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 check_preempt_curr(rq, p, wake_flags);
1287 trace_sched_wakeup(p, true);
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 if (WARN_ON_ONCE(rq != this_rq()) ||
1497 WARN_ON_ONCE(p == current))
1500 lockdep_assert_held(&rq->lock);
1502 if (!raw_spin_trylock(&p->pi_lock)) {
1503 raw_spin_unlock(&rq->lock);
1504 raw_spin_lock(&p->pi_lock);
1505 raw_spin_lock(&rq->lock);
1508 if (!(p->state & TASK_NORMAL))
1512 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1514 ttwu_do_wakeup(rq, p, 0);
1515 ttwu_stat(p, smp_processor_id(), 0);
1517 raw_spin_unlock(&p->pi_lock);
1521 * wake_up_process - Wake up a specific process
1522 * @p: The process to be woken up.
1524 * Attempt to wake up the nominated process and move it to the set of runnable
1525 * processes. Returns 1 if the process was woken up, 0 if it was already
1528 * It may be assumed that this function implies a write memory barrier before
1529 * changing the task state if and only if any tasks are woken up.
1531 int wake_up_process(struct task_struct *p)
1533 WARN_ON(task_is_stopped_or_traced(p));
1534 return try_to_wake_up(p, TASK_NORMAL, 0);
1536 EXPORT_SYMBOL(wake_up_process);
1538 int wake_up_state(struct task_struct *p, unsigned int state)
1540 return try_to_wake_up(p, state, 0);
1544 * Perform scheduler related setup for a newly forked process p.
1545 * p is forked by current.
1547 * __sched_fork() is basic setup used by init_idle() too:
1549 static void __sched_fork(struct task_struct *p)
1554 p->se.exec_start = 0;
1555 p->se.sum_exec_runtime = 0;
1556 p->se.prev_sum_exec_runtime = 0;
1557 p->se.nr_migrations = 0;
1559 INIT_LIST_HEAD(&p->se.group_node);
1562 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1563 * removed when useful for applications beyond shares distribution (e.g.
1566 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1567 p->se.avg.runnable_avg_period = 0;
1568 p->se.avg.runnable_avg_sum = 0;
1570 #ifdef CONFIG_SCHEDSTATS
1571 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1574 INIT_LIST_HEAD(&p->rt.run_list);
1576 #ifdef CONFIG_PREEMPT_NOTIFIERS
1577 INIT_HLIST_HEAD(&p->preempt_notifiers);
1580 #ifdef CONFIG_NUMA_BALANCING
1581 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1582 p->mm->numa_next_scan = jiffies;
1583 p->mm->numa_next_reset = jiffies;
1584 p->mm->numa_scan_seq = 0;
1587 p->node_stamp = 0ULL;
1588 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1589 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1590 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1591 p->numa_work.next = &p->numa_work;
1592 #endif /* CONFIG_NUMA_BALANCING */
1595 #ifdef CONFIG_NUMA_BALANCING
1596 #ifdef CONFIG_SCHED_DEBUG
1597 void set_numabalancing_state(bool enabled)
1600 sched_feat_set("NUMA");
1602 sched_feat_set("NO_NUMA");
1605 __read_mostly bool numabalancing_enabled;
1607 void set_numabalancing_state(bool enabled)
1609 numabalancing_enabled = enabled;
1611 #endif /* CONFIG_SCHED_DEBUG */
1612 #endif /* CONFIG_NUMA_BALANCING */
1615 * fork()/clone()-time setup:
1617 void sched_fork(struct task_struct *p)
1619 unsigned long flags;
1620 int cpu = get_cpu();
1624 * We mark the process as running here. This guarantees that
1625 * nobody will actually run it, and a signal or other external
1626 * event cannot wake it up and insert it on the runqueue either.
1628 p->state = TASK_RUNNING;
1631 * Make sure we do not leak PI boosting priority to the child.
1633 p->prio = current->normal_prio;
1636 * Revert to default priority/policy on fork if requested.
1638 if (unlikely(p->sched_reset_on_fork)) {
1639 if (task_has_rt_policy(p)) {
1640 p->policy = SCHED_NORMAL;
1641 p->static_prio = NICE_TO_PRIO(0);
1643 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1644 p->static_prio = NICE_TO_PRIO(0);
1646 p->prio = p->normal_prio = __normal_prio(p);
1650 * We don't need the reset flag anymore after the fork. It has
1651 * fulfilled its duty:
1653 p->sched_reset_on_fork = 0;
1656 if (!rt_prio(p->prio))
1657 p->sched_class = &fair_sched_class;
1659 if (p->sched_class->task_fork)
1660 p->sched_class->task_fork(p);
1663 * The child is not yet in the pid-hash so no cgroup attach races,
1664 * and the cgroup is pinned to this child due to cgroup_fork()
1665 * is ran before sched_fork().
1667 * Silence PROVE_RCU.
1669 raw_spin_lock_irqsave(&p->pi_lock, flags);
1670 set_task_cpu(p, cpu);
1671 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1673 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1674 if (likely(sched_info_on()))
1675 memset(&p->sched_info, 0, sizeof(p->sched_info));
1677 #if defined(CONFIG_SMP)
1680 #ifdef CONFIG_PREEMPT_COUNT
1681 /* Want to start with kernel preemption disabled. */
1682 task_thread_info(p)->preempt_count = 1;
1685 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1692 * wake_up_new_task - wake up a newly created task for the first time.
1694 * This function will do some initial scheduler statistics housekeeping
1695 * that must be done for every newly created context, then puts the task
1696 * on the runqueue and wakes it.
1698 void wake_up_new_task(struct task_struct *p)
1700 unsigned long flags;
1703 raw_spin_lock_irqsave(&p->pi_lock, flags);
1706 * Fork balancing, do it here and not earlier because:
1707 * - cpus_allowed can change in the fork path
1708 * - any previously selected cpu might disappear through hotplug
1710 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1713 rq = __task_rq_lock(p);
1714 activate_task(rq, p, 0);
1716 trace_sched_wakeup_new(p, true);
1717 check_preempt_curr(rq, p, WF_FORK);
1719 if (p->sched_class->task_woken)
1720 p->sched_class->task_woken(rq, p);
1722 task_rq_unlock(rq, p, &flags);
1725 #ifdef CONFIG_PREEMPT_NOTIFIERS
1728 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1729 * @notifier: notifier struct to register
1731 void preempt_notifier_register(struct preempt_notifier *notifier)
1733 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1735 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1738 * preempt_notifier_unregister - no longer interested in preemption notifications
1739 * @notifier: notifier struct to unregister
1741 * This is safe to call from within a preemption notifier.
1743 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1745 hlist_del(¬ifier->link);
1747 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1749 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1751 struct preempt_notifier *notifier;
1753 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1754 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1758 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1759 struct task_struct *next)
1761 struct preempt_notifier *notifier;
1763 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1764 notifier->ops->sched_out(notifier, next);
1767 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1769 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1774 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1775 struct task_struct *next)
1779 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1782 * prepare_task_switch - prepare to switch tasks
1783 * @rq: the runqueue preparing to switch
1784 * @prev: the current task that is being switched out
1785 * @next: the task we are going to switch to.
1787 * This is called with the rq lock held and interrupts off. It must
1788 * be paired with a subsequent finish_task_switch after the context
1791 * prepare_task_switch sets up locking and calls architecture specific
1795 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1796 struct task_struct *next)
1798 trace_sched_switch(prev, next);
1799 sched_info_switch(prev, next);
1800 perf_event_task_sched_out(prev, next);
1801 fire_sched_out_preempt_notifiers(prev, next);
1802 prepare_lock_switch(rq, next);
1803 prepare_arch_switch(next);
1807 * finish_task_switch - clean up after a task-switch
1808 * @rq: runqueue associated with task-switch
1809 * @prev: the thread we just switched away from.
1811 * finish_task_switch must be called after the context switch, paired
1812 * with a prepare_task_switch call before the context switch.
1813 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1814 * and do any other architecture-specific cleanup actions.
1816 * Note that we may have delayed dropping an mm in context_switch(). If
1817 * so, we finish that here outside of the runqueue lock. (Doing it
1818 * with the lock held can cause deadlocks; see schedule() for
1821 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1822 __releases(rq->lock)
1824 struct mm_struct *mm = rq->prev_mm;
1830 * A task struct has one reference for the use as "current".
1831 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1832 * schedule one last time. The schedule call will never return, and
1833 * the scheduled task must drop that reference.
1834 * The test for TASK_DEAD must occur while the runqueue locks are
1835 * still held, otherwise prev could be scheduled on another cpu, die
1836 * there before we look at prev->state, and then the reference would
1838 * Manfred Spraul <manfred@colorfullife.com>
1840 prev_state = prev->state;
1841 vtime_task_switch(prev);
1842 finish_arch_switch(prev);
1843 perf_event_task_sched_in(prev, current);
1844 finish_lock_switch(rq, prev);
1845 finish_arch_post_lock_switch();
1847 fire_sched_in_preempt_notifiers(current);
1850 if (unlikely(prev_state == TASK_DEAD)) {
1852 * Remove function-return probe instances associated with this
1853 * task and put them back on the free list.
1855 kprobe_flush_task(prev);
1856 put_task_struct(prev);
1862 /* assumes rq->lock is held */
1863 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1865 if (prev->sched_class->pre_schedule)
1866 prev->sched_class->pre_schedule(rq, prev);
1869 /* rq->lock is NOT held, but preemption is disabled */
1870 static inline void post_schedule(struct rq *rq)
1872 if (rq->post_schedule) {
1873 unsigned long flags;
1875 raw_spin_lock_irqsave(&rq->lock, flags);
1876 if (rq->curr->sched_class->post_schedule)
1877 rq->curr->sched_class->post_schedule(rq);
1878 raw_spin_unlock_irqrestore(&rq->lock, flags);
1880 rq->post_schedule = 0;
1886 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1890 static inline void post_schedule(struct rq *rq)
1897 * schedule_tail - first thing a freshly forked thread must call.
1898 * @prev: the thread we just switched away from.
1900 asmlinkage void schedule_tail(struct task_struct *prev)
1901 __releases(rq->lock)
1903 struct rq *rq = this_rq();
1905 finish_task_switch(rq, prev);
1908 * FIXME: do we need to worry about rq being invalidated by the
1913 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1914 /* In this case, finish_task_switch does not reenable preemption */
1917 if (current->set_child_tid)
1918 put_user(task_pid_vnr(current), current->set_child_tid);
1922 * context_switch - switch to the new MM and the new
1923 * thread's register state.
1926 context_switch(struct rq *rq, struct task_struct *prev,
1927 struct task_struct *next)
1929 struct mm_struct *mm, *oldmm;
1931 prepare_task_switch(rq, prev, next);
1934 oldmm = prev->active_mm;
1936 * For paravirt, this is coupled with an exit in switch_to to
1937 * combine the page table reload and the switch backend into
1940 arch_start_context_switch(prev);
1943 next->active_mm = oldmm;
1944 atomic_inc(&oldmm->mm_count);
1945 enter_lazy_tlb(oldmm, next);
1947 switch_mm(oldmm, mm, next);
1950 prev->active_mm = NULL;
1951 rq->prev_mm = oldmm;
1954 * Since the runqueue lock will be released by the next
1955 * task (which is an invalid locking op but in the case
1956 * of the scheduler it's an obvious special-case), so we
1957 * do an early lockdep release here:
1959 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1960 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1963 context_tracking_task_switch(prev, next);
1964 /* Here we just switch the register state and the stack. */
1965 switch_to(prev, next, prev);
1969 * this_rq must be evaluated again because prev may have moved
1970 * CPUs since it called schedule(), thus the 'rq' on its stack
1971 * frame will be invalid.
1973 finish_task_switch(this_rq(), prev);
1977 * nr_running and nr_context_switches:
1979 * externally visible scheduler statistics: current number of runnable
1980 * threads, total number of context switches performed since bootup.
1982 unsigned long nr_running(void)
1984 unsigned long i, sum = 0;
1986 for_each_online_cpu(i)
1987 sum += cpu_rq(i)->nr_running;
1992 unsigned long long nr_context_switches(void)
1995 unsigned long long sum = 0;
1997 for_each_possible_cpu(i)
1998 sum += cpu_rq(i)->nr_switches;
2003 unsigned long nr_iowait(void)
2005 unsigned long i, sum = 0;
2007 for_each_possible_cpu(i)
2008 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2013 unsigned long nr_iowait_cpu(int cpu)
2015 struct rq *this = cpu_rq(cpu);
2016 return atomic_read(&this->nr_iowait);
2019 unsigned long this_cpu_load(void)
2021 struct rq *this = this_rq();
2022 return this->cpu_load[0];
2027 * Global load-average calculations
2029 * We take a distributed and async approach to calculating the global load-avg
2030 * in order to minimize overhead.
2032 * The global load average is an exponentially decaying average of nr_running +
2033 * nr_uninterruptible.
2035 * Once every LOAD_FREQ:
2038 * for_each_possible_cpu(cpu)
2039 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2041 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2043 * Due to a number of reasons the above turns in the mess below:
2045 * - for_each_possible_cpu() is prohibitively expensive on machines with
2046 * serious number of cpus, therefore we need to take a distributed approach
2047 * to calculating nr_active.
2049 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2050 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2052 * So assuming nr_active := 0 when we start out -- true per definition, we
2053 * can simply take per-cpu deltas and fold those into a global accumulate
2054 * to obtain the same result. See calc_load_fold_active().
2056 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2057 * across the machine, we assume 10 ticks is sufficient time for every
2058 * cpu to have completed this task.
2060 * This places an upper-bound on the IRQ-off latency of the machine. Then
2061 * again, being late doesn't loose the delta, just wrecks the sample.
2063 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2064 * this would add another cross-cpu cacheline miss and atomic operation
2065 * to the wakeup path. Instead we increment on whatever cpu the task ran
2066 * when it went into uninterruptible state and decrement on whatever cpu
2067 * did the wakeup. This means that only the sum of nr_uninterruptible over
2068 * all cpus yields the correct result.
2070 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2073 /* Variables and functions for calc_load */
2074 static atomic_long_t calc_load_tasks;
2075 static unsigned long calc_load_update;
2076 unsigned long avenrun[3];
2077 EXPORT_SYMBOL(avenrun); /* should be removed */
2080 * get_avenrun - get the load average array
2081 * @loads: pointer to dest load array
2082 * @offset: offset to add
2083 * @shift: shift count to shift the result left
2085 * These values are estimates at best, so no need for locking.
2087 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2089 loads[0] = (avenrun[0] + offset) << shift;
2090 loads[1] = (avenrun[1] + offset) << shift;
2091 loads[2] = (avenrun[2] + offset) << shift;
2094 static long calc_load_fold_active(struct rq *this_rq)
2096 long nr_active, delta = 0;
2098 nr_active = this_rq->nr_running;
2099 nr_active += (long) this_rq->nr_uninterruptible;
2101 if (nr_active != this_rq->calc_load_active) {
2102 delta = nr_active - this_rq->calc_load_active;
2103 this_rq->calc_load_active = nr_active;
2110 * a1 = a0 * e + a * (1 - e)
2112 static unsigned long
2113 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2116 load += active * (FIXED_1 - exp);
2117 load += 1UL << (FSHIFT - 1);
2118 return load >> FSHIFT;
2123 * Handle NO_HZ for the global load-average.
2125 * Since the above described distributed algorithm to compute the global
2126 * load-average relies on per-cpu sampling from the tick, it is affected by
2129 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2130 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2131 * when we read the global state.
2133 * Obviously reality has to ruin such a delightfully simple scheme:
2135 * - When we go NO_HZ idle during the window, we can negate our sample
2136 * contribution, causing under-accounting.
2138 * We avoid this by keeping two idle-delta counters and flipping them
2139 * when the window starts, thus separating old and new NO_HZ load.
2141 * The only trick is the slight shift in index flip for read vs write.
2145 * |-|-----------|-|-----------|-|-----------|-|
2146 * r:0 0 1 1 0 0 1 1 0
2147 * w:0 1 1 0 0 1 1 0 0
2149 * This ensures we'll fold the old idle contribution in this window while
2150 * accumlating the new one.
2152 * - When we wake up from NO_HZ idle during the window, we push up our
2153 * contribution, since we effectively move our sample point to a known
2156 * This is solved by pushing the window forward, and thus skipping the
2157 * sample, for this cpu (effectively using the idle-delta for this cpu which
2158 * was in effect at the time the window opened). This also solves the issue
2159 * of having to deal with a cpu having been in NOHZ idle for multiple
2160 * LOAD_FREQ intervals.
2162 * When making the ILB scale, we should try to pull this in as well.
2164 static atomic_long_t calc_load_idle[2];
2165 static int calc_load_idx;
2167 static inline int calc_load_write_idx(void)
2169 int idx = calc_load_idx;
2172 * See calc_global_nohz(), if we observe the new index, we also
2173 * need to observe the new update time.
2178 * If the folding window started, make sure we start writing in the
2181 if (!time_before(jiffies, calc_load_update))
2187 static inline int calc_load_read_idx(void)
2189 return calc_load_idx & 1;
2192 void calc_load_enter_idle(void)
2194 struct rq *this_rq = this_rq();
2198 * We're going into NOHZ mode, if there's any pending delta, fold it
2199 * into the pending idle delta.
