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(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(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);
1965 finish_arch_post_lock_switch();
1967 fire_sched_in_preempt_notifiers(current);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1982 /* assumes rq->lock is held */
1983 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1989 /* rq->lock is NOT held, but preemption is disabled */
1990 static inline void post_schedule(struct rq *rq)
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
2000 rq->post_schedule = 0;
2006 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2010 static inline void post_schedule(struct rq *rq)
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2020 asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2023 struct rq *rq = this_rq();
2025 finish_task_switch(rq, prev);
2028 * FIXME: do we need to worry about rq being invalidated by the
2033 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2046 context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2049 struct mm_struct *mm, *oldmm;
2051 prepare_task_switch(rq, prev, next);
2054 oldmm = prev->active_mm;
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2060 arch_start_context_switch(prev);
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2067 switch_mm(oldmm, mm, next);
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2079 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2083 /* Here we just switch the register state and the stack. */
2084 switch_to(prev, next, prev);
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2092 finish_task_switch(this_rq(), prev);
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2102 unsigned long nr_running(void)
2104 unsigned long i, sum = 0;
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2112 unsigned long nr_uninterruptible(void)
2114 unsigned long i, sum = 0;
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2123 if (unlikely((long)sum < 0))
2129 unsigned long long nr_context_switches(void)
2132 unsigned long long sum = 0;
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2140 unsigned long nr_iowait(void)
2142 unsigned long i, sum = 0;
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2150 unsigned long nr_iowait_cpu(int cpu)
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2156 unsigned long this_cpu_load(void)
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2163 /* Variables and functions for calc_load */
2164 static atomic_long_t calc_load_tasks;
2165 static unsigned long calc_load_update;
2166 unsigned long avenrun[3];
2167 EXPORT_SYMBOL(avenrun);
2169 static long calc_load_fold_active(struct rq *this_rq)
2171 long nr_active, delta = 0;
2173 nr_active = this_rq->nr_running;
2174 nr_active += (long) this_rq->nr_uninterruptible;
2176 if (nr_active != this_rq->calc_load_active) {
2177 delta = nr_active - this_rq->calc_load_active;
2178 this_rq->calc_load_active = nr_active;
2184 static unsigned long
2185 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2188 load += active * (FIXED_1 - exp);
2189 load += 1UL << (FSHIFT - 1);
2190 return load >> FSHIFT;
2195 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2197 * When making the ILB scale, we should try to pull this in as well.
2199 static atomic_long_t calc_load_tasks_idle;
2201 void calc_load_account_idle(struct rq *this_rq)
2205 delta = calc_load_fold_active(this_rq);
2207 atomic_long_add(delta, &calc_load_tasks_idle);
2210 static long calc_load_fold_idle(void)
2215 * Its got a race, we don't care...
2217 if (atomic_long_read(&calc_load_tasks_idle))
2218 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2224 * fixed_power_int - compute: x^n, in O(log n) time
2226 * @x: base of the power
2227 * @frac_bits: fractional bits of @x
2228 * @n: power to raise @x to.
2230 * By exploiting the relation between the definition of the natural power
2231 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2232 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2233 * (where: n_i \elem {0, 1}, the binary vector representing n),
2234 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2235 * of course trivially computable in O(log_2 n), the length of our binary
2238 static unsigned long
2239 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2241 unsigned long result = 1UL << frac_bits;
2246 result += 1UL << (frac_bits - 1);
2247 result >>= frac_bits;
2253 x += 1UL << (frac_bits - 1);
2261 * a1 = a0 * e + a * (1 - e)
2263 * a2 = a1 * e + a * (1 - e)
2264 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2265 * = a0 * e^2 + a * (1 - e) * (1 + e)
2267 * a3 = a2 * e + a * (1 - e)
2268 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2269 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2273 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2274 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2275 * = a0 * e^n + a * (1 - e^n)
2277 * [1] application of the geometric series:
2280 * S_n := \Sum x^i = -------------
2283 static unsigned long
2284 calc_load_n(unsigned long load, unsigned long exp,
2285 unsigned long active, unsigned int n)
2288 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2292 * NO_HZ can leave us missing all per-cpu ticks calling
2293 * calc_load_account_active(), but since an idle CPU folds its delta into
2294 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2295 * in the pending idle delta if our idle period crossed a load cycle boundary.
2297 * Once we've updated the global active value, we need to apply the exponential
2298 * weights adjusted to the number of cycles missed.
2300 static void calc_global_nohz(void)
2302 long delta, active, n;
2305 * If we crossed a calc_load_update boundary, make sure to fold
2306 * any pending idle changes, the respective CPUs might have
2307 * missed the tick driven calc_load_account_active() update
2310 delta = calc_load_fold_idle();
2312 atomic_long_add(delta, &calc_load_tasks);
2315 * It could be the one fold was all it took, we done!
2317 if (time_before(jiffies, calc_load_update + 10))
2321 * Catch-up, fold however many we are behind still
2323 delta = jiffies - calc_load_update - 10;
2324 n = 1 + (delta / LOAD_FREQ);
2326 active = atomic_long_read(&calc_load_tasks);
2327 active = active > 0 ? active * FIXED_1 : 0;
2329 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2330 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2331 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2333 calc_load_update += n * LOAD_FREQ;
2336 void calc_load_account_idle(struct rq *this_rq)
2340 static inline long calc_load_fold_idle(void)
2345 static void calc_global_nohz(void)
2351 * get_avenrun - get the load average array
2352 * @loads: pointer to dest load array
2353 * @offset: offset to add
2354 * @shift: shift count to shift the result left
2356 * These values are estimates at best, so no need for locking.
2358 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2360 loads[0] = (avenrun[0] + offset) << shift;
2361 loads[1] = (avenrun[1] + offset) << shift;
2362 loads[2] = (avenrun[2] + offset) << shift;
2366 * calc_load - update the avenrun load estimates 10 ticks after the
2367 * CPUs have updated calc_load_tasks.
2369 void calc_global_load(unsigned long ticks)
2373 if (time_before(jiffies, calc_load_update + 10))
2376 active = atomic_long_read(&calc_load_tasks);
2377 active = active > 0 ? active * FIXED_1 : 0;
2379 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2380 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2381 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2383 calc_load_update += LOAD_FREQ;
2386 * Account one period with whatever state we found before
2387 * folding in the nohz state and ageing the entire idle period.
2389 * This avoids loosing a sample when we go idle between
2390 * calc_load_account_active() (10 ticks ago) and now and thus
2397 * Called from update_cpu_load() to periodically update this CPU's
2400 static void calc_load_account_active(struct rq *this_rq)
2404 if (time_before(jiffies, this_rq->calc_load_update))
2407 delta = calc_load_fold_active(this_rq);
2408 delta += calc_load_fold_idle();
2410 atomic_long_add(delta, &calc_load_tasks);
2412 this_rq->calc_load_update += LOAD_FREQ;
2416 * The exact cpuload at various idx values, calculated at every tick would be
2417 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2419 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2420 * on nth tick when cpu may be busy, then we have:
2421 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2422 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2424 * decay_load_missed() below does efficient calculation of
2425 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2426 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2428 * The calculation is approximated on a 128 point scale.
2429 * degrade_zero_ticks is the number of ticks after which load at any
2430 * particular idx is approximated to be zero.
2431 * degrade_factor is a precomputed table, a row for each load idx.
2432 * Each column corresponds to degradation factor for a power of two ticks,
2433 * based on 128 point scale.
2435 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2436 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2438 * With this power of 2 load factors, we can degrade the load n times
2439 * by looking at 1 bits in n and doing as many mult/shift instead of
2440 * n mult/shifts needed by the exact degradation.
2442 #define DEGRADE_SHIFT 7
2443 static const unsigned char
2444 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2445 static const unsigned char
2446 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2447 {0, 0, 0, 0, 0, 0, 0, 0},
2448 {64, 32, 8, 0, 0, 0, 0, 0},
2449 {96, 72, 40, 12, 1, 0, 0},
2450 {112, 98, 75, 43, 15, 1, 0},
2451 {120, 112, 98, 76, 45, 16, 2} };
2454 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2455 * would be when CPU is idle and so we just decay the old load without
2456 * adding any new load.
2458 static unsigned long
2459 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2463 if (!missed_updates)
2466 if (missed_updates >= degrade_zero_ticks[idx])
2470 return load >> missed_updates;
2472 while (missed_updates) {
2473 if (missed_updates % 2)
2474 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2476 missed_updates >>= 1;
2483 * Update rq->cpu_load[] statistics. This function is usually called every
2484 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2485 * every tick. We fix it up based on jiffies.
2487 void update_cpu_load(struct rq *this_rq)
2489 unsigned long this_load = this_rq->load.weight;
2490 unsigned long curr_jiffies = jiffies;
2491 unsigned long pending_updates;
2494 this_rq->nr_load_updates++;
2496 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2497 if (curr_jiffies == this_rq->last_load_update_tick)
2500 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2501 this_rq->last_load_update_tick = curr_jiffies;
2503 /* Update our load: */
2504 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2505 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2506 unsigned long old_load, new_load;
2508 /* scale is effectively 1 << i now, and >> i divides by scale */
2510 old_load = this_rq->cpu_load[i];
2511 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2512 new_load = this_load;
2514 * Round up the averaging division if load is increasing. This
2515 * prevents us from getting stuck on 9 if the load is 10, for
2518 if (new_load > old_load)
2519 new_load += scale - 1;
2521 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2524 sched_avg_update(this_rq);
2527 static void update_cpu_load_active(struct rq *this_rq)
2529 update_cpu_load(this_rq);
2531 calc_load_account_active(this_rq);
2537 * sched_exec - execve() is a valuable balancing opportunity, because at
2538 * this point the task has the smallest effective memory and cache footprint.
2540 void sched_exec(void)
2542 struct task_struct *p = current;
2543 unsigned long flags;
2546 raw_spin_lock_irqsave(&p->pi_lock, flags);
2547 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2548 if (dest_cpu == smp_processor_id())
2551 if (likely(cpu_active(dest_cpu))) {
2552 struct migration_arg arg = { p, dest_cpu };
2554 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2555 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2559 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2564 DEFINE_PER_CPU(struct kernel_stat, kstat);
2565 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2567 EXPORT_PER_CPU_SYMBOL(kstat);
2568 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2571 * Return any ns on the sched_clock that have not yet been accounted in
2572 * @p in case that task is currently running.
2574 * Called with task_rq_lock() held on @rq.
2576 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2580 if (task_current(rq, p)) {
2581 update_rq_clock(rq);
2582 ns = rq->clock_task - p->se.exec_start;
2590 unsigned long long task_delta_exec(struct task_struct *p)
2592 unsigned long flags;
2596 rq = task_rq_lock(p, &flags);
2597 ns = do_task_delta_exec(p, rq);
2598 task_rq_unlock(rq, p, &flags);
2604 * Return accounted runtime for the task.
2605 * In case the task is currently running, return the runtime plus current's
2606 * pending runtime that have not been accounted yet.
2608 unsigned long long task_sched_runtime(struct task_struct *p)
2610 unsigned long flags;
2614 rq = task_rq_lock(p, &flags);
2615 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2616 task_rq_unlock(rq, p, &flags);
2621 #ifdef CONFIG_CGROUP_CPUACCT
2622 struct cgroup_subsys cpuacct_subsys;
2623 struct cpuacct root_cpuacct;
2626 static inline void task_group_account_field(struct task_struct *p, int index,
2629 #ifdef CONFIG_CGROUP_CPUACCT
2630 struct kernel_cpustat *kcpustat;
2634 * Since all updates are sure to touch the root cgroup, we
2635 * get ourselves ahead and touch it first. If the root cgroup
2636 * is the only cgroup, then nothing else should be necessary.
2639 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2641 #ifdef CONFIG_CGROUP_CPUACCT
2642 if (unlikely(!cpuacct_subsys.active))
2647 while (ca && (ca != &root_cpuacct)) {
2648 kcpustat = this_cpu_ptr(ca->cpustat);
2649 kcpustat->cpustat[index] += tmp;
2658 * Account user cpu time to a process.
2659 * @p: the process that the cpu time gets accounted to
2660 * @cputime: the cpu time spent in user space since the last update
2661 * @cputime_scaled: cputime scaled by cpu frequency
2663 void account_user_time(struct task_struct *p, cputime_t cputime,
2664 cputime_t cputime_scaled)
2668 /* Add user time to process. */
2669 p->utime += cputime;
2670 p->utimescaled += cputime_scaled;
2671 account_group_user_time(p, cputime);
2673 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2675 /* Add user time to cpustat. */
2676 task_group_account_field(p, index, (__force u64) cputime);
2678 /* Account for user time used */
2679 acct_update_integrals(p);
2683 * Account guest cpu time to a process.
2684 * @p: the process that the cpu time gets accounted to
2685 * @cputime: the cpu time spent in virtual machine since the last update
2686 * @cputime_scaled: cputime scaled by cpu frequency
2688 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2689 cputime_t cputime_scaled)
2691 u64 *cpustat = kcpustat_this_cpu->cpustat;
2693 /* Add guest time to process. */
2694 p->utime += cputime;
2695 p->utimescaled += cputime_scaled;
2696 account_group_user_time(p, cputime);
2697 p->gtime += cputime;
2699 /* Add guest time to cpustat. */
2700 if (TASK_NICE(p) > 0) {
2701 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2702 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2704 cpustat[CPUTIME_USER] += (__force u64) cputime;
2705 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2710 * Account system cpu time to a process and desired cpustat field
2711 * @p: the process that the cpu time gets accounted to
2712 * @cputime: the cpu time spent in kernel space since the last update
2713 * @cputime_scaled: cputime scaled by cpu frequency
2714 * @target_cputime64: pointer to cpustat field that has to be updated
2717 void __account_system_time(struct task_struct *p, cputime_t cputime,
2718 cputime_t cputime_scaled, int index)
2720 /* Add system time to process. */
2721 p->stime += cputime;
2722 p->stimescaled += cputime_scaled;
2723 account_group_system_time(p, cputime);
2725 /* Add system time to cpustat. */
2726 task_group_account_field(p, index, (__force u64) cputime);
2728 /* Account for system time used */
2729 acct_update_integrals(p);
2733 * Account system cpu time to a process.
2734 * @p: the process that the cpu time gets accounted to
2735 * @hardirq_offset: the offset to subtract from hardirq_count()
2736 * @cputime: the cpu time spent in kernel space since the last update
2737 * @cputime_scaled: cputime scaled by cpu frequency
2739 void account_system_time(struct task_struct *p, int hardirq_offset,
2740 cputime_t cputime, cputime_t cputime_scaled)
2744 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2745 account_guest_time(p, cputime, cputime_scaled);
2749 if (hardirq_count() - hardirq_offset)
2750 index = CPUTIME_IRQ;
2751 else if (in_serving_softirq())
2752 index = CPUTIME_SOFTIRQ;
2754 index = CPUTIME_SYSTEM;
2756 __account_system_time(p, cputime, cputime_scaled, index);
2760 * Account for involuntary wait time.
2761 * @cputime: the cpu time spent in involuntary wait
2763 void account_steal_time(cputime_t cputime)
2765 u64 *cpustat = kcpustat_this_cpu->cpustat;
2767 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2771 * Account for idle time.
2772 * @cputime: the cpu time spent in idle wait
2774 void account_idle_time(cputime_t cputime)
2776 u64 *cpustat = kcpustat_this_cpu->cpustat;
2777 struct rq *rq = this_rq();
2779 if (atomic_read(&rq->nr_iowait) > 0)
2780 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2782 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2785 static __always_inline bool steal_account_process_tick(void)
2787 #ifdef CONFIG_PARAVIRT
2788 if (static_branch(¶virt_steal_enabled)) {
2791 steal = paravirt_steal_clock(smp_processor_id());
2792 steal -= this_rq()->prev_steal_time;
2794 st = steal_ticks(steal);
2795 this_rq()->prev_steal_time += st * TICK_NSEC;
2797 account_steal_time(st);
2804 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2806 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2808 * Account a tick to a process and cpustat
2809 * @p: the process that the cpu time gets accounted to
2810 * @user_tick: is the tick from userspace
2811 * @rq: the pointer to rq
2813 * Tick demultiplexing follows the order
2814 * - pending hardirq update
2815 * - pending softirq update
2819 * - check for guest_time
2820 * - else account as system_time
2822 * Check for hardirq is done both for system and user time as there is
2823 * no timer going off while we are on hardirq and hence we may never get an
2824 * opportunity to update it solely in system time.
