4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 ktime_t soft, hard, now;
97 if (hrtimer_active(period_timer))
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
111 DEFINE_MUTEX(sched_domains_mutex);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 void update_rq_clock(struct rq *rq)
120 if (rq->skip_clock_update > 0)
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 update_rq_clock_task(rq, delta);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug unsigned int sysctl_sched_features =
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static __read_mostly char *sched_feat_names[] = {
146 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
197 sched_feat_write(struct file *filp, const char __user *ubuf,
198 size_t cnt, loff_t *ppos)
208 if (copy_from_user(&buf, ubuf, cnt))
214 if (strncmp(cmp, "NO_", 3) == 0) {
219 for (i = 0; i < __SCHED_FEAT_NR; i++) {
220 if (strcmp(cmp, sched_feat_names[i]) == 0) {
222 sysctl_sched_features &= ~(1UL << i);
223 sched_feat_disable(i);
225 sysctl_sched_features |= (1UL << i);
226 sched_feat_enable(i);
232 if (i == __SCHED_FEAT_NR)
240 static int sched_feat_open(struct inode *inode, struct file *filp)
242 return single_open(filp, sched_feat_show, NULL);
245 static const struct file_operations sched_feat_fops = {
246 .open = sched_feat_open,
247 .write = sched_feat_write,
250 .release = single_release,
253 static __init int sched_init_debug(void)
255 debugfs_create_file("sched_features", 0644, NULL, NULL,
260 late_initcall(sched_init_debug);
261 #endif /* CONFIG_SCHED_DEBUG */
264 * Number of tasks to iterate in a single balance run.
265 * Limited because this is done with IRQs disabled.
267 const_debug unsigned int sysctl_sched_nr_migrate = 32;
270 * period over which we average the RT time consumption, measured
275 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
278 * period over which we measure -rt task cpu usage in us.
281 unsigned int sysctl_sched_rt_period = 1000000;
283 __read_mostly int scheduler_running;
286 * part of the period that we allow rt tasks to run in us.
289 int sysctl_sched_rt_runtime = 950000;
294 * __task_rq_lock - lock the rq @p resides on.
296 static inline struct rq *__task_rq_lock(struct task_struct *p)
301 lockdep_assert_held(&p->pi_lock);
305 raw_spin_lock(&rq->lock);
306 if (likely(rq == task_rq(p)))
308 raw_spin_unlock(&rq->lock);
313 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
315 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
316 __acquires(p->pi_lock)
322 raw_spin_lock_irqsave(&p->pi_lock, *flags);
324 raw_spin_lock(&rq->lock);
325 if (likely(rq == task_rq(p)))
327 raw_spin_unlock(&rq->lock);
328 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
332 static void __task_rq_unlock(struct rq *rq)
335 raw_spin_unlock(&rq->lock);
339 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
341 __releases(p->pi_lock)
343 raw_spin_unlock(&rq->lock);
344 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
348 * this_rq_lock - lock this runqueue and disable interrupts.
350 static struct rq *this_rq_lock(void)
357 raw_spin_lock(&rq->lock);
362 #ifdef CONFIG_SCHED_HRTICK
364 * Use HR-timers to deliver accurate preemption points.
366 * Its all a bit involved since we cannot program an hrt while holding the
367 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
370 * When we get rescheduled we reprogram the hrtick_timer outside of the
374 static void hrtick_clear(struct rq *rq)
376 if (hrtimer_active(&rq->hrtick_timer))
377 hrtimer_cancel(&rq->hrtick_timer);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
386 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
390 raw_spin_lock(&rq->lock);
392 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393 raw_spin_unlock(&rq->lock);
395 return HRTIMER_NORESTART;
400 * called from hardirq (IPI) context
402 static void __hrtick_start(void *arg)
406 raw_spin_lock(&rq->lock);
407 hrtimer_restart(&rq->hrtick_timer);
408 rq->hrtick_csd_pending = 0;
409 raw_spin_unlock(&rq->lock);
413 * Called to set the hrtick timer state.
415 * called with rq->lock held and irqs disabled
417 void hrtick_start(struct rq *rq, u64 delay)
419 struct hrtimer *timer = &rq->hrtick_timer;
420 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
422 hrtimer_set_expires(timer, time);
424 if (rq == this_rq()) {
425 hrtimer_restart(timer);
426 } else if (!rq->hrtick_csd_pending) {
427 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
428 rq->hrtick_csd_pending = 1;
433 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
435 int cpu = (int)(long)hcpu;
438 case CPU_UP_CANCELED:
439 case CPU_UP_CANCELED_FROZEN:
440 case CPU_DOWN_PREPARE:
441 case CPU_DOWN_PREPARE_FROZEN:
443 case CPU_DEAD_FROZEN:
444 hrtick_clear(cpu_rq(cpu));
451 static __init void init_hrtick(void)
453 hotcpu_notifier(hotplug_hrtick, 0);
457 * Called to set the hrtick timer state.
459 * called with rq->lock held and irqs disabled
461 void hrtick_start(struct rq *rq, u64 delay)
463 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
464 HRTIMER_MODE_REL_PINNED, 0);
467 static inline void init_hrtick(void)
470 #endif /* CONFIG_SMP */
472 static void init_rq_hrtick(struct rq *rq)
475 rq->hrtick_csd_pending = 0;
477 rq->hrtick_csd.flags = 0;
478 rq->hrtick_csd.func = __hrtick_start;
479 rq->hrtick_csd.info = rq;
482 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
483 rq->hrtick_timer.function = hrtick;
485 #else /* CONFIG_SCHED_HRTICK */
486 static inline void hrtick_clear(struct rq *rq)
490 static inline void init_rq_hrtick(struct rq *rq)
494 static inline void init_hrtick(void)
497 #endif /* CONFIG_SCHED_HRTICK */
500 * resched_task - mark a task 'to be rescheduled now'.
502 * On UP this means the setting of the need_resched flag, on SMP it
503 * might also involve a cross-CPU call to trigger the scheduler on
508 #ifndef tsk_is_polling
509 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
512 void resched_task(struct task_struct *p)
516 assert_raw_spin_locked(&task_rq(p)->lock);
518 if (test_tsk_need_resched(p))
521 set_tsk_need_resched(p);
524 if (cpu == smp_processor_id())
527 /* NEED_RESCHED must be visible before we test polling */
529 if (!tsk_is_polling(p))
530 smp_send_reschedule(cpu);
533 void resched_cpu(int cpu)
535 struct rq *rq = cpu_rq(cpu);
538 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
540 resched_task(cpu_curr(cpu));
541 raw_spin_unlock_irqrestore(&rq->lock, flags);
546 * In the semi idle case, use the nearest busy cpu for migrating timers
547 * from an idle cpu. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle cpu will add more delays to the timers than intended
551 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int cpu = smp_processor_id();
557 struct sched_domain *sd;
560 for_each_domain(cpu, sd) {
561 for_each_cpu(i, sched_domain_span(sd)) {
573 * When add_timer_on() enqueues a timer into the timer wheel of an
574 * idle CPU then this timer might expire before the next timer event
575 * which is scheduled to wake up that CPU. In case of a completely
576 * idle system the next event might even be infinite time into the
577 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
578 * leaves the inner idle loop so the newly added timer is taken into
579 * account when the CPU goes back to idle and evaluates the timer
580 * wheel for the next timer event.
582 void wake_up_idle_cpu(int cpu)
584 struct rq *rq = cpu_rq(cpu);
586 if (cpu == smp_processor_id())
590 * This is safe, as this function is called with the timer
591 * wheel base lock of (cpu) held. When the CPU is on the way
592 * to idle and has not yet set rq->curr to idle then it will
593 * be serialized on the timer wheel base lock and take the new
594 * timer into account automatically.
596 if (rq->curr != rq->idle)
600 * We can set TIF_RESCHED on the idle task of the other CPU
601 * lockless. The worst case is that the other CPU runs the
602 * idle task through an additional NOOP schedule()
604 set_tsk_need_resched(rq->idle);
606 /* NEED_RESCHED must be visible before we test polling */
608 if (!tsk_is_polling(rq->idle))
609 smp_send_reschedule(cpu);
612 static inline bool got_nohz_idle_kick(void)
614 int cpu = smp_processor_id();
615 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
618 #else /* CONFIG_NO_HZ */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ */
627 void sched_avg_update(struct rq *rq)
629 s64 period = sched_avg_period();
631 while ((s64)(rq->clock - rq->age_stamp) > period) {
633 * Inline assembly required to prevent the compiler
634 * optimising this loop into a divmod call.
635 * See __iter_div_u64_rem() for another example of this.
637 asm("" : "+rm" (rq->age_stamp));
638 rq->age_stamp += period;
643 #else /* !CONFIG_SMP */
644 void resched_task(struct task_struct *p)
646 assert_raw_spin_locked(&task_rq(p)->lock);
647 set_tsk_need_resched(p);
649 #endif /* CONFIG_SMP */
651 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
652 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
654 * Iterate task_group tree rooted at *from, calling @down when first entering a
655 * node and @up when leaving it for the final time.
657 * Caller must hold rcu_lock or sufficient equivalent.
659 int walk_tg_tree_from(struct task_group *from,
660 tg_visitor down, tg_visitor up, void *data)
662 struct task_group *parent, *child;
668 ret = (*down)(parent, data);
671 list_for_each_entry_rcu(child, &parent->children, siblings) {
678 ret = (*up)(parent, data);
679 if (ret || parent == from)
683 parent = parent->parent;
690 int tg_nop(struct task_group *tg, void *data)
696 static void set_load_weight(struct task_struct *p)
698 int prio = p->static_prio - MAX_RT_PRIO;
699 struct load_weight *load = &p->se.load;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p->policy == SCHED_IDLE) {
705 load->weight = scale_load(WEIGHT_IDLEPRIO);
706 load->inv_weight = WMULT_IDLEPRIO;
710 load->weight = scale_load(prio_to_weight[prio]);
711 load->inv_weight = prio_to_wmult[prio];
714 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
717 sched_info_queued(p);
718 p->sched_class->enqueue_task(rq, p, flags);
721 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
724 sched_info_dequeued(p);
725 p->sched_class->dequeue_task(rq, p, flags);
728 void activate_task(struct rq *rq, struct task_struct *p, int flags)
730 if (task_contributes_to_load(p))
731 rq->nr_uninterruptible--;
733 enqueue_task(rq, p, flags);
736 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible++;
741 dequeue_task(rq, p, flags);
744 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
747 * There are no locks covering percpu hardirq/softirq time.
748 * They are only modified in account_system_vtime, on corresponding CPU
749 * with interrupts disabled. So, writes are safe.
750 * They are read and saved off onto struct rq in update_rq_clock().
751 * This may result in other CPU reading this CPU's irq time and can
752 * race with irq/account_system_vtime on this CPU. We would either get old
753 * or new value with a side effect of accounting a slice of irq time to wrong
754 * task when irq is in progress while we read rq->clock. That is a worthy
755 * compromise in place of having locks on each irq in account_system_time.
757 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
758 static DEFINE_PER_CPU(u64, cpu_softirq_time);
760 static DEFINE_PER_CPU(u64, irq_start_time);
761 static int sched_clock_irqtime;
763 void enable_sched_clock_irqtime(void)
765 sched_clock_irqtime = 1;
768 void disable_sched_clock_irqtime(void)
770 sched_clock_irqtime = 0;
774 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
776 static inline void irq_time_write_begin(void)
778 __this_cpu_inc(irq_time_seq.sequence);
782 static inline void irq_time_write_end(void)
785 __this_cpu_inc(irq_time_seq.sequence);
788 static inline u64 irq_time_read(int cpu)
794 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
795 irq_time = per_cpu(cpu_softirq_time, cpu) +
796 per_cpu(cpu_hardirq_time, cpu);
797 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
801 #else /* CONFIG_64BIT */
802 static inline void irq_time_write_begin(void)
806 static inline void irq_time_write_end(void)
810 static inline u64 irq_time_read(int cpu)
812 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
814 #endif /* CONFIG_64BIT */
817 * Called before incrementing preempt_count on {soft,}irq_enter
818 * and before decrementing preempt_count on {soft,}irq_exit.
820 void account_system_vtime(struct task_struct *curr)
826 if (!sched_clock_irqtime)
829 local_irq_save(flags);
831 cpu = smp_processor_id();
832 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
833 __this_cpu_add(irq_start_time, delta);
835 irq_time_write_begin();
837 * We do not account for softirq time from ksoftirqd here.
838 * We want to continue accounting softirq time to ksoftirqd thread
839 * in that case, so as not to confuse scheduler with a special task
840 * that do not consume any time, but still wants to run.
843 __this_cpu_add(cpu_hardirq_time, delta);
844 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
845 __this_cpu_add(cpu_softirq_time, delta);
847 irq_time_write_end();
848 local_irq_restore(flags);
850 EXPORT_SYMBOL_GPL(account_system_vtime);
852 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
854 #ifdef CONFIG_PARAVIRT
855 static inline u64 steal_ticks(u64 steal)
857 if (unlikely(steal > NSEC_PER_SEC))
858 return div_u64(steal, TICK_NSEC);
860 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
864 static void update_rq_clock_task(struct rq *rq, s64 delta)
867 * In theory, the compile should just see 0 here, and optimize out the call
868 * to sched_rt_avg_update. But I don't trust it...
870 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
871 s64 steal = 0, irq_delta = 0;
873 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
874 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
877 * Since irq_time is only updated on {soft,}irq_exit, we might run into
878 * this case when a previous update_rq_clock() happened inside a
881 * When this happens, we stop ->clock_task and only update the
882 * prev_irq_time stamp to account for the part that fit, so that a next
883 * update will consume the rest. This ensures ->clock_task is
886 * It does however cause some slight miss-attribution of {soft,}irq
887 * time, a more accurate solution would be to update the irq_time using
888 * the current rq->clock timestamp, except that would require using
891 if (irq_delta > delta)
894 rq->prev_irq_time += irq_delta;
897 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
898 if (static_key_false((¶virt_steal_rq_enabled))) {
901 steal = paravirt_steal_clock(cpu_of(rq));
902 steal -= rq->prev_steal_time_rq;
904 if (unlikely(steal > delta))
907 st = steal_ticks(steal);
908 steal = st * TICK_NSEC;
910 rq->prev_steal_time_rq += steal;
916 rq->clock_task += delta;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
920 sched_rt_avg_update(rq, irq_delta + steal);
924 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
925 static int irqtime_account_hi_update(void)
927 u64 *cpustat = kcpustat_this_cpu->cpustat;
932 local_irq_save(flags);
933 latest_ns = this_cpu_read(cpu_hardirq_time);
934 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
936 local_irq_restore(flags);
940 static int irqtime_account_si_update(void)
942 u64 *cpustat = kcpustat_this_cpu->cpustat;
947 local_irq_save(flags);
948 latest_ns = this_cpu_read(cpu_softirq_time);
949 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
951 local_irq_restore(flags);
955 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
957 #define sched_clock_irqtime (0)
961 void sched_set_stop_task(int cpu, struct task_struct *stop)
963 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
964 struct task_struct *old_stop = cpu_rq(cpu)->stop;
968 * Make it appear like a SCHED_FIFO task, its something
969 * userspace knows about and won't get confused about.
971 * Also, it will make PI more or less work without too
972 * much confusion -- but then, stop work should not
973 * rely on PI working anyway.
975 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
977 stop->sched_class = &stop_sched_class;
980 cpu_rq(cpu)->stop = stop;
984 * Reset it back to a normal scheduling class so that
985 * it can die in pieces.
987 old_stop->sched_class = &rt_sched_class;
992 * __normal_prio - return the priority that is based on the static prio
994 static inline int __normal_prio(struct task_struct *p)
996 return p->static_prio;
1000 * Calculate the expected normal priority: i.e. priority
1001 * without taking RT-inheritance into account. Might be
1002 * boosted by interactivity modifiers. Changes upon fork,
1003 * setprio syscalls, and whenever the interactivity
1004 * estimator recalculates.
1006 static inline int normal_prio(struct task_struct *p)
1010 if (task_has_rt_policy(p))
1011 prio = MAX_RT_PRIO-1 - p->rt_priority;
1013 prio = __normal_prio(p);
1018 * Calculate the current priority, i.e. the priority
1019 * taken into account by the scheduler. This value might
1020 * be boosted by RT tasks, or might be boosted by
1021 * interactivity modifiers. Will be RT if the task got
1022 * RT-boosted. If not then it returns p->normal_prio.
1024 static int effective_prio(struct task_struct *p)
1026 p->normal_prio = normal_prio(p);
1028 * If we are RT tasks or we were boosted to RT priority,
1029 * keep the priority unchanged. Otherwise, update priority
1030 * to the normal priority:
1032 if (!rt_prio(p->prio))
1033 return p->normal_prio;
1038 * task_curr - is this task currently executing on a CPU?
1039 * @p: the task in question.
1041 inline int task_curr(const struct task_struct *p)
1043 return cpu_curr(task_cpu(p)) == p;
1046 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1047 const struct sched_class *prev_class,
1050 if (prev_class != p->sched_class) {
1051 if (prev_class->switched_from)
1052 prev_class->switched_from(rq, p);
1053 p->sched_class->switched_to(rq, p);
1054 } else if (oldprio != p->prio)
1055 p->sched_class->prio_changed(rq, p, oldprio);
1058 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1060 const struct sched_class *class;
1062 if (p->sched_class == rq->curr->sched_class) {
1063 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1065 for_each_class(class) {
1066 if (class == rq->curr->sched_class)
1068 if (class == p->sched_class) {
1069 resched_task(rq->curr);
1076 * A queue event has occurred, and we're going to schedule. In
1077 * this case, we can save a useless back to back clock update.
1079 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1080 rq->skip_clock_update = 1;
1084 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1086 #ifdef CONFIG_SCHED_DEBUG
1088 * We should never call set_task_cpu() on a blocked task,
1089 * ttwu() will sort out the placement.
1091 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1092 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1094 #ifdef CONFIG_LOCKDEP
1096 * The caller should hold either p->pi_lock or rq->lock, when changing
1097 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1099 * sched_move_task() holds both and thus holding either pins the cgroup,
1100 * see set_task_rq().
1102 * Furthermore, all task_rq users should acquire both locks, see
1105 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1106 lockdep_is_held(&task_rq(p)->lock)));
1110 trace_sched_migrate_task(p, new_cpu);
1112 if (task_cpu(p) != new_cpu) {
1113 p->se.nr_migrations++;
1114 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1117 __set_task_cpu(p, new_cpu);
1120 struct migration_arg {
1121 struct task_struct *task;
1125 static int migration_cpu_stop(void *data);
1128 * wait_task_inactive - wait for a thread to unschedule.
1130 * If @match_state is nonzero, it's the @p->state value just checked and
1131 * not expected to change. If it changes, i.e. @p might have woken up,
1132 * then return zero. When we succeed in waiting for @p to be off its CPU,
1133 * we return a positive number (its total switch count). If a second call
1134 * a short while later returns the same number, the caller can be sure that
1135 * @p has remained unscheduled the whole time.
1137 * The caller must ensure that the task *will* unschedule sometime soon,
1138 * else this function might spin for a *long* time. This function can't
1139 * be called with interrupts off, or it may introduce deadlock with
1140 * smp_call_function() if an IPI is sent by the same process we are
1141 * waiting to become inactive.
1143 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1145 unsigned long flags;
1152 * We do the initial early heuristics without holding
1153 * any task-queue locks at all. We'll only try to get
1154 * the runqueue lock when things look like they will
1160 * If the task is actively running on another CPU
1161 * still, just relax and busy-wait without holding
1164 * NOTE! Since we don't hold any locks, it's not
1165 * even sure that "rq" stays as the right runqueue!
1166 * But we don't care, since "task_running()" will
1167 * return false if the runqueue has changed and p
1168 * is actually now running somewhere else!
1170 while (task_running(rq, p)) {
1171 if (match_state && unlikely(p->state != match_state))
1177 * Ok, time to look more closely! We need the rq
1178 * lock now, to be *sure*. If we're wrong, we'll
1179 * just go back and repeat.
1181 rq = task_rq_lock(p, &flags);
1182 trace_sched_wait_task(p);
1183 running = task_running(rq, p);
1186 if (!match_state || p->state == match_state)
1187 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1188 task_rq_unlock(rq, p, &flags);
1191 * If it changed from the expected state, bail out now.
1193 if (unlikely(!ncsw))
1197 * Was it really running after all now that we
1198 * checked with the proper locks actually held?
1200 * Oops. Go back and try again..
1202 if (unlikely(running)) {
1208 * It's not enough that it's not actively running,
1209 * it must be off the runqueue _entirely_, and not
1212 * So if it was still runnable (but just not actively
1213 * running right now), it's preempted, and we should
1214 * yield - it could be a while.
1216 if (unlikely(on_rq)) {
1217 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1219 set_current_state(TASK_UNINTERRUPTIBLE);
1220 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1225 * Ahh, all good. It wasn't running, and it wasn't
1226 * runnable, which means that it will never become
1227 * running in the future either. We're all done!