2201 delta = calc_load_fold_active(this_rq);
2203 int idx = calc_load_write_idx();
2204 atomic_long_add(delta, &calc_load_idle[idx]);
2208 void calc_load_exit_idle(void)
2210 struct rq *this_rq = this_rq();
2213 * If we're still before the sample window, we're done.
2215 if (time_before(jiffies, this_rq->calc_load_update))
2219 * We woke inside or after the sample window, this means we're already
2220 * accounted through the nohz accounting, so skip the entire deal and
2221 * sync up for the next window.
2223 this_rq->calc_load_update = calc_load_update;
2224 if (time_before(jiffies, this_rq->calc_load_update + 10))
2225 this_rq->calc_load_update += LOAD_FREQ;
2228 static long calc_load_fold_idle(void)
2230 int idx = calc_load_read_idx();
2233 if (atomic_long_read(&calc_load_idle[idx]))
2234 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2240 * fixed_power_int - compute: x^n, in O(log n) time
2242 * @x: base of the power
2243 * @frac_bits: fractional bits of @x
2244 * @n: power to raise @x to.
2246 * By exploiting the relation between the definition of the natural power
2247 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2248 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2249 * (where: n_i \elem {0, 1}, the binary vector representing n),
2250 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2251 * of course trivially computable in O(log_2 n), the length of our binary
2254 static unsigned long
2255 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2257 unsigned long result = 1UL << frac_bits;
2262 result += 1UL << (frac_bits - 1);
2263 result >>= frac_bits;
2269 x += 1UL << (frac_bits - 1);
2277 * a1 = a0 * e + a * (1 - e)
2279 * a2 = a1 * e + a * (1 - e)
2280 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2281 * = a0 * e^2 + a * (1 - e) * (1 + e)
2283 * a3 = a2 * e + a * (1 - e)
2284 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2285 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2289 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2290 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2291 * = a0 * e^n + a * (1 - e^n)
2293 * [1] application of the geometric series:
2296 * S_n := \Sum x^i = -------------
2299 static unsigned long
2300 calc_load_n(unsigned long load, unsigned long exp,
2301 unsigned long active, unsigned int n)
2304 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2308 * NO_HZ can leave us missing all per-cpu ticks calling
2309 * calc_load_account_active(), but since an idle CPU folds its delta into
2310 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2311 * in the pending idle delta if our idle period crossed a load cycle boundary.
2313 * Once we've updated the global active value, we need to apply the exponential
2314 * weights adjusted to the number of cycles missed.
2316 static void calc_global_nohz(void)
2318 long delta, active, n;
2320 if (!time_before(jiffies, calc_load_update + 10)) {
2322 * Catch-up, fold however many we are behind still
2324 delta = jiffies - calc_load_update - 10;
2325 n = 1 + (delta / LOAD_FREQ);
2327 active = atomic_long_read(&calc_load_tasks);
2328 active = active > 0 ? active * FIXED_1 : 0;
2330 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2331 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2332 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2334 calc_load_update += n * LOAD_FREQ;
2338 * Flip the idle index...
2340 * Make sure we first write the new time then flip the index, so that
2341 * calc_load_write_idx() will see the new time when it reads the new
2342 * index, this avoids a double flip messing things up.
2347 #else /* !CONFIG_NO_HZ */
2349 static inline long calc_load_fold_idle(void) { return 0; }
2350 static inline void calc_global_nohz(void) { }
2352 #endif /* CONFIG_NO_HZ */
2355 * calc_load - update the avenrun load estimates 10 ticks after the
2356 * CPUs have updated calc_load_tasks.
2358 void calc_global_load(unsigned long ticks)
2362 if (time_before(jiffies, calc_load_update + 10))
2366 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2368 delta = calc_load_fold_idle();
2370 atomic_long_add(delta, &calc_load_tasks);
2372 active = atomic_long_read(&calc_load_tasks);
2373 active = active > 0 ? active * FIXED_1 : 0;
2375 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2376 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2377 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2379 calc_load_update += LOAD_FREQ;
2382 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2388 * Called from update_cpu_load() to periodically update this CPU's
2391 static void calc_load_account_active(struct rq *this_rq)
2395 if (time_before(jiffies, this_rq->calc_load_update))
2398 delta = calc_load_fold_active(this_rq);
2400 atomic_long_add(delta, &calc_load_tasks);
2402 this_rq->calc_load_update += LOAD_FREQ;
2406 * End of global load-average stuff
2410 * The exact cpuload at various idx values, calculated at every tick would be
2411 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2413 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2414 * on nth tick when cpu may be busy, then we have:
2415 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2416 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2418 * decay_load_missed() below does efficient calculation of
2419 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2420 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2422 * The calculation is approximated on a 128 point scale.
2423 * degrade_zero_ticks is the number of ticks after which load at any
2424 * particular idx is approximated to be zero.
2425 * degrade_factor is a precomputed table, a row for each load idx.
2426 * Each column corresponds to degradation factor for a power of two ticks,
2427 * based on 128 point scale.
2429 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2430 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2432 * With this power of 2 load factors, we can degrade the load n times
2433 * by looking at 1 bits in n and doing as many mult/shift instead of
2434 * n mult/shifts needed by the exact degradation.
2436 #define DEGRADE_SHIFT 7
2437 static const unsigned char
2438 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2439 static const unsigned char
2440 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2441 {0, 0, 0, 0, 0, 0, 0, 0},
2442 {64, 32, 8, 0, 0, 0, 0, 0},
2443 {96, 72, 40, 12, 1, 0, 0},
2444 {112, 98, 75, 43, 15, 1, 0},
2445 {120, 112, 98, 76, 45, 16, 2} };
2448 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2449 * would be when CPU is idle and so we just decay the old load without
2450 * adding any new load.
2452 static unsigned long
2453 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2457 if (!missed_updates)
2460 if (missed_updates >= degrade_zero_ticks[idx])
2464 return load >> missed_updates;
2466 while (missed_updates) {
2467 if (missed_updates % 2)
2468 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2470 missed_updates >>= 1;
2477 * Update rq->cpu_load[] statistics. This function is usually called every
2478 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2479 * every tick. We fix it up based on jiffies.
2481 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2482 unsigned long pending_updates)
2486 this_rq->nr_load_updates++;
2488 /* Update our load: */
2489 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2490 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2491 unsigned long old_load, new_load;
2493 /* scale is effectively 1 << i now, and >> i divides by scale */
2495 old_load = this_rq->cpu_load[i];
2496 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2497 new_load = this_load;
2499 * Round up the averaging division if load is increasing. This
2500 * prevents us from getting stuck on 9 if the load is 10, for
2503 if (new_load > old_load)
2504 new_load += scale - 1;
2506 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2509 sched_avg_update(this_rq);
2514 * There is no sane way to deal with nohz on smp when using jiffies because the
2515 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2516 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2518 * Therefore we cannot use the delta approach from the regular tick since that
2519 * would seriously skew the load calculation. However we'll make do for those
2520 * updates happening while idle (nohz_idle_balance) or coming out of idle
2521 * (tick_nohz_idle_exit).
2523 * This means we might still be one tick off for nohz periods.
2527 * Called from nohz_idle_balance() to update the load ratings before doing the
2530 void update_idle_cpu_load(struct rq *this_rq)
2532 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2533 unsigned long load = this_rq->load.weight;
2534 unsigned long pending_updates;
2537 * bail if there's load or we're actually up-to-date.
2539 if (load || curr_jiffies == this_rq->last_load_update_tick)
2542 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2543 this_rq->last_load_update_tick = curr_jiffies;
2545 __update_cpu_load(this_rq, load, pending_updates);
2549 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2551 void update_cpu_load_nohz(void)
2553 struct rq *this_rq = this_rq();
2554 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2555 unsigned long pending_updates;
2557 if (curr_jiffies == this_rq->last_load_update_tick)
2560 raw_spin_lock(&this_rq->lock);
2561 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2562 if (pending_updates) {
2563 this_rq->last_load_update_tick = curr_jiffies;
2565 * We were idle, this means load 0, the current load might be
2566 * !0 due to remote wakeups and the sort.
2568 __update_cpu_load(this_rq, 0, pending_updates);
2570 raw_spin_unlock(&this_rq->lock);
2572 #endif /* CONFIG_NO_HZ */
2575 * Called from scheduler_tick()
2577 static void update_cpu_load_active(struct rq *this_rq)
2580 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2582 this_rq->last_load_update_tick = jiffies;
2583 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2585 calc_load_account_active(this_rq);
2591 * sched_exec - execve() is a valuable balancing opportunity, because at
2592 * this point the task has the smallest effective memory and cache footprint.
2594 void sched_exec(void)
2596 struct task_struct *p = current;
2597 unsigned long flags;
2600 raw_spin_lock_irqsave(&p->pi_lock, flags);
2601 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2602 if (dest_cpu == smp_processor_id())
2605 if (likely(cpu_active(dest_cpu))) {
2606 struct migration_arg arg = { p, dest_cpu };
2608 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2609 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2613 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2618 DEFINE_PER_CPU(struct kernel_stat, kstat);
2619 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2621 EXPORT_PER_CPU_SYMBOL(kstat);
2622 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2625 * Return any ns on the sched_clock that have not yet been accounted in
2626 * @p in case that task is currently running.
2628 * Called with task_rq_lock() held on @rq.
2630 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2634 if (task_current(rq, p)) {
2635 update_rq_clock(rq);
2636 ns = rq->clock_task - p->se.exec_start;
2644 unsigned long long task_delta_exec(struct task_struct *p)
2646 unsigned long flags;
2650 rq = task_rq_lock(p, &flags);
2651 ns = do_task_delta_exec(p, rq);
2652 task_rq_unlock(rq, p, &flags);
2658 * Return accounted runtime for the task.
2659 * In case the task is currently running, return the runtime plus current's
2660 * pending runtime that have not been accounted yet.
2662 unsigned long long task_sched_runtime(struct task_struct *p)
2664 unsigned long flags;
2668 rq = task_rq_lock(p, &flags);
2669 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2670 task_rq_unlock(rq, p, &flags);
2676 * This function gets called by the timer code, with HZ frequency.
2677 * We call it with interrupts disabled.
2679 void scheduler_tick(void)
2681 int cpu = smp_processor_id();
2682 struct rq *rq = cpu_rq(cpu);
2683 struct task_struct *curr = rq->curr;
2687 raw_spin_lock(&rq->lock);
2688 update_rq_clock(rq);
2689 update_cpu_load_active(rq);
2690 curr->sched_class->task_tick(rq, curr, 0);
2691 raw_spin_unlock(&rq->lock);
2693 perf_event_task_tick();
2696 rq->idle_balance = idle_cpu(cpu);
2697 trigger_load_balance(rq, cpu);
2701 notrace unsigned long get_parent_ip(unsigned long addr)
2703 if (in_lock_functions(addr)) {
2704 addr = CALLER_ADDR2;
2705 if (in_lock_functions(addr))
2706 addr = CALLER_ADDR3;
2711 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2712 defined(CONFIG_PREEMPT_TRACER))
2714 void __kprobes add_preempt_count(int val)
2716 #ifdef CONFIG_DEBUG_PREEMPT
2720 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2723 preempt_count() += val;
2724 #ifdef CONFIG_DEBUG_PREEMPT
2726 * Spinlock count overflowing soon?
2728 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2731 if (preempt_count() == val)
2732 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2734 EXPORT_SYMBOL(add_preempt_count);
2736 void __kprobes sub_preempt_count(int val)
2738 #ifdef CONFIG_DEBUG_PREEMPT
2742 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2745 * Is the spinlock portion underflowing?
2747 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2748 !(preempt_count() & PREEMPT_MASK)))
2752 if (preempt_count() == val)
2753 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2754 preempt_count() -= val;
2756 EXPORT_SYMBOL(sub_preempt_count);
2761 * Print scheduling while atomic bug:
2763 static noinline void __schedule_bug(struct task_struct *prev)
2765 if (oops_in_progress)
2768 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2769 prev->comm, prev->pid, preempt_count());
2771 debug_show_held_locks(prev);
2773 if (irqs_disabled())
2774 print_irqtrace_events(prev);
2776 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2780 * Various schedule()-time debugging checks and statistics:
2782 static inline void schedule_debug(struct task_struct *prev)
2785 * Test if we are atomic. Since do_exit() needs to call into
2786 * schedule() atomically, we ignore that path for now.
2787 * Otherwise, whine if we are scheduling when we should not be.
2789 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2790 __schedule_bug(prev);
2793 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2795 schedstat_inc(this_rq(), sched_count);
2798 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2800 if (prev->on_rq || rq->skip_clock_update < 0)
2801 update_rq_clock(rq);
2802 prev->sched_class->put_prev_task(rq, prev);
2806 * Pick up the highest-prio task:
2808 static inline struct task_struct *
2809 pick_next_task(struct rq *rq)
2811 const struct sched_class *class;
2812 struct task_struct *p;
2815 * Optimization: we know that if all tasks are in
2816 * the fair class we can call that function directly:
2818 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2819 p = fair_sched_class.pick_next_task(rq);
2824 for_each_class(class) {
2825 p = class->pick_next_task(rq);
2830 BUG(); /* the idle class will always have a runnable task */
2834 * __schedule() is the main scheduler function.
2836 * The main means of driving the scheduler and thus entering this function are:
2838 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2840 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2841 * paths. For example, see arch/x86/entry_64.S.
2843 * To drive preemption between tasks, the scheduler sets the flag in timer
2844 * interrupt handler scheduler_tick().
2846 * 3. Wakeups don't really cause entry into schedule(). They add a
2847 * task to the run-queue and that's it.
2849 * Now, if the new task added to the run-queue preempts the current
2850 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2851 * called on the nearest possible occasion:
2853 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2855 * - in syscall or exception context, at the next outmost
2856 * preempt_enable(). (this might be as soon as the wake_up()'s
2859 * - in IRQ context, return from interrupt-handler to
2860 * preemptible context
2862 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2865 * - cond_resched() call
2866 * - explicit schedule() call
2867 * - return from syscall or exception to user-space
2868 * - return from interrupt-handler to user-space
2870 static void __sched __schedule(void)
2872 struct task_struct *prev, *next;
2873 unsigned long *switch_count;
2879 cpu = smp_processor_id();
2881 rcu_note_context_switch(cpu);
2884 schedule_debug(prev);
2886 if (sched_feat(HRTICK))
2889 raw_spin_lock_irq(&rq->lock);
2891 switch_count = &prev->nivcsw;
2892 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2893 if (unlikely(signal_pending_state(prev->state, prev))) {
2894 prev->state = TASK_RUNNING;
2896 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2900 * If a worker went to sleep, notify and ask workqueue
2901 * whether it wants to wake up a task to maintain
2904 if (prev->flags & PF_WQ_WORKER) {
2905 struct task_struct *to_wakeup;
2907 to_wakeup = wq_worker_sleeping(prev, cpu);
2909 try_to_wake_up_local(to_wakeup);
2912 switch_count = &prev->nvcsw;
2915 pre_schedule(rq, prev);
2917 if (unlikely(!rq->nr_running))
2918 idle_balance(cpu, rq);
2920 put_prev_task(rq, prev);
2921 next = pick_next_task(rq);
2922 clear_tsk_need_resched(prev);
2923 rq->skip_clock_update = 0;
2925 if (likely(prev != next)) {
2930 context_switch(rq, prev, next); /* unlocks the rq */
2932 * The context switch have flipped the stack from under us
2933 * and restored the local variables which were saved when
2934 * this task called schedule() in the past. prev == current
2935 * is still correct, but it can be moved to another cpu/rq.
2937 cpu = smp_processor_id();
2940 raw_spin_unlock_irq(&rq->lock);
2944 sched_preempt_enable_no_resched();
2949 static inline void sched_submit_work(struct task_struct *tsk)
2951 if (!tsk->state || tsk_is_pi_blocked(tsk))
2954 * If we are going to sleep and we have plugged IO queued,
2955 * make sure to submit it to avoid deadlocks.
2957 if (blk_needs_flush_plug(tsk))
2958 blk_schedule_flush_plug(tsk);
2961 asmlinkage void __sched schedule(void)
2963 struct task_struct *tsk = current;
2965 sched_submit_work(tsk);
2968 EXPORT_SYMBOL(schedule);
2970 #ifdef CONFIG_CONTEXT_TRACKING
2971 asmlinkage void __sched schedule_user(void)
2974 * If we come here after a random call to set_need_resched(),
2975 * or we have been woken up remotely but the IPI has not yet arrived,
2976 * we haven't yet exited the RCU idle mode. Do it here manually until
2977 * we find a better solution.
2986 * schedule_preempt_disabled - called with preemption disabled
2988 * Returns with preemption disabled. Note: preempt_count must be 1
2990 void __sched schedule_preempt_disabled(void)
2992 sched_preempt_enable_no_resched();
2997 #ifdef CONFIG_PREEMPT
2999 * this is the entry point to schedule() from in-kernel preemption
3000 * off of preempt_enable. Kernel preemptions off return from interrupt
3001 * occur there and call schedule directly.
3003 asmlinkage void __sched notrace preempt_schedule(void)
3005 struct thread_info *ti = current_thread_info();
3008 * If there is a non-zero preempt_count or interrupts are disabled,
3009 * we do not want to preempt the current task. Just return..
3011 if (likely(ti->preempt_count || irqs_disabled()))
3015 add_preempt_count_notrace(PREEMPT_ACTIVE);
3017 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3020 * Check again in case we missed a preemption opportunity
3021 * between schedule and now.