2825 * p->stime and friends are only updated on system time and not on irq
2826 * softirq as those do not count in task exec_runtime any more.
2828 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2831 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2832 u64 *cpustat = kcpustat_this_cpu->cpustat;
2834 if (steal_account_process_tick())
2837 if (irqtime_account_hi_update()) {
2838 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2839 } else if (irqtime_account_si_update()) {
2840 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2841 } else if (this_cpu_ksoftirqd() == p) {
2843 * ksoftirqd time do not get accounted in cpu_softirq_time.
2844 * So, we have to handle it separately here.
2845 * Also, p->stime needs to be updated for ksoftirqd.
2847 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2849 } else if (user_tick) {
2850 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2851 } else if (p == rq->idle) {
2852 account_idle_time(cputime_one_jiffy);
2853 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2854 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2856 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2861 static void irqtime_account_idle_ticks(int ticks)
2864 struct rq *rq = this_rq();
2866 for (i = 0; i < ticks; i++)
2867 irqtime_account_process_tick(current, 0, rq);
2869 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2870 static void irqtime_account_idle_ticks(int ticks) {}
2871 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2873 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2876 * Account a single tick of cpu time.
2877 * @p: the process that the cpu time gets accounted to
2878 * @user_tick: indicates if the tick is a user or a system tick
2880 void account_process_tick(struct task_struct *p, int user_tick)
2882 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2883 struct rq *rq = this_rq();
2885 if (sched_clock_irqtime) {
2886 irqtime_account_process_tick(p, user_tick, rq);
2890 if (steal_account_process_tick())
2894 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2895 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2896 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2899 account_idle_time(cputime_one_jiffy);
2903 * Account multiple ticks of steal time.
2904 * @p: the process from which the cpu time has been stolen
2905 * @ticks: number of stolen ticks
2907 void account_steal_ticks(unsigned long ticks)
2909 account_steal_time(jiffies_to_cputime(ticks));
2913 * Account multiple ticks of idle time.
2914 * @ticks: number of stolen ticks
2916 void account_idle_ticks(unsigned long ticks)
2919 if (sched_clock_irqtime) {
2920 irqtime_account_idle_ticks(ticks);
2924 account_idle_time(jiffies_to_cputime(ticks));
2930 * Use precise platform statistics if available:
2932 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2933 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2939 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2941 struct task_cputime cputime;
2943 thread_group_cputime(p, &cputime);
2945 *ut = cputime.utime;
2946 *st = cputime.stime;
2950 #ifndef nsecs_to_cputime
2951 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2954 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2956 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2959 * Use CFS's precise accounting:
2961 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2964 u64 temp = (__force u64) rtime;
2966 temp *= (__force u64) utime;
2967 do_div(temp, (__force u32) total);
2968 utime = (__force cputime_t) temp;
2973 * Compare with previous values, to keep monotonicity:
2975 p->prev_utime = max(p->prev_utime, utime);
2976 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2978 *ut = p->prev_utime;
2979 *st = p->prev_stime;
2983 * Must be called with siglock held.
2985 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2987 struct signal_struct *sig = p->signal;
2988 struct task_cputime cputime;
2989 cputime_t rtime, utime, total;
2991 thread_group_cputime(p, &cputime);
2993 total = cputime.utime + cputime.stime;
2994 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2997 u64 temp = (__force u64) rtime;
2999 temp *= (__force u64) cputime.utime;
3000 do_div(temp, (__force u32) total);
3001 utime = (__force cputime_t) temp;
3005 sig->prev_utime = max(sig->prev_utime, utime);
3006 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3008 *ut = sig->prev_utime;
3009 *st = sig->prev_stime;
3014 * This function gets called by the timer code, with HZ frequency.
3015 * We call it with interrupts disabled.
3017 void scheduler_tick(void)
3019 int cpu = smp_processor_id();
3020 struct rq *rq = cpu_rq(cpu);
3021 struct task_struct *curr = rq->curr;
3025 raw_spin_lock(&rq->lock);
3026 update_rq_clock(rq);
3027 update_cpu_load_active(rq);
3028 curr->sched_class->task_tick(rq, curr, 0);
3029 raw_spin_unlock(&rq->lock);
3031 perf_event_task_tick();
3034 rq->idle_balance = idle_cpu(cpu);
3035 trigger_load_balance(rq, cpu);
3039 notrace unsigned long get_parent_ip(unsigned long addr)
3041 if (in_lock_functions(addr)) {
3042 addr = CALLER_ADDR2;
3043 if (in_lock_functions(addr))
3044 addr = CALLER_ADDR3;
3049 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3050 defined(CONFIG_PREEMPT_TRACER))
3052 void __kprobes add_preempt_count(int val)
3054 #ifdef CONFIG_DEBUG_PREEMPT
3058 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3061 preempt_count() += val;
3062 #ifdef CONFIG_DEBUG_PREEMPT
3064 * Spinlock count overflowing soon?
3066 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3069 if (preempt_count() == val)
3070 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3072 EXPORT_SYMBOL(add_preempt_count);
3074 void __kprobes sub_preempt_count(int val)
3076 #ifdef CONFIG_DEBUG_PREEMPT
3080 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3083 * Is the spinlock portion underflowing?
3085 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3086 !(preempt_count() & PREEMPT_MASK)))
3090 if (preempt_count() == val)
3091 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3092 preempt_count() -= val;
3094 EXPORT_SYMBOL(sub_preempt_count);
3099 * Print scheduling while atomic bug:
3101 static noinline void __schedule_bug(struct task_struct *prev)
3103 if (oops_in_progress)
3106 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3107 prev->comm, prev->pid, preempt_count());
3109 debug_show_held_locks(prev);
3111 if (irqs_disabled())
3112 print_irqtrace_events(prev);
3117 * Various schedule()-time debugging checks and statistics:
3119 static inline void schedule_debug(struct task_struct *prev)
3122 * Test if we are atomic. Since do_exit() needs to call into
3123 * schedule() atomically, we ignore that path for now.
3124 * Otherwise, whine if we are scheduling when we should not be.
3126 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3127 __schedule_bug(prev);
3130 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3132 schedstat_inc(this_rq(), sched_count);
3135 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3137 if (prev->on_rq || rq->skip_clock_update < 0)
3138 update_rq_clock(rq);
3139 prev->sched_class->put_prev_task(rq, prev);
3143 * Pick up the highest-prio task:
3145 static inline struct task_struct *
3146 pick_next_task(struct rq *rq)
3148 const struct sched_class *class;
3149 struct task_struct *p;
3152 * Optimization: we know that if all tasks are in
3153 * the fair class we can call that function directly:
3155 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3156 p = fair_sched_class.pick_next_task(rq);
3161 for_each_class(class) {
3162 p = class->pick_next_task(rq);
3167 BUG(); /* the idle class will always have a runnable task */
3171 * __schedule() is the main scheduler function.
3173 static void __sched __schedule(void)
3175 struct task_struct *prev, *next;
3176 unsigned long *switch_count;
3182 cpu = smp_processor_id();
3184 rcu_note_context_switch(cpu);
3187 schedule_debug(prev);
3189 if (sched_feat(HRTICK))
3192 raw_spin_lock_irq(&rq->lock);
3194 switch_count = &prev->nivcsw;
3195 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3196 if (unlikely(signal_pending_state(prev->state, prev))) {
3197 prev->state = TASK_RUNNING;
3199 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3203 * If a worker went to sleep, notify and ask workqueue
3204 * whether it wants to wake up a task to maintain
3207 if (prev->flags & PF_WQ_WORKER) {
3208 struct task_struct *to_wakeup;
3210 to_wakeup = wq_worker_sleeping(prev, cpu);
3212 try_to_wake_up_local(to_wakeup);
3215 switch_count = &prev->nvcsw;
3218 pre_schedule(rq, prev);
3220 if (unlikely(!rq->nr_running))
3221 idle_balance(cpu, rq);
3223 put_prev_task(rq, prev);
3224 next = pick_next_task(rq);
3225 clear_tsk_need_resched(prev);
3226 rq->skip_clock_update = 0;
3228 if (likely(prev != next)) {
3233 context_switch(rq, prev, next); /* unlocks the rq */
3235 * The context switch have flipped the stack from under us
3236 * and restored the local variables which were saved when
3237 * this task called schedule() in the past. prev == current
3238 * is still correct, but it can be moved to another cpu/rq.
3240 cpu = smp_processor_id();
3243 raw_spin_unlock_irq(&rq->lock);
3247 sched_preempt_enable_no_resched();
3252 static inline void sched_submit_work(struct task_struct *tsk)
3254 if (!tsk->state || tsk_is_pi_blocked(tsk))
3257 * If we are going to sleep and we have plugged IO queued,
3258 * make sure to submit it to avoid deadlocks.
3260 if (blk_needs_flush_plug(tsk))
3261 blk_schedule_flush_plug(tsk);
3264 asmlinkage void __sched schedule(void)
3266 struct task_struct *tsk = current;
3268 sched_submit_work(tsk);
3271 EXPORT_SYMBOL(schedule);
3274 * schedule_preempt_disabled - called with preemption disabled
3276 * Returns with preemption disabled. Note: preempt_count must be 1
3278 void __sched schedule_preempt_disabled(void)
3280 sched_preempt_enable_no_resched();
3285 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3287 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3289 if (lock->owner != owner)
3293 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3294 * lock->owner still matches owner, if that fails, owner might
3295 * point to free()d memory, if it still matches, the rcu_read_lock()
3296 * ensures the memory stays valid.
3300 return owner->on_cpu;
3304 * Look out! "owner" is an entirely speculative pointer
3305 * access and not reliable.
3307 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3309 if (!sched_feat(OWNER_SPIN))
3313 while (owner_running(lock, owner)) {
3317 arch_mutex_cpu_relax();
3322 * We break out the loop above on need_resched() and when the
3323 * owner changed, which is a sign for heavy contention. Return
3324 * success only when lock->owner is NULL.
3326 return lock->owner == NULL;
3330 #ifdef CONFIG_PREEMPT
3332 * this is the entry point to schedule() from in-kernel preemption
3333 * off of preempt_enable. Kernel preemptions off return from interrupt
3334 * occur there and call schedule directly.
3336 asmlinkage void __sched notrace preempt_schedule(void)
3338 struct thread_info *ti = current_thread_info();
3341 * If there is a non-zero preempt_count or interrupts are disabled,
3342 * we do not want to preempt the current task. Just return..
3344 if (likely(ti->preempt_count || irqs_disabled()))
3348 add_preempt_count_notrace(PREEMPT_ACTIVE);
3350 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3353 * Check again in case we missed a preemption opportunity
3354 * between schedule and now.
3357 } while (need_resched());
3359 EXPORT_SYMBOL(preempt_schedule);
3362 * this is the entry point to schedule() from kernel preemption
3363 * off of irq context.
3364 * Note, that this is called and return with irqs disabled. This will
3365 * protect us against recursive calling from irq.
3367 asmlinkage void __sched preempt_schedule_irq(void)
3369 struct thread_info *ti = current_thread_info();
3371 /* Catch callers which need to be fixed */
3372 BUG_ON(ti->preempt_count || !irqs_disabled());
3375 add_preempt_count(PREEMPT_ACTIVE);
3378 local_irq_disable();
3379 sub_preempt_count(PREEMPT_ACTIVE);
3382 * Check again in case we missed a preemption opportunity
3383 * between schedule and now.
3386 } while (need_resched());
3389 #endif /* CONFIG_PREEMPT */
3391 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3394 return try_to_wake_up(curr->private, mode, wake_flags);
3396 EXPORT_SYMBOL(default_wake_function);
3399 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3400 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3401 * number) then we wake all the non-exclusive tasks and one exclusive task.
3403 * There are circumstances in which we can try to wake a task which has already
3404 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3405 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3407 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3408 int nr_exclusive, int wake_flags, void *key)
3410 wait_queue_t *curr, *next;
3412 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3413 unsigned flags = curr->flags;
3415 if (curr->func(curr, mode, wake_flags, key) &&
3416 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3422 * __wake_up - wake up threads blocked on a waitqueue.
3424 * @mode: which threads
3425 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3426 * @key: is directly passed to the wakeup function
3428 * It may be assumed that this function implies a write memory barrier before
3429 * changing the task state if and only if any tasks are woken up.
3431 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3432 int nr_exclusive, void *key)
3434 unsigned long flags;
3436 spin_lock_irqsave(&q->lock, flags);
3437 __wake_up_common(q, mode, nr_exclusive, 0, key);
3438 spin_unlock_irqrestore(&q->lock, flags);
3440 EXPORT_SYMBOL(__wake_up);
3443 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3445 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3447 __wake_up_common(q, mode, nr, 0, NULL);
3449 EXPORT_SYMBOL_GPL(__wake_up_locked);
3451 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3453 __wake_up_common(q, mode, 1, 0, key);
3455 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3458 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3460 * @mode: which threads
3461 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3462 * @key: opaque value to be passed to wakeup targets
3464 * The sync wakeup differs that the waker knows that it will schedule
3465 * away soon, so while the target thread will be woken up, it will not
3466 * be migrated to another CPU - ie. the two threads are 'synchronized'
3467 * with each other. This can prevent needless bouncing between CPUs.
3469 * On UP it can prevent extra preemption.
3471 * It may be assumed that this function implies a write memory barrier before
3472 * changing the task state if and only if any tasks are woken up.
3474 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3475 int nr_exclusive, void *key)
3477 unsigned long flags;
3478 int wake_flags = WF_SYNC;
3483 if (unlikely(!nr_exclusive))
3486 spin_lock_irqsave(&q->lock, flags);
3487 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3488 spin_unlock_irqrestore(&q->lock, flags);
3490 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3493 * __wake_up_sync - see __wake_up_sync_key()
3495 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3497 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3499 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3502 * complete: - signals a single thread waiting on this completion
3503 * @x: holds the state of this particular completion
3505 * This will wake up a single thread waiting on this completion. Threads will be
3506 * awakened in the same order in which they were queued.
3508 * See also complete_all(), wait_for_completion() and related routines.
3510 * It may be assumed that this function implies a write memory barrier before
3511 * changing the task state if and only if any tasks are woken up.
3513 void complete(struct completion *x)
3515 unsigned long flags;
3517 spin_lock_irqsave(&x->wait.lock, flags);
3519 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3520 spin_unlock_irqrestore(&x->wait.lock, flags);
3522 EXPORT_SYMBOL(complete);
3525 * complete_all: - signals all threads waiting on this completion
3526 * @x: holds the state of this particular completion
3528 * This will wake up all threads waiting on this particular completion event.
3530 * It may be assumed that this function implies a write memory barrier before
3531 * changing the task state if and only if any tasks are woken up.
3533 void complete_all(struct completion *x)
3535 unsigned long flags;
3537 spin_lock_irqsave(&x->wait.lock, flags);
3538 x->done += UINT_MAX/2;
3539 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3540 spin_unlock_irqrestore(&x->wait.lock, flags);
3542 EXPORT_SYMBOL(complete_all);
3544 static inline long __sched
3545 do_wait_for_common(struct completion *x, long timeout, int state)
3548 DECLARE_WAITQUEUE(wait, current);
3550 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3552 if (signal_pending_state(state, current)) {
3553 timeout = -ERESTARTSYS;
3556 __set_current_state(state);
3557 spin_unlock_irq(&x->wait.lock);
3558 timeout = schedule_timeout(timeout);
3559 spin_lock_irq(&x->wait.lock);
3560 } while (!x->done && timeout);
3561 __remove_wait_queue(&x->wait, &wait);
3566 return timeout ?: 1;
3570 wait_for_common(struct completion *x, long timeout, int state)
3574 spin_lock_irq(&x->wait.lock);
3575 timeout = do_wait_for_common(x, timeout, state);
3576 spin_unlock_irq(&x->wait.lock);
3581 * wait_for_completion: - waits for completion of a task
3582 * @x: holds the state of this particular completion
3584 * This waits to be signaled for completion of a specific task. It is NOT
3585 * interruptible and there is no timeout.
