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>
76 #include <asm/irq_regs.h>
77 #ifdef CONFIG_PARAVIRT
78 #include <asm/paravirt.h>
82 #include "workqueue_sched.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
87 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
90 ktime_t soft, hard, now;
93 if (hrtimer_active(period_timer))
96 now = hrtimer_cb_get_time(period_timer);
97 hrtimer_forward(period_timer, now, period);
99 soft = hrtimer_get_softexpires(period_timer);
100 hard = hrtimer_get_expires(period_timer);
101 delta = ktime_to_ns(ktime_sub(hard, soft));
102 __hrtimer_start_range_ns(period_timer, soft, delta,
103 HRTIMER_MODE_ABS_PINNED, 0);
107 DEFINE_MUTEX(sched_domains_mutex);
108 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
110 static void update_rq_clock_task(struct rq *rq, s64 delta);
112 void update_rq_clock(struct rq *rq)
116 if (rq->skip_clock_update > 0)
119 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
121 update_rq_clock_task(rq, delta);
125 * Debugging: various feature bits
128 #define SCHED_FEAT(name, enabled) \
129 (1UL << __SCHED_FEAT_##name) * enabled |
131 const_debug unsigned int sysctl_sched_features =
132 #include "sched_features.h"
137 #ifdef CONFIG_SCHED_DEBUG
138 #define SCHED_FEAT(name, enabled) \
141 static __read_mostly char *sched_feat_names[] = {
142 #include "sched_features.h"
148 static int sched_feat_show(struct seq_file *m, void *v)
152 for (i = 0; sched_feat_names[i]; i++) {
153 if (!(sysctl_sched_features & (1UL << i)))
155 seq_printf(m, "%s ", sched_feat_names[i]);
163 sched_feat_write(struct file *filp, const char __user *ubuf,
164 size_t cnt, loff_t *ppos)
174 if (copy_from_user(&buf, ubuf, cnt))
180 if (strncmp(cmp, "NO_", 3) == 0) {
185 for (i = 0; sched_feat_names[i]; i++) {
186 if (strcmp(cmp, sched_feat_names[i]) == 0) {
188 sysctl_sched_features &= ~(1UL << i);
190 sysctl_sched_features |= (1UL << i);
195 if (!sched_feat_names[i])
203 static int sched_feat_open(struct inode *inode, struct file *filp)
205 return single_open(filp, sched_feat_show, NULL);
208 static const struct file_operations sched_feat_fops = {
209 .open = sched_feat_open,
210 .write = sched_feat_write,
213 .release = single_release,
216 static __init int sched_init_debug(void)
218 debugfs_create_file("sched_features", 0644, NULL, NULL,
223 late_initcall(sched_init_debug);
228 * Number of tasks to iterate in a single balance run.
229 * Limited because this is done with IRQs disabled.
231 const_debug unsigned int sysctl_sched_nr_migrate = 32;
234 * period over which we average the RT time consumption, measured
239 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
242 * period over which we measure -rt task cpu usage in us.
245 unsigned int sysctl_sched_rt_period = 1000000;
247 __read_mostly int scheduler_running;
250 * part of the period that we allow rt tasks to run in us.
253 int sysctl_sched_rt_runtime = 950000;
258 * __task_rq_lock - lock the rq @p resides on.
260 static inline struct rq *__task_rq_lock(struct task_struct *p)
265 lockdep_assert_held(&p->pi_lock);
269 raw_spin_lock(&rq->lock);
270 if (likely(rq == task_rq(p)))
272 raw_spin_unlock(&rq->lock);
277 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
279 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
280 __acquires(p->pi_lock)
286 raw_spin_lock_irqsave(&p->pi_lock, *flags);
288 raw_spin_lock(&rq->lock);
289 if (likely(rq == task_rq(p)))
291 raw_spin_unlock(&rq->lock);
292 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
296 static void __task_rq_unlock(struct rq *rq)
299 raw_spin_unlock(&rq->lock);
303 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
305 __releases(p->pi_lock)
307 raw_spin_unlock(&rq->lock);
308 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
312 * this_rq_lock - lock this runqueue and disable interrupts.
314 static struct rq *this_rq_lock(void)
321 raw_spin_lock(&rq->lock);
326 #ifdef CONFIG_SCHED_HRTICK
328 * Use HR-timers to deliver accurate preemption points.
330 * Its all a bit involved since we cannot program an hrt while holding the
331 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
334 * When we get rescheduled we reprogram the hrtick_timer outside of the
338 static void hrtick_clear(struct rq *rq)
340 if (hrtimer_active(&rq->hrtick_timer))
341 hrtimer_cancel(&rq->hrtick_timer);
345 * High-resolution timer tick.
346 * Runs from hardirq context with interrupts disabled.
348 static enum hrtimer_restart hrtick(struct hrtimer *timer)
350 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
352 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
354 raw_spin_lock(&rq->lock);
356 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
357 raw_spin_unlock(&rq->lock);
359 return HRTIMER_NORESTART;
364 * called from hardirq (IPI) context
366 static void __hrtick_start(void *arg)
370 raw_spin_lock(&rq->lock);
371 hrtimer_restart(&rq->hrtick_timer);
372 rq->hrtick_csd_pending = 0;
373 raw_spin_unlock(&rq->lock);
377 * Called to set the hrtick timer state.
379 * called with rq->lock held and irqs disabled
381 void hrtick_start(struct rq *rq, u64 delay)
383 struct hrtimer *timer = &rq->hrtick_timer;
384 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
386 hrtimer_set_expires(timer, time);
388 if (rq == this_rq()) {
389 hrtimer_restart(timer);
390 } else if (!rq->hrtick_csd_pending) {
391 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
392 rq->hrtick_csd_pending = 1;
397 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
399 int cpu = (int)(long)hcpu;
402 case CPU_UP_CANCELED:
403 case CPU_UP_CANCELED_FROZEN:
404 case CPU_DOWN_PREPARE:
405 case CPU_DOWN_PREPARE_FROZEN:
407 case CPU_DEAD_FROZEN:
408 hrtick_clear(cpu_rq(cpu));
415 static __init void init_hrtick(void)
417 hotcpu_notifier(hotplug_hrtick, 0);
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 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
428 HRTIMER_MODE_REL_PINNED, 0);
431 static inline void init_hrtick(void)
434 #endif /* CONFIG_SMP */
436 static void init_rq_hrtick(struct rq *rq)
439 rq->hrtick_csd_pending = 0;
441 rq->hrtick_csd.flags = 0;
442 rq->hrtick_csd.func = __hrtick_start;
443 rq->hrtick_csd.info = rq;
446 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
447 rq->hrtick_timer.function = hrtick;
449 #else /* CONFIG_SCHED_HRTICK */
450 static inline void hrtick_clear(struct rq *rq)
454 static inline void init_rq_hrtick(struct rq *rq)
458 static inline void init_hrtick(void)
461 #endif /* CONFIG_SCHED_HRTICK */
464 * resched_task - mark a task 'to be rescheduled now'.
466 * On UP this means the setting of the need_resched flag, on SMP it
467 * might also involve a cross-CPU call to trigger the scheduler on
472 #ifndef tsk_is_polling
473 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
476 void resched_task(struct task_struct *p)
480 assert_raw_spin_locked(&task_rq(p)->lock);
482 if (test_tsk_need_resched(p))
485 set_tsk_need_resched(p);
488 if (cpu == smp_processor_id())
491 /* NEED_RESCHED must be visible before we test polling */
493 if (!tsk_is_polling(p))
494 smp_send_reschedule(cpu);
497 void resched_cpu(int cpu)
499 struct rq *rq = cpu_rq(cpu);
502 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
504 resched_task(cpu_curr(cpu));
505 raw_spin_unlock_irqrestore(&rq->lock, flags);
510 * In the semi idle case, use the nearest busy cpu for migrating timers
511 * from an idle cpu. This is good for power-savings.
513 * We don't do similar optimization for completely idle system, as
514 * selecting an idle cpu will add more delays to the timers than intended
515 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
517 int get_nohz_timer_target(void)
519 int cpu = smp_processor_id();
521 struct sched_domain *sd;
524 for_each_domain(cpu, sd) {
525 for_each_cpu(i, sched_domain_span(sd)) {
537 * When add_timer_on() enqueues a timer into the timer wheel of an
538 * idle CPU then this timer might expire before the next timer event
539 * which is scheduled to wake up that CPU. In case of a completely
540 * idle system the next event might even be infinite time into the
541 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
542 * leaves the inner idle loop so the newly added timer is taken into
543 * account when the CPU goes back to idle and evaluates the timer
544 * wheel for the next timer event.
546 void wake_up_idle_cpu(int cpu)
548 struct rq *rq = cpu_rq(cpu);
550 if (cpu == smp_processor_id())
554 * This is safe, as this function is called with the timer
555 * wheel base lock of (cpu) held. When the CPU is on the way
556 * to idle and has not yet set rq->curr to idle then it will
557 * be serialized on the timer wheel base lock and take the new
558 * timer into account automatically.
560 if (rq->curr != rq->idle)
564 * We can set TIF_RESCHED on the idle task of the other CPU
565 * lockless. The worst case is that the other CPU runs the
566 * idle task through an additional NOOP schedule()
568 set_tsk_need_resched(rq->idle);
570 /* NEED_RESCHED must be visible before we test polling */
572 if (!tsk_is_polling(rq->idle))
573 smp_send_reschedule(cpu);
576 static inline bool got_nohz_idle_kick(void)
578 return idle_cpu(smp_processor_id()) && this_rq()->nohz_balance_kick;
581 #else /* CONFIG_NO_HZ */
583 static inline bool got_nohz_idle_kick(void)
588 #endif /* CONFIG_NO_HZ */
590 void sched_avg_update(struct rq *rq)
592 s64 period = sched_avg_period();
594 while ((s64)(rq->clock - rq->age_stamp) > period) {
596 * Inline assembly required to prevent the compiler
597 * optimising this loop into a divmod call.
598 * See __iter_div_u64_rem() for another example of this.
600 asm("" : "+rm" (rq->age_stamp));
601 rq->age_stamp += period;
606 #else /* !CONFIG_SMP */
607 void resched_task(struct task_struct *p)
609 assert_raw_spin_locked(&task_rq(p)->lock);
610 set_tsk_need_resched(p);
612 #endif /* CONFIG_SMP */
614 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
615 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
617 * Iterate task_group tree rooted at *from, calling @down when first entering a
618 * node and @up when leaving it for the final time.
620 * Caller must hold rcu_lock or sufficient equivalent.
622 int walk_tg_tree_from(struct task_group *from,
623 tg_visitor down, tg_visitor up, void *data)
625 struct task_group *parent, *child;
631 ret = (*down)(parent, data);
634 list_for_each_entry_rcu(child, &parent->children, siblings) {
641 ret = (*up)(parent, data);
642 if (ret || parent == from)
646 parent = parent->parent;
653 int tg_nop(struct task_group *tg, void *data)
659 void update_cpu_load(struct rq *this_rq);
661 static void set_load_weight(struct task_struct *p)
663 int prio = p->static_prio - MAX_RT_PRIO;
664 struct load_weight *load = &p->se.load;
667 * SCHED_IDLE tasks get minimal weight:
669 if (p->policy == SCHED_IDLE) {
670 load->weight = scale_load(WEIGHT_IDLEPRIO);
671 load->inv_weight = WMULT_IDLEPRIO;
675 load->weight = scale_load(prio_to_weight[prio]);
676 load->inv_weight = prio_to_wmult[prio];
679 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
682 sched_info_queued(p);
683 p->sched_class->enqueue_task(rq, p, flags);
686 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
689 sched_info_dequeued(p);
690 p->sched_class->dequeue_task(rq, p, flags);
694 * activate_task - move a task to the runqueue.
696 void activate_task(struct rq *rq, struct task_struct *p, int flags)
698 if (task_contributes_to_load(p))
699 rq->nr_uninterruptible--;
701 enqueue_task(rq, p, flags);
705 * deactivate_task - remove a task from the runqueue.
707 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
709 if (task_contributes_to_load(p))
710 rq->nr_uninterruptible++;
712 dequeue_task(rq, p, flags);
715 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
718 * There are no locks covering percpu hardirq/softirq time.
719 * They are only modified in account_system_vtime, on corresponding CPU
720 * with interrupts disabled. So, writes are safe.
721 * They are read and saved off onto struct rq in update_rq_clock().
722 * This may result in other CPU reading this CPU's irq time and can
723 * race with irq/account_system_vtime on this CPU. We would either get old
724 * or new value with a side effect of accounting a slice of irq time to wrong
725 * task when irq is in progress while we read rq->clock. That is a worthy
726 * compromise in place of having locks on each irq in account_system_time.
728 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
729 static DEFINE_PER_CPU(u64, cpu_softirq_time);
731 static DEFINE_PER_CPU(u64, irq_start_time);
732 static int sched_clock_irqtime;
734 void enable_sched_clock_irqtime(void)
736 sched_clock_irqtime = 1;
739 void disable_sched_clock_irqtime(void)
741 sched_clock_irqtime = 0;
745 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
747 static inline void irq_time_write_begin(void)
749 __this_cpu_inc(irq_time_seq.sequence);
753 static inline void irq_time_write_end(void)
756 __this_cpu_inc(irq_time_seq.sequence);
759 static inline u64 irq_time_read(int cpu)
765 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
766 irq_time = per_cpu(cpu_softirq_time, cpu) +
767 per_cpu(cpu_hardirq_time, cpu);
768 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
772 #else /* CONFIG_64BIT */
773 static inline void irq_time_write_begin(void)
777 static inline void irq_time_write_end(void)
781 static inline u64 irq_time_read(int cpu)
783 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
785 #endif /* CONFIG_64BIT */
788 * Called before incrementing preempt_count on {soft,}irq_enter
789 * and before decrementing preempt_count on {soft,}irq_exit.
791 void account_system_vtime(struct task_struct *curr)
797 if (!sched_clock_irqtime)
800 local_irq_save(flags);
802 cpu = smp_processor_id();
803 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
804 __this_cpu_add(irq_start_time, delta);
806 irq_time_write_begin();
808 * We do not account for softirq time from ksoftirqd here.
809 * We want to continue accounting softirq time to ksoftirqd thread
810 * in that case, so as not to confuse scheduler with a special task
811 * that do not consume any time, but still wants to run.
814 __this_cpu_add(cpu_hardirq_time, delta);
815 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
816 __this_cpu_add(cpu_softirq_time, delta);
818 irq_time_write_end();
819 local_irq_restore(flags);
821 EXPORT_SYMBOL_GPL(account_system_vtime);
823 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
825 #ifdef CONFIG_PARAVIRT
826 static inline u64 steal_ticks(u64 steal)
828 if (unlikely(steal > NSEC_PER_SEC))
829 return div_u64(steal, TICK_NSEC);
831 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
835 static void update_rq_clock_task(struct rq *rq, s64 delta)
838 * In theory, the compile should just see 0 here, and optimize out the call
839 * to sched_rt_avg_update. But I don't trust it...
841 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
842 s64 steal = 0, irq_delta = 0;
844 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
845 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
848 * Since irq_time is only updated on {soft,}irq_exit, we might run into
849 * this case when a previous update_rq_clock() happened inside a
852 * When this happens, we stop ->clock_task and only update the
853 * prev_irq_time stamp to account for the part that fit, so that a next
854 * update will consume the rest. This ensures ->clock_task is
857 * It does however cause some slight miss-attribution of {soft,}irq
858 * time, a more accurate solution would be to update the irq_time using
859 * the current rq->clock timestamp, except that would require using
862 if (irq_delta > delta)
865 rq->prev_irq_time += irq_delta;
868 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
869 if (static_branch((¶virt_steal_rq_enabled))) {
872 steal = paravirt_steal_clock(cpu_of(rq));
873 steal -= rq->prev_steal_time_rq;
875 if (unlikely(steal > delta))
878 st = steal_ticks(steal);
879 steal = st * TICK_NSEC;
881 rq->prev_steal_time_rq += steal;
887 rq->clock_task += delta;
889 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
890 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
891 sched_rt_avg_update(rq, irq_delta + steal);
895 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
896 static int irqtime_account_hi_update(void)
898 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
903 local_irq_save(flags);
904 latest_ns = this_cpu_read(cpu_hardirq_time);
905 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
907 local_irq_restore(flags);
911 static int irqtime_account_si_update(void)
913 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
918 local_irq_save(flags);
919 latest_ns = this_cpu_read(cpu_softirq_time);
920 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
922 local_irq_restore(flags);
926 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
928 #define sched_clock_irqtime (0)
932 void sched_set_stop_task(int cpu, struct task_struct *stop)
934 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
935 struct task_struct *old_stop = cpu_rq(cpu)->stop;
939 * Make it appear like a SCHED_FIFO task, its something
940 * userspace knows about and won't get confused about.
942 * Also, it will make PI more or less work without too
943 * much confusion -- but then, stop work should not
944 * rely on PI working anyway.
946 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
948 stop->sched_class = &stop_sched_class;
951 cpu_rq(cpu)->stop = stop;
955 * Reset it back to a normal scheduling class so that
956 * it can die in pieces.
958 old_stop->sched_class = &rt_sched_class;
963 * __normal_prio - return the priority that is based on the static prio
965 static inline int __normal_prio(struct task_struct *p)
967 return p->static_prio;
971 * Calculate the expected normal priority: i.e. priority
972 * without taking RT-inheritance into account. Might be
973 * boosted by interactivity modifiers. Changes upon fork,
974 * setprio syscalls, and whenever the interactivity
975 * estimator recalculates.
977 static inline int normal_prio(struct task_struct *p)
981 if (task_has_rt_policy(p))
982 prio = MAX_RT_PRIO-1 - p->rt_priority;
984 prio = __normal_prio(p);
989 * Calculate the current priority, i.e. the priority
990 * taken into account by the scheduler. This value might
991 * be boosted by RT tasks, or might be boosted by
992 * interactivity modifiers. Will be RT if the task got
993 * RT-boosted. If not then it returns p->normal_prio.
995 static int effective_prio(struct task_struct *p)
997 p->normal_prio = normal_prio(p);
999 * If we are RT tasks or we were boosted to RT priority,
1000 * keep the priority unchanged. Otherwise, update priority
1001 * to the normal priority:
1003 if (!rt_prio(p->prio))
1004 return p->normal_prio;
1009 * task_curr - is this task currently executing on a CPU?
1010 * @p: the task in question.
1012 inline int task_curr(const struct task_struct *p)
1014 return cpu_curr(task_cpu(p)) == p;
1017 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1018 const struct sched_class *prev_class,
1021 if (prev_class != p->sched_class) {
1022 if (prev_class->switched_from)
1023 prev_class->switched_from(rq, p);
1024 p->sched_class->switched_to(rq, p);
1025 } else if (oldprio != p->prio)
1026 p->sched_class->prio_changed(rq, p, oldprio);
1029 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1031 const struct sched_class *class;
1033 if (p->sched_class == rq->curr->sched_class) {
1034 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1036 for_each_class(class) {
1037 if (class == rq->curr->sched_class)
1039 if (class == p->sched_class) {
1040 resched_task(rq->curr);
1047 * A queue event has occurred, and we're going to schedule. In
1048 * this case, we can save a useless back to back clock update.
1050 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1051 rq->skip_clock_update = 1;
1055 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1057 #ifdef CONFIG_SCHED_DEBUG
1059 * We should never call set_task_cpu() on a blocked task,
1060 * ttwu() will sort out the placement.
1062 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1063 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1065 #ifdef CONFIG_LOCKDEP
1067 * The caller should hold either p->pi_lock or rq->lock, when changing
1068 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1070 * sched_move_task() holds both and thus holding either pins the cgroup,
1071 * see set_task_rq().
1073 * Furthermore, all task_rq users should acquire both locks, see
1076 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1077 lockdep_is_held(&task_rq(p)->lock)));
1081 trace_sched_migrate_task(p, new_cpu);
1083 if (task_cpu(p) != new_cpu) {
1084 p->se.nr_migrations++;
1085 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1088 __set_task_cpu(p, new_cpu);
1091 struct migration_arg {
1092 struct task_struct *task;
1096 static int migration_cpu_stop(void *data);
1099 * wait_task_inactive - wait for a thread to unschedule.
1101 * If @match_state is nonzero, it's the @p->state value just checked and
1102 * not expected to change. If it changes, i.e. @p might have woken up,
1103 * then return zero. When we succeed in waiting for @p to be off its CPU,
1104 * we return a positive number (its total switch count). If a second call
1105 * a short while later returns the same number, the caller can be sure that
1106 * @p has remained unscheduled the whole time.
1108 * The caller must ensure that the task *will* unschedule sometime soon,
1109 * else this function might spin for a *long* time. This function can't
1110 * be called with interrupts off, or it may introduce deadlock with
1111 * smp_call_function() if an IPI is sent by the same process we are
1112 * waiting to become inactive.
