
5182 struct root_domain *rd;
5192 struct sched_domain_topology_level;
5194 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5195 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5197 #define SDTL_OVERLAP 0x01
5199 struct sched_domain_topology_level {
5200 sched_domain_init_f init;
5201 sched_domain_mask_f mask;
5204 struct sd_data data;
5208 * Build an iteration mask that can exclude certain CPUs from the upwards
5211 * Asymmetric node setups can result in situations where the domain tree is of
5212 * unequal depth, make sure to skip domains that already cover the entire
5215 * In that case build_sched_domains() will have terminated the iteration early
5216 * and our sibling sd spans will be empty. Domains should always include the
5217 * cpu they're built on, so check that.
5220 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5222 const struct cpumask *span = sched_domain_span(sd);
5223 struct sd_data *sdd = sd->private;
5224 struct sched_domain *sibling;
5227 for_each_cpu(i, span) {
5228 sibling = *per_cpu_ptr(sdd->sd, i);
5229 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5232 cpumask_set_cpu(i, sched_group_mask(sg));
5237 * Return the canonical balance cpu for this group, this is the first cpu
5238 * of this group that's also in the iteration mask.
5240 int group_balance_cpu(struct sched_group *sg)
5242 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5246 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5248 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5249 const struct cpumask *span = sched_domain_span(sd);
5250 struct cpumask *covered = sched_domains_tmpmask;
5251 struct sd_data *sdd = sd->private;
5252 struct sched_domain *child;
5255 cpumask_clear(covered);
5257 for_each_cpu(i, span) {
5258 struct cpumask *sg_span;
5260 if (cpumask_test_cpu(i, covered))
5263 child = *per_cpu_ptr(sdd->sd, i);
5265 /* See the comment near build_group_mask(). */
5266 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5269 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5270 GFP_KERNEL, cpu_to_node(cpu));
5275 sg_span = sched_group_cpus(sg);
5277 child = child->child;
5278 cpumask_copy(sg_span, sched_domain_span(child));
5280 cpumask_set_cpu(i, sg_span);
5282 cpumask_or(covered, covered, sg_span);
5284 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5285 if (atomic_inc_return(&sg->sgp->ref) == 1)
5286 build_group_mask(sd, sg);
5289 * Initialize sgp->power such that even if we mess up the
5290 * domains and no possible iteration will get us here, we won't
5293 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5296 * Make sure the first group of this domain contains the
5297 * canonical balance cpu. Otherwise the sched_domain iteration
5298 * breaks. See update_sg_lb_stats().
5300 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5301 group_balance_cpu(sg) == cpu)
5311 sd->groups = groups;
5316 free_sched_groups(first, 0);
5321 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5323 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5324 struct sched_domain *child = sd->child;
5327 cpu = cpumask_first(sched_domain_span(child));
5330 *sg = *per_cpu_ptr(sdd->sg, cpu);
5331 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5332 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5339 * build_sched_groups will build a circular linked list of the groups
5340 * covered by the given span, and will set each group's ->cpumask correctly,
5341 * and ->cpu_power to 0.
5343 * Assumes the sched_domain tree is fully constructed
5346 build_sched_groups(struct sched_domain *sd, int cpu)
5348 struct sched_group *first = NULL, *last = NULL;
5349 struct sd_data *sdd = sd->private;
5350 const struct cpumask *span = sched_domain_span(sd);
5351 struct cpumask *covered;
5354 get_group(cpu, sdd, &sd->groups);
5355 atomic_inc(&sd->groups->ref);
5357 if (cpu != cpumask_first(span))
5360 lockdep_assert_held(&sched_domains_mutex);
5361 covered = sched_domains_tmpmask;
5363 cpumask_clear(covered);
5365 for_each_cpu(i, span) {
5366 struct sched_group *sg;
5369 if (cpumask_test_cpu(i, covered))
5372 group = get_group(i, sdd, &sg);
5373 cpumask_clear(sched_group_cpus(sg));
5375 cpumask_setall(sched_group_mask(sg));
5377 for_each_cpu(j, span) {
5378 if (get_group(j, sdd, NULL) != group)
5381 cpumask_set_cpu(j, covered);
5382 cpumask_set_cpu(j, sched_group_cpus(sg));
5397 * Initialize sched groups cpu_power.
5399 * cpu_power indicates the capacity of sched group, which is used while
5400 * distributing the load between different sched groups in a sched domain.
5401 * Typically cpu_power for all the groups in a sched domain will be same unless
5402 * there are asymmetries in the topology. If there are asymmetries, group
5403 * having more cpu_power will pickup more load compared to the group having
5406 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5408 struct sched_group *sg = sd->groups;
5413 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5415 } while (sg != sd->groups);
5417 if (cpu != group_balance_cpu(sg))
5420 update_group_power(sd, cpu);
5421 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5424 int __weak arch_sd_sibling_asym_packing(void)
5426 return 0*SD_ASYM_PACKING;
5430 * Initializers for schedule domains
5431 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5434 #ifdef CONFIG_SCHED_DEBUG
5435 # define SD_INIT_NAME(sd, type) sd->name = #type
5437 # define SD_INIT_NAME(sd, type) do { } while (0)
5440 #define SD_INIT_FUNC(type) \
5441 static noinline struct sched_domain * \
5442 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5444 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5445 *sd = SD_##type##_INIT; \
5446 SD_INIT_NAME(sd, type); \
5447 sd->private = &tl->data; \
5452 #ifdef CONFIG_SCHED_SMT
5453 SD_INIT_FUNC(SIBLING)
5455 #ifdef CONFIG_SCHED_MC
5458 #ifdef CONFIG_SCHED_BOOK
5462 static int default_relax_domain_level = -1;
5463 int sched_domain_level_max;
5465 static int __init setup_relax_domain_level(char *str)
5467 if (kstrtoint(str, 0, &default_relax_domain_level))
5468 pr_warn("Unable to set relax_domain_level\n");
5472 __setup("relax_domain_level=", setup_relax_domain_level);
5474 static void set_domain_attribute(struct sched_domain *sd,
5475 struct sched_domain_attr *attr)
5479 if (!attr || attr->relax_domain_level < 0) {
5480 if (default_relax_domain_level < 0)
5483 request = default_relax_domain_level;
5485 request = attr->relax_domain_level;
5486 if (request < sd->level) {
5487 /* turn off idle balance on this domain */
5488 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5490 /* turn on idle balance on this domain */
5491 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5495 static void __sdt_free(const struct cpumask *cpu_map);
5496 static int __sdt_alloc(const struct cpumask *cpu_map);
5498 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5499 const struct cpumask *cpu_map)
5503 if (!atomic_read(&d->rd->refcount))
5504 free_rootdomain(&d->rd->rcu); /* fall through */
5506 free_percpu(d->sd); /* fall through */
5508 __sdt_free(cpu_map); /* fall through */
5514 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5515 const struct cpumask *cpu_map)
5517 memset(d, 0, sizeof(*d));
5519 if (__sdt_alloc(cpu_map))
5520 return sa_sd_storage;
5521 d->sd = alloc_percpu(struct sched_domain *);
5523 return sa_sd_storage;
5524 d->rd = alloc_rootdomain();
5527 return sa_rootdomain;
5531 * NULL the sd_data elements we've used to build the sched_domain and
5532 * sched_group structure so that the subsequent __free_domain_allocs()
5533 * will not free the data we're using.
