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>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_sched.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
516 #ifndef tsk_is_polling
517 #define tsk_is_polling(t) 0
520 void resched_task(struct task_struct *p)
524 assert_raw_spin_locked(&task_rq(p)->lock);
526 if (test_tsk_need_resched(p))
529 set_tsk_need_resched(p);
532 if (cpu == smp_processor_id())
535 /* NEED_RESCHED must be visible before we test polling */
537 if (!tsk_is_polling(p))
538 smp_send_reschedule(cpu);
541 void resched_cpu(int cpu)
543 struct rq *rq = cpu_rq(cpu);
546 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
548 resched_task(cpu_curr(cpu));
549 raw_spin_unlock_irqrestore(&rq->lock, flags);
554 * In the semi idle case, use the nearest busy cpu for migrating timers
555 * from an idle cpu. This is good for power-savings.
557 * We don't do similar optimization for completely idle system, as
558 * selecting an idle cpu will add more delays to the timers than intended
559 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
561 int get_nohz_timer_target(void)
563 int cpu = smp_processor_id();
565 struct sched_domain *sd;
568 for_each_domain(cpu, sd) {
569 for_each_cpu(i, sched_domain_span(sd)) {
581 * When add_timer_on() enqueues a timer into the timer wheel of an
582 * idle CPU then this timer might expire before the next timer event
583 * which is scheduled to wake up that CPU. In case of a completely
584 * idle system the next event might even be infinite time into the
585 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
586 * leaves the inner idle loop so the newly added timer is taken into
587 * account when the CPU goes back to idle and evaluates the timer
588 * wheel for the next timer event.
590 void wake_up_idle_cpu(int cpu)
592 struct rq *rq = cpu_rq(cpu);
594 if (cpu == smp_processor_id())
598 * This is safe, as this function is called with the timer
599 * wheel base lock of (cpu) held. When the CPU is on the way
600 * to idle and has not yet set rq->curr to idle then it will
601 * be serialized on the timer wheel base lock and take the new
602 * timer into account automatically.
604 if (rq->curr != rq->idle)
608 * We can set TIF_RESCHED on the idle task of the other CPU
609 * lockless. The worst case is that the other CPU runs the
610 * idle task through an additional NOOP schedule()
612 set_tsk_need_resched(rq->idle);
614 /* NEED_RESCHED must be visible before we test polling */
616 if (!tsk_is_polling(rq->idle))
617 smp_send_reschedule(cpu);
620 static inline bool got_nohz_idle_kick(void)
622 int cpu = smp_processor_id();
623 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
626 #else /* CONFIG_NO_HZ */
628 static inline bool got_nohz_idle_kick(void)
633 #endif /* CONFIG_NO_HZ */
635 void sched_avg_update(struct rq *rq)
637 s64 period = sched_avg_period();
639 while ((s64)(rq->clock - rq->age_stamp) > period) {
641 * Inline assembly required to prevent the compiler
642 * optimising this loop into a divmod call.
643 * See __iter_div_u64_rem() for another example of this.
645 asm("" : "+rm" (rq->age_stamp));
646 rq->age_stamp += period;
651 #else /* !CONFIG_SMP */
652 void resched_task(struct task_struct *p)
654 assert_raw_spin_locked(&task_rq(p)->lock);
655 set_tsk_need_resched(p);
657 #endif /* CONFIG_SMP */
659 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
660 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
662 * Iterate task_group tree rooted at *from, calling @down when first entering a
663 * node and @up when leaving it for the final time.
665 * Caller must hold rcu_lock or sufficient equivalent.
667 int walk_tg_tree_from(struct task_group *from,
668 tg_visitor down, tg_visitor up, void *data)
670 struct task_group *parent, *child;
676 ret = (*down)(parent, data);
679 list_for_each_entry_rcu(child, &parent->children, siblings) {
686 ret = (*up)(parent, data);
687 if (ret || parent == from)
691 parent = parent->parent;
698 int tg_nop(struct task_group *tg, void *data)
704 static void set_load_weight(struct task_struct *p)
706 int prio = p->static_prio - MAX_RT_PRIO;
707 struct load_weight *load = &p->se.load;
710 * SCHED_IDLE tasks get minimal weight:
712 if (p->policy == SCHED_IDLE) {
713 load->weight = scale_load(WEIGHT_IDLEPRIO);
714 load->inv_weight = WMULT_IDLEPRIO;
718 load->weight = scale_load(prio_to_weight[prio]);
719 load->inv_weight = prio_to_wmult[prio];
722 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
725 sched_info_queued(p);
726 p->sched_class->enqueue_task(rq, p, flags);
729 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
732 sched_info_dequeued(p);
733 p->sched_class->dequeue_task(rq, p, flags);
736 void activate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible--;
741 enqueue_task(rq, p, flags);
744 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
746 if (task_contributes_to_load(p))
747 rq->nr_uninterruptible++;
749 dequeue_task(rq, p, flags);
752 static void update_rq_clock_task(struct rq *rq, s64 delta)
755 * In theory, the compile should just see 0 here, and optimize out the call
756 * to sched_rt_avg_update. But I don't trust it...
758 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
759 s64 steal = 0, irq_delta = 0;
761 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
762 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
765 * Since irq_time is only updated on {soft,}irq_exit, we might run into
766 * this case when a previous update_rq_clock() happened inside a
769 * When this happens, we stop ->clock_task and only update the
770 * prev_irq_time stamp to account for the part that fit, so that a next
771 * update will consume the rest. This ensures ->clock_task is
774 * It does however cause some slight miss-attribution of {soft,}irq
775 * time, a more accurate solution would be to update the irq_time using
776 * the current rq->clock timestamp, except that would require using
779 if (irq_delta > delta)
782 rq->prev_irq_time += irq_delta;
785 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
786 if (static_key_false((¶virt_steal_rq_enabled))) {
789 steal = paravirt_steal_clock(cpu_of(rq));
790 steal -= rq->prev_steal_time_rq;
792 if (unlikely(steal > delta))
795 st = steal_ticks(steal);
796 steal = st * TICK_NSEC;
798 rq->prev_steal_time_rq += steal;
804 rq->clock_task += delta;
806 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
807 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
808 sched_rt_avg_update(rq, irq_delta + steal);
812 void sched_set_stop_task(int cpu, struct task_struct *stop)
814 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
815 struct task_struct *old_stop = cpu_rq(cpu)->stop;
819 * Make it appear like a SCHED_FIFO task, its something
820 * userspace knows about and won't get confused about.
822 * Also, it will make PI more or less work without too
823 * much confusion -- but then, stop work should not
824 * rely on PI working anyway.
826 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
828 stop->sched_class = &stop_sched_class;
831 cpu_rq(cpu)->stop = stop;
835 * Reset it back to a normal scheduling class so that
836 * it can die in pieces.
838 old_stop->sched_class = &rt_sched_class;
843 * __normal_prio - return the priority that is based on the static prio
845 static inline int __normal_prio(struct task_struct *p)
847 return p->static_prio;
851 * Calculate the expected normal priority: i.e. priority
852 * without taking RT-inheritance into account. Might be
853 * boosted by interactivity modifiers. Changes upon fork,
854 * setprio syscalls, and whenever the interactivity
855 * estimator recalculates.
857 static inline int normal_prio(struct task_struct *p)
861 if (task_has_rt_policy(p))
862 prio = MAX_RT_PRIO-1 - p->rt_priority;
864 prio = __normal_prio(p);
869 * Calculate the current priority, i.e. the priority
870 * taken into account by the scheduler. This value might
871 * be boosted by RT tasks, or might be boosted by
872 * interactivity modifiers. Will be RT if the task got
873 * RT-boosted. If not then it returns p->normal_prio.
875 static int effective_prio(struct task_struct *p)
877 p->normal_prio = normal_prio(p);
879 * If we are RT tasks or we were boosted to RT priority,
880 * keep the priority unchanged. Otherwise, update priority
881 * to the normal priority:
883 if (!rt_prio(p->prio))
884 return p->normal_prio;
889 * task_curr - is this task currently executing on a CPU?
890 * @p: the task in question.
892 inline int task_curr(const struct task_struct *p)
894 return cpu_curr(task_cpu(p)) == p;
897 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
898 const struct sched_class *prev_class,
901 if (prev_class != p->sched_class) {
902 if (prev_class->switched_from)
903 prev_class->switched_from(rq, p);
904 p->sched_class->switched_to(rq, p);
905 } else if (oldprio != p->prio)
906 p->sched_class->prio_changed(rq, p, oldprio);
909 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
911 const struct sched_class *class;
913 if (p->sched_class == rq->curr->sched_class) {
914 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
916 for_each_class(class) {
917 if (class == rq->curr->sched_class)
919 if (class == p->sched_class) {
920 resched_task(rq->curr);
927 * A queue event has occurred, and we're going to schedule. In
928 * this case, we can save a useless back to back clock update.
930 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
931 rq->skip_clock_update = 1;
934 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
936 void register_task_migration_notifier(struct notifier_block *n)
938 atomic_notifier_chain_register(&task_migration_notifier, n);
942 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
944 #ifdef CONFIG_SCHED_DEBUG
946 * We should never call set_task_cpu() on a blocked task,
947 * ttwu() will sort out the placement.
949 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
950 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
952 #ifdef CONFIG_LOCKDEP
954 * The caller should hold either p->pi_lock or rq->lock, when changing
955 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
957 * sched_move_task() holds both and thus holding either pins the cgroup,
960 * Furthermore, all task_rq users should acquire both locks, see
963 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
964 lockdep_is_held(&task_rq(p)->lock)));
968 trace_sched_migrate_task(p, new_cpu);
970 if (task_cpu(p) != new_cpu) {
971 struct task_migration_notifier tmn;
973 if (p->sched_class->migrate_task_rq)
974 p->sched_class->migrate_task_rq(p, new_cpu);
975 p->se.nr_migrations++;
976 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
979 tmn.from_cpu = task_cpu(p);
980 tmn.to_cpu = new_cpu;
982 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
985 __set_task_cpu(p, new_cpu);
988 struct migration_arg {
989 struct task_struct *task;
993 static int migration_cpu_stop(void *data);
996 * wait_task_inactive - wait for a thread to unschedule.
998 * If @match_state is nonzero, it's the @p->state value just checked and
999 * not expected to change. If it changes, i.e. @p might have woken up,
1000 * then return zero. When we succeed in waiting for @p to be off its CPU,
1001 * we return a positive number (its total switch count). If a second call
1002 * a short while later returns the same number, the caller can be sure that
1003 * @p has remained unscheduled the whole time.
1005 * The caller must ensure that the task *will* unschedule sometime soon,
1006 * else this function might spin for a *long* time. This function can't
1007 * be called with interrupts off, or it may introduce deadlock with
1008 * smp_call_function() if an IPI is sent by the same process we are
1009 * waiting to become inactive.
1011 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1013 unsigned long flags;
1020 * We do the initial early heuristics without holding
1021 * any task-queue locks at all. We'll only try to get
1022 * the runqueue lock when things look like they will
1028 * If the task is actively running on another CPU
1029 * still, just relax and busy-wait without holding
1032 * NOTE! Since we don't hold any locks, it's not
1033 * even sure that "rq" stays as the right runqueue!
1034 * But we don't care, since "task_running()" will
1035 * return false if the runqueue has changed and p
1036 * is actually now running somewhere else!
1038 while (task_running(rq, p)) {
1039 if (match_state && unlikely(p->state != match_state))
1045 * Ok, time to look more closely! We need the rq
1046 * lock now, to be *sure*. If we're wrong, we'll
1047 * just go back and repeat.
1049 rq = task_rq_lock(p, &flags);
1050 trace_sched_wait_task(p);
1051 running = task_running(rq, p);
1054 if (!match_state || p->state == match_state)
1055 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1056 task_rq_unlock(rq, p, &flags);
1059 * If it changed from the expected state, bail out now.
1061 if (unlikely(!ncsw))
1065 * Was it really running after all now that we
1066 * checked with the proper locks actually held?
1068 * Oops. Go back and try again..
1070 if (unlikely(running)) {
1076 * It's not enough that it's not actively running,
1077 * it must be off the runqueue _entirely_, and not
1080 * So if it was still runnable (but just not actively
1081 * running right now), it's preempted, and we should
1082 * yield - it could be a while.
1084 if (unlikely(on_rq)) {
1085 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1087 set_current_state(TASK_UNINTERRUPTIBLE);
1088 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1093 * Ahh, all good. It wasn't running, and it wasn't
1094 * runnable, which means that it will never become
1095 * running in the future either. We're all done!
1104 * kick_process - kick a running thread to enter/exit the kernel
1105 * @p: the to-be-kicked thread
1107 * Cause a process which is running on another CPU to enter
1108 * kernel-mode, without any delay. (to get signals handled.)
1110 * NOTE: this function doesn't have to take the runqueue lock,
1111 * because all it wants to ensure is that the remote task enters
1112 * the kernel. If the IPI races and the task has been migrated
1113 * to another CPU then no harm is done and the purpose has been
1116 void kick_process(struct task_struct *p)
1122 if ((cpu != smp_processor_id()) && task_curr(p))
1123 smp_send_reschedule(cpu);
1126 EXPORT_SYMBOL_GPL(kick_process);
1127 #endif /* CONFIG_SMP */
1131 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1133 static int select_fallback_rq(int cpu, struct task_struct *p)
1135 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1136 enum { cpuset, possible, fail } state = cpuset;
1139 /* Look for allowed, online CPU in same node. */
1140 for_each_cpu(dest_cpu, nodemask) {
1141 if (!cpu_online(dest_cpu))
1143 if (!cpu_active(dest_cpu))
1145 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1150 /* Any allowed, online CPU? */
1151 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1152 if (!cpu_online(dest_cpu))
1154 if (!cpu_active(dest_cpu))
1161 /* No more Mr. Nice Guy. */
1162 cpuset_cpus_allowed_fallback(p);
1167 do_set_cpus_allowed(p, cpu_possible_mask);
1178 if (state != cpuset) {
1180 * Don't tell them about moving exiting tasks or
1181 * kernel threads (both mm NULL), since they never
1184 if (p->mm && printk_ratelimit()) {
1185 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1186 task_pid_nr(p), p->comm, cpu);
1194 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1197 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1199 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1202 * In order not to call set_task_cpu() on a blocking task we need
1203 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1206 * Since this is common to all placement strategies, this lives here.
1208 * [ this allows ->select_task() to simply return task_cpu(p) and
1209 * not worry about this generic constraint ]
1211 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1213 cpu = select_fallback_rq(task_cpu(p), p);
1218 static void update_avg(u64 *avg, u64 sample)
1220 s64 diff = sample - *avg;
1226 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1228 #ifdef CONFIG_SCHEDSTATS
1229 struct rq *rq = this_rq();
1232 int this_cpu = smp_processor_id();
1234 if (cpu == this_cpu) {
1235 schedstat_inc(rq, ttwu_local);
1236 schedstat_inc(p, se.statistics.nr_wakeups_local);
1238 struct sched_domain *sd;
1240 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1242 for_each_domain(this_cpu, sd) {
1243 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1244 schedstat_inc(sd, ttwu_wake_remote);
1251 if (wake_flags & WF_MIGRATED)
1252 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1254 #endif /* CONFIG_SMP */
1256 schedstat_inc(rq, ttwu_count);
1257 schedstat_inc(p, se.statistics.nr_wakeups);
1259 if (wake_flags & WF_SYNC)
1260 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1262 #endif /* CONFIG_SCHEDSTATS */
1265 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1267 activate_task(rq, p, en_flags);
1270 /* if a worker is waking up, notify workqueue */
1271 if (p->flags & PF_WQ_WORKER)
1272 wq_worker_waking_up(p, cpu_of(rq));
1276 * Mark the task runnable and perform wakeup-preemption.
1279 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1281 trace_sched_wakeup(p, true);
1282 check_preempt_curr(rq, p, wake_flags);
1284 p->state = TASK_RUNNING;
1286 if (p->sched_class->task_woken)
1287 p->sched_class->task_woken(rq, p);
1289 if (rq->idle_stamp) {
1290 u64 delta = rq->clock - rq->idle_stamp;
1291 u64 max = 2*sysctl_sched_migration_cost;
1296 update_avg(&rq->avg_idle, delta);
1303 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1306 if (p->sched_contributes_to_load)
1307 rq->nr_uninterruptible--;
1310 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1311 ttwu_do_wakeup(rq, p, wake_flags);
1315 * Called in case the task @p isn't fully descheduled from its runqueue,
1316 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1317 * since all we need to do is flip p->state to TASK_RUNNING, since
1318 * the task is still ->on_rq.
1320 static int ttwu_remote(struct task_struct *p, int wake_flags)
1325 rq = __task_rq_lock(p);
1327 ttwu_do_wakeup(rq, p, wake_flags);
1330 __task_rq_unlock(rq);
1336 static void sched_ttwu_pending(void)
1338 struct rq *rq = this_rq();
1339 struct llist_node *llist = llist_del_all(&rq->wake_list);
1340 struct task_struct *p;
1342 raw_spin_lock(&rq->lock);
1345 p = llist_entry(llist, struct task_struct, wake_entry);
1346 llist = llist_next(llist);
1347 ttwu_do_activate(rq, p, 0);
1350 raw_spin_unlock(&rq->lock);
1353 void scheduler_ipi(void)
1355 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1359 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1360 * traditionally all their work was done from the interrupt return
1361 * path. Now that we actually do some work, we need to make sure
1364 * Some archs already do call them, luckily irq_enter/exit nest
1367 * Arguably we should visit all archs and update all handlers,
1368 * however a fair share of IPIs are still resched only so this would
1369 * somewhat pessimize the simple resched case.
1372 sched_ttwu_pending();
1375 * Check if someone kicked us for doing the nohz idle load balance.