1114 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1116 unsigned long flags;
1123 * We do the initial early heuristics without holding
1124 * any task-queue locks at all. We'll only try to get
1125 * the runqueue lock when things look like they will
1131 * If the task is actively running on another CPU
1132 * still, just relax and busy-wait without holding
1135 * NOTE! Since we don't hold any locks, it's not
1136 * even sure that "rq" stays as the right runqueue!
1137 * But we don't care, since "task_running()" will
1138 * return false if the runqueue has changed and p
1139 * is actually now running somewhere else!
1141 while (task_running(rq, p)) {
1142 if (match_state && unlikely(p->state != match_state))
1148 * Ok, time to look more closely! We need the rq
1149 * lock now, to be *sure*. If we're wrong, we'll
1150 * just go back and repeat.
1152 rq = task_rq_lock(p, &flags);
1153 trace_sched_wait_task(p);
1154 running = task_running(rq, p);
1157 if (!match_state || p->state == match_state)
1158 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1159 task_rq_unlock(rq, p, &flags);
1162 * If it changed from the expected state, bail out now.
1164 if (unlikely(!ncsw))
1168 * Was it really running after all now that we
1169 * checked with the proper locks actually held?
1171 * Oops. Go back and try again..
1173 if (unlikely(running)) {
1179 * It's not enough that it's not actively running,
1180 * it must be off the runqueue _entirely_, and not
1183 * So if it was still runnable (but just not actively
1184 * running right now), it's preempted, and we should
1185 * yield - it could be a while.
1187 if (unlikely(on_rq)) {
1188 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1190 set_current_state(TASK_UNINTERRUPTIBLE);
1191 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1196 * Ahh, all good. It wasn't running, and it wasn't
1197 * runnable, which means that it will never become
1198 * running in the future either. We're all done!
1207 * kick_process - kick a running thread to enter/exit the kernel
1208 * @p: the to-be-kicked thread
1210 * Cause a process which is running on another CPU to enter
1211 * kernel-mode, without any delay. (to get signals handled.)
1213 * NOTE: this function doesn't have to take the runqueue lock,
1214 * because all it wants to ensure is that the remote task enters
1215 * the kernel. If the IPI races and the task has been migrated
1216 * to another CPU then no harm is done and the purpose has been
1219 void kick_process(struct task_struct *p)
1225 if ((cpu != smp_processor_id()) && task_curr(p))
1226 smp_send_reschedule(cpu);
1229 EXPORT_SYMBOL_GPL(kick_process);
1230 #endif /* CONFIG_SMP */
1234 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1236 static int select_fallback_rq(int cpu, struct task_struct *p)
1239 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1241 /* Look for allowed, online CPU in same node. */
1242 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1243 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1246 /* Any allowed, online CPU? */
1247 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1248 if (dest_cpu < nr_cpu_ids)
1251 /* No more Mr. Nice Guy. */
1252 dest_cpu = cpuset_cpus_allowed_fallback(p);
1254 * Don't tell them about moving exiting tasks or
1255 * kernel threads (both mm NULL), since they never
1258 if (p->mm && printk_ratelimit()) {
1259 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1260 task_pid_nr(p), p->comm, cpu);
1267 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1270 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1272 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1275 * In order not to call set_task_cpu() on a blocking task we need
1276 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1279 * Since this is common to all placement strategies, this lives here.
1281 * [ this allows ->select_task() to simply return task_cpu(p) and
1282 * not worry about this generic constraint ]
1284 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1286 cpu = select_fallback_rq(task_cpu(p), p);
1291 static void update_avg(u64 *avg, u64 sample)
1293 s64 diff = sample - *avg;
1299 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1301 #ifdef CONFIG_SCHEDSTATS
1302 struct rq *rq = this_rq();
1305 int this_cpu = smp_processor_id();
1307 if (cpu == this_cpu) {
1308 schedstat_inc(rq, ttwu_local);
1309 schedstat_inc(p, se.statistics.nr_wakeups_local);
1311 struct sched_domain *sd;
1313 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1315 for_each_domain(this_cpu, sd) {
1316 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1317 schedstat_inc(sd, ttwu_wake_remote);
1324 if (wake_flags & WF_MIGRATED)
1325 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1327 #endif /* CONFIG_SMP */
1329 schedstat_inc(rq, ttwu_count);
1330 schedstat_inc(p, se.statistics.nr_wakeups);
1332 if (wake_flags & WF_SYNC)
1333 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1335 #endif /* CONFIG_SCHEDSTATS */
1338 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1340 activate_task(rq, p, en_flags);
1343 /* if a worker is waking up, notify workqueue */
1344 if (p->flags & PF_WQ_WORKER)
1345 wq_worker_waking_up(p, cpu_of(rq));
1349 * Mark the task runnable and perform wakeup-preemption.
1352 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1354 trace_sched_wakeup(p, true);
1355 check_preempt_curr(rq, p, wake_flags);
1357 p->state = TASK_RUNNING;
1359 if (p->sched_class->task_woken)
1360 p->sched_class->task_woken(rq, p);
1362 if (rq->idle_stamp) {
1363 u64 delta = rq->clock - rq->idle_stamp;
1364 u64 max = 2*sysctl_sched_migration_cost;
1369 update_avg(&rq->avg_idle, delta);
1376 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1379 if (p->sched_contributes_to_load)
1380 rq->nr_uninterruptible--;
1383 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1384 ttwu_do_wakeup(rq, p, wake_flags);
1388 * Called in case the task @p isn't fully descheduled from its runqueue,
1389 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1390 * since all we need to do is flip p->state to TASK_RUNNING, since
1391 * the task is still ->on_rq.
1393 static int ttwu_remote(struct task_struct *p, int wake_flags)
1398 rq = __task_rq_lock(p);
1400 ttwu_do_wakeup(rq, p, wake_flags);
1403 __task_rq_unlock(rq);
1409 static void sched_ttwu_pending(void)
1411 struct rq *rq = this_rq();
1412 struct llist_node *llist = llist_del_all(&rq->wake_list);
1413 struct task_struct *p;
1415 raw_spin_lock(&rq->lock);
1418 p = llist_entry(llist, struct task_struct, wake_entry);
1419 llist = llist_next(llist);
1420 ttwu_do_activate(rq, p, 0);
1423 raw_spin_unlock(&rq->lock);
1426 void scheduler_ipi(void)
1428 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1432 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1433 * traditionally all their work was done from the interrupt return
1434 * path. Now that we actually do some work, we need to make sure
1437 * Some archs already do call them, luckily irq_enter/exit nest
1440 * Arguably we should visit all archs and update all handlers,
1441 * however a fair share of IPIs are still resched only so this would
1442 * somewhat pessimize the simple resched case.
1445 sched_ttwu_pending();
1448 * Check if someone kicked us for doing the nohz idle load balance.
1450 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1451 this_rq()->idle_balance = 1;
1452 raise_softirq_irqoff(SCHED_SOFTIRQ);
1457 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1459 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1460 smp_send_reschedule(cpu);
1463 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1464 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1469 rq = __task_rq_lock(p);
1471 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1472 ttwu_do_wakeup(rq, p, wake_flags);
1475 __task_rq_unlock(rq);
1480 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1481 #endif /* CONFIG_SMP */
1483 static void ttwu_queue(struct task_struct *p, int cpu)
1485 struct rq *rq = cpu_rq(cpu);
1487 #if defined(CONFIG_SMP)
1488 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
1489 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1490 ttwu_queue_remote(p, cpu);
1495 raw_spin_lock(&rq->lock);
1496 ttwu_do_activate(rq, p, 0);
1497 raw_spin_unlock(&rq->lock);
1501 * try_to_wake_up - wake up a thread
1502 * @p: the thread to be awakened
1503 * @state: the mask of task states that can be woken
1504 * @wake_flags: wake modifier flags (WF_*)
1506 * Put it on the run-queue if it's not already there. The "current"
1507 * thread is always on the run-queue (except when the actual
1508 * re-schedule is in progress), and as such you're allowed to do
1509 * the simpler "current->state = TASK_RUNNING" to mark yourself
1510 * runnable without the overhead of this.
1512 * Returns %true if @p was woken up, %false if it was already running
1513 * or @state didn't match @p's state.
1516 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1518 unsigned long flags;
1519 int cpu, success = 0;
1522 raw_spin_lock_irqsave(&p->pi_lock, flags);
1523 if (!(p->state & state))
1526 success = 1; /* we're going to change ->state */
1529 if (p->on_rq && ttwu_remote(p, wake_flags))
1534 * If the owning (remote) cpu is still in the middle of schedule() with
1535 * this task as prev, wait until its done referencing the task.
1538 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1540 * In case the architecture enables interrupts in
1541 * context_switch(), we cannot busy wait, since that
1542 * would lead to deadlocks when an interrupt hits and
1543 * tries to wake up @prev. So bail and do a complete
1546 if (ttwu_activate_remote(p, wake_flags))
1553 * Pairs with the smp_wmb() in finish_lock_switch().
1557 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1558 p->state = TASK_WAKING;
1560 if (p->sched_class->task_waking)
1561 p->sched_class->task_waking(p);
1563 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1564 if (task_cpu(p) != cpu) {
1565 wake_flags |= WF_MIGRATED;
1566 set_task_cpu(p, cpu);
1568 #endif /* CONFIG_SMP */
1572 ttwu_stat(p, cpu, wake_flags);
1574 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1580 * try_to_wake_up_local - try to wake up a local task with rq lock held
1581 * @p: the thread to be awakened
1583 * Put @p on the run-queue if it's not already there. The caller must
1584 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1587 static void try_to_wake_up_local(struct task_struct *p)
1589 struct rq *rq = task_rq(p);
1591 BUG_ON(rq != this_rq());
1592 BUG_ON(p == current);
1593 lockdep_assert_held(&rq->lock);
1595 if (!raw_spin_trylock(&p->pi_lock)) {
1596 raw_spin_unlock(&rq->lock);
1597 raw_spin_lock(&p->pi_lock);
1598 raw_spin_lock(&rq->lock);
1601 if (!(p->state & TASK_NORMAL))
1605 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1607 ttwu_do_wakeup(rq, p, 0);
1608 ttwu_stat(p, smp_processor_id(), 0);
1610 raw_spin_unlock(&p->pi_lock);
1614 * wake_up_process - Wake up a specific process
1615 * @p: The process to be woken up.
1617 * Attempt to wake up the nominated process and move it to the set of runnable
1618 * processes. Returns 1 if the process was woken up, 0 if it was already
1621 * It may be assumed that this function implies a write memory barrier before
1622 * changing the task state if and only if any tasks are woken up.
1624 int wake_up_process(struct task_struct *p)
1626 return try_to_wake_up(p, TASK_ALL, 0);
1628 EXPORT_SYMBOL(wake_up_process);
1630 int wake_up_state(struct task_struct *p, unsigned int state)
1632 return try_to_wake_up(p, state, 0);
1636 * Perform scheduler related setup for a newly forked process p.
1637 * p is forked by current.
1639 * __sched_fork() is basic setup used by init_idle() too:
1641 static void __sched_fork(struct task_struct *p)
1646 p->se.exec_start = 0;
1647 p->se.sum_exec_runtime = 0;
1648 p->se.prev_sum_exec_runtime = 0;
1649 p->se.nr_migrations = 0;
1651 INIT_LIST_HEAD(&p->se.group_node);
1653 #ifdef CONFIG_SCHEDSTATS
1654 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1657 INIT_LIST_HEAD(&p->rt.run_list);
1659 #ifdef CONFIG_PREEMPT_NOTIFIERS
1660 INIT_HLIST_HEAD(&p->preempt_notifiers);
1665 * fork()/clone()-time setup:
1667 void sched_fork(struct task_struct *p)
1669 unsigned long flags;
1670 int cpu = get_cpu();
1674 * We mark the process as running here. This guarantees that
1675 * nobody will actually run it, and a signal or other external
1676 * event cannot wake it up and insert it on the runqueue either.
1678 p->state = TASK_RUNNING;
1681 * Make sure we do not leak PI boosting priority to the child.
1683 p->prio = current->normal_prio;
1686 * Revert to default priority/policy on fork if requested.
1688 if (unlikely(p->sched_reset_on_fork)) {
1689 if (task_has_rt_policy(p)) {
1690 p->policy = SCHED_NORMAL;
1691 p->static_prio = NICE_TO_PRIO(0);
1693 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1694 p->static_prio = NICE_TO_PRIO(0);
1696 p->prio = p->normal_prio = __normal_prio(p);
1700 * We don't need the reset flag anymore after the fork. It has
1701 * fulfilled its duty:
1703 p->sched_reset_on_fork = 0;
1706 if (!rt_prio(p->prio))
1707 p->sched_class = &fair_sched_class;
1709 if (p->sched_class->task_fork)
1710 p->sched_class->task_fork(p);
1713 * The child is not yet in the pid-hash so no cgroup attach races,
1714 * and the cgroup is pinned to this child due to cgroup_fork()
1715 * is ran before sched_fork().
1717 * Silence PROVE_RCU.
1719 raw_spin_lock_irqsave(&p->pi_lock, flags);
1720 set_task_cpu(p, cpu);
1721 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1723 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1724 if (likely(sched_info_on()))
1725 memset(&p->sched_info, 0, sizeof(p->sched_info));
1727 #if defined(CONFIG_SMP)
1730 #ifdef CONFIG_PREEMPT_COUNT
1731 /* Want to start with kernel preemption disabled. */
1732 task_thread_info(p)->preempt_count = 1;
1735 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1742 * wake_up_new_task - wake up a newly created task for the first time.
1744 * This function will do some initial scheduler statistics housekeeping
1745 * that must be done for every newly created context, then puts the task
1746 * on the runqueue and wakes it.
1748 void wake_up_new_task(struct task_struct *p)
1750 unsigned long flags;
1753 raw_spin_lock_irqsave(&p->pi_lock, flags);
1756 * Fork balancing, do it here and not earlier because:
1757 * - cpus_allowed can change in the fork path
1758 * - any previously selected cpu might disappear through hotplug
1760 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1763 rq = __task_rq_lock(p);
1764 activate_task(rq, p, 0);
1766 trace_sched_wakeup_new(p, true);
1767 check_preempt_curr(rq, p, WF_FORK);
1769 if (p->sched_class->task_woken)
1770 p->sched_class->task_woken(rq, p);
1772 task_rq_unlock(rq, p, &flags);
1775 #ifdef CONFIG_PREEMPT_NOTIFIERS
1778 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1779 * @notifier: notifier struct to register
1781 void preempt_notifier_register(struct preempt_notifier *notifier)
1783 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1785 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1788 * preempt_notifier_unregister - no longer interested in preemption notifications
1789 * @notifier: notifier struct to unregister
1791 * This is safe to call from within a preemption notifier.
1793 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1795 hlist_del(¬ifier->link);
1797 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1799 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1801 struct preempt_notifier *notifier;
1802 struct hlist_node *node;
1804 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1805 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1809 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1810 struct task_struct *next)
1812 struct preempt_notifier *notifier;
1813 struct hlist_node *node;
1815 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1816 notifier->ops->sched_out(notifier, next);
1819 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1821 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1826 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1827 struct task_struct *next)
1831 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1834 * prepare_task_switch - prepare to switch tasks
1835 * @rq: the runqueue preparing to switch
1836 * @prev: the current task that is being switched out
1837 * @next: the task we are going to switch to.
1839 * This is called with the rq lock held and interrupts off. It must
1840 * be paired with a subsequent finish_task_switch after the context
1843 * prepare_task_switch sets up locking and calls architecture specific
1847 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1848 struct task_struct *next)
1850 sched_info_switch(prev, next);
1851 perf_event_task_sched_out(prev, next);
1852 fire_sched_out_preempt_notifiers(prev, next);
1853 prepare_lock_switch(rq, next);
1854 prepare_arch_switch(next);
1855 trace_sched_switch(prev, next);
1859 * finish_task_switch - clean up after a task-switch
1860 * @rq: runqueue associated with task-switch
1861 * @prev: the thread we just switched away from.
1863 * finish_task_switch must be called after the context switch, paired
1864 * with a prepare_task_switch call before the context switch.
1865 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1866 * and do any other architecture-specific cleanup actions.
1868 * Note that we may have delayed dropping an mm in context_switch(). If
1869 * so, we finish that here outside of the runqueue lock. (Doing it
1870 * with the lock held can cause deadlocks; see schedule() for
1873 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1874 __releases(rq->lock)
1876 struct mm_struct *mm = rq->prev_mm;
1882 * A task struct has one reference for the use as "current".
1883 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1884 * schedule one last time. The schedule call will never return, and
1885 * the scheduled task must drop that reference.
1886 * The test for TASK_DEAD must occur while the runqueue locks are
1887 * still held, otherwise prev could be scheduled on another cpu, die
1888 * there before we look at prev->state, and then the reference would
1890 * Manfred Spraul <manfred@colorfullife.com>
1892 prev_state = prev->state;
1893 finish_arch_switch(prev);
1894 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1895 local_irq_disable();
1896 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1897 perf_event_task_sched_in(prev, current);
1898 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1900 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1901 finish_lock_switch(rq, prev);
1903 fire_sched_in_preempt_notifiers(current);
1906 if (unlikely(prev_state == TASK_DEAD)) {
1908 * Remove function-return probe instances associated with this
1909 * task and put them back on the free list.
1911 kprobe_flush_task(prev);
1912 put_task_struct(prev);
1918 /* assumes rq->lock is held */
1919 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1921 if (prev->sched_class->pre_schedule)
1922 prev->sched_class->pre_schedule(rq, prev);
1925 /* rq->lock is NOT held, but preemption is disabled */
1926 static inline void post_schedule(struct rq *rq)
1928 if (rq->post_schedule) {
1929 unsigned long flags;
1931 raw_spin_lock_irqsave(&rq->lock, flags);
1932 if (rq->curr->sched_class->post_schedule)
1933 rq->curr->sched_class->post_schedule(rq);
1934 raw_spin_unlock_irqrestore(&rq->lock, flags);
1936 rq->post_schedule = 0;
1942 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1946 static inline void post_schedule(struct rq *rq)
1953 * schedule_tail - first thing a freshly forked thread must call.
1954 * @prev: the thread we just switched away from.
1956 asmlinkage void schedule_tail(struct task_struct *prev)
1957 __releases(rq->lock)
1959 struct rq *rq = this_rq();
1961 finish_task_switch(rq, prev);
1964 * FIXME: do we need to worry about rq being invalidated by the
1969 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1970 /* In this case, finish_task_switch does not reenable preemption */
1973 if (current->set_child_tid)
1974 put_user(task_pid_vnr(current), current->set_child_tid);
1978 * context_switch - switch to the new MM and the new
1979 * thread's register state.
1982 context_switch(struct rq *rq, struct task_struct *prev,
1983 struct task_struct *next)
1985 struct mm_struct *mm, *oldmm;
1987 prepare_task_switch(rq, prev, next);
1990 oldmm = prev->active_mm;
1992 * For paravirt, this is coupled with an exit in switch_to to
1993 * combine the page table reload and the switch backend into
1996 arch_start_context_switch(prev);
1999 next->active_mm = oldmm;
2000 atomic_inc(&oldmm->mm_count);
2001 enter_lazy_tlb(oldmm, next);
2003 switch_mm(oldmm, mm, next);
2006 prev->active_mm = NULL;
2007 rq->prev_mm = oldmm;
2010 * Since the runqueue lock will be released by the next
2011 * task (which is an invalid locking op but in the case
2012 * of the scheduler it's an obvious special-case), so we
2013 * do an early lockdep release here:
2015 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2016 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2019 /* Here we just switch the register state and the stack. */
2020 switch_to(prev, next, prev);
2024 * this_rq must be evaluated again because prev may have moved
2025 * CPUs since it called schedule(), thus the 'rq' on its stack
2026 * frame will be invalid.
2028 finish_task_switch(this_rq(), prev);
2032 * nr_running, nr_uninterruptible and nr_context_switches:
2034 * externally visible scheduler statistics: current number of runnable
2035 * threads, current number of uninterruptible-sleeping threads, total
2036 * number of context switches performed since bootup.
