4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
76 #include <asm/irq_regs.h>
77 #ifdef CONFIG_PARAVIRT
78 #include <asm/paravirt.h>
82 #include "../workqueue_sched.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
87 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
90 ktime_t soft, hard, now;
93 if (hrtimer_active(period_timer))
96 now = hrtimer_cb_get_time(period_timer);
97 hrtimer_forward(period_timer, now, period);
99 soft = hrtimer_get_softexpires(period_timer);
100 hard = hrtimer_get_expires(period_timer);
101 delta = ktime_to_ns(ktime_sub(hard, soft));
102 __hrtimer_start_range_ns(period_timer, soft, delta,
103 HRTIMER_MODE_ABS_PINNED, 0);
107 DEFINE_MUTEX(sched_domains_mutex);
108 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
110 static void update_rq_clock_task(struct rq *rq, s64 delta);
112 void update_rq_clock(struct rq *rq)
116 if (rq->skip_clock_update > 0)
119 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
121 update_rq_clock_task(rq, delta);
125 * Debugging: various feature bits
128 #define SCHED_FEAT(name, enabled) \
129 (1UL << __SCHED_FEAT_##name) * enabled |
131 const_debug unsigned int sysctl_sched_features =
132 #include "features.h"
137 #ifdef CONFIG_SCHED_DEBUG
138 #define SCHED_FEAT(name, enabled) \
141 static __read_mostly char *sched_feat_names[] = {
142 #include "features.h"
148 static int sched_feat_show(struct seq_file *m, void *v)
152 for (i = 0; i < __SCHED_FEAT_NR; i++) {
153 if (!(sysctl_sched_features & (1UL << i)))
155 seq_printf(m, "%s ", sched_feat_names[i]);
162 #ifdef HAVE_JUMP_LABEL
164 #define jump_label_key__true jump_label_key_enabled
165 #define jump_label_key__false jump_label_key_disabled
167 #define SCHED_FEAT(name, enabled) \
168 jump_label_key__##enabled ,
170 struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = {
171 #include "features.h"
176 static void sched_feat_disable(int i)
178 if (jump_label_enabled(&sched_feat_keys[i]))
179 jump_label_dec(&sched_feat_keys[i]);
182 static void sched_feat_enable(int i)
184 if (!jump_label_enabled(&sched_feat_keys[i]))
185 jump_label_inc(&sched_feat_keys[i]);
188 static void sched_feat_disable(int i) { };
189 static void sched_feat_enable(int i) { };
190 #endif /* HAVE_JUMP_LABEL */
193 sched_feat_write(struct file *filp, const char __user *ubuf,
194 size_t cnt, loff_t *ppos)
204 if (copy_from_user(&buf, ubuf, cnt))
210 if (strncmp(cmp, "NO_", 3) == 0) {
215 for (i = 0; i < __SCHED_FEAT_NR; i++) {
216 if (strcmp(cmp, sched_feat_names[i]) == 0) {
218 sysctl_sched_features &= ~(1UL << i);
219 sched_feat_disable(i);
221 sysctl_sched_features |= (1UL << i);
222 sched_feat_enable(i);
228 if (i == __SCHED_FEAT_NR)
236 static int sched_feat_open(struct inode *inode, struct file *filp)
238 return single_open(filp, sched_feat_show, NULL);
241 static const struct file_operations sched_feat_fops = {
242 .open = sched_feat_open,
243 .write = sched_feat_write,
246 .release = single_release,
249 static __init int sched_init_debug(void)
251 debugfs_create_file("sched_features", 0644, NULL, NULL,
256 late_initcall(sched_init_debug);
257 #endif /* CONFIG_SCHED_DEBUG */
260 * Number of tasks to iterate in a single balance run.
261 * Limited because this is done with IRQs disabled.
263 const_debug unsigned int sysctl_sched_nr_migrate = 32;
266 * period over which we average the RT time consumption, measured
271 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
274 * period over which we measure -rt task cpu usage in us.
277 unsigned int sysctl_sched_rt_period = 1000000;
279 __read_mostly int scheduler_running;
282 * part of the period that we allow rt tasks to run in us.
285 int sysctl_sched_rt_runtime = 950000;
290 * __task_rq_lock - lock the rq @p resides on.
292 static inline struct rq *__task_rq_lock(struct task_struct *p)
297 lockdep_assert_held(&p->pi_lock);
301 raw_spin_lock(&rq->lock);
302 if (likely(rq == task_rq(p)))
304 raw_spin_unlock(&rq->lock);
309 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
311 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
312 __acquires(p->pi_lock)
318 raw_spin_lock_irqsave(&p->pi_lock, *flags);
320 raw_spin_lock(&rq->lock);
321 if (likely(rq == task_rq(p)))
323 raw_spin_unlock(&rq->lock);
324 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
328 static void __task_rq_unlock(struct rq *rq)
331 raw_spin_unlock(&rq->lock);
335 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
337 __releases(p->pi_lock)
339 raw_spin_unlock(&rq->lock);
340 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
344 * this_rq_lock - lock this runqueue and disable interrupts.
346 static struct rq *this_rq_lock(void)
353 raw_spin_lock(&rq->lock);
358 #ifdef CONFIG_SCHED_HRTICK
360 * Use HR-timers to deliver accurate preemption points.
362 * Its all a bit involved since we cannot program an hrt while holding the
363 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
366 * When we get rescheduled we reprogram the hrtick_timer outside of the
370 static void hrtick_clear(struct rq *rq)
372 if (hrtimer_active(&rq->hrtick_timer))
373 hrtimer_cancel(&rq->hrtick_timer);
377 * High-resolution timer tick.
378 * Runs from hardirq context with interrupts disabled.
380 static enum hrtimer_restart hrtick(struct hrtimer *timer)
382 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
384 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
386 raw_spin_lock(&rq->lock);
388 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
389 raw_spin_unlock(&rq->lock);
391 return HRTIMER_NORESTART;
396 * called from hardirq (IPI) context
398 static void __hrtick_start(void *arg)
402 raw_spin_lock(&rq->lock);
403 hrtimer_restart(&rq->hrtick_timer);
404 rq->hrtick_csd_pending = 0;
405 raw_spin_unlock(&rq->lock);
409 * Called to set the hrtick timer state.
411 * called with rq->lock held and irqs disabled
413 void hrtick_start(struct rq *rq, u64 delay)
415 struct hrtimer *timer = &rq->hrtick_timer;
416 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
418 hrtimer_set_expires(timer, time);
420 if (rq == this_rq()) {
421 hrtimer_restart(timer);
422 } else if (!rq->hrtick_csd_pending) {
423 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
424 rq->hrtick_csd_pending = 1;
429 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
431 int cpu = (int)(long)hcpu;
434 case CPU_UP_CANCELED:
435 case CPU_UP_CANCELED_FROZEN:
436 case CPU_DOWN_PREPARE:
437 case CPU_DOWN_PREPARE_FROZEN:
439 case CPU_DEAD_FROZEN:
440 hrtick_clear(cpu_rq(cpu));
447 static __init void init_hrtick(void)
449 hotcpu_notifier(hotplug_hrtick, 0);
453 * Called to set the hrtick timer state.
455 * called with rq->lock held and irqs disabled
457 void hrtick_start(struct rq *rq, u64 delay)
459 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
460 HRTIMER_MODE_REL_PINNED, 0);
463 static inline void init_hrtick(void)
466 #endif /* CONFIG_SMP */
468 static void init_rq_hrtick(struct rq *rq)
471 rq->hrtick_csd_pending = 0;
473 rq->hrtick_csd.flags = 0;
474 rq->hrtick_csd.func = __hrtick_start;
475 rq->hrtick_csd.info = rq;
478 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
479 rq->hrtick_timer.function = hrtick;
481 #else /* CONFIG_SCHED_HRTICK */
482 static inline void hrtick_clear(struct rq *rq)
486 static inline void init_rq_hrtick(struct rq *rq)
490 static inline void init_hrtick(void)
493 #endif /* CONFIG_SCHED_HRTICK */
496 * resched_task - mark a task 'to be rescheduled now'.
498 * On UP this means the setting of the need_resched flag, on SMP it
499 * might also involve a cross-CPU call to trigger the scheduler on
504 #ifndef tsk_is_polling
505 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
508 void resched_task(struct task_struct *p)
512 assert_raw_spin_locked(&task_rq(p)->lock);
514 if (test_tsk_need_resched(p))
517 set_tsk_need_resched(p);
520 if (cpu == smp_processor_id())
523 /* NEED_RESCHED must be visible before we test polling */
525 if (!tsk_is_polling(p))
526 smp_send_reschedule(cpu);
529 void resched_cpu(int cpu)
531 struct rq *rq = cpu_rq(cpu);
534 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
536 resched_task(cpu_curr(cpu));
537 raw_spin_unlock_irqrestore(&rq->lock, flags);
542 * In the semi idle case, use the nearest busy cpu for migrating timers
543 * from an idle cpu. This is good for power-savings.
545 * We don't do similar optimization for completely idle system, as
546 * selecting an idle cpu will add more delays to the timers than intended
547 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
549 int get_nohz_timer_target(void)
551 int cpu = smp_processor_id();
553 struct sched_domain *sd;
556 for_each_domain(cpu, sd) {
557 for_each_cpu(i, sched_domain_span(sd)) {
569 * When add_timer_on() enqueues a timer into the timer wheel of an
570 * idle CPU then this timer might expire before the next timer event
571 * which is scheduled to wake up that CPU. In case of a completely
572 * idle system the next event might even be infinite time into the
573 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
574 * leaves the inner idle loop so the newly added timer is taken into
575 * account when the CPU goes back to idle and evaluates the timer
576 * wheel for the next timer event.
578 void wake_up_idle_cpu(int cpu)
580 struct rq *rq = cpu_rq(cpu);
582 if (cpu == smp_processor_id())
586 * This is safe, as this function is called with the timer
587 * wheel base lock of (cpu) held. When the CPU is on the way
588 * to idle and has not yet set rq->curr to idle then it will
589 * be serialized on the timer wheel base lock and take the new
590 * timer into account automatically.
592 if (rq->curr != rq->idle)
596 * We can set TIF_RESCHED on the idle task of the other CPU
597 * lockless. The worst case is that the other CPU runs the
598 * idle task through an additional NOOP schedule()
600 set_tsk_need_resched(rq->idle);
602 /* NEED_RESCHED must be visible before we test polling */
604 if (!tsk_is_polling(rq->idle))
605 smp_send_reschedule(cpu);
608 static inline bool got_nohz_idle_kick(void)
610 int cpu = smp_processor_id();
611 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
614 #else /* CONFIG_NO_HZ */
616 static inline bool got_nohz_idle_kick(void)
621 #endif /* CONFIG_NO_HZ */
623 void sched_avg_update(struct rq *rq)
625 s64 period = sched_avg_period();
627 while ((s64)(rq->clock - rq->age_stamp) > period) {
629 * Inline assembly required to prevent the compiler
630 * optimising this loop into a divmod call.
631 * See __iter_div_u64_rem() for another example of this.
633 asm("" : "+rm" (rq->age_stamp));
634 rq->age_stamp += period;
639 #else /* !CONFIG_SMP */
640 void resched_task(struct task_struct *p)
642 assert_raw_spin_locked(&task_rq(p)->lock);
643 set_tsk_need_resched(p);
645 #endif /* CONFIG_SMP */
647 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
648 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
650 * Iterate task_group tree rooted at *from, calling @down when first entering a
651 * node and @up when leaving it for the final time.
653 * Caller must hold rcu_lock or sufficient equivalent.
655 int walk_tg_tree_from(struct task_group *from,
656 tg_visitor down, tg_visitor up, void *data)
658 struct task_group *parent, *child;
664 ret = (*down)(parent, data);
667 list_for_each_entry_rcu(child, &parent->children, siblings) {
674 ret = (*up)(parent, data);
675 if (ret || parent == from)
679 parent = parent->parent;
686 int tg_nop(struct task_group *tg, void *data)
692 void update_cpu_load(struct rq *this_rq);
694 static void set_load_weight(struct task_struct *p)
696 int prio = p->static_prio - MAX_RT_PRIO;
697 struct load_weight *load = &p->se.load;
700 * SCHED_IDLE tasks get minimal weight:
702 if (p->policy == SCHED_IDLE) {
703 load->weight = scale_load(WEIGHT_IDLEPRIO);
704 load->inv_weight = WMULT_IDLEPRIO;
708 load->weight = scale_load(prio_to_weight[prio]);
709 load->inv_weight = prio_to_wmult[prio];
712 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
715 sched_info_queued(p);
716 p->sched_class->enqueue_task(rq, p, flags);
719 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
722 sched_info_dequeued(p);
723 p->sched_class->dequeue_task(rq, p, flags);
727 * activate_task - move a task to the runqueue.
729 void activate_task(struct rq *rq, struct task_struct *p, int flags)
731 if (task_contributes_to_load(p))
732 rq->nr_uninterruptible--;
734 enqueue_task(rq, p, flags);
738 * deactivate_task - remove a task from the runqueue.
740 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
742 if (task_contributes_to_load(p))
743 rq->nr_uninterruptible++;
745 dequeue_task(rq, p, flags);
748 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
751 * There are no locks covering percpu hardirq/softirq time.
752 * They are only modified in account_system_vtime, on corresponding CPU
753 * with interrupts disabled. So, writes are safe.
754 * They are read and saved off onto struct rq in update_rq_clock().
755 * This may result in other CPU reading this CPU's irq time and can
756 * race with irq/account_system_vtime on this CPU. We would either get old
757 * or new value with a side effect of accounting a slice of irq time to wrong
758 * task when irq is in progress while we read rq->clock. That is a worthy
759 * compromise in place of having locks on each irq in account_system_time.
761 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
762 static DEFINE_PER_CPU(u64, cpu_softirq_time);
764 static DEFINE_PER_CPU(u64, irq_start_time);
765 static int sched_clock_irqtime;
767 void enable_sched_clock_irqtime(void)
769 sched_clock_irqtime = 1;
772 void disable_sched_clock_irqtime(void)
774 sched_clock_irqtime = 0;
778 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
780 static inline void irq_time_write_begin(void)
782 __this_cpu_inc(irq_time_seq.sequence);
786 static inline void irq_time_write_end(void)
789 __this_cpu_inc(irq_time_seq.sequence);
792 static inline u64 irq_time_read(int cpu)
798 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
799 irq_time = per_cpu(cpu_softirq_time, cpu) +
800 per_cpu(cpu_hardirq_time, cpu);
801 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
805 #else /* CONFIG_64BIT */
806 static inline void irq_time_write_begin(void)
810 static inline void irq_time_write_end(void)
814 static inline u64 irq_time_read(int cpu)
816 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
818 #endif /* CONFIG_64BIT */
821 * Called before incrementing preempt_count on {soft,}irq_enter
822 * and before decrementing preempt_count on {soft,}irq_exit.
824 void account_system_vtime(struct task_struct *curr)
830 if (!sched_clock_irqtime)
833 local_irq_save(flags);
835 cpu = smp_processor_id();
836 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
837 __this_cpu_add(irq_start_time, delta);
839 irq_time_write_begin();
841 * We do not account for softirq time from ksoftirqd here.
842 * We want to continue accounting softirq time to ksoftirqd thread
843 * in that case, so as not to confuse scheduler with a special task
844 * that do not consume any time, but still wants to run.
847 __this_cpu_add(cpu_hardirq_time, delta);
848 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
849 __this_cpu_add(cpu_softirq_time, delta);
851 irq_time_write_end();
852 local_irq_restore(flags);
854 EXPORT_SYMBOL_GPL(account_system_vtime);
856 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
858 #ifdef CONFIG_PARAVIRT
859 static inline u64 steal_ticks(u64 steal)
861 if (unlikely(steal > NSEC_PER_SEC))
862 return div_u64(steal, TICK_NSEC);
864 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
868 static void update_rq_clock_task(struct rq *rq, s64 delta)
871 * In theory, the compile should just see 0 here, and optimize out the call
872 * to sched_rt_avg_update. But I don't trust it...
874 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
875 s64 steal = 0, irq_delta = 0;
877 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
878 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
881 * Since irq_time is only updated on {soft,}irq_exit, we might run into
882 * this case when a previous update_rq_clock() happened inside a
885 * When this happens, we stop ->clock_task and only update the
886 * prev_irq_time stamp to account for the part that fit, so that a next
887 * update will consume the rest. This ensures ->clock_task is
890 * It does however cause some slight miss-attribution of {soft,}irq
891 * time, a more accurate solution would be to update the irq_time using
892 * the current rq->clock timestamp, except that would require using
895 if (irq_delta > delta)
898 rq->prev_irq_time += irq_delta;
901 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
902 if (static_branch((¶virt_steal_rq_enabled))) {
905 steal = paravirt_steal_clock(cpu_of(rq));
906 steal -= rq->prev_steal_time_rq;
908 if (unlikely(steal > delta))
911 st = steal_ticks(steal);
912 steal = st * TICK_NSEC;
914 rq->prev_steal_time_rq += steal;
920 rq->clock_task += delta;
922 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
923 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
924 sched_rt_avg_update(rq, irq_delta + steal);
928 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
929 static int irqtime_account_hi_update(void)
931 u64 *cpustat = kcpustat_this_cpu->cpustat;
936 local_irq_save(flags);
937 latest_ns = this_cpu_read(cpu_hardirq_time);
938 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
940 local_irq_restore(flags);
944 static int irqtime_account_si_update(void)
946 u64 *cpustat = kcpustat_this_cpu->cpustat;
951 local_irq_save(flags);
952 latest_ns = this_cpu_read(cpu_softirq_time);
953 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
955 local_irq_restore(flags);
959 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
961 #define sched_clock_irqtime (0)
965 void sched_set_stop_task(int cpu, struct task_struct *stop)
967 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
968 struct task_struct *old_stop = cpu_rq(cpu)->stop;
972 * Make it appear like a SCHED_FIFO task, its something
973 * userspace knows about and won't get confused about.
975 * Also, it will make PI more or less work without too
976 * much confusion -- but then, stop work should not
977 * rely on PI working anyway.
979 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
981 stop->sched_class = &stop_sched_class;
984 cpu_rq(cpu)->stop = stop;
988 * Reset it back to a normal scheduling class so that
989 * it can die in pieces.
991 old_stop->sched_class = &rt_sched_class;
996 * __normal_prio - return the priority that is based on the static prio
998 static inline int __normal_prio(struct task_struct *p)
1000 return p->static_prio;
1004 * Calculate the expected normal priority: i.e. priority
1005 * without taking RT-inheritance into account. Might be
1006 * boosted by interactivity modifiers. Changes upon fork,
1007 * setprio syscalls, and whenever the interactivity
1008 * estimator recalculates.
1010 static inline int normal_prio(struct task_struct *p)
1014 if (task_has_rt_policy(p))
1015 prio = MAX_RT_PRIO-1 - p->rt_priority;
1017 prio = __normal_prio(p);
1022 * Calculate the current priority, i.e. the priority
1023 * taken into account by the scheduler. This value might
1024 * be boosted by RT tasks, or might be boosted by
1025 * interactivity modifiers. Will be RT if the task got
1026 * RT-boosted. If not then it returns p->normal_prio.
1028 static int effective_prio(struct task_struct *p)
1030 p->normal_prio = normal_prio(p);
1032 * If we are RT tasks or we were boosted to RT priority,
1033 * keep the priority unchanged. Otherwise, update priority
1034 * to the normal priority:
1036 if (!rt_prio(p->prio))
1037 return p->normal_prio;
1042 * task_curr - is this task currently executing on a CPU?
1043 * @p: the task in question.
1045 inline int task_curr(const struct task_struct *p)
1047 return cpu_curr(task_cpu(p)) == p;
1050 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1051 const struct sched_class *prev_class,
1054 if (prev_class != p->sched_class) {
1055 if (prev_class->switched_from)
1056 prev_class->switched_from(rq, p);
1057 p->sched_class->switched_to(rq, p);
1058 } else if (oldprio != p->prio)
1059 p->sched_class->prio_changed(rq, p, oldprio);
1062 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1064 const struct sched_class *class;
1066 if (p->sched_class == rq->curr->sched_class) {
1067 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1069 for_each_class(class) {
1070 if (class == rq->curr->sched_class)
1072 if (class == p->sched_class) {
1073 resched_task(rq->curr);
1080 * A queue event has occurred, and we're going to schedule. In
1081 * this case, we can save a useless back to back clock update.
1083 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1084 rq->skip_clock_update = 1;
1088 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1090 #ifdef CONFIG_SCHED_DEBUG
1092 * We should never call set_task_cpu() on a blocked task,
1093 * ttwu() will sort out the placement.
1095 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1096 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1098 #ifdef CONFIG_LOCKDEP
1100 * The caller should hold either p->pi_lock or rq->lock, when changing
1101 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1103 * sched_move_task() holds both and thus holding either pins the cgroup,
1104 * see set_task_rq().
