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 switch_to(prev, next, prev);
2088 * this_rq must be evaluated again because prev may have moved
2089 * CPUs since it called schedule(), thus the 'rq' on its stack
2090 * frame will be invalid.
2092 finish_task_switch(this_rq(), prev);
2096 * nr_running, nr_uninterruptible and nr_context_switches:
2098 * externally visible scheduler statistics: current number of runnable
2099 * threads, current number of uninterruptible-sleeping threads, total
2100 * number of context switches performed since bootup.
2102 unsigned long nr_running(void)
2104 unsigned long i, sum = 0;
2106 for_each_online_cpu(i)
2107 sum += cpu_rq(i)->nr_running;
2112 unsigned long nr_uninterruptible(void)
2114 unsigned long i, sum = 0;
2116 for_each_possible_cpu(i)
2117 sum += cpu_rq(i)->nr_uninterruptible;
2120 * Since we read the counters lockless, it might be slightly
2121 * inaccurate. Do not allow it to go below zero though:
2123 if (unlikely((long)sum < 0))
2129 unsigned long long nr_context_switches(void)
2132 unsigned long long sum = 0;
2134 for_each_possible_cpu(i)
2135 sum += cpu_rq(i)->nr_switches;
2140 unsigned long nr_iowait(void)
2142 unsigned long i, sum = 0;
2144 for_each_possible_cpu(i)
2145 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2150 unsigned long nr_iowait_cpu(int cpu)
2152 struct rq *this = cpu_rq(cpu);
2153 return atomic_read(&this->nr_iowait);
2156 unsigned long this_cpu_load(void)
2158 struct rq *this = this_rq();
2159 return this->cpu_load[0];
2164 * Global load-average calculations
2166 * We take a distributed and async approach to calculating the global load-avg
2167 * in order to minimize overhead.
2169 * The global load average is an exponentially decaying average of nr_running +
2170 * nr_uninterruptible.
2172 * Once every LOAD_FREQ:
2175 * for_each_possible_cpu(cpu)
2176 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2178 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2180 * Due to a number of reasons the above turns in the mess below:
2182 * - for_each_possible_cpu() is prohibitively expensive on machines with
2183 * serious number of cpus, therefore we need to take a distributed approach
2184 * to calculating nr_active.
2186 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2187 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2189 * So assuming nr_active := 0 when we start out -- true per definition, we
2190 * can simply take per-cpu deltas and fold those into a global accumulate
2191 * to obtain the same result. See calc_load_fold_active().
2193 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2194 * across the machine, we assume 10 ticks is sufficient time for every
2195 * cpu to have completed this task.
2197 * This places an upper-bound on the IRQ-off latency of the machine. Then
2198 * again, being late doesn't loose the delta, just wrecks the sample.
2200 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2201 * this would add another cross-cpu cacheline miss and atomic operation
2202 * to the wakeup path. Instead we increment on whatever cpu the task ran
2203 * when it went into uninterruptible state and decrement on whatever cpu
2204 * did the wakeup. This means that only the sum of nr_uninterruptible over
2205 * all cpus yields the correct result.
2207 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2210 /* Variables and functions for calc_load */
2211 static atomic_long_t calc_load_tasks;
2212 static unsigned long calc_load_update;
2213 unsigned long avenrun[3];
2214 EXPORT_SYMBOL(avenrun); /* should be removed */
2217 * get_avenrun - get the load average array
2218 * @loads: pointer to dest load array
2219 * @offset: offset to add
2220 * @shift: shift count to shift the result left
2222 * These values are estimates at best, so no need for locking.
2224 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2226 loads[0] = (avenrun[0] + offset) << shift;
2227 loads[1] = (avenrun[1] + offset) << shift;
2228 loads[2] = (avenrun[2] + offset) << shift;
2231 static long calc_load_fold_active(struct rq *this_rq)
2233 long nr_active, delta = 0;
2235 nr_active = this_rq->nr_running;
2236 nr_active += (long) this_rq->nr_uninterruptible;
2238 if (nr_active != this_rq->calc_load_active) {
2239 delta = nr_active - this_rq->calc_load_active;
2240 this_rq->calc_load_active = nr_active;
2247 * a1 = a0 * e + a * (1 - e)
2249 static unsigned long
2250 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2253 load += active * (FIXED_1 - exp);
2254 load += 1UL << (FSHIFT - 1);
2255 return load >> FSHIFT;
2260 * Handle NO_HZ for the global load-average.
2262 * Since the above described distributed algorithm to compute the global
2263 * load-average relies on per-cpu sampling from the tick, it is affected by
2266 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2267 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2268 * when we read the global state.
2270 * Obviously reality has to ruin such a delightfully simple scheme:
2272 * - When we go NO_HZ idle during the window, we can negate our sample
2273 * contribution, causing under-accounting.
2275 * We avoid this by keeping two idle-delta counters and flipping them
2276 * when the window starts, thus separating old and new NO_HZ load.
2278 * The only trick is the slight shift in index flip for read vs write.
2282 * |-|-----------|-|-----------|-|-----------|-|
2283 * r:0 0 1 1 0 0 1 1 0
2284 * w:0 1 1 0 0 1 1 0 0
2286 * This ensures we'll fold the old idle contribution in this window while
2287 * accumlating the new one.
2289 * - When we wake up from NO_HZ idle during the window, we push up our
2290 * contribution, since we effectively move our sample point to a known
2293 * This is solved by pushing the window forward, and thus skipping the
2294 * sample, for this cpu (effectively using the idle-delta for this cpu which
2295 * was in effect at the time the window opened). This also solves the issue
2296 * of having to deal with a cpu having been in NOHZ idle for multiple
2297 * LOAD_FREQ intervals.
2299 * When making the ILB scale, we should try to pull this in as well.
2301 static atomic_long_t calc_load_idle[2];
2302 static int calc_load_idx;
2304 static inline int calc_load_write_idx(void)
2306 int idx = calc_load_idx;
2309 * See calc_global_nohz(), if we observe the new index, we also
2310 * need to observe the new update time.
2315 * If the folding window started, make sure we start writing in the
2318 if (!time_before(jiffies, calc_load_update))
2324 static inline int calc_load_read_idx(void)
2326 return calc_load_idx & 1;
2329 void calc_load_enter_idle(void)
2331 struct rq *this_rq = this_rq();
2335 * We're going into NOHZ mode, if there's any pending delta, fold it
2336 * into the pending idle delta.
2338 delta = calc_load_fold_active(this_rq);
2340 int idx = calc_load_write_idx();
2341 atomic_long_add(delta, &calc_load_idle[idx]);
2345 void calc_load_exit_idle(void)
2347 struct rq *this_rq = this_rq();
2350 * If we're still before the sample window, we're done.
2352 if (time_before(jiffies, this_rq->calc_load_update))
2356 * We woke inside or after the sample window, this means we're already
2357 * accounted through the nohz accounting, so skip the entire deal and
2358 * sync up for the next window.
2360 this_rq->calc_load_update = calc_load_update;
2361 if (time_before(jiffies, this_rq->calc_load_update + 10))
2362 this_rq->calc_load_update += LOAD_FREQ;
2365 static long calc_load_fold_idle(void)
2367 int idx = calc_load_read_idx();
2370 if (atomic_long_read(&calc_load_idle[idx]))
2371 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2377 * fixed_power_int - compute: x^n, in O(log n) time
2379 * @x: base of the power
2380 * @frac_bits: fractional bits of @x
2381 * @n: power to raise @x to.
2383 * By exploiting the relation between the definition of the natural power
2384 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2385 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2386 * (where: n_i \elem {0, 1}, the binary vector representing n),
2387 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2388 * of course trivially computable in O(log_2 n), the length of our binary
2391 static unsigned long
2392 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2394 unsigned long result = 1UL << frac_bits;
2399 result += 1UL << (frac_bits - 1);
2400 result >>= frac_bits;
2406 x += 1UL << (frac_bits - 1);
2414 * a1 = a0 * e + a * (1 - e)
2416 * a2 = a1 * e + a * (1 - e)
2417 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2418 * = a0 * e^2 + a * (1 - e) * (1 + e)
2420 * a3 = a2 * e + a * (1 - e)
2421 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2422 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2426 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2427 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2428 * = a0 * e^n + a * (1 - e^n)
2430 * [1] application of the geometric series:
2433 * S_n := \Sum x^i = -------------
2436 static unsigned long
2437 calc_load_n(unsigned long load, unsigned long exp,
2438 unsigned long active, unsigned int n)
2441 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2445 * NO_HZ can leave us missing all per-cpu ticks calling
2446 * calc_load_account_active(), but since an idle CPU folds its delta into
2447 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2448 * in the pending idle delta if our idle period crossed a load cycle boundary.
2450 * Once we've updated the global active value, we need to apply the exponential
2451 * weights adjusted to the number of cycles missed.
2453 static void calc_global_nohz(void)
2455 long delta, active, n;
2457 if (!time_before(jiffies, calc_load_update + 10)) {
2459 * Catch-up, fold however many we are behind still
2461 delta = jiffies - calc_load_update - 10;
2462 n = 1 + (delta / LOAD_FREQ);
2464 active = atomic_long_read(&calc_load_tasks);
2465 active = active > 0 ? active * FIXED_1 : 0;
2467 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2468 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2469 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2471 calc_load_update += n * LOAD_FREQ;
2475 * Flip the idle index...
2477 * Make sure we first write the new time then flip the index, so that
2478 * calc_load_write_idx() will see the new time when it reads the new
2479 * index, this avoids a double flip messing things up.
2484 #else /* !CONFIG_NO_HZ */
2486 static inline long calc_load_fold_idle(void) { return 0; }
2487 static inline void calc_global_nohz(void) { }
2489 #endif /* CONFIG_NO_HZ */
2492 * calc_load - update the avenrun load estimates 10 ticks after the
2493 * CPUs have updated calc_load_tasks.
2495 void calc_global_load(unsigned long ticks)
2499 if (time_before(jiffies, calc_load_update + 10))
2503 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2505 delta = calc_load_fold_idle();
2507 atomic_long_add(delta, &calc_load_tasks);
2509 active = atomic_long_read(&calc_load_tasks);
2510 active = active > 0 ? active * FIXED_1 : 0;
2512 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2513 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2514 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2516 calc_load_update += LOAD_FREQ;
2519 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2525 * Called from update_cpu_load() to periodically update this CPU's
2528 static void calc_load_account_active(struct rq *this_rq)
2532 if (time_before(jiffies, this_rq->calc_load_update))
2535 delta = calc_load_fold_active(this_rq);
2537 atomic_long_add(delta, &calc_load_tasks);
2539 this_rq->calc_load_update += LOAD_FREQ;
2543 * End of global load-average stuff
2547 * The exact cpuload at various idx values, calculated at every tick would be
2548 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2550 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2551 * on nth tick when cpu may be busy, then we have:
2552 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2553 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2555 * decay_load_missed() below does efficient calculation of
2556 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2557 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2559 * The calculation is approximated on a 128 point scale.
2560 * degrade_zero_ticks is the number of ticks after which load at any
2561 * particular idx is approximated to be zero.
2562 * degrade_factor is a precomputed table, a row for each load idx.
2563 * Each column corresponds to degradation factor for a power of two ticks,
2564 * based on 128 point scale.
2566 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2567 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2569 * With this power of 2 load factors, we can degrade the load n times
2570 * by looking at 1 bits in n and doing as many mult/shift instead of
2571 * n mult/shifts needed by the exact degradation.
2573 #define DEGRADE_SHIFT 7
2574 static const unsigned char
2575 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2576 static const unsigned char
2577 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2578 {0, 0, 0, 0, 0, 0, 0, 0},
2579 {64, 32, 8, 0, 0, 0, 0, 0},
2580 {96, 72, 40, 12, 1, 0, 0},
2581 {112, 98, 75, 43, 15, 1, 0},
2582 {120, 112, 98, 76, 45, 16, 2} };
2585 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2586 * would be when CPU is idle and so we just decay the old load without
2587 * adding any new load.
2589 static unsigned long
2590 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2594 if (!missed_updates)
2597 if (missed_updates >= degrade_zero_ticks[idx])
2601 return load >> missed_updates;
2603 while (missed_updates) {
2604 if (missed_updates % 2)
2605 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2607 missed_updates >>= 1;
2614 * Update rq->cpu_load[] statistics. This function is usually called every
2615 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2616 * every tick. We fix it up based on jiffies.
2618 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2619 unsigned long pending_updates)
2623 this_rq->nr_load_updates++;
2625 /* Update our load: */
2626 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2627 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2628 unsigned long old_load, new_load;
2630 /* scale is effectively 1 << i now, and >> i divides by scale */
2632 old_load = this_rq->cpu_load[i];
2633 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2634 new_load = this_load;
2636 * Round up the averaging division if load is increasing. This
2637 * prevents us from getting stuck on 9 if the load is 10, for
2640 if (new_load > old_load)
2641 new_load += scale - 1;
2643 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2646 sched_avg_update(this_rq);
2651 * There is no sane way to deal with nohz on smp when using jiffies because the
2652 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2653 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2655 * Therefore we cannot use the delta approach from the regular tick since that
2656 * would seriously skew the load calculation. However we'll make do for those
2657 * updates happening while idle (nohz_idle_balance) or coming out of idle
2658 * (tick_nohz_idle_exit).
2660 * This means we might still be one tick off for nohz periods.
2664 * Called from nohz_idle_balance() to update the load ratings before doing the
2667 void update_idle_cpu_load(struct rq *this_rq)
2669 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2670 unsigned long load = this_rq->load.weight;
2671 unsigned long pending_updates;
2674 * bail if there's load or we're actually up-to-date.
2676 if (load || curr_jiffies == this_rq->last_load_update_tick)
2679 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2680 this_rq->last_load_update_tick = curr_jiffies;
2682 __update_cpu_load(this_rq, load, pending_updates);
2686 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2688 void update_cpu_load_nohz(void)
2690 struct rq *this_rq = this_rq();
2691 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2692 unsigned long pending_updates;
2694 if (curr_jiffies == this_rq->last_load_update_tick)
2697 raw_spin_lock(&this_rq->lock);
2698 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2699 if (pending_updates) {
2700 this_rq->last_load_update_tick = curr_jiffies;
2702 * We were idle, this means load 0, the current load might be
2703 * !0 due to remote wakeups and the sort.
2705 __update_cpu_load(this_rq, 0, pending_updates);
2707 raw_spin_unlock(&this_rq->lock);
2709 #endif /* CONFIG_NO_HZ */
2712 * Called from scheduler_tick()
2714 static void update_cpu_load_active(struct rq *this_rq)
2717 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2719 this_rq->last_load_update_tick = jiffies;
2720 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2722 calc_load_account_active(this_rq);
2728 * sched_exec - execve() is a valuable balancing opportunity, because at
2729 * this point the task has the smallest effective memory and cache footprint.
2731 void sched_exec(void)
2733 struct task_struct *p = current;
2734 unsigned long flags;
2737 raw_spin_lock_irqsave(&p->pi_lock, flags);
2738 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2739 if (dest_cpu == smp_processor_id())
2742 if (likely(cpu_active(dest_cpu))) {
2743 struct migration_arg arg = { p, dest_cpu };
2745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2746 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2750 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2755 DEFINE_PER_CPU(struct kernel_stat, kstat);
2756 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2758 EXPORT_PER_CPU_SYMBOL(kstat);
2759 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2762 * Return any ns on the sched_clock that have not yet been accounted in
2763 * @p in case that task is currently running.
2765 * Called with task_rq_lock() held on @rq.
2767 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2771 if (task_current(rq, p)) {
2772 update_rq_clock(rq);
2773 ns = rq->clock_task - p->se.exec_start;
2781 unsigned long long task_delta_exec(struct task_struct *p)
2783 unsigned long flags;
2787 rq = task_rq_lock(p, &flags);
2788 ns = do_task_delta_exec(p, rq);
2789 task_rq_unlock(rq, p, &flags);
2795 * Return accounted runtime for the task.
2796 * In case the task is currently running, return the runtime plus current's
2797 * pending runtime that have not been accounted yet.
2799 unsigned long long task_sched_runtime(struct task_struct *p)
2801 unsigned long flags;
2805 rq = task_rq_lock(p, &flags);
2806 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2807 task_rq_unlock(rq, p, &flags);
2812 #ifdef CONFIG_CGROUP_CPUACCT
2813 struct cgroup_subsys cpuacct_subsys;
2814 struct cpuacct root_cpuacct;
2817 static inline void task_group_account_field(struct task_struct *p, int index,
2820 #ifdef CONFIG_CGROUP_CPUACCT
2821 struct kernel_cpustat *kcpustat;
2825 * Since all updates are sure to touch the root cgroup, we
2826 * get ourselves ahead and touch it first. If the root cgroup
2827 * is the only cgroup, then nothing else should be necessary.
2830 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp;
2832 #ifdef CONFIG_CGROUP_CPUACCT
2833 if (unlikely(!cpuacct_subsys.active))
2838 while (ca && (ca != &root_cpuacct)) {
2839 kcpustat = this_cpu_ptr(ca->cpustat);
2840 kcpustat->cpustat[index] += tmp;
2849 * Account user cpu time to a process.
