4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
87 #define CREATE_TRACE_POINTS
88 #include <trace/events/sched.h>
90 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
93 ktime_t soft, hard, now;
96 if (hrtimer_active(period_timer))
99 now = hrtimer_cb_get_time(period_timer);
100 hrtimer_forward(period_timer, now, period);
102 soft = hrtimer_get_softexpires(period_timer);
103 hard = hrtimer_get_expires(period_timer);
104 delta = ktime_to_ns(ktime_sub(hard, soft));
105 __hrtimer_start_range_ns(period_timer, soft, delta,
106 HRTIMER_MODE_ABS_PINNED, 0);
110 DEFINE_MUTEX(sched_domains_mutex);
111 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
113 static void update_rq_clock_task(struct rq *rq, s64 delta);
115 void update_rq_clock(struct rq *rq)
119 if (rq->skip_clock_update > 0)
122 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
124 update_rq_clock_task(rq, delta);
128 * Debugging: various feature bits
131 #define SCHED_FEAT(name, enabled) \
132 (1UL << __SCHED_FEAT_##name) * enabled |
134 const_debug unsigned int sysctl_sched_features =
135 #include "features.h"
140 #ifdef CONFIG_SCHED_DEBUG
141 #define SCHED_FEAT(name, enabled) \
144 static __read_mostly char *sched_feat_names[] = {
145 #include "features.h"
151 static int sched_feat_show(struct seq_file *m, void *v)
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
158 seq_printf(m, "%s ", sched_feat_names[i]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174 #include "features.h"
179 static void sched_feat_disable(int i)
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
185 static void sched_feat_enable(int i)
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
191 static void sched_feat_disable(int i) { };
192 static void sched_feat_enable(int i) { };
193 #endif /* HAVE_JUMP_LABEL */
196 sched_feat_write(struct file *filp, const char __user *ubuf,
197 size_t cnt, loff_t *ppos)
207 if (copy_from_user(&buf, ubuf, cnt))
213 if (strncmp(cmp, "NO_", 3) == 0) {
218 for (i = 0; i < __SCHED_FEAT_NR; i++) {
219 if (strcmp(cmp, sched_feat_names[i]) == 0) {
221 sysctl_sched_features &= ~(1UL << i);
222 sched_feat_disable(i);
224 sysctl_sched_features |= (1UL << i);
225 sched_feat_enable(i);
231 if (i == __SCHED_FEAT_NR)
239 static int sched_feat_open(struct inode *inode, struct file *filp)
241 return single_open(filp, sched_feat_show, NULL);
244 static const struct file_operations sched_feat_fops = {
245 .open = sched_feat_open,
246 .write = sched_feat_write,
249 .release = single_release,
252 static __init int sched_init_debug(void)
254 debugfs_create_file("sched_features", 0644, NULL, NULL,
259 late_initcall(sched_init_debug);
260 #endif /* CONFIG_SCHED_DEBUG */
263 * Number of tasks to iterate in a single balance run.
264 * Limited because this is done with IRQs disabled.
266 const_debug unsigned int sysctl_sched_nr_migrate = 32;
269 * period over which we average the RT time consumption, measured
274 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
277 * period over which we measure -rt task cpu usage in us.
280 unsigned int sysctl_sched_rt_period = 1000000;
282 __read_mostly int scheduler_running;
285 * part of the period that we allow rt tasks to run in us.
288 int sysctl_sched_rt_runtime = 950000;
293 * __task_rq_lock - lock the rq @p resides on.
295 static inline struct rq *__task_rq_lock(struct task_struct *p)
300 lockdep_assert_held(&p->pi_lock);
304 raw_spin_lock(&rq->lock);
305 if (likely(rq == task_rq(p)))
307 raw_spin_unlock(&rq->lock);
312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
314 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
315 __acquires(p->pi_lock)
321 raw_spin_lock_irqsave(&p->pi_lock, *flags);
323 raw_spin_lock(&rq->lock);
324 if (likely(rq == task_rq(p)))
326 raw_spin_unlock(&rq->lock);
327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
331 static void __task_rq_unlock(struct rq *rq)
334 raw_spin_unlock(&rq->lock);
338 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
340 __releases(p->pi_lock)
342 raw_spin_unlock(&rq->lock);
343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
347 * this_rq_lock - lock this runqueue and disable interrupts.
349 static struct rq *this_rq_lock(void)
356 raw_spin_lock(&rq->lock);
361 #ifdef CONFIG_SCHED_HRTICK
363 * Use HR-timers to deliver accurate preemption points.
365 * Its all a bit involved since we cannot program an hrt while holding the
366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
369 * When we get rescheduled we reprogram the hrtick_timer outside of the
373 static void hrtick_clear(struct rq *rq)
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 raw_spin_lock(&rq->lock);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
394 return HRTIMER_NORESTART;
399 * called from hardirq (IPI) context
401 static void __hrtick_start(void *arg)
405 raw_spin_lock(&rq->lock);
406 hrtimer_restart(&rq->hrtick_timer);
407 rq->hrtick_csd_pending = 0;
408 raw_spin_unlock(&rq->lock);
412 * Called to set the hrtick timer state.
414 * called with rq->lock held and irqs disabled
416 void hrtick_start(struct rq *rq, u64 delay)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
421 hrtimer_set_expires(timer, time);
423 if (rq == this_rq()) {
424 hrtimer_restart(timer);
425 } else if (!rq->hrtick_csd_pending) {
426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
427 rq->hrtick_csd_pending = 1;
432 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
434 int cpu = (int)(long)hcpu;
437 case CPU_UP_CANCELED:
438 case CPU_UP_CANCELED_FROZEN:
439 case CPU_DOWN_PREPARE:
440 case CPU_DOWN_PREPARE_FROZEN:
442 case CPU_DEAD_FROZEN:
443 hrtick_clear(cpu_rq(cpu));
450 static __init void init_hrtick(void)
452 hotcpu_notifier(hotplug_hrtick, 0);
456 * Called to set the hrtick timer state.
458 * called with rq->lock held and irqs disabled
460 void hrtick_start(struct rq *rq, u64 delay)
462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
463 HRTIMER_MODE_REL_PINNED, 0);
466 static inline void init_hrtick(void)
469 #endif /* CONFIG_SMP */
471 static void init_rq_hrtick(struct rq *rq)
474 rq->hrtick_csd_pending = 0;
476 rq->hrtick_csd.flags = 0;
477 rq->hrtick_csd.func = __hrtick_start;
478 rq->hrtick_csd.info = rq;
481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
482 rq->hrtick_timer.function = hrtick;
484 #else /* CONFIG_SCHED_HRTICK */
485 static inline void hrtick_clear(struct rq *rq)
489 static inline void init_rq_hrtick(struct rq *rq)
493 static inline void init_hrtick(void)
496 #endif /* CONFIG_SCHED_HRTICK */
499 * resched_task - mark a task 'to be rescheduled now'.
501 * On UP this means the setting of the need_resched flag, on SMP it
502 * might also involve a cross-CPU call to trigger the scheduler on
507 #ifndef tsk_is_polling
508 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
511 void resched_task(struct task_struct *p)
515 assert_raw_spin_locked(&task_rq(p)->lock);
517 if (test_tsk_need_resched(p))
520 set_tsk_need_resched(p);
523 if (cpu == smp_processor_id())
526 /* NEED_RESCHED must be visible before we test polling */
528 if (!tsk_is_polling(p))
529 smp_send_reschedule(cpu);
532 void resched_cpu(int cpu)
534 struct rq *rq = cpu_rq(cpu);
537 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
539 resched_task(cpu_curr(cpu));
540 raw_spin_unlock_irqrestore(&rq->lock, flags);
545 * In the semi idle case, use the nearest busy cpu for migrating timers
546 * from an idle cpu. This is good for power-savings.
548 * We don't do similar optimization for completely idle system, as
549 * selecting an idle cpu will add more delays to the timers than intended
550 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
552 int get_nohz_timer_target(void)
554 int cpu = smp_processor_id();
556 struct sched_domain *sd;
559 for_each_domain(cpu, sd) {
560 for_each_cpu(i, sched_domain_span(sd)) {
572 * When add_timer_on() enqueues a timer into the timer wheel of an
573 * idle CPU then this timer might expire before the next timer event
574 * which is scheduled to wake up that CPU. In case of a completely
575 * idle system the next event might even be infinite time into the
576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
577 * leaves the inner idle loop so the newly added timer is taken into
578 * account when the CPU goes back to idle and evaluates the timer
579 * wheel for the next timer event.
581 void wake_up_idle_cpu(int cpu)
583 struct rq *rq = cpu_rq(cpu);
585 if (cpu == smp_processor_id())
589 * This is safe, as this function is called with the timer
590 * wheel base lock of (cpu) held. When the CPU is on the way
591 * to idle and has not yet set rq->curr to idle then it will
592 * be serialized on the timer wheel base lock and take the new
593 * timer into account automatically.
595 if (rq->curr != rq->idle)
599 * We can set TIF_RESCHED on the idle task of the other CPU
600 * lockless. The worst case is that the other CPU runs the
601 * idle task through an additional NOOP schedule()
603 set_tsk_need_resched(rq->idle);
605 /* NEED_RESCHED must be visible before we test polling */
607 if (!tsk_is_polling(rq->idle))
608 smp_send_reschedule(cpu);
611 static inline bool got_nohz_idle_kick(void)
613 int cpu = smp_processor_id();
614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
617 #else /* CONFIG_NO_HZ */
619 static inline bool got_nohz_idle_kick(void)
624 #endif /* CONFIG_NO_HZ */
626 void sched_avg_update(struct rq *rq)
628 s64 period = sched_avg_period();
630 while ((s64)(rq->clock - rq->age_stamp) > period) {
632 * Inline assembly required to prevent the compiler
633 * optimising this loop into a divmod call.
634 * See __iter_div_u64_rem() for another example of this.
636 asm("" : "+rm" (rq->age_stamp));
637 rq->age_stamp += period;
642 #else /* !CONFIG_SMP */
643 void resched_task(struct task_struct *p)
645 assert_raw_spin_locked(&task_rq(p)->lock);
646 set_tsk_need_resched(p);
648 #endif /* CONFIG_SMP */
650 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
653 * Iterate task_group tree rooted at *from, calling @down when first entering a
654 * node and @up when leaving it for the final time.
656 * Caller must hold rcu_lock or sufficient equivalent.
658 int walk_tg_tree_from(struct task_group *from,
659 tg_visitor down, tg_visitor up, void *data)
661 struct task_group *parent, *child;
667 ret = (*down)(parent, data);
670 list_for_each_entry_rcu(child, &parent->children, siblings) {
677 ret = (*up)(parent, data);
678 if (ret || parent == from)
682 parent = parent->parent;
689 int tg_nop(struct task_group *tg, void *data)
695 void update_cpu_load(struct rq *this_rq);
697 static void set_load_weight(struct task_struct *p)
699 int prio = p->static_prio - MAX_RT_PRIO;
700 struct load_weight *load = &p->se.load;
703 * SCHED_IDLE tasks get minimal weight:
705 if (p->policy == SCHED_IDLE) {
706 load->weight = scale_load(WEIGHT_IDLEPRIO);
707 load->inv_weight = WMULT_IDLEPRIO;
711 load->weight = scale_load(prio_to_weight[prio]);
712 load->inv_weight = prio_to_wmult[prio];
715 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
718 sched_info_queued(p);
719 p->sched_class->enqueue_task(rq, p, flags);
722 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
725 sched_info_dequeued(p);
726 p->sched_class->dequeue_task(rq, p, flags);
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);
737 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
739 if (task_contributes_to_load(p))
740 rq->nr_uninterruptible++;
742 dequeue_task(rq, p, flags);
745 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
748 * There are no locks covering percpu hardirq/softirq time.
749 * They are only modified in account_system_vtime, on corresponding CPU
750 * with interrupts disabled. So, writes are safe.
751 * They are read and saved off onto struct rq in update_rq_clock().
752 * This may result in other CPU reading this CPU's irq time and can
753 * race with irq/account_system_vtime on this CPU. We would either get old
754 * or new value with a side effect of accounting a slice of irq time to wrong
755 * task when irq is in progress while we read rq->clock. That is a worthy
756 * compromise in place of having locks on each irq in account_system_time.
758 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
759 static DEFINE_PER_CPU(u64, cpu_softirq_time);
761 static DEFINE_PER_CPU(u64, irq_start_time);
762 static int sched_clock_irqtime;
764 void enable_sched_clock_irqtime(void)
766 sched_clock_irqtime = 1;
769 void disable_sched_clock_irqtime(void)
771 sched_clock_irqtime = 0;
775 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
777 static inline void irq_time_write_begin(void)
779 __this_cpu_inc(irq_time_seq.sequence);
783 static inline void irq_time_write_end(void)
786 __this_cpu_inc(irq_time_seq.sequence);
789 static inline u64 irq_time_read(int cpu)
795 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
796 irq_time = per_cpu(cpu_softirq_time, cpu) +
797 per_cpu(cpu_hardirq_time, cpu);
798 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
802 #else /* CONFIG_64BIT */
803 static inline void irq_time_write_begin(void)
807 static inline void irq_time_write_end(void)
811 static inline u64 irq_time_read(int cpu)
813 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
815 #endif /* CONFIG_64BIT */
818 * Called before incrementing preempt_count on {soft,}irq_enter
819 * and before decrementing preempt_count on {soft,}irq_exit.
821 void account_system_vtime(struct task_struct *curr)
827 if (!sched_clock_irqtime)
830 local_irq_save(flags);
832 cpu = smp_processor_id();
833 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
834 __this_cpu_add(irq_start_time, delta);
836 irq_time_write_begin();
838 * We do not account for softirq time from ksoftirqd here.
839 * We want to continue accounting softirq time to ksoftirqd thread
840 * in that case, so as not to confuse scheduler with a special task
841 * that do not consume any time, but still wants to run.
844 __this_cpu_add(cpu_hardirq_time, delta);
845 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
846 __this_cpu_add(cpu_softirq_time, delta);
848 irq_time_write_end();
849 local_irq_restore(flags);
851 EXPORT_SYMBOL_GPL(account_system_vtime);
853 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
855 #ifdef CONFIG_PARAVIRT
856 static inline u64 steal_ticks(u64 steal)
858 if (unlikely(steal > NSEC_PER_SEC))
859 return div_u64(steal, TICK_NSEC);
861 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
865 static void update_rq_clock_task(struct rq *rq, s64 delta)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal = 0, irq_delta = 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta > delta)
895 rq->prev_irq_time += irq_delta;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled))) {
902 steal = paravirt_steal_clock(cpu_of(rq));
903 steal -= rq->prev_steal_time_rq;
905 if (unlikely(steal > delta))
908 st = steal_ticks(steal);
909 steal = st * TICK_NSEC;
911 rq->prev_steal_time_rq += steal;
917 rq->clock_task += delta;
919 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
920 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
921 sched_rt_avg_update(rq, irq_delta + steal);
925 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
926 static int irqtime_account_hi_update(void)
928 u64 *cpustat = kcpustat_this_cpu->cpustat;
933 local_irq_save(flags);
934 latest_ns = this_cpu_read(cpu_hardirq_time);
935 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
937 local_irq_restore(flags);
941 static int irqtime_account_si_update(void)
943 u64 *cpustat = kcpustat_this_cpu->cpustat;
948 local_irq_save(flags);
949 latest_ns = this_cpu_read(cpu_softirq_time);
950 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
952 local_irq_restore(flags);
956 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
958 #define sched_clock_irqtime (0)
962 void sched_set_stop_task(int cpu, struct task_struct *stop)
964 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
965 struct task_struct *old_stop = cpu_rq(cpu)->stop;
969 * Make it appear like a SCHED_FIFO task, its something
970 * userspace knows about and won't get confused about.
972 * Also, it will make PI more or less work without too
973 * much confusion -- but then, stop work should not
974 * rely on PI working anyway.
976 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
978 stop->sched_class = &stop_sched_class;
981 cpu_rq(cpu)->stop = stop;
985 * Reset it back to a normal scheduling class so that
986 * it can die in pieces.
988 old_stop->sched_class = &rt_sched_class;
993 * __normal_prio - return the priority that is based on the static prio
995 static inline int __normal_prio(struct task_struct *p)
997 return p->static_prio;
1001 * Calculate the expected normal priority: i.e. priority
1002 * without taking RT-inheritance into account. Might be
1003 * boosted by interactivity modifiers. Changes upon fork,
1004 * setprio syscalls, and whenever the interactivity
1005 * estimator recalculates.
1007 static inline int normal_prio(struct task_struct *p)
1011 if (task_has_rt_policy(p))
1012 prio = MAX_RT_PRIO-1 - p->rt_priority;
1014 prio = __normal_prio(p);
1019 * Calculate the current priority, i.e. the priority
1020 * taken into account by the scheduler. This value might
1021 * be boosted by RT tasks, or might be boosted by
1022 * interactivity modifiers. Will be RT if the task got
1023 * RT-boosted. If not then it returns p->normal_prio.
1025 static int effective_prio(struct task_struct *p)
1027 p->normal_prio = normal_prio(p);
1029 * If we are RT tasks or we were boosted to RT priority,
1030 * keep the priority unchanged. Otherwise, update priority
1031 * to the normal priority:
1033 if (!rt_prio(p->prio))
1034 return p->normal_prio;
1039 * task_curr - is this task currently executing on a CPU?
1040 * @p: the task in question.
1042 inline int task_curr(const struct task_struct *p)
1044 return cpu_curr(task_cpu(p)) == p;
1047 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1048 const struct sched_class *prev_class,
1051 if (prev_class != p->sched_class) {
1052 if (prev_class->switched_from)
1053 prev_class->switched_from(rq, p);
1054 p->sched_class->switched_to(rq, p);
1055 } else if (oldprio != p->prio)
1056 p->sched_class->prio_changed(rq, p, oldprio);
1059 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1061 const struct sched_class *class;
1063 if (p->sched_class == rq->curr->sched_class) {
1064 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1066 for_each_class(class) {
1067 if (class == rq->curr->sched_class)
1069 if (class == p->sched_class) {
1070 resched_task(rq->curr);
1077 * A queue event has occurred, and we're going to schedule. In
1078 * this case, we can save a useless back to back clock update.
1080 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1081 rq->skip_clock_update = 1;
1085 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1087 #ifdef CONFIG_SCHED_DEBUG
1089 * We should never call set_task_cpu() on a blocked task,
1090 * ttwu() will sort out the placement.
1092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1095 #ifdef CONFIG_LOCKDEP
1097 * The caller should hold either p->pi_lock or rq->lock, when changing
1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1100 * sched_move_task() holds both and thus holding either pins the cgroup,
1101 * see set_task_rq().
1103 * Furthermore, all task_rq users should acquire both locks, see
1106 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1107 lockdep_is_held(&task_rq(p)->lock)));
1111 trace_sched_migrate_task(p, new_cpu);
1113 if (task_cpu(p) != new_cpu) {
1114 p->se.nr_migrations++;
1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1118 __set_task_cpu(p, new_cpu);
1121 struct migration_arg {
1122 struct task_struct *task;
1126 static int migration_cpu_stop(void *data);
1129 * wait_task_inactive - wait for a thread to unschedule.
1131 * If @match_state is nonzero, it's the @p->state value just checked and
1132 * not expected to change. If it changes, i.e. @p might have woken up,
1133 * then return zero. When we succeed in waiting for @p to be off its CPU,
1134 * we return a positive number (its total switch count). If a second call
1135 * a short while later returns the same number, the caller can be sure that
1136 * @p has remained unscheduled the whole time.
1138 * The caller must ensure that the task *will* unschedule sometime soon,
1139 * else this function might spin for a *long* time. This function can't
1140 * be called with interrupts off, or it may introduce deadlock with
1141 * smp_call_function() if an IPI is sent by the same process we are
1142 * waiting to become inactive.
1144 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1146 unsigned long flags;
1153 * We do the initial early heuristics without holding
1154 * any task-queue locks at all. We'll only try to get
1155 * the runqueue lock when things look like they will
1161 * If the task is actively running on another CPU
1162 * still, just relax and busy-wait without holding
1165 * NOTE! Since we don't hold any locks, it's not
1166 * even sure that "rq" stays as the right runqueue!
1167 * But we don't care, since "task_running()" will
1168 * return false if the runqueue has changed and p
1169 * is actually now running somewhere else!
1171 while (task_running(rq, p)) {
1172 if (match_state && unlikely(p->state != match_state))
1178 * Ok, time to look more closely! We need the rq
1179 * lock now, to be *sure*. If we're wrong, we'll
1180 * just go back and repeat.
1182 rq = task_rq_lock(p, &flags);
1183 trace_sched_wait_task(p);
1184 running = task_running(rq, p);
1187 if (!match_state || p->state == match_state)
1188 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1189 task_rq_unlock(rq, p, &flags);
1192 * If it changed from the expected state, bail out now.
1194 if (unlikely(!ncsw))
1198 * Was it really running after all now that we
1199 * checked with the proper locks actually held?
1201 * Oops. Go back and try again..
1203 if (unlikely(running)) {
1209 * It's not enough that it's not actively running,
1210 * it must be off the runqueue _entirely_, and not
1213 * So if it was still runnable (but just not actively
1214 * running right now), it's preempted, and we should
1215 * yield - it could be a while.
