4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
83 #include "../workqueue_sched.h"
85 #define CREATE_TRACE_POINTS
86 #include <trace/events/sched.h>
88 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
91 ktime_t soft, hard, now;
94 if (hrtimer_active(period_timer))
97 now = hrtimer_cb_get_time(period_timer);
98 hrtimer_forward(period_timer, now, period);
100 soft = hrtimer_get_softexpires(period_timer);
101 hard = hrtimer_get_expires(period_timer);
102 delta = ktime_to_ns(ktime_sub(hard, soft));
103 __hrtimer_start_range_ns(period_timer, soft, delta,
104 HRTIMER_MODE_ABS_PINNED, 0);
108 DEFINE_MUTEX(sched_domains_mutex);
109 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
111 static void update_rq_clock_task(struct rq *rq, s64 delta);
113 void update_rq_clock(struct rq *rq)
117 if (rq->skip_clock_update > 0)
120 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
122 update_rq_clock_task(rq, delta);
126 * Debugging: various feature bits
129 #define SCHED_FEAT(name, enabled) \
130 (1UL << __SCHED_FEAT_##name) * enabled |
132 const_debug unsigned int sysctl_sched_features =
133 #include "features.h"
138 #ifdef CONFIG_SCHED_DEBUG
139 #define SCHED_FEAT(name, enabled) \
142 static __read_mostly char *sched_feat_names[] = {
143 #include "features.h"
149 static int sched_feat_show(struct seq_file *m, void *v)
153 for (i = 0; i < __SCHED_FEAT_NR; i++) {
154 if (!(sysctl_sched_features & (1UL << i)))
156 seq_printf(m, "%s ", sched_feat_names[i]);
163 #ifdef HAVE_JUMP_LABEL
165 #define jump_label_key__true jump_label_key_enabled
166 #define jump_label_key__false jump_label_key_disabled
168 #define SCHED_FEAT(name, enabled) \
169 jump_label_key__##enabled ,
171 struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = {
172 #include "features.h"
177 static void sched_feat_disable(int i)
179 if (jump_label_enabled(&sched_feat_keys[i]))
180 jump_label_dec(&sched_feat_keys[i]);
183 static void sched_feat_enable(int i)
185 if (!jump_label_enabled(&sched_feat_keys[i]))
186 jump_label_inc(&sched_feat_keys[i]);
189 static void sched_feat_disable(int i) { };
190 static void sched_feat_enable(int i) { };
191 #endif /* HAVE_JUMP_LABEL */
194 sched_feat_write(struct file *filp, const char __user *ubuf,
195 size_t cnt, loff_t *ppos)
205 if (copy_from_user(&buf, ubuf, cnt))
211 if (strncmp(cmp, "NO_", 3) == 0) {
216 for (i = 0; i < __SCHED_FEAT_NR; i++) {
217 if (strcmp(cmp, sched_feat_names[i]) == 0) {
219 sysctl_sched_features &= ~(1UL << i);
220 sched_feat_disable(i);
222 sysctl_sched_features |= (1UL << i);
223 sched_feat_enable(i);
229 if (i == __SCHED_FEAT_NR)
237 static int sched_feat_open(struct inode *inode, struct file *filp)
239 return single_open(filp, sched_feat_show, NULL);
242 static const struct file_operations sched_feat_fops = {
243 .open = sched_feat_open,
244 .write = sched_feat_write,
247 .release = single_release,
250 static __init int sched_init_debug(void)
252 debugfs_create_file("sched_features", 0644, NULL, NULL,
257 late_initcall(sched_init_debug);
258 #endif /* CONFIG_SCHED_DEBUG */
261 * Number of tasks to iterate in a single balance run.
262 * Limited because this is done with IRQs disabled.
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
267 * period over which we average the RT time consumption, measured
272 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275 * period over which we measure -rt task cpu usage in us.
278 unsigned int sysctl_sched_rt_period = 1000000;
280 __read_mostly int scheduler_running;
283 * part of the period that we allow rt tasks to run in us.
286 int sysctl_sched_rt_runtime = 950000;
291 * __task_rq_lock - lock the rq @p resides on.
293 static inline struct rq *__task_rq_lock(struct task_struct *p)
298 lockdep_assert_held(&p->pi_lock);
302 raw_spin_lock(&rq->lock);
303 if (likely(rq == task_rq(p)))
305 raw_spin_unlock(&rq->lock);
310 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
312 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
313 __acquires(p->pi_lock)
319 raw_spin_lock_irqsave(&p->pi_lock, *flags);
321 raw_spin_lock(&rq->lock);
322 if (likely(rq == task_rq(p)))
324 raw_spin_unlock(&rq->lock);
325 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
329 static void __task_rq_unlock(struct rq *rq)
332 raw_spin_unlock(&rq->lock);
336 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
338 __releases(p->pi_lock)
340 raw_spin_unlock(&rq->lock);
341 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
345 * this_rq_lock - lock this runqueue and disable interrupts.
347 static struct rq *this_rq_lock(void)
354 raw_spin_lock(&rq->lock);
359 #ifdef CONFIG_SCHED_HRTICK
361 * Use HR-timers to deliver accurate preemption points.
363 * Its all a bit involved since we cannot program an hrt while holding the
364 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
367 * When we get rescheduled we reprogram the hrtick_timer outside of the
371 static void hrtick_clear(struct rq *rq)
373 if (hrtimer_active(&rq->hrtick_timer))
374 hrtimer_cancel(&rq->hrtick_timer);
378 * High-resolution timer tick.
379 * Runs from hardirq context with interrupts disabled.
381 static enum hrtimer_restart hrtick(struct hrtimer *timer)
383 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
385 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
387 raw_spin_lock(&rq->lock);
389 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
390 raw_spin_unlock(&rq->lock);
392 return HRTIMER_NORESTART;
397 * called from hardirq (IPI) context
399 static void __hrtick_start(void *arg)
403 raw_spin_lock(&rq->lock);
404 hrtimer_restart(&rq->hrtick_timer);
405 rq->hrtick_csd_pending = 0;
406 raw_spin_unlock(&rq->lock);
410 * Called to set the hrtick timer state.
412 * called with rq->lock held and irqs disabled
414 void hrtick_start(struct rq *rq, u64 delay)
416 struct hrtimer *timer = &rq->hrtick_timer;
417 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
419 hrtimer_set_expires(timer, time);
421 if (rq == this_rq()) {
422 hrtimer_restart(timer);
423 } else if (!rq->hrtick_csd_pending) {
424 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
425 rq->hrtick_csd_pending = 1;
430 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
432 int cpu = (int)(long)hcpu;
435 case CPU_UP_CANCELED:
436 case CPU_UP_CANCELED_FROZEN:
437 case CPU_DOWN_PREPARE:
438 case CPU_DOWN_PREPARE_FROZEN:
440 case CPU_DEAD_FROZEN:
441 hrtick_clear(cpu_rq(cpu));
448 static __init void init_hrtick(void)
450 hotcpu_notifier(hotplug_hrtick, 0);
454 * Called to set the hrtick timer state.
456 * called with rq->lock held and irqs disabled
458 void hrtick_start(struct rq *rq, u64 delay)
460 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
461 HRTIMER_MODE_REL_PINNED, 0);
464 static inline void init_hrtick(void)
467 #endif /* CONFIG_SMP */
469 static void init_rq_hrtick(struct rq *rq)
472 rq->hrtick_csd_pending = 0;
474 rq->hrtick_csd.flags = 0;
475 rq->hrtick_csd.func = __hrtick_start;
476 rq->hrtick_csd.info = rq;
479 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
480 rq->hrtick_timer.function = hrtick;
482 #else /* CONFIG_SCHED_HRTICK */
483 static inline void hrtick_clear(struct rq *rq)
487 static inline void init_rq_hrtick(struct rq *rq)
491 static inline void init_hrtick(void)
494 #endif /* CONFIG_SCHED_HRTICK */
497 * resched_task - mark a task 'to be rescheduled now'.
499 * On UP this means the setting of the need_resched flag, on SMP it
500 * might also involve a cross-CPU call to trigger the scheduler on
505 #ifndef tsk_is_polling
506 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
509 void resched_task(struct task_struct *p)
513 assert_raw_spin_locked(&task_rq(p)->lock);
515 if (test_tsk_need_resched(p))
518 set_tsk_need_resched(p);
521 if (cpu == smp_processor_id())
524 /* NEED_RESCHED must be visible before we test polling */
526 if (!tsk_is_polling(p))
527 smp_send_reschedule(cpu);
530 void resched_cpu(int cpu)
532 struct rq *rq = cpu_rq(cpu);
535 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
537 resched_task(cpu_curr(cpu));
538 raw_spin_unlock_irqrestore(&rq->lock, flags);
543 * In the semi idle case, use the nearest busy cpu for migrating timers
544 * from an idle cpu. This is good for power-savings.
546 * We don't do similar optimization for completely idle system, as
547 * selecting an idle cpu will add more delays to the timers than intended
548 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
550 int get_nohz_timer_target(void)
552 int cpu = smp_processor_id();
554 struct sched_domain *sd;
557 for_each_domain(cpu, sd) {
558 for_each_cpu(i, sched_domain_span(sd)) {
570 * When add_timer_on() enqueues a timer into the timer wheel of an
571 * idle CPU then this timer might expire before the next timer event
572 * which is scheduled to wake up that CPU. In case of a completely
573 * idle system the next event might even be infinite time into the
574 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
575 * leaves the inner idle loop so the newly added timer is taken into
576 * account when the CPU goes back to idle and evaluates the timer
577 * wheel for the next timer event.
579 void wake_up_idle_cpu(int cpu)
581 struct rq *rq = cpu_rq(cpu);
583 if (cpu == smp_processor_id())
587 * This is safe, as this function is called with the timer
588 * wheel base lock of (cpu) held. When the CPU is on the way
589 * to idle and has not yet set rq->curr to idle then it will
590 * be serialized on the timer wheel base lock and take the new
591 * timer into account automatically.
593 if (rq->curr != rq->idle)
597 * We can set TIF_RESCHED on the idle task of the other CPU
598 * lockless. The worst case is that the other CPU runs the
599 * idle task through an additional NOOP schedule()
601 set_tsk_need_resched(rq->idle);
603 /* NEED_RESCHED must be visible before we test polling */
605 if (!tsk_is_polling(rq->idle))
606 smp_send_reschedule(cpu);
609 static inline bool got_nohz_idle_kick(void)
611 int cpu = smp_processor_id();
612 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
615 #else /* CONFIG_NO_HZ */
617 static inline bool got_nohz_idle_kick(void)
622 #endif /* CONFIG_NO_HZ */
624 void sched_avg_update(struct rq *rq)
626 s64 period = sched_avg_period();
628 while ((s64)(rq->clock - rq->age_stamp) > period) {
630 * Inline assembly required to prevent the compiler
631 * optimising this loop into a divmod call.
632 * See __iter_div_u64_rem() for another example of this.
634 asm("" : "+rm" (rq->age_stamp));
635 rq->age_stamp += period;
640 #else /* !CONFIG_SMP */
641 void resched_task(struct task_struct *p)
643 assert_raw_spin_locked(&task_rq(p)->lock);
644 set_tsk_need_resched(p);
646 #endif /* CONFIG_SMP */
648 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
651 * Iterate task_group tree rooted at *from, calling @down when first entering a
652 * node and @up when leaving it for the final time.
654 * Caller must hold rcu_lock or sufficient equivalent.
656 int walk_tg_tree_from(struct task_group *from,
657 tg_visitor down, tg_visitor up, void *data)
659 struct task_group *parent, *child;
665 ret = (*down)(parent, data);
668 list_for_each_entry_rcu(child, &parent->children, siblings) {
675 ret = (*up)(parent, data);
676 if (ret || parent == from)
680 parent = parent->parent;
687 int tg_nop(struct task_group *tg, void *data)
693 void update_cpu_load(struct rq *this_rq);
695 static void set_load_weight(struct task_struct *p)
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
713 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
720 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
727 void activate_task(struct rq *rq, struct task_struct *p, int flags)
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
732 enqueue_task(rq, p, flags);
735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
740 dequeue_task(rq, p, flags);
743 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
756 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757 static DEFINE_PER_CPU(u64, cpu_softirq_time);
759 static DEFINE_PER_CPU(u64, irq_start_time);
760 static int sched_clock_irqtime;
762 void enable_sched_clock_irqtime(void)
764 sched_clock_irqtime = 1;
767 void disable_sched_clock_irqtime(void)
769 sched_clock_irqtime = 0;
773 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
775 static inline void irq_time_write_begin(void)
777 __this_cpu_inc(irq_time_seq.sequence);
781 static inline void irq_time_write_end(void)
784 __this_cpu_inc(irq_time_seq.sequence);
787 static inline u64 irq_time_read(int cpu)
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
800 #else /* CONFIG_64BIT */
801 static inline void irq_time_write_begin(void)
805 static inline void irq_time_write_end(void)
809 static inline u64 irq_time_read(int cpu)
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
813 #endif /* CONFIG_64BIT */
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
819 void account_system_vtime(struct task_struct *curr)
825 if (!sched_clock_irqtime)
828 local_irq_save(flags);
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
834 irq_time_write_begin();
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
846 irq_time_write_end();
847 local_irq_restore(flags);
849 EXPORT_SYMBOL_GPL(account_system_vtime);
851 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
853 #ifdef CONFIG_PARAVIRT
854 static inline u64 steal_ticks(u64 steal)
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
863 static void update_rq_clock_task(struct rq *rq, s64 delta)
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
890 if (irq_delta > delta)
893 rq->prev_irq_time += irq_delta;
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_branch((¶virt_steal_rq_enabled))) {
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
903 if (unlikely(steal > delta))
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
909 rq->prev_steal_time_rq += steal;
915 rq->clock_task += delta;
917 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
923 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
924 static int irqtime_account_hi_update(void)
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
935 local_irq_restore(flags);
939 static int irqtime_account_si_update(void)
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
950 local_irq_restore(flags);
954 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
956 #define sched_clock_irqtime (0)
960 void sched_set_stop_task(int cpu, struct task_struct *stop)
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
976 stop->sched_class = &stop_sched_class;
979 cpu_rq(cpu)->stop = stop;
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
986 old_stop->sched_class = &rt_sched_class;
991 * __normal_prio - return the priority that is based on the static prio
993 static inline int __normal_prio(struct task_struct *p)
995 return p->static_prio;
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1005 static inline int normal_prio(struct task_struct *p)
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1012 prio = __normal_prio(p);
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1023 static int effective_prio(struct task_struct *p)
1025 p->normal_prio = normal_prio(p);
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1040 inline int task_curr(const struct task_struct *p)
1042 return cpu_curr(task_cpu(p)) == p;
1045 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1057 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1059 const struct sched_class *class;
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1085 #ifdef CONFIG_SCHED_DEBUG
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1093 #ifdef CONFIG_LOCKDEP
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1099 * see set_task_rq().
1101 * Furthermore, all task_rq users should acquire both locks, see
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1109 trace_sched_migrate_task(p, new_cpu);
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1116 __set_task_cpu(p, new_cpu);
1119 struct migration_arg {
1120 struct task_struct *task;
1124 static int migration_cpu_stop(void *data);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1144 unsigned long flags;
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1190 * If it changed from the expected state, bail out now.
1192 if (unlikely(!ncsw))
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1199 * Oops. Go back and try again..
1201 if (unlikely(running)) {
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1247 void kick_process(struct task_struct *p)
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1257 EXPORT_SYMBOL_GPL(kick_process);
1258 #endif /* CONFIG_SMP */
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1264 static int select_fallback_rq(int cpu, struct task_struct *p)
1267 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1269 /* Look for allowed, online CPU in same node. */
1270 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1271 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1274 /* Any allowed, online CPU? */
1275 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1276 if (dest_cpu < nr_cpu_ids)
1279 /* No more Mr. Nice Guy. */
1280 dest_cpu = cpuset_cpus_allowed_fallback(p);
1282 * Don't tell them about moving exiting tasks or
1283 * kernel threads (both mm NULL), since they never
1286 if (p->mm && printk_ratelimit()) {
1287 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1288 task_pid_nr(p), p->comm, cpu);
1295 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1298 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1300 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1303 * In order not to call set_task_cpu() on a blocking task we need
1304 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1307 * Since this is common to all placement strategies, this lives here.
1309 * [ this allows ->select_task() to simply return task_cpu(p) and
1310 * not worry about this generic constraint ]
1312 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1314 cpu = select_fallback_rq(task_cpu(p), p);
1319 static void update_avg(u64 *avg, u64 sample)
1321 s64 diff = sample - *avg;
1327 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1329 #ifdef CONFIG_SCHEDSTATS
1330 struct rq *rq = this_rq();
1333 int this_cpu = smp_processor_id();
1335 if (cpu == this_cpu) {
1336 schedstat_inc(rq, ttwu_local);
1337 schedstat_inc(p, se.statistics.nr_wakeups_local);
1339 struct sched_domain *sd;
1341 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1343 for_each_domain(this_cpu, sd) {
1344 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1345 schedstat_inc(sd, ttwu_wake_remote);
1352 if (wake_flags & WF_MIGRATED)
1353 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1355 #endif /* CONFIG_SMP */
1357 schedstat_inc(rq, ttwu_count);
1358 schedstat_inc(p, se.statistics.nr_wakeups);
1360 if (wake_flags & WF_SYNC)
1361 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1363 #endif /* CONFIG_SCHEDSTATS */
1366 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1368 activate_task(rq, p, en_flags);
1371 /* if a worker is waking up, notify workqueue */
1372 if (p->flags & PF_WQ_WORKER)
1373 wq_worker_waking_up(p, cpu_of(rq));
1377 * Mark the task runnable and perform wakeup-preemption.
1380 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1382 trace_sched_wakeup(p, true);
1383 check_preempt_curr(rq, p, wake_flags);
1385 p->state = TASK_RUNNING;
1387 if (p->sched_class->task_woken)
1388 p->sched_class->task_woken(rq, p);
1390 if (rq->idle_stamp) {
1391 u64 delta = rq->clock - rq->idle_stamp;
1392 u64 max = 2*sysctl_sched_migration_cost;
1397 update_avg(&rq->avg_idle, delta);
1404 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1407 if (p->sched_contributes_to_load)
1408 rq->nr_uninterruptible--;
1411 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1412 ttwu_do_wakeup(rq, p, wake_flags);
1416 * Called in case the task @p isn't fully descheduled from its runqueue,
1417 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1418 * since all we need to do is flip p->state to TASK_RUNNING, since
1419 * the task is still ->on_rq.
1421 static int ttwu_remote(struct task_struct *p, int wake_flags)
1426 rq = __task_rq_lock(p);
1428 ttwu_do_wakeup(rq, p, wake_flags);
1431 __task_rq_unlock(rq);
1437 static void sched_ttwu_pending(void)
1439 struct rq *rq = this_rq();
1440 struct llist_node *llist = llist_del_all(&rq->wake_list);
1441 struct task_struct *p;
1443 raw_spin_lock(&rq->lock);
1446 p = llist_entry(llist, struct task_struct, wake_entry);
1447 llist = llist_next(llist);
1448 ttwu_do_activate(rq, p, 0);
1451 raw_spin_unlock(&rq->lock);
1454 void scheduler_ipi(void)
1456 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1460 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1461 * traditionally all their work was done from the interrupt return
1462 * path. Now that we actually do some work, we need to make sure
1465 * Some archs already do call them, luckily irq_enter/exit nest
1468 * Arguably we should visit all archs and update all handlers,
1469 * however a fair share of IPIs are still resched only so this would
1470 * somewhat pessimize the simple resched case.
1473 sched_ttwu_pending();
1476 * Check if someone kicked us for doing the nohz idle load balance.
