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 "features.h"
137 #ifdef CONFIG_SCHED_DEBUG
138 #define SCHED_FEAT(name, enabled) \
141 static __read_mostly char *sched_feat_names[] = {
142 #include "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 int cpu = smp_processor_id();
579 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
582 #else /* CONFIG_NO_HZ */
584 static inline bool got_nohz_idle_kick(void)
589 #endif /* CONFIG_NO_HZ */
591 void sched_avg_update(struct rq *rq)
593 s64 period = sched_avg_period();
595 while ((s64)(rq->clock - rq->age_stamp) > period) {
597 * Inline assembly required to prevent the compiler
598 * optimising this loop into a divmod call.
599 * See __iter_div_u64_rem() for another example of this.
601 asm("" : "+rm" (rq->age_stamp));
602 rq->age_stamp += period;
607 #else /* !CONFIG_SMP */
608 void resched_task(struct task_struct *p)
610 assert_raw_spin_locked(&task_rq(p)->lock);
611 set_tsk_need_resched(p);
613 #endif /* CONFIG_SMP */
615 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
616 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
618 * Iterate task_group tree rooted at *from, calling @down when first entering a
619 * node and @up when leaving it for the final time.
621 * Caller must hold rcu_lock or sufficient equivalent.
623 int walk_tg_tree_from(struct task_group *from,
624 tg_visitor down, tg_visitor up, void *data)
626 struct task_group *parent, *child;
632 ret = (*down)(parent, data);
635 list_for_each_entry_rcu(child, &parent->children, siblings) {
642 ret = (*up)(parent, data);
643 if (ret || parent == from)
647 parent = parent->parent;
654 int tg_nop(struct task_group *tg, void *data)
660 void update_cpu_load(struct rq *this_rq);
662 static void set_load_weight(struct task_struct *p)
664 int prio = p->static_prio - MAX_RT_PRIO;
665 struct load_weight *load = &p->se.load;
668 * SCHED_IDLE tasks get minimal weight:
670 if (p->policy == SCHED_IDLE) {
671 load->weight = scale_load(WEIGHT_IDLEPRIO);
672 load->inv_weight = WMULT_IDLEPRIO;
676 load->weight = scale_load(prio_to_weight[prio]);
677 load->inv_weight = prio_to_wmult[prio];
680 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
683 sched_info_queued(p);
684 p->sched_class->enqueue_task(rq, p, flags);
687 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
690 sched_info_dequeued(p);
691 p->sched_class->dequeue_task(rq, p, flags);
695 * activate_task - move a task to the runqueue.
697 void activate_task(struct rq *rq, struct task_struct *p, int flags)
699 if (task_contributes_to_load(p))
700 rq->nr_uninterruptible--;
702 enqueue_task(rq, p, flags);
706 * deactivate_task - remove a task from the runqueue.
708 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
710 if (task_contributes_to_load(p))
711 rq->nr_uninterruptible++;
713 dequeue_task(rq, p, flags);
716 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
719 * There are no locks covering percpu hardirq/softirq time.
720 * They are only modified in account_system_vtime, on corresponding CPU
721 * with interrupts disabled. So, writes are safe.
722 * They are read and saved off onto struct rq in update_rq_clock().
723 * This may result in other CPU reading this CPU's irq time and can
724 * race with irq/account_system_vtime on this CPU. We would either get old
725 * or new value with a side effect of accounting a slice of irq time to wrong
726 * task when irq is in progress while we read rq->clock. That is a worthy
727 * compromise in place of having locks on each irq in account_system_time.
729 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
730 static DEFINE_PER_CPU(u64, cpu_softirq_time);
732 static DEFINE_PER_CPU(u64, irq_start_time);
733 static int sched_clock_irqtime;
735 void enable_sched_clock_irqtime(void)
737 sched_clock_irqtime = 1;
740 void disable_sched_clock_irqtime(void)
742 sched_clock_irqtime = 0;
746 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
748 static inline void irq_time_write_begin(void)
750 __this_cpu_inc(irq_time_seq.sequence);
754 static inline void irq_time_write_end(void)
757 __this_cpu_inc(irq_time_seq.sequence);
760 static inline u64 irq_time_read(int cpu)
766 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
767 irq_time = per_cpu(cpu_softirq_time, cpu) +
768 per_cpu(cpu_hardirq_time, cpu);
769 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
773 #else /* CONFIG_64BIT */
774 static inline void irq_time_write_begin(void)
778 static inline void irq_time_write_end(void)
782 static inline u64 irq_time_read(int cpu)
784 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
786 #endif /* CONFIG_64BIT */
789 * Called before incrementing preempt_count on {soft,}irq_enter
790 * and before decrementing preempt_count on {soft,}irq_exit.
792 void account_system_vtime(struct task_struct *curr)
798 if (!sched_clock_irqtime)
801 local_irq_save(flags);
803 cpu = smp_processor_id();
804 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
805 __this_cpu_add(irq_start_time, delta);
807 irq_time_write_begin();
809 * We do not account for softirq time from ksoftirqd here.
810 * We want to continue accounting softirq time to ksoftirqd thread
811 * in that case, so as not to confuse scheduler with a special task
812 * that do not consume any time, but still wants to run.
815 __this_cpu_add(cpu_hardirq_time, delta);
816 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
817 __this_cpu_add(cpu_softirq_time, delta);
819 irq_time_write_end();
820 local_irq_restore(flags);
822 EXPORT_SYMBOL_GPL(account_system_vtime);
824 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
826 #ifdef CONFIG_PARAVIRT
827 static inline u64 steal_ticks(u64 steal)
829 if (unlikely(steal > NSEC_PER_SEC))
830 return div_u64(steal, TICK_NSEC);
832 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
836 static void update_rq_clock_task(struct rq *rq, s64 delta)
839 * In theory, the compile should just see 0 here, and optimize out the call
840 * to sched_rt_avg_update. But I don't trust it...
842 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
843 s64 steal = 0, irq_delta = 0;
845 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
846 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
849 * Since irq_time is only updated on {soft,}irq_exit, we might run into
850 * this case when a previous update_rq_clock() happened inside a
853 * When this happens, we stop ->clock_task and only update the
854 * prev_irq_time stamp to account for the part that fit, so that a next
855 * update will consume the rest. This ensures ->clock_task is
858 * It does however cause some slight miss-attribution of {soft,}irq
859 * time, a more accurate solution would be to update the irq_time using
860 * the current rq->clock timestamp, except that would require using
863 if (irq_delta > delta)
866 rq->prev_irq_time += irq_delta;
869 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
870 if (static_branch((¶virt_steal_rq_enabled))) {
873 steal = paravirt_steal_clock(cpu_of(rq));
874 steal -= rq->prev_steal_time_rq;
876 if (unlikely(steal > delta))
879 st = steal_ticks(steal);
880 steal = st * TICK_NSEC;
882 rq->prev_steal_time_rq += steal;
888 rq->clock_task += delta;
890 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
891 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
892 sched_rt_avg_update(rq, irq_delta + steal);
896 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
897 static int irqtime_account_hi_update(void)
899 u64 *cpustat = kcpustat_this_cpu->cpustat;
904 local_irq_save(flags);
905 latest_ns = this_cpu_read(cpu_hardirq_time);
906 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat[CPUTIME_IRQ]))
908 local_irq_restore(flags);
912 static int irqtime_account_si_update(void)
914 u64 *cpustat = kcpustat_this_cpu->cpustat;
919 local_irq_save(flags);
920 latest_ns = this_cpu_read(cpu_softirq_time);
921 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat[CPUTIME_SOFTIRQ]))
923 local_irq_restore(flags);
927 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
929 #define sched_clock_irqtime (0)
933 void sched_set_stop_task(int cpu, struct task_struct *stop)
935 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
936 struct task_struct *old_stop = cpu_rq(cpu)->stop;
940 * Make it appear like a SCHED_FIFO task, its something
941 * userspace knows about and won't get confused about.
943 * Also, it will make PI more or less work without too
944 * much confusion -- but then, stop work should not
945 * rely on PI working anyway.
947 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
949 stop->sched_class = &stop_sched_class;
952 cpu_rq(cpu)->stop = stop;
956 * Reset it back to a normal scheduling class so that
957 * it can die in pieces.
959 old_stop->sched_class = &rt_sched_class;
964 * __normal_prio - return the priority that is based on the static prio
966 static inline int __normal_prio(struct task_struct *p)
968 return p->static_prio;
972 * Calculate the expected normal priority: i.e. priority
973 * without taking RT-inheritance into account. Might be
974 * boosted by interactivity modifiers. Changes upon fork,
975 * setprio syscalls, and whenever the interactivity
976 * estimator recalculates.
978 static inline int normal_prio(struct task_struct *p)
982 if (task_has_rt_policy(p))
983 prio = MAX_RT_PRIO-1 - p->rt_priority;
985 prio = __normal_prio(p);
990 * Calculate the current priority, i.e. the priority
991 * taken into account by the scheduler. This value might
992 * be boosted by RT tasks, or might be boosted by
993 * interactivity modifiers. Will be RT if the task got
994 * RT-boosted. If not then it returns p->normal_prio.
996 static int effective_prio(struct task_struct *p)
998 p->normal_prio = normal_prio(p);
1000 * If we are RT tasks or we were boosted to RT priority,
1001 * keep the priority unchanged. Otherwise, update priority
1002 * to the normal priority:
1004 if (!rt_prio(p->prio))
1005 return p->normal_prio;
1010 * task_curr - is this task currently executing on a CPU?
1011 * @p: the task in question.
1013 inline int task_curr(const struct task_struct *p)
1015 return cpu_curr(task_cpu(p)) == p;
1018 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1019 const struct sched_class *prev_class,
1022 if (prev_class != p->sched_class) {
1023 if (prev_class->switched_from)
1024 prev_class->switched_from(rq, p);
1025 p->sched_class->switched_to(rq, p);
1026 } else if (oldprio != p->prio)
1027 p->sched_class->prio_changed(rq, p, oldprio);
1030 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1032 const struct sched_class *class;
1034 if (p->sched_class == rq->curr->sched_class) {
1035 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1037 for_each_class(class) {
1038 if (class == rq->curr->sched_class)
1040 if (class == p->sched_class) {
1041 resched_task(rq->curr);
1048 * A queue event has occurred, and we're going to schedule. In
1049 * this case, we can save a useless back to back clock update.
1051 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1052 rq->skip_clock_update = 1;
1056 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1058 #ifdef CONFIG_SCHED_DEBUG
1060 * We should never call set_task_cpu() on a blocked task,
1061 * ttwu() will sort out the placement.
1063 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1064 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1066 #ifdef CONFIG_LOCKDEP
1068 * The caller should hold either p->pi_lock or rq->lock, when changing
1069 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1071 * sched_move_task() holds both and thus holding either pins the cgroup,
1072 * see set_task_rq().
1074 * Furthermore, all task_rq users should acquire both locks, see
1077 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1078 lockdep_is_held(&task_rq(p)->lock)));
1082 trace_sched_migrate_task(p, new_cpu);
1084 if (task_cpu(p) != new_cpu) {
1085 p->se.nr_migrations++;
1086 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1089 __set_task_cpu(p, new_cpu);
1092 struct migration_arg {
1093 struct task_struct *task;
1097 static int migration_cpu_stop(void *data);
1100 * wait_task_inactive - wait for a thread to unschedule.
1102 * If @match_state is nonzero, it's the @p->state value just checked and
1103 * not expected to change. If it changes, i.e. @p might have woken up,
1104 * then return zero. When we succeed in waiting for @p to be off its CPU,
1105 * we return a positive number (its total switch count). If a second call
1106 * a short while later returns the same number, the caller can be sure that
1107 * @p has remained unscheduled the whole time.
1109 * The caller must ensure that the task *will* unschedule sometime soon,
1110 * else this function might spin for a *long* time. This function can't
1111 * be called with interrupts off, or it may introduce deadlock with
1112 * smp_call_function() if an IPI is sent by the same process we are
1113 * waiting to become inactive.
1115 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1117 unsigned long flags;
1124 * We do the initial early heuristics without holding
1125 * any task-queue locks at all. We'll only try to get
1126 * the runqueue lock when things look like they will
1132 * If the task is actively running on another CPU
1133 * still, just relax and busy-wait without holding
1136 * NOTE! Since we don't hold any locks, it's not
1137 * even sure that "rq" stays as the right runqueue!
1138 * But we don't care, since "task_running()" will
1139 * return false if the runqueue has changed and p
1140 * is actually now running somewhere else!
1142 while (task_running(rq, p)) {
1143 if (match_state && unlikely(p->state != match_state))
1149 * Ok, time to look more closely! We need the rq
1150 * lock now, to be *sure*. If we're wrong, we'll
1151 * just go back and repeat.
1153 rq = task_rq_lock(p, &flags);
1154 trace_sched_wait_task(p);
1155 running = task_running(rq, p);
1158 if (!match_state || p->state == match_state)
1159 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1160 task_rq_unlock(rq, p, &flags);
1163 * If it changed from the expected state, bail out now.
1165 if (unlikely(!ncsw))
1169 * Was it really running after all now that we
1170 * checked with the proper locks actually held?
1172 * Oops. Go back and try again..
1174 if (unlikely(running)) {
1180 * It's not enough that it's not actively running,
1181 * it must be off the runqueue _entirely_, and not
1184 * So if it was still runnable (but just not actively
1185 * running right now), it's preempted, and we should
1186 * yield - it could be a while.
1188 if (unlikely(on_rq)) {
1189 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1191 set_current_state(TASK_UNINTERRUPTIBLE);
1192 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1197 * Ahh, all good. It wasn't running, and it wasn't
1198 * runnable, which means that it will never become
1199 * running in the future either. We're all done!
1208 * kick_process - kick a running thread to enter/exit the kernel
1209 * @p: the to-be-kicked thread
1211 * Cause a process which is running on another CPU to enter
1212 * kernel-mode, without any delay. (to get signals handled.)
1214 * NOTE: this function doesn't have to take the runqueue lock,
1215 * because all it wants to ensure is that the remote task enters
1216 * the kernel. If the IPI races and the task has been migrated
1217 * to another CPU then no harm is done and the purpose has been
1220 void kick_process(struct task_struct *p)
1226 if ((cpu != smp_processor_id()) && task_curr(p))
1227 smp_send_reschedule(cpu);
1230 EXPORT_SYMBOL_GPL(kick_process);
1231 #endif /* CONFIG_SMP */
1235 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1237 static int select_fallback_rq(int cpu, struct task_struct *p)
1240 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1242 /* Look for allowed, online CPU in same node. */
1243 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1244 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1247 /* Any allowed, online CPU? */
1248 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1249 if (dest_cpu < nr_cpu_ids)
1252 /* No more Mr. Nice Guy. */
1253 dest_cpu = cpuset_cpus_allowed_fallback(p);
1255 * Don't tell them about moving exiting tasks or
1256 * kernel threads (both mm NULL), since they never
1259 if (p->mm && printk_ratelimit()) {
1260 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1261 task_pid_nr(p), p->comm, cpu);
1268 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1271 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1273 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1276 * In order not to call set_task_cpu() on a blocking task we need
1277 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1280 * Since this is common to all placement strategies, this lives here.
1282 * [ this allows ->select_task() to simply return task_cpu(p) and
1283 * not worry about this generic constraint ]
1285 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1287 cpu = select_fallback_rq(task_cpu(p), p);
1292 static void update_avg(u64 *avg, u64 sample)
1294 s64 diff = sample - *avg;
1300 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1302 #ifdef CONFIG_SCHEDSTATS
1303 struct rq *rq = this_rq();
1306 int this_cpu = smp_processor_id();
1308 if (cpu == this_cpu) {
1309 schedstat_inc(rq, ttwu_local);
1310 schedstat_inc(p, se.statistics.nr_wakeups_local);
1312 struct sched_domain *sd;
1314 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1316 for_each_domain(this_cpu, sd) {
1317 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1318 schedstat_inc(sd, ttwu_wake_remote);
1325 if (wake_flags & WF_MIGRATED)
1326 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1328 #endif /* CONFIG_SMP */
1330 schedstat_inc(rq, ttwu_count);
1331 schedstat_inc(p, se.statistics.nr_wakeups);
1333 if (wake_flags & WF_SYNC)
1334 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1336 #endif /* CONFIG_SCHEDSTATS */
1339 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1341 activate_task(rq, p, en_flags);
1344 /* if a worker is waking up, notify workqueue */
1345 if (p->flags & PF_WQ_WORKER)
1346 wq_worker_waking_up(p, cpu_of(rq));
1350 * Mark the task runnable and perform wakeup-preemption.
1353 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1355 trace_sched_wakeup(p, true);
1356 check_preempt_curr(rq, p, wake_flags);
1358 p->state = TASK_RUNNING;
1360 if (p->sched_class->task_woken)
1361 p->sched_class->task_woken(rq, p);
1363 if (rq->idle_stamp) {
1364 u64 delta = rq->clock - rq->idle_stamp;
1365 u64 max = 2*sysctl_sched_migration_cost;
1370 update_avg(&rq->avg_idle, delta);
1377 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1380 if (p->sched_contributes_to_load)
1381 rq->nr_uninterruptible--;
1384 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1385 ttwu_do_wakeup(rq, p, wake_flags);
1389 * Called in case the task @p isn't fully descheduled from its runqueue,
1390 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1391 * since all we need to do is flip p->state to TASK_RUNNING, since
1392 * the task is still ->on_rq.
1394 static int ttwu_remote(struct task_struct *p, int wake_flags)
1399 rq = __task_rq_lock(p);
1401 ttwu_do_wakeup(rq, p, wake_flags);
1404 __task_rq_unlock(rq);
1410 static void sched_ttwu_pending(void)
1412 struct rq *rq = this_rq();
1413 struct llist_node *llist = llist_del_all(&rq->wake_list);
1414 struct task_struct *p;
1416 raw_spin_lock(&rq->lock);
1419 p = llist_entry(llist, struct task_struct, wake_entry);
1420 llist = llist_next(llist);
1421 ttwu_do_activate(rq, p, 0);
1424 raw_spin_unlock(&rq->lock);
1427 void scheduler_ipi(void)
1429 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1433 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1434 * traditionally all their work was done from the interrupt return
1435 * path. Now that we actually do some work, we need to make sure
1438 * Some archs already do call them, luckily irq_enter/exit nest
1441 * Arguably we should visit all archs and update all handlers,
1442 * however a fair share of IPIs are still resched only so this would
1443 * somewhat pessimize the simple resched case.
1446 sched_ttwu_pending();
1449 * Check if someone kicked us for doing the nohz idle load balance.
1451 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1452 this_rq()->idle_balance = 1;
1453 raise_softirq_irqoff(SCHED_SOFTIRQ);
1458 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1460 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1461 smp_send_reschedule(cpu);
1464 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1465 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1470 rq = __task_rq_lock(p);
1472 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1473 ttwu_do_wakeup(rq, p, wake_flags);
1476 __task_rq_unlock(rq);
1481 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1482 #endif /* CONFIG_SMP */
1484 static void ttwu_queue(struct task_struct *p, int cpu)
1486 struct rq *rq = cpu_rq(cpu);
1488 #if defined(CONFIG_SMP)
1489 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
1490 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1491 ttwu_queue_remote(p, cpu);
1496 raw_spin_lock(&rq->lock);
1497 ttwu_do_activate(rq, p, 0);
1498 raw_spin_unlock(&rq->lock);
1502 * try_to_wake_up - wake up a thread
1503 * @p: the thread to be awakened
1504 * @state: the mask of task states that can be woken
1505 * @wake_flags: wake modifier flags (WF_*)
1507 * Put it on the run-queue if it's not already there. The "current"
1508 * thread is always on the run-queue (except when the actual
1509 * re-schedule is in progress), and as such you're allowed to do
1510 * the simpler "current->state = TASK_RUNNING" to mark yourself
1511 * runnable without the overhead of this.