1377 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1378 this_rq()->idle_balance = 1;
1379 raise_softirq_irqoff(SCHED_SOFTIRQ);
1384 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1386 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1387 smp_send_reschedule(cpu);
1390 bool cpus_share_cache(int this_cpu, int that_cpu)
1392 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1394 #endif /* CONFIG_SMP */
1396 static void ttwu_queue(struct task_struct *p, int cpu)
1398 struct rq *rq = cpu_rq(cpu);
1400 #if defined(CONFIG_SMP)
1401 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1402 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1403 ttwu_queue_remote(p, cpu);
1408 raw_spin_lock(&rq->lock);
1409 ttwu_do_activate(rq, p, 0);
1410 raw_spin_unlock(&rq->lock);
1414 * try_to_wake_up - wake up a thread
1415 * @p: the thread to be awakened
1416 * @state: the mask of task states that can be woken
1417 * @wake_flags: wake modifier flags (WF_*)
1419 * Put it on the run-queue if it's not already there. The "current"
1420 * thread is always on the run-queue (except when the actual
1421 * re-schedule is in progress), and as such you're allowed to do
1422 * the simpler "current->state = TASK_RUNNING" to mark yourself
1423 * runnable without the overhead of this.
1425 * Returns %true if @p was woken up, %false if it was already running
1426 * or @state didn't match @p's state.
1429 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1431 unsigned long flags;
1432 int cpu, success = 0;
1435 raw_spin_lock_irqsave(&p->pi_lock, flags);
1436 if (!(p->state & state))
1439 success = 1; /* we're going to change ->state */
1442 if (p->on_rq && ttwu_remote(p, wake_flags))
1447 * If the owning (remote) cpu is still in the middle of schedule() with
1448 * this task as prev, wait until its done referencing the task.
1453 * Pairs with the smp_wmb() in finish_lock_switch().
1457 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1458 p->state = TASK_WAKING;
1460 if (p->sched_class->task_waking)
1461 p->sched_class->task_waking(p);
1463 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1464 if (task_cpu(p) != cpu) {
1465 wake_flags |= WF_MIGRATED;
1466 set_task_cpu(p, cpu);
1468 #endif /* CONFIG_SMP */
1472 ttwu_stat(p, cpu, wake_flags);
1474 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1480 * try_to_wake_up_local - try to wake up a local task with rq lock held
1481 * @p: the thread to be awakened
1483 * Put @p on the run-queue if it's not already there. The caller must
1484 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1487 static void try_to_wake_up_local(struct task_struct *p)
1489 struct rq *rq = task_rq(p);
1491 BUG_ON(rq != this_rq());
1492 BUG_ON(p == current);
1493 lockdep_assert_held(&rq->lock);
1495 if (!raw_spin_trylock(&p->pi_lock)) {
1496 raw_spin_unlock(&rq->lock);
1497 raw_spin_lock(&p->pi_lock);
1498 raw_spin_lock(&rq->lock);
1501 if (!(p->state & TASK_NORMAL))
1505 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1507 ttwu_do_wakeup(rq, p, 0);
1508 ttwu_stat(p, smp_processor_id(), 0);
1510 raw_spin_unlock(&p->pi_lock);
1514 * wake_up_process - Wake up a specific process
1515 * @p: The process to be woken up.
1517 * Attempt to wake up the nominated process and move it to the set of runnable
1518 * processes. Returns 1 if the process was woken up, 0 if it was already
1521 * It may be assumed that this function implies a write memory barrier before
1522 * changing the task state if and only if any tasks are woken up.
1524 int wake_up_process(struct task_struct *p)
1526 WARN_ON(task_is_stopped_or_traced(p));
1527 return try_to_wake_up(p, TASK_NORMAL, 0);
1529 EXPORT_SYMBOL(wake_up_process);
1531 int wake_up_state(struct task_struct *p, unsigned int state)
1533 return try_to_wake_up(p, state, 0);
1537 * Perform scheduler related setup for a newly forked process p.
1538 * p is forked by current.
1540 * __sched_fork() is basic setup used by init_idle() too:
1542 static void __sched_fork(struct task_struct *p)
1547 p->se.exec_start = 0;
1548 p->se.sum_exec_runtime = 0;
1549 p->se.prev_sum_exec_runtime = 0;
1550 p->se.nr_migrations = 0;
1552 INIT_LIST_HEAD(&p->se.group_node);
1555 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1556 * removed when useful for applications beyond shares distribution (e.g.
1559 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1560 p->se.avg.runnable_avg_period = 0;
1561 p->se.avg.runnable_avg_sum = 0;
1563 #ifdef CONFIG_SCHEDSTATS
1564 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1567 INIT_LIST_HEAD(&p->rt.run_list);
1569 #ifdef CONFIG_PREEMPT_NOTIFIERS
1570 INIT_HLIST_HEAD(&p->preempt_notifiers);
1573 #ifdef CONFIG_NUMA_BALANCING
1574 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1575 p->mm->numa_next_scan = jiffies;
1576 p->mm->numa_next_reset = jiffies;
1577 p->mm->numa_scan_seq = 0;
1580 p->node_stamp = 0ULL;
1581 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1582 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1583 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1584 p->numa_work.next = &p->numa_work;
1585 #endif /* CONFIG_NUMA_BALANCING */
1588 #ifdef CONFIG_NUMA_BALANCING
1589 #ifdef CONFIG_SCHED_DEBUG
1590 void set_numabalancing_state(bool enabled)
1593 sched_feat_set("NUMA");
1595 sched_feat_set("NO_NUMA");
1598 __read_mostly bool numabalancing_enabled;
1600 void set_numabalancing_state(bool enabled)
1602 numabalancing_enabled = enabled;
1604 #endif /* CONFIG_SCHED_DEBUG */
1605 #endif /* CONFIG_NUMA_BALANCING */
1608 * fork()/clone()-time setup:
1610 void sched_fork(struct task_struct *p)
1612 unsigned long flags;
1613 int cpu = get_cpu();
1617 * We mark the process as running here. This guarantees that
1618 * nobody will actually run it, and a signal or other external
1619 * event cannot wake it up and insert it on the runqueue either.
1621 p->state = TASK_RUNNING;
1624 * Make sure we do not leak PI boosting priority to the child.
1626 p->prio = current->normal_prio;
1629 * Revert to default priority/policy on fork if requested.
1631 if (unlikely(p->sched_reset_on_fork)) {
1632 if (task_has_rt_policy(p)) {
1633 p->policy = SCHED_NORMAL;
1634 p->static_prio = NICE_TO_PRIO(0);
1636 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1637 p->static_prio = NICE_TO_PRIO(0);
1639 p->prio = p->normal_prio = __normal_prio(p);
1643 * We don't need the reset flag anymore after the fork. It has
1644 * fulfilled its duty:
1646 p->sched_reset_on_fork = 0;
1649 if (!rt_prio(p->prio))
1650 p->sched_class = &fair_sched_class;
1652 if (p->sched_class->task_fork)
1653 p->sched_class->task_fork(p);
1656 * The child is not yet in the pid-hash so no cgroup attach races,
1657 * and the cgroup is pinned to this child due to cgroup_fork()
1658 * is ran before sched_fork().
1660 * Silence PROVE_RCU.
1662 raw_spin_lock_irqsave(&p->pi_lock, flags);
1663 set_task_cpu(p, cpu);
1664 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1666 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1667 if (likely(sched_info_on()))
1668 memset(&p->sched_info, 0, sizeof(p->sched_info));
1670 #if defined(CONFIG_SMP)
1673 #ifdef CONFIG_PREEMPT_COUNT
1674 /* Want to start with kernel preemption disabled. */
1675 task_thread_info(p)->preempt_count = 1;
1678 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1685 * wake_up_new_task - wake up a newly created task for the first time.
1687 * This function will do some initial scheduler statistics housekeeping
1688 * that must be done for every newly created context, then puts the task
1689 * on the runqueue and wakes it.
1691 void wake_up_new_task(struct task_struct *p)
1693 unsigned long flags;
1696 raw_spin_lock_irqsave(&p->pi_lock, flags);
1699 * Fork balancing, do it here and not earlier because:
1700 * - cpus_allowed can change in the fork path
1701 * - any previously selected cpu might disappear through hotplug
1703 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1706 rq = __task_rq_lock(p);
1707 activate_task(rq, p, 0);
1709 trace_sched_wakeup_new(p, true);
1710 check_preempt_curr(rq, p, WF_FORK);
1712 if (p->sched_class->task_woken)
1713 p->sched_class->task_woken(rq, p);
1715 task_rq_unlock(rq, p, &flags);
1718 #ifdef CONFIG_PREEMPT_NOTIFIERS
1721 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1722 * @notifier: notifier struct to register
1724 void preempt_notifier_register(struct preempt_notifier *notifier)
1726 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1728 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1731 * preempt_notifier_unregister - no longer interested in preemption notifications
1732 * @notifier: notifier struct to unregister
1734 * This is safe to call from within a preemption notifier.
1736 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1738 hlist_del(¬ifier->link);
1740 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1742 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1744 struct preempt_notifier *notifier;
1745 struct hlist_node *node;
1747 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1748 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1752 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1753 struct task_struct *next)
1755 struct preempt_notifier *notifier;
1756 struct hlist_node *node;
1758 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1759 notifier->ops->sched_out(notifier, next);
1762 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1764 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1769 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1770 struct task_struct *next)
1774 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1777 * prepare_task_switch - prepare to switch tasks
1778 * @rq: the runqueue preparing to switch
1779 * @prev: the current task that is being switched out
1780 * @next: the task we are going to switch to.
1782 * This is called with the rq lock held and interrupts off. It must
1783 * be paired with a subsequent finish_task_switch after the context
1786 * prepare_task_switch sets up locking and calls architecture specific
1790 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1791 struct task_struct *next)
1793 trace_sched_switch(prev, next);
1794 sched_info_switch(prev, next);
1795 perf_event_task_sched_out(prev, next);
1796 fire_sched_out_preempt_notifiers(prev, next);
1797 prepare_lock_switch(rq, next);
1798 prepare_arch_switch(next);
1802 * finish_task_switch - clean up after a task-switch
1803 * @rq: runqueue associated with task-switch
1804 * @prev: the thread we just switched away from.
1806 * finish_task_switch must be called after the context switch, paired
1807 * with a prepare_task_switch call before the context switch.
1808 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1809 * and do any other architecture-specific cleanup actions.
1811 * Note that we may have delayed dropping an mm in context_switch(). If
1812 * so, we finish that here outside of the runqueue lock. (Doing it
1813 * with the lock held can cause deadlocks; see schedule() for
1816 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1817 __releases(rq->lock)
1819 struct mm_struct *mm = rq->prev_mm;
1825 * A task struct has one reference for the use as "current".
1826 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1827 * schedule one last time. The schedule call will never return, and
1828 * the scheduled task must drop that reference.
1829 * The test for TASK_DEAD must occur while the runqueue locks are
1830 * still held, otherwise prev could be scheduled on another cpu, die
1831 * there before we look at prev->state, and then the reference would
1833 * Manfred Spraul <manfred@colorfullife.com>
1835 prev_state = prev->state;
1836 vtime_task_switch(prev);
1837 finish_arch_switch(prev);
1838 perf_event_task_sched_in(prev, current);
1839 finish_lock_switch(rq, prev);
1840 finish_arch_post_lock_switch();
1842 fire_sched_in_preempt_notifiers(current);
1845 if (unlikely(prev_state == TASK_DEAD)) {
1847 * Remove function-return probe instances associated with this
1848 * task and put them back on the free list.
1850 kprobe_flush_task(prev);
1851 put_task_struct(prev);
1857 /* assumes rq->lock is held */
1858 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1860 if (prev->sched_class->pre_schedule)
1861 prev->sched_class->pre_schedule(rq, prev);
1864 /* rq->lock is NOT held, but preemption is disabled */
1865 static inline void post_schedule(struct rq *rq)
1867 if (rq->post_schedule) {
1868 unsigned long flags;
1870 raw_spin_lock_irqsave(&rq->lock, flags);
1871 if (rq->curr->sched_class->post_schedule)
1872 rq->curr->sched_class->post_schedule(rq);
1873 raw_spin_unlock_irqrestore(&rq->lock, flags);
1875 rq->post_schedule = 0;
1881 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1885 static inline void post_schedule(struct rq *rq)
1892 * schedule_tail - first thing a freshly forked thread must call.
1893 * @prev: the thread we just switched away from.
1895 asmlinkage void schedule_tail(struct task_struct *prev)
1896 __releases(rq->lock)
1898 struct rq *rq = this_rq();
1900 finish_task_switch(rq, prev);
1903 * FIXME: do we need to worry about rq being invalidated by the
1908 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1909 /* In this case, finish_task_switch does not reenable preemption */
1912 if (current->set_child_tid)
1913 put_user(task_pid_vnr(current), current->set_child_tid);
1917 * context_switch - switch to the new MM and the new
1918 * thread's register state.
1921 context_switch(struct rq *rq, struct task_struct *prev,
1922 struct task_struct *next)
1924 struct mm_struct *mm, *oldmm;
1926 prepare_task_switch(rq, prev, next);
1929 oldmm = prev->active_mm;
1931 * For paravirt, this is coupled with an exit in switch_to to
1932 * combine the page table reload and the switch backend into
1935 arch_start_context_switch(prev);
1938 next->active_mm = oldmm;
1939 atomic_inc(&oldmm->mm_count);
1940 enter_lazy_tlb(oldmm, next);
1942 switch_mm(oldmm, mm, next);
1945 prev->active_mm = NULL;
1946 rq->prev_mm = oldmm;
1949 * Since the runqueue lock will be released by the next
1950 * task (which is an invalid locking op but in the case
1951 * of the scheduler it's an obvious special-case), so we
1952 * do an early lockdep release here:
1954 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1955 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1958 context_tracking_task_switch(prev, next);
1959 /* Here we just switch the register state and the stack. */
1960 switch_to(prev, next, prev);
1964 * this_rq must be evaluated again because prev may have moved
1965 * CPUs since it called schedule(), thus the 'rq' on its stack
1966 * frame will be invalid.
1968 finish_task_switch(this_rq(), prev);
1972 * nr_running, nr_uninterruptible and nr_context_switches:
1974 * externally visible scheduler statistics: current number of runnable
1975 * threads, current number of uninterruptible-sleeping threads, total
1976 * number of context switches performed since bootup.
1978 unsigned long nr_running(void)
1980 unsigned long i, sum = 0;
1982 for_each_online_cpu(i)
1983 sum += cpu_rq(i)->nr_running;
1988 unsigned long nr_uninterruptible(void)
1990 unsigned long i, sum = 0;
1992 for_each_possible_cpu(i)
1993 sum += cpu_rq(i)->nr_uninterruptible;
1996 * Since we read the counters lockless, it might be slightly
1997 * inaccurate. Do not allow it to go below zero though:
1999 if (unlikely((long)sum < 0))
2005 unsigned long long nr_context_switches(void)
2008 unsigned long long sum = 0;
2010 for_each_possible_cpu(i)
2011 sum += cpu_rq(i)->nr_switches;
2016 unsigned long nr_iowait(void)
2018 unsigned long i, sum = 0;
2020 for_each_possible_cpu(i)
2021 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2026 unsigned long nr_iowait_cpu(int cpu)
2028 struct rq *this = cpu_rq(cpu);
2029 return atomic_read(&this->nr_iowait);
2032 unsigned long this_cpu_load(void)
2034 struct rq *this = this_rq();
2035 return this->cpu_load[0];
2040 * Global load-average calculations
2042 * We take a distributed and async approach to calculating the global load-avg
2043 * in order to minimize overhead.
2045 * The global load average is an exponentially decaying average of nr_running +
2046 * nr_uninterruptible.
2048 * Once every LOAD_FREQ:
2051 * for_each_possible_cpu(cpu)
2052 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2054 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2056 * Due to a number of reasons the above turns in the mess below:
2058 * - for_each_possible_cpu() is prohibitively expensive on machines with
2059 * serious number of cpus, therefore we need to take a distributed approach
2060 * to calculating nr_active.
2062 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2063 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2065 * So assuming nr_active := 0 when we start out -- true per definition, we
2066 * can simply take per-cpu deltas and fold those into a global accumulate
2067 * to obtain the same result. See calc_load_fold_active().
2069 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2070 * across the machine, we assume 10 ticks is sufficient time for every
2071 * cpu to have completed this task.
2073 * This places an upper-bound on the IRQ-off latency of the machine. Then
2074 * again, being late doesn't loose the delta, just wrecks the sample.
2076 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2077 * this would add another cross-cpu cacheline miss and atomic operation
2078 * to the wakeup path. Instead we increment on whatever cpu the task ran
2079 * when it went into uninterruptible state and decrement on whatever cpu
2080 * did the wakeup. This means that only the sum of nr_uninterruptible over
2081 * all cpus yields the correct result.
2083 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2086 /* Variables and functions for calc_load */
2087 static atomic_long_t calc_load_tasks;
2088 static unsigned long calc_load_update;
2089 unsigned long avenrun[3];
2090 EXPORT_SYMBOL(avenrun); /* should be removed */
2093 * get_avenrun - get the load average array
2094 * @loads: pointer to dest load array
2095 * @offset: offset to add
2096 * @shift: shift count to shift the result left
2098 * These values are estimates at best, so no need for locking.
2100 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2102 loads[0] = (avenrun[0] + offset) << shift;
2103 loads[1] = (avenrun[1] + offset) << shift;
2104 loads[2] = (avenrun[2] + offset) << shift;
2107 static long calc_load_fold_active(struct rq *this_rq)
2109 long nr_active, delta = 0;
2111 nr_active = this_rq->nr_running;
2112 nr_active += (long) this_rq->nr_uninterruptible;
2114 if (nr_active != this_rq->calc_load_active) {
2115 delta = nr_active - this_rq->calc_load_active;
2116 this_rq->calc_load_active = nr_active;
2123 * a1 = a0 * e + a * (1 - e)
2125 static unsigned long
2126 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2129 load += active * (FIXED_1 - exp);
2130 load += 1UL << (FSHIFT - 1);
2131 return load >> FSHIFT;
2136 * Handle NO_HZ for the global load-average.
2138 * Since the above described distributed algorithm to compute the global
2139 * load-average relies on per-cpu sampling from the tick, it is affected by
2142 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2143 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2144 * when we read the global state.