2850 * @p: the process that the cpu time gets accounted to
2851 * @cputime: the cpu time spent in user space since the last update
2852 * @cputime_scaled: cputime scaled by cpu frequency
2854 void account_user_time(struct task_struct *p, cputime_t cputime,
2855 cputime_t cputime_scaled)
2859 /* Add user time to process. */
2860 p->utime += cputime;
2861 p->utimescaled += cputime_scaled;
2862 account_group_user_time(p, cputime);
2864 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER;
2866 /* Add user time to cpustat. */
2867 task_group_account_field(p, index, (__force u64) cputime);
2869 /* Account for user time used */
2870 acct_update_integrals(p);
2874 * Account guest cpu time to a process.
2875 * @p: the process that the cpu time gets accounted to
2876 * @cputime: the cpu time spent in virtual machine since the last update
2877 * @cputime_scaled: cputime scaled by cpu frequency
2879 static void account_guest_time(struct task_struct *p, cputime_t cputime,
2880 cputime_t cputime_scaled)
2882 u64 *cpustat = kcpustat_this_cpu->cpustat;
2884 /* Add guest time to process. */
2885 p->utime += cputime;
2886 p->utimescaled += cputime_scaled;
2887 account_group_user_time(p, cputime);
2888 p->gtime += cputime;
2890 /* Add guest time to cpustat. */
2891 if (TASK_NICE(p) > 0) {
2892 cpustat[CPUTIME_NICE] += (__force u64) cputime;
2893 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime;
2895 cpustat[CPUTIME_USER] += (__force u64) cputime;
2896 cpustat[CPUTIME_GUEST] += (__force u64) cputime;
2901 * Account system cpu time to a process and desired cpustat field
2902 * @p: the process that the cpu time gets accounted to
2903 * @cputime: the cpu time spent in kernel space since the last update
2904 * @cputime_scaled: cputime scaled by cpu frequency
2905 * @target_cputime64: pointer to cpustat field that has to be updated
2908 void __account_system_time(struct task_struct *p, cputime_t cputime,
2909 cputime_t cputime_scaled, int index)
2911 /* Add system time to process. */
2912 p->stime += cputime;
2913 p->stimescaled += cputime_scaled;
2914 account_group_system_time(p, cputime);
2916 /* Add system time to cpustat. */
2917 task_group_account_field(p, index, (__force u64) cputime);
2919 /* Account for system time used */
2920 acct_update_integrals(p);
2924 * Account system cpu time to a process.
2925 * @p: the process that the cpu time gets accounted to
2926 * @hardirq_offset: the offset to subtract from hardirq_count()
2927 * @cputime: the cpu time spent in kernel space since the last update
2928 * @cputime_scaled: cputime scaled by cpu frequency
2930 void account_system_time(struct task_struct *p, int hardirq_offset,
2931 cputime_t cputime, cputime_t cputime_scaled)
2935 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
2936 account_guest_time(p, cputime, cputime_scaled);
2940 if (hardirq_count() - hardirq_offset)
2941 index = CPUTIME_IRQ;
2942 else if (in_serving_softirq())
2943 index = CPUTIME_SOFTIRQ;
2945 index = CPUTIME_SYSTEM;
2947 __account_system_time(p, cputime, cputime_scaled, index);
2951 * Account for involuntary wait time.
2952 * @cputime: the cpu time spent in involuntary wait
2954 void account_steal_time(cputime_t cputime)
2956 u64 *cpustat = kcpustat_this_cpu->cpustat;
2958 cpustat[CPUTIME_STEAL] += (__force u64) cputime;
2962 * Account for idle time.
2963 * @cputime: the cpu time spent in idle wait
2965 void account_idle_time(cputime_t cputime)
2967 u64 *cpustat = kcpustat_this_cpu->cpustat;
2968 struct rq *rq = this_rq();
2970 if (atomic_read(&rq->nr_iowait) > 0)
2971 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime;
2973 cpustat[CPUTIME_IDLE] += (__force u64) cputime;
2976 static __always_inline bool steal_account_process_tick(void)
2978 #ifdef CONFIG_PARAVIRT
2979 if (static_key_false(¶virt_steal_enabled)) {
2982 steal = paravirt_steal_clock(smp_processor_id());
2983 steal -= this_rq()->prev_steal_time;
2985 st = steal_ticks(steal);
2986 this_rq()->prev_steal_time += st * TICK_NSEC;
2988 account_steal_time(st);
2995 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
2997 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2999 * Account a tick to a process and cpustat
3000 * @p: the process that the cpu time gets accounted to
3001 * @user_tick: is the tick from userspace
3002 * @rq: the pointer to rq
3004 * Tick demultiplexing follows the order
3005 * - pending hardirq update
3006 * - pending softirq update
3010 * - check for guest_time
3011 * - else account as system_time
3013 * Check for hardirq is done both for system and user time as there is
3014 * no timer going off while we are on hardirq and hence we may never get an
3015 * opportunity to update it solely in system time.
3016 * p->stime and friends are only updated on system time and not on irq
3017 * softirq as those do not count in task exec_runtime any more.
3019 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3022 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3023 u64 *cpustat = kcpustat_this_cpu->cpustat;
3025 if (steal_account_process_tick())
3028 if (irqtime_account_hi_update()) {
3029 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy;
3030 } else if (irqtime_account_si_update()) {
3031 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy;
3032 } else if (this_cpu_ksoftirqd() == p) {
3034 * ksoftirqd time do not get accounted in cpu_softirq_time.
3035 * So, we have to handle it separately here.
3036 * Also, p->stime needs to be updated for ksoftirqd.
3038 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3040 } else if (user_tick) {
3041 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3042 } else if (p == rq->idle) {
3043 account_idle_time(cputime_one_jiffy);
3044 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3045 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3047 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3052 static void irqtime_account_idle_ticks(int ticks)
3055 struct rq *rq = this_rq();
3057 for (i = 0; i < ticks; i++)
3058 irqtime_account_process_tick(current, 0, rq);
3060 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3061 static void irqtime_account_idle_ticks(int ticks) {}
3062 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3064 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3067 * Account a single tick of cpu time.
3068 * @p: the process that the cpu time gets accounted to
3069 * @user_tick: indicates if the tick is a user or a system tick
3071 void account_process_tick(struct task_struct *p, int user_tick)
3073 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3074 struct rq *rq = this_rq();
3076 if (sched_clock_irqtime) {
3077 irqtime_account_process_tick(p, user_tick, rq);
3081 if (steal_account_process_tick())
3085 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3086 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3087 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3090 account_idle_time(cputime_one_jiffy);
3094 * Account multiple ticks of steal time.
3095 * @p: the process from which the cpu time has been stolen
3096 * @ticks: number of stolen ticks
3098 void account_steal_ticks(unsigned long ticks)
3100 account_steal_time(jiffies_to_cputime(ticks));
3104 * Account multiple ticks of idle time.
3105 * @ticks: number of stolen ticks
3107 void account_idle_ticks(unsigned long ticks)
3110 if (sched_clock_irqtime) {
3111 irqtime_account_idle_ticks(ticks);
3115 account_idle_time(jiffies_to_cputime(ticks));
3121 * Use precise platform statistics if available:
3123 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3124 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3130 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3132 struct task_cputime cputime;
3134 thread_group_cputime(p, &cputime);
3136 *ut = cputime.utime;
3137 *st = cputime.stime;
3141 #ifndef nsecs_to_cputime
3142 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3145 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3147 cputime_t rtime, utime = p->utime, total = utime + p->stime;
3150 * Use CFS's precise accounting:
3152 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3155 u64 temp = (__force u64) rtime;
3157 temp *= (__force u64) utime;
3158 do_div(temp, (__force u32) total);
3159 utime = (__force cputime_t) temp;
3164 * Compare with previous values, to keep monotonicity:
3166 p->prev_utime = max(p->prev_utime, utime);
3167 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime);
3169 *ut = p->prev_utime;
3170 *st = p->prev_stime;
3174 * Must be called with siglock held.
3176 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3178 struct signal_struct *sig = p->signal;
3179 struct task_cputime cputime;
3180 cputime_t rtime, utime, total;
3182 thread_group_cputime(p, &cputime);
3184 total = cputime.utime + cputime.stime;
3185 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3188 u64 temp = (__force u64) rtime;
3190 temp *= (__force u64) cputime.utime;
3191 do_div(temp, (__force u32) total);
3192 utime = (__force cputime_t) temp;
3196 sig->prev_utime = max(sig->prev_utime, utime);
3197 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime);
3199 *ut = sig->prev_utime;
3200 *st = sig->prev_stime;
3205 * This function gets called by the timer code, with HZ frequency.
3206 * We call it with interrupts disabled.
3208 void scheduler_tick(void)
3210 int cpu = smp_processor_id();
3211 struct rq *rq = cpu_rq(cpu);
3212 struct task_struct *curr = rq->curr;
3216 raw_spin_lock(&rq->lock);
3217 update_rq_clock(rq);
3218 update_cpu_load_active(rq);
3219 curr->sched_class->task_tick(rq, curr, 0);
3220 raw_spin_unlock(&rq->lock);
3222 perf_event_task_tick();
3225 rq->idle_balance = idle_cpu(cpu);
3226 trigger_load_balance(rq, cpu);
3230 notrace unsigned long get_parent_ip(unsigned long addr)
3232 if (in_lock_functions(addr)) {
3233 addr = CALLER_ADDR2;
3234 if (in_lock_functions(addr))
3235 addr = CALLER_ADDR3;
3240 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3241 defined(CONFIG_PREEMPT_TRACER))
3243 void __kprobes add_preempt_count(int val)
3245 #ifdef CONFIG_DEBUG_PREEMPT
3249 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3252 preempt_count() += val;
3253 #ifdef CONFIG_DEBUG_PREEMPT
3255 * Spinlock count overflowing soon?
3257 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3260 if (preempt_count() == val)
3261 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3263 EXPORT_SYMBOL(add_preempt_count);
3265 void __kprobes sub_preempt_count(int val)
3267 #ifdef CONFIG_DEBUG_PREEMPT
3271 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3274 * Is the spinlock portion underflowing?
3276 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3277 !(preempt_count() & PREEMPT_MASK)))
3281 if (preempt_count() == val)
3282 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3283 preempt_count() -= val;
3285 EXPORT_SYMBOL(sub_preempt_count);
3290 * Print scheduling while atomic bug:
3292 static noinline void __schedule_bug(struct task_struct *prev)
3294 if (oops_in_progress)
3297 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3298 prev->comm, prev->pid, preempt_count());
3300 debug_show_held_locks(prev);
3302 if (irqs_disabled())
3303 print_irqtrace_events(prev);
3305 add_taint(TAINT_WARN);
3309 * Various schedule()-time debugging checks and statistics:
3311 static inline void schedule_debug(struct task_struct *prev)
3314 * Test if we are atomic. Since do_exit() needs to call into
3315 * schedule() atomically, we ignore that path for now.
3316 * Otherwise, whine if we are scheduling when we should not be.
3318 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3319 __schedule_bug(prev);
3322 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3324 schedstat_inc(this_rq(), sched_count);
3327 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3329 if (prev->on_rq || rq->skip_clock_update < 0)
3330 update_rq_clock(rq);
3331 prev->sched_class->put_prev_task(rq, prev);
3335 * Pick up the highest-prio task:
3337 static inline struct task_struct *
3338 pick_next_task(struct rq *rq)
3340 const struct sched_class *class;
3341 struct task_struct *p;
3344 * Optimization: we know that if all tasks are in
3345 * the fair class we can call that function directly:
3347 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
3348 p = fair_sched_class.pick_next_task(rq);
3353 for_each_class(class) {
3354 p = class->pick_next_task(rq);
3359 BUG(); /* the idle class will always have a runnable task */
3363 * __schedule() is the main scheduler function.
3365 static void __sched __schedule(void)
3367 struct task_struct *prev, *next;
3368 unsigned long *switch_count;
3374 cpu = smp_processor_id();
3376 rcu_note_context_switch(cpu);
3379 schedule_debug(prev);
3381 if (sched_feat(HRTICK))
3384 raw_spin_lock_irq(&rq->lock);
3386 switch_count = &prev->nivcsw;
3387 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3388 if (unlikely(signal_pending_state(prev->state, prev))) {
3389 prev->state = TASK_RUNNING;
3391 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3395 * If a worker went to sleep, notify and ask workqueue
3396 * whether it wants to wake up a task to maintain
3399 if (prev->flags & PF_WQ_WORKER) {
3400 struct task_struct *to_wakeup;
3402 to_wakeup = wq_worker_sleeping(prev, cpu);
3404 try_to_wake_up_local(to_wakeup);
3407 switch_count = &prev->nvcsw;
3410 pre_schedule(rq, prev);
3412 if (unlikely(!rq->nr_running))
3413 idle_balance(cpu, rq);
3415 put_prev_task(rq, prev);
3416 next = pick_next_task(rq);
3417 clear_tsk_need_resched(prev);
3418 rq->skip_clock_update = 0;
3420 if (likely(prev != next)) {
3425 context_switch(rq, prev, next); /* unlocks the rq */
3427 * The context switch have flipped the stack from under us
3428 * and restored the local variables which were saved when
3429 * this task called schedule() in the past. prev == current
3430 * is still correct, but it can be moved to another cpu/rq.
3432 cpu = smp_processor_id();
3435 raw_spin_unlock_irq(&rq->lock);
3439 sched_preempt_enable_no_resched();
3444 static inline void sched_submit_work(struct task_struct *tsk)
3446 if (!tsk->state || tsk_is_pi_blocked(tsk))
3449 * If we are going to sleep and we have plugged IO queued,
3450 * make sure to submit it to avoid deadlocks.
3452 if (blk_needs_flush_plug(tsk))
3453 blk_schedule_flush_plug(tsk);
3456 asmlinkage void __sched schedule(void)
3458 struct task_struct *tsk = current;
3460 sched_submit_work(tsk);
3463 EXPORT_SYMBOL(schedule);
3466 * schedule_preempt_disabled - called with preemption disabled
3468 * Returns with preemption disabled. Note: preempt_count must be 1
3470 void __sched schedule_preempt_disabled(void)
3472 sched_preempt_enable_no_resched();
3477 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3479 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3481 if (lock->owner != owner)
3485 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3486 * lock->owner still matches owner, if that fails, owner might
3487 * point to free()d memory, if it still matches, the rcu_read_lock()
3488 * ensures the memory stays valid.
3492 return owner->on_cpu;
3496 * Look out! "owner" is an entirely speculative pointer
3497 * access and not reliable.
3499 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3501 if (!sched_feat(OWNER_SPIN))
3505 while (owner_running(lock, owner)) {
3509 arch_mutex_cpu_relax();
3514 * We break out the loop above on need_resched() and when the
3515 * owner changed, which is a sign for heavy contention. Return
3516 * success only when lock->owner is NULL.
3518 return lock->owner == NULL;
3522 #ifdef CONFIG_PREEMPT
3524 * this is the entry point to schedule() from in-kernel preemption
3525 * off of preempt_enable. Kernel preemptions off return from interrupt
3526 * occur there and call schedule directly.
3528 asmlinkage void __sched notrace preempt_schedule(void)
3530 struct thread_info *ti = current_thread_info();
3533 * If there is a non-zero preempt_count or interrupts are disabled,
3534 * we do not want to preempt the current task. Just return..
3536 if (likely(ti->preempt_count || irqs_disabled()))
3540 add_preempt_count_notrace(PREEMPT_ACTIVE);
3542 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3545 * Check again in case we missed a preemption opportunity
3546 * between schedule and now.
3549 } while (need_resched());
3551 EXPORT_SYMBOL(preempt_schedule);
3554 * this is the entry point to schedule() from kernel preemption
3555 * off of irq context.
3556 * Note, that this is called and return with irqs disabled. This will
3557 * protect us against recursive calling from irq.
3559 asmlinkage void __sched preempt_schedule_irq(void)
3561 struct thread_info *ti = current_thread_info();
3563 /* Catch callers which need to be fixed */
3564 BUG_ON(ti->preempt_count || !irqs_disabled());
3567 add_preempt_count(PREEMPT_ACTIVE);
3570 local_irq_disable();
3571 sub_preempt_count(PREEMPT_ACTIVE);
3574 * Check again in case we missed a preemption opportunity
3575 * between schedule and now.
3578 } while (need_resched());
3581 #endif /* CONFIG_PREEMPT */
3583 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3586 return try_to_wake_up(curr->private, mode, wake_flags);
3588 EXPORT_SYMBOL(default_wake_function);
3591 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3592 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3593 * number) then we wake all the non-exclusive tasks and one exclusive task.
3595 * There are circumstances in which we can try to wake a task which has already
3596 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3597 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3599 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3600 int nr_exclusive, int wake_flags, void *key)
3602 wait_queue_t *curr, *next;
3604 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3605 unsigned flags = curr->flags;
3607 if (curr->func(curr, mode, wake_flags, key) &&
3608 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3614 * __wake_up - wake up threads blocked on a waitqueue.
3616 * @mode: which threads
3617 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3618 * @key: is directly passed to the wakeup function
3620 * It may be assumed that this function implies a write memory barrier before
3621 * changing the task state if and only if any tasks are woken up.