5535 static void claim_allocations(int cpu, struct sched_domain *sd)
5537 struct sd_data *sdd = sd->private;
5539 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5540 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5542 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5543 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5545 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5546 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5549 #ifdef CONFIG_SCHED_SMT
5550 static const struct cpumask *cpu_smt_mask(int cpu)
5552 return topology_thread_cpumask(cpu);
5557 * Topology list, bottom-up.
5559 static struct sched_domain_topology_level default_topology[] = {
5560 #ifdef CONFIG_SCHED_SMT
5561 { sd_init_SIBLING, cpu_smt_mask, },
5563 #ifdef CONFIG_SCHED_MC
5564 { sd_init_MC, cpu_coregroup_mask, },
5566 #ifdef CONFIG_SCHED_BOOK
5567 { sd_init_BOOK, cpu_book_mask, },
5569 { sd_init_CPU, cpu_cpu_mask, },
5573 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5575 #define for_each_sd_topology(tl) \
5576 for (tl = sched_domain_topology; tl->init; tl++)
5580 static int sched_domains_numa_levels;
5581 static int *sched_domains_numa_distance;
5582 static struct cpumask ***sched_domains_numa_masks;
5583 static int sched_domains_curr_level;
5585 static inline int sd_local_flags(int level)
5587 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5590 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5593 static struct sched_domain *
5594 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5596 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5597 int level = tl->numa_level;
5598 int sd_weight = cpumask_weight(
5599 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5601 *sd = (struct sched_domain){
5602 .min_interval = sd_weight,
5603 .max_interval = 2*sd_weight,
5605 .imbalance_pct = 125,
5606 .cache_nice_tries = 2,
5613 .flags = 1*SD_LOAD_BALANCE
5614 | 1*SD_BALANCE_NEWIDLE
5619 | 0*SD_SHARE_CPUPOWER
5620 | 0*SD_SHARE_PKG_RESOURCES
5622 | 0*SD_PREFER_SIBLING
5623 | sd_local_flags(level)
5625 .last_balance = jiffies,
5626 .balance_interval = sd_weight,
5628 SD_INIT_NAME(sd, NUMA);
5629 sd->private = &tl->data;
5632 * Ugly hack to pass state to sd_numa_mask()...
5634 sched_domains_curr_level = tl->numa_level;
5639 static const struct cpumask *sd_numa_mask(int cpu)
5641 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5644 static void sched_numa_warn(const char *str)
5646 static int done = false;
5654 printk(KERN_WARNING "ERROR: %s\n\n", str);
5656 for (i = 0; i < nr_node_ids; i++) {
5657 printk(KERN_WARNING " ");
5658 for (j = 0; j < nr_node_ids; j++)
5659 printk(KERN_CONT "%02d ", node_distance(i,j));
5660 printk(KERN_CONT "\n");
5662 printk(KERN_WARNING "\n");
5665 static bool find_numa_distance(int distance)
5669 if (distance == node_distance(0, 0))
5672 for (i = 0; i < sched_domains_numa_levels; i++) {
5673 if (sched_domains_numa_distance[i] == distance)
5680 static void sched_init_numa(void)
5682 int next_distance, curr_distance = node_distance(0, 0);
5683 struct sched_domain_topology_level *tl;
5687 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5688 if (!sched_domains_numa_distance)
5692 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5693 * unique distances in the node_distance() table.
5695 * Assumes node_distance(0,j) includes all distances in
5696 * node_distance(i,j) in order to avoid cubic time.
5698 next_distance = curr_distance;
5699 for (i = 0; i < nr_node_ids; i++) {
5700 for (j = 0; j < nr_node_ids; j++) {
5701 for (k = 0; k < nr_node_ids; k++) {
5702 int distance = node_distance(i, k);
5704 if (distance > curr_distance &&
5705 (distance < next_distance ||
5706 next_distance == curr_distance))
5707 next_distance = distance;
5710 * While not a strong assumption it would be nice to know
5711 * about cases where if node A is connected to B, B is not
5712 * equally connected to A.
5714 if (sched_debug() && node_distance(k, i) != distance)
5715 sched_numa_warn("Node-distance not symmetric");
5717 if (sched_debug() && i && !find_numa_distance(distance))
5718 sched_numa_warn("Node-0 not representative");
5720 if (next_distance != curr_distance) {
5721 sched_domains_numa_distance[level++] = next_distance;
5722 sched_domains_numa_levels = level;
5723 curr_distance = next_distance;
5728 * In case of sched_debug() we verify the above assumption.
5734 * 'level' contains the number of unique distances, excluding the
5735 * identity distance node_distance(i,i).
5737 * The sched_domains_numa_distance[] array includes the actual distance
5742 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5743 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5744 * the array will contain less then 'level' members. This could be
5745 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5746 * in other functions.
5748 * We reset it to 'level' at the end of this function.
5750 sched_domains_numa_levels = 0;
5752 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5753 if (!sched_domains_numa_masks)
5757 * Now for each level, construct a mask per node which contains all
5758 * cpus of nodes that are that many hops away from us.
5760 for (i = 0; i < level; i++) {
5761 sched_domains_numa_masks[i] =
5762 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5763 if (!sched_domains_numa_masks[i])
5766 for (j = 0; j < nr_node_ids; j++) {
5767 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5771 sched_domains_numa_masks[i][j] = mask;
5773 for (k = 0; k < nr_node_ids; k++) {
5774 if (node_distance(j, k) > sched_domains_numa_distance[i])
5777 cpumask_or(mask, mask, cpumask_of_node(k));
5782 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5783 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5788 * Copy the default topology bits..
