4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
90 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
93 ktime_t soft, hard, now;
96 if (hrtimer_active(period_timer))
99 now = hrtimer_cb_get_time(period_timer);
100 hrtimer_forward(period_timer, now, period);
102 soft = hrtimer_get_softexpires(period_timer);
103 hard = hrtimer_get_expires(period_timer);
104 delta = ktime_to_ns(ktime_sub(hard, soft));
105 __hrtimer_start_range_ns(period_timer, soft, delta,
106 HRTIMER_MODE_ABS_PINNED, 0);
110 DEFINE_MUTEX(sched_domains_mutex);
111 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113 static void update_rq_clock_task(struct rq *rq, s64 delta);
115 void update_rq_clock(struct rq *rq)
119 if (rq->skip_clock_update > 0)
122 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 update_rq_clock_task(rq, delta);
128 * Debugging: various feature bits
131 #define SCHED_FEAT(name, enabled) \
132 (1UL << __SCHED_FEAT_##name) * enabled |
134 const_debug unsigned int sysctl_sched_features =
135 #include "features.h"
140 #ifdef CONFIG_SCHED_DEBUG
141 #define SCHED_FEAT(name, enabled) \
144 static __read_mostly char *sched_feat_names[] = {
145 #include "features.h"
151 static int sched_feat_show(struct seq_file *m, void *v)
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
158 seq_printf(m, "%s ", sched_feat_names[i]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174 #include "features.h"
179 static void sched_feat_disable(int i)
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
185 static void sched_feat_enable(int i)
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
191 static void sched_feat_disable(int i) { };
192 static void sched_feat_enable(int i) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
207 if (copy_from_user(&buf, ubuf, cnt))
213 if (strncmp(cmp, "NO_", 3) == 0) {
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
231 if (i == __SCHED_FEAT_NR)
239 static int sched_feat_open(struct inode *inode, struct file *filp)
241 return single_open(filp, sched_feat_show, NULL);
244 static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
249 .release = single_release,
252 static __init int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
259 late_initcall(sched_init_debug);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug unsigned int sysctl_sched_nr_migrate = 32;
269 * period over which we average the RT time consumption, measured
274 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period = 1000000;
282 __read_mostly int scheduler_running;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime = 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq *__task_rq_lock(struct task_struct *p)
300 lockdep_assert_held(&p->pi_lock);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
307 raw_spin_unlock(&rq->lock);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
331 static void __task_rq_unlock(struct rq *rq)
334 raw_spin_unlock(&rq->lock);
338 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
340 __releases(p->pi_lock)
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq *this_rq_lock(void)
356 raw_spin_lock(&rq->lock);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq *rq)
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 raw_spin_lock(&rq->lock);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
394 return HRTIMER_NORESTART;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg)
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq *rq, u64 delay)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
421 hrtimer_set_expires(timer, time);
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
432 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
434 int cpu = (int)(long)hcpu;
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
450 static __init void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq *rq, u64 delay)
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq *rq)
474 rq->hrtick_csd_pending = 0;
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq *rq)
489 static inline void init_rq_hrtick(struct rq *rq)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct *p)
515 assert_raw_spin_locked(&task_rq(p)->lock);
517 if (test_tsk_need_resched(p))
520 set_tsk_need_resched(p);
523 if (cpu == smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
532 void resched_cpu(int cpu)
534 struct rq *rq = cpu_rq(cpu);
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu = smp_processor_id();
556 struct sched_domain *sd;
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu)
583 struct rq *rq = cpu_rq(cpu);
585 if (cpu == smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq->curr != rq->idle)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq->idle);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq *rq)
628 s64 period = sched_avg_period();
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct *p)
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
661 struct task_group *parent, *child;
667 ret = (*down)(parent, data);
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
682 parent = parent->parent;
689 int tg_nop(struct task_group *tg, void *data)
695 void update_cpu_load(struct rq *this_rq);
697 static void set_load_weight(struct task_struct *p)
699 int prio = p->static_prio - MAX_RT_PRIO;
700 struct load_weight *load = &p->se.load;
703 * SCHED_IDLE tasks get minimal weight:
705 if (p->policy == SCHED_IDLE) {
706 load->weight = scale_load(WEIGHT_IDLEPRIO);
707 load->inv_weight = WMULT_IDLEPRIO;
711 load->weight = scale_load(prio_to_weight[prio]);
712 load->inv_weight = prio_to_wmult[prio];
715 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
718 sched_info_queued(p);
719 p->sched_class->enqueue_task(rq, p, flags);
722 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
725 sched_info_dequeued(p);
726 p->sched_class->dequeue_task(rq, p, flags);
729 void activate_task(struct rq *rq, struct task_struct *p, int flags)
731 if (task_contributes_to_load(p))
732 rq->nr_uninterruptible--;
734 enqueue_task(rq, p, flags);
737 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
739 if (task_contributes_to_load(p))
740 rq->nr_uninterruptible++;
742 dequeue_task(rq, p, flags);
745 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
748 * There are no locks covering percpu hardirq/softirq time.
749 * They are only modified in account_system_vtime, on corresponding CPU
750 * with interrupts disabled. So, writes are safe.
751 * They are read and saved off onto struct rq in update_rq_clock().
752 * This may result in other CPU reading this CPU's irq time and can
753 * race with irq/account_system_vtime on this CPU. We would either get old
754 * or new value with a side effect of accounting a slice of irq time to wrong
755 * task when irq is in progress while we read rq->clock. That is a worthy
756 * compromise in place of having locks on each irq in account_system_time.
758 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
759 static DEFINE_PER_CPU(u64, cpu_softirq_time);
761 static DEFINE_PER_CPU(u64, irq_start_time);
762 static int sched_clock_irqtime;
764 void enable_sched_clock_irqtime(void)
766 sched_clock_irqtime = 1;
769 void disable_sched_clock_irqtime(void)
771 sched_clock_irqtime = 0;
775 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
777 static inline void irq_time_write_begin(void)
779 __this_cpu_inc(irq_time_seq.sequence);
783 static inline void irq_time_write_end(void)
786 __this_cpu_inc(irq_time_seq.sequence);
789 static inline u64 irq_time_read(int cpu)
795 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
796 irq_time = per_cpu(cpu_softirq_time, cpu) +
797 per_cpu(cpu_hardirq_time, cpu);
798 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
802 #else /* CONFIG_64BIT */
803 static inline void irq_time_write_begin(void)
807 static inline void irq_time_write_end(void)
811 static inline u64 irq_time_read(int cpu)
813 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
815 #endif /* CONFIG_64BIT */
818 * Called before incrementing preempt_count on {soft,}irq_enter
819 * and before decrementing preempt_count on {soft,}irq_exit.
821 void account_system_vtime(struct task_struct *curr)
827 if (!sched_clock_irqtime)
830 local_irq_save(flags);
832 cpu = smp_processor_id();
833 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
834 __this_cpu_add(irq_start_time, delta);
836 irq_time_write_begin();
838 * We do not account for softirq time from ksoftirqd here.
839 * We want to continue accounting softirq time to ksoftirqd thread
840 * in that case, so as not to confuse scheduler with a special task
841 * that do not consume any time, but still wants to run.
844 __this_cpu_add(cpu_hardirq_time, delta);
845 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
846 __this_cpu_add(cpu_softirq_time, delta);
848 irq_time_write_end();
849 local_irq_restore(flags);
851 EXPORT_SYMBOL_GPL(account_system_vtime);
853 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
855 #ifdef CONFIG_PARAVIRT
856 static inline u64 steal_ticks(u64 steal)
858 if (unlikely(steal > NSEC_PER_SEC))
859 return div_u64(steal, TICK_NSEC);
861 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
865 static void update_rq_clock_task(struct rq *rq, s64 delta)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal = 0, irq_delta = 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta > delta)
895 rq->prev_irq_time += irq_delta;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled))) {
902 steal = paravirt_steal_clock(cpu_of(rq));
903 steal -= rq->prev_steal_time_rq;
905 if (unlikely(steal > delta))
908 st = steal_ticks(steal);
909 steal = st * TICK_NSEC;
911 rq->prev_steal_time_rq += steal;
917 rq->clock_task += delta;
919 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
920 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
921 sched_rt_avg_update(rq, irq_delta + steal);
925 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
926 static int irqtime_account_hi_update(void)
928 u64 *cpustat = kcpustat_this_cpu->cpustat;
933 local_irq_save(flags);
934 latest_ns = this_cpu_read(cpu_hardirq_time);
935 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
937 local_irq_restore(flags);
941 static int irqtime_account_si_update(void)
943 u64 *cpustat = kcpustat_this_cpu->cpustat;
948 local_irq_save(flags);
949 latest_ns = this_cpu_read(cpu_softirq_time);
950 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
952 local_irq_restore(flags);
956 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
958 #define sched_clock_irqtime (0)
962 void sched_set_stop_task(int cpu, struct task_struct *stop)
964 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
965 struct task_struct *old_stop = cpu_rq(cpu)->stop;
969 * Make it appear like a SCHED_FIFO task, its something
970 * userspace knows about and won't get confused about.
972 * Also, it will make PI more or less work without too
973 * much confusion -- but then, stop work should not
974 * rely on PI working anyway.
976 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
978 stop->sched_class = &stop_sched_class;
981 cpu_rq(cpu)->stop = stop;
985 * Reset it back to a normal scheduling class so that
986 * it can die in pieces.
988 old_stop->sched_class = &rt_sched_class;
993 * __normal_prio - return the priority that is based on the static prio
995 static inline int __normal_prio(struct task_struct *p)
997 return p->static_prio;
1001 * Calculate the expected normal priority: i.e. priority
1002 * without taking RT-inheritance into account. Might be
1003 * boosted by interactivity modifiers. Changes upon fork,
1004 * setprio syscalls, and whenever the interactivity
1005 * estimator recalculates.
1007 static inline int normal_prio(struct task_struct *p)
1011 if (task_has_rt_policy(p))
1012 prio = MAX_RT_PRIO-1 - p->rt_priority;
1014 prio = __normal_prio(p);
1019 * Calculate the current priority, i.e. the priority
1020 * taken into account by the scheduler. This value might
1021 * be boosted by RT tasks, or might be boosted by
1022 * interactivity modifiers. Will be RT if the task got
1023 * RT-boosted. If not then it returns p->normal_prio.
1025 static int effective_prio(struct task_struct *p)
1027 p->normal_prio = normal_prio(p);
1029 * If we are RT tasks or we were boosted to RT priority,
1030 * keep the priority unchanged. Otherwise, update priority
1031 * to the normal priority:
1033 if (!rt_prio(p->prio))
1034 return p->normal_prio;
1039 * task_curr - is this task currently executing on a CPU?
1040 * @p: the task in question.
1042 inline int task_curr(const struct task_struct *p)
1044 return cpu_curr(task_cpu(p)) == p;
1047 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1048 const struct sched_class *prev_class,
1051 if (prev_class != p->sched_class) {
1052 if (prev_class->switched_from)
1053 prev_class->switched_from(rq, p);
1054 p->sched_class->switched_to(rq, p);
1055 } else if (oldprio != p->prio)
1056 p->sched_class->prio_changed(rq, p, oldprio);
1059 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1061 const struct sched_class *class;
1063 if (p->sched_class == rq->curr->sched_class) {
1064 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1066 for_each_class(class) {
1067 if (class == rq->curr->sched_class)
1069 if (class == p->sched_class) {
1070 resched_task(rq->curr);
1077 * A queue event has occurred, and we're going to schedule. In
1078 * this case, we can save a useless back to back clock update.
1080 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1081 rq->skip_clock_update = 1;
1085 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1087 #ifdef CONFIG_SCHED_DEBUG
1089 * We should never call set_task_cpu() on a blocked task,
1090 * ttwu() will sort out the placement.
1092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1095 #ifdef CONFIG_LOCKDEP
1097 * The caller should hold either p->pi_lock or rq->lock, when changing
1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1100 * sched_move_task() holds both and thus holding either pins the cgroup,
1101 * see set_task_rq().
1103 * Furthermore, all task_rq users should acquire both locks, see
1106 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1107 lockdep_is_held(&task_rq(p)->lock)));
1111 trace_sched_migrate_task(p, new_cpu);
1113 if (task_cpu(p) != new_cpu) {
1114 p->se.nr_migrations++;
1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1118 __set_task_cpu(p, new_cpu);
1121 struct migration_arg {
1122 struct task_struct *task;
1126 static int migration_cpu_stop(void *data);
1129 * wait_task_inactive - wait for a thread to unschedule.
1131 * If @match_state is nonzero, it's the @p->state value just checked and
1132 * not expected to change. If it changes, i.e. @p might have woken up,
1133 * then return zero. When we succeed in waiting for @p to be off its CPU,
1134 * we return a positive number (its total switch count). If a second call
1135 * a short while later returns the same number, the caller can be sure that
1136 * @p has remained unscheduled the whole time.
1138 * The caller must ensure that the task *will* unschedule sometime soon,
1139 * else this function might spin for a *long* time. This function can't
1140 * be called with interrupts off, or it may introduce deadlock with
1141 * smp_call_function() if an IPI is sent by the same process we are
1142 * waiting to become inactive.
1144 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1146 unsigned long flags;
1153 * We do the initial early heuristics without holding
1154 * any task-queue locks at all. We'll only try to get
1155 * the runqueue lock when things look like they will
1161 * If the task is actively running on another CPU
1162 * still, just relax and busy-wait without holding
1165 * NOTE! Since we don't hold any locks, it's not
1166 * even sure that "rq" stays as the right runqueue!
1167 * But we don't care, since "task_running()" will
1168 * return false if the runqueue has changed and p
1169 * is actually now running somewhere else!
1171 while (task_running(rq, p)) {
1172 if (match_state && unlikely(p->state != match_state))
1178 * Ok, time to look more closely! We need the rq
1179 * lock now, to be *sure*. If we're wrong, we'll
1180 * just go back and repeat.
1182 rq = task_rq_lock(p, &flags);
1183 trace_sched_wait_task(p);
1184 running = task_running(rq, p);
1187 if (!match_state || p->state == match_state)
1188 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1189 task_rq_unlock(rq, p, &flags);
1192 * If it changed from the expected state, bail out now.
1194 if (unlikely(!ncsw))
1198 * Was it really running after all now that we
1199 * checked with the proper locks actually held?
1201 * Oops. Go back and try again..
1203 if (unlikely(running)) {
1209 * It's not enough that it's not actively running,
1210 * it must be off the runqueue _entirely_, and not
1213 * So if it was still runnable (but just not actively
1214 * running right now), it's preempted, and we should
1215 * yield - it could be a while.
1217 if (unlikely(on_rq)) {
1218 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1220 set_current_state(TASK_UNINTERRUPTIBLE);
1221 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1226 * Ahh, all good. It wasn't running, and it wasn't
1227 * runnable, which means that it will never become
1228 * running in the future either. We're all done!
1237 * kick_process - kick a running thread to enter/exit the kernel
1238 * @p: the to-be-kicked thread
1240 * Cause a process which is running on another CPU to enter
1241 * kernel-mode, without any delay. (to get signals handled.)
1243 * NOTE: this function doesn't have to take the runqueue lock,
1244 * because all it wants to ensure is that the remote task enters
1245 * the kernel. If the IPI races and the task has been migrated
1246 * to another CPU then no harm is done and the purpose has been
1249 void kick_process(struct task_struct *p)
1255 if ((cpu != smp_processor_id()) && task_curr(p))
1256 smp_send_reschedule(cpu);
1259 EXPORT_SYMBOL_GPL(kick_process);
1260 #endif /* CONFIG_SMP */
1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1266 static int select_fallback_rq(int cpu, struct task_struct *p)
1268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1269 enum { cpuset, possible, fail } state = cpuset;
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu, nodemask) {
1274 if (!cpu_online(dest_cpu))
1276 if (!cpu_active(dest_cpu))
1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283 /* Any allowed, online CPU? */
1284 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1285 if (!cpu_online(dest_cpu))
1287 if (!cpu_active(dest_cpu))
1294 /* No more Mr. Nice Guy. */
1295 cpuset_cpus_allowed_fallback(p);
1300 do_set_cpus_allowed(p, cpu_possible_mask);
1311 if (state != cpuset) {
1313 * Don't tell them about moving exiting tasks or
1314 * kernel threads (both mm NULL), since they never
1317 if (p->mm && printk_ratelimit()) {
1318 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1319 task_pid_nr(p), p->comm, cpu);
1327 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1330 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1332 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1335 * In order not to call set_task_cpu() on a blocking task we need
1336 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1339 * Since this is common to all placement strategies, this lives here.
1341 * [ this allows ->select_task() to simply return task_cpu(p) and
1342 * not worry about this generic constraint ]
1344 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1346 cpu = select_fallback_rq(task_cpu(p), p);
1351 static void update_avg(u64 *avg, u64 sample)
1353 s64 diff = sample - *avg;
1359 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1361 #ifdef CONFIG_SCHEDSTATS
1362 struct rq *rq = this_rq();
1365 int this_cpu = smp_processor_id();
1367 if (cpu == this_cpu) {
1368 schedstat_inc(rq, ttwu_local);
1369 schedstat_inc(p, se.statistics.nr_wakeups_local);
1371 struct sched_domain *sd;
1373 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1375 for_each_domain(this_cpu, sd) {
1376 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1377 schedstat_inc(sd, ttwu_wake_remote);
1384 if (wake_flags & WF_MIGRATED)
1385 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1387 #endif /* CONFIG_SMP */
1389 schedstat_inc(rq, ttwu_count);
1390 schedstat_inc(p, se.statistics.nr_wakeups);
1392 if (wake_flags & WF_SYNC)
1393 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1395 #endif /* CONFIG_SCHEDSTATS */
1398 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1400 activate_task(rq, p, en_flags);
1403 /* if a worker is waking up, notify workqueue */
1404 if (p->flags & PF_WQ_WORKER)
1405 wq_worker_waking_up(p, cpu_of(rq));
1409 * Mark the task runnable and perform wakeup-preemption.
1412 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1414 trace_sched_wakeup(p, true);
1415 check_preempt_curr(rq, p, wake_flags);
1417 p->state = TASK_RUNNING;
1419 if (p->sched_class->task_woken)
1420 p->sched_class->task_woken(rq, p);
1422 if (rq->idle_stamp) {
1423 u64 delta = rq->clock - rq->idle_stamp;
1424 u64 max = 2*sysctl_sched_migration_cost;
1429 update_avg(&rq->avg_idle, delta);
1436 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1439 if (p->sched_contributes_to_load)
1440 rq->nr_uninterruptible--;
1443 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1444 ttwu_do_wakeup(rq, p, wake_flags);
1448 * Called in case the task @p isn't fully descheduled from its runqueue,
1449 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1450 * since all we need to do is flip p->state to TASK_RUNNING, since
1451 * the task is still ->on_rq.
1453 static int ttwu_remote(struct task_struct *p, int wake_flags)
1458 rq = __task_rq_lock(p);
1460 ttwu_do_wakeup(rq, p, wake_flags);
1463 __task_rq_unlock(rq);
1469 static void sched_ttwu_pending(void)
1471 struct rq *rq = this_rq();
1472 struct llist_node *llist = llist_del_all(&rq->wake_list);
1473 struct task_struct *p;
1475 raw_spin_lock(&rq->lock);
1478 p = llist_entry(llist, struct task_struct, wake_entry);
1479 llist = llist_next(llist);
1480 ttwu_do_activate(rq, p, 0);
1483 raw_spin_unlock(&rq->lock);
1486 void scheduler_ipi(void)
1488 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1492 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1493 * traditionally all their work was done from the interrupt return
1494 * path. Now that we actually do some work, we need to make sure
1497 * Some archs already do call them, luckily irq_enter/exit nest
1500 * Arguably we should visit all archs and update all handlers,
1501 * however a fair share of IPIs are still resched only so this would
1502 * somewhat pessimize the simple resched case.
1505 sched_ttwu_pending();
1508 * Check if someone kicked us for doing the nohz idle load balance.