2038 unsigned long nr_running(void)
2040 unsigned long i, sum = 0;
2042 for_each_online_cpu(i)
2043 sum += cpu_rq(i)->nr_running;
2048 unsigned long nr_uninterruptible(void)
2050 unsigned long i, sum = 0;
2052 for_each_possible_cpu(i)
2053 sum += cpu_rq(i)->nr_uninterruptible;
2056 * Since we read the counters lockless, it might be slightly
2057 * inaccurate. Do not allow it to go below zero though:
2059 if (unlikely((long)sum < 0))
2065 unsigned long long nr_context_switches(void)
2068 unsigned long long sum = 0;
2070 for_each_possible_cpu(i)
2071 sum += cpu_rq(i)->nr_switches;
2076 unsigned long nr_iowait(void)
2078 unsigned long i, sum = 0;
2080 for_each_possible_cpu(i)
2081 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2086 unsigned long nr_iowait_cpu(int cpu)
2088 struct rq *this = cpu_rq(cpu);
2089 return atomic_read(&this->nr_iowait);
2092 unsigned long this_cpu_load(void)
2094 struct rq *this = this_rq();
2095 return this->cpu_load[0];
2099 /* Variables and functions for calc_load */
2100 static atomic_long_t calc_load_tasks;
2101 static unsigned long calc_load_update;
2102 unsigned long avenrun[3];
2103 EXPORT_SYMBOL(avenrun);
2105 static long calc_load_fold_active(struct rq *this_rq)
2107 long nr_active, delta = 0;
2109 nr_active = this_rq->nr_running;
2110 nr_active += (long) this_rq->nr_uninterruptible;
2112 if (nr_active != this_rq->calc_load_active) {
2113 delta = nr_active - this_rq->calc_load_active;
2114 this_rq->calc_load_active = nr_active;
2120 static unsigned long
2121 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2124 load += active * (FIXED_1 - exp);
2125 load += 1UL << (FSHIFT - 1);
2126 return load >> FSHIFT;
2131 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2133 * When making the ILB scale, we should try to pull this in as well.
2135 static atomic_long_t calc_load_tasks_idle;
2137 void calc_load_account_idle(struct rq *this_rq)
2141 delta = calc_load_fold_active(this_rq);
2143 atomic_long_add(delta, &calc_load_tasks_idle);
2146 static long calc_load_fold_idle(void)
2151 * Its got a race, we don't care...
2153 if (atomic_long_read(&calc_load_tasks_idle))
2154 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2160 * fixed_power_int - compute: x^n, in O(log n) time
2162 * @x: base of the power
2163 * @frac_bits: fractional bits of @x
2164 * @n: power to raise @x to.
2166 * By exploiting the relation between the definition of the natural power
2167 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2168 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2169 * (where: n_i \elem {0, 1}, the binary vector representing n),
2170 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2171 * of course trivially computable in O(log_2 n), the length of our binary
2174 static unsigned long
2175 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2177 unsigned long result = 1UL << frac_bits;
2182 result += 1UL << (frac_bits - 1);
2183 result >>= frac_bits;
2189 x += 1UL << (frac_bits - 1);
2197 * a1 = a0 * e + a * (1 - e)
2199 * a2 = a1 * e + a * (1 - e)
2200 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2201 * = a0 * e^2 + a * (1 - e) * (1 + e)
2203 * a3 = a2 * e + a * (1 - e)
2204 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2205 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2209 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2210 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2211 * = a0 * e^n + a * (1 - e^n)
2213 * [1] application of the geometric series:
2216 * S_n := \Sum x^i = -------------
2219 static unsigned long
2220 calc_load_n(unsigned long load, unsigned long exp,
2221 unsigned long active, unsigned int n)
2224 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2228 * NO_HZ can leave us missing all per-cpu ticks calling
2229 * calc_load_account_active(), but since an idle CPU folds its delta into
2230 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2231 * in the pending idle delta if our idle period crossed a load cycle boundary.
2233 * Once we've updated the global active value, we need to apply the exponential
2234 * weights adjusted to the number of cycles missed.
2236 static void calc_global_nohz(unsigned long ticks)
2238 long delta, active, n;
2240 if (time_before(jiffies, calc_load_update))
2244 * If we crossed a calc_load_update boundary, make sure to fold
2245 * any pending idle changes, the respective CPUs might have
2246 * missed the tick driven calc_load_account_active() update
2249 delta = calc_load_fold_idle();
2251 atomic_long_add(delta, &calc_load_tasks);
2254 * If we were idle for multiple load cycles, apply them.
2256 if (ticks >= LOAD_FREQ) {
2257 n = ticks / LOAD_FREQ;
2259 active = atomic_long_read(&calc_load_tasks);
2260 active = active > 0 ? active * FIXED_1 : 0;
2262 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2263 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2264 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2266 calc_load_update += n * LOAD_FREQ;
2270 * Its possible the remainder of the above division also crosses
2271 * a LOAD_FREQ period, the regular check in calc_global_load()
2272 * which comes after this will take care of that.
2274 * Consider us being 11 ticks before a cycle completion, and us
2275 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2276 * age us 4 cycles, and the test in calc_global_load() will
2277 * pick up the final one.
2281 void calc_load_account_idle(struct rq *this_rq)
2285 static inline long calc_load_fold_idle(void)
2290 static void calc_global_nohz(unsigned long ticks)
2296 * get_avenrun - get the load average array
2297 * @loads: pointer to dest load array
2298 * @offset: offset to add
2299 * @shift: shift count to shift the result left
2301 * These values are estimates at best, so no need for locking.
2303 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2305 loads[0] = (avenrun[0] + offset) << shift;
2306 loads[1] = (avenrun[1] + offset) << shift;
2307 loads[2] = (avenrun[2] + offset) << shift;
2311 * calc_load - update the avenrun load estimates 10 ticks after the
2312 * CPUs have updated calc_load_tasks.
2314 void calc_global_load(unsigned long ticks)
2318 calc_global_nohz(ticks);
2320 if (time_before(jiffies, calc_load_update + 10))
2323 active = atomic_long_read(&calc_load_tasks);
2324 active = active > 0 ? active * FIXED_1 : 0;
2326 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2327 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2328 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2330 calc_load_update += LOAD_FREQ;
2334 * Called from update_cpu_load() to periodically update this CPU's
2337 static void calc_load_account_active(struct rq *this_rq)
2341 if (time_before(jiffies, this_rq->calc_load_update))
2344 delta = calc_load_fold_active(this_rq);
2345 delta += calc_load_fold_idle();
2347 atomic_long_add(delta, &calc_load_tasks);
2349 this_rq->calc_load_update += LOAD_FREQ;
2353 * The exact cpuload at various idx values, calculated at every tick would be
2354 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2356 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2357 * on nth tick when cpu may be busy, then we have:
2358 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2359 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2361 * decay_load_missed() below does efficient calculation of
2362 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2363 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2365 * The calculation is approximated on a 128 point scale.
2366 * degrade_zero_ticks is the number of ticks after which load at any
2367 * particular idx is approximated to be zero.
2368 * degrade_factor is a precomputed table, a row for each load idx.
2369 * Each column corresponds to degradation factor for a power of two ticks,
2370 * based on 128 point scale.
2372 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2373 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2375 * With this power of 2 load factors, we can degrade the load n times
2376 * by looking at 1 bits in n and doing as many mult/shift instead of
2377 * n mult/shifts needed by the exact degradation.
2379 #define DEGRADE_SHIFT 7
2380 static const unsigned char
2381 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2382 static const unsigned char
2383 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2384 {0, 0, 0, 0, 0, 0, 0, 0},
2385 {64, 32, 8, 0, 0, 0, 0, 0},
2386 {96, 72, 40, 12, 1, 0, 0},
2387 {112, 98, 75, 43, 15, 1, 0},
2388 {120, 112, 98, 76, 45, 16, 2} };
2391 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2392 * would be when CPU is idle and so we just decay the old load without
2393 * adding any new load.
2395 static unsigned long
2396 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2400 if (!missed_updates)
2403 if (missed_updates >= degrade_zero_ticks[idx])
2407 return load >> missed_updates;
2409 while (missed_updates) {
2410 if (missed_updates % 2)
2411 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2413 missed_updates >>= 1;
2420 * Update rq->cpu_load[] statistics. This function is usually called every
2421 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2422 * every tick. We fix it up based on jiffies.
2424 void update_cpu_load(struct rq *this_rq)
2426 unsigned long this_load = this_rq->load.weight;
2427 unsigned long curr_jiffies = jiffies;
2428 unsigned long pending_updates;
2431 this_rq->nr_load_updates++;
2433 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2434 if (curr_jiffies == this_rq->last_load_update_tick)
2437 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2438 this_rq->last_load_update_tick = curr_jiffies;
2440 /* Update our load: */
2441 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2442 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2443 unsigned long old_load, new_load;
2445 /* scale is effectively 1 << i now, and >> i divides by scale */
2447 old_load = this_rq->cpu_load[i];
2448 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2449 new_load = this_load;
2451 * Round up the averaging division if load is increasing. This
2452 * prevents us from getting stuck on 9 if the load is 10, for
2455 if (new_load > old_load)
2456 new_load += scale - 1;
2458 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2461 sched_avg_update(this_rq);
2464 static void update_cpu_load_active(struct rq *this_rq)
2466 update_cpu_load(this_rq);
2468 calc_load_account_active(this_rq);
2474 * sched_exec - execve() is a valuable balancing opportunity, because at
2475 * this point the task has the smallest effective memory and cache footprint.
2477 void sched_exec(void)
2479 struct task_struct *p = current;
2480 unsigned long flags;
2483 raw_spin_lock_irqsave(&p->pi_lock, flags);
2484 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2485 if (dest_cpu == smp_processor_id())
2488 if (likely(cpu_active(dest_cpu))) {
2489 struct migration_arg arg = { p, dest_cpu };
2491 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2492 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2496 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2501 DEFINE_PER_CPU(struct kernel_stat, kstat);
2503 EXPORT_PER_CPU_SYMBOL(kstat);
2506 * Return any ns on the sched_clock that have not yet been accounted in
2507 * @p in case that task is currently running.
2509 * Called with task_rq_lock() held on @rq.
2511 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2515 if (task_current(rq, p)) {
2516 update_rq_clock(rq);
2517 ns = rq->clock_task - p->se.exec_start;
2525 unsigned long long task_delta_exec(struct task_struct *p)
2527 unsigned long flags;
2531 rq = task_rq_lock(p, &flags);
2532 ns = do_task_delta_exec(p, rq);
2533 task_rq_unlock(rq, p, &flags);
2539 * Return accounted runtime for the task.
2540 * In case the task is currently running, return the runtime plus current's
2541 * pending runtime that have not been accounted yet.
2543 unsigned long long task_sched_runtime(struct task_struct *p)
2545 unsigned long flags;
2549 rq = task_rq_lock(p, &flags);
2550 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2551 task_rq_unlock(rq, p, &flags);
2557 * Account user cpu time to a process.
2558 * @p: the process that the cpu time gets accounted to
2559 * @cputime: the cpu time spent in user space since the last update
2560 * @cputime_scaled: cputime scaled by cpu frequency
2562 void account_user_time(struct task_struct *p, cputime_t cputime,
2563 cputime_t cputime_scaled)
2565 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2568 /* Add user time to process. */
2569 p->utime = cputime_add(p->utime, cputime);
2570 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
2571 account_group_user_time(p, cputime);
2573 /* Add user time to cpustat. */
2574 tmp = cputime_to_cputime64(cputime);
2575 if (TASK_NICE(p) > 0)
2576 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2578 cpustat->user = cputime64_add(cpustat->user, tmp);
2580 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
2581 /* Account for user time used */
2582 acct_update_integrals(p);
2586 * Account guest cpu time to a process.
2587 * @p: the process that the cpu time gets accounted to
2588 * @cputime: the cpu time spent in virtual machine since the last update
2589 * @cputime_scaled: cputime scaled by cpu frequency
2591 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2592 cputime_t cputime_scaled)
2595 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2597 tmp = cputime_to_cputime64(cputime);
2599 /* Add guest time to process. */
2600 p->utime = cputime_add(p->utime, cputime);
2601 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
2602 account_group_user_time(p, cputime);
2603 p->gtime = cputime_add(p->gtime, cputime);
2605 /* Add guest time to cpustat. */
2606 if (TASK_NICE(p) > 0) {
2607 cpustat->nice = cputime64_add(cpustat->nice, tmp);
2608 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
2610 cpustat->user = cputime64_add(cpustat->user, tmp);
2611 cpustat->guest = cputime64_add(cpustat->guest, tmp);
2616 * Account system cpu time to a process and desired cpustat field
2617 * @p: the process that the cpu time gets accounted to
2618 * @cputime: the cpu time spent in kernel space since the last update
2619 * @cputime_scaled: cputime scaled by cpu frequency
2620 * @target_cputime64: pointer to cpustat field that has to be updated
2623 void __account_system_time(struct task_struct *p, cputime_t cputime,
2624 cputime_t cputime_scaled, cputime64_t *target_cputime64)
2626 cputime64_t tmp = cputime_to_cputime64(cputime);
2628 /* Add system time to process. */
2629 p->stime = cputime_add(p->stime, cputime);
2630 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
2631 account_group_system_time(p, cputime);
2633 /* Add system time to cpustat. */
2634 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
2635 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
2637 /* Account for system time used */
2638 acct_update_integrals(p);
2642 * Account system cpu time to a process.
2643 * @p: the process that the cpu time gets accounted to
2644 * @hardirq_offset: the offset to subtract from hardirq_count()
2645 * @cputime: the cpu time spent in kernel space since the last update
2646 * @cputime_scaled: cputime scaled by cpu frequency
2648 void account_system_time(struct task_struct *p, int hardirq_offset,
2649 cputime_t cputime, cputime_t cputime_scaled)
2651 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2652 cputime64_t *target_cputime64;
2654 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2655 account_guest_time(p, cputime, cputime_scaled);
2659 if (hardirq_count() - hardirq_offset)
2660 target_cputime64 = &cpustat->irq;
2661 else if (in_serving_softirq())
2662 target_cputime64 = &cpustat->softirq;
2664 target_cputime64 = &cpustat->system;
2666 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
2670 * Account for involuntary wait time.
2671 * @cputime: the cpu time spent in involuntary wait
2673 void account_steal_time(cputime_t cputime)
2675 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2676 cputime64_t cputime64 = cputime_to_cputime64(cputime);
2678 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
2682 * Account for idle time.
2683 * @cputime: the cpu time spent in idle wait
2685 void account_idle_time(cputime_t cputime)
2687 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2688 cputime64_t cputime64 = cputime_to_cputime64(cputime);
2689 struct rq *rq = this_rq();
2691 if (atomic_read(&rq->nr_iowait) > 0)
2692 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
2694 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
2697 static __always_inline bool steal_account_process_tick(void)
2699 #ifdef CONFIG_PARAVIRT
2700 if (static_branch(¶virt_steal_enabled)) {
2703 steal = paravirt_steal_clock(smp_processor_id());
2704 steal -= this_rq()->prev_steal_time;
2706 st = steal_ticks(steal);
2707 this_rq()->prev_steal_time += st * TICK_NSEC;
2709 account_steal_time(st);
2716 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2718 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2720 * Account a tick to a process and cpustat
2721 * @p: the process that the cpu time gets accounted to
2722 * @user_tick: is the tick from userspace
2723 * @rq: the pointer to rq
2725 * Tick demultiplexing follows the order
2726 * - pending hardirq update
2727 * - pending softirq update
2731 * - check for guest_time
2732 * - else account as system_time
2734 * Check for hardirq is done both for system and user time as there is
2735 * no timer going off while we are on hardirq and hence we may never get an
2736 * opportunity to update it solely in system time.
2737 * p->stime and friends are only updated on system time and not on irq
2738 * softirq as those do not count in task exec_runtime any more.
2740 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2743 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2744 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
2745 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2747 if (steal_account_process_tick())
2750 if (irqtime_account_hi_update()) {
2751 cpustat->irq = cputime64_add(cpustat->irq, tmp);
2752 } else if (irqtime_account_si_update()) {
2753 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
2754 } else if (this_cpu_ksoftirqd() == p) {
2756 * ksoftirqd time do not get accounted in cpu_softirq_time.
2757 * So, we have to handle it separately here.
2758 * Also, p->stime needs to be updated for ksoftirqd.
2760 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2762 } else if (user_tick) {
2763 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2764 } else if (p == rq->idle) {
2765 account_idle_time(cputime_one_jiffy);
2766 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2767 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2769 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2774 static void irqtime_account_idle_ticks(int ticks)
2777 struct rq *rq = this_rq();
2779 for (i = 0; i < ticks; i++)
2780 irqtime_account_process_tick(current, 0, rq);
2782 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2783 static void irqtime_account_idle_ticks(int ticks) {}
2784 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2786 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2789 * Account a single tick of cpu time.
2790 * @p: the process that the cpu time gets accounted to
2791 * @user_tick: indicates if the tick is a user or a system tick
2793 void account_process_tick(struct task_struct *p, int user_tick)
2795 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2796 struct rq *rq = this_rq();
2798 if (sched_clock_irqtime) {
2799 irqtime_account_process_tick(p, user_tick, rq);
2803 if (steal_account_process_tick())
2807 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2808 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2809 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2812 account_idle_time(cputime_one_jiffy);
2816 * Account multiple ticks of steal time.
2817 * @p: the process from which the cpu time has been stolen
2818 * @ticks: number of stolen ticks
2820 void account_steal_ticks(unsigned long ticks)
2822 account_steal_time(jiffies_to_cputime(ticks));
2826 * Account multiple ticks of idle time.
2827 * @ticks: number of stolen ticks
2829 void account_idle_ticks(unsigned long ticks)
2832 if (sched_clock_irqtime) {
2833 irqtime_account_idle_ticks(ticks);
2837 account_idle_time(jiffies_to_cputime(ticks));
2843 * Use precise platform statistics if available:
2845 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2846 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2852 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2854 struct task_cputime cputime;
2856 thread_group_cputime(p, &cputime);
2858 *ut = cputime.utime;
2859 *st = cputime.stime;
2863 #ifndef nsecs_to_cputime
2864 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2867 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2869 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
2872 * Use CFS's precise accounting:
2874 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2880 do_div(temp, total);
2881 utime = (cputime_t)temp;
2886 * Compare with previous values, to keep monotonicity:
2888 p->prev_utime = max(p->prev_utime, utime);
2889 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
2891 *ut = p->prev_utime;
2892 *st = p->prev_stime;
2896 * Must be called with siglock held.
2898 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2900 struct signal_struct *sig = p->signal;
2901 struct task_cputime cputime;
2902 cputime_t rtime, utime, total;
2904 thread_group_cputime(p, &cputime);
2906 total = cputime_add(cputime.utime, cputime.stime);
2907 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2912 temp *= cputime.utime;
2913 do_div(temp, total);
2914 utime = (cputime_t)temp;
2918 sig->prev_utime = max(sig->prev_utime, utime);
2919 sig->prev_stime = max(sig->prev_stime,
2920 cputime_sub(rtime, sig->prev_utime));
2922 *ut = sig->prev_utime;
2923 *st = sig->prev_stime;
2928 * This function gets called by the timer code, with HZ frequency.
2929 * We call it with interrupts disabled.
2931 void scheduler_tick(void)
2933 int cpu = smp_processor_id();
2934 struct rq *rq = cpu_rq(cpu);
2935 struct task_struct *curr = rq->curr;
2939 raw_spin_lock(&rq->lock);
2940 update_rq_clock(rq);
2941 update_cpu_load_active(rq);
2942 curr->sched_class->task_tick(rq, curr, 0);
2943 raw_spin_unlock(&rq->lock);
2945 perf_event_task_tick();
2948 rq->idle_balance = idle_cpu(cpu);
2949 trigger_load_balance(rq, cpu);
2953 notrace unsigned long get_parent_ip(unsigned long addr)
2955 if (in_lock_functions(addr)) {
2956 addr = CALLER_ADDR2;
2957 if (in_lock_functions(addr))
2958 addr = CALLER_ADDR3;
2963 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2964 defined(CONFIG_PREEMPT_TRACER))
2966 void __kprobes add_preempt_count(int val)
2968 #ifdef CONFIG_DEBUG_PREEMPT
2972 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2975 preempt_count() += val;
2976 #ifdef CONFIG_DEBUG_PREEMPT
2978 * Spinlock count overflowing soon?
2980 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2983 if (preempt_count() == val)
2984 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2986 EXPORT_SYMBOL(add_preempt_count);
2988 void __kprobes sub_preempt_count(int val)
2990 #ifdef CONFIG_DEBUG_PREEMPT
2994 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2997 * Is the spinlock portion underflowing?
2999 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3000 !(preempt_count() & PREEMPT_MASK)))
3004 if (preempt_count() == val)
3005 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3006 preempt_count() -= val;
3008 EXPORT_SYMBOL(sub_preempt_count);
3013 * Print scheduling while atomic bug:
3015 static noinline void __schedule_bug(struct task_struct *prev)
3017 struct pt_regs *regs = get_irq_regs();
3019 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3020 prev->comm, prev->pid, preempt_count());
3022 debug_show_held_locks(prev);
3024 if (irqs_disabled())
3025 print_irqtrace_events(prev);
3034 * Various schedule()-time debugging checks and statistics:
3036 static inline void schedule_debug(struct task_struct *prev)
3039 * Test if we are atomic. Since do_exit() needs to call into
3040 * schedule() atomically, we ignore that path for now.