1513 * Returns %true if @p was woken up, %false if it was already running
1514 * or @state didn't match @p's state.
1517 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1519 unsigned long flags;
1520 int cpu, success = 0;
1523 raw_spin_lock_irqsave(&p->pi_lock, flags);
1524 if (!(p->state & state))
1527 success = 1; /* we're going to change ->state */
1530 if (p->on_rq && ttwu_remote(p, wake_flags))
1535 * If the owning (remote) cpu is still in the middle of schedule() with
1536 * this task as prev, wait until its done referencing the task.
1539 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1541 * In case the architecture enables interrupts in
1542 * context_switch(), we cannot busy wait, since that
1543 * would lead to deadlocks when an interrupt hits and
1544 * tries to wake up @prev. So bail and do a complete
1547 if (ttwu_activate_remote(p, wake_flags))
1554 * Pairs with the smp_wmb() in finish_lock_switch().
1558 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1559 p->state = TASK_WAKING;
1561 if (p->sched_class->task_waking)
1562 p->sched_class->task_waking(p);
1564 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1565 if (task_cpu(p) != cpu) {
1566 wake_flags |= WF_MIGRATED;
1567 set_task_cpu(p, cpu);
1569 #endif /* CONFIG_SMP */
1573 ttwu_stat(p, cpu, wake_flags);
1575 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1581 * try_to_wake_up_local - try to wake up a local task with rq lock held
1582 * @p: the thread to be awakened
1584 * Put @p on the run-queue if it's not already there. The caller must
1585 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1588 static void try_to_wake_up_local(struct task_struct *p)
1590 struct rq *rq = task_rq(p);
1592 BUG_ON(rq != this_rq());
1593 BUG_ON(p == current);
1594 lockdep_assert_held(&rq->lock);
1596 if (!raw_spin_trylock(&p->pi_lock)) {
1597 raw_spin_unlock(&rq->lock);
1598 raw_spin_lock(&p->pi_lock);
1599 raw_spin_lock(&rq->lock);
1602 if (!(p->state & TASK_NORMAL))
1606 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1608 ttwu_do_wakeup(rq, p, 0);
1609 ttwu_stat(p, smp_processor_id(), 0);
1611 raw_spin_unlock(&p->pi_lock);
1615 * wake_up_process - Wake up a specific process
1616 * @p: The process to be woken up.
1618 * Attempt to wake up the nominated process and move it to the set of runnable
1619 * processes. Returns 1 if the process was woken up, 0 if it was already
1622 * It may be assumed that this function implies a write memory barrier before
1623 * changing the task state if and only if any tasks are woken up.
1625 int wake_up_process(struct task_struct *p)
1627 return try_to_wake_up(p, TASK_ALL, 0);
1629 EXPORT_SYMBOL(wake_up_process);
1631 int wake_up_state(struct task_struct *p, unsigned int state)
1633 return try_to_wake_up(p, state, 0);
1637 * Perform scheduler related setup for a newly forked process p.
1638 * p is forked by current.
1640 * __sched_fork() is basic setup used by init_idle() too:
1642 static void __sched_fork(struct task_struct *p)
1647 p->se.exec_start = 0;
1648 p->se.sum_exec_runtime = 0;
1649 p->se.prev_sum_exec_runtime = 0;
1650 p->se.nr_migrations = 0;
1652 INIT_LIST_HEAD(&p->se.group_node);
1654 #ifdef CONFIG_SCHEDSTATS
1655 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1658 INIT_LIST_HEAD(&p->rt.run_list);
1660 #ifdef CONFIG_PREEMPT_NOTIFIERS
1661 INIT_HLIST_HEAD(&p->preempt_notifiers);
1666 * fork()/clone()-time setup:
1668 void sched_fork(struct task_struct *p)
1670 unsigned long flags;
1671 int cpu = get_cpu();
1675 * We mark the process as running here. This guarantees that
1676 * nobody will actually run it, and a signal or other external
1677 * event cannot wake it up and insert it on the runqueue either.
1679 p->state = TASK_RUNNING;
1682 * Make sure we do not leak PI boosting priority to the child.
1684 p->prio = current->normal_prio;
1687 * Revert to default priority/policy on fork if requested.
1689 if (unlikely(p->sched_reset_on_fork)) {
1690 if (task_has_rt_policy(p)) {
1691 p->policy = SCHED_NORMAL;
1692 p->static_prio = NICE_TO_PRIO(0);
1694 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1695 p->static_prio = NICE_TO_PRIO(0);
1697 p->prio = p->normal_prio = __normal_prio(p);
1701 * We don't need the reset flag anymore after the fork. It has
1702 * fulfilled its duty:
1704 p->sched_reset_on_fork = 0;
1707 if (!rt_prio(p->prio))
1708 p->sched_class = &fair_sched_class;
1710 if (p->sched_class->task_fork)
1711 p->sched_class->task_fork(p);
1714 * The child is not yet in the pid-hash so no cgroup attach races,
1715 * and the cgroup is pinned to this child due to cgroup_fork()
1716 * is ran before sched_fork().
1718 * Silence PROVE_RCU.
1720 raw_spin_lock_irqsave(&p->pi_lock, flags);
1721 set_task_cpu(p, cpu);
1722 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1724 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1725 if (likely(sched_info_on()))
1726 memset(&p->sched_info, 0, sizeof(p->sched_info));
1728 #if defined(CONFIG_SMP)
1731 #ifdef CONFIG_PREEMPT_COUNT
1732 /* Want to start with kernel preemption disabled. */
1733 task_thread_info(p)->preempt_count = 1;
1736 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1743 * wake_up_new_task - wake up a newly created task for the first time.
1745 * This function will do some initial scheduler statistics housekeeping
1746 * that must be done for every newly created context, then puts the task
1747 * on the runqueue and wakes it.
1749 void wake_up_new_task(struct task_struct *p)
1751 unsigned long flags;
1754 raw_spin_lock_irqsave(&p->pi_lock, flags);
1757 * Fork balancing, do it here and not earlier because:
1758 * - cpus_allowed can change in the fork path
1759 * - any previously selected cpu might disappear through hotplug
1761 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1764 rq = __task_rq_lock(p);
1765 activate_task(rq, p, 0);
1767 trace_sched_wakeup_new(p, true);
1768 check_preempt_curr(rq, p, WF_FORK);
1770 if (p->sched_class->task_woken)
1771 p->sched_class->task_woken(rq, p);
1773 task_rq_unlock(rq, p, &flags);
1776 #ifdef CONFIG_PREEMPT_NOTIFIERS
1779 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1780 * @notifier: notifier struct to register
1782 void preempt_notifier_register(struct preempt_notifier *notifier)
1784 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1786 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1789 * preempt_notifier_unregister - no longer interested in preemption notifications
1790 * @notifier: notifier struct to unregister
1792 * This is safe to call from within a preemption notifier.
1794 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1796 hlist_del(¬ifier->link);
1798 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1800 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1802 struct preempt_notifier *notifier;
1803 struct hlist_node *node;
1805 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1806 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1810 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1811 struct task_struct *next)
1813 struct preempt_notifier *notifier;
1814 struct hlist_node *node;
1816 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1817 notifier->ops->sched_out(notifier, next);
1820 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1822 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1827 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1828 struct task_struct *next)
1832 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1835 * prepare_task_switch - prepare to switch tasks
1836 * @rq: the runqueue preparing to switch
1837 * @prev: the current task that is being switched out
1838 * @next: the task we are going to switch to.
1840 * This is called with the rq lock held and interrupts off. It must
1841 * be paired with a subsequent finish_task_switch after the context
1844 * prepare_task_switch sets up locking and calls architecture specific
1848 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1849 struct task_struct *next)
1851 sched_info_switch(prev, next);
1852 perf_event_task_sched_out(prev, next);
1853 fire_sched_out_preempt_notifiers(prev, next);
1854 prepare_lock_switch(rq, next);
1855 prepare_arch_switch(next);
1856 trace_sched_switch(prev, next);
1860 * finish_task_switch - clean up after a task-switch
1861 * @rq: runqueue associated with task-switch
1862 * @prev: the thread we just switched away from.
1864 * finish_task_switch must be called after the context switch, paired
1865 * with a prepare_task_switch call before the context switch.
1866 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1867 * and do any other architecture-specific cleanup actions.
1869 * Note that we may have delayed dropping an mm in context_switch(). If
1870 * so, we finish that here outside of the runqueue lock. (Doing it
1871 * with the lock held can cause deadlocks; see schedule() for
1874 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1875 __releases(rq->lock)
1877 struct mm_struct *mm = rq->prev_mm;
1883 * A task struct has one reference for the use as "current".
1884 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1885 * schedule one last time. The schedule call will never return, and
1886 * the scheduled task must drop that reference.
1887 * The test for TASK_DEAD must occur while the runqueue locks are
1888 * still held, otherwise prev could be scheduled on another cpu, die
1889 * there before we look at prev->state, and then the reference would
1891 * Manfred Spraul <manfred@colorfullife.com>
1893 prev_state = prev->state;
1894 finish_arch_switch(prev);
1895 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1896 local_irq_disable();
1897 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1898 perf_event_task_sched_in(prev, current);
1899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1901 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1902 finish_lock_switch(rq, prev);
1904 fire_sched_in_preempt_notifiers(current);
1907 if (unlikely(prev_state == TASK_DEAD)) {
1909 * Remove function-return probe instances associated with this
1910 * task and put them back on the free list.
1912 kprobe_flush_task(prev);
1913 put_task_struct(prev);
1919 /* assumes rq->lock is held */
1920 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1922 if (prev->sched_class->pre_schedule)
1923 prev->sched_class->pre_schedule(rq, prev);
1926 /* rq->lock is NOT held, but preemption is disabled */
1927 static inline void post_schedule(struct rq *rq)
1929 if (rq->post_schedule) {
1930 unsigned long flags;
1932 raw_spin_lock_irqsave(&rq->lock, flags);
1933 if (rq->curr->sched_class->post_schedule)
1934 rq->curr->sched_class->post_schedule(rq);
1935 raw_spin_unlock_irqrestore(&rq->lock, flags);
1937 rq->post_schedule = 0;
1943 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1947 static inline void post_schedule(struct rq *rq)
1954 * schedule_tail - first thing a freshly forked thread must call.
1955 * @prev: the thread we just switched away from.
1957 asmlinkage void schedule_tail(struct task_struct *prev)
1958 __releases(rq->lock)
1960 struct rq *rq = this_rq();
1962 finish_task_switch(rq, prev);
1965 * FIXME: do we need to worry about rq being invalidated by the
1970 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1971 /* In this case, finish_task_switch does not reenable preemption */
1974 if (current->set_child_tid)
1975 put_user(task_pid_vnr(current), current->set_child_tid);
1979 * context_switch - switch to the new MM and the new
1980 * thread's register state.
1983 context_switch(struct rq *rq, struct task_struct *prev,
1984 struct task_struct *next)
1986 struct mm_struct *mm, *oldmm;
1988 prepare_task_switch(rq, prev, next);
1991 oldmm = prev->active_mm;
1993 * For paravirt, this is coupled with an exit in switch_to to
1994 * combine the page table reload and the switch backend into
1997 arch_start_context_switch(prev);
2000 next->active_mm = oldmm;
2001 atomic_inc(&oldmm->mm_count);
2002 enter_lazy_tlb(oldmm, next);
2004 switch_mm(oldmm, mm, next);
2007 prev->active_mm = NULL;
2008 rq->prev_mm = oldmm;
2011 * Since the runqueue lock will be released by the next
2012 * task (which is an invalid locking op but in the case
2013 * of the scheduler it's an obvious special-case), so we
2014 * do an early lockdep release here:
2016 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2017 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2020 /* Here we just switch the register state and the stack. */
2021 switch_to(prev, next, prev);
2025 * this_rq must be evaluated again because prev may have moved
2026 * CPUs since it called schedule(), thus the 'rq' on its stack
2027 * frame will be invalid.
2029 finish_task_switch(this_rq(), prev);
2033 * nr_running, nr_uninterruptible and nr_context_switches:
2035 * externally visible scheduler statistics: current number of runnable
2036 * threads, current number of uninterruptible-sleeping threads, total
2037 * number of context switches performed since bootup.
2039 unsigned long nr_running(void)
2041 unsigned long i, sum = 0;
2043 for_each_online_cpu(i)
2044 sum += cpu_rq(i)->nr_running;
2049 unsigned long nr_uninterruptible(void)
2051 unsigned long i, sum = 0;
2053 for_each_possible_cpu(i)
2054 sum += cpu_rq(i)->nr_uninterruptible;
2057 * Since we read the counters lockless, it might be slightly
2058 * inaccurate. Do not allow it to go below zero though:
2060 if (unlikely((long)sum < 0))
2066 unsigned long long nr_context_switches(void)
2069 unsigned long long sum = 0;
2071 for_each_possible_cpu(i)
2072 sum += cpu_rq(i)->nr_switches;
2077 unsigned long nr_iowait(void)
2079 unsigned long i, sum = 0;
2081 for_each_possible_cpu(i)
2082 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2087 unsigned long nr_iowait_cpu(int cpu)
2089 struct rq *this = cpu_rq(cpu);
2090 return atomic_read(&this->nr_iowait);
2093 unsigned long this_cpu_load(void)
2095 struct rq *this = this_rq();
2096 return this->cpu_load[0];
2100 /* Variables and functions for calc_load */
2101 static atomic_long_t calc_load_tasks;
2102 static unsigned long calc_load_update;
2103 unsigned long avenrun[3];
2104 EXPORT_SYMBOL(avenrun);
2106 static long calc_load_fold_active(struct rq *this_rq)
2108 long nr_active, delta = 0;
2110 nr_active = this_rq->nr_running;
2111 nr_active += (long) this_rq->nr_uninterruptible;
2113 if (nr_active != this_rq->calc_load_active) {
2114 delta = nr_active - this_rq->calc_load_active;
2115 this_rq->calc_load_active = nr_active;
2121 static unsigned long
2122 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2125 load += active * (FIXED_1 - exp);
2126 load += 1UL << (FSHIFT - 1);
2127 return load >> FSHIFT;
2132 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2134 * When making the ILB scale, we should try to pull this in as well.
2136 static atomic_long_t calc_load_tasks_idle;
2138 void calc_load_account_idle(struct rq *this_rq)
2142 delta = calc_load_fold_active(this_rq);
2144 atomic_long_add(delta, &calc_load_tasks_idle);
2147 static long calc_load_fold_idle(void)
2152 * Its got a race, we don't care...
2154 if (atomic_long_read(&calc_load_tasks_idle))
2155 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2161 * fixed_power_int - compute: x^n, in O(log n) time
2163 * @x: base of the power
2164 * @frac_bits: fractional bits of @x
2165 * @n: power to raise @x to.
2167 * By exploiting the relation between the definition of the natural power
2168 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2169 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2170 * (where: n_i \elem {0, 1}, the binary vector representing n),
2171 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2172 * of course trivially computable in O(log_2 n), the length of our binary
2175 static unsigned long
2176 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2178 unsigned long result = 1UL << frac_bits;
2183 result += 1UL << (frac_bits - 1);
2184 result >>= frac_bits;
2190 x += 1UL << (frac_bits - 1);
2198 * a1 = a0 * e + a * (1 - e)
2200 * a2 = a1 * e + a * (1 - e)
2201 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2202 * = a0 * e^2 + a * (1 - e) * (1 + e)
2204 * a3 = a2 * e + a * (1 - e)
2205 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2206 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2210 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2211 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2212 * = a0 * e^n + a * (1 - e^n)
2214 * [1] application of the geometric series:
2217 * S_n := \Sum x^i = -------------
2220 static unsigned long
2221 calc_load_n(unsigned long load, unsigned long exp,
2222 unsigned long active, unsigned int n)
2225 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2229 * NO_HZ can leave us missing all per-cpu ticks calling
2230 * calc_load_account_active(), but since an idle CPU folds its delta into
2231 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2232 * in the pending idle delta if our idle period crossed a load cycle boundary.
2234 * Once we've updated the global active value, we need to apply the exponential
2235 * weights adjusted to the number of cycles missed.
2237 static void calc_global_nohz(unsigned long ticks)
2239 long delta, active, n;
2241 if (time_before(jiffies, calc_load_update))
2245 * If we crossed a calc_load_update boundary, make sure to fold
2246 * any pending idle changes, the respective CPUs might have
2247 * missed the tick driven calc_load_account_active() update
2250 delta = calc_load_fold_idle();
2252 atomic_long_add(delta, &calc_load_tasks);
2255 * If we were idle for multiple load cycles, apply them.
2257 if (ticks >= LOAD_FREQ) {
2258 n = ticks / LOAD_FREQ;
2260 active = atomic_long_read(&calc_load_tasks);
2261 active = active > 0 ? active * FIXED_1 : 0;
2263 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2264 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2265 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2267 calc_load_update += n * LOAD_FREQ;
2271 * Its possible the remainder of the above division also crosses
2272 * a LOAD_FREQ period, the regular check in calc_global_load()
2273 * which comes after this will take care of that.
2275 * Consider us being 11 ticks before a cycle completion, and us
2276 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2277 * age us 4 cycles, and the test in calc_global_load() will
2278 * pick up the final one.
2282 void calc_load_account_idle(struct rq *this_rq)
2286 static inline long calc_load_fold_idle(void)
2291 static void calc_global_nohz(unsigned long ticks)
2297 * get_avenrun - get the load average array
2298 * @loads: pointer to dest load array
2299 * @offset: offset to add
2300 * @shift: shift count to shift the result left
2302 * These values are estimates at best, so no need for locking.
2304 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2306 loads[0] = (avenrun[0] + offset) << shift;
2307 loads[1] = (avenrun[1] + offset) << shift;
2308 loads[2] = (avenrun[2] + offset) << shift;
2312 * calc_load - update the avenrun load estimates 10 ticks after the
2313 * CPUs have updated calc_load_tasks.
2315 void calc_global_load(unsigned long ticks)
2319 calc_global_nohz(ticks);
2321 if (time_before(jiffies, calc_load_update + 10))
2324 active = atomic_long_read(&calc_load_tasks);
2325 active = active > 0 ? active * FIXED_1 : 0;
2327 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2328 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2329 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2331 calc_load_update += LOAD_FREQ;
2335 * Called from update_cpu_load() to periodically update this CPU's
2338 static void calc_load_account_active(struct rq *this_rq)
2342 if (time_before(jiffies, this_rq->calc_load_update))
2345 delta = calc_load_fold_active(this_rq);
2346 delta += calc_load_fold_idle();
2348 atomic_long_add(delta, &calc_load_tasks);
2350 this_rq->calc_load_update += LOAD_FREQ;
2354 * The exact cpuload at various idx values, calculated at every tick would be
2355 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2357 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2358 * on nth tick when cpu may be busy, then we have:
2359 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2360 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2362 * decay_load_missed() below does efficient calculation of
2363 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2364 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2366 * The calculation is approximated on a 128 point scale.
2367 * degrade_zero_ticks is the number of ticks after which load at any
2368 * particular idx is approximated to be zero.
2369 * degrade_factor is a precomputed table, a row for each load idx.
2370 * Each column corresponds to degradation factor for a power of two ticks,
2371 * based on 128 point scale.
2373 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2374 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2376 * With this power of 2 load factors, we can degrade the load n times
2377 * by looking at 1 bits in n and doing as many mult/shift instead of
2378 * n mult/shifts needed by the exact degradation.
2380 #define DEGRADE_SHIFT 7
2381 static const unsigned char
2382 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2383 static const unsigned char
2384 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2385 {0, 0, 0, 0, 0, 0, 0, 0},
2386 {64, 32, 8, 0, 0, 0, 0, 0},
2387 {96, 72, 40, 12, 1, 0, 0},
2388 {112, 98, 75, 43, 15, 1, 0},
2389 {120, 112, 98, 76, 45, 16, 2} };
2392 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2393 * would be when CPU is idle and so we just decay the old load without
2394 * adding any new load.