1478 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1479 this_rq()->idle_balance = 1;
1480 raise_softirq_irqoff(SCHED_SOFTIRQ);
1485 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1487 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1488 smp_send_reschedule(cpu);
1491 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1492 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1497 rq = __task_rq_lock(p);
1499 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1500 ttwu_do_wakeup(rq, p, wake_flags);
1503 __task_rq_unlock(rq);
1508 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1510 static inline int ttwu_share_cache(int this_cpu, int that_cpu)
1512 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1514 #endif /* CONFIG_SMP */
1516 static void ttwu_queue(struct task_struct *p, int cpu)
1518 struct rq *rq = cpu_rq(cpu);
1520 #if defined(CONFIG_SMP)
1521 if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
1522 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1523 ttwu_queue_remote(p, cpu);
1528 raw_spin_lock(&rq->lock);
1529 ttwu_do_activate(rq, p, 0);
1530 raw_spin_unlock(&rq->lock);
1534 * try_to_wake_up - wake up a thread
1535 * @p: the thread to be awakened
1536 * @state: the mask of task states that can be woken
1537 * @wake_flags: wake modifier flags (WF_*)
1539 * Put it on the run-queue if it's not already there. The "current"
1540 * thread is always on the run-queue (except when the actual
1541 * re-schedule is in progress), and as such you're allowed to do
1542 * the simpler "current->state = TASK_RUNNING" to mark yourself
1543 * runnable without the overhead of this.
1545 * Returns %true if @p was woken up, %false if it was already running
1546 * or @state didn't match @p's state.
1549 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1551 unsigned long flags;
1552 int cpu, success = 0;
1555 raw_spin_lock_irqsave(&p->pi_lock, flags);
1556 if (!(p->state & state))
1559 success = 1; /* we're going to change ->state */
1562 if (p->on_rq && ttwu_remote(p, wake_flags))
1567 * If the owning (remote) cpu is still in the middle of schedule() with
1568 * this task as prev, wait until its done referencing the task.
1571 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1573 * In case the architecture enables interrupts in
1574 * context_switch(), we cannot busy wait, since that
1575 * would lead to deadlocks when an interrupt hits and
1576 * tries to wake up @prev. So bail and do a complete
1579 if (ttwu_activate_remote(p, wake_flags))
1586 * Pairs with the smp_wmb() in finish_lock_switch().
1590 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1591 p->state = TASK_WAKING;
1593 if (p->sched_class->task_waking)
1594 p->sched_class->task_waking(p);
1596 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1597 if (task_cpu(p) != cpu) {
1598 wake_flags |= WF_MIGRATED;
1599 set_task_cpu(p, cpu);
1601 #endif /* CONFIG_SMP */
1605 ttwu_stat(p, cpu, wake_flags);
1607 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1613 * try_to_wake_up_local - try to wake up a local task with rq lock held
1614 * @p: the thread to be awakened
1616 * Put @p on the run-queue if it's not already there. The caller must
1617 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1620 static void try_to_wake_up_local(struct task_struct *p)
1622 struct rq *rq = task_rq(p);
1624 BUG_ON(rq != this_rq());
1625 BUG_ON(p == current);
1626 lockdep_assert_held(&rq->lock);
1628 if (!raw_spin_trylock(&p->pi_lock)) {
1629 raw_spin_unlock(&rq->lock);
1630 raw_spin_lock(&p->pi_lock);
1631 raw_spin_lock(&rq->lock);
1634 if (!(p->state & TASK_NORMAL))
1638 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1640 ttwu_do_wakeup(rq, p, 0);
1641 ttwu_stat(p, smp_processor_id(), 0);
1643 raw_spin_unlock(&p->pi_lock);
1647 * wake_up_process - Wake up a specific process
1648 * @p: The process to be woken up.
1650 * Attempt to wake up the nominated process and move it to the set of runnable
1651 * processes. Returns 1 if the process was woken up, 0 if it was already
1654 * It may be assumed that this function implies a write memory barrier before
1655 * changing the task state if and only if any tasks are woken up.
1657 int wake_up_process(struct task_struct *p)
1659 return try_to_wake_up(p, TASK_ALL, 0);
1661 EXPORT_SYMBOL(wake_up_process);
1663 int wake_up_state(struct task_struct *p, unsigned int state)
1665 return try_to_wake_up(p, state, 0);
1669 * Perform scheduler related setup for a newly forked process p.
1670 * p is forked by current.
1672 * __sched_fork() is basic setup used by init_idle() too:
1674 static void __sched_fork(struct task_struct *p)
1679 p->se.exec_start = 0;
1680 p->se.sum_exec_runtime = 0;
1681 p->se.prev_sum_exec_runtime = 0;
1682 p->se.nr_migrations = 0;
1684 INIT_LIST_HEAD(&p->se.group_node);
1686 #ifdef CONFIG_SCHEDSTATS
1687 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1690 INIT_LIST_HEAD(&p->rt.run_list);
1692 #ifdef CONFIG_PREEMPT_NOTIFIERS
1693 INIT_HLIST_HEAD(&p->preempt_notifiers);
1698 * fork()/clone()-time setup:
1700 void sched_fork(struct task_struct *p)
1702 unsigned long flags;
1703 int cpu = get_cpu();
1707 * We mark the process as running here. This guarantees that
1708 * nobody will actually run it, and a signal or other external
1709 * event cannot wake it up and insert it on the runqueue either.
1711 p->state = TASK_RUNNING;
1714 * Make sure we do not leak PI boosting priority to the child.
1716 p->prio = current->normal_prio;
1719 * Revert to default priority/policy on fork if requested.
1721 if (unlikely(p->sched_reset_on_fork)) {
1722 if (task_has_rt_policy(p)) {
1723 p->policy = SCHED_NORMAL;
1724 p->static_prio = NICE_TO_PRIO(0);
1726 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1727 p->static_prio = NICE_TO_PRIO(0);
1729 p->prio = p->normal_prio = __normal_prio(p);
1733 * We don't need the reset flag anymore after the fork. It has
1734 * fulfilled its duty:
1736 p->sched_reset_on_fork = 0;
1739 if (!rt_prio(p->prio))
1740 p->sched_class = &fair_sched_class;
1742 if (p->sched_class->task_fork)
1743 p->sched_class->task_fork(p);
1746 * The child is not yet in the pid-hash so no cgroup attach races,
1747 * and the cgroup is pinned to this child due to cgroup_fork()
1748 * is ran before sched_fork().
1750 * Silence PROVE_RCU.
1752 raw_spin_lock_irqsave(&p->pi_lock, flags);
1753 set_task_cpu(p, cpu);
1754 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1756 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1757 if (likely(sched_info_on()))
1758 memset(&p->sched_info, 0, sizeof(p->sched_info));
1760 #if defined(CONFIG_SMP)
1763 #ifdef CONFIG_PREEMPT_COUNT
1764 /* Want to start with kernel preemption disabled. */
1765 task_thread_info(p)->preempt_count = 1;
1768 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1775 * wake_up_new_task - wake up a newly created task for the first time.
1777 * This function will do some initial scheduler statistics housekeeping
1778 * that must be done for every newly created context, then puts the task
1779 * on the runqueue and wakes it.
1781 void wake_up_new_task(struct task_struct *p)
1783 unsigned long flags;
1786 raw_spin_lock_irqsave(&p->pi_lock, flags);
1789 * Fork balancing, do it here and not earlier because:
1790 * - cpus_allowed can change in the fork path
1791 * - any previously selected cpu might disappear through hotplug
1793 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1796 rq = __task_rq_lock(p);
1797 activate_task(rq, p, 0);
1799 trace_sched_wakeup_new(p, true);
1800 check_preempt_curr(rq, p, WF_FORK);
1802 if (p->sched_class->task_woken)
1803 p->sched_class->task_woken(rq, p);
1805 task_rq_unlock(rq, p, &flags);
1808 #ifdef CONFIG_PREEMPT_NOTIFIERS
1811 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1812 * @notifier: notifier struct to register
1814 void preempt_notifier_register(struct preempt_notifier *notifier)
1816 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1818 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1821 * preempt_notifier_unregister - no longer interested in preemption notifications
1822 * @notifier: notifier struct to unregister
1824 * This is safe to call from within a preemption notifier.
1826 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1828 hlist_del(¬ifier->link);
1830 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1832 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1834 struct preempt_notifier *notifier;
1835 struct hlist_node *node;
1837 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1838 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1842 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1843 struct task_struct *next)
1845 struct preempt_notifier *notifier;
1846 struct hlist_node *node;
1848 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1849 notifier->ops->sched_out(notifier, next);
1852 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1854 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1859 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1860 struct task_struct *next)
1864 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1867 * prepare_task_switch - prepare to switch tasks
1868 * @rq: the runqueue preparing to switch
1869 * @prev: the current task that is being switched out
1870 * @next: the task we are going to switch to.
1872 * This is called with the rq lock held and interrupts off. It must
1873 * be paired with a subsequent finish_task_switch after the context
1876 * prepare_task_switch sets up locking and calls architecture specific
1880 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1881 struct task_struct *next)
1883 sched_info_switch(prev, next);
1884 perf_event_task_sched_out(prev, next);
1885 fire_sched_out_preempt_notifiers(prev, next);
1886 prepare_lock_switch(rq, next);
1887 prepare_arch_switch(next);
1888 trace_sched_switch(prev, next);
1892 * finish_task_switch - clean up after a task-switch
1893 * @rq: runqueue associated with task-switch
1894 * @prev: the thread we just switched away from.
1896 * finish_task_switch must be called after the context switch, paired
1897 * with a prepare_task_switch call before the context switch.
1898 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1899 * and do any other architecture-specific cleanup actions.
1901 * Note that we may have delayed dropping an mm in context_switch(). If
1902 * so, we finish that here outside of the runqueue lock. (Doing it
1903 * with the lock held can cause deadlocks; see schedule() for
1906 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1907 __releases(rq->lock)
1909 struct mm_struct *mm = rq->prev_mm;
1915 * A task struct has one reference for the use as "current".
1916 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1917 * schedule one last time. The schedule call will never return, and
1918 * the scheduled task must drop that reference.
1919 * The test for TASK_DEAD must occur while the runqueue locks are
1920 * still held, otherwise prev could be scheduled on another cpu, die
1921 * there before we look at prev->state, and then the reference would
1923 * Manfred Spraul <manfred@colorfullife.com>
1925 prev_state = prev->state;
1926 finish_arch_switch(prev);
1927 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1928 local_irq_disable();
1929 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1930 perf_event_task_sched_in(prev, current);
1931 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1933 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1934 finish_lock_switch(rq, prev);
1935 trace_sched_stat_sleeptime(current, rq->clock);
1937 fire_sched_in_preempt_notifiers(current);
1940 if (unlikely(prev_state == TASK_DEAD)) {
1942 * Remove function-return probe instances associated with this
1943 * task and put them back on the free list.
1945 kprobe_flush_task(prev);
1946 put_task_struct(prev);
1952 /* assumes rq->lock is held */
1953 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1955 if (prev->sched_class->pre_schedule)
1956 prev->sched_class->pre_schedule(rq, prev);
1959 /* rq->lock is NOT held, but preemption is disabled */
1960 static inline void post_schedule(struct rq *rq)
1962 if (rq->post_schedule) {
1963 unsigned long flags;
1965 raw_spin_lock_irqsave(&rq->lock, flags);
1966 if (rq->curr->sched_class->post_schedule)
1967 rq->curr->sched_class->post_schedule(rq);
1968 raw_spin_unlock_irqrestore(&rq->lock, flags);
1970 rq->post_schedule = 0;
1976 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1980 static inline void post_schedule(struct rq *rq)
1987 * schedule_tail - first thing a freshly forked thread must call.
1988 * @prev: the thread we just switched away from.
1990 asmlinkage void schedule_tail(struct task_struct *prev)
1991 __releases(rq->lock)
1993 struct rq *rq = this_rq();
1995 finish_task_switch(rq, prev);
1998 * FIXME: do we need to worry about rq being invalidated by the
2003 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2004 /* In this case, finish_task_switch does not reenable preemption */
2007 if (current->set_child_tid)
2008 put_user(task_pid_vnr(current), current->set_child_tid);
2012 * context_switch - switch to the new MM and the new
2013 * thread's register state.
2016 context_switch(struct rq *rq, struct task_struct *prev,
2017 struct task_struct *next)
2019 struct mm_struct *mm, *oldmm;
2021 prepare_task_switch(rq, prev, next);
2024 oldmm = prev->active_mm;
2026 * For paravirt, this is coupled with an exit in switch_to to
2027 * combine the page table reload and the switch backend into
2030 arch_start_context_switch(prev);
2033 next->active_mm = oldmm;
2034 atomic_inc(&oldmm->mm_count);
2035 enter_lazy_tlb(oldmm, next);
2037 switch_mm(oldmm, mm, next);
2040 prev->active_mm = NULL;
2041 rq->prev_mm = oldmm;
2044 * Since the runqueue lock will be released by the next
2045 * task (which is an invalid locking op but in the case
2046 * of the scheduler it's an obvious special-case), so we
2047 * do an early lockdep release here:
2049 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2050 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2053 /* Here we just switch the register state and the stack. */
2054 switch_to(prev, next, prev);
2058 * this_rq must be evaluated again because prev may have moved
2059 * CPUs since it called schedule(), thus the 'rq' on its stack
2060 * frame will be invalid.
2062 finish_task_switch(this_rq(), prev);
2066 * nr_running, nr_uninterruptible and nr_context_switches:
2068 * externally visible scheduler statistics: current number of runnable
2069 * threads, current number of uninterruptible-sleeping threads, total
2070 * number of context switches performed since bootup.
2072 unsigned long nr_running(void)
2074 unsigned long i, sum = 0;
2076 for_each_online_cpu(i)
2077 sum += cpu_rq(i)->nr_running;
2082 unsigned long nr_uninterruptible(void)
2084 unsigned long i, sum = 0;
2086 for_each_possible_cpu(i)
2087 sum += cpu_rq(i)->nr_uninterruptible;
2090 * Since we read the counters lockless, it might be slightly
2091 * inaccurate. Do not allow it to go below zero though:
2093 if (unlikely((long)sum < 0))
2099 unsigned long long nr_context_switches(void)
2102 unsigned long long sum = 0;
2104 for_each_possible_cpu(i)
2105 sum += cpu_rq(i)->nr_switches;
2110 unsigned long nr_iowait(void)
2112 unsigned long i, sum = 0;
2114 for_each_possible_cpu(i)
2115 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2120 unsigned long nr_iowait_cpu(int cpu)
2122 struct rq *this = cpu_rq(cpu);
2123 return atomic_read(&this->nr_iowait);
2126 unsigned long this_cpu_load(void)
2128 struct rq *this = this_rq();
2129 return this->cpu_load[0];
2133 /* Variables and functions for calc_load */
2134 static atomic_long_t calc_load_tasks;
2135 static unsigned long calc_load_update;
2136 unsigned long avenrun[3];
2137 EXPORT_SYMBOL(avenrun);
2139 static long calc_load_fold_active(struct rq *this_rq)
2141 long nr_active, delta = 0;
2143 nr_active = this_rq->nr_running;
2144 nr_active += (long) this_rq->nr_uninterruptible;
2146 if (nr_active != this_rq->calc_load_active) {
2147 delta = nr_active - this_rq->calc_load_active;
2148 this_rq->calc_load_active = nr_active;
2154 static unsigned long
2155 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2158 load += active * (FIXED_1 - exp);
2159 load += 1UL << (FSHIFT - 1);
2160 return load >> FSHIFT;
2165 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2167 * When making the ILB scale, we should try to pull this in as well.
2169 static atomic_long_t calc_load_tasks_idle;
2171 void calc_load_account_idle(struct rq *this_rq)
2175 delta = calc_load_fold_active(this_rq);
2177 atomic_long_add(delta, &calc_load_tasks_idle);
2180 static long calc_load_fold_idle(void)
2185 * Its got a race, we don't care...
2187 if (atomic_long_read(&calc_load_tasks_idle))
2188 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2194 * fixed_power_int - compute: x^n, in O(log n) time
2196 * @x: base of the power
2197 * @frac_bits: fractional bits of @x
2198 * @n: power to raise @x to.
2200 * By exploiting the relation between the definition of the natural power
2201 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2202 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2203 * (where: n_i \elem {0, 1}, the binary vector representing n),
2204 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2205 * of course trivially computable in O(log_2 n), the length of our binary
2208 static unsigned long
2209 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2211 unsigned long result = 1UL << frac_bits;
2216 result += 1UL << (frac_bits - 1);
2217 result >>= frac_bits;
2223 x += 1UL << (frac_bits - 1);
2231 * a1 = a0 * e + a * (1 - e)
2233 * a2 = a1 * e + a * (1 - e)
2234 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2235 * = a0 * e^2 + a * (1 - e) * (1 + e)
2237 * a3 = a2 * e + a * (1 - e)
2238 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2239 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2243 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2244 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2245 * = a0 * e^n + a * (1 - e^n)
2247 * [1] application of the geometric series:
2250 * S_n := \Sum x^i = -------------
2253 static unsigned long
2254 calc_load_n(unsigned long load, unsigned long exp,
2255 unsigned long active, unsigned int n)
2258 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2262 * NO_HZ can leave us missing all per-cpu ticks calling
2263 * calc_load_account_active(), but since an idle CPU folds its delta into
2264 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2265 * in the pending idle delta if our idle period crossed a load cycle boundary.
2267 * Once we've updated the global active value, we need to apply the exponential
2268 * weights adjusted to the number of cycles missed.
2270 static void calc_global_nohz(unsigned long ticks)
2272 long delta, active, n;
2274 if (time_before(jiffies, calc_load_update))
2278 * If we crossed a calc_load_update boundary, make sure to fold
2279 * any pending idle changes, the respective CPUs might have
2280 * missed the tick driven calc_load_account_active() update
2283 delta = calc_load_fold_idle();
2285 atomic_long_add(delta, &calc_load_tasks);
2288 * If we were idle for multiple load cycles, apply them.
2290 if (ticks >= LOAD_FREQ) {
2291 n = ticks / LOAD_FREQ;
2293 active = atomic_long_read(&calc_load_tasks);
2294 active = active > 0 ? active * FIXED_1 : 0;
2296 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2297 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2298 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2300 calc_load_update += n * LOAD_FREQ;
2304 * Its possible the remainder of the above division also crosses
2305 * a LOAD_FREQ period, the regular check in calc_global_load()
2306 * which comes after this will take care of that.
2308 * Consider us being 11 ticks before a cycle completion, and us
2309 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2310 * age us 4 cycles, and the test in calc_global_load() will
2311 * pick up the final one.
2315 void calc_load_account_idle(struct rq *this_rq)
2319 static inline long calc_load_fold_idle(void)
2324 static void calc_global_nohz(unsigned long ticks)
2330 * get_avenrun - get the load average array
2331 * @loads: pointer to dest load array
2332 * @offset: offset to add
2333 * @shift: shift count to shift the result left
2335 * These values are estimates at best, so no need for locking.
2337 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2339 loads[0] = (avenrun[0] + offset) << shift;
2340 loads[1] = (avenrun[1] + offset) << shift;
2341 loads[2] = (avenrun[2] + offset) << shift;
2345 * calc_load - update the avenrun load estimates 10 ticks after the
2346 * CPUs have updated calc_load_tasks.
2348 void calc_global_load(unsigned long ticks)
2352 calc_global_nohz(ticks);
2354 if (time_before(jiffies, calc_load_update + 10))
2357 active = atomic_long_read(&calc_load_tasks);
2358 active = active > 0 ? active * FIXED_1 : 0;
2360 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2361 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2362 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2364 calc_load_update += LOAD_FREQ;
2368 * Called from update_cpu_load() to periodically update this CPU's
2371 static void calc_load_account_active(struct rq *this_rq)
2375 if (time_before(jiffies, this_rq->calc_load_update))
2378 delta = calc_load_fold_active(this_rq);
2379 delta += calc_load_fold_idle();
2381 atomic_long_add(delta, &calc_load_tasks);
2383 this_rq->calc_load_update += LOAD_FREQ;
2387 * The exact cpuload at various idx values, calculated at every tick would be
2388 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2390 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2391 * on nth tick when cpu may be busy, then we have:
2392 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2393 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2395 * decay_load_missed() below does efficient calculation of
2396 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2397 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2399 * The calculation is approximated on a 128 point scale.