3024 } while (need_resched());
3026 EXPORT_SYMBOL(preempt_schedule);
3029 * this is the entry point to schedule() from kernel preemption
3030 * off of irq context.
3031 * Note, that this is called and return with irqs disabled. This will
3032 * protect us against recursive calling from irq.
3034 asmlinkage void __sched preempt_schedule_irq(void)
3036 struct thread_info *ti = current_thread_info();
3037 enum ctx_state prev_state;
3039 /* Catch callers which need to be fixed */
3040 BUG_ON(ti->preempt_count || !irqs_disabled());
3042 prev_state = exception_enter();
3045 add_preempt_count(PREEMPT_ACTIVE);
3048 local_irq_disable();
3049 sub_preempt_count(PREEMPT_ACTIVE);
3052 * Check again in case we missed a preemption opportunity
3053 * between schedule and now.
3056 } while (need_resched());
3058 exception_exit(prev_state);
3061 #endif /* CONFIG_PREEMPT */
3063 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3066 return try_to_wake_up(curr->private, mode, wake_flags);
3068 EXPORT_SYMBOL(default_wake_function);
3071 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3072 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3073 * number) then we wake all the non-exclusive tasks and one exclusive task.
3075 * There are circumstances in which we can try to wake a task which has already
3076 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3077 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3079 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3080 int nr_exclusive, int wake_flags, void *key)
3082 wait_queue_t *curr, *next;
3084 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3085 unsigned flags = curr->flags;
3087 if (curr->func(curr, mode, wake_flags, key) &&
3088 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3094 * __wake_up - wake up threads blocked on a waitqueue.
3096 * @mode: which threads
3097 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3098 * @key: is directly passed to the wakeup function
3100 * It may be assumed that this function implies a write memory barrier before
3101 * changing the task state if and only if any tasks are woken up.
3103 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3104 int nr_exclusive, void *key)
3106 unsigned long flags;
3108 spin_lock_irqsave(&q->lock, flags);
3109 __wake_up_common(q, mode, nr_exclusive, 0, key);
3110 spin_unlock_irqrestore(&q->lock, flags);
3112 EXPORT_SYMBOL(__wake_up);
3115 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3117 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3119 __wake_up_common(q, mode, nr, 0, NULL);
3121 EXPORT_SYMBOL_GPL(__wake_up_locked);
3123 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3125 __wake_up_common(q, mode, 1, 0, key);
3127 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3130 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3132 * @mode: which threads
3133 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3134 * @key: opaque value to be passed to wakeup targets
3136 * The sync wakeup differs that the waker knows that it will schedule
3137 * away soon, so while the target thread will be woken up, it will not
3138 * be migrated to another CPU - ie. the two threads are 'synchronized'
3139 * with each other. This can prevent needless bouncing between CPUs.
3141 * On UP it can prevent extra preemption.
3143 * It may be assumed that this function implies a write memory barrier before
3144 * changing the task state if and only if any tasks are woken up.
3146 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3147 int nr_exclusive, void *key)
3149 unsigned long flags;
3150 int wake_flags = WF_SYNC;
3155 if (unlikely(!nr_exclusive))
3158 spin_lock_irqsave(&q->lock, flags);
3159 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3160 spin_unlock_irqrestore(&q->lock, flags);
3162 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3165 * __wake_up_sync - see __wake_up_sync_key()
3167 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3169 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3171 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3174 * complete: - signals a single thread waiting on this completion
3175 * @x: holds the state of this particular completion
3177 * This will wake up a single thread waiting on this completion. Threads will be
3178 * awakened in the same order in which they were queued.
3180 * See also complete_all(), wait_for_completion() and related routines.
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 complete(struct completion *x)
3187 unsigned long flags;
3189 spin_lock_irqsave(&x->wait.lock, flags);
3191 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3192 spin_unlock_irqrestore(&x->wait.lock, flags);
3194 EXPORT_SYMBOL(complete);
3197 * complete_all: - signals all threads waiting on this completion
3198 * @x: holds the state of this particular completion
3200 * This will wake up all threads waiting on this particular completion event.
3202 * It may be assumed that this function implies a write memory barrier before
3203 * changing the task state if and only if any tasks are woken up.
3205 void complete_all(struct completion *x)
3207 unsigned long flags;
3209 spin_lock_irqsave(&x->wait.lock, flags);
3210 x->done += UINT_MAX/2;
3211 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3212 spin_unlock_irqrestore(&x->wait.lock, flags);
3214 EXPORT_SYMBOL(complete_all);
3216 static inline long __sched
3217 do_wait_for_common(struct completion *x,
3218 long (*action)(long), long timeout, int state)
3221 DECLARE_WAITQUEUE(wait, current);
3223 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3225 if (signal_pending_state(state, current)) {
3226 timeout = -ERESTARTSYS;
3229 __set_current_state(state);
3230 spin_unlock_irq(&x->wait.lock);
3231 timeout = action(timeout);
3232 spin_lock_irq(&x->wait.lock);
3233 } while (!x->done && timeout);
3234 __remove_wait_queue(&x->wait, &wait);
3239 return timeout ?: 1;
3242 static inline long __sched
3243 __wait_for_common(struct completion *x,
3244 long (*action)(long), long timeout, int state)
3248 spin_lock_irq(&x->wait.lock);
3249 timeout = do_wait_for_common(x, action, timeout, state);
3250 spin_unlock_irq(&x->wait.lock);
3255 wait_for_common(struct completion *x, long timeout, int state)
3257 return __wait_for_common(x, schedule_timeout, timeout, state);
3261 wait_for_common_io(struct completion *x, long timeout, int state)
3263 return __wait_for_common(x, io_schedule_timeout, timeout, state);
3267 * wait_for_completion: - waits for completion of a task
3268 * @x: holds the state of this particular completion
3270 * This waits to be signaled for completion of a specific task. It is NOT
3271 * interruptible and there is no timeout.
3273 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3274 * and interrupt capability. Also see complete().
3276 void __sched wait_for_completion(struct completion *x)
3278 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3280 EXPORT_SYMBOL(wait_for_completion);
3283 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3284 * @x: holds the state of this particular completion
3285 * @timeout: timeout value in jiffies
3287 * This waits for either a completion of a specific task to be signaled or for a
3288 * specified timeout to expire. The timeout is in jiffies. It is not
3291 * The return value is 0 if timed out, and positive (at least 1, or number of
3292 * jiffies left till timeout) if completed.
3294 unsigned long __sched
3295 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3297 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3299 EXPORT_SYMBOL(wait_for_completion_timeout);
3302 * wait_for_completion_io: - waits for completion of a task
3303 * @x: holds the state of this particular completion
3305 * This waits to be signaled for completion of a specific task. It is NOT
3306 * interruptible and there is no timeout. The caller is accounted as waiting
3309 void __sched wait_for_completion_io(struct completion *x)
3311 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3313 EXPORT_SYMBOL(wait_for_completion_io);
3316 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3317 * @x: holds the state of this particular completion
3318 * @timeout: timeout value in jiffies
3320 * This waits for either a completion of a specific task to be signaled or for a
3321 * specified timeout to expire. The timeout is in jiffies. It is not
3322 * interruptible. The caller is accounted as waiting for IO.
3324 * The return value is 0 if timed out, and positive (at least 1, or number of
3325 * jiffies left till timeout) if completed.
3327 unsigned long __sched
3328 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
3330 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
3332 EXPORT_SYMBOL(wait_for_completion_io_timeout);
3335 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3336 * @x: holds the state of this particular completion
3338 * This waits for completion of a specific task to be signaled. It is
3341 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3343 int __sched wait_for_completion_interruptible(struct completion *x)
3345 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3346 if (t == -ERESTARTSYS)
3350 EXPORT_SYMBOL(wait_for_completion_interruptible);
3353 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3354 * @x: holds the state of this particular completion
3355 * @timeout: timeout value in jiffies
3357 * This waits for either a completion of a specific task to be signaled or for a
3358 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3360 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3361 * positive (at least 1, or number of jiffies left till timeout) if completed.
3364 wait_for_completion_interruptible_timeout(struct completion *x,
3365 unsigned long timeout)
3367 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3369 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3372 * wait_for_completion_killable: - waits for completion of a task (killable)
3373 * @x: holds the state of this particular completion
3375 * This waits to be signaled for completion of a specific task. It can be
3376 * interrupted by a kill signal.
3378 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3380 int __sched wait_for_completion_killable(struct completion *x)
3382 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3383 if (t == -ERESTARTSYS)
3387 EXPORT_SYMBOL(wait_for_completion_killable);
3390 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3391 * @x: holds the state of this particular completion
3392 * @timeout: timeout value in jiffies
3394 * This waits for either a completion of a specific task to be
3395 * signaled or for a specified timeout to expire. It can be
3396 * interrupted by a kill signal. The timeout is in jiffies.
3398 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3399 * positive (at least 1, or number of jiffies left till timeout) if completed.
3402 wait_for_completion_killable_timeout(struct completion *x,
3403 unsigned long timeout)
3405 return wait_for_common(x, timeout, TASK_KILLABLE);
3407 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3410 * try_wait_for_completion - try to decrement a completion without blocking
3411 * @x: completion structure
3413 * Returns: 0 if a decrement cannot be done without blocking
3414 * 1 if a decrement succeeded.
3416 * If a completion is being used as a counting completion,
3417 * attempt to decrement the counter without blocking. This
3418 * enables us to avoid waiting if the resource the completion
3419 * is protecting is not available.
3421 bool try_wait_for_completion(struct completion *x)
3423 unsigned long flags;
3426 spin_lock_irqsave(&x->wait.lock, flags);
3431 spin_unlock_irqrestore(&x->wait.lock, flags);
3434 EXPORT_SYMBOL(try_wait_for_completion);
3437 * completion_done - Test to see if a completion has any waiters
3438 * @x: completion structure
3440 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3441 * 1 if there are no waiters.
3444 bool completion_done(struct completion *x)
3446 unsigned long flags;
3449 spin_lock_irqsave(&x->wait.lock, flags);
3452 spin_unlock_irqrestore(&x->wait.lock, flags);
3455 EXPORT_SYMBOL(completion_done);
3458 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3460 unsigned long flags;
3463 init_waitqueue_entry(&wait, current);
3465 __set_current_state(state);
3467 spin_lock_irqsave(&q->lock, flags);
3468 __add_wait_queue(q, &wait);
3469 spin_unlock(&q->lock);
3470 timeout = schedule_timeout(timeout);
3471 spin_lock_irq(&q->lock);
3472 __remove_wait_queue(q, &wait);
3473 spin_unlock_irqrestore(&q->lock, flags);
3478 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3480 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3482 EXPORT_SYMBOL(interruptible_sleep_on);
3485 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3487 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3489 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3491 void __sched sleep_on(wait_queue_head_t *q)
3493 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3495 EXPORT_SYMBOL(sleep_on);
3497 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3499 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3501 EXPORT_SYMBOL(sleep_on_timeout);
3503 #ifdef CONFIG_RT_MUTEXES
3506 * rt_mutex_setprio - set the current priority of a task
3508 * @prio: prio value (kernel-internal form)
3510 * This function changes the 'effective' priority of a task. It does
3511 * not touch ->normal_prio like __setscheduler().
3513 * Used by the rt_mutex code to implement priority inheritance logic.
3515 void rt_mutex_setprio(struct task_struct *p, int prio)
3517 int oldprio, on_rq, running;
3519 const struct sched_class *prev_class;
3521 BUG_ON(prio < 0 || prio > MAX_PRIO);
3523 rq = __task_rq_lock(p);
3526 * Idle task boosting is a nono in general. There is one
3527 * exception, when PREEMPT_RT and NOHZ is active:
3529 * The idle task calls get_next_timer_interrupt() and holds
3530 * the timer wheel base->lock on the CPU and another CPU wants
3531 * to access the timer (probably to cancel it). We can safely
3532 * ignore the boosting request, as the idle CPU runs this code
3533 * with interrupts disabled and will complete the lock
3534 * protected section without being interrupted. So there is no
3535 * real need to boost.
3537 if (unlikely(p == rq->idle)) {
3538 WARN_ON(p != rq->curr);
3539 WARN_ON(p->pi_blocked_on);
3543 trace_sched_pi_setprio(p, prio);
3545 prev_class = p->sched_class;
3547 running = task_current(rq, p);
3549 dequeue_task(rq, p, 0);
3551 p->sched_class->put_prev_task(rq, p);
3554 p->sched_class = &rt_sched_class;
3556 p->sched_class = &fair_sched_class;
3561 p->sched_class->set_curr_task(rq);
3563 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3565 check_class_changed(rq, p, prev_class, oldprio);
3567 __task_rq_unlock(rq);
3570 void set_user_nice(struct task_struct *p, long nice)
3572 int old_prio, delta, on_rq;
3573 unsigned long flags;
3576 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3579 * We have to be careful, if called from sys_setpriority(),
3580 * the task might be in the middle of scheduling on another CPU.
3582 rq = task_rq_lock(p, &flags);
3584 * The RT priorities are set via sched_setscheduler(), but we still
3585 * allow the 'normal' nice value to be set - but as expected
3586 * it wont have any effect on scheduling until the task is
3587 * SCHED_FIFO/SCHED_RR:
3589 if (task_has_rt_policy(p)) {
3590 p->static_prio = NICE_TO_PRIO(nice);
3595 dequeue_task(rq, p, 0);
3597 p->static_prio = NICE_TO_PRIO(nice);
3600 p->prio = effective_prio(p);
3601 delta = p->prio - old_prio;
3604 enqueue_task(rq, p, 0);
3606 * If the task increased its priority or is running and
3607 * lowered its priority, then reschedule its CPU:
3609 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3610 resched_task(rq->curr);
3613 task_rq_unlock(rq, p, &flags);
3615 EXPORT_SYMBOL(set_user_nice);
3618 * can_nice - check if a task can reduce its nice value
3622 int can_nice(const struct task_struct *p, const int nice)
3624 /* convert nice value [19,-20] to rlimit style value [1,40] */
3625 int nice_rlim = 20 - nice;
3627 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3628 capable(CAP_SYS_NICE));
3631 #ifdef __ARCH_WANT_SYS_NICE
3634 * sys_nice - change the priority of the current process.
3635 * @increment: priority increment
3637 * sys_setpriority is a more generic, but much slower function that
3638 * does similar things.
3640 SYSCALL_DEFINE1(nice, int, increment)
3645 * Setpriority might change our priority at the same moment.
3646 * We don't have to worry. Conceptually one call occurs first
3647 * and we have a single winner.
3649 if (increment < -40)
3654 nice = TASK_NICE(current) + increment;
3660 if (increment < 0 && !can_nice(current, nice))
3663 retval = security_task_setnice(current, nice);
3667 set_user_nice(current, nice);
3674 * task_prio - return the priority value of a given task.
3675 * @p: the task in question.
3677 * This is the priority value as seen by users in /proc.
3678 * RT tasks are offset by -200. Normal tasks are centered
3679 * around 0, value goes from -16 to +15.
3681 int task_prio(const struct task_struct *p)
3683 return p->prio - MAX_RT_PRIO;
3687 * task_nice - return the nice value of a given task.
3688 * @p: the task in question.
3690 int task_nice(const struct task_struct *p)
3692 return TASK_NICE(p);
3694 EXPORT_SYMBOL(task_nice);
3697 * idle_cpu - is a given cpu idle currently?
3698 * @cpu: the processor in question.
3700 int idle_cpu(int cpu)
3702 struct rq *rq = cpu_rq(cpu);
3704 if (rq->curr != rq->idle)
3711 if (!llist_empty(&rq->wake_list))
3719 * idle_task - return the idle task for a given cpu.
3720 * @cpu: the processor in question.
3722 struct task_struct *idle_task(int cpu)
3724 return cpu_rq(cpu)->idle;
3728 * find_process_by_pid - find a process with a matching PID value.
3729 * @pid: the pid in question.
3731 static struct task_struct *find_process_by_pid(pid_t pid)
3733 return pid ? find_task_by_vpid(pid) : current;
3736 /* Actually do priority change: must hold rq lock. */
3738 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3741 p->rt_priority = prio;
3742 p->normal_prio = normal_prio(p);
3743 /* we are holding p->pi_lock already */
3744 p->prio = rt_mutex_getprio(p);
3745 if (rt_prio(p->prio))
3746 p->sched_class = &rt_sched_class;
3748 p->sched_class = &fair_sched_class;
3753 * check the target process has a UID that matches the current process's
3755 static bool check_same_owner(struct task_struct *p)
3757 const struct cred *cred = current_cred(), *pcred;
3761 pcred = __task_cred(p);
3762 match = (uid_eq(cred->euid, pcred->euid) ||
3763 uid_eq(cred->euid, pcred->uid));
3768 static int __sched_setscheduler(struct task_struct *p, int policy,
3769 const struct sched_param *param, bool user)
3771 int retval, oldprio, oldpolicy = -1, on_rq, running;
3772 unsigned long flags;
3773 const struct sched_class *prev_class;
3777 /* may grab non-irq protected spin_locks */
3778 BUG_ON(in_interrupt());
3780 /* double check policy once rq lock held */
3782 reset_on_fork = p->sched_reset_on_fork;
3783 policy = oldpolicy = p->policy;
3785 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3786 policy &= ~SCHED_RESET_ON_FORK;
3788 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3789 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3790 policy != SCHED_IDLE)
3795 * Valid priorities for SCHED_FIFO and SCHED_RR are
3796 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3797 * SCHED_BATCH and SCHED_IDLE is 0.