3623 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3624 int nr_exclusive, void *key)
3626 unsigned long flags;
3628 spin_lock_irqsave(&q->lock, flags);
3629 __wake_up_common(q, mode, nr_exclusive, 0, key);
3630 spin_unlock_irqrestore(&q->lock, flags);
3632 EXPORT_SYMBOL(__wake_up);
3635 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3637 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3639 __wake_up_common(q, mode, nr, 0, NULL);
3641 EXPORT_SYMBOL_GPL(__wake_up_locked);
3643 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3645 __wake_up_common(q, mode, 1, 0, key);
3647 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3650 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3652 * @mode: which threads
3653 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3654 * @key: opaque value to be passed to wakeup targets
3656 * The sync wakeup differs that the waker knows that it will schedule
3657 * away soon, so while the target thread will be woken up, it will not
3658 * be migrated to another CPU - ie. the two threads are 'synchronized'
3659 * with each other. This can prevent needless bouncing between CPUs.
3661 * On UP it can prevent extra preemption.
3663 * It may be assumed that this function implies a write memory barrier before
3664 * changing the task state if and only if any tasks are woken up.
3666 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3667 int nr_exclusive, void *key)
3669 unsigned long flags;
3670 int wake_flags = WF_SYNC;
3675 if (unlikely(!nr_exclusive))
3678 spin_lock_irqsave(&q->lock, flags);
3679 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3680 spin_unlock_irqrestore(&q->lock, flags);
3682 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3685 * __wake_up_sync - see __wake_up_sync_key()
3687 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3689 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3691 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3694 * complete: - signals a single thread waiting on this completion
3695 * @x: holds the state of this particular completion
3697 * This will wake up a single thread waiting on this completion. Threads will be
3698 * awakened in the same order in which they were queued.
3700 * See also complete_all(), wait_for_completion() and related routines.
3702 * It may be assumed that this function implies a write memory barrier before
3703 * changing the task state if and only if any tasks are woken up.
3705 void complete(struct completion *x)
3707 unsigned long flags;
3709 spin_lock_irqsave(&x->wait.lock, flags);
3711 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3712 spin_unlock_irqrestore(&x->wait.lock, flags);
3714 EXPORT_SYMBOL(complete);
3717 * complete_all: - signals all threads waiting on this completion
3718 * @x: holds the state of this particular completion
3720 * This will wake up all threads waiting on this particular completion event.
3722 * It may be assumed that this function implies a write memory barrier before
3723 * changing the task state if and only if any tasks are woken up.
3725 void complete_all(struct completion *x)
3727 unsigned long flags;
3729 spin_lock_irqsave(&x->wait.lock, flags);
3730 x->done += UINT_MAX/2;
3731 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3732 spin_unlock_irqrestore(&x->wait.lock, flags);
3734 EXPORT_SYMBOL(complete_all);
3736 static inline long __sched
3737 do_wait_for_common(struct completion *x, long timeout, int state)
3740 DECLARE_WAITQUEUE(wait, current);
3742 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3744 if (signal_pending_state(state, current)) {
3745 timeout = -ERESTARTSYS;
3748 __set_current_state(state);
3749 spin_unlock_irq(&x->wait.lock);
3750 timeout = schedule_timeout(timeout);
3751 spin_lock_irq(&x->wait.lock);
3752 } while (!x->done && timeout);
3753 __remove_wait_queue(&x->wait, &wait);
3758 return timeout ?: 1;
3762 wait_for_common(struct completion *x, long timeout, int state)
3766 spin_lock_irq(&x->wait.lock);
3767 timeout = do_wait_for_common(x, timeout, state);
3768 spin_unlock_irq(&x->wait.lock);
3773 * wait_for_completion: - waits for completion of a task
3774 * @x: holds the state of this particular completion
3776 * This waits to be signaled for completion of a specific task. It is NOT
3777 * interruptible and there is no timeout.
3779 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3780 * and interrupt capability. Also see complete().
3782 void __sched wait_for_completion(struct completion *x)
3784 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3786 EXPORT_SYMBOL(wait_for_completion);
3789 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3790 * @x: holds the state of this particular completion
3791 * @timeout: timeout value in jiffies
3793 * This waits for either a completion of a specific task to be signaled or for a
3794 * specified timeout to expire. The timeout is in jiffies. It is not
3797 * The return value is 0 if timed out, and positive (at least 1, or number of
3798 * jiffies left till timeout) if completed.
3800 unsigned long __sched
3801 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3803 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3805 EXPORT_SYMBOL(wait_for_completion_timeout);
3808 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3809 * @x: holds the state of this particular completion
3811 * This waits for completion of a specific task to be signaled. It is
3814 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3816 int __sched wait_for_completion_interruptible(struct completion *x)
3818 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3819 if (t == -ERESTARTSYS)
3823 EXPORT_SYMBOL(wait_for_completion_interruptible);
3826 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3827 * @x: holds the state of this particular completion
3828 * @timeout: timeout value in jiffies
3830 * This waits for either a completion of a specific task to be signaled or for a
3831 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3833 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3834 * positive (at least 1, or number of jiffies left till timeout) if completed.
3837 wait_for_completion_interruptible_timeout(struct completion *x,
3838 unsigned long timeout)
3840 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3842 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3845 * wait_for_completion_killable: - waits for completion of a task (killable)
3846 * @x: holds the state of this particular completion
3848 * This waits to be signaled for completion of a specific task. It can be
3849 * interrupted by a kill signal.
3851 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3853 int __sched wait_for_completion_killable(struct completion *x)
3855 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3856 if (t == -ERESTARTSYS)
3860 EXPORT_SYMBOL(wait_for_completion_killable);
3863 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3864 * @x: holds the state of this particular completion
3865 * @timeout: timeout value in jiffies
3867 * This waits for either a completion of a specific task to be
3868 * signaled or for a specified timeout to expire. It can be
3869 * interrupted by a kill signal. The timeout is in jiffies.
3871 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3872 * positive (at least 1, or number of jiffies left till timeout) if completed.
3875 wait_for_completion_killable_timeout(struct completion *x,
3876 unsigned long timeout)
3878 return wait_for_common(x, timeout, TASK_KILLABLE);
3880 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3883 * try_wait_for_completion - try to decrement a completion without blocking
3884 * @x: completion structure
3886 * Returns: 0 if a decrement cannot be done without blocking
3887 * 1 if a decrement succeeded.
3889 * If a completion is being used as a counting completion,
3890 * attempt to decrement the counter without blocking. This
3891 * enables us to avoid waiting if the resource the completion
3892 * is protecting is not available.
3894 bool try_wait_for_completion(struct completion *x)
3896 unsigned long flags;
3899 spin_lock_irqsave(&x->wait.lock, flags);
3904 spin_unlock_irqrestore(&x->wait.lock, flags);
3907 EXPORT_SYMBOL(try_wait_for_completion);
3910 * completion_done - Test to see if a completion has any waiters
3911 * @x: completion structure
3913 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3914 * 1 if there are no waiters.
3917 bool completion_done(struct completion *x)
3919 unsigned long flags;
3922 spin_lock_irqsave(&x->wait.lock, flags);
3925 spin_unlock_irqrestore(&x->wait.lock, flags);
3928 EXPORT_SYMBOL(completion_done);
3931 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3933 unsigned long flags;
3936 init_waitqueue_entry(&wait, current);
3938 __set_current_state(state);
3940 spin_lock_irqsave(&q->lock, flags);
3941 __add_wait_queue(q, &wait);
3942 spin_unlock(&q->lock);
3943 timeout = schedule_timeout(timeout);
3944 spin_lock_irq(&q->lock);
3945 __remove_wait_queue(q, &wait);
3946 spin_unlock_irqrestore(&q->lock, flags);
3951 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3953 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3955 EXPORT_SYMBOL(interruptible_sleep_on);
3958 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3960 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3962 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3964 void __sched sleep_on(wait_queue_head_t *q)
3966 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3968 EXPORT_SYMBOL(sleep_on);
3970 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3972 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3974 EXPORT_SYMBOL(sleep_on_timeout);
3976 #ifdef CONFIG_RT_MUTEXES
3979 * rt_mutex_setprio - set the current priority of a task
3981 * @prio: prio value (kernel-internal form)
3983 * This function changes the 'effective' priority of a task. It does
3984 * not touch ->normal_prio like __setscheduler().
3986 * Used by the rt_mutex code to implement priority inheritance logic.
3988 void rt_mutex_setprio(struct task_struct *p, int prio)
3990 int oldprio, on_rq, running;
3992 const struct sched_class *prev_class;
3994 BUG_ON(prio < 0 || prio > MAX_PRIO);
3996 rq = __task_rq_lock(p);
3999 * Idle task boosting is a nono in general. There is one
4000 * exception, when PREEMPT_RT and NOHZ is active:
4002 * The idle task calls get_next_timer_interrupt() and holds
4003 * the timer wheel base->lock on the CPU and another CPU wants
4004 * to access the timer (probably to cancel it). We can safely
4005 * ignore the boosting request, as the idle CPU runs this code
4006 * with interrupts disabled and will complete the lock
4007 * protected section without being interrupted. So there is no
4008 * real need to boost.
4010 if (unlikely(p == rq->idle)) {
4011 WARN_ON(p != rq->curr);
4012 WARN_ON(p->pi_blocked_on);
4016 trace_sched_pi_setprio(p, prio);
4018 prev_class = p->sched_class;
4020 running = task_current(rq, p);
4022 dequeue_task(rq, p, 0);
4024 p->sched_class->put_prev_task(rq, p);
4027 p->sched_class = &rt_sched_class;
4029 p->sched_class = &fair_sched_class;
4034 p->sched_class->set_curr_task(rq);
4036 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4038 check_class_changed(rq, p, prev_class, oldprio);
4040 __task_rq_unlock(rq);
4043 void set_user_nice(struct task_struct *p, long nice)
4045 int old_prio, delta, on_rq;
4046 unsigned long flags;
4049 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4052 * We have to be careful, if called from sys_setpriority(),
4053 * the task might be in the middle of scheduling on another CPU.
4055 rq = task_rq_lock(p, &flags);
4057 * The RT priorities are set via sched_setscheduler(), but we still
4058 * allow the 'normal' nice value to be set - but as expected
4059 * it wont have any effect on scheduling until the task is
4060 * SCHED_FIFO/SCHED_RR:
4062 if (task_has_rt_policy(p)) {
4063 p->static_prio = NICE_TO_PRIO(nice);
4068 dequeue_task(rq, p, 0);
4070 p->static_prio = NICE_TO_PRIO(nice);
4073 p->prio = effective_prio(p);
4074 delta = p->prio - old_prio;
4077 enqueue_task(rq, p, 0);
4079 * If the task increased its priority or is running and
4080 * lowered its priority, then reschedule its CPU:
4082 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4083 resched_task(rq->curr);
4086 task_rq_unlock(rq, p, &flags);
4088 EXPORT_SYMBOL(set_user_nice);
4091 * can_nice - check if a task can reduce its nice value
4095 int can_nice(const struct task_struct *p, const int nice)
4097 /* convert nice value [19,-20] to rlimit style value [1,40] */
4098 int nice_rlim = 20 - nice;
4100 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4101 capable(CAP_SYS_NICE));
4104 #ifdef __ARCH_WANT_SYS_NICE
4107 * sys_nice - change the priority of the current process.
4108 * @increment: priority increment
4110 * sys_setpriority is a more generic, but much slower function that
4111 * does similar things.
4113 SYSCALL_DEFINE1(nice, int, increment)
4118 * Setpriority might change our priority at the same moment.
4119 * We don't have to worry. Conceptually one call occurs first
4120 * and we have a single winner.
4122 if (increment < -40)
4127 nice = TASK_NICE(current) + increment;
4133 if (increment < 0 && !can_nice(current, nice))
4136 retval = security_task_setnice(current, nice);
4140 set_user_nice(current, nice);
4147 * task_prio - return the priority value of a given task.
4148 * @p: the task in question.
4150 * This is the priority value as seen by users in /proc.
4151 * RT tasks are offset by -200. Normal tasks are centered
4152 * around 0, value goes from -16 to +15.
4154 int task_prio(const struct task_struct *p)
4156 return p->prio - MAX_RT_PRIO;
4160 * task_nice - return the nice value of a given task.
4161 * @p: the task in question.
4163 int task_nice(const struct task_struct *p)
4165 return TASK_NICE(p);
4167 EXPORT_SYMBOL(task_nice);
4170 * idle_cpu - is a given cpu idle currently?
4171 * @cpu: the processor in question.
4173 int idle_cpu(int cpu)
4175 struct rq *rq = cpu_rq(cpu);
4177 if (rq->curr != rq->idle)
4184 if (!llist_empty(&rq->wake_list))
4192 * idle_task - return the idle task for a given cpu.
4193 * @cpu: the processor in question.
4195 struct task_struct *idle_task(int cpu)
4197 return cpu_rq(cpu)->idle;
4201 * find_process_by_pid - find a process with a matching PID value.
4202 * @pid: the pid in question.
4204 static struct task_struct *find_process_by_pid(pid_t pid)
4206 return pid ? find_task_by_vpid(pid) : current;
4209 /* Actually do priority change: must hold rq lock. */
4211 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4214 p->rt_priority = prio;
4215 p->normal_prio = normal_prio(p);
4216 /* we are holding p->pi_lock already */
4217 p->prio = rt_mutex_getprio(p);
4218 if (rt_prio(p->prio))
4219 p->sched_class = &rt_sched_class;
4221 p->sched_class = &fair_sched_class;
4226 * check the target process has a UID that matches the current process's
4228 static bool check_same_owner(struct task_struct *p)
4230 const struct cred *cred = current_cred(), *pcred;
4234 pcred = __task_cred(p);
4235 match = (uid_eq(cred->euid, pcred->euid) ||
4236 uid_eq(cred->euid, pcred->uid));
4241 static int __sched_setscheduler(struct task_struct *p, int policy,
4242 const struct sched_param *param, bool user)
4244 int retval, oldprio, oldpolicy = -1, on_rq, running;
4245 unsigned long flags;
4246 const struct sched_class *prev_class;
4250 /* may grab non-irq protected spin_locks */
4251 BUG_ON(in_interrupt());
4253 /* double check policy once rq lock held */
4255 reset_on_fork = p->sched_reset_on_fork;
4256 policy = oldpolicy = p->policy;
4258 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4259 policy &= ~SCHED_RESET_ON_FORK;
4261 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4262 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4263 policy != SCHED_IDLE)
4268 * Valid priorities for SCHED_FIFO and SCHED_RR are
4269 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4270 * SCHED_BATCH and SCHED_IDLE is 0.
4272 if (param->sched_priority < 0 ||
4273 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4274 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4276 if (rt_policy(policy) != (param->sched_priority != 0))
4280 * Allow unprivileged RT tasks to decrease priority:
4282 if (user && !capable(CAP_SYS_NICE)) {
4283 if (rt_policy(policy)) {
4284 unsigned long rlim_rtprio =
4285 task_rlimit(p, RLIMIT_RTPRIO);
4287 /* can't set/change the rt policy */
4288 if (policy != p->policy && !rlim_rtprio)
4291 /* can't increase priority */
4292 if (param->sched_priority > p->rt_priority &&
4293 param->sched_priority > rlim_rtprio)
4298 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4299 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4301 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4302 if (!can_nice(p, TASK_NICE(p)))
4306 /* can't change other user's priorities */
4307 if (!check_same_owner(p))
4310 /* Normal users shall not reset the sched_reset_on_fork flag */
4311 if (p->sched_reset_on_fork && !reset_on_fork)
4316 retval = security_task_setscheduler(p);
4322 * make sure no PI-waiters arrive (or leave) while we are
4323 * changing the priority of the task:
4325 * To be able to change p->policy safely, the appropriate
4326 * runqueue lock must be held.
4328 rq = task_rq_lock(p, &flags);
4331 * Changing the policy of the stop threads its a very bad idea
4333 if (p == rq->stop) {
4334 task_rq_unlock(rq, p, &flags);
4339 * If not changing anything there's no need to proceed further:
4341 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
4342 param->sched_priority == p->rt_priority))) {
4343 task_rq_unlock(rq, p, &flags);
4347 #ifdef CONFIG_RT_GROUP_SCHED
4350 * Do not allow realtime tasks into groups that have no runtime
4353 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4354 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4355 !task_group_is_autogroup(task_group(p))) {
4356 task_rq_unlock(rq, p, &flags);
4362 /* recheck policy now with rq lock held */
4363 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4364 policy = oldpolicy = -1;
4365 task_rq_unlock(rq, p, &flags);
4369 running = task_current(rq, p);
4371 dequeue_task(rq, p, 0);
4373 p->sched_class->put_prev_task(rq, p);
4375 p->sched_reset_on_fork = reset_on_fork;
4378 prev_class = p->sched_class;
4379 __setscheduler(rq, p, policy, param->sched_priority);
4382 p->sched_class->set_curr_task(rq);
4384 enqueue_task(rq, p, 0);
4386 check_class_changed(rq, p, prev_class, oldprio);
4387 task_rq_unlock(rq, p, &flags);
4389 rt_mutex_adjust_pi(p);
4395 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4396 * @p: the task in question.
4397 * @policy: new policy.
4398 * @param: structure containing the new RT priority.
4400 * NOTE that the task may be already dead.
4402 int sched_setscheduler(struct task_struct *p, int policy,
4403 const struct sched_param *param)
4405 return __sched_setscheduler(p, policy, param, true);
4407 EXPORT_SYMBOL_GPL(sched_setscheduler);
4410 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4411 * @p: the task in question.