1510 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1511 this_rq()->idle_balance = 1;
1512 raise_softirq_irqoff(SCHED_SOFTIRQ);
1517 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1519 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1520 smp_send_reschedule(cpu);
1523 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1524 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1529 rq = __task_rq_lock(p);
1531 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1532 ttwu_do_wakeup(rq, p, wake_flags);
1535 __task_rq_unlock(rq);
1540 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1542 bool cpus_share_cache(int this_cpu, int that_cpu)
1544 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1546 #endif /* CONFIG_SMP */
1548 static void ttwu_queue(struct task_struct *p, int cpu)
1550 struct rq *rq = cpu_rq(cpu);
1552 #if defined(CONFIG_SMP)
1553 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1554 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1555 ttwu_queue_remote(p, cpu);
1560 raw_spin_lock(&rq->lock);
1561 ttwu_do_activate(rq, p, 0);
1562 raw_spin_unlock(&rq->lock);
1566 * try_to_wake_up - wake up a thread
1567 * @p: the thread to be awakened
1568 * @state: the mask of task states that can be woken
1569 * @wake_flags: wake modifier flags (WF_*)
1571 * Put it on the run-queue if it's not already there. The "current"
1572 * thread is always on the run-queue (except when the actual
1573 * re-schedule is in progress), and as such you're allowed to do
1574 * the simpler "current->state = TASK_RUNNING" to mark yourself
1575 * runnable without the overhead of this.
1577 * Returns %true if @p was woken up, %false if it was already running
1578 * or @state didn't match @p's state.
1581 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1583 unsigned long flags;
1584 int cpu, success = 0;
1587 raw_spin_lock_irqsave(&p->pi_lock, flags);
1588 if (!(p->state & state))
1591 success = 1; /* we're going to change ->state */
1594 if (p->on_rq && ttwu_remote(p, wake_flags))
1599 * If the owning (remote) cpu is still in the middle of schedule() with
1600 * this task as prev, wait until its done referencing the task.
1603 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1605 * In case the architecture enables interrupts in
1606 * context_switch(), we cannot busy wait, since that
1607 * would lead to deadlocks when an interrupt hits and
1608 * tries to wake up @prev. So bail and do a complete
1611 if (ttwu_activate_remote(p, wake_flags))
1618 * Pairs with the smp_wmb() in finish_lock_switch().
1622 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1623 p->state = TASK_WAKING;
1625 if (p->sched_class->task_waking)
1626 p->sched_class->task_waking(p);
1628 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1629 if (task_cpu(p) != cpu) {
1630 wake_flags |= WF_MIGRATED;
1631 set_task_cpu(p, cpu);
1633 #endif /* CONFIG_SMP */
1637 ttwu_stat(p, cpu, wake_flags);
1639 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1645 * try_to_wake_up_local - try to wake up a local task with rq lock held
1646 * @p: the thread to be awakened
1648 * Put @p on the run-queue if it's not already there. The caller must
1649 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1652 static void try_to_wake_up_local(struct task_struct *p)
1654 struct rq *rq = task_rq(p);
1656 BUG_ON(rq != this_rq());
1657 BUG_ON(p == current);
1658 lockdep_assert_held(&rq->lock);
1660 if (!raw_spin_trylock(&p->pi_lock)) {
1661 raw_spin_unlock(&rq->lock);
1662 raw_spin_lock(&p->pi_lock);
1663 raw_spin_lock(&rq->lock);
1666 if (!(p->state & TASK_NORMAL))
1670 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1672 ttwu_do_wakeup(rq, p, 0);
1673 ttwu_stat(p, smp_processor_id(), 0);
1675 raw_spin_unlock(&p->pi_lock);
1679 * wake_up_process - Wake up a specific process
1680 * @p: The process to be woken up.
1682 * Attempt to wake up the nominated process and move it to the set of runnable
1683 * processes. Returns 1 if the process was woken up, 0 if it was already
1686 * It may be assumed that this function implies a write memory barrier before
1687 * changing the task state if and only if any tasks are woken up.
1689 int wake_up_process(struct task_struct *p)
1691 return try_to_wake_up(p, TASK_ALL, 0);
1693 EXPORT_SYMBOL(wake_up_process);
1695 int wake_up_state(struct task_struct *p, unsigned int state)
1697 return try_to_wake_up(p, state, 0);
1701 * Perform scheduler related setup for a newly forked process p.
1702 * p is forked by current.
1704 * __sched_fork() is basic setup used by init_idle() too:
1706 static void __sched_fork(struct task_struct *p)
1711 p->se.exec_start = 0;
1712 p->se.sum_exec_runtime = 0;
1713 p->se.prev_sum_exec_runtime = 0;
1714 p->se.nr_migrations = 0;
1716 INIT_LIST_HEAD(&p->se.group_node);
1718 #ifdef CONFIG_SCHEDSTATS
1719 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1722 INIT_LIST_HEAD(&p->rt.run_list);
1724 #ifdef CONFIG_PREEMPT_NOTIFIERS
1725 INIT_HLIST_HEAD(&p->preempt_notifiers);
1730 * fork()/clone()-time setup:
1732 void sched_fork(struct task_struct *p)
1734 unsigned long flags;
1735 int cpu = get_cpu();
1739 * We mark the process as running here. This guarantees that
1740 * nobody will actually run it, and a signal or other external
1741 * event cannot wake it up and insert it on the runqueue either.
1743 p->state = TASK_RUNNING;
1746 * Make sure we do not leak PI boosting priority to the child.
1748 p->prio = current->normal_prio;
1751 * Revert to default priority/policy on fork if requested.
1753 if (unlikely(p->sched_reset_on_fork)) {
1754 if (task_has_rt_policy(p)) {
1755 p->policy = SCHED_NORMAL;
1756 p->static_prio = NICE_TO_PRIO(0);
1758 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1759 p->static_prio = NICE_TO_PRIO(0);
1761 p->prio = p->normal_prio = __normal_prio(p);
1765 * We don't need the reset flag anymore after the fork. It has
1766 * fulfilled its duty:
1768 p->sched_reset_on_fork = 0;
1771 if (!rt_prio(p->prio))
1772 p->sched_class = &fair_sched_class;
1774 if (p->sched_class->task_fork)
1775 p->sched_class->task_fork(p);
1778 * The child is not yet in the pid-hash so no cgroup attach races,
1779 * and the cgroup is pinned to this child due to cgroup_fork()
1780 * is ran before sched_fork().
1782 * Silence PROVE_RCU.
1784 raw_spin_lock_irqsave(&p->pi_lock, flags);
1785 set_task_cpu(p, cpu);
1786 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1788 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1789 if (likely(sched_info_on()))
1790 memset(&p->sched_info, 0, sizeof(p->sched_info));
1792 #if defined(CONFIG_SMP)
1795 #ifdef CONFIG_PREEMPT_COUNT
1796 /* Want to start with kernel preemption disabled. */
1797 task_thread_info(p)->preempt_count = 1;
1800 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1807 * wake_up_new_task - wake up a newly created task for the first time.
1809 * This function will do some initial scheduler statistics housekeeping
1810 * that must be done for every newly created context, then puts the task
1811 * on the runqueue and wakes it.
1813 void wake_up_new_task(struct task_struct *p)
1815 unsigned long flags;
1818 raw_spin_lock_irqsave(&p->pi_lock, flags);
1821 * Fork balancing, do it here and not earlier because:
1822 * - cpus_allowed can change in the fork path
1823 * - any previously selected cpu might disappear through hotplug
1825 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1828 rq = __task_rq_lock(p);
1829 activate_task(rq, p, 0);
1831 trace_sched_wakeup_new(p, true);
1832 check_preempt_curr(rq, p, WF_FORK);
1834 if (p->sched_class->task_woken)
1835 p->sched_class->task_woken(rq, p);
1837 task_rq_unlock(rq, p, &flags);
1840 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1844 * @notifier: notifier struct to register
1846 void preempt_notifier_register(struct preempt_notifier *notifier)
1848 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1850 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1853 * preempt_notifier_unregister - no longer interested in preemption notifications
1854 * @notifier: notifier struct to unregister
1856 * This is safe to call from within a preemption notifier.
1858 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1860 hlist_del(¬ifier->link);
1862 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1864 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1866 struct preempt_notifier *notifier;
1867 struct hlist_node *node;
1869 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1870 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1874 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1875 struct task_struct *next)
1877 struct preempt_notifier *notifier;
1878 struct hlist_node *node;
1880 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1881 notifier->ops->sched_out(notifier, next);
1884 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1886 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1891 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1892 struct task_struct *next)
1896 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1899 * prepare_task_switch - prepare to switch tasks
1900 * @rq: the runqueue preparing to switch
1901 * @prev: the current task that is being switched out
1902 * @next: the task we are going to switch to.
1904 * This is called with the rq lock held and interrupts off. It must
1905 * be paired with a subsequent finish_task_switch after the context
1908 * prepare_task_switch sets up locking and calls architecture specific
1912 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1913 struct task_struct *next)
1915 sched_info_switch(prev, next);
1916 perf_event_task_sched(prev, next);
1917 fire_sched_out_preempt_notifiers(prev, next);
1918 prepare_lock_switch(rq, next);
1919 prepare_arch_switch(next);
1920 trace_sched_switch(prev, next);
1924 * finish_task_switch - clean up after a task-switch
1925 * @rq: runqueue associated with task-switch
1926 * @prev: the thread we just switched away from.
1928 * finish_task_switch must be called after the context switch, paired
1929 * with a prepare_task_switch call before the context switch.
1930 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1931 * and do any other architecture-specific cleanup actions.
1933 * Note that we may have delayed dropping an mm in context_switch(). If
1934 * so, we finish that here outside of the runqueue lock. (Doing it
1935 * with the lock held can cause deadlocks; see schedule() for
1938 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1939 __releases(rq->lock)
1941 struct mm_struct *mm = rq->prev_mm;
1947 * A task struct has one reference for the use as "current".
1948 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1949 * schedule one last time. The schedule call will never return, and
1950 * the scheduled task must drop that reference.
1951 * The test for TASK_DEAD must occur while the runqueue locks are
1952 * still held, otherwise prev could be scheduled on another cpu, die
1953 * there before we look at prev->state, and then the reference would
1955 * Manfred Spraul <manfred@colorfullife.com>
1957 prev_state = prev->state;
1958 finish_arch_switch(prev);
1959 finish_lock_switch(rq, prev);
1960 finish_arch_post_lock_switch();
1962 fire_sched_in_preempt_notifiers(current);
1965 if (unlikely(prev_state == TASK_DEAD)) {
1967 * Remove function-return probe instances associated with this
1968 * task and put them back on the free list.
1970 kprobe_flush_task(prev);
1971 put_task_struct(prev);
1977 /* assumes rq->lock is held */
1978 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1980 if (prev->sched_class->pre_schedule)
1981 prev->sched_class->pre_schedule(rq, prev);
1984 /* rq->lock is NOT held, but preemption is disabled */
1985 static inline void post_schedule(struct rq *rq)
1987 if (rq->post_schedule) {
1988 unsigned long flags;
1990 raw_spin_lock_irqsave(&rq->lock, flags);
1991 if (rq->curr->sched_class->post_schedule)
1992 rq->curr->sched_class->post_schedule(rq);
1993 raw_spin_unlock_irqrestore(&rq->lock, flags);
1995 rq->post_schedule = 0;
2001 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2005 static inline void post_schedule(struct rq *rq)
2012 * schedule_tail - first thing a freshly forked thread must call.
2013 * @prev: the thread we just switched away from.
2015 asmlinkage void schedule_tail(struct task_struct *prev)
2016 __releases(rq->lock)
2018 struct rq *rq = this_rq();
2020 finish_task_switch(rq, prev);
2023 * FIXME: do we need to worry about rq being invalidated by the
2028 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2029 /* In this case, finish_task_switch does not reenable preemption */
2032 if (current->set_child_tid)
2033 put_user(task_pid_vnr(current), current->set_child_tid);
2037 * context_switch - switch to the new MM and the new
2038 * thread's register state.
2041 context_switch(struct rq *rq, struct task_struct *prev,
2042 struct task_struct *next)
2044 struct mm_struct *mm, *oldmm;
2046 prepare_task_switch(rq, prev, next);
2049 oldmm = prev->active_mm;
2051 * For paravirt, this is coupled with an exit in switch_to to
2052 * combine the page table reload and the switch backend into
2055 arch_start_context_switch(prev);
2058 next->active_mm = oldmm;
2059 atomic_inc(&oldmm->mm_count);
2060 enter_lazy_tlb(oldmm, next);
2062 switch_mm(oldmm, mm, next);
2065 prev->active_mm = NULL;
2066 rq->prev_mm = oldmm;
2069 * Since the runqueue lock will be released by the next
2070 * task (which is an invalid locking op but in the case
2071 * of the scheduler it's an obvious special-case), so we
2072 * do an early lockdep release here:
2074 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2075 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2078 /* Here we just switch the register state and the stack. */
2079 switch_to(prev, next, prev);
2083 * this_rq must be evaluated again because prev may have moved
2084 * CPUs since it called schedule(), thus the 'rq' on its stack
2085 * frame will be invalid.
2087 finish_task_switch(this_rq(), prev);
2091 * nr_running, nr_uninterruptible and nr_context_switches:
2093 * externally visible scheduler statistics: current number of runnable
2094 * threads, current number of uninterruptible-sleeping threads, total
2095 * number of context switches performed since bootup.
2097 unsigned long nr_running(void)
2099 unsigned long i, sum = 0;
2101 for_each_online_cpu(i)
2102 sum += cpu_rq(i)->nr_running;
2107 unsigned long nr_uninterruptible(void)
2109 unsigned long i, sum = 0;
2111 for_each_possible_cpu(i)
2112 sum += cpu_rq(i)->nr_uninterruptible;
2115 * Since we read the counters lockless, it might be slightly
2116 * inaccurate. Do not allow it to go below zero though:
2118 if (unlikely((long)sum < 0))
2124 unsigned long long nr_context_switches(void)
2127 unsigned long long sum = 0;
2129 for_each_possible_cpu(i)
2130 sum += cpu_rq(i)->nr_switches;
2135 unsigned long nr_iowait(void)
2137 unsigned long i, sum = 0;
2139 for_each_possible_cpu(i)
2140 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2145 unsigned long nr_iowait_cpu(int cpu)
2147 struct rq *this = cpu_rq(cpu);
2148 return atomic_read(&this->nr_iowait);
2151 unsigned long this_cpu_load(void)
2153 struct rq *this = this_rq();
2154 return this->cpu_load[0];
2158 /* Variables and functions for calc_load */
2159 static atomic_long_t calc_load_tasks;
2160 static unsigned long calc_load_update;
2161 unsigned long avenrun[3];
2162 EXPORT_SYMBOL(avenrun);
2164 static long calc_load_fold_active(struct rq *this_rq)
2166 long nr_active, delta = 0;
2168 nr_active = this_rq->nr_running;
2169 nr_active += (long) this_rq->nr_uninterruptible;
2171 if (nr_active != this_rq->calc_load_active) {
2172 delta = nr_active - this_rq->calc_load_active;
2173 this_rq->calc_load_active = nr_active;
2179 static unsigned long
2180 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2183 load += active * (FIXED_1 - exp);
2184 load += 1UL << (FSHIFT - 1);
2185 return load >> FSHIFT;
2190 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2192 * When making the ILB scale, we should try to pull this in as well.
2194 static atomic_long_t calc_load_tasks_idle;
2196 void calc_load_account_idle(struct rq *this_rq)
2200 delta = calc_load_fold_active(this_rq);
2202 atomic_long_add(delta, &calc_load_tasks_idle);
2205 static long calc_load_fold_idle(void)
2210 * Its got a race, we don't care...
2212 if (atomic_long_read(&calc_load_tasks_idle))
2213 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2219 * fixed_power_int - compute: x^n, in O(log n) time
2221 * @x: base of the power
2222 * @frac_bits: fractional bits of @x
2223 * @n: power to raise @x to.
2225 * By exploiting the relation between the definition of the natural power
2226 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2227 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2228 * (where: n_i \elem {0, 1}, the binary vector representing n),
2229 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2230 * of course trivially computable in O(log_2 n), the length of our binary
2233 static unsigned long
2234 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2236 unsigned long result = 1UL << frac_bits;
2241 result += 1UL << (frac_bits - 1);
2242 result >>= frac_bits;
2248 x += 1UL << (frac_bits - 1);
2256 * a1 = a0 * e + a * (1 - e)
2258 * a2 = a1 * e + a * (1 - e)
2259 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2260 * = a0 * e^2 + a * (1 - e) * (1 + e)
2262 * a3 = a2 * e + a * (1 - e)
2263 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2264 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2268 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2269 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2270 * = a0 * e^n + a * (1 - e^n)
2272 * [1] application of the geometric series:
2275 * S_n := \Sum x^i = -------------
2278 static unsigned long
2279 calc_load_n(unsigned long load, unsigned long exp,
2280 unsigned long active, unsigned int n)
2283 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2287 * NO_HZ can leave us missing all per-cpu ticks calling
2288 * calc_load_account_active(), but since an idle CPU folds its delta into
2289 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2290 * in the pending idle delta if our idle period crossed a load cycle boundary.
2292 * Once we've updated the global active value, we need to apply the exponential
2293 * weights adjusted to the number of cycles missed.
2295 static void calc_global_nohz(void)
2297 long delta, active, n;
2300 * If we crossed a calc_load_update boundary, make sure to fold
2301 * any pending idle changes, the respective CPUs might have
2302 * missed the tick driven calc_load_account_active() update
2305 delta = calc_load_fold_idle();
2307 atomic_long_add(delta, &calc_load_tasks);
2310 * It could be the one fold was all it took, we done!
2312 if (time_before(jiffies, calc_load_update + 10))
2316 * Catch-up, fold however many we are behind still
2318 delta = jiffies - calc_load_update - 10;
2319 n = 1 + (delta / LOAD_FREQ);
2321 active = atomic_long_read(&calc_load_tasks);
2322 active = active > 0 ? active * FIXED_1 : 0;
2324 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2325 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2326 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2328 calc_load_update += n * LOAD_FREQ;
2331 void calc_load_account_idle(struct rq *this_rq)
2335 static inline long calc_load_fold_idle(void)
2340 static void calc_global_nohz(void)
2346 * get_avenrun - get the load average array
2347 * @loads: pointer to dest load array
2348 * @offset: offset to add
2349 * @shift: shift count to shift the result left
2351 * These values are estimates at best, so no need for locking.
2353 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2355 loads[0] = (avenrun[0] + offset) << shift;
2356 loads[1] = (avenrun[1] + offset) << shift;
2357 loads[2] = (avenrun[2] + offset) << shift;
2361 * calc_load - update the avenrun load estimates 10 ticks after the
2362 * CPUs have updated calc_load_tasks.
2364 void calc_global_load(unsigned long ticks)
2368 if (time_before(jiffies, calc_load_update + 10))
2371 active = atomic_long_read(&calc_load_tasks);
2372 active = active > 0 ? active * FIXED_1 : 0;
2374 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2375 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2376 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2378 calc_load_update += LOAD_FREQ;
2381 * Account one period with whatever state we found before
2382 * folding in the nohz state and ageing the entire idle period.
2384 * This avoids loosing a sample when we go idle between
2385 * calc_load_account_active() (10 ticks ago) and now and thus
2392 * Called from update_cpu_load() to periodically update this CPU's
2395 static void calc_load_account_active(struct rq *this_rq)
2399 if (time_before(jiffies, this_rq->calc_load_update))
2402 delta = calc_load_fold_active(this_rq);
2403 delta += calc_load_fold_idle();
2405 atomic_long_add(delta, &calc_load_tasks);
2407 this_rq->calc_load_update += LOAD_FREQ;
2411 * The exact cpuload at various idx values, calculated at every tick would be
2412 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2414 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2415 * on nth tick when cpu may be busy, then we have:
2416 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2417 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2419 * decay_load_missed() below does efficient calculation of
2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2421 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2423 * The calculation is approximated on a 128 point scale.
2424 * degrade_zero_ticks is the number of ticks after which load at any
2425 * particular idx is approximated to be zero.
2426 * degrade_factor is a precomputed table, a row for each load idx.
