4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 assert_raw_spin_locked(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p))
533 smp_send_reschedule(cpu);
536 void resched_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
541 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 resched_task(cpu_curr(cpu));
544 raw_spin_unlock_irqrestore(&rq->lock, flags);
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu = smp_processor_id();
560 struct sched_domain *sd;
563 for_each_domain(cpu, sd) {
564 for_each_cpu(i, sched_domain_span(sd)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 void wake_up_idle_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
589 if (cpu == smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq->curr != rq->idle)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq->idle);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq->idle))
612 smp_send_reschedule(cpu);
615 static inline bool got_nohz_idle_kick(void)
617 int cpu = smp_processor_id();
618 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
621 #else /* CONFIG_NO_HZ */
623 static inline bool got_nohz_idle_kick(void)
628 #endif /* CONFIG_NO_HZ */
630 void sched_avg_update(struct rq *rq)
632 s64 period = sched_avg_period();
634 while ((s64)(rq->clock - rq->age_stamp) > period) {
636 * Inline assembly required to prevent the compiler
637 * optimising this loop into a divmod call.
638 * See __iter_div_u64_rem() for another example of this.
640 asm("" : "+rm" (rq->age_stamp));
641 rq->age_stamp += period;
646 #else /* !CONFIG_SMP */
647 void resched_task(struct task_struct *p)
649 assert_raw_spin_locked(&task_rq(p)->lock);
650 set_tsk_need_resched(p);
652 #endif /* CONFIG_SMP */
654 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
655 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
657 * Iterate task_group tree rooted at *from, calling @down when first entering a
658 * node and @up when leaving it for the final time.
660 * Caller must hold rcu_lock or sufficient equivalent.
662 int walk_tg_tree_from(struct task_group *from,
663 tg_visitor down, tg_visitor up, void *data)
665 struct task_group *parent, *child;
671 ret = (*down)(parent, data);
674 list_for_each_entry_rcu(child, &parent->children, siblings) {
681 ret = (*up)(parent, data);
682 if (ret || parent == from)
686 parent = parent->parent;
693 int tg_nop(struct task_group *tg, void *data)
699 static void set_load_weight(struct task_struct *p)
701 int prio = p->static_prio - MAX_RT_PRIO;
702 struct load_weight *load = &p->se.load;
705 * SCHED_IDLE tasks get minimal weight:
707 if (p->policy == SCHED_IDLE) {
708 load->weight = scale_load(WEIGHT_IDLEPRIO);
709 load->inv_weight = WMULT_IDLEPRIO;
713 load->weight = scale_load(prio_to_weight[prio]);
714 load->inv_weight = prio_to_wmult[prio];
717 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
720 sched_info_queued(p);
721 p->sched_class->enqueue_task(rq, p, flags);
724 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
727 sched_info_dequeued(p);
728 p->sched_class->dequeue_task(rq, p, flags);
731 void activate_task(struct rq *rq, struct task_struct *p, int flags)
733 if (task_contributes_to_load(p))
734 rq->nr_uninterruptible--;
736 enqueue_task(rq, p, flags);
739 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
741 if (task_contributes_to_load(p))
742 rq->nr_uninterruptible++;
744 dequeue_task(rq, p, flags);
747 static void update_rq_clock_task(struct rq *rq, s64 delta)
750 * In theory, the compile should just see 0 here, and optimize out the call
751 * to sched_rt_avg_update. But I don't trust it...
753 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
754 s64 steal = 0, irq_delta = 0;
756 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
757 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
760 * Since irq_time is only updated on {soft,}irq_exit, we might run into
761 * this case when a previous update_rq_clock() happened inside a
764 * When this happens, we stop ->clock_task and only update the
765 * prev_irq_time stamp to account for the part that fit, so that a next
766 * update will consume the rest. This ensures ->clock_task is
769 * It does however cause some slight miss-attribution of {soft,}irq
770 * time, a more accurate solution would be to update the irq_time using
771 * the current rq->clock timestamp, except that would require using
774 if (irq_delta > delta)
777 rq->prev_irq_time += irq_delta;
780 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
781 if (static_key_false((¶virt_steal_rq_enabled))) {
784 steal = paravirt_steal_clock(cpu_of(rq));
785 steal -= rq->prev_steal_time_rq;
787 if (unlikely(steal > delta))
790 st = steal_ticks(steal);
791 steal = st * TICK_NSEC;
793 rq->prev_steal_time_rq += steal;
799 rq->clock_task += delta;
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
803 sched_rt_avg_update(rq, irq_delta + steal);
807 void sched_set_stop_task(int cpu, struct task_struct *stop)
809 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
810 struct task_struct *old_stop = cpu_rq(cpu)->stop;
814 * Make it appear like a SCHED_FIFO task, its something
815 * userspace knows about and won't get confused about.
817 * Also, it will make PI more or less work without too
818 * much confusion -- but then, stop work should not
819 * rely on PI working anyway.
821 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
823 stop->sched_class = &stop_sched_class;
826 cpu_rq(cpu)->stop = stop;
830 * Reset it back to a normal scheduling class so that
831 * it can die in pieces.
833 old_stop->sched_class = &rt_sched_class;
838 * __normal_prio - return the priority that is based on the static prio
840 static inline int __normal_prio(struct task_struct *p)
842 return p->static_prio;
846 * Calculate the expected normal priority: i.e. priority
847 * without taking RT-inheritance into account. Might be
848 * boosted by interactivity modifiers. Changes upon fork,
849 * setprio syscalls, and whenever the interactivity
850 * estimator recalculates.
852 static inline int normal_prio(struct task_struct *p)
856 if (task_has_rt_policy(p))
857 prio = MAX_RT_PRIO-1 - p->rt_priority;
859 prio = __normal_prio(p);
864 * Calculate the current priority, i.e. the priority
865 * taken into account by the scheduler. This value might
866 * be boosted by RT tasks, or might be boosted by
867 * interactivity modifiers. Will be RT if the task got
868 * RT-boosted. If not then it returns p->normal_prio.
870 static int effective_prio(struct task_struct *p)
872 p->normal_prio = normal_prio(p);
874 * If we are RT tasks or we were boosted to RT priority,
875 * keep the priority unchanged. Otherwise, update priority
876 * to the normal priority:
878 if (!rt_prio(p->prio))
879 return p->normal_prio;
884 * task_curr - is this task currently executing on a CPU?
885 * @p: the task in question.
887 inline int task_curr(const struct task_struct *p)
889 return cpu_curr(task_cpu(p)) == p;
892 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
893 const struct sched_class *prev_class,
896 if (prev_class != p->sched_class) {
897 if (prev_class->switched_from)
898 prev_class->switched_from(rq, p);
899 p->sched_class->switched_to(rq, p);
900 } else if (oldprio != p->prio)
901 p->sched_class->prio_changed(rq, p, oldprio);
904 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
906 const struct sched_class *class;
908 if (p->sched_class == rq->curr->sched_class) {
909 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
911 for_each_class(class) {
912 if (class == rq->curr->sched_class)
914 if (class == p->sched_class) {
915 resched_task(rq->curr);
922 * A queue event has occurred, and we're going to schedule. In
923 * this case, we can save a useless back to back clock update.
925 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
926 rq->skip_clock_update = 1;
929 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
931 void register_task_migration_notifier(struct notifier_block *n)
933 atomic_notifier_chain_register(&task_migration_notifier, n);
937 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
939 #ifdef CONFIG_SCHED_DEBUG
941 * We should never call set_task_cpu() on a blocked task,
942 * ttwu() will sort out the placement.
944 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
945 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
947 #ifdef CONFIG_LOCKDEP
949 * The caller should hold either p->pi_lock or rq->lock, when changing
950 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
952 * sched_move_task() holds both and thus holding either pins the cgroup,
955 * Furthermore, all task_rq users should acquire both locks, see
958 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
959 lockdep_is_held(&task_rq(p)->lock)));
963 trace_sched_migrate_task(p, new_cpu);
965 if (task_cpu(p) != new_cpu) {
966 struct task_migration_notifier tmn;
968 if (p->sched_class->migrate_task_rq)
969 p->sched_class->migrate_task_rq(p, new_cpu);
970 p->se.nr_migrations++;
971 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
974 tmn.from_cpu = task_cpu(p);
975 tmn.to_cpu = new_cpu;
977 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
980 __set_task_cpu(p, new_cpu);
983 struct migration_arg {
984 struct task_struct *task;
988 static int migration_cpu_stop(void *data);
991 * wait_task_inactive - wait for a thread to unschedule.
993 * If @match_state is nonzero, it's the @p->state value just checked and
994 * not expected to change. If it changes, i.e. @p might have woken up,
995 * then return zero. When we succeed in waiting for @p to be off its CPU,
996 * we return a positive number (its total switch count). If a second call
997 * a short while later returns the same number, the caller can be sure that
998 * @p has remained unscheduled the whole time.
1000 * The caller must ensure that the task *will* unschedule sometime soon,
1001 * else this function might spin for a *long* time. This function can't
1002 * be called with interrupts off, or it may introduce deadlock with
1003 * smp_call_function() if an IPI is sent by the same process we are
1004 * waiting to become inactive.
1006 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1008 unsigned long flags;
1015 * We do the initial early heuristics without holding
1016 * any task-queue locks at all. We'll only try to get
1017 * the runqueue lock when things look like they will
1023 * If the task is actively running on another CPU
1024 * still, just relax and busy-wait without holding
1027 * NOTE! Since we don't hold any locks, it's not
1028 * even sure that "rq" stays as the right runqueue!
1029 * But we don't care, since "task_running()" will
1030 * return false if the runqueue has changed and p
1031 * is actually now running somewhere else!
1033 while (task_running(rq, p)) {
1034 if (match_state && unlikely(p->state != match_state))
1040 * Ok, time to look more closely! We need the rq
1041 * lock now, to be *sure*. If we're wrong, we'll
1042 * just go back and repeat.
1044 rq = task_rq_lock(p, &flags);
1045 trace_sched_wait_task(p);
1046 running = task_running(rq, p);
1049 if (!match_state || p->state == match_state)
1050 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1051 task_rq_unlock(rq, p, &flags);
1054 * If it changed from the expected state, bail out now.
1056 if (unlikely(!ncsw))
1060 * Was it really running after all now that we
1061 * checked with the proper locks actually held?
1063 * Oops. Go back and try again..
1065 if (unlikely(running)) {
1071 * It's not enough that it's not actively running,
1072 * it must be off the runqueue _entirely_, and not
1075 * So if it was still runnable (but just not actively
1076 * running right now), it's preempted, and we should
1077 * yield - it could be a while.
1079 if (unlikely(on_rq)) {
1080 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1082 set_current_state(TASK_UNINTERRUPTIBLE);
1083 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1088 * Ahh, all good. It wasn't running, and it wasn't
1089 * runnable, which means that it will never become
1090 * running in the future either. We're all done!
1099 * kick_process - kick a running thread to enter/exit the kernel
1100 * @p: the to-be-kicked thread
1102 * Cause a process which is running on another CPU to enter
1103 * kernel-mode, without any delay. (to get signals handled.)
1105 * NOTE: this function doesn't have to take the runqueue lock,
1106 * because all it wants to ensure is that the remote task enters
1107 * the kernel. If the IPI races and the task has been migrated
1108 * to another CPU then no harm is done and the purpose has been
1111 void kick_process(struct task_struct *p)
1117 if ((cpu != smp_processor_id()) && task_curr(p))
1118 smp_send_reschedule(cpu);
1121 EXPORT_SYMBOL_GPL(kick_process);
1122 #endif /* CONFIG_SMP */
1126 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1128 static int select_fallback_rq(int cpu, struct task_struct *p)
1130 int nid = cpu_to_node(cpu);
1131 const struct cpumask *nodemask = NULL;
1132 enum { cpuset, possible, fail } state = cpuset;
1136 * If the node that the cpu is on has been offlined, cpu_to_node()
1137 * will return -1. There is no cpu on the node, and we should
1138 * select the cpu on the other node.
1141 nodemask = cpumask_of_node(nid);
1143 /* Look for allowed, online CPU in same node. */
1144 for_each_cpu(dest_cpu, nodemask) {
1145 if (!cpu_online(dest_cpu))
1147 if (!cpu_active(dest_cpu))
1149 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1155 /* Any allowed, online CPU? */
1156 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1157 if (!cpu_online(dest_cpu))
1159 if (!cpu_active(dest_cpu))
1166 /* No more Mr. Nice Guy. */
1167 cpuset_cpus_allowed_fallback(p);
1172 do_set_cpus_allowed(p, cpu_possible_mask);
1183 if (state != cpuset) {
1185 * Don't tell them about moving exiting tasks or
1186 * kernel threads (both mm NULL), since they never
1189 if (p->mm && printk_ratelimit()) {
1190 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1191 task_pid_nr(p), p->comm, cpu);
1199 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1202 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1204 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1207 * In order not to call set_task_cpu() on a blocking task we need
1208 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1211 * Since this is common to all placement strategies, this lives here.
1213 * [ this allows ->select_task() to simply return task_cpu(p) and
1214 * not worry about this generic constraint ]
1216 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1218 cpu = select_fallback_rq(task_cpu(p), p);
1223 static void update_avg(u64 *avg, u64 sample)
1225 s64 diff = sample - *avg;
1231 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1233 #ifdef CONFIG_SCHEDSTATS
1234 struct rq *rq = this_rq();
1237 int this_cpu = smp_processor_id();
1239 if (cpu == this_cpu) {
1240 schedstat_inc(rq, ttwu_local);
1241 schedstat_inc(p, se.statistics.nr_wakeups_local);
1243 struct sched_domain *sd;
1245 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1247 for_each_domain(this_cpu, sd) {
1248 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1249 schedstat_inc(sd, ttwu_wake_remote);
1256 if (wake_flags & WF_MIGRATED)
1257 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1259 #endif /* CONFIG_SMP */
1261 schedstat_inc(rq, ttwu_count);
1262 schedstat_inc(p, se.statistics.nr_wakeups);
1264 if (wake_flags & WF_SYNC)
1265 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1267 #endif /* CONFIG_SCHEDSTATS */
1270 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1272 activate_task(rq, p, en_flags);
1275 /* if a worker is waking up, notify workqueue */
1276 if (p->flags & PF_WQ_WORKER)
1277 wq_worker_waking_up(p, cpu_of(rq));
1281 * Mark the task runnable and perform wakeup-preemption.
1284 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1286 trace_sched_wakeup(p, true);
1287 check_preempt_curr(rq, p, wake_flags);
1289 p->state = TASK_RUNNING;
1291 if (p->sched_class->task_woken)
1292 p->sched_class->task_woken(rq, p);
1294 if (rq->idle_stamp) {
1295 u64 delta = rq->clock - rq->idle_stamp;
1296 u64 max = 2*sysctl_sched_migration_cost;
1301 update_avg(&rq->avg_idle, delta);
1308 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1311 if (p->sched_contributes_to_load)
1312 rq->nr_uninterruptible--;
1315 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1316 ttwu_do_wakeup(rq, p, wake_flags);
1320 * Called in case the task @p isn't fully descheduled from its runqueue,
1321 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1322 * since all we need to do is flip p->state to TASK_RUNNING, since
1323 * the task is still ->on_rq.
1325 static int ttwu_remote(struct task_struct *p, int wake_flags)
1330 rq = __task_rq_lock(p);
1332 ttwu_do_wakeup(rq, p, wake_flags);
1335 __task_rq_unlock(rq);
1341 static void sched_ttwu_pending(void)
1343 struct rq *rq = this_rq();
1344 struct llist_node *llist = llist_del_all(&rq->wake_list);
1345 struct task_struct *p;
1347 raw_spin_lock(&rq->lock);
1350 p = llist_entry(llist, struct task_struct, wake_entry);
1351 llist = llist_next(llist);
1352 ttwu_do_activate(rq, p, 0);
1355 raw_spin_unlock(&rq->lock);
1358 void scheduler_ipi(void)
1360 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1364 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1365 * traditionally all their work was done from the interrupt return
1366 * path. Now that we actually do some work, we need to make sure
1369 * Some archs already do call them, luckily irq_enter/exit nest
1372 * Arguably we should visit all archs and update all handlers,
1373 * however a fair share of IPIs are still resched only so this would
1374 * somewhat pessimize the simple resched case.
1377 sched_ttwu_pending();
1380 * Check if someone kicked us for doing the nohz idle load balance.
1382 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1383 this_rq()->idle_balance = 1;
1384 raise_softirq_irqoff(SCHED_SOFTIRQ);
1389 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1391 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1392 smp_send_reschedule(cpu);
1395 bool cpus_share_cache(int this_cpu, int that_cpu)
1397 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1399 #endif /* CONFIG_SMP */
1401 static void ttwu_queue(struct task_struct *p, int cpu)
1403 struct rq *rq = cpu_rq(cpu);
1405 #if defined(CONFIG_SMP)
1406 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1407 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1408 ttwu_queue_remote(p, cpu);
1413 raw_spin_lock(&rq->lock);
1414 ttwu_do_activate(rq, p, 0);
1415 raw_spin_unlock(&rq->lock);
1419 * try_to_wake_up - wake up a thread
1420 * @p: the thread to be awakened
1421 * @state: the mask of task states that can be woken
1422 * @wake_flags: wake modifier flags (WF_*)
1424 * Put it on the run-queue if it's not already there. The "current"
1425 * thread is always on the run-queue (except when the actual
1426 * re-schedule is in progress), and as such you're allowed to do
1427 * the simpler "current->state = TASK_RUNNING" to mark yourself
1428 * runnable without the overhead of this.
1430 * Returns %true if @p was woken up, %false if it was already running
1431 * or @state didn't match @p's state.
1434 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1436 unsigned long flags;
1437 int cpu, success = 0;
1440 raw_spin_lock_irqsave(&p->pi_lock, flags);
1441 if (!(p->state & state))
1444 success = 1; /* we're going to change ->state */
1447 if (p->on_rq && ttwu_remote(p, wake_flags))
1452 * If the owning (remote) cpu is still in the middle of schedule() with
1453 * this task as prev, wait until its done referencing the task.
1458 * Pairs with the smp_wmb() in finish_lock_switch().
1462 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1463 p->state = TASK_WAKING;
1465 if (p->sched_class->task_waking)
1466 p->sched_class->task_waking(p);
1468 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1469 if (task_cpu(p) != cpu) {
1470 wake_flags |= WF_MIGRATED;
1471 set_task_cpu(p, cpu);
1473 #endif /* CONFIG_SMP */
1477 ttwu_stat(p, cpu, wake_flags);
1479 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1485 * try_to_wake_up_local - try to wake up a local task with rq lock held
1486 * @p: the thread to be awakened
1488 * Put @p on the run-queue if it's not already there. The caller must
1489 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1492 static void try_to_wake_up_local(struct task_struct *p)
1494 struct rq *rq = task_rq(p);
1496 BUG_ON(rq != this_rq());
1497 BUG_ON(p == current);
1498 lockdep_assert_held(&rq->lock);
1500 if (!raw_spin_trylock(&p->pi_lock)) {
1501 raw_spin_unlock(&rq->lock);
1502 raw_spin_lock(&p->pi_lock);
1503 raw_spin_lock(&rq->lock);
1506 if (!(p->state & TASK_NORMAL))
1510 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1512 ttwu_do_wakeup(rq, p, 0);
1513 ttwu_stat(p, smp_processor_id(), 0);
1515 raw_spin_unlock(&p->pi_lock);
1519 * wake_up_process - Wake up a specific process
1520 * @p: The process to be woken up.
1522 * Attempt to wake up the nominated process and move it to the set of runnable
1523 * processes. Returns 1 if the process was woken up, 0 if it was already
1526 * It may be assumed that this function implies a write memory barrier before
1527 * changing the task state if and only if any tasks are woken up.
1529 int wake_up_process(struct task_struct *p)
1531 WARN_ON(task_is_stopped_or_traced(p));
1532 return try_to_wake_up(p, TASK_NORMAL, 0);
1534 EXPORT_SYMBOL(wake_up_process);
1536 int wake_up_state(struct task_struct *p, unsigned int state)
1538 return try_to_wake_up(p, state, 0);
1542 * Perform scheduler related setup for a newly forked process p.
1543 * p is forked by current.