4412 * @policy: new policy.
4413 * @param: structure containing the new RT priority.
4415 * Just like sched_setscheduler, only don't bother checking if the
4416 * current context has permission. For example, this is needed in
4417 * stop_machine(): we create temporary high priority worker threads,
4418 * but our caller might not have that capability.
4420 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4421 const struct sched_param *param)
4423 return __sched_setscheduler(p, policy, param, false);
4427 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4429 struct sched_param lparam;
4430 struct task_struct *p;
4433 if (!param || pid < 0)
4435 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4440 p = find_process_by_pid(pid);
4442 retval = sched_setscheduler(p, policy, &lparam);
4449 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4450 * @pid: the pid in question.
4451 * @policy: new policy.
4452 * @param: structure containing the new RT priority.
4454 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4455 struct sched_param __user *, param)
4457 /* negative values for policy are not valid */
4461 return do_sched_setscheduler(pid, policy, param);
4465 * sys_sched_setparam - set/change the RT priority of a thread
4466 * @pid: the pid in question.
4467 * @param: structure containing the new RT priority.
4469 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4471 return do_sched_setscheduler(pid, -1, param);
4475 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4476 * @pid: the pid in question.
4478 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4480 struct task_struct *p;
4488 p = find_process_by_pid(pid);
4490 retval = security_task_getscheduler(p);
4493 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4500 * sys_sched_getparam - get the RT priority of a thread
4501 * @pid: the pid in question.
4502 * @param: structure containing the RT priority.
4504 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4506 struct sched_param lp;
4507 struct task_struct *p;
4510 if (!param || pid < 0)
4514 p = find_process_by_pid(pid);
4519 retval = security_task_getscheduler(p);
4523 lp.sched_priority = p->rt_priority;
4527 * This one might sleep, we cannot do it with a spinlock held ...
4529 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4538 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4540 cpumask_var_t cpus_allowed, new_mask;
4541 struct task_struct *p;
4547 p = find_process_by_pid(pid);
4554 /* Prevent p going away */
4558 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4562 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4564 goto out_free_cpus_allowed;
4567 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4570 retval = security_task_setscheduler(p);
4574 cpuset_cpus_allowed(p, cpus_allowed);
4575 cpumask_and(new_mask, in_mask, cpus_allowed);
4577 retval = set_cpus_allowed_ptr(p, new_mask);
4580 cpuset_cpus_allowed(p, cpus_allowed);
4581 if (!cpumask_subset(new_mask, cpus_allowed)) {
4583 * We must have raced with a concurrent cpuset
4584 * update. Just reset the cpus_allowed to the
4585 * cpuset's cpus_allowed
4587 cpumask_copy(new_mask, cpus_allowed);
4592 free_cpumask_var(new_mask);
4593 out_free_cpus_allowed:
4594 free_cpumask_var(cpus_allowed);
4601 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4602 struct cpumask *new_mask)
4604 if (len < cpumask_size())
4605 cpumask_clear(new_mask);
4606 else if (len > cpumask_size())
4607 len = cpumask_size();
4609 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4613 * sys_sched_setaffinity - set the cpu affinity of a process
4614 * @pid: pid of the process
4615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4616 * @user_mask_ptr: user-space pointer to the new cpu mask
4618 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4619 unsigned long __user *, user_mask_ptr)
4621 cpumask_var_t new_mask;
4624 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4627 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4629 retval = sched_setaffinity(pid, new_mask);
4630 free_cpumask_var(new_mask);
4634 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4636 struct task_struct *p;
4637 unsigned long flags;
4644 p = find_process_by_pid(pid);
4648 retval = security_task_getscheduler(p);
4652 raw_spin_lock_irqsave(&p->pi_lock, flags);
4653 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4654 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4664 * sys_sched_getaffinity - get the cpu affinity of a process
4665 * @pid: pid of the process
4666 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4667 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4669 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4670 unsigned long __user *, user_mask_ptr)
4675 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4677 if (len & (sizeof(unsigned long)-1))
4680 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4683 ret = sched_getaffinity(pid, mask);
4685 size_t retlen = min_t(size_t, len, cpumask_size());
4687 if (copy_to_user(user_mask_ptr, mask, retlen))
4692 free_cpumask_var(mask);
4698 * sys_sched_yield - yield the current processor to other threads.
4700 * This function yields the current CPU to other tasks. If there are no
4701 * other threads running on this CPU then this function will return.
4703 SYSCALL_DEFINE0(sched_yield)
4705 struct rq *rq = this_rq_lock();
4707 schedstat_inc(rq, yld_count);
4708 current->sched_class->yield_task(rq);
4711 * Since we are going to call schedule() anyway, there's
4712 * no need to preempt or enable interrupts:
4714 __release(rq->lock);
4715 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4716 do_raw_spin_unlock(&rq->lock);
4717 sched_preempt_enable_no_resched();
4724 static inline int should_resched(void)
4726 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4729 static void __cond_resched(void)
4731 add_preempt_count(PREEMPT_ACTIVE);
4733 sub_preempt_count(PREEMPT_ACTIVE);
4736 int __sched _cond_resched(void)
4738 if (should_resched()) {
4744 EXPORT_SYMBOL(_cond_resched);
4747 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4748 * call schedule, and on return reacquire the lock.
4750 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4751 * operations here to prevent schedule() from being called twice (once via
4752 * spin_unlock(), once by hand).
4754 int __cond_resched_lock(spinlock_t *lock)
4756 int resched = should_resched();
4759 lockdep_assert_held(lock);
4761 if (spin_needbreak(lock) || resched) {
4772 EXPORT_SYMBOL(__cond_resched_lock);
4774 int __sched __cond_resched_softirq(void)
4776 BUG_ON(!in_softirq());
4778 if (should_resched()) {
4786 EXPORT_SYMBOL(__cond_resched_softirq);
4789 * yield - yield the current processor to other threads.
4791 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4793 * The scheduler is at all times free to pick the calling task as the most
4794 * eligible task to run, if removing the yield() call from your code breaks
4795 * it, its already broken.
4797 * Typical broken usage is:
4802 * where one assumes that yield() will let 'the other' process run that will
4803 * make event true. If the current task is a SCHED_FIFO task that will never
4804 * happen. Never use yield() as a progress guarantee!!
4806 * If you want to use yield() to wait for something, use wait_event().
4807 * If you want to use yield() to be 'nice' for others, use cond_resched().
4808 * If you still want to use yield(), do not!
4810 void __sched yield(void)
4812 set_current_state(TASK_RUNNING);
4815 EXPORT_SYMBOL(yield);
4818 * yield_to - yield the current processor to another thread in
4819 * your thread group, or accelerate that thread toward the
4820 * processor it's on.
4822 * @preempt: whether task preemption is allowed or not
4824 * It's the caller's job to ensure that the target task struct
4825 * can't go away on us before we can do any checks.
4827 * Returns true if we indeed boosted the target task.
4829 bool __sched yield_to(struct task_struct *p, bool preempt)
4831 struct task_struct *curr = current;
4832 struct rq *rq, *p_rq;
4833 unsigned long flags;
4836 local_irq_save(flags);
4841 double_rq_lock(rq, p_rq);
4842 while (task_rq(p) != p_rq) {
4843 double_rq_unlock(rq, p_rq);
4847 if (!curr->sched_class->yield_to_task)
4850 if (curr->sched_class != p->sched_class)
4853 if (task_running(p_rq, p) || p->state)
4856 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4858 schedstat_inc(rq, yld_count);
4860 * Make p's CPU reschedule; pick_next_entity takes care of
4863 if (preempt && rq != p_rq)
4864 resched_task(p_rq->curr);
4867 * We might have set it in task_yield_fair(), but are
4868 * not going to schedule(), so don't want to skip
4871 rq->skip_clock_update = 0;
4875 double_rq_unlock(rq, p_rq);
4876 local_irq_restore(flags);
4883 EXPORT_SYMBOL_GPL(yield_to);
4886 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4887 * that process accounting knows that this is a task in IO wait state.
4889 void __sched io_schedule(void)
4891 struct rq *rq = raw_rq();
4893 delayacct_blkio_start();
4894 atomic_inc(&rq->nr_iowait);
4895 blk_flush_plug(current);
4896 current->in_iowait = 1;
4898 current->in_iowait = 0;
4899 atomic_dec(&rq->nr_iowait);
4900 delayacct_blkio_end();
4902 EXPORT_SYMBOL(io_schedule);
4904 long __sched io_schedule_timeout(long timeout)
4906 struct rq *rq = raw_rq();
4909 delayacct_blkio_start();
4910 atomic_inc(&rq->nr_iowait);
4911 blk_flush_plug(current);
4912 current->in_iowait = 1;
4913 ret = schedule_timeout(timeout);
4914 current->in_iowait = 0;
4915 atomic_dec(&rq->nr_iowait);
4916 delayacct_blkio_end();
4921 * sys_sched_get_priority_max - return maximum RT priority.
4922 * @policy: scheduling class.
4924 * this syscall returns the maximum rt_priority that can be used
4925 * by a given scheduling class.
4927 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4934 ret = MAX_USER_RT_PRIO-1;
4946 * sys_sched_get_priority_min - return minimum RT priority.
4947 * @policy: scheduling class.
4949 * this syscall returns the minimum rt_priority that can be used
4950 * by a given scheduling class.
4952 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4970 * sys_sched_rr_get_interval - return the default timeslice of a process.
4971 * @pid: pid of the process.
4972 * @interval: userspace pointer to the timeslice value.
4974 * this syscall writes the default timeslice value of a given process
4975 * into the user-space timespec buffer. A value of '0' means infinity.
4977 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4978 struct timespec __user *, interval)
4980 struct task_struct *p;
4981 unsigned int time_slice;
4982 unsigned long flags;
4992 p = find_process_by_pid(pid);
4996 retval = security_task_getscheduler(p);
5000 rq = task_rq_lock(p, &flags);
5001 time_slice = p->sched_class->get_rr_interval(rq, p);
5002 task_rq_unlock(rq, p, &flags);
5005 jiffies_to_timespec(time_slice, &t);
5006 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5014 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5016 void sched_show_task(struct task_struct *p)
5018 unsigned long free = 0;
5021 state = p->state ? __ffs(p->state) + 1 : 0;
5022 printk(KERN_INFO "%-15.15s %c", p->comm,
5023 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5024 #if BITS_PER_LONG == 32
5025 if (state == TASK_RUNNING)
5026 printk(KERN_CONT " running ");
5028 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5030 if (state == TASK_RUNNING)
5031 printk(KERN_CONT " running task ");
5033 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5035 #ifdef CONFIG_DEBUG_STACK_USAGE
5036 free = stack_not_used(p);
5038 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5039 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
5040 (unsigned long)task_thread_info(p)->flags);
5042 show_stack(p, NULL);
5045 void show_state_filter(unsigned long state_filter)
5047 struct task_struct *g, *p;
5049 #if BITS_PER_LONG == 32
5051 " task PC stack pid father\n");
5054 " task PC stack pid father\n");
5057 do_each_thread(g, p) {
5059 * reset the NMI-timeout, listing all files on a slow
5060 * console might take a lot of time:
5062 touch_nmi_watchdog();
5063 if (!state_filter || (p->state & state_filter))
5065 } while_each_thread(g, p);
5067 touch_all_softlockup_watchdogs();
5069 #ifdef CONFIG_SCHED_DEBUG
5070 sysrq_sched_debug_show();
5074 * Only show locks if all tasks are dumped:
5077 debug_show_all_locks();
5080 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5082 idle->sched_class = &idle_sched_class;
5086 * init_idle - set up an idle thread for a given CPU
5087 * @idle: task in question
5088 * @cpu: cpu the idle task belongs to
5090 * NOTE: this function does not set the idle thread's NEED_RESCHED
5091 * flag, to make booting more robust.
5093 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5095 struct rq *rq = cpu_rq(cpu);
5096 unsigned long flags;
5098 raw_spin_lock_irqsave(&rq->lock, flags);
5101 idle->state = TASK_RUNNING;
5102 idle->se.exec_start = sched_clock();
5104 do_set_cpus_allowed(idle, cpumask_of(cpu));
5106 * We're having a chicken and egg problem, even though we are
5107 * holding rq->lock, the cpu isn't yet set to this cpu so the
5108 * lockdep check in task_group() will fail.
5110 * Similar case to sched_fork(). / Alternatively we could
5111 * use task_rq_lock() here and obtain the other rq->lock.
5116 __set_task_cpu(idle, cpu);
5119 rq->curr = rq->idle = idle;
5120 #if defined(CONFIG_SMP)
5123 raw_spin_unlock_irqrestore(&rq->lock, flags);
5125 /* Set the preempt count _outside_ the spinlocks! */
5126 task_thread_info(idle)->preempt_count = 0;
5129 * The idle tasks have their own, simple scheduling class:
5131 idle->sched_class = &idle_sched_class;
5132 ftrace_graph_init_idle_task(idle, cpu);
5133 #if defined(CONFIG_SMP)
5134 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5139 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
5141 if (p->sched_class && p->sched_class->set_cpus_allowed)
5142 p->sched_class->set_cpus_allowed(p, new_mask);
5144 cpumask_copy(&p->cpus_allowed, new_mask);
5145 p->nr_cpus_allowed = cpumask_weight(new_mask);
5149 * This is how migration works:
5151 * 1) we invoke migration_cpu_stop() on the target CPU using
5153 * 2) stopper starts to run (implicitly forcing the migrated thread
5155 * 3) it checks whether the migrated task is still in the wrong runqueue.
5156 * 4) if it's in the wrong runqueue then the migration thread removes
5157 * it and puts it into the right queue.
5158 * 5) stopper completes and stop_one_cpu() returns and the migration
5163 * Change a given task's CPU affinity. Migrate the thread to a
5164 * proper CPU and schedule it away if the CPU it's executing on
5165 * is removed from the allowed bitmask.
5167 * NOTE: the caller must have a valid reference to the task, the
5168 * task must not exit() & deallocate itself prematurely. The
5169 * call is not atomic; no spinlocks may be held.
5171 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5173 unsigned long flags;
5175 unsigned int dest_cpu;
5178 rq = task_rq_lock(p, &flags);
5180 if (cpumask_equal(&p->cpus_allowed, new_mask))
5183 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5188 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
5193 do_set_cpus_allowed(p, new_mask);
5195 /* Can the task run on the task's current CPU? If so, we're done */
5196 if (cpumask_test_cpu(task_cpu(p), new_mask))
5199 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5201 struct migration_arg arg = { p, dest_cpu };
5202 /* Need help from migration thread: drop lock and wait. */
5203 task_rq_unlock(rq, p, &flags);
5204 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5205 tlb_migrate_finish(p->mm);
5209 task_rq_unlock(rq, p, &flags);
5213 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5216 * Move (not current) task off this cpu, onto dest cpu. We're doing
5217 * this because either it can't run here any more (set_cpus_allowed()
5218 * away from this CPU, or CPU going down), or because we're
5219 * attempting to rebalance this task on exec (sched_exec).
5221 * So we race with normal scheduler movements, but that's OK, as long
5222 * as the task is no longer on this CPU.
5224 * Returns non-zero if task was successfully migrated.
5226 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5228 struct rq *rq_dest, *rq_src;
5231 if (unlikely(!cpu_active(dest_cpu)))
5234 rq_src = cpu_rq(src_cpu);
5235 rq_dest = cpu_rq(dest_cpu);
5237 raw_spin_lock(&p->pi_lock);
5238 double_rq_lock(rq_src, rq_dest);
5239 /* Already moved. */
5240 if (task_cpu(p) != src_cpu)
5242 /* Affinity changed (again). */
5243 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
5247 * If we're not on a rq, the next wake-up will ensure we're
5251 dequeue_task(rq_src, p, 0);
5252 set_task_cpu(p, dest_cpu);
5253 enqueue_task(rq_dest, p, 0);
5254 check_preempt_curr(rq_dest, p, 0);
5259 double_rq_unlock(rq_src, rq_dest);
5260 raw_spin_unlock(&p->pi_lock);
5265 * migration_cpu_stop - this will be executed by a highprio stopper thread
5266 * and performs thread migration by bumping thread off CPU then
5267 * 'pushing' onto another runqueue.
5269 static int migration_cpu_stop(void *data)
5271 struct migration_arg *arg = data;
5274 * The original target cpu might have gone down and we might
5275 * be on another cpu but it doesn't matter.
5277 local_irq_disable();
5278 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5283 #ifdef CONFIG_HOTPLUG_CPU
5286 * Ensures that the idle task is using init_mm right before its cpu goes
5289 void idle_task_exit(void)
5291 struct mm_struct *mm = current->active_mm;
5293 BUG_ON(cpu_online(smp_processor_id()));
5296 switch_mm(mm, &init_mm, current);
5301 * While a dead CPU has no uninterruptible tasks queued at this point,
5302 * it might still have a nonzero ->nr_uninterruptible counter, because
5303 * for performance reasons the counter is not stricly tracking tasks to
5304 * their home CPUs. So we just add the counter to another CPU's counter,
5305 * to keep the global sum constant after CPU-down:
5307 static void migrate_nr_uninterruptible(struct rq *rq_src)
5309 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5311 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5312 rq_src->nr_uninterruptible = 0;
5316 * remove the tasks which were accounted by rq from calc_load_tasks.