1106 * Furthermore, all task_rq users should acquire both locks, see
1109 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1110 lockdep_is_held(&task_rq(p)->lock)));
1114 trace_sched_migrate_task(p, new_cpu);
1116 if (task_cpu(p) != new_cpu) {
1117 p->se.nr_migrations++;
1118 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1121 __set_task_cpu(p, new_cpu);
1124 struct migration_arg {
1125 struct task_struct *task;
1129 static int migration_cpu_stop(void *data);
1132 * wait_task_inactive - wait for a thread to unschedule.
1134 * If @match_state is nonzero, it's the @p->state value just checked and
1135 * not expected to change. If it changes, i.e. @p might have woken up,
1136 * then return zero. When we succeed in waiting for @p to be off its CPU,
1137 * we return a positive number (its total switch count). If a second call
1138 * a short while later returns the same number, the caller can be sure that
1139 * @p has remained unscheduled the whole time.
1141 * The caller must ensure that the task *will* unschedule sometime soon,
1142 * else this function might spin for a *long* time. This function can't
1143 * be called with interrupts off, or it may introduce deadlock with
1144 * smp_call_function() if an IPI is sent by the same process we are
1145 * waiting to become inactive.
1147 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1149 unsigned long flags;
1156 * We do the initial early heuristics without holding
1157 * any task-queue locks at all. We'll only try to get
1158 * the runqueue lock when things look like they will
1164 * If the task is actively running on another CPU
1165 * still, just relax and busy-wait without holding
1168 * NOTE! Since we don't hold any locks, it's not
1169 * even sure that "rq" stays as the right runqueue!
1170 * But we don't care, since "task_running()" will
1171 * return false if the runqueue has changed and p
1172 * is actually now running somewhere else!
1174 while (task_running(rq, p)) {
1175 if (match_state && unlikely(p->state != match_state))
1181 * Ok, time to look more closely! We need the rq
1182 * lock now, to be *sure*. If we're wrong, we'll
1183 * just go back and repeat.
1185 rq = task_rq_lock(p, &flags);
1186 trace_sched_wait_task(p);
1187 running = task_running(rq, p);
1190 if (!match_state || p->state == match_state)
1191 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1192 task_rq_unlock(rq, p, &flags);
1195 * If it changed from the expected state, bail out now.
1197 if (unlikely(!ncsw))
1201 * Was it really running after all now that we
1202 * checked with the proper locks actually held?
1204 * Oops. Go back and try again..
1206 if (unlikely(running)) {
1212 * It's not enough that it's not actively running,
1213 * it must be off the runqueue _entirely_, and not
1216 * So if it was still runnable (but just not actively
1217 * running right now), it's preempted, and we should
1218 * yield - it could be a while.
1220 if (unlikely(on_rq)) {
1221 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1223 set_current_state(TASK_UNINTERRUPTIBLE);
1224 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1229 * Ahh, all good. It wasn't running, and it wasn't
1230 * runnable, which means that it will never become
1231 * running in the future either. We're all done!
1240 * kick_process - kick a running thread to enter/exit the kernel
1241 * @p: the to-be-kicked thread
1243 * Cause a process which is running on another CPU to enter
1244 * kernel-mode, without any delay. (to get signals handled.)
1246 * NOTE: this function doesn't have to take the runqueue lock,
1247 * because all it wants to ensure is that the remote task enters
1248 * the kernel. If the IPI races and the task has been migrated
1249 * to another CPU then no harm is done and the purpose has been
1252 void kick_process(struct task_struct *p)
1258 if ((cpu != smp_processor_id()) && task_curr(p))
1259 smp_send_reschedule(cpu);
1262 EXPORT_SYMBOL_GPL(kick_process);
1263 #endif /* CONFIG_SMP */
1267 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1269 static int select_fallback_rq(int cpu, struct task_struct *p)
1272 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1274 /* Look for allowed, online CPU in same node. */
1275 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1279 /* Any allowed, online CPU? */
1280 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask);
1281 if (dest_cpu < nr_cpu_ids)
1284 /* No more Mr. Nice Guy. */
1285 dest_cpu = cpuset_cpus_allowed_fallback(p);
1287 * Don't tell them about moving exiting tasks or
1288 * kernel threads (both mm NULL), since they never
1291 if (p->mm && printk_ratelimit()) {
1292 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
1293 task_pid_nr(p), p->comm, cpu);
1300 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1303 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1305 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1308 * In order not to call set_task_cpu() on a blocking task we need
1309 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1312 * Since this is common to all placement strategies, this lives here.
1314 * [ this allows ->select_task() to simply return task_cpu(p) and
1315 * not worry about this generic constraint ]
1317 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1319 cpu = select_fallback_rq(task_cpu(p), p);
1324 static void update_avg(u64 *avg, u64 sample)
1326 s64 diff = sample - *avg;
1332 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1334 #ifdef CONFIG_SCHEDSTATS
1335 struct rq *rq = this_rq();
1338 int this_cpu = smp_processor_id();
1340 if (cpu == this_cpu) {
1341 schedstat_inc(rq, ttwu_local);
1342 schedstat_inc(p, se.statistics.nr_wakeups_local);
1344 struct sched_domain *sd;
1346 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1348 for_each_domain(this_cpu, sd) {
1349 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1350 schedstat_inc(sd, ttwu_wake_remote);
1357 if (wake_flags & WF_MIGRATED)
1358 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1360 #endif /* CONFIG_SMP */
1362 schedstat_inc(rq, ttwu_count);
1363 schedstat_inc(p, se.statistics.nr_wakeups);
1365 if (wake_flags & WF_SYNC)
1366 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1368 #endif /* CONFIG_SCHEDSTATS */
1371 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1373 activate_task(rq, p, en_flags);
1376 /* if a worker is waking up, notify workqueue */
1377 if (p->flags & PF_WQ_WORKER)
1378 wq_worker_waking_up(p, cpu_of(rq));
1382 * Mark the task runnable and perform wakeup-preemption.
1385 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1387 trace_sched_wakeup(p, true);
1388 check_preempt_curr(rq, p, wake_flags);
1390 p->state = TASK_RUNNING;
1392 if (p->sched_class->task_woken)
1393 p->sched_class->task_woken(rq, p);
1395 if (rq->idle_stamp) {
1396 u64 delta = rq->clock - rq->idle_stamp;
1397 u64 max = 2*sysctl_sched_migration_cost;
1402 update_avg(&rq->avg_idle, delta);
1409 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1412 if (p->sched_contributes_to_load)
1413 rq->nr_uninterruptible--;
1416 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1417 ttwu_do_wakeup(rq, p, wake_flags);
1421 * Called in case the task @p isn't fully descheduled from its runqueue,
1422 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1423 * since all we need to do is flip p->state to TASK_RUNNING, since
1424 * the task is still ->on_rq.
1426 static int ttwu_remote(struct task_struct *p, int wake_flags)
1431 rq = __task_rq_lock(p);
1433 ttwu_do_wakeup(rq, p, wake_flags);
1436 __task_rq_unlock(rq);
1442 static void sched_ttwu_pending(void)
1444 struct rq *rq = this_rq();
1445 struct llist_node *llist = llist_del_all(&rq->wake_list);
1446 struct task_struct *p;
1448 raw_spin_lock(&rq->lock);
1451 p = llist_entry(llist, struct task_struct, wake_entry);
1452 llist = llist_next(llist);
1453 ttwu_do_activate(rq, p, 0);
1456 raw_spin_unlock(&rq->lock);
1459 void scheduler_ipi(void)
1461 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1465 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1466 * traditionally all their work was done from the interrupt return
1467 * path. Now that we actually do some work, we need to make sure
1470 * Some archs already do call them, luckily irq_enter/exit nest
1473 * Arguably we should visit all archs and update all handlers,
1474 * however a fair share of IPIs are still resched only so this would
1475 * somewhat pessimize the simple resched case.
1478 sched_ttwu_pending();
1481 * Check if someone kicked us for doing the nohz idle load balance.
1483 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1484 this_rq()->idle_balance = 1;
1485 raise_softirq_irqoff(SCHED_SOFTIRQ);
1490 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1492 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1493 smp_send_reschedule(cpu);
1496 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1497 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1502 rq = __task_rq_lock(p);
1504 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1505 ttwu_do_wakeup(rq, p, wake_flags);
1508 __task_rq_unlock(rq);
1513 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1515 static inline int ttwu_share_cache(int this_cpu, int that_cpu)
1517 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1519 #endif /* CONFIG_SMP */
1521 static void ttwu_queue(struct task_struct *p, int cpu)
1523 struct rq *rq = cpu_rq(cpu);
1525 #if defined(CONFIG_SMP)
1526 if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
1527 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1528 ttwu_queue_remote(p, cpu);
1533 raw_spin_lock(&rq->lock);
1534 ttwu_do_activate(rq, p, 0);
1535 raw_spin_unlock(&rq->lock);
1539 * try_to_wake_up - wake up a thread
1540 * @p: the thread to be awakened
1541 * @state: the mask of task states that can be woken
1542 * @wake_flags: wake modifier flags (WF_*)
1544 * Put it on the run-queue if it's not already there. The "current"
1545 * thread is always on the run-queue (except when the actual
1546 * re-schedule is in progress), and as such you're allowed to do
1547 * the simpler "current->state = TASK_RUNNING" to mark yourself
1548 * runnable without the overhead of this.
1550 * Returns %true if @p was woken up, %false if it was already running
1551 * or @state didn't match @p's state.
1554 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1556 unsigned long flags;
1557 int cpu, success = 0;
1560 raw_spin_lock_irqsave(&p->pi_lock, flags);
1561 if (!(p->state & state))
1564 success = 1; /* we're going to change ->state */
1567 if (p->on_rq && ttwu_remote(p, wake_flags))
1572 * If the owning (remote) cpu is still in the middle of schedule() with
1573 * this task as prev, wait until its done referencing the task.
1576 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1578 * In case the architecture enables interrupts in
1579 * context_switch(), we cannot busy wait, since that
1580 * would lead to deadlocks when an interrupt hits and
1581 * tries to wake up @prev. So bail and do a complete
1584 if (ttwu_activate_remote(p, wake_flags))
1591 * Pairs with the smp_wmb() in finish_lock_switch().
1595 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1596 p->state = TASK_WAKING;
1598 if (p->sched_class->task_waking)
1599 p->sched_class->task_waking(p);
1601 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1602 if (task_cpu(p) != cpu) {
1603 wake_flags |= WF_MIGRATED;
1604 set_task_cpu(p, cpu);
1606 #endif /* CONFIG_SMP */
1610 ttwu_stat(p, cpu, wake_flags);
1612 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1618 * try_to_wake_up_local - try to wake up a local task with rq lock held
1619 * @p: the thread to be awakened
1621 * Put @p on the run-queue if it's not already there. The caller must
1622 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1625 static void try_to_wake_up_local(struct task_struct *p)
1627 struct rq *rq = task_rq(p);
1629 BUG_ON(rq != this_rq());
1630 BUG_ON(p == current);
1631 lockdep_assert_held(&rq->lock);
1633 if (!raw_spin_trylock(&p->pi_lock)) {
1634 raw_spin_unlock(&rq->lock);
1635 raw_spin_lock(&p->pi_lock);
1636 raw_spin_lock(&rq->lock);
1639 if (!(p->state & TASK_NORMAL))
1643 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1645 ttwu_do_wakeup(rq, p, 0);
1646 ttwu_stat(p, smp_processor_id(), 0);
1648 raw_spin_unlock(&p->pi_lock);
1652 * wake_up_process - Wake up a specific process
1653 * @p: The process to be woken up.
1655 * Attempt to wake up the nominated process and move it to the set of runnable
1656 * processes. Returns 1 if the process was woken up, 0 if it was already
1659 * It may be assumed that this function implies a write memory barrier before
1660 * changing the task state if and only if any tasks are woken up.
1662 int wake_up_process(struct task_struct *p)
1664 return try_to_wake_up(p, TASK_ALL, 0);
1666 EXPORT_SYMBOL(wake_up_process);
1668 int wake_up_state(struct task_struct *p, unsigned int state)
1670 return try_to_wake_up(p, state, 0);
1674 * Perform scheduler related setup for a newly forked process p.
1675 * p is forked by current.
1677 * __sched_fork() is basic setup used by init_idle() too:
1679 static void __sched_fork(struct task_struct *p)
1684 p->se.exec_start = 0;
1685 p->se.sum_exec_runtime = 0;
1686 p->se.prev_sum_exec_runtime = 0;
1687 p->se.nr_migrations = 0;
1689 INIT_LIST_HEAD(&p->se.group_node);
1691 #ifdef CONFIG_SCHEDSTATS
1692 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1695 INIT_LIST_HEAD(&p->rt.run_list);
1697 #ifdef CONFIG_PREEMPT_NOTIFIERS
1698 INIT_HLIST_HEAD(&p->preempt_notifiers);
1703 * fork()/clone()-time setup:
1705 void sched_fork(struct task_struct *p)
1707 unsigned long flags;
1708 int cpu = get_cpu();
1712 * We mark the process as running here. This guarantees that
1713 * nobody will actually run it, and a signal or other external
1714 * event cannot wake it up and insert it on the runqueue either.
1716 p->state = TASK_RUNNING;
1719 * Make sure we do not leak PI boosting priority to the child.
1721 p->prio = current->normal_prio;
1724 * Revert to default priority/policy on fork if requested.
1726 if (unlikely(p->sched_reset_on_fork)) {
1727 if (task_has_rt_policy(p)) {
1728 p->policy = SCHED_NORMAL;
1729 p->static_prio = NICE_TO_PRIO(0);
1731 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1732 p->static_prio = NICE_TO_PRIO(0);
1734 p->prio = p->normal_prio = __normal_prio(p);
1738 * We don't need the reset flag anymore after the fork. It has
1739 * fulfilled its duty:
1741 p->sched_reset_on_fork = 0;
1744 if (!rt_prio(p->prio))
1745 p->sched_class = &fair_sched_class;
1747 if (p->sched_class->task_fork)
1748 p->sched_class->task_fork(p);
1751 * The child is not yet in the pid-hash so no cgroup attach races,
1752 * and the cgroup is pinned to this child due to cgroup_fork()
1753 * is ran before sched_fork().
1755 * Silence PROVE_RCU.
1757 raw_spin_lock_irqsave(&p->pi_lock, flags);
1758 set_task_cpu(p, cpu);
1759 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1761 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1762 if (likely(sched_info_on()))
1763 memset(&p->sched_info, 0, sizeof(p->sched_info));
1765 #if defined(CONFIG_SMP)
1768 #ifdef CONFIG_PREEMPT_COUNT
1769 /* Want to start with kernel preemption disabled. */
1770 task_thread_info(p)->preempt_count = 1;
1773 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1780 * wake_up_new_task - wake up a newly created task for the first time.
1782 * This function will do some initial scheduler statistics housekeeping
1783 * that must be done for every newly created context, then puts the task
1784 * on the runqueue and wakes it.
1786 void wake_up_new_task(struct task_struct *p)
1788 unsigned long flags;
1791 raw_spin_lock_irqsave(&p->pi_lock, flags);
1794 * Fork balancing, do it here and not earlier because:
1795 * - cpus_allowed can change in the fork path
1796 * - any previously selected cpu might disappear through hotplug
1798 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1801 rq = __task_rq_lock(p);
1802 activate_task(rq, p, 0);
1804 trace_sched_wakeup_new(p, true);
1805 check_preempt_curr(rq, p, WF_FORK);
1807 if (p->sched_class->task_woken)
1808 p->sched_class->task_woken(rq, p);
1810 task_rq_unlock(rq, p, &flags);
1813 #ifdef CONFIG_PREEMPT_NOTIFIERS
1816 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1817 * @notifier: notifier struct to register
1819 void preempt_notifier_register(struct preempt_notifier *notifier)
1821 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1823 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1826 * preempt_notifier_unregister - no longer interested in preemption notifications
1827 * @notifier: notifier struct to unregister
1829 * This is safe to call from within a preemption notifier.
1831 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1833 hlist_del(¬ifier->link);
1835 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1837 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1839 struct preempt_notifier *notifier;
1840 struct hlist_node *node;
1842 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1843 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1847 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1848 struct task_struct *next)
1850 struct preempt_notifier *notifier;
1851 struct hlist_node *node;
1853 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1854 notifier->ops->sched_out(notifier, next);
1857 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1859 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1864 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1865 struct task_struct *next)
1869 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1872 * prepare_task_switch - prepare to switch tasks
1873 * @rq: the runqueue preparing to switch
1874 * @prev: the current task that is being switched out
1875 * @next: the task we are going to switch to.
1877 * This is called with the rq lock held and interrupts off. It must
1878 * be paired with a subsequent finish_task_switch after the context
1881 * prepare_task_switch sets up locking and calls architecture specific
1885 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1886 struct task_struct *next)
1888 sched_info_switch(prev, next);
1889 perf_event_task_sched_out(prev, next);
1890 fire_sched_out_preempt_notifiers(prev, next);
1891 prepare_lock_switch(rq, next);
1892 prepare_arch_switch(next);
1893 trace_sched_switch(prev, next);
1897 * finish_task_switch - clean up after a task-switch
1898 * @rq: runqueue associated with task-switch
1899 * @prev: the thread we just switched away from.
1901 * finish_task_switch must be called after the context switch, paired
1902 * with a prepare_task_switch call before the context switch.
1903 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1904 * and do any other architecture-specific cleanup actions.
1906 * Note that we may have delayed dropping an mm in context_switch(). If
1907 * so, we finish that here outside of the runqueue lock. (Doing it
1908 * with the lock held can cause deadlocks; see schedule() for
1911 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1912 __releases(rq->lock)
1914 struct mm_struct *mm = rq->prev_mm;
1920 * A task struct has one reference for the use as "current".
1921 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1922 * schedule one last time. The schedule call will never return, and
1923 * the scheduled task must drop that reference.
1924 * The test for TASK_DEAD must occur while the runqueue locks are
1925 * still held, otherwise prev could be scheduled on another cpu, die
1926 * there before we look at prev->state, and then the reference would
1928 * Manfred Spraul <manfred@colorfullife.com>
1930 prev_state = prev->state;
1931 finish_arch_switch(prev);
1932 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1933 local_irq_disable();
1934 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1935 perf_event_task_sched_in(prev, current);
1936 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1938 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1939 finish_lock_switch(rq, prev);
1940 trace_sched_stat_sleeptime(current, rq->clock);
1942 fire_sched_in_preempt_notifiers(current);
1945 if (unlikely(prev_state == TASK_DEAD)) {
1947 * Remove function-return probe instances associated with this
1948 * task and put them back on the free list.
1950 kprobe_flush_task(prev);
1951 put_task_struct(prev);
1957 /* assumes rq->lock is held */
1958 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1960 if (prev->sched_class->pre_schedule)
1961 prev->sched_class->pre_schedule(rq, prev);
1964 /* rq->lock is NOT held, but preemption is disabled */
1965 static inline void post_schedule(struct rq *rq)
1967 if (rq->post_schedule) {
1968 unsigned long flags;
1970 raw_spin_lock_irqsave(&rq->lock, flags);
1971 if (rq->curr->sched_class->post_schedule)
1972 rq->curr->sched_class->post_schedule(rq);
1973 raw_spin_unlock_irqrestore(&rq->lock, flags);
1975 rq->post_schedule = 0;
1981 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1985 static inline void post_schedule(struct rq *rq)
1992 * schedule_tail - first thing a freshly forked thread must call.
1993 * @prev: the thread we just switched away from.
1995 asmlinkage void schedule_tail(struct task_struct *prev)
1996 __releases(rq->lock)
1998 struct rq *rq = this_rq();
2000 finish_task_switch(rq, prev);
2003 * FIXME: do we need to worry about rq being invalidated by the
2008 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2009 /* In this case, finish_task_switch does not reenable preemption */
2012 if (current->set_child_tid)
2013 put_user(task_pid_vnr(current), current->set_child_tid);
2017 * context_switch - switch to the new MM and the new
2018 * thread's register state.
2021 context_switch(struct rq *rq, struct task_struct *prev,
2022 struct task_struct *next)
2024 struct mm_struct *mm, *oldmm;
2026 prepare_task_switch(rq, prev, next);
2029 oldmm = prev->active_mm;
2031 * For paravirt, this is coupled with an exit in switch_to to
2032 * combine the page table reload and the switch backend into
2035 arch_start_context_switch(prev);
2038 next->active_mm = oldmm;
2039 atomic_inc(&oldmm->mm_count);
2040 enter_lazy_tlb(oldmm, next);
2042 switch_mm(oldmm, mm, next);
2045 prev->active_mm = NULL;
2046 rq->prev_mm = oldmm;
2049 * Since the runqueue lock will be released by the next
2050 * task (which is an invalid locking op but in the case
2051 * of the scheduler it's an obvious special-case), so we
2052 * do an early lockdep release here:
2054 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2055 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2058 /* Here we just switch the register state and the stack. */
2059 switch_to(prev, next, prev);
2063 * this_rq must be evaluated again because prev may have moved
2064 * CPUs since it called schedule(), thus the 'rq' on its stack
2065 * frame will be invalid.