1545 * __sched_fork() is basic setup used by init_idle() too:
1547 static void __sched_fork(struct task_struct *p)
1552 p->se.exec_start = 0;
1553 p->se.sum_exec_runtime = 0;
1554 p->se.prev_sum_exec_runtime = 0;
1555 p->se.nr_migrations = 0;
1557 INIT_LIST_HEAD(&p->se.group_node);
1560 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1561 * removed when useful for applications beyond shares distribution (e.g.
1564 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1565 p->se.avg.runnable_avg_period = 0;
1566 p->se.avg.runnable_avg_sum = 0;
1568 #ifdef CONFIG_SCHEDSTATS
1569 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1572 INIT_LIST_HEAD(&p->rt.run_list);
1574 #ifdef CONFIG_PREEMPT_NOTIFIERS
1575 INIT_HLIST_HEAD(&p->preempt_notifiers);
1578 #ifdef CONFIG_NUMA_BALANCING
1579 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1580 p->mm->numa_next_scan = jiffies;
1581 p->mm->numa_next_reset = jiffies;
1582 p->mm->numa_scan_seq = 0;
1585 p->node_stamp = 0ULL;
1586 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1587 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1588 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1589 p->numa_work.next = &p->numa_work;
1590 #endif /* CONFIG_NUMA_BALANCING */
1593 #ifdef CONFIG_NUMA_BALANCING
1594 #ifdef CONFIG_SCHED_DEBUG
1595 void set_numabalancing_state(bool enabled)
1598 sched_feat_set("NUMA");
1600 sched_feat_set("NO_NUMA");
1603 __read_mostly bool numabalancing_enabled;
1605 void set_numabalancing_state(bool enabled)
1607 numabalancing_enabled = enabled;
1609 #endif /* CONFIG_SCHED_DEBUG */
1610 #endif /* CONFIG_NUMA_BALANCING */
1613 * fork()/clone()-time setup:
1615 void sched_fork(struct task_struct *p)
1617 unsigned long flags;
1618 int cpu = get_cpu();
1622 * We mark the process as running here. This guarantees that
1623 * nobody will actually run it, and a signal or other external
1624 * event cannot wake it up and insert it on the runqueue either.
1626 p->state = TASK_RUNNING;
1629 * Make sure we do not leak PI boosting priority to the child.
1631 p->prio = current->normal_prio;
1634 * Revert to default priority/policy on fork if requested.
1636 if (unlikely(p->sched_reset_on_fork)) {
1637 if (task_has_rt_policy(p)) {
1638 p->policy = SCHED_NORMAL;
1639 p->static_prio = NICE_TO_PRIO(0);
1641 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1642 p->static_prio = NICE_TO_PRIO(0);
1644 p->prio = p->normal_prio = __normal_prio(p);
1648 * We don't need the reset flag anymore after the fork. It has
1649 * fulfilled its duty:
1651 p->sched_reset_on_fork = 0;
1654 if (!rt_prio(p->prio))
1655 p->sched_class = &fair_sched_class;
1657 if (p->sched_class->task_fork)
1658 p->sched_class->task_fork(p);
1661 * The child is not yet in the pid-hash so no cgroup attach races,
1662 * and the cgroup is pinned to this child due to cgroup_fork()
1663 * is ran before sched_fork().
1665 * Silence PROVE_RCU.
1667 raw_spin_lock_irqsave(&p->pi_lock, flags);
1668 set_task_cpu(p, cpu);
1669 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1671 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1672 if (likely(sched_info_on()))
1673 memset(&p->sched_info, 0, sizeof(p->sched_info));
1675 #if defined(CONFIG_SMP)
1678 #ifdef CONFIG_PREEMPT_COUNT
1679 /* Want to start with kernel preemption disabled. */
1680 task_thread_info(p)->preempt_count = 1;
1683 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1690 * wake_up_new_task - wake up a newly created task for the first time.
1692 * This function will do some initial scheduler statistics housekeeping
1693 * that must be done for every newly created context, then puts the task
1694 * on the runqueue and wakes it.
1696 void wake_up_new_task(struct task_struct *p)
1698 unsigned long flags;
1701 raw_spin_lock_irqsave(&p->pi_lock, flags);
1704 * Fork balancing, do it here and not earlier because:
1705 * - cpus_allowed can change in the fork path
1706 * - any previously selected cpu might disappear through hotplug
1708 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1711 rq = __task_rq_lock(p);
1712 activate_task(rq, p, 0);
1714 trace_sched_wakeup_new(p, true);
1715 check_preempt_curr(rq, p, WF_FORK);
1717 if (p->sched_class->task_woken)
1718 p->sched_class->task_woken(rq, p);
1720 task_rq_unlock(rq, p, &flags);
1723 #ifdef CONFIG_PREEMPT_NOTIFIERS
1726 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1727 * @notifier: notifier struct to register
1729 void preempt_notifier_register(struct preempt_notifier *notifier)
1731 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1733 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1736 * preempt_notifier_unregister - no longer interested in preemption notifications
1737 * @notifier: notifier struct to unregister
1739 * This is safe to call from within a preemption notifier.
1741 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1743 hlist_del(¬ifier->link);
1745 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1747 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1749 struct preempt_notifier *notifier;
1751 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1752 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1756 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1757 struct task_struct *next)
1759 struct preempt_notifier *notifier;
1761 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1762 notifier->ops->sched_out(notifier, next);
1765 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1767 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1772 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1773 struct task_struct *next)
1777 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1780 * prepare_task_switch - prepare to switch tasks
1781 * @rq: the runqueue preparing to switch
1782 * @prev: the current task that is being switched out
1783 * @next: the task we are going to switch to.
1785 * This is called with the rq lock held and interrupts off. It must
1786 * be paired with a subsequent finish_task_switch after the context
1789 * prepare_task_switch sets up locking and calls architecture specific
1793 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1794 struct task_struct *next)
1796 trace_sched_switch(prev, next);
1797 sched_info_switch(prev, next);
1798 perf_event_task_sched_out(prev, next);
1799 fire_sched_out_preempt_notifiers(prev, next);
1800 prepare_lock_switch(rq, next);
1801 prepare_arch_switch(next);
1805 * finish_task_switch - clean up after a task-switch
1806 * @rq: runqueue associated with task-switch
1807 * @prev: the thread we just switched away from.
1809 * finish_task_switch must be called after the context switch, paired
1810 * with a prepare_task_switch call before the context switch.
1811 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1812 * and do any other architecture-specific cleanup actions.
1814 * Note that we may have delayed dropping an mm in context_switch(). If
1815 * so, we finish that here outside of the runqueue lock. (Doing it
1816 * with the lock held can cause deadlocks; see schedule() for
1819 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1820 __releases(rq->lock)
1822 struct mm_struct *mm = rq->prev_mm;
1828 * A task struct has one reference for the use as "current".
1829 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1830 * schedule one last time. The schedule call will never return, and
1831 * the scheduled task must drop that reference.
1832 * The test for TASK_DEAD must occur while the runqueue locks are
1833 * still held, otherwise prev could be scheduled on another cpu, die
1834 * there before we look at prev->state, and then the reference would
1836 * Manfred Spraul <manfred@colorfullife.com>
1838 prev_state = prev->state;
1839 vtime_task_switch(prev);
1840 finish_arch_switch(prev);
1841 perf_event_task_sched_in(prev, current);
1842 finish_lock_switch(rq, prev);
1843 finish_arch_post_lock_switch();
1845 fire_sched_in_preempt_notifiers(current);
1848 if (unlikely(prev_state == TASK_DEAD)) {
1850 * Remove function-return probe instances associated with this
1851 * task and put them back on the free list.
1853 kprobe_flush_task(prev);
1854 put_task_struct(prev);
1860 /* assumes rq->lock is held */
1861 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1863 if (prev->sched_class->pre_schedule)
1864 prev->sched_class->pre_schedule(rq, prev);
1867 /* rq->lock is NOT held, but preemption is disabled */
1868 static inline void post_schedule(struct rq *rq)
1870 if (rq->post_schedule) {
1871 unsigned long flags;
1873 raw_spin_lock_irqsave(&rq->lock, flags);
1874 if (rq->curr->sched_class->post_schedule)
1875 rq->curr->sched_class->post_schedule(rq);
1876 raw_spin_unlock_irqrestore(&rq->lock, flags);
1878 rq->post_schedule = 0;
1884 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1888 static inline void post_schedule(struct rq *rq)
1895 * schedule_tail - first thing a freshly forked thread must call.
1896 * @prev: the thread we just switched away from.
1898 asmlinkage void schedule_tail(struct task_struct *prev)
1899 __releases(rq->lock)
1901 struct rq *rq = this_rq();
1903 finish_task_switch(rq, prev);
1906 * FIXME: do we need to worry about rq being invalidated by the
1911 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1912 /* In this case, finish_task_switch does not reenable preemption */
1915 if (current->set_child_tid)
1916 put_user(task_pid_vnr(current), current->set_child_tid);
1920 * context_switch - switch to the new MM and the new
1921 * thread's register state.
1924 context_switch(struct rq *rq, struct task_struct *prev,
1925 struct task_struct *next)
1927 struct mm_struct *mm, *oldmm;
1929 prepare_task_switch(rq, prev, next);
1932 oldmm = prev->active_mm;
1934 * For paravirt, this is coupled with an exit in switch_to to
1935 * combine the page table reload and the switch backend into
1938 arch_start_context_switch(prev);
1941 next->active_mm = oldmm;
1942 atomic_inc(&oldmm->mm_count);
1943 enter_lazy_tlb(oldmm, next);
1945 switch_mm(oldmm, mm, next);
1948 prev->active_mm = NULL;
1949 rq->prev_mm = oldmm;
1952 * Since the runqueue lock will be released by the next
1953 * task (which is an invalid locking op but in the case
1954 * of the scheduler it's an obvious special-case), so we
1955 * do an early lockdep release here:
1957 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1958 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1961 context_tracking_task_switch(prev, next);
1962 /* Here we just switch the register state and the stack. */
1963 switch_to(prev, next, prev);
1967 * this_rq must be evaluated again because prev may have moved
1968 * CPUs since it called schedule(), thus the 'rq' on its stack
1969 * frame will be invalid.
1971 finish_task_switch(this_rq(), prev);
1975 * nr_running and nr_context_switches:
1977 * externally visible scheduler statistics: current number of runnable
1978 * threads, total number of context switches performed since bootup.
1980 unsigned long nr_running(void)
1982 unsigned long i, sum = 0;
1984 for_each_online_cpu(i)
1985 sum += cpu_rq(i)->nr_running;
1990 unsigned long long nr_context_switches(void)
1993 unsigned long long sum = 0;
1995 for_each_possible_cpu(i)
1996 sum += cpu_rq(i)->nr_switches;
2001 unsigned long nr_iowait(void)
2003 unsigned long i, sum = 0;
2005 for_each_possible_cpu(i)
2006 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2011 unsigned long nr_iowait_cpu(int cpu)
2013 struct rq *this = cpu_rq(cpu);
2014 return atomic_read(&this->nr_iowait);
2017 unsigned long this_cpu_load(void)
2019 struct rq *this = this_rq();
2020 return this->cpu_load[0];
2025 * Global load-average calculations
2027 * We take a distributed and async approach to calculating the global load-avg
2028 * in order to minimize overhead.
2030 * The global load average is an exponentially decaying average of nr_running +
2031 * nr_uninterruptible.
2033 * Once every LOAD_FREQ:
2036 * for_each_possible_cpu(cpu)
2037 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2039 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2041 * Due to a number of reasons the above turns in the mess below:
2043 * - for_each_possible_cpu() is prohibitively expensive on machines with
2044 * serious number of cpus, therefore we need to take a distributed approach
2045 * to calculating nr_active.
2047 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2048 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2050 * So assuming nr_active := 0 when we start out -- true per definition, we
2051 * can simply take per-cpu deltas and fold those into a global accumulate
2052 * to obtain the same result. See calc_load_fold_active().
2054 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2055 * across the machine, we assume 10 ticks is sufficient time for every
2056 * cpu to have completed this task.
2058 * This places an upper-bound on the IRQ-off latency of the machine. Then
2059 * again, being late doesn't loose the delta, just wrecks the sample.
2061 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2062 * this would add another cross-cpu cacheline miss and atomic operation
2063 * to the wakeup path. Instead we increment on whatever cpu the task ran
2064 * when it went into uninterruptible state and decrement on whatever cpu
2065 * did the wakeup. This means that only the sum of nr_uninterruptible over
2066 * all cpus yields the correct result.
2068 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2071 /* Variables and functions for calc_load */
2072 static atomic_long_t calc_load_tasks;
2073 static unsigned long calc_load_update;
2074 unsigned long avenrun[3];
2075 EXPORT_SYMBOL(avenrun); /* should be removed */
2078 * get_avenrun - get the load average array
2079 * @loads: pointer to dest load array
2080 * @offset: offset to add
2081 * @shift: shift count to shift the result left
2083 * These values are estimates at best, so no need for locking.
2085 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2087 loads[0] = (avenrun[0] + offset) << shift;
2088 loads[1] = (avenrun[1] + offset) << shift;
2089 loads[2] = (avenrun[2] + offset) << shift;
2092 static long calc_load_fold_active(struct rq *this_rq)
2094 long nr_active, delta = 0;
2096 nr_active = this_rq->nr_running;
2097 nr_active += (long) this_rq->nr_uninterruptible;
2099 if (nr_active != this_rq->calc_load_active) {
2100 delta = nr_active - this_rq->calc_load_active;
2101 this_rq->calc_load_active = nr_active;
2108 * a1 = a0 * e + a * (1 - e)
2110 static unsigned long
2111 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2114 load += active * (FIXED_1 - exp);
2115 load += 1UL << (FSHIFT - 1);
2116 return load >> FSHIFT;
2121 * Handle NO_HZ for the global load-average.
2123 * Since the above described distributed algorithm to compute the global
2124 * load-average relies on per-cpu sampling from the tick, it is affected by
2127 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2128 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2129 * when we read the global state.
2131 * Obviously reality has to ruin such a delightfully simple scheme:
2133 * - When we go NO_HZ idle during the window, we can negate our sample
2134 * contribution, causing under-accounting.
2136 * We avoid this by keeping two idle-delta counters and flipping them
2137 * when the window starts, thus separating old and new NO_HZ load.
2139 * The only trick is the slight shift in index flip for read vs write.
2143 * |-|-----------|-|-----------|-|-----------|-|
2144 * r:0 0 1 1 0 0 1 1 0
2145 * w:0 1 1 0 0 1 1 0 0
2147 * This ensures we'll fold the old idle contribution in this window while
2148 * accumlating the new one.
2150 * - When we wake up from NO_HZ idle during the window, we push up our
2151 * contribution, since we effectively move our sample point to a known
2154 * This is solved by pushing the window forward, and thus skipping the
2155 * sample, for this cpu (effectively using the idle-delta for this cpu which
2156 * was in effect at the time the window opened). This also solves the issue
2157 * of having to deal with a cpu having been in NOHZ idle for multiple
2158 * LOAD_FREQ intervals.
2160 * When making the ILB scale, we should try to pull this in as well.
2162 static atomic_long_t calc_load_idle[2];
2163 static int calc_load_idx;
2165 static inline int calc_load_write_idx(void)
2167 int idx = calc_load_idx;
2170 * See calc_global_nohz(), if we observe the new index, we also
2171 * need to observe the new update time.
2176 * If the folding window started, make sure we start writing in the
2179 if (!time_before(jiffies, calc_load_update))
2185 static inline int calc_load_read_idx(void)
2187 return calc_load_idx & 1;
2190 void calc_load_enter_idle(void)
2192 struct rq *this_rq = this_rq();
2196 * We're going into NOHZ mode, if there's any pending delta, fold it
2197 * into the pending idle delta.
2199 delta = calc_load_fold_active(this_rq);
2201 int idx = calc_load_write_idx();
2202 atomic_long_add(delta, &calc_load_idle[idx]);
2206 void calc_load_exit_idle(void)
2208 struct rq *this_rq = this_rq();
2211 * If we're still before the sample window, we're done.
2213 if (time_before(jiffies, this_rq->calc_load_update))
2217 * We woke inside or after the sample window, this means we're already
2218 * accounted through the nohz accounting, so skip the entire deal and
2219 * sync up for the next window.
2221 this_rq->calc_load_update = calc_load_update;
2222 if (time_before(jiffies, this_rq->calc_load_update + 10))
2223 this_rq->calc_load_update += LOAD_FREQ;
2226 static long calc_load_fold_idle(void)
2228 int idx = calc_load_read_idx();
2231 if (atomic_long_read(&calc_load_idle[idx]))
2232 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2238 * fixed_power_int - compute: x^n, in O(log n) time
2240 * @x: base of the power
2241 * @frac_bits: fractional bits of @x
2242 * @n: power to raise @x to.
2244 * By exploiting the relation between the definition of the natural power
2245 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2246 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2247 * (where: n_i \elem {0, 1}, the binary vector representing n),
2248 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2249 * of course trivially computable in O(log_2 n), the length of our binary
2252 static unsigned long
2253 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2255 unsigned long result = 1UL << frac_bits;
2260 result += 1UL << (frac_bits - 1);
2261 result >>= frac_bits;
2267 x += 1UL << (frac_bits - 1);
2275 * a1 = a0 * e + a * (1 - e)
2277 * a2 = a1 * e + a * (1 - e)
2278 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2279 * = a0 * e^2 + a * (1 - e) * (1 + e)
2281 * a3 = a2 * e + a * (1 - e)
2282 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2283 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2287 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2288 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2289 * = a0 * e^n + a * (1 - e^n)
2291 * [1] application of the geometric series:
2294 * S_n := \Sum x^i = -------------
2297 static unsigned long
2298 calc_load_n(unsigned long load, unsigned long exp,
2299 unsigned long active, unsigned int n)
2302 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2306 * NO_HZ can leave us missing all per-cpu ticks calling
2307 * calc_load_account_active(), but since an idle CPU folds its delta into
2308 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2309 * in the pending idle delta if our idle period crossed a load cycle boundary.
2311 * Once we've updated the global active value, we need to apply the exponential
2312 * weights adjusted to the number of cycles missed.
2314 static void calc_global_nohz(void)
2316 long delta, active, n;
2318 if (!time_before(jiffies, calc_load_update + 10)) {
2320 * Catch-up, fold however many we are behind still
2322 delta = jiffies - calc_load_update - 10;
2323 n = 1 + (delta / LOAD_FREQ);
2325 active = atomic_long_read(&calc_load_tasks);
2326 active = active > 0 ? active * FIXED_1 : 0;
2328 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2329 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2330 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2332 calc_load_update += n * LOAD_FREQ;
2336 * Flip the idle index...
2338 * Make sure we first write the new time then flip the index, so that
2339 * calc_load_write_idx() will see the new time when it reads the new
2340 * index, this avoids a double flip messing things up.
2345 #else /* !CONFIG_NO_HZ */
2347 static inline long calc_load_fold_idle(void) { return 0; }
2348 static inline void calc_global_nohz(void) { }
2350 #endif /* CONFIG_NO_HZ */
2353 * calc_load - update the avenrun load estimates 10 ticks after the
2354 * CPUs have updated calc_load_tasks.
2356 void calc_global_load(unsigned long ticks)
2360 if (time_before(jiffies, calc_load_update + 10))
2364 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2366 delta = calc_load_fold_idle();
2368 atomic_long_add(delta, &calc_load_tasks);
2370 active = atomic_long_read(&calc_load_tasks);
2371 active = active > 0 ? active * FIXED_1 : 0;
2373 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2374 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2375 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2377 calc_load_update += LOAD_FREQ;
2380 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2386 * Called from update_cpu_load() to periodically update this CPU's
2389 static void calc_load_account_active(struct rq *this_rq)
2393 if (time_before(jiffies, this_rq->calc_load_update))
2396 delta = calc_load_fold_active(this_rq);
2398 atomic_long_add(delta, &calc_load_tasks);
2400 this_rq->calc_load_update += LOAD_FREQ;
2404 * End of global load-average stuff
2408 * The exact cpuload at various idx values, calculated at every tick would be
2409 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2411 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2412 * on nth tick when cpu may be busy, then we have:
2413 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2414 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2416 * decay_load_missed() below does efficient calculation of
2417 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2418 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2420 * The calculation is approximated on a 128 point scale.