5318 static void calc_global_load_remove(struct rq *rq)
5320 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5321 rq->calc_load_active = 0;
5325 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5326 * try_to_wake_up()->select_task_rq().
5328 * Called with rq->lock held even though we'er in stop_machine() and
5329 * there's no concurrency possible, we hold the required locks anyway
5330 * because of lock validation efforts.
5332 static void migrate_tasks(unsigned int dead_cpu)
5334 struct rq *rq = cpu_rq(dead_cpu);
5335 struct task_struct *next, *stop = rq->stop;
5339 * Fudge the rq selection such that the below task selection loop
5340 * doesn't get stuck on the currently eligible stop task.
5342 * We're currently inside stop_machine() and the rq is either stuck
5343 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5344 * either way we should never end up calling schedule() until we're
5349 /* Ensure any throttled groups are reachable by pick_next_task */
5350 unthrottle_offline_cfs_rqs(rq);
5354 * There's this thread running, bail when that's the only
5357 if (rq->nr_running == 1)
5360 next = pick_next_task(rq);
5362 next->sched_class->put_prev_task(rq, next);
5364 /* Find suitable destination for @next, with force if needed. */
5365 dest_cpu = select_fallback_rq(dead_cpu, next);
5366 raw_spin_unlock(&rq->lock);
5368 __migrate_task(next, dead_cpu, dest_cpu);
5370 raw_spin_lock(&rq->lock);
5376 #endif /* CONFIG_HOTPLUG_CPU */
5378 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5380 static struct ctl_table sd_ctl_dir[] = {
5382 .procname = "sched_domain",
5388 static struct ctl_table sd_ctl_root[] = {
5390 .procname = "kernel",
5392 .child = sd_ctl_dir,
5397 static struct ctl_table *sd_alloc_ctl_entry(int n)
5399 struct ctl_table *entry =
5400 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5405 static void sd_free_ctl_entry(struct ctl_table **tablep)
5407 struct ctl_table *entry;
5410 * In the intermediate directories, both the child directory and
5411 * procname are dynamically allocated and could fail but the mode
5412 * will always be set. In the lowest directory the names are
5413 * static strings and all have proc handlers.
5415 for (entry = *tablep; entry->mode; entry++) {
5417 sd_free_ctl_entry(&entry->child);
5418 if (entry->proc_handler == NULL)
5419 kfree(entry->procname);
5427 set_table_entry(struct ctl_table *entry,
5428 const char *procname, void *data, int maxlen,
5429 umode_t mode, proc_handler *proc_handler)
5431 entry->procname = procname;
5433 entry->maxlen = maxlen;
5435 entry->proc_handler = proc_handler;
5438 static struct ctl_table *
5439 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5441 struct ctl_table *table = sd_alloc_ctl_entry(13);
5446 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5447 sizeof(long), 0644, proc_doulongvec_minmax);
5448 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5449 sizeof(long), 0644, proc_doulongvec_minmax);
5450 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5451 sizeof(int), 0644, proc_dointvec_minmax);
5452 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5453 sizeof(int), 0644, proc_dointvec_minmax);
5454 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5455 sizeof(int), 0644, proc_dointvec_minmax);
5456 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5457 sizeof(int), 0644, proc_dointvec_minmax);
5458 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5459 sizeof(int), 0644, proc_dointvec_minmax);
5460 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5461 sizeof(int), 0644, proc_dointvec_minmax);
5462 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5463 sizeof(int), 0644, proc_dointvec_minmax);
5464 set_table_entry(&table[9], "cache_nice_tries",
5465 &sd->cache_nice_tries,
5466 sizeof(int), 0644, proc_dointvec_minmax);
5467 set_table_entry(&table[10], "flags", &sd->flags,
5468 sizeof(int), 0644, proc_dointvec_minmax);
5469 set_table_entry(&table[11], "name", sd->name,
5470 CORENAME_MAX_SIZE, 0444, proc_dostring);
5471 /* &table[12] is terminator */
5476 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5478 struct ctl_table *entry, *table;
5479 struct sched_domain *sd;
5480 int domain_num = 0, i;
5483 for_each_domain(cpu, sd)
5485 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5490 for_each_domain(cpu, sd) {
5491 snprintf(buf, 32, "domain%d", i);
5492 entry->procname = kstrdup(buf, GFP_KERNEL);
5494 entry->child = sd_alloc_ctl_domain_table(sd);
5501 static struct ctl_table_header *sd_sysctl_header;
5502 static void register_sched_domain_sysctl(void)
5504 int i, cpu_num = num_possible_cpus();
5505 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5508 WARN_ON(sd_ctl_dir[0].child);
5509 sd_ctl_dir[0].child = entry;
5514 for_each_possible_cpu(i) {
5515 snprintf(buf, 32, "cpu%d", i);
5516 entry->procname = kstrdup(buf, GFP_KERNEL);
5518 entry->child = sd_alloc_ctl_cpu_table(i);
5522 WARN_ON(sd_sysctl_header);
5523 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5526 /* may be called multiple times per register */
5527 static void unregister_sched_domain_sysctl(void)
5529 if (sd_sysctl_header)
5530 unregister_sysctl_table(sd_sysctl_header);
5531 sd_sysctl_header = NULL;
5532 if (sd_ctl_dir[0].child)
5533 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5536 static void register_sched_domain_sysctl(void)
5539 static void unregister_sched_domain_sysctl(void)
5544 static void set_rq_online(struct rq *rq)
5547 const struct sched_class *class;
5549 cpumask_set_cpu(rq->cpu, rq->rd->online);
5552 for_each_class(class) {
5553 if (class->rq_online)
5554 class->rq_online(rq);
5559 static void set_rq_offline(struct rq *rq)
5562 const struct sched_class *class;
5564 for_each_class(class) {
5565 if (class->rq_offline)
5566 class->rq_offline(rq);
5569 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5575 * migration_call - callback that gets triggered when a CPU is added.
5576 * Here we can start up the necessary migration thread for the new CPU.
5578 static int __cpuinit
5579 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5581 int cpu = (long)hcpu;
5582 unsigned long flags;
5583 struct rq *rq = cpu_rq(cpu);
5585 switch (action & ~CPU_TASKS_FROZEN) {
5587 case CPU_UP_PREPARE:
5588 rq->calc_load_update = calc_load_update;
5592 /* Update our root-domain */
5593 raw_spin_lock_irqsave(&rq->lock, flags);
5595 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5599 raw_spin_unlock_irqrestore(&rq->lock, flags);
5602 #ifdef CONFIG_HOTPLUG_CPU
5604 sched_ttwu_pending();
5605 /* Update our root-domain */
5606 raw_spin_lock_irqsave(&rq->lock, flags);
5608 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5612 BUG_ON(rq->nr_running != 1); /* the migration thread */
5613 raw_spin_unlock_irqrestore(&rq->lock, flags);
5615 migrate_nr_uninterruptible(rq);
5616 calc_global_load_remove(rq);
5621 update_max_interval();
5627 * Register at high priority so that task migration (migrate_all_tasks)
5628 * happens before everything else. This has to be lower priority than
5629 * the notifier in the perf_event subsystem, though.
5631 static struct notifier_block __cpuinitdata migration_notifier = {
5632 .notifier_call = migration_call,
5633 .priority = CPU_PRI_MIGRATION,
5636 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5637 unsigned long action, void *hcpu)
5639 switch (action & ~CPU_TASKS_FROZEN) {
5641 case CPU_DOWN_FAILED:
5642 set_cpu_active((long)hcpu, true);
5649 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5650 unsigned long action, void *hcpu)
5652 switch (action & ~CPU_TASKS_FROZEN) {
5653 case CPU_DOWN_PREPARE:
5654 set_cpu_active((long)hcpu, false);
5661 static int __init migration_init(void)
5663 void *cpu = (void *)(long)smp_processor_id();
5666 /* Initialize migration for the boot CPU */
5667 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5668 BUG_ON(err == NOTIFY_BAD);
5669 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5670 register_cpu_notifier(&migration_notifier);
5672 /* Register cpu active notifiers */
5673 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5674 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5678 early_initcall(migration_init);
5683 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5685 #ifdef CONFIG_SCHED_DEBUG
5687 static __read_mostly int sched_debug_enabled;
5689 static int __init sched_debug_setup(char *str)
5691 sched_debug_enabled = 1;
5695 early_param("sched_debug", sched_debug_setup);
5697 static inline bool sched_debug(void)
5699 return sched_debug_enabled;
5702 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5703 struct cpumask *groupmask)
5705 struct sched_group *group = sd->groups;
5708 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5709 cpumask_clear(groupmask);
5711 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5713 if (!(sd->flags & SD_LOAD_BALANCE)) {
5714 printk("does not load-balance\n");
5716 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5721 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5723 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5724 printk(KERN_ERR "ERROR: domain->span does not contain "
5727 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5728 printk(KERN_ERR "ERROR: domain->groups does not contain"
5732 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5736 printk(KERN_ERR "ERROR: group is NULL\n");
5741 * Even though we initialize ->power to something semi-sane,
5742 * we leave power_orig unset. This allows us to detect if
5743 * domain iteration is still funny without causing /0 traps.
5745 if (!group->sgp->power_orig) {
5746 printk(KERN_CONT "\n");
5747 printk(KERN_ERR "ERROR: domain->cpu_power not "
5752 if (!cpumask_weight(sched_group_cpus(group))) {
5753 printk(KERN_CONT "\n");
5754 printk(KERN_ERR "ERROR: empty group\n");
5758 if (!(sd->flags & SD_OVERLAP) &&
5759 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5760 printk(KERN_CONT "\n");
5761 printk(KERN_ERR "ERROR: repeated CPUs\n");
5765 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5767 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5769 printk(KERN_CONT " %s", str);
5770 if (group->sgp->power != SCHED_POWER_SCALE) {
5771 printk(KERN_CONT " (cpu_power = %d)",
5775 group = group->next;
5776 } while (group != sd->groups);
5777 printk(KERN_CONT "\n");
5779 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5780 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5783 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5784 printk(KERN_ERR "ERROR: parent span is not a superset "
5785 "of domain->span\n");
5789 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5793 if (!sched_debug_enabled)
5797 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5801 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5804 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5812 #else /* !CONFIG_SCHED_DEBUG */
5813 # define sched_domain_debug(sd, cpu) do { } while (0)
5814 static inline bool sched_debug(void)
5818 #endif /* CONFIG_SCHED_DEBUG */
5820 static int sd_degenerate(struct sched_domain *sd)
5822 if (cpumask_weight(sched_domain_span(sd)) == 1)
5825 /* Following flags need at least 2 groups */
5826 if (sd->flags & (SD_LOAD_BALANCE |
5827 SD_BALANCE_NEWIDLE |
5831 SD_SHARE_PKG_RESOURCES)) {
5832 if (sd->groups != sd->groups->next)
5836 /* Following flags don't use groups */
5837 if (sd->flags & (SD_WAKE_AFFINE))
5844 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5846 unsigned long cflags = sd->flags, pflags = parent->flags;
5848 if (sd_degenerate(parent))
5851 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5854 /* Flags needing groups don't count if only 1 group in parent */
5855 if (parent->groups == parent->groups->next) {
5856 pflags &= ~(SD_LOAD_BALANCE |
5857 SD_BALANCE_NEWIDLE |
5861 SD_SHARE_PKG_RESOURCES);
5862 if (nr_node_ids == 1)
5863 pflags &= ~SD_SERIALIZE;
5865 if (~cflags & pflags)
5871 static void free_rootdomain(struct rcu_head *rcu)
5873 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5875 cpupri_cleanup(&rd->cpupri);
5876 free_cpumask_var(rd->rto_mask);
5877 free_cpumask_var(rd->online);
5878 free_cpumask_var(rd->span);
5882 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5884 struct root_domain *old_rd = NULL;
5885 unsigned long flags;
5887 raw_spin_lock_irqsave(&rq->lock, flags);
5892 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5895 cpumask_clear_cpu(rq->cpu, old_rd->span);
5898 * If we dont want to free the old_rt yet then
5899 * set old_rd to NULL to skip the freeing later
5902 if (!atomic_dec_and_test(&old_rd->refcount))
5906 atomic_inc(&rd->refcount);
5909 cpumask_set_cpu(rq->cpu, rd->span);
5910 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5913 raw_spin_unlock_irqrestore(&rq->lock, flags);
5916 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5919 static int init_rootdomain(struct root_domain *rd)
5921 memset(rd, 0, sizeof(*rd));
5923 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5925 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5927 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5930 if (cpupri_init(&rd->cpupri) != 0)
5935 free_cpumask_var(rd->rto_mask);
5937 free_cpumask_var(rd->online);
5939 free_cpumask_var(rd->span);
5945 * By default the system creates a single root-domain with all cpus as
5946 * members (mimicking the global state we have today).
5948 struct root_domain def_root_domain;
5950 static void init_defrootdomain(void)
5952 init_rootdomain(&def_root_domain);
5954 atomic_set(&def_root_domain.refcount, 1);
5957 static struct root_domain *alloc_rootdomain(void)
5959 struct root_domain *rd;
5961 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5965 if (init_rootdomain(rd) != 0) {
5973 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5975 struct sched_group *tmp, *first;
5984 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5989 } while (sg != first);
5992 static void free_sched_domain(struct rcu_head *rcu)
5994 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5997 * If its an overlapping domain it has private groups, iterate and
6000 if (sd->flags & SD_OVERLAP) {
6001 free_sched_groups(sd->groups, 1);
6002 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6003 kfree(sd->groups->sgp);
6009 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6011 call_rcu(&sd->rcu, free_sched_domain);
6014 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6016 for (; sd; sd = sd->parent)
6017 destroy_sched_domain(sd, cpu);
6021 * Keep a special pointer to the highest sched_domain that has
6022 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6023 * allows us to avoid some pointer chasing select_idle_sibling().
6025 * Iterate domains and sched_groups downward, assigning CPUs to be
6026 * select_idle_sibling() hw buddy. Cross-wiring hw makes bouncing
6027 * due to random perturbation self canceling, ie sw buddies pull
6028 * their counterpart to their CPU's hw counterpart.
6030 * Also keep a unique ID per domain (we use the first cpu number in
6031 * the cpumask of the domain), this allows us to quickly tell if
6032 * two cpus are in the same cache domain, see cpus_share_cache().
6034 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6035 DEFINE_PER_CPU(int, sd_llc_id);
6037 static void update_top_cache_domain(int cpu)
6039 struct sched_domain *sd;
6042 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6044 struct sched_domain *tmp = sd;
6045 struct sched_group *sg, *prev;
6049 * Traverse to first CPU in group, and count hops
6050 * to cpu from there, switching direction on each
6051 * hop, never ever pointing the last CPU rightward.
6054 id = cpumask_first(sched_domain_span(tmp));
6055 prev = sg = tmp->groups;
6058 while (cpumask_first(sched_group_cpus(sg)) != id)
6061 while (!cpumask_test_cpu(cpu, sched_group_cpus(sg))) {
6067 /* A CPU went down, never point back to domain start. */
6068 if (right && cpumask_first(sched_group_cpus(sg->next)) == id)
6071 sg = right ? sg->next : prev;
6072 tmp->idle_buddy = cpumask_first(sched_group_cpus(sg));
6073 } while ((tmp = tmp->child));
6075 id = cpumask_first(sched_domain_span(sd));
6078 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6079 per_cpu(sd_llc_id, cpu) = id;
6083 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6084 * hold the hotplug lock.
6087 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6089 struct rq *rq = cpu_rq(cpu);
6090 struct sched_domain *tmp;
6092 /* Remove the sched domains which do not contribute to scheduling. */
6093 for (tmp = sd; tmp; ) {
6094 struct sched_domain *parent = tmp->parent;
6098 if (sd_parent_degenerate(tmp, parent)) {
6099 tmp->parent = parent->parent;
6101 parent->parent->child = tmp;
6102 destroy_sched_domain(parent, cpu);
6107 if (sd && sd_degenerate(sd)) {
6110 destroy_sched_domain(tmp, cpu);
6115 sched_domain_debug(sd, cpu);
6117 rq_attach_root(rq, rd);
6119 rcu_assign_pointer(rq->sd, sd);
6120 destroy_sched_domains(tmp, cpu);
6122 update_top_cache_domain(cpu);
6125 /* cpus with isolated domains */
6126 static cpumask_var_t cpu_isolated_map;
6128 /* Setup the mask of cpus configured for isolated domains */
6129 static int __init isolated_cpu_setup(char *str)
6131 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6132 cpulist_parse(str, cpu_isolated_map);
6136 __setup("isolcpus=", isolated_cpu_setup);
6138 static const struct cpumask *cpu_cpu_mask(int cpu)
6140 return cpumask_of_node(cpu_to_node(cpu));
6144 struct sched_domain **__percpu sd;
6145 struct sched_group **__percpu sg;
6146 struct sched_group_power **__percpu sgp;
6150 struct sched_domain ** __percpu sd;
6151 struct root_domain *rd;
6161 struct sched_domain_topology_level;
6163 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
6164 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
6166 #define SDTL_OVERLAP 0x01
6168 struct sched_domain_topology_level {
6169 sched_domain_init_f init;
6170 sched_domain_mask_f mask;
6173 struct sd_data data;
6177 * Build an iteration mask that can exclude certain CPUs from the upwards
6180 * Asymmetric node setups can result in situations where the domain tree is of
6181 * unequal depth, make sure to skip domains that already cover the entire
6184 * In that case build_sched_domains() will have terminated the iteration early
6185 * and our sibling sd spans will be empty. Domains should always include the
6186 * cpu they're built on, so check that.
