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
516 #ifndef tsk_is_polling
517 #define tsk_is_polling(t) 0
520 void resched_task(struct task_struct *p)
524 assert_raw_spin_locked(&task_rq(p)->lock);
526 if (test_tsk_need_resched(p))
529 set_tsk_need_resched(p);
532 if (cpu == smp_processor_id())
535 /* NEED_RESCHED must be visible before we test polling */
537 if (!tsk_is_polling(p))
538 smp_send_reschedule(cpu);
541 void resched_cpu(int cpu)
543 struct rq *rq = cpu_rq(cpu);
546 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
548 resched_task(cpu_curr(cpu));
549 raw_spin_unlock_irqrestore(&rq->lock, flags);
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
561 int get_nohz_timer_target(void)
563 int cpu = smp_processor_id();
565 struct sched_domain *sd;
568 for_each_domain(cpu, sd) {
569 for_each_cpu(i, sched_domain_span(sd)) {
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 void wake_up_idle_cpu(int cpu)
592 struct rq *rq = cpu_rq(cpu);
594 if (cpu == smp_processor_id())
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
604 if (rq->curr != rq->idle)
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
612 set_tsk_need_resched(rq->idle);
614 /* NEED_RESCHED must be visible before we test polling */
616 if (!tsk_is_polling(rq->idle))
617 smp_send_reschedule(cpu);
620 static inline bool got_nohz_idle_kick(void)
622 int cpu = smp_processor_id();
623 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
626 #else /* CONFIG_NO_HZ */
628 static inline bool got_nohz_idle_kick(void)
633 #endif /* CONFIG_NO_HZ */
635 void sched_avg_update(struct rq *rq)
637 s64 period = sched_avg_period();
639 while ((s64)(rq->clock - rq->age_stamp) > period) {
641 * Inline assembly required to prevent the compiler
642 * optimising this loop into a divmod call.
643 * See __iter_div_u64_rem() for another example of this.
645 asm("" : "+rm" (rq->age_stamp));
646 rq->age_stamp += period;
651 #else /* !CONFIG_SMP */
652 void resched_task(struct task_struct *p)
654 assert_raw_spin_locked(&task_rq(p)->lock);
655 set_tsk_need_resched(p);
657 #endif /* CONFIG_SMP */
659 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
662 * Iterate task_group tree rooted at *from, calling @down when first entering a
663 * node and @up when leaving it for the final time.
665 * Caller must hold rcu_lock or sufficient equivalent.
667 int walk_tg_tree_from(struct task_group *from,
668 tg_visitor down, tg_visitor up, void *data)
670 struct task_group *parent, *child;
676 ret = (*down)(parent, data);
679 list_for_each_entry_rcu(child, &parent->children, siblings) {
686 ret = (*up)(parent, data);
687 if (ret || parent == from)
691 parent = parent->parent;
698 int tg_nop(struct task_group *tg, void *data)
704 static void set_load_weight(struct task_struct *p)
706 int prio = p->static_prio - MAX_RT_PRIO;
707 struct load_weight *load = &p->se.load;
710 * SCHED_IDLE tasks get minimal weight:
712 if (p->policy == SCHED_IDLE) {
713 load->weight = scale_load(WEIGHT_IDLEPRIO);
714 load->inv_weight = WMULT_IDLEPRIO;
718 load->weight = scale_load(prio_to_weight[prio]);
719 load->inv_weight = prio_to_wmult[prio];
722 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
725 sched_info_queued(p);
726 p->sched_class->enqueue_task(rq, p, flags);
729 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
732 sched_info_dequeued(p);
733 p->sched_class->dequeue_task(rq, p, flags);
736 void activate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible--;
741 enqueue_task(rq, p, flags);
744 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
746 if (task_contributes_to_load(p))
747 rq->nr_uninterruptible++;
749 dequeue_task(rq, p, flags);
752 static void update_rq_clock_task(struct rq *rq, s64 delta)
755 * In theory, the compile should just see 0 here, and optimize out the call
756 * to sched_rt_avg_update. But I don't trust it...
758 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
759 s64 steal = 0, irq_delta = 0;
761 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
762 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
765 * Since irq_time is only updated on {soft,}irq_exit, we might run into
766 * this case when a previous update_rq_clock() happened inside a
769 * When this happens, we stop ->clock_task and only update the
770 * prev_irq_time stamp to account for the part that fit, so that a next
771 * update will consume the rest. This ensures ->clock_task is
774 * It does however cause some slight miss-attribution of {soft,}irq
775 * time, a more accurate solution would be to update the irq_time using
776 * the current rq->clock timestamp, except that would require using
779 if (irq_delta > delta)
782 rq->prev_irq_time += irq_delta;
785 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
786 if (static_key_false((¶virt_steal_rq_enabled))) {
789 steal = paravirt_steal_clock(cpu_of(rq));
790 steal -= rq->prev_steal_time_rq;
792 if (unlikely(steal > delta))
795 st = steal_ticks(steal);
796 steal = st * TICK_NSEC;
798 rq->prev_steal_time_rq += steal;
804 rq->clock_task += delta;
806 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
807 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
808 sched_rt_avg_update(rq, irq_delta + steal);
812 void sched_set_stop_task(int cpu, struct task_struct *stop)
814 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
815 struct task_struct *old_stop = cpu_rq(cpu)->stop;
819 * Make it appear like a SCHED_FIFO task, its something
820 * userspace knows about and won't get confused about.
822 * Also, it will make PI more or less work without too
823 * much confusion -- but then, stop work should not
824 * rely on PI working anyway.
826 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
828 stop->sched_class = &stop_sched_class;
831 cpu_rq(cpu)->stop = stop;
835 * Reset it back to a normal scheduling class so that
836 * it can die in pieces.
838 old_stop->sched_class = &rt_sched_class;
843 * __normal_prio - return the priority that is based on the static prio
845 static inline int __normal_prio(struct task_struct *p)
847 return p->static_prio;
851 * Calculate the expected normal priority: i.e. priority
852 * without taking RT-inheritance into account. Might be
853 * boosted by interactivity modifiers. Changes upon fork,
854 * setprio syscalls, and whenever the interactivity
855 * estimator recalculates.
857 static inline int normal_prio(struct task_struct *p)
861 if (task_has_rt_policy(p))
862 prio = MAX_RT_PRIO-1 - p->rt_priority;
864 prio = __normal_prio(p);
869 * Calculate the current priority, i.e. the priority
870 * taken into account by the scheduler. This value might
871 * be boosted by RT tasks, or might be boosted by
872 * interactivity modifiers. Will be RT if the task got
873 * RT-boosted. If not then it returns p->normal_prio.
875 static int effective_prio(struct task_struct *p)
877 p->normal_prio = normal_prio(p);
879 * If we are RT tasks or we were boosted to RT priority,
880 * keep the priority unchanged. Otherwise, update priority
881 * to the normal priority:
883 if (!rt_prio(p->prio))
884 return p->normal_prio;
889 * task_curr - is this task currently executing on a CPU?
890 * @p: the task in question.
892 inline int task_curr(const struct task_struct *p)
894 return cpu_curr(task_cpu(p)) == p;
897 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
898 const struct sched_class *prev_class,
901 if (prev_class != p->sched_class) {
902 if (prev_class->switched_from)
903 prev_class->switched_from(rq, p);
904 p->sched_class->switched_to(rq, p);
905 } else if (oldprio != p->prio)
906 p->sched_class->prio_changed(rq, p, oldprio);
909 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
911 const struct sched_class *class;
913 if (p->sched_class == rq->curr->sched_class) {
914 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
916 for_each_class(class) {
917 if (class == rq->curr->sched_class)
919 if (class == p->sched_class) {
920 resched_task(rq->curr);
927 * A queue event has occurred, and we're going to schedule. In
928 * this case, we can save a useless back to back clock update.
930 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
931 rq->skip_clock_update = 1;
934 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
936 void register_task_migration_notifier(struct notifier_block *n)
938 atomic_notifier_chain_register(&task_migration_notifier, n);
942 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
944 #ifdef CONFIG_SCHED_DEBUG
946 * We should never call set_task_cpu() on a blocked task,
947 * ttwu() will sort out the placement.
949 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
950 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
952 #ifdef CONFIG_LOCKDEP
954 * The caller should hold either p->pi_lock or rq->lock, when changing
955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
957 * sched_move_task() holds both and thus holding either pins the cgroup,
960 * Furthermore, all task_rq users should acquire both locks, see
963 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
964 lockdep_is_held(&task_rq(p)->lock)));
968 trace_sched_migrate_task(p, new_cpu);
970 if (task_cpu(p) != new_cpu) {
971 struct task_migration_notifier tmn;
973 if (p->sched_class->migrate_task_rq)
974 p->sched_class->migrate_task_rq(p, new_cpu);
975 p->se.nr_migrations++;
976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
979 tmn.from_cpu = task_cpu(p);
980 tmn.to_cpu = new_cpu;
982 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
985 __set_task_cpu(p, new_cpu);
988 struct migration_arg {
989 struct task_struct *task;
993 static int migration_cpu_stop(void *data);
996 * wait_task_inactive - wait for a thread to unschedule.
998 * If @match_state is nonzero, it's the @p->state value just checked and
999 * not expected to change. If it changes, i.e. @p might have woken up,
1000 * then return zero. When we succeed in waiting for @p to be off its CPU,
1001 * we return a positive number (its total switch count). If a second call
1002 * a short while later returns the same number, the caller can be sure that
1003 * @p has remained unscheduled the whole time.
1005 * The caller must ensure that the task *will* unschedule sometime soon,
1006 * else this function might spin for a *long* time. This function can't
1007 * be called with interrupts off, or it may introduce deadlock with
1008 * smp_call_function() if an IPI is sent by the same process we are
1009 * waiting to become inactive.
1011 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1013 unsigned long flags;
1020 * We do the initial early heuristics without holding
1021 * any task-queue locks at all. We'll only try to get
1022 * the runqueue lock when things look like they will
1028 * If the task is actively running on another CPU
1029 * still, just relax and busy-wait without holding
1032 * NOTE! Since we don't hold any locks, it's not
1033 * even sure that "rq" stays as the right runqueue!
1034 * But we don't care, since "task_running()" will
1035 * return false if the runqueue has changed and p
1036 * is actually now running somewhere else!
1038 while (task_running(rq, p)) {
1039 if (match_state && unlikely(p->state != match_state))
1045 * Ok, time to look more closely! We need the rq
1046 * lock now, to be *sure*. If we're wrong, we'll
1047 * just go back and repeat.
1049 rq = task_rq_lock(p, &flags);
1050 trace_sched_wait_task(p);
1051 running = task_running(rq, p);
1054 if (!match_state || p->state == match_state)
1055 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1056 task_rq_unlock(rq, p, &flags);
1059 * If it changed from the expected state, bail out now.
1061 if (unlikely(!ncsw))
1065 * Was it really running after all now that we
1066 * checked with the proper locks actually held?
1068 * Oops. Go back and try again..
1070 if (unlikely(running)) {
1076 * It's not enough that it's not actively running,
1077 * it must be off the runqueue _entirely_, and not
1080 * So if it was still runnable (but just not actively
1081 * running right now), it's preempted, and we should
1082 * yield - it could be a while.
1084 if (unlikely(on_rq)) {
1085 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1087 set_current_state(TASK_UNINTERRUPTIBLE);
1088 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1093 * Ahh, all good. It wasn't running, and it wasn't
1094 * runnable, which means that it will never become
1095 * running in the future either. We're all done!
1104 * kick_process - kick a running thread to enter/exit the kernel
1105 * @p: the to-be-kicked thread
1107 * Cause a process which is running on another CPU to enter
1108 * kernel-mode, without any delay. (to get signals handled.)
1110 * NOTE: this function doesn't have to take the runqueue lock,
1111 * because all it wants to ensure is that the remote task enters
1112 * the kernel. If the IPI races and the task has been migrated
1113 * to another CPU then no harm is done and the purpose has been
1116 void kick_process(struct task_struct *p)
1122 if ((cpu != smp_processor_id()) && task_curr(p))
1123 smp_send_reschedule(cpu);
1126 EXPORT_SYMBOL_GPL(kick_process);
1127 #endif /* CONFIG_SMP */
1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1133 static int select_fallback_rq(int cpu, struct task_struct *p)
1135 int nid = cpu_to_node(cpu);
1136 const struct cpumask *nodemask = NULL;
1137 enum { cpuset, possible, fail } state = cpuset;
1141 * If the node that the cpu is on has been offlined, cpu_to_node()
1142 * will return -1. There is no cpu on the node, and we should
1143 * select the cpu on the other node.
1146 nodemask = cpumask_of_node(nid);
1148 /* Look for allowed, online CPU in same node. */
1149 for_each_cpu(dest_cpu, nodemask) {
1150 if (!cpu_online(dest_cpu))
1152 if (!cpu_active(dest_cpu))
1154 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1160 /* Any allowed, online CPU? */
1161 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1162 if (!cpu_online(dest_cpu))
1164 if (!cpu_active(dest_cpu))
1171 /* No more Mr. Nice Guy. */
1172 cpuset_cpus_allowed_fallback(p);
1177 do_set_cpus_allowed(p, cpu_possible_mask);
1188 if (state != cpuset) {
1190 * Don't tell them about moving exiting tasks or
1191 * kernel threads (both mm NULL), since they never
1194 if (p->mm && printk_ratelimit()) {
1195 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1196 task_pid_nr(p), p->comm, cpu);
1204 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1207 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1209 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1212 * In order not to call set_task_cpu() on a blocking task we need
1213 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1216 * Since this is common to all placement strategies, this lives here.
1218 * [ this allows ->select_task() to simply return task_cpu(p) and
1219 * not worry about this generic constraint ]
1221 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1223 cpu = select_fallback_rq(task_cpu(p), p);
1228 static void update_avg(u64 *avg, u64 sample)
1230 s64 diff = sample - *avg;
1236 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1238 #ifdef CONFIG_SCHEDSTATS
1239 struct rq *rq = this_rq();
1242 int this_cpu = smp_processor_id();
1244 if (cpu == this_cpu) {
1245 schedstat_inc(rq, ttwu_local);
1246 schedstat_inc(p, se.statistics.nr_wakeups_local);
1248 struct sched_domain *sd;
1250 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1252 for_each_domain(this_cpu, sd) {
1253 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1254 schedstat_inc(sd, ttwu_wake_remote);
1261 if (wake_flags & WF_MIGRATED)
1262 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1264 #endif /* CONFIG_SMP */
1266 schedstat_inc(rq, ttwu_count);
1267 schedstat_inc(p, se.statistics.nr_wakeups);
1269 if (wake_flags & WF_SYNC)
1270 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1272 #endif /* CONFIG_SCHEDSTATS */
1275 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1277 activate_task(rq, p, en_flags);
1280 /* if a worker is waking up, notify workqueue */
1281 if (p->flags & PF_WQ_WORKER)
1282 wq_worker_waking_up(p, cpu_of(rq));
1286 * Mark the task runnable and perform wakeup-preemption.
1289 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1291 trace_sched_wakeup(p, true);
1292 check_preempt_curr(rq, p, wake_flags);
1294 p->state = TASK_RUNNING;
1296 if (p->sched_class->task_woken)
1297 p->sched_class->task_woken(rq, p);
1299 if (rq->idle_stamp) {
1300 u64 delta = rq->clock - rq->idle_stamp;
1301 u64 max = 2*sysctl_sched_migration_cost;
1306 update_avg(&rq->avg_idle, delta);
1313 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1316 if (p->sched_contributes_to_load)
1317 rq->nr_uninterruptible--;
1320 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1321 ttwu_do_wakeup(rq, p, wake_flags);
1325 * Called in case the task @p isn't fully descheduled from its runqueue,
1326 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1327 * since all we need to do is flip p->state to TASK_RUNNING, since
1328 * the task is still ->on_rq.
1330 static int ttwu_remote(struct task_struct *p, int wake_flags)
1335 rq = __task_rq_lock(p);
1337 ttwu_do_wakeup(rq, p, wake_flags);
1340 __task_rq_unlock(rq);
1346 static void sched_ttwu_pending(void)
1348 struct rq *rq = this_rq();
1349 struct llist_node *llist = llist_del_all(&rq->wake_list);
1350 struct task_struct *p;
1352 raw_spin_lock(&rq->lock);
1355 p = llist_entry(llist, struct task_struct, wake_entry);
1356 llist = llist_next(llist);
1357 ttwu_do_activate(rq, p, 0);
1360 raw_spin_unlock(&rq->lock);
1363 void scheduler_ipi(void)
1365 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1369 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1370 * traditionally all their work was done from the interrupt return
1371 * path. Now that we actually do some work, we need to make sure
1374 * Some archs already do call them, luckily irq_enter/exit nest
1377 * Arguably we should visit all archs and update all handlers,
1378 * however a fair share of IPIs are still resched only so this would
1379 * somewhat pessimize the simple resched case.
1382 sched_ttwu_pending();
1385 * Check if someone kicked us for doing the nohz idle load balance.
1387 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1388 this_rq()->idle_balance = 1;
1389 raise_softirq_irqoff(SCHED_SOFTIRQ);
1394 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1396 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1397 smp_send_reschedule(cpu);
1400 bool cpus_share_cache(int this_cpu, int that_cpu)
1402 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1404 #endif /* CONFIG_SMP */
1406 static void ttwu_queue(struct task_struct *p, int cpu)
1408 struct rq *rq = cpu_rq(cpu);
1410 #if defined(CONFIG_SMP)
1411 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1412 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1413 ttwu_queue_remote(p, cpu);
1418 raw_spin_lock(&rq->lock);
1419 ttwu_do_activate(rq, p, 0);
1420 raw_spin_unlock(&rq->lock);
1424 * try_to_wake_up - wake up a thread
1425 * @p: the thread to be awakened
1426 * @state: the mask of task states that can be woken
1427 * @wake_flags: wake modifier flags (WF_*)
1429 * Put it on the run-queue if it's not already there. The "current"
1430 * thread is always on the run-queue (except when the actual
1431 * re-schedule is in progress), and as such you're allowed to do
1432 * the simpler "current->state = TASK_RUNNING" to mark yourself
1433 * runnable without the overhead of this.
1435 * Returns %true if @p was woken up, %false if it was already running
1436 * or @state didn't match @p's state.
1439 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1441 unsigned long flags;
1442 int cpu, success = 0;
1445 raw_spin_lock_irqsave(&p->pi_lock, flags);
1446 if (!(p->state & state))
1449 success = 1; /* we're going to change ->state */
1452 if (p->on_rq && ttwu_remote(p, wake_flags))
1457 * If the owning (remote) cpu is still in the middle of schedule() with
1458 * this task as prev, wait until its done referencing the task.
1463 * Pairs with the smp_wmb() in finish_lock_switch().
1467 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1468 p->state = TASK_WAKING;
1470 if (p->sched_class->task_waking)
1471 p->sched_class->task_waking(p);
1473 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1474 if (task_cpu(p) != cpu) {
1475 wake_flags |= WF_MIGRATED;
1476 set_task_cpu(p, cpu);
1478 #endif /* CONFIG_SMP */
1482 ttwu_stat(p, cpu, wake_flags);
1484 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1490 * try_to_wake_up_local - try to wake up a local task with rq lock held
1491 * @p: the thread to be awakened
1493 * Put @p on the run-queue if it's not already there. The caller must
1494 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1497 static void try_to_wake_up_local(struct task_struct *p)
1499 struct rq *rq = task_rq(p);
1501 BUG_ON(rq != this_rq());
1502 BUG_ON(p == current);
1503 lockdep_assert_held(&rq->lock);
1505 if (!raw_spin_trylock(&p->pi_lock)) {
1506 raw_spin_unlock(&rq->lock);
1507 raw_spin_lock(&p->pi_lock);
1508 raw_spin_lock(&rq->lock);
1511 if (!(p->state & TASK_NORMAL))
1515 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1517 ttwu_do_wakeup(rq, p, 0);
1518 ttwu_stat(p, smp_processor_id(), 0);
1520 raw_spin_unlock(&p->pi_lock);
1524 * wake_up_process - Wake up a specific process
1525 * @p: The process to be woken up.