2146 * Obviously reality has to ruin such a delightfully simple scheme:
2148 * - When we go NO_HZ idle during the window, we can negate our sample
2149 * contribution, causing under-accounting.
2151 * We avoid this by keeping two idle-delta counters and flipping them
2152 * when the window starts, thus separating old and new NO_HZ load.
2154 * The only trick is the slight shift in index flip for read vs write.
2158 * |-|-----------|-|-----------|-|-----------|-|
2159 * r:0 0 1 1 0 0 1 1 0
2160 * w:0 1 1 0 0 1 1 0 0
2162 * This ensures we'll fold the old idle contribution in this window while
2163 * accumlating the new one.
2165 * - When we wake up from NO_HZ idle during the window, we push up our
2166 * contribution, since we effectively move our sample point to a known
2169 * This is solved by pushing the window forward, and thus skipping the
2170 * sample, for this cpu (effectively using the idle-delta for this cpu which
2171 * was in effect at the time the window opened). This also solves the issue
2172 * of having to deal with a cpu having been in NOHZ idle for multiple
2173 * LOAD_FREQ intervals.
2175 * When making the ILB scale, we should try to pull this in as well.
2177 static atomic_long_t calc_load_idle[2];
2178 static int calc_load_idx;
2180 static inline int calc_load_write_idx(void)
2182 int idx = calc_load_idx;
2185 * See calc_global_nohz(), if we observe the new index, we also
2186 * need to observe the new update time.
2191 * If the folding window started, make sure we start writing in the
2194 if (!time_before(jiffies, calc_load_update))
2200 static inline int calc_load_read_idx(void)
2202 return calc_load_idx & 1;
2205 void calc_load_enter_idle(void)
2207 struct rq *this_rq = this_rq();
2211 * We're going into NOHZ mode, if there's any pending delta, fold it
2212 * into the pending idle delta.
2214 delta = calc_load_fold_active(this_rq);
2216 int idx = calc_load_write_idx();
2217 atomic_long_add(delta, &calc_load_idle[idx]);
2221 void calc_load_exit_idle(void)
2223 struct rq *this_rq = this_rq();
2226 * If we're still before the sample window, we're done.
2228 if (time_before(jiffies, this_rq->calc_load_update))
2232 * We woke inside or after the sample window, this means we're already
2233 * accounted through the nohz accounting, so skip the entire deal and
2234 * sync up for the next window.
2236 this_rq->calc_load_update = calc_load_update;
2237 if (time_before(jiffies, this_rq->calc_load_update + 10))
2238 this_rq->calc_load_update += LOAD_FREQ;
2241 static long calc_load_fold_idle(void)
2243 int idx = calc_load_read_idx();
2246 if (atomic_long_read(&calc_load_idle[idx]))
2247 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2253 * fixed_power_int - compute: x^n, in O(log n) time
2255 * @x: base of the power
2256 * @frac_bits: fractional bits of @x
2257 * @n: power to raise @x to.
2259 * By exploiting the relation between the definition of the natural power
2260 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2261 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2262 * (where: n_i \elem {0, 1}, the binary vector representing n),
2263 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2264 * of course trivially computable in O(log_2 n), the length of our binary
2267 static unsigned long
2268 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2270 unsigned long result = 1UL << frac_bits;
2275 result += 1UL << (frac_bits - 1);
2276 result >>= frac_bits;
2282 x += 1UL << (frac_bits - 1);
2290 * a1 = a0 * e + a * (1 - e)
2292 * a2 = a1 * e + a * (1 - e)
2293 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2294 * = a0 * e^2 + a * (1 - e) * (1 + e)
2296 * a3 = a2 * e + a * (1 - e)
2297 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2298 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2302 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2303 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2304 * = a0 * e^n + a * (1 - e^n)
2306 * [1] application of the geometric series:
2309 * S_n := \Sum x^i = -------------
2312 static unsigned long
2313 calc_load_n(unsigned long load, unsigned long exp,
2314 unsigned long active, unsigned int n)
2317 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2321 * NO_HZ can leave us missing all per-cpu ticks calling
2322 * calc_load_account_active(), but since an idle CPU folds its delta into
2323 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2324 * in the pending idle delta if our idle period crossed a load cycle boundary.
2326 * Once we've updated the global active value, we need to apply the exponential
2327 * weights adjusted to the number of cycles missed.
2329 static void calc_global_nohz(void)
2331 long delta, active, n;
2333 if (!time_before(jiffies, calc_load_update + 10)) {
2335 * Catch-up, fold however many we are behind still
2337 delta = jiffies - calc_load_update - 10;
2338 n = 1 + (delta / LOAD_FREQ);
2340 active = atomic_long_read(&calc_load_tasks);
2341 active = active > 0 ? active * FIXED_1 : 0;
2343 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2344 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2345 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2347 calc_load_update += n * LOAD_FREQ;
2351 * Flip the idle index...
2353 * Make sure we first write the new time then flip the index, so that
2354 * calc_load_write_idx() will see the new time when it reads the new
2355 * index, this avoids a double flip messing things up.
2360 #else /* !CONFIG_NO_HZ */
2362 static inline long calc_load_fold_idle(void) { return 0; }
2363 static inline void calc_global_nohz(void) { }
2365 #endif /* CONFIG_NO_HZ */
2368 * calc_load - update the avenrun load estimates 10 ticks after the
2369 * CPUs have updated calc_load_tasks.
2371 void calc_global_load(unsigned long ticks)
2375 if (time_before(jiffies, calc_load_update + 10))
2379 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2381 delta = calc_load_fold_idle();
2383 atomic_long_add(delta, &calc_load_tasks);
2385 active = atomic_long_read(&calc_load_tasks);
2386 active = active > 0 ? active * FIXED_1 : 0;
2388 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2389 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2390 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2392 calc_load_update += LOAD_FREQ;
2395 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2401 * Called from update_cpu_load() to periodically update this CPU's
2404 static void calc_load_account_active(struct rq *this_rq)
2408 if (time_before(jiffies, this_rq->calc_load_update))
2411 delta = calc_load_fold_active(this_rq);
2413 atomic_long_add(delta, &calc_load_tasks);
2415 this_rq->calc_load_update += LOAD_FREQ;
2419 * End of global load-average stuff
2423 * The exact cpuload at various idx values, calculated at every tick would be
2424 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2426 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2427 * on nth tick when cpu may be busy, then we have:
2428 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2429 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2431 * decay_load_missed() below does efficient calculation of
2432 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2433 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2435 * The calculation is approximated on a 128 point scale.
2436 * degrade_zero_ticks is the number of ticks after which load at any
2437 * particular idx is approximated to be zero.
2438 * degrade_factor is a precomputed table, a row for each load idx.
2439 * Each column corresponds to degradation factor for a power of two ticks,
2440 * based on 128 point scale.
2442 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2443 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2445 * With this power of 2 load factors, we can degrade the load n times
2446 * by looking at 1 bits in n and doing as many mult/shift instead of
2447 * n mult/shifts needed by the exact degradation.
2449 #define DEGRADE_SHIFT 7
2450 static const unsigned char
2451 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2452 static const unsigned char
2453 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2454 {0, 0, 0, 0, 0, 0, 0, 0},
2455 {64, 32, 8, 0, 0, 0, 0, 0},
2456 {96, 72, 40, 12, 1, 0, 0},
2457 {112, 98, 75, 43, 15, 1, 0},
2458 {120, 112, 98, 76, 45, 16, 2} };
2461 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2462 * would be when CPU is idle and so we just decay the old load without
2463 * adding any new load.
2465 static unsigned long
2466 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2470 if (!missed_updates)
2473 if (missed_updates >= degrade_zero_ticks[idx])
2477 return load >> missed_updates;
2479 while (missed_updates) {
2480 if (missed_updates % 2)
2481 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2483 missed_updates >>= 1;
2490 * Update rq->cpu_load[] statistics. This function is usually called every
2491 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2492 * every tick. We fix it up based on jiffies.
2494 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2495 unsigned long pending_updates)
2499 this_rq->nr_load_updates++;
2501 /* Update our load: */
2502 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2503 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2504 unsigned long old_load, new_load;
2506 /* scale is effectively 1 << i now, and >> i divides by scale */
2508 old_load = this_rq->cpu_load[i];
2509 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2510 new_load = this_load;
2512 * Round up the averaging division if load is increasing. This
2513 * prevents us from getting stuck on 9 if the load is 10, for
2516 if (new_load > old_load)
2517 new_load += scale - 1;
2519 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2522 sched_avg_update(this_rq);
2527 * There is no sane way to deal with nohz on smp when using jiffies because the
2528 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2529 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2531 * Therefore we cannot use the delta approach from the regular tick since that
2532 * would seriously skew the load calculation. However we'll make do for those
2533 * updates happening while idle (nohz_idle_balance) or coming out of idle
2534 * (tick_nohz_idle_exit).
2536 * This means we might still be one tick off for nohz periods.
2540 * Called from nohz_idle_balance() to update the load ratings before doing the
2543 void update_idle_cpu_load(struct rq *this_rq)
2545 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2546 unsigned long load = this_rq->load.weight;
2547 unsigned long pending_updates;
2550 * bail if there's load or we're actually up-to-date.
2552 if (load || curr_jiffies == this_rq->last_load_update_tick)
2555 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2556 this_rq->last_load_update_tick = curr_jiffies;
2558 __update_cpu_load(this_rq, load, pending_updates);
2562 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2564 void update_cpu_load_nohz(void)
2566 struct rq *this_rq = this_rq();
2567 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2568 unsigned long pending_updates;
2570 if (curr_jiffies == this_rq->last_load_update_tick)
2573 raw_spin_lock(&this_rq->lock);
2574 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2575 if (pending_updates) {
2576 this_rq->last_load_update_tick = curr_jiffies;
2578 * We were idle, this means load 0, the current load might be
2579 * !0 due to remote wakeups and the sort.
2581 __update_cpu_load(this_rq, 0, pending_updates);
2583 raw_spin_unlock(&this_rq->lock);
2585 #endif /* CONFIG_NO_HZ */
2588 * Called from scheduler_tick()
2590 static void update_cpu_load_active(struct rq *this_rq)
2593 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2595 this_rq->last_load_update_tick = jiffies;
2596 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2598 calc_load_account_active(this_rq);
2604 * sched_exec - execve() is a valuable balancing opportunity, because at
2605 * this point the task has the smallest effective memory and cache footprint.
2607 void sched_exec(void)
2609 struct task_struct *p = current;
2610 unsigned long flags;
2613 raw_spin_lock_irqsave(&p->pi_lock, flags);
2614 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2615 if (dest_cpu == smp_processor_id())
2618 if (likely(cpu_active(dest_cpu))) {
2619 struct migration_arg arg = { p, dest_cpu };
2621 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2622 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2626 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2631 DEFINE_PER_CPU(struct kernel_stat, kstat);
2632 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2634 EXPORT_PER_CPU_SYMBOL(kstat);
2635 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2638 * Return any ns on the sched_clock that have not yet been accounted in
2639 * @p in case that task is currently running.
2641 * Called with task_rq_lock() held on @rq.
2643 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2647 if (task_current(rq, p)) {
2648 update_rq_clock(rq);
2649 ns = rq->clock_task - p->se.exec_start;
2657 unsigned long long task_delta_exec(struct task_struct *p)
2659 unsigned long flags;
2663 rq = task_rq_lock(p, &flags);
2664 ns = do_task_delta_exec(p, rq);
2665 task_rq_unlock(rq, p, &flags);
2671 * Return accounted runtime for the task.
2672 * In case the task is currently running, return the runtime plus current's
2673 * pending runtime that have not been accounted yet.
2675 unsigned long long task_sched_runtime(struct task_struct *p)
2677 unsigned long flags;
2681 rq = task_rq_lock(p, &flags);
2682 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2683 task_rq_unlock(rq, p, &flags);
2689 * This function gets called by the timer code, with HZ frequency.
2690 * We call it with interrupts disabled.
2692 void scheduler_tick(void)
2694 int cpu = smp_processor_id();
2695 struct rq *rq = cpu_rq(cpu);
2696 struct task_struct *curr = rq->curr;
2700 raw_spin_lock(&rq->lock);
2701 update_rq_clock(rq);
2702 update_cpu_load_active(rq);
2703 curr->sched_class->task_tick(rq, curr, 0);
2704 raw_spin_unlock(&rq->lock);
2706 perf_event_task_tick();
2709 rq->idle_balance = idle_cpu(cpu);
2710 trigger_load_balance(rq, cpu);
2714 notrace unsigned long get_parent_ip(unsigned long addr)
2716 if (in_lock_functions(addr)) {
2717 addr = CALLER_ADDR2;
2718 if (in_lock_functions(addr))
2719 addr = CALLER_ADDR3;
2724 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2725 defined(CONFIG_PREEMPT_TRACER))
2727 void __kprobes add_preempt_count(int val)
2729 #ifdef CONFIG_DEBUG_PREEMPT
2733 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2736 preempt_count() += val;
2737 #ifdef CONFIG_DEBUG_PREEMPT
2739 * Spinlock count overflowing soon?
2741 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2744 if (preempt_count() == val)
2745 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2747 EXPORT_SYMBOL(add_preempt_count);
2749 void __kprobes sub_preempt_count(int val)
2751 #ifdef CONFIG_DEBUG_PREEMPT
2755 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2758 * Is the spinlock portion underflowing?
2760 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2761 !(preempt_count() & PREEMPT_MASK)))
2765 if (preempt_count() == val)
2766 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2767 preempt_count() -= val;
2769 EXPORT_SYMBOL(sub_preempt_count);
2774 * Print scheduling while atomic bug:
2776 static noinline void __schedule_bug(struct task_struct *prev)
2778 if (oops_in_progress)
2781 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2782 prev->comm, prev->pid, preempt_count());
2784 debug_show_held_locks(prev);
2786 if (irqs_disabled())
2787 print_irqtrace_events(prev);
2789 add_taint(TAINT_WARN);
2793 * Various schedule()-time debugging checks and statistics:
2795 static inline void schedule_debug(struct task_struct *prev)
2798 * Test if we are atomic. Since do_exit() needs to call into
2799 * schedule() atomically, we ignore that path for now.
2800 * Otherwise, whine if we are scheduling when we should not be.
2802 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2803 __schedule_bug(prev);
2806 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2808 schedstat_inc(this_rq(), sched_count);
2811 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2813 if (prev->on_rq || rq->skip_clock_update < 0)
2814 update_rq_clock(rq);
2815 prev->sched_class->put_prev_task(rq, prev);
2819 * Pick up the highest-prio task:
2821 static inline struct task_struct *
2822 pick_next_task(struct rq *rq)
2824 const struct sched_class *class;
2825 struct task_struct *p;
2828 * Optimization: we know that if all tasks are in
2829 * the fair class we can call that function directly:
2831 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2832 p = fair_sched_class.pick_next_task(rq);
2837 for_each_class(class) {
2838 p = class->pick_next_task(rq);
2843 BUG(); /* the idle class will always have a runnable task */
2847 * __schedule() is the main scheduler function.
2849 * The main means of driving the scheduler and thus entering this function are:
2851 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2853 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2854 * paths. For example, see arch/x86/entry_64.S.
2856 * To drive preemption between tasks, the scheduler sets the flag in timer
2857 * interrupt handler scheduler_tick().
2859 * 3. Wakeups don't really cause entry into schedule(). They add a
2860 * task to the run-queue and that's it.
2862 * Now, if the new task added to the run-queue preempts the current
2863 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2864 * called on the nearest possible occasion:
2866 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2868 * - in syscall or exception context, at the next outmost
2869 * preempt_enable(). (this might be as soon as the wake_up()'s
2872 * - in IRQ context, return from interrupt-handler to
2873 * preemptible context
2875 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2878 * - cond_resched() call
2879 * - explicit schedule() call
2880 * - return from syscall or exception to user-space
2881 * - return from interrupt-handler to user-space
2883 static void __sched __schedule(void)
2885 struct task_struct *prev, *next;
2886 unsigned long *switch_count;
2892 cpu = smp_processor_id();
2894 rcu_note_context_switch(cpu);
2897 schedule_debug(prev);
2899 if (sched_feat(HRTICK))
2902 raw_spin_lock_irq(&rq->lock);
2904 switch_count = &prev->nivcsw;
2905 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2906 if (unlikely(signal_pending_state(prev->state, prev))) {
2907 prev->state = TASK_RUNNING;
2909 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2913 * If a worker went to sleep, notify and ask workqueue
2914 * whether it wants to wake up a task to maintain
2917 if (prev->flags & PF_WQ_WORKER) {
2918 struct task_struct *to_wakeup;
2920 to_wakeup = wq_worker_sleeping(prev, cpu);
2922 try_to_wake_up_local(to_wakeup);
2925 switch_count = &prev->nvcsw;
2928 pre_schedule(rq, prev);
2930 if (unlikely(!rq->nr_running))
2931 idle_balance(cpu, rq);
2933 put_prev_task(rq, prev);
2934 next = pick_next_task(rq);
2935 clear_tsk_need_resched(prev);
2936 rq->skip_clock_update = 0;
2938 if (likely(prev != next)) {
2943 context_switch(rq, prev, next); /* unlocks the rq */
2945 * The context switch have flipped the stack from under us
2946 * and restored the local variables which were saved when
2947 * this task called schedule() in the past. prev == current
2948 * is still correct, but it can be moved to another cpu/rq.
2950 cpu = smp_processor_id();
2953 raw_spin_unlock_irq(&rq->lock);
2957 sched_preempt_enable_no_resched();
2962 static inline void sched_submit_work(struct task_struct *tsk)
2964 if (!tsk->state || tsk_is_pi_blocked(tsk))
2967 * If we are going to sleep and we have plugged IO queued,
2968 * make sure to submit it to avoid deadlocks.