2396 static unsigned long
2397 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2401 if (!missed_updates)
2404 if (missed_updates >= degrade_zero_ticks[idx])
2408 return load >> missed_updates;
2410 while (missed_updates) {
2411 if (missed_updates % 2)
2412 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2414 missed_updates >>= 1;
2421 * Update rq->cpu_load[] statistics. This function is usually called every
2422 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2423 * every tick. We fix it up based on jiffies.
2425 void update_cpu_load(struct rq *this_rq)
2427 unsigned long this_load = this_rq->load.weight;
2428 unsigned long curr_jiffies = jiffies;
2429 unsigned long pending_updates;
2432 this_rq->nr_load_updates++;
2434 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2435 if (curr_jiffies == this_rq->last_load_update_tick)
2438 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2439 this_rq->last_load_update_tick = curr_jiffies;
2441 /* Update our load: */
2442 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2443 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2444 unsigned long old_load, new_load;
2446 /* scale is effectively 1 << i now, and >> i divides by scale */
2448 old_load = this_rq->cpu_load[i];
2449 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2450 new_load = this_load;
2452 * Round up the averaging division if load is increasing. This
2453 * prevents us from getting stuck on 9 if the load is 10, for
2456 if (new_load > old_load)
2457 new_load += scale - 1;
2459 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2462 sched_avg_update(this_rq);
2465 static void update_cpu_load_active(struct rq *this_rq)
2467 update_cpu_load(this_rq);
2469 calc_load_account_active(this_rq);
2475 * sched_exec - execve() is a valuable balancing opportunity, because at
2476 * this point the task has the smallest effective memory and cache footprint.
2478 void sched_exec(void)
2480 struct task_struct *p = current;
2481 unsigned long flags;
2484 raw_spin_lock_irqsave(&p->pi_lock, flags);
2485 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2486 if (dest_cpu == smp_processor_id())
2489 if (likely(cpu_active(dest_cpu))) {
2490 struct migration_arg arg = { p, dest_cpu };
2492 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2493 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2497 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2502 DEFINE_PER_CPU(struct kernel_stat, kstat);
2503 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2505 EXPORT_PER_CPU_SYMBOL(kstat);
2506 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2509 * Return any ns on the sched_clock that have not yet been accounted in
2510 * @p in case that task is currently running.
2512 * Called with task_rq_lock() held on @rq.
2514 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2518 if (task_current(rq, p)) {
2519 update_rq_clock(rq);
2520 ns = rq->clock_task - p->se.exec_start;
2528 unsigned long long task_delta_exec(struct task_struct *p)
2530 unsigned long flags;
2534 rq = task_rq_lock(p, &flags);
2535 ns = do_task_delta_exec(p, rq);
2536 task_rq_unlock(rq, p, &flags);
2542 * Return accounted runtime for the task.
2543 * In case the task is currently running, return the runtime plus current's
2544 * pending runtime that have not been accounted yet.
2546 unsigned long long task_sched_runtime(struct task_struct *p)
2548 unsigned long flags;
2552 rq = task_rq_lock(p, &flags);
2553 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2554 task_rq_unlock(rq, p, &flags);
2560 * Account user cpu time to a process.
2561 * @p: the process that the cpu time gets accounted to
2562 * @cputime: the cpu time spent in user space since the last update
2563 * @cputime_scaled: cputime scaled by cpu frequency
2565 void account_user_time(struct task_struct *p, cputime_t cputime,
2566 cputime_t cputime_scaled)
2568 u64 *cpustat = kcpustat_this_cpu->cpustat;
2572 /* Add user time to process. */
2573 p->utime = cputime_add(p->utime, cputime);
2574 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
2575 account_group_user_time(p, cputime);
2577 /* Add user time to cpustat. */
2578 tmp = cputime_to_cputime64(cputime);
2580 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2581 cpustat[index] += tmp;
2583 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
2584 /* Account for user time used */
2585 acct_update_integrals(p);
2589 * Account guest cpu time to a process.
2590 * @p: the process that the cpu time gets accounted to
2591 * @cputime: the cpu time spent in virtual machine since the last update
2592 * @cputime_scaled: cputime scaled by cpu frequency
2594 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2595 cputime_t cputime_scaled)
2598 u64 *cpustat = kcpustat_this_cpu->cpustat;
2600 tmp = cputime_to_cputime64(cputime);
2602 /* Add guest time to process. */
2603 p->utime = cputime_add(p->utime, cputime);
2604 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
2605 account_group_user_time(p, cputime);
2606 p->gtime = cputime_add(p->gtime, cputime);
2608 /* Add guest time to cpustat. */
2609 if (TASK_NICE(p) > 0) {
2610 cpustat[CPUTIME_NICE] += tmp;
2611 cpustat[CPUTIME_GUEST_NICE] += tmp;
2613 cpustat[CPUTIME_USER] += tmp;
2614 cpustat[CPUTIME_GUEST] += tmp;
2619 * Account system cpu time to a process and desired cpustat field
2620 * @p: the process that the cpu time gets accounted to
2621 * @cputime: the cpu time spent in kernel space since the last update
2622 * @cputime_scaled: cputime scaled by cpu frequency
2623 * @target_cputime64: pointer to cpustat field that has to be updated
2626 void __account_system_time(struct task_struct *p, cputime_t cputime,
2627 cputime_t cputime_scaled, int index)
2629 u64 tmp = cputime_to_cputime64(cputime);
2630 u64 *cpustat = kcpustat_this_cpu->cpustat;
2632 /* Add system time to process. */
2633 p->stime = cputime_add(p->stime, cputime);
2634 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
2635 account_group_system_time(p, cputime);
2637 /* Add system time to cpustat. */
2638 cpustat[index] += tmp;
2639 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
2641 /* Account for system time used */
2642 acct_update_integrals(p);
2646 * Account system cpu time to a process.
2647 * @p: the process that the cpu time gets accounted to
2648 * @hardirq_offset: the offset to subtract from hardirq_count()
2649 * @cputime: the cpu time spent in kernel space since the last update
2650 * @cputime_scaled: cputime scaled by cpu frequency
2652 void account_system_time(struct task_struct *p, int hardirq_offset,
2653 cputime_t cputime, cputime_t cputime_scaled)
2657 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2658 account_guest_time(p, cputime, cputime_scaled);
2662 if (hardirq_count() - hardirq_offset)
2663 index = CPUTIME_IRQ;
2664 else if (in_serving_softirq())
2665 index = CPUTIME_SOFTIRQ;
2667 index = CPUTIME_SYSTEM;
2669 __account_system_time(p, cputime, cputime_scaled, index);
2673 * Account for involuntary wait time.
2674 * @cputime: the cpu time spent in involuntary wait
2676 void account_steal_time(cputime_t cputime)
2678 u64 *cpustat = kcpustat_this_cpu->cpustat;
2679 u64 cputime64 = cputime_to_cputime64(cputime);
2681 cpustat[CPUTIME_STEAL] += cputime64;
2685 * Account for idle time.
2686 * @cputime: the cpu time spent in idle wait
2688 void account_idle_time(cputime_t cputime)
2690 u64 *cpustat = kcpustat_this_cpu->cpustat;
2691 u64 cputime64 = cputime_to_cputime64(cputime);
2692 struct rq *rq = this_rq();
2694 if (atomic_read(&rq->nr_iowait) > 0)
2695 cpustat[CPUTIME_IOWAIT] += cputime64;
2697 cpustat[CPUTIME_IDLE] += cputime64;
2700 static __always_inline bool steal_account_process_tick(void)
2702 #ifdef CONFIG_PARAVIRT
2703 if (static_branch(¶virt_steal_enabled)) {
2706 steal = paravirt_steal_clock(smp_processor_id());
2707 steal -= this_rq()->prev_steal_time;
2709 st = steal_ticks(steal);
2710 this_rq()->prev_steal_time += st * TICK_NSEC;
2712 account_steal_time(st);
2719 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2721 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2723 * Account a tick to a process and cpustat
2724 * @p: the process that the cpu time gets accounted to
2725 * @user_tick: is the tick from userspace
2726 * @rq: the pointer to rq
2728 * Tick demultiplexing follows the order
2729 * - pending hardirq update
2730 * - pending softirq update
2734 * - check for guest_time
2735 * - else account as system_time
2737 * Check for hardirq is done both for system and user time as there is
2738 * no timer going off while we are on hardirq and hence we may never get an
2739 * opportunity to update it solely in system time.
2740 * p->stime and friends are only updated on system time and not on irq
2741 * softirq as those do not count in task exec_runtime any more.
2743 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2746 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2747 u64 tmp = cputime_to_cputime64(cputime_one_jiffy);
2748 u64 *cpustat = kcpustat_this_cpu->cpustat;
2750 if (steal_account_process_tick())
2753 if (irqtime_account_hi_update()) {
2754 cpustat[CPUTIME_IRQ] += tmp;
2755 } else if (irqtime_account_si_update()) {
2756 cpustat[CPUTIME_SOFTIRQ] += tmp;
2757 } else if (this_cpu_ksoftirqd() == p) {
2759 * ksoftirqd time do not get accounted in cpu_softirq_time.
2760 * So, we have to handle it separately here.
2761 * Also, p->stime needs to be updated for ksoftirqd.
2763 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2765 } else if (user_tick) {
2766 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2767 } else if (p == rq->idle) {
2768 account_idle_time(cputime_one_jiffy);
2769 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2770 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2772 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2777 static void irqtime_account_idle_ticks(int ticks)
2780 struct rq *rq = this_rq();
2782 for (i = 0; i < ticks; i++)
2783 irqtime_account_process_tick(current, 0, rq);
2785 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2786 static void irqtime_account_idle_ticks(int ticks) {}
2787 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2789 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2792 * Account a single tick of cpu time.
2793 * @p: the process that the cpu time gets accounted to
2794 * @user_tick: indicates if the tick is a user or a system tick
2796 void account_process_tick(struct task_struct *p, int user_tick)
2798 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2799 struct rq *rq = this_rq();
2801 if (sched_clock_irqtime) {
2802 irqtime_account_process_tick(p, user_tick, rq);
2806 if (steal_account_process_tick())
2810 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2811 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2812 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2815 account_idle_time(cputime_one_jiffy);
2819 * Account multiple ticks of steal time.
2820 * @p: the process from which the cpu time has been stolen
2821 * @ticks: number of stolen ticks
2823 void account_steal_ticks(unsigned long ticks)
2825 account_steal_time(jiffies_to_cputime(ticks));
2829 * Account multiple ticks of idle time.
2830 * @ticks: number of stolen ticks
2832 void account_idle_ticks(unsigned long ticks)
2835 if (sched_clock_irqtime) {
2836 irqtime_account_idle_ticks(ticks);
2840 account_idle_time(jiffies_to_cputime(ticks));
2846 * Use precise platform statistics if available:
2848 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2849 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2855 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2857 struct task_cputime cputime;
2859 thread_group_cputime(p, &cputime);
2861 *ut = cputime.utime;
2862 *st = cputime.stime;
2866 #ifndef nsecs_to_cputime
2867 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2870 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2872 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
2875 * Use CFS's precise accounting:
2877 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2883 do_div(temp, total);
2884 utime = (cputime_t)temp;
2889 * Compare with previous values, to keep monotonicity:
2891 p->prev_utime = max(p->prev_utime, utime);
2892 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
2894 *ut = p->prev_utime;
2895 *st = p->prev_stime;
2899 * Must be called with siglock held.
2901 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2903 struct signal_struct *sig = p->signal;
2904 struct task_cputime cputime;
2905 cputime_t rtime, utime, total;
2907 thread_group_cputime(p, &cputime);
2909 total = cputime_add(cputime.utime, cputime.stime);
2910 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2915 temp *= cputime.utime;
2916 do_div(temp, total);
2917 utime = (cputime_t)temp;
2921 sig->prev_utime = max(sig->prev_utime, utime);
2922 sig->prev_stime = max(sig->prev_stime,
2923 cputime_sub(rtime, sig->prev_utime));
2925 *ut = sig->prev_utime;
2926 *st = sig->prev_stime;
2931 * This function gets called by the timer code, with HZ frequency.
2932 * We call it with interrupts disabled.
2934 void scheduler_tick(void)
2936 int cpu = smp_processor_id();
2937 struct rq *rq = cpu_rq(cpu);
2938 struct task_struct *curr = rq->curr;
2942 raw_spin_lock(&rq->lock);
2943 update_rq_clock(rq);
2944 update_cpu_load_active(rq);
2945 curr->sched_class->task_tick(rq, curr, 0);
2946 raw_spin_unlock(&rq->lock);
2948 perf_event_task_tick();
2951 rq->idle_balance = idle_cpu(cpu);
2952 trigger_load_balance(rq, cpu);
2956 notrace unsigned long get_parent_ip(unsigned long addr)
2958 if (in_lock_functions(addr)) {
2959 addr = CALLER_ADDR2;
2960 if (in_lock_functions(addr))
2961 addr = CALLER_ADDR3;
2966 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2967 defined(CONFIG_PREEMPT_TRACER))
2969 void __kprobes add_preempt_count(int val)
2971 #ifdef CONFIG_DEBUG_PREEMPT
2975 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2978 preempt_count() += val;
2979 #ifdef CONFIG_DEBUG_PREEMPT
2981 * Spinlock count overflowing soon?
2983 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2986 if (preempt_count() == val)
2987 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2989 EXPORT_SYMBOL(add_preempt_count);
2991 void __kprobes sub_preempt_count(int val)
2993 #ifdef CONFIG_DEBUG_PREEMPT
2997 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3000 * Is the spinlock portion underflowing?
3002 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3003 !(preempt_count() & PREEMPT_MASK)))
3007 if (preempt_count() == val)
3008 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3009 preempt_count() -= val;
3011 EXPORT_SYMBOL(sub_preempt_count);
3016 * Print scheduling while atomic bug:
3018 static noinline void __schedule_bug(struct task_struct *prev)
3020 struct pt_regs *regs = get_irq_regs();
3022 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3023 prev->comm, prev->pid, preempt_count());
3025 debug_show_held_locks(prev);
3027 if (irqs_disabled())
3028 print_irqtrace_events(prev);
3037 * Various schedule()-time debugging checks and statistics:
3039 static inline void schedule_debug(struct task_struct *prev)
3042 * Test if we are atomic. Since do_exit() needs to call into
3043 * schedule() atomically, we ignore that path for now.
3044 * Otherwise, whine if we are scheduling when we should not be.
3046 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3047 __schedule_bug(prev);
3050 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3052 schedstat_inc(this_rq(), sched_count);
3055 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3057 if (prev->on_rq || rq->skip_clock_update < 0)
3058 update_rq_clock(rq);
3059 prev->sched_class->put_prev_task(rq, prev);
3063 * Pick up the highest-prio task:
3065 static inline struct task_struct *
3066 pick_next_task(struct rq *rq)
3068 const struct sched_class *class;
3069 struct task_struct *p;
3072 * Optimization: we know that if all tasks are in
3073 * the fair class we can call that function directly:
3075 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3076 p = fair_sched_class.pick_next_task(rq);
3081 for_each_class(class) {
3082 p = class->pick_next_task(rq);
3087 BUG(); /* the idle class will always have a runnable task */
3091 * __schedule() is the main scheduler function.
3093 static void __sched __schedule(void)
3095 struct task_struct *prev, *next;
3096 unsigned long *switch_count;
3102 cpu = smp_processor_id();
3104 rcu_note_context_switch(cpu);
3107 schedule_debug(prev);
3109 if (sched_feat(HRTICK))
3112 raw_spin_lock_irq(&rq->lock);
3114 switch_count = &prev->nivcsw;
3115 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3116 if (unlikely(signal_pending_state(prev->state, prev))) {
3117 prev->state = TASK_RUNNING;
3119 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3123 * If a worker went to sleep, notify and ask workqueue
3124 * whether it wants to wake up a task to maintain
3127 if (prev->flags & PF_WQ_WORKER) {
3128 struct task_struct *to_wakeup;
3130 to_wakeup = wq_worker_sleeping(prev, cpu);
3132 try_to_wake_up_local(to_wakeup);
3135 switch_count = &prev->nvcsw;
3138 pre_schedule(rq, prev);
3140 if (unlikely(!rq->nr_running))
3141 idle_balance(cpu, rq);
3143 put_prev_task(rq, prev);
3144 next = pick_next_task(rq);
3145 clear_tsk_need_resched(prev);
3146 rq->skip_clock_update = 0;
3148 if (likely(prev != next)) {
3153 context_switch(rq, prev, next); /* unlocks the rq */
3155 * The context switch have flipped the stack from under us
3156 * and restored the local variables which were saved when
3157 * this task called schedule() in the past. prev == current
3158 * is still correct, but it can be moved to another cpu/rq.
3160 cpu = smp_processor_id();
3163 raw_spin_unlock_irq(&rq->lock);
3167 preempt_enable_no_resched();
3172 static inline void sched_submit_work(struct task_struct *tsk)
3177 * If we are going to sleep and we have plugged IO queued,
3178 * make sure to submit it to avoid deadlocks.
3180 if (blk_needs_flush_plug(tsk))
3181 blk_schedule_flush_plug(tsk);
3184 asmlinkage void __sched schedule(void)
3186 struct task_struct *tsk = current;
3188 sched_submit_work(tsk);
3191 EXPORT_SYMBOL(schedule);
3193 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3195 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3197 if (lock->owner != owner)
3201 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3202 * lock->owner still matches owner, if that fails, owner might
3203 * point to free()d memory, if it still matches, the rcu_read_lock()
3204 * ensures the memory stays valid.
3208 return owner->on_cpu;
3212 * Look out! "owner" is an entirely speculative pointer
3213 * access and not reliable.
3215 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3217 if (!sched_feat(OWNER_SPIN))
3221 while (owner_running(lock, owner)) {
3225 arch_mutex_cpu_relax();
3230 * We break out the loop above on need_resched() and when the
3231 * owner changed, which is a sign for heavy contention. Return
3232 * success only when lock->owner is NULL.
3234 return lock->owner == NULL;
3238 #ifdef CONFIG_PREEMPT
3240 * this is the entry point to schedule() from in-kernel preemption
3241 * off of preempt_enable. Kernel preemptions off return from interrupt
3242 * occur there and call schedule directly.
3244 asmlinkage void __sched notrace preempt_schedule(void)
3246 struct thread_info *ti = current_thread_info();
3249 * If there is a non-zero preempt_count or interrupts are disabled,
3250 * we do not want to preempt the current task. Just return..
3252 if (likely(ti->preempt_count || irqs_disabled()))
3256 add_preempt_count_notrace(PREEMPT_ACTIVE);
3258 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3261 * Check again in case we missed a preemption opportunity
3262 * between schedule and now.
3265 } while (need_resched());
3267 EXPORT_SYMBOL(preempt_schedule);
3270 * this is the entry point to schedule() from kernel preemption
3271 * off of irq context.
3272 * Note, that this is called and return with irqs disabled. This will
3273 * protect us against recursive calling from irq.
3275 asmlinkage void __sched preempt_schedule_irq(void)
3277 struct thread_info *ti = current_thread_info();
3279 /* Catch callers which need to be fixed */
3280 BUG_ON(ti->preempt_count || !irqs_disabled());
3283 add_preempt_count(PREEMPT_ACTIVE);
3286 local_irq_disable();
3287 sub_preempt_count(PREEMPT_ACTIVE);
3290 * Check again in case we missed a preemption opportunity
3291 * between schedule and now.