2421 * degrade_zero_ticks is the number of ticks after which load at any
2422 * particular idx is approximated to be zero.
2423 * degrade_factor is a precomputed table, a row for each load idx.
2424 * Each column corresponds to degradation factor for a power of two ticks,
2425 * based on 128 point scale.
2427 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2428 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2430 * With this power of 2 load factors, we can degrade the load n times
2431 * by looking at 1 bits in n and doing as many mult/shift instead of
2432 * n mult/shifts needed by the exact degradation.
2434 #define DEGRADE_SHIFT 7
2435 static const unsigned char
2436 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2437 static const unsigned char
2438 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2439 {0, 0, 0, 0, 0, 0, 0, 0},
2440 {64, 32, 8, 0, 0, 0, 0, 0},
2441 {96, 72, 40, 12, 1, 0, 0},
2442 {112, 98, 75, 43, 15, 1, 0},
2443 {120, 112, 98, 76, 45, 16, 2} };
2446 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2447 * would be when CPU is idle and so we just decay the old load without
2448 * adding any new load.
2450 static unsigned long
2451 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2455 if (!missed_updates)
2458 if (missed_updates >= degrade_zero_ticks[idx])
2462 return load >> missed_updates;
2464 while (missed_updates) {
2465 if (missed_updates % 2)
2466 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2468 missed_updates >>= 1;
2475 * Update rq->cpu_load[] statistics. This function is usually called every
2476 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2477 * every tick. We fix it up based on jiffies.
2479 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2480 unsigned long pending_updates)
2484 this_rq->nr_load_updates++;
2486 /* Update our load: */
2487 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2488 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2489 unsigned long old_load, new_load;
2491 /* scale is effectively 1 << i now, and >> i divides by scale */
2493 old_load = this_rq->cpu_load[i];
2494 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2495 new_load = this_load;
2497 * Round up the averaging division if load is increasing. This
2498 * prevents us from getting stuck on 9 if the load is 10, for
2501 if (new_load > old_load)
2502 new_load += scale - 1;
2504 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2507 sched_avg_update(this_rq);
2512 * There is no sane way to deal with nohz on smp when using jiffies because the
2513 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2514 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2516 * Therefore we cannot use the delta approach from the regular tick since that
2517 * would seriously skew the load calculation. However we'll make do for those
2518 * updates happening while idle (nohz_idle_balance) or coming out of idle
2519 * (tick_nohz_idle_exit).
2521 * This means we might still be one tick off for nohz periods.
2525 * Called from nohz_idle_balance() to update the load ratings before doing the
2528 void update_idle_cpu_load(struct rq *this_rq)
2530 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2531 unsigned long load = this_rq->load.weight;
2532 unsigned long pending_updates;
2535 * bail if there's load or we're actually up-to-date.
2537 if (load || curr_jiffies == this_rq->last_load_update_tick)
2540 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2541 this_rq->last_load_update_tick = curr_jiffies;
2543 __update_cpu_load(this_rq, load, pending_updates);
2547 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2549 void update_cpu_load_nohz(void)
2551 struct rq *this_rq = this_rq();
2552 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2553 unsigned long pending_updates;
2555 if (curr_jiffies == this_rq->last_load_update_tick)
2558 raw_spin_lock(&this_rq->lock);
2559 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2560 if (pending_updates) {
2561 this_rq->last_load_update_tick = curr_jiffies;
2563 * We were idle, this means load 0, the current load might be
2564 * !0 due to remote wakeups and the sort.
2566 __update_cpu_load(this_rq, 0, pending_updates);
2568 raw_spin_unlock(&this_rq->lock);
2570 #endif /* CONFIG_NO_HZ */
2573 * Called from scheduler_tick()
2575 static void update_cpu_load_active(struct rq *this_rq)
2578 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2580 this_rq->last_load_update_tick = jiffies;
2581 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2583 calc_load_account_active(this_rq);
2589 * sched_exec - execve() is a valuable balancing opportunity, because at
2590 * this point the task has the smallest effective memory and cache footprint.
2592 void sched_exec(void)
2594 struct task_struct *p = current;
2595 unsigned long flags;
2598 raw_spin_lock_irqsave(&p->pi_lock, flags);
2599 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2600 if (dest_cpu == smp_processor_id())
2603 if (likely(cpu_active(dest_cpu))) {
2604 struct migration_arg arg = { p, dest_cpu };
2606 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2607 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2611 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2616 DEFINE_PER_CPU(struct kernel_stat, kstat);
2617 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2619 EXPORT_PER_CPU_SYMBOL(kstat);
2620 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2623 * Return any ns on the sched_clock that have not yet been accounted in
2624 * @p in case that task is currently running.
2626 * Called with task_rq_lock() held on @rq.
2628 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2632 if (task_current(rq, p)) {
2633 update_rq_clock(rq);
2634 ns = rq->clock_task - p->se.exec_start;
2642 unsigned long long task_delta_exec(struct task_struct *p)
2644 unsigned long flags;
2648 rq = task_rq_lock(p, &flags);
2649 ns = do_task_delta_exec(p, rq);
2650 task_rq_unlock(rq, p, &flags);
2656 * Return accounted runtime for the task.
2657 * In case the task is currently running, return the runtime plus current's
2658 * pending runtime that have not been accounted yet.
2660 unsigned long long task_sched_runtime(struct task_struct *p)
2662 unsigned long flags;
2666 rq = task_rq_lock(p, &flags);
2667 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2668 task_rq_unlock(rq, p, &flags);
2674 * This function gets called by the timer code, with HZ frequency.
2675 * We call it with interrupts disabled.
2677 void scheduler_tick(void)
2679 int cpu = smp_processor_id();
2680 struct rq *rq = cpu_rq(cpu);
2681 struct task_struct *curr = rq->curr;
2685 raw_spin_lock(&rq->lock);
2686 update_rq_clock(rq);
2687 update_cpu_load_active(rq);
2688 curr->sched_class->task_tick(rq, curr, 0);
2689 raw_spin_unlock(&rq->lock);
2691 perf_event_task_tick();
2694 rq->idle_balance = idle_cpu(cpu);
2695 trigger_load_balance(rq, cpu);
2699 notrace unsigned long get_parent_ip(unsigned long addr)
2701 if (in_lock_functions(addr)) {
2702 addr = CALLER_ADDR2;
2703 if (in_lock_functions(addr))
2704 addr = CALLER_ADDR3;
2709 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2710 defined(CONFIG_PREEMPT_TRACER))
2712 void __kprobes add_preempt_count(int val)
2714 #ifdef CONFIG_DEBUG_PREEMPT
2718 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2721 preempt_count() += val;
2722 #ifdef CONFIG_DEBUG_PREEMPT
2724 * Spinlock count overflowing soon?
2726 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2729 if (preempt_count() == val)
2730 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2732 EXPORT_SYMBOL(add_preempt_count);
2734 void __kprobes sub_preempt_count(int val)
2736 #ifdef CONFIG_DEBUG_PREEMPT
2740 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2743 * Is the spinlock portion underflowing?
2745 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2746 !(preempt_count() & PREEMPT_MASK)))
2750 if (preempt_count() == val)
2751 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2752 preempt_count() -= val;
2754 EXPORT_SYMBOL(sub_preempt_count);
2759 * Print scheduling while atomic bug:
2761 static noinline void __schedule_bug(struct task_struct *prev)
2763 if (oops_in_progress)
2766 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2767 prev->comm, prev->pid, preempt_count());
2769 debug_show_held_locks(prev);
2771 if (irqs_disabled())
2772 print_irqtrace_events(prev);
2774 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2778 * Various schedule()-time debugging checks and statistics:
2780 static inline void schedule_debug(struct task_struct *prev)
2783 * Test if we are atomic. Since do_exit() needs to call into
2784 * schedule() atomically, we ignore that path for now.
2785 * Otherwise, whine if we are scheduling when we should not be.
2787 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2788 __schedule_bug(prev);
2791 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2793 schedstat_inc(this_rq(), sched_count);
2796 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2798 if (prev->on_rq || rq->skip_clock_update < 0)
2799 update_rq_clock(rq);
2800 prev->sched_class->put_prev_task(rq, prev);
2804 * Pick up the highest-prio task:
2806 static inline struct task_struct *
2807 pick_next_task(struct rq *rq)
2809 const struct sched_class *class;
2810 struct task_struct *p;
2813 * Optimization: we know that if all tasks are in
2814 * the fair class we can call that function directly:
2816 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2817 p = fair_sched_class.pick_next_task(rq);
2822 for_each_class(class) {
2823 p = class->pick_next_task(rq);
2828 BUG(); /* the idle class will always have a runnable task */
2832 * __schedule() is the main scheduler function.
2834 * The main means of driving the scheduler and thus entering this function are:
2836 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2838 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2839 * paths. For example, see arch/x86/entry_64.S.
2841 * To drive preemption between tasks, the scheduler sets the flag in timer
2842 * interrupt handler scheduler_tick().
2844 * 3. Wakeups don't really cause entry into schedule(). They add a
2845 * task to the run-queue and that's it.
2847 * Now, if the new task added to the run-queue preempts the current
2848 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2849 * called on the nearest possible occasion:
2851 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2853 * - in syscall or exception context, at the next outmost
2854 * preempt_enable(). (this might be as soon as the wake_up()'s
2857 * - in IRQ context, return from interrupt-handler to
2858 * preemptible context
2860 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2863 * - cond_resched() call
2864 * - explicit schedule() call
2865 * - return from syscall or exception to user-space
2866 * - return from interrupt-handler to user-space
2868 static void __sched __schedule(void)
2870 struct task_struct *prev, *next;
2871 unsigned long *switch_count;
2877 cpu = smp_processor_id();
2879 rcu_note_context_switch(cpu);
2882 schedule_debug(prev);
2884 if (sched_feat(HRTICK))
2887 raw_spin_lock_irq(&rq->lock);
2889 switch_count = &prev->nivcsw;
2890 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2891 if (unlikely(signal_pending_state(prev->state, prev))) {
2892 prev->state = TASK_RUNNING;
2894 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2898 * If a worker went to sleep, notify and ask workqueue
2899 * whether it wants to wake up a task to maintain
2902 if (prev->flags & PF_WQ_WORKER) {
2903 struct task_struct *to_wakeup;
2905 to_wakeup = wq_worker_sleeping(prev, cpu);
2907 try_to_wake_up_local(to_wakeup);
2910 switch_count = &prev->nvcsw;
2913 pre_schedule(rq, prev);
2915 if (unlikely(!rq->nr_running))
2916 idle_balance(cpu, rq);
2918 put_prev_task(rq, prev);
2919 next = pick_next_task(rq);
2920 clear_tsk_need_resched(prev);
2921 rq->skip_clock_update = 0;
2923 if (likely(prev != next)) {
2928 context_switch(rq, prev, next); /* unlocks the rq */
2930 * The context switch have flipped the stack from under us
2931 * and restored the local variables which were saved when
2932 * this task called schedule() in the past. prev == current
2933 * is still correct, but it can be moved to another cpu/rq.
2935 cpu = smp_processor_id();
2938 raw_spin_unlock_irq(&rq->lock);
2942 sched_preempt_enable_no_resched();
2947 static inline void sched_submit_work(struct task_struct *tsk)
2949 if (!tsk->state || tsk_is_pi_blocked(tsk))
2952 * If we are going to sleep and we have plugged IO queued,
2953 * make sure to submit it to avoid deadlocks.
2955 if (blk_needs_flush_plug(tsk))
2956 blk_schedule_flush_plug(tsk);
2959 asmlinkage void __sched schedule(void)
2961 struct task_struct *tsk = current;
2963 sched_submit_work(tsk);
2966 EXPORT_SYMBOL(schedule);
2968 #ifdef CONFIG_CONTEXT_TRACKING
2969 asmlinkage void __sched schedule_user(void)
2972 * If we come here after a random call to set_need_resched(),
2973 * or we have been woken up remotely but the IPI has not yet arrived,
2974 * we haven't yet exited the RCU idle mode. Do it here manually until
2975 * we find a better solution.
2984 * schedule_preempt_disabled - called with preemption disabled
2986 * Returns with preemption disabled. Note: preempt_count must be 1
2988 void __sched schedule_preempt_disabled(void)
2990 sched_preempt_enable_no_resched();
2995 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2997 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2999 if (lock->owner != owner)
3003 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3004 * lock->owner still matches owner, if that fails, owner might
3005 * point to free()d memory, if it still matches, the rcu_read_lock()
3006 * ensures the memory stays valid.
3010 return owner->on_cpu;
3014 * Look out! "owner" is an entirely speculative pointer
3015 * access and not reliable.
3017 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3019 if (!sched_feat(OWNER_SPIN))
3023 while (owner_running(lock, owner)) {
3027 arch_mutex_cpu_relax();
3032 * We break out the loop above on need_resched() and when the
3033 * owner changed, which is a sign for heavy contention. Return
3034 * success only when lock->owner is NULL.
3036 return lock->owner == NULL;
3040 #ifdef CONFIG_PREEMPT
3042 * this is the entry point to schedule() from in-kernel preemption
3043 * off of preempt_enable. Kernel preemptions off return from interrupt
3044 * occur there and call schedule directly.
3046 asmlinkage void __sched notrace preempt_schedule(void)
3048 struct thread_info *ti = current_thread_info();
3051 * If there is a non-zero preempt_count or interrupts are disabled,
3052 * we do not want to preempt the current task. Just return..
3054 if (likely(ti->preempt_count || irqs_disabled()))
3058 add_preempt_count_notrace(PREEMPT_ACTIVE);
3060 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3063 * Check again in case we missed a preemption opportunity
3064 * between schedule and now.
3067 } while (need_resched());
3069 EXPORT_SYMBOL(preempt_schedule);
3072 * this is the entry point to schedule() from kernel preemption
3073 * off of irq context.
3074 * Note, that this is called and return with irqs disabled. This will
3075 * protect us against recursive calling from irq.
3077 asmlinkage void __sched preempt_schedule_irq(void)
3079 struct thread_info *ti = current_thread_info();
3081 /* Catch callers which need to be fixed */
3082 BUG_ON(ti->preempt_count || !irqs_disabled());
3086 add_preempt_count(PREEMPT_ACTIVE);
3089 local_irq_disable();
3090 sub_preempt_count(PREEMPT_ACTIVE);
3093 * Check again in case we missed a preemption opportunity
3094 * between schedule and now.
3097 } while (need_resched());
3100 #endif /* CONFIG_PREEMPT */
3102 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3105 return try_to_wake_up(curr->private, mode, wake_flags);
3107 EXPORT_SYMBOL(default_wake_function);
3110 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3111 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3112 * number) then we wake all the non-exclusive tasks and one exclusive task.
3114 * There are circumstances in which we can try to wake a task which has already
3115 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3116 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3118 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3119 int nr_exclusive, int wake_flags, void *key)
3121 wait_queue_t *curr, *next;
3123 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3124 unsigned flags = curr->flags;
3126 if (curr->func(curr, mode, wake_flags, key) &&
3127 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3133 * __wake_up - wake up threads blocked on a waitqueue.
3135 * @mode: which threads
3136 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3137 * @key: is directly passed to the wakeup function
3139 * It may be assumed that this function implies a write memory barrier before
3140 * changing the task state if and only if any tasks are woken up.
3142 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3143 int nr_exclusive, void *key)
3145 unsigned long flags;
3147 spin_lock_irqsave(&q->lock, flags);
3148 __wake_up_common(q, mode, nr_exclusive, 0, key);
3149 spin_unlock_irqrestore(&q->lock, flags);
3151 EXPORT_SYMBOL(__wake_up);
3154 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3156 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3158 __wake_up_common(q, mode, nr, 0, NULL);
3160 EXPORT_SYMBOL_GPL(__wake_up_locked);
3162 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3164 __wake_up_common(q, mode, 1, 0, key);
3166 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3169 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3171 * @mode: which threads
3172 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3173 * @key: opaque value to be passed to wakeup targets
3175 * The sync wakeup differs that the waker knows that it will schedule
3176 * away soon, so while the target thread will be woken up, it will not
3177 * be migrated to another CPU - ie. the two threads are 'synchronized'
3178 * with each other. This can prevent needless bouncing between CPUs.
3180 * On UP it can prevent extra preemption.
3182 * It may be assumed that this function implies a write memory barrier before
3183 * changing the task state if and only if any tasks are woken up.
3185 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3186 int nr_exclusive, void *key)
3188 unsigned long flags;
3189 int wake_flags = WF_SYNC;
3194 if (unlikely(!nr_exclusive))
3197 spin_lock_irqsave(&q->lock, flags);
3198 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3199 spin_unlock_irqrestore(&q->lock, flags);
3201 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3204 * __wake_up_sync - see __wake_up_sync_key()
3206 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3208 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3210 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3213 * complete: - signals a single thread waiting on this completion
3214 * @x: holds the state of this particular completion
3216 * This will wake up a single thread waiting on this completion. Threads will be
3217 * awakened in the same order in which they were queued.
3219 * See also complete_all(), wait_for_completion() and related routines.
3221 * It may be assumed that this function implies a write memory barrier before
3222 * changing the task state if and only if any tasks are woken up.
3224 void complete(struct completion *x)
3226 unsigned long flags;
3228 spin_lock_irqsave(&x->wait.lock, flags);
3230 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3231 spin_unlock_irqrestore(&x->wait.lock, flags);
3233 EXPORT_SYMBOL(complete);
3236 * complete_all: - signals all threads waiting on this completion
3237 * @x: holds the state of this particular completion
3239 * This will wake up all threads waiting on this particular completion event.
3241 * It may be assumed that this function implies a write memory barrier before
3242 * changing the task state if and only if any tasks are woken up.
3244 void complete_all(struct completion *x)
3246 unsigned long flags;
3248 spin_lock_irqsave(&x->wait.lock, flags);
3249 x->done += UINT_MAX/2;
3250 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3251 spin_unlock_irqrestore(&x->wait.lock, flags);
3253 EXPORT_SYMBOL(complete_all);
3255 static inline long __sched
3256 do_wait_for_common(struct completion *x,
3257 long (*action)(long), long timeout, int state)
3260 DECLARE_WAITQUEUE(wait, current);
3262 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3264 if (signal_pending_state(state, current)) {
3265 timeout = -ERESTARTSYS;
3268 __set_current_state(state);
3269 spin_unlock_irq(&x->wait.lock);
3270 timeout = action(timeout);
3271 spin_lock_irq(&x->wait.lock);
3272 } while (!x->done && timeout);
3273 __remove_wait_queue(&x->wait, &wait);
3278 return timeout ?: 1;
3281 static inline long __sched
3282 __wait_for_common(struct completion *x,
3283 long (*action)(long), long timeout, int state)
3287 spin_lock_irq(&x->wait.lock);
3288 timeout = do_wait_for_common(x, action, timeout, state);
3289 spin_unlock_irq(&x->wait.lock);
3294 wait_for_common(struct completion *x, long timeout, int state)
3296 return __wait_for_common(x, schedule_timeout, timeout, state);
3300 wait_for_common_io(struct completion *x, long timeout, int state)
3302 return __wait_for_common(x, io_schedule_timeout, timeout, state);
3306 * wait_for_completion: - waits for completion of a task
3307 * @x: holds the state of this particular completion
3309 * This waits to be signaled for completion of a specific task. It is NOT
3310 * interruptible and there is no timeout.
3312 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3313 * and interrupt capability. Also see complete().
3315 void __sched wait_for_completion(struct completion *x)
3317 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3319 EXPORT_SYMBOL(wait_for_completion);
3322 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3323 * @x: holds the state of this particular completion
3324 * @timeout: timeout value in jiffies
3326 * This waits for either a completion of a specific task to be signaled or for a
3327 * specified timeout to expire. The timeout is in jiffies. It is not
3330 * The return value is 0 if timed out, and positive (at least 1, or number of
3331 * jiffies left till timeout) if completed.
