4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
83 #include "../workqueue_sched.h"
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/sched.h>
88 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
91 ktime_t soft, hard, now;
94 if (hrtimer_active(period_timer))
97 now = hrtimer_cb_get_time(period_timer);
98 hrtimer_forward(period_timer, now, period);
100 soft = hrtimer_get_softexpires(period_timer);
101 hard = hrtimer_get_expires(period_timer);
102 delta = ktime_to_ns(ktime_sub(hard, soft));
103 __hrtimer_start_range_ns(period_timer, soft, delta,
104 HRTIMER_MODE_ABS_PINNED, 0);
108 DEFINE_MUTEX(sched_domains_mutex);
109 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
111 static void update_rq_clock_task(struct rq *rq, s64 delta);
113 void update_rq_clock(struct rq *rq)
117 if (rq->skip_clock_update > 0)
120 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
122 update_rq_clock_task(rq, delta);
126 * Debugging: various feature bits
129 #define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
132 const_debug unsigned int sysctl_sched_features =
133 #include "features.h"
138 #ifdef CONFIG_SCHED_DEBUG
139 #define SCHED_FEAT(name, enabled) \
142 static __read_mostly char *sched_feat_names[] = {
143 #include "features.h"
149 static int sched_feat_show(struct seq_file *m, void *v)
153 for (i = 0; i < __SCHED_FEAT_NR; i++) {
154 if (!(sysctl_sched_features & (1UL << i)))
156 seq_printf(m, "%s ", sched_feat_names[i]);
163 #ifdef HAVE_JUMP_LABEL
165 #define jump_label_key__true jump_label_key_enabled
166 #define jump_label_key__false jump_label_key_disabled
168 #define SCHED_FEAT(name, enabled) \
169 jump_label_key__##enabled ,
171 struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = {
172 #include "features.h"
177 static void sched_feat_disable(int i)
179 if (jump_label_enabled(&sched_feat_keys[i]))
180 jump_label_dec(&sched_feat_keys[i]);
183 static void sched_feat_enable(int i)
185 if (!jump_label_enabled(&sched_feat_keys[i]))
186 jump_label_inc(&sched_feat_keys[i]);
189 static void sched_feat_disable(int i) { };
190 static void sched_feat_enable(int i) { };
191 #endif /* HAVE_JUMP_LABEL */
194 sched_feat_write(struct file *filp, const char __user *ubuf,
195 size_t cnt, loff_t *ppos)
205 if (copy_from_user(&buf, ubuf, cnt))
211 if (strncmp(cmp, "NO_", 3) == 0) {
216 for (i = 0; i < __SCHED_FEAT_NR; i++) {
217 if (strcmp(cmp, sched_feat_names[i]) == 0) {
219 sysctl_sched_features &= ~(1UL << i);
220 sched_feat_disable(i);
222 sysctl_sched_features |= (1UL << i);
223 sched_feat_enable(i);
229 if (i == __SCHED_FEAT_NR)
237 static int sched_feat_open(struct inode *inode, struct file *filp)
239 return single_open(filp, sched_feat_show, NULL);
242 static const struct file_operations sched_feat_fops = {
243 .open = sched_feat_open,
244 .write = sched_feat_write,
247 .release = single_release,
250 static __init int sched_init_debug(void)
252 debugfs_create_file("sched_features", 0644, NULL, NULL,
257 late_initcall(sched_init_debug);
258 #endif /* CONFIG_SCHED_DEBUG */
261 * Number of tasks to iterate in a single balance run.
262 * Limited because this is done with IRQs disabled.
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
267 * period over which we average the RT time consumption, measured
272 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275 * period over which we measure -rt task cpu usage in us.
278 unsigned int sysctl_sched_rt_period = 1000000;
280 __read_mostly int scheduler_running;
283 * part of the period that we allow rt tasks to run in us.
286 int sysctl_sched_rt_runtime = 950000;
291 * __task_rq_lock - lock the rq @p resides on.
293 static inline struct rq *__task_rq_lock(struct task_struct *p)
298 lockdep_assert_held(&p->pi_lock);
302 raw_spin_lock(&rq->lock);
303 if (likely(rq == task_rq(p)))
305 raw_spin_unlock(&rq->lock);
310 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
312 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
313 __acquires(p->pi_lock)
319 raw_spin_lock_irqsave(&p->pi_lock, *flags);
321 raw_spin_lock(&rq->lock);
322 if (likely(rq == task_rq(p)))
324 raw_spin_unlock(&rq->lock);
325 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
329 static void __task_rq_unlock(struct rq *rq)
332 raw_spin_unlock(&rq->lock);
336 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
338 __releases(p->pi_lock)
340 raw_spin_unlock(&rq->lock);
341 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
345 * this_rq_lock - lock this runqueue and disable interrupts.
347 static struct rq *this_rq_lock(void)
354 raw_spin_lock(&rq->lock);
359 #ifdef CONFIG_SCHED_HRTICK
361 * Use HR-timers to deliver accurate preemption points.
363 * Its all a bit involved since we cannot program an hrt while holding the
364 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * When we get rescheduled we reprogram the hrtick_timer outside of the
371 static void hrtick_clear(struct rq *rq)
373 if (hrtimer_active(&rq->hrtick_timer))
374 hrtimer_cancel(&rq->hrtick_timer);
378 * High-resolution timer tick.
379 * Runs from hardirq context with interrupts disabled.
381 static enum hrtimer_restart hrtick(struct hrtimer *timer)
383 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
385 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
387 raw_spin_lock(&rq->lock);
389 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
390 raw_spin_unlock(&rq->lock);
392 return HRTIMER_NORESTART;
397 * called from hardirq (IPI) context
399 static void __hrtick_start(void *arg)
403 raw_spin_lock(&rq->lock);
404 hrtimer_restart(&rq->hrtick_timer);
405 rq->hrtick_csd_pending = 0;
406 raw_spin_unlock(&rq->lock);
410 * Called to set the hrtick timer state.
412 * called with rq->lock held and irqs disabled
414 void hrtick_start(struct rq *rq, u64 delay)
416 struct hrtimer *timer = &rq->hrtick_timer;
417 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
419 hrtimer_set_expires(timer, time);
421 if (rq == this_rq()) {
422 hrtimer_restart(timer);
423 } else if (!rq->hrtick_csd_pending) {
424 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
425 rq->hrtick_csd_pending = 1;
430 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
432 int cpu = (int)(long)hcpu;
435 case CPU_UP_CANCELED:
436 case CPU_UP_CANCELED_FROZEN:
437 case CPU_DOWN_PREPARE:
438 case CPU_DOWN_PREPARE_FROZEN:
440 case CPU_DEAD_FROZEN:
441 hrtick_clear(cpu_rq(cpu));
448 static __init void init_hrtick(void)
450 hotcpu_notifier(hotplug_hrtick, 0);
454 * Called to set the hrtick timer state.
456 * called with rq->lock held and irqs disabled
458 void hrtick_start(struct rq *rq, u64 delay)
460 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
461 HRTIMER_MODE_REL_PINNED, 0);
464 static inline void init_hrtick(void)
467 #endif /* CONFIG_SMP */
469 static void init_rq_hrtick(struct rq *rq)
472 rq->hrtick_csd_pending = 0;
474 rq->hrtick_csd.flags = 0;
475 rq->hrtick_csd.func = __hrtick_start;
476 rq->hrtick_csd.info = rq;
479 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
480 rq->hrtick_timer.function = hrtick;
482 #else /* CONFIG_SCHED_HRTICK */
483 static inline void hrtick_clear(struct rq *rq)
487 static inline void init_rq_hrtick(struct rq *rq)
491 static inline void init_hrtick(void)
494 #endif /* CONFIG_SCHED_HRTICK */
497 * resched_task - mark a task 'to be rescheduled now'.
499 * On UP this means the setting of the need_resched flag, on SMP it
500 * might also involve a cross-CPU call to trigger the scheduler on
505 #ifndef tsk_is_polling
506 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509 void resched_task(struct task_struct *p)
513 assert_raw_spin_locked(&task_rq(p)->lock);
515 if (test_tsk_need_resched(p))
518 set_tsk_need_resched(p);
521 if (cpu == smp_processor_id())
524 /* NEED_RESCHED must be visible before we test polling */
526 if (!tsk_is_polling(p))
527 smp_send_reschedule(cpu);
530 void resched_cpu(int cpu)
532 struct rq *rq = cpu_rq(cpu);
535 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
537 resched_task(cpu_curr(cpu));
538 raw_spin_unlock_irqrestore(&rq->lock, flags);
543 * In the semi idle case, use the nearest busy cpu for migrating timers
544 * from an idle cpu. This is good for power-savings.
546 * We don't do similar optimization for completely idle system, as
547 * selecting an idle cpu will add more delays to the timers than intended
548 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
550 int get_nohz_timer_target(void)
552 int cpu = smp_processor_id();
554 struct sched_domain *sd;
557 for_each_domain(cpu, sd) {
558 for_each_cpu(i, sched_domain_span(sd)) {
570 * When add_timer_on() enqueues a timer into the timer wheel of an
571 * idle CPU then this timer might expire before the next timer event
572 * which is scheduled to wake up that CPU. In case of a completely
573 * idle system the next event might even be infinite time into the
574 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
575 * leaves the inner idle loop so the newly added timer is taken into
576 * account when the CPU goes back to idle and evaluates the timer
577 * wheel for the next timer event.
579 void wake_up_idle_cpu(int cpu)
581 struct rq *rq = cpu_rq(cpu);
583 if (cpu == smp_processor_id())
587 * This is safe, as this function is called with the timer
588 * wheel base lock of (cpu) held. When the CPU is on the way
589 * to idle and has not yet set rq->curr to idle then it will
590 * be serialized on the timer wheel base lock and take the new
591 * timer into account automatically.
593 if (rq->curr != rq->idle)
597 * We can set TIF_RESCHED on the idle task of the other CPU
598 * lockless. The worst case is that the other CPU runs the
599 * idle task through an additional NOOP schedule()
601 set_tsk_need_resched(rq->idle);
603 /* NEED_RESCHED must be visible before we test polling */
605 if (!tsk_is_polling(rq->idle))
606 smp_send_reschedule(cpu);
609 static inline bool got_nohz_idle_kick(void)
611 int cpu = smp_processor_id();
612 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615 #else /* CONFIG_NO_HZ */
617 static inline bool got_nohz_idle_kick(void)
622 #endif /* CONFIG_NO_HZ */
624 void sched_avg_update(struct rq *rq)
626 s64 period = sched_avg_period();
628 while ((s64)(rq->clock - rq->age_stamp) > period) {
630 * Inline assembly required to prevent the compiler
631 * optimising this loop into a divmod call.
632 * See __iter_div_u64_rem() for another example of this.
634 asm("" : "+rm" (rq->age_stamp));
635 rq->age_stamp += period;
640 #else /* !CONFIG_SMP */
641 void resched_task(struct task_struct *p)
643 assert_raw_spin_locked(&task_rq(p)->lock);
644 set_tsk_need_resched(p);
646 #endif /* CONFIG_SMP */
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
651 * Iterate task_group tree rooted at *from, calling @down when first entering a
652 * node and @up when leaving it for the final time.
654 * Caller must hold rcu_lock or sufficient equivalent.
656 int walk_tg_tree_from(struct task_group *from,
657 tg_visitor down, tg_visitor up, void *data)
659 struct task_group *parent, *child;
665 ret = (*down)(parent, data);
668 list_for_each_entry_rcu(child, &parent->children, siblings) {
675 ret = (*up)(parent, data);
676 if (ret || parent == from)
680 parent = parent->parent;
687 int tg_nop(struct task_group *tg, void *data)
693 void update_cpu_load(struct rq *this_rq);
695 static void set_load_weight(struct task_struct *p)
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
713 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
720 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
727 void activate_task(struct rq *rq, struct task_struct *p, int flags)
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
732 enqueue_task(rq, p, flags);
735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
740 dequeue_task(rq, p, flags);
743 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
756 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757 static DEFINE_PER_CPU(u64, cpu_softirq_time);
759 static DEFINE_PER_CPU(u64, irq_start_time);
760 static int sched_clock_irqtime;
762 void enable_sched_clock_irqtime(void)
764 sched_clock_irqtime = 1;
767 void disable_sched_clock_irqtime(void)
769 sched_clock_irqtime = 0;
773 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
775 static inline void irq_time_write_begin(void)
777 __this_cpu_inc(irq_time_seq.sequence);
781 static inline void irq_time_write_end(void)
784 __this_cpu_inc(irq_time_seq.sequence);
787 static inline u64 irq_time_read(int cpu)
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
800 #else /* CONFIG_64BIT */
801 static inline void irq_time_write_begin(void)
805 static inline void irq_time_write_end(void)
809 static inline u64 irq_time_read(int cpu)
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
813 #endif /* CONFIG_64BIT */
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
819 void account_system_vtime(struct task_struct *curr)
825 if (!sched_clock_irqtime)
828 local_irq_save(flags);
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
834 irq_time_write_begin();
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
846 irq_time_write_end();
847 local_irq_restore(flags);
849 EXPORT_SYMBOL_GPL(account_system_vtime);
851 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
853 #ifdef CONFIG_PARAVIRT
854 static inline u64 steal_ticks(u64 steal)
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
863 static void update_rq_clock_task(struct rq *rq, s64 delta)
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
890 if (irq_delta > delta)
893 rq->prev_irq_time += irq_delta;
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_branch((¶virt_steal_rq_enabled))) {
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
903 if (unlikely(steal > delta))
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
909 rq->prev_steal_time_rq += steal;
915 rq->clock_task += delta;
917 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
923 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
924 static int irqtime_account_hi_update(void)
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
935 local_irq_restore(flags);
939 static int irqtime_account_si_update(void)
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
950 local_irq_restore(flags);
954 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
956 #define sched_clock_irqtime (0)
960 void sched_set_stop_task(int cpu, struct task_struct *stop)
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
976 stop->sched_class = &stop_sched_class;
979 cpu_rq(cpu)->stop = stop;
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
986 old_stop->sched_class = &rt_sched_class;
991 * __normal_prio - return the priority that is based on the static prio
993 static inline int __normal_prio(struct task_struct *p)
995 return p->static_prio;
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1005 static inline int normal_prio(struct task_struct *p)
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1012 prio = __normal_prio(p);
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1023 static int effective_prio(struct task_struct *p)
1025 p->normal_prio = normal_prio(p);
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1040 inline int task_curr(const struct task_struct *p)
1042 return cpu_curr(task_cpu(p)) == p;
1045 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1057 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1059 const struct sched_class *class;
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1085 #ifdef CONFIG_SCHED_DEBUG
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1093 #ifdef CONFIG_LOCKDEP
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see set_task_rq().
1101 * Furthermore, all task_rq users should acquire both locks, see
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1109 trace_sched_migrate_task(p, new_cpu);
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1116 __set_task_cpu(p, new_cpu);
1119 struct migration_arg {
1120 struct task_struct *task;
1124 static int migration_cpu_stop(void *data);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1144 unsigned long flags;
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1190 * If it changed from the expected state, bail out now.
1192 if (unlikely(!ncsw))
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1199 * Oops. Go back and try again..
1201 if (unlikely(running)) {
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1247 void kick_process(struct task_struct *p)
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1257 EXPORT_SYMBOL_GPL(kick_process);
1258 #endif /* CONFIG_SMP */
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1264 static int select_fallback_rq(int cpu, struct task_struct *p)
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu_mask(dest_cpu, *nodemask) {
1272 if (!cpu_online(dest_cpu))
1274 if (!cpu_active(dest_cpu))
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1281 /* Any allowed, online CPU? */
1282 for_each_cpu_mask(dest_cpu, *tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1285 if (!cpu_active(dest_cpu))
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1309 if (state != cpuset) {
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1328 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1337 * Since this is common to all placement strategies, this lives here.
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1344 cpu = select_fallback_rq(task_cpu(p), p);
1349 static void update_avg(u64 *avg, u64 sample)
1351 s64 diff = sample - *avg;
1357 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1359 #ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1363 int this_cpu = smp_processor_id();
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1369 struct sched_domain *sd;
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1385 #endif /* CONFIG_SMP */
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1393 #endif /* CONFIG_SCHEDSTATS */
1396 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1398 activate_task(rq, p, en_flags);
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1407 * Mark the task runnable and perform wakeup-preemption.
1410 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1415 p->state = TASK_RUNNING;
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1427 update_avg(&rq->avg_idle, delta);
1434 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1451 static int ttwu_remote(struct task_struct *p, int wake_flags)
1456 rq = __task_rq_lock(p);
1458 ttwu_do_wakeup(rq, p, wake_flags);
1461 __task_rq_unlock(rq);
1467 static void sched_ttwu_pending(void)
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1473 raw_spin_lock(&rq->lock);
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1481 raw_spin_unlock(&rq->lock);
1484 void scheduler_ipi(void)
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1495 * Some archs already do call them, luckily irq_enter/exit nest
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1503 sched_ttwu_pending();
1506 * Check if someone kicked us for doing the nohz idle load balance.
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1515 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1521 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1527 rq = __task_rq_lock(p);
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1533 __task_rq_unlock(rq);
1538 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1540 bool cpus_share_cache(int this_cpu, int that_cpu)
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544 #endif /* CONFIG_SMP */
1546 static void ttwu_queue(struct task_struct *p, int cpu)
1548 struct rq *rq = cpu_rq(cpu);
1550 #if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1579 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581 unsigned long flags;
1582 int cpu, success = 0;
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1589 success = 1; /* we're going to change ->state */
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1601 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1609 if (ttwu_activate_remote(p, wake_flags))
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1631 #endif /* CONFIG_SMP */
1635 ttwu_stat(p, cpu, wake_flags);
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1650 static void try_to_wake_up_local(struct task_struct *p)
1652 struct rq *rq = task_rq(p);
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1664 if (!(p->state & TASK_NORMAL))
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1673 raw_spin_unlock(&p->pi_lock);
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1687 int wake_up_process(struct task_struct *p)
1689 return try_to_wake_up(p, TASK_ALL, 0);
1691 EXPORT_SYMBOL(wake_up_process);
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1695 return try_to_wake_up(p, state, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(struct task_struct *p)
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1720 INIT_LIST_HEAD(&p->rt.run_list);
1722 #ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1728 * fork()/clone()-time setup:
1730 void sched_fork(struct task_struct *p)
1732 unsigned long flags;
1733 int cpu = get_cpu();
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1741 p->state = TASK_RUNNING;
1744 * Make sure we do not leak PI boosting priority to the child.
1746 p->prio = current->normal_prio;
1749 * Revert to default priority/policy on fork if requested.
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1759 p->prio = p->normal_prio = __normal_prio(p);
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1766 p->sched_reset_on_fork = 0;
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1780 * Silence PROVE_RCU.