1217 if (unlikely(on_rq)) {
1218 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1220 set_current_state(TASK_UNINTERRUPTIBLE);
1221 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1226 * Ahh, all good. It wasn't running, and it wasn't
1227 * runnable, which means that it will never become
1228 * running in the future either. We're all done!
1237 * kick_process - kick a running thread to enter/exit the kernel
1238 * @p: the to-be-kicked thread
1240 * Cause a process which is running on another CPU to enter
1241 * kernel-mode, without any delay. (to get signals handled.)
1243 * NOTE: this function doesn't have to take the runqueue lock,
1244 * because all it wants to ensure is that the remote task enters
1245 * the kernel. If the IPI races and the task has been migrated
1246 * to another CPU then no harm is done and the purpose has been
1249 void kick_process(struct task_struct *p)
1255 if ((cpu != smp_processor_id()) && task_curr(p))
1256 smp_send_reschedule(cpu);
1259 EXPORT_SYMBOL_GPL(kick_process);
1260 #endif /* CONFIG_SMP */
1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1266 static int select_fallback_rq(int cpu, struct task_struct *p)
1268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1269 enum { cpuset, possible, fail } state = cpuset;
1272 /* Look for allowed, online CPU in same node. */
1273 for_each_cpu(dest_cpu, nodemask) {
1274 if (!cpu_online(dest_cpu))
1276 if (!cpu_active(dest_cpu))
1278 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1283 /* Any allowed, online CPU? */
1284 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1285 if (!cpu_online(dest_cpu))
1287 if (!cpu_active(dest_cpu))
1294 /* No more Mr. Nice Guy. */
1295 cpuset_cpus_allowed_fallback(p);
1300 do_set_cpus_allowed(p, cpu_possible_mask);
1311 if (state != cpuset) {
1313 * Don't tell them about moving exiting tasks or
1314 * kernel threads (both mm NULL), since they never
1317 if (p->mm && printk_ratelimit()) {
1318 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1319 task_pid_nr(p), p->comm, cpu);
1327 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1330 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1332 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1335 * In order not to call set_task_cpu() on a blocking task we need
1336 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1339 * Since this is common to all placement strategies, this lives here.
1341 * [ this allows ->select_task() to simply return task_cpu(p) and
1342 * not worry about this generic constraint ]
1344 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1346 cpu = select_fallback_rq(task_cpu(p), p);
1351 static void update_avg(u64 *avg, u64 sample)
1353 s64 diff = sample - *avg;
1359 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1361 #ifdef CONFIG_SCHEDSTATS
1362 struct rq *rq = this_rq();
1365 int this_cpu = smp_processor_id();
1367 if (cpu == this_cpu) {
1368 schedstat_inc(rq, ttwu_local);
1369 schedstat_inc(p, se.statistics.nr_wakeups_local);
1371 struct sched_domain *sd;
1373 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1375 for_each_domain(this_cpu, sd) {
1376 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1377 schedstat_inc(sd, ttwu_wake_remote);
1384 if (wake_flags & WF_MIGRATED)
1385 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1387 #endif /* CONFIG_SMP */
1389 schedstat_inc(rq, ttwu_count);
1390 schedstat_inc(p, se.statistics.nr_wakeups);
1392 if (wake_flags & WF_SYNC)
1393 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1395 #endif /* CONFIG_SCHEDSTATS */
1398 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1400 activate_task(rq, p, en_flags);
1403 /* if a worker is waking up, notify workqueue */
1404 if (p->flags & PF_WQ_WORKER)
1405 wq_worker_waking_up(p, cpu_of(rq));
1409 * Mark the task runnable and perform wakeup-preemption.
1412 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1414 trace_sched_wakeup(p, true);
1415 check_preempt_curr(rq, p, wake_flags);
1417 p->state = TASK_RUNNING;
1419 if (p->sched_class->task_woken)
1420 p->sched_class->task_woken(rq, p);
1422 if (rq->idle_stamp) {
1423 u64 delta = rq->clock - rq->idle_stamp;
1424 u64 max = 2*sysctl_sched_migration_cost;
1429 update_avg(&rq->avg_idle, delta);
1436 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1439 if (p->sched_contributes_to_load)
1440 rq->nr_uninterruptible--;
1443 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1444 ttwu_do_wakeup(rq, p, wake_flags);
1448 * Called in case the task @p isn't fully descheduled from its runqueue,
1449 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1450 * since all we need to do is flip p->state to TASK_RUNNING, since
1451 * the task is still ->on_rq.
1453 static int ttwu_remote(struct task_struct *p, int wake_flags)
1458 rq = __task_rq_lock(p);
1460 ttwu_do_wakeup(rq, p, wake_flags);
1463 __task_rq_unlock(rq);
1469 static void sched_ttwu_pending(void)
1471 struct rq *rq = this_rq();
1472 struct llist_node *llist = llist_del_all(&rq->wake_list);
1473 struct task_struct *p;
1475 raw_spin_lock(&rq->lock);
1478 p = llist_entry(llist, struct task_struct, wake_entry);
1479 llist = llist_next(llist);
1480 ttwu_do_activate(rq, p, 0);
1483 raw_spin_unlock(&rq->lock);
1486 void scheduler_ipi(void)
1488 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1492 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1493 * traditionally all their work was done from the interrupt return
1494 * path. Now that we actually do some work, we need to make sure
1497 * Some archs already do call them, luckily irq_enter/exit nest
1500 * Arguably we should visit all archs and update all handlers,
1501 * however a fair share of IPIs are still resched only so this would
1502 * somewhat pessimize the simple resched case.
1505 sched_ttwu_pending();
1508 * Check if someone kicked us for doing the nohz idle load balance.
1510 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1511 this_rq()->idle_balance = 1;
1512 raise_softirq_irqoff(SCHED_SOFTIRQ);
1517 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1519 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1520 smp_send_reschedule(cpu);
1523 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1524 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1529 rq = __task_rq_lock(p);
1531 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1532 ttwu_do_wakeup(rq, p, wake_flags);
1535 __task_rq_unlock(rq);
1540 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1542 bool cpus_share_cache(int this_cpu, int that_cpu)
1544 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1546 #endif /* CONFIG_SMP */
1548 static void ttwu_queue(struct task_struct *p, int cpu)
1550 struct rq *rq = cpu_rq(cpu);
1552 #if defined(CONFIG_SMP)
1553 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1554 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1555 ttwu_queue_remote(p, cpu);
1560 raw_spin_lock(&rq->lock);
1561 ttwu_do_activate(rq, p, 0);
1562 raw_spin_unlock(&rq->lock);
1566 * try_to_wake_up - wake up a thread
1567 * @p: the thread to be awakened
1568 * @state: the mask of task states that can be woken
1569 * @wake_flags: wake modifier flags (WF_*)
1571 * Put it on the run-queue if it's not already there. The "current"
1572 * thread is always on the run-queue (except when the actual
1573 * re-schedule is in progress), and as such you're allowed to do
1574 * the simpler "current->state = TASK_RUNNING" to mark yourself
1575 * runnable without the overhead of this.
1577 * Returns %true if @p was woken up, %false if it was already running
1578 * or @state didn't match @p's state.
1581 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1583 unsigned long flags;
1584 int cpu, success = 0;
1587 raw_spin_lock_irqsave(&p->pi_lock, flags);
1588 if (!(p->state & state))
1591 success = 1; /* we're going to change ->state */
1594 if (p->on_rq && ttwu_remote(p, wake_flags))
1599 * If the owning (remote) cpu is still in the middle of schedule() with
1600 * this task as prev, wait until its done referencing the task.
1603 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1605 * In case the architecture enables interrupts in
1606 * context_switch(), we cannot busy wait, since that
1607 * would lead to deadlocks when an interrupt hits and
1608 * tries to wake up @prev. So bail and do a complete
1611 if (ttwu_activate_remote(p, wake_flags))
1618 * Pairs with the smp_wmb() in finish_lock_switch().
1622 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1623 p->state = TASK_WAKING;
1625 if (p->sched_class->task_waking)
1626 p->sched_class->task_waking(p);
1628 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1629 if (task_cpu(p) != cpu) {
1630 wake_flags |= WF_MIGRATED;
1631 set_task_cpu(p, cpu);
1633 #endif /* CONFIG_SMP */
1637 ttwu_stat(p, cpu, wake_flags);
1639 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1645 * try_to_wake_up_local - try to wake up a local task with rq lock held
1646 * @p: the thread to be awakened
1648 * Put @p on the run-queue if it's not already there. The caller must
1649 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1652 static void try_to_wake_up_local(struct task_struct *p)
1654 struct rq *rq = task_rq(p);
1656 BUG_ON(rq != this_rq());
1657 BUG_ON(p == current);
1658 lockdep_assert_held(&rq->lock);
1660 if (!raw_spin_trylock(&p->pi_lock)) {
1661 raw_spin_unlock(&rq->lock);
1662 raw_spin_lock(&p->pi_lock);
1663 raw_spin_lock(&rq->lock);
1666 if (!(p->state & TASK_NORMAL))
1670 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1672 ttwu_do_wakeup(rq, p, 0);
1673 ttwu_stat(p, smp_processor_id(), 0);
1675 raw_spin_unlock(&p->pi_lock);
1679 * wake_up_process - Wake up a specific process
1680 * @p: The process to be woken up.
1682 * Attempt to wake up the nominated process and move it to the set of runnable
1683 * processes. Returns 1 if the process was woken up, 0 if it was already
1686 * It may be assumed that this function implies a write memory barrier before
1687 * changing the task state if and only if any tasks are woken up.
1689 int wake_up_process(struct task_struct *p)
1691 return try_to_wake_up(p, TASK_ALL, 0);
1693 EXPORT_SYMBOL(wake_up_process);
1695 int wake_up_state(struct task_struct *p, unsigned int state)
1697 return try_to_wake_up(p, state, 0);
1701 * Perform scheduler related setup for a newly forked process p.
1702 * p is forked by current.
1704 * __sched_fork() is basic setup used by init_idle() too:
1706 static void __sched_fork(struct task_struct *p)
1711 p->se.exec_start = 0;
1712 p->se.sum_exec_runtime = 0;
1713 p->se.prev_sum_exec_runtime = 0;
1714 p->se.nr_migrations = 0;
1716 INIT_LIST_HEAD(&p->se.group_node);
1718 #ifdef CONFIG_SCHEDSTATS
1719 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1722 INIT_LIST_HEAD(&p->rt.run_list);
1724 #ifdef CONFIG_PREEMPT_NOTIFIERS
1725 INIT_HLIST_HEAD(&p->preempt_notifiers);
1730 * fork()/clone()-time setup:
1732 void sched_fork(struct task_struct *p)
1734 unsigned long flags;
1735 int cpu = get_cpu();
1739 * We mark the process as running here. This guarantees that
1740 * nobody will actually run it, and a signal or other external
1741 * event cannot wake it up and insert it on the runqueue either.
1743 p->state = TASK_RUNNING;
1746 * Make sure we do not leak PI boosting priority to the child.
1748 p->prio = current->normal_prio;
1751 * Revert to default priority/policy on fork if requested.
1753 if (unlikely(p->sched_reset_on_fork)) {
1754 if (task_has_rt_policy(p)) {
1755 p->policy = SCHED_NORMAL;
1756 p->static_prio = NICE_TO_PRIO(0);
1758 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1759 p->static_prio = NICE_TO_PRIO(0);
1761 p->prio = p->normal_prio = __normal_prio(p);
1765 * We don't need the reset flag anymore after the fork. It has
1766 * fulfilled its duty:
1768 p->sched_reset_on_fork = 0;
1771 if (!rt_prio(p->prio))
1772 p->sched_class = &fair_sched_class;
1774 if (p->sched_class->task_fork)
1775 p->sched_class->task_fork(p);
1778 * The child is not yet in the pid-hash so no cgroup attach races,
1779 * and the cgroup is pinned to this child due to cgroup_fork()
1780 * is ran before sched_fork().
1782 * Silence PROVE_RCU.
1784 raw_spin_lock_irqsave(&p->pi_lock, flags);
1785 set_task_cpu(p, cpu);
1786 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1788 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1789 if (likely(sched_info_on()))
1790 memset(&p->sched_info, 0, sizeof(p->sched_info));
1792 #if defined(CONFIG_SMP)
1795 #ifdef CONFIG_PREEMPT_COUNT
1796 /* Want to start with kernel preemption disabled. */
1797 task_thread_info(p)->preempt_count = 1;
1800 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1807 * wake_up_new_task - wake up a newly created task for the first time.
1809 * This function will do some initial scheduler statistics housekeeping
1810 * that must be done for every newly created context, then puts the task
1811 * on the runqueue and wakes it.
1813 void wake_up_new_task(struct task_struct *p)
1815 unsigned long flags;
1818 raw_spin_lock_irqsave(&p->pi_lock, flags);
1821 * Fork balancing, do it here and not earlier because:
1822 * - cpus_allowed can change in the fork path
1823 * - any previously selected cpu might disappear through hotplug
1825 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1828 rq = __task_rq_lock(p);
1829 activate_task(rq, p, 0);
1831 trace_sched_wakeup_new(p, true);
1832 check_preempt_curr(rq, p, WF_FORK);
1834 if (p->sched_class->task_woken)
1835 p->sched_class->task_woken(rq, p);
1837 task_rq_unlock(rq, p, &flags);
1840 #ifdef CONFIG_PREEMPT_NOTIFIERS
1843 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1844 * @notifier: notifier struct to register
1846 void preempt_notifier_register(struct preempt_notifier *notifier)
1848 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1850 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1853 * preempt_notifier_unregister - no longer interested in preemption notifications
1854 * @notifier: notifier struct to unregister
1856 * This is safe to call from within a preemption notifier.
1858 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1860 hlist_del(¬ifier->link);
1862 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1864 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1866 struct preempt_notifier *notifier;
1867 struct hlist_node *node;
1869 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1870 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1874 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1875 struct task_struct *next)
1877 struct preempt_notifier *notifier;
1878 struct hlist_node *node;
1880 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1881 notifier->ops->sched_out(notifier, next);
1884 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1886 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1891 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1892 struct task_struct *next)
1896 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1899 * prepare_task_switch - prepare to switch tasks
1900 * @rq: the runqueue preparing to switch
1901 * @prev: the current task that is being switched out
1902 * @next: the task we are going to switch to.
1904 * This is called with the rq lock held and interrupts off. It must
1905 * be paired with a subsequent finish_task_switch after the context
1908 * prepare_task_switch sets up locking and calls architecture specific
1912 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1913 struct task_struct *next)
1915 sched_info_switch(prev, next);
1916 perf_event_task_sched_out(prev, next);
1917 fire_sched_out_preempt_notifiers(prev, next);
1918 prepare_lock_switch(rq, next);
1919 prepare_arch_switch(next);
1920 trace_sched_switch(prev, next);
1924 * finish_task_switch - clean up after a task-switch
1925 * @rq: runqueue associated with task-switch
1926 * @prev: the thread we just switched away from.
1928 * finish_task_switch must be called after the context switch, paired
1929 * with a prepare_task_switch call before the context switch.
1930 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1931 * and do any other architecture-specific cleanup actions.
1933 * Note that we may have delayed dropping an mm in context_switch(). If
1934 * so, we finish that here outside of the runqueue lock. (Doing it
1935 * with the lock held can cause deadlocks; see schedule() for
1938 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1939 __releases(rq->lock)
1941 struct mm_struct *mm = rq->prev_mm;
1947 * A task struct has one reference for the use as "current".
1948 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1949 * schedule one last time. The schedule call will never return, and
1950 * the scheduled task must drop that reference.
1951 * The test for TASK_DEAD must occur while the runqueue locks are
1952 * still held, otherwise prev could be scheduled on another cpu, die
1953 * there before we look at prev->state, and then the reference would
1955 * Manfred Spraul <manfred@colorfullife.com>
1957 prev_state = prev->state;
1958 finish_arch_switch(prev);
1959 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1960 local_irq_disable();
1961 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1962 perf_event_task_sched_in(prev, current);
1963 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1965 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1966 finish_lock_switch(rq, prev);
1967 finish_arch_post_lock_switch();
1969 fire_sched_in_preempt_notifiers(current);
1972 if (unlikely(prev_state == TASK_DEAD)) {
1974 * Remove function-return probe instances associated with this
1975 * task and put them back on the free list.
1977 kprobe_flush_task(prev);
1978 put_task_struct(prev);
1984 /* assumes rq->lock is held */
1985 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1987 if (prev->sched_class->pre_schedule)
1988 prev->sched_class->pre_schedule(rq, prev);
1991 /* rq->lock is NOT held, but preemption is disabled */
1992 static inline void post_schedule(struct rq *rq)
1994 if (rq->post_schedule) {
1995 unsigned long flags;
1997 raw_spin_lock_irqsave(&rq->lock, flags);
1998 if (rq->curr->sched_class->post_schedule)
1999 rq->curr->sched_class->post_schedule(rq);
2000 raw_spin_unlock_irqrestore(&rq->lock, flags);
2002 rq->post_schedule = 0;
2008 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2012 static inline void post_schedule(struct rq *rq)
2019 * schedule_tail - first thing a freshly forked thread must call.
2020 * @prev: the thread we just switched away from.
2022 asmlinkage void schedule_tail(struct task_struct *prev)
2023 __releases(rq->lock)
2025 struct rq *rq = this_rq();
2027 finish_task_switch(rq, prev);
2030 * FIXME: do we need to worry about rq being invalidated by the
2035 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2036 /* In this case, finish_task_switch does not reenable preemption */
2039 if (current->set_child_tid)
2040 put_user(task_pid_vnr(current), current->set_child_tid);
2044 * context_switch - switch to the new MM and the new
2045 * thread's register state.
2048 context_switch(struct rq *rq, struct task_struct *prev,
2049 struct task_struct *next)
2051 struct mm_struct *mm, *oldmm;
2053 prepare_task_switch(rq, prev, next);
2056 oldmm = prev->active_mm;
2058 * For paravirt, this is coupled with an exit in switch_to to
2059 * combine the page table reload and the switch backend into
2062 arch_start_context_switch(prev);
2065 next->active_mm = oldmm;
2066 atomic_inc(&oldmm->mm_count);
2067 enter_lazy_tlb(oldmm, next);
2069 switch_mm(oldmm, mm, next);
2072 prev->active_mm = NULL;
2073 rq->prev_mm = oldmm;
2076 * Since the runqueue lock will be released by the next
2077 * task (which is an invalid locking op but in the case
2078 * of the scheduler it's an obvious special-case), so we
2079 * do an early lockdep release here:
2081 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2082 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2085 /* Here we just switch the register state and the stack. */
2086 switch_to(prev, next, prev);
2090 * this_rq must be evaluated again because prev may have moved
2091 * CPUs since it called schedule(), thus the 'rq' on its stack
2092 * frame will be invalid.
2094 finish_task_switch(this_rq(), prev);
2098 * nr_running, nr_uninterruptible and nr_context_switches:
2100 * externally visible scheduler statistics: current number of runnable
2101 * threads, current number of uninterruptible-sleeping threads, total
2102 * number of context switches performed since bootup.
2104 unsigned long nr_running(void)
2106 unsigned long i, sum = 0;
2108 for_each_online_cpu(i)
2109 sum += cpu_rq(i)->nr_running;
2114 unsigned long nr_uninterruptible(void)
2116 unsigned long i, sum = 0;
2118 for_each_possible_cpu(i)
2119 sum += cpu_rq(i)->nr_uninterruptible;
2122 * Since we read the counters lockless, it might be slightly
2123 * inaccurate. Do not allow it to go below zero though:
2125 if (unlikely((long)sum < 0))
2131 unsigned long long nr_context_switches(void)
2134 unsigned long long sum = 0;
2136 for_each_possible_cpu(i)
2137 sum += cpu_rq(i)->nr_switches;
2142 unsigned long nr_iowait(void)
2144 unsigned long i, sum = 0;
2146 for_each_possible_cpu(i)
2147 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2152 unsigned long nr_iowait_cpu(int cpu)
2154 struct rq *this = cpu_rq(cpu);
2155 return atomic_read(&this->nr_iowait);
2158 unsigned long this_cpu_load(void)
2160 struct rq *this = this_rq();
2161 return this->cpu_load[0];
2165 /* Variables and functions for calc_load */
2166 static atomic_long_t calc_load_tasks;
2167 static unsigned long calc_load_update;
2168 unsigned long avenrun[3];
2169 EXPORT_SYMBOL(avenrun);
2171 static long calc_load_fold_active(struct rq *this_rq)
2173 long nr_active, delta = 0;
2175 nr_active = this_rq->nr_running;
2176 nr_active += (long) this_rq->nr_uninterruptible;
2178 if (nr_active != this_rq->calc_load_active) {
2179 delta = nr_active - this_rq->calc_load_active;
2180 this_rq->calc_load_active = nr_active;
2186 static unsigned long
2187 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2190 load += active * (FIXED_1 - exp);
2191 load += 1UL << (FSHIFT - 1);
2192 return load >> FSHIFT;
2197 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2199 * When making the ILB scale, we should try to pull this in as well.
