4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 ktime_t soft, hard, now;
97 if (hrtimer_active(period_timer))
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
111 DEFINE_MUTEX(sched_domains_mutex);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 void update_rq_clock(struct rq *rq)
120 if (rq->skip_clock_update > 0)
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 update_rq_clock_task(rq, delta);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug unsigned int sysctl_sched_features =
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names[] = {
146 #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 static void set_load_weight(struct task_struct *p)
697 int prio = p->static_prio - MAX_RT_PRIO;
698 struct load_weight *load = &p->se.load;
701 * SCHED_IDLE tasks get minimal weight:
703 if (p->policy == SCHED_IDLE) {
704 load->weight = scale_load(WEIGHT_IDLEPRIO);
705 load->inv_weight = WMULT_IDLEPRIO;
709 load->weight = scale_load(prio_to_weight[prio]);
710 load->inv_weight = prio_to_wmult[prio];
713 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
716 sched_info_queued(p);
717 p->sched_class->enqueue_task(rq, p, flags);
720 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
723 sched_info_dequeued(p);
724 p->sched_class->dequeue_task(rq, p, flags);
727 void activate_task(struct rq *rq, struct task_struct *p, int flags)
729 if (task_contributes_to_load(p))
730 rq->nr_uninterruptible--;
732 enqueue_task(rq, p, flags);
735 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible++;
740 dequeue_task(rq, p, flags);
743 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
746 * There are no locks covering percpu hardirq/softirq time.
747 * They are only modified in account_system_vtime, on corresponding CPU
748 * with interrupts disabled. So, writes are safe.
749 * They are read and saved off onto struct rq in update_rq_clock().
750 * This may result in other CPU reading this CPU's irq time and can
751 * race with irq/account_system_vtime on this CPU. We would either get old
752 * or new value with a side effect of accounting a slice of irq time to wrong
753 * task when irq is in progress while we read rq->clock. That is a worthy
754 * compromise in place of having locks on each irq in account_system_time.
756 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
757 static DEFINE_PER_CPU(u64, cpu_softirq_time);
759 static DEFINE_PER_CPU(u64, irq_start_time);
760 static int sched_clock_irqtime;
762 void enable_sched_clock_irqtime(void)
764 sched_clock_irqtime = 1;
767 void disable_sched_clock_irqtime(void)
769 sched_clock_irqtime = 0;
773 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
775 static inline void irq_time_write_begin(void)
777 __this_cpu_inc(irq_time_seq.sequence);
781 static inline void irq_time_write_end(void)
784 __this_cpu_inc(irq_time_seq.sequence);
787 static inline u64 irq_time_read(int cpu)
793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
794 irq_time = per_cpu(cpu_softirq_time, cpu) +
795 per_cpu(cpu_hardirq_time, cpu);
796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
800 #else /* CONFIG_64BIT */
801 static inline void irq_time_write_begin(void)
805 static inline void irq_time_write_end(void)
809 static inline u64 irq_time_read(int cpu)
811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
813 #endif /* CONFIG_64BIT */
816 * Called before incrementing preempt_count on {soft,}irq_enter
817 * and before decrementing preempt_count on {soft,}irq_exit.
819 void account_system_vtime(struct task_struct *curr)
825 if (!sched_clock_irqtime)
828 local_irq_save(flags);
830 cpu = smp_processor_id();
831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
832 __this_cpu_add(irq_start_time, delta);
834 irq_time_write_begin();
836 * We do not account for softirq time from ksoftirqd here.
837 * We want to continue accounting softirq time to ksoftirqd thread
838 * in that case, so as not to confuse scheduler with a special task
839 * that do not consume any time, but still wants to run.
842 __this_cpu_add(cpu_hardirq_time, delta);
843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
844 __this_cpu_add(cpu_softirq_time, delta);
846 irq_time_write_end();
847 local_irq_restore(flags);
849 EXPORT_SYMBOL_GPL(account_system_vtime);
851 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
853 #ifdef CONFIG_PARAVIRT
854 static inline u64 steal_ticks(u64 steal)
856 if (unlikely(steal > NSEC_PER_SEC))
857 return div_u64(steal, TICK_NSEC);
859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
863 static void update_rq_clock_task(struct rq *rq, s64 delta)
866 * In theory, the compile should just see 0 here, and optimize out the call
867 * to sched_rt_avg_update. But I don't trust it...
869 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
870 s64 steal = 0, irq_delta = 0;
872 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
876 * Since irq_time is only updated on {soft,}irq_exit, we might run into
877 * this case when a previous update_rq_clock() happened inside a
880 * When this happens, we stop ->clock_task and only update the
881 * prev_irq_time stamp to account for the part that fit, so that a next
882 * update will consume the rest. This ensures ->clock_task is
885 * It does however cause some slight miss-attribution of {soft,}irq
886 * time, a more accurate solution would be to update the irq_time using
887 * the current rq->clock timestamp, except that would require using
890 if (irq_delta > delta)
893 rq->prev_irq_time += irq_delta;
896 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
897 if (static_key_false((¶virt_steal_rq_enabled))) {
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
903 if (unlikely(steal > delta))
906 st = steal_ticks(steal);
907 steal = st * TICK_NSEC;
909 rq->prev_steal_time_rq += steal;
915 rq->clock_task += delta;
917 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
919 sched_rt_avg_update(rq, irq_delta + steal);
923 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
924 static int irqtime_account_hi_update(void)
926 u64 *cpustat = kcpustat_this_cpu->cpustat;
931 local_irq_save(flags);
932 latest_ns = this_cpu_read(cpu_hardirq_time);
933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ])
935 local_irq_restore(flags);
939 static int irqtime_account_si_update(void)
941 u64 *cpustat = kcpustat_this_cpu->cpustat;
946 local_irq_save(flags);
947 latest_ns = this_cpu_read(cpu_softirq_time);
948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ])
950 local_irq_restore(flags);
954 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
956 #define sched_clock_irqtime (0)
960 void sched_set_stop_task(int cpu, struct task_struct *stop)
962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
963 struct task_struct *old_stop = cpu_rq(cpu)->stop;
967 * Make it appear like a SCHED_FIFO task, its something
968 * userspace knows about and won't get confused about.
970 * Also, it will make PI more or less work without too
971 * much confusion -- but then, stop work should not
972 * rely on PI working anyway.
974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
976 stop->sched_class = &stop_sched_class;
979 cpu_rq(cpu)->stop = stop;
983 * Reset it back to a normal scheduling class so that
984 * it can die in pieces.
986 old_stop->sched_class = &rt_sched_class;
991 * __normal_prio - return the priority that is based on the static prio
993 static inline int __normal_prio(struct task_struct *p)
995 return p->static_prio;
999 * Calculate the expected normal priority: i.e. priority
1000 * without taking RT-inheritance into account. Might be
1001 * boosted by interactivity modifiers. Changes upon fork,
1002 * setprio syscalls, and whenever the interactivity
1003 * estimator recalculates.
1005 static inline int normal_prio(struct task_struct *p)
1009 if (task_has_rt_policy(p))
1010 prio = MAX_RT_PRIO-1 - p->rt_priority;
1012 prio = __normal_prio(p);
1017 * Calculate the current priority, i.e. the priority
1018 * taken into account by the scheduler. This value might
1019 * be boosted by RT tasks, or might be boosted by
1020 * interactivity modifiers. Will be RT if the task got
1021 * RT-boosted. If not then it returns p->normal_prio.
1023 static int effective_prio(struct task_struct *p)
1025 p->normal_prio = normal_prio(p);
1027 * If we are RT tasks or we were boosted to RT priority,
1028 * keep the priority unchanged. Otherwise, update priority
1029 * to the normal priority:
1031 if (!rt_prio(p->prio))
1032 return p->normal_prio;
1037 * task_curr - is this task currently executing on a CPU?
1038 * @p: the task in question.
1040 inline int task_curr(const struct task_struct *p)
1042 return cpu_curr(task_cpu(p)) == p;
1045 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1046 const struct sched_class *prev_class,
1049 if (prev_class != p->sched_class) {
1050 if (prev_class->switched_from)
1051 prev_class->switched_from(rq, p);
1052 p->sched_class->switched_to(rq, p);
1053 } else if (oldprio != p->prio)
1054 p->sched_class->prio_changed(rq, p, oldprio);
1057 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1059 const struct sched_class *class;
1061 if (p->sched_class == rq->curr->sched_class) {
1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1064 for_each_class(class) {
1065 if (class == rq->curr->sched_class)
1067 if (class == p->sched_class) {
1068 resched_task(rq->curr);
1075 * A queue event has occurred, and we're going to schedule. In
1076 * this case, we can save a useless back to back clock update.
1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1079 rq->skip_clock_update = 1;
1083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1085 #ifdef CONFIG_SCHED_DEBUG
1087 * We should never call set_task_cpu() on a blocked task,
1088 * ttwu() will sort out the placement.
1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
1093 #ifdef CONFIG_LOCKDEP
1095 * The caller should hold either p->pi_lock or rq->lock, when changing
1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1098 * sched_move_task() holds both and thus holding either pins the cgroup,
1101 * Furthermore, all task_rq users should acquire both locks, see
1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1105 lockdep_is_held(&task_rq(p)->lock)));
1109 trace_sched_migrate_task(p, new_cpu);
1111 if (task_cpu(p) != new_cpu) {
1112 p->se.nr_migrations++;
1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1116 __set_task_cpu(p, new_cpu);
1119 struct migration_arg {
1120 struct task_struct *task;
1124 static int migration_cpu_stop(void *data);
1127 * wait_task_inactive - wait for a thread to unschedule.
1129 * If @match_state is nonzero, it's the @p->state value just checked and
1130 * not expected to change. If it changes, i.e. @p might have woken up,
1131 * then return zero. When we succeed in waiting for @p to be off its CPU,
1132 * we return a positive number (its total switch count). If a second call
1133 * a short while later returns the same number, the caller can be sure that
1134 * @p has remained unscheduled the whole time.
1136 * The caller must ensure that the task *will* unschedule sometime soon,
1137 * else this function might spin for a *long* time. This function can't
1138 * be called with interrupts off, or it may introduce deadlock with
1139 * smp_call_function() if an IPI is sent by the same process we are
1140 * waiting to become inactive.
1142 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1144 unsigned long flags;
1151 * We do the initial early heuristics without holding
1152 * any task-queue locks at all. We'll only try to get
1153 * the runqueue lock when things look like they will
1159 * If the task is actively running on another CPU
1160 * still, just relax and busy-wait without holding
1163 * NOTE! Since we don't hold any locks, it's not
1164 * even sure that "rq" stays as the right runqueue!
1165 * But we don't care, since "task_running()" will
1166 * return false if the runqueue has changed and p
1167 * is actually now running somewhere else!
1169 while (task_running(rq, p)) {
1170 if (match_state && unlikely(p->state != match_state))
1176 * Ok, time to look more closely! We need the rq
1177 * lock now, to be *sure*. If we're wrong, we'll
1178 * just go back and repeat.
1180 rq = task_rq_lock(p, &flags);
1181 trace_sched_wait_task(p);
1182 running = task_running(rq, p);
1185 if (!match_state || p->state == match_state)
1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1187 task_rq_unlock(rq, p, &flags);
1190 * If it changed from the expected state, bail out now.
1192 if (unlikely(!ncsw))
1196 * Was it really running after all now that we
1197 * checked with the proper locks actually held?
1199 * Oops. Go back and try again..
1201 if (unlikely(running)) {
1207 * It's not enough that it's not actively running,
1208 * it must be off the runqueue _entirely_, and not
1211 * So if it was still runnable (but just not actively
1212 * running right now), it's preempted, and we should
1213 * yield - it could be a while.
1215 if (unlikely(on_rq)) {
1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1218 set_current_state(TASK_UNINTERRUPTIBLE);
1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1224 * Ahh, all good. It wasn't running, and it wasn't
1225 * runnable, which means that it will never become
1226 * running in the future either. We're all done!
1235 * kick_process - kick a running thread to enter/exit the kernel
1236 * @p: the to-be-kicked thread
1238 * Cause a process which is running on another CPU to enter
1239 * kernel-mode, without any delay. (to get signals handled.)
1241 * NOTE: this function doesn't have to take the runqueue lock,
1242 * because all it wants to ensure is that the remote task enters
1243 * the kernel. If the IPI races and the task has been migrated
1244 * to another CPU then no harm is done and the purpose has been
1247 void kick_process(struct task_struct *p)
1253 if ((cpu != smp_processor_id()) && task_curr(p))
1254 smp_send_reschedule(cpu);
1257 EXPORT_SYMBOL_GPL(kick_process);
1258 #endif /* CONFIG_SMP */
1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1264 static int select_fallback_rq(int cpu, struct task_struct *p)
1266 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1267 enum { cpuset, possible, fail } state = cpuset;
1270 /* Look for allowed, online CPU in same node. */
1271 for_each_cpu(dest_cpu, nodemask) {
1272 if (!cpu_online(dest_cpu))
1274 if (!cpu_active(dest_cpu))
1276 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1281 /* Any allowed, online CPU? */
1282 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1283 if (!cpu_online(dest_cpu))
1285 if (!cpu_active(dest_cpu))
1292 /* No more Mr. Nice Guy. */
1293 cpuset_cpus_allowed_fallback(p);
1298 do_set_cpus_allowed(p, cpu_possible_mask);
1309 if (state != cpuset) {
1311 * Don't tell them about moving exiting tasks or
1312 * kernel threads (both mm NULL), since they never
1315 if (p->mm && printk_ratelimit()) {
1316 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1317 task_pid_nr(p), p->comm, cpu);
1325 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1328 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1330 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1333 * In order not to call set_task_cpu() on a blocking task we need
1334 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1337 * Since this is common to all placement strategies, this lives here.
1339 * [ this allows ->select_task() to simply return task_cpu(p) and
1340 * not worry about this generic constraint ]
1342 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1344 cpu = select_fallback_rq(task_cpu(p), p);
1349 static void update_avg(u64 *avg, u64 sample)
1351 s64 diff = sample - *avg;
1357 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1359 #ifdef CONFIG_SCHEDSTATS
1360 struct rq *rq = this_rq();
1363 int this_cpu = smp_processor_id();
1365 if (cpu == this_cpu) {
1366 schedstat_inc(rq, ttwu_local);
1367 schedstat_inc(p, se.statistics.nr_wakeups_local);
1369 struct sched_domain *sd;
1371 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1373 for_each_domain(this_cpu, sd) {
1374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1375 schedstat_inc(sd, ttwu_wake_remote);
1382 if (wake_flags & WF_MIGRATED)
1383 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1385 #endif /* CONFIG_SMP */
1387 schedstat_inc(rq, ttwu_count);
1388 schedstat_inc(p, se.statistics.nr_wakeups);
1390 if (wake_flags & WF_SYNC)
1391 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1393 #endif /* CONFIG_SCHEDSTATS */
1396 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1398 activate_task(rq, p, en_flags);
1401 /* if a worker is waking up, notify workqueue */
1402 if (p->flags & PF_WQ_WORKER)
1403 wq_worker_waking_up(p, cpu_of(rq));
1407 * Mark the task runnable and perform wakeup-preemption.
1410 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1412 trace_sched_wakeup(p, true);
1413 check_preempt_curr(rq, p, wake_flags);
1415 p->state = TASK_RUNNING;
1417 if (p->sched_class->task_woken)
1418 p->sched_class->task_woken(rq, p);
1420 if (rq->idle_stamp) {
1421 u64 delta = rq->clock - rq->idle_stamp;
1422 u64 max = 2*sysctl_sched_migration_cost;
1427 update_avg(&rq->avg_idle, delta);
1434 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1437 if (p->sched_contributes_to_load)
1438 rq->nr_uninterruptible--;
1441 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1442 ttwu_do_wakeup(rq, p, wake_flags);
1446 * Called in case the task @p isn't fully descheduled from its runqueue,
1447 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1448 * since all we need to do is flip p->state to TASK_RUNNING, since
1449 * the task is still ->on_rq.
1451 static int ttwu_remote(struct task_struct *p, int wake_flags)
1456 rq = __task_rq_lock(p);
1458 ttwu_do_wakeup(rq, p, wake_flags);
1461 __task_rq_unlock(rq);
1467 static void sched_ttwu_pending(void)
1469 struct rq *rq = this_rq();
1470 struct llist_node *llist = llist_del_all(&rq->wake_list);
1471 struct task_struct *p;
1473 raw_spin_lock(&rq->lock);
1476 p = llist_entry(llist, struct task_struct, wake_entry);
1477 llist = llist_next(llist);
1478 ttwu_do_activate(rq, p, 0);
1481 raw_spin_unlock(&rq->lock);
1484 void scheduler_ipi(void)
1486 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1490 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1491 * traditionally all their work was done from the interrupt return
1492 * path. Now that we actually do some work, we need to make sure
1495 * Some archs already do call them, luckily irq_enter/exit nest
1498 * Arguably we should visit all archs and update all handlers,
1499 * however a fair share of IPIs are still resched only so this would
1500 * somewhat pessimize the simple resched case.
1503 sched_ttwu_pending();
1506 * Check if someone kicked us for doing the nohz idle load balance.
1508 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1509 this_rq()->idle_balance = 1;
1510 raise_softirq_irqoff(SCHED_SOFTIRQ);
1515 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1517 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1518 smp_send_reschedule(cpu);
1521 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1522 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
1527 rq = __task_rq_lock(p);
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1533 __task_rq_unlock(rq);
1538 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1540 bool cpus_share_cache(int this_cpu, int that_cpu)
1542 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1544 #endif /* CONFIG_SMP */
1546 static void ttwu_queue(struct task_struct *p, int cpu)
1548 struct rq *rq = cpu_rq(cpu);
1550 #if defined(CONFIG_SMP)
1551 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1552 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1553 ttwu_queue_remote(p, cpu);
1558 raw_spin_lock(&rq->lock);
1559 ttwu_do_activate(rq, p, 0);
1560 raw_spin_unlock(&rq->lock);
1564 * try_to_wake_up - wake up a thread
1565 * @p: the thread to be awakened
1566 * @state: the mask of task states that can be woken
1567 * @wake_flags: wake modifier flags (WF_*)
1569 * Put it on the run-queue if it's not already there. The "current"
1570 * thread is always on the run-queue (except when the actual
1571 * re-schedule is in progress), and as such you're allowed to do
1572 * the simpler "current->state = TASK_RUNNING" to mark yourself
1573 * runnable without the overhead of this.
1575 * Returns %true if @p was woken up, %false if it was already running
1576 * or @state didn't match @p's state.
1579 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1581 unsigned long flags;
1582 int cpu, success = 0;
1585 raw_spin_lock_irqsave(&p->pi_lock, flags);
1586 if (!(p->state & state))
1589 success = 1; /* we're going to change ->state */
1592 if (p->on_rq && ttwu_remote(p, wake_flags))
1597 * If the owning (remote) cpu is still in the middle of schedule() with
1598 * this task as prev, wait until its done referencing the task.
1601 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1603 * In case the architecture enables interrupts in
1604 * context_switch(), we cannot busy wait, since that
1605 * would lead to deadlocks when an interrupt hits and
1606 * tries to wake up @prev. So bail and do a complete
1609 if (ttwu_activate_remote(p, wake_flags))
1616 * Pairs with the smp_wmb() in finish_lock_switch().
1620 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1621 p->state = TASK_WAKING;
1623 if (p->sched_class->task_waking)
1624 p->sched_class->task_waking(p);
1626 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1627 if (task_cpu(p) != cpu) {
1628 wake_flags |= WF_MIGRATED;
1629 set_task_cpu(p, cpu);
1631 #endif /* CONFIG_SMP */
1635 ttwu_stat(p, cpu, wake_flags);
1637 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1643 * try_to_wake_up_local - try to wake up a local task with rq lock held
1644 * @p: the thread to be awakened
1646 * Put @p on the run-queue if it's not already there. The caller must
1647 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1650 static void try_to_wake_up_local(struct task_struct *p)
1652 struct rq *rq = task_rq(p);
1654 BUG_ON(rq != this_rq());
1655 BUG_ON(p == current);
1656 lockdep_assert_held(&rq->lock);
1658 if (!raw_spin_trylock(&p->pi_lock)) {
1659 raw_spin_unlock(&rq->lock);
1660 raw_spin_lock(&p->pi_lock);
1661 raw_spin_lock(&rq->lock);
1664 if (!(p->state & TASK_NORMAL))
1668 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1670 ttwu_do_wakeup(rq, p, 0);
1671 ttwu_stat(p, smp_processor_id(), 0);
1673 raw_spin_unlock(&p->pi_lock);
1677 * wake_up_process - Wake up a specific process
1678 * @p: The process to be woken up.