2427 * Each column corresponds to degradation factor for a power of two ticks,
2428 * based on 128 point scale.
2430 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2431 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2433 * With this power of 2 load factors, we can degrade the load n times
2434 * by looking at 1 bits in n and doing as many mult/shift instead of
2435 * n mult/shifts needed by the exact degradation.
2437 #define DEGRADE_SHIFT 7
2438 static const unsigned char
2439 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2440 static const unsigned char
2441 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2442 {0, 0, 0, 0, 0, 0, 0, 0},
2443 {64, 32, 8, 0, 0, 0, 0, 0},
2444 {96, 72, 40, 12, 1, 0, 0},
2445 {112, 98, 75, 43, 15, 1, 0},
2446 {120, 112, 98, 76, 45, 16, 2} };
2449 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2450 * would be when CPU is idle and so we just decay the old load without
2451 * adding any new load.
2453 static unsigned long
2454 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2458 if (!missed_updates)
2461 if (missed_updates >= degrade_zero_ticks[idx])
2465 return load >> missed_updates;
2467 while (missed_updates) {
2468 if (missed_updates % 2)
2469 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2471 missed_updates >>= 1;
2478 * Update rq->cpu_load[] statistics. This function is usually called every
2479 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2480 * every tick. We fix it up based on jiffies.
2482 void update_cpu_load(struct rq *this_rq)
2484 unsigned long this_load = this_rq->load.weight;
2485 unsigned long curr_jiffies = jiffies;
2486 unsigned long pending_updates;
2489 this_rq->nr_load_updates++;
2491 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2492 if (curr_jiffies == this_rq->last_load_update_tick)
2495 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2496 this_rq->last_load_update_tick = curr_jiffies;
2498 /* Update our load: */
2499 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2500 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2501 unsigned long old_load, new_load;
2503 /* scale is effectively 1 << i now, and >> i divides by scale */
2505 old_load = this_rq->cpu_load[i];
2506 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2507 new_load = this_load;
2509 * Round up the averaging division if load is increasing. This
2510 * prevents us from getting stuck on 9 if the load is 10, for
2513 if (new_load > old_load)
2514 new_load += scale - 1;
2516 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2519 sched_avg_update(this_rq);
2522 static void update_cpu_load_active(struct rq *this_rq)
2524 update_cpu_load(this_rq);
2526 calc_load_account_active(this_rq);
2532 * sched_exec - execve() is a valuable balancing opportunity, because at
2533 * this point the task has the smallest effective memory and cache footprint.
2535 void sched_exec(void)
2537 struct task_struct *p = current;
2538 unsigned long flags;
2541 raw_spin_lock_irqsave(&p->pi_lock, flags);
2542 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2543 if (dest_cpu == smp_processor_id())
2546 if (likely(cpu_active(dest_cpu))) {
2547 struct migration_arg arg = { p, dest_cpu };
2549 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2550 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2554 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2559 DEFINE_PER_CPU(struct kernel_stat, kstat);
2560 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2562 EXPORT_PER_CPU_SYMBOL(kstat);
2563 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2566 * Return any ns on the sched_clock that have not yet been accounted in
2567 * @p in case that task is currently running.
2569 * Called with task_rq_lock() held on @rq.
2571 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2575 if (task_current(rq, p)) {
2576 update_rq_clock(rq);
2577 ns = rq->clock_task - p->se.exec_start;
2585 unsigned long long task_delta_exec(struct task_struct *p)
2587 unsigned long flags;
2591 rq = task_rq_lock(p, &flags);
2592 ns = do_task_delta_exec(p, rq);
2593 task_rq_unlock(rq, p, &flags);
2599 * Return accounted runtime for the task.
2600 * In case the task is currently running, return the runtime plus current's
2601 * pending runtime that have not been accounted yet.
2603 unsigned long long task_sched_runtime(struct task_struct *p)
2605 unsigned long flags;
2609 rq = task_rq_lock(p, &flags);
2610 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2611 task_rq_unlock(rq, p, &flags);
2616 #ifdef CONFIG_CGROUP_CPUACCT
2617 struct cgroup_subsys cpuacct_subsys;
2618 struct cpuacct root_cpuacct;
2621 static inline void task_group_account_field(struct task_struct *p, int index,
2624 #ifdef CONFIG_CGROUP_CPUACCT
2625 struct kernel_cpustat *kcpustat;
2629 * Since all updates are sure to touch the root cgroup, we
2630 * get ourselves ahead and touch it first. If the root cgroup
2631 * is the only cgroup, then nothing else should be necessary.
2634 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2636 #ifdef CONFIG_CGROUP_CPUACCT
2637 if (unlikely(!cpuacct_subsys.active))
2642 while (ca && (ca != &root_cpuacct)) {
2643 kcpustat = this_cpu_ptr(ca->cpustat);
2644 kcpustat->cpustat[index] += tmp;
2653 * Account user cpu time to a process.
2654 * @p: the process that the cpu time gets accounted to
2655 * @cputime: the cpu time spent in user space since the last update
2656 * @cputime_scaled: cputime scaled by cpu frequency
2658 void account_user_time(struct task_struct *p, cputime_t cputime,
2659 cputime_t cputime_scaled)
2663 /* Add user time to process. */
2664 p->utime += cputime;
2665 p->utimescaled += cputime_scaled;
2666 account_group_user_time(p, cputime);
2668 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2670 /* Add user time to cpustat. */
2671 task_group_account_field(p, index, (__force u64) cputime);
2673 /* Account for user time used */
2674 acct_update_integrals(p);
2678 * Account guest cpu time to a process.
2679 * @p: the process that the cpu time gets accounted to
2680 * @cputime: the cpu time spent in virtual machine since the last update
2681 * @cputime_scaled: cputime scaled by cpu frequency
2683 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2684 cputime_t cputime_scaled)
2686 u64 *cpustat = kcpustat_this_cpu->cpustat;
2688 /* Add guest time to process. */
2689 p->utime += cputime;
2690 p->utimescaled += cputime_scaled;
2691 account_group_user_time(p, cputime);
2692 p->gtime += cputime;
2694 /* Add guest time to cpustat. */
2695 if (TASK_NICE(p) > 0) {
2696 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2697 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2699 cpustat[CPUTIME_USER] += (__force u64) cputime;
2700 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2705 * Account system cpu time to a process and desired cpustat field
2706 * @p: the process that the cpu time gets accounted to
2707 * @cputime: the cpu time spent in kernel space since the last update
2708 * @cputime_scaled: cputime scaled by cpu frequency
2709 * @target_cputime64: pointer to cpustat field that has to be updated
2712 void __account_system_time(struct task_struct *p, cputime_t cputime,
2713 cputime_t cputime_scaled, int index)
2715 /* Add system time to process. */
2716 p->stime += cputime;
2717 p->stimescaled += cputime_scaled;
2718 account_group_system_time(p, cputime);
2720 /* Add system time to cpustat. */
2721 task_group_account_field(p, index, (__force u64) cputime);
2723 /* Account for system time used */
2724 acct_update_integrals(p);
2728 * Account system cpu time to a process.
2729 * @p: the process that the cpu time gets accounted to
2730 * @hardirq_offset: the offset to subtract from hardirq_count()
2731 * @cputime: the cpu time spent in kernel space since the last update
2732 * @cputime_scaled: cputime scaled by cpu frequency
2734 void account_system_time(struct task_struct *p, int hardirq_offset,
2735 cputime_t cputime, cputime_t cputime_scaled)
2739 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2740 account_guest_time(p, cputime, cputime_scaled);
2744 if (hardirq_count() - hardirq_offset)
2745 index = CPUTIME_IRQ;
2746 else if (in_serving_softirq())
2747 index = CPUTIME_SOFTIRQ;
2749 index = CPUTIME_SYSTEM;
2751 __account_system_time(p, cputime, cputime_scaled, index);
2755 * Account for involuntary wait time.
2756 * @cputime: the cpu time spent in involuntary wait
2758 void account_steal_time(cputime_t cputime)
2760 u64 *cpustat = kcpustat_this_cpu->cpustat;
2762 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2766 * Account for idle time.
2767 * @cputime: the cpu time spent in idle wait
2769 void account_idle_time(cputime_t cputime)
2771 u64 *cpustat = kcpustat_this_cpu->cpustat;
2772 struct rq *rq = this_rq();
2774 if (atomic_read(&rq->nr_iowait) > 0)
2775 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2777 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2780 static __always_inline bool steal_account_process_tick(void)
2782 #ifdef CONFIG_PARAVIRT
2783 if (static_key_false(¶virt_steal_enabled)) {
2786 steal = paravirt_steal_clock(smp_processor_id());
2787 steal -= this_rq()->prev_steal_time;
2789 st = steal_ticks(steal);
2790 this_rq()->prev_steal_time += st * TICK_NSEC;
2792 account_steal_time(st);
2799 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2801 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2803 * Account a tick to a process and cpustat
2804 * @p: the process that the cpu time gets accounted to
2805 * @user_tick: is the tick from userspace
2806 * @rq: the pointer to rq
2808 * Tick demultiplexing follows the order
2809 * - pending hardirq update
2810 * - pending softirq update
2814 * - check for guest_time
2815 * - else account as system_time
2817 * Check for hardirq is done both for system and user time as there is
2818 * no timer going off while we are on hardirq and hence we may never get an
2819 * opportunity to update it solely in system time.
2820 * p->stime and friends are only updated on system time and not on irq
2821 * softirq as those do not count in task exec_runtime any more.
2823 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2826 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2827 u64 *cpustat = kcpustat_this_cpu->cpustat;
2829 if (steal_account_process_tick())
2832 if (irqtime_account_hi_update()) {
2833 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2834 } else if (irqtime_account_si_update()) {
2835 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2836 } else if (this_cpu_ksoftirqd() == p) {
2838 * ksoftirqd time do not get accounted in cpu_softirq_time.
2839 * So, we have to handle it separately here.
2840 * Also, p->stime needs to be updated for ksoftirqd.
2842 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2844 } else if (user_tick) {
2845 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2846 } else if (p == rq->idle) {
2847 account_idle_time(cputime_one_jiffy);
2848 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2849 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2851 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2856 static void irqtime_account_idle_ticks(int ticks)
2859 struct rq *rq = this_rq();
2861 for (i = 0; i < ticks; i++)
2862 irqtime_account_process_tick(current, 0, rq);
2864 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2865 static void irqtime_account_idle_ticks(int ticks) {}
2866 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2868 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2871 * Account a single tick of cpu time.
2872 * @p: the process that the cpu time gets accounted to
2873 * @user_tick: indicates if the tick is a user or a system tick
2875 void account_process_tick(struct task_struct *p, int user_tick)
2877 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2878 struct rq *rq = this_rq();
2880 if (sched_clock_irqtime) {
2881 irqtime_account_process_tick(p, user_tick, rq);
2885 if (steal_account_process_tick())
2889 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2890 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2891 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2894 account_idle_time(cputime_one_jiffy);
2898 * Account multiple ticks of steal time.
2899 * @p: the process from which the cpu time has been stolen
2900 * @ticks: number of stolen ticks
2902 void account_steal_ticks(unsigned long ticks)
2904 account_steal_time(jiffies_to_cputime(ticks));
2908 * Account multiple ticks of idle time.
2909 * @ticks: number of stolen ticks
2911 void account_idle_ticks(unsigned long ticks)
2914 if (sched_clock_irqtime) {
2915 irqtime_account_idle_ticks(ticks);
2919 account_idle_time(jiffies_to_cputime(ticks));
2925 * Use precise platform statistics if available:
2927 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2928 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2934 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2936 struct task_cputime cputime;
2938 thread_group_cputime(p, &cputime);
2940 *ut = cputime.utime;
2941 *st = cputime.stime;
2945 #ifndef nsecs_to_cputime
2946 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2949 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2951 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2954 * Use CFS's precise accounting:
2956 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2959 u64 temp = (__force u64) rtime;
2961 temp *= (__force u64) utime;
2962 do_div(temp, (__force u32) total);
2963 utime = (__force cputime_t) temp;
2968 * Compare with previous values, to keep monotonicity:
2970 p->prev_utime = max(p->prev_utime, utime);
2971 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2973 *ut = p->prev_utime;
2974 *st = p->prev_stime;
2978 * Must be called with siglock held.
2980 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2982 struct signal_struct *sig = p->signal;
2983 struct task_cputime cputime;
2984 cputime_t rtime, utime, total;
2986 thread_group_cputime(p, &cputime);
2988 total = cputime.utime + cputime.stime;
2989 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2992 u64 temp = (__force u64) rtime;
2994 temp *= (__force u64) cputime.utime;
2995 do_div(temp, (__force u32) total);
2996 utime = (__force cputime_t) temp;
3000 sig->prev_utime = max(sig->prev_utime, utime);
3001 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3003 *ut = sig->prev_utime;
3004 *st = sig->prev_stime;
3009 * This function gets called by the timer code, with HZ frequency.
3010 * We call it with interrupts disabled.
3012 void scheduler_tick(void)
3014 int cpu = smp_processor_id();
3015 struct rq *rq = cpu_rq(cpu);
3016 struct task_struct *curr = rq->curr;
3020 raw_spin_lock(&rq->lock);
3021 update_rq_clock(rq);
3022 update_cpu_load_active(rq);
3023 curr->sched_class->task_tick(rq, curr, 0);
3024 raw_spin_unlock(&rq->lock);
3026 perf_event_task_tick();
3029 rq->idle_balance = idle_cpu(cpu);
3030 trigger_load_balance(rq, cpu);
3034 notrace unsigned long get_parent_ip(unsigned long addr)
3036 if (in_lock_functions(addr)) {
3037 addr = CALLER_ADDR2;
3038 if (in_lock_functions(addr))
3039 addr = CALLER_ADDR3;
3044 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3045 defined(CONFIG_PREEMPT_TRACER))
3047 void __kprobes add_preempt_count(int val)
3049 #ifdef CONFIG_DEBUG_PREEMPT
3053 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3056 preempt_count() += val;
3057 #ifdef CONFIG_DEBUG_PREEMPT
3059 * Spinlock count overflowing soon?
3061 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3064 if (preempt_count() == val)
3065 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3067 EXPORT_SYMBOL(add_preempt_count);
3069 void __kprobes sub_preempt_count(int val)
3071 #ifdef CONFIG_DEBUG_PREEMPT
3075 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3078 * Is the spinlock portion underflowing?
3080 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3081 !(preempt_count() & PREEMPT_MASK)))
3085 if (preempt_count() == val)
3086 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3087 preempt_count() -= val;
3089 EXPORT_SYMBOL(sub_preempt_count);
3094 * Print scheduling while atomic bug:
3096 static noinline void __schedule_bug(struct task_struct *prev)
3098 if (oops_in_progress)
3101 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3102 prev->comm, prev->pid, preempt_count());
3104 debug_show_held_locks(prev);
3106 if (irqs_disabled())
3107 print_irqtrace_events(prev);
3112 * Various schedule()-time debugging checks and statistics:
3114 static inline void schedule_debug(struct task_struct *prev)
3117 * Test if we are atomic. Since do_exit() needs to call into
3118 * schedule() atomically, we ignore that path for now.
3119 * Otherwise, whine if we are scheduling when we should not be.
3121 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3122 __schedule_bug(prev);
3125 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3127 schedstat_inc(this_rq(), sched_count);
3130 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3132 if (prev->on_rq || rq->skip_clock_update < 0)
3133 update_rq_clock(rq);
3134 prev->sched_class->put_prev_task(rq, prev);
3138 * Pick up the highest-prio task:
3140 static inline struct task_struct *
3141 pick_next_task(struct rq *rq)
3143 const struct sched_class *class;
3144 struct task_struct *p;
3147 * Optimization: we know that if all tasks are in
3148 * the fair class we can call that function directly:
3150 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3151 p = fair_sched_class.pick_next_task(rq);
3156 for_each_class(class) {
3157 p = class->pick_next_task(rq);
3162 BUG(); /* the idle class will always have a runnable task */
3166 * __schedule() is the main scheduler function.
3168 static void __sched __schedule(void)
3170 struct task_struct *prev, *next;
3171 unsigned long *switch_count;
3177 cpu = smp_processor_id();
3179 rcu_note_context_switch(cpu);
3182 schedule_debug(prev);
3184 if (sched_feat(HRTICK))
3187 raw_spin_lock_irq(&rq->lock);
3189 switch_count = &prev->nivcsw;
3190 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3191 if (unlikely(signal_pending_state(prev->state, prev))) {
3192 prev->state = TASK_RUNNING;
3194 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3198 * If a worker went to sleep, notify and ask workqueue
3199 * whether it wants to wake up a task to maintain
3202 if (prev->flags & PF_WQ_WORKER) {
3203 struct task_struct *to_wakeup;
3205 to_wakeup = wq_worker_sleeping(prev, cpu);
3207 try_to_wake_up_local(to_wakeup);
3210 switch_count = &prev->nvcsw;
3213 pre_schedule(rq, prev);
3215 if (unlikely(!rq->nr_running))
3216 idle_balance(cpu, rq);
3218 put_prev_task(rq, prev);
3219 next = pick_next_task(rq);
3220 clear_tsk_need_resched(prev);
3221 rq->skip_clock_update = 0;
3223 if (likely(prev != next)) {
3228 context_switch(rq, prev, next); /* unlocks the rq */
3230 * The context switch have flipped the stack from under us
3231 * and restored the local variables which were saved when
3232 * this task called schedule() in the past. prev == current
3233 * is still correct, but it can be moved to another cpu/rq.
3235 cpu = smp_processor_id();
3238 raw_spin_unlock_irq(&rq->lock);
3242 sched_preempt_enable_no_resched();
3247 static inline void sched_submit_work(struct task_struct *tsk)
3249 if (!tsk->state || tsk_is_pi_blocked(tsk))
3252 * If we are going to sleep and we have plugged IO queued,
3253 * make sure to submit it to avoid deadlocks.
3255 if (blk_needs_flush_plug(tsk))
3256 blk_schedule_flush_plug(tsk);
3259 asmlinkage void __sched schedule(void)
3261 struct task_struct *tsk = current;
3263 sched_submit_work(tsk);
3266 EXPORT_SYMBOL(schedule);
3269 * schedule_preempt_disabled - called with preemption disabled
3271 * Returns with preemption disabled. Note: preempt_count must be 1
3273 void __sched schedule_preempt_disabled(void)
3275 sched_preempt_enable_no_resched();
3280 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3282 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3284 if (lock->owner != owner)
3288 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3289 * lock->owner still matches owner, if that fails, owner might
3290 * point to free()d memory, if it still matches, the rcu_read_lock()
3291 * ensures the memory stays valid.
3295 return owner->on_cpu;
3299 * Look out! "owner" is an entirely speculative pointer
3300 * access and not reliable.
3302 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3304 if (!sched_feat(OWNER_SPIN))
3308 while (owner_running(lock, owner)) {
3312 arch_mutex_cpu_relax();
3317 * We break out the loop above on need_resched() and when the
3318 * owner changed, which is a sign for heavy contention. Return
3319 * success only when lock->owner is NULL.
3321 return lock->owner == NULL;
3325 #ifdef CONFIG_PREEMPT
3327 * this is the entry point to schedule() from in-kernel preemption
3328 * off of preempt_enable. Kernel preemptions off return from interrupt
3329 * occur there and call schedule directly.
3331 asmlinkage void __sched notrace preempt_schedule(void)
3333 struct thread_info *ti = current_thread_info();
3336 * If there is a non-zero preempt_count or interrupts are disabled,
3337 * we do not want to preempt the current task. Just return..
3339 if (likely(ti->preempt_count || irqs_disabled()))
3343 add_preempt_count_notrace(PREEMPT_ACTIVE);
3345 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3348 * Check again in case we missed a preemption opportunity
3349 * between schedule and now.
3352 } while (need_resched());
3354 EXPORT_SYMBOL(preempt_schedule);
3357 * this is the entry point to schedule() from kernel preemption
3358 * off of irq context.
3359 * Note, that this is called and return with irqs disabled. This will
3360 * protect us against recursive calling from irq.