1527 * Attempt to wake up the nominated process and move it to the set of runnable
1528 * processes. Returns 1 if the process was woken up, 0 if it was already
1531 * It may be assumed that this function implies a write memory barrier before
1532 * changing the task state if and only if any tasks are woken up.
1534 int wake_up_process(struct task_struct *p)
1536 WARN_ON(task_is_stopped_or_traced(p));
1537 return try_to_wake_up(p, TASK_NORMAL, 0);
1539 EXPORT_SYMBOL(wake_up_process);
1541 int wake_up_state(struct task_struct *p, unsigned int state)
1543 return try_to_wake_up(p, state, 0);
1547 * Perform scheduler related setup for a newly forked process p.
1548 * p is forked by current.
1550 * __sched_fork() is basic setup used by init_idle() too:
1552 static void __sched_fork(struct task_struct *p)
1557 p->se.exec_start = 0;
1558 p->se.sum_exec_runtime = 0;
1559 p->se.prev_sum_exec_runtime = 0;
1560 p->se.nr_migrations = 0;
1562 INIT_LIST_HEAD(&p->se.group_node);
1565 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1566 * removed when useful for applications beyond shares distribution (e.g.
1569 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1570 p->se.avg.runnable_avg_period = 0;
1571 p->se.avg.runnable_avg_sum = 0;
1573 #ifdef CONFIG_SCHEDSTATS
1574 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1577 INIT_LIST_HEAD(&p->rt.run_list);
1579 #ifdef CONFIG_PREEMPT_NOTIFIERS
1580 INIT_HLIST_HEAD(&p->preempt_notifiers);
1583 #ifdef CONFIG_NUMA_BALANCING
1584 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1585 p->mm->numa_next_scan = jiffies;
1586 p->mm->numa_next_reset = jiffies;
1587 p->mm->numa_scan_seq = 0;
1590 p->node_stamp = 0ULL;
1591 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1592 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1593 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1594 p->numa_work.next = &p->numa_work;
1595 #endif /* CONFIG_NUMA_BALANCING */
1598 #ifdef CONFIG_NUMA_BALANCING
1599 #ifdef CONFIG_SCHED_DEBUG
1600 void set_numabalancing_state(bool enabled)
1603 sched_feat_set("NUMA");
1605 sched_feat_set("NO_NUMA");
1608 __read_mostly bool numabalancing_enabled;
1610 void set_numabalancing_state(bool enabled)
1612 numabalancing_enabled = enabled;
1614 #endif /* CONFIG_SCHED_DEBUG */
1615 #endif /* CONFIG_NUMA_BALANCING */
1618 * fork()/clone()-time setup:
1620 void sched_fork(struct task_struct *p)
1622 unsigned long flags;
1623 int cpu = get_cpu();
1627 * We mark the process as running here. This guarantees that
1628 * nobody will actually run it, and a signal or other external
1629 * event cannot wake it up and insert it on the runqueue either.
1631 p->state = TASK_RUNNING;
1634 * Make sure we do not leak PI boosting priority to the child.
1636 p->prio = current->normal_prio;
1639 * Revert to default priority/policy on fork if requested.
1641 if (unlikely(p->sched_reset_on_fork)) {
1642 if (task_has_rt_policy(p)) {
1643 p->policy = SCHED_NORMAL;
1644 p->static_prio = NICE_TO_PRIO(0);
1646 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1647 p->static_prio = NICE_TO_PRIO(0);
1649 p->prio = p->normal_prio = __normal_prio(p);
1653 * We don't need the reset flag anymore after the fork. It has
1654 * fulfilled its duty:
1656 p->sched_reset_on_fork = 0;
1659 if (!rt_prio(p->prio))
1660 p->sched_class = &fair_sched_class;
1662 if (p->sched_class->task_fork)
1663 p->sched_class->task_fork(p);
1666 * The child is not yet in the pid-hash so no cgroup attach races,
1667 * and the cgroup is pinned to this child due to cgroup_fork()
1668 * is ran before sched_fork().
1670 * Silence PROVE_RCU.
1672 raw_spin_lock_irqsave(&p->pi_lock, flags);
1673 set_task_cpu(p, cpu);
1674 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1676 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1677 if (likely(sched_info_on()))
1678 memset(&p->sched_info, 0, sizeof(p->sched_info));
1680 #if defined(CONFIG_SMP)
1683 #ifdef CONFIG_PREEMPT_COUNT
1684 /* Want to start with kernel preemption disabled. */
1685 task_thread_info(p)->preempt_count = 1;
1688 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1695 * wake_up_new_task - wake up a newly created task for the first time.
1697 * This function will do some initial scheduler statistics housekeeping
1698 * that must be done for every newly created context, then puts the task
1699 * on the runqueue and wakes it.
1701 void wake_up_new_task(struct task_struct *p)
1703 unsigned long flags;
1706 raw_spin_lock_irqsave(&p->pi_lock, flags);
1709 * Fork balancing, do it here and not earlier because:
1710 * - cpus_allowed can change in the fork path
1711 * - any previously selected cpu might disappear through hotplug
1713 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1716 rq = __task_rq_lock(p);
1717 activate_task(rq, p, 0);
1719 trace_sched_wakeup_new(p, true);
1720 check_preempt_curr(rq, p, WF_FORK);
1722 if (p->sched_class->task_woken)
1723 p->sched_class->task_woken(rq, p);
1725 task_rq_unlock(rq, p, &flags);
1728 #ifdef CONFIG_PREEMPT_NOTIFIERS
1731 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1732 * @notifier: notifier struct to register
1734 void preempt_notifier_register(struct preempt_notifier *notifier)
1736 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1738 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1741 * preempt_notifier_unregister - no longer interested in preemption notifications
1742 * @notifier: notifier struct to unregister
1744 * This is safe to call from within a preemption notifier.
1746 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1748 hlist_del(¬ifier->link);
1750 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1752 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1754 struct preempt_notifier *notifier;
1755 struct hlist_node *node;
1757 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1758 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1762 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1763 struct task_struct *next)
1765 struct preempt_notifier *notifier;
1766 struct hlist_node *node;
1768 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1769 notifier->ops->sched_out(notifier, next);
1772 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1774 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1779 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1780 struct task_struct *next)
1784 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1787 * prepare_task_switch - prepare to switch tasks
1788 * @rq: the runqueue preparing to switch
1789 * @prev: the current task that is being switched out
1790 * @next: the task we are going to switch to.
1792 * This is called with the rq lock held and interrupts off. It must
1793 * be paired with a subsequent finish_task_switch after the context
1796 * prepare_task_switch sets up locking and calls architecture specific
1800 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1801 struct task_struct *next)
1803 trace_sched_switch(prev, next);
1804 sched_info_switch(prev, next);
1805 perf_event_task_sched_out(prev, next);
1806 fire_sched_out_preempt_notifiers(prev, next);
1807 prepare_lock_switch(rq, next);
1808 prepare_arch_switch(next);
1812 * finish_task_switch - clean up after a task-switch
1813 * @rq: runqueue associated with task-switch
1814 * @prev: the thread we just switched away from.
1816 * finish_task_switch must be called after the context switch, paired
1817 * with a prepare_task_switch call before the context switch.
1818 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1819 * and do any other architecture-specific cleanup actions.
1821 * Note that we may have delayed dropping an mm in context_switch(). If
1822 * so, we finish that here outside of the runqueue lock. (Doing it
1823 * with the lock held can cause deadlocks; see schedule() for
1826 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1827 __releases(rq->lock)
1829 struct mm_struct *mm = rq->prev_mm;
1835 * A task struct has one reference for the use as "current".
1836 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1837 * schedule one last time. The schedule call will never return, and
1838 * the scheduled task must drop that reference.
1839 * The test for TASK_DEAD must occur while the runqueue locks are
1840 * still held, otherwise prev could be scheduled on another cpu, die
1841 * there before we look at prev->state, and then the reference would
1843 * Manfred Spraul <manfred@colorfullife.com>
1845 prev_state = prev->state;
1846 vtime_task_switch(prev);
1847 finish_arch_switch(prev);
1848 perf_event_task_sched_in(prev, current);
1849 finish_lock_switch(rq, prev);
1850 finish_arch_post_lock_switch();
1852 fire_sched_in_preempt_notifiers(current);
1855 if (unlikely(prev_state == TASK_DEAD)) {
1857 * Remove function-return probe instances associated with this
1858 * task and put them back on the free list.
1860 kprobe_flush_task(prev);
1861 put_task_struct(prev);
1867 /* assumes rq->lock is held */
1868 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1870 if (prev->sched_class->pre_schedule)
1871 prev->sched_class->pre_schedule(rq, prev);
1874 /* rq->lock is NOT held, but preemption is disabled */
1875 static inline void post_schedule(struct rq *rq)
1877 if (rq->post_schedule) {
1878 unsigned long flags;
1880 raw_spin_lock_irqsave(&rq->lock, flags);
1881 if (rq->curr->sched_class->post_schedule)
1882 rq->curr->sched_class->post_schedule(rq);
1883 raw_spin_unlock_irqrestore(&rq->lock, flags);
1885 rq->post_schedule = 0;
1891 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1895 static inline void post_schedule(struct rq *rq)
1902 * schedule_tail - first thing a freshly forked thread must call.
1903 * @prev: the thread we just switched away from.
1905 asmlinkage void schedule_tail(struct task_struct *prev)
1906 __releases(rq->lock)
1908 struct rq *rq = this_rq();
1910 finish_task_switch(rq, prev);
1913 * FIXME: do we need to worry about rq being invalidated by the
1918 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1919 /* In this case, finish_task_switch does not reenable preemption */
1922 if (current->set_child_tid)
1923 put_user(task_pid_vnr(current), current->set_child_tid);
1927 * context_switch - switch to the new MM and the new
1928 * thread's register state.
1931 context_switch(struct rq *rq, struct task_struct *prev,
1932 struct task_struct *next)
1934 struct mm_struct *mm, *oldmm;
1936 prepare_task_switch(rq, prev, next);
1939 oldmm = prev->active_mm;
1941 * For paravirt, this is coupled with an exit in switch_to to
1942 * combine the page table reload and the switch backend into
1945 arch_start_context_switch(prev);
1948 next->active_mm = oldmm;
1949 atomic_inc(&oldmm->mm_count);
1950 enter_lazy_tlb(oldmm, next);
1952 switch_mm(oldmm, mm, next);
1955 prev->active_mm = NULL;
1956 rq->prev_mm = oldmm;
1959 * Since the runqueue lock will be released by the next
1960 * task (which is an invalid locking op but in the case
1961 * of the scheduler it's an obvious special-case), so we
1962 * do an early lockdep release here:
1964 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1965 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1968 context_tracking_task_switch(prev, next);
1969 /* Here we just switch the register state and the stack. */
1970 switch_to(prev, next, prev);
1974 * this_rq must be evaluated again because prev may have moved
1975 * CPUs since it called schedule(), thus the 'rq' on its stack
1976 * frame will be invalid.
1978 finish_task_switch(this_rq(), prev);
1982 * nr_running and nr_context_switches:
1984 * externally visible scheduler statistics: current number of runnable
1985 * threads, total number of context switches performed since bootup.
1987 unsigned long nr_running(void)
1989 unsigned long i, sum = 0;
1991 for_each_online_cpu(i)
1992 sum += cpu_rq(i)->nr_running;
1997 unsigned long long nr_context_switches(void)
2000 unsigned long long sum = 0;
2002 for_each_possible_cpu(i)
2003 sum += cpu_rq(i)->nr_switches;
2008 unsigned long nr_iowait(void)
2010 unsigned long i, sum = 0;
2012 for_each_possible_cpu(i)
2013 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2018 unsigned long nr_iowait_cpu(int cpu)
2020 struct rq *this = cpu_rq(cpu);
2021 return atomic_read(&this->nr_iowait);
2024 unsigned long this_cpu_load(void)
2026 struct rq *this = this_rq();
2027 return this->cpu_load[0];
2032 * Global load-average calculations
2034 * We take a distributed and async approach to calculating the global load-avg
2035 * in order to minimize overhead.
2037 * The global load average is an exponentially decaying average of nr_running +
2038 * nr_uninterruptible.
2040 * Once every LOAD_FREQ:
2043 * for_each_possible_cpu(cpu)
2044 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2046 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2048 * Due to a number of reasons the above turns in the mess below:
2050 * - for_each_possible_cpu() is prohibitively expensive on machines with
2051 * serious number of cpus, therefore we need to take a distributed approach
2052 * to calculating nr_active.
2054 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2055 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2057 * So assuming nr_active := 0 when we start out -- true per definition, we
2058 * can simply take per-cpu deltas and fold those into a global accumulate
2059 * to obtain the same result. See calc_load_fold_active().
2061 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2062 * across the machine, we assume 10 ticks is sufficient time for every
2063 * cpu to have completed this task.
2065 * This places an upper-bound on the IRQ-off latency of the machine. Then
2066 * again, being late doesn't loose the delta, just wrecks the sample.
2068 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2069 * this would add another cross-cpu cacheline miss and atomic operation
2070 * to the wakeup path. Instead we increment on whatever cpu the task ran
2071 * when it went into uninterruptible state and decrement on whatever cpu
2072 * did the wakeup. This means that only the sum of nr_uninterruptible over
2073 * all cpus yields the correct result.
2075 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2078 /* Variables and functions for calc_load */
2079 static atomic_long_t calc_load_tasks;
2080 static unsigned long calc_load_update;
2081 unsigned long avenrun[3];
2082 EXPORT_SYMBOL(avenrun); /* should be removed */
2085 * get_avenrun - get the load average array
2086 * @loads: pointer to dest load array
2087 * @offset: offset to add
2088 * @shift: shift count to shift the result left
2090 * These values are estimates at best, so no need for locking.
2092 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2094 loads[0] = (avenrun[0] + offset) << shift;
2095 loads[1] = (avenrun[1] + offset) << shift;
2096 loads[2] = (avenrun[2] + offset) << shift;
2099 static long calc_load_fold_active(struct rq *this_rq)
2101 long nr_active, delta = 0;
2103 nr_active = this_rq->nr_running;
2104 nr_active += (long) this_rq->nr_uninterruptible;
2106 if (nr_active != this_rq->calc_load_active) {
2107 delta = nr_active - this_rq->calc_load_active;
2108 this_rq->calc_load_active = nr_active;
2115 * a1 = a0 * e + a * (1 - e)
2117 static unsigned long
2118 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2121 load += active * (FIXED_1 - exp);
2122 load += 1UL << (FSHIFT - 1);
2123 return load >> FSHIFT;
2128 * Handle NO_HZ for the global load-average.
2130 * Since the above described distributed algorithm to compute the global
2131 * load-average relies on per-cpu sampling from the tick, it is affected by
2134 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2135 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2136 * when we read the global state.
2138 * Obviously reality has to ruin such a delightfully simple scheme:
2140 * - When we go NO_HZ idle during the window, we can negate our sample
2141 * contribution, causing under-accounting.
2143 * We avoid this by keeping two idle-delta counters and flipping them
2144 * when the window starts, thus separating old and new NO_HZ load.
2146 * The only trick is the slight shift in index flip for read vs write.
2150 * |-|-----------|-|-----------|-|-----------|-|
2151 * r:0 0 1 1 0 0 1 1 0
2152 * w:0 1 1 0 0 1 1 0 0
2154 * This ensures we'll fold the old idle contribution in this window while
2155 * accumlating the new one.
2157 * - When we wake up from NO_HZ idle during the window, we push up our
2158 * contribution, since we effectively move our sample point to a known
2161 * This is solved by pushing the window forward, and thus skipping the
2162 * sample, for this cpu (effectively using the idle-delta for this cpu which
2163 * was in effect at the time the window opened). This also solves the issue
2164 * of having to deal with a cpu having been in NOHZ idle for multiple
2165 * LOAD_FREQ intervals.
2167 * When making the ILB scale, we should try to pull this in as well.
2169 static atomic_long_t calc_load_idle[2];
2170 static int calc_load_idx;
2172 static inline int calc_load_write_idx(void)
2174 int idx = calc_load_idx;
2177 * See calc_global_nohz(), if we observe the new index, we also
2178 * need to observe the new update time.
2183 * If the folding window started, make sure we start writing in the
2186 if (!time_before(jiffies, calc_load_update))
2192 static inline int calc_load_read_idx(void)
2194 return calc_load_idx & 1;
2197 void calc_load_enter_idle(void)
2199 struct rq *this_rq = this_rq();
2203 * We're going into NOHZ mode, if there's any pending delta, fold it
2204 * into the pending idle delta.
2206 delta = calc_load_fold_active(this_rq);
2208 int idx = calc_load_write_idx();
2209 atomic_long_add(delta, &calc_load_idle[idx]);
2213 void calc_load_exit_idle(void)
2215 struct rq *this_rq = this_rq();
2218 * If we're still before the sample window, we're done.
2220 if (time_before(jiffies, this_rq->calc_load_update))
2224 * We woke inside or after the sample window, this means we're already
2225 * accounted through the nohz accounting, so skip the entire deal and
2226 * sync up for the next window.
2228 this_rq->calc_load_update = calc_load_update;
2229 if (time_before(jiffies, this_rq->calc_load_update + 10))
2230 this_rq->calc_load_update += LOAD_FREQ;
2233 static long calc_load_fold_idle(void)
2235 int idx = calc_load_read_idx();
2238 if (atomic_long_read(&calc_load_idle[idx]))
2239 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2245 * fixed_power_int - compute: x^n, in O(log n) time
2247 * @x: base of the power
2248 * @frac_bits: fractional bits of @x
2249 * @n: power to raise @x to.
2251 * By exploiting the relation between the definition of the natural power
2252 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2253 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2254 * (where: n_i \elem {0, 1}, the binary vector representing n),
2255 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2256 * of course trivially computable in O(log_2 n), the length of our binary
2259 static unsigned long
2260 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2262 unsigned long result = 1UL << frac_bits;
2267 result += 1UL << (frac_bits - 1);
2268 result >>= frac_bits;
2274 x += 1UL << (frac_bits - 1);
2282 * a1 = a0 * e + a * (1 - e)
2284 * a2 = a1 * e + a * (1 - e)
2285 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2286 * = a0 * e^2 + a * (1 - e) * (1 + e)
2288 * a3 = a2 * e + a * (1 - e)
2289 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2290 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2294 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2295 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2296 * = a0 * e^n + a * (1 - e^n)
2298 * [1] application of the geometric series:
2301 * S_n := \Sum x^i = -------------
2304 static unsigned long
2305 calc_load_n(unsigned long load, unsigned long exp,
2306 unsigned long active, unsigned int n)
2309 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2313 * NO_HZ can leave us missing all per-cpu ticks calling
2314 * calc_load_account_active(), but since an idle CPU folds its delta into
2315 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2316 * in the pending idle delta if our idle period crossed a load cycle boundary.
2318 * Once we've updated the global active value, we need to apply the exponential
2319 * weights adjusted to the number of cycles missed.
2321 static void calc_global_nohz(void)
2323 long delta, active, n;
2325 if (!time_before(jiffies, calc_load_update + 10)) {
2327 * Catch-up, fold however many we are behind still
2329 delta = jiffies - calc_load_update - 10;
2330 n = 1 + (delta / LOAD_FREQ);
2332 active = atomic_long_read(&calc_load_tasks);
2333 active = active > 0 ? active * FIXED_1 : 0;
2335 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2336 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2337 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2339 calc_load_update += n * LOAD_FREQ;
2343 * Flip the idle index...
2345 * Make sure we first write the new time then flip the index, so that
2346 * calc_load_write_idx() will see the new time when it reads the new
2347 * index, this avoids a double flip messing things up.
2352 #else /* !CONFIG_NO_HZ */
2354 static inline long calc_load_fold_idle(void) { return 0; }
2355 static inline void calc_global_nohz(void) { }
2357 #endif /* CONFIG_NO_HZ */
2360 * calc_load - update the avenrun load estimates 10 ticks after the
2361 * CPUs have updated calc_load_tasks.