3294 } while (need_resched());
3297 #endif /* CONFIG_PREEMPT */
3299 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3302 return try_to_wake_up(curr->private, mode, wake_flags);
3304 EXPORT_SYMBOL(default_wake_function);
3307 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3308 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3309 * number) then we wake all the non-exclusive tasks and one exclusive task.
3311 * There are circumstances in which we can try to wake a task which has already
3312 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3313 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3315 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3316 int nr_exclusive, int wake_flags, void *key)
3318 wait_queue_t *curr, *next;
3320 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3321 unsigned flags = curr->flags;
3323 if (curr->func(curr, mode, wake_flags, key) &&
3324 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3330 * __wake_up - wake up threads blocked on a waitqueue.
3332 * @mode: which threads
3333 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3334 * @key: is directly passed to the wakeup function
3336 * It may be assumed that this function implies a write memory barrier before
3337 * changing the task state if and only if any tasks are woken up.
3339 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3340 int nr_exclusive, void *key)
3342 unsigned long flags;
3344 spin_lock_irqsave(&q->lock, flags);
3345 __wake_up_common(q, mode, nr_exclusive, 0, key);
3346 spin_unlock_irqrestore(&q->lock, flags);
3348 EXPORT_SYMBOL(__wake_up);
3351 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3353 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3355 __wake_up_common(q, mode, 1, 0, NULL);
3357 EXPORT_SYMBOL_GPL(__wake_up_locked);
3359 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3361 __wake_up_common(q, mode, 1, 0, key);
3363 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3366 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3368 * @mode: which threads
3369 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3370 * @key: opaque value to be passed to wakeup targets
3372 * The sync wakeup differs that the waker knows that it will schedule
3373 * away soon, so while the target thread will be woken up, it will not
3374 * be migrated to another CPU - ie. the two threads are 'synchronized'
3375 * with each other. This can prevent needless bouncing between CPUs.
3377 * On UP it can prevent extra preemption.
3379 * It may be assumed that this function implies a write memory barrier before
3380 * changing the task state if and only if any tasks are woken up.
3382 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3383 int nr_exclusive, void *key)
3385 unsigned long flags;
3386 int wake_flags = WF_SYNC;
3391 if (unlikely(!nr_exclusive))
3394 spin_lock_irqsave(&q->lock, flags);
3395 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3396 spin_unlock_irqrestore(&q->lock, flags);
3398 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3401 * __wake_up_sync - see __wake_up_sync_key()
3403 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3405 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3407 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3410 * complete: - signals a single thread waiting on this completion
3411 * @x: holds the state of this particular completion
3413 * This will wake up a single thread waiting on this completion. Threads will be
3414 * awakened in the same order in which they were queued.
3416 * See also complete_all(), wait_for_completion() and related routines.
3418 * It may be assumed that this function implies a write memory barrier before
3419 * changing the task state if and only if any tasks are woken up.
3421 void complete(struct completion *x)
3423 unsigned long flags;
3425 spin_lock_irqsave(&x->wait.lock, flags);
3427 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3428 spin_unlock_irqrestore(&x->wait.lock, flags);
3430 EXPORT_SYMBOL(complete);
3433 * complete_all: - signals all threads waiting on this completion
3434 * @x: holds the state of this particular completion
3436 * This will wake up all threads waiting on this particular completion event.
3438 * It may be assumed that this function implies a write memory barrier before
3439 * changing the task state if and only if any tasks are woken up.
3441 void complete_all(struct completion *x)
3443 unsigned long flags;
3445 spin_lock_irqsave(&x->wait.lock, flags);
3446 x->done += UINT_MAX/2;
3447 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3448 spin_unlock_irqrestore(&x->wait.lock, flags);
3450 EXPORT_SYMBOL(complete_all);
3452 static inline long __sched
3453 do_wait_for_common(struct completion *x, long timeout, int state)
3456 DECLARE_WAITQUEUE(wait, current);
3458 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3460 if (signal_pending_state(state, current)) {
3461 timeout = -ERESTARTSYS;
3464 __set_current_state(state);
3465 spin_unlock_irq(&x->wait.lock);
3466 timeout = schedule_timeout(timeout);
3467 spin_lock_irq(&x->wait.lock);
3468 } while (!x->done && timeout);
3469 __remove_wait_queue(&x->wait, &wait);
3474 return timeout ?: 1;
3478 wait_for_common(struct completion *x, long timeout, int state)
3482 spin_lock_irq(&x->wait.lock);
3483 timeout = do_wait_for_common(x, timeout, state);
3484 spin_unlock_irq(&x->wait.lock);
3489 * wait_for_completion: - waits for completion of a task
3490 * @x: holds the state of this particular completion
3492 * This waits to be signaled for completion of a specific task. It is NOT
3493 * interruptible and there is no timeout.
3495 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3496 * and interrupt capability. Also see complete().
3498 void __sched wait_for_completion(struct completion *x)
3500 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3502 EXPORT_SYMBOL(wait_for_completion);
3505 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3506 * @x: holds the state of this particular completion
3507 * @timeout: timeout value in jiffies
3509 * This waits for either a completion of a specific task to be signaled or for a
3510 * specified timeout to expire. The timeout is in jiffies. It is not
3513 * The return value is 0 if timed out, and positive (at least 1, or number of
3514 * jiffies left till timeout) if completed.
3516 unsigned long __sched
3517 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3519 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3521 EXPORT_SYMBOL(wait_for_completion_timeout);
3524 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3525 * @x: holds the state of this particular completion
3527 * This waits for completion of a specific task to be signaled. It is
3530 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3532 int __sched wait_for_completion_interruptible(struct completion *x)
3534 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3535 if (t == -ERESTARTSYS)
3539 EXPORT_SYMBOL(wait_for_completion_interruptible);
3542 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3543 * @x: holds the state of this particular completion
3544 * @timeout: timeout value in jiffies
3546 * This waits for either a completion of a specific task to be signaled or for a
3547 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3549 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3550 * positive (at least 1, or number of jiffies left till timeout) if completed.
3553 wait_for_completion_interruptible_timeout(struct completion *x,
3554 unsigned long timeout)
3556 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3558 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3561 * wait_for_completion_killable: - waits for completion of a task (killable)
3562 * @x: holds the state of this particular completion
3564 * This waits to be signaled for completion of a specific task. It can be
3565 * interrupted by a kill signal.
3567 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3569 int __sched wait_for_completion_killable(struct completion *x)
3571 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3572 if (t == -ERESTARTSYS)
3576 EXPORT_SYMBOL(wait_for_completion_killable);
3579 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3580 * @x: holds the state of this particular completion
3581 * @timeout: timeout value in jiffies
3583 * This waits for either a completion of a specific task to be
3584 * signaled or for a specified timeout to expire. It can be
3585 * interrupted by a kill signal. The timeout is in jiffies.
3587 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3588 * positive (at least 1, or number of jiffies left till timeout) if completed.
3591 wait_for_completion_killable_timeout(struct completion *x,
3592 unsigned long timeout)
3594 return wait_for_common(x, timeout, TASK_KILLABLE);
3596 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3599 * try_wait_for_completion - try to decrement a completion without blocking
3600 * @x: completion structure
3602 * Returns: 0 if a decrement cannot be done without blocking
3603 * 1 if a decrement succeeded.
3605 * If a completion is being used as a counting completion,
3606 * attempt to decrement the counter without blocking. This
3607 * enables us to avoid waiting if the resource the completion
3608 * is protecting is not available.
3610 bool try_wait_for_completion(struct completion *x)
3612 unsigned long flags;
3615 spin_lock_irqsave(&x->wait.lock, flags);
3620 spin_unlock_irqrestore(&x->wait.lock, flags);
3623 EXPORT_SYMBOL(try_wait_for_completion);
3626 * completion_done - Test to see if a completion has any waiters
3627 * @x: completion structure
3629 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3630 * 1 if there are no waiters.
3633 bool completion_done(struct completion *x)
3635 unsigned long flags;
3638 spin_lock_irqsave(&x->wait.lock, flags);
3641 spin_unlock_irqrestore(&x->wait.lock, flags);
3644 EXPORT_SYMBOL(completion_done);
3647 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3649 unsigned long flags;
3652 init_waitqueue_entry(&wait, current);
3654 __set_current_state(state);
3656 spin_lock_irqsave(&q->lock, flags);
3657 __add_wait_queue(q, &wait);
3658 spin_unlock(&q->lock);
3659 timeout = schedule_timeout(timeout);
3660 spin_lock_irq(&q->lock);
3661 __remove_wait_queue(q, &wait);
3662 spin_unlock_irqrestore(&q->lock, flags);
3667 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3669 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3671 EXPORT_SYMBOL(interruptible_sleep_on);
3674 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3676 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3678 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3680 void __sched sleep_on(wait_queue_head_t *q)
3682 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3684 EXPORT_SYMBOL(sleep_on);
3686 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3688 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3690 EXPORT_SYMBOL(sleep_on_timeout);
3692 #ifdef CONFIG_RT_MUTEXES
3695 * rt_mutex_setprio - set the current priority of a task
3697 * @prio: prio value (kernel-internal form)
3699 * This function changes the 'effective' priority of a task. It does
3700 * not touch ->normal_prio like __setscheduler().
3702 * Used by the rt_mutex code to implement priority inheritance logic.
3704 void rt_mutex_setprio(struct task_struct *p, int prio)
3706 int oldprio, on_rq, running;
3708 const struct sched_class *prev_class;
3710 BUG_ON(prio < 0 || prio > MAX_PRIO);
3712 rq = __task_rq_lock(p);
3714 trace_sched_pi_setprio(p, prio);
3716 prev_class = p->sched_class;
3718 running = task_current(rq, p);
3720 dequeue_task(rq, p, 0);
3722 p->sched_class->put_prev_task(rq, p);
3725 p->sched_class = &rt_sched_class;
3727 p->sched_class = &fair_sched_class;
3732 p->sched_class->set_curr_task(rq);
3734 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3736 check_class_changed(rq, p, prev_class, oldprio);
3737 __task_rq_unlock(rq);
3742 void set_user_nice(struct task_struct *p, long nice)
3744 int old_prio, delta, on_rq;
3745 unsigned long flags;
3748 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3751 * We have to be careful, if called from sys_setpriority(),
3752 * the task might be in the middle of scheduling on another CPU.
3754 rq = task_rq_lock(p, &flags);
3756 * The RT priorities are set via sched_setscheduler(), but we still
3757 * allow the 'normal' nice value to be set - but as expected
3758 * it wont have any effect on scheduling until the task is
3759 * SCHED_FIFO/SCHED_RR:
3761 if (task_has_rt_policy(p)) {
3762 p->static_prio = NICE_TO_PRIO(nice);
3767 dequeue_task(rq, p, 0);
3769 p->static_prio = NICE_TO_PRIO(nice);
3772 p->prio = effective_prio(p);
3773 delta = p->prio - old_prio;
3776 enqueue_task(rq, p, 0);
3778 * If the task increased its priority or is running and
3779 * lowered its priority, then reschedule its CPU:
3781 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3782 resched_task(rq->curr);
3785 task_rq_unlock(rq, p, &flags);
3787 EXPORT_SYMBOL(set_user_nice);
3790 * can_nice - check if a task can reduce its nice value
3794 int can_nice(const struct task_struct *p, const int nice)
3796 /* convert nice value [19,-20] to rlimit style value [1,40] */
3797 int nice_rlim = 20 - nice;
3799 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3800 capable(CAP_SYS_NICE));
3803 #ifdef __ARCH_WANT_SYS_NICE
3806 * sys_nice - change the priority of the current process.
3807 * @increment: priority increment
3809 * sys_setpriority is a more generic, but much slower function that
3810 * does similar things.
3812 SYSCALL_DEFINE1(nice, int, increment)
3817 * Setpriority might change our priority at the same moment.
3818 * We don't have to worry. Conceptually one call occurs first
3819 * and we have a single winner.
3821 if (increment < -40)
3826 nice = TASK_NICE(current) + increment;
3832 if (increment < 0 && !can_nice(current, nice))
3835 retval = security_task_setnice(current, nice);
3839 set_user_nice(current, nice);
3846 * task_prio - return the priority value of a given task.
3847 * @p: the task in question.
3849 * This is the priority value as seen by users in /proc.
3850 * RT tasks are offset by -200. Normal tasks are centered
3851 * around 0, value goes from -16 to +15.
3853 int task_prio(const struct task_struct *p)
3855 return p->prio - MAX_RT_PRIO;
3859 * task_nice - return the nice value of a given task.
3860 * @p: the task in question.
3862 int task_nice(const struct task_struct *p)
3864 return TASK_NICE(p);
3866 EXPORT_SYMBOL(task_nice);
3869 * idle_cpu - is a given cpu idle currently?
3870 * @cpu: the processor in question.
3872 int idle_cpu(int cpu)
3874 struct rq *rq = cpu_rq(cpu);
3876 if (rq->curr != rq->idle)
3883 if (!llist_empty(&rq->wake_list))
3891 * idle_task - return the idle task for a given cpu.
3892 * @cpu: the processor in question.
3894 struct task_struct *idle_task(int cpu)
3896 return cpu_rq(cpu)->idle;
3900 * find_process_by_pid - find a process with a matching PID value.
3901 * @pid: the pid in question.
3903 static struct task_struct *find_process_by_pid(pid_t pid)
3905 return pid ? find_task_by_vpid(pid) : current;
3908 /* Actually do priority change: must hold rq lock. */
3910 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3913 p->rt_priority = prio;
3914 p->normal_prio = normal_prio(p);
3915 /* we are holding p->pi_lock already */
3916 p->prio = rt_mutex_getprio(p);
3917 if (rt_prio(p->prio))
3918 p->sched_class = &rt_sched_class;
3920 p->sched_class = &fair_sched_class;
3925 * check the target process has a UID that matches the current process's
3927 static bool check_same_owner(struct task_struct *p)
3929 const struct cred *cred = current_cred(), *pcred;
3933 pcred = __task_cred(p);
3934 if (cred->user->user_ns == pcred->user->user_ns)
3935 match = (cred->euid == pcred->euid ||
3936 cred->euid == pcred->uid);
3943 static int __sched_setscheduler(struct task_struct *p, int policy,
3944 const struct sched_param *param, bool user)
3946 int retval, oldprio, oldpolicy = -1, on_rq, running;
3947 unsigned long flags;
3948 const struct sched_class *prev_class;
3952 /* may grab non-irq protected spin_locks */
3953 BUG_ON(in_interrupt());
3955 /* double check policy once rq lock held */
3957 reset_on_fork = p->sched_reset_on_fork;
3958 policy = oldpolicy = p->policy;
3960 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3961 policy &= ~SCHED_RESET_ON_FORK;
3963 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3964 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3965 policy != SCHED_IDLE)
3970 * Valid priorities for SCHED_FIFO and SCHED_RR are
3971 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3972 * SCHED_BATCH and SCHED_IDLE is 0.
3974 if (param->sched_priority < 0 ||
3975 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3976 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3978 if (rt_policy(policy) != (param->sched_priority != 0))
3982 * Allow unprivileged RT tasks to decrease priority:
3984 if (user && !capable(CAP_SYS_NICE)) {
3985 if (rt_policy(policy)) {
3986 unsigned long rlim_rtprio =
3987 task_rlimit(p, RLIMIT_RTPRIO);
3989 /* can't set/change the rt policy */
3990 if (policy != p->policy && !rlim_rtprio)
3993 /* can't increase priority */
3994 if (param->sched_priority > p->rt_priority &&
3995 param->sched_priority > rlim_rtprio)
4000 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4001 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4003 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4004 if (!can_nice(p, TASK_NICE(p)))
4008 /* can't change other user's priorities */
4009 if (!check_same_owner(p))
4012 /* Normal users shall not reset the sched_reset_on_fork flag */
4013 if (p->sched_reset_on_fork && !reset_on_fork)
4018 retval = security_task_setscheduler(p);
4024 * make sure no PI-waiters arrive (or leave) while we are
4025 * changing the priority of the task:
4027 * To be able to change p->policy safely, the appropriate
4028 * runqueue lock must be held.
4030 rq = task_rq_lock(p, &flags);
4033 * Changing the policy of the stop threads its a very bad idea
4035 if (p == rq->stop) {
4036 task_rq_unlock(rq, p, &flags);
4041 * If not changing anything there's no need to proceed further:
4043 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4044 param->sched_priority == p->rt_priority))) {
4046 __task_rq_unlock(rq);
4047 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4051 #ifdef CONFIG_RT_GROUP_SCHED
4054 * Do not allow realtime tasks into groups that have no runtime
4057 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4058 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4059 !task_group_is_autogroup(task_group(p))) {
4060 task_rq_unlock(rq, p, &flags);
4066 /* recheck policy now with rq lock held */
4067 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4068 policy = oldpolicy = -1;
4069 task_rq_unlock(rq, p, &flags);
4073 running = task_current(rq, p);
4075 deactivate_task(rq, p, 0);
4077 p->sched_class->put_prev_task(rq, p);
4079 p->sched_reset_on_fork = reset_on_fork;
4082 prev_class = p->sched_class;
4083 __setscheduler(rq, p, policy, param->sched_priority);
4086 p->sched_class->set_curr_task(rq);
4088 activate_task(rq, p, 0);
4090 check_class_changed(rq, p, prev_class, oldprio);
4091 task_rq_unlock(rq, p, &flags);
4093 rt_mutex_adjust_pi(p);
4099 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4100 * @p: the task in question.
4101 * @policy: new policy.
4102 * @param: structure containing the new RT priority.
4104 * NOTE that the task may be already dead.
4106 int sched_setscheduler(struct task_struct *p, int policy,
4107 const struct sched_param *param)
4109 return __sched_setscheduler(p, policy, param, true);
4111 EXPORT_SYMBOL_GPL(sched_setscheduler);
4114 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4115 * @p: the task in question.
4116 * @policy: new policy.
4117 * @param: structure containing the new RT priority.
4119 * Just like sched_setscheduler, only don't bother checking if the
4120 * current context has permission. For example, this is needed in
4121 * stop_machine(): we create temporary high priority worker threads,
4122 * but our caller might not have that capability.
4124 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4125 const struct sched_param *param)
4127 return __sched_setscheduler(p, policy, param, false);
4131 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4133 struct sched_param lparam;
4134 struct task_struct *p;
4137 if (!param || pid < 0)
4139 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4144 p = find_process_by_pid(pid);
4146 retval = sched_setscheduler(p, policy, &lparam);
4153 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4154 * @pid: the pid in question.
4155 * @policy: new policy.
4156 * @param: structure containing the new RT priority.
4158 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4159 struct sched_param __user *, param)
4161 /* negative values for policy are not valid */
4165 return do_sched_setscheduler(pid, policy, param);
4169 * sys_sched_setparam - set/change the RT priority of a thread
4170 * @pid: the pid in question.
4171 * @param: structure containing the new RT priority.
4173 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4175 return do_sched_setscheduler(pid, -1, param);
4179 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4180 * @pid: the pid in question.
4182 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4184 struct task_struct *p;
4192 p = find_process_by_pid(pid);
4194 retval = security_task_getscheduler(p);
4197 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4204 * sys_sched_getparam - get the RT priority of a thread
4205 * @pid: the pid in question.
4206 * @param: structure containing the RT priority.
4208 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4210 struct sched_param lp;
4211 struct task_struct *p;
4214 if (!param || pid < 0)
4218 p = find_process_by_pid(pid);
4223 retval = security_task_getscheduler(p);
4227 lp.sched_priority = p->rt_priority;
4231 * This one might sleep, we cannot do it with a spinlock held ...