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1786 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1790 #if defined(CONFIG_SMP)
1793 #ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1805 * wake_up_new_task - wake up a newly created task for the first time.
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1811 void wake_up_new_task(struct task_struct *p)
1813 unsigned long flags;
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1835 task_rq_unlock(rq, p, &flags);
1838 #ifdef CONFIG_PREEMPT_NOTIFIERS
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1844 void preempt_notifier_register(struct preempt_notifier *notifier)
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1848 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1854 * This is safe to call from within a preemption notifier.
1856 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1858 hlist_del(¬ifier->link);
1860 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1862 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1872 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1882 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1884 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1889 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1894 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1906 * prepare_task_switch sets up locking and calls architecture specific
1910 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1913 sched_info_switch(prev, next);
1914 perf_event_task_sched_out(prev, next);
1915 fire_sched_out_preempt_notifiers(prev, next);
1916 prepare_lock_switch(rq, next);
1917 prepare_arch_switch(next);
1918 trace_sched_switch(prev, next);
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1936 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1939 struct mm_struct *mm = rq->prev_mm;
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1953 * Manfred Spraul <manfred@colorfullife.com>
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1963 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1966 fire_sched_in_preempt_notifiers(current);
1969 if (unlikely(prev_state == TASK_DEAD)) {
1971 * Remove function-return probe instances associated with this
1972 * task and put them back on the free list.
1974 kprobe_flush_task(prev);
1975 put_task_struct(prev);
1981 /* assumes rq->lock is held */
1982 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1984 if (prev->sched_class->pre_schedule)
1985 prev->sched_class->pre_schedule(rq, prev);
1988 /* rq->lock is NOT held, but preemption is disabled */
1989 static inline void post_schedule(struct rq *rq)
1991 if (rq->post_schedule) {
1992 unsigned long flags;
1994 raw_spin_lock_irqsave(&rq->lock, flags);
1995 if (rq->curr->sched_class->post_schedule)
1996 rq->curr->sched_class->post_schedule(rq);
1997 raw_spin_unlock_irqrestore(&rq->lock, flags);
1999 rq->post_schedule = 0;
2005 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2009 static inline void post_schedule(struct rq *rq)
2016 * schedule_tail - first thing a freshly forked thread must call.
2017 * @prev: the thread we just switched away from.
2019 asmlinkage void schedule_tail(struct task_struct *prev)
2020 __releases(rq->lock)
2022 struct rq *rq = this_rq();
2024 finish_task_switch(rq, prev);
2027 * FIXME: do we need to worry about rq being invalidated by the
2032 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2033 /* In this case, finish_task_switch does not reenable preemption */
2036 if (current->set_child_tid)
2037 put_user(task_pid_vnr(current), current->set_child_tid);
2041 * context_switch - switch to the new MM and the new
2042 * thread's register state.
2045 context_switch(struct rq *rq, struct task_struct *prev,
2046 struct task_struct *next)
2048 struct mm_struct *mm, *oldmm;
2050 prepare_task_switch(rq, prev, next);
2053 oldmm = prev->active_mm;
2055 * For paravirt, this is coupled with an exit in switch_to to
2056 * combine the page table reload and the switch backend into
2059 arch_start_context_switch(prev);
2062 next->active_mm = oldmm;
2063 atomic_inc(&oldmm->mm_count);
2064 enter_lazy_tlb(oldmm, next);
2066 switch_mm(oldmm, mm, next);
2069 prev->active_mm = NULL;
2070 rq->prev_mm = oldmm;
2073 * Since the runqueue lock will be released by the next
2074 * task (which is an invalid locking op but in the case
2075 * of the scheduler it's an obvious special-case), so we
2076 * do an early lockdep release here:
2078 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2079 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2082 /* Here we just switch the register state and the stack. */
2083 switch_to(prev, next, prev);
2087 * this_rq must be evaluated again because prev may have moved
2088 * CPUs since it called schedule(), thus the 'rq' on its stack
2089 * frame will be invalid.
2091 finish_task_switch(this_rq(), prev);
2095 * nr_running, nr_uninterruptible and nr_context_switches:
2097 * externally visible scheduler statistics: current number of runnable
2098 * threads, current number of uninterruptible-sleeping threads, total
2099 * number of context switches performed since bootup.
2101 unsigned long nr_running(void)
2103 unsigned long i, sum = 0;
2105 for_each_online_cpu(i)
2106 sum += cpu_rq(i)->nr_running;
2111 unsigned long nr_uninterruptible(void)
2113 unsigned long i, sum = 0;
2115 for_each_possible_cpu(i)
2116 sum += cpu_rq(i)->nr_uninterruptible;
2119 * Since we read the counters lockless, it might be slightly
2120 * inaccurate. Do not allow it to go below zero though:
2122 if (unlikely((long)sum < 0))
2128 unsigned long long nr_context_switches(void)
2131 unsigned long long sum = 0;
2133 for_each_possible_cpu(i)
2134 sum += cpu_rq(i)->nr_switches;
2139 unsigned long nr_iowait(void)
2141 unsigned long i, sum = 0;
2143 for_each_possible_cpu(i)
2144 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2149 unsigned long nr_iowait_cpu(int cpu)
2151 struct rq *this = cpu_rq(cpu);
2152 return atomic_read(&this->nr_iowait);
2155 unsigned long this_cpu_load(void)
2157 struct rq *this = this_rq();
2158 return this->cpu_load[0];
2162 /* Variables and functions for calc_load */
2163 static atomic_long_t calc_load_tasks;
2164 static unsigned long calc_load_update;
2165 unsigned long avenrun[3];
2166 EXPORT_SYMBOL(avenrun);
2168 static long calc_load_fold_active(struct rq *this_rq)
2170 long nr_active, delta = 0;
2172 nr_active = this_rq->nr_running;
2173 nr_active += (long) this_rq->nr_uninterruptible;
2175 if (nr_active != this_rq->calc_load_active) {
2176 delta = nr_active - this_rq->calc_load_active;
2177 this_rq->calc_load_active = nr_active;
2183 static unsigned long
2184 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2187 load += active * (FIXED_1 - exp);
2188 load += 1UL << (FSHIFT - 1);
2189 return load >> FSHIFT;
2194 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2196 * When making the ILB scale, we should try to pull this in as well.
2198 static atomic_long_t calc_load_tasks_idle;
2200 void calc_load_account_idle(struct rq *this_rq)
2204 delta = calc_load_fold_active(this_rq);
2206 atomic_long_add(delta, &calc_load_tasks_idle);
2209 static long calc_load_fold_idle(void)
2214 * Its got a race, we don't care...
2216 if (atomic_long_read(&calc_load_tasks_idle))
2217 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2223 * fixed_power_int - compute: x^n, in O(log n) time
2225 * @x: base of the power
2226 * @frac_bits: fractional bits of @x
2227 * @n: power to raise @x to.
2229 * By exploiting the relation between the definition of the natural power
2230 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2231 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2232 * (where: n_i \elem {0, 1}, the binary vector representing n),
2233 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2234 * of course trivially computable in O(log_2 n), the length of our binary
2237 static unsigned long
2238 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2240 unsigned long result = 1UL << frac_bits;
2245 result += 1UL << (frac_bits - 1);
2246 result >>= frac_bits;
2252 x += 1UL << (frac_bits - 1);
2260 * a1 = a0 * e + a * (1 - e)
2262 * a2 = a1 * e + a * (1 - e)
2263 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2264 * = a0 * e^2 + a * (1 - e) * (1 + e)
2266 * a3 = a2 * e + a * (1 - e)
2267 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2268 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2272 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2273 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2274 * = a0 * e^n + a * (1 - e^n)
2276 * [1] application of the geometric series:
2279 * S_n := \Sum x^i = -------------
2282 static unsigned long
2283 calc_load_n(unsigned long load, unsigned long exp,
2284 unsigned long active, unsigned int n)
2287 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2291 * NO_HZ can leave us missing all per-cpu ticks calling
2292 * calc_load_account_active(), but since an idle CPU folds its delta into
2293 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2294 * in the pending idle delta if our idle period crossed a load cycle boundary.
2296 * Once we've updated the global active value, we need to apply the exponential
2297 * weights adjusted to the number of cycles missed.
2299 static void calc_global_nohz(void)
2301 long delta, active, n;
2304 * If we crossed a calc_load_update boundary, make sure to fold
2305 * any pending idle changes, the respective CPUs might have
2306 * missed the tick driven calc_load_account_active() update
2309 delta = calc_load_fold_idle();
2311 atomic_long_add(delta, &calc_load_tasks);
2314 * It could be the one fold was all it took, we done!
2316 if (time_before(jiffies, calc_load_update + 10))
2320 * Catch-up, fold however many we are behind still
2322 delta = jiffies - calc_load_update - 10;
2323 n = 1 + (delta / LOAD_FREQ);
2325 active = atomic_long_read(&calc_load_tasks);
2326 active = active > 0 ? active * FIXED_1 : 0;
2328 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2329 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2330 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2332 calc_load_update += n * LOAD_FREQ;
2335 void calc_load_account_idle(struct rq *this_rq)
2339 static inline long calc_load_fold_idle(void)
2344 static void calc_global_nohz(void)
2350 * get_avenrun - get the load average array
2351 * @loads: pointer to dest load array
2352 * @offset: offset to add
2353 * @shift: shift count to shift the result left
2355 * These values are estimates at best, so no need for locking.
2357 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2359 loads[0] = (avenrun[0] + offset) << shift;
2360 loads[1] = (avenrun[1] + offset) << shift;
2361 loads[2] = (avenrun[2] + offset) << shift;
2365 * calc_load - update the avenrun load estimates 10 ticks after the
2366 * CPUs have updated calc_load_tasks.
2368 void calc_global_load(unsigned long ticks)
2372 if (time_before(jiffies, calc_load_update + 10))
2375 active = atomic_long_read(&calc_load_tasks);
2376 active = active > 0 ? active * FIXED_1 : 0;
2378 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2379 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2380 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2382 calc_load_update += LOAD_FREQ;
2385 * Account one period with whatever state we found before
2386 * folding in the nohz state and ageing the entire idle period.
2388 * This avoids loosing a sample when we go idle between
2389 * calc_load_account_active() (10 ticks ago) and now and thus
2396 * Called from update_cpu_load() to periodically update this CPU's
2399 static void calc_load_account_active(struct rq *this_rq)
2403 if (time_before(jiffies, this_rq->calc_load_update))
2406 delta = calc_load_fold_active(this_rq);
2407 delta += calc_load_fold_idle();
2409 atomic_long_add(delta, &calc_load_tasks);
2411 this_rq->calc_load_update += LOAD_FREQ;
2415 * The exact cpuload at various idx values, calculated at every tick would be
2416 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2418 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2419 * on nth tick when cpu may be busy, then we have:
2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2421 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2423 * decay_load_missed() below does efficient calculation of
2424 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2425 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2427 * The calculation is approximated on a 128 point scale.
2428 * degrade_zero_ticks is the number of ticks after which load at any
2429 * particular idx is approximated to be zero.
2430 * degrade_factor is a precomputed table, a row for each load idx.
2431 * Each column corresponds to degradation factor for a power of two ticks,
2432 * based on 128 point scale.
2434 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2435 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2437 * With this power of 2 load factors, we can degrade the load n times
2438 * by looking at 1 bits in n and doing as many mult/shift instead of
2439 * n mult/shifts needed by the exact degradation.
2441 #define DEGRADE_SHIFT 7
2442 static const unsigned char
2443 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2444 static const unsigned char
2445 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2446 {0, 0, 0, 0, 0, 0, 0, 0},
2447 {64, 32, 8, 0, 0, 0, 0, 0},
2448 {96, 72, 40, 12, 1, 0, 0},
2449 {112, 98, 75, 43, 15, 1, 0},
2450 {120, 112, 98, 76, 45, 16, 2} };
2453 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2454 * would be when CPU is idle and so we just decay the old load without
2455 * adding any new load.
2457 static unsigned long
2458 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2462 if (!missed_updates)
2465 if (missed_updates >= degrade_zero_ticks[idx])
2469 return load >> missed_updates;
2471 while (missed_updates) {
2472 if (missed_updates % 2)
2473 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2475 missed_updates >>= 1;
2482 * Update rq->cpu_load[] statistics. This function is usually called every
2483 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2484 * every tick. We fix it up based on jiffies.
2486 void update_cpu_load(struct rq *this_rq)
2488 unsigned long this_load = this_rq->load.weight;
2489 unsigned long curr_jiffies = jiffies;
2490 unsigned long pending_updates;
2493 this_rq->nr_load_updates++;
2495 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2496 if (curr_jiffies == this_rq->last_load_update_tick)
2499 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2500 this_rq->last_load_update_tick = curr_jiffies;
2502 /* Update our load: */
2503 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2504 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2505 unsigned long old_load, new_load;
2507 /* scale is effectively 1 << i now, and >> i divides by scale */
2509 old_load = this_rq->cpu_load[i];
2510 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2511 new_load = this_load;
2513 * Round up the averaging division if load is increasing. This
2514 * prevents us from getting stuck on 9 if the load is 10, for
2517 if (new_load > old_load)
2518 new_load += scale - 1;
2520 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2523 sched_avg_update(this_rq);
2526 static void update_cpu_load_active(struct rq *this_rq)
2528 update_cpu_load(this_rq);
2530 calc_load_account_active(this_rq);
2536 * sched_exec - execve() is a valuable balancing opportunity, because at
2537 * this point the task has the smallest effective memory and cache footprint.
2539 void sched_exec(void)
2541 struct task_struct *p = current;
2542 unsigned long flags;
2545 raw_spin_lock_irqsave(&p->pi_lock, flags);
2546 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2547 if (dest_cpu == smp_processor_id())
2550 if (likely(cpu_active(dest_cpu))) {
2551 struct migration_arg arg = { p, dest_cpu };
2553 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2554 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2558 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2563 DEFINE_PER_CPU(struct kernel_stat, kstat);
2564 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2566 EXPORT_PER_CPU_SYMBOL(kstat);
2567 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2570 * Return any ns on the sched_clock that have not yet been accounted in
2571 * @p in case that task is currently running.
2573 * Called with task_rq_lock() held on @rq.
2575 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2579 if (task_current(rq, p)) {
2580 update_rq_clock(rq);
2581 ns = rq->clock_task - p->se.exec_start;
2589 unsigned long long task_delta_exec(struct task_struct *p)
2591 unsigned long flags;
2595 rq = task_rq_lock(p, &flags);
2596 ns = do_task_delta_exec(p, rq);
2597 task_rq_unlock(rq, p, &flags);
2603 * Return accounted runtime for the task.
2604 * In case the task is currently running, return the runtime plus current's
2605 * pending runtime that have not been accounted yet.
2607 unsigned long long task_sched_runtime(struct task_struct *p)
2609 unsigned long flags;
2613 rq = task_rq_lock(p, &flags);
2614 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2615 task_rq_unlock(rq, p, &flags);
2620 #ifdef CONFIG_CGROUP_CPUACCT
2621 struct cgroup_subsys cpuacct_subsys;
2622 struct cpuacct root_cpuacct;
2625 static inline void task_group_account_field(struct task_struct *p, int index,
2628 #ifdef CONFIG_CGROUP_CPUACCT
2629 struct kernel_cpustat *kcpustat;
2633 * Since all updates are sure to touch the root cgroup, we
2634 * get ourselves ahead and touch it first. If the root cgroup
2635 * is the only cgroup, then nothing else should be necessary.
2638 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2640 #ifdef CONFIG_CGROUP_CPUACCT
2641 if (unlikely(!cpuacct_subsys.active))
2646 while (ca && (ca != &root_cpuacct)) {
2647 kcpustat = this_cpu_ptr(ca->cpustat);
2648 kcpustat->cpustat[index] += tmp;
2657 * Account user cpu time to a process.
2658 * @p: the process that the cpu time gets accounted to
2659 * @cputime: the cpu time spent in user space since the last update
2660 * @cputime_scaled: cputime scaled by cpu frequency
2662 void account_user_time(struct task_struct *p, cputime_t cputime,
2663 cputime_t cputime_scaled)
2667 /* Add user time to process. */
2668 p->utime += cputime;
2669 p->utimescaled += cputime_scaled;
2670 account_group_user_time(p, cputime);
2672 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2674 /* Add user time to cpustat. */
2675 task_group_account_field(p, index, (__force u64) cputime);
2677 /* Account for user time used */
2678 acct_update_integrals(p);
2682 * Account guest cpu time to a process.
2683 * @p: the process that the cpu time gets accounted to
2684 * @cputime: the cpu time spent in virtual machine since the last update
2685 * @cputime_scaled: cputime scaled by cpu frequency
2687 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2688 cputime_t cputime_scaled)
2690 u64 *cpustat = kcpustat_this_cpu->cpustat;
2692 /* Add guest time to process. */
2693 p->utime += cputime;
2694 p->utimescaled += cputime_scaled;
2695 account_group_user_time(p, cputime);
2696 p->gtime += cputime;
2698 /* Add guest time to cpustat. */
2699 if (TASK_NICE(p) > 0) {
2700 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2701 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2703 cpustat[CPUTIME_USER] += (__force u64) cputime;
2704 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2709 * Account system cpu time to a process and desired cpustat field
2710 * @p: the process that the cpu time gets accounted to
2711 * @cputime: the cpu time spent in kernel space since the last update
2712 * @cputime_scaled: cputime scaled by cpu frequency
2713 * @target_cputime64: pointer to cpustat field that has to be updated
2716 void __account_system_time(struct task_struct *p, cputime_t cputime,
2717 cputime_t cputime_scaled, int index)
2719 /* Add system time to process. */
2720 p->stime += cputime;
2721 p->stimescaled += cputime_scaled;
2722 account_group_system_time(p, cputime);
2724 /* Add system time to cpustat. */
2725 task_group_account_field(p, index, (__force u64) cputime);
2727 /* Account for system time used */
2728 acct_update_integrals(p);
2732 * Account system cpu time to a process.
2733 * @p: the process that the cpu time gets accounted to
2734 * @hardirq_offset: the offset to subtract from hardirq_count()
2735 * @cputime: the cpu time spent in kernel space since the last update
2736 * @cputime_scaled: cputime scaled by cpu frequency
2738 void account_system_time(struct task_struct *p, int hardirq_offset,
2739 cputime_t cputime, cputime_t cputime_scaled)
2743 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2744 account_guest_time(p, cputime, cputime_scaled);
2748 if (hardirq_count() - hardirq_offset)
2749 index = CPUTIME_IRQ;
2750 else if (in_serving_softirq())
2751 index = CPUTIME_SOFTIRQ;
2753 index = CPUTIME_SYSTEM;
2755 __account_system_time(p, cputime, cputime_scaled, index);
2759 * Account for involuntary wait time.