1680 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * processes. Returns 1 if the process was woken up, 0 if it was already
1684 * It may be assumed that this function implies a write memory barrier before
1685 * changing the task state if and only if any tasks are woken up.
1687 int wake_up_process(struct task_struct *p)
1689 return try_to_wake_up(p, TASK_ALL, 0);
1691 EXPORT_SYMBOL(wake_up_process);
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1695 return try_to_wake_up(p, state, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(struct task_struct *p)
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1720 INIT_LIST_HEAD(&p->rt.run_list);
1722 #ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1728 * fork()/clone()-time setup:
1730 void sched_fork(struct task_struct *p)
1732 unsigned long flags;
1733 int cpu = get_cpu();
1737 * We mark the process as running here. This guarantees that
1738 * nobody will actually run it, and a signal or other external
1739 * event cannot wake it up and insert it on the runqueue either.
1741 p->state = TASK_RUNNING;
1744 * Make sure we do not leak PI boosting priority to the child.
1746 p->prio = current->normal_prio;
1749 * Revert to default priority/policy on fork if requested.
1751 if (unlikely(p->sched_reset_on_fork)) {
1752 if (task_has_rt_policy(p)) {
1753 p->policy = SCHED_NORMAL;
1754 p->static_prio = NICE_TO_PRIO(0);
1756 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1757 p->static_prio = NICE_TO_PRIO(0);
1759 p->prio = p->normal_prio = __normal_prio(p);
1763 * We don't need the reset flag anymore after the fork. It has
1764 * fulfilled its duty:
1766 p->sched_reset_on_fork = 0;
1769 if (!rt_prio(p->prio))
1770 p->sched_class = &fair_sched_class;
1772 if (p->sched_class->task_fork)
1773 p->sched_class->task_fork(p);
1776 * The child is not yet in the pid-hash so no cgroup attach races,
1777 * and the cgroup is pinned to this child due to cgroup_fork()
1778 * is ran before sched_fork().
1780 * Silence PROVE_RCU.
1782 raw_spin_lock_irqsave(&p->pi_lock, flags);
1783 set_task_cpu(p, cpu);
1784 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1786 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1787 if (likely(sched_info_on()))
1788 memset(&p->sched_info, 0, sizeof(p->sched_info));
1790 #if defined(CONFIG_SMP)
1793 #ifdef CONFIG_PREEMPT_COUNT
1794 /* Want to start with kernel preemption disabled. */
1795 task_thread_info(p)->preempt_count = 1;
1798 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1805 * wake_up_new_task - wake up a newly created task for the first time.
1807 * This function will do some initial scheduler statistics housekeeping
1808 * that must be done for every newly created context, then puts the task
1809 * on the runqueue and wakes it.
1811 void wake_up_new_task(struct task_struct *p)
1813 unsigned long flags;
1816 raw_spin_lock_irqsave(&p->pi_lock, flags);
1819 * Fork balancing, do it here and not earlier because:
1820 * - cpus_allowed can change in the fork path
1821 * - any previously selected cpu might disappear through hotplug
1823 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1826 rq = __task_rq_lock(p);
1827 activate_task(rq, p, 0);
1829 trace_sched_wakeup_new(p, true);
1830 check_preempt_curr(rq, p, WF_FORK);
1832 if (p->sched_class->task_woken)
1833 p->sched_class->task_woken(rq, p);
1835 task_rq_unlock(rq, p, &flags);
1838 #ifdef CONFIG_PREEMPT_NOTIFIERS
1841 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1842 * @notifier: notifier struct to register
1844 void preempt_notifier_register(struct preempt_notifier *notifier)
1846 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1848 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1851 * preempt_notifier_unregister - no longer interested in preemption notifications
1852 * @notifier: notifier struct to unregister
1854 * This is safe to call from within a preemption notifier.
1856 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1858 hlist_del(¬ifier->link);
1860 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1862 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1864 struct preempt_notifier *notifier;
1865 struct hlist_node *node;
1867 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1868 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1872 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1873 struct task_struct *next)
1875 struct preempt_notifier *notifier;
1876 struct hlist_node *node;
1878 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1879 notifier->ops->sched_out(notifier, next);
1882 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1884 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1889 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1890 struct task_struct *next)
1894 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1897 * prepare_task_switch - prepare to switch tasks
1898 * @rq: the runqueue preparing to switch
1899 * @prev: the current task that is being switched out
1900 * @next: the task we are going to switch to.
1902 * This is called with the rq lock held and interrupts off. It must
1903 * be paired with a subsequent finish_task_switch after the context
1906 * prepare_task_switch sets up locking and calls architecture specific
1910 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1911 struct task_struct *next)
1913 trace_sched_switch(prev, next);
1914 sched_info_switch(prev, next);
1915 perf_event_task_sched_out(prev, next);
1916 fire_sched_out_preempt_notifiers(prev, next);
1917 prepare_lock_switch(rq, next);
1918 prepare_arch_switch(next);
1922 * finish_task_switch - clean up after a task-switch
1923 * @rq: runqueue associated with task-switch
1924 * @prev: the thread we just switched away from.
1926 * finish_task_switch must be called after the context switch, paired
1927 * with a prepare_task_switch call before the context switch.
1928 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1929 * and do any other architecture-specific cleanup actions.
1931 * Note that we may have delayed dropping an mm in context_switch(). If
1932 * so, we finish that here outside of the runqueue lock. (Doing it
1933 * with the lock held can cause deadlocks; see schedule() for
1936 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1937 __releases(rq->lock)
1939 struct mm_struct *mm = rq->prev_mm;
1945 * A task struct has one reference for the use as "current".
1946 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1947 * schedule one last time. The schedule call will never return, and
1948 * the scheduled task must drop that reference.
1949 * The test for TASK_DEAD must occur while the runqueue locks are
1950 * still held, otherwise prev could be scheduled on another cpu, die
1951 * there before we look at prev->state, and then the reference would
1953 * Manfred Spraul <manfred@colorfullife.com>
1955 prev_state = prev->state;
1956 finish_arch_switch(prev);
1957 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1958 local_irq_disable();
1959 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1960 perf_event_task_sched_in(prev, current);
1961 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1963 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
1964 finish_lock_switch(rq, prev);
1965 finish_arch_post_lock_switch();
1967 fire_sched_in_preempt_notifiers(current);
1970 if (unlikely(prev_state == TASK_DEAD)) {
1972 * Remove function-return probe instances associated with this
1973 * task and put them back on the free list.
1975 kprobe_flush_task(prev);
1976 put_task_struct(prev);
1982 /* assumes rq->lock is held */
1983 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1985 if (prev->sched_class->pre_schedule)
1986 prev->sched_class->pre_schedule(rq, prev);
1989 /* rq->lock is NOT held, but preemption is disabled */
1990 static inline void post_schedule(struct rq *rq)
1992 if (rq->post_schedule) {
1993 unsigned long flags;
1995 raw_spin_lock_irqsave(&rq->lock, flags);
1996 if (rq->curr->sched_class->post_schedule)
1997 rq->curr->sched_class->post_schedule(rq);
1998 raw_spin_unlock_irqrestore(&rq->lock, flags);
2000 rq->post_schedule = 0;
2006 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2010 static inline void post_schedule(struct rq *rq)
2017 * schedule_tail - first thing a freshly forked thread must call.
2018 * @prev: the thread we just switched away from.
2020 asmlinkage void schedule_tail(struct task_struct *prev)
2021 __releases(rq->lock)
2023 struct rq *rq = this_rq();
2025 finish_task_switch(rq, prev);
2028 * FIXME: do we need to worry about rq being invalidated by the
2033 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2034 /* In this case, finish_task_switch does not reenable preemption */
2037 if (current->set_child_tid)
2038 put_user(task_pid_vnr(current), current->set_child_tid);
2042 * context_switch - switch to the new MM and the new
2043 * thread's register state.
2046 context_switch(struct rq *rq, struct task_struct *prev,
2047 struct task_struct *next)
2049 struct mm_struct *mm, *oldmm;
2051 prepare_task_switch(rq, prev, next);
2054 oldmm = prev->active_mm;
2056 * For paravirt, this is coupled with an exit in switch_to to
2057 * combine the page table reload and the switch backend into
2060 arch_start_context_switch(prev);
2063 next->active_mm = oldmm;
2064 atomic_inc(&oldmm->mm_count);
2065 enter_lazy_tlb(oldmm, next);
2067 switch_mm(oldmm, mm, next);
2070 prev->active_mm = NULL;
2071 rq->prev_mm = oldmm;
2074 * Since the runqueue lock will be released by the next
2075 * task (which is an invalid locking op but in the case
2076 * of the scheduler it's an obvious special-case), so we
2077 * do an early lockdep release here:
2079 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2080 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2083 /* Here we just switch the register state and the stack. */
2084 rcu_switch(prev, next);
2085 switch_to(prev, next, prev);
2089 * this_rq must be evaluated again because prev may have moved
2090 * CPUs since it called schedule(), thus the 'rq' on its stack
2091 * frame will be invalid.
2093 finish_task_switch(this_rq(), prev);
2097 * nr_running, nr_uninterruptible and nr_context_switches:
2099 * externally visible scheduler statistics: current number of runnable
2100 * threads, current number of uninterruptible-sleeping threads, total
2101 * number of context switches performed since bootup.
2103 unsigned long nr_running(void)
2105 unsigned long i, sum = 0;
2107 for_each_online_cpu(i)
2108 sum += cpu_rq(i)->nr_running;
2113 unsigned long nr_uninterruptible(void)
2115 unsigned long i, sum = 0;
2117 for_each_possible_cpu(i)
2118 sum += cpu_rq(i)->nr_uninterruptible;
2121 * Since we read the counters lockless, it might be slightly
2122 * inaccurate. Do not allow it to go below zero though:
2124 if (unlikely((long)sum < 0))
2130 unsigned long long nr_context_switches(void)
2133 unsigned long long sum = 0;
2135 for_each_possible_cpu(i)
2136 sum += cpu_rq(i)->nr_switches;
2141 unsigned long nr_iowait(void)
2143 unsigned long i, sum = 0;
2145 for_each_possible_cpu(i)
2146 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2151 unsigned long nr_iowait_cpu(int cpu)
2153 struct rq *this = cpu_rq(cpu);
2154 return atomic_read(&this->nr_iowait);
2157 unsigned long this_cpu_load(void)
2159 struct rq *this = this_rq();
2160 return this->cpu_load[0];
2165 * Global load-average calculations
2167 * We take a distributed and async approach to calculating the global load-avg
2168 * in order to minimize overhead.
2170 * The global load average is an exponentially decaying average of nr_running +
2171 * nr_uninterruptible.
2173 * Once every LOAD_FREQ:
2176 * for_each_possible_cpu(cpu)
2177 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2179 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2181 * Due to a number of reasons the above turns in the mess below:
2183 * - for_each_possible_cpu() is prohibitively expensive on machines with
2184 * serious number of cpus, therefore we need to take a distributed approach
2185 * to calculating nr_active.
2187 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2188 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2190 * So assuming nr_active := 0 when we start out -- true per definition, we
2191 * can simply take per-cpu deltas and fold those into a global accumulate
2192 * to obtain the same result. See calc_load_fold_active().
2194 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2195 * across the machine, we assume 10 ticks is sufficient time for every
2196 * cpu to have completed this task.
2198 * This places an upper-bound on the IRQ-off latency of the machine. Then
2199 * again, being late doesn't loose the delta, just wrecks the sample.
2201 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2202 * this would add another cross-cpu cacheline miss and atomic operation
2203 * to the wakeup path. Instead we increment on whatever cpu the task ran
2204 * when it went into uninterruptible state and decrement on whatever cpu
2205 * did the wakeup. This means that only the sum of nr_uninterruptible over
2206 * all cpus yields the correct result.
2208 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2211 /* Variables and functions for calc_load */
2212 static atomic_long_t calc_load_tasks;
2213 static unsigned long calc_load_update;
2214 unsigned long avenrun[3];
2215 EXPORT_SYMBOL(avenrun); /* should be removed */
2218 * get_avenrun - get the load average array
2219 * @loads: pointer to dest load array
2220 * @offset: offset to add
2221 * @shift: shift count to shift the result left
2223 * These values are estimates at best, so no need for locking.
2225 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2227 loads[0] = (avenrun[0] + offset) << shift;
2228 loads[1] = (avenrun[1] + offset) << shift;
2229 loads[2] = (avenrun[2] + offset) << shift;
2232 static long calc_load_fold_active(struct rq *this_rq)
2234 long nr_active, delta = 0;
2236 nr_active = this_rq->nr_running;
2237 nr_active += (long) this_rq->nr_uninterruptible;
2239 if (nr_active != this_rq->calc_load_active) {
2240 delta = nr_active - this_rq->calc_load_active;
2241 this_rq->calc_load_active = nr_active;
2248 * a1 = a0 * e + a * (1 - e)
2250 static unsigned long
2251 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2254 load += active * (FIXED_1 - exp);
2255 load += 1UL << (FSHIFT - 1);
2256 return load >> FSHIFT;
2261 * Handle NO_HZ for the global load-average.
2263 * Since the above described distributed algorithm to compute the global
2264 * load-average relies on per-cpu sampling from the tick, it is affected by
2267 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2268 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2269 * when we read the global state.
2271 * Obviously reality has to ruin such a delightfully simple scheme:
2273 * - When we go NO_HZ idle during the window, we can negate our sample
2274 * contribution, causing under-accounting.
2276 * We avoid this by keeping two idle-delta counters and flipping them
2277 * when the window starts, thus separating old and new NO_HZ load.
2279 * The only trick is the slight shift in index flip for read vs write.
2283 * |-|-----------|-|-----------|-|-----------|-|
2284 * r:0 0 1 1 0 0 1 1 0
2285 * w:0 1 1 0 0 1 1 0 0
2287 * This ensures we'll fold the old idle contribution in this window while
2288 * accumlating the new one.
2290 * - When we wake up from NO_HZ idle during the window, we push up our
2291 * contribution, since we effectively move our sample point to a known
2294 * This is solved by pushing the window forward, and thus skipping the
2295 * sample, for this cpu (effectively using the idle-delta for this cpu which
2296 * was in effect at the time the window opened). This also solves the issue
2297 * of having to deal with a cpu having been in NOHZ idle for multiple
2298 * LOAD_FREQ intervals.
2300 * When making the ILB scale, we should try to pull this in as well.
2302 static atomic_long_t calc_load_idle[2];
2303 static int calc_load_idx;
2305 static inline int calc_load_write_idx(void)
2307 int idx = calc_load_idx;
2310 * See calc_global_nohz(), if we observe the new index, we also
2311 * need to observe the new update time.
2316 * If the folding window started, make sure we start writing in the
2319 if (!time_before(jiffies, calc_load_update))
2325 static inline int calc_load_read_idx(void)
2327 return calc_load_idx & 1;
2330 void calc_load_enter_idle(void)
2332 struct rq *this_rq = this_rq();
2336 * We're going into NOHZ mode, if there's any pending delta, fold it
2337 * into the pending idle delta.
2339 delta = calc_load_fold_active(this_rq);
2341 int idx = calc_load_write_idx();
2342 atomic_long_add(delta, &calc_load_idle[idx]);
2346 void calc_load_exit_idle(void)
2348 struct rq *this_rq = this_rq();
2351 * If we're still before the sample window, we're done.
2353 if (time_before(jiffies, this_rq->calc_load_update))
2357 * We woke inside or after the sample window, this means we're already
2358 * accounted through the nohz accounting, so skip the entire deal and
2359 * sync up for the next window.
2361 this_rq->calc_load_update = calc_load_update;
2362 if (time_before(jiffies, this_rq->calc_load_update + 10))
2363 this_rq->calc_load_update += LOAD_FREQ;
2366 static long calc_load_fold_idle(void)
2368 int idx = calc_load_read_idx();
2371 if (atomic_long_read(&calc_load_idle[idx]))
2372 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2378 * fixed_power_int - compute: x^n, in O(log n) time
2380 * @x: base of the power
2381 * @frac_bits: fractional bits of @x
2382 * @n: power to raise @x to.
2384 * By exploiting the relation between the definition of the natural power
2385 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2386 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2387 * (where: n_i \elem {0, 1}, the binary vector representing n),
2388 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2389 * of course trivially computable in O(log_2 n), the length of our binary
2392 static unsigned long
2393 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2395 unsigned long result = 1UL << frac_bits;
2400 result += 1UL << (frac_bits - 1);
2401 result >>= frac_bits;
2407 x += 1UL << (frac_bits - 1);
2415 * a1 = a0 * e + a * (1 - e)
2417 * a2 = a1 * e + a * (1 - e)
2418 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2419 * = a0 * e^2 + a * (1 - e) * (1 + e)
2421 * a3 = a2 * e + a * (1 - e)
2422 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2423 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2427 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2428 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2429 * = a0 * e^n + a * (1 - e^n)
2431 * [1] application of the geometric series:
2434 * S_n := \Sum x^i = -------------
2437 static unsigned long
2438 calc_load_n(unsigned long load, unsigned long exp,
2439 unsigned long active, unsigned int n)
2442 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2446 * NO_HZ can leave us missing all per-cpu ticks calling
2447 * calc_load_account_active(), but since an idle CPU folds its delta into
2448 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2449 * in the pending idle delta if our idle period crossed a load cycle boundary.
2451 * Once we've updated the global active value, we need to apply the exponential
2452 * weights adjusted to the number of cycles missed.
2454 static void calc_global_nohz(void)
2456 long delta, active, n;
2458 if (!time_before(jiffies, calc_load_update + 10)) {
2460 * Catch-up, fold however many we are behind still
2462 delta = jiffies - calc_load_update - 10;
2463 n = 1 + (delta / LOAD_FREQ);
2465 active = atomic_long_read(&calc_load_tasks);
2466 active = active > 0 ? active * FIXED_1 : 0;
2468 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2469 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2470 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2472 calc_load_update += n * LOAD_FREQ;
2476 * Flip the idle index...
2478 * Make sure we first write the new time then flip the index, so that
2479 * calc_load_write_idx() will see the new time when it reads the new
2480 * index, this avoids a double flip messing things up.
2485 #else /* !CONFIG_NO_HZ */
2487 static inline long calc_load_fold_idle(void) { return 0; }
2488 static inline void calc_global_nohz(void) { }
2490 #endif /* CONFIG_NO_HZ */
2493 * calc_load - update the avenrun load estimates 10 ticks after the
2494 * CPUs have updated calc_load_tasks.
2496 void calc_global_load(unsigned long ticks)
2500 if (time_before(jiffies, calc_load_update + 10))
2504 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2506 delta = calc_load_fold_idle();
2508 atomic_long_add(delta, &calc_load_tasks);
2510 active = atomic_long_read(&calc_load_tasks);
2511 active = active > 0 ? active * FIXED_1 : 0;
2513 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2514 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2515 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2517 calc_load_update += LOAD_FREQ;
2520 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2526 * Called from update_cpu_load() to periodically update this CPU's
2529 static void calc_load_account_active(struct rq *this_rq)
2533 if (time_before(jiffies, this_rq->calc_load_update))
2536 delta = calc_load_fold_active(this_rq);
2538 atomic_long_add(delta, &calc_load_tasks);
2540 this_rq->calc_load_update += LOAD_FREQ;
2544 * End of global load-average stuff
2548 * The exact cpuload at various idx values, calculated at every tick would be
2549 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2551 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2552 * on nth tick when cpu may be busy, then we have:
2553 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2554 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2556 * decay_load_missed() below does efficient calculation of
2557 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2558 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2560 * The calculation is approximated on a 128 point scale.
2561 * degrade_zero_ticks is the number of ticks after which load at any
2562 * particular idx is approximated to be zero.
2563 * degrade_factor is a precomputed table, a row for each load idx.
2564 * Each column corresponds to degradation factor for a power of two ticks,
2565 * based on 128 point scale.
2567 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2568 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2570 * With this power of 2 load factors, we can degrade the load n times
2571 * by looking at 1 bits in n and doing as many mult/shift instead of
2572 * n mult/shifts needed by the exact degradation.
2574 #define DEGRADE_SHIFT 7
2575 static const unsigned char
2576 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2577 static const unsigned char
2578 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2579 {0, 0, 0, 0, 0, 0, 0, 0},
2580 {64, 32, 8, 0, 0, 0, 0, 0},
2581 {96, 72, 40, 12, 1, 0, 0},
2582 {112, 98, 75, 43, 15, 1, 0},
2583 {120, 112, 98, 76, 45, 16, 2} };
2586 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2587 * would be when CPU is idle and so we just decay the old load without
2588 * adding any new load.