5790 for (i = 0; default_topology[i].init; i++)
5791 tl[i] = default_topology[i];
5794 * .. and append 'j' levels of NUMA goodness.
5796 for (j = 0; j < level; i++, j++) {
5797 tl[i] = (struct sched_domain_topology_level){
5798 .init = sd_numa_init,
5799 .mask = sd_numa_mask,
5800 .flags = SDTL_OVERLAP,
5805 sched_domain_topology = tl;
5807 sched_domains_numa_levels = level;
5810 static void sched_domains_numa_masks_set(int cpu)
5813 int node = cpu_to_node(cpu);
5815 for (i = 0; i < sched_domains_numa_levels; i++) {
5816 for (j = 0; j < nr_node_ids; j++) {
5817 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5818 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5823 static void sched_domains_numa_masks_clear(int cpu)
5826 for (i = 0; i < sched_domains_numa_levels; i++) {
5827 for (j = 0; j < nr_node_ids; j++)
5828 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5833 * Update sched_domains_numa_masks[level][node] array when new cpus
5836 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5837 unsigned long action,
5840 int cpu = (long)hcpu;
5842 switch (action & ~CPU_TASKS_FROZEN) {
5844 sched_domains_numa_masks_set(cpu);
5848 sched_domains_numa_masks_clear(cpu);
5858 static inline void sched_init_numa(void)
5862 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5863 unsigned long action,
5868 #endif /* CONFIG_NUMA */
5870 static int __sdt_alloc(const struct cpumask *cpu_map)
5872 struct sched_domain_topology_level *tl;
5875 for_each_sd_topology(tl) {
5876 struct sd_data *sdd = &tl->data;
5878 sdd->sd = alloc_percpu(struct sched_domain *);
5882 sdd->sg = alloc_percpu(struct sched_group *);
5886 sdd->sgp = alloc_percpu(struct sched_group_power *);
5890 for_each_cpu(j, cpu_map) {
5891 struct sched_domain *sd;
5892 struct sched_group *sg;
5893 struct sched_group_power *sgp;
5895 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5896 GFP_KERNEL, cpu_to_node(j));
5900 *per_cpu_ptr(sdd->sd, j) = sd;
5902 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5903 GFP_KERNEL, cpu_to_node(j));
5909 *per_cpu_ptr(sdd->sg, j) = sg;
5911 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5912 GFP_KERNEL, cpu_to_node(j));
5916 *per_cpu_ptr(sdd->sgp, j) = sgp;
5923 static void __sdt_free(const struct cpumask *cpu_map)
5925 struct sched_domain_topology_level *tl;
5928 for_each_sd_topology(tl) {
5929 struct sd_data *sdd = &tl->data;
5931 for_each_cpu(j, cpu_map) {
5932 struct sched_domain *sd;
5935 sd = *per_cpu_ptr(sdd->sd, j);
5936 if (sd && (sd->flags & SD_OVERLAP))
5937 free_sched_groups(sd->groups, 0);
5938 kfree(*per_cpu_ptr(sdd->sd, j));
5942 kfree(*per_cpu_ptr(sdd->sg, j));
5944 kfree(*per_cpu_ptr(sdd->sgp, j));
5946 free_percpu(sdd->sd);
5948 free_percpu(sdd->sg);
5950 free_percpu(sdd->sgp);
5955 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5956 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5957 struct sched_domain *child, int cpu)
5959 struct sched_domain *sd = tl->init(tl, cpu);
5963 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5965 sd->level = child->level + 1;
5966 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5970 set_domain_attribute(sd, attr);
5976 * Build sched domains for a given set of cpus and attach the sched domains
5977 * to the individual cpus
5979 static int build_sched_domains(const struct cpumask *cpu_map,
5980 struct sched_domain_attr *attr)
5982 enum s_alloc alloc_state;
5983 struct sched_domain *sd;
5985 int i, ret = -ENOMEM;
5987 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5988 if (alloc_state != sa_rootdomain)
5991 /* Set up domains for cpus specified by the cpu_map. */
5992 for_each_cpu(i, cpu_map) {
5993 struct sched_domain_topology_level *tl;
5996 for_each_sd_topology(tl) {
5997 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
5998 if (tl == sched_domain_topology)
5999 *per_cpu_ptr(d.sd, i) = sd;
6000 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6001 sd->flags |= SD_OVERLAP;
6002 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6007 /* Build the groups for the domains */
6008 for_each_cpu(i, cpu_map) {
6009 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6010 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6011 if (sd->flags & SD_OVERLAP) {
6012 if (build_overlap_sched_groups(sd, i))
6015 if (build_sched_groups(sd, i))
6021 /* Calculate CPU power for physical packages and nodes */
6022 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6023 if (!cpumask_test_cpu(i, cpu_map))
6026 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6027 claim_allocations(i, sd);
6028 init_sched_groups_power(i, sd);
6032 /* Attach the domains */
6034 for_each_cpu(i, cpu_map) {
6035 sd = *per_cpu_ptr(d.sd, i);
6036 cpu_attach_domain(sd, d.rd, i);
6042 __free_domain_allocs(&d, alloc_state, cpu_map);
6046 static cpumask_var_t *doms_cur; /* current sched domains */
6047 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6048 static struct sched_domain_attr *dattr_cur;
6049 /* attribues of custom domains in 'doms_cur' */
6052 * Special case: If a kmalloc of a doms_cur partition (array of
6053 * cpumask) fails, then fallback to a single sched domain,
6054 * as determined by the single cpumask fallback_doms.
6056 static cpumask_var_t fallback_doms;
6059 * arch_update_cpu_topology lets virtualized architectures update the
6060 * cpu core maps. It is supposed to return 1 if the topology changed
6061 * or 0 if it stayed the same.
6063 int __attribute__((weak)) arch_update_cpu_topology(void)
6068 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6071 cpumask_var_t *doms;
6073 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6076 for (i = 0; i < ndoms; i++) {
6077 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6078 free_sched_domains(doms, i);
6085 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6088 for (i = 0; i < ndoms; i++)
6089 free_cpumask_var(doms[i]);
6094 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6095 * For now this just excludes isolated cpus, but could be used to
6096 * exclude other special cases in the future.