2760 * @cputime: the cpu time spent in involuntary wait
2762 void account_steal_time(cputime_t cputime)
2764 u64 *cpustat = kcpustat_this_cpu->cpustat;
2766 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2770 * Account for idle time.
2771 * @cputime: the cpu time spent in idle wait
2773 void account_idle_time(cputime_t cputime)
2775 u64 *cpustat = kcpustat_this_cpu->cpustat;
2776 struct rq *rq = this_rq();
2778 if (atomic_read(&rq->nr_iowait) > 0)
2779 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2781 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2784 static __always_inline bool steal_account_process_tick(void)
2786 #ifdef CONFIG_PARAVIRT
2787 if (static_branch(¶virt_steal_enabled)) {
2790 steal = paravirt_steal_clock(smp_processor_id());
2791 steal -= this_rq()->prev_steal_time;
2793 st = steal_ticks(steal);
2794 this_rq()->prev_steal_time += st * TICK_NSEC;
2796 account_steal_time(st);
2803 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2805 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2807 * Account a tick to a process and cpustat
2808 * @p: the process that the cpu time gets accounted to
2809 * @user_tick: is the tick from userspace
2810 * @rq: the pointer to rq
2812 * Tick demultiplexing follows the order
2813 * - pending hardirq update
2814 * - pending softirq update
2818 * - check for guest_time
2819 * - else account as system_time
2821 * Check for hardirq is done both for system and user time as there is
2822 * no timer going off while we are on hardirq and hence we may never get an
2823 * opportunity to update it solely in system time.
2824 * p->stime and friends are only updated on system time and not on irq
2825 * softirq as those do not count in task exec_runtime any more.
2827 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2830 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2831 u64 *cpustat = kcpustat_this_cpu->cpustat;
2833 if (steal_account_process_tick())
2836 if (irqtime_account_hi_update()) {
2837 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2838 } else if (irqtime_account_si_update()) {
2839 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2840 } else if (this_cpu_ksoftirqd() == p) {
2842 * ksoftirqd time do not get accounted in cpu_softirq_time.
2843 * So, we have to handle it separately here.
2844 * Also, p->stime needs to be updated for ksoftirqd.
2846 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2848 } else if (user_tick) {
2849 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2850 } else if (p == rq->idle) {
2851 account_idle_time(cputime_one_jiffy);
2852 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2853 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2855 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2860 static void irqtime_account_idle_ticks(int ticks)
2863 struct rq *rq = this_rq();
2865 for (i = 0; i < ticks; i++)
2866 irqtime_account_process_tick(current, 0, rq);
2868 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2869 static void irqtime_account_idle_ticks(int ticks) {}
2870 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2872 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2875 * Account a single tick of cpu time.
2876 * @p: the process that the cpu time gets accounted to
2877 * @user_tick: indicates if the tick is a user or a system tick
2879 void account_process_tick(struct task_struct *p, int user_tick)
2881 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2882 struct rq *rq = this_rq();
2884 if (sched_clock_irqtime) {
2885 irqtime_account_process_tick(p, user_tick, rq);
2889 if (steal_account_process_tick())
2893 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2894 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2895 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2898 account_idle_time(cputime_one_jiffy);
2902 * Account multiple ticks of steal time.
2903 * @p: the process from which the cpu time has been stolen
2904 * @ticks: number of stolen ticks
2906 void account_steal_ticks(unsigned long ticks)
2908 account_steal_time(jiffies_to_cputime(ticks));
2912 * Account multiple ticks of idle time.
2913 * @ticks: number of stolen ticks
2915 void account_idle_ticks(unsigned long ticks)
2918 if (sched_clock_irqtime) {
2919 irqtime_account_idle_ticks(ticks);
2923 account_idle_time(jiffies_to_cputime(ticks));
2929 * Use precise platform statistics if available:
2931 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2932 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2938 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2940 struct task_cputime cputime;
2942 thread_group_cputime(p, &cputime);
2944 *ut = cputime.utime;
2945 *st = cputime.stime;
2949 #ifndef nsecs_to_cputime
2950 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2953 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2955 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2958 * Use CFS's precise accounting:
2960 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2963 u64 temp = (__force u64) rtime;
2965 temp *= (__force u64) utime;
2966 do_div(temp, (__force u32) total);
2967 utime = (__force cputime_t) temp;
2972 * Compare with previous values, to keep monotonicity:
2974 p->prev_utime = max(p->prev_utime, utime);
2975 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2977 *ut = p->prev_utime;
2978 *st = p->prev_stime;
2982 * Must be called with siglock held.
2984 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2986 struct signal_struct *sig = p->signal;
2987 struct task_cputime cputime;
2988 cputime_t rtime, utime, total;
2990 thread_group_cputime(p, &cputime);
2992 total = cputime.utime + cputime.stime;
2993 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2996 u64 temp = (__force u64) rtime;
2998 temp *= (__force u64) cputime.utime;
2999 do_div(temp, (__force u32) total);
3000 utime = (__force cputime_t) temp;
3004 sig->prev_utime = max(sig->prev_utime, utime);
3005 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3007 *ut = sig->prev_utime;
3008 *st = sig->prev_stime;
3013 * This function gets called by the timer code, with HZ frequency.
3014 * We call it with interrupts disabled.
3016 void scheduler_tick(void)
3018 int cpu = smp_processor_id();
3019 struct rq *rq = cpu_rq(cpu);
3020 struct task_struct *curr = rq->curr;
3024 raw_spin_lock(&rq->lock);
3025 update_rq_clock(rq);
3026 update_cpu_load_active(rq);
3027 curr->sched_class->task_tick(rq, curr, 0);
3028 raw_spin_unlock(&rq->lock);
3030 perf_event_task_tick();
3033 rq->idle_balance = idle_cpu(cpu);
3034 trigger_load_balance(rq, cpu);
3038 notrace unsigned long get_parent_ip(unsigned long addr)
3040 if (in_lock_functions(addr)) {
3041 addr = CALLER_ADDR2;
3042 if (in_lock_functions(addr))
3043 addr = CALLER_ADDR3;
3048 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3049 defined(CONFIG_PREEMPT_TRACER))
3051 void __kprobes add_preempt_count(int val)
3053 #ifdef CONFIG_DEBUG_PREEMPT
3057 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3060 preempt_count() += val;
3061 #ifdef CONFIG_DEBUG_PREEMPT
3063 * Spinlock count overflowing soon?
3065 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3068 if (preempt_count() == val)
3069 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3071 EXPORT_SYMBOL(add_preempt_count);
3073 void __kprobes sub_preempt_count(int val)
3075 #ifdef CONFIG_DEBUG_PREEMPT
3079 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3082 * Is the spinlock portion underflowing?
3084 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3085 !(preempt_count() & PREEMPT_MASK)))
3089 if (preempt_count() == val)
3090 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3091 preempt_count() -= val;
3093 EXPORT_SYMBOL(sub_preempt_count);
3098 * Print scheduling while atomic bug:
3100 static noinline void __schedule_bug(struct task_struct *prev)
3102 struct pt_regs *regs = get_irq_regs();
3104 if (oops_in_progress)
3107 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3108 prev->comm, prev->pid, preempt_count());
3110 debug_show_held_locks(prev);
3112 if (irqs_disabled())
3113 print_irqtrace_events(prev);
3122 * Various schedule()-time debugging checks and statistics:
3124 static inline void schedule_debug(struct task_struct *prev)
3127 * Test if we are atomic. Since do_exit() needs to call into
3128 * schedule() atomically, we ignore that path for now.
3129 * Otherwise, whine if we are scheduling when we should not be.
3131 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3132 __schedule_bug(prev);
3135 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3137 schedstat_inc(this_rq(), sched_count);
3140 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3142 if (prev->on_rq || rq->skip_clock_update < 0)
3143 update_rq_clock(rq);
3144 prev->sched_class->put_prev_task(rq, prev);
3148 * Pick up the highest-prio task:
3150 static inline struct task_struct *
3151 pick_next_task(struct rq *rq)
3153 const struct sched_class *class;
3154 struct task_struct *p;
3157 * Optimization: we know that if all tasks are in
3158 * the fair class we can call that function directly:
3160 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3161 p = fair_sched_class.pick_next_task(rq);
3166 for_each_class(class) {
3167 p = class->pick_next_task(rq);
3172 BUG(); /* the idle class will always have a runnable task */
3176 * __schedule() is the main scheduler function.
3178 static void __sched __schedule(void)
3180 struct task_struct *prev, *next;
3181 unsigned long *switch_count;
3187 cpu = smp_processor_id();
3189 rcu_note_context_switch(cpu);
3192 schedule_debug(prev);
3194 if (sched_feat(HRTICK))
3197 raw_spin_lock_irq(&rq->lock);
3199 switch_count = &prev->nivcsw;
3200 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3201 if (unlikely(signal_pending_state(prev->state, prev))) {
3202 prev->state = TASK_RUNNING;
3204 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3208 * If a worker went to sleep, notify and ask workqueue
3209 * whether it wants to wake up a task to maintain
3212 if (prev->flags & PF_WQ_WORKER) {
3213 struct task_struct *to_wakeup;
3215 to_wakeup = wq_worker_sleeping(prev, cpu);
3217 try_to_wake_up_local(to_wakeup);
3220 switch_count = &prev->nvcsw;
3223 pre_schedule(rq, prev);
3225 if (unlikely(!rq->nr_running))
3226 idle_balance(cpu, rq);
3228 put_prev_task(rq, prev);
3229 next = pick_next_task(rq);
3230 clear_tsk_need_resched(prev);
3231 rq->skip_clock_update = 0;
3233 if (likely(prev != next)) {
3238 context_switch(rq, prev, next); /* unlocks the rq */
3240 * The context switch have flipped the stack from under us
3241 * and restored the local variables which were saved when
3242 * this task called schedule() in the past. prev == current
3243 * is still correct, but it can be moved to another cpu/rq.
3245 cpu = smp_processor_id();
3248 raw_spin_unlock_irq(&rq->lock);
3252 sched_preempt_enable_no_resched();
3257 static inline void sched_submit_work(struct task_struct *tsk)
3259 if (!tsk->state || tsk_is_pi_blocked(tsk))
3262 * If we are going to sleep and we have plugged IO queued,
3263 * make sure to submit it to avoid deadlocks.
3265 if (blk_needs_flush_plug(tsk))
3266 blk_schedule_flush_plug(tsk);
3269 asmlinkage void __sched schedule(void)
3271 struct task_struct *tsk = current;
3273 sched_submit_work(tsk);
3276 EXPORT_SYMBOL(schedule);
3279 * schedule_preempt_disabled - called with preemption disabled
3281 * Returns with preemption disabled. Note: preempt_count must be 1
3283 void __sched schedule_preempt_disabled(void)
3285 sched_preempt_enable_no_resched();
3290 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3292 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3294 if (lock->owner != owner)
3298 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3299 * lock->owner still matches owner, if that fails, owner might
3300 * point to free()d memory, if it still matches, the rcu_read_lock()
3301 * ensures the memory stays valid.
3305 return owner->on_cpu;
3309 * Look out! "owner" is an entirely speculative pointer
3310 * access and not reliable.
3312 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3314 if (!sched_feat(OWNER_SPIN))
3318 while (owner_running(lock, owner)) {
3322 arch_mutex_cpu_relax();
3327 * We break out the loop above on need_resched() and when the
3328 * owner changed, which is a sign for heavy contention. Return
3329 * success only when lock->owner is NULL.
3331 return lock->owner == NULL;
3335 #ifdef CONFIG_PREEMPT
3337 * this is the entry point to schedule() from in-kernel preemption
3338 * off of preempt_enable. Kernel preemptions off return from interrupt
3339 * occur there and call schedule directly.
3341 asmlinkage void __sched notrace preempt_schedule(void)
3343 struct thread_info *ti = current_thread_info();
3346 * If there is a non-zero preempt_count or interrupts are disabled,
3347 * we do not want to preempt the current task. Just return..
3349 if (likely(ti->preempt_count || irqs_disabled()))
3353 add_preempt_count_notrace(PREEMPT_ACTIVE);
3355 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3358 * Check again in case we missed a preemption opportunity
3359 * between schedule and now.
3362 } while (need_resched());
3364 EXPORT_SYMBOL(preempt_schedule);
3367 * this is the entry point to schedule() from kernel preemption
3368 * off of irq context.
3369 * Note, that this is called and return with irqs disabled. This will
3370 * protect us against recursive calling from irq.
3372 asmlinkage void __sched preempt_schedule_irq(void)
3374 struct thread_info *ti = current_thread_info();
3376 /* Catch callers which need to be fixed */
3377 BUG_ON(ti->preempt_count || !irqs_disabled());
3380 add_preempt_count(PREEMPT_ACTIVE);
3383 local_irq_disable();
3384 sub_preempt_count(PREEMPT_ACTIVE);
3387 * Check again in case we missed a preemption opportunity
3388 * between schedule and now.
3391 } while (need_resched());
3394 #endif /* CONFIG_PREEMPT */
3396 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3399 return try_to_wake_up(curr->private, mode, wake_flags);
3401 EXPORT_SYMBOL(default_wake_function);
3404 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3405 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3406 * number) then we wake all the non-exclusive tasks and one exclusive task.
3408 * There are circumstances in which we can try to wake a task which has already
3409 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3410 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3412 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3413 int nr_exclusive, int wake_flags, void *key)
3415 wait_queue_t *curr, *next;
3417 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3418 unsigned flags = curr->flags;
3420 if (curr->func(curr, mode, wake_flags, key) &&
3421 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3427 * __wake_up - wake up threads blocked on a waitqueue.
3429 * @mode: which threads
3430 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3431 * @key: is directly passed to the wakeup function
3433 * It may be assumed that this function implies a write memory barrier before
3434 * changing the task state if and only if any tasks are woken up.
3436 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3437 int nr_exclusive, void *key)
3439 unsigned long flags;
3441 spin_lock_irqsave(&q->lock, flags);
3442 __wake_up_common(q, mode, nr_exclusive, 0, key);
3443 spin_unlock_irqrestore(&q->lock, flags);
3445 EXPORT_SYMBOL(__wake_up);
3448 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3450 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3452 __wake_up_common(q, mode, nr, 0, NULL);
3454 EXPORT_SYMBOL_GPL(__wake_up_locked);
3456 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3458 __wake_up_common(q, mode, 1, 0, key);
3460 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3463 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3465 * @mode: which threads
3466 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3467 * @key: opaque value to be passed to wakeup targets
3469 * The sync wakeup differs that the waker knows that it will schedule
3470 * away soon, so while the target thread will be woken up, it will not
3471 * be migrated to another CPU - ie. the two threads are 'synchronized'
3472 * with each other. This can prevent needless bouncing between CPUs.
3474 * On UP it can prevent extra preemption.
3476 * It may be assumed that this function implies a write memory barrier before
3477 * changing the task state if and only if any tasks are woken up.
3479 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3480 int nr_exclusive, void *key)
3482 unsigned long flags;
3483 int wake_flags = WF_SYNC;
3488 if (unlikely(!nr_exclusive))
3491 spin_lock_irqsave(&q->lock, flags);
3492 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3493 spin_unlock_irqrestore(&q->lock, flags);
3495 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3498 * __wake_up_sync - see __wake_up_sync_key()
3500 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3502 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3504 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3507 * complete: - signals a single thread waiting on this completion
3508 * @x: holds the state of this particular completion
3510 * This will wake up a single thread waiting on this completion. Threads will be
3511 * awakened in the same order in which they were queued.
3513 * See also complete_all(), wait_for_completion() and related routines.
3515 * It may be assumed that this function implies a write memory barrier before
3516 * changing the task state if and only if any tasks are woken up.
3518 void complete(struct completion *x)
3520 unsigned long flags;
3522 spin_lock_irqsave(&x->wait.lock, flags);
3524 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3525 spin_unlock_irqrestore(&x->wait.lock, flags);
3527 EXPORT_SYMBOL(complete);
3530 * complete_all: - signals all threads waiting on this completion
3531 * @x: holds the state of this particular completion
3533 * This will wake up all threads waiting on this particular completion event.
3535 * It may be assumed that this function implies a write memory barrier before
3536 * changing the task state if and only if any tasks are woken up.
3538 void complete_all(struct completion *x)
3540 unsigned long flags;
3542 spin_lock_irqsave(&x->wait.lock, flags);
3543 x->done += UINT_MAX/2;
3544 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3545 spin_unlock_irqrestore(&x->wait.lock, flags);
3547 EXPORT_SYMBOL(complete_all);
3549 static inline long __sched
3550 do_wait_for_common(struct completion *x, long timeout, int state)
3553 DECLARE_WAITQUEUE(wait, current);
3555 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3557 if (signal_pending_state(state, current)) {
3558 timeout = -ERESTARTSYS;
3561 __set_current_state(state);
3562 spin_unlock_irq(&x->wait.lock);
3563 timeout = schedule_timeout(timeout);
3564 spin_lock_irq(&x->wait.lock);
3565 } while (!x->done && timeout);
3566 __remove_wait_queue(&x->wait, &wait);
3571 return timeout ?: 1;
3575 wait_for_common(struct completion *x, long timeout, int state)
3579 spin_lock_irq(&x->wait.lock);
3580 timeout = do_wait_for_common(x, timeout, state);
3581 spin_unlock_irq(&x->wait.lock);
3586 * wait_for_completion: - waits for completion of a task
3587 * @x: holds the state of this particular completion
3589 * This waits to be signaled for completion of a specific task. It is NOT
3590 * interruptible and there is no timeout.
3592 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3593 * and interrupt capability. Also see complete().
3595 void __sched wait_for_completion(struct completion *x)
3597 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3599 EXPORT_SYMBOL(wait_for_completion);
3602 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3603 * @x: holds the state of this particular completion
3604 * @timeout: timeout value in jiffies
3606 * This waits for either a completion of a specific task to be signaled or for a
3607 * specified timeout to expire. The timeout is in jiffies. It is not
3610 * The return value is 0 if timed out, and positive (at least 1, or number of
3611 * jiffies left till timeout) if completed.
3613 unsigned long __sched
3614 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3616 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3618 EXPORT_SYMBOL(wait_for_completion_timeout);
3621 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3622 * @x: holds the state of this particular completion
3624 * This waits for completion of a specific task to be signaled. It is
3627 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3629 int __sched wait_for_completion_interruptible(struct completion *x)
3631 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3632 if (t == -ERESTARTSYS)
3636 EXPORT_SYMBOL(wait_for_completion_interruptible);
3639 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3640 * @x: holds the state of this particular completion
3641 * @timeout: timeout value in jiffies
3643 * This waits for either a completion of a specific task to be signaled or for a
3644 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3646 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3647 * positive (at least 1, or number of jiffies left till timeout) if completed.
3650 wait_for_completion_interruptible_timeout(struct completion *x,
3651 unsigned long timeout)
3653 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3655 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3658 * wait_for_completion_killable: - waits for completion of a task (killable)
3659 * @x: holds the state of this particular completion
3661 * This waits to be signaled for completion of a specific task. It can be
3662 * interrupted by a kill signal.