2363 void calc_global_load(unsigned long ticks)
2367 if (time_before(jiffies, calc_load_update + 10))
2371 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2373 delta = calc_load_fold_idle();
2375 atomic_long_add(delta, &calc_load_tasks);
2377 active = atomic_long_read(&calc_load_tasks);
2378 active = active > 0 ? active * FIXED_1 : 0;
2380 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2381 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2382 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2384 calc_load_update += LOAD_FREQ;
2387 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2393 * Called from update_cpu_load() to periodically update this CPU's
2396 static void calc_load_account_active(struct rq *this_rq)
2400 if (time_before(jiffies, this_rq->calc_load_update))
2403 delta = calc_load_fold_active(this_rq);
2405 atomic_long_add(delta, &calc_load_tasks);
2407 this_rq->calc_load_update += LOAD_FREQ;
2411 * End of global load-average stuff
2415 * The exact cpuload at various idx values, calculated at every tick would be
2416 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2418 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2419 * on nth tick when cpu may be busy, then we have:
2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2421 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2423 * decay_load_missed() below does efficient calculation of
2424 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2425 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2427 * The calculation is approximated on a 128 point scale.
2428 * degrade_zero_ticks is the number of ticks after which load at any
2429 * particular idx is approximated to be zero.
2430 * degrade_factor is a precomputed table, a row for each load idx.
2431 * Each column corresponds to degradation factor for a power of two ticks,
2432 * based on 128 point scale.
2434 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2435 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2437 * With this power of 2 load factors, we can degrade the load n times
2438 * by looking at 1 bits in n and doing as many mult/shift instead of
2439 * n mult/shifts needed by the exact degradation.
2441 #define DEGRADE_SHIFT 7
2442 static const unsigned char
2443 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2444 static const unsigned char
2445 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2446 {0, 0, 0, 0, 0, 0, 0, 0},
2447 {64, 32, 8, 0, 0, 0, 0, 0},
2448 {96, 72, 40, 12, 1, 0, 0},
2449 {112, 98, 75, 43, 15, 1, 0},
2450 {120, 112, 98, 76, 45, 16, 2} };
2453 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2454 * would be when CPU is idle and so we just decay the old load without
2455 * adding any new load.
2457 static unsigned long
2458 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2462 if (!missed_updates)
2465 if (missed_updates >= degrade_zero_ticks[idx])
2469 return load >> missed_updates;
2471 while (missed_updates) {
2472 if (missed_updates % 2)
2473 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2475 missed_updates >>= 1;
2482 * Update rq->cpu_load[] statistics. This function is usually called every
2483 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2484 * every tick. We fix it up based on jiffies.
2486 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2487 unsigned long pending_updates)
2491 this_rq->nr_load_updates++;
2493 /* Update our load: */
2494 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2495 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2496 unsigned long old_load, new_load;
2498 /* scale is effectively 1 << i now, and >> i divides by scale */
2500 old_load = this_rq->cpu_load[i];
2501 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2502 new_load = this_load;
2504 * Round up the averaging division if load is increasing. This
2505 * prevents us from getting stuck on 9 if the load is 10, for
2508 if (new_load > old_load)
2509 new_load += scale - 1;
2511 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2514 sched_avg_update(this_rq);
2519 * There is no sane way to deal with nohz on smp when using jiffies because the
2520 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2521 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2523 * Therefore we cannot use the delta approach from the regular tick since that
2524 * would seriously skew the load calculation. However we'll make do for those
2525 * updates happening while idle (nohz_idle_balance) or coming out of idle
2526 * (tick_nohz_idle_exit).
2528 * This means we might still be one tick off for nohz periods.
2532 * Called from nohz_idle_balance() to update the load ratings before doing the
2535 void update_idle_cpu_load(struct rq *this_rq)
2537 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2538 unsigned long load = this_rq->load.weight;
2539 unsigned long pending_updates;
2542 * bail if there's load or we're actually up-to-date.
2544 if (load || curr_jiffies == this_rq->last_load_update_tick)
2547 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2548 this_rq->last_load_update_tick = curr_jiffies;
2550 __update_cpu_load(this_rq, load, pending_updates);
2554 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2556 void update_cpu_load_nohz(void)
2558 struct rq *this_rq = this_rq();
2559 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2560 unsigned long pending_updates;
2562 if (curr_jiffies == this_rq->last_load_update_tick)
2565 raw_spin_lock(&this_rq->lock);
2566 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2567 if (pending_updates) {
2568 this_rq->last_load_update_tick = curr_jiffies;
2570 * We were idle, this means load 0, the current load might be
2571 * !0 due to remote wakeups and the sort.
2573 __update_cpu_load(this_rq, 0, pending_updates);
2575 raw_spin_unlock(&this_rq->lock);
2577 #endif /* CONFIG_NO_HZ */
2580 * Called from scheduler_tick()
2582 static void update_cpu_load_active(struct rq *this_rq)
2585 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2587 this_rq->last_load_update_tick = jiffies;
2588 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2590 calc_load_account_active(this_rq);
2596 * sched_exec - execve() is a valuable balancing opportunity, because at
2597 * this point the task has the smallest effective memory and cache footprint.
2599 void sched_exec(void)
2601 struct task_struct *p = current;
2602 unsigned long flags;
2605 raw_spin_lock_irqsave(&p->pi_lock, flags);
2606 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2607 if (dest_cpu == smp_processor_id())
2610 if (likely(cpu_active(dest_cpu))) {
2611 struct migration_arg arg = { p, dest_cpu };
2613 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2614 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2618 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2623 DEFINE_PER_CPU(struct kernel_stat, kstat);
2624 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2626 EXPORT_PER_CPU_SYMBOL(kstat);
2627 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2630 * Return any ns on the sched_clock that have not yet been accounted in
2631 * @p in case that task is currently running.
2633 * Called with task_rq_lock() held on @rq.
2635 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2639 if (task_current(rq, p)) {
2640 update_rq_clock(rq);
2641 ns = rq->clock_task - p->se.exec_start;
2649 unsigned long long task_delta_exec(struct task_struct *p)
2651 unsigned long flags;
2655 rq = task_rq_lock(p, &flags);
2656 ns = do_task_delta_exec(p, rq);
2657 task_rq_unlock(rq, p, &flags);
2663 * Return accounted runtime for the task.
2664 * In case the task is currently running, return the runtime plus current's
2665 * pending runtime that have not been accounted yet.
2667 unsigned long long task_sched_runtime(struct task_struct *p)
2669 unsigned long flags;
2673 rq = task_rq_lock(p, &flags);
2674 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2675 task_rq_unlock(rq, p, &flags);
2681 * This function gets called by the timer code, with HZ frequency.
2682 * We call it with interrupts disabled.
2684 void scheduler_tick(void)
2686 int cpu = smp_processor_id();
2687 struct rq *rq = cpu_rq(cpu);
2688 struct task_struct *curr = rq->curr;
2692 raw_spin_lock(&rq->lock);
2693 update_rq_clock(rq);
2694 update_cpu_load_active(rq);
2695 curr->sched_class->task_tick(rq, curr, 0);
2696 raw_spin_unlock(&rq->lock);
2698 perf_event_task_tick();
2701 rq->idle_balance = idle_cpu(cpu);
2702 trigger_load_balance(rq, cpu);
2706 notrace unsigned long get_parent_ip(unsigned long addr)
2708 if (in_lock_functions(addr)) {
2709 addr = CALLER_ADDR2;
2710 if (in_lock_functions(addr))
2711 addr = CALLER_ADDR3;
2716 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2717 defined(CONFIG_PREEMPT_TRACER))
2719 void __kprobes add_preempt_count(int val)
2721 #ifdef CONFIG_DEBUG_PREEMPT
2725 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2728 preempt_count() += val;
2729 #ifdef CONFIG_DEBUG_PREEMPT
2731 * Spinlock count overflowing soon?
2733 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2736 if (preempt_count() == val)
2737 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2739 EXPORT_SYMBOL(add_preempt_count);
2741 void __kprobes sub_preempt_count(int val)
2743 #ifdef CONFIG_DEBUG_PREEMPT
2747 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2750 * Is the spinlock portion underflowing?
2752 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2753 !(preempt_count() & PREEMPT_MASK)))
2757 if (preempt_count() == val)
2758 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2759 preempt_count() -= val;
2761 EXPORT_SYMBOL(sub_preempt_count);
2766 * Print scheduling while atomic bug:
2768 static noinline void __schedule_bug(struct task_struct *prev)
2770 if (oops_in_progress)
2773 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2774 prev->comm, prev->pid, preempt_count());
2776 debug_show_held_locks(prev);
2778 if (irqs_disabled())
2779 print_irqtrace_events(prev);
2781 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2785 * Various schedule()-time debugging checks and statistics:
2787 static inline void schedule_debug(struct task_struct *prev)
2790 * Test if we are atomic. Since do_exit() needs to call into
2791 * schedule() atomically, we ignore that path for now.
2792 * Otherwise, whine if we are scheduling when we should not be.
2794 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2795 __schedule_bug(prev);
2798 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2800 schedstat_inc(this_rq(), sched_count);
2803 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2805 if (prev->on_rq || rq->skip_clock_update < 0)
2806 update_rq_clock(rq);
2807 prev->sched_class->put_prev_task(rq, prev);
2811 * Pick up the highest-prio task:
2813 static inline struct task_struct *
2814 pick_next_task(struct rq *rq)
2816 const struct sched_class *class;
2817 struct task_struct *p;
2820 * Optimization: we know that if all tasks are in
2821 * the fair class we can call that function directly:
2823 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2824 p = fair_sched_class.pick_next_task(rq);
2829 for_each_class(class) {
2830 p = class->pick_next_task(rq);
2835 BUG(); /* the idle class will always have a runnable task */
2839 * __schedule() is the main scheduler function.
2841 * The main means of driving the scheduler and thus entering this function are:
2843 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2845 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2846 * paths. For example, see arch/x86/entry_64.S.
2848 * To drive preemption between tasks, the scheduler sets the flag in timer
2849 * interrupt handler scheduler_tick().
2851 * 3. Wakeups don't really cause entry into schedule(). They add a
2852 * task to the run-queue and that's it.
2854 * Now, if the new task added to the run-queue preempts the current
2855 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2856 * called on the nearest possible occasion:
2858 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2860 * - in syscall or exception context, at the next outmost
2861 * preempt_enable(). (this might be as soon as the wake_up()'s
2864 * - in IRQ context, return from interrupt-handler to
2865 * preemptible context
2867 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2870 * - cond_resched() call
2871 * - explicit schedule() call
2872 * - return from syscall or exception to user-space
2873 * - return from interrupt-handler to user-space
2875 static void __sched __schedule(void)
2877 struct task_struct *prev, *next;
2878 unsigned long *switch_count;
2884 cpu = smp_processor_id();
2886 rcu_note_context_switch(cpu);
2889 schedule_debug(prev);
2891 if (sched_feat(HRTICK))
2894 raw_spin_lock_irq(&rq->lock);
2896 switch_count = &prev->nivcsw;
2897 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2898 if (unlikely(signal_pending_state(prev->state, prev))) {
2899 prev->state = TASK_RUNNING;
2901 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2905 * If a worker went to sleep, notify and ask workqueue
2906 * whether it wants to wake up a task to maintain
2909 if (prev->flags & PF_WQ_WORKER) {
2910 struct task_struct *to_wakeup;
2912 to_wakeup = wq_worker_sleeping(prev, cpu);
2914 try_to_wake_up_local(to_wakeup);
2917 switch_count = &prev->nvcsw;
2920 pre_schedule(rq, prev);
2922 if (unlikely(!rq->nr_running))
2923 idle_balance(cpu, rq);
2925 put_prev_task(rq, prev);
2926 next = pick_next_task(rq);
2927 clear_tsk_need_resched(prev);
2928 rq->skip_clock_update = 0;
2930 if (likely(prev != next)) {
2935 context_switch(rq, prev, next); /* unlocks the rq */
2937 * The context switch have flipped the stack from under us
2938 * and restored the local variables which were saved when
2939 * this task called schedule() in the past. prev == current
2940 * is still correct, but it can be moved to another cpu/rq.
2942 cpu = smp_processor_id();
2945 raw_spin_unlock_irq(&rq->lock);
2949 sched_preempt_enable_no_resched();
2954 static inline void sched_submit_work(struct task_struct *tsk)
2956 if (!tsk->state || tsk_is_pi_blocked(tsk))
2959 * If we are going to sleep and we have plugged IO queued,
2960 * make sure to submit it to avoid deadlocks.
2962 if (blk_needs_flush_plug(tsk))
2963 blk_schedule_flush_plug(tsk);
2966 asmlinkage void __sched schedule(void)
2968 struct task_struct *tsk = current;
2970 sched_submit_work(tsk);
2973 EXPORT_SYMBOL(schedule);
2975 #ifdef CONFIG_CONTEXT_TRACKING
2976 asmlinkage void __sched schedule_user(void)
2979 * If we come here after a random call to set_need_resched(),
2980 * or we have been woken up remotely but the IPI has not yet arrived,
2981 * we haven't yet exited the RCU idle mode. Do it here manually until
2982 * we find a better solution.
2991 * schedule_preempt_disabled - called with preemption disabled
2993 * Returns with preemption disabled. Note: preempt_count must be 1
2995 void __sched schedule_preempt_disabled(void)
2997 sched_preempt_enable_no_resched();
3002 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3004 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3006 if (lock->owner != owner)
3010 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3011 * lock->owner still matches owner, if that fails, owner might
3012 * point to free()d memory, if it still matches, the rcu_read_lock()
3013 * ensures the memory stays valid.
3017 return owner->on_cpu;
3021 * Look out! "owner" is an entirely speculative pointer
3022 * access and not reliable.
3024 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3026 if (!sched_feat(OWNER_SPIN))
3030 while (owner_running(lock, owner)) {
3034 arch_mutex_cpu_relax();
3039 * We break out the loop above on need_resched() and when the
3040 * owner changed, which is a sign for heavy contention. Return
3041 * success only when lock->owner is NULL.
3043 return lock->owner == NULL;
3047 #ifdef CONFIG_PREEMPT
3049 * this is the entry point to schedule() from in-kernel preemption
3050 * off of preempt_enable. Kernel preemptions off return from interrupt
3051 * occur there and call schedule directly.
3053 asmlinkage void __sched notrace preempt_schedule(void)
3055 struct thread_info *ti = current_thread_info();
3058 * If there is a non-zero preempt_count or interrupts are disabled,
3059 * we do not want to preempt the current task. Just return..
3061 if (likely(ti->preempt_count || irqs_disabled()))
3065 add_preempt_count_notrace(PREEMPT_ACTIVE);
3067 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3070 * Check again in case we missed a preemption opportunity
3071 * between schedule and now.
3074 } while (need_resched());
3076 EXPORT_SYMBOL(preempt_schedule);
3079 * this is the entry point to schedule() from kernel preemption
3080 * off of irq context.
3081 * Note, that this is called and return with irqs disabled. This will
3082 * protect us against recursive calling from irq.
3084 asmlinkage void __sched preempt_schedule_irq(void)
3086 struct thread_info *ti = current_thread_info();
3088 /* Catch callers which need to be fixed */
3089 BUG_ON(ti->preempt_count || !irqs_disabled());
3093 add_preempt_count(PREEMPT_ACTIVE);
3096 local_irq_disable();
3097 sub_preempt_count(PREEMPT_ACTIVE);
3100 * Check again in case we missed a preemption opportunity
3101 * between schedule and now.
3104 } while (need_resched());
3107 #endif /* CONFIG_PREEMPT */
3109 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3112 return try_to_wake_up(curr->private, mode, wake_flags);
3114 EXPORT_SYMBOL(default_wake_function);
3117 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3118 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3119 * number) then we wake all the non-exclusive tasks and one exclusive task.
3121 * There are circumstances in which we can try to wake a task which has already
3122 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3123 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3125 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3126 int nr_exclusive, int wake_flags, void *key)
3128 wait_queue_t *curr, *next;
3130 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3131 unsigned flags = curr->flags;
3133 if (curr->func(curr, mode, wake_flags, key) &&
3134 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3140 * __wake_up - wake up threads blocked on a waitqueue.
3142 * @mode: which threads
3143 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3144 * @key: is directly passed to the wakeup function
3146 * It may be assumed that this function implies a write memory barrier before
3147 * changing the task state if and only if any tasks are woken up.
3149 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3150 int nr_exclusive, void *key)
3152 unsigned long flags;
3154 spin_lock_irqsave(&q->lock, flags);
3155 __wake_up_common(q, mode, nr_exclusive, 0, key);
3156 spin_unlock_irqrestore(&q->lock, flags);
3158 EXPORT_SYMBOL(__wake_up);
3161 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3163 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3165 __wake_up_common(q, mode, nr, 0, NULL);
3167 EXPORT_SYMBOL_GPL(__wake_up_locked);
3169 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3171 __wake_up_common(q, mode, 1, 0, key);
3173 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3176 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3178 * @mode: which threads
3179 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3180 * @key: opaque value to be passed to wakeup targets
3182 * The sync wakeup differs that the waker knows that it will schedule
3183 * away soon, so while the target thread will be woken up, it will not
3184 * be migrated to another CPU - ie. the two threads are 'synchronized'
3185 * with each other. This can prevent needless bouncing between CPUs.
3187 * On UP it can prevent extra preemption.
3189 * It may be assumed that this function implies a write memory barrier before
3190 * changing the task state if and only if any tasks are woken up.
3192 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3193 int nr_exclusive, void *key)
3195 unsigned long flags;
3196 int wake_flags = WF_SYNC;
3201 if (unlikely(!nr_exclusive))
3204 spin_lock_irqsave(&q->lock, flags);
3205 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3206 spin_unlock_irqrestore(&q->lock, flags);
3208 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3211 * __wake_up_sync - see __wake_up_sync_key()
3213 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3215 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3217 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3220 * complete: - signals a single thread waiting on this completion
3221 * @x: holds the state of this particular completion
3223 * This will wake up a single thread waiting on this completion. Threads will be
3224 * awakened in the same order in which they were queued.
3226 * See also complete_all(), wait_for_completion() and related routines.
3228 * It may be assumed that this function implies a write memory barrier before
3229 * changing the task state if and only if any tasks are woken up.
3231 void complete(struct completion *x)
3233 unsigned long flags;
3235 spin_lock_irqsave(&x->wait.lock, flags);
3237 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3238 spin_unlock_irqrestore(&x->wait.lock, flags);
3240 EXPORT_SYMBOL(complete);
3243 * complete_all: - signals all threads waiting on this completion
3244 * @x: holds the state of this particular completion
3246 * This will wake up all threads waiting on this particular completion event.
3248 * It may be assumed that this function implies a write memory barrier before
3249 * changing the task state if and only if any tasks are woken up.
3251 void complete_all(struct completion *x)
3253 unsigned long flags;
3255 spin_lock_irqsave(&x->wait.lock, flags);
3256 x->done += UINT_MAX/2;
3257 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3258 spin_unlock_irqrestore(&x->wait.lock, flags);
3260 EXPORT_SYMBOL(complete_all);
3262 static inline long __sched
3263 do_wait_for_common(struct completion *x, long timeout, int state)
3266 DECLARE_WAITQUEUE(wait, current);
3268 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3270 if (signal_pending_state(state, current)) {
3271 timeout = -ERESTARTSYS;
3274 __set_current_state(state);
3275 spin_unlock_irq(&x->wait.lock);
3276 timeout = schedule_timeout(timeout);
3277 spin_lock_irq(&x->wait.lock);
3278 } while (!x->done && timeout);
3279 __remove_wait_queue(&x->wait, &wait);
3284 return timeout ?: 1;
3288 wait_for_common(struct completion *x, long timeout, int state)
3292 spin_lock_irq(&x->wait.lock);
3293 timeout = do_wait_for_common(x, timeout, state);
3294 spin_unlock_irq(&x->wait.lock);
3299 * wait_for_completion: - waits for completion of a task
3300 * @x: holds the state of this particular completion
3302 * This waits to be signaled for completion of a specific task. It is NOT
3303 * interruptible and there is no timeout.
3305 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3306 * and interrupt capability. Also see complete().