2590 static unsigned long
2591 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2595 if (!missed_updates)
2598 if (missed_updates >= degrade_zero_ticks[idx])
2602 return load >> missed_updates;
2604 while (missed_updates) {
2605 if (missed_updates % 2)
2606 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2608 missed_updates >>= 1;
2615 * Update rq->cpu_load[] statistics. This function is usually called every
2616 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2617 * every tick. We fix it up based on jiffies.
2619 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2620 unsigned long pending_updates)
2624 this_rq->nr_load_updates++;
2626 /* Update our load: */
2627 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2628 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2629 unsigned long old_load, new_load;
2631 /* scale is effectively 1 << i now, and >> i divides by scale */
2633 old_load = this_rq->cpu_load[i];
2634 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2635 new_load = this_load;
2637 * Round up the averaging division if load is increasing. This
2638 * prevents us from getting stuck on 9 if the load is 10, for
2641 if (new_load > old_load)
2642 new_load += scale - 1;
2644 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2647 sched_avg_update(this_rq);
2652 * There is no sane way to deal with nohz on smp when using jiffies because the
2653 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2654 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2656 * Therefore we cannot use the delta approach from the regular tick since that
2657 * would seriously skew the load calculation. However we'll make do for those
2658 * updates happening while idle (nohz_idle_balance) or coming out of idle
2659 * (tick_nohz_idle_exit).
2661 * This means we might still be one tick off for nohz periods.
2665 * Called from nohz_idle_balance() to update the load ratings before doing the
2668 void update_idle_cpu_load(struct rq *this_rq)
2670 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2671 unsigned long load = this_rq->load.weight;
2672 unsigned long pending_updates;
2675 * bail if there's load or we're actually up-to-date.
2677 if (load || curr_jiffies == this_rq->last_load_update_tick)
2680 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2681 this_rq->last_load_update_tick = curr_jiffies;
2683 __update_cpu_load(this_rq, load, pending_updates);
2687 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2689 void update_cpu_load_nohz(void)
2691 struct rq *this_rq = this_rq();
2692 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2693 unsigned long pending_updates;
2695 if (curr_jiffies == this_rq->last_load_update_tick)
2698 raw_spin_lock(&this_rq->lock);
2699 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2700 if (pending_updates) {
2701 this_rq->last_load_update_tick = curr_jiffies;
2703 * We were idle, this means load 0, the current load might be
2704 * !0 due to remote wakeups and the sort.
2706 __update_cpu_load(this_rq, 0, pending_updates);
2708 raw_spin_unlock(&this_rq->lock);
2710 #endif /* CONFIG_NO_HZ */
2713 * Called from scheduler_tick()
2715 static void update_cpu_load_active(struct rq *this_rq)
2718 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2720 this_rq->last_load_update_tick = jiffies;
2721 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2723 calc_load_account_active(this_rq);
2729 * sched_exec - execve() is a valuable balancing opportunity, because at
2730 * this point the task has the smallest effective memory and cache footprint.
2732 void sched_exec(void)
2734 struct task_struct *p = current;
2735 unsigned long flags;
2738 raw_spin_lock_irqsave(&p->pi_lock, flags);
2739 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2740 if (dest_cpu == smp_processor_id())
2743 if (likely(cpu_active(dest_cpu))) {
2744 struct migration_arg arg = { p, dest_cpu };
2746 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2747 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2751 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2756 DEFINE_PER_CPU(struct kernel_stat, kstat);
2757 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2759 EXPORT_PER_CPU_SYMBOL(kstat);
2760 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2763 * Return any ns on the sched_clock that have not yet been accounted in
2764 * @p in case that task is currently running.
2766 * Called with task_rq_lock() held on @rq.
2768 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2772 if (task_current(rq, p)) {
2773 update_rq_clock(rq);
2774 ns = rq->clock_task - p->se.exec_start;
2782 unsigned long long task_delta_exec(struct task_struct *p)
2784 unsigned long flags;
2788 rq = task_rq_lock(p, &flags);
2789 ns = do_task_delta_exec(p, rq);
2790 task_rq_unlock(rq, p, &flags);
2796 * Return accounted runtime for the task.
2797 * In case the task is currently running, return the runtime plus current's
2798 * pending runtime that have not been accounted yet.
2800 unsigned long long task_sched_runtime(struct task_struct *p)
2802 unsigned long flags;
2806 rq = task_rq_lock(p, &flags);
2807 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2808 task_rq_unlock(rq, p, &flags);
2813 #ifdef CONFIG_CGROUP_CPUACCT
2814 struct cgroup_subsys cpuacct_subsys;
2815 struct cpuacct root_cpuacct;
2818 static inline void task_group_account_field(struct task_struct *p, int index,
2821 #ifdef CONFIG_CGROUP_CPUACCT
2822 struct kernel_cpustat *kcpustat;
2826 * Since all updates are sure to touch the root cgroup, we
2827 * get ourselves ahead and touch it first. If the root cgroup
2828 * is the only cgroup, then nothing else should be necessary.
2831 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2833 #ifdef CONFIG_CGROUP_CPUACCT
2834 if (unlikely(!cpuacct_subsys.active))
2839 while (ca && (ca != &root_cpuacct)) {
2840 kcpustat = this_cpu_ptr(ca->cpustat);
2841 kcpustat->cpustat[index] += tmp;
2850 * Account user cpu time to a process.
2851 * @p: the process that the cpu time gets accounted to
2852 * @cputime: the cpu time spent in user space since the last update
2853 * @cputime_scaled: cputime scaled by cpu frequency
2855 void account_user_time(struct task_struct *p, cputime_t cputime,
2856 cputime_t cputime_scaled)
2860 /* Add user time to process. */
2861 p->utime += cputime;
2862 p->utimescaled += cputime_scaled;
2863 account_group_user_time(p, cputime);
2865 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2867 /* Add user time to cpustat. */
2868 task_group_account_field(p, index, (__force u64) cputime);
2870 /* Account for user time used */
2871 acct_update_integrals(p);
2875 * Account guest cpu time to a process.
2876 * @p: the process that the cpu time gets accounted to
2877 * @cputime: the cpu time spent in virtual machine since the last update
2878 * @cputime_scaled: cputime scaled by cpu frequency
2880 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2881 cputime_t cputime_scaled)
2883 u64 *cpustat = kcpustat_this_cpu->cpustat;
2885 /* Add guest time to process. */
2886 p->utime += cputime;
2887 p->utimescaled += cputime_scaled;
2888 account_group_user_time(p, cputime);
2889 p->gtime += cputime;
2891 /* Add guest time to cpustat. */
2892 if (TASK_NICE(p) > 0) {
2893 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2894 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2896 cpustat[CPUTIME_USER] += (__force u64) cputime;
2897 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2902 * Account system cpu time to a process and desired cpustat field
2903 * @p: the process that the cpu time gets accounted to
2904 * @cputime: the cpu time spent in kernel space since the last update
2905 * @cputime_scaled: cputime scaled by cpu frequency
2906 * @target_cputime64: pointer to cpustat field that has to be updated
2909 void __account_system_time(struct task_struct *p, cputime_t cputime,
2910 cputime_t cputime_scaled, int index)
2912 /* Add system time to process. */
2913 p->stime += cputime;
2914 p->stimescaled += cputime_scaled;
2915 account_group_system_time(p, cputime);
2917 /* Add system time to cpustat. */
2918 task_group_account_field(p, index, (__force u64) cputime);
2920 /* Account for system time used */
2921 acct_update_integrals(p);
2925 * Account system cpu time to a process.
2926 * @p: the process that the cpu time gets accounted to
2927 * @hardirq_offset: the offset to subtract from hardirq_count()
2928 * @cputime: the cpu time spent in kernel space since the last update
2929 * @cputime_scaled: cputime scaled by cpu frequency
2931 void account_system_time(struct task_struct *p, int hardirq_offset,
2932 cputime_t cputime, cputime_t cputime_scaled)
2936 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2937 account_guest_time(p, cputime, cputime_scaled);
2941 if (hardirq_count() - hardirq_offset)
2942 index = CPUTIME_IRQ;
2943 else if (in_serving_softirq())
2944 index = CPUTIME_SOFTIRQ;
2946 index = CPUTIME_SYSTEM;
2948 __account_system_time(p, cputime, cputime_scaled, index);
2952 * Account for involuntary wait time.
2953 * @cputime: the cpu time spent in involuntary wait
2955 void account_steal_time(cputime_t cputime)
2957 u64 *cpustat = kcpustat_this_cpu->cpustat;
2959 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2963 * Account for idle time.
2964 * @cputime: the cpu time spent in idle wait
2966 void account_idle_time(cputime_t cputime)
2968 u64 *cpustat = kcpustat_this_cpu->cpustat;
2969 struct rq *rq = this_rq();
2971 if (atomic_read(&rq->nr_iowait) > 0)
2972 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2974 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2977 static __always_inline bool steal_account_process_tick(void)
2979 #ifdef CONFIG_PARAVIRT
2980 if (static_key_false(¶virt_steal_enabled)) {
2983 steal = paravirt_steal_clock(smp_processor_id());
2984 steal -= this_rq()->prev_steal_time;
2986 st = steal_ticks(steal);
2987 this_rq()->prev_steal_time += st * TICK_NSEC;
2989 account_steal_time(st);
2996 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2998 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3000 * Account a tick to a process and cpustat
3001 * @p: the process that the cpu time gets accounted to
3002 * @user_tick: is the tick from userspace
3003 * @rq: the pointer to rq
3005 * Tick demultiplexing follows the order
3006 * - pending hardirq update
3007 * - pending softirq update
3011 * - check for guest_time
3012 * - else account as system_time
3014 * Check for hardirq is done both for system and user time as there is
3015 * no timer going off while we are on hardirq and hence we may never get an
3016 * opportunity to update it solely in system time.
3017 * p->stime and friends are only updated on system time and not on irq
3018 * softirq as those do not count in task exec_runtime any more.
3020 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3023 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3024 u64 *cpustat = kcpustat_this_cpu->cpustat;
3026 if (steal_account_process_tick())
3029 if (irqtime_account_hi_update()) {
3030 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3031 } else if (irqtime_account_si_update()) {
3032 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3033 } else if (this_cpu_ksoftirqd() == p) {
3035 * ksoftirqd time do not get accounted in cpu_softirq_time.
3036 * So, we have to handle it separately here.
3037 * Also, p->stime needs to be updated for ksoftirqd.
3039 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3041 } else if (user_tick) {
3042 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3043 } else if (p == rq->idle) {
3044 account_idle_time(cputime_one_jiffy);
3045 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3046 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3048 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3053 static void irqtime_account_idle_ticks(int ticks)
3056 struct rq *rq = this_rq();
3058 for (i = 0; i < ticks; i++)
3059 irqtime_account_process_tick(current, 0, rq);
3061 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3062 static void irqtime_account_idle_ticks(int ticks) {}
3063 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3065 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3068 * Account a single tick of cpu time.
3069 * @p: the process that the cpu time gets accounted to
3070 * @user_tick: indicates if the tick is a user or a system tick
3072 void account_process_tick(struct task_struct *p, int user_tick)
3074 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3075 struct rq *rq = this_rq();
3077 if (sched_clock_irqtime) {
3078 irqtime_account_process_tick(p, user_tick, rq);
3082 if (steal_account_process_tick())
3086 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3087 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3088 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3091 account_idle_time(cputime_one_jiffy);
3095 * Account multiple ticks of steal time.
3096 * @p: the process from which the cpu time has been stolen
3097 * @ticks: number of stolen ticks
3099 void account_steal_ticks(unsigned long ticks)
3101 account_steal_time(jiffies_to_cputime(ticks));
3105 * Account multiple ticks of idle time.
3106 * @ticks: number of stolen ticks
3108 void account_idle_ticks(unsigned long ticks)
3111 if (sched_clock_irqtime) {
3112 irqtime_account_idle_ticks(ticks);
3116 account_idle_time(jiffies_to_cputime(ticks));
3122 * Use precise platform statistics if available:
3124 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3125 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3131 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3133 struct task_cputime cputime;
3135 thread_group_cputime(p, &cputime);
3137 *ut = cputime.utime;
3138 *st = cputime.stime;
3142 #ifndef nsecs_to_cputime
3143 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3146 static cputime_t scale_utime(cputime_t utime, cputime_t rtime, cputime_t total)
3148 u64 temp = (__force u64) rtime;
3150 temp *= (__force u64) utime;
3152 if (sizeof(cputime_t) == 4)
3153 temp = div_u64(temp, (__force u32) total);
3155 temp = div64_u64(temp, (__force u64) total);
3157 return (__force cputime_t) temp;
3160 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3162 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3165 * Use CFS's precise accounting:
3167 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3170 utime = scale_utime(utime, rtime, total);
3175 * Compare with previous values, to keep monotonicity:
3177 p->prev_utime = max(p->prev_utime, utime);
3178 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3180 *ut = p->prev_utime;
3181 *st = p->prev_stime;
3185 * Must be called with siglock held.
3187 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3189 struct signal_struct *sig = p->signal;
3190 struct task_cputime cputime;
3191 cputime_t rtime, utime, total;
3193 thread_group_cputime(p, &cputime);
3195 total = cputime.utime + cputime.stime;
3196 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3199 utime = scale_utime(cputime.utime, rtime, total);
3203 sig->prev_utime = max(sig->prev_utime, utime);
3204 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3206 *ut = sig->prev_utime;
3207 *st = sig->prev_stime;
3212 * This function gets called by the timer code, with HZ frequency.
3213 * We call it with interrupts disabled.
3215 void scheduler_tick(void)
3217 int cpu = smp_processor_id();
3218 struct rq *rq = cpu_rq(cpu);
3219 struct task_struct *curr = rq->curr;
3223 raw_spin_lock(&rq->lock);
3224 update_rq_clock(rq);
3225 update_cpu_load_active(rq);
3226 curr->sched_class->task_tick(rq, curr, 0);
3227 raw_spin_unlock(&rq->lock);
3229 perf_event_task_tick();
3232 rq->idle_balance = idle_cpu(cpu);
3233 trigger_load_balance(rq, cpu);
3237 notrace unsigned long get_parent_ip(unsigned long addr)
3239 if (in_lock_functions(addr)) {
3240 addr = CALLER_ADDR2;
3241 if (in_lock_functions(addr))
3242 addr = CALLER_ADDR3;
3247 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3248 defined(CONFIG_PREEMPT_TRACER))
3250 void __kprobes add_preempt_count(int val)
3252 #ifdef CONFIG_DEBUG_PREEMPT
3256 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3259 preempt_count() += val;
3260 #ifdef CONFIG_DEBUG_PREEMPT
3262 * Spinlock count overflowing soon?
3264 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3267 if (preempt_count() == val)
3268 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3270 EXPORT_SYMBOL(add_preempt_count);
3272 void __kprobes sub_preempt_count(int val)
3274 #ifdef CONFIG_DEBUG_PREEMPT
3278 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3281 * Is the spinlock portion underflowing?
3283 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3284 !(preempt_count() & PREEMPT_MASK)))
3288 if (preempt_count() == val)
3289 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3290 preempt_count() -= val;
3292 EXPORT_SYMBOL(sub_preempt_count);
3297 * Print scheduling while atomic bug:
3299 static noinline void __schedule_bug(struct task_struct *prev)
3301 if (oops_in_progress)
3304 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3305 prev->comm, prev->pid, preempt_count());
3307 debug_show_held_locks(prev);
3309 if (irqs_disabled())
3310 print_irqtrace_events(prev);
3312 add_taint(TAINT_WARN);
3316 * Various schedule()-time debugging checks and statistics:
3318 static inline void schedule_debug(struct task_struct *prev)
3321 * Test if we are atomic. Since do_exit() needs to call into
3322 * schedule() atomically, we ignore that path for now.
3323 * Otherwise, whine if we are scheduling when we should not be.
3325 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3326 __schedule_bug(prev);
3329 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3331 schedstat_inc(this_rq(), sched_count);
3334 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3336 if (prev->on_rq || rq->skip_clock_update < 0)
3337 update_rq_clock(rq);
3338 prev->sched_class->put_prev_task(rq, prev);
3342 * Pick up the highest-prio task:
3344 static inline struct task_struct *
3345 pick_next_task(struct rq *rq)
3347 const struct sched_class *class;
3348 struct task_struct *p;
3351 * Optimization: we know that if all tasks are in
3352 * the fair class we can call that function directly:
3354 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3355 p = fair_sched_class.pick_next_task(rq);
3360 for_each_class(class) {
3361 p = class->pick_next_task(rq);
3366 BUG(); /* the idle class will always have a runnable task */
3370 * __schedule() is the main scheduler function.
3372 static void __sched __schedule(void)
3374 struct task_struct *prev, *next;
3375 unsigned long *switch_count;
3381 cpu = smp_processor_id();
3383 rcu_note_context_switch(cpu);
3386 schedule_debug(prev);
3388 if (sched_feat(HRTICK))
3391 raw_spin_lock_irq(&rq->lock);
3393 switch_count = &prev->nivcsw;
3394 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3395 if (unlikely(signal_pending_state(prev->state, prev))) {
3396 prev->state = TASK_RUNNING;
3398 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3402 * If a worker went to sleep, notify and ask workqueue
3403 * whether it wants to wake up a task to maintain
3406 if (prev->flags & PF_WQ_WORKER) {
3407 struct task_struct *to_wakeup;
3409 to_wakeup = wq_worker_sleeping(prev, cpu);
3411 try_to_wake_up_local(to_wakeup);
3414 switch_count = &prev->nvcsw;
3417 pre_schedule(rq, prev);
3419 if (unlikely(!rq->nr_running))
3420 idle_balance(cpu, rq);
3422 put_prev_task(rq, prev);
3423 next = pick_next_task(rq);
3424 clear_tsk_need_resched(prev);
3425 rq->skip_clock_update = 0;
3427 if (likely(prev != next)) {
3432 context_switch(rq, prev, next); /* unlocks the rq */
3434 * The context switch have flipped the stack from under us
3435 * and restored the local variables which were saved when
3436 * this task called schedule() in the past. prev == current
3437 * is still correct, but it can be moved to another cpu/rq.
3439 cpu = smp_processor_id();
3442 raw_spin_unlock_irq(&rq->lock);
3446 sched_preempt_enable_no_resched();
3451 static inline void sched_submit_work(struct task_struct *tsk)
3453 if (!tsk->state || tsk_is_pi_blocked(tsk))
3456 * If we are going to sleep and we have plugged IO queued,
3457 * make sure to submit it to avoid deadlocks.
3459 if (blk_needs_flush_plug(tsk))
3460 blk_schedule_flush_plug(tsk);
3463 asmlinkage void __sched schedule(void)
3465 struct task_struct *tsk = current;
3467 sched_submit_work(tsk);
3470 EXPORT_SYMBOL(schedule);
3472 #ifdef CONFIG_RCU_USER_QS
3473 asmlinkage void __sched schedule_user(void)
3476 * If we come here after a random call to set_need_resched(),
3477 * or we have been woken up remotely but the IPI has not yet arrived,
3478 * we haven't yet exited the RCU idle mode. Do it here manually until
3479 * we find a better solution.
3488 * schedule_preempt_disabled - called with preemption disabled
3490 * Returns with preemption disabled. Note: preempt_count must be 1
3492 void __sched schedule_preempt_disabled(void)
3494 sched_preempt_enable_no_resched();
3499 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3501 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3503 if (lock->owner != owner)
3507 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3508 * lock->owner still matches owner, if that fails, owner might
3509 * point to free()d memory, if it still matches, the rcu_read_lock()
3510 * ensures the memory stays valid.
3514 return owner->on_cpu;
3518 * Look out! "owner" is an entirely speculative pointer
3519 * access and not reliable.
3521 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3523 if (!sched_feat(OWNER_SPIN))
3527 while (owner_running(lock, owner)) {
3531 arch_mutex_cpu_relax();
3536 * We break out the loop above on need_resched() and when the
3537 * owner changed, which is a sign for heavy contention. Return
3538 * success only when lock->owner is NULL.
3540 return lock->owner == NULL;
3544 #ifdef CONFIG_PREEMPT
3546 * this is the entry point to schedule() from in-kernel preemption
3547 * off of preempt_enable. Kernel preemptions off return from interrupt
3548 * occur there and call schedule directly.
3550 asmlinkage void __sched notrace preempt_schedule(void)
3552 struct thread_info *ti = current_thread_info();
3555 * If there is a non-zero preempt_count or interrupts are disabled,
3556 * we do not want to preempt the current task. Just return..
3558 if (likely(ti->preempt_count || irqs_disabled()))
3562 add_preempt_count_notrace(PREEMPT_ACTIVE);
3564 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3567 * Check again in case we missed a preemption opportunity
3568 * between schedule and now.