3664 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3666 int __sched wait_for_completion_killable(struct completion *x)
3668 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3669 if (t == -ERESTARTSYS)
3673 EXPORT_SYMBOL(wait_for_completion_killable);
3676 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3677 * @x: holds the state of this particular completion
3678 * @timeout: timeout value in jiffies
3680 * This waits for either a completion of a specific task to be
3681 * signaled or for a specified timeout to expire. It can be
3682 * interrupted by a kill signal. The timeout is in jiffies.
3684 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3685 * positive (at least 1, or number of jiffies left till timeout) if completed.
3688 wait_for_completion_killable_timeout(struct completion *x,
3689 unsigned long timeout)
3691 return wait_for_common(x, timeout, TASK_KILLABLE);
3693 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3696 * try_wait_for_completion - try to decrement a completion without blocking
3697 * @x: completion structure
3699 * Returns: 0 if a decrement cannot be done without blocking
3700 * 1 if a decrement succeeded.
3702 * If a completion is being used as a counting completion,
3703 * attempt to decrement the counter without blocking. This
3704 * enables us to avoid waiting if the resource the completion
3705 * is protecting is not available.
3707 bool try_wait_for_completion(struct completion *x)
3709 unsigned long flags;
3712 spin_lock_irqsave(&x->wait.lock, flags);
3717 spin_unlock_irqrestore(&x->wait.lock, flags);
3720 EXPORT_SYMBOL(try_wait_for_completion);
3723 * completion_done - Test to see if a completion has any waiters
3724 * @x: completion structure
3726 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3727 * 1 if there are no waiters.
3730 bool completion_done(struct completion *x)
3732 unsigned long flags;
3735 spin_lock_irqsave(&x->wait.lock, flags);
3738 spin_unlock_irqrestore(&x->wait.lock, flags);
3741 EXPORT_SYMBOL(completion_done);
3744 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3746 unsigned long flags;
3749 init_waitqueue_entry(&wait, current);
3751 __set_current_state(state);
3753 spin_lock_irqsave(&q->lock, flags);
3754 __add_wait_queue(q, &wait);
3755 spin_unlock(&q->lock);
3756 timeout = schedule_timeout(timeout);
3757 spin_lock_irq(&q->lock);
3758 __remove_wait_queue(q, &wait);
3759 spin_unlock_irqrestore(&q->lock, flags);
3764 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3766 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3768 EXPORT_SYMBOL(interruptible_sleep_on);
3771 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3773 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3775 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3777 void __sched sleep_on(wait_queue_head_t *q)
3779 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3781 EXPORT_SYMBOL(sleep_on);
3783 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3785 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3787 EXPORT_SYMBOL(sleep_on_timeout);
3789 #ifdef CONFIG_RT_MUTEXES
3792 * rt_mutex_setprio - set the current priority of a task
3794 * @prio: prio value (kernel-internal form)
3796 * This function changes the 'effective' priority of a task. It does
3797 * not touch ->normal_prio like __setscheduler().
3799 * Used by the rt_mutex code to implement priority inheritance logic.
3801 void rt_mutex_setprio(struct task_struct *p, int prio)
3803 int oldprio, on_rq, running;
3805 const struct sched_class *prev_class;
3807 BUG_ON(prio < 0 || prio > MAX_PRIO);
3809 rq = __task_rq_lock(p);
3812 * Idle task boosting is a nono in general. There is one
3813 * exception, when PREEMPT_RT and NOHZ is active:
3815 * The idle task calls get_next_timer_interrupt() and holds
3816 * the timer wheel base->lock on the CPU and another CPU wants
3817 * to access the timer (probably to cancel it). We can safely
3818 * ignore the boosting request, as the idle CPU runs this code
3819 * with interrupts disabled and will complete the lock
3820 * protected section without being interrupted. So there is no
3821 * real need to boost.
3823 if (unlikely(p == rq->idle)) {
3824 WARN_ON(p != rq->curr);
3825 WARN_ON(p->pi_blocked_on);
3829 trace_sched_pi_setprio(p, prio);
3831 prev_class = p->sched_class;
3833 running = task_current(rq, p);
3835 dequeue_task(rq, p, 0);
3837 p->sched_class->put_prev_task(rq, p);
3840 p->sched_class = &rt_sched_class;
3842 p->sched_class = &fair_sched_class;
3847 p->sched_class->set_curr_task(rq);
3849 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3851 check_class_changed(rq, p, prev_class, oldprio);
3853 __task_rq_unlock(rq);
3856 void set_user_nice(struct task_struct *p, long nice)
3858 int old_prio, delta, on_rq;
3859 unsigned long flags;
3862 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3865 * We have to be careful, if called from sys_setpriority(),
3866 * the task might be in the middle of scheduling on another CPU.
3868 rq = task_rq_lock(p, &flags);
3870 * The RT priorities are set via sched_setscheduler(), but we still
3871 * allow the 'normal' nice value to be set - but as expected
3872 * it wont have any effect on scheduling until the task is
3873 * SCHED_FIFO/SCHED_RR:
3875 if (task_has_rt_policy(p)) {
3876 p->static_prio = NICE_TO_PRIO(nice);
3881 dequeue_task(rq, p, 0);
3883 p->static_prio = NICE_TO_PRIO(nice);
3886 p->prio = effective_prio(p);
3887 delta = p->prio - old_prio;
3890 enqueue_task(rq, p, 0);
3892 * If the task increased its priority or is running and
3893 * lowered its priority, then reschedule its CPU:
3895 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3896 resched_task(rq->curr);
3899 task_rq_unlock(rq, p, &flags);
3901 EXPORT_SYMBOL(set_user_nice);
3904 * can_nice - check if a task can reduce its nice value
3908 int can_nice(const struct task_struct *p, const int nice)
3910 /* convert nice value [19,-20] to rlimit style value [1,40] */
3911 int nice_rlim = 20 - nice;
3913 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3914 capable(CAP_SYS_NICE));
3917 #ifdef __ARCH_WANT_SYS_NICE
3920 * sys_nice - change the priority of the current process.
3921 * @increment: priority increment
3923 * sys_setpriority is a more generic, but much slower function that
3924 * does similar things.
3926 SYSCALL_DEFINE1(nice, int, increment)
3931 * Setpriority might change our priority at the same moment.
3932 * We don't have to worry. Conceptually one call occurs first
3933 * and we have a single winner.
3935 if (increment < -40)
3940 nice = TASK_NICE(current) + increment;
3946 if (increment < 0 && !can_nice(current, nice))
3949 retval = security_task_setnice(current, nice);
3953 set_user_nice(current, nice);
3960 * task_prio - return the priority value of a given task.
3961 * @p: the task in question.
3963 * This is the priority value as seen by users in /proc.
3964 * RT tasks are offset by -200. Normal tasks are centered
3965 * around 0, value goes from -16 to +15.
3967 int task_prio(const struct task_struct *p)
3969 return p->prio - MAX_RT_PRIO;
3973 * task_nice - return the nice value of a given task.
3974 * @p: the task in question.
3976 int task_nice(const struct task_struct *p)
3978 return TASK_NICE(p);
3980 EXPORT_SYMBOL(task_nice);
3983 * idle_cpu - is a given cpu idle currently?
3984 * @cpu: the processor in question.
3986 int idle_cpu(int cpu)
3988 struct rq *rq = cpu_rq(cpu);
3990 if (rq->curr != rq->idle)
3997 if (!llist_empty(&rq->wake_list))
4005 * idle_task - return the idle task for a given cpu.
4006 * @cpu: the processor in question.
4008 struct task_struct *idle_task(int cpu)
4010 return cpu_rq(cpu)->idle;
4014 * find_process_by_pid - find a process with a matching PID value.
4015 * @pid: the pid in question.
4017 static struct task_struct *find_process_by_pid(pid_t pid)
4019 return pid ? find_task_by_vpid(pid) : current;
4022 /* Actually do priority change: must hold rq lock. */
4024 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4027 p->rt_priority = prio;
4028 p->normal_prio = normal_prio(p);
4029 /* we are holding p->pi_lock already */
4030 p->prio = rt_mutex_getprio(p);
4031 if (rt_prio(p->prio))
4032 p->sched_class = &rt_sched_class;
4034 p->sched_class = &fair_sched_class;
4039 * check the target process has a UID that matches the current process's
4041 static bool check_same_owner(struct task_struct *p)
4043 const struct cred *cred = current_cred(), *pcred;
4047 pcred = __task_cred(p);
4048 if (cred->user->user_ns == pcred->user->user_ns)
4049 match = (cred->euid == pcred->euid ||
4050 cred->euid == pcred->uid);
4057 static int __sched_setscheduler(struct task_struct *p, int policy,
4058 const struct sched_param *param, bool user)
4060 int retval, oldprio, oldpolicy = -1, on_rq, running;
4061 unsigned long flags;
4062 const struct sched_class *prev_class;
4066 /* may grab non-irq protected spin_locks */
4067 BUG_ON(in_interrupt());
4069 /* double check policy once rq lock held */
4071 reset_on_fork = p->sched_reset_on_fork;
4072 policy = oldpolicy = p->policy;
4074 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4075 policy &= ~SCHED_RESET_ON_FORK;
4077 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4078 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4079 policy != SCHED_IDLE)
4084 * Valid priorities for SCHED_FIFO and SCHED_RR are
4085 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4086 * SCHED_BATCH and SCHED_IDLE is 0.
4088 if (param->sched_priority < 0 ||
4089 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4090 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4092 if (rt_policy(policy) != (param->sched_priority != 0))
4096 * Allow unprivileged RT tasks to decrease priority:
4098 if (user && !capable(CAP_SYS_NICE)) {
4099 if (rt_policy(policy)) {
4100 unsigned long rlim_rtprio =
4101 task_rlimit(p, RLIMIT_RTPRIO);
4103 /* can't set/change the rt policy */
4104 if (policy != p->policy && !rlim_rtprio)
4107 /* can't increase priority */
4108 if (param->sched_priority > p->rt_priority &&
4109 param->sched_priority > rlim_rtprio)
4114 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4115 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4117 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4118 if (!can_nice(p, TASK_NICE(p)))
4122 /* can't change other user's priorities */
4123 if (!check_same_owner(p))
4126 /* Normal users shall not reset the sched_reset_on_fork flag */
4127 if (p->sched_reset_on_fork && !reset_on_fork)
4132 retval = security_task_setscheduler(p);
4138 * make sure no PI-waiters arrive (or leave) while we are
4139 * changing the priority of the task:
4141 * To be able to change p->policy safely, the appropriate
4142 * runqueue lock must be held.
4144 rq = task_rq_lock(p, &flags);
4147 * Changing the policy of the stop threads its a very bad idea
4149 if (p == rq->stop) {
4150 task_rq_unlock(rq, p, &flags);
4155 * If not changing anything there's no need to proceed further:
4157 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4158 param->sched_priority == p->rt_priority))) {
4160 __task_rq_unlock(rq);
4161 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4165 #ifdef CONFIG_RT_GROUP_SCHED
4168 * Do not allow realtime tasks into groups that have no runtime
4171 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4172 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4173 !task_group_is_autogroup(task_group(p))) {
4174 task_rq_unlock(rq, p, &flags);
4180 /* recheck policy now with rq lock held */
4181 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4182 policy = oldpolicy = -1;
4183 task_rq_unlock(rq, p, &flags);
4187 running = task_current(rq, p);
4189 dequeue_task(rq, p, 0);
4191 p->sched_class->put_prev_task(rq, p);
4193 p->sched_reset_on_fork = reset_on_fork;
4196 prev_class = p->sched_class;
4197 __setscheduler(rq, p, policy, param->sched_priority);
4200 p->sched_class->set_curr_task(rq);
4202 enqueue_task(rq, p, 0);
4204 check_class_changed(rq, p, prev_class, oldprio);
4205 task_rq_unlock(rq, p, &flags);
4207 rt_mutex_adjust_pi(p);
4213 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4214 * @p: the task in question.
4215 * @policy: new policy.
4216 * @param: structure containing the new RT priority.
4218 * NOTE that the task may be already dead.
4220 int sched_setscheduler(struct task_struct *p, int policy,
4221 const struct sched_param *param)
4223 return __sched_setscheduler(p, policy, param, true);
4225 EXPORT_SYMBOL_GPL(sched_setscheduler);
4228 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4229 * @p: the task in question.
4230 * @policy: new policy.
4231 * @param: structure containing the new RT priority.
4233 * Just like sched_setscheduler, only don't bother checking if the
4234 * current context has permission. For example, this is needed in
4235 * stop_machine(): we create temporary high priority worker threads,
4236 * but our caller might not have that capability.
4238 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4239 const struct sched_param *param)
4241 return __sched_setscheduler(p, policy, param, false);
4245 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4247 struct sched_param lparam;
4248 struct task_struct *p;
4251 if (!param || pid < 0)
4253 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4258 p = find_process_by_pid(pid);
4260 retval = sched_setscheduler(p, policy, &lparam);
4267 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4268 * @pid: the pid in question.
4269 * @policy: new policy.
4270 * @param: structure containing the new RT priority.
4272 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4273 struct sched_param __user *, param)
4275 /* negative values for policy are not valid */
4279 return do_sched_setscheduler(pid, policy, param);
4283 * sys_sched_setparam - set/change the RT priority of a thread
4284 * @pid: the pid in question.
4285 * @param: structure containing the new RT priority.
4287 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4289 return do_sched_setscheduler(pid, -1, param);
4293 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4294 * @pid: the pid in question.
4296 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4298 struct task_struct *p;
4306 p = find_process_by_pid(pid);
4308 retval = security_task_getscheduler(p);
4311 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4318 * sys_sched_getparam - get the RT priority of a thread
4319 * @pid: the pid in question.
4320 * @param: structure containing the RT priority.
4322 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4324 struct sched_param lp;
4325 struct task_struct *p;
4328 if (!param || pid < 0)
4332 p = find_process_by_pid(pid);
4337 retval = security_task_getscheduler(p);
4341 lp.sched_priority = p->rt_priority;
4345 * This one might sleep, we cannot do it with a spinlock held ...
4347 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4356 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4358 cpumask_var_t cpus_allowed, new_mask;
4359 struct task_struct *p;
4365 p = find_process_by_pid(pid);
4372 /* Prevent p going away */
4376 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4380 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4382 goto out_free_cpus_allowed;
4385 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4388 retval = security_task_setscheduler(p);
4392 cpuset_cpus_allowed(p, cpus_allowed);
4393 cpumask_and(new_mask, in_mask, cpus_allowed);
4395 retval = set_cpus_allowed_ptr(p, new_mask);
4398 cpuset_cpus_allowed(p, cpus_allowed);
4399 if (!cpumask_subset(new_mask, cpus_allowed)) {
4401 * We must have raced with a concurrent cpuset
4402 * update. Just reset the cpus_allowed to the
4403 * cpuset's cpus_allowed
4405 cpumask_copy(new_mask, cpus_allowed);
4410 free_cpumask_var(new_mask);
4411 out_free_cpus_allowed:
4412 free_cpumask_var(cpus_allowed);
4419 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4420 struct cpumask *new_mask)
4422 if (len < cpumask_size())
4423 cpumask_clear(new_mask);
4424 else if (len > cpumask_size())
4425 len = cpumask_size();
4427 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4431 * sys_sched_setaffinity - set the cpu affinity of a process
4432 * @pid: pid of the process
4433 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4434 * @user_mask_ptr: user-space pointer to the new cpu mask
4436 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4437 unsigned long __user *, user_mask_ptr)
4439 cpumask_var_t new_mask;
4442 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4445 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4447 retval = sched_setaffinity(pid, new_mask);
4448 free_cpumask_var(new_mask);
4452 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4454 struct task_struct *p;
4455 unsigned long flags;
4462 p = find_process_by_pid(pid);
4466 retval = security_task_getscheduler(p);
4470 raw_spin_lock_irqsave(&p->pi_lock, flags);
4471 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4472 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4482 * sys_sched_getaffinity - get the cpu affinity of a process
4483 * @pid: pid of the process
4484 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4485 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4487 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4488 unsigned long __user *, user_mask_ptr)
4493 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4495 if (len & (sizeof(unsigned long)-1))
4498 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4501 ret = sched_getaffinity(pid, mask);
4503 size_t retlen = min_t(size_t, len, cpumask_size());
4505 if (copy_to_user(user_mask_ptr, mask, retlen))
4510 free_cpumask_var(mask);
4516 * sys_sched_yield - yield the current processor to other threads.
4518 * This function yields the current CPU to other tasks. If there are no
4519 * other threads running on this CPU then this function will return.
4521 SYSCALL_DEFINE0(sched_yield)
4523 struct rq *rq = this_rq_lock();
4525 schedstat_inc(rq, yld_count);
4526 current->sched_class->yield_task(rq);
4529 * Since we are going to call schedule() anyway, there's
4530 * no need to preempt or enable interrupts:
4532 __release(rq->lock);
4533 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4534 do_raw_spin_unlock(&rq->lock);
4535 sched_preempt_enable_no_resched();
4542 static inline int should_resched(void)
4544 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4547 static void __cond_resched(void)
4549 add_preempt_count(PREEMPT_ACTIVE);
4551 sub_preempt_count(PREEMPT_ACTIVE);
4554 int __sched _cond_resched(void)
4556 if (should_resched()) {
4562 EXPORT_SYMBOL(_cond_resched);
4565 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4566 * call schedule, and on return reacquire the lock.
4568 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4569 * operations here to prevent schedule() from being called twice (once via
4570 * spin_unlock(), once by hand).
4572 int __cond_resched_lock(spinlock_t *lock)
4574 int resched = should_resched();
4577 lockdep_assert_held(lock);
4579 if (spin_needbreak(lock) || resched) {
4590 EXPORT_SYMBOL(__cond_resched_lock);
4592 int __sched __cond_resched_softirq(void)
4594 BUG_ON(!in_softirq());
4596 if (should_resched()) {
4604 EXPORT_SYMBOL(__cond_resched_softirq);
4607 * yield - yield the current processor to other threads.
4609 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4611 * The scheduler is at all times free to pick the calling task as the most
4612 * eligible task to run, if removing the yield() call from your code breaks
4613 * it, its already broken.
4615 * Typical broken usage is:
4620 * where one assumes that yield() will let 'the other' process run that will
4621 * make event true. If the current task is a SCHED_FIFO task that will never
4622 * happen. Never use yield() as a progress guarantee!!
4624 * If you want to use yield() to wait for something, use wait_event().
4625 * If you want to use yield() to be 'nice' for others, use cond_resched().
4626 * If you still want to use yield(), do not!
