4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
79 #ifdef CONFIG_PARAVIRT
80 #include <asm/paravirt.h>
84 #include "../workqueue_sched.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
89 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
92 ktime_t soft, hard, now;
95 if (hrtimer_active(period_timer))
98 now = hrtimer_cb_get_time(period_timer);
99 hrtimer_forward(period_timer, now, period);
101 soft = hrtimer_get_softexpires(period_timer);
102 hard = hrtimer_get_expires(period_timer);
103 delta = ktime_to_ns(ktime_sub(hard, soft));
104 __hrtimer_start_range_ns(period_timer, soft, delta,
105 HRTIMER_MODE_ABS_PINNED, 0);
109 DEFINE_MUTEX(sched_domains_mutex);
110 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
112 static void update_rq_clock_task(struct rq *rq, s64 delta);
114 void update_rq_clock(struct rq *rq)
118 if (rq->skip_clock_update > 0)
121 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
123 update_rq_clock_task(rq, delta);
127 * Debugging: various feature bits
130 #define SCHED_FEAT(name, enabled) \
131 (1UL << __SCHED_FEAT_##name) * enabled |
133 const_debug unsigned int sysctl_sched_features =
134 #include "features.h"
139 #ifdef CONFIG_SCHED_DEBUG
140 #define SCHED_FEAT(name, enabled) \
143 static __read_mostly char *sched_feat_names[] = {
144 #include "features.h"
150 static int sched_feat_show(struct seq_file *m, void *v)
154 for (i = 0; i < __SCHED_FEAT_NR; i++) {
155 if (!(sysctl_sched_features & (1UL << i)))
157 seq_printf(m, "%s ", sched_feat_names[i]);
164 #ifdef HAVE_JUMP_LABEL
166 #define jump_label_key__true jump_label_key_enabled
167 #define jump_label_key__false jump_label_key_disabled
169 #define SCHED_FEAT(name, enabled) \
170 jump_label_key__##enabled ,
172 struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = {
173 #include "features.h"
178 static void sched_feat_disable(int i)
180 if (jump_label_enabled(&sched_feat_keys[i]))
181 jump_label_dec(&sched_feat_keys[i]);
184 static void sched_feat_enable(int i)
186 if (!jump_label_enabled(&sched_feat_keys[i]))
187 jump_label_inc(&sched_feat_keys[i]);
190 static void sched_feat_disable(int i) { };
191 static void sched_feat_enable(int i) { };
192 #endif /* HAVE_JUMP_LABEL */
195 sched_feat_write(struct file *filp, const char __user *ubuf,
196 size_t cnt, loff_t *ppos)
206 if (copy_from_user(&buf, ubuf, cnt))
212 if (strncmp(cmp, "NO_", 3) == 0) {
217 for (i = 0; i < __SCHED_FEAT_NR; i++) {
218 if (strcmp(cmp, sched_feat_names[i]) == 0) {
220 sysctl_sched_features &= ~(1UL << i);
221 sched_feat_disable(i);
223 sysctl_sched_features |= (1UL << i);
224 sched_feat_enable(i);
230 if (i == __SCHED_FEAT_NR)
238 static int sched_feat_open(struct inode *inode, struct file *filp)
240 return single_open(filp, sched_feat_show, NULL);
243 static const struct file_operations sched_feat_fops = {
244 .open = sched_feat_open,
245 .write = sched_feat_write,
248 .release = single_release,
251 static __init int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL, NULL,
258 late_initcall(sched_init_debug);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug unsigned int sysctl_sched_nr_migrate = 32;
268 * period over which we average the RT time consumption, measured
273 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period = 1000000;
281 __read_mostly int scheduler_running;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime = 950000;
292 * __task_rq_lock - lock the rq @p resides on.
294 static inline struct rq *__task_rq_lock(struct task_struct *p)
299 lockdep_assert_held(&p->pi_lock);
303 raw_spin_lock(&rq->lock);
304 if (likely(rq == task_rq(p)))
306 raw_spin_unlock(&rq->lock);
311 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
313 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
314 __acquires(p->pi_lock)
320 raw_spin_lock_irqsave(&p->pi_lock, *flags);
322 raw_spin_lock(&rq->lock);
323 if (likely(rq == task_rq(p)))
325 raw_spin_unlock(&rq->lock);
326 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
330 static void __task_rq_unlock(struct rq *rq)
333 raw_spin_unlock(&rq->lock);
337 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
339 __releases(p->pi_lock)
341 raw_spin_unlock(&rq->lock);
342 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
346 * this_rq_lock - lock this runqueue and disable interrupts.
348 static struct rq *this_rq_lock(void)
355 raw_spin_lock(&rq->lock);
360 #ifdef CONFIG_SCHED_HRTICK
362 * Use HR-timers to deliver accurate preemption points.
364 * Its all a bit involved since we cannot program an hrt while holding the
365 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
368 * When we get rescheduled we reprogram the hrtick_timer outside of the
372 static void hrtick_clear(struct rq *rq)
374 if (hrtimer_active(&rq->hrtick_timer))
375 hrtimer_cancel(&rq->hrtick_timer);
379 * High-resolution timer tick.
380 * Runs from hardirq context with interrupts disabled.
382 static enum hrtimer_restart hrtick(struct hrtimer *timer)
384 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
386 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
388 raw_spin_lock(&rq->lock);
390 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
391 raw_spin_unlock(&rq->lock);
393 return HRTIMER_NORESTART;
398 * called from hardirq (IPI) context
400 static void __hrtick_start(void *arg)
404 raw_spin_lock(&rq->lock);
405 hrtimer_restart(&rq->hrtick_timer);
406 rq->hrtick_csd_pending = 0;
407 raw_spin_unlock(&rq->lock);
411 * Called to set the hrtick timer state.
413 * called with rq->lock held and irqs disabled
415 void hrtick_start(struct rq *rq, u64 delay)
417 struct hrtimer *timer = &rq->hrtick_timer;
418 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
420 hrtimer_set_expires(timer, time);
422 if (rq == this_rq()) {
423 hrtimer_restart(timer);
424 } else if (!rq->hrtick_csd_pending) {
425 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
426 rq->hrtick_csd_pending = 1;
431 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
433 int cpu = (int)(long)hcpu;
436 case CPU_UP_CANCELED:
437 case CPU_UP_CANCELED_FROZEN:
438 case CPU_DOWN_PREPARE:
439 case CPU_DOWN_PREPARE_FROZEN:
441 case CPU_DEAD_FROZEN:
442 hrtick_clear(cpu_rq(cpu));
449 static __init void init_hrtick(void)
451 hotcpu_notifier(hotplug_hrtick, 0);
455 * Called to set the hrtick timer state.
457 * called with rq->lock held and irqs disabled
459 void hrtick_start(struct rq *rq, u64 delay)
461 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
462 HRTIMER_MODE_REL_PINNED, 0);
465 static inline void init_hrtick(void)
468 #endif /* CONFIG_SMP */
470 static void init_rq_hrtick(struct rq *rq)
473 rq->hrtick_csd_pending = 0;
475 rq->hrtick_csd.flags = 0;
476 rq->hrtick_csd.func = __hrtick_start;
477 rq->hrtick_csd.info = rq;
480 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
481 rq->hrtick_timer.function = hrtick;
483 #else /* CONFIG_SCHED_HRTICK */
484 static inline void hrtick_clear(struct rq *rq)
488 static inline void init_rq_hrtick(struct rq *rq)
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SCHED_HRTICK */
498 * resched_task - mark a task 'to be rescheduled now'.
500 * On UP this means the setting of the need_resched flag, on SMP it
501 * might also involve a cross-CPU call to trigger the scheduler on
506 #ifndef tsk_is_polling
507 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
510 void resched_task(struct task_struct *p)
514 assert_raw_spin_locked(&task_rq(p)->lock);
516 if (test_tsk_need_resched(p))
519 set_tsk_need_resched(p);
522 if (cpu == smp_processor_id())
525 /* NEED_RESCHED must be visible before we test polling */
527 if (!tsk_is_polling(p))
528 smp_send_reschedule(cpu);
531 void resched_cpu(int cpu)
533 struct rq *rq = cpu_rq(cpu);
536 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
538 resched_task(cpu_curr(cpu));
539 raw_spin_unlock_irqrestore(&rq->lock, flags);
544 * In the semi idle case, use the nearest busy cpu for migrating timers
545 * from an idle cpu. This is good for power-savings.
547 * We don't do similar optimization for completely idle system, as
548 * selecting an idle cpu will add more delays to the timers than intended
549 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
551 int get_nohz_timer_target(void)
553 int cpu = smp_processor_id();
555 struct sched_domain *sd;
558 for_each_domain(cpu, sd) {
559 for_each_cpu(i, sched_domain_span(sd)) {
571 * When add_timer_on() enqueues a timer into the timer wheel of an
572 * idle CPU then this timer might expire before the next timer event
573 * which is scheduled to wake up that CPU. In case of a completely
574 * idle system the next event might even be infinite time into the
575 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
576 * leaves the inner idle loop so the newly added timer is taken into
577 * account when the CPU goes back to idle and evaluates the timer
578 * wheel for the next timer event.
580 void wake_up_idle_cpu(int cpu)
582 struct rq *rq = cpu_rq(cpu);
584 if (cpu == smp_processor_id())
588 * This is safe, as this function is called with the timer
589 * wheel base lock of (cpu) held. When the CPU is on the way
590 * to idle and has not yet set rq->curr to idle then it will
591 * be serialized on the timer wheel base lock and take the new
592 * timer into account automatically.
594 if (rq->curr != rq->idle)
598 * We can set TIF_RESCHED on the idle task of the other CPU
599 * lockless. The worst case is that the other CPU runs the
600 * idle task through an additional NOOP schedule()
602 set_tsk_need_resched(rq->idle);
604 /* NEED_RESCHED must be visible before we test polling */
606 if (!tsk_is_polling(rq->idle))
607 smp_send_reschedule(cpu);
610 static inline bool got_nohz_idle_kick(void)
612 int cpu = smp_processor_id();
613 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
616 #else /* CONFIG_NO_HZ */
618 static inline bool got_nohz_idle_kick(void)
623 #endif /* CONFIG_NO_HZ */
625 void sched_avg_update(struct rq *rq)
627 s64 period = sched_avg_period();
629 while ((s64)(rq->clock - rq->age_stamp) > period) {
631 * Inline assembly required to prevent the compiler
632 * optimising this loop into a divmod call.
633 * See __iter_div_u64_rem() for another example of this.
635 asm("" : "+rm" (rq->age_stamp));
636 rq->age_stamp += period;
641 #else /* !CONFIG_SMP */
642 void resched_task(struct task_struct *p)
644 assert_raw_spin_locked(&task_rq(p)->lock);
645 set_tsk_need_resched(p);
647 #endif /* CONFIG_SMP */
649 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
650 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
652 * Iterate task_group tree rooted at *from, calling @down when first entering a
653 * node and @up when leaving it for the final time.
655 * Caller must hold rcu_lock or sufficient equivalent.
657 int walk_tg_tree_from(struct task_group *from,
658 tg_visitor down, tg_visitor up, void *data)
660 struct task_group *parent, *child;
666 ret = (*down)(parent, data);
669 list_for_each_entry_rcu(child, &parent->children, siblings) {
676 ret = (*up)(parent, data);
677 if (ret || parent == from)
681 parent = parent->parent;
688 int tg_nop(struct task_group *tg, void *data)
694 void update_cpu_load(struct rq *this_rq);
696 static void set_load_weight(struct task_struct *p)
698 int prio = p->static_prio - MAX_RT_PRIO;
699 struct load_weight *load = &p->se.load;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p->policy == SCHED_IDLE) {
705 load->weight = scale_load(WEIGHT_IDLEPRIO);
706 load->inv_weight = WMULT_IDLEPRIO;
710 load->weight = scale_load(prio_to_weight[prio]);
711 load->inv_weight = prio_to_wmult[prio];
714 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
717 sched_info_queued(p);
718 p->sched_class->enqueue_task(rq, p, flags);
721 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
724 sched_info_dequeued(p);
725 p->sched_class->dequeue_task(rq, p, flags);
728 void activate_task(struct rq *rq, struct task_struct *p, int flags)
730 if (task_contributes_to_load(p))
731 rq->nr_uninterruptible--;
733 enqueue_task(rq, p, flags);
736 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible++;
741 dequeue_task(rq, p, flags);
744 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
747 * There are no locks covering percpu hardirq/softirq time.
748 * They are only modified in account_system_vtime, on corresponding CPU
749 * with interrupts disabled. So, writes are safe.
750 * They are read and saved off onto struct rq in update_rq_clock().
751 * This may result in other CPU reading this CPU's irq time and can
752 * race with irq/account_system_vtime on this CPU. We would either get old
753 * or new value with a side effect of accounting a slice of irq time to wrong
754 * task when irq is in progress while we read rq->clock. That is a worthy
755 * compromise in place of having locks on each irq in account_system_time.
757 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
758 static DEFINE_PER_CPU(u64, cpu_softirq_time);
760 static DEFINE_PER_CPU(u64, irq_start_time);
761 static int sched_clock_irqtime;
763 void enable_sched_clock_irqtime(void)
765 sched_clock_irqtime = 1;
768 void disable_sched_clock_irqtime(void)
770 sched_clock_irqtime = 0;
774 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
776 static inline void irq_time_write_begin(void)
778 __this_cpu_inc(irq_time_seq.sequence);
782 static inline void irq_time_write_end(void)
785 __this_cpu_inc(irq_time_seq.sequence);
788 static inline u64 irq_time_read(int cpu)
794 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
795 irq_time = per_cpu(cpu_softirq_time, cpu) +
796 per_cpu(cpu_hardirq_time, cpu);
797 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
801 #else /* CONFIG_64BIT */
802 static inline void irq_time_write_begin(void)
806 static inline void irq_time_write_end(void)
810 static inline u64 irq_time_read(int cpu)
812 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
814 #endif /* CONFIG_64BIT */
817 * Called before incrementing preempt_count on {soft,}irq_enter
818 * and before decrementing preempt_count on {soft,}irq_exit.
820 void account_system_vtime(struct task_struct *curr)
826 if (!sched_clock_irqtime)
829 local_irq_save(flags);
831 cpu = smp_processor_id();
832 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
833 __this_cpu_add(irq_start_time, delta);
835 irq_time_write_begin();
837 * We do not account for softirq time from ksoftirqd here.
838 * We want to continue accounting softirq time to ksoftirqd thread
839 * in that case, so as not to confuse scheduler with a special task
840 * that do not consume any time, but still wants to run.
843 __this_cpu_add(cpu_hardirq_time, delta);
844 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
845 __this_cpu_add(cpu_softirq_time, delta);
847 irq_time_write_end();
848 local_irq_restore(flags);
850 EXPORT_SYMBOL_GPL(account_system_vtime);
852 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
854 #ifdef CONFIG_PARAVIRT
855 static inline u64 steal_ticks(u64 steal)
857 if (unlikely(steal > NSEC_PER_SEC))
858 return div_u64(steal, TICK_NSEC);
860 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
864 static void update_rq_clock_task(struct rq *rq, s64 delta)
867 * In theory, the compile should just see 0 here, and optimize out the call
868 * to sched_rt_avg_update. But I don't trust it...
870 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
871 s64 steal = 0, irq_delta = 0;
873 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
874 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
877 * Since irq_time is only updated on {soft,}irq_exit, we might run into
878 * this case when a previous update_rq_clock() happened inside a
881 * When this happens, we stop ->clock_task and only update the
882 * prev_irq_time stamp to account for the part that fit, so that a next
883 * update will consume the rest. This ensures ->clock_task is
886 * It does however cause some slight miss-attribution of {soft,}irq
887 * time, a more accurate solution would be to update the irq_time using
888 * the current rq->clock timestamp, except that would require using
891 if (irq_delta > delta)
894 rq->prev_irq_time += irq_delta;
897 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
898 if (static_branch((¶virt_steal_rq_enabled))) {
901 steal = paravirt_steal_clock(cpu_of(rq));
902 steal -= rq->prev_steal_time_rq;
904 if (unlikely(steal > delta))
907 st = steal_ticks(steal);
908 steal = st * TICK_NSEC;
910 rq->prev_steal_time_rq += steal;
916 rq->clock_task += delta;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
920 sched_rt_avg_update(rq, irq_delta + steal);
924 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
925 static int irqtime_account_hi_update(void)
927 u64 *cpustat = kcpustat_this_cpu->cpustat;
932 local_irq_save(flags);
933 latest_ns = this_cpu_read(cpu_hardirq_time);
934 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
936 local_irq_restore(flags);
940 static int irqtime_account_si_update(void)
942 u64 *cpustat = kcpustat_this_cpu->cpustat;
947 local_irq_save(flags);
948 latest_ns = this_cpu_read(cpu_softirq_time);
949 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
951 local_irq_restore(flags);
955 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
957 #define sched_clock_irqtime (0)
961 void sched_set_stop_task(int cpu, struct task_struct *stop)
963 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
964 struct task_struct *old_stop = cpu_rq(cpu)->stop;
968 * Make it appear like a SCHED_FIFO task, its something
969 * userspace knows about and won't get confused about.
971 * Also, it will make PI more or less work without too
972 * much confusion -- but then, stop work should not
973 * rely on PI working anyway.
975 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
977 stop->sched_class = &stop_sched_class;
980 cpu_rq(cpu)->stop = stop;
984 * Reset it back to a normal scheduling class so that
985 * it can die in pieces.
987 old_stop->sched_class = &rt_sched_class;
992 * __normal_prio - return the priority that is based on the static prio
994 static inline int __normal_prio(struct task_struct *p)
996 return p->static_prio;
1000 * Calculate the expected normal priority: i.e. priority
1001 * without taking RT-inheritance into account. Might be
1002 * boosted by interactivity modifiers. Changes upon fork,
1003 * setprio syscalls, and whenever the interactivity
1004 * estimator recalculates.
1006 static inline int normal_prio(struct task_struct *p)
1010 if (task_has_rt_policy(p))
1011 prio = MAX_RT_PRIO-1 - p->rt_priority;
1013 prio = __normal_prio(p);
1018 * Calculate the current priority, i.e. the priority
1019 * taken into account by the scheduler. This value might
1020 * be boosted by RT tasks, or might be boosted by
1021 * interactivity modifiers. Will be RT if the task got
1022 * RT-boosted. If not then it returns p->normal_prio.
1024 static int effective_prio(struct task_struct *p)
1026 p->normal_prio = normal_prio(p);
1028 * If we are RT tasks or we were boosted to RT priority,
1029 * keep the priority unchanged. Otherwise, update priority
1030 * to the normal priority:
1032 if (!rt_prio(p->prio))
1033 return p->normal_prio;
1038 * task_curr - is this task currently executing on a CPU?
1039 * @p: the task in question.
1041 inline int task_curr(const struct task_struct *p)
1043 return cpu_curr(task_cpu(p)) == p;
1046 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1047 const struct sched_class *prev_class,
1050 if (prev_class != p->sched_class) {
1051 if (prev_class->switched_from)
1052 prev_class->switched_from(rq, p);
1053 p->sched_class->switched_to(rq, p);
1054 } else if (oldprio != p->prio)
1055 p->sched_class->prio_changed(rq, p, oldprio);
1058 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1060 const struct sched_class *class;
1062 if (p->sched_class == rq->curr->sched_class) {
1063 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1065 for_each_class(class) {
1066 if (class == rq->curr->sched_class)
1068 if (class == p->sched_class) {
1069 resched_task(rq->curr);
1076 * A queue event has occurred, and we're going to schedule. In
1077 * this case, we can save a useless back to back clock update.
1079 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1080 rq->skip_clock_update = 1;
1084 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1086 #ifdef CONFIG_SCHED_DEBUG
1088 * We should never call set_task_cpu() on a blocked task,
1089 * ttwu() will sort out the placement.
1091 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1092 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1094 #ifdef CONFIG_LOCKDEP
1096 * The caller should hold either p->pi_lock or rq->lock, when changing
1097 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1099 * sched_move_task() holds both and thus holding either pins the cgroup,
1100 * see set_task_rq().
1102 * Furthermore, all task_rq users should acquire both locks, see
1105 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1106 lockdep_is_held(&task_rq(p)->lock)));
1110 trace_sched_migrate_task(p, new_cpu);
1112 if (task_cpu(p) != new_cpu) {
1113 p->se.nr_migrations++;
1114 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1117 __set_task_cpu(p, new_cpu);
1120 struct migration_arg {
1121 struct task_struct *task;
1125 static int migration_cpu_stop(void *data);
1128 * wait_task_inactive - wait for a thread to unschedule.
1130 * If @match_state is nonzero, it's the @p->state value just checked and
1131 * not expected to change. If it changes, i.e. @p might have woken up,
1132 * then return zero. When we succeed in waiting for @p to be off its CPU,
1133 * we return a positive number (its total switch count). If a second call
1134 * a short while later returns the same number, the caller can be sure that
1135 * @p has remained unscheduled the whole time.