3362 asmlinkage void __sched preempt_schedule_irq(void)
3364 struct thread_info *ti = current_thread_info();
3366 /* Catch callers which need to be fixed */
3367 BUG_ON(ti->preempt_count || !irqs_disabled());
3370 add_preempt_count(PREEMPT_ACTIVE);
3373 local_irq_disable();
3374 sub_preempt_count(PREEMPT_ACTIVE);
3377 * Check again in case we missed a preemption opportunity
3378 * between schedule and now.
3381 } while (need_resched());
3384 #endif /* CONFIG_PREEMPT */
3386 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3389 return try_to_wake_up(curr->private, mode, wake_flags);
3391 EXPORT_SYMBOL(default_wake_function);
3394 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3395 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3396 * number) then we wake all the non-exclusive tasks and one exclusive task.
3398 * There are circumstances in which we can try to wake a task which has already
3399 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3400 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3402 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3403 int nr_exclusive, int wake_flags, void *key)
3405 wait_queue_t *curr, *next;
3407 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3408 unsigned flags = curr->flags;
3410 if (curr->func(curr, mode, wake_flags, key) &&
3411 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3417 * __wake_up - wake up threads blocked on a waitqueue.
3419 * @mode: which threads
3420 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3421 * @key: is directly passed to the wakeup function
3423 * It may be assumed that this function implies a write memory barrier before
3424 * changing the task state if and only if any tasks are woken up.
3426 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3427 int nr_exclusive, void *key)
3429 unsigned long flags;
3431 spin_lock_irqsave(&q->lock, flags);
3432 __wake_up_common(q, mode, nr_exclusive, 0, key);
3433 spin_unlock_irqrestore(&q->lock, flags);
3435 EXPORT_SYMBOL(__wake_up);
3438 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3440 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3442 __wake_up_common(q, mode, nr, 0, NULL);
3444 EXPORT_SYMBOL_GPL(__wake_up_locked);
3446 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3448 __wake_up_common(q, mode, 1, 0, key);
3450 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3453 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3455 * @mode: which threads
3456 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3457 * @key: opaque value to be passed to wakeup targets
3459 * The sync wakeup differs that the waker knows that it will schedule
3460 * away soon, so while the target thread will be woken up, it will not
3461 * be migrated to another CPU - ie. the two threads are 'synchronized'
3462 * with each other. This can prevent needless bouncing between CPUs.
3464 * On UP it can prevent extra preemption.
3466 * It may be assumed that this function implies a write memory barrier before
3467 * changing the task state if and only if any tasks are woken up.
3469 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3470 int nr_exclusive, void *key)
3472 unsigned long flags;
3473 int wake_flags = WF_SYNC;
3478 if (unlikely(!nr_exclusive))
3481 spin_lock_irqsave(&q->lock, flags);
3482 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3483 spin_unlock_irqrestore(&q->lock, flags);
3485 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3488 * __wake_up_sync - see __wake_up_sync_key()
3490 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3492 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3494 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3497 * complete: - signals a single thread waiting on this completion
3498 * @x: holds the state of this particular completion
3500 * This will wake up a single thread waiting on this completion. Threads will be
3501 * awakened in the same order in which they were queued.
3503 * See also complete_all(), wait_for_completion() and related routines.
3505 * It may be assumed that this function implies a write memory barrier before
3506 * changing the task state if and only if any tasks are woken up.
3508 void complete(struct completion *x)
3510 unsigned long flags;
3512 spin_lock_irqsave(&x->wait.lock, flags);
3514 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3515 spin_unlock_irqrestore(&x->wait.lock, flags);
3517 EXPORT_SYMBOL(complete);
3520 * complete_all: - signals all threads waiting on this completion
3521 * @x: holds the state of this particular completion
3523 * This will wake up all threads waiting on this particular completion event.
3525 * It may be assumed that this function implies a write memory barrier before
3526 * changing the task state if and only if any tasks are woken up.
3528 void complete_all(struct completion *x)
3530 unsigned long flags;
3532 spin_lock_irqsave(&x->wait.lock, flags);
3533 x->done += UINT_MAX/2;
3534 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3535 spin_unlock_irqrestore(&x->wait.lock, flags);
3537 EXPORT_SYMBOL(complete_all);
3539 static inline long __sched
3540 do_wait_for_common(struct completion *x, long timeout, int state)
3543 DECLARE_WAITQUEUE(wait, current);
3545 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3547 if (signal_pending_state(state, current)) {
3548 timeout = -ERESTARTSYS;
3551 __set_current_state(state);
3552 spin_unlock_irq(&x->wait.lock);
3553 timeout = schedule_timeout(timeout);
3554 spin_lock_irq(&x->wait.lock);
3555 } while (!x->done && timeout);
3556 __remove_wait_queue(&x->wait, &wait);
3561 return timeout ?: 1;
3565 wait_for_common(struct completion *x, long timeout, int state)
3569 spin_lock_irq(&x->wait.lock);
3570 timeout = do_wait_for_common(x, timeout, state);
3571 spin_unlock_irq(&x->wait.lock);
3576 * wait_for_completion: - waits for completion of a task
3577 * @x: holds the state of this particular completion
3579 * This waits to be signaled for completion of a specific task. It is NOT
3580 * interruptible and there is no timeout.
3582 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3583 * and interrupt capability. Also see complete().
3585 void __sched wait_for_completion(struct completion *x)
3587 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3589 EXPORT_SYMBOL(wait_for_completion);
3592 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3593 * @x: holds the state of this particular completion
3594 * @timeout: timeout value in jiffies
3596 * This waits for either a completion of a specific task to be signaled or for a
3597 * specified timeout to expire. The timeout is in jiffies. It is not
3600 * The return value is 0 if timed out, and positive (at least 1, or number of
3601 * jiffies left till timeout) if completed.
3603 unsigned long __sched
3604 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3606 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3608 EXPORT_SYMBOL(wait_for_completion_timeout);
3611 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3612 * @x: holds the state of this particular completion
3614 * This waits for completion of a specific task to be signaled. It is
3617 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3619 int __sched wait_for_completion_interruptible(struct completion *x)
3621 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3622 if (t == -ERESTARTSYS)
3626 EXPORT_SYMBOL(wait_for_completion_interruptible);
3629 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3630 * @x: holds the state of this particular completion
3631 * @timeout: timeout value in jiffies
3633 * This waits for either a completion of a specific task to be signaled or for a
3634 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3636 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3637 * positive (at least 1, or number of jiffies left till timeout) if completed.
3640 wait_for_completion_interruptible_timeout(struct completion *x,
3641 unsigned long timeout)
3643 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3645 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3648 * wait_for_completion_killable: - waits for completion of a task (killable)
3649 * @x: holds the state of this particular completion
3651 * This waits to be signaled for completion of a specific task. It can be
3652 * interrupted by a kill signal.
3654 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3656 int __sched wait_for_completion_killable(struct completion *x)
3658 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3659 if (t == -ERESTARTSYS)
3663 EXPORT_SYMBOL(wait_for_completion_killable);
3666 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3667 * @x: holds the state of this particular completion
3668 * @timeout: timeout value in jiffies
3670 * This waits for either a completion of a specific task to be
3671 * signaled or for a specified timeout to expire. It can be
3672 * interrupted by a kill signal. The timeout is in jiffies.
3674 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3675 * positive (at least 1, or number of jiffies left till timeout) if completed.
3678 wait_for_completion_killable_timeout(struct completion *x,
3679 unsigned long timeout)
3681 return wait_for_common(x, timeout, TASK_KILLABLE);
3683 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3686 * try_wait_for_completion - try to decrement a completion without blocking
3687 * @x: completion structure
3689 * Returns: 0 if a decrement cannot be done without blocking
3690 * 1 if a decrement succeeded.
3692 * If a completion is being used as a counting completion,
3693 * attempt to decrement the counter without blocking. This
3694 * enables us to avoid waiting if the resource the completion
3695 * is protecting is not available.
3697 bool try_wait_for_completion(struct completion *x)
3699 unsigned long flags;
3702 spin_lock_irqsave(&x->wait.lock, flags);
3707 spin_unlock_irqrestore(&x->wait.lock, flags);
3710 EXPORT_SYMBOL(try_wait_for_completion);
3713 * completion_done - Test to see if a completion has any waiters
3714 * @x: completion structure
3716 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3717 * 1 if there are no waiters.
3720 bool completion_done(struct completion *x)
3722 unsigned long flags;
3725 spin_lock_irqsave(&x->wait.lock, flags);
3728 spin_unlock_irqrestore(&x->wait.lock, flags);
3731 EXPORT_SYMBOL(completion_done);
3734 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3736 unsigned long flags;
3739 init_waitqueue_entry(&wait, current);
3741 __set_current_state(state);
3743 spin_lock_irqsave(&q->lock, flags);
3744 __add_wait_queue(q, &wait);
3745 spin_unlock(&q->lock);
3746 timeout = schedule_timeout(timeout);
3747 spin_lock_irq(&q->lock);
3748 __remove_wait_queue(q, &wait);
3749 spin_unlock_irqrestore(&q->lock, flags);
3754 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3756 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3758 EXPORT_SYMBOL(interruptible_sleep_on);
3761 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3763 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3765 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3767 void __sched sleep_on(wait_queue_head_t *q)
3769 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3771 EXPORT_SYMBOL(sleep_on);
3773 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3775 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3777 EXPORT_SYMBOL(sleep_on_timeout);
3779 #ifdef CONFIG_RT_MUTEXES
3782 * rt_mutex_setprio - set the current priority of a task
3784 * @prio: prio value (kernel-internal form)
3786 * This function changes the 'effective' priority of a task. It does
3787 * not touch ->normal_prio like __setscheduler().
3789 * Used by the rt_mutex code to implement priority inheritance logic.
3791 void rt_mutex_setprio(struct task_struct *p, int prio)
3793 int oldprio, on_rq, running;
3795 const struct sched_class *prev_class;
3797 BUG_ON(prio < 0 || prio > MAX_PRIO);
3799 rq = __task_rq_lock(p);
3802 * Idle task boosting is a nono in general. There is one
3803 * exception, when PREEMPT_RT and NOHZ is active:
3805 * The idle task calls get_next_timer_interrupt() and holds
3806 * the timer wheel base->lock on the CPU and another CPU wants
3807 * to access the timer (probably to cancel it). We can safely
3808 * ignore the boosting request, as the idle CPU runs this code
3809 * with interrupts disabled and will complete the lock
3810 * protected section without being interrupted. So there is no
3811 * real need to boost.
3813 if (unlikely(p == rq->idle)) {
3814 WARN_ON(p != rq->curr);
3815 WARN_ON(p->pi_blocked_on);
3819 trace_sched_pi_setprio(p, prio);
3821 prev_class = p->sched_class;
3823 running = task_current(rq, p);
3825 dequeue_task(rq, p, 0);
3827 p->sched_class->put_prev_task(rq, p);
3830 p->sched_class = &rt_sched_class;
3832 p->sched_class = &fair_sched_class;
3837 p->sched_class->set_curr_task(rq);
3839 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3841 check_class_changed(rq, p, prev_class, oldprio);
3843 __task_rq_unlock(rq);
3846 void set_user_nice(struct task_struct *p, long nice)
3848 int old_prio, delta, on_rq;
3849 unsigned long flags;
3852 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3855 * We have to be careful, if called from sys_setpriority(),
3856 * the task might be in the middle of scheduling on another CPU.
3858 rq = task_rq_lock(p, &flags);
3860 * The RT priorities are set via sched_setscheduler(), but we still
3861 * allow the 'normal' nice value to be set - but as expected
3862 * it wont have any effect on scheduling until the task is
3863 * SCHED_FIFO/SCHED_RR:
3865 if (task_has_rt_policy(p)) {
3866 p->static_prio = NICE_TO_PRIO(nice);
3871 dequeue_task(rq, p, 0);
3873 p->static_prio = NICE_TO_PRIO(nice);
3876 p->prio = effective_prio(p);
3877 delta = p->prio - old_prio;
3880 enqueue_task(rq, p, 0);
3882 * If the task increased its priority or is running and
3883 * lowered its priority, then reschedule its CPU:
3885 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3886 resched_task(rq->curr);
3889 task_rq_unlock(rq, p, &flags);
3891 EXPORT_SYMBOL(set_user_nice);
3894 * can_nice - check if a task can reduce its nice value
3898 int can_nice(const struct task_struct *p, const int nice)
3900 /* convert nice value [19,-20] to rlimit style value [1,40] */
3901 int nice_rlim = 20 - nice;
3903 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3904 capable(CAP_SYS_NICE));
3907 #ifdef __ARCH_WANT_SYS_NICE
3910 * sys_nice - change the priority of the current process.
3911 * @increment: priority increment
3913 * sys_setpriority is a more generic, but much slower function that
3914 * does similar things.
3916 SYSCALL_DEFINE1(nice, int, increment)
3921 * Setpriority might change our priority at the same moment.
3922 * We don't have to worry. Conceptually one call occurs first
3923 * and we have a single winner.
3925 if (increment < -40)
3930 nice = TASK_NICE(current) + increment;
3936 if (increment < 0 && !can_nice(current, nice))
3939 retval = security_task_setnice(current, nice);
3943 set_user_nice(current, nice);
3950 * task_prio - return the priority value of a given task.
3951 * @p: the task in question.
3953 * This is the priority value as seen by users in /proc.
3954 * RT tasks are offset by -200. Normal tasks are centered
3955 * around 0, value goes from -16 to +15.
3957 int task_prio(const struct task_struct *p)
3959 return p->prio - MAX_RT_PRIO;
3963 * task_nice - return the nice value of a given task.
3964 * @p: the task in question.
3966 int task_nice(const struct task_struct *p)
3968 return TASK_NICE(p);
3970 EXPORT_SYMBOL(task_nice);
3973 * idle_cpu - is a given cpu idle currently?
3974 * @cpu: the processor in question.
3976 int idle_cpu(int cpu)
3978 struct rq *rq = cpu_rq(cpu);
3980 if (rq->curr != rq->idle)
3987 if (!llist_empty(&rq->wake_list))
3995 * idle_task - return the idle task for a given cpu.
3996 * @cpu: the processor in question.
3998 struct task_struct *idle_task(int cpu)
4000 return cpu_rq(cpu)->idle;
4004 * find_process_by_pid - find a process with a matching PID value.
4005 * @pid: the pid in question.
4007 static struct task_struct *find_process_by_pid(pid_t pid)
4009 return pid ? find_task_by_vpid(pid) : current;
4012 /* Actually do priority change: must hold rq lock. */
4014 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4017 p->rt_priority = prio;
4018 p->normal_prio = normal_prio(p);
4019 /* we are holding p->pi_lock already */
4020 p->prio = rt_mutex_getprio(p);
4021 if (rt_prio(p->prio))
4022 p->sched_class = &rt_sched_class;
4024 p->sched_class = &fair_sched_class;
4029 * check the target process has a UID that matches the current process's
4031 static bool check_same_owner(struct task_struct *p)
4033 const struct cred *cred = current_cred(), *pcred;
4037 pcred = __task_cred(p);
4038 if (cred->user->user_ns == pcred->user->user_ns)
4039 match = (cred->euid == pcred->euid ||
4040 cred->euid == pcred->uid);
4047 static int __sched_setscheduler(struct task_struct *p, int policy,
4048 const struct sched_param *param, bool user)
4050 int retval, oldprio, oldpolicy = -1, on_rq, running;
4051 unsigned long flags;
4052 const struct sched_class *prev_class;
4056 /* may grab non-irq protected spin_locks */
4057 BUG_ON(in_interrupt());
4059 /* double check policy once rq lock held */
4061 reset_on_fork = p->sched_reset_on_fork;
4062 policy = oldpolicy = p->policy;
4064 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4065 policy &= ~SCHED_RESET_ON_FORK;
4067 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4068 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4069 policy != SCHED_IDLE)
4074 * Valid priorities for SCHED_FIFO and SCHED_RR are
4075 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4076 * SCHED_BATCH and SCHED_IDLE is 0.
4078 if (param->sched_priority < 0 ||
4079 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4080 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4082 if (rt_policy(policy) != (param->sched_priority != 0))
4086 * Allow unprivileged RT tasks to decrease priority:
4088 if (user && !capable(CAP_SYS_NICE)) {
4089 if (rt_policy(policy)) {
4090 unsigned long rlim_rtprio =
4091 task_rlimit(p, RLIMIT_RTPRIO);
4093 /* can't set/change the rt policy */
4094 if (policy != p->policy && !rlim_rtprio)
4097 /* can't increase priority */
4098 if (param->sched_priority > p->rt_priority &&
4099 param->sched_priority > rlim_rtprio)
4104 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4105 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4107 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4108 if (!can_nice(p, TASK_NICE(p)))
4112 /* can't change other user's priorities */
4113 if (!check_same_owner(p))
4116 /* Normal users shall not reset the sched_reset_on_fork flag */
4117 if (p->sched_reset_on_fork && !reset_on_fork)
4122 retval = security_task_setscheduler(p);
4128 * make sure no PI-waiters arrive (or leave) while we are
4129 * changing the priority of the task:
4131 * To be able to change p->policy safely, the appropriate
4132 * runqueue lock must be held.
4134 rq = task_rq_lock(p, &flags);
4137 * Changing the policy of the stop threads its a very bad idea
4139 if (p == rq->stop) {
4140 task_rq_unlock(rq, p, &flags);
4145 * If not changing anything there's no need to proceed further:
4147 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4148 param->sched_priority == p->rt_priority))) {
4150 __task_rq_unlock(rq);
4151 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4155 #ifdef CONFIG_RT_GROUP_SCHED
4158 * Do not allow realtime tasks into groups that have no runtime
4161 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4162 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4163 !task_group_is_autogroup(task_group(p))) {
4164 task_rq_unlock(rq, p, &flags);
4170 /* recheck policy now with rq lock held */
4171 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4172 policy = oldpolicy = -1;
4173 task_rq_unlock(rq, p, &flags);
4177 running = task_current(rq, p);
4179 dequeue_task(rq, p, 0);
4181 p->sched_class->put_prev_task(rq, p);
4183 p->sched_reset_on_fork = reset_on_fork;
4186 prev_class = p->sched_class;
4187 __setscheduler(rq, p, policy, param->sched_priority);
4190 p->sched_class->set_curr_task(rq);
4192 enqueue_task(rq, p, 0);
4194 check_class_changed(rq, p, prev_class, oldprio);
4195 task_rq_unlock(rq, p, &flags);
4197 rt_mutex_adjust_pi(p);
4203 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4204 * @p: the task in question.
4205 * @policy: new policy.
4206 * @param: structure containing the new RT priority.
4208 * NOTE that the task may be already dead.
4210 int sched_setscheduler(struct task_struct *p, int policy,
4211 const struct sched_param *param)
4213 return __sched_setscheduler(p, policy, param, true);
4215 EXPORT_SYMBOL_GPL(sched_setscheduler);
4218 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4219 * @p: the task in question.
4220 * @policy: new policy.
4221 * @param: structure containing the new RT priority.
4223 * Just like sched_setscheduler, only don't bother checking if the
4224 * current context has permission. For example, this is needed in
4225 * stop_machine(): we create temporary high priority worker threads,
4226 * but our caller might not have that capability.
4228 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4229 const struct sched_param *param)
4231 return __sched_setscheduler(p, policy, param, false);
4235 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4237 struct sched_param lparam;
4238 struct task_struct *p;
4241 if (!param || pid < 0)
4243 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4248 p = find_process_by_pid(pid);
4250 retval = sched_setscheduler(p, policy, &lparam);
4257 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4258 * @pid: the pid in question.
4259 * @policy: new policy.
4260 * @param: structure containing the new RT priority.
4262 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4263 struct sched_param __user *, param)
4265 /* negative values for policy are not valid */
4269 return do_sched_setscheduler(pid, policy, param);
4273 * sys_sched_setparam - set/change the RT priority of a thread
4274 * @pid: the pid in question.
4275 * @param: structure containing the new RT priority.
4277 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4279 return do_sched_setscheduler(pid, -1, param);
4283 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4284 * @pid: the pid in question.