6098 static int init_sched_domains(const struct cpumask *cpu_map)
6102 arch_update_cpu_topology();
6104 doms_cur = alloc_sched_domains(ndoms_cur);
6106 doms_cur = &fallback_doms;
6107 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6108 err = build_sched_domains(doms_cur[0], NULL);
6109 register_sched_domain_sysctl();
6115 * Detach sched domains from a group of cpus specified in cpu_map
6116 * These cpus will now be attached to the NULL domain
6118 static void detach_destroy_domains(const struct cpumask *cpu_map)
6123 for_each_cpu(i, cpu_map)
6124 cpu_attach_domain(NULL, &def_root_domain, i);
6128 /* handle null as "default" */
6129 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6130 struct sched_domain_attr *new, int idx_new)
6132 struct sched_domain_attr tmp;
6139 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6140 new ? (new + idx_new) : &tmp,
6141 sizeof(struct sched_domain_attr));
6145 * Partition sched domains as specified by the 'ndoms_new'
6146 * cpumasks in the array doms_new[] of cpumasks. This compares
6147 * doms_new[] to the current sched domain partitioning, doms_cur[].
6148 * It destroys each deleted domain and builds each new domain.
6150 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6151 * The masks don't intersect (don't overlap.) We should setup one
6152 * sched domain for each mask. CPUs not in any of the cpumasks will
6153 * not be load balanced. If the same cpumask appears both in the
6154 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6157 * The passed in 'doms_new' should be allocated using
6158 * alloc_sched_domains. This routine takes ownership of it and will
6159 * free_sched_domains it when done with it. If the caller failed the
6160 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6161 * and partition_sched_domains() will fallback to the single partition
6162 * 'fallback_doms', it also forces the domains to be rebuilt.
6164 * If doms_new == NULL it will be replaced with cpu_online_mask.
6165 * ndoms_new == 0 is a special case for destroying existing domains,
6166 * and it will not create the default domain.
6168 * Call with hotplug lock held
6170 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6171 struct sched_domain_attr *dattr_new)
6176 mutex_lock(&sched_domains_mutex);
6178 /* always unregister in case we don't destroy any domains */
6179 unregister_sched_domain_sysctl();
6181 /* Let architecture update cpu core mappings. */
6182 new_topology = arch_update_cpu_topology();
6184 n = doms_new ? ndoms_new : 0;
6186 /* Destroy deleted domains */
6187 for (i = 0; i < ndoms_cur; i++) {
6188 for (j = 0; j < n && !new_topology; j++) {
6189 if (cpumask_equal(doms_cur[i], doms_new[j])
6190 && dattrs_equal(dattr_cur, i, dattr_new, j))
6193 /* no match - a current sched domain not in new doms_new[] */
6194 detach_destroy_domains(doms_cur[i]);
6199 if (doms_new == NULL) {
6201 doms_new = &fallback_doms;
6202 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6203 WARN_ON_ONCE(dattr_new);
6206 /* Build new domains */
6207 for (i = 0; i < ndoms_new; i++) {
6208 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6209 if (cpumask_equal(doms_new[i], doms_cur[j])
6210 && dattrs_equal(dattr_new, i, dattr_cur, j))
6213 /* no match - add a new doms_new */
6214 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6219 /* Remember the new sched domains */
6220 if (doms_cur != &fallback_doms)
6221 free_sched_domains(doms_cur, ndoms_cur);
6222 kfree(dattr_cur); /* kfree(NULL) is safe */
6223 doms_cur = doms_new;
6224 dattr_cur = dattr_new;
6225 ndoms_cur = ndoms_new;
6227 register_sched_domain_sysctl();
6229 mutex_unlock(&sched_domains_mutex);
6232 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6235 * Update cpusets according to cpu_active mask. If cpusets are
6236 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6237 * around partition_sched_domains().
6239 * If we come here as part of a suspend/resume, don't touch cpusets because we
6240 * want to restore it back to its original state upon resume anyway.
6242 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6246 case CPU_ONLINE_FROZEN:
6247 case CPU_DOWN_FAILED_FROZEN:
6250 * num_cpus_frozen tracks how many CPUs are involved in suspend
6251 * resume sequence. As long as this is not the last online
6252 * operation in the resume sequence, just build a single sched
6253 * domain, ignoring cpusets.
6256 if (likely(num_cpus_frozen)) {
6257 partition_sched_domains(1, NULL, NULL);
6262 * This is the last CPU online operation. So fall through and
6263 * restore the original sched domains by considering the
6264 * cpuset configurations.
6268 case CPU_DOWN_FAILED:
6269 cpuset_update_active_cpus(true);
6277 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6281 case CPU_DOWN_PREPARE:
6282 cpuset_update_active_cpus(false);
6284 case CPU_DOWN_PREPARE_FROZEN:
6286 partition_sched_domains(1, NULL, NULL);
6294 void __init sched_init_smp(void)
6296 cpumask_var_t non_isolated_cpus;
6298 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6299 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6304 mutex_lock(&sched_domains_mutex);
6305 init_sched_domains(cpu_active_mask);
6306 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6307 if (cpumask_empty(non_isolated_cpus))
6308 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6309 mutex_unlock(&sched_domains_mutex);
6312 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6313 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6314 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6318 /* Move init over to a non-isolated CPU */
6319 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6321 sched_init_granularity();
6322 free_cpumask_var(non_isolated_cpus);
6324 init_sched_rt_class();
6327 void __init sched_init_smp(void)
6329 sched_init_granularity();
6331 #endif /* CONFIG_SMP */
6333 const_debug unsigned int sysctl_timer_migration = 1;
6335 int in_sched_functions(unsigned long addr)
6337 return in_lock_functions(addr) ||
6338 (addr >= (unsigned long)__sched_text_start
6339 && addr < (unsigned long)__sched_text_end);
6342 #ifdef CONFIG_CGROUP_SCHED
6344 * Default task group.
6345 * Every task in system belongs to this group at bootup.