2201 static atomic_long_t calc_load_tasks_idle;
2203 void calc_load_account_idle(struct rq *this_rq)
2207 delta = calc_load_fold_active(this_rq);
2209 atomic_long_add(delta, &calc_load_tasks_idle);
2212 static long calc_load_fold_idle(void)
2217 * Its got a race, we don't care...
2219 if (atomic_long_read(&calc_load_tasks_idle))
2220 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2226 * fixed_power_int - compute: x^n, in O(log n) time
2228 * @x: base of the power
2229 * @frac_bits: fractional bits of @x
2230 * @n: power to raise @x to.
2232 * By exploiting the relation between the definition of the natural power
2233 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2234 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2235 * (where: n_i \elem {0, 1}, the binary vector representing n),
2236 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2237 * of course trivially computable in O(log_2 n), the length of our binary
2240 static unsigned long
2241 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2243 unsigned long result = 1UL << frac_bits;
2248 result += 1UL << (frac_bits - 1);
2249 result >>= frac_bits;
2255 x += 1UL << (frac_bits - 1);
2263 * a1 = a0 * e + a * (1 - e)
2265 * a2 = a1 * e + a * (1 - e)
2266 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2267 * = a0 * e^2 + a * (1 - e) * (1 + e)
2269 * a3 = a2 * e + a * (1 - e)
2270 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2271 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2275 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2276 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2277 * = a0 * e^n + a * (1 - e^n)
2279 * [1] application of the geometric series:
2282 * S_n := \Sum x^i = -------------
2285 static unsigned long
2286 calc_load_n(unsigned long load, unsigned long exp,
2287 unsigned long active, unsigned int n)
2290 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2294 * NO_HZ can leave us missing all per-cpu ticks calling
2295 * calc_load_account_active(), but since an idle CPU folds its delta into
2296 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2297 * in the pending idle delta if our idle period crossed a load cycle boundary.
2299 * Once we've updated the global active value, we need to apply the exponential
2300 * weights adjusted to the number of cycles missed.
2302 static void calc_global_nohz(void)
2304 long delta, active, n;
2307 * If we crossed a calc_load_update boundary, make sure to fold
2308 * any pending idle changes, the respective CPUs might have
2309 * missed the tick driven calc_load_account_active() update
2312 delta = calc_load_fold_idle();
2314 atomic_long_add(delta, &calc_load_tasks);
2317 * It could be the one fold was all it took, we done!
2319 if (time_before(jiffies, calc_load_update + 10))
2323 * Catch-up, fold however many we are behind still
2325 delta = jiffies - calc_load_update - 10;
2326 n = 1 + (delta / LOAD_FREQ);
2328 active = atomic_long_read(&calc_load_tasks);
2329 active = active > 0 ? active * FIXED_1 : 0;
2331 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2332 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2333 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2335 calc_load_update += n * LOAD_FREQ;
2338 void calc_load_account_idle(struct rq *this_rq)
2342 static inline long calc_load_fold_idle(void)
2347 static void calc_global_nohz(void)
2353 * get_avenrun - get the load average array
2354 * @loads: pointer to dest load array
2355 * @offset: offset to add
2356 * @shift: shift count to shift the result left
2358 * These values are estimates at best, so no need for locking.
2360 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2362 loads[0] = (avenrun[0] + offset) << shift;
2363 loads[1] = (avenrun[1] + offset) << shift;
2364 loads[2] = (avenrun[2] + offset) << shift;
2368 * calc_load - update the avenrun load estimates 10 ticks after the
2369 * CPUs have updated calc_load_tasks.
2371 void calc_global_load(unsigned long ticks)
2375 if (time_before(jiffies, calc_load_update + 10))
2378 active = atomic_long_read(&calc_load_tasks);
2379 active = active > 0 ? active * FIXED_1 : 0;
2381 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2382 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2383 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2385 calc_load_update += LOAD_FREQ;
2388 * Account one period with whatever state we found before
2389 * folding in the nohz state and ageing the entire idle period.
2391 * This avoids loosing a sample when we go idle between
2392 * calc_load_account_active() (10 ticks ago) and now and thus
2399 * Called from update_cpu_load() to periodically update this CPU's
2402 static void calc_load_account_active(struct rq *this_rq)
2406 if (time_before(jiffies, this_rq->calc_load_update))
2409 delta = calc_load_fold_active(this_rq);
2410 delta += calc_load_fold_idle();
2412 atomic_long_add(delta, &calc_load_tasks);
2414 this_rq->calc_load_update += LOAD_FREQ;
2418 * The exact cpuload at various idx values, calculated at every tick would be
2419 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2421 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2422 * on nth tick when cpu may be busy, then we have:
2423 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2424 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2426 * decay_load_missed() below does efficient calculation of
2427 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2428 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2430 * The calculation is approximated on a 128 point scale.
2431 * degrade_zero_ticks is the number of ticks after which load at any
2432 * particular idx is approximated to be zero.
2433 * degrade_factor is a precomputed table, a row for each load idx.
2434 * Each column corresponds to degradation factor for a power of two ticks,
2435 * based on 128 point scale.
2437 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2438 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2440 * With this power of 2 load factors, we can degrade the load n times
2441 * by looking at 1 bits in n and doing as many mult/shift instead of
2442 * n mult/shifts needed by the exact degradation.
2444 #define DEGRADE_SHIFT 7
2445 static const unsigned char
2446 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2447 static const unsigned char
2448 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2449 {0, 0, 0, 0, 0, 0, 0, 0},
2450 {64, 32, 8, 0, 0, 0, 0, 0},
2451 {96, 72, 40, 12, 1, 0, 0},
2452 {112, 98, 75, 43, 15, 1, 0},
2453 {120, 112, 98, 76, 45, 16, 2} };
2456 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2457 * would be when CPU is idle and so we just decay the old load without
2458 * adding any new load.
2460 static unsigned long
2461 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2465 if (!missed_updates)
2468 if (missed_updates >= degrade_zero_ticks[idx])
2472 return load >> missed_updates;
2474 while (missed_updates) {
2475 if (missed_updates % 2)
2476 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2478 missed_updates >>= 1;
2485 * Update rq->cpu_load[] statistics. This function is usually called every
2486 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2487 * every tick. We fix it up based on jiffies.
2489 void update_cpu_load(struct rq *this_rq)
2491 unsigned long this_load = this_rq->load.weight;
2492 unsigned long curr_jiffies = jiffies;
2493 unsigned long pending_updates;
2496 this_rq->nr_load_updates++;
2498 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
2499 if (curr_jiffies == this_rq->last_load_update_tick)
2502 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2503 this_rq->last_load_update_tick = curr_jiffies;
2505 /* Update our load: */
2506 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2507 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2508 unsigned long old_load, new_load;
2510 /* scale is effectively 1 << i now, and >> i divides by scale */
2512 old_load = this_rq->cpu_load[i];
2513 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2514 new_load = this_load;
2516 * Round up the averaging division if load is increasing. This
2517 * prevents us from getting stuck on 9 if the load is 10, for
2520 if (new_load > old_load)
2521 new_load += scale - 1;
2523 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2526 sched_avg_update(this_rq);
2529 static void update_cpu_load_active(struct rq *this_rq)
2531 update_cpu_load(this_rq);
2533 calc_load_account_active(this_rq);
2539 * sched_exec - execve() is a valuable balancing opportunity, because at
2540 * this point the task has the smallest effective memory and cache footprint.
2542 void sched_exec(void)
2544 struct task_struct *p = current;
2545 unsigned long flags;
2548 raw_spin_lock_irqsave(&p->pi_lock, flags);
2549 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2550 if (dest_cpu == smp_processor_id())
2553 if (likely(cpu_active(dest_cpu))) {
2554 struct migration_arg arg = { p, dest_cpu };
2556 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2557 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2561 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2566 DEFINE_PER_CPU(struct kernel_stat, kstat);
2567 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2569 EXPORT_PER_CPU_SYMBOL(kstat);
2570 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2573 * Return any ns on the sched_clock that have not yet been accounted in
2574 * @p in case that task is currently running.
2576 * Called with task_rq_lock() held on @rq.
2578 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2582 if (task_current(rq, p)) {
2583 update_rq_clock(rq);
2584 ns = rq->clock_task - p->se.exec_start;
2592 unsigned long long task_delta_exec(struct task_struct *p)
2594 unsigned long flags;
2598 rq = task_rq_lock(p, &flags);
2599 ns = do_task_delta_exec(p, rq);
2600 task_rq_unlock(rq, p, &flags);
2606 * Return accounted runtime for the task.
2607 * In case the task is currently running, return the runtime plus current's
2608 * pending runtime that have not been accounted yet.
2610 unsigned long long task_sched_runtime(struct task_struct *p)
2612 unsigned long flags;
2616 rq = task_rq_lock(p, &flags);
2617 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2618 task_rq_unlock(rq, p, &flags);
2623 #ifdef CONFIG_CGROUP_CPUACCT
2624 struct cgroup_subsys cpuacct_subsys;
2625 struct cpuacct root_cpuacct;
2628 static inline void task_group_account_field(struct task_struct *p, int index,
2631 #ifdef CONFIG_CGROUP_CPUACCT
2632 struct kernel_cpustat *kcpustat;
2636 * Since all updates are sure to touch the root cgroup, we
2637 * get ourselves ahead and touch it first. If the root cgroup
2638 * is the only cgroup, then nothing else should be necessary.
2641 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2643 #ifdef CONFIG_CGROUP_CPUACCT
2644 if (unlikely(!cpuacct_subsys.active))
2649 while (ca && (ca != &root_cpuacct)) {
2650 kcpustat = this_cpu_ptr(ca->cpustat);
2651 kcpustat->cpustat[index] += tmp;
2659 #if !defined(CONFIG_XEN) || defined(CONFIG_VIRT_CPU_ACCOUNTING)
2660 # define cputime_to_u64(t) ((__force u64)(t))
2662 # include <linux/syscore_ops.h>
2663 # define NS_PER_TICK (1000000000 / HZ)
2665 static DEFINE_PER_CPU(u64, steal_snapshot);
2666 static DEFINE_PER_CPU(unsigned int, steal_residual);
2668 static u64 cputime_to_u64(cputime_t t)
2670 u64 s = this_vcpu_read(runstate.time[RUNSTATE_runnable]);
2671 unsigned long adj = div_u64_rem(s - __this_cpu_read(steal_snapshot)
2672 + __this_cpu_read(steal_residual),
2674 &__get_cpu_var(steal_residual));
2676 __this_cpu_write(steal_snapshot, s);
2677 if (t < jiffies_to_cputime(adj))
2680 return (__force u64)(t - jiffies_to_cputime(adj));
2683 static void steal_resume(void)
2685 cputime_to_u64(((cputime_t)1 << (BITS_PER_LONG * sizeof(cputime_t)
2686 / sizeof(long) - 1)) - 1);
2689 static struct syscore_ops steal_syscore_ops = {
2690 .resume = steal_resume,
2693 static int __init steal_register(void)
2695 register_syscore_ops(&steal_syscore_ops);
2698 core_initcall(steal_register);
2702 * Account user cpu time to a process.
2703 * @p: the process that the cpu time gets accounted to
2704 * @cputime: the cpu time spent in user space since the last update
2705 * @cputime_scaled: cputime scaled by cpu frequency
2707 void account_user_time(struct task_struct *p, cputime_t cputime,
2708 cputime_t cputime_scaled)
2712 /* Add user time to process. */
2713 p->utime += cputime;
2714 p->utimescaled += cputime_scaled;
2715 account_group_user_time(p, cputime);
2717 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2719 /* Add user time to cpustat. */
2720 task_group_account_field(p, index, cputime_to_u64(cputime));
2722 /* Account for user time used */
2723 acct_update_integrals(p);
2727 * Account guest cpu time to a process.
2728 * @p: the process that the cpu time gets accounted to
2729 * @cputime: the cpu time spent in virtual machine since the last update
2730 * @cputime_scaled: cputime scaled by cpu frequency
2732 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2733 cputime_t cputime_scaled)
2735 u64 *cpustat = kcpustat_this_cpu->cpustat;
2737 /* Add guest time to process. */
2738 p->utime += cputime;
2739 p->utimescaled += cputime_scaled;
2740 account_group_user_time(p, cputime);
2741 p->gtime += cputime;
2743 /* Add guest time to cpustat. */
2744 if (TASK_NICE(p) > 0) {
2745 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2746 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2748 cpustat[CPUTIME_USER] += (__force u64) cputime;
2749 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2754 * Account system cpu time to a process and desired cpustat field
2755 * @p: the process that the cpu time gets accounted to
2756 * @cputime: the cpu time spent in kernel space since the last update
2757 * @cputime_scaled: cputime scaled by cpu frequency
2758 * @target_cputime64: pointer to cpustat field that has to be updated
2761 void __account_system_time(struct task_struct *p, cputime_t cputime,
2762 cputime_t cputime_scaled, int index)
2764 /* Add system time to process. */
2765 p->stime += cputime;
2766 p->stimescaled += cputime_scaled;
2767 account_group_system_time(p, cputime);
2769 /* Add system time to cpustat. */
2770 task_group_account_field(p, index, cputime_to_u64(cputime));
2772 /* Account for system time used */
2773 acct_update_integrals(p);
2777 * Account system cpu time to a process.
2778 * @p: the process that the cpu time gets accounted to
2779 * @hardirq_offset: the offset to subtract from hardirq_count()
2780 * @cputime: the cpu time spent in kernel space since the last update
2781 * @cputime_scaled: cputime scaled by cpu frequency
2783 void account_system_time(struct task_struct *p, int hardirq_offset,
2784 cputime_t cputime, cputime_t cputime_scaled)
2788 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2789 account_guest_time(p, cputime, cputime_scaled);
2793 if (hardirq_count() - hardirq_offset)
2794 index = CPUTIME_IRQ;
2795 else if (in_serving_softirq())
2796 index = CPUTIME_SOFTIRQ;
2798 index = CPUTIME_SYSTEM;
2800 __account_system_time(p, cputime, cputime_scaled, index);
2804 * Account for involuntary wait time.
2805 * @cputime: the cpu time spent in involuntary wait
2807 void account_steal_time(cputime_t cputime)
2809 u64 *cpustat = kcpustat_this_cpu->cpustat;
2811 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2815 * Account for idle time.
2816 * @cputime: the cpu time spent in idle wait
2818 void account_idle_time(cputime_t cputime)
2820 u64 *cpustat = kcpustat_this_cpu->cpustat;
2821 struct rq *rq = this_rq();
2823 if (atomic_read(&rq->nr_iowait) > 0)
2824 cpustat[CPUTIME_IOWAIT] += cputime_to_u64(cputime);
2826 cpustat[CPUTIME_IDLE] += cputime_to_u64(cputime);
2829 static __always_inline bool steal_account_process_tick(void)
2831 #ifdef CONFIG_PARAVIRT
2832 if (static_key_false(¶virt_steal_enabled)) {
2835 steal = paravirt_steal_clock(smp_processor_id());
2836 steal -= this_rq()->prev_steal_time;
2838 st = steal_ticks(steal);
2839 this_rq()->prev_steal_time += st * TICK_NSEC;
2841 account_steal_time(st);
2848 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2850 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2852 * Account a tick to a process and cpustat
2853 * @p: the process that the cpu time gets accounted to
2854 * @user_tick: is the tick from userspace
2855 * @rq: the pointer to rq
2857 * Tick demultiplexing follows the order
2858 * - pending hardirq update
2859 * - pending softirq update
2863 * - check for guest_time
2864 * - else account as system_time
2866 * Check for hardirq is done both for system and user time as there is
2867 * no timer going off while we are on hardirq and hence we may never get an
2868 * opportunity to update it solely in system time.
2869 * p->stime and friends are only updated on system time and not on irq
2870 * softirq as those do not count in task exec_runtime any more.
2872 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2875 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2876 u64 *cpustat = kcpustat_this_cpu->cpustat;
2878 if (steal_account_process_tick())
2881 if (irqtime_account_hi_update()) {
2882 cpustat[CPUTIME_IRQ] += cputime_to_u64(cputime_one_jiffy);
2883 } else if (irqtime_account_si_update()) {
2884 cpustat[CPUTIME_SOFTIRQ] += cputime_to_u64(cputime_one_jiffy);
2885 } else if (this_cpu_ksoftirqd() == p) {
2887 * ksoftirqd time do not get accounted in cpu_softirq_time.
2888 * So, we have to handle it separately here.
2889 * Also, p->stime needs to be updated for ksoftirqd.
2891 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2893 } else if (user_tick) {
2894 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2895 } else if (p == rq->idle) {
2896 account_idle_time(cputime_one_jiffy);
2897 } else if (p->flags & PF_VCPU) { /* System time or guest time */
2898 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
2900 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
2905 static void irqtime_account_idle_ticks(int ticks)
2908 struct rq *rq = this_rq();
2910 for (i = 0; i < ticks; i++)
2911 irqtime_account_process_tick(current, 0, rq);
2913 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2914 static void irqtime_account_idle_ticks(int ticks) {}
2915 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
2917 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2920 * Account a single tick of cpu time.
2921 * @p: the process that the cpu time gets accounted to
2922 * @user_tick: indicates if the tick is a user or a system tick
2924 void account_process_tick(struct task_struct *p, int user_tick)
2926 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
2927 struct rq *rq = this_rq();
2929 if (sched_clock_irqtime) {
2930 irqtime_account_process_tick(p, user_tick, rq);
2934 if (steal_account_process_tick())
2938 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
2939 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
2940 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
2943 account_idle_time(cputime_one_jiffy);
2947 * Account multiple ticks of steal time.
2948 * @p: the process from which the cpu time has been stolen
2949 * @ticks: number of stolen ticks
2951 void account_steal_ticks(unsigned long ticks)
2953 account_steal_time(jiffies_to_cputime(ticks));
2957 * Account multiple ticks of idle time.
2958 * @ticks: number of stolen ticks
2960 void account_idle_ticks(unsigned long ticks)
2963 if (sched_clock_irqtime) {
2964 irqtime_account_idle_ticks(ticks);
2968 account_idle_time(jiffies_to_cputime(ticks));
2974 * Use precise platform statistics if available:
2976 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
2977 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2983 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
2985 struct task_cputime cputime;
2987 thread_group_cputime(p, &cputime);
2989 *ut = cputime.utime;
2990 *st = cputime.stime;
2994 #ifndef nsecs_to_cputime
2995 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
2998 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3000 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3003 * Use CFS's precise accounting:
3005 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3008 u64 temp = (__force u64) rtime;
3010 temp *= (__force u64) utime;
3011 do_div(temp, (__force u32) total);
3012 utime = (__force cputime_t) temp;
3017 * Compare with previous values, to keep monotonicity:
3019 p->prev_utime = max(p->prev_utime, utime);
3020 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3022 *ut = p->prev_utime;
3023 *st = p->prev_stime;
3027 * Must be called with siglock held.
3029 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3031 struct signal_struct *sig = p->signal;
3032 struct task_cputime cputime;
3033 cputime_t rtime, utime, total;
3035 thread_group_cputime(p, &cputime);
3037 total = cputime.utime + cputime.stime;
3038 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3041 u64 temp = (__force u64) rtime;
3043 temp *= (__force u64) cputime.utime;
3044 do_div(temp, (__force u32) total);
3045 utime = (__force cputime_t) temp;
3049 sig->prev_utime = max(sig->prev_utime, utime);
3050 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3052 *ut = sig->prev_utime;
3053 *st = sig->prev_stime;
3058 * This function gets called by the timer code, with HZ frequency.
3059 * We call it with interrupts disabled.
3061 void scheduler_tick(void)
3063 int cpu = smp_processor_id();
3064 struct rq *rq = cpu_rq(cpu);
3065 struct task_struct *curr = rq->curr;
3069 raw_spin_lock(&rq->lock);
3070 update_rq_clock(rq);
3071 update_cpu_load_active(rq);
3072 curr->sched_class->task_tick(rq, curr, 0);
3073 raw_spin_unlock(&rq->lock);
3075 perf_event_task_tick();
3078 rq->idle_balance = idle_cpu(cpu);
3079 trigger_load_balance(rq, cpu);
3083 notrace unsigned long get_parent_ip(unsigned long addr)
3085 if (in_lock_functions(addr)) {
3086 addr = CALLER_ADDR2;
3087 if (in_lock_functions(addr))
3088 addr = CALLER_ADDR3;
3093 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3094 defined(CONFIG_PREEMPT_TRACER))
3096 void __kprobes add_preempt_count(int val)
3098 #ifdef CONFIG_DEBUG_PREEMPT
3102 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3105 preempt_count() += val;
3106 #ifdef CONFIG_DEBUG_PREEMPT
3108 * Spinlock count overflowing soon?
3110 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3113 if (preempt_count() == val)
3114 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3116 EXPORT_SYMBOL(add_preempt_count);
3118 void __kprobes sub_preempt_count(int val)
3120 #ifdef CONFIG_DEBUG_PREEMPT
3124 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3127 * Is the spinlock portion underflowing?
3129 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3130 !(preempt_count() & PREEMPT_MASK)))
3134 if (preempt_count() == val)
3135 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3136 preempt_count() -= val;
3138 EXPORT_SYMBOL(sub_preempt_count);
3143 * Print scheduling while atomic bug:
3145 static noinline void __schedule_bug(struct task_struct *prev)
3147 if (oops_in_progress)
3150 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3151 prev->comm, prev->pid, preempt_count());
3153 debug_show_held_locks(prev);
3155 if (irqs_disabled())
3156 print_irqtrace_events(prev);
3161 * Various schedule()-time debugging checks and statistics:
3163 static inline void schedule_debug(struct task_struct *prev)
3166 * Test if we are atomic. Since do_exit() needs to call into
3167 * schedule() atomically, we ignore that path for now.