2067 finish_task_switch(this_rq(), prev);
2071 * nr_running, nr_uninterruptible and nr_context_switches:
2073 * externally visible scheduler statistics: current number of runnable
2074 * threads, current number of uninterruptible-sleeping threads, total
2075 * number of context switches performed since bootup.
2077 unsigned long nr_running(void)
2079 unsigned long i, sum = 0;
2081 for_each_online_cpu(i)
2082 sum += cpu_rq(i)->nr_running;
2087 unsigned long nr_uninterruptible(void)
2089 unsigned long i, sum = 0;
2091 for_each_possible_cpu(i)
2092 sum += cpu_rq(i)->nr_uninterruptible;
2095 * Since we read the counters lockless, it might be slightly
2096 * inaccurate. Do not allow it to go below zero though:
2098 if (unlikely((long)sum < 0))
2104 unsigned long long nr_context_switches(void)
2107 unsigned long long sum = 0;
2109 for_each_possible_cpu(i)
2110 sum += cpu_rq(i)->nr_switches;
2115 unsigned long nr_iowait(void)
2117 unsigned long i, sum = 0;
2119 for_each_possible_cpu(i)
2120 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2125 unsigned long nr_iowait_cpu(int cpu)
2127 struct rq *this = cpu_rq(cpu);
2128 return atomic_read(&this->nr_iowait);
2131 unsigned long this_cpu_load(void)
2133 struct rq *this = this_rq();
2134 return this->cpu_load[0];
2138 /* Variables and functions for calc_load */
2139 static atomic_long_t calc_load_tasks;
2140 static unsigned long calc_load_update;
2141 unsigned long avenrun[3];
2142 EXPORT_SYMBOL(avenrun);
2144 static long calc_load_fold_active(struct rq *this_rq)
2146 long nr_active, delta = 0;
2148 nr_active = this_rq->nr_running;
2149 nr_active += (long) this_rq->nr_uninterruptible;
2151 if (nr_active != this_rq->calc_load_active) {
2152 delta = nr_active - this_rq->calc_load_active;
2153 this_rq->calc_load_active = nr_active;
2159 static unsigned long
2160 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2163 load += active * (FIXED_1 - exp);
2164 load += 1UL << (FSHIFT - 1);
2165 return load >> FSHIFT;
2170 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2172 * When making the ILB scale, we should try to pull this in as well.
2174 static atomic_long_t calc_load_tasks_idle;
2176 void calc_load_account_idle(struct rq *this_rq)
2180 delta = calc_load_fold_active(this_rq);
2182 atomic_long_add(delta, &calc_load_tasks_idle);
2185 static long calc_load_fold_idle(void)
2190 * Its got a race, we don't care...
2192 if (atomic_long_read(&calc_load_tasks_idle))
2193 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2199 * fixed_power_int - compute: x^n, in O(log n) time
2201 * @x: base of the power
2202 * @frac_bits: fractional bits of @x
2203 * @n: power to raise @x to.
2205 * By exploiting the relation between the definition of the natural power
2206 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2207 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2208 * (where: n_i \elem {0, 1}, the binary vector representing n),
2209 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2210 * of course trivially computable in O(log_2 n), the length of our binary
2213 static unsigned long
2214 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2216 unsigned long result = 1UL << frac_bits;
2221 result += 1UL << (frac_bits - 1);
2222 result >>= frac_bits;
2228 x += 1UL << (frac_bits - 1);
2236 * a1 = a0 * e + a * (1 - e)
2238 * a2 = a1 * e + a * (1 - e)
2239 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2240 * = a0 * e^2 + a * (1 - e) * (1 + e)
2242 * a3 = a2 * e + a * (1 - e)
2243 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2244 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2248 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2249 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2250 * = a0 * e^n + a * (1 - e^n)
2252 * [1] application of the geometric series:
2255 * S_n := \Sum x^i = -------------
2258 static unsigned long
2259 calc_load_n(unsigned long load, unsigned long exp,
2260 unsigned long active, unsigned int n)
2263 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2267 * NO_HZ can leave us missing all per-cpu ticks calling
2268 * calc_load_account_active(), but since an idle CPU folds its delta into
2269 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2270 * in the pending idle delta if our idle period crossed a load cycle boundary.
2272 * Once we've updated the global active value, we need to apply the exponential
2273 * weights adjusted to the number of cycles missed.
2275 static void calc_global_nohz(unsigned long ticks)
2277 long delta, active, n;
2279 if (time_before(jiffies, calc_load_update))
2283 * If we crossed a calc_load_update boundary, make sure to fold
2284 * any pending idle changes, the respective CPUs might have
2285 * missed the tick driven calc_load_account_active() update
2288 delta = calc_load_fold_idle();
2290 atomic_long_add(delta, &calc_load_tasks);
2293 * If we were idle for multiple load cycles, apply them.
2295 if (ticks >= LOAD_FREQ) {
2296 n = ticks / LOAD_FREQ;
2298 active = atomic_long_read(&calc_load_tasks);
2299 active = active > 0 ? active * FIXED_1 : 0;
2301 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2302 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2303 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2305 calc_load_update += n * LOAD_FREQ;
2309 * Its possible the remainder of the above division also crosses
2310 * a LOAD_FREQ period, the regular check in calc_global_load()
2311 * which comes after this will take care of that.
2313 * Consider us being 11 ticks before a cycle completion, and us
2314 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
2315 * age us 4 cycles, and the test in calc_global_load() will
2316 * pick up the final one.
2320 void calc_load_account_idle(struct rq *this_rq)
2324 static inline long calc_load_fold_idle(void)
2329 static void calc_global_nohz(unsigned long ticks)
2335 * get_avenrun - get the load average array
2336 * @loads: pointer to dest load array
2337 * @offset: offset to add
2338 * @shift: shift count to shift the result left
2340 * These values are estimates at best, so no need for locking.
2342 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2344 loads[0] = (avenrun[0] + offset) << shift;
2345 loads[1] = (avenrun[1] + offset) << shift;
2346 loads[2] = (avenrun[2] + offset) << shift;
2350 * calc_load - update the avenrun load estimates 10 ticks after the
2351 * CPUs have updated calc_load_tasks.
2353 void calc_global_load(unsigned long ticks)
2357 calc_global_nohz(ticks);
2359 if (time_before(jiffies, calc_load_update + 10))
2362 active = atomic_long_read(&calc_load_tasks);
2363 active = active > 0 ? active * FIXED_1 : 0;
2365 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2366 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2367 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2369 calc_load_update += LOAD_FREQ;
2373 * Called from update_cpu_load() to periodically update this CPU's
2376 static void calc_load_account_active(struct rq *this_rq)
2380 if (time_before(jiffies, this_rq->calc_load_update))
2383 delta = calc_load_fold_active(this_rq);
2384 delta += calc_load_fold_idle();
2386 atomic_long_add(delta, &calc_load_tasks);
2388 this_rq->calc_load_update += LOAD_FREQ;
2392 * The exact cpuload at various idx values, calculated at every tick would be
2393 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2395 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2396 * on nth tick when cpu may be busy, then we have:
2397 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2398 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2400 * decay_load_missed() below does efficient calculation of
2401 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2402 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2404 * The calculation is approximated on a 128 point scale.
2405 * degrade_zero_ticks is the number of ticks after which load at any
2406 * particular idx is approximated to be zero.
2407 * degrade_factor is a precomputed table, a row for each load idx.
2408 * Each column corresponds to degradation factor for a power of two ticks,
2409 * based on 128 point scale.
2411 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2412 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2414 * With this power of 2 load factors, we can degrade the load n times
2415 * by looking at 1 bits in n and doing as many mult/shift instead of
2416 * n mult/shifts needed by the exact degradation.
2418 #define DEGRADE_SHIFT 7
2419 static const unsigned char
2420 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2421 static const unsigned char
2422 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2423 {0, 0, 0, 0, 0, 0, 0, 0},
2424 {64, 32, 8, 0, 0, 0, 0, 0},
2425 {96, 72, 40, 12, 1, 0, 0},
2426 {112, 98, 75, 43, 15, 1, 0},
2427 {120, 112, 98, 76, 45, 16, 2} };
2430 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2431 * would be when CPU is idle and so we just decay the old load without
2432 * adding any new load.
2434 static unsigned long
2435 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2439 if (!missed_updates)
2442 if (missed_updates >= degrade_zero_ticks[idx])
2446 return load >> missed_updates;
2448 while (missed_updates) {
2449 if (missed_updates % 2)
2450 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2452 missed_updates >>= 1;
2459 * Update rq->cpu_load[] statistics. This function is usually called every
2460 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2461 * every tick. We fix it up based on jiffies.
2463 void update_cpu_load(struct rq *this_rq)
2465 unsigned long this_load = this_rq->load.weight;
2466 unsigned long curr_jiffies = jiffies;
2467 unsigned long pending_updates;
2470 this_rq->nr_load_updates++;
2472 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2473 if (curr_jiffies == this_rq->last_load_update_tick)
2476 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2477 this_rq->last_load_update_tick = curr_jiffies;
2479 /* Update our load: */
2480 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2481 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2482 unsigned long old_load, new_load;
2484 /* scale is effectively 1 << i now, and >> i divides by scale */
2486 old_load = this_rq->cpu_load[i];
2487 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2488 new_load = this_load;
2490 * Round up the averaging division if load is increasing. This
2491 * prevents us from getting stuck on 9 if the load is 10, for
2494 if (new_load > old_load)
2495 new_load += scale - 1;
2497 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2500 sched_avg_update(this_rq);
2503 static void update_cpu_load_active(struct rq *this_rq)
2505 update_cpu_load(this_rq);
2507 calc_load_account_active(this_rq);
2513 * sched_exec - execve() is a valuable balancing opportunity, because at
2514 * this point the task has the smallest effective memory and cache footprint.
2516 void sched_exec(void)
2518 struct task_struct *p = current;
2519 unsigned long flags;
2522 raw_spin_lock_irqsave(&p->pi_lock, flags);
2523 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2524 if (dest_cpu == smp_processor_id())
2527 if (likely(cpu_active(dest_cpu))) {
2528 struct migration_arg arg = { p, dest_cpu };
2530 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2531 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2535 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2540 DEFINE_PER_CPU(struct kernel_stat, kstat);
2541 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2543 EXPORT_PER_CPU_SYMBOL(kstat);
2544 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2547 * Return any ns on the sched_clock that have not yet been accounted in
2548 * @p in case that task is currently running.
2550 * Called with task_rq_lock() held on @rq.
2552 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2556 if (task_current(rq, p)) {
2557 update_rq_clock(rq);
2558 ns = rq->clock_task - p->se.exec_start;
2566 unsigned long long task_delta_exec(struct task_struct *p)
2568 unsigned long flags;
2572 rq = task_rq_lock(p, &flags);
2573 ns = do_task_delta_exec(p, rq);
2574 task_rq_unlock(rq, p, &flags);
2580 * Return accounted runtime for the task.
2581 * In case the task is currently running, return the runtime plus current's
2582 * pending runtime that have not been accounted yet.
2584 unsigned long long task_sched_runtime(struct task_struct *p)
2586 unsigned long flags;
2590 rq = task_rq_lock(p, &flags);
2591 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2592 task_rq_unlock(rq, p, &flags);
2597 #ifdef CONFIG_CGROUP_CPUACCT
2598 struct cgroup_subsys cpuacct_subsys;
2599 struct cpuacct root_cpuacct;
2602 static inline void task_group_account_field(struct task_struct *p, int index,
2605 #ifdef CONFIG_CGROUP_CPUACCT
2606 struct kernel_cpustat *kcpustat;
2610 * Since all updates are sure to touch the root cgroup, we
2611 * get ourselves ahead and touch it first. If the root cgroup
2612 * is the only cgroup, then nothing else should be necessary.
2615 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2617 #ifdef CONFIG_CGROUP_CPUACCT
2618 if (unlikely(!cpuacct_subsys.active))
2623 while (ca && (ca != &root_cpuacct)) {
2624 kcpustat = this_cpu_ptr(ca->cpustat);
2625 kcpustat->cpustat[index] += tmp;
2634 * Account user cpu time to a process.
2635 * @p: the process that the cpu time gets accounted to
2636 * @cputime: the cpu time spent in user space since the last update
2637 * @cputime_scaled: cputime scaled by cpu frequency
2639 void account_user_time(struct task_struct *p, cputime_t cputime,
2640 cputime_t cputime_scaled)
2644 /* Add user time to process. */
2645 p->utime += cputime;
2646 p->utimescaled += cputime_scaled;
2647 account_group_user_time(p, cputime);
2649 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2651 /* Add user time to cpustat. */
2652 task_group_account_field(p, index, (__force u64) cputime);
2654 /* Account for user time used */
2655 acct_update_integrals(p);
2659 * Account guest cpu time to a process.
2660 * @p: the process that the cpu time gets accounted to
2661 * @cputime: the cpu time spent in virtual machine since the last update
2662 * @cputime_scaled: cputime scaled by cpu frequency
2664 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2665 cputime_t cputime_scaled)
2667 u64 *cpustat = kcpustat_this_cpu->cpustat;
2669 /* Add guest time to process. */
2670 p->utime += cputime;
2671 p->utimescaled += cputime_scaled;
2672 account_group_user_time(p, cputime);
2673 p->gtime += cputime;
2675 /* Add guest time to cpustat. */
2676 if (TASK_NICE(p) > 0) {
2677 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2678 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2680 cpustat[CPUTIME_USER] += (__force u64) cputime;
2681 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2686 * Account system cpu time to a process and desired cpustat field
2687 * @p: the process that the cpu time gets accounted to
2688 * @cputime: the cpu time spent in kernel space since the last update
2689 * @cputime_scaled: cputime scaled by cpu frequency
2690 * @target_cputime64: pointer to cpustat field that has to be updated
2693 void __account_system_time(struct task_struct *p, cputime_t cputime,
2694 cputime_t cputime_scaled, int index)
2696 /* Add system time to process. */
2697 p->stime += cputime;
2698 p->stimescaled += cputime_scaled;
2699 account_group_system_time(p, cputime);
2701 /* Add system time to cpustat. */
2702 task_group_account_field(p, index, (__force u64) cputime);
2704 /* Account for system time used */
2705 acct_update_integrals(p);
2709 * Account system cpu time to a process.
2710 * @p: the process that the cpu time gets accounted to
2711 * @hardirq_offset: the offset to subtract from hardirq_count()
2712 * @cputime: the cpu time spent in kernel space since the last update
2713 * @cputime_scaled: cputime scaled by cpu frequency
2715 void account_system_time(struct task_struct *p, int hardirq_offset,
2716 cputime_t cputime, cputime_t cputime_scaled)
2720 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2721 account_guest_time(p, cputime, cputime_scaled);
2725 if (hardirq_count() - hardirq_offset)
2726 index = CPUTIME_IRQ;
2727 else if (in_serving_softirq())
2728 index = CPUTIME_SOFTIRQ;
2730 index = CPUTIME_SYSTEM;
2732 __account_system_time(p, cputime, cputime_scaled, index);
2736 * Account for involuntary wait time.
2737 * @cputime: the cpu time spent in involuntary wait
2739 void account_steal_time(cputime_t cputime)
2741 u64 *cpustat = kcpustat_this_cpu->cpustat;
2743 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2747 * Account for idle time.
2748 * @cputime: the cpu time spent in idle wait
2750 void account_idle_time(cputime_t cputime)
2752 u64 *cpustat = kcpustat_this_cpu->cpustat;
2753 struct rq *rq = this_rq();
2755 if (atomic_read(&rq->nr_iowait) > 0)
2756 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2758 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2761 static __always_inline bool steal_account_process_tick(void)
2763 #ifdef CONFIG_PARAVIRT
2764 if (static_branch(¶virt_steal_enabled)) {
2767 steal = paravirt_steal_clock(smp_processor_id());
2768 steal -= this_rq()->prev_steal_time;
2770 st = steal_ticks(steal);
2771 this_rq()->prev_steal_time += st * TICK_NSEC;
2773 account_steal_time(st);
2780 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2782 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2784 * Account a tick to a process and cpustat
2785 * @p: the process that the cpu time gets accounted to
2786 * @user_tick: is the tick from userspace
2787 * @rq: the pointer to rq
2789 * Tick demultiplexing follows the order
2790 * - pending hardirq update
2791 * - pending softirq update
2795 * - check for guest_time
2796 * - else account as system_time
2798 * Check for hardirq is done both for system and user time as there is
2799 * no timer going off while we are on hardirq and hence we may never get an
2800 * opportunity to update it solely in system time.
2801 * p->stime and friends are only updated on system time and not on irq
2802 * softirq as those do not count in task exec_runtime any more.
2804 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2807 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2808 u64 *cpustat = kcpustat_this_cpu->cpustat;
2810 if (steal_account_process_tick())
2813 if (irqtime_account_hi_update()) {
2814 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
2815 } else if (irqtime_account_si_update()) {
2816 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
2817 } else if (this_cpu_ksoftirqd() == p) {
2819 * ksoftirqd time do not get accounted in cpu_softirq_time.
2820 * So, we have to handle it separately here.
2821 * Also, p->stime needs to be updated for ksoftirqd.
2823 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2825 } else if (user_tick) {
2826 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2827 } else if (p == rq->idle) {
2828 account_idle_time(cputime_one_jiffy);
2829 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2830 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2832 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2837 static void irqtime_account_idle_ticks(int ticks)
2840 struct rq *rq = this_rq();
2842 for (i = 0; i < ticks; i++)
2843 irqtime_account_process_tick(current, 0, rq);
2845 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2846 static void irqtime_account_idle_ticks(int ticks) {}
2847 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2849 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2852 * Account a single tick of cpu time.
2853 * @p: the process that the cpu time gets accounted to
2854 * @user_tick: indicates if the tick is a user or a system tick
2856 void account_process_tick(struct task_struct *p, int user_tick)
2858 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2859 struct rq *rq = this_rq();
2861 if (sched_clock_irqtime) {
2862 irqtime_account_process_tick(p, user_tick, rq);
2866 if (steal_account_process_tick())
2870 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2871 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2872 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2875 account_idle_time(cputime_one_jiffy);
2879 * Account multiple ticks of steal time.
2880 * @p: the process from which the cpu time has been stolen
2881 * @ticks: number of stolen ticks
2883 void account_steal_ticks(unsigned long ticks)
2885 account_steal_time(jiffies_to_cputime(ticks));
2889 * Account multiple ticks of idle time.
2890 * @ticks: number of stolen ticks
2892 void account_idle_ticks(unsigned long ticks)
2895 if (sched_clock_irqtime) {
2896 irqtime_account_idle_ticks(ticks);
2900 account_idle_time(jiffies_to_cputime(ticks));
2906 * Use precise platform statistics if available:
2908 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2909 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2915 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2917 struct task_cputime cputime;
2919 thread_group_cputime(p, &cputime);
2921 *ut = cputime.utime;
2922 *st = cputime.stime;
2926 #ifndef nsecs_to_cputime
2927 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2930 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2932 cputime_t rtime, utime = p->utime, total = utime + p->stime;
2935 * Use CFS's precise accounting:
2937 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
2940 u64 temp = (__force u64) rtime;
2942 temp *= (__force u64) utime;
2943 do_div(temp, (__force u32) total);
2944 utime = (__force cputime_t) temp;
2949 * Compare with previous values, to keep monotonicity:
2951 p->prev_utime = max(p->prev_utime, utime);
2952 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
2954 *ut = p->prev_utime;
2955 *st = p->prev_stime;
2959 * Must be called with siglock held.
2961 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2963 struct signal_struct *sig = p->signal;
2964 struct task_cputime cputime;
2965 cputime_t rtime, utime, total;
2967 thread_group_cputime(p, &cputime);
2969 total = cputime.utime + cputime.stime;
2970 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
2973 u64 temp = (__force u64) rtime;
2975 temp *= (__force u64) cputime.utime;
2976 do_div(temp, (__force u32) total);
2977 utime = (__force cputime_t) temp;
2981 sig->prev_utime = max(sig->prev_utime, utime);
2982 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
2984 *ut = sig->prev_utime;
2985 *st = sig->prev_stime;
2990 * This function gets called by the timer code, with HZ frequency.
2991 * We call it with interrupts disabled.
2993 void scheduler_tick(void)
2995 int cpu = smp_processor_id();
2996 struct rq *rq = cpu_rq(cpu);
2997 struct task_struct *curr = rq->curr;
3001 raw_spin_lock(&rq->lock);
3002 update_rq_clock(rq);
3003 update_cpu_load_active(rq);
3004 curr->sched_class->task_tick(rq, curr, 0);
3005 raw_spin_unlock(&rq->lock);
3007 perf_event_task_tick();
3010 rq->idle_balance = idle_cpu(cpu);
3011 trigger_load_balance(rq, cpu);
3015 notrace unsigned long get_parent_ip(unsigned long addr)
3017 if (in_lock_functions(addr)) {
3018 addr = CALLER_ADDR2;
3019 if (in_lock_functions(addr))
3020 addr = CALLER_ADDR3;
3025 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3026 defined(CONFIG_PREEMPT_TRACER))
3028 void __kprobes add_preempt_count(int val)
3030 #ifdef CONFIG_DEBUG_PREEMPT
3034 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3037 preempt_count() += val;
3038 #ifdef CONFIG_DEBUG_PREEMPT
3040 * Spinlock count overflowing soon?