6189 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6191 const struct cpumask *span = sched_domain_span(sd);
6192 struct sd_data *sdd = sd->private;
6193 struct sched_domain *sibling;
6196 for_each_cpu(i, span) {
6197 sibling = *per_cpu_ptr(sdd->sd, i);
6198 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6201 cpumask_set_cpu(i, sched_group_mask(sg));
6206 * Return the canonical balance cpu for this group, this is the first cpu
6207 * of this group that's also in the iteration mask.
6209 int group_balance_cpu(struct sched_group *sg)
6211 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6215 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6217 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6218 const struct cpumask *span = sched_domain_span(sd);
6219 struct cpumask *covered = sched_domains_tmpmask;
6220 struct sd_data *sdd = sd->private;
6221 struct sched_domain *child;
6224 cpumask_clear(covered);
6226 for_each_cpu(i, span) {
6227 struct cpumask *sg_span;
6229 if (cpumask_test_cpu(i, covered))
6232 child = *per_cpu_ptr(sdd->sd, i);
6234 /* See the comment near build_group_mask(). */
6235 if (!cpumask_test_cpu(i, sched_domain_span(child)))
6238 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6239 GFP_KERNEL, cpu_to_node(cpu));
6244 sg_span = sched_group_cpus(sg);
6246 child = child->child;
6247 cpumask_copy(sg_span, sched_domain_span(child));
6249 cpumask_set_cpu(i, sg_span);
6251 cpumask_or(covered, covered, sg_span);
6253 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
6254 if (atomic_inc_return(&sg->sgp->ref) == 1)
6255 build_group_mask(sd, sg);
6258 * Initialize sgp->power such that even if we mess up the
6259 * domains and no possible iteration will get us here, we won't
6262 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
6265 * Make sure the first group of this domain contains the
6266 * canonical balance cpu. Otherwise the sched_domain iteration
6267 * breaks. See update_sg_lb_stats().
6269 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6270 group_balance_cpu(sg) == cpu)
6280 sd->groups = groups;
6285 free_sched_groups(first, 0);
6290 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6292 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6293 struct sched_domain *child = sd->child;
6296 cpu = cpumask_first(sched_domain_span(child));
6299 *sg = *per_cpu_ptr(sdd->sg, cpu);
6300 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
6301 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
6308 * build_sched_groups will build a circular linked list of the groups
6309 * covered by the given span, and will set each group's ->cpumask correctly,
6310 * and ->cpu_power to 0.
6312 * Assumes the sched_domain tree is fully constructed
6315 build_sched_groups(struct sched_domain *sd, int cpu)
6317 struct sched_group *first = NULL, *last = NULL;
6318 struct sd_data *sdd = sd->private;
6319 const struct cpumask *span = sched_domain_span(sd);
6320 struct cpumask *covered;
6323 get_group(cpu, sdd, &sd->groups);
6324 atomic_inc(&sd->groups->ref);
6326 if (cpu != cpumask_first(sched_domain_span(sd)))
6329 lockdep_assert_held(&sched_domains_mutex);
6330 covered = sched_domains_tmpmask;
6332 cpumask_clear(covered);
6334 for_each_cpu(i, span) {
6335 struct sched_group *sg;
6336 int group = get_group(i, sdd, &sg);
6339 if (cpumask_test_cpu(i, covered))
6342 cpumask_clear(sched_group_cpus(sg));
6344 cpumask_setall(sched_group_mask(sg));
6346 for_each_cpu(j, span) {
6347 if (get_group(j, sdd, NULL) != group)
6350 cpumask_set_cpu(j, covered);
6351 cpumask_set_cpu(j, sched_group_cpus(sg));
6366 * Initialize sched groups cpu_power.
6368 * cpu_power indicates the capacity of sched group, which is used while
6369 * distributing the load between different sched groups in a sched domain.
6370 * Typically cpu_power for all the groups in a sched domain will be same unless
6371 * there are asymmetries in the topology. If there are asymmetries, group
6372 * having more cpu_power will pickup more load compared to the group having
6375 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6377 struct sched_group *sg = sd->groups;
6379 WARN_ON(!sd || !sg);
6382 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6384 } while (sg != sd->groups);
6386 if (cpu != group_balance_cpu(sg))
6389 update_group_power(sd, cpu);
6390 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6393 int __weak arch_sd_sibling_asym_packing(void)
6395 return 0*SD_ASYM_PACKING;
6399 * Initializers for schedule domains
6400 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6403 #ifdef CONFIG_SCHED_DEBUG
6404 # define SD_INIT_NAME(sd, type) sd->name = #type
6406 # define SD_INIT_NAME(sd, type) do { } while (0)
6409 #define SD_INIT_FUNC(type) \
6410 static noinline struct sched_domain * \
6411 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6413 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6414 *sd = SD_##type##_INIT; \
6415 SD_INIT_NAME(sd, type); \
6416 sd->private = &tl->data; \
6421 #ifdef CONFIG_SCHED_SMT
6422 SD_INIT_FUNC(SIBLING)
6424 #ifdef CONFIG_SCHED_MC
6427 #ifdef CONFIG_SCHED_BOOK
6431 static int default_relax_domain_level = -1;
6432 int sched_domain_level_max;
6434 static int __init setup_relax_domain_level(char *str)
6436 if (kstrtoint(str, 0, &default_relax_domain_level))
6437 pr_warn("Unable to set relax_domain_level\n");
6441 __setup("relax_domain_level=", setup_relax_domain_level);
6443 static void set_domain_attribute(struct sched_domain *sd,
6444 struct sched_domain_attr *attr)
6448 if (!attr || attr->relax_domain_level < 0) {
6449 if (default_relax_domain_level < 0)
6452 request = default_relax_domain_level;
6454 request = attr->relax_domain_level;
6455 if (request < sd->level) {
6456 /* turn off idle balance on this domain */
6457 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6459 /* turn on idle balance on this domain */
6460 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6464 static void __sdt_free(const struct cpumask *cpu_map);
6465 static int __sdt_alloc(const struct cpumask *cpu_map);
6467 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6468 const struct cpumask *cpu_map)
6472 if (!atomic_read(&d->rd->refcount))
6473 free_rootdomain(&d->rd->rcu); /* fall through */
6475 free_percpu(d->sd); /* fall through */
6477 __sdt_free(cpu_map); /* fall through */
6483 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6484 const struct cpumask *cpu_map)
6486 memset(d, 0, sizeof(*d));
6488 if (__sdt_alloc(cpu_map))
6489 return sa_sd_storage;
6490 d->sd = alloc_percpu(struct sched_domain *);
6492 return sa_sd_storage;
6493 d->rd = alloc_rootdomain();
6496 return sa_rootdomain;
6500 * NULL the sd_data elements we've used to build the sched_domain and
6501 * sched_group structure so that the subsequent __free_domain_allocs()
6502 * will not free the data we're using.
6504 static void claim_allocations(int cpu, struct sched_domain *sd)
6506 struct sd_data *sdd = sd->private;
6508 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6509 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6511 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6512 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6514 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6515 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6518 #ifdef CONFIG_SCHED_SMT
6519 static const struct cpumask *cpu_smt_mask(int cpu)
6521 return topology_thread_cpumask(cpu);
6526 * Topology list, bottom-up.
6528 static struct sched_domain_topology_level default_topology[] = {
6529 #ifdef CONFIG_SCHED_SMT
6530 { sd_init_SIBLING, cpu_smt_mask, },
6532 #ifdef CONFIG_SCHED_MC
6533 { sd_init_MC, cpu_coregroup_mask, },
6535 #ifdef CONFIG_SCHED_BOOK
6536 { sd_init_BOOK, cpu_book_mask, },
6538 { sd_init_CPU, cpu_cpu_mask, },
6542 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6546 static int sched_domains_numa_levels;
6547 static int *sched_domains_numa_distance;
6548 static struct cpumask ***sched_domains_numa_masks;
6549 static int sched_domains_curr_level;
6551 static inline int sd_local_flags(int level)
6553 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6556 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6559 static struct sched_domain *
6560 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6562 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6563 int level = tl->numa_level;
6564 int sd_weight = cpumask_weight(
6565 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6567 *sd = (struct sched_domain){
6568 .min_interval = sd_weight,
6569 .max_interval = 2*sd_weight,
6571 .imbalance_pct = 125,
6572 .cache_nice_tries = 2,
6579 .flags = 1*SD_LOAD_BALANCE
6580 | 1*SD_BALANCE_NEWIDLE
6586 | 0*SD_SHARE_CPUPOWER
6587 | 0*SD_SHARE_PKG_RESOURCES
6589 | 0*SD_PREFER_SIBLING
6590 | sd_local_flags(level)
6592 .last_balance = jiffies,
6593 .balance_interval = sd_weight,
6595 SD_INIT_NAME(sd, NUMA);
6596 sd->private = &tl->data;
6599 * Ugly hack to pass state to sd_numa_mask()...
6601 sched_domains_curr_level = tl->numa_level;
6606 static const struct cpumask *sd_numa_mask(int cpu)
6608 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6611 static void sched_numa_warn(const char *str)
6613 static int done = false;
6621 printk(KERN_WARNING "ERROR: %s\n\n", str);
6623 for (i = 0; i < nr_node_ids; i++) {
6624 printk(KERN_WARNING " ");
6625 for (j = 0; j < nr_node_ids; j++)
6626 printk(KERN_CONT "%02d ", node_distance(i,j));
6627 printk(KERN_CONT "\n");
6629 printk(KERN_WARNING "\n");
6632 static bool find_numa_distance(int distance)
6636 if (distance == node_distance(0, 0))
6639 for (i = 0; i < sched_domains_numa_levels; i++) {
6640 if (sched_domains_numa_distance[i] == distance)
6647 static void sched_init_numa(void)
6649 int next_distance, curr_distance = node_distance(0, 0);
6650 struct sched_domain_topology_level *tl;
6654 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6655 if (!sched_domains_numa_distance)
6659 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6660 * unique distances in the node_distance() table.
6662 * Assumes node_distance(0,j) includes all distances in
6663 * node_distance(i,j) in order to avoid cubic time.
6665 next_distance = curr_distance;
6666 for (i = 0; i < nr_node_ids; i++) {
6667 for (j = 0; j < nr_node_ids; j++) {
6668 for (k = 0; k < nr_node_ids; k++) {
6669 int distance = node_distance(i, k);
6671 if (distance > curr_distance &&
6672 (distance < next_distance ||
6673 next_distance == curr_distance))
6674 next_distance = distance;
6677 * While not a strong assumption it would be nice to know
6678 * about cases where if node A is connected to B, B is not
6679 * equally connected to A.
6681 if (sched_debug() && node_distance(k, i) != distance)
6682 sched_numa_warn("Node-distance not symmetric");
6684 if (sched_debug() && i && !find_numa_distance(distance))
6685 sched_numa_warn("Node-0 not representative");
6687 if (next_distance != curr_distance) {
6688 sched_domains_numa_distance[level++] = next_distance;
6689 sched_domains_numa_levels = level;
6690 curr_distance = next_distance;
6695 * In case of sched_debug() we verify the above assumption.
6701 * 'level' contains the number of unique distances, excluding the
6702 * identity distance node_distance(i,i).
6704 * The sched_domains_nume_distance[] array includes the actual distance
6708 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6709 if (!sched_domains_numa_masks)
6713 * Now for each level, construct a mask per node which contains all
6714 * cpus of nodes that are that many hops away from us.
6716 for (i = 0; i < level; i++) {
6717 sched_domains_numa_masks[i] =
6718 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6719 if (!sched_domains_numa_masks[i])
6722 for (j = 0; j < nr_node_ids; j++) {
6723 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6727 sched_domains_numa_masks[i][j] = mask;
6729 for (k = 0; k < nr_node_ids; k++) {
6730 if (node_distance(j, k) > sched_domains_numa_distance[i])
6733 cpumask_or(mask, mask, cpumask_of_node(k));
6738 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6739 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6744 * Copy the default topology bits..
6746 for (i = 0; default_topology[i].init; i++)
6747 tl[i] = default_topology[i];
6750 * .. and append 'j' levels of NUMA goodness.
6752 for (j = 0; j < level; i++, j++) {
6753 tl[i] = (struct sched_domain_topology_level){
6754 .init = sd_numa_init,
6755 .mask = sd_numa_mask,
6756 .flags = SDTL_OVERLAP,
6761 sched_domain_topology = tl;
6764 static inline void sched_init_numa(void)
6767 #endif /* CONFIG_NUMA */
6769 static int __sdt_alloc(const struct cpumask *cpu_map)
6771 struct sched_domain_topology_level *tl;
6774 for (tl = sched_domain_topology; tl->init; tl++) {
6775 struct sd_data *sdd = &tl->data;
6777 sdd->sd = alloc_percpu(struct sched_domain *);
6781 sdd->sg = alloc_percpu(struct sched_group *);
6785 sdd->sgp = alloc_percpu(struct sched_group_power *);
6789 for_each_cpu(j, cpu_map) {
6790 struct sched_domain *sd;
6791 struct sched_group *sg;
6792 struct sched_group_power *sgp;
6794 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6795 GFP_KERNEL, cpu_to_node(j));
6799 *per_cpu_ptr(sdd->sd, j) = sd;
6801 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6802 GFP_KERNEL, cpu_to_node(j));
6808 *per_cpu_ptr(sdd->sg, j) = sg;
6810 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6811 GFP_KERNEL, cpu_to_node(j));
6815 *per_cpu_ptr(sdd->sgp, j) = sgp;
6822 static void __sdt_free(const struct cpumask *cpu_map)
6824 struct sched_domain_topology_level *tl;
6827 for (tl = sched_domain_topology; tl->init; tl++) {
6828 struct sd_data *sdd = &tl->data;
6830 for_each_cpu(j, cpu_map) {
6831 struct sched_domain *sd;
6834 sd = *per_cpu_ptr(sdd->sd, j);
6835 if (sd && (sd->flags & SD_OVERLAP))
6836 free_sched_groups(sd->groups, 0);
6837 kfree(*per_cpu_ptr(sdd->sd, j));
6841 kfree(*per_cpu_ptr(sdd->sg, j));
6843 kfree(*per_cpu_ptr(sdd->sgp, j));
6845 free_percpu(sdd->sd);
6847 free_percpu(sdd->sg);
6849 free_percpu(sdd->sgp);
6854 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6855 struct s_data *d, const struct cpumask *cpu_map,
6856 struct sched_domain_attr *attr, struct sched_domain *child,
6859 struct sched_domain *sd = tl->init(tl, cpu);
6863 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6865 sd->level = child->level + 1;
6866 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6870 set_domain_attribute(sd, attr);
6876 * Build sched domains for a given set of cpus and attach the sched domains
6877 * to the individual cpus
6879 static int build_sched_domains(const struct cpumask *cpu_map,
6880 struct sched_domain_attr *attr)
6882 enum s_alloc alloc_state = sa_none;
6883 struct sched_domain *sd;
6885 int i, ret = -ENOMEM;
6887 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6888 if (alloc_state != sa_rootdomain)
6891 /* Set up domains for cpus specified by the cpu_map. */
6892 for_each_cpu(i, cpu_map) {
6893 struct sched_domain_topology_level *tl;
6896 for (tl = sched_domain_topology; tl->init; tl++) {
6897 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6898 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6899 sd->flags |= SD_OVERLAP;
6900 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6907 *per_cpu_ptr(d.sd, i) = sd;
6910 /* Build the groups for the domains */
6911 for_each_cpu(i, cpu_map) {
6912 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6913 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6914 if (sd->flags & SD_OVERLAP) {
6915 if (build_overlap_sched_groups(sd, i))
6918 if (build_sched_groups(sd, i))
6924 /* Calculate CPU power for physical packages and nodes */
6925 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6926 if (!cpumask_test_cpu(i, cpu_map))
6929 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6930 claim_allocations(i, sd);
6931 init_sched_groups_power(i, sd);
6935 /* Attach the domains */
6937 for_each_cpu(i, cpu_map) {
6938 sd = *per_cpu_ptr(d.sd, i);
6939 cpu_attach_domain(sd, d.rd, i);
6945 __free_domain_allocs(&d, alloc_state, cpu_map);
6949 static cpumask_var_t *doms_cur; /* current sched domains */
6950 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6951 static struct sched_domain_attr *dattr_cur;
6952 /* attribues of custom domains in 'doms_cur' */
6955 * Special case: If a kmalloc of a doms_cur partition (array of
6956 * cpumask) fails, then fallback to a single sched domain,
6957 * as determined by the single cpumask fallback_doms.
6959 static cpumask_var_t fallback_doms;
6962 * arch_update_cpu_topology lets virtualized architectures update the
6963 * cpu core maps. It is supposed to return 1 if the topology changed
6964 * or 0 if it stayed the same.
6966 int __attribute__((weak)) arch_update_cpu_topology(void)
6971 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6974 cpumask_var_t *doms;
6976 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6979 for (i = 0; i < ndoms; i++) {
6980 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6981 free_sched_domains(doms, i);
6988 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6991 for (i = 0; i < ndoms; i++)
6992 free_cpumask_var(doms[i]);
6997 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6998 * For now this just excludes isolated cpus, but could be used to
6999 * exclude other special cases in the future.