3571 } while (need_resched());
3573 EXPORT_SYMBOL(preempt_schedule);
3576 * this is the entry point to schedule() from kernel preemption
3577 * off of irq context.
3578 * Note, that this is called and return with irqs disabled. This will
3579 * protect us against recursive calling from irq.
3581 asmlinkage void __sched preempt_schedule_irq(void)
3583 struct thread_info *ti = current_thread_info();
3585 /* Catch callers which need to be fixed */
3586 BUG_ON(ti->preempt_count || !irqs_disabled());
3590 add_preempt_count(PREEMPT_ACTIVE);
3593 local_irq_disable();
3594 sub_preempt_count(PREEMPT_ACTIVE);
3597 * Check again in case we missed a preemption opportunity
3598 * between schedule and now.
3601 } while (need_resched());
3604 #endif /* CONFIG_PREEMPT */
3606 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3609 return try_to_wake_up(curr->private, mode, wake_flags);
3611 EXPORT_SYMBOL(default_wake_function);
3614 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3615 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3616 * number) then we wake all the non-exclusive tasks and one exclusive task.
3618 * There are circumstances in which we can try to wake a task which has already
3619 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3620 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3622 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3623 int nr_exclusive, int wake_flags, void *key)
3625 wait_queue_t *curr, *next;
3627 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3628 unsigned flags = curr->flags;
3630 if (curr->func(curr, mode, wake_flags, key) &&
3631 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3637 * __wake_up - wake up threads blocked on a waitqueue.
3639 * @mode: which threads
3640 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3641 * @key: is directly passed to the wakeup function
3643 * It may be assumed that this function implies a write memory barrier before
3644 * changing the task state if and only if any tasks are woken up.
3646 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3647 int nr_exclusive, void *key)
3649 unsigned long flags;
3651 spin_lock_irqsave(&q->lock, flags);
3652 __wake_up_common(q, mode, nr_exclusive, 0, key);
3653 spin_unlock_irqrestore(&q->lock, flags);
3655 EXPORT_SYMBOL(__wake_up);
3658 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3660 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3662 __wake_up_common(q, mode, nr, 0, NULL);
3664 EXPORT_SYMBOL_GPL(__wake_up_locked);
3666 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3668 __wake_up_common(q, mode, 1, 0, key);
3670 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3673 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3675 * @mode: which threads
3676 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3677 * @key: opaque value to be passed to wakeup targets
3679 * The sync wakeup differs that the waker knows that it will schedule
3680 * away soon, so while the target thread will be woken up, it will not
3681 * be migrated to another CPU - ie. the two threads are 'synchronized'
3682 * with each other. This can prevent needless bouncing between CPUs.
3684 * On UP it can prevent extra preemption.
3686 * It may be assumed that this function implies a write memory barrier before
3687 * changing the task state if and only if any tasks are woken up.
3689 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3690 int nr_exclusive, void *key)
3692 unsigned long flags;
3693 int wake_flags = WF_SYNC;
3698 if (unlikely(!nr_exclusive))
3701 spin_lock_irqsave(&q->lock, flags);
3702 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3703 spin_unlock_irqrestore(&q->lock, flags);
3705 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3708 * __wake_up_sync - see __wake_up_sync_key()
3710 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3712 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3714 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3717 * complete: - signals a single thread waiting on this completion
3718 * @x: holds the state of this particular completion
3720 * This will wake up a single thread waiting on this completion. Threads will be
3721 * awakened in the same order in which they were queued.
3723 * See also complete_all(), wait_for_completion() and related routines.
3725 * It may be assumed that this function implies a write memory barrier before
3726 * changing the task state if and only if any tasks are woken up.
3728 void complete(struct completion *x)
3730 unsigned long flags;
3732 spin_lock_irqsave(&x->wait.lock, flags);
3734 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3735 spin_unlock_irqrestore(&x->wait.lock, flags);
3737 EXPORT_SYMBOL(complete);
3740 * complete_all: - signals all threads waiting on this completion
3741 * @x: holds the state of this particular completion
3743 * This will wake up all threads waiting on this particular completion event.
3745 * It may be assumed that this function implies a write memory barrier before
3746 * changing the task state if and only if any tasks are woken up.
3748 void complete_all(struct completion *x)
3750 unsigned long flags;
3752 spin_lock_irqsave(&x->wait.lock, flags);
3753 x->done += UINT_MAX/2;
3754 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3755 spin_unlock_irqrestore(&x->wait.lock, flags);
3757 EXPORT_SYMBOL(complete_all);
3759 static inline long __sched
3760 do_wait_for_common(struct completion *x, long timeout, int state)
3763 DECLARE_WAITQUEUE(wait, current);
3765 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3767 if (signal_pending_state(state, current)) {
3768 timeout = -ERESTARTSYS;
3771 __set_current_state(state);
3772 spin_unlock_irq(&x->wait.lock);
3773 timeout = schedule_timeout(timeout);
3774 spin_lock_irq(&x->wait.lock);
3775 } while (!x->done && timeout);
3776 __remove_wait_queue(&x->wait, &wait);
3781 return timeout ?: 1;
3785 wait_for_common(struct completion *x, long timeout, int state)
3789 spin_lock_irq(&x->wait.lock);
3790 timeout = do_wait_for_common(x, timeout, state);
3791 spin_unlock_irq(&x->wait.lock);
3796 * wait_for_completion: - waits for completion of a task
3797 * @x: holds the state of this particular completion
3799 * This waits to be signaled for completion of a specific task. It is NOT
3800 * interruptible and there is no timeout.
3802 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3803 * and interrupt capability. Also see complete().
3805 void __sched wait_for_completion(struct completion *x)
3807 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3809 EXPORT_SYMBOL(wait_for_completion);
3812 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3813 * @x: holds the state of this particular completion
3814 * @timeout: timeout value in jiffies
3816 * This waits for either a completion of a specific task to be signaled or for a
3817 * specified timeout to expire. The timeout is in jiffies. It is not
3820 * The return value is 0 if timed out, and positive (at least 1, or number of
3821 * jiffies left till timeout) if completed.
3823 unsigned long __sched
3824 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3826 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3828 EXPORT_SYMBOL(wait_for_completion_timeout);
3831 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3832 * @x: holds the state of this particular completion
3834 * This waits for completion of a specific task to be signaled. It is
3837 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3839 int __sched wait_for_completion_interruptible(struct completion *x)
3841 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3842 if (t == -ERESTARTSYS)
3846 EXPORT_SYMBOL(wait_for_completion_interruptible);
3849 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3850 * @x: holds the state of this particular completion
3851 * @timeout: timeout value in jiffies
3853 * This waits for either a completion of a specific task to be signaled or for a
3854 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3856 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3857 * positive (at least 1, or number of jiffies left till timeout) if completed.
3860 wait_for_completion_interruptible_timeout(struct completion *x,
3861 unsigned long timeout)
3863 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3865 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3868 * wait_for_completion_killable: - waits for completion of a task (killable)
3869 * @x: holds the state of this particular completion
3871 * This waits to be signaled for completion of a specific task. It can be
3872 * interrupted by a kill signal.
3874 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3876 int __sched wait_for_completion_killable(struct completion *x)
3878 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3879 if (t == -ERESTARTSYS)
3883 EXPORT_SYMBOL(wait_for_completion_killable);
3886 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3887 * @x: holds the state of this particular completion
3888 * @timeout: timeout value in jiffies
3890 * This waits for either a completion of a specific task to be
3891 * signaled or for a specified timeout to expire. It can be
3892 * interrupted by a kill signal. The timeout is in jiffies.
3894 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3895 * positive (at least 1, or number of jiffies left till timeout) if completed.
3898 wait_for_completion_killable_timeout(struct completion *x,
3899 unsigned long timeout)
3901 return wait_for_common(x, timeout, TASK_KILLABLE);
3903 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3906 * try_wait_for_completion - try to decrement a completion without blocking
3907 * @x: completion structure
3909 * Returns: 0 if a decrement cannot be done without blocking
3910 * 1 if a decrement succeeded.
3912 * If a completion is being used as a counting completion,
3913 * attempt to decrement the counter without blocking. This
3914 * enables us to avoid waiting if the resource the completion
3915 * is protecting is not available.
3917 bool try_wait_for_completion(struct completion *x)
3919 unsigned long flags;
3922 spin_lock_irqsave(&x->wait.lock, flags);
3927 spin_unlock_irqrestore(&x->wait.lock, flags);
3930 EXPORT_SYMBOL(try_wait_for_completion);
3933 * completion_done - Test to see if a completion has any waiters
3934 * @x: completion structure
3936 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3937 * 1 if there are no waiters.
3940 bool completion_done(struct completion *x)
3942 unsigned long flags;
3945 spin_lock_irqsave(&x->wait.lock, flags);
3948 spin_unlock_irqrestore(&x->wait.lock, flags);
3951 EXPORT_SYMBOL(completion_done);
3954 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3956 unsigned long flags;
3959 init_waitqueue_entry(&wait, current);
3961 __set_current_state(state);
3963 spin_lock_irqsave(&q->lock, flags);
3964 __add_wait_queue(q, &wait);
3965 spin_unlock(&q->lock);
3966 timeout = schedule_timeout(timeout);
3967 spin_lock_irq(&q->lock);
3968 __remove_wait_queue(q, &wait);
3969 spin_unlock_irqrestore(&q->lock, flags);
3974 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3976 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3978 EXPORT_SYMBOL(interruptible_sleep_on);
3981 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3983 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3985 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3987 void __sched sleep_on(wait_queue_head_t *q)
3989 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3991 EXPORT_SYMBOL(sleep_on);
3993 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3995 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3997 EXPORT_SYMBOL(sleep_on_timeout);
3999 #ifdef CONFIG_RT_MUTEXES
4002 * rt_mutex_setprio - set the current priority of a task
4004 * @prio: prio value (kernel-internal form)
4006 * This function changes the 'effective' priority of a task. It does
4007 * not touch ->normal_prio like __setscheduler().
4009 * Used by the rt_mutex code to implement priority inheritance logic.
4011 void rt_mutex_setprio(struct task_struct *p, int prio)
4013 int oldprio, on_rq, running;
4015 const struct sched_class *prev_class;
4017 BUG_ON(prio < 0 || prio > MAX_PRIO);
4019 rq = __task_rq_lock(p);
4022 * Idle task boosting is a nono in general. There is one
4023 * exception, when PREEMPT_RT and NOHZ is active:
4025 * The idle task calls get_next_timer_interrupt() and holds
4026 * the timer wheel base->lock on the CPU and another CPU wants
4027 * to access the timer (probably to cancel it). We can safely
4028 * ignore the boosting request, as the idle CPU runs this code
4029 * with interrupts disabled and will complete the lock
4030 * protected section without being interrupted. So there is no
4031 * real need to boost.
4033 if (unlikely(p == rq->idle)) {
4034 WARN_ON(p != rq->curr);
4035 WARN_ON(p->pi_blocked_on);
4039 trace_sched_pi_setprio(p, prio);
4041 prev_class = p->sched_class;
4043 running = task_current(rq, p);
4045 dequeue_task(rq, p, 0);
4047 p->sched_class->put_prev_task(rq, p);
4050 p->sched_class = &rt_sched_class;
4052 p->sched_class = &fair_sched_class;
4057 p->sched_class->set_curr_task(rq);
4059 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4061 check_class_changed(rq, p, prev_class, oldprio);
4063 __task_rq_unlock(rq);
4066 void set_user_nice(struct task_struct *p, long nice)
4068 int old_prio, delta, on_rq;
4069 unsigned long flags;
4072 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4075 * We have to be careful, if called from sys_setpriority(),
4076 * the task might be in the middle of scheduling on another CPU.
4078 rq = task_rq_lock(p, &flags);
4080 * The RT priorities are set via sched_setscheduler(), but we still
4081 * allow the 'normal' nice value to be set - but as expected
4082 * it wont have any effect on scheduling until the task is
4083 * SCHED_FIFO/SCHED_RR:
4085 if (task_has_rt_policy(p)) {
4086 p->static_prio = NICE_TO_PRIO(nice);
4091 dequeue_task(rq, p, 0);
4093 p->static_prio = NICE_TO_PRIO(nice);
4096 p->prio = effective_prio(p);
4097 delta = p->prio - old_prio;
4100 enqueue_task(rq, p, 0);
4102 * If the task increased its priority or is running and
4103 * lowered its priority, then reschedule its CPU:
4105 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4106 resched_task(rq->curr);
4109 task_rq_unlock(rq, p, &flags);
4111 EXPORT_SYMBOL(set_user_nice);
4114 * can_nice - check if a task can reduce its nice value
4118 int can_nice(const struct task_struct *p, const int nice)
4120 /* convert nice value [19,-20] to rlimit style value [1,40] */
4121 int nice_rlim = 20 - nice;
4123 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4124 capable(CAP_SYS_NICE));
4127 #ifdef __ARCH_WANT_SYS_NICE
4130 * sys_nice - change the priority of the current process.
4131 * @increment: priority increment
4133 * sys_setpriority is a more generic, but much slower function that
4134 * does similar things.
4136 SYSCALL_DEFINE1(nice, int, increment)
4141 * Setpriority might change our priority at the same moment.
4142 * We don't have to worry. Conceptually one call occurs first
4143 * and we have a single winner.
4145 if (increment < -40)
4150 nice = TASK_NICE(current) + increment;
4156 if (increment < 0 && !can_nice(current, nice))
4159 retval = security_task_setnice(current, nice);
4163 set_user_nice(current, nice);
4170 * task_prio - return the priority value of a given task.
4171 * @p: the task in question.
4173 * This is the priority value as seen by users in /proc.
4174 * RT tasks are offset by -200. Normal tasks are centered
4175 * around 0, value goes from -16 to +15.
4177 int task_prio(const struct task_struct *p)
4179 return p->prio - MAX_RT_PRIO;
4183 * task_nice - return the nice value of a given task.
4184 * @p: the task in question.
4186 int task_nice(const struct task_struct *p)
4188 return TASK_NICE(p);
4190 EXPORT_SYMBOL(task_nice);
4193 * idle_cpu - is a given cpu idle currently?
4194 * @cpu: the processor in question.
4196 int idle_cpu(int cpu)
4198 struct rq *rq = cpu_rq(cpu);
4200 if (rq->curr != rq->idle)
4207 if (!llist_empty(&rq->wake_list))
4215 * idle_task - return the idle task for a given cpu.
4216 * @cpu: the processor in question.
4218 struct task_struct *idle_task(int cpu)
4220 return cpu_rq(cpu)->idle;
4224 * find_process_by_pid - find a process with a matching PID value.
4225 * @pid: the pid in question.
4227 static struct task_struct *find_process_by_pid(pid_t pid)
4229 return pid ? find_task_by_vpid(pid) : current;
4232 /* Actually do priority change: must hold rq lock. */
4234 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4237 p->rt_priority = prio;
4238 p->normal_prio = normal_prio(p);
4239 /* we are holding p->pi_lock already */
4240 p->prio = rt_mutex_getprio(p);
4241 if (rt_prio(p->prio))
4242 p->sched_class = &rt_sched_class;
4244 p->sched_class = &fair_sched_class;
4249 * check the target process has a UID that matches the current process's
4251 static bool check_same_owner(struct task_struct *p)
4253 const struct cred *cred = current_cred(), *pcred;
4257 pcred = __task_cred(p);
4258 match = (uid_eq(cred->euid, pcred->euid) ||
4259 uid_eq(cred->euid, pcred->uid));
4264 static int __sched_setscheduler(struct task_struct *p, int policy,
4265 const struct sched_param *param, bool user)
4267 int retval, oldprio, oldpolicy = -1, on_rq, running;
4268 unsigned long flags;
4269 const struct sched_class *prev_class;
4273 /* may grab non-irq protected spin_locks */
4274 BUG_ON(in_interrupt());
4276 /* double check policy once rq lock held */
4278 reset_on_fork = p->sched_reset_on_fork;
4279 policy = oldpolicy = p->policy;
4281 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4282 policy &= ~SCHED_RESET_ON_FORK;
4284 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4285 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4286 policy != SCHED_IDLE)
4291 * Valid priorities for SCHED_FIFO and SCHED_RR are
4292 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4293 * SCHED_BATCH and SCHED_IDLE is 0.
4295 if (param->sched_priority < 0 ||
4296 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4297 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4299 if (rt_policy(policy) != (param->sched_priority != 0))
4303 * Allow unprivileged RT tasks to decrease priority:
4305 if (user && !capable(CAP_SYS_NICE)) {
4306 if (rt_policy(policy)) {
4307 unsigned long rlim_rtprio =
4308 task_rlimit(p, RLIMIT_RTPRIO);
4310 /* can't set/change the rt policy */
4311 if (policy != p->policy && !rlim_rtprio)
4314 /* can't increase priority */
4315 if (param->sched_priority > p->rt_priority &&
4316 param->sched_priority > rlim_rtprio)
4321 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4322 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4324 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4325 if (!can_nice(p, TASK_NICE(p)))
4329 /* can't change other user's priorities */
4330 if (!check_same_owner(p))
4333 /* Normal users shall not reset the sched_reset_on_fork flag */
4334 if (p->sched_reset_on_fork && !reset_on_fork)
4339 retval = security_task_setscheduler(p);
4345 * make sure no PI-waiters arrive (or leave) while we are
4346 * changing the priority of the task:
4348 * To be able to change p->policy safely, the appropriate
4349 * runqueue lock must be held.
4351 rq = task_rq_lock(p, &flags);
4354 * Changing the policy of the stop threads its a very bad idea
4356 if (p == rq->stop) {
4357 task_rq_unlock(rq, p, &flags);
4362 * If not changing anything there's no need to proceed further:
4364 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4365 param->sched_priority == p->rt_priority))) {
4366 task_rq_unlock(rq, p, &flags);
4370 #ifdef CONFIG_RT_GROUP_SCHED
4373 * Do not allow realtime tasks into groups that have no runtime
4376 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4377 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4378 !task_group_is_autogroup(task_group(p))) {
4379 task_rq_unlock(rq, p, &flags);
4385 /* recheck policy now with rq lock held */
4386 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4387 policy = oldpolicy = -1;
4388 task_rq_unlock(rq, p, &flags);
4392 running = task_current(rq, p);
4394 dequeue_task(rq, p, 0);
4396 p->sched_class->put_prev_task(rq, p);
4398 p->sched_reset_on_fork = reset_on_fork;
4401 prev_class = p->sched_class;
4402 __setscheduler(rq, p, policy, param->sched_priority);
4405 p->sched_class->set_curr_task(rq);
4407 enqueue_task(rq, p, 0);
4409 check_class_changed(rq, p, prev_class, oldprio);
4410 task_rq_unlock(rq, p, &flags);
4412 rt_mutex_adjust_pi(p);
4418 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4419 * @p: the task in question.
4420 * @policy: new policy.
4421 * @param: structure containing the new RT priority.
4423 * NOTE that the task may be already dead.
4425 int sched_setscheduler(struct task_struct *p, int policy,
4426 const struct sched_param *param)
4428 return __sched_setscheduler(p, policy, param, true);
4430 EXPORT_SYMBOL_GPL(sched_setscheduler);
4433 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4434 * @p: the task in question.
4435 * @policy: new policy.
4436 * @param: structure containing the new RT priority.
4438 * Just like sched_setscheduler, only don't bother checking if the
4439 * current context has permission. For example, this is needed in
4440 * stop_machine(): we create temporary high priority worker threads,
4441 * but our caller might not have that capability.
4443 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4444 const struct sched_param *param)
4446 return __sched_setscheduler(p, policy, param, false);
4450 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4452 struct sched_param lparam;
4453 struct task_struct *p;
4456 if (!param || pid < 0)
4458 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4463 p = find_process_by_pid(pid);
4465 retval = sched_setscheduler(p, policy, &lparam);
4472 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4473 * @pid: the pid in question.
4474 * @policy: new policy.
4475 * @param: structure containing the new RT priority.
4477 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4478 struct sched_param __user *, param)
4480 /* negative values for policy are not valid */
4484 return do_sched_setscheduler(pid, policy, param);
4488 * sys_sched_setparam - set/change the RT priority of a thread
4489 * @pid: the pid in question.
4490 * @param: structure containing the new RT priority.
4492 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4494 return do_sched_setscheduler(pid, -1, param);
4498 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4499 * @pid: the pid in question.
4501 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4503 struct task_struct *p;
4511 p = find_process_by_pid(pid);
4513 retval = security_task_getscheduler(p);
4516 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4523 * sys_sched_getparam - get the RT priority of a thread
4524 * @pid: the pid in question.
4525 * @param: structure containing the RT priority.