2400 * degrade_zero_ticks is the number of ticks after which load at any
2401 * particular idx is approximated to be zero.
2402 * degrade_factor is a precomputed table, a row for each load idx.
2403 * Each column corresponds to degradation factor for a power of two ticks,
2404 * based on 128 point scale.
2406 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2407 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2409 * With this power of 2 load factors, we can degrade the load n times
2410 * by looking at 1 bits in n and doing as many mult/shift instead of
2411 * n mult/shifts needed by the exact degradation.
2413 #define DEGRADE_SHIFT 7
2414 static const unsigned char
2415 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2416 static const unsigned char
2417 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2418 {0, 0, 0, 0, 0, 0, 0, 0},
2419 {64, 32, 8, 0, 0, 0, 0, 0},
2420 {96, 72, 40, 12, 1, 0, 0},
2421 {112, 98, 75, 43, 15, 1, 0},
2422 {120, 112, 98, 76, 45, 16, 2} };
2425 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2426 * would be when CPU is idle and so we just decay the old load without
2427 * adding any new load.
2429 static unsigned long
2430 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2434 if (!missed_updates)
2437 if (missed_updates >= degrade_zero_ticks[idx])
2441 return load >> missed_updates;
2443 while (missed_updates) {
2444 if (missed_updates % 2)
2445 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2447 missed_updates >>= 1;
2454 * Update rq->cpu_load[] statistics. This function is usually called every
2455 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2456 * every tick. We fix it up based on jiffies.
2458 void update_cpu_load(struct rq *this_rq)
2460 unsigned long this_load = this_rq->load.weight;
2461 unsigned long curr_jiffies = jiffies;
2462 unsigned long pending_updates;
2465 this_rq->nr_load_updates++;
2467 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2468 if (curr_jiffies == this_rq->last_load_update_tick)
2471 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2472 this_rq->last_load_update_tick = curr_jiffies;
2474 /* Update our load: */
2475 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2476 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2477 unsigned long old_load, new_load;
2479 /* scale is effectively 1 << i now, and >> i divides by scale */
2481 old_load = this_rq->cpu_load[i];
2482 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2483 new_load = this_load;
2485 * Round up the averaging division if load is increasing. This
2486 * prevents us from getting stuck on 9 if the load is 10, for
2489 if (new_load > old_load)
2490 new_load += scale - 1;
2492 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2495 sched_avg_update(this_rq);
2498 static void update_cpu_load_active(struct rq *this_rq)
2500 update_cpu_load(this_rq);
2502 calc_load_account_active(this_rq);
2508 * sched_exec - execve() is a valuable balancing opportunity, because at
2509 * this point the task has the smallest effective memory and cache footprint.
2511 void sched_exec(void)
2513 struct task_struct *p = current;
2514 unsigned long flags;
2517 raw_spin_lock_irqsave(&p->pi_lock, flags);
2518 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2519 if (dest_cpu == smp_processor_id())
2522 if (likely(cpu_active(dest_cpu))) {
2523 struct migration_arg arg = { p, dest_cpu };
2525 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2526 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2530 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2535 DEFINE_PER_CPU(struct kernel_stat, kstat);
2536 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2538 EXPORT_PER_CPU_SYMBOL(kstat);
2539 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2542 * Return any ns on the sched_clock that have not yet been accounted in
2543 * @p in case that task is currently running.
2545 * Called with task_rq_lock() held on @rq.
2547 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2551 if (task_current(rq, p)) {
2552 update_rq_clock(rq);
2553 ns = rq->clock_task - p->se.exec_start;
2561 unsigned long long task_delta_exec(struct task_struct *p)
2563 unsigned long flags;
2567 rq = task_rq_lock(p, &flags);
2568 ns = do_task_delta_exec(p, rq);
2569 task_rq_unlock(rq, p, &flags);
2575 * Return accounted runtime for the task.
2576 * In case the task is currently running, return the runtime plus current's
2577 * pending runtime that have not been accounted yet.
2579 unsigned long long task_sched_runtime(struct task_struct *p)
2581 unsigned long flags;
2585 rq = task_rq_lock(p, &flags);
2586 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2587 task_rq_unlock(rq, p, &flags);
2592 #ifdef CONFIG_CGROUP_CPUACCT
2593 struct cgroup_subsys cpuacct_subsys;
2594 struct cpuacct root_cpuacct;
2597 static inline void task_group_account_field(struct task_struct *p, int index,
2600 #ifdef CONFIG_CGROUP_CPUACCT
2601 struct kernel_cpustat *kcpustat;
2605 * Since all updates are sure to touch the root cgroup, we
2606 * get ourselves ahead and touch it first. If the root cgroup
2607 * is the only cgroup, then nothing else should be necessary.
2610 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2612 #ifdef CONFIG_CGROUP_CPUACCT
2613 if (unlikely(!cpuacct_subsys.active))
2618 while (ca && (ca != &root_cpuacct)) {
2619 kcpustat = this_cpu_ptr(ca->cpustat);
2620 kcpustat->cpustat[index] += tmp;
2629 * Account user cpu time to a process.
2630 * @p: the process that the cpu time gets accounted to
2631 * @cputime: the cpu time spent in user space since the last update
2632 * @cputime_scaled: cputime scaled by cpu frequency
2634 void account_user_time(struct task_struct *p, cputime_t cputime,
2635 cputime_t cputime_scaled)
2639 /* Add user time to process. */
2640 p->utime += cputime;
2641 p->utimescaled += cputime_scaled;
2642 account_group_user_time(p, cputime);
2644 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2646 /* Add user time to cpustat. */
2647 task_group_account_field(p, index, (__force u64) cputime);
2649 /* Account for user time used */
2650 acct_update_integrals(p);
2654 * Account guest cpu time to a process.
2655 * @p: the process that the cpu time gets accounted to
2656 * @cputime: the cpu time spent in virtual machine since the last update
2657 * @cputime_scaled: cputime scaled by cpu frequency
2659 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2660 cputime_t cputime_scaled)
2662 u64 *cpustat = kcpustat_this_cpu->cpustat;
2664 /* Add guest time to process. */
2665 p->utime += cputime;
2666 p->utimescaled += cputime_scaled;
2667 account_group_user_time(p, cputime);
2668 p->gtime += cputime;
2670 /* Add guest time to cpustat. */
2671 if (TASK_NICE(p) > 0) {
2672 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2673 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2675 cpustat[CPUTIME_USER] += (__force u64) cputime;
2676 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2681 * Account system cpu time to a process and desired cpustat field
2682 * @p: the process that the cpu time gets accounted to
2683 * @cputime: the cpu time spent in kernel space since the last update
2684 * @cputime_scaled: cputime scaled by cpu frequency
2685 * @target_cputime64: pointer to cpustat field that has to be updated
2688 void __account_system_time(struct task_struct *p, cputime_t cputime,
2689 cputime_t cputime_scaled, int index)
2691 /* Add system time to process. */
2692 p->stime += cputime;
2693 p->stimescaled += cputime_scaled;
2694 account_group_system_time(p, cputime);
2696 /* Add system time to cpustat. */
2697 task_group_account_field(p, index, (__force u64) cputime);
2699 /* Account for system time used */
2700 acct_update_integrals(p);
2704 * Account system cpu time to a process.
2705 * @p: the process that the cpu time gets accounted to
2706 * @hardirq_offset: the offset to subtract from hardirq_count()
2707 * @cputime: the cpu time spent in kernel space since the last update
2708 * @cputime_scaled: cputime scaled by cpu frequency
2710 void account_system_time(struct task_struct *p, int hardirq_offset,
2711 cputime_t cputime, cputime_t cputime_scaled)
2715 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2716 account_guest_time(p, cputime, cputime_scaled);
2720 if (hardirq_count() - hardirq_offset)
2721 index = CPUTIME_IRQ;
2722 else if (in_serving_softirq())
2723 index = CPUTIME_SOFTIRQ;
2725 index = CPUTIME_SYSTEM;
2727 __account_system_time(p, cputime, cputime_scaled, index);
2731 * Account for involuntary wait time.
2732 * @cputime: the cpu time spent in involuntary wait
2734 void account_steal_time(cputime_t cputime)
2736 u64 *cpustat = kcpustat_this_cpu->cpustat;
2738 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2742 * Account for idle time.
2743 * @cputime: the cpu time spent in idle wait
2745 void account_idle_time(cputime_t cputime)
2747 u64 *cpustat = kcpustat_this_cpu->cpustat;
2748 struct rq *rq = this_rq();
2750 if (atomic_read(&rq->nr_iowait) > 0)
2751 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2753 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2756 static __always_inline bool steal_account_process_tick(void)
2758 #ifdef CONFIG_PARAVIRT
2759 if (static_branch(¶virt_steal_enabled)) {
2762 steal = paravirt_steal_clock(smp_processor_id());
2763 steal -= this_rq()->prev_steal_time;
2765 st = steal_ticks(steal);
2766 this_rq()->prev_steal_time += st * TICK_NSEC;
2768 account_steal_time(st);
2775 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2777 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2779 * Account a tick to a process and cpustat
2780 * @p: the process that the cpu time gets accounted to
2781 * @user_tick: is the tick from userspace
2782 * @rq: the pointer to rq
2784 * Tick demultiplexing follows the order
2785 * - pending hardirq update
2786 * - pending softirq update
2790 * - check for guest_time
2791 * - else account as system_time
2793 * Check for hardirq is done both for system and user time as there is
2794 * no timer going off while we are on hardirq and hence we may never get an
2795 * opportunity to update it solely in system time.
2796 * p->stime and friends are only updated on system time and not on irq
2797 * softirq as those do not count in task exec_runtime any more.
2799 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2802 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2803 u64 *cpustat = kcpustat_this_cpu->cpustat;
2805 if (steal_account_process_tick())
2808 if (irqtime_account_hi_update()) {
2809 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2810 } else if (irqtime_account_si_update()) {
2811 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2812 } else if (this_cpu_ksoftirqd() == p) {
2814 * ksoftirqd time do not get accounted in cpu_softirq_time.
2815 * So, we have to handle it separately here.
2816 * Also, p->stime needs to be updated for ksoftirqd.
2818 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2820 } else if (user_tick) {
2821 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2822 } else if (p == rq->idle) {
2823 account_idle_time(cputime_one_jiffy);
2824 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2825 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2827 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2832 static void irqtime_account_idle_ticks(int ticks)
2835 struct rq *rq = this_rq();
2837 for (i = 0; i < ticks; i++)
2838 irqtime_account_process_tick(current, 0, rq);
2840 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2841 static void irqtime_account_idle_ticks(int ticks) {}
2842 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2844 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2847 * Account a single tick of cpu time.
2848 * @p: the process that the cpu time gets accounted to
2849 * @user_tick: indicates if the tick is a user or a system tick
2851 void account_process_tick(struct task_struct *p, int user_tick)
2853 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2854 struct rq *rq = this_rq();
2856 if (sched_clock_irqtime) {
2857 irqtime_account_process_tick(p, user_tick, rq);
2861 if (steal_account_process_tick())
2865 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2866 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2867 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2870 account_idle_time(cputime_one_jiffy);
2874 * Account multiple ticks of steal time.
2875 * @p: the process from which the cpu time has been stolen
2876 * @ticks: number of stolen ticks
2878 void account_steal_ticks(unsigned long ticks)
2880 account_steal_time(jiffies_to_cputime(ticks));
2884 * Account multiple ticks of idle time.
2885 * @ticks: number of stolen ticks
2887 void account_idle_ticks(unsigned long ticks)
2890 if (sched_clock_irqtime) {
2891 irqtime_account_idle_ticks(ticks);
2895 account_idle_time(jiffies_to_cputime(ticks));
2901 * Use precise platform statistics if available:
2903 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2904 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2910 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2912 struct task_cputime cputime;
2914 thread_group_cputime(p, &cputime);
2916 *ut = cputime.utime;
2917 *st = cputime.stime;
2921 #ifndef nsecs_to_cputime
2922 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2925 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2927 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2930 * Use CFS's precise accounting:
2932 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2935 u64 temp = (__force u64) rtime;
2937 temp *= (__force u64) utime;
2938 do_div(temp, (__force u32) total);
2939 utime = (__force cputime_t) temp;
2944 * Compare with previous values, to keep monotonicity:
2946 p->prev_utime = max(p->prev_utime, utime);
2947 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2949 *ut = p->prev_utime;
2950 *st = p->prev_stime;
2954 * Must be called with siglock held.
2956 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2958 struct signal_struct *sig = p->signal;
2959 struct task_cputime cputime;
2960 cputime_t rtime, utime, total;
2962 thread_group_cputime(p, &cputime);
2964 total = cputime.utime + cputime.stime;
2965 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2968 u64 temp = (__force u64) rtime;
2970 temp *= (__force u64) cputime.utime;
2971 do_div(temp, (__force u32) total);
2972 utime = (__force cputime_t) temp;
2976 sig->prev_utime = max(sig->prev_utime, utime);
2977 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
2979 *ut = sig->prev_utime;
2980 *st = sig->prev_stime;
2985 * This function gets called by the timer code, with HZ frequency.
2986 * We call it with interrupts disabled.
2988 void scheduler_tick(void)
2990 int cpu = smp_processor_id();
2991 struct rq *rq = cpu_rq(cpu);
2992 struct task_struct *curr = rq->curr;
2996 raw_spin_lock(&rq->lock);
2997 update_rq_clock(rq);
2998 update_cpu_load_active(rq);
2999 curr->sched_class->task_tick(rq, curr, 0);
3000 raw_spin_unlock(&rq->lock);
3002 perf_event_task_tick();
3005 rq->idle_balance = idle_cpu(cpu);
3006 trigger_load_balance(rq, cpu);
3010 notrace unsigned long get_parent_ip(unsigned long addr)
3012 if (in_lock_functions(addr)) {
3013 addr = CALLER_ADDR2;
3014 if (in_lock_functions(addr))
3015 addr = CALLER_ADDR3;
3020 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3021 defined(CONFIG_PREEMPT_TRACER))
3023 void __kprobes add_preempt_count(int val)
3025 #ifdef CONFIG_DEBUG_PREEMPT
3029 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3032 preempt_count() += val;
3033 #ifdef CONFIG_DEBUG_PREEMPT
3035 * Spinlock count overflowing soon?
3037 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3040 if (preempt_count() == val)
3041 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3043 EXPORT_SYMBOL(add_preempt_count);
3045 void __kprobes sub_preempt_count(int val)
3047 #ifdef CONFIG_DEBUG_PREEMPT
3051 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3054 * Is the spinlock portion underflowing?
3056 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3057 !(preempt_count() & PREEMPT_MASK)))
3061 if (preempt_count() == val)
3062 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3063 preempt_count() -= val;
3065 EXPORT_SYMBOL(sub_preempt_count);
3070 * Print scheduling while atomic bug:
3072 static noinline void __schedule_bug(struct task_struct *prev)
3074 struct pt_regs *regs = get_irq_regs();
3076 if (oops_in_progress)
3079 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3080 prev->comm, prev->pid, preempt_count());
3082 debug_show_held_locks(prev);
3084 if (irqs_disabled())
3085 print_irqtrace_events(prev);
3094 * Various schedule()-time debugging checks and statistics:
3096 static inline void schedule_debug(struct task_struct *prev)
3099 * Test if we are atomic. Since do_exit() needs to call into
3100 * schedule() atomically, we ignore that path for now.
3101 * Otherwise, whine if we are scheduling when we should not be.
3103 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3104 __schedule_bug(prev);
3107 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3109 schedstat_inc(this_rq(), sched_count);
3112 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3114 if (prev->on_rq || rq->skip_clock_update < 0)
3115 update_rq_clock(rq);
3116 prev->sched_class->put_prev_task(rq, prev);
3120 * Pick up the highest-prio task:
3122 static inline struct task_struct *
3123 pick_next_task(struct rq *rq)
3125 const struct sched_class *class;
3126 struct task_struct *p;
3129 * Optimization: we know that if all tasks are in
3130 * the fair class we can call that function directly:
3132 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3133 p = fair_sched_class.pick_next_task(rq);
3138 for_each_class(class) {
3139 p = class->pick_next_task(rq);
3144 BUG(); /* the idle class will always have a runnable task */
3148 * __schedule() is the main scheduler function.
3150 static void __sched __schedule(void)
3152 struct task_struct *prev, *next;
3153 unsigned long *switch_count;
3159 cpu = smp_processor_id();
3161 rcu_note_context_switch(cpu);
3164 schedule_debug(prev);
3166 if (sched_feat(HRTICK))
3169 raw_spin_lock_irq(&rq->lock);
3171 switch_count = &prev->nivcsw;
3172 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3173 if (unlikely(signal_pending_state(prev->state, prev))) {
3174 prev->state = TASK_RUNNING;
3176 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3180 * If a worker went to sleep, notify and ask workqueue
3181 * whether it wants to wake up a task to maintain
3184 if (prev->flags & PF_WQ_WORKER) {
3185 struct task_struct *to_wakeup;
3187 to_wakeup = wq_worker_sleeping(prev, cpu);
3189 try_to_wake_up_local(to_wakeup);
3192 switch_count = &prev->nvcsw;
3195 pre_schedule(rq, prev);
3197 if (unlikely(!rq->nr_running))
3198 idle_balance(cpu, rq);
3200 put_prev_task(rq, prev);
3201 next = pick_next_task(rq);
3202 clear_tsk_need_resched(prev);
3203 rq->skip_clock_update = 0;
3205 if (likely(prev != next)) {
3210 context_switch(rq, prev, next); /* unlocks the rq */
3212 * The context switch have flipped the stack from under us
3213 * and restored the local variables which were saved when
3214 * this task called schedule() in the past. prev == current
3215 * is still correct, but it can be moved to another cpu/rq.
3217 cpu = smp_processor_id();
3220 raw_spin_unlock_irq(&rq->lock);
3224 preempt_enable_no_resched();
3229 static inline void sched_submit_work(struct task_struct *tsk)
3234 * If we are going to sleep and we have plugged IO queued,
3235 * make sure to submit it to avoid deadlocks.
3237 if (blk_needs_flush_plug(tsk))
3238 blk_schedule_flush_plug(tsk);
3241 asmlinkage void __sched schedule(void)
3243 struct task_struct *tsk = current;
3245 sched_submit_work(tsk);
3248 EXPORT_SYMBOL(schedule);
3250 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3252 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3254 if (lock->owner != owner)
3258 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3259 * lock->owner still matches owner, if that fails, owner might
3260 * point to free()d memory, if it still matches, the rcu_read_lock()
3261 * ensures the memory stays valid.
3265 return owner->on_cpu;
3269 * Look out! "owner" is an entirely speculative pointer
3270 * access and not reliable.
3272 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3274 if (!sched_feat(OWNER_SPIN))
3278 while (owner_running(lock, owner)) {
3282 arch_mutex_cpu_relax();
3287 * We break out the loop above on need_resched() and when the
3288 * owner changed, which is a sign for heavy contention. Return
3289 * success only when lock->owner is NULL.
3291 return lock->owner == NULL;
3295 #ifdef CONFIG_PREEMPT
3297 * this is the entry point to schedule() from in-kernel preemption
3298 * off of preempt_enable. Kernel preemptions off return from interrupt
3299 * occur there and call schedule directly.
3301 asmlinkage void __sched notrace preempt_schedule(void)
3303 struct thread_info *ti = current_thread_info();
3306 * If there is a non-zero preempt_count or interrupts are disabled,
3307 * we do not want to preempt the current task. Just return..
3309 if (likely(ti->preempt_count || irqs_disabled()))
3313 add_preempt_count_notrace(PREEMPT_ACTIVE);
3315 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3318 * Check again in case we missed a preemption opportunity
3319 * between schedule and now.