6347 struct task_group root_task_group;
6348 LIST_HEAD(task_groups);
6351 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6353 void __init sched_init(void)
6356 unsigned long alloc_size = 0, ptr;
6358 #ifdef CONFIG_FAIR_GROUP_SCHED
6359 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6361 #ifdef CONFIG_RT_GROUP_SCHED
6362 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6364 #ifdef CONFIG_CPUMASK_OFFSTACK
6365 alloc_size += num_possible_cpus() * cpumask_size();
6368 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6370 #ifdef CONFIG_FAIR_GROUP_SCHED
6371 root_task_group.se = (struct sched_entity **)ptr;
6372 ptr += nr_cpu_ids * sizeof(void **);
6374 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6375 ptr += nr_cpu_ids * sizeof(void **);
6377 #endif /* CONFIG_FAIR_GROUP_SCHED */
6378 #ifdef CONFIG_RT_GROUP_SCHED
6379 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6380 ptr += nr_cpu_ids * sizeof(void **);
6382 root_task_group.rt_rq = (struct rt_rq **)ptr;
6383 ptr += nr_cpu_ids * sizeof(void **);
6385 #endif /* CONFIG_RT_GROUP_SCHED */
6386 #ifdef CONFIG_CPUMASK_OFFSTACK
6387 for_each_possible_cpu(i) {
6388 per_cpu(load_balance_mask, i) = (void *)ptr;
6389 ptr += cpumask_size();
6391 #endif /* CONFIG_CPUMASK_OFFSTACK */
6395 init_defrootdomain();
6398 init_rt_bandwidth(&def_rt_bandwidth,
6399 global_rt_period(), global_rt_runtime());
6401 #ifdef CONFIG_RT_GROUP_SCHED
6402 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6403 global_rt_period(), global_rt_runtime());
6404 #endif /* CONFIG_RT_GROUP_SCHED */
6406 #ifdef CONFIG_CGROUP_SCHED
6407 list_add(&root_task_group.list, &task_groups);
6408 INIT_LIST_HEAD(&root_task_group.children);
6409 INIT_LIST_HEAD(&root_task_group.siblings);
6410 autogroup_init(&init_task);
6412 #endif /* CONFIG_CGROUP_SCHED */
6414 for_each_possible_cpu(i) {
6418 raw_spin_lock_init(&rq->lock);
6420 rq->calc_load_active = 0;
6421 rq->calc_load_update = jiffies + LOAD_FREQ;
6422 init_cfs_rq(&rq->cfs);
6423 init_rt_rq(&rq->rt, rq);
6424 #ifdef CONFIG_FAIR_GROUP_SCHED
6425 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6426 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6428 * How much cpu bandwidth does root_task_group get?
6430 * In case of task-groups formed thr' the cgroup filesystem, it
6431 * gets 100% of the cpu resources in the system. This overall
6432 * system cpu resource is divided among the tasks of
6433 * root_task_group and its child task-groups in a fair manner,
6434 * based on each entity's (task or task-group's) weight
6435 * (se->load.weight).
6437 * In other words, if root_task_group has 10 tasks of weight
6438 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6439 * then A0's share of the cpu resource is:
6441 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6443 * We achieve this by letting root_task_group's tasks sit
6444 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6446 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6447 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6448 #endif /* CONFIG_FAIR_GROUP_SCHED */
6450 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6451 #ifdef CONFIG_RT_GROUP_SCHED
6452 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6453 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6456 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6457 rq->cpu_load[j] = 0;
6459 rq->last_load_update_tick = jiffies;
6464 rq->cpu_power = SCHED_POWER_SCALE;
6465 rq->post_schedule = 0;
6466 rq->active_balance = 0;
6467 rq->next_balance = jiffies;
6472 rq->avg_idle = 2*sysctl_sched_migration_cost;
6474 INIT_LIST_HEAD(&rq->cfs_tasks);
6476 rq_attach_root(rq, &def_root_domain);
6477 #ifdef CONFIG_NO_HZ_COMMON
6480 #ifdef CONFIG_NO_HZ_FULL
6481 rq->last_sched_tick = 0;
6485 atomic_set(&rq->nr_iowait, 0);
6488 set_load_weight(&init_task);
6490 #ifdef CONFIG_PREEMPT_NOTIFIERS
6491 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6494 #ifdef CONFIG_RT_MUTEXES
6495 plist_head_init(&init_task.pi_waiters);
6499 * The boot idle thread does lazy MMU switching as well:
6501 atomic_inc(&init_mm.mm_count);
6502 enter_lazy_tlb(&init_mm, current);
6505 * Make us the idle thread. Technically, schedule() should not be
6506 * called from this thread, however somewhere below it might be,
6507 * but because we are the idle thread, we just pick up running again
6508 * when this runqueue becomes "idle".
6510 init_idle(current, smp_processor_id());
6512 calc_load_update = jiffies + LOAD_FREQ;
6515 * During early bootup we pretend to be a normal task:
6517 current->sched_class = &fair_sched_class;
6520 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6521 /* May be allocated at isolcpus cmdline parse time */
6522 if (cpu_isolated_map == NULL)
6523 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6524 idle_thread_set_boot_cpu();
6526 init_sched_fair_class();
6528 scheduler_running = 1;
6531 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6532 static inline int preempt_count_equals(int preempt_offset)
6534 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6536 return (nested == preempt_offset);
6539 void __might_sleep(const char *file, int line, int preempt_offset)
6541 static unsigned long prev_jiffy; /* ratelimiting */
6543 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6544 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6545 system_state != SYSTEM_RUNNING || oops_in_progress)
6547 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6549 prev_jiffy = jiffies;
6552 "BUG: sleeping function called from invalid context at %s:%d\n",
6555 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6556 in_atomic(), irqs_disabled(),
6557 current->pid, current->comm);
6559 debug_show_held_locks(current);
6560 if (irqs_disabled())
6561 print_irqtrace_events(current);
6564 EXPORT_SYMBOL(__might_sleep);
6567 #ifdef CONFIG_MAGIC_SYSRQ
6568 static void normalize_task(struct rq *rq, struct task_struct *p)
6570 const struct sched_class *prev_class = p->sched_class;
6571 int old_prio = p->prio;
6576 dequeue_task(rq, p, 0);
6577 __setscheduler(rq, p, SCHED_NORMAL, 0);
6579 enqueue_task(rq, p, 0);
6580 resched_task(rq->curr);
6583 check_class_changed(rq, p, prev_class, old_prio);
6586 void normalize_rt_tasks(void)
6588 struct task_struct *g, *p;
6589 unsigned long flags;
6592 read_lock_irqsave(&tasklist_lock, flags);
6593 do_each_thread(g, p) {
6595 * Only normalize user tasks:
6600 p->se.exec_start = 0;
6601 #ifdef CONFIG_SCHEDSTATS
6602 p->se.statistics.wait_start = 0;
6603 p->se.statistics.sleep_start = 0;
6604 p->se.statistics.block_start = 0;
6609 * Renice negative nice level userspace
6612 if (TASK_NICE(p) < 0 && p->mm)
6613 set_user_nice(p, 0);
6617 raw_spin_lock(&p->pi_lock);
6618 rq = __task_rq_lock(p);
6620 normalize_task(rq, p);
6622 __task_rq_unlock(rq);
6623 raw_spin_unlock(&p->pi_lock);
6624 } while_each_thread(g, p);
6626 read_unlock_irqrestore(&tasklist_lock, flags);
6629 #endif /* CONFIG_MAGIC_SYSRQ */
6631 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6633 * These functions are only useful for the IA64 MCA handling, or kdb.