1236 * kick_process - kick a running thread to enter/exit the kernel
1237 * @p: the to-be-kicked thread
1239 * Cause a process which is running on another CPU to enter
1240 * kernel-mode, without any delay. (to get signals handled.)
1242 * NOTE: this function doesn't have to take the runqueue lock,
1243 * because all it wants to ensure is that the remote task enters
1244 * the kernel. If the IPI races and the task has been migrated
1245 * to another CPU then no harm is done and the purpose has been
1248 void kick_process(struct task_struct *p)
1254 if ((cpu != smp_processor_id()) && task_curr(p))
1255 smp_send_reschedule(cpu);
1258 EXPORT_SYMBOL_GPL(kick_process);
1259 #endif /* CONFIG_SMP */
1263 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1265 static int select_fallback_rq(int cpu, struct task_struct *p)
1267 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1268 enum { cpuset, possible, fail } state = cpuset;
1271 /* Look for allowed, online CPU in same node. */
1272 for_each_cpu(dest_cpu, nodemask) {
1273 if (!cpu_online(dest_cpu))
1275 if (!cpu_active(dest_cpu))
1277 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1282 /* Any allowed, online CPU? */
1283 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1284 if (!cpu_online(dest_cpu))
1286 if (!cpu_active(dest_cpu))
1293 /* No more Mr. Nice Guy. */
1294 cpuset_cpus_allowed_fallback(p);
1299 do_set_cpus_allowed(p, cpu_possible_mask);
1310 if (state != cpuset) {
1312 * Don't tell them about moving exiting tasks or
1313 * kernel threads (both mm NULL), since they never
1316 if (p->mm && printk_ratelimit()) {
1317 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1318 task_pid_nr(p), p->comm, cpu);
1326 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1329 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1331 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1334 * In order not to call set_task_cpu() on a blocking task we need
1335 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1338 * Since this is common to all placement strategies, this lives here.
1340 * [ this allows ->select_task() to simply return task_cpu(p) and
1341 * not worry about this generic constraint ]
1343 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1345 cpu = select_fallback_rq(task_cpu(p), p);
1350 static void update_avg(u64 *avg, u64 sample)
1352 s64 diff = sample - *avg;
1358 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1360 #ifdef CONFIG_SCHEDSTATS
1361 struct rq *rq = this_rq();
1364 int this_cpu = smp_processor_id();
1366 if (cpu == this_cpu) {
1367 schedstat_inc(rq, ttwu_local);
1368 schedstat_inc(p, se.statistics.nr_wakeups_local);
1370 struct sched_domain *sd;
1372 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1374 for_each_domain(this_cpu, sd) {
1375 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1376 schedstat_inc(sd, ttwu_wake_remote);
1383 if (wake_flags & WF_MIGRATED)
1384 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1386 #endif /* CONFIG_SMP */
1388 schedstat_inc(rq, ttwu_count);
1389 schedstat_inc(p, se.statistics.nr_wakeups);
1391 if (wake_flags & WF_SYNC)
1392 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1394 #endif /* CONFIG_SCHEDSTATS */
1397 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1399 activate_task(rq, p, en_flags);
1402 /* if a worker is waking up, notify workqueue */
1403 if (p->flags & PF_WQ_WORKER)
1404 wq_worker_waking_up(p, cpu_of(rq));
1408 * Mark the task runnable and perform wakeup-preemption.
1411 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1413 trace_sched_wakeup(p, true);
1414 check_preempt_curr(rq, p, wake_flags);
1416 p->state = TASK_RUNNING;
1418 if (p->sched_class->task_woken)
1419 p->sched_class->task_woken(rq, p);
1421 if (rq->idle_stamp) {
1422 u64 delta = rq->clock - rq->idle_stamp;
1423 u64 max = 2*sysctl_sched_migration_cost;
1428 update_avg(&rq->avg_idle, delta);
1435 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1438 if (p->sched_contributes_to_load)
1439 rq->nr_uninterruptible--;
1442 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1443 ttwu_do_wakeup(rq, p, wake_flags);
1447 * Called in case the task @p isn't fully descheduled from its runqueue,
1448 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1449 * since all we need to do is flip p->state to TASK_RUNNING, since
1450 * the task is still ->on_rq.
1452 static int ttwu_remote(struct task_struct *p, int wake_flags)
1457 rq = __task_rq_lock(p);
1459 ttwu_do_wakeup(rq, p, wake_flags);
1462 __task_rq_unlock(rq);
1468 static void sched_ttwu_pending(void)
1470 struct rq *rq = this_rq();
1471 struct llist_node *llist = llist_del_all(&rq->wake_list);
1472 struct task_struct *p;
1474 raw_spin_lock(&rq->lock);
1477 p = llist_entry(llist, struct task_struct, wake_entry);
1478 llist = llist_next(llist);
1479 ttwu_do_activate(rq, p, 0);
1482 raw_spin_unlock(&rq->lock);
1485 void scheduler_ipi(void)
1487 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1491 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1492 * traditionally all their work was done from the interrupt return
1493 * path. Now that we actually do some work, we need to make sure
1496 * Some archs already do call them, luckily irq_enter/exit nest
1499 * Arguably we should visit all archs and update all handlers,
1500 * however a fair share of IPIs are still resched only so this would
1501 * somewhat pessimize the simple resched case.
1504 sched_ttwu_pending();
1507 * Check if someone kicked us for doing the nohz idle load balance.
1509 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1510 this_rq()->idle_balance = 1;
1511 raise_softirq_irqoff(SCHED_SOFTIRQ);
1516 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1518 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1519 smp_send_reschedule(cpu);
1522 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1523 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1528 rq = __task_rq_lock(p);
1530 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1531 ttwu_do_wakeup(rq, p, wake_flags);
1534 __task_rq_unlock(rq);
1539 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1541 bool cpus_share_cache(int this_cpu, int that_cpu)
1543 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1545 #endif /* CONFIG_SMP */
1547 static void ttwu_queue(struct task_struct *p, int cpu)
1549 struct rq *rq = cpu_rq(cpu);
1551 #if defined(CONFIG_SMP)
1552 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554 ttwu_queue_remote(p, cpu);
1559 raw_spin_lock(&rq->lock);
1560 ttwu_do_activate(rq, p, 0);
1561 raw_spin_unlock(&rq->lock);
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1576 * Returns %true if @p was woken up, %false if it was already running
1577 * or @state didn't match @p's state.
1580 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1582 unsigned long flags;
1583 int cpu, success = 0;
1586 raw_spin_lock_irqsave(&p->pi_lock, flags);
1587 if (!(p->state & state))
1590 success = 1; /* we're going to change ->state */
1593 if (p->on_rq && ttwu_remote(p, wake_flags))
1598 * If the owning (remote) cpu is still in the middle of schedule() with
1599 * this task as prev, wait until its done referencing the task.
1602 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1604 * In case the architecture enables interrupts in
1605 * context_switch(), we cannot busy wait, since that
1606 * would lead to deadlocks when an interrupt hits and
1607 * tries to wake up @prev. So bail and do a complete
1610 if (ttwu_activate_remote(p, wake_flags))
1617 * Pairs with the smp_wmb() in finish_lock_switch().
1621 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1622 p->state = TASK_WAKING;
1624 if (p->sched_class->task_waking)
1625 p->sched_class->task_waking(p);
1627 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1628 if (task_cpu(p) != cpu) {
1629 wake_flags |= WF_MIGRATED;
1630 set_task_cpu(p, cpu);
1632 #endif /* CONFIG_SMP */
1636 ttwu_stat(p, cpu, wake_flags);
1638 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1644 * try_to_wake_up_local - try to wake up a local task with rq lock held
1645 * @p: the thread to be awakened
1647 * Put @p on the run-queue if it's not already there. The caller must
1648 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1651 static void try_to_wake_up_local(struct task_struct *p)
1653 struct rq *rq = task_rq(p);
1655 BUG_ON(rq != this_rq());
1656 BUG_ON(p == current);
1657 lockdep_assert_held(&rq->lock);
1659 if (!raw_spin_trylock(&p->pi_lock)) {
1660 raw_spin_unlock(&rq->lock);
1661 raw_spin_lock(&p->pi_lock);
1662 raw_spin_lock(&rq->lock);
1665 if (!(p->state & TASK_NORMAL))
1669 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1671 ttwu_do_wakeup(rq, p, 0);
1672 ttwu_stat(p, smp_processor_id(), 0);
1674 raw_spin_unlock(&p->pi_lock);
1678 * wake_up_process - Wake up a specific process
1679 * @p: The process to be woken up.
1681 * Attempt to wake up the nominated process and move it to the set of runnable
1682 * processes. Returns 1 if the process was woken up, 0 if it was already
1685 * It may be assumed that this function implies a write memory barrier before
1686 * changing the task state if and only if any tasks are woken up.
1688 int wake_up_process(struct task_struct *p)
1690 return try_to_wake_up(p, TASK_ALL, 0);
1692 EXPORT_SYMBOL(wake_up_process);
1694 int wake_up_state(struct task_struct *p, unsigned int state)
1696 return try_to_wake_up(p, state, 0);
1700 * Perform scheduler related setup for a newly forked process p.
1701 * p is forked by current.
1703 * __sched_fork() is basic setup used by init_idle() too:
1705 static void __sched_fork(struct task_struct *p)
1710 p->se.exec_start = 0;
1711 p->se.sum_exec_runtime = 0;
1712 p->se.prev_sum_exec_runtime = 0;
1713 p->se.nr_migrations = 0;
1715 INIT_LIST_HEAD(&p->se.group_node);
1717 #ifdef CONFIG_SCHEDSTATS
1718 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1721 INIT_LIST_HEAD(&p->rt.run_list);
1723 #ifdef CONFIG_PREEMPT_NOTIFIERS
1724 INIT_HLIST_HEAD(&p->preempt_notifiers);
1729 * fork()/clone()-time setup:
1731 void sched_fork(struct task_struct *p)
1733 unsigned long flags;
1734 int cpu = get_cpu();
1738 * We mark the process as running here. This guarantees that
1739 * nobody will actually run it, and a signal or other external
1740 * event cannot wake it up and insert it on the runqueue either.
1742 p->state = TASK_RUNNING;
1745 * Make sure we do not leak PI boosting priority to the child.
1747 p->prio = current->normal_prio;
1750 * Revert to default priority/policy on fork if requested.
1752 if (unlikely(p->sched_reset_on_fork)) {
1753 if (task_has_rt_policy(p)) {
1754 p->policy = SCHED_NORMAL;
1755 p->static_prio = NICE_TO_PRIO(0);
1757 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1758 p->static_prio = NICE_TO_PRIO(0);
1760 p->prio = p->normal_prio = __normal_prio(p);
1764 * We don't need the reset flag anymore after the fork. It has
1765 * fulfilled its duty:
1767 p->sched_reset_on_fork = 0;
1770 if (!rt_prio(p->prio))
1771 p->sched_class = &fair_sched_class;
1773 if (p->sched_class->task_fork)
1774 p->sched_class->task_fork(p);
1777 * The child is not yet in the pid-hash so no cgroup attach races,
1778 * and the cgroup is pinned to this child due to cgroup_fork()
1779 * is ran before sched_fork().
1781 * Silence PROVE_RCU.
1783 raw_spin_lock_irqsave(&p->pi_lock, flags);
1784 set_task_cpu(p, cpu);
1785 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1787 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1788 if (likely(sched_info_on()))
1789 memset(&p->sched_info, 0, sizeof(p->sched_info));
1791 #if defined(CONFIG_SMP)
1794 #ifdef CONFIG_PREEMPT_COUNT
1795 /* Want to start with kernel preemption disabled. */
1796 task_thread_info(p)->preempt_count = 1;
1799 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1806 * wake_up_new_task - wake up a newly created task for the first time.
1808 * This function will do some initial scheduler statistics housekeeping
1809 * that must be done for every newly created context, then puts the task
1810 * on the runqueue and wakes it.
1812 void wake_up_new_task(struct task_struct *p)
1814 unsigned long flags;
1817 raw_spin_lock_irqsave(&p->pi_lock, flags);
1820 * Fork balancing, do it here and not earlier because:
1821 * - cpus_allowed can change in the fork path
1822 * - any previously selected cpu might disappear through hotplug
1824 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1827 rq = __task_rq_lock(p);
1828 activate_task(rq, p, 0);
1830 trace_sched_wakeup_new(p, true);
1831 check_preempt_curr(rq, p, WF_FORK);
1833 if (p->sched_class->task_woken)
1834 p->sched_class->task_woken(rq, p);
1836 task_rq_unlock(rq, p, &flags);
1839 #ifdef CONFIG_PREEMPT_NOTIFIERS
1842 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1843 * @notifier: notifier struct to register
1845 void preempt_notifier_register(struct preempt_notifier *notifier)
1847 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1849 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1852 * preempt_notifier_unregister - no longer interested in preemption notifications
1853 * @notifier: notifier struct to unregister
1855 * This is safe to call from within a preemption notifier.
1857 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1859 hlist_del(¬ifier->link);
1861 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1863 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1865 struct preempt_notifier *notifier;
1866 struct hlist_node *node;
1868 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1869 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1873 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1874 struct task_struct *next)
1876 struct preempt_notifier *notifier;
1877 struct hlist_node *node;
1879 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1880 notifier->ops->sched_out(notifier, next);
1883 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1885 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1890 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1891 struct task_struct *next)
1895 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1898 * prepare_task_switch - prepare to switch tasks
1899 * @rq: the runqueue preparing to switch
1900 * @prev: the current task that is being switched out
1901 * @next: the task we are going to switch to.
1903 * This is called with the rq lock held and interrupts off. It must
1904 * be paired with a subsequent finish_task_switch after the context
1907 * prepare_task_switch sets up locking and calls architecture specific
1911 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1912 struct task_struct *next)
1914 sched_info_switch(prev, next);
1915 perf_event_task_sched(prev, next);
1916 fire_sched_out_preempt_notifiers(prev, next);
1917 prepare_lock_switch(rq, next);
1918 prepare_arch_switch(next);
1919 trace_sched_switch(prev, next);
1923 * finish_task_switch - clean up after a task-switch
1924 * @rq: runqueue associated with task-switch
1925 * @prev: the thread we just switched away from.
1927 * finish_task_switch must be called after the context switch, paired
1928 * with a prepare_task_switch call before the context switch.
1929 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1930 * and do any other architecture-specific cleanup actions.
1932 * Note that we may have delayed dropping an mm in context_switch(). If
1933 * so, we finish that here outside of the runqueue lock. (Doing it
1934 * with the lock held can cause deadlocks; see schedule() for
1937 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1938 __releases(rq->lock)
1940 struct mm_struct *mm = rq->prev_mm;
1946 * A task struct has one reference for the use as "current".
1947 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1948 * schedule one last time. The schedule call will never return, and
1949 * the scheduled task must drop that reference.
1950 * The test for TASK_DEAD must occur while the runqueue locks are
1951 * still held, otherwise prev could be scheduled on another cpu, die
1952 * there before we look at prev->state, and then the reference would
1954 * Manfred Spraul <manfred@colorfullife.com>
1956 prev_state = prev->state;
1957 finish_arch_switch(prev);
1958 finish_lock_switch(rq, prev);
1959 finish_arch_post_lock_switch();
1961 fire_sched_in_preempt_notifiers(current);
1964 if (unlikely(prev_state == TASK_DEAD)) {
1966 * Remove function-return probe instances associated with this
1967 * task and put them back on the free list.
1969 kprobe_flush_task(prev);
1970 put_task_struct(prev);
1976 /* assumes rq->lock is held */
1977 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1979 if (prev->sched_class->pre_schedule)
1980 prev->sched_class->pre_schedule(rq, prev);
1983 /* rq->lock is NOT held, but preemption is disabled */
1984 static inline void post_schedule(struct rq *rq)
1986 if (rq->post_schedule) {
1987 unsigned long flags;
1989 raw_spin_lock_irqsave(&rq->lock, flags);
1990 if (rq->curr->sched_class->post_schedule)
1991 rq->curr->sched_class->post_schedule(rq);
1992 raw_spin_unlock_irqrestore(&rq->lock, flags);
1994 rq->post_schedule = 0;
2000 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2004 static inline void post_schedule(struct rq *rq)
2011 * schedule_tail - first thing a freshly forked thread must call.
2012 * @prev: the thread we just switched away from.
2014 asmlinkage void schedule_tail(struct task_struct *prev)
2015 __releases(rq->lock)
2017 struct rq *rq = this_rq();
2019 finish_task_switch(rq, prev);
2022 * FIXME: do we need to worry about rq being invalidated by the
2027 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2028 /* In this case, finish_task_switch does not reenable preemption */
2031 if (current->set_child_tid)
2032 put_user(task_pid_vnr(current), current->set_child_tid);
2036 * context_switch - switch to the new MM and the new
2037 * thread's register state.
2040 context_switch(struct rq *rq, struct task_struct *prev,
2041 struct task_struct *next)
2043 struct mm_struct *mm, *oldmm;
2045 prepare_task_switch(rq, prev, next);
2048 oldmm = prev->active_mm;
2050 * For paravirt, this is coupled with an exit in switch_to to
2051 * combine the page table reload and the switch backend into
2054 arch_start_context_switch(prev);
2057 next->active_mm = oldmm;
2058 atomic_inc(&oldmm->mm_count);
2059 enter_lazy_tlb(oldmm, next);
2061 switch_mm(oldmm, mm, next);
2064 prev->active_mm = NULL;
2065 rq->prev_mm = oldmm;
2068 * Since the runqueue lock will be released by the next
2069 * task (which is an invalid locking op but in the case
2070 * of the scheduler it's an obvious special-case), so we
2071 * do an early lockdep release here:
2073 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2074 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2077 /* Here we just switch the register state and the stack. */
2078 rcu_switch_from(prev);
2079 switch_to(prev, next, prev);
2083 * this_rq must be evaluated again because prev may have moved
2084 * CPUs since it called schedule(), thus the 'rq' on its stack
2085 * frame will be invalid.
2087 finish_task_switch(this_rq(), prev);
2091 * nr_running, nr_uninterruptible and nr_context_switches:
2093 * externally visible scheduler statistics: current number of runnable
2094 * threads, current number of uninterruptible-sleeping threads, total
2095 * number of context switches performed since bootup.
2097 unsigned long nr_running(void)
2099 unsigned long i, sum = 0;
2101 for_each_online_cpu(i)
2102 sum += cpu_rq(i)->nr_running;
2107 unsigned long nr_uninterruptible(void)
2109 unsigned long i, sum = 0;
2111 for_each_possible_cpu(i)
2112 sum += cpu_rq(i)->nr_uninterruptible;
2115 * Since we read the counters lockless, it might be slightly
2116 * inaccurate. Do not allow it to go below zero though:
2118 if (unlikely((long)sum < 0))
2124 unsigned long long nr_context_switches(void)
2127 unsigned long long sum = 0;
2129 for_each_possible_cpu(i)
2130 sum += cpu_rq(i)->nr_switches;
2135 unsigned long nr_iowait(void)
2137 unsigned long i, sum = 0;
2139 for_each_possible_cpu(i)
2140 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2145 unsigned long nr_iowait_cpu(int cpu)
2147 struct rq *this = cpu_rq(cpu);
2148 return atomic_read(&this->nr_iowait);
2151 unsigned long this_cpu_load(void)
2153 struct rq *this = this_rq();
2154 return this->cpu_load[0];
2158 /* Variables and functions for calc_load */
2159 static atomic_long_t calc_load_tasks;
2160 static unsigned long calc_load_update;
2161 unsigned long avenrun[3];
2162 EXPORT_SYMBOL(avenrun);
2164 static long calc_load_fold_active(struct rq *this_rq)
2166 long nr_active, delta = 0;
2168 nr_active = this_rq->nr_running;
2169 nr_active += (long) this_rq->nr_uninterruptible;
2171 if (nr_active != this_rq->calc_load_active) {
2172 delta = nr_active - this_rq->calc_load_active;
2173 this_rq->calc_load_active = nr_active;
2179 static unsigned long
2180 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2183 load += active * (FIXED_1 - exp);
2184 load += 1UL << (FSHIFT - 1);
2185 return load >> FSHIFT;
2190 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2192 * When making the ILB scale, we should try to pull this in as well.
2194 static atomic_long_t calc_load_tasks_idle;
2196 void calc_load_account_idle(struct rq *this_rq)
2200 delta = calc_load_fold_active(this_rq);
2202 atomic_long_add(delta, &calc_load_tasks_idle);
2205 static long calc_load_fold_idle(void)
2210 * Its got a race, we don't care...
2212 if (atomic_long_read(&calc_load_tasks_idle))
2213 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2219 * fixed_power_int - compute: x^n, in O(log n) time
2221 * @x: base of the power
2222 * @frac_bits: fractional bits of @x
2223 * @n: power to raise @x to.