4527 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4529 struct sched_param lp;
4530 struct task_struct *p;
4533 if (!param || pid < 0)
4537 p = find_process_by_pid(pid);
4542 retval = security_task_getscheduler(p);
4546 lp.sched_priority = p->rt_priority;
4550 * This one might sleep, we cannot do it with a spinlock held ...
4552 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4561 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4563 cpumask_var_t cpus_allowed, new_mask;
4564 struct task_struct *p;
4570 p = find_process_by_pid(pid);
4577 /* Prevent p going away */
4581 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4585 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4587 goto out_free_cpus_allowed;
4590 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4593 retval = security_task_setscheduler(p);
4597 cpuset_cpus_allowed(p, cpus_allowed);
4598 cpumask_and(new_mask, in_mask, cpus_allowed);
4600 retval = set_cpus_allowed_ptr(p, new_mask);
4603 cpuset_cpus_allowed(p, cpus_allowed);
4604 if (!cpumask_subset(new_mask, cpus_allowed)) {
4606 * We must have raced with a concurrent cpuset
4607 * update. Just reset the cpus_allowed to the
4608 * cpuset's cpus_allowed
4610 cpumask_copy(new_mask, cpus_allowed);
4615 free_cpumask_var(new_mask);
4616 out_free_cpus_allowed:
4617 free_cpumask_var(cpus_allowed);
4624 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4625 struct cpumask *new_mask)
4627 if (len < cpumask_size())
4628 cpumask_clear(new_mask);
4629 else if (len > cpumask_size())
4630 len = cpumask_size();
4632 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4636 * sys_sched_setaffinity - set the cpu affinity of a process
4637 * @pid: pid of the process
4638 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4639 * @user_mask_ptr: user-space pointer to the new cpu mask
4641 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4642 unsigned long __user *, user_mask_ptr)
4644 cpumask_var_t new_mask;
4647 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4650 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4652 retval = sched_setaffinity(pid, new_mask);
4653 free_cpumask_var(new_mask);
4657 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4659 struct task_struct *p;
4660 unsigned long flags;
4667 p = find_process_by_pid(pid);
4671 retval = security_task_getscheduler(p);
4675 raw_spin_lock_irqsave(&p->pi_lock, flags);
4676 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4677 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4687 * sys_sched_getaffinity - get the cpu affinity of a process
4688 * @pid: pid of the process
4689 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4690 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4692 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4693 unsigned long __user *, user_mask_ptr)
4698 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4700 if (len & (sizeof(unsigned long)-1))
4703 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4706 ret = sched_getaffinity(pid, mask);
4708 size_t retlen = min_t(size_t, len, cpumask_size());
4710 if (copy_to_user(user_mask_ptr, mask, retlen))
4715 free_cpumask_var(mask);
4721 * sys_sched_yield - yield the current processor to other threads.
4723 * This function yields the current CPU to other tasks. If there are no
4724 * other threads running on this CPU then this function will return.
4726 SYSCALL_DEFINE0(sched_yield)
4728 struct rq *rq = this_rq_lock();
4730 schedstat_inc(rq, yld_count);
4731 current->sched_class->yield_task(rq);
4734 * Since we are going to call schedule() anyway, there's
4735 * no need to preempt or enable interrupts:
4737 __release(rq->lock);
4738 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4739 do_raw_spin_unlock(&rq->lock);
4740 sched_preempt_enable_no_resched();
4747 static inline int should_resched(void)
4749 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4752 static void __cond_resched(void)
4754 add_preempt_count(PREEMPT_ACTIVE);
4756 sub_preempt_count(PREEMPT_ACTIVE);
4759 int __sched _cond_resched(void)
4761 if (should_resched()) {
4767 EXPORT_SYMBOL(_cond_resched);
4770 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4771 * call schedule, and on return reacquire the lock.
4773 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4774 * operations here to prevent schedule() from being called twice (once via
4775 * spin_unlock(), once by hand).
4777 int __cond_resched_lock(spinlock_t *lock)
4779 int resched = should_resched();
4782 lockdep_assert_held(lock);
4784 if (spin_needbreak(lock) || resched) {
4795 EXPORT_SYMBOL(__cond_resched_lock);
4797 int __sched __cond_resched_softirq(void)
4799 BUG_ON(!in_softirq());
4801 if (should_resched()) {
4809 EXPORT_SYMBOL(__cond_resched_softirq);
4812 * yield - yield the current processor to other threads.
4814 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4816 * The scheduler is at all times free to pick the calling task as the most
4817 * eligible task to run, if removing the yield() call from your code breaks
4818 * it, its already broken.
4820 * Typical broken usage is:
4825 * where one assumes that yield() will let 'the other' process run that will
4826 * make event true. If the current task is a SCHED_FIFO task that will never
4827 * happen. Never use yield() as a progress guarantee!!
4829 * If you want to use yield() to wait for something, use wait_event().
4830 * If you want to use yield() to be 'nice' for others, use cond_resched().
4831 * If you still want to use yield(), do not!
4833 void __sched yield(void)
4835 set_current_state(TASK_RUNNING);
4838 EXPORT_SYMBOL(yield);
4841 * yield_to - yield the current processor to another thread in
4842 * your thread group, or accelerate that thread toward the
4843 * processor it's on.
4845 * @preempt: whether task preemption is allowed or not
4847 * It's the caller's job to ensure that the target task struct
4848 * can't go away on us before we can do any checks.
4850 * Returns true if we indeed boosted the target task.
4852 bool __sched yield_to(struct task_struct *p, bool preempt)
4854 struct task_struct *curr = current;
4855 struct rq *rq, *p_rq;
4856 unsigned long flags;
4859 local_irq_save(flags);
4864 double_rq_lock(rq, p_rq);
4865 while (task_rq(p) != p_rq) {
4866 double_rq_unlock(rq, p_rq);
4870 if (!curr->sched_class->yield_to_task)
4873 if (curr->sched_class != p->sched_class)
4876 if (task_running(p_rq, p) || p->state)
4879 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4881 schedstat_inc(rq, yld_count);
4883 * Make p's CPU reschedule; pick_next_entity takes care of
4886 if (preempt && rq != p_rq)
4887 resched_task(p_rq->curr);
4890 * We might have set it in task_yield_fair(), but are
4891 * not going to schedule(), so don't want to skip
4894 rq->skip_clock_update = 0;
4898 double_rq_unlock(rq, p_rq);
4899 local_irq_restore(flags);
4906 EXPORT_SYMBOL_GPL(yield_to);
4909 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4910 * that process accounting knows that this is a task in IO wait state.
4912 void __sched io_schedule(void)
4914 struct rq *rq = raw_rq();
4916 delayacct_blkio_start();
4917 atomic_inc(&rq->nr_iowait);
4918 blk_flush_plug(current);
4919 current->in_iowait = 1;
4921 current->in_iowait = 0;
4922 atomic_dec(&rq->nr_iowait);
4923 delayacct_blkio_end();
4925 EXPORT_SYMBOL(io_schedule);
4927 long __sched io_schedule_timeout(long timeout)
4929 struct rq *rq = raw_rq();
4932 delayacct_blkio_start();
4933 atomic_inc(&rq->nr_iowait);
4934 blk_flush_plug(current);
4935 current->in_iowait = 1;
4936 ret = schedule_timeout(timeout);
4937 current->in_iowait = 0;
4938 atomic_dec(&rq->nr_iowait);
4939 delayacct_blkio_end();
4944 * sys_sched_get_priority_max - return maximum RT priority.
4945 * @policy: scheduling class.
4947 * this syscall returns the maximum rt_priority that can be used
4948 * by a given scheduling class.
4950 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4957 ret = MAX_USER_RT_PRIO-1;
4969 * sys_sched_get_priority_min - return minimum RT priority.
4970 * @policy: scheduling class.
4972 * this syscall returns the minimum rt_priority that can be used
4973 * by a given scheduling class.
4975 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4993 * sys_sched_rr_get_interval - return the default timeslice of a process.
4994 * @pid: pid of the process.
4995 * @interval: userspace pointer to the timeslice value.
4997 * this syscall writes the default timeslice value of a given process
4998 * into the user-space timespec buffer. A value of '0' means infinity.
5000 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5001 struct timespec __user *, interval)
5003 struct task_struct *p;
5004 unsigned int time_slice;
5005 unsigned long flags;
5015 p = find_process_by_pid(pid);
5019 retval = security_task_getscheduler(p);
5023 rq = task_rq_lock(p, &flags);
5024 time_slice = p->sched_class->get_rr_interval(rq, p);
5025 task_rq_unlock(rq, p, &flags);
5028 jiffies_to_timespec(time_slice, &t);
5029 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5037 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5039 void sched_show_task(struct task_struct *p)
5041 unsigned long free = 0;
5044 state = p->state ? __ffs(p->state) + 1 : 0;
5045 printk(KERN_INFO "%-15.15s %c", p->comm,
5046 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5047 #if BITS_PER_LONG == 32
5048 if (state == TASK_RUNNING)
5049 printk(KERN_CONT " running ");
5051 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5053 if (state == TASK_RUNNING)
5054 printk(KERN_CONT " running task ");
5056 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5058 #ifdef CONFIG_DEBUG_STACK_USAGE
5059 free = stack_not_used(p);
5061 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5062 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5063 (unsigned long)task_thread_info(p)->flags);
5065 show_stack(p, NULL);
5068 void show_state_filter(unsigned long state_filter)
5070 struct task_struct *g, *p;
5072 #if BITS_PER_LONG == 32
5074 " task PC stack pid father\n");
5077 " task PC stack pid father\n");
5080 do_each_thread(g, p) {
5082 * reset the NMI-timeout, listing all files on a slow
5083 * console might take a lot of time:
5085 touch_nmi_watchdog();
5086 if (!state_filter || (p->state & state_filter))
5088 } while_each_thread(g, p);
5090 touch_all_softlockup_watchdogs();
5092 #ifdef CONFIG_SCHED_DEBUG
5093 sysrq_sched_debug_show();
5097 * Only show locks if all tasks are dumped:
5100 debug_show_all_locks();
5103 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5105 idle->sched_class = &idle_sched_class;
5109 * init_idle - set up an idle thread for a given CPU
5110 * @idle: task in question
5111 * @cpu: cpu the idle task belongs to
5113 * NOTE: this function does not set the idle thread's NEED_RESCHED
5114 * flag, to make booting more robust.
5116 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5118 struct rq *rq = cpu_rq(cpu);
5119 unsigned long flags;
5121 raw_spin_lock_irqsave(&rq->lock, flags);
5124 idle->state = TASK_RUNNING;
5125 idle->se.exec_start = sched_clock();
5127 do_set_cpus_allowed(idle, cpumask_of(cpu));
5129 * We're having a chicken and egg problem, even though we are
5130 * holding rq->lock, the cpu isn't yet set to this cpu so the
5131 * lockdep check in task_group() will fail.
5133 * Similar case to sched_fork(). / Alternatively we could
5134 * use task_rq_lock() here and obtain the other rq->lock.
5139 __set_task_cpu(idle, cpu);
5142 rq->curr = rq->idle = idle;
5143 #if defined(CONFIG_SMP)
5146 raw_spin_unlock_irqrestore(&rq->lock, flags);
5148 /* Set the preempt count _outside_ the spinlocks! */
5149 task_thread_info(idle)->preempt_count = 0;
5152 * The idle tasks have their own, simple scheduling class:
5154 idle->sched_class = &idle_sched_class;
5155 ftrace_graph_init_idle_task(idle, cpu);
5156 #if defined(CONFIG_SMP)
5157 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5162 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5164 if (p->sched_class && p->sched_class->set_cpus_allowed)
5165 p->sched_class->set_cpus_allowed(p, new_mask);
5167 cpumask_copy(&p->cpus_allowed, new_mask);
5168 p->nr_cpus_allowed = cpumask_weight(new_mask);
5172 * This is how migration works:
5174 * 1) we invoke migration_cpu_stop() on the target CPU using
5176 * 2) stopper starts to run (implicitly forcing the migrated thread
5178 * 3) it checks whether the migrated task is still in the wrong runqueue.
5179 * 4) if it's in the wrong runqueue then the migration thread removes
5180 * it and puts it into the right queue.
5181 * 5) stopper completes and stop_one_cpu() returns and the migration
5186 * Change a given task's CPU affinity. Migrate the thread to a
5187 * proper CPU and schedule it away if the CPU it's executing on
5188 * is removed from the allowed bitmask.
5190 * NOTE: the caller must have a valid reference to the task, the
5191 * task must not exit() & deallocate itself prematurely. The
5192 * call is not atomic; no spinlocks may be held.
5194 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5196 unsigned long flags;
5198 unsigned int dest_cpu;
5201 rq = task_rq_lock(p, &flags);
5203 if (cpumask_equal(&p->cpus_allowed, new_mask))
5206 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5211 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5216 do_set_cpus_allowed(p, new_mask);
5218 /* Can the task run on the task's current CPU? If so, we're done */
5219 if (cpumask_test_cpu(task_cpu(p), new_mask))
5222 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5224 struct migration_arg arg = { p, dest_cpu };
5225 /* Need help from migration thread: drop lock and wait. */
5226 task_rq_unlock(rq, p, &flags);
5227 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5228 tlb_migrate_finish(p->mm);
5232 task_rq_unlock(rq, p, &flags);
5236 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5239 * Move (not current) task off this cpu, onto dest cpu. We're doing
5240 * this because either it can't run here any more (set_cpus_allowed()
5241 * away from this CPU, or CPU going down), or because we're
5242 * attempting to rebalance this task on exec (sched_exec).
5244 * So we race with normal scheduler movements, but that's OK, as long
5245 * as the task is no longer on this CPU.
5247 * Returns non-zero if task was successfully migrated.
5249 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5251 struct rq *rq_dest, *rq_src;
5254 if (unlikely(!cpu_active(dest_cpu)))
5257 rq_src = cpu_rq(src_cpu);
5258 rq_dest = cpu_rq(dest_cpu);
5260 raw_spin_lock(&p->pi_lock);
5261 double_rq_lock(rq_src, rq_dest);
5262 /* Already moved. */
5263 if (task_cpu(p) != src_cpu)
5265 /* Affinity changed (again). */
5266 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5270 * If we're not on a rq, the next wake-up will ensure we're
5274 dequeue_task(rq_src, p, 0);
5275 set_task_cpu(p, dest_cpu);
5276 enqueue_task(rq_dest, p, 0);
5277 check_preempt_curr(rq_dest, p, 0);
5282 double_rq_unlock(rq_src, rq_dest);
5283 raw_spin_unlock(&p->pi_lock);
5288 * migration_cpu_stop - this will be executed by a highprio stopper thread
5289 * and performs thread migration by bumping thread off CPU then
5290 * 'pushing' onto another runqueue.
5292 static int migration_cpu_stop(void *data)
5294 struct migration_arg *arg = data;
5297 * The original target cpu might have gone down and we might
5298 * be on another cpu but it doesn't matter.
5300 local_irq_disable();
5301 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5306 #ifdef CONFIG_HOTPLUG_CPU
5309 * Ensures that the idle task is using init_mm right before its cpu goes
5312 void idle_task_exit(void)
5314 struct mm_struct *mm = current->active_mm;
5316 BUG_ON(cpu_online(smp_processor_id()));
5319 switch_mm(mm, &init_mm, current);
5324 * Since this CPU is going 'away' for a while, fold any nr_active delta
5325 * we might have. Assumes we're called after migrate_tasks() so that the
5326 * nr_active count is stable.
5328 * Also see the comment "Global load-average calculations".
5330 static void calc_load_migrate(struct rq *rq)
5332 long delta = calc_load_fold_active(rq);
5334 atomic_long_add(delta, &calc_load_tasks);
5338 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5339 * try_to_wake_up()->select_task_rq().
5341 * Called with rq->lock held even though we'er in stop_machine() and
5342 * there's no concurrency possible, we hold the required locks anyway
5343 * because of lock validation efforts.
5345 static void migrate_tasks(unsigned int dead_cpu)
5347 struct rq *rq = cpu_rq(dead_cpu);
5348 struct task_struct *next, *stop = rq->stop;
5352 * Fudge the rq selection such that the below task selection loop
5353 * doesn't get stuck on the currently eligible stop task.
5355 * We're currently inside stop_machine() and the rq is either stuck
5356 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5357 * either way we should never end up calling schedule() until we're
5364 * There's this thread running, bail when that's the only
5367 if (rq->nr_running == 1)
5370 next = pick_next_task(rq);
5372 next->sched_class->put_prev_task(rq, next);
5374 /* Find suitable destination for @next, with force if needed. */
5375 dest_cpu = select_fallback_rq(dead_cpu, next);
5376 raw_spin_unlock(&rq->lock);
5378 __migrate_task(next, dead_cpu, dest_cpu);
5380 raw_spin_lock(&rq->lock);
5386 #endif /* CONFIG_HOTPLUG_CPU */
5388 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5390 static struct ctl_table sd_ctl_dir[] = {
5392 .procname = "sched_domain",
5398 static struct ctl_table sd_ctl_root[] = {
5400 .procname = "kernel",
5402 .child = sd_ctl_dir,
5407 static struct ctl_table *sd_alloc_ctl_entry(int n)
5409 struct ctl_table *entry =
5410 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5415 static void sd_free_ctl_entry(struct ctl_table **tablep)
5417 struct ctl_table *entry;
5420 * In the intermediate directories, both the child directory and
5421 * procname are dynamically allocated and could fail but the mode
5422 * will always be set. In the lowest directory the names are
5423 * static strings and all have proc handlers.
5425 for (entry = *tablep; entry->mode; entry++) {
5427 sd_free_ctl_entry(&entry->child);
5428 if (entry->proc_handler == NULL)
5429 kfree(entry->procname);
5437 set_table_entry(struct ctl_table *entry,
5438 const char *procname, void *data, int maxlen,
5439 umode_t mode, proc_handler *proc_handler)
5441 entry->procname = procname;
5443 entry->maxlen = maxlen;
5445 entry->proc_handler = proc_handler;
5448 static struct ctl_table *
5449 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5451 struct ctl_table *table = sd_alloc_ctl_entry(13);
5456 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5457 sizeof(long), 0644, proc_doulongvec_minmax);
5458 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5459 sizeof(long), 0644, proc_doulongvec_minmax);
5460 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5465 sizeof(int), 0644, proc_dointvec_minmax);
5466 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5467 sizeof(int), 0644, proc_dointvec_minmax);
5468 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5469 sizeof(int), 0644, proc_dointvec_minmax);
5470 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5471 sizeof(int), 0644, proc_dointvec_minmax);
5472 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5473 sizeof(int), 0644, proc_dointvec_minmax);
5474 set_table_entry(&table[9], "cache_nice_tries",
5475 &sd->cache_nice_tries,
5476 sizeof(int), 0644, proc_dointvec_minmax);
5477 set_table_entry(&table[10], "flags", &sd->flags,
5478 sizeof(int), 0644, proc_dointvec_minmax);
5479 set_table_entry(&table[11], "name", sd->name,
5480 CORENAME_MAX_SIZE, 0444, proc_dostring);
5481 /* &table[12] is terminator */
5486 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5488 struct ctl_table *entry, *table;
5489 struct sched_domain *sd;
5490 int domain_num = 0, i;
5493 for_each_domain(cpu, sd)
5495 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5500 for_each_domain(cpu, sd) {
5501 snprintf(buf, 32, "domain%d", i);
5502 entry->procname = kstrdup(buf, GFP_KERNEL);
5504 entry->child = sd_alloc_ctl_domain_table(sd);
5511 static struct ctl_table_header *sd_sysctl_header;
5512 static void register_sched_domain_sysctl(void)
5514 int i, cpu_num = num_possible_cpus();
5515 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5518 WARN_ON(sd_ctl_dir[0].child);
5519 sd_ctl_dir[0].child = entry;
5524 for_each_possible_cpu(i) {
5525 snprintf(buf, 32, "cpu%d", i);
5526 entry->procname = kstrdup(buf, GFP_KERNEL);
5528 entry->child = sd_alloc_ctl_cpu_table(i);
5532 WARN_ON(sd_sysctl_header);
5533 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5536 /* may be called multiple times per register */
5537 static void unregister_sched_domain_sysctl(void)
5539 if (sd_sysctl_header)
5540 unregister_sysctl_table(sd_sysctl_header);
5541 sd_sysctl_header = NULL;
5542 if (sd_ctl_dir[0].child)
5543 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5546 static void register_sched_domain_sysctl(void)
5549 static void unregister_sched_domain_sysctl(void)
5554 static void set_rq_online(struct rq *rq)
5557 const struct sched_class *class;
5559 cpumask_set_cpu(rq->cpu, rq->rd->online);
5562 for_each_class(class) {
5563 if (class->rq_online)
5564 class->rq_online(rq);
5569 static void set_rq_offline(struct rq *rq)
5572 const struct sched_class *class;
5574 for_each_class(class) {
5575 if (class->rq_offline)
5576 class->rq_offline(rq);
5579 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5585 * migration_call - callback that gets triggered when a CPU is added.