3587 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3588 * and interrupt capability. Also see complete().
3590 void __sched wait_for_completion(struct completion *x)
3592 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3594 EXPORT_SYMBOL(wait_for_completion);
3597 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3598 * @x: holds the state of this particular completion
3599 * @timeout: timeout value in jiffies
3601 * This waits for either a completion of a specific task to be signaled or for a
3602 * specified timeout to expire. The timeout is in jiffies. It is not
3605 * The return value is 0 if timed out, and positive (at least 1, or number of
3606 * jiffies left till timeout) if completed.
3608 unsigned long __sched
3609 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3611 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3613 EXPORT_SYMBOL(wait_for_completion_timeout);
3616 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3617 * @x: holds the state of this particular completion
3619 * This waits for completion of a specific task to be signaled. It is
3622 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3624 int __sched wait_for_completion_interruptible(struct completion *x)
3626 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3627 if (t == -ERESTARTSYS)
3631 EXPORT_SYMBOL(wait_for_completion_interruptible);
3634 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3635 * @x: holds the state of this particular completion
3636 * @timeout: timeout value in jiffies
3638 * This waits for either a completion of a specific task to be signaled or for a
3639 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3641 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3642 * positive (at least 1, or number of jiffies left till timeout) if completed.
3645 wait_for_completion_interruptible_timeout(struct completion *x,
3646 unsigned long timeout)
3648 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3650 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3653 * wait_for_completion_killable: - waits for completion of a task (killable)
3654 * @x: holds the state of this particular completion
3656 * This waits to be signaled for completion of a specific task. It can be
3657 * interrupted by a kill signal.
3659 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3661 int __sched wait_for_completion_killable(struct completion *x)
3663 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3664 if (t == -ERESTARTSYS)
3668 EXPORT_SYMBOL(wait_for_completion_killable);
3671 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3672 * @x: holds the state of this particular completion
3673 * @timeout: timeout value in jiffies
3675 * This waits for either a completion of a specific task to be
3676 * signaled or for a specified timeout to expire. It can be
3677 * interrupted by a kill signal. The timeout is in jiffies.
3679 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3680 * positive (at least 1, or number of jiffies left till timeout) if completed.
3683 wait_for_completion_killable_timeout(struct completion *x,
3684 unsigned long timeout)
3686 return wait_for_common(x, timeout, TASK_KILLABLE);
3688 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3691 * try_wait_for_completion - try to decrement a completion without blocking
3692 * @x: completion structure
3694 * Returns: 0 if a decrement cannot be done without blocking
3695 * 1 if a decrement succeeded.
3697 * If a completion is being used as a counting completion,
3698 * attempt to decrement the counter without blocking. This
3699 * enables us to avoid waiting if the resource the completion
3700 * is protecting is not available.
3702 bool try_wait_for_completion(struct completion *x)
3704 unsigned long flags;
3707 spin_lock_irqsave(&x->wait.lock, flags);
3712 spin_unlock_irqrestore(&x->wait.lock, flags);
3715 EXPORT_SYMBOL(try_wait_for_completion);
3718 * completion_done - Test to see if a completion has any waiters
3719 * @x: completion structure
3721 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3722 * 1 if there are no waiters.
3725 bool completion_done(struct completion *x)
3727 unsigned long flags;
3730 spin_lock_irqsave(&x->wait.lock, flags);
3733 spin_unlock_irqrestore(&x->wait.lock, flags);
3736 EXPORT_SYMBOL(completion_done);
3739 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3741 unsigned long flags;
3744 init_waitqueue_entry(&wait, current);
3746 __set_current_state(state);
3748 spin_lock_irqsave(&q->lock, flags);
3749 __add_wait_queue(q, &wait);
3750 spin_unlock(&q->lock);
3751 timeout = schedule_timeout(timeout);
3752 spin_lock_irq(&q->lock);
3753 __remove_wait_queue(q, &wait);
3754 spin_unlock_irqrestore(&q->lock, flags);
3759 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3761 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3763 EXPORT_SYMBOL(interruptible_sleep_on);
3766 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3768 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3770 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3772 void __sched sleep_on(wait_queue_head_t *q)
3774 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3776 EXPORT_SYMBOL(sleep_on);
3778 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3780 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3782 EXPORT_SYMBOL(sleep_on_timeout);
3784 #ifdef CONFIG_RT_MUTEXES
3787 * rt_mutex_setprio - set the current priority of a task
3789 * @prio: prio value (kernel-internal form)
3791 * This function changes the 'effective' priority of a task. It does
3792 * not touch ->normal_prio like __setscheduler().
3794 * Used by the rt_mutex code to implement priority inheritance logic.
3796 void rt_mutex_setprio(struct task_struct *p, int prio)
3798 int oldprio, on_rq, running;
3800 const struct sched_class *prev_class;
3802 BUG_ON(prio < 0 || prio > MAX_PRIO);
3804 rq = __task_rq_lock(p);
3807 * Idle task boosting is a nono in general. There is one
3808 * exception, when PREEMPT_RT and NOHZ is active:
3810 * The idle task calls get_next_timer_interrupt() and holds
3811 * the timer wheel base->lock on the CPU and another CPU wants
3812 * to access the timer (probably to cancel it). We can safely
3813 * ignore the boosting request, as the idle CPU runs this code
3814 * with interrupts disabled and will complete the lock
3815 * protected section without being interrupted. So there is no
3816 * real need to boost.
3818 if (unlikely(p == rq->idle)) {
3819 WARN_ON(p != rq->curr);
3820 WARN_ON(p->pi_blocked_on);
3824 trace_sched_pi_setprio(p, prio);
3826 prev_class = p->sched_class;
3828 running = task_current(rq, p);
3830 dequeue_task(rq, p, 0);
3832 p->sched_class->put_prev_task(rq, p);
3835 p->sched_class = &rt_sched_class;
3837 p->sched_class = &fair_sched_class;
3842 p->sched_class->set_curr_task(rq);
3844 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3846 check_class_changed(rq, p, prev_class, oldprio);
3848 __task_rq_unlock(rq);
3851 void set_user_nice(struct task_struct *p, long nice)
3853 int old_prio, delta, on_rq;
3854 unsigned long flags;
3857 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3860 * We have to be careful, if called from sys_setpriority(),
3861 * the task might be in the middle of scheduling on another CPU.
3863 rq = task_rq_lock(p, &flags);
3865 * The RT priorities are set via sched_setscheduler(), but we still
3866 * allow the 'normal' nice value to be set - but as expected
3867 * it wont have any effect on scheduling until the task is
3868 * SCHED_FIFO/SCHED_RR:
3870 if (task_has_rt_policy(p)) {
3871 p->static_prio = NICE_TO_PRIO(nice);
3876 dequeue_task(rq, p, 0);
3878 p->static_prio = NICE_TO_PRIO(nice);
3881 p->prio = effective_prio(p);
3882 delta = p->prio - old_prio;
3885 enqueue_task(rq, p, 0);
3887 * If the task increased its priority or is running and
3888 * lowered its priority, then reschedule its CPU:
3890 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3891 resched_task(rq->curr);
3894 task_rq_unlock(rq, p, &flags);
3896 EXPORT_SYMBOL(set_user_nice);
3899 * can_nice - check if a task can reduce its nice value
3903 int can_nice(const struct task_struct *p, const int nice)
3905 /* convert nice value [19,-20] to rlimit style value [1,40] */
3906 int nice_rlim = 20 - nice;
3908 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3909 capable(CAP_SYS_NICE));
3912 #ifdef __ARCH_WANT_SYS_NICE
3915 * sys_nice - change the priority of the current process.
3916 * @increment: priority increment
3918 * sys_setpriority is a more generic, but much slower function that
3919 * does similar things.
3921 SYSCALL_DEFINE1(nice, int, increment)
3926 * Setpriority might change our priority at the same moment.
3927 * We don't have to worry. Conceptually one call occurs first
3928 * and we have a single winner.
3930 if (increment < -40)
3935 nice = TASK_NICE(current) + increment;
3941 if (increment < 0 && !can_nice(current, nice))
3944 retval = security_task_setnice(current, nice);
3948 set_user_nice(current, nice);
3955 * task_prio - return the priority value of a given task.
3956 * @p: the task in question.
3958 * This is the priority value as seen by users in /proc.
3959 * RT tasks are offset by -200. Normal tasks are centered
3960 * around 0, value goes from -16 to +15.
3962 int task_prio(const struct task_struct *p)
3964 return p->prio - MAX_RT_PRIO;
3968 * task_nice - return the nice value of a given task.
3969 * @p: the task in question.
3971 int task_nice(const struct task_struct *p)
3973 return TASK_NICE(p);
3975 EXPORT_SYMBOL(task_nice);
3978 * idle_cpu - is a given cpu idle currently?
3979 * @cpu: the processor in question.
3981 int idle_cpu(int cpu)
3983 struct rq *rq = cpu_rq(cpu);
3985 if (rq->curr != rq->idle)
3992 if (!llist_empty(&rq->wake_list))
4000 * idle_task - return the idle task for a given cpu.
4001 * @cpu: the processor in question.
4003 struct task_struct *idle_task(int cpu)
4005 return cpu_rq(cpu)->idle;
4009 * find_process_by_pid - find a process with a matching PID value.
4010 * @pid: the pid in question.
4012 static struct task_struct *find_process_by_pid(pid_t pid)
4014 return pid ? find_task_by_vpid(pid) : current;
4017 /* Actually do priority change: must hold rq lock. */
4019 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4022 p->rt_priority = prio;
4023 p->normal_prio = normal_prio(p);
4024 /* we are holding p->pi_lock already */
4025 p->prio = rt_mutex_getprio(p);
4026 if (rt_prio(p->prio))
4027 p->sched_class = &rt_sched_class;
4029 p->sched_class = &fair_sched_class;
4034 * check the target process has a UID that matches the current process's
4036 static bool check_same_owner(struct task_struct *p)
4038 const struct cred *cred = current_cred(), *pcred;
4042 pcred = __task_cred(p);
4043 if (cred->user->user_ns == pcred->user->user_ns)
4044 match = (cred->euid == pcred->euid ||
4045 cred->euid == pcred->uid);
4052 static int __sched_setscheduler(struct task_struct *p, int policy,
4053 const struct sched_param *param, bool user)
4055 int retval, oldprio, oldpolicy = -1, on_rq, running;
4056 unsigned long flags;
4057 const struct sched_class *prev_class;
4061 /* may grab non-irq protected spin_locks */
4062 BUG_ON(in_interrupt());
4064 /* double check policy once rq lock held */
4066 reset_on_fork = p->sched_reset_on_fork;
4067 policy = oldpolicy = p->policy;
4069 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4070 policy &= ~SCHED_RESET_ON_FORK;
4072 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4073 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4074 policy != SCHED_IDLE)
4079 * Valid priorities for SCHED_FIFO and SCHED_RR are
4080 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4081 * SCHED_BATCH and SCHED_IDLE is 0.
4083 if (param->sched_priority < 0 ||
4084 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4085 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4087 if (rt_policy(policy) != (param->sched_priority != 0))
4091 * Allow unprivileged RT tasks to decrease priority:
4093 if (user && !capable(CAP_SYS_NICE)) {
4094 if (rt_policy(policy)) {
4095 unsigned long rlim_rtprio =
4096 task_rlimit(p, RLIMIT_RTPRIO);
4098 /* can't set/change the rt policy */
4099 if (policy != p->policy && !rlim_rtprio)
4102 /* can't increase priority */
4103 if (param->sched_priority > p->rt_priority &&
4104 param->sched_priority > rlim_rtprio)
4109 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4110 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4112 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4113 if (!can_nice(p, TASK_NICE(p)))
4117 /* can't change other user's priorities */
4118 if (!check_same_owner(p))
4121 /* Normal users shall not reset the sched_reset_on_fork flag */
4122 if (p->sched_reset_on_fork && !reset_on_fork)
4127 retval = security_task_setscheduler(p);
4133 * make sure no PI-waiters arrive (or leave) while we are
4134 * changing the priority of the task:
4136 * To be able to change p->policy safely, the appropriate
4137 * runqueue lock must be held.
4139 rq = task_rq_lock(p, &flags);
4142 * Changing the policy of the stop threads its a very bad idea
4144 if (p == rq->stop) {
4145 task_rq_unlock(rq, p, &flags);
4150 * If not changing anything there's no need to proceed further:
4152 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4153 param->sched_priority == p->rt_priority))) {
4155 __task_rq_unlock(rq);
4156 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4160 #ifdef CONFIG_RT_GROUP_SCHED
4163 * Do not allow realtime tasks into groups that have no runtime
4166 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4167 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4168 !task_group_is_autogroup(task_group(p))) {
4169 task_rq_unlock(rq, p, &flags);
4175 /* recheck policy now with rq lock held */
4176 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4177 policy = oldpolicy = -1;
4178 task_rq_unlock(rq, p, &flags);
4182 running = task_current(rq, p);
4184 dequeue_task(rq, p, 0);
4186 p->sched_class->put_prev_task(rq, p);
4188 p->sched_reset_on_fork = reset_on_fork;
4191 prev_class = p->sched_class;
4192 __setscheduler(rq, p, policy, param->sched_priority);
4195 p->sched_class->set_curr_task(rq);
4197 enqueue_task(rq, p, 0);
4199 check_class_changed(rq, p, prev_class, oldprio);
4200 task_rq_unlock(rq, p, &flags);
4202 rt_mutex_adjust_pi(p);
4208 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4209 * @p: the task in question.
4210 * @policy: new policy.
4211 * @param: structure containing the new RT priority.
4213 * NOTE that the task may be already dead.
4215 int sched_setscheduler(struct task_struct *p, int policy,
4216 const struct sched_param *param)
4218 return __sched_setscheduler(p, policy, param, true);
4220 EXPORT_SYMBOL_GPL(sched_setscheduler);
4223 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4224 * @p: the task in question.
4225 * @policy: new policy.
4226 * @param: structure containing the new RT priority.
4228 * Just like sched_setscheduler, only don't bother checking if the
4229 * current context has permission. For example, this is needed in
4230 * stop_machine(): we create temporary high priority worker threads,
4231 * but our caller might not have that capability.
4233 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4234 const struct sched_param *param)
4236 return __sched_setscheduler(p, policy, param, false);
4240 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4242 struct sched_param lparam;
4243 struct task_struct *p;
4246 if (!param || pid < 0)
4248 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4253 p = find_process_by_pid(pid);
4255 retval = sched_setscheduler(p, policy, &lparam);
4262 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4263 * @pid: the pid in question.
4264 * @policy: new policy.
4265 * @param: structure containing the new RT priority.
4267 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4268 struct sched_param __user *, param)
4270 /* negative values for policy are not valid */
4274 return do_sched_setscheduler(pid, policy, param);
4278 * sys_sched_setparam - set/change the RT priority of a thread
4279 * @pid: the pid in question.
4280 * @param: structure containing the new RT priority.
4282 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4284 return do_sched_setscheduler(pid, -1, param);
4288 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4289 * @pid: the pid in question.
4291 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4293 struct task_struct *p;
4301 p = find_process_by_pid(pid);
4303 retval = security_task_getscheduler(p);
4306 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4313 * sys_sched_getparam - get the RT priority of a thread
4314 * @pid: the pid in question.
4315 * @param: structure containing the RT priority.
4317 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4319 struct sched_param lp;
4320 struct task_struct *p;
4323 if (!param || pid < 0)
4327 p = find_process_by_pid(pid);
4332 retval = security_task_getscheduler(p);
4336 lp.sched_priority = p->rt_priority;
4340 * This one might sleep, we cannot do it with a spinlock held ...