2970 if (blk_needs_flush_plug(tsk))
2971 blk_schedule_flush_plug(tsk);
2974 asmlinkage void __sched schedule(void)
2976 struct task_struct *tsk = current;
2978 sched_submit_work(tsk);
2981 EXPORT_SYMBOL(schedule);
2983 #ifdef CONFIG_CONTEXT_TRACKING
2984 asmlinkage void __sched schedule_user(void)
2987 * If we come here after a random call to set_need_resched(),
2988 * or we have been woken up remotely but the IPI has not yet arrived,
2989 * we haven't yet exited the RCU idle mode. Do it here manually until
2990 * we find a better solution.
2999 * schedule_preempt_disabled - called with preemption disabled
3001 * Returns with preemption disabled. Note: preempt_count must be 1
3003 void __sched schedule_preempt_disabled(void)
3005 sched_preempt_enable_no_resched();
3010 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3012 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
3014 if (lock->owner != owner)
3018 * Ensure we emit the owner->on_cpu, dereference _after_ checking
3019 * lock->owner still matches owner, if that fails, owner might
3020 * point to free()d memory, if it still matches, the rcu_read_lock()
3021 * ensures the memory stays valid.
3025 return owner->on_cpu;
3029 * Look out! "owner" is an entirely speculative pointer
3030 * access and not reliable.
3032 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3034 if (!sched_feat(OWNER_SPIN))
3038 while (owner_running(lock, owner)) {
3042 arch_mutex_cpu_relax();
3047 * We break out the loop above on need_resched() and when the
3048 * owner changed, which is a sign for heavy contention. Return
3049 * success only when lock->owner is NULL.
3051 return lock->owner == NULL;
3055 #ifdef CONFIG_PREEMPT
3057 * this is the entry point to schedule() from in-kernel preemption
3058 * off of preempt_enable. Kernel preemptions off return from interrupt
3059 * occur there and call schedule directly.
3061 asmlinkage void __sched notrace preempt_schedule(void)
3063 struct thread_info *ti = current_thread_info();
3066 * If there is a non-zero preempt_count or interrupts are disabled,
3067 * we do not want to preempt the current task. Just return..
3069 if (likely(ti->preempt_count || irqs_disabled()))
3073 add_preempt_count_notrace(PREEMPT_ACTIVE);
3075 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3078 * Check again in case we missed a preemption opportunity
3079 * between schedule and now.
3082 } while (need_resched());
3084 EXPORT_SYMBOL(preempt_schedule);
3087 * this is the entry point to schedule() from kernel preemption
3088 * off of irq context.
3089 * Note, that this is called and return with irqs disabled. This will
3090 * protect us against recursive calling from irq.
3092 asmlinkage void __sched preempt_schedule_irq(void)
3094 struct thread_info *ti = current_thread_info();
3096 /* Catch callers which need to be fixed */
3097 BUG_ON(ti->preempt_count || !irqs_disabled());
3101 add_preempt_count(PREEMPT_ACTIVE);
3104 local_irq_disable();
3105 sub_preempt_count(PREEMPT_ACTIVE);
3108 * Check again in case we missed a preemption opportunity
3109 * between schedule and now.
3112 } while (need_resched());
3115 #endif /* CONFIG_PREEMPT */
3117 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3120 return try_to_wake_up(curr->private, mode, wake_flags);
3122 EXPORT_SYMBOL(default_wake_function);
3125 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3126 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3127 * number) then we wake all the non-exclusive tasks and one exclusive task.
3129 * There are circumstances in which we can try to wake a task which has already
3130 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3131 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3133 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3134 int nr_exclusive, int wake_flags, void *key)
3136 wait_queue_t *curr, *next;
3138 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3139 unsigned flags = curr->flags;
3141 if (curr->func(curr, mode, wake_flags, key) &&
3142 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3148 * __wake_up - wake up threads blocked on a waitqueue.
3150 * @mode: which threads
3151 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3152 * @key: is directly passed to the wakeup function
3154 * It may be assumed that this function implies a write memory barrier before
3155 * changing the task state if and only if any tasks are woken up.
3157 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3158 int nr_exclusive, void *key)
3160 unsigned long flags;
3162 spin_lock_irqsave(&q->lock, flags);
3163 __wake_up_common(q, mode, nr_exclusive, 0, key);
3164 spin_unlock_irqrestore(&q->lock, flags);
3166 EXPORT_SYMBOL(__wake_up);
3169 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3171 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3173 __wake_up_common(q, mode, nr, 0, NULL);
3175 EXPORT_SYMBOL_GPL(__wake_up_locked);
3177 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3179 __wake_up_common(q, mode, 1, 0, key);
3181 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3184 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3186 * @mode: which threads
3187 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3188 * @key: opaque value to be passed to wakeup targets
3190 * The sync wakeup differs that the waker knows that it will schedule
3191 * away soon, so while the target thread will be woken up, it will not
3192 * be migrated to another CPU - ie. the two threads are 'synchronized'
3193 * with each other. This can prevent needless bouncing between CPUs.
3195 * On UP it can prevent extra preemption.
3197 * It may be assumed that this function implies a write memory barrier before
3198 * changing the task state if and only if any tasks are woken up.
3200 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3201 int nr_exclusive, void *key)
3203 unsigned long flags;
3204 int wake_flags = WF_SYNC;
3209 if (unlikely(!nr_exclusive))
3212 spin_lock_irqsave(&q->lock, flags);
3213 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3214 spin_unlock_irqrestore(&q->lock, flags);
3216 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3219 * __wake_up_sync - see __wake_up_sync_key()
3221 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3223 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3225 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3228 * complete: - signals a single thread waiting on this completion
3229 * @x: holds the state of this particular completion
3231 * This will wake up a single thread waiting on this completion. Threads will be
3232 * awakened in the same order in which they were queued.
3234 * See also complete_all(), wait_for_completion() and related routines.
3236 * It may be assumed that this function implies a write memory barrier before
3237 * changing the task state if and only if any tasks are woken up.
3239 void complete(struct completion *x)
3241 unsigned long flags;
3243 spin_lock_irqsave(&x->wait.lock, flags);
3245 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3246 spin_unlock_irqrestore(&x->wait.lock, flags);
3248 EXPORT_SYMBOL(complete);
3251 * complete_all: - signals all threads waiting on this completion
3252 * @x: holds the state of this particular completion
3254 * This will wake up all threads waiting on this particular completion event.
3256 * It may be assumed that this function implies a write memory barrier before
3257 * changing the task state if and only if any tasks are woken up.
3259 void complete_all(struct completion *x)
3261 unsigned long flags;
3263 spin_lock_irqsave(&x->wait.lock, flags);
3264 x->done += UINT_MAX/2;
3265 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3266 spin_unlock_irqrestore(&x->wait.lock, flags);
3268 EXPORT_SYMBOL(complete_all);
3270 static inline long __sched
3271 do_wait_for_common(struct completion *x, long timeout, int state)
3274 DECLARE_WAITQUEUE(wait, current);
3276 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3278 if (signal_pending_state(state, current)) {
3279 timeout = -ERESTARTSYS;
3282 __set_current_state(state);
3283 spin_unlock_irq(&x->wait.lock);
3284 timeout = schedule_timeout(timeout);
3285 spin_lock_irq(&x->wait.lock);
3286 } while (!x->done && timeout);
3287 __remove_wait_queue(&x->wait, &wait);
3292 return timeout ?: 1;
3296 wait_for_common(struct completion *x, long timeout, int state)
3300 spin_lock_irq(&x->wait.lock);
3301 timeout = do_wait_for_common(x, timeout, state);
3302 spin_unlock_irq(&x->wait.lock);
3307 * wait_for_completion: - waits for completion of a task
3308 * @x: holds the state of this particular completion
3310 * This waits to be signaled for completion of a specific task. It is NOT
3311 * interruptible and there is no timeout.
3313 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3314 * and interrupt capability. Also see complete().
3316 void __sched wait_for_completion(struct completion *x)
3318 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3320 EXPORT_SYMBOL(wait_for_completion);
3323 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3324 * @x: holds the state of this particular completion
3325 * @timeout: timeout value in jiffies
3327 * This waits for either a completion of a specific task to be signaled or for a
3328 * specified timeout to expire. The timeout is in jiffies. It is not
3331 * The return value is 0 if timed out, and positive (at least 1, or number of
3332 * jiffies left till timeout) if completed.
3334 unsigned long __sched
3335 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3337 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3339 EXPORT_SYMBOL(wait_for_completion_timeout);
3342 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3343 * @x: holds the state of this particular completion
3345 * This waits for completion of a specific task to be signaled. It is
3348 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3350 int __sched wait_for_completion_interruptible(struct completion *x)
3352 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3353 if (t == -ERESTARTSYS)
3357 EXPORT_SYMBOL(wait_for_completion_interruptible);
3360 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3361 * @x: holds the state of this particular completion
3362 * @timeout: timeout value in jiffies
3364 * This waits for either a completion of a specific task to be signaled or for a
3365 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3367 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3368 * positive (at least 1, or number of jiffies left till timeout) if completed.
3371 wait_for_completion_interruptible_timeout(struct completion *x,
3372 unsigned long timeout)
3374 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3376 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3379 * wait_for_completion_killable: - waits for completion of a task (killable)
3380 * @x: holds the state of this particular completion
3382 * This waits to be signaled for completion of a specific task. It can be
3383 * interrupted by a kill signal.
3385 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3387 int __sched wait_for_completion_killable(struct completion *x)
3389 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3390 if (t == -ERESTARTSYS)
3394 EXPORT_SYMBOL(wait_for_completion_killable);
3397 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3398 * @x: holds the state of this particular completion
3399 * @timeout: timeout value in jiffies
3401 * This waits for either a completion of a specific task to be
3402 * signaled or for a specified timeout to expire. It can be
3403 * interrupted by a kill signal. The timeout is in jiffies.
3405 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3406 * positive (at least 1, or number of jiffies left till timeout) if completed.
3409 wait_for_completion_killable_timeout(struct completion *x,
3410 unsigned long timeout)
3412 return wait_for_common(x, timeout, TASK_KILLABLE);
3414 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3417 * try_wait_for_completion - try to decrement a completion without blocking
3418 * @x: completion structure
3420 * Returns: 0 if a decrement cannot be done without blocking
3421 * 1 if a decrement succeeded.
3423 * If a completion is being used as a counting completion,
3424 * attempt to decrement the counter without blocking. This
3425 * enables us to avoid waiting if the resource the completion
3426 * is protecting is not available.
3428 bool try_wait_for_completion(struct completion *x)
3430 unsigned long flags;
3433 spin_lock_irqsave(&x->wait.lock, flags);
3438 spin_unlock_irqrestore(&x->wait.lock, flags);
3441 EXPORT_SYMBOL(try_wait_for_completion);
3444 * completion_done - Test to see if a completion has any waiters
3445 * @x: completion structure
3447 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3448 * 1 if there are no waiters.
3451 bool completion_done(struct completion *x)
3453 unsigned long flags;
3456 spin_lock_irqsave(&x->wait.lock, flags);
3459 spin_unlock_irqrestore(&x->wait.lock, flags);
3462 EXPORT_SYMBOL(completion_done);
3465 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3467 unsigned long flags;
3470 init_waitqueue_entry(&wait, current);
3472 __set_current_state(state);
3474 spin_lock_irqsave(&q->lock, flags);
3475 __add_wait_queue(q, &wait);
3476 spin_unlock(&q->lock);
3477 timeout = schedule_timeout(timeout);
3478 spin_lock_irq(&q->lock);
3479 __remove_wait_queue(q, &wait);
3480 spin_unlock_irqrestore(&q->lock, flags);
3485 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3487 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3489 EXPORT_SYMBOL(interruptible_sleep_on);
3492 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3494 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3496 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3498 void __sched sleep_on(wait_queue_head_t *q)
3500 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3502 EXPORT_SYMBOL(sleep_on);
3504 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3506 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3508 EXPORT_SYMBOL(sleep_on_timeout);
3510 #ifdef CONFIG_RT_MUTEXES
3513 * rt_mutex_setprio - set the current priority of a task
3515 * @prio: prio value (kernel-internal form)
3517 * This function changes the 'effective' priority of a task. It does
3518 * not touch ->normal_prio like __setscheduler().
3520 * Used by the rt_mutex code to implement priority inheritance logic.
3522 void rt_mutex_setprio(struct task_struct *p, int prio)
3524 int oldprio, on_rq, running;
3526 const struct sched_class *prev_class;
3528 BUG_ON(prio < 0 || prio > MAX_PRIO);
3530 rq = __task_rq_lock(p);
3533 * Idle task boosting is a nono in general. There is one
3534 * exception, when PREEMPT_RT and NOHZ is active:
3536 * The idle task calls get_next_timer_interrupt() and holds
3537 * the timer wheel base->lock on the CPU and another CPU wants
3538 * to access the timer (probably to cancel it). We can safely
3539 * ignore the boosting request, as the idle CPU runs this code
3540 * with interrupts disabled and will complete the lock
3541 * protected section without being interrupted. So there is no
3542 * real need to boost.
3544 if (unlikely(p == rq->idle)) {
3545 WARN_ON(p != rq->curr);
3546 WARN_ON(p->pi_blocked_on);
3550 trace_sched_pi_setprio(p, prio);
3552 prev_class = p->sched_class;
3554 running = task_current(rq, p);
3556 dequeue_task(rq, p, 0);
3558 p->sched_class->put_prev_task(rq, p);
3561 p->sched_class = &rt_sched_class;
3563 p->sched_class = &fair_sched_class;
3568 p->sched_class->set_curr_task(rq);
3570 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3572 check_class_changed(rq, p, prev_class, oldprio);
3574 __task_rq_unlock(rq);
3577 void set_user_nice(struct task_struct *p, long nice)
3579 int old_prio, delta, on_rq;
3580 unsigned long flags;
3583 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3586 * We have to be careful, if called from sys_setpriority(),
3587 * the task might be in the middle of scheduling on another CPU.
3589 rq = task_rq_lock(p, &flags);
3591 * The RT priorities are set via sched_setscheduler(), but we still
3592 * allow the 'normal' nice value to be set - but as expected
3593 * it wont have any effect on scheduling until the task is
3594 * SCHED_FIFO/SCHED_RR:
3596 if (task_has_rt_policy(p)) {
3597 p->static_prio = NICE_TO_PRIO(nice);
3602 dequeue_task(rq, p, 0);
3604 p->static_prio = NICE_TO_PRIO(nice);
3607 p->prio = effective_prio(p);
3608 delta = p->prio - old_prio;
3611 enqueue_task(rq, p, 0);
3613 * If the task increased its priority or is running and
3614 * lowered its priority, then reschedule its CPU:
3616 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3617 resched_task(rq->curr);
3620 task_rq_unlock(rq, p, &flags);
3622 EXPORT_SYMBOL(set_user_nice);
3625 * can_nice - check if a task can reduce its nice value
3629 int can_nice(const struct task_struct *p, const int nice)
3631 /* convert nice value [19,-20] to rlimit style value [1,40] */
3632 int nice_rlim = 20 - nice;
3634 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3635 capable(CAP_SYS_NICE));
3638 #ifdef __ARCH_WANT_SYS_NICE
3641 * sys_nice - change the priority of the current process.
3642 * @increment: priority increment
3644 * sys_setpriority is a more generic, but much slower function that
3645 * does similar things.
3647 SYSCALL_DEFINE1(nice, int, increment)
3652 * Setpriority might change our priority at the same moment.
3653 * We don't have to worry. Conceptually one call occurs first
3654 * and we have a single winner.
3656 if (increment < -40)
3661 nice = TASK_NICE(current) + increment;
3667 if (increment < 0 && !can_nice(current, nice))
3670 retval = security_task_setnice(current, nice);
3674 set_user_nice(current, nice);
3681 * task_prio - return the priority value of a given task.
3682 * @p: the task in question.
3684 * This is the priority value as seen by users in /proc.
3685 * RT tasks are offset by -200. Normal tasks are centered
3686 * around 0, value goes from -16 to +15.
3688 int task_prio(const struct task_struct *p)
3690 return p->prio - MAX_RT_PRIO;
3694 * task_nice - return the nice value of a given task.
3695 * @p: the task in question.
3697 int task_nice(const struct task_struct *p)
3699 return TASK_NICE(p);
3701 EXPORT_SYMBOL(task_nice);
3704 * idle_cpu - is a given cpu idle currently?
3705 * @cpu: the processor in question.
3707 int idle_cpu(int cpu)
3709 struct rq *rq = cpu_rq(cpu);
3711 if (rq->curr != rq->idle)
3718 if (!llist_empty(&rq->wake_list))
3726 * idle_task - return the idle task for a given cpu.
3727 * @cpu: the processor in question.
3729 struct task_struct *idle_task(int cpu)
3731 return cpu_rq(cpu)->idle;
3735 * find_process_by_pid - find a process with a matching PID value.
3736 * @pid: the pid in question.
3738 static struct task_struct *find_process_by_pid(pid_t pid)
3740 return pid ? find_task_by_vpid(pid) : current;
3743 /* Actually do priority change: must hold rq lock. */
3745 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3748 p->rt_priority = prio;
3749 p->normal_prio = normal_prio(p);
3750 /* we are holding p->pi_lock already */
3751 p->prio = rt_mutex_getprio(p);
3752 if (rt_prio(p->prio))
3753 p->sched_class = &rt_sched_class;
3755 p->sched_class = &fair_sched_class;
3760 * check the target process has a UID that matches the current process's
3762 static bool check_same_owner(struct task_struct *p)
3764 const struct cred *cred = current_cred(), *pcred;
3768 pcred = __task_cred(p);
3769 match = (uid_eq(cred->euid, pcred->euid) ||
3770 uid_eq(cred->euid, pcred->uid));
3775 static int __sched_setscheduler(struct task_struct *p, int policy,
3776 const struct sched_param *param, bool user)
3778 int retval, oldprio, oldpolicy = -1, on_rq, running;
3779 unsigned long flags;
3780 const struct sched_class *prev_class;
3784 /* may grab non-irq protected spin_locks */
3785 BUG_ON(in_interrupt());
3787 /* double check policy once rq lock held */
3789 reset_on_fork = p->sched_reset_on_fork;
3790 policy = oldpolicy = p->policy;
3792 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3793 policy &= ~SCHED_RESET_ON_FORK;
3795 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3796 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3797 policy != SCHED_IDLE)
3802 * Valid priorities for SCHED_FIFO and SCHED_RR are
3803 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3804 * SCHED_BATCH and SCHED_IDLE is 0.