3322 } while (need_resched());
3324 EXPORT_SYMBOL(preempt_schedule);
3327 * this is the entry point to schedule() from kernel preemption
3328 * off of irq context.
3329 * Note, that this is called and return with irqs disabled. This will
3330 * protect us against recursive calling from irq.
3332 asmlinkage void __sched preempt_schedule_irq(void)
3334 struct thread_info *ti = current_thread_info();
3336 /* Catch callers which need to be fixed */
3337 BUG_ON(ti->preempt_count || !irqs_disabled());
3340 add_preempt_count(PREEMPT_ACTIVE);
3343 local_irq_disable();
3344 sub_preempt_count(PREEMPT_ACTIVE);
3347 * Check again in case we missed a preemption opportunity
3348 * between schedule and now.
3351 } while (need_resched());
3354 #endif /* CONFIG_PREEMPT */
3356 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3359 return try_to_wake_up(curr->private, mode, wake_flags);
3361 EXPORT_SYMBOL(default_wake_function);
3364 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3365 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3366 * number) then we wake all the non-exclusive tasks and one exclusive task.
3368 * There are circumstances in which we can try to wake a task which has already
3369 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3370 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3372 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3373 int nr_exclusive, int wake_flags, void *key)
3375 wait_queue_t *curr, *next;
3377 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3378 unsigned flags = curr->flags;
3380 if (curr->func(curr, mode, wake_flags, key) &&
3381 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3387 * __wake_up - wake up threads blocked on a waitqueue.
3389 * @mode: which threads
3390 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3391 * @key: is directly passed to the wakeup function
3393 * It may be assumed that this function implies a write memory barrier before
3394 * changing the task state if and only if any tasks are woken up.
3396 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3397 int nr_exclusive, void *key)
3399 unsigned long flags;
3401 spin_lock_irqsave(&q->lock, flags);
3402 __wake_up_common(q, mode, nr_exclusive, 0, key);
3403 spin_unlock_irqrestore(&q->lock, flags);
3405 EXPORT_SYMBOL(__wake_up);
3408 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3410 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3412 __wake_up_common(q, mode, 1, 0, NULL);
3414 EXPORT_SYMBOL_GPL(__wake_up_locked);
3416 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3418 __wake_up_common(q, mode, 1, 0, key);
3420 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3423 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3425 * @mode: which threads
3426 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3427 * @key: opaque value to be passed to wakeup targets
3429 * The sync wakeup differs that the waker knows that it will schedule
3430 * away soon, so while the target thread will be woken up, it will not
3431 * be migrated to another CPU - ie. the two threads are 'synchronized'
3432 * with each other. This can prevent needless bouncing between CPUs.
3434 * On UP it can prevent extra preemption.
3436 * It may be assumed that this function implies a write memory barrier before
3437 * changing the task state if and only if any tasks are woken up.
3439 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3440 int nr_exclusive, void *key)
3442 unsigned long flags;
3443 int wake_flags = WF_SYNC;
3448 if (unlikely(!nr_exclusive))
3451 spin_lock_irqsave(&q->lock, flags);
3452 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3453 spin_unlock_irqrestore(&q->lock, flags);
3455 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3458 * __wake_up_sync - see __wake_up_sync_key()
3460 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3462 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3464 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3467 * complete: - signals a single thread waiting on this completion
3468 * @x: holds the state of this particular completion
3470 * This will wake up a single thread waiting on this completion. Threads will be
3471 * awakened in the same order in which they were queued.
3473 * See also complete_all(), wait_for_completion() and related routines.
3475 * It may be assumed that this function implies a write memory barrier before
3476 * changing the task state if and only if any tasks are woken up.
3478 void complete(struct completion *x)
3480 unsigned long flags;
3482 spin_lock_irqsave(&x->wait.lock, flags);
3484 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3485 spin_unlock_irqrestore(&x->wait.lock, flags);
3487 EXPORT_SYMBOL(complete);
3490 * complete_all: - signals all threads waiting on this completion
3491 * @x: holds the state of this particular completion
3493 * This will wake up all threads waiting on this particular completion event.
3495 * It may be assumed that this function implies a write memory barrier before
3496 * changing the task state if and only if any tasks are woken up.
3498 void complete_all(struct completion *x)
3500 unsigned long flags;
3502 spin_lock_irqsave(&x->wait.lock, flags);
3503 x->done += UINT_MAX/2;
3504 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3505 spin_unlock_irqrestore(&x->wait.lock, flags);
3507 EXPORT_SYMBOL(complete_all);
3509 static inline long __sched
3510 do_wait_for_common(struct completion *x, long timeout, int state)
3513 DECLARE_WAITQUEUE(wait, current);
3515 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3517 if (signal_pending_state(state, current)) {
3518 timeout = -ERESTARTSYS;
3521 __set_current_state(state);
3522 spin_unlock_irq(&x->wait.lock);
3523 timeout = schedule_timeout(timeout);
3524 spin_lock_irq(&x->wait.lock);
3525 } while (!x->done && timeout);
3526 __remove_wait_queue(&x->wait, &wait);
3531 return timeout ?: 1;
3535 wait_for_common(struct completion *x, long timeout, int state)
3539 spin_lock_irq(&x->wait.lock);
3540 timeout = do_wait_for_common(x, timeout, state);
3541 spin_unlock_irq(&x->wait.lock);
3546 * wait_for_completion: - waits for completion of a task
3547 * @x: holds the state of this particular completion
3549 * This waits to be signaled for completion of a specific task. It is NOT
3550 * interruptible and there is no timeout.
3552 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3553 * and interrupt capability. Also see complete().
3555 void __sched wait_for_completion(struct completion *x)
3557 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3559 EXPORT_SYMBOL(wait_for_completion);
3562 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3563 * @x: holds the state of this particular completion
3564 * @timeout: timeout value in jiffies
3566 * This waits for either a completion of a specific task to be signaled or for a
3567 * specified timeout to expire. The timeout is in jiffies. It is not
3570 * The return value is 0 if timed out, and positive (at least 1, or number of
3571 * jiffies left till timeout) if completed.
3573 unsigned long __sched
3574 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3576 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3578 EXPORT_SYMBOL(wait_for_completion_timeout);
3581 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3582 * @x: holds the state of this particular completion
3584 * This waits for completion of a specific task to be signaled. It is
3587 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3589 int __sched wait_for_completion_interruptible(struct completion *x)
3591 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3592 if (t == -ERESTARTSYS)
3596 EXPORT_SYMBOL(wait_for_completion_interruptible);
3599 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3600 * @x: holds the state of this particular completion
3601 * @timeout: timeout value in jiffies
3603 * This waits for either a completion of a specific task to be signaled or for a
3604 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3606 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3607 * positive (at least 1, or number of jiffies left till timeout) if completed.
3610 wait_for_completion_interruptible_timeout(struct completion *x,
3611 unsigned long timeout)
3613 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3615 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3618 * wait_for_completion_killable: - waits for completion of a task (killable)
3619 * @x: holds the state of this particular completion
3621 * This waits to be signaled for completion of a specific task. It can be
3622 * interrupted by a kill signal.
3624 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3626 int __sched wait_for_completion_killable(struct completion *x)
3628 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3629 if (t == -ERESTARTSYS)
3633 EXPORT_SYMBOL(wait_for_completion_killable);
3636 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3637 * @x: holds the state of this particular completion
3638 * @timeout: timeout value in jiffies
3640 * This waits for either a completion of a specific task to be
3641 * signaled or for a specified timeout to expire. It can be
3642 * interrupted by a kill signal. The timeout is in jiffies.
3644 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3645 * positive (at least 1, or number of jiffies left till timeout) if completed.
3648 wait_for_completion_killable_timeout(struct completion *x,
3649 unsigned long timeout)
3651 return wait_for_common(x, timeout, TASK_KILLABLE);
3653 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3656 * try_wait_for_completion - try to decrement a completion without blocking
3657 * @x: completion structure
3659 * Returns: 0 if a decrement cannot be done without blocking
3660 * 1 if a decrement succeeded.
3662 * If a completion is being used as a counting completion,
3663 * attempt to decrement the counter without blocking. This
3664 * enables us to avoid waiting if the resource the completion
3665 * is protecting is not available.
3667 bool try_wait_for_completion(struct completion *x)
3669 unsigned long flags;
3672 spin_lock_irqsave(&x->wait.lock, flags);
3677 spin_unlock_irqrestore(&x->wait.lock, flags);
3680 EXPORT_SYMBOL(try_wait_for_completion);
3683 * completion_done - Test to see if a completion has any waiters
3684 * @x: completion structure
3686 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3687 * 1 if there are no waiters.
3690 bool completion_done(struct completion *x)
3692 unsigned long flags;
3695 spin_lock_irqsave(&x->wait.lock, flags);
3698 spin_unlock_irqrestore(&x->wait.lock, flags);
3701 EXPORT_SYMBOL(completion_done);
3704 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3706 unsigned long flags;
3709 init_waitqueue_entry(&wait, current);
3711 __set_current_state(state);
3713 spin_lock_irqsave(&q->lock, flags);
3714 __add_wait_queue(q, &wait);
3715 spin_unlock(&q->lock);
3716 timeout = schedule_timeout(timeout);
3717 spin_lock_irq(&q->lock);
3718 __remove_wait_queue(q, &wait);
3719 spin_unlock_irqrestore(&q->lock, flags);
3724 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3726 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3728 EXPORT_SYMBOL(interruptible_sleep_on);
3731 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3733 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3735 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3737 void __sched sleep_on(wait_queue_head_t *q)
3739 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3741 EXPORT_SYMBOL(sleep_on);
3743 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3745 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3747 EXPORT_SYMBOL(sleep_on_timeout);
3749 #ifdef CONFIG_RT_MUTEXES
3752 * rt_mutex_setprio - set the current priority of a task
3754 * @prio: prio value (kernel-internal form)
3756 * This function changes the 'effective' priority of a task. It does
3757 * not touch ->normal_prio like __setscheduler().
3759 * Used by the rt_mutex code to implement priority inheritance logic.
3761 void rt_mutex_setprio(struct task_struct *p, int prio)
3763 int oldprio, on_rq, running;
3765 const struct sched_class *prev_class;
3767 BUG_ON(prio < 0 || prio > MAX_PRIO);
3769 rq = __task_rq_lock(p);
3771 trace_sched_pi_setprio(p, prio);
3773 prev_class = p->sched_class;
3775 running = task_current(rq, p);
3777 dequeue_task(rq, p, 0);
3779 p->sched_class->put_prev_task(rq, p);
3782 p->sched_class = &rt_sched_class;
3784 p->sched_class = &fair_sched_class;
3789 p->sched_class->set_curr_task(rq);
3791 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3793 check_class_changed(rq, p, prev_class, oldprio);
3794 __task_rq_unlock(rq);
3799 void set_user_nice(struct task_struct *p, long nice)
3801 int old_prio, delta, on_rq;
3802 unsigned long flags;
3805 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3808 * We have to be careful, if called from sys_setpriority(),
3809 * the task might be in the middle of scheduling on another CPU.
3811 rq = task_rq_lock(p, &flags);
3813 * The RT priorities are set via sched_setscheduler(), but we still
3814 * allow the 'normal' nice value to be set - but as expected
3815 * it wont have any effect on scheduling until the task is
3816 * SCHED_FIFO/SCHED_RR:
3818 if (task_has_rt_policy(p)) {
3819 p->static_prio = NICE_TO_PRIO(nice);
3824 dequeue_task(rq, p, 0);
3826 p->static_prio = NICE_TO_PRIO(nice);
3829 p->prio = effective_prio(p);
3830 delta = p->prio - old_prio;
3833 enqueue_task(rq, p, 0);
3835 * If the task increased its priority or is running and
3836 * lowered its priority, then reschedule its CPU:
3838 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3839 resched_task(rq->curr);
3842 task_rq_unlock(rq, p, &flags);
3844 EXPORT_SYMBOL(set_user_nice);
3847 * can_nice - check if a task can reduce its nice value
3851 int can_nice(const struct task_struct *p, const int nice)
3853 /* convert nice value [19,-20] to rlimit style value [1,40] */
3854 int nice_rlim = 20 - nice;
3856 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3857 capable(CAP_SYS_NICE));
3860 #ifdef __ARCH_WANT_SYS_NICE
3863 * sys_nice - change the priority of the current process.
3864 * @increment: priority increment
3866 * sys_setpriority is a more generic, but much slower function that
3867 * does similar things.
3869 SYSCALL_DEFINE1(nice, int, increment)
3874 * Setpriority might change our priority at the same moment.
3875 * We don't have to worry. Conceptually one call occurs first
3876 * and we have a single winner.
3878 if (increment < -40)
3883 nice = TASK_NICE(current) + increment;
3889 if (increment < 0 && !can_nice(current, nice))
3892 retval = security_task_setnice(current, nice);
3896 set_user_nice(current, nice);
3903 * task_prio - return the priority value of a given task.
3904 * @p: the task in question.
3906 * This is the priority value as seen by users in /proc.
3907 * RT tasks are offset by -200. Normal tasks are centered
3908 * around 0, value goes from -16 to +15.
3910 int task_prio(const struct task_struct *p)
3912 return p->prio - MAX_RT_PRIO;
3916 * task_nice - return the nice value of a given task.
3917 * @p: the task in question.
3919 int task_nice(const struct task_struct *p)
3921 return TASK_NICE(p);
3923 EXPORT_SYMBOL(task_nice);
3926 * idle_cpu - is a given cpu idle currently?
3927 * @cpu: the processor in question.
3929 int idle_cpu(int cpu)
3931 struct rq *rq = cpu_rq(cpu);
3933 if (rq->curr != rq->idle)
3940 if (!llist_empty(&rq->wake_list))
3948 * idle_task - return the idle task for a given cpu.
3949 * @cpu: the processor in question.
3951 struct task_struct *idle_task(int cpu)
3953 return cpu_rq(cpu)->idle;
3957 * find_process_by_pid - find a process with a matching PID value.
3958 * @pid: the pid in question.
3960 static struct task_struct *find_process_by_pid(pid_t pid)
3962 return pid ? find_task_by_vpid(pid) : current;
3965 /* Actually do priority change: must hold rq lock. */
3967 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3970 p->rt_priority = prio;
3971 p->normal_prio = normal_prio(p);
3972 /* we are holding p->pi_lock already */
3973 p->prio = rt_mutex_getprio(p);
3974 if (rt_prio(p->prio))
3975 p->sched_class = &rt_sched_class;
3977 p->sched_class = &fair_sched_class;
3982 * check the target process has a UID that matches the current process's
3984 static bool check_same_owner(struct task_struct *p)
3986 const struct cred *cred = current_cred(), *pcred;
3990 pcred = __task_cred(p);
3991 if (cred->user->user_ns == pcred->user->user_ns)
3992 match = (cred->euid == pcred->euid ||
3993 cred->euid == pcred->uid);
4000 static int __sched_setscheduler(struct task_struct *p, int policy,
4001 const struct sched_param *param, bool user)
4003 int retval, oldprio, oldpolicy = -1, on_rq, running;
4004 unsigned long flags;
4005 const struct sched_class *prev_class;
4009 /* may grab non-irq protected spin_locks */
4010 BUG_ON(in_interrupt());
4012 /* double check policy once rq lock held */
4014 reset_on_fork = p->sched_reset_on_fork;
4015 policy = oldpolicy = p->policy;
4017 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4018 policy &= ~SCHED_RESET_ON_FORK;
4020 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4021 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4022 policy != SCHED_IDLE)
4027 * Valid priorities for SCHED_FIFO and SCHED_RR are
4028 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4029 * SCHED_BATCH and SCHED_IDLE is 0.
4031 if (param->sched_priority < 0 ||
4032 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4033 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4035 if (rt_policy(policy) != (param->sched_priority != 0))
4039 * Allow unprivileged RT tasks to decrease priority:
4041 if (user && !capable(CAP_SYS_NICE)) {
4042 if (rt_policy(policy)) {
4043 unsigned long rlim_rtprio =
4044 task_rlimit(p, RLIMIT_RTPRIO);
4046 /* can't set/change the rt policy */
4047 if (policy != p->policy && !rlim_rtprio)
4050 /* can't increase priority */
4051 if (param->sched_priority > p->rt_priority &&
4052 param->sched_priority > rlim_rtprio)
4057 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4058 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4060 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4061 if (!can_nice(p, TASK_NICE(p)))
4065 /* can't change other user's priorities */
4066 if (!check_same_owner(p))
4069 /* Normal users shall not reset the sched_reset_on_fork flag */
4070 if (p->sched_reset_on_fork && !reset_on_fork)
4075 retval = security_task_setscheduler(p);
4081 * make sure no PI-waiters arrive (or leave) while we are
4082 * changing the priority of the task:
4084 * To be able to change p->policy safely, the appropriate
4085 * runqueue lock must be held.
4087 rq = task_rq_lock(p, &flags);
4090 * Changing the policy of the stop threads its a very bad idea
4092 if (p == rq->stop) {
4093 task_rq_unlock(rq, p, &flags);
4098 * If not changing anything there's no need to proceed further:
4100 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4101 param->sched_priority == p->rt_priority))) {
4103 __task_rq_unlock(rq);
4104 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4108 #ifdef CONFIG_RT_GROUP_SCHED
4111 * Do not allow realtime tasks into groups that have no runtime
4114 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4115 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4116 !task_group_is_autogroup(task_group(p))) {
4117 task_rq_unlock(rq, p, &flags);
4123 /* recheck policy now with rq lock held */
4124 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4125 policy = oldpolicy = -1;
4126 task_rq_unlock(rq, p, &flags);
4130 running = task_current(rq, p);
4132 dequeue_task(rq, p, 0);
4134 p->sched_class->put_prev_task(rq, p);
4136 p->sched_reset_on_fork = reset_on_fork;
4139 prev_class = p->sched_class;
4140 __setscheduler(rq, p, policy, param->sched_priority);
4143 p->sched_class->set_curr_task(rq);
4145 enqueue_task(rq, p, 0);
4147 check_class_changed(rq, p, prev_class, oldprio);
4148 task_rq_unlock(rq, p, &flags);
4150 rt_mutex_adjust_pi(p);
4156 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4157 * @p: the task in question.
4158 * @policy: new policy.
4159 * @param: structure containing the new RT priority.
4161 * NOTE that the task may be already dead.
4163 int sched_setscheduler(struct task_struct *p, int policy,
4164 const struct sched_param *param)
4166 return __sched_setscheduler(p, policy, param, true);
4168 EXPORT_SYMBOL_GPL(sched_setscheduler);
4171 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4172 * @p: the task in question.
4173 * @policy: new policy.
4174 * @param: structure containing the new RT priority.
4176 * Just like sched_setscheduler, only don't bother checking if the
4177 * current context has permission. For example, this is needed in
4178 * stop_machine(): we create temporary high priority worker threads,
4179 * but our caller might not have that capability.
4181 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4182 const struct sched_param *param)
4184 return __sched_setscheduler(p, policy, param, false);
4188 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4190 struct sched_param lparam;
4191 struct task_struct *p;
4194 if (!param || pid < 0)
4196 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4201 p = find_process_by_pid(pid);
4203 retval = sched_setscheduler(p, policy, &lparam);
4210 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4211 * @pid: the pid in question.
4212 * @policy: new policy.
4213 * @param: structure containing the new RT priority.
4215 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4216 struct sched_param __user *, param)
4218 /* negative values for policy are not valid */
4222 return do_sched_setscheduler(pid, policy, param);
4226 * sys_sched_setparam - set/change the RT priority of a thread
4227 * @pid: the pid in question.
4228 * @param: structure containing the new RT priority.
4230 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4232 return do_sched_setscheduler(pid, -1, param);
4236 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4237 * @pid: the pid in question.
4239 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4241 struct task_struct *p;
4249 p = find_process_by_pid(pid);
4251 retval = security_task_getscheduler(p);
4254 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4261 * sys_sched_getparam - get the RT priority of a thread
4262 * @pid: the pid in question.