4342 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4351 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4353 cpumask_var_t cpus_allowed, new_mask;
4354 struct task_struct *p;
4360 p = find_process_by_pid(pid);
4367 /* Prevent p going away */
4371 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4375 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4377 goto out_free_cpus_allowed;
4380 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4383 retval = security_task_setscheduler(p);
4387 cpuset_cpus_allowed(p, cpus_allowed);
4388 cpumask_and(new_mask, in_mask, cpus_allowed);
4390 retval = set_cpus_allowed_ptr(p, new_mask);
4393 cpuset_cpus_allowed(p, cpus_allowed);
4394 if (!cpumask_subset(new_mask, cpus_allowed)) {
4396 * We must have raced with a concurrent cpuset
4397 * update. Just reset the cpus_allowed to the
4398 * cpuset's cpus_allowed
4400 cpumask_copy(new_mask, cpus_allowed);
4405 free_cpumask_var(new_mask);
4406 out_free_cpus_allowed:
4407 free_cpumask_var(cpus_allowed);
4414 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4415 struct cpumask *new_mask)
4417 if (len < cpumask_size())
4418 cpumask_clear(new_mask);
4419 else if (len > cpumask_size())
4420 len = cpumask_size();
4422 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4426 * sys_sched_setaffinity - set the cpu affinity of a process
4427 * @pid: pid of the process
4428 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4429 * @user_mask_ptr: user-space pointer to the new cpu mask
4431 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4432 unsigned long __user *, user_mask_ptr)
4434 cpumask_var_t new_mask;
4437 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4440 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4442 retval = sched_setaffinity(pid, new_mask);
4443 free_cpumask_var(new_mask);
4447 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4449 struct task_struct *p;
4450 unsigned long flags;
4457 p = find_process_by_pid(pid);
4461 retval = security_task_getscheduler(p);
4465 raw_spin_lock_irqsave(&p->pi_lock, flags);
4466 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4467 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4477 * sys_sched_getaffinity - get the cpu affinity of a process
4478 * @pid: pid of the process
4479 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4480 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4482 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4483 unsigned long __user *, user_mask_ptr)
4488 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4490 if (len & (sizeof(unsigned long)-1))
4493 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4496 ret = sched_getaffinity(pid, mask);
4498 size_t retlen = min_t(size_t, len, cpumask_size());
4500 if (copy_to_user(user_mask_ptr, mask, retlen))
4505 free_cpumask_var(mask);
4511 * sys_sched_yield - yield the current processor to other threads.
4513 * This function yields the current CPU to other tasks. If there are no
4514 * other threads running on this CPU then this function will return.
4516 SYSCALL_DEFINE0(sched_yield)
4518 struct rq *rq = this_rq_lock();
4520 schedstat_inc(rq, yld_count);
4521 current->sched_class->yield_task(rq);
4524 * Since we are going to call schedule() anyway, there's
4525 * no need to preempt or enable interrupts:
4527 __release(rq->lock);
4528 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4529 do_raw_spin_unlock(&rq->lock);
4530 sched_preempt_enable_no_resched();
4537 static inline int should_resched(void)
4539 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4542 static void __cond_resched(void)
4544 add_preempt_count(PREEMPT_ACTIVE);
4546 sub_preempt_count(PREEMPT_ACTIVE);
4549 int __sched _cond_resched(void)
4551 if (should_resched()) {
4557 EXPORT_SYMBOL(_cond_resched);
4560 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4561 * call schedule, and on return reacquire the lock.
4563 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4564 * operations here to prevent schedule() from being called twice (once via
4565 * spin_unlock(), once by hand).
4567 int __cond_resched_lock(spinlock_t *lock)
4569 int resched = should_resched();
4572 lockdep_assert_held(lock);
4574 if (spin_needbreak(lock) || resched) {
4585 EXPORT_SYMBOL(__cond_resched_lock);
4587 int __sched __cond_resched_softirq(void)
4589 BUG_ON(!in_softirq());
4591 if (should_resched()) {
4599 EXPORT_SYMBOL(__cond_resched_softirq);
4602 * yield - yield the current processor to other threads.
4604 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4606 * The scheduler is at all times free to pick the calling task as the most
4607 * eligible task to run, if removing the yield() call from your code breaks
4608 * it, its already broken.
4610 * Typical broken usage is:
4615 * where one assumes that yield() will let 'the other' process run that will
4616 * make event true. If the current task is a SCHED_FIFO task that will never
4617 * happen. Never use yield() as a progress guarantee!!
4619 * If you want to use yield() to wait for something, use wait_event().
4620 * If you want to use yield() to be 'nice' for others, use cond_resched().
4621 * If you still want to use yield(), do not!
4623 void __sched yield(void)
4625 set_current_state(TASK_RUNNING);
4628 EXPORT_SYMBOL(yield);
4631 * yield_to - yield the current processor to another thread in
4632 * your thread group, or accelerate that thread toward the
4633 * processor it's on.
4635 * @preempt: whether task preemption is allowed or not
4637 * It's the caller's job to ensure that the target task struct
4638 * can't go away on us before we can do any checks.
4640 * Returns true if we indeed boosted the target task.
4642 bool __sched yield_to(struct task_struct *p, bool preempt)
4644 struct task_struct *curr = current;
4645 struct rq *rq, *p_rq;
4646 unsigned long flags;
4649 local_irq_save(flags);
4654 double_rq_lock(rq, p_rq);
4655 while (task_rq(p) != p_rq) {
4656 double_rq_unlock(rq, p_rq);
4660 if (!curr->sched_class->yield_to_task)
4663 if (curr->sched_class != p->sched_class)
4666 if (task_running(p_rq, p) || p->state)
4669 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4671 schedstat_inc(rq, yld_count);
4673 * Make p's CPU reschedule; pick_next_entity takes care of
4676 if (preempt && rq != p_rq)
4677 resched_task(p_rq->curr);
4680 * We might have set it in task_yield_fair(), but are
4681 * not going to schedule(), so don't want to skip
4684 rq->skip_clock_update = 0;
4688 double_rq_unlock(rq, p_rq);
4689 local_irq_restore(flags);
4696 EXPORT_SYMBOL_GPL(yield_to);
4699 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4700 * that process accounting knows that this is a task in IO wait state.
4702 void __sched io_schedule(void)
4704 struct rq *rq = raw_rq();
4706 delayacct_blkio_start();
4707 atomic_inc(&rq->nr_iowait);
4708 blk_flush_plug(current);
4709 current->in_iowait = 1;
4711 current->in_iowait = 0;
4712 atomic_dec(&rq->nr_iowait);
4713 delayacct_blkio_end();
4715 EXPORT_SYMBOL(io_schedule);
4717 long __sched io_schedule_timeout(long timeout)
4719 struct rq *rq = raw_rq();
4722 delayacct_blkio_start();
4723 atomic_inc(&rq->nr_iowait);
4724 blk_flush_plug(current);
4725 current->in_iowait = 1;
4726 ret = schedule_timeout(timeout);
4727 current->in_iowait = 0;
4728 atomic_dec(&rq->nr_iowait);
4729 delayacct_blkio_end();
4734 * sys_sched_get_priority_max - return maximum RT priority.
4735 * @policy: scheduling class.
4737 * this syscall returns the maximum rt_priority that can be used
4738 * by a given scheduling class.
4740 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4747 ret = MAX_USER_RT_PRIO-1;
4759 * sys_sched_get_priority_min - return minimum RT priority.
4760 * @policy: scheduling class.
4762 * this syscall returns the minimum rt_priority that can be used
4763 * by a given scheduling class.
4765 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4783 * sys_sched_rr_get_interval - return the default timeslice of a process.
4784 * @pid: pid of the process.
4785 * @interval: userspace pointer to the timeslice value.
4787 * this syscall writes the default timeslice value of a given process
4788 * into the user-space timespec buffer. A value of '0' means infinity.
4790 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4791 struct timespec __user *, interval)
4793 struct task_struct *p;
4794 unsigned int time_slice;
4795 unsigned long flags;
4805 p = find_process_by_pid(pid);
4809 retval = security_task_getscheduler(p);
4813 rq = task_rq_lock(p, &flags);
4814 time_slice = p->sched_class->get_rr_interval(rq, p);
4815 task_rq_unlock(rq, p, &flags);
4818 jiffies_to_timespec(time_slice, &t);
4819 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4827 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4829 void sched_show_task(struct task_struct *p)
4831 unsigned long free = 0;
4834 state = p->state ? __ffs(p->state) + 1 : 0;
4835 printk(KERN_INFO "%-15.15s %c", p->comm,
4836 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4837 #if BITS_PER_LONG == 32
4838 if (state == TASK_RUNNING)
4839 printk(KERN_CONT " running ");
4841 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4843 if (state == TASK_RUNNING)
4844 printk(KERN_CONT " running task ");
4846 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4848 #ifdef CONFIG_DEBUG_STACK_USAGE
4849 free = stack_not_used(p);
4851 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4852 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4853 (unsigned long)task_thread_info(p)->flags);
4855 show_stack(p, NULL);
4858 void show_state_filter(unsigned long state_filter)
4860 struct task_struct *g, *p;
4862 #if BITS_PER_LONG == 32
4864 " task PC stack pid father\n");
4867 " task PC stack pid father\n");
4870 do_each_thread(g, p) {
4872 * reset the NMI-timeout, listing all files on a slow
4873 * console might take a lot of time:
4875 touch_nmi_watchdog();
4876 if (!state_filter || (p->state & state_filter))
4878 } while_each_thread(g, p);
4880 touch_all_softlockup_watchdogs();
4882 #ifdef CONFIG_SCHED_DEBUG
4883 sysrq_sched_debug_show();
4887 * Only show locks if all tasks are dumped:
4890 debug_show_all_locks();
4893 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4895 idle->sched_class = &idle_sched_class;
4899 * init_idle - set up an idle thread for a given CPU
4900 * @idle: task in question
4901 * @cpu: cpu the idle task belongs to
4903 * NOTE: this function does not set the idle thread's NEED_RESCHED
4904 * flag, to make booting more robust.
4906 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4908 struct rq *rq = cpu_rq(cpu);
4909 unsigned long flags;
4911 raw_spin_lock_irqsave(&rq->lock, flags);
4914 idle->state = TASK_RUNNING;
4915 idle->se.exec_start = sched_clock();
4917 do_set_cpus_allowed(idle, cpumask_of(cpu));
4919 * We're having a chicken and egg problem, even though we are
4920 * holding rq->lock, the cpu isn't yet set to this cpu so the
4921 * lockdep check in task_group() will fail.
4923 * Similar case to sched_fork(). / Alternatively we could
4924 * use task_rq_lock() here and obtain the other rq->lock.
4929 __set_task_cpu(idle, cpu);
4932 rq->curr = rq->idle = idle;
4933 #if defined(CONFIG_SMP)
4936 raw_spin_unlock_irqrestore(&rq->lock, flags);
4938 /* Set the preempt count _outside_ the spinlocks! */
4939 task_thread_info(idle)->preempt_count = 0;
4942 * The idle tasks have their own, simple scheduling class:
4944 idle->sched_class = &idle_sched_class;
4945 ftrace_graph_init_idle_task(idle, cpu);
4946 #if defined(CONFIG_SMP)
4947 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4952 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4954 if (p->sched_class && p->sched_class->set_cpus_allowed)
4955 p->sched_class->set_cpus_allowed(p, new_mask);
4957 cpumask_copy(&p->cpus_allowed, new_mask);
4958 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4962 * This is how migration works:
4964 * 1) we invoke migration_cpu_stop() on the target CPU using
4966 * 2) stopper starts to run (implicitly forcing the migrated thread
4968 * 3) it checks whether the migrated task is still in the wrong runqueue.
4969 * 4) if it's in the wrong runqueue then the migration thread removes
4970 * it and puts it into the right queue.
4971 * 5) stopper completes and stop_one_cpu() returns and the migration
4976 * Change a given task's CPU affinity. Migrate the thread to a
4977 * proper CPU and schedule it away if the CPU it's executing on
4978 * is removed from the allowed bitmask.
4980 * NOTE: the caller must have a valid reference to the task, the
4981 * task must not exit() & deallocate itself prematurely. The
4982 * call is not atomic; no spinlocks may be held.
4984 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4986 unsigned long flags;
4988 unsigned int dest_cpu;
4991 rq = task_rq_lock(p, &flags);
4993 if (cpumask_equal(&p->cpus_allowed, new_mask))
4996 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5001 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5006 do_set_cpus_allowed(p, new_mask);
5008 /* Can the task run on the task's current CPU? If so, we're done */
5009 if (cpumask_test_cpu(task_cpu(p), new_mask))
5012 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5014 struct migration_arg arg = { p, dest_cpu };
5015 /* Need help from migration thread: drop lock and wait. */
5016 task_rq_unlock(rq, p, &flags);
5017 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5018 tlb_migrate_finish(p->mm);
5022 task_rq_unlock(rq, p, &flags);
5026 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5029 * Move (not current) task off this cpu, onto dest cpu. We're doing
5030 * this because either it can't run here any more (set_cpus_allowed()
5031 * away from this CPU, or CPU going down), or because we're
5032 * attempting to rebalance this task on exec (sched_exec).
5034 * So we race with normal scheduler movements, but that's OK, as long
5035 * as the task is no longer on this CPU.
5037 * Returns non-zero if task was successfully migrated.
5039 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5041 struct rq *rq_dest, *rq_src;
5044 if (unlikely(!cpu_active(dest_cpu)))
5047 rq_src = cpu_rq(src_cpu);
5048 rq_dest = cpu_rq(dest_cpu);
5050 raw_spin_lock(&p->pi_lock);
5051 double_rq_lock(rq_src, rq_dest);
5052 /* Already moved. */
5053 if (task_cpu(p) != src_cpu)
5055 /* Affinity changed (again). */
5056 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5060 * If we're not on a rq, the next wake-up will ensure we're
5064 dequeue_task(rq_src, p, 0);
5065 set_task_cpu(p, dest_cpu);
5066 enqueue_task(rq_dest, p, 0);
5067 check_preempt_curr(rq_dest, p, 0);
5072 double_rq_unlock(rq_src, rq_dest);
5073 raw_spin_unlock(&p->pi_lock);
5078 * migration_cpu_stop - this will be executed by a highprio stopper thread
5079 * and performs thread migration by bumping thread off CPU then
5080 * 'pushing' onto another runqueue.
5082 static int migration_cpu_stop(void *data)
5084 struct migration_arg *arg = data;
5087 * The original target cpu might have gone down and we might
5088 * be on another cpu but it doesn't matter.
5090 local_irq_disable();
5091 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5096 #ifdef CONFIG_HOTPLUG_CPU
5099 * Ensures that the idle task is using init_mm right before its cpu goes
5102 void idle_task_exit(void)
5104 struct mm_struct *mm = current->active_mm;
5106 BUG_ON(cpu_online(smp_processor_id()));
5109 switch_mm(mm, &init_mm, current);
5114 * While a dead CPU has no uninterruptible tasks queued at this point,
5115 * it might still have a nonzero ->nr_uninterruptible counter, because
5116 * for performance reasons the counter is not stricly tracking tasks to
5117 * their home CPUs. So we just add the counter to another CPU's counter,
5118 * to keep the global sum constant after CPU-down:
5120 static void migrate_nr_uninterruptible(struct rq *rq_src)
5122 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5124 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5125 rq_src->nr_uninterruptible = 0;
5129 * remove the tasks which were accounted by rq from calc_load_tasks.
5131 static void calc_global_load_remove(struct rq *rq)
5133 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5134 rq->calc_load_active = 0;
5138 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5139 * try_to_wake_up()->select_task_rq().
5141 * Called with rq->lock held even though we'er in stop_machine() and
5142 * there's no concurrency possible, we hold the required locks anyway
5143 * because of lock validation efforts.
5145 static void migrate_tasks(unsigned int dead_cpu)
5147 struct rq *rq = cpu_rq(dead_cpu);
5148 struct task_struct *next, *stop = rq->stop;
5152 * Fudge the rq selection such that the below task selection loop
5153 * doesn't get stuck on the currently eligible stop task.