3308 void __sched wait_for_completion(struct completion *x)
3310 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3312 EXPORT_SYMBOL(wait_for_completion);
3315 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3316 * @x: holds the state of this particular completion
3317 * @timeout: timeout value in jiffies
3319 * This waits for either a completion of a specific task to be signaled or for a
3320 * specified timeout to expire. The timeout is in jiffies. It is not
3323 * The return value is 0 if timed out, and positive (at least 1, or number of
3324 * jiffies left till timeout) if completed.
3326 unsigned long __sched
3327 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3329 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3331 EXPORT_SYMBOL(wait_for_completion_timeout);
3334 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3335 * @x: holds the state of this particular completion
3337 * This waits for completion of a specific task to be signaled. It is
3340 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3342 int __sched wait_for_completion_interruptible(struct completion *x)
3344 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3345 if (t == -ERESTARTSYS)
3349 EXPORT_SYMBOL(wait_for_completion_interruptible);
3352 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3353 * @x: holds the state of this particular completion
3354 * @timeout: timeout value in jiffies
3356 * This waits for either a completion of a specific task to be signaled or for a
3357 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3359 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3360 * positive (at least 1, or number of jiffies left till timeout) if completed.
3363 wait_for_completion_interruptible_timeout(struct completion *x,
3364 unsigned long timeout)
3366 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3368 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3371 * wait_for_completion_killable: - waits for completion of a task (killable)
3372 * @x: holds the state of this particular completion
3374 * This waits to be signaled for completion of a specific task. It can be
3375 * interrupted by a kill signal.
3377 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3379 int __sched wait_for_completion_killable(struct completion *x)
3381 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3382 if (t == -ERESTARTSYS)
3386 EXPORT_SYMBOL(wait_for_completion_killable);
3389 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3390 * @x: holds the state of this particular completion
3391 * @timeout: timeout value in jiffies
3393 * This waits for either a completion of a specific task to be
3394 * signaled or for a specified timeout to expire. It can be
3395 * interrupted by a kill signal. The timeout is in jiffies.
3397 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3398 * positive (at least 1, or number of jiffies left till timeout) if completed.
3401 wait_for_completion_killable_timeout(struct completion *x,
3402 unsigned long timeout)
3404 return wait_for_common(x, timeout, TASK_KILLABLE);
3406 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3409 * try_wait_for_completion - try to decrement a completion without blocking
3410 * @x: completion structure
3412 * Returns: 0 if a decrement cannot be done without blocking
3413 * 1 if a decrement succeeded.
3415 * If a completion is being used as a counting completion,
3416 * attempt to decrement the counter without blocking. This
3417 * enables us to avoid waiting if the resource the completion
3418 * is protecting is not available.
3420 bool try_wait_for_completion(struct completion *x)
3422 unsigned long flags;
3425 spin_lock_irqsave(&x->wait.lock, flags);
3430 spin_unlock_irqrestore(&x->wait.lock, flags);
3433 EXPORT_SYMBOL(try_wait_for_completion);
3436 * completion_done - Test to see if a completion has any waiters
3437 * @x: completion structure
3439 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3440 * 1 if there are no waiters.
3443 bool completion_done(struct completion *x)
3445 unsigned long flags;
3448 spin_lock_irqsave(&x->wait.lock, flags);
3451 spin_unlock_irqrestore(&x->wait.lock, flags);
3454 EXPORT_SYMBOL(completion_done);
3457 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3459 unsigned long flags;
3462 init_waitqueue_entry(&wait, current);
3464 __set_current_state(state);
3466 spin_lock_irqsave(&q->lock, flags);
3467 __add_wait_queue(q, &wait);
3468 spin_unlock(&q->lock);
3469 timeout = schedule_timeout(timeout);
3470 spin_lock_irq(&q->lock);
3471 __remove_wait_queue(q, &wait);
3472 spin_unlock_irqrestore(&q->lock, flags);
3477 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3479 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3481 EXPORT_SYMBOL(interruptible_sleep_on);
3484 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3486 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3488 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3490 void __sched sleep_on(wait_queue_head_t *q)
3492 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3494 EXPORT_SYMBOL(sleep_on);
3496 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3498 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3500 EXPORT_SYMBOL(sleep_on_timeout);
3502 #ifdef CONFIG_RT_MUTEXES
3505 * rt_mutex_setprio - set the current priority of a task
3507 * @prio: prio value (kernel-internal form)
3509 * This function changes the 'effective' priority of a task. It does
3510 * not touch ->normal_prio like __setscheduler().
3512 * Used by the rt_mutex code to implement priority inheritance logic.
3514 void rt_mutex_setprio(struct task_struct *p, int prio)
3516 int oldprio, on_rq, running;
3518 const struct sched_class *prev_class;
3520 BUG_ON(prio < 0 || prio > MAX_PRIO);
3522 rq = __task_rq_lock(p);
3525 * Idle task boosting is a nono in general. There is one
3526 * exception, when PREEMPT_RT and NOHZ is active:
3528 * The idle task calls get_next_timer_interrupt() and holds
3529 * the timer wheel base->lock on the CPU and another CPU wants
3530 * to access the timer (probably to cancel it). We can safely
3531 * ignore the boosting request, as the idle CPU runs this code
3532 * with interrupts disabled and will complete the lock
3533 * protected section without being interrupted. So there is no
3534 * real need to boost.
3536 if (unlikely(p == rq->idle)) {
3537 WARN_ON(p != rq->curr);
3538 WARN_ON(p->pi_blocked_on);
3542 trace_sched_pi_setprio(p, prio);
3544 prev_class = p->sched_class;
3546 running = task_current(rq, p);
3548 dequeue_task(rq, p, 0);
3550 p->sched_class->put_prev_task(rq, p);
3553 p->sched_class = &rt_sched_class;
3555 p->sched_class = &fair_sched_class;
3560 p->sched_class->set_curr_task(rq);
3562 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3564 check_class_changed(rq, p, prev_class, oldprio);
3566 __task_rq_unlock(rq);
3569 void set_user_nice(struct task_struct *p, long nice)
3571 int old_prio, delta, on_rq;
3572 unsigned long flags;
3575 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3578 * We have to be careful, if called from sys_setpriority(),
3579 * the task might be in the middle of scheduling on another CPU.
3581 rq = task_rq_lock(p, &flags);
3583 * The RT priorities are set via sched_setscheduler(), but we still
3584 * allow the 'normal' nice value to be set - but as expected
3585 * it wont have any effect on scheduling until the task is
3586 * SCHED_FIFO/SCHED_RR:
3588 if (task_has_rt_policy(p)) {
3589 p->static_prio = NICE_TO_PRIO(nice);
3594 dequeue_task(rq, p, 0);
3596 p->static_prio = NICE_TO_PRIO(nice);
3599 p->prio = effective_prio(p);
3600 delta = p->prio - old_prio;
3603 enqueue_task(rq, p, 0);
3605 * If the task increased its priority or is running and
3606 * lowered its priority, then reschedule its CPU:
3608 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3609 resched_task(rq->curr);
3612 task_rq_unlock(rq, p, &flags);
3614 EXPORT_SYMBOL(set_user_nice);
3617 * can_nice - check if a task can reduce its nice value
3621 int can_nice(const struct task_struct *p, const int nice)
3623 /* convert nice value [19,-20] to rlimit style value [1,40] */
3624 int nice_rlim = 20 - nice;
3626 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3627 capable(CAP_SYS_NICE));
3630 #ifdef __ARCH_WANT_SYS_NICE
3633 * sys_nice - change the priority of the current process.
3634 * @increment: priority increment
3636 * sys_setpriority is a more generic, but much slower function that
3637 * does similar things.
3639 SYSCALL_DEFINE1(nice, int, increment)
3644 * Setpriority might change our priority at the same moment.
3645 * We don't have to worry. Conceptually one call occurs first
3646 * and we have a single winner.
3648 if (increment < -40)
3653 nice = TASK_NICE(current) + increment;
3659 if (increment < 0 && !can_nice(current, nice))
3662 retval = security_task_setnice(current, nice);
3666 set_user_nice(current, nice);
3673 * task_prio - return the priority value of a given task.
3674 * @p: the task in question.
3676 * This is the priority value as seen by users in /proc.
3677 * RT tasks are offset by -200. Normal tasks are centered
3678 * around 0, value goes from -16 to +15.
3680 int task_prio(const struct task_struct *p)
3682 return p->prio - MAX_RT_PRIO;
3686 * task_nice - return the nice value of a given task.
3687 * @p: the task in question.
3689 int task_nice(const struct task_struct *p)
3691 return TASK_NICE(p);
3693 EXPORT_SYMBOL(task_nice);
3696 * idle_cpu - is a given cpu idle currently?
3697 * @cpu: the processor in question.
3699 int idle_cpu(int cpu)
3701 struct rq *rq = cpu_rq(cpu);
3703 if (rq->curr != rq->idle)
3710 if (!llist_empty(&rq->wake_list))
3718 * idle_task - return the idle task for a given cpu.
3719 * @cpu: the processor in question.
3721 struct task_struct *idle_task(int cpu)
3723 return cpu_rq(cpu)->idle;
3727 * find_process_by_pid - find a process with a matching PID value.
3728 * @pid: the pid in question.
3730 static struct task_struct *find_process_by_pid(pid_t pid)
3732 return pid ? find_task_by_vpid(pid) : current;
3735 /* Actually do priority change: must hold rq lock. */
3737 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3740 p->rt_priority = prio;
3741 p->normal_prio = normal_prio(p);
3742 /* we are holding p->pi_lock already */
3743 p->prio = rt_mutex_getprio(p);
3744 if (rt_prio(p->prio))
3745 p->sched_class = &rt_sched_class;
3747 p->sched_class = &fair_sched_class;
3752 * check the target process has a UID that matches the current process's
3754 static bool check_same_owner(struct task_struct *p)
3756 const struct cred *cred = current_cred(), *pcred;
3760 pcred = __task_cred(p);
3761 match = (uid_eq(cred->euid, pcred->euid) ||
3762 uid_eq(cred->euid, pcred->uid));
3767 static int __sched_setscheduler(struct task_struct *p, int policy,
3768 const struct sched_param *param, bool user)
3770 int retval, oldprio, oldpolicy = -1, on_rq, running;
3771 unsigned long flags;
3772 const struct sched_class *prev_class;
3776 /* may grab non-irq protected spin_locks */
3777 BUG_ON(in_interrupt());
3779 /* double check policy once rq lock held */
3781 reset_on_fork = p->sched_reset_on_fork;
3782 policy = oldpolicy = p->policy;
3784 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3785 policy &= ~SCHED_RESET_ON_FORK;
3787 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3788 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3789 policy != SCHED_IDLE)
3794 * Valid priorities for SCHED_FIFO and SCHED_RR are
3795 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3796 * SCHED_BATCH and SCHED_IDLE is 0.
3798 if (param->sched_priority < 0 ||
3799 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3800 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3802 if (rt_policy(policy) != (param->sched_priority != 0))
3806 * Allow unprivileged RT tasks to decrease priority:
3808 if (user && !capable(CAP_SYS_NICE)) {
3809 if (rt_policy(policy)) {
3810 unsigned long rlim_rtprio =
3811 task_rlimit(p, RLIMIT_RTPRIO);
3813 /* can't set/change the rt policy */
3814 if (policy != p->policy && !rlim_rtprio)
3817 /* can't increase priority */
3818 if (param->sched_priority > p->rt_priority &&
3819 param->sched_priority > rlim_rtprio)
3824 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3825 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3827 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3828 if (!can_nice(p, TASK_NICE(p)))
3832 /* can't change other user's priorities */
3833 if (!check_same_owner(p))
3836 /* Normal users shall not reset the sched_reset_on_fork flag */
3837 if (p->sched_reset_on_fork && !reset_on_fork)
3842 retval = security_task_setscheduler(p);
3848 * make sure no PI-waiters arrive (or leave) while we are
3849 * changing the priority of the task:
3851 * To be able to change p->policy safely, the appropriate
3852 * runqueue lock must be held.
3854 rq = task_rq_lock(p, &flags);
3857 * Changing the policy of the stop threads its a very bad idea
3859 if (p == rq->stop) {
3860 task_rq_unlock(rq, p, &flags);
3865 * If not changing anything there's no need to proceed further:
3867 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3868 param->sched_priority == p->rt_priority))) {
3869 task_rq_unlock(rq, p, &flags);
3873 #ifdef CONFIG_RT_GROUP_SCHED
3876 * Do not allow realtime tasks into groups that have no runtime
3879 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3880 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3881 !task_group_is_autogroup(task_group(p))) {
3882 task_rq_unlock(rq, p, &flags);
3888 /* recheck policy now with rq lock held */
3889 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3890 policy = oldpolicy = -1;
3891 task_rq_unlock(rq, p, &flags);
3895 running = task_current(rq, p);
3897 dequeue_task(rq, p, 0);
3899 p->sched_class->put_prev_task(rq, p);
3901 p->sched_reset_on_fork = reset_on_fork;
3904 prev_class = p->sched_class;
3905 __setscheduler(rq, p, policy, param->sched_priority);
3908 p->sched_class->set_curr_task(rq);
3910 enqueue_task(rq, p, 0);
3912 check_class_changed(rq, p, prev_class, oldprio);
3913 task_rq_unlock(rq, p, &flags);
3915 rt_mutex_adjust_pi(p);
3921 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3922 * @p: the task in question.
3923 * @policy: new policy.
3924 * @param: structure containing the new RT priority.
3926 * NOTE that the task may be already dead.
3928 int sched_setscheduler(struct task_struct *p, int policy,
3929 const struct sched_param *param)
3931 return __sched_setscheduler(p, policy, param, true);
3933 EXPORT_SYMBOL_GPL(sched_setscheduler);
3936 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3937 * @p: the task in question.
3938 * @policy: new policy.
3939 * @param: structure containing the new RT priority.
3941 * Just like sched_setscheduler, only don't bother checking if the
3942 * current context has permission. For example, this is needed in
3943 * stop_machine(): we create temporary high priority worker threads,
3944 * but our caller might not have that capability.
3946 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3947 const struct sched_param *param)
3949 return __sched_setscheduler(p, policy, param, false);
3953 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3955 struct sched_param lparam;
3956 struct task_struct *p;
3959 if (!param || pid < 0)
3961 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3966 p = find_process_by_pid(pid);
3968 retval = sched_setscheduler(p, policy, &lparam);
3975 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3976 * @pid: the pid in question.
3977 * @policy: new policy.
3978 * @param: structure containing the new RT priority.
3980 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3981 struct sched_param __user *, param)
3983 /* negative values for policy are not valid */
3987 return do_sched_setscheduler(pid, policy, param);
3991 * sys_sched_setparam - set/change the RT priority of a thread
3992 * @pid: the pid in question.
3993 * @param: structure containing the new RT priority.
3995 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3997 return do_sched_setscheduler(pid, -1, param);
4001 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4002 * @pid: the pid in question.
4004 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4006 struct task_struct *p;
4014 p = find_process_by_pid(pid);
4016 retval = security_task_getscheduler(p);
4019 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4026 * sys_sched_getparam - get the RT priority of a thread
4027 * @pid: the pid in question.
4028 * @param: structure containing the RT priority.
4030 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4032 struct sched_param lp;
4033 struct task_struct *p;
4036 if (!param || pid < 0)
4040 p = find_process_by_pid(pid);
4045 retval = security_task_getscheduler(p);
4049 lp.sched_priority = p->rt_priority;
4053 * This one might sleep, we cannot do it with a spinlock held ...
4055 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4064 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4066 cpumask_var_t cpus_allowed, new_mask;
4067 struct task_struct *p;
4073 p = find_process_by_pid(pid);
4080 /* Prevent p going away */
4084 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4088 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4090 goto out_free_cpus_allowed;
4093 if (!check_same_owner(p)) {
4095 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4102 retval = security_task_setscheduler(p);
4106 cpuset_cpus_allowed(p, cpus_allowed);
4107 cpumask_and(new_mask, in_mask, cpus_allowed);
4109 retval = set_cpus_allowed_ptr(p, new_mask);
4112 cpuset_cpus_allowed(p, cpus_allowed);
4113 if (!cpumask_subset(new_mask, cpus_allowed)) {
4115 * We must have raced with a concurrent cpuset
4116 * update. Just reset the cpus_allowed to the
4117 * cpuset's cpus_allowed
4119 cpumask_copy(new_mask, cpus_allowed);
4124 free_cpumask_var(new_mask);
4125 out_free_cpus_allowed:
4126 free_cpumask_var(cpus_allowed);
4133 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4134 struct cpumask *new_mask)
4136 if (len < cpumask_size())
4137 cpumask_clear(new_mask);
4138 else if (len > cpumask_size())
4139 len = cpumask_size();
4141 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4145 * sys_sched_setaffinity - set the cpu affinity of a process
4146 * @pid: pid of the process
4147 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4148 * @user_mask_ptr: user-space pointer to the new cpu mask
4150 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4151 unsigned long __user *, user_mask_ptr)
4153 cpumask_var_t new_mask;
4156 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4159 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4161 retval = sched_setaffinity(pid, new_mask);
4162 free_cpumask_var(new_mask);
4166 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4168 struct task_struct *p;
4169 unsigned long flags;
4176 p = find_process_by_pid(pid);
4180 retval = security_task_getscheduler(p);
4184 raw_spin_lock_irqsave(&p->pi_lock, flags);
4185 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4186 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4196 * sys_sched_getaffinity - get the cpu affinity of a process
4197 * @pid: pid of the process
4198 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4199 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4201 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4202 unsigned long __user *, user_mask_ptr)
4207 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4209 if (len & (sizeof(unsigned long)-1))
4212 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4215 ret = sched_getaffinity(pid, mask);
4217 size_t retlen = min_t(size_t, len, cpumask_size());
4219 if (copy_to_user(user_mask_ptr, mask, retlen))
4224 free_cpumask_var(mask);
4230 * sys_sched_yield - yield the current processor to other threads.
4232 * This function yields the current CPU to other tasks. If there are no
4233 * other threads running on this CPU then this function will return.
4235 SYSCALL_DEFINE0(sched_yield)
4237 struct rq *rq = this_rq_lock();
4239 schedstat_inc(rq, yld_count);
4240 current->sched_class->yield_task(rq);
4243 * Since we are going to call schedule() anyway, there's
4244 * no need to preempt or enable interrupts:
4246 __release(rq->lock);
4247 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4248 do_raw_spin_unlock(&rq->lock);
4249 sched_preempt_enable_no_resched();
4256 static inline int should_resched(void)
4258 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4261 static void __cond_resched(void)
4263 add_preempt_count(PREEMPT_ACTIVE);
4265 sub_preempt_count(PREEMPT_ACTIVE);
4268 int __sched _cond_resched(void)
4270 if (should_resched()) {
4276 EXPORT_SYMBOL(_cond_resched);
4279 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4280 * call schedule, and on return reacquire the lock.
4282 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4283 * operations here to prevent schedule() from being called twice (once via
4284 * spin_unlock(), once by hand).
4286 int __cond_resched_lock(spinlock_t *lock)
4288 int resched = should_resched();
4291 lockdep_assert_held(lock);
4293 if (spin_needbreak(lock) || resched) {
4304 EXPORT_SYMBOL(__cond_resched_lock);
4306 int __sched __cond_resched_softirq(void)
4308 BUG_ON(!in_softirq());
4310 if (should_resched()) {
4318 EXPORT_SYMBOL(__cond_resched_softirq);
4321 * yield - yield the current processor to other threads.
4323 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4325 * The scheduler is at all times free to pick the calling task as the most
4326 * eligible task to run, if removing the yield() call from your code breaks
4327 * it, its already broken.
4329 * Typical broken usage is:
4334 * where one assumes that yield() will let 'the other' process run that will
4335 * make event true. If the current task is a SCHED_FIFO task that will never
4336 * happen. Never use yield() as a progress guarantee!!
4338 * If you want to use yield() to wait for something, use wait_event().
4339 * If you want to use yield() to be 'nice' for others, use cond_resched().
4340 * If you still want to use yield(), do not!
4342 void __sched yield(void)
4344 set_current_state(TASK_RUNNING);
4347 EXPORT_SYMBOL(yield);
4350 * yield_to - yield the current processor to another thread in
4351 * your thread group, or accelerate that thread toward the
4352 * processor it's on.