3333 unsigned long __sched
3334 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3336 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3338 EXPORT_SYMBOL(wait_for_completion_timeout);
3341 * wait_for_completion_io: - waits for completion of a task
3342 * @x: holds the state of this particular completion
3344 * This waits to be signaled for completion of a specific task. It is NOT
3345 * interruptible and there is no timeout. The caller is accounted as waiting
3348 void __sched wait_for_completion_io(struct completion *x)
3350 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3352 EXPORT_SYMBOL(wait_for_completion_io);
3355 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3356 * @x: holds the state of this particular completion
3357 * @timeout: timeout value in jiffies
3359 * This waits for either a completion of a specific task to be signaled or for a
3360 * specified timeout to expire. The timeout is in jiffies. It is not
3361 * interruptible. The caller is accounted as waiting for IO.
3363 * The return value is 0 if timed out, and positive (at least 1, or number of
3364 * jiffies left till timeout) if completed.
3366 unsigned long __sched
3367 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
3369 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
3371 EXPORT_SYMBOL(wait_for_completion_io_timeout);
3374 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3375 * @x: holds the state of this particular completion
3377 * This waits for completion of a specific task to be signaled. It is
3380 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3382 int __sched wait_for_completion_interruptible(struct completion *x)
3384 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3385 if (t == -ERESTARTSYS)
3389 EXPORT_SYMBOL(wait_for_completion_interruptible);
3392 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3393 * @x: holds the state of this particular completion
3394 * @timeout: timeout value in jiffies
3396 * This waits for either a completion of a specific task to be signaled or for a
3397 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3399 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3400 * positive (at least 1, or number of jiffies left till timeout) if completed.
3403 wait_for_completion_interruptible_timeout(struct completion *x,
3404 unsigned long timeout)
3406 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3408 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3411 * wait_for_completion_killable: - waits for completion of a task (killable)
3412 * @x: holds the state of this particular completion
3414 * This waits to be signaled for completion of a specific task. It can be
3415 * interrupted by a kill signal.
3417 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3419 int __sched wait_for_completion_killable(struct completion *x)
3421 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3422 if (t == -ERESTARTSYS)
3426 EXPORT_SYMBOL(wait_for_completion_killable);
3429 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3430 * @x: holds the state of this particular completion
3431 * @timeout: timeout value in jiffies
3433 * This waits for either a completion of a specific task to be
3434 * signaled or for a specified timeout to expire. It can be
3435 * interrupted by a kill signal. The timeout is in jiffies.
3437 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3438 * positive (at least 1, or number of jiffies left till timeout) if completed.
3441 wait_for_completion_killable_timeout(struct completion *x,
3442 unsigned long timeout)
3444 return wait_for_common(x, timeout, TASK_KILLABLE);
3446 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3449 * try_wait_for_completion - try to decrement a completion without blocking
3450 * @x: completion structure
3452 * Returns: 0 if a decrement cannot be done without blocking
3453 * 1 if a decrement succeeded.
3455 * If a completion is being used as a counting completion,
3456 * attempt to decrement the counter without blocking. This
3457 * enables us to avoid waiting if the resource the completion
3458 * is protecting is not available.
3460 bool try_wait_for_completion(struct completion *x)
3462 unsigned long flags;
3465 spin_lock_irqsave(&x->wait.lock, flags);
3470 spin_unlock_irqrestore(&x->wait.lock, flags);
3473 EXPORT_SYMBOL(try_wait_for_completion);
3476 * completion_done - Test to see if a completion has any waiters
3477 * @x: completion structure
3479 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3480 * 1 if there are no waiters.
3483 bool completion_done(struct completion *x)
3485 unsigned long flags;
3488 spin_lock_irqsave(&x->wait.lock, flags);
3491 spin_unlock_irqrestore(&x->wait.lock, flags);
3494 EXPORT_SYMBOL(completion_done);
3497 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3499 unsigned long flags;
3502 init_waitqueue_entry(&wait, current);
3504 __set_current_state(state);
3506 spin_lock_irqsave(&q->lock, flags);
3507 __add_wait_queue(q, &wait);
3508 spin_unlock(&q->lock);
3509 timeout = schedule_timeout(timeout);
3510 spin_lock_irq(&q->lock);
3511 __remove_wait_queue(q, &wait);
3512 spin_unlock_irqrestore(&q->lock, flags);
3517 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3519 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3521 EXPORT_SYMBOL(interruptible_sleep_on);
3524 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3526 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3528 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3530 void __sched sleep_on(wait_queue_head_t *q)
3532 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3534 EXPORT_SYMBOL(sleep_on);
3536 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3538 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3540 EXPORT_SYMBOL(sleep_on_timeout);
3542 #ifdef CONFIG_RT_MUTEXES
3545 * rt_mutex_setprio - set the current priority of a task
3547 * @prio: prio value (kernel-internal form)
3549 * This function changes the 'effective' priority of a task. It does
3550 * not touch ->normal_prio like __setscheduler().
3552 * Used by the rt_mutex code to implement priority inheritance logic.
3554 void rt_mutex_setprio(struct task_struct *p, int prio)
3556 int oldprio, on_rq, running;
3558 const struct sched_class *prev_class;
3560 BUG_ON(prio < 0 || prio > MAX_PRIO);
3562 rq = __task_rq_lock(p);
3565 * Idle task boosting is a nono in general. There is one
3566 * exception, when PREEMPT_RT and NOHZ is active:
3568 * The idle task calls get_next_timer_interrupt() and holds
3569 * the timer wheel base->lock on the CPU and another CPU wants
3570 * to access the timer (probably to cancel it). We can safely
3571 * ignore the boosting request, as the idle CPU runs this code
3572 * with interrupts disabled and will complete the lock
3573 * protected section without being interrupted. So there is no
3574 * real need to boost.
3576 if (unlikely(p == rq->idle)) {
3577 WARN_ON(p != rq->curr);
3578 WARN_ON(p->pi_blocked_on);
3582 trace_sched_pi_setprio(p, prio);
3584 prev_class = p->sched_class;
3586 running = task_current(rq, p);
3588 dequeue_task(rq, p, 0);
3590 p->sched_class->put_prev_task(rq, p);
3593 p->sched_class = &rt_sched_class;
3595 p->sched_class = &fair_sched_class;
3600 p->sched_class->set_curr_task(rq);
3602 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3604 check_class_changed(rq, p, prev_class, oldprio);
3606 __task_rq_unlock(rq);
3609 void set_user_nice(struct task_struct *p, long nice)
3611 int old_prio, delta, on_rq;
3612 unsigned long flags;
3615 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3618 * We have to be careful, if called from sys_setpriority(),
3619 * the task might be in the middle of scheduling on another CPU.
3621 rq = task_rq_lock(p, &flags);
3623 * The RT priorities are set via sched_setscheduler(), but we still
3624 * allow the 'normal' nice value to be set - but as expected
3625 * it wont have any effect on scheduling until the task is
3626 * SCHED_FIFO/SCHED_RR:
3628 if (task_has_rt_policy(p)) {
3629 p->static_prio = NICE_TO_PRIO(nice);
3634 dequeue_task(rq, p, 0);
3636 p->static_prio = NICE_TO_PRIO(nice);
3639 p->prio = effective_prio(p);
3640 delta = p->prio - old_prio;
3643 enqueue_task(rq, p, 0);
3645 * If the task increased its priority or is running and
3646 * lowered its priority, then reschedule its CPU:
3648 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3649 resched_task(rq->curr);
3652 task_rq_unlock(rq, p, &flags);
3654 EXPORT_SYMBOL(set_user_nice);
3657 * can_nice - check if a task can reduce its nice value
3661 int can_nice(const struct task_struct *p, const int nice)
3663 /* convert nice value [19,-20] to rlimit style value [1,40] */
3664 int nice_rlim = 20 - nice;
3666 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3667 capable(CAP_SYS_NICE));
3670 #ifdef __ARCH_WANT_SYS_NICE
3673 * sys_nice - change the priority of the current process.
3674 * @increment: priority increment
3676 * sys_setpriority is a more generic, but much slower function that
3677 * does similar things.
3679 SYSCALL_DEFINE1(nice, int, increment)
3684 * Setpriority might change our priority at the same moment.
3685 * We don't have to worry. Conceptually one call occurs first
3686 * and we have a single winner.
3688 if (increment < -40)
3693 nice = TASK_NICE(current) + increment;
3699 if (increment < 0 && !can_nice(current, nice))
3702 retval = security_task_setnice(current, nice);
3706 set_user_nice(current, nice);
3713 * task_prio - return the priority value of a given task.
3714 * @p: the task in question.
3716 * This is the priority value as seen by users in /proc.
3717 * RT tasks are offset by -200. Normal tasks are centered
3718 * around 0, value goes from -16 to +15.
3720 int task_prio(const struct task_struct *p)
3722 return p->prio - MAX_RT_PRIO;
3726 * task_nice - return the nice value of a given task.
3727 * @p: the task in question.
3729 int task_nice(const struct task_struct *p)
3731 return TASK_NICE(p);
3733 EXPORT_SYMBOL(task_nice);
3736 * idle_cpu - is a given cpu idle currently?
3737 * @cpu: the processor in question.
3739 int idle_cpu(int cpu)
3741 struct rq *rq = cpu_rq(cpu);
3743 if (rq->curr != rq->idle)
3750 if (!llist_empty(&rq->wake_list))
3758 * idle_task - return the idle task for a given cpu.
3759 * @cpu: the processor in question.
3761 struct task_struct *idle_task(int cpu)
3763 return cpu_rq(cpu)->idle;
3767 * find_process_by_pid - find a process with a matching PID value.
3768 * @pid: the pid in question.
3770 static struct task_struct *find_process_by_pid(pid_t pid)
3772 return pid ? find_task_by_vpid(pid) : current;
3775 /* Actually do priority change: must hold rq lock. */
3777 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3780 p->rt_priority = prio;
3781 p->normal_prio = normal_prio(p);
3782 /* we are holding p->pi_lock already */
3783 p->prio = rt_mutex_getprio(p);
3784 if (rt_prio(p->prio))
3785 p->sched_class = &rt_sched_class;
3787 p->sched_class = &fair_sched_class;
3792 * check the target process has a UID that matches the current process's
3794 static bool check_same_owner(struct task_struct *p)
3796 const struct cred *cred = current_cred(), *pcred;
3800 pcred = __task_cred(p);
3801 match = (uid_eq(cred->euid, pcred->euid) ||
3802 uid_eq(cred->euid, pcred->uid));
3807 static int __sched_setscheduler(struct task_struct *p, int policy,
3808 const struct sched_param *param, bool user)
3810 int retval, oldprio, oldpolicy = -1, on_rq, running;
3811 unsigned long flags;
3812 const struct sched_class *prev_class;
3816 /* may grab non-irq protected spin_locks */
3817 BUG_ON(in_interrupt());
3819 /* double check policy once rq lock held */
3821 reset_on_fork = p->sched_reset_on_fork;
3822 policy = oldpolicy = p->policy;
3824 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3825 policy &= ~SCHED_RESET_ON_FORK;
3827 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3828 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3829 policy != SCHED_IDLE)
3834 * Valid priorities for SCHED_FIFO and SCHED_RR are
3835 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3836 * SCHED_BATCH and SCHED_IDLE is 0.
3838 if (param->sched_priority < 0 ||
3839 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3840 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3842 if (rt_policy(policy) != (param->sched_priority != 0))
3846 * Allow unprivileged RT tasks to decrease priority:
3848 if (user && !capable(CAP_SYS_NICE)) {
3849 if (rt_policy(policy)) {
3850 unsigned long rlim_rtprio =
3851 task_rlimit(p, RLIMIT_RTPRIO);
3853 /* can't set/change the rt policy */
3854 if (policy != p->policy && !rlim_rtprio)
3857 /* can't increase priority */
3858 if (param->sched_priority > p->rt_priority &&
3859 param->sched_priority > rlim_rtprio)
3864 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3865 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3867 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3868 if (!can_nice(p, TASK_NICE(p)))
3872 /* can't change other user's priorities */
3873 if (!check_same_owner(p))
3876 /* Normal users shall not reset the sched_reset_on_fork flag */
3877 if (p->sched_reset_on_fork && !reset_on_fork)
3882 retval = security_task_setscheduler(p);
3888 * make sure no PI-waiters arrive (or leave) while we are
3889 * changing the priority of the task:
3891 * To be able to change p->policy safely, the appropriate
3892 * runqueue lock must be held.
3894 rq = task_rq_lock(p, &flags);
3897 * Changing the policy of the stop threads its a very bad idea
3899 if (p == rq->stop) {
3900 task_rq_unlock(rq, p, &flags);
3905 * If not changing anything there's no need to proceed further:
3907 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3908 param->sched_priority == p->rt_priority))) {
3909 task_rq_unlock(rq, p, &flags);
3913 #ifdef CONFIG_RT_GROUP_SCHED
3916 * Do not allow realtime tasks into groups that have no runtime
3919 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3920 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3921 !task_group_is_autogroup(task_group(p))) {
3922 task_rq_unlock(rq, p, &flags);
3928 /* recheck policy now with rq lock held */
3929 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3930 policy = oldpolicy = -1;
3931 task_rq_unlock(rq, p, &flags);
3935 running = task_current(rq, p);
3937 dequeue_task(rq, p, 0);
3939 p->sched_class->put_prev_task(rq, p);
3941 p->sched_reset_on_fork = reset_on_fork;
3944 prev_class = p->sched_class;
3945 __setscheduler(rq, p, policy, param->sched_priority);
3948 p->sched_class->set_curr_task(rq);
3950 enqueue_task(rq, p, 0);
3952 check_class_changed(rq, p, prev_class, oldprio);
3953 task_rq_unlock(rq, p, &flags);
3955 rt_mutex_adjust_pi(p);
3961 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3962 * @p: the task in question.
3963 * @policy: new policy.
3964 * @param: structure containing the new RT priority.
3966 * NOTE that the task may be already dead.
3968 int sched_setscheduler(struct task_struct *p, int policy,
3969 const struct sched_param *param)
3971 return __sched_setscheduler(p, policy, param, true);
3973 EXPORT_SYMBOL_GPL(sched_setscheduler);
3976 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3977 * @p: the task in question.
3978 * @policy: new policy.
3979 * @param: structure containing the new RT priority.
3981 * Just like sched_setscheduler, only don't bother checking if the
3982 * current context has permission. For example, this is needed in
3983 * stop_machine(): we create temporary high priority worker threads,
3984 * but our caller might not have that capability.
3986 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3987 const struct sched_param *param)
3989 return __sched_setscheduler(p, policy, param, false);
3993 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3995 struct sched_param lparam;
3996 struct task_struct *p;
3999 if (!param || pid < 0)
4001 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4006 p = find_process_by_pid(pid);
4008 retval = sched_setscheduler(p, policy, &lparam);
4015 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4016 * @pid: the pid in question.
4017 * @policy: new policy.
4018 * @param: structure containing the new RT priority.
4020 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4021 struct sched_param __user *, param)
4023 /* negative values for policy are not valid */
4027 return do_sched_setscheduler(pid, policy, param);
4031 * sys_sched_setparam - set/change the RT priority of a thread
4032 * @pid: the pid in question.
4033 * @param: structure containing the new RT priority.
4035 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4037 return do_sched_setscheduler(pid, -1, param);
4041 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4042 * @pid: the pid in question.
4044 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4046 struct task_struct *p;
4054 p = find_process_by_pid(pid);
4056 retval = security_task_getscheduler(p);
4059 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4066 * sys_sched_getparam - get the RT priority of a thread
4067 * @pid: the pid in question.
4068 * @param: structure containing the RT priority.
4070 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4072 struct sched_param lp;
4073 struct task_struct *p;
4076 if (!param || pid < 0)
4080 p = find_process_by_pid(pid);
4085 retval = security_task_getscheduler(p);
4089 lp.sched_priority = p->rt_priority;
4093 * This one might sleep, we cannot do it with a spinlock held ...
4095 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4104 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4106 cpumask_var_t cpus_allowed, new_mask;
4107 struct task_struct *p;
4113 p = find_process_by_pid(pid);
4120 /* Prevent p going away */
4124 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4128 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4130 goto out_free_cpus_allowed;
4133 if (!check_same_owner(p)) {
4135 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4142 retval = security_task_setscheduler(p);
4146 cpuset_cpus_allowed(p, cpus_allowed);
4147 cpumask_and(new_mask, in_mask, cpus_allowed);
4149 retval = set_cpus_allowed_ptr(p, new_mask);
4152 cpuset_cpus_allowed(p, cpus_allowed);
4153 if (!cpumask_subset(new_mask, cpus_allowed)) {
4155 * We must have raced with a concurrent cpuset
4156 * update. Just reset the cpus_allowed to the
4157 * cpuset's cpus_allowed
4159 cpumask_copy(new_mask, cpus_allowed);
4164 free_cpumask_var(new_mask);
4165 out_free_cpus_allowed:
4166 free_cpumask_var(cpus_allowed);
4173 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4174 struct cpumask *new_mask)
4176 if (len < cpumask_size())
4177 cpumask_clear(new_mask);
4178 else if (len > cpumask_size())
4179 len = cpumask_size();
4181 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4185 * sys_sched_setaffinity - set the cpu affinity of a process
4186 * @pid: pid of the process
4187 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4188 * @user_mask_ptr: user-space pointer to the new cpu mask
4190 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4191 unsigned long __user *, user_mask_ptr)
4193 cpumask_var_t new_mask;
4196 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4199 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4201 retval = sched_setaffinity(pid, new_mask);
4202 free_cpumask_var(new_mask);
4206 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4208 struct task_struct *p;
4209 unsigned long flags;
4216 p = find_process_by_pid(pid);
4220 retval = security_task_getscheduler(p);
4224 raw_spin_lock_irqsave(&p->pi_lock, flags);
4225 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4226 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4236 * sys_sched_getaffinity - get the cpu affinity of a process
4237 * @pid: pid of the process
4238 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4239 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4241 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4242 unsigned long __user *, user_mask_ptr)
4247 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4249 if (len & (sizeof(unsigned long)-1))
4252 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4255 ret = sched_getaffinity(pid, mask);
4257 size_t retlen = min_t(size_t, len, cpumask_size());
4259 if (copy_to_user(user_mask_ptr, mask, retlen))
4264 free_cpumask_var(mask);
4270 * sys_sched_yield - yield the current processor to other threads.
4272 * This function yields the current CPU to other tasks. If there are no
4273 * other threads running on this CPU then this function will return.
4275 SYSCALL_DEFINE0(sched_yield)
4277 struct rq *rq = this_rq_lock();
4279 schedstat_inc(rq, yld_count);
4280 current->sched_class->yield_task(rq);
4283 * Since we are going to call schedule() anyway, there's
4284 * no need to preempt or enable interrupts:
4286 __release(rq->lock);
4287 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4288 do_raw_spin_unlock(&rq->lock);
4289 sched_preempt_enable_no_resched();
4296 static inline int should_resched(void)
4298 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4301 static void __cond_resched(void)
4303 add_preempt_count(PREEMPT_ACTIVE);
4305 sub_preempt_count(PREEMPT_ACTIVE);
4308 int __sched _cond_resched(void)
4310 if (should_resched()) {
4316 EXPORT_SYMBOL(_cond_resched);
4319 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4320 * call schedule, and on return reacquire the lock.
4322 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4323 * operations here to prevent schedule() from being called twice (once via
4324 * spin_unlock(), once by hand).
4326 int __cond_resched_lock(spinlock_t *lock)
4328 int resched = should_resched();
4331 lockdep_assert_held(lock);
4333 if (spin_needbreak(lock) || resched) {
4344 EXPORT_SYMBOL(__cond_resched_lock);
4346 int __sched __cond_resched_softirq(void)
4348 BUG_ON(!in_softirq());
4350 if (should_resched()) {
4358 EXPORT_SYMBOL(__cond_resched_softirq);
4361 * yield - yield the current processor to other threads.
4363 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4365 * The scheduler is at all times free to pick the calling task as the most
4366 * eligible task to run, if removing the yield() call from your code breaks
4367 * it, its already broken.
4369 * Typical broken usage is:
4374 * where one assumes that yield() will let 'the other' process run that will
4375 * make event true. If the current task is a SCHED_FIFO task that will never
4376 * happen. Never use yield() as a progress guarantee!!
4378 * If you want to use yield() to wait for something, use wait_event().
4379 * If you want to use yield() to be 'nice' for others, use cond_resched().
4380 * If you still want to use yield(), do not!