6635 * They can only be called when the whole system has been
6636 * stopped - every CPU needs to be quiescent, and no scheduling
6637 * activity can take place. Using them for anything else would
6638 * be a serious bug, and as a result, they aren't even visible
6639 * under any other configuration.
6643 * curr_task - return the current task for a given cpu.
6644 * @cpu: the processor in question.
6646 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6648 struct task_struct *curr_task(int cpu)
6650 return cpu_curr(cpu);
6653 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6657 * set_curr_task - set the current task for a given cpu.
6658 * @cpu: the processor in question.
6659 * @p: the task pointer to set.
6661 * Description: This function must only be used when non-maskable interrupts
6662 * are serviced on a separate stack. It allows the architecture to switch the
6663 * notion of the current task on a cpu in a non-blocking manner. This function
6664 * must be called with all CPU's synchronized, and interrupts disabled, the
6665 * and caller must save the original value of the current task (see
6666 * curr_task() above) and restore that value before reenabling interrupts and
6667 * re-starting the system.
6669 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6671 void set_curr_task(int cpu, struct task_struct *p)
6678 #ifdef CONFIG_CGROUP_SCHED
6679 /* task_group_lock serializes the addition/removal of task groups */
6680 static DEFINE_SPINLOCK(task_group_lock);
6682 static void free_sched_group(struct task_group *tg)
6684 free_fair_sched_group(tg);
6685 free_rt_sched_group(tg);
6690 /* allocate runqueue etc for a new task group */
6691 struct task_group *sched_create_group(struct task_group *parent)
6693 struct task_group *tg;
6695 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6697 return ERR_PTR(-ENOMEM);
6699 if (!alloc_fair_sched_group(tg, parent))
6702 if (!alloc_rt_sched_group(tg, parent))
6708 free_sched_group(tg);
6709 return ERR_PTR(-ENOMEM);
6712 void sched_online_group(struct task_group *tg, struct task_group *parent)
6714 unsigned long flags;
6716 spin_lock_irqsave(&task_group_lock, flags);
6717 list_add_rcu(&tg->list, &task_groups);
6719 WARN_ON(!parent); /* root should already exist */
6721 tg->parent = parent;
6722 INIT_LIST_HEAD(&tg->children);
6723 list_add_rcu(&tg->siblings, &parent->children);
6724 spin_unlock_irqrestore(&task_group_lock, flags);
6727 /* rcu callback to free various structures associated with a task group */
6728 static void free_sched_group_rcu(struct rcu_head *rhp)
6730 /* now it should be safe to free those cfs_rqs */
6731 free_sched_group(container_of(rhp, struct task_group, rcu));
6734 /* Destroy runqueue etc associated with a task group */
6735 void sched_destroy_group(struct task_group *tg)
6737 /* wait for possible concurrent references to cfs_rqs complete */
6738 call_rcu(&tg->rcu, free_sched_group_rcu);
6741 void sched_offline_group(struct task_group *tg)
6743 unsigned long flags;
6746 /* end participation in shares distribution */
6747 for_each_possible_cpu(i)
6748 unregister_fair_sched_group(tg, i);
6750 spin_lock_irqsave(&task_group_lock, flags);
6751 list_del_rcu(&tg->list);
6752 list_del_rcu(&tg->siblings);
6753 spin_unlock_irqrestore(&task_group_lock, flags);
6756 /* change task's runqueue when it moves between groups.
6757 * The caller of this function should have put the task in its new group
6758 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6759 * reflect its new group.
6761 void sched_move_task(struct task_struct *tsk)
6763 struct task_group *tg;
6765 unsigned long flags;
6768 rq = task_rq_lock(tsk, &flags);
6770 running = task_current(rq, tsk);
6774 dequeue_task(rq, tsk, 0);
6775 if (unlikely(running))
6776 tsk->sched_class->put_prev_task(rq, tsk);
6778 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
6779 lockdep_is_held(&tsk->sighand->siglock)),
6780 struct task_group, css);
6781 tg = autogroup_task_group(tsk, tg);
6782 tsk->sched_task_group = tg;
6784 #ifdef CONFIG_FAIR_GROUP_SCHED
6785 if (tsk->sched_class->task_move_group)
6786 tsk->sched_class->task_move_group(tsk, on_rq);
6789 set_task_rq(tsk, task_cpu(tsk));
6791 if (unlikely(running))
6792 tsk->sched_class->set_curr_task(rq);
6794 enqueue_task(rq, tsk, 0);
6796 task_rq_unlock(rq, tsk, &flags);
6798 #endif /* CONFIG_CGROUP_SCHED */
6800 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6801 static unsigned long to_ratio(u64 period, u64 runtime)
6803 if (runtime == RUNTIME_INF)
6806 return div64_u64(runtime << 20, period);
6810 #ifdef CONFIG_RT_GROUP_SCHED
6812 * Ensure that the real time constraints are schedulable.
6814 static DEFINE_MUTEX(rt_constraints_mutex);
6816 /* Must be called with tasklist_lock held */
6817 static inline int tg_has_rt_tasks(struct task_group *tg)
6819 struct task_struct *g, *p;
6821 do_each_thread(g, p) {
6822 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6824 } while_each_thread(g, p);
6829 struct rt_schedulable_data {
6830 struct task_group *tg;
6835 static int tg_rt_schedulable(struct task_group *tg, void *data)
6837 struct rt_schedulable_data *d = data;
6838 struct task_group *child;
6839 unsigned long total, sum = 0;
6840 u64 period, runtime;
6842 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6843 runtime = tg->rt_bandwidth.rt_runtime;
6846 period = d->rt_period;
6847 runtime = d->rt_runtime;
6851 * Cannot have more runtime than the period.
6853 if (runtime > period && runtime != RUNTIME_INF)
6857 * Ensure we don't starve existing RT tasks.
6859 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6862 total = to_ratio(period, runtime);
6865 * Nobody can have more than the global setting allows.
6867 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6871 * The sum of our children's runtime should not exceed our own.