5155 * We're currently inside stop_machine() and the rq is either stuck
5156 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5157 * either way we should never end up calling schedule() until we're
5162 /* Ensure any throttled groups are reachable by pick_next_task */
5163 unthrottle_offline_cfs_rqs(rq);
5167 * There's this thread running, bail when that's the only
5170 if (rq->nr_running == 1)
5173 next = pick_next_task(rq);
5175 next->sched_class->put_prev_task(rq, next);
5177 /* Find suitable destination for @next, with force if needed. */
5178 dest_cpu = select_fallback_rq(dead_cpu, next);
5179 raw_spin_unlock(&rq->lock);
5181 __migrate_task(next, dead_cpu, dest_cpu);
5183 raw_spin_lock(&rq->lock);
5189 #endif /* CONFIG_HOTPLUG_CPU */
5191 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5193 static struct ctl_table sd_ctl_dir[] = {
5195 .procname = "sched_domain",
5201 static struct ctl_table sd_ctl_root[] = {
5203 .procname = "kernel",
5205 .child = sd_ctl_dir,
5210 static struct ctl_table *sd_alloc_ctl_entry(int n)
5212 struct ctl_table *entry =
5213 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5218 static void sd_free_ctl_entry(struct ctl_table **tablep)
5220 struct ctl_table *entry;
5223 * In the intermediate directories, both the child directory and
5224 * procname are dynamically allocated and could fail but the mode
5225 * will always be set. In the lowest directory the names are
5226 * static strings and all have proc handlers.
5228 for (entry = *tablep; entry->mode; entry++) {
5230 sd_free_ctl_entry(&entry->child);
5231 if (entry->proc_handler == NULL)
5232 kfree(entry->procname);
5240 set_table_entry(struct ctl_table *entry,
5241 const char *procname, void *data, int maxlen,
5242 umode_t mode, proc_handler *proc_handler)
5244 entry->procname = procname;
5246 entry->maxlen = maxlen;
5248 entry->proc_handler = proc_handler;
5251 static struct ctl_table *
5252 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5254 struct ctl_table *table = sd_alloc_ctl_entry(13);
5259 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5260 sizeof(long), 0644, proc_doulongvec_minmax);
5261 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5262 sizeof(long), 0644, proc_doulongvec_minmax);
5263 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5264 sizeof(int), 0644, proc_dointvec_minmax);
5265 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5266 sizeof(int), 0644, proc_dointvec_minmax);
5267 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5268 sizeof(int), 0644, proc_dointvec_minmax);
5269 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5270 sizeof(int), 0644, proc_dointvec_minmax);
5271 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5272 sizeof(int), 0644, proc_dointvec_minmax);
5273 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5274 sizeof(int), 0644, proc_dointvec_minmax);
5275 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5276 sizeof(int), 0644, proc_dointvec_minmax);
5277 set_table_entry(&table[9], "cache_nice_tries",
5278 &sd->cache_nice_tries,
5279 sizeof(int), 0644, proc_dointvec_minmax);
5280 set_table_entry(&table[10], "flags", &sd->flags,
5281 sizeof(int), 0644, proc_dointvec_minmax);
5282 set_table_entry(&table[11], "name", sd->name,
5283 CORENAME_MAX_SIZE, 0444, proc_dostring);
5284 /* &table[12] is terminator */
5289 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5291 struct ctl_table *entry, *table;
5292 struct sched_domain *sd;
5293 int domain_num = 0, i;
5296 for_each_domain(cpu, sd)
5298 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5303 for_each_domain(cpu, sd) {
5304 snprintf(buf, 32, "domain%d", i);
5305 entry->procname = kstrdup(buf, GFP_KERNEL);
5307 entry->child = sd_alloc_ctl_domain_table(sd);
5314 static struct ctl_table_header *sd_sysctl_header;
5315 static void register_sched_domain_sysctl(void)
5317 int i, cpu_num = num_possible_cpus();
5318 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5321 WARN_ON(sd_ctl_dir[0].child);
5322 sd_ctl_dir[0].child = entry;
5327 for_each_possible_cpu(i) {
5328 snprintf(buf, 32, "cpu%d", i);
5329 entry->procname = kstrdup(buf, GFP_KERNEL);
5331 entry->child = sd_alloc_ctl_cpu_table(i);
5335 WARN_ON(sd_sysctl_header);
5336 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5339 /* may be called multiple times per register */
5340 static void unregister_sched_domain_sysctl(void)
5342 if (sd_sysctl_header)
5343 unregister_sysctl_table(sd_sysctl_header);
5344 sd_sysctl_header = NULL;
5345 if (sd_ctl_dir[0].child)
5346 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5349 static void register_sched_domain_sysctl(void)
5352 static void unregister_sched_domain_sysctl(void)
5357 static void set_rq_online(struct rq *rq)
5360 const struct sched_class *class;
5362 cpumask_set_cpu(rq->cpu, rq->rd->online);
5365 for_each_class(class) {
5366 if (class->rq_online)
5367 class->rq_online(rq);
5372 static void set_rq_offline(struct rq *rq)
5375 const struct sched_class *class;
5377 for_each_class(class) {
5378 if (class->rq_offline)
5379 class->rq_offline(rq);
5382 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5388 * migration_call - callback that gets triggered when a CPU is added.
5389 * Here we can start up the necessary migration thread for the new CPU.
5391 static int __cpuinit
5392 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5394 int cpu = (long)hcpu;
5395 unsigned long flags;
5396 struct rq *rq = cpu_rq(cpu);
5398 switch (action & ~CPU_TASKS_FROZEN) {
5400 case CPU_UP_PREPARE:
5401 rq->calc_load_update = calc_load_update;
5405 /* Update our root-domain */
5406 raw_spin_lock_irqsave(&rq->lock, flags);
5408 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5412 raw_spin_unlock_irqrestore(&rq->lock, flags);
5415 #ifdef CONFIG_HOTPLUG_CPU
5417 sched_ttwu_pending();
5418 /* Update our root-domain */
5419 raw_spin_lock_irqsave(&rq->lock, flags);
5421 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5425 BUG_ON(rq->nr_running != 1); /* the migration thread */
5426 raw_spin_unlock_irqrestore(&rq->lock, flags);
5428 migrate_nr_uninterruptible(rq);
5429 calc_global_load_remove(rq);
5434 update_max_interval();
5440 * Register at high priority so that task migration (migrate_all_tasks)
5441 * happens before everything else. This has to be lower priority than
5442 * the notifier in the perf_event subsystem, though.
5444 static struct notifier_block __cpuinitdata migration_notifier = {
5445 .notifier_call = migration_call,
5446 .priority = CPU_PRI_MIGRATION,
5449 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5450 unsigned long action, void *hcpu)
5452 switch (action & ~CPU_TASKS_FROZEN) {
5454 case CPU_DOWN_FAILED:
5455 set_cpu_active((long)hcpu, true);
5462 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5463 unsigned long action, void *hcpu)
5465 switch (action & ~CPU_TASKS_FROZEN) {
5466 case CPU_DOWN_PREPARE:
5467 set_cpu_active((long)hcpu, false);
5474 static int __init migration_init(void)
5476 void *cpu = (void *)(long)smp_processor_id();
5479 /* Initialize migration for the boot CPU */
5480 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5481 BUG_ON(err == NOTIFY_BAD);
5482 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5483 register_cpu_notifier(&migration_notifier);
5485 /* Register cpu active notifiers */
5486 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5487 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5491 early_initcall(migration_init);
5496 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5498 #ifdef CONFIG_SCHED_DEBUG
5500 static __read_mostly int sched_domain_debug_enabled;
5502 static int __init sched_domain_debug_setup(char *str)
5504 sched_domain_debug_enabled = 1;
5508 early_param("sched_debug", sched_domain_debug_setup);
5510 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5511 struct cpumask *groupmask)
5513 struct sched_group *group = sd->groups;
5516 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5517 cpumask_clear(groupmask);
5519 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5521 if (!(sd->flags & SD_LOAD_BALANCE)) {
5522 printk("does not load-balance\n");
5524 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5529 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5531 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5532 printk(KERN_ERR "ERROR: domain->span does not contain "
5535 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5536 printk(KERN_ERR "ERROR: domain->groups does not contain"
5540 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5544 printk(KERN_ERR "ERROR: group is NULL\n");
5548 if (!group->sgp->power) {
5549 printk(KERN_CONT "\n");
5550 printk(KERN_ERR "ERROR: domain->cpu_power not "
5555 if (!cpumask_weight(sched_group_cpus(group))) {
5556 printk(KERN_CONT "\n");
5557 printk(KERN_ERR "ERROR: empty group\n");
5561 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5562 printk(KERN_CONT "\n");
5563 printk(KERN_ERR "ERROR: repeated CPUs\n");
5567 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5569 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5571 printk(KERN_CONT " %s", str);
5572 if (group->sgp->power != SCHED_POWER_SCALE) {
5573 printk(KERN_CONT " (cpu_power = %d)",
5577 group = group->next;
5578 } while (group != sd->groups);
5579 printk(KERN_CONT "\n");
5581 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5582 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5585 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5586 printk(KERN_ERR "ERROR: parent span is not a superset "
5587 "of domain->span\n");
5591 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5595 if (!sched_domain_debug_enabled)
5599 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5603 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5606 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5614 #else /* !CONFIG_SCHED_DEBUG */
5615 # define sched_domain_debug(sd, cpu) do { } while (0)
5616 #endif /* CONFIG_SCHED_DEBUG */
5618 static int sd_degenerate(struct sched_domain *sd)
5620 if (cpumask_weight(sched_domain_span(sd)) == 1)
5623 /* Following flags need at least 2 groups */
5624 if (sd->flags & (SD_LOAD_BALANCE |
5625 SD_BALANCE_NEWIDLE |
5629 SD_SHARE_PKG_RESOURCES)) {
5630 if (sd->groups != sd->groups->next)
5634 /* Following flags don't use groups */
5635 if (sd->flags & (SD_WAKE_AFFINE))
5642 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5644 unsigned long cflags = sd->flags, pflags = parent->flags;
5646 if (sd_degenerate(parent))
5649 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5652 /* Flags needing groups don't count if only 1 group in parent */
5653 if (parent->groups == parent->groups->next) {
5654 pflags &= ~(SD_LOAD_BALANCE |
5655 SD_BALANCE_NEWIDLE |
5659 SD_SHARE_PKG_RESOURCES);
5660 if (nr_node_ids == 1)
5661 pflags &= ~SD_SERIALIZE;
5663 if (~cflags & pflags)
5669 static void free_rootdomain(struct rcu_head *rcu)
5671 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5673 cpupri_cleanup(&rd->cpupri);
5674 free_cpumask_var(rd->rto_mask);
5675 free_cpumask_var(rd->online);
5676 free_cpumask_var(rd->span);
5680 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5682 struct root_domain *old_rd = NULL;
5683 unsigned long flags;
5685 raw_spin_lock_irqsave(&rq->lock, flags);
5690 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5693 cpumask_clear_cpu(rq->cpu, old_rd->span);
5696 * If we dont want to free the old_rt yet then
5697 * set old_rd to NULL to skip the freeing later
5700 if (!atomic_dec_and_test(&old_rd->refcount))
5704 atomic_inc(&rd->refcount);
5707 cpumask_set_cpu(rq->cpu, rd->span);
5708 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5711 raw_spin_unlock_irqrestore(&rq->lock, flags);
5714 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5717 static int init_rootdomain(struct root_domain *rd)
5719 memset(rd, 0, sizeof(*rd));
5721 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5723 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5725 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5728 if (cpupri_init(&rd->cpupri) != 0)
5733 free_cpumask_var(rd->rto_mask);
5735 free_cpumask_var(rd->online);
5737 free_cpumask_var(rd->span);
5743 * By default the system creates a single root-domain with all cpus as
5744 * members (mimicking the global state we have today).
5746 struct root_domain def_root_domain;
5748 static void init_defrootdomain(void)
5750 init_rootdomain(&def_root_domain);
5752 atomic_set(&def_root_domain.refcount, 1);
5755 static struct root_domain *alloc_rootdomain(void)
5757 struct root_domain *rd;
5759 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5763 if (init_rootdomain(rd) != 0) {
5771 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5773 struct sched_group *tmp, *first;
5782 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5787 } while (sg != first);
5790 static void free_sched_domain(struct rcu_head *rcu)
5792 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5795 * If its an overlapping domain it has private groups, iterate and
5798 if (sd->flags & SD_OVERLAP) {
5799 free_sched_groups(sd->groups, 1);
5800 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5801 kfree(sd->groups->sgp);
5807 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5809 call_rcu(&sd->rcu, free_sched_domain);
5812 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5814 for (; sd; sd = sd->parent)
5815 destroy_sched_domain(sd, cpu);
5819 * Keep a special pointer to the highest sched_domain that has
5820 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5821 * allows us to avoid some pointer chasing select_idle_sibling().
5823 * Also keep a unique ID per domain (we use the first cpu number in
5824 * the cpumask of the domain), this allows us to quickly tell if
5825 * two cpus are in the same cache domain, see cpus_share_cache().
5827 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5828 DEFINE_PER_CPU(int, sd_llc_id);
5830 static void update_top_cache_domain(int cpu)
5832 struct sched_domain *sd;
5835 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5837 id = cpumask_first(sched_domain_span(sd));
5839 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5840 per_cpu(sd_llc_id, cpu) = id;
5844 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5845 * hold the hotplug lock.
5848 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5850 struct rq *rq = cpu_rq(cpu);
5851 struct sched_domain *tmp;
5853 /* Remove the sched domains which do not contribute to scheduling. */
5854 for (tmp = sd; tmp; ) {
5855 struct sched_domain *parent = tmp->parent;
5859 if (sd_parent_degenerate(tmp, parent)) {
5860 tmp->parent = parent->parent;
5862 parent->parent->child = tmp;
5863 destroy_sched_domain(parent, cpu);
5868 if (sd && sd_degenerate(sd)) {
5871 destroy_sched_domain(tmp, cpu);
5876 sched_domain_debug(sd, cpu);
5878 rq_attach_root(rq, rd);
5880 rcu_assign_pointer(rq->sd, sd);
5881 destroy_sched_domains(tmp, cpu);
5883 update_top_cache_domain(cpu);
5886 /* cpus with isolated domains */
5887 static cpumask_var_t cpu_isolated_map;
5889 /* Setup the mask of cpus configured for isolated domains */
5890 static int __init isolated_cpu_setup(char *str)
5892 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5893 cpulist_parse(str, cpu_isolated_map);
5897 __setup("isolcpus=", isolated_cpu_setup);
5902 * find_next_best_node - find the next node to include in a sched_domain
5903 * @node: node whose sched_domain we're building
5904 * @used_nodes: nodes already in the sched_domain
5906 * Find the next node to include in a given scheduling domain. Simply
5907 * finds the closest node not already in the @used_nodes map.
5909 * Should use nodemask_t.