2225 * By exploiting the relation between the definition of the natural power
2226 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2227 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2228 * (where: n_i \elem {0, 1}, the binary vector representing n),
2229 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2230 * of course trivially computable in O(log_2 n), the length of our binary
2233 static unsigned long
2234 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2236 unsigned long result = 1UL << frac_bits;
2241 result += 1UL << (frac_bits - 1);
2242 result >>= frac_bits;
2248 x += 1UL << (frac_bits - 1);
2256 * a1 = a0 * e + a * (1 - e)
2258 * a2 = a1 * e + a * (1 - e)
2259 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2260 * = a0 * e^2 + a * (1 - e) * (1 + e)
2262 * a3 = a2 * e + a * (1 - e)
2263 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2264 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2268 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2269 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2270 * = a0 * e^n + a * (1 - e^n)
2272 * [1] application of the geometric series:
2275 * S_n := \Sum x^i = -------------
2278 static unsigned long
2279 calc_load_n(unsigned long load, unsigned long exp,
2280 unsigned long active, unsigned int n)
2283 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2287 * NO_HZ can leave us missing all per-cpu ticks calling
2288 * calc_load_account_active(), but since an idle CPU folds its delta into
2289 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2290 * in the pending idle delta if our idle period crossed a load cycle boundary.
2292 * Once we've updated the global active value, we need to apply the exponential
2293 * weights adjusted to the number of cycles missed.
2295 static void calc_global_nohz(void)
2297 long delta, active, n;
2300 * If we crossed a calc_load_update boundary, make sure to fold
2301 * any pending idle changes, the respective CPUs might have
2302 * missed the tick driven calc_load_account_active() update
2305 delta = calc_load_fold_idle();
2307 atomic_long_add(delta, &calc_load_tasks);
2310 * It could be the one fold was all it took, we done!
2312 if (time_before(jiffies, calc_load_update + 10))
2316 * Catch-up, fold however many we are behind still
2318 delta = jiffies - calc_load_update - 10;
2319 n = 1 + (delta / LOAD_FREQ);
2321 active = atomic_long_read(&calc_load_tasks);
2322 active = active > 0 ? active * FIXED_1 : 0;
2324 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2325 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2326 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2328 calc_load_update += n * LOAD_FREQ;
2331 void calc_load_account_idle(struct rq *this_rq)
2335 static inline long calc_load_fold_idle(void)
2340 static void calc_global_nohz(void)
2346 * get_avenrun - get the load average array
2347 * @loads: pointer to dest load array
2348 * @offset: offset to add
2349 * @shift: shift count to shift the result left
2351 * These values are estimates at best, so no need for locking.
2353 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2355 loads[0] = (avenrun[0] + offset) << shift;
2356 loads[1] = (avenrun[1] + offset) << shift;
2357 loads[2] = (avenrun[2] + offset) << shift;
2361 * calc_load - update the avenrun load estimates 10 ticks after the
2362 * CPUs have updated calc_load_tasks.
2364 void calc_global_load(unsigned long ticks)
2368 if (time_before(jiffies, calc_load_update + 10))
2371 active = atomic_long_read(&calc_load_tasks);
2372 active = active > 0 ? active * FIXED_1 : 0;
2374 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2375 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2376 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2378 calc_load_update += LOAD_FREQ;
2381 * Account one period with whatever state we found before
2382 * folding in the nohz state and ageing the entire idle period.
2384 * This avoids loosing a sample when we go idle between
2385 * calc_load_account_active() (10 ticks ago) and now and thus
2392 * Called from update_cpu_load() to periodically update this CPU's
2395 static void calc_load_account_active(struct rq *this_rq)
2399 if (time_before(jiffies, this_rq->calc_load_update))
2402 delta = calc_load_fold_active(this_rq);
2403 delta += calc_load_fold_idle();
2405 atomic_long_add(delta, &calc_load_tasks);
2407 this_rq->calc_load_update += LOAD_FREQ;
2411 * The exact cpuload at various idx values, calculated at every tick would be
2412 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2414 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2415 * on nth tick when cpu may be busy, then we have:
2416 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2417 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2419 * decay_load_missed() below does efficient calculation of
2420 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2421 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2423 * The calculation is approximated on a 128 point scale.
2424 * degrade_zero_ticks is the number of ticks after which load at any
2425 * particular idx is approximated to be zero.
2426 * degrade_factor is a precomputed table, a row for each load idx.
2427 * Each column corresponds to degradation factor for a power of two ticks,
2428 * based on 128 point scale.
2430 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2431 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2433 * With this power of 2 load factors, we can degrade the load n times
2434 * by looking at 1 bits in n and doing as many mult/shift instead of
2435 * n mult/shifts needed by the exact degradation.
2437 #define DEGRADE_SHIFT 7
2438 static const unsigned char
2439 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2440 static const unsigned char
2441 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2442 {0, 0, 0, 0, 0, 0, 0, 0},
2443 {64, 32, 8, 0, 0, 0, 0, 0},
2444 {96, 72, 40, 12, 1, 0, 0},
2445 {112, 98, 75, 43, 15, 1, 0},
2446 {120, 112, 98, 76, 45, 16, 2} };
2449 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2450 * would be when CPU is idle and so we just decay the old load without
2451 * adding any new load.
2453 static unsigned long
2454 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2458 if (!missed_updates)
2461 if (missed_updates >= degrade_zero_ticks[idx])
2465 return load >> missed_updates;
2467 while (missed_updates) {
2468 if (missed_updates % 2)
2469 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2471 missed_updates >>= 1;
2478 * Update rq->cpu_load[] statistics. This function is usually called every
2479 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2480 * every tick. We fix it up based on jiffies.
2482 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2483 unsigned long pending_updates)
2487 this_rq->nr_load_updates++;
2489 /* Update our load: */
2490 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2491 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2492 unsigned long old_load, new_load;
2494 /* scale is effectively 1 << i now, and >> i divides by scale */
2496 old_load = this_rq->cpu_load[i];
2497 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2498 new_load = this_load;
2500 * Round up the averaging division if load is increasing. This
2501 * prevents us from getting stuck on 9 if the load is 10, for
2504 if (new_load > old_load)
2505 new_load += scale - 1;
2507 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2510 sched_avg_update(this_rq);
2514 * Called from nohz_idle_balance() to update the load ratings before doing the
2517 void update_idle_cpu_load(struct rq *this_rq)
2519 unsigned long curr_jiffies = jiffies;
2520 unsigned long load = this_rq->load.weight;
2521 unsigned long pending_updates;
2524 * Bloody broken means of dealing with nohz, but better than nothing..
2525 * jiffies is updated by one cpu, another cpu can drift wrt the jiffy
2526 * update and see 0 difference the one time and 2 the next, even though
2527 * we ticked at roughtly the same rate.
2529 * Hence we only use this from nohz_idle_balance() and skip this
2530 * nonsense when called from the scheduler_tick() since that's
2531 * guaranteed a stable rate.
2533 if (load || curr_jiffies == this_rq->last_load_update_tick)
2536 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2537 this_rq->last_load_update_tick = curr_jiffies;
2539 __update_cpu_load(this_rq, load, pending_updates);
2543 * Called from scheduler_tick()
2545 static void update_cpu_load_active(struct rq *this_rq)
2548 * See the mess in update_idle_cpu_load().
2550 this_rq->last_load_update_tick = jiffies;
2551 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2553 calc_load_account_active(this_rq);
2559 * sched_exec - execve() is a valuable balancing opportunity, because at
2560 * this point the task has the smallest effective memory and cache footprint.
2562 void sched_exec(void)
2564 struct task_struct *p = current;
2565 unsigned long flags;
2568 raw_spin_lock_irqsave(&p->pi_lock, flags);
2569 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2570 if (dest_cpu == smp_processor_id())
2573 if (likely(cpu_active(dest_cpu))) {
2574 struct migration_arg arg = { p, dest_cpu };
2576 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2577 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2581 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2586 DEFINE_PER_CPU(struct kernel_stat, kstat);
2587 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2589 EXPORT_PER_CPU_SYMBOL(kstat);
2590 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2593 * Return any ns on the sched_clock that have not yet been accounted in
2594 * @p in case that task is currently running.
2596 * Called with task_rq_lock() held on @rq.
2598 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2602 if (task_current(rq, p)) {
2603 update_rq_clock(rq);
2604 ns = rq->clock_task - p->se.exec_start;
2612 unsigned long long task_delta_exec(struct task_struct *p)
2614 unsigned long flags;
2618 rq = task_rq_lock(p, &flags);
2619 ns = do_task_delta_exec(p, rq);
2620 task_rq_unlock(rq, p, &flags);
2626 * Return accounted runtime for the task.
2627 * In case the task is currently running, return the runtime plus current's
2628 * pending runtime that have not been accounted yet.
2630 unsigned long long task_sched_runtime(struct task_struct *p)
2632 unsigned long flags;
2636 rq = task_rq_lock(p, &flags);
2637 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2638 task_rq_unlock(rq, p, &flags);
2643 #ifdef CONFIG_CGROUP_CPUACCT
2644 struct cgroup_subsys cpuacct_subsys;
2645 struct cpuacct root_cpuacct;
2648 static inline void task_group_account_field(struct task_struct *p, int index,
2651 #ifdef CONFIG_CGROUP_CPUACCT
2652 struct kernel_cpustat *kcpustat;
2656 * Since all updates are sure to touch the root cgroup, we
2657 * get ourselves ahead and touch it first. If the root cgroup
2658 * is the only cgroup, then nothing else should be necessary.
2661 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2663 #ifdef CONFIG_CGROUP_CPUACCT
2664 if (unlikely(!cpuacct_subsys.active))
2669 while (ca && (ca != &root_cpuacct)) {
2670 kcpustat = this_cpu_ptr(ca->cpustat);
2671 kcpustat->cpustat[index] += tmp;
2680 * Account user cpu time to a process.
2681 * @p: the process that the cpu time gets accounted to
2682 * @cputime: the cpu time spent in user space since the last update
2683 * @cputime_scaled: cputime scaled by cpu frequency
2685 void account_user_time(struct task_struct *p, cputime_t cputime,
2686 cputime_t cputime_scaled)
2690 /* Add user time to process. */
2691 p->utime += cputime;
2692 p->utimescaled += cputime_scaled;
2693 account_group_user_time(p, cputime);
2695 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2697 /* Add user time to cpustat. */
2698 task_group_account_field(p, index, (__force u64) cputime);
2700 /* Account for user time used */
2701 acct_update_integrals(p);
2705 * Account guest cpu time to a process.
2706 * @p: the process that the cpu time gets accounted to
2707 * @cputime: the cpu time spent in virtual machine since the last update
2708 * @cputime_scaled: cputime scaled by cpu frequency
2710 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2711 cputime_t cputime_scaled)
2713 u64 *cpustat = kcpustat_this_cpu->cpustat;
2715 /* Add guest time to process. */
2716 p->utime += cputime;
2717 p->utimescaled += cputime_scaled;
2718 account_group_user_time(p, cputime);
2719 p->gtime += cputime;
2721 /* Add guest time to cpustat. */
2722 if (TASK_NICE(p) > 0) {
2723 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2724 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2726 cpustat[CPUTIME_USER] += (__force u64) cputime;
2727 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2732 * Account system cpu time to a process and desired cpustat field
2733 * @p: the process that the cpu time gets accounted to
2734 * @cputime: the cpu time spent in kernel space since the last update
2735 * @cputime_scaled: cputime scaled by cpu frequency
2736 * @target_cputime64: pointer to cpustat field that has to be updated
2739 void __account_system_time(struct task_struct *p, cputime_t cputime,
2740 cputime_t cputime_scaled, int index)
2742 /* Add system time to process. */
2743 p->stime += cputime;
2744 p->stimescaled += cputime_scaled;
2745 account_group_system_time(p, cputime);
2747 /* Add system time to cpustat. */
2748 task_group_account_field(p, index, (__force u64) cputime);
2750 /* Account for system time used */
2751 acct_update_integrals(p);
2755 * Account system cpu time to a process.
2756 * @p: the process that the cpu time gets accounted to
2757 * @hardirq_offset: the offset to subtract from hardirq_count()
2758 * @cputime: the cpu time spent in kernel space since the last update
2759 * @cputime_scaled: cputime scaled by cpu frequency
2761 void account_system_time(struct task_struct *p, int hardirq_offset,
2762 cputime_t cputime, cputime_t cputime_scaled)
2766 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2767 account_guest_time(p, cputime, cputime_scaled);
2771 if (hardirq_count() - hardirq_offset)
2772 index = CPUTIME_IRQ;
2773 else if (in_serving_softirq())
2774 index = CPUTIME_SOFTIRQ;
2776 index = CPUTIME_SYSTEM;
2778 __account_system_time(p, cputime, cputime_scaled, index);
2782 * Account for involuntary wait time.
2783 * @cputime: the cpu time spent in involuntary wait
2785 void account_steal_time(cputime_t cputime)
2787 u64 *cpustat = kcpustat_this_cpu->cpustat;
2789 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2793 * Account for idle time.
2794 * @cputime: the cpu time spent in idle wait
2796 void account_idle_time(cputime_t cputime)
2798 u64 *cpustat = kcpustat_this_cpu->cpustat;
2799 struct rq *rq = this_rq();
2801 if (atomic_read(&rq->nr_iowait) > 0)
2802 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2804 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2807 static __always_inline bool steal_account_process_tick(void)
2809 #ifdef CONFIG_PARAVIRT
2810 if (static_key_false(¶virt_steal_enabled)) {
2813 steal = paravirt_steal_clock(smp_processor_id());
2814 steal -= this_rq()->prev_steal_time;
2816 st = steal_ticks(steal);
2817 this_rq()->prev_steal_time += st * TICK_NSEC;
2819 account_steal_time(st);
2826 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2828 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2830 * Account a tick to a process and cpustat
2831 * @p: the process that the cpu time gets accounted to
2832 * @user_tick: is the tick from userspace
2833 * @rq: the pointer to rq
2835 * Tick demultiplexing follows the order
2836 * - pending hardirq update
2837 * - pending softirq update
2841 * - check for guest_time
2842 * - else account as system_time
2844 * Check for hardirq is done both for system and user time as there is
2845 * no timer going off while we are on hardirq and hence we may never get an
2846 * opportunity to update it solely in system time.
2847 * p->stime and friends are only updated on system time and not on irq
2848 * softirq as those do not count in task exec_runtime any more.
2850 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2853 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2854 u64 *cpustat = kcpustat_this_cpu->cpustat;
2856 if (steal_account_process_tick())
2859 if (irqtime_account_hi_update()) {
2860 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2861 } else if (irqtime_account_si_update()) {
2862 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2863 } else if (this_cpu_ksoftirqd() == p) {
2865 * ksoftirqd time do not get accounted in cpu_softirq_time.
2866 * So, we have to handle it separately here.
2867 * Also, p->stime needs to be updated for ksoftirqd.
2869 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2871 } else if (user_tick) {
2872 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2873 } else if (p == rq->idle) {
2874 account_idle_time(cputime_one_jiffy);
2875 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2876 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2878 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2883 static void irqtime_account_idle_ticks(int ticks)
2886 struct rq *rq = this_rq();
2888 for (i = 0; i < ticks; i++)
2889 irqtime_account_process_tick(current, 0, rq);
2891 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2892 static void irqtime_account_idle_ticks(int ticks) {}
2893 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2895 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2898 * Account a single tick of cpu time.
2899 * @p: the process that the cpu time gets accounted to
2900 * @user_tick: indicates if the tick is a user or a system tick
2902 void account_process_tick(struct task_struct *p, int user_tick)
2904 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2905 struct rq *rq = this_rq();
2907 if (sched_clock_irqtime) {
2908 irqtime_account_process_tick(p, user_tick, rq);
2912 if (steal_account_process_tick())
2916 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2917 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2918 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2921 account_idle_time(cputime_one_jiffy);
2925 * Account multiple ticks of steal time.
2926 * @p: the process from which the cpu time has been stolen
2927 * @ticks: number of stolen ticks
2929 void account_steal_ticks(unsigned long ticks)
2931 account_steal_time(jiffies_to_cputime(ticks));
2935 * Account multiple ticks of idle time.
2936 * @ticks: number of stolen ticks
2938 void account_idle_ticks(unsigned long ticks)
2941 if (sched_clock_irqtime) {
2942 irqtime_account_idle_ticks(ticks);
2946 account_idle_time(jiffies_to_cputime(ticks));
2952 * Use precise platform statistics if available:
2954 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2955 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2961 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2963 struct task_cputime cputime;
2965 thread_group_cputime(p, &cputime);
2967 *ut = cputime.utime;
2968 *st = cputime.stime;
2972 #ifndef nsecs_to_cputime
2973 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2976 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2978 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2981 * Use CFS's precise accounting:
2983 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2986 u64 temp = (__force u64) rtime;
2988 temp *= (__force u64) utime;
2989 do_div(temp, (__force u32) total);
2990 utime = (__force cputime_t) temp;
2995 * Compare with previous values, to keep monotonicity:
2997 p->prev_utime = max(p->prev_utime, utime);
2998 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3000 *ut = p->prev_utime;
3001 *st = p->prev_stime;
3005 * Must be called with siglock held.
3007 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3009 struct signal_struct *sig = p->signal;
3010 struct task_cputime cputime;
3011 cputime_t rtime, utime, total;
3013 thread_group_cputime(p, &cputime);
3015 total = cputime.utime + cputime.stime;
3016 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3019 u64 temp = (__force u64) rtime;
3021 temp *= (__force u64) cputime.utime;
3022 do_div(temp, (__force u32) total);
3023 utime = (__force cputime_t) temp;
3027 sig->prev_utime = max(sig->prev_utime, utime);
3028 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3030 *ut = sig->prev_utime;
3031 *st = sig->prev_stime;
3036 * This function gets called by the timer code, with HZ frequency.
3037 * We call it with interrupts disabled.
3039 void scheduler_tick(void)
3041 int cpu = smp_processor_id();
3042 struct rq *rq = cpu_rq(cpu);
3043 struct task_struct *curr = rq->curr;
3047 raw_spin_lock(&rq->lock);
3048 update_rq_clock(rq);
3049 update_cpu_load_active(rq);
3050 curr->sched_class->task_tick(rq, curr, 0);
3051 raw_spin_unlock(&rq->lock);
3053 perf_event_task_tick();
3056 rq->idle_balance = idle_cpu(cpu);
3057 trigger_load_balance(rq, cpu);
3061 notrace unsigned long get_parent_ip(unsigned long addr)
3063 if (in_lock_functions(addr)) {
3064 addr = CALLER_ADDR2;
3065 if (in_lock_functions(addr))
3066 addr = CALLER_ADDR3;
3071 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3072 defined(CONFIG_PREEMPT_TRACER))
3074 void __kprobes add_preempt_count(int val)
3076 #ifdef CONFIG_DEBUG_PREEMPT
3080 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3083 preempt_count() += val;
3084 #ifdef CONFIG_DEBUG_PREEMPT
3086 * Spinlock count overflowing soon?
3088 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3091 if (preempt_count() == val)
3092 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3094 EXPORT_SYMBOL(add_preempt_count);
3096 void __kprobes sub_preempt_count(int val)
3098 #ifdef CONFIG_DEBUG_PREEMPT
3102 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3105 * Is the spinlock portion underflowing?
3107 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3108 !(preempt_count() & PREEMPT_MASK)))
3112 if (preempt_count() == val)
3113 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3114 preempt_count() -= val;
3116 EXPORT_SYMBOL(sub_preempt_count);
3121 * Print scheduling while atomic bug:
3123 static noinline void __schedule_bug(struct task_struct *prev)
3125 if (oops_in_progress)
3128 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3129 prev->comm, prev->pid, preempt_count());
3131 debug_show_held_locks(prev);
3133 if (irqs_disabled())
3134 print_irqtrace_events(prev);
3136 add_taint(TAINT_WARN);
3140 * Various schedule()-time debugging checks and statistics:
3142 static inline void schedule_debug(struct task_struct *prev)
3145 * Test if we are atomic. Since do_exit() needs to call into
3146 * schedule() atomically, we ignore that path for now.
3147 * Otherwise, whine if we are scheduling when we should not be.
3149 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3150 __schedule_bug(prev);
3153 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3155 schedstat_inc(this_rq(), sched_count);
3158 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3160 if (prev->on_rq || rq->skip_clock_update < 0)
3161 update_rq_clock(rq);
3162 prev->sched_class->put_prev_task(rq, prev);
3166 * Pick up the highest-prio task:
3168 static inline struct task_struct *
3169 pick_next_task(struct rq *rq)
3171 const struct sched_class *class;
3172 struct task_struct *p;
3175 * Optimization: we know that if all tasks are in
3176 * the fair class we can call that function directly:
3178 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3179 p = fair_sched_class.pick_next_task(rq);
3184 for_each_class(class) {
3185 p = class->pick_next_task(rq);
3190 BUG(); /* the idle class will always have a runnable task */
3194 * __schedule() is the main scheduler function.