4263 * @param: structure containing the RT priority.
4265 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4267 struct sched_param lp;
4268 struct task_struct *p;
4271 if (!param || pid < 0)
4275 p = find_process_by_pid(pid);
4280 retval = security_task_getscheduler(p);
4284 lp.sched_priority = p->rt_priority;
4288 * This one might sleep, we cannot do it with a spinlock held ...
4290 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4299 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4301 cpumask_var_t cpus_allowed, new_mask;
4302 struct task_struct *p;
4308 p = find_process_by_pid(pid);
4315 /* Prevent p going away */
4319 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4323 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4325 goto out_free_cpus_allowed;
4328 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4331 retval = security_task_setscheduler(p);
4335 cpuset_cpus_allowed(p, cpus_allowed);
4336 cpumask_and(new_mask, in_mask, cpus_allowed);
4338 retval = set_cpus_allowed_ptr(p, new_mask);
4341 cpuset_cpus_allowed(p, cpus_allowed);
4342 if (!cpumask_subset(new_mask, cpus_allowed)) {
4344 * We must have raced with a concurrent cpuset
4345 * update. Just reset the cpus_allowed to the
4346 * cpuset's cpus_allowed
4348 cpumask_copy(new_mask, cpus_allowed);
4353 free_cpumask_var(new_mask);
4354 out_free_cpus_allowed:
4355 free_cpumask_var(cpus_allowed);
4362 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4363 struct cpumask *new_mask)
4365 if (len < cpumask_size())
4366 cpumask_clear(new_mask);
4367 else if (len > cpumask_size())
4368 len = cpumask_size();
4370 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4374 * sys_sched_setaffinity - set the cpu affinity of a process
4375 * @pid: pid of the process
4376 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4377 * @user_mask_ptr: user-space pointer to the new cpu mask
4379 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4380 unsigned long __user *, user_mask_ptr)
4382 cpumask_var_t new_mask;
4385 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4388 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4390 retval = sched_setaffinity(pid, new_mask);
4391 free_cpumask_var(new_mask);
4395 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4397 struct task_struct *p;
4398 unsigned long flags;
4405 p = find_process_by_pid(pid);
4409 retval = security_task_getscheduler(p);
4413 raw_spin_lock_irqsave(&p->pi_lock, flags);
4414 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4415 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4425 * sys_sched_getaffinity - get the cpu affinity of a process
4426 * @pid: pid of the process
4427 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4428 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4430 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4431 unsigned long __user *, user_mask_ptr)
4436 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4438 if (len & (sizeof(unsigned long)-1))
4441 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4444 ret = sched_getaffinity(pid, mask);
4446 size_t retlen = min_t(size_t, len, cpumask_size());
4448 if (copy_to_user(user_mask_ptr, mask, retlen))
4453 free_cpumask_var(mask);
4459 * sys_sched_yield - yield the current processor to other threads.
4461 * This function yields the current CPU to other tasks. If there are no
4462 * other threads running on this CPU then this function will return.
4464 SYSCALL_DEFINE0(sched_yield)
4466 struct rq *rq = this_rq_lock();
4468 schedstat_inc(rq, yld_count);
4469 current->sched_class->yield_task(rq);
4472 * Since we are going to call schedule() anyway, there's
4473 * no need to preempt or enable interrupts:
4475 __release(rq->lock);
4476 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4477 do_raw_spin_unlock(&rq->lock);
4478 preempt_enable_no_resched();
4485 static inline int should_resched(void)
4487 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4490 static void __cond_resched(void)
4492 add_preempt_count(PREEMPT_ACTIVE);
4494 sub_preempt_count(PREEMPT_ACTIVE);
4497 int __sched _cond_resched(void)
4499 if (should_resched()) {
4505 EXPORT_SYMBOL(_cond_resched);
4508 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4509 * call schedule, and on return reacquire the lock.
4511 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4512 * operations here to prevent schedule() from being called twice (once via
4513 * spin_unlock(), once by hand).
4515 int __cond_resched_lock(spinlock_t *lock)
4517 int resched = should_resched();
4520 lockdep_assert_held(lock);
4522 if (spin_needbreak(lock) || resched) {
4533 EXPORT_SYMBOL(__cond_resched_lock);
4535 int __sched __cond_resched_softirq(void)
4537 BUG_ON(!in_softirq());
4539 if (should_resched()) {
4547 EXPORT_SYMBOL(__cond_resched_softirq);
4550 * yield - yield the current processor to other threads.
4552 * This is a shortcut for kernel-space yielding - it marks the
4553 * thread runnable and calls sys_sched_yield().
4555 void __sched yield(void)
4557 set_current_state(TASK_RUNNING);
4560 EXPORT_SYMBOL(yield);
4563 * yield_to - yield the current processor to another thread in
4564 * your thread group, or accelerate that thread toward the
4565 * processor it's on.
4567 * @preempt: whether task preemption is allowed or not
4569 * It's the caller's job to ensure that the target task struct
4570 * can't go away on us before we can do any checks.
4572 * Returns true if we indeed boosted the target task.
4574 bool __sched yield_to(struct task_struct *p, bool preempt)
4576 struct task_struct *curr = current;
4577 struct rq *rq, *p_rq;
4578 unsigned long flags;
4581 local_irq_save(flags);
4586 double_rq_lock(rq, p_rq);
4587 while (task_rq(p) != p_rq) {
4588 double_rq_unlock(rq, p_rq);
4592 if (!curr->sched_class->yield_to_task)
4595 if (curr->sched_class != p->sched_class)
4598 if (task_running(p_rq, p) || p->state)
4601 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4603 schedstat_inc(rq, yld_count);
4605 * Make p's CPU reschedule; pick_next_entity takes care of
4608 if (preempt && rq != p_rq)
4609 resched_task(p_rq->curr);
4612 * We might have set it in task_yield_fair(), but are
4613 * not going to schedule(), so don't want to skip
4616 rq->skip_clock_update = 0;
4620 double_rq_unlock(rq, p_rq);
4621 local_irq_restore(flags);
4628 EXPORT_SYMBOL_GPL(yield_to);
4631 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4632 * that process accounting knows that this is a task in IO wait state.
4634 void __sched io_schedule(void)
4636 struct rq *rq = raw_rq();
4638 delayacct_blkio_start();
4639 atomic_inc(&rq->nr_iowait);
4640 blk_flush_plug(current);
4641 current->in_iowait = 1;
4643 current->in_iowait = 0;
4644 atomic_dec(&rq->nr_iowait);
4645 delayacct_blkio_end();
4647 EXPORT_SYMBOL(io_schedule);
4649 long __sched io_schedule_timeout(long timeout)
4651 struct rq *rq = raw_rq();
4654 delayacct_blkio_start();
4655 atomic_inc(&rq->nr_iowait);
4656 blk_flush_plug(current);
4657 current->in_iowait = 1;
4658 ret = schedule_timeout(timeout);
4659 current->in_iowait = 0;
4660 atomic_dec(&rq->nr_iowait);
4661 delayacct_blkio_end();
4666 * sys_sched_get_priority_max - return maximum RT priority.
4667 * @policy: scheduling class.
4669 * this syscall returns the maximum rt_priority that can be used
4670 * by a given scheduling class.
4672 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4679 ret = MAX_USER_RT_PRIO-1;
4691 * sys_sched_get_priority_min - return minimum RT priority.
4692 * @policy: scheduling class.
4694 * this syscall returns the minimum rt_priority that can be used
4695 * by a given scheduling class.
4697 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4715 * sys_sched_rr_get_interval - return the default timeslice of a process.
4716 * @pid: pid of the process.
4717 * @interval: userspace pointer to the timeslice value.
4719 * this syscall writes the default timeslice value of a given process
4720 * into the user-space timespec buffer. A value of '0' means infinity.
4722 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4723 struct timespec __user *, interval)
4725 struct task_struct *p;
4726 unsigned int time_slice;
4727 unsigned long flags;
4737 p = find_process_by_pid(pid);
4741 retval = security_task_getscheduler(p);
4745 rq = task_rq_lock(p, &flags);
4746 time_slice = p->sched_class->get_rr_interval(rq, p);
4747 task_rq_unlock(rq, p, &flags);
4750 jiffies_to_timespec(time_slice, &t);
4751 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4759 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4761 void sched_show_task(struct task_struct *p)
4763 unsigned long free = 0;
4766 state = p->state ? __ffs(p->state) + 1 : 0;
4767 printk(KERN_INFO "%-15.15s %c", p->comm,
4768 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4769 #if BITS_PER_LONG == 32
4770 if (state == TASK_RUNNING)
4771 printk(KERN_CONT " running ");
4773 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4775 if (state == TASK_RUNNING)
4776 printk(KERN_CONT " running task ");
4778 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4780 #ifdef CONFIG_DEBUG_STACK_USAGE
4781 free = stack_not_used(p);
4783 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4784 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4785 (unsigned long)task_thread_info(p)->flags);
4787 show_stack(p, NULL);
4790 void show_state_filter(unsigned long state_filter)
4792 struct task_struct *g, *p;
4794 #if BITS_PER_LONG == 32
4796 " task PC stack pid father\n");
4799 " task PC stack pid father\n");
4802 do_each_thread(g, p) {
4804 * reset the NMI-timeout, listing all files on a slow
4805 * console might take a lot of time:
4807 touch_nmi_watchdog();
4808 if (!state_filter || (p->state & state_filter))
4810 } while_each_thread(g, p);
4812 touch_all_softlockup_watchdogs();
4814 #ifdef CONFIG_SCHED_DEBUG
4815 sysrq_sched_debug_show();
4819 * Only show locks if all tasks are dumped:
4822 debug_show_all_locks();
4825 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4827 idle->sched_class = &idle_sched_class;
4831 * init_idle - set up an idle thread for a given CPU
4832 * @idle: task in question
4833 * @cpu: cpu the idle task belongs to
4835 * NOTE: this function does not set the idle thread's NEED_RESCHED
4836 * flag, to make booting more robust.
4838 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4840 struct rq *rq = cpu_rq(cpu);
4841 unsigned long flags;
4843 raw_spin_lock_irqsave(&rq->lock, flags);
4846 idle->state = TASK_RUNNING;
4847 idle->se.exec_start = sched_clock();
4849 do_set_cpus_allowed(idle, cpumask_of(cpu));
4851 * We're having a chicken and egg problem, even though we are
4852 * holding rq->lock, the cpu isn't yet set to this cpu so the
4853 * lockdep check in task_group() will fail.
4855 * Similar case to sched_fork(). / Alternatively we could
4856 * use task_rq_lock() here and obtain the other rq->lock.
4861 __set_task_cpu(idle, cpu);
4864 rq->curr = rq->idle = idle;
4865 #if defined(CONFIG_SMP)
4868 raw_spin_unlock_irqrestore(&rq->lock, flags);
4870 /* Set the preempt count _outside_ the spinlocks! */
4871 task_thread_info(idle)->preempt_count = 0;
4874 * The idle tasks have their own, simple scheduling class:
4876 idle->sched_class = &idle_sched_class;
4877 ftrace_graph_init_idle_task(idle, cpu);
4878 #if defined(CONFIG_SMP)
4879 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4884 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4886 if (p->sched_class && p->sched_class->set_cpus_allowed)
4887 p->sched_class->set_cpus_allowed(p, new_mask);
4889 cpumask_copy(&p->cpus_allowed, new_mask);
4890 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4894 * This is how migration works:
4896 * 1) we invoke migration_cpu_stop() on the target CPU using
4898 * 2) stopper starts to run (implicitly forcing the migrated thread
4900 * 3) it checks whether the migrated task is still in the wrong runqueue.
4901 * 4) if it's in the wrong runqueue then the migration thread removes
4902 * it and puts it into the right queue.
4903 * 5) stopper completes and stop_one_cpu() returns and the migration
4908 * Change a given task's CPU affinity. Migrate the thread to a
4909 * proper CPU and schedule it away if the CPU it's executing on
4910 * is removed from the allowed bitmask.
4912 * NOTE: the caller must have a valid reference to the task, the
4913 * task must not exit() & deallocate itself prematurely. The
4914 * call is not atomic; no spinlocks may be held.
4916 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4918 unsigned long flags;
4920 unsigned int dest_cpu;
4923 rq = task_rq_lock(p, &flags);
4925 if (cpumask_equal(&p->cpus_allowed, new_mask))
4928 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4933 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4938 do_set_cpus_allowed(p, new_mask);
4940 /* Can the task run on the task's current CPU? If so, we're done */
4941 if (cpumask_test_cpu(task_cpu(p), new_mask))
4944 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4946 struct migration_arg arg = { p, dest_cpu };
4947 /* Need help from migration thread: drop lock and wait. */
4948 task_rq_unlock(rq, p, &flags);
4949 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4950 tlb_migrate_finish(p->mm);
4954 task_rq_unlock(rq, p, &flags);
4958 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4961 * Move (not current) task off this cpu, onto dest cpu. We're doing
4962 * this because either it can't run here any more (set_cpus_allowed()
4963 * away from this CPU, or CPU going down), or because we're
4964 * attempting to rebalance this task on exec (sched_exec).
4966 * So we race with normal scheduler movements, but that's OK, as long
4967 * as the task is no longer on this CPU.
4969 * Returns non-zero if task was successfully migrated.
4971 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4973 struct rq *rq_dest, *rq_src;
4976 if (unlikely(!cpu_active(dest_cpu)))
4979 rq_src = cpu_rq(src_cpu);
4980 rq_dest = cpu_rq(dest_cpu);
4982 raw_spin_lock(&p->pi_lock);
4983 double_rq_lock(rq_src, rq_dest);
4984 /* Already moved. */
4985 if (task_cpu(p) != src_cpu)
4987 /* Affinity changed (again). */
4988 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4992 * If we're not on a rq, the next wake-up will ensure we're
4996 dequeue_task(rq_src, p, 0);
4997 set_task_cpu(p, dest_cpu);
4998 enqueue_task(rq_dest, p, 0);
4999 check_preempt_curr(rq_dest, p, 0);
5004 double_rq_unlock(rq_src, rq_dest);
5005 raw_spin_unlock(&p->pi_lock);
5010 * migration_cpu_stop - this will be executed by a highprio stopper thread
5011 * and performs thread migration by bumping thread off CPU then
5012 * 'pushing' onto another runqueue.
5014 static int migration_cpu_stop(void *data)
5016 struct migration_arg *arg = data;
5019 * The original target cpu might have gone down and we might
5020 * be on another cpu but it doesn't matter.
5022 local_irq_disable();
5023 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5028 #ifdef CONFIG_HOTPLUG_CPU
5031 * Ensures that the idle task is using init_mm right before its cpu goes
5034 void idle_task_exit(void)
5036 struct mm_struct *mm = current->active_mm;
5038 BUG_ON(cpu_online(smp_processor_id()));
5041 switch_mm(mm, &init_mm, current);
5046 * While a dead CPU has no uninterruptible tasks queued at this point,
5047 * it might still have a nonzero ->nr_uninterruptible counter, because
5048 * for performance reasons the counter is not stricly tracking tasks to
5049 * their home CPUs. So we just add the counter to another CPU's counter,
5050 * to keep the global sum constant after CPU-down:
5052 static void migrate_nr_uninterruptible(struct rq *rq_src)
5054 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5056 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5057 rq_src->nr_uninterruptible = 0;
5061 * remove the tasks which were accounted by rq from calc_load_tasks.
5063 static void calc_global_load_remove(struct rq *rq)
5065 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5066 rq->calc_load_active = 0;
5070 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5071 * try_to_wake_up()->select_task_rq().
5073 * Called with rq->lock held even though we'er in stop_machine() and
5074 * there's no concurrency possible, we hold the required locks anyway
5075 * because of lock validation efforts.
5077 static void migrate_tasks(unsigned int dead_cpu)
5079 struct rq *rq = cpu_rq(dead_cpu);
5080 struct task_struct *next, *stop = rq->stop;
5084 * Fudge the rq selection such that the below task selection loop
5085 * doesn't get stuck on the currently eligible stop task.
5087 * We're currently inside stop_machine() and the rq is either stuck
5088 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5089 * either way we should never end up calling schedule() until we're
5094 /* Ensure any throttled groups are reachable by pick_next_task */
5095 unthrottle_offline_cfs_rqs(rq);
5099 * There's this thread running, bail when that's the only
5102 if (rq->nr_running == 1)
5105 next = pick_next_task(rq);
5107 next->sched_class->put_prev_task(rq, next);
5109 /* Find suitable destination for @next, with force if needed. */
5110 dest_cpu = select_fallback_rq(dead_cpu, next);
5111 raw_spin_unlock(&rq->lock);
5113 __migrate_task(next, dead_cpu, dest_cpu);
5115 raw_spin_lock(&rq->lock);
5121 #endif /* CONFIG_HOTPLUG_CPU */
5123 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5125 static struct ctl_table sd_ctl_dir[] = {
5127 .procname = "sched_domain",
5133 static struct ctl_table sd_ctl_root[] = {
5135 .procname = "kernel",
5137 .child = sd_ctl_dir,
5142 static struct ctl_table *sd_alloc_ctl_entry(int n)
5144 struct ctl_table *entry =
5145 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5150 static void sd_free_ctl_entry(struct ctl_table **tablep)
5152 struct ctl_table *entry;
5155 * In the intermediate directories, both the child directory and
5156 * procname are dynamically allocated and could fail but the mode
5157 * will always be set. In the lowest directory the names are
5158 * static strings and all have proc handlers.
5160 for (entry = *tablep; entry->mode; entry++) {
5162 sd_free_ctl_entry(&entry->child);
5163 if (entry->proc_handler == NULL)
5164 kfree(entry->procname);
5172 set_table_entry(struct ctl_table *entry,
5173 const char *procname, void *data, int maxlen,
5174 umode_t mode, proc_handler *proc_handler)
5176 entry->procname = procname;
5178 entry->maxlen = maxlen;
5180 entry->proc_handler = proc_handler;
5183 static struct ctl_table *
5184 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5186 struct ctl_table *table = sd_alloc_ctl_entry(13);
5191 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5192 sizeof(long), 0644, proc_doulongvec_minmax);
5193 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5194 sizeof(long), 0644, proc_doulongvec_minmax);
5195 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5196 sizeof(int), 0644, proc_dointvec_minmax);
5197 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5198 sizeof(int), 0644, proc_dointvec_minmax);
5199 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5200 sizeof(int), 0644, proc_dointvec_minmax);
5201 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5202 sizeof(int), 0644, proc_dointvec_minmax);
5203 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5204 sizeof(int), 0644, proc_dointvec_minmax);
5205 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5206 sizeof(int), 0644, proc_dointvec_minmax);
5207 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5208 sizeof(int), 0644, proc_dointvec_minmax);
5209 set_table_entry(&table[9], "cache_nice_tries",
5210 &sd->cache_nice_tries,
5211 sizeof(int), 0644, proc_dointvec_minmax);
5212 set_table_entry(&table[10], "flags", &sd->flags,
5213 sizeof(int), 0644, proc_dointvec_minmax);
5214 set_table_entry(&table[11], "name", sd->name,
5215 CORENAME_MAX_SIZE, 0444, proc_dostring);
5216 /* &table[12] is terminator */
5221 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5223 struct ctl_table *entry, *table;
5224 struct sched_domain *sd;
5225 int domain_num = 0, i;
5228 for_each_domain(cpu, sd)
5230 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5235 for_each_domain(cpu, sd) {
5236 snprintf(buf, 32, "domain%d", i);
5237 entry->procname = kstrdup(buf, GFP_KERNEL);
5239 entry->child = sd_alloc_ctl_domain_table(sd);
5246 static struct ctl_table_header *sd_sysctl_header;
5247 static void register_sched_domain_sysctl(void)
5249 int i, cpu_num = num_possible_cpus();
5250 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5253 WARN_ON(sd_ctl_dir[0].child);
5254 sd_ctl_dir[0].child = entry;
5259 for_each_possible_cpu(i) {
5260 snprintf(buf, 32, "cpu%d", i);
5261 entry->procname = kstrdup(buf, GFP_KERNEL);
5263 entry->child = sd_alloc_ctl_cpu_table(i);
5267 WARN_ON(sd_sysctl_header);
5268 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5271 /* may be called multiple times per register */
5272 static void unregister_sched_domain_sysctl(void)
5274 if (sd_sysctl_header)
5275 unregister_sysctl_table(sd_sysctl_header);
5276 sd_sysctl_header = NULL;
5277 if (sd_ctl_dir[0].child)
5278 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5281 static void register_sched_domain_sysctl(void)
5284 static void unregister_sched_domain_sysctl(void)
5289 static void set_rq_online(struct rq *rq)
5292 const struct sched_class *class;
5294 cpumask_set_cpu(rq->cpu, rq->rd->online);
5297 for_each_class(class) {
5298 if (class->rq_online)
5299 class->rq_online(rq);
5304 static void set_rq_offline(struct rq *rq)
5307 const struct sched_class *class;
5309 for_each_class(class) {
5310 if (class->rq_offline)
5311 class->rq_offline(rq);
5314 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5320 * migration_call - callback that gets triggered when a CPU is added.