5911 static int find_next_best_node(int node, nodemask_t *used_nodes)
5913 int i, n, val, min_val, best_node = -1;
5917 for (i = 0; i < nr_node_ids; i++) {
5918 /* Start at @node */
5919 n = (node + i) % nr_node_ids;
5921 if (!nr_cpus_node(n))
5924 /* Skip already used nodes */
5925 if (node_isset(n, *used_nodes))
5928 /* Simple min distance search */
5929 val = node_distance(node, n);
5931 if (val < min_val) {
5937 if (best_node != -1)
5938 node_set(best_node, *used_nodes);
5943 * sched_domain_node_span - get a cpumask for a node's sched_domain
5944 * @node: node whose cpumask we're constructing
5945 * @span: resulting cpumask
5947 * Given a node, construct a good cpumask for its sched_domain to span. It
5948 * should be one that prevents unnecessary balancing, but also spreads tasks
5951 static void sched_domain_node_span(int node, struct cpumask *span)
5953 nodemask_t used_nodes;
5956 cpumask_clear(span);
5957 nodes_clear(used_nodes);
5959 cpumask_or(span, span, cpumask_of_node(node));
5960 node_set(node, used_nodes);
5962 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5963 int next_node = find_next_best_node(node, &used_nodes);
5966 cpumask_or(span, span, cpumask_of_node(next_node));
5970 static const struct cpumask *cpu_node_mask(int cpu)
5972 lockdep_assert_held(&sched_domains_mutex);
5974 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5976 return sched_domains_tmpmask;
5979 static const struct cpumask *cpu_allnodes_mask(int cpu)
5981 return cpu_possible_mask;
5983 #endif /* CONFIG_NUMA */
5985 static const struct cpumask *cpu_cpu_mask(int cpu)
5987 return cpumask_of_node(cpu_to_node(cpu));
5990 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5993 struct sched_domain **__percpu sd;
5994 struct sched_group **__percpu sg;
5995 struct sched_group_power **__percpu sgp;
5999 struct sched_domain ** __percpu sd;
6000 struct root_domain *rd;
6010 struct sched_domain_topology_level;
6012 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6013 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6015 #define SDTL_OVERLAP 0x01
6017 struct sched_domain_topology_level {
6018 sched_domain_init_f init;
6019 sched_domain_mask_f mask;
6021 struct sd_data data;
6025 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6027 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6028 const struct cpumask *span = sched_domain_span(sd);
6029 struct cpumask *covered = sched_domains_tmpmask;
6030 struct sd_data *sdd = sd->private;
6031 struct sched_domain *child;
6034 cpumask_clear(covered);
6036 for_each_cpu(i, span) {
6037 struct cpumask *sg_span;
6039 if (cpumask_test_cpu(i, covered))
6042 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6043 GFP_KERNEL, cpu_to_node(cpu));
6048 sg_span = sched_group_cpus(sg);
6050 child = *per_cpu_ptr(sdd->sd, i);
6052 child = child->child;
6053 cpumask_copy(sg_span, sched_domain_span(child));
6055 cpumask_set_cpu(i, sg_span);
6057 cpumask_or(covered, covered, sg_span);
6059 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
6060 atomic_inc(&sg->sgp->ref);
6062 if (cpumask_test_cpu(cpu, sg_span))
6072 sd->groups = groups;
6077 free_sched_groups(first, 0);
6082 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6084 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6085 struct sched_domain *child = sd->child;
6088 cpu = cpumask_first(sched_domain_span(child));
6091 *sg = *per_cpu_ptr(sdd->sg, cpu);
6092 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6093 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6100 * build_sched_groups will build a circular linked list of the groups
6101 * covered by the given span, and will set each group's ->cpumask correctly,
6102 * and ->cpu_power to 0.
6104 * Assumes the sched_domain tree is fully constructed
6107 build_sched_groups(struct sched_domain *sd, int cpu)
6109 struct sched_group *first = NULL, *last = NULL;
6110 struct sd_data *sdd = sd->private;
6111 const struct cpumask *span = sched_domain_span(sd);
6112 struct cpumask *covered;
6115 get_group(cpu, sdd, &sd->groups);
6116 atomic_inc(&sd->groups->ref);
6118 if (cpu != cpumask_first(sched_domain_span(sd)))
6121 lockdep_assert_held(&sched_domains_mutex);
6122 covered = sched_domains_tmpmask;
6124 cpumask_clear(covered);
6126 for_each_cpu(i, span) {
6127 struct sched_group *sg;
6128 int group = get_group(i, sdd, &sg);
6131 if (cpumask_test_cpu(i, covered))
6134 cpumask_clear(sched_group_cpus(sg));
6137 for_each_cpu(j, span) {
6138 if (get_group(j, sdd, NULL) != group)
6141 cpumask_set_cpu(j, covered);
6142 cpumask_set_cpu(j, sched_group_cpus(sg));
6157 * Initialize sched groups cpu_power.
6159 * cpu_power indicates the capacity of sched group, which is used while
6160 * distributing the load between different sched groups in a sched domain.
6161 * Typically cpu_power for all the groups in a sched domain will be same unless
6162 * there are asymmetries in the topology. If there are asymmetries, group
6163 * having more cpu_power will pickup more load compared to the group having
6166 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6168 struct sched_group *sg = sd->groups;
6170 WARN_ON(!sd || !sg);
6173 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6175 } while (sg != sd->groups);
6177 if (cpu != group_first_cpu(sg))
6180 update_group_power(sd, cpu);
6181 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6184 int __weak arch_sd_sibling_asym_packing(void)
6186 return 0*SD_ASYM_PACKING;
6190 * Initializers for schedule domains
6191 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6194 #ifdef CONFIG_SCHED_DEBUG
6195 # define SD_INIT_NAME(sd, type) sd->name = #type
6197 # define SD_INIT_NAME(sd, type) do { } while (0)
6200 #define SD_INIT_FUNC(type) \
6201 static noinline struct sched_domain * \
6202 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6204 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6205 *sd = SD_##type##_INIT; \
6206 SD_INIT_NAME(sd, type); \
6207 sd->private = &tl->data; \
6213 SD_INIT_FUNC(ALLNODES)
6216 #ifdef CONFIG_SCHED_SMT
6217 SD_INIT_FUNC(SIBLING)
6219 #ifdef CONFIG_SCHED_MC
6222 #ifdef CONFIG_SCHED_BOOK
6226 static int default_relax_domain_level = -1;
6227 int sched_domain_level_max;
6229 static int __init setup_relax_domain_level(char *str)
6233 val = simple_strtoul(str, NULL, 0);
6234 if (val < sched_domain_level_max)
6235 default_relax_domain_level = val;
6239 __setup("relax_domain_level=", setup_relax_domain_level);
6241 static void set_domain_attribute(struct sched_domain *sd,
6242 struct sched_domain_attr *attr)
6246 if (!attr || attr->relax_domain_level < 0) {
6247 if (default_relax_domain_level < 0)
6250 request = default_relax_domain_level;
6252 request = attr->relax_domain_level;
6253 if (request < sd->level) {
6254 /* turn off idle balance on this domain */
6255 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6257 /* turn on idle balance on this domain */
6258 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6262 static void __sdt_free(const struct cpumask *cpu_map);
6263 static int __sdt_alloc(const struct cpumask *cpu_map);
6265 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6266 const struct cpumask *cpu_map)
6270 if (!atomic_read(&d->rd->refcount))
6271 free_rootdomain(&d->rd->rcu); /* fall through */
6273 free_percpu(d->sd); /* fall through */
6275 __sdt_free(cpu_map); /* fall through */
6281 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6282 const struct cpumask *cpu_map)
6284 memset(d, 0, sizeof(*d));
6286 if (__sdt_alloc(cpu_map))
6287 return sa_sd_storage;
6288 d->sd = alloc_percpu(struct sched_domain *);
6290 return sa_sd_storage;
6291 d->rd = alloc_rootdomain();
6294 return sa_rootdomain;
6298 * NULL the sd_data elements we've used to build the sched_domain and
6299 * sched_group structure so that the subsequent __free_domain_allocs()
6300 * will not free the data we're using.
6302 static void claim_allocations(int cpu, struct sched_domain *sd)
6304 struct sd_data *sdd = sd->private;
6306 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6307 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6309 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6310 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6312 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6313 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6316 #ifdef CONFIG_SCHED_SMT
6317 static const struct cpumask *cpu_smt_mask(int cpu)
6319 return topology_thread_cpumask(cpu);
6324 * Topology list, bottom-up.
6326 static struct sched_domain_topology_level default_topology[] = {
6327 #ifdef CONFIG_SCHED_SMT
6328 { sd_init_SIBLING, cpu_smt_mask, },
6330 #ifdef CONFIG_SCHED_MC
6331 { sd_init_MC, cpu_coregroup_mask, },
6333 #ifdef CONFIG_SCHED_BOOK
6334 { sd_init_BOOK, cpu_book_mask, },
6336 { sd_init_CPU, cpu_cpu_mask, },
6338 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6339 { sd_init_ALLNODES, cpu_allnodes_mask, },
6344 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6346 static int __sdt_alloc(const struct cpumask *cpu_map)
6348 struct sched_domain_topology_level *tl;
6351 for (tl = sched_domain_topology; tl->init; tl++) {
6352 struct sd_data *sdd = &tl->data;
6354 sdd->sd = alloc_percpu(struct sched_domain *);
6358 sdd->sg = alloc_percpu(struct sched_group *);
6362 sdd->sgp = alloc_percpu(struct sched_group_power *);
6366 for_each_cpu(j, cpu_map) {
6367 struct sched_domain *sd;
6368 struct sched_group *sg;
6369 struct sched_group_power *sgp;
6371 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6372 GFP_KERNEL, cpu_to_node(j));
6376 *per_cpu_ptr(sdd->sd, j) = sd;
6378 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6379 GFP_KERNEL, cpu_to_node(j));
6383 *per_cpu_ptr(sdd->sg, j) = sg;
6385 sgp = kzalloc_node(sizeof(struct sched_group_power),
6386 GFP_KERNEL, cpu_to_node(j));
6390 *per_cpu_ptr(sdd->sgp, j) = sgp;
6397 static void __sdt_free(const struct cpumask *cpu_map)
6399 struct sched_domain_topology_level *tl;
6402 for (tl = sched_domain_topology; tl->init; tl++) {
6403 struct sd_data *sdd = &tl->data;
6405 for_each_cpu(j, cpu_map) {
6406 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6407 if (sd && (sd->flags & SD_OVERLAP))
6408 free_sched_groups(sd->groups, 0);
6409 kfree(*per_cpu_ptr(sdd->sd, j));
6410 kfree(*per_cpu_ptr(sdd->sg, j));
6411 kfree(*per_cpu_ptr(sdd->sgp, j));
6413 free_percpu(sdd->sd);
6414 free_percpu(sdd->sg);
6415 free_percpu(sdd->sgp);
6419 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6420 struct s_data *d, const struct cpumask *cpu_map,
6421 struct sched_domain_attr *attr, struct sched_domain *child,
6424 struct sched_domain *sd = tl->init(tl, cpu);
6428 set_domain_attribute(sd, attr);
6429 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6431 sd->level = child->level + 1;
6432 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6441 * Build sched domains for a given set of cpus and attach the sched domains
6442 * to the individual cpus
6444 static int build_sched_domains(const struct cpumask *cpu_map,
6445 struct sched_domain_attr *attr)
6447 enum s_alloc alloc_state = sa_none;
6448 struct sched_domain *sd;
6450 int i, ret = -ENOMEM;
6452 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6453 if (alloc_state != sa_rootdomain)
6456 /* Set up domains for cpus specified by the cpu_map. */
6457 for_each_cpu(i, cpu_map) {
6458 struct sched_domain_topology_level *tl;
6461 for (tl = sched_domain_topology; tl->init; tl++) {
6462 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6463 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6464 sd->flags |= SD_OVERLAP;
6465 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6472 *per_cpu_ptr(d.sd, i) = sd;
6475 /* Build the groups for the domains */
6476 for_each_cpu(i, cpu_map) {
6477 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6478 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6479 if (sd->flags & SD_OVERLAP) {
6480 if (build_overlap_sched_groups(sd, i))
6483 if (build_sched_groups(sd, i))
6489 /* Calculate CPU power for physical packages and nodes */
6490 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6491 if (!cpumask_test_cpu(i, cpu_map))
6494 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6495 claim_allocations(i, sd);
6496 init_sched_groups_power(i, sd);
6500 /* Attach the domains */
6502 for_each_cpu(i, cpu_map) {
6503 sd = *per_cpu_ptr(d.sd, i);
6504 cpu_attach_domain(sd, d.rd, i);
6510 __free_domain_allocs(&d, alloc_state, cpu_map);
6514 static cpumask_var_t *doms_cur; /* current sched domains */
6515 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6516 static struct sched_domain_attr *dattr_cur;
6517 /* attribues of custom domains in 'doms_cur' */
6520 * Special case: If a kmalloc of a doms_cur partition (array of
6521 * cpumask) fails, then fallback to a single sched domain,
6522 * as determined by the single cpumask fallback_doms.
6524 static cpumask_var_t fallback_doms;
6527 * arch_update_cpu_topology lets virtualized architectures update the
6528 * cpu core maps. It is supposed to return 1 if the topology changed
6529 * or 0 if it stayed the same.
6531 int __attribute__((weak)) arch_update_cpu_topology(void)
6536 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6539 cpumask_var_t *doms;
6541 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6544 for (i = 0; i < ndoms; i++) {
6545 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6546 free_sched_domains(doms, i);
6553 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6556 for (i = 0; i < ndoms; i++)
6557 free_cpumask_var(doms[i]);
6562 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6563 * For now this just excludes isolated cpus, but could be used to
6564 * exclude other special cases in the future.
6566 static int init_sched_domains(const struct cpumask *cpu_map)
6570 arch_update_cpu_topology();
6572 doms_cur = alloc_sched_domains(ndoms_cur);
6574 doms_cur = &fallback_doms;
6575 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6577 err = build_sched_domains(doms_cur[0], NULL);
6578 register_sched_domain_sysctl();
6584 * Detach sched domains from a group of cpus specified in cpu_map
6585 * These cpus will now be attached to the NULL domain
6587 static void detach_destroy_domains(const struct cpumask *cpu_map)
6592 for_each_cpu(i, cpu_map)
6593 cpu_attach_domain(NULL, &def_root_domain, i);
6597 /* handle null as "default" */
6598 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6599 struct sched_domain_attr *new, int idx_new)
6601 struct sched_domain_attr tmp;
6608 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6609 new ? (new + idx_new) : &tmp,
6610 sizeof(struct sched_domain_attr));
6614 * Partition sched domains as specified by the 'ndoms_new'
6615 * cpumasks in the array doms_new[] of cpumasks. This compares
6616 * doms_new[] to the current sched domain partitioning, doms_cur[].
6617 * It destroys each deleted domain and builds each new domain.
6619 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6620 * The masks don't intersect (don't overlap.) We should setup one
6621 * sched domain for each mask. CPUs not in any of the cpumasks will
6622 * not be load balanced. If the same cpumask appears both in the
6623 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6626 * The passed in 'doms_new' should be allocated using
6627 * alloc_sched_domains. This routine takes ownership of it and will
6628 * free_sched_domains it when done with it. If the caller failed the
6629 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6630 * and partition_sched_domains() will fallback to the single partition
6631 * 'fallback_doms', it also forces the domains to be rebuilt.
6633 * If doms_new == NULL it will be replaced with cpu_online_mask.
6634 * ndoms_new == 0 is a special case for destroying existing domains,
6635 * and it will not create the default domain.
6637 * Call with hotplug lock held
6639 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6640 struct sched_domain_attr *dattr_new)
6645 mutex_lock(&sched_domains_mutex);
6647 /* always unregister in case we don't destroy any domains */
6648 unregister_sched_domain_sysctl();
6650 /* Let architecture update cpu core mappings. */
6651 new_topology = arch_update_cpu_topology();
6653 n = doms_new ? ndoms_new : 0;
6655 /* Destroy deleted domains */
6656 for (i = 0; i < ndoms_cur; i++) {
6657 for (j = 0; j < n && !new_topology; j++) {
6658 if (cpumask_equal(doms_cur[i], doms_new[j])
6659 && dattrs_equal(dattr_cur, i, dattr_new, j))
6662 /* no match - a current sched domain not in new doms_new[] */
6663 detach_destroy_domains(doms_cur[i]);
6668 if (doms_new == NULL) {
6670 doms_new = &fallback_doms;
6671 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6672 WARN_ON_ONCE(dattr_new);
6675 /* Build new domains */
6676 for (i = 0; i < ndoms_new; i++) {
6677 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6678 if (cpumask_equal(doms_new[i], doms_cur[j])
6679 && dattrs_equal(dattr_new, i, dattr_cur, j))
6682 /* no match - add a new doms_new */
6683 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6688 /* Remember the new sched domains */
6689 if (doms_cur != &fallback_doms)
6690 free_sched_domains(doms_cur, ndoms_cur);
6691 kfree(dattr_cur); /* kfree(NULL) is safe */
6692 doms_cur = doms_new;
6693 dattr_cur = dattr_new;
6694 ndoms_cur = ndoms_new;
6696 register_sched_domain_sysctl();
6698 mutex_unlock(&sched_domains_mutex);
6701 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6702 static void reinit_sched_domains(void)
6706 /* Destroy domains first to force the rebuild */
6707 partition_sched_domains(0, NULL, NULL);
6709 rebuild_sched_domains();
6713 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6715 unsigned int level = 0;
6717 if (sscanf(buf, "%u", &level) != 1)
6721 * level is always be positive so don't check for
6722 * level < POWERSAVINGS_BALANCE_NONE which is 0
6723 * What happens on 0 or 1 byte write,
6724 * need to check for count as well?