4233 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4242 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4244 cpumask_var_t cpus_allowed, new_mask;
4245 struct task_struct *p;
4251 p = find_process_by_pid(pid);
4258 /* Prevent p going away */
4262 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4266 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4268 goto out_free_cpus_allowed;
4271 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
4274 retval = security_task_setscheduler(p);
4278 cpuset_cpus_allowed(p, cpus_allowed);
4279 cpumask_and(new_mask, in_mask, cpus_allowed);
4281 retval = set_cpus_allowed_ptr(p, new_mask);
4284 cpuset_cpus_allowed(p, cpus_allowed);
4285 if (!cpumask_subset(new_mask, cpus_allowed)) {
4287 * We must have raced with a concurrent cpuset
4288 * update. Just reset the cpus_allowed to the
4289 * cpuset's cpus_allowed
4291 cpumask_copy(new_mask, cpus_allowed);
4296 free_cpumask_var(new_mask);
4297 out_free_cpus_allowed:
4298 free_cpumask_var(cpus_allowed);
4305 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4306 struct cpumask *new_mask)
4308 if (len < cpumask_size())
4309 cpumask_clear(new_mask);
4310 else if (len > cpumask_size())
4311 len = cpumask_size();
4313 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4317 * sys_sched_setaffinity - set the cpu affinity of a process
4318 * @pid: pid of the process
4319 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4320 * @user_mask_ptr: user-space pointer to the new cpu mask
4322 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4323 unsigned long __user *, user_mask_ptr)
4325 cpumask_var_t new_mask;
4328 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4331 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4333 retval = sched_setaffinity(pid, new_mask);
4334 free_cpumask_var(new_mask);
4338 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4340 struct task_struct *p;
4341 unsigned long flags;
4348 p = find_process_by_pid(pid);
4352 retval = security_task_getscheduler(p);
4356 raw_spin_lock_irqsave(&p->pi_lock, flags);
4357 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4358 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4368 * sys_sched_getaffinity - get the cpu affinity of a process
4369 * @pid: pid of the process
4370 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4371 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4373 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4374 unsigned long __user *, user_mask_ptr)
4379 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4381 if (len & (sizeof(unsigned long)-1))
4384 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4387 ret = sched_getaffinity(pid, mask);
4389 size_t retlen = min_t(size_t, len, cpumask_size());
4391 if (copy_to_user(user_mask_ptr, mask, retlen))
4396 free_cpumask_var(mask);
4402 * sys_sched_yield - yield the current processor to other threads.
4404 * This function yields the current CPU to other tasks. If there are no
4405 * other threads running on this CPU then this function will return.
4407 SYSCALL_DEFINE0(sched_yield)
4409 struct rq *rq = this_rq_lock();
4411 schedstat_inc(rq, yld_count);
4412 current->sched_class->yield_task(rq);
4415 * Since we are going to call schedule() anyway, there's
4416 * no need to preempt or enable interrupts:
4418 __release(rq->lock);
4419 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4420 do_raw_spin_unlock(&rq->lock);
4421 preempt_enable_no_resched();
4428 static inline int should_resched(void)
4430 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4433 static void __cond_resched(void)
4435 add_preempt_count(PREEMPT_ACTIVE);
4437 sub_preempt_count(PREEMPT_ACTIVE);
4440 int __sched _cond_resched(void)
4442 if (should_resched()) {
4448 EXPORT_SYMBOL(_cond_resched);
4451 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4452 * call schedule, and on return reacquire the lock.
4454 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4455 * operations here to prevent schedule() from being called twice (once via
4456 * spin_unlock(), once by hand).
4458 int __cond_resched_lock(spinlock_t *lock)
4460 int resched = should_resched();
4463 lockdep_assert_held(lock);
4465 if (spin_needbreak(lock) || resched) {
4476 EXPORT_SYMBOL(__cond_resched_lock);
4478 int __sched __cond_resched_softirq(void)
4480 BUG_ON(!in_softirq());
4482 if (should_resched()) {
4490 EXPORT_SYMBOL(__cond_resched_softirq);
4493 * yield - yield the current processor to other threads.
4495 * This is a shortcut for kernel-space yielding - it marks the
4496 * thread runnable and calls sys_sched_yield().
4498 void __sched yield(void)
4500 set_current_state(TASK_RUNNING);
4503 EXPORT_SYMBOL(yield);
4506 * yield_to - yield the current processor to another thread in
4507 * your thread group, or accelerate that thread toward the
4508 * processor it's on.
4510 * @preempt: whether task preemption is allowed or not
4512 * It's the caller's job to ensure that the target task struct
4513 * can't go away on us before we can do any checks.
4515 * Returns true if we indeed boosted the target task.
4517 bool __sched yield_to(struct task_struct *p, bool preempt)
4519 struct task_struct *curr = current;
4520 struct rq *rq, *p_rq;
4521 unsigned long flags;
4524 local_irq_save(flags);
4529 double_rq_lock(rq, p_rq);
4530 while (task_rq(p) != p_rq) {
4531 double_rq_unlock(rq, p_rq);
4535 if (!curr->sched_class->yield_to_task)
4538 if (curr->sched_class != p->sched_class)
4541 if (task_running(p_rq, p) || p->state)
4544 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4546 schedstat_inc(rq, yld_count);
4548 * Make p's CPU reschedule; pick_next_entity takes care of
4551 if (preempt && rq != p_rq)
4552 resched_task(p_rq->curr);
4555 * We might have set it in task_yield_fair(), but are
4556 * not going to schedule(), so don't want to skip
4559 rq->skip_clock_update = 0;
4563 double_rq_unlock(rq, p_rq);
4564 local_irq_restore(flags);
4571 EXPORT_SYMBOL_GPL(yield_to);
4574 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4575 * that process accounting knows that this is a task in IO wait state.
4577 void __sched io_schedule(void)
4579 struct rq *rq = raw_rq();
4581 delayacct_blkio_start();
4582 atomic_inc(&rq->nr_iowait);
4583 blk_flush_plug(current);
4584 current->in_iowait = 1;
4586 current->in_iowait = 0;
4587 atomic_dec(&rq->nr_iowait);
4588 delayacct_blkio_end();
4590 EXPORT_SYMBOL(io_schedule);
4592 long __sched io_schedule_timeout(long timeout)
4594 struct rq *rq = raw_rq();
4597 delayacct_blkio_start();
4598 atomic_inc(&rq->nr_iowait);
4599 blk_flush_plug(current);
4600 current->in_iowait = 1;
4601 ret = schedule_timeout(timeout);
4602 current->in_iowait = 0;
4603 atomic_dec(&rq->nr_iowait);
4604 delayacct_blkio_end();
4609 * sys_sched_get_priority_max - return maximum RT priority.
4610 * @policy: scheduling class.
4612 * this syscall returns the maximum rt_priority that can be used
4613 * by a given scheduling class.
4615 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4622 ret = MAX_USER_RT_PRIO-1;
4634 * sys_sched_get_priority_min - return minimum RT priority.
4635 * @policy: scheduling class.
4637 * this syscall returns the minimum rt_priority that can be used
4638 * by a given scheduling class.
4640 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4658 * sys_sched_rr_get_interval - return the default timeslice of a process.
4659 * @pid: pid of the process.
4660 * @interval: userspace pointer to the timeslice value.
4662 * this syscall writes the default timeslice value of a given process
4663 * into the user-space timespec buffer. A value of '0' means infinity.
4665 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4666 struct timespec __user *, interval)
4668 struct task_struct *p;
4669 unsigned int time_slice;
4670 unsigned long flags;
4680 p = find_process_by_pid(pid);
4684 retval = security_task_getscheduler(p);
4688 rq = task_rq_lock(p, &flags);
4689 time_slice = p->sched_class->get_rr_interval(rq, p);
4690 task_rq_unlock(rq, p, &flags);
4693 jiffies_to_timespec(time_slice, &t);
4694 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4702 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4704 void sched_show_task(struct task_struct *p)
4706 unsigned long free = 0;
4709 state = p->state ? __ffs(p->state) + 1 : 0;
4710 printk(KERN_INFO "%-15.15s %c", p->comm,
4711 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4712 #if BITS_PER_LONG == 32
4713 if (state == TASK_RUNNING)
4714 printk(KERN_CONT " running ");
4716 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4718 if (state == TASK_RUNNING)
4719 printk(KERN_CONT " running task ");
4721 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4723 #ifdef CONFIG_DEBUG_STACK_USAGE
4724 free = stack_not_used(p);
4726 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4727 task_pid_nr(p), task_pid_nr(p->real_parent),
4728 (unsigned long)task_thread_info(p)->flags);
4730 show_stack(p, NULL);
4733 void show_state_filter(unsigned long state_filter)
4735 struct task_struct *g, *p;
4737 #if BITS_PER_LONG == 32
4739 " task PC stack pid father\n");
4742 " task PC stack pid father\n");
4745 do_each_thread(g, p) {
4747 * reset the NMI-timeout, listing all files on a slow
4748 * console might take a lot of time:
4750 touch_nmi_watchdog();
4751 if (!state_filter || (p->state & state_filter))
4753 } while_each_thread(g, p);
4755 touch_all_softlockup_watchdogs();
4757 #ifdef CONFIG_SCHED_DEBUG
4758 sysrq_sched_debug_show();
4762 * Only show locks if all tasks are dumped:
4765 debug_show_all_locks();
4768 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4770 idle->sched_class = &idle_sched_class;
4774 * init_idle - set up an idle thread for a given CPU
4775 * @idle: task in question
4776 * @cpu: cpu the idle task belongs to
4778 * NOTE: this function does not set the idle thread's NEED_RESCHED
4779 * flag, to make booting more robust.
4781 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4783 struct rq *rq = cpu_rq(cpu);
4784 unsigned long flags;
4786 raw_spin_lock_irqsave(&rq->lock, flags);
4789 idle->state = TASK_RUNNING;
4790 idle->se.exec_start = sched_clock();
4792 do_set_cpus_allowed(idle, cpumask_of(cpu));
4794 * We're having a chicken and egg problem, even though we are
4795 * holding rq->lock, the cpu isn't yet set to this cpu so the
4796 * lockdep check in task_group() will fail.
4798 * Similar case to sched_fork(). / Alternatively we could
4799 * use task_rq_lock() here and obtain the other rq->lock.
4804 __set_task_cpu(idle, cpu);
4807 rq->curr = rq->idle = idle;
4808 #if defined(CONFIG_SMP)
4811 raw_spin_unlock_irqrestore(&rq->lock, flags);
4813 /* Set the preempt count _outside_ the spinlocks! */
4814 task_thread_info(idle)->preempt_count = 0;
4817 * The idle tasks have their own, simple scheduling class:
4819 idle->sched_class = &idle_sched_class;
4820 ftrace_graph_init_idle_task(idle, cpu);
4821 #if defined(CONFIG_SMP)
4822 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4827 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4829 if (p->sched_class && p->sched_class->set_cpus_allowed)
4830 p->sched_class->set_cpus_allowed(p, new_mask);
4832 cpumask_copy(&p->cpus_allowed, new_mask);
4833 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4837 * This is how migration works:
4839 * 1) we invoke migration_cpu_stop() on the target CPU using
4841 * 2) stopper starts to run (implicitly forcing the migrated thread
4843 * 3) it checks whether the migrated task is still in the wrong runqueue.
4844 * 4) if it's in the wrong runqueue then the migration thread removes
4845 * it and puts it into the right queue.
4846 * 5) stopper completes and stop_one_cpu() returns and the migration
4851 * Change a given task's CPU affinity. Migrate the thread to a
4852 * proper CPU and schedule it away if the CPU it's executing on
4853 * is removed from the allowed bitmask.
4855 * NOTE: the caller must have a valid reference to the task, the
4856 * task must not exit() & deallocate itself prematurely. The
4857 * call is not atomic; no spinlocks may be held.
4859 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4861 unsigned long flags;
4863 unsigned int dest_cpu;
4866 rq = task_rq_lock(p, &flags);
4868 if (cpumask_equal(&p->cpus_allowed, new_mask))
4871 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4876 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4881 do_set_cpus_allowed(p, new_mask);
4883 /* Can the task run on the task's current CPU? If so, we're done */
4884 if (cpumask_test_cpu(task_cpu(p), new_mask))
4887 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4889 struct migration_arg arg = { p, dest_cpu };
4890 /* Need help from migration thread: drop lock and wait. */
4891 task_rq_unlock(rq, p, &flags);
4892 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4893 tlb_migrate_finish(p->mm);
4897 task_rq_unlock(rq, p, &flags);
4901 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4904 * Move (not current) task off this cpu, onto dest cpu. We're doing
4905 * this because either it can't run here any more (set_cpus_allowed()
4906 * away from this CPU, or CPU going down), or because we're
4907 * attempting to rebalance this task on exec (sched_exec).
4909 * So we race with normal scheduler movements, but that's OK, as long
4910 * as the task is no longer on this CPU.
4912 * Returns non-zero if task was successfully migrated.
4914 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4916 struct rq *rq_dest, *rq_src;
4919 if (unlikely(!cpu_active(dest_cpu)))
4922 rq_src = cpu_rq(src_cpu);
4923 rq_dest = cpu_rq(dest_cpu);
4925 raw_spin_lock(&p->pi_lock);
4926 double_rq_lock(rq_src, rq_dest);
4927 /* Already moved. */
4928 if (task_cpu(p) != src_cpu)
4930 /* Affinity changed (again). */
4931 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4935 * If we're not on a rq, the next wake-up will ensure we're
4939 deactivate_task(rq_src, p, 0);
4940 set_task_cpu(p, dest_cpu);
4941 activate_task(rq_dest, p, 0);
4942 check_preempt_curr(rq_dest, p, 0);
4947 double_rq_unlock(rq_src, rq_dest);
4948 raw_spin_unlock(&p->pi_lock);
4953 * migration_cpu_stop - this will be executed by a highprio stopper thread
4954 * and performs thread migration by bumping thread off CPU then
4955 * 'pushing' onto another runqueue.
4957 static int migration_cpu_stop(void *data)
4959 struct migration_arg *arg = data;
4962 * The original target cpu might have gone down and we might
4963 * be on another cpu but it doesn't matter.
4965 local_irq_disable();
4966 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4971 #ifdef CONFIG_HOTPLUG_CPU
4974 * Ensures that the idle task is using init_mm right before its cpu goes
4977 void idle_task_exit(void)
4979 struct mm_struct *mm = current->active_mm;
4981 BUG_ON(cpu_online(smp_processor_id()));
4984 switch_mm(mm, &init_mm, current);
4989 * While a dead CPU has no uninterruptible tasks queued at this point,
4990 * it might still have a nonzero ->nr_uninterruptible counter, because
4991 * for performance reasons the counter is not stricly tracking tasks to
4992 * their home CPUs. So we just add the counter to another CPU's counter,
4993 * to keep the global sum constant after CPU-down:
4995 static void migrate_nr_uninterruptible(struct rq *rq_src)
4997 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
4999 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5000 rq_src->nr_uninterruptible = 0;
5004 * remove the tasks which were accounted by rq from calc_load_tasks.
5006 static void calc_global_load_remove(struct rq *rq)
5008 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5009 rq->calc_load_active = 0;
5013 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5014 * try_to_wake_up()->select_task_rq().
5016 * Called with rq->lock held even though we'er in stop_machine() and
5017 * there's no concurrency possible, we hold the required locks anyway
5018 * because of lock validation efforts.
5020 static void migrate_tasks(unsigned int dead_cpu)
5022 struct rq *rq = cpu_rq(dead_cpu);
5023 struct task_struct *next, *stop = rq->stop;
5027 * Fudge the rq selection such that the below task selection loop
5028 * doesn't get stuck on the currently eligible stop task.
5030 * We're currently inside stop_machine() and the rq is either stuck
5031 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5032 * either way we should never end up calling schedule() until we're
5037 /* Ensure any throttled groups are reachable by pick_next_task */
5038 unthrottle_offline_cfs_rqs(rq);
5042 * There's this thread running, bail when that's the only
5045 if (rq->nr_running == 1)
5048 next = pick_next_task(rq);
5050 next->sched_class->put_prev_task(rq, next);
5052 /* Find suitable destination for @next, with force if needed. */
5053 dest_cpu = select_fallback_rq(dead_cpu, next);
5054 raw_spin_unlock(&rq->lock);
5056 __migrate_task(next, dead_cpu, dest_cpu);
5058 raw_spin_lock(&rq->lock);
5064 #endif /* CONFIG_HOTPLUG_CPU */
5066 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5068 static struct ctl_table sd_ctl_dir[] = {
5070 .procname = "sched_domain",
5076 static struct ctl_table sd_ctl_root[] = {
5078 .procname = "kernel",
5080 .child = sd_ctl_dir,
5085 static struct ctl_table *sd_alloc_ctl_entry(int n)
5087 struct ctl_table *entry =
5088 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5093 static void sd_free_ctl_entry(struct ctl_table **tablep)
5095 struct ctl_table *entry;
5098 * In the intermediate directories, both the child directory and
5099 * procname are dynamically allocated and could fail but the mode
5100 * will always be set. In the lowest directory the names are
5101 * static strings and all have proc handlers.
5103 for (entry = *tablep; entry->mode; entry++) {
5105 sd_free_ctl_entry(&entry->child);
5106 if (entry->proc_handler == NULL)
5107 kfree(entry->procname);
5115 set_table_entry(struct ctl_table *entry,
5116 const char *procname, void *data, int maxlen,
5117 mode_t mode, proc_handler *proc_handler)
5119 entry->procname = procname;
5121 entry->maxlen = maxlen;
5123 entry->proc_handler = proc_handler;
5126 static struct ctl_table *
5127 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5129 struct ctl_table *table = sd_alloc_ctl_entry(13);
5134 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5135 sizeof(long), 0644, proc_doulongvec_minmax);
5136 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5137 sizeof(long), 0644, proc_doulongvec_minmax);
5138 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5139 sizeof(int), 0644, proc_dointvec_minmax);
5140 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5141 sizeof(int), 0644, proc_dointvec_minmax);
5142 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5143 sizeof(int), 0644, proc_dointvec_minmax);
5144 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5145 sizeof(int), 0644, proc_dointvec_minmax);
5146 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5147 sizeof(int), 0644, proc_dointvec_minmax);
5148 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5149 sizeof(int), 0644, proc_dointvec_minmax);
5150 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5151 sizeof(int), 0644, proc_dointvec_minmax);
5152 set_table_entry(&table[9], "cache_nice_tries",
5153 &sd->cache_nice_tries,
5154 sizeof(int), 0644, proc_dointvec_minmax);
5155 set_table_entry(&table[10], "flags", &sd->flags,
5156 sizeof(int), 0644, proc_dointvec_minmax);
5157 set_table_entry(&table[11], "name", sd->name,
5158 CORENAME_MAX_SIZE, 0444, proc_dostring);
5159 /* &table[12] is terminator */
5164 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5166 struct ctl_table *entry, *table;
5167 struct sched_domain *sd;
5168 int domain_num = 0, i;
5171 for_each_domain(cpu, sd)
5173 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5178 for_each_domain(cpu, sd) {
5179 snprintf(buf, 32, "domain%d", i);
5180 entry->procname = kstrdup(buf, GFP_KERNEL);
5182 entry->child = sd_alloc_ctl_domain_table(sd);
5189 static struct ctl_table_header *sd_sysctl_header;
5190 static void register_sched_domain_sysctl(void)
5192 int i, cpu_num = num_possible_cpus();
5193 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5196 WARN_ON(sd_ctl_dir[0].child);
5197 sd_ctl_dir[0].child = entry;
5202 for_each_possible_cpu(i) {
5203 snprintf(buf, 32, "cpu%d", i);
5204 entry->procname = kstrdup(buf, GFP_KERNEL);
5206 entry->child = sd_alloc_ctl_cpu_table(i);
5210 WARN_ON(sd_sysctl_header);
5211 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5214 /* may be called multiple times per register */
5215 static void unregister_sched_domain_sysctl(void)
5217 if (sd_sysctl_header)
5218 unregister_sysctl_table(sd_sysctl_header);
5219 sd_sysctl_header = NULL;
5220 if (sd_ctl_dir[0].child)
5221 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5224 static void register_sched_domain_sysctl(void)
5227 static void unregister_sched_domain_sysctl(void)
5232 static void set_rq_online(struct rq *rq)
5235 const struct sched_class *class;
5237 cpumask_set_cpu(rq->cpu, rq->rd->online);
5240 for_each_class(class) {
5241 if (class->rq_online)
5242 class->rq_online(rq);
5247 static void set_rq_offline(struct rq *rq)
5250 const struct sched_class *class;
5252 for_each_class(class) {
5253 if (class->rq_offline)
5254 class->rq_offline(rq);
5257 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5263 * migration_call - callback that gets triggered when a CPU is added.