6873 list_for_each_entry_rcu(child, &tg->children, siblings) {
6874 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6875 runtime = child->rt_bandwidth.rt_runtime;
6877 if (child == d->tg) {
6878 period = d->rt_period;
6879 runtime = d->rt_runtime;
6882 sum += to_ratio(period, runtime);
6891 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6895 struct rt_schedulable_data data = {
6897 .rt_period = period,
6898 .rt_runtime = runtime,
6902 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6908 static int tg_set_rt_bandwidth(struct task_group *tg,
6909 u64 rt_period, u64 rt_runtime)
6913 mutex_lock(&rt_constraints_mutex);
6914 read_lock(&tasklist_lock);
6915 err = __rt_schedulable(tg, rt_period, rt_runtime);
6919 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6920 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6921 tg->rt_bandwidth.rt_runtime = rt_runtime;
6923 for_each_possible_cpu(i) {
6924 struct rt_rq *rt_rq = tg->rt_rq[i];
6926 raw_spin_lock(&rt_rq->rt_runtime_lock);
6927 rt_rq->rt_runtime = rt_runtime;
6928 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6930 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6932 read_unlock(&tasklist_lock);
6933 mutex_unlock(&rt_constraints_mutex);
6938 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6940 u64 rt_runtime, rt_period;
6942 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6943 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6944 if (rt_runtime_us < 0)
6945 rt_runtime = RUNTIME_INF;
6947 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6950 static long sched_group_rt_runtime(struct task_group *tg)
6954 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6957 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6958 do_div(rt_runtime_us, NSEC_PER_USEC);
6959 return rt_runtime_us;
6962 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6964 u64 rt_runtime, rt_period;
6966 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6967 rt_runtime = tg->rt_bandwidth.rt_runtime;
6972 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6975 static long sched_group_rt_period(struct task_group *tg)
6979 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6980 do_div(rt_period_us, NSEC_PER_USEC);
6981 return rt_period_us;
6984 static int sched_rt_global_constraints(void)
6986 u64 runtime, period;
6989 if (sysctl_sched_rt_period <= 0)
6992 runtime = global_rt_runtime();
6993 period = global_rt_period();
6996 * Sanity check on the sysctl variables.
6998 if (runtime > period && runtime != RUNTIME_INF)
7001 mutex_lock(&rt_constraints_mutex);
7002 read_lock(&tasklist_lock);
7003 ret = __rt_schedulable(NULL, 0, 0);
7004 read_unlock(&tasklist_lock);
7005 mutex_unlock(&rt_constraints_mutex);
7010 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7012 /* Don't accept realtime tasks when there is no way for them to run */
7013 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7019 #else /* !CONFIG_RT_GROUP_SCHED */
7020 static int sched_rt_global_constraints(void)
7022 unsigned long flags;
7025 if (sysctl_sched_rt_period <= 0)
7029 * There's always some RT tasks in the root group
7030 * -- migration, kstopmachine etc..
7032 if (sysctl_sched_rt_runtime == 0)
7035 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7036 for_each_possible_cpu(i) {
7037 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7039 raw_spin_lock(&rt_rq->rt_runtime_lock);
7040 rt_rq->rt_runtime = global_rt_runtime();
7041 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7043 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7047 #endif /* CONFIG_RT_GROUP_SCHED */
7049 int sched_rr_handler(struct ctl_table *table, int write,
7050 void __user *buffer, size_t *lenp,
7054 static DEFINE_MUTEX(mutex);
7057 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7058 /* make sure that internally we keep jiffies */
7059 /* also, writing zero resets timeslice to default */
7060 if (!ret && write) {
7061 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7062 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7064 mutex_unlock(&mutex);
7068 int sched_rt_handler(struct ctl_table *table, int write,
7069 void __user *buffer, size_t *lenp,
7073 int old_period, old_runtime;
7074 static DEFINE_MUTEX(mutex);
7077 old_period = sysctl_sched_rt_period;
7078 old_runtime = sysctl_sched_rt_runtime;
7080 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7082 if (!ret && write) {
7083 ret = sched_rt_global_constraints();
7085 sysctl_sched_rt_period = old_period;
7086 sysctl_sched_rt_runtime = old_runtime;
7088 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7089 def_rt_bandwidth.rt_period =
7090 ns_to_ktime(global_rt_period());
7093 mutex_unlock(&mutex);
7098 #ifdef CONFIG_CGROUP_SCHED
7100 /* return corresponding task_group object of a cgroup */
7101 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7103 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7104 struct task_group, css);
7107 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7109 struct task_group *tg, *parent;
7111 if (!cgrp->parent) {
7112 /* This is early initialization for the top cgroup */
7113 return &root_task_group.css;
7116 parent = cgroup_tg(cgrp->parent);
7117 tg = sched_create_group(parent);
7119 return ERR_PTR(-ENOMEM);
7124 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7126 struct task_group *tg = cgroup_tg(cgrp);
7127 struct task_group *parent;
7132 parent = cgroup_tg(cgrp->parent);
7133 sched_online_group(tg, parent);
7137 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7139 struct task_group *tg = cgroup_tg(cgrp);
7141 sched_destroy_group(tg);
7144 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7146 struct task_group *tg = cgroup_tg(cgrp);
7148 sched_offline_group(tg);
7151 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7152 struct cgroup_taskset *tset)
7154 struct task_struct *task;
7156 cgroup_taskset_for_each(task, cgrp, tset) {
7157 #ifdef CONFIG_RT_GROUP_SCHED
7158 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7161 /* We don't support RT-tasks being in separate groups */
7162 if (task->sched_class != &fair_sched_class)
7169 static void cpu_cgroup_attach(struct cgroup *cgrp,
7170 struct cgroup_taskset *tset)
7172 struct task_struct *task;
7174 cgroup_taskset_for_each(task, cgrp, tset)
7175 sched_move_task(task);
7179 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7180 struct task_struct *task)
7183 * cgroup_exit() is called in the copy_process() failure path.
7184 * Ignore this case since the task hasn't ran yet, this avoids
7185 * trying to poke a half freed task state from generic code.
7187 if (!(task->flags & PF_EXITING))
7190 sched_move_task(task);
7193 #ifdef CONFIG_FAIR_GROUP_SCHED
7194 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7197 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7200 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7202 struct task_group *tg = cgroup_tg(cgrp);
7204 return (u64) scale_load_down(tg->shares);
7207 #ifdef CONFIG_CFS_BANDWIDTH
7208 static DEFINE_MUTEX(cfs_constraints_mutex);
7210 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7211 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7213 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7215 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7217 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7218 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7220 if (tg == &root_task_group)
7224 * Ensure we have at some amount of bandwidth every period. This is
7225 * to prevent reaching a state of large arrears when throttled via
7226 * entity_tick() resulting in prolonged exit starvation.