4382 void __sched yield(void)
4384 set_current_state(TASK_RUNNING);
4387 EXPORT_SYMBOL(yield);
4390 * yield_to - yield the current processor to another thread in
4391 * your thread group, or accelerate that thread toward the
4392 * processor it's on.
4394 * @preempt: whether task preemption is allowed or not
4396 * It's the caller's job to ensure that the target task struct
4397 * can't go away on us before we can do any checks.
4400 * true (>0) if we indeed boosted the target task.
4401 * false (0) if we failed to boost the target.
4402 * -ESRCH if there's no task to yield to.
4404 bool __sched yield_to(struct task_struct *p, bool preempt)
4406 struct task_struct *curr = current;
4407 struct rq *rq, *p_rq;
4408 unsigned long flags;
4411 local_irq_save(flags);
4417 * If we're the only runnable task on the rq and target rq also
4418 * has only one task, there's absolutely no point in yielding.
4420 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4425 double_rq_lock(rq, p_rq);
4426 while (task_rq(p) != p_rq) {
4427 double_rq_unlock(rq, p_rq);
4431 if (!curr->sched_class->yield_to_task)
4434 if (curr->sched_class != p->sched_class)
4437 if (task_running(p_rq, p) || p->state)
4440 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4442 schedstat_inc(rq, yld_count);
4444 * Make p's CPU reschedule; pick_next_entity takes care of
4447 if (preempt && rq != p_rq)
4448 resched_task(p_rq->curr);
4452 double_rq_unlock(rq, p_rq);
4454 local_irq_restore(flags);
4461 EXPORT_SYMBOL_GPL(yield_to);
4464 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4465 * that process accounting knows that this is a task in IO wait state.
4467 void __sched io_schedule(void)
4469 struct rq *rq = raw_rq();
4471 delayacct_blkio_start();
4472 atomic_inc(&rq->nr_iowait);
4473 blk_flush_plug(current);
4474 current->in_iowait = 1;
4476 current->in_iowait = 0;
4477 atomic_dec(&rq->nr_iowait);
4478 delayacct_blkio_end();
4480 EXPORT_SYMBOL(io_schedule);
4482 long __sched io_schedule_timeout(long timeout)
4484 struct rq *rq = raw_rq();
4487 delayacct_blkio_start();
4488 atomic_inc(&rq->nr_iowait);
4489 blk_flush_plug(current);
4490 current->in_iowait = 1;
4491 ret = schedule_timeout(timeout);
4492 current->in_iowait = 0;
4493 atomic_dec(&rq->nr_iowait);
4494 delayacct_blkio_end();
4499 * sys_sched_get_priority_max - return maximum RT priority.
4500 * @policy: scheduling class.
4502 * this syscall returns the maximum rt_priority that can be used
4503 * by a given scheduling class.
4505 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4512 ret = MAX_USER_RT_PRIO-1;
4524 * sys_sched_get_priority_min - return minimum RT priority.
4525 * @policy: scheduling class.
4527 * this syscall returns the minimum rt_priority that can be used
4528 * by a given scheduling class.
4530 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4548 * sys_sched_rr_get_interval - return the default timeslice of a process.
4549 * @pid: pid of the process.
4550 * @interval: userspace pointer to the timeslice value.
4552 * this syscall writes the default timeslice value of a given process
4553 * into the user-space timespec buffer. A value of '0' means infinity.
4555 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4556 struct timespec __user *, interval)
4558 struct task_struct *p;
4559 unsigned int time_slice;
4560 unsigned long flags;
4570 p = find_process_by_pid(pid);
4574 retval = security_task_getscheduler(p);
4578 rq = task_rq_lock(p, &flags);
4579 time_slice = p->sched_class->get_rr_interval(rq, p);
4580 task_rq_unlock(rq, p, &flags);
4583 jiffies_to_timespec(time_slice, &t);
4584 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4592 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4594 void sched_show_task(struct task_struct *p)
4596 unsigned long free = 0;
4600 state = p->state ? __ffs(p->state) + 1 : 0;
4601 printk(KERN_INFO "%-15.15s %c", p->comm,
4602 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4603 #if BITS_PER_LONG == 32
4604 if (state == TASK_RUNNING)
4605 printk(KERN_CONT " running ");
4607 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4609 if (state == TASK_RUNNING)
4610 printk(KERN_CONT " running task ");
4612 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4614 #ifdef CONFIG_DEBUG_STACK_USAGE
4615 free = stack_not_used(p);
4618 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4620 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4621 task_pid_nr(p), ppid,
4622 (unsigned long)task_thread_info(p)->flags);
4624 show_stack(p, NULL);
4627 void show_state_filter(unsigned long state_filter)
4629 struct task_struct *g, *p;
4631 #if BITS_PER_LONG == 32
4633 " task PC stack pid father\n");
4636 " task PC stack pid father\n");
4639 do_each_thread(g, p) {
4641 * reset the NMI-timeout, listing all files on a slow
4642 * console might take a lot of time:
4644 touch_nmi_watchdog();
4645 if (!state_filter || (p->state & state_filter))
4647 } while_each_thread(g, p);
4649 touch_all_softlockup_watchdogs();
4651 #ifdef CONFIG_SCHED_DEBUG
4652 sysrq_sched_debug_show();
4656 * Only show locks if all tasks are dumped:
4659 debug_show_all_locks();
4662 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4664 idle->sched_class = &idle_sched_class;
4668 * init_idle - set up an idle thread for a given CPU
4669 * @idle: task in question
4670 * @cpu: cpu the idle task belongs to
4672 * NOTE: this function does not set the idle thread's NEED_RESCHED
4673 * flag, to make booting more robust.
4675 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4677 struct rq *rq = cpu_rq(cpu);
4678 unsigned long flags;
4680 raw_spin_lock_irqsave(&rq->lock, flags);
4683 idle->state = TASK_RUNNING;
4684 idle->se.exec_start = sched_clock();
4686 do_set_cpus_allowed(idle, cpumask_of(cpu));
4688 * We're having a chicken and egg problem, even though we are
4689 * holding rq->lock, the cpu isn't yet set to this cpu so the
4690 * lockdep check in task_group() will fail.
4692 * Similar case to sched_fork(). / Alternatively we could
4693 * use task_rq_lock() here and obtain the other rq->lock.
4698 __set_task_cpu(idle, cpu);
4701 rq->curr = rq->idle = idle;
4702 #if defined(CONFIG_SMP)
4705 raw_spin_unlock_irqrestore(&rq->lock, flags);
4707 /* Set the preempt count _outside_ the spinlocks! */
4708 task_thread_info(idle)->preempt_count = 0;
4711 * The idle tasks have their own, simple scheduling class:
4713 idle->sched_class = &idle_sched_class;
4714 ftrace_graph_init_idle_task(idle, cpu);
4715 vtime_init_idle(idle);
4716 #if defined(CONFIG_SMP)
4717 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4722 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4724 if (p->sched_class && p->sched_class->set_cpus_allowed)
4725 p->sched_class->set_cpus_allowed(p, new_mask);
4727 cpumask_copy(&p->cpus_allowed, new_mask);
4728 p->nr_cpus_allowed = cpumask_weight(new_mask);
4732 * This is how migration works:
4734 * 1) we invoke migration_cpu_stop() on the target CPU using
4736 * 2) stopper starts to run (implicitly forcing the migrated thread
4738 * 3) it checks whether the migrated task is still in the wrong runqueue.
4739 * 4) if it's in the wrong runqueue then the migration thread removes
4740 * it and puts it into the right queue.
4741 * 5) stopper completes and stop_one_cpu() returns and the migration
4746 * Change a given task's CPU affinity. Migrate the thread to a
4747 * proper CPU and schedule it away if the CPU it's executing on
4748 * is removed from the allowed bitmask.
4750 * NOTE: the caller must have a valid reference to the task, the
4751 * task must not exit() & deallocate itself prematurely. The
4752 * call is not atomic; no spinlocks may be held.
4754 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4756 unsigned long flags;
4758 unsigned int dest_cpu;
4761 rq = task_rq_lock(p, &flags);
4763 if (cpumask_equal(&p->cpus_allowed, new_mask))
4766 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4771 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4776 do_set_cpus_allowed(p, new_mask);
4778 /* Can the task run on the task's current CPU? If so, we're done */
4779 if (cpumask_test_cpu(task_cpu(p), new_mask))
4782 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4784 struct migration_arg arg = { p, dest_cpu };
4785 /* Need help from migration thread: drop lock and wait. */
4786 task_rq_unlock(rq, p, &flags);
4787 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4788 tlb_migrate_finish(p->mm);
4792 task_rq_unlock(rq, p, &flags);
4796 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4799 * Move (not current) task off this cpu, onto dest cpu. We're doing
4800 * this because either it can't run here any more (set_cpus_allowed()
4801 * away from this CPU, or CPU going down), or because we're
4802 * attempting to rebalance this task on exec (sched_exec).
4804 * So we race with normal scheduler movements, but that's OK, as long
4805 * as the task is no longer on this CPU.
4807 * Returns non-zero if task was successfully migrated.
4809 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4811 struct rq *rq_dest, *rq_src;
4814 if (unlikely(!cpu_active(dest_cpu)))
4817 rq_src = cpu_rq(src_cpu);
4818 rq_dest = cpu_rq(dest_cpu);
4820 raw_spin_lock(&p->pi_lock);
4821 double_rq_lock(rq_src, rq_dest);
4822 /* Already moved. */
4823 if (task_cpu(p) != src_cpu)
4825 /* Affinity changed (again). */
4826 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4830 * If we're not on a rq, the next wake-up will ensure we're
4834 dequeue_task(rq_src, p, 0);
4835 set_task_cpu(p, dest_cpu);
4836 enqueue_task(rq_dest, p, 0);
4837 check_preempt_curr(rq_dest, p, 0);
4842 double_rq_unlock(rq_src, rq_dest);
4843 raw_spin_unlock(&p->pi_lock);
4848 * migration_cpu_stop - this will be executed by a highprio stopper thread
4849 * and performs thread migration by bumping thread off CPU then
4850 * 'pushing' onto another runqueue.
4852 static int migration_cpu_stop(void *data)
4854 struct migration_arg *arg = data;
4857 * The original target cpu might have gone down and we might
4858 * be on another cpu but it doesn't matter.
4860 local_irq_disable();
4861 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4866 #ifdef CONFIG_HOTPLUG_CPU
4869 * Ensures that the idle task is using init_mm right before its cpu goes
4872 void idle_task_exit(void)
4874 struct mm_struct *mm = current->active_mm;
4876 BUG_ON(cpu_online(smp_processor_id()));
4879 switch_mm(mm, &init_mm, current);
4884 * Since this CPU is going 'away' for a while, fold any nr_active delta
4885 * we might have. Assumes we're called after migrate_tasks() so that the
4886 * nr_active count is stable.
4888 * Also see the comment "Global load-average calculations".
4890 static void calc_load_migrate(struct rq *rq)
4892 long delta = calc_load_fold_active(rq);
4894 atomic_long_add(delta, &calc_load_tasks);
4898 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4899 * try_to_wake_up()->select_task_rq().
4901 * Called with rq->lock held even though we'er in stop_machine() and
4902 * there's no concurrency possible, we hold the required locks anyway
4903 * because of lock validation efforts.
4905 static void migrate_tasks(unsigned int dead_cpu)
4907 struct rq *rq = cpu_rq(dead_cpu);
4908 struct task_struct *next, *stop = rq->stop;
4912 * Fudge the rq selection such that the below task selection loop
4913 * doesn't get stuck on the currently eligible stop task.
4915 * We're currently inside stop_machine() and the rq is either stuck
4916 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4917 * either way we should never end up calling schedule() until we're
4924 * There's this thread running, bail when that's the only
4927 if (rq->nr_running == 1)
4930 next = pick_next_task(rq);
4932 next->sched_class->put_prev_task(rq, next);
4934 /* Find suitable destination for @next, with force if needed. */
4935 dest_cpu = select_fallback_rq(dead_cpu, next);
4936 raw_spin_unlock(&rq->lock);
4938 __migrate_task(next, dead_cpu, dest_cpu);
4940 raw_spin_lock(&rq->lock);
4946 #endif /* CONFIG_HOTPLUG_CPU */
4948 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4950 static struct ctl_table sd_ctl_dir[] = {
4952 .procname = "sched_domain",
4958 static struct ctl_table sd_ctl_root[] = {
4960 .procname = "kernel",
4962 .child = sd_ctl_dir,
4967 static struct ctl_table *sd_alloc_ctl_entry(int n)
4969 struct ctl_table *entry =
4970 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4975 static void sd_free_ctl_entry(struct ctl_table **tablep)
4977 struct ctl_table *entry;
4980 * In the intermediate directories, both the child directory and
4981 * procname are dynamically allocated and could fail but the mode
4982 * will always be set. In the lowest directory the names are
4983 * static strings and all have proc handlers.
4985 for (entry = *tablep; entry->mode; entry++) {
4987 sd_free_ctl_entry(&entry->child);
4988 if (entry->proc_handler == NULL)
4989 kfree(entry->procname);
4996 static int min_load_idx = 0;
4997 static int max_load_idx = CPU_LOAD_IDX_MAX;
5000 set_table_entry(struct ctl_table *entry,
5001 const char *procname, void *data, int maxlen,
5002 umode_t mode, proc_handler *proc_handler,
5005 entry->procname = procname;
5007 entry->maxlen = maxlen;
5009 entry->proc_handler = proc_handler;
5012 entry->extra1 = &min_load_idx;
5013 entry->extra2 = &max_load_idx;
5017 static struct ctl_table *
5018 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5020 struct ctl_table *table = sd_alloc_ctl_entry(13);
5025 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5026 sizeof(long), 0644, proc_doulongvec_minmax, false);
5027 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5028 sizeof(long), 0644, proc_doulongvec_minmax, false);
5029 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5030 sizeof(int), 0644, proc_dointvec_minmax, true);
5031 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5032 sizeof(int), 0644, proc_dointvec_minmax, true);
5033 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5034 sizeof(int), 0644, proc_dointvec_minmax, true);
5035 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5036 sizeof(int), 0644, proc_dointvec_minmax, true);
5037 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5038 sizeof(int), 0644, proc_dointvec_minmax, true);
5039 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5040 sizeof(int), 0644, proc_dointvec_minmax, false);
5041 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5042 sizeof(int), 0644, proc_dointvec_minmax, false);
5043 set_table_entry(&table[9], "cache_nice_tries",
5044 &sd->cache_nice_tries,
5045 sizeof(int), 0644, proc_dointvec_minmax, false);
5046 set_table_entry(&table[10], "flags", &sd->flags,
5047 sizeof(int), 0644, proc_dointvec_minmax, false);
5048 set_table_entry(&table[11], "name", sd->name,
5049 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5050 /* &table[12] is terminator */
5055 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5057 struct ctl_table *entry, *table;
5058 struct sched_domain *sd;
5059 int domain_num = 0, i;
5062 for_each_domain(cpu, sd)
5064 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5069 for_each_domain(cpu, sd) {
5070 snprintf(buf, 32, "domain%d", i);
5071 entry->procname = kstrdup(buf, GFP_KERNEL);
5073 entry->child = sd_alloc_ctl_domain_table(sd);
5080 static struct ctl_table_header *sd_sysctl_header;
5081 static void register_sched_domain_sysctl(void)
5083 int i, cpu_num = num_possible_cpus();
5084 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5087 WARN_ON(sd_ctl_dir[0].child);
5088 sd_ctl_dir[0].child = entry;
5093 for_each_possible_cpu(i) {
5094 snprintf(buf, 32, "cpu%d", i);
5095 entry->procname = kstrdup(buf, GFP_KERNEL);
5097 entry->child = sd_alloc_ctl_cpu_table(i);
5101 WARN_ON(sd_sysctl_header);
5102 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5105 /* may be called multiple times per register */
5106 static void unregister_sched_domain_sysctl(void)
5108 if (sd_sysctl_header)
5109 unregister_sysctl_table(sd_sysctl_header);
5110 sd_sysctl_header = NULL;
5111 if (sd_ctl_dir[0].child)
5112 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5115 static void register_sched_domain_sysctl(void)
5118 static void unregister_sched_domain_sysctl(void)
5123 static void set_rq_online(struct rq *rq)
5126 const struct sched_class *class;
5128 cpumask_set_cpu(rq->cpu, rq->rd->online);
5131 for_each_class(class) {
5132 if (class->rq_online)
5133 class->rq_online(rq);
5138 static void set_rq_offline(struct rq *rq)
5141 const struct sched_class *class;
5143 for_each_class(class) {
5144 if (class->rq_offline)
5145 class->rq_offline(rq);
5148 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5154 * migration_call - callback that gets triggered when a CPU is added.
5155 * Here we can start up the necessary migration thread for the new CPU.
5157 static int __cpuinit
5158 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5160 int cpu = (long)hcpu;
5161 unsigned long flags;
5162 struct rq *rq = cpu_rq(cpu);
5164 switch (action & ~CPU_TASKS_FROZEN) {
5166 case CPU_UP_PREPARE:
5167 rq->calc_load_update = calc_load_update;
5171 /* Update our root-domain */
5172 raw_spin_lock_irqsave(&rq->lock, flags);
5174 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5178 raw_spin_unlock_irqrestore(&rq->lock, flags);
5181 #ifdef CONFIG_HOTPLUG_CPU
5183 sched_ttwu_pending();
5184 /* Update our root-domain */
5185 raw_spin_lock_irqsave(&rq->lock, flags);
5187 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5191 BUG_ON(rq->nr_running != 1); /* the migration thread */
5192 raw_spin_unlock_irqrestore(&rq->lock, flags);
5196 calc_load_migrate(rq);
5201 update_max_interval();
5207 * Register at high priority so that task migration (migrate_all_tasks)
5208 * happens before everything else. This has to be lower priority than
5209 * the notifier in the perf_event subsystem, though.
5211 static struct notifier_block __cpuinitdata migration_notifier = {
5212 .notifier_call = migration_call,
5213 .priority = CPU_PRI_MIGRATION,
5216 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5217 unsigned long action, void *hcpu)
5219 switch (action & ~CPU_TASKS_FROZEN) {
5221 case CPU_DOWN_FAILED:
5222 set_cpu_active((long)hcpu, true);
5229 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5230 unsigned long action, void *hcpu)
5232 switch (action & ~CPU_TASKS_FROZEN) {
5233 case CPU_DOWN_PREPARE:
5234 set_cpu_active((long)hcpu, false);
5241 static int __init migration_init(void)
5243 void *cpu = (void *)(long)smp_processor_id();
5246 /* Initialize migration for the boot CPU */
5247 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5248 BUG_ON(err == NOTIFY_BAD);
5249 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5250 register_cpu_notifier(&migration_notifier);
5252 /* Register cpu active notifiers */
5253 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5254 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5258 early_initcall(migration_init);
5263 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5265 #ifdef CONFIG_SCHED_DEBUG
5267 static __read_mostly int sched_debug_enabled;
5269 static int __init sched_debug_setup(char *str)
5271 sched_debug_enabled = 1;
5275 early_param("sched_debug", sched_debug_setup);
5277 static inline bool sched_debug(void)
5279 return sched_debug_enabled;
5282 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5283 struct cpumask *groupmask)
5285 struct sched_group *group = sd->groups;
5288 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5289 cpumask_clear(groupmask);
5291 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5293 if (!(sd->flags & SD_LOAD_BALANCE)) {
5294 printk("does not load-balance\n");
5296 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5301 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5303 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5304 printk(KERN_ERR "ERROR: domain->span does not contain "
5307 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5308 printk(KERN_ERR "ERROR: domain->groups does not contain"
5312 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5316 printk(KERN_ERR "ERROR: group is NULL\n");
5321 * Even though we initialize ->power to something semi-sane,
5322 * we leave power_orig unset. This allows us to detect if
5323 * domain iteration is still funny without causing /0 traps.