3042 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3045 if (preempt_count() == val)
3046 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3048 EXPORT_SYMBOL(add_preempt_count);
3050 void __kprobes sub_preempt_count(int val)
3052 #ifdef CONFIG_DEBUG_PREEMPT
3056 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3059 * Is the spinlock portion underflowing?
3061 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3062 !(preempt_count() & PREEMPT_MASK)))
3066 if (preempt_count() == val)
3067 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3068 preempt_count() -= val;
3070 EXPORT_SYMBOL(sub_preempt_count);
3075 * Print scheduling while atomic bug:
3077 static noinline void __schedule_bug(struct task_struct *prev)
3079 struct pt_regs *regs = get_irq_regs();
3081 if (oops_in_progress)
3084 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3085 prev->comm, prev->pid, preempt_count());
3087 debug_show_held_locks(prev);
3089 if (irqs_disabled())
3090 print_irqtrace_events(prev);
3099 * Various schedule()-time debugging checks and statistics:
3101 static inline void schedule_debug(struct task_struct *prev)
3104 * Test if we are atomic. Since do_exit() needs to call into
3105 * schedule() atomically, we ignore that path for now.
3106 * Otherwise, whine if we are scheduling when we should not be.
3108 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3109 __schedule_bug(prev);
3112 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3114 schedstat_inc(this_rq(), sched_count);
3117 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3119 if (prev->on_rq || rq->skip_clock_update < 0)
3120 update_rq_clock(rq);
3121 prev->sched_class->put_prev_task(rq, prev);
3125 * Pick up the highest-prio task:
3127 static inline struct task_struct *
3128 pick_next_task(struct rq *rq)
3130 const struct sched_class *class;
3131 struct task_struct *p;
3134 * Optimization: we know that if all tasks are in
3135 * the fair class we can call that function directly:
3137 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3138 p = fair_sched_class.pick_next_task(rq);
3143 for_each_class(class) {
3144 p = class->pick_next_task(rq);
3149 BUG(); /* the idle class will always have a runnable task */
3153 * __schedule() is the main scheduler function.
3155 static void __sched __schedule(void)
3157 struct task_struct *prev, *next;
3158 unsigned long *switch_count;
3164 cpu = smp_processor_id();
3166 rcu_note_context_switch(cpu);
3169 schedule_debug(prev);
3171 if (sched_feat(HRTICK))
3174 raw_spin_lock_irq(&rq->lock);
3176 switch_count = &prev->nivcsw;
3177 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3178 if (unlikely(signal_pending_state(prev->state, prev))) {
3179 prev->state = TASK_RUNNING;
3181 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3185 * If a worker went to sleep, notify and ask workqueue
3186 * whether it wants to wake up a task to maintain
3189 if (prev->flags & PF_WQ_WORKER) {
3190 struct task_struct *to_wakeup;
3192 to_wakeup = wq_worker_sleeping(prev, cpu);
3194 try_to_wake_up_local(to_wakeup);
3197 switch_count = &prev->nvcsw;
3200 pre_schedule(rq, prev);
3202 if (unlikely(!rq->nr_running))
3203 idle_balance(cpu, rq);
3205 put_prev_task(rq, prev);
3206 next = pick_next_task(rq);
3207 clear_tsk_need_resched(prev);
3208 rq->skip_clock_update = 0;
3210 if (likely(prev != next)) {
3215 context_switch(rq, prev, next); /* unlocks the rq */
3217 * The context switch have flipped the stack from under us
3218 * and restored the local variables which were saved when
3219 * this task called schedule() in the past. prev == current
3220 * is still correct, but it can be moved to another cpu/rq.
3222 cpu = smp_processor_id();
3225 raw_spin_unlock_irq(&rq->lock);
3229 preempt_enable_no_resched();
3234 static inline void sched_submit_work(struct task_struct *tsk)
3239 * If we are going to sleep and we have plugged IO queued,
3240 * make sure to submit it to avoid deadlocks.
3242 if (blk_needs_flush_plug(tsk))
3243 blk_schedule_flush_plug(tsk);
3246 asmlinkage void __sched schedule(void)
3248 struct task_struct *tsk = current;
3250 sched_submit_work(tsk);
3253 EXPORT_SYMBOL(schedule);
3255 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3257 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3259 if (lock->owner != owner)
3263 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3264 * lock->owner still matches owner, if that fails, owner might
3265 * point to free()d memory, if it still matches, the rcu_read_lock()
3266 * ensures the memory stays valid.
3270 return owner->on_cpu;
3274 * Look out! "owner" is an entirely speculative pointer
3275 * access and not reliable.
3277 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3279 if (!sched_feat(OWNER_SPIN))
3283 while (owner_running(lock, owner)) {
3287 arch_mutex_cpu_relax();
3292 * We break out the loop above on need_resched() and when the
3293 * owner changed, which is a sign for heavy contention. Return
3294 * success only when lock->owner is NULL.
3296 return lock->owner == NULL;
3300 #ifdef CONFIG_PREEMPT
3302 * this is the entry point to schedule() from in-kernel preemption
3303 * off of preempt_enable. Kernel preemptions off return from interrupt
3304 * occur there and call schedule directly.
3306 asmlinkage void __sched notrace preempt_schedule(void)
3308 struct thread_info *ti = current_thread_info();
3311 * If there is a non-zero preempt_count or interrupts are disabled,
3312 * we do not want to preempt the current task. Just return..
3314 if (likely(ti->preempt_count || irqs_disabled()))
3318 add_preempt_count_notrace(PREEMPT_ACTIVE);
3320 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3323 * Check again in case we missed a preemption opportunity
3324 * between schedule and now.
3327 } while (need_resched());
3329 EXPORT_SYMBOL(preempt_schedule);
3332 * this is the entry point to schedule() from kernel preemption
3333 * off of irq context.
3334 * Note, that this is called and return with irqs disabled. This will
3335 * protect us against recursive calling from irq.
3337 asmlinkage void __sched preempt_schedule_irq(void)
3339 struct thread_info *ti = current_thread_info();
3341 /* Catch callers which need to be fixed */
3342 BUG_ON(ti->preempt_count || !irqs_disabled());
3345 add_preempt_count(PREEMPT_ACTIVE);
3348 local_irq_disable();
3349 sub_preempt_count(PREEMPT_ACTIVE);
3352 * Check again in case we missed a preemption opportunity
3353 * between schedule and now.
3356 } while (need_resched());
3359 #endif /* CONFIG_PREEMPT */
3361 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3364 return try_to_wake_up(curr->private, mode, wake_flags);
3366 EXPORT_SYMBOL(default_wake_function);
3369 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3370 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3371 * number) then we wake all the non-exclusive tasks and one exclusive task.
3373 * There are circumstances in which we can try to wake a task which has already
3374 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3375 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3377 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3378 int nr_exclusive, int wake_flags, void *key)
3380 wait_queue_t *curr, *next;
3382 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3383 unsigned flags = curr->flags;
3385 if (curr->func(curr, mode, wake_flags, key) &&
3386 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3392 * __wake_up - wake up threads blocked on a waitqueue.
3394 * @mode: which threads
3395 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3396 * @key: is directly passed to the wakeup function
3398 * It may be assumed that this function implies a write memory barrier before
3399 * changing the task state if and only if any tasks are woken up.
3401 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3402 int nr_exclusive, void *key)
3404 unsigned long flags;
3406 spin_lock_irqsave(&q->lock, flags);
3407 __wake_up_common(q, mode, nr_exclusive, 0, key);
3408 spin_unlock_irqrestore(&q->lock, flags);
3410 EXPORT_SYMBOL(__wake_up);
3413 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3415 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3417 __wake_up_common(q, mode, 1, 0, NULL);
3419 EXPORT_SYMBOL_GPL(__wake_up_locked);
3421 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3423 __wake_up_common(q, mode, 1, 0, key);
3425 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3428 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3430 * @mode: which threads
3431 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3432 * @key: opaque value to be passed to wakeup targets
3434 * The sync wakeup differs that the waker knows that it will schedule
3435 * away soon, so while the target thread will be woken up, it will not
3436 * be migrated to another CPU - ie. the two threads are 'synchronized'
3437 * with each other. This can prevent needless bouncing between CPUs.
3439 * On UP it can prevent extra preemption.
3441 * It may be assumed that this function implies a write memory barrier before
3442 * changing the task state if and only if any tasks are woken up.
3444 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3445 int nr_exclusive, void *key)
3447 unsigned long flags;
3448 int wake_flags = WF_SYNC;
3453 if (unlikely(!nr_exclusive))
3456 spin_lock_irqsave(&q->lock, flags);
3457 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3458 spin_unlock_irqrestore(&q->lock, flags);
3460 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3463 * __wake_up_sync - see __wake_up_sync_key()
3465 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3467 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3469 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3472 * complete: - signals a single thread waiting on this completion
3473 * @x: holds the state of this particular completion
3475 * This will wake up a single thread waiting on this completion. Threads will be
3476 * awakened in the same order in which they were queued.
3478 * See also complete_all(), wait_for_completion() and related routines.
3480 * It may be assumed that this function implies a write memory barrier before
3481 * changing the task state if and only if any tasks are woken up.
3483 void complete(struct completion *x)
3485 unsigned long flags;
3487 spin_lock_irqsave(&x->wait.lock, flags);
3489 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3490 spin_unlock_irqrestore(&x->wait.lock, flags);
3492 EXPORT_SYMBOL(complete);
3495 * complete_all: - signals all threads waiting on this completion
3496 * @x: holds the state of this particular completion
3498 * This will wake up all threads waiting on this particular completion event.
3500 * It may be assumed that this function implies a write memory barrier before
3501 * changing the task state if and only if any tasks are woken up.
3503 void complete_all(struct completion *x)
3505 unsigned long flags;
3507 spin_lock_irqsave(&x->wait.lock, flags);
3508 x->done += UINT_MAX/2;
3509 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3510 spin_unlock_irqrestore(&x->wait.lock, flags);
3512 EXPORT_SYMBOL(complete_all);
3514 static inline long __sched
3515 do_wait_for_common(struct completion *x, long timeout, int state)
3518 DECLARE_WAITQUEUE(wait, current);
3520 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3522 if (signal_pending_state(state, current)) {
3523 timeout = -ERESTARTSYS;
3526 __set_current_state(state);
3527 spin_unlock_irq(&x->wait.lock);
3528 timeout = schedule_timeout(timeout);
3529 spin_lock_irq(&x->wait.lock);
3530 } while (!x->done && timeout);
3531 __remove_wait_queue(&x->wait, &wait);
3536 return timeout ?: 1;
3540 wait_for_common(struct completion *x, long timeout, int state)
3544 spin_lock_irq(&x->wait.lock);
3545 timeout = do_wait_for_common(x, timeout, state);
3546 spin_unlock_irq(&x->wait.lock);
3551 * wait_for_completion: - waits for completion of a task
3552 * @x: holds the state of this particular completion
3554 * This waits to be signaled for completion of a specific task. It is NOT
3555 * interruptible and there is no timeout.
3557 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3558 * and interrupt capability. Also see complete().
3560 void __sched wait_for_completion(struct completion *x)
3562 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3564 EXPORT_SYMBOL(wait_for_completion);
3567 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3568 * @x: holds the state of this particular completion
3569 * @timeout: timeout value in jiffies
3571 * This waits for either a completion of a specific task to be signaled or for a
3572 * specified timeout to expire. The timeout is in jiffies. It is not
3575 * The return value is 0 if timed out, and positive (at least 1, or number of
3576 * jiffies left till timeout) if completed.
3578 unsigned long __sched
3579 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3581 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3583 EXPORT_SYMBOL(wait_for_completion_timeout);
3586 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3587 * @x: holds the state of this particular completion
3589 * This waits for completion of a specific task to be signaled. It is
3592 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3594 int __sched wait_for_completion_interruptible(struct completion *x)
3596 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3597 if (t == -ERESTARTSYS)
3601 EXPORT_SYMBOL(wait_for_completion_interruptible);
3604 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3605 * @x: holds the state of this particular completion
3606 * @timeout: timeout value in jiffies
3608 * This waits for either a completion of a specific task to be signaled or for a
3609 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3611 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3612 * positive (at least 1, or number of jiffies left till timeout) if completed.
3615 wait_for_completion_interruptible_timeout(struct completion *x,
3616 unsigned long timeout)
3618 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3620 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3623 * wait_for_completion_killable: - waits for completion of a task (killable)
3624 * @x: holds the state of this particular completion
3626 * This waits to be signaled for completion of a specific task. It can be
3627 * interrupted by a kill signal.
3629 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3631 int __sched wait_for_completion_killable(struct completion *x)
3633 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3634 if (t == -ERESTARTSYS)
3638 EXPORT_SYMBOL(wait_for_completion_killable);
3641 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3642 * @x: holds the state of this particular completion
3643 * @timeout: timeout value in jiffies
3645 * This waits for either a completion of a specific task to be
3646 * signaled or for a specified timeout to expire. It can be
3647 * interrupted by a kill signal. The timeout is in jiffies.
3649 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3650 * positive (at least 1, or number of jiffies left till timeout) if completed.
3653 wait_for_completion_killable_timeout(struct completion *x,
3654 unsigned long timeout)
3656 return wait_for_common(x, timeout, TASK_KILLABLE);
3658 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3661 * try_wait_for_completion - try to decrement a completion without blocking
3662 * @x: completion structure
3664 * Returns: 0 if a decrement cannot be done without blocking
3665 * 1 if a decrement succeeded.
3667 * If a completion is being used as a counting completion,
3668 * attempt to decrement the counter without blocking. This
3669 * enables us to avoid waiting if the resource the completion
3670 * is protecting is not available.
3672 bool try_wait_for_completion(struct completion *x)
3674 unsigned long flags;
3677 spin_lock_irqsave(&x->wait.lock, flags);
3682 spin_unlock_irqrestore(&x->wait.lock, flags);
3685 EXPORT_SYMBOL(try_wait_for_completion);
3688 * completion_done - Test to see if a completion has any waiters
3689 * @x: completion structure
3691 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3692 * 1 if there are no waiters.
3695 bool completion_done(struct completion *x)
3697 unsigned long flags;
3700 spin_lock_irqsave(&x->wait.lock, flags);
3703 spin_unlock_irqrestore(&x->wait.lock, flags);
3706 EXPORT_SYMBOL(completion_done);
3709 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3711 unsigned long flags;
3714 init_waitqueue_entry(&wait, current);
3716 __set_current_state(state);
3718 spin_lock_irqsave(&q->lock, flags);
3719 __add_wait_queue(q, &wait);
3720 spin_unlock(&q->lock);
3721 timeout = schedule_timeout(timeout);
3722 spin_lock_irq(&q->lock);
3723 __remove_wait_queue(q, &wait);
3724 spin_unlock_irqrestore(&q->lock, flags);
3729 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3731 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3733 EXPORT_SYMBOL(interruptible_sleep_on);
3736 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3738 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3740 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3742 void __sched sleep_on(wait_queue_head_t *q)
3744 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3746 EXPORT_SYMBOL(sleep_on);
3748 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3750 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3752 EXPORT_SYMBOL(sleep_on_timeout);
3754 #ifdef CONFIG_RT_MUTEXES
3757 * rt_mutex_setprio - set the current priority of a task
3759 * @prio: prio value (kernel-internal form)
3761 * This function changes the 'effective' priority of a task. It does
3762 * not touch ->normal_prio like __setscheduler().
3764 * Used by the rt_mutex code to implement priority inheritance logic.
3766 void rt_mutex_setprio(struct task_struct *p, int prio)
3768 int oldprio, on_rq, running;
3770 const struct sched_class *prev_class;
3772 BUG_ON(prio < 0 || prio > MAX_PRIO);
3774 rq = __task_rq_lock(p);
3776 trace_sched_pi_setprio(p, prio);
3778 prev_class = p->sched_class;
3780 running = task_current(rq, p);
3782 dequeue_task(rq, p, 0);
3784 p->sched_class->put_prev_task(rq, p);
3787 p->sched_class = &rt_sched_class;
3789 p->sched_class = &fair_sched_class;
3794 p->sched_class->set_curr_task(rq);
3796 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3798 check_class_changed(rq, p, prev_class, oldprio);
3799 __task_rq_unlock(rq);
3804 void set_user_nice(struct task_struct *p, long nice)
3806 int old_prio, delta, on_rq;
3807 unsigned long flags;
3810 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3813 * We have to be careful, if called from sys_setpriority(),
3814 * the task might be in the middle of scheduling on another CPU.
3816 rq = task_rq_lock(p, &flags);
3818 * The RT priorities are set via sched_setscheduler(), but we still
3819 * allow the 'normal' nice value to be set - but as expected
3820 * it wont have any effect on scheduling until the task is
3821 * SCHED_FIFO/SCHED_RR:
3823 if (task_has_rt_policy(p)) {
3824 p->static_prio = NICE_TO_PRIO(nice);
3829 dequeue_task(rq, p, 0);
3831 p->static_prio = NICE_TO_PRIO(nice);
3834 p->prio = effective_prio(p);
3835 delta = p->prio - old_prio;
3838 enqueue_task(rq, p, 0);
3840 * If the task increased its priority or is running and
3841 * lowered its priority, then reschedule its CPU:
3843 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3844 resched_task(rq->curr);
3847 task_rq_unlock(rq, p, &flags);
3849 EXPORT_SYMBOL(set_user_nice);
3852 * can_nice - check if a task can reduce its nice value
3856 int can_nice(const struct task_struct *p, const int nice)
3858 /* convert nice value [19,-20] to rlimit style value [1,40] */
3859 int nice_rlim = 20 - nice;
3861 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3862 capable(CAP_SYS_NICE));
3865 #ifdef __ARCH_WANT_SYS_NICE
3868 * sys_nice - change the priority of the current process.
3869 * @increment: priority increment
3871 * sys_setpriority is a more generic, but much slower function that
3872 * does similar things.
3874 SYSCALL_DEFINE1(nice, int, increment)
3879 * Setpriority might change our priority at the same moment.
3880 * We don't have to worry. Conceptually one call occurs first
3881 * and we have a single winner.
3883 if (increment < -40)
3888 nice = TASK_NICE(current) + increment;
3894 if (increment < 0 && !can_nice(current, nice))
3897 retval = security_task_setnice(current, nice);
3901 set_user_nice(current, nice);
3908 * task_prio - return the priority value of a given task.
3909 * @p: the task in question.
3911 * This is the priority value as seen by users in /proc.
3912 * RT tasks are offset by -200. Normal tasks are centered
3913 * around 0, value goes from -16 to +15.
3915 int task_prio(const struct task_struct *p)
3917 return p->prio - MAX_RT_PRIO;
3921 * task_nice - return the nice value of a given task.
3922 * @p: the task in question.
3924 int task_nice(const struct task_struct *p)
3926 return TASK_NICE(p);
3928 EXPORT_SYMBOL(task_nice);
3931 * idle_cpu - is a given cpu idle currently?
3932 * @cpu: the processor in question.
3934 int idle_cpu(int cpu)
3936 struct rq *rq = cpu_rq(cpu);
3938 if (rq->curr != rq->idle)
3945 if (!llist_empty(&rq->wake_list))
3953 * idle_task - return the idle task for a given cpu.
3954 * @cpu: the processor in question.
3956 struct task_struct *idle_task(int cpu)
3958 return cpu_rq(cpu)->idle;
3962 * find_process_by_pid - find a process with a matching PID value.
3963 * @pid: the pid in question.
3965 static struct task_struct *find_process_by_pid(pid_t pid)
3967 return pid ? find_task_by_vpid(pid) : current;
3970 /* Actually do priority change: must hold rq lock. */
3972 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3975 p->rt_priority = prio;
3976 p->normal_prio = normal_prio(p);
3977 /* we are holding p->pi_lock already */
3978 p->prio = rt_mutex_getprio(p);
3979 if (rt_prio(p->prio))
3980 p->sched_class = &rt_sched_class;
3982 p->sched_class = &fair_sched_class;
3987 * check the target process has a UID that matches the current process's
3989 static bool check_same_owner(struct task_struct *p)
3991 const struct cred *cred = current_cred(), *pcred;
3995 pcred = __task_cred(p);
3996 if (cred->user->user_ns == pcred->user->user_ns)
3997 match = (cred->euid == pcred->euid ||
3998 cred->euid == pcred->uid);
4005 static int __sched_setscheduler(struct task_struct *p, int policy,
4006 const struct sched_param *param, bool user)
4008 int retval, oldprio, oldpolicy = -1, on_rq, running;
4009 unsigned long flags;
4010 const struct sched_class *prev_class;
4014 /* may grab non-irq protected spin_locks */
4015 BUG_ON(in_interrupt());
4017 /* double check policy once rq lock held */
4019 reset_on_fork = p->sched_reset_on_fork;
4020 policy = oldpolicy = p->policy;
4022 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4023 policy &= ~SCHED_RESET_ON_FORK;
4025 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4026 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4027 policy != SCHED_IDLE)
4032 * Valid priorities for SCHED_FIFO and SCHED_RR are
4033 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4034 * SCHED_BATCH and SCHED_IDLE is 0.