3196 static void __sched __schedule(void)
3198 struct task_struct *prev, *next;
3199 unsigned long *switch_count;
3205 cpu = smp_processor_id();
3207 rcu_note_context_switch(cpu);
3210 schedule_debug(prev);
3212 if (sched_feat(HRTICK))
3215 raw_spin_lock_irq(&rq->lock);
3217 switch_count = &prev->nivcsw;
3218 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3219 if (unlikely(signal_pending_state(prev->state, prev))) {
3220 prev->state = TASK_RUNNING;
3222 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3226 * If a worker went to sleep, notify and ask workqueue
3227 * whether it wants to wake up a task to maintain
3230 if (prev->flags & PF_WQ_WORKER) {
3231 struct task_struct *to_wakeup;
3233 to_wakeup = wq_worker_sleeping(prev, cpu);
3235 try_to_wake_up_local(to_wakeup);
3238 switch_count = &prev->nvcsw;
3241 pre_schedule(rq, prev);
3243 if (unlikely(!rq->nr_running))
3244 idle_balance(cpu, rq);
3246 put_prev_task(rq, prev);
3247 next = pick_next_task(rq);
3248 clear_tsk_need_resched(prev);
3249 rq->skip_clock_update = 0;
3251 if (likely(prev != next)) {
3256 context_switch(rq, prev, next); /* unlocks the rq */
3258 * The context switch have flipped the stack from under us
3259 * and restored the local variables which were saved when
3260 * this task called schedule() in the past. prev == current
3261 * is still correct, but it can be moved to another cpu/rq.
3263 cpu = smp_processor_id();
3266 raw_spin_unlock_irq(&rq->lock);
3270 sched_preempt_enable_no_resched();
3275 static inline void sched_submit_work(struct task_struct *tsk)
3277 if (!tsk->state || tsk_is_pi_blocked(tsk))
3280 * If we are going to sleep and we have plugged IO queued,
3281 * make sure to submit it to avoid deadlocks.
3283 if (blk_needs_flush_plug(tsk))
3284 blk_schedule_flush_plug(tsk);
3287 asmlinkage void __sched schedule(void)
3289 struct task_struct *tsk = current;
3291 sched_submit_work(tsk);
3294 EXPORT_SYMBOL(schedule);
3297 * schedule_preempt_disabled - called with preemption disabled
3299 * Returns with preemption disabled. Note: preempt_count must be 1
3301 void __sched schedule_preempt_disabled(void)
3303 sched_preempt_enable_no_resched();
3308 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3310 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3312 if (lock->owner != owner)
3316 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3317 * lock->owner still matches owner, if that fails, owner might
3318 * point to free()d memory, if it still matches, the rcu_read_lock()
3319 * ensures the memory stays valid.
3323 return owner->on_cpu;
3327 * Look out! "owner" is an entirely speculative pointer
3328 * access and not reliable.
3330 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3332 if (!sched_feat(OWNER_SPIN))
3336 while (owner_running(lock, owner)) {
3340 arch_mutex_cpu_relax();
3345 * We break out the loop above on need_resched() and when the
3346 * owner changed, which is a sign for heavy contention. Return
3347 * success only when lock->owner is NULL.
3349 return lock->owner == NULL;
3353 #ifdef CONFIG_PREEMPT
3355 * this is the entry point to schedule() from in-kernel preemption
3356 * off of preempt_enable. Kernel preemptions off return from interrupt
3357 * occur there and call schedule directly.
3359 asmlinkage void __sched notrace preempt_schedule(void)
3361 struct thread_info *ti = current_thread_info();
3364 * If there is a non-zero preempt_count or interrupts are disabled,
3365 * we do not want to preempt the current task. Just return..
3367 if (likely(ti->preempt_count || irqs_disabled()))
3371 add_preempt_count_notrace(PREEMPT_ACTIVE);
3373 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3376 * Check again in case we missed a preemption opportunity
3377 * between schedule and now.
3380 } while (need_resched());
3382 EXPORT_SYMBOL(preempt_schedule);
3385 * this is the entry point to schedule() from kernel preemption
3386 * off of irq context.
3387 * Note, that this is called and return with irqs disabled. This will
3388 * protect us against recursive calling from irq.
3390 asmlinkage void __sched preempt_schedule_irq(void)
3392 struct thread_info *ti = current_thread_info();
3394 /* Catch callers which need to be fixed */
3395 BUG_ON(ti->preempt_count || !irqs_disabled());
3398 add_preempt_count(PREEMPT_ACTIVE);
3401 local_irq_disable();
3402 sub_preempt_count(PREEMPT_ACTIVE);
3405 * Check again in case we missed a preemption opportunity
3406 * between schedule and now.
3409 } while (need_resched());
3412 #endif /* CONFIG_PREEMPT */
3414 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3417 return try_to_wake_up(curr->private, mode, wake_flags);
3419 EXPORT_SYMBOL(default_wake_function);
3422 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3423 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3424 * number) then we wake all the non-exclusive tasks and one exclusive task.
3426 * There are circumstances in which we can try to wake a task which has already
3427 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3428 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3430 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3431 int nr_exclusive, int wake_flags, void *key)
3433 wait_queue_t *curr, *next;
3435 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3436 unsigned flags = curr->flags;
3438 if (curr->func(curr, mode, wake_flags, key) &&
3439 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3445 * __wake_up - wake up threads blocked on a waitqueue.
3447 * @mode: which threads
3448 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3449 * @key: is directly passed to the wakeup function
3451 * It may be assumed that this function implies a write memory barrier before
3452 * changing the task state if and only if any tasks are woken up.
3454 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3455 int nr_exclusive, void *key)
3457 unsigned long flags;
3459 spin_lock_irqsave(&q->lock, flags);
3460 __wake_up_common(q, mode, nr_exclusive, 0, key);
3461 spin_unlock_irqrestore(&q->lock, flags);
3463 EXPORT_SYMBOL(__wake_up);
3466 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3468 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3470 __wake_up_common(q, mode, nr, 0, NULL);
3472 EXPORT_SYMBOL_GPL(__wake_up_locked);
3474 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3476 __wake_up_common(q, mode, 1, 0, key);
3478 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3481 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3483 * @mode: which threads
3484 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3485 * @key: opaque value to be passed to wakeup targets
3487 * The sync wakeup differs that the waker knows that it will schedule
3488 * away soon, so while the target thread will be woken up, it will not
3489 * be migrated to another CPU - ie. the two threads are 'synchronized'
3490 * with each other. This can prevent needless bouncing between CPUs.
3492 * On UP it can prevent extra preemption.
3494 * It may be assumed that this function implies a write memory barrier before
3495 * changing the task state if and only if any tasks are woken up.
3497 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3498 int nr_exclusive, void *key)
3500 unsigned long flags;
3501 int wake_flags = WF_SYNC;
3506 if (unlikely(!nr_exclusive))
3509 spin_lock_irqsave(&q->lock, flags);
3510 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3511 spin_unlock_irqrestore(&q->lock, flags);
3513 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3516 * __wake_up_sync - see __wake_up_sync_key()
3518 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3520 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3522 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3525 * complete: - signals a single thread waiting on this completion
3526 * @x: holds the state of this particular completion
3528 * This will wake up a single thread waiting on this completion. Threads will be
3529 * awakened in the same order in which they were queued.
3531 * See also complete_all(), wait_for_completion() and related routines.
3533 * It may be assumed that this function implies a write memory barrier before
3534 * changing the task state if and only if any tasks are woken up.
3536 void complete(struct completion *x)
3538 unsigned long flags;
3540 spin_lock_irqsave(&x->wait.lock, flags);
3542 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3543 spin_unlock_irqrestore(&x->wait.lock, flags);
3545 EXPORT_SYMBOL(complete);
3548 * complete_all: - signals all threads waiting on this completion
3549 * @x: holds the state of this particular completion
3551 * This will wake up all threads waiting on this particular completion event.
3553 * It may be assumed that this function implies a write memory barrier before
3554 * changing the task state if and only if any tasks are woken up.
3556 void complete_all(struct completion *x)
3558 unsigned long flags;
3560 spin_lock_irqsave(&x->wait.lock, flags);
3561 x->done += UINT_MAX/2;
3562 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3563 spin_unlock_irqrestore(&x->wait.lock, flags);
3565 EXPORT_SYMBOL(complete_all);
3567 static inline long __sched
3568 do_wait_for_common(struct completion *x, long timeout, int state)
3571 DECLARE_WAITQUEUE(wait, current);
3573 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3575 if (signal_pending_state(state, current)) {
3576 timeout = -ERESTARTSYS;
3579 __set_current_state(state);
3580 spin_unlock_irq(&x->wait.lock);
3581 timeout = schedule_timeout(timeout);
3582 spin_lock_irq(&x->wait.lock);
3583 } while (!x->done && timeout);
3584 __remove_wait_queue(&x->wait, &wait);
3589 return timeout ?: 1;
3593 wait_for_common(struct completion *x, long timeout, int state)
3597 spin_lock_irq(&x->wait.lock);
3598 timeout = do_wait_for_common(x, timeout, state);
3599 spin_unlock_irq(&x->wait.lock);
3604 * wait_for_completion: - waits for completion of a task
3605 * @x: holds the state of this particular completion
3607 * This waits to be signaled for completion of a specific task. It is NOT
3608 * interruptible and there is no timeout.
3610 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3611 * and interrupt capability. Also see complete().
3613 void __sched wait_for_completion(struct completion *x)
3615 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3617 EXPORT_SYMBOL(wait_for_completion);
3620 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3621 * @x: holds the state of this particular completion
3622 * @timeout: timeout value in jiffies
3624 * This waits for either a completion of a specific task to be signaled or for a
3625 * specified timeout to expire. The timeout is in jiffies. It is not
3628 * The return value is 0 if timed out, and positive (at least 1, or number of
3629 * jiffies left till timeout) if completed.
3631 unsigned long __sched
3632 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3634 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3636 EXPORT_SYMBOL(wait_for_completion_timeout);
3639 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3640 * @x: holds the state of this particular completion
3642 * This waits for completion of a specific task to be signaled. It is
3645 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3647 int __sched wait_for_completion_interruptible(struct completion *x)
3649 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3650 if (t == -ERESTARTSYS)
3654 EXPORT_SYMBOL(wait_for_completion_interruptible);
3657 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3658 * @x: holds the state of this particular completion
3659 * @timeout: timeout value in jiffies
3661 * This waits for either a completion of a specific task to be signaled or for a
3662 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3664 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3665 * positive (at least 1, or number of jiffies left till timeout) if completed.
3668 wait_for_completion_interruptible_timeout(struct completion *x,
3669 unsigned long timeout)
3671 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3673 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3676 * wait_for_completion_killable: - waits for completion of a task (killable)
3677 * @x: holds the state of this particular completion
3679 * This waits to be signaled for completion of a specific task. It can be
3680 * interrupted by a kill signal.
3682 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3684 int __sched wait_for_completion_killable(struct completion *x)
3686 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3687 if (t == -ERESTARTSYS)
3691 EXPORT_SYMBOL(wait_for_completion_killable);
3694 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3695 * @x: holds the state of this particular completion
3696 * @timeout: timeout value in jiffies
3698 * This waits for either a completion of a specific task to be
3699 * signaled or for a specified timeout to expire. It can be
3700 * interrupted by a kill signal. The timeout is in jiffies.
3702 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3703 * positive (at least 1, or number of jiffies left till timeout) if completed.
3706 wait_for_completion_killable_timeout(struct completion *x,
3707 unsigned long timeout)
3709 return wait_for_common(x, timeout, TASK_KILLABLE);
3711 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3714 * try_wait_for_completion - try to decrement a completion without blocking
3715 * @x: completion structure
3717 * Returns: 0 if a decrement cannot be done without blocking
3718 * 1 if a decrement succeeded.
3720 * If a completion is being used as a counting completion,
3721 * attempt to decrement the counter without blocking. This
3722 * enables us to avoid waiting if the resource the completion
3723 * is protecting is not available.
3725 bool try_wait_for_completion(struct completion *x)
3727 unsigned long flags;
3730 spin_lock_irqsave(&x->wait.lock, flags);
3735 spin_unlock_irqrestore(&x->wait.lock, flags);
3738 EXPORT_SYMBOL(try_wait_for_completion);
3741 * completion_done - Test to see if a completion has any waiters
3742 * @x: completion structure
3744 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3745 * 1 if there are no waiters.
3748 bool completion_done(struct completion *x)
3750 unsigned long flags;
3753 spin_lock_irqsave(&x->wait.lock, flags);
3756 spin_unlock_irqrestore(&x->wait.lock, flags);
3759 EXPORT_SYMBOL(completion_done);
3762 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3764 unsigned long flags;
3767 init_waitqueue_entry(&wait, current);
3769 __set_current_state(state);
3771 spin_lock_irqsave(&q->lock, flags);
3772 __add_wait_queue(q, &wait);
3773 spin_unlock(&q->lock);
3774 timeout = schedule_timeout(timeout);
3775 spin_lock_irq(&q->lock);
3776 __remove_wait_queue(q, &wait);
3777 spin_unlock_irqrestore(&q->lock, flags);
3782 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3784 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3786 EXPORT_SYMBOL(interruptible_sleep_on);
3789 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3791 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3793 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3795 void __sched sleep_on(wait_queue_head_t *q)
3797 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3799 EXPORT_SYMBOL(sleep_on);
3801 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3803 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3805 EXPORT_SYMBOL(sleep_on_timeout);
3807 #ifdef CONFIG_RT_MUTEXES
3810 * rt_mutex_setprio - set the current priority of a task
3812 * @prio: prio value (kernel-internal form)
3814 * This function changes the 'effective' priority of a task. It does
3815 * not touch ->normal_prio like __setscheduler().
3817 * Used by the rt_mutex code to implement priority inheritance logic.
3819 void rt_mutex_setprio(struct task_struct *p, int prio)
3821 int oldprio, on_rq, running;
3823 const struct sched_class *prev_class;
3825 BUG_ON(prio < 0 || prio > MAX_PRIO);
3827 rq = __task_rq_lock(p);
3830 * Idle task boosting is a nono in general. There is one
3831 * exception, when PREEMPT_RT and NOHZ is active:
3833 * The idle task calls get_next_timer_interrupt() and holds
3834 * the timer wheel base->lock on the CPU and another CPU wants
3835 * to access the timer (probably to cancel it). We can safely
3836 * ignore the boosting request, as the idle CPU runs this code
3837 * with interrupts disabled and will complete the lock
3838 * protected section without being interrupted. So there is no
3839 * real need to boost.
3841 if (unlikely(p == rq->idle)) {
3842 WARN_ON(p != rq->curr);
3843 WARN_ON(p->pi_blocked_on);
3847 trace_sched_pi_setprio(p, prio);
3849 prev_class = p->sched_class;
3851 running = task_current(rq, p);
3853 dequeue_task(rq, p, 0);
3855 p->sched_class->put_prev_task(rq, p);
3858 p->sched_class = &rt_sched_class;
3860 p->sched_class = &fair_sched_class;
3865 p->sched_class->set_curr_task(rq);
3867 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3869 check_class_changed(rq, p, prev_class, oldprio);
3871 __task_rq_unlock(rq);
3874 void set_user_nice(struct task_struct *p, long nice)
3876 int old_prio, delta, on_rq;
3877 unsigned long flags;
3880 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3883 * We have to be careful, if called from sys_setpriority(),
3884 * the task might be in the middle of scheduling on another CPU.
3886 rq = task_rq_lock(p, &flags);
3888 * The RT priorities are set via sched_setscheduler(), but we still
3889 * allow the 'normal' nice value to be set - but as expected
3890 * it wont have any effect on scheduling until the task is
3891 * SCHED_FIFO/SCHED_RR:
3893 if (task_has_rt_policy(p)) {
3894 p->static_prio = NICE_TO_PRIO(nice);
3899 dequeue_task(rq, p, 0);
3901 p->static_prio = NICE_TO_PRIO(nice);
3904 p->prio = effective_prio(p);
3905 delta = p->prio - old_prio;
3908 enqueue_task(rq, p, 0);
3910 * If the task increased its priority or is running and
3911 * lowered its priority, then reschedule its CPU:
3913 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3914 resched_task(rq->curr);
3917 task_rq_unlock(rq, p, &flags);
3919 EXPORT_SYMBOL(set_user_nice);
3922 * can_nice - check if a task can reduce its nice value
3926 int can_nice(const struct task_struct *p, const int nice)
3928 /* convert nice value [19,-20] to rlimit style value [1,40] */
3929 int nice_rlim = 20 - nice;
3931 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3932 capable(CAP_SYS_NICE));
3935 #ifdef __ARCH_WANT_SYS_NICE
3938 * sys_nice - change the priority of the current process.
3939 * @increment: priority increment
3941 * sys_setpriority is a more generic, but much slower function that
3942 * does similar things.
3944 SYSCALL_DEFINE1(nice, int, increment)
3949 * Setpriority might change our priority at the same moment.
3950 * We don't have to worry. Conceptually one call occurs first
3951 * and we have a single winner.
3953 if (increment < -40)
3958 nice = TASK_NICE(current) + increment;
3964 if (increment < 0 && !can_nice(current, nice))
3967 retval = security_task_setnice(current, nice);
3971 set_user_nice(current, nice);
3978 * task_prio - return the priority value of a given task.
3979 * @p: the task in question.
3981 * This is the priority value as seen by users in /proc.
3982 * RT tasks are offset by -200. Normal tasks are centered
3983 * around 0, value goes from -16 to +15.
3985 int task_prio(const struct task_struct *p)
3987 return p->prio - MAX_RT_PRIO;
3991 * task_nice - return the nice value of a given task.
3992 * @p: the task in question.
3994 int task_nice(const struct task_struct *p)
3996 return TASK_NICE(p);
3998 EXPORT_SYMBOL(task_nice);
4001 * idle_cpu - is a given cpu idle currently?
4002 * @cpu: the processor in question.
4004 int idle_cpu(int cpu)
4006 struct rq *rq = cpu_rq(cpu);
4008 if (rq->curr != rq->idle)
4015 if (!llist_empty(&rq->wake_list))
4023 * idle_task - return the idle task for a given cpu.
4024 * @cpu: the processor in question.
4026 struct task_struct *idle_task(int cpu)
4028 return cpu_rq(cpu)->idle;
4032 * find_process_by_pid - find a process with a matching PID value.
4033 * @pid: the pid in question.
4035 static struct task_struct *find_process_by_pid(pid_t pid)
4037 return pid ? find_task_by_vpid(pid) : current;
4040 /* Actually do priority change: must hold rq lock. */
4042 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4045 p->rt_priority = prio;
4046 p->normal_prio = normal_prio(p);
4047 /* we are holding p->pi_lock already */
4048 p->prio = rt_mutex_getprio(p);
4049 if (rt_prio(p->prio))
4050 p->sched_class = &rt_sched_class;
4052 p->sched_class = &fair_sched_class;
4057 * check the target process has a UID that matches the current process's
4059 static bool check_same_owner(struct task_struct *p)
4061 const struct cred *cred = current_cred(), *pcred;
4065 pcred = __task_cred(p);
4066 if (cred->user->user_ns == pcred->user->user_ns)
4067 match = (cred->euid == pcred->euid ||
4068 cred->euid == pcred->uid);
4075 static int __sched_setscheduler(struct task_struct *p, int policy,
4076 const struct sched_param *param, bool user)
4078 int retval, oldprio, oldpolicy = -1, on_rq, running;
4079 unsigned long flags;
4080 const struct sched_class *prev_class;
4084 /* may grab non-irq protected spin_locks */
4085 BUG_ON(in_interrupt());
4087 /* double check policy once rq lock held */
4089 reset_on_fork = p->sched_reset_on_fork;
4090 policy = oldpolicy = p->policy;
4092 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4093 policy &= ~SCHED_RESET_ON_FORK;
4095 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4096 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4097 policy != SCHED_IDLE)
4102 * Valid priorities for SCHED_FIFO and SCHED_RR are
4103 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4104 * SCHED_BATCH and SCHED_IDLE is 0.
4106 if (param->sched_priority < 0 ||
4107 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4108 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4110 if (rt_policy(policy) != (param->sched_priority != 0))
4114 * Allow unprivileged RT tasks to decrease priority:
4116 if (user && !capable(CAP_SYS_NICE)) {
4117 if (rt_policy(policy)) {
4118 unsigned long rlim_rtprio =
4119 task_rlimit(p, RLIMIT_RTPRIO);
4121 /* can't set/change the rt policy */
4122 if (policy != p->policy && !rlim_rtprio)
4125 /* can't increase priority */
4126 if (param->sched_priority > p->rt_priority &&
4127 param->sched_priority > rlim_rtprio)
4132 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4133 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4135 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4136 if (!can_nice(p, TASK_NICE(p)))
4140 /* can't change other user's priorities */
4141 if (!check_same_owner(p))
4144 /* Normal users shall not reset the sched_reset_on_fork flag */
4145 if (p->sched_reset_on_fork && !reset_on_fork)
4150 retval = security_task_setscheduler(p);
4156 * make sure no PI-waiters arrive (or leave) while we are
4157 * changing the priority of the task:
4159 * To be able to change p->policy safely, the appropriate
4160 * runqueue lock must be held.