6727 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6731 sched_smt_power_savings = level;
6733 sched_mc_power_savings = level;
6735 reinit_sched_domains();
6740 #ifdef CONFIG_SCHED_MC
6741 static ssize_t sched_mc_power_savings_show(struct device *dev,
6742 struct device_attribute *attr,
6745 return sprintf(buf, "%u\n", sched_mc_power_savings);
6747 static ssize_t sched_mc_power_savings_store(struct device *dev,
6748 struct device_attribute *attr,
6749 const char *buf, size_t count)
6751 return sched_power_savings_store(buf, count, 0);
6753 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6754 sched_mc_power_savings_show,
6755 sched_mc_power_savings_store);
6758 #ifdef CONFIG_SCHED_SMT
6759 static ssize_t sched_smt_power_savings_show(struct device *dev,
6760 struct device_attribute *attr,
6763 return sprintf(buf, "%u\n", sched_smt_power_savings);
6765 static ssize_t sched_smt_power_savings_store(struct device *dev,
6766 struct device_attribute *attr,
6767 const char *buf, size_t count)
6769 return sched_power_savings_store(buf, count, 1);
6771 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6772 sched_smt_power_savings_show,
6773 sched_smt_power_savings_store);
6776 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6780 #ifdef CONFIG_SCHED_SMT
6782 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6784 #ifdef CONFIG_SCHED_MC
6785 if (!err && mc_capable())
6786 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6790 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6793 * Update cpusets according to cpu_active mask. If cpusets are
6794 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6795 * around partition_sched_domains().
6797 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6800 switch (action & ~CPU_TASKS_FROZEN) {
6802 case CPU_DOWN_FAILED:
6803 cpuset_update_active_cpus();
6810 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6813 switch (action & ~CPU_TASKS_FROZEN) {
6814 case CPU_DOWN_PREPARE:
6815 cpuset_update_active_cpus();
6822 void __init sched_init_smp(void)
6824 cpumask_var_t non_isolated_cpus;
6826 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6827 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6830 mutex_lock(&sched_domains_mutex);
6831 init_sched_domains(cpu_active_mask);
6832 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6833 if (cpumask_empty(non_isolated_cpus))
6834 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6835 mutex_unlock(&sched_domains_mutex);
6838 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6839 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6841 /* RT runtime code needs to handle some hotplug events */
6842 hotcpu_notifier(update_runtime, 0);
6846 /* Move init over to a non-isolated CPU */
6847 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6849 sched_init_granularity();
6850 free_cpumask_var(non_isolated_cpus);
6852 init_sched_rt_class();
6855 void __init sched_init_smp(void)
6857 sched_init_granularity();
6859 #endif /* CONFIG_SMP */
6861 const_debug unsigned int sysctl_timer_migration = 1;
6863 int in_sched_functions(unsigned long addr)
6865 return in_lock_functions(addr) ||
6866 (addr >= (unsigned long)__sched_text_start
6867 && addr < (unsigned long)__sched_text_end);
6870 #ifdef CONFIG_CGROUP_SCHED
6871 struct task_group root_task_group;
6874 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6876 void __init sched_init(void)
6879 unsigned long alloc_size = 0, ptr;
6881 #ifdef CONFIG_FAIR_GROUP_SCHED
6882 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6884 #ifdef CONFIG_RT_GROUP_SCHED
6885 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6887 #ifdef CONFIG_CPUMASK_OFFSTACK
6888 alloc_size += num_possible_cpus() * cpumask_size();
6891 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6893 #ifdef CONFIG_FAIR_GROUP_SCHED
6894 root_task_group.se = (struct sched_entity **)ptr;
6895 ptr += nr_cpu_ids * sizeof(void **);
6897 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6898 ptr += nr_cpu_ids * sizeof(void **);
6900 #endif /* CONFIG_FAIR_GROUP_SCHED */
6901 #ifdef CONFIG_RT_GROUP_SCHED
6902 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6903 ptr += nr_cpu_ids * sizeof(void **);
6905 root_task_group.rt_rq = (struct rt_rq **)ptr;
6906 ptr += nr_cpu_ids * sizeof(void **);
6908 #endif /* CONFIG_RT_GROUP_SCHED */
6909 #ifdef CONFIG_CPUMASK_OFFSTACK
6910 for_each_possible_cpu(i) {
6911 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6912 ptr += cpumask_size();
6914 #endif /* CONFIG_CPUMASK_OFFSTACK */
6918 init_defrootdomain();
6921 init_rt_bandwidth(&def_rt_bandwidth,
6922 global_rt_period(), global_rt_runtime());
6924 #ifdef CONFIG_RT_GROUP_SCHED
6925 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6926 global_rt_period(), global_rt_runtime());
6927 #endif /* CONFIG_RT_GROUP_SCHED */
6929 #ifdef CONFIG_CGROUP_SCHED
6930 list_add(&root_task_group.list, &task_groups);
6931 INIT_LIST_HEAD(&root_task_group.children);
6932 INIT_LIST_HEAD(&root_task_group.siblings);
6933 autogroup_init(&init_task);
6935 #endif /* CONFIG_CGROUP_SCHED */
6937 #ifdef CONFIG_CGROUP_CPUACCT
6938 root_cpuacct.cpustat = &kernel_cpustat;
6939 root_cpuacct.cpuusage = alloc_percpu(u64);
6940 /* Too early, not expected to fail */
6941 BUG_ON(!root_cpuacct.cpuusage);
6943 for_each_possible_cpu(i) {
6947 raw_spin_lock_init(&rq->lock);
6949 rq->calc_load_active = 0;
6950 rq->calc_load_update = jiffies + LOAD_FREQ;
6951 init_cfs_rq(&rq->cfs);
6952 init_rt_rq(&rq->rt, rq);
6953 #ifdef CONFIG_FAIR_GROUP_SCHED
6954 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6955 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6957 * How much cpu bandwidth does root_task_group get?
6959 * In case of task-groups formed thr' the cgroup filesystem, it
6960 * gets 100% of the cpu resources in the system. This overall
6961 * system cpu resource is divided among the tasks of
6962 * root_task_group and its child task-groups in a fair manner,
6963 * based on each entity's (task or task-group's) weight
6964 * (se->load.weight).
6966 * In other words, if root_task_group has 10 tasks of weight
6967 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6968 * then A0's share of the cpu resource is:
6970 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6972 * We achieve this by letting root_task_group's tasks sit
6973 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6975 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6976 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6977 #endif /* CONFIG_FAIR_GROUP_SCHED */
6979 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6980 #ifdef CONFIG_RT_GROUP_SCHED
6981 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6982 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6985 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6986 rq->cpu_load[j] = 0;
6988 rq->last_load_update_tick = jiffies;
6993 rq->cpu_power = SCHED_POWER_SCALE;
6994 rq->post_schedule = 0;
6995 rq->active_balance = 0;
6996 rq->next_balance = jiffies;
7001 rq->avg_idle = 2*sysctl_sched_migration_cost;
7003 INIT_LIST_HEAD(&rq->cfs_tasks);
7005 rq_attach_root(rq, &def_root_domain);
7011 atomic_set(&rq->nr_iowait, 0);
7014 set_load_weight(&init_task);
7016 #ifdef CONFIG_PREEMPT_NOTIFIERS
7017 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7020 #ifdef CONFIG_RT_MUTEXES
7021 plist_head_init(&init_task.pi_waiters);
7025 * The boot idle thread does lazy MMU switching as well:
7027 atomic_inc(&init_mm.mm_count);
7028 enter_lazy_tlb(&init_mm, current);
7031 * Make us the idle thread. Technically, schedule() should not be
7032 * called from this thread, however somewhere below it might be,
7033 * but because we are the idle thread, we just pick up running again
7034 * when this runqueue becomes "idle".
7036 init_idle(current, smp_processor_id());
7038 calc_load_update = jiffies + LOAD_FREQ;
7041 * During early bootup we pretend to be a normal task:
7043 current->sched_class = &fair_sched_class;
7046 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7047 /* May be allocated at isolcpus cmdline parse time */
7048 if (cpu_isolated_map == NULL)
7049 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7051 init_sched_fair_class();
7053 scheduler_running = 1;
7056 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7057 static inline int preempt_count_equals(int preempt_offset)
7059 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7061 return (nested == preempt_offset);
7064 void __might_sleep(const char *file, int line, int preempt_offset)
7066 static unsigned long prev_jiffy; /* ratelimiting */
7068 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7069 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7070 system_state != SYSTEM_RUNNING || oops_in_progress)
7072 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7074 prev_jiffy = jiffies;
7077 "BUG: sleeping function called from invalid context at %s:%d\n",
7080 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7081 in_atomic(), irqs_disabled(),
7082 current->pid, current->comm);
7084 debug_show_held_locks(current);
7085 if (irqs_disabled())
7086 print_irqtrace_events(current);
7089 EXPORT_SYMBOL(__might_sleep);
7092 #ifdef CONFIG_MAGIC_SYSRQ
7093 static void normalize_task(struct rq *rq, struct task_struct *p)
7095 const struct sched_class *prev_class = p->sched_class;
7096 int old_prio = p->prio;
7101 dequeue_task(rq, p, 0);
7102 __setscheduler(rq, p, SCHED_NORMAL, 0);
7104 enqueue_task(rq, p, 0);
7105 resched_task(rq->curr);
7108 check_class_changed(rq, p, prev_class, old_prio);
7111 void normalize_rt_tasks(void)
7113 struct task_struct *g, *p;
7114 unsigned long flags;
7117 read_lock_irqsave(&tasklist_lock, flags);
7118 do_each_thread(g, p) {
7120 * Only normalize user tasks:
7125 p->se.exec_start = 0;
7126 #ifdef CONFIG_SCHEDSTATS
7127 p->se.statistics.wait_start = 0;
7128 p->se.statistics.sleep_start = 0;
7129 p->se.statistics.block_start = 0;
7134 * Renice negative nice level userspace
7137 if (TASK_NICE(p) < 0 && p->mm)
7138 set_user_nice(p, 0);
7142 raw_spin_lock(&p->pi_lock);
7143 rq = __task_rq_lock(p);
7145 normalize_task(rq, p);
7147 __task_rq_unlock(rq);
7148 raw_spin_unlock(&p->pi_lock);
7149 } while_each_thread(g, p);
7151 read_unlock_irqrestore(&tasklist_lock, flags);
7154 #endif /* CONFIG_MAGIC_SYSRQ */
7156 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7158 * These functions are only useful for the IA64 MCA handling, or kdb.
7160 * They can only be called when the whole system has been
7161 * stopped - every CPU needs to be quiescent, and no scheduling
7162 * activity can take place. Using them for anything else would
7163 * be a serious bug, and as a result, they aren't even visible
7164 * under any other configuration.
7168 * curr_task - return the current task for a given cpu.
7169 * @cpu: the processor in question.
7171 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7173 struct task_struct *curr_task(int cpu)
7175 return cpu_curr(cpu);
7178 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7182 * set_curr_task - set the current task for a given cpu.
7183 * @cpu: the processor in question.
7184 * @p: the task pointer to set.
7186 * Description: This function must only be used when non-maskable interrupts
7187 * are serviced on a separate stack. It allows the architecture to switch the
7188 * notion of the current task on a cpu in a non-blocking manner. This function
7189 * must be called with all CPU's synchronized, and interrupts disabled, the
7190 * and caller must save the original value of the current task (see
7191 * curr_task() above) and restore that value before reenabling interrupts and
7192 * re-starting the system.
7194 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7196 void set_curr_task(int cpu, struct task_struct *p)
7203 #ifdef CONFIG_CGROUP_SCHED
7204 /* task_group_lock serializes the addition/removal of task groups */
7205 static DEFINE_SPINLOCK(task_group_lock);
7207 static void free_sched_group(struct task_group *tg)
7209 free_fair_sched_group(tg);
7210 free_rt_sched_group(tg);
7215 /* allocate runqueue etc for a new task group */
7216 struct task_group *sched_create_group(struct task_group *parent)
7218 struct task_group *tg;
7219 unsigned long flags;
7221 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7223 return ERR_PTR(-ENOMEM);
7225 if (!alloc_fair_sched_group(tg, parent))
7228 if (!alloc_rt_sched_group(tg, parent))
7231 spin_lock_irqsave(&task_group_lock, flags);
7232 list_add_rcu(&tg->list, &task_groups);
7234 WARN_ON(!parent); /* root should already exist */
7236 tg->parent = parent;
7237 INIT_LIST_HEAD(&tg->children);
7238 list_add_rcu(&tg->siblings, &parent->children);
7239 spin_unlock_irqrestore(&task_group_lock, flags);
7244 free_sched_group(tg);
7245 return ERR_PTR(-ENOMEM);
7248 /* rcu callback to free various structures associated with a task group */
7249 static void free_sched_group_rcu(struct rcu_head *rhp)
7251 /* now it should be safe to free those cfs_rqs */
7252 free_sched_group(container_of(rhp, struct task_group, rcu));
7255 /* Destroy runqueue etc associated with a task group */
7256 void sched_destroy_group(struct task_group *tg)
7258 unsigned long flags;
7261 /* end participation in shares distribution */
7262 for_each_possible_cpu(i)
7263 unregister_fair_sched_group(tg, i);
7265 spin_lock_irqsave(&task_group_lock, flags);
7266 list_del_rcu(&tg->list);
7267 list_del_rcu(&tg->siblings);
7268 spin_unlock_irqrestore(&task_group_lock, flags);
7270 /* wait for possible concurrent references to cfs_rqs complete */
7271 call_rcu(&tg->rcu, free_sched_group_rcu);
7274 /* change task's runqueue when it moves between groups.
7275 * The caller of this function should have put the task in its new group
7276 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7277 * reflect its new group.
7279 void sched_move_task(struct task_struct *tsk)
7282 unsigned long flags;
7285 rq = task_rq_lock(tsk, &flags);
7287 running = task_current(rq, tsk);
7291 dequeue_task(rq, tsk, 0);
7292 if (unlikely(running))
7293 tsk->sched_class->put_prev_task(rq, tsk);
7295 #ifdef CONFIG_FAIR_GROUP_SCHED
7296 if (tsk->sched_class->task_move_group)
7297 tsk->sched_class->task_move_group(tsk, on_rq);
7300 set_task_rq(tsk, task_cpu(tsk));
7302 if (unlikely(running))
7303 tsk->sched_class->set_curr_task(rq);
7305 enqueue_task(rq, tsk, 0);
7307 task_rq_unlock(rq, tsk, &flags);
7309 #endif /* CONFIG_CGROUP_SCHED */
7311 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7312 static unsigned long to_ratio(u64 period, u64 runtime)
7314 if (runtime == RUNTIME_INF)
7317 return div64_u64(runtime << 20, period);
7321 #ifdef CONFIG_RT_GROUP_SCHED
7323 * Ensure that the real time constraints are schedulable.
7325 static DEFINE_MUTEX(rt_constraints_mutex);
7327 /* Must be called with tasklist_lock held */
7328 static inline int tg_has_rt_tasks(struct task_group *tg)
7330 struct task_struct *g, *p;
7332 do_each_thread(g, p) {
7333 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7335 } while_each_thread(g, p);
7340 struct rt_schedulable_data {
7341 struct task_group *tg;
7346 static int tg_rt_schedulable(struct task_group *tg, void *data)
7348 struct rt_schedulable_data *d = data;
7349 struct task_group *child;
7350 unsigned long total, sum = 0;
7351 u64 period, runtime;
7353 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7354 runtime = tg->rt_bandwidth.rt_runtime;
7357 period = d->rt_period;
7358 runtime = d->rt_runtime;
7362 * Cannot have more runtime than the period.