4628 void __sched yield(void)
4630 set_current_state(TASK_RUNNING);
4633 EXPORT_SYMBOL(yield);
4636 * yield_to - yield the current processor to another thread in
4637 * your thread group, or accelerate that thread toward the
4638 * processor it's on.
4640 * @preempt: whether task preemption is allowed or not
4642 * It's the caller's job to ensure that the target task struct
4643 * can't go away on us before we can do any checks.
4645 * Returns true if we indeed boosted the target task.
4647 bool __sched yield_to(struct task_struct *p, bool preempt)
4649 struct task_struct *curr = current;
4650 struct rq *rq, *p_rq;
4651 unsigned long flags;
4654 local_irq_save(flags);
4659 double_rq_lock(rq, p_rq);
4660 while (task_rq(p) != p_rq) {
4661 double_rq_unlock(rq, p_rq);
4665 if (!curr->sched_class->yield_to_task)
4668 if (curr->sched_class != p->sched_class)
4671 if (task_running(p_rq, p) || p->state)
4674 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4676 schedstat_inc(rq, yld_count);
4678 * Make p's CPU reschedule; pick_next_entity takes care of
4681 if (preempt && rq != p_rq)
4682 resched_task(p_rq->curr);
4685 * We might have set it in task_yield_fair(), but are
4686 * not going to schedule(), so don't want to skip
4689 rq->skip_clock_update = 0;
4693 double_rq_unlock(rq, p_rq);
4694 local_irq_restore(flags);
4701 EXPORT_SYMBOL_GPL(yield_to);
4704 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4705 * that process accounting knows that this is a task in IO wait state.
4707 void __sched io_schedule(void)
4709 struct rq *rq = raw_rq();
4711 delayacct_blkio_start();
4712 atomic_inc(&rq->nr_iowait);
4713 blk_flush_plug(current);
4714 current->in_iowait = 1;
4716 current->in_iowait = 0;
4717 atomic_dec(&rq->nr_iowait);
4718 delayacct_blkio_end();
4720 EXPORT_SYMBOL(io_schedule);
4722 long __sched io_schedule_timeout(long timeout)
4724 struct rq *rq = raw_rq();
4727 delayacct_blkio_start();
4728 atomic_inc(&rq->nr_iowait);
4729 blk_flush_plug(current);
4730 current->in_iowait = 1;
4731 ret = schedule_timeout(timeout);
4732 current->in_iowait = 0;
4733 atomic_dec(&rq->nr_iowait);
4734 delayacct_blkio_end();
4739 * sys_sched_get_priority_max - return maximum RT priority.
4740 * @policy: scheduling class.
4742 * this syscall returns the maximum rt_priority that can be used
4743 * by a given scheduling class.
4745 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4752 ret = MAX_USER_RT_PRIO-1;
4764 * sys_sched_get_priority_min - return minimum RT priority.
4765 * @policy: scheduling class.
4767 * this syscall returns the minimum rt_priority that can be used
4768 * by a given scheduling class.
4770 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4788 * sys_sched_rr_get_interval - return the default timeslice of a process.
4789 * @pid: pid of the process.
4790 * @interval: userspace pointer to the timeslice value.
4792 * this syscall writes the default timeslice value of a given process
4793 * into the user-space timespec buffer. A value of '0' means infinity.
4795 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4796 struct timespec __user *, interval)
4798 struct task_struct *p;
4799 unsigned int time_slice;
4800 unsigned long flags;
4810 p = find_process_by_pid(pid);
4814 retval = security_task_getscheduler(p);
4818 rq = task_rq_lock(p, &flags);
4819 time_slice = p->sched_class->get_rr_interval(rq, p);
4820 task_rq_unlock(rq, p, &flags);
4823 jiffies_to_timespec(time_slice, &t);
4824 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4832 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4834 void sched_show_task(struct task_struct *p)
4836 unsigned long free = 0;
4839 state = p->state ? __ffs(p->state) + 1 : 0;
4840 printk(KERN_INFO "%-15.15s %c", p->comm,
4841 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4842 #if BITS_PER_LONG == 32
4843 if (state == TASK_RUNNING)
4844 printk(KERN_CONT " running ");
4846 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4848 if (state == TASK_RUNNING)
4849 printk(KERN_CONT " running task ");
4851 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4853 #ifdef CONFIG_DEBUG_STACK_USAGE
4854 free = stack_not_used(p);
4856 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4857 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4858 (unsigned long)task_thread_info(p)->flags);
4860 show_stack(p, NULL);
4863 void show_state_filter(unsigned long state_filter)
4865 struct task_struct *g, *p;
4867 #if BITS_PER_LONG == 32
4869 " task PC stack pid father\n");
4872 " task PC stack pid father\n");
4875 do_each_thread(g, p) {
4877 * reset the NMI-timeout, listing all files on a slow
4878 * console might take a lot of time:
4880 touch_nmi_watchdog();
4881 if (!state_filter || (p->state & state_filter))
4883 } while_each_thread(g, p);
4885 touch_all_softlockup_watchdogs();
4887 #ifdef CONFIG_SCHED_DEBUG
4888 sysrq_sched_debug_show();
4892 * Only show locks if all tasks are dumped:
4895 debug_show_all_locks();
4898 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4900 idle->sched_class = &idle_sched_class;
4904 * init_idle - set up an idle thread for a given CPU
4905 * @idle: task in question
4906 * @cpu: cpu the idle task belongs to
4908 * NOTE: this function does not set the idle thread's NEED_RESCHED
4909 * flag, to make booting more robust.
4911 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4913 struct rq *rq = cpu_rq(cpu);
4914 unsigned long flags;
4916 raw_spin_lock_irqsave(&rq->lock, flags);
4919 idle->state = TASK_RUNNING;
4920 idle->se.exec_start = sched_clock();
4922 do_set_cpus_allowed(idle, cpumask_of(cpu));
4924 * We're having a chicken and egg problem, even though we are
4925 * holding rq->lock, the cpu isn't yet set to this cpu so the
4926 * lockdep check in task_group() will fail.
4928 * Similar case to sched_fork(). / Alternatively we could
4929 * use task_rq_lock() here and obtain the other rq->lock.
4934 __set_task_cpu(idle, cpu);
4937 rq->curr = rq->idle = idle;
4938 #if defined(CONFIG_SMP)
4941 raw_spin_unlock_irqrestore(&rq->lock, flags);
4943 /* Set the preempt count _outside_ the spinlocks! */
4944 task_thread_info(idle)->preempt_count = 0;
4947 * The idle tasks have their own, simple scheduling class:
4949 idle->sched_class = &idle_sched_class;
4950 ftrace_graph_init_idle_task(idle, cpu);
4951 #if defined(CONFIG_SMP)
4952 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4957 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4959 if (p->sched_class && p->sched_class->set_cpus_allowed)
4960 p->sched_class->set_cpus_allowed(p, new_mask);
4962 cpumask_copy(&p->cpus_allowed, new_mask);
4963 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4967 * This is how migration works:
4969 * 1) we invoke migration_cpu_stop() on the target CPU using
4971 * 2) stopper starts to run (implicitly forcing the migrated thread
4973 * 3) it checks whether the migrated task is still in the wrong runqueue.
4974 * 4) if it's in the wrong runqueue then the migration thread removes
4975 * it and puts it into the right queue.
4976 * 5) stopper completes and stop_one_cpu() returns and the migration
4981 * Change a given task's CPU affinity. Migrate the thread to a
4982 * proper CPU and schedule it away if the CPU it's executing on
4983 * is removed from the allowed bitmask.
4985 * NOTE: the caller must have a valid reference to the task, the
4986 * task must not exit() & deallocate itself prematurely. The
4987 * call is not atomic; no spinlocks may be held.
4989 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4991 unsigned long flags;
4993 unsigned int dest_cpu;
4996 rq = task_rq_lock(p, &flags);
4998 if (cpumask_equal(&p->cpus_allowed, new_mask))
5001 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5006 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5011 do_set_cpus_allowed(p, new_mask);
5013 /* Can the task run on the task's current CPU? If so, we're done */
5014 if (cpumask_test_cpu(task_cpu(p), new_mask))
5017 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5019 struct migration_arg arg = { p, dest_cpu };
5020 /* Need help from migration thread: drop lock and wait. */
5021 task_rq_unlock(rq, p, &flags);
5022 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5023 tlb_migrate_finish(p->mm);
5027 task_rq_unlock(rq, p, &flags);
5031 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5034 * Move (not current) task off this cpu, onto dest cpu. We're doing
5035 * this because either it can't run here any more (set_cpus_allowed()
5036 * away from this CPU, or CPU going down), or because we're
5037 * attempting to rebalance this task on exec (sched_exec).
5039 * So we race with normal scheduler movements, but that's OK, as long
5040 * as the task is no longer on this CPU.
5042 * Returns non-zero if task was successfully migrated.
5044 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5046 struct rq *rq_dest, *rq_src;
5049 if (unlikely(!cpu_active(dest_cpu)))
5052 rq_src = cpu_rq(src_cpu);
5053 rq_dest = cpu_rq(dest_cpu);
5055 raw_spin_lock(&p->pi_lock);
5056 double_rq_lock(rq_src, rq_dest);
5057 /* Already moved. */
5058 if (task_cpu(p) != src_cpu)
5060 /* Affinity changed (again). */
5061 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5065 * If we're not on a rq, the next wake-up will ensure we're
5069 dequeue_task(rq_src, p, 0);
5070 set_task_cpu(p, dest_cpu);
5071 enqueue_task(rq_dest, p, 0);
5072 check_preempt_curr(rq_dest, p, 0);
5077 double_rq_unlock(rq_src, rq_dest);
5078 raw_spin_unlock(&p->pi_lock);
5083 * migration_cpu_stop - this will be executed by a highprio stopper thread
5084 * and performs thread migration by bumping thread off CPU then
5085 * 'pushing' onto another runqueue.
5087 static int migration_cpu_stop(void *data)
5089 struct migration_arg *arg = data;
5092 * The original target cpu might have gone down and we might
5093 * be on another cpu but it doesn't matter.
5095 local_irq_disable();
5096 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5101 #ifdef CONFIG_HOTPLUG_CPU
5104 * Ensures that the idle task is using init_mm right before its cpu goes
5107 void idle_task_exit(void)
5109 struct mm_struct *mm = current->active_mm;
5111 BUG_ON(cpu_online(smp_processor_id()));
5114 switch_mm(mm, &init_mm, current);
5119 * While a dead CPU has no uninterruptible tasks queued at this point,
5120 * it might still have a nonzero ->nr_uninterruptible counter, because
5121 * for performance reasons the counter is not stricly tracking tasks to
5122 * their home CPUs. So we just add the counter to another CPU's counter,
5123 * to keep the global sum constant after CPU-down:
5125 static void migrate_nr_uninterruptible(struct rq *rq_src)
5127 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5129 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5130 rq_src->nr_uninterruptible = 0;
5134 * remove the tasks which were accounted by rq from calc_load_tasks.
5136 static void calc_global_load_remove(struct rq *rq)
5138 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5139 rq->calc_load_active = 0;
5143 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5144 * try_to_wake_up()->select_task_rq().
5146 * Called with rq->lock held even though we'er in stop_machine() and
5147 * there's no concurrency possible, we hold the required locks anyway
5148 * because of lock validation efforts.
5150 static void migrate_tasks(unsigned int dead_cpu)
5152 struct rq *rq = cpu_rq(dead_cpu);
5153 struct task_struct *next, *stop = rq->stop;
5157 * Fudge the rq selection such that the below task selection loop
5158 * doesn't get stuck on the currently eligible stop task.
5160 * We're currently inside stop_machine() and the rq is either stuck
5161 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5162 * either way we should never end up calling schedule() until we're
5167 /* Ensure any throttled groups are reachable by pick_next_task */
5168 unthrottle_offline_cfs_rqs(rq);
5172 * There's this thread running, bail when that's the only
5175 if (rq->nr_running == 1)
5178 next = pick_next_task(rq);
5180 next->sched_class->put_prev_task(rq, next);
5182 /* Find suitable destination for @next, with force if needed. */
5183 dest_cpu = select_fallback_rq(dead_cpu, next);
5184 raw_spin_unlock(&rq->lock);
5186 __migrate_task(next, dead_cpu, dest_cpu);
5188 raw_spin_lock(&rq->lock);
5194 #endif /* CONFIG_HOTPLUG_CPU */
5196 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5198 static struct ctl_table sd_ctl_dir[] = {
5200 .procname = "sched_domain",
5206 static struct ctl_table sd_ctl_root[] = {
5208 .procname = "kernel",
5210 .child = sd_ctl_dir,
5215 static struct ctl_table *sd_alloc_ctl_entry(int n)
5217 struct ctl_table *entry =
5218 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5223 static void sd_free_ctl_entry(struct ctl_table **tablep)
5225 struct ctl_table *entry;
5228 * In the intermediate directories, both the child directory and
5229 * procname are dynamically allocated and could fail but the mode
5230 * will always be set. In the lowest directory the names are
5231 * static strings and all have proc handlers.
5233 for (entry = *tablep; entry->mode; entry++) {
5235 sd_free_ctl_entry(&entry->child);
5236 if (entry->proc_handler == NULL)
5237 kfree(entry->procname);
5245 set_table_entry(struct ctl_table *entry,
5246 const char *procname, void *data, int maxlen,
5247 umode_t mode, proc_handler *proc_handler)
5249 entry->procname = procname;
5251 entry->maxlen = maxlen;
5253 entry->proc_handler = proc_handler;
5256 static struct ctl_table *
5257 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5259 struct ctl_table *table = sd_alloc_ctl_entry(13);
5264 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5265 sizeof(long), 0644, proc_doulongvec_minmax);
5266 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5267 sizeof(long), 0644, proc_doulongvec_minmax);
5268 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5269 sizeof(int), 0644, proc_dointvec_minmax);
5270 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5271 sizeof(int), 0644, proc_dointvec_minmax);
5272 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5273 sizeof(int), 0644, proc_dointvec_minmax);
5274 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5275 sizeof(int), 0644, proc_dointvec_minmax);
5276 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5277 sizeof(int), 0644, proc_dointvec_minmax);
5278 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5279 sizeof(int), 0644, proc_dointvec_minmax);
5280 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5281 sizeof(int), 0644, proc_dointvec_minmax);
5282 set_table_entry(&table[9], "cache_nice_tries",
5283 &sd->cache_nice_tries,
5284 sizeof(int), 0644, proc_dointvec_minmax);
5285 set_table_entry(&table[10], "flags", &sd->flags,
5286 sizeof(int), 0644, proc_dointvec_minmax);
5287 set_table_entry(&table[11], "name", sd->name,
5288 CORENAME_MAX_SIZE, 0444, proc_dostring);
5289 /* &table[12] is terminator */
5294 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5296 struct ctl_table *entry, *table;
5297 struct sched_domain *sd;
5298 int domain_num = 0, i;
5301 for_each_domain(cpu, sd)
5303 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5308 for_each_domain(cpu, sd) {
5309 snprintf(buf, 32, "domain%d", i);
5310 entry->procname = kstrdup(buf, GFP_KERNEL);
5312 entry->child = sd_alloc_ctl_domain_table(sd);
5319 static struct ctl_table_header *sd_sysctl_header;
5320 static void register_sched_domain_sysctl(void)
5322 int i, cpu_num = num_possible_cpus();
5323 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5326 WARN_ON(sd_ctl_dir[0].child);
5327 sd_ctl_dir[0].child = entry;
5332 for_each_possible_cpu(i) {
5333 snprintf(buf, 32, "cpu%d", i);
5334 entry->procname = kstrdup(buf, GFP_KERNEL);
5336 entry->child = sd_alloc_ctl_cpu_table(i);
5340 WARN_ON(sd_sysctl_header);
5341 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5344 /* may be called multiple times per register */
5345 static void unregister_sched_domain_sysctl(void)
5347 if (sd_sysctl_header)
5348 unregister_sysctl_table(sd_sysctl_header);
5349 sd_sysctl_header = NULL;
5350 if (sd_ctl_dir[0].child)
5351 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5354 static void register_sched_domain_sysctl(void)
5357 static void unregister_sched_domain_sysctl(void)
5362 static void set_rq_online(struct rq *rq)
5365 const struct sched_class *class;
5367 cpumask_set_cpu(rq->cpu, rq->rd->online);
5370 for_each_class(class) {
5371 if (class->rq_online)
5372 class->rq_online(rq);
5377 static void set_rq_offline(struct rq *rq)
5380 const struct sched_class *class;
5382 for_each_class(class) {
5383 if (class->rq_offline)
5384 class->rq_offline(rq);
5387 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5393 * migration_call - callback that gets triggered when a CPU is added.
5394 * Here we can start up the necessary migration thread for the new CPU.
5396 static int __cpuinit
5397 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5399 int cpu = (long)hcpu;
5400 unsigned long flags;
5401 struct rq *rq = cpu_rq(cpu);
5403 switch (action & ~CPU_TASKS_FROZEN) {
5405 case CPU_UP_PREPARE:
5406 rq->calc_load_update = calc_load_update;
5410 /* Update our root-domain */
5411 raw_spin_lock_irqsave(&rq->lock, flags);
5413 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5417 raw_spin_unlock_irqrestore(&rq->lock, flags);
5420 #ifdef CONFIG_HOTPLUG_CPU
5422 sched_ttwu_pending();
5423 /* Update our root-domain */
5424 raw_spin_lock_irqsave(&rq->lock, flags);
5426 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5430 BUG_ON(rq->nr_running != 1); /* the migration thread */
5431 raw_spin_unlock_irqrestore(&rq->lock, flags);
5433 migrate_nr_uninterruptible(rq);
5434 calc_global_load_remove(rq);
5439 update_max_interval();
5445 * Register at high priority so that task migration (migrate_all_tasks)
5446 * happens before everything else. This has to be lower priority than
5447 * the notifier in the perf_event subsystem, though.