3799 if (param->sched_priority < 0 ||
3800 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3801 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3803 if (rt_policy(policy) != (param->sched_priority != 0))
3807 * Allow unprivileged RT tasks to decrease priority:
3809 if (user && !capable(CAP_SYS_NICE)) {
3810 if (rt_policy(policy)) {
3811 unsigned long rlim_rtprio =
3812 task_rlimit(p, RLIMIT_RTPRIO);
3814 /* can't set/change the rt policy */
3815 if (policy != p->policy && !rlim_rtprio)
3818 /* can't increase priority */
3819 if (param->sched_priority > p->rt_priority &&
3820 param->sched_priority > rlim_rtprio)
3825 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3826 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3828 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3829 if (!can_nice(p, TASK_NICE(p)))
3833 /* can't change other user's priorities */
3834 if (!check_same_owner(p))
3837 /* Normal users shall not reset the sched_reset_on_fork flag */
3838 if (p->sched_reset_on_fork && !reset_on_fork)
3843 retval = security_task_setscheduler(p);
3849 * make sure no PI-waiters arrive (or leave) while we are
3850 * changing the priority of the task:
3852 * To be able to change p->policy safely, the appropriate
3853 * runqueue lock must be held.
3855 rq = task_rq_lock(p, &flags);
3858 * Changing the policy of the stop threads its a very bad idea
3860 if (p == rq->stop) {
3861 task_rq_unlock(rq, p, &flags);
3866 * If not changing anything there's no need to proceed further:
3868 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3869 param->sched_priority == p->rt_priority))) {
3870 task_rq_unlock(rq, p, &flags);
3874 #ifdef CONFIG_RT_GROUP_SCHED
3877 * Do not allow realtime tasks into groups that have no runtime
3880 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3881 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3882 !task_group_is_autogroup(task_group(p))) {
3883 task_rq_unlock(rq, p, &flags);
3889 /* recheck policy now with rq lock held */
3890 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3891 policy = oldpolicy = -1;
3892 task_rq_unlock(rq, p, &flags);
3896 running = task_current(rq, p);
3898 dequeue_task(rq, p, 0);
3900 p->sched_class->put_prev_task(rq, p);
3902 p->sched_reset_on_fork = reset_on_fork;
3905 prev_class = p->sched_class;
3906 __setscheduler(rq, p, policy, param->sched_priority);
3909 p->sched_class->set_curr_task(rq);
3911 enqueue_task(rq, p, 0);
3913 check_class_changed(rq, p, prev_class, oldprio);
3914 task_rq_unlock(rq, p, &flags);
3916 rt_mutex_adjust_pi(p);
3922 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3923 * @p: the task in question.
3924 * @policy: new policy.
3925 * @param: structure containing the new RT priority.
3927 * NOTE that the task may be already dead.
3929 int sched_setscheduler(struct task_struct *p, int policy,
3930 const struct sched_param *param)
3932 return __sched_setscheduler(p, policy, param, true);
3934 EXPORT_SYMBOL_GPL(sched_setscheduler);
3937 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3938 * @p: the task in question.
3939 * @policy: new policy.
3940 * @param: structure containing the new RT priority.
3942 * Just like sched_setscheduler, only don't bother checking if the
3943 * current context has permission. For example, this is needed in
3944 * stop_machine(): we create temporary high priority worker threads,
3945 * but our caller might not have that capability.
3947 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3948 const struct sched_param *param)
3950 return __sched_setscheduler(p, policy, param, false);
3954 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3956 struct sched_param lparam;
3957 struct task_struct *p;
3960 if (!param || pid < 0)
3962 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3967 p = find_process_by_pid(pid);
3969 retval = sched_setscheduler(p, policy, &lparam);
3976 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3977 * @pid: the pid in question.
3978 * @policy: new policy.
3979 * @param: structure containing the new RT priority.
3981 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3982 struct sched_param __user *, param)
3984 /* negative values for policy are not valid */
3988 return do_sched_setscheduler(pid, policy, param);
3992 * sys_sched_setparam - set/change the RT priority of a thread
3993 * @pid: the pid in question.
3994 * @param: structure containing the new RT priority.
3996 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3998 return do_sched_setscheduler(pid, -1, param);
4002 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4003 * @pid: the pid in question.
4005 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4007 struct task_struct *p;
4015 p = find_process_by_pid(pid);
4017 retval = security_task_getscheduler(p);
4020 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4027 * sys_sched_getparam - get the RT priority of a thread
4028 * @pid: the pid in question.
4029 * @param: structure containing the RT priority.
4031 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4033 struct sched_param lp;
4034 struct task_struct *p;
4037 if (!param || pid < 0)
4041 p = find_process_by_pid(pid);
4046 retval = security_task_getscheduler(p);
4050 lp.sched_priority = p->rt_priority;
4054 * This one might sleep, we cannot do it with a spinlock held ...
4056 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4065 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4067 cpumask_var_t cpus_allowed, new_mask;
4068 struct task_struct *p;
4074 p = find_process_by_pid(pid);
4081 /* Prevent p going away */
4085 if (p->flags & PF_NO_SETAFFINITY) {
4089 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4093 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4095 goto out_free_cpus_allowed;
4098 if (!check_same_owner(p)) {
4100 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4107 retval = security_task_setscheduler(p);
4111 cpuset_cpus_allowed(p, cpus_allowed);
4112 cpumask_and(new_mask, in_mask, cpus_allowed);
4114 retval = set_cpus_allowed_ptr(p, new_mask);
4117 cpuset_cpus_allowed(p, cpus_allowed);
4118 if (!cpumask_subset(new_mask, cpus_allowed)) {
4120 * We must have raced with a concurrent cpuset
4121 * update. Just reset the cpus_allowed to the
4122 * cpuset's cpus_allowed
4124 cpumask_copy(new_mask, cpus_allowed);
4129 free_cpumask_var(new_mask);
4130 out_free_cpus_allowed:
4131 free_cpumask_var(cpus_allowed);
4138 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4139 struct cpumask *new_mask)
4141 if (len < cpumask_size())
4142 cpumask_clear(new_mask);
4143 else if (len > cpumask_size())
4144 len = cpumask_size();
4146 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4150 * sys_sched_setaffinity - set the cpu affinity of a process
4151 * @pid: pid of the process
4152 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4153 * @user_mask_ptr: user-space pointer to the new cpu mask
4155 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4156 unsigned long __user *, user_mask_ptr)
4158 cpumask_var_t new_mask;
4161 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4164 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4166 retval = sched_setaffinity(pid, new_mask);
4167 free_cpumask_var(new_mask);
4171 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4173 struct task_struct *p;
4174 unsigned long flags;
4181 p = find_process_by_pid(pid);
4185 retval = security_task_getscheduler(p);
4189 raw_spin_lock_irqsave(&p->pi_lock, flags);
4190 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4191 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4201 * sys_sched_getaffinity - get the cpu affinity of a process
4202 * @pid: pid of the process
4203 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4204 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4206 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4207 unsigned long __user *, user_mask_ptr)
4212 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4214 if (len & (sizeof(unsigned long)-1))
4217 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4220 ret = sched_getaffinity(pid, mask);
4222 size_t retlen = min_t(size_t, len, cpumask_size());
4224 if (copy_to_user(user_mask_ptr, mask, retlen))
4229 free_cpumask_var(mask);
4235 * sys_sched_yield - yield the current processor to other threads.
4237 * This function yields the current CPU to other tasks. If there are no
4238 * other threads running on this CPU then this function will return.
4240 SYSCALL_DEFINE0(sched_yield)
4242 struct rq *rq = this_rq_lock();
4244 schedstat_inc(rq, yld_count);
4245 current->sched_class->yield_task(rq);
4248 * Since we are going to call schedule() anyway, there's
4249 * no need to preempt or enable interrupts:
4251 __release(rq->lock);
4252 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4253 do_raw_spin_unlock(&rq->lock);
4254 sched_preempt_enable_no_resched();
4261 static inline int should_resched(void)
4263 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4266 static void __cond_resched(void)
4268 add_preempt_count(PREEMPT_ACTIVE);
4270 sub_preempt_count(PREEMPT_ACTIVE);
4273 int __sched _cond_resched(void)
4275 if (should_resched()) {
4281 EXPORT_SYMBOL(_cond_resched);
4284 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4285 * call schedule, and on return reacquire the lock.
4287 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4288 * operations here to prevent schedule() from being called twice (once via
4289 * spin_unlock(), once by hand).
4291 int __cond_resched_lock(spinlock_t *lock)
4293 int resched = should_resched();
4296 lockdep_assert_held(lock);
4298 if (spin_needbreak(lock) || resched) {
4309 EXPORT_SYMBOL(__cond_resched_lock);
4311 int __sched __cond_resched_softirq(void)
4313 BUG_ON(!in_softirq());
4315 if (should_resched()) {
4323 EXPORT_SYMBOL(__cond_resched_softirq);
4326 * yield - yield the current processor to other threads.
4328 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4330 * The scheduler is at all times free to pick the calling task as the most
4331 * eligible task to run, if removing the yield() call from your code breaks
4332 * it, its already broken.
4334 * Typical broken usage is:
4339 * where one assumes that yield() will let 'the other' process run that will
4340 * make event true. If the current task is a SCHED_FIFO task that will never
4341 * happen. Never use yield() as a progress guarantee!!
4343 * If you want to use yield() to wait for something, use wait_event().
4344 * If you want to use yield() to be 'nice' for others, use cond_resched().
4345 * If you still want to use yield(), do not!
4347 void __sched yield(void)
4349 set_current_state(TASK_RUNNING);
4352 EXPORT_SYMBOL(yield);
4355 * yield_to - yield the current processor to another thread in
4356 * your thread group, or accelerate that thread toward the
4357 * processor it's on.
4359 * @preempt: whether task preemption is allowed or not
4361 * It's the caller's job to ensure that the target task struct
4362 * can't go away on us before we can do any checks.
4365 * true (>0) if we indeed boosted the target task.
4366 * false (0) if we failed to boost the target.
4367 * -ESRCH if there's no task to yield to.
4369 bool __sched yield_to(struct task_struct *p, bool preempt)
4371 struct task_struct *curr = current;
4372 struct rq *rq, *p_rq;
4373 unsigned long flags;
4376 local_irq_save(flags);
4382 * If we're the only runnable task on the rq and target rq also
4383 * has only one task, there's absolutely no point in yielding.
4385 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4390 double_rq_lock(rq, p_rq);
4391 while (task_rq(p) != p_rq) {
4392 double_rq_unlock(rq, p_rq);
4396 if (!curr->sched_class->yield_to_task)
4399 if (curr->sched_class != p->sched_class)
4402 if (task_running(p_rq, p) || p->state)
4405 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4407 schedstat_inc(rq, yld_count);
4409 * Make p's CPU reschedule; pick_next_entity takes care of
4412 if (preempt && rq != p_rq)
4413 resched_task(p_rq->curr);
4417 double_rq_unlock(rq, p_rq);
4419 local_irq_restore(flags);
4426 EXPORT_SYMBOL_GPL(yield_to);
4429 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4430 * that process accounting knows that this is a task in IO wait state.
4432 void __sched io_schedule(void)
4434 struct rq *rq = raw_rq();
4436 delayacct_blkio_start();
4437 atomic_inc(&rq->nr_iowait);
4438 blk_flush_plug(current);
4439 current->in_iowait = 1;
4441 current->in_iowait = 0;
4442 atomic_dec(&rq->nr_iowait);
4443 delayacct_blkio_end();
4445 EXPORT_SYMBOL(io_schedule);
4447 long __sched io_schedule_timeout(long timeout)
4449 struct rq *rq = raw_rq();
4452 delayacct_blkio_start();
4453 atomic_inc(&rq->nr_iowait);
4454 blk_flush_plug(current);
4455 current->in_iowait = 1;
4456 ret = schedule_timeout(timeout);
4457 current->in_iowait = 0;
4458 atomic_dec(&rq->nr_iowait);
4459 delayacct_blkio_end();
4464 * sys_sched_get_priority_max - return maximum RT priority.
4465 * @policy: scheduling class.
4467 * this syscall returns the maximum rt_priority that can be used
4468 * by a given scheduling class.
4470 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4477 ret = MAX_USER_RT_PRIO-1;
4489 * sys_sched_get_priority_min - return minimum RT priority.
4490 * @policy: scheduling class.
4492 * this syscall returns the minimum rt_priority that can be used
4493 * by a given scheduling class.
4495 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4513 * sys_sched_rr_get_interval - return the default timeslice of a process.
4514 * @pid: pid of the process.
4515 * @interval: userspace pointer to the timeslice value.
4517 * this syscall writes the default timeslice value of a given process
4518 * into the user-space timespec buffer. A value of '0' means infinity.
4520 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4521 struct timespec __user *, interval)
4523 struct task_struct *p;
4524 unsigned int time_slice;
4525 unsigned long flags;
4535 p = find_process_by_pid(pid);
4539 retval = security_task_getscheduler(p);
4543 rq = task_rq_lock(p, &flags);
4544 time_slice = p->sched_class->get_rr_interval(rq, p);
4545 task_rq_unlock(rq, p, &flags);
4548 jiffies_to_timespec(time_slice, &t);
4549 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4557 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4559 void sched_show_task(struct task_struct *p)
4561 unsigned long free = 0;
4565 state = p->state ? __ffs(p->state) + 1 : 0;
4566 printk(KERN_INFO "%-15.15s %c", p->comm,
4567 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4568 #if BITS_PER_LONG == 32
4569 if (state == TASK_RUNNING)
4570 printk(KERN_CONT " running ");
4572 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4574 if (state == TASK_RUNNING)
4575 printk(KERN_CONT " running task ");
4577 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4579 #ifdef CONFIG_DEBUG_STACK_USAGE
4580 free = stack_not_used(p);
4583 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4585 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4586 task_pid_nr(p), ppid,
4587 (unsigned long)task_thread_info(p)->flags);
4589 print_worker_info(KERN_INFO, p);
4590 show_stack(p, NULL);
4593 void show_state_filter(unsigned long state_filter)
4595 struct task_struct *g, *p;
4597 #if BITS_PER_LONG == 32
4599 " task PC stack pid father\n");
4602 " task PC stack pid father\n");
4605 do_each_thread(g, p) {
4607 * reset the NMI-timeout, listing all files on a slow
4608 * console might take a lot of time:
4610 touch_nmi_watchdog();
4611 if (!state_filter || (p->state & state_filter))
4613 } while_each_thread(g, p);
4615 touch_all_softlockup_watchdogs();
4617 #ifdef CONFIG_SCHED_DEBUG
4618 sysrq_sched_debug_show();
4622 * Only show locks if all tasks are dumped:
4625 debug_show_all_locks();
4628 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4630 idle->sched_class = &idle_sched_class;
4634 * init_idle - set up an idle thread for a given CPU
4635 * @idle: task in question
4636 * @cpu: cpu the idle task belongs to
4638 * NOTE: this function does not set the idle thread's NEED_RESCHED
4639 * flag, to make booting more robust.
4641 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4643 struct rq *rq = cpu_rq(cpu);
4644 unsigned long flags;
4646 raw_spin_lock_irqsave(&rq->lock, flags);
4649 idle->state = TASK_RUNNING;
4650 idle->se.exec_start = sched_clock();
4652 do_set_cpus_allowed(idle, cpumask_of(cpu));
4654 * We're having a chicken and egg problem, even though we are
4655 * holding rq->lock, the cpu isn't yet set to this cpu so the
4656 * lockdep check in task_group() will fail.
4658 * Similar case to sched_fork(). / Alternatively we could
4659 * use task_rq_lock() here and obtain the other rq->lock.
4664 __set_task_cpu(idle, cpu);
4667 rq->curr = rq->idle = idle;
4668 #if defined(CONFIG_SMP)
4671 raw_spin_unlock_irqrestore(&rq->lock, flags);
4673 /* Set the preempt count _outside_ the spinlocks! */
4674 task_thread_info(idle)->preempt_count = 0;
4677 * The idle tasks have their own, simple scheduling class:
4679 idle->sched_class = &idle_sched_class;
4680 ftrace_graph_init_idle_task(idle, cpu);
4681 vtime_init_idle(idle);
4682 #if defined(CONFIG_SMP)
4683 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4688 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4690 if (p->sched_class && p->sched_class->set_cpus_allowed)
4691 p->sched_class->set_cpus_allowed(p, new_mask);
4693 cpumask_copy(&p->cpus_allowed, new_mask);
4694 p->nr_cpus_allowed = cpumask_weight(new_mask);
4698 * This is how migration works:
4700 * 1) we invoke migration_cpu_stop() on the target CPU using
4702 * 2) stopper starts to run (implicitly forcing the migrated thread
4704 * 3) it checks whether the migrated task is still in the wrong runqueue.
4705 * 4) if it's in the wrong runqueue then the migration thread removes
4706 * it and puts it into the right queue.
4707 * 5) stopper completes and stop_one_cpu() returns and the migration
4712 * Change a given task's CPU affinity. Migrate the thread to a
4713 * proper CPU and schedule it away if the CPU it's executing on
4714 * is removed from the allowed bitmask.
4716 * NOTE: the caller must have a valid reference to the task, the
4717 * task must not exit() & deallocate itself prematurely. The
4718 * call is not atomic; no spinlocks may be held.