4036 if (param->sched_priority < 0 ||
4037 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4038 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4040 if (rt_policy(policy) != (param->sched_priority != 0))
4044 * Allow unprivileged RT tasks to decrease priority:
4046 if (user && !capable(CAP_SYS_NICE)) {
4047 if (rt_policy(policy)) {
4048 unsigned long rlim_rtprio =
4049 task_rlimit(p, RLIMIT_RTPRIO);
4051 /* can't set/change the rt policy */
4052 if (policy != p->policy && !rlim_rtprio)
4055 /* can't increase priority */
4056 if (param->sched_priority > p->rt_priority &&
4057 param->sched_priority > rlim_rtprio)
4062 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4063 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4065 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4066 if (!can_nice(p, TASK_NICE(p)))
4070 /* can't change other user's priorities */
4071 if (!check_same_owner(p))
4074 /* Normal users shall not reset the sched_reset_on_fork flag */
4075 if (p->sched_reset_on_fork && !reset_on_fork)
4080 retval = security_task_setscheduler(p);
4086 * make sure no PI-waiters arrive (or leave) while we are
4087 * changing the priority of the task:
4089 * To be able to change p->policy safely, the appropriate
4090 * runqueue lock must be held.
4092 rq = task_rq_lock(p, &flags);
4095 * Changing the policy of the stop threads its a very bad idea
4097 if (p == rq->stop) {
4098 task_rq_unlock(rq, p, &flags);
4103 * If not changing anything there's no need to proceed further:
4105 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4106 param->sched_priority == p->rt_priority))) {
4108 __task_rq_unlock(rq);
4109 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4113 #ifdef CONFIG_RT_GROUP_SCHED
4116 * Do not allow realtime tasks into groups that have no runtime
4119 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4120 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4121 !task_group_is_autogroup(task_group(p))) {
4122 task_rq_unlock(rq, p, &flags);
4128 /* recheck policy now with rq lock held */
4129 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4130 policy = oldpolicy = -1;
4131 task_rq_unlock(rq, p, &flags);
4135 running = task_current(rq, p);
4137 deactivate_task(rq, p, 0);
4139 p->sched_class->put_prev_task(rq, p);
4141 p->sched_reset_on_fork = reset_on_fork;
4144 prev_class = p->sched_class;
4145 __setscheduler(rq, p, policy, param->sched_priority);
4148 p->sched_class->set_curr_task(rq);
4150 activate_task(rq, p, 0);
4152 check_class_changed(rq, p, prev_class, oldprio);
4153 task_rq_unlock(rq, p, &flags);
4155 rt_mutex_adjust_pi(p);
4161 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4162 * @p: the task in question.
4163 * @policy: new policy.
4164 * @param: structure containing the new RT priority.
4166 * NOTE that the task may be already dead.
4168 int sched_setscheduler(struct task_struct *p, int policy,
4169 const struct sched_param *param)
4171 return __sched_setscheduler(p, policy, param, true);
4173 EXPORT_SYMBOL_GPL(sched_setscheduler);
4176 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4177 * @p: the task in question.
4178 * @policy: new policy.
4179 * @param: structure containing the new RT priority.
4181 * Just like sched_setscheduler, only don't bother checking if the
4182 * current context has permission. For example, this is needed in
4183 * stop_machine(): we create temporary high priority worker threads,
4184 * but our caller might not have that capability.
4186 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4187 const struct sched_param *param)
4189 return __sched_setscheduler(p, policy, param, false);
4193 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4195 struct sched_param lparam;
4196 struct task_struct *p;
4199 if (!param || pid < 0)
4201 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4206 p = find_process_by_pid(pid);
4208 retval = sched_setscheduler(p, policy, &lparam);
4215 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4216 * @pid: the pid in question.
4217 * @policy: new policy.
4218 * @param: structure containing the new RT priority.
4220 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4221 struct sched_param __user *, param)
4223 /* negative values for policy are not valid */
4227 return do_sched_setscheduler(pid, policy, param);
4231 * sys_sched_setparam - set/change the RT priority of a thread
4232 * @pid: the pid in question.
4233 * @param: structure containing the new RT priority.
4235 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4237 return do_sched_setscheduler(pid, -1, param);
4241 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4242 * @pid: the pid in question.
4244 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4246 struct task_struct *p;
4254 p = find_process_by_pid(pid);
4256 retval = security_task_getscheduler(p);
4259 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4266 * sys_sched_getparam - get the RT priority of a thread
4267 * @pid: the pid in question.
4268 * @param: structure containing the RT priority.
4270 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4272 struct sched_param lp;
4273 struct task_struct *p;
4276 if (!param || pid < 0)
4280 p = find_process_by_pid(pid);
4285 retval = security_task_getscheduler(p);
4289 lp.sched_priority = p->rt_priority;
4293 * This one might sleep, we cannot do it with a spinlock held ...
4295 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4304 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4306 cpumask_var_t cpus_allowed, new_mask;
4307 struct task_struct *p;
4313 p = find_process_by_pid(pid);
4320 /* Prevent p going away */
4324 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4328 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4330 goto out_free_cpus_allowed;
4333 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4336 retval = security_task_setscheduler(p);
4340 cpuset_cpus_allowed(p, cpus_allowed);
4341 cpumask_and(new_mask, in_mask, cpus_allowed);
4343 retval = set_cpus_allowed_ptr(p, new_mask);
4346 cpuset_cpus_allowed(p, cpus_allowed);
4347 if (!cpumask_subset(new_mask, cpus_allowed)) {
4349 * We must have raced with a concurrent cpuset
4350 * update. Just reset the cpus_allowed to the
4351 * cpuset's cpus_allowed
4353 cpumask_copy(new_mask, cpus_allowed);
4358 free_cpumask_var(new_mask);
4359 out_free_cpus_allowed:
4360 free_cpumask_var(cpus_allowed);
4367 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4368 struct cpumask *new_mask)
4370 if (len < cpumask_size())
4371 cpumask_clear(new_mask);
4372 else if (len > cpumask_size())
4373 len = cpumask_size();
4375 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4379 * sys_sched_setaffinity - set the cpu affinity of a process
4380 * @pid: pid of the process
4381 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4382 * @user_mask_ptr: user-space pointer to the new cpu mask
4384 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4385 unsigned long __user *, user_mask_ptr)
4387 cpumask_var_t new_mask;
4390 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4393 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4395 retval = sched_setaffinity(pid, new_mask);
4396 free_cpumask_var(new_mask);
4400 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4402 struct task_struct *p;
4403 unsigned long flags;
4410 p = find_process_by_pid(pid);
4414 retval = security_task_getscheduler(p);
4418 raw_spin_lock_irqsave(&p->pi_lock, flags);
4419 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4420 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4430 * sys_sched_getaffinity - get the cpu affinity of a process
4431 * @pid: pid of the process
4432 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4433 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4435 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4436 unsigned long __user *, user_mask_ptr)
4441 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4443 if (len & (sizeof(unsigned long)-1))
4446 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4449 ret = sched_getaffinity(pid, mask);
4451 size_t retlen = min_t(size_t, len, cpumask_size());
4453 if (copy_to_user(user_mask_ptr, mask, retlen))
4458 free_cpumask_var(mask);
4464 * sys_sched_yield - yield the current processor to other threads.
4466 * This function yields the current CPU to other tasks. If there are no
4467 * other threads running on this CPU then this function will return.
4469 SYSCALL_DEFINE0(sched_yield)
4471 struct rq *rq = this_rq_lock();
4473 schedstat_inc(rq, yld_count);
4474 current->sched_class->yield_task(rq);
4477 * Since we are going to call schedule() anyway, there's
4478 * no need to preempt or enable interrupts:
4480 __release(rq->lock);
4481 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4482 do_raw_spin_unlock(&rq->lock);
4483 preempt_enable_no_resched();
4490 static inline int should_resched(void)
4492 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4495 static void __cond_resched(void)
4497 add_preempt_count(PREEMPT_ACTIVE);
4499 sub_preempt_count(PREEMPT_ACTIVE);
4502 int __sched _cond_resched(void)
4504 if (should_resched()) {
4510 EXPORT_SYMBOL(_cond_resched);
4513 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4514 * call schedule, and on return reacquire the lock.
4516 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4517 * operations here to prevent schedule() from being called twice (once via
4518 * spin_unlock(), once by hand).
4520 int __cond_resched_lock(spinlock_t *lock)
4522 int resched = should_resched();
4525 lockdep_assert_held(lock);
4527 if (spin_needbreak(lock) || resched) {
4538 EXPORT_SYMBOL(__cond_resched_lock);
4540 int __sched __cond_resched_softirq(void)
4542 BUG_ON(!in_softirq());
4544 if (should_resched()) {
4552 EXPORT_SYMBOL(__cond_resched_softirq);
4555 * yield - yield the current processor to other threads.
4557 * This is a shortcut for kernel-space yielding - it marks the
4558 * thread runnable and calls sys_sched_yield().
4560 void __sched yield(void)
4562 set_current_state(TASK_RUNNING);
4565 EXPORT_SYMBOL(yield);
4568 * yield_to - yield the current processor to another thread in
4569 * your thread group, or accelerate that thread toward the
4570 * processor it's on.
4572 * @preempt: whether task preemption is allowed or not
4574 * It's the caller's job to ensure that the target task struct
4575 * can't go away on us before we can do any checks.
4577 * Returns true if we indeed boosted the target task.
4579 bool __sched yield_to(struct task_struct *p, bool preempt)
4581 struct task_struct *curr = current;
4582 struct rq *rq, *p_rq;
4583 unsigned long flags;
4586 local_irq_save(flags);
4591 double_rq_lock(rq, p_rq);
4592 while (task_rq(p) != p_rq) {
4593 double_rq_unlock(rq, p_rq);
4597 if (!curr->sched_class->yield_to_task)
4600 if (curr->sched_class != p->sched_class)
4603 if (task_running(p_rq, p) || p->state)
4606 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4608 schedstat_inc(rq, yld_count);
4610 * Make p's CPU reschedule; pick_next_entity takes care of
4613 if (preempt && rq != p_rq)
4614 resched_task(p_rq->curr);
4617 * We might have set it in task_yield_fair(), but are
4618 * not going to schedule(), so don't want to skip
4621 rq->skip_clock_update = 0;
4625 double_rq_unlock(rq, p_rq);
4626 local_irq_restore(flags);
4633 EXPORT_SYMBOL_GPL(yield_to);
4636 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4637 * that process accounting knows that this is a task in IO wait state.
4639 void __sched io_schedule(void)
4641 struct rq *rq = raw_rq();
4643 delayacct_blkio_start();
4644 atomic_inc(&rq->nr_iowait);
4645 blk_flush_plug(current);
4646 current->in_iowait = 1;
4648 current->in_iowait = 0;
4649 atomic_dec(&rq->nr_iowait);
4650 delayacct_blkio_end();
4652 EXPORT_SYMBOL(io_schedule);
4654 long __sched io_schedule_timeout(long timeout)
4656 struct rq *rq = raw_rq();
4659 delayacct_blkio_start();
4660 atomic_inc(&rq->nr_iowait);
4661 blk_flush_plug(current);
4662 current->in_iowait = 1;
4663 ret = schedule_timeout(timeout);
4664 current->in_iowait = 0;
4665 atomic_dec(&rq->nr_iowait);
4666 delayacct_blkio_end();
4671 * sys_sched_get_priority_max - return maximum RT priority.
4672 * @policy: scheduling class.
4674 * this syscall returns the maximum rt_priority that can be used
4675 * by a given scheduling class.
4677 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4684 ret = MAX_USER_RT_PRIO-1;
4696 * sys_sched_get_priority_min - return minimum RT priority.
4697 * @policy: scheduling class.
4699 * this syscall returns the minimum rt_priority that can be used
4700 * by a given scheduling class.
4702 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4720 * sys_sched_rr_get_interval - return the default timeslice of a process.
4721 * @pid: pid of the process.
4722 * @interval: userspace pointer to the timeslice value.
4724 * this syscall writes the default timeslice value of a given process
4725 * into the user-space timespec buffer. A value of '0' means infinity.
4727 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4728 struct timespec __user *, interval)
4730 struct task_struct *p;
4731 unsigned int time_slice;
4732 unsigned long flags;
4742 p = find_process_by_pid(pid);
4746 retval = security_task_getscheduler(p);
4750 rq = task_rq_lock(p, &flags);
4751 time_slice = p->sched_class->get_rr_interval(rq, p);
4752 task_rq_unlock(rq, p, &flags);
4755 jiffies_to_timespec(time_slice, &t);
4756 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4764 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4766 void sched_show_task(struct task_struct *p)
4768 unsigned long free = 0;
4771 state = p->state ? __ffs(p->state) + 1 : 0;
4772 printk(KERN_INFO "%-15.15s %c", p->comm,
4773 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4774 #if BITS_PER_LONG == 32
4775 if (state == TASK_RUNNING)
4776 printk(KERN_CONT " running ");
4778 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4780 if (state == TASK_RUNNING)
4781 printk(KERN_CONT " running task ");
4783 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4785 #ifdef CONFIG_DEBUG_STACK_USAGE
4786 free = stack_not_used(p);
4788 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4789 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4790 (unsigned long)task_thread_info(p)->flags);
4792 show_stack(p, NULL);
4795 void show_state_filter(unsigned long state_filter)
4797 struct task_struct *g, *p;
4799 #if BITS_PER_LONG == 32
4801 " task PC stack pid father\n");
4804 " task PC stack pid father\n");
4807 do_each_thread(g, p) {
4809 * reset the NMI-timeout, listing all files on a slow
4810 * console might take a lot of time:
4812 touch_nmi_watchdog();
4813 if (!state_filter || (p->state & state_filter))
4815 } while_each_thread(g, p);
4817 touch_all_softlockup_watchdogs();
4819 #ifdef CONFIG_SCHED_DEBUG
4820 sysrq_sched_debug_show();
4824 * Only show locks if all tasks are dumped:
4827 debug_show_all_locks();
4830 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4832 idle->sched_class = &idle_sched_class;
4836 * init_idle - set up an idle thread for a given CPU
4837 * @idle: task in question
4838 * @cpu: cpu the idle task belongs to
4840 * NOTE: this function does not set the idle thread's NEED_RESCHED
4841 * flag, to make booting more robust.
4843 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4845 struct rq *rq = cpu_rq(cpu);
4846 unsigned long flags;
4848 raw_spin_lock_irqsave(&rq->lock, flags);
4851 idle->state = TASK_RUNNING;
4852 idle->se.exec_start = sched_clock();
4854 do_set_cpus_allowed(idle, cpumask_of(cpu));
4856 * We're having a chicken and egg problem, even though we are
4857 * holding rq->lock, the cpu isn't yet set to this cpu so the
4858 * lockdep check in task_group() will fail.
4860 * Similar case to sched_fork(). / Alternatively we could
4861 * use task_rq_lock() here and obtain the other rq->lock.
4866 __set_task_cpu(idle, cpu);
4869 rq->curr = rq->idle = idle;
4870 #if defined(CONFIG_SMP)
4873 raw_spin_unlock_irqrestore(&rq->lock, flags);
4875 /* Set the preempt count _outside_ the spinlocks! */
4876 task_thread_info(idle)->preempt_count = 0;
4879 * The idle tasks have their own, simple scheduling class:
4881 idle->sched_class = &idle_sched_class;
4882 ftrace_graph_init_idle_task(idle, cpu);
4883 #if defined(CONFIG_SMP)
4884 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4889 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4891 if (p->sched_class && p->sched_class->set_cpus_allowed)
4892 p->sched_class->set_cpus_allowed(p, new_mask);
4894 cpumask_copy(&p->cpus_allowed, new_mask);
4895 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
4899 * This is how migration works:
4901 * 1) we invoke migration_cpu_stop() on the target CPU using
4903 * 2) stopper starts to run (implicitly forcing the migrated thread
4905 * 3) it checks whether the migrated task is still in the wrong runqueue.
4906 * 4) if it's in the wrong runqueue then the migration thread removes
4907 * it and puts it into the right queue.
4908 * 5) stopper completes and stop_one_cpu() returns and the migration
4913 * Change a given task's CPU affinity. Migrate the thread to a
4914 * proper CPU and schedule it away if the CPU it's executing on
4915 * is removed from the allowed bitmask.
4917 * NOTE: the caller must have a valid reference to the task, the
4918 * task must not exit() & deallocate itself prematurely. The
4919 * call is not atomic; no spinlocks may be held.
4921 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4923 unsigned long flags;
4925 unsigned int dest_cpu;
4928 rq = task_rq_lock(p, &flags);
4930 if (cpumask_equal(&p->cpus_allowed, new_mask))
4933 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4938 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4943 do_set_cpus_allowed(p, new_mask);
4945 /* Can the task run on the task's current CPU? If so, we're done */
4946 if (cpumask_test_cpu(task_cpu(p), new_mask))
4949 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4951 struct migration_arg arg = { p, dest_cpu };
4952 /* Need help from migration thread: drop lock and wait. */
4953 task_rq_unlock(rq, p, &flags);
4954 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4955 tlb_migrate_finish(p->mm);
4959 task_rq_unlock(rq, p, &flags);
4963 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4966 * Move (not current) task off this cpu, onto dest cpu. We're doing
4967 * this because either it can't run here any more (set_cpus_allowed()
4968 * away from this CPU, or CPU going down), or because we're
4969 * attempting to rebalance this task on exec (sched_exec).
4971 * So we race with normal scheduler movements, but that's OK, as long
4972 * as the task is no longer on this CPU.
4974 * Returns non-zero if task was successfully migrated.
4976 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4978 struct rq *rq_dest, *rq_src;
4981 if (unlikely(!cpu_active(dest_cpu)))
4984 rq_src = cpu_rq(src_cpu);
4985 rq_dest = cpu_rq(dest_cpu);
4987 raw_spin_lock(&p->pi_lock);
4988 double_rq_lock(rq_src, rq_dest);
4989 /* Already moved. */
4990 if (task_cpu(p) != src_cpu)
4992 /* Affinity changed (again). */
4993 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4997 * If we're not on a rq, the next wake-up will ensure we're
5001 deactivate_task(rq_src, p, 0);
5002 set_task_cpu(p, dest_cpu);
5003 activate_task(rq_dest, p, 0);
5004 check_preempt_curr(rq_dest, p, 0);
5009 double_rq_unlock(rq_src, rq_dest);
5010 raw_spin_unlock(&p->pi_lock);
5015 * migration_cpu_stop - this will be executed by a highprio stopper thread
5016 * and performs thread migration by bumping thread off CPU then
5017 * 'pushing' onto another runqueue.
5019 static int migration_cpu_stop(void *data)
5021 struct migration_arg *arg = data;
5024 * The original target cpu might have gone down and we might
5025 * be on another cpu but it doesn't matter.
5027 local_irq_disable();
5028 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5033 #ifdef CONFIG_HOTPLUG_CPU
5036 * Ensures that the idle task is using init_mm right before its cpu goes
5039 void idle_task_exit(void)
5041 struct mm_struct *mm = current->active_mm;
5043 BUG_ON(cpu_online(smp_processor_id()));
5046 switch_mm(mm, &init_mm, current);
5051 * While a dead CPU has no uninterruptible tasks queued at this point,
5052 * it might still have a nonzero ->nr_uninterruptible counter, because
5053 * for performance reasons the counter is not stricly tracking tasks to
5054 * their home CPUs. So we just add the counter to another CPU's counter,
5055 * to keep the global sum constant after CPU-down:
5057 static void migrate_nr_uninterruptible(struct rq *rq_src)
5059 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5061 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5062 rq_src->nr_uninterruptible = 0;
5066 * remove the tasks which were accounted by rq from calc_load_tasks.
5068 static void calc_global_load_remove(struct rq *rq)
5070 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5071 rq->calc_load_active = 0;
5075 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5076 * try_to_wake_up()->select_task_rq().
5078 * Called with rq->lock held even though we'er in stop_machine() and
5079 * there's no concurrency possible, we hold the required locks anyway
5080 * because of lock validation efforts.
5082 static void migrate_tasks(unsigned int dead_cpu)
5084 struct rq *rq = cpu_rq(dead_cpu);
5085 struct task_struct *next, *stop = rq->stop;
5089 * Fudge the rq selection such that the below task selection loop
5090 * doesn't get stuck on the currently eligible stop task.