3041 * Otherwise, whine if we are scheduling when we should not be.
3043 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3044 __schedule_bug(prev);
3047 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3049 schedstat_inc(this_rq(), sched_count);
3052 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3054 if (prev->on_rq || rq->skip_clock_update < 0)
3055 update_rq_clock(rq);
3056 prev->sched_class->put_prev_task(rq, prev);
3060 * Pick up the highest-prio task:
3062 static inline struct task_struct *
3063 pick_next_task(struct rq *rq)
3065 const struct sched_class *class;
3066 struct task_struct *p;
3069 * Optimization: we know that if all tasks are in
3070 * the fair class we can call that function directly:
3072 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3073 p = fair_sched_class.pick_next_task(rq);
3078 for_each_class(class) {
3079 p = class->pick_next_task(rq);
3084 BUG(); /* the idle class will always have a runnable task */
3088 * __schedule() is the main scheduler function.
3090 static void __sched __schedule(void)
3092 struct task_struct *prev, *next;
3093 unsigned long *switch_count;
3099 cpu = smp_processor_id();
3101 rcu_note_context_switch(cpu);
3104 schedule_debug(prev);
3106 if (sched_feat(HRTICK))
3109 raw_spin_lock_irq(&rq->lock);
3111 switch_count = &prev->nivcsw;
3112 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3113 if (unlikely(signal_pending_state(prev->state, prev))) {
3114 prev->state = TASK_RUNNING;
3116 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3120 * If a worker went to sleep, notify and ask workqueue
3121 * whether it wants to wake up a task to maintain
3124 if (prev->flags & PF_WQ_WORKER) {
3125 struct task_struct *to_wakeup;
3127 to_wakeup = wq_worker_sleeping(prev, cpu);
3129 try_to_wake_up_local(to_wakeup);
3132 switch_count = &prev->nvcsw;
3135 pre_schedule(rq, prev);
3137 if (unlikely(!rq->nr_running))
3138 idle_balance(cpu, rq);
3140 put_prev_task(rq, prev);
3141 next = pick_next_task(rq);
3142 clear_tsk_need_resched(prev);
3143 rq->skip_clock_update = 0;
3145 if (likely(prev != next)) {
3150 context_switch(rq, prev, next); /* unlocks the rq */
3152 * The context switch have flipped the stack from under us
3153 * and restored the local variables which were saved when
3154 * this task called schedule() in the past. prev == current
3155 * is still correct, but it can be moved to another cpu/rq.
3157 cpu = smp_processor_id();
3160 raw_spin_unlock_irq(&rq->lock);
3164 preempt_enable_no_resched();
3169 static inline void sched_submit_work(struct task_struct *tsk)
3174 * If we are going to sleep and we have plugged IO queued,
3175 * make sure to submit it to avoid deadlocks.
3177 if (blk_needs_flush_plug(tsk))
3178 blk_schedule_flush_plug(tsk);
3181 asmlinkage void __sched schedule(void)
3183 struct task_struct *tsk = current;
3185 sched_submit_work(tsk);
3188 EXPORT_SYMBOL(schedule);
3190 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3192 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3194 if (lock->owner != owner)
3198 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3199 * lock->owner still matches owner, if that fails, owner might
3200 * point to free()d memory, if it still matches, the rcu_read_lock()
3201 * ensures the memory stays valid.
3205 return owner->on_cpu;
3209 * Look out! "owner" is an entirely speculative pointer
3210 * access and not reliable.
3212 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3214 if (!sched_feat(OWNER_SPIN))
3218 while (owner_running(lock, owner)) {
3222 arch_mutex_cpu_relax();
3227 * We break out the loop above on need_resched() and when the
3228 * owner changed, which is a sign for heavy contention. Return
3229 * success only when lock->owner is NULL.
3231 return lock->owner == NULL;
3235 #ifdef CONFIG_PREEMPT
3237 * this is the entry point to schedule() from in-kernel preemption
3238 * off of preempt_enable. Kernel preemptions off return from interrupt
3239 * occur there and call schedule directly.
3241 asmlinkage void __sched notrace preempt_schedule(void)
3243 struct thread_info *ti = current_thread_info();
3246 * If there is a non-zero preempt_count or interrupts are disabled,
3247 * we do not want to preempt the current task. Just return..
3249 if (likely(ti->preempt_count || irqs_disabled()))
3253 add_preempt_count_notrace(PREEMPT_ACTIVE);
3255 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3258 * Check again in case we missed a preemption opportunity
3259 * between schedule and now.
3262 } while (need_resched());
3264 EXPORT_SYMBOL(preempt_schedule);
3267 * this is the entry point to schedule() from kernel preemption
3268 * off of irq context.
3269 * Note, that this is called and return with irqs disabled. This will
3270 * protect us against recursive calling from irq.
3272 asmlinkage void __sched preempt_schedule_irq(void)
3274 struct thread_info *ti = current_thread_info();
3276 /* Catch callers which need to be fixed */
3277 BUG_ON(ti->preempt_count || !irqs_disabled());
3280 add_preempt_count(PREEMPT_ACTIVE);
3283 local_irq_disable();
3284 sub_preempt_count(PREEMPT_ACTIVE);
3287 * Check again in case we missed a preemption opportunity
3288 * between schedule and now.
3291 } while (need_resched());
3294 #endif /* CONFIG_PREEMPT */
3296 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3299 return try_to_wake_up(curr->private, mode, wake_flags);
3301 EXPORT_SYMBOL(default_wake_function);
3304 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3305 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3306 * number) then we wake all the non-exclusive tasks and one exclusive task.
3308 * There are circumstances in which we can try to wake a task which has already
3309 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3310 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3312 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3313 int nr_exclusive, int wake_flags, void *key)
3315 wait_queue_t *curr, *next;
3317 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3318 unsigned flags = curr->flags;
3320 if (curr->func(curr, mode, wake_flags, key) &&
3321 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3327 * __wake_up - wake up threads blocked on a waitqueue.
3329 * @mode: which threads
3330 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3331 * @key: is directly passed to the wakeup function
3333 * It may be assumed that this function implies a write memory barrier before
3334 * changing the task state if and only if any tasks are woken up.
3336 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3337 int nr_exclusive, void *key)
3339 unsigned long flags;
3341 spin_lock_irqsave(&q->lock, flags);
3342 __wake_up_common(q, mode, nr_exclusive, 0, key);
3343 spin_unlock_irqrestore(&q->lock, flags);
3345 EXPORT_SYMBOL(__wake_up);
3348 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3350 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3352 __wake_up_common(q, mode, 1, 0, NULL);
3354 EXPORT_SYMBOL_GPL(__wake_up_locked);
3356 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3358 __wake_up_common(q, mode, 1, 0, key);
3360 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3363 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3365 * @mode: which threads
3366 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3367 * @key: opaque value to be passed to wakeup targets
3369 * The sync wakeup differs that the waker knows that it will schedule
3370 * away soon, so while the target thread will be woken up, it will not
3371 * be migrated to another CPU - ie. the two threads are 'synchronized'
3372 * with each other. This can prevent needless bouncing between CPUs.
3374 * On UP it can prevent extra preemption.
3376 * It may be assumed that this function implies a write memory barrier before
3377 * changing the task state if and only if any tasks are woken up.
3379 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3380 int nr_exclusive, void *key)
3382 unsigned long flags;
3383 int wake_flags = WF_SYNC;
3388 if (unlikely(!nr_exclusive))
3391 spin_lock_irqsave(&q->lock, flags);
3392 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3393 spin_unlock_irqrestore(&q->lock, flags);
3395 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3398 * __wake_up_sync - see __wake_up_sync_key()
3400 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3402 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3404 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3407 * complete: - signals a single thread waiting on this completion
3408 * @x: holds the state of this particular completion
3410 * This will wake up a single thread waiting on this completion. Threads will be
3411 * awakened in the same order in which they were queued.
3413 * See also complete_all(), wait_for_completion() and related routines.
3415 * It may be assumed that this function implies a write memory barrier before
3416 * changing the task state if and only if any tasks are woken up.
3418 void complete(struct completion *x)
3420 unsigned long flags;
3422 spin_lock_irqsave(&x->wait.lock, flags);
3424 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3425 spin_unlock_irqrestore(&x->wait.lock, flags);
3427 EXPORT_SYMBOL(complete);
3430 * complete_all: - signals all threads waiting on this completion
3431 * @x: holds the state of this particular completion
3433 * This will wake up all threads waiting on this particular completion event.
3435 * It may be assumed that this function implies a write memory barrier before
3436 * changing the task state if and only if any tasks are woken up.
3438 void complete_all(struct completion *x)
3440 unsigned long flags;
3442 spin_lock_irqsave(&x->wait.lock, flags);
3443 x->done += UINT_MAX/2;
3444 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3445 spin_unlock_irqrestore(&x->wait.lock, flags);
3447 EXPORT_SYMBOL(complete_all);
3449 static inline long __sched
3450 do_wait_for_common(struct completion *x, long timeout, int state)
3453 DECLARE_WAITQUEUE(wait, current);
3455 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3457 if (signal_pending_state(state, current)) {
3458 timeout = -ERESTARTSYS;
3461 __set_current_state(state);
3462 spin_unlock_irq(&x->wait.lock);
3463 timeout = schedule_timeout(timeout);
3464 spin_lock_irq(&x->wait.lock);
3465 } while (!x->done && timeout);
3466 __remove_wait_queue(&x->wait, &wait);
3471 return timeout ?: 1;
3475 wait_for_common(struct completion *x, long timeout, int state)
3479 spin_lock_irq(&x->wait.lock);
3480 timeout = do_wait_for_common(x, timeout, state);
3481 spin_unlock_irq(&x->wait.lock);
3486 * wait_for_completion: - waits for completion of a task
3487 * @x: holds the state of this particular completion
3489 * This waits to be signaled for completion of a specific task. It is NOT
3490 * interruptible and there is no timeout.
3492 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3493 * and interrupt capability. Also see complete().
3495 void __sched wait_for_completion(struct completion *x)
3497 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3499 EXPORT_SYMBOL(wait_for_completion);
3502 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3503 * @x: holds the state of this particular completion
3504 * @timeout: timeout value in jiffies
3506 * This waits for either a completion of a specific task to be signaled or for a
3507 * specified timeout to expire. The timeout is in jiffies. It is not
3510 * The return value is 0 if timed out, and positive (at least 1, or number of
3511 * jiffies left till timeout) if completed.
3513 unsigned long __sched
3514 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3516 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3518 EXPORT_SYMBOL(wait_for_completion_timeout);
3521 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3522 * @x: holds the state of this particular completion
3524 * This waits for completion of a specific task to be signaled. It is
3527 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3529 int __sched wait_for_completion_interruptible(struct completion *x)
3531 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3532 if (t == -ERESTARTSYS)
3536 EXPORT_SYMBOL(wait_for_completion_interruptible);
3539 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3540 * @x: holds the state of this particular completion
3541 * @timeout: timeout value in jiffies
3543 * This waits for either a completion of a specific task to be signaled or for a
3544 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3546 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3547 * positive (at least 1, or number of jiffies left till timeout) if completed.
3550 wait_for_completion_interruptible_timeout(struct completion *x,
3551 unsigned long timeout)
3553 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3555 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3558 * wait_for_completion_killable: - waits for completion of a task (killable)
3559 * @x: holds the state of this particular completion
3561 * This waits to be signaled for completion of a specific task. It can be
3562 * interrupted by a kill signal.
3564 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3566 int __sched wait_for_completion_killable(struct completion *x)
3568 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3569 if (t == -ERESTARTSYS)
3573 EXPORT_SYMBOL(wait_for_completion_killable);
3576 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3577 * @x: holds the state of this particular completion
3578 * @timeout: timeout value in jiffies
3580 * This waits for either a completion of a specific task to be
3581 * signaled or for a specified timeout to expire. It can be
3582 * interrupted by a kill signal. The timeout is in jiffies.
3584 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3585 * positive (at least 1, or number of jiffies left till timeout) if completed.
3588 wait_for_completion_killable_timeout(struct completion *x,
3589 unsigned long timeout)
3591 return wait_for_common(x, timeout, TASK_KILLABLE);
3593 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3596 * try_wait_for_completion - try to decrement a completion without blocking
3597 * @x: completion structure
3599 * Returns: 0 if a decrement cannot be done without blocking
3600 * 1 if a decrement succeeded.
3602 * If a completion is being used as a counting completion,
3603 * attempt to decrement the counter without blocking. This
3604 * enables us to avoid waiting if the resource the completion
3605 * is protecting is not available.
3607 bool try_wait_for_completion(struct completion *x)
3609 unsigned long flags;
3612 spin_lock_irqsave(&x->wait.lock, flags);
3617 spin_unlock_irqrestore(&x->wait.lock, flags);
3620 EXPORT_SYMBOL(try_wait_for_completion);
3623 * completion_done - Test to see if a completion has any waiters
3624 * @x: completion structure
3626 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3627 * 1 if there are no waiters.
3630 bool completion_done(struct completion *x)
3632 unsigned long flags;
3635 spin_lock_irqsave(&x->wait.lock, flags);
3638 spin_unlock_irqrestore(&x->wait.lock, flags);
3641 EXPORT_SYMBOL(completion_done);
3644 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3646 unsigned long flags;
3649 init_waitqueue_entry(&wait, current);
3651 __set_current_state(state);
3653 spin_lock_irqsave(&q->lock, flags);
3654 __add_wait_queue(q, &wait);
3655 spin_unlock(&q->lock);
3656 timeout = schedule_timeout(timeout);
3657 spin_lock_irq(&q->lock);
3658 __remove_wait_queue(q, &wait);
3659 spin_unlock_irqrestore(&q->lock, flags);
3664 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3666 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3668 EXPORT_SYMBOL(interruptible_sleep_on);
3671 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3673 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3675 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3677 void __sched sleep_on(wait_queue_head_t *q)
3679 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3681 EXPORT_SYMBOL(sleep_on);
3683 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3685 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3687 EXPORT_SYMBOL(sleep_on_timeout);
3689 #ifdef CONFIG_RT_MUTEXES
3692 * rt_mutex_setprio - set the current priority of a task
3694 * @prio: prio value (kernel-internal form)
3696 * This function changes the 'effective' priority of a task. It does
3697 * not touch ->normal_prio like __setscheduler().
3699 * Used by the rt_mutex code to implement priority inheritance logic.
3701 void rt_mutex_setprio(struct task_struct *p, int prio)
3703 int oldprio, on_rq, running;
3705 const struct sched_class *prev_class;
3707 BUG_ON(prio < 0 || prio > MAX_PRIO);
3709 rq = __task_rq_lock(p);
3711 trace_sched_pi_setprio(p, prio);
3713 prev_class = p->sched_class;
3715 running = task_current(rq, p);
3717 dequeue_task(rq, p, 0);
3719 p->sched_class->put_prev_task(rq, p);
3722 p->sched_class = &rt_sched_class;
3724 p->sched_class = &fair_sched_class;
3729 p->sched_class->set_curr_task(rq);
3731 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3733 check_class_changed(rq, p, prev_class, oldprio);
3734 __task_rq_unlock(rq);
3739 void set_user_nice(struct task_struct *p, long nice)
3741 int old_prio, delta, on_rq;
3742 unsigned long flags;
3745 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3748 * We have to be careful, if called from sys_setpriority(),
3749 * the task might be in the middle of scheduling on another CPU.
3751 rq = task_rq_lock(p, &flags);
3753 * The RT priorities are set via sched_setscheduler(), but we still
3754 * allow the 'normal' nice value to be set - but as expected
3755 * it wont have any effect on scheduling until the task is
3756 * SCHED_FIFO/SCHED_RR:
3758 if (task_has_rt_policy(p)) {
3759 p->static_prio = NICE_TO_PRIO(nice);
3764 dequeue_task(rq, p, 0);
3766 p->static_prio = NICE_TO_PRIO(nice);
3769 p->prio = effective_prio(p);
3770 delta = p->prio - old_prio;
3773 enqueue_task(rq, p, 0);
3775 * If the task increased its priority or is running and
3776 * lowered its priority, then reschedule its CPU:
3778 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3779 resched_task(rq->curr);
3782 task_rq_unlock(rq, p, &flags);
3784 EXPORT_SYMBOL(set_user_nice);
3787 * can_nice - check if a task can reduce its nice value
3791 int can_nice(const struct task_struct *p, const int nice)
3793 /* convert nice value [19,-20] to rlimit style value [1,40] */
3794 int nice_rlim = 20 - nice;
3796 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3797 capable(CAP_SYS_NICE));
3800 #ifdef __ARCH_WANT_SYS_NICE
3803 * sys_nice - change the priority of the current process.
3804 * @increment: priority increment
3806 * sys_setpriority is a more generic, but much slower function that
3807 * does similar things.
3809 SYSCALL_DEFINE1(nice, int, increment)
3814 * Setpriority might change our priority at the same moment.
3815 * We don't have to worry. Conceptually one call occurs first
3816 * and we have a single winner.
3818 if (increment < -40)
3823 nice = TASK_NICE(current) + increment;
3829 if (increment < 0 && !can_nice(current, nice))
3832 retval = security_task_setnice(current, nice);
3836 set_user_nice(current, nice);
3843 * task_prio - return the priority value of a given task.
3844 * @p: the task in question.
3846 * This is the priority value as seen by users in /proc.
3847 * RT tasks are offset by -200. Normal tasks are centered
3848 * around 0, value goes from -16 to +15.
3850 int task_prio(const struct task_struct *p)
3852 return p->prio - MAX_RT_PRIO;
3856 * task_nice - return the nice value of a given task.
3857 * @p: the task in question.
3859 int task_nice(const struct task_struct *p)
3861 return TASK_NICE(p);
3863 EXPORT_SYMBOL(task_nice);
3866 * idle_cpu - is a given cpu idle currently?
3867 * @cpu: the processor in question.
3869 int idle_cpu(int cpu)
3871 struct rq *rq = cpu_rq(cpu);
3873 if (rq->curr != rq->idle)
3880 if (!llist_empty(&rq->wake_list))
3888 * idle_task - return the idle task for a given cpu.
3889 * @cpu: the processor in question.
3891 struct task_struct *idle_task(int cpu)
3893 return cpu_rq(cpu)->idle;
3897 * find_process_by_pid - find a process with a matching PID value.
3898 * @pid: the pid in question.
3900 static struct task_struct *find_process_by_pid(pid_t pid)
3902 return pid ? find_task_by_vpid(pid) : current;
3905 /* Actually do priority change: must hold rq lock. */
3907 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3910 p->rt_priority = prio;
3911 p->normal_prio = normal_prio(p);
3912 /* we are holding p->pi_lock already */
3913 p->prio = rt_mutex_getprio(p);
3914 if (rt_prio(p->prio))
3915 p->sched_class = &rt_sched_class;
3917 p->sched_class = &fair_sched_class;
3922 * check the target process has a UID that matches the current process's
3924 static bool check_same_owner(struct task_struct *p)
3926 const struct cred *cred = current_cred(), *pcred;
3930 pcred = __task_cred(p);
3931 if (cred->user->user_ns == pcred->user->user_ns)
3932 match = (cred->euid == pcred->euid ||
3933 cred->euid == pcred->uid);
3940 static int __sched_setscheduler(struct task_struct *p, int policy,
3941 const struct sched_param *param, bool user)
3943 int retval, oldprio, oldpolicy = -1, on_rq, running;
3944 unsigned long flags;
3945 const struct sched_class *prev_class;
3949 /* may grab non-irq protected spin_locks */
3950 BUG_ON(in_interrupt());
3952 /* double check policy once rq lock held */
3954 reset_on_fork = p->sched_reset_on_fork;
3955 policy = oldpolicy = p->policy;
3957 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3958 policy &= ~SCHED_RESET_ON_FORK;
3960 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3961 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3962 policy != SCHED_IDLE)
3967 * Valid priorities for SCHED_FIFO and SCHED_RR are
3968 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3969 * SCHED_BATCH and SCHED_IDLE is 0.
3971 if (param->sched_priority < 0 ||
3972 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3973 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3975 if (rt_policy(policy) != (param->sched_priority != 0))
3979 * Allow unprivileged RT tasks to decrease priority:
3981 if (user && !capable(CAP_SYS_NICE)) {
3982 if (rt_policy(policy)) {
3983 unsigned long rlim_rtprio =
3984 task_rlimit(p, RLIMIT_RTPRIO);
3986 /* can't set/change the rt policy */
3987 if (policy != p->policy && !rlim_rtprio)
3990 /* can't increase priority */
3991 if (param->sched_priority > p->rt_priority &&
3992 param->sched_priority > rlim_rtprio)
3997 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3998 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4000 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4001 if (!can_nice(p, TASK_NICE(p)))
4005 /* can't change other user's priorities */
4006 if (!check_same_owner(p))
4009 /* Normal users shall not reset the sched_reset_on_fork flag */
4010 if (p->sched_reset_on_fork && !reset_on_fork)
4015 retval = security_task_setscheduler(p);
4021 * make sure no PI-waiters arrive (or leave) while we are
4022 * changing the priority of the task:
4024 * To be able to change p->policy safely, the appropriate
4025 * runqueue lock must be held.