4720 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4722 unsigned long flags;
4724 unsigned int dest_cpu;
4727 rq = task_rq_lock(p, &flags);
4729 if (cpumask_equal(&p->cpus_allowed, new_mask))
4732 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4737 do_set_cpus_allowed(p, new_mask);
4739 /* Can the task run on the task's current CPU? If so, we're done */
4740 if (cpumask_test_cpu(task_cpu(p), new_mask))
4743 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4745 struct migration_arg arg = { p, dest_cpu };
4746 /* Need help from migration thread: drop lock and wait. */
4747 task_rq_unlock(rq, p, &flags);
4748 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4749 tlb_migrate_finish(p->mm);
4753 task_rq_unlock(rq, p, &flags);
4757 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4760 * Move (not current) task off this cpu, onto dest cpu. We're doing
4761 * this because either it can't run here any more (set_cpus_allowed()
4762 * away from this CPU, or CPU going down), or because we're
4763 * attempting to rebalance this task on exec (sched_exec).
4765 * So we race with normal scheduler movements, but that's OK, as long
4766 * as the task is no longer on this CPU.
4768 * Returns non-zero if task was successfully migrated.
4770 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4772 struct rq *rq_dest, *rq_src;
4775 if (unlikely(!cpu_active(dest_cpu)))
4778 rq_src = cpu_rq(src_cpu);
4779 rq_dest = cpu_rq(dest_cpu);
4781 raw_spin_lock(&p->pi_lock);
4782 double_rq_lock(rq_src, rq_dest);
4783 /* Already moved. */
4784 if (task_cpu(p) != src_cpu)
4786 /* Affinity changed (again). */
4787 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4791 * If we're not on a rq, the next wake-up will ensure we're
4795 dequeue_task(rq_src, p, 0);
4796 set_task_cpu(p, dest_cpu);
4797 enqueue_task(rq_dest, p, 0);
4798 check_preempt_curr(rq_dest, p, 0);
4803 double_rq_unlock(rq_src, rq_dest);
4804 raw_spin_unlock(&p->pi_lock);
4809 * migration_cpu_stop - this will be executed by a highprio stopper thread
4810 * and performs thread migration by bumping thread off CPU then
4811 * 'pushing' onto another runqueue.
4813 static int migration_cpu_stop(void *data)
4815 struct migration_arg *arg = data;
4818 * The original target cpu might have gone down and we might
4819 * be on another cpu but it doesn't matter.
4821 local_irq_disable();
4822 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4827 #ifdef CONFIG_HOTPLUG_CPU
4830 * Ensures that the idle task is using init_mm right before its cpu goes
4833 void idle_task_exit(void)
4835 struct mm_struct *mm = current->active_mm;
4837 BUG_ON(cpu_online(smp_processor_id()));
4840 switch_mm(mm, &init_mm, current);
4845 * Since this CPU is going 'away' for a while, fold any nr_active delta
4846 * we might have. Assumes we're called after migrate_tasks() so that the
4847 * nr_active count is stable.
4849 * Also see the comment "Global load-average calculations".
4851 static void calc_load_migrate(struct rq *rq)
4853 long delta = calc_load_fold_active(rq);
4855 atomic_long_add(delta, &calc_load_tasks);
4859 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4860 * try_to_wake_up()->select_task_rq().
4862 * Called with rq->lock held even though we'er in stop_machine() and
4863 * there's no concurrency possible, we hold the required locks anyway
4864 * because of lock validation efforts.
4866 static void migrate_tasks(unsigned int dead_cpu)
4868 struct rq *rq = cpu_rq(dead_cpu);
4869 struct task_struct *next, *stop = rq->stop;
4873 * Fudge the rq selection such that the below task selection loop
4874 * doesn't get stuck on the currently eligible stop task.
4876 * We're currently inside stop_machine() and the rq is either stuck
4877 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4878 * either way we should never end up calling schedule() until we're
4885 * There's this thread running, bail when that's the only
4888 if (rq->nr_running == 1)
4891 next = pick_next_task(rq);
4893 next->sched_class->put_prev_task(rq, next);
4895 /* Find suitable destination for @next, with force if needed. */
4896 dest_cpu = select_fallback_rq(dead_cpu, next);
4897 raw_spin_unlock(&rq->lock);
4899 __migrate_task(next, dead_cpu, dest_cpu);
4901 raw_spin_lock(&rq->lock);
4907 #endif /* CONFIG_HOTPLUG_CPU */
4909 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4911 static struct ctl_table sd_ctl_dir[] = {
4913 .procname = "sched_domain",
4919 static struct ctl_table sd_ctl_root[] = {
4921 .procname = "kernel",
4923 .child = sd_ctl_dir,
4928 static struct ctl_table *sd_alloc_ctl_entry(int n)
4930 struct ctl_table *entry =
4931 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4936 static void sd_free_ctl_entry(struct ctl_table **tablep)
4938 struct ctl_table *entry;
4941 * In the intermediate directories, both the child directory and
4942 * procname are dynamically allocated and could fail but the mode
4943 * will always be set. In the lowest directory the names are
4944 * static strings and all have proc handlers.
4946 for (entry = *tablep; entry->mode; entry++) {
4948 sd_free_ctl_entry(&entry->child);
4949 if (entry->proc_handler == NULL)
4950 kfree(entry->procname);
4957 static int min_load_idx = 0;
4958 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4961 set_table_entry(struct ctl_table *entry,
4962 const char *procname, void *data, int maxlen,
4963 umode_t mode, proc_handler *proc_handler,
4966 entry->procname = procname;
4968 entry->maxlen = maxlen;
4970 entry->proc_handler = proc_handler;
4973 entry->extra1 = &min_load_idx;
4974 entry->extra2 = &max_load_idx;
4978 static struct ctl_table *
4979 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4981 struct ctl_table *table = sd_alloc_ctl_entry(13);
4986 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4987 sizeof(long), 0644, proc_doulongvec_minmax, false);
4988 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4989 sizeof(long), 0644, proc_doulongvec_minmax, false);
4990 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4991 sizeof(int), 0644, proc_dointvec_minmax, true);
4992 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4993 sizeof(int), 0644, proc_dointvec_minmax, true);
4994 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4995 sizeof(int), 0644, proc_dointvec_minmax, true);
4996 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4997 sizeof(int), 0644, proc_dointvec_minmax, true);
4998 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4999 sizeof(int), 0644, proc_dointvec_minmax, true);
5000 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5001 sizeof(int), 0644, proc_dointvec_minmax, false);
5002 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5003 sizeof(int), 0644, proc_dointvec_minmax, false);
5004 set_table_entry(&table[9], "cache_nice_tries",
5005 &sd->cache_nice_tries,
5006 sizeof(int), 0644, proc_dointvec_minmax, false);
5007 set_table_entry(&table[10], "flags", &sd->flags,
5008 sizeof(int), 0644, proc_dointvec_minmax, false);
5009 set_table_entry(&table[11], "name", sd->name,
5010 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5011 /* &table[12] is terminator */
5016 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5018 struct ctl_table *entry, *table;
5019 struct sched_domain *sd;
5020 int domain_num = 0, i;
5023 for_each_domain(cpu, sd)
5025 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5030 for_each_domain(cpu, sd) {
5031 snprintf(buf, 32, "domain%d", i);
5032 entry->procname = kstrdup(buf, GFP_KERNEL);
5034 entry->child = sd_alloc_ctl_domain_table(sd);
5041 static struct ctl_table_header *sd_sysctl_header;
5042 static void register_sched_domain_sysctl(void)
5044 int i, cpu_num = num_possible_cpus();
5045 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5048 WARN_ON(sd_ctl_dir[0].child);
5049 sd_ctl_dir[0].child = entry;
5054 for_each_possible_cpu(i) {
5055 snprintf(buf, 32, "cpu%d", i);
5056 entry->procname = kstrdup(buf, GFP_KERNEL);
5058 entry->child = sd_alloc_ctl_cpu_table(i);
5062 WARN_ON(sd_sysctl_header);
5063 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5066 /* may be called multiple times per register */
5067 static void unregister_sched_domain_sysctl(void)
5069 if (sd_sysctl_header)
5070 unregister_sysctl_table(sd_sysctl_header);
5071 sd_sysctl_header = NULL;
5072 if (sd_ctl_dir[0].child)
5073 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5076 static void register_sched_domain_sysctl(void)
5079 static void unregister_sched_domain_sysctl(void)
5084 static void set_rq_online(struct rq *rq)
5087 const struct sched_class *class;
5089 cpumask_set_cpu(rq->cpu, rq->rd->online);
5092 for_each_class(class) {
5093 if (class->rq_online)
5094 class->rq_online(rq);
5099 static void set_rq_offline(struct rq *rq)
5102 const struct sched_class *class;
5104 for_each_class(class) {
5105 if (class->rq_offline)
5106 class->rq_offline(rq);
5109 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5115 * migration_call - callback that gets triggered when a CPU is added.
5116 * Here we can start up the necessary migration thread for the new CPU.
5118 static int __cpuinit
5119 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5121 int cpu = (long)hcpu;
5122 unsigned long flags;
5123 struct rq *rq = cpu_rq(cpu);
5125 switch (action & ~CPU_TASKS_FROZEN) {
5127 case CPU_UP_PREPARE:
5128 rq->calc_load_update = calc_load_update;
5132 /* Update our root-domain */
5133 raw_spin_lock_irqsave(&rq->lock, flags);
5135 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5139 raw_spin_unlock_irqrestore(&rq->lock, flags);
5142 #ifdef CONFIG_HOTPLUG_CPU
5144 sched_ttwu_pending();
5145 /* Update our root-domain */
5146 raw_spin_lock_irqsave(&rq->lock, flags);
5148 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5152 BUG_ON(rq->nr_running != 1); /* the migration thread */
5153 raw_spin_unlock_irqrestore(&rq->lock, flags);
5157 calc_load_migrate(rq);
5162 update_max_interval();
5168 * Register at high priority so that task migration (migrate_all_tasks)
5169 * happens before everything else. This has to be lower priority than
5170 * the notifier in the perf_event subsystem, though.
5172 static struct notifier_block __cpuinitdata migration_notifier = {
5173 .notifier_call = migration_call,
5174 .priority = CPU_PRI_MIGRATION,
5177 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5178 unsigned long action, void *hcpu)
5180 switch (action & ~CPU_TASKS_FROZEN) {
5182 case CPU_DOWN_FAILED:
5183 set_cpu_active((long)hcpu, true);
5190 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5191 unsigned long action, void *hcpu)
5193 switch (action & ~CPU_TASKS_FROZEN) {
5194 case CPU_DOWN_PREPARE:
5195 set_cpu_active((long)hcpu, false);
5202 static int __init migration_init(void)
5204 void *cpu = (void *)(long)smp_processor_id();
5207 /* Initialize migration for the boot CPU */
5208 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5209 BUG_ON(err == NOTIFY_BAD);
5210 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5211 register_cpu_notifier(&migration_notifier);
5213 /* Register cpu active notifiers */
5214 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5215 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5219 early_initcall(migration_init);
5224 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5226 #ifdef CONFIG_SCHED_DEBUG
5228 static __read_mostly int sched_debug_enabled;
5230 static int __init sched_debug_setup(char *str)
5232 sched_debug_enabled = 1;
5236 early_param("sched_debug", sched_debug_setup);
5238 static inline bool sched_debug(void)
5240 return sched_debug_enabled;
5243 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5244 struct cpumask *groupmask)
5246 struct sched_group *group = sd->groups;
5249 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5250 cpumask_clear(groupmask);
5252 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5254 if (!(sd->flags & SD_LOAD_BALANCE)) {
5255 printk("does not load-balance\n");
5257 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5262 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5264 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5265 printk(KERN_ERR "ERROR: domain->span does not contain "
5268 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5269 printk(KERN_ERR "ERROR: domain->groups does not contain"
5273 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5277 printk(KERN_ERR "ERROR: group is NULL\n");
5282 * Even though we initialize ->power to something semi-sane,
5283 * we leave power_orig unset. This allows us to detect if
5284 * domain iteration is still funny without causing /0 traps.
5286 if (!group->sgp->power_orig) {
5287 printk(KERN_CONT "\n");
5288 printk(KERN_ERR "ERROR: domain->cpu_power not "
5293 if (!cpumask_weight(sched_group_cpus(group))) {
5294 printk(KERN_CONT "\n");
5295 printk(KERN_ERR "ERROR: empty group\n");
5299 if (!(sd->flags & SD_OVERLAP) &&
5300 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5301 printk(KERN_CONT "\n");
5302 printk(KERN_ERR "ERROR: repeated CPUs\n");
5306 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5308 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5310 printk(KERN_CONT " %s", str);
5311 if (group->sgp->power != SCHED_POWER_SCALE) {
5312 printk(KERN_CONT " (cpu_power = %d)",
5316 group = group->next;
5317 } while (group != sd->groups);
5318 printk(KERN_CONT "\n");
5320 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5321 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5324 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5325 printk(KERN_ERR "ERROR: parent span is not a superset "
5326 "of domain->span\n");
5330 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5334 if (!sched_debug_enabled)
5338 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5342 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5345 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5353 #else /* !CONFIG_SCHED_DEBUG */
5354 # define sched_domain_debug(sd, cpu) do { } while (0)
5355 static inline bool sched_debug(void)
5359 #endif /* CONFIG_SCHED_DEBUG */
5361 static int sd_degenerate(struct sched_domain *sd)
5363 if (cpumask_weight(sched_domain_span(sd)) == 1)
5366 /* Following flags need at least 2 groups */
5367 if (sd->flags & (SD_LOAD_BALANCE |
5368 SD_BALANCE_NEWIDLE |
5372 SD_SHARE_PKG_RESOURCES)) {
5373 if (sd->groups != sd->groups->next)
5377 /* Following flags don't use groups */
5378 if (sd->flags & (SD_WAKE_AFFINE))
5385 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5387 unsigned long cflags = sd->flags, pflags = parent->flags;
5389 if (sd_degenerate(parent))
5392 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5395 /* Flags needing groups don't count if only 1 group in parent */
5396 if (parent->groups == parent->groups->next) {
5397 pflags &= ~(SD_LOAD_BALANCE |
5398 SD_BALANCE_NEWIDLE |
5402 SD_SHARE_PKG_RESOURCES);
5403 if (nr_node_ids == 1)
5404 pflags &= ~SD_SERIALIZE;
5406 if (~cflags & pflags)
5412 static void free_rootdomain(struct rcu_head *rcu)
5414 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5416 cpupri_cleanup(&rd->cpupri);
5417 free_cpumask_var(rd->rto_mask);
5418 free_cpumask_var(rd->online);
5419 free_cpumask_var(rd->span);
5423 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5425 struct root_domain *old_rd = NULL;
5426 unsigned long flags;
5428 raw_spin_lock_irqsave(&rq->lock, flags);
5433 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5436 cpumask_clear_cpu(rq->cpu, old_rd->span);
5439 * If we dont want to free the old_rt yet then
5440 * set old_rd to NULL to skip the freeing later
5443 if (!atomic_dec_and_test(&old_rd->refcount))
5447 atomic_inc(&rd->refcount);
5450 cpumask_set_cpu(rq->cpu, rd->span);
5451 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5454 raw_spin_unlock_irqrestore(&rq->lock, flags);
5457 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5460 static int init_rootdomain(struct root_domain *rd)
5462 memset(rd, 0, sizeof(*rd));
5464 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5466 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5468 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5471 if (cpupri_init(&rd->cpupri) != 0)
5476 free_cpumask_var(rd->rto_mask);
5478 free_cpumask_var(rd->online);
5480 free_cpumask_var(rd->span);
5486 * By default the system creates a single root-domain with all cpus as
5487 * members (mimicking the global state we have today).
5489 struct root_domain def_root_domain;
5491 static void init_defrootdomain(void)
5493 init_rootdomain(&def_root_domain);
5495 atomic_set(&def_root_domain.refcount, 1);
5498 static struct root_domain *alloc_rootdomain(void)
5500 struct root_domain *rd;
5502 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5506 if (init_rootdomain(rd) != 0) {
5514 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5516 struct sched_group *tmp, *first;
5525 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5530 } while (sg != first);
5533 static void free_sched_domain(struct rcu_head *rcu)
5535 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5538 * If its an overlapping domain it has private groups, iterate and
5541 if (sd->flags & SD_OVERLAP) {
5542 free_sched_groups(sd->groups, 1);
5543 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5544 kfree(sd->groups->sgp);
5550 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5552 call_rcu(&sd->rcu, free_sched_domain);
5555 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5557 for (; sd; sd = sd->parent)
5558 destroy_sched_domain(sd, cpu);
5562 * Keep a special pointer to the highest sched_domain that has
5563 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5564 * allows us to avoid some pointer chasing select_idle_sibling().
5566 * Also keep a unique ID per domain (we use the first cpu number in
5567 * the cpumask of the domain), this allows us to quickly tell if
5568 * two cpus are in the same cache domain, see cpus_share_cache().
5570 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5571 DEFINE_PER_CPU(int, sd_llc_id);
5573 static void update_top_cache_domain(int cpu)
5575 struct sched_domain *sd;
5578 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5580 id = cpumask_first(sched_domain_span(sd));
5582 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5583 per_cpu(sd_llc_id, cpu) = id;
5587 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5588 * hold the hotplug lock.