5586 * Here we can start up the necessary migration thread for the new CPU.
5588 static int __cpuinit
5589 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5591 int cpu = (long)hcpu;
5592 unsigned long flags;
5593 struct rq *rq = cpu_rq(cpu);
5595 switch (action & ~CPU_TASKS_FROZEN) {
5597 case CPU_UP_PREPARE:
5598 rq->calc_load_update = calc_load_update;
5602 /* Update our root-domain */
5603 raw_spin_lock_irqsave(&rq->lock, flags);
5605 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5609 raw_spin_unlock_irqrestore(&rq->lock, flags);
5612 #ifdef CONFIG_HOTPLUG_CPU
5614 sched_ttwu_pending();
5615 /* Update our root-domain */
5616 raw_spin_lock_irqsave(&rq->lock, flags);
5618 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5622 BUG_ON(rq->nr_running != 1); /* the migration thread */
5623 raw_spin_unlock_irqrestore(&rq->lock, flags);
5627 calc_load_migrate(rq);
5632 update_max_interval();
5638 * Register at high priority so that task migration (migrate_all_tasks)
5639 * happens before everything else. This has to be lower priority than
5640 * the notifier in the perf_event subsystem, though.
5642 static struct notifier_block __cpuinitdata migration_notifier = {
5643 .notifier_call = migration_call,
5644 .priority = CPU_PRI_MIGRATION,
5647 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5648 unsigned long action, void *hcpu)
5650 switch (action & ~CPU_TASKS_FROZEN) {
5652 case CPU_DOWN_FAILED:
5653 set_cpu_active((long)hcpu, true);
5660 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5661 unsigned long action, void *hcpu)
5663 switch (action & ~CPU_TASKS_FROZEN) {
5664 case CPU_DOWN_PREPARE:
5665 set_cpu_active((long)hcpu, false);
5672 static int __init migration_init(void)
5674 void *cpu = (void *)(long)smp_processor_id();
5677 /* Initialize migration for the boot CPU */
5678 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5679 BUG_ON(err == NOTIFY_BAD);
5680 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5681 register_cpu_notifier(&migration_notifier);
5683 /* Register cpu active notifiers */
5684 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5685 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5689 early_initcall(migration_init);
5694 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5696 #ifdef CONFIG_SCHED_DEBUG
5698 static __read_mostly int sched_debug_enabled;
5700 static int __init sched_debug_setup(char *str)
5702 sched_debug_enabled = 1;
5706 early_param("sched_debug", sched_debug_setup);
5708 static inline bool sched_debug(void)
5710 return sched_debug_enabled;
5713 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5714 struct cpumask *groupmask)
5716 struct sched_group *group = sd->groups;
5719 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5720 cpumask_clear(groupmask);
5722 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5724 if (!(sd->flags & SD_LOAD_BALANCE)) {
5725 printk("does not load-balance\n");
5727 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5732 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5734 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5735 printk(KERN_ERR "ERROR: domain->span does not contain "
5738 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5739 printk(KERN_ERR "ERROR: domain->groups does not contain"
5743 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5747 printk(KERN_ERR "ERROR: group is NULL\n");
5752 * Even though we initialize ->power to something semi-sane,
5753 * we leave power_orig unset. This allows us to detect if
5754 * domain iteration is still funny without causing /0 traps.
5756 if (!group->sgp->power_orig) {
5757 printk(KERN_CONT "\n");
5758 printk(KERN_ERR "ERROR: domain->cpu_power not "
5763 if (!cpumask_weight(sched_group_cpus(group))) {
5764 printk(KERN_CONT "\n");
5765 printk(KERN_ERR "ERROR: empty group\n");
5769 if (!(sd->flags & SD_OVERLAP) &&
5770 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5771 printk(KERN_CONT "\n");
5772 printk(KERN_ERR "ERROR: repeated CPUs\n");
5776 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5778 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5780 printk(KERN_CONT " %s", str);
5781 if (group->sgp->power != SCHED_POWER_SCALE) {
5782 printk(KERN_CONT " (cpu_power = %d)",
5786 group = group->next;
5787 } while (group != sd->groups);
5788 printk(KERN_CONT "\n");
5790 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5791 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5794 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5795 printk(KERN_ERR "ERROR: parent span is not a superset "
5796 "of domain->span\n");
5800 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5804 if (!sched_debug_enabled)
5808 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5812 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5815 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5823 #else /* !CONFIG_SCHED_DEBUG */
5824 # define sched_domain_debug(sd, cpu) do { } while (0)
5825 static inline bool sched_debug(void)
5829 #endif /* CONFIG_SCHED_DEBUG */
5831 static int sd_degenerate(struct sched_domain *sd)
5833 if (cpumask_weight(sched_domain_span(sd)) == 1)
5836 /* Following flags need at least 2 groups */
5837 if (sd->flags & (SD_LOAD_BALANCE |
5838 SD_BALANCE_NEWIDLE |
5842 SD_SHARE_PKG_RESOURCES)) {
5843 if (sd->groups != sd->groups->next)
5847 /* Following flags don't use groups */
5848 if (sd->flags & (SD_WAKE_AFFINE))
5855 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5857 unsigned long cflags = sd->flags, pflags = parent->flags;
5859 if (sd_degenerate(parent))
5862 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5865 /* Flags needing groups don't count if only 1 group in parent */
5866 if (parent->groups == parent->groups->next) {
5867 pflags &= ~(SD_LOAD_BALANCE |
5868 SD_BALANCE_NEWIDLE |
5872 SD_SHARE_PKG_RESOURCES);
5873 if (nr_node_ids == 1)
5874 pflags &= ~SD_SERIALIZE;
5876 if (~cflags & pflags)
5882 static void free_rootdomain(struct rcu_head *rcu)
5884 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5886 cpupri_cleanup(&rd->cpupri);
5887 free_cpumask_var(rd->rto_mask);
5888 free_cpumask_var(rd->online);
5889 free_cpumask_var(rd->span);
5893 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5895 struct root_domain *old_rd = NULL;
5896 unsigned long flags;
5898 raw_spin_lock_irqsave(&rq->lock, flags);
5903 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5906 cpumask_clear_cpu(rq->cpu, old_rd->span);
5909 * If we dont want to free the old_rt yet then
5910 * set old_rd to NULL to skip the freeing later
5913 if (!atomic_dec_and_test(&old_rd->refcount))
5917 atomic_inc(&rd->refcount);
5920 cpumask_set_cpu(rq->cpu, rd->span);
5921 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5924 raw_spin_unlock_irqrestore(&rq->lock, flags);
5927 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5930 static int init_rootdomain(struct root_domain *rd)
5932 memset(rd, 0, sizeof(*rd));
5934 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5936 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5938 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5941 if (cpupri_init(&rd->cpupri) != 0)
5946 free_cpumask_var(rd->rto_mask);
5948 free_cpumask_var(rd->online);
5950 free_cpumask_var(rd->span);
5956 * By default the system creates a single root-domain with all cpus as
5957 * members (mimicking the global state we have today).
5959 struct root_domain def_root_domain;
5961 static void init_defrootdomain(void)
5963 init_rootdomain(&def_root_domain);
5965 atomic_set(&def_root_domain.refcount, 1);
5968 static struct root_domain *alloc_rootdomain(void)
5970 struct root_domain *rd;
5972 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5976 if (init_rootdomain(rd) != 0) {
5984 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5986 struct sched_group *tmp, *first;
5995 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6000 } while (sg != first);
6003 static void free_sched_domain(struct rcu_head *rcu)
6005 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6008 * If its an overlapping domain it has private groups, iterate and
6011 if (sd->flags & SD_OVERLAP) {
6012 free_sched_groups(sd->groups, 1);
6013 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6014 kfree(sd->groups->sgp);
6020 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6022 call_rcu(&sd->rcu, free_sched_domain);
6025 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6027 for (; sd; sd = sd->parent)
6028 destroy_sched_domain(sd, cpu);
6032 * Keep a special pointer to the highest sched_domain that has
6033 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6034 * allows us to avoid some pointer chasing select_idle_sibling().
6036 * Also keep a unique ID per domain (we use the first cpu number in
6037 * the cpumask of the domain), this allows us to quickly tell if
6038 * two cpus are in the same cache domain, see cpus_share_cache().
6040 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6041 DEFINE_PER_CPU(int, sd_llc_id);
6043 static void update_top_cache_domain(int cpu)
6045 struct sched_domain *sd;
6048 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6050 id = cpumask_first(sched_domain_span(sd));
6052 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6053 per_cpu(sd_llc_id, cpu) = id;
6057 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6058 * hold the hotplug lock.
6061 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6063 struct rq *rq = cpu_rq(cpu);
6064 struct sched_domain *tmp;
6066 /* Remove the sched domains which do not contribute to scheduling. */
6067 for (tmp = sd; tmp; ) {
6068 struct sched_domain *parent = tmp->parent;
6072 if (sd_parent_degenerate(tmp, parent)) {
6073 tmp->parent = parent->parent;
6075 parent->parent->child = tmp;
6076 destroy_sched_domain(parent, cpu);
6081 if (sd && sd_degenerate(sd)) {
6084 destroy_sched_domain(tmp, cpu);
6089 sched_domain_debug(sd, cpu);
6091 rq_attach_root(rq, rd);
6093 rcu_assign_pointer(rq->sd, sd);
6094 destroy_sched_domains(tmp, cpu);
6096 update_top_cache_domain(cpu);
6099 /* cpus with isolated domains */
6100 static cpumask_var_t cpu_isolated_map;
6102 /* Setup the mask of cpus configured for isolated domains */
6103 static int __init isolated_cpu_setup(char *str)
6105 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6106 cpulist_parse(str, cpu_isolated_map);
6110 __setup("isolcpus=", isolated_cpu_setup);
6112 static const struct cpumask *cpu_cpu_mask(int cpu)
6114 return cpumask_of_node(cpu_to_node(cpu));
6118 struct sched_domain **__percpu sd;
6119 struct sched_group **__percpu sg;
6120 struct sched_group_power **__percpu sgp;
6124 struct sched_domain ** __percpu sd;
6125 struct root_domain *rd;
6135 struct sched_domain_topology_level;
6137 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6138 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6140 #define SDTL_OVERLAP 0x01
6142 struct sched_domain_topology_level {
6143 sched_domain_init_f init;
6144 sched_domain_mask_f mask;
6147 struct sd_data data;
6151 * Build an iteration mask that can exclude certain CPUs from the upwards
6154 * Asymmetric node setups can result in situations where the domain tree is of
6155 * unequal depth, make sure to skip domains that already cover the entire
6158 * In that case build_sched_domains() will have terminated the iteration early
6159 * and our sibling sd spans will be empty. Domains should always include the
6160 * cpu they're built on, so check that.
6163 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6165 const struct cpumask *span = sched_domain_span(sd);
6166 struct sd_data *sdd = sd->private;
6167 struct sched_domain *sibling;
6170 for_each_cpu(i, span) {
6171 sibling = *per_cpu_ptr(sdd->sd, i);
6172 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6175 cpumask_set_cpu(i, sched_group_mask(sg));
6180 * Return the canonical balance cpu for this group, this is the first cpu
6181 * of this group that's also in the iteration mask.
6183 int group_balance_cpu(struct sched_group *sg)
6185 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6189 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6191 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6192 const struct cpumask *span = sched_domain_span(sd);
6193 struct cpumask *covered = sched_domains_tmpmask;
6194 struct sd_data *sdd = sd->private;
6195 struct sched_domain *child;
6198 cpumask_clear(covered);
6200 for_each_cpu(i, span) {
6201 struct cpumask *sg_span;
6203 if (cpumask_test_cpu(i, covered))
6206 child = *per_cpu_ptr(sdd->sd, i);
6208 /* See the comment near build_group_mask(). */
6209 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6212 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6213 GFP_KERNEL, cpu_to_node(cpu));
6218 sg_span = sched_group_cpus(sg);
6220 child = child->child;
6221 cpumask_copy(sg_span, sched_domain_span(child));
6223 cpumask_set_cpu(i, sg_span);
6225 cpumask_or(covered, covered, sg_span);
6227 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6228 if (atomic_inc_return(&sg->sgp->ref) == 1)
6229 build_group_mask(sd, sg);
6232 * Initialize sgp->power such that even if we mess up the
6233 * domains and no possible iteration will get us here, we won't
6236 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6239 * Make sure the first group of this domain contains the
6240 * canonical balance cpu. Otherwise the sched_domain iteration
6241 * breaks. See update_sg_lb_stats().
6243 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6244 group_balance_cpu(sg) == cpu)
6254 sd->groups = groups;
6259 free_sched_groups(first, 0);
6264 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6266 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6267 struct sched_domain *child = sd->child;
6270 cpu = cpumask_first(sched_domain_span(child));
6273 *sg = *per_cpu_ptr(sdd->sg, cpu);
6274 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6275 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6282 * build_sched_groups will build a circular linked list of the groups
6283 * covered by the given span, and will set each group's ->cpumask correctly,
6284 * and ->cpu_power to 0.
6286 * Assumes the sched_domain tree is fully constructed
6289 build_sched_groups(struct sched_domain *sd, int cpu)
6291 struct sched_group *first = NULL, *last = NULL;
6292 struct sd_data *sdd = sd->private;
6293 const struct cpumask *span = sched_domain_span(sd);
6294 struct cpumask *covered;
6297 get_group(cpu, sdd, &sd->groups);
6298 atomic_inc(&sd->groups->ref);
6300 if (cpu != cpumask_first(sched_domain_span(sd)))
6303 lockdep_assert_held(&sched_domains_mutex);
6304 covered = sched_domains_tmpmask;
6306 cpumask_clear(covered);
6308 for_each_cpu(i, span) {
6309 struct sched_group *sg;
6310 int group = get_group(i, sdd, &sg);
6313 if (cpumask_test_cpu(i, covered))
6316 cpumask_clear(sched_group_cpus(sg));
6318 cpumask_setall(sched_group_mask(sg));
6320 for_each_cpu(j, span) {
6321 if (get_group(j, sdd, NULL) != group)
6324 cpumask_set_cpu(j, covered);
6325 cpumask_set_cpu(j, sched_group_cpus(sg));
6340 * Initialize sched groups cpu_power.
6342 * cpu_power indicates the capacity of sched group, which is used while
6343 * distributing the load between different sched groups in a sched domain.
6344 * Typically cpu_power for all the groups in a sched domain will be same unless
6345 * there are asymmetries in the topology. If there are asymmetries, group
6346 * having more cpu_power will pickup more load compared to the group having
6349 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6351 struct sched_group *sg = sd->groups;
6353 WARN_ON(!sd || !sg);
6356 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6358 } while (sg != sd->groups);
6360 if (cpu != group_balance_cpu(sg))
6363 update_group_power(sd, cpu);
6364 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6367 int __weak arch_sd_sibling_asym_packing(void)
6369 return 0*SD_ASYM_PACKING;
6373 * Initializers for schedule domains
6374 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6377 #ifdef CONFIG_SCHED_DEBUG
6378 # define SD_INIT_NAME(sd, type) sd->name = #type
6380 # define SD_INIT_NAME(sd, type) do { } while (0)
6383 #define SD_INIT_FUNC(type) \
6384 static noinline struct sched_domain * \
6385 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6387 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6388 *sd = SD_##type##_INIT; \
6389 SD_INIT_NAME(sd, type); \
6390 sd->private = &tl->data; \
6395 #ifdef CONFIG_SCHED_SMT
6396 SD_INIT_FUNC(SIBLING)
6398 #ifdef CONFIG_SCHED_MC
6401 #ifdef CONFIG_SCHED_BOOK
6405 static int default_relax_domain_level = -1;
6406 int sched_domain_level_max;
6408 static int __init setup_relax_domain_level(char *str)
6410 if (kstrtoint(str, 0, &default_relax_domain_level))
6411 pr_warn("Unable to set relax_domain_level\n");
6415 __setup("relax_domain_level=", setup_relax_domain_level);
6417 static void set_domain_attribute(struct sched_domain *sd,
6418 struct sched_domain_attr *attr)
6422 if (!attr || attr->relax_domain_level < 0) {
6423 if (default_relax_domain_level < 0)
6426 request = default_relax_domain_level;
6428 request = attr->relax_domain_level;
6429 if (request < sd->level) {
6430 /* turn off idle balance on this domain */
6431 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6433 /* turn on idle balance on this domain */
6434 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6438 static void __sdt_free(const struct cpumask *cpu_map);
6439 static int __sdt_alloc(const struct cpumask *cpu_map);
6441 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6442 const struct cpumask *cpu_map)
6446 if (!atomic_read(&d->rd->refcount))
6447 free_rootdomain(&d->rd->rcu); /* fall through */
6449 free_percpu(d->sd); /* fall through */
6451 __sdt_free(cpu_map); /* fall through */
6457 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6458 const struct cpumask *cpu_map)
6460 memset(d, 0, sizeof(*d));
6462 if (__sdt_alloc(cpu_map))
6463 return sa_sd_storage;
6464 d->sd = alloc_percpu(struct sched_domain *);
6466 return sa_sd_storage;
6467 d->rd = alloc_rootdomain();
6470 return sa_rootdomain;
6474 * NULL the sd_data elements we've used to build the sched_domain and
6475 * sched_group structure so that the subsequent __free_domain_allocs()
6476 * will not free the data we're using.
6478 static void claim_allocations(int cpu, struct sched_domain *sd)
6480 struct sd_data *sdd = sd->private;
6482 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6483 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6485 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6486 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6488 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6489 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6492 #ifdef CONFIG_SCHED_SMT
6493 static const struct cpumask *cpu_smt_mask(int cpu)
6495 return topology_thread_cpumask(cpu);
6500 * Topology list, bottom-up.
6502 static struct sched_domain_topology_level default_topology[] = {
6503 #ifdef CONFIG_SCHED_SMT
6504 { sd_init_SIBLING, cpu_smt_mask, },
6506 #ifdef CONFIG_SCHED_MC
6507 { sd_init_MC, cpu_coregroup_mask, },
6509 #ifdef CONFIG_SCHED_BOOK
6510 { sd_init_BOOK, cpu_book_mask, },
6512 { sd_init_CPU, cpu_cpu_mask, },
6516 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6520 static int sched_domains_numa_levels;
6521 static int *sched_domains_numa_distance;
6522 static struct cpumask ***sched_domains_numa_masks;
6523 static int sched_domains_curr_level;
6525 static inline int sd_local_flags(int level)
6527 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6530 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6533 static struct sched_domain *
6534 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6536 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6537 int level = tl->numa_level;
6538 int sd_weight = cpumask_weight(
6539 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6541 *sd = (struct sched_domain){
6542 .min_interval = sd_weight,
6543 .max_interval = 2*sd_weight,
6545 .imbalance_pct = 125,
6546 .cache_nice_tries = 2,
6553 .flags = 1*SD_LOAD_BALANCE
6554 | 1*SD_BALANCE_NEWIDLE
6560 | 0*SD_SHARE_CPUPOWER
6561 | 0*SD_SHARE_PKG_RESOURCES
6563 | 0*SD_PREFER_SIBLING
6564 | sd_local_flags(level)
6566 .last_balance = jiffies,
6567 .balance_interval = sd_weight,
6569 SD_INIT_NAME(sd, NUMA);
6570 sd->private = &tl->data;
6573 * Ugly hack to pass state to sd_numa_mask()...
6575 sched_domains_curr_level = tl->numa_level;
6580 static const struct cpumask *sd_numa_mask(int cpu)
6582 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6585 static void sched_numa_warn(const char *str)
6587 static int done = false;
6595 printk(KERN_WARNING "ERROR: %s\n\n", str);
6597 for (i = 0; i < nr_node_ids; i++) {
6598 printk(KERN_WARNING " ");
6599 for (j = 0; j < nr_node_ids; j++)
6600 printk(KERN_CONT "%02d ", node_distance(i,j));
6601 printk(KERN_CONT "\n");
6603 printk(KERN_WARNING "\n");
6606 static bool find_numa_distance(int distance)
6610 if (distance == node_distance(0, 0))
6613 for (i = 0; i < sched_domains_numa_levels; i++) {
6614 if (sched_domains_numa_distance[i] == distance)
6621 static void sched_init_numa(void)
6623 int next_distance, curr_distance = node_distance(0, 0);
6624 struct sched_domain_topology_level *tl;
6628 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6629 if (!sched_domains_numa_distance)
6633 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6634 * unique distances in the node_distance() table.
6636 * Assumes node_distance(0,j) includes all distances in
6637 * node_distance(i,j) in order to avoid cubic time.