1137 * The caller must ensure that the task *will* unschedule sometime soon,
1138 * else this function might spin for a *long* time. This function can't
1139 * be called with interrupts off, or it may introduce deadlock with
1140 * smp_call_function() if an IPI is sent by the same process we are
1141 * waiting to become inactive.
1143 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1145 unsigned long flags;
1152 * We do the initial early heuristics without holding
1153 * any task-queue locks at all. We'll only try to get
1154 * the runqueue lock when things look like they will
1160 * If the task is actively running on another CPU
1161 * still, just relax and busy-wait without holding
1164 * NOTE! Since we don't hold any locks, it's not
1165 * even sure that "rq" stays as the right runqueue!
1166 * But we don't care, since "task_running()" will
1167 * return false if the runqueue has changed and p
1168 * is actually now running somewhere else!
1170 while (task_running(rq, p)) {
1171 if (match_state && unlikely(p->state != match_state))
1177 * Ok, time to look more closely! We need the rq
1178 * lock now, to be *sure*. If we're wrong, we'll
1179 * just go back and repeat.
1181 rq = task_rq_lock(p, &flags);
1182 trace_sched_wait_task(p);
1183 running = task_running(rq, p);
1186 if (!match_state || p->state == match_state)
1187 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1188 task_rq_unlock(rq, p, &flags);
1191 * If it changed from the expected state, bail out now.
1193 if (unlikely(!ncsw))
1197 * Was it really running after all now that we
1198 * checked with the proper locks actually held?
1200 * Oops. Go back and try again..
1202 if (unlikely(running)) {
1208 * It's not enough that it's not actively running,
1209 * it must be off the runqueue _entirely_, and not
1212 * So if it was still runnable (but just not actively
1213 * running right now), it's preempted, and we should
1214 * yield - it could be a while.
1216 if (unlikely(on_rq)) {
1217 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1219 set_current_state(TASK_UNINTERRUPTIBLE);
1220 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1225 * Ahh, all good. It wasn't running, and it wasn't
1226 * runnable, which means that it will never become
1227 * running in the future either. We're all done!
1236 * kick_process - kick a running thread to enter/exit the kernel
1237 * @p: the to-be-kicked thread
1239 * Cause a process which is running on another CPU to enter
1240 * kernel-mode, without any delay. (to get signals handled.)
1242 * NOTE: this function doesn't have to take the runqueue lock,
1243 * because all it wants to ensure is that the remote task enters
1244 * the kernel. If the IPI races and the task has been migrated
1245 * to another CPU then no harm is done and the purpose has been
1248 void kick_process(struct task_struct *p)
1254 if ((cpu != smp_processor_id()) && task_curr(p))
1255 smp_send_reschedule(cpu);
1258 EXPORT_SYMBOL_GPL(kick_process);
1259 #endif /* CONFIG_SMP */
1263 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1265 static int select_fallback_rq(int cpu, struct task_struct *p)
1268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1272 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1275 /* Any allowed, online CPU? */
1276 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1277 if (dest_cpu < nr_cpu_ids)
1280 /* No more Mr. Nice Guy. */
1281 dest_cpu = cpuset_cpus_allowed_fallback(p);
1283 * Don't tell them about moving exiting tasks or
1284 * kernel threads (both mm NULL), since they never
1287 if (p->mm && printk_ratelimit()) {
1288 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1289 task_pid_nr(p), p->comm, cpu);
1296 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1299 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1301 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1304 * In order not to call set_task_cpu() on a blocking task we need
1305 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1308 * Since this is common to all placement strategies, this lives here.
1310 * [ this allows ->select_task() to simply return task_cpu(p) and
1311 * not worry about this generic constraint ]
1313 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1315 cpu = select_fallback_rq(task_cpu(p), p);
1320 static void update_avg(u64 *avg, u64 sample)
1322 s64 diff = sample - *avg;
1328 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1330 #ifdef CONFIG_SCHEDSTATS
1331 struct rq *rq = this_rq();
1334 int this_cpu = smp_processor_id();
1336 if (cpu == this_cpu) {
1337 schedstat_inc(rq, ttwu_local);
1338 schedstat_inc(p, se.statistics.nr_wakeups_local);
1340 struct sched_domain *sd;
1342 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1344 for_each_domain(this_cpu, sd) {
1345 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1346 schedstat_inc(sd, ttwu_wake_remote);
1353 if (wake_flags & WF_MIGRATED)
1354 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1356 #endif /* CONFIG_SMP */
1358 schedstat_inc(rq, ttwu_count);
1359 schedstat_inc(p, se.statistics.nr_wakeups);
1361 if (wake_flags & WF_SYNC)
1362 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1364 #endif /* CONFIG_SCHEDSTATS */
1367 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1369 activate_task(rq, p, en_flags);
1372 /* if a worker is waking up, notify workqueue */
1373 if (p->flags & PF_WQ_WORKER)
1374 wq_worker_waking_up(p, cpu_of(rq));
1378 * Mark the task runnable and perform wakeup-preemption.
1381 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1383 trace_sched_wakeup(p, true);
1384 check_preempt_curr(rq, p, wake_flags);
1386 p->state = TASK_RUNNING;
1388 if (p->sched_class->task_woken)
1389 p->sched_class->task_woken(rq, p);
1391 if (rq->idle_stamp) {
1392 u64 delta = rq->clock - rq->idle_stamp;
1393 u64 max = 2*sysctl_sched_migration_cost;
1398 update_avg(&rq->avg_idle, delta);
1405 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1408 if (p->sched_contributes_to_load)
1409 rq->nr_uninterruptible--;
1412 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1413 ttwu_do_wakeup(rq, p, wake_flags);
1417 * Called in case the task @p isn't fully descheduled from its runqueue,
1418 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1419 * since all we need to do is flip p->state to TASK_RUNNING, since
1420 * the task is still ->on_rq.
1422 static int ttwu_remote(struct task_struct *p, int wake_flags)
1427 rq = __task_rq_lock(p);
1429 ttwu_do_wakeup(rq, p, wake_flags);
1432 __task_rq_unlock(rq);
1438 static void sched_ttwu_pending(void)
1440 struct rq *rq = this_rq();
1441 struct llist_node *llist = llist_del_all(&rq->wake_list);
1442 struct task_struct *p;
1444 raw_spin_lock(&rq->lock);
1447 p = llist_entry(llist, struct task_struct, wake_entry);
1448 llist = llist_next(llist);
1449 ttwu_do_activate(rq, p, 0);
1452 raw_spin_unlock(&rq->lock);
1455 void scheduler_ipi(void)
1457 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1461 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1462 * traditionally all their work was done from the interrupt return
1463 * path. Now that we actually do some work, we need to make sure
1466 * Some archs already do call them, luckily irq_enter/exit nest
1469 * Arguably we should visit all archs and update all handlers,
1470 * however a fair share of IPIs are still resched only so this would
1471 * somewhat pessimize the simple resched case.
1474 sched_ttwu_pending();
1477 * Check if someone kicked us for doing the nohz idle load balance.
1479 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1480 this_rq()->idle_balance = 1;
1481 raise_softirq_irqoff(SCHED_SOFTIRQ);
1486 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1488 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1489 smp_send_reschedule(cpu);
1492 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1493 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1498 rq = __task_rq_lock(p);
1500 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1501 ttwu_do_wakeup(rq, p, wake_flags);
1504 __task_rq_unlock(rq);
1509 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1511 static inline int ttwu_share_cache(int this_cpu, int that_cpu)
1513 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1515 #endif /* CONFIG_SMP */
1517 static void ttwu_queue(struct task_struct *p, int cpu)
1519 struct rq *rq = cpu_rq(cpu);
1521 #if defined(CONFIG_SMP)
1522 if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
1523 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1524 ttwu_queue_remote(p, cpu);
1529 raw_spin_lock(&rq->lock);
1530 ttwu_do_activate(rq, p, 0);
1531 raw_spin_unlock(&rq->lock);
1535 * try_to_wake_up - wake up a thread
1536 * @p: the thread to be awakened
1537 * @state: the mask of task states that can be woken
1538 * @wake_flags: wake modifier flags (WF_*)
1540 * Put it on the run-queue if it's not already there. The "current"
1541 * thread is always on the run-queue (except when the actual
1542 * re-schedule is in progress), and as such you're allowed to do
1543 * the simpler "current->state = TASK_RUNNING" to mark yourself
1544 * runnable without the overhead of this.
1546 * Returns %true if @p was woken up, %false if it was already running
1547 * or @state didn't match @p's state.
1550 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1552 unsigned long flags;
1553 int cpu, success = 0;
1556 raw_spin_lock_irqsave(&p->pi_lock, flags);
1557 if (!(p->state & state))
1560 success = 1; /* we're going to change ->state */
1563 if (p->on_rq && ttwu_remote(p, wake_flags))
1568 * If the owning (remote) cpu is still in the middle of schedule() with
1569 * this task as prev, wait until its done referencing the task.
1572 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1574 * In case the architecture enables interrupts in
1575 * context_switch(), we cannot busy wait, since that
1576 * would lead to deadlocks when an interrupt hits and
1577 * tries to wake up @prev. So bail and do a complete
1580 if (ttwu_activate_remote(p, wake_flags))
1587 * Pairs with the smp_wmb() in finish_lock_switch().
1591 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1592 p->state = TASK_WAKING;
1594 if (p->sched_class->task_waking)
1595 p->sched_class->task_waking(p);
1597 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1598 if (task_cpu(p) != cpu) {
1599 wake_flags |= WF_MIGRATED;
1600 set_task_cpu(p, cpu);
1602 #endif /* CONFIG_SMP */
1606 ttwu_stat(p, cpu, wake_flags);
1608 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1614 * try_to_wake_up_local - try to wake up a local task with rq lock held
1615 * @p: the thread to be awakened
1617 * Put @p on the run-queue if it's not already there. The caller must
1618 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1621 static void try_to_wake_up_local(struct task_struct *p)
1623 struct rq *rq = task_rq(p);
1625 BUG_ON(rq != this_rq());
1626 BUG_ON(p == current);
1627 lockdep_assert_held(&rq->lock);
1629 if (!raw_spin_trylock(&p->pi_lock)) {
1630 raw_spin_unlock(&rq->lock);
1631 raw_spin_lock(&p->pi_lock);
1632 raw_spin_lock(&rq->lock);
1635 if (!(p->state & TASK_NORMAL))
1639 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1641 ttwu_do_wakeup(rq, p, 0);
1642 ttwu_stat(p, smp_processor_id(), 0);
1644 raw_spin_unlock(&p->pi_lock);
1648 * wake_up_process - Wake up a specific process
1649 * @p: The process to be woken up.
1651 * Attempt to wake up the nominated process and move it to the set of runnable
1652 * processes. Returns 1 if the process was woken up, 0 if it was already
1655 * It may be assumed that this function implies a write memory barrier before
1656 * changing the task state if and only if any tasks are woken up.
1658 int wake_up_process(struct task_struct *p)
1660 return try_to_wake_up(p, TASK_ALL, 0);
1662 EXPORT_SYMBOL(wake_up_process);
1664 int wake_up_state(struct task_struct *p, unsigned int state)
1666 return try_to_wake_up(p, state, 0);
1670 * Perform scheduler related setup for a newly forked process p.
1671 * p is forked by current.
1673 * __sched_fork() is basic setup used by init_idle() too:
1675 static void __sched_fork(struct task_struct *p)
1680 p->se.exec_start = 0;
1681 p->se.sum_exec_runtime = 0;
1682 p->se.prev_sum_exec_runtime = 0;
1683 p->se.nr_migrations = 0;
1685 INIT_LIST_HEAD(&p->se.group_node);
1687 #ifdef CONFIG_SCHEDSTATS
1688 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1691 INIT_LIST_HEAD(&p->rt.run_list);
1693 #ifdef CONFIG_PREEMPT_NOTIFIERS
1694 INIT_HLIST_HEAD(&p->preempt_notifiers);
1699 * fork()/clone()-time setup:
1701 void sched_fork(struct task_struct *p)
1703 unsigned long flags;
1704 int cpu = get_cpu();
1708 * We mark the process as running here. This guarantees that
1709 * nobody will actually run it, and a signal or other external
1710 * event cannot wake it up and insert it on the runqueue either.
1712 p->state = TASK_RUNNING;
1715 * Make sure we do not leak PI boosting priority to the child.
1717 p->prio = current->normal_prio;
1720 * Revert to default priority/policy on fork if requested.
1722 if (unlikely(p->sched_reset_on_fork)) {
1723 if (task_has_rt_policy(p)) {
1724 p->policy = SCHED_NORMAL;
1725 p->static_prio = NICE_TO_PRIO(0);
1727 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1728 p->static_prio = NICE_TO_PRIO(0);
1730 p->prio = p->normal_prio = __normal_prio(p);
1734 * We don't need the reset flag anymore after the fork. It has
1735 * fulfilled its duty:
1737 p->sched_reset_on_fork = 0;
1740 if (!rt_prio(p->prio))
1741 p->sched_class = &fair_sched_class;
1743 if (p->sched_class->task_fork)
1744 p->sched_class->task_fork(p);
1747 * The child is not yet in the pid-hash so no cgroup attach races,
1748 * and the cgroup is pinned to this child due to cgroup_fork()
1749 * is ran before sched_fork().
1751 * Silence PROVE_RCU.
1753 raw_spin_lock_irqsave(&p->pi_lock, flags);
1754 set_task_cpu(p, cpu);
1755 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1757 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1758 if (likely(sched_info_on()))
1759 memset(&p->sched_info, 0, sizeof(p->sched_info));
1761 #if defined(CONFIG_SMP)
1764 #ifdef CONFIG_PREEMPT_COUNT
1765 /* Want to start with kernel preemption disabled. */
1766 task_thread_info(p)->preempt_count = 1;
1769 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1776 * wake_up_new_task - wake up a newly created task for the first time.
1778 * This function will do some initial scheduler statistics housekeeping
1779 * that must be done for every newly created context, then puts the task
1780 * on the runqueue and wakes it.
1782 void wake_up_new_task(struct task_struct *p)
1784 unsigned long flags;
1787 raw_spin_lock_irqsave(&p->pi_lock, flags);
1790 * Fork balancing, do it here and not earlier because:
1791 * - cpus_allowed can change in the fork path
1792 * - any previously selected cpu might disappear through hotplug
1794 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1797 rq = __task_rq_lock(p);
1798 activate_task(rq, p, 0);
1800 trace_sched_wakeup_new(p, true);
1801 check_preempt_curr(rq, p, WF_FORK);
1803 if (p->sched_class->task_woken)
1804 p->sched_class->task_woken(rq, p);
1806 task_rq_unlock(rq, p, &flags);
1809 #ifdef CONFIG_PREEMPT_NOTIFIERS
1812 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1813 * @notifier: notifier struct to register
1815 void preempt_notifier_register(struct preempt_notifier *notifier)
1817 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1819 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1822 * preempt_notifier_unregister - no longer interested in preemption notifications
1823 * @notifier: notifier struct to unregister
1825 * This is safe to call from within a preemption notifier.
1827 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1829 hlist_del(¬ifier->link);
1831 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1833 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1835 struct preempt_notifier *notifier;
1836 struct hlist_node *node;
1838 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1839 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1843 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1844 struct task_struct *next)
1846 struct preempt_notifier *notifier;
1847 struct hlist_node *node;
1849 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1850 notifier->ops->sched_out(notifier, next);
1853 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1855 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1860 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1861 struct task_struct *next)
1865 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1868 * prepare_task_switch - prepare to switch tasks
1869 * @rq: the runqueue preparing to switch
1870 * @prev: the current task that is being switched out
1871 * @next: the task we are going to switch to.
1873 * This is called with the rq lock held and interrupts off. It must
1874 * be paired with a subsequent finish_task_switch after the context
1877 * prepare_task_switch sets up locking and calls architecture specific
1881 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1882 struct task_struct *next)
1884 sched_info_switch(prev, next);
1885 perf_event_task_sched_out(prev, next);
1886 fire_sched_out_preempt_notifiers(prev, next);
1887 prepare_lock_switch(rq, next);
1888 prepare_arch_switch(next);
1889 trace_sched_switch(prev, next);
1893 * finish_task_switch - clean up after a task-switch
1894 * @rq: runqueue associated with task-switch
1895 * @prev: the thread we just switched away from.
1897 * finish_task_switch must be called after the context switch, paired
1898 * with a prepare_task_switch call before the context switch.
1899 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1900 * and do any other architecture-specific cleanup actions.
1902 * Note that we may have delayed dropping an mm in context_switch(). If
1903 * so, we finish that here outside of the runqueue lock. (Doing it
1904 * with the lock held can cause deadlocks; see schedule() for
1907 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1908 __releases(rq->lock)
1910 struct mm_struct *mm = rq->prev_mm;
1916 * A task struct has one reference for the use as "current".
1917 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1918 * schedule one last time. The schedule call will never return, and
1919 * the scheduled task must drop that reference.
1920 * The test for TASK_DEAD must occur while the runqueue locks are
1921 * still held, otherwise prev could be scheduled on another cpu, die
1922 * there before we look at prev->state, and then the reference would
1924 * Manfred Spraul <manfred@colorfullife.com>
1926 prev_state = prev->state;
1927 finish_arch_switch(prev);
1928 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1929 local_irq_disable();
1930 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1931 perf_event_task_sched_in(prev, current);
1932 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1934 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1935 finish_lock_switch(rq, prev);
1936 trace_sched_stat_sleeptime(current, rq->clock);
1938 fire_sched_in_preempt_notifiers(current);
1941 if (unlikely(prev_state == TASK_DEAD)) {
1943 * Remove function-return probe instances associated with this
1944 * task and put them back on the free list.
1946 kprobe_flush_task(prev);
1947 put_task_struct(prev);
1953 /* assumes rq->lock is held */
1954 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1956 if (prev->sched_class->pre_schedule)
1957 prev->sched_class->pre_schedule(rq, prev);
1960 /* rq->lock is NOT held, but preemption is disabled */
1961 static inline void post_schedule(struct rq *rq)
1963 if (rq->post_schedule) {
1964 unsigned long flags;
1966 raw_spin_lock_irqsave(&rq->lock, flags);
1967 if (rq->curr->sched_class->post_schedule)
1968 rq->curr->sched_class->post_schedule(rq);
1969 raw_spin_unlock_irqrestore(&rq->lock, flags);
1971 rq->post_schedule = 0;
1977 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1981 static inline void post_schedule(struct rq *rq)
1988 * schedule_tail - first thing a freshly forked thread must call.
1989 * @prev: the thread we just switched away from.
1991 asmlinkage void schedule_tail(struct task_struct *prev)
1992 __releases(rq->lock)
1994 struct rq *rq = this_rq();
1996 finish_task_switch(rq, prev);
1999 * FIXME: do we need to worry about rq being invalidated by the
2004 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2005 /* In this case, finish_task_switch does not reenable preemption */
2008 if (current->set_child_tid)
2009 put_user(task_pid_vnr(current), current->set_child_tid);
2013 * context_switch - switch to the new MM and the new
2014 * thread's register state.
2017 context_switch(struct rq *rq, struct task_struct *prev,
2018 struct task_struct *next)
2020 struct mm_struct *mm, *oldmm;
2022 prepare_task_switch(rq, prev, next);
2025 oldmm = prev->active_mm;
2027 * For paravirt, this is coupled with an exit in switch_to to
2028 * combine the page table reload and the switch backend into
2031 arch_start_context_switch(prev);
2034 next->active_mm = oldmm;
2035 atomic_inc(&oldmm->mm_count);
2036 enter_lazy_tlb(oldmm, next);
2038 switch_mm(oldmm, mm, next);
2041 prev->active_mm = NULL;
2042 rq->prev_mm = oldmm;
2045 * Since the runqueue lock will be released by the next
2046 * task (which is an invalid locking op but in the case
2047 * of the scheduler it's an obvious special-case), so we
2048 * do an early lockdep release here:
2050 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2051 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2054 /* Here we just switch the register state and the stack. */
2055 switch_to(prev, next, prev);
2059 * this_rq must be evaluated again because prev may have moved
2060 * CPUs since it called schedule(), thus the 'rq' on its stack
2061 * frame will be invalid.
2063 finish_task_switch(this_rq(), prev);
2067 * nr_running, nr_uninterruptible and nr_context_switches:
2069 * externally visible scheduler statistics: current number of runnable
2070 * threads, current number of uninterruptible-sleeping threads, total
2071 * number of context switches performed since bootup.