5591 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5593 struct rq *rq = cpu_rq(cpu);
5594 struct sched_domain *tmp;
5596 /* Remove the sched domains which do not contribute to scheduling. */
5597 for (tmp = sd; tmp; ) {
5598 struct sched_domain *parent = tmp->parent;
5602 if (sd_parent_degenerate(tmp, parent)) {
5603 tmp->parent = parent->parent;
5605 parent->parent->child = tmp;
5606 destroy_sched_domain(parent, cpu);
5611 if (sd && sd_degenerate(sd)) {
5614 destroy_sched_domain(tmp, cpu);
5619 sched_domain_debug(sd, cpu);
5621 rq_attach_root(rq, rd);
5623 rcu_assign_pointer(rq->sd, sd);
5624 destroy_sched_domains(tmp, cpu);
5626 update_top_cache_domain(cpu);
5629 /* cpus with isolated domains */
5630 static cpumask_var_t cpu_isolated_map;
5632 /* Setup the mask of cpus configured for isolated domains */
5633 static int __init isolated_cpu_setup(char *str)
5635 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5636 cpulist_parse(str, cpu_isolated_map);
5640 __setup("isolcpus=", isolated_cpu_setup);
5642 static const struct cpumask *cpu_cpu_mask(int cpu)
5644 return cpumask_of_node(cpu_to_node(cpu));
5648 struct sched_domain **__percpu sd;
5649 struct sched_group **__percpu sg;
5650 struct sched_group_power **__percpu sgp;
5654 struct sched_domain ** __percpu sd;
5655 struct root_domain *rd;
5665 struct sched_domain_topology_level;
5667 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5668 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5670 #define SDTL_OVERLAP 0x01
5672 struct sched_domain_topology_level {
5673 sched_domain_init_f init;
5674 sched_domain_mask_f mask;
5677 struct sd_data data;
5681 * Build an iteration mask that can exclude certain CPUs from the upwards
5684 * Asymmetric node setups can result in situations where the domain tree is of
5685 * unequal depth, make sure to skip domains that already cover the entire
5688 * In that case build_sched_domains() will have terminated the iteration early
5689 * and our sibling sd spans will be empty. Domains should always include the
5690 * cpu they're built on, so check that.
5693 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5695 const struct cpumask *span = sched_domain_span(sd);
5696 struct sd_data *sdd = sd->private;
5697 struct sched_domain *sibling;
5700 for_each_cpu(i, span) {
5701 sibling = *per_cpu_ptr(sdd->sd, i);
5702 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5705 cpumask_set_cpu(i, sched_group_mask(sg));
5710 * Return the canonical balance cpu for this group, this is the first cpu
5711 * of this group that's also in the iteration mask.
5713 int group_balance_cpu(struct sched_group *sg)
5715 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5719 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5721 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5722 const struct cpumask *span = sched_domain_span(sd);
5723 struct cpumask *covered = sched_domains_tmpmask;
5724 struct sd_data *sdd = sd->private;
5725 struct sched_domain *child;
5728 cpumask_clear(covered);
5730 for_each_cpu(i, span) {
5731 struct cpumask *sg_span;
5733 if (cpumask_test_cpu(i, covered))
5736 child = *per_cpu_ptr(sdd->sd, i);
5738 /* See the comment near build_group_mask(). */
5739 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5742 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5743 GFP_KERNEL, cpu_to_node(cpu));
5748 sg_span = sched_group_cpus(sg);
5750 child = child->child;
5751 cpumask_copy(sg_span, sched_domain_span(child));
5753 cpumask_set_cpu(i, sg_span);
5755 cpumask_or(covered, covered, sg_span);
5757 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5758 if (atomic_inc_return(&sg->sgp->ref) == 1)
5759 build_group_mask(sd, sg);
5762 * Initialize sgp->power such that even if we mess up the
5763 * domains and no possible iteration will get us here, we won't
5766 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5769 * Make sure the first group of this domain contains the
5770 * canonical balance cpu. Otherwise the sched_domain iteration
5771 * breaks. See update_sg_lb_stats().
5773 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5774 group_balance_cpu(sg) == cpu)
5784 sd->groups = groups;
5789 free_sched_groups(first, 0);
5794 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5796 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5797 struct sched_domain *child = sd->child;
5800 cpu = cpumask_first(sched_domain_span(child));
5803 *sg = *per_cpu_ptr(sdd->sg, cpu);
5804 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5805 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5812 * build_sched_groups will build a circular linked list of the groups
5813 * covered by the given span, and will set each group's ->cpumask correctly,
5814 * and ->cpu_power to 0.
5816 * Assumes the sched_domain tree is fully constructed
5819 build_sched_groups(struct sched_domain *sd, int cpu)
5821 struct sched_group *first = NULL, *last = NULL;
5822 struct sd_data *sdd = sd->private;
5823 const struct cpumask *span = sched_domain_span(sd);
5824 struct cpumask *covered;
5827 get_group(cpu, sdd, &sd->groups);
5828 atomic_inc(&sd->groups->ref);
5830 if (cpu != cpumask_first(sched_domain_span(sd)))
5833 lockdep_assert_held(&sched_domains_mutex);
5834 covered = sched_domains_tmpmask;
5836 cpumask_clear(covered);
5838 for_each_cpu(i, span) {
5839 struct sched_group *sg;
5840 int group = get_group(i, sdd, &sg);
5843 if (cpumask_test_cpu(i, covered))
5846 cpumask_clear(sched_group_cpus(sg));
5848 cpumask_setall(sched_group_mask(sg));
5850 for_each_cpu(j, span) {
5851 if (get_group(j, sdd, NULL) != group)
5854 cpumask_set_cpu(j, covered);
5855 cpumask_set_cpu(j, sched_group_cpus(sg));
5870 * Initialize sched groups cpu_power.
5872 * cpu_power indicates the capacity of sched group, which is used while
5873 * distributing the load between different sched groups in a sched domain.
5874 * Typically cpu_power for all the groups in a sched domain will be same unless
5875 * there are asymmetries in the topology. If there are asymmetries, group
5876 * having more cpu_power will pickup more load compared to the group having
5879 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5881 struct sched_group *sg = sd->groups;
5883 WARN_ON(!sd || !sg);
5886 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5888 } while (sg != sd->groups);
5890 if (cpu != group_balance_cpu(sg))
5893 update_group_power(sd, cpu);
5894 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5897 int __weak arch_sd_sibling_asym_packing(void)
5899 return 0*SD_ASYM_PACKING;
5903 * Initializers for schedule domains
5904 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5907 #ifdef CONFIG_SCHED_DEBUG
5908 # define SD_INIT_NAME(sd, type) sd->name = #type
5910 # define SD_INIT_NAME(sd, type) do { } while (0)
5913 #define SD_INIT_FUNC(type) \
5914 static noinline struct sched_domain * \
5915 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5917 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5918 *sd = SD_##type##_INIT; \
5919 SD_INIT_NAME(sd, type); \
5920 sd->private = &tl->data; \
5925 #ifdef CONFIG_SCHED_SMT
5926 SD_INIT_FUNC(SIBLING)
5928 #ifdef CONFIG_SCHED_MC
5931 #ifdef CONFIG_SCHED_BOOK
5935 static int default_relax_domain_level = -1;
5936 int sched_domain_level_max;
5938 static int __init setup_relax_domain_level(char *str)
5940 if (kstrtoint(str, 0, &default_relax_domain_level))
5941 pr_warn("Unable to set relax_domain_level\n");
5945 __setup("relax_domain_level=", setup_relax_domain_level);
5947 static void set_domain_attribute(struct sched_domain *sd,
5948 struct sched_domain_attr *attr)
5952 if (!attr || attr->relax_domain_level < 0) {
5953 if (default_relax_domain_level < 0)
5956 request = default_relax_domain_level;
5958 request = attr->relax_domain_level;
5959 if (request < sd->level) {
5960 /* turn off idle balance on this domain */
5961 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5963 /* turn on idle balance on this domain */
5964 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5968 static void __sdt_free(const struct cpumask *cpu_map);
5969 static int __sdt_alloc(const struct cpumask *cpu_map);
5971 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5972 const struct cpumask *cpu_map)
5976 if (!atomic_read(&d->rd->refcount))
5977 free_rootdomain(&d->rd->rcu); /* fall through */
5979 free_percpu(d->sd); /* fall through */
5981 __sdt_free(cpu_map); /* fall through */
5987 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5988 const struct cpumask *cpu_map)
5990 memset(d, 0, sizeof(*d));
5992 if (__sdt_alloc(cpu_map))
5993 return sa_sd_storage;
5994 d->sd = alloc_percpu(struct sched_domain *);
5996 return sa_sd_storage;
5997 d->rd = alloc_rootdomain();
6000 return sa_rootdomain;
6004 * NULL the sd_data elements we've used to build the sched_domain and
6005 * sched_group structure so that the subsequent __free_domain_allocs()
6006 * will not free the data we're using.
6008 static void claim_allocations(int cpu, struct sched_domain *sd)
6010 struct sd_data *sdd = sd->private;
6012 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6013 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6015 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6016 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6018 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6019 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6022 #ifdef CONFIG_SCHED_SMT
6023 static const struct cpumask *cpu_smt_mask(int cpu)
6025 return topology_thread_cpumask(cpu);
6030 * Topology list, bottom-up.
6032 static struct sched_domain_topology_level default_topology[] = {
6033 #ifdef CONFIG_SCHED_SMT
6034 { sd_init_SIBLING, cpu_smt_mask, },
6036 #ifdef CONFIG_SCHED_MC
6037 { sd_init_MC, cpu_coregroup_mask, },
6039 #ifdef CONFIG_SCHED_BOOK
6040 { sd_init_BOOK, cpu_book_mask, },
6042 { sd_init_CPU, cpu_cpu_mask, },
6046 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6050 static int sched_domains_numa_levels;
6051 static int *sched_domains_numa_distance;
6052 static struct cpumask ***sched_domains_numa_masks;
6053 static int sched_domains_curr_level;
6055 static inline int sd_local_flags(int level)
6057 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6060 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6063 static struct sched_domain *
6064 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6066 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6067 int level = tl->numa_level;
6068 int sd_weight = cpumask_weight(
6069 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6071 *sd = (struct sched_domain){
6072 .min_interval = sd_weight,
6073 .max_interval = 2*sd_weight,
6075 .imbalance_pct = 125,
6076 .cache_nice_tries = 2,
6083 .flags = 1*SD_LOAD_BALANCE
6084 | 1*SD_BALANCE_NEWIDLE
6089 | 0*SD_SHARE_CPUPOWER
6090 | 0*SD_SHARE_PKG_RESOURCES
6092 | 0*SD_PREFER_SIBLING
6093 | sd_local_flags(level)
6095 .last_balance = jiffies,
6096 .balance_interval = sd_weight,
6098 SD_INIT_NAME(sd, NUMA);
6099 sd->private = &tl->data;
6102 * Ugly hack to pass state to sd_numa_mask()...
6104 sched_domains_curr_level = tl->numa_level;
6109 static const struct cpumask *sd_numa_mask(int cpu)
6111 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6114 static void sched_numa_warn(const char *str)
6116 static int done = false;
6124 printk(KERN_WARNING "ERROR: %s\n\n", str);
6126 for (i = 0; i < nr_node_ids; i++) {
6127 printk(KERN_WARNING " ");
6128 for (j = 0; j < nr_node_ids; j++)
6129 printk(KERN_CONT "%02d ", node_distance(i,j));
6130 printk(KERN_CONT "\n");
6132 printk(KERN_WARNING "\n");
6135 static bool find_numa_distance(int distance)
6139 if (distance == node_distance(0, 0))
6142 for (i = 0; i < sched_domains_numa_levels; i++) {
6143 if (sched_domains_numa_distance[i] == distance)
6150 static void sched_init_numa(void)
6152 int next_distance, curr_distance = node_distance(0, 0);
6153 struct sched_domain_topology_level *tl;
6157 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6158 if (!sched_domains_numa_distance)
6162 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6163 * unique distances in the node_distance() table.
6165 * Assumes node_distance(0,j) includes all distances in
6166 * node_distance(i,j) in order to avoid cubic time.
6168 next_distance = curr_distance;
6169 for (i = 0; i < nr_node_ids; i++) {
6170 for (j = 0; j < nr_node_ids; j++) {
6171 for (k = 0; k < nr_node_ids; k++) {
6172 int distance = node_distance(i, k);
6174 if (distance > curr_distance &&
6175 (distance < next_distance ||
6176 next_distance == curr_distance))
6177 next_distance = distance;
6180 * While not a strong assumption it would be nice to know
6181 * about cases where if node A is connected to B, B is not
6182 * equally connected to A.
6184 if (sched_debug() && node_distance(k, i) != distance)
6185 sched_numa_warn("Node-distance not symmetric");
6187 if (sched_debug() && i && !find_numa_distance(distance))
6188 sched_numa_warn("Node-0 not representative");
6190 if (next_distance != curr_distance) {
6191 sched_domains_numa_distance[level++] = next_distance;
6192 sched_domains_numa_levels = level;
6193 curr_distance = next_distance;
6198 * In case of sched_debug() we verify the above assumption.
6204 * 'level' contains the number of unique distances, excluding the
6205 * identity distance node_distance(i,i).
6207 * The sched_domains_numa_distance[] array includes the actual distance
6212 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6213 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6214 * the array will contain less then 'level' members. This could be
6215 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6216 * in other functions.
6218 * We reset it to 'level' at the end of this function.
6220 sched_domains_numa_levels = 0;
6222 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6223 if (!sched_domains_numa_masks)
6227 * Now for each level, construct a mask per node which contains all
6228 * cpus of nodes that are that many hops away from us.
6230 for (i = 0; i < level; i++) {
6231 sched_domains_numa_masks[i] =
6232 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6233 if (!sched_domains_numa_masks[i])
6236 for (j = 0; j < nr_node_ids; j++) {
6237 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6241 sched_domains_numa_masks[i][j] = mask;
6243 for (k = 0; k < nr_node_ids; k++) {
6244 if (node_distance(j, k) > sched_domains_numa_distance[i])
6247 cpumask_or(mask, mask, cpumask_of_node(k));
6252 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6253 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6258 * Copy the default topology bits..
6260 for (i = 0; default_topology[i].init; i++)
6261 tl[i] = default_topology[i];
6264 * .. and append 'j' levels of NUMA goodness.
6266 for (j = 0; j < level; i++, j++) {
6267 tl[i] = (struct sched_domain_topology_level){
6268 .init = sd_numa_init,
6269 .mask = sd_numa_mask,
6270 .flags = SDTL_OVERLAP,
6275 sched_domain_topology = tl;
6277 sched_domains_numa_levels = level;
6280 static void sched_domains_numa_masks_set(int cpu)
6283 int node = cpu_to_node(cpu);
6285 for (i = 0; i < sched_domains_numa_levels; i++) {
6286 for (j = 0; j < nr_node_ids; j++) {
6287 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6288 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6293 static void sched_domains_numa_masks_clear(int cpu)
6296 for (i = 0; i < sched_domains_numa_levels; i++) {
6297 for (j = 0; j < nr_node_ids; j++)
6298 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6303 * Update sched_domains_numa_masks[level][node] array when new cpus
6306 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6307 unsigned long action,
6310 int cpu = (long)hcpu;
6312 switch (action & ~CPU_TASKS_FROZEN) {
6314 sched_domains_numa_masks_set(cpu);
6318 sched_domains_numa_masks_clear(cpu);
6328 static inline void sched_init_numa(void)
6332 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6333 unsigned long action,
6338 #endif /* CONFIG_NUMA */
6340 static int __sdt_alloc(const struct cpumask *cpu_map)
6342 struct sched_domain_topology_level *tl;
6345 for (tl = sched_domain_topology; tl->init; tl++) {
6346 struct sd_data *sdd = &tl->data;
6348 sdd->sd = alloc_percpu(struct sched_domain *);
6352 sdd->sg = alloc_percpu(struct sched_group *);
6356 sdd->sgp = alloc_percpu(struct sched_group_power *);
6360 for_each_cpu(j, cpu_map) {
6361 struct sched_domain *sd;
6362 struct sched_group *sg;
6363 struct sched_group_power *sgp;
6365 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6366 GFP_KERNEL, cpu_to_node(j));
6370 *per_cpu_ptr(sdd->sd, j) = sd;
6372 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6373 GFP_KERNEL, cpu_to_node(j));
6379 *per_cpu_ptr(sdd->sg, j) = sg;
6381 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6382 GFP_KERNEL, cpu_to_node(j));
6386 *per_cpu_ptr(sdd->sgp, j) = sgp;
6393 static void __sdt_free(const struct cpumask *cpu_map)
6395 struct sched_domain_topology_level *tl;
6398 for (tl = sched_domain_topology; tl->init; tl++) {
6399 struct sd_data *sdd = &tl->data;
6401 for_each_cpu(j, cpu_map) {
6402 struct sched_domain *sd;
6405 sd = *per_cpu_ptr(sdd->sd, j);
6406 if (sd && (sd->flags & SD_OVERLAP))
6407 free_sched_groups(sd->groups, 0);
6408 kfree(*per_cpu_ptr(sdd->sd, j));
6412 kfree(*per_cpu_ptr(sdd->sg, j));
6414 kfree(*per_cpu_ptr(sdd->sgp, j));
6416 free_percpu(sdd->sd);
6418 free_percpu(sdd->sg);
6420 free_percpu(sdd->sgp);
6425 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6426 struct s_data *d, const struct cpumask *cpu_map,
6427 struct sched_domain_attr *attr, struct sched_domain *child,
6430 struct sched_domain *sd = tl->init(tl, cpu);
6434 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6436 sd->level = child->level + 1;
6437 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6441 set_domain_attribute(sd, attr);
6447 * Build sched domains for a given set of cpus and attach the sched domains
6448 * to the individual cpus
6450 static int build_sched_domains(const struct cpumask *cpu_map,
6451 struct sched_domain_attr *attr)
6453 enum s_alloc alloc_state = sa_none;
6454 struct sched_domain *sd;
6456 int i, ret = -ENOMEM;
6458 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6459 if (alloc_state != sa_rootdomain)
6462 /* Set up domains for cpus specified by the cpu_map. */
6463 for_each_cpu(i, cpu_map) {
6464 struct sched_domain_topology_level *tl;
6467 for (tl = sched_domain_topology; tl->init; tl++) {
6468 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6469 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6470 sd->flags |= SD_OVERLAP;
6471 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6478 *per_cpu_ptr(d.sd, i) = sd;
6481 /* Build the groups for the domains */
6482 for_each_cpu(i, cpu_map) {
6483 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6484 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6485 if (sd->flags & SD_OVERLAP) {
6486 if (build_overlap_sched_groups(sd, i))
6489 if (build_sched_groups(sd, i))
6495 /* Calculate CPU power for physical packages and nodes */
6496 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6497 if (!cpumask_test_cpu(i, cpu_map))
6500 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6501 claim_allocations(i, sd);
6502 init_sched_groups_power(i, sd);
6506 /* Attach the domains */
6508 for_each_cpu(i, cpu_map) {
6509 sd = *per_cpu_ptr(d.sd, i);
6510 cpu_attach_domain(sd, d.rd, i);
6516 __free_domain_allocs(&d, alloc_state, cpu_map);
6520 static cpumask_var_t *doms_cur; /* current sched domains */
6521 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6522 static struct sched_domain_attr *dattr_cur;
6523 /* attribues of custom domains in 'doms_cur' */
6526 * Special case: If a kmalloc of a doms_cur partition (array of
6527 * cpumask) fails, then fallback to a single sched domain,
6528 * as determined by the single cpumask fallback_doms.