3806 if (param->sched_priority < 0 ||
3807 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3808 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3810 if (rt_policy(policy) != (param->sched_priority != 0))
3814 * Allow unprivileged RT tasks to decrease priority:
3816 if (user && !capable(CAP_SYS_NICE)) {
3817 if (rt_policy(policy)) {
3818 unsigned long rlim_rtprio =
3819 task_rlimit(p, RLIMIT_RTPRIO);
3821 /* can't set/change the rt policy */
3822 if (policy != p->policy && !rlim_rtprio)
3825 /* can't increase priority */
3826 if (param->sched_priority > p->rt_priority &&
3827 param->sched_priority > rlim_rtprio)
3832 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3833 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3835 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3836 if (!can_nice(p, TASK_NICE(p)))
3840 /* can't change other user's priorities */
3841 if (!check_same_owner(p))
3844 /* Normal users shall not reset the sched_reset_on_fork flag */
3845 if (p->sched_reset_on_fork && !reset_on_fork)
3850 retval = security_task_setscheduler(p);
3856 * make sure no PI-waiters arrive (or leave) while we are
3857 * changing the priority of the task:
3859 * To be able to change p->policy safely, the appropriate
3860 * runqueue lock must be held.
3862 rq = task_rq_lock(p, &flags);
3865 * Changing the policy of the stop threads its a very bad idea
3867 if (p == rq->stop) {
3868 task_rq_unlock(rq, p, &flags);
3873 * If not changing anything there's no need to proceed further:
3875 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3876 param->sched_priority == p->rt_priority))) {
3877 task_rq_unlock(rq, p, &flags);
3881 #ifdef CONFIG_RT_GROUP_SCHED
3884 * Do not allow realtime tasks into groups that have no runtime
3887 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3888 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3889 !task_group_is_autogroup(task_group(p))) {
3890 task_rq_unlock(rq, p, &flags);
3896 /* recheck policy now with rq lock held */
3897 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3898 policy = oldpolicy = -1;
3899 task_rq_unlock(rq, p, &flags);
3903 running = task_current(rq, p);
3905 dequeue_task(rq, p, 0);
3907 p->sched_class->put_prev_task(rq, p);
3909 p->sched_reset_on_fork = reset_on_fork;
3912 prev_class = p->sched_class;
3913 __setscheduler(rq, p, policy, param->sched_priority);
3916 p->sched_class->set_curr_task(rq);
3918 enqueue_task(rq, p, 0);
3920 check_class_changed(rq, p, prev_class, oldprio);
3921 task_rq_unlock(rq, p, &flags);
3923 rt_mutex_adjust_pi(p);
3929 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3930 * @p: the task in question.
3931 * @policy: new policy.
3932 * @param: structure containing the new RT priority.
3934 * NOTE that the task may be already dead.
3936 int sched_setscheduler(struct task_struct *p, int policy,
3937 const struct sched_param *param)
3939 return __sched_setscheduler(p, policy, param, true);
3941 EXPORT_SYMBOL_GPL(sched_setscheduler);
3944 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3945 * @p: the task in question.
3946 * @policy: new policy.
3947 * @param: structure containing the new RT priority.
3949 * Just like sched_setscheduler, only don't bother checking if the
3950 * current context has permission. For example, this is needed in
3951 * stop_machine(): we create temporary high priority worker threads,
3952 * but our caller might not have that capability.
3954 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3955 const struct sched_param *param)
3957 return __sched_setscheduler(p, policy, param, false);
3961 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3963 struct sched_param lparam;
3964 struct task_struct *p;
3967 if (!param || pid < 0)
3969 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3974 p = find_process_by_pid(pid);
3976 retval = sched_setscheduler(p, policy, &lparam);
3983 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3984 * @pid: the pid in question.
3985 * @policy: new policy.
3986 * @param: structure containing the new RT priority.
3988 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3989 struct sched_param __user *, param)
3991 /* negative values for policy are not valid */
3995 return do_sched_setscheduler(pid, policy, param);
3999 * sys_sched_setparam - set/change the RT priority of a thread
4000 * @pid: the pid in question.
4001 * @param: structure containing the new RT priority.
4003 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4005 return do_sched_setscheduler(pid, -1, param);
4009 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4010 * @pid: the pid in question.
4012 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4014 struct task_struct *p;
4022 p = find_process_by_pid(pid);
4024 retval = security_task_getscheduler(p);
4027 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4034 * sys_sched_getparam - get the RT priority of a thread
4035 * @pid: the pid in question.
4036 * @param: structure containing the RT priority.
4038 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4040 struct sched_param lp;
4041 struct task_struct *p;
4044 if (!param || pid < 0)
4048 p = find_process_by_pid(pid);
4053 retval = security_task_getscheduler(p);
4057 lp.sched_priority = p->rt_priority;
4061 * This one might sleep, we cannot do it with a spinlock held ...
4063 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4072 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4074 cpumask_var_t cpus_allowed, new_mask;
4075 struct task_struct *p;
4081 p = find_process_by_pid(pid);
4088 /* Prevent p going away */
4092 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4096 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4098 goto out_free_cpus_allowed;
4101 if (!check_same_owner(p)) {
4103 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4110 retval = security_task_setscheduler(p);
4114 cpuset_cpus_allowed(p, cpus_allowed);
4115 cpumask_and(new_mask, in_mask, cpus_allowed);
4117 retval = set_cpus_allowed_ptr(p, new_mask);
4120 cpuset_cpus_allowed(p, cpus_allowed);
4121 if (!cpumask_subset(new_mask, cpus_allowed)) {
4123 * We must have raced with a concurrent cpuset
4124 * update. Just reset the cpus_allowed to the
4125 * cpuset's cpus_allowed
4127 cpumask_copy(new_mask, cpus_allowed);
4132 free_cpumask_var(new_mask);
4133 out_free_cpus_allowed:
4134 free_cpumask_var(cpus_allowed);
4141 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4142 struct cpumask *new_mask)
4144 if (len < cpumask_size())
4145 cpumask_clear(new_mask);
4146 else if (len > cpumask_size())
4147 len = cpumask_size();
4149 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4153 * sys_sched_setaffinity - set the cpu affinity of a process
4154 * @pid: pid of the process
4155 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4156 * @user_mask_ptr: user-space pointer to the new cpu mask
4158 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4159 unsigned long __user *, user_mask_ptr)
4161 cpumask_var_t new_mask;
4164 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4167 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4169 retval = sched_setaffinity(pid, new_mask);
4170 free_cpumask_var(new_mask);
4174 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4176 struct task_struct *p;
4177 unsigned long flags;
4184 p = find_process_by_pid(pid);
4188 retval = security_task_getscheduler(p);
4192 raw_spin_lock_irqsave(&p->pi_lock, flags);
4193 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4194 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4204 * sys_sched_getaffinity - get the cpu affinity of a process
4205 * @pid: pid of the process
4206 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4207 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4209 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4210 unsigned long __user *, user_mask_ptr)
4215 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4217 if (len & (sizeof(unsigned long)-1))
4220 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4223 ret = sched_getaffinity(pid, mask);
4225 size_t retlen = min_t(size_t, len, cpumask_size());
4227 if (copy_to_user(user_mask_ptr, mask, retlen))
4232 free_cpumask_var(mask);
4238 * sys_sched_yield - yield the current processor to other threads.
4240 * This function yields the current CPU to other tasks. If there are no
4241 * other threads running on this CPU then this function will return.
4243 SYSCALL_DEFINE0(sched_yield)
4245 struct rq *rq = this_rq_lock();
4247 schedstat_inc(rq, yld_count);
4248 current->sched_class->yield_task(rq);
4251 * Since we are going to call schedule() anyway, there's
4252 * no need to preempt or enable interrupts:
4254 __release(rq->lock);
4255 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4256 do_raw_spin_unlock(&rq->lock);
4257 sched_preempt_enable_no_resched();
4264 static inline int should_resched(void)
4266 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4269 static void __cond_resched(void)
4271 add_preempt_count(PREEMPT_ACTIVE);
4273 sub_preempt_count(PREEMPT_ACTIVE);
4276 int __sched _cond_resched(void)
4278 if (should_resched()) {
4284 EXPORT_SYMBOL(_cond_resched);
4287 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4288 * call schedule, and on return reacquire the lock.
4290 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4291 * operations here to prevent schedule() from being called twice (once via
4292 * spin_unlock(), once by hand).
4294 int __cond_resched_lock(spinlock_t *lock)
4296 int resched = should_resched();
4299 lockdep_assert_held(lock);
4301 if (spin_needbreak(lock) || resched) {
4312 EXPORT_SYMBOL(__cond_resched_lock);
4314 int __sched __cond_resched_softirq(void)
4316 BUG_ON(!in_softirq());
4318 if (should_resched()) {
4326 EXPORT_SYMBOL(__cond_resched_softirq);
4329 * yield - yield the current processor to other threads.
4331 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4333 * The scheduler is at all times free to pick the calling task as the most
4334 * eligible task to run, if removing the yield() call from your code breaks
4335 * it, its already broken.
4337 * Typical broken usage is:
4342 * where one assumes that yield() will let 'the other' process run that will
4343 * make event true. If the current task is a SCHED_FIFO task that will never
4344 * happen. Never use yield() as a progress guarantee!!
4346 * If you want to use yield() to wait for something, use wait_event().
4347 * If you want to use yield() to be 'nice' for others, use cond_resched().
4348 * If you still want to use yield(), do not!
4350 void __sched yield(void)
4352 set_current_state(TASK_RUNNING);
4355 EXPORT_SYMBOL(yield);
4358 * yield_to - yield the current processor to another thread in
4359 * your thread group, or accelerate that thread toward the
4360 * processor it's on.
4362 * @preempt: whether task preemption is allowed or not
4364 * It's the caller's job to ensure that the target task struct
4365 * can't go away on us before we can do any checks.
4367 * Returns true if we indeed boosted the target task.
4369 bool __sched yield_to(struct task_struct *p, bool preempt)
4371 struct task_struct *curr = current;
4372 struct rq *rq, *p_rq;
4373 unsigned long flags;
4376 local_irq_save(flags);
4381 double_rq_lock(rq, p_rq);
4382 while (task_rq(p) != p_rq) {
4383 double_rq_unlock(rq, p_rq);
4387 if (!curr->sched_class->yield_to_task)
4390 if (curr->sched_class != p->sched_class)
4393 if (task_running(p_rq, p) || p->state)
4396 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4398 schedstat_inc(rq, yld_count);
4400 * Make p's CPU reschedule; pick_next_entity takes care of
4403 if (preempt && rq != p_rq)
4404 resched_task(p_rq->curr);
4408 double_rq_unlock(rq, p_rq);
4409 local_irq_restore(flags);
4416 EXPORT_SYMBOL_GPL(yield_to);
4419 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4420 * that process accounting knows that this is a task in IO wait state.
4422 void __sched io_schedule(void)
4424 struct rq *rq = raw_rq();
4426 delayacct_blkio_start();
4427 atomic_inc(&rq->nr_iowait);
4428 blk_flush_plug(current);
4429 current->in_iowait = 1;
4431 current->in_iowait = 0;
4432 atomic_dec(&rq->nr_iowait);
4433 delayacct_blkio_end();
4435 EXPORT_SYMBOL(io_schedule);
4437 long __sched io_schedule_timeout(long timeout)
4439 struct rq *rq = raw_rq();
4442 delayacct_blkio_start();
4443 atomic_inc(&rq->nr_iowait);
4444 blk_flush_plug(current);
4445 current->in_iowait = 1;
4446 ret = schedule_timeout(timeout);
4447 current->in_iowait = 0;
4448 atomic_dec(&rq->nr_iowait);
4449 delayacct_blkio_end();
4454 * sys_sched_get_priority_max - return maximum RT priority.
4455 * @policy: scheduling class.
4457 * this syscall returns the maximum rt_priority that can be used
4458 * by a given scheduling class.
4460 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4467 ret = MAX_USER_RT_PRIO-1;
4479 * sys_sched_get_priority_min - return minimum RT priority.
4480 * @policy: scheduling class.
4482 * this syscall returns the minimum rt_priority that can be used
4483 * by a given scheduling class.
4485 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4503 * sys_sched_rr_get_interval - return the default timeslice of a process.
4504 * @pid: pid of the process.
4505 * @interval: userspace pointer to the timeslice value.
4507 * this syscall writes the default timeslice value of a given process
4508 * into the user-space timespec buffer. A value of '0' means infinity.
4510 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4511 struct timespec __user *, interval)
4513 struct task_struct *p;
4514 unsigned int time_slice;
4515 unsigned long flags;
4525 p = find_process_by_pid(pid);
4529 retval = security_task_getscheduler(p);
4533 rq = task_rq_lock(p, &flags);
4534 time_slice = p->sched_class->get_rr_interval(rq, p);
4535 task_rq_unlock(rq, p, &flags);
4538 jiffies_to_timespec(time_slice, &t);
4539 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4547 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4549 void sched_show_task(struct task_struct *p)
4551 unsigned long free = 0;
4555 state = p->state ? __ffs(p->state) + 1 : 0;
4556 printk(KERN_INFO "%-15.15s %c", p->comm,
4557 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4558 #if BITS_PER_LONG == 32
4559 if (state == TASK_RUNNING)
4560 printk(KERN_CONT " running ");
4562 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4564 if (state == TASK_RUNNING)
4565 printk(KERN_CONT " running task ");
4567 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4569 #ifdef CONFIG_DEBUG_STACK_USAGE
4570 free = stack_not_used(p);
4573 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4575 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4576 task_pid_nr(p), ppid,
4577 (unsigned long)task_thread_info(p)->flags);
4579 show_stack(p, NULL);
4582 void show_state_filter(unsigned long state_filter)
4584 struct task_struct *g, *p;
4586 #if BITS_PER_LONG == 32
4588 " task PC stack pid father\n");
4591 " task PC stack pid father\n");
4594 do_each_thread(g, p) {
4596 * reset the NMI-timeout, listing all files on a slow
4597 * console might take a lot of time:
4599 touch_nmi_watchdog();
4600 if (!state_filter || (p->state & state_filter))
4602 } while_each_thread(g, p);
4604 touch_all_softlockup_watchdogs();
4606 #ifdef CONFIG_SCHED_DEBUG
4607 sysrq_sched_debug_show();
4611 * Only show locks if all tasks are dumped:
4614 debug_show_all_locks();
4617 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4619 idle->sched_class = &idle_sched_class;
4623 * init_idle - set up an idle thread for a given CPU
4624 * @idle: task in question
4625 * @cpu: cpu the idle task belongs to
4627 * NOTE: this function does not set the idle thread's NEED_RESCHED
4628 * flag, to make booting more robust.
4630 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4632 struct rq *rq = cpu_rq(cpu);
4633 unsigned long flags;
4635 raw_spin_lock_irqsave(&rq->lock, flags);
4638 idle->state = TASK_RUNNING;
4639 idle->se.exec_start = sched_clock();
4641 do_set_cpus_allowed(idle, cpumask_of(cpu));
4643 * We're having a chicken and egg problem, even though we are
4644 * holding rq->lock, the cpu isn't yet set to this cpu so the
4645 * lockdep check in task_group() will fail.
4647 * Similar case to sched_fork(). / Alternatively we could
4648 * use task_rq_lock() here and obtain the other rq->lock.
4653 __set_task_cpu(idle, cpu);
4656 rq->curr = rq->idle = idle;
4657 #if defined(CONFIG_SMP)
4660 raw_spin_unlock_irqrestore(&rq->lock, flags);
4662 /* Set the preempt count _outside_ the spinlocks! */
4663 task_thread_info(idle)->preempt_count = 0;
4666 * The idle tasks have their own, simple scheduling class:
4668 idle->sched_class = &idle_sched_class;
4669 ftrace_graph_init_idle_task(idle, cpu);
4670 #if defined(CONFIG_SMP)
4671 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4676 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4678 if (p->sched_class && p->sched_class->set_cpus_allowed)
4679 p->sched_class->set_cpus_allowed(p, new_mask);
4681 cpumask_copy(&p->cpus_allowed, new_mask);
4682 p->nr_cpus_allowed = cpumask_weight(new_mask);
4686 * This is how migration works:
4688 * 1) we invoke migration_cpu_stop() on the target CPU using
4690 * 2) stopper starts to run (implicitly forcing the migrated thread
4692 * 3) it checks whether the migrated task is still in the wrong runqueue.
4693 * 4) if it's in the wrong runqueue then the migration thread removes
4694 * it and puts it into the right queue.
4695 * 5) stopper completes and stop_one_cpu() returns and the migration
4700 * Change a given task's CPU affinity. Migrate the thread to a
4701 * proper CPU and schedule it away if the CPU it's executing on
4702 * is removed from the allowed bitmask.
4704 * NOTE: the caller must have a valid reference to the task, the
4705 * task must not exit() & deallocate itself prematurely. The
4706 * call is not atomic; no spinlocks may be held.
4708 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4710 unsigned long flags;
4712 unsigned int dest_cpu;
4715 rq = task_rq_lock(p, &flags);
4717 if (cpumask_equal(&p->cpus_allowed, new_mask))
4720 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4725 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4730 do_set_cpus_allowed(p, new_mask);
4732 /* Can the task run on the task's current CPU? If so, we're done */
4733 if (cpumask_test_cpu(task_cpu(p), new_mask))
4736 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4738 struct migration_arg arg = { p, dest_cpu };
4739 /* Need help from migration thread: drop lock and wait. */
4740 task_rq_unlock(rq, p, &flags);
4741 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4742 tlb_migrate_finish(p->mm);
4746 task_rq_unlock(rq, p, &flags);
4750 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4753 * Move (not current) task off this cpu, onto dest cpu. We're doing
4754 * this because either it can't run here any more (set_cpus_allowed()
4755 * away from this CPU, or CPU going down), or because we're
4756 * attempting to rebalance this task on exec (sched_exec).