5325 if (!group->sgp->power_orig) {
5326 printk(KERN_CONT "\n");
5327 printk(KERN_ERR "ERROR: domain->cpu_power not "
5332 if (!cpumask_weight(sched_group_cpus(group))) {
5333 printk(KERN_CONT "\n");
5334 printk(KERN_ERR "ERROR: empty group\n");
5338 if (!(sd->flags & SD_OVERLAP) &&
5339 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5340 printk(KERN_CONT "\n");
5341 printk(KERN_ERR "ERROR: repeated CPUs\n");
5345 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5347 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5349 printk(KERN_CONT " %s", str);
5350 if (group->sgp->power != SCHED_POWER_SCALE) {
5351 printk(KERN_CONT " (cpu_power = %d)",
5355 group = group->next;
5356 } while (group != sd->groups);
5357 printk(KERN_CONT "\n");
5359 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5360 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5363 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5364 printk(KERN_ERR "ERROR: parent span is not a superset "
5365 "of domain->span\n");
5369 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5373 if (!sched_debug_enabled)
5377 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5381 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5384 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5392 #else /* !CONFIG_SCHED_DEBUG */
5393 # define sched_domain_debug(sd, cpu) do { } while (0)
5394 static inline bool sched_debug(void)
5398 #endif /* CONFIG_SCHED_DEBUG */
5400 static int sd_degenerate(struct sched_domain *sd)
5402 if (cpumask_weight(sched_domain_span(sd)) == 1)
5405 /* Following flags need at least 2 groups */
5406 if (sd->flags & (SD_LOAD_BALANCE |
5407 SD_BALANCE_NEWIDLE |
5411 SD_SHARE_PKG_RESOURCES)) {
5412 if (sd->groups != sd->groups->next)
5416 /* Following flags don't use groups */
5417 if (sd->flags & (SD_WAKE_AFFINE))
5424 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5426 unsigned long cflags = sd->flags, pflags = parent->flags;
5428 if (sd_degenerate(parent))
5431 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5434 /* Flags needing groups don't count if only 1 group in parent */
5435 if (parent->groups == parent->groups->next) {
5436 pflags &= ~(SD_LOAD_BALANCE |
5437 SD_BALANCE_NEWIDLE |
5441 SD_SHARE_PKG_RESOURCES);
5442 if (nr_node_ids == 1)
5443 pflags &= ~SD_SERIALIZE;
5445 if (~cflags & pflags)
5451 static void free_rootdomain(struct rcu_head *rcu)
5453 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5455 cpupri_cleanup(&rd->cpupri);
5456 free_cpumask_var(rd->rto_mask);
5457 free_cpumask_var(rd->online);
5458 free_cpumask_var(rd->span);
5462 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5464 struct root_domain *old_rd = NULL;
5465 unsigned long flags;
5467 raw_spin_lock_irqsave(&rq->lock, flags);
5472 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5475 cpumask_clear_cpu(rq->cpu, old_rd->span);
5478 * If we dont want to free the old_rt yet then
5479 * set old_rd to NULL to skip the freeing later
5482 if (!atomic_dec_and_test(&old_rd->refcount))
5486 atomic_inc(&rd->refcount);
5489 cpumask_set_cpu(rq->cpu, rd->span);
5490 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5493 raw_spin_unlock_irqrestore(&rq->lock, flags);
5496 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5499 static int init_rootdomain(struct root_domain *rd)
5501 memset(rd, 0, sizeof(*rd));
5503 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5505 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5507 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5510 if (cpupri_init(&rd->cpupri) != 0)
5515 free_cpumask_var(rd->rto_mask);
5517 free_cpumask_var(rd->online);
5519 free_cpumask_var(rd->span);
5525 * By default the system creates a single root-domain with all cpus as
5526 * members (mimicking the global state we have today).
5528 struct root_domain def_root_domain;
5530 static void init_defrootdomain(void)
5532 init_rootdomain(&def_root_domain);
5534 atomic_set(&def_root_domain.refcount, 1);
5537 static struct root_domain *alloc_rootdomain(void)
5539 struct root_domain *rd;
5541 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5545 if (init_rootdomain(rd) != 0) {
5553 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5555 struct sched_group *tmp, *first;
5564 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5569 } while (sg != first);
5572 static void free_sched_domain(struct rcu_head *rcu)
5574 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5577 * If its an overlapping domain it has private groups, iterate and
5580 if (sd->flags & SD_OVERLAP) {
5581 free_sched_groups(sd->groups, 1);
5582 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5583 kfree(sd->groups->sgp);
5589 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5591 call_rcu(&sd->rcu, free_sched_domain);
5594 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5596 for (; sd; sd = sd->parent)
5597 destroy_sched_domain(sd, cpu);
5601 * Keep a special pointer to the highest sched_domain that has
5602 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5603 * allows us to avoid some pointer chasing select_idle_sibling().
5605 * Also keep a unique ID per domain (we use the first cpu number in
5606 * the cpumask of the domain), this allows us to quickly tell if
5607 * two cpus are in the same cache domain, see cpus_share_cache().
5609 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5610 DEFINE_PER_CPU(int, sd_llc_id);
5612 static void update_top_cache_domain(int cpu)
5614 struct sched_domain *sd;
5617 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5619 id = cpumask_first(sched_domain_span(sd));
5621 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5622 per_cpu(sd_llc_id, cpu) = id;
5626 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5627 * hold the hotplug lock.
5630 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5632 struct rq *rq = cpu_rq(cpu);
5633 struct sched_domain *tmp;
5635 /* Remove the sched domains which do not contribute to scheduling. */
5636 for (tmp = sd; tmp; ) {
5637 struct sched_domain *parent = tmp->parent;
5641 if (sd_parent_degenerate(tmp, parent)) {
5642 tmp->parent = parent->parent;
5644 parent->parent->child = tmp;
5645 destroy_sched_domain(parent, cpu);
5650 if (sd && sd_degenerate(sd)) {
5653 destroy_sched_domain(tmp, cpu);
5658 sched_domain_debug(sd, cpu);
5660 rq_attach_root(rq, rd);
5662 rcu_assign_pointer(rq->sd, sd);
5663 destroy_sched_domains(tmp, cpu);
5665 update_top_cache_domain(cpu);
5668 /* cpus with isolated domains */
5669 static cpumask_var_t cpu_isolated_map;
5671 /* Setup the mask of cpus configured for isolated domains */
5672 static int __init isolated_cpu_setup(char *str)
5674 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5675 cpulist_parse(str, cpu_isolated_map);
5679 __setup("isolcpus=", isolated_cpu_setup);
5681 static const struct cpumask *cpu_cpu_mask(int cpu)
5683 return cpumask_of_node(cpu_to_node(cpu));
5687 struct sched_domain **__percpu sd;
5688 struct sched_group **__percpu sg;
5689 struct sched_group_power **__percpu sgp;
5693 struct sched_domain ** __percpu sd;
5694 struct root_domain *rd;
5704 struct sched_domain_topology_level;
5706 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5707 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5709 #define SDTL_OVERLAP 0x01
5711 struct sched_domain_topology_level {
5712 sched_domain_init_f init;
5713 sched_domain_mask_f mask;
5716 struct sd_data data;
5720 * Build an iteration mask that can exclude certain CPUs from the upwards
5723 * Asymmetric node setups can result in situations where the domain tree is of
5724 * unequal depth, make sure to skip domains that already cover the entire
5727 * In that case build_sched_domains() will have terminated the iteration early
5728 * and our sibling sd spans will be empty. Domains should always include the
5729 * cpu they're built on, so check that.
5732 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5734 const struct cpumask *span = sched_domain_span(sd);
5735 struct sd_data *sdd = sd->private;
5736 struct sched_domain *sibling;
5739 for_each_cpu(i, span) {
5740 sibling = *per_cpu_ptr(sdd->sd, i);
5741 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5744 cpumask_set_cpu(i, sched_group_mask(sg));
5749 * Return the canonical balance cpu for this group, this is the first cpu
5750 * of this group that's also in the iteration mask.
5752 int group_balance_cpu(struct sched_group *sg)
5754 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5758 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5760 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5761 const struct cpumask *span = sched_domain_span(sd);
5762 struct cpumask *covered = sched_domains_tmpmask;
5763 struct sd_data *sdd = sd->private;
5764 struct sched_domain *child;
5767 cpumask_clear(covered);
5769 for_each_cpu(i, span) {
5770 struct cpumask *sg_span;
5772 if (cpumask_test_cpu(i, covered))
5775 child = *per_cpu_ptr(sdd->sd, i);
5777 /* See the comment near build_group_mask(). */
5778 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5781 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5782 GFP_KERNEL, cpu_to_node(cpu));
5787 sg_span = sched_group_cpus(sg);
5789 child = child->child;
5790 cpumask_copy(sg_span, sched_domain_span(child));
5792 cpumask_set_cpu(i, sg_span);
5794 cpumask_or(covered, covered, sg_span);
5796 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5797 if (atomic_inc_return(&sg->sgp->ref) == 1)
5798 build_group_mask(sd, sg);
5801 * Initialize sgp->power such that even if we mess up the
5802 * domains and no possible iteration will get us here, we won't
5805 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5808 * Make sure the first group of this domain contains the
5809 * canonical balance cpu. Otherwise the sched_domain iteration
5810 * breaks. See update_sg_lb_stats().
5812 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5813 group_balance_cpu(sg) == cpu)
5823 sd->groups = groups;
5828 free_sched_groups(first, 0);
5833 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5835 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5836 struct sched_domain *child = sd->child;
5839 cpu = cpumask_first(sched_domain_span(child));
5842 *sg = *per_cpu_ptr(sdd->sg, cpu);
5843 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5844 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5851 * build_sched_groups will build a circular linked list of the groups
5852 * covered by the given span, and will set each group's ->cpumask correctly,
5853 * and ->cpu_power to 0.
5855 * Assumes the sched_domain tree is fully constructed
5858 build_sched_groups(struct sched_domain *sd, int cpu)
5860 struct sched_group *first = NULL, *last = NULL;
5861 struct sd_data *sdd = sd->private;
5862 const struct cpumask *span = sched_domain_span(sd);
5863 struct cpumask *covered;
5866 get_group(cpu, sdd, &sd->groups);
5867 atomic_inc(&sd->groups->ref);
5869 if (cpu != cpumask_first(sched_domain_span(sd)))
5872 lockdep_assert_held(&sched_domains_mutex);
5873 covered = sched_domains_tmpmask;
5875 cpumask_clear(covered);
5877 for_each_cpu(i, span) {
5878 struct sched_group *sg;
5879 int group = get_group(i, sdd, &sg);
5882 if (cpumask_test_cpu(i, covered))
5885 cpumask_clear(sched_group_cpus(sg));
5887 cpumask_setall(sched_group_mask(sg));
5889 for_each_cpu(j, span) {
5890 if (get_group(j, sdd, NULL) != group)
5893 cpumask_set_cpu(j, covered);
5894 cpumask_set_cpu(j, sched_group_cpus(sg));
5909 * Initialize sched groups cpu_power.
5911 * cpu_power indicates the capacity of sched group, which is used while
5912 * distributing the load between different sched groups in a sched domain.
5913 * Typically cpu_power for all the groups in a sched domain will be same unless
5914 * there are asymmetries in the topology. If there are asymmetries, group
5915 * having more cpu_power will pickup more load compared to the group having
5918 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5920 struct sched_group *sg = sd->groups;
5922 WARN_ON(!sd || !sg);
5925 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5927 } while (sg != sd->groups);
5929 if (cpu != group_balance_cpu(sg))
5932 update_group_power(sd, cpu);
5933 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5936 int __weak arch_sd_sibling_asym_packing(void)
5938 return 0*SD_ASYM_PACKING;
5942 * Initializers for schedule domains
5943 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5946 #ifdef CONFIG_SCHED_DEBUG
5947 # define SD_INIT_NAME(sd, type) sd->name = #type
5949 # define SD_INIT_NAME(sd, type) do { } while (0)
5952 #define SD_INIT_FUNC(type) \
5953 static noinline struct sched_domain * \
5954 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5956 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5957 *sd = SD_##type##_INIT; \
5958 SD_INIT_NAME(sd, type); \
5959 sd->private = &tl->data; \
5964 #ifdef CONFIG_SCHED_SMT
5965 SD_INIT_FUNC(SIBLING)
5967 #ifdef CONFIG_SCHED_MC
5970 #ifdef CONFIG_SCHED_BOOK
5974 static int default_relax_domain_level = -1;
5975 int sched_domain_level_max;
5977 static int __init setup_relax_domain_level(char *str)
5979 if (kstrtoint(str, 0, &default_relax_domain_level))
5980 pr_warn("Unable to set relax_domain_level\n");
5984 __setup("relax_domain_level=", setup_relax_domain_level);
5986 static void set_domain_attribute(struct sched_domain *sd,
5987 struct sched_domain_attr *attr)
5991 if (!attr || attr->relax_domain_level < 0) {
5992 if (default_relax_domain_level < 0)
5995 request = default_relax_domain_level;
5997 request = attr->relax_domain_level;
5998 if (request < sd->level) {
5999 /* turn off idle balance on this domain */
6000 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6002 /* turn on idle balance on this domain */
6003 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6007 static void __sdt_free(const struct cpumask *cpu_map);
6008 static int __sdt_alloc(const struct cpumask *cpu_map);
6010 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6011 const struct cpumask *cpu_map)
6015 if (!atomic_read(&d->rd->refcount))
6016 free_rootdomain(&d->rd->rcu); /* fall through */
6018 free_percpu(d->sd); /* fall through */
6020 __sdt_free(cpu_map); /* fall through */
6026 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6027 const struct cpumask *cpu_map)
6029 memset(d, 0, sizeof(*d));
6031 if (__sdt_alloc(cpu_map))
6032 return sa_sd_storage;
6033 d->sd = alloc_percpu(struct sched_domain *);
6035 return sa_sd_storage;
6036 d->rd = alloc_rootdomain();
6039 return sa_rootdomain;
6043 * NULL the sd_data elements we've used to build the sched_domain and
6044 * sched_group structure so that the subsequent __free_domain_allocs()
6045 * will not free the data we're using.
6047 static void claim_allocations(int cpu, struct sched_domain *sd)
6049 struct sd_data *sdd = sd->private;
6051 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6052 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6054 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6055 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6057 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6058 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6061 #ifdef CONFIG_SCHED_SMT
6062 static const struct cpumask *cpu_smt_mask(int cpu)
6064 return topology_thread_cpumask(cpu);
6069 * Topology list, bottom-up.
6071 static struct sched_domain_topology_level default_topology[] = {
6072 #ifdef CONFIG_SCHED_SMT
6073 { sd_init_SIBLING, cpu_smt_mask, },
6075 #ifdef CONFIG_SCHED_MC
6076 { sd_init_MC, cpu_coregroup_mask, },
6078 #ifdef CONFIG_SCHED_BOOK
6079 { sd_init_BOOK, cpu_book_mask, },
6081 { sd_init_CPU, cpu_cpu_mask, },
6085 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6089 static int sched_domains_numa_levels;
6090 static int *sched_domains_numa_distance;
6091 static struct cpumask ***sched_domains_numa_masks;
6092 static int sched_domains_curr_level;
6094 static inline int sd_local_flags(int level)
6096 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6099 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6102 static struct sched_domain *
6103 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6105 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6106 int level = tl->numa_level;
6107 int sd_weight = cpumask_weight(
6108 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6110 *sd = (struct sched_domain){
6111 .min_interval = sd_weight,
6112 .max_interval = 2*sd_weight,
6114 .imbalance_pct = 125,
6115 .cache_nice_tries = 2,
6122 .flags = 1*SD_LOAD_BALANCE
6123 | 1*SD_BALANCE_NEWIDLE
6128 | 0*SD_SHARE_CPUPOWER
6129 | 0*SD_SHARE_PKG_RESOURCES
6131 | 0*SD_PREFER_SIBLING
6132 | sd_local_flags(level)
6134 .last_balance = jiffies,
6135 .balance_interval = sd_weight,
6137 SD_INIT_NAME(sd, NUMA);
6138 sd->private = &tl->data;
6141 * Ugly hack to pass state to sd_numa_mask()...
6143 sched_domains_curr_level = tl->numa_level;
6148 static const struct cpumask *sd_numa_mask(int cpu)
6150 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6153 static void sched_numa_warn(const char *str)
6155 static int done = false;
6163 printk(KERN_WARNING "ERROR: %s\n\n", str);
6165 for (i = 0; i < nr_node_ids; i++) {
6166 printk(KERN_WARNING " ");
6167 for (j = 0; j < nr_node_ids; j++)
6168 printk(KERN_CONT "%02d ", node_distance(i,j));
6169 printk(KERN_CONT "\n");
6171 printk(KERN_WARNING "\n");
6174 static bool find_numa_distance(int distance)
6178 if (distance == node_distance(0, 0))
6181 for (i = 0; i < sched_domains_numa_levels; i++) {
6182 if (sched_domains_numa_distance[i] == distance)
6189 static void sched_init_numa(void)
6191 int next_distance, curr_distance = node_distance(0, 0);
6192 struct sched_domain_topology_level *tl;
6196 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6197 if (!sched_domains_numa_distance)
6201 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6202 * unique distances in the node_distance() table.
6204 * Assumes node_distance(0,j) includes all distances in
6205 * node_distance(i,j) in order to avoid cubic time.
6207 next_distance = curr_distance;
6208 for (i = 0; i < nr_node_ids; i++) {
6209 for (j = 0; j < nr_node_ids; j++) {
6210 for (k = 0; k < nr_node_ids; k++) {
6211 int distance = node_distance(i, k);
6213 if (distance > curr_distance &&
6214 (distance < next_distance ||
6215 next_distance == curr_distance))
6216 next_distance = distance;
6219 * While not a strong assumption it would be nice to know
6220 * about cases where if node A is connected to B, B is not
6221 * equally connected to A.
6223 if (sched_debug() && node_distance(k, i) != distance)
6224 sched_numa_warn("Node-distance not symmetric");
6226 if (sched_debug() && i && !find_numa_distance(distance))
6227 sched_numa_warn("Node-0 not representative");
6229 if (next_distance != curr_distance) {
6230 sched_domains_numa_distance[level++] = next_distance;
6231 sched_domains_numa_levels = level;
6232 curr_distance = next_distance;
6237 * In case of sched_debug() we verify the above assumption.
6243 * 'level' contains the number of unique distances, excluding the
6244 * identity distance node_distance(i,i).
6246 * The sched_domains_nume_distance[] array includes the actual distance
6251 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6252 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6253 * the array will contain less then 'level' members. This could be
6254 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6255 * in other functions.
6257 * We reset it to 'level' at the end of this function.
6259 sched_domains_numa_levels = 0;
6261 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6262 if (!sched_domains_numa_masks)
6266 * Now for each level, construct a mask per node which contains all
6267 * cpus of nodes that are that many hops away from us.
6269 for (i = 0; i < level; i++) {
6270 sched_domains_numa_masks[i] =
6271 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6272 if (!sched_domains_numa_masks[i])
6275 for (j = 0; j < nr_node_ids; j++) {
6276 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6280 sched_domains_numa_masks[i][j] = mask;
6282 for (k = 0; k < nr_node_ids; k++) {
6283 if (node_distance(j, k) > sched_domains_numa_distance[i])
6286 cpumask_or(mask, mask, cpumask_of_node(k));
6291 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6292 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6297 * Copy the default topology bits..
6299 for (i = 0; default_topology[i].init; i++)
6300 tl[i] = default_topology[i];
6303 * .. and append 'j' levels of NUMA goodness.