5092 * We're currently inside stop_machine() and the rq is either stuck
5093 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5094 * either way we should never end up calling schedule() until we're
5099 /* Ensure any throttled groups are reachable by pick_next_task */
5100 unthrottle_offline_cfs_rqs(rq);
5104 * There's this thread running, bail when that's the only
5107 if (rq->nr_running == 1)
5110 next = pick_next_task(rq);
5112 next->sched_class->put_prev_task(rq, next);
5114 /* Find suitable destination for @next, with force if needed. */
5115 dest_cpu = select_fallback_rq(dead_cpu, next);
5116 raw_spin_unlock(&rq->lock);
5118 __migrate_task(next, dead_cpu, dest_cpu);
5120 raw_spin_lock(&rq->lock);
5126 #endif /* CONFIG_HOTPLUG_CPU */
5128 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5130 static struct ctl_table sd_ctl_dir[] = {
5132 .procname = "sched_domain",
5138 static struct ctl_table sd_ctl_root[] = {
5140 .procname = "kernel",
5142 .child = sd_ctl_dir,
5147 static struct ctl_table *sd_alloc_ctl_entry(int n)
5149 struct ctl_table *entry =
5150 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5155 static void sd_free_ctl_entry(struct ctl_table **tablep)
5157 struct ctl_table *entry;
5160 * In the intermediate directories, both the child directory and
5161 * procname are dynamically allocated and could fail but the mode
5162 * will always be set. In the lowest directory the names are
5163 * static strings and all have proc handlers.
5165 for (entry = *tablep; entry->mode; entry++) {
5167 sd_free_ctl_entry(&entry->child);
5168 if (entry->proc_handler == NULL)
5169 kfree(entry->procname);
5177 set_table_entry(struct ctl_table *entry,
5178 const char *procname, void *data, int maxlen,
5179 umode_t mode, proc_handler *proc_handler)
5181 entry->procname = procname;
5183 entry->maxlen = maxlen;
5185 entry->proc_handler = proc_handler;
5188 static struct ctl_table *
5189 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5191 struct ctl_table *table = sd_alloc_ctl_entry(13);
5196 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5197 sizeof(long), 0644, proc_doulongvec_minmax);
5198 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5199 sizeof(long), 0644, proc_doulongvec_minmax);
5200 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5201 sizeof(int), 0644, proc_dointvec_minmax);
5202 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5203 sizeof(int), 0644, proc_dointvec_minmax);
5204 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5205 sizeof(int), 0644, proc_dointvec_minmax);
5206 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5207 sizeof(int), 0644, proc_dointvec_minmax);
5208 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5209 sizeof(int), 0644, proc_dointvec_minmax);
5210 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5211 sizeof(int), 0644, proc_dointvec_minmax);
5212 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5213 sizeof(int), 0644, proc_dointvec_minmax);
5214 set_table_entry(&table[9], "cache_nice_tries",
5215 &sd->cache_nice_tries,
5216 sizeof(int), 0644, proc_dointvec_minmax);
5217 set_table_entry(&table[10], "flags", &sd->flags,
5218 sizeof(int), 0644, proc_dointvec_minmax);
5219 set_table_entry(&table[11], "name", sd->name,
5220 CORENAME_MAX_SIZE, 0444, proc_dostring);
5221 /* &table[12] is terminator */
5226 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5228 struct ctl_table *entry, *table;
5229 struct sched_domain *sd;
5230 int domain_num = 0, i;
5233 for_each_domain(cpu, sd)
5235 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5240 for_each_domain(cpu, sd) {
5241 snprintf(buf, 32, "domain%d", i);
5242 entry->procname = kstrdup(buf, GFP_KERNEL);
5244 entry->child = sd_alloc_ctl_domain_table(sd);
5251 static struct ctl_table_header *sd_sysctl_header;
5252 static void register_sched_domain_sysctl(void)
5254 int i, cpu_num = num_possible_cpus();
5255 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5258 WARN_ON(sd_ctl_dir[0].child);
5259 sd_ctl_dir[0].child = entry;
5264 for_each_possible_cpu(i) {
5265 snprintf(buf, 32, "cpu%d", i);
5266 entry->procname = kstrdup(buf, GFP_KERNEL);
5268 entry->child = sd_alloc_ctl_cpu_table(i);
5272 WARN_ON(sd_sysctl_header);
5273 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5276 /* may be called multiple times per register */
5277 static void unregister_sched_domain_sysctl(void)
5279 if (sd_sysctl_header)
5280 unregister_sysctl_table(sd_sysctl_header);
5281 sd_sysctl_header = NULL;
5282 if (sd_ctl_dir[0].child)
5283 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5286 static void register_sched_domain_sysctl(void)
5289 static void unregister_sched_domain_sysctl(void)
5294 static void set_rq_online(struct rq *rq)
5297 const struct sched_class *class;
5299 cpumask_set_cpu(rq->cpu, rq->rd->online);
5302 for_each_class(class) {
5303 if (class->rq_online)
5304 class->rq_online(rq);
5309 static void set_rq_offline(struct rq *rq)
5312 const struct sched_class *class;
5314 for_each_class(class) {
5315 if (class->rq_offline)
5316 class->rq_offline(rq);
5319 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5325 * migration_call - callback that gets triggered when a CPU is added.
5326 * Here we can start up the necessary migration thread for the new CPU.
5328 static int __cpuinit
5329 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5331 int cpu = (long)hcpu;
5332 unsigned long flags;
5333 struct rq *rq = cpu_rq(cpu);
5335 switch (action & ~CPU_TASKS_FROZEN) {
5337 case CPU_UP_PREPARE:
5338 rq->calc_load_update = calc_load_update;
5342 /* Update our root-domain */
5343 raw_spin_lock_irqsave(&rq->lock, flags);
5345 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5349 raw_spin_unlock_irqrestore(&rq->lock, flags);
5352 #ifdef CONFIG_HOTPLUG_CPU
5354 sched_ttwu_pending();
5355 /* Update our root-domain */
5356 raw_spin_lock_irqsave(&rq->lock, flags);
5358 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5362 BUG_ON(rq->nr_running != 1); /* the migration thread */
5363 raw_spin_unlock_irqrestore(&rq->lock, flags);
5365 migrate_nr_uninterruptible(rq);
5366 calc_global_load_remove(rq);
5371 update_max_interval();
5377 * Register at high priority so that task migration (migrate_all_tasks)
5378 * happens before everything else. This has to be lower priority than
5379 * the notifier in the perf_event subsystem, though.
5381 static struct notifier_block __cpuinitdata migration_notifier = {
5382 .notifier_call = migration_call,
5383 .priority = CPU_PRI_MIGRATION,
5386 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5387 unsigned long action, void *hcpu)
5389 switch (action & ~CPU_TASKS_FROZEN) {
5391 case CPU_DOWN_FAILED:
5392 set_cpu_active((long)hcpu, true);
5399 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5400 unsigned long action, void *hcpu)
5402 switch (action & ~CPU_TASKS_FROZEN) {
5403 case CPU_DOWN_PREPARE:
5404 set_cpu_active((long)hcpu, false);
5411 static int __init migration_init(void)
5413 void *cpu = (void *)(long)smp_processor_id();
5416 /* Initialize migration for the boot CPU */
5417 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5418 BUG_ON(err == NOTIFY_BAD);
5419 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5420 register_cpu_notifier(&migration_notifier);
5422 /* Register cpu active notifiers */
5423 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5424 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5428 early_initcall(migration_init);
5433 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5435 #ifdef CONFIG_SCHED_DEBUG
5437 static __read_mostly int sched_domain_debug_enabled;
5439 static int __init sched_domain_debug_setup(char *str)
5441 sched_domain_debug_enabled = 1;
5445 early_param("sched_debug", sched_domain_debug_setup);
5447 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5448 struct cpumask *groupmask)
5450 struct sched_group *group = sd->groups;
5453 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5454 cpumask_clear(groupmask);
5456 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5458 if (!(sd->flags & SD_LOAD_BALANCE)) {
5459 printk("does not load-balance\n");
5461 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5466 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5468 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5469 printk(KERN_ERR "ERROR: domain->span does not contain "
5472 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5473 printk(KERN_ERR "ERROR: domain->groups does not contain"
5477 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5481 printk(KERN_ERR "ERROR: group is NULL\n");
5485 if (!group->sgp->power) {
5486 printk(KERN_CONT "\n");
5487 printk(KERN_ERR "ERROR: domain->cpu_power not "
5492 if (!cpumask_weight(sched_group_cpus(group))) {
5493 printk(KERN_CONT "\n");
5494 printk(KERN_ERR "ERROR: empty group\n");
5498 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5499 printk(KERN_CONT "\n");
5500 printk(KERN_ERR "ERROR: repeated CPUs\n");
5504 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5506 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5508 printk(KERN_CONT " %s", str);
5509 if (group->sgp->power != SCHED_POWER_SCALE) {
5510 printk(KERN_CONT " (cpu_power = %d)",
5514 group = group->next;
5515 } while (group != sd->groups);
5516 printk(KERN_CONT "\n");
5518 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5519 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5522 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5523 printk(KERN_ERR "ERROR: parent span is not a superset "
5524 "of domain->span\n");
5528 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5532 if (!sched_domain_debug_enabled)
5536 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5540 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5543 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5551 #else /* !CONFIG_SCHED_DEBUG */
5552 # define sched_domain_debug(sd, cpu) do { } while (0)
5553 #endif /* CONFIG_SCHED_DEBUG */
5555 static int sd_degenerate(struct sched_domain *sd)
5557 if (cpumask_weight(sched_domain_span(sd)) == 1)
5560 /* Following flags need at least 2 groups */
5561 if (sd->flags & (SD_LOAD_BALANCE |
5562 SD_BALANCE_NEWIDLE |
5566 SD_SHARE_PKG_RESOURCES)) {
5567 if (sd->groups != sd->groups->next)
5571 /* Following flags don't use groups */
5572 if (sd->flags & (SD_WAKE_AFFINE))
5579 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5581 unsigned long cflags = sd->flags, pflags = parent->flags;
5583 if (sd_degenerate(parent))
5586 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5589 /* Flags needing groups don't count if only 1 group in parent */
5590 if (parent->groups == parent->groups->next) {
5591 pflags &= ~(SD_LOAD_BALANCE |
5592 SD_BALANCE_NEWIDLE |
5596 SD_SHARE_PKG_RESOURCES);
5597 if (nr_node_ids == 1)
5598 pflags &= ~SD_SERIALIZE;
5600 if (~cflags & pflags)
5606 static void free_rootdomain(struct rcu_head *rcu)
5608 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5610 cpupri_cleanup(&rd->cpupri);
5611 free_cpumask_var(rd->rto_mask);
5612 free_cpumask_var(rd->online);
5613 free_cpumask_var(rd->span);
5617 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5619 struct root_domain *old_rd = NULL;
5620 unsigned long flags;
5622 raw_spin_lock_irqsave(&rq->lock, flags);
5627 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5630 cpumask_clear_cpu(rq->cpu, old_rd->span);
5633 * If we dont want to free the old_rt yet then
5634 * set old_rd to NULL to skip the freeing later
5637 if (!atomic_dec_and_test(&old_rd->refcount))
5641 atomic_inc(&rd->refcount);
5644 cpumask_set_cpu(rq->cpu, rd->span);
5645 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5648 raw_spin_unlock_irqrestore(&rq->lock, flags);
5651 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5654 static int init_rootdomain(struct root_domain *rd)
5656 memset(rd, 0, sizeof(*rd));
5658 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5660 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5662 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5665 if (cpupri_init(&rd->cpupri) != 0)
5670 free_cpumask_var(rd->rto_mask);
5672 free_cpumask_var(rd->online);
5674 free_cpumask_var(rd->span);
5680 * By default the system creates a single root-domain with all cpus as
5681 * members (mimicking the global state we have today).
5683 struct root_domain def_root_domain;
5685 static void init_defrootdomain(void)
5687 init_rootdomain(&def_root_domain);
5689 atomic_set(&def_root_domain.refcount, 1);
5692 static struct root_domain *alloc_rootdomain(void)
5694 struct root_domain *rd;
5696 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5700 if (init_rootdomain(rd) != 0) {
5708 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5710 struct sched_group *tmp, *first;
5719 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5724 } while (sg != first);
5727 static void free_sched_domain(struct rcu_head *rcu)
5729 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5732 * If its an overlapping domain it has private groups, iterate and
5735 if (sd->flags & SD_OVERLAP) {
5736 free_sched_groups(sd->groups, 1);
5737 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5738 kfree(sd->groups->sgp);
5744 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5746 call_rcu(&sd->rcu, free_sched_domain);
5749 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5751 for (; sd; sd = sd->parent)
5752 destroy_sched_domain(sd, cpu);
5756 * Keep a special pointer to the highest sched_domain that has
5757 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5758 * allows us to avoid some pointer chasing select_idle_sibling().
5760 * Also keep a unique ID per domain (we use the first cpu number in
5761 * the cpumask of the domain), this allows us to quickly tell if
5762 * two cpus are in the same cache domain, see ttwu_share_cache().
5764 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5765 DEFINE_PER_CPU(int, sd_llc_id);
5767 static void update_top_cache_domain(int cpu)
5769 struct sched_domain *sd;
5772 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5774 id = cpumask_first(sched_domain_span(sd));
5776 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5777 per_cpu(sd_llc_id, cpu) = id;
5781 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5782 * hold the hotplug lock.
5785 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5787 struct rq *rq = cpu_rq(cpu);
5788 struct sched_domain *tmp;
5790 /* Remove the sched domains which do not contribute to scheduling. */
5791 for (tmp = sd; tmp; ) {
5792 struct sched_domain *parent = tmp->parent;
5796 if (sd_parent_degenerate(tmp, parent)) {
5797 tmp->parent = parent->parent;
5799 parent->parent->child = tmp;
5800 destroy_sched_domain(parent, cpu);
5805 if (sd && sd_degenerate(sd)) {
5808 destroy_sched_domain(tmp, cpu);
5813 sched_domain_debug(sd, cpu);
5815 rq_attach_root(rq, rd);
5817 rcu_assign_pointer(rq->sd, sd);
5818 destroy_sched_domains(tmp, cpu);
5820 update_top_cache_domain(cpu);
5823 /* cpus with isolated domains */
5824 static cpumask_var_t cpu_isolated_map;
5826 /* Setup the mask of cpus configured for isolated domains */
5827 static int __init isolated_cpu_setup(char *str)
5829 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5830 cpulist_parse(str, cpu_isolated_map);
5834 __setup("isolcpus=", isolated_cpu_setup);
5839 * find_next_best_node - find the next node to include in a sched_domain
5840 * @node: node whose sched_domain we're building
5841 * @used_nodes: nodes already in the sched_domain
5843 * Find the next node to include in a given scheduling domain. Simply
5844 * finds the closest node not already in the @used_nodes map.
5846 * Should use nodemask_t.
5848 static int find_next_best_node(int node, nodemask_t *used_nodes)
5850 int i, n, val, min_val, best_node = -1;
5854 for (i = 0; i < nr_node_ids; i++) {
5855 /* Start at @node */
5856 n = (node + i) % nr_node_ids;
5858 if (!nr_cpus_node(n))
5861 /* Skip already used nodes */
5862 if (node_isset(n, *used_nodes))
5865 /* Simple min distance search */
5866 val = node_distance(node, n);
5868 if (val < min_val) {
5874 if (best_node != -1)
5875 node_set(best_node, *used_nodes);
5880 * sched_domain_node_span - get a cpumask for a node's sched_domain
5881 * @node: node whose cpumask we're constructing
5882 * @span: resulting cpumask
5884 * Given a node, construct a good cpumask for its sched_domain to span. It
5885 * should be one that prevents unnecessary balancing, but also spreads tasks
5888 static void sched_domain_node_span(int node, struct cpumask *span)
5890 nodemask_t used_nodes;
5893 cpumask_clear(span);
5894 nodes_clear(used_nodes);
5896 cpumask_or(span, span, cpumask_of_node(node));
5897 node_set(node, used_nodes);
5899 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
5900 int next_node = find_next_best_node(node, &used_nodes);
5903 cpumask_or(span, span, cpumask_of_node(next_node));
5907 static const struct cpumask *cpu_node_mask(int cpu)
5909 lockdep_assert_held(&sched_domains_mutex);
5911 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
5913 return sched_domains_tmpmask;
5916 static const struct cpumask *cpu_allnodes_mask(int cpu)
5918 return cpu_possible_mask;
5920 #endif /* CONFIG_NUMA */
5922 static const struct cpumask *cpu_cpu_mask(int cpu)
5924 return cpumask_of_node(cpu_to_node(cpu));
5927 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
5930 struct sched_domain **__percpu sd;
5931 struct sched_group **__percpu sg;
5932 struct sched_group_power **__percpu sgp;
5936 struct sched_domain ** __percpu sd;
5937 struct root_domain *rd;
5947 struct sched_domain_topology_level;
5949 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5950 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5952 #define SDTL_OVERLAP 0x01
5954 struct sched_domain_topology_level {
5955 sched_domain_init_f init;
5956 sched_domain_mask_f mask;
5958 struct sd_data data;
5962 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5964 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5965 const struct cpumask *span = sched_domain_span(sd);
5966 struct cpumask *covered = sched_domains_tmpmask;
5967 struct sd_data *sdd = sd->private;
5968 struct sched_domain *child;
5971 cpumask_clear(covered);
5973 for_each_cpu(i, span) {
5974 struct cpumask *sg_span;
5976 if (cpumask_test_cpu(i, covered))
5979 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5980 GFP_KERNEL, cpu_to_node(cpu));
5985 sg_span = sched_group_cpus(sg);
5987 child = *per_cpu_ptr(sdd->sd, i);
5989 child = child->child;
5990 cpumask_copy(sg_span, sched_domain_span(child));
5992 cpumask_set_cpu(i, sg_span);
5994 cpumask_or(covered, covered, sg_span);
5996 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
5997 atomic_inc(&sg->sgp->ref);
5999 if (cpumask_test_cpu(cpu, sg_span))
6009 sd->groups = groups;
6014 free_sched_groups(first, 0);
6019 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6021 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6022 struct sched_domain *child = sd->child;
6025 cpu = cpumask_first(sched_domain_span(child));
6028 *sg = *per_cpu_ptr(sdd->sg, cpu);
6029 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6030 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6037 * build_sched_groups will build a circular linked list of the groups
6038 * covered by the given span, and will set each group's ->cpumask correctly,
6039 * and ->cpu_power to 0.
6041 * Assumes the sched_domain tree is fully constructed
6044 build_sched_groups(struct sched_domain *sd, int cpu)
6046 struct sched_group *first = NULL, *last = NULL;
6047 struct sd_data *sdd = sd->private;
6048 const struct cpumask *span = sched_domain_span(sd);
6049 struct cpumask *covered;
6052 get_group(cpu, sdd, &sd->groups);
6053 atomic_inc(&sd->groups->ref);
6055 if (cpu != cpumask_first(sched_domain_span(sd)))
6058 lockdep_assert_held(&sched_domains_mutex);
6059 covered = sched_domains_tmpmask;
6061 cpumask_clear(covered);
6063 for_each_cpu(i, span) {
6064 struct sched_group *sg;
6065 int group = get_group(i, sdd, &sg);
6068 if (cpumask_test_cpu(i, covered))
6071 cpumask_clear(sched_group_cpus(sg));
6074 for_each_cpu(j, span) {
6075 if (get_group(j, sdd, NULL) != group)
6078 cpumask_set_cpu(j, covered);
6079 cpumask_set_cpu(j, sched_group_cpus(sg));
6094 * Initialize sched groups cpu_power.
6096 * cpu_power indicates the capacity of sched group, which is used while
6097 * distributing the load between different sched groups in a sched domain.
6098 * Typically cpu_power for all the groups in a sched domain will be same unless
6099 * there are asymmetries in the topology. If there are asymmetries, group
6100 * having more cpu_power will pickup more load compared to the group having
6103 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6105 struct sched_group *sg = sd->groups;
6107 WARN_ON(!sd || !sg);
6110 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6112 } while (sg != sd->groups);
6114 if (cpu != group_first_cpu(sg))
6117 update_group_power(sd, cpu);
6118 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6121 int __weak arch_sd_sibling_asym_packing(void)
6123 return 0*SD_ASYM_PACKING;
6127 * Initializers for schedule domains
6128 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6131 #ifdef CONFIG_SCHED_DEBUG
6132 # define SD_INIT_NAME(sd, type) sd->name = #type
6134 # define SD_INIT_NAME(sd, type) do { } while (0)
6137 #define SD_INIT_FUNC(type) \
6138 static noinline struct sched_domain * \
6139 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6141 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6142 *sd = SD_##type##_INIT; \
6143 SD_INIT_NAME(sd, type); \
6144 sd->private = &tl->data; \
6150 SD_INIT_FUNC(ALLNODES)
6153 #ifdef CONFIG_SCHED_SMT
6154 SD_INIT_FUNC(SIBLING)
6156 #ifdef CONFIG_SCHED_MC
6159 #ifdef CONFIG_SCHED_BOOK
6163 static int default_relax_domain_level = -1;
6164 int sched_domain_level_max;
6166 static int __init setup_relax_domain_level(char *str)
6170 val = simple_strtoul(str, NULL, 0);
6171 if (val < sched_domain_level_max)
6172 default_relax_domain_level = val;
6176 __setup("relax_domain_level=", setup_relax_domain_level);
6178 static void set_domain_attribute(struct sched_domain *sd,
6179 struct sched_domain_attr *attr)
6183 if (!attr || attr->relax_domain_level < 0) {
6184 if (default_relax_domain_level < 0)
6187 request = default_relax_domain_level;
6189 request = attr->relax_domain_level;
6190 if (request < sd->level) {
6191 /* turn off idle balance on this domain */
6192 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6194 /* turn on idle balance on this domain */
6195 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6199 static void __sdt_free(const struct cpumask *cpu_map);
6200 static int __sdt_alloc(const struct cpumask *cpu_map);
6202 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6203 const struct cpumask *cpu_map)
6207 if (!atomic_read(&d->rd->refcount))
6208 free_rootdomain(&d->rd->rcu); /* fall through */
6210 free_percpu(d->sd); /* fall through */
6212 __sdt_free(cpu_map); /* fall through */
6218 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6219 const struct cpumask *cpu_map)
6221 memset(d, 0, sizeof(*d));
6223 if (__sdt_alloc(cpu_map))
6224 return sa_sd_storage;
6225 d->sd = alloc_percpu(struct sched_domain *);
6227 return sa_sd_storage;
6228 d->rd = alloc_rootdomain();
6231 return sa_rootdomain;
6235 * NULL the sd_data elements we've used to build the sched_domain and
6236 * sched_group structure so that the subsequent __free_domain_allocs()
6237 * will not free the data we're using.