5264 * Here we can start up the necessary migration thread for the new CPU.
5266 static int __cpuinit
5267 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5269 int cpu = (long)hcpu;
5270 unsigned long flags;
5271 struct rq *rq = cpu_rq(cpu);
5273 switch (action & ~CPU_TASKS_FROZEN) {
5275 case CPU_UP_PREPARE:
5276 rq->calc_load_update = calc_load_update;
5280 /* Update our root-domain */
5281 raw_spin_lock_irqsave(&rq->lock, flags);
5283 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5287 raw_spin_unlock_irqrestore(&rq->lock, flags);
5290 #ifdef CONFIG_HOTPLUG_CPU
5292 sched_ttwu_pending();
5293 /* Update our root-domain */
5294 raw_spin_lock_irqsave(&rq->lock, flags);
5296 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5300 BUG_ON(rq->nr_running != 1); /* the migration thread */
5301 raw_spin_unlock_irqrestore(&rq->lock, flags);
5303 migrate_nr_uninterruptible(rq);
5304 calc_global_load_remove(rq);
5309 update_max_interval();
5315 * Register at high priority so that task migration (migrate_all_tasks)
5316 * happens before everything else. This has to be lower priority than
5317 * the notifier in the perf_event subsystem, though.
5319 static struct notifier_block __cpuinitdata migration_notifier = {
5320 .notifier_call = migration_call,
5321 .priority = CPU_PRI_MIGRATION,
5324 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5325 unsigned long action, void *hcpu)
5327 switch (action & ~CPU_TASKS_FROZEN) {
5329 case CPU_DOWN_FAILED:
5330 set_cpu_active((long)hcpu, true);
5337 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5338 unsigned long action, void *hcpu)
5340 switch (action & ~CPU_TASKS_FROZEN) {
5341 case CPU_DOWN_PREPARE:
5342 set_cpu_active((long)hcpu, false);
5349 static int __init migration_init(void)
5351 void *cpu = (void *)(long)smp_processor_id();
5354 /* Initialize migration for the boot CPU */
5355 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5356 BUG_ON(err == NOTIFY_BAD);
5357 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5358 register_cpu_notifier(&migration_notifier);
5360 /* Register cpu active notifiers */
5361 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5362 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5366 early_initcall(migration_init);
5371 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5373 #ifdef CONFIG_SCHED_DEBUG
5375 static __read_mostly int sched_domain_debug_enabled;
5377 static int __init sched_domain_debug_setup(char *str)
5379 sched_domain_debug_enabled = 1;
5383 early_param("sched_debug", sched_domain_debug_setup);
5385 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5386 struct cpumask *groupmask)
5388 struct sched_group *group = sd->groups;
5391 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5392 cpumask_clear(groupmask);
5394 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5396 if (!(sd->flags & SD_LOAD_BALANCE)) {
5397 printk("does not load-balance\n");
5399 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5404 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5406 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5407 printk(KERN_ERR "ERROR: domain->span does not contain "
5410 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5411 printk(KERN_ERR "ERROR: domain->groups does not contain"
5415 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5419 printk(KERN_ERR "ERROR: group is NULL\n");
5423 if (!group->sgp->power) {
5424 printk(KERN_CONT "\n");
5425 printk(KERN_ERR "ERROR: domain->cpu_power not "
5430 if (!cpumask_weight(sched_group_cpus(group))) {
5431 printk(KERN_CONT "\n");
5432 printk(KERN_ERR "ERROR: empty group\n");
5436 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5437 printk(KERN_CONT "\n");
5438 printk(KERN_ERR "ERROR: repeated CPUs\n");
5442 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5444 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5446 printk(KERN_CONT " %s", str);
5447 if (group->sgp->power != SCHED_POWER_SCALE) {
5448 printk(KERN_CONT " (cpu_power = %d)",
5452 group = group->next;
5453 } while (group != sd->groups);
5454 printk(KERN_CONT "\n");
5456 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5457 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5460 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5461 printk(KERN_ERR "ERROR: parent span is not a superset "
5462 "of domain->span\n");
5466 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5470 if (!sched_domain_debug_enabled)
5474 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5478 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5481 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5489 #else /* !CONFIG_SCHED_DEBUG */
5490 # define sched_domain_debug(sd, cpu) do { } while (0)
5491 #endif /* CONFIG_SCHED_DEBUG */
5493 static int sd_degenerate(struct sched_domain *sd)
5495 if (cpumask_weight(sched_domain_span(sd)) == 1)
5498 /* Following flags need at least 2 groups */
5499 if (sd->flags & (SD_LOAD_BALANCE |
5500 SD_BALANCE_NEWIDLE |
5504 SD_SHARE_PKG_RESOURCES)) {
5505 if (sd->groups != sd->groups->next)
5509 /* Following flags don't use groups */
5510 if (sd->flags & (SD_WAKE_AFFINE))
5517 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5519 unsigned long cflags = sd->flags, pflags = parent->flags;
5521 if (sd_degenerate(parent))
5524 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5527 /* Flags needing groups don't count if only 1 group in parent */
5528 if (parent->groups == parent->groups->next) {
5529 pflags &= ~(SD_LOAD_BALANCE |
5530 SD_BALANCE_NEWIDLE |
5534 SD_SHARE_PKG_RESOURCES);
5535 if (nr_node_ids == 1)
5536 pflags &= ~SD_SERIALIZE;
5538 if (~cflags & pflags)
5544 static void free_rootdomain(struct rcu_head *rcu)
5546 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5548 cpupri_cleanup(&rd->cpupri);
5549 free_cpumask_var(rd->rto_mask);
5550 free_cpumask_var(rd->online);
5551 free_cpumask_var(rd->span);
5555 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5557 struct root_domain *old_rd = NULL;
5558 unsigned long flags;
5560 raw_spin_lock_irqsave(&rq->lock, flags);
5565 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5568 cpumask_clear_cpu(rq->cpu, old_rd->span);
5571 * If we dont want to free the old_rt yet then
5572 * set old_rd to NULL to skip the freeing later
5575 if (!atomic_dec_and_test(&old_rd->refcount))
5579 atomic_inc(&rd->refcount);
5582 cpumask_set_cpu(rq->cpu, rd->span);
5583 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5586 raw_spin_unlock_irqrestore(&rq->lock, flags);
5589 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5592 static int init_rootdomain(struct root_domain *rd)
5594 memset(rd, 0, sizeof(*rd));
5596 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5598 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5600 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5603 if (cpupri_init(&rd->cpupri) != 0)
5608 free_cpumask_var(rd->rto_mask);
5610 free_cpumask_var(rd->online);
5612 free_cpumask_var(rd->span);
5618 * By default the system creates a single root-domain with all cpus as
5619 * members (mimicking the global state we have today).
5621 struct root_domain def_root_domain;
5623 static void init_defrootdomain(void)
5625 init_rootdomain(&def_root_domain);
5627 atomic_set(&def_root_domain.refcount, 1);
5630 static struct root_domain *alloc_rootdomain(void)
5632 struct root_domain *rd;
5634 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5638 if (init_rootdomain(rd) != 0) {
5646 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5648 struct sched_group *tmp, *first;
5657 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5662 } while (sg != first);
5665 static void free_sched_domain(struct rcu_head *rcu)
5667 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5670 * If its an overlapping domain it has private groups, iterate and
5673 if (sd->flags & SD_OVERLAP) {
5674 free_sched_groups(sd->groups, 1);
5675 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5676 kfree(sd->groups->sgp);
5682 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5684 call_rcu(&sd->rcu, free_sched_domain);
5687 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5689 for (; sd; sd = sd->parent)
5690 destroy_sched_domain(sd, cpu);
5694 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5695 * hold the hotplug lock.
5698 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5700 struct rq *rq = cpu_rq(cpu);
5701 struct sched_domain *tmp;
5703 /* Remove the sched domains which do not contribute to scheduling. */
5704 for (tmp = sd; tmp; ) {
5705 struct sched_domain *parent = tmp->parent;
5709 if (sd_parent_degenerate(tmp, parent)) {
5710 tmp->parent = parent->parent;
5712 parent->parent->child = tmp;
5713 destroy_sched_domain(parent, cpu);
5718 if (sd && sd_degenerate(sd)) {
5721 destroy_sched_domain(tmp, cpu);
5726 sched_domain_debug(sd, cpu);
5728 rq_attach_root(rq, rd);
5730 rcu_assign_pointer(rq->sd, sd);
5731 destroy_sched_domains(tmp, cpu);
5734 /* cpus with isolated domains */
5735 static cpumask_var_t cpu_isolated_map;
5737 /* Setup the mask of cpus configured for isolated domains */
5738 static int __init isolated_cpu_setup(char *str)
5740 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5741 cpulist_parse(str, cpu_isolated_map);
5745 __setup("isolcpus=", isolated_cpu_setup);
5750 * find_next_best_node - find the next node to include in a sched_domain
5751 * @node: node whose sched_domain we're building
5752 * @used_nodes: nodes already in the sched_domain
5754 * Find the next node to include in a given scheduling domain. Simply
5755 * finds the closest node not already in the @used_nodes map.
5757 * Should use nodemask_t.
5759 static int find_next_best_node(int node, nodemask_t *used_nodes)
5761 int i, n, val, min_val, best_node = -1;
5765 for (i = 0; i < nr_node_ids; i++) {
5766 /* Start at @node */
5767 n = (node + i) % nr_node_ids;
5769 if (!nr_cpus_node(n))
5772 /* Skip already used nodes */
5773 if (node_isset(n, *used_nodes))
5776 /* Simple min distance search */
5777 val = node_distance(node, n);
5779 if (val < min_val) {
5785 if (best_node != -1)
5786 node_set(best_node, *used_nodes);
5791 * sched_domain_node_span - get a cpumask for a node's sched_domain
5792 * @node: node whose cpumask we're constructing
5793 * @span: resulting cpumask
5795 * Given a node, construct a good cpumask for its sched_domain to span. It
5796 * should be one that prevents unnecessary balancing, but also spreads tasks
5799 static void sched_domain_node_span(int node, struct cpumask *span)
5801 nodemask_t used_nodes;
5804 cpumask_clear(span);
5805 nodes_clear(used_nodes);
5807 cpumask_or(span, span, cpumask_of_node(node));
5808 node_set(node, used_nodes);
5810 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5811 int next_node = find_next_best_node(node, &used_nodes);
5814 cpumask_or(span, span, cpumask_of_node(next_node));
5818 static const struct cpumask *cpu_node_mask(int cpu)
5820 lockdep_assert_held(&sched_domains_mutex);
5822 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5824 return sched_domains_tmpmask;
5827 static const struct cpumask *cpu_allnodes_mask(int cpu)
5829 return cpu_possible_mask;
5831 #endif /* CONFIG_NUMA */
5833 static const struct cpumask *cpu_cpu_mask(int cpu)
5835 return cpumask_of_node(cpu_to_node(cpu));
5838 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5841 struct sched_domain **__percpu sd;
5842 struct sched_group **__percpu sg;
5843 struct sched_group_power **__percpu sgp;
5847 struct sched_domain ** __percpu sd;
5848 struct root_domain *rd;
5858 struct sched_domain_topology_level;
5860 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5861 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5863 #define SDTL_OVERLAP 0x01
5865 struct sched_domain_topology_level {
5866 sched_domain_init_f init;
5867 sched_domain_mask_f mask;
5869 struct sd_data data;
5873 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5875 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5876 const struct cpumask *span = sched_domain_span(sd);
5877 struct cpumask *covered = sched_domains_tmpmask;
5878 struct sd_data *sdd = sd->private;
5879 struct sched_domain *child;
5882 cpumask_clear(covered);
5884 for_each_cpu(i, span) {
5885 struct cpumask *sg_span;
5887 if (cpumask_test_cpu(i, covered))
5890 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5891 GFP_KERNEL, cpu_to_node(cpu));
5896 sg_span = sched_group_cpus(sg);
5898 child = *per_cpu_ptr(sdd->sd, i);
5900 child = child->child;
5901 cpumask_copy(sg_span, sched_domain_span(child));
5903 cpumask_set_cpu(i, sg_span);
5905 cpumask_or(covered, covered, sg_span);
5907 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5908 atomic_inc(&sg->sgp->ref);
5910 if (cpumask_test_cpu(cpu, sg_span))
5920 sd->groups = groups;
5925 free_sched_groups(first, 0);
5930 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5932 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5933 struct sched_domain *child = sd->child;
5936 cpu = cpumask_first(sched_domain_span(child));
5939 *sg = *per_cpu_ptr(sdd->sg, cpu);
5940 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5941 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5948 * build_sched_groups will build a circular linked list of the groups
5949 * covered by the given span, and will set each group's ->cpumask correctly,
5950 * and ->cpu_power to 0.
5952 * Assumes the sched_domain tree is fully constructed
5955 build_sched_groups(struct sched_domain *sd, int cpu)
5957 struct sched_group *first = NULL, *last = NULL;
5958 struct sd_data *sdd = sd->private;
5959 const struct cpumask *span = sched_domain_span(sd);
5960 struct cpumask *covered;
5963 get_group(cpu, sdd, &sd->groups);
5964 atomic_inc(&sd->groups->ref);
5966 if (cpu != cpumask_first(sched_domain_span(sd)))
5969 lockdep_assert_held(&sched_domains_mutex);
5970 covered = sched_domains_tmpmask;
5972 cpumask_clear(covered);
5974 for_each_cpu(i, span) {
5975 struct sched_group *sg;
5976 int group = get_group(i, sdd, &sg);
5979 if (cpumask_test_cpu(i, covered))
5982 cpumask_clear(sched_group_cpus(sg));
5985 for_each_cpu(j, span) {
5986 if (get_group(j, sdd, NULL) != group)
5989 cpumask_set_cpu(j, covered);
5990 cpumask_set_cpu(j, sched_group_cpus(sg));
6005 * Initialize sched groups cpu_power.
6007 * cpu_power indicates the capacity of sched group, which is used while
6008 * distributing the load between different sched groups in a sched domain.
6009 * Typically cpu_power for all the groups in a sched domain will be same unless
6010 * there are asymmetries in the topology. If there are asymmetries, group
6011 * having more cpu_power will pickup more load compared to the group having
6014 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6016 struct sched_group *sg = sd->groups;
6018 WARN_ON(!sd || !sg);
6021 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6023 } while (sg != sd->groups);
6025 if (cpu != group_first_cpu(sg))
6028 update_group_power(sd, cpu);
6029 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6032 int __weak arch_sd_sibling_asym_packing(void)
6034 return 0*SD_ASYM_PACKING;
6038 * Initializers for schedule domains
6039 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6042 #ifdef CONFIG_SCHED_DEBUG
6043 # define SD_INIT_NAME(sd, type) sd->name = #type
6045 # define SD_INIT_NAME(sd, type) do { } while (0)
6048 #define SD_INIT_FUNC(type) \
6049 static noinline struct sched_domain * \
6050 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6052 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6053 *sd = SD_##type##_INIT; \
6054 SD_INIT_NAME(sd, type); \
6055 sd->private = &tl->data; \
6061 SD_INIT_FUNC(ALLNODES)
6064 #ifdef CONFIG_SCHED_SMT
6065 SD_INIT_FUNC(SIBLING)
6067 #ifdef CONFIG_SCHED_MC
6070 #ifdef CONFIG_SCHED_BOOK
6074 static int default_relax_domain_level = -1;
6075 int sched_domain_level_max;
6077 static int __init setup_relax_domain_level(char *str)
6081 val = simple_strtoul(str, NULL, 0);
6082 if (val < sched_domain_level_max)
6083 default_relax_domain_level = val;
6087 __setup("relax_domain_level=", setup_relax_domain_level);
6089 static void set_domain_attribute(struct sched_domain *sd,
6090 struct sched_domain_attr *attr)
6094 if (!attr || attr->relax_domain_level < 0) {
6095 if (default_relax_domain_level < 0)
6098 request = default_relax_domain_level;
6100 request = attr->relax_domain_level;
6101 if (request < sd->level) {
6102 /* turn off idle balance on this domain */
6103 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6105 /* turn on idle balance on this domain */
6106 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6110 static void __sdt_free(const struct cpumask *cpu_map);
6111 static int __sdt_alloc(const struct cpumask *cpu_map);
6113 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6114 const struct cpumask *cpu_map)
6118 if (!atomic_read(&d->rd->refcount))
6119 free_rootdomain(&d->rd->rcu); /* fall through */
6121 free_percpu(d->sd); /* fall through */
6123 __sdt_free(cpu_map); /* fall through */
6129 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6130 const struct cpumask *cpu_map)
6132 memset(d, 0, sizeof(*d));
6134 if (__sdt_alloc(cpu_map))
6135 return sa_sd_storage;
6136 d->sd = alloc_percpu(struct sched_domain *);
6138 return sa_sd_storage;
6139 d->rd = alloc_rootdomain();
6142 return sa_rootdomain;
6146 * NULL the sd_data elements we've used to build the sched_domain and
6147 * sched_group structure so that the subsequent __free_domain_allocs()
6148 * will not free the data we're using.
6150 static void claim_allocations(int cpu, struct sched_domain *sd)
6152 struct sd_data *sdd = sd->private;
6154 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6155 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6157 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6158 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6160 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6161 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6164 #ifdef CONFIG_SCHED_SMT
6165 static const struct cpumask *cpu_smt_mask(int cpu)
6167 return topology_thread_cpumask(cpu);
6172 * Topology list, bottom-up.