5321 * Here we can start up the necessary migration thread for the new CPU.
5323 static int __cpuinit
5324 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5326 int cpu = (long)hcpu;
5327 unsigned long flags;
5328 struct rq *rq = cpu_rq(cpu);
5330 switch (action & ~CPU_TASKS_FROZEN) {
5332 case CPU_UP_PREPARE:
5333 rq->calc_load_update = calc_load_update;
5337 /* Update our root-domain */
5338 raw_spin_lock_irqsave(&rq->lock, flags);
5340 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5344 raw_spin_unlock_irqrestore(&rq->lock, flags);
5347 #ifdef CONFIG_HOTPLUG_CPU
5349 sched_ttwu_pending();
5350 /* Update our root-domain */
5351 raw_spin_lock_irqsave(&rq->lock, flags);
5353 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5357 BUG_ON(rq->nr_running != 1); /* the migration thread */
5358 raw_spin_unlock_irqrestore(&rq->lock, flags);
5360 migrate_nr_uninterruptible(rq);
5361 calc_global_load_remove(rq);
5366 update_max_interval();
5372 * Register at high priority so that task migration (migrate_all_tasks)
5373 * happens before everything else. This has to be lower priority than
5374 * the notifier in the perf_event subsystem, though.
5376 static struct notifier_block __cpuinitdata migration_notifier = {
5377 .notifier_call = migration_call,
5378 .priority = CPU_PRI_MIGRATION,
5381 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5382 unsigned long action, void *hcpu)
5384 switch (action & ~CPU_TASKS_FROZEN) {
5386 case CPU_DOWN_FAILED:
5387 set_cpu_active((long)hcpu, true);
5394 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5395 unsigned long action, void *hcpu)
5397 switch (action & ~CPU_TASKS_FROZEN) {
5398 case CPU_DOWN_PREPARE:
5399 set_cpu_active((long)hcpu, false);
5406 static int __init migration_init(void)
5408 void *cpu = (void *)(long)smp_processor_id();
5411 /* Initialize migration for the boot CPU */
5412 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5413 BUG_ON(err == NOTIFY_BAD);
5414 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5415 register_cpu_notifier(&migration_notifier);
5417 /* Register cpu active notifiers */
5418 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5419 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5423 early_initcall(migration_init);
5428 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5430 #ifdef CONFIG_SCHED_DEBUG
5432 static __read_mostly int sched_domain_debug_enabled;
5434 static int __init sched_domain_debug_setup(char *str)
5436 sched_domain_debug_enabled = 1;
5440 early_param("sched_debug", sched_domain_debug_setup);
5442 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5443 struct cpumask *groupmask)
5445 struct sched_group *group = sd->groups;
5448 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5449 cpumask_clear(groupmask);
5451 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5453 if (!(sd->flags & SD_LOAD_BALANCE)) {
5454 printk("does not load-balance\n");
5456 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5461 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5463 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5464 printk(KERN_ERR "ERROR: domain->span does not contain "
5467 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5468 printk(KERN_ERR "ERROR: domain->groups does not contain"
5472 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5476 printk(KERN_ERR "ERROR: group is NULL\n");
5480 if (!group->sgp->power) {
5481 printk(KERN_CONT "\n");
5482 printk(KERN_ERR "ERROR: domain->cpu_power not "
5487 if (!cpumask_weight(sched_group_cpus(group))) {
5488 printk(KERN_CONT "\n");
5489 printk(KERN_ERR "ERROR: empty group\n");
5493 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5494 printk(KERN_CONT "\n");
5495 printk(KERN_ERR "ERROR: repeated CPUs\n");
5499 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5501 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5503 printk(KERN_CONT " %s", str);
5504 if (group->sgp->power != SCHED_POWER_SCALE) {
5505 printk(KERN_CONT " (cpu_power = %d)",
5509 group = group->next;
5510 } while (group != sd->groups);
5511 printk(KERN_CONT "\n");
5513 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5514 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5517 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5518 printk(KERN_ERR "ERROR: parent span is not a superset "
5519 "of domain->span\n");
5523 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5527 if (!sched_domain_debug_enabled)
5531 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5535 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5538 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5546 #else /* !CONFIG_SCHED_DEBUG */
5547 # define sched_domain_debug(sd, cpu) do { } while (0)
5548 #endif /* CONFIG_SCHED_DEBUG */
5550 static int sd_degenerate(struct sched_domain *sd)
5552 if (cpumask_weight(sched_domain_span(sd)) == 1)
5555 /* Following flags need at least 2 groups */
5556 if (sd->flags & (SD_LOAD_BALANCE |
5557 SD_BALANCE_NEWIDLE |
5561 SD_SHARE_PKG_RESOURCES)) {
5562 if (sd->groups != sd->groups->next)
5566 /* Following flags don't use groups */
5567 if (sd->flags & (SD_WAKE_AFFINE))
5574 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5576 unsigned long cflags = sd->flags, pflags = parent->flags;
5578 if (sd_degenerate(parent))
5581 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5584 /* Flags needing groups don't count if only 1 group in parent */
5585 if (parent->groups == parent->groups->next) {
5586 pflags &= ~(SD_LOAD_BALANCE |
5587 SD_BALANCE_NEWIDLE |
5591 SD_SHARE_PKG_RESOURCES);
5592 if (nr_node_ids == 1)
5593 pflags &= ~SD_SERIALIZE;
5595 if (~cflags & pflags)
5601 static void free_rootdomain(struct rcu_head *rcu)
5603 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5605 cpupri_cleanup(&rd->cpupri);
5606 free_cpumask_var(rd->rto_mask);
5607 free_cpumask_var(rd->online);
5608 free_cpumask_var(rd->span);
5612 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5614 struct root_domain *old_rd = NULL;
5615 unsigned long flags;
5617 raw_spin_lock_irqsave(&rq->lock, flags);
5622 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5625 cpumask_clear_cpu(rq->cpu, old_rd->span);
5628 * If we dont want to free the old_rt yet then
5629 * set old_rd to NULL to skip the freeing later
5632 if (!atomic_dec_and_test(&old_rd->refcount))
5636 atomic_inc(&rd->refcount);
5639 cpumask_set_cpu(rq->cpu, rd->span);
5640 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5643 raw_spin_unlock_irqrestore(&rq->lock, flags);
5646 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5649 static int init_rootdomain(struct root_domain *rd)
5651 memset(rd, 0, sizeof(*rd));
5653 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5655 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5657 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5660 if (cpupri_init(&rd->cpupri) != 0)
5665 free_cpumask_var(rd->rto_mask);
5667 free_cpumask_var(rd->online);
5669 free_cpumask_var(rd->span);
5675 * By default the system creates a single root-domain with all cpus as
5676 * members (mimicking the global state we have today).
5678 struct root_domain def_root_domain;
5680 static void init_defrootdomain(void)
5682 init_rootdomain(&def_root_domain);
5684 atomic_set(&def_root_domain.refcount, 1);
5687 static struct root_domain *alloc_rootdomain(void)
5689 struct root_domain *rd;
5691 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5695 if (init_rootdomain(rd) != 0) {
5703 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5705 struct sched_group *tmp, *first;
5714 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5719 } while (sg != first);
5722 static void free_sched_domain(struct rcu_head *rcu)
5724 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5727 * If its an overlapping domain it has private groups, iterate and
5730 if (sd->flags & SD_OVERLAP) {
5731 free_sched_groups(sd->groups, 1);
5732 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5733 kfree(sd->groups->sgp);
5739 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5741 call_rcu(&sd->rcu, free_sched_domain);
5744 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5746 for (; sd; sd = sd->parent)
5747 destroy_sched_domain(sd, cpu);
5751 * Keep a special pointer to the highest sched_domain that has
5752 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5753 * allows us to avoid some pointer chasing select_idle_sibling().
5755 * Also keep a unique ID per domain (we use the first cpu number in
5756 * the cpumask of the domain), this allows us to quickly tell if
5757 * two cpus are in the same cache domain, see ttwu_share_cache().
5759 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5760 DEFINE_PER_CPU(int, sd_llc_id);
5762 static void update_top_cache_domain(int cpu)
5764 struct sched_domain *sd;
5767 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5769 id = cpumask_first(sched_domain_span(sd));
5771 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5772 per_cpu(sd_llc_id, cpu) = id;
5776 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5777 * hold the hotplug lock.
5780 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5782 struct rq *rq = cpu_rq(cpu);
5783 struct sched_domain *tmp;
5785 /* Remove the sched domains which do not contribute to scheduling. */
5786 for (tmp = sd; tmp; ) {
5787 struct sched_domain *parent = tmp->parent;
5791 if (sd_parent_degenerate(tmp, parent)) {
5792 tmp->parent = parent->parent;
5794 parent->parent->child = tmp;
5795 destroy_sched_domain(parent, cpu);
5800 if (sd && sd_degenerate(sd)) {
5803 destroy_sched_domain(tmp, cpu);
5808 sched_domain_debug(sd, cpu);
5810 rq_attach_root(rq, rd);
5812 rcu_assign_pointer(rq->sd, sd);
5813 destroy_sched_domains(tmp, cpu);
5815 update_top_cache_domain(cpu);
5818 /* cpus with isolated domains */
5819 static cpumask_var_t cpu_isolated_map;
5821 /* Setup the mask of cpus configured for isolated domains */
5822 static int __init isolated_cpu_setup(char *str)
5824 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5825 cpulist_parse(str, cpu_isolated_map);
5829 __setup("isolcpus=", isolated_cpu_setup);
5834 * find_next_best_node - find the next node to include in a sched_domain
5835 * @node: node whose sched_domain we're building
5836 * @used_nodes: nodes already in the sched_domain
5838 * Find the next node to include in a given scheduling domain. Simply
5839 * finds the closest node not already in the @used_nodes map.
5841 * Should use nodemask_t.
5843 static int find_next_best_node(int node, nodemask_t *used_nodes)
5845 int i, n, val, min_val, best_node = -1;
5849 for (i = 0; i < nr_node_ids; i++) {
5850 /* Start at @node */
5851 n = (node + i) % nr_node_ids;
5853 if (!nr_cpus_node(n))
5856 /* Skip already used nodes */
5857 if (node_isset(n, *used_nodes))
5860 /* Simple min distance search */
5861 val = node_distance(node, n);
5863 if (val < min_val) {
5869 if (best_node != -1)
5870 node_set(best_node, *used_nodes);
5875 * sched_domain_node_span - get a cpumask for a node's sched_domain
5876 * @node: node whose cpumask we're constructing
5877 * @span: resulting cpumask
5879 * Given a node, construct a good cpumask for its sched_domain to span. It
5880 * should be one that prevents unnecessary balancing, but also spreads tasks
5883 static void sched_domain_node_span(int node, struct cpumask *span)
5885 nodemask_t used_nodes;
5888 cpumask_clear(span);
5889 nodes_clear(used_nodes);
5891 cpumask_or(span, span, cpumask_of_node(node));
5892 node_set(node, used_nodes);
5894 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5895 int next_node = find_next_best_node(node, &used_nodes);
5898 cpumask_or(span, span, cpumask_of_node(next_node));
5902 static const struct cpumask *cpu_node_mask(int cpu)
5904 lockdep_assert_held(&sched_domains_mutex);
5906 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5908 return sched_domains_tmpmask;
5911 static const struct cpumask *cpu_allnodes_mask(int cpu)
5913 return cpu_possible_mask;
5915 #endif /* CONFIG_NUMA */
5917 static const struct cpumask *cpu_cpu_mask(int cpu)
5919 return cpumask_of_node(cpu_to_node(cpu));
5922 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5925 struct sched_domain **__percpu sd;
5926 struct sched_group **__percpu sg;
5927 struct sched_group_power **__percpu sgp;
5931 struct sched_domain ** __percpu sd;
5932 struct root_domain *rd;
5942 struct sched_domain_topology_level;
5944 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5945 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5947 #define SDTL_OVERLAP 0x01
5949 struct sched_domain_topology_level {
5950 sched_domain_init_f init;
5951 sched_domain_mask_f mask;
5953 struct sd_data data;
5957 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5959 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5960 const struct cpumask *span = sched_domain_span(sd);
5961 struct cpumask *covered = sched_domains_tmpmask;
5962 struct sd_data *sdd = sd->private;
5963 struct sched_domain *child;
5966 cpumask_clear(covered);
5968 for_each_cpu(i, span) {
5969 struct cpumask *sg_span;
5971 if (cpumask_test_cpu(i, covered))
5974 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5975 GFP_KERNEL, cpu_to_node(cpu));
5980 sg_span = sched_group_cpus(sg);
5982 child = *per_cpu_ptr(sdd->sd, i);
5984 child = child->child;
5985 cpumask_copy(sg_span, sched_domain_span(child));
5987 cpumask_set_cpu(i, sg_span);
5989 cpumask_or(covered, covered, sg_span);
5991 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5992 atomic_inc(&sg->sgp->ref);
5994 if (cpumask_test_cpu(cpu, sg_span))
6004 sd->groups = groups;
6009 free_sched_groups(first, 0);
6014 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6016 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6017 struct sched_domain *child = sd->child;
6020 cpu = cpumask_first(sched_domain_span(child));
6023 *sg = *per_cpu_ptr(sdd->sg, cpu);
6024 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6025 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6032 * build_sched_groups will build a circular linked list of the groups
6033 * covered by the given span, and will set each group's ->cpumask correctly,
6034 * and ->cpu_power to 0.
6036 * Assumes the sched_domain tree is fully constructed
6039 build_sched_groups(struct sched_domain *sd, int cpu)
6041 struct sched_group *first = NULL, *last = NULL;
6042 struct sd_data *sdd = sd->private;
6043 const struct cpumask *span = sched_domain_span(sd);
6044 struct cpumask *covered;
6047 get_group(cpu, sdd, &sd->groups);
6048 atomic_inc(&sd->groups->ref);
6050 if (cpu != cpumask_first(sched_domain_span(sd)))
6053 lockdep_assert_held(&sched_domains_mutex);
6054 covered = sched_domains_tmpmask;
6056 cpumask_clear(covered);
6058 for_each_cpu(i, span) {
6059 struct sched_group *sg;
6060 int group = get_group(i, sdd, &sg);
6063 if (cpumask_test_cpu(i, covered))
6066 cpumask_clear(sched_group_cpus(sg));
6069 for_each_cpu(j, span) {
6070 if (get_group(j, sdd, NULL) != group)
6073 cpumask_set_cpu(j, covered);
6074 cpumask_set_cpu(j, sched_group_cpus(sg));
6089 * Initialize sched groups cpu_power.
6091 * cpu_power indicates the capacity of sched group, which is used while
6092 * distributing the load between different sched groups in a sched domain.
6093 * Typically cpu_power for all the groups in a sched domain will be same unless
6094 * there are asymmetries in the topology. If there are asymmetries, group
6095 * having more cpu_power will pickup more load compared to the group having
6098 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6100 struct sched_group *sg = sd->groups;
6102 WARN_ON(!sd || !sg);
6105 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6107 } while (sg != sd->groups);
6109 if (cpu != group_first_cpu(sg))
6112 update_group_power(sd, cpu);
6113 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6116 int __weak arch_sd_sibling_asym_packing(void)
6118 return 0*SD_ASYM_PACKING;
6122 * Initializers for schedule domains
6123 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6126 #ifdef CONFIG_SCHED_DEBUG
6127 # define SD_INIT_NAME(sd, type) sd->name = #type
6129 # define SD_INIT_NAME(sd, type) do { } while (0)
6132 #define SD_INIT_FUNC(type) \
6133 static noinline struct sched_domain * \
6134 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6136 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6137 *sd = SD_##type##_INIT; \
6138 SD_INIT_NAME(sd, type); \
6139 sd->private = &tl->data; \
6145 SD_INIT_FUNC(ALLNODES)
6148 #ifdef CONFIG_SCHED_SMT
6149 SD_INIT_FUNC(SIBLING)
6151 #ifdef CONFIG_SCHED_MC
6154 #ifdef CONFIG_SCHED_BOOK
6158 static int default_relax_domain_level = -1;
6159 int sched_domain_level_max;
6161 static int __init setup_relax_domain_level(char *str)
6165 val = simple_strtoul(str, NULL, 0);
6166 if (val < sched_domain_level_max)
6167 default_relax_domain_level = val;
6171 __setup("relax_domain_level=", setup_relax_domain_level);
6173 static void set_domain_attribute(struct sched_domain *sd,
6174 struct sched_domain_attr *attr)
6178 if (!attr || attr->relax_domain_level < 0) {
6179 if (default_relax_domain_level < 0)
6182 request = default_relax_domain_level;
6184 request = attr->relax_domain_level;
6185 if (request < sd->level) {
6186 /* turn off idle balance on this domain */
6187 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6189 /* turn on idle balance on this domain */
6190 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6194 static void __sdt_free(const struct cpumask *cpu_map);
6195 static int __sdt_alloc(const struct cpumask *cpu_map);
6197 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6198 const struct cpumask *cpu_map)
6202 if (!atomic_read(&d->rd->refcount))
6203 free_rootdomain(&d->rd->rcu); /* fall through */
6205 free_percpu(d->sd); /* fall through */
6207 __sdt_free(cpu_map); /* fall through */
6213 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6214 const struct cpumask *cpu_map)
6216 memset(d, 0, sizeof(*d));
6218 if (__sdt_alloc(cpu_map))
6219 return sa_sd_storage;
6220 d->sd = alloc_percpu(struct sched_domain *);
6222 return sa_sd_storage;
6223 d->rd = alloc_rootdomain();
6226 return sa_rootdomain;
6230 * NULL the sd_data elements we've used to build the sched_domain and
6231 * sched_group structure so that the subsequent __free_domain_allocs()
6232 * will not free the data we're using.
6234 static void claim_allocations(int cpu, struct sched_domain *sd)
6236 struct sd_data *sdd = sd->private;
6238 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6239 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6241 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6242 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6244 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6245 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6248 #ifdef CONFIG_SCHED_SMT
6249 static const struct cpumask *cpu_smt_mask(int cpu)
6251 return topology_thread_cpumask(cpu);
6256 * Topology list, bottom-up.