7228 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7232 * Likewise, bound things on the otherside by preventing insane quota
7233 * periods. This also allows us to normalize in computing quota
7236 if (period > max_cfs_quota_period)
7239 mutex_lock(&cfs_constraints_mutex);
7240 ret = __cfs_schedulable(tg, period, quota);
7244 runtime_enabled = quota != RUNTIME_INF;
7245 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7246 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7247 raw_spin_lock_irq(&cfs_b->lock);
7248 cfs_b->period = ns_to_ktime(period);
7249 cfs_b->quota = quota;
7251 __refill_cfs_bandwidth_runtime(cfs_b);
7252 /* restart the period timer (if active) to handle new period expiry */
7253 if (runtime_enabled && cfs_b->timer_active) {
7254 /* force a reprogram */
7255 cfs_b->timer_active = 0;
7256 __start_cfs_bandwidth(cfs_b);
7258 raw_spin_unlock_irq(&cfs_b->lock);
7260 for_each_possible_cpu(i) {
7261 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7262 struct rq *rq = cfs_rq->rq;
7264 raw_spin_lock_irq(&rq->lock);
7265 cfs_rq->runtime_enabled = runtime_enabled;
7266 cfs_rq->runtime_remaining = 0;
7268 if (cfs_rq->throttled)
7269 unthrottle_cfs_rq(cfs_rq);
7270 raw_spin_unlock_irq(&rq->lock);
7273 mutex_unlock(&cfs_constraints_mutex);
7278 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7282 period = ktime_to_ns(tg->cfs_bandwidth.period);
7283 if (cfs_quota_us < 0)
7284 quota = RUNTIME_INF;
7286 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7288 return tg_set_cfs_bandwidth(tg, period, quota);
7291 long tg_get_cfs_quota(struct task_group *tg)
7295 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7298 quota_us = tg->cfs_bandwidth.quota;
7299 do_div(quota_us, NSEC_PER_USEC);
7304 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7308 period = (u64)cfs_period_us * NSEC_PER_USEC;
7309 quota = tg->cfs_bandwidth.quota;
7311 return tg_set_cfs_bandwidth(tg, period, quota);
7314 long tg_get_cfs_period(struct task_group *tg)
7318 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7319 do_div(cfs_period_us, NSEC_PER_USEC);
7321 return cfs_period_us;
7324 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7326 return tg_get_cfs_quota(cgroup_tg(cgrp));
7329 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7332 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7335 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7337 return tg_get_cfs_period(cgroup_tg(cgrp));
7340 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7343 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7346 struct cfs_schedulable_data {
7347 struct task_group *tg;
7352 * normalize group quota/period to be quota/max_period
7353 * note: units are usecs
7355 static u64 normalize_cfs_quota(struct task_group *tg,
7356 struct cfs_schedulable_data *d)
7364 period = tg_get_cfs_period(tg);
7365 quota = tg_get_cfs_quota(tg);
7368 /* note: these should typically be equivalent */
7369 if (quota == RUNTIME_INF || quota == -1)
7372 return to_ratio(period, quota);
7375 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7377 struct cfs_schedulable_data *d = data;
7378 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7379 s64 quota = 0, parent_quota = -1;
7382 quota = RUNTIME_INF;
7384 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7386 quota = normalize_cfs_quota(tg, d);
7387 parent_quota = parent_b->hierarchal_quota;
7390 * ensure max(child_quota) <= parent_quota, inherit when no
7393 if (quota == RUNTIME_INF)
7394 quota = parent_quota;
7395 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7398 cfs_b->hierarchal_quota = quota;
7403 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7406 struct cfs_schedulable_data data = {
7412 if (quota != RUNTIME_INF) {
7413 do_div(data.period, NSEC_PER_USEC);
7414 do_div(data.quota, NSEC_PER_USEC);
7418 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7424 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7425 struct cgroup_map_cb *cb)
7427 struct task_group *tg = cgroup_tg(cgrp);
7428 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7430 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7431 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7432 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7436 #endif /* CONFIG_CFS_BANDWIDTH */
7437 #endif /* CONFIG_FAIR_GROUP_SCHED */
7439 #ifdef CONFIG_RT_GROUP_SCHED
7440 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7443 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7446 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7448 return sched_group_rt_runtime(cgroup_tg(cgrp));
7451 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7454 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7457 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7459 return sched_group_rt_period(cgroup_tg(cgrp));
7461 #endif /* CONFIG_RT_GROUP_SCHED */
7463 static struct cftype cpu_files[] = {
7464 #ifdef CONFIG_FAIR_GROUP_SCHED
7467 .read_u64 = cpu_shares_read_u64,
7468 .write_u64 = cpu_shares_write_u64,
7471 #ifdef CONFIG_CFS_BANDWIDTH
7473 .name = "cfs_quota_us",
7474 .read_s64 = cpu_cfs_quota_read_s64,
7475 .write_s64 = cpu_cfs_quota_write_s64,
7478 .name = "cfs_period_us",
7479 .read_u64 = cpu_cfs_period_read_u64,
7480 .write_u64 = cpu_cfs_period_write_u64,
7484 .read_map = cpu_stats_show,
7487 #ifdef CONFIG_RT_GROUP_SCHED
7489 .name = "rt_runtime_us",
7490 .read_s64 = cpu_rt_runtime_read,
7491 .write_s64 = cpu_rt_runtime_write,
7494 .name = "rt_period_us",
7495 .read_u64 = cpu_rt_period_read_uint,
7496 .write_u64 = cpu_rt_period_write_uint,
7502 struct cgroup_subsys cpu_cgroup_subsys = {
7504 .css_alloc = cpu_cgroup_css_alloc,
7505 .css_free = cpu_cgroup_css_free,
7506 .css_online = cpu_cgroup_css_online,
7507 .css_offline = cpu_cgroup_css_offline,
7508 .can_attach = cpu_cgroup_can_attach,
7509 .attach = cpu_cgroup_attach,
7510 .exit = cpu_cgroup_exit,
7511 .subsys_id = cpu_cgroup_subsys_id,
7512 .base_cftypes = cpu_files,
7516 #endif /* CONFIG_CGROUP_SCHED */
7518 void dump_cpu_task(int cpu)
7520 pr_info("Task dump for CPU %d:\n", cpu);
7521 sched_show_task(cpu_curr(cpu));