4162 rq = task_rq_lock(p, &flags);
4165 * Changing the policy of the stop threads its a very bad idea
4167 if (p == rq->stop) {
4168 task_rq_unlock(rq, p, &flags);
4173 * If not changing anything there's no need to proceed further:
4175 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4176 param->sched_priority == p->rt_priority))) {
4178 __task_rq_unlock(rq);
4179 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4183 #ifdef CONFIG_RT_GROUP_SCHED
4186 * Do not allow realtime tasks into groups that have no runtime
4189 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4190 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4191 !task_group_is_autogroup(task_group(p))) {
4192 task_rq_unlock(rq, p, &flags);
4198 /* recheck policy now with rq lock held */
4199 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4200 policy = oldpolicy = -1;
4201 task_rq_unlock(rq, p, &flags);
4205 running = task_current(rq, p);
4207 dequeue_task(rq, p, 0);
4209 p->sched_class->put_prev_task(rq, p);
4211 p->sched_reset_on_fork = reset_on_fork;
4214 prev_class = p->sched_class;
4215 __setscheduler(rq, p, policy, param->sched_priority);
4218 p->sched_class->set_curr_task(rq);
4220 enqueue_task(rq, p, 0);
4222 check_class_changed(rq, p, prev_class, oldprio);
4223 task_rq_unlock(rq, p, &flags);
4225 rt_mutex_adjust_pi(p);
4231 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4232 * @p: the task in question.
4233 * @policy: new policy.
4234 * @param: structure containing the new RT priority.
4236 * NOTE that the task may be already dead.
4238 int sched_setscheduler(struct task_struct *p, int policy,
4239 const struct sched_param *param)
4241 return __sched_setscheduler(p, policy, param, true);
4243 EXPORT_SYMBOL_GPL(sched_setscheduler);
4246 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4247 * @p: the task in question.
4248 * @policy: new policy.
4249 * @param: structure containing the new RT priority.
4251 * Just like sched_setscheduler, only don't bother checking if the
4252 * current context has permission. For example, this is needed in
4253 * stop_machine(): we create temporary high priority worker threads,
4254 * but our caller might not have that capability.
4256 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4257 const struct sched_param *param)
4259 return __sched_setscheduler(p, policy, param, false);
4263 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4265 struct sched_param lparam;
4266 struct task_struct *p;
4269 if (!param || pid < 0)
4271 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4276 p = find_process_by_pid(pid);
4278 retval = sched_setscheduler(p, policy, &lparam);
4285 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4286 * @pid: the pid in question.
4287 * @policy: new policy.
4288 * @param: structure containing the new RT priority.
4290 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4291 struct sched_param __user *, param)
4293 /* negative values for policy are not valid */
4297 return do_sched_setscheduler(pid, policy, param);
4301 * sys_sched_setparam - set/change the RT priority of a thread
4302 * @pid: the pid in question.
4303 * @param: structure containing the new RT priority.
4305 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4307 return do_sched_setscheduler(pid, -1, param);
4311 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4312 * @pid: the pid in question.
4314 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4316 struct task_struct *p;
4324 p = find_process_by_pid(pid);
4326 retval = security_task_getscheduler(p);
4329 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4336 * sys_sched_getparam - get the RT priority of a thread
4337 * @pid: the pid in question.
4338 * @param: structure containing the RT priority.
4340 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4342 struct sched_param lp;
4343 struct task_struct *p;
4346 if (!param || pid < 0)
4350 p = find_process_by_pid(pid);
4355 retval = security_task_getscheduler(p);
4359 lp.sched_priority = p->rt_priority;
4363 * This one might sleep, we cannot do it with a spinlock held ...
4365 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4374 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4376 cpumask_var_t cpus_allowed, new_mask;
4377 struct task_struct *p;
4383 p = find_process_by_pid(pid);
4390 /* Prevent p going away */
4394 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4398 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4400 goto out_free_cpus_allowed;
4403 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4406 retval = security_task_setscheduler(p);
4410 cpuset_cpus_allowed(p, cpus_allowed);
4411 cpumask_and(new_mask, in_mask, cpus_allowed);
4413 retval = set_cpus_allowed_ptr(p, new_mask);
4416 cpuset_cpus_allowed(p, cpus_allowed);
4417 if (!cpumask_subset(new_mask, cpus_allowed)) {
4419 * We must have raced with a concurrent cpuset
4420 * update. Just reset the cpus_allowed to the
4421 * cpuset's cpus_allowed
4423 cpumask_copy(new_mask, cpus_allowed);
4428 free_cpumask_var(new_mask);
4429 out_free_cpus_allowed:
4430 free_cpumask_var(cpus_allowed);
4437 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4438 struct cpumask *new_mask)
4440 if (len < cpumask_size())
4441 cpumask_clear(new_mask);
4442 else if (len > cpumask_size())
4443 len = cpumask_size();
4445 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4449 * sys_sched_setaffinity - set the cpu affinity of a process
4450 * @pid: pid of the process
4451 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4452 * @user_mask_ptr: user-space pointer to the new cpu mask
4454 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4455 unsigned long __user *, user_mask_ptr)
4457 cpumask_var_t new_mask;
4460 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4463 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4465 retval = sched_setaffinity(pid, new_mask);
4466 free_cpumask_var(new_mask);
4470 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4472 struct task_struct *p;
4473 unsigned long flags;
4480 p = find_process_by_pid(pid);
4484 retval = security_task_getscheduler(p);
4488 raw_spin_lock_irqsave(&p->pi_lock, flags);
4489 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4490 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4500 * sys_sched_getaffinity - get the cpu affinity of a process
4501 * @pid: pid of the process
4502 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4503 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4505 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4506 unsigned long __user *, user_mask_ptr)
4511 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4513 if (len & (sizeof(unsigned long)-1))
4516 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4519 ret = sched_getaffinity(pid, mask);
4521 size_t retlen = min_t(size_t, len, cpumask_size());
4523 if (copy_to_user(user_mask_ptr, mask, retlen))
4528 free_cpumask_var(mask);
4534 * sys_sched_yield - yield the current processor to other threads.
4536 * This function yields the current CPU to other tasks. If there are no
4537 * other threads running on this CPU then this function will return.
4539 SYSCALL_DEFINE0(sched_yield)
4541 struct rq *rq = this_rq_lock();
4543 schedstat_inc(rq, yld_count);
4544 current->sched_class->yield_task(rq);
4547 * Since we are going to call schedule() anyway, there's
4548 * no need to preempt or enable interrupts:
4550 __release(rq->lock);
4551 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4552 do_raw_spin_unlock(&rq->lock);
4553 sched_preempt_enable_no_resched();
4560 static inline int should_resched(void)
4562 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4565 static void __cond_resched(void)
4567 add_preempt_count(PREEMPT_ACTIVE);
4569 sub_preempt_count(PREEMPT_ACTIVE);
4572 int __sched _cond_resched(void)
4574 if (should_resched()) {
4580 EXPORT_SYMBOL(_cond_resched);
4583 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4584 * call schedule, and on return reacquire the lock.
4586 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4587 * operations here to prevent schedule() from being called twice (once via
4588 * spin_unlock(), once by hand).
4590 int __cond_resched_lock(spinlock_t *lock)
4592 int resched = should_resched();
4595 lockdep_assert_held(lock);
4597 if (spin_needbreak(lock) || resched) {
4608 EXPORT_SYMBOL(__cond_resched_lock);
4610 int __sched __cond_resched_softirq(void)
4612 BUG_ON(!in_softirq());
4614 if (should_resched()) {
4622 EXPORT_SYMBOL(__cond_resched_softirq);
4625 * yield - yield the current processor to other threads.
4627 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4629 * The scheduler is at all times free to pick the calling task as the most
4630 * eligible task to run, if removing the yield() call from your code breaks
4631 * it, its already broken.
4633 * Typical broken usage is:
4638 * where one assumes that yield() will let 'the other' process run that will
4639 * make event true. If the current task is a SCHED_FIFO task that will never
4640 * happen. Never use yield() as a progress guarantee!!
4642 * If you want to use yield() to wait for something, use wait_event().
4643 * If you want to use yield() to be 'nice' for others, use cond_resched().
4644 * If you still want to use yield(), do not!
4646 void __sched yield(void)
4648 set_current_state(TASK_RUNNING);
4651 EXPORT_SYMBOL(yield);
4654 * yield_to - yield the current processor to another thread in
4655 * your thread group, or accelerate that thread toward the
4656 * processor it's on.
4658 * @preempt: whether task preemption is allowed or not
4660 * It's the caller's job to ensure that the target task struct
4661 * can't go away on us before we can do any checks.
4663 * Returns true if we indeed boosted the target task.
4665 bool __sched yield_to(struct task_struct *p, bool preempt)
4667 struct task_struct *curr = current;
4668 struct rq *rq, *p_rq;
4669 unsigned long flags;
4672 local_irq_save(flags);
4677 double_rq_lock(rq, p_rq);
4678 while (task_rq(p) != p_rq) {
4679 double_rq_unlock(rq, p_rq);
4683 if (!curr->sched_class->yield_to_task)
4686 if (curr->sched_class != p->sched_class)
4689 if (task_running(p_rq, p) || p->state)
4692 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4694 schedstat_inc(rq, yld_count);
4696 * Make p's CPU reschedule; pick_next_entity takes care of
4699 if (preempt && rq != p_rq)
4700 resched_task(p_rq->curr);
4703 * We might have set it in task_yield_fair(), but are
4704 * not going to schedule(), so don't want to skip
4707 rq->skip_clock_update = 0;
4711 double_rq_unlock(rq, p_rq);
4712 local_irq_restore(flags);
4719 EXPORT_SYMBOL_GPL(yield_to);
4722 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4723 * that process accounting knows that this is a task in IO wait state.
4725 void __sched io_schedule(void)
4727 struct rq *rq = raw_rq();
4729 delayacct_blkio_start();
4730 atomic_inc(&rq->nr_iowait);
4731 blk_flush_plug(current);
4732 current->in_iowait = 1;
4734 current->in_iowait = 0;
4735 atomic_dec(&rq->nr_iowait);
4736 delayacct_blkio_end();
4738 EXPORT_SYMBOL(io_schedule);
4740 long __sched io_schedule_timeout(long timeout)
4742 struct rq *rq = raw_rq();
4745 delayacct_blkio_start();
4746 atomic_inc(&rq->nr_iowait);
4747 blk_flush_plug(current);
4748 current->in_iowait = 1;
4749 ret = schedule_timeout(timeout);
4750 current->in_iowait = 0;
4751 atomic_dec(&rq->nr_iowait);
4752 delayacct_blkio_end();
4757 * sys_sched_get_priority_max - return maximum RT priority.
4758 * @policy: scheduling class.
4760 * this syscall returns the maximum rt_priority that can be used
4761 * by a given scheduling class.
4763 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4770 ret = MAX_USER_RT_PRIO-1;
4782 * sys_sched_get_priority_min - return minimum RT priority.
4783 * @policy: scheduling class.
4785 * this syscall returns the minimum rt_priority that can be used
4786 * by a given scheduling class.
4788 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4806 * sys_sched_rr_get_interval - return the default timeslice of a process.
4807 * @pid: pid of the process.
4808 * @interval: userspace pointer to the timeslice value.
4810 * this syscall writes the default timeslice value of a given process
4811 * into the user-space timespec buffer. A value of '0' means infinity.
4813 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4814 struct timespec __user *, interval)
4816 struct task_struct *p;
4817 unsigned int time_slice;
4818 unsigned long flags;
4828 p = find_process_by_pid(pid);
4832 retval = security_task_getscheduler(p);
4836 rq = task_rq_lock(p, &flags);
4837 time_slice = p->sched_class->get_rr_interval(rq, p);
4838 task_rq_unlock(rq, p, &flags);
4841 jiffies_to_timespec(time_slice, &t);
4842 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4850 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4852 void sched_show_task(struct task_struct *p)
4854 unsigned long free = 0;
4857 state = p->state ? __ffs(p->state) + 1 : 0;
4858 printk(KERN_INFO "%-15.15s %c", p->comm,
4859 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4860 #if BITS_PER_LONG == 32
4861 if (state == TASK_RUNNING)
4862 printk(KERN_CONT " running ");
4864 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4866 if (state == TASK_RUNNING)
4867 printk(KERN_CONT " running task ");
4869 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4871 #ifdef CONFIG_DEBUG_STACK_USAGE
4872 free = stack_not_used(p);
4874 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4875 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4876 (unsigned long)task_thread_info(p)->flags);
4878 show_stack(p, NULL);
4881 void show_state_filter(unsigned long state_filter)
4883 struct task_struct *g, *p;
4885 #if BITS_PER_LONG == 32
4887 " task PC stack pid father\n");
4890 " task PC stack pid father\n");
4893 do_each_thread(g, p) {
4895 * reset the NMI-timeout, listing all files on a slow
4896 * console might take a lot of time:
4898 touch_nmi_watchdog();
4899 if (!state_filter || (p->state & state_filter))
4901 } while_each_thread(g, p);
4903 touch_all_softlockup_watchdogs();
4905 #ifdef CONFIG_SCHED_DEBUG
4906 sysrq_sched_debug_show();
4910 * Only show locks if all tasks are dumped:
4913 debug_show_all_locks();
4916 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4918 idle->sched_class = &idle_sched_class;
4922 * init_idle - set up an idle thread for a given CPU
4923 * @idle: task in question
4924 * @cpu: cpu the idle task belongs to
4926 * NOTE: this function does not set the idle thread's NEED_RESCHED
4927 * flag, to make booting more robust.
4929 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4931 struct rq *rq = cpu_rq(cpu);
4932 unsigned long flags;
4934 raw_spin_lock_irqsave(&rq->lock, flags);
4937 idle->state = TASK_RUNNING;
4938 idle->se.exec_start = sched_clock();
4940 do_set_cpus_allowed(idle, cpumask_of(cpu));
4942 * We're having a chicken and egg problem, even though we are
4943 * holding rq->lock, the cpu isn't yet set to this cpu so the
4944 * lockdep check in task_group() will fail.
4946 * Similar case to sched_fork(). / Alternatively we could
4947 * use task_rq_lock() here and obtain the other rq->lock.
4952 __set_task_cpu(idle, cpu);
4955 rq->curr = rq->idle = idle;
4956 #if defined(CONFIG_SMP)
4959 raw_spin_unlock_irqrestore(&rq->lock, flags);
4961 /* Set the preempt count _outside_ the spinlocks! */
4962 task_thread_info(idle)->preempt_count = 0;
4965 * The idle tasks have their own, simple scheduling class:
4967 idle->sched_class = &idle_sched_class;
4968 ftrace_graph_init_idle_task(idle, cpu);
4969 #if defined(CONFIG_SMP)
4970 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4975 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4977 if (p->sched_class && p->sched_class->set_cpus_allowed)
4978 p->sched_class->set_cpus_allowed(p, new_mask);
4980 cpumask_copy(&p->cpus_allowed, new_mask);
4981 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4985 * This is how migration works:
4987 * 1) we invoke migration_cpu_stop() on the target CPU using
4989 * 2) stopper starts to run (implicitly forcing the migrated thread
4991 * 3) it checks whether the migrated task is still in the wrong runqueue.
4992 * 4) if it's in the wrong runqueue then the migration thread removes
4993 * it and puts it into the right queue.
4994 * 5) stopper completes and stop_one_cpu() returns and the migration
4999 * Change a given task's CPU affinity. Migrate the thread to a
5000 * proper CPU and schedule it away if the CPU it's executing on
5001 * is removed from the allowed bitmask.
5003 * NOTE: the caller must have a valid reference to the task, the
5004 * task must not exit() & deallocate itself prematurely. The
5005 * call is not atomic; no spinlocks may be held.
5007 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5009 unsigned long flags;
5011 unsigned int dest_cpu;
5014 rq = task_rq_lock(p, &flags);
5016 if (cpumask_equal(&p->cpus_allowed, new_mask))
5019 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5024 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5029 do_set_cpus_allowed(p, new_mask);
5031 /* Can the task run on the task's current CPU? If so, we're done */
5032 if (cpumask_test_cpu(task_cpu(p), new_mask))
5035 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5037 struct migration_arg arg = { p, dest_cpu };
5038 /* Need help from migration thread: drop lock and wait. */
5039 task_rq_unlock(rq, p, &flags);
5040 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5041 tlb_migrate_finish(p->mm);
5045 task_rq_unlock(rq, p, &flags);
5049 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5052 * Move (not current) task off this cpu, onto dest cpu. We're doing
5053 * this because either it can't run here any more (set_cpus_allowed()
5054 * away from this CPU, or CPU going down), or because we're
5055 * attempting to rebalance this task on exec (sched_exec).
5057 * So we race with normal scheduler movements, but that's OK, as long
5058 * as the task is no longer on this CPU.
5060 * Returns non-zero if task was successfully migrated.
5062 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5064 struct rq *rq_dest, *rq_src;
5067 if (unlikely(!cpu_active(dest_cpu)))
5070 rq_src = cpu_rq(src_cpu);
5071 rq_dest = cpu_rq(dest_cpu);
5073 raw_spin_lock(&p->pi_lock);
5074 double_rq_lock(rq_src, rq_dest);
5075 /* Already moved. */
5076 if (task_cpu(p) != src_cpu)
5078 /* Affinity changed (again). */
5079 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5083 * If we're not on a rq, the next wake-up will ensure we're
5087 dequeue_task(rq_src, p, 0);
5088 set_task_cpu(p, dest_cpu);
5089 enqueue_task(rq_dest, p, 0);
5090 check_preempt_curr(rq_dest, p, 0);
5095 double_rq_unlock(rq_src, rq_dest);
5096 raw_spin_unlock(&p->pi_lock);
5101 * migration_cpu_stop - this will be executed by a highprio stopper thread
5102 * and performs thread migration by bumping thread off CPU then
5103 * 'pushing' onto another runqueue.
5105 static int migration_cpu_stop(void *data)
5107 struct migration_arg *arg = data;
5110 * The original target cpu might have gone down and we might
5111 * be on another cpu but it doesn't matter.
5113 local_irq_disable();
5114 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5119 #ifdef CONFIG_HOTPLUG_CPU
5122 * Ensures that the idle task is using init_mm right before its cpu goes
5125 void idle_task_exit(void)
5127 struct mm_struct *mm = current->active_mm;
5129 BUG_ON(cpu_online(smp_processor_id()));
5132 switch_mm(mm, &init_mm, current);
5137 * While a dead CPU has no uninterruptible tasks queued at this point,
5138 * it might still have a nonzero ->nr_uninterruptible counter, because
5139 * for performance reasons the counter is not stricly tracking tasks to
5140 * their home CPUs. So we just add the counter to another CPU's counter,
5141 * to keep the global sum constant after CPU-down:
5143 static void migrate_nr_uninterruptible(struct rq *rq_src)
5145 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5147 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5148 rq_src->nr_uninterruptible = 0;
5152 * remove the tasks which were accounted by rq from calc_load_tasks.
5154 static void calc_global_load_remove(struct rq *rq)
5156 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5157 rq->calc_load_active = 0;
5161 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5162 * try_to_wake_up()->select_task_rq().
5164 * Called with rq->lock held even though we'er in stop_machine() and
5165 * there's no concurrency possible, we hold the required locks anyway
5166 * because of lock validation efforts.
5168 static void migrate_tasks(unsigned int dead_cpu)
5170 struct rq *rq = cpu_rq(dead_cpu);
5171 struct task_struct *next, *stop = rq->stop;
5175 * Fudge the rq selection such that the below task selection loop
5176 * doesn't get stuck on the currently eligible stop task.
5178 * We're currently inside stop_machine() and the rq is either stuck
5179 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5180 * either way we should never end up calling schedule() until we're
5185 /* Ensure any throttled groups are reachable by pick_next_task */
5186 unthrottle_offline_cfs_rqs(rq);
5190 * There's this thread running, bail when that's the only
5193 if (rq->nr_running == 1)
5196 next = pick_next_task(rq);
5198 next->sched_class->put_prev_task(rq, next);
5200 /* Find suitable destination for @next, with force if needed. */
5201 dest_cpu = select_fallback_rq(dead_cpu, next);
5202 raw_spin_unlock(&rq->lock);
5204 __migrate_task(next, dead_cpu, dest_cpu);
5206 raw_spin_lock(&rq->lock);
5212 #endif /* CONFIG_HOTPLUG_CPU */
5214 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5216 static struct ctl_table sd_ctl_dir[] = {
5218 .procname = "sched_domain",
5224 static struct ctl_table sd_ctl_root[] = {
5226 .procname = "kernel",
5228 .child = sd_ctl_dir,
5233 static struct ctl_table *sd_alloc_ctl_entry(int n)
5235 struct ctl_table *entry =
5236 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5241 static void sd_free_ctl_entry(struct ctl_table **tablep)
5243 struct ctl_table *entry;
5246 * In the intermediate directories, both the child directory and
5247 * procname are dynamically allocated and could fail but the mode
5248 * will always be set. In the lowest directory the names are
5249 * static strings and all have proc handlers.