6239 static void claim_allocations(int cpu, struct sched_domain *sd)
6241 struct sd_data *sdd = sd->private;
6243 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6244 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6246 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6247 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6249 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6250 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6253 #ifdef CONFIG_SCHED_SMT
6254 static const struct cpumask *cpu_smt_mask(int cpu)
6256 return topology_thread_cpumask(cpu);
6261 * Topology list, bottom-up.
6263 static struct sched_domain_topology_level default_topology[] = {
6264 #ifdef CONFIG_SCHED_SMT
6265 { sd_init_SIBLING, cpu_smt_mask, },
6267 #ifdef CONFIG_SCHED_MC
6268 { sd_init_MC, cpu_coregroup_mask, },
6270 #ifdef CONFIG_SCHED_BOOK
6271 { sd_init_BOOK, cpu_book_mask, },
6273 { sd_init_CPU, cpu_cpu_mask, },
6275 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6276 { sd_init_ALLNODES, cpu_allnodes_mask, },
6281 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6283 static int __sdt_alloc(const struct cpumask *cpu_map)
6285 struct sched_domain_topology_level *tl;
6288 for (tl = sched_domain_topology; tl->init; tl++) {
6289 struct sd_data *sdd = &tl->data;
6291 sdd->sd = alloc_percpu(struct sched_domain *);
6295 sdd->sg = alloc_percpu(struct sched_group *);
6299 sdd->sgp = alloc_percpu(struct sched_group_power *);
6303 for_each_cpu(j, cpu_map) {
6304 struct sched_domain *sd;
6305 struct sched_group *sg;
6306 struct sched_group_power *sgp;
6308 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6309 GFP_KERNEL, cpu_to_node(j));
6313 *per_cpu_ptr(sdd->sd, j) = sd;
6315 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6316 GFP_KERNEL, cpu_to_node(j));
6320 *per_cpu_ptr(sdd->sg, j) = sg;
6322 sgp = kzalloc_node(sizeof(struct sched_group_power),
6323 GFP_KERNEL, cpu_to_node(j));
6327 *per_cpu_ptr(sdd->sgp, j) = sgp;
6334 static void __sdt_free(const struct cpumask *cpu_map)
6336 struct sched_domain_topology_level *tl;
6339 for (tl = sched_domain_topology; tl->init; tl++) {
6340 struct sd_data *sdd = &tl->data;
6342 for_each_cpu(j, cpu_map) {
6343 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
6344 if (sd && (sd->flags & SD_OVERLAP))
6345 free_sched_groups(sd->groups, 0);
6346 kfree(*per_cpu_ptr(sdd->sd, j));
6347 kfree(*per_cpu_ptr(sdd->sg, j));
6348 kfree(*per_cpu_ptr(sdd->sgp, j));
6350 free_percpu(sdd->sd);
6351 free_percpu(sdd->sg);
6352 free_percpu(sdd->sgp);
6356 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6357 struct s_data *d, const struct cpumask *cpu_map,
6358 struct sched_domain_attr *attr, struct sched_domain *child,
6361 struct sched_domain *sd = tl->init(tl, cpu);
6365 set_domain_attribute(sd, attr);
6366 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6368 sd->level = child->level + 1;
6369 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6378 * Build sched domains for a given set of cpus and attach the sched domains
6379 * to the individual cpus
6381 static int build_sched_domains(const struct cpumask *cpu_map,
6382 struct sched_domain_attr *attr)
6384 enum s_alloc alloc_state = sa_none;
6385 struct sched_domain *sd;
6387 int i, ret = -ENOMEM;
6389 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6390 if (alloc_state != sa_rootdomain)
6393 /* Set up domains for cpus specified by the cpu_map. */
6394 for_each_cpu(i, cpu_map) {
6395 struct sched_domain_topology_level *tl;
6398 for (tl = sched_domain_topology; tl->init; tl++) {
6399 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6400 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6401 sd->flags |= SD_OVERLAP;
6402 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6409 *per_cpu_ptr(d.sd, i) = sd;
6412 /* Build the groups for the domains */
6413 for_each_cpu(i, cpu_map) {
6414 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6415 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6416 if (sd->flags & SD_OVERLAP) {
6417 if (build_overlap_sched_groups(sd, i))
6420 if (build_sched_groups(sd, i))
6426 /* Calculate CPU power for physical packages and nodes */
6427 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6428 if (!cpumask_test_cpu(i, cpu_map))
6431 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6432 claim_allocations(i, sd);
6433 init_sched_groups_power(i, sd);
6437 /* Attach the domains */
6439 for_each_cpu(i, cpu_map) {
6440 sd = *per_cpu_ptr(d.sd, i);
6441 cpu_attach_domain(sd, d.rd, i);
6447 __free_domain_allocs(&d, alloc_state, cpu_map);
6451 static cpumask_var_t *doms_cur; /* current sched domains */
6452 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6453 static struct sched_domain_attr *dattr_cur;
6454 /* attribues of custom domains in 'doms_cur' */
6457 * Special case: If a kmalloc of a doms_cur partition (array of
6458 * cpumask) fails, then fallback to a single sched domain,
6459 * as determined by the single cpumask fallback_doms.
6461 static cpumask_var_t fallback_doms;
6464 * arch_update_cpu_topology lets virtualized architectures update the
6465 * cpu core maps. It is supposed to return 1 if the topology changed
6466 * or 0 if it stayed the same.
6468 int __attribute__((weak)) arch_update_cpu_topology(void)
6473 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6476 cpumask_var_t *doms;
6478 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6481 for (i = 0; i < ndoms; i++) {
6482 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6483 free_sched_domains(doms, i);
6490 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6493 for (i = 0; i < ndoms; i++)
6494 free_cpumask_var(doms[i]);
6499 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6500 * For now this just excludes isolated cpus, but could be used to
6501 * exclude other special cases in the future.
6503 static int init_sched_domains(const struct cpumask *cpu_map)
6507 arch_update_cpu_topology();
6509 doms_cur = alloc_sched_domains(ndoms_cur);
6511 doms_cur = &fallback_doms;
6512 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6514 err = build_sched_domains(doms_cur[0], NULL);
6515 register_sched_domain_sysctl();
6521 * Detach sched domains from a group of cpus specified in cpu_map
6522 * These cpus will now be attached to the NULL domain
6524 static void detach_destroy_domains(const struct cpumask *cpu_map)
6529 for_each_cpu(i, cpu_map)
6530 cpu_attach_domain(NULL, &def_root_domain, i);
6534 /* handle null as "default" */
6535 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6536 struct sched_domain_attr *new, int idx_new)
6538 struct sched_domain_attr tmp;
6545 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6546 new ? (new + idx_new) : &tmp,
6547 sizeof(struct sched_domain_attr));
6551 * Partition sched domains as specified by the 'ndoms_new'
6552 * cpumasks in the array doms_new[] of cpumasks. This compares
6553 * doms_new[] to the current sched domain partitioning, doms_cur[].
6554 * It destroys each deleted domain and builds each new domain.
6556 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6557 * The masks don't intersect (don't overlap.) We should setup one
6558 * sched domain for each mask. CPUs not in any of the cpumasks will
6559 * not be load balanced. If the same cpumask appears both in the
6560 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6563 * The passed in 'doms_new' should be allocated using
6564 * alloc_sched_domains. This routine takes ownership of it and will
6565 * free_sched_domains it when done with it. If the caller failed the
6566 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6567 * and partition_sched_domains() will fallback to the single partition
6568 * 'fallback_doms', it also forces the domains to be rebuilt.
6570 * If doms_new == NULL it will be replaced with cpu_online_mask.
6571 * ndoms_new == 0 is a special case for destroying existing domains,
6572 * and it will not create the default domain.
6574 * Call with hotplug lock held
6576 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6577 struct sched_domain_attr *dattr_new)
6582 mutex_lock(&sched_domains_mutex);
6584 /* always unregister in case we don't destroy any domains */
6585 unregister_sched_domain_sysctl();
6587 /* Let architecture update cpu core mappings. */
6588 new_topology = arch_update_cpu_topology();
6590 n = doms_new ? ndoms_new : 0;
6592 /* Destroy deleted domains */
6593 for (i = 0; i < ndoms_cur; i++) {
6594 for (j = 0; j < n && !new_topology; j++) {
6595 if (cpumask_equal(doms_cur[i], doms_new[j])
6596 && dattrs_equal(dattr_cur, i, dattr_new, j))
6599 /* no match - a current sched domain not in new doms_new[] */
6600 detach_destroy_domains(doms_cur[i]);
6605 if (doms_new == NULL) {
6607 doms_new = &fallback_doms;
6608 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6609 WARN_ON_ONCE(dattr_new);
6612 /* Build new domains */
6613 for (i = 0; i < ndoms_new; i++) {
6614 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6615 if (cpumask_equal(doms_new[i], doms_cur[j])
6616 && dattrs_equal(dattr_new, i, dattr_cur, j))
6619 /* no match - add a new doms_new */
6620 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6625 /* Remember the new sched domains */
6626 if (doms_cur != &fallback_doms)
6627 free_sched_domains(doms_cur, ndoms_cur);
6628 kfree(dattr_cur); /* kfree(NULL) is safe */
6629 doms_cur = doms_new;
6630 dattr_cur = dattr_new;
6631 ndoms_cur = ndoms_new;
6633 register_sched_domain_sysctl();
6635 mutex_unlock(&sched_domains_mutex);
6638 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6639 static void reinit_sched_domains(void)
6643 /* Destroy domains first to force the rebuild */
6644 partition_sched_domains(0, NULL, NULL);
6646 rebuild_sched_domains();
6650 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6652 unsigned int level = 0;
6654 if (sscanf(buf, "%u", &level) != 1)
6658 * level is always be positive so don't check for
6659 * level < POWERSAVINGS_BALANCE_NONE which is 0
6660 * What happens on 0 or 1 byte write,
6661 * need to check for count as well?
6664 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6668 sched_smt_power_savings = level;
6670 sched_mc_power_savings = level;
6672 reinit_sched_domains();
6677 #ifdef CONFIG_SCHED_MC
6678 static ssize_t sched_mc_power_savings_show(struct device *dev,
6679 struct device_attribute *attr,
6682 return sprintf(buf, "%u\n", sched_mc_power_savings);
6684 static ssize_t sched_mc_power_savings_store(struct device *dev,
6685 struct device_attribute *attr,
6686 const char *buf, size_t count)
6688 return sched_power_savings_store(buf, count, 0);
6690 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6691 sched_mc_power_savings_show,
6692 sched_mc_power_savings_store);
6695 #ifdef CONFIG_SCHED_SMT
6696 static ssize_t sched_smt_power_savings_show(struct device *dev,
6697 struct device_attribute *attr,
6700 return sprintf(buf, "%u\n", sched_smt_power_savings);
6702 static ssize_t sched_smt_power_savings_store(struct device *dev,
6703 struct device_attribute *attr,
6704 const char *buf, size_t count)
6706 return sched_power_savings_store(buf, count, 1);
6708 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6709 sched_smt_power_savings_show,
6710 sched_smt_power_savings_store);
6713 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6717 #ifdef CONFIG_SCHED_SMT
6719 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6721 #ifdef CONFIG_SCHED_MC
6722 if (!err && mc_capable())
6723 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6727 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6730 * Update cpusets according to cpu_active mask. If cpusets are
6731 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6732 * around partition_sched_domains().
6734 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6737 switch (action & ~CPU_TASKS_FROZEN) {
6739 case CPU_DOWN_FAILED:
6740 cpuset_update_active_cpus();
6747 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6750 switch (action & ~CPU_TASKS_FROZEN) {
6751 case CPU_DOWN_PREPARE:
6752 cpuset_update_active_cpus();
6759 void __init sched_init_smp(void)
6761 cpumask_var_t non_isolated_cpus;
6763 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6764 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6767 mutex_lock(&sched_domains_mutex);
6768 init_sched_domains(cpu_active_mask);
6769 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6770 if (cpumask_empty(non_isolated_cpus))
6771 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6772 mutex_unlock(&sched_domains_mutex);
6775 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6776 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6778 /* RT runtime code needs to handle some hotplug events */
6779 hotcpu_notifier(update_runtime, 0);
6783 /* Move init over to a non-isolated CPU */
6784 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6786 sched_init_granularity();
6787 free_cpumask_var(non_isolated_cpus);
6789 init_sched_rt_class();
6792 void __init sched_init_smp(void)
6794 sched_init_granularity();
6796 #endif /* CONFIG_SMP */
6798 const_debug unsigned int sysctl_timer_migration = 1;
6800 int in_sched_functions(unsigned long addr)
6802 return in_lock_functions(addr) ||
6803 (addr >= (unsigned long)__sched_text_start
6804 && addr < (unsigned long)__sched_text_end);
6807 #ifdef CONFIG_CGROUP_SCHED
6808 struct task_group root_task_group;
6811 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6813 void __init sched_init(void)
6816 unsigned long alloc_size = 0, ptr;
6818 #ifdef CONFIG_FAIR_GROUP_SCHED
6819 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6821 #ifdef CONFIG_RT_GROUP_SCHED
6822 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6824 #ifdef CONFIG_CPUMASK_OFFSTACK
6825 alloc_size += num_possible_cpus() * cpumask_size();
6828 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6830 #ifdef CONFIG_FAIR_GROUP_SCHED
6831 root_task_group.se = (struct sched_entity **)ptr;
6832 ptr += nr_cpu_ids * sizeof(void **);
6834 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6835 ptr += nr_cpu_ids * sizeof(void **);
6837 #endif /* CONFIG_FAIR_GROUP_SCHED */
6838 #ifdef CONFIG_RT_GROUP_SCHED
6839 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6840 ptr += nr_cpu_ids * sizeof(void **);
6842 root_task_group.rt_rq = (struct rt_rq **)ptr;
6843 ptr += nr_cpu_ids * sizeof(void **);
6845 #endif /* CONFIG_RT_GROUP_SCHED */
6846 #ifdef CONFIG_CPUMASK_OFFSTACK
6847 for_each_possible_cpu(i) {
6848 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6849 ptr += cpumask_size();
6851 #endif /* CONFIG_CPUMASK_OFFSTACK */
6855 init_defrootdomain();
6858 init_rt_bandwidth(&def_rt_bandwidth,
6859 global_rt_period(), global_rt_runtime());
6861 #ifdef CONFIG_RT_GROUP_SCHED
6862 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6863 global_rt_period(), global_rt_runtime());
6864 #endif /* CONFIG_RT_GROUP_SCHED */
6866 #ifdef CONFIG_CGROUP_SCHED
6867 list_add(&root_task_group.list, &task_groups);
6868 INIT_LIST_HEAD(&root_task_group.children);
6869 INIT_LIST_HEAD(&root_task_group.siblings);
6870 autogroup_init(&init_task);
6872 #endif /* CONFIG_CGROUP_SCHED */
6874 #ifdef CONFIG_CGROUP_CPUACCT
6875 root_cpuacct.cpustat = &kernel_cpustat;
6876 root_cpuacct.cpuusage = alloc_percpu(u64);
6877 /* Too early, not expected to fail */
6878 BUG_ON(!root_cpuacct.cpuusage);
6880 for_each_possible_cpu(i) {
6884 raw_spin_lock_init(&rq->lock);
6886 rq->calc_load_active = 0;
6887 rq->calc_load_update = jiffies + LOAD_FREQ;
6888 init_cfs_rq(&rq->cfs);
6889 init_rt_rq(&rq->rt, rq);
6890 #ifdef CONFIG_FAIR_GROUP_SCHED
6891 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6892 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6894 * How much cpu bandwidth does root_task_group get?
6896 * In case of task-groups formed thr' the cgroup filesystem, it
6897 * gets 100% of the cpu resources in the system. This overall
6898 * system cpu resource is divided among the tasks of
6899 * root_task_group and its child task-groups in a fair manner,
6900 * based on each entity's (task or task-group's) weight
6901 * (se->load.weight).
6903 * In other words, if root_task_group has 10 tasks of weight
6904 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6905 * then A0's share of the cpu resource is:
6907 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6909 * We achieve this by letting root_task_group's tasks sit
6910 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6912 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6913 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6914 #endif /* CONFIG_FAIR_GROUP_SCHED */
6916 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6917 #ifdef CONFIG_RT_GROUP_SCHED
6918 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6919 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6922 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6923 rq->cpu_load[j] = 0;
6925 rq->last_load_update_tick = jiffies;
6930 rq->cpu_power = SCHED_POWER_SCALE;
6931 rq->post_schedule = 0;
6932 rq->active_balance = 0;
6933 rq->next_balance = jiffies;
6938 rq->avg_idle = 2*sysctl_sched_migration_cost;
6939 rq_attach_root(rq, &def_root_domain);
6945 atomic_set(&rq->nr_iowait, 0);
6948 set_load_weight(&init_task);
6950 #ifdef CONFIG_PREEMPT_NOTIFIERS
6951 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6954 #ifdef CONFIG_RT_MUTEXES
6955 plist_head_init(&init_task.pi_waiters);
6959 * The boot idle thread does lazy MMU switching as well:
6961 atomic_inc(&init_mm.mm_count);
6962 enter_lazy_tlb(&init_mm, current);
6965 * Make us the idle thread. Technically, schedule() should not be
6966 * called from this thread, however somewhere below it might be,
6967 * but because we are the idle thread, we just pick up running again
6968 * when this runqueue becomes "idle".
6970 init_idle(current, smp_processor_id());
6972 calc_load_update = jiffies + LOAD_FREQ;
6975 * During early bootup we pretend to be a normal task:
6977 current->sched_class = &fair_sched_class;
6980 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6981 /* May be allocated at isolcpus cmdline parse time */
6982 if (cpu_isolated_map == NULL)
6983 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6985 init_sched_fair_class();
6987 scheduler_running = 1;
6990 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6991 static inline int preempt_count_equals(int preempt_offset)
6993 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6995 return (nested == preempt_offset);
6998 void __might_sleep(const char *file, int line, int preempt_offset)
7000 static unsigned long prev_jiffy; /* ratelimiting */
7002 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7003 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7004 system_state != SYSTEM_RUNNING || oops_in_progress)
7006 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7008 prev_jiffy = jiffies;
7011 "BUG: sleeping function called from invalid context at %s:%d\n",
7014 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7015 in_atomic(), irqs_disabled(),
7016 current->pid, current->comm);
7018 debug_show_held_locks(current);
7019 if (irqs_disabled())
7020 print_irqtrace_events(current);
7023 EXPORT_SYMBOL(__might_sleep);
7026 #ifdef CONFIG_MAGIC_SYSRQ
7027 static void normalize_task(struct rq *rq, struct task_struct *p)
7029 const struct sched_class *prev_class = p->sched_class;
7030 int old_prio = p->prio;
7035 deactivate_task(rq, p, 0);
7036 __setscheduler(rq, p, SCHED_NORMAL, 0);
7038 activate_task(rq, p, 0);
7039 resched_task(rq->curr);
7042 check_class_changed(rq, p, prev_class, old_prio);
7045 void normalize_rt_tasks(void)
7047 struct task_struct *g, *p;
7048 unsigned long flags;
7051 read_lock_irqsave(&tasklist_lock, flags);
7052 do_each_thread(g, p) {
7054 * Only normalize user tasks:
7059 p->se.exec_start = 0;
7060 #ifdef CONFIG_SCHEDSTATS
7061 p->se.statistics.wait_start = 0;
7062 p->se.statistics.sleep_start = 0;
7063 p->se.statistics.block_start = 0;
7068 * Renice negative nice level userspace
7071 if (TASK_NICE(p) < 0 && p->mm)
7072 set_user_nice(p, 0);
7076 raw_spin_lock(&p->pi_lock);
7077 rq = __task_rq_lock(p);
7079 normalize_task(rq, p);
7081 __task_rq_unlock(rq);
7082 raw_spin_unlock(&p->pi_lock);
7083 } while_each_thread(g, p);
7085 read_unlock_irqrestore(&tasklist_lock, flags);
7088 #endif /* CONFIG_MAGIC_SYSRQ */
7090 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7092 * These functions are only useful for the IA64 MCA handling, or kdb.