4354 * @preempt: whether task preemption is allowed or not
4356 * It's the caller's job to ensure that the target task struct
4357 * can't go away on us before we can do any checks.
4360 * true (>0) if we indeed boosted the target task.
4361 * false (0) if we failed to boost the target.
4362 * -ESRCH if there's no task to yield to.
4364 bool __sched yield_to(struct task_struct *p, bool preempt)
4366 struct task_struct *curr = current;
4367 struct rq *rq, *p_rq;
4368 unsigned long flags;
4371 local_irq_save(flags);
4377 * If we're the only runnable task on the rq and target rq also
4378 * has only one task, there's absolutely no point in yielding.
4380 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4385 double_rq_lock(rq, p_rq);
4386 while (task_rq(p) != p_rq) {
4387 double_rq_unlock(rq, p_rq);
4391 if (!curr->sched_class->yield_to_task)
4394 if (curr->sched_class != p->sched_class)
4397 if (task_running(p_rq, p) || p->state)
4400 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4402 schedstat_inc(rq, yld_count);
4404 * Make p's CPU reschedule; pick_next_entity takes care of
4407 if (preempt && rq != p_rq)
4408 resched_task(p_rq->curr);
4412 double_rq_unlock(rq, p_rq);
4414 local_irq_restore(flags);
4421 EXPORT_SYMBOL_GPL(yield_to);
4424 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4425 * that process accounting knows that this is a task in IO wait state.
4427 void __sched io_schedule(void)
4429 struct rq *rq = raw_rq();
4431 delayacct_blkio_start();
4432 atomic_inc(&rq->nr_iowait);
4433 blk_flush_plug(current);
4434 current->in_iowait = 1;
4436 current->in_iowait = 0;
4437 atomic_dec(&rq->nr_iowait);
4438 delayacct_blkio_end();
4440 EXPORT_SYMBOL(io_schedule);
4442 long __sched io_schedule_timeout(long timeout)
4444 struct rq *rq = raw_rq();
4447 delayacct_blkio_start();
4448 atomic_inc(&rq->nr_iowait);
4449 blk_flush_plug(current);
4450 current->in_iowait = 1;
4451 ret = schedule_timeout(timeout);
4452 current->in_iowait = 0;
4453 atomic_dec(&rq->nr_iowait);
4454 delayacct_blkio_end();
4459 * sys_sched_get_priority_max - return maximum RT priority.
4460 * @policy: scheduling class.
4462 * this syscall returns the maximum rt_priority that can be used
4463 * by a given scheduling class.
4465 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4472 ret = MAX_USER_RT_PRIO-1;
4484 * sys_sched_get_priority_min - return minimum RT priority.
4485 * @policy: scheduling class.
4487 * this syscall returns the minimum rt_priority that can be used
4488 * by a given scheduling class.
4490 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4508 * sys_sched_rr_get_interval - return the default timeslice of a process.
4509 * @pid: pid of the process.
4510 * @interval: userspace pointer to the timeslice value.
4512 * this syscall writes the default timeslice value of a given process
4513 * into the user-space timespec buffer. A value of '0' means infinity.
4515 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4516 struct timespec __user *, interval)
4518 struct task_struct *p;
4519 unsigned int time_slice;
4520 unsigned long flags;
4530 p = find_process_by_pid(pid);
4534 retval = security_task_getscheduler(p);
4538 rq = task_rq_lock(p, &flags);
4539 time_slice = p->sched_class->get_rr_interval(rq, p);
4540 task_rq_unlock(rq, p, &flags);
4543 jiffies_to_timespec(time_slice, &t);
4544 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4552 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4554 void sched_show_task(struct task_struct *p)
4556 unsigned long free = 0;
4560 state = p->state ? __ffs(p->state) + 1 : 0;
4561 printk(KERN_INFO "%-15.15s %c", p->comm,
4562 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4563 #if BITS_PER_LONG == 32
4564 if (state == TASK_RUNNING)
4565 printk(KERN_CONT " running ");
4567 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4569 if (state == TASK_RUNNING)
4570 printk(KERN_CONT " running task ");
4572 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4574 #ifdef CONFIG_DEBUG_STACK_USAGE
4575 free = stack_not_used(p);
4578 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4580 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4581 task_pid_nr(p), ppid,
4582 (unsigned long)task_thread_info(p)->flags);
4584 show_stack(p, NULL);
4587 void show_state_filter(unsigned long state_filter)
4589 struct task_struct *g, *p;
4591 #if BITS_PER_LONG == 32
4593 " task PC stack pid father\n");
4596 " task PC stack pid father\n");
4599 do_each_thread(g, p) {
4601 * reset the NMI-timeout, listing all files on a slow
4602 * console might take a lot of time:
4604 touch_nmi_watchdog();
4605 if (!state_filter || (p->state & state_filter))
4607 } while_each_thread(g, p);
4609 touch_all_softlockup_watchdogs();
4611 #ifdef CONFIG_SCHED_DEBUG
4612 sysrq_sched_debug_show();
4616 * Only show locks if all tasks are dumped:
4619 debug_show_all_locks();
4622 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4624 idle->sched_class = &idle_sched_class;
4628 * init_idle - set up an idle thread for a given CPU
4629 * @idle: task in question
4630 * @cpu: cpu the idle task belongs to
4632 * NOTE: this function does not set the idle thread's NEED_RESCHED
4633 * flag, to make booting more robust.
4635 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4637 struct rq *rq = cpu_rq(cpu);
4638 unsigned long flags;
4640 raw_spin_lock_irqsave(&rq->lock, flags);
4643 idle->state = TASK_RUNNING;
4644 idle->se.exec_start = sched_clock();
4646 do_set_cpus_allowed(idle, cpumask_of(cpu));
4648 * We're having a chicken and egg problem, even though we are
4649 * holding rq->lock, the cpu isn't yet set to this cpu so the
4650 * lockdep check in task_group() will fail.
4652 * Similar case to sched_fork(). / Alternatively we could
4653 * use task_rq_lock() here and obtain the other rq->lock.
4658 __set_task_cpu(idle, cpu);
4661 rq->curr = rq->idle = idle;
4662 #if defined(CONFIG_SMP)
4665 raw_spin_unlock_irqrestore(&rq->lock, flags);
4667 /* Set the preempt count _outside_ the spinlocks! */
4668 task_thread_info(idle)->preempt_count = 0;
4671 * The idle tasks have their own, simple scheduling class:
4673 idle->sched_class = &idle_sched_class;
4674 ftrace_graph_init_idle_task(idle, cpu);
4675 vtime_init_idle(idle);
4676 #if defined(CONFIG_SMP)
4677 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4682 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4684 if (p->sched_class && p->sched_class->set_cpus_allowed)
4685 p->sched_class->set_cpus_allowed(p, new_mask);
4687 cpumask_copy(&p->cpus_allowed, new_mask);
4688 p->nr_cpus_allowed = cpumask_weight(new_mask);
4692 * This is how migration works:
4694 * 1) we invoke migration_cpu_stop() on the target CPU using
4696 * 2) stopper starts to run (implicitly forcing the migrated thread
4698 * 3) it checks whether the migrated task is still in the wrong runqueue.
4699 * 4) if it's in the wrong runqueue then the migration thread removes
4700 * it and puts it into the right queue.
4701 * 5) stopper completes and stop_one_cpu() returns and the migration
4706 * Change a given task's CPU affinity. Migrate the thread to a
4707 * proper CPU and schedule it away if the CPU it's executing on
4708 * is removed from the allowed bitmask.
4710 * NOTE: the caller must have a valid reference to the task, the
4711 * task must not exit() & deallocate itself prematurely. The
4712 * call is not atomic; no spinlocks may be held.
4714 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4716 unsigned long flags;
4718 unsigned int dest_cpu;
4721 rq = task_rq_lock(p, &flags);
4723 if (cpumask_equal(&p->cpus_allowed, new_mask))
4726 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4731 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4736 do_set_cpus_allowed(p, new_mask);
4738 /* Can the task run on the task's current CPU? If so, we're done */
4739 if (cpumask_test_cpu(task_cpu(p), new_mask))
4742 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4744 struct migration_arg arg = { p, dest_cpu };
4745 /* Need help from migration thread: drop lock and wait. */
4746 task_rq_unlock(rq, p, &flags);
4747 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4748 tlb_migrate_finish(p->mm);
4752 task_rq_unlock(rq, p, &flags);
4756 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4759 * Move (not current) task off this cpu, onto dest cpu. We're doing
4760 * this because either it can't run here any more (set_cpus_allowed()
4761 * away from this CPU, or CPU going down), or because we're
4762 * attempting to rebalance this task on exec (sched_exec).
4764 * So we race with normal scheduler movements, but that's OK, as long
4765 * as the task is no longer on this CPU.
4767 * Returns non-zero if task was successfully migrated.
4769 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4771 struct rq *rq_dest, *rq_src;
4774 if (unlikely(!cpu_active(dest_cpu)))
4777 rq_src = cpu_rq(src_cpu);
4778 rq_dest = cpu_rq(dest_cpu);
4780 raw_spin_lock(&p->pi_lock);
4781 double_rq_lock(rq_src, rq_dest);
4782 /* Already moved. */
4783 if (task_cpu(p) != src_cpu)
4785 /* Affinity changed (again). */
4786 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4790 * If we're not on a rq, the next wake-up will ensure we're
4794 dequeue_task(rq_src, p, 0);
4795 set_task_cpu(p, dest_cpu);
4796 enqueue_task(rq_dest, p, 0);
4797 check_preempt_curr(rq_dest, p, 0);
4802 double_rq_unlock(rq_src, rq_dest);
4803 raw_spin_unlock(&p->pi_lock);
4808 * migration_cpu_stop - this will be executed by a highprio stopper thread
4809 * and performs thread migration by bumping thread off CPU then
4810 * 'pushing' onto another runqueue.
4812 static int migration_cpu_stop(void *data)
4814 struct migration_arg *arg = data;
4817 * The original target cpu might have gone down and we might
4818 * be on another cpu but it doesn't matter.
4820 local_irq_disable();
4821 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4826 #ifdef CONFIG_HOTPLUG_CPU
4829 * Ensures that the idle task is using init_mm right before its cpu goes
4832 void idle_task_exit(void)
4834 struct mm_struct *mm = current->active_mm;
4836 BUG_ON(cpu_online(smp_processor_id()));
4839 switch_mm(mm, &init_mm, current);
4844 * Since this CPU is going 'away' for a while, fold any nr_active delta
4845 * we might have. Assumes we're called after migrate_tasks() so that the
4846 * nr_active count is stable.
4848 * Also see the comment "Global load-average calculations".
4850 static void calc_load_migrate(struct rq *rq)
4852 long delta = calc_load_fold_active(rq);
4854 atomic_long_add(delta, &calc_load_tasks);
4858 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4859 * try_to_wake_up()->select_task_rq().
4861 * Called with rq->lock held even though we'er in stop_machine() and
4862 * there's no concurrency possible, we hold the required locks anyway
4863 * because of lock validation efforts.
4865 static void migrate_tasks(unsigned int dead_cpu)
4867 struct rq *rq = cpu_rq(dead_cpu);
4868 struct task_struct *next, *stop = rq->stop;
4872 * Fudge the rq selection such that the below task selection loop
4873 * doesn't get stuck on the currently eligible stop task.
4875 * We're currently inside stop_machine() and the rq is either stuck
4876 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4877 * either way we should never end up calling schedule() until we're
4884 * There's this thread running, bail when that's the only
4887 if (rq->nr_running == 1)
4890 next = pick_next_task(rq);
4892 next->sched_class->put_prev_task(rq, next);
4894 /* Find suitable destination for @next, with force if needed. */
4895 dest_cpu = select_fallback_rq(dead_cpu, next);
4896 raw_spin_unlock(&rq->lock);
4898 __migrate_task(next, dead_cpu, dest_cpu);
4900 raw_spin_lock(&rq->lock);
4906 #endif /* CONFIG_HOTPLUG_CPU */
4908 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4910 static struct ctl_table sd_ctl_dir[] = {
4912 .procname = "sched_domain",
4918 static struct ctl_table sd_ctl_root[] = {
4920 .procname = "kernel",
4922 .child = sd_ctl_dir,
4927 static struct ctl_table *sd_alloc_ctl_entry(int n)
4929 struct ctl_table *entry =
4930 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4935 static void sd_free_ctl_entry(struct ctl_table **tablep)
4937 struct ctl_table *entry;
4940 * In the intermediate directories, both the child directory and
4941 * procname are dynamically allocated and could fail but the mode
4942 * will always be set. In the lowest directory the names are
4943 * static strings and all have proc handlers.
4945 for (entry = *tablep; entry->mode; entry++) {
4947 sd_free_ctl_entry(&entry->child);
4948 if (entry->proc_handler == NULL)
4949 kfree(entry->procname);
4956 static int min_load_idx = 0;
4957 static int max_load_idx = CPU_LOAD_IDX_MAX;
4960 set_table_entry(struct ctl_table *entry,
4961 const char *procname, void *data, int maxlen,
4962 umode_t mode, proc_handler *proc_handler,
4965 entry->procname = procname;
4967 entry->maxlen = maxlen;
4969 entry->proc_handler = proc_handler;
4972 entry->extra1 = &min_load_idx;
4973 entry->extra2 = &max_load_idx;
4977 static struct ctl_table *
4978 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4980 struct ctl_table *table = sd_alloc_ctl_entry(13);
4985 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4986 sizeof(long), 0644, proc_doulongvec_minmax, false);
4987 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4988 sizeof(long), 0644, proc_doulongvec_minmax, false);
4989 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4990 sizeof(int), 0644, proc_dointvec_minmax, true);
4991 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4992 sizeof(int), 0644, proc_dointvec_minmax, true);
4993 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4994 sizeof(int), 0644, proc_dointvec_minmax, true);
4995 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4996 sizeof(int), 0644, proc_dointvec_minmax, true);
4997 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4998 sizeof(int), 0644, proc_dointvec_minmax, true);
4999 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5000 sizeof(int), 0644, proc_dointvec_minmax, false);
5001 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5002 sizeof(int), 0644, proc_dointvec_minmax, false);
5003 set_table_entry(&table[9], "cache_nice_tries",
5004 &sd->cache_nice_tries,
5005 sizeof(int), 0644, proc_dointvec_minmax, false);
5006 set_table_entry(&table[10], "flags", &sd->flags,
5007 sizeof(int), 0644, proc_dointvec_minmax, false);
5008 set_table_entry(&table[11], "name", sd->name,
5009 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5010 /* &table[12] is terminator */
5015 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5017 struct ctl_table *entry, *table;
5018 struct sched_domain *sd;
5019 int domain_num = 0, i;
5022 for_each_domain(cpu, sd)
5024 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5029 for_each_domain(cpu, sd) {
5030 snprintf(buf, 32, "domain%d", i);
5031 entry->procname = kstrdup(buf, GFP_KERNEL);
5033 entry->child = sd_alloc_ctl_domain_table(sd);
5040 static struct ctl_table_header *sd_sysctl_header;
5041 static void register_sched_domain_sysctl(void)
5043 int i, cpu_num = num_possible_cpus();
5044 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5047 WARN_ON(sd_ctl_dir[0].child);
5048 sd_ctl_dir[0].child = entry;
5053 for_each_possible_cpu(i) {
5054 snprintf(buf, 32, "cpu%d", i);
5055 entry->procname = kstrdup(buf, GFP_KERNEL);
5057 entry->child = sd_alloc_ctl_cpu_table(i);
5061 WARN_ON(sd_sysctl_header);
5062 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5065 /* may be called multiple times per register */
5066 static void unregister_sched_domain_sysctl(void)
5068 if (sd_sysctl_header)
5069 unregister_sysctl_table(sd_sysctl_header);
5070 sd_sysctl_header = NULL;
5071 if (sd_ctl_dir[0].child)
5072 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5075 static void register_sched_domain_sysctl(void)
5078 static void unregister_sched_domain_sysctl(void)
5083 static void set_rq_online(struct rq *rq)
5086 const struct sched_class *class;
5088 cpumask_set_cpu(rq->cpu, rq->rd->online);
5091 for_each_class(class) {
5092 if (class->rq_online)
5093 class->rq_online(rq);
5098 static void set_rq_offline(struct rq *rq)
5101 const struct sched_class *class;
5103 for_each_class(class) {
5104 if (class->rq_offline)
5105 class->rq_offline(rq);
5108 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5114 * migration_call - callback that gets triggered when a CPU is added.
5115 * Here we can start up the necessary migration thread for the new CPU.
5117 static int __cpuinit
5118 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5120 int cpu = (long)hcpu;
5121 unsigned long flags;
5122 struct rq *rq = cpu_rq(cpu);
5124 switch (action & ~CPU_TASKS_FROZEN) {
5126 case CPU_UP_PREPARE:
5127 rq->calc_load_update = calc_load_update;
5131 /* Update our root-domain */
5132 raw_spin_lock_irqsave(&rq->lock, flags);
5134 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5138 raw_spin_unlock_irqrestore(&rq->lock, flags);
5141 #ifdef CONFIG_HOTPLUG_CPU
5143 sched_ttwu_pending();
5144 /* Update our root-domain */
5145 raw_spin_lock_irqsave(&rq->lock, flags);
5147 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5151 BUG_ON(rq->nr_running != 1); /* the migration thread */
5152 raw_spin_unlock_irqrestore(&rq->lock, flags);
5156 calc_load_migrate(rq);
5161 update_max_interval();
5167 * Register at high priority so that task migration (migrate_all_tasks)
5168 * happens before everything else. This has to be lower priority than
5169 * the notifier in the perf_event subsystem, though.
5171 static struct notifier_block __cpuinitdata migration_notifier = {
5172 .notifier_call = migration_call,
5173 .priority = CPU_PRI_MIGRATION,
5176 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5177 unsigned long action, void *hcpu)
5179 switch (action & ~CPU_TASKS_FROZEN) {
5181 case CPU_DOWN_FAILED:
5182 set_cpu_active((long)hcpu, true);
5189 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5190 unsigned long action, void *hcpu)
5192 switch (action & ~CPU_TASKS_FROZEN) {
5193 case CPU_DOWN_PREPARE:
5194 set_cpu_active((long)hcpu, false);
5201 static int __init migration_init(void)
5203 void *cpu = (void *)(long)smp_processor_id();
5206 /* Initialize migration for the boot CPU */
5207 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5208 BUG_ON(err == NOTIFY_BAD);
5209 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5210 register_cpu_notifier(&migration_notifier);
5212 /* Register cpu active notifiers */
5213 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5214 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5218 early_initcall(migration_init);
5223 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5225 #ifdef CONFIG_SCHED_DEBUG
5227 static __read_mostly int sched_debug_enabled;
5229 static int __init sched_debug_setup(char *str)
5231 sched_debug_enabled = 1;
5235 early_param("sched_debug", sched_debug_setup);
5237 static inline bool sched_debug(void)
5239 return sched_debug_enabled;
5242 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5243 struct cpumask *groupmask)
5245 struct sched_group *group = sd->groups;
5248 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5249 cpumask_clear(groupmask);
5251 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5253 if (!(sd->flags & SD_LOAD_BALANCE)) {
5254 printk("does not load-balance\n");
5256 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5261 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5263 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5264 printk(KERN_ERR "ERROR: domain->span does not contain "
5267 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5268 printk(KERN_ERR "ERROR: domain->groups does not contain"
5272 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5276 printk(KERN_ERR "ERROR: group is NULL\n");
5281 * Even though we initialize ->power to something semi-sane,
5282 * we leave power_orig unset. This allows us to detect if
5283 * domain iteration is still funny without causing /0 traps.