2073 unsigned long nr_running(void)
2075 unsigned long i, sum = 0;
2077 for_each_online_cpu(i)
2078 sum += cpu_rq(i)->nr_running;
2083 unsigned long nr_uninterruptible(void)
2085 unsigned long i, sum = 0;
2087 for_each_possible_cpu(i)
2088 sum += cpu_rq(i)->nr_uninterruptible;
2091 * Since we read the counters lockless, it might be slightly
2092 * inaccurate. Do not allow it to go below zero though:
2094 if (unlikely((long)sum < 0))
2100 unsigned long long nr_context_switches(void)
2103 unsigned long long sum = 0;
2105 for_each_possible_cpu(i)
2106 sum += cpu_rq(i)->nr_switches;
2111 unsigned long nr_iowait(void)
2113 unsigned long i, sum = 0;
2115 for_each_possible_cpu(i)
2116 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2121 unsigned long nr_iowait_cpu(int cpu)
2123 struct rq *this = cpu_rq(cpu);
2124 return atomic_read(&this->nr_iowait);
2127 unsigned long this_cpu_load(void)
2129 struct rq *this = this_rq();
2130 return this->cpu_load[0];
2134 /* Variables and functions for calc_load */
2135 static atomic_long_t calc_load_tasks;
2136 static unsigned long calc_load_update;
2137 unsigned long avenrun[3];
2138 EXPORT_SYMBOL(avenrun);
2140 static long calc_load_fold_active(struct rq *this_rq)
2142 long nr_active, delta = 0;
2144 nr_active = this_rq->nr_running;
2145 nr_active += (long) this_rq->nr_uninterruptible;
2147 if (nr_active != this_rq->calc_load_active) {
2148 delta = nr_active - this_rq->calc_load_active;
2149 this_rq->calc_load_active = nr_active;
2155 static unsigned long
2156 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2159 load += active * (FIXED_1 - exp);
2160 load += 1UL << (FSHIFT - 1);
2161 return load >> FSHIFT;
2166 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2168 * When making the ILB scale, we should try to pull this in as well.
2170 static atomic_long_t calc_load_tasks_idle;
2172 void calc_load_account_idle(struct rq *this_rq)
2176 delta = calc_load_fold_active(this_rq);
2178 atomic_long_add(delta, &calc_load_tasks_idle);
2181 static long calc_load_fold_idle(void)
2186 * Its got a race, we don't care...
2188 if (atomic_long_read(&calc_load_tasks_idle))
2189 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2195 * fixed_power_int - compute: x^n, in O(log n) time
2197 * @x: base of the power
2198 * @frac_bits: fractional bits of @x
2199 * @n: power to raise @x to.
2201 * By exploiting the relation between the definition of the natural power
2202 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2203 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2204 * (where: n_i \elem {0, 1}, the binary vector representing n),
2205 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2206 * of course trivially computable in O(log_2 n), the length of our binary
2209 static unsigned long
2210 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2212 unsigned long result = 1UL << frac_bits;
2217 result += 1UL << (frac_bits - 1);
2218 result >>= frac_bits;
2224 x += 1UL << (frac_bits - 1);
2232 * a1 = a0 * e + a * (1 - e)
2234 * a2 = a1 * e + a * (1 - e)
2235 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2236 * = a0 * e^2 + a * (1 - e) * (1 + e)
2238 * a3 = a2 * e + a * (1 - e)
2239 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2240 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2244 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2245 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2246 * = a0 * e^n + a * (1 - e^n)
2248 * [1] application of the geometric series:
2251 * S_n := \Sum x^i = -------------
2254 static unsigned long
2255 calc_load_n(unsigned long load, unsigned long exp,
2256 unsigned long active, unsigned int n)
2259 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2263 * NO_HZ can leave us missing all per-cpu ticks calling
2264 * calc_load_account_active(), but since an idle CPU folds its delta into
2265 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2266 * in the pending idle delta if our idle period crossed a load cycle boundary.
2268 * Once we've updated the global active value, we need to apply the exponential
2269 * weights adjusted to the number of cycles missed.
2271 static void calc_global_nohz(unsigned long ticks)
2273 long delta, active, n;
2275 if (time_before(jiffies, calc_load_update))
2279 * If we crossed a calc_load_update boundary, make sure to fold
2280 * any pending idle changes, the respective CPUs might have
2281 * missed the tick driven calc_load_account_active() update
2284 delta = calc_load_fold_idle();
2286 atomic_long_add(delta, &calc_load_tasks);
2289 * If we were idle for multiple load cycles, apply them.
2291 if (ticks >= LOAD_FREQ) {
2292 n = ticks / LOAD_FREQ;
2294 active = atomic_long_read(&calc_load_tasks);
2295 active = active > 0 ? active * FIXED_1 : 0;
2297 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2298 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2299 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2301 calc_load_update += n * LOAD_FREQ;
2305 * Its possible the remainder of the above division also crosses
2306 * a LOAD_FREQ period, the regular check in calc_global_load()
2307 * which comes after this will take care of that.
2309 * Consider us being 11 ticks before a cycle completion, and us
2310 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2311 * age us 4 cycles, and the test in calc_global_load() will
2312 * pick up the final one.
2316 void calc_load_account_idle(struct rq *this_rq)
2320 static inline long calc_load_fold_idle(void)
2325 static void calc_global_nohz(unsigned long ticks)
2331 * get_avenrun - get the load average array
2332 * @loads: pointer to dest load array
2333 * @offset: offset to add
2334 * @shift: shift count to shift the result left
2336 * These values are estimates at best, so no need for locking.
2338 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2340 loads[0] = (avenrun[0] + offset) << shift;
2341 loads[1] = (avenrun[1] + offset) << shift;
2342 loads[2] = (avenrun[2] + offset) << shift;
2346 * calc_load - update the avenrun load estimates 10 ticks after the
2347 * CPUs have updated calc_load_tasks.
2349 void calc_global_load(unsigned long ticks)
2353 calc_global_nohz(ticks);
2355 if (time_before(jiffies, calc_load_update + 10))
2358 active = atomic_long_read(&calc_load_tasks);
2359 active = active > 0 ? active * FIXED_1 : 0;
2361 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2362 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2363 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2365 calc_load_update += LOAD_FREQ;
2369 * Called from update_cpu_load() to periodically update this CPU's
2372 static void calc_load_account_active(struct rq *this_rq)
2376 if (time_before(jiffies, this_rq->calc_load_update))
2379 delta = calc_load_fold_active(this_rq);
2380 delta += calc_load_fold_idle();
2382 atomic_long_add(delta, &calc_load_tasks);
2384 this_rq->calc_load_update += LOAD_FREQ;
2388 * The exact cpuload at various idx values, calculated at every tick would be
2389 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2391 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2392 * on nth tick when cpu may be busy, then we have:
2393 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2394 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2396 * decay_load_missed() below does efficient calculation of
2397 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2398 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2400 * The calculation is approximated on a 128 point scale.
2401 * degrade_zero_ticks is the number of ticks after which load at any
2402 * particular idx is approximated to be zero.
2403 * degrade_factor is a precomputed table, a row for each load idx.
2404 * Each column corresponds to degradation factor for a power of two ticks,
2405 * based on 128 point scale.
2407 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2408 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2410 * With this power of 2 load factors, we can degrade the load n times
2411 * by looking at 1 bits in n and doing as many mult/shift instead of
2412 * n mult/shifts needed by the exact degradation.
2414 #define DEGRADE_SHIFT 7
2415 static const unsigned char
2416 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2417 static const unsigned char
2418 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2419 {0, 0, 0, 0, 0, 0, 0, 0},
2420 {64, 32, 8, 0, 0, 0, 0, 0},
2421 {96, 72, 40, 12, 1, 0, 0},
2422 {112, 98, 75, 43, 15, 1, 0},
2423 {120, 112, 98, 76, 45, 16, 2} };
2426 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2427 * would be when CPU is idle and so we just decay the old load without
2428 * adding any new load.
2430 static unsigned long
2431 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2435 if (!missed_updates)
2438 if (missed_updates >= degrade_zero_ticks[idx])
2442 return load >> missed_updates;
2444 while (missed_updates) {
2445 if (missed_updates % 2)
2446 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2448 missed_updates >>= 1;
2455 * Update rq->cpu_load[] statistics. This function is usually called every
2456 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2457 * every tick. We fix it up based on jiffies.
2459 void update_cpu_load(struct rq *this_rq)
2461 unsigned long this_load = this_rq->load.weight;
2462 unsigned long curr_jiffies = jiffies;
2463 unsigned long pending_updates;
2466 this_rq->nr_load_updates++;
2468 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2469 if (curr_jiffies == this_rq->last_load_update_tick)
2472 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2473 this_rq->last_load_update_tick = curr_jiffies;
2475 /* Update our load: */
2476 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2477 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2478 unsigned long old_load, new_load;
2480 /* scale is effectively 1 << i now, and >> i divides by scale */
2482 old_load = this_rq->cpu_load[i];
2483 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2484 new_load = this_load;
2486 * Round up the averaging division if load is increasing. This
2487 * prevents us from getting stuck on 9 if the load is 10, for
2490 if (new_load > old_load)
2491 new_load += scale - 1;
2493 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2496 sched_avg_update(this_rq);
2499 static void update_cpu_load_active(struct rq *this_rq)
2501 update_cpu_load(this_rq);
2503 calc_load_account_active(this_rq);
2509 * sched_exec - execve() is a valuable balancing opportunity, because at
2510 * this point the task has the smallest effective memory and cache footprint.
2512 void sched_exec(void)
2514 struct task_struct *p = current;
2515 unsigned long flags;
2518 raw_spin_lock_irqsave(&p->pi_lock, flags);
2519 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2520 if (dest_cpu == smp_processor_id())
2523 if (likely(cpu_active(dest_cpu))) {
2524 struct migration_arg arg = { p, dest_cpu };
2526 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2527 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2531 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2536 DEFINE_PER_CPU(struct kernel_stat, kstat);
2537 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2539 EXPORT_PER_CPU_SYMBOL(kstat);
2540 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2543 * Return any ns on the sched_clock that have not yet been accounted in
2544 * @p in case that task is currently running.
2546 * Called with task_rq_lock() held on @rq.
2548 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2552 if (task_current(rq, p)) {
2553 update_rq_clock(rq);
2554 ns = rq->clock_task - p->se.exec_start;
2562 unsigned long long task_delta_exec(struct task_struct *p)
2564 unsigned long flags;
2568 rq = task_rq_lock(p, &flags);
2569 ns = do_task_delta_exec(p, rq);
2570 task_rq_unlock(rq, p, &flags);
2576 * Return accounted runtime for the task.
2577 * In case the task is currently running, return the runtime plus current's
2578 * pending runtime that have not been accounted yet.
2580 unsigned long long task_sched_runtime(struct task_struct *p)
2582 unsigned long flags;
2586 rq = task_rq_lock(p, &flags);
2587 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2588 task_rq_unlock(rq, p, &flags);
2593 #ifdef CONFIG_CGROUP_CPUACCT
2594 struct cgroup_subsys cpuacct_subsys;
2595 struct cpuacct root_cpuacct;
2598 static inline void task_group_account_field(struct task_struct *p, int index,
2601 #ifdef CONFIG_CGROUP_CPUACCT
2602 struct kernel_cpustat *kcpustat;
2606 * Since all updates are sure to touch the root cgroup, we
2607 * get ourselves ahead and touch it first. If the root cgroup
2608 * is the only cgroup, then nothing else should be necessary.
2611 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2613 #ifdef CONFIG_CGROUP_CPUACCT
2614 if (unlikely(!cpuacct_subsys.active))
2619 while (ca && (ca != &root_cpuacct)) {
2620 kcpustat = this_cpu_ptr(ca->cpustat);
2621 kcpustat->cpustat[index] += tmp;
2630 * Account user cpu time to a process.
2631 * @p: the process that the cpu time gets accounted to
2632 * @cputime: the cpu time spent in user space since the last update
2633 * @cputime_scaled: cputime scaled by cpu frequency
2635 void account_user_time(struct task_struct *p, cputime_t cputime,
2636 cputime_t cputime_scaled)
2640 /* Add user time to process. */
2641 p->utime += cputime;
2642 p->utimescaled += cputime_scaled;
2643 account_group_user_time(p, cputime);
2645 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2647 /* Add user time to cpustat. */
2648 task_group_account_field(p, index, (__force u64) cputime);
2650 /* Account for user time used */
2651 acct_update_integrals(p);
2655 * Account guest cpu time to a process.
2656 * @p: the process that the cpu time gets accounted to
2657 * @cputime: the cpu time spent in virtual machine since the last update
2658 * @cputime_scaled: cputime scaled by cpu frequency
2660 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2661 cputime_t cputime_scaled)
2663 u64 *cpustat = kcpustat_this_cpu->cpustat;
2665 /* Add guest time to process. */
2666 p->utime += cputime;
2667 p->utimescaled += cputime_scaled;
2668 account_group_user_time(p, cputime);
2669 p->gtime += cputime;
2671 /* Add guest time to cpustat. */
2672 if (TASK_NICE(p) > 0) {
2673 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2674 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2676 cpustat[CPUTIME_USER] += (__force u64) cputime;
2677 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2682 * Account system cpu time to a process and desired cpustat field
2683 * @p: the process that the cpu time gets accounted to
2684 * @cputime: the cpu time spent in kernel space since the last update
2685 * @cputime_scaled: cputime scaled by cpu frequency
2686 * @target_cputime64: pointer to cpustat field that has to be updated
2689 void __account_system_time(struct task_struct *p, cputime_t cputime,
2690 cputime_t cputime_scaled, int index)
2692 /* Add system time to process. */
2693 p->stime += cputime;
2694 p->stimescaled += cputime_scaled;
2695 account_group_system_time(p, cputime);
2697 /* Add system time to cpustat. */
2698 task_group_account_field(p, index, (__force u64) cputime);
2700 /* Account for system time used */
2701 acct_update_integrals(p);
2705 * Account system cpu time to a process.
2706 * @p: the process that the cpu time gets accounted to
2707 * @hardirq_offset: the offset to subtract from hardirq_count()
2708 * @cputime: the cpu time spent in kernel space since the last update
2709 * @cputime_scaled: cputime scaled by cpu frequency
2711 void account_system_time(struct task_struct *p, int hardirq_offset,
2712 cputime_t cputime, cputime_t cputime_scaled)
2716 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2717 account_guest_time(p, cputime, cputime_scaled);
2721 if (hardirq_count() - hardirq_offset)
2722 index = CPUTIME_IRQ;
2723 else if (in_serving_softirq())
2724 index = CPUTIME_SOFTIRQ;
2726 index = CPUTIME_SYSTEM;
2728 __account_system_time(p, cputime, cputime_scaled, index);
2732 * Account for involuntary wait time.
2733 * @cputime: the cpu time spent in involuntary wait
2735 void account_steal_time(cputime_t cputime)
2737 u64 *cpustat = kcpustat_this_cpu->cpustat;
2739 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2743 * Account for idle time.
2744 * @cputime: the cpu time spent in idle wait
2746 void account_idle_time(cputime_t cputime)
2748 u64 *cpustat = kcpustat_this_cpu->cpustat;
2749 struct rq *rq = this_rq();
2751 if (atomic_read(&rq->nr_iowait) > 0)
2752 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2754 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2757 static __always_inline bool steal_account_process_tick(void)
2759 #ifdef CONFIG_PARAVIRT
2760 if (static_branch(¶virt_steal_enabled)) {
2763 steal = paravirt_steal_clock(smp_processor_id());
2764 steal -= this_rq()->prev_steal_time;
2766 st = steal_ticks(steal);
2767 this_rq()->prev_steal_time += st * TICK_NSEC;
2769 account_steal_time(st);
2776 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2778 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2780 * Account a tick to a process and cpustat
2781 * @p: the process that the cpu time gets accounted to
2782 * @user_tick: is the tick from userspace
2783 * @rq: the pointer to rq
2785 * Tick demultiplexing follows the order
2786 * - pending hardirq update
2787 * - pending softirq update
2791 * - check for guest_time
2792 * - else account as system_time
2794 * Check for hardirq is done both for system and user time as there is
2795 * no timer going off while we are on hardirq and hence we may never get an
2796 * opportunity to update it solely in system time.
2797 * p->stime and friends are only updated on system time and not on irq
2798 * softirq as those do not count in task exec_runtime any more.
2800 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2803 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2804 u64 *cpustat = kcpustat_this_cpu->cpustat;
2806 if (steal_account_process_tick())
2809 if (irqtime_account_hi_update()) {
2810 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2811 } else if (irqtime_account_si_update()) {
2812 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2813 } else if (this_cpu_ksoftirqd() == p) {
2815 * ksoftirqd time do not get accounted in cpu_softirq_time.
2816 * So, we have to handle it separately here.
2817 * Also, p->stime needs to be updated for ksoftirqd.
2819 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2821 } else if (user_tick) {
2822 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2823 } else if (p == rq->idle) {
2824 account_idle_time(cputime_one_jiffy);
2825 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2826 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2828 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2833 static void irqtime_account_idle_ticks(int ticks)
2836 struct rq *rq = this_rq();
2838 for (i = 0; i < ticks; i++)
2839 irqtime_account_process_tick(current, 0, rq);
2841 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2842 static void irqtime_account_idle_ticks(int ticks) {}
2843 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2845 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2848 * Account a single tick of cpu time.
2849 * @p: the process that the cpu time gets accounted to
2850 * @user_tick: indicates if the tick is a user or a system tick
2852 void account_process_tick(struct task_struct *p, int user_tick)
2854 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2855 struct rq *rq = this_rq();
2857 if (sched_clock_irqtime) {
2858 irqtime_account_process_tick(p, user_tick, rq);
2862 if (steal_account_process_tick())
2866 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2867 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2868 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2871 account_idle_time(cputime_one_jiffy);
2875 * Account multiple ticks of steal time.
2876 * @p: the process from which the cpu time has been stolen
2877 * @ticks: number of stolen ticks
2879 void account_steal_ticks(unsigned long ticks)
2881 account_steal_time(jiffies_to_cputime(ticks));
2885 * Account multiple ticks of idle time.
2886 * @ticks: number of stolen ticks
2888 void account_idle_ticks(unsigned long ticks)
2891 if (sched_clock_irqtime) {
2892 irqtime_account_idle_ticks(ticks);
2896 account_idle_time(jiffies_to_cputime(ticks));
2902 * Use precise platform statistics if available:
2904 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2905 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2911 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2913 struct task_cputime cputime;
2915 thread_group_cputime(p, &cputime);
2917 *ut = cputime.utime;
2918 *st = cputime.stime;
2922 #ifndef nsecs_to_cputime
2923 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2926 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2928 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2931 * Use CFS's precise accounting:
2933 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2936 u64 temp = (__force u64) rtime;
2938 temp *= (__force u64) utime;
2939 do_div(temp, (__force u32) total);
2940 utime = (__force cputime_t) temp;
2945 * Compare with previous values, to keep monotonicity:
2947 p->prev_utime = max(p->prev_utime, utime);
2948 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2950 *ut = p->prev_utime;
2951 *st = p->prev_stime;
2955 * Must be called with siglock held.
2957 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2959 struct signal_struct *sig = p->signal;
2960 struct task_cputime cputime;
2961 cputime_t rtime, utime, total;
2963 thread_group_cputime(p, &cputime);
2965 total = cputime.utime + cputime.stime;
2966 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2969 u64 temp = (__force u64) rtime;
2971 temp *= (__force u64) cputime.utime;
2972 do_div(temp, (__force u32) total);
2973 utime = (__force cputime_t) temp;
2977 sig->prev_utime = max(sig->prev_utime, utime);
2978 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
2980 *ut = sig->prev_utime;
2981 *st = sig->prev_stime;
2986 * This function gets called by the timer code, with HZ frequency.
2987 * We call it with interrupts disabled.
2989 void scheduler_tick(void)
2991 int cpu = smp_processor_id();
2992 struct rq *rq = cpu_rq(cpu);
2993 struct task_struct *curr = rq->curr;
2997 raw_spin_lock(&rq->lock);
2998 update_rq_clock(rq);
2999 update_cpu_load_active(rq);
3000 curr->sched_class->task_tick(rq, curr, 0);
3001 raw_spin_unlock(&rq->lock);
3003 perf_event_task_tick();
3006 rq->idle_balance = idle_cpu(cpu);
3007 trigger_load_balance(rq, cpu);
3011 notrace unsigned long get_parent_ip(unsigned long addr)
3013 if (in_lock_functions(addr)) {
3014 addr = CALLER_ADDR2;
3015 if (in_lock_functions(addr))
3016 addr = CALLER_ADDR3;
3021 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3022 defined(CONFIG_PREEMPT_TRACER))
3024 void __kprobes add_preempt_count(int val)
3026 #ifdef CONFIG_DEBUG_PREEMPT
3030 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3033 preempt_count() += val;
3034 #ifdef CONFIG_DEBUG_PREEMPT
3036 * Spinlock count overflowing soon?
3038 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3041 if (preempt_count() == val)
3042 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3044 EXPORT_SYMBOL(add_preempt_count);
3046 void __kprobes sub_preempt_count(int val)
3048 #ifdef CONFIG_DEBUG_PREEMPT
3052 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3055 * Is the spinlock portion underflowing?