3168 * Otherwise, whine if we are scheduling when we should not be.
3170 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3171 __schedule_bug(prev);
3174 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3176 schedstat_inc(this_rq(), sched_count);
3179 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3181 if (prev->on_rq || rq->skip_clock_update < 0)
3182 update_rq_clock(rq);
3183 prev->sched_class->put_prev_task(rq, prev);
3187 * Pick up the highest-prio task:
3189 static inline struct task_struct *
3190 pick_next_task(struct rq *rq)
3192 const struct sched_class *class;
3193 struct task_struct *p;
3196 * Optimization: we know that if all tasks are in
3197 * the fair class we can call that function directly:
3199 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3200 p = fair_sched_class.pick_next_task(rq);
3205 for_each_class(class) {
3206 p = class->pick_next_task(rq);
3211 BUG(); /* the idle class will always have a runnable task */
3215 * __schedule() is the main scheduler function.
3217 static void __sched __schedule(void)
3219 struct task_struct *prev, *next;
3220 unsigned long *switch_count;
3226 cpu = smp_processor_id();
3228 rcu_note_context_switch(cpu);
3231 schedule_debug(prev);
3233 if (sched_feat(HRTICK))
3236 raw_spin_lock_irq(&rq->lock);
3238 switch_count = &prev->nivcsw;
3239 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3240 if (unlikely(signal_pending_state(prev->state, prev))) {
3241 prev->state = TASK_RUNNING;
3243 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3247 * If a worker went to sleep, notify and ask workqueue
3248 * whether it wants to wake up a task to maintain
3251 if (prev->flags & PF_WQ_WORKER) {
3252 struct task_struct *to_wakeup;
3254 to_wakeup = wq_worker_sleeping(prev, cpu);
3256 try_to_wake_up_local(to_wakeup);
3259 switch_count = &prev->nvcsw;
3262 pre_schedule(rq, prev);
3264 if (unlikely(!rq->nr_running))
3265 idle_balance(cpu, rq);
3267 put_prev_task(rq, prev);
3268 next = pick_next_task(rq);
3269 clear_tsk_need_resched(prev);
3270 rq->skip_clock_update = 0;
3272 if (likely(prev != next)) {
3277 context_switch(rq, prev, next); /* unlocks the rq */
3279 * The context switch have flipped the stack from under us
3280 * and restored the local variables which were saved when
3281 * this task called schedule() in the past. prev == current
3282 * is still correct, but it can be moved to another cpu/rq.
3284 cpu = smp_processor_id();
3287 raw_spin_unlock_irq(&rq->lock);
3291 sched_preempt_enable_no_resched();
3296 static inline void sched_submit_work(struct task_struct *tsk)
3298 if (!tsk->state || tsk_is_pi_blocked(tsk))
3301 * If we are going to sleep and we have plugged IO queued,
3302 * make sure to submit it to avoid deadlocks.
3304 if (blk_needs_flush_plug(tsk))
3305 blk_schedule_flush_plug(tsk);
3308 asmlinkage void __sched schedule(void)
3310 struct task_struct *tsk = current;
3312 sched_submit_work(tsk);
3315 EXPORT_SYMBOL(schedule);
3318 * schedule_preempt_disabled - called with preemption disabled
3320 * Returns with preemption disabled. Note: preempt_count must be 1
3322 void __sched schedule_preempt_disabled(void)
3324 sched_preempt_enable_no_resched();
3329 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3330 #include <asm/mutex.h>
3332 #ifndef arch_cpu_is_running
3333 #define arch_cpu_is_running(cpu) true
3336 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3338 if (lock->owner != owner)
3342 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3343 * lock->owner still matches owner, if that fails, owner might
3344 * point to free()d memory, if it still matches, the rcu_read_lock()
3345 * ensures the memory stays valid.
3349 return owner->on_cpu
3350 && arch_cpu_is_running(task_thread_info(owner)->cpu);
3354 * Look out! "owner" is an entirely speculative pointer
3355 * access and not reliable.
3357 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3359 if (!sched_feat(OWNER_SPIN))
3363 while (owner_running(lock, owner)) {
3367 arch_mutex_cpu_relax();
3372 * We break out the loop above on need_resched() and when the
3373 * owner changed, which is a sign for heavy contention. Return
3374 * success only when lock->owner is NULL.
3376 return lock->owner == NULL;
3380 #ifdef CONFIG_PREEMPT
3382 * this is the entry point to schedule() from in-kernel preemption
3383 * off of preempt_enable. Kernel preemptions off return from interrupt
3384 * occur there and call schedule directly.
3386 asmlinkage void __sched notrace preempt_schedule(void)
3388 struct thread_info *ti = current_thread_info();
3391 * If there is a non-zero preempt_count or interrupts are disabled,
3392 * we do not want to preempt the current task. Just return..
3394 if (likely(ti->preempt_count || irqs_disabled()))
3398 add_preempt_count_notrace(PREEMPT_ACTIVE);
3400 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3403 * Check again in case we missed a preemption opportunity
3404 * between schedule and now.
3407 } while (need_resched());
3409 EXPORT_SYMBOL(preempt_schedule);
3412 * this is the entry point to schedule() from kernel preemption
3413 * off of irq context.
3414 * Note, that this is called and return with irqs disabled. This will
3415 * protect us against recursive calling from irq.
3417 asmlinkage void __sched preempt_schedule_irq(void)
3419 struct thread_info *ti = current_thread_info();
3421 /* Catch callers which need to be fixed */
3422 BUG_ON(ti->preempt_count || !irqs_disabled());
3425 add_preempt_count(PREEMPT_ACTIVE);
3428 local_irq_disable();
3429 sub_preempt_count(PREEMPT_ACTIVE);
3432 * Check again in case we missed a preemption opportunity
3433 * between schedule and now.
3436 } while (need_resched());
3439 #endif /* CONFIG_PREEMPT */
3441 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3444 return try_to_wake_up(curr->private, mode, wake_flags);
3446 EXPORT_SYMBOL(default_wake_function);
3449 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3450 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3451 * number) then we wake all the non-exclusive tasks and one exclusive task.
3453 * There are circumstances in which we can try to wake a task which has already
3454 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3455 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3457 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3458 int nr_exclusive, int wake_flags, void *key)
3460 wait_queue_t *curr, *next;
3462 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3463 unsigned flags = curr->flags;
3465 if (curr->func(curr, mode, wake_flags, key) &&
3466 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3472 * __wake_up - wake up threads blocked on a waitqueue.
3474 * @mode: which threads
3475 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3476 * @key: is directly passed to the wakeup function
3478 * It may be assumed that this function implies a write memory barrier before
3479 * changing the task state if and only if any tasks are woken up.
3481 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3482 int nr_exclusive, void *key)
3484 unsigned long flags;
3486 spin_lock_irqsave(&q->lock, flags);
3487 __wake_up_common(q, mode, nr_exclusive, 0, key);
3488 spin_unlock_irqrestore(&q->lock, flags);
3490 EXPORT_SYMBOL(__wake_up);
3493 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3495 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3497 __wake_up_common(q, mode, nr, 0, NULL);
3499 EXPORT_SYMBOL_GPL(__wake_up_locked);
3501 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3503 __wake_up_common(q, mode, 1, 0, key);
3505 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3508 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3510 * @mode: which threads
3511 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3512 * @key: opaque value to be passed to wakeup targets
3514 * The sync wakeup differs that the waker knows that it will schedule
3515 * away soon, so while the target thread will be woken up, it will not
3516 * be migrated to another CPU - ie. the two threads are 'synchronized'
3517 * with each other. This can prevent needless bouncing between CPUs.
3519 * On UP it can prevent extra preemption.
3521 * It may be assumed that this function implies a write memory barrier before
3522 * changing the task state if and only if any tasks are woken up.
3524 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3525 int nr_exclusive, void *key)
3527 unsigned long flags;
3528 int wake_flags = WF_SYNC;
3533 if (unlikely(!nr_exclusive))
3536 spin_lock_irqsave(&q->lock, flags);
3537 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3538 spin_unlock_irqrestore(&q->lock, flags);
3540 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3543 * __wake_up_sync - see __wake_up_sync_key()
3545 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3547 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3549 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3552 * complete: - signals a single thread waiting on this completion
3553 * @x: holds the state of this particular completion
3555 * This will wake up a single thread waiting on this completion. Threads will be
3556 * awakened in the same order in which they were queued.
3558 * See also complete_all(), wait_for_completion() and related routines.
3560 * It may be assumed that this function implies a write memory barrier before
3561 * changing the task state if and only if any tasks are woken up.
3563 void complete(struct completion *x)
3565 unsigned long flags;
3567 spin_lock_irqsave(&x->wait.lock, flags);
3569 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3570 spin_unlock_irqrestore(&x->wait.lock, flags);
3572 EXPORT_SYMBOL(complete);
3575 * complete_all: - signals all threads waiting on this completion
3576 * @x: holds the state of this particular completion
3578 * This will wake up all threads waiting on this particular completion event.
3580 * It may be assumed that this function implies a write memory barrier before
3581 * changing the task state if and only if any tasks are woken up.
3583 void complete_all(struct completion *x)
3585 unsigned long flags;
3587 spin_lock_irqsave(&x->wait.lock, flags);
3588 x->done += UINT_MAX/2;
3589 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3590 spin_unlock_irqrestore(&x->wait.lock, flags);
3592 EXPORT_SYMBOL(complete_all);
3594 static inline long __sched
3595 do_wait_for_common(struct completion *x, long timeout, int state)
3598 DECLARE_WAITQUEUE(wait, current);
3600 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3602 if (signal_pending_state(state, current)) {
3603 timeout = -ERESTARTSYS;
3606 __set_current_state(state);
3607 spin_unlock_irq(&x->wait.lock);
3608 timeout = schedule_timeout(timeout);
3609 spin_lock_irq(&x->wait.lock);
3610 } while (!x->done && timeout);
3611 __remove_wait_queue(&x->wait, &wait);
3616 return timeout ?: 1;
3620 wait_for_common(struct completion *x, long timeout, int state)
3624 spin_lock_irq(&x->wait.lock);
3625 timeout = do_wait_for_common(x, timeout, state);
3626 spin_unlock_irq(&x->wait.lock);
3631 * wait_for_completion: - waits for completion of a task
3632 * @x: holds the state of this particular completion
3634 * This waits to be signaled for completion of a specific task. It is NOT
3635 * interruptible and there is no timeout.
3637 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3638 * and interrupt capability. Also see complete().
3640 void __sched wait_for_completion(struct completion *x)
3642 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3644 EXPORT_SYMBOL(wait_for_completion);
3647 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3648 * @x: holds the state of this particular completion
3649 * @timeout: timeout value in jiffies
3651 * This waits for either a completion of a specific task to be signaled or for a
3652 * specified timeout to expire. The timeout is in jiffies. It is not
3655 * The return value is 0 if timed out, and positive (at least 1, or number of
3656 * jiffies left till timeout) if completed.
3658 unsigned long __sched
3659 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3661 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3663 EXPORT_SYMBOL(wait_for_completion_timeout);
3666 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3667 * @x: holds the state of this particular completion
3669 * This waits for completion of a specific task to be signaled. It is
3672 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3674 int __sched wait_for_completion_interruptible(struct completion *x)
3676 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3677 if (t == -ERESTARTSYS)
3681 EXPORT_SYMBOL(wait_for_completion_interruptible);
3684 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3685 * @x: holds the state of this particular completion
3686 * @timeout: timeout value in jiffies
3688 * This waits for either a completion of a specific task to be signaled or for a
3689 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3691 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3692 * positive (at least 1, or number of jiffies left till timeout) if completed.
3695 wait_for_completion_interruptible_timeout(struct completion *x,
3696 unsigned long timeout)
3698 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3700 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3703 * wait_for_completion_killable: - waits for completion of a task (killable)
3704 * @x: holds the state of this particular completion
3706 * This waits to be signaled for completion of a specific task. It can be
3707 * interrupted by a kill signal.
3709 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3711 int __sched wait_for_completion_killable(struct completion *x)
3713 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3714 if (t == -ERESTARTSYS)
3718 EXPORT_SYMBOL(wait_for_completion_killable);
3721 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3722 * @x: holds the state of this particular completion
3723 * @timeout: timeout value in jiffies
3725 * This waits for either a completion of a specific task to be
3726 * signaled or for a specified timeout to expire. It can be
3727 * interrupted by a kill signal. The timeout is in jiffies.
3729 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3730 * positive (at least 1, or number of jiffies left till timeout) if completed.
3733 wait_for_completion_killable_timeout(struct completion *x,
3734 unsigned long timeout)
3736 return wait_for_common(x, timeout, TASK_KILLABLE);
3738 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3741 * try_wait_for_completion - try to decrement a completion without blocking
3742 * @x: completion structure
3744 * Returns: 0 if a decrement cannot be done without blocking
3745 * 1 if a decrement succeeded.
3747 * If a completion is being used as a counting completion,
3748 * attempt to decrement the counter without blocking. This
3749 * enables us to avoid waiting if the resource the completion
3750 * is protecting is not available.
3752 bool try_wait_for_completion(struct completion *x)
3754 unsigned long flags;
3757 spin_lock_irqsave(&x->wait.lock, flags);
3762 spin_unlock_irqrestore(&x->wait.lock, flags);
3765 EXPORT_SYMBOL(try_wait_for_completion);
3768 * completion_done - Test to see if a completion has any waiters
3769 * @x: completion structure
3771 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3772 * 1 if there are no waiters.
3775 bool completion_done(struct completion *x)
3777 unsigned long flags;
3780 spin_lock_irqsave(&x->wait.lock, flags);
3783 spin_unlock_irqrestore(&x->wait.lock, flags);
3786 EXPORT_SYMBOL(completion_done);
3789 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3791 unsigned long flags;
3794 init_waitqueue_entry(&wait, current);
3796 __set_current_state(state);
3798 spin_lock_irqsave(&q->lock, flags);
3799 __add_wait_queue(q, &wait);
3800 spin_unlock(&q->lock);
3801 timeout = schedule_timeout(timeout);
3802 spin_lock_irq(&q->lock);
3803 __remove_wait_queue(q, &wait);
3804 spin_unlock_irqrestore(&q->lock, flags);
3809 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3811 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3813 EXPORT_SYMBOL(interruptible_sleep_on);
3816 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3818 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3820 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3822 void __sched sleep_on(wait_queue_head_t *q)
3824 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3826 EXPORT_SYMBOL(sleep_on);
3828 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3830 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3832 EXPORT_SYMBOL(sleep_on_timeout);
3834 #ifdef CONFIG_RT_MUTEXES
3837 * rt_mutex_setprio - set the current priority of a task
3839 * @prio: prio value (kernel-internal form)
3841 * This function changes the 'effective' priority of a task. It does
3842 * not touch ->normal_prio like __setscheduler().
3844 * Used by the rt_mutex code to implement priority inheritance logic.
3846 void rt_mutex_setprio(struct task_struct *p, int prio)
3848 int oldprio, on_rq, running;
3850 const struct sched_class *prev_class;
3852 BUG_ON(prio < 0 || prio > MAX_PRIO);
3854 rq = __task_rq_lock(p);
3857 * Idle task boosting is a nono in general. There is one
3858 * exception, when PREEMPT_RT and NOHZ is active:
3860 * The idle task calls get_next_timer_interrupt() and holds
3861 * the timer wheel base->lock on the CPU and another CPU wants
3862 * to access the timer (probably to cancel it). We can safely
3863 * ignore the boosting request, as the idle CPU runs this code
3864 * with interrupts disabled and will complete the lock
3865 * protected section without being interrupted. So there is no
3866 * real need to boost.
3868 if (unlikely(p == rq->idle)) {
3869 WARN_ON(p != rq->curr);
3870 WARN_ON(p->pi_blocked_on);
3874 trace_sched_pi_setprio(p, prio);
3876 prev_class = p->sched_class;
3878 running = task_current(rq, p);
3880 dequeue_task(rq, p, 0);
3882 p->sched_class->put_prev_task(rq, p);
3885 p->sched_class = &rt_sched_class;
3887 p->sched_class = &fair_sched_class;
3892 p->sched_class->set_curr_task(rq);
3894 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3896 check_class_changed(rq, p, prev_class, oldprio);
3898 __task_rq_unlock(rq);
3901 void set_user_nice(struct task_struct *p, long nice)
3903 int old_prio, delta, on_rq;
3904 unsigned long flags;
3907 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3910 * We have to be careful, if called from sys_setpriority(),
3911 * the task might be in the middle of scheduling on another CPU.
3913 rq = task_rq_lock(p, &flags);
3915 * The RT priorities are set via sched_setscheduler(), but we still
3916 * allow the 'normal' nice value to be set - but as expected
3917 * it wont have any effect on scheduling until the task is
3918 * SCHED_FIFO/SCHED_RR:
3920 if (task_has_rt_policy(p)) {
3921 p->static_prio = NICE_TO_PRIO(nice);
3926 dequeue_task(rq, p, 0);
3928 p->static_prio = NICE_TO_PRIO(nice);
3931 p->prio = effective_prio(p);
3932 delta = p->prio - old_prio;
3935 enqueue_task(rq, p, 0);
3937 * If the task increased its priority or is running and
3938 * lowered its priority, then reschedule its CPU:
3940 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3941 resched_task(rq->curr);
3944 task_rq_unlock(rq, p, &flags);
3946 EXPORT_SYMBOL(set_user_nice);
3949 * can_nice - check if a task can reduce its nice value
3953 int can_nice(const struct task_struct *p, const int nice)
3955 /* convert nice value [19,-20] to rlimit style value [1,40] */
3956 int nice_rlim = 20 - nice;
3958 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3959 capable(CAP_SYS_NICE));
3962 #ifdef __ARCH_WANT_SYS_NICE
3965 * sys_nice - change the priority of the current process.
3966 * @increment: priority increment
3968 * sys_setpriority is a more generic, but much slower function that
3969 * does similar things.
3971 SYSCALL_DEFINE1(nice, int, increment)
3976 * Setpriority might change our priority at the same moment.
3977 * We don't have to worry. Conceptually one call occurs first
3978 * and we have a single winner.
3980 if (increment < -40)
3985 nice = TASK_NICE(current) + increment;
3991 if (increment < 0 && !can_nice(current, nice))
3994 retval = security_task_setnice(current, nice);
3998 set_user_nice(current, nice);
4005 * task_prio - return the priority value of a given task.
4006 * @p: the task in question.
4008 * This is the priority value as seen by users in /proc.
4009 * RT tasks are offset by -200. Normal tasks are centered
4010 * around 0, value goes from -16 to +15.
4012 int task_prio(const struct task_struct *p)
4014 return p->prio - MAX_RT_PRIO;
4018 * task_nice - return the nice value of a given task.
4019 * @p: the task in question.
4021 int task_nice(const struct task_struct *p)
4023 return TASK_NICE(p);
4025 EXPORT_SYMBOL(task_nice);
4028 * idle_cpu - is a given cpu idle currently?
4029 * @cpu: the processor in question.
4031 int idle_cpu(int cpu)
4033 struct rq *rq = cpu_rq(cpu);
4035 if (rq->curr != rq->idle)
4042 if (!llist_empty(&rq->wake_list))
4050 * idle_task - return the idle task for a given cpu.
4051 * @cpu: the processor in question.
4053 struct task_struct *idle_task(int cpu)
4055 return cpu_rq(cpu)->idle;
4059 * find_process_by_pid - find a process with a matching PID value.
4060 * @pid: the pid in question.
4062 static struct task_struct *find_process_by_pid(pid_t pid)
4064 return pid ? find_task_by_vpid(pid) : current;
4067 /* Actually do priority change: must hold rq lock. */
4069 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4072 p->rt_priority = prio;
4073 p->normal_prio = normal_prio(p);
4074 /* we are holding p->pi_lock already */
4075 p->prio = rt_mutex_getprio(p);
4076 if (rt_prio(p->prio))
4077 p->sched_class = &rt_sched_class;
4079 p->sched_class = &fair_sched_class;
4084 * check the target process has a UID that matches the current process's
4086 static bool check_same_owner(struct task_struct *p)
4088 const struct cred *cred = current_cred(), *pcred;
4092 pcred = __task_cred(p);
4093 if (cred->user->user_ns == pcred->user->user_ns)
4094 match = (cred->euid == pcred->euid ||
4095 cred->euid == pcred->uid);
4102 static int __sched_setscheduler(struct task_struct *p, int policy,
4103 const struct sched_param *param, bool user)
4105 int retval, oldprio, oldpolicy = -1, on_rq, running;
4106 unsigned long flags;
4107 const struct sched_class *prev_class;
4111 /* may grab non-irq protected spin_locks */
4112 BUG_ON(in_interrupt());
4114 /* double check policy once rq lock held */
4116 reset_on_fork = p->sched_reset_on_fork;
4117 policy = oldpolicy = p->policy;
4119 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4120 policy &= ~SCHED_RESET_ON_FORK;
4122 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4123 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4124 policy != SCHED_IDLE)
4129 * Valid priorities for SCHED_FIFO and SCHED_RR are
4130 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4131 * SCHED_BATCH and SCHED_IDLE is 0.