4758 * So we race with normal scheduler movements, but that's OK, as long
4759 * as the task is no longer on this CPU.
4761 * Returns non-zero if task was successfully migrated.
4763 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4765 struct rq *rq_dest, *rq_src;
4768 if (unlikely(!cpu_active(dest_cpu)))
4771 rq_src = cpu_rq(src_cpu);
4772 rq_dest = cpu_rq(dest_cpu);
4774 raw_spin_lock(&p->pi_lock);
4775 double_rq_lock(rq_src, rq_dest);
4776 /* Already moved. */
4777 if (task_cpu(p) != src_cpu)
4779 /* Affinity changed (again). */
4780 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4784 * If we're not on a rq, the next wake-up will ensure we're
4788 dequeue_task(rq_src, p, 0);
4789 set_task_cpu(p, dest_cpu);
4790 enqueue_task(rq_dest, p, 0);
4791 check_preempt_curr(rq_dest, p, 0);
4796 double_rq_unlock(rq_src, rq_dest);
4797 raw_spin_unlock(&p->pi_lock);
4802 * migration_cpu_stop - this will be executed by a highprio stopper thread
4803 * and performs thread migration by bumping thread off CPU then
4804 * 'pushing' onto another runqueue.
4806 static int migration_cpu_stop(void *data)
4808 struct migration_arg *arg = data;
4811 * The original target cpu might have gone down and we might
4812 * be on another cpu but it doesn't matter.
4814 local_irq_disable();
4815 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4820 #ifdef CONFIG_HOTPLUG_CPU
4823 * Ensures that the idle task is using init_mm right before its cpu goes
4826 void idle_task_exit(void)
4828 struct mm_struct *mm = current->active_mm;
4830 BUG_ON(cpu_online(smp_processor_id()));
4833 switch_mm(mm, &init_mm, current);
4838 * Since this CPU is going 'away' for a while, fold any nr_active delta
4839 * we might have. Assumes we're called after migrate_tasks() so that the
4840 * nr_active count is stable.
4842 * Also see the comment "Global load-average calculations".
4844 static void calc_load_migrate(struct rq *rq)
4846 long delta = calc_load_fold_active(rq);
4848 atomic_long_add(delta, &calc_load_tasks);
4852 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4853 * try_to_wake_up()->select_task_rq().
4855 * Called with rq->lock held even though we'er in stop_machine() and
4856 * there's no concurrency possible, we hold the required locks anyway
4857 * because of lock validation efforts.
4859 static void migrate_tasks(unsigned int dead_cpu)
4861 struct rq *rq = cpu_rq(dead_cpu);
4862 struct task_struct *next, *stop = rq->stop;
4866 * Fudge the rq selection such that the below task selection loop
4867 * doesn't get stuck on the currently eligible stop task.
4869 * We're currently inside stop_machine() and the rq is either stuck
4870 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4871 * either way we should never end up calling schedule() until we're
4878 * There's this thread running, bail when that's the only
4881 if (rq->nr_running == 1)
4884 next = pick_next_task(rq);
4886 next->sched_class->put_prev_task(rq, next);
4888 /* Find suitable destination for @next, with force if needed. */
4889 dest_cpu = select_fallback_rq(dead_cpu, next);
4890 raw_spin_unlock(&rq->lock);
4892 __migrate_task(next, dead_cpu, dest_cpu);
4894 raw_spin_lock(&rq->lock);
4900 #endif /* CONFIG_HOTPLUG_CPU */
4902 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4904 static struct ctl_table sd_ctl_dir[] = {
4906 .procname = "sched_domain",
4912 static struct ctl_table sd_ctl_root[] = {
4914 .procname = "kernel",
4916 .child = sd_ctl_dir,
4921 static struct ctl_table *sd_alloc_ctl_entry(int n)
4923 struct ctl_table *entry =
4924 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4929 static void sd_free_ctl_entry(struct ctl_table **tablep)
4931 struct ctl_table *entry;
4934 * In the intermediate directories, both the child directory and
4935 * procname are dynamically allocated and could fail but the mode
4936 * will always be set. In the lowest directory the names are
4937 * static strings and all have proc handlers.
4939 for (entry = *tablep; entry->mode; entry++) {
4941 sd_free_ctl_entry(&entry->child);
4942 if (entry->proc_handler == NULL)
4943 kfree(entry->procname);
4950 static int min_load_idx = 0;
4951 static int max_load_idx = CPU_LOAD_IDX_MAX;
4954 set_table_entry(struct ctl_table *entry,
4955 const char *procname, void *data, int maxlen,
4956 umode_t mode, proc_handler *proc_handler,
4959 entry->procname = procname;
4961 entry->maxlen = maxlen;
4963 entry->proc_handler = proc_handler;
4966 entry->extra1 = &min_load_idx;
4967 entry->extra2 = &max_load_idx;
4971 static struct ctl_table *
4972 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4974 struct ctl_table *table = sd_alloc_ctl_entry(13);
4979 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4980 sizeof(long), 0644, proc_doulongvec_minmax, false);
4981 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4982 sizeof(long), 0644, proc_doulongvec_minmax, false);
4983 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4984 sizeof(int), 0644, proc_dointvec_minmax, true);
4985 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4986 sizeof(int), 0644, proc_dointvec_minmax, true);
4987 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4988 sizeof(int), 0644, proc_dointvec_minmax, true);
4989 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4990 sizeof(int), 0644, proc_dointvec_minmax, true);
4991 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4992 sizeof(int), 0644, proc_dointvec_minmax, true);
4993 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4994 sizeof(int), 0644, proc_dointvec_minmax, false);
4995 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4996 sizeof(int), 0644, proc_dointvec_minmax, false);
4997 set_table_entry(&table[9], "cache_nice_tries",
4998 &sd->cache_nice_tries,
4999 sizeof(int), 0644, proc_dointvec_minmax, false);
5000 set_table_entry(&table[10], "flags", &sd->flags,
5001 sizeof(int), 0644, proc_dointvec_minmax, false);
5002 set_table_entry(&table[11], "name", sd->name,
5003 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5004 /* &table[12] is terminator */
5009 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5011 struct ctl_table *entry, *table;
5012 struct sched_domain *sd;
5013 int domain_num = 0, i;
5016 for_each_domain(cpu, sd)
5018 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5023 for_each_domain(cpu, sd) {
5024 snprintf(buf, 32, "domain%d", i);
5025 entry->procname = kstrdup(buf, GFP_KERNEL);
5027 entry->child = sd_alloc_ctl_domain_table(sd);
5034 static struct ctl_table_header *sd_sysctl_header;
5035 static void register_sched_domain_sysctl(void)
5037 int i, cpu_num = num_possible_cpus();
5038 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5041 WARN_ON(sd_ctl_dir[0].child);
5042 sd_ctl_dir[0].child = entry;
5047 for_each_possible_cpu(i) {
5048 snprintf(buf, 32, "cpu%d", i);
5049 entry->procname = kstrdup(buf, GFP_KERNEL);
5051 entry->child = sd_alloc_ctl_cpu_table(i);
5055 WARN_ON(sd_sysctl_header);
5056 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5059 /* may be called multiple times per register */
5060 static void unregister_sched_domain_sysctl(void)
5062 if (sd_sysctl_header)
5063 unregister_sysctl_table(sd_sysctl_header);
5064 sd_sysctl_header = NULL;
5065 if (sd_ctl_dir[0].child)
5066 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5069 static void register_sched_domain_sysctl(void)
5072 static void unregister_sched_domain_sysctl(void)
5077 static void set_rq_online(struct rq *rq)
5080 const struct sched_class *class;
5082 cpumask_set_cpu(rq->cpu, rq->rd->online);
5085 for_each_class(class) {
5086 if (class->rq_online)
5087 class->rq_online(rq);
5092 static void set_rq_offline(struct rq *rq)
5095 const struct sched_class *class;
5097 for_each_class(class) {
5098 if (class->rq_offline)
5099 class->rq_offline(rq);
5102 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5108 * migration_call - callback that gets triggered when a CPU is added.
5109 * Here we can start up the necessary migration thread for the new CPU.
5111 static int __cpuinit
5112 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5114 int cpu = (long)hcpu;
5115 unsigned long flags;
5116 struct rq *rq = cpu_rq(cpu);
5118 switch (action & ~CPU_TASKS_FROZEN) {
5120 case CPU_UP_PREPARE:
5121 rq->calc_load_update = calc_load_update;
5125 /* Update our root-domain */
5126 raw_spin_lock_irqsave(&rq->lock, flags);
5128 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5132 raw_spin_unlock_irqrestore(&rq->lock, flags);
5135 #ifdef CONFIG_HOTPLUG_CPU
5137 sched_ttwu_pending();
5138 /* Update our root-domain */
5139 raw_spin_lock_irqsave(&rq->lock, flags);
5141 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5145 BUG_ON(rq->nr_running != 1); /* the migration thread */
5146 raw_spin_unlock_irqrestore(&rq->lock, flags);
5150 calc_load_migrate(rq);
5155 update_max_interval();
5161 * Register at high priority so that task migration (migrate_all_tasks)
5162 * happens before everything else. This has to be lower priority than
5163 * the notifier in the perf_event subsystem, though.
5165 static struct notifier_block __cpuinitdata migration_notifier = {
5166 .notifier_call = migration_call,
5167 .priority = CPU_PRI_MIGRATION,
5170 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5171 unsigned long action, void *hcpu)
5173 switch (action & ~CPU_TASKS_FROZEN) {
5175 case CPU_DOWN_FAILED:
5176 set_cpu_active((long)hcpu, true);
5183 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5184 unsigned long action, void *hcpu)
5186 switch (action & ~CPU_TASKS_FROZEN) {
5187 case CPU_DOWN_PREPARE:
5188 set_cpu_active((long)hcpu, false);
5195 static int __init migration_init(void)
5197 void *cpu = (void *)(long)smp_processor_id();
5200 /* Initialize migration for the boot CPU */
5201 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5202 BUG_ON(err == NOTIFY_BAD);
5203 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5204 register_cpu_notifier(&migration_notifier);
5206 /* Register cpu active notifiers */
5207 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5208 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5212 early_initcall(migration_init);
5217 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5219 #ifdef CONFIG_SCHED_DEBUG
5221 static __read_mostly int sched_debug_enabled;
5223 static int __init sched_debug_setup(char *str)
5225 sched_debug_enabled = 1;
5229 early_param("sched_debug", sched_debug_setup);
5231 static inline bool sched_debug(void)
5233 return sched_debug_enabled;
5236 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5237 struct cpumask *groupmask)
5239 struct sched_group *group = sd->groups;
5242 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5243 cpumask_clear(groupmask);
5245 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5247 if (!(sd->flags & SD_LOAD_BALANCE)) {
5248 printk("does not load-balance\n");
5250 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5255 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5257 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5258 printk(KERN_ERR "ERROR: domain->span does not contain "
5261 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5262 printk(KERN_ERR "ERROR: domain->groups does not contain"
5266 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5270 printk(KERN_ERR "ERROR: group is NULL\n");
5275 * Even though we initialize ->power to something semi-sane,
5276 * we leave power_orig unset. This allows us to detect if
5277 * domain iteration is still funny without causing /0 traps.
5279 if (!group->sgp->power_orig) {
5280 printk(KERN_CONT "\n");
5281 printk(KERN_ERR "ERROR: domain->cpu_power not "
5286 if (!cpumask_weight(sched_group_cpus(group))) {
5287 printk(KERN_CONT "\n");
5288 printk(KERN_ERR "ERROR: empty group\n");
5292 if (!(sd->flags & SD_OVERLAP) &&
5293 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5294 printk(KERN_CONT "\n");
5295 printk(KERN_ERR "ERROR: repeated CPUs\n");
5299 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5301 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5303 printk(KERN_CONT " %s", str);
5304 if (group->sgp->power != SCHED_POWER_SCALE) {
5305 printk(KERN_CONT " (cpu_power = %d)",
5309 group = group->next;
5310 } while (group != sd->groups);
5311 printk(KERN_CONT "\n");
5313 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5314 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5317 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5318 printk(KERN_ERR "ERROR: parent span is not a superset "
5319 "of domain->span\n");
5323 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5327 if (!sched_debug_enabled)
5331 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5335 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5338 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5346 #else /* !CONFIG_SCHED_DEBUG */
5347 # define sched_domain_debug(sd, cpu) do { } while (0)
5348 static inline bool sched_debug(void)
5352 #endif /* CONFIG_SCHED_DEBUG */
5354 static int sd_degenerate(struct sched_domain *sd)
5356 if (cpumask_weight(sched_domain_span(sd)) == 1)
5359 /* Following flags need at least 2 groups */
5360 if (sd->flags & (SD_LOAD_BALANCE |
5361 SD_BALANCE_NEWIDLE |
5365 SD_SHARE_PKG_RESOURCES)) {
5366 if (sd->groups != sd->groups->next)
5370 /* Following flags don't use groups */
5371 if (sd->flags & (SD_WAKE_AFFINE))
5378 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5380 unsigned long cflags = sd->flags, pflags = parent->flags;
5382 if (sd_degenerate(parent))
5385 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5388 /* Flags needing groups don't count if only 1 group in parent */
5389 if (parent->groups == parent->groups->next) {
5390 pflags &= ~(SD_LOAD_BALANCE |
5391 SD_BALANCE_NEWIDLE |
5395 SD_SHARE_PKG_RESOURCES);
5396 if (nr_node_ids == 1)
5397 pflags &= ~SD_SERIALIZE;
5399 if (~cflags & pflags)
5405 static void free_rootdomain(struct rcu_head *rcu)
5407 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5409 cpupri_cleanup(&rd->cpupri);
5410 free_cpumask_var(rd->rto_mask);
5411 free_cpumask_var(rd->online);
5412 free_cpumask_var(rd->span);
5416 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5418 struct root_domain *old_rd = NULL;
5419 unsigned long flags;
5421 raw_spin_lock_irqsave(&rq->lock, flags);
5426 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5429 cpumask_clear_cpu(rq->cpu, old_rd->span);
5432 * If we dont want to free the old_rt yet then
5433 * set old_rd to NULL to skip the freeing later
5436 if (!atomic_dec_and_test(&old_rd->refcount))
5440 atomic_inc(&rd->refcount);
5443 cpumask_set_cpu(rq->cpu, rd->span);
5444 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5447 raw_spin_unlock_irqrestore(&rq->lock, flags);
5450 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5453 static int init_rootdomain(struct root_domain *rd)
5455 memset(rd, 0, sizeof(*rd));
5457 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5459 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5461 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5464 if (cpupri_init(&rd->cpupri) != 0)
5469 free_cpumask_var(rd->rto_mask);
5471 free_cpumask_var(rd->online);
5473 free_cpumask_var(rd->span);
5479 * By default the system creates a single root-domain with all cpus as
5480 * members (mimicking the global state we have today).
5482 struct root_domain def_root_domain;
5484 static void init_defrootdomain(void)
5486 init_rootdomain(&def_root_domain);
5488 atomic_set(&def_root_domain.refcount, 1);
5491 static struct root_domain *alloc_rootdomain(void)
5493 struct root_domain *rd;
5495 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5499 if (init_rootdomain(rd) != 0) {
5507 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5509 struct sched_group *tmp, *first;
5518 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5523 } while (sg != first);
5526 static void free_sched_domain(struct rcu_head *rcu)
5528 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5531 * If its an overlapping domain it has private groups, iterate and
5534 if (sd->flags & SD_OVERLAP) {
5535 free_sched_groups(sd->groups, 1);
5536 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5537 kfree(sd->groups->sgp);
5543 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5545 call_rcu(&sd->rcu, free_sched_domain);
5548 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5550 for (; sd; sd = sd->parent)
5551 destroy_sched_domain(sd, cpu);
5555 * Keep a special pointer to the highest sched_domain that has
5556 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5557 * allows us to avoid some pointer chasing select_idle_sibling().
5559 * Also keep a unique ID per domain (we use the first cpu number in
5560 * the cpumask of the domain), this allows us to quickly tell if
5561 * two cpus are in the same cache domain, see cpus_share_cache().
5563 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5564 DEFINE_PER_CPU(int, sd_llc_id);
5566 static void update_top_cache_domain(int cpu)
5568 struct sched_domain *sd;
5571 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5573 id = cpumask_first(sched_domain_span(sd));
5575 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5576 per_cpu(sd_llc_id, cpu) = id;
5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5581 * hold the hotplug lock.
5584 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5586 struct rq *rq = cpu_rq(cpu);
5587 struct sched_domain *tmp;
5589 /* Remove the sched domains which do not contribute to scheduling. */
5590 for (tmp = sd; tmp; ) {
5591 struct sched_domain *parent = tmp->parent;
5595 if (sd_parent_degenerate(tmp, parent)) {
5596 tmp->parent = parent->parent;
5598 parent->parent->child = tmp;
5599 destroy_sched_domain(parent, cpu);
5604 if (sd && sd_degenerate(sd)) {
5607 destroy_sched_domain(tmp, cpu);
5612 sched_domain_debug(sd, cpu);
5614 rq_attach_root(rq, rd);
5616 rcu_assign_pointer(rq->sd, sd);
5617 destroy_sched_domains(tmp, cpu);
5619 update_top_cache_domain(cpu);
5622 /* cpus with isolated domains */
5623 static cpumask_var_t cpu_isolated_map;
5625 /* Setup the mask of cpus configured for isolated domains */
5626 static int __init isolated_cpu_setup(char *str)
5628 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5629 cpulist_parse(str, cpu_isolated_map);
5633 __setup("isolcpus=", isolated_cpu_setup);
5635 static const struct cpumask *cpu_cpu_mask(int cpu)
5637 return cpumask_of_node(cpu_to_node(cpu));
5641 struct sched_domain **__percpu sd;
5642 struct sched_group **__percpu sg;
5643 struct sched_group_power **__percpu sgp;
5647 struct sched_domain ** __percpu sd;
5648 struct root_domain *rd;
5658 struct sched_domain_topology_level;
5660 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5661 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5663 #define SDTL_OVERLAP 0x01
5665 struct sched_domain_topology_level {
5666 sched_domain_init_f init;
5667 sched_domain_mask_f mask;
5670 struct sd_data data;
5674 * Build an iteration mask that can exclude certain CPUs from the upwards
5677 * Asymmetric node setups can result in situations where the domain tree is of
5678 * unequal depth, make sure to skip domains that already cover the entire
5681 * In that case build_sched_domains() will have terminated the iteration early
5682 * and our sibling sd spans will be empty. Domains should always include the
5683 * cpu they're built on, so check that.