6174 static struct sched_domain_topology_level default_topology[] = {
6175 #ifdef CONFIG_SCHED_SMT
6176 { sd_init_SIBLING, cpu_smt_mask, },
6178 #ifdef CONFIG_SCHED_MC
6179 { sd_init_MC, cpu_coregroup_mask, },
6181 #ifdef CONFIG_SCHED_BOOK
6182 { sd_init_BOOK, cpu_book_mask, },
6184 { sd_init_CPU, cpu_cpu_mask, },
6186 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6187 { sd_init_ALLNODES, cpu_allnodes_mask, },
6192 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6194 static int __sdt_alloc(const struct cpumask *cpu_map)
6196 struct sched_domain_topology_level *tl;
6199 for (tl = sched_domain_topology; tl->init; tl++) {
6200 struct sd_data *sdd = &tl->data;
6202 sdd->sd = alloc_percpu(struct sched_domain *);
6206 sdd->sg = alloc_percpu(struct sched_group *);
6210 sdd->sgp = alloc_percpu(struct sched_group_power *);
6214 for_each_cpu(j, cpu_map) {
6215 struct sched_domain *sd;
6216 struct sched_group *sg;
6217 struct sched_group_power *sgp;
6219 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6220 GFP_KERNEL, cpu_to_node(j));
6224 *per_cpu_ptr(sdd->sd, j) = sd;
6226 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6227 GFP_KERNEL, cpu_to_node(j));
6231 *per_cpu_ptr(sdd->sg, j) = sg;
6233 sgp = kzalloc_node(sizeof(struct sched_group_power),
6234 GFP_KERNEL, cpu_to_node(j));
6238 *per_cpu_ptr(sdd->sgp, j) = sgp;
6245 static void __sdt_free(const struct cpumask *cpu_map)
6247 struct sched_domain_topology_level *tl;
6250 for (tl = sched_domain_topology; tl->init; tl++) {
6251 struct sd_data *sdd = &tl->data;
6253 for_each_cpu(j, cpu_map) {
6254 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6255 if (sd && (sd->flags & SD_OVERLAP))
6256 free_sched_groups(sd->groups, 0);
6257 kfree(*per_cpu_ptr(sdd->sd, j));
6258 kfree(*per_cpu_ptr(sdd->sg, j));
6259 kfree(*per_cpu_ptr(sdd->sgp, j));
6261 free_percpu(sdd->sd);
6262 free_percpu(sdd->sg);
6263 free_percpu(sdd->sgp);
6267 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6268 struct s_data *d, const struct cpumask *cpu_map,
6269 struct sched_domain_attr *attr, struct sched_domain *child,
6272 struct sched_domain *sd = tl->init(tl, cpu);
6276 set_domain_attribute(sd, attr);
6277 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6279 sd->level = child->level + 1;
6280 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6289 * Build sched domains for a given set of cpus and attach the sched domains
6290 * to the individual cpus
6292 static int build_sched_domains(const struct cpumask *cpu_map,
6293 struct sched_domain_attr *attr)
6295 enum s_alloc alloc_state = sa_none;
6296 struct sched_domain *sd;
6298 int i, ret = -ENOMEM;
6300 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6301 if (alloc_state != sa_rootdomain)
6304 /* Set up domains for cpus specified by the cpu_map. */
6305 for_each_cpu(i, cpu_map) {
6306 struct sched_domain_topology_level *tl;
6309 for (tl = sched_domain_topology; tl->init; tl++) {
6310 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6311 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6312 sd->flags |= SD_OVERLAP;
6313 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6320 *per_cpu_ptr(d.sd, i) = sd;
6323 /* Build the groups for the domains */
6324 for_each_cpu(i, cpu_map) {
6325 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6326 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6327 if (sd->flags & SD_OVERLAP) {
6328 if (build_overlap_sched_groups(sd, i))
6331 if (build_sched_groups(sd, i))
6337 /* Calculate CPU power for physical packages and nodes */
6338 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6339 if (!cpumask_test_cpu(i, cpu_map))
6342 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6343 claim_allocations(i, sd);
6344 init_sched_groups_power(i, sd);
6348 /* Attach the domains */
6350 for_each_cpu(i, cpu_map) {
6351 sd = *per_cpu_ptr(d.sd, i);
6352 cpu_attach_domain(sd, d.rd, i);
6358 __free_domain_allocs(&d, alloc_state, cpu_map);
6362 static cpumask_var_t *doms_cur; /* current sched domains */
6363 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6364 static struct sched_domain_attr *dattr_cur;
6365 /* attribues of custom domains in 'doms_cur' */
6368 * Special case: If a kmalloc of a doms_cur partition (array of
6369 * cpumask) fails, then fallback to a single sched domain,
6370 * as determined by the single cpumask fallback_doms.
6372 static cpumask_var_t fallback_doms;
6375 * arch_update_cpu_topology lets virtualized architectures update the
6376 * cpu core maps. It is supposed to return 1 if the topology changed
6377 * or 0 if it stayed the same.
6379 int __attribute__((weak)) arch_update_cpu_topology(void)
6384 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6387 cpumask_var_t *doms;
6389 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6392 for (i = 0; i < ndoms; i++) {
6393 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6394 free_sched_domains(doms, i);
6401 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6404 for (i = 0; i < ndoms; i++)
6405 free_cpumask_var(doms[i]);
6410 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6411 * For now this just excludes isolated cpus, but could be used to
6412 * exclude other special cases in the future.
6414 static int init_sched_domains(const struct cpumask *cpu_map)
6418 arch_update_cpu_topology();
6420 doms_cur = alloc_sched_domains(ndoms_cur);
6422 doms_cur = &fallback_doms;
6423 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6425 err = build_sched_domains(doms_cur[0], NULL);
6426 register_sched_domain_sysctl();
6432 * Detach sched domains from a group of cpus specified in cpu_map
6433 * These cpus will now be attached to the NULL domain
6435 static void detach_destroy_domains(const struct cpumask *cpu_map)
6440 for_each_cpu(i, cpu_map)
6441 cpu_attach_domain(NULL, &def_root_domain, i);
6445 /* handle null as "default" */
6446 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6447 struct sched_domain_attr *new, int idx_new)
6449 struct sched_domain_attr tmp;
6456 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6457 new ? (new + idx_new) : &tmp,
6458 sizeof(struct sched_domain_attr));
6462 * Partition sched domains as specified by the 'ndoms_new'
6463 * cpumasks in the array doms_new[] of cpumasks. This compares
6464 * doms_new[] to the current sched domain partitioning, doms_cur[].
6465 * It destroys each deleted domain and builds each new domain.
6467 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6468 * The masks don't intersect (don't overlap.) We should setup one
6469 * sched domain for each mask. CPUs not in any of the cpumasks will
6470 * not be load balanced. If the same cpumask appears both in the
6471 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6474 * The passed in 'doms_new' should be allocated using
6475 * alloc_sched_domains. This routine takes ownership of it and will
6476 * free_sched_domains it when done with it. If the caller failed the
6477 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6478 * and partition_sched_domains() will fallback to the single partition
6479 * 'fallback_doms', it also forces the domains to be rebuilt.
6481 * If doms_new == NULL it will be replaced with cpu_online_mask.
6482 * ndoms_new == 0 is a special case for destroying existing domains,
6483 * and it will not create the default domain.
6485 * Call with hotplug lock held
6487 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6488 struct sched_domain_attr *dattr_new)
6493 mutex_lock(&sched_domains_mutex);
6495 /* always unregister in case we don't destroy any domains */
6496 unregister_sched_domain_sysctl();
6498 /* Let architecture update cpu core mappings. */
6499 new_topology = arch_update_cpu_topology();
6501 n = doms_new ? ndoms_new : 0;
6503 /* Destroy deleted domains */
6504 for (i = 0; i < ndoms_cur; i++) {
6505 for (j = 0; j < n && !new_topology; j++) {
6506 if (cpumask_equal(doms_cur[i], doms_new[j])
6507 && dattrs_equal(dattr_cur, i, dattr_new, j))
6510 /* no match - a current sched domain not in new doms_new[] */
6511 detach_destroy_domains(doms_cur[i]);
6516 if (doms_new == NULL) {
6518 doms_new = &fallback_doms;
6519 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6520 WARN_ON_ONCE(dattr_new);
6523 /* Build new domains */
6524 for (i = 0; i < ndoms_new; i++) {
6525 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6526 if (cpumask_equal(doms_new[i], doms_cur[j])
6527 && dattrs_equal(dattr_new, i, dattr_cur, j))
6530 /* no match - add a new doms_new */
6531 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6536 /* Remember the new sched domains */
6537 if (doms_cur != &fallback_doms)
6538 free_sched_domains(doms_cur, ndoms_cur);
6539 kfree(dattr_cur); /* kfree(NULL) is safe */
6540 doms_cur = doms_new;
6541 dattr_cur = dattr_new;
6542 ndoms_cur = ndoms_new;
6544 register_sched_domain_sysctl();
6546 mutex_unlock(&sched_domains_mutex);
6549 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6550 static void reinit_sched_domains(void)
6554 /* Destroy domains first to force the rebuild */
6555 partition_sched_domains(0, NULL, NULL);
6557 rebuild_sched_domains();
6561 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6563 unsigned int level = 0;
6565 if (sscanf(buf, "%u", &level) != 1)
6569 * level is always be positive so don't check for
6570 * level < POWERSAVINGS_BALANCE_NONE which is 0
6571 * What happens on 0 or 1 byte write,
6572 * need to check for count as well?
6575 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6579 sched_smt_power_savings = level;
6581 sched_mc_power_savings = level;
6583 reinit_sched_domains();
6588 #ifdef CONFIG_SCHED_MC
6589 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
6590 struct sysdev_class_attribute *attr,
6593 return sprintf(page, "%u\n", sched_mc_power_savings);
6595 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
6596 struct sysdev_class_attribute *attr,
6597 const char *buf, size_t count)
6599 return sched_power_savings_store(buf, count, 0);
6601 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
6602 sched_mc_power_savings_show,
6603 sched_mc_power_savings_store);
6606 #ifdef CONFIG_SCHED_SMT
6607 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
6608 struct sysdev_class_attribute *attr,
6611 return sprintf(page, "%u\n", sched_smt_power_savings);
6613 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
6614 struct sysdev_class_attribute *attr,
6615 const char *buf, size_t count)
6617 return sched_power_savings_store(buf, count, 1);
6619 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
6620 sched_smt_power_savings_show,
6621 sched_smt_power_savings_store);
6624 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
6628 #ifdef CONFIG_SCHED_SMT
6630 err = sysfs_create_file(&cls->kset.kobj,
6631 &attr_sched_smt_power_savings.attr);
6633 #ifdef CONFIG_SCHED_MC
6634 if (!err && mc_capable())
6635 err = sysfs_create_file(&cls->kset.kobj,
6636 &attr_sched_mc_power_savings.attr);
6640 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6643 * Update cpusets according to cpu_active mask. If cpusets are
6644 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6645 * around partition_sched_domains().
6647 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6650 switch (action & ~CPU_TASKS_FROZEN) {
6652 case CPU_DOWN_FAILED:
6653 cpuset_update_active_cpus();
6660 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6663 switch (action & ~CPU_TASKS_FROZEN) {
6664 case CPU_DOWN_PREPARE:
6665 cpuset_update_active_cpus();
6672 void __init sched_init_smp(void)
6674 cpumask_var_t non_isolated_cpus;
6676 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6677 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6680 mutex_lock(&sched_domains_mutex);
6681 init_sched_domains(cpu_active_mask);
6682 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6683 if (cpumask_empty(non_isolated_cpus))
6684 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6685 mutex_unlock(&sched_domains_mutex);
6688 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6689 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6691 /* RT runtime code needs to handle some hotplug events */
6692 hotcpu_notifier(update_runtime, 0);
6696 /* Move init over to a non-isolated CPU */
6697 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6699 sched_init_granularity();
6700 free_cpumask_var(non_isolated_cpus);
6702 init_sched_rt_class();
6705 void __init sched_init_smp(void)
6707 sched_init_granularity();
6709 #endif /* CONFIG_SMP */
6711 const_debug unsigned int sysctl_timer_migration = 1;
6713 int in_sched_functions(unsigned long addr)
6715 return in_lock_functions(addr) ||
6716 (addr >= (unsigned long)__sched_text_start
6717 && addr < (unsigned long)__sched_text_end);
6720 #ifdef CONFIG_CGROUP_SCHED
6721 struct task_group root_task_group;
6724 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6726 void __init sched_init(void)
6729 unsigned long alloc_size = 0, ptr;
6731 #ifdef CONFIG_FAIR_GROUP_SCHED
6732 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6734 #ifdef CONFIG_RT_GROUP_SCHED
6735 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6737 #ifdef CONFIG_CPUMASK_OFFSTACK
6738 alloc_size += num_possible_cpus() * cpumask_size();
6741 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6743 #ifdef CONFIG_FAIR_GROUP_SCHED
6744 root_task_group.se = (struct sched_entity **)ptr;
6745 ptr += nr_cpu_ids * sizeof(void **);
6747 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6748 ptr += nr_cpu_ids * sizeof(void **);
6750 #endif /* CONFIG_FAIR_GROUP_SCHED */
6751 #ifdef CONFIG_RT_GROUP_SCHED
6752 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6753 ptr += nr_cpu_ids * sizeof(void **);
6755 root_task_group.rt_rq = (struct rt_rq **)ptr;
6756 ptr += nr_cpu_ids * sizeof(void **);
6758 #endif /* CONFIG_RT_GROUP_SCHED */
6759 #ifdef CONFIG_CPUMASK_OFFSTACK
6760 for_each_possible_cpu(i) {
6761 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6762 ptr += cpumask_size();
6764 #endif /* CONFIG_CPUMASK_OFFSTACK */
6768 init_defrootdomain();
6771 init_rt_bandwidth(&def_rt_bandwidth,
6772 global_rt_period(), global_rt_runtime());
6774 #ifdef CONFIG_RT_GROUP_SCHED
6775 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6776 global_rt_period(), global_rt_runtime());
6777 #endif /* CONFIG_RT_GROUP_SCHED */
6779 #ifdef CONFIG_CGROUP_SCHED
6780 list_add(&root_task_group.list, &task_groups);
6781 INIT_LIST_HEAD(&root_task_group.children);
6782 INIT_LIST_HEAD(&root_task_group.siblings);
6783 autogroup_init(&init_task);
6784 #endif /* CONFIG_CGROUP_SCHED */
6786 for_each_possible_cpu(i) {
6790 raw_spin_lock_init(&rq->lock);
6792 rq->calc_load_active = 0;
6793 rq->calc_load_update = jiffies + LOAD_FREQ;
6794 init_cfs_rq(&rq->cfs);
6795 init_rt_rq(&rq->rt, rq);
6796 #ifdef CONFIG_FAIR_GROUP_SCHED
6797 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6798 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6800 * How much cpu bandwidth does root_task_group get?
6802 * In case of task-groups formed thr' the cgroup filesystem, it
6803 * gets 100% of the cpu resources in the system. This overall
6804 * system cpu resource is divided among the tasks of
6805 * root_task_group and its child task-groups in a fair manner,
6806 * based on each entity's (task or task-group's) weight
6807 * (se->load.weight).
6809 * In other words, if root_task_group has 10 tasks of weight
6810 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6811 * then A0's share of the cpu resource is:
6813 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6815 * We achieve this by letting root_task_group's tasks sit
6816 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6818 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6819 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6820 #endif /* CONFIG_FAIR_GROUP_SCHED */
6822 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6823 #ifdef CONFIG_RT_GROUP_SCHED
6824 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6825 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6828 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6829 rq->cpu_load[j] = 0;
6831 rq->last_load_update_tick = jiffies;
6836 rq->cpu_power = SCHED_POWER_SCALE;
6837 rq->post_schedule = 0;
6838 rq->active_balance = 0;
6839 rq->next_balance = jiffies;
6844 rq->avg_idle = 2*sysctl_sched_migration_cost;
6845 rq_attach_root(rq, &def_root_domain);
6851 atomic_set(&rq->nr_iowait, 0);
6854 set_load_weight(&init_task);
6856 #ifdef CONFIG_PREEMPT_NOTIFIERS
6857 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6860 #ifdef CONFIG_RT_MUTEXES
6861 plist_head_init(&init_task.pi_waiters);
6865 * The boot idle thread does lazy MMU switching as well:
6867 atomic_inc(&init_mm.mm_count);
6868 enter_lazy_tlb(&init_mm, current);
6871 * Make us the idle thread. Technically, schedule() should not be
6872 * called from this thread, however somewhere below it might be,
6873 * but because we are the idle thread, we just pick up running again
6874 * when this runqueue becomes "idle".
6876 init_idle(current, smp_processor_id());
6878 calc_load_update = jiffies + LOAD_FREQ;
6881 * During early bootup we pretend to be a normal task:
6883 current->sched_class = &fair_sched_class;
6886 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6887 /* May be allocated at isolcpus cmdline parse time */
6888 if (cpu_isolated_map == NULL)
6889 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6891 init_sched_fair_class();
6893 scheduler_running = 1;
6896 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6897 static inline int preempt_count_equals(int preempt_offset)
6899 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6901 return (nested == preempt_offset);
6904 void __might_sleep(const char *file, int line, int preempt_offset)
6906 static unsigned long prev_jiffy; /* ratelimiting */
6908 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6909 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6910 system_state != SYSTEM_RUNNING || oops_in_progress)
6912 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6914 prev_jiffy = jiffies;
6917 "BUG: sleeping function called from invalid context at %s:%d\n",
6920 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6921 in_atomic(), irqs_disabled(),
6922 current->pid, current->comm);
6924 debug_show_held_locks(current);
6925 if (irqs_disabled())
6926 print_irqtrace_events(current);
6929 EXPORT_SYMBOL(__might_sleep);
6932 #ifdef CONFIG_MAGIC_SYSRQ
6933 static void normalize_task(struct rq *rq, struct task_struct *p)
6935 const struct sched_class *prev_class = p->sched_class;
6936 int old_prio = p->prio;
6941 deactivate_task(rq, p, 0);
6942 __setscheduler(rq, p, SCHED_NORMAL, 0);
6944 activate_task(rq, p, 0);
6945 resched_task(rq->curr);
6948 check_class_changed(rq, p, prev_class, old_prio);
6951 void normalize_rt_tasks(void)
6953 struct task_struct *g, *p;
6954 unsigned long flags;
6957 read_lock_irqsave(&tasklist_lock, flags);
6958 do_each_thread(g, p) {
6960 * Only normalize user tasks:
6965 p->se.exec_start = 0;
6966 #ifdef CONFIG_SCHEDSTATS
6967 p->se.statistics.wait_start = 0;
6968 p->se.statistics.sleep_start = 0;
6969 p->se.statistics.block_start = 0;
6974 * Renice negative nice level userspace
6977 if (TASK_NICE(p) < 0 && p->mm)
6978 set_user_nice(p, 0);
6982 raw_spin_lock(&p->pi_lock);
6983 rq = __task_rq_lock(p);
6985 normalize_task(rq, p);
6987 __task_rq_unlock(rq);
6988 raw_spin_unlock(&p->pi_lock);
6989 } while_each_thread(g, p);
6991 read_unlock_irqrestore(&tasklist_lock, flags);
6994 #endif /* CONFIG_MAGIC_SYSRQ */
6996 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6998 * These functions are only useful for the IA64 MCA handling, or kdb.
7000 * They can only be called when the whole system has been
7001 * stopped - every CPU needs to be quiescent, and no scheduling
7002 * activity can take place. Using them for anything else would
7003 * be a serious bug, and as a result, they aren't even visible
7004 * under any other configuration.
7008 * curr_task - return the current task for a given cpu.
7009 * @cpu: the processor in question.
7011 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7013 struct task_struct *curr_task(int cpu)
7015 return cpu_curr(cpu);
7018 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7022 * set_curr_task - set the current task for a given cpu.
7023 * @cpu: the processor in question.
7024 * @p: the task pointer to set.
7026 * Description: This function must only be used when non-maskable interrupts
7027 * are serviced on a separate stack. It allows the architecture to switch the
7028 * notion of the current task on a cpu in a non-blocking manner. This function
7029 * must be called with all CPU's synchronized, and interrupts disabled, the
7030 * and caller must save the original value of the current task (see
7031 * curr_task() above) and restore that value before reenabling interrupts and
7032 * re-starting the system.