4133 if (param->sched_priority < 0 ||
4134 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4135 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4137 if (rt_policy(policy) != (param->sched_priority != 0))
4141 * Allow unprivileged RT tasks to decrease priority:
4143 if (user && !capable(CAP_SYS_NICE)) {
4144 if (rt_policy(policy)) {
4145 unsigned long rlim_rtprio =
4146 task_rlimit(p, RLIMIT_RTPRIO);
4148 /* can't set/change the rt policy */
4149 if (policy != p->policy && !rlim_rtprio)
4152 /* can't increase priority */
4153 if (param->sched_priority > p->rt_priority &&
4154 param->sched_priority > rlim_rtprio)
4159 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4160 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4162 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4163 if (!can_nice(p, TASK_NICE(p)))
4167 /* can't change other user's priorities */
4168 if (!check_same_owner(p))
4171 /* Normal users shall not reset the sched_reset_on_fork flag */
4172 if (p->sched_reset_on_fork && !reset_on_fork)
4177 retval = security_task_setscheduler(p);
4183 * make sure no PI-waiters arrive (or leave) while we are
4184 * changing the priority of the task:
4186 * To be able to change p->policy safely, the appropriate
4187 * runqueue lock must be held.
4189 rq = task_rq_lock(p, &flags);
4192 * Changing the policy of the stop threads its a very bad idea
4194 if (p == rq->stop) {
4195 task_rq_unlock(rq, p, &flags);
4200 * If not changing anything there's no need to proceed further:
4202 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4203 param->sched_priority == p->rt_priority))) {
4205 __task_rq_unlock(rq);
4206 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4210 #ifdef CONFIG_RT_GROUP_SCHED
4213 * Do not allow realtime tasks into groups that have no runtime
4216 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4217 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4218 !task_group_is_autogroup(task_group(p))) {
4219 task_rq_unlock(rq, p, &flags);
4225 /* recheck policy now with rq lock held */
4226 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4227 policy = oldpolicy = -1;
4228 task_rq_unlock(rq, p, &flags);
4232 running = task_current(rq, p);
4234 dequeue_task(rq, p, 0);
4236 p->sched_class->put_prev_task(rq, p);
4238 p->sched_reset_on_fork = reset_on_fork;
4241 prev_class = p->sched_class;
4242 __setscheduler(rq, p, policy, param->sched_priority);
4245 p->sched_class->set_curr_task(rq);
4247 enqueue_task(rq, p, 0);
4249 check_class_changed(rq, p, prev_class, oldprio);
4250 task_rq_unlock(rq, p, &flags);
4252 rt_mutex_adjust_pi(p);
4258 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4259 * @p: the task in question.
4260 * @policy: new policy.
4261 * @param: structure containing the new RT priority.
4263 * NOTE that the task may be already dead.
4265 int sched_setscheduler(struct task_struct *p, int policy,
4266 const struct sched_param *param)
4268 return __sched_setscheduler(p, policy, param, true);
4270 EXPORT_SYMBOL_GPL(sched_setscheduler);
4273 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4274 * @p: the task in question.
4275 * @policy: new policy.
4276 * @param: structure containing the new RT priority.
4278 * Just like sched_setscheduler, only don't bother checking if the
4279 * current context has permission. For example, this is needed in
4280 * stop_machine(): we create temporary high priority worker threads,
4281 * but our caller might not have that capability.
4283 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4284 const struct sched_param *param)
4286 return __sched_setscheduler(p, policy, param, false);
4290 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4292 struct sched_param lparam;
4293 struct task_struct *p;
4296 if (!param || pid < 0)
4298 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4303 p = find_process_by_pid(pid);
4305 retval = sched_setscheduler(p, policy, &lparam);
4312 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4313 * @pid: the pid in question.
4314 * @policy: new policy.
4315 * @param: structure containing the new RT priority.
4317 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4318 struct sched_param __user *, param)
4320 /* negative values for policy are not valid */
4324 return do_sched_setscheduler(pid, policy, param);
4328 * sys_sched_setparam - set/change the RT priority of a thread
4329 * @pid: the pid in question.
4330 * @param: structure containing the new RT priority.
4332 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4334 return do_sched_setscheduler(pid, -1, param);
4338 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4339 * @pid: the pid in question.
4341 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4343 struct task_struct *p;
4351 p = find_process_by_pid(pid);
4353 retval = security_task_getscheduler(p);
4356 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4363 * sys_sched_getparam - get the RT priority of a thread
4364 * @pid: the pid in question.
4365 * @param: structure containing the RT priority.
4367 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4369 struct sched_param lp;
4370 struct task_struct *p;
4373 if (!param || pid < 0)
4377 p = find_process_by_pid(pid);
4382 retval = security_task_getscheduler(p);
4386 lp.sched_priority = p->rt_priority;
4390 * This one might sleep, we cannot do it with a spinlock held ...
4392 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4401 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4403 cpumask_var_t cpus_allowed, new_mask;
4404 struct task_struct *p;
4410 p = find_process_by_pid(pid);
4417 /* Prevent p going away */
4421 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4425 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4427 goto out_free_cpus_allowed;
4430 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4433 retval = security_task_setscheduler(p);
4437 cpuset_cpus_allowed(p, cpus_allowed);
4438 cpumask_and(new_mask, in_mask, cpus_allowed);
4440 retval = set_cpus_allowed_ptr(p, new_mask);
4443 cpuset_cpus_allowed(p, cpus_allowed);
4444 if (!cpumask_subset(new_mask, cpus_allowed)) {
4446 * We must have raced with a concurrent cpuset
4447 * update. Just reset the cpus_allowed to the
4448 * cpuset's cpus_allowed
4450 cpumask_copy(new_mask, cpus_allowed);
4455 free_cpumask_var(new_mask);
4456 out_free_cpus_allowed:
4457 free_cpumask_var(cpus_allowed);
4464 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4465 struct cpumask *new_mask)
4467 if (len < cpumask_size())
4468 cpumask_clear(new_mask);
4469 else if (len > cpumask_size())
4470 len = cpumask_size();
4472 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4476 * sys_sched_setaffinity - set the cpu affinity of a process
4477 * @pid: pid of the process
4478 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4479 * @user_mask_ptr: user-space pointer to the new cpu mask
4481 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4482 unsigned long __user *, user_mask_ptr)
4484 cpumask_var_t new_mask;
4487 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4490 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4492 retval = sched_setaffinity(pid, new_mask);
4493 free_cpumask_var(new_mask);
4497 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4499 struct task_struct *p;
4500 unsigned long flags;
4507 p = find_process_by_pid(pid);
4511 retval = security_task_getscheduler(p);
4515 raw_spin_lock_irqsave(&p->pi_lock, flags);
4516 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4517 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4527 * sys_sched_getaffinity - get the cpu affinity of a process
4528 * @pid: pid of the process
4529 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4530 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4532 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4533 unsigned long __user *, user_mask_ptr)
4538 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4540 if (len & (sizeof(unsigned long)-1))
4543 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4546 ret = sched_getaffinity(pid, mask);
4548 size_t retlen = min_t(size_t, len, cpumask_size());
4550 if (copy_to_user(user_mask_ptr, mask, retlen))
4555 free_cpumask_var(mask);
4561 * sys_sched_yield - yield the current processor to other threads.
4563 * This function yields the current CPU to other tasks. If there are no
4564 * other threads running on this CPU then this function will return.
4566 SYSCALL_DEFINE0(sched_yield)
4568 struct rq *rq = this_rq_lock();
4570 schedstat_inc(rq, yld_count);
4571 current->sched_class->yield_task(rq);
4574 * Since we are going to call schedule() anyway, there's
4575 * no need to preempt or enable interrupts:
4577 __release(rq->lock);
4578 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4579 do_raw_spin_unlock(&rq->lock);
4580 sched_preempt_enable_no_resched();
4587 static inline int should_resched(void)
4589 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4592 static void __cond_resched(void)
4594 add_preempt_count(PREEMPT_ACTIVE);
4596 sub_preempt_count(PREEMPT_ACTIVE);
4599 int __sched _cond_resched(void)
4601 if (should_resched()) {
4607 EXPORT_SYMBOL(_cond_resched);
4610 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4611 * call schedule, and on return reacquire the lock.
4613 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4614 * operations here to prevent schedule() from being called twice (once via
4615 * spin_unlock(), once by hand).
4617 int __cond_resched_lock(spinlock_t *lock)
4619 int resched = should_resched();
4622 lockdep_assert_held(lock);
4624 if (spin_needbreak(lock) || resched) {
4635 EXPORT_SYMBOL(__cond_resched_lock);
4637 int __sched __cond_resched_softirq(void)
4639 BUG_ON(!in_softirq());
4641 if (should_resched()) {
4649 EXPORT_SYMBOL(__cond_resched_softirq);
4652 * yield - yield the current processor to other threads.
4654 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4656 * The scheduler is at all times free to pick the calling task as the most
4657 * eligible task to run, if removing the yield() call from your code breaks
4658 * it, its already broken.
4660 * Typical broken usage is:
4665 * where one assumes that yield() will let 'the other' process run that will
4666 * make event true. If the current task is a SCHED_FIFO task that will never
4667 * happen. Never use yield() as a progress guarantee!!
4669 * If you want to use yield() to wait for something, use wait_event().
4670 * If you want to use yield() to be 'nice' for others, use cond_resched().
4671 * If you still want to use yield(), do not!
4673 void __sched yield(void)
4675 set_current_state(TASK_RUNNING);
4678 EXPORT_SYMBOL(yield);
4681 * yield_to - yield the current processor to another thread in
4682 * your thread group, or accelerate that thread toward the
4683 * processor it's on.
4685 * @preempt: whether task preemption is allowed or not
4687 * It's the caller's job to ensure that the target task struct
4688 * can't go away on us before we can do any checks.
4690 * Returns true if we indeed boosted the target task.
4692 bool __sched yield_to(struct task_struct *p, bool preempt)
4694 struct task_struct *curr = current;
4695 struct rq *rq, *p_rq;
4696 unsigned long flags;
4699 local_irq_save(flags);
4704 double_rq_lock(rq, p_rq);
4705 while (task_rq(p) != p_rq) {
4706 double_rq_unlock(rq, p_rq);
4710 if (!curr->sched_class->yield_to_task)
4713 if (curr->sched_class != p->sched_class)
4716 if (task_running(p_rq, p) || p->state)
4719 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4721 schedstat_inc(rq, yld_count);
4723 * Make p's CPU reschedule; pick_next_entity takes care of
4726 if (preempt && rq != p_rq)
4727 resched_task(p_rq->curr);
4730 * We might have set it in task_yield_fair(), but are
4731 * not going to schedule(), so don't want to skip
4734 rq->skip_clock_update = 0;
4738 double_rq_unlock(rq, p_rq);
4739 local_irq_restore(flags);
4746 EXPORT_SYMBOL_GPL(yield_to);
4749 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4750 * that process accounting knows that this is a task in IO wait state.
4752 void __sched io_schedule(void)
4754 struct rq *rq = raw_rq();
4756 delayacct_blkio_start();
4757 atomic_inc(&rq->nr_iowait);
4758 blk_flush_plug(current);
4759 current->in_iowait = 1;
4761 current->in_iowait = 0;
4762 atomic_dec(&rq->nr_iowait);
4763 delayacct_blkio_end();
4765 EXPORT_SYMBOL(io_schedule);
4767 long __sched io_schedule_timeout(long timeout)
4769 struct rq *rq = raw_rq();
4772 delayacct_blkio_start();
4773 atomic_inc(&rq->nr_iowait);
4774 blk_flush_plug(current);
4775 current->in_iowait = 1;
4776 ret = schedule_timeout(timeout);
4777 current->in_iowait = 0;
4778 atomic_dec(&rq->nr_iowait);
4779 delayacct_blkio_end();
4784 * sys_sched_get_priority_max - return maximum RT priority.
4785 * @policy: scheduling class.
4787 * this syscall returns the maximum rt_priority that can be used
4788 * by a given scheduling class.
4790 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4797 ret = MAX_USER_RT_PRIO-1;
4809 * sys_sched_get_priority_min - return minimum RT priority.
4810 * @policy: scheduling class.
4812 * this syscall returns the minimum rt_priority that can be used
4813 * by a given scheduling class.
4815 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4833 * sys_sched_rr_get_interval - return the default timeslice of a process.
4834 * @pid: pid of the process.
4835 * @interval: userspace pointer to the timeslice value.
4837 * this syscall writes the default timeslice value of a given process
4838 * into the user-space timespec buffer. A value of '0' means infinity.
4840 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4841 struct timespec __user *, interval)
4843 struct task_struct *p;
4844 unsigned int time_slice;
4845 unsigned long flags;
4855 p = find_process_by_pid(pid);
4859 retval = security_task_getscheduler(p);
4863 rq = task_rq_lock(p, &flags);
4864 time_slice = p->sched_class->get_rr_interval(rq, p);
4865 task_rq_unlock(rq, p, &flags);
4868 jiffies_to_timespec(time_slice, &t);
4869 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4877 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4879 void sched_show_task(struct task_struct *p)
4881 unsigned long free = 0;
4884 state = p->state ? __ffs(p->state) + 1 : 0;
4885 printk(KERN_INFO "%-15.15s %c", p->comm,
4886 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4887 #if BITS_PER_LONG == 32
4888 if (state == TASK_RUNNING)
4889 printk(KERN_CONT " running ");
4891 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4893 if (state == TASK_RUNNING)
4894 printk(KERN_CONT " running task ");
4896 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4898 #ifdef CONFIG_DEBUG_STACK_USAGE
4899 free = stack_not_used(p);
4901 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4902 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4903 (unsigned long)task_thread_info(p)->flags);
4905 show_stack(p, NULL);
4908 void show_state_filter(unsigned long state_filter)
4910 struct task_struct *g, *p;
4912 #if BITS_PER_LONG == 32
4914 " task PC stack pid father\n");
4917 " task PC stack pid father\n");
4920 do_each_thread(g, p) {
4922 * reset the NMI-timeout, listing all files on a slow
4923 * console might take a lot of time:
4925 touch_nmi_watchdog();
4926 if (!state_filter || (p->state & state_filter))
4928 } while_each_thread(g, p);
4930 touch_all_softlockup_watchdogs();
4932 #ifdef CONFIG_SCHED_DEBUG
4933 sysrq_sched_debug_show();
4937 * Only show locks if all tasks are dumped:
4940 debug_show_all_locks();
4943 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4945 idle->sched_class = &idle_sched_class;
4949 * init_idle - set up an idle thread for a given CPU
4950 * @idle: task in question
4951 * @cpu: cpu the idle task belongs to
4953 * NOTE: this function does not set the idle thread's NEED_RESCHED
4954 * flag, to make booting more robust.
4956 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4958 struct rq *rq = cpu_rq(cpu);
4959 unsigned long flags;
4961 raw_spin_lock_irqsave(&rq->lock, flags);
4964 idle->state = TASK_RUNNING;
4965 idle->se.exec_start = sched_clock();
4967 do_set_cpus_allowed(idle, cpumask_of(cpu));
4969 * We're having a chicken and egg problem, even though we are
4970 * holding rq->lock, the cpu isn't yet set to this cpu so the
4971 * lockdep check in task_group() will fail.
4973 * Similar case to sched_fork(). / Alternatively we could
4974 * use task_rq_lock() here and obtain the other rq->lock.
4979 __set_task_cpu(idle, cpu);
4982 rq->curr = rq->idle = idle;
4983 #if defined(CONFIG_SMP)
4986 raw_spin_unlock_irqrestore(&rq->lock, flags);
4988 /* Set the preempt count _outside_ the spinlocks! */
4989 task_thread_info(idle)->preempt_count = 0;
4992 * The idle tasks have their own, simple scheduling class:
4994 idle->sched_class = &idle_sched_class;
4995 ftrace_graph_init_idle_task(idle, cpu);
4996 #if defined(CONFIG_SMP)
4997 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5002 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5004 if (p->sched_class && p->sched_class->set_cpus_allowed)
5005 p->sched_class->set_cpus_allowed(p, new_mask);
5007 cpumask_copy(&p->cpus_allowed, new_mask);
5008 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5012 * This is how migration works:
5014 * 1) we invoke migration_cpu_stop() on the target CPU using
5016 * 2) stopper starts to run (implicitly forcing the migrated thread
5018 * 3) it checks whether the migrated task is still in the wrong runqueue.
5019 * 4) if it's in the wrong runqueue then the migration thread removes
5020 * it and puts it into the right queue.
5021 * 5) stopper completes and stop_one_cpu() returns and the migration
5026 * Change a given task's CPU affinity. Migrate the thread to a
5027 * proper CPU and schedule it away if the CPU it's executing on
5028 * is removed from the allowed bitmask.
5030 * NOTE: the caller must have a valid reference to the task, the
5031 * task must not exit() & deallocate itself prematurely. The
5032 * call is not atomic; no spinlocks may be held.
5034 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5036 unsigned long flags;
5038 unsigned int dest_cpu;
5041 rq = task_rq_lock(p, &flags);
5043 if (cpumask_equal(&p->cpus_allowed, new_mask))
5046 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5051 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5056 do_set_cpus_allowed(p, new_mask);
5058 /* Can the task run on the task's current CPU? If so, we're done */
5059 if (cpumask_test_cpu(task_cpu(p), new_mask))
5062 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5064 struct migration_arg arg = { p, dest_cpu };
5065 /* Need help from migration thread: drop lock and wait. */
5066 task_rq_unlock(rq, p, &flags);
5067 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5068 tlb_migrate_finish(p->mm);
5072 task_rq_unlock(rq, p, &flags);
5076 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5079 * Move (not current) task off this cpu, onto dest cpu. We're doing
5080 * this because either it can't run here any more (set_cpus_allowed()
5081 * away from this CPU, or CPU going down), or because we're
5082 * attempting to rebalance this task on exec (sched_exec).
5084 * So we race with normal scheduler movements, but that's OK, as long
5085 * as the task is no longer on this CPU.
5087 * Returns non-zero if task was successfully migrated.
5089 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5091 struct rq *rq_dest, *rq_src;
5094 if (unlikely(!cpu_active(dest_cpu)))
5097 rq_src = cpu_rq(src_cpu);
5098 rq_dest = cpu_rq(dest_cpu);
5100 raw_spin_lock(&p->pi_lock);
5101 double_rq_lock(rq_src, rq_dest);
5102 /* Already moved. */
5103 if (task_cpu(p) != src_cpu)
5105 /* Affinity changed (again). */
5106 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5110 * If we're not on a rq, the next wake-up will ensure we're
5114 dequeue_task(rq_src, p, 0);
5115 set_task_cpu(p, dest_cpu);
5116 enqueue_task(rq_dest, p, 0);
5117 check_preempt_curr(rq_dest, p, 0);
5122 double_rq_unlock(rq_src, rq_dest);
5123 raw_spin_unlock(&p->pi_lock);
5128 * migration_cpu_stop - this will be executed by a highprio stopper thread
5129 * and performs thread migration by bumping thread off CPU then
5130 * 'pushing' onto another runqueue.
5132 static int migration_cpu_stop(void *data)
5134 struct migration_arg *arg = data;
5137 * The original target cpu might have gone down and we might
5138 * be on another cpu but it doesn't matter.
5140 local_irq_disable();
5141 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5146 #ifdef CONFIG_HOTPLUG_CPU
5149 * Ensures that the idle task is using init_mm right before its cpu goes
5152 void idle_task_exit(void)
5154 struct mm_struct *mm = current->active_mm;
5156 BUG_ON(cpu_online(smp_processor_id()));
5159 switch_mm(mm, &init_mm, current);
5164 * While a dead CPU has no uninterruptible tasks queued at this point,
5165 * it might still have a nonzero ->nr_uninterruptible counter, because
5166 * for performance reasons the counter is not stricly tracking tasks to
5167 * their home CPUs. So we just add the counter to another CPU's counter,
5168 * to keep the global sum constant after CPU-down:
5170 static void migrate_nr_uninterruptible(struct rq *rq_src)
5172 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5174 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5175 rq_src->nr_uninterruptible = 0;
5179 * remove the tasks which were accounted by rq from calc_load_tasks.
5181 static void calc_global_load_remove(struct rq *rq)
5183 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5184 rq->calc_load_active = 0;
5188 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5189 * try_to_wake_up()->select_task_rq().
5191 * Called with rq->lock held even though we'er in stop_machine() and
5192 * there's no concurrency possible, we hold the required locks anyway
5193 * because of lock validation efforts.
5195 static void migrate_tasks(unsigned int dead_cpu)
5197 struct rq *rq = cpu_rq(dead_cpu);
5198 struct task_struct *next, *stop = rq->stop;
5202 * Fudge the rq selection such that the below task selection loop
5203 * doesn't get stuck on the currently eligible stop task.