4027 rq = task_rq_lock(p, &flags);
4030 * Changing the policy of the stop threads its a very bad idea
4032 if (p == rq->stop) {
4033 task_rq_unlock(rq, p, &flags);
4038 * If not changing anything there's no need to proceed further:
4040 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4041 param->sched_priority == p->rt_priority))) {
4043 __task_rq_unlock(rq);
4044 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4048 #ifdef CONFIG_RT_GROUP_SCHED
4051 * Do not allow realtime tasks into groups that have no runtime
4054 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4055 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4056 !task_group_is_autogroup(task_group(p))) {
4057 task_rq_unlock(rq, p, &flags);
4063 /* recheck policy now with rq lock held */
4064 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4065 policy = oldpolicy = -1;
4066 task_rq_unlock(rq, p, &flags);
4070 running = task_current(rq, p);
4072 deactivate_task(rq, p, 0);
4074 p->sched_class->put_prev_task(rq, p);
4076 p->sched_reset_on_fork = reset_on_fork;
4079 prev_class = p->sched_class;
4080 __setscheduler(rq, p, policy, param->sched_priority);
4083 p->sched_class->set_curr_task(rq);
4085 activate_task(rq, p, 0);
4087 check_class_changed(rq, p, prev_class, oldprio);
4088 task_rq_unlock(rq, p, &flags);
4090 rt_mutex_adjust_pi(p);
4096 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4097 * @p: the task in question.
4098 * @policy: new policy.
4099 * @param: structure containing the new RT priority.
4101 * NOTE that the task may be already dead.
4103 int sched_setscheduler(struct task_struct *p, int policy,
4104 const struct sched_param *param)
4106 return __sched_setscheduler(p, policy, param, true);
4108 EXPORT_SYMBOL_GPL(sched_setscheduler);
4111 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4112 * @p: the task in question.
4113 * @policy: new policy.
4114 * @param: structure containing the new RT priority.
4116 * Just like sched_setscheduler, only don't bother checking if the
4117 * current context has permission. For example, this is needed in
4118 * stop_machine(): we create temporary high priority worker threads,
4119 * but our caller might not have that capability.
4121 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4122 const struct sched_param *param)
4124 return __sched_setscheduler(p, policy, param, false);
4128 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4130 struct sched_param lparam;
4131 struct task_struct *p;
4134 if (!param || pid < 0)
4136 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4141 p = find_process_by_pid(pid);
4143 retval = sched_setscheduler(p, policy, &lparam);
4150 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4151 * @pid: the pid in question.
4152 * @policy: new policy.
4153 * @param: structure containing the new RT priority.
4155 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4156 struct sched_param __user *, param)
4158 /* negative values for policy are not valid */
4162 return do_sched_setscheduler(pid, policy, param);
4166 * sys_sched_setparam - set/change the RT priority of a thread
4167 * @pid: the pid in question.
4168 * @param: structure containing the new RT priority.
4170 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4172 return do_sched_setscheduler(pid, -1, param);
4176 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4177 * @pid: the pid in question.
4179 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4181 struct task_struct *p;
4189 p = find_process_by_pid(pid);
4191 retval = security_task_getscheduler(p);
4194 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4201 * sys_sched_getparam - get the RT priority of a thread
4202 * @pid: the pid in question.
4203 * @param: structure containing the RT priority.
4205 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4207 struct sched_param lp;
4208 struct task_struct *p;
4211 if (!param || pid < 0)
4215 p = find_process_by_pid(pid);
4220 retval = security_task_getscheduler(p);
4224 lp.sched_priority = p->rt_priority;
4228 * This one might sleep, we cannot do it with a spinlock held ...
4230 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4239 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4241 cpumask_var_t cpus_allowed, new_mask;
4242 struct task_struct *p;
4248 p = find_process_by_pid(pid);
4255 /* Prevent p going away */
4259 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4263 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4265 goto out_free_cpus_allowed;
4268 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
4271 retval = security_task_setscheduler(p);
4275 cpuset_cpus_allowed(p, cpus_allowed);
4276 cpumask_and(new_mask, in_mask, cpus_allowed);
4278 retval = set_cpus_allowed_ptr(p, new_mask);
4281 cpuset_cpus_allowed(p, cpus_allowed);
4282 if (!cpumask_subset(new_mask, cpus_allowed)) {
4284 * We must have raced with a concurrent cpuset
4285 * update. Just reset the cpus_allowed to the
4286 * cpuset's cpus_allowed
4288 cpumask_copy(new_mask, cpus_allowed);
4293 free_cpumask_var(new_mask);
4294 out_free_cpus_allowed:
4295 free_cpumask_var(cpus_allowed);
4302 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4303 struct cpumask *new_mask)
4305 if (len < cpumask_size())
4306 cpumask_clear(new_mask);
4307 else if (len > cpumask_size())
4308 len = cpumask_size();
4310 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4314 * sys_sched_setaffinity - set the cpu affinity of a process
4315 * @pid: pid of the process
4316 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4317 * @user_mask_ptr: user-space pointer to the new cpu mask
4319 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4320 unsigned long __user *, user_mask_ptr)
4322 cpumask_var_t new_mask;
4325 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4328 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4330 retval = sched_setaffinity(pid, new_mask);
4331 free_cpumask_var(new_mask);
4335 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4337 struct task_struct *p;
4338 unsigned long flags;
4345 p = find_process_by_pid(pid);
4349 retval = security_task_getscheduler(p);
4353 raw_spin_lock_irqsave(&p->pi_lock, flags);
4354 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4355 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4365 * sys_sched_getaffinity - get the cpu affinity of a process
4366 * @pid: pid of the process
4367 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4368 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4370 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4371 unsigned long __user *, user_mask_ptr)
4376 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4378 if (len & (sizeof(unsigned long)-1))
4381 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4384 ret = sched_getaffinity(pid, mask);
4386 size_t retlen = min_t(size_t, len, cpumask_size());
4388 if (copy_to_user(user_mask_ptr, mask, retlen))
4393 free_cpumask_var(mask);
4399 * sys_sched_yield - yield the current processor to other threads.
4401 * This function yields the current CPU to other tasks. If there are no
4402 * other threads running on this CPU then this function will return.
4404 SYSCALL_DEFINE0(sched_yield)
4406 struct rq *rq = this_rq_lock();
4408 schedstat_inc(rq, yld_count);
4409 current->sched_class->yield_task(rq);
4412 * Since we are going to call schedule() anyway, there's
4413 * no need to preempt or enable interrupts:
4415 __release(rq->lock);
4416 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4417 do_raw_spin_unlock(&rq->lock);
4418 preempt_enable_no_resched();
4425 static inline int should_resched(void)
4427 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4430 static void __cond_resched(void)
4432 add_preempt_count(PREEMPT_ACTIVE);
4434 sub_preempt_count(PREEMPT_ACTIVE);
4437 int __sched _cond_resched(void)
4439 if (should_resched()) {
4445 EXPORT_SYMBOL(_cond_resched);
4448 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4449 * call schedule, and on return reacquire the lock.
4451 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4452 * operations here to prevent schedule() from being called twice (once via
4453 * spin_unlock(), once by hand).
4455 int __cond_resched_lock(spinlock_t *lock)
4457 int resched = should_resched();
4460 lockdep_assert_held(lock);
4462 if (spin_needbreak(lock) || resched) {
4473 EXPORT_SYMBOL(__cond_resched_lock);
4475 int __sched __cond_resched_softirq(void)
4477 BUG_ON(!in_softirq());
4479 if (should_resched()) {
4487 EXPORT_SYMBOL(__cond_resched_softirq);
4490 * yield - yield the current processor to other threads.
4492 * This is a shortcut for kernel-space yielding - it marks the
4493 * thread runnable and calls sys_sched_yield().
4495 void __sched yield(void)
4497 set_current_state(TASK_RUNNING);
4500 EXPORT_SYMBOL(yield);
4503 * yield_to - yield the current processor to another thread in
4504 * your thread group, or accelerate that thread toward the
4505 * processor it's on.
4507 * @preempt: whether task preemption is allowed or not
4509 * It's the caller's job to ensure that the target task struct
4510 * can't go away on us before we can do any checks.
4512 * Returns true if we indeed boosted the target task.
4514 bool __sched yield_to(struct task_struct *p, bool preempt)
4516 struct task_struct *curr = current;
4517 struct rq *rq, *p_rq;
4518 unsigned long flags;
4521 local_irq_save(flags);
4526 double_rq_lock(rq, p_rq);
4527 while (task_rq(p) != p_rq) {
4528 double_rq_unlock(rq, p_rq);
4532 if (!curr->sched_class->yield_to_task)
4535 if (curr->sched_class != p->sched_class)
4538 if (task_running(p_rq, p) || p->state)
4541 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4543 schedstat_inc(rq, yld_count);
4545 * Make p's CPU reschedule; pick_next_entity takes care of
4548 if (preempt && rq != p_rq)
4549 resched_task(p_rq->curr);
4553 double_rq_unlock(rq, p_rq);
4554 local_irq_restore(flags);
4561 EXPORT_SYMBOL_GPL(yield_to);
4564 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4565 * that process accounting knows that this is a task in IO wait state.
4567 void __sched io_schedule(void)
4569 struct rq *rq = raw_rq();
4571 delayacct_blkio_start();
4572 atomic_inc(&rq->nr_iowait);
4573 blk_flush_plug(current);
4574 current->in_iowait = 1;
4576 current->in_iowait = 0;
4577 atomic_dec(&rq->nr_iowait);
4578 delayacct_blkio_end();
4580 EXPORT_SYMBOL(io_schedule);
4582 long __sched io_schedule_timeout(long timeout)
4584 struct rq *rq = raw_rq();
4587 delayacct_blkio_start();
4588 atomic_inc(&rq->nr_iowait);
4589 blk_flush_plug(current);
4590 current->in_iowait = 1;
4591 ret = schedule_timeout(timeout);
4592 current->in_iowait = 0;
4593 atomic_dec(&rq->nr_iowait);
4594 delayacct_blkio_end();
4599 * sys_sched_get_priority_max - return maximum RT priority.
4600 * @policy: scheduling class.
4602 * this syscall returns the maximum rt_priority that can be used
4603 * by a given scheduling class.
4605 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4612 ret = MAX_USER_RT_PRIO-1;
4624 * sys_sched_get_priority_min - return minimum RT priority.
4625 * @policy: scheduling class.
4627 * this syscall returns the minimum rt_priority that can be used
4628 * by a given scheduling class.
4630 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4648 * sys_sched_rr_get_interval - return the default timeslice of a process.
4649 * @pid: pid of the process.
4650 * @interval: userspace pointer to the timeslice value.
4652 * this syscall writes the default timeslice value of a given process
4653 * into the user-space timespec buffer. A value of '0' means infinity.
4655 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4656 struct timespec __user *, interval)
4658 struct task_struct *p;
4659 unsigned int time_slice;
4660 unsigned long flags;
4670 p = find_process_by_pid(pid);
4674 retval = security_task_getscheduler(p);
4678 rq = task_rq_lock(p, &flags);
4679 time_slice = p->sched_class->get_rr_interval(rq, p);
4680 task_rq_unlock(rq, p, &flags);
4683 jiffies_to_timespec(time_slice, &t);
4684 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4692 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4694 void sched_show_task(struct task_struct *p)
4696 unsigned long free = 0;
4699 state = p->state ? __ffs(p->state) + 1 : 0;
4700 printk(KERN_INFO "%-15.15s %c", p->comm,
4701 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4702 #if BITS_PER_LONG == 32
4703 if (state == TASK_RUNNING)
4704 printk(KERN_CONT " running ");
4706 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4708 if (state == TASK_RUNNING)
4709 printk(KERN_CONT " running task ");
4711 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4713 #ifdef CONFIG_DEBUG_STACK_USAGE
4714 free = stack_not_used(p);
4716 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4717 task_pid_nr(p), task_pid_nr(p->real_parent),
4718 (unsigned long)task_thread_info(p)->flags);
4720 show_stack(p, NULL);
4723 void show_state_filter(unsigned long state_filter)
4725 struct task_struct *g, *p;
4727 #if BITS_PER_LONG == 32
4729 " task PC stack pid father\n");
4732 " task PC stack pid father\n");
4735 do_each_thread(g, p) {
4737 * reset the NMI-timeout, listing all files on a slow
4738 * console might take a lot of time:
4740 touch_nmi_watchdog();
4741 if (!state_filter || (p->state & state_filter))
4743 } while_each_thread(g, p);
4745 touch_all_softlockup_watchdogs();
4747 #ifdef CONFIG_SCHED_DEBUG
4748 sysrq_sched_debug_show();
4752 * Only show locks if all tasks are dumped:
4755 debug_show_all_locks();
4758 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4760 idle->sched_class = &idle_sched_class;
4764 * init_idle - set up an idle thread for a given CPU
4765 * @idle: task in question
4766 * @cpu: cpu the idle task belongs to
4768 * NOTE: this function does not set the idle thread's NEED_RESCHED
4769 * flag, to make booting more robust.
4771 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4773 struct rq *rq = cpu_rq(cpu);
4774 unsigned long flags;
4776 raw_spin_lock_irqsave(&rq->lock, flags);
4779 idle->state = TASK_RUNNING;
4780 idle->se.exec_start = sched_clock();
4782 do_set_cpus_allowed(idle, cpumask_of(cpu));
4784 * We're having a chicken and egg problem, even though we are
4785 * holding rq->lock, the cpu isn't yet set to this cpu so the
4786 * lockdep check in task_group() will fail.
4788 * Similar case to sched_fork(). / Alternatively we could
4789 * use task_rq_lock() here and obtain the other rq->lock.
4794 __set_task_cpu(idle, cpu);
4797 rq->curr = rq->idle = idle;
4798 #if defined(CONFIG_SMP)
4801 raw_spin_unlock_irqrestore(&rq->lock, flags);
4803 /* Set the preempt count _outside_ the spinlocks! */
4804 task_thread_info(idle)->preempt_count = 0;
4807 * The idle tasks have their own, simple scheduling class:
4809 idle->sched_class = &idle_sched_class;
4810 ftrace_graph_init_idle_task(idle, cpu);
4811 #if defined(CONFIG_SMP)
4812 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4817 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4819 if (p->sched_class && p->sched_class->set_cpus_allowed)
4820 p->sched_class->set_cpus_allowed(p, new_mask);
4822 cpumask_copy(&p->cpus_allowed, new_mask);
4823 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4827 * This is how migration works:
4829 * 1) we invoke migration_cpu_stop() on the target CPU using
4831 * 2) stopper starts to run (implicitly forcing the migrated thread
4833 * 3) it checks whether the migrated task is still in the wrong runqueue.
4834 * 4) if it's in the wrong runqueue then the migration thread removes
4835 * it and puts it into the right queue.
4836 * 5) stopper completes and stop_one_cpu() returns and the migration
4841 * Change a given task's CPU affinity. Migrate the thread to a
4842 * proper CPU and schedule it away if the CPU it's executing on
4843 * is removed from the allowed bitmask.
4845 * NOTE: the caller must have a valid reference to the task, the
4846 * task must not exit() & deallocate itself prematurely. The
4847 * call is not atomic; no spinlocks may be held.
4849 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4851 unsigned long flags;
4853 unsigned int dest_cpu;
4856 rq = task_rq_lock(p, &flags);
4858 if (cpumask_equal(&p->cpus_allowed, new_mask))
4861 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4866 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4871 do_set_cpus_allowed(p, new_mask);
4873 /* Can the task run on the task's current CPU? If so, we're done */
4874 if (cpumask_test_cpu(task_cpu(p), new_mask))
4877 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4879 struct migration_arg arg = { p, dest_cpu };
4880 /* Need help from migration thread: drop lock and wait. */
4881 task_rq_unlock(rq, p, &flags);
4882 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4883 tlb_migrate_finish(p->mm);
4887 task_rq_unlock(rq, p, &flags);
4891 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4894 * Move (not current) task off this cpu, onto dest cpu. We're doing
4895 * this because either it can't run here any more (set_cpus_allowed()
4896 * away from this CPU, or CPU going down), or because we're
4897 * attempting to rebalance this task on exec (sched_exec).
4899 * So we race with normal scheduler movements, but that's OK, as long
4900 * as the task is no longer on this CPU.
4902 * Returns non-zero if task was successfully migrated.
4904 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4906 struct rq *rq_dest, *rq_src;
4909 if (unlikely(!cpu_active(dest_cpu)))
4912 rq_src = cpu_rq(src_cpu);
4913 rq_dest = cpu_rq(dest_cpu);
4915 raw_spin_lock(&p->pi_lock);
4916 double_rq_lock(rq_src, rq_dest);
4917 /* Already moved. */
4918 if (task_cpu(p) != src_cpu)
4920 /* Affinity changed (again). */
4921 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4925 * If we're not on a rq, the next wake-up will ensure we're
4929 deactivate_task(rq_src, p, 0);
4930 set_task_cpu(p, dest_cpu);
4931 activate_task(rq_dest, p, 0);
4932 check_preempt_curr(rq_dest, p, 0);
4937 double_rq_unlock(rq_src, rq_dest);
4938 raw_spin_unlock(&p->pi_lock);
4943 * migration_cpu_stop - this will be executed by a highprio stopper thread
4944 * and performs thread migration by bumping thread off CPU then
4945 * 'pushing' onto another runqueue.
4947 static int migration_cpu_stop(void *data)
4949 struct migration_arg *arg = data;
4952 * The original target cpu might have gone down and we might
4953 * be on another cpu but it doesn't matter.
4955 local_irq_disable();
4956 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4961 #ifdef CONFIG_HOTPLUG_CPU
4964 * Ensures that the idle task is using init_mm right before its cpu goes
4967 void idle_task_exit(void)
4969 struct mm_struct *mm = current->active_mm;
4971 BUG_ON(cpu_online(smp_processor_id()));
4974 switch_mm(mm, &init_mm, current);
4979 * While a dead CPU has no uninterruptible tasks queued at this point,
4980 * it might still have a nonzero ->nr_uninterruptible counter, because
4981 * for performance reasons the counter is not stricly tracking tasks to
4982 * their home CPUs. So we just add the counter to another CPU's counter,
4983 * to keep the global sum constant after CPU-down:
4985 static void migrate_nr_uninterruptible(struct rq *rq_src)
4987 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
4989 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
4990 rq_src->nr_uninterruptible = 0;
4994 * remove the tasks which were accounted by rq from calc_load_tasks.
4996 static void calc_global_load_remove(struct rq *rq)
4998 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
4999 rq->calc_load_active = 0;
5003 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5004 * try_to_wake_up()->select_task_rq().
5006 * Called with rq->lock held even though we'er in stop_machine() and
5007 * there's no concurrency possible, we hold the required locks anyway
5008 * because of lock validation efforts.
5010 static void migrate_tasks(unsigned int dead_cpu)
5012 struct rq *rq = cpu_rq(dead_cpu);
5013 struct task_struct *next, *stop = rq->stop;
5017 * Fudge the rq selection such that the below task selection loop
5018 * doesn't get stuck on the currently eligible stop task.
5020 * We're currently inside stop_machine() and the rq is either stuck
5021 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5022 * either way we should never end up calling schedule() until we're
5027 /* Ensure any throttled groups are reachable by pick_next_task */
5028 unthrottle_offline_cfs_rqs(rq);
5032 * There's this thread running, bail when that's the only
5035 if (rq->nr_running == 1)
5038 next = pick_next_task(rq);
5040 next->sched_class->put_prev_task(rq, next);
5042 /* Find suitable destination for @next, with force if needed. */
5043 dest_cpu = select_fallback_rq(dead_cpu, next);
5044 raw_spin_unlock(&rq->lock);
5046 __migrate_task(next, dead_cpu, dest_cpu);
5048 raw_spin_lock(&rq->lock);
5054 #endif /* CONFIG_HOTPLUG_CPU */
5056 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5058 static struct ctl_table sd_ctl_dir[] = {
5060 .procname = "sched_domain",
5066 static struct ctl_table sd_ctl_root[] = {
5068 .procname = "kernel",
5070 .child = sd_ctl_dir,
5075 static struct ctl_table *sd_alloc_ctl_entry(int n)
5077 struct ctl_table *entry =
5078 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5083 static void sd_free_ctl_entry(struct ctl_table **tablep)
5085 struct ctl_table *entry;
5088 * In the intermediate directories, both the child directory and
5089 * procname are dynamically allocated and could fail but the mode
5090 * will always be set. In the lowest directory the names are
5091 * static strings and all have proc handlers.