6305 for (j = 0; j < level; i++, j++) {
6306 tl[i] = (struct sched_domain_topology_level){
6307 .init = sd_numa_init,
6308 .mask = sd_numa_mask,
6309 .flags = SDTL_OVERLAP,
6314 sched_domain_topology = tl;
6316 sched_domains_numa_levels = level;
6319 static void sched_domains_numa_masks_set(int cpu)
6322 int node = cpu_to_node(cpu);
6324 for (i = 0; i < sched_domains_numa_levels; i++) {
6325 for (j = 0; j < nr_node_ids; j++) {
6326 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6327 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6332 static void sched_domains_numa_masks_clear(int cpu)
6335 for (i = 0; i < sched_domains_numa_levels; i++) {
6336 for (j = 0; j < nr_node_ids; j++)
6337 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6342 * Update sched_domains_numa_masks[level][node] array when new cpus
6345 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6346 unsigned long action,
6349 int cpu = (long)hcpu;
6351 switch (action & ~CPU_TASKS_FROZEN) {
6353 sched_domains_numa_masks_set(cpu);
6357 sched_domains_numa_masks_clear(cpu);
6367 static inline void sched_init_numa(void)
6371 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6372 unsigned long action,
6377 #endif /* CONFIG_NUMA */
6379 static int __sdt_alloc(const struct cpumask *cpu_map)
6381 struct sched_domain_topology_level *tl;
6384 for (tl = sched_domain_topology; tl->init; tl++) {
6385 struct sd_data *sdd = &tl->data;
6387 sdd->sd = alloc_percpu(struct sched_domain *);
6391 sdd->sg = alloc_percpu(struct sched_group *);
6395 sdd->sgp = alloc_percpu(struct sched_group_power *);
6399 for_each_cpu(j, cpu_map) {
6400 struct sched_domain *sd;
6401 struct sched_group *sg;
6402 struct sched_group_power *sgp;
6404 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6405 GFP_KERNEL, cpu_to_node(j));
6409 *per_cpu_ptr(sdd->sd, j) = sd;
6411 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6412 GFP_KERNEL, cpu_to_node(j));
6418 *per_cpu_ptr(sdd->sg, j) = sg;
6420 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6421 GFP_KERNEL, cpu_to_node(j));
6425 *per_cpu_ptr(sdd->sgp, j) = sgp;
6432 static void __sdt_free(const struct cpumask *cpu_map)
6434 struct sched_domain_topology_level *tl;
6437 for (tl = sched_domain_topology; tl->init; tl++) {
6438 struct sd_data *sdd = &tl->data;
6440 for_each_cpu(j, cpu_map) {
6441 struct sched_domain *sd;
6444 sd = *per_cpu_ptr(sdd->sd, j);
6445 if (sd && (sd->flags & SD_OVERLAP))
6446 free_sched_groups(sd->groups, 0);
6447 kfree(*per_cpu_ptr(sdd->sd, j));
6451 kfree(*per_cpu_ptr(sdd->sg, j));
6453 kfree(*per_cpu_ptr(sdd->sgp, j));
6455 free_percpu(sdd->sd);
6457 free_percpu(sdd->sg);
6459 free_percpu(sdd->sgp);
6464 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6465 struct s_data *d, const struct cpumask *cpu_map,
6466 struct sched_domain_attr *attr, struct sched_domain *child,
6469 struct sched_domain *sd = tl->init(tl, cpu);
6473 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6475 sd->level = child->level + 1;
6476 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6480 set_domain_attribute(sd, attr);
6486 * Build sched domains for a given set of cpus and attach the sched domains
6487 * to the individual cpus
6489 static int build_sched_domains(const struct cpumask *cpu_map,
6490 struct sched_domain_attr *attr)
6492 enum s_alloc alloc_state = sa_none;
6493 struct sched_domain *sd;
6495 int i, ret = -ENOMEM;
6497 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6498 if (alloc_state != sa_rootdomain)
6501 /* Set up domains for cpus specified by the cpu_map. */
6502 for_each_cpu(i, cpu_map) {
6503 struct sched_domain_topology_level *tl;
6506 for (tl = sched_domain_topology; tl->init; tl++) {
6507 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6508 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6509 sd->flags |= SD_OVERLAP;
6510 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6517 *per_cpu_ptr(d.sd, i) = sd;
6520 /* Build the groups for the domains */
6521 for_each_cpu(i, cpu_map) {
6522 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6523 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6524 if (sd->flags & SD_OVERLAP) {
6525 if (build_overlap_sched_groups(sd, i))
6528 if (build_sched_groups(sd, i))
6534 /* Calculate CPU power for physical packages and nodes */
6535 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6536 if (!cpumask_test_cpu(i, cpu_map))
6539 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6540 claim_allocations(i, sd);
6541 init_sched_groups_power(i, sd);
6545 /* Attach the domains */
6547 for_each_cpu(i, cpu_map) {
6548 sd = *per_cpu_ptr(d.sd, i);
6549 cpu_attach_domain(sd, d.rd, i);
6555 __free_domain_allocs(&d, alloc_state, cpu_map);
6559 static cpumask_var_t *doms_cur; /* current sched domains */
6560 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6561 static struct sched_domain_attr *dattr_cur;
6562 /* attribues of custom domains in 'doms_cur' */
6565 * Special case: If a kmalloc of a doms_cur partition (array of
6566 * cpumask) fails, then fallback to a single sched domain,
6567 * as determined by the single cpumask fallback_doms.
6569 static cpumask_var_t fallback_doms;
6572 * arch_update_cpu_topology lets virtualized architectures update the
6573 * cpu core maps. It is supposed to return 1 if the topology changed
6574 * or 0 if it stayed the same.
6576 int __attribute__((weak)) arch_update_cpu_topology(void)
6581 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6584 cpumask_var_t *doms;
6586 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6589 for (i = 0; i < ndoms; i++) {
6590 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6591 free_sched_domains(doms, i);
6598 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6601 for (i = 0; i < ndoms; i++)
6602 free_cpumask_var(doms[i]);
6607 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6608 * For now this just excludes isolated cpus, but could be used to
6609 * exclude other special cases in the future.
6611 static int init_sched_domains(const struct cpumask *cpu_map)
6615 arch_update_cpu_topology();
6617 doms_cur = alloc_sched_domains(ndoms_cur);
6619 doms_cur = &fallback_doms;
6620 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6621 err = build_sched_domains(doms_cur[0], NULL);
6622 register_sched_domain_sysctl();
6628 * Detach sched domains from a group of cpus specified in cpu_map
6629 * These cpus will now be attached to the NULL domain
6631 static void detach_destroy_domains(const struct cpumask *cpu_map)
6636 for_each_cpu(i, cpu_map)
6637 cpu_attach_domain(NULL, &def_root_domain, i);
6641 /* handle null as "default" */
6642 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6643 struct sched_domain_attr *new, int idx_new)
6645 struct sched_domain_attr tmp;
6652 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6653 new ? (new + idx_new) : &tmp,
6654 sizeof(struct sched_domain_attr));
6658 * Partition sched domains as specified by the 'ndoms_new'
6659 * cpumasks in the array doms_new[] of cpumasks. This compares
6660 * doms_new[] to the current sched domain partitioning, doms_cur[].
6661 * It destroys each deleted domain and builds each new domain.
6663 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6664 * The masks don't intersect (don't overlap.) We should setup one
6665 * sched domain for each mask. CPUs not in any of the cpumasks will
6666 * not be load balanced. If the same cpumask appears both in the
6667 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6670 * The passed in 'doms_new' should be allocated using
6671 * alloc_sched_domains. This routine takes ownership of it and will
6672 * free_sched_domains it when done with it. If the caller failed the
6673 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6674 * and partition_sched_domains() will fallback to the single partition
6675 * 'fallback_doms', it also forces the domains to be rebuilt.
6677 * If doms_new == NULL it will be replaced with cpu_online_mask.
6678 * ndoms_new == 0 is a special case for destroying existing domains,
6679 * and it will not create the default domain.
6681 * Call with hotplug lock held
6683 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6684 struct sched_domain_attr *dattr_new)
6689 mutex_lock(&sched_domains_mutex);
6691 /* always unregister in case we don't destroy any domains */
6692 unregister_sched_domain_sysctl();
6694 /* Let architecture update cpu core mappings. */
6695 new_topology = arch_update_cpu_topology();
6697 n = doms_new ? ndoms_new : 0;
6699 /* Destroy deleted domains */
6700 for (i = 0; i < ndoms_cur; i++) {
6701 for (j = 0; j < n && !new_topology; j++) {
6702 if (cpumask_equal(doms_cur[i], doms_new[j])
6703 && dattrs_equal(dattr_cur, i, dattr_new, j))
6706 /* no match - a current sched domain not in new doms_new[] */
6707 detach_destroy_domains(doms_cur[i]);
6712 if (doms_new == NULL) {
6714 doms_new = &fallback_doms;
6715 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6716 WARN_ON_ONCE(dattr_new);
6719 /* Build new domains */
6720 for (i = 0; i < ndoms_new; i++) {
6721 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6722 if (cpumask_equal(doms_new[i], doms_cur[j])
6723 && dattrs_equal(dattr_new, i, dattr_cur, j))
6726 /* no match - add a new doms_new */
6727 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6732 /* Remember the new sched domains */
6733 if (doms_cur != &fallback_doms)
6734 free_sched_domains(doms_cur, ndoms_cur);
6735 kfree(dattr_cur); /* kfree(NULL) is safe */
6736 doms_cur = doms_new;
6737 dattr_cur = dattr_new;
6738 ndoms_cur = ndoms_new;
6740 register_sched_domain_sysctl();
6742 mutex_unlock(&sched_domains_mutex);
6745 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6748 * Update cpusets according to cpu_active mask. If cpusets are
6749 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6750 * around partition_sched_domains().
6752 * If we come here as part of a suspend/resume, don't touch cpusets because we
6753 * want to restore it back to its original state upon resume anyway.
6755 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6759 case CPU_ONLINE_FROZEN:
6760 case CPU_DOWN_FAILED_FROZEN:
6763 * num_cpus_frozen tracks how many CPUs are involved in suspend
6764 * resume sequence. As long as this is not the last online
6765 * operation in the resume sequence, just build a single sched
6766 * domain, ignoring cpusets.
6769 if (likely(num_cpus_frozen)) {
6770 partition_sched_domains(1, NULL, NULL);
6775 * This is the last CPU online operation. So fall through and
6776 * restore the original sched domains by considering the
6777 * cpuset configurations.
6781 case CPU_DOWN_FAILED:
6782 cpuset_update_active_cpus(true);
6790 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6794 case CPU_DOWN_PREPARE:
6795 cpuset_update_active_cpus(false);
6797 case CPU_DOWN_PREPARE_FROZEN:
6799 partition_sched_domains(1, NULL, NULL);
6807 void __init sched_init_smp(void)
6809 cpumask_var_t non_isolated_cpus;
6811 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6812 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6817 mutex_lock(&sched_domains_mutex);
6818 init_sched_domains(cpu_active_mask);
6819 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6820 if (cpumask_empty(non_isolated_cpus))
6821 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6822 mutex_unlock(&sched_domains_mutex);
6825 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6826 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6827 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6829 /* RT runtime code needs to handle some hotplug events */
6830 hotcpu_notifier(update_runtime, 0);
6834 /* Move init over to a non-isolated CPU */
6835 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6837 sched_init_granularity();
6838 free_cpumask_var(non_isolated_cpus);
6840 init_sched_rt_class();
6843 void __init sched_init_smp(void)
6845 sched_init_granularity();
6847 #endif /* CONFIG_SMP */
6849 const_debug unsigned int sysctl_timer_migration = 1;
6851 int in_sched_functions(unsigned long addr)
6853 return in_lock_functions(addr) ||
6854 (addr >= (unsigned long)__sched_text_start
6855 && addr < (unsigned long)__sched_text_end);
6858 #ifdef CONFIG_CGROUP_SCHED
6859 struct task_group root_task_group;
6860 LIST_HEAD(task_groups);
6863 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6865 void __init sched_init(void)
6868 unsigned long alloc_size = 0, ptr;
6870 #ifdef CONFIG_FAIR_GROUP_SCHED
6871 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6873 #ifdef CONFIG_RT_GROUP_SCHED
6874 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6876 #ifdef CONFIG_CPUMASK_OFFSTACK
6877 alloc_size += num_possible_cpus() * cpumask_size();
6880 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6882 #ifdef CONFIG_FAIR_GROUP_SCHED
6883 root_task_group.se = (struct sched_entity **)ptr;
6884 ptr += nr_cpu_ids * sizeof(void **);
6886 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6887 ptr += nr_cpu_ids * sizeof(void **);
6889 #endif /* CONFIG_FAIR_GROUP_SCHED */
6890 #ifdef CONFIG_RT_GROUP_SCHED
6891 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6892 ptr += nr_cpu_ids * sizeof(void **);
6894 root_task_group.rt_rq = (struct rt_rq **)ptr;
6895 ptr += nr_cpu_ids * sizeof(void **);
6897 #endif /* CONFIG_RT_GROUP_SCHED */
6898 #ifdef CONFIG_CPUMASK_OFFSTACK
6899 for_each_possible_cpu(i) {
6900 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6901 ptr += cpumask_size();
6903 #endif /* CONFIG_CPUMASK_OFFSTACK */
6907 init_defrootdomain();
6910 init_rt_bandwidth(&def_rt_bandwidth,
6911 global_rt_period(), global_rt_runtime());
6913 #ifdef CONFIG_RT_GROUP_SCHED
6914 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6915 global_rt_period(), global_rt_runtime());
6916 #endif /* CONFIG_RT_GROUP_SCHED */
6918 #ifdef CONFIG_CGROUP_SCHED
6919 list_add(&root_task_group.list, &task_groups);
6920 INIT_LIST_HEAD(&root_task_group.children);
6921 INIT_LIST_HEAD(&root_task_group.siblings);
6922 autogroup_init(&init_task);
6924 #endif /* CONFIG_CGROUP_SCHED */
6926 #ifdef CONFIG_CGROUP_CPUACCT
6927 root_cpuacct.cpustat = &kernel_cpustat;
6928 root_cpuacct.cpuusage = alloc_percpu(u64);
6929 /* Too early, not expected to fail */
6930 BUG_ON(!root_cpuacct.cpuusage);
6932 for_each_possible_cpu(i) {
6936 raw_spin_lock_init(&rq->lock);
6938 rq->calc_load_active = 0;
6939 rq->calc_load_update = jiffies + LOAD_FREQ;
6940 init_cfs_rq(&rq->cfs);
6941 init_rt_rq(&rq->rt, rq);
6942 #ifdef CONFIG_FAIR_GROUP_SCHED
6943 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6944 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6946 * How much cpu bandwidth does root_task_group get?
6948 * In case of task-groups formed thr' the cgroup filesystem, it
6949 * gets 100% of the cpu resources in the system. This overall
6950 * system cpu resource is divided among the tasks of
6951 * root_task_group and its child task-groups in a fair manner,
6952 * based on each entity's (task or task-group's) weight
6953 * (se->load.weight).
6955 * In other words, if root_task_group has 10 tasks of weight
6956 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6957 * then A0's share of the cpu resource is:
6959 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6961 * We achieve this by letting root_task_group's tasks sit
6962 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6964 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6965 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6966 #endif /* CONFIG_FAIR_GROUP_SCHED */
6968 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6969 #ifdef CONFIG_RT_GROUP_SCHED
6970 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6971 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6974 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6975 rq->cpu_load[j] = 0;
6977 rq->last_load_update_tick = jiffies;
6982 rq->cpu_power = SCHED_POWER_SCALE;
6983 rq->post_schedule = 0;
6984 rq->active_balance = 0;
6985 rq->next_balance = jiffies;
6990 rq->avg_idle = 2*sysctl_sched_migration_cost;
6992 INIT_LIST_HEAD(&rq->cfs_tasks);
6994 rq_attach_root(rq, &def_root_domain);
7000 atomic_set(&rq->nr_iowait, 0);
7003 set_load_weight(&init_task);
7005 #ifdef CONFIG_PREEMPT_NOTIFIERS
7006 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7009 #ifdef CONFIG_RT_MUTEXES
7010 plist_head_init(&init_task.pi_waiters);
7014 * The boot idle thread does lazy MMU switching as well:
7016 atomic_inc(&init_mm.mm_count);
7017 enter_lazy_tlb(&init_mm, current);
7020 * Make us the idle thread. Technically, schedule() should not be
7021 * called from this thread, however somewhere below it might be,
7022 * but because we are the idle thread, we just pick up running again
7023 * when this runqueue becomes "idle".
7025 init_idle(current, smp_processor_id());
7027 calc_load_update = jiffies + LOAD_FREQ;
7030 * During early bootup we pretend to be a normal task:
7032 current->sched_class = &fair_sched_class;
7035 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7036 /* May be allocated at isolcpus cmdline parse time */
7037 if (cpu_isolated_map == NULL)
7038 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7039 idle_thread_set_boot_cpu();
7041 init_sched_fair_class();
7043 scheduler_running = 1;
7046 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7047 static inline int preempt_count_equals(int preempt_offset)
7049 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7051 return (nested == preempt_offset);
7054 void __might_sleep(const char *file, int line, int preempt_offset)
7056 static unsigned long prev_jiffy; /* ratelimiting */
7058 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7059 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7060 system_state != SYSTEM_RUNNING || oops_in_progress)
7062 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7064 prev_jiffy = jiffies;
7067 "BUG: sleeping function called from invalid context at %s:%d\n",
7070 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7071 in_atomic(), irqs_disabled(),
7072 current->pid, current->comm);
7074 debug_show_held_locks(current);
7075 if (irqs_disabled())
7076 print_irqtrace_events(current);
7079 EXPORT_SYMBOL(__might_sleep);
7082 #ifdef CONFIG_MAGIC_SYSRQ
7083 static void normalize_task(struct rq *rq, struct task_struct *p)
7085 const struct sched_class *prev_class = p->sched_class;
7086 int old_prio = p->prio;
7091 dequeue_task(rq, p, 0);
7092 __setscheduler(rq, p, SCHED_NORMAL, 0);
7094 enqueue_task(rq, p, 0);
7095 resched_task(rq->curr);
7098 check_class_changed(rq, p, prev_class, old_prio);
7101 void normalize_rt_tasks(void)
7103 struct task_struct *g, *p;
7104 unsigned long flags;
7107 read_lock_irqsave(&tasklist_lock, flags);
7108 do_each_thread(g, p) {
7110 * Only normalize user tasks:
7115 p->se.exec_start = 0;
7116 #ifdef CONFIG_SCHEDSTATS
7117 p->se.statistics.wait_start = 0;
7118 p->se.statistics.sleep_start = 0;
7119 p->se.statistics.block_start = 0;
7124 * Renice negative nice level userspace
7127 if (TASK_NICE(p) < 0 && p->mm)
7128 set_user_nice(p, 0);
7132 raw_spin_lock(&p->pi_lock);
7133 rq = __task_rq_lock(p);
7135 normalize_task(rq, p);
7137 __task_rq_unlock(rq);
7138 raw_spin_unlock(&p->pi_lock);
7139 } while_each_thread(g, p);
7141 read_unlock_irqrestore(&tasklist_lock, flags);
7144 #endif /* CONFIG_MAGIC_SYSRQ */
7146 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7148 * These functions are only useful for the IA64 MCA handling, or kdb.
7150 * They can only be called when the whole system has been
7151 * stopped - every CPU needs to be quiescent, and no scheduling
7152 * activity can take place. Using them for anything else would
7153 * be a serious bug, and as a result, they aren't even visible
7154 * under any other configuration.
7158 * curr_task - return the current task for a given cpu.
7159 * @cpu: the processor in question.
7161 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7163 struct task_struct *curr_task(int cpu)
7165 return cpu_curr(cpu);
7168 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7172 * set_curr_task - set the current task for a given cpu.
7173 * @cpu: the processor in question.
7174 * @p: the task pointer to set.
7176 * Description: This function must only be used when non-maskable interrupts
7177 * are serviced on a separate stack. It allows the architecture to switch the
7178 * notion of the current task on a cpu in a non-blocking manner. This function
7179 * must be called with all CPU's synchronized, and interrupts disabled, the
7180 * and caller must save the original value of the current task (see
7181 * curr_task() above) and restore that value before reenabling interrupts and
7182 * re-starting the system.