3057 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3058 !(preempt_count() & PREEMPT_MASK)))
3062 if (preempt_count() == val)
3063 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3064 preempt_count() -= val;
3066 EXPORT_SYMBOL(sub_preempt_count);
3071 * Print scheduling while atomic bug:
3073 static noinline void __schedule_bug(struct task_struct *prev)
3075 struct pt_regs *regs = get_irq_regs();
3077 if (oops_in_progress)
3080 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3081 prev->comm, prev->pid, preempt_count());
3083 debug_show_held_locks(prev);
3085 if (irqs_disabled())
3086 print_irqtrace_events(prev);
3095 * Various schedule()-time debugging checks and statistics:
3097 static inline void schedule_debug(struct task_struct *prev)
3100 * Test if we are atomic. Since do_exit() needs to call into
3101 * schedule() atomically, we ignore that path for now.
3102 * Otherwise, whine if we are scheduling when we should not be.
3104 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3105 __schedule_bug(prev);
3108 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3110 schedstat_inc(this_rq(), sched_count);
3113 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3115 if (prev->on_rq || rq->skip_clock_update < 0)
3116 update_rq_clock(rq);
3117 prev->sched_class->put_prev_task(rq, prev);
3121 * Pick up the highest-prio task:
3123 static inline struct task_struct *
3124 pick_next_task(struct rq *rq)
3126 const struct sched_class *class;
3127 struct task_struct *p;
3130 * Optimization: we know that if all tasks are in
3131 * the fair class we can call that function directly:
3133 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3134 p = fair_sched_class.pick_next_task(rq);
3139 for_each_class(class) {
3140 p = class->pick_next_task(rq);
3145 BUG(); /* the idle class will always have a runnable task */
3149 * __schedule() is the main scheduler function.
3151 static void __sched __schedule(void)
3153 struct task_struct *prev, *next;
3154 unsigned long *switch_count;
3160 cpu = smp_processor_id();
3162 rcu_note_context_switch(cpu);
3165 schedule_debug(prev);
3167 if (sched_feat(HRTICK))
3170 raw_spin_lock_irq(&rq->lock);
3172 switch_count = &prev->nivcsw;
3173 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3174 if (unlikely(signal_pending_state(prev->state, prev))) {
3175 prev->state = TASK_RUNNING;
3177 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3181 * If a worker went to sleep, notify and ask workqueue
3182 * whether it wants to wake up a task to maintain
3185 if (prev->flags & PF_WQ_WORKER) {
3186 struct task_struct *to_wakeup;
3188 to_wakeup = wq_worker_sleeping(prev, cpu);
3190 try_to_wake_up_local(to_wakeup);
3193 switch_count = &prev->nvcsw;
3196 pre_schedule(rq, prev);
3198 if (unlikely(!rq->nr_running))
3199 idle_balance(cpu, rq);
3201 put_prev_task(rq, prev);
3202 next = pick_next_task(rq);
3203 clear_tsk_need_resched(prev);
3204 rq->skip_clock_update = 0;
3206 if (likely(prev != next)) {
3211 context_switch(rq, prev, next); /* unlocks the rq */
3213 * The context switch have flipped the stack from under us
3214 * and restored the local variables which were saved when
3215 * this task called schedule() in the past. prev == current
3216 * is still correct, but it can be moved to another cpu/rq.
3218 cpu = smp_processor_id();
3221 raw_spin_unlock_irq(&rq->lock);
3225 preempt_enable_no_resched();
3230 static inline void sched_submit_work(struct task_struct *tsk)
3235 * If we are going to sleep and we have plugged IO queued,
3236 * make sure to submit it to avoid deadlocks.
3238 if (blk_needs_flush_plug(tsk))
3239 blk_schedule_flush_plug(tsk);
3242 asmlinkage void __sched schedule(void)
3244 struct task_struct *tsk = current;
3246 sched_submit_work(tsk);
3249 EXPORT_SYMBOL(schedule);
3251 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3253 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3255 if (lock->owner != owner)
3259 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3260 * lock->owner still matches owner, if that fails, owner might
3261 * point to free()d memory, if it still matches, the rcu_read_lock()
3262 * ensures the memory stays valid.
3266 return owner->on_cpu;
3270 * Look out! "owner" is an entirely speculative pointer
3271 * access and not reliable.
3273 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3275 if (!sched_feat(OWNER_SPIN))
3279 while (owner_running(lock, owner)) {
3283 arch_mutex_cpu_relax();
3288 * We break out the loop above on need_resched() and when the
3289 * owner changed, which is a sign for heavy contention. Return
3290 * success only when lock->owner is NULL.
3292 return lock->owner == NULL;
3296 #ifdef CONFIG_PREEMPT
3298 * this is the entry point to schedule() from in-kernel preemption
3299 * off of preempt_enable. Kernel preemptions off return from interrupt
3300 * occur there and call schedule directly.
3302 asmlinkage void __sched notrace preempt_schedule(void)
3304 struct thread_info *ti = current_thread_info();
3307 * If there is a non-zero preempt_count or interrupts are disabled,
3308 * we do not want to preempt the current task. Just return..
3310 if (likely(ti->preempt_count || irqs_disabled()))
3314 add_preempt_count_notrace(PREEMPT_ACTIVE);
3316 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3319 * Check again in case we missed a preemption opportunity
3320 * between schedule and now.
3323 } while (need_resched());
3325 EXPORT_SYMBOL(preempt_schedule);
3328 * this is the entry point to schedule() from kernel preemption
3329 * off of irq context.
3330 * Note, that this is called and return with irqs disabled. This will
3331 * protect us against recursive calling from irq.
3333 asmlinkage void __sched preempt_schedule_irq(void)
3335 struct thread_info *ti = current_thread_info();
3337 /* Catch callers which need to be fixed */
3338 BUG_ON(ti->preempt_count || !irqs_disabled());
3341 add_preempt_count(PREEMPT_ACTIVE);
3344 local_irq_disable();
3345 sub_preempt_count(PREEMPT_ACTIVE);
3348 * Check again in case we missed a preemption opportunity
3349 * between schedule and now.
3352 } while (need_resched());
3355 #endif /* CONFIG_PREEMPT */
3357 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3360 return try_to_wake_up(curr->private, mode, wake_flags);
3362 EXPORT_SYMBOL(default_wake_function);
3365 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3366 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3367 * number) then we wake all the non-exclusive tasks and one exclusive task.
3369 * There are circumstances in which we can try to wake a task which has already
3370 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3371 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3373 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3374 int nr_exclusive, int wake_flags, void *key)
3376 wait_queue_t *curr, *next;
3378 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3379 unsigned flags = curr->flags;
3381 if (curr->func(curr, mode, wake_flags, key) &&
3382 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3388 * __wake_up - wake up threads blocked on a waitqueue.
3390 * @mode: which threads
3391 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3392 * @key: is directly passed to the wakeup function
3394 * It may be assumed that this function implies a write memory barrier before
3395 * changing the task state if and only if any tasks are woken up.
3397 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3398 int nr_exclusive, void *key)
3400 unsigned long flags;
3402 spin_lock_irqsave(&q->lock, flags);
3403 __wake_up_common(q, mode, nr_exclusive, 0, key);
3404 spin_unlock_irqrestore(&q->lock, flags);
3406 EXPORT_SYMBOL(__wake_up);
3409 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3411 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3413 __wake_up_common(q, mode, 1, 0, NULL);
3415 EXPORT_SYMBOL_GPL(__wake_up_locked);
3417 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3419 __wake_up_common(q, mode, 1, 0, key);
3421 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3424 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3426 * @mode: which threads
3427 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3428 * @key: opaque value to be passed to wakeup targets
3430 * The sync wakeup differs that the waker knows that it will schedule
3431 * away soon, so while the target thread will be woken up, it will not
3432 * be migrated to another CPU - ie. the two threads are 'synchronized'
3433 * with each other. This can prevent needless bouncing between CPUs.
3435 * On UP it can prevent extra preemption.
3437 * It may be assumed that this function implies a write memory barrier before
3438 * changing the task state if and only if any tasks are woken up.
3440 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3441 int nr_exclusive, void *key)
3443 unsigned long flags;
3444 int wake_flags = WF_SYNC;
3449 if (unlikely(!nr_exclusive))
3452 spin_lock_irqsave(&q->lock, flags);
3453 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3454 spin_unlock_irqrestore(&q->lock, flags);
3456 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3459 * __wake_up_sync - see __wake_up_sync_key()
3461 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3463 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3465 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3468 * complete: - signals a single thread waiting on this completion
3469 * @x: holds the state of this particular completion
3471 * This will wake up a single thread waiting on this completion. Threads will be
3472 * awakened in the same order in which they were queued.
3474 * See also complete_all(), wait_for_completion() and related routines.
3476 * It may be assumed that this function implies a write memory barrier before
3477 * changing the task state if and only if any tasks are woken up.
3479 void complete(struct completion *x)
3481 unsigned long flags;
3483 spin_lock_irqsave(&x->wait.lock, flags);
3485 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3486 spin_unlock_irqrestore(&x->wait.lock, flags);
3488 EXPORT_SYMBOL(complete);
3491 * complete_all: - signals all threads waiting on this completion
3492 * @x: holds the state of this particular completion
3494 * This will wake up all threads waiting on this particular completion event.
3496 * It may be assumed that this function implies a write memory barrier before
3497 * changing the task state if and only if any tasks are woken up.
3499 void complete_all(struct completion *x)
3501 unsigned long flags;
3503 spin_lock_irqsave(&x->wait.lock, flags);
3504 x->done += UINT_MAX/2;
3505 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3506 spin_unlock_irqrestore(&x->wait.lock, flags);
3508 EXPORT_SYMBOL(complete_all);
3510 static inline long __sched
3511 do_wait_for_common(struct completion *x, long timeout, int state)
3514 DECLARE_WAITQUEUE(wait, current);
3516 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3518 if (signal_pending_state(state, current)) {
3519 timeout = -ERESTARTSYS;
3522 __set_current_state(state);
3523 spin_unlock_irq(&x->wait.lock);
3524 timeout = schedule_timeout(timeout);
3525 spin_lock_irq(&x->wait.lock);
3526 } while (!x->done && timeout);
3527 __remove_wait_queue(&x->wait, &wait);
3532 return timeout ?: 1;
3536 wait_for_common(struct completion *x, long timeout, int state)
3540 spin_lock_irq(&x->wait.lock);
3541 timeout = do_wait_for_common(x, timeout, state);
3542 spin_unlock_irq(&x->wait.lock);
3547 * wait_for_completion: - waits for completion of a task
3548 * @x: holds the state of this particular completion
3550 * This waits to be signaled for completion of a specific task. It is NOT
3551 * interruptible and there is no timeout.
3553 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3554 * and interrupt capability. Also see complete().
3556 void __sched wait_for_completion(struct completion *x)
3558 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3560 EXPORT_SYMBOL(wait_for_completion);
3563 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3564 * @x: holds the state of this particular completion
3565 * @timeout: timeout value in jiffies
3567 * This waits for either a completion of a specific task to be signaled or for a
3568 * specified timeout to expire. The timeout is in jiffies. It is not
3571 * The return value is 0 if timed out, and positive (at least 1, or number of
3572 * jiffies left till timeout) if completed.
3574 unsigned long __sched
3575 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3577 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3579 EXPORT_SYMBOL(wait_for_completion_timeout);
3582 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3583 * @x: holds the state of this particular completion
3585 * This waits for completion of a specific task to be signaled. It is
3588 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3590 int __sched wait_for_completion_interruptible(struct completion *x)
3592 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3593 if (t == -ERESTARTSYS)
3597 EXPORT_SYMBOL(wait_for_completion_interruptible);
3600 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3601 * @x: holds the state of this particular completion
3602 * @timeout: timeout value in jiffies
3604 * This waits for either a completion of a specific task to be signaled or for a
3605 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3607 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3608 * positive (at least 1, or number of jiffies left till timeout) if completed.
3611 wait_for_completion_interruptible_timeout(struct completion *x,
3612 unsigned long timeout)
3614 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3616 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3619 * wait_for_completion_killable: - waits for completion of a task (killable)
3620 * @x: holds the state of this particular completion
3622 * This waits to be signaled for completion of a specific task. It can be
3623 * interrupted by a kill signal.
3625 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3627 int __sched wait_for_completion_killable(struct completion *x)
3629 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3630 if (t == -ERESTARTSYS)
3634 EXPORT_SYMBOL(wait_for_completion_killable);
3637 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3638 * @x: holds the state of this particular completion
3639 * @timeout: timeout value in jiffies
3641 * This waits for either a completion of a specific task to be
3642 * signaled or for a specified timeout to expire. It can be
3643 * interrupted by a kill signal. The timeout is in jiffies.
3645 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3646 * positive (at least 1, or number of jiffies left till timeout) if completed.
3649 wait_for_completion_killable_timeout(struct completion *x,
3650 unsigned long timeout)
3652 return wait_for_common(x, timeout, TASK_KILLABLE);
3654 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3657 * try_wait_for_completion - try to decrement a completion without blocking
3658 * @x: completion structure
3660 * Returns: 0 if a decrement cannot be done without blocking
3661 * 1 if a decrement succeeded.
3663 * If a completion is being used as a counting completion,
3664 * attempt to decrement the counter without blocking. This
3665 * enables us to avoid waiting if the resource the completion
3666 * is protecting is not available.
3668 bool try_wait_for_completion(struct completion *x)
3670 unsigned long flags;
3673 spin_lock_irqsave(&x->wait.lock, flags);
3678 spin_unlock_irqrestore(&x->wait.lock, flags);
3681 EXPORT_SYMBOL(try_wait_for_completion);
3684 * completion_done - Test to see if a completion has any waiters
3685 * @x: completion structure
3687 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3688 * 1 if there are no waiters.
3691 bool completion_done(struct completion *x)
3693 unsigned long flags;
3696 spin_lock_irqsave(&x->wait.lock, flags);
3699 spin_unlock_irqrestore(&x->wait.lock, flags);
3702 EXPORT_SYMBOL(completion_done);
3705 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3707 unsigned long flags;
3710 init_waitqueue_entry(&wait, current);
3712 __set_current_state(state);
3714 spin_lock_irqsave(&q->lock, flags);
3715 __add_wait_queue(q, &wait);
3716 spin_unlock(&q->lock);
3717 timeout = schedule_timeout(timeout);
3718 spin_lock_irq(&q->lock);
3719 __remove_wait_queue(q, &wait);
3720 spin_unlock_irqrestore(&q->lock, flags);
3725 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3727 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3729 EXPORT_SYMBOL(interruptible_sleep_on);
3732 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3734 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3736 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3738 void __sched sleep_on(wait_queue_head_t *q)
3740 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3742 EXPORT_SYMBOL(sleep_on);
3744 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3746 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3748 EXPORT_SYMBOL(sleep_on_timeout);
3750 #ifdef CONFIG_RT_MUTEXES
3753 * rt_mutex_setprio - set the current priority of a task
3755 * @prio: prio value (kernel-internal form)
3757 * This function changes the 'effective' priority of a task. It does
3758 * not touch ->normal_prio like __setscheduler().
3760 * Used by the rt_mutex code to implement priority inheritance logic.
3762 void rt_mutex_setprio(struct task_struct *p, int prio)
3764 int oldprio, on_rq, running;
3766 const struct sched_class *prev_class;
3768 BUG_ON(prio < 0 || prio > MAX_PRIO);
3770 rq = __task_rq_lock(p);
3772 trace_sched_pi_setprio(p, prio);
3774 prev_class = p->sched_class;
3776 running = task_current(rq, p);
3778 dequeue_task(rq, p, 0);
3780 p->sched_class->put_prev_task(rq, p);
3783 p->sched_class = &rt_sched_class;
3785 p->sched_class = &fair_sched_class;
3790 p->sched_class->set_curr_task(rq);
3792 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3794 check_class_changed(rq, p, prev_class, oldprio);
3795 __task_rq_unlock(rq);
3800 void set_user_nice(struct task_struct *p, long nice)
3802 int old_prio, delta, on_rq;
3803 unsigned long flags;
3806 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3809 * We have to be careful, if called from sys_setpriority(),
3810 * the task might be in the middle of scheduling on another CPU.
3812 rq = task_rq_lock(p, &flags);
3814 * The RT priorities are set via sched_setscheduler(), but we still
3815 * allow the 'normal' nice value to be set - but as expected
3816 * it wont have any effect on scheduling until the task is
3817 * SCHED_FIFO/SCHED_RR:
3819 if (task_has_rt_policy(p)) {
3820 p->static_prio = NICE_TO_PRIO(nice);
3825 dequeue_task(rq, p, 0);
3827 p->static_prio = NICE_TO_PRIO(nice);
3830 p->prio = effective_prio(p);
3831 delta = p->prio - old_prio;
3834 enqueue_task(rq, p, 0);
3836 * If the task increased its priority or is running and
3837 * lowered its priority, then reschedule its CPU:
3839 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3840 resched_task(rq->curr);
3843 task_rq_unlock(rq, p, &flags);
3845 EXPORT_SYMBOL(set_user_nice);
3848 * can_nice - check if a task can reduce its nice value
3852 int can_nice(const struct task_struct *p, const int nice)
3854 /* convert nice value [19,-20] to rlimit style value [1,40] */
3855 int nice_rlim = 20 - nice;
3857 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3858 capable(CAP_SYS_NICE));
3861 #ifdef __ARCH_WANT_SYS_NICE
3864 * sys_nice - change the priority of the current process.
3865 * @increment: priority increment
3867 * sys_setpriority is a more generic, but much slower function that
3868 * does similar things.
3870 SYSCALL_DEFINE1(nice, int, increment)
3875 * Setpriority might change our priority at the same moment.
3876 * We don't have to worry. Conceptually one call occurs first
3877 * and we have a single winner.
3879 if (increment < -40)
3884 nice = TASK_NICE(current) + increment;
3890 if (increment < 0 && !can_nice(current, nice))
3893 retval = security_task_setnice(current, nice);
3897 set_user_nice(current, nice);
3904 * task_prio - return the priority value of a given task.
3905 * @p: the task in question.
3907 * This is the priority value as seen by users in /proc.
3908 * RT tasks are offset by -200. Normal tasks are centered
3909 * around 0, value goes from -16 to +15.
3911 int task_prio(const struct task_struct *p)
3913 return p->prio - MAX_RT_PRIO;
3917 * task_nice - return the nice value of a given task.
3918 * @p: the task in question.
3920 int task_nice(const struct task_struct *p)
3922 return TASK_NICE(p);
3924 EXPORT_SYMBOL(task_nice);
3927 * idle_cpu - is a given cpu idle currently?
3928 * @cpu: the processor in question.
3930 int idle_cpu(int cpu)
3932 struct rq *rq = cpu_rq(cpu);
3934 if (rq->curr != rq->idle)
3941 if (!llist_empty(&rq->wake_list))
3949 * idle_task - return the idle task for a given cpu.
3950 * @cpu: the processor in question.
3952 struct task_struct *idle_task(int cpu)
3954 return cpu_rq(cpu)->idle;
3958 * find_process_by_pid - find a process with a matching PID value.
3959 * @pid: the pid in question.
3961 static struct task_struct *find_process_by_pid(pid_t pid)
3963 return pid ? find_task_by_vpid(pid) : current;
3966 /* Actually do priority change: must hold rq lock. */
3968 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3971 p->rt_priority = prio;
3972 p->normal_prio = normal_prio(p);
3973 /* we are holding p->pi_lock already */
3974 p->prio = rt_mutex_getprio(p);
3975 if (rt_prio(p->prio))
3976 p->sched_class = &rt_sched_class;
3978 p->sched_class = &fair_sched_class;
3983 * check the target process has a UID that matches the current process's
3985 static bool check_same_owner(struct task_struct *p)
3987 const struct cred *cred = current_cred(), *pcred;
3991 pcred = __task_cred(p);
3992 if (cred->user->user_ns == pcred->user->user_ns)
3993 match = (cred->euid == pcred->euid ||
3994 cred->euid == pcred->uid);
4001 static int __sched_setscheduler(struct task_struct *p, int policy,
4002 const struct sched_param *param, bool user)
4004 int retval, oldprio, oldpolicy = -1, on_rq, running;
4005 unsigned long flags;
4006 const struct sched_class *prev_class;
4010 /* may grab non-irq protected spin_locks */
4011 BUG_ON(in_interrupt());
4013 /* double check policy once rq lock held */
4015 reset_on_fork = p->sched_reset_on_fork;
4016 policy = oldpolicy = p->policy;
4018 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4019 policy &= ~SCHED_RESET_ON_FORK;
4021 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4022 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4023 policy != SCHED_IDLE)
4028 * Valid priorities for SCHED_FIFO and SCHED_RR are
4029 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4030 * SCHED_BATCH and SCHED_IDLE is 0.