5093 for (entry = *tablep; entry->mode; entry++) {
5095 sd_free_ctl_entry(&entry->child);
5096 if (entry->proc_handler == NULL)
5097 kfree(entry->procname);
5105 set_table_entry(struct ctl_table *entry,
5106 const char *procname, void *data, int maxlen,
5107 mode_t mode, proc_handler *proc_handler)
5109 entry->procname = procname;
5111 entry->maxlen = maxlen;
5113 entry->proc_handler = proc_handler;
5116 static struct ctl_table *
5117 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5119 struct ctl_table *table = sd_alloc_ctl_entry(13);
5124 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5125 sizeof(long), 0644, proc_doulongvec_minmax);
5126 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5127 sizeof(long), 0644, proc_doulongvec_minmax);
5128 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5129 sizeof(int), 0644, proc_dointvec_minmax);
5130 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5131 sizeof(int), 0644, proc_dointvec_minmax);
5132 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5133 sizeof(int), 0644, proc_dointvec_minmax);
5134 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5135 sizeof(int), 0644, proc_dointvec_minmax);
5136 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5137 sizeof(int), 0644, proc_dointvec_minmax);
5138 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5139 sizeof(int), 0644, proc_dointvec_minmax);
5140 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5141 sizeof(int), 0644, proc_dointvec_minmax);
5142 set_table_entry(&table[9], "cache_nice_tries",
5143 &sd->cache_nice_tries,
5144 sizeof(int), 0644, proc_dointvec_minmax);
5145 set_table_entry(&table[10], "flags", &sd->flags,
5146 sizeof(int), 0644, proc_dointvec_minmax);
5147 set_table_entry(&table[11], "name", sd->name,
5148 CORENAME_MAX_SIZE, 0444, proc_dostring);
5149 /* &table[12] is terminator */
5154 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5156 struct ctl_table *entry, *table;
5157 struct sched_domain *sd;
5158 int domain_num = 0, i;
5161 for_each_domain(cpu, sd)
5163 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5168 for_each_domain(cpu, sd) {
5169 snprintf(buf, 32, "domain%d", i);
5170 entry->procname = kstrdup(buf, GFP_KERNEL);
5172 entry->child = sd_alloc_ctl_domain_table(sd);
5179 static struct ctl_table_header *sd_sysctl_header;
5180 static void register_sched_domain_sysctl(void)
5182 int i, cpu_num = num_possible_cpus();
5183 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5186 WARN_ON(sd_ctl_dir[0].child);
5187 sd_ctl_dir[0].child = entry;
5192 for_each_possible_cpu(i) {
5193 snprintf(buf, 32, "cpu%d", i);
5194 entry->procname = kstrdup(buf, GFP_KERNEL);
5196 entry->child = sd_alloc_ctl_cpu_table(i);
5200 WARN_ON(sd_sysctl_header);
5201 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5204 /* may be called multiple times per register */
5205 static void unregister_sched_domain_sysctl(void)
5207 if (sd_sysctl_header)
5208 unregister_sysctl_table(sd_sysctl_header);
5209 sd_sysctl_header = NULL;
5210 if (sd_ctl_dir[0].child)
5211 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5214 static void register_sched_domain_sysctl(void)
5217 static void unregister_sched_domain_sysctl(void)
5222 static void set_rq_online(struct rq *rq)
5225 const struct sched_class *class;
5227 cpumask_set_cpu(rq->cpu, rq->rd->online);
5230 for_each_class(class) {
5231 if (class->rq_online)
5232 class->rq_online(rq);
5237 static void set_rq_offline(struct rq *rq)
5240 const struct sched_class *class;
5242 for_each_class(class) {
5243 if (class->rq_offline)
5244 class->rq_offline(rq);
5247 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5253 * migration_call - callback that gets triggered when a CPU is added.
5254 * Here we can start up the necessary migration thread for the new CPU.
5256 static int __cpuinit
5257 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5259 int cpu = (long)hcpu;
5260 unsigned long flags;
5261 struct rq *rq = cpu_rq(cpu);
5263 switch (action & ~CPU_TASKS_FROZEN) {
5265 case CPU_UP_PREPARE:
5266 rq->calc_load_update = calc_load_update;
5270 /* Update our root-domain */
5271 raw_spin_lock_irqsave(&rq->lock, flags);
5273 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5277 raw_spin_unlock_irqrestore(&rq->lock, flags);
5280 #ifdef CONFIG_HOTPLUG_CPU
5282 sched_ttwu_pending();
5283 /* Update our root-domain */
5284 raw_spin_lock_irqsave(&rq->lock, flags);
5286 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5290 BUG_ON(rq->nr_running != 1); /* the migration thread */
5291 raw_spin_unlock_irqrestore(&rq->lock, flags);
5293 migrate_nr_uninterruptible(rq);
5294 calc_global_load_remove(rq);
5299 update_max_interval();
5305 * Register at high priority so that task migration (migrate_all_tasks)
5306 * happens before everything else. This has to be lower priority than
5307 * the notifier in the perf_event subsystem, though.
5309 static struct notifier_block __cpuinitdata migration_notifier = {
5310 .notifier_call = migration_call,
5311 .priority = CPU_PRI_MIGRATION,
5314 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5315 unsigned long action, void *hcpu)
5317 switch (action & ~CPU_TASKS_FROZEN) {
5319 case CPU_DOWN_FAILED:
5320 set_cpu_active((long)hcpu, true);
5327 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5328 unsigned long action, void *hcpu)
5330 switch (action & ~CPU_TASKS_FROZEN) {
5331 case CPU_DOWN_PREPARE:
5332 set_cpu_active((long)hcpu, false);
5339 static int __init migration_init(void)
5341 void *cpu = (void *)(long)smp_processor_id();
5344 /* Initialize migration for the boot CPU */
5345 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5346 BUG_ON(err == NOTIFY_BAD);
5347 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5348 register_cpu_notifier(&migration_notifier);
5350 /* Register cpu active notifiers */
5351 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5352 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5356 early_initcall(migration_init);
5361 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5363 #ifdef CONFIG_SCHED_DEBUG
5365 static __read_mostly int sched_domain_debug_enabled;
5367 static int __init sched_domain_debug_setup(char *str)
5369 sched_domain_debug_enabled = 1;
5373 early_param("sched_debug", sched_domain_debug_setup);
5375 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5376 struct cpumask *groupmask)
5378 struct sched_group *group = sd->groups;
5381 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5382 cpumask_clear(groupmask);
5384 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5386 if (!(sd->flags & SD_LOAD_BALANCE)) {
5387 printk("does not load-balance\n");
5389 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5394 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5396 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5397 printk(KERN_ERR "ERROR: domain->span does not contain "
5400 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5401 printk(KERN_ERR "ERROR: domain->groups does not contain"
5405 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5409 printk(KERN_ERR "ERROR: group is NULL\n");
5413 if (!group->sgp->power) {
5414 printk(KERN_CONT "\n");
5415 printk(KERN_ERR "ERROR: domain->cpu_power not "
5420 if (!cpumask_weight(sched_group_cpus(group))) {
5421 printk(KERN_CONT "\n");
5422 printk(KERN_ERR "ERROR: empty group\n");
5426 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5427 printk(KERN_CONT "\n");
5428 printk(KERN_ERR "ERROR: repeated CPUs\n");
5432 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5434 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5436 printk(KERN_CONT " %s", str);
5437 if (group->sgp->power != SCHED_POWER_SCALE) {
5438 printk(KERN_CONT " (cpu_power = %d)",
5442 group = group->next;
5443 } while (group != sd->groups);
5444 printk(KERN_CONT "\n");
5446 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5447 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5450 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5451 printk(KERN_ERR "ERROR: parent span is not a superset "
5452 "of domain->span\n");
5456 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5460 if (!sched_domain_debug_enabled)
5464 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5468 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5471 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5479 #else /* !CONFIG_SCHED_DEBUG */
5480 # define sched_domain_debug(sd, cpu) do { } while (0)
5481 #endif /* CONFIG_SCHED_DEBUG */
5483 static int sd_degenerate(struct sched_domain *sd)
5485 if (cpumask_weight(sched_domain_span(sd)) == 1)
5488 /* Following flags need at least 2 groups */
5489 if (sd->flags & (SD_LOAD_BALANCE |
5490 SD_BALANCE_NEWIDLE |
5494 SD_SHARE_PKG_RESOURCES)) {
5495 if (sd->groups != sd->groups->next)
5499 /* Following flags don't use groups */
5500 if (sd->flags & (SD_WAKE_AFFINE))
5507 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5509 unsigned long cflags = sd->flags, pflags = parent->flags;
5511 if (sd_degenerate(parent))
5514 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5517 /* Flags needing groups don't count if only 1 group in parent */
5518 if (parent->groups == parent->groups->next) {
5519 pflags &= ~(SD_LOAD_BALANCE |
5520 SD_BALANCE_NEWIDLE |
5524 SD_SHARE_PKG_RESOURCES);
5525 if (nr_node_ids == 1)
5526 pflags &= ~SD_SERIALIZE;
5528 if (~cflags & pflags)
5534 static void free_rootdomain(struct rcu_head *rcu)
5536 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5538 cpupri_cleanup(&rd->cpupri);
5539 free_cpumask_var(rd->rto_mask);
5540 free_cpumask_var(rd->online);
5541 free_cpumask_var(rd->span);
5545 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5547 struct root_domain *old_rd = NULL;
5548 unsigned long flags;
5550 raw_spin_lock_irqsave(&rq->lock, flags);
5555 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5558 cpumask_clear_cpu(rq->cpu, old_rd->span);
5561 * If we dont want to free the old_rt yet then
5562 * set old_rd to NULL to skip the freeing later
5565 if (!atomic_dec_and_test(&old_rd->refcount))
5569 atomic_inc(&rd->refcount);
5572 cpumask_set_cpu(rq->cpu, rd->span);
5573 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5576 raw_spin_unlock_irqrestore(&rq->lock, flags);
5579 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5582 static int init_rootdomain(struct root_domain *rd)
5584 memset(rd, 0, sizeof(*rd));
5586 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5588 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5590 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5593 if (cpupri_init(&rd->cpupri) != 0)
5598 free_cpumask_var(rd->rto_mask);
5600 free_cpumask_var(rd->online);
5602 free_cpumask_var(rd->span);
5608 * By default the system creates a single root-domain with all cpus as
5609 * members (mimicking the global state we have today).
5611 struct root_domain def_root_domain;
5613 static void init_defrootdomain(void)
5615 init_rootdomain(&def_root_domain);
5617 atomic_set(&def_root_domain.refcount, 1);
5620 static struct root_domain *alloc_rootdomain(void)
5622 struct root_domain *rd;
5624 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5628 if (init_rootdomain(rd) != 0) {
5636 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5638 struct sched_group *tmp, *first;
5647 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5652 } while (sg != first);
5655 static void free_sched_domain(struct rcu_head *rcu)
5657 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5660 * If its an overlapping domain it has private groups, iterate and
5663 if (sd->flags & SD_OVERLAP) {
5664 free_sched_groups(sd->groups, 1);
5665 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5666 kfree(sd->groups->sgp);
5672 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5674 call_rcu(&sd->rcu, free_sched_domain);
5677 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5679 for (; sd; sd = sd->parent)
5680 destroy_sched_domain(sd, cpu);
5684 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5685 * hold the hotplug lock.
5688 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5690 struct rq *rq = cpu_rq(cpu);
5691 struct sched_domain *tmp;
5693 /* Remove the sched domains which do not contribute to scheduling. */
5694 for (tmp = sd; tmp; ) {
5695 struct sched_domain *parent = tmp->parent;
5699 if (sd_parent_degenerate(tmp, parent)) {
5700 tmp->parent = parent->parent;
5702 parent->parent->child = tmp;
5703 destroy_sched_domain(parent, cpu);
5708 if (sd && sd_degenerate(sd)) {
5711 destroy_sched_domain(tmp, cpu);
5716 sched_domain_debug(sd, cpu);
5718 rq_attach_root(rq, rd);
5720 rcu_assign_pointer(rq->sd, sd);
5721 destroy_sched_domains(tmp, cpu);
5724 /* cpus with isolated domains */
5725 static cpumask_var_t cpu_isolated_map;
5727 /* Setup the mask of cpus configured for isolated domains */
5728 static int __init isolated_cpu_setup(char *str)
5730 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5731 cpulist_parse(str, cpu_isolated_map);
5735 __setup("isolcpus=", isolated_cpu_setup);
5740 * find_next_best_node - find the next node to include in a sched_domain
5741 * @node: node whose sched_domain we're building
5742 * @used_nodes: nodes already in the sched_domain
5744 * Find the next node to include in a given scheduling domain. Simply
5745 * finds the closest node not already in the @used_nodes map.
5747 * Should use nodemask_t.
5749 static int find_next_best_node(int node, nodemask_t *used_nodes)
5751 int i, n, val, min_val, best_node = -1;
5755 for (i = 0; i < nr_node_ids; i++) {
5756 /* Start at @node */
5757 n = (node + i) % nr_node_ids;
5759 if (!nr_cpus_node(n))
5762 /* Skip already used nodes */
5763 if (node_isset(n, *used_nodes))
5766 /* Simple min distance search */
5767 val = node_distance(node, n);
5769 if (val < min_val) {
5775 if (best_node != -1)
5776 node_set(best_node, *used_nodes);
5781 * sched_domain_node_span - get a cpumask for a node's sched_domain
5782 * @node: node whose cpumask we're constructing
5783 * @span: resulting cpumask
5785 * Given a node, construct a good cpumask for its sched_domain to span. It
5786 * should be one that prevents unnecessary balancing, but also spreads tasks
5789 static void sched_domain_node_span(int node, struct cpumask *span)
5791 nodemask_t used_nodes;
5794 cpumask_clear(span);
5795 nodes_clear(used_nodes);
5797 cpumask_or(span, span, cpumask_of_node(node));
5798 node_set(node, used_nodes);
5800 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5801 int next_node = find_next_best_node(node, &used_nodes);
5804 cpumask_or(span, span, cpumask_of_node(next_node));
5808 static const struct cpumask *cpu_node_mask(int cpu)
5810 lockdep_assert_held(&sched_domains_mutex);
5812 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5814 return sched_domains_tmpmask;
5817 static const struct cpumask *cpu_allnodes_mask(int cpu)
5819 return cpu_possible_mask;
5821 #endif /* CONFIG_NUMA */
5823 static const struct cpumask *cpu_cpu_mask(int cpu)
5825 return cpumask_of_node(cpu_to_node(cpu));
5828 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5831 struct sched_domain **__percpu sd;
5832 struct sched_group **__percpu sg;
5833 struct sched_group_power **__percpu sgp;
5837 struct sched_domain ** __percpu sd;
5838 struct root_domain *rd;
5848 struct sched_domain_topology_level;
5850 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5851 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5853 #define SDTL_OVERLAP 0x01
5855 struct sched_domain_topology_level {
5856 sched_domain_init_f init;
5857 sched_domain_mask_f mask;
5859 struct sd_data data;
5863 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5865 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5866 const struct cpumask *span = sched_domain_span(sd);
5867 struct cpumask *covered = sched_domains_tmpmask;
5868 struct sd_data *sdd = sd->private;
5869 struct sched_domain *child;
5872 cpumask_clear(covered);
5874 for_each_cpu(i, span) {
5875 struct cpumask *sg_span;
5877 if (cpumask_test_cpu(i, covered))
5880 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5881 GFP_KERNEL, cpu_to_node(i));
5886 sg_span = sched_group_cpus(sg);
5888 child = *per_cpu_ptr(sdd->sd, i);
5890 child = child->child;
5891 cpumask_copy(sg_span, sched_domain_span(child));
5893 cpumask_set_cpu(i, sg_span);
5895 cpumask_or(covered, covered, sg_span);
5897 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5898 atomic_inc(&sg->sgp->ref);
5900 if (cpumask_test_cpu(cpu, sg_span))
5910 sd->groups = groups;
5915 free_sched_groups(first, 0);
5920 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5922 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5923 struct sched_domain *child = sd->child;
5926 cpu = cpumask_first(sched_domain_span(child));
5929 *sg = *per_cpu_ptr(sdd->sg, cpu);
5930 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5931 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5938 * build_sched_groups will build a circular linked list of the groups
5939 * covered by the given span, and will set each group's ->cpumask correctly,
5940 * and ->cpu_power to 0.
5942 * Assumes the sched_domain tree is fully constructed
5945 build_sched_groups(struct sched_domain *sd, int cpu)
5947 struct sched_group *first = NULL, *last = NULL;
5948 struct sd_data *sdd = sd->private;
5949 const struct cpumask *span = sched_domain_span(sd);
5950 struct cpumask *covered;
5953 get_group(cpu, sdd, &sd->groups);
5954 atomic_inc(&sd->groups->ref);
5956 if (cpu != cpumask_first(sched_domain_span(sd)))
5959 lockdep_assert_held(&sched_domains_mutex);
5960 covered = sched_domains_tmpmask;
5962 cpumask_clear(covered);
5964 for_each_cpu(i, span) {
5965 struct sched_group *sg;
5966 int group = get_group(i, sdd, &sg);
5969 if (cpumask_test_cpu(i, covered))
5972 cpumask_clear(sched_group_cpus(sg));
5975 for_each_cpu(j, span) {
5976 if (get_group(j, sdd, NULL) != group)
5979 cpumask_set_cpu(j, covered);
5980 cpumask_set_cpu(j, sched_group_cpus(sg));
5995 * Initialize sched groups cpu_power.
5997 * cpu_power indicates the capacity of sched group, which is used while
5998 * distributing the load between different sched groups in a sched domain.
5999 * Typically cpu_power for all the groups in a sched domain will be same unless
6000 * there are asymmetries in the topology. If there are asymmetries, group
6001 * having more cpu_power will pickup more load compared to the group having
6004 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6006 struct sched_group *sg = sd->groups;
6008 WARN_ON(!sd || !sg);
6011 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6013 } while (sg != sd->groups);
6015 if (cpu != group_first_cpu(sg))
6018 update_group_power(sd, cpu);
6021 int __weak arch_sd_sibling_asym_packing(void)
6023 return 0*SD_ASYM_PACKING;
6027 * Initializers for schedule domains
6028 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6031 #ifdef CONFIG_SCHED_DEBUG
6032 # define SD_INIT_NAME(sd, type) sd->name = #type
6034 # define SD_INIT_NAME(sd, type) do { } while (0)
6037 #define SD_INIT_FUNC(type) \
6038 static noinline struct sched_domain * \
6039 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6041 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6042 *sd = SD_##type##_INIT; \
6043 SD_INIT_NAME(sd, type); \
6044 sd->private = &tl->data; \
6050 SD_INIT_FUNC(ALLNODES)
6053 #ifdef CONFIG_SCHED_SMT
6054 SD_INIT_FUNC(SIBLING)
6056 #ifdef CONFIG_SCHED_MC
6059 #ifdef CONFIG_SCHED_BOOK
6063 static int default_relax_domain_level = -1;
6064 int sched_domain_level_max;
6066 static int __init setup_relax_domain_level(char *str)
6070 val = simple_strtoul(str, NULL, 0);
6071 if (val < sched_domain_level_max)
6072 default_relax_domain_level = val;
6076 __setup("relax_domain_level=", setup_relax_domain_level);
6078 static void set_domain_attribute(struct sched_domain *sd,
6079 struct sched_domain_attr *attr)
6083 if (!attr || attr->relax_domain_level < 0) {
6084 if (default_relax_domain_level < 0)
6087 request = default_relax_domain_level;
6089 request = attr->relax_domain_level;
6090 if (request < sd->level) {
6091 /* turn off idle balance on this domain */
6092 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6094 /* turn on idle balance on this domain */
6095 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6099 static void __sdt_free(const struct cpumask *cpu_map);
6100 static int __sdt_alloc(const struct cpumask *cpu_map);
6102 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6103 const struct cpumask *cpu_map)
6107 if (!atomic_read(&d->rd->refcount))
6108 free_rootdomain(&d->rd->rcu); /* fall through */
6110 free_percpu(d->sd); /* fall through */
6112 __sdt_free(cpu_map); /* fall through */
6118 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6119 const struct cpumask *cpu_map)
6121 memset(d, 0, sizeof(*d));
6123 if (__sdt_alloc(cpu_map))
6124 return sa_sd_storage;
6125 d->sd = alloc_percpu(struct sched_domain *);
6127 return sa_sd_storage;
6128 d->rd = alloc_rootdomain();
6131 return sa_rootdomain;
6135 * NULL the sd_data elements we've used to build the sched_domain and
6136 * sched_group structure so that the subsequent __free_domain_allocs()
6137 * will not free the data we're using.