4286 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4288 struct task_struct *p;
4296 p = find_process_by_pid(pid);
4298 retval = security_task_getscheduler(p);
4301 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4308 * sys_sched_getparam - get the RT priority of a thread
4309 * @pid: the pid in question.
4310 * @param: structure containing the RT priority.
4312 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4314 struct sched_param lp;
4315 struct task_struct *p;
4318 if (!param || pid < 0)
4322 p = find_process_by_pid(pid);
4327 retval = security_task_getscheduler(p);
4331 lp.sched_priority = p->rt_priority;
4335 * This one might sleep, we cannot do it with a spinlock held ...
4337 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4346 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4348 cpumask_var_t cpus_allowed, new_mask;
4349 struct task_struct *p;
4355 p = find_process_by_pid(pid);
4362 /* Prevent p going away */
4366 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4370 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4372 goto out_free_cpus_allowed;
4375 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4378 retval = security_task_setscheduler(p);
4382 cpuset_cpus_allowed(p, cpus_allowed);
4383 cpumask_and(new_mask, in_mask, cpus_allowed);
4385 retval = set_cpus_allowed_ptr(p, new_mask);
4388 cpuset_cpus_allowed(p, cpus_allowed);
4389 if (!cpumask_subset(new_mask, cpus_allowed)) {
4391 * We must have raced with a concurrent cpuset
4392 * update. Just reset the cpus_allowed to the
4393 * cpuset's cpus_allowed
4395 cpumask_copy(new_mask, cpus_allowed);
4400 free_cpumask_var(new_mask);
4401 out_free_cpus_allowed:
4402 free_cpumask_var(cpus_allowed);
4409 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4410 struct cpumask *new_mask)
4412 if (len < cpumask_size())
4413 cpumask_clear(new_mask);
4414 else if (len > cpumask_size())
4415 len = cpumask_size();
4417 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4421 * sys_sched_setaffinity - set the cpu affinity of a process
4422 * @pid: pid of the process
4423 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4424 * @user_mask_ptr: user-space pointer to the new cpu mask
4426 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4427 unsigned long __user *, user_mask_ptr)
4429 cpumask_var_t new_mask;
4432 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4435 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4437 retval = sched_setaffinity(pid, new_mask);
4438 free_cpumask_var(new_mask);
4442 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4444 struct task_struct *p;
4445 unsigned long flags;
4452 p = find_process_by_pid(pid);
4456 retval = security_task_getscheduler(p);
4460 raw_spin_lock_irqsave(&p->pi_lock, flags);
4461 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4462 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4472 * sys_sched_getaffinity - get the cpu affinity of a process
4473 * @pid: pid of the process
4474 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4475 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4477 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4478 unsigned long __user *, user_mask_ptr)
4483 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4485 if (len & (sizeof(unsigned long)-1))
4488 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4491 ret = sched_getaffinity(pid, mask);
4493 size_t retlen = min_t(size_t, len, cpumask_size());
4495 if (copy_to_user(user_mask_ptr, mask, retlen))
4500 free_cpumask_var(mask);
4506 * sys_sched_yield - yield the current processor to other threads.
4508 * This function yields the current CPU to other tasks. If there are no
4509 * other threads running on this CPU then this function will return.
4511 SYSCALL_DEFINE0(sched_yield)
4513 struct rq *rq = this_rq_lock();
4515 schedstat_inc(rq, yld_count);
4516 current->sched_class->yield_task(rq);
4519 * Since we are going to call schedule() anyway, there's
4520 * no need to preempt or enable interrupts:
4522 __release(rq->lock);
4523 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4524 do_raw_spin_unlock(&rq->lock);
4525 sched_preempt_enable_no_resched();
4532 static inline int should_resched(void)
4534 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4537 static void __cond_resched(void)
4539 add_preempt_count(PREEMPT_ACTIVE);
4541 sub_preempt_count(PREEMPT_ACTIVE);
4544 int __sched _cond_resched(void)
4546 if (should_resched()) {
4552 EXPORT_SYMBOL(_cond_resched);
4555 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4556 * call schedule, and on return reacquire the lock.
4558 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4559 * operations here to prevent schedule() from being called twice (once via
4560 * spin_unlock(), once by hand).
4562 int __cond_resched_lock(spinlock_t *lock)
4564 int resched = should_resched();
4567 lockdep_assert_held(lock);
4569 if (spin_needbreak(lock) || resched) {
4580 EXPORT_SYMBOL(__cond_resched_lock);
4582 int __sched __cond_resched_softirq(void)
4584 BUG_ON(!in_softirq());
4586 if (should_resched()) {
4594 EXPORT_SYMBOL(__cond_resched_softirq);
4597 * yield - yield the current processor to other threads.
4599 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4601 * The scheduler is at all times free to pick the calling task as the most
4602 * eligible task to run, if removing the yield() call from your code breaks
4603 * it, its already broken.
4605 * Typical broken usage is:
4610 * where one assumes that yield() will let 'the other' process run that will
4611 * make event true. If the current task is a SCHED_FIFO task that will never
4612 * happen. Never use yield() as a progress guarantee!!
4614 * If you want to use yield() to wait for something, use wait_event().
4615 * If you want to use yield() to be 'nice' for others, use cond_resched().
4616 * If you still want to use yield(), do not!
4618 void __sched yield(void)
4620 set_current_state(TASK_RUNNING);
4623 EXPORT_SYMBOL(yield);
4626 * yield_to - yield the current processor to another thread in
4627 * your thread group, or accelerate that thread toward the
4628 * processor it's on.
4630 * @preempt: whether task preemption is allowed or not
4632 * It's the caller's job to ensure that the target task struct
4633 * can't go away on us before we can do any checks.
4635 * Returns true if we indeed boosted the target task.
4637 bool __sched yield_to(struct task_struct *p, bool preempt)
4639 struct task_struct *curr = current;
4640 struct rq *rq, *p_rq;
4641 unsigned long flags;
4644 local_irq_save(flags);
4649 double_rq_lock(rq, p_rq);
4650 while (task_rq(p) != p_rq) {
4651 double_rq_unlock(rq, p_rq);
4655 if (!curr->sched_class->yield_to_task)
4658 if (curr->sched_class != p->sched_class)
4661 if (task_running(p_rq, p) || p->state)
4664 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4666 schedstat_inc(rq, yld_count);
4668 * Make p's CPU reschedule; pick_next_entity takes care of
4671 if (preempt && rq != p_rq)
4672 resched_task(p_rq->curr);
4675 * We might have set it in task_yield_fair(), but are
4676 * not going to schedule(), so don't want to skip
4679 rq->skip_clock_update = 0;
4683 double_rq_unlock(rq, p_rq);
4684 local_irq_restore(flags);
4691 EXPORT_SYMBOL_GPL(yield_to);
4694 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4695 * that process accounting knows that this is a task in IO wait state.
4697 void __sched io_schedule(void)
4699 struct rq *rq = raw_rq();
4701 delayacct_blkio_start();
4702 atomic_inc(&rq->nr_iowait);
4703 blk_flush_plug(current);
4704 current->in_iowait = 1;
4706 current->in_iowait = 0;
4707 atomic_dec(&rq->nr_iowait);
4708 delayacct_blkio_end();
4710 EXPORT_SYMBOL(io_schedule);
4712 long __sched io_schedule_timeout(long timeout)
4714 struct rq *rq = raw_rq();
4717 delayacct_blkio_start();
4718 atomic_inc(&rq->nr_iowait);
4719 blk_flush_plug(current);
4720 current->in_iowait = 1;
4721 ret = schedule_timeout(timeout);
4722 current->in_iowait = 0;
4723 atomic_dec(&rq->nr_iowait);
4724 delayacct_blkio_end();
4729 * sys_sched_get_priority_max - return maximum RT priority.
4730 * @policy: scheduling class.
4732 * this syscall returns the maximum rt_priority that can be used
4733 * by a given scheduling class.
4735 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4742 ret = MAX_USER_RT_PRIO-1;
4754 * sys_sched_get_priority_min - return minimum RT priority.
4755 * @policy: scheduling class.
4757 * this syscall returns the minimum rt_priority that can be used
4758 * by a given scheduling class.
4760 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4778 * sys_sched_rr_get_interval - return the default timeslice of a process.
4779 * @pid: pid of the process.
4780 * @interval: userspace pointer to the timeslice value.
4782 * this syscall writes the default timeslice value of a given process
4783 * into the user-space timespec buffer. A value of '0' means infinity.
4785 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4786 struct timespec __user *, interval)
4788 struct task_struct *p;
4789 unsigned int time_slice;
4790 unsigned long flags;
4800 p = find_process_by_pid(pid);
4804 retval = security_task_getscheduler(p);
4808 rq = task_rq_lock(p, &flags);
4809 time_slice = p->sched_class->get_rr_interval(rq, p);
4810 task_rq_unlock(rq, p, &flags);
4813 jiffies_to_timespec(time_slice, &t);
4814 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4822 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4824 void sched_show_task(struct task_struct *p)
4826 unsigned long free = 0;
4829 state = p->state ? __ffs(p->state) + 1 : 0;
4830 printk(KERN_INFO "%-15.15s %c", p->comm,
4831 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4832 #if BITS_PER_LONG == 32
4833 if (state == TASK_RUNNING)
4834 printk(KERN_CONT " running ");
4836 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4838 if (state == TASK_RUNNING)
4839 printk(KERN_CONT " running task ");
4841 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4843 #ifdef CONFIG_DEBUG_STACK_USAGE
4844 free = stack_not_used(p);
4846 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4847 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4848 (unsigned long)task_thread_info(p)->flags);
4850 show_stack(p, NULL);
4853 void show_state_filter(unsigned long state_filter)
4855 struct task_struct *g, *p;
4857 #if BITS_PER_LONG == 32
4859 " task PC stack pid father\n");
4862 " task PC stack pid father\n");
4865 do_each_thread(g, p) {
4867 * reset the NMI-timeout, listing all files on a slow
4868 * console might take a lot of time:
4870 touch_nmi_watchdog();
4871 if (!state_filter || (p->state & state_filter))
4873 } while_each_thread(g, p);
4875 touch_all_softlockup_watchdogs();
4877 #ifdef CONFIG_SCHED_DEBUG
4878 sysrq_sched_debug_show();
4882 * Only show locks if all tasks are dumped:
4885 debug_show_all_locks();
4888 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4890 idle->sched_class = &idle_sched_class;
4894 * init_idle - set up an idle thread for a given CPU
4895 * @idle: task in question
4896 * @cpu: cpu the idle task belongs to
4898 * NOTE: this function does not set the idle thread's NEED_RESCHED
4899 * flag, to make booting more robust.
4901 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4903 struct rq *rq = cpu_rq(cpu);
4904 unsigned long flags;
4906 raw_spin_lock_irqsave(&rq->lock, flags);
4909 idle->state = TASK_RUNNING;
4910 idle->se.exec_start = sched_clock();
4912 do_set_cpus_allowed(idle, cpumask_of(cpu));
4914 * We're having a chicken and egg problem, even though we are
4915 * holding rq->lock, the cpu isn't yet set to this cpu so the
4916 * lockdep check in task_group() will fail.
4918 * Similar case to sched_fork(). / Alternatively we could
4919 * use task_rq_lock() here and obtain the other rq->lock.
4924 __set_task_cpu(idle, cpu);
4927 rq->curr = rq->idle = idle;
4928 #if defined(CONFIG_SMP)
4931 raw_spin_unlock_irqrestore(&rq->lock, flags);
4933 /* Set the preempt count _outside_ the spinlocks! */
4934 task_thread_info(idle)->preempt_count = 0;
4937 * The idle tasks have their own, simple scheduling class:
4939 idle->sched_class = &idle_sched_class;
4940 ftrace_graph_init_idle_task(idle, cpu);
4941 #if defined(CONFIG_SMP)
4942 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4947 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4949 if (p->sched_class && p->sched_class->set_cpus_allowed)
4950 p->sched_class->set_cpus_allowed(p, new_mask);
4952 cpumask_copy(&p->cpus_allowed, new_mask);
4953 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4957 * This is how migration works:
4959 * 1) we invoke migration_cpu_stop() on the target CPU using
4961 * 2) stopper starts to run (implicitly forcing the migrated thread
4963 * 3) it checks whether the migrated task is still in the wrong runqueue.
4964 * 4) if it's in the wrong runqueue then the migration thread removes
4965 * it and puts it into the right queue.
4966 * 5) stopper completes and stop_one_cpu() returns and the migration
4971 * Change a given task's CPU affinity. Migrate the thread to a
4972 * proper CPU and schedule it away if the CPU it's executing on
4973 * is removed from the allowed bitmask.
4975 * NOTE: the caller must have a valid reference to the task, the
4976 * task must not exit() & deallocate itself prematurely. The
4977 * call is not atomic; no spinlocks may be held.
4979 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4981 unsigned long flags;
4983 unsigned int dest_cpu;
4986 rq = task_rq_lock(p, &flags);
4988 if (cpumask_equal(&p->cpus_allowed, new_mask))
4991 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4996 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5001 do_set_cpus_allowed(p, new_mask);
5003 /* Can the task run on the task's current CPU? If so, we're done */
5004 if (cpumask_test_cpu(task_cpu(p), new_mask))
5007 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5009 struct migration_arg arg = { p, dest_cpu };
5010 /* Need help from migration thread: drop lock and wait. */
5011 task_rq_unlock(rq, p, &flags);
5012 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5013 tlb_migrate_finish(p->mm);
5017 task_rq_unlock(rq, p, &flags);
5021 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5024 * Move (not current) task off this cpu, onto dest cpu. We're doing
5025 * this because either it can't run here any more (set_cpus_allowed()
5026 * away from this CPU, or CPU going down), or because we're
5027 * attempting to rebalance this task on exec (sched_exec).
5029 * So we race with normal scheduler movements, but that's OK, as long
5030 * as the task is no longer on this CPU.
5032 * Returns non-zero if task was successfully migrated.
5034 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5036 struct rq *rq_dest, *rq_src;
5039 if (unlikely(!cpu_active(dest_cpu)))
5042 rq_src = cpu_rq(src_cpu);
5043 rq_dest = cpu_rq(dest_cpu);
5045 raw_spin_lock(&p->pi_lock);
5046 double_rq_lock(rq_src, rq_dest);
5047 /* Already moved. */
5048 if (task_cpu(p) != src_cpu)
5050 /* Affinity changed (again). */
5051 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5055 * If we're not on a rq, the next wake-up will ensure we're
5059 dequeue_task(rq_src, p, 0);
5060 set_task_cpu(p, dest_cpu);
5061 enqueue_task(rq_dest, p, 0);
5062 check_preempt_curr(rq_dest, p, 0);
5067 double_rq_unlock(rq_src, rq_dest);
5068 raw_spin_unlock(&p->pi_lock);
5073 * migration_cpu_stop - this will be executed by a highprio stopper thread
5074 * and performs thread migration by bumping thread off CPU then
5075 * 'pushing' onto another runqueue.
5077 static int migration_cpu_stop(void *data)
5079 struct migration_arg *arg = data;
5082 * The original target cpu might have gone down and we might
5083 * be on another cpu but it doesn't matter.
5085 local_irq_disable();
5086 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5091 #ifdef CONFIG_HOTPLUG_CPU
5094 * Ensures that the idle task is using init_mm right before its cpu goes
5097 void idle_task_exit(void)
5099 struct mm_struct *mm = current->active_mm;
5101 BUG_ON(cpu_online(smp_processor_id()));
5104 switch_mm(mm, &init_mm, current);
5109 * While a dead CPU has no uninterruptible tasks queued at this point,
5110 * it might still have a nonzero ->nr_uninterruptible counter, because
5111 * for performance reasons the counter is not stricly tracking tasks to
5112 * their home CPUs. So we just add the counter to another CPU's counter,
5113 * to keep the global sum constant after CPU-down:
5115 static void migrate_nr_uninterruptible(struct rq *rq_src)
5117 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5119 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5120 rq_src->nr_uninterruptible = 0;
5124 * remove the tasks which were accounted by rq from calc_load_tasks.
5126 static void calc_global_load_remove(struct rq *rq)
5128 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5129 rq->calc_load_active = 0;
5133 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5134 * try_to_wake_up()->select_task_rq().
5136 * Called with rq->lock held even though we'er in stop_machine() and
5137 * there's no concurrency possible, we hold the required locks anyway
5138 * because of lock validation efforts.
5140 static void migrate_tasks(unsigned int dead_cpu)
5142 struct rq *rq = cpu_rq(dead_cpu);
5143 struct task_struct *next, *stop = rq->stop;
5147 * Fudge the rq selection such that the below task selection loop
5148 * doesn't get stuck on the currently eligible stop task.
5150 * We're currently inside stop_machine() and the rq is either stuck
5151 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5152 * either way we should never end up calling schedule() until we're
5157 /* Ensure any throttled groups are reachable by pick_next_task */
5158 unthrottle_offline_cfs_rqs(rq);
5162 * There's this thread running, bail when that's the only
5165 if (rq->nr_running == 1)
5168 next = pick_next_task(rq);
5170 next->sched_class->put_prev_task(rq, next);
5172 /* Find suitable destination for @next, with force if needed. */
5173 dest_cpu = select_fallback_rq(dead_cpu, next);
5174 raw_spin_unlock(&rq->lock);
5176 __migrate_task(next, dead_cpu, dest_cpu);
5178 raw_spin_lock(&rq->lock);
5184 #endif /* CONFIG_HOTPLUG_CPU */
5186 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5188 static struct ctl_table sd_ctl_dir[] = {
5190 .procname = "sched_domain",
5196 static struct ctl_table sd_ctl_root[] = {
5198 .procname = "kernel",
5200 .child = sd_ctl_dir,
5205 static struct ctl_table *sd_alloc_ctl_entry(int n)
5207 struct ctl_table *entry =
5208 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5213 static void sd_free_ctl_entry(struct ctl_table **tablep)
5215 struct ctl_table *entry;
5218 * In the intermediate directories, both the child directory and
5219 * procname are dynamically allocated and could fail but the mode
5220 * will always be set. In the lowest directory the names are
5221 * static strings and all have proc handlers.
5223 for (entry = *tablep; entry->mode; entry++) {
5225 sd_free_ctl_entry(&entry->child);
5226 if (entry->proc_handler == NULL)
5227 kfree(entry->procname);
5235 set_table_entry(struct ctl_table *entry,
5236 const char *procname, void *data, int maxlen,
5237 umode_t mode, proc_handler *proc_handler)
5239 entry->procname = procname;
5241 entry->maxlen = maxlen;
5243 entry->proc_handler = proc_handler;
5246 static struct ctl_table *
5247 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5249 struct ctl_table *table = sd_alloc_ctl_entry(13);
5254 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5255 sizeof(long), 0644, proc_doulongvec_minmax);
5256 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5257 sizeof(long), 0644, proc_doulongvec_minmax);
5258 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5259 sizeof(int), 0644, proc_dointvec_minmax);
5260 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5261 sizeof(int), 0644, proc_dointvec_minmax);
5262 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5263 sizeof(int), 0644, proc_dointvec_minmax);
5264 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5265 sizeof(int), 0644, proc_dointvec_minmax);
5266 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5267 sizeof(int), 0644, proc_dointvec_minmax);
5268 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5269 sizeof(int), 0644, proc_dointvec_minmax);
5270 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5271 sizeof(int), 0644, proc_dointvec_minmax);
5272 set_table_entry(&table[9], "cache_nice_tries",
5273 &sd->cache_nice_tries,
5274 sizeof(int), 0644, proc_dointvec_minmax);
5275 set_table_entry(&table[10], "flags", &sd->flags,
5276 sizeof(int), 0644, proc_dointvec_minmax);
5277 set_table_entry(&table[11], "name", sd->name,
5278 CORENAME_MAX_SIZE, 0444, proc_dostring);
5279 /* &table[12] is terminator */
5284 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5286 struct ctl_table *entry, *table;
5287 struct sched_domain *sd;
5288 int domain_num = 0, i;
5291 for_each_domain(cpu, sd)
5293 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5298 for_each_domain(cpu, sd) {
5299 snprintf(buf, 32, "domain%d", i);
5300 entry->procname = kstrdup(buf, GFP_KERNEL);
5302 entry->child = sd_alloc_ctl_domain_table(sd);
5309 static struct ctl_table_header *sd_sysctl_header;
5310 static void register_sched_domain_sysctl(void)
5312 int i, cpu_num = num_possible_cpus();
5313 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5316 WARN_ON(sd_ctl_dir[0].child);
5317 sd_ctl_dir[0].child = entry;
5322 for_each_possible_cpu(i) {
5323 snprintf(buf, 32, "cpu%d", i);
5324 entry->procname = kstrdup(buf, GFP_KERNEL);
5326 entry->child = sd_alloc_ctl_cpu_table(i);
5330 WARN_ON(sd_sysctl_header);
5331 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5334 /* may be called multiple times per register */
5335 static void unregister_sched_domain_sysctl(void)
5337 if (sd_sysctl_header)
5338 unregister_sysctl_table(sd_sysctl_header);
5339 sd_sysctl_header = NULL;
5340 if (sd_ctl_dir[0].child)
5341 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5344 static void register_sched_domain_sysctl(void)
5347 static void unregister_sched_domain_sysctl(void)
5352 static void set_rq_online(struct rq *rq)
5355 const struct sched_class *class;
5357 cpumask_set_cpu(rq->cpu, rq->rd->online);
5360 for_each_class(class) {
5361 if (class->rq_online)
5362 class->rq_online(rq);
5367 static void set_rq_offline(struct rq *rq)
5370 const struct sched_class *class;
5372 for_each_class(class) {
5373 if (class->rq_offline)
5374 class->rq_offline(rq);
5377 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5383 * migration_call - callback that gets triggered when a CPU is added.