6530 static cpumask_var_t fallback_doms;
6533 * arch_update_cpu_topology lets virtualized architectures update the
6534 * cpu core maps. It is supposed to return 1 if the topology changed
6535 * or 0 if it stayed the same.
6537 int __attribute__((weak)) arch_update_cpu_topology(void)
6542 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6545 cpumask_var_t *doms;
6547 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6550 for (i = 0; i < ndoms; i++) {
6551 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6552 free_sched_domains(doms, i);
6559 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6562 for (i = 0; i < ndoms; i++)
6563 free_cpumask_var(doms[i]);
6568 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6569 * For now this just excludes isolated cpus, but could be used to
6570 * exclude other special cases in the future.
6572 static int init_sched_domains(const struct cpumask *cpu_map)
6576 arch_update_cpu_topology();
6578 doms_cur = alloc_sched_domains(ndoms_cur);
6580 doms_cur = &fallback_doms;
6581 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6582 err = build_sched_domains(doms_cur[0], NULL);
6583 register_sched_domain_sysctl();
6589 * Detach sched domains from a group of cpus specified in cpu_map
6590 * These cpus will now be attached to the NULL domain
6592 static void detach_destroy_domains(const struct cpumask *cpu_map)
6597 for_each_cpu(i, cpu_map)
6598 cpu_attach_domain(NULL, &def_root_domain, i);
6602 /* handle null as "default" */
6603 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6604 struct sched_domain_attr *new, int idx_new)
6606 struct sched_domain_attr tmp;
6613 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6614 new ? (new + idx_new) : &tmp,
6615 sizeof(struct sched_domain_attr));
6619 * Partition sched domains as specified by the 'ndoms_new'
6620 * cpumasks in the array doms_new[] of cpumasks. This compares
6621 * doms_new[] to the current sched domain partitioning, doms_cur[].
6622 * It destroys each deleted domain and builds each new domain.
6624 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6625 * The masks don't intersect (don't overlap.) We should setup one
6626 * sched domain for each mask. CPUs not in any of the cpumasks will
6627 * not be load balanced. If the same cpumask appears both in the
6628 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6631 * The passed in 'doms_new' should be allocated using
6632 * alloc_sched_domains. This routine takes ownership of it and will
6633 * free_sched_domains it when done with it. If the caller failed the
6634 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6635 * and partition_sched_domains() will fallback to the single partition
6636 * 'fallback_doms', it also forces the domains to be rebuilt.
6638 * If doms_new == NULL it will be replaced with cpu_online_mask.
6639 * ndoms_new == 0 is a special case for destroying existing domains,
6640 * and it will not create the default domain.
6642 * Call with hotplug lock held
6644 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6645 struct sched_domain_attr *dattr_new)
6650 mutex_lock(&sched_domains_mutex);
6652 /* always unregister in case we don't destroy any domains */
6653 unregister_sched_domain_sysctl();
6655 /* Let architecture update cpu core mappings. */
6656 new_topology = arch_update_cpu_topology();
6658 n = doms_new ? ndoms_new : 0;
6660 /* Destroy deleted domains */
6661 for (i = 0; i < ndoms_cur; i++) {
6662 for (j = 0; j < n && !new_topology; j++) {
6663 if (cpumask_equal(doms_cur[i], doms_new[j])
6664 && dattrs_equal(dattr_cur, i, dattr_new, j))
6667 /* no match - a current sched domain not in new doms_new[] */
6668 detach_destroy_domains(doms_cur[i]);
6673 if (doms_new == NULL) {
6675 doms_new = &fallback_doms;
6676 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6677 WARN_ON_ONCE(dattr_new);
6680 /* Build new domains */
6681 for (i = 0; i < ndoms_new; i++) {
6682 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6683 if (cpumask_equal(doms_new[i], doms_cur[j])
6684 && dattrs_equal(dattr_new, i, dattr_cur, j))
6687 /* no match - add a new doms_new */
6688 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6693 /* Remember the new sched domains */
6694 if (doms_cur != &fallback_doms)
6695 free_sched_domains(doms_cur, ndoms_cur);
6696 kfree(dattr_cur); /* kfree(NULL) is safe */
6697 doms_cur = doms_new;
6698 dattr_cur = dattr_new;
6699 ndoms_cur = ndoms_new;
6701 register_sched_domain_sysctl();
6703 mutex_unlock(&sched_domains_mutex);
6706 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6709 * Update cpusets according to cpu_active mask. If cpusets are
6710 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6711 * around partition_sched_domains().
6713 * If we come here as part of a suspend/resume, don't touch cpusets because we
6714 * want to restore it back to its original state upon resume anyway.
6716 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6720 case CPU_ONLINE_FROZEN:
6721 case CPU_DOWN_FAILED_FROZEN:
6724 * num_cpus_frozen tracks how many CPUs are involved in suspend
6725 * resume sequence. As long as this is not the last online
6726 * operation in the resume sequence, just build a single sched
6727 * domain, ignoring cpusets.
6730 if (likely(num_cpus_frozen)) {
6731 partition_sched_domains(1, NULL, NULL);
6736 * This is the last CPU online operation. So fall through and
6737 * restore the original sched domains by considering the
6738 * cpuset configurations.
6742 case CPU_DOWN_FAILED:
6743 cpuset_update_active_cpus(true);
6751 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6755 case CPU_DOWN_PREPARE:
6756 cpuset_update_active_cpus(false);
6758 case CPU_DOWN_PREPARE_FROZEN:
6760 partition_sched_domains(1, NULL, NULL);
6768 void __init sched_init_smp(void)
6770 cpumask_var_t non_isolated_cpus;
6772 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6773 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6778 mutex_lock(&sched_domains_mutex);
6779 init_sched_domains(cpu_active_mask);
6780 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6781 if (cpumask_empty(non_isolated_cpus))
6782 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6783 mutex_unlock(&sched_domains_mutex);
6786 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6787 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6788 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6790 /* RT runtime code needs to handle some hotplug events */
6791 hotcpu_notifier(update_runtime, 0);
6795 /* Move init over to a non-isolated CPU */
6796 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6798 sched_init_granularity();
6799 free_cpumask_var(non_isolated_cpus);
6801 init_sched_rt_class();
6804 void __init sched_init_smp(void)
6806 sched_init_granularity();
6808 #endif /* CONFIG_SMP */
6810 const_debug unsigned int sysctl_timer_migration = 1;
6812 int in_sched_functions(unsigned long addr)
6814 return in_lock_functions(addr) ||
6815 (addr >= (unsigned long)__sched_text_start
6816 && addr < (unsigned long)__sched_text_end);
6819 #ifdef CONFIG_CGROUP_SCHED
6821 * Default task group.
6822 * Every task in system belongs to this group at bootup.
6824 struct task_group root_task_group;
6825 LIST_HEAD(task_groups);
6828 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6830 void __init sched_init(void)
6833 unsigned long alloc_size = 0, ptr;
6835 #ifdef CONFIG_FAIR_GROUP_SCHED
6836 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6838 #ifdef CONFIG_RT_GROUP_SCHED
6839 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6841 #ifdef CONFIG_CPUMASK_OFFSTACK
6842 alloc_size += num_possible_cpus() * cpumask_size();
6845 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6847 #ifdef CONFIG_FAIR_GROUP_SCHED
6848 root_task_group.se = (struct sched_entity **)ptr;
6849 ptr += nr_cpu_ids * sizeof(void **);
6851 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6852 ptr += nr_cpu_ids * sizeof(void **);
6854 #endif /* CONFIG_FAIR_GROUP_SCHED */
6855 #ifdef CONFIG_RT_GROUP_SCHED
6856 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6857 ptr += nr_cpu_ids * sizeof(void **);
6859 root_task_group.rt_rq = (struct rt_rq **)ptr;
6860 ptr += nr_cpu_ids * sizeof(void **);
6862 #endif /* CONFIG_RT_GROUP_SCHED */
6863 #ifdef CONFIG_CPUMASK_OFFSTACK
6864 for_each_possible_cpu(i) {
6865 per_cpu(load_balance_mask, i) = (void *)ptr;
6866 ptr += cpumask_size();
6868 #endif /* CONFIG_CPUMASK_OFFSTACK */
6872 init_defrootdomain();
6875 init_rt_bandwidth(&def_rt_bandwidth,
6876 global_rt_period(), global_rt_runtime());
6878 #ifdef CONFIG_RT_GROUP_SCHED
6879 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6880 global_rt_period(), global_rt_runtime());
6881 #endif /* CONFIG_RT_GROUP_SCHED */
6883 #ifdef CONFIG_CGROUP_SCHED
6884 list_add(&root_task_group.list, &task_groups);
6885 INIT_LIST_HEAD(&root_task_group.children);
6886 INIT_LIST_HEAD(&root_task_group.siblings);
6887 autogroup_init(&init_task);
6889 #endif /* CONFIG_CGROUP_SCHED */
6891 for_each_possible_cpu(i) {
6895 raw_spin_lock_init(&rq->lock);
6897 rq->calc_load_active = 0;
6898 rq->calc_load_update = jiffies + LOAD_FREQ;
6899 init_cfs_rq(&rq->cfs);
6900 init_rt_rq(&rq->rt, rq);
6901 #ifdef CONFIG_FAIR_GROUP_SCHED
6902 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6903 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6905 * How much cpu bandwidth does root_task_group get?
6907 * In case of task-groups formed thr' the cgroup filesystem, it
6908 * gets 100% of the cpu resources in the system. This overall
6909 * system cpu resource is divided among the tasks of
6910 * root_task_group and its child task-groups in a fair manner,
6911 * based on each entity's (task or task-group's) weight
6912 * (se->load.weight).
6914 * In other words, if root_task_group has 10 tasks of weight
6915 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6916 * then A0's share of the cpu resource is:
6918 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6920 * We achieve this by letting root_task_group's tasks sit
6921 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6923 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6924 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6925 #endif /* CONFIG_FAIR_GROUP_SCHED */
6927 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6928 #ifdef CONFIG_RT_GROUP_SCHED
6929 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6930 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6933 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6934 rq->cpu_load[j] = 0;
6936 rq->last_load_update_tick = jiffies;
6941 rq->cpu_power = SCHED_POWER_SCALE;
6942 rq->post_schedule = 0;
6943 rq->active_balance = 0;
6944 rq->next_balance = jiffies;
6949 rq->avg_idle = 2*sysctl_sched_migration_cost;
6951 INIT_LIST_HEAD(&rq->cfs_tasks);
6953 rq_attach_root(rq, &def_root_domain);
6959 atomic_set(&rq->nr_iowait, 0);
6962 set_load_weight(&init_task);
6964 #ifdef CONFIG_PREEMPT_NOTIFIERS
6965 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6968 #ifdef CONFIG_RT_MUTEXES
6969 plist_head_init(&init_task.pi_waiters);
6973 * The boot idle thread does lazy MMU switching as well:
6975 atomic_inc(&init_mm.mm_count);
6976 enter_lazy_tlb(&init_mm, current);
6979 * Make us the idle thread. Technically, schedule() should not be
6980 * called from this thread, however somewhere below it might be,
6981 * but because we are the idle thread, we just pick up running again
6982 * when this runqueue becomes "idle".
6984 init_idle(current, smp_processor_id());
6986 calc_load_update = jiffies + LOAD_FREQ;
6989 * During early bootup we pretend to be a normal task:
6991 current->sched_class = &fair_sched_class;
6994 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6995 /* May be allocated at isolcpus cmdline parse time */
6996 if (cpu_isolated_map == NULL)
6997 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6998 idle_thread_set_boot_cpu();
7000 init_sched_fair_class();
7002 scheduler_running = 1;
7005 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7006 static inline int preempt_count_equals(int preempt_offset)
7008 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7010 return (nested == preempt_offset);
7013 void __might_sleep(const char *file, int line, int preempt_offset)
7015 static unsigned long prev_jiffy; /* ratelimiting */
7017 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7018 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7019 system_state != SYSTEM_RUNNING || oops_in_progress)
7021 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7023 prev_jiffy = jiffies;
7026 "BUG: sleeping function called from invalid context at %s:%d\n",
7029 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7030 in_atomic(), irqs_disabled(),
7031 current->pid, current->comm);
7033 debug_show_held_locks(current);
7034 if (irqs_disabled())
7035 print_irqtrace_events(current);
7038 EXPORT_SYMBOL(__might_sleep);
7041 #ifdef CONFIG_MAGIC_SYSRQ
7042 static void normalize_task(struct rq *rq, struct task_struct *p)
7044 const struct sched_class *prev_class = p->sched_class;
7045 int old_prio = p->prio;
7050 dequeue_task(rq, p, 0);
7051 __setscheduler(rq, p, SCHED_NORMAL, 0);
7053 enqueue_task(rq, p, 0);
7054 resched_task(rq->curr);
7057 check_class_changed(rq, p, prev_class, old_prio);
7060 void normalize_rt_tasks(void)
7062 struct task_struct *g, *p;
7063 unsigned long flags;
7066 read_lock_irqsave(&tasklist_lock, flags);
7067 do_each_thread(g, p) {
7069 * Only normalize user tasks:
7074 p->se.exec_start = 0;
7075 #ifdef CONFIG_SCHEDSTATS
7076 p->se.statistics.wait_start = 0;
7077 p->se.statistics.sleep_start = 0;
7078 p->se.statistics.block_start = 0;
7083 * Renice negative nice level userspace
7086 if (TASK_NICE(p) < 0 && p->mm)
7087 set_user_nice(p, 0);
7091 raw_spin_lock(&p->pi_lock);
7092 rq = __task_rq_lock(p);
7094 normalize_task(rq, p);
7096 __task_rq_unlock(rq);
7097 raw_spin_unlock(&p->pi_lock);
7098 } while_each_thread(g, p);
7100 read_unlock_irqrestore(&tasklist_lock, flags);
7103 #endif /* CONFIG_MAGIC_SYSRQ */
7105 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7107 * These functions are only useful for the IA64 MCA handling, or kdb.
7109 * They can only be called when the whole system has been
7110 * stopped - every CPU needs to be quiescent, and no scheduling
7111 * activity can take place. Using them for anything else would
7112 * be a serious bug, and as a result, they aren't even visible
7113 * under any other configuration.
7117 * curr_task - return the current task for a given cpu.
7118 * @cpu: the processor in question.
7120 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7122 struct task_struct *curr_task(int cpu)
7124 return cpu_curr(cpu);
7127 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7131 * set_curr_task - set the current task for a given cpu.
7132 * @cpu: the processor in question.
7133 * @p: the task pointer to set.
7135 * Description: This function must only be used when non-maskable interrupts
7136 * are serviced on a separate stack. It allows the architecture to switch the
7137 * notion of the current task on a cpu in a non-blocking manner. This function
7138 * must be called with all CPU's synchronized, and interrupts disabled, the
7139 * and caller must save the original value of the current task (see
7140 * curr_task() above) and restore that value before reenabling interrupts and
7141 * re-starting the system.