5449 static struct notifier_block __cpuinitdata migration_notifier = {
5450 .notifier_call = migration_call,
5451 .priority = CPU_PRI_MIGRATION,
5454 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5455 unsigned long action, void *hcpu)
5457 switch (action & ~CPU_TASKS_FROZEN) {
5459 case CPU_DOWN_FAILED:
5460 set_cpu_active((long)hcpu, true);
5467 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5468 unsigned long action, void *hcpu)
5470 switch (action & ~CPU_TASKS_FROZEN) {
5471 case CPU_DOWN_PREPARE:
5472 set_cpu_active((long)hcpu, false);
5479 static int __init migration_init(void)
5481 void *cpu = (void *)(long)smp_processor_id();
5484 /* Initialize migration for the boot CPU */
5485 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5486 BUG_ON(err == NOTIFY_BAD);
5487 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5488 register_cpu_notifier(&migration_notifier);
5490 /* Register cpu active notifiers */
5491 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5492 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5496 early_initcall(migration_init);
5501 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5503 #ifdef CONFIG_SCHED_DEBUG
5505 static __read_mostly int sched_domain_debug_enabled;
5507 static int __init sched_domain_debug_setup(char *str)
5509 sched_domain_debug_enabled = 1;
5513 early_param("sched_debug", sched_domain_debug_setup);
5515 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5516 struct cpumask *groupmask)
5518 struct sched_group *group = sd->groups;
5521 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5522 cpumask_clear(groupmask);
5524 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5526 if (!(sd->flags & SD_LOAD_BALANCE)) {
5527 printk("does not load-balance\n");
5529 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5534 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5536 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5537 printk(KERN_ERR "ERROR: domain->span does not contain "
5540 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5541 printk(KERN_ERR "ERROR: domain->groups does not contain"
5545 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5549 printk(KERN_ERR "ERROR: group is NULL\n");
5553 if (!group->sgp->power) {
5554 printk(KERN_CONT "\n");
5555 printk(KERN_ERR "ERROR: domain->cpu_power not "
5560 if (!cpumask_weight(sched_group_cpus(group))) {
5561 printk(KERN_CONT "\n");
5562 printk(KERN_ERR "ERROR: empty group\n");
5566 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5567 printk(KERN_CONT "\n");
5568 printk(KERN_ERR "ERROR: repeated CPUs\n");
5572 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5574 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5576 printk(KERN_CONT " %s", str);
5577 if (group->sgp->power != SCHED_POWER_SCALE) {
5578 printk(KERN_CONT " (cpu_power = %d)",
5582 group = group->next;
5583 } while (group != sd->groups);
5584 printk(KERN_CONT "\n");
5586 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5587 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5590 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5591 printk(KERN_ERR "ERROR: parent span is not a superset "
5592 "of domain->span\n");
5596 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5600 if (!sched_domain_debug_enabled)
5604 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5608 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5611 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5619 #else /* !CONFIG_SCHED_DEBUG */
5620 # define sched_domain_debug(sd, cpu) do { } while (0)
5621 #endif /* CONFIG_SCHED_DEBUG */
5623 static int sd_degenerate(struct sched_domain *sd)
5625 if (cpumask_weight(sched_domain_span(sd)) == 1)
5628 /* Following flags need at least 2 groups */
5629 if (sd->flags & (SD_LOAD_BALANCE |
5630 SD_BALANCE_NEWIDLE |
5634 SD_SHARE_PKG_RESOURCES)) {
5635 if (sd->groups != sd->groups->next)
5639 /* Following flags don't use groups */
5640 if (sd->flags & (SD_WAKE_AFFINE))
5647 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5649 unsigned long cflags = sd->flags, pflags = parent->flags;
5651 if (sd_degenerate(parent))
5654 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5657 /* Flags needing groups don't count if only 1 group in parent */
5658 if (parent->groups == parent->groups->next) {
5659 pflags &= ~(SD_LOAD_BALANCE |
5660 SD_BALANCE_NEWIDLE |
5664 SD_SHARE_PKG_RESOURCES);
5665 if (nr_node_ids == 1)
5666 pflags &= ~SD_SERIALIZE;
5668 if (~cflags & pflags)
5674 static void free_rootdomain(struct rcu_head *rcu)
5676 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5678 cpupri_cleanup(&rd->cpupri);
5679 free_cpumask_var(rd->rto_mask);
5680 free_cpumask_var(rd->online);
5681 free_cpumask_var(rd->span);
5685 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5687 struct root_domain *old_rd = NULL;
5688 unsigned long flags;
5690 raw_spin_lock_irqsave(&rq->lock, flags);
5695 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5698 cpumask_clear_cpu(rq->cpu, old_rd->span);
5701 * If we dont want to free the old_rt yet then
5702 * set old_rd to NULL to skip the freeing later
5705 if (!atomic_dec_and_test(&old_rd->refcount))
5709 atomic_inc(&rd->refcount);
5712 cpumask_set_cpu(rq->cpu, rd->span);
5713 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5716 raw_spin_unlock_irqrestore(&rq->lock, flags);
5719 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5722 static int init_rootdomain(struct root_domain *rd)
5724 memset(rd, 0, sizeof(*rd));
5726 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5728 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5730 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5733 if (cpupri_init(&rd->cpupri) != 0)
5738 free_cpumask_var(rd->rto_mask);
5740 free_cpumask_var(rd->online);
5742 free_cpumask_var(rd->span);
5748 * By default the system creates a single root-domain with all cpus as
5749 * members (mimicking the global state we have today).
5751 struct root_domain def_root_domain;
5753 static void init_defrootdomain(void)
5755 init_rootdomain(&def_root_domain);
5757 atomic_set(&def_root_domain.refcount, 1);
5760 static struct root_domain *alloc_rootdomain(void)
5762 struct root_domain *rd;
5764 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5768 if (init_rootdomain(rd) != 0) {
5776 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5778 struct sched_group *tmp, *first;
5787 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5792 } while (sg != first);
5795 static void free_sched_domain(struct rcu_head *rcu)
5797 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5800 * If its an overlapping domain it has private groups, iterate and
5803 if (sd->flags & SD_OVERLAP) {
5804 free_sched_groups(sd->groups, 1);
5805 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5806 kfree(sd->groups->sgp);
5812 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5814 call_rcu(&sd->rcu, free_sched_domain);
5817 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5819 for (; sd; sd = sd->parent)
5820 destroy_sched_domain(sd, cpu);
5824 * Keep a special pointer to the highest sched_domain that has
5825 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5826 * allows us to avoid some pointer chasing select_idle_sibling().
5828 * Also keep a unique ID per domain (we use the first cpu number in
5829 * the cpumask of the domain), this allows us to quickly tell if
5830 * two cpus are in the same cache domain, see cpus_share_cache().
5832 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5833 DEFINE_PER_CPU(int, sd_llc_id);
5835 static void update_top_cache_domain(int cpu)
5837 struct sched_domain *sd;
5840 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5842 id = cpumask_first(sched_domain_span(sd));
5844 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5845 per_cpu(sd_llc_id, cpu) = id;
5849 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5850 * hold the hotplug lock.
5853 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5855 struct rq *rq = cpu_rq(cpu);
5856 struct sched_domain *tmp;
5858 /* Remove the sched domains which do not contribute to scheduling. */
5859 for (tmp = sd; tmp; ) {
5860 struct sched_domain *parent = tmp->parent;
5864 if (sd_parent_degenerate(tmp, parent)) {
5865 tmp->parent = parent->parent;
5867 parent->parent->child = tmp;
5868 destroy_sched_domain(parent, cpu);
5873 if (sd && sd_degenerate(sd)) {
5876 destroy_sched_domain(tmp, cpu);
5881 sched_domain_debug(sd, cpu);
5883 rq_attach_root(rq, rd);
5885 rcu_assign_pointer(rq->sd, sd);
5886 destroy_sched_domains(tmp, cpu);
5888 update_top_cache_domain(cpu);
5891 /* cpus with isolated domains */
5892 static cpumask_var_t cpu_isolated_map;
5894 /* Setup the mask of cpus configured for isolated domains */
5895 static int __init isolated_cpu_setup(char *str)
5897 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5898 cpulist_parse(str, cpu_isolated_map);
5902 __setup("isolcpus=", isolated_cpu_setup);
5907 * find_next_best_node - find the next node to include in a sched_domain
5908 * @node: node whose sched_domain we're building
5909 * @used_nodes: nodes already in the sched_domain
5911 * Find the next node to include in a given scheduling domain. Simply
5912 * finds the closest node not already in the @used_nodes map.
5914 * Should use nodemask_t.
5916 static int find_next_best_node(int node, nodemask_t *used_nodes)
5918 int i, n, val, min_val, best_node = -1;
5922 for (i = 0; i < nr_node_ids; i++) {
5923 /* Start at @node */
5924 n = (node + i) % nr_node_ids;
5926 if (!nr_cpus_node(n))
5929 /* Skip already used nodes */
5930 if (node_isset(n, *used_nodes))
5933 /* Simple min distance search */
5934 val = node_distance(node, n);
5936 if (val < min_val) {
5942 if (best_node != -1)
5943 node_set(best_node, *used_nodes);
5948 * sched_domain_node_span - get a cpumask for a node's sched_domain
5949 * @node: node whose cpumask we're constructing
5950 * @span: resulting cpumask
5952 * Given a node, construct a good cpumask for its sched_domain to span. It
5953 * should be one that prevents unnecessary balancing, but also spreads tasks
5956 static void sched_domain_node_span(int node, struct cpumask *span)
5958 nodemask_t used_nodes;
5961 cpumask_clear(span);
5962 nodes_clear(used_nodes);
5964 cpumask_or(span, span, cpumask_of_node(node));
5965 node_set(node, used_nodes);
5967 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5968 int next_node = find_next_best_node(node, &used_nodes);
5971 cpumask_or(span, span, cpumask_of_node(next_node));
5975 static const struct cpumask *cpu_node_mask(int cpu)
5977 lockdep_assert_held(&sched_domains_mutex);
5979 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5981 return sched_domains_tmpmask;
5984 static const struct cpumask *cpu_allnodes_mask(int cpu)
5986 return cpu_possible_mask;
5988 #endif /* CONFIG_NUMA */
5990 static const struct cpumask *cpu_cpu_mask(int cpu)
5992 return cpumask_of_node(cpu_to_node(cpu));
5995 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5998 struct sched_domain **__percpu sd;
5999 struct sched_group **__percpu sg;
6000 struct sched_group_power **__percpu sgp;
6004 struct sched_domain ** __percpu sd;
6005 struct root_domain *rd;
6015 struct sched_domain_topology_level;
6017 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6018 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6020 #define SDTL_OVERLAP 0x01
6022 struct sched_domain_topology_level {
6023 sched_domain_init_f init;
6024 sched_domain_mask_f mask;
6026 struct sd_data data;
6030 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6032 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6033 const struct cpumask *span = sched_domain_span(sd);
6034 struct cpumask *covered = sched_domains_tmpmask;
6035 struct sd_data *sdd = sd->private;
6036 struct sched_domain *child;
6039 cpumask_clear(covered);
6041 for_each_cpu(i, span) {
6042 struct cpumask *sg_span;
6044 if (cpumask_test_cpu(i, covered))
6047 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6048 GFP_KERNEL, cpu_to_node(cpu));
6053 sg_span = sched_group_cpus(sg);
6055 child = *per_cpu_ptr(sdd->sd, i);
6057 child = child->child;
6058 cpumask_copy(sg_span, sched_domain_span(child));
6060 cpumask_set_cpu(i, sg_span);
6062 cpumask_or(covered, covered, sg_span);
6064 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
6065 atomic_inc(&sg->sgp->ref);
6067 if (cpumask_test_cpu(cpu, sg_span))
6077 sd->groups = groups;
6082 free_sched_groups(first, 0);
6087 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6089 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6090 struct sched_domain *child = sd->child;
6093 cpu = cpumask_first(sched_domain_span(child));
6096 *sg = *per_cpu_ptr(sdd->sg, cpu);
6097 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6098 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6105 * build_sched_groups will build a circular linked list of the groups
6106 * covered by the given span, and will set each group's ->cpumask correctly,
6107 * and ->cpu_power to 0.
6109 * Assumes the sched_domain tree is fully constructed
6112 build_sched_groups(struct sched_domain *sd, int cpu)
6114 struct sched_group *first = NULL, *last = NULL;
6115 struct sd_data *sdd = sd->private;
6116 const struct cpumask *span = sched_domain_span(sd);
6117 struct cpumask *covered;
6120 get_group(cpu, sdd, &sd->groups);
6121 atomic_inc(&sd->groups->ref);
6123 if (cpu != cpumask_first(sched_domain_span(sd)))
6126 lockdep_assert_held(&sched_domains_mutex);
6127 covered = sched_domains_tmpmask;
6129 cpumask_clear(covered);
6131 for_each_cpu(i, span) {
6132 struct sched_group *sg;
6133 int group = get_group(i, sdd, &sg);
6136 if (cpumask_test_cpu(i, covered))
6139 cpumask_clear(sched_group_cpus(sg));
6142 for_each_cpu(j, span) {
6143 if (get_group(j, sdd, NULL) != group)
6146 cpumask_set_cpu(j, covered);
6147 cpumask_set_cpu(j, sched_group_cpus(sg));
6162 * Initialize sched groups cpu_power.
6164 * cpu_power indicates the capacity of sched group, which is used while
6165 * distributing the load between different sched groups in a sched domain.
6166 * Typically cpu_power for all the groups in a sched domain will be same unless
6167 * there are asymmetries in the topology. If there are asymmetries, group
6168 * having more cpu_power will pickup more load compared to the group having
6171 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6173 struct sched_group *sg = sd->groups;
6175 WARN_ON(!sd || !sg);
6178 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6180 } while (sg != sd->groups);
6182 if (cpu != group_first_cpu(sg))
6185 update_group_power(sd, cpu);
6186 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6189 int __weak arch_sd_sibling_asym_packing(void)
6191 return 0*SD_ASYM_PACKING;
6195 * Initializers for schedule domains
6196 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6199 #ifdef CONFIG_SCHED_DEBUG
6200 # define SD_INIT_NAME(sd, type) sd->name = #type
6202 # define SD_INIT_NAME(sd, type) do { } while (0)
6205 #define SD_INIT_FUNC(type) \
6206 static noinline struct sched_domain * \
6207 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6209 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6210 *sd = SD_##type##_INIT; \
6211 SD_INIT_NAME(sd, type); \
6212 sd->private = &tl->data; \
6218 SD_INIT_FUNC(ALLNODES)
6221 #ifdef CONFIG_SCHED_SMT
6222 SD_INIT_FUNC(SIBLING)
6224 #ifdef CONFIG_SCHED_MC
6227 #ifdef CONFIG_SCHED_BOOK
6231 static int default_relax_domain_level = -1;
6232 int sched_domain_level_max;
6234 static int __init setup_relax_domain_level(char *str)
6238 val = simple_strtoul(str, NULL, 0);
6239 if (val < sched_domain_level_max)
6240 default_relax_domain_level = val;
6244 __setup("relax_domain_level=", setup_relax_domain_level);
6246 static void set_domain_attribute(struct sched_domain *sd,
6247 struct sched_domain_attr *attr)
6251 if (!attr || attr->relax_domain_level < 0) {
6252 if (default_relax_domain_level < 0)
6255 request = default_relax_domain_level;
6257 request = attr->relax_domain_level;
6258 if (request < sd->level) {
6259 /* turn off idle balance on this domain */
6260 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6262 /* turn on idle balance on this domain */
6263 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6267 static void __sdt_free(const struct cpumask *cpu_map);
6268 static int __sdt_alloc(const struct cpumask *cpu_map);
6270 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6271 const struct cpumask *cpu_map)
6275 if (!atomic_read(&d->rd->refcount))
6276 free_rootdomain(&d->rd->rcu); /* fall through */
6278 free_percpu(d->sd); /* fall through */
6280 __sdt_free(cpu_map); /* fall through */
6286 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6287 const struct cpumask *cpu_map)
6289 memset(d, 0, sizeof(*d));
6291 if (__sdt_alloc(cpu_map))
6292 return sa_sd_storage;
6293 d->sd = alloc_percpu(struct sched_domain *);
6295 return sa_sd_storage;
6296 d->rd = alloc_rootdomain();
6299 return sa_rootdomain;
6303 * NULL the sd_data elements we've used to build the sched_domain and
6304 * sched_group structure so that the subsequent __free_domain_allocs()
6305 * will not free the data we're using.
6307 static void claim_allocations(int cpu, struct sched_domain *sd)
6309 struct sd_data *sdd = sd->private;
6311 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6312 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6314 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6315 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6317 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6318 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6321 #ifdef CONFIG_SCHED_SMT
6322 static const struct cpumask *cpu_smt_mask(int cpu)
6324 return topology_thread_cpumask(cpu);
6329 * Topology list, bottom-up.
6331 static struct sched_domain_topology_level default_topology[] = {
6332 #ifdef CONFIG_SCHED_SMT
6333 { sd_init_SIBLING, cpu_smt_mask, },
6335 #ifdef CONFIG_SCHED_MC
6336 { sd_init_MC, cpu_coregroup_mask, },
6338 #ifdef CONFIG_SCHED_BOOK
6339 { sd_init_BOOK, cpu_book_mask, },
6341 { sd_init_CPU, cpu_cpu_mask, },
6343 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6344 { sd_init_ALLNODES, cpu_allnodes_mask, },
6349 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6351 static int __sdt_alloc(const struct cpumask *cpu_map)
6353 struct sched_domain_topology_level *tl;
6356 for (tl = sched_domain_topology; tl->init; tl++) {
6357 struct sd_data *sdd = &tl->data;
6359 sdd->sd = alloc_percpu(struct sched_domain *);
6363 sdd->sg = alloc_percpu(struct sched_group *);
6367 sdd->sgp = alloc_percpu(struct sched_group_power *);
6371 for_each_cpu(j, cpu_map) {
6372 struct sched_domain *sd;
6373 struct sched_group *sg;
6374 struct sched_group_power *sgp;
6376 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6377 GFP_KERNEL, cpu_to_node(j));
6381 *per_cpu_ptr(sdd->sd, j) = sd;
6383 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6384 GFP_KERNEL, cpu_to_node(j));
6388 *per_cpu_ptr(sdd->sg, j) = sg;
6390 sgp = kzalloc_node(sizeof(struct sched_group_power),
6391 GFP_KERNEL, cpu_to_node(j));
6395 *per_cpu_ptr(sdd->sgp, j) = sgp;
6402 static void __sdt_free(const struct cpumask *cpu_map)
6404 struct sched_domain_topology_level *tl;
6407 for (tl = sched_domain_topology; tl->init; tl++) {
6408 struct sd_data *sdd = &tl->data;
6410 for_each_cpu(j, cpu_map) {
6411 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6412 if (sd && (sd->flags & SD_OVERLAP))
6413 free_sched_groups(sd->groups, 0);
6414 kfree(*per_cpu_ptr(sdd->sd, j));
6415 kfree(*per_cpu_ptr(sdd->sg, j));
6416 kfree(*per_cpu_ptr(sdd->sgp, j));
6418 free_percpu(sdd->sd);
6419 free_percpu(sdd->sg);
6420 free_percpu(sdd->sgp);
6424 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6425 struct s_data *d, const struct cpumask *cpu_map,
6426 struct sched_domain_attr *attr, struct sched_domain *child,
6429 struct sched_domain *sd = tl->init(tl, cpu);
6433 set_domain_attribute(sd, attr);
6434 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6436 sd->level = child->level + 1;
6437 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6446 * Build sched domains for a given set of cpus and attach the sched domains
6447 * to the individual cpus
6449 static int build_sched_domains(const struct cpumask *cpu_map,
6450 struct sched_domain_attr *attr)
6452 enum s_alloc alloc_state = sa_none;
6453 struct sched_domain *sd;
6455 int i, ret = -ENOMEM;
6457 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6458 if (alloc_state != sa_rootdomain)
6461 /* Set up domains for cpus specified by the cpu_map. */
6462 for_each_cpu(i, cpu_map) {
6463 struct sched_domain_topology_level *tl;
6466 for (tl = sched_domain_topology; tl->init; tl++) {
6467 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6468 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6469 sd->flags |= SD_OVERLAP;
6470 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6477 *per_cpu_ptr(d.sd, i) = sd;
6480 /* Build the groups for the domains */
6481 for_each_cpu(i, cpu_map) {
6482 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6483 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6484 if (sd->flags & SD_OVERLAP) {
6485 if (build_overlap_sched_groups(sd, i))
6488 if (build_sched_groups(sd, i))
6494 /* Calculate CPU power for physical packages and nodes */
6495 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6496 if (!cpumask_test_cpu(i, cpu_map))
6499 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6500 claim_allocations(i, sd);
6501 init_sched_groups_power(i, sd);
6505 /* Attach the domains */
6507 for_each_cpu(i, cpu_map) {
6508 sd = *per_cpu_ptr(d.sd, i);
6509 cpu_attach_domain(sd, d.rd, i);
6515 __free_domain_allocs(&d, alloc_state, cpu_map);
6519 static cpumask_var_t *doms_cur; /* current sched domains */
6520 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6521 static struct sched_domain_attr *dattr_cur;
6522 /* attribues of custom domains in 'doms_cur' */
6525 * Special case: If a kmalloc of a doms_cur partition (array of
6526 * cpumask) fails, then fallback to a single sched domain,
6527 * as determined by the single cpumask fallback_doms.