5384 * Here we can start up the necessary migration thread for the new CPU.
5386 static int __cpuinit
5387 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5389 int cpu = (long)hcpu;
5390 unsigned long flags;
5391 struct rq *rq = cpu_rq(cpu);
5393 switch (action & ~CPU_TASKS_FROZEN) {
5395 case CPU_UP_PREPARE:
5396 rq->calc_load_update = calc_load_update;
5400 /* Update our root-domain */
5401 raw_spin_lock_irqsave(&rq->lock, flags);
5403 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5407 raw_spin_unlock_irqrestore(&rq->lock, flags);
5410 #ifdef CONFIG_HOTPLUG_CPU
5412 sched_ttwu_pending();
5413 /* Update our root-domain */
5414 raw_spin_lock_irqsave(&rq->lock, flags);
5416 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5420 BUG_ON(rq->nr_running != 1); /* the migration thread */
5421 raw_spin_unlock_irqrestore(&rq->lock, flags);
5423 migrate_nr_uninterruptible(rq);
5424 calc_global_load_remove(rq);
5429 update_max_interval();
5435 * Register at high priority so that task migration (migrate_all_tasks)
5436 * happens before everything else. This has to be lower priority than
5437 * the notifier in the perf_event subsystem, though.
5439 static struct notifier_block __cpuinitdata migration_notifier = {
5440 .notifier_call = migration_call,
5441 .priority = CPU_PRI_MIGRATION,
5444 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5445 unsigned long action, void *hcpu)
5447 switch (action & ~CPU_TASKS_FROZEN) {
5449 case CPU_DOWN_FAILED:
5450 set_cpu_active((long)hcpu, true);
5457 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5458 unsigned long action, void *hcpu)
5460 switch (action & ~CPU_TASKS_FROZEN) {
5461 case CPU_DOWN_PREPARE:
5462 set_cpu_active((long)hcpu, false);
5469 static int __init migration_init(void)
5471 void *cpu = (void *)(long)smp_processor_id();
5474 /* Initialize migration for the boot CPU */
5475 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5476 BUG_ON(err == NOTIFY_BAD);
5477 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5478 register_cpu_notifier(&migration_notifier);
5480 /* Register cpu active notifiers */
5481 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5482 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5486 early_initcall(migration_init);
5491 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5493 #ifdef CONFIG_SCHED_DEBUG
5495 static __read_mostly int sched_domain_debug_enabled;
5497 static int __init sched_domain_debug_setup(char *str)
5499 sched_domain_debug_enabled = 1;
5503 early_param("sched_debug", sched_domain_debug_setup);
5505 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5506 struct cpumask *groupmask)
5508 struct sched_group *group = sd->groups;
5511 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5512 cpumask_clear(groupmask);
5514 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5516 if (!(sd->flags & SD_LOAD_BALANCE)) {
5517 printk("does not load-balance\n");
5519 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5524 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5526 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5527 printk(KERN_ERR "ERROR: domain->span does not contain "
5530 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5531 printk(KERN_ERR "ERROR: domain->groups does not contain"
5535 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5539 printk(KERN_ERR "ERROR: group is NULL\n");
5543 if (!group->sgp->power) {
5544 printk(KERN_CONT "\n");
5545 printk(KERN_ERR "ERROR: domain->cpu_power not "
5550 if (!cpumask_weight(sched_group_cpus(group))) {
5551 printk(KERN_CONT "\n");
5552 printk(KERN_ERR "ERROR: empty group\n");
5556 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5557 printk(KERN_CONT "\n");
5558 printk(KERN_ERR "ERROR: repeated CPUs\n");
5562 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5564 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5566 printk(KERN_CONT " %s", str);
5567 if (group->sgp->power != SCHED_POWER_SCALE) {
5568 printk(KERN_CONT " (cpu_power = %d)",
5572 group = group->next;
5573 } while (group != sd->groups);
5574 printk(KERN_CONT "\n");
5576 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5577 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5580 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5581 printk(KERN_ERR "ERROR: parent span is not a superset "
5582 "of domain->span\n");
5586 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5590 if (!sched_domain_debug_enabled)
5594 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5598 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5601 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5609 #else /* !CONFIG_SCHED_DEBUG */
5610 # define sched_domain_debug(sd, cpu) do { } while (0)
5611 #endif /* CONFIG_SCHED_DEBUG */
5613 static int sd_degenerate(struct sched_domain *sd)
5615 if (cpumask_weight(sched_domain_span(sd)) == 1)
5618 /* Following flags need at least 2 groups */
5619 if (sd->flags & (SD_LOAD_BALANCE |
5620 SD_BALANCE_NEWIDLE |
5624 SD_SHARE_PKG_RESOURCES)) {
5625 if (sd->groups != sd->groups->next)
5629 /* Following flags don't use groups */
5630 if (sd->flags & (SD_WAKE_AFFINE))
5637 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5639 unsigned long cflags = sd->flags, pflags = parent->flags;
5641 if (sd_degenerate(parent))
5644 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5647 /* Flags needing groups don't count if only 1 group in parent */
5648 if (parent->groups == parent->groups->next) {
5649 pflags &= ~(SD_LOAD_BALANCE |
5650 SD_BALANCE_NEWIDLE |
5654 SD_SHARE_PKG_RESOURCES);
5655 if (nr_node_ids == 1)
5656 pflags &= ~SD_SERIALIZE;
5658 if (~cflags & pflags)
5664 static void free_rootdomain(struct rcu_head *rcu)
5666 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5668 cpupri_cleanup(&rd->cpupri);
5669 free_cpumask_var(rd->rto_mask);
5670 free_cpumask_var(rd->online);
5671 free_cpumask_var(rd->span);
5675 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5677 struct root_domain *old_rd = NULL;
5678 unsigned long flags;
5680 raw_spin_lock_irqsave(&rq->lock, flags);
5685 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5688 cpumask_clear_cpu(rq->cpu, old_rd->span);
5691 * If we dont want to free the old_rt yet then
5692 * set old_rd to NULL to skip the freeing later
5695 if (!atomic_dec_and_test(&old_rd->refcount))
5699 atomic_inc(&rd->refcount);
5702 cpumask_set_cpu(rq->cpu, rd->span);
5703 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5706 raw_spin_unlock_irqrestore(&rq->lock, flags);
5709 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5712 static int init_rootdomain(struct root_domain *rd)
5714 memset(rd, 0, sizeof(*rd));
5716 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5718 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5720 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5723 if (cpupri_init(&rd->cpupri) != 0)
5728 free_cpumask_var(rd->rto_mask);
5730 free_cpumask_var(rd->online);
5732 free_cpumask_var(rd->span);
5738 * By default the system creates a single root-domain with all cpus as
5739 * members (mimicking the global state we have today).
5741 struct root_domain def_root_domain;
5743 static void init_defrootdomain(void)
5745 init_rootdomain(&def_root_domain);
5747 atomic_set(&def_root_domain.refcount, 1);
5750 static struct root_domain *alloc_rootdomain(void)
5752 struct root_domain *rd;
5754 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5758 if (init_rootdomain(rd) != 0) {
5766 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5768 struct sched_group *tmp, *first;
5777 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5782 } while (sg != first);
5785 static void free_sched_domain(struct rcu_head *rcu)
5787 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5790 * If its an overlapping domain it has private groups, iterate and
5793 if (sd->flags & SD_OVERLAP) {
5794 free_sched_groups(sd->groups, 1);
5795 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5796 kfree(sd->groups->sgp);
5802 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5804 call_rcu(&sd->rcu, free_sched_domain);
5807 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5809 for (; sd; sd = sd->parent)
5810 destroy_sched_domain(sd, cpu);
5814 * Keep a special pointer to the highest sched_domain that has
5815 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5816 * allows us to avoid some pointer chasing select_idle_sibling().
5818 * Also keep a unique ID per domain (we use the first cpu number in
5819 * the cpumask of the domain), this allows us to quickly tell if
5820 * two cpus are in the same cache domain, see cpus_share_cache().
5822 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5823 DEFINE_PER_CPU(int, sd_llc_id);
5825 static void update_top_cache_domain(int cpu)
5827 struct sched_domain *sd;
5830 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5832 id = cpumask_first(sched_domain_span(sd));
5834 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5835 per_cpu(sd_llc_id, cpu) = id;
5839 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5840 * hold the hotplug lock.
5843 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5845 struct rq *rq = cpu_rq(cpu);
5846 struct sched_domain *tmp;
5848 /* Remove the sched domains which do not contribute to scheduling. */
5849 for (tmp = sd; tmp; ) {
5850 struct sched_domain *parent = tmp->parent;
5854 if (sd_parent_degenerate(tmp, parent)) {
5855 tmp->parent = parent->parent;
5857 parent->parent->child = tmp;
5858 destroy_sched_domain(parent, cpu);
5863 if (sd && sd_degenerate(sd)) {
5866 destroy_sched_domain(tmp, cpu);
5871 sched_domain_debug(sd, cpu);
5873 rq_attach_root(rq, rd);
5875 rcu_assign_pointer(rq->sd, sd);
5876 destroy_sched_domains(tmp, cpu);
5878 update_top_cache_domain(cpu);
5881 /* cpus with isolated domains */
5882 static cpumask_var_t cpu_isolated_map;
5884 /* Setup the mask of cpus configured for isolated domains */
5885 static int __init isolated_cpu_setup(char *str)
5887 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5888 cpulist_parse(str, cpu_isolated_map);
5892 __setup("isolcpus=", isolated_cpu_setup);
5897 * find_next_best_node - find the next node to include in a sched_domain
5898 * @node: node whose sched_domain we're building
5899 * @used_nodes: nodes already in the sched_domain
5901 * Find the next node to include in a given scheduling domain. Simply
5902 * finds the closest node not already in the @used_nodes map.
5904 * Should use nodemask_t.
5906 static int find_next_best_node(int node, nodemask_t *used_nodes)
5908 int i, n, val, min_val, best_node = -1;
5912 for (i = 0; i < nr_node_ids; i++) {
5913 /* Start at @node */
5914 n = (node + i) % nr_node_ids;
5916 if (!nr_cpus_node(n))
5919 /* Skip already used nodes */
5920 if (node_isset(n, *used_nodes))
5923 /* Simple min distance search */
5924 val = node_distance(node, n);
5926 if (val < min_val) {
5932 if (best_node != -1)
5933 node_set(best_node, *used_nodes);
5938 * sched_domain_node_span - get a cpumask for a node's sched_domain
5939 * @node: node whose cpumask we're constructing
5940 * @span: resulting cpumask
5942 * Given a node, construct a good cpumask for its sched_domain to span. It
5943 * should be one that prevents unnecessary balancing, but also spreads tasks
5946 static void sched_domain_node_span(int node, struct cpumask *span)
5948 nodemask_t used_nodes;
5951 cpumask_clear(span);
5952 nodes_clear(used_nodes);
5954 cpumask_or(span, span, cpumask_of_node(node));
5955 node_set(node, used_nodes);
5957 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5958 int next_node = find_next_best_node(node, &used_nodes);
5961 cpumask_or(span, span, cpumask_of_node(next_node));
5965 static const struct cpumask *cpu_node_mask(int cpu)
5967 lockdep_assert_held(&sched_domains_mutex);
5969 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5971 return sched_domains_tmpmask;
5974 static const struct cpumask *cpu_allnodes_mask(int cpu)
5976 return cpu_possible_mask;
5978 #endif /* CONFIG_NUMA */
5980 static const struct cpumask *cpu_cpu_mask(int cpu)
5982 return cpumask_of_node(cpu_to_node(cpu));
5985 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5988 struct sched_domain **__percpu sd;
5989 struct sched_group **__percpu sg;
5990 struct sched_group_power **__percpu sgp;
5994 struct sched_domain ** __percpu sd;
5995 struct root_domain *rd;
6005 struct sched_domain_topology_level;
6007 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6008 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6010 #define SDTL_OVERLAP 0x01
6012 struct sched_domain_topology_level {
6013 sched_domain_init_f init;
6014 sched_domain_mask_f mask;
6016 struct sd_data data;
6020 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6022 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6023 const struct cpumask *span = sched_domain_span(sd);
6024 struct cpumask *covered = sched_domains_tmpmask;
6025 struct sd_data *sdd = sd->private;
6026 struct sched_domain *child;
6029 cpumask_clear(covered);
6031 for_each_cpu(i, span) {
6032 struct cpumask *sg_span;
6034 if (cpumask_test_cpu(i, covered))
6037 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6038 GFP_KERNEL, cpu_to_node(cpu));
6043 sg_span = sched_group_cpus(sg);
6045 child = *per_cpu_ptr(sdd->sd, i);
6047 child = child->child;
6048 cpumask_copy(sg_span, sched_domain_span(child));
6050 cpumask_set_cpu(i, sg_span);
6052 cpumask_or(covered, covered, sg_span);
6054 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
6055 atomic_inc(&sg->sgp->ref);
6057 if (cpumask_test_cpu(cpu, sg_span))
6067 sd->groups = groups;
6072 free_sched_groups(first, 0);
6077 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6079 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6080 struct sched_domain *child = sd->child;
6083 cpu = cpumask_first(sched_domain_span(child));
6086 *sg = *per_cpu_ptr(sdd->sg, cpu);
6087 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6088 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6095 * build_sched_groups will build a circular linked list of the groups
6096 * covered by the given span, and will set each group's ->cpumask correctly,
6097 * and ->cpu_power to 0.
6099 * Assumes the sched_domain tree is fully constructed
6102 build_sched_groups(struct sched_domain *sd, int cpu)
6104 struct sched_group *first = NULL, *last = NULL;
6105 struct sd_data *sdd = sd->private;
6106 const struct cpumask *span = sched_domain_span(sd);
6107 struct cpumask *covered;
6110 get_group(cpu, sdd, &sd->groups);
6111 atomic_inc(&sd->groups->ref);
6113 if (cpu != cpumask_first(sched_domain_span(sd)))
6116 lockdep_assert_held(&sched_domains_mutex);
6117 covered = sched_domains_tmpmask;
6119 cpumask_clear(covered);
6121 for_each_cpu(i, span) {
6122 struct sched_group *sg;
6123 int group = get_group(i, sdd, &sg);
6126 if (cpumask_test_cpu(i, covered))
6129 cpumask_clear(sched_group_cpus(sg));
6132 for_each_cpu(j, span) {
6133 if (get_group(j, sdd, NULL) != group)
6136 cpumask_set_cpu(j, covered);
6137 cpumask_set_cpu(j, sched_group_cpus(sg));
6152 * Initialize sched groups cpu_power.
6154 * cpu_power indicates the capacity of sched group, which is used while
6155 * distributing the load between different sched groups in a sched domain.
6156 * Typically cpu_power for all the groups in a sched domain will be same unless
6157 * there are asymmetries in the topology. If there are asymmetries, group
6158 * having more cpu_power will pickup more load compared to the group having
6161 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6163 struct sched_group *sg = sd->groups;
6165 WARN_ON(!sd || !sg);
6168 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6170 } while (sg != sd->groups);
6172 if (cpu != group_first_cpu(sg))
6175 update_group_power(sd, cpu);
6176 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6179 int __weak arch_sd_sibling_asym_packing(void)
6181 return 0*SD_ASYM_PACKING;
6185 * Initializers for schedule domains
6186 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6189 #ifdef CONFIG_SCHED_DEBUG
6190 # define SD_INIT_NAME(sd, type) sd->name = #type
6192 # define SD_INIT_NAME(sd, type) do { } while (0)
6195 #define SD_INIT_FUNC(type) \
6196 static noinline struct sched_domain * \
6197 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6199 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6200 *sd = SD_##type##_INIT; \
6201 SD_INIT_NAME(sd, type); \
6202 sd->private = &tl->data; \
6208 SD_INIT_FUNC(ALLNODES)
6211 #ifdef CONFIG_SCHED_SMT
6212 SD_INIT_FUNC(SIBLING)
6214 #ifdef CONFIG_SCHED_MC
6217 #ifdef CONFIG_SCHED_BOOK
6221 static int default_relax_domain_level = -1;
6222 int sched_domain_level_max;
6224 static int __init setup_relax_domain_level(char *str)
6228 val = simple_strtoul(str, NULL, 0);
6229 if (val < sched_domain_level_max)
6230 default_relax_domain_level = val;
6234 __setup("relax_domain_level=", setup_relax_domain_level);
6236 static void set_domain_attribute(struct sched_domain *sd,
6237 struct sched_domain_attr *attr)
6241 if (!attr || attr->relax_domain_level < 0) {
6242 if (default_relax_domain_level < 0)
6245 request = default_relax_domain_level;
6247 request = attr->relax_domain_level;
6248 if (request < sd->level) {
6249 /* turn off idle balance on this domain */
6250 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6252 /* turn on idle balance on this domain */
6253 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6257 static void __sdt_free(const struct cpumask *cpu_map);
6258 static int __sdt_alloc(const struct cpumask *cpu_map);
6260 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6261 const struct cpumask *cpu_map)
6265 if (!atomic_read(&d->rd->refcount))
6266 free_rootdomain(&d->rd->rcu); /* fall through */
6268 free_percpu(d->sd); /* fall through */
6270 __sdt_free(cpu_map); /* fall through */
6276 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6277 const struct cpumask *cpu_map)
6279 memset(d, 0, sizeof(*d));
6281 if (__sdt_alloc(cpu_map))
6282 return sa_sd_storage;
6283 d->sd = alloc_percpu(struct sched_domain *);
6285 return sa_sd_storage;
6286 d->rd = alloc_rootdomain();
6289 return sa_rootdomain;
6293 * NULL the sd_data elements we've used to build the sched_domain and
6294 * sched_group structure so that the subsequent __free_domain_allocs()
6295 * will not free the data we're using.
6297 static void claim_allocations(int cpu, struct sched_domain *sd)
6299 struct sd_data *sdd = sd->private;
6301 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6302 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6304 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6305 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6307 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6308 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6311 #ifdef CONFIG_SCHED_SMT
6312 static const struct cpumask *cpu_smt_mask(int cpu)
6314 return topology_thread_cpumask(cpu);
6319 * Topology list, bottom-up.