7364 if (runtime > period && runtime != RUNTIME_INF)
7368 * Ensure we don't starve existing RT tasks.
7370 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7373 total = to_ratio(period, runtime);
7376 * Nobody can have more than the global setting allows.
7378 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7382 * The sum of our children's runtime should not exceed our own.
7384 list_for_each_entry_rcu(child, &tg->children, siblings) {
7385 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7386 runtime = child->rt_bandwidth.rt_runtime;
7388 if (child == d->tg) {
7389 period = d->rt_period;
7390 runtime = d->rt_runtime;
7393 sum += to_ratio(period, runtime);
7402 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7406 struct rt_schedulable_data data = {
7408 .rt_period = period,
7409 .rt_runtime = runtime,
7413 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7419 static int tg_set_rt_bandwidth(struct task_group *tg,
7420 u64 rt_period, u64 rt_runtime)
7424 mutex_lock(&rt_constraints_mutex);
7425 read_lock(&tasklist_lock);
7426 err = __rt_schedulable(tg, rt_period, rt_runtime);
7430 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7431 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7432 tg->rt_bandwidth.rt_runtime = rt_runtime;
7434 for_each_possible_cpu(i) {
7435 struct rt_rq *rt_rq = tg->rt_rq[i];
7437 raw_spin_lock(&rt_rq->rt_runtime_lock);
7438 rt_rq->rt_runtime = rt_runtime;
7439 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7441 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7443 read_unlock(&tasklist_lock);
7444 mutex_unlock(&rt_constraints_mutex);
7449 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7451 u64 rt_runtime, rt_period;
7453 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7454 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7455 if (rt_runtime_us < 0)
7456 rt_runtime = RUNTIME_INF;
7458 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7461 long sched_group_rt_runtime(struct task_group *tg)
7465 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7468 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7469 do_div(rt_runtime_us, NSEC_PER_USEC);
7470 return rt_runtime_us;
7473 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7475 u64 rt_runtime, rt_period;
7477 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7478 rt_runtime = tg->rt_bandwidth.rt_runtime;
7483 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7486 long sched_group_rt_period(struct task_group *tg)
7490 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7491 do_div(rt_period_us, NSEC_PER_USEC);
7492 return rt_period_us;
7495 static int sched_rt_global_constraints(void)
7497 u64 runtime, period;
7500 if (sysctl_sched_rt_period <= 0)
7503 runtime = global_rt_runtime();
7504 period = global_rt_period();
7507 * Sanity check on the sysctl variables.
7509 if (runtime > period && runtime != RUNTIME_INF)
7512 mutex_lock(&rt_constraints_mutex);
7513 read_lock(&tasklist_lock);
7514 ret = __rt_schedulable(NULL, 0, 0);
7515 read_unlock(&tasklist_lock);
7516 mutex_unlock(&rt_constraints_mutex);
7521 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7523 /* Don't accept realtime tasks when there is no way for them to run */
7524 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7530 #else /* !CONFIG_RT_GROUP_SCHED */
7531 static int sched_rt_global_constraints(void)
7533 unsigned long flags;
7536 if (sysctl_sched_rt_period <= 0)
7540 * There's always some RT tasks in the root group
7541 * -- migration, kstopmachine etc..
7543 if (sysctl_sched_rt_runtime == 0)
7546 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7547 for_each_possible_cpu(i) {
7548 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7550 raw_spin_lock(&rt_rq->rt_runtime_lock);
7551 rt_rq->rt_runtime = global_rt_runtime();
7552 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7554 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7558 #endif /* CONFIG_RT_GROUP_SCHED */
7560 int sched_rt_handler(struct ctl_table *table, int write,
7561 void __user *buffer, size_t *lenp,
7565 int old_period, old_runtime;
7566 static DEFINE_MUTEX(mutex);
7569 old_period = sysctl_sched_rt_period;
7570 old_runtime = sysctl_sched_rt_runtime;
7572 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7574 if (!ret && write) {
7575 ret = sched_rt_global_constraints();
7577 sysctl_sched_rt_period = old_period;
7578 sysctl_sched_rt_runtime = old_runtime;
7580 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7581 def_rt_bandwidth.rt_period =
7582 ns_to_ktime(global_rt_period());
7585 mutex_unlock(&mutex);
7590 #ifdef CONFIG_CGROUP_SCHED
7592 /* return corresponding task_group object of a cgroup */
7593 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7595 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7596 struct task_group, css);
7599 static struct cgroup_subsys_state *
7600 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7602 struct task_group *tg, *parent;
7604 if (!cgrp->parent) {
7605 /* This is early initialization for the top cgroup */
7606 return &root_task_group.css;
7609 parent = cgroup_tg(cgrp->parent);
7610 tg = sched_create_group(parent);
7612 return ERR_PTR(-ENOMEM);
7618 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7620 struct task_group *tg = cgroup_tg(cgrp);
7622 sched_destroy_group(tg);
7625 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7626 struct cgroup_taskset *tset)
7628 struct task_struct *task;
7630 cgroup_taskset_for_each(task, cgrp, tset) {
7631 #ifdef CONFIG_RT_GROUP_SCHED
7632 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7635 /* We don't support RT-tasks being in separate groups */
7636 if (task->sched_class != &fair_sched_class)
7643 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7644 struct cgroup_taskset *tset)
7646 struct task_struct *task;
7648 cgroup_taskset_for_each(task, cgrp, tset)
7649 sched_move_task(task);
7653 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7654 struct cgroup *old_cgrp, struct task_struct *task)
7657 * cgroup_exit() is called in the copy_process() failure path.
7658 * Ignore this case since the task hasn't ran yet, this avoids
7659 * trying to poke a half freed task state from generic code.
7661 if (!(task->flags & PF_EXITING))
7664 sched_move_task(task);
7667 #ifdef CONFIG_FAIR_GROUP_SCHED
7668 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7671 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7674 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7676 struct task_group *tg = cgroup_tg(cgrp);
7678 return (u64) scale_load_down(tg->shares);
7681 #ifdef CONFIG_CFS_BANDWIDTH
7682 static DEFINE_MUTEX(cfs_constraints_mutex);
7684 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7685 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7687 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7689 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7691 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7692 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7694 if (tg == &root_task_group)
7698 * Ensure we have at some amount of bandwidth every period. This is
7699 * to prevent reaching a state of large arrears when throttled via
7700 * entity_tick() resulting in prolonged exit starvation.
7702 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7706 * Likewise, bound things on the otherside by preventing insane quota
7707 * periods. This also allows us to normalize in computing quota
7710 if (period > max_cfs_quota_period)
7713 mutex_lock(&cfs_constraints_mutex);
7714 ret = __cfs_schedulable(tg, period, quota);
7718 runtime_enabled = quota != RUNTIME_INF;
7719 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7720 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7721 raw_spin_lock_irq(&cfs_b->lock);
7722 cfs_b->period = ns_to_ktime(period);
7723 cfs_b->quota = quota;
7725 __refill_cfs_bandwidth_runtime(cfs_b);
7726 /* restart the period timer (if active) to handle new period expiry */
7727 if (runtime_enabled && cfs_b->timer_active) {
7728 /* force a reprogram */
7729 cfs_b->timer_active = 0;
7730 __start_cfs_bandwidth(cfs_b);
7732 raw_spin_unlock_irq(&cfs_b->lock);
7734 for_each_possible_cpu(i) {
7735 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7736 struct rq *rq = cfs_rq->rq;
7738 raw_spin_lock_irq(&rq->lock);
7739 cfs_rq->runtime_enabled = runtime_enabled;
7740 cfs_rq->runtime_remaining = 0;
7742 if (cfs_rq->throttled)
7743 unthrottle_cfs_rq(cfs_rq);
7744 raw_spin_unlock_irq(&rq->lock);
7747 mutex_unlock(&cfs_constraints_mutex);
7752 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7756 period = ktime_to_ns(tg->cfs_bandwidth.period);
7757 if (cfs_quota_us < 0)
7758 quota = RUNTIME_INF;
7760 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7762 return tg_set_cfs_bandwidth(tg, period, quota);
7765 long tg_get_cfs_quota(struct task_group *tg)
7769 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7772 quota_us = tg->cfs_bandwidth.quota;
7773 do_div(quota_us, NSEC_PER_USEC);
7778 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7782 period = (u64)cfs_period_us * NSEC_PER_USEC;
7783 quota = tg->cfs_bandwidth.quota;
7785 return tg_set_cfs_bandwidth(tg, period, quota);
7788 long tg_get_cfs_period(struct task_group *tg)
7792 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7793 do_div(cfs_period_us, NSEC_PER_USEC);
7795 return cfs_period_us;
7798 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7800 return tg_get_cfs_quota(cgroup_tg(cgrp));
7803 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7806 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7809 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7811 return tg_get_cfs_period(cgroup_tg(cgrp));
7814 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7817 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7820 struct cfs_schedulable_data {
7821 struct task_group *tg;
7826 * normalize group quota/period to be quota/max_period
7827 * note: units are usecs
7829 static u64 normalize_cfs_quota(struct task_group *tg,
7830 struct cfs_schedulable_data *d)
7838 period = tg_get_cfs_period(tg);
7839 quota = tg_get_cfs_quota(tg);
7842 /* note: these should typically be equivalent */
7843 if (quota == RUNTIME_INF || quota == -1)
7846 return to_ratio(period, quota);
7849 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7851 struct cfs_schedulable_data *d = data;
7852 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7853 s64 quota = 0, parent_quota = -1;
7856 quota = RUNTIME_INF;
7858 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7860 quota = normalize_cfs_quota(tg, d);
7861 parent_quota = parent_b->hierarchal_quota;
7864 * ensure max(child_quota) <= parent_quota, inherit when no
7867 if (quota == RUNTIME_INF)
7868 quota = parent_quota;
7869 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7872 cfs_b->hierarchal_quota = quota;
7877 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7880 struct cfs_schedulable_data data = {
7886 if (quota != RUNTIME_INF) {
7887 do_div(data.period, NSEC_PER_USEC);
7888 do_div(data.quota, NSEC_PER_USEC);
7892 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7898 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7899 struct cgroup_map_cb *cb)
7901 struct task_group *tg = cgroup_tg(cgrp);
7902 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7904 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7905 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7906 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7910 #endif /* CONFIG_CFS_BANDWIDTH */
7911 #endif /* CONFIG_FAIR_GROUP_SCHED */
7913 #ifdef CONFIG_RT_GROUP_SCHED
7914 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7917 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7920 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7922 return sched_group_rt_runtime(cgroup_tg(cgrp));
7925 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7928 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7931 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7933 return sched_group_rt_period(cgroup_tg(cgrp));
7935 #endif /* CONFIG_RT_GROUP_SCHED */
7937 static struct cftype cpu_files[] = {
7938 #ifdef CONFIG_FAIR_GROUP_SCHED
7941 .read_u64 = cpu_shares_read_u64,
7942 .write_u64 = cpu_shares_write_u64,
7945 #ifdef CONFIG_CFS_BANDWIDTH
7947 .name = "cfs_quota_us",
7948 .read_s64 = cpu_cfs_quota_read_s64,
7949 .write_s64 = cpu_cfs_quota_write_s64,
7952 .name = "cfs_period_us",
7953 .read_u64 = cpu_cfs_period_read_u64,
7954 .write_u64 = cpu_cfs_period_write_u64,
7958 .read_map = cpu_stats_show,
7961 #ifdef CONFIG_RT_GROUP_SCHED
7963 .name = "rt_runtime_us",
7964 .read_s64 = cpu_rt_runtime_read,
7965 .write_s64 = cpu_rt_runtime_write,
7968 .name = "rt_period_us",
7969 .read_u64 = cpu_rt_period_read_uint,
7970 .write_u64 = cpu_rt_period_write_uint,
7975 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7977 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7980 struct cgroup_subsys cpu_cgroup_subsys = {
7982 .create = cpu_cgroup_create,
7983 .destroy = cpu_cgroup_destroy,
7984 .can_attach = cpu_cgroup_can_attach,
7985 .attach = cpu_cgroup_attach,
7986 .exit = cpu_cgroup_exit,
7987 .populate = cpu_cgroup_populate,
7988 .subsys_id = cpu_cgroup_subsys_id,
7992 #endif /* CONFIG_CGROUP_SCHED */
7994 #ifdef CONFIG_CGROUP_CPUACCT
7997 * CPU accounting code for task groups.
7999 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8000 * (balbir@in.ibm.com).
8003 /* create a new cpu accounting group */
8004 static struct cgroup_subsys_state *cpuacct_create(
8005 struct cgroup_subsys *ss, struct cgroup *cgrp)
8010 return &root_cpuacct.css;
8012 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8016 ca->cpuusage = alloc_percpu(u64);
8020 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8022 goto out_free_cpuusage;
8027 free_percpu(ca->cpuusage);
8031 return ERR_PTR(-ENOMEM);
8034 /* destroy an existing cpu accounting group */
8036 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8038 struct cpuacct *ca = cgroup_ca(cgrp);
8040 free_percpu(ca->cpustat);
8041 free_percpu(ca->cpuusage);
8045 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8047 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8050 #ifndef CONFIG_64BIT
8052 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8054 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8056 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8064 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8066 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8068 #ifndef CONFIG_64BIT
8070 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8072 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8074 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8080 /* return total cpu usage (in nanoseconds) of a group */
8081 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8083 struct cpuacct *ca = cgroup_ca(cgrp);
8084 u64 totalcpuusage = 0;
8087 for_each_present_cpu(i)
8088 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8090 return totalcpuusage;
8093 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8096 struct cpuacct *ca = cgroup_ca(cgrp);
8105 for_each_present_cpu(i)
8106 cpuacct_cpuusage_write(ca, i, 0);
8112 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8115 struct cpuacct *ca = cgroup_ca(cgroup);
8119 for_each_present_cpu(i) {
8120 percpu = cpuacct_cpuusage_read(ca, i);
8121 seq_printf(m, "%llu ", (unsigned long long) percpu);
8123 seq_printf(m, "\n");
8127 static const char *cpuacct_stat_desc[] = {
8128 [CPUACCT_STAT_USER] = "user",
8129 [CPUACCT_STAT_SYSTEM] = "system",
8132 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8133 struct cgroup_map_cb *cb)
8135 struct cpuacct *ca = cgroup_ca(cgrp);
8139 for_each_online_cpu(cpu) {
8140 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8141 val += kcpustat->cpustat[CPUTIME_USER];
8142 val += kcpustat->cpustat[CPUTIME_NICE];
8144 val = cputime64_to_clock_t(val);
8145 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8148 for_each_online_cpu(cpu) {
8149 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8150 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8151 val += kcpustat->cpustat[CPUTIME_IRQ];
8152 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8155 val = cputime64_to_clock_t(val);
8156 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8161 static struct cftype files[] = {
8164 .read_u64 = cpuusage_read,
8165 .write_u64 = cpuusage_write,
8168 .name = "usage_percpu",
8169 .read_seq_string = cpuacct_percpu_seq_read,
8173 .read_map = cpuacct_stats_show,
8177 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8179 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8183 * charge this task's execution time to its accounting group.
8185 * called with rq->lock held.
8187 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8192 if (unlikely(!cpuacct_subsys.active))
8195 cpu = task_cpu(tsk);
8201 for (; ca; ca = parent_ca(ca)) {
8202 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8203 *cpuusage += cputime;
8209 struct cgroup_subsys cpuacct_subsys = {
8211 .create = cpuacct_create,
8212 .destroy = cpuacct_destroy,
8213 .populate = cpuacct_populate,
8214 .subsys_id = cpuacct_subsys_id,
8216 #endif /* CONFIG_CGROUP_CPUACCT */