7034 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7036 void set_curr_task(int cpu, struct task_struct *p)
7043 #ifdef CONFIG_RT_GROUP_SCHED
7044 #else /* !CONFIG_RT_GROUP_SCHED */
7045 #endif /* CONFIG_RT_GROUP_SCHED */
7047 #ifdef CONFIG_CGROUP_SCHED
7048 /* task_group_lock serializes the addition/removal of task groups */
7049 static DEFINE_SPINLOCK(task_group_lock);
7051 static void free_sched_group(struct task_group *tg)
7053 free_fair_sched_group(tg);
7054 free_rt_sched_group(tg);
7059 /* allocate runqueue etc for a new task group */
7060 struct task_group *sched_create_group(struct task_group *parent)
7062 struct task_group *tg;
7063 unsigned long flags;
7065 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7067 return ERR_PTR(-ENOMEM);
7069 if (!alloc_fair_sched_group(tg, parent))
7072 if (!alloc_rt_sched_group(tg, parent))
7075 spin_lock_irqsave(&task_group_lock, flags);
7076 list_add_rcu(&tg->list, &task_groups);
7078 WARN_ON(!parent); /* root should already exist */
7080 tg->parent = parent;
7081 INIT_LIST_HEAD(&tg->children);
7082 list_add_rcu(&tg->siblings, &parent->children);
7083 spin_unlock_irqrestore(&task_group_lock, flags);
7088 free_sched_group(tg);
7089 return ERR_PTR(-ENOMEM);
7092 /* rcu callback to free various structures associated with a task group */
7093 static void free_sched_group_rcu(struct rcu_head *rhp)
7095 /* now it should be safe to free those cfs_rqs */
7096 free_sched_group(container_of(rhp, struct task_group, rcu));
7099 /* Destroy runqueue etc associated with a task group */
7100 void sched_destroy_group(struct task_group *tg)
7102 unsigned long flags;
7105 /* end participation in shares distribution */
7106 for_each_possible_cpu(i)
7107 unregister_fair_sched_group(tg, i);
7109 spin_lock_irqsave(&task_group_lock, flags);
7110 list_del_rcu(&tg->list);
7111 list_del_rcu(&tg->siblings);
7112 spin_unlock_irqrestore(&task_group_lock, flags);
7114 /* wait for possible concurrent references to cfs_rqs complete */
7115 call_rcu(&tg->rcu, free_sched_group_rcu);
7118 /* change task's runqueue when it moves between groups.
7119 * The caller of this function should have put the task in its new group
7120 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7121 * reflect its new group.
7123 void sched_move_task(struct task_struct *tsk)
7126 unsigned long flags;
7129 rq = task_rq_lock(tsk, &flags);
7131 running = task_current(rq, tsk);
7135 dequeue_task(rq, tsk, 0);
7136 if (unlikely(running))
7137 tsk->sched_class->put_prev_task(rq, tsk);
7139 #ifdef CONFIG_FAIR_GROUP_SCHED
7140 if (tsk->sched_class->task_move_group)
7141 tsk->sched_class->task_move_group(tsk, on_rq);
7144 set_task_rq(tsk, task_cpu(tsk));
7146 if (unlikely(running))
7147 tsk->sched_class->set_curr_task(rq);
7149 enqueue_task(rq, tsk, 0);
7151 task_rq_unlock(rq, tsk, &flags);
7153 #endif /* CONFIG_CGROUP_SCHED */
7155 #ifdef CONFIG_FAIR_GROUP_SCHED
7158 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7159 static unsigned long to_ratio(u64 period, u64 runtime)
7161 if (runtime == RUNTIME_INF)
7164 return div64_u64(runtime << 20, period);
7168 #ifdef CONFIG_RT_GROUP_SCHED
7170 * Ensure that the real time constraints are schedulable.
7172 static DEFINE_MUTEX(rt_constraints_mutex);
7174 /* Must be called with tasklist_lock held */
7175 static inline int tg_has_rt_tasks(struct task_group *tg)
7177 struct task_struct *g, *p;
7179 do_each_thread(g, p) {
7180 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7182 } while_each_thread(g, p);
7187 struct rt_schedulable_data {
7188 struct task_group *tg;
7193 static int tg_rt_schedulable(struct task_group *tg, void *data)
7195 struct rt_schedulable_data *d = data;
7196 struct task_group *child;
7197 unsigned long total, sum = 0;
7198 u64 period, runtime;
7200 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7201 runtime = tg->rt_bandwidth.rt_runtime;
7204 period = d->rt_period;
7205 runtime = d->rt_runtime;
7209 * Cannot have more runtime than the period.
7211 if (runtime > period && runtime != RUNTIME_INF)
7215 * Ensure we don't starve existing RT tasks.
7217 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7220 total = to_ratio(period, runtime);
7223 * Nobody can have more than the global setting allows.
7225 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7229 * The sum of our children's runtime should not exceed our own.
7231 list_for_each_entry_rcu(child, &tg->children, siblings) {
7232 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7233 runtime = child->rt_bandwidth.rt_runtime;
7235 if (child == d->tg) {
7236 period = d->rt_period;
7237 runtime = d->rt_runtime;
7240 sum += to_ratio(period, runtime);
7249 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7253 struct rt_schedulable_data data = {
7255 .rt_period = period,
7256 .rt_runtime = runtime,
7260 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7266 static int tg_set_rt_bandwidth(struct task_group *tg,
7267 u64 rt_period, u64 rt_runtime)
7271 mutex_lock(&rt_constraints_mutex);
7272 read_lock(&tasklist_lock);
7273 err = __rt_schedulable(tg, rt_period, rt_runtime);
7277 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7278 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7279 tg->rt_bandwidth.rt_runtime = rt_runtime;
7281 for_each_possible_cpu(i) {
7282 struct rt_rq *rt_rq = tg->rt_rq[i];
7284 raw_spin_lock(&rt_rq->rt_runtime_lock);
7285 rt_rq->rt_runtime = rt_runtime;
7286 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7288 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7290 read_unlock(&tasklist_lock);
7291 mutex_unlock(&rt_constraints_mutex);
7296 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7298 u64 rt_runtime, rt_period;
7300 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7301 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7302 if (rt_runtime_us < 0)
7303 rt_runtime = RUNTIME_INF;
7305 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7308 long sched_group_rt_runtime(struct task_group *tg)
7312 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7315 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7316 do_div(rt_runtime_us, NSEC_PER_USEC);
7317 return rt_runtime_us;
7320 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7322 u64 rt_runtime, rt_period;
7324 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7325 rt_runtime = tg->rt_bandwidth.rt_runtime;
7330 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7333 long sched_group_rt_period(struct task_group *tg)
7337 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7338 do_div(rt_period_us, NSEC_PER_USEC);
7339 return rt_period_us;
7342 static int sched_rt_global_constraints(void)
7344 u64 runtime, period;
7347 if (sysctl_sched_rt_period <= 0)
7350 runtime = global_rt_runtime();
7351 period = global_rt_period();
7354 * Sanity check on the sysctl variables.
7356 if (runtime > period && runtime != RUNTIME_INF)
7359 mutex_lock(&rt_constraints_mutex);
7360 read_lock(&tasklist_lock);
7361 ret = __rt_schedulable(NULL, 0, 0);
7362 read_unlock(&tasklist_lock);
7363 mutex_unlock(&rt_constraints_mutex);
7368 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7370 /* Don't accept realtime tasks when there is no way for them to run */
7371 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7377 #else /* !CONFIG_RT_GROUP_SCHED */
7378 static int sched_rt_global_constraints(void)
7380 unsigned long flags;
7383 if (sysctl_sched_rt_period <= 0)
7387 * There's always some RT tasks in the root group
7388 * -- migration, kstopmachine etc..
7390 if (sysctl_sched_rt_runtime == 0)
7393 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7394 for_each_possible_cpu(i) {
7395 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7397 raw_spin_lock(&rt_rq->rt_runtime_lock);
7398 rt_rq->rt_runtime = global_rt_runtime();
7399 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7401 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7405 #endif /* CONFIG_RT_GROUP_SCHED */
7407 int sched_rt_handler(struct ctl_table *table, int write,
7408 void __user *buffer, size_t *lenp,
7412 int old_period, old_runtime;
7413 static DEFINE_MUTEX(mutex);
7416 old_period = sysctl_sched_rt_period;
7417 old_runtime = sysctl_sched_rt_runtime;
7419 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7421 if (!ret && write) {
7422 ret = sched_rt_global_constraints();
7424 sysctl_sched_rt_period = old_period;
7425 sysctl_sched_rt_runtime = old_runtime;
7427 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7428 def_rt_bandwidth.rt_period =
7429 ns_to_ktime(global_rt_period());
7432 mutex_unlock(&mutex);
7437 #ifdef CONFIG_CGROUP_SCHED
7439 /* return corresponding task_group object of a cgroup */
7440 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7442 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7443 struct task_group, css);
7446 static struct cgroup_subsys_state *
7447 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7449 struct task_group *tg, *parent;
7451 if (!cgrp->parent) {
7452 /* This is early initialization for the top cgroup */
7453 return &root_task_group.css;
7456 parent = cgroup_tg(cgrp->parent);
7457 tg = sched_create_group(parent);
7459 return ERR_PTR(-ENOMEM);
7465 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7467 struct task_group *tg = cgroup_tg(cgrp);
7469 sched_destroy_group(tg);
7473 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
7475 #ifdef CONFIG_RT_GROUP_SCHED
7476 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
7479 /* We don't support RT-tasks being in separate groups */
7480 if (tsk->sched_class != &fair_sched_class)
7487 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
7489 sched_move_task(tsk);
7493 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7494 struct cgroup *old_cgrp, struct task_struct *task)
7497 * cgroup_exit() is called in the copy_process() failure path.
7498 * Ignore this case since the task hasn't ran yet, this avoids
7499 * trying to poke a half freed task state from generic code.
7501 if (!(task->flags & PF_EXITING))
7504 sched_move_task(task);
7507 #ifdef CONFIG_FAIR_GROUP_SCHED
7508 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7511 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7514 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7516 struct task_group *tg = cgroup_tg(cgrp);
7518 return (u64) scale_load_down(tg->shares);
7521 #ifdef CONFIG_CFS_BANDWIDTH
7522 static DEFINE_MUTEX(cfs_constraints_mutex);
7524 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7525 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7527 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7529 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7531 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7532 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7534 if (tg == &root_task_group)
7538 * Ensure we have at some amount of bandwidth every period. This is
7539 * to prevent reaching a state of large arrears when throttled via
7540 * entity_tick() resulting in prolonged exit starvation.
7542 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7546 * Likewise, bound things on the otherside by preventing insane quota
7547 * periods. This also allows us to normalize in computing quota
7550 if (period > max_cfs_quota_period)
7553 mutex_lock(&cfs_constraints_mutex);
7554 ret = __cfs_schedulable(tg, period, quota);
7558 runtime_enabled = quota != RUNTIME_INF;
7559 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7560 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7561 raw_spin_lock_irq(&cfs_b->lock);
7562 cfs_b->period = ns_to_ktime(period);
7563 cfs_b->quota = quota;
7565 __refill_cfs_bandwidth_runtime(cfs_b);
7566 /* restart the period timer (if active) to handle new period expiry */
7567 if (runtime_enabled && cfs_b->timer_active) {
7568 /* force a reprogram */
7569 cfs_b->timer_active = 0;
7570 __start_cfs_bandwidth(cfs_b);
7572 raw_spin_unlock_irq(&cfs_b->lock);
7574 for_each_possible_cpu(i) {
7575 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7576 struct rq *rq = cfs_rq->rq;
7578 raw_spin_lock_irq(&rq->lock);
7579 cfs_rq->runtime_enabled = runtime_enabled;
7580 cfs_rq->runtime_remaining = 0;
7582 if (cfs_rq->throttled)
7583 unthrottle_cfs_rq(cfs_rq);
7584 raw_spin_unlock_irq(&rq->lock);
7587 mutex_unlock(&cfs_constraints_mutex);
7592 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7596 period = ktime_to_ns(tg->cfs_bandwidth.period);
7597 if (cfs_quota_us < 0)
7598 quota = RUNTIME_INF;
7600 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7602 return tg_set_cfs_bandwidth(tg, period, quota);
7605 long tg_get_cfs_quota(struct task_group *tg)
7609 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7612 quota_us = tg->cfs_bandwidth.quota;
7613 do_div(quota_us, NSEC_PER_USEC);
7618 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7622 period = (u64)cfs_period_us * NSEC_PER_USEC;
7623 quota = tg->cfs_bandwidth.quota;
7628 return tg_set_cfs_bandwidth(tg, period, quota);
7631 long tg_get_cfs_period(struct task_group *tg)
7635 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7636 do_div(cfs_period_us, NSEC_PER_USEC);
7638 return cfs_period_us;
7641 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7643 return tg_get_cfs_quota(cgroup_tg(cgrp));
7646 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7649 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7652 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7654 return tg_get_cfs_period(cgroup_tg(cgrp));
7657 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7660 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7663 struct cfs_schedulable_data {
7664 struct task_group *tg;
7669 * normalize group quota/period to be quota/max_period
7670 * note: units are usecs
7672 static u64 normalize_cfs_quota(struct task_group *tg,
7673 struct cfs_schedulable_data *d)
7681 period = tg_get_cfs_period(tg);
7682 quota = tg_get_cfs_quota(tg);
7685 /* note: these should typically be equivalent */
7686 if (quota == RUNTIME_INF || quota == -1)
7689 return to_ratio(period, quota);
7692 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7694 struct cfs_schedulable_data *d = data;
7695 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7696 s64 quota = 0, parent_quota = -1;
7699 quota = RUNTIME_INF;
7701 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7703 quota = normalize_cfs_quota(tg, d);
7704 parent_quota = parent_b->hierarchal_quota;
7707 * ensure max(child_quota) <= parent_quota, inherit when no
7710 if (quota == RUNTIME_INF)
7711 quota = parent_quota;
7712 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7715 cfs_b->hierarchal_quota = quota;
7720 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7723 struct cfs_schedulable_data data = {
7729 if (quota != RUNTIME_INF) {
7730 do_div(data.period, NSEC_PER_USEC);
7731 do_div(data.quota, NSEC_PER_USEC);
7735 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7741 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7742 struct cgroup_map_cb *cb)
7744 struct task_group *tg = cgroup_tg(cgrp);
7745 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7747 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7748 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7749 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7753 #endif /* CONFIG_CFS_BANDWIDTH */
7754 #endif /* CONFIG_FAIR_GROUP_SCHED */
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7760 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7763 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7765 return sched_group_rt_runtime(cgroup_tg(cgrp));
7768 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7771 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7774 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7776 return sched_group_rt_period(cgroup_tg(cgrp));
7778 #endif /* CONFIG_RT_GROUP_SCHED */
7780 static struct cftype cpu_files[] = {
7781 #ifdef CONFIG_FAIR_GROUP_SCHED
7784 .read_u64 = cpu_shares_read_u64,
7785 .write_u64 = cpu_shares_write_u64,
7788 #ifdef CONFIG_CFS_BANDWIDTH
7790 .name = "cfs_quota_us",
7791 .read_s64 = cpu_cfs_quota_read_s64,
7792 .write_s64 = cpu_cfs_quota_write_s64,
7795 .name = "cfs_period_us",
7796 .read_u64 = cpu_cfs_period_read_u64,
7797 .write_u64 = cpu_cfs_period_write_u64,
7801 .read_map = cpu_stats_show,
7804 #ifdef CONFIG_RT_GROUP_SCHED
7806 .name = "rt_runtime_us",
7807 .read_s64 = cpu_rt_runtime_read,
7808 .write_s64 = cpu_rt_runtime_write,
7811 .name = "rt_period_us",
7812 .read_u64 = cpu_rt_period_read_uint,
7813 .write_u64 = cpu_rt_period_write_uint,
7818 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7820 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7823 struct cgroup_subsys cpu_cgroup_subsys = {
7825 .create = cpu_cgroup_create,
7826 .destroy = cpu_cgroup_destroy,
7827 .can_attach_task = cpu_cgroup_can_attach_task,
7828 .attach_task = cpu_cgroup_attach_task,
7829 .exit = cpu_cgroup_exit,
7830 .populate = cpu_cgroup_populate,
7831 .subsys_id = cpu_cgroup_subsys_id,
7835 #endif /* CONFIG_CGROUP_SCHED */
7837 #ifdef CONFIG_CGROUP_CPUACCT
7840 * CPU accounting code for task groups.
7842 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7843 * (balbir@in.ibm.com).
7846 /* track cpu usage of a group of tasks and its child groups */
7848 struct cgroup_subsys_state css;
7849 /* cpuusage holds pointer to a u64-type object on every cpu */
7850 u64 __percpu *cpuusage;
7851 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
7852 struct cpuacct *parent;
7855 struct cgroup_subsys cpuacct_subsys;
7857 /* return cpu accounting group corresponding to this container */
7858 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
7860 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
7861 struct cpuacct, css);
7864 /* return cpu accounting group to which this task belongs */
7865 static inline struct cpuacct *task_ca(struct task_struct *tsk)
7867 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
7868 struct cpuacct, css);
7871 /* create a new cpu accounting group */
7872 static struct cgroup_subsys_state *cpuacct_create(
7873 struct cgroup_subsys *ss, struct cgroup *cgrp)
7875 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7881 ca->cpuusage = alloc_percpu(u64);
7885 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
7886 if (percpu_counter_init(&ca->cpustat[i], 0))
7887 goto out_free_counters;
7890 ca->parent = cgroup_ca(cgrp->parent);
7896 percpu_counter_destroy(&ca->cpustat[i]);
7897 free_percpu(ca->cpuusage);
7901 return ERR_PTR(-ENOMEM);
7904 /* destroy an existing cpu accounting group */
7906 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7908 struct cpuacct *ca = cgroup_ca(cgrp);
7911 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
7912 percpu_counter_destroy(&ca->cpustat[i]);
7913 free_percpu(ca->cpuusage);
7917 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7919 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7922 #ifndef CONFIG_64BIT
7924 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7926 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7928 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7936 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7938 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7940 #ifndef CONFIG_64BIT
7942 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7944 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7946 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7952 /* return total cpu usage (in nanoseconds) of a group */
7953 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7955 struct cpuacct *ca = cgroup_ca(cgrp);
7956 u64 totalcpuusage = 0;
7959 for_each_present_cpu(i)
7960 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7962 return totalcpuusage;
7965 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7968 struct cpuacct *ca = cgroup_ca(cgrp);
7977 for_each_present_cpu(i)
7978 cpuacct_cpuusage_write(ca, i, 0);
7984 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7987 struct cpuacct *ca = cgroup_ca(cgroup);
7991 for_each_present_cpu(i) {
7992 percpu = cpuacct_cpuusage_read(ca, i);
7993 seq_printf(m, "%llu ", (unsigned long long) percpu);
7995 seq_printf(m, "\n");
7999 static const char *cpuacct_stat_desc[] = {
8000 [CPUACCT_STAT_USER] = "user",
8001 [CPUACCT_STAT_SYSTEM] = "system",
8004 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8005 struct cgroup_map_cb *cb)
8007 struct cpuacct *ca = cgroup_ca(cgrp);
8010 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8011 s64 val = percpu_counter_read(&ca->cpustat[i]);
8012 val = cputime64_to_clock_t(val);
8013 cb->fill(cb, cpuacct_stat_desc[i], val);
8018 static struct cftype files[] = {
8021 .read_u64 = cpuusage_read,
8022 .write_u64 = cpuusage_write,
8025 .name = "usage_percpu",
8026 .read_seq_string = cpuacct_percpu_seq_read,
8030 .read_map = cpuacct_stats_show,
8034 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8036 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8040 * charge this task's execution time to its accounting group.
8042 * called with rq->lock held.
8044 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8049 if (unlikely(!cpuacct_subsys.active))
8052 cpu = task_cpu(tsk);
8058 for (; ca; ca = ca->parent) {
8059 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8060 *cpuusage += cputime;
8067 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8068 * in cputime_t units. As a result, cpuacct_update_stats calls
8069 * percpu_counter_add with values large enough to always overflow the
8070 * per cpu batch limit causing bad SMP scalability.
8072 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8073 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8074 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8077 #define CPUACCT_BATCH \
8078 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8080 #define CPUACCT_BATCH 0
8084 * Charge the system/user time to the task's accounting group.
8086 void cpuacct_update_stats(struct task_struct *tsk,
8087 enum cpuacct_stat_index idx, cputime_t val)
8090 int batch = CPUACCT_BATCH;
8092 if (unlikely(!cpuacct_subsys.active))
8099 __percpu_counter_add(&ca->cpustat[idx], val, batch);
8105 struct cgroup_subsys cpuacct_subsys = {
8107 .create = cpuacct_create,
8108 .destroy = cpuacct_destroy,
8109 .populate = cpuacct_populate,
8110 .subsys_id = cpuacct_subsys_id,
8112 #endif /* CONFIG_CGROUP_CPUACCT */