6258 static struct sched_domain_topology_level default_topology[] = {
6259 #ifdef CONFIG_SCHED_SMT
6260 { sd_init_SIBLING, cpu_smt_mask, },
6262 #ifdef CONFIG_SCHED_MC
6263 { sd_init_MC, cpu_coregroup_mask, },
6265 #ifdef CONFIG_SCHED_BOOK
6266 { sd_init_BOOK, cpu_book_mask, },
6268 { sd_init_CPU, cpu_cpu_mask, },
6270 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6271 { sd_init_ALLNODES, cpu_allnodes_mask, },
6276 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6278 static int __sdt_alloc(const struct cpumask *cpu_map)
6280 struct sched_domain_topology_level *tl;
6283 for (tl = sched_domain_topology; tl->init; tl++) {
6284 struct sd_data *sdd = &tl->data;
6286 sdd->sd = alloc_percpu(struct sched_domain *);
6290 sdd->sg = alloc_percpu(struct sched_group *);
6294 sdd->sgp = alloc_percpu(struct sched_group_power *);
6298 for_each_cpu(j, cpu_map) {
6299 struct sched_domain *sd;
6300 struct sched_group *sg;
6301 struct sched_group_power *sgp;
6303 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6304 GFP_KERNEL, cpu_to_node(j));
6308 *per_cpu_ptr(sdd->sd, j) = sd;
6310 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6311 GFP_KERNEL, cpu_to_node(j));
6315 *per_cpu_ptr(sdd->sg, j) = sg;
6317 sgp = kzalloc_node(sizeof(struct sched_group_power),
6318 GFP_KERNEL, cpu_to_node(j));
6322 *per_cpu_ptr(sdd->sgp, j) = sgp;
6329 static void __sdt_free(const struct cpumask *cpu_map)
6331 struct sched_domain_topology_level *tl;
6334 for (tl = sched_domain_topology; tl->init; tl++) {
6335 struct sd_data *sdd = &tl->data;
6337 for_each_cpu(j, cpu_map) {
6338 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6339 if (sd && (sd->flags & SD_OVERLAP))
6340 free_sched_groups(sd->groups, 0);
6341 kfree(*per_cpu_ptr(sdd->sd, j));
6342 kfree(*per_cpu_ptr(sdd->sg, j));
6343 kfree(*per_cpu_ptr(sdd->sgp, j));
6345 free_percpu(sdd->sd);
6346 free_percpu(sdd->sg);
6347 free_percpu(sdd->sgp);
6351 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6352 struct s_data *d, const struct cpumask *cpu_map,
6353 struct sched_domain_attr *attr, struct sched_domain *child,
6356 struct sched_domain *sd = tl->init(tl, cpu);
6360 set_domain_attribute(sd, attr);
6361 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6363 sd->level = child->level + 1;
6364 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6373 * Build sched domains for a given set of cpus and attach the sched domains
6374 * to the individual cpus
6376 static int build_sched_domains(const struct cpumask *cpu_map,
6377 struct sched_domain_attr *attr)
6379 enum s_alloc alloc_state = sa_none;
6380 struct sched_domain *sd;
6382 int i, ret = -ENOMEM;
6384 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6385 if (alloc_state != sa_rootdomain)
6388 /* Set up domains for cpus specified by the cpu_map. */
6389 for_each_cpu(i, cpu_map) {
6390 struct sched_domain_topology_level *tl;
6393 for (tl = sched_domain_topology; tl->init; tl++) {
6394 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6395 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6396 sd->flags |= SD_OVERLAP;
6397 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6404 *per_cpu_ptr(d.sd, i) = sd;
6407 /* Build the groups for the domains */
6408 for_each_cpu(i, cpu_map) {
6409 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6410 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6411 if (sd->flags & SD_OVERLAP) {
6412 if (build_overlap_sched_groups(sd, i))
6415 if (build_sched_groups(sd, i))
6421 /* Calculate CPU power for physical packages and nodes */
6422 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6423 if (!cpumask_test_cpu(i, cpu_map))
6426 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6427 claim_allocations(i, sd);
6428 init_sched_groups_power(i, sd);
6432 /* Attach the domains */
6434 for_each_cpu(i, cpu_map) {
6435 sd = *per_cpu_ptr(d.sd, i);
6436 cpu_attach_domain(sd, d.rd, i);
6442 __free_domain_allocs(&d, alloc_state, cpu_map);
6446 static cpumask_var_t *doms_cur; /* current sched domains */
6447 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6448 static struct sched_domain_attr *dattr_cur;
6449 /* attribues of custom domains in 'doms_cur' */
6452 * Special case: If a kmalloc of a doms_cur partition (array of
6453 * cpumask) fails, then fallback to a single sched domain,
6454 * as determined by the single cpumask fallback_doms.
6456 static cpumask_var_t fallback_doms;
6459 * arch_update_cpu_topology lets virtualized architectures update the
6460 * cpu core maps. It is supposed to return 1 if the topology changed
6461 * or 0 if it stayed the same.
6463 int __attribute__((weak)) arch_update_cpu_topology(void)
6468 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6471 cpumask_var_t *doms;
6473 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6476 for (i = 0; i < ndoms; i++) {
6477 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6478 free_sched_domains(doms, i);
6485 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6488 for (i = 0; i < ndoms; i++)
6489 free_cpumask_var(doms[i]);
6494 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6495 * For now this just excludes isolated cpus, but could be used to
6496 * exclude other special cases in the future.
6498 static int init_sched_domains(const struct cpumask *cpu_map)
6502 arch_update_cpu_topology();
6504 doms_cur = alloc_sched_domains(ndoms_cur);
6506 doms_cur = &fallback_doms;
6507 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6509 err = build_sched_domains(doms_cur[0], NULL);
6510 register_sched_domain_sysctl();
6516 * Detach sched domains from a group of cpus specified in cpu_map
6517 * These cpus will now be attached to the NULL domain
6519 static void detach_destroy_domains(const struct cpumask *cpu_map)
6524 for_each_cpu(i, cpu_map)
6525 cpu_attach_domain(NULL, &def_root_domain, i);
6529 /* handle null as "default" */
6530 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6531 struct sched_domain_attr *new, int idx_new)
6533 struct sched_domain_attr tmp;
6540 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6541 new ? (new + idx_new) : &tmp,
6542 sizeof(struct sched_domain_attr));
6546 * Partition sched domains as specified by the 'ndoms_new'
6547 * cpumasks in the array doms_new[] of cpumasks. This compares
6548 * doms_new[] to the current sched domain partitioning, doms_cur[].
6549 * It destroys each deleted domain and builds each new domain.
6551 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6552 * The masks don't intersect (don't overlap.) We should setup one
6553 * sched domain for each mask. CPUs not in any of the cpumasks will
6554 * not be load balanced. If the same cpumask appears both in the
6555 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6558 * The passed in 'doms_new' should be allocated using
6559 * alloc_sched_domains. This routine takes ownership of it and will
6560 * free_sched_domains it when done with it. If the caller failed the
6561 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6562 * and partition_sched_domains() will fallback to the single partition
6563 * 'fallback_doms', it also forces the domains to be rebuilt.
6565 * If doms_new == NULL it will be replaced with cpu_online_mask.
6566 * ndoms_new == 0 is a special case for destroying existing domains,
6567 * and it will not create the default domain.
6569 * Call with hotplug lock held
6571 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6572 struct sched_domain_attr *dattr_new)
6577 mutex_lock(&sched_domains_mutex);
6579 /* always unregister in case we don't destroy any domains */
6580 unregister_sched_domain_sysctl();
6582 /* Let architecture update cpu core mappings. */
6583 new_topology = arch_update_cpu_topology();
6585 n = doms_new ? ndoms_new : 0;
6587 /* Destroy deleted domains */
6588 for (i = 0; i < ndoms_cur; i++) {
6589 for (j = 0; j < n && !new_topology; j++) {
6590 if (cpumask_equal(doms_cur[i], doms_new[j])
6591 && dattrs_equal(dattr_cur, i, dattr_new, j))
6594 /* no match - a current sched domain not in new doms_new[] */
6595 detach_destroy_domains(doms_cur[i]);
6600 if (doms_new == NULL) {
6602 doms_new = &fallback_doms;
6603 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6604 WARN_ON_ONCE(dattr_new);
6607 /* Build new domains */
6608 for (i = 0; i < ndoms_new; i++) {
6609 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6610 if (cpumask_equal(doms_new[i], doms_cur[j])
6611 && dattrs_equal(dattr_new, i, dattr_cur, j))
6614 /* no match - add a new doms_new */
6615 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6620 /* Remember the new sched domains */
6621 if (doms_cur != &fallback_doms)
6622 free_sched_domains(doms_cur, ndoms_cur);
6623 kfree(dattr_cur); /* kfree(NULL) is safe */
6624 doms_cur = doms_new;
6625 dattr_cur = dattr_new;
6626 ndoms_cur = ndoms_new;
6628 register_sched_domain_sysctl();
6630 mutex_unlock(&sched_domains_mutex);
6633 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6634 static void reinit_sched_domains(void)
6638 /* Destroy domains first to force the rebuild */
6639 partition_sched_domains(0, NULL, NULL);
6641 rebuild_sched_domains();
6645 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6647 unsigned int level = 0;
6649 if (sscanf(buf, "%u", &level) != 1)
6653 * level is always be positive so don't check for
6654 * level < POWERSAVINGS_BALANCE_NONE which is 0
6655 * What happens on 0 or 1 byte write,
6656 * need to check for count as well?
6659 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6663 sched_smt_power_savings = level;
6665 sched_mc_power_savings = level;
6667 reinit_sched_domains();
6672 #ifdef CONFIG_SCHED_MC
6673 static ssize_t sched_mc_power_savings_show(struct device *dev,
6674 struct device_attribute *attr,
6677 return sprintf(buf, "%u\n", sched_mc_power_savings);
6679 static ssize_t sched_mc_power_savings_store(struct device *dev,
6680 struct device_attribute *attr,
6681 const char *buf, size_t count)
6683 return sched_power_savings_store(buf, count, 0);
6685 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6686 sched_mc_power_savings_show,
6687 sched_mc_power_savings_store);
6690 #ifdef CONFIG_SCHED_SMT
6691 static ssize_t sched_smt_power_savings_show(struct device *dev,
6692 struct device_attribute *attr,
6695 return sprintf(buf, "%u\n", sched_smt_power_savings);
6697 static ssize_t sched_smt_power_savings_store(struct device *dev,
6698 struct device_attribute *attr,
6699 const char *buf, size_t count)
6701 return sched_power_savings_store(buf, count, 1);
6703 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6704 sched_smt_power_savings_show,
6705 sched_smt_power_savings_store);
6708 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6712 #ifdef CONFIG_SCHED_SMT
6714 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6716 #ifdef CONFIG_SCHED_MC
6717 if (!err && mc_capable())
6718 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6722 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6725 * Update cpusets according to cpu_active mask. If cpusets are
6726 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6727 * around partition_sched_domains().
6729 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6732 switch (action & ~CPU_TASKS_FROZEN) {
6734 case CPU_DOWN_FAILED:
6735 cpuset_update_active_cpus();
6742 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6745 switch (action & ~CPU_TASKS_FROZEN) {
6746 case CPU_DOWN_PREPARE:
6747 cpuset_update_active_cpus();
6754 void __init sched_init_smp(void)
6756 cpumask_var_t non_isolated_cpus;
6758 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6759 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6762 mutex_lock(&sched_domains_mutex);
6763 init_sched_domains(cpu_active_mask);
6764 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6765 if (cpumask_empty(non_isolated_cpus))
6766 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6767 mutex_unlock(&sched_domains_mutex);
6770 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6771 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6773 /* RT runtime code needs to handle some hotplug events */
6774 hotcpu_notifier(update_runtime, 0);
6778 /* Move init over to a non-isolated CPU */
6779 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6781 sched_init_granularity();
6782 free_cpumask_var(non_isolated_cpus);
6784 init_sched_rt_class();
6787 void __init sched_init_smp(void)
6789 sched_init_granularity();
6791 #endif /* CONFIG_SMP */
6793 const_debug unsigned int sysctl_timer_migration = 1;
6795 int in_sched_functions(unsigned long addr)
6797 return in_lock_functions(addr) ||
6798 (addr >= (unsigned long)__sched_text_start
6799 && addr < (unsigned long)__sched_text_end);
6802 #ifdef CONFIG_CGROUP_SCHED
6803 struct task_group root_task_group;
6806 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6808 void __init sched_init(void)
6811 unsigned long alloc_size = 0, ptr;
6813 #ifdef CONFIG_FAIR_GROUP_SCHED
6814 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6816 #ifdef CONFIG_RT_GROUP_SCHED
6817 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6819 #ifdef CONFIG_CPUMASK_OFFSTACK
6820 alloc_size += num_possible_cpus() * cpumask_size();
6823 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6825 #ifdef CONFIG_FAIR_GROUP_SCHED
6826 root_task_group.se = (struct sched_entity **)ptr;
6827 ptr += nr_cpu_ids * sizeof(void **);
6829 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6830 ptr += nr_cpu_ids * sizeof(void **);
6832 #endif /* CONFIG_FAIR_GROUP_SCHED */
6833 #ifdef CONFIG_RT_GROUP_SCHED
6834 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6835 ptr += nr_cpu_ids * sizeof(void **);
6837 root_task_group.rt_rq = (struct rt_rq **)ptr;
6838 ptr += nr_cpu_ids * sizeof(void **);
6840 #endif /* CONFIG_RT_GROUP_SCHED */
6841 #ifdef CONFIG_CPUMASK_OFFSTACK
6842 for_each_possible_cpu(i) {
6843 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6844 ptr += cpumask_size();
6846 #endif /* CONFIG_CPUMASK_OFFSTACK */
6850 init_defrootdomain();
6853 init_rt_bandwidth(&def_rt_bandwidth,
6854 global_rt_period(), global_rt_runtime());
6856 #ifdef CONFIG_RT_GROUP_SCHED
6857 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6858 global_rt_period(), global_rt_runtime());
6859 #endif /* CONFIG_RT_GROUP_SCHED */
6861 #ifdef CONFIG_CGROUP_SCHED
6862 list_add(&root_task_group.list, &task_groups);
6863 INIT_LIST_HEAD(&root_task_group.children);
6864 INIT_LIST_HEAD(&root_task_group.siblings);
6865 autogroup_init(&init_task);
6867 #endif /* CONFIG_CGROUP_SCHED */
6869 #ifdef CONFIG_CGROUP_CPUACCT
6870 root_cpuacct.cpustat = &kernel_cpustat;
6871 root_cpuacct.cpuusage = alloc_percpu(u64);
6872 /* Too early, not expected to fail */
6873 BUG_ON(!root_cpuacct.cpuusage);
6875 for_each_possible_cpu(i) {
6879 raw_spin_lock_init(&rq->lock);
6881 rq->calc_load_active = 0;
6882 rq->calc_load_update = jiffies + LOAD_FREQ;
6883 init_cfs_rq(&rq->cfs);
6884 init_rt_rq(&rq->rt, rq);
6885 #ifdef CONFIG_FAIR_GROUP_SCHED
6886 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6887 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6889 * How much cpu bandwidth does root_task_group get?
6891 * In case of task-groups formed thr' the cgroup filesystem, it
6892 * gets 100% of the cpu resources in the system. This overall
6893 * system cpu resource is divided among the tasks of
6894 * root_task_group and its child task-groups in a fair manner,
6895 * based on each entity's (task or task-group's) weight
6896 * (se->load.weight).
6898 * In other words, if root_task_group has 10 tasks of weight
6899 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6900 * then A0's share of the cpu resource is:
6902 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6904 * We achieve this by letting root_task_group's tasks sit
6905 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6907 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6908 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6909 #endif /* CONFIG_FAIR_GROUP_SCHED */
6911 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6912 #ifdef CONFIG_RT_GROUP_SCHED
6913 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6914 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6917 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6918 rq->cpu_load[j] = 0;
6920 rq->last_load_update_tick = jiffies;
6925 rq->cpu_power = SCHED_POWER_SCALE;
6926 rq->post_schedule = 0;
6927 rq->active_balance = 0;
6928 rq->next_balance = jiffies;
6933 rq->avg_idle = 2*sysctl_sched_migration_cost;
6934 rq_attach_root(rq, &def_root_domain);
6940 atomic_set(&rq->nr_iowait, 0);
6943 set_load_weight(&init_task);
6945 #ifdef CONFIG_PREEMPT_NOTIFIERS
6946 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6949 #ifdef CONFIG_RT_MUTEXES
6950 plist_head_init(&init_task.pi_waiters);
6954 * The boot idle thread does lazy MMU switching as well:
6956 atomic_inc(&init_mm.mm_count);
6957 enter_lazy_tlb(&init_mm, current);
6960 * Make us the idle thread. Technically, schedule() should not be
6961 * called from this thread, however somewhere below it might be,
6962 * but because we are the idle thread, we just pick up running again
6963 * when this runqueue becomes "idle".
6965 init_idle(current, smp_processor_id());
6967 calc_load_update = jiffies + LOAD_FREQ;
6970 * During early bootup we pretend to be a normal task:
6972 current->sched_class = &fair_sched_class;
6975 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6976 /* May be allocated at isolcpus cmdline parse time */
6977 if (cpu_isolated_map == NULL)
6978 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6980 init_sched_fair_class();
6982 scheduler_running = 1;
6985 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6986 static inline int preempt_count_equals(int preempt_offset)
6988 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6990 return (nested == preempt_offset);
6993 void __might_sleep(const char *file, int line, int preempt_offset)
6995 static unsigned long prev_jiffy; /* ratelimiting */
6997 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6998 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6999 system_state != SYSTEM_RUNNING || oops_in_progress)
7001 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7003 prev_jiffy = jiffies;
7006 "BUG: sleeping function called from invalid context at %s:%d\n",
7009 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7010 in_atomic(), irqs_disabled(),
7011 current->pid, current->comm);
7013 debug_show_held_locks(current);
7014 if (irqs_disabled())
7015 print_irqtrace_events(current);
7018 EXPORT_SYMBOL(__might_sleep);
7021 #ifdef CONFIG_MAGIC_SYSRQ
7022 static void normalize_task(struct rq *rq, struct task_struct *p)
7024 const struct sched_class *prev_class = p->sched_class;
7025 int old_prio = p->prio;
7030 dequeue_task(rq, p, 0);
7031 __setscheduler(rq, p, SCHED_NORMAL, 0);
7033 enqueue_task(rq, p, 0);
7034 resched_task(rq->curr);
7037 check_class_changed(rq, p, prev_class, old_prio);
7040 void normalize_rt_tasks(void)
7042 struct task_struct *g, *p;
7043 unsigned long flags;
7046 read_lock_irqsave(&tasklist_lock, flags);
7047 do_each_thread(g, p) {
7049 * Only normalize user tasks:
7054 p->se.exec_start = 0;
7055 #ifdef CONFIG_SCHEDSTATS
7056 p->se.statistics.wait_start = 0;
7057 p->se.statistics.sleep_start = 0;
7058 p->se.statistics.block_start = 0;
7063 * Renice negative nice level userspace
7066 if (TASK_NICE(p) < 0 && p->mm)
7067 set_user_nice(p, 0);
7071 raw_spin_lock(&p->pi_lock);
7072 rq = __task_rq_lock(p);
7074 normalize_task(rq, p);
7076 __task_rq_unlock(rq);
7077 raw_spin_unlock(&p->pi_lock);
7078 } while_each_thread(g, p);
7080 read_unlock_irqrestore(&tasklist_lock, flags);
7083 #endif /* CONFIG_MAGIC_SYSRQ */
7085 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7087 * These functions are only useful for the IA64 MCA handling, or kdb.
7089 * They can only be called when the whole system has been
7090 * stopped - every CPU needs to be quiescent, and no scheduling
7091 * activity can take place. Using them for anything else would
7092 * be a serious bug, and as a result, they aren't even visible
7093 * under any other configuration.
7097 * curr_task - return the current task for a given cpu.
7098 * @cpu: the processor in question.
7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7102 struct task_struct *curr_task(int cpu)
7104 return cpu_curr(cpu);
7107 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7111 * set_curr_task - set the current task for a given cpu.
7112 * @cpu: the processor in question.
7113 * @p: the task pointer to set.
7115 * Description: This function must only be used when non-maskable interrupts
7116 * are serviced on a separate stack. It allows the architecture to switch the
7117 * notion of the current task on a cpu in a non-blocking manner. This function
7118 * must be called with all CPU's synchronized, and interrupts disabled, the
7119 * and caller must save the original value of the current task (see
7120 * curr_task() above) and restore that value before reenabling interrupts and
7121 * re-starting the system.