5686 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5688 const struct cpumask *span = sched_domain_span(sd);
5689 struct sd_data *sdd = sd->private;
5690 struct sched_domain *sibling;
5693 for_each_cpu(i, span) {
5694 sibling = *per_cpu_ptr(sdd->sd, i);
5695 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5698 cpumask_set_cpu(i, sched_group_mask(sg));
5703 * Return the canonical balance cpu for this group, this is the first cpu
5704 * of this group that's also in the iteration mask.
5706 int group_balance_cpu(struct sched_group *sg)
5708 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5712 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5714 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5715 const struct cpumask *span = sched_domain_span(sd);
5716 struct cpumask *covered = sched_domains_tmpmask;
5717 struct sd_data *sdd = sd->private;
5718 struct sched_domain *child;
5721 cpumask_clear(covered);
5723 for_each_cpu(i, span) {
5724 struct cpumask *sg_span;
5726 if (cpumask_test_cpu(i, covered))
5729 child = *per_cpu_ptr(sdd->sd, i);
5731 /* See the comment near build_group_mask(). */
5732 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5735 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5736 GFP_KERNEL, cpu_to_node(cpu));
5741 sg_span = sched_group_cpus(sg);
5743 child = child->child;
5744 cpumask_copy(sg_span, sched_domain_span(child));
5746 cpumask_set_cpu(i, sg_span);
5748 cpumask_or(covered, covered, sg_span);
5750 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5751 if (atomic_inc_return(&sg->sgp->ref) == 1)
5752 build_group_mask(sd, sg);
5755 * Initialize sgp->power such that even if we mess up the
5756 * domains and no possible iteration will get us here, we won't
5759 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5762 * Make sure the first group of this domain contains the
5763 * canonical balance cpu. Otherwise the sched_domain iteration
5764 * breaks. See update_sg_lb_stats().
5766 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5767 group_balance_cpu(sg) == cpu)
5777 sd->groups = groups;
5782 free_sched_groups(first, 0);
5787 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5789 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5790 struct sched_domain *child = sd->child;
5793 cpu = cpumask_first(sched_domain_span(child));
5796 *sg = *per_cpu_ptr(sdd->sg, cpu);
5797 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5798 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5805 * build_sched_groups will build a circular linked list of the groups
5806 * covered by the given span, and will set each group's ->cpumask correctly,
5807 * and ->cpu_power to 0.
5809 * Assumes the sched_domain tree is fully constructed
5812 build_sched_groups(struct sched_domain *sd, int cpu)
5814 struct sched_group *first = NULL, *last = NULL;
5815 struct sd_data *sdd = sd->private;
5816 const struct cpumask *span = sched_domain_span(sd);
5817 struct cpumask *covered;
5820 get_group(cpu, sdd, &sd->groups);
5821 atomic_inc(&sd->groups->ref);
5823 if (cpu != cpumask_first(sched_domain_span(sd)))
5826 lockdep_assert_held(&sched_domains_mutex);
5827 covered = sched_domains_tmpmask;
5829 cpumask_clear(covered);
5831 for_each_cpu(i, span) {
5832 struct sched_group *sg;
5833 int group = get_group(i, sdd, &sg);
5836 if (cpumask_test_cpu(i, covered))
5839 cpumask_clear(sched_group_cpus(sg));
5841 cpumask_setall(sched_group_mask(sg));
5843 for_each_cpu(j, span) {
5844 if (get_group(j, sdd, NULL) != group)
5847 cpumask_set_cpu(j, covered);
5848 cpumask_set_cpu(j, sched_group_cpus(sg));
5863 * Initialize sched groups cpu_power.
5865 * cpu_power indicates the capacity of sched group, which is used while
5866 * distributing the load between different sched groups in a sched domain.
5867 * Typically cpu_power for all the groups in a sched domain will be same unless
5868 * there are asymmetries in the topology. If there are asymmetries, group
5869 * having more cpu_power will pickup more load compared to the group having
5872 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5874 struct sched_group *sg = sd->groups;
5876 WARN_ON(!sd || !sg);
5879 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5881 } while (sg != sd->groups);
5883 if (cpu != group_balance_cpu(sg))
5886 update_group_power(sd, cpu);
5887 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5890 int __weak arch_sd_sibling_asym_packing(void)
5892 return 0*SD_ASYM_PACKING;
5896 * Initializers for schedule domains
5897 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5900 #ifdef CONFIG_SCHED_DEBUG
5901 # define SD_INIT_NAME(sd, type) sd->name = #type
5903 # define SD_INIT_NAME(sd, type) do { } while (0)
5906 #define SD_INIT_FUNC(type) \
5907 static noinline struct sched_domain * \
5908 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5910 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5911 *sd = SD_##type##_INIT; \
5912 SD_INIT_NAME(sd, type); \
5913 sd->private = &tl->data; \
5918 #ifdef CONFIG_SCHED_SMT
5919 SD_INIT_FUNC(SIBLING)
5921 #ifdef CONFIG_SCHED_MC
5924 #ifdef CONFIG_SCHED_BOOK
5928 static int default_relax_domain_level = -1;
5929 int sched_domain_level_max;
5931 static int __init setup_relax_domain_level(char *str)
5933 if (kstrtoint(str, 0, &default_relax_domain_level))
5934 pr_warn("Unable to set relax_domain_level\n");
5938 __setup("relax_domain_level=", setup_relax_domain_level);
5940 static void set_domain_attribute(struct sched_domain *sd,
5941 struct sched_domain_attr *attr)
5945 if (!attr || attr->relax_domain_level < 0) {
5946 if (default_relax_domain_level < 0)
5949 request = default_relax_domain_level;
5951 request = attr->relax_domain_level;
5952 if (request < sd->level) {
5953 /* turn off idle balance on this domain */
5954 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5956 /* turn on idle balance on this domain */
5957 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5961 static void __sdt_free(const struct cpumask *cpu_map);
5962 static int __sdt_alloc(const struct cpumask *cpu_map);
5964 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5965 const struct cpumask *cpu_map)
5969 if (!atomic_read(&d->rd->refcount))
5970 free_rootdomain(&d->rd->rcu); /* fall through */
5972 free_percpu(d->sd); /* fall through */
5974 __sdt_free(cpu_map); /* fall through */
5980 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5981 const struct cpumask *cpu_map)
5983 memset(d, 0, sizeof(*d));
5985 if (__sdt_alloc(cpu_map))
5986 return sa_sd_storage;
5987 d->sd = alloc_percpu(struct sched_domain *);
5989 return sa_sd_storage;
5990 d->rd = alloc_rootdomain();
5993 return sa_rootdomain;
5997 * NULL the sd_data elements we've used to build the sched_domain and
5998 * sched_group structure so that the subsequent __free_domain_allocs()
5999 * will not free the data we're using.
6001 static void claim_allocations(int cpu, struct sched_domain *sd)
6003 struct sd_data *sdd = sd->private;
6005 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6006 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6008 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6009 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6011 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6012 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6015 #ifdef CONFIG_SCHED_SMT
6016 static const struct cpumask *cpu_smt_mask(int cpu)
6018 return topology_thread_cpumask(cpu);
6023 * Topology list, bottom-up.
6025 static struct sched_domain_topology_level default_topology[] = {
6026 #ifdef CONFIG_SCHED_SMT
6027 { sd_init_SIBLING, cpu_smt_mask, },
6029 #ifdef CONFIG_SCHED_MC
6030 { sd_init_MC, cpu_coregroup_mask, },
6032 #ifdef CONFIG_SCHED_BOOK
6033 { sd_init_BOOK, cpu_book_mask, },
6035 { sd_init_CPU, cpu_cpu_mask, },
6039 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6043 static int sched_domains_numa_levels;
6044 static int *sched_domains_numa_distance;
6045 static struct cpumask ***sched_domains_numa_masks;
6046 static int sched_domains_curr_level;
6048 static inline int sd_local_flags(int level)
6050 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6053 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6056 static struct sched_domain *
6057 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6059 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6060 int level = tl->numa_level;
6061 int sd_weight = cpumask_weight(
6062 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6064 *sd = (struct sched_domain){
6065 .min_interval = sd_weight,
6066 .max_interval = 2*sd_weight,
6068 .imbalance_pct = 125,
6069 .cache_nice_tries = 2,
6076 .flags = 1*SD_LOAD_BALANCE
6077 | 1*SD_BALANCE_NEWIDLE
6082 | 0*SD_SHARE_CPUPOWER
6083 | 0*SD_SHARE_PKG_RESOURCES
6085 | 0*SD_PREFER_SIBLING
6086 | sd_local_flags(level)
6088 .last_balance = jiffies,
6089 .balance_interval = sd_weight,
6091 SD_INIT_NAME(sd, NUMA);
6092 sd->private = &tl->data;
6095 * Ugly hack to pass state to sd_numa_mask()...
6097 sched_domains_curr_level = tl->numa_level;
6102 static const struct cpumask *sd_numa_mask(int cpu)
6104 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6107 static void sched_numa_warn(const char *str)
6109 static int done = false;
6117 printk(KERN_WARNING "ERROR: %s\n\n", str);
6119 for (i = 0; i < nr_node_ids; i++) {
6120 printk(KERN_WARNING " ");
6121 for (j = 0; j < nr_node_ids; j++)
6122 printk(KERN_CONT "%02d ", node_distance(i,j));
6123 printk(KERN_CONT "\n");
6125 printk(KERN_WARNING "\n");
6128 static bool find_numa_distance(int distance)
6132 if (distance == node_distance(0, 0))
6135 for (i = 0; i < sched_domains_numa_levels; i++) {
6136 if (sched_domains_numa_distance[i] == distance)
6143 static void sched_init_numa(void)
6145 int next_distance, curr_distance = node_distance(0, 0);
6146 struct sched_domain_topology_level *tl;
6150 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6151 if (!sched_domains_numa_distance)
6155 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6156 * unique distances in the node_distance() table.
6158 * Assumes node_distance(0,j) includes all distances in
6159 * node_distance(i,j) in order to avoid cubic time.
6161 next_distance = curr_distance;
6162 for (i = 0; i < nr_node_ids; i++) {
6163 for (j = 0; j < nr_node_ids; j++) {
6164 for (k = 0; k < nr_node_ids; k++) {
6165 int distance = node_distance(i, k);
6167 if (distance > curr_distance &&
6168 (distance < next_distance ||
6169 next_distance == curr_distance))
6170 next_distance = distance;
6173 * While not a strong assumption it would be nice to know
6174 * about cases where if node A is connected to B, B is not
6175 * equally connected to A.
6177 if (sched_debug() && node_distance(k, i) != distance)
6178 sched_numa_warn("Node-distance not symmetric");
6180 if (sched_debug() && i && !find_numa_distance(distance))
6181 sched_numa_warn("Node-0 not representative");
6183 if (next_distance != curr_distance) {
6184 sched_domains_numa_distance[level++] = next_distance;
6185 sched_domains_numa_levels = level;
6186 curr_distance = next_distance;
6191 * In case of sched_debug() we verify the above assumption.
6197 * 'level' contains the number of unique distances, excluding the
6198 * identity distance node_distance(i,i).
6200 * The sched_domains_nume_distance[] array includes the actual distance
6205 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6206 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6207 * the array will contain less then 'level' members. This could be
6208 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6209 * in other functions.
6211 * We reset it to 'level' at the end of this function.
6213 sched_domains_numa_levels = 0;
6215 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6216 if (!sched_domains_numa_masks)
6220 * Now for each level, construct a mask per node which contains all
6221 * cpus of nodes that are that many hops away from us.
6223 for (i = 0; i < level; i++) {
6224 sched_domains_numa_masks[i] =
6225 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6226 if (!sched_domains_numa_masks[i])
6229 for (j = 0; j < nr_node_ids; j++) {
6230 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6234 sched_domains_numa_masks[i][j] = mask;
6236 for (k = 0; k < nr_node_ids; k++) {
6237 if (node_distance(j, k) > sched_domains_numa_distance[i])
6240 cpumask_or(mask, mask, cpumask_of_node(k));
6245 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6246 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6251 * Copy the default topology bits..
6253 for (i = 0; default_topology[i].init; i++)
6254 tl[i] = default_topology[i];
6257 * .. and append 'j' levels of NUMA goodness.
6259 for (j = 0; j < level; i++, j++) {
6260 tl[i] = (struct sched_domain_topology_level){
6261 .init = sd_numa_init,
6262 .mask = sd_numa_mask,
6263 .flags = SDTL_OVERLAP,
6268 sched_domain_topology = tl;
6270 sched_domains_numa_levels = level;
6273 static void sched_domains_numa_masks_set(int cpu)
6276 int node = cpu_to_node(cpu);
6278 for (i = 0; i < sched_domains_numa_levels; i++) {
6279 for (j = 0; j < nr_node_ids; j++) {
6280 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6281 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6286 static void sched_domains_numa_masks_clear(int cpu)
6289 for (i = 0; i < sched_domains_numa_levels; i++) {
6290 for (j = 0; j < nr_node_ids; j++)
6291 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6296 * Update sched_domains_numa_masks[level][node] array when new cpus
6299 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6300 unsigned long action,
6303 int cpu = (long)hcpu;
6305 switch (action & ~CPU_TASKS_FROZEN) {
6307 sched_domains_numa_masks_set(cpu);
6311 sched_domains_numa_masks_clear(cpu);
6321 static inline void sched_init_numa(void)
6325 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6326 unsigned long action,
6331 #endif /* CONFIG_NUMA */
6333 static int __sdt_alloc(const struct cpumask *cpu_map)
6335 struct sched_domain_topology_level *tl;
6338 for (tl = sched_domain_topology; tl->init; tl++) {
6339 struct sd_data *sdd = &tl->data;
6341 sdd->sd = alloc_percpu(struct sched_domain *);
6345 sdd->sg = alloc_percpu(struct sched_group *);
6349 sdd->sgp = alloc_percpu(struct sched_group_power *);
6353 for_each_cpu(j, cpu_map) {
6354 struct sched_domain *sd;
6355 struct sched_group *sg;
6356 struct sched_group_power *sgp;
6358 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6359 GFP_KERNEL, cpu_to_node(j));
6363 *per_cpu_ptr(sdd->sd, j) = sd;
6365 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6366 GFP_KERNEL, cpu_to_node(j));
6372 *per_cpu_ptr(sdd->sg, j) = sg;
6374 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6375 GFP_KERNEL, cpu_to_node(j));
6379 *per_cpu_ptr(sdd->sgp, j) = sgp;
6386 static void __sdt_free(const struct cpumask *cpu_map)
6388 struct sched_domain_topology_level *tl;
6391 for (tl = sched_domain_topology; tl->init; tl++) {
6392 struct sd_data *sdd = &tl->data;
6394 for_each_cpu(j, cpu_map) {
6395 struct sched_domain *sd;
6398 sd = *per_cpu_ptr(sdd->sd, j);
6399 if (sd && (sd->flags & SD_OVERLAP))
6400 free_sched_groups(sd->groups, 0);
6401 kfree(*per_cpu_ptr(sdd->sd, j));
6405 kfree(*per_cpu_ptr(sdd->sg, j));
6407 kfree(*per_cpu_ptr(sdd->sgp, j));
6409 free_percpu(sdd->sd);
6411 free_percpu(sdd->sg);
6413 free_percpu(sdd->sgp);
6418 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6419 struct s_data *d, const struct cpumask *cpu_map,
6420 struct sched_domain_attr *attr, struct sched_domain *child,
6423 struct sched_domain *sd = tl->init(tl, cpu);
6427 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6429 sd->level = child->level + 1;
6430 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6434 set_domain_attribute(sd, attr);
6440 * Build sched domains for a given set of cpus and attach the sched domains
6441 * to the individual cpus
6443 static int build_sched_domains(const struct cpumask *cpu_map,
6444 struct sched_domain_attr *attr)
6446 enum s_alloc alloc_state = sa_none;
6447 struct sched_domain *sd;
6449 int i, ret = -ENOMEM;
6451 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6452 if (alloc_state != sa_rootdomain)
6455 /* Set up domains for cpus specified by the cpu_map. */
6456 for_each_cpu(i, cpu_map) {
6457 struct sched_domain_topology_level *tl;
6460 for (tl = sched_domain_topology; tl->init; tl++) {
6461 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6462 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6463 sd->flags |= SD_OVERLAP;
6464 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6471 *per_cpu_ptr(d.sd, i) = sd;
6474 /* Build the groups for the domains */
6475 for_each_cpu(i, cpu_map) {
6476 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6477 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6478 if (sd->flags & SD_OVERLAP) {
6479 if (build_overlap_sched_groups(sd, i))
6482 if (build_sched_groups(sd, i))
6488 /* Calculate CPU power for physical packages and nodes */
6489 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6490 if (!cpumask_test_cpu(i, cpu_map))
6493 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6494 claim_allocations(i, sd);
6495 init_sched_groups_power(i, sd);
6499 /* Attach the domains */
6501 for_each_cpu(i, cpu_map) {
6502 sd = *per_cpu_ptr(d.sd, i);
6503 cpu_attach_domain(sd, d.rd, i);
6509 __free_domain_allocs(&d, alloc_state, cpu_map);
6513 static cpumask_var_t *doms_cur; /* current sched domains */
6514 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6515 static struct sched_domain_attr *dattr_cur;
6516 /* attribues of custom domains in 'doms_cur' */
6519 * Special case: If a kmalloc of a doms_cur partition (array of
6520 * cpumask) fails, then fallback to a single sched domain,
6521 * as determined by the single cpumask fallback_doms.