7001 static int init_sched_domains(const struct cpumask *cpu_map)
7005 arch_update_cpu_topology();
7007 doms_cur = alloc_sched_domains(ndoms_cur);
7009 doms_cur = &fallback_doms;
7010 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7011 err = build_sched_domains(doms_cur[0], NULL);
7012 register_sched_domain_sysctl();
7018 * Detach sched domains from a group of cpus specified in cpu_map
7019 * These cpus will now be attached to the NULL domain
7021 static void detach_destroy_domains(const struct cpumask *cpu_map)
7026 for_each_cpu(i, cpu_map)
7027 cpu_attach_domain(NULL, &def_root_domain, i);
7031 /* handle null as "default" */
7032 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7033 struct sched_domain_attr *new, int idx_new)
7035 struct sched_domain_attr tmp;
7042 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7043 new ? (new + idx_new) : &tmp,
7044 sizeof(struct sched_domain_attr));
7048 * Partition sched domains as specified by the 'ndoms_new'
7049 * cpumasks in the array doms_new[] of cpumasks. This compares
7050 * doms_new[] to the current sched domain partitioning, doms_cur[].
7051 * It destroys each deleted domain and builds each new domain.
7053 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7054 * The masks don't intersect (don't overlap.) We should setup one
7055 * sched domain for each mask. CPUs not in any of the cpumasks will
7056 * not be load balanced. If the same cpumask appears both in the
7057 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7060 * The passed in 'doms_new' should be allocated using
7061 * alloc_sched_domains. This routine takes ownership of it and will
7062 * free_sched_domains it when done with it. If the caller failed the
7063 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7064 * and partition_sched_domains() will fallback to the single partition
7065 * 'fallback_doms', it also forces the domains to be rebuilt.
7067 * If doms_new == NULL it will be replaced with cpu_online_mask.
7068 * ndoms_new == 0 is a special case for destroying existing domains,
7069 * and it will not create the default domain.
7071 * Call with hotplug lock held
7073 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7074 struct sched_domain_attr *dattr_new)
7079 mutex_lock(&sched_domains_mutex);
7081 /* always unregister in case we don't destroy any domains */
7082 unregister_sched_domain_sysctl();
7084 /* Let architecture update cpu core mappings. */
7085 new_topology = arch_update_cpu_topology();
7087 n = doms_new ? ndoms_new : 0;
7089 /* Destroy deleted domains */
7090 for (i = 0; i < ndoms_cur; i++) {
7091 for (j = 0; j < n && !new_topology; j++) {
7092 if (cpumask_equal(doms_cur[i], doms_new[j])
7093 && dattrs_equal(dattr_cur, i, dattr_new, j))
7096 /* no match - a current sched domain not in new doms_new[] */
7097 detach_destroy_domains(doms_cur[i]);
7102 if (doms_new == NULL) {
7104 doms_new = &fallback_doms;
7105 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7106 WARN_ON_ONCE(dattr_new);
7109 /* Build new domains */
7110 for (i = 0; i < ndoms_new; i++) {
7111 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7112 if (cpumask_equal(doms_new[i], doms_cur[j])
7113 && dattrs_equal(dattr_new, i, dattr_cur, j))
7116 /* no match - add a new doms_new */
7117 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7122 /* Remember the new sched domains */
7123 if (doms_cur != &fallback_doms)
7124 free_sched_domains(doms_cur, ndoms_cur);
7125 kfree(dattr_cur); /* kfree(NULL) is safe */
7126 doms_cur = doms_new;
7127 dattr_cur = dattr_new;
7128 ndoms_cur = ndoms_new;
7130 register_sched_domain_sysctl();
7132 mutex_unlock(&sched_domains_mutex);
7135 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7138 * Update cpusets according to cpu_active mask. If cpusets are
7139 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7140 * around partition_sched_domains().
7142 * If we come here as part of a suspend/resume, don't touch cpusets because we
7143 * want to restore it back to its original state upon resume anyway.
7145 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7149 case CPU_ONLINE_FROZEN:
7150 case CPU_DOWN_FAILED_FROZEN:
7153 * num_cpus_frozen tracks how many CPUs are involved in suspend
7154 * resume sequence. As long as this is not the last online
7155 * operation in the resume sequence, just build a single sched
7156 * domain, ignoring cpusets.
7159 if (likely(num_cpus_frozen)) {
7160 partition_sched_domains(1, NULL, NULL);
7165 * This is the last CPU online operation. So fall through and
7166 * restore the original sched domains by considering the
7167 * cpuset configurations.
7171 case CPU_DOWN_FAILED:
7172 cpuset_update_active_cpus(true);
7180 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7184 case CPU_DOWN_PREPARE:
7185 cpuset_update_active_cpus(false);
7187 case CPU_DOWN_PREPARE_FROZEN:
7189 partition_sched_domains(1, NULL, NULL);
7197 void __init sched_init_smp(void)
7199 cpumask_var_t non_isolated_cpus;
7201 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7202 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7207 mutex_lock(&sched_domains_mutex);
7208 init_sched_domains(cpu_active_mask);
7209 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7210 if (cpumask_empty(non_isolated_cpus))
7211 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7212 mutex_unlock(&sched_domains_mutex);
7215 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7216 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7218 /* RT runtime code needs to handle some hotplug events */
7219 hotcpu_notifier(update_runtime, 0);
7223 /* Move init over to a non-isolated CPU */
7224 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7226 sched_init_granularity();
7227 free_cpumask_var(non_isolated_cpus);
7229 init_sched_rt_class();
7232 void __init sched_init_smp(void)
7234 sched_init_granularity();
7236 #endif /* CONFIG_SMP */
7238 const_debug unsigned int sysctl_timer_migration = 1;
7240 int in_sched_functions(unsigned long addr)
7242 return in_lock_functions(addr) ||
7243 (addr >= (unsigned long)__sched_text_start
7244 && addr < (unsigned long)__sched_text_end);
7247 #ifdef CONFIG_CGROUP_SCHED
7248 struct task_group root_task_group;
7251 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
7253 void __init sched_init(void)
7256 unsigned long alloc_size = 0, ptr;
7258 #ifdef CONFIG_FAIR_GROUP_SCHED
7259 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7261 #ifdef CONFIG_RT_GROUP_SCHED
7262 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7264 #ifdef CONFIG_CPUMASK_OFFSTACK
7265 alloc_size += num_possible_cpus() * cpumask_size();
7268 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7270 #ifdef CONFIG_FAIR_GROUP_SCHED
7271 root_task_group.se = (struct sched_entity **)ptr;
7272 ptr += nr_cpu_ids * sizeof(void **);
7274 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7275 ptr += nr_cpu_ids * sizeof(void **);
7277 #endif /* CONFIG_FAIR_GROUP_SCHED */
7278 #ifdef CONFIG_RT_GROUP_SCHED
7279 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7280 ptr += nr_cpu_ids * sizeof(void **);
7282 root_task_group.rt_rq = (struct rt_rq **)ptr;
7283 ptr += nr_cpu_ids * sizeof(void **);
7285 #endif /* CONFIG_RT_GROUP_SCHED */
7286 #ifdef CONFIG_CPUMASK_OFFSTACK
7287 for_each_possible_cpu(i) {
7288 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7289 ptr += cpumask_size();
7291 #endif /* CONFIG_CPUMASK_OFFSTACK */
7295 init_defrootdomain();
7298 init_rt_bandwidth(&def_rt_bandwidth,
7299 global_rt_period(), global_rt_runtime());
7301 #ifdef CONFIG_RT_GROUP_SCHED
7302 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7303 global_rt_period(), global_rt_runtime());
7304 #endif /* CONFIG_RT_GROUP_SCHED */
7306 #ifdef CONFIG_CGROUP_SCHED
7307 list_add(&root_task_group.list, &task_groups);
7308 INIT_LIST_HEAD(&root_task_group.children);
7309 INIT_LIST_HEAD(&root_task_group.siblings);
7310 autogroup_init(&init_task);
7312 #endif /* CONFIG_CGROUP_SCHED */
7314 #ifdef CONFIG_CGROUP_CPUACCT
7315 root_cpuacct.cpustat = &kernel_cpustat;
7316 root_cpuacct.cpuusage = alloc_percpu(u64);
7317 /* Too early, not expected to fail */
7318 BUG_ON(!root_cpuacct.cpuusage);
7320 for_each_possible_cpu(i) {
7324 raw_spin_lock_init(&rq->lock);
7326 rq->calc_load_active = 0;
7327 rq->calc_load_update = jiffies + LOAD_FREQ;
7328 init_cfs_rq(&rq->cfs);
7329 init_rt_rq(&rq->rt, rq);
7330 #ifdef CONFIG_FAIR_GROUP_SCHED
7331 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7332 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7334 * How much cpu bandwidth does root_task_group get?
7336 * In case of task-groups formed thr' the cgroup filesystem, it
7337 * gets 100% of the cpu resources in the system. This overall
7338 * system cpu resource is divided among the tasks of
7339 * root_task_group and its child task-groups in a fair manner,
7340 * based on each entity's (task or task-group's) weight
7341 * (se->load.weight).
7343 * In other words, if root_task_group has 10 tasks of weight
7344 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7345 * then A0's share of the cpu resource is:
7347 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7349 * We achieve this by letting root_task_group's tasks sit
7350 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7352 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7353 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7354 #endif /* CONFIG_FAIR_GROUP_SCHED */
7356 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7357 #ifdef CONFIG_RT_GROUP_SCHED
7358 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7359 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7362 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7363 rq->cpu_load[j] = 0;
7365 rq->last_load_update_tick = jiffies;
7370 rq->cpu_power = SCHED_POWER_SCALE;
7371 rq->post_schedule = 0;
7372 rq->active_balance = 0;
7373 rq->next_balance = jiffies;
7378 rq->avg_idle = 2*sysctl_sched_migration_cost;
7380 INIT_LIST_HEAD(&rq->cfs_tasks);
7382 rq_attach_root(rq, &def_root_domain);
7388 atomic_set(&rq->nr_iowait, 0);
7391 set_load_weight(&init_task);
7393 #ifdef CONFIG_PREEMPT_NOTIFIERS
7394 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7397 #ifdef CONFIG_RT_MUTEXES
7398 plist_head_init(&init_task.pi_waiters);
7402 * The boot idle thread does lazy MMU switching as well:
7404 atomic_inc(&init_mm.mm_count);
7405 enter_lazy_tlb(&init_mm, current);
7408 * Make us the idle thread. Technically, schedule() should not be
7409 * called from this thread, however somewhere below it might be,
7410 * but because we are the idle thread, we just pick up running again
7411 * when this runqueue becomes "idle".
7413 init_idle(current, smp_processor_id());
7415 calc_load_update = jiffies + LOAD_FREQ;
7418 * During early bootup we pretend to be a normal task:
7420 current->sched_class = &fair_sched_class;
7423 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7424 /* May be allocated at isolcpus cmdline parse time */
7425 if (cpu_isolated_map == NULL)
7426 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7427 idle_thread_set_boot_cpu();
7429 init_sched_fair_class();
7431 scheduler_running = 1;
7434 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7435 static inline int preempt_count_equals(int preempt_offset)
7437 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7439 return (nested == preempt_offset);
7442 void __might_sleep(const char *file, int line, int preempt_offset)
7444 static unsigned long prev_jiffy; /* ratelimiting */
7446 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7447 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7448 system_state != SYSTEM_RUNNING || oops_in_progress)
7450 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7452 prev_jiffy = jiffies;
7455 "BUG: sleeping function called from invalid context at %s:%d\n",
7458 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7459 in_atomic(), irqs_disabled(),
7460 current->pid, current->comm);
7462 debug_show_held_locks(current);
7463 if (irqs_disabled())
7464 print_irqtrace_events(current);
7467 EXPORT_SYMBOL(__might_sleep);
7470 #ifdef CONFIG_MAGIC_SYSRQ
7471 static void normalize_task(struct rq *rq, struct task_struct *p)
7473 const struct sched_class *prev_class = p->sched_class;
7474 int old_prio = p->prio;
7479 dequeue_task(rq, p, 0);
7480 __setscheduler(rq, p, SCHED_NORMAL, 0);
7482 enqueue_task(rq, p, 0);
7483 resched_task(rq->curr);
7486 check_class_changed(rq, p, prev_class, old_prio);
7489 void normalize_rt_tasks(void)
7491 struct task_struct *g, *p;
7492 unsigned long flags;
7495 read_lock_irqsave(&tasklist_lock, flags);
7496 do_each_thread(g, p) {
7498 * Only normalize user tasks:
7503 p->se.exec_start = 0;
7504 #ifdef CONFIG_SCHEDSTATS
7505 p->se.statistics.wait_start = 0;
7506 p->se.statistics.sleep_start = 0;
7507 p->se.statistics.block_start = 0;
7512 * Renice negative nice level userspace
7515 if (TASK_NICE(p) < 0 && p->mm)
7516 set_user_nice(p, 0);
7520 raw_spin_lock(&p->pi_lock);
7521 rq = __task_rq_lock(p);
7523 normalize_task(rq, p);
7525 __task_rq_unlock(rq);
7526 raw_spin_unlock(&p->pi_lock);
7527 } while_each_thread(g, p);
7529 read_unlock_irqrestore(&tasklist_lock, flags);
7532 #endif /* CONFIG_MAGIC_SYSRQ */
7534 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7536 * These functions are only useful for the IA64 MCA handling, or kdb.
7538 * They can only be called when the whole system has been
7539 * stopped - every CPU needs to be quiescent, and no scheduling
7540 * activity can take place. Using them for anything else would
7541 * be a serious bug, and as a result, they aren't even visible
7542 * under any other configuration.
7546 * curr_task - return the current task for a given cpu.
7547 * @cpu: the processor in question.
7549 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7551 struct task_struct *curr_task(int cpu)
7553 return cpu_curr(cpu);
7556 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7560 * set_curr_task - set the current task for a given cpu.
7561 * @cpu: the processor in question.
7562 * @p: the task pointer to set.
7564 * Description: This function must only be used when non-maskable interrupts
7565 * are serviced on a separate stack. It allows the architecture to switch the
7566 * notion of the current task on a cpu in a non-blocking manner. This function
7567 * must be called with all CPU's synchronized, and interrupts disabled, the
7568 * and caller must save the original value of the current task (see
7569 * curr_task() above) and restore that value before reenabling interrupts and
7570 * re-starting the system.
7572 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7574 void set_curr_task(int cpu, struct task_struct *p)
7581 #ifdef CONFIG_CGROUP_SCHED
7582 /* task_group_lock serializes the addition/removal of task groups */
7583 static DEFINE_SPINLOCK(task_group_lock);
7585 static void free_sched_group(struct task_group *tg)
7587 free_fair_sched_group(tg);
7588 free_rt_sched_group(tg);
7593 /* allocate runqueue etc for a new task group */
7594 struct task_group *sched_create_group(struct task_group *parent)
7596 struct task_group *tg;
7597 unsigned long flags;
7599 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7601 return ERR_PTR(-ENOMEM);
7603 if (!alloc_fair_sched_group(tg, parent))
7606 if (!alloc_rt_sched_group(tg, parent))
7609 spin_lock_irqsave(&task_group_lock, flags);
7610 list_add_rcu(&tg->list, &task_groups);
7612 WARN_ON(!parent); /* root should already exist */
7614 tg->parent = parent;
7615 INIT_LIST_HEAD(&tg->children);
7616 list_add_rcu(&tg->siblings, &parent->children);
7617 spin_unlock_irqrestore(&task_group_lock, flags);
7622 free_sched_group(tg);
7623 return ERR_PTR(-ENOMEM);
7626 /* rcu callback to free various structures associated with a task group */
7627 static void free_sched_group_rcu(struct rcu_head *rhp)
7629 /* now it should be safe to free those cfs_rqs */
7630 free_sched_group(container_of(rhp, struct task_group, rcu));
7633 /* Destroy runqueue etc associated with a task group */
7634 void sched_destroy_group(struct task_group *tg)
7636 unsigned long flags;
7639 /* end participation in shares distribution */
7640 for_each_possible_cpu(i)
7641 unregister_fair_sched_group(tg, i);
7643 spin_lock_irqsave(&task_group_lock, flags);
7644 list_del_rcu(&tg->list);
7645 list_del_rcu(&tg->siblings);
7646 spin_unlock_irqrestore(&task_group_lock, flags);
7648 /* wait for possible concurrent references to cfs_rqs complete */
7649 call_rcu(&tg->rcu, free_sched_group_rcu);
7652 /* change task's runqueue when it moves between groups.
7653 * The caller of this function should have put the task in its new group
7654 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7655 * reflect its new group.