7143 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7145 void set_curr_task(int cpu, struct task_struct *p)
7152 #ifdef CONFIG_CGROUP_SCHED
7153 /* task_group_lock serializes the addition/removal of task groups */
7154 static DEFINE_SPINLOCK(task_group_lock);
7156 static void free_sched_group(struct task_group *tg)
7158 free_fair_sched_group(tg);
7159 free_rt_sched_group(tg);
7164 /* allocate runqueue etc for a new task group */
7165 struct task_group *sched_create_group(struct task_group *parent)
7167 struct task_group *tg;
7169 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7171 return ERR_PTR(-ENOMEM);
7173 if (!alloc_fair_sched_group(tg, parent))
7176 if (!alloc_rt_sched_group(tg, parent))
7182 free_sched_group(tg);
7183 return ERR_PTR(-ENOMEM);
7186 void sched_online_group(struct task_group *tg, struct task_group *parent)
7188 unsigned long flags;
7190 spin_lock_irqsave(&task_group_lock, flags);
7191 list_add_rcu(&tg->list, &task_groups);
7193 WARN_ON(!parent); /* root should already exist */
7195 tg->parent = parent;
7196 INIT_LIST_HEAD(&tg->children);
7197 list_add_rcu(&tg->siblings, &parent->children);
7198 spin_unlock_irqrestore(&task_group_lock, flags);
7201 /* rcu callback to free various structures associated with a task group */
7202 static void free_sched_group_rcu(struct rcu_head *rhp)
7204 /* now it should be safe to free those cfs_rqs */
7205 free_sched_group(container_of(rhp, struct task_group, rcu));
7208 /* Destroy runqueue etc associated with a task group */
7209 void sched_destroy_group(struct task_group *tg)
7211 /* wait for possible concurrent references to cfs_rqs complete */
7212 call_rcu(&tg->rcu, free_sched_group_rcu);
7215 void sched_offline_group(struct task_group *tg)
7217 unsigned long flags;
7220 /* end participation in shares distribution */
7221 for_each_possible_cpu(i)
7222 unregister_fair_sched_group(tg, i);
7224 spin_lock_irqsave(&task_group_lock, flags);
7225 list_del_rcu(&tg->list);
7226 list_del_rcu(&tg->siblings);
7227 spin_unlock_irqrestore(&task_group_lock, flags);
7230 /* change task's runqueue when it moves between groups.
7231 * The caller of this function should have put the task in its new group
7232 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7233 * reflect its new group.
7235 void sched_move_task(struct task_struct *tsk)
7237 struct task_group *tg;
7239 unsigned long flags;
7242 rq = task_rq_lock(tsk, &flags);
7244 running = task_current(rq, tsk);
7248 dequeue_task(rq, tsk, 0);
7249 if (unlikely(running))
7250 tsk->sched_class->put_prev_task(rq, tsk);
7252 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7253 lockdep_is_held(&tsk->sighand->siglock)),
7254 struct task_group, css);
7255 tg = autogroup_task_group(tsk, tg);
7256 tsk->sched_task_group = tg;
7258 #ifdef CONFIG_FAIR_GROUP_SCHED
7259 if (tsk->sched_class->task_move_group)
7260 tsk->sched_class->task_move_group(tsk, on_rq);
7263 set_task_rq(tsk, task_cpu(tsk));
7265 if (unlikely(running))
7266 tsk->sched_class->set_curr_task(rq);
7268 enqueue_task(rq, tsk, 0);
7270 task_rq_unlock(rq, tsk, &flags);
7272 #endif /* CONFIG_CGROUP_SCHED */
7274 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7275 static unsigned long to_ratio(u64 period, u64 runtime)
7277 if (runtime == RUNTIME_INF)
7280 return div64_u64(runtime << 20, period);
7284 #ifdef CONFIG_RT_GROUP_SCHED
7286 * Ensure that the real time constraints are schedulable.
7288 static DEFINE_MUTEX(rt_constraints_mutex);
7290 /* Must be called with tasklist_lock held */
7291 static inline int tg_has_rt_tasks(struct task_group *tg)
7293 struct task_struct *g, *p;
7295 do_each_thread(g, p) {
7296 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7298 } while_each_thread(g, p);
7303 struct rt_schedulable_data {
7304 struct task_group *tg;
7309 static int tg_rt_schedulable(struct task_group *tg, void *data)
7311 struct rt_schedulable_data *d = data;
7312 struct task_group *child;
7313 unsigned long total, sum = 0;
7314 u64 period, runtime;
7316 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7317 runtime = tg->rt_bandwidth.rt_runtime;
7320 period = d->rt_period;
7321 runtime = d->rt_runtime;
7325 * Cannot have more runtime than the period.
7327 if (runtime > period && runtime != RUNTIME_INF)
7331 * Ensure we don't starve existing RT tasks.
7333 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7336 total = to_ratio(period, runtime);
7339 * Nobody can have more than the global setting allows.
7341 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7345 * The sum of our children's runtime should not exceed our own.
7347 list_for_each_entry_rcu(child, &tg->children, siblings) {
7348 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7349 runtime = child->rt_bandwidth.rt_runtime;
7351 if (child == d->tg) {
7352 period = d->rt_period;
7353 runtime = d->rt_runtime;
7356 sum += to_ratio(period, runtime);
7365 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7369 struct rt_schedulable_data data = {
7371 .rt_period = period,
7372 .rt_runtime = runtime,
7376 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7382 static int tg_set_rt_bandwidth(struct task_group *tg,
7383 u64 rt_period, u64 rt_runtime)
7387 mutex_lock(&rt_constraints_mutex);
7388 read_lock(&tasklist_lock);
7389 err = __rt_schedulable(tg, rt_period, rt_runtime);
7393 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7394 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7395 tg->rt_bandwidth.rt_runtime = rt_runtime;
7397 for_each_possible_cpu(i) {
7398 struct rt_rq *rt_rq = tg->rt_rq[i];
7400 raw_spin_lock(&rt_rq->rt_runtime_lock);
7401 rt_rq->rt_runtime = rt_runtime;
7402 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7404 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7406 read_unlock(&tasklist_lock);
7407 mutex_unlock(&rt_constraints_mutex);
7412 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7414 u64 rt_runtime, rt_period;
7416 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7417 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7418 if (rt_runtime_us < 0)
7419 rt_runtime = RUNTIME_INF;
7421 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7424 static long sched_group_rt_runtime(struct task_group *tg)
7428 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7431 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7432 do_div(rt_runtime_us, NSEC_PER_USEC);
7433 return rt_runtime_us;
7436 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7438 u64 rt_runtime, rt_period;
7440 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7441 rt_runtime = tg->rt_bandwidth.rt_runtime;
7446 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7449 static long sched_group_rt_period(struct task_group *tg)
7453 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7454 do_div(rt_period_us, NSEC_PER_USEC);
7455 return rt_period_us;
7458 static int sched_rt_global_constraints(void)
7460 u64 runtime, period;
7463 if (sysctl_sched_rt_period <= 0)
7466 runtime = global_rt_runtime();
7467 period = global_rt_period();
7470 * Sanity check on the sysctl variables.
7472 if (runtime > period && runtime != RUNTIME_INF)
7475 mutex_lock(&rt_constraints_mutex);
7476 read_lock(&tasklist_lock);
7477 ret = __rt_schedulable(NULL, 0, 0);
7478 read_unlock(&tasklist_lock);
7479 mutex_unlock(&rt_constraints_mutex);
7484 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7486 /* Don't accept realtime tasks when there is no way for them to run */
7487 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7493 #else /* !CONFIG_RT_GROUP_SCHED */
7494 static int sched_rt_global_constraints(void)
7496 unsigned long flags;
7499 if (sysctl_sched_rt_period <= 0)
7503 * There's always some RT tasks in the root group
7504 * -- migration, kstopmachine etc..
7506 if (sysctl_sched_rt_runtime == 0)
7509 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7510 for_each_possible_cpu(i) {
7511 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7513 raw_spin_lock(&rt_rq->rt_runtime_lock);
7514 rt_rq->rt_runtime = global_rt_runtime();
7515 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7517 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7521 #endif /* CONFIG_RT_GROUP_SCHED */
7523 int sched_rr_handler(struct ctl_table *table, int write,
7524 void __user *buffer, size_t *lenp,
7528 static DEFINE_MUTEX(mutex);
7531 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7532 /* make sure that internally we keep jiffies */
7533 /* also, writing zero resets timeslice to default */
7534 if (!ret && write) {
7535 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7536 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7538 mutex_unlock(&mutex);
7542 int sched_rt_handler(struct ctl_table *table, int write,
7543 void __user *buffer, size_t *lenp,
7547 int old_period, old_runtime;
7548 static DEFINE_MUTEX(mutex);
7551 old_period = sysctl_sched_rt_period;
7552 old_runtime = sysctl_sched_rt_runtime;
7554 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7556 if (!ret && write) {
7557 ret = sched_rt_global_constraints();
7559 sysctl_sched_rt_period = old_period;
7560 sysctl_sched_rt_runtime = old_runtime;
7562 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7563 def_rt_bandwidth.rt_period =
7564 ns_to_ktime(global_rt_period());
7567 mutex_unlock(&mutex);
7572 #ifdef CONFIG_CGROUP_SCHED
7574 /* return corresponding task_group object of a cgroup */
7575 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7577 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7578 struct task_group, css);
7581 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7583 struct task_group *tg, *parent;
7585 if (!cgrp->parent) {
7586 /* This is early initialization for the top cgroup */
7587 return &root_task_group.css;
7590 parent = cgroup_tg(cgrp->parent);
7591 tg = sched_create_group(parent);
7593 return ERR_PTR(-ENOMEM);
7598 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7600 struct task_group *tg = cgroup_tg(cgrp);
7601 struct task_group *parent;
7606 parent = cgroup_tg(cgrp->parent);
7607 sched_online_group(tg, parent);
7611 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7613 struct task_group *tg = cgroup_tg(cgrp);
7615 sched_destroy_group(tg);
7618 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7620 struct task_group *tg = cgroup_tg(cgrp);
7622 sched_offline_group(tg);
7625 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7626 struct cgroup_taskset *tset)
7628 struct task_struct *task;
7630 cgroup_taskset_for_each(task, cgrp, tset) {
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7635 /* We don't support RT-tasks being in separate groups */
7636 if (task->sched_class != &fair_sched_class)
7643 static void cpu_cgroup_attach(struct cgroup *cgrp,
7644 struct cgroup_taskset *tset)
7646 struct task_struct *task;
7648 cgroup_taskset_for_each(task, cgrp, tset)
7649 sched_move_task(task);
7653 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7654 struct task_struct *task)
7657 * cgroup_exit() is called in the copy_process() failure path.
7658 * Ignore this case since the task hasn't ran yet, this avoids
7659 * trying to poke a half freed task state from generic code.
7661 if (!(task->flags & PF_EXITING))
7664 sched_move_task(task);
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7671 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7674 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7676 struct task_group *tg = cgroup_tg(cgrp);
7678 return (u64) scale_load_down(tg->shares);
7681 #ifdef CONFIG_CFS_BANDWIDTH
7682 static DEFINE_MUTEX(cfs_constraints_mutex);
7684 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7685 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7687 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7689 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7691 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7692 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7694 if (tg == &root_task_group)
7698 * Ensure we have at some amount of bandwidth every period. This is
7699 * to prevent reaching a state of large arrears when throttled via
7700 * entity_tick() resulting in prolonged exit starvation.
7702 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7706 * Likewise, bound things on the otherside by preventing insane quota
7707 * periods. This also allows us to normalize in computing quota
7710 if (period > max_cfs_quota_period)
7713 mutex_lock(&cfs_constraints_mutex);
7714 ret = __cfs_schedulable(tg, period, quota);
7718 runtime_enabled = quota != RUNTIME_INF;
7719 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7720 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7721 raw_spin_lock_irq(&cfs_b->lock);
7722 cfs_b->period = ns_to_ktime(period);
7723 cfs_b->quota = quota;
7725 __refill_cfs_bandwidth_runtime(cfs_b);
7726 /* restart the period timer (if active) to handle new period expiry */
7727 if (runtime_enabled && cfs_b->timer_active) {
7728 /* force a reprogram */
7729 cfs_b->timer_active = 0;
7730 __start_cfs_bandwidth(cfs_b);
7732 raw_spin_unlock_irq(&cfs_b->lock);
7734 for_each_possible_cpu(i) {
7735 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7736 struct rq *rq = cfs_rq->rq;
7738 raw_spin_lock_irq(&rq->lock);
7739 cfs_rq->runtime_enabled = runtime_enabled;
7740 cfs_rq->runtime_remaining = 0;
7742 if (cfs_rq->throttled)
7743 unthrottle_cfs_rq(cfs_rq);
7744 raw_spin_unlock_irq(&rq->lock);
7747 mutex_unlock(&cfs_constraints_mutex);
7752 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7756 period = ktime_to_ns(tg->cfs_bandwidth.period);
7757 if (cfs_quota_us < 0)
7758 quota = RUNTIME_INF;
7760 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7762 return tg_set_cfs_bandwidth(tg, period, quota);
7765 long tg_get_cfs_quota(struct task_group *tg)
7769 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7772 quota_us = tg->cfs_bandwidth.quota;
7773 do_div(quota_us, NSEC_PER_USEC);
7778 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7782 period = (u64)cfs_period_us * NSEC_PER_USEC;
7783 quota = tg->cfs_bandwidth.quota;
7785 return tg_set_cfs_bandwidth(tg, period, quota);
7788 long tg_get_cfs_period(struct task_group *tg)
7792 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7793 do_div(cfs_period_us, NSEC_PER_USEC);
7795 return cfs_period_us;
7798 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7800 return tg_get_cfs_quota(cgroup_tg(cgrp));
7803 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7806 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7809 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7811 return tg_get_cfs_period(cgroup_tg(cgrp));
7814 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7817 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7820 struct cfs_schedulable_data {
7821 struct task_group *tg;
7826 * normalize group quota/period to be quota/max_period
7827 * note: units are usecs
7829 static u64 normalize_cfs_quota(struct task_group *tg,
7830 struct cfs_schedulable_data *d)
7838 period = tg_get_cfs_period(tg);
7839 quota = tg_get_cfs_quota(tg);
7842 /* note: these should typically be equivalent */
7843 if (quota == RUNTIME_INF || quota == -1)
7846 return to_ratio(period, quota);
7849 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7851 struct cfs_schedulable_data *d = data;
7852 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7853 s64 quota = 0, parent_quota = -1;
7856 quota = RUNTIME_INF;
7858 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7860 quota = normalize_cfs_quota(tg, d);
7861 parent_quota = parent_b->hierarchal_quota;
7864 * ensure max(child_quota) <= parent_quota, inherit when no
7867 if (quota == RUNTIME_INF)
7868 quota = parent_quota;
7869 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7872 cfs_b->hierarchal_quota = quota;
7877 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7880 struct cfs_schedulable_data data = {
7886 if (quota != RUNTIME_INF) {
7887 do_div(data.period, NSEC_PER_USEC);
7888 do_div(data.quota, NSEC_PER_USEC);
7892 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7898 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7899 struct cgroup_map_cb *cb)
7901 struct task_group *tg = cgroup_tg(cgrp);
7902 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7904 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7905 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7906 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7910 #endif /* CONFIG_CFS_BANDWIDTH */
7911 #endif /* CONFIG_FAIR_GROUP_SCHED */
7913 #ifdef CONFIG_RT_GROUP_SCHED
7914 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7917 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7920 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7922 return sched_group_rt_runtime(cgroup_tg(cgrp));
7925 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7928 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7931 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7933 return sched_group_rt_period(cgroup_tg(cgrp));
7935 #endif /* CONFIG_RT_GROUP_SCHED */
7937 static struct cftype cpu_files[] = {
7938 #ifdef CONFIG_FAIR_GROUP_SCHED
7941 .read_u64 = cpu_shares_read_u64,
7942 .write_u64 = cpu_shares_write_u64,
7945 #ifdef CONFIG_CFS_BANDWIDTH
7947 .name = "cfs_quota_us",
7948 .read_s64 = cpu_cfs_quota_read_s64,
7949 .write_s64 = cpu_cfs_quota_write_s64,
7952 .name = "cfs_period_us",
7953 .read_u64 = cpu_cfs_period_read_u64,
7954 .write_u64 = cpu_cfs_period_write_u64,
7958 .read_map = cpu_stats_show,
7961 #ifdef CONFIG_RT_GROUP_SCHED
7963 .name = "rt_runtime_us",
7964 .read_s64 = cpu_rt_runtime_read,
7965 .write_s64 = cpu_rt_runtime_write,
7968 .name = "rt_period_us",
7969 .read_u64 = cpu_rt_period_read_uint,
7970 .write_u64 = cpu_rt_period_write_uint,
7976 struct cgroup_subsys cpu_cgroup_subsys = {
7978 .css_alloc = cpu_cgroup_css_alloc,
7979 .css_free = cpu_cgroup_css_free,
7980 .css_online = cpu_cgroup_css_online,
7981 .css_offline = cpu_cgroup_css_offline,
7982 .can_attach = cpu_cgroup_can_attach,
7983 .attach = cpu_cgroup_attach,
7984 .exit = cpu_cgroup_exit,
7985 .subsys_id = cpu_cgroup_subsys_id,
7986 .base_cftypes = cpu_files,
7990 #endif /* CONFIG_CGROUP_SCHED */
7992 void dump_cpu_task(int cpu)
7994 pr_info("Task dump for CPU %d:\n", cpu);
7995 sched_show_task(cpu_curr(cpu));