5251 for (entry = *tablep; entry->mode; entry++) {
5253 sd_free_ctl_entry(&entry->child);
5254 if (entry->proc_handler == NULL)
5255 kfree(entry->procname);
5263 set_table_entry(struct ctl_table *entry,
5264 const char *procname, void *data, int maxlen,
5265 umode_t mode, proc_handler *proc_handler)
5267 entry->procname = procname;
5269 entry->maxlen = maxlen;
5271 entry->proc_handler = proc_handler;
5274 static struct ctl_table *
5275 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5277 struct ctl_table *table = sd_alloc_ctl_entry(13);
5282 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5283 sizeof(long), 0644, proc_doulongvec_minmax);
5284 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5285 sizeof(long), 0644, proc_doulongvec_minmax);
5286 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5287 sizeof(int), 0644, proc_dointvec_minmax);
5288 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5289 sizeof(int), 0644, proc_dointvec_minmax);
5290 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5291 sizeof(int), 0644, proc_dointvec_minmax);
5292 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5293 sizeof(int), 0644, proc_dointvec_minmax);
5294 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5295 sizeof(int), 0644, proc_dointvec_minmax);
5296 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5297 sizeof(int), 0644, proc_dointvec_minmax);
5298 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5299 sizeof(int), 0644, proc_dointvec_minmax);
5300 set_table_entry(&table[9], "cache_nice_tries",
5301 &sd->cache_nice_tries,
5302 sizeof(int), 0644, proc_dointvec_minmax);
5303 set_table_entry(&table[10], "flags", &sd->flags,
5304 sizeof(int), 0644, proc_dointvec_minmax);
5305 set_table_entry(&table[11], "name", sd->name,
5306 CORENAME_MAX_SIZE, 0444, proc_dostring);
5307 /* &table[12] is terminator */
5312 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5314 struct ctl_table *entry, *table;
5315 struct sched_domain *sd;
5316 int domain_num = 0, i;
5319 for_each_domain(cpu, sd)
5321 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5326 for_each_domain(cpu, sd) {
5327 snprintf(buf, 32, "domain%d", i);
5328 entry->procname = kstrdup(buf, GFP_KERNEL);
5330 entry->child = sd_alloc_ctl_domain_table(sd);
5337 static struct ctl_table_header *sd_sysctl_header;
5338 static void register_sched_domain_sysctl(void)
5340 int i, cpu_num = num_possible_cpus();
5341 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5344 WARN_ON(sd_ctl_dir[0].child);
5345 sd_ctl_dir[0].child = entry;
5350 for_each_possible_cpu(i) {
5351 snprintf(buf, 32, "cpu%d", i);
5352 entry->procname = kstrdup(buf, GFP_KERNEL);
5354 entry->child = sd_alloc_ctl_cpu_table(i);
5358 WARN_ON(sd_sysctl_header);
5359 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5362 /* may be called multiple times per register */
5363 static void unregister_sched_domain_sysctl(void)
5365 if (sd_sysctl_header)
5366 unregister_sysctl_table(sd_sysctl_header);
5367 sd_sysctl_header = NULL;
5368 if (sd_ctl_dir[0].child)
5369 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5372 static void register_sched_domain_sysctl(void)
5375 static void unregister_sched_domain_sysctl(void)
5380 static void set_rq_online(struct rq *rq)
5383 const struct sched_class *class;
5385 cpumask_set_cpu(rq->cpu, rq->rd->online);
5388 for_each_class(class) {
5389 if (class->rq_online)
5390 class->rq_online(rq);
5395 static void set_rq_offline(struct rq *rq)
5398 const struct sched_class *class;
5400 for_each_class(class) {
5401 if (class->rq_offline)
5402 class->rq_offline(rq);
5405 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5411 * migration_call - callback that gets triggered when a CPU is added.
5412 * Here we can start up the necessary migration thread for the new CPU.
5414 static int __cpuinit
5415 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5417 int cpu = (long)hcpu;
5418 unsigned long flags;
5419 struct rq *rq = cpu_rq(cpu);
5421 switch (action & ~CPU_TASKS_FROZEN) {
5423 case CPU_UP_PREPARE:
5424 rq->calc_load_update = calc_load_update;
5428 /* Update our root-domain */
5429 raw_spin_lock_irqsave(&rq->lock, flags);
5431 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5435 raw_spin_unlock_irqrestore(&rq->lock, flags);
5438 #ifdef CONFIG_HOTPLUG_CPU
5440 sched_ttwu_pending();
5441 /* Update our root-domain */
5442 raw_spin_lock_irqsave(&rq->lock, flags);
5444 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5448 BUG_ON(rq->nr_running != 1); /* the migration thread */
5449 raw_spin_unlock_irqrestore(&rq->lock, flags);
5451 migrate_nr_uninterruptible(rq);
5452 calc_global_load_remove(rq);
5457 update_max_interval();
5463 * Register at high priority so that task migration (migrate_all_tasks)
5464 * happens before everything else. This has to be lower priority than
5465 * the notifier in the perf_event subsystem, though.
5467 static struct notifier_block __cpuinitdata migration_notifier = {
5468 .notifier_call = migration_call,
5469 .priority = CPU_PRI_MIGRATION,
5472 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5473 unsigned long action, void *hcpu)
5475 switch (action & ~CPU_TASKS_FROZEN) {
5477 case CPU_DOWN_FAILED:
5478 set_cpu_active((long)hcpu, true);
5485 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5486 unsigned long action, void *hcpu)
5488 switch (action & ~CPU_TASKS_FROZEN) {
5489 case CPU_DOWN_PREPARE:
5490 set_cpu_active((long)hcpu, false);
5497 static int __init migration_init(void)
5499 void *cpu = (void *)(long)smp_processor_id();
5502 /* Initialize migration for the boot CPU */
5503 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5504 BUG_ON(err == NOTIFY_BAD);
5505 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5506 register_cpu_notifier(&migration_notifier);
5508 /* Register cpu active notifiers */
5509 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5510 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5514 early_initcall(migration_init);
5519 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5521 #ifdef CONFIG_SCHED_DEBUG
5523 static __read_mostly int sched_domain_debug_enabled;
5525 static int __init sched_domain_debug_setup(char *str)
5527 sched_domain_debug_enabled = 1;
5531 early_param("sched_debug", sched_domain_debug_setup);
5533 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5534 struct cpumask *groupmask)
5536 struct sched_group *group = sd->groups;
5539 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5540 cpumask_clear(groupmask);
5542 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5544 if (!(sd->flags & SD_LOAD_BALANCE)) {
5545 printk("does not load-balance\n");
5547 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5552 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5554 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5555 printk(KERN_ERR "ERROR: domain->span does not contain "
5558 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5559 printk(KERN_ERR "ERROR: domain->groups does not contain"
5563 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5567 printk(KERN_ERR "ERROR: group is NULL\n");
5571 if (!group->sgp->power) {
5572 printk(KERN_CONT "\n");
5573 printk(KERN_ERR "ERROR: domain->cpu_power not "
5578 if (!cpumask_weight(sched_group_cpus(group))) {
5579 printk(KERN_CONT "\n");
5580 printk(KERN_ERR "ERROR: empty group\n");
5584 if (!(sd->flags & SD_OVERLAP) &&
5585 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5586 printk(KERN_CONT "\n");
5587 printk(KERN_ERR "ERROR: repeated CPUs\n");
5591 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5593 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5595 printk(KERN_CONT " %s", str);
5596 if (group->sgp->power != SCHED_POWER_SCALE) {
5597 printk(KERN_CONT " (cpu_power = %d)",
5601 group = group->next;
5602 } while (group != sd->groups);
5603 printk(KERN_CONT "\n");
5605 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5606 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5609 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5610 printk(KERN_ERR "ERROR: parent span is not a superset "
5611 "of domain->span\n");
5615 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5619 if (!sched_domain_debug_enabled)
5623 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5627 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5630 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5638 #else /* !CONFIG_SCHED_DEBUG */
5639 # define sched_domain_debug(sd, cpu) do { } while (0)
5640 #endif /* CONFIG_SCHED_DEBUG */
5642 static int sd_degenerate(struct sched_domain *sd)
5644 if (cpumask_weight(sched_domain_span(sd)) == 1)
5647 /* Following flags need at least 2 groups */
5648 if (sd->flags & (SD_LOAD_BALANCE |
5649 SD_BALANCE_NEWIDLE |
5653 SD_SHARE_PKG_RESOURCES)) {
5654 if (sd->groups != sd->groups->next)
5658 /* Following flags don't use groups */
5659 if (sd->flags & (SD_WAKE_AFFINE))
5666 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5668 unsigned long cflags = sd->flags, pflags = parent->flags;
5670 if (sd_degenerate(parent))
5673 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5676 /* Flags needing groups don't count if only 1 group in parent */
5677 if (parent->groups == parent->groups->next) {
5678 pflags &= ~(SD_LOAD_BALANCE |
5679 SD_BALANCE_NEWIDLE |
5683 SD_SHARE_PKG_RESOURCES);
5684 if (nr_node_ids == 1)
5685 pflags &= ~SD_SERIALIZE;
5687 if (~cflags & pflags)
5693 static void free_rootdomain(struct rcu_head *rcu)
5695 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5697 cpupri_cleanup(&rd->cpupri);
5698 free_cpumask_var(rd->rto_mask);
5699 free_cpumask_var(rd->online);
5700 free_cpumask_var(rd->span);
5704 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5706 struct root_domain *old_rd = NULL;
5707 unsigned long flags;
5709 raw_spin_lock_irqsave(&rq->lock, flags);
5714 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5717 cpumask_clear_cpu(rq->cpu, old_rd->span);
5720 * If we dont want to free the old_rt yet then
5721 * set old_rd to NULL to skip the freeing later
5724 if (!atomic_dec_and_test(&old_rd->refcount))
5728 atomic_inc(&rd->refcount);
5731 cpumask_set_cpu(rq->cpu, rd->span);
5732 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5735 raw_spin_unlock_irqrestore(&rq->lock, flags);
5738 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5741 static int init_rootdomain(struct root_domain *rd)
5743 memset(rd, 0, sizeof(*rd));
5745 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5747 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5749 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5752 if (cpupri_init(&rd->cpupri) != 0)
5757 free_cpumask_var(rd->rto_mask);
5759 free_cpumask_var(rd->online);
5761 free_cpumask_var(rd->span);
5767 * By default the system creates a single root-domain with all cpus as
5768 * members (mimicking the global state we have today).
5770 struct root_domain def_root_domain;
5772 static void init_defrootdomain(void)
5774 init_rootdomain(&def_root_domain);
5776 atomic_set(&def_root_domain.refcount, 1);
5779 static struct root_domain *alloc_rootdomain(void)
5781 struct root_domain *rd;
5783 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5787 if (init_rootdomain(rd) != 0) {
5795 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5797 struct sched_group *tmp, *first;
5806 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5811 } while (sg != first);
5814 static void free_sched_domain(struct rcu_head *rcu)
5816 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5819 * If its an overlapping domain it has private groups, iterate and
5822 if (sd->flags & SD_OVERLAP) {
5823 free_sched_groups(sd->groups, 1);
5824 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5825 kfree(sd->groups->sgp);
5831 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5833 call_rcu(&sd->rcu, free_sched_domain);
5836 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5838 for (; sd; sd = sd->parent)
5839 destroy_sched_domain(sd, cpu);
5843 * Keep a special pointer to the highest sched_domain that has
5844 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5845 * allows us to avoid some pointer chasing select_idle_sibling().
5847 * Also keep a unique ID per domain (we use the first cpu number in
5848 * the cpumask of the domain), this allows us to quickly tell if
5849 * two cpus are in the same cache domain, see cpus_share_cache().
5851 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5852 DEFINE_PER_CPU(int, sd_llc_id);
5854 static void update_top_cache_domain(int cpu)
5856 struct sched_domain *sd;
5859 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5861 id = cpumask_first(sched_domain_span(sd));
5863 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5864 per_cpu(sd_llc_id, cpu) = id;
5868 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5869 * hold the hotplug lock.
5872 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5874 struct rq *rq = cpu_rq(cpu);
5875 struct sched_domain *tmp;
5877 /* Remove the sched domains which do not contribute to scheduling. */
5878 for (tmp = sd; tmp; ) {
5879 struct sched_domain *parent = tmp->parent;
5883 if (sd_parent_degenerate(tmp, parent)) {
5884 tmp->parent = parent->parent;
5886 parent->parent->child = tmp;
5887 destroy_sched_domain(parent, cpu);
5892 if (sd && sd_degenerate(sd)) {
5895 destroy_sched_domain(tmp, cpu);
5900 sched_domain_debug(sd, cpu);
5902 rq_attach_root(rq, rd);
5904 rcu_assign_pointer(rq->sd, sd);
5905 destroy_sched_domains(tmp, cpu);
5907 update_top_cache_domain(cpu);
5910 /* cpus with isolated domains */
5911 static cpumask_var_t cpu_isolated_map;
5913 /* Setup the mask of cpus configured for isolated domains */
5914 static int __init isolated_cpu_setup(char *str)
5916 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5917 cpulist_parse(str, cpu_isolated_map);
5921 __setup("isolcpus=", isolated_cpu_setup);
5923 static const struct cpumask *cpu_cpu_mask(int cpu)
5925 return cpumask_of_node(cpu_to_node(cpu));
5929 struct sched_domain **__percpu sd;
5930 struct sched_group **__percpu sg;
5931 struct sched_group_power **__percpu sgp;
5935 struct sched_domain ** __percpu sd;
5936 struct root_domain *rd;
5946 struct sched_domain_topology_level;
5948 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5949 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5951 #define SDTL_OVERLAP 0x01
5953 struct sched_domain_topology_level {
5954 sched_domain_init_f init;
5955 sched_domain_mask_f mask;
5958 struct sd_data data;
5962 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5964 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5965 const struct cpumask *span = sched_domain_span(sd);
5966 struct cpumask *covered = sched_domains_tmpmask;
5967 struct sd_data *sdd = sd->private;
5968 struct sched_domain *child;
5971 cpumask_clear(covered);
5973 for_each_cpu(i, span) {
5974 struct cpumask *sg_span;
5976 if (cpumask_test_cpu(i, covered))
5979 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5980 GFP_KERNEL, cpu_to_node(cpu));
5985 sg_span = sched_group_cpus(sg);
5987 child = *per_cpu_ptr(sdd->sd, i);
5989 child = child->child;
5990 cpumask_copy(sg_span, sched_domain_span(child));
5992 cpumask_set_cpu(i, sg_span);
5994 cpumask_or(covered, covered, sg_span);
5996 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5997 atomic_inc(&sg->sgp->ref);
5999 if (cpumask_test_cpu(cpu, sg_span))
6009 sd->groups = groups;
6014 free_sched_groups(first, 0);
6019 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6021 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6022 struct sched_domain *child = sd->child;
6025 cpu = cpumask_first(sched_domain_span(child));
6028 *sg = *per_cpu_ptr(sdd->sg, cpu);
6029 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6030 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6037 * build_sched_groups will build a circular linked list of the groups
6038 * covered by the given span, and will set each group's ->cpumask correctly,
6039 * and ->cpu_power to 0.
6041 * Assumes the sched_domain tree is fully constructed
6044 build_sched_groups(struct sched_domain *sd, int cpu)
6046 struct sched_group *first = NULL, *last = NULL;
6047 struct sd_data *sdd = sd->private;
6048 const struct cpumask *span = sched_domain_span(sd);
6049 struct cpumask *covered;
6052 get_group(cpu, sdd, &sd->groups);
6053 atomic_inc(&sd->groups->ref);
6055 if (cpu != cpumask_first(sched_domain_span(sd)))
6058 lockdep_assert_held(&sched_domains_mutex);
6059 covered = sched_domains_tmpmask;
6061 cpumask_clear(covered);
6063 for_each_cpu(i, span) {
6064 struct sched_group *sg;
6065 int group = get_group(i, sdd, &sg);
6068 if (cpumask_test_cpu(i, covered))
6071 cpumask_clear(sched_group_cpus(sg));
6074 for_each_cpu(j, span) {
6075 if (get_group(j, sdd, NULL) != group)
6078 cpumask_set_cpu(j, covered);
6079 cpumask_set_cpu(j, sched_group_cpus(sg));
6094 * Initialize sched groups cpu_power.
6096 * cpu_power indicates the capacity of sched group, which is used while
6097 * distributing the load between different sched groups in a sched domain.
6098 * Typically cpu_power for all the groups in a sched domain will be same unless
6099 * there are asymmetries in the topology. If there are asymmetries, group
6100 * having more cpu_power will pickup more load compared to the group having
6103 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6105 struct sched_group *sg = sd->groups;
6107 WARN_ON(!sd || !sg);
6110 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6112 } while (sg != sd->groups);
6114 if (cpu != group_first_cpu(sg))
6117 update_group_power(sd, cpu);
6118 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6121 int __weak arch_sd_sibling_asym_packing(void)
6123 return 0*SD_ASYM_PACKING;
6127 * Initializers for schedule domains
6128 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6131 #ifdef CONFIG_SCHED_DEBUG
6132 # define SD_INIT_NAME(sd, type) sd->name = #type
6134 # define SD_INIT_NAME(sd, type) do { } while (0)
6137 #define SD_INIT_FUNC(type) \
6138 static noinline struct sched_domain * \
6139 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6141 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6142 *sd = SD_##type##_INIT; \
6143 SD_INIT_NAME(sd, type); \
6144 sd->private = &tl->data; \
6149 #ifdef CONFIG_SCHED_SMT
6150 SD_INIT_FUNC(SIBLING)
6152 #ifdef CONFIG_SCHED_MC
6155 #ifdef CONFIG_SCHED_BOOK
6159 static int default_relax_domain_level = -1;
6160 int sched_domain_level_max;
6162 static int __init setup_relax_domain_level(char *str)
6166 val = simple_strtoul(str, NULL, 0);
6167 if (val < sched_domain_level_max)
6168 default_relax_domain_level = val;
6172 __setup("relax_domain_level=", setup_relax_domain_level);
6174 static void set_domain_attribute(struct sched_domain *sd,
6175 struct sched_domain_attr *attr)
6179 if (!attr || attr->relax_domain_level < 0) {
6180 if (default_relax_domain_level < 0)
6183 request = default_relax_domain_level;
6185 request = attr->relax_domain_level;
6186 if (request < sd->level) {
6187 /* turn off idle balance on this domain */
6188 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6190 /* turn on idle balance on this domain */
6191 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6195 static void __sdt_free(const struct cpumask *cpu_map);
6196 static int __sdt_alloc(const struct cpumask *cpu_map);
6198 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6199 const struct cpumask *cpu_map)
6203 if (!atomic_read(&d->rd->refcount))
6204 free_rootdomain(&d->rd->rcu); /* fall through */
6206 free_percpu(d->sd); /* fall through */
6208 __sdt_free(cpu_map); /* fall through */
6214 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6215 const struct cpumask *cpu_map)
6217 memset(d, 0, sizeof(*d));
6219 if (__sdt_alloc(cpu_map))
6220 return sa_sd_storage;
6221 d->sd = alloc_percpu(struct sched_domain *);
6223 return sa_sd_storage;
6224 d->rd = alloc_rootdomain();
6227 return sa_rootdomain;
6231 * NULL the sd_data elements we've used to build the sched_domain and
6232 * sched_group structure so that the subsequent __free_domain_allocs()
6233 * will not free the data we're using.
6235 static void claim_allocations(int cpu, struct sched_domain *sd)
6237 struct sd_data *sdd = sd->private;
6239 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6240 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6242 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6243 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6245 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6246 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6249 #ifdef CONFIG_SCHED_SMT
6250 static const struct cpumask *cpu_smt_mask(int cpu)
6252 return topology_thread_cpumask(cpu);
6257 * Topology list, bottom-up.