7657 void sched_move_task(struct task_struct *tsk)
7659 struct task_group *tg;
7661 unsigned long flags;
7664 rq = task_rq_lock(tsk, &flags);
7666 running = task_current(rq, tsk);
7670 dequeue_task(rq, tsk, 0);
7671 if (unlikely(running))
7672 tsk->sched_class->put_prev_task(rq, tsk);
7674 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7675 lockdep_is_held(&tsk->sighand->siglock)),
7676 struct task_group, css);
7677 tg = autogroup_task_group(tsk, tg);
7678 tsk->sched_task_group = tg;
7680 #ifdef CONFIG_FAIR_GROUP_SCHED
7681 if (tsk->sched_class->task_move_group)
7682 tsk->sched_class->task_move_group(tsk, on_rq);
7685 set_task_rq(tsk, task_cpu(tsk));
7687 if (unlikely(running))
7688 tsk->sched_class->set_curr_task(rq);
7690 enqueue_task(rq, tsk, 0);
7692 task_rq_unlock(rq, tsk, &flags);
7694 #endif /* CONFIG_CGROUP_SCHED */
7696 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7697 static unsigned long to_ratio(u64 period, u64 runtime)
7699 if (runtime == RUNTIME_INF)
7702 return div64_u64(runtime << 20, period);
7706 #ifdef CONFIG_RT_GROUP_SCHED
7708 * Ensure that the real time constraints are schedulable.
7710 static DEFINE_MUTEX(rt_constraints_mutex);
7712 /* Must be called with tasklist_lock held */
7713 static inline int tg_has_rt_tasks(struct task_group *tg)
7715 struct task_struct *g, *p;
7717 do_each_thread(g, p) {
7718 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7720 } while_each_thread(g, p);
7725 struct rt_schedulable_data {
7726 struct task_group *tg;
7731 static int tg_rt_schedulable(struct task_group *tg, void *data)
7733 struct rt_schedulable_data *d = data;
7734 struct task_group *child;
7735 unsigned long total, sum = 0;
7736 u64 period, runtime;
7738 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7739 runtime = tg->rt_bandwidth.rt_runtime;
7742 period = d->rt_period;
7743 runtime = d->rt_runtime;
7747 * Cannot have more runtime than the period.
7749 if (runtime > period && runtime != RUNTIME_INF)
7753 * Ensure we don't starve existing RT tasks.
7755 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7758 total = to_ratio(period, runtime);
7761 * Nobody can have more than the global setting allows.
7763 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7767 * The sum of our children's runtime should not exceed our own.
7769 list_for_each_entry_rcu(child, &tg->children, siblings) {
7770 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7771 runtime = child->rt_bandwidth.rt_runtime;
7773 if (child == d->tg) {
7774 period = d->rt_period;
7775 runtime = d->rt_runtime;
7778 sum += to_ratio(period, runtime);
7787 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7791 struct rt_schedulable_data data = {
7793 .rt_period = period,
7794 .rt_runtime = runtime,
7798 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7804 static int tg_set_rt_bandwidth(struct task_group *tg,
7805 u64 rt_period, u64 rt_runtime)
7809 mutex_lock(&rt_constraints_mutex);
7810 read_lock(&tasklist_lock);
7811 err = __rt_schedulable(tg, rt_period, rt_runtime);
7815 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7816 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7817 tg->rt_bandwidth.rt_runtime = rt_runtime;
7819 for_each_possible_cpu(i) {
7820 struct rt_rq *rt_rq = tg->rt_rq[i];
7822 raw_spin_lock(&rt_rq->rt_runtime_lock);
7823 rt_rq->rt_runtime = rt_runtime;
7824 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7826 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7828 read_unlock(&tasklist_lock);
7829 mutex_unlock(&rt_constraints_mutex);
7834 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7836 u64 rt_runtime, rt_period;
7838 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7839 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7840 if (rt_runtime_us < 0)
7841 rt_runtime = RUNTIME_INF;
7843 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7846 long sched_group_rt_runtime(struct task_group *tg)
7850 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7853 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7854 do_div(rt_runtime_us, NSEC_PER_USEC);
7855 return rt_runtime_us;
7858 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7860 u64 rt_runtime, rt_period;
7862 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7863 rt_runtime = tg->rt_bandwidth.rt_runtime;
7868 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7871 long sched_group_rt_period(struct task_group *tg)
7875 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7876 do_div(rt_period_us, NSEC_PER_USEC);
7877 return rt_period_us;
7880 static int sched_rt_global_constraints(void)
7882 u64 runtime, period;
7885 if (sysctl_sched_rt_period <= 0)
7888 runtime = global_rt_runtime();
7889 period = global_rt_period();
7892 * Sanity check on the sysctl variables.
7894 if (runtime > period && runtime != RUNTIME_INF)
7897 mutex_lock(&rt_constraints_mutex);
7898 read_lock(&tasklist_lock);
7899 ret = __rt_schedulable(NULL, 0, 0);
7900 read_unlock(&tasklist_lock);
7901 mutex_unlock(&rt_constraints_mutex);
7906 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7908 /* Don't accept realtime tasks when there is no way for them to run */
7909 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7915 #else /* !CONFIG_RT_GROUP_SCHED */
7916 static int sched_rt_global_constraints(void)
7918 unsigned long flags;
7921 if (sysctl_sched_rt_period <= 0)
7925 * There's always some RT tasks in the root group
7926 * -- migration, kstopmachine etc..
7928 if (sysctl_sched_rt_runtime == 0)
7931 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7932 for_each_possible_cpu(i) {
7933 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7935 raw_spin_lock(&rt_rq->rt_runtime_lock);
7936 rt_rq->rt_runtime = global_rt_runtime();
7937 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7939 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7943 #endif /* CONFIG_RT_GROUP_SCHED */
7945 int sched_rt_handler(struct ctl_table *table, int write,
7946 void __user *buffer, size_t *lenp,
7950 int old_period, old_runtime;
7951 static DEFINE_MUTEX(mutex);
7954 old_period = sysctl_sched_rt_period;
7955 old_runtime = sysctl_sched_rt_runtime;
7957 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7959 if (!ret && write) {
7960 ret = sched_rt_global_constraints();
7962 sysctl_sched_rt_period = old_period;
7963 sysctl_sched_rt_runtime = old_runtime;
7965 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7966 def_rt_bandwidth.rt_period =
7967 ns_to_ktime(global_rt_period());
7970 mutex_unlock(&mutex);
7975 #ifdef CONFIG_CGROUP_SCHED
7977 /* return corresponding task_group object of a cgroup */
7978 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7980 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7981 struct task_group, css);
7984 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7986 struct task_group *tg, *parent;
7988 if (!cgrp->parent) {
7989 /* This is early initialization for the top cgroup */
7990 return &root_task_group.css;
7993 parent = cgroup_tg(cgrp->parent);
7994 tg = sched_create_group(parent);
7996 return ERR_PTR(-ENOMEM);
8001 static void cpu_cgroup_destroy(struct cgroup *cgrp)
8003 struct task_group *tg = cgroup_tg(cgrp);
8005 sched_destroy_group(tg);
8008 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
8009 struct cgroup_taskset *tset)
8011 struct task_struct *task;
8013 cgroup_taskset_for_each(task, cgrp, tset) {
8014 #ifdef CONFIG_RT_GROUP_SCHED
8015 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
8018 /* We don't support RT-tasks being in separate groups */
8019 if (task->sched_class != &fair_sched_class)
8026 static void cpu_cgroup_attach(struct cgroup *cgrp,
8027 struct cgroup_taskset *tset)
8029 struct task_struct *task;
8031 cgroup_taskset_for_each(task, cgrp, tset)
8032 sched_move_task(task);
8036 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
8037 struct task_struct *task)
8040 * cgroup_exit() is called in the copy_process() failure path.
8041 * Ignore this case since the task hasn't ran yet, this avoids
8042 * trying to poke a half freed task state from generic code.
8044 if (!(task->flags & PF_EXITING))
8047 sched_move_task(task);
8050 #ifdef CONFIG_FAIR_GROUP_SCHED
8051 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8054 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
8057 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8059 struct task_group *tg = cgroup_tg(cgrp);
8061 return (u64) scale_load_down(tg->shares);
8064 #ifdef CONFIG_CFS_BANDWIDTH
8065 static DEFINE_MUTEX(cfs_constraints_mutex);
8067 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8068 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8070 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8072 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8074 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8075 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8077 if (tg == &root_task_group)
8081 * Ensure we have at some amount of bandwidth every period. This is
8082 * to prevent reaching a state of large arrears when throttled via
8083 * entity_tick() resulting in prolonged exit starvation.
8085 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8089 * Likewise, bound things on the otherside by preventing insane quota
8090 * periods. This also allows us to normalize in computing quota
8093 if (period > max_cfs_quota_period)
8096 mutex_lock(&cfs_constraints_mutex);
8097 ret = __cfs_schedulable(tg, period, quota);
8101 runtime_enabled = quota != RUNTIME_INF;
8102 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8103 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
8104 raw_spin_lock_irq(&cfs_b->lock);
8105 cfs_b->period = ns_to_ktime(period);
8106 cfs_b->quota = quota;
8108 __refill_cfs_bandwidth_runtime(cfs_b);
8109 /* restart the period timer (if active) to handle new period expiry */
8110 if (runtime_enabled && cfs_b->timer_active) {
8111 /* force a reprogram */
8112 cfs_b->timer_active = 0;
8113 __start_cfs_bandwidth(cfs_b);
8115 raw_spin_unlock_irq(&cfs_b->lock);
8117 for_each_possible_cpu(i) {
8118 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8119 struct rq *rq = cfs_rq->rq;
8121 raw_spin_lock_irq(&rq->lock);
8122 cfs_rq->runtime_enabled = runtime_enabled;
8123 cfs_rq->runtime_remaining = 0;
8125 if (cfs_rq->throttled)
8126 unthrottle_cfs_rq(cfs_rq);
8127 raw_spin_unlock_irq(&rq->lock);
8130 mutex_unlock(&cfs_constraints_mutex);
8135 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8139 period = ktime_to_ns(tg->cfs_bandwidth.period);
8140 if (cfs_quota_us < 0)
8141 quota = RUNTIME_INF;
8143 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8145 return tg_set_cfs_bandwidth(tg, period, quota);
8148 long tg_get_cfs_quota(struct task_group *tg)
8152 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8155 quota_us = tg->cfs_bandwidth.quota;
8156 do_div(quota_us, NSEC_PER_USEC);
8161 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8165 period = (u64)cfs_period_us * NSEC_PER_USEC;
8166 quota = tg->cfs_bandwidth.quota;
8168 return tg_set_cfs_bandwidth(tg, period, quota);
8171 long tg_get_cfs_period(struct task_group *tg)
8175 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8176 do_div(cfs_period_us, NSEC_PER_USEC);
8178 return cfs_period_us;
8181 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
8183 return tg_get_cfs_quota(cgroup_tg(cgrp));
8186 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
8189 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
8192 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
8194 return tg_get_cfs_period(cgroup_tg(cgrp));
8197 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8200 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
8203 struct cfs_schedulable_data {
8204 struct task_group *tg;
8209 * normalize group quota/period to be quota/max_period
8210 * note: units are usecs
8212 static u64 normalize_cfs_quota(struct task_group *tg,
8213 struct cfs_schedulable_data *d)
8221 period = tg_get_cfs_period(tg);
8222 quota = tg_get_cfs_quota(tg);
8225 /* note: these should typically be equivalent */
8226 if (quota == RUNTIME_INF || quota == -1)
8229 return to_ratio(period, quota);
8232 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8234 struct cfs_schedulable_data *d = data;
8235 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8236 s64 quota = 0, parent_quota = -1;
8239 quota = RUNTIME_INF;
8241 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8243 quota = normalize_cfs_quota(tg, d);
8244 parent_quota = parent_b->hierarchal_quota;
8247 * ensure max(child_quota) <= parent_quota, inherit when no
8250 if (quota == RUNTIME_INF)
8251 quota = parent_quota;
8252 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8255 cfs_b->hierarchal_quota = quota;
8260 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8263 struct cfs_schedulable_data data = {
8269 if (quota != RUNTIME_INF) {
8270 do_div(data.period, NSEC_PER_USEC);
8271 do_div(data.quota, NSEC_PER_USEC);
8275 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8281 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8282 struct cgroup_map_cb *cb)
8284 struct task_group *tg = cgroup_tg(cgrp);
8285 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8287 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8288 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8289 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8293 #endif /* CONFIG_CFS_BANDWIDTH */
8294 #endif /* CONFIG_FAIR_GROUP_SCHED */
8296 #ifdef CONFIG_RT_GROUP_SCHED
8297 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8300 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8303 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8305 return sched_group_rt_runtime(cgroup_tg(cgrp));
8308 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8311 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8314 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8316 return sched_group_rt_period(cgroup_tg(cgrp));
8318 #endif /* CONFIG_RT_GROUP_SCHED */
8320 static struct cftype cpu_files[] = {
8321 #ifdef CONFIG_FAIR_GROUP_SCHED
8324 .read_u64 = cpu_shares_read_u64,
8325 .write_u64 = cpu_shares_write_u64,
8328 #ifdef CONFIG_CFS_BANDWIDTH
8330 .name = "cfs_quota_us",
8331 .read_s64 = cpu_cfs_quota_read_s64,
8332 .write_s64 = cpu_cfs_quota_write_s64,
8335 .name = "cfs_period_us",
8336 .read_u64 = cpu_cfs_period_read_u64,
8337 .write_u64 = cpu_cfs_period_write_u64,
8341 .read_map = cpu_stats_show,
8344 #ifdef CONFIG_RT_GROUP_SCHED
8346 .name = "rt_runtime_us",
8347 .read_s64 = cpu_rt_runtime_read,
8348 .write_s64 = cpu_rt_runtime_write,
8351 .name = "rt_period_us",
8352 .read_u64 = cpu_rt_period_read_uint,
8353 .write_u64 = cpu_rt_period_write_uint,
8359 struct cgroup_subsys cpu_cgroup_subsys = {
8361 .create = cpu_cgroup_create,
8362 .destroy = cpu_cgroup_destroy,
8363 .can_attach = cpu_cgroup_can_attach,
8364 .attach = cpu_cgroup_attach,
8365 .exit = cpu_cgroup_exit,
8366 .subsys_id = cpu_cgroup_subsys_id,
8367 .base_cftypes = cpu_files,
8371 #endif /* CONFIG_CGROUP_SCHED */
8373 #ifdef CONFIG_CGROUP_CPUACCT
8376 * CPU accounting code for task groups.
8378 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8379 * (balbir@in.ibm.com).
8382 /* create a new cpu accounting group */
8383 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
8388 return &root_cpuacct.css;
8390 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8394 ca->cpuusage = alloc_percpu(u64);
8398 ca->cpustat = alloc_percpu(struct kernel_cpustat);
8400 goto out_free_cpuusage;
8405 free_percpu(ca->cpuusage);
8409 return ERR_PTR(-ENOMEM);
8412 /* destroy an existing cpu accounting group */
8413 static void cpuacct_destroy(struct cgroup *cgrp)
8415 struct cpuacct *ca = cgroup_ca(cgrp);
8417 free_percpu(ca->cpustat);
8418 free_percpu(ca->cpuusage);
8422 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8424 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8427 #ifndef CONFIG_64BIT
8429 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8431 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8433 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8441 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8443 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8445 #ifndef CONFIG_64BIT
8447 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8449 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8451 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8457 /* return total cpu usage (in nanoseconds) of a group */
8458 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8460 struct cpuacct *ca = cgroup_ca(cgrp);
8461 u64 totalcpuusage = 0;
8464 for_each_present_cpu(i)
8465 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8467 return totalcpuusage;
8470 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8473 struct cpuacct *ca = cgroup_ca(cgrp);
8482 for_each_present_cpu(i)
8483 cpuacct_cpuusage_write(ca, i, 0);
8489 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8492 struct cpuacct *ca = cgroup_ca(cgroup);
8496 for_each_present_cpu(i) {
8497 percpu = cpuacct_cpuusage_read(ca, i);
8498 seq_printf(m, "%llu ", (unsigned long long) percpu);
8500 seq_printf(m, "\n");
8504 static const char *cpuacct_stat_desc[] = {
8505 [CPUACCT_STAT_USER] = "user",
8506 [CPUACCT_STAT_SYSTEM] = "system",
8509 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8510 struct cgroup_map_cb *cb)
8512 struct cpuacct *ca = cgroup_ca(cgrp);
8516 for_each_online_cpu(cpu) {
8517 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8518 val += kcpustat->cpustat[CPUTIME_USER];
8519 val += kcpustat->cpustat[CPUTIME_NICE];
8521 val = cputime64_to_clock_t(val);
8522 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8525 for_each_online_cpu(cpu) {
8526 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8527 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8528 val += kcpustat->cpustat[CPUTIME_IRQ];
8529 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8532 val = cputime64_to_clock_t(val);
8533 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8538 static struct cftype files[] = {
8541 .read_u64 = cpuusage_read,
8542 .write_u64 = cpuusage_write,
8545 .name = "usage_percpu",
8546 .read_seq_string = cpuacct_percpu_seq_read,
8550 .read_map = cpuacct_stats_show,
8556 * charge this task's execution time to its accounting group.
8558 * called with rq->lock held.
8560 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8565 if (unlikely(!cpuacct_subsys.active))
8568 cpu = task_cpu(tsk);
8574 for (; ca; ca = parent_ca(ca)) {
8575 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8576 *cpuusage += cputime;
8582 struct cgroup_subsys cpuacct_subsys = {
8584 .create = cpuacct_create,
8585 .destroy = cpuacct_destroy,
8586 .subsys_id = cpuacct_subsys_id,
8587 .base_cftypes = files,
8589 #endif /* CONFIG_CGROUP_CPUACCT */