5205 * We're currently inside stop_machine() and the rq is either stuck
5206 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5207 * either way we should never end up calling schedule() until we're
5212 /* Ensure any throttled groups are reachable by pick_next_task */
5213 unthrottle_offline_cfs_rqs(rq);
5217 * There's this thread running, bail when that's the only
5220 if (rq->nr_running == 1)
5223 next = pick_next_task(rq);
5225 next->sched_class->put_prev_task(rq, next);
5227 /* Find suitable destination for @next, with force if needed. */
5228 dest_cpu = select_fallback_rq(dead_cpu, next);
5229 raw_spin_unlock(&rq->lock);
5231 __migrate_task(next, dead_cpu, dest_cpu);
5233 raw_spin_lock(&rq->lock);
5239 #endif /* CONFIG_HOTPLUG_CPU */
5241 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5243 static struct ctl_table sd_ctl_dir[] = {
5245 .procname = "sched_domain",
5251 static struct ctl_table sd_ctl_root[] = {
5253 .procname = "kernel",
5255 .child = sd_ctl_dir,
5260 static struct ctl_table *sd_alloc_ctl_entry(int n)
5262 struct ctl_table *entry =
5263 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5268 static void sd_free_ctl_entry(struct ctl_table **tablep)
5270 struct ctl_table *entry;
5273 * In the intermediate directories, both the child directory and
5274 * procname are dynamically allocated and could fail but the mode
5275 * will always be set. In the lowest directory the names are
5276 * static strings and all have proc handlers.
5278 for (entry = *tablep; entry->mode; entry++) {
5280 sd_free_ctl_entry(&entry->child);
5281 if (entry->proc_handler == NULL)
5282 kfree(entry->procname);
5290 set_table_entry(struct ctl_table *entry,
5291 const char *procname, void *data, int maxlen,
5292 umode_t mode, proc_handler *proc_handler)
5294 entry->procname = procname;
5296 entry->maxlen = maxlen;
5298 entry->proc_handler = proc_handler;
5301 static struct ctl_table *
5302 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5304 struct ctl_table *table = sd_alloc_ctl_entry(13);
5309 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5310 sizeof(long), 0644, proc_doulongvec_minmax);
5311 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5312 sizeof(long), 0644, proc_doulongvec_minmax);
5313 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5314 sizeof(int), 0644, proc_dointvec_minmax);
5315 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5316 sizeof(int), 0644, proc_dointvec_minmax);
5317 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5318 sizeof(int), 0644, proc_dointvec_minmax);
5319 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5320 sizeof(int), 0644, proc_dointvec_minmax);
5321 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5322 sizeof(int), 0644, proc_dointvec_minmax);
5323 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5324 sizeof(int), 0644, proc_dointvec_minmax);
5325 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5326 sizeof(int), 0644, proc_dointvec_minmax);
5327 set_table_entry(&table[9], "cache_nice_tries",
5328 &sd->cache_nice_tries,
5329 sizeof(int), 0644, proc_dointvec_minmax);
5330 set_table_entry(&table[10], "flags", &sd->flags,
5331 sizeof(int), 0644, proc_dointvec_minmax);
5332 set_table_entry(&table[11], "name", sd->name,
5333 CORENAME_MAX_SIZE, 0444, proc_dostring);
5334 /* &table[12] is terminator */
5339 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5341 struct ctl_table *entry, *table;
5342 struct sched_domain *sd;
5343 int domain_num = 0, i;
5346 for_each_domain(cpu, sd)
5348 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5353 for_each_domain(cpu, sd) {
5354 snprintf(buf, 32, "domain%d", i);
5355 entry->procname = kstrdup(buf, GFP_KERNEL);
5357 entry->child = sd_alloc_ctl_domain_table(sd);
5364 static struct ctl_table_header *sd_sysctl_header;
5365 static void register_sched_domain_sysctl(void)
5367 int i, cpu_num = num_possible_cpus();
5368 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5371 WARN_ON(sd_ctl_dir[0].child);
5372 sd_ctl_dir[0].child = entry;
5377 for_each_possible_cpu(i) {
5378 snprintf(buf, 32, "cpu%d", i);
5379 entry->procname = kstrdup(buf, GFP_KERNEL);
5381 entry->child = sd_alloc_ctl_cpu_table(i);
5385 WARN_ON(sd_sysctl_header);
5386 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5389 /* may be called multiple times per register */
5390 static void unregister_sched_domain_sysctl(void)
5392 if (sd_sysctl_header)
5393 unregister_sysctl_table(sd_sysctl_header);
5394 sd_sysctl_header = NULL;
5395 if (sd_ctl_dir[0].child)
5396 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5399 static void register_sched_domain_sysctl(void)
5402 static void unregister_sched_domain_sysctl(void)
5407 static void set_rq_online(struct rq *rq)
5410 const struct sched_class *class;
5412 cpumask_set_cpu(rq->cpu, rq->rd->online);
5415 for_each_class(class) {
5416 if (class->rq_online)
5417 class->rq_online(rq);
5422 static void set_rq_offline(struct rq *rq)
5425 const struct sched_class *class;
5427 for_each_class(class) {
5428 if (class->rq_offline)
5429 class->rq_offline(rq);
5432 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5438 * migration_call - callback that gets triggered when a CPU is added.
5439 * Here we can start up the necessary migration thread for the new CPU.
5441 static int __cpuinit
5442 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5444 int cpu = (long)hcpu;
5445 unsigned long flags;
5446 struct rq *rq = cpu_rq(cpu);
5448 switch (action & ~CPU_TASKS_FROZEN) {
5450 case CPU_UP_PREPARE:
5451 rq->calc_load_update = calc_load_update;
5455 /* Update our root-domain */
5456 raw_spin_lock_irqsave(&rq->lock, flags);
5458 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5462 raw_spin_unlock_irqrestore(&rq->lock, flags);
5465 #ifdef CONFIG_HOTPLUG_CPU
5467 sched_ttwu_pending();
5468 /* Update our root-domain */
5469 raw_spin_lock_irqsave(&rq->lock, flags);
5471 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5475 BUG_ON(rq->nr_running != 1); /* the migration thread */
5476 raw_spin_unlock_irqrestore(&rq->lock, flags);
5478 migrate_nr_uninterruptible(rq);
5479 calc_global_load_remove(rq);
5484 update_max_interval();
5490 * Register at high priority so that task migration (migrate_all_tasks)
5491 * happens before everything else. This has to be lower priority than
5492 * the notifier in the perf_event subsystem, though.
5494 static struct notifier_block __cpuinitdata migration_notifier = {
5495 .notifier_call = migration_call,
5496 .priority = CPU_PRI_MIGRATION,
5499 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5500 unsigned long action, void *hcpu)
5502 switch (action & ~CPU_TASKS_FROZEN) {
5504 case CPU_DOWN_FAILED:
5505 set_cpu_active((long)hcpu, true);
5512 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5513 unsigned long action, void *hcpu)
5515 switch (action & ~CPU_TASKS_FROZEN) {
5516 case CPU_DOWN_PREPARE:
5517 set_cpu_active((long)hcpu, false);
5524 static int __init migration_init(void)
5526 void *cpu = (void *)(long)smp_processor_id();
5529 /* Initialize migration for the boot CPU */
5530 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5531 BUG_ON(err == NOTIFY_BAD);
5532 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5533 register_cpu_notifier(&migration_notifier);
5535 /* Register cpu active notifiers */
5536 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5537 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5541 early_initcall(migration_init);
5546 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5548 #ifdef CONFIG_SCHED_DEBUG
5550 static __read_mostly int sched_domain_debug_enabled;
5552 static int __init sched_domain_debug_setup(char *str)
5554 sched_domain_debug_enabled = 1;
5558 early_param("sched_debug", sched_domain_debug_setup);
5560 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5561 struct cpumask *groupmask)
5563 struct sched_group *group = sd->groups;
5566 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5567 cpumask_clear(groupmask);
5569 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5571 if (!(sd->flags & SD_LOAD_BALANCE)) {
5572 printk("does not load-balance\n");
5574 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5579 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5581 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5582 printk(KERN_ERR "ERROR: domain->span does not contain "
5585 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5586 printk(KERN_ERR "ERROR: domain->groups does not contain"
5590 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5594 printk(KERN_ERR "ERROR: group is NULL\n");
5598 if (!group->sgp->power) {
5599 printk(KERN_CONT "\n");
5600 printk(KERN_ERR "ERROR: domain->cpu_power not "
5605 if (!cpumask_weight(sched_group_cpus(group))) {
5606 printk(KERN_CONT "\n");
5607 printk(KERN_ERR "ERROR: empty group\n");
5611 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
5612 printk(KERN_CONT "\n");
5613 printk(KERN_ERR "ERROR: repeated CPUs\n");
5617 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5619 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5621 printk(KERN_CONT " %s", str);
5622 if (group->sgp->power != SCHED_POWER_SCALE) {
5623 printk(KERN_CONT " (cpu_power = %d)",
5627 group = group->next;
5628 } while (group != sd->groups);
5629 printk(KERN_CONT "\n");
5631 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5632 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5635 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5636 printk(KERN_ERR "ERROR: parent span is not a superset "
5637 "of domain->span\n");
5641 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5645 if (!sched_domain_debug_enabled)
5649 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5653 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5656 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5664 #else /* !CONFIG_SCHED_DEBUG */
5665 # define sched_domain_debug(sd, cpu) do { } while (0)
5666 #endif /* CONFIG_SCHED_DEBUG */
5668 static int sd_degenerate(struct sched_domain *sd)
5670 if (cpumask_weight(sched_domain_span(sd)) == 1)
5673 /* Following flags need at least 2 groups */
5674 if (sd->flags & (SD_LOAD_BALANCE |
5675 SD_BALANCE_NEWIDLE |
5679 SD_SHARE_PKG_RESOURCES)) {
5680 if (sd->groups != sd->groups->next)
5684 /* Following flags don't use groups */
5685 if (sd->flags & (SD_WAKE_AFFINE))
5692 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5694 unsigned long cflags = sd->flags, pflags = parent->flags;
5696 if (sd_degenerate(parent))
5699 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5702 /* Flags needing groups don't count if only 1 group in parent */
5703 if (parent->groups == parent->groups->next) {
5704 pflags &= ~(SD_LOAD_BALANCE |
5705 SD_BALANCE_NEWIDLE |
5709 SD_SHARE_PKG_RESOURCES);
5710 if (nr_node_ids == 1)
5711 pflags &= ~SD_SERIALIZE;
5713 if (~cflags & pflags)
5719 static void free_rootdomain(struct rcu_head *rcu)
5721 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5723 cpupri_cleanup(&rd->cpupri);
5724 free_cpumask_var(rd->rto_mask);
5725 free_cpumask_var(rd->online);
5726 free_cpumask_var(rd->span);
5730 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5732 struct root_domain *old_rd = NULL;
5733 unsigned long flags;
5735 raw_spin_lock_irqsave(&rq->lock, flags);
5740 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5743 cpumask_clear_cpu(rq->cpu, old_rd->span);
5746 * If we dont want to free the old_rt yet then
5747 * set old_rd to NULL to skip the freeing later
5750 if (!atomic_dec_and_test(&old_rd->refcount))
5754 atomic_inc(&rd->refcount);
5757 cpumask_set_cpu(rq->cpu, rd->span);
5758 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5761 raw_spin_unlock_irqrestore(&rq->lock, flags);
5764 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5767 static int init_rootdomain(struct root_domain *rd)
5769 memset(rd, 0, sizeof(*rd));
5771 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5773 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5775 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5778 if (cpupri_init(&rd->cpupri) != 0)
5783 free_cpumask_var(rd->rto_mask);
5785 free_cpumask_var(rd->online);
5787 free_cpumask_var(rd->span);
5793 * By default the system creates a single root-domain with all cpus as
5794 * members (mimicking the global state we have today).
5796 struct root_domain def_root_domain;
5798 static void init_defrootdomain(void)
5800 init_rootdomain(&def_root_domain);
5802 atomic_set(&def_root_domain.refcount, 1);
5805 static struct root_domain *alloc_rootdomain(void)
5807 struct root_domain *rd;
5809 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5813 if (init_rootdomain(rd) != 0) {
5821 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5823 struct sched_group *tmp, *first;
5832 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5837 } while (sg != first);
5840 static void free_sched_domain(struct rcu_head *rcu)
5842 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5845 * If its an overlapping domain it has private groups, iterate and
5848 if (sd->flags & SD_OVERLAP) {
5849 free_sched_groups(sd->groups, 1);
5850 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5851 kfree(sd->groups->sgp);
5857 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5859 call_rcu(&sd->rcu, free_sched_domain);
5862 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5864 for (; sd; sd = sd->parent)
5865 destroy_sched_domain(sd, cpu);
5869 * Keep a special pointer to the highest sched_domain that has
5870 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5871 * allows us to avoid some pointer chasing select_idle_sibling().
5873 * Also keep a unique ID per domain (we use the first cpu number in
5874 * the cpumask of the domain), this allows us to quickly tell if
5875 * two cpus are in the same cache domain, see cpus_share_cache().
5877 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5878 DEFINE_PER_CPU(int, sd_llc_id);
5880 static void update_top_cache_domain(int cpu)
5882 struct sched_domain *sd;
5885 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5887 id = cpumask_first(sched_domain_span(sd));
5889 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5890 per_cpu(sd_llc_id, cpu) = id;
5894 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5895 * hold the hotplug lock.
5898 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5900 struct rq *rq = cpu_rq(cpu);
5901 struct sched_domain *tmp;
5903 /* Remove the sched domains which do not contribute to scheduling. */
5904 for (tmp = sd; tmp; ) {
5905 struct sched_domain *parent = tmp->parent;
5909 if (sd_parent_degenerate(tmp, parent)) {
5910 tmp->parent = parent->parent;
5912 parent->parent->child = tmp;
5913 destroy_sched_domain(parent, cpu);
5918 if (sd && sd_degenerate(sd)) {
5921 destroy_sched_domain(tmp, cpu);
5926 sched_domain_debug(sd, cpu);
5928 rq_attach_root(rq, rd);
5930 rcu_assign_pointer(rq->sd, sd);
5931 destroy_sched_domains(tmp, cpu);
5933 update_top_cache_domain(cpu);
5936 /* cpus with isolated domains */
5937 static cpumask_var_t cpu_isolated_map;
5939 /* Setup the mask of cpus configured for isolated domains */
5940 static int __init isolated_cpu_setup(char *str)
5942 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5943 cpulist_parse(str, cpu_isolated_map);
5947 __setup("isolcpus=", isolated_cpu_setup);
5952 * find_next_best_node - find the next node to include in a sched_domain
5953 * @node: node whose sched_domain we're building
5954 * @used_nodes: nodes already in the sched_domain
5956 * Find the next node to include in a given scheduling domain. Simply
5957 * finds the closest node not already in the @used_nodes map.
5959 * Should use nodemask_t.
5961 static int find_next_best_node(int node, nodemask_t *used_nodes)
5963 int i, n, val, min_val, best_node = -1;
5967 for (i = 0; i < nr_node_ids; i++) {
5968 /* Start at @node */
5969 n = (node + i) % nr_node_ids;
5971 if (!nr_cpus_node(n))
5974 /* Skip already used nodes */
5975 if (node_isset(n, *used_nodes))
5978 /* Simple min distance search */
5979 val = node_distance(node, n);
5981 if (val < min_val) {
5987 if (best_node != -1)
5988 node_set(best_node, *used_nodes);
5993 * sched_domain_node_span - get a cpumask for a node's sched_domain
5994 * @node: node whose cpumask we're constructing
5995 * @span: resulting cpumask
5997 * Given a node, construct a good cpumask for its sched_domain to span. It
5998 * should be one that prevents unnecessary balancing, but also spreads tasks
6001 static void sched_domain_node_span(int node, struct cpumask *span)
6003 nodemask_t used_nodes;
6006 cpumask_clear(span);
6007 nodes_clear(used_nodes);
6009 cpumask_or(span, span, cpumask_of_node(node));
6010 node_set(node, used_nodes);
6012 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6013 int next_node = find_next_best_node(node, &used_nodes);
6016 cpumask_or(span, span, cpumask_of_node(next_node));
6020 static const struct cpumask *cpu_node_mask(int cpu)
6022 lockdep_assert_held(&sched_domains_mutex);
6024 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
6026 return sched_domains_tmpmask;
6029 static const struct cpumask *cpu_allnodes_mask(int cpu)
6031 return cpu_possible_mask;
6033 #endif /* CONFIG_NUMA */
6035 static const struct cpumask *cpu_cpu_mask(int cpu)
6037 return cpumask_of_node(cpu_to_node(cpu));
6040 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6043 struct sched_domain **__percpu sd;
6044 struct sched_group **__percpu sg;
6045 struct sched_group_power **__percpu sgp;
6049 struct sched_domain ** __percpu sd;
6050 struct root_domain *rd;
6060 struct sched_domain_topology_level;
6062 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6063 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6065 #define SDTL_OVERLAP 0x01
6067 struct sched_domain_topology_level {
6068 sched_domain_init_f init;
6069 sched_domain_mask_f mask;
6071 struct sd_data data;
6075 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6077 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6078 const struct cpumask *span = sched_domain_span(sd);
6079 struct cpumask *covered = sched_domains_tmpmask;
6080 struct sd_data *sdd = sd->private;
6081 struct sched_domain *child;
6084 cpumask_clear(covered);
6086 for_each_cpu(i, span) {
6087 struct cpumask *sg_span;
6089 if (cpumask_test_cpu(i, covered))
6092 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6093 GFP_KERNEL, cpu_to_node(cpu));
6098 sg_span = sched_group_cpus(sg);
6100 child = *per_cpu_ptr(sdd->sd, i);
6102 child = child->child;
6103 cpumask_copy(sg_span, sched_domain_span(child));
6105 cpumask_set_cpu(i, sg_span);
6107 cpumask_or(covered, covered, sg_span);
6109 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
6110 atomic_inc(&sg->sgp->ref);
6112 if (cpumask_test_cpu(cpu, sg_span))
6122 sd->groups = groups;
6127 free_sched_groups(first, 0);
6132 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6134 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6135 struct sched_domain *child = sd->child;
6138 cpu = cpumask_first(sched_domain_span(child));
6141 *sg = *per_cpu_ptr(sdd->sg, cpu);
6142 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6143 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6150 * build_sched_groups will build a circular linked list of the groups
6151 * covered by the given span, and will set each group's ->cpumask correctly,
6152 * and ->cpu_power to 0.
6154 * Assumes the sched_domain tree is fully constructed
6157 build_sched_groups(struct sched_domain *sd, int cpu)
6159 struct sched_group *first = NULL, *last = NULL;
6160 struct sd_data *sdd = sd->private;
6161 const struct cpumask *span = sched_domain_span(sd);
6162 struct cpumask *covered;
6165 get_group(cpu, sdd, &sd->groups);
6166 atomic_inc(&sd->groups->ref);
6168 if (cpu != cpumask_first(sched_domain_span(sd)))
6171 lockdep_assert_held(&sched_domains_mutex);
6172 covered = sched_domains_tmpmask;
6174 cpumask_clear(covered);
6176 for_each_cpu(i, span) {
6177 struct sched_group *sg;
6178 int group = get_group(i, sdd, &sg);
6181 if (cpumask_test_cpu(i, covered))
6184 cpumask_clear(sched_group_cpus(sg));
6187 for_each_cpu(j, span) {
6188 if (get_group(j, sdd, NULL) != group)
6191 cpumask_set_cpu(j, covered);
6192 cpumask_set_cpu(j, sched_group_cpus(sg));
6207 * Initialize sched groups cpu_power.
6209 * cpu_power indicates the capacity of sched group, which is used while
6210 * distributing the load between different sched groups in a sched domain.
6211 * Typically cpu_power for all the groups in a sched domain will be same unless
6212 * there are asymmetries in the topology. If there are asymmetries, group
6213 * having more cpu_power will pickup more load compared to the group having
6216 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6218 struct sched_group *sg = sd->groups;
6220 WARN_ON(!sd || !sg);
6223 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6225 } while (sg != sd->groups);
6227 if (cpu != group_first_cpu(sg))
6230 update_group_power(sd, cpu);
6231 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6234 int __weak arch_sd_sibling_asym_packing(void)
6236 return 0*SD_ASYM_PACKING;
6240 * Initializers for schedule domains
6241 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6244 #ifdef CONFIG_SCHED_DEBUG
6245 # define SD_INIT_NAME(sd, type) sd->name = #type
6247 # define SD_INIT_NAME(sd, type) do { } while (0)
6250 #define SD_INIT_FUNC(type) \
6251 static noinline struct sched_domain * \
6252 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6254 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6255 *sd = SD_##type##_INIT; \
6256 SD_INIT_NAME(sd, type); \
6257 sd->private = &tl->data; \
6263 SD_INIT_FUNC(ALLNODES)
6266 #ifdef CONFIG_SCHED_SMT
6267 SD_INIT_FUNC(SIBLING)
6269 #ifdef CONFIG_SCHED_MC
6272 #ifdef CONFIG_SCHED_BOOK
6276 static int default_relax_domain_level = -1;
6277 int sched_domain_level_max;
6279 static int __init setup_relax_domain_level(char *str)
6283 val = simple_strtoul(str, NULL, 0);
6284 if (val < sched_domain_level_max)
6285 default_relax_domain_level = val;
6289 __setup("relax_domain_level=", setup_relax_domain_level);
6291 static void set_domain_attribute(struct sched_domain *sd,
6292 struct sched_domain_attr *attr)
6296 if (!attr || attr->relax_domain_level < 0) {
6297 if (default_relax_domain_level < 0)
6300 request = default_relax_domain_level;
6302 request = attr->relax_domain_level;
6303 if (request < sd->level) {
6304 /* turn off idle balance on this domain */
6305 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6307 /* turn on idle balance on this domain */
6308 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6312 static void __sdt_free(const struct cpumask *cpu_map);
6313 static int __sdt_alloc(const struct cpumask *cpu_map);
6315 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6316 const struct cpumask *cpu_map)
6320 if (!atomic_read(&d->rd->refcount))
6321 free_rootdomain(&d->rd->rcu); /* fall through */
6323 free_percpu(d->sd); /* fall through */
6325 __sdt_free(cpu_map); /* fall through */
6331 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6332 const struct cpumask *cpu_map)
6334 memset(d, 0, sizeof(*d));
6336 if (__sdt_alloc(cpu_map))
6337 return sa_sd_storage;
6338 d->sd = alloc_percpu(struct sched_domain *);
6340 return sa_sd_storage;
6341 d->rd = alloc_rootdomain();
6344 return sa_rootdomain;
6348 * NULL the sd_data elements we've used to build the sched_domain and
6349 * sched_group structure so that the subsequent __free_domain_allocs()
6350 * will not free the data we're using.