6321 static struct sched_domain_topology_level default_topology[] = {
6322 #ifdef CONFIG_SCHED_SMT
6323 { sd_init_SIBLING, cpu_smt_mask, },
6325 #ifdef CONFIG_SCHED_MC
6326 { sd_init_MC, cpu_coregroup_mask, },
6328 #ifdef CONFIG_SCHED_BOOK
6329 { sd_init_BOOK, cpu_book_mask, },
6331 { sd_init_CPU, cpu_cpu_mask, },
6333 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6334 { sd_init_ALLNODES, cpu_allnodes_mask, },
6339 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6341 static int __sdt_alloc(const struct cpumask *cpu_map)
6343 struct sched_domain_topology_level *tl;
6346 for (tl = sched_domain_topology; tl->init; tl++) {
6347 struct sd_data *sdd = &tl->data;
6349 sdd->sd = alloc_percpu(struct sched_domain *);
6353 sdd->sg = alloc_percpu(struct sched_group *);
6357 sdd->sgp = alloc_percpu(struct sched_group_power *);
6361 for_each_cpu(j, cpu_map) {
6362 struct sched_domain *sd;
6363 struct sched_group *sg;
6364 struct sched_group_power *sgp;
6366 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6367 GFP_KERNEL, cpu_to_node(j));
6371 *per_cpu_ptr(sdd->sd, j) = sd;
6373 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6374 GFP_KERNEL, cpu_to_node(j));
6378 *per_cpu_ptr(sdd->sg, j) = sg;
6380 sgp = kzalloc_node(sizeof(struct sched_group_power),
6381 GFP_KERNEL, cpu_to_node(j));
6385 *per_cpu_ptr(sdd->sgp, j) = sgp;
6392 static void __sdt_free(const struct cpumask *cpu_map)
6394 struct sched_domain_topology_level *tl;
6397 for (tl = sched_domain_topology; tl->init; tl++) {
6398 struct sd_data *sdd = &tl->data;
6400 for_each_cpu(j, cpu_map) {
6401 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6402 if (sd && (sd->flags & SD_OVERLAP))
6403 free_sched_groups(sd->groups, 0);
6404 kfree(*per_cpu_ptr(sdd->sd, j));
6405 kfree(*per_cpu_ptr(sdd->sg, j));
6406 kfree(*per_cpu_ptr(sdd->sgp, j));
6408 free_percpu(sdd->sd);
6409 free_percpu(sdd->sg);
6410 free_percpu(sdd->sgp);
6414 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6415 struct s_data *d, const struct cpumask *cpu_map,
6416 struct sched_domain_attr *attr, struct sched_domain *child,
6419 struct sched_domain *sd = tl->init(tl, cpu);
6423 set_domain_attribute(sd, attr);
6424 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6426 sd->level = child->level + 1;
6427 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6436 * Build sched domains for a given set of cpus and attach the sched domains
6437 * to the individual cpus
6439 static int build_sched_domains(const struct cpumask *cpu_map,
6440 struct sched_domain_attr *attr)
6442 enum s_alloc alloc_state = sa_none;
6443 struct sched_domain *sd;
6445 int i, ret = -ENOMEM;
6447 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6448 if (alloc_state != sa_rootdomain)
6451 /* Set up domains for cpus specified by the cpu_map. */
6452 for_each_cpu(i, cpu_map) {
6453 struct sched_domain_topology_level *tl;
6456 for (tl = sched_domain_topology; tl->init; tl++) {
6457 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6458 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6459 sd->flags |= SD_OVERLAP;
6460 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6467 *per_cpu_ptr(d.sd, i) = sd;
6470 /* Build the groups for the domains */
6471 for_each_cpu(i, cpu_map) {
6472 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6473 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6474 if (sd->flags & SD_OVERLAP) {
6475 if (build_overlap_sched_groups(sd, i))
6478 if (build_sched_groups(sd, i))
6484 /* Calculate CPU power for physical packages and nodes */
6485 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6486 if (!cpumask_test_cpu(i, cpu_map))
6489 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6490 claim_allocations(i, sd);
6491 init_sched_groups_power(i, sd);
6495 /* Attach the domains */
6497 for_each_cpu(i, cpu_map) {
6498 sd = *per_cpu_ptr(d.sd, i);
6499 cpu_attach_domain(sd, d.rd, i);
6505 __free_domain_allocs(&d, alloc_state, cpu_map);
6509 static cpumask_var_t *doms_cur; /* current sched domains */
6510 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6511 static struct sched_domain_attr *dattr_cur;
6512 /* attribues of custom domains in 'doms_cur' */
6515 * Special case: If a kmalloc of a doms_cur partition (array of
6516 * cpumask) fails, then fallback to a single sched domain,
6517 * as determined by the single cpumask fallback_doms.
6519 static cpumask_var_t fallback_doms;
6522 * arch_update_cpu_topology lets virtualized architectures update the
6523 * cpu core maps. It is supposed to return 1 if the topology changed
6524 * or 0 if it stayed the same.
6526 int __attribute__((weak)) arch_update_cpu_topology(void)
6531 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6534 cpumask_var_t *doms;
6536 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6539 for (i = 0; i < ndoms; i++) {
6540 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6541 free_sched_domains(doms, i);
6548 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6551 for (i = 0; i < ndoms; i++)
6552 free_cpumask_var(doms[i]);
6557 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6558 * For now this just excludes isolated cpus, but could be used to
6559 * exclude other special cases in the future.
6561 static int init_sched_domains(const struct cpumask *cpu_map)
6565 arch_update_cpu_topology();
6567 doms_cur = alloc_sched_domains(ndoms_cur);
6569 doms_cur = &fallback_doms;
6570 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6572 err = build_sched_domains(doms_cur[0], NULL);
6573 register_sched_domain_sysctl();
6579 * Detach sched domains from a group of cpus specified in cpu_map
6580 * These cpus will now be attached to the NULL domain
6582 static void detach_destroy_domains(const struct cpumask *cpu_map)
6587 for_each_cpu(i, cpu_map)
6588 cpu_attach_domain(NULL, &def_root_domain, i);
6592 /* handle null as "default" */
6593 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6594 struct sched_domain_attr *new, int idx_new)
6596 struct sched_domain_attr tmp;
6603 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6604 new ? (new + idx_new) : &tmp,
6605 sizeof(struct sched_domain_attr));
6609 * Partition sched domains as specified by the 'ndoms_new'
6610 * cpumasks in the array doms_new[] of cpumasks. This compares
6611 * doms_new[] to the current sched domain partitioning, doms_cur[].
6612 * It destroys each deleted domain and builds each new domain.
6614 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6615 * The masks don't intersect (don't overlap.) We should setup one
6616 * sched domain for each mask. CPUs not in any of the cpumasks will
6617 * not be load balanced. If the same cpumask appears both in the
6618 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6621 * The passed in 'doms_new' should be allocated using
6622 * alloc_sched_domains. This routine takes ownership of it and will
6623 * free_sched_domains it when done with it. If the caller failed the
6624 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6625 * and partition_sched_domains() will fallback to the single partition
6626 * 'fallback_doms', it also forces the domains to be rebuilt.
6628 * If doms_new == NULL it will be replaced with cpu_online_mask.
6629 * ndoms_new == 0 is a special case for destroying existing domains,
6630 * and it will not create the default domain.
6632 * Call with hotplug lock held
6634 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6635 struct sched_domain_attr *dattr_new)
6640 mutex_lock(&sched_domains_mutex);
6642 /* always unregister in case we don't destroy any domains */
6643 unregister_sched_domain_sysctl();
6645 /* Let architecture update cpu core mappings. */
6646 new_topology = arch_update_cpu_topology();
6648 n = doms_new ? ndoms_new : 0;
6650 /* Destroy deleted domains */
6651 for (i = 0; i < ndoms_cur; i++) {
6652 for (j = 0; j < n && !new_topology; j++) {
6653 if (cpumask_equal(doms_cur[i], doms_new[j])
6654 && dattrs_equal(dattr_cur, i, dattr_new, j))
6657 /* no match - a current sched domain not in new doms_new[] */
6658 detach_destroy_domains(doms_cur[i]);
6663 if (doms_new == NULL) {
6665 doms_new = &fallback_doms;
6666 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6667 WARN_ON_ONCE(dattr_new);
6670 /* Build new domains */
6671 for (i = 0; i < ndoms_new; i++) {
6672 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6673 if (cpumask_equal(doms_new[i], doms_cur[j])
6674 && dattrs_equal(dattr_new, i, dattr_cur, j))
6677 /* no match - add a new doms_new */
6678 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6683 /* Remember the new sched domains */
6684 if (doms_cur != &fallback_doms)
6685 free_sched_domains(doms_cur, ndoms_cur);
6686 kfree(dattr_cur); /* kfree(NULL) is safe */
6687 doms_cur = doms_new;
6688 dattr_cur = dattr_new;
6689 ndoms_cur = ndoms_new;
6691 register_sched_domain_sysctl();
6693 mutex_unlock(&sched_domains_mutex);
6696 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6697 static void reinit_sched_domains(void)
6701 /* Destroy domains first to force the rebuild */
6702 partition_sched_domains(0, NULL, NULL);
6704 rebuild_sched_domains();
6708 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6710 unsigned int level = 0;
6712 if (sscanf(buf, "%u", &level) != 1)
6716 * level is always be positive so don't check for
6717 * level < POWERSAVINGS_BALANCE_NONE which is 0
6718 * What happens on 0 or 1 byte write,
6719 * need to check for count as well?
6722 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6726 sched_smt_power_savings = level;
6728 sched_mc_power_savings = level;
6730 reinit_sched_domains();
6735 #ifdef CONFIG_SCHED_MC
6736 static ssize_t sched_mc_power_savings_show(struct device *dev,
6737 struct device_attribute *attr,
6740 return sprintf(buf, "%u\n", sched_mc_power_savings);
6742 static ssize_t sched_mc_power_savings_store(struct device *dev,
6743 struct device_attribute *attr,
6744 const char *buf, size_t count)
6746 return sched_power_savings_store(buf, count, 0);
6748 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6749 sched_mc_power_savings_show,
6750 sched_mc_power_savings_store);
6753 #ifdef CONFIG_SCHED_SMT
6754 static ssize_t sched_smt_power_savings_show(struct device *dev,
6755 struct device_attribute *attr,
6758 return sprintf(buf, "%u\n", sched_smt_power_savings);
6760 static ssize_t sched_smt_power_savings_store(struct device *dev,
6761 struct device_attribute *attr,
6762 const char *buf, size_t count)
6764 return sched_power_savings_store(buf, count, 1);
6766 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6767 sched_smt_power_savings_show,
6768 sched_smt_power_savings_store);
6771 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6775 #ifdef CONFIG_SCHED_SMT
6777 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6779 #ifdef CONFIG_SCHED_MC
6780 if (!err && mc_capable())
6781 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6785 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6788 * Update cpusets according to cpu_active mask. If cpusets are
6789 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6790 * around partition_sched_domains().
6792 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6795 switch (action & ~CPU_TASKS_FROZEN) {
6797 case CPU_DOWN_FAILED:
6798 cpuset_update_active_cpus();
6805 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6808 switch (action & ~CPU_TASKS_FROZEN) {
6809 case CPU_DOWN_PREPARE:
6810 cpuset_update_active_cpus();
6817 void __init sched_init_smp(void)
6819 cpumask_var_t non_isolated_cpus;
6821 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6822 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6825 mutex_lock(&sched_domains_mutex);
6826 init_sched_domains(cpu_active_mask);
6827 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6828 if (cpumask_empty(non_isolated_cpus))
6829 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6830 mutex_unlock(&sched_domains_mutex);
6833 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6834 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6836 /* RT runtime code needs to handle some hotplug events */
6837 hotcpu_notifier(update_runtime, 0);
6841 /* Move init over to a non-isolated CPU */
6842 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6844 sched_init_granularity();
6845 free_cpumask_var(non_isolated_cpus);
6847 init_sched_rt_class();
6850 void __init sched_init_smp(void)
6852 sched_init_granularity();
6854 #endif /* CONFIG_SMP */
6856 const_debug unsigned int sysctl_timer_migration = 1;
6858 int in_sched_functions(unsigned long addr)
6860 return in_lock_functions(addr) ||
6861 (addr >= (unsigned long)__sched_text_start
6862 && addr < (unsigned long)__sched_text_end);
6865 #ifdef CONFIG_CGROUP_SCHED
6866 struct task_group root_task_group;
6869 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6871 void __init sched_init(void)
6874 unsigned long alloc_size = 0, ptr;
6876 #ifdef CONFIG_FAIR_GROUP_SCHED
6877 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6879 #ifdef CONFIG_RT_GROUP_SCHED
6880 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6882 #ifdef CONFIG_CPUMASK_OFFSTACK
6883 alloc_size += num_possible_cpus() * cpumask_size();
6886 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6888 #ifdef CONFIG_FAIR_GROUP_SCHED
6889 root_task_group.se = (struct sched_entity **)ptr;
6890 ptr += nr_cpu_ids * sizeof(void **);
6892 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6893 ptr += nr_cpu_ids * sizeof(void **);
6895 #endif /* CONFIG_FAIR_GROUP_SCHED */
6896 #ifdef CONFIG_RT_GROUP_SCHED
6897 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6898 ptr += nr_cpu_ids * sizeof(void **);
6900 root_task_group.rt_rq = (struct rt_rq **)ptr;
6901 ptr += nr_cpu_ids * sizeof(void **);
6903 #endif /* CONFIG_RT_GROUP_SCHED */
6904 #ifdef CONFIG_CPUMASK_OFFSTACK
6905 for_each_possible_cpu(i) {
6906 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6907 ptr += cpumask_size();
6909 #endif /* CONFIG_CPUMASK_OFFSTACK */
6913 init_defrootdomain();
6916 init_rt_bandwidth(&def_rt_bandwidth,
6917 global_rt_period(), global_rt_runtime());
6919 #ifdef CONFIG_RT_GROUP_SCHED
6920 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6921 global_rt_period(), global_rt_runtime());
6922 #endif /* CONFIG_RT_GROUP_SCHED */
6924 #ifdef CONFIG_CGROUP_SCHED
6925 list_add(&root_task_group.list, &task_groups);
6926 INIT_LIST_HEAD(&root_task_group.children);
6927 INIT_LIST_HEAD(&root_task_group.siblings);
6928 autogroup_init(&init_task);
6930 #endif /* CONFIG_CGROUP_SCHED */
6932 #ifdef CONFIG_CGROUP_CPUACCT
6933 root_cpuacct.cpustat = &kernel_cpustat;
6934 root_cpuacct.cpuusage = alloc_percpu(u64);
6935 /* Too early, not expected to fail */
6936 BUG_ON(!root_cpuacct.cpuusage);
6938 for_each_possible_cpu(i) {
6942 raw_spin_lock_init(&rq->lock);
6944 rq->calc_load_active = 0;
6945 rq->calc_load_update = jiffies + LOAD_FREQ;
6946 init_cfs_rq(&rq->cfs);
6947 init_rt_rq(&rq->rt, rq);
6948 #ifdef CONFIG_FAIR_GROUP_SCHED
6949 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6950 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6952 * How much cpu bandwidth does root_task_group get?
6954 * In case of task-groups formed thr' the cgroup filesystem, it
6955 * gets 100% of the cpu resources in the system. This overall
6956 * system cpu resource is divided among the tasks of
6957 * root_task_group and its child task-groups in a fair manner,
6958 * based on each entity's (task or task-group's) weight
6959 * (se->load.weight).
6961 * In other words, if root_task_group has 10 tasks of weight
6962 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6963 * then A0's share of the cpu resource is:
6965 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6967 * We achieve this by letting root_task_group's tasks sit
6968 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6970 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6971 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6972 #endif /* CONFIG_FAIR_GROUP_SCHED */
6974 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6975 #ifdef CONFIG_RT_GROUP_SCHED
6976 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6977 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6980 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6981 rq->cpu_load[j] = 0;
6983 rq->last_load_update_tick = jiffies;
6988 rq->cpu_power = SCHED_POWER_SCALE;
6989 rq->post_schedule = 0;
6990 rq->active_balance = 0;
6991 rq->next_balance = jiffies;
6996 rq->avg_idle = 2*sysctl_sched_migration_cost;
6998 INIT_LIST_HEAD(&rq->cfs_tasks);
7000 rq_attach_root(rq, &def_root_domain);
7006 atomic_set(&rq->nr_iowait, 0);
7009 set_load_weight(&init_task);
7011 #ifdef CONFIG_PREEMPT_NOTIFIERS
7012 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7015 #ifdef CONFIG_RT_MUTEXES
7016 plist_head_init(&init_task.pi_waiters);
7020 * The boot idle thread does lazy MMU switching as well:
7022 atomic_inc(&init_mm.mm_count);
7023 enter_lazy_tlb(&init_mm, current);
7026 * Make us the idle thread. Technically, schedule() should not be
7027 * called from this thread, however somewhere below it might be,
7028 * but because we are the idle thread, we just pick up running again
7029 * when this runqueue becomes "idle".
7031 init_idle(current, smp_processor_id());
7033 calc_load_update = jiffies + LOAD_FREQ;
7036 * During early bootup we pretend to be a normal task:
7038 current->sched_class = &fair_sched_class;
7041 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7042 /* May be allocated at isolcpus cmdline parse time */
7043 if (cpu_isolated_map == NULL)
7044 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7046 init_sched_fair_class();
7048 scheduler_running = 1;
7051 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7052 static inline int preempt_count_equals(int preempt_offset)
7054 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7056 return (nested == preempt_offset);
7059 void __might_sleep(const char *file, int line, int preempt_offset)
7061 static unsigned long prev_jiffy; /* ratelimiting */
7063 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7064 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7065 system_state != SYSTEM_RUNNING || oops_in_progress)
7067 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7069 prev_jiffy = jiffies;
7072 "BUG: sleeping function called from invalid context at %s:%d\n",
7075 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7076 in_atomic(), irqs_disabled(),
7077 current->pid, current->comm);
7079 debug_show_held_locks(current);
7080 if (irqs_disabled())
7081 print_irqtrace_events(current);
7084 EXPORT_SYMBOL(__might_sleep);
7087 #ifdef CONFIG_MAGIC_SYSRQ
7088 static void normalize_task(struct rq *rq, struct task_struct *p)
7090 const struct sched_class *prev_class = p->sched_class;
7091 int old_prio = p->prio;
7096 dequeue_task(rq, p, 0);
7097 __setscheduler(rq, p, SCHED_NORMAL, 0);
7099 enqueue_task(rq, p, 0);
7100 resched_task(rq->curr);
7103 check_class_changed(rq, p, prev_class, old_prio);
7106 void normalize_rt_tasks(void)
7108 struct task_struct *g, *p;
7109 unsigned long flags;
7112 read_lock_irqsave(&tasklist_lock, flags);
7113 do_each_thread(g, p) {
7115 * Only normalize user tasks:
7120 p->se.exec_start = 0;
7121 #ifdef CONFIG_SCHEDSTATS
7122 p->se.statistics.wait_start = 0;
7123 p->se.statistics.sleep_start = 0;
7124 p->se.statistics.block_start = 0;
7129 * Renice negative nice level userspace
7132 if (TASK_NICE(p) < 0 && p->mm)
7133 set_user_nice(p, 0);
7137 raw_spin_lock(&p->pi_lock);
7138 rq = __task_rq_lock(p);
7140 normalize_task(rq, p);
7142 __task_rq_unlock(rq);
7143 raw_spin_unlock(&p->pi_lock);
7144 } while_each_thread(g, p);
7146 read_unlock_irqrestore(&tasklist_lock, flags);
7149 #endif /* CONFIG_MAGIC_SYSRQ */
7151 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7153 * These functions are only useful for the IA64 MCA handling, or kdb.
7155 * They can only be called when the whole system has been
7156 * stopped - every CPU needs to be quiescent, and no scheduling
7157 * activity can take place. Using them for anything else would
7158 * be a serious bug, and as a result, they aren't even visible
7159 * under any other configuration.
7163 * curr_task - return the current task for a given cpu.
7164 * @cpu: the processor in question.
7166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7168 struct task_struct *curr_task(int cpu)
7170 return cpu_curr(cpu);
7173 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7177 * set_curr_task - set the current task for a given cpu.
7178 * @cpu: the processor in question.
7179 * @p: the task pointer to set.