4032 if (param->sched_priority < 0 ||
4033 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4034 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4036 if (rt_policy(policy) != (param->sched_priority != 0))
4040 * Allow unprivileged RT tasks to decrease priority:
4042 if (user && !capable(CAP_SYS_NICE)) {
4043 if (rt_policy(policy)) {
4044 unsigned long rlim_rtprio =
4045 task_rlimit(p, RLIMIT_RTPRIO);
4047 /* can't set/change the rt policy */
4048 if (policy != p->policy && !rlim_rtprio)
4051 /* can't increase priority */
4052 if (param->sched_priority > p->rt_priority &&
4053 param->sched_priority > rlim_rtprio)
4058 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4059 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4061 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4062 if (!can_nice(p, TASK_NICE(p)))
4066 /* can't change other user's priorities */
4067 if (!check_same_owner(p))
4070 /* Normal users shall not reset the sched_reset_on_fork flag */
4071 if (p->sched_reset_on_fork && !reset_on_fork)
4076 retval = security_task_setscheduler(p);
4082 * make sure no PI-waiters arrive (or leave) while we are
4083 * changing the priority of the task:
4085 * To be able to change p->policy safely, the appropriate
4086 * runqueue lock must be held.
4088 rq = task_rq_lock(p, &flags);
4091 * Changing the policy of the stop threads its a very bad idea
4093 if (p == rq->stop) {
4094 task_rq_unlock(rq, p, &flags);
4099 * If not changing anything there's no need to proceed further:
4101 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4102 param->sched_priority == p->rt_priority))) {
4104 __task_rq_unlock(rq);
4105 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4109 #ifdef CONFIG_RT_GROUP_SCHED
4112 * Do not allow realtime tasks into groups that have no runtime
4115 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4116 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4117 !task_group_is_autogroup(task_group(p))) {
4118 task_rq_unlock(rq, p, &flags);
4124 /* recheck policy now with rq lock held */
4125 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4126 policy = oldpolicy = -1;
4127 task_rq_unlock(rq, p, &flags);
4131 running = task_current(rq, p);
4133 dequeue_task(rq, p, 0);
4135 p->sched_class->put_prev_task(rq, p);
4137 p->sched_reset_on_fork = reset_on_fork;
4140 prev_class = p->sched_class;
4141 __setscheduler(rq, p, policy, param->sched_priority);
4144 p->sched_class->set_curr_task(rq);
4146 enqueue_task(rq, p, 0);
4148 check_class_changed(rq, p, prev_class, oldprio);
4149 task_rq_unlock(rq, p, &flags);
4151 rt_mutex_adjust_pi(p);
4157 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4158 * @p: the task in question.
4159 * @policy: new policy.
4160 * @param: structure containing the new RT priority.
4162 * NOTE that the task may be already dead.
4164 int sched_setscheduler(struct task_struct *p, int policy,
4165 const struct sched_param *param)
4167 return __sched_setscheduler(p, policy, param, true);
4169 EXPORT_SYMBOL_GPL(sched_setscheduler);
4172 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4173 * @p: the task in question.
4174 * @policy: new policy.
4175 * @param: structure containing the new RT priority.
4177 * Just like sched_setscheduler, only don't bother checking if the
4178 * current context has permission. For example, this is needed in
4179 * stop_machine(): we create temporary high priority worker threads,
4180 * but our caller might not have that capability.
4182 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4183 const struct sched_param *param)
4185 return __sched_setscheduler(p, policy, param, false);
4189 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4191 struct sched_param lparam;
4192 struct task_struct *p;
4195 if (!param || pid < 0)
4197 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4202 p = find_process_by_pid(pid);
4204 retval = sched_setscheduler(p, policy, &lparam);
4211 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4212 * @pid: the pid in question.
4213 * @policy: new policy.
4214 * @param: structure containing the new RT priority.
4216 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4217 struct sched_param __user *, param)
4219 /* negative values for policy are not valid */
4223 return do_sched_setscheduler(pid, policy, param);
4227 * sys_sched_setparam - set/change the RT priority of a thread
4228 * @pid: the pid in question.
4229 * @param: structure containing the new RT priority.
4231 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4233 return do_sched_setscheduler(pid, -1, param);
4237 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4238 * @pid: the pid in question.
4240 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4242 struct task_struct *p;
4250 p = find_process_by_pid(pid);
4252 retval = security_task_getscheduler(p);
4255 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4262 * sys_sched_getparam - get the RT priority of a thread
4263 * @pid: the pid in question.
4264 * @param: structure containing the RT priority.
4266 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4268 struct sched_param lp;
4269 struct task_struct *p;
4272 if (!param || pid < 0)
4276 p = find_process_by_pid(pid);
4281 retval = security_task_getscheduler(p);
4285 lp.sched_priority = p->rt_priority;
4289 * This one might sleep, we cannot do it with a spinlock held ...
4291 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4300 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4302 cpumask_var_t cpus_allowed, new_mask;
4303 struct task_struct *p;
4309 p = find_process_by_pid(pid);
4316 /* Prevent p going away */
4320 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4324 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4326 goto out_free_cpus_allowed;
4329 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4332 retval = security_task_setscheduler(p);
4336 cpuset_cpus_allowed(p, cpus_allowed);
4337 cpumask_and(new_mask, in_mask, cpus_allowed);
4339 retval = set_cpus_allowed_ptr(p, new_mask);
4342 cpuset_cpus_allowed(p, cpus_allowed);
4343 if (!cpumask_subset(new_mask, cpus_allowed)) {
4345 * We must have raced with a concurrent cpuset
4346 * update. Just reset the cpus_allowed to the
4347 * cpuset's cpus_allowed
4349 cpumask_copy(new_mask, cpus_allowed);
4354 free_cpumask_var(new_mask);
4355 out_free_cpus_allowed:
4356 free_cpumask_var(cpus_allowed);
4363 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4364 struct cpumask *new_mask)
4366 if (len < cpumask_size())
4367 cpumask_clear(new_mask);
4368 else if (len > cpumask_size())
4369 len = cpumask_size();
4371 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4375 * sys_sched_setaffinity - set the cpu affinity of a process
4376 * @pid: pid of the process
4377 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4378 * @user_mask_ptr: user-space pointer to the new cpu mask
4380 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4381 unsigned long __user *, user_mask_ptr)
4383 cpumask_var_t new_mask;
4386 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4389 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4391 retval = sched_setaffinity(pid, new_mask);
4392 free_cpumask_var(new_mask);
4396 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4398 struct task_struct *p;
4399 unsigned long flags;
4406 p = find_process_by_pid(pid);
4410 retval = security_task_getscheduler(p);
4414 raw_spin_lock_irqsave(&p->pi_lock, flags);
4415 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4416 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4426 * sys_sched_getaffinity - get the cpu affinity of a process
4427 * @pid: pid of the process
4428 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4429 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4431 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4432 unsigned long __user *, user_mask_ptr)
4437 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4439 if (len & (sizeof(unsigned long)-1))
4442 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4445 ret = sched_getaffinity(pid, mask);
4447 size_t retlen = min_t(size_t, len, cpumask_size());
4449 if (copy_to_user(user_mask_ptr, mask, retlen))
4454 free_cpumask_var(mask);
4460 * sys_sched_yield - yield the current processor to other threads.
4462 * This function yields the current CPU to other tasks. If there are no
4463 * other threads running on this CPU then this function will return.
4465 SYSCALL_DEFINE0(sched_yield)
4467 struct rq *rq = this_rq_lock();
4469 schedstat_inc(rq, yld_count);
4470 current->sched_class->yield_task(rq);
4473 * Since we are going to call schedule() anyway, there's
4474 * no need to preempt or enable interrupts:
4476 __release(rq->lock);
4477 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4478 do_raw_spin_unlock(&rq->lock);
4479 preempt_enable_no_resched();
4486 static inline int should_resched(void)
4488 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4491 static void __cond_resched(void)
4493 add_preempt_count(PREEMPT_ACTIVE);
4495 sub_preempt_count(PREEMPT_ACTIVE);
4498 int __sched _cond_resched(void)
4500 if (should_resched()) {
4506 EXPORT_SYMBOL(_cond_resched);
4509 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4510 * call schedule, and on return reacquire the lock.
4512 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4513 * operations here to prevent schedule() from being called twice (once via
4514 * spin_unlock(), once by hand).
4516 int __cond_resched_lock(spinlock_t *lock)
4518 int resched = should_resched();
4521 lockdep_assert_held(lock);
4523 if (spin_needbreak(lock) || resched) {
4534 EXPORT_SYMBOL(__cond_resched_lock);
4536 int __sched __cond_resched_softirq(void)
4538 BUG_ON(!in_softirq());
4540 if (should_resched()) {
4548 EXPORT_SYMBOL(__cond_resched_softirq);
4551 * yield - yield the current processor to other threads.
4553 * This is a shortcut for kernel-space yielding - it marks the
4554 * thread runnable and calls sys_sched_yield().
4556 void __sched yield(void)
4558 set_current_state(TASK_RUNNING);
4561 EXPORT_SYMBOL(yield);
4564 * yield_to - yield the current processor to another thread in
4565 * your thread group, or accelerate that thread toward the
4566 * processor it's on.
4568 * @preempt: whether task preemption is allowed or not
4570 * It's the caller's job to ensure that the target task struct
4571 * can't go away on us before we can do any checks.
4573 * Returns true if we indeed boosted the target task.
4575 bool __sched yield_to(struct task_struct *p, bool preempt)
4577 struct task_struct *curr = current;
4578 struct rq *rq, *p_rq;
4579 unsigned long flags;
4582 local_irq_save(flags);
4587 double_rq_lock(rq, p_rq);
4588 while (task_rq(p) != p_rq) {
4589 double_rq_unlock(rq, p_rq);
4593 if (!curr->sched_class->yield_to_task)
4596 if (curr->sched_class != p->sched_class)
4599 if (task_running(p_rq, p) || p->state)
4602 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4604 schedstat_inc(rq, yld_count);
4606 * Make p's CPU reschedule; pick_next_entity takes care of
4609 if (preempt && rq != p_rq)
4610 resched_task(p_rq->curr);
4613 * We might have set it in task_yield_fair(), but are
4614 * not going to schedule(), so don't want to skip
4617 rq->skip_clock_update = 0;
4621 double_rq_unlock(rq, p_rq);
4622 local_irq_restore(flags);
4629 EXPORT_SYMBOL_GPL(yield_to);
4632 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4633 * that process accounting knows that this is a task in IO wait state.
4635 void __sched io_schedule(void)
4637 struct rq *rq = raw_rq();
4639 delayacct_blkio_start();
4640 atomic_inc(&rq->nr_iowait);
4641 blk_flush_plug(current);
4642 current->in_iowait = 1;
4644 current->in_iowait = 0;
4645 atomic_dec(&rq->nr_iowait);
4646 delayacct_blkio_end();
4648 EXPORT_SYMBOL(io_schedule);
4650 long __sched io_schedule_timeout(long timeout)
4652 struct rq *rq = raw_rq();
4655 delayacct_blkio_start();
4656 atomic_inc(&rq->nr_iowait);
4657 blk_flush_plug(current);
4658 current->in_iowait = 1;
4659 ret = schedule_timeout(timeout);
4660 current->in_iowait = 0;
4661 atomic_dec(&rq->nr_iowait);
4662 delayacct_blkio_end();
4667 * sys_sched_get_priority_max - return maximum RT priority.
4668 * @policy: scheduling class.
4670 * this syscall returns the maximum rt_priority that can be used
4671 * by a given scheduling class.
4673 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4680 ret = MAX_USER_RT_PRIO-1;
4692 * sys_sched_get_priority_min - return minimum RT priority.
4693 * @policy: scheduling class.
4695 * this syscall returns the minimum rt_priority that can be used
4696 * by a given scheduling class.
4698 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4716 * sys_sched_rr_get_interval - return the default timeslice of a process.
4717 * @pid: pid of the process.
4718 * @interval: userspace pointer to the timeslice value.
4720 * this syscall writes the default timeslice value of a given process
4721 * into the user-space timespec buffer. A value of '0' means infinity.
4723 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4724 struct timespec __user *, interval)
4726 struct task_struct *p;
4727 unsigned int time_slice;
4728 unsigned long flags;
4738 p = find_process_by_pid(pid);
4742 retval = security_task_getscheduler(p);
4746 rq = task_rq_lock(p, &flags);
4747 time_slice = p->sched_class->get_rr_interval(rq, p);
4748 task_rq_unlock(rq, p, &flags);
4751 jiffies_to_timespec(time_slice, &t);
4752 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4760 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4762 void sched_show_task(struct task_struct *p)
4764 unsigned long free = 0;
4767 state = p->state ? __ffs(p->state) + 1 : 0;
4768 printk(KERN_INFO "%-15.15s %c", p->comm,
4769 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4770 #if BITS_PER_LONG == 32
4771 if (state == TASK_RUNNING)
4772 printk(KERN_CONT " running ");
4774 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4776 if (state == TASK_RUNNING)
4777 printk(KERN_CONT " running task ");
4779 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4781 #ifdef CONFIG_DEBUG_STACK_USAGE
4782 free = stack_not_used(p);
4784 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4785 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4786 (unsigned long)task_thread_info(p)->flags);
4788 show_stack(p, NULL);
4791 void show_state_filter(unsigned long state_filter)
4793 struct task_struct *g, *p;
4795 #if BITS_PER_LONG == 32
4797 " task PC stack pid father\n");
4800 " task PC stack pid father\n");
4803 do_each_thread(g, p) {
4805 * reset the NMI-timeout, listing all files on a slow
4806 * console might take a lot of time:
4808 touch_nmi_watchdog();
4809 if (!state_filter || (p->state & state_filter))
4811 } while_each_thread(g, p);
4813 touch_all_softlockup_watchdogs();
4815 #ifdef CONFIG_SCHED_DEBUG
4816 sysrq_sched_debug_show();
4820 * Only show locks if all tasks are dumped:
4823 debug_show_all_locks();
4826 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4828 idle->sched_class = &idle_sched_class;
4832 * init_idle - set up an idle thread for a given CPU
4833 * @idle: task in question
4834 * @cpu: cpu the idle task belongs to
4836 * NOTE: this function does not set the idle thread's NEED_RESCHED
4837 * flag, to make booting more robust.
4839 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4841 struct rq *rq = cpu_rq(cpu);
4842 unsigned long flags;
4844 raw_spin_lock_irqsave(&rq->lock, flags);
4847 idle->state = TASK_RUNNING;
4848 idle->se.exec_start = sched_clock();
4850 do_set_cpus_allowed(idle, cpumask_of(cpu));
4852 * We're having a chicken and egg problem, even though we are
4853 * holding rq->lock, the cpu isn't yet set to this cpu so the
4854 * lockdep check in task_group() will fail.
4856 * Similar case to sched_fork(). / Alternatively we could
4857 * use task_rq_lock() here and obtain the other rq->lock.
4862 __set_task_cpu(idle, cpu);
4865 rq->curr = rq->idle = idle;
4866 #if defined(CONFIG_SMP)
4869 raw_spin_unlock_irqrestore(&rq->lock, flags);
4871 /* Set the preempt count _outside_ the spinlocks! */
4872 task_thread_info(idle)->preempt_count = 0;
4875 * The idle tasks have their own, simple scheduling class:
4877 idle->sched_class = &idle_sched_class;
4878 ftrace_graph_init_idle_task(idle, cpu);
4879 #if defined(CONFIG_SMP)
4880 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4885 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4887 if (p->sched_class && p->sched_class->set_cpus_allowed)
4888 p->sched_class->set_cpus_allowed(p, new_mask);
4890 cpumask_copy(&p->cpus_allowed, new_mask);
4891 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4895 * This is how migration works:
4897 * 1) we invoke migration_cpu_stop() on the target CPU using
4899 * 2) stopper starts to run (implicitly forcing the migrated thread
4901 * 3) it checks whether the migrated task is still in the wrong runqueue.
4902 * 4) if it's in the wrong runqueue then the migration thread removes
4903 * it and puts it into the right queue.
4904 * 5) stopper completes and stop_one_cpu() returns and the migration
4909 * Change a given task's CPU affinity. Migrate the thread to a
4910 * proper CPU and schedule it away if the CPU it's executing on
4911 * is removed from the allowed bitmask.
4913 * NOTE: the caller must have a valid reference to the task, the
4914 * task must not exit() & deallocate itself prematurely. The
4915 * call is not atomic; no spinlocks may be held.
4917 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4919 unsigned long flags;
4921 unsigned int dest_cpu;
4924 rq = task_rq_lock(p, &flags);
4926 if (cpumask_equal(&p->cpus_allowed, new_mask))
4929 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4934 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4939 do_set_cpus_allowed(p, new_mask);
4941 /* Can the task run on the task's current CPU? If so, we're done */
4942 if (cpumask_test_cpu(task_cpu(p), new_mask))
4945 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4947 struct migration_arg arg = { p, dest_cpu };
4948 /* Need help from migration thread: drop lock and wait. */
4949 task_rq_unlock(rq, p, &flags);
4950 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4951 tlb_migrate_finish(p->mm);
4955 task_rq_unlock(rq, p, &flags);
4959 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4962 * Move (not current) task off this cpu, onto dest cpu. We're doing
4963 * this because either it can't run here any more (set_cpus_allowed()
4964 * away from this CPU, or CPU going down), or because we're
4965 * attempting to rebalance this task on exec (sched_exec).
4967 * So we race with normal scheduler movements, but that's OK, as long
4968 * as the task is no longer on this CPU.
4970 * Returns non-zero if task was successfully migrated.
4972 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4974 struct rq *rq_dest, *rq_src;
4977 if (unlikely(!cpu_active(dest_cpu)))
4980 rq_src = cpu_rq(src_cpu);
4981 rq_dest = cpu_rq(dest_cpu);
4983 raw_spin_lock(&p->pi_lock);
4984 double_rq_lock(rq_src, rq_dest);
4985 /* Already moved. */
4986 if (task_cpu(p) != src_cpu)
4988 /* Affinity changed (again). */
4989 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4993 * If we're not on a rq, the next wake-up will ensure we're
4997 dequeue_task(rq_src, p, 0);
4998 set_task_cpu(p, dest_cpu);
4999 enqueue_task(rq_dest, p, 0);
5000 check_preempt_curr(rq_dest, p, 0);
5005 double_rq_unlock(rq_src, rq_dest);
5006 raw_spin_unlock(&p->pi_lock);
5011 * migration_cpu_stop - this will be executed by a highprio stopper thread
5012 * and performs thread migration by bumping thread off CPU then
5013 * 'pushing' onto another runqueue.
5015 static int migration_cpu_stop(void *data)
5017 struct migration_arg *arg = data;
5020 * The original target cpu might have gone down and we might
5021 * be on another cpu but it doesn't matter.
5023 local_irq_disable();
5024 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5029 #ifdef CONFIG_HOTPLUG_CPU
5032 * Ensures that the idle task is using init_mm right before its cpu goes
5035 void idle_task_exit(void)
5037 struct mm_struct *mm = current->active_mm;
5039 BUG_ON(cpu_online(smp_processor_id()));
5042 switch_mm(mm, &init_mm, current);
5047 * While a dead CPU has no uninterruptible tasks queued at this point,
5048 * it might still have a nonzero ->nr_uninterruptible counter, because
5049 * for performance reasons the counter is not stricly tracking tasks to
5050 * their home CPUs. So we just add the counter to another CPU's counter,
5051 * to keep the global sum constant after CPU-down:
5053 static void migrate_nr_uninterruptible(struct rq *rq_src)
5055 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5057 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5058 rq_src->nr_uninterruptible = 0;
5062 * remove the tasks which were accounted by rq from calc_load_tasks.
5064 static void calc_global_load_remove(struct rq *rq)
5066 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5067 rq->calc_load_active = 0;
5071 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5072 * try_to_wake_up()->select_task_rq().
5074 * Called with rq->lock held even though we'er in stop_machine() and
5075 * there's no concurrency possible, we hold the required locks anyway
5076 * because of lock validation efforts.
5078 static void migrate_tasks(unsigned int dead_cpu)
5080 struct rq *rq = cpu_rq(dead_cpu);
5081 struct task_struct *next, *stop = rq->stop;
5085 * Fudge the rq selection such that the below task selection loop
5086 * doesn't get stuck on the currently eligible stop task.
5088 * We're currently inside stop_machine() and the rq is either stuck
5089 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5090 * either way we should never end up calling schedule() until we're
5095 /* Ensure any throttled groups are reachable by pick_next_task */
5096 unthrottle_offline_cfs_rqs(rq);
5100 * There's this thread running, bail when that's the only
5103 if (rq->nr_running == 1)
5106 next = pick_next_task(rq);
5108 next->sched_class->put_prev_task(rq, next);
5110 /* Find suitable destination for @next, with force if needed. */
5111 dest_cpu = select_fallback_rq(dead_cpu, next);
5112 raw_spin_unlock(&rq->lock);
5114 __migrate_task(next, dead_cpu, dest_cpu);
5116 raw_spin_lock(&rq->lock);
5122 #endif /* CONFIG_HOTPLUG_CPU */
5124 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5126 static struct ctl_table sd_ctl_dir[] = {
5128 .procname = "sched_domain",
5134 static struct ctl_table sd_ctl_root[] = {
5136 .procname = "kernel",
5138 .child = sd_ctl_dir,
5143 static struct ctl_table *sd_alloc_ctl_entry(int n)
5145 struct ctl_table *entry =
5146 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5151 static void sd_free_ctl_entry(struct ctl_table **tablep)
5153 struct ctl_table *entry;
5156 * In the intermediate directories, both the child directory and
5157 * procname are dynamically allocated and could fail but the mode
5158 * will always be set. In the lowest directory the names are
5159 * static strings and all have proc handlers.