6139 static void claim_allocations(int cpu, struct sched_domain *sd)
6141 struct sd_data *sdd = sd->private;
6143 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6144 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6146 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6147 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6149 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6150 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6153 #ifdef CONFIG_SCHED_SMT
6154 static const struct cpumask *cpu_smt_mask(int cpu)
6156 return topology_thread_cpumask(cpu);
6161 * Topology list, bottom-up.
6163 static struct sched_domain_topology_level default_topology[] = {
6164 #ifdef CONFIG_SCHED_SMT
6165 { sd_init_SIBLING, cpu_smt_mask, },
6167 #ifdef CONFIG_SCHED_MC
6168 { sd_init_MC, cpu_coregroup_mask, },
6170 #ifdef CONFIG_SCHED_BOOK
6171 { sd_init_BOOK, cpu_book_mask, },
6173 { sd_init_CPU, cpu_cpu_mask, },
6175 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6176 { sd_init_ALLNODES, cpu_allnodes_mask, },
6181 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6183 static int __sdt_alloc(const struct cpumask *cpu_map)
6185 struct sched_domain_topology_level *tl;
6188 for (tl = sched_domain_topology; tl->init; tl++) {
6189 struct sd_data *sdd = &tl->data;
6191 sdd->sd = alloc_percpu(struct sched_domain *);
6195 sdd->sg = alloc_percpu(struct sched_group *);
6199 sdd->sgp = alloc_percpu(struct sched_group_power *);
6203 for_each_cpu(j, cpu_map) {
6204 struct sched_domain *sd;
6205 struct sched_group *sg;
6206 struct sched_group_power *sgp;
6208 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6209 GFP_KERNEL, cpu_to_node(j));
6213 *per_cpu_ptr(sdd->sd, j) = sd;
6215 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6216 GFP_KERNEL, cpu_to_node(j));
6220 *per_cpu_ptr(sdd->sg, j) = sg;
6222 sgp = kzalloc_node(sizeof(struct sched_group_power),
6223 GFP_KERNEL, cpu_to_node(j));
6227 *per_cpu_ptr(sdd->sgp, j) = sgp;
6234 static void __sdt_free(const struct cpumask *cpu_map)
6236 struct sched_domain_topology_level *tl;
6239 for (tl = sched_domain_topology; tl->init; tl++) {
6240 struct sd_data *sdd = &tl->data;
6242 for_each_cpu(j, cpu_map) {
6243 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6244 if (sd && (sd->flags & SD_OVERLAP))
6245 free_sched_groups(sd->groups, 0);
6246 kfree(*per_cpu_ptr(sdd->sd, j));
6247 kfree(*per_cpu_ptr(sdd->sg, j));
6248 kfree(*per_cpu_ptr(sdd->sgp, j));
6250 free_percpu(sdd->sd);
6251 free_percpu(sdd->sg);
6252 free_percpu(sdd->sgp);
6256 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6257 struct s_data *d, const struct cpumask *cpu_map,
6258 struct sched_domain_attr *attr, struct sched_domain *child,
6261 struct sched_domain *sd = tl->init(tl, cpu);
6265 set_domain_attribute(sd, attr);
6266 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6268 sd->level = child->level + 1;
6269 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6278 * Build sched domains for a given set of cpus and attach the sched domains
6279 * to the individual cpus
6281 static int build_sched_domains(const struct cpumask *cpu_map,
6282 struct sched_domain_attr *attr)
6284 enum s_alloc alloc_state = sa_none;
6285 struct sched_domain *sd;
6287 int i, ret = -ENOMEM;
6289 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6290 if (alloc_state != sa_rootdomain)
6293 /* Set up domains for cpus specified by the cpu_map. */
6294 for_each_cpu(i, cpu_map) {
6295 struct sched_domain_topology_level *tl;
6298 for (tl = sched_domain_topology; tl->init; tl++) {
6299 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6300 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6301 sd->flags |= SD_OVERLAP;
6302 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6309 *per_cpu_ptr(d.sd, i) = sd;
6312 /* Build the groups for the domains */
6313 for_each_cpu(i, cpu_map) {
6314 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6315 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6316 if (sd->flags & SD_OVERLAP) {
6317 if (build_overlap_sched_groups(sd, i))
6320 if (build_sched_groups(sd, i))
6326 /* Calculate CPU power for physical packages and nodes */
6327 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6328 if (!cpumask_test_cpu(i, cpu_map))
6331 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6332 claim_allocations(i, sd);
6333 init_sched_groups_power(i, sd);
6337 /* Attach the domains */
6339 for_each_cpu(i, cpu_map) {
6340 sd = *per_cpu_ptr(d.sd, i);
6341 cpu_attach_domain(sd, d.rd, i);
6347 __free_domain_allocs(&d, alloc_state, cpu_map);
6351 static cpumask_var_t *doms_cur; /* current sched domains */
6352 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6353 static struct sched_domain_attr *dattr_cur;
6354 /* attribues of custom domains in 'doms_cur' */
6357 * Special case: If a kmalloc of a doms_cur partition (array of
6358 * cpumask) fails, then fallback to a single sched domain,
6359 * as determined by the single cpumask fallback_doms.
6361 static cpumask_var_t fallback_doms;
6364 * arch_update_cpu_topology lets virtualized architectures update the
6365 * cpu core maps. It is supposed to return 1 if the topology changed
6366 * or 0 if it stayed the same.
6368 int __attribute__((weak)) arch_update_cpu_topology(void)
6373 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6376 cpumask_var_t *doms;
6378 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6381 for (i = 0; i < ndoms; i++) {
6382 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6383 free_sched_domains(doms, i);
6390 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6393 for (i = 0; i < ndoms; i++)
6394 free_cpumask_var(doms[i]);
6399 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6400 * For now this just excludes isolated cpus, but could be used to
6401 * exclude other special cases in the future.
6403 static int init_sched_domains(const struct cpumask *cpu_map)
6407 arch_update_cpu_topology();
6409 doms_cur = alloc_sched_domains(ndoms_cur);
6411 doms_cur = &fallback_doms;
6412 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6414 err = build_sched_domains(doms_cur[0], NULL);
6415 register_sched_domain_sysctl();
6421 * Detach sched domains from a group of cpus specified in cpu_map
6422 * These cpus will now be attached to the NULL domain
6424 static void detach_destroy_domains(const struct cpumask *cpu_map)
6429 for_each_cpu(i, cpu_map)
6430 cpu_attach_domain(NULL, &def_root_domain, i);
6434 /* handle null as "default" */
6435 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6436 struct sched_domain_attr *new, int idx_new)
6438 struct sched_domain_attr tmp;
6445 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6446 new ? (new + idx_new) : &tmp,
6447 sizeof(struct sched_domain_attr));
6451 * Partition sched domains as specified by the 'ndoms_new'
6452 * cpumasks in the array doms_new[] of cpumasks. This compares
6453 * doms_new[] to the current sched domain partitioning, doms_cur[].
6454 * It destroys each deleted domain and builds each new domain.
6456 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6457 * The masks don't intersect (don't overlap.) We should setup one
6458 * sched domain for each mask. CPUs not in any of the cpumasks will
6459 * not be load balanced. If the same cpumask appears both in the
6460 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6463 * The passed in 'doms_new' should be allocated using
6464 * alloc_sched_domains. This routine takes ownership of it and will
6465 * free_sched_domains it when done with it. If the caller failed the
6466 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6467 * and partition_sched_domains() will fallback to the single partition
6468 * 'fallback_doms', it also forces the domains to be rebuilt.
6470 * If doms_new == NULL it will be replaced with cpu_online_mask.
6471 * ndoms_new == 0 is a special case for destroying existing domains,
6472 * and it will not create the default domain.
6474 * Call with hotplug lock held
6476 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6477 struct sched_domain_attr *dattr_new)
6482 mutex_lock(&sched_domains_mutex);
6484 /* always unregister in case we don't destroy any domains */
6485 unregister_sched_domain_sysctl();
6487 /* Let architecture update cpu core mappings. */
6488 new_topology = arch_update_cpu_topology();
6490 n = doms_new ? ndoms_new : 0;
6492 /* Destroy deleted domains */
6493 for (i = 0; i < ndoms_cur; i++) {
6494 for (j = 0; j < n && !new_topology; j++) {
6495 if (cpumask_equal(doms_cur[i], doms_new[j])
6496 && dattrs_equal(dattr_cur, i, dattr_new, j))
6499 /* no match - a current sched domain not in new doms_new[] */
6500 detach_destroy_domains(doms_cur[i]);
6505 if (doms_new == NULL) {
6507 doms_new = &fallback_doms;
6508 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6509 WARN_ON_ONCE(dattr_new);
6512 /* Build new domains */
6513 for (i = 0; i < ndoms_new; i++) {
6514 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6515 if (cpumask_equal(doms_new[i], doms_cur[j])
6516 && dattrs_equal(dattr_new, i, dattr_cur, j))
6519 /* no match - add a new doms_new */
6520 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6525 /* Remember the new sched domains */
6526 if (doms_cur != &fallback_doms)
6527 free_sched_domains(doms_cur, ndoms_cur);
6528 kfree(dattr_cur); /* kfree(NULL) is safe */
6529 doms_cur = doms_new;
6530 dattr_cur = dattr_new;
6531 ndoms_cur = ndoms_new;
6533 register_sched_domain_sysctl();
6535 mutex_unlock(&sched_domains_mutex);
6538 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6539 static void reinit_sched_domains(void)
6543 /* Destroy domains first to force the rebuild */
6544 partition_sched_domains(0, NULL, NULL);
6546 rebuild_sched_domains();
6550 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6552 unsigned int level = 0;
6554 if (sscanf(buf, "%u", &level) != 1)
6558 * level is always be positive so don't check for
6559 * level < POWERSAVINGS_BALANCE_NONE which is 0
6560 * What happens on 0 or 1 byte write,
6561 * need to check for count as well?
6564 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6568 sched_smt_power_savings = level;
6570 sched_mc_power_savings = level;
6572 reinit_sched_domains();
6577 #ifdef CONFIG_SCHED_MC
6578 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
6579 struct sysdev_class_attribute *attr,
6582 return sprintf(page, "%u\n", sched_mc_power_savings);
6584 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
6585 struct sysdev_class_attribute *attr,
6586 const char *buf, size_t count)
6588 return sched_power_savings_store(buf, count, 0);
6590 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
6591 sched_mc_power_savings_show,
6592 sched_mc_power_savings_store);
6595 #ifdef CONFIG_SCHED_SMT
6596 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
6597 struct sysdev_class_attribute *attr,
6600 return sprintf(page, "%u\n", sched_smt_power_savings);
6602 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
6603 struct sysdev_class_attribute *attr,
6604 const char *buf, size_t count)
6606 return sched_power_savings_store(buf, count, 1);
6608 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
6609 sched_smt_power_savings_show,
6610 sched_smt_power_savings_store);
6613 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6617 #ifdef CONFIG_SCHED_SMT
6619 err = sysfs_create_file(&cls->kset.kobj,
6620 &attr_sched_smt_power_savings.attr);
6622 #ifdef CONFIG_SCHED_MC
6623 if (!err && mc_capable())
6624 err = sysfs_create_file(&cls->kset.kobj,
6625 &attr_sched_mc_power_savings.attr);
6629 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6632 * Update cpusets according to cpu_active mask. If cpusets are
6633 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6634 * around partition_sched_domains().
6636 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6639 switch (action & ~CPU_TASKS_FROZEN) {
6641 case CPU_DOWN_FAILED:
6642 cpuset_update_active_cpus();
6649 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6652 switch (action & ~CPU_TASKS_FROZEN) {
6653 case CPU_DOWN_PREPARE:
6654 cpuset_update_active_cpus();
6661 void __init sched_init_smp(void)
6663 cpumask_var_t non_isolated_cpus;
6665 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6666 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6669 mutex_lock(&sched_domains_mutex);
6670 init_sched_domains(cpu_active_mask);
6671 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6672 if (cpumask_empty(non_isolated_cpus))
6673 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6674 mutex_unlock(&sched_domains_mutex);
6677 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6678 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6680 /* RT runtime code needs to handle some hotplug events */
6681 hotcpu_notifier(update_runtime, 0);
6685 /* Move init over to a non-isolated CPU */
6686 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6688 sched_init_granularity();
6689 free_cpumask_var(non_isolated_cpus);
6691 init_sched_rt_class();
6694 void __init sched_init_smp(void)
6696 sched_init_granularity();
6698 #endif /* CONFIG_SMP */
6700 const_debug unsigned int sysctl_timer_migration = 1;
6702 int in_sched_functions(unsigned long addr)
6704 return in_lock_functions(addr) ||
6705 (addr >= (unsigned long)__sched_text_start
6706 && addr < (unsigned long)__sched_text_end);
6709 #ifdef CONFIG_CGROUP_SCHED
6710 struct task_group root_task_group;
6713 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6715 void __init sched_init(void)
6718 unsigned long alloc_size = 0, ptr;
6720 #ifdef CONFIG_FAIR_GROUP_SCHED
6721 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6723 #ifdef CONFIG_RT_GROUP_SCHED
6724 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6726 #ifdef CONFIG_CPUMASK_OFFSTACK
6727 alloc_size += num_possible_cpus() * cpumask_size();
6730 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6732 #ifdef CONFIG_FAIR_GROUP_SCHED
6733 root_task_group.se = (struct sched_entity **)ptr;
6734 ptr += nr_cpu_ids * sizeof(void **);
6736 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6737 ptr += nr_cpu_ids * sizeof(void **);
6739 #endif /* CONFIG_FAIR_GROUP_SCHED */
6740 #ifdef CONFIG_RT_GROUP_SCHED
6741 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6742 ptr += nr_cpu_ids * sizeof(void **);
6744 root_task_group.rt_rq = (struct rt_rq **)ptr;
6745 ptr += nr_cpu_ids * sizeof(void **);
6747 #endif /* CONFIG_RT_GROUP_SCHED */
6748 #ifdef CONFIG_CPUMASK_OFFSTACK
6749 for_each_possible_cpu(i) {
6750 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6751 ptr += cpumask_size();
6753 #endif /* CONFIG_CPUMASK_OFFSTACK */
6757 init_defrootdomain();
6760 init_rt_bandwidth(&def_rt_bandwidth,
6761 global_rt_period(), global_rt_runtime());
6763 #ifdef CONFIG_RT_GROUP_SCHED
6764 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6765 global_rt_period(), global_rt_runtime());
6766 #endif /* CONFIG_RT_GROUP_SCHED */
6768 #ifdef CONFIG_CGROUP_SCHED
6769 list_add(&root_task_group.list, &task_groups);
6770 INIT_LIST_HEAD(&root_task_group.children);
6771 INIT_LIST_HEAD(&root_task_group.siblings);
6772 autogroup_init(&init_task);
6773 #endif /* CONFIG_CGROUP_SCHED */
6775 for_each_possible_cpu(i) {
6779 raw_spin_lock_init(&rq->lock);
6781 rq->calc_load_active = 0;
6782 rq->calc_load_update = jiffies + LOAD_FREQ;
6783 init_cfs_rq(&rq->cfs);
6784 init_rt_rq(&rq->rt, rq);
6785 #ifdef CONFIG_FAIR_GROUP_SCHED
6786 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6787 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6789 * How much cpu bandwidth does root_task_group get?
6791 * In case of task-groups formed thr' the cgroup filesystem, it
6792 * gets 100% of the cpu resources in the system. This overall
6793 * system cpu resource is divided among the tasks of
6794 * root_task_group and its child task-groups in a fair manner,
6795 * based on each entity's (task or task-group's) weight
6796 * (se->load.weight).
6798 * In other words, if root_task_group has 10 tasks of weight
6799 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6800 * then A0's share of the cpu resource is:
6802 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6804 * We achieve this by letting root_task_group's tasks sit
6805 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6807 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6808 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6809 #endif /* CONFIG_FAIR_GROUP_SCHED */
6811 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6812 #ifdef CONFIG_RT_GROUP_SCHED
6813 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6814 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6817 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6818 rq->cpu_load[j] = 0;
6820 rq->last_load_update_tick = jiffies;
6825 rq->cpu_power = SCHED_POWER_SCALE;
6826 rq->post_schedule = 0;
6827 rq->active_balance = 0;
6828 rq->next_balance = jiffies;
6833 rq->avg_idle = 2*sysctl_sched_migration_cost;
6834 rq_attach_root(rq, &def_root_domain);
6836 rq->nohz_balance_kick = 0;
6840 atomic_set(&rq->nr_iowait, 0);
6843 set_load_weight(&init_task);
6845 #ifdef CONFIG_PREEMPT_NOTIFIERS
6846 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6849 #ifdef CONFIG_RT_MUTEXES
6850 plist_head_init(&init_task.pi_waiters);
6854 * The boot idle thread does lazy MMU switching as well:
6856 atomic_inc(&init_mm.mm_count);
6857 enter_lazy_tlb(&init_mm, current);
6860 * Make us the idle thread. Technically, schedule() should not be
6861 * called from this thread, however somewhere below it might be,
6862 * but because we are the idle thread, we just pick up running again
6863 * when this runqueue becomes "idle".
6865 init_idle(current, smp_processor_id());
6867 calc_load_update = jiffies + LOAD_FREQ;
6870 * During early bootup we pretend to be a normal task:
6872 current->sched_class = &fair_sched_class;
6875 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6876 /* May be allocated at isolcpus cmdline parse time */
6877 if (cpu_isolated_map == NULL)
6878 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6880 init_sched_fair_class();
6882 scheduler_running = 1;
6885 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6886 static inline int preempt_count_equals(int preempt_offset)
6888 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6890 return (nested == preempt_offset);
6893 void __might_sleep(const char *file, int line, int preempt_offset)
6895 static unsigned long prev_jiffy; /* ratelimiting */
6897 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6898 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6899 system_state != SYSTEM_RUNNING || oops_in_progress)
6901 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6903 prev_jiffy = jiffies;
6906 "BUG: sleeping function called from invalid context at %s:%d\n",
6909 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6910 in_atomic(), irqs_disabled(),
6911 current->pid, current->comm);
6913 debug_show_held_locks(current);
6914 if (irqs_disabled())
6915 print_irqtrace_events(current);
6918 EXPORT_SYMBOL(__might_sleep);
6921 #ifdef CONFIG_MAGIC_SYSRQ
6922 static void normalize_task(struct rq *rq, struct task_struct *p)
6924 const struct sched_class *prev_class = p->sched_class;
6925 int old_prio = p->prio;
6930 deactivate_task(rq, p, 0);
6931 __setscheduler(rq, p, SCHED_NORMAL, 0);
6933 activate_task(rq, p, 0);
6934 resched_task(rq->curr);
6937 check_class_changed(rq, p, prev_class, old_prio);
6940 void normalize_rt_tasks(void)
6942 struct task_struct *g, *p;
6943 unsigned long flags;
6946 read_lock_irqsave(&tasklist_lock, flags);
6947 do_each_thread(g, p) {
6949 * Only normalize user tasks:
6954 p->se.exec_start = 0;
6955 #ifdef CONFIG_SCHEDSTATS
6956 p->se.statistics.wait_start = 0;
6957 p->se.statistics.sleep_start = 0;
6958 p->se.statistics.block_start = 0;
6963 * Renice negative nice level userspace
6966 if (TASK_NICE(p) < 0 && p->mm)
6967 set_user_nice(p, 0);
6971 raw_spin_lock(&p->pi_lock);
6972 rq = __task_rq_lock(p);
6974 normalize_task(rq, p);
6976 __task_rq_unlock(rq);
6977 raw_spin_unlock(&p->pi_lock);
6978 } while_each_thread(g, p);
6980 read_unlock_irqrestore(&tasklist_lock, flags);
6983 #endif /* CONFIG_MAGIC_SYSRQ */
6985 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6987 * These functions are only useful for the IA64 MCA handling, or kdb.
6989 * They can only be called when the whole system has been
6990 * stopped - every CPU needs to be quiescent, and no scheduling
6991 * activity can take place. Using them for anything else would
6992 * be a serious bug, and as a result, they aren't even visible
6993 * under any other configuration.
6997 * curr_task - return the current task for a given cpu.
6998 * @cpu: the processor in question.
7000 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7002 struct task_struct *curr_task(int cpu)
7004 return cpu_curr(cpu);
7007 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7011 * set_curr_task - set the current task for a given cpu.
7012 * @cpu: the processor in question.
7013 * @p: the task pointer to set.
7015 * Description: This function must only be used when non-maskable interrupts
7016 * are serviced on a separate stack. It allows the architecture to switch the
7017 * notion of the current task on a cpu in a non-blocking manner. This function
7018 * must be called with all CPU's synchronized, and interrupts disabled, the
7019 * and caller must save the original value of the current task (see
7020 * curr_task() above) and restore that value before reenabling interrupts and
7021 * re-starting the system.