5285 if (!group->sgp->power_orig) {
5286 printk(KERN_CONT "\n");
5287 printk(KERN_ERR "ERROR: domain->cpu_power not "
5292 if (!cpumask_weight(sched_group_cpus(group))) {
5293 printk(KERN_CONT "\n");
5294 printk(KERN_ERR "ERROR: empty group\n");
5298 if (!(sd->flags & SD_OVERLAP) &&
5299 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5300 printk(KERN_CONT "\n");
5301 printk(KERN_ERR "ERROR: repeated CPUs\n");
5305 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5307 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5309 printk(KERN_CONT " %s", str);
5310 if (group->sgp->power != SCHED_POWER_SCALE) {
5311 printk(KERN_CONT " (cpu_power = %d)",
5315 group = group->next;
5316 } while (group != sd->groups);
5317 printk(KERN_CONT "\n");
5319 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5320 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5323 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5324 printk(KERN_ERR "ERROR: parent span is not a superset "
5325 "of domain->span\n");
5329 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5333 if (!sched_debug_enabled)
5337 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5341 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5344 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5352 #else /* !CONFIG_SCHED_DEBUG */
5353 # define sched_domain_debug(sd, cpu) do { } while (0)
5354 static inline bool sched_debug(void)
5358 #endif /* CONFIG_SCHED_DEBUG */
5360 static int sd_degenerate(struct sched_domain *sd)
5362 if (cpumask_weight(sched_domain_span(sd)) == 1)
5365 /* Following flags need at least 2 groups */
5366 if (sd->flags & (SD_LOAD_BALANCE |
5367 SD_BALANCE_NEWIDLE |
5371 SD_SHARE_PKG_RESOURCES)) {
5372 if (sd->groups != sd->groups->next)
5376 /* Following flags don't use groups */
5377 if (sd->flags & (SD_WAKE_AFFINE))
5384 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5386 unsigned long cflags = sd->flags, pflags = parent->flags;
5388 if (sd_degenerate(parent))
5391 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5394 /* Flags needing groups don't count if only 1 group in parent */
5395 if (parent->groups == parent->groups->next) {
5396 pflags &= ~(SD_LOAD_BALANCE |
5397 SD_BALANCE_NEWIDLE |
5401 SD_SHARE_PKG_RESOURCES);
5402 if (nr_node_ids == 1)
5403 pflags &= ~SD_SERIALIZE;
5405 if (~cflags & pflags)
5411 static void free_rootdomain(struct rcu_head *rcu)
5413 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5415 cpupri_cleanup(&rd->cpupri);
5416 free_cpumask_var(rd->rto_mask);
5417 free_cpumask_var(rd->online);
5418 free_cpumask_var(rd->span);
5422 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5424 struct root_domain *old_rd = NULL;
5425 unsigned long flags;
5427 raw_spin_lock_irqsave(&rq->lock, flags);
5432 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5435 cpumask_clear_cpu(rq->cpu, old_rd->span);
5438 * If we dont want to free the old_rt yet then
5439 * set old_rd to NULL to skip the freeing later
5442 if (!atomic_dec_and_test(&old_rd->refcount))
5446 atomic_inc(&rd->refcount);
5449 cpumask_set_cpu(rq->cpu, rd->span);
5450 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5453 raw_spin_unlock_irqrestore(&rq->lock, flags);
5456 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5459 static int init_rootdomain(struct root_domain *rd)
5461 memset(rd, 0, sizeof(*rd));
5463 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5465 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5467 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5470 if (cpupri_init(&rd->cpupri) != 0)
5475 free_cpumask_var(rd->rto_mask);
5477 free_cpumask_var(rd->online);
5479 free_cpumask_var(rd->span);
5485 * By default the system creates a single root-domain with all cpus as
5486 * members (mimicking the global state we have today).
5488 struct root_domain def_root_domain;
5490 static void init_defrootdomain(void)
5492 init_rootdomain(&def_root_domain);
5494 atomic_set(&def_root_domain.refcount, 1);
5497 static struct root_domain *alloc_rootdomain(void)
5499 struct root_domain *rd;
5501 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5505 if (init_rootdomain(rd) != 0) {
5513 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5515 struct sched_group *tmp, *first;
5524 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5529 } while (sg != first);
5532 static void free_sched_domain(struct rcu_head *rcu)
5534 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5537 * If its an overlapping domain it has private groups, iterate and
5540 if (sd->flags & SD_OVERLAP) {
5541 free_sched_groups(sd->groups, 1);
5542 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5543 kfree(sd->groups->sgp);
5549 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5551 call_rcu(&sd->rcu, free_sched_domain);
5554 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5556 for (; sd; sd = sd->parent)
5557 destroy_sched_domain(sd, cpu);
5561 * Keep a special pointer to the highest sched_domain that has
5562 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5563 * allows us to avoid some pointer chasing select_idle_sibling().
5565 * Also keep a unique ID per domain (we use the first cpu number in
5566 * the cpumask of the domain), this allows us to quickly tell if
5567 * two cpus are in the same cache domain, see cpus_share_cache().
5569 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5570 DEFINE_PER_CPU(int, sd_llc_id);
5572 static void update_top_cache_domain(int cpu)
5574 struct sched_domain *sd;
5577 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5579 id = cpumask_first(sched_domain_span(sd));
5581 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5582 per_cpu(sd_llc_id, cpu) = id;
5586 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5587 * hold the hotplug lock.
5590 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5592 struct rq *rq = cpu_rq(cpu);
5593 struct sched_domain *tmp;
5595 /* Remove the sched domains which do not contribute to scheduling. */
5596 for (tmp = sd; tmp; ) {
5597 struct sched_domain *parent = tmp->parent;
5601 if (sd_parent_degenerate(tmp, parent)) {
5602 tmp->parent = parent->parent;
5604 parent->parent->child = tmp;
5605 destroy_sched_domain(parent, cpu);
5610 if (sd && sd_degenerate(sd)) {
5613 destroy_sched_domain(tmp, cpu);
5618 sched_domain_debug(sd, cpu);
5620 rq_attach_root(rq, rd);
5622 rcu_assign_pointer(rq->sd, sd);
5623 destroy_sched_domains(tmp, cpu);
5625 update_top_cache_domain(cpu);
5628 /* cpus with isolated domains */
5629 static cpumask_var_t cpu_isolated_map;
5631 /* Setup the mask of cpus configured for isolated domains */
5632 static int __init isolated_cpu_setup(char *str)
5634 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5635 cpulist_parse(str, cpu_isolated_map);
5639 __setup("isolcpus=", isolated_cpu_setup);
5641 static const struct cpumask *cpu_cpu_mask(int cpu)
5643 return cpumask_of_node(cpu_to_node(cpu));
5647 struct sched_domain **__percpu sd;
5648 struct sched_group **__percpu sg;
5649 struct sched_group_power **__percpu sgp;
5653 struct sched_domain ** __percpu sd;
5654 struct root_domain *rd;
5664 struct sched_domain_topology_level;
5666 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5667 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5669 #define SDTL_OVERLAP 0x01
5671 struct sched_domain_topology_level {
5672 sched_domain_init_f init;
5673 sched_domain_mask_f mask;
5676 struct sd_data data;
5680 * Build an iteration mask that can exclude certain CPUs from the upwards
5683 * Asymmetric node setups can result in situations where the domain tree is of
5684 * unequal depth, make sure to skip domains that already cover the entire
5687 * In that case build_sched_domains() will have terminated the iteration early
5688 * and our sibling sd spans will be empty. Domains should always include the
5689 * cpu they're built on, so check that.
5692 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5694 const struct cpumask *span = sched_domain_span(sd);
5695 struct sd_data *sdd = sd->private;
5696 struct sched_domain *sibling;
5699 for_each_cpu(i, span) {
5700 sibling = *per_cpu_ptr(sdd->sd, i);
5701 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5704 cpumask_set_cpu(i, sched_group_mask(sg));
5709 * Return the canonical balance cpu for this group, this is the first cpu
5710 * of this group that's also in the iteration mask.
5712 int group_balance_cpu(struct sched_group *sg)
5714 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5718 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5720 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5721 const struct cpumask *span = sched_domain_span(sd);
5722 struct cpumask *covered = sched_domains_tmpmask;
5723 struct sd_data *sdd = sd->private;
5724 struct sched_domain *child;
5727 cpumask_clear(covered);
5729 for_each_cpu(i, span) {
5730 struct cpumask *sg_span;
5732 if (cpumask_test_cpu(i, covered))
5735 child = *per_cpu_ptr(sdd->sd, i);
5737 /* See the comment near build_group_mask(). */
5738 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5741 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5742 GFP_KERNEL, cpu_to_node(cpu));
5747 sg_span = sched_group_cpus(sg);
5749 child = child->child;
5750 cpumask_copy(sg_span, sched_domain_span(child));
5752 cpumask_set_cpu(i, sg_span);
5754 cpumask_or(covered, covered, sg_span);
5756 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5757 if (atomic_inc_return(&sg->sgp->ref) == 1)
5758 build_group_mask(sd, sg);
5761 * Initialize sgp->power such that even if we mess up the
5762 * domains and no possible iteration will get us here, we won't
5765 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5768 * Make sure the first group of this domain contains the
5769 * canonical balance cpu. Otherwise the sched_domain iteration
5770 * breaks. See update_sg_lb_stats().
5772 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5773 group_balance_cpu(sg) == cpu)
5783 sd->groups = groups;
5788 free_sched_groups(first, 0);
5793 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5795 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5796 struct sched_domain *child = sd->child;
5799 cpu = cpumask_first(sched_domain_span(child));
5802 *sg = *per_cpu_ptr(sdd->sg, cpu);
5803 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5804 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5811 * build_sched_groups will build a circular linked list of the groups
5812 * covered by the given span, and will set each group's ->cpumask correctly,
5813 * and ->cpu_power to 0.
5815 * Assumes the sched_domain tree is fully constructed
5818 build_sched_groups(struct sched_domain *sd, int cpu)
5820 struct sched_group *first = NULL, *last = NULL;
5821 struct sd_data *sdd = sd->private;
5822 const struct cpumask *span = sched_domain_span(sd);
5823 struct cpumask *covered;
5826 get_group(cpu, sdd, &sd->groups);
5827 atomic_inc(&sd->groups->ref);
5829 if (cpu != cpumask_first(sched_domain_span(sd)))
5832 lockdep_assert_held(&sched_domains_mutex);
5833 covered = sched_domains_tmpmask;
5835 cpumask_clear(covered);
5837 for_each_cpu(i, span) {
5838 struct sched_group *sg;
5839 int group = get_group(i, sdd, &sg);
5842 if (cpumask_test_cpu(i, covered))
5845 cpumask_clear(sched_group_cpus(sg));
5847 cpumask_setall(sched_group_mask(sg));
5849 for_each_cpu(j, span) {
5850 if (get_group(j, sdd, NULL) != group)
5853 cpumask_set_cpu(j, covered);
5854 cpumask_set_cpu(j, sched_group_cpus(sg));
5869 * Initialize sched groups cpu_power.
5871 * cpu_power indicates the capacity of sched group, which is used while
5872 * distributing the load between different sched groups in a sched domain.
5873 * Typically cpu_power for all the groups in a sched domain will be same unless
5874 * there are asymmetries in the topology. If there are asymmetries, group
5875 * having more cpu_power will pickup more load compared to the group having
5878 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5880 struct sched_group *sg = sd->groups;
5882 WARN_ON(!sd || !sg);
5885 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5887 } while (sg != sd->groups);
5889 if (cpu != group_balance_cpu(sg))
5892 update_group_power(sd, cpu);
5893 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5896 int __weak arch_sd_sibling_asym_packing(void)
5898 return 0*SD_ASYM_PACKING;
5902 * Initializers for schedule domains
5903 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5906 #ifdef CONFIG_SCHED_DEBUG
5907 # define SD_INIT_NAME(sd, type) sd->name = #type
5909 # define SD_INIT_NAME(sd, type) do { } while (0)
5912 #define SD_INIT_FUNC(type) \
5913 static noinline struct sched_domain * \
5914 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5916 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5917 *sd = SD_##type##_INIT; \
5918 SD_INIT_NAME(sd, type); \
5919 sd->private = &tl->data; \
5924 #ifdef CONFIG_SCHED_SMT
5925 SD_INIT_FUNC(SIBLING)
5927 #ifdef CONFIG_SCHED_MC
5930 #ifdef CONFIG_SCHED_BOOK
5934 static int default_relax_domain_level = -1;
5935 int sched_domain_level_max;
5937 static int __init setup_relax_domain_level(char *str)
5939 if (kstrtoint(str, 0, &default_relax_domain_level))
5940 pr_warn("Unable to set relax_domain_level\n");
5944 __setup("relax_domain_level=", setup_relax_domain_level);
5946 static void set_domain_attribute(struct sched_domain *sd,
5947 struct sched_domain_attr *attr)
5951 if (!attr || attr->relax_domain_level < 0) {
5952 if (default_relax_domain_level < 0)
5955 request = default_relax_domain_level;
5957 request = attr->relax_domain_level;
5958 if (request < sd->level) {
5959 /* turn off idle balance on this domain */
5960 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5962 /* turn on idle balance on this domain */
5963 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5967 static void __sdt_free(const struct cpumask *cpu_map);
5968 static int __sdt_alloc(const struct cpumask *cpu_map);
5970 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5971 const struct cpumask *cpu_map)
5975 if (!atomic_read(&d->rd->refcount))
5976 free_rootdomain(&d->rd->rcu); /* fall through */
5978 free_percpu(d->sd); /* fall through */
5980 __sdt_free(cpu_map); /* fall through */
5986 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5987 const struct cpumask *cpu_map)
5989 memset(d, 0, sizeof(*d));
5991 if (__sdt_alloc(cpu_map))
5992 return sa_sd_storage;
5993 d->sd = alloc_percpu(struct sched_domain *);
5995 return sa_sd_storage;
5996 d->rd = alloc_rootdomain();
5999 return sa_rootdomain;
6003 * NULL the sd_data elements we've used to build the sched_domain and
6004 * sched_group structure so that the subsequent __free_domain_allocs()
6005 * will not free the data we're using.
6007 static void claim_allocations(int cpu, struct sched_domain *sd)
6009 struct sd_data *sdd = sd->private;
6011 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6012 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6014 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6015 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6017 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6018 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6021 #ifdef CONFIG_SCHED_SMT
6022 static const struct cpumask *cpu_smt_mask(int cpu)
6024 return topology_thread_cpumask(cpu);
6029 * Topology list, bottom-up.
6031 static struct sched_domain_topology_level default_topology[] = {
6032 #ifdef CONFIG_SCHED_SMT
6033 { sd_init_SIBLING, cpu_smt_mask, },
6035 #ifdef CONFIG_SCHED_MC
6036 { sd_init_MC, cpu_coregroup_mask, },
6038 #ifdef CONFIG_SCHED_BOOK
6039 { sd_init_BOOK, cpu_book_mask, },
6041 { sd_init_CPU, cpu_cpu_mask, },
6045 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6049 static int sched_domains_numa_levels;
6050 static int *sched_domains_numa_distance;
6051 static struct cpumask ***sched_domains_numa_masks;
6052 static int sched_domains_curr_level;
6054 static inline int sd_local_flags(int level)
6056 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6059 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6062 static struct sched_domain *
6063 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6065 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6066 int level = tl->numa_level;
6067 int sd_weight = cpumask_weight(
6068 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6070 *sd = (struct sched_domain){
6071 .min_interval = sd_weight,
6072 .max_interval = 2*sd_weight,
6074 .imbalance_pct = 125,
6075 .cache_nice_tries = 2,
6082 .flags = 1*SD_LOAD_BALANCE
6083 | 1*SD_BALANCE_NEWIDLE
6088 | 0*SD_SHARE_CPUPOWER
6089 | 0*SD_SHARE_PKG_RESOURCES
6091 | 0*SD_PREFER_SIBLING
6092 | sd_local_flags(level)
6094 .last_balance = jiffies,
6095 .balance_interval = sd_weight,
6097 SD_INIT_NAME(sd, NUMA);
6098 sd->private = &tl->data;
6101 * Ugly hack to pass state to sd_numa_mask()...
6103 sched_domains_curr_level = tl->numa_level;
6108 static const struct cpumask *sd_numa_mask(int cpu)
6110 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6113 static void sched_numa_warn(const char *str)
6115 static int done = false;
6123 printk(KERN_WARNING "ERROR: %s\n\n", str);
6125 for (i = 0; i < nr_node_ids; i++) {
6126 printk(KERN_WARNING " ");
6127 for (j = 0; j < nr_node_ids; j++)
6128 printk(KERN_CONT "%02d ", node_distance(i,j));
6129 printk(KERN_CONT "\n");
6131 printk(KERN_WARNING "\n");
6134 static bool find_numa_distance(int distance)
6138 if (distance == node_distance(0, 0))
6141 for (i = 0; i < sched_domains_numa_levels; i++) {
6142 if (sched_domains_numa_distance[i] == distance)
6149 static void sched_init_numa(void)
6151 int next_distance, curr_distance = node_distance(0, 0);
6152 struct sched_domain_topology_level *tl;
6156 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6157 if (!sched_domains_numa_distance)
6161 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6162 * unique distances in the node_distance() table.
6164 * Assumes node_distance(0,j) includes all distances in
6165 * node_distance(i,j) in order to avoid cubic time.
6167 next_distance = curr_distance;
6168 for (i = 0; i < nr_node_ids; i++) {
6169 for (j = 0; j < nr_node_ids; j++) {
6170 for (k = 0; k < nr_node_ids; k++) {
6171 int distance = node_distance(i, k);
6173 if (distance > curr_distance &&
6174 (distance < next_distance ||
6175 next_distance == curr_distance))
6176 next_distance = distance;
6179 * While not a strong assumption it would be nice to know
6180 * about cases where if node A is connected to B, B is not
6181 * equally connected to A.
6183 if (sched_debug() && node_distance(k, i) != distance)
6184 sched_numa_warn("Node-distance not symmetric");
6186 if (sched_debug() && i && !find_numa_distance(distance))
6187 sched_numa_warn("Node-0 not representative");
6189 if (next_distance != curr_distance) {
6190 sched_domains_numa_distance[level++] = next_distance;
6191 sched_domains_numa_levels = level;
6192 curr_distance = next_distance;
6197 * In case of sched_debug() we verify the above assumption.
6203 * 'level' contains the number of unique distances, excluding the
6204 * identity distance node_distance(i,i).
6206 * The sched_domains_nume_distance[] array includes the actual distance
6211 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6212 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6213 * the array will contain less then 'level' members. This could be
6214 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6215 * in other functions.
6217 * We reset it to 'level' at the end of this function.
6219 sched_domains_numa_levels = 0;
6221 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6222 if (!sched_domains_numa_masks)
6226 * Now for each level, construct a mask per node which contains all
6227 * cpus of nodes that are that many hops away from us.
6229 for (i = 0; i < level; i++) {
6230 sched_domains_numa_masks[i] =
6231 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6232 if (!sched_domains_numa_masks[i])
6235 for (j = 0; j < nr_node_ids; j++) {
6236 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6240 sched_domains_numa_masks[i][j] = mask;
6242 for (k = 0; k < nr_node_ids; k++) {
6243 if (node_distance(j, k) > sched_domains_numa_distance[i])
6246 cpumask_or(mask, mask, cpumask_of_node(k));
6251 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6252 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6257 * Copy the default topology bits..
6259 for (i = 0; default_topology[i].init; i++)
6260 tl[i] = default_topology[i];
6263 * .. and append 'j' levels of NUMA goodness.