6639 next_distance = curr_distance;
6640 for (i = 0; i < nr_node_ids; i++) {
6641 for (j = 0; j < nr_node_ids; j++) {
6642 for (k = 0; k < nr_node_ids; k++) {
6643 int distance = node_distance(i, k);
6645 if (distance > curr_distance &&
6646 (distance < next_distance ||
6647 next_distance == curr_distance))
6648 next_distance = distance;
6651 * While not a strong assumption it would be nice to know
6652 * about cases where if node A is connected to B, B is not
6653 * equally connected to A.
6655 if (sched_debug() && node_distance(k, i) != distance)
6656 sched_numa_warn("Node-distance not symmetric");
6658 if (sched_debug() && i && !find_numa_distance(distance))
6659 sched_numa_warn("Node-0 not representative");
6661 if (next_distance != curr_distance) {
6662 sched_domains_numa_distance[level++] = next_distance;
6663 sched_domains_numa_levels = level;
6664 curr_distance = next_distance;
6669 * In case of sched_debug() we verify the above assumption.
6675 * 'level' contains the number of unique distances, excluding the
6676 * identity distance node_distance(i,i).
6678 * The sched_domains_nume_distance[] array includes the actual distance
6682 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6683 if (!sched_domains_numa_masks)
6687 * Now for each level, construct a mask per node which contains all
6688 * cpus of nodes that are that many hops away from us.
6690 for (i = 0; i < level; i++) {
6691 sched_domains_numa_masks[i] =
6692 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6693 if (!sched_domains_numa_masks[i])
6696 for (j = 0; j < nr_node_ids; j++) {
6697 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6701 sched_domains_numa_masks[i][j] = mask;
6703 for (k = 0; k < nr_node_ids; k++) {
6704 if (node_distance(j, k) > sched_domains_numa_distance[i])
6707 cpumask_or(mask, mask, cpumask_of_node(k));
6712 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6713 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6718 * Copy the default topology bits..
6720 for (i = 0; default_topology[i].init; i++)
6721 tl[i] = default_topology[i];
6724 * .. and append 'j' levels of NUMA goodness.
6726 for (j = 0; j < level; i++, j++) {
6727 tl[i] = (struct sched_domain_topology_level){
6728 .init = sd_numa_init,
6729 .mask = sd_numa_mask,
6730 .flags = SDTL_OVERLAP,
6735 sched_domain_topology = tl;
6738 static inline void sched_init_numa(void)
6741 #endif /* CONFIG_NUMA */
6743 static int __sdt_alloc(const struct cpumask *cpu_map)
6745 struct sched_domain_topology_level *tl;
6748 for (tl = sched_domain_topology; tl->init; tl++) {
6749 struct sd_data *sdd = &tl->data;
6751 sdd->sd = alloc_percpu(struct sched_domain *);
6755 sdd->sg = alloc_percpu(struct sched_group *);
6759 sdd->sgp = alloc_percpu(struct sched_group_power *);
6763 for_each_cpu(j, cpu_map) {
6764 struct sched_domain *sd;
6765 struct sched_group *sg;
6766 struct sched_group_power *sgp;
6768 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6769 GFP_KERNEL, cpu_to_node(j));
6773 *per_cpu_ptr(sdd->sd, j) = sd;
6775 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6776 GFP_KERNEL, cpu_to_node(j));
6782 *per_cpu_ptr(sdd->sg, j) = sg;
6784 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6785 GFP_KERNEL, cpu_to_node(j));
6789 *per_cpu_ptr(sdd->sgp, j) = sgp;
6796 static void __sdt_free(const struct cpumask *cpu_map)
6798 struct sched_domain_topology_level *tl;
6801 for (tl = sched_domain_topology; tl->init; tl++) {
6802 struct sd_data *sdd = &tl->data;
6804 for_each_cpu(j, cpu_map) {
6805 struct sched_domain *sd;
6808 sd = *per_cpu_ptr(sdd->sd, j);
6809 if (sd && (sd->flags & SD_OVERLAP))
6810 free_sched_groups(sd->groups, 0);
6811 kfree(*per_cpu_ptr(sdd->sd, j));
6815 kfree(*per_cpu_ptr(sdd->sg, j));
6817 kfree(*per_cpu_ptr(sdd->sgp, j));
6819 free_percpu(sdd->sd);
6821 free_percpu(sdd->sg);
6823 free_percpu(sdd->sgp);
6828 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6829 struct s_data *d, const struct cpumask *cpu_map,
6830 struct sched_domain_attr *attr, struct sched_domain *child,
6833 struct sched_domain *sd = tl->init(tl, cpu);
6837 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6839 sd->level = child->level + 1;
6840 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6844 set_domain_attribute(sd, attr);
6850 * Build sched domains for a given set of cpus and attach the sched domains
6851 * to the individual cpus
6853 static int build_sched_domains(const struct cpumask *cpu_map,
6854 struct sched_domain_attr *attr)
6856 enum s_alloc alloc_state = sa_none;
6857 struct sched_domain *sd;
6859 int i, ret = -ENOMEM;
6861 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6862 if (alloc_state != sa_rootdomain)
6865 /* Set up domains for cpus specified by the cpu_map. */
6866 for_each_cpu(i, cpu_map) {
6867 struct sched_domain_topology_level *tl;
6870 for (tl = sched_domain_topology; tl->init; tl++) {
6871 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6872 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6873 sd->flags |= SD_OVERLAP;
6874 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6881 *per_cpu_ptr(d.sd, i) = sd;
6884 /* Build the groups for the domains */
6885 for_each_cpu(i, cpu_map) {
6886 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6887 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6888 if (sd->flags & SD_OVERLAP) {
6889 if (build_overlap_sched_groups(sd, i))
6892 if (build_sched_groups(sd, i))
6898 /* Calculate CPU power for physical packages and nodes */
6899 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6900 if (!cpumask_test_cpu(i, cpu_map))
6903 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6904 claim_allocations(i, sd);
6905 init_sched_groups_power(i, sd);
6909 /* Attach the domains */
6911 for_each_cpu(i, cpu_map) {
6912 sd = *per_cpu_ptr(d.sd, i);
6913 cpu_attach_domain(sd, d.rd, i);
6919 __free_domain_allocs(&d, alloc_state, cpu_map);
6923 static cpumask_var_t *doms_cur; /* current sched domains */
6924 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6925 static struct sched_domain_attr *dattr_cur;
6926 /* attribues of custom domains in 'doms_cur' */
6929 * Special case: If a kmalloc of a doms_cur partition (array of
6930 * cpumask) fails, then fallback to a single sched domain,
6931 * as determined by the single cpumask fallback_doms.
6933 static cpumask_var_t fallback_doms;
6936 * arch_update_cpu_topology lets virtualized architectures update the
6937 * cpu core maps. It is supposed to return 1 if the topology changed
6938 * or 0 if it stayed the same.
6940 int __attribute__((weak)) arch_update_cpu_topology(void)
6945 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6948 cpumask_var_t *doms;
6950 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6953 for (i = 0; i < ndoms; i++) {
6954 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6955 free_sched_domains(doms, i);
6962 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6965 for (i = 0; i < ndoms; i++)
6966 free_cpumask_var(doms[i]);
6971 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6972 * For now this just excludes isolated cpus, but could be used to
6973 * exclude other special cases in the future.
6975 static int init_sched_domains(const struct cpumask *cpu_map)
6979 arch_update_cpu_topology();
6981 doms_cur = alloc_sched_domains(ndoms_cur);
6983 doms_cur = &fallback_doms;
6984 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6985 err = build_sched_domains(doms_cur[0], NULL);
6986 register_sched_domain_sysctl();
6992 * Detach sched domains from a group of cpus specified in cpu_map
6993 * These cpus will now be attached to the NULL domain
6995 static void detach_destroy_domains(const struct cpumask *cpu_map)
7000 for_each_cpu(i, cpu_map)
7001 cpu_attach_domain(NULL, &def_root_domain, i);
7005 /* handle null as "default" */
7006 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7007 struct sched_domain_attr *new, int idx_new)
7009 struct sched_domain_attr tmp;
7016 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7017 new ? (new + idx_new) : &tmp,
7018 sizeof(struct sched_domain_attr));
7022 * Partition sched domains as specified by the 'ndoms_new'
7023 * cpumasks in the array doms_new[] of cpumasks. This compares
7024 * doms_new[] to the current sched domain partitioning, doms_cur[].
7025 * It destroys each deleted domain and builds each new domain.
7027 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7028 * The masks don't intersect (don't overlap.) We should setup one
7029 * sched domain for each mask. CPUs not in any of the cpumasks will
7030 * not be load balanced. If the same cpumask appears both in the
7031 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7034 * The passed in 'doms_new' should be allocated using
7035 * alloc_sched_domains. This routine takes ownership of it and will
7036 * free_sched_domains it when done with it. If the caller failed the
7037 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7038 * and partition_sched_domains() will fallback to the single partition
7039 * 'fallback_doms', it also forces the domains to be rebuilt.
7041 * If doms_new == NULL it will be replaced with cpu_online_mask.
7042 * ndoms_new == 0 is a special case for destroying existing domains,
7043 * and it will not create the default domain.
7045 * Call with hotplug lock held
7047 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7048 struct sched_domain_attr *dattr_new)
7053 mutex_lock(&sched_domains_mutex);
7055 /* always unregister in case we don't destroy any domains */
7056 unregister_sched_domain_sysctl();
7058 /* Let architecture update cpu core mappings. */
7059 new_topology = arch_update_cpu_topology();
7061 n = doms_new ? ndoms_new : 0;
7063 /* Destroy deleted domains */
7064 for (i = 0; i < ndoms_cur; i++) {
7065 for (j = 0; j < n && !new_topology; j++) {
7066 if (cpumask_equal(doms_cur[i], doms_new[j])
7067 && dattrs_equal(dattr_cur, i, dattr_new, j))
7070 /* no match - a current sched domain not in new doms_new[] */
7071 detach_destroy_domains(doms_cur[i]);
7076 if (doms_new == NULL) {
7078 doms_new = &fallback_doms;
7079 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7080 WARN_ON_ONCE(dattr_new);
7083 /* Build new domains */
7084 for (i = 0; i < ndoms_new; i++) {
7085 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7086 if (cpumask_equal(doms_new[i], doms_cur[j])
7087 && dattrs_equal(dattr_new, i, dattr_cur, j))
7090 /* no match - add a new doms_new */
7091 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7096 /* Remember the new sched domains */
7097 if (doms_cur != &fallback_doms)
7098 free_sched_domains(doms_cur, ndoms_cur);
7099 kfree(dattr_cur); /* kfree(NULL) is safe */
7100 doms_cur = doms_new;
7101 dattr_cur = dattr_new;
7102 ndoms_cur = ndoms_new;
7104 register_sched_domain_sysctl();
7106 mutex_unlock(&sched_domains_mutex);
7109 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7112 * Update cpusets according to cpu_active mask. If cpusets are
7113 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7114 * around partition_sched_domains().
7116 * If we come here as part of a suspend/resume, don't touch cpusets because we
7117 * want to restore it back to its original state upon resume anyway.
7119 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7123 case CPU_ONLINE_FROZEN:
7124 case CPU_DOWN_FAILED_FROZEN:
7127 * num_cpus_frozen tracks how many CPUs are involved in suspend
7128 * resume sequence. As long as this is not the last online
7129 * operation in the resume sequence, just build a single sched
7130 * domain, ignoring cpusets.
7133 if (likely(num_cpus_frozen)) {
7134 partition_sched_domains(1, NULL, NULL);
7139 * This is the last CPU online operation. So fall through and
7140 * restore the original sched domains by considering the
7141 * cpuset configurations.
7145 case CPU_DOWN_FAILED:
7146 cpuset_update_active_cpus(true);
7154 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7158 case CPU_DOWN_PREPARE:
7159 cpuset_update_active_cpus(false);
7161 case CPU_DOWN_PREPARE_FROZEN:
7163 partition_sched_domains(1, NULL, NULL);
7171 void __init sched_init_smp(void)
7173 cpumask_var_t non_isolated_cpus;
7175 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7176 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7181 mutex_lock(&sched_domains_mutex);
7182 init_sched_domains(cpu_active_mask);
7183 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7184 if (cpumask_empty(non_isolated_cpus))
7185 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7186 mutex_unlock(&sched_domains_mutex);
7189 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7190 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7192 /* RT runtime code needs to handle some hotplug events */
7193 hotcpu_notifier(update_runtime, 0);
7197 /* Move init over to a non-isolated CPU */
7198 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7200 sched_init_granularity();
7201 free_cpumask_var(non_isolated_cpus);
7203 init_sched_rt_class();
7206 void __init sched_init_smp(void)
7208 sched_init_granularity();
7210 #endif /* CONFIG_SMP */
7212 const_debug unsigned int sysctl_timer_migration = 1;
7214 int in_sched_functions(unsigned long addr)
7216 return in_lock_functions(addr) ||
7217 (addr >= (unsigned long)__sched_text_start
7218 && addr < (unsigned long)__sched_text_end);
7221 #ifdef CONFIG_CGROUP_SCHED
7222 struct task_group root_task_group;
7223 LIST_HEAD(task_groups);
7226 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7228 void __init sched_init(void)
7231 unsigned long alloc_size = 0, ptr;
7233 #ifdef CONFIG_FAIR_GROUP_SCHED
7234 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7236 #ifdef CONFIG_RT_GROUP_SCHED
7237 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7239 #ifdef CONFIG_CPUMASK_OFFSTACK
7240 alloc_size += num_possible_cpus() * cpumask_size();
7243 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7245 #ifdef CONFIG_FAIR_GROUP_SCHED
7246 root_task_group.se = (struct sched_entity **)ptr;
7247 ptr += nr_cpu_ids * sizeof(void **);
7249 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7250 ptr += nr_cpu_ids * sizeof(void **);
7252 #endif /* CONFIG_FAIR_GROUP_SCHED */
7253 #ifdef CONFIG_RT_GROUP_SCHED
7254 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7255 ptr += nr_cpu_ids * sizeof(void **);
7257 root_task_group.rt_rq = (struct rt_rq **)ptr;
7258 ptr += nr_cpu_ids * sizeof(void **);
7260 #endif /* CONFIG_RT_GROUP_SCHED */
7261 #ifdef CONFIG_CPUMASK_OFFSTACK
7262 for_each_possible_cpu(i) {
7263 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7264 ptr += cpumask_size();
7266 #endif /* CONFIG_CPUMASK_OFFSTACK */
7270 init_defrootdomain();
7273 init_rt_bandwidth(&def_rt_bandwidth,
7274 global_rt_period(), global_rt_runtime());
7276 #ifdef CONFIG_RT_GROUP_SCHED
7277 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7278 global_rt_period(), global_rt_runtime());
7279 #endif /* CONFIG_RT_GROUP_SCHED */
7281 #ifdef CONFIG_CGROUP_SCHED
7282 list_add(&root_task_group.list, &task_groups);
7283 INIT_LIST_HEAD(&root_task_group.children);
7284 INIT_LIST_HEAD(&root_task_group.siblings);
7285 autogroup_init(&init_task);
7287 #endif /* CONFIG_CGROUP_SCHED */
7289 #ifdef CONFIG_CGROUP_CPUACCT
7290 root_cpuacct.cpustat = &kernel_cpustat;
7291 root_cpuacct.cpuusage = alloc_percpu(u64);
7292 /* Too early, not expected to fail */
7293 BUG_ON(!root_cpuacct.cpuusage);
7295 for_each_possible_cpu(i) {
7299 raw_spin_lock_init(&rq->lock);
7301 rq->calc_load_active = 0;
7302 rq->calc_load_update = jiffies + LOAD_FREQ;
7303 init_cfs_rq(&rq->cfs);
7304 init_rt_rq(&rq->rt, rq);
7305 #ifdef CONFIG_FAIR_GROUP_SCHED
7306 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7307 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7309 * How much cpu bandwidth does root_task_group get?
7311 * In case of task-groups formed thr' the cgroup filesystem, it
7312 * gets 100% of the cpu resources in the system. This overall
7313 * system cpu resource is divided among the tasks of
7314 * root_task_group and its child task-groups in a fair manner,
7315 * based on each entity's (task or task-group's) weight
7316 * (se->load.weight).
7318 * In other words, if root_task_group has 10 tasks of weight
7319 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7320 * then A0's share of the cpu resource is:
7322 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7324 * We achieve this by letting root_task_group's tasks sit
7325 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7327 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7328 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7329 #endif /* CONFIG_FAIR_GROUP_SCHED */
7331 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7332 #ifdef CONFIG_RT_GROUP_SCHED
7333 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7334 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7337 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7338 rq->cpu_load[j] = 0;
7340 rq->last_load_update_tick = jiffies;
7345 rq->cpu_power = SCHED_POWER_SCALE;
7346 rq->post_schedule = 0;
7347 rq->active_balance = 0;
7348 rq->next_balance = jiffies;
7353 rq->avg_idle = 2*sysctl_sched_migration_cost;
7355 INIT_LIST_HEAD(&rq->cfs_tasks);
7357 rq_attach_root(rq, &def_root_domain);
7363 atomic_set(&rq->nr_iowait, 0);
7366 set_load_weight(&init_task);
7368 #ifdef CONFIG_PREEMPT_NOTIFIERS
7369 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7372 #ifdef CONFIG_RT_MUTEXES
7373 plist_head_init(&init_task.pi_waiters);
7377 * The boot idle thread does lazy MMU switching as well:
7379 atomic_inc(&init_mm.mm_count);
7380 enter_lazy_tlb(&init_mm, current);
7383 * Make us the idle thread. Technically, schedule() should not be
7384 * called from this thread, however somewhere below it might be,
7385 * but because we are the idle thread, we just pick up running again
7386 * when this runqueue becomes "idle".
7388 init_idle(current, smp_processor_id());
7390 calc_load_update = jiffies + LOAD_FREQ;
7393 * During early bootup we pretend to be a normal task:
7395 current->sched_class = &fair_sched_class;
7398 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7399 /* May be allocated at isolcpus cmdline parse time */
7400 if (cpu_isolated_map == NULL)
7401 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7402 idle_thread_set_boot_cpu();
7404 init_sched_fair_class();
7406 scheduler_running = 1;
7409 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7410 static inline int preempt_count_equals(int preempt_offset)
7412 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7414 return (nested == preempt_offset);
7417 void __might_sleep(const char *file, int line, int preempt_offset)
7419 static unsigned long prev_jiffy; /* ratelimiting */
7421 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7422 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7423 system_state != SYSTEM_RUNNING || oops_in_progress)
7425 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7427 prev_jiffy = jiffies;
7430 "BUG: sleeping function called from invalid context at %s:%d\n",
7433 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7434 in_atomic(), irqs_disabled(),
7435 current->pid, current->comm);
7437 debug_show_held_locks(current);
7438 if (irqs_disabled())
7439 print_irqtrace_events(current);
7442 EXPORT_SYMBOL(__might_sleep);
7445 #ifdef CONFIG_MAGIC_SYSRQ
7446 static void normalize_task(struct rq *rq, struct task_struct *p)
7448 const struct sched_class *prev_class = p->sched_class;
7449 int old_prio = p->prio;
7454 dequeue_task(rq, p, 0);
7455 __setscheduler(rq, p, SCHED_NORMAL, 0);
7457 enqueue_task(rq, p, 0);
7458 resched_task(rq->curr);
7461 check_class_changed(rq, p, prev_class, old_prio);
7464 void normalize_rt_tasks(void)
7466 struct task_struct *g, *p;
7467 unsigned long flags;
7470 read_lock_irqsave(&tasklist_lock, flags);
7471 do_each_thread(g, p) {
7473 * Only normalize user tasks:
7478 p->se.exec_start = 0;
7479 #ifdef CONFIG_SCHEDSTATS
7480 p->se.statistics.wait_start = 0;
7481 p->se.statistics.sleep_start = 0;
7482 p->se.statistics.block_start = 0;
7487 * Renice negative nice level userspace
7490 if (TASK_NICE(p) < 0 && p->mm)
7491 set_user_nice(p, 0);
7495 raw_spin_lock(&p->pi_lock);
7496 rq = __task_rq_lock(p);
7498 normalize_task(rq, p);
7500 __task_rq_unlock(rq);
7501 raw_spin_unlock(&p->pi_lock);
7502 } while_each_thread(g, p);
7504 read_unlock_irqrestore(&tasklist_lock, flags);
7507 #endif /* CONFIG_MAGIC_SYSRQ */
7509 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7511 * These functions are only useful for the IA64 MCA handling, or kdb.