7184 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7186 void set_curr_task(int cpu, struct task_struct *p)
7193 #ifdef CONFIG_CGROUP_SCHED
7194 /* task_group_lock serializes the addition/removal of task groups */
7195 static DEFINE_SPINLOCK(task_group_lock);
7197 static void free_sched_group(struct task_group *tg)
7199 free_fair_sched_group(tg);
7200 free_rt_sched_group(tg);
7205 /* allocate runqueue etc for a new task group */
7206 struct task_group *sched_create_group(struct task_group *parent)
7208 struct task_group *tg;
7210 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7212 return ERR_PTR(-ENOMEM);
7214 if (!alloc_fair_sched_group(tg, parent))
7217 if (!alloc_rt_sched_group(tg, parent))
7223 free_sched_group(tg);
7224 return ERR_PTR(-ENOMEM);
7227 void sched_online_group(struct task_group *tg, struct task_group *parent)
7229 unsigned long flags;
7231 spin_lock_irqsave(&task_group_lock, flags);
7232 list_add_rcu(&tg->list, &task_groups);
7234 WARN_ON(!parent); /* root should already exist */
7236 tg->parent = parent;
7237 INIT_LIST_HEAD(&tg->children);
7238 list_add_rcu(&tg->siblings, &parent->children);
7239 spin_unlock_irqrestore(&task_group_lock, flags);
7242 /* rcu callback to free various structures associated with a task group */
7243 static void free_sched_group_rcu(struct rcu_head *rhp)
7245 /* now it should be safe to free those cfs_rqs */
7246 free_sched_group(container_of(rhp, struct task_group, rcu));
7249 /* Destroy runqueue etc associated with a task group */
7250 void sched_destroy_group(struct task_group *tg)
7252 /* wait for possible concurrent references to cfs_rqs complete */
7253 call_rcu(&tg->rcu, free_sched_group_rcu);
7256 void sched_offline_group(struct task_group *tg)
7258 unsigned long flags;
7261 /* end participation in shares distribution */
7262 for_each_possible_cpu(i)
7263 unregister_fair_sched_group(tg, i);
7265 spin_lock_irqsave(&task_group_lock, flags);
7266 list_del_rcu(&tg->list);
7267 list_del_rcu(&tg->siblings);
7268 spin_unlock_irqrestore(&task_group_lock, flags);
7271 /* change task's runqueue when it moves between groups.
7272 * The caller of this function should have put the task in its new group
7273 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7274 * reflect its new group.
7276 void sched_move_task(struct task_struct *tsk)
7278 struct task_group *tg;
7280 unsigned long flags;
7283 rq = task_rq_lock(tsk, &flags);
7285 running = task_current(rq, tsk);
7289 dequeue_task(rq, tsk, 0);
7290 if (unlikely(running))
7291 tsk->sched_class->put_prev_task(rq, tsk);
7293 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7294 lockdep_is_held(&tsk->sighand->siglock)),
7295 struct task_group, css);
7296 tg = autogroup_task_group(tsk, tg);
7297 tsk->sched_task_group = tg;
7299 #ifdef CONFIG_FAIR_GROUP_SCHED
7300 if (tsk->sched_class->task_move_group)
7301 tsk->sched_class->task_move_group(tsk, on_rq);
7304 set_task_rq(tsk, task_cpu(tsk));
7306 if (unlikely(running))
7307 tsk->sched_class->set_curr_task(rq);
7309 enqueue_task(rq, tsk, 0);
7311 task_rq_unlock(rq, tsk, &flags);
7313 #endif /* CONFIG_CGROUP_SCHED */
7315 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7316 static unsigned long to_ratio(u64 period, u64 runtime)
7318 if (runtime == RUNTIME_INF)
7321 return div64_u64(runtime << 20, period);
7325 #ifdef CONFIG_RT_GROUP_SCHED
7327 * Ensure that the real time constraints are schedulable.
7329 static DEFINE_MUTEX(rt_constraints_mutex);
7331 /* Must be called with tasklist_lock held */
7332 static inline int tg_has_rt_tasks(struct task_group *tg)
7334 struct task_struct *g, *p;
7336 do_each_thread(g, p) {
7337 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7339 } while_each_thread(g, p);
7344 struct rt_schedulable_data {
7345 struct task_group *tg;
7350 static int tg_rt_schedulable(struct task_group *tg, void *data)
7352 struct rt_schedulable_data *d = data;
7353 struct task_group *child;
7354 unsigned long total, sum = 0;
7355 u64 period, runtime;
7357 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7358 runtime = tg->rt_bandwidth.rt_runtime;
7361 period = d->rt_period;
7362 runtime = d->rt_runtime;
7366 * Cannot have more runtime than the period.
7368 if (runtime > period && runtime != RUNTIME_INF)
7372 * Ensure we don't starve existing RT tasks.
7374 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7377 total = to_ratio(period, runtime);
7380 * Nobody can have more than the global setting allows.
7382 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7386 * The sum of our children's runtime should not exceed our own.
7388 list_for_each_entry_rcu(child, &tg->children, siblings) {
7389 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7390 runtime = child->rt_bandwidth.rt_runtime;
7392 if (child == d->tg) {
7393 period = d->rt_period;
7394 runtime = d->rt_runtime;
7397 sum += to_ratio(period, runtime);
7406 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7410 struct rt_schedulable_data data = {
7412 .rt_period = period,
7413 .rt_runtime = runtime,
7417 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7423 static int tg_set_rt_bandwidth(struct task_group *tg,
7424 u64 rt_period, u64 rt_runtime)
7428 mutex_lock(&rt_constraints_mutex);
7429 read_lock(&tasklist_lock);
7430 err = __rt_schedulable(tg, rt_period, rt_runtime);
7434 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7435 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7436 tg->rt_bandwidth.rt_runtime = rt_runtime;
7438 for_each_possible_cpu(i) {
7439 struct rt_rq *rt_rq = tg->rt_rq[i];
7441 raw_spin_lock(&rt_rq->rt_runtime_lock);
7442 rt_rq->rt_runtime = rt_runtime;
7443 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7445 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7447 read_unlock(&tasklist_lock);
7448 mutex_unlock(&rt_constraints_mutex);
7453 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7455 u64 rt_runtime, rt_period;
7457 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7458 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7459 if (rt_runtime_us < 0)
7460 rt_runtime = RUNTIME_INF;
7462 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7465 long sched_group_rt_runtime(struct task_group *tg)
7469 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7472 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7473 do_div(rt_runtime_us, NSEC_PER_USEC);
7474 return rt_runtime_us;
7477 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7479 u64 rt_runtime, rt_period;
7481 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7482 rt_runtime = tg->rt_bandwidth.rt_runtime;
7487 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7490 long sched_group_rt_period(struct task_group *tg)
7494 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7495 do_div(rt_period_us, NSEC_PER_USEC);
7496 return rt_period_us;
7499 static int sched_rt_global_constraints(void)
7501 u64 runtime, period;
7504 if (sysctl_sched_rt_period <= 0)
7507 runtime = global_rt_runtime();
7508 period = global_rt_period();
7511 * Sanity check on the sysctl variables.
7513 if (runtime > period && runtime != RUNTIME_INF)
7516 mutex_lock(&rt_constraints_mutex);
7517 read_lock(&tasklist_lock);
7518 ret = __rt_schedulable(NULL, 0, 0);
7519 read_unlock(&tasklist_lock);
7520 mutex_unlock(&rt_constraints_mutex);
7525 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7527 /* Don't accept realtime tasks when there is no way for them to run */
7528 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7534 #else /* !CONFIG_RT_GROUP_SCHED */
7535 static int sched_rt_global_constraints(void)
7537 unsigned long flags;
7540 if (sysctl_sched_rt_period <= 0)
7544 * There's always some RT tasks in the root group
7545 * -- migration, kstopmachine etc..
7547 if (sysctl_sched_rt_runtime == 0)
7550 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7551 for_each_possible_cpu(i) {
7552 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7554 raw_spin_lock(&rt_rq->rt_runtime_lock);
7555 rt_rq->rt_runtime = global_rt_runtime();
7556 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7558 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7562 #endif /* CONFIG_RT_GROUP_SCHED */
7564 int sched_rr_handler(struct ctl_table *table, int write,
7565 void __user *buffer, size_t *lenp,
7569 static DEFINE_MUTEX(mutex);
7572 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7573 /* make sure that internally we keep jiffies */
7574 /* also, writing zero resets timeslice to default */
7575 if (!ret && write) {
7576 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7577 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7579 mutex_unlock(&mutex);
7583 int sched_rt_handler(struct ctl_table *table, int write,
7584 void __user *buffer, size_t *lenp,
7588 int old_period, old_runtime;
7589 static DEFINE_MUTEX(mutex);
7592 old_period = sysctl_sched_rt_period;
7593 old_runtime = sysctl_sched_rt_runtime;
7595 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7597 if (!ret && write) {
7598 ret = sched_rt_global_constraints();
7600 sysctl_sched_rt_period = old_period;
7601 sysctl_sched_rt_runtime = old_runtime;
7603 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7604 def_rt_bandwidth.rt_period =
7605 ns_to_ktime(global_rt_period());
7608 mutex_unlock(&mutex);
7613 #ifdef CONFIG_CGROUP_SCHED
7615 /* return corresponding task_group object of a cgroup */
7616 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7618 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7619 struct task_group, css);
7622 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7624 struct task_group *tg, *parent;
7626 if (!cgrp->parent) {
7627 /* This is early initialization for the top cgroup */
7628 return &root_task_group.css;
7631 parent = cgroup_tg(cgrp->parent);
7632 tg = sched_create_group(parent);
7634 return ERR_PTR(-ENOMEM);
7639 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7641 struct task_group *tg = cgroup_tg(cgrp);
7642 struct task_group *parent;
7647 parent = cgroup_tg(cgrp->parent);
7648 sched_online_group(tg, parent);
7652 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7654 struct task_group *tg = cgroup_tg(cgrp);
7656 sched_destroy_group(tg);
7659 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7661 struct task_group *tg = cgroup_tg(cgrp);
7663 sched_offline_group(tg);
7666 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7667 struct cgroup_taskset *tset)
7669 struct task_struct *task;
7671 cgroup_taskset_for_each(task, cgrp, tset) {
7672 #ifdef CONFIG_RT_GROUP_SCHED
7673 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7676 /* We don't support RT-tasks being in separate groups */
7677 if (task->sched_class != &fair_sched_class)
7684 static void cpu_cgroup_attach(struct cgroup *cgrp,
7685 struct cgroup_taskset *tset)
7687 struct task_struct *task;
7689 cgroup_taskset_for_each(task, cgrp, tset)
7690 sched_move_task(task);
7694 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7695 struct task_struct *task)
7698 * cgroup_exit() is called in the copy_process() failure path.
7699 * Ignore this case since the task hasn't ran yet, this avoids
7700 * trying to poke a half freed task state from generic code.
7702 if (!(task->flags & PF_EXITING))
7705 sched_move_task(task);
7708 #ifdef CONFIG_FAIR_GROUP_SCHED
7709 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7712 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7715 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7717 struct task_group *tg = cgroup_tg(cgrp);
7719 return (u64) scale_load_down(tg->shares);
7722 #ifdef CONFIG_CFS_BANDWIDTH
7723 static DEFINE_MUTEX(cfs_constraints_mutex);
7725 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7726 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7728 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7730 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7732 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7733 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7735 if (tg == &root_task_group)
7739 * Ensure we have at some amount of bandwidth every period. This is
7740 * to prevent reaching a state of large arrears when throttled via
7741 * entity_tick() resulting in prolonged exit starvation.
7743 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7747 * Likewise, bound things on the otherside by preventing insane quota
7748 * periods. This also allows us to normalize in computing quota
7751 if (period > max_cfs_quota_period)
7754 mutex_lock(&cfs_constraints_mutex);
7755 ret = __cfs_schedulable(tg, period, quota);
7759 runtime_enabled = quota != RUNTIME_INF;
7760 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7761 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7762 raw_spin_lock_irq(&cfs_b->lock);
7763 cfs_b->period = ns_to_ktime(period);
7764 cfs_b->quota = quota;
7766 __refill_cfs_bandwidth_runtime(cfs_b);
7767 /* restart the period timer (if active) to handle new period expiry */
7768 if (runtime_enabled && cfs_b->timer_active) {
7769 /* force a reprogram */
7770 cfs_b->timer_active = 0;
7771 __start_cfs_bandwidth(cfs_b);
7773 raw_spin_unlock_irq(&cfs_b->lock);
7775 for_each_possible_cpu(i) {
7776 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7777 struct rq *rq = cfs_rq->rq;
7779 raw_spin_lock_irq(&rq->lock);
7780 cfs_rq->runtime_enabled = runtime_enabled;
7781 cfs_rq->runtime_remaining = 0;
7783 if (cfs_rq->throttled)
7784 unthrottle_cfs_rq(cfs_rq);
7785 raw_spin_unlock_irq(&rq->lock);
7788 mutex_unlock(&cfs_constraints_mutex);
7793 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7797 period = ktime_to_ns(tg->cfs_bandwidth.period);
7798 if (cfs_quota_us < 0)
7799 quota = RUNTIME_INF;
7801 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7803 return tg_set_cfs_bandwidth(tg, period, quota);
7806 long tg_get_cfs_quota(struct task_group *tg)
7810 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7813 quota_us = tg->cfs_bandwidth.quota;
7814 do_div(quota_us, NSEC_PER_USEC);
7819 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7823 period = (u64)cfs_period_us * NSEC_PER_USEC;
7824 quota = tg->cfs_bandwidth.quota;
7826 return tg_set_cfs_bandwidth(tg, period, quota);
7829 long tg_get_cfs_period(struct task_group *tg)
7833 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7834 do_div(cfs_period_us, NSEC_PER_USEC);
7836 return cfs_period_us;
7839 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7841 return tg_get_cfs_quota(cgroup_tg(cgrp));
7844 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7847 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7850 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7852 return tg_get_cfs_period(cgroup_tg(cgrp));
7855 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7858 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7861 struct cfs_schedulable_data {
7862 struct task_group *tg;
7867 * normalize group quota/period to be quota/max_period
7868 * note: units are usecs
7870 static u64 normalize_cfs_quota(struct task_group *tg,
7871 struct cfs_schedulable_data *d)
7879 period = tg_get_cfs_period(tg);
7880 quota = tg_get_cfs_quota(tg);
7883 /* note: these should typically be equivalent */
7884 if (quota == RUNTIME_INF || quota == -1)
7887 return to_ratio(period, quota);
7890 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7892 struct cfs_schedulable_data *d = data;
7893 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7894 s64 quota = 0, parent_quota = -1;
7897 quota = RUNTIME_INF;
7899 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7901 quota = normalize_cfs_quota(tg, d);
7902 parent_quota = parent_b->hierarchal_quota;
7905 * ensure max(child_quota) <= parent_quota, inherit when no
7908 if (quota == RUNTIME_INF)
7909 quota = parent_quota;
7910 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7913 cfs_b->hierarchal_quota = quota;
7918 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7921 struct cfs_schedulable_data data = {
7927 if (quota != RUNTIME_INF) {
7928 do_div(data.period, NSEC_PER_USEC);
7929 do_div(data.quota, NSEC_PER_USEC);
7933 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7939 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7940 struct cgroup_map_cb *cb)
7942 struct task_group *tg = cgroup_tg(cgrp);
7943 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7945 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7946 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7947 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7951 #endif /* CONFIG_CFS_BANDWIDTH */
7952 #endif /* CONFIG_FAIR_GROUP_SCHED */
7954 #ifdef CONFIG_RT_GROUP_SCHED
7955 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7958 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7961 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7963 return sched_group_rt_runtime(cgroup_tg(cgrp));
7966 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7969 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7972 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7974 return sched_group_rt_period(cgroup_tg(cgrp));
7976 #endif /* CONFIG_RT_GROUP_SCHED */
7978 static struct cftype cpu_files[] = {
7979 #ifdef CONFIG_FAIR_GROUP_SCHED
7982 .read_u64 = cpu_shares_read_u64,
7983 .write_u64 = cpu_shares_write_u64,
7986 #ifdef CONFIG_CFS_BANDWIDTH
7988 .name = "cfs_quota_us",
7989 .read_s64 = cpu_cfs_quota_read_s64,
7990 .write_s64 = cpu_cfs_quota_write_s64,
7993 .name = "cfs_period_us",
7994 .read_u64 = cpu_cfs_period_read_u64,
7995 .write_u64 = cpu_cfs_period_write_u64,
7999 .read_map = cpu_stats_show,
8002 #ifdef CONFIG_RT_GROUP_SCHED
8004 .name = "rt_runtime_us",
8005 .read_s64 = cpu_rt_runtime_read,
8006 .write_s64 = cpu_rt_runtime_write,
8009 .name = "rt_period_us",
8010 .read_u64 = cpu_rt_period_read_uint,
8011 .write_u64 = cpu_rt_period_write_uint,
8017 struct cgroup_subsys cpu_cgroup_subsys = {
8019 .css_alloc = cpu_cgroup_css_alloc,
8020 .css_free = cpu_cgroup_css_free,
8021 .css_online = cpu_cgroup_css_online,
8022 .css_offline = cpu_cgroup_css_offline,
8023 .can_attach = cpu_cgroup_can_attach,
8024 .attach = cpu_cgroup_attach,
8025 .exit = cpu_cgroup_exit,
8026 .subsys_id = cpu_cgroup_subsys_id,
8027 .base_cftypes = cpu_files,
8031 #endif /* CONFIG_CGROUP_SCHED */
8033 #ifdef CONFIG_CGROUP_CPUACCT
8036 * CPU accounting code for task groups.
8038 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8039 * (balbir@in.ibm.com).
8042 struct cpuacct root_cpuacct;
8044 /* create a new cpu accounting group */
8045 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
8050 return &root_cpuacct.css;
8052 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8056 ca->cpuusage = alloc_percpu(u64);
8060 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8062 goto out_free_cpuusage;
8067 free_percpu(ca->cpuusage);
8071 return ERR_PTR(-ENOMEM);
8074 /* destroy an existing cpu accounting group */
8075 static void cpuacct_css_free(struct cgroup *cgrp)
8077 struct cpuacct *ca = cgroup_ca(cgrp);
8079 free_percpu(ca->cpustat);
8080 free_percpu(ca->cpuusage);
8084 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8086 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8089 #ifndef CONFIG_64BIT
8091 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8093 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8095 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8103 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8105 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8107 #ifndef CONFIG_64BIT
8109 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8111 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8113 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8119 /* return total cpu usage (in nanoseconds) of a group */
8120 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8122 struct cpuacct *ca = cgroup_ca(cgrp);
8123 u64 totalcpuusage = 0;
8126 for_each_present_cpu(i)
8127 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8129 return totalcpuusage;
8132 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8135 struct cpuacct *ca = cgroup_ca(cgrp);
8144 for_each_present_cpu(i)
8145 cpuacct_cpuusage_write(ca, i, 0);
8151 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8154 struct cpuacct *ca = cgroup_ca(cgroup);
8158 for_each_present_cpu(i) {
8159 percpu = cpuacct_cpuusage_read(ca, i);
8160 seq_printf(m, "%llu ", (unsigned long long) percpu);
8162 seq_printf(m, "\n");
8166 static const char *cpuacct_stat_desc[] = {
8167 [CPUACCT_STAT_USER] = "user",
8168 [CPUACCT_STAT_SYSTEM] = "system",
8171 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8172 struct cgroup_map_cb *cb)
8174 struct cpuacct *ca = cgroup_ca(cgrp);
8178 for_each_online_cpu(cpu) {
8179 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8180 val += kcpustat->cpustat[CPUTIME_USER];
8181 val += kcpustat->cpustat[CPUTIME_NICE];
8183 val = cputime64_to_clock_t(val);
8184 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8187 for_each_online_cpu(cpu) {
8188 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8189 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8190 val += kcpustat->cpustat[CPUTIME_IRQ];
8191 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8194 val = cputime64_to_clock_t(val);
8195 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8200 static struct cftype files[] = {
8203 .read_u64 = cpuusage_read,
8204 .write_u64 = cpuusage_write,
8207 .name = "usage_percpu",
8208 .read_seq_string = cpuacct_percpu_seq_read,
8212 .read_map = cpuacct_stats_show,
8218 * charge this task's execution time to its accounting group.
8220 * called with rq->lock held.
8222 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8227 if (unlikely(!cpuacct_subsys.active))
8230 cpu = task_cpu(tsk);
8236 for (; ca; ca = parent_ca(ca)) {
8237 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8238 *cpuusage += cputime;
8244 struct cgroup_subsys cpuacct_subsys = {
8246 .css_alloc = cpuacct_css_alloc,
8247 .css_free = cpuacct_css_free,
8248 .subsys_id = cpuacct_subsys_id,
8249 .base_cftypes = files,
8251 #endif /* CONFIG_CGROUP_CPUACCT */
8253 void dump_cpu_task(int cpu)
8255 pr_info("Task dump for CPU %d:\n", cpu);
8256 sched_show_task(cpu_curr(cpu));