7181 * Description: This function must only be used when non-maskable interrupts
7182 * are serviced on a separate stack. It allows the architecture to switch the
7183 * notion of the current task on a cpu in a non-blocking manner. This function
7184 * must be called with all CPU's synchronized, and interrupts disabled, the
7185 * and caller must save the original value of the current task (see
7186 * curr_task() above) and restore that value before reenabling interrupts and
7187 * re-starting the system.
7189 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7191 void set_curr_task(int cpu, struct task_struct *p)
7198 #ifdef CONFIG_CGROUP_SCHED
7199 /* task_group_lock serializes the addition/removal of task groups */
7200 static DEFINE_SPINLOCK(task_group_lock);
7202 static void free_sched_group(struct task_group *tg)
7204 free_fair_sched_group(tg);
7205 free_rt_sched_group(tg);
7210 /* allocate runqueue etc for a new task group */
7211 struct task_group *sched_create_group(struct task_group *parent)
7213 struct task_group *tg;
7214 unsigned long flags;
7216 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7218 return ERR_PTR(-ENOMEM);
7220 if (!alloc_fair_sched_group(tg, parent))
7223 if (!alloc_rt_sched_group(tg, parent))
7226 spin_lock_irqsave(&task_group_lock, flags);
7227 list_add_rcu(&tg->list, &task_groups);
7229 WARN_ON(!parent); /* root should already exist */
7231 tg->parent = parent;
7232 INIT_LIST_HEAD(&tg->children);
7233 list_add_rcu(&tg->siblings, &parent->children);
7234 spin_unlock_irqrestore(&task_group_lock, flags);
7239 free_sched_group(tg);
7240 return ERR_PTR(-ENOMEM);
7243 /* rcu callback to free various structures associated with a task group */
7244 static void free_sched_group_rcu(struct rcu_head *rhp)
7246 /* now it should be safe to free those cfs_rqs */
7247 free_sched_group(container_of(rhp, struct task_group, rcu));
7250 /* Destroy runqueue etc associated with a task group */
7251 void sched_destroy_group(struct task_group *tg)
7253 unsigned long flags;
7256 /* end participation in shares distribution */
7257 for_each_possible_cpu(i)
7258 unregister_fair_sched_group(tg, i);
7260 spin_lock_irqsave(&task_group_lock, flags);
7261 list_del_rcu(&tg->list);
7262 list_del_rcu(&tg->siblings);
7263 spin_unlock_irqrestore(&task_group_lock, flags);
7265 /* wait for possible concurrent references to cfs_rqs complete */
7266 call_rcu(&tg->rcu, free_sched_group_rcu);
7269 /* change task's runqueue when it moves between groups.
7270 * The caller of this function should have put the task in its new group
7271 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7272 * reflect its new group.
7274 void sched_move_task(struct task_struct *tsk)
7277 unsigned long flags;
7280 rq = task_rq_lock(tsk, &flags);
7282 running = task_current(rq, tsk);
7286 dequeue_task(rq, tsk, 0);
7287 if (unlikely(running))
7288 tsk->sched_class->put_prev_task(rq, tsk);
7290 #ifdef CONFIG_FAIR_GROUP_SCHED
7291 if (tsk->sched_class->task_move_group)
7292 tsk->sched_class->task_move_group(tsk, on_rq);
7295 set_task_rq(tsk, task_cpu(tsk));
7297 if (unlikely(running))
7298 tsk->sched_class->set_curr_task(rq);
7300 enqueue_task(rq, tsk, 0);
7302 task_rq_unlock(rq, tsk, &flags);
7304 #endif /* CONFIG_CGROUP_SCHED */
7306 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7307 static unsigned long to_ratio(u64 period, u64 runtime)
7309 if (runtime == RUNTIME_INF)
7312 return div64_u64(runtime << 20, period);
7316 #ifdef CONFIG_RT_GROUP_SCHED
7318 * Ensure that the real time constraints are schedulable.
7320 static DEFINE_MUTEX(rt_constraints_mutex);
7322 /* Must be called with tasklist_lock held */
7323 static inline int tg_has_rt_tasks(struct task_group *tg)
7325 struct task_struct *g, *p;
7327 do_each_thread(g, p) {
7328 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7330 } while_each_thread(g, p);
7335 struct rt_schedulable_data {
7336 struct task_group *tg;
7341 static int tg_rt_schedulable(struct task_group *tg, void *data)
7343 struct rt_schedulable_data *d = data;
7344 struct task_group *child;
7345 unsigned long total, sum = 0;
7346 u64 period, runtime;
7348 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7349 runtime = tg->rt_bandwidth.rt_runtime;
7352 period = d->rt_period;
7353 runtime = d->rt_runtime;
7357 * Cannot have more runtime than the period.
7359 if (runtime > period && runtime != RUNTIME_INF)
7363 * Ensure we don't starve existing RT tasks.
7365 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7368 total = to_ratio(period, runtime);
7371 * Nobody can have more than the global setting allows.
7373 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7377 * The sum of our children's runtime should not exceed our own.
7379 list_for_each_entry_rcu(child, &tg->children, siblings) {
7380 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7381 runtime = child->rt_bandwidth.rt_runtime;
7383 if (child == d->tg) {
7384 period = d->rt_period;
7385 runtime = d->rt_runtime;
7388 sum += to_ratio(period, runtime);
7397 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7401 struct rt_schedulable_data data = {
7403 .rt_period = period,
7404 .rt_runtime = runtime,
7408 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7414 static int tg_set_rt_bandwidth(struct task_group *tg,
7415 u64 rt_period, u64 rt_runtime)
7419 mutex_lock(&rt_constraints_mutex);
7420 read_lock(&tasklist_lock);
7421 err = __rt_schedulable(tg, rt_period, rt_runtime);
7425 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7426 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7427 tg->rt_bandwidth.rt_runtime = rt_runtime;
7429 for_each_possible_cpu(i) {
7430 struct rt_rq *rt_rq = tg->rt_rq[i];
7432 raw_spin_lock(&rt_rq->rt_runtime_lock);
7433 rt_rq->rt_runtime = rt_runtime;
7434 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7436 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7438 read_unlock(&tasklist_lock);
7439 mutex_unlock(&rt_constraints_mutex);
7444 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7446 u64 rt_runtime, rt_period;
7448 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7449 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7450 if (rt_runtime_us < 0)
7451 rt_runtime = RUNTIME_INF;
7453 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7456 long sched_group_rt_runtime(struct task_group *tg)
7460 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7463 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7464 do_div(rt_runtime_us, NSEC_PER_USEC);
7465 return rt_runtime_us;
7468 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7470 u64 rt_runtime, rt_period;
7472 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7473 rt_runtime = tg->rt_bandwidth.rt_runtime;
7478 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7481 long sched_group_rt_period(struct task_group *tg)
7485 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7486 do_div(rt_period_us, NSEC_PER_USEC);
7487 return rt_period_us;
7490 static int sched_rt_global_constraints(void)
7492 u64 runtime, period;
7495 if (sysctl_sched_rt_period <= 0)
7498 runtime = global_rt_runtime();
7499 period = global_rt_period();
7502 * Sanity check on the sysctl variables.
7504 if (runtime > period && runtime != RUNTIME_INF)
7507 mutex_lock(&rt_constraints_mutex);
7508 read_lock(&tasklist_lock);
7509 ret = __rt_schedulable(NULL, 0, 0);
7510 read_unlock(&tasklist_lock);
7511 mutex_unlock(&rt_constraints_mutex);
7516 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7518 /* Don't accept realtime tasks when there is no way for them to run */
7519 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7525 #else /* !CONFIG_RT_GROUP_SCHED */
7526 static int sched_rt_global_constraints(void)
7528 unsigned long flags;
7531 if (sysctl_sched_rt_period <= 0)
7535 * There's always some RT tasks in the root group
7536 * -- migration, kstopmachine etc..
7538 if (sysctl_sched_rt_runtime == 0)
7541 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7542 for_each_possible_cpu(i) {
7543 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7545 raw_spin_lock(&rt_rq->rt_runtime_lock);
7546 rt_rq->rt_runtime = global_rt_runtime();
7547 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7549 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7553 #endif /* CONFIG_RT_GROUP_SCHED */
7555 int sched_rt_handler(struct ctl_table *table, int write,
7556 void __user *buffer, size_t *lenp,
7560 int old_period, old_runtime;
7561 static DEFINE_MUTEX(mutex);
7564 old_period = sysctl_sched_rt_period;
7565 old_runtime = sysctl_sched_rt_runtime;
7567 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7569 if (!ret && write) {
7570 ret = sched_rt_global_constraints();
7572 sysctl_sched_rt_period = old_period;
7573 sysctl_sched_rt_runtime = old_runtime;
7575 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7576 def_rt_bandwidth.rt_period =
7577 ns_to_ktime(global_rt_period());
7580 mutex_unlock(&mutex);
7585 #ifdef CONFIG_CGROUP_SCHED
7587 /* return corresponding task_group object of a cgroup */
7588 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7590 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7591 struct task_group, css);
7594 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7596 struct task_group *tg, *parent;
7598 if (!cgrp->parent) {
7599 /* This is early initialization for the top cgroup */
7600 return &root_task_group.css;
7603 parent = cgroup_tg(cgrp->parent);
7604 tg = sched_create_group(parent);
7606 return ERR_PTR(-ENOMEM);
7611 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7613 struct task_group *tg = cgroup_tg(cgrp);
7615 sched_destroy_group(tg);
7618 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7619 struct cgroup_taskset *tset)
7621 struct task_struct *task;
7623 cgroup_taskset_for_each(task, cgrp, tset) {
7624 #ifdef CONFIG_RT_GROUP_SCHED
7625 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7628 /* We don't support RT-tasks being in separate groups */
7629 if (task->sched_class != &fair_sched_class)
7636 static void cpu_cgroup_attach(struct cgroup *cgrp,
7637 struct cgroup_taskset *tset)
7639 struct task_struct *task;
7641 cgroup_taskset_for_each(task, cgrp, tset)
7642 sched_move_task(task);
7646 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7647 struct task_struct *task)
7650 * cgroup_exit() is called in the copy_process() failure path.
7651 * Ignore this case since the task hasn't ran yet, this avoids
7652 * trying to poke a half freed task state from generic code.
7654 if (!(task->flags & PF_EXITING))
7657 sched_move_task(task);
7660 #ifdef CONFIG_FAIR_GROUP_SCHED
7661 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7664 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7667 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7669 struct task_group *tg = cgroup_tg(cgrp);
7671 return (u64) scale_load_down(tg->shares);
7674 #ifdef CONFIG_CFS_BANDWIDTH
7675 static DEFINE_MUTEX(cfs_constraints_mutex);
7677 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7678 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7680 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7682 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7684 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7685 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7687 if (tg == &root_task_group)
7691 * Ensure we have at some amount of bandwidth every period. This is
7692 * to prevent reaching a state of large arrears when throttled via
7693 * entity_tick() resulting in prolonged exit starvation.
7695 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7699 * Likewise, bound things on the otherside by preventing insane quota
7700 * periods. This also allows us to normalize in computing quota
7703 if (period > max_cfs_quota_period)
7706 mutex_lock(&cfs_constraints_mutex);
7707 ret = __cfs_schedulable(tg, period, quota);
7711 runtime_enabled = quota != RUNTIME_INF;
7712 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7713 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7714 raw_spin_lock_irq(&cfs_b->lock);
7715 cfs_b->period = ns_to_ktime(period);
7716 cfs_b->quota = quota;
7718 __refill_cfs_bandwidth_runtime(cfs_b);
7719 /* restart the period timer (if active) to handle new period expiry */
7720 if (runtime_enabled && cfs_b->timer_active) {
7721 /* force a reprogram */
7722 cfs_b->timer_active = 0;
7723 __start_cfs_bandwidth(cfs_b);
7725 raw_spin_unlock_irq(&cfs_b->lock);
7727 for_each_possible_cpu(i) {
7728 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7729 struct rq *rq = cfs_rq->rq;
7731 raw_spin_lock_irq(&rq->lock);
7732 cfs_rq->runtime_enabled = runtime_enabled;
7733 cfs_rq->runtime_remaining = 0;
7735 if (cfs_rq->throttled)
7736 unthrottle_cfs_rq(cfs_rq);
7737 raw_spin_unlock_irq(&rq->lock);
7740 mutex_unlock(&cfs_constraints_mutex);
7745 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7749 period = ktime_to_ns(tg->cfs_bandwidth.period);
7750 if (cfs_quota_us < 0)
7751 quota = RUNTIME_INF;
7753 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7755 return tg_set_cfs_bandwidth(tg, period, quota);
7758 long tg_get_cfs_quota(struct task_group *tg)
7762 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7765 quota_us = tg->cfs_bandwidth.quota;
7766 do_div(quota_us, NSEC_PER_USEC);
7771 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7775 period = (u64)cfs_period_us * NSEC_PER_USEC;
7776 quota = tg->cfs_bandwidth.quota;
7778 return tg_set_cfs_bandwidth(tg, period, quota);
7781 long tg_get_cfs_period(struct task_group *tg)
7785 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7786 do_div(cfs_period_us, NSEC_PER_USEC);
7788 return cfs_period_us;
7791 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7793 return tg_get_cfs_quota(cgroup_tg(cgrp));
7796 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7799 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7802 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7804 return tg_get_cfs_period(cgroup_tg(cgrp));
7807 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7810 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7813 struct cfs_schedulable_data {
7814 struct task_group *tg;
7819 * normalize group quota/period to be quota/max_period
7820 * note: units are usecs
7822 static u64 normalize_cfs_quota(struct task_group *tg,
7823 struct cfs_schedulable_data *d)
7831 period = tg_get_cfs_period(tg);
7832 quota = tg_get_cfs_quota(tg);
7835 /* note: these should typically be equivalent */
7836 if (quota == RUNTIME_INF || quota == -1)
7839 return to_ratio(period, quota);
7842 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7844 struct cfs_schedulable_data *d = data;
7845 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7846 s64 quota = 0, parent_quota = -1;
7849 quota = RUNTIME_INF;
7851 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7853 quota = normalize_cfs_quota(tg, d);
7854 parent_quota = parent_b->hierarchal_quota;
7857 * ensure max(child_quota) <= parent_quota, inherit when no
7860 if (quota == RUNTIME_INF)
7861 quota = parent_quota;
7862 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7865 cfs_b->hierarchal_quota = quota;
7870 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7873 struct cfs_schedulable_data data = {
7879 if (quota != RUNTIME_INF) {
7880 do_div(data.period, NSEC_PER_USEC);
7881 do_div(data.quota, NSEC_PER_USEC);
7885 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7891 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7892 struct cgroup_map_cb *cb)
7894 struct task_group *tg = cgroup_tg(cgrp);
7895 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7897 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7898 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7899 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7903 #endif /* CONFIG_CFS_BANDWIDTH */
7904 #endif /* CONFIG_FAIR_GROUP_SCHED */
7906 #ifdef CONFIG_RT_GROUP_SCHED
7907 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7910 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7913 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7915 return sched_group_rt_runtime(cgroup_tg(cgrp));
7918 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7921 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7924 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7926 return sched_group_rt_period(cgroup_tg(cgrp));
7928 #endif /* CONFIG_RT_GROUP_SCHED */
7930 static struct cftype cpu_files[] = {
7931 #ifdef CONFIG_FAIR_GROUP_SCHED
7934 .read_u64 = cpu_shares_read_u64,
7935 .write_u64 = cpu_shares_write_u64,
7938 #ifdef CONFIG_CFS_BANDWIDTH
7940 .name = "cfs_quota_us",
7941 .read_s64 = cpu_cfs_quota_read_s64,
7942 .write_s64 = cpu_cfs_quota_write_s64,
7945 .name = "cfs_period_us",
7946 .read_u64 = cpu_cfs_period_read_u64,
7947 .write_u64 = cpu_cfs_period_write_u64,
7951 .read_map = cpu_stats_show,
7954 #ifdef CONFIG_RT_GROUP_SCHED
7956 .name = "rt_runtime_us",
7957 .read_s64 = cpu_rt_runtime_read,
7958 .write_s64 = cpu_rt_runtime_write,
7961 .name = "rt_period_us",
7962 .read_u64 = cpu_rt_period_read_uint,
7963 .write_u64 = cpu_rt_period_write_uint,
7968 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7970 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7973 struct cgroup_subsys cpu_cgroup_subsys = {
7975 .create = cpu_cgroup_create,
7976 .destroy = cpu_cgroup_destroy,
7977 .can_attach = cpu_cgroup_can_attach,
7978 .attach = cpu_cgroup_attach,
7979 .exit = cpu_cgroup_exit,
7980 .populate = cpu_cgroup_populate,
7981 .subsys_id = cpu_cgroup_subsys_id,
7985 #endif /* CONFIG_CGROUP_SCHED */
7987 #ifdef CONFIG_CGROUP_CPUACCT
7990 * CPU accounting code for task groups.
7992 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7993 * (balbir@in.ibm.com).
7996 /* create a new cpu accounting group */
7997 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8002 return &root_cpuacct.css;
8004 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8008 ca->cpuusage = alloc_percpu(u64);
8012 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8014 goto out_free_cpuusage;
8019 free_percpu(ca->cpuusage);
8023 return ERR_PTR(-ENOMEM);
8026 /* destroy an existing cpu accounting group */
8027 static void cpuacct_destroy(struct cgroup *cgrp)
8029 struct cpuacct *ca = cgroup_ca(cgrp);
8031 free_percpu(ca->cpustat);
8032 free_percpu(ca->cpuusage);
8036 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8038 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8041 #ifndef CONFIG_64BIT
8043 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8045 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8047 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8055 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8057 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8059 #ifndef CONFIG_64BIT
8061 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8063 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8065 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8071 /* return total cpu usage (in nanoseconds) of a group */
8072 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8074 struct cpuacct *ca = cgroup_ca(cgrp);
8075 u64 totalcpuusage = 0;
8078 for_each_present_cpu(i)
8079 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8081 return totalcpuusage;
8084 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8087 struct cpuacct *ca = cgroup_ca(cgrp);
8096 for_each_present_cpu(i)
8097 cpuacct_cpuusage_write(ca, i, 0);
8103 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8106 struct cpuacct *ca = cgroup_ca(cgroup);
8110 for_each_present_cpu(i) {
8111 percpu = cpuacct_cpuusage_read(ca, i);
8112 seq_printf(m, "%llu ", (unsigned long long) percpu);
8114 seq_printf(m, "\n");
8118 static const char *cpuacct_stat_desc[] = {
8119 [CPUACCT_STAT_USER] = "user",
8120 [CPUACCT_STAT_SYSTEM] = "system",
8123 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8124 struct cgroup_map_cb *cb)
8126 struct cpuacct *ca = cgroup_ca(cgrp);
8130 for_each_online_cpu(cpu) {
8131 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8132 val += kcpustat->cpustat[CPUTIME_USER];
8133 val += kcpustat->cpustat[CPUTIME_NICE];
8135 val = cputime64_to_clock_t(val);
8136 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8139 for_each_online_cpu(cpu) {
8140 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8141 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8142 val += kcpustat->cpustat[CPUTIME_IRQ];
8143 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8146 val = cputime64_to_clock_t(val);
8147 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8152 static struct cftype files[] = {
8155 .read_u64 = cpuusage_read,
8156 .write_u64 = cpuusage_write,
8159 .name = "usage_percpu",
8160 .read_seq_string = cpuacct_percpu_seq_read,
8164 .read_map = cpuacct_stats_show,
8168 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8170 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8174 * charge this task's execution time to its accounting group.
8176 * called with rq->lock held.
8178 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8183 if (unlikely(!cpuacct_subsys.active))
8186 cpu = task_cpu(tsk);
8192 for (; ca; ca = parent_ca(ca)) {
8193 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8194 *cpuusage += cputime;
8200 struct cgroup_subsys cpuacct_subsys = {
8202 .create = cpuacct_create,
8203 .destroy = cpuacct_destroy,
8204 .populate = cpuacct_populate,
8205 .subsys_id = cpuacct_subsys_id,
8207 #endif /* CONFIG_CGROUP_CPUACCT */