6265 for (j = 0; j < level; i++, j++) {
6266 tl[i] = (struct sched_domain_topology_level){
6267 .init = sd_numa_init,
6268 .mask = sd_numa_mask,
6269 .flags = SDTL_OVERLAP,
6274 sched_domain_topology = tl;
6276 sched_domains_numa_levels = level;
6279 static void sched_domains_numa_masks_set(int cpu)
6282 int node = cpu_to_node(cpu);
6284 for (i = 0; i < sched_domains_numa_levels; i++) {
6285 for (j = 0; j < nr_node_ids; j++) {
6286 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6287 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6292 static void sched_domains_numa_masks_clear(int cpu)
6295 for (i = 0; i < sched_domains_numa_levels; i++) {
6296 for (j = 0; j < nr_node_ids; j++)
6297 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6302 * Update sched_domains_numa_masks[level][node] array when new cpus
6305 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6306 unsigned long action,
6309 int cpu = (long)hcpu;
6311 switch (action & ~CPU_TASKS_FROZEN) {
6313 sched_domains_numa_masks_set(cpu);
6317 sched_domains_numa_masks_clear(cpu);
6327 static inline void sched_init_numa(void)
6331 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6332 unsigned long action,
6337 #endif /* CONFIG_NUMA */
6339 static int __sdt_alloc(const struct cpumask *cpu_map)
6341 struct sched_domain_topology_level *tl;
6344 for (tl = sched_domain_topology; tl->init; tl++) {
6345 struct sd_data *sdd = &tl->data;
6347 sdd->sd = alloc_percpu(struct sched_domain *);
6351 sdd->sg = alloc_percpu(struct sched_group *);
6355 sdd->sgp = alloc_percpu(struct sched_group_power *);
6359 for_each_cpu(j, cpu_map) {
6360 struct sched_domain *sd;
6361 struct sched_group *sg;
6362 struct sched_group_power *sgp;
6364 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6365 GFP_KERNEL, cpu_to_node(j));
6369 *per_cpu_ptr(sdd->sd, j) = sd;
6371 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6372 GFP_KERNEL, cpu_to_node(j));
6378 *per_cpu_ptr(sdd->sg, j) = sg;
6380 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6381 GFP_KERNEL, cpu_to_node(j));
6385 *per_cpu_ptr(sdd->sgp, j) = sgp;
6392 static void __sdt_free(const struct cpumask *cpu_map)
6394 struct sched_domain_topology_level *tl;
6397 for (tl = sched_domain_topology; tl->init; tl++) {
6398 struct sd_data *sdd = &tl->data;
6400 for_each_cpu(j, cpu_map) {
6401 struct sched_domain *sd;
6404 sd = *per_cpu_ptr(sdd->sd, j);
6405 if (sd && (sd->flags & SD_OVERLAP))
6406 free_sched_groups(sd->groups, 0);
6407 kfree(*per_cpu_ptr(sdd->sd, j));
6411 kfree(*per_cpu_ptr(sdd->sg, j));
6413 kfree(*per_cpu_ptr(sdd->sgp, j));
6415 free_percpu(sdd->sd);
6417 free_percpu(sdd->sg);
6419 free_percpu(sdd->sgp);
6424 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6425 struct s_data *d, const struct cpumask *cpu_map,
6426 struct sched_domain_attr *attr, struct sched_domain *child,
6429 struct sched_domain *sd = tl->init(tl, cpu);
6433 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6435 sd->level = child->level + 1;
6436 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6440 set_domain_attribute(sd, attr);
6446 * Build sched domains for a given set of cpus and attach the sched domains
6447 * to the individual cpus
6449 static int build_sched_domains(const struct cpumask *cpu_map,
6450 struct sched_domain_attr *attr)
6452 enum s_alloc alloc_state = sa_none;
6453 struct sched_domain *sd;
6455 int i, ret = -ENOMEM;
6457 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6458 if (alloc_state != sa_rootdomain)
6461 /* Set up domains for cpus specified by the cpu_map. */
6462 for_each_cpu(i, cpu_map) {
6463 struct sched_domain_topology_level *tl;
6466 for (tl = sched_domain_topology; tl->init; tl++) {
6467 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6468 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6469 sd->flags |= SD_OVERLAP;
6470 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6477 *per_cpu_ptr(d.sd, i) = sd;
6480 /* Build the groups for the domains */
6481 for_each_cpu(i, cpu_map) {
6482 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6483 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6484 if (sd->flags & SD_OVERLAP) {
6485 if (build_overlap_sched_groups(sd, i))
6488 if (build_sched_groups(sd, i))
6494 /* Calculate CPU power for physical packages and nodes */
6495 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6496 if (!cpumask_test_cpu(i, cpu_map))
6499 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6500 claim_allocations(i, sd);
6501 init_sched_groups_power(i, sd);
6505 /* Attach the domains */
6507 for_each_cpu(i, cpu_map) {
6508 sd = *per_cpu_ptr(d.sd, i);
6509 cpu_attach_domain(sd, d.rd, i);
6515 __free_domain_allocs(&d, alloc_state, cpu_map);
6519 static cpumask_var_t *doms_cur; /* current sched domains */
6520 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6521 static struct sched_domain_attr *dattr_cur;
6522 /* attribues of custom domains in 'doms_cur' */
6525 * Special case: If a kmalloc of a doms_cur partition (array of
6526 * cpumask) fails, then fallback to a single sched domain,
6527 * as determined by the single cpumask fallback_doms.
6529 static cpumask_var_t fallback_doms;
6532 * arch_update_cpu_topology lets virtualized architectures update the
6533 * cpu core maps. It is supposed to return 1 if the topology changed
6534 * or 0 if it stayed the same.
6536 int __attribute__((weak)) arch_update_cpu_topology(void)
6541 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6544 cpumask_var_t *doms;
6546 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6549 for (i = 0; i < ndoms; i++) {
6550 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6551 free_sched_domains(doms, i);
6558 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6561 for (i = 0; i < ndoms; i++)
6562 free_cpumask_var(doms[i]);
6567 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6568 * For now this just excludes isolated cpus, but could be used to
6569 * exclude other special cases in the future.
6571 static int init_sched_domains(const struct cpumask *cpu_map)
6575 arch_update_cpu_topology();
6577 doms_cur = alloc_sched_domains(ndoms_cur);
6579 doms_cur = &fallback_doms;
6580 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6581 err = build_sched_domains(doms_cur[0], NULL);
6582 register_sched_domain_sysctl();
6588 * Detach sched domains from a group of cpus specified in cpu_map
6589 * These cpus will now be attached to the NULL domain
6591 static void detach_destroy_domains(const struct cpumask *cpu_map)
6596 for_each_cpu(i, cpu_map)
6597 cpu_attach_domain(NULL, &def_root_domain, i);
6601 /* handle null as "default" */
6602 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6603 struct sched_domain_attr *new, int idx_new)
6605 struct sched_domain_attr tmp;
6612 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6613 new ? (new + idx_new) : &tmp,
6614 sizeof(struct sched_domain_attr));
6618 * Partition sched domains as specified by the 'ndoms_new'
6619 * cpumasks in the array doms_new[] of cpumasks. This compares
6620 * doms_new[] to the current sched domain partitioning, doms_cur[].
6621 * It destroys each deleted domain and builds each new domain.
6623 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6624 * The masks don't intersect (don't overlap.) We should setup one
6625 * sched domain for each mask. CPUs not in any of the cpumasks will
6626 * not be load balanced. If the same cpumask appears both in the
6627 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6630 * The passed in 'doms_new' should be allocated using
6631 * alloc_sched_domains. This routine takes ownership of it and will
6632 * free_sched_domains it when done with it. If the caller failed the
6633 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6634 * and partition_sched_domains() will fallback to the single partition
6635 * 'fallback_doms', it also forces the domains to be rebuilt.
6637 * If doms_new == NULL it will be replaced with cpu_online_mask.
6638 * ndoms_new == 0 is a special case for destroying existing domains,
6639 * and it will not create the default domain.
6641 * Call with hotplug lock held
6643 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6644 struct sched_domain_attr *dattr_new)
6649 mutex_lock(&sched_domains_mutex);
6651 /* always unregister in case we don't destroy any domains */
6652 unregister_sched_domain_sysctl();
6654 /* Let architecture update cpu core mappings. */
6655 new_topology = arch_update_cpu_topology();
6657 n = doms_new ? ndoms_new : 0;
6659 /* Destroy deleted domains */
6660 for (i = 0; i < ndoms_cur; i++) {
6661 for (j = 0; j < n && !new_topology; j++) {
6662 if (cpumask_equal(doms_cur[i], doms_new[j])
6663 && dattrs_equal(dattr_cur, i, dattr_new, j))
6666 /* no match - a current sched domain not in new doms_new[] */
6667 detach_destroy_domains(doms_cur[i]);
6672 if (doms_new == NULL) {
6674 doms_new = &fallback_doms;
6675 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6676 WARN_ON_ONCE(dattr_new);
6679 /* Build new domains */
6680 for (i = 0; i < ndoms_new; i++) {
6681 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6682 if (cpumask_equal(doms_new[i], doms_cur[j])
6683 && dattrs_equal(dattr_new, i, dattr_cur, j))
6686 /* no match - add a new doms_new */
6687 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6692 /* Remember the new sched domains */
6693 if (doms_cur != &fallback_doms)
6694 free_sched_domains(doms_cur, ndoms_cur);
6695 kfree(dattr_cur); /* kfree(NULL) is safe */
6696 doms_cur = doms_new;
6697 dattr_cur = dattr_new;
6698 ndoms_cur = ndoms_new;
6700 register_sched_domain_sysctl();
6702 mutex_unlock(&sched_domains_mutex);
6705 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6708 * Update cpusets according to cpu_active mask. If cpusets are
6709 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6710 * around partition_sched_domains().
6712 * If we come here as part of a suspend/resume, don't touch cpusets because we
6713 * want to restore it back to its original state upon resume anyway.
6715 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6719 case CPU_ONLINE_FROZEN:
6720 case CPU_DOWN_FAILED_FROZEN:
6723 * num_cpus_frozen tracks how many CPUs are involved in suspend
6724 * resume sequence. As long as this is not the last online
6725 * operation in the resume sequence, just build a single sched
6726 * domain, ignoring cpusets.
6729 if (likely(num_cpus_frozen)) {
6730 partition_sched_domains(1, NULL, NULL);
6735 * This is the last CPU online operation. So fall through and
6736 * restore the original sched domains by considering the
6737 * cpuset configurations.
6741 case CPU_DOWN_FAILED:
6742 cpuset_update_active_cpus(true);
6750 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6754 case CPU_DOWN_PREPARE:
6755 cpuset_update_active_cpus(false);
6757 case CPU_DOWN_PREPARE_FROZEN:
6759 partition_sched_domains(1, NULL, NULL);
6767 void __init sched_init_smp(void)
6769 cpumask_var_t non_isolated_cpus;
6771 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6772 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6777 mutex_lock(&sched_domains_mutex);
6778 init_sched_domains(cpu_active_mask);
6779 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6780 if (cpumask_empty(non_isolated_cpus))
6781 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6782 mutex_unlock(&sched_domains_mutex);
6785 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6786 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6787 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6789 /* RT runtime code needs to handle some hotplug events */
6790 hotcpu_notifier(update_runtime, 0);
6794 /* Move init over to a non-isolated CPU */
6795 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6797 sched_init_granularity();
6798 free_cpumask_var(non_isolated_cpus);
6800 init_sched_rt_class();
6803 void __init sched_init_smp(void)
6805 sched_init_granularity();
6807 #endif /* CONFIG_SMP */
6809 const_debug unsigned int sysctl_timer_migration = 1;
6811 int in_sched_functions(unsigned long addr)
6813 return in_lock_functions(addr) ||
6814 (addr >= (unsigned long)__sched_text_start
6815 && addr < (unsigned long)__sched_text_end);
6818 #ifdef CONFIG_CGROUP_SCHED
6819 struct task_group root_task_group;
6820 LIST_HEAD(task_groups);
6823 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6825 void __init sched_init(void)
6828 unsigned long alloc_size = 0, ptr;
6830 #ifdef CONFIG_FAIR_GROUP_SCHED
6831 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6833 #ifdef CONFIG_RT_GROUP_SCHED
6834 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6836 #ifdef CONFIG_CPUMASK_OFFSTACK
6837 alloc_size += num_possible_cpus() * cpumask_size();
6840 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6842 #ifdef CONFIG_FAIR_GROUP_SCHED
6843 root_task_group.se = (struct sched_entity **)ptr;
6844 ptr += nr_cpu_ids * sizeof(void **);
6846 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6847 ptr += nr_cpu_ids * sizeof(void **);
6849 #endif /* CONFIG_FAIR_GROUP_SCHED */
6850 #ifdef CONFIG_RT_GROUP_SCHED
6851 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6852 ptr += nr_cpu_ids * sizeof(void **);
6854 root_task_group.rt_rq = (struct rt_rq **)ptr;
6855 ptr += nr_cpu_ids * sizeof(void **);
6857 #endif /* CONFIG_RT_GROUP_SCHED */
6858 #ifdef CONFIG_CPUMASK_OFFSTACK
6859 for_each_possible_cpu(i) {
6860 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6861 ptr += cpumask_size();
6863 #endif /* CONFIG_CPUMASK_OFFSTACK */
6867 init_defrootdomain();
6870 init_rt_bandwidth(&def_rt_bandwidth,
6871 global_rt_period(), global_rt_runtime());
6873 #ifdef CONFIG_RT_GROUP_SCHED
6874 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6875 global_rt_period(), global_rt_runtime());
6876 #endif /* CONFIG_RT_GROUP_SCHED */
6878 #ifdef CONFIG_CGROUP_SCHED
6879 list_add(&root_task_group.list, &task_groups);
6880 INIT_LIST_HEAD(&root_task_group.children);
6881 INIT_LIST_HEAD(&root_task_group.siblings);
6882 autogroup_init(&init_task);
6884 #endif /* CONFIG_CGROUP_SCHED */
6886 #ifdef CONFIG_CGROUP_CPUACCT
6887 root_cpuacct.cpustat = &kernel_cpustat;
6888 root_cpuacct.cpuusage = alloc_percpu(u64);
6889 /* Too early, not expected to fail */
6890 BUG_ON(!root_cpuacct.cpuusage);
6892 for_each_possible_cpu(i) {
6896 raw_spin_lock_init(&rq->lock);
6898 rq->calc_load_active = 0;
6899 rq->calc_load_update = jiffies + LOAD_FREQ;
6900 init_cfs_rq(&rq->cfs);
6901 init_rt_rq(&rq->rt, rq);
6902 #ifdef CONFIG_FAIR_GROUP_SCHED
6903 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6904 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6906 * How much cpu bandwidth does root_task_group get?
6908 * In case of task-groups formed thr' the cgroup filesystem, it
6909 * gets 100% of the cpu resources in the system. This overall
6910 * system cpu resource is divided among the tasks of
6911 * root_task_group and its child task-groups in a fair manner,
6912 * based on each entity's (task or task-group's) weight
6913 * (se->load.weight).
6915 * In other words, if root_task_group has 10 tasks of weight
6916 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6917 * then A0's share of the cpu resource is:
6919 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6921 * We achieve this by letting root_task_group's tasks sit
6922 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6924 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6925 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6926 #endif /* CONFIG_FAIR_GROUP_SCHED */
6928 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6929 #ifdef CONFIG_RT_GROUP_SCHED
6930 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6931 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6934 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6935 rq->cpu_load[j] = 0;
6937 rq->last_load_update_tick = jiffies;
6942 rq->cpu_power = SCHED_POWER_SCALE;
6943 rq->post_schedule = 0;
6944 rq->active_balance = 0;
6945 rq->next_balance = jiffies;
6950 rq->avg_idle = 2*sysctl_sched_migration_cost;
6952 INIT_LIST_HEAD(&rq->cfs_tasks);
6954 rq_attach_root(rq, &def_root_domain);
6960 atomic_set(&rq->nr_iowait, 0);
6963 set_load_weight(&init_task);
6965 #ifdef CONFIG_PREEMPT_NOTIFIERS
6966 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6969 #ifdef CONFIG_RT_MUTEXES
6970 plist_head_init(&init_task.pi_waiters);
6974 * The boot idle thread does lazy MMU switching as well:
6976 atomic_inc(&init_mm.mm_count);
6977 enter_lazy_tlb(&init_mm, current);
6980 * Make us the idle thread. Technically, schedule() should not be
6981 * called from this thread, however somewhere below it might be,
6982 * but because we are the idle thread, we just pick up running again
6983 * when this runqueue becomes "idle".
6985 init_idle(current, smp_processor_id());
6987 calc_load_update = jiffies + LOAD_FREQ;
6990 * During early bootup we pretend to be a normal task:
6992 current->sched_class = &fair_sched_class;
6995 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6996 /* May be allocated at isolcpus cmdline parse time */
6997 if (cpu_isolated_map == NULL)
6998 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6999 idle_thread_set_boot_cpu();
7001 init_sched_fair_class();
7003 scheduler_running = 1;
7006 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7007 static inline int preempt_count_equals(int preempt_offset)
7009 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7011 return (nested == preempt_offset);
7014 void __might_sleep(const char *file, int line, int preempt_offset)
7016 static unsigned long prev_jiffy; /* ratelimiting */
7018 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7019 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7020 system_state != SYSTEM_RUNNING || oops_in_progress)
7022 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7024 prev_jiffy = jiffies;
7027 "BUG: sleeping function called from invalid context at %s:%d\n",
7030 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7031 in_atomic(), irqs_disabled(),
7032 current->pid, current->comm);
7034 debug_show_held_locks(current);
7035 if (irqs_disabled())
7036 print_irqtrace_events(current);
7039 EXPORT_SYMBOL(__might_sleep);
7042 #ifdef CONFIG_MAGIC_SYSRQ
7043 static void normalize_task(struct rq *rq, struct task_struct *p)
7045 const struct sched_class *prev_class = p->sched_class;
7046 int old_prio = p->prio;
7051 dequeue_task(rq, p, 0);
7052 __setscheduler(rq, p, SCHED_NORMAL, 0);
7054 enqueue_task(rq, p, 0);
7055 resched_task(rq->curr);
7058 check_class_changed(rq, p, prev_class, old_prio);
7061 void normalize_rt_tasks(void)
7063 struct task_struct *g, *p;
7064 unsigned long flags;
7067 read_lock_irqsave(&tasklist_lock, flags);
7068 do_each_thread(g, p) {
7070 * Only normalize user tasks:
7075 p->se.exec_start = 0;
7076 #ifdef CONFIG_SCHEDSTATS
7077 p->se.statistics.wait_start = 0;
7078 p->se.statistics.sleep_start = 0;
7079 p->se.statistics.block_start = 0;
7084 * Renice negative nice level userspace
7087 if (TASK_NICE(p) < 0 && p->mm)
7088 set_user_nice(p, 0);
7092 raw_spin_lock(&p->pi_lock);
7093 rq = __task_rq_lock(p);
7095 normalize_task(rq, p);
7097 __task_rq_unlock(rq);
7098 raw_spin_unlock(&p->pi_lock);
7099 } while_each_thread(g, p);
7101 read_unlock_irqrestore(&tasklist_lock, flags);
7104 #endif /* CONFIG_MAGIC_SYSRQ */
7106 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7108 * These functions are only useful for the IA64 MCA handling, or kdb.
7110 * They can only be called when the whole system has been
7111 * stopped - every CPU needs to be quiescent, and no scheduling
7112 * activity can take place. Using them for anything else would
7113 * be a serious bug, and as a result, they aren't even visible
7114 * under any other configuration.
7118 * curr_task - return the current task for a given cpu.
7119 * @cpu: the processor in question.
7121 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7123 struct task_struct *curr_task(int cpu)
7125 return cpu_curr(cpu);
7128 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7132 * set_curr_task - set the current task for a given cpu.
7133 * @cpu: the processor in question.
7134 * @p: the task pointer to set.
7136 * Description: This function must only be used when non-maskable interrupts
7137 * are serviced on a separate stack. It allows the architecture to switch the
7138 * notion of the current task on a cpu in a non-blocking manner. This function
7139 * must be called with all CPU's synchronized, and interrupts disabled, the
7140 * and caller must save the original value of the current task (see
7141 * curr_task() above) and restore that value before reenabling interrupts and
7142 * re-starting the system.