5161 for (entry = *tablep; entry->mode; entry++) {
5163 sd_free_ctl_entry(&entry->child);
5164 if (entry->proc_handler == NULL)
5165 kfree(entry->procname);
5173 set_table_entry(struct ctl_table *entry,
5174 const char *procname, void *data, int maxlen,
5175 umode_t mode, proc_handler *proc_handler)
5177 entry->procname = procname;
5179 entry->maxlen = maxlen;
5181 entry->proc_handler = proc_handler;
5184 static struct ctl_table *
5185 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5187 struct ctl_table *table = sd_alloc_ctl_entry(13);
5192 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5193 sizeof(long), 0644, proc_doulongvec_minmax);
5194 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5195 sizeof(long), 0644, proc_doulongvec_minmax);
5196 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5197 sizeof(int), 0644, proc_dointvec_minmax);
5198 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5199 sizeof(int), 0644, proc_dointvec_minmax);
5200 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5201 sizeof(int), 0644, proc_dointvec_minmax);
5202 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5203 sizeof(int), 0644, proc_dointvec_minmax);
5204 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5205 sizeof(int), 0644, proc_dointvec_minmax);
5206 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5207 sizeof(int), 0644, proc_dointvec_minmax);
5208 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5209 sizeof(int), 0644, proc_dointvec_minmax);
5210 set_table_entry(&table[9], "cache_nice_tries",
5211 &sd->cache_nice_tries,
5212 sizeof(int), 0644, proc_dointvec_minmax);
5213 set_table_entry(&table[10], "flags", &sd->flags,
5214 sizeof(int), 0644, proc_dointvec_minmax);
5215 set_table_entry(&table[11], "name", sd->name,
5216 CORENAME_MAX_SIZE, 0444, proc_dostring);
5217 /* &table[12] is terminator */
5222 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5224 struct ctl_table *entry, *table;
5225 struct sched_domain *sd;
5226 int domain_num = 0, i;
5229 for_each_domain(cpu, sd)
5231 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5236 for_each_domain(cpu, sd) {
5237 snprintf(buf, 32, "domain%d", i);
5238 entry->procname = kstrdup(buf, GFP_KERNEL);
5240 entry->child = sd_alloc_ctl_domain_table(sd);
5247 static struct ctl_table_header *sd_sysctl_header;
5248 static void register_sched_domain_sysctl(void)
5250 int i, cpu_num = num_possible_cpus();
5251 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5254 WARN_ON(sd_ctl_dir[0].child);
5255 sd_ctl_dir[0].child = entry;
5260 for_each_possible_cpu(i) {
5261 snprintf(buf, 32, "cpu%d", i);
5262 entry->procname = kstrdup(buf, GFP_KERNEL);
5264 entry->child = sd_alloc_ctl_cpu_table(i);
5268 WARN_ON(sd_sysctl_header);
5269 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5272 /* may be called multiple times per register */
5273 static void unregister_sched_domain_sysctl(void)
5275 if (sd_sysctl_header)
5276 unregister_sysctl_table(sd_sysctl_header);
5277 sd_sysctl_header = NULL;
5278 if (sd_ctl_dir[0].child)
5279 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5282 static void register_sched_domain_sysctl(void)
5285 static void unregister_sched_domain_sysctl(void)
5290 static void set_rq_online(struct rq *rq)
5293 const struct sched_class *class;
5295 cpumask_set_cpu(rq->cpu, rq->rd->online);
5298 for_each_class(class) {
5299 if (class->rq_online)
5300 class->rq_online(rq);
5305 static void set_rq_offline(struct rq *rq)
5308 const struct sched_class *class;
5310 for_each_class(class) {
5311 if (class->rq_offline)
5312 class->rq_offline(rq);
5315 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5321 * migration_call - callback that gets triggered when a CPU is added.
5322 * Here we can start up the necessary migration thread for the new CPU.
5324 static int __cpuinit
5325 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5327 int cpu = (long)hcpu;
5328 unsigned long flags;
5329 struct rq *rq = cpu_rq(cpu);
5331 switch (action & ~CPU_TASKS_FROZEN) {
5333 case CPU_UP_PREPARE:
5334 rq->calc_load_update = calc_load_update;
5338 /* Update our root-domain */
5339 raw_spin_lock_irqsave(&rq->lock, flags);
5341 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5345 raw_spin_unlock_irqrestore(&rq->lock, flags);
5348 #ifdef CONFIG_HOTPLUG_CPU
5350 sched_ttwu_pending();
5351 /* Update our root-domain */
5352 raw_spin_lock_irqsave(&rq->lock, flags);
5354 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5358 BUG_ON(rq->nr_running != 1); /* the migration thread */
5359 raw_spin_unlock_irqrestore(&rq->lock, flags);
5361 migrate_nr_uninterruptible(rq);
5362 calc_global_load_remove(rq);
5367 update_max_interval();
5373 * Register at high priority so that task migration (migrate_all_tasks)
5374 * happens before everything else. This has to be lower priority than
5375 * the notifier in the perf_event subsystem, though.
5377 static struct notifier_block __cpuinitdata migration_notifier = {
5378 .notifier_call = migration_call,
5379 .priority = CPU_PRI_MIGRATION,
5382 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5383 unsigned long action, void *hcpu)
5385 switch (action & ~CPU_TASKS_FROZEN) {
5387 case CPU_DOWN_FAILED:
5388 set_cpu_active((long)hcpu, true);
5395 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5396 unsigned long action, void *hcpu)
5398 switch (action & ~CPU_TASKS_FROZEN) {
5399 case CPU_DOWN_PREPARE:
5400 set_cpu_active((long)hcpu, false);
5407 static int __init migration_init(void)
5409 void *cpu = (void *)(long)smp_processor_id();
5412 /* Initialize migration for the boot CPU */
5413 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5414 BUG_ON(err == NOTIFY_BAD);
5415 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5416 register_cpu_notifier(&migration_notifier);
5418 /* Register cpu active notifiers */
5419 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5420 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5424 early_initcall(migration_init);
5429 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5431 #ifdef CONFIG_SCHED_DEBUG
5433 static __read_mostly int sched_domain_debug_enabled;
5435 static int __init sched_domain_debug_setup(char *str)
5437 sched_domain_debug_enabled = 1;
5441 early_param("sched_debug", sched_domain_debug_setup);
5443 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5444 struct cpumask *groupmask)
5446 struct sched_group *group = sd->groups;
5449 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5450 cpumask_clear(groupmask);
5452 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5454 if (!(sd->flags & SD_LOAD_BALANCE)) {
5455 printk("does not load-balance\n");
5457 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5462 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5464 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5465 printk(KERN_ERR "ERROR: domain->span does not contain "
5468 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5469 printk(KERN_ERR "ERROR: domain->groups does not contain"
5473 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5477 printk(KERN_ERR "ERROR: group is NULL\n");
5481 if (!group->sgp->power) {
5482 printk(KERN_CONT "\n");
5483 printk(KERN_ERR "ERROR: domain->cpu_power not "
5488 if (!cpumask_weight(sched_group_cpus(group))) {
5489 printk(KERN_CONT "\n");
5490 printk(KERN_ERR "ERROR: empty group\n");
5494 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5495 printk(KERN_CONT "\n");
5496 printk(KERN_ERR "ERROR: repeated CPUs\n");
5500 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5502 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5504 printk(KERN_CONT " %s", str);
5505 if (group->sgp->power != SCHED_POWER_SCALE) {
5506 printk(KERN_CONT " (cpu_power = %d)",
5510 group = group->next;
5511 } while (group != sd->groups);
5512 printk(KERN_CONT "\n");
5514 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5515 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5518 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5519 printk(KERN_ERR "ERROR: parent span is not a superset "
5520 "of domain->span\n");
5524 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5528 if (!sched_domain_debug_enabled)
5532 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5536 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5539 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5547 #else /* !CONFIG_SCHED_DEBUG */
5548 # define sched_domain_debug(sd, cpu) do { } while (0)
5549 #endif /* CONFIG_SCHED_DEBUG */
5551 static int sd_degenerate(struct sched_domain *sd)
5553 if (cpumask_weight(sched_domain_span(sd)) == 1)
5556 /* Following flags need at least 2 groups */
5557 if (sd->flags & (SD_LOAD_BALANCE |
5558 SD_BALANCE_NEWIDLE |
5562 SD_SHARE_PKG_RESOURCES)) {
5563 if (sd->groups != sd->groups->next)
5567 /* Following flags don't use groups */
5568 if (sd->flags & (SD_WAKE_AFFINE))
5575 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5577 unsigned long cflags = sd->flags, pflags = parent->flags;
5579 if (sd_degenerate(parent))
5582 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5585 /* Flags needing groups don't count if only 1 group in parent */
5586 if (parent->groups == parent->groups->next) {
5587 pflags &= ~(SD_LOAD_BALANCE |
5588 SD_BALANCE_NEWIDLE |
5592 SD_SHARE_PKG_RESOURCES);
5593 if (nr_node_ids == 1)
5594 pflags &= ~SD_SERIALIZE;
5596 if (~cflags & pflags)
5602 static void free_rootdomain(struct rcu_head *rcu)
5604 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5606 cpupri_cleanup(&rd->cpupri);
5607 free_cpumask_var(rd->rto_mask);
5608 free_cpumask_var(rd->online);
5609 free_cpumask_var(rd->span);
5613 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5615 struct root_domain *old_rd = NULL;
5616 unsigned long flags;
5618 raw_spin_lock_irqsave(&rq->lock, flags);
5623 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5626 cpumask_clear_cpu(rq->cpu, old_rd->span);
5629 * If we dont want to free the old_rt yet then
5630 * set old_rd to NULL to skip the freeing later
5633 if (!atomic_dec_and_test(&old_rd->refcount))
5637 atomic_inc(&rd->refcount);
5640 cpumask_set_cpu(rq->cpu, rd->span);
5641 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5644 raw_spin_unlock_irqrestore(&rq->lock, flags);
5647 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5650 static int init_rootdomain(struct root_domain *rd)
5652 memset(rd, 0, sizeof(*rd));
5654 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5656 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5658 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5661 if (cpupri_init(&rd->cpupri) != 0)
5666 free_cpumask_var(rd->rto_mask);
5668 free_cpumask_var(rd->online);
5670 free_cpumask_var(rd->span);
5676 * By default the system creates a single root-domain with all cpus as
5677 * members (mimicking the global state we have today).
5679 struct root_domain def_root_domain;
5681 static void init_defrootdomain(void)
5683 init_rootdomain(&def_root_domain);
5685 atomic_set(&def_root_domain.refcount, 1);
5688 static struct root_domain *alloc_rootdomain(void)
5690 struct root_domain *rd;
5692 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5696 if (init_rootdomain(rd) != 0) {
5704 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5706 struct sched_group *tmp, *first;
5715 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5720 } while (sg != first);
5723 static void free_sched_domain(struct rcu_head *rcu)
5725 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5728 * If its an overlapping domain it has private groups, iterate and
5731 if (sd->flags & SD_OVERLAP) {
5732 free_sched_groups(sd->groups, 1);
5733 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5734 kfree(sd->groups->sgp);
5740 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5742 call_rcu(&sd->rcu, free_sched_domain);
5745 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5747 for (; sd; sd = sd->parent)
5748 destroy_sched_domain(sd, cpu);
5752 * Keep a special pointer to the highest sched_domain that has
5753 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5754 * allows us to avoid some pointer chasing select_idle_sibling().
5756 * Also keep a unique ID per domain (we use the first cpu number in
5757 * the cpumask of the domain), this allows us to quickly tell if
5758 * two cpus are in the same cache domain, see ttwu_share_cache().
5760 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5761 DEFINE_PER_CPU(int, sd_llc_id);
5763 static void update_top_cache_domain(int cpu)
5765 struct sched_domain *sd;
5768 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5770 id = cpumask_first(sched_domain_span(sd));
5772 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5773 per_cpu(sd_llc_id, cpu) = id;
5777 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5778 * hold the hotplug lock.
5781 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5783 struct rq *rq = cpu_rq(cpu);
5784 struct sched_domain *tmp;
5786 /* Remove the sched domains which do not contribute to scheduling. */
5787 for (tmp = sd; tmp; ) {
5788 struct sched_domain *parent = tmp->parent;
5792 if (sd_parent_degenerate(tmp, parent)) {
5793 tmp->parent = parent->parent;
5795 parent->parent->child = tmp;
5796 destroy_sched_domain(parent, cpu);
5801 if (sd && sd_degenerate(sd)) {
5804 destroy_sched_domain(tmp, cpu);
5809 sched_domain_debug(sd, cpu);
5811 rq_attach_root(rq, rd);
5813 rcu_assign_pointer(rq->sd, sd);
5814 destroy_sched_domains(tmp, cpu);
5816 update_top_cache_domain(cpu);
5819 /* cpus with isolated domains */
5820 static cpumask_var_t cpu_isolated_map;
5822 /* Setup the mask of cpus configured for isolated domains */
5823 static int __init isolated_cpu_setup(char *str)
5825 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5826 cpulist_parse(str, cpu_isolated_map);
5830 __setup("isolcpus=", isolated_cpu_setup);
5835 * find_next_best_node - find the next node to include in a sched_domain
5836 * @node: node whose sched_domain we're building
5837 * @used_nodes: nodes already in the sched_domain
5839 * Find the next node to include in a given scheduling domain. Simply
5840 * finds the closest node not already in the @used_nodes map.
5842 * Should use nodemask_t.
5844 static int find_next_best_node(int node, nodemask_t *used_nodes)
5846 int i, n, val, min_val, best_node = -1;
5850 for (i = 0; i < nr_node_ids; i++) {
5851 /* Start at @node */
5852 n = (node + i) % nr_node_ids;
5854 if (!nr_cpus_node(n))
5857 /* Skip already used nodes */
5858 if (node_isset(n, *used_nodes))
5861 /* Simple min distance search */
5862 val = node_distance(node, n);
5864 if (val < min_val) {
5870 if (best_node != -1)
5871 node_set(best_node, *used_nodes);
5876 * sched_domain_node_span - get a cpumask for a node's sched_domain
5877 * @node: node whose cpumask we're constructing
5878 * @span: resulting cpumask
5880 * Given a node, construct a good cpumask for its sched_domain to span. It
5881 * should be one that prevents unnecessary balancing, but also spreads tasks
5884 static void sched_domain_node_span(int node, struct cpumask *span)
5886 nodemask_t used_nodes;
5889 cpumask_clear(span);
5890 nodes_clear(used_nodes);
5892 cpumask_or(span, span, cpumask_of_node(node));
5893 node_set(node, used_nodes);
5895 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5896 int next_node = find_next_best_node(node, &used_nodes);
5899 cpumask_or(span, span, cpumask_of_node(next_node));
5903 static const struct cpumask *cpu_node_mask(int cpu)
5905 lockdep_assert_held(&sched_domains_mutex);
5907 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5909 return sched_domains_tmpmask;
5912 static const struct cpumask *cpu_allnodes_mask(int cpu)
5914 return cpu_possible_mask;
5916 #endif /* CONFIG_NUMA */
5918 static const struct cpumask *cpu_cpu_mask(int cpu)
5920 return cpumask_of_node(cpu_to_node(cpu));
5923 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5926 struct sched_domain **__percpu sd;
5927 struct sched_group **__percpu sg;
5928 struct sched_group_power **__percpu sgp;
5932 struct sched_domain ** __percpu sd;
5933 struct root_domain *rd;
5943 struct sched_domain_topology_level;
5945 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5946 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5948 #define SDTL_OVERLAP 0x01
5950 struct sched_domain_topology_level {
5951 sched_domain_init_f init;
5952 sched_domain_mask_f mask;
5954 struct sd_data data;
5958 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5960 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5961 const struct cpumask *span = sched_domain_span(sd);
5962 struct cpumask *covered = sched_domains_tmpmask;
5963 struct sd_data *sdd = sd->private;
5964 struct sched_domain *child;
5967 cpumask_clear(covered);
5969 for_each_cpu(i, span) {
5970 struct cpumask *sg_span;
5972 if (cpumask_test_cpu(i, covered))
5975 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5976 GFP_KERNEL, cpu_to_node(cpu));
5981 sg_span = sched_group_cpus(sg);
5983 child = *per_cpu_ptr(sdd->sd, i);
5985 child = child->child;
5986 cpumask_copy(sg_span, sched_domain_span(child));
5988 cpumask_set_cpu(i, sg_span);
5990 cpumask_or(covered, covered, sg_span);
5992 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5993 atomic_inc(&sg->sgp->ref);
5995 if (cpumask_test_cpu(cpu, sg_span))
6005 sd->groups = groups;
6010 free_sched_groups(first, 0);
6015 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6017 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6018 struct sched_domain *child = sd->child;
6021 cpu = cpumask_first(sched_domain_span(child));
6024 *sg = *per_cpu_ptr(sdd->sg, cpu);
6025 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6026 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6033 * build_sched_groups will build a circular linked list of the groups
6034 * covered by the given span, and will set each group's ->cpumask correctly,
6035 * and ->cpu_power to 0.
6037 * Assumes the sched_domain tree is fully constructed
6040 build_sched_groups(struct sched_domain *sd, int cpu)
6042 struct sched_group *first = NULL, *last = NULL;
6043 struct sd_data *sdd = sd->private;
6044 const struct cpumask *span = sched_domain_span(sd);
6045 struct cpumask *covered;
6048 get_group(cpu, sdd, &sd->groups);
6049 atomic_inc(&sd->groups->ref);
6051 if (cpu != cpumask_first(sched_domain_span(sd)))
6054 lockdep_assert_held(&sched_domains_mutex);
6055 covered = sched_domains_tmpmask;
6057 cpumask_clear(covered);
6059 for_each_cpu(i, span) {
6060 struct sched_group *sg;
6061 int group = get_group(i, sdd, &sg);
6064 if (cpumask_test_cpu(i, covered))
6067 cpumask_clear(sched_group_cpus(sg));
6070 for_each_cpu(j, span) {
6071 if (get_group(j, sdd, NULL) != group)
6074 cpumask_set_cpu(j, covered);
6075 cpumask_set_cpu(j, sched_group_cpus(sg));
6090 * Initialize sched groups cpu_power.
6092 * cpu_power indicates the capacity of sched group, which is used while
6093 * distributing the load between different sched groups in a sched domain.
6094 * Typically cpu_power for all the groups in a sched domain will be same unless
6095 * there are asymmetries in the topology. If there are asymmetries, group
6096 * having more cpu_power will pickup more load compared to the group having
6099 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6101 struct sched_group *sg = sd->groups;
6103 WARN_ON(!sd || !sg);
6106 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6108 } while (sg != sd->groups);
6110 if (cpu != group_first_cpu(sg))
6113 update_group_power(sd, cpu);
6114 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6117 int __weak arch_sd_sibling_asym_packing(void)
6119 return 0*SD_ASYM_PACKING;
6123 * Initializers for schedule domains
6124 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6127 #ifdef CONFIG_SCHED_DEBUG
6128 # define SD_INIT_NAME(sd, type) sd->name = #type
6130 # define SD_INIT_NAME(sd, type) do { } while (0)
6133 #define SD_INIT_FUNC(type) \
6134 static noinline struct sched_domain * \
6135 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6137 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6138 *sd = SD_##type##_INIT; \
6139 SD_INIT_NAME(sd, type); \
6140 sd->private = &tl->data; \
6146 SD_INIT_FUNC(ALLNODES)
6149 #ifdef CONFIG_SCHED_SMT
6150 SD_INIT_FUNC(SIBLING)
6152 #ifdef CONFIG_SCHED_MC
6155 #ifdef CONFIG_SCHED_BOOK
6159 static int default_relax_domain_level = -1;
6160 int sched_domain_level_max;
6162 static int __init setup_relax_domain_level(char *str)
6166 val = simple_strtoul(str, NULL, 0);
6167 if (val < sched_domain_level_max)
6168 default_relax_domain_level = val;
6172 __setup("relax_domain_level=", setup_relax_domain_level);
6174 static void set_domain_attribute(struct sched_domain *sd,
6175 struct sched_domain_attr *attr)
6179 if (!attr || attr->relax_domain_level < 0) {
6180 if (default_relax_domain_level < 0)
6183 request = default_relax_domain_level;
6185 request = attr->relax_domain_level;
6186 if (request < sd->level) {
6187 /* turn off idle balance on this domain */
6188 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6190 /* turn on idle balance on this domain */
6191 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6195 static void __sdt_free(const struct cpumask *cpu_map);
6196 static int __sdt_alloc(const struct cpumask *cpu_map);
6198 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6199 const struct cpumask *cpu_map)
6203 if (!atomic_read(&d->rd->refcount))
6204 free_rootdomain(&d->rd->rcu); /* fall through */
6206 free_percpu(d->sd); /* fall through */
6208 __sdt_free(cpu_map); /* fall through */
6214 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6215 const struct cpumask *cpu_map)
6217 memset(d, 0, sizeof(*d));
6219 if (__sdt_alloc(cpu_map))
6220 return sa_sd_storage;
6221 d->sd = alloc_percpu(struct sched_domain *);
6223 return sa_sd_storage;
6224 d->rd = alloc_rootdomain();
6227 return sa_rootdomain;
6231 * NULL the sd_data elements we've used to build the sched_domain and
6232 * sched_group structure so that the subsequent __free_domain_allocs()
6233 * will not free the data we're using.