7513 * They can only be called when the whole system has been
7514 * stopped - every CPU needs to be quiescent, and no scheduling
7515 * activity can take place. Using them for anything else would
7516 * be a serious bug, and as a result, they aren't even visible
7517 * under any other configuration.
7521 * curr_task - return the current task for a given cpu.
7522 * @cpu: the processor in question.
7524 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7526 struct task_struct *curr_task(int cpu)
7528 return cpu_curr(cpu);
7531 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7535 * set_curr_task - set the current task for a given cpu.
7536 * @cpu: the processor in question.
7537 * @p: the task pointer to set.
7539 * Description: This function must only be used when non-maskable interrupts
7540 * are serviced on a separate stack. It allows the architecture to switch the
7541 * notion of the current task on a cpu in a non-blocking manner. This function
7542 * must be called with all CPU's synchronized, and interrupts disabled, the
7543 * and caller must save the original value of the current task (see
7544 * curr_task() above) and restore that value before reenabling interrupts and
7545 * re-starting the system.
7547 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7549 void set_curr_task(int cpu, struct task_struct *p)
7556 #ifdef CONFIG_CGROUP_SCHED
7557 /* task_group_lock serializes the addition/removal of task groups */
7558 static DEFINE_SPINLOCK(task_group_lock);
7560 static void free_sched_group(struct task_group *tg)
7562 free_fair_sched_group(tg);
7563 free_rt_sched_group(tg);
7568 /* allocate runqueue etc for a new task group */
7569 struct task_group *sched_create_group(struct task_group *parent)
7571 struct task_group *tg;
7572 unsigned long flags;
7574 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7576 return ERR_PTR(-ENOMEM);
7578 if (!alloc_fair_sched_group(tg, parent))
7581 if (!alloc_rt_sched_group(tg, parent))
7584 spin_lock_irqsave(&task_group_lock, flags);
7585 list_add_rcu(&tg->list, &task_groups);
7587 WARN_ON(!parent); /* root should already exist */
7589 tg->parent = parent;
7590 INIT_LIST_HEAD(&tg->children);
7591 list_add_rcu(&tg->siblings, &parent->children);
7592 spin_unlock_irqrestore(&task_group_lock, flags);
7597 free_sched_group(tg);
7598 return ERR_PTR(-ENOMEM);
7601 /* rcu callback to free various structures associated with a task group */
7602 static void free_sched_group_rcu(struct rcu_head *rhp)
7604 /* now it should be safe to free those cfs_rqs */
7605 free_sched_group(container_of(rhp, struct task_group, rcu));
7608 /* Destroy runqueue etc associated with a task group */
7609 void sched_destroy_group(struct task_group *tg)
7611 unsigned long flags;
7614 /* end participation in shares distribution */
7615 for_each_possible_cpu(i)
7616 unregister_fair_sched_group(tg, i);
7618 spin_lock_irqsave(&task_group_lock, flags);
7619 list_del_rcu(&tg->list);
7620 list_del_rcu(&tg->siblings);
7621 spin_unlock_irqrestore(&task_group_lock, flags);
7623 /* wait for possible concurrent references to cfs_rqs complete */
7624 call_rcu(&tg->rcu, free_sched_group_rcu);
7627 /* change task's runqueue when it moves between groups.
7628 * The caller of this function should have put the task in its new group
7629 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7630 * reflect its new group.
7632 void sched_move_task(struct task_struct *tsk)
7634 struct task_group *tg;
7636 unsigned long flags;
7639 rq = task_rq_lock(tsk, &flags);
7641 running = task_current(rq, tsk);
7645 dequeue_task(rq, tsk, 0);
7646 if (unlikely(running))
7647 tsk->sched_class->put_prev_task(rq, tsk);
7649 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7650 lockdep_is_held(&tsk->sighand->siglock)),
7651 struct task_group, css);
7652 tg = autogroup_task_group(tsk, tg);
7653 tsk->sched_task_group = tg;
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 if (tsk->sched_class->task_move_group)
7657 tsk->sched_class->task_move_group(tsk, on_rq);
7660 set_task_rq(tsk, task_cpu(tsk));
7662 if (unlikely(running))
7663 tsk->sched_class->set_curr_task(rq);
7665 enqueue_task(rq, tsk, 0);
7667 task_rq_unlock(rq, tsk, &flags);
7669 #endif /* CONFIG_CGROUP_SCHED */
7671 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7672 static unsigned long to_ratio(u64 period, u64 runtime)
7674 if (runtime == RUNTIME_INF)
7677 return div64_u64(runtime << 20, period);
7681 #ifdef CONFIG_RT_GROUP_SCHED
7683 * Ensure that the real time constraints are schedulable.
7685 static DEFINE_MUTEX(rt_constraints_mutex);
7687 /* Must be called with tasklist_lock held */
7688 static inline int tg_has_rt_tasks(struct task_group *tg)
7690 struct task_struct *g, *p;
7692 do_each_thread(g, p) {
7693 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7695 } while_each_thread(g, p);
7700 struct rt_schedulable_data {
7701 struct task_group *tg;
7706 static int tg_rt_schedulable(struct task_group *tg, void *data)
7708 struct rt_schedulable_data *d = data;
7709 struct task_group *child;
7710 unsigned long total, sum = 0;
7711 u64 period, runtime;
7713 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7714 runtime = tg->rt_bandwidth.rt_runtime;
7717 period = d->rt_period;
7718 runtime = d->rt_runtime;
7722 * Cannot have more runtime than the period.
7724 if (runtime > period && runtime != RUNTIME_INF)
7728 * Ensure we don't starve existing RT tasks.
7730 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7733 total = to_ratio(period, runtime);
7736 * Nobody can have more than the global setting allows.
7738 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7742 * The sum of our children's runtime should not exceed our own.
7744 list_for_each_entry_rcu(child, &tg->children, siblings) {
7745 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7746 runtime = child->rt_bandwidth.rt_runtime;
7748 if (child == d->tg) {
7749 period = d->rt_period;
7750 runtime = d->rt_runtime;
7753 sum += to_ratio(period, runtime);
7762 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7766 struct rt_schedulable_data data = {
7768 .rt_period = period,
7769 .rt_runtime = runtime,
7773 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7779 static int tg_set_rt_bandwidth(struct task_group *tg,
7780 u64 rt_period, u64 rt_runtime)
7784 mutex_lock(&rt_constraints_mutex);
7785 read_lock(&tasklist_lock);
7786 err = __rt_schedulable(tg, rt_period, rt_runtime);
7790 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7791 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7792 tg->rt_bandwidth.rt_runtime = rt_runtime;
7794 for_each_possible_cpu(i) {
7795 struct rt_rq *rt_rq = tg->rt_rq[i];
7797 raw_spin_lock(&rt_rq->rt_runtime_lock);
7798 rt_rq->rt_runtime = rt_runtime;
7799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7801 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7803 read_unlock(&tasklist_lock);
7804 mutex_unlock(&rt_constraints_mutex);
7809 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7811 u64 rt_runtime, rt_period;
7813 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7814 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7815 if (rt_runtime_us < 0)
7816 rt_runtime = RUNTIME_INF;
7818 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7821 long sched_group_rt_runtime(struct task_group *tg)
7825 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7828 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7829 do_div(rt_runtime_us, NSEC_PER_USEC);
7830 return rt_runtime_us;
7833 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7835 u64 rt_runtime, rt_period;
7837 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7838 rt_runtime = tg->rt_bandwidth.rt_runtime;
7843 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7846 long sched_group_rt_period(struct task_group *tg)
7850 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7851 do_div(rt_period_us, NSEC_PER_USEC);
7852 return rt_period_us;
7855 static int sched_rt_global_constraints(void)
7857 u64 runtime, period;
7860 if (sysctl_sched_rt_period <= 0)
7863 runtime = global_rt_runtime();
7864 period = global_rt_period();
7867 * Sanity check on the sysctl variables.
7869 if (runtime > period && runtime != RUNTIME_INF)
7872 mutex_lock(&rt_constraints_mutex);
7873 read_lock(&tasklist_lock);
7874 ret = __rt_schedulable(NULL, 0, 0);
7875 read_unlock(&tasklist_lock);
7876 mutex_unlock(&rt_constraints_mutex);
7881 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7883 /* Don't accept realtime tasks when there is no way for them to run */
7884 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7890 #else /* !CONFIG_RT_GROUP_SCHED */
7891 static int sched_rt_global_constraints(void)
7893 unsigned long flags;
7896 if (sysctl_sched_rt_period <= 0)
7900 * There's always some RT tasks in the root group
7901 * -- migration, kstopmachine etc..
7903 if (sysctl_sched_rt_runtime == 0)
7906 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7907 for_each_possible_cpu(i) {
7908 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7910 raw_spin_lock(&rt_rq->rt_runtime_lock);
7911 rt_rq->rt_runtime = global_rt_runtime();
7912 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7914 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7918 #endif /* CONFIG_RT_GROUP_SCHED */
7920 int sched_rt_handler(struct ctl_table *table, int write,
7921 void __user *buffer, size_t *lenp,
7925 int old_period, old_runtime;
7926 static DEFINE_MUTEX(mutex);
7929 old_period = sysctl_sched_rt_period;
7930 old_runtime = sysctl_sched_rt_runtime;
7932 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7934 if (!ret && write) {
7935 ret = sched_rt_global_constraints();
7937 sysctl_sched_rt_period = old_period;
7938 sysctl_sched_rt_runtime = old_runtime;
7940 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7941 def_rt_bandwidth.rt_period =
7942 ns_to_ktime(global_rt_period());
7945 mutex_unlock(&mutex);
7950 #ifdef CONFIG_CGROUP_SCHED
7952 /* return corresponding task_group object of a cgroup */
7953 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7955 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7956 struct task_group, css);
7959 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7961 struct task_group *tg, *parent;
7963 if (!cgrp->parent) {
7964 /* This is early initialization for the top cgroup */
7965 return &root_task_group.css;
7968 parent = cgroup_tg(cgrp->parent);
7969 tg = sched_create_group(parent);
7971 return ERR_PTR(-ENOMEM);
7976 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7978 struct task_group *tg = cgroup_tg(cgrp);
7980 sched_destroy_group(tg);
7983 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7984 struct cgroup_taskset *tset)
7986 struct task_struct *task;
7988 cgroup_taskset_for_each(task, cgrp, tset) {
7989 #ifdef CONFIG_RT_GROUP_SCHED
7990 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7993 /* We don't support RT-tasks being in separate groups */
7994 if (task->sched_class != &fair_sched_class)
8001 static void cpu_cgroup_attach(struct cgroup *cgrp,
8002 struct cgroup_taskset *tset)
8004 struct task_struct *task;
8006 cgroup_taskset_for_each(task, cgrp, tset)
8007 sched_move_task(task);
8011 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
8012 struct task_struct *task)
8015 * cgroup_exit() is called in the copy_process() failure path.
8016 * Ignore this case since the task hasn't ran yet, this avoids
8017 * trying to poke a half freed task state from generic code.
8019 if (!(task->flags & PF_EXITING))
8022 sched_move_task(task);
8025 #ifdef CONFIG_FAIR_GROUP_SCHED
8026 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8029 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8032 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8034 struct task_group *tg = cgroup_tg(cgrp);
8036 return (u64) scale_load_down(tg->shares);
8039 #ifdef CONFIG_CFS_BANDWIDTH
8040 static DEFINE_MUTEX(cfs_constraints_mutex);
8042 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8043 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8045 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8047 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8049 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8050 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8052 if (tg == &root_task_group)
8056 * Ensure we have at some amount of bandwidth every period. This is
8057 * to prevent reaching a state of large arrears when throttled via
8058 * entity_tick() resulting in prolonged exit starvation.
8060 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8064 * Likewise, bound things on the otherside by preventing insane quota
8065 * periods. This also allows us to normalize in computing quota
8068 if (period > max_cfs_quota_period)
8071 mutex_lock(&cfs_constraints_mutex);
8072 ret = __cfs_schedulable(tg, period, quota);
8076 runtime_enabled = quota != RUNTIME_INF;
8077 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8078 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8079 raw_spin_lock_irq(&cfs_b->lock);
8080 cfs_b->period = ns_to_ktime(period);
8081 cfs_b->quota = quota;
8083 __refill_cfs_bandwidth_runtime(cfs_b);
8084 /* restart the period timer (if active) to handle new period expiry */
8085 if (runtime_enabled && cfs_b->timer_active) {
8086 /* force a reprogram */
8087 cfs_b->timer_active = 0;
8088 __start_cfs_bandwidth(cfs_b);
8090 raw_spin_unlock_irq(&cfs_b->lock);
8092 for_each_possible_cpu(i) {
8093 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8094 struct rq *rq = cfs_rq->rq;
8096 raw_spin_lock_irq(&rq->lock);
8097 cfs_rq->runtime_enabled = runtime_enabled;
8098 cfs_rq->runtime_remaining = 0;
8100 if (cfs_rq->throttled)
8101 unthrottle_cfs_rq(cfs_rq);
8102 raw_spin_unlock_irq(&rq->lock);
8105 mutex_unlock(&cfs_constraints_mutex);
8110 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8114 period = ktime_to_ns(tg->cfs_bandwidth.period);
8115 if (cfs_quota_us < 0)
8116 quota = RUNTIME_INF;
8118 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8120 return tg_set_cfs_bandwidth(tg, period, quota);
8123 long tg_get_cfs_quota(struct task_group *tg)
8127 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8130 quota_us = tg->cfs_bandwidth.quota;
8131 do_div(quota_us, NSEC_PER_USEC);
8136 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8140 period = (u64)cfs_period_us * NSEC_PER_USEC;
8141 quota = tg->cfs_bandwidth.quota;
8143 return tg_set_cfs_bandwidth(tg, period, quota);
8146 long tg_get_cfs_period(struct task_group *tg)
8150 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8151 do_div(cfs_period_us, NSEC_PER_USEC);
8153 return cfs_period_us;
8156 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8158 return tg_get_cfs_quota(cgroup_tg(cgrp));
8161 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8164 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8167 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8169 return tg_get_cfs_period(cgroup_tg(cgrp));
8172 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8175 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8178 struct cfs_schedulable_data {
8179 struct task_group *tg;
8184 * normalize group quota/period to be quota/max_period
8185 * note: units are usecs
8187 static u64 normalize_cfs_quota(struct task_group *tg,
8188 struct cfs_schedulable_data *d)
8196 period = tg_get_cfs_period(tg);
8197 quota = tg_get_cfs_quota(tg);
8200 /* note: these should typically be equivalent */
8201 if (quota == RUNTIME_INF || quota == -1)
8204 return to_ratio(period, quota);
8207 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8209 struct cfs_schedulable_data *d = data;
8210 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8211 s64 quota = 0, parent_quota = -1;
8214 quota = RUNTIME_INF;
8216 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8218 quota = normalize_cfs_quota(tg, d);
8219 parent_quota = parent_b->hierarchal_quota;
8222 * ensure max(child_quota) <= parent_quota, inherit when no
8225 if (quota == RUNTIME_INF)
8226 quota = parent_quota;
8227 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8230 cfs_b->hierarchal_quota = quota;
8235 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8238 struct cfs_schedulable_data data = {
8244 if (quota != RUNTIME_INF) {
8245 do_div(data.period, NSEC_PER_USEC);
8246 do_div(data.quota, NSEC_PER_USEC);
8250 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8256 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8257 struct cgroup_map_cb *cb)
8259 struct task_group *tg = cgroup_tg(cgrp);
8260 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8262 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8263 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8264 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8268 #endif /* CONFIG_CFS_BANDWIDTH */
8269 #endif /* CONFIG_FAIR_GROUP_SCHED */
8271 #ifdef CONFIG_RT_GROUP_SCHED
8272 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8275 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8278 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8280 return sched_group_rt_runtime(cgroup_tg(cgrp));
8283 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8286 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8289 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8291 return sched_group_rt_period(cgroup_tg(cgrp));
8293 #endif /* CONFIG_RT_GROUP_SCHED */
8295 static struct cftype cpu_files[] = {
8296 #ifdef CONFIG_FAIR_GROUP_SCHED
8299 .read_u64 = cpu_shares_read_u64,
8300 .write_u64 = cpu_shares_write_u64,
8303 #ifdef CONFIG_CFS_BANDWIDTH
8305 .name = "cfs_quota_us",
8306 .read_s64 = cpu_cfs_quota_read_s64,
8307 .write_s64 = cpu_cfs_quota_write_s64,
8310 .name = "cfs_period_us",
8311 .read_u64 = cpu_cfs_period_read_u64,
8312 .write_u64 = cpu_cfs_period_write_u64,
8316 .read_map = cpu_stats_show,
8319 #ifdef CONFIG_RT_GROUP_SCHED
8321 .name = "rt_runtime_us",
8322 .read_s64 = cpu_rt_runtime_read,
8323 .write_s64 = cpu_rt_runtime_write,
8326 .name = "rt_period_us",
8327 .read_u64 = cpu_rt_period_read_uint,
8328 .write_u64 = cpu_rt_period_write_uint,
8334 struct cgroup_subsys cpu_cgroup_subsys = {
8336 .create = cpu_cgroup_create,
8337 .destroy = cpu_cgroup_destroy,
8338 .can_attach = cpu_cgroup_can_attach,
8339 .attach = cpu_cgroup_attach,
8340 .exit = cpu_cgroup_exit,
8341 .subsys_id = cpu_cgroup_subsys_id,
8342 .base_cftypes = cpu_files,
8346 #endif /* CONFIG_CGROUP_SCHED */
8348 #ifdef CONFIG_CGROUP_CPUACCT
8351 * CPU accounting code for task groups.
8353 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8354 * (balbir@in.ibm.com).
8357 /* create a new cpu accounting group */
8358 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8363 return &root_cpuacct.css;
8365 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8369 ca->cpuusage = alloc_percpu(u64);
8373 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8375 goto out_free_cpuusage;
8380 free_percpu(ca->cpuusage);
8384 return ERR_PTR(-ENOMEM);
8387 /* destroy an existing cpu accounting group */
8388 static void cpuacct_destroy(struct cgroup *cgrp)
8390 struct cpuacct *ca = cgroup_ca(cgrp);
8392 free_percpu(ca->cpustat);
8393 free_percpu(ca->cpuusage);
8397 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8399 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8402 #ifndef CONFIG_64BIT
8404 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8406 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8408 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8416 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8418 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8420 #ifndef CONFIG_64BIT
8422 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8424 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8426 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8432 /* return total cpu usage (in nanoseconds) of a group */
8433 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8435 struct cpuacct *ca = cgroup_ca(cgrp);
8436 u64 totalcpuusage = 0;
8439 for_each_present_cpu(i)
8440 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8442 return totalcpuusage;
8445 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8448 struct cpuacct *ca = cgroup_ca(cgrp);
8457 for_each_present_cpu(i)
8458 cpuacct_cpuusage_write(ca, i, 0);
8464 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8467 struct cpuacct *ca = cgroup_ca(cgroup);
8471 for_each_present_cpu(i) {
8472 percpu = cpuacct_cpuusage_read(ca, i);
8473 seq_printf(m, "%llu ", (unsigned long long) percpu);
8475 seq_printf(m, "\n");
8479 static const char *cpuacct_stat_desc[] = {
8480 [CPUACCT_STAT_USER] = "user",
8481 [CPUACCT_STAT_SYSTEM] = "system",
8484 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8485 struct cgroup_map_cb *cb)
8487 struct cpuacct *ca = cgroup_ca(cgrp);
8491 for_each_online_cpu(cpu) {
8492 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8493 val += kcpustat->cpustat[CPUTIME_USER];
8494 val += kcpustat->cpustat[CPUTIME_NICE];
8496 val = cputime64_to_clock_t(val);
8497 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8500 for_each_online_cpu(cpu) {
8501 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8502 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8503 val += kcpustat->cpustat[CPUTIME_IRQ];
8504 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8507 val = cputime64_to_clock_t(val);
8508 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8513 static struct cftype files[] = {
8516 .read_u64 = cpuusage_read,
8517 .write_u64 = cpuusage_write,
8520 .name = "usage_percpu",
8521 .read_seq_string = cpuacct_percpu_seq_read,
8525 .read_map = cpuacct_stats_show,
8531 * charge this task's execution time to its accounting group.
8533 * called with rq->lock held.
8535 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8540 if (unlikely(!cpuacct_subsys.active))
8543 cpu = task_cpu(tsk);
8549 for (; ca; ca = parent_ca(ca)) {
8550 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8551 *cpuusage += cputime;
8557 struct cgroup_subsys cpuacct_subsys = {
8559 .create = cpuacct_create,
8560 .destroy = cpuacct_destroy,
8561 .subsys_id = cpuacct_subsys_id,
8562 .base_cftypes = files,
8564 #endif /* CONFIG_CGROUP_CPUACCT */