7144 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7146 void set_curr_task(int cpu, struct task_struct *p)
7153 #ifdef CONFIG_CGROUP_SCHED
7154 /* task_group_lock serializes the addition/removal of task groups */
7155 static DEFINE_SPINLOCK(task_group_lock);
7157 static void free_sched_group(struct task_group *tg)
7159 free_fair_sched_group(tg);
7160 free_rt_sched_group(tg);
7165 /* allocate runqueue etc for a new task group */
7166 struct task_group *sched_create_group(struct task_group *parent)
7168 struct task_group *tg;
7170 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7172 return ERR_PTR(-ENOMEM);
7174 if (!alloc_fair_sched_group(tg, parent))
7177 if (!alloc_rt_sched_group(tg, parent))
7183 free_sched_group(tg);
7184 return ERR_PTR(-ENOMEM);
7187 void sched_online_group(struct task_group *tg, struct task_group *parent)
7189 unsigned long flags;
7191 spin_lock_irqsave(&task_group_lock, flags);
7192 list_add_rcu(&tg->list, &task_groups);
7194 WARN_ON(!parent); /* root should already exist */
7196 tg->parent = parent;
7197 INIT_LIST_HEAD(&tg->children);
7198 list_add_rcu(&tg->siblings, &parent->children);
7199 spin_unlock_irqrestore(&task_group_lock, flags);
7202 /* rcu callback to free various structures associated with a task group */
7203 static void free_sched_group_rcu(struct rcu_head *rhp)
7205 /* now it should be safe to free those cfs_rqs */
7206 free_sched_group(container_of(rhp, struct task_group, rcu));
7209 /* Destroy runqueue etc associated with a task group */
7210 void sched_destroy_group(struct task_group *tg)
7212 /* wait for possible concurrent references to cfs_rqs complete */
7213 call_rcu(&tg->rcu, free_sched_group_rcu);
7216 void sched_offline_group(struct task_group *tg)
7218 unsigned long flags;
7221 /* end participation in shares distribution */
7222 for_each_possible_cpu(i)
7223 unregister_fair_sched_group(tg, i);
7225 spin_lock_irqsave(&task_group_lock, flags);
7226 list_del_rcu(&tg->list);
7227 list_del_rcu(&tg->siblings);
7228 spin_unlock_irqrestore(&task_group_lock, flags);
7231 /* change task's runqueue when it moves between groups.
7232 * The caller of this function should have put the task in its new group
7233 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7234 * reflect its new group.
7236 void sched_move_task(struct task_struct *tsk)
7238 struct task_group *tg;
7240 unsigned long flags;
7243 rq = task_rq_lock(tsk, &flags);
7245 running = task_current(rq, tsk);
7249 dequeue_task(rq, tsk, 0);
7250 if (unlikely(running))
7251 tsk->sched_class->put_prev_task(rq, tsk);
7253 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7254 lockdep_is_held(&tsk->sighand->siglock)),
7255 struct task_group, css);
7256 tg = autogroup_task_group(tsk, tg);
7257 tsk->sched_task_group = tg;
7259 #ifdef CONFIG_FAIR_GROUP_SCHED
7260 if (tsk->sched_class->task_move_group)
7261 tsk->sched_class->task_move_group(tsk, on_rq);
7264 set_task_rq(tsk, task_cpu(tsk));
7266 if (unlikely(running))
7267 tsk->sched_class->set_curr_task(rq);
7269 enqueue_task(rq, tsk, 0);
7271 task_rq_unlock(rq, tsk, &flags);
7273 #endif /* CONFIG_CGROUP_SCHED */
7275 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7276 static unsigned long to_ratio(u64 period, u64 runtime)
7278 if (runtime == RUNTIME_INF)
7281 return div64_u64(runtime << 20, period);
7285 #ifdef CONFIG_RT_GROUP_SCHED
7287 * Ensure that the real time constraints are schedulable.
7289 static DEFINE_MUTEX(rt_constraints_mutex);
7291 /* Must be called with tasklist_lock held */
7292 static inline int tg_has_rt_tasks(struct task_group *tg)
7294 struct task_struct *g, *p;
7296 do_each_thread(g, p) {
7297 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7299 } while_each_thread(g, p);
7304 struct rt_schedulable_data {
7305 struct task_group *tg;
7310 static int tg_rt_schedulable(struct task_group *tg, void *data)
7312 struct rt_schedulable_data *d = data;
7313 struct task_group *child;
7314 unsigned long total, sum = 0;
7315 u64 period, runtime;
7317 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7318 runtime = tg->rt_bandwidth.rt_runtime;
7321 period = d->rt_period;
7322 runtime = d->rt_runtime;
7326 * Cannot have more runtime than the period.
7328 if (runtime > period && runtime != RUNTIME_INF)
7332 * Ensure we don't starve existing RT tasks.
7334 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7337 total = to_ratio(period, runtime);
7340 * Nobody can have more than the global setting allows.
7342 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7346 * The sum of our children's runtime should not exceed our own.
7348 list_for_each_entry_rcu(child, &tg->children, siblings) {
7349 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7350 runtime = child->rt_bandwidth.rt_runtime;
7352 if (child == d->tg) {
7353 period = d->rt_period;
7354 runtime = d->rt_runtime;
7357 sum += to_ratio(period, runtime);
7366 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7370 struct rt_schedulable_data data = {
7372 .rt_period = period,
7373 .rt_runtime = runtime,
7377 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7383 static int tg_set_rt_bandwidth(struct task_group *tg,
7384 u64 rt_period, u64 rt_runtime)
7388 mutex_lock(&rt_constraints_mutex);
7389 read_lock(&tasklist_lock);
7390 err = __rt_schedulable(tg, rt_period, rt_runtime);
7394 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7395 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7396 tg->rt_bandwidth.rt_runtime = rt_runtime;
7398 for_each_possible_cpu(i) {
7399 struct rt_rq *rt_rq = tg->rt_rq[i];
7401 raw_spin_lock(&rt_rq->rt_runtime_lock);
7402 rt_rq->rt_runtime = rt_runtime;
7403 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7405 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7407 read_unlock(&tasklist_lock);
7408 mutex_unlock(&rt_constraints_mutex);
7413 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7415 u64 rt_runtime, rt_period;
7417 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7418 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7419 if (rt_runtime_us < 0)
7420 rt_runtime = RUNTIME_INF;
7422 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7425 long sched_group_rt_runtime(struct task_group *tg)
7429 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7432 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7433 do_div(rt_runtime_us, NSEC_PER_USEC);
7434 return rt_runtime_us;
7437 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7439 u64 rt_runtime, rt_period;
7441 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7442 rt_runtime = tg->rt_bandwidth.rt_runtime;
7447 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7450 long sched_group_rt_period(struct task_group *tg)
7454 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7455 do_div(rt_period_us, NSEC_PER_USEC);
7456 return rt_period_us;
7459 static int sched_rt_global_constraints(void)
7461 u64 runtime, period;
7464 if (sysctl_sched_rt_period <= 0)
7467 runtime = global_rt_runtime();
7468 period = global_rt_period();
7471 * Sanity check on the sysctl variables.
7473 if (runtime > period && runtime != RUNTIME_INF)
7476 mutex_lock(&rt_constraints_mutex);
7477 read_lock(&tasklist_lock);
7478 ret = __rt_schedulable(NULL, 0, 0);
7479 read_unlock(&tasklist_lock);
7480 mutex_unlock(&rt_constraints_mutex);
7485 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7487 /* Don't accept realtime tasks when there is no way for them to run */
7488 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7494 #else /* !CONFIG_RT_GROUP_SCHED */
7495 static int sched_rt_global_constraints(void)
7497 unsigned long flags;
7500 if (sysctl_sched_rt_period <= 0)
7504 * There's always some RT tasks in the root group
7505 * -- migration, kstopmachine etc..
7507 if (sysctl_sched_rt_runtime == 0)
7510 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7511 for_each_possible_cpu(i) {
7512 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7514 raw_spin_lock(&rt_rq->rt_runtime_lock);
7515 rt_rq->rt_runtime = global_rt_runtime();
7516 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7518 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7522 #endif /* CONFIG_RT_GROUP_SCHED */
7524 int sched_rr_handler(struct ctl_table *table, int write,
7525 void __user *buffer, size_t *lenp,
7529 static DEFINE_MUTEX(mutex);
7532 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7533 /* make sure that internally we keep jiffies */
7534 /* also, writing zero resets timeslice to default */
7535 if (!ret && write) {
7536 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7537 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7539 mutex_unlock(&mutex);
7543 int sched_rt_handler(struct ctl_table *table, int write,
7544 void __user *buffer, size_t *lenp,
7548 int old_period, old_runtime;
7549 static DEFINE_MUTEX(mutex);
7552 old_period = sysctl_sched_rt_period;
7553 old_runtime = sysctl_sched_rt_runtime;
7555 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7557 if (!ret && write) {
7558 ret = sched_rt_global_constraints();
7560 sysctl_sched_rt_period = old_period;
7561 sysctl_sched_rt_runtime = old_runtime;
7563 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7564 def_rt_bandwidth.rt_period =
7565 ns_to_ktime(global_rt_period());
7568 mutex_unlock(&mutex);
7573 #ifdef CONFIG_CGROUP_SCHED
7575 /* return corresponding task_group object of a cgroup */
7576 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7578 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7579 struct task_group, css);
7582 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7584 struct task_group *tg, *parent;
7586 if (!cgrp->parent) {
7587 /* This is early initialization for the top cgroup */
7588 return &root_task_group.css;
7591 parent = cgroup_tg(cgrp->parent);
7592 tg = sched_create_group(parent);
7594 return ERR_PTR(-ENOMEM);
7599 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7601 struct task_group *tg = cgroup_tg(cgrp);
7602 struct task_group *parent;
7607 parent = cgroup_tg(cgrp->parent);
7608 sched_online_group(tg, parent);
7612 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7614 struct task_group *tg = cgroup_tg(cgrp);
7616 sched_destroy_group(tg);
7619 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7621 struct task_group *tg = cgroup_tg(cgrp);
7623 sched_offline_group(tg);
7626 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7627 struct cgroup_taskset *tset)
7629 struct task_struct *task;
7631 cgroup_taskset_for_each(task, cgrp, tset) {
7632 #ifdef CONFIG_RT_GROUP_SCHED
7633 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7636 /* We don't support RT-tasks being in separate groups */
7637 if (task->sched_class != &fair_sched_class)
7644 static void cpu_cgroup_attach(struct cgroup *cgrp,
7645 struct cgroup_taskset *tset)
7647 struct task_struct *task;
7649 cgroup_taskset_for_each(task, cgrp, tset)
7650 sched_move_task(task);
7654 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7655 struct task_struct *task)
7658 * cgroup_exit() is called in the copy_process() failure path.
7659 * Ignore this case since the task hasn't ran yet, this avoids
7660 * trying to poke a half freed task state from generic code.
7662 if (!(task->flags & PF_EXITING))
7665 sched_move_task(task);
7668 #ifdef CONFIG_FAIR_GROUP_SCHED
7669 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7672 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7675 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7677 struct task_group *tg = cgroup_tg(cgrp);
7679 return (u64) scale_load_down(tg->shares);
7682 #ifdef CONFIG_CFS_BANDWIDTH
7683 static DEFINE_MUTEX(cfs_constraints_mutex);
7685 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7686 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7688 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7690 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7692 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7693 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7695 if (tg == &root_task_group)
7699 * Ensure we have at some amount of bandwidth every period. This is
7700 * to prevent reaching a state of large arrears when throttled via
7701 * entity_tick() resulting in prolonged exit starvation.
7703 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7707 * Likewise, bound things on the otherside by preventing insane quota
7708 * periods. This also allows us to normalize in computing quota
7711 if (period > max_cfs_quota_period)
7714 mutex_lock(&cfs_constraints_mutex);
7715 ret = __cfs_schedulable(tg, period, quota);
7719 runtime_enabled = quota != RUNTIME_INF;
7720 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7721 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7722 raw_spin_lock_irq(&cfs_b->lock);
7723 cfs_b->period = ns_to_ktime(period);
7724 cfs_b->quota = quota;
7726 __refill_cfs_bandwidth_runtime(cfs_b);
7727 /* restart the period timer (if active) to handle new period expiry */
7728 if (runtime_enabled && cfs_b->timer_active) {
7729 /* force a reprogram */
7730 cfs_b->timer_active = 0;
7731 __start_cfs_bandwidth(cfs_b);
7733 raw_spin_unlock_irq(&cfs_b->lock);
7735 for_each_possible_cpu(i) {
7736 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7737 struct rq *rq = cfs_rq->rq;
7739 raw_spin_lock_irq(&rq->lock);
7740 cfs_rq->runtime_enabled = runtime_enabled;
7741 cfs_rq->runtime_remaining = 0;
7743 if (cfs_rq->throttled)
7744 unthrottle_cfs_rq(cfs_rq);
7745 raw_spin_unlock_irq(&rq->lock);
7748 mutex_unlock(&cfs_constraints_mutex);
7753 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7757 period = ktime_to_ns(tg->cfs_bandwidth.period);
7758 if (cfs_quota_us < 0)
7759 quota = RUNTIME_INF;
7761 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7763 return tg_set_cfs_bandwidth(tg, period, quota);
7766 long tg_get_cfs_quota(struct task_group *tg)
7770 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7773 quota_us = tg->cfs_bandwidth.quota;
7774 do_div(quota_us, NSEC_PER_USEC);
7779 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7783 period = (u64)cfs_period_us * NSEC_PER_USEC;
7784 quota = tg->cfs_bandwidth.quota;
7786 return tg_set_cfs_bandwidth(tg, period, quota);
7789 long tg_get_cfs_period(struct task_group *tg)
7793 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7794 do_div(cfs_period_us, NSEC_PER_USEC);
7796 return cfs_period_us;
7799 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7801 return tg_get_cfs_quota(cgroup_tg(cgrp));
7804 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7807 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7810 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7812 return tg_get_cfs_period(cgroup_tg(cgrp));
7815 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7818 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7821 struct cfs_schedulable_data {
7822 struct task_group *tg;
7827 * normalize group quota/period to be quota/max_period
7828 * note: units are usecs
7830 static u64 normalize_cfs_quota(struct task_group *tg,
7831 struct cfs_schedulable_data *d)
7839 period = tg_get_cfs_period(tg);
7840 quota = tg_get_cfs_quota(tg);
7843 /* note: these should typically be equivalent */
7844 if (quota == RUNTIME_INF || quota == -1)
7847 return to_ratio(period, quota);
7850 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7852 struct cfs_schedulable_data *d = data;
7853 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7854 s64 quota = 0, parent_quota = -1;
7857 quota = RUNTIME_INF;
7859 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7861 quota = normalize_cfs_quota(tg, d);
7862 parent_quota = parent_b->hierarchal_quota;
7865 * ensure max(child_quota) <= parent_quota, inherit when no
7868 if (quota == RUNTIME_INF)
7869 quota = parent_quota;
7870 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7873 cfs_b->hierarchal_quota = quota;
7878 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7881 struct cfs_schedulable_data data = {
7887 if (quota != RUNTIME_INF) {
7888 do_div(data.period, NSEC_PER_USEC);
7889 do_div(data.quota, NSEC_PER_USEC);
7893 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7899 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7900 struct cgroup_map_cb *cb)
7902 struct task_group *tg = cgroup_tg(cgrp);
7903 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7905 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7906 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7907 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7911 #endif /* CONFIG_CFS_BANDWIDTH */
7912 #endif /* CONFIG_FAIR_GROUP_SCHED */
7914 #ifdef CONFIG_RT_GROUP_SCHED
7915 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7918 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7921 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7923 return sched_group_rt_runtime(cgroup_tg(cgrp));
7926 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7929 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7932 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7934 return sched_group_rt_period(cgroup_tg(cgrp));
7936 #endif /* CONFIG_RT_GROUP_SCHED */
7938 static struct cftype cpu_files[] = {
7939 #ifdef CONFIG_FAIR_GROUP_SCHED
7942 .read_u64 = cpu_shares_read_u64,
7943 .write_u64 = cpu_shares_write_u64,
7946 #ifdef CONFIG_CFS_BANDWIDTH
7948 .name = "cfs_quota_us",
7949 .read_s64 = cpu_cfs_quota_read_s64,
7950 .write_s64 = cpu_cfs_quota_write_s64,
7953 .name = "cfs_period_us",
7954 .read_u64 = cpu_cfs_period_read_u64,
7955 .write_u64 = cpu_cfs_period_write_u64,
7959 .read_map = cpu_stats_show,
7962 #ifdef CONFIG_RT_GROUP_SCHED
7964 .name = "rt_runtime_us",
7965 .read_s64 = cpu_rt_runtime_read,
7966 .write_s64 = cpu_rt_runtime_write,
7969 .name = "rt_period_us",
7970 .read_u64 = cpu_rt_period_read_uint,
7971 .write_u64 = cpu_rt_period_write_uint,
7977 struct cgroup_subsys cpu_cgroup_subsys = {
7979 .css_alloc = cpu_cgroup_css_alloc,
7980 .css_free = cpu_cgroup_css_free,
7981 .css_online = cpu_cgroup_css_online,
7982 .css_offline = cpu_cgroup_css_offline,
7983 .can_attach = cpu_cgroup_can_attach,
7984 .attach = cpu_cgroup_attach,
7985 .exit = cpu_cgroup_exit,
7986 .subsys_id = cpu_cgroup_subsys_id,
7987 .base_cftypes = cpu_files,
7991 #endif /* CONFIG_CGROUP_SCHED */
7993 #ifdef CONFIG_CGROUP_CPUACCT
7996 * CPU accounting code for task groups.
7998 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7999 * (balbir@in.ibm.com).
8002 struct cpuacct root_cpuacct;
8004 /* create a new cpu accounting group */
8005 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
8010 return &root_cpuacct.css;
8012 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8016 ca->cpuusage = alloc_percpu(u64);
8020 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8022 goto out_free_cpuusage;
8027 free_percpu(ca->cpuusage);
8031 return ERR_PTR(-ENOMEM);
8034 /* destroy an existing cpu accounting group */
8035 static void cpuacct_css_free(struct cgroup *cgrp)
8037 struct cpuacct *ca = cgroup_ca(cgrp);
8039 free_percpu(ca->cpustat);
8040 free_percpu(ca->cpuusage);
8044 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8046 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8049 #ifndef CONFIG_64BIT
8051 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8053 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8055 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8063 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8065 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8067 #ifndef CONFIG_64BIT
8069 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8071 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8073 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8079 /* return total cpu usage (in nanoseconds) of a group */
8080 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8082 struct cpuacct *ca = cgroup_ca(cgrp);
8083 u64 totalcpuusage = 0;
8086 for_each_present_cpu(i)
8087 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8089 return totalcpuusage;
8092 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8095 struct cpuacct *ca = cgroup_ca(cgrp);
8104 for_each_present_cpu(i)
8105 cpuacct_cpuusage_write(ca, i, 0);
8111 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8114 struct cpuacct *ca = cgroup_ca(cgroup);
8118 for_each_present_cpu(i) {
8119 percpu = cpuacct_cpuusage_read(ca, i);
8120 seq_printf(m, "%llu ", (unsigned long long) percpu);
8122 seq_printf(m, "\n");
8126 static const char *cpuacct_stat_desc[] = {
8127 [CPUACCT_STAT_USER] = "user",
8128 [CPUACCT_STAT_SYSTEM] = "system",
8131 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8132 struct cgroup_map_cb *cb)
8134 struct cpuacct *ca = cgroup_ca(cgrp);
8138 for_each_online_cpu(cpu) {
8139 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8140 val += kcpustat->cpustat[CPUTIME_USER];
8141 val += kcpustat->cpustat[CPUTIME_NICE];
8143 val = cputime64_to_clock_t(val);
8144 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8147 for_each_online_cpu(cpu) {
8148 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8149 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8150 val += kcpustat->cpustat[CPUTIME_IRQ];
8151 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8154 val = cputime64_to_clock_t(val);
8155 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8160 static struct cftype files[] = {
8163 .read_u64 = cpuusage_read,
8164 .write_u64 = cpuusage_write,
8167 .name = "usage_percpu",
8168 .read_seq_string = cpuacct_percpu_seq_read,
8172 .read_map = cpuacct_stats_show,
8178 * charge this task's execution time to its accounting group.
8180 * called with rq->lock held.
8182 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8187 if (unlikely(!cpuacct_subsys.active))
8190 cpu = task_cpu(tsk);
8196 for (; ca; ca = parent_ca(ca)) {
8197 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8198 *cpuusage += cputime;
8204 struct cgroup_subsys cpuacct_subsys = {
8206 .css_alloc = cpuacct_css_alloc,
8207 .css_free = cpuacct_css_free,
8208 .subsys_id = cpuacct_subsys_id,
8209 .base_cftypes = files,
8211 #endif /* CONFIG_CGROUP_CPUACCT */
8213 void dump_cpu_task(int cpu)
8215 pr_info("Task dump for CPU %d:\n", cpu);
8216 sched_show_task(cpu_curr(cpu));