6235 static void claim_allocations(int cpu, struct sched_domain *sd)
6237 struct sd_data *sdd = sd->private;
6239 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6240 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6242 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6243 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6245 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6246 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6249 #ifdef CONFIG_SCHED_SMT
6250 static const struct cpumask *cpu_smt_mask(int cpu)
6252 return topology_thread_cpumask(cpu);
6257 * Topology list, bottom-up.
6259 static struct sched_domain_topology_level default_topology[] = {
6260 #ifdef CONFIG_SCHED_SMT
6261 { sd_init_SIBLING, cpu_smt_mask, },
6263 #ifdef CONFIG_SCHED_MC
6264 { sd_init_MC, cpu_coregroup_mask, },
6266 #ifdef CONFIG_SCHED_BOOK
6267 { sd_init_BOOK, cpu_book_mask, },
6269 { sd_init_CPU, cpu_cpu_mask, },
6271 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6272 { sd_init_ALLNODES, cpu_allnodes_mask, },
6277 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6279 static int __sdt_alloc(const struct cpumask *cpu_map)
6281 struct sched_domain_topology_level *tl;
6284 for (tl = sched_domain_topology; tl->init; tl++) {
6285 struct sd_data *sdd = &tl->data;
6287 sdd->sd = alloc_percpu(struct sched_domain *);
6291 sdd->sg = alloc_percpu(struct sched_group *);
6295 sdd->sgp = alloc_percpu(struct sched_group_power *);
6299 for_each_cpu(j, cpu_map) {
6300 struct sched_domain *sd;
6301 struct sched_group *sg;
6302 struct sched_group_power *sgp;
6304 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6305 GFP_KERNEL, cpu_to_node(j));
6309 *per_cpu_ptr(sdd->sd, j) = sd;
6311 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6312 GFP_KERNEL, cpu_to_node(j));
6316 *per_cpu_ptr(sdd->sg, j) = sg;
6318 sgp = kzalloc_node(sizeof(struct sched_group_power),
6319 GFP_KERNEL, cpu_to_node(j));
6323 *per_cpu_ptr(sdd->sgp, j) = sgp;
6330 static void __sdt_free(const struct cpumask *cpu_map)
6332 struct sched_domain_topology_level *tl;
6335 for (tl = sched_domain_topology; tl->init; tl++) {
6336 struct sd_data *sdd = &tl->data;
6338 for_each_cpu(j, cpu_map) {
6339 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6340 if (sd && (sd->flags & SD_OVERLAP))
6341 free_sched_groups(sd->groups, 0);
6342 kfree(*per_cpu_ptr(sdd->sd, j));
6343 kfree(*per_cpu_ptr(sdd->sg, j));
6344 kfree(*per_cpu_ptr(sdd->sgp, j));
6346 free_percpu(sdd->sd);
6347 free_percpu(sdd->sg);
6348 free_percpu(sdd->sgp);
6352 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6353 struct s_data *d, const struct cpumask *cpu_map,
6354 struct sched_domain_attr *attr, struct sched_domain *child,
6357 struct sched_domain *sd = tl->init(tl, cpu);
6361 set_domain_attribute(sd, attr);
6362 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6364 sd->level = child->level + 1;
6365 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6374 * Build sched domains for a given set of cpus and attach the sched domains
6375 * to the individual cpus
6377 static int build_sched_domains(const struct cpumask *cpu_map,
6378 struct sched_domain_attr *attr)
6380 enum s_alloc alloc_state = sa_none;
6381 struct sched_domain *sd;
6383 int i, ret = -ENOMEM;
6385 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6386 if (alloc_state != sa_rootdomain)
6389 /* Set up domains for cpus specified by the cpu_map. */
6390 for_each_cpu(i, cpu_map) {
6391 struct sched_domain_topology_level *tl;
6394 for (tl = sched_domain_topology; tl->init; tl++) {
6395 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6396 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6397 sd->flags |= SD_OVERLAP;
6398 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6405 *per_cpu_ptr(d.sd, i) = sd;
6408 /* Build the groups for the domains */
6409 for_each_cpu(i, cpu_map) {
6410 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6411 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6412 if (sd->flags & SD_OVERLAP) {
6413 if (build_overlap_sched_groups(sd, i))
6416 if (build_sched_groups(sd, i))
6422 /* Calculate CPU power for physical packages and nodes */
6423 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6424 if (!cpumask_test_cpu(i, cpu_map))
6427 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6428 claim_allocations(i, sd);
6429 init_sched_groups_power(i, sd);
6433 /* Attach the domains */
6435 for_each_cpu(i, cpu_map) {
6436 sd = *per_cpu_ptr(d.sd, i);
6437 cpu_attach_domain(sd, d.rd, i);
6443 __free_domain_allocs(&d, alloc_state, cpu_map);
6447 static cpumask_var_t *doms_cur; /* current sched domains */
6448 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6449 static struct sched_domain_attr *dattr_cur;
6450 /* attribues of custom domains in 'doms_cur' */
6453 * Special case: If a kmalloc of a doms_cur partition (array of
6454 * cpumask) fails, then fallback to a single sched domain,
6455 * as determined by the single cpumask fallback_doms.
6457 static cpumask_var_t fallback_doms;
6460 * arch_update_cpu_topology lets virtualized architectures update the
6461 * cpu core maps. It is supposed to return 1 if the topology changed
6462 * or 0 if it stayed the same.
6464 int __attribute__((weak)) arch_update_cpu_topology(void)
6469 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6472 cpumask_var_t *doms;
6474 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6477 for (i = 0; i < ndoms; i++) {
6478 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6479 free_sched_domains(doms, i);
6486 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6489 for (i = 0; i < ndoms; i++)
6490 free_cpumask_var(doms[i]);
6495 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6496 * For now this just excludes isolated cpus, but could be used to
6497 * exclude other special cases in the future.
6499 static int init_sched_domains(const struct cpumask *cpu_map)
6503 arch_update_cpu_topology();
6505 doms_cur = alloc_sched_domains(ndoms_cur);
6507 doms_cur = &fallback_doms;
6508 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6510 err = build_sched_domains(doms_cur[0], NULL);
6511 register_sched_domain_sysctl();
6517 * Detach sched domains from a group of cpus specified in cpu_map
6518 * These cpus will now be attached to the NULL domain
6520 static void detach_destroy_domains(const struct cpumask *cpu_map)
6525 for_each_cpu(i, cpu_map)
6526 cpu_attach_domain(NULL, &def_root_domain, i);
6530 /* handle null as "default" */
6531 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6532 struct sched_domain_attr *new, int idx_new)
6534 struct sched_domain_attr tmp;
6541 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6542 new ? (new + idx_new) : &tmp,
6543 sizeof(struct sched_domain_attr));
6547 * Partition sched domains as specified by the 'ndoms_new'
6548 * cpumasks in the array doms_new[] of cpumasks. This compares
6549 * doms_new[] to the current sched domain partitioning, doms_cur[].
6550 * It destroys each deleted domain and builds each new domain.
6552 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6553 * The masks don't intersect (don't overlap.) We should setup one
6554 * sched domain for each mask. CPUs not in any of the cpumasks will
6555 * not be load balanced. If the same cpumask appears both in the
6556 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6559 * The passed in 'doms_new' should be allocated using
6560 * alloc_sched_domains. This routine takes ownership of it and will
6561 * free_sched_domains it when done with it. If the caller failed the
6562 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6563 * and partition_sched_domains() will fallback to the single partition
6564 * 'fallback_doms', it also forces the domains to be rebuilt.
6566 * If doms_new == NULL it will be replaced with cpu_online_mask.
6567 * ndoms_new == 0 is a special case for destroying existing domains,
6568 * and it will not create the default domain.
6570 * Call with hotplug lock held
6572 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6573 struct sched_domain_attr *dattr_new)
6578 mutex_lock(&sched_domains_mutex);
6580 /* always unregister in case we don't destroy any domains */
6581 unregister_sched_domain_sysctl();
6583 /* Let architecture update cpu core mappings. */
6584 new_topology = arch_update_cpu_topology();
6586 n = doms_new ? ndoms_new : 0;
6588 /* Destroy deleted domains */
6589 for (i = 0; i < ndoms_cur; i++) {
6590 for (j = 0; j < n && !new_topology; j++) {
6591 if (cpumask_equal(doms_cur[i], doms_new[j])
6592 && dattrs_equal(dattr_cur, i, dattr_new, j))
6595 /* no match - a current sched domain not in new doms_new[] */
6596 detach_destroy_domains(doms_cur[i]);
6601 if (doms_new == NULL) {
6603 doms_new = &fallback_doms;
6604 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6605 WARN_ON_ONCE(dattr_new);
6608 /* Build new domains */
6609 for (i = 0; i < ndoms_new; i++) {
6610 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6611 if (cpumask_equal(doms_new[i], doms_cur[j])
6612 && dattrs_equal(dattr_new, i, dattr_cur, j))
6615 /* no match - add a new doms_new */
6616 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6621 /* Remember the new sched domains */
6622 if (doms_cur != &fallback_doms)
6623 free_sched_domains(doms_cur, ndoms_cur);
6624 kfree(dattr_cur); /* kfree(NULL) is safe */
6625 doms_cur = doms_new;
6626 dattr_cur = dattr_new;
6627 ndoms_cur = ndoms_new;
6629 register_sched_domain_sysctl();
6631 mutex_unlock(&sched_domains_mutex);
6634 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6635 static void reinit_sched_domains(void)
6639 /* Destroy domains first to force the rebuild */
6640 partition_sched_domains(0, NULL, NULL);
6642 rebuild_sched_domains();
6646 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6648 unsigned int level = 0;
6650 if (sscanf(buf, "%u", &level) != 1)
6654 * level is always be positive so don't check for
6655 * level < POWERSAVINGS_BALANCE_NONE which is 0
6656 * What happens on 0 or 1 byte write,
6657 * need to check for count as well?
6660 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6664 sched_smt_power_savings = level;
6666 sched_mc_power_savings = level;
6668 reinit_sched_domains();
6673 #ifdef CONFIG_SCHED_MC
6674 static ssize_t sched_mc_power_savings_show(struct device *dev,
6675 struct device_attribute *attr,
6678 return sprintf(buf, "%u\n", sched_mc_power_savings);
6680 static ssize_t sched_mc_power_savings_store(struct device *dev,
6681 struct device_attribute *attr,
6682 const char *buf, size_t count)
6684 return sched_power_savings_store(buf, count, 0);
6686 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6687 sched_mc_power_savings_show,
6688 sched_mc_power_savings_store);
6691 #ifdef CONFIG_SCHED_SMT
6692 static ssize_t sched_smt_power_savings_show(struct device *dev,
6693 struct device_attribute *attr,
6696 return sprintf(buf, "%u\n", sched_smt_power_savings);
6698 static ssize_t sched_smt_power_savings_store(struct device *dev,
6699 struct device_attribute *attr,
6700 const char *buf, size_t count)
6702 return sched_power_savings_store(buf, count, 1);
6704 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6705 sched_smt_power_savings_show,
6706 sched_smt_power_savings_store);
6709 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6713 #ifdef CONFIG_SCHED_SMT
6715 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6717 #ifdef CONFIG_SCHED_MC
6718 if (!err && mc_capable())
6719 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6723 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6726 * Update cpusets according to cpu_active mask. If cpusets are
6727 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6728 * around partition_sched_domains().
6730 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6733 switch (action & ~CPU_TASKS_FROZEN) {
6735 case CPU_DOWN_FAILED:
6736 cpuset_update_active_cpus();
6743 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6746 switch (action & ~CPU_TASKS_FROZEN) {
6747 case CPU_DOWN_PREPARE:
6748 cpuset_update_active_cpus();
6755 void __init sched_init_smp(void)
6757 cpumask_var_t non_isolated_cpus;
6759 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6760 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6763 mutex_lock(&sched_domains_mutex);
6764 init_sched_domains(cpu_active_mask);
6765 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6766 if (cpumask_empty(non_isolated_cpus))
6767 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6768 mutex_unlock(&sched_domains_mutex);
6771 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6772 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6774 /* RT runtime code needs to handle some hotplug events */
6775 hotcpu_notifier(update_runtime, 0);
6779 /* Move init over to a non-isolated CPU */
6780 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6782 sched_init_granularity();
6783 free_cpumask_var(non_isolated_cpus);
6785 init_sched_rt_class();
6788 void __init sched_init_smp(void)
6790 sched_init_granularity();
6792 #endif /* CONFIG_SMP */
6794 const_debug unsigned int sysctl_timer_migration = 1;
6796 int in_sched_functions(unsigned long addr)
6798 return in_lock_functions(addr) ||
6799 (addr >= (unsigned long)__sched_text_start
6800 && addr < (unsigned long)__sched_text_end);
6803 #ifdef CONFIG_CGROUP_SCHED
6804 struct task_group root_task_group;
6807 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6809 void __init sched_init(void)
6812 unsigned long alloc_size = 0, ptr;
6814 #ifdef CONFIG_FAIR_GROUP_SCHED
6815 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6817 #ifdef CONFIG_RT_GROUP_SCHED
6818 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6820 #ifdef CONFIG_CPUMASK_OFFSTACK
6821 alloc_size += num_possible_cpus() * cpumask_size();
6824 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6826 #ifdef CONFIG_FAIR_GROUP_SCHED
6827 root_task_group.se = (struct sched_entity **)ptr;
6828 ptr += nr_cpu_ids * sizeof(void **);
6830 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6831 ptr += nr_cpu_ids * sizeof(void **);
6833 #endif /* CONFIG_FAIR_GROUP_SCHED */
6834 #ifdef CONFIG_RT_GROUP_SCHED
6835 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6836 ptr += nr_cpu_ids * sizeof(void **);
6838 root_task_group.rt_rq = (struct rt_rq **)ptr;
6839 ptr += nr_cpu_ids * sizeof(void **);
6841 #endif /* CONFIG_RT_GROUP_SCHED */
6842 #ifdef CONFIG_CPUMASK_OFFSTACK
6843 for_each_possible_cpu(i) {
6844 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6845 ptr += cpumask_size();
6847 #endif /* CONFIG_CPUMASK_OFFSTACK */
6851 init_defrootdomain();
6854 init_rt_bandwidth(&def_rt_bandwidth,
6855 global_rt_period(), global_rt_runtime());
6857 #ifdef CONFIG_RT_GROUP_SCHED
6858 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6859 global_rt_period(), global_rt_runtime());
6860 #endif /* CONFIG_RT_GROUP_SCHED */
6862 #ifdef CONFIG_CGROUP_SCHED
6863 list_add(&root_task_group.list, &task_groups);
6864 INIT_LIST_HEAD(&root_task_group.children);
6865 INIT_LIST_HEAD(&root_task_group.siblings);
6866 autogroup_init(&init_task);
6868 #endif /* CONFIG_CGROUP_SCHED */
6870 #ifdef CONFIG_CGROUP_CPUACCT
6871 root_cpuacct.cpustat = &kernel_cpustat;
6872 root_cpuacct.cpuusage = alloc_percpu(u64);
6873 /* Too early, not expected to fail */
6874 BUG_ON(!root_cpuacct.cpuusage);
6876 for_each_possible_cpu(i) {
6880 raw_spin_lock_init(&rq->lock);
6882 rq->calc_load_active = 0;
6883 rq->calc_load_update = jiffies + LOAD_FREQ;
6884 init_cfs_rq(&rq->cfs);
6885 init_rt_rq(&rq->rt, rq);
6886 #ifdef CONFIG_FAIR_GROUP_SCHED
6887 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6888 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6890 * How much cpu bandwidth does root_task_group get?
6892 * In case of task-groups formed thr' the cgroup filesystem, it
6893 * gets 100% of the cpu resources in the system. This overall
6894 * system cpu resource is divided among the tasks of
6895 * root_task_group and its child task-groups in a fair manner,
6896 * based on each entity's (task or task-group's) weight
6897 * (se->load.weight).
6899 * In other words, if root_task_group has 10 tasks of weight
6900 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6901 * then A0's share of the cpu resource is:
6903 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6905 * We achieve this by letting root_task_group's tasks sit
6906 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6908 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6909 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6910 #endif /* CONFIG_FAIR_GROUP_SCHED */
6912 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6913 #ifdef CONFIG_RT_GROUP_SCHED
6914 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6915 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6918 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6919 rq->cpu_load[j] = 0;
6921 rq->last_load_update_tick = jiffies;
6926 rq->cpu_power = SCHED_POWER_SCALE;
6927 rq->post_schedule = 0;
6928 rq->active_balance = 0;
6929 rq->next_balance = jiffies;
6934 rq->avg_idle = 2*sysctl_sched_migration_cost;
6935 rq_attach_root(rq, &def_root_domain);
6941 atomic_set(&rq->nr_iowait, 0);
6944 set_load_weight(&init_task);
6946 #ifdef CONFIG_PREEMPT_NOTIFIERS
6947 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6950 #ifdef CONFIG_RT_MUTEXES
6951 plist_head_init(&init_task.pi_waiters);
6955 * The boot idle thread does lazy MMU switching as well:
6957 atomic_inc(&init_mm.mm_count);
6958 enter_lazy_tlb(&init_mm, current);
6961 * Make us the idle thread. Technically, schedule() should not be
6962 * called from this thread, however somewhere below it might be,
6963 * but because we are the idle thread, we just pick up running again
6964 * when this runqueue becomes "idle".
6966 init_idle(current, smp_processor_id());
6968 calc_load_update = jiffies + LOAD_FREQ;
6971 * During early bootup we pretend to be a normal task:
6973 current->sched_class = &fair_sched_class;
6976 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6977 /* May be allocated at isolcpus cmdline parse time */
6978 if (cpu_isolated_map == NULL)
6979 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6981 init_sched_fair_class();
6983 scheduler_running = 1;
6986 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6987 static inline int preempt_count_equals(int preempt_offset)
6989 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6991 return (nested == preempt_offset);
6994 void __might_sleep(const char *file, int line, int preempt_offset)
6996 static unsigned long prev_jiffy; /* ratelimiting */
6998 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6999 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7000 system_state != SYSTEM_RUNNING || oops_in_progress)
7002 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7004 prev_jiffy = jiffies;
7007 "BUG: sleeping function called from invalid context at %s:%d\n",
7010 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7011 in_atomic(), irqs_disabled(),
7012 current->pid, current->comm);
7014 debug_show_held_locks(current);
7015 if (irqs_disabled())
7016 print_irqtrace_events(current);
7019 EXPORT_SYMBOL(__might_sleep);
7022 #ifdef CONFIG_MAGIC_SYSRQ
7023 static void normalize_task(struct rq *rq, struct task_struct *p)
7025 const struct sched_class *prev_class = p->sched_class;
7026 int old_prio = p->prio;
7031 dequeue_task(rq, p, 0);
7032 __setscheduler(rq, p, SCHED_NORMAL, 0);
7034 enqueue_task(rq, p, 0);
7035 resched_task(rq->curr);
7038 check_class_changed(rq, p, prev_class, old_prio);
7041 void normalize_rt_tasks(void)
7043 struct task_struct *g, *p;
7044 unsigned long flags;
7047 read_lock_irqsave(&tasklist_lock, flags);
7048 do_each_thread(g, p) {
7050 * Only normalize user tasks:
7055 p->se.exec_start = 0;
7056 #ifdef CONFIG_SCHEDSTATS
7057 p->se.statistics.wait_start = 0;
7058 p->se.statistics.sleep_start = 0;
7059 p->se.statistics.block_start = 0;
7064 * Renice negative nice level userspace
7067 if (TASK_NICE(p) < 0 && p->mm)
7068 set_user_nice(p, 0);
7072 raw_spin_lock(&p->pi_lock);
7073 rq = __task_rq_lock(p);
7075 normalize_task(rq, p);
7077 __task_rq_unlock(rq);
7078 raw_spin_unlock(&p->pi_lock);
7079 } while_each_thread(g, p);
7081 read_unlock_irqrestore(&tasklist_lock, flags);
7084 #endif /* CONFIG_MAGIC_SYSRQ */
7086 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7088 * These functions are only useful for the IA64 MCA handling, or kdb.