7123 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7125 void set_curr_task(int cpu, struct task_struct *p)
7132 #ifdef CONFIG_CGROUP_SCHED
7133 /* task_group_lock serializes the addition/removal of task groups */
7134 static DEFINE_SPINLOCK(task_group_lock);
7136 static void free_sched_group(struct task_group *tg)
7138 free_fair_sched_group(tg);
7139 free_rt_sched_group(tg);
7144 /* allocate runqueue etc for a new task group */
7145 struct task_group *sched_create_group(struct task_group *parent)
7147 struct task_group *tg;
7148 unsigned long flags;
7150 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7152 return ERR_PTR(-ENOMEM);
7154 if (!alloc_fair_sched_group(tg, parent))
7157 if (!alloc_rt_sched_group(tg, parent))
7160 spin_lock_irqsave(&task_group_lock, flags);
7161 list_add_rcu(&tg->list, &task_groups);
7163 WARN_ON(!parent); /* root should already exist */
7165 tg->parent = parent;
7166 INIT_LIST_HEAD(&tg->children);
7167 list_add_rcu(&tg->siblings, &parent->children);
7168 spin_unlock_irqrestore(&task_group_lock, flags);
7173 free_sched_group(tg);
7174 return ERR_PTR(-ENOMEM);
7177 /* rcu callback to free various structures associated with a task group */
7178 static void free_sched_group_rcu(struct rcu_head *rhp)
7180 /* now it should be safe to free those cfs_rqs */
7181 free_sched_group(container_of(rhp, struct task_group, rcu));
7184 /* Destroy runqueue etc associated with a task group */
7185 void sched_destroy_group(struct task_group *tg)
7187 unsigned long flags;
7190 /* end participation in shares distribution */
7191 for_each_possible_cpu(i)
7192 unregister_fair_sched_group(tg, i);
7194 spin_lock_irqsave(&task_group_lock, flags);
7195 list_del_rcu(&tg->list);
7196 list_del_rcu(&tg->siblings);
7197 spin_unlock_irqrestore(&task_group_lock, flags);
7199 /* wait for possible concurrent references to cfs_rqs complete */
7200 call_rcu(&tg->rcu, free_sched_group_rcu);
7203 /* change task's runqueue when it moves between groups.
7204 * The caller of this function should have put the task in its new group
7205 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7206 * reflect its new group.
7208 void sched_move_task(struct task_struct *tsk)
7211 unsigned long flags;
7214 rq = task_rq_lock(tsk, &flags);
7216 running = task_current(rq, tsk);
7220 dequeue_task(rq, tsk, 0);
7221 if (unlikely(running))
7222 tsk->sched_class->put_prev_task(rq, tsk);
7224 #ifdef CONFIG_FAIR_GROUP_SCHED
7225 if (tsk->sched_class->task_move_group)
7226 tsk->sched_class->task_move_group(tsk, on_rq);
7229 set_task_rq(tsk, task_cpu(tsk));
7231 if (unlikely(running))
7232 tsk->sched_class->set_curr_task(rq);
7234 enqueue_task(rq, tsk, 0);
7236 task_rq_unlock(rq, tsk, &flags);
7238 #endif /* CONFIG_CGROUP_SCHED */
7240 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7241 static unsigned long to_ratio(u64 period, u64 runtime)
7243 if (runtime == RUNTIME_INF)
7246 return div64_u64(runtime << 20, period);
7250 #ifdef CONFIG_RT_GROUP_SCHED
7252 * Ensure that the real time constraints are schedulable.
7254 static DEFINE_MUTEX(rt_constraints_mutex);
7256 /* Must be called with tasklist_lock held */
7257 static inline int tg_has_rt_tasks(struct task_group *tg)
7259 struct task_struct *g, *p;
7261 do_each_thread(g, p) {
7262 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7264 } while_each_thread(g, p);
7269 struct rt_schedulable_data {
7270 struct task_group *tg;
7275 static int tg_rt_schedulable(struct task_group *tg, void *data)
7277 struct rt_schedulable_data *d = data;
7278 struct task_group *child;
7279 unsigned long total, sum = 0;
7280 u64 period, runtime;
7282 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7283 runtime = tg->rt_bandwidth.rt_runtime;
7286 period = d->rt_period;
7287 runtime = d->rt_runtime;
7291 * Cannot have more runtime than the period.
7293 if (runtime > period && runtime != RUNTIME_INF)
7297 * Ensure we don't starve existing RT tasks.
7299 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7302 total = to_ratio(period, runtime);
7305 * Nobody can have more than the global setting allows.
7307 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7311 * The sum of our children's runtime should not exceed our own.
7313 list_for_each_entry_rcu(child, &tg->children, siblings) {
7314 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7315 runtime = child->rt_bandwidth.rt_runtime;
7317 if (child == d->tg) {
7318 period = d->rt_period;
7319 runtime = d->rt_runtime;
7322 sum += to_ratio(period, runtime);
7331 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7335 struct rt_schedulable_data data = {
7337 .rt_period = period,
7338 .rt_runtime = runtime,
7342 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7348 static int tg_set_rt_bandwidth(struct task_group *tg,
7349 u64 rt_period, u64 rt_runtime)
7353 mutex_lock(&rt_constraints_mutex);
7354 read_lock(&tasklist_lock);
7355 err = __rt_schedulable(tg, rt_period, rt_runtime);
7359 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7360 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7361 tg->rt_bandwidth.rt_runtime = rt_runtime;
7363 for_each_possible_cpu(i) {
7364 struct rt_rq *rt_rq = tg->rt_rq[i];
7366 raw_spin_lock(&rt_rq->rt_runtime_lock);
7367 rt_rq->rt_runtime = rt_runtime;
7368 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7370 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7372 read_unlock(&tasklist_lock);
7373 mutex_unlock(&rt_constraints_mutex);
7378 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7380 u64 rt_runtime, rt_period;
7382 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7383 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7384 if (rt_runtime_us < 0)
7385 rt_runtime = RUNTIME_INF;
7387 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7390 long sched_group_rt_runtime(struct task_group *tg)
7394 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7397 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7398 do_div(rt_runtime_us, NSEC_PER_USEC);
7399 return rt_runtime_us;
7402 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7404 u64 rt_runtime, rt_period;
7406 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7407 rt_runtime = tg->rt_bandwidth.rt_runtime;
7412 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7415 long sched_group_rt_period(struct task_group *tg)
7419 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7420 do_div(rt_period_us, NSEC_PER_USEC);
7421 return rt_period_us;
7424 static int sched_rt_global_constraints(void)
7426 u64 runtime, period;
7429 if (sysctl_sched_rt_period <= 0)
7432 runtime = global_rt_runtime();
7433 period = global_rt_period();
7436 * Sanity check on the sysctl variables.
7438 if (runtime > period && runtime != RUNTIME_INF)
7441 mutex_lock(&rt_constraints_mutex);
7442 read_lock(&tasklist_lock);
7443 ret = __rt_schedulable(NULL, 0, 0);
7444 read_unlock(&tasklist_lock);
7445 mutex_unlock(&rt_constraints_mutex);
7450 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7452 /* Don't accept realtime tasks when there is no way for them to run */
7453 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7459 #else /* !CONFIG_RT_GROUP_SCHED */
7460 static int sched_rt_global_constraints(void)
7462 unsigned long flags;
7465 if (sysctl_sched_rt_period <= 0)
7469 * There's always some RT tasks in the root group
7470 * -- migration, kstopmachine etc..
7472 if (sysctl_sched_rt_runtime == 0)
7475 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7476 for_each_possible_cpu(i) {
7477 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7479 raw_spin_lock(&rt_rq->rt_runtime_lock);
7480 rt_rq->rt_runtime = global_rt_runtime();
7481 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7483 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7487 #endif /* CONFIG_RT_GROUP_SCHED */
7489 int sched_rt_handler(struct ctl_table *table, int write,
7490 void __user *buffer, size_t *lenp,
7494 int old_period, old_runtime;
7495 static DEFINE_MUTEX(mutex);
7498 old_period = sysctl_sched_rt_period;
7499 old_runtime = sysctl_sched_rt_runtime;
7501 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7503 if (!ret && write) {
7504 ret = sched_rt_global_constraints();
7506 sysctl_sched_rt_period = old_period;
7507 sysctl_sched_rt_runtime = old_runtime;
7509 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7510 def_rt_bandwidth.rt_period =
7511 ns_to_ktime(global_rt_period());
7514 mutex_unlock(&mutex);
7519 #ifdef CONFIG_CGROUP_SCHED
7521 /* return corresponding task_group object of a cgroup */
7522 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7524 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7525 struct task_group, css);
7528 static struct cgroup_subsys_state *
7529 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7531 struct task_group *tg, *parent;
7533 if (!cgrp->parent) {
7534 /* This is early initialization for the top cgroup */
7535 return &root_task_group.css;
7538 parent = cgroup_tg(cgrp->parent);
7539 tg = sched_create_group(parent);
7541 return ERR_PTR(-ENOMEM);
7547 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7549 struct task_group *tg = cgroup_tg(cgrp);
7551 sched_destroy_group(tg);
7554 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7555 struct cgroup_taskset *tset)
7557 struct task_struct *task;
7559 cgroup_taskset_for_each(task, cgrp, tset) {
7560 #ifdef CONFIG_RT_GROUP_SCHED
7561 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7564 /* We don't support RT-tasks being in separate groups */
7565 if (task->sched_class != &fair_sched_class)
7572 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7573 struct cgroup_taskset *tset)
7575 struct task_struct *task;
7577 cgroup_taskset_for_each(task, cgrp, tset)
7578 sched_move_task(task);
7582 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7583 struct cgroup *old_cgrp, struct task_struct *task)
7586 * cgroup_exit() is called in the copy_process() failure path.
7587 * Ignore this case since the task hasn't ran yet, this avoids
7588 * trying to poke a half freed task state from generic code.
7590 if (!(task->flags & PF_EXITING))
7593 sched_move_task(task);
7596 #ifdef CONFIG_FAIR_GROUP_SCHED
7597 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7600 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7603 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7605 struct task_group *tg = cgroup_tg(cgrp);
7607 return (u64) scale_load_down(tg->shares);
7610 #ifdef CONFIG_CFS_BANDWIDTH
7611 static DEFINE_MUTEX(cfs_constraints_mutex);
7613 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7614 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7616 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7618 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7620 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7621 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7623 if (tg == &root_task_group)
7627 * Ensure we have at some amount of bandwidth every period. This is
7628 * to prevent reaching a state of large arrears when throttled via
7629 * entity_tick() resulting in prolonged exit starvation.
7631 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7635 * Likewise, bound things on the otherside by preventing insane quota
7636 * periods. This also allows us to normalize in computing quota
7639 if (period > max_cfs_quota_period)
7642 mutex_lock(&cfs_constraints_mutex);
7643 ret = __cfs_schedulable(tg, period, quota);
7647 runtime_enabled = quota != RUNTIME_INF;
7648 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7649 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7650 raw_spin_lock_irq(&cfs_b->lock);
7651 cfs_b->period = ns_to_ktime(period);
7652 cfs_b->quota = quota;
7654 __refill_cfs_bandwidth_runtime(cfs_b);
7655 /* restart the period timer (if active) to handle new period expiry */
7656 if (runtime_enabled && cfs_b->timer_active) {
7657 /* force a reprogram */
7658 cfs_b->timer_active = 0;
7659 __start_cfs_bandwidth(cfs_b);
7661 raw_spin_unlock_irq(&cfs_b->lock);
7663 for_each_possible_cpu(i) {
7664 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7665 struct rq *rq = cfs_rq->rq;
7667 raw_spin_lock_irq(&rq->lock);
7668 cfs_rq->runtime_enabled = runtime_enabled;
7669 cfs_rq->runtime_remaining = 0;
7671 if (cfs_rq->throttled)
7672 unthrottle_cfs_rq(cfs_rq);
7673 raw_spin_unlock_irq(&rq->lock);
7676 mutex_unlock(&cfs_constraints_mutex);
7681 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7685 period = ktime_to_ns(tg->cfs_bandwidth.period);
7686 if (cfs_quota_us < 0)
7687 quota = RUNTIME_INF;
7689 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7691 return tg_set_cfs_bandwidth(tg, period, quota);
7694 long tg_get_cfs_quota(struct task_group *tg)
7698 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7701 quota_us = tg->cfs_bandwidth.quota;
7702 do_div(quota_us, NSEC_PER_USEC);
7707 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7711 period = (u64)cfs_period_us * NSEC_PER_USEC;
7712 quota = tg->cfs_bandwidth.quota;
7714 return tg_set_cfs_bandwidth(tg, period, quota);
7717 long tg_get_cfs_period(struct task_group *tg)
7721 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7722 do_div(cfs_period_us, NSEC_PER_USEC);
7724 return cfs_period_us;
7727 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7729 return tg_get_cfs_quota(cgroup_tg(cgrp));
7732 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7735 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7738 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7740 return tg_get_cfs_period(cgroup_tg(cgrp));
7743 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7746 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7749 struct cfs_schedulable_data {
7750 struct task_group *tg;
7755 * normalize group quota/period to be quota/max_period
7756 * note: units are usecs
7758 static u64 normalize_cfs_quota(struct task_group *tg,
7759 struct cfs_schedulable_data *d)
7767 period = tg_get_cfs_period(tg);
7768 quota = tg_get_cfs_quota(tg);
7771 /* note: these should typically be equivalent */
7772 if (quota == RUNTIME_INF || quota == -1)
7775 return to_ratio(period, quota);
7778 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7780 struct cfs_schedulable_data *d = data;
7781 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7782 s64 quota = 0, parent_quota = -1;
7785 quota = RUNTIME_INF;
7787 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7789 quota = normalize_cfs_quota(tg, d);
7790 parent_quota = parent_b->hierarchal_quota;
7793 * ensure max(child_quota) <= parent_quota, inherit when no
7796 if (quota == RUNTIME_INF)
7797 quota = parent_quota;
7798 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7801 cfs_b->hierarchal_quota = quota;
7806 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7809 struct cfs_schedulable_data data = {
7815 if (quota != RUNTIME_INF) {
7816 do_div(data.period, NSEC_PER_USEC);
7817 do_div(data.quota, NSEC_PER_USEC);
7821 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7827 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7828 struct cgroup_map_cb *cb)
7830 struct task_group *tg = cgroup_tg(cgrp);
7831 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7833 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7834 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7835 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7839 #endif /* CONFIG_CFS_BANDWIDTH */
7840 #endif /* CONFIG_FAIR_GROUP_SCHED */
7842 #ifdef CONFIG_RT_GROUP_SCHED
7843 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7846 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7849 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7851 return sched_group_rt_runtime(cgroup_tg(cgrp));
7854 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7857 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7860 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7862 return sched_group_rt_period(cgroup_tg(cgrp));
7864 #endif /* CONFIG_RT_GROUP_SCHED */
7866 static struct cftype cpu_files[] = {
7867 #ifdef CONFIG_FAIR_GROUP_SCHED
7870 .read_u64 = cpu_shares_read_u64,
7871 .write_u64 = cpu_shares_write_u64,
7874 #ifdef CONFIG_CFS_BANDWIDTH
7876 .name = "cfs_quota_us",
7877 .read_s64 = cpu_cfs_quota_read_s64,
7878 .write_s64 = cpu_cfs_quota_write_s64,
7881 .name = "cfs_period_us",
7882 .read_u64 = cpu_cfs_period_read_u64,
7883 .write_u64 = cpu_cfs_period_write_u64,
7887 .read_map = cpu_stats_show,
7890 #ifdef CONFIG_RT_GROUP_SCHED
7892 .name = "rt_runtime_us",
7893 .read_s64 = cpu_rt_runtime_read,
7894 .write_s64 = cpu_rt_runtime_write,
7897 .name = "rt_period_us",
7898 .read_u64 = cpu_rt_period_read_uint,
7899 .write_u64 = cpu_rt_period_write_uint,
7904 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7906 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7909 struct cgroup_subsys cpu_cgroup_subsys = {
7911 .create = cpu_cgroup_create,
7912 .destroy = cpu_cgroup_destroy,
7913 .can_attach = cpu_cgroup_can_attach,
7914 .attach = cpu_cgroup_attach,
7915 .exit = cpu_cgroup_exit,
7916 .populate = cpu_cgroup_populate,
7917 .subsys_id = cpu_cgroup_subsys_id,
7921 #endif /* CONFIG_CGROUP_SCHED */
7923 #ifdef CONFIG_CGROUP_CPUACCT
7926 * CPU accounting code for task groups.
7928 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7929 * (balbir@in.ibm.com).
7932 /* create a new cpu accounting group */
7933 static struct cgroup_subsys_state *cpuacct_create(
7934 struct cgroup_subsys *ss, struct cgroup *cgrp)
7939 return &root_cpuacct.css;
7941 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7945 ca->cpuusage = alloc_percpu(u64);
7949 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7951 goto out_free_cpuusage;
7956 free_percpu(ca->cpuusage);
7960 return ERR_PTR(-ENOMEM);
7963 /* destroy an existing cpu accounting group */
7965 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7967 struct cpuacct *ca = cgroup_ca(cgrp);
7969 free_percpu(ca->cpustat);
7970 free_percpu(ca->cpuusage);
7974 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7976 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7979 #ifndef CONFIG_64BIT
7981 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7983 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7985 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7993 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7995 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7997 #ifndef CONFIG_64BIT
7999 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8001 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8003 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8009 /* return total cpu usage (in nanoseconds) of a group */
8010 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8012 struct cpuacct *ca = cgroup_ca(cgrp);
8013 u64 totalcpuusage = 0;
8016 for_each_present_cpu(i)
8017 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8019 return totalcpuusage;
8022 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8025 struct cpuacct *ca = cgroup_ca(cgrp);
8034 for_each_present_cpu(i)
8035 cpuacct_cpuusage_write(ca, i, 0);
8041 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8044 struct cpuacct *ca = cgroup_ca(cgroup);
8048 for_each_present_cpu(i) {
8049 percpu = cpuacct_cpuusage_read(ca, i);
8050 seq_printf(m, "%llu ", (unsigned long long) percpu);
8052 seq_printf(m, "\n");
8056 static const char *cpuacct_stat_desc[] = {
8057 [CPUACCT_STAT_USER] = "user",
8058 [CPUACCT_STAT_SYSTEM] = "system",
8061 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8062 struct cgroup_map_cb *cb)
8064 struct cpuacct *ca = cgroup_ca(cgrp);
8068 for_each_online_cpu(cpu) {
8069 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8070 val += kcpustat->cpustat[CPUTIME_USER];
8071 val += kcpustat->cpustat[CPUTIME_NICE];
8073 val = cputime64_to_clock_t(val);
8074 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8077 for_each_online_cpu(cpu) {
8078 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8079 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8080 val += kcpustat->cpustat[CPUTIME_IRQ];
8081 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8084 val = cputime64_to_clock_t(val);
8085 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8090 static struct cftype files[] = {
8093 .read_u64 = cpuusage_read,
8094 .write_u64 = cpuusage_write,
8097 .name = "usage_percpu",
8098 .read_seq_string = cpuacct_percpu_seq_read,
8102 .read_map = cpuacct_stats_show,
8106 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8108 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8112 * charge this task's execution time to its accounting group.
8114 * called with rq->lock held.
8116 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8121 if (unlikely(!cpuacct_subsys.active))
8124 cpu = task_cpu(tsk);
8130 for (; ca; ca = parent_ca(ca)) {
8131 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8132 *cpuusage += cputime;
8138 struct cgroup_subsys cpuacct_subsys = {
8140 .create = cpuacct_create,
8141 .destroy = cpuacct_destroy,
8142 .populate = cpuacct_populate,
8143 .subsys_id = cpuacct_subsys_id,
8145 #endif /* CONFIG_CGROUP_CPUACCT */