6529 static cpumask_var_t fallback_doms;
6532 * arch_update_cpu_topology lets virtualized architectures update the
6533 * cpu core maps. It is supposed to return 1 if the topology changed
6534 * or 0 if it stayed the same.
6536 int __attribute__((weak)) arch_update_cpu_topology(void)
6541 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6544 cpumask_var_t *doms;
6546 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6549 for (i = 0; i < ndoms; i++) {
6550 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6551 free_sched_domains(doms, i);
6558 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6561 for (i = 0; i < ndoms; i++)
6562 free_cpumask_var(doms[i]);
6567 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6568 * For now this just excludes isolated cpus, but could be used to
6569 * exclude other special cases in the future.
6571 static int init_sched_domains(const struct cpumask *cpu_map)
6575 arch_update_cpu_topology();
6577 doms_cur = alloc_sched_domains(ndoms_cur);
6579 doms_cur = &fallback_doms;
6580 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6582 err = build_sched_domains(doms_cur[0], NULL);
6583 register_sched_domain_sysctl();
6589 * Detach sched domains from a group of cpus specified in cpu_map
6590 * These cpus will now be attached to the NULL domain
6592 static void detach_destroy_domains(const struct cpumask *cpu_map)
6597 for_each_cpu(i, cpu_map)
6598 cpu_attach_domain(NULL, &def_root_domain, i);
6602 /* handle null as "default" */
6603 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6604 struct sched_domain_attr *new, int idx_new)
6606 struct sched_domain_attr tmp;
6613 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6614 new ? (new + idx_new) : &tmp,
6615 sizeof(struct sched_domain_attr));
6619 * Partition sched domains as specified by the 'ndoms_new'
6620 * cpumasks in the array doms_new[] of cpumasks. This compares
6621 * doms_new[] to the current sched domain partitioning, doms_cur[].
6622 * It destroys each deleted domain and builds each new domain.
6624 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6625 * The masks don't intersect (don't overlap.) We should setup one
6626 * sched domain for each mask. CPUs not in any of the cpumasks will
6627 * not be load balanced. If the same cpumask appears both in the
6628 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6631 * The passed in 'doms_new' should be allocated using
6632 * alloc_sched_domains. This routine takes ownership of it and will
6633 * free_sched_domains it when done with it. If the caller failed the
6634 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6635 * and partition_sched_domains() will fallback to the single partition
6636 * 'fallback_doms', it also forces the domains to be rebuilt.
6638 * If doms_new == NULL it will be replaced with cpu_online_mask.
6639 * ndoms_new == 0 is a special case for destroying existing domains,
6640 * and it will not create the default domain.
6642 * Call with hotplug lock held
6644 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6645 struct sched_domain_attr *dattr_new)
6650 mutex_lock(&sched_domains_mutex);
6652 /* always unregister in case we don't destroy any domains */
6653 unregister_sched_domain_sysctl();
6655 /* Let architecture update cpu core mappings. */
6656 new_topology = arch_update_cpu_topology();
6658 n = doms_new ? ndoms_new : 0;
6660 /* Destroy deleted domains */
6661 for (i = 0; i < ndoms_cur; i++) {
6662 for (j = 0; j < n && !new_topology; j++) {
6663 if (cpumask_equal(doms_cur[i], doms_new[j])
6664 && dattrs_equal(dattr_cur, i, dattr_new, j))
6667 /* no match - a current sched domain not in new doms_new[] */
6668 detach_destroy_domains(doms_cur[i]);
6673 if (doms_new == NULL) {
6675 doms_new = &fallback_doms;
6676 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6677 WARN_ON_ONCE(dattr_new);
6680 /* Build new domains */
6681 for (i = 0; i < ndoms_new; i++) {
6682 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6683 if (cpumask_equal(doms_new[i], doms_cur[j])
6684 && dattrs_equal(dattr_new, i, dattr_cur, j))
6687 /* no match - add a new doms_new */
6688 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6693 /* Remember the new sched domains */
6694 if (doms_cur != &fallback_doms)
6695 free_sched_domains(doms_cur, ndoms_cur);
6696 kfree(dattr_cur); /* kfree(NULL) is safe */
6697 doms_cur = doms_new;
6698 dattr_cur = dattr_new;
6699 ndoms_cur = ndoms_new;
6701 register_sched_domain_sysctl();
6703 mutex_unlock(&sched_domains_mutex);
6706 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6707 static void reinit_sched_domains(void)
6711 /* Destroy domains first to force the rebuild */
6712 partition_sched_domains(0, NULL, NULL);
6714 rebuild_sched_domains();
6718 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6720 unsigned int level = 0;
6722 if (sscanf(buf, "%u", &level) != 1)
6726 * level is always be positive so don't check for
6727 * level < POWERSAVINGS_BALANCE_NONE which is 0
6728 * What happens on 0 or 1 byte write,
6729 * need to check for count as well?
6732 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6736 sched_smt_power_savings = level;
6738 sched_mc_power_savings = level;
6740 reinit_sched_domains();
6745 #ifdef CONFIG_SCHED_MC
6746 static ssize_t sched_mc_power_savings_show(struct device *dev,
6747 struct device_attribute *attr,
6750 return sprintf(buf, "%u\n", sched_mc_power_savings);
6752 static ssize_t sched_mc_power_savings_store(struct device *dev,
6753 struct device_attribute *attr,
6754 const char *buf, size_t count)
6756 return sched_power_savings_store(buf, count, 0);
6758 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6759 sched_mc_power_savings_show,
6760 sched_mc_power_savings_store);
6763 #ifdef CONFIG_SCHED_SMT
6764 static ssize_t sched_smt_power_savings_show(struct device *dev,
6765 struct device_attribute *attr,
6768 return sprintf(buf, "%u\n", sched_smt_power_savings);
6770 static ssize_t sched_smt_power_savings_store(struct device *dev,
6771 struct device_attribute *attr,
6772 const char *buf, size_t count)
6774 return sched_power_savings_store(buf, count, 1);
6776 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6777 sched_smt_power_savings_show,
6778 sched_smt_power_savings_store);
6781 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6785 #ifdef CONFIG_SCHED_SMT
6787 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6789 #ifdef CONFIG_SCHED_MC
6790 if (!err && mc_capable())
6791 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6795 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6798 * Update cpusets according to cpu_active mask. If cpusets are
6799 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6800 * around partition_sched_domains().
6802 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6805 switch (action & ~CPU_TASKS_FROZEN) {
6807 case CPU_DOWN_FAILED:
6808 cpuset_update_active_cpus();
6815 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6818 switch (action & ~CPU_TASKS_FROZEN) {
6819 case CPU_DOWN_PREPARE:
6820 cpuset_update_active_cpus();
6827 void __init sched_init_smp(void)
6829 cpumask_var_t non_isolated_cpus;
6831 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6832 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6835 mutex_lock(&sched_domains_mutex);
6836 init_sched_domains(cpu_active_mask);
6837 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6838 if (cpumask_empty(non_isolated_cpus))
6839 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6840 mutex_unlock(&sched_domains_mutex);
6843 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6844 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6846 /* RT runtime code needs to handle some hotplug events */
6847 hotcpu_notifier(update_runtime, 0);
6851 /* Move init over to a non-isolated CPU */
6852 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6854 sched_init_granularity();
6855 free_cpumask_var(non_isolated_cpus);
6857 init_sched_rt_class();
6860 void __init sched_init_smp(void)
6862 sched_init_granularity();
6864 #endif /* CONFIG_SMP */
6866 const_debug unsigned int sysctl_timer_migration = 1;
6868 int in_sched_functions(unsigned long addr)
6870 return in_lock_functions(addr) ||
6871 (addr >= (unsigned long)__sched_text_start
6872 && addr < (unsigned long)__sched_text_end);
6875 #ifdef CONFIG_CGROUP_SCHED
6876 struct task_group root_task_group;
6879 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6881 void __init sched_init(void)
6884 unsigned long alloc_size = 0, ptr;
6886 #ifdef CONFIG_FAIR_GROUP_SCHED
6887 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6889 #ifdef CONFIG_RT_GROUP_SCHED
6890 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6892 #ifdef CONFIG_CPUMASK_OFFSTACK
6893 alloc_size += num_possible_cpus() * cpumask_size();
6896 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6898 #ifdef CONFIG_FAIR_GROUP_SCHED
6899 root_task_group.se = (struct sched_entity **)ptr;
6900 ptr += nr_cpu_ids * sizeof(void **);
6902 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6903 ptr += nr_cpu_ids * sizeof(void **);
6905 #endif /* CONFIG_FAIR_GROUP_SCHED */
6906 #ifdef CONFIG_RT_GROUP_SCHED
6907 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6908 ptr += nr_cpu_ids * sizeof(void **);
6910 root_task_group.rt_rq = (struct rt_rq **)ptr;
6911 ptr += nr_cpu_ids * sizeof(void **);
6913 #endif /* CONFIG_RT_GROUP_SCHED */
6914 #ifdef CONFIG_CPUMASK_OFFSTACK
6915 for_each_possible_cpu(i) {
6916 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6917 ptr += cpumask_size();
6919 #endif /* CONFIG_CPUMASK_OFFSTACK */
6923 init_defrootdomain();
6926 init_rt_bandwidth(&def_rt_bandwidth,
6927 global_rt_period(), global_rt_runtime());
6929 #ifdef CONFIG_RT_GROUP_SCHED
6930 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6931 global_rt_period(), global_rt_runtime());
6932 #endif /* CONFIG_RT_GROUP_SCHED */
6934 #ifdef CONFIG_CGROUP_SCHED
6935 list_add(&root_task_group.list, &task_groups);
6936 INIT_LIST_HEAD(&root_task_group.children);
6937 INIT_LIST_HEAD(&root_task_group.siblings);
6938 autogroup_init(&init_task);
6940 #endif /* CONFIG_CGROUP_SCHED */
6942 #ifdef CONFIG_CGROUP_CPUACCT
6943 root_cpuacct.cpustat = &kernel_cpustat;
6944 root_cpuacct.cpuusage = alloc_percpu(u64);
6945 /* Too early, not expected to fail */
6946 BUG_ON(!root_cpuacct.cpuusage);
6948 for_each_possible_cpu(i) {
6952 raw_spin_lock_init(&rq->lock);
6954 rq->calc_load_active = 0;
6955 rq->calc_load_update = jiffies + LOAD_FREQ;
6956 init_cfs_rq(&rq->cfs);
6957 init_rt_rq(&rq->rt, rq);
6958 #ifdef CONFIG_FAIR_GROUP_SCHED
6959 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6960 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6962 * How much cpu bandwidth does root_task_group get?
6964 * In case of task-groups formed thr' the cgroup filesystem, it
6965 * gets 100% of the cpu resources in the system. This overall
6966 * system cpu resource is divided among the tasks of
6967 * root_task_group and its child task-groups in a fair manner,
6968 * based on each entity's (task or task-group's) weight
6969 * (se->load.weight).
6971 * In other words, if root_task_group has 10 tasks of weight
6972 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6973 * then A0's share of the cpu resource is:
6975 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6977 * We achieve this by letting root_task_group's tasks sit
6978 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6980 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6981 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6982 #endif /* CONFIG_FAIR_GROUP_SCHED */
6984 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6985 #ifdef CONFIG_RT_GROUP_SCHED
6986 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6987 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6990 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6991 rq->cpu_load[j] = 0;
6993 rq->last_load_update_tick = jiffies;
6998 rq->cpu_power = SCHED_POWER_SCALE;
6999 rq->post_schedule = 0;
7000 rq->active_balance = 0;
7001 rq->next_balance = jiffies;
7006 rq->avg_idle = 2*sysctl_sched_migration_cost;
7008 INIT_LIST_HEAD(&rq->cfs_tasks);
7010 rq_attach_root(rq, &def_root_domain);
7016 atomic_set(&rq->nr_iowait, 0);
7019 set_load_weight(&init_task);
7021 #ifdef CONFIG_PREEMPT_NOTIFIERS
7022 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7025 #ifdef CONFIG_RT_MUTEXES
7026 plist_head_init(&init_task.pi_waiters);
7030 * The boot idle thread does lazy MMU switching as well:
7032 atomic_inc(&init_mm.mm_count);
7033 enter_lazy_tlb(&init_mm, current);
7036 * Make us the idle thread. Technically, schedule() should not be
7037 * called from this thread, however somewhere below it might be,
7038 * but because we are the idle thread, we just pick up running again
7039 * when this runqueue becomes "idle".
7041 init_idle(current, smp_processor_id());
7043 calc_load_update = jiffies + LOAD_FREQ;
7046 * During early bootup we pretend to be a normal task:
7048 current->sched_class = &fair_sched_class;
7051 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7052 /* May be allocated at isolcpus cmdline parse time */
7053 if (cpu_isolated_map == NULL)
7054 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7056 init_sched_fair_class();
7058 scheduler_running = 1;
7061 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7062 static inline int preempt_count_equals(int preempt_offset)
7064 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7066 return (nested == preempt_offset);
7069 void __might_sleep(const char *file, int line, int preempt_offset)
7071 static unsigned long prev_jiffy; /* ratelimiting */
7073 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7074 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7075 system_state != SYSTEM_RUNNING || oops_in_progress)
7077 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7079 prev_jiffy = jiffies;
7082 "BUG: sleeping function called from invalid context at %s:%d\n",
7085 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7086 in_atomic(), irqs_disabled(),
7087 current->pid, current->comm);
7089 debug_show_held_locks(current);
7090 if (irqs_disabled())
7091 print_irqtrace_events(current);
7094 EXPORT_SYMBOL(__might_sleep);
7097 #ifdef CONFIG_MAGIC_SYSRQ
7098 static void normalize_task(struct rq *rq, struct task_struct *p)
7100 const struct sched_class *prev_class = p->sched_class;
7101 int old_prio = p->prio;
7106 dequeue_task(rq, p, 0);
7107 __setscheduler(rq, p, SCHED_NORMAL, 0);
7109 enqueue_task(rq, p, 0);
7110 resched_task(rq->curr);
7113 check_class_changed(rq, p, prev_class, old_prio);
7116 void normalize_rt_tasks(void)
7118 struct task_struct *g, *p;
7119 unsigned long flags;
7122 read_lock_irqsave(&tasklist_lock, flags);
7123 do_each_thread(g, p) {
7125 * Only normalize user tasks:
7130 p->se.exec_start = 0;
7131 #ifdef CONFIG_SCHEDSTATS
7132 p->se.statistics.wait_start = 0;
7133 p->se.statistics.sleep_start = 0;
7134 p->se.statistics.block_start = 0;
7139 * Renice negative nice level userspace
7142 if (TASK_NICE(p) < 0 && p->mm)
7143 set_user_nice(p, 0);
7147 raw_spin_lock(&p->pi_lock);
7148 rq = __task_rq_lock(p);
7150 normalize_task(rq, p);
7152 __task_rq_unlock(rq);
7153 raw_spin_unlock(&p->pi_lock);
7154 } while_each_thread(g, p);
7156 read_unlock_irqrestore(&tasklist_lock, flags);
7159 #endif /* CONFIG_MAGIC_SYSRQ */
7161 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7163 * These functions are only useful for the IA64 MCA handling, or kdb.
7165 * They can only be called when the whole system has been
7166 * stopped - every CPU needs to be quiescent, and no scheduling
7167 * activity can take place. Using them for anything else would
7168 * be a serious bug, and as a result, they aren't even visible
7169 * under any other configuration.
7173 * curr_task - return the current task for a given cpu.
7174 * @cpu: the processor in question.
7176 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7178 struct task_struct *curr_task(int cpu)
7180 return cpu_curr(cpu);
7183 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7187 * set_curr_task - set the current task for a given cpu.
7188 * @cpu: the processor in question.
7189 * @p: the task pointer to set.
7191 * Description: This function must only be used when non-maskable interrupts
7192 * are serviced on a separate stack. It allows the architecture to switch the
7193 * notion of the current task on a cpu in a non-blocking manner. This function
7194 * must be called with all CPU's synchronized, and interrupts disabled, the
7195 * and caller must save the original value of the current task (see
7196 * curr_task() above) and restore that value before reenabling interrupts and
7197 * re-starting the system.