6259 static struct sched_domain_topology_level default_topology[] = {
6260 #ifdef CONFIG_SCHED_SMT
6261 { sd_init_SIBLING, cpu_smt_mask, },
6263 #ifdef CONFIG_SCHED_MC
6264 { sd_init_MC, cpu_coregroup_mask, },
6266 #ifdef CONFIG_SCHED_BOOK
6267 { sd_init_BOOK, cpu_book_mask, },
6269 { sd_init_CPU, cpu_cpu_mask, },
6273 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6277 static int sched_domains_numa_levels;
6278 static int sched_domains_numa_scale;
6279 static int *sched_domains_numa_distance;
6280 static struct cpumask ***sched_domains_numa_masks;
6281 static int sched_domains_curr_level;
6283 static inline int sd_local_flags(int level)
6285 if (sched_domains_numa_distance[level] > REMOTE_DISTANCE)
6288 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6291 static struct sched_domain *
6292 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6294 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6295 int level = tl->numa_level;
6296 int sd_weight = cpumask_weight(
6297 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6299 *sd = (struct sched_domain){
6300 .min_interval = sd_weight,
6301 .max_interval = 2*sd_weight,
6303 .imbalance_pct = 125,
6304 .cache_nice_tries = 2,
6311 .flags = 1*SD_LOAD_BALANCE
6312 | 1*SD_BALANCE_NEWIDLE
6318 | 0*SD_SHARE_CPUPOWER
6319 | 0*SD_SHARE_PKG_RESOURCES
6321 | 0*SD_PREFER_SIBLING
6322 | sd_local_flags(level)
6324 .last_balance = jiffies,
6325 .balance_interval = sd_weight,
6327 SD_INIT_NAME(sd, NUMA);
6328 sd->private = &tl->data;
6331 * Ugly hack to pass state to sd_numa_mask()...
6333 sched_domains_curr_level = tl->numa_level;
6338 static const struct cpumask *sd_numa_mask(int cpu)
6340 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6343 static void sched_init_numa(void)
6345 int next_distance, curr_distance = node_distance(0, 0);
6346 struct sched_domain_topology_level *tl;
6350 sched_domains_numa_scale = curr_distance;
6351 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6352 if (!sched_domains_numa_distance)
6356 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6357 * unique distances in the node_distance() table.
6359 * Assumes node_distance(0,j) includes all distances in
6360 * node_distance(i,j) in order to avoid cubic time.
6362 * XXX: could be optimized to O(n log n) by using sort()
6364 next_distance = curr_distance;
6365 for (i = 0; i < nr_node_ids; i++) {
6366 for (j = 0; j < nr_node_ids; j++) {
6367 int distance = node_distance(0, j);
6368 if (distance > curr_distance &&
6369 (distance < next_distance ||
6370 next_distance == curr_distance))
6371 next_distance = distance;
6373 if (next_distance != curr_distance) {
6374 sched_domains_numa_distance[level++] = next_distance;
6375 sched_domains_numa_levels = level;
6376 curr_distance = next_distance;
6380 * 'level' contains the number of unique distances, excluding the
6381 * identity distance node_distance(i,i).
6383 * The sched_domains_nume_distance[] array includes the actual distance
6387 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6388 if (!sched_domains_numa_masks)
6392 * Now for each level, construct a mask per node which contains all
6393 * cpus of nodes that are that many hops away from us.
6395 for (i = 0; i < level; i++) {
6396 sched_domains_numa_masks[i] =
6397 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6398 if (!sched_domains_numa_masks[i])
6401 for (j = 0; j < nr_node_ids; j++) {
6402 struct cpumask *mask = kzalloc_node(cpumask_size(), GFP_KERNEL, j);
6406 sched_domains_numa_masks[i][j] = mask;
6408 for (k = 0; k < nr_node_ids; k++) {
6409 if (node_distance(j, k) > sched_domains_numa_distance[i])
6412 cpumask_or(mask, mask, cpumask_of_node(k));
6417 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6418 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6423 * Copy the default topology bits..
6425 for (i = 0; default_topology[i].init; i++)
6426 tl[i] = default_topology[i];
6429 * .. and append 'j' levels of NUMA goodness.
6431 for (j = 0; j < level; i++, j++) {
6432 tl[i] = (struct sched_domain_topology_level){
6433 .init = sd_numa_init,
6434 .mask = sd_numa_mask,
6435 .flags = SDTL_OVERLAP,
6440 sched_domain_topology = tl;
6443 static inline void sched_init_numa(void)
6446 #endif /* CONFIG_NUMA */
6448 static int __sdt_alloc(const struct cpumask *cpu_map)
6450 struct sched_domain_topology_level *tl;
6453 for (tl = sched_domain_topology; tl->init; tl++) {
6454 struct sd_data *sdd = &tl->data;
6456 sdd->sd = alloc_percpu(struct sched_domain *);
6460 sdd->sg = alloc_percpu(struct sched_group *);
6464 sdd->sgp = alloc_percpu(struct sched_group_power *);
6468 for_each_cpu(j, cpu_map) {
6469 struct sched_domain *sd;
6470 struct sched_group *sg;
6471 struct sched_group_power *sgp;
6473 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6474 GFP_KERNEL, cpu_to_node(j));
6478 *per_cpu_ptr(sdd->sd, j) = sd;
6480 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6481 GFP_KERNEL, cpu_to_node(j));
6487 *per_cpu_ptr(sdd->sg, j) = sg;
6489 sgp = kzalloc_node(sizeof(struct sched_group_power),
6490 GFP_KERNEL, cpu_to_node(j));
6494 *per_cpu_ptr(sdd->sgp, j) = sgp;
6501 static void __sdt_free(const struct cpumask *cpu_map)
6503 struct sched_domain_topology_level *tl;
6506 for (tl = sched_domain_topology; tl->init; tl++) {
6507 struct sd_data *sdd = &tl->data;
6509 for_each_cpu(j, cpu_map) {
6510 struct sched_domain *sd;
6513 sd = *per_cpu_ptr(sdd->sd, j);
6514 if (sd && (sd->flags & SD_OVERLAP))
6515 free_sched_groups(sd->groups, 0);
6516 kfree(*per_cpu_ptr(sdd->sd, j));
6520 kfree(*per_cpu_ptr(sdd->sg, j));
6522 kfree(*per_cpu_ptr(sdd->sgp, j));
6524 free_percpu(sdd->sd);
6526 free_percpu(sdd->sg);
6528 free_percpu(sdd->sgp);
6533 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6534 struct s_data *d, const struct cpumask *cpu_map,
6535 struct sched_domain_attr *attr, struct sched_domain *child,
6538 struct sched_domain *sd = tl->init(tl, cpu);
6542 set_domain_attribute(sd, attr);
6543 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6545 sd->level = child->level + 1;
6546 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6555 * Build sched domains for a given set of cpus and attach the sched domains
6556 * to the individual cpus
6558 static int build_sched_domains(const struct cpumask *cpu_map,
6559 struct sched_domain_attr *attr)
6561 enum s_alloc alloc_state = sa_none;
6562 struct sched_domain *sd;
6564 int i, ret = -ENOMEM;
6566 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6567 if (alloc_state != sa_rootdomain)
6570 /* Set up domains for cpus specified by the cpu_map. */
6571 for_each_cpu(i, cpu_map) {
6572 struct sched_domain_topology_level *tl;
6575 for (tl = sched_domain_topology; tl->init; tl++) {
6576 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6577 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6578 sd->flags |= SD_OVERLAP;
6579 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6586 *per_cpu_ptr(d.sd, i) = sd;
6589 /* Build the groups for the domains */
6590 for_each_cpu(i, cpu_map) {
6591 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6592 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6593 if (sd->flags & SD_OVERLAP) {
6594 if (build_overlap_sched_groups(sd, i))
6597 if (build_sched_groups(sd, i))
6603 /* Calculate CPU power for physical packages and nodes */
6604 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6605 if (!cpumask_test_cpu(i, cpu_map))
6608 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6609 claim_allocations(i, sd);
6610 init_sched_groups_power(i, sd);
6614 /* Attach the domains */
6616 for_each_cpu(i, cpu_map) {
6617 sd = *per_cpu_ptr(d.sd, i);
6618 cpu_attach_domain(sd, d.rd, i);
6624 __free_domain_allocs(&d, alloc_state, cpu_map);
6628 static cpumask_var_t *doms_cur; /* current sched domains */
6629 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6630 static struct sched_domain_attr *dattr_cur;
6631 /* attribues of custom domains in 'doms_cur' */
6634 * Special case: If a kmalloc of a doms_cur partition (array of
6635 * cpumask) fails, then fallback to a single sched domain,
6636 * as determined by the single cpumask fallback_doms.
6638 static cpumask_var_t fallback_doms;
6641 * arch_update_cpu_topology lets virtualized architectures update the
6642 * cpu core maps. It is supposed to return 1 if the topology changed
6643 * or 0 if it stayed the same.
6645 int __attribute__((weak)) arch_update_cpu_topology(void)
6650 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6653 cpumask_var_t *doms;
6655 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6658 for (i = 0; i < ndoms; i++) {
6659 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6660 free_sched_domains(doms, i);
6667 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6670 for (i = 0; i < ndoms; i++)
6671 free_cpumask_var(doms[i]);
6676 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6677 * For now this just excludes isolated cpus, but could be used to
6678 * exclude other special cases in the future.
6680 static int init_sched_domains(const struct cpumask *cpu_map)
6684 arch_update_cpu_topology();
6686 doms_cur = alloc_sched_domains(ndoms_cur);
6688 doms_cur = &fallback_doms;
6689 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6691 err = build_sched_domains(doms_cur[0], NULL);
6692 register_sched_domain_sysctl();
6698 * Detach sched domains from a group of cpus specified in cpu_map
6699 * These cpus will now be attached to the NULL domain
6701 static void detach_destroy_domains(const struct cpumask *cpu_map)
6706 for_each_cpu(i, cpu_map)
6707 cpu_attach_domain(NULL, &def_root_domain, i);
6711 /* handle null as "default" */
6712 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6713 struct sched_domain_attr *new, int idx_new)
6715 struct sched_domain_attr tmp;
6722 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6723 new ? (new + idx_new) : &tmp,
6724 sizeof(struct sched_domain_attr));
6728 * Partition sched domains as specified by the 'ndoms_new'
6729 * cpumasks in the array doms_new[] of cpumasks. This compares
6730 * doms_new[] to the current sched domain partitioning, doms_cur[].
6731 * It destroys each deleted domain and builds each new domain.
6733 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6734 * The masks don't intersect (don't overlap.) We should setup one
6735 * sched domain for each mask. CPUs not in any of the cpumasks will
6736 * not be load balanced. If the same cpumask appears both in the
6737 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6740 * The passed in 'doms_new' should be allocated using
6741 * alloc_sched_domains. This routine takes ownership of it and will
6742 * free_sched_domains it when done with it. If the caller failed the
6743 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6744 * and partition_sched_domains() will fallback to the single partition
6745 * 'fallback_doms', it also forces the domains to be rebuilt.
6747 * If doms_new == NULL it will be replaced with cpu_online_mask.
6748 * ndoms_new == 0 is a special case for destroying existing domains,
6749 * and it will not create the default domain.
6751 * Call with hotplug lock held
6753 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6754 struct sched_domain_attr *dattr_new)
6759 mutex_lock(&sched_domains_mutex);
6761 /* always unregister in case we don't destroy any domains */
6762 unregister_sched_domain_sysctl();
6764 /* Let architecture update cpu core mappings. */
6765 new_topology = arch_update_cpu_topology();
6767 n = doms_new ? ndoms_new : 0;
6769 /* Destroy deleted domains */
6770 for (i = 0; i < ndoms_cur; i++) {
6771 for (j = 0; j < n && !new_topology; j++) {
6772 if (cpumask_equal(doms_cur[i], doms_new[j])
6773 && dattrs_equal(dattr_cur, i, dattr_new, j))
6776 /* no match - a current sched domain not in new doms_new[] */
6777 detach_destroy_domains(doms_cur[i]);
6782 if (doms_new == NULL) {
6784 doms_new = &fallback_doms;
6785 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6786 WARN_ON_ONCE(dattr_new);
6789 /* Build new domains */
6790 for (i = 0; i < ndoms_new; i++) {
6791 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6792 if (cpumask_equal(doms_new[i], doms_cur[j])
6793 && dattrs_equal(dattr_new, i, dattr_cur, j))
6796 /* no match - add a new doms_new */
6797 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6802 /* Remember the new sched domains */
6803 if (doms_cur != &fallback_doms)
6804 free_sched_domains(doms_cur, ndoms_cur);
6805 kfree(dattr_cur); /* kfree(NULL) is safe */
6806 doms_cur = doms_new;
6807 dattr_cur = dattr_new;
6808 ndoms_cur = ndoms_new;
6810 register_sched_domain_sysctl();
6812 mutex_unlock(&sched_domains_mutex);
6816 * Update cpusets according to cpu_active mask. If cpusets are
6817 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6818 * around partition_sched_domains().
6820 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6823 switch (action & ~CPU_TASKS_FROZEN) {
6825 case CPU_DOWN_FAILED:
6826 cpuset_update_active_cpus();
6833 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6836 switch (action & ~CPU_TASKS_FROZEN) {
6837 case CPU_DOWN_PREPARE:
6838 cpuset_update_active_cpus();
6845 void __init sched_init_smp(void)
6847 cpumask_var_t non_isolated_cpus;
6849 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6850 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6855 mutex_lock(&sched_domains_mutex);
6856 init_sched_domains(cpu_active_mask);
6857 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6858 if (cpumask_empty(non_isolated_cpus))
6859 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6860 mutex_unlock(&sched_domains_mutex);
6863 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6864 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6866 /* RT runtime code needs to handle some hotplug events */
6867 hotcpu_notifier(update_runtime, 0);
6871 /* Move init over to a non-isolated CPU */
6872 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6874 sched_init_granularity();
6875 free_cpumask_var(non_isolated_cpus);
6877 init_sched_rt_class();
6880 void __init sched_init_smp(void)
6882 sched_init_granularity();
6884 #endif /* CONFIG_SMP */
6886 const_debug unsigned int sysctl_timer_migration = 1;
6888 int in_sched_functions(unsigned long addr)
6890 return in_lock_functions(addr) ||
6891 (addr >= (unsigned long)__sched_text_start
6892 && addr < (unsigned long)__sched_text_end);
6895 #ifdef CONFIG_CGROUP_SCHED
6896 struct task_group root_task_group;
6899 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6901 void __init sched_init(void)
6904 unsigned long alloc_size = 0, ptr;
6906 #ifdef CONFIG_FAIR_GROUP_SCHED
6907 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6909 #ifdef CONFIG_RT_GROUP_SCHED
6910 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6912 #ifdef CONFIG_CPUMASK_OFFSTACK
6913 alloc_size += num_possible_cpus() * cpumask_size();
6916 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6918 #ifdef CONFIG_FAIR_GROUP_SCHED
6919 root_task_group.se = (struct sched_entity **)ptr;
6920 ptr += nr_cpu_ids * sizeof(void **);
6922 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6923 ptr += nr_cpu_ids * sizeof(void **);
6925 #endif /* CONFIG_FAIR_GROUP_SCHED */
6926 #ifdef CONFIG_RT_GROUP_SCHED
6927 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6928 ptr += nr_cpu_ids * sizeof(void **);
6930 root_task_group.rt_rq = (struct rt_rq **)ptr;
6931 ptr += nr_cpu_ids * sizeof(void **);
6933 #endif /* CONFIG_RT_GROUP_SCHED */
6934 #ifdef CONFIG_CPUMASK_OFFSTACK
6935 for_each_possible_cpu(i) {
6936 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6937 ptr += cpumask_size();
6939 #endif /* CONFIG_CPUMASK_OFFSTACK */
6943 init_defrootdomain();
6946 init_rt_bandwidth(&def_rt_bandwidth,
6947 global_rt_period(), global_rt_runtime());
6949 #ifdef CONFIG_RT_GROUP_SCHED
6950 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6951 global_rt_period(), global_rt_runtime());
6952 #endif /* CONFIG_RT_GROUP_SCHED */
6954 #ifdef CONFIG_CGROUP_SCHED
6955 list_add(&root_task_group.list, &task_groups);
6956 INIT_LIST_HEAD(&root_task_group.children);
6957 INIT_LIST_HEAD(&root_task_group.siblings);
6958 autogroup_init(&init_task);
6960 #endif /* CONFIG_CGROUP_SCHED */
6962 #ifdef CONFIG_CGROUP_CPUACCT
6963 root_cpuacct.cpustat = &kernel_cpustat;
6964 root_cpuacct.cpuusage = alloc_percpu(u64);
6965 /* Too early, not expected to fail */
6966 BUG_ON(!root_cpuacct.cpuusage);
6968 for_each_possible_cpu(i) {
6972 raw_spin_lock_init(&rq->lock);
6974 rq->calc_load_active = 0;
6975 rq->calc_load_update = jiffies + LOAD_FREQ;
6976 init_cfs_rq(&rq->cfs);
6977 init_rt_rq(&rq->rt, rq);
6978 #ifdef CONFIG_FAIR_GROUP_SCHED
6979 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6980 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6982 * How much cpu bandwidth does root_task_group get?
6984 * In case of task-groups formed thr' the cgroup filesystem, it
6985 * gets 100% of the cpu resources in the system. This overall
6986 * system cpu resource is divided among the tasks of
6987 * root_task_group and its child task-groups in a fair manner,
6988 * based on each entity's (task or task-group's) weight
6989 * (se->load.weight).
6991 * In other words, if root_task_group has 10 tasks of weight
6992 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6993 * then A0's share of the cpu resource is:
6995 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6997 * We achieve this by letting root_task_group's tasks sit
6998 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7000 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7001 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7002 #endif /* CONFIG_FAIR_GROUP_SCHED */
7004 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7005 #ifdef CONFIG_RT_GROUP_SCHED
7006 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7007 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7010 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7011 rq->cpu_load[j] = 0;
7013 rq->last_load_update_tick = jiffies;
7018 rq->cpu_power = SCHED_POWER_SCALE;
7019 rq->post_schedule = 0;
7020 rq->active_balance = 0;
7021 rq->next_balance = jiffies;
7026 rq->avg_idle = 2*sysctl_sched_migration_cost;
7028 INIT_LIST_HEAD(&rq->cfs_tasks);
7030 rq_attach_root(rq, &def_root_domain);
7036 atomic_set(&rq->nr_iowait, 0);
7039 set_load_weight(&init_task);
7041 #ifdef CONFIG_PREEMPT_NOTIFIERS
7042 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7045 #ifdef CONFIG_RT_MUTEXES
7046 plist_head_init(&init_task.pi_waiters);
7050 * The boot idle thread does lazy MMU switching as well:
7052 atomic_inc(&init_mm.mm_count);
7053 enter_lazy_tlb(&init_mm, current);
7056 * Make us the idle thread. Technically, schedule() should not be
7057 * called from this thread, however somewhere below it might be,
7058 * but because we are the idle thread, we just pick up running again
7059 * when this runqueue becomes "idle".
7061 init_idle(current, smp_processor_id());
7063 calc_load_update = jiffies + LOAD_FREQ;
7066 * During early bootup we pretend to be a normal task:
7068 current->sched_class = &fair_sched_class;
7071 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7072 /* May be allocated at isolcpus cmdline parse time */
7073 if (cpu_isolated_map == NULL)
7074 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7075 idle_thread_set_boot_cpu();
7077 init_sched_fair_class();
7079 scheduler_running = 1;
7082 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7083 static inline int preempt_count_equals(int preempt_offset)
7085 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7087 return (nested == preempt_offset);
7090 void __might_sleep(const char *file, int line, int preempt_offset)
7092 static unsigned long prev_jiffy; /* ratelimiting */
7094 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7095 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7096 system_state != SYSTEM_RUNNING || oops_in_progress)
7098 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7100 prev_jiffy = jiffies;
7103 "BUG: sleeping function called from invalid context at %s:%d\n",
7106 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7107 in_atomic(), irqs_disabled(),
7108 current->pid, current->comm);
7110 debug_show_held_locks(current);
7111 if (irqs_disabled())
7112 print_irqtrace_events(current);
7115 EXPORT_SYMBOL(__might_sleep);
7118 #ifdef CONFIG_MAGIC_SYSRQ
7119 static void normalize_task(struct rq *rq, struct task_struct *p)
7121 const struct sched_class *prev_class = p->sched_class;
7122 int old_prio = p->prio;
7127 dequeue_task(rq, p, 0);
7128 __setscheduler(rq, p, SCHED_NORMAL, 0);
7130 enqueue_task(rq, p, 0);
7131 resched_task(rq->curr);
7134 check_class_changed(rq, p, prev_class, old_prio);
7137 void normalize_rt_tasks(void)
7139 struct task_struct *g, *p;
7140 unsigned long flags;
7143 read_lock_irqsave(&tasklist_lock, flags);
7144 do_each_thread(g, p) {
7146 * Only normalize user tasks:
7151 p->se.exec_start = 0;
7152 #ifdef CONFIG_SCHEDSTATS
7153 p->se.statistics.wait_start = 0;
7154 p->se.statistics.sleep_start = 0;
7155 p->se.statistics.block_start = 0;
7160 * Renice negative nice level userspace
7163 if (TASK_NICE(p) < 0 && p->mm)
7164 set_user_nice(p, 0);
7168 raw_spin_lock(&p->pi_lock);
7169 rq = __task_rq_lock(p);
7171 normalize_task(rq, p);
7173 __task_rq_unlock(rq);
7174 raw_spin_unlock(&p->pi_lock);
7175 } while_each_thread(g, p);
7177 read_unlock_irqrestore(&tasklist_lock, flags);
7180 #endif /* CONFIG_MAGIC_SYSRQ */
7182 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7184 * These functions are only useful for the IA64 MCA handling, or kdb.