6352 static void claim_allocations(int cpu, struct sched_domain *sd)
6354 struct sd_data *sdd = sd->private;
6356 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6357 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6359 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6360 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6362 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6363 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6366 #ifdef CONFIG_SCHED_SMT
6367 static const struct cpumask *cpu_smt_mask(int cpu)
6369 return topology_thread_cpumask(cpu);
6374 * Topology list, bottom-up.
6376 static struct sched_domain_topology_level default_topology[] = {
6377 #ifdef CONFIG_SCHED_SMT
6378 { sd_init_SIBLING, cpu_smt_mask, },
6380 #ifdef CONFIG_SCHED_MC
6381 { sd_init_MC, cpu_coregroup_mask, },
6383 #ifdef CONFIG_SCHED_BOOK
6384 { sd_init_BOOK, cpu_book_mask, },
6386 { sd_init_CPU, cpu_cpu_mask, },
6388 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
6389 { sd_init_ALLNODES, cpu_allnodes_mask, },
6394 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6396 static int __sdt_alloc(const struct cpumask *cpu_map)
6398 struct sched_domain_topology_level *tl;
6401 for (tl = sched_domain_topology; tl->init; tl++) {
6402 struct sd_data *sdd = &tl->data;
6404 sdd->sd = alloc_percpu(struct sched_domain *);
6408 sdd->sg = alloc_percpu(struct sched_group *);
6412 sdd->sgp = alloc_percpu(struct sched_group_power *);
6416 for_each_cpu(j, cpu_map) {
6417 struct sched_domain *sd;
6418 struct sched_group *sg;
6419 struct sched_group_power *sgp;
6421 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6422 GFP_KERNEL, cpu_to_node(j));
6426 *per_cpu_ptr(sdd->sd, j) = sd;
6428 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6429 GFP_KERNEL, cpu_to_node(j));
6435 *per_cpu_ptr(sdd->sg, j) = sg;
6437 sgp = kzalloc_node(sizeof(struct sched_group_power),
6438 GFP_KERNEL, cpu_to_node(j));
6442 *per_cpu_ptr(sdd->sgp, j) = sgp;
6449 static void __sdt_free(const struct cpumask *cpu_map)
6451 struct sched_domain_topology_level *tl;
6454 for (tl = sched_domain_topology; tl->init; tl++) {
6455 struct sd_data *sdd = &tl->data;
6457 for_each_cpu(j, cpu_map) {
6458 struct sched_domain *sd;
6461 sd = *per_cpu_ptr(sdd->sd, j);
6462 if (sd && (sd->flags & SD_OVERLAP))
6463 free_sched_groups(sd->groups, 0);
6464 kfree(*per_cpu_ptr(sdd->sd, j));
6468 kfree(*per_cpu_ptr(sdd->sg, j));
6470 kfree(*per_cpu_ptr(sdd->sgp, j));
6472 free_percpu(sdd->sd);
6474 free_percpu(sdd->sg);
6476 free_percpu(sdd->sgp);
6481 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6482 struct s_data *d, const struct cpumask *cpu_map,
6483 struct sched_domain_attr *attr, struct sched_domain *child,
6486 struct sched_domain *sd = tl->init(tl, cpu);
6490 set_domain_attribute(sd, attr);
6491 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6493 sd->level = child->level + 1;
6494 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6503 * Build sched domains for a given set of cpus and attach the sched domains
6504 * to the individual cpus
6506 static int build_sched_domains(const struct cpumask *cpu_map,
6507 struct sched_domain_attr *attr)
6509 enum s_alloc alloc_state = sa_none;
6510 struct sched_domain *sd;
6512 int i, ret = -ENOMEM;
6514 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6515 if (alloc_state != sa_rootdomain)
6518 /* Set up domains for cpus specified by the cpu_map. */
6519 for_each_cpu(i, cpu_map) {
6520 struct sched_domain_topology_level *tl;
6523 for (tl = sched_domain_topology; tl->init; tl++) {
6524 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6525 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6526 sd->flags |= SD_OVERLAP;
6527 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6534 *per_cpu_ptr(d.sd, i) = sd;
6537 /* Build the groups for the domains */
6538 for_each_cpu(i, cpu_map) {
6539 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6540 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6541 if (sd->flags & SD_OVERLAP) {
6542 if (build_overlap_sched_groups(sd, i))
6545 if (build_sched_groups(sd, i))
6551 /* Calculate CPU power for physical packages and nodes */
6552 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6553 if (!cpumask_test_cpu(i, cpu_map))
6556 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6557 claim_allocations(i, sd);
6558 init_sched_groups_power(i, sd);
6562 /* Attach the domains */
6564 for_each_cpu(i, cpu_map) {
6565 sd = *per_cpu_ptr(d.sd, i);
6566 cpu_attach_domain(sd, d.rd, i);
6572 __free_domain_allocs(&d, alloc_state, cpu_map);
6576 static cpumask_var_t *doms_cur; /* current sched domains */
6577 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6578 static struct sched_domain_attr *dattr_cur;
6579 /* attribues of custom domains in 'doms_cur' */
6582 * Special case: If a kmalloc of a doms_cur partition (array of
6583 * cpumask) fails, then fallback to a single sched domain,
6584 * as determined by the single cpumask fallback_doms.
6586 static cpumask_var_t fallback_doms;
6589 * arch_update_cpu_topology lets virtualized architectures update the
6590 * cpu core maps. It is supposed to return 1 if the topology changed
6591 * or 0 if it stayed the same.
6593 int __attribute__((weak)) arch_update_cpu_topology(void)
6598 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6601 cpumask_var_t *doms;
6603 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6606 for (i = 0; i < ndoms; i++) {
6607 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6608 free_sched_domains(doms, i);
6615 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6618 for (i = 0; i < ndoms; i++)
6619 free_cpumask_var(doms[i]);
6624 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6625 * For now this just excludes isolated cpus, but could be used to
6626 * exclude other special cases in the future.
6628 static int init_sched_domains(const struct cpumask *cpu_map)
6632 arch_update_cpu_topology();
6634 doms_cur = alloc_sched_domains(ndoms_cur);
6636 doms_cur = &fallback_doms;
6637 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6639 err = build_sched_domains(doms_cur[0], NULL);
6640 register_sched_domain_sysctl();
6646 * Detach sched domains from a group of cpus specified in cpu_map
6647 * These cpus will now be attached to the NULL domain
6649 static void detach_destroy_domains(const struct cpumask *cpu_map)
6654 for_each_cpu(i, cpu_map)
6655 cpu_attach_domain(NULL, &def_root_domain, i);
6659 /* handle null as "default" */
6660 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6661 struct sched_domain_attr *new, int idx_new)
6663 struct sched_domain_attr tmp;
6670 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6671 new ? (new + idx_new) : &tmp,
6672 sizeof(struct sched_domain_attr));
6676 * Partition sched domains as specified by the 'ndoms_new'
6677 * cpumasks in the array doms_new[] of cpumasks. This compares
6678 * doms_new[] to the current sched domain partitioning, doms_cur[].
6679 * It destroys each deleted domain and builds each new domain.
6681 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6682 * The masks don't intersect (don't overlap.) We should setup one
6683 * sched domain for each mask. CPUs not in any of the cpumasks will
6684 * not be load balanced. If the same cpumask appears both in the
6685 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6688 * The passed in 'doms_new' should be allocated using
6689 * alloc_sched_domains. This routine takes ownership of it and will
6690 * free_sched_domains it when done with it. If the caller failed the
6691 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6692 * and partition_sched_domains() will fallback to the single partition
6693 * 'fallback_doms', it also forces the domains to be rebuilt.
6695 * If doms_new == NULL it will be replaced with cpu_online_mask.
6696 * ndoms_new == 0 is a special case for destroying existing domains,
6697 * and it will not create the default domain.
6699 * Call with hotplug lock held
6701 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6702 struct sched_domain_attr *dattr_new)
6707 mutex_lock(&sched_domains_mutex);
6709 /* always unregister in case we don't destroy any domains */
6710 unregister_sched_domain_sysctl();
6712 /* Let architecture update cpu core mappings. */
6713 new_topology = arch_update_cpu_topology();
6715 n = doms_new ? ndoms_new : 0;
6717 /* Destroy deleted domains */
6718 for (i = 0; i < ndoms_cur; i++) {
6719 for (j = 0; j < n && !new_topology; j++) {
6720 if (cpumask_equal(doms_cur[i], doms_new[j])
6721 && dattrs_equal(dattr_cur, i, dattr_new, j))
6724 /* no match - a current sched domain not in new doms_new[] */
6725 detach_destroy_domains(doms_cur[i]);
6730 if (doms_new == NULL) {
6732 doms_new = &fallback_doms;
6733 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6734 WARN_ON_ONCE(dattr_new);
6737 /* Build new domains */
6738 for (i = 0; i < ndoms_new; i++) {
6739 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6740 if (cpumask_equal(doms_new[i], doms_cur[j])
6741 && dattrs_equal(dattr_new, i, dattr_cur, j))
6744 /* no match - add a new doms_new */
6745 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6750 /* Remember the new sched domains */
6751 if (doms_cur != &fallback_doms)
6752 free_sched_domains(doms_cur, ndoms_cur);
6753 kfree(dattr_cur); /* kfree(NULL) is safe */
6754 doms_cur = doms_new;
6755 dattr_cur = dattr_new;
6756 ndoms_cur = ndoms_new;
6758 register_sched_domain_sysctl();
6760 mutex_unlock(&sched_domains_mutex);
6763 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6764 static void reinit_sched_domains(void)
6768 /* Destroy domains first to force the rebuild */
6769 partition_sched_domains(0, NULL, NULL);
6771 rebuild_sched_domains();
6775 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
6777 unsigned int level = 0;
6779 if (sscanf(buf, "%u", &level) != 1)
6783 * level is always be positive so don't check for
6784 * level < POWERSAVINGS_BALANCE_NONE which is 0
6785 * What happens on 0 or 1 byte write,
6786 * need to check for count as well?
6789 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
6793 sched_smt_power_savings = level;
6795 sched_mc_power_savings = level;
6797 reinit_sched_domains();
6802 #ifdef CONFIG_SCHED_MC
6803 static ssize_t sched_mc_power_savings_show(struct device *dev,
6804 struct device_attribute *attr,
6807 return sprintf(buf, "%u\n", sched_mc_power_savings);
6809 static ssize_t sched_mc_power_savings_store(struct device *dev,
6810 struct device_attribute *attr,
6811 const char *buf, size_t count)
6813 return sched_power_savings_store(buf, count, 0);
6815 static DEVICE_ATTR(sched_mc_power_savings, 0644,
6816 sched_mc_power_savings_show,
6817 sched_mc_power_savings_store);
6820 #ifdef CONFIG_SCHED_SMT
6821 static ssize_t sched_smt_power_savings_show(struct device *dev,
6822 struct device_attribute *attr,
6825 return sprintf(buf, "%u\n", sched_smt_power_savings);
6827 static ssize_t sched_smt_power_savings_store(struct device *dev,
6828 struct device_attribute *attr,
6829 const char *buf, size_t count)
6831 return sched_power_savings_store(buf, count, 1);
6833 static DEVICE_ATTR(sched_smt_power_savings, 0644,
6834 sched_smt_power_savings_show,
6835 sched_smt_power_savings_store);
6838 int __init sched_create_sysfs_power_savings_entries(struct device *dev)
6842 #ifdef CONFIG_SCHED_SMT
6844 err = device_create_file(dev, &dev_attr_sched_smt_power_savings);
6846 #ifdef CONFIG_SCHED_MC
6847 if (!err && mc_capable())
6848 err = device_create_file(dev, &dev_attr_sched_mc_power_savings);
6852 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
6855 * Update cpusets according to cpu_active mask. If cpusets are
6856 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6857 * around partition_sched_domains().
6859 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6862 switch (action & ~CPU_TASKS_FROZEN) {
6864 case CPU_DOWN_FAILED:
6865 cpuset_update_active_cpus();
6872 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6875 switch (action & ~CPU_TASKS_FROZEN) {
6876 case CPU_DOWN_PREPARE:
6877 cpuset_update_active_cpus();
6884 void __init sched_init_smp(void)
6886 cpumask_var_t non_isolated_cpus;
6888 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6889 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6892 mutex_lock(&sched_domains_mutex);
6893 init_sched_domains(cpu_active_mask);
6894 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6895 if (cpumask_empty(non_isolated_cpus))
6896 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6897 mutex_unlock(&sched_domains_mutex);
6900 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6901 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6903 /* RT runtime code needs to handle some hotplug events */
6904 hotcpu_notifier(update_runtime, 0);
6908 /* Move init over to a non-isolated CPU */
6909 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6911 sched_init_granularity();
6912 free_cpumask_var(non_isolated_cpus);
6914 init_sched_rt_class();
6917 void __init sched_init_smp(void)
6919 sched_init_granularity();
6921 #endif /* CONFIG_SMP */
6923 const_debug unsigned int sysctl_timer_migration = 1;
6925 int in_sched_functions(unsigned long addr)
6927 return in_lock_functions(addr) ||
6928 (addr >= (unsigned long)__sched_text_start
6929 && addr < (unsigned long)__sched_text_end);
6932 #ifdef CONFIG_CGROUP_SCHED
6933 struct task_group root_task_group;
6936 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6938 void __init sched_init(void)
6941 unsigned long alloc_size = 0, ptr;
6943 #ifdef CONFIG_FAIR_GROUP_SCHED
6944 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6946 #ifdef CONFIG_RT_GROUP_SCHED
6947 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6949 #ifdef CONFIG_CPUMASK_OFFSTACK
6950 alloc_size += num_possible_cpus() * cpumask_size();
6953 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6955 #ifdef CONFIG_FAIR_GROUP_SCHED
6956 root_task_group.se = (struct sched_entity **)ptr;
6957 ptr += nr_cpu_ids * sizeof(void **);
6959 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6960 ptr += nr_cpu_ids * sizeof(void **);
6962 #endif /* CONFIG_FAIR_GROUP_SCHED */
6963 #ifdef CONFIG_RT_GROUP_SCHED
6964 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6965 ptr += nr_cpu_ids * sizeof(void **);
6967 root_task_group.rt_rq = (struct rt_rq **)ptr;
6968 ptr += nr_cpu_ids * sizeof(void **);
6970 #endif /* CONFIG_RT_GROUP_SCHED */
6971 #ifdef CONFIG_CPUMASK_OFFSTACK
6972 for_each_possible_cpu(i) {
6973 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6974 ptr += cpumask_size();
6976 #endif /* CONFIG_CPUMASK_OFFSTACK */
6980 init_defrootdomain();
6983 init_rt_bandwidth(&def_rt_bandwidth,
6984 global_rt_period(), global_rt_runtime());
6986 #ifdef CONFIG_RT_GROUP_SCHED
6987 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6988 global_rt_period(), global_rt_runtime());
6989 #endif /* CONFIG_RT_GROUP_SCHED */
6991 #ifdef CONFIG_CGROUP_SCHED
6992 list_add(&root_task_group.list, &task_groups);
6993 INIT_LIST_HEAD(&root_task_group.children);
6994 INIT_LIST_HEAD(&root_task_group.siblings);
6995 autogroup_init(&init_task);
6997 #endif /* CONFIG_CGROUP_SCHED */
6999 #ifdef CONFIG_CGROUP_CPUACCT
7000 root_cpuacct.cpustat = &kernel_cpustat;
7001 root_cpuacct.cpuusage = alloc_percpu(u64);
7002 /* Too early, not expected to fail */
7003 BUG_ON(!root_cpuacct.cpuusage);
7005 for_each_possible_cpu(i) {
7009 raw_spin_lock_init(&rq->lock);
7011 rq->calc_load_active = 0;
7012 rq->calc_load_update = jiffies + LOAD_FREQ;
7013 init_cfs_rq(&rq->cfs);
7014 init_rt_rq(&rq->rt, rq);
7015 #ifdef CONFIG_FAIR_GROUP_SCHED
7016 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7017 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7019 * How much cpu bandwidth does root_task_group get?
7021 * In case of task-groups formed thr' the cgroup filesystem, it
7022 * gets 100% of the cpu resources in the system. This overall
7023 * system cpu resource is divided among the tasks of
7024 * root_task_group and its child task-groups in a fair manner,
7025 * based on each entity's (task or task-group's) weight
7026 * (se->load.weight).
7028 * In other words, if root_task_group has 10 tasks of weight
7029 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7030 * then A0's share of the cpu resource is:
7032 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7034 * We achieve this by letting root_task_group's tasks sit
7035 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7037 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7038 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7039 #endif /* CONFIG_FAIR_GROUP_SCHED */
7041 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7042 #ifdef CONFIG_RT_GROUP_SCHED
7043 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7044 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7047 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7048 rq->cpu_load[j] = 0;
7050 rq->last_load_update_tick = jiffies;
7055 rq->cpu_power = SCHED_POWER_SCALE;
7056 rq->post_schedule = 0;
7057 rq->active_balance = 0;
7058 rq->next_balance = jiffies;
7063 rq->avg_idle = 2*sysctl_sched_migration_cost;
7065 INIT_LIST_HEAD(&rq->cfs_tasks);
7067 rq_attach_root(rq, &def_root_domain);
7073 atomic_set(&rq->nr_iowait, 0);
7076 set_load_weight(&init_task);
7078 #ifdef CONFIG_PREEMPT_NOTIFIERS
7079 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7082 #ifdef CONFIG_RT_MUTEXES
7083 plist_head_init(&init_task.pi_waiters);
7087 * The boot idle thread does lazy MMU switching as well:
7089 atomic_inc(&init_mm.mm_count);
7090 enter_lazy_tlb(&init_mm, current);
7093 * Make us the idle thread. Technically, schedule() should not be
7094 * called from this thread, however somewhere below it might be,
7095 * but because we are the idle thread, we just pick up running again
7096 * when this runqueue becomes "idle".
7098 init_idle(current, smp_processor_id());
7100 calc_load_update = jiffies + LOAD_FREQ;
7103 * During early bootup we pretend to be a normal task:
7105 current->sched_class = &fair_sched_class;
7108 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7109 /* May be allocated at isolcpus cmdline parse time */
7110 if (cpu_isolated_map == NULL)
7111 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7113 init_sched_fair_class();
7115 scheduler_running = 1;
7118 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7119 static inline int preempt_count_equals(int preempt_offset)
7121 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7123 return (nested == preempt_offset);
7126 void __might_sleep(const char *file, int line, int preempt_offset)
7128 static unsigned long prev_jiffy; /* ratelimiting */
7130 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7131 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7132 system_state != SYSTEM_RUNNING || oops_in_progress)
7134 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7136 prev_jiffy = jiffies;
7139 "BUG: sleeping function called from invalid context at %s:%d\n",
7142 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7143 in_atomic(), irqs_disabled(),
7144 current->pid, current->comm);
7146 debug_show_held_locks(current);
7147 if (irqs_disabled())
7148 print_irqtrace_events(current);
7151 EXPORT_SYMBOL(__might_sleep);
7154 #ifdef CONFIG_MAGIC_SYSRQ
7155 static void normalize_task(struct rq *rq, struct task_struct *p)
7157 const struct sched_class *prev_class = p->sched_class;
7158 int old_prio = p->prio;
7163 dequeue_task(rq, p, 0);
7164 __setscheduler(rq, p, SCHED_NORMAL, 0);
7166 enqueue_task(rq, p, 0);
7167 resched_task(rq->curr);
7170 check_class_changed(rq, p, prev_class, old_prio);
7173 void normalize_rt_tasks(void)
7175 struct task_struct *g, *p;
7176 unsigned long flags;
7179 read_lock_irqsave(&tasklist_lock, flags);
7180 do_each_thread(g, p) {
7182 * Only normalize user tasks:
7187 p->se.exec_start = 0;
7188 #ifdef CONFIG_SCHEDSTATS
7189 p->se.statistics.wait_start = 0;
7190 p->se.statistics.sleep_start = 0;
7191 p->se.statistics.block_start = 0;
7196 * Renice negative nice level userspace
7199 if (TASK_NICE(p) < 0 && p->mm)
7200 set_user_nice(p, 0);
7204 raw_spin_lock(&p->pi_lock);
7205 rq = __task_rq_lock(p);
7207 normalize_task(rq, p);
7209 __task_rq_unlock(rq);
7210 raw_spin_unlock(&p->pi_lock);
7211 } while_each_thread(g, p);
7213 read_unlock_irqrestore(&tasklist_lock, flags);
7216 #endif /* CONFIG_MAGIC_SYSRQ */
7218 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7220 * These functions are only useful for the IA64 MCA handling, or kdb.