7090 * They can only be called when the whole system has been
7091 * stopped - every CPU needs to be quiescent, and no scheduling
7092 * activity can take place. Using them for anything else would
7093 * be a serious bug, and as a result, they aren't even visible
7094 * under any other configuration.
7098 * curr_task - return the current task for a given cpu.
7099 * @cpu: the processor in question.
7101 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7103 struct task_struct *curr_task(int cpu)
7105 return cpu_curr(cpu);
7108 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7112 * set_curr_task - set the current task for a given cpu.
7113 * @cpu: the processor in question.
7114 * @p: the task pointer to set.
7116 * Description: This function must only be used when non-maskable interrupts
7117 * are serviced on a separate stack. It allows the architecture to switch the
7118 * notion of the current task on a cpu in a non-blocking manner. This function
7119 * must be called with all CPU's synchronized, and interrupts disabled, the
7120 * and caller must save the original value of the current task (see
7121 * curr_task() above) and restore that value before reenabling interrupts and
7122 * re-starting the system.
7124 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7126 void set_curr_task(int cpu, struct task_struct *p)
7133 #ifdef CONFIG_CGROUP_SCHED
7134 /* task_group_lock serializes the addition/removal of task groups */
7135 static DEFINE_SPINLOCK(task_group_lock);
7137 static void free_sched_group(struct task_group *tg)
7139 free_fair_sched_group(tg);
7140 free_rt_sched_group(tg);
7145 /* allocate runqueue etc for a new task group */
7146 struct task_group *sched_create_group(struct task_group *parent)
7148 struct task_group *tg;
7149 unsigned long flags;
7151 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7153 return ERR_PTR(-ENOMEM);
7155 if (!alloc_fair_sched_group(tg, parent))
7158 if (!alloc_rt_sched_group(tg, parent))
7161 spin_lock_irqsave(&task_group_lock, flags);
7162 list_add_rcu(&tg->list, &task_groups);
7164 WARN_ON(!parent); /* root should already exist */
7166 tg->parent = parent;
7167 INIT_LIST_HEAD(&tg->children);
7168 list_add_rcu(&tg->siblings, &parent->children);
7169 spin_unlock_irqrestore(&task_group_lock, flags);
7174 free_sched_group(tg);
7175 return ERR_PTR(-ENOMEM);
7178 /* rcu callback to free various structures associated with a task group */
7179 static void free_sched_group_rcu(struct rcu_head *rhp)
7181 /* now it should be safe to free those cfs_rqs */
7182 free_sched_group(container_of(rhp, struct task_group, rcu));
7185 /* Destroy runqueue etc associated with a task group */
7186 void sched_destroy_group(struct task_group *tg)
7188 unsigned long flags;
7191 /* end participation in shares distribution */
7192 for_each_possible_cpu(i)
7193 unregister_fair_sched_group(tg, i);
7195 spin_lock_irqsave(&task_group_lock, flags);
7196 list_del_rcu(&tg->list);
7197 list_del_rcu(&tg->siblings);
7198 spin_unlock_irqrestore(&task_group_lock, flags);
7200 /* wait for possible concurrent references to cfs_rqs complete */
7201 call_rcu(&tg->rcu, free_sched_group_rcu);
7204 /* change task's runqueue when it moves between groups.
7205 * The caller of this function should have put the task in its new group
7206 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7207 * reflect its new group.
7209 void sched_move_task(struct task_struct *tsk)
7212 unsigned long flags;
7215 rq = task_rq_lock(tsk, &flags);
7217 running = task_current(rq, tsk);
7221 dequeue_task(rq, tsk, 0);
7222 if (unlikely(running))
7223 tsk->sched_class->put_prev_task(rq, tsk);
7225 #ifdef CONFIG_FAIR_GROUP_SCHED
7226 if (tsk->sched_class->task_move_group)
7227 tsk->sched_class->task_move_group(tsk, on_rq);
7230 set_task_rq(tsk, task_cpu(tsk));
7232 if (unlikely(running))
7233 tsk->sched_class->set_curr_task(rq);
7235 enqueue_task(rq, tsk, 0);
7237 task_rq_unlock(rq, tsk, &flags);
7239 #endif /* CONFIG_CGROUP_SCHED */
7241 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7242 static unsigned long to_ratio(u64 period, u64 runtime)
7244 if (runtime == RUNTIME_INF)
7247 return div64_u64(runtime << 20, period);
7251 #ifdef CONFIG_RT_GROUP_SCHED
7253 * Ensure that the real time constraints are schedulable.
7255 static DEFINE_MUTEX(rt_constraints_mutex);
7257 /* Must be called with tasklist_lock held */
7258 static inline int tg_has_rt_tasks(struct task_group *tg)
7260 struct task_struct *g, *p;
7262 do_each_thread(g, p) {
7263 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7265 } while_each_thread(g, p);
7270 struct rt_schedulable_data {
7271 struct task_group *tg;
7276 static int tg_rt_schedulable(struct task_group *tg, void *data)
7278 struct rt_schedulable_data *d = data;
7279 struct task_group *child;
7280 unsigned long total, sum = 0;
7281 u64 period, runtime;
7283 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7284 runtime = tg->rt_bandwidth.rt_runtime;
7287 period = d->rt_period;
7288 runtime = d->rt_runtime;
7292 * Cannot have more runtime than the period.
7294 if (runtime > period && runtime != RUNTIME_INF)
7298 * Ensure we don't starve existing RT tasks.
7300 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7303 total = to_ratio(period, runtime);
7306 * Nobody can have more than the global setting allows.
7308 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7312 * The sum of our children's runtime should not exceed our own.
7314 list_for_each_entry_rcu(child, &tg->children, siblings) {
7315 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7316 runtime = child->rt_bandwidth.rt_runtime;
7318 if (child == d->tg) {
7319 period = d->rt_period;
7320 runtime = d->rt_runtime;
7323 sum += to_ratio(period, runtime);
7332 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7336 struct rt_schedulable_data data = {
7338 .rt_period = period,
7339 .rt_runtime = runtime,
7343 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7349 static int tg_set_rt_bandwidth(struct task_group *tg,
7350 u64 rt_period, u64 rt_runtime)
7354 mutex_lock(&rt_constraints_mutex);
7355 read_lock(&tasklist_lock);
7356 err = __rt_schedulable(tg, rt_period, rt_runtime);
7360 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7361 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7362 tg->rt_bandwidth.rt_runtime = rt_runtime;
7364 for_each_possible_cpu(i) {
7365 struct rt_rq *rt_rq = tg->rt_rq[i];
7367 raw_spin_lock(&rt_rq->rt_runtime_lock);
7368 rt_rq->rt_runtime = rt_runtime;
7369 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7371 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7373 read_unlock(&tasklist_lock);
7374 mutex_unlock(&rt_constraints_mutex);
7379 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7381 u64 rt_runtime, rt_period;
7383 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7384 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7385 if (rt_runtime_us < 0)
7386 rt_runtime = RUNTIME_INF;
7388 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7391 long sched_group_rt_runtime(struct task_group *tg)
7395 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7398 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7399 do_div(rt_runtime_us, NSEC_PER_USEC);
7400 return rt_runtime_us;
7403 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7405 u64 rt_runtime, rt_period;
7407 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7408 rt_runtime = tg->rt_bandwidth.rt_runtime;
7413 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7416 long sched_group_rt_period(struct task_group *tg)
7420 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7421 do_div(rt_period_us, NSEC_PER_USEC);
7422 return rt_period_us;
7425 static int sched_rt_global_constraints(void)
7427 u64 runtime, period;
7430 if (sysctl_sched_rt_period <= 0)
7433 runtime = global_rt_runtime();
7434 period = global_rt_period();
7437 * Sanity check on the sysctl variables.
7439 if (runtime > period && runtime != RUNTIME_INF)
7442 mutex_lock(&rt_constraints_mutex);
7443 read_lock(&tasklist_lock);
7444 ret = __rt_schedulable(NULL, 0, 0);
7445 read_unlock(&tasklist_lock);
7446 mutex_unlock(&rt_constraints_mutex);
7451 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7453 /* Don't accept realtime tasks when there is no way for them to run */
7454 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7460 #else /* !CONFIG_RT_GROUP_SCHED */
7461 static int sched_rt_global_constraints(void)
7463 unsigned long flags;
7466 if (sysctl_sched_rt_period <= 0)
7470 * There's always some RT tasks in the root group
7471 * -- migration, kstopmachine etc..
7473 if (sysctl_sched_rt_runtime == 0)
7476 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7477 for_each_possible_cpu(i) {
7478 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7480 raw_spin_lock(&rt_rq->rt_runtime_lock);
7481 rt_rq->rt_runtime = global_rt_runtime();
7482 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7484 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7488 #endif /* CONFIG_RT_GROUP_SCHED */
7490 int sched_rt_handler(struct ctl_table *table, int write,
7491 void __user *buffer, size_t *lenp,
7495 int old_period, old_runtime;
7496 static DEFINE_MUTEX(mutex);
7499 old_period = sysctl_sched_rt_period;
7500 old_runtime = sysctl_sched_rt_runtime;
7502 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7504 if (!ret && write) {
7505 ret = sched_rt_global_constraints();
7507 sysctl_sched_rt_period = old_period;
7508 sysctl_sched_rt_runtime = old_runtime;
7510 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7511 def_rt_bandwidth.rt_period =
7512 ns_to_ktime(global_rt_period());
7515 mutex_unlock(&mutex);
7520 #ifdef CONFIG_CGROUP_SCHED
7522 /* return corresponding task_group object of a cgroup */
7523 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7525 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7526 struct task_group, css);
7529 static struct cgroup_subsys_state *
7530 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7532 struct task_group *tg, *parent;
7534 if (!cgrp->parent) {
7535 /* This is early initialization for the top cgroup */
7536 return &root_task_group.css;
7539 parent = cgroup_tg(cgrp->parent);
7540 tg = sched_create_group(parent);
7542 return ERR_PTR(-ENOMEM);
7548 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7550 struct task_group *tg = cgroup_tg(cgrp);
7552 sched_destroy_group(tg);
7555 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7556 struct cgroup_taskset *tset)
7558 struct task_struct *task;
7560 cgroup_taskset_for_each(task, cgrp, tset) {
7561 #ifdef CONFIG_RT_GROUP_SCHED
7562 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7565 /* We don't support RT-tasks being in separate groups */
7566 if (task->sched_class != &fair_sched_class)
7573 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7574 struct cgroup_taskset *tset)
7576 struct task_struct *task;
7578 cgroup_taskset_for_each(task, cgrp, tset)
7579 sched_move_task(task);
7583 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7584 struct cgroup *old_cgrp, struct task_struct *task)
7587 * cgroup_exit() is called in the copy_process() failure path.
7588 * Ignore this case since the task hasn't ran yet, this avoids
7589 * trying to poke a half freed task state from generic code.
7591 if (!(task->flags & PF_EXITING))
7594 sched_move_task(task);
7597 #ifdef CONFIG_FAIR_GROUP_SCHED
7598 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7601 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7604 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7606 struct task_group *tg = cgroup_tg(cgrp);
7608 return (u64) scale_load_down(tg->shares);
7611 #ifdef CONFIG_CFS_BANDWIDTH
7612 static DEFINE_MUTEX(cfs_constraints_mutex);
7614 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7615 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7617 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7619 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7621 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7622 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7624 if (tg == &root_task_group)
7628 * Ensure we have at some amount of bandwidth every period. This is
7629 * to prevent reaching a state of large arrears when throttled via
7630 * entity_tick() resulting in prolonged exit starvation.
7632 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7636 * Likewise, bound things on the otherside by preventing insane quota
7637 * periods. This also allows us to normalize in computing quota
7640 if (period > max_cfs_quota_period)
7643 mutex_lock(&cfs_constraints_mutex);
7644 ret = __cfs_schedulable(tg, period, quota);
7648 runtime_enabled = quota != RUNTIME_INF;
7649 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7650 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7651 raw_spin_lock_irq(&cfs_b->lock);
7652 cfs_b->period = ns_to_ktime(period);
7653 cfs_b->quota = quota;
7655 __refill_cfs_bandwidth_runtime(cfs_b);
7656 /* restart the period timer (if active) to handle new period expiry */
7657 if (runtime_enabled && cfs_b->timer_active) {
7658 /* force a reprogram */
7659 cfs_b->timer_active = 0;
7660 __start_cfs_bandwidth(cfs_b);
7662 raw_spin_unlock_irq(&cfs_b->lock);
7664 for_each_possible_cpu(i) {
7665 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7666 struct rq *rq = cfs_rq->rq;
7668 raw_spin_lock_irq(&rq->lock);
7669 cfs_rq->runtime_enabled = runtime_enabled;
7670 cfs_rq->runtime_remaining = 0;
7672 if (cfs_rq->throttled)
7673 unthrottle_cfs_rq(cfs_rq);
7674 raw_spin_unlock_irq(&rq->lock);
7677 mutex_unlock(&cfs_constraints_mutex);
7682 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7686 period = ktime_to_ns(tg->cfs_bandwidth.period);
7687 if (cfs_quota_us < 0)
7688 quota = RUNTIME_INF;
7690 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7692 return tg_set_cfs_bandwidth(tg, period, quota);
7695 long tg_get_cfs_quota(struct task_group *tg)
7699 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7702 quota_us = tg->cfs_bandwidth.quota;
7703 do_div(quota_us, NSEC_PER_USEC);
7708 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7712 period = (u64)cfs_period_us * NSEC_PER_USEC;
7713 quota = tg->cfs_bandwidth.quota;
7715 return tg_set_cfs_bandwidth(tg, period, quota);
7718 long tg_get_cfs_period(struct task_group *tg)
7722 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7723 do_div(cfs_period_us, NSEC_PER_USEC);
7725 return cfs_period_us;
7728 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7730 return tg_get_cfs_quota(cgroup_tg(cgrp));
7733 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7736 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7739 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7741 return tg_get_cfs_period(cgroup_tg(cgrp));
7744 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7747 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7750 struct cfs_schedulable_data {
7751 struct task_group *tg;
7756 * normalize group quota/period to be quota/max_period
7757 * note: units are usecs
7759 static u64 normalize_cfs_quota(struct task_group *tg,
7760 struct cfs_schedulable_data *d)
7768 period = tg_get_cfs_period(tg);
7769 quota = tg_get_cfs_quota(tg);
7772 /* note: these should typically be equivalent */
7773 if (quota == RUNTIME_INF || quota == -1)
7776 return to_ratio(period, quota);
7779 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7781 struct cfs_schedulable_data *d = data;
7782 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7783 s64 quota = 0, parent_quota = -1;
7786 quota = RUNTIME_INF;
7788 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7790 quota = normalize_cfs_quota(tg, d);
7791 parent_quota = parent_b->hierarchal_quota;
7794 * ensure max(child_quota) <= parent_quota, inherit when no
7797 if (quota == RUNTIME_INF)
7798 quota = parent_quota;
7799 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7802 cfs_b->hierarchal_quota = quota;
7807 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7810 struct cfs_schedulable_data data = {
7816 if (quota != RUNTIME_INF) {
7817 do_div(data.period, NSEC_PER_USEC);
7818 do_div(data.quota, NSEC_PER_USEC);
7822 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7828 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7829 struct cgroup_map_cb *cb)
7831 struct task_group *tg = cgroup_tg(cgrp);
7832 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7834 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7835 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7836 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7840 #endif /* CONFIG_CFS_BANDWIDTH */
7841 #endif /* CONFIG_FAIR_GROUP_SCHED */
7843 #ifdef CONFIG_RT_GROUP_SCHED
7844 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7847 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7850 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7852 return sched_group_rt_runtime(cgroup_tg(cgrp));
7855 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7858 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7861 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7863 return sched_group_rt_period(cgroup_tg(cgrp));
7865 #endif /* CONFIG_RT_GROUP_SCHED */
7867 static struct cftype cpu_files[] = {
7868 #ifdef CONFIG_FAIR_GROUP_SCHED
7871 .read_u64 = cpu_shares_read_u64,
7872 .write_u64 = cpu_shares_write_u64,
7875 #ifdef CONFIG_CFS_BANDWIDTH
7877 .name = "cfs_quota_us",
7878 .read_s64 = cpu_cfs_quota_read_s64,
7879 .write_s64 = cpu_cfs_quota_write_s64,
7882 .name = "cfs_period_us",
7883 .read_u64 = cpu_cfs_period_read_u64,
7884 .write_u64 = cpu_cfs_period_write_u64,
7888 .read_map = cpu_stats_show,
7891 #ifdef CONFIG_RT_GROUP_SCHED
7893 .name = "rt_runtime_us",
7894 .read_s64 = cpu_rt_runtime_read,
7895 .write_s64 = cpu_rt_runtime_write,
7898 .name = "rt_period_us",
7899 .read_u64 = cpu_rt_period_read_uint,
7900 .write_u64 = cpu_rt_period_write_uint,
7905 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7907 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7910 struct cgroup_subsys cpu_cgroup_subsys = {
7912 .create = cpu_cgroup_create,
7913 .destroy = cpu_cgroup_destroy,
7914 .can_attach = cpu_cgroup_can_attach,
7915 .attach = cpu_cgroup_attach,
7916 .exit = cpu_cgroup_exit,
7917 .populate = cpu_cgroup_populate,
7918 .subsys_id = cpu_cgroup_subsys_id,
7922 #endif /* CONFIG_CGROUP_SCHED */
7924 #ifdef CONFIG_CGROUP_CPUACCT
7927 * CPU accounting code for task groups.
7929 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7930 * (balbir@in.ibm.com).
7933 /* create a new cpu accounting group */
7934 static struct cgroup_subsys_state *cpuacct_create(
7935 struct cgroup_subsys *ss, struct cgroup *cgrp)
7940 return &root_cpuacct.css;
7942 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7946 ca->cpuusage = alloc_percpu(u64);
7950 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7952 goto out_free_cpuusage;
7957 free_percpu(ca->cpuusage);
7961 return ERR_PTR(-ENOMEM);
7964 /* destroy an existing cpu accounting group */
7966 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7968 struct cpuacct *ca = cgroup_ca(cgrp);
7970 free_percpu(ca->cpustat);
7971 free_percpu(ca->cpuusage);
7975 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7977 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7980 #ifndef CONFIG_64BIT
7982 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7984 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7986 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7994 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7996 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7998 #ifndef CONFIG_64BIT
8000 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8002 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8004 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8010 /* return total cpu usage (in nanoseconds) of a group */
8011 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8013 struct cpuacct *ca = cgroup_ca(cgrp);
8014 u64 totalcpuusage = 0;
8017 for_each_present_cpu(i)
8018 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8020 return totalcpuusage;
8023 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8026 struct cpuacct *ca = cgroup_ca(cgrp);
8035 for_each_present_cpu(i)
8036 cpuacct_cpuusage_write(ca, i, 0);
8042 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8045 struct cpuacct *ca = cgroup_ca(cgroup);
8049 for_each_present_cpu(i) {
8050 percpu = cpuacct_cpuusage_read(ca, i);
8051 seq_printf(m, "%llu ", (unsigned long long) percpu);
8053 seq_printf(m, "\n");
8057 static const char *cpuacct_stat_desc[] = {
8058 [CPUACCT_STAT_USER] = "user",
8059 [CPUACCT_STAT_SYSTEM] = "system",
8062 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8063 struct cgroup_map_cb *cb)
8065 struct cpuacct *ca = cgroup_ca(cgrp);
8069 for_each_online_cpu(cpu) {
8070 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8071 val += kcpustat->cpustat[CPUTIME_USER];
8072 val += kcpustat->cpustat[CPUTIME_NICE];
8074 val = cputime64_to_clock_t(val);
8075 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8078 for_each_online_cpu(cpu) {
8079 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8080 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8081 val += kcpustat->cpustat[CPUTIME_IRQ];
8082 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8085 val = cputime64_to_clock_t(val);
8086 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8091 static struct cftype files[] = {
8094 .read_u64 = cpuusage_read,
8095 .write_u64 = cpuusage_write,
8098 .name = "usage_percpu",
8099 .read_seq_string = cpuacct_percpu_seq_read,
8103 .read_map = cpuacct_stats_show,
8107 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8109 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8113 * charge this task's execution time to its accounting group.
8115 * called with rq->lock held.
8117 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8122 if (unlikely(!cpuacct_subsys.active))
8125 cpu = task_cpu(tsk);
8131 for (; ca; ca = parent_ca(ca)) {
8132 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8133 *cpuusage += cputime;
8139 struct cgroup_subsys cpuacct_subsys = {
8141 .create = cpuacct_create,
8142 .destroy = cpuacct_destroy,
8143 .populate = cpuacct_populate,
8144 .subsys_id = cpuacct_subsys_id,
8146 #endif /* CONFIG_CGROUP_CPUACCT */