7094 * They can only be called when the whole system has been
7095 * stopped - every CPU needs to be quiescent, and no scheduling
7096 * activity can take place. Using them for anything else would
7097 * be a serious bug, and as a result, they aren't even visible
7098 * under any other configuration.
7102 * curr_task - return the current task for a given cpu.
7103 * @cpu: the processor in question.
7105 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7107 struct task_struct *curr_task(int cpu)
7109 return cpu_curr(cpu);
7112 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7116 * set_curr_task - set the current task for a given cpu.
7117 * @cpu: the processor in question.
7118 * @p: the task pointer to set.
7120 * Description: This function must only be used when non-maskable interrupts
7121 * are serviced on a separate stack. It allows the architecture to switch the
7122 * notion of the current task on a cpu in a non-blocking manner. This function
7123 * must be called with all CPU's synchronized, and interrupts disabled, the
7124 * and caller must save the original value of the current task (see
7125 * curr_task() above) and restore that value before reenabling interrupts and
7126 * re-starting the system.
7128 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7130 void set_curr_task(int cpu, struct task_struct *p)
7137 #ifdef CONFIG_CGROUP_SCHED
7138 /* task_group_lock serializes the addition/removal of task groups */
7139 static DEFINE_SPINLOCK(task_group_lock);
7141 static void free_sched_group(struct task_group *tg)
7143 free_fair_sched_group(tg);
7144 free_rt_sched_group(tg);
7149 /* allocate runqueue etc for a new task group */
7150 struct task_group *sched_create_group(struct task_group *parent)
7152 struct task_group *tg;
7153 unsigned long flags;
7155 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7157 return ERR_PTR(-ENOMEM);
7159 if (!alloc_fair_sched_group(tg, parent))
7162 if (!alloc_rt_sched_group(tg, parent))
7165 spin_lock_irqsave(&task_group_lock, flags);
7166 list_add_rcu(&tg->list, &task_groups);
7168 WARN_ON(!parent); /* root should already exist */
7170 tg->parent = parent;
7171 INIT_LIST_HEAD(&tg->children);
7172 list_add_rcu(&tg->siblings, &parent->children);
7173 spin_unlock_irqrestore(&task_group_lock, flags);
7178 free_sched_group(tg);
7179 return ERR_PTR(-ENOMEM);
7182 /* rcu callback to free various structures associated with a task group */
7183 static void free_sched_group_rcu(struct rcu_head *rhp)
7185 /* now it should be safe to free those cfs_rqs */
7186 free_sched_group(container_of(rhp, struct task_group, rcu));
7189 /* Destroy runqueue etc associated with a task group */
7190 void sched_destroy_group(struct task_group *tg)
7192 unsigned long flags;
7195 /* end participation in shares distribution */
7196 for_each_possible_cpu(i)
7197 unregister_fair_sched_group(tg, i);
7199 spin_lock_irqsave(&task_group_lock, flags);
7200 list_del_rcu(&tg->list);
7201 list_del_rcu(&tg->siblings);
7202 spin_unlock_irqrestore(&task_group_lock, flags);
7204 /* wait for possible concurrent references to cfs_rqs complete */
7205 call_rcu(&tg->rcu, free_sched_group_rcu);
7208 /* change task's runqueue when it moves between groups.
7209 * The caller of this function should have put the task in its new group
7210 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7211 * reflect its new group.
7213 void sched_move_task(struct task_struct *tsk)
7216 unsigned long flags;
7219 rq = task_rq_lock(tsk, &flags);
7221 running = task_current(rq, tsk);
7225 dequeue_task(rq, tsk, 0);
7226 if (unlikely(running))
7227 tsk->sched_class->put_prev_task(rq, tsk);
7229 #ifdef CONFIG_FAIR_GROUP_SCHED
7230 if (tsk->sched_class->task_move_group)
7231 tsk->sched_class->task_move_group(tsk, on_rq);
7234 set_task_rq(tsk, task_cpu(tsk));
7236 if (unlikely(running))
7237 tsk->sched_class->set_curr_task(rq);
7239 enqueue_task(rq, tsk, 0);
7241 task_rq_unlock(rq, tsk, &flags);
7243 #endif /* CONFIG_CGROUP_SCHED */
7245 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7246 static unsigned long to_ratio(u64 period, u64 runtime)
7248 if (runtime == RUNTIME_INF)
7251 return div64_u64(runtime << 20, period);
7255 #ifdef CONFIG_RT_GROUP_SCHED
7257 * Ensure that the real time constraints are schedulable.
7259 static DEFINE_MUTEX(rt_constraints_mutex);
7261 /* Must be called with tasklist_lock held */
7262 static inline int tg_has_rt_tasks(struct task_group *tg)
7264 struct task_struct *g, *p;
7266 do_each_thread(g, p) {
7267 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7269 } while_each_thread(g, p);
7274 struct rt_schedulable_data {
7275 struct task_group *tg;
7280 static int tg_rt_schedulable(struct task_group *tg, void *data)
7282 struct rt_schedulable_data *d = data;
7283 struct task_group *child;
7284 unsigned long total, sum = 0;
7285 u64 period, runtime;
7287 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7288 runtime = tg->rt_bandwidth.rt_runtime;
7291 period = d->rt_period;
7292 runtime = d->rt_runtime;
7296 * Cannot have more runtime than the period.
7298 if (runtime > period && runtime != RUNTIME_INF)
7302 * Ensure we don't starve existing RT tasks.
7304 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7307 total = to_ratio(period, runtime);
7310 * Nobody can have more than the global setting allows.
7312 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7316 * The sum of our children's runtime should not exceed our own.
7318 list_for_each_entry_rcu(child, &tg->children, siblings) {
7319 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7320 runtime = child->rt_bandwidth.rt_runtime;
7322 if (child == d->tg) {
7323 period = d->rt_period;
7324 runtime = d->rt_runtime;
7327 sum += to_ratio(period, runtime);
7336 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7340 struct rt_schedulable_data data = {
7342 .rt_period = period,
7343 .rt_runtime = runtime,
7347 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7353 static int tg_set_rt_bandwidth(struct task_group *tg,
7354 u64 rt_period, u64 rt_runtime)
7358 mutex_lock(&rt_constraints_mutex);
7359 read_lock(&tasklist_lock);
7360 err = __rt_schedulable(tg, rt_period, rt_runtime);
7364 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7365 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7366 tg->rt_bandwidth.rt_runtime = rt_runtime;
7368 for_each_possible_cpu(i) {
7369 struct rt_rq *rt_rq = tg->rt_rq[i];
7371 raw_spin_lock(&rt_rq->rt_runtime_lock);
7372 rt_rq->rt_runtime = rt_runtime;
7373 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7375 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7377 read_unlock(&tasklist_lock);
7378 mutex_unlock(&rt_constraints_mutex);
7383 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7385 u64 rt_runtime, rt_period;
7387 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7388 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7389 if (rt_runtime_us < 0)
7390 rt_runtime = RUNTIME_INF;
7392 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7395 long sched_group_rt_runtime(struct task_group *tg)
7399 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7402 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7403 do_div(rt_runtime_us, NSEC_PER_USEC);
7404 return rt_runtime_us;
7407 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7409 u64 rt_runtime, rt_period;
7411 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7412 rt_runtime = tg->rt_bandwidth.rt_runtime;
7417 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7420 long sched_group_rt_period(struct task_group *tg)
7424 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7425 do_div(rt_period_us, NSEC_PER_USEC);
7426 return rt_period_us;
7429 static int sched_rt_global_constraints(void)
7431 u64 runtime, period;
7434 if (sysctl_sched_rt_period <= 0)
7437 runtime = global_rt_runtime();
7438 period = global_rt_period();
7441 * Sanity check on the sysctl variables.
7443 if (runtime > period && runtime != RUNTIME_INF)
7446 mutex_lock(&rt_constraints_mutex);
7447 read_lock(&tasklist_lock);
7448 ret = __rt_schedulable(NULL, 0, 0);
7449 read_unlock(&tasklist_lock);
7450 mutex_unlock(&rt_constraints_mutex);
7455 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7457 /* Don't accept realtime tasks when there is no way for them to run */
7458 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7464 #else /* !CONFIG_RT_GROUP_SCHED */
7465 static int sched_rt_global_constraints(void)
7467 unsigned long flags;
7470 if (sysctl_sched_rt_period <= 0)
7474 * There's always some RT tasks in the root group
7475 * -- migration, kstopmachine etc..
7477 if (sysctl_sched_rt_runtime == 0)
7480 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7481 for_each_possible_cpu(i) {
7482 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7484 raw_spin_lock(&rt_rq->rt_runtime_lock);
7485 rt_rq->rt_runtime = global_rt_runtime();
7486 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7488 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7492 #endif /* CONFIG_RT_GROUP_SCHED */
7494 int sched_rt_handler(struct ctl_table *table, int write,
7495 void __user *buffer, size_t *lenp,
7499 int old_period, old_runtime;
7500 static DEFINE_MUTEX(mutex);
7503 old_period = sysctl_sched_rt_period;
7504 old_runtime = sysctl_sched_rt_runtime;
7506 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7508 if (!ret && write) {
7509 ret = sched_rt_global_constraints();
7511 sysctl_sched_rt_period = old_period;
7512 sysctl_sched_rt_runtime = old_runtime;
7514 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7515 def_rt_bandwidth.rt_period =
7516 ns_to_ktime(global_rt_period());
7519 mutex_unlock(&mutex);
7524 #ifdef CONFIG_CGROUP_SCHED
7526 /* return corresponding task_group object of a cgroup */
7527 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7529 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7530 struct task_group, css);
7533 static struct cgroup_subsys_state *
7534 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
7536 struct task_group *tg, *parent;
7538 if (!cgrp->parent) {
7539 /* This is early initialization for the top cgroup */
7540 return &root_task_group.css;
7543 parent = cgroup_tg(cgrp->parent);
7544 tg = sched_create_group(parent);
7546 return ERR_PTR(-ENOMEM);
7552 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7554 struct task_group *tg = cgroup_tg(cgrp);
7556 sched_destroy_group(tg);
7559 static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7560 struct cgroup_taskset *tset)
7562 struct task_struct *task;
7564 cgroup_taskset_for_each(task, cgrp, tset) {
7565 #ifdef CONFIG_RT_GROUP_SCHED
7566 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7569 /* We don't support RT-tasks being in separate groups */
7570 if (task->sched_class != &fair_sched_class)
7577 static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
7578 struct cgroup_taskset *tset)
7580 struct task_struct *task;
7582 cgroup_taskset_for_each(task, cgrp, tset)
7583 sched_move_task(task);
7587 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
7588 struct cgroup *old_cgrp, struct task_struct *task)
7591 * cgroup_exit() is called in the copy_process() failure path.
7592 * Ignore this case since the task hasn't ran yet, this avoids
7593 * trying to poke a half freed task state from generic code.
7595 if (!(task->flags & PF_EXITING))
7598 sched_move_task(task);
7601 #ifdef CONFIG_FAIR_GROUP_SCHED
7602 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7605 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7608 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7610 struct task_group *tg = cgroup_tg(cgrp);
7612 return (u64) scale_load_down(tg->shares);
7615 #ifdef CONFIG_CFS_BANDWIDTH
7616 static DEFINE_MUTEX(cfs_constraints_mutex);
7618 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7619 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7621 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7623 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7625 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7626 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7628 if (tg == &root_task_group)
7632 * Ensure we have at some amount of bandwidth every period. This is
7633 * to prevent reaching a state of large arrears when throttled via
7634 * entity_tick() resulting in prolonged exit starvation.
7636 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7640 * Likewise, bound things on the otherside by preventing insane quota
7641 * periods. This also allows us to normalize in computing quota
7644 if (period > max_cfs_quota_period)
7647 mutex_lock(&cfs_constraints_mutex);
7648 ret = __cfs_schedulable(tg, period, quota);
7652 runtime_enabled = quota != RUNTIME_INF;
7653 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7654 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7655 raw_spin_lock_irq(&cfs_b->lock);
7656 cfs_b->period = ns_to_ktime(period);
7657 cfs_b->quota = quota;
7659 __refill_cfs_bandwidth_runtime(cfs_b);
7660 /* restart the period timer (if active) to handle new period expiry */
7661 if (runtime_enabled && cfs_b->timer_active) {
7662 /* force a reprogram */
7663 cfs_b->timer_active = 0;
7664 __start_cfs_bandwidth(cfs_b);
7666 raw_spin_unlock_irq(&cfs_b->lock);
7668 for_each_possible_cpu(i) {
7669 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7670 struct rq *rq = cfs_rq->rq;
7672 raw_spin_lock_irq(&rq->lock);
7673 cfs_rq->runtime_enabled = runtime_enabled;
7674 cfs_rq->runtime_remaining = 0;
7676 if (cfs_rq->throttled)
7677 unthrottle_cfs_rq(cfs_rq);
7678 raw_spin_unlock_irq(&rq->lock);
7681 mutex_unlock(&cfs_constraints_mutex);
7686 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7690 period = ktime_to_ns(tg->cfs_bandwidth.period);
7691 if (cfs_quota_us < 0)
7692 quota = RUNTIME_INF;
7694 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7696 return tg_set_cfs_bandwidth(tg, period, quota);
7699 long tg_get_cfs_quota(struct task_group *tg)
7703 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7706 quota_us = tg->cfs_bandwidth.quota;
7707 do_div(quota_us, NSEC_PER_USEC);
7712 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7716 period = (u64)cfs_period_us * NSEC_PER_USEC;
7717 quota = tg->cfs_bandwidth.quota;
7719 return tg_set_cfs_bandwidth(tg, period, quota);
7722 long tg_get_cfs_period(struct task_group *tg)
7726 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7727 do_div(cfs_period_us, NSEC_PER_USEC);
7729 return cfs_period_us;
7732 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7734 return tg_get_cfs_quota(cgroup_tg(cgrp));
7737 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7740 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7743 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7745 return tg_get_cfs_period(cgroup_tg(cgrp));
7748 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7751 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7754 struct cfs_schedulable_data {
7755 struct task_group *tg;
7760 * normalize group quota/period to be quota/max_period
7761 * note: units are usecs
7763 static u64 normalize_cfs_quota(struct task_group *tg,
7764 struct cfs_schedulable_data *d)
7772 period = tg_get_cfs_period(tg);
7773 quota = tg_get_cfs_quota(tg);
7776 /* note: these should typically be equivalent */
7777 if (quota == RUNTIME_INF || quota == -1)
7780 return to_ratio(period, quota);
7783 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7785 struct cfs_schedulable_data *d = data;
7786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7787 s64 quota = 0, parent_quota = -1;
7790 quota = RUNTIME_INF;
7792 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7794 quota = normalize_cfs_quota(tg, d);
7795 parent_quota = parent_b->hierarchal_quota;
7798 * ensure max(child_quota) <= parent_quota, inherit when no
7801 if (quota == RUNTIME_INF)
7802 quota = parent_quota;
7803 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7806 cfs_b->hierarchal_quota = quota;
7811 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7814 struct cfs_schedulable_data data = {
7820 if (quota != RUNTIME_INF) {
7821 do_div(data.period, NSEC_PER_USEC);
7822 do_div(data.quota, NSEC_PER_USEC);
7826 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7832 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7833 struct cgroup_map_cb *cb)
7835 struct task_group *tg = cgroup_tg(cgrp);
7836 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7838 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7839 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7840 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7844 #endif /* CONFIG_CFS_BANDWIDTH */
7845 #endif /* CONFIG_FAIR_GROUP_SCHED */
7847 #ifdef CONFIG_RT_GROUP_SCHED
7848 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7851 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7854 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7856 return sched_group_rt_runtime(cgroup_tg(cgrp));
7859 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7862 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7865 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7867 return sched_group_rt_period(cgroup_tg(cgrp));
7869 #endif /* CONFIG_RT_GROUP_SCHED */
7871 static struct cftype cpu_files[] = {
7872 #ifdef CONFIG_FAIR_GROUP_SCHED
7875 .read_u64 = cpu_shares_read_u64,
7876 .write_u64 = cpu_shares_write_u64,
7879 #ifdef CONFIG_CFS_BANDWIDTH
7881 .name = "cfs_quota_us",
7882 .read_s64 = cpu_cfs_quota_read_s64,
7883 .write_s64 = cpu_cfs_quota_write_s64,
7886 .name = "cfs_period_us",
7887 .read_u64 = cpu_cfs_period_read_u64,
7888 .write_u64 = cpu_cfs_period_write_u64,
7892 .read_map = cpu_stats_show,
7895 #ifdef CONFIG_RT_GROUP_SCHED
7897 .name = "rt_runtime_us",
7898 .read_s64 = cpu_rt_runtime_read,
7899 .write_s64 = cpu_rt_runtime_write,
7902 .name = "rt_period_us",
7903 .read_u64 = cpu_rt_period_read_uint,
7904 .write_u64 = cpu_rt_period_write_uint,
7909 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
7911 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
7914 struct cgroup_subsys cpu_cgroup_subsys = {
7916 .create = cpu_cgroup_create,
7917 .destroy = cpu_cgroup_destroy,
7918 .can_attach = cpu_cgroup_can_attach,
7919 .attach = cpu_cgroup_attach,
7920 .exit = cpu_cgroup_exit,
7921 .populate = cpu_cgroup_populate,
7922 .subsys_id = cpu_cgroup_subsys_id,
7926 #endif /* CONFIG_CGROUP_SCHED */
7928 #ifdef CONFIG_CGROUP_CPUACCT
7931 * CPU accounting code for task groups.
7933 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7934 * (balbir@in.ibm.com).
7937 /* create a new cpu accounting group */
7938 static struct cgroup_subsys_state *cpuacct_create(
7939 struct cgroup_subsys *ss, struct cgroup *cgrp)
7944 return &root_cpuacct.css;
7946 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7950 ca->cpuusage = alloc_percpu(u64);
7954 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7956 goto out_free_cpuusage;
7961 free_percpu(ca->cpuusage);
7965 return ERR_PTR(-ENOMEM);
7968 /* destroy an existing cpu accounting group */
7970 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
7972 struct cpuacct *ca = cgroup_ca(cgrp);
7974 free_percpu(ca->cpustat);
7975 free_percpu(ca->cpuusage);
7979 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7981 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7984 #ifndef CONFIG_64BIT
7986 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7988 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7990 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7998 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8000 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8002 #ifndef CONFIG_64BIT
8004 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8006 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8008 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8014 /* return total cpu usage (in nanoseconds) of a group */
8015 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8017 struct cpuacct *ca = cgroup_ca(cgrp);
8018 u64 totalcpuusage = 0;
8021 for_each_present_cpu(i)
8022 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8024 return totalcpuusage;
8027 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8030 struct cpuacct *ca = cgroup_ca(cgrp);
8039 for_each_present_cpu(i)
8040 cpuacct_cpuusage_write(ca, i, 0);
8046 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8049 struct cpuacct *ca = cgroup_ca(cgroup);
8053 for_each_present_cpu(i) {
8054 percpu = cpuacct_cpuusage_read(ca, i);
8055 seq_printf(m, "%llu ", (unsigned long long) percpu);
8057 seq_printf(m, "\n");
8061 static const char *cpuacct_stat_desc[] = {
8062 [CPUACCT_STAT_USER] = "user",
8063 [CPUACCT_STAT_SYSTEM] = "system",
8066 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8067 struct cgroup_map_cb *cb)
8069 struct cpuacct *ca = cgroup_ca(cgrp);
8073 for_each_online_cpu(cpu) {
8074 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8075 val += kcpustat->cpustat[CPUTIME_USER];
8076 val += kcpustat->cpustat[CPUTIME_NICE];
8078 val = cputime64_to_clock_t(val);
8079 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8082 for_each_online_cpu(cpu) {
8083 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8084 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8085 val += kcpustat->cpustat[CPUTIME_IRQ];
8086 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8089 val = cputime64_to_clock_t(val);
8090 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8095 static struct cftype files[] = {
8098 .read_u64 = cpuusage_read,
8099 .write_u64 = cpuusage_write,
8102 .name = "usage_percpu",
8103 .read_seq_string = cpuacct_percpu_seq_read,
8107 .read_map = cpuacct_stats_show,
8111 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8113 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8117 * charge this task's execution time to its accounting group.
8119 * called with rq->lock held.
8121 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8126 if (unlikely(!cpuacct_subsys.active))
8129 cpu = task_cpu(tsk);
8135 for (; ca; ca = parent_ca(ca)) {
8136 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8137 *cpuusage += cputime;
8143 struct cgroup_subsys cpuacct_subsys = {
8145 .create = cpuacct_create,
8146 .destroy = cpuacct_destroy,
8147 .populate = cpuacct_populate,
8148 .subsys_id = cpuacct_subsys_id,
8150 #endif /* CONFIG_CGROUP_CPUACCT */