7186 * They can only be called when the whole system has been
7187 * stopped - every CPU needs to be quiescent, and no scheduling
7188 * activity can take place. Using them for anything else would
7189 * be a serious bug, and as a result, they aren't even visible
7190 * under any other configuration.
7194 * curr_task - return the current task for a given cpu.
7195 * @cpu: the processor in question.
7197 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7199 struct task_struct *curr_task(int cpu)
7201 return cpu_curr(cpu);
7204 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7208 * set_curr_task - set the current task for a given cpu.
7209 * @cpu: the processor in question.
7210 * @p: the task pointer to set.
7212 * Description: This function must only be used when non-maskable interrupts
7213 * are serviced on a separate stack. It allows the architecture to switch the
7214 * notion of the current task on a cpu in a non-blocking manner. This function
7215 * must be called with all CPU's synchronized, and interrupts disabled, the
7216 * and caller must save the original value of the current task (see
7217 * curr_task() above) and restore that value before reenabling interrupts and
7218 * re-starting the system.
7220 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7222 void set_curr_task(int cpu, struct task_struct *p)
7229 #ifdef CONFIG_CGROUP_SCHED
7230 /* task_group_lock serializes the addition/removal of task groups */
7231 static DEFINE_SPINLOCK(task_group_lock);
7233 static void free_sched_group(struct task_group *tg)
7235 free_fair_sched_group(tg);
7236 free_rt_sched_group(tg);
7241 /* allocate runqueue etc for a new task group */
7242 struct task_group *sched_create_group(struct task_group *parent)
7244 struct task_group *tg;
7245 unsigned long flags;
7247 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7249 return ERR_PTR(-ENOMEM);
7251 if (!alloc_fair_sched_group(tg, parent))
7254 if (!alloc_rt_sched_group(tg, parent))
7257 spin_lock_irqsave(&task_group_lock, flags);
7258 list_add_rcu(&tg->list, &task_groups);
7260 WARN_ON(!parent); /* root should already exist */
7262 tg->parent = parent;
7263 INIT_LIST_HEAD(&tg->children);
7264 list_add_rcu(&tg->siblings, &parent->children);
7265 spin_unlock_irqrestore(&task_group_lock, flags);
7270 free_sched_group(tg);
7271 return ERR_PTR(-ENOMEM);
7274 /* rcu callback to free various structures associated with a task group */
7275 static void free_sched_group_rcu(struct rcu_head *rhp)
7277 /* now it should be safe to free those cfs_rqs */
7278 free_sched_group(container_of(rhp, struct task_group, rcu));
7281 /* Destroy runqueue etc associated with a task group */
7282 void sched_destroy_group(struct task_group *tg)
7284 unsigned long flags;
7287 /* end participation in shares distribution */
7288 for_each_possible_cpu(i)
7289 unregister_fair_sched_group(tg, i);
7291 spin_lock_irqsave(&task_group_lock, flags);
7292 list_del_rcu(&tg->list);
7293 list_del_rcu(&tg->siblings);
7294 spin_unlock_irqrestore(&task_group_lock, flags);
7296 /* wait for possible concurrent references to cfs_rqs complete */
7297 call_rcu(&tg->rcu, free_sched_group_rcu);
7300 /* change task's runqueue when it moves between groups.
7301 * The caller of this function should have put the task in its new group
7302 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7303 * reflect its new group.
7305 void sched_move_task(struct task_struct *tsk)
7308 unsigned long flags;
7311 rq = task_rq_lock(tsk, &flags);
7313 running = task_current(rq, tsk);
7317 dequeue_task(rq, tsk, 0);
7318 if (unlikely(running))
7319 tsk->sched_class->put_prev_task(rq, tsk);
7321 #ifdef CONFIG_FAIR_GROUP_SCHED
7322 if (tsk->sched_class->task_move_group)
7323 tsk->sched_class->task_move_group(tsk, on_rq);
7326 set_task_rq(tsk, task_cpu(tsk));
7328 if (unlikely(running))
7329 tsk->sched_class->set_curr_task(rq);
7331 enqueue_task(rq, tsk, 0);
7333 task_rq_unlock(rq, tsk, &flags);
7335 #endif /* CONFIG_CGROUP_SCHED */
7337 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7338 static unsigned long to_ratio(u64 period, u64 runtime)
7340 if (runtime == RUNTIME_INF)
7343 return div64_u64(runtime << 20, period);
7347 #ifdef CONFIG_RT_GROUP_SCHED
7349 * Ensure that the real time constraints are schedulable.
7351 static DEFINE_MUTEX(rt_constraints_mutex);
7353 /* Must be called with tasklist_lock held */
7354 static inline int tg_has_rt_tasks(struct task_group *tg)
7356 struct task_struct *g, *p;
7358 do_each_thread(g, p) {
7359 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7361 } while_each_thread(g, p);
7366 struct rt_schedulable_data {
7367 struct task_group *tg;
7372 static int tg_rt_schedulable(struct task_group *tg, void *data)
7374 struct rt_schedulable_data *d = data;
7375 struct task_group *child;
7376 unsigned long total, sum = 0;
7377 u64 period, runtime;
7379 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7380 runtime = tg->rt_bandwidth.rt_runtime;
7383 period = d->rt_period;
7384 runtime = d->rt_runtime;
7388 * Cannot have more runtime than the period.
7390 if (runtime > period && runtime != RUNTIME_INF)
7394 * Ensure we don't starve existing RT tasks.
7396 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7399 total = to_ratio(period, runtime);
7402 * Nobody can have more than the global setting allows.
7404 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7408 * The sum of our children's runtime should not exceed our own.
7410 list_for_each_entry_rcu(child, &tg->children, siblings) {
7411 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7412 runtime = child->rt_bandwidth.rt_runtime;
7414 if (child == d->tg) {
7415 period = d->rt_period;
7416 runtime = d->rt_runtime;
7419 sum += to_ratio(period, runtime);
7428 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7432 struct rt_schedulable_data data = {
7434 .rt_period = period,
7435 .rt_runtime = runtime,
7439 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7445 static int tg_set_rt_bandwidth(struct task_group *tg,
7446 u64 rt_period, u64 rt_runtime)
7450 mutex_lock(&rt_constraints_mutex);
7451 read_lock(&tasklist_lock);
7452 err = __rt_schedulable(tg, rt_period, rt_runtime);
7456 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7457 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7458 tg->rt_bandwidth.rt_runtime = rt_runtime;
7460 for_each_possible_cpu(i) {
7461 struct rt_rq *rt_rq = tg->rt_rq[i];
7463 raw_spin_lock(&rt_rq->rt_runtime_lock);
7464 rt_rq->rt_runtime = rt_runtime;
7465 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7467 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7469 read_unlock(&tasklist_lock);
7470 mutex_unlock(&rt_constraints_mutex);
7475 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7477 u64 rt_runtime, rt_period;
7479 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7480 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7481 if (rt_runtime_us < 0)
7482 rt_runtime = RUNTIME_INF;
7484 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7487 long sched_group_rt_runtime(struct task_group *tg)
7491 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7494 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7495 do_div(rt_runtime_us, NSEC_PER_USEC);
7496 return rt_runtime_us;
7499 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7501 u64 rt_runtime, rt_period;
7503 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7504 rt_runtime = tg->rt_bandwidth.rt_runtime;
7509 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7512 long sched_group_rt_period(struct task_group *tg)
7516 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7517 do_div(rt_period_us, NSEC_PER_USEC);
7518 return rt_period_us;
7521 static int sched_rt_global_constraints(void)
7523 u64 runtime, period;
7526 if (sysctl_sched_rt_period <= 0)
7529 runtime = global_rt_runtime();
7530 period = global_rt_period();
7533 * Sanity check on the sysctl variables.
7535 if (runtime > period && runtime != RUNTIME_INF)
7538 mutex_lock(&rt_constraints_mutex);
7539 read_lock(&tasklist_lock);
7540 ret = __rt_schedulable(NULL, 0, 0);
7541 read_unlock(&tasklist_lock);
7542 mutex_unlock(&rt_constraints_mutex);
7547 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7549 /* Don't accept realtime tasks when there is no way for them to run */
7550 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7556 #else /* !CONFIG_RT_GROUP_SCHED */
7557 static int sched_rt_global_constraints(void)
7559 unsigned long flags;
7562 if (sysctl_sched_rt_period <= 0)
7566 * There's always some RT tasks in the root group
7567 * -- migration, kstopmachine etc..
7569 if (sysctl_sched_rt_runtime == 0)
7572 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7573 for_each_possible_cpu(i) {
7574 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7576 raw_spin_lock(&rt_rq->rt_runtime_lock);
7577 rt_rq->rt_runtime = global_rt_runtime();
7578 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7580 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7584 #endif /* CONFIG_RT_GROUP_SCHED */
7586 int sched_rt_handler(struct ctl_table *table, int write,
7587 void __user *buffer, size_t *lenp,
7591 int old_period, old_runtime;
7592 static DEFINE_MUTEX(mutex);
7595 old_period = sysctl_sched_rt_period;
7596 old_runtime = sysctl_sched_rt_runtime;
7598 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7600 if (!ret && write) {
7601 ret = sched_rt_global_constraints();
7603 sysctl_sched_rt_period = old_period;
7604 sysctl_sched_rt_runtime = old_runtime;
7606 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7607 def_rt_bandwidth.rt_period =
7608 ns_to_ktime(global_rt_period());
7611 mutex_unlock(&mutex);
7616 #ifdef CONFIG_CGROUP_SCHED
7618 /* return corresponding task_group object of a cgroup */
7619 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7621 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7622 struct task_group, css);
7625 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7627 struct task_group *tg, *parent;
7629 if (!cgrp->parent) {
7630 /* This is early initialization for the top cgroup */
7631 return &root_task_group.css;
7634 parent = cgroup_tg(cgrp->parent);
7635 tg = sched_create_group(parent);
7637 return ERR_PTR(-ENOMEM);
7642 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7644 struct task_group *tg = cgroup_tg(cgrp);
7646 sched_destroy_group(tg);
7649 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7650 struct cgroup_taskset *tset)
7652 struct task_struct *task;
7654 cgroup_taskset_for_each(task, cgrp, tset) {
7655 #ifdef CONFIG_RT_GROUP_SCHED
7656 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7659 /* We don't support RT-tasks being in separate groups */
7660 if (task->sched_class != &fair_sched_class)
7667 static void cpu_cgroup_attach(struct cgroup *cgrp,
7668 struct cgroup_taskset *tset)
7670 struct task_struct *task;
7672 cgroup_taskset_for_each(task, cgrp, tset)
7673 sched_move_task(task);
7677 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7678 struct task_struct *task)
7681 * cgroup_exit() is called in the copy_process() failure path.
7682 * Ignore this case since the task hasn't ran yet, this avoids
7683 * trying to poke a half freed task state from generic code.
7685 if (!(task->flags & PF_EXITING))
7688 sched_move_task(task);
7691 #ifdef CONFIG_FAIR_GROUP_SCHED
7692 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7695 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7698 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7700 struct task_group *tg = cgroup_tg(cgrp);
7702 return (u64) scale_load_down(tg->shares);
7705 #ifdef CONFIG_CFS_BANDWIDTH
7706 static DEFINE_MUTEX(cfs_constraints_mutex);
7708 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7709 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7711 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7713 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7715 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7716 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7718 if (tg == &root_task_group)
7722 * Ensure we have at some amount of bandwidth every period. This is
7723 * to prevent reaching a state of large arrears when throttled via
7724 * entity_tick() resulting in prolonged exit starvation.
7726 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7730 * Likewise, bound things on the otherside by preventing insane quota
7731 * periods. This also allows us to normalize in computing quota
7734 if (period > max_cfs_quota_period)
7737 mutex_lock(&cfs_constraints_mutex);
7738 ret = __cfs_schedulable(tg, period, quota);
7742 runtime_enabled = quota != RUNTIME_INF;
7743 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7744 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7745 raw_spin_lock_irq(&cfs_b->lock);
7746 cfs_b->period = ns_to_ktime(period);
7747 cfs_b->quota = quota;
7749 __refill_cfs_bandwidth_runtime(cfs_b);
7750 /* restart the period timer (if active) to handle new period expiry */
7751 if (runtime_enabled && cfs_b->timer_active) {
7752 /* force a reprogram */
7753 cfs_b->timer_active = 0;
7754 __start_cfs_bandwidth(cfs_b);
7756 raw_spin_unlock_irq(&cfs_b->lock);
7758 for_each_possible_cpu(i) {
7759 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7760 struct rq *rq = cfs_rq->rq;
7762 raw_spin_lock_irq(&rq->lock);
7763 cfs_rq->runtime_enabled = runtime_enabled;
7764 cfs_rq->runtime_remaining = 0;
7766 if (cfs_rq->throttled)
7767 unthrottle_cfs_rq(cfs_rq);
7768 raw_spin_unlock_irq(&rq->lock);
7771 mutex_unlock(&cfs_constraints_mutex);
7776 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7780 period = ktime_to_ns(tg->cfs_bandwidth.period);
7781 if (cfs_quota_us < 0)
7782 quota = RUNTIME_INF;
7784 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7786 return tg_set_cfs_bandwidth(tg, period, quota);
7789 long tg_get_cfs_quota(struct task_group *tg)
7793 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7796 quota_us = tg->cfs_bandwidth.quota;
7797 do_div(quota_us, NSEC_PER_USEC);
7802 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7806 period = (u64)cfs_period_us * NSEC_PER_USEC;
7807 quota = tg->cfs_bandwidth.quota;
7809 return tg_set_cfs_bandwidth(tg, period, quota);
7812 long tg_get_cfs_period(struct task_group *tg)
7816 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7817 do_div(cfs_period_us, NSEC_PER_USEC);
7819 return cfs_period_us;
7822 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7824 return tg_get_cfs_quota(cgroup_tg(cgrp));
7827 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7830 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7833 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7835 return tg_get_cfs_period(cgroup_tg(cgrp));
7838 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7841 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7844 struct cfs_schedulable_data {
7845 struct task_group *tg;
7850 * normalize group quota/period to be quota/max_period
7851 * note: units are usecs
7853 static u64 normalize_cfs_quota(struct task_group *tg,
7854 struct cfs_schedulable_data *d)
7862 period = tg_get_cfs_period(tg);
7863 quota = tg_get_cfs_quota(tg);
7866 /* note: these should typically be equivalent */
7867 if (quota == RUNTIME_INF || quota == -1)
7870 return to_ratio(period, quota);
7873 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7875 struct cfs_schedulable_data *d = data;
7876 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7877 s64 quota = 0, parent_quota = -1;
7880 quota = RUNTIME_INF;
7882 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7884 quota = normalize_cfs_quota(tg, d);
7885 parent_quota = parent_b->hierarchal_quota;
7888 * ensure max(child_quota) <= parent_quota, inherit when no
7891 if (quota == RUNTIME_INF)
7892 quota = parent_quota;
7893 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7896 cfs_b->hierarchal_quota = quota;
7901 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7904 struct cfs_schedulable_data data = {
7910 if (quota != RUNTIME_INF) {
7911 do_div(data.period, NSEC_PER_USEC);
7912 do_div(data.quota, NSEC_PER_USEC);
7916 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7922 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7923 struct cgroup_map_cb *cb)
7925 struct task_group *tg = cgroup_tg(cgrp);
7926 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7928 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7929 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7930 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7934 #endif /* CONFIG_CFS_BANDWIDTH */
7935 #endif /* CONFIG_FAIR_GROUP_SCHED */
7937 #ifdef CONFIG_RT_GROUP_SCHED
7938 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7941 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7944 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7946 return sched_group_rt_runtime(cgroup_tg(cgrp));
7949 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7952 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7955 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7957 return sched_group_rt_period(cgroup_tg(cgrp));
7959 #endif /* CONFIG_RT_GROUP_SCHED */
7961 static struct cftype cpu_files[] = {
7962 #ifdef CONFIG_FAIR_GROUP_SCHED
7965 .read_u64 = cpu_shares_read_u64,
7966 .write_u64 = cpu_shares_write_u64,
7969 #ifdef CONFIG_CFS_BANDWIDTH
7971 .name = "cfs_quota_us",
7972 .read_s64 = cpu_cfs_quota_read_s64,
7973 .write_s64 = cpu_cfs_quota_write_s64,
7976 .name = "cfs_period_us",
7977 .read_u64 = cpu_cfs_period_read_u64,
7978 .write_u64 = cpu_cfs_period_write_u64,
7982 .read_map = cpu_stats_show,
7985 #ifdef CONFIG_RT_GROUP_SCHED
7987 .name = "rt_runtime_us",
7988 .read_s64 = cpu_rt_runtime_read,
7989 .write_s64 = cpu_rt_runtime_write,
7992 .name = "rt_period_us",
7993 .read_u64 = cpu_rt_period_read_uint,
7994 .write_u64 = cpu_rt_period_write_uint,
8000 struct cgroup_subsys cpu_cgroup_subsys = {
8002 .create = cpu_cgroup_create,
8003 .destroy = cpu_cgroup_destroy,
8004 .can_attach = cpu_cgroup_can_attach,
8005 .attach = cpu_cgroup_attach,
8006 .exit = cpu_cgroup_exit,
8007 .subsys_id = cpu_cgroup_subsys_id,
8008 .base_cftypes = cpu_files,
8012 #endif /* CONFIG_CGROUP_SCHED */
8014 #ifdef CONFIG_CGROUP_CPUACCT
8017 * CPU accounting code for task groups.
8019 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8020 * (balbir@in.ibm.com).
8023 /* create a new cpu accounting group */
8024 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8029 return &root_cpuacct.css;
8031 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8035 ca->cpuusage = alloc_percpu(u64);
8039 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8041 goto out_free_cpuusage;
8046 free_percpu(ca->cpuusage);
8050 return ERR_PTR(-ENOMEM);
8053 /* destroy an existing cpu accounting group */
8054 static void cpuacct_destroy(struct cgroup *cgrp)
8056 struct cpuacct *ca = cgroup_ca(cgrp);
8058 free_percpu(ca->cpustat);
8059 free_percpu(ca->cpuusage);
8063 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8065 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8068 #ifndef CONFIG_64BIT
8070 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8072 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8074 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8082 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8084 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8086 #ifndef CONFIG_64BIT
8088 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8090 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8092 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8098 /* return total cpu usage (in nanoseconds) of a group */
8099 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8101 struct cpuacct *ca = cgroup_ca(cgrp);
8102 u64 totalcpuusage = 0;
8105 for_each_present_cpu(i)
8106 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8108 return totalcpuusage;
8111 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8114 struct cpuacct *ca = cgroup_ca(cgrp);
8123 for_each_present_cpu(i)
8124 cpuacct_cpuusage_write(ca, i, 0);
8130 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8133 struct cpuacct *ca = cgroup_ca(cgroup);
8137 for_each_present_cpu(i) {
8138 percpu = cpuacct_cpuusage_read(ca, i);
8139 seq_printf(m, "%llu ", (unsigned long long) percpu);
8141 seq_printf(m, "\n");
8145 static const char *cpuacct_stat_desc[] = {
8146 [CPUACCT_STAT_USER] = "user",
8147 [CPUACCT_STAT_SYSTEM] = "system",
8150 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8151 struct cgroup_map_cb *cb)
8153 struct cpuacct *ca = cgroup_ca(cgrp);
8157 for_each_online_cpu(cpu) {
8158 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8159 val += kcpustat->cpustat[CPUTIME_USER];
8160 val += kcpustat->cpustat[CPUTIME_NICE];
8162 val = cputime64_to_clock_t(val);
8163 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8166 for_each_online_cpu(cpu) {
8167 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8168 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8169 val += kcpustat->cpustat[CPUTIME_IRQ];
8170 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8173 val = cputime64_to_clock_t(val);
8174 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8179 static struct cftype files[] = {
8182 .read_u64 = cpuusage_read,
8183 .write_u64 = cpuusage_write,
8186 .name = "usage_percpu",
8187 .read_seq_string = cpuacct_percpu_seq_read,
8191 .read_map = cpuacct_stats_show,
8197 * charge this task's execution time to its accounting group.
8199 * called with rq->lock held.
8201 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8206 if (unlikely(!cpuacct_subsys.active))
8209 cpu = task_cpu(tsk);
8215 for (; ca; ca = parent_ca(ca)) {
8216 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8217 *cpuusage += cputime;
8223 struct cgroup_subsys cpuacct_subsys = {
8225 .create = cpuacct_create,
8226 .destroy = cpuacct_destroy,
8227 .subsys_id = cpuacct_subsys_id,
8228 .base_cftypes = files,
8230 #endif /* CONFIG_CGROUP_CPUACCT */