7023 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7025 void set_curr_task(int cpu, struct task_struct *p)
7032 #ifdef CONFIG_RT_GROUP_SCHED
7033 #else /* !CONFIG_RT_GROUP_SCHED */
7034 #endif /* CONFIG_RT_GROUP_SCHED */
7036 #ifdef CONFIG_CGROUP_SCHED
7037 /* task_group_lock serializes the addition/removal of task groups */
7038 static DEFINE_SPINLOCK(task_group_lock);
7040 static void free_sched_group(struct task_group *tg)
7042 free_fair_sched_group(tg);
7043 free_rt_sched_group(tg);
7048 /* allocate runqueue etc for a new task group */
7049 struct task_group *sched_create_group(struct task_group *parent)
7051 struct task_group *tg;
7052 unsigned long flags;
7054 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7056 return ERR_PTR(-ENOMEM);
7058 if (!alloc_fair_sched_group(tg, parent))
7061 if (!alloc_rt_sched_group(tg, parent))
7064 spin_lock_irqsave(&task_group_lock, flags);
7065 list_add_rcu(&tg->list, &task_groups);
7067 WARN_ON(!parent); /* root should already exist */
7069 tg->parent = parent;
7070 INIT_LIST_HEAD(&tg->children);
7071 list_add_rcu(&tg->siblings, &parent->children);
7072 spin_unlock_irqrestore(&task_group_lock, flags);
7077 free_sched_group(tg);
7078 return ERR_PTR(-ENOMEM);
7081 /* rcu callback to free various structures associated with a task group */
7082 static void free_sched_group_rcu(struct rcu_head *rhp)
7084 /* now it should be safe to free those cfs_rqs */
7085 free_sched_group(container_of(rhp, struct task_group, rcu));
7088 /* Destroy runqueue etc associated with a task group */
7089 void sched_destroy_group(struct task_group *tg)
7091 unsigned long flags;
7094 /* end participation in shares distribution */
7095 for_each_possible_cpu(i)
7096 unregister_fair_sched_group(tg, i);
7098 spin_lock_irqsave(&task_group_lock, flags);
7099 list_del_rcu(&tg->list);
7100 list_del_rcu(&tg->siblings);
7101 spin_unlock_irqrestore(&task_group_lock, flags);
7103 /* wait for possible concurrent references to cfs_rqs complete */
7104 call_rcu(&tg->rcu, free_sched_group_rcu);
7107 /* change task's runqueue when it moves between groups.
7108 * The caller of this function should have put the task in its new group
7109 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7110 * reflect its new group.
7112 void sched_move_task(struct task_struct *tsk)
7115 unsigned long flags;
7118 rq = task_rq_lock(tsk, &flags);
7120 running = task_current(rq, tsk);
7124 dequeue_task(rq, tsk, 0);
7125 if (unlikely(running))
7126 tsk->sched_class->put_prev_task(rq, tsk);
7128 #ifdef CONFIG_FAIR_GROUP_SCHED
7129 if (tsk->sched_class->task_move_group)
7130 tsk->sched_class->task_move_group(tsk, on_rq);
7133 set_task_rq(tsk, task_cpu(tsk));
7135 if (unlikely(running))
7136 tsk->sched_class->set_curr_task(rq);
7138 enqueue_task(rq, tsk, 0);
7140 task_rq_unlock(rq, tsk, &flags);
7142 #endif /* CONFIG_CGROUP_SCHED */
7144 #ifdef CONFIG_FAIR_GROUP_SCHED
7147 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7148 static unsigned long to_ratio(u64 period, u64 runtime)
7150 if (runtime == RUNTIME_INF)
7153 return div64_u64(runtime << 20, period);
7157 #ifdef CONFIG_RT_GROUP_SCHED
7159 * Ensure that the real time constraints are schedulable.
7161 static DEFINE_MUTEX(rt_constraints_mutex);
7163 /* Must be called with tasklist_lock held */
7164 static inline int tg_has_rt_tasks(struct task_group *tg)
7166 struct task_struct *g, *p;
7168 do_each_thread(g, p) {
7169 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7171 } while_each_thread(g, p);
7176 struct rt_schedulable_data {
7177 struct task_group *tg;
7182 static int tg_rt_schedulable(struct task_group *tg, void *data)
7184 struct rt_schedulable_data *d = data;
7185 struct task_group *child;
7186 unsigned long total, sum = 0;
7187 u64 period, runtime;
7189 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7190 runtime = tg->rt_bandwidth.rt_runtime;
7193 period = d->rt_period;
7194 runtime = d->rt_runtime;
7198 * Cannot have more runtime than the period.
7200 if (runtime > period && runtime != RUNTIME_INF)
7204 * Ensure we don't starve existing RT tasks.
7206 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7209 total = to_ratio(period, runtime);
7212 * Nobody can have more than the global setting allows.
7214 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7218 * The sum of our children's runtime should not exceed our own.
7220 list_for_each_entry_rcu(child, &tg->children, siblings) {
7221 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7222 runtime = child->rt_bandwidth.rt_runtime;
7224 if (child == d->tg) {
7225 period = d->rt_period;
7226 runtime = d->rt_runtime;
7229 sum += to_ratio(period, runtime);
7238 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7242 struct rt_schedulable_data data = {
7244 .rt_period = period,
7245 .rt_runtime = runtime,
7249 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7255 static int tg_set_rt_bandwidth(struct task_group *tg,
7256 u64 rt_period, u64 rt_runtime)
7260 mutex_lock(&rt_constraints_mutex);
7261 read_lock(&tasklist_lock);
7262 err = __rt_schedulable(tg, rt_period, rt_runtime);
7266 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7267 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7268 tg->rt_bandwidth.rt_runtime = rt_runtime;
7270 for_each_possible_cpu(i) {
7271 struct rt_rq *rt_rq = tg->rt_rq[i];
7273 raw_spin_lock(&rt_rq->rt_runtime_lock);
7274 rt_rq->rt_runtime = rt_runtime;
7275 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7277 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7279 read_unlock(&tasklist_lock);
7280 mutex_unlock(&rt_constraints_mutex);
7285 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7287 u64 rt_runtime, rt_period;
7289 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7290 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7291 if (rt_runtime_us < 0)
7292 rt_runtime = RUNTIME_INF;
7294 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7297 long sched_group_rt_runtime(struct task_group *tg)
7301 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7304 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7305 do_div(rt_runtime_us, NSEC_PER_USEC);
7306 return rt_runtime_us;
7309 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7311 u64 rt_runtime, rt_period;
7313 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7314 rt_runtime = tg->rt_bandwidth.rt_runtime;
7319 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7322 long sched_group_rt_period(struct task_group *tg)
7326 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7327 do_div(rt_period_us, NSEC_PER_USEC);
7328 return rt_period_us;
7331 static int sched_rt_global_constraints(void)
7333 u64 runtime, period;
7336 if (sysctl_sched_rt_period <= 0)
7339 runtime = global_rt_runtime();
7340 period = global_rt_period();
7343 * Sanity check on the sysctl variables.
7345 if (runtime > period && runtime != RUNTIME_INF)
7348 mutex_lock(&rt_constraints_mutex);
7349 read_lock(&tasklist_lock);
7350 ret = __rt_schedulable(NULL, 0, 0);
7351 read_unlock(&tasklist_lock);
7352 mutex_unlock(&rt_constraints_mutex);
7357 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7359 /* Don't accept realtime tasks when there is no way for them to run */
7360 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7366 #else /* !CONFIG_RT_GROUP_SCHED */
7367 static int sched_rt_global_constraints(void)
7369 unsigned long flags;
7372 if (sysctl_sched_rt_period <= 0)
7376 * There's always some RT tasks in the root group
7377 * -- migration, kstopmachine etc..
7379 if (sysctl_sched_rt_runtime == 0)
7382 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7383 for_each_possible_cpu(i) {
7384 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7386 raw_spin_lock(&rt_rq->rt_runtime_lock);
7387 rt_rq->rt_runtime = global_rt_runtime();
7388 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7390 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7394 #endif /* CONFIG_RT_GROUP_SCHED */
7396 int sched_rt_handler(struct ctl_table *table, int write,
7397 void __user *buffer, size_t *lenp,
7401 int old_period, old_runtime;
7402 static DEFINE_MUTEX(mutex);
7405 old_period = sysctl_sched_rt_period;
7406 old_runtime = sysctl_sched_rt_runtime;
7408 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7410 if (!ret && write) {
7411 ret = sched_rt_global_constraints();
7413 sysctl_sched_rt_period = old_period;
7414 sysctl_sched_rt_runtime = old_runtime;
7416 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7417 def_rt_bandwidth.rt_period =
7418 ns_to_ktime(global_rt_period());
7421 mutex_unlock(&mutex);
7426 #ifdef CONFIG_CGROUP_SCHED
7428 /* return corresponding task_group object of a cgroup */
7429 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7431 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7432 struct task_group, css);
7435 static struct cgroup_subsys_state *
7436 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7438 struct task_group *tg, *parent;
7440 if (!cgrp->parent) {
7441 /* This is early initialization for the top cgroup */
7442 return &root_task_group.css;
7445 parent = cgroup_tg(cgrp->parent);
7446 tg = sched_create_group(parent);
7448 return ERR_PTR(-ENOMEM);
7454 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7456 struct task_group *tg = cgroup_tg(cgrp);
7458 sched_destroy_group(tg);
7462 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
7464 #ifdef CONFIG_RT_GROUP_SCHED
7465 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
7468 /* We don't support RT-tasks being in separate groups */
7469 if (tsk->sched_class != &fair_sched_class)
7476 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
7478 sched_move_task(tsk);
7482 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7483 struct cgroup *old_cgrp, struct task_struct *task)
7486 * cgroup_exit() is called in the copy_process() failure path.
7487 * Ignore this case since the task hasn't ran yet, this avoids
7488 * trying to poke a half freed task state from generic code.
7490 if (!(task->flags & PF_EXITING))
7493 sched_move_task(task);
7496 #ifdef CONFIG_FAIR_GROUP_SCHED
7497 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7500 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7503 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7505 struct task_group *tg = cgroup_tg(cgrp);
7507 return (u64) scale_load_down(tg->shares);
7510 #ifdef CONFIG_CFS_BANDWIDTH
7511 static DEFINE_MUTEX(cfs_constraints_mutex);
7513 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7514 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7516 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7518 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7520 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7521 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7523 if (tg == &root_task_group)
7527 * Ensure we have at some amount of bandwidth every period. This is
7528 * to prevent reaching a state of large arrears when throttled via
7529 * entity_tick() resulting in prolonged exit starvation.
7531 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7535 * Likewise, bound things on the otherside by preventing insane quota
7536 * periods. This also allows us to normalize in computing quota
7539 if (period > max_cfs_quota_period)
7542 mutex_lock(&cfs_constraints_mutex);
7543 ret = __cfs_schedulable(tg, period, quota);
7547 runtime_enabled = quota != RUNTIME_INF;
7548 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7549 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7550 raw_spin_lock_irq(&cfs_b->lock);
7551 cfs_b->period = ns_to_ktime(period);
7552 cfs_b->quota = quota;
7554 __refill_cfs_bandwidth_runtime(cfs_b);
7555 /* restart the period timer (if active) to handle new period expiry */
7556 if (runtime_enabled && cfs_b->timer_active) {
7557 /* force a reprogram */
7558 cfs_b->timer_active = 0;
7559 __start_cfs_bandwidth(cfs_b);
7561 raw_spin_unlock_irq(&cfs_b->lock);
7563 for_each_possible_cpu(i) {
7564 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7565 struct rq *rq = cfs_rq->rq;
7567 raw_spin_lock_irq(&rq->lock);
7568 cfs_rq->runtime_enabled = runtime_enabled;
7569 cfs_rq->runtime_remaining = 0;
7571 if (cfs_rq->throttled)
7572 unthrottle_cfs_rq(cfs_rq);
7573 raw_spin_unlock_irq(&rq->lock);
7576 mutex_unlock(&cfs_constraints_mutex);
7581 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7585 period = ktime_to_ns(tg->cfs_bandwidth.period);
7586 if (cfs_quota_us < 0)
7587 quota = RUNTIME_INF;
7589 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7591 return tg_set_cfs_bandwidth(tg, period, quota);
7594 long tg_get_cfs_quota(struct task_group *tg)
7598 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7601 quota_us = tg->cfs_bandwidth.quota;
7602 do_div(quota_us, NSEC_PER_USEC);
7607 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7611 period = (u64)cfs_period_us * NSEC_PER_USEC;
7612 quota = tg->cfs_bandwidth.quota;
7617 return tg_set_cfs_bandwidth(tg, period, quota);
7620 long tg_get_cfs_period(struct task_group *tg)
7624 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7625 do_div(cfs_period_us, NSEC_PER_USEC);
7627 return cfs_period_us;
7630 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7632 return tg_get_cfs_quota(cgroup_tg(cgrp));
7635 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7638 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7641 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7643 return tg_get_cfs_period(cgroup_tg(cgrp));
7646 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7649 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7652 struct cfs_schedulable_data {
7653 struct task_group *tg;
7658 * normalize group quota/period to be quota/max_period
7659 * note: units are usecs
7661 static u64 normalize_cfs_quota(struct task_group *tg,
7662 struct cfs_schedulable_data *d)
7670 period = tg_get_cfs_period(tg);
7671 quota = tg_get_cfs_quota(tg);
7674 /* note: these should typically be equivalent */
7675 if (quota == RUNTIME_INF || quota == -1)
7678 return to_ratio(period, quota);
7681 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7683 struct cfs_schedulable_data *d = data;
7684 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7685 s64 quota = 0, parent_quota = -1;
7688 quota = RUNTIME_INF;
7690 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7692 quota = normalize_cfs_quota(tg, d);
7693 parent_quota = parent_b->hierarchal_quota;
7696 * ensure max(child_quota) <= parent_quota, inherit when no
7699 if (quota == RUNTIME_INF)
7700 quota = parent_quota;
7701 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7704 cfs_b->hierarchal_quota = quota;
7709 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7712 struct cfs_schedulable_data data = {
7718 if (quota != RUNTIME_INF) {
7719 do_div(data.period, NSEC_PER_USEC);
7720 do_div(data.quota, NSEC_PER_USEC);
7724 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7730 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7731 struct cgroup_map_cb *cb)
7733 struct task_group *tg = cgroup_tg(cgrp);
7734 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7736 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7737 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7738 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7742 #endif /* CONFIG_CFS_BANDWIDTH */
7743 #endif /* CONFIG_FAIR_GROUP_SCHED */
7745 #ifdef CONFIG_RT_GROUP_SCHED
7746 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7749 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7752 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7754 return sched_group_rt_runtime(cgroup_tg(cgrp));
7757 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7760 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7763 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7765 return sched_group_rt_period(cgroup_tg(cgrp));
7767 #endif /* CONFIG_RT_GROUP_SCHED */
7769 static struct cftype cpu_files[] = {
7770 #ifdef CONFIG_FAIR_GROUP_SCHED
7773 .read_u64 = cpu_shares_read_u64,
7774 .write_u64 = cpu_shares_write_u64,
7777 #ifdef CONFIG_CFS_BANDWIDTH
7779 .name = "cfs_quota_us",
7780 .read_s64 = cpu_cfs_quota_read_s64,
7781 .write_s64 = cpu_cfs_quota_write_s64,
7784 .name = "cfs_period_us",
7785 .read_u64 = cpu_cfs_period_read_u64,
7786 .write_u64 = cpu_cfs_period_write_u64,
7790 .read_map = cpu_stats_show,
7793 #ifdef CONFIG_RT_GROUP_SCHED
7795 .name = "rt_runtime_us",
7796 .read_s64 = cpu_rt_runtime_read,
7797 .write_s64 = cpu_rt_runtime_write,
7800 .name = "rt_period_us",
7801 .read_u64 = cpu_rt_period_read_uint,
7802 .write_u64 = cpu_rt_period_write_uint,
7807 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7809 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7812 struct cgroup_subsys cpu_cgroup_subsys = {
7814 .create = cpu_cgroup_create,
7815 .destroy = cpu_cgroup_destroy,
7816 .can_attach_task = cpu_cgroup_can_attach_task,
7817 .attach_task = cpu_cgroup_attach_task,
7818 .exit = cpu_cgroup_exit,
7819 .populate = cpu_cgroup_populate,
7820 .subsys_id = cpu_cgroup_subsys_id,
7824 #endif /* CONFIG_CGROUP_SCHED */
7826 #ifdef CONFIG_CGROUP_CPUACCT
7829 * CPU accounting code for task groups.
7831 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7832 * (balbir@in.ibm.com).
7835 /* track cpu usage of a group of tasks and its child groups */
7837 struct cgroup_subsys_state css;
7838 /* cpuusage holds pointer to a u64-type object on every cpu */
7839 u64 __percpu *cpuusage;
7840 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
7841 struct cpuacct *parent;
7844 struct cgroup_subsys cpuacct_subsys;
7846 /* return cpu accounting group corresponding to this container */
7847 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
7849 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
7850 struct cpuacct, css);
7853 /* return cpu accounting group to which this task belongs */
7854 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7856 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7857 struct cpuacct, css);
7860 /* create a new cpu accounting group */
7861 static struct cgroup_subsys_state *cpuacct_create(
7862 struct cgroup_subsys *ss, struct cgroup *cgrp)
7864 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7870 ca->cpuusage = alloc_percpu(u64);
7874 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
7875 if (percpu_counter_init(&ca->cpustat[i], 0))
7876 goto out_free_counters;
7879 ca->parent = cgroup_ca(cgrp->parent);
7885 percpu_counter_destroy(&ca->cpustat[i]);
7886 free_percpu(ca->cpuusage);
7890 return ERR_PTR(-ENOMEM);
7893 /* destroy an existing cpu accounting group */
7895 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7897 struct cpuacct *ca = cgroup_ca(cgrp);
7900 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
7901 percpu_counter_destroy(&ca->cpustat[i]);
7902 free_percpu(ca->cpuusage);
7906 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7908 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7911 #ifndef CONFIG_64BIT
7913 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7915 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7917 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7925 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7927 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7929 #ifndef CONFIG_64BIT
7931 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7933 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7935 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7941 /* return total cpu usage (in nanoseconds) of a group */
7942 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7944 struct cpuacct *ca = cgroup_ca(cgrp);
7945 u64 totalcpuusage = 0;
7948 for_each_present_cpu(i)
7949 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7951 return totalcpuusage;
7954 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7957 struct cpuacct *ca = cgroup_ca(cgrp);
7966 for_each_present_cpu(i)
7967 cpuacct_cpuusage_write(ca, i, 0);
7973 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7976 struct cpuacct *ca = cgroup_ca(cgroup);
7980 for_each_present_cpu(i) {
7981 percpu = cpuacct_cpuusage_read(ca, i);
7982 seq_printf(m, "%llu ", (unsigned long long) percpu);
7984 seq_printf(m, "\n");
7988 static const char *cpuacct_stat_desc[] = {
7989 [CPUACCT_STAT_USER] = "user",
7990 [CPUACCT_STAT_SYSTEM] = "system",
7993 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
7994 struct cgroup_map_cb *cb)
7996 struct cpuacct *ca = cgroup_ca(cgrp);
7999 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8000 s64 val = percpu_counter_read(&ca->cpustat[i]);
8001 val = cputime64_to_clock_t(val);
8002 cb->fill(cb, cpuacct_stat_desc[i], val);
8007 static struct cftype files[] = {
8010 .read_u64 = cpuusage_read,
8011 .write_u64 = cpuusage_write,
8014 .name = "usage_percpu",
8015 .read_seq_string = cpuacct_percpu_seq_read,
8019 .read_map = cpuacct_stats_show,
8023 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8025 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8029 * charge this task's execution time to its accounting group.
8031 * called with rq->lock held.
8033 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8038 if (unlikely(!cpuacct_subsys.active))
8041 cpu = task_cpu(tsk);
8047 for (; ca; ca = ca->parent) {
8048 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8049 *cpuusage += cputime;
8056 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8057 * in cputime_t units. As a result, cpuacct_update_stats calls
8058 * percpu_counter_add with values large enough to always overflow the
8059 * per cpu batch limit causing bad SMP scalability.
8061 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8062 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8063 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8066 #define CPUACCT_BATCH \
8067 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8069 #define CPUACCT_BATCH 0
8073 * Charge the system/user time to the task's accounting group.
8075 void cpuacct_update_stats(struct task_struct *tsk,
8076 enum cpuacct_stat_index idx, cputime_t val)
8079 int batch = CPUACCT_BATCH;
8081 if (unlikely(!cpuacct_subsys.active))
8088 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8094 struct cgroup_subsys cpuacct_subsys = {
8096 .create = cpuacct_create,
8097 .destroy = cpuacct_destroy,
8098 .populate = cpuacct_populate,
8099 .subsys_id = cpuacct_subsys_id,
8101 #endif /* CONFIG_CGROUP_CPUACCT */