7199 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7201 void set_curr_task(int cpu, struct task_struct *p)
7208 #ifdef CONFIG_CGROUP_SCHED
7209 /* task_group_lock serializes the addition/removal of task groups */
7210 static DEFINE_SPINLOCK(task_group_lock);
7212 static void free_sched_group(struct task_group *tg)
7214 free_fair_sched_group(tg);
7215 free_rt_sched_group(tg);
7220 /* allocate runqueue etc for a new task group */
7221 struct task_group *sched_create_group(struct task_group *parent)
7223 struct task_group *tg;
7224 unsigned long flags;
7226 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7228 return ERR_PTR(-ENOMEM);
7230 if (!alloc_fair_sched_group(tg, parent))
7233 if (!alloc_rt_sched_group(tg, parent))
7236 spin_lock_irqsave(&task_group_lock, flags);
7237 list_add_rcu(&tg->list, &task_groups);
7239 WARN_ON(!parent); /* root should already exist */
7241 tg->parent = parent;
7242 INIT_LIST_HEAD(&tg->children);
7243 list_add_rcu(&tg->siblings, &parent->children);
7244 spin_unlock_irqrestore(&task_group_lock, flags);
7249 free_sched_group(tg);
7250 return ERR_PTR(-ENOMEM);
7253 /* rcu callback to free various structures associated with a task group */
7254 static void free_sched_group_rcu(struct rcu_head *rhp)
7256 /* now it should be safe to free those cfs_rqs */
7257 free_sched_group(container_of(rhp, struct task_group, rcu));
7260 /* Destroy runqueue etc associated with a task group */
7261 void sched_destroy_group(struct task_group *tg)
7263 unsigned long flags;
7266 /* end participation in shares distribution */
7267 for_each_possible_cpu(i)
7268 unregister_fair_sched_group(tg, i);
7270 spin_lock_irqsave(&task_group_lock, flags);
7271 list_del_rcu(&tg->list);
7272 list_del_rcu(&tg->siblings);
7273 spin_unlock_irqrestore(&task_group_lock, flags);
7275 /* wait for possible concurrent references to cfs_rqs complete */
7276 call_rcu(&tg->rcu, free_sched_group_rcu);
7279 /* change task's runqueue when it moves between groups.
7280 * The caller of this function should have put the task in its new group
7281 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7282 * reflect its new group.
7284 void sched_move_task(struct task_struct *tsk)
7287 unsigned long flags;
7290 rq = task_rq_lock(tsk, &flags);
7292 running = task_current(rq, tsk);
7296 dequeue_task(rq, tsk, 0);
7297 if (unlikely(running))
7298 tsk->sched_class->put_prev_task(rq, tsk);
7300 #ifdef CONFIG_FAIR_GROUP_SCHED
7301 if (tsk->sched_class->task_move_group)
7302 tsk->sched_class->task_move_group(tsk, on_rq);
7305 set_task_rq(tsk, task_cpu(tsk));
7307 if (unlikely(running))
7308 tsk->sched_class->set_curr_task(rq);
7310 enqueue_task(rq, tsk, 0);
7312 task_rq_unlock(rq, tsk, &flags);
7314 #endif /* CONFIG_CGROUP_SCHED */
7316 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7317 static unsigned long to_ratio(u64 period, u64 runtime)
7319 if (runtime == RUNTIME_INF)
7322 return div64_u64(runtime << 20, period);
7326 #ifdef CONFIG_RT_GROUP_SCHED
7328 * Ensure that the real time constraints are schedulable.
7330 static DEFINE_MUTEX(rt_constraints_mutex);
7332 /* Must be called with tasklist_lock held */
7333 static inline int tg_has_rt_tasks(struct task_group *tg)
7335 struct task_struct *g, *p;
7337 do_each_thread(g, p) {
7338 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7340 } while_each_thread(g, p);
7345 struct rt_schedulable_data {
7346 struct task_group *tg;
7351 static int tg_rt_schedulable(struct task_group *tg, void *data)
7353 struct rt_schedulable_data *d = data;
7354 struct task_group *child;
7355 unsigned long total, sum = 0;
7356 u64 period, runtime;
7358 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7359 runtime = tg->rt_bandwidth.rt_runtime;
7362 period = d->rt_period;
7363 runtime = d->rt_runtime;
7367 * Cannot have more runtime than the period.
7369 if (runtime > period && runtime != RUNTIME_INF)
7373 * Ensure we don't starve existing RT tasks.
7375 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7378 total = to_ratio(period, runtime);
7381 * Nobody can have more than the global setting allows.
7383 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7387 * The sum of our children's runtime should not exceed our own.
7389 list_for_each_entry_rcu(child, &tg->children, siblings) {
7390 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7391 runtime = child->rt_bandwidth.rt_runtime;
7393 if (child == d->tg) {
7394 period = d->rt_period;
7395 runtime = d->rt_runtime;
7398 sum += to_ratio(period, runtime);
7407 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7411 struct rt_schedulable_data data = {
7413 .rt_period = period,
7414 .rt_runtime = runtime,
7418 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7424 static int tg_set_rt_bandwidth(struct task_group *tg,
7425 u64 rt_period, u64 rt_runtime)
7429 mutex_lock(&rt_constraints_mutex);
7430 read_lock(&tasklist_lock);
7431 err = __rt_schedulable(tg, rt_period, rt_runtime);
7435 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7436 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7437 tg->rt_bandwidth.rt_runtime = rt_runtime;
7439 for_each_possible_cpu(i) {
7440 struct rt_rq *rt_rq = tg->rt_rq[i];
7442 raw_spin_lock(&rt_rq->rt_runtime_lock);
7443 rt_rq->rt_runtime = rt_runtime;
7444 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7446 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7448 read_unlock(&tasklist_lock);
7449 mutex_unlock(&rt_constraints_mutex);
7454 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7456 u64 rt_runtime, rt_period;
7458 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7459 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7460 if (rt_runtime_us < 0)
7461 rt_runtime = RUNTIME_INF;
7463 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7466 long sched_group_rt_runtime(struct task_group *tg)
7470 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7473 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7474 do_div(rt_runtime_us, NSEC_PER_USEC);
7475 return rt_runtime_us;
7478 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7480 u64 rt_runtime, rt_period;
7482 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7483 rt_runtime = tg->rt_bandwidth.rt_runtime;
7488 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7491 long sched_group_rt_period(struct task_group *tg)
7495 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7496 do_div(rt_period_us, NSEC_PER_USEC);
7497 return rt_period_us;
7500 static int sched_rt_global_constraints(void)
7502 u64 runtime, period;
7505 if (sysctl_sched_rt_period <= 0)
7508 runtime = global_rt_runtime();
7509 period = global_rt_period();
7512 * Sanity check on the sysctl variables.
7514 if (runtime > period && runtime != RUNTIME_INF)
7517 mutex_lock(&rt_constraints_mutex);
7518 read_lock(&tasklist_lock);
7519 ret = __rt_schedulable(NULL, 0, 0);
7520 read_unlock(&tasklist_lock);
7521 mutex_unlock(&rt_constraints_mutex);
7526 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7528 /* Don't accept realtime tasks when there is no way for them to run */
7529 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7535 #else /* !CONFIG_RT_GROUP_SCHED */
7536 static int sched_rt_global_constraints(void)
7538 unsigned long flags;
7541 if (sysctl_sched_rt_period <= 0)
7545 * There's always some RT tasks in the root group
7546 * -- migration, kstopmachine etc..
7548 if (sysctl_sched_rt_runtime == 0)
7551 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7552 for_each_possible_cpu(i) {
7553 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7555 raw_spin_lock(&rt_rq->rt_runtime_lock);
7556 rt_rq->rt_runtime = global_rt_runtime();
7557 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7559 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7563 #endif /* CONFIG_RT_GROUP_SCHED */
7565 int sched_rt_handler(struct ctl_table *table, int write,
7566 void __user *buffer, size_t *lenp,
7570 int old_period, old_runtime;
7571 static DEFINE_MUTEX(mutex);
7574 old_period = sysctl_sched_rt_period;
7575 old_runtime = sysctl_sched_rt_runtime;
7577 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7579 if (!ret && write) {
7580 ret = sched_rt_global_constraints();
7582 sysctl_sched_rt_period = old_period;
7583 sysctl_sched_rt_runtime = old_runtime;
7585 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7586 def_rt_bandwidth.rt_period =
7587 ns_to_ktime(global_rt_period());
7590 mutex_unlock(&mutex);
7595 #ifdef CONFIG_CGROUP_SCHED
7597 /* return corresponding task_group object of a cgroup */
7598 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7600 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7601 struct task_group, css);
7604 static struct cgroup_subsys_state *
7605 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7607 struct task_group *tg, *parent;
7609 if (!cgrp->parent) {
7610 /* This is early initialization for the top cgroup */
7611 return &root_task_group.css;
7614 parent = cgroup_tg(cgrp->parent);
7615 tg = sched_create_group(parent);
7617 return ERR_PTR(-ENOMEM);
7623 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7625 struct task_group *tg = cgroup_tg(cgrp);
7627 sched_destroy_group(tg);
7630 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7631 struct cgroup_taskset *tset)
7633 struct task_struct *task;
7635 cgroup_taskset_for_each(task, cgrp, tset) {
7636 #ifdef CONFIG_RT_GROUP_SCHED
7637 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7640 /* We don't support RT-tasks being in separate groups */
7641 if (task->sched_class != &fair_sched_class)
7648 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7649 struct cgroup_taskset *tset)
7651 struct task_struct *task;
7653 cgroup_taskset_for_each(task, cgrp, tset)
7654 sched_move_task(task);
7658 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7659 struct cgroup *old_cgrp, struct task_struct *task)
7662 * cgroup_exit() is called in the copy_process() failure path.
7663 * Ignore this case since the task hasn't ran yet, this avoids
7664 * trying to poke a half freed task state from generic code.
7666 if (!(task->flags & PF_EXITING))
7669 sched_move_task(task);
7672 #ifdef CONFIG_FAIR_GROUP_SCHED
7673 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7676 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7679 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7681 struct task_group *tg = cgroup_tg(cgrp);
7683 return (u64) scale_load_down(tg->shares);
7686 #ifdef CONFIG_CFS_BANDWIDTH
7687 static DEFINE_MUTEX(cfs_constraints_mutex);
7689 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7690 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7692 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7694 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7696 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7697 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7699 if (tg == &root_task_group)
7703 * Ensure we have at some amount of bandwidth every period. This is
7704 * to prevent reaching a state of large arrears when throttled via
7705 * entity_tick() resulting in prolonged exit starvation.
7707 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7711 * Likewise, bound things on the otherside by preventing insane quota
7712 * periods. This also allows us to normalize in computing quota
7715 if (period > max_cfs_quota_period)
7718 mutex_lock(&cfs_constraints_mutex);
7719 ret = __cfs_schedulable(tg, period, quota);
7723 runtime_enabled = quota != RUNTIME_INF;
7724 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7725 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7726 raw_spin_lock_irq(&cfs_b->lock);
7727 cfs_b->period = ns_to_ktime(period);
7728 cfs_b->quota = quota;
7730 __refill_cfs_bandwidth_runtime(cfs_b);
7731 /* restart the period timer (if active) to handle new period expiry */
7732 if (runtime_enabled && cfs_b->timer_active) {
7733 /* force a reprogram */
7734 cfs_b->timer_active = 0;
7735 __start_cfs_bandwidth(cfs_b);
7737 raw_spin_unlock_irq(&cfs_b->lock);
7739 for_each_possible_cpu(i) {
7740 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7741 struct rq *rq = cfs_rq->rq;
7743 raw_spin_lock_irq(&rq->lock);
7744 cfs_rq->runtime_enabled = runtime_enabled;
7745 cfs_rq->runtime_remaining = 0;
7747 if (cfs_rq->throttled)
7748 unthrottle_cfs_rq(cfs_rq);
7749 raw_spin_unlock_irq(&rq->lock);
7752 mutex_unlock(&cfs_constraints_mutex);
7757 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7761 period = ktime_to_ns(tg->cfs_bandwidth.period);
7762 if (cfs_quota_us < 0)
7763 quota = RUNTIME_INF;
7765 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7767 return tg_set_cfs_bandwidth(tg, period, quota);
7770 long tg_get_cfs_quota(struct task_group *tg)
7774 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7777 quota_us = tg->cfs_bandwidth.quota;
7778 do_div(quota_us, NSEC_PER_USEC);
7783 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7787 period = (u64)cfs_period_us * NSEC_PER_USEC;
7788 quota = tg->cfs_bandwidth.quota;
7790 return tg_set_cfs_bandwidth(tg, period, quota);
7793 long tg_get_cfs_period(struct task_group *tg)
7797 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7798 do_div(cfs_period_us, NSEC_PER_USEC);
7800 return cfs_period_us;
7803 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7805 return tg_get_cfs_quota(cgroup_tg(cgrp));
7808 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7811 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7814 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7816 return tg_get_cfs_period(cgroup_tg(cgrp));
7819 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7822 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7825 struct cfs_schedulable_data {
7826 struct task_group *tg;
7831 * normalize group quota/period to be quota/max_period
7832 * note: units are usecs
7834 static u64 normalize_cfs_quota(struct task_group *tg,
7835 struct cfs_schedulable_data *d)
7843 period = tg_get_cfs_period(tg);
7844 quota = tg_get_cfs_quota(tg);
7847 /* note: these should typically be equivalent */
7848 if (quota == RUNTIME_INF || quota == -1)
7851 return to_ratio(period, quota);
7854 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7856 struct cfs_schedulable_data *d = data;
7857 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7858 s64 quota = 0, parent_quota = -1;
7861 quota = RUNTIME_INF;
7863 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7865 quota = normalize_cfs_quota(tg, d);
7866 parent_quota = parent_b->hierarchal_quota;
7869 * ensure max(child_quota) <= parent_quota, inherit when no
7872 if (quota == RUNTIME_INF)
7873 quota = parent_quota;
7874 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7877 cfs_b->hierarchal_quota = quota;
7882 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7885 struct cfs_schedulable_data data = {
7891 if (quota != RUNTIME_INF) {
7892 do_div(data.period, NSEC_PER_USEC);
7893 do_div(data.quota, NSEC_PER_USEC);
7897 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7903 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7904 struct cgroup_map_cb *cb)
7906 struct task_group *tg = cgroup_tg(cgrp);
7907 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7909 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7910 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7911 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7915 #endif /* CONFIG_CFS_BANDWIDTH */
7916 #endif /* CONFIG_FAIR_GROUP_SCHED */
7918 #ifdef CONFIG_RT_GROUP_SCHED
7919 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7922 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7925 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7927 return sched_group_rt_runtime(cgroup_tg(cgrp));
7930 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7933 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7936 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7938 return sched_group_rt_period(cgroup_tg(cgrp));
7940 #endif /* CONFIG_RT_GROUP_SCHED */
7942 static struct cftype cpu_files[] = {
7943 #ifdef CONFIG_FAIR_GROUP_SCHED
7946 .read_u64 = cpu_shares_read_u64,
7947 .write_u64 = cpu_shares_write_u64,
7950 #ifdef CONFIG_CFS_BANDWIDTH
7952 .name = "cfs_quota_us",
7953 .read_s64 = cpu_cfs_quota_read_s64,
7954 .write_s64 = cpu_cfs_quota_write_s64,
7957 .name = "cfs_period_us",
7958 .read_u64 = cpu_cfs_period_read_u64,
7959 .write_u64 = cpu_cfs_period_write_u64,
7963 .read_map = cpu_stats_show,
7966 #ifdef CONFIG_RT_GROUP_SCHED
7968 .name = "rt_runtime_us",
7969 .read_s64 = cpu_rt_runtime_read,
7970 .write_s64 = cpu_rt_runtime_write,
7973 .name = "rt_period_us",
7974 .read_u64 = cpu_rt_period_read_uint,
7975 .write_u64 = cpu_rt_period_write_uint,
7980 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7982 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7985 struct cgroup_subsys cpu_cgroup_subsys = {
7987 .create = cpu_cgroup_create,
7988 .destroy = cpu_cgroup_destroy,
7989 .can_attach = cpu_cgroup_can_attach,
7990 .attach = cpu_cgroup_attach,
7991 .exit = cpu_cgroup_exit,
7992 .populate = cpu_cgroup_populate,
7993 .subsys_id = cpu_cgroup_subsys_id,
7997 #endif /* CONFIG_CGROUP_SCHED */
7999 #ifdef CONFIG_CGROUP_CPUACCT
8002 * CPU accounting code for task groups.
8004 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8005 * (balbir@in.ibm.com).
8008 /* create a new cpu accounting group */
8009 static struct cgroup_subsys_state *cpuacct_create(
8010 struct cgroup_subsys *ss, struct cgroup *cgrp)
8015 return &root_cpuacct.css;
8017 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8021 ca->cpuusage = alloc_percpu(u64);
8025 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8027 goto out_free_cpuusage;
8032 free_percpu(ca->cpuusage);
8036 return ERR_PTR(-ENOMEM);
8039 /* destroy an existing cpu accounting group */
8041 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8043 struct cpuacct *ca = cgroup_ca(cgrp);
8045 free_percpu(ca->cpustat);
8046 free_percpu(ca->cpuusage);
8050 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8052 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8055 #ifndef CONFIG_64BIT
8057 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8059 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8061 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8069 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8071 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8073 #ifndef CONFIG_64BIT
8075 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8077 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8079 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8085 /* return total cpu usage (in nanoseconds) of a group */
8086 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8088 struct cpuacct *ca = cgroup_ca(cgrp);
8089 u64 totalcpuusage = 0;
8092 for_each_present_cpu(i)
8093 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8095 return totalcpuusage;
8098 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8101 struct cpuacct *ca = cgroup_ca(cgrp);
8110 for_each_present_cpu(i)
8111 cpuacct_cpuusage_write(ca, i, 0);
8117 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8120 struct cpuacct *ca = cgroup_ca(cgroup);
8124 for_each_present_cpu(i) {
8125 percpu = cpuacct_cpuusage_read(ca, i);
8126 seq_printf(m, "%llu ", (unsigned long long) percpu);
8128 seq_printf(m, "\n");
8132 static const char *cpuacct_stat_desc[] = {
8133 [CPUACCT_STAT_USER] = "user",
8134 [CPUACCT_STAT_SYSTEM] = "system",
8137 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8138 struct cgroup_map_cb *cb)
8140 struct cpuacct *ca = cgroup_ca(cgrp);
8144 for_each_online_cpu(cpu) {
8145 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8146 val += kcpustat->cpustat[CPUTIME_USER];
8147 val += kcpustat->cpustat[CPUTIME_NICE];
8149 val = cputime64_to_clock_t(val);
8150 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8153 for_each_online_cpu(cpu) {
8154 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8155 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8156 val += kcpustat->cpustat[CPUTIME_IRQ];
8157 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8160 val = cputime64_to_clock_t(val);
8161 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8166 static struct cftype files[] = {
8169 .read_u64 = cpuusage_read,
8170 .write_u64 = cpuusage_write,
8173 .name = "usage_percpu",
8174 .read_seq_string = cpuacct_percpu_seq_read,
8178 .read_map = cpuacct_stats_show,
8182 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8184 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8188 * charge this task's execution time to its accounting group.
8190 * called with rq->lock held.
8192 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8197 if (unlikely(!cpuacct_subsys.active))
8200 cpu = task_cpu(tsk);
8206 for (; ca; ca = parent_ca(ca)) {
8207 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8208 *cpuusage += cputime;
8214 struct cgroup_subsys cpuacct_subsys = {
8216 .create = cpuacct_create,
8217 .destroy = cpuacct_destroy,
8218 .populate = cpuacct_populate,
8219 .subsys_id = cpuacct_subsys_id,
8221 #endif /* CONFIG_CGROUP_CPUACCT */