7222 * They can only be called when the whole system has been
7223 * stopped - every CPU needs to be quiescent, and no scheduling
7224 * activity can take place. Using them for anything else would
7225 * be a serious bug, and as a result, they aren't even visible
7226 * under any other configuration.
7230 * curr_task - return the current task for a given cpu.
7231 * @cpu: the processor in question.
7233 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7235 struct task_struct *curr_task(int cpu)
7237 return cpu_curr(cpu);
7240 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7244 * set_curr_task - set the current task for a given cpu.
7245 * @cpu: the processor in question.
7246 * @p: the task pointer to set.
7248 * Description: This function must only be used when non-maskable interrupts
7249 * are serviced on a separate stack. It allows the architecture to switch the
7250 * notion of the current task on a cpu in a non-blocking manner. This function
7251 * must be called with all CPU's synchronized, and interrupts disabled, the
7252 * and caller must save the original value of the current task (see
7253 * curr_task() above) and restore that value before reenabling interrupts and
7254 * re-starting the system.
7256 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7258 void set_curr_task(int cpu, struct task_struct *p)
7265 #ifdef CONFIG_CGROUP_SCHED
7266 /* task_group_lock serializes the addition/removal of task groups */
7267 static DEFINE_SPINLOCK(task_group_lock);
7269 static void free_sched_group(struct task_group *tg)
7271 free_fair_sched_group(tg);
7272 free_rt_sched_group(tg);
7277 /* allocate runqueue etc for a new task group */
7278 struct task_group *sched_create_group(struct task_group *parent)
7280 struct task_group *tg;
7281 unsigned long flags;
7283 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7285 return ERR_PTR(-ENOMEM);
7287 if (!alloc_fair_sched_group(tg, parent))
7290 if (!alloc_rt_sched_group(tg, parent))
7293 spin_lock_irqsave(&task_group_lock, flags);
7294 list_add_rcu(&tg->list, &task_groups);
7296 WARN_ON(!parent); /* root should already exist */
7298 tg->parent = parent;
7299 INIT_LIST_HEAD(&tg->children);
7300 list_add_rcu(&tg->siblings, &parent->children);
7301 spin_unlock_irqrestore(&task_group_lock, flags);
7306 free_sched_group(tg);
7307 return ERR_PTR(-ENOMEM);
7310 /* rcu callback to free various structures associated with a task group */
7311 static void free_sched_group_rcu(struct rcu_head *rhp)
7313 /* now it should be safe to free those cfs_rqs */
7314 free_sched_group(container_of(rhp, struct task_group, rcu));
7317 /* Destroy runqueue etc associated with a task group */
7318 void sched_destroy_group(struct task_group *tg)
7320 unsigned long flags;
7323 /* end participation in shares distribution */
7324 for_each_possible_cpu(i)
7325 unregister_fair_sched_group(tg, i);
7327 spin_lock_irqsave(&task_group_lock, flags);
7328 list_del_rcu(&tg->list);
7329 list_del_rcu(&tg->siblings);
7330 spin_unlock_irqrestore(&task_group_lock, flags);
7332 /* wait for possible concurrent references to cfs_rqs complete */
7333 call_rcu(&tg->rcu, free_sched_group_rcu);
7336 /* change task's runqueue when it moves between groups.
7337 * The caller of this function should have put the task in its new group
7338 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7339 * reflect its new group.
7341 void sched_move_task(struct task_struct *tsk)
7344 unsigned long flags;
7347 rq = task_rq_lock(tsk, &flags);
7349 running = task_current(rq, tsk);
7353 dequeue_task(rq, tsk, 0);
7354 if (unlikely(running))
7355 tsk->sched_class->put_prev_task(rq, tsk);
7357 #ifdef CONFIG_FAIR_GROUP_SCHED
7358 if (tsk->sched_class->task_move_group)
7359 tsk->sched_class->task_move_group(tsk, on_rq);
7362 set_task_rq(tsk, task_cpu(tsk));
7364 if (unlikely(running))
7365 tsk->sched_class->set_curr_task(rq);
7367 enqueue_task(rq, tsk, 0);
7369 task_rq_unlock(rq, tsk, &flags);
7371 #endif /* CONFIG_CGROUP_SCHED */
7373 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7374 static unsigned long to_ratio(u64 period, u64 runtime)
7376 if (runtime == RUNTIME_INF)
7379 return div64_u64(runtime << 20, period);
7383 #ifdef CONFIG_RT_GROUP_SCHED
7385 * Ensure that the real time constraints are schedulable.
7387 static DEFINE_MUTEX(rt_constraints_mutex);
7389 /* Must be called with tasklist_lock held */
7390 static inline int tg_has_rt_tasks(struct task_group *tg)
7392 struct task_struct *g, *p;
7394 do_each_thread(g, p) {
7395 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7397 } while_each_thread(g, p);
7402 struct rt_schedulable_data {
7403 struct task_group *tg;
7408 static int tg_rt_schedulable(struct task_group *tg, void *data)
7410 struct rt_schedulable_data *d = data;
7411 struct task_group *child;
7412 unsigned long total, sum = 0;
7413 u64 period, runtime;
7415 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7416 runtime = tg->rt_bandwidth.rt_runtime;
7419 period = d->rt_period;
7420 runtime = d->rt_runtime;
7424 * Cannot have more runtime than the period.
7426 if (runtime > period && runtime != RUNTIME_INF)
7430 * Ensure we don't starve existing RT tasks.
7432 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7435 total = to_ratio(period, runtime);
7438 * Nobody can have more than the global setting allows.
7440 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7444 * The sum of our children's runtime should not exceed our own.
7446 list_for_each_entry_rcu(child, &tg->children, siblings) {
7447 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7448 runtime = child->rt_bandwidth.rt_runtime;
7450 if (child == d->tg) {
7451 period = d->rt_period;
7452 runtime = d->rt_runtime;
7455 sum += to_ratio(period, runtime);
7464 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7468 struct rt_schedulable_data data = {
7470 .rt_period = period,
7471 .rt_runtime = runtime,
7475 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7481 static int tg_set_rt_bandwidth(struct task_group *tg,
7482 u64 rt_period, u64 rt_runtime)
7486 mutex_lock(&rt_constraints_mutex);
7487 read_lock(&tasklist_lock);
7488 err = __rt_schedulable(tg, rt_period, rt_runtime);
7492 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7493 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7494 tg->rt_bandwidth.rt_runtime = rt_runtime;
7496 for_each_possible_cpu(i) {
7497 struct rt_rq *rt_rq = tg->rt_rq[i];
7499 raw_spin_lock(&rt_rq->rt_runtime_lock);
7500 rt_rq->rt_runtime = rt_runtime;
7501 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7503 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7505 read_unlock(&tasklist_lock);
7506 mutex_unlock(&rt_constraints_mutex);
7511 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7513 u64 rt_runtime, rt_period;
7515 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7516 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7517 if (rt_runtime_us < 0)
7518 rt_runtime = RUNTIME_INF;
7520 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7523 long sched_group_rt_runtime(struct task_group *tg)
7527 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7530 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7531 do_div(rt_runtime_us, NSEC_PER_USEC);
7532 return rt_runtime_us;
7535 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7537 u64 rt_runtime, rt_period;
7539 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7540 rt_runtime = tg->rt_bandwidth.rt_runtime;
7545 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7548 long sched_group_rt_period(struct task_group *tg)
7552 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7553 do_div(rt_period_us, NSEC_PER_USEC);
7554 return rt_period_us;
7557 static int sched_rt_global_constraints(void)
7559 u64 runtime, period;
7562 if (sysctl_sched_rt_period <= 0)
7565 runtime = global_rt_runtime();
7566 period = global_rt_period();
7569 * Sanity check on the sysctl variables.
7571 if (runtime > period && runtime != RUNTIME_INF)
7574 mutex_lock(&rt_constraints_mutex);
7575 read_lock(&tasklist_lock);
7576 ret = __rt_schedulable(NULL, 0, 0);
7577 read_unlock(&tasklist_lock);
7578 mutex_unlock(&rt_constraints_mutex);
7583 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7585 /* Don't accept realtime tasks when there is no way for them to run */
7586 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7592 #else /* !CONFIG_RT_GROUP_SCHED */
7593 static int sched_rt_global_constraints(void)
7595 unsigned long flags;
7598 if (sysctl_sched_rt_period <= 0)
7602 * There's always some RT tasks in the root group
7603 * -- migration, kstopmachine etc..
7605 if (sysctl_sched_rt_runtime == 0)
7608 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7609 for_each_possible_cpu(i) {
7610 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7612 raw_spin_lock(&rt_rq->rt_runtime_lock);
7613 rt_rq->rt_runtime = global_rt_runtime();
7614 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7616 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7620 #endif /* CONFIG_RT_GROUP_SCHED */
7622 int sched_rt_handler(struct ctl_table *table, int write,
7623 void __user *buffer, size_t *lenp,
7627 int old_period, old_runtime;
7628 static DEFINE_MUTEX(mutex);
7631 old_period = sysctl_sched_rt_period;
7632 old_runtime = sysctl_sched_rt_runtime;
7634 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7636 if (!ret && write) {
7637 ret = sched_rt_global_constraints();
7639 sysctl_sched_rt_period = old_period;
7640 sysctl_sched_rt_runtime = old_runtime;
7642 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7643 def_rt_bandwidth.rt_period =
7644 ns_to_ktime(global_rt_period());
7647 mutex_unlock(&mutex);
7652 #ifdef CONFIG_CGROUP_SCHED
7654 /* return corresponding task_group object of a cgroup */
7655 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7657 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7658 struct task_group, css);
7661 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7663 struct task_group *tg, *parent;
7665 if (!cgrp->parent) {
7666 /* This is early initialization for the top cgroup */
7667 return &root_task_group.css;
7670 parent = cgroup_tg(cgrp->parent);
7671 tg = sched_create_group(parent);
7673 return ERR_PTR(-ENOMEM);
7678 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7680 struct task_group *tg = cgroup_tg(cgrp);
7682 sched_destroy_group(tg);
7685 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7686 struct cgroup_taskset *tset)
7688 struct task_struct *task;
7690 cgroup_taskset_for_each(task, cgrp, tset) {
7691 #ifdef CONFIG_RT_GROUP_SCHED
7692 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7695 /* We don't support RT-tasks being in separate groups */
7696 if (task->sched_class != &fair_sched_class)
7703 static void cpu_cgroup_attach(struct cgroup *cgrp,
7704 struct cgroup_taskset *tset)
7706 struct task_struct *task;
7708 cgroup_taskset_for_each(task, cgrp, tset)
7709 sched_move_task(task);
7713 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7714 struct task_struct *task)
7717 * cgroup_exit() is called in the copy_process() failure path.
7718 * Ignore this case since the task hasn't ran yet, this avoids
7719 * trying to poke a half freed task state from generic code.
7721 if (!(task->flags & PF_EXITING))
7724 sched_move_task(task);
7727 #ifdef CONFIG_FAIR_GROUP_SCHED
7728 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7731 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7734 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7736 struct task_group *tg = cgroup_tg(cgrp);
7738 return (u64) scale_load_down(tg->shares);
7741 #ifdef CONFIG_CFS_BANDWIDTH
7742 static DEFINE_MUTEX(cfs_constraints_mutex);
7744 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7745 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7747 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7749 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7751 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7752 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7754 if (tg == &root_task_group)
7758 * Ensure we have at some amount of bandwidth every period. This is
7759 * to prevent reaching a state of large arrears when throttled via
7760 * entity_tick() resulting in prolonged exit starvation.
7762 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7766 * Likewise, bound things on the otherside by preventing insane quota
7767 * periods. This also allows us to normalize in computing quota
7770 if (period > max_cfs_quota_period)
7773 mutex_lock(&cfs_constraints_mutex);
7774 ret = __cfs_schedulable(tg, period, quota);
7778 runtime_enabled = quota != RUNTIME_INF;
7779 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7780 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7781 raw_spin_lock_irq(&cfs_b->lock);
7782 cfs_b->period = ns_to_ktime(period);
7783 cfs_b->quota = quota;
7785 __refill_cfs_bandwidth_runtime(cfs_b);
7786 /* restart the period timer (if active) to handle new period expiry */
7787 if (runtime_enabled && cfs_b->timer_active) {
7788 /* force a reprogram */
7789 cfs_b->timer_active = 0;
7790 __start_cfs_bandwidth(cfs_b);
7792 raw_spin_unlock_irq(&cfs_b->lock);
7794 for_each_possible_cpu(i) {
7795 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7796 struct rq *rq = cfs_rq->rq;
7798 raw_spin_lock_irq(&rq->lock);
7799 cfs_rq->runtime_enabled = runtime_enabled;
7800 cfs_rq->runtime_remaining = 0;
7802 if (cfs_rq->throttled)
7803 unthrottle_cfs_rq(cfs_rq);
7804 raw_spin_unlock_irq(&rq->lock);
7807 mutex_unlock(&cfs_constraints_mutex);
7812 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7816 period = ktime_to_ns(tg->cfs_bandwidth.period);
7817 if (cfs_quota_us < 0)
7818 quota = RUNTIME_INF;
7820 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7822 return tg_set_cfs_bandwidth(tg, period, quota);
7825 long tg_get_cfs_quota(struct task_group *tg)
7829 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7832 quota_us = tg->cfs_bandwidth.quota;
7833 do_div(quota_us, NSEC_PER_USEC);
7838 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7842 period = (u64)cfs_period_us * NSEC_PER_USEC;
7843 quota = tg->cfs_bandwidth.quota;
7845 return tg_set_cfs_bandwidth(tg, period, quota);
7848 long tg_get_cfs_period(struct task_group *tg)
7852 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7853 do_div(cfs_period_us, NSEC_PER_USEC);
7855 return cfs_period_us;
7858 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7860 return tg_get_cfs_quota(cgroup_tg(cgrp));
7863 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7866 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7869 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7871 return tg_get_cfs_period(cgroup_tg(cgrp));
7874 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7877 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7880 struct cfs_schedulable_data {
7881 struct task_group *tg;
7886 * normalize group quota/period to be quota/max_period
7887 * note: units are usecs
7889 static u64 normalize_cfs_quota(struct task_group *tg,
7890 struct cfs_schedulable_data *d)
7898 period = tg_get_cfs_period(tg);
7899 quota = tg_get_cfs_quota(tg);
7902 /* note: these should typically be equivalent */
7903 if (quota == RUNTIME_INF || quota == -1)
7906 return to_ratio(period, quota);
7909 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7911 struct cfs_schedulable_data *d = data;
7912 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7913 s64 quota = 0, parent_quota = -1;
7916 quota = RUNTIME_INF;
7918 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7920 quota = normalize_cfs_quota(tg, d);
7921 parent_quota = parent_b->hierarchal_quota;
7924 * ensure max(child_quota) <= parent_quota, inherit when no
7927 if (quota == RUNTIME_INF)
7928 quota = parent_quota;
7929 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7932 cfs_b->hierarchal_quota = quota;
7937 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7940 struct cfs_schedulable_data data = {
7946 if (quota != RUNTIME_INF) {
7947 do_div(data.period, NSEC_PER_USEC);
7948 do_div(data.quota, NSEC_PER_USEC);
7952 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7958 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7959 struct cgroup_map_cb *cb)
7961 struct task_group *tg = cgroup_tg(cgrp);
7962 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7964 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7965 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7966 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7970 #endif /* CONFIG_CFS_BANDWIDTH */
7971 #endif /* CONFIG_FAIR_GROUP_SCHED */
7973 #ifdef CONFIG_RT_GROUP_SCHED
7974 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7977 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7980 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7982 return sched_group_rt_runtime(cgroup_tg(cgrp));
7985 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7988 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7991 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7993 return sched_group_rt_period(cgroup_tg(cgrp));
7995 #endif /* CONFIG_RT_GROUP_SCHED */
7997 static struct cftype cpu_files[] = {
7998 #ifdef CONFIG_FAIR_GROUP_SCHED
8001 .read_u64 = cpu_shares_read_u64,
8002 .write_u64 = cpu_shares_write_u64,
8005 #ifdef CONFIG_CFS_BANDWIDTH
8007 .name = "cfs_quota_us",
8008 .read_s64 = cpu_cfs_quota_read_s64,
8009 .write_s64 = cpu_cfs_quota_write_s64,
8012 .name = "cfs_period_us",
8013 .read_u64 = cpu_cfs_period_read_u64,
8014 .write_u64 = cpu_cfs_period_write_u64,
8018 .read_map = cpu_stats_show,
8021 #ifdef CONFIG_RT_GROUP_SCHED
8023 .name = "rt_runtime_us",
8024 .read_s64 = cpu_rt_runtime_read,
8025 .write_s64 = cpu_rt_runtime_write,
8028 .name = "rt_period_us",
8029 .read_u64 = cpu_rt_period_read_uint,
8030 .write_u64 = cpu_rt_period_write_uint,
8035 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8037 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8040 struct cgroup_subsys cpu_cgroup_subsys = {
8042 .create = cpu_cgroup_create,
8043 .destroy = cpu_cgroup_destroy,
8044 .can_attach = cpu_cgroup_can_attach,
8045 .attach = cpu_cgroup_attach,
8046 .exit = cpu_cgroup_exit,
8047 .populate = cpu_cgroup_populate,
8048 .subsys_id = cpu_cgroup_subsys_id,
8052 #endif /* CONFIG_CGROUP_SCHED */
8054 #ifdef CONFIG_CGROUP_CPUACCT
8057 * CPU accounting code for task groups.
8059 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8060 * (balbir@in.ibm.com).
8063 /* create a new cpu accounting group */
8064 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8069 return &root_cpuacct.css;
8071 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8075 ca->cpuusage = alloc_percpu(u64);
8079 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8081 goto out_free_cpuusage;
8086 free_percpu(ca->cpuusage);
8090 return ERR_PTR(-ENOMEM);
8093 /* destroy an existing cpu accounting group */
8094 static void cpuacct_destroy(struct cgroup *cgrp)
8096 struct cpuacct *ca = cgroup_ca(cgrp);
8098 free_percpu(ca->cpustat);
8099 free_percpu(ca->cpuusage);
8103 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8105 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8108 #ifndef CONFIG_64BIT
8110 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8112 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8114 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8122 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8124 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8126 #ifndef CONFIG_64BIT
8128 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8130 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8132 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8138 /* return total cpu usage (in nanoseconds) of a group */
8139 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8141 struct cpuacct *ca = cgroup_ca(cgrp);
8142 u64 totalcpuusage = 0;
8145 for_each_present_cpu(i)
8146 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8148 return totalcpuusage;
8151 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8154 struct cpuacct *ca = cgroup_ca(cgrp);
8163 for_each_present_cpu(i)
8164 cpuacct_cpuusage_write(ca, i, 0);
8170 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8173 struct cpuacct *ca = cgroup_ca(cgroup);
8177 for_each_present_cpu(i) {
8178 percpu = cpuacct_cpuusage_read(ca, i);
8179 seq_printf(m, "%llu ", (unsigned long long) percpu);
8181 seq_printf(m, "\n");
8185 static const char *cpuacct_stat_desc[] = {
8186 [CPUACCT_STAT_USER] = "user",
8187 [CPUACCT_STAT_SYSTEM] = "system",
8190 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8191 struct cgroup_map_cb *cb)
8193 struct cpuacct *ca = cgroup_ca(cgrp);
8197 for_each_online_cpu(cpu) {
8198 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8199 val += kcpustat->cpustat[CPUTIME_USER];
8200 val += kcpustat->cpustat[CPUTIME_NICE];
8202 val = cputime64_to_clock_t(val);
8203 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8206 for_each_online_cpu(cpu) {
8207 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8208 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8209 val += kcpustat->cpustat[CPUTIME_IRQ];
8210 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8213 val = cputime64_to_clock_t(val);
8214 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8219 static struct cftype files[] = {
8222 .read_u64 = cpuusage_read,
8223 .write_u64 = cpuusage_write,
8226 .name = "usage_percpu",
8227 .read_seq_string = cpuacct_percpu_seq_read,
8231 .read_map = cpuacct_stats_show,
8235 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8237 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8241 * charge this task's execution time to its accounting group.
8243 * called with rq->lock held.
8245 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8250 if (unlikely(!cpuacct_subsys.active))
8253 cpu = task_cpu(tsk);
8259 for (; ca; ca = parent_ca(ca)) {
8260 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8261 *cpuusage += cputime;
8267 struct cgroup_subsys cpuacct_subsys = {
8269 .create = cpuacct_create,
8270 .destroy = cpuacct_destroy,
8271 .populate = cpuacct_populate,
8272 .subsys_id = cpuacct_subsys_id,
8274 #endif /* CONFIG_CGROUP_CPUACCT */