6523 static cpumask_var_t fallback_doms;
6526 * arch_update_cpu_topology lets virtualized architectures update the
6527 * cpu core maps. It is supposed to return 1 if the topology changed
6528 * or 0 if it stayed the same.
6530 int __attribute__((weak)) arch_update_cpu_topology(void)
6535 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6538 cpumask_var_t *doms;
6540 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6543 for (i = 0; i < ndoms; i++) {
6544 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6545 free_sched_domains(doms, i);
6552 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6555 for (i = 0; i < ndoms; i++)
6556 free_cpumask_var(doms[i]);
6561 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6562 * For now this just excludes isolated cpus, but could be used to
6563 * exclude other special cases in the future.
6565 static int init_sched_domains(const struct cpumask *cpu_map)
6569 arch_update_cpu_topology();
6571 doms_cur = alloc_sched_domains(ndoms_cur);
6573 doms_cur = &fallback_doms;
6574 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6575 err = build_sched_domains(doms_cur[0], NULL);
6576 register_sched_domain_sysctl();
6582 * Detach sched domains from a group of cpus specified in cpu_map
6583 * These cpus will now be attached to the NULL domain
6585 static void detach_destroy_domains(const struct cpumask *cpu_map)
6590 for_each_cpu(i, cpu_map)
6591 cpu_attach_domain(NULL, &def_root_domain, i);
6595 /* handle null as "default" */
6596 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6597 struct sched_domain_attr *new, int idx_new)
6599 struct sched_domain_attr tmp;
6606 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6607 new ? (new + idx_new) : &tmp,
6608 sizeof(struct sched_domain_attr));
6612 * Partition sched domains as specified by the 'ndoms_new'
6613 * cpumasks in the array doms_new[] of cpumasks. This compares
6614 * doms_new[] to the current sched domain partitioning, doms_cur[].
6615 * It destroys each deleted domain and builds each new domain.
6617 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6618 * The masks don't intersect (don't overlap.) We should setup one
6619 * sched domain for each mask. CPUs not in any of the cpumasks will
6620 * not be load balanced. If the same cpumask appears both in the
6621 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6624 * The passed in 'doms_new' should be allocated using
6625 * alloc_sched_domains. This routine takes ownership of it and will
6626 * free_sched_domains it when done with it. If the caller failed the
6627 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6628 * and partition_sched_domains() will fallback to the single partition
6629 * 'fallback_doms', it also forces the domains to be rebuilt.
6631 * If doms_new == NULL it will be replaced with cpu_online_mask.
6632 * ndoms_new == 0 is a special case for destroying existing domains,
6633 * and it will not create the default domain.
6635 * Call with hotplug lock held
6637 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6638 struct sched_domain_attr *dattr_new)
6643 mutex_lock(&sched_domains_mutex);
6645 /* always unregister in case we don't destroy any domains */
6646 unregister_sched_domain_sysctl();
6648 /* Let architecture update cpu core mappings. */
6649 new_topology = arch_update_cpu_topology();
6651 n = doms_new ? ndoms_new : 0;
6653 /* Destroy deleted domains */
6654 for (i = 0; i < ndoms_cur; i++) {
6655 for (j = 0; j < n && !new_topology; j++) {
6656 if (cpumask_equal(doms_cur[i], doms_new[j])
6657 && dattrs_equal(dattr_cur, i, dattr_new, j))
6660 /* no match - a current sched domain not in new doms_new[] */
6661 detach_destroy_domains(doms_cur[i]);
6666 if (doms_new == NULL) {
6668 doms_new = &fallback_doms;
6669 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6670 WARN_ON_ONCE(dattr_new);
6673 /* Build new domains */
6674 for (i = 0; i < ndoms_new; i++) {
6675 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6676 if (cpumask_equal(doms_new[i], doms_cur[j])
6677 && dattrs_equal(dattr_new, i, dattr_cur, j))
6680 /* no match - add a new doms_new */
6681 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6686 /* Remember the new sched domains */
6687 if (doms_cur != &fallback_doms)
6688 free_sched_domains(doms_cur, ndoms_cur);
6689 kfree(dattr_cur); /* kfree(NULL) is safe */
6690 doms_cur = doms_new;
6691 dattr_cur = dattr_new;
6692 ndoms_cur = ndoms_new;
6694 register_sched_domain_sysctl();
6696 mutex_unlock(&sched_domains_mutex);
6699 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6702 * Update cpusets according to cpu_active mask. If cpusets are
6703 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6704 * around partition_sched_domains().
6706 * If we come here as part of a suspend/resume, don't touch cpusets because we
6707 * want to restore it back to its original state upon resume anyway.
6709 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6713 case CPU_ONLINE_FROZEN:
6714 case CPU_DOWN_FAILED_FROZEN:
6717 * num_cpus_frozen tracks how many CPUs are involved in suspend
6718 * resume sequence. As long as this is not the last online
6719 * operation in the resume sequence, just build a single sched
6720 * domain, ignoring cpusets.
6723 if (likely(num_cpus_frozen)) {
6724 partition_sched_domains(1, NULL, NULL);
6729 * This is the last CPU online operation. So fall through and
6730 * restore the original sched domains by considering the
6731 * cpuset configurations.
6735 case CPU_DOWN_FAILED:
6736 cpuset_update_active_cpus(true);
6744 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6748 case CPU_DOWN_PREPARE:
6749 cpuset_update_active_cpus(false);
6751 case CPU_DOWN_PREPARE_FROZEN:
6753 partition_sched_domains(1, NULL, NULL);
6761 void __init sched_init_smp(void)
6763 cpumask_var_t non_isolated_cpus;
6765 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6766 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6771 mutex_lock(&sched_domains_mutex);
6772 init_sched_domains(cpu_active_mask);
6773 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6774 if (cpumask_empty(non_isolated_cpus))
6775 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6776 mutex_unlock(&sched_domains_mutex);
6779 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6780 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6781 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6783 /* RT runtime code needs to handle some hotplug events */
6784 hotcpu_notifier(update_runtime, 0);
6788 /* Move init over to a non-isolated CPU */
6789 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6791 sched_init_granularity();
6792 free_cpumask_var(non_isolated_cpus);
6794 init_sched_rt_class();
6797 void __init sched_init_smp(void)
6799 sched_init_granularity();
6801 #endif /* CONFIG_SMP */
6803 const_debug unsigned int sysctl_timer_migration = 1;
6805 int in_sched_functions(unsigned long addr)
6807 return in_lock_functions(addr) ||
6808 (addr >= (unsigned long)__sched_text_start
6809 && addr < (unsigned long)__sched_text_end);
6812 #ifdef CONFIG_CGROUP_SCHED
6813 struct task_group root_task_group;
6814 LIST_HEAD(task_groups);
6817 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6819 void __init sched_init(void)
6822 unsigned long alloc_size = 0, ptr;
6824 #ifdef CONFIG_FAIR_GROUP_SCHED
6825 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6827 #ifdef CONFIG_RT_GROUP_SCHED
6828 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6830 #ifdef CONFIG_CPUMASK_OFFSTACK
6831 alloc_size += num_possible_cpus() * cpumask_size();
6834 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6836 #ifdef CONFIG_FAIR_GROUP_SCHED
6837 root_task_group.se = (struct sched_entity **)ptr;
6838 ptr += nr_cpu_ids * sizeof(void **);
6840 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6841 ptr += nr_cpu_ids * sizeof(void **);
6843 #endif /* CONFIG_FAIR_GROUP_SCHED */
6844 #ifdef CONFIG_RT_GROUP_SCHED
6845 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6846 ptr += nr_cpu_ids * sizeof(void **);
6848 root_task_group.rt_rq = (struct rt_rq **)ptr;
6849 ptr += nr_cpu_ids * sizeof(void **);
6851 #endif /* CONFIG_RT_GROUP_SCHED */
6852 #ifdef CONFIG_CPUMASK_OFFSTACK
6853 for_each_possible_cpu(i) {
6854 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6855 ptr += cpumask_size();
6857 #endif /* CONFIG_CPUMASK_OFFSTACK */
6861 init_defrootdomain();
6864 init_rt_bandwidth(&def_rt_bandwidth,
6865 global_rt_period(), global_rt_runtime());
6867 #ifdef CONFIG_RT_GROUP_SCHED
6868 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6869 global_rt_period(), global_rt_runtime());
6870 #endif /* CONFIG_RT_GROUP_SCHED */
6872 #ifdef CONFIG_CGROUP_SCHED
6873 list_add(&root_task_group.list, &task_groups);
6874 INIT_LIST_HEAD(&root_task_group.children);
6875 INIT_LIST_HEAD(&root_task_group.siblings);
6876 autogroup_init(&init_task);
6878 #endif /* CONFIG_CGROUP_SCHED */
6880 #ifdef CONFIG_CGROUP_CPUACCT
6881 root_cpuacct.cpustat = &kernel_cpustat;
6882 root_cpuacct.cpuusage = alloc_percpu(u64);
6883 /* Too early, not expected to fail */
6884 BUG_ON(!root_cpuacct.cpuusage);
6886 for_each_possible_cpu(i) {
6890 raw_spin_lock_init(&rq->lock);
6892 rq->calc_load_active = 0;
6893 rq->calc_load_update = jiffies + LOAD_FREQ;
6894 init_cfs_rq(&rq->cfs);
6895 init_rt_rq(&rq->rt, rq);
6896 #ifdef CONFIG_FAIR_GROUP_SCHED
6897 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6898 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6900 * How much cpu bandwidth does root_task_group get?
6902 * In case of task-groups formed thr' the cgroup filesystem, it
6903 * gets 100% of the cpu resources in the system. This overall
6904 * system cpu resource is divided among the tasks of
6905 * root_task_group and its child task-groups in a fair manner,
6906 * based on each entity's (task or task-group's) weight
6907 * (se->load.weight).
6909 * In other words, if root_task_group has 10 tasks of weight
6910 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6911 * then A0's share of the cpu resource is:
6913 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6915 * We achieve this by letting root_task_group's tasks sit
6916 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6918 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6919 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6920 #endif /* CONFIG_FAIR_GROUP_SCHED */
6922 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6923 #ifdef CONFIG_RT_GROUP_SCHED
6924 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6925 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6928 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6929 rq->cpu_load[j] = 0;
6931 rq->last_load_update_tick = jiffies;
6936 rq->cpu_power = SCHED_POWER_SCALE;
6937 rq->post_schedule = 0;
6938 rq->active_balance = 0;
6939 rq->next_balance = jiffies;
6944 rq->avg_idle = 2*sysctl_sched_migration_cost;
6946 INIT_LIST_HEAD(&rq->cfs_tasks);
6948 rq_attach_root(rq, &def_root_domain);
6954 atomic_set(&rq->nr_iowait, 0);
6957 set_load_weight(&init_task);
6959 #ifdef CONFIG_PREEMPT_NOTIFIERS
6960 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6963 #ifdef CONFIG_RT_MUTEXES
6964 plist_head_init(&init_task.pi_waiters);
6968 * The boot idle thread does lazy MMU switching as well:
6970 atomic_inc(&init_mm.mm_count);
6971 enter_lazy_tlb(&init_mm, current);
6974 * Make us the idle thread. Technically, schedule() should not be
6975 * called from this thread, however somewhere below it might be,
6976 * but because we are the idle thread, we just pick up running again
6977 * when this runqueue becomes "idle".
6979 init_idle(current, smp_processor_id());
6981 calc_load_update = jiffies + LOAD_FREQ;
6984 * During early bootup we pretend to be a normal task:
6986 current->sched_class = &fair_sched_class;
6989 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6990 /* May be allocated at isolcpus cmdline parse time */
6991 if (cpu_isolated_map == NULL)
6992 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6993 idle_thread_set_boot_cpu();
6995 init_sched_fair_class();
6997 scheduler_running = 1;
7000 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7001 static inline int preempt_count_equals(int preempt_offset)
7003 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7005 return (nested == preempt_offset);
7008 void __might_sleep(const char *file, int line, int preempt_offset)
7010 static unsigned long prev_jiffy; /* ratelimiting */
7012 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7013 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7014 system_state != SYSTEM_RUNNING || oops_in_progress)
7016 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7018 prev_jiffy = jiffies;
7021 "BUG: sleeping function called from invalid context at %s:%d\n",
7024 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7025 in_atomic(), irqs_disabled(),
7026 current->pid, current->comm);
7028 debug_show_held_locks(current);
7029 if (irqs_disabled())
7030 print_irqtrace_events(current);
7033 EXPORT_SYMBOL(__might_sleep);
7036 #ifdef CONFIG_MAGIC_SYSRQ
7037 static void normalize_task(struct rq *rq, struct task_struct *p)
7039 const struct sched_class *prev_class = p->sched_class;
7040 int old_prio = p->prio;
7045 dequeue_task(rq, p, 0);
7046 __setscheduler(rq, p, SCHED_NORMAL, 0);
7048 enqueue_task(rq, p, 0);
7049 resched_task(rq->curr);
7052 check_class_changed(rq, p, prev_class, old_prio);
7055 void normalize_rt_tasks(void)
7057 struct task_struct *g, *p;
7058 unsigned long flags;
7061 read_lock_irqsave(&tasklist_lock, flags);
7062 do_each_thread(g, p) {
7064 * Only normalize user tasks:
7069 p->se.exec_start = 0;
7070 #ifdef CONFIG_SCHEDSTATS
7071 p->se.statistics.wait_start = 0;
7072 p->se.statistics.sleep_start = 0;
7073 p->se.statistics.block_start = 0;
7078 * Renice negative nice level userspace
7081 if (TASK_NICE(p) < 0 && p->mm)
7082 set_user_nice(p, 0);
7086 raw_spin_lock(&p->pi_lock);
7087 rq = __task_rq_lock(p);
7089 normalize_task(rq, p);
7091 __task_rq_unlock(rq);
7092 raw_spin_unlock(&p->pi_lock);
7093 } while_each_thread(g, p);
7095 read_unlock_irqrestore(&tasklist_lock, flags);
7098 #endif /* CONFIG_MAGIC_SYSRQ */
7100 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7102 * These functions are only useful for the IA64 MCA handling, or kdb.
7104 * They can only be called when the whole system has been
7105 * stopped - every CPU needs to be quiescent, and no scheduling
7106 * activity can take place. Using them for anything else would
7107 * be a serious bug, and as a result, they aren't even visible
7108 * under any other configuration.
7112 * curr_task - return the current task for a given cpu.
7113 * @cpu: the processor in question.
7115 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7117 struct task_struct *curr_task(int cpu)
7119 return cpu_curr(cpu);
7122 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7126 * set_curr_task - set the current task for a given cpu.
7127 * @cpu: the processor in question.
7128 * @p: the task pointer to set.
7130 * Description: This function must only be used when non-maskable interrupts
7131 * are serviced on a separate stack. It allows the architecture to switch the
7132 * notion of the current task on a cpu in a non-blocking manner. This function
7133 * must be called with all CPU's synchronized, and interrupts disabled, the
7134 * and caller must save the original value of the current task (see
7135 * curr_task() above) and restore that value before reenabling interrupts and
7136 * re-starting the system.
7138 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7140 void set_curr_task(int cpu, struct task_struct *p)
7147 #ifdef CONFIG_CGROUP_SCHED
7148 /* task_group_lock serializes the addition/removal of task groups */
7149 static DEFINE_SPINLOCK(task_group_lock);
7151 static void free_sched_group(struct task_group *tg)
7153 free_fair_sched_group(tg);
7154 free_rt_sched_group(tg);
7159 /* allocate runqueue etc for a new task group */
7160 struct task_group *sched_create_group(struct task_group *parent)
7162 struct task_group *tg;
7163 unsigned long flags;
7165 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7167 return ERR_PTR(-ENOMEM);
7169 if (!alloc_fair_sched_group(tg, parent))
7172 if (!alloc_rt_sched_group(tg, parent))
7175 spin_lock_irqsave(&task_group_lock, flags);
7176 list_add_rcu(&tg->list, &task_groups);
7178 WARN_ON(!parent); /* root should already exist */
7180 tg->parent = parent;
7181 INIT_LIST_HEAD(&tg->children);
7182 list_add_rcu(&tg->siblings, &parent->children);
7183 spin_unlock_irqrestore(&task_group_lock, flags);
7188 free_sched_group(tg);
7189 return ERR_PTR(-ENOMEM);
7192 /* rcu callback to free various structures associated with a task group */
7193 static void free_sched_group_rcu(struct rcu_head *rhp)
7195 /* now it should be safe to free those cfs_rqs */
7196 free_sched_group(container_of(rhp, struct task_group, rcu));
7199 /* Destroy runqueue etc associated with a task group */
7200 void sched_destroy_group(struct task_group *tg)
7202 unsigned long flags;
7205 /* end participation in shares distribution */
7206 for_each_possible_cpu(i)
7207 unregister_fair_sched_group(tg, i);
7209 spin_lock_irqsave(&task_group_lock, flags);
7210 list_del_rcu(&tg->list);
7211 list_del_rcu(&tg->siblings);
7212 spin_unlock_irqrestore(&task_group_lock, flags);
7214 /* wait for possible concurrent references to cfs_rqs complete */
7215 call_rcu(&tg->rcu, free_sched_group_rcu);
7218 /* change task's runqueue when it moves between groups.
7219 * The caller of this function should have put the task in its new group
7220 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7221 * reflect its new group.
7223 void sched_move_task(struct task_struct *tsk)
7225 struct task_group *tg;
7227 unsigned long flags;
7230 rq = task_rq_lock(tsk, &flags);
7232 running = task_current(rq, tsk);
7236 dequeue_task(rq, tsk, 0);
7237 if (unlikely(running))
7238 tsk->sched_class->put_prev_task(rq, tsk);
7240 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7241 lockdep_is_held(&tsk->sighand->siglock)),
7242 struct task_group, css);
7243 tg = autogroup_task_group(tsk, tg);
7244 tsk->sched_task_group = tg;
7246 #ifdef CONFIG_FAIR_GROUP_SCHED
7247 if (tsk->sched_class->task_move_group)
7248 tsk->sched_class->task_move_group(tsk, on_rq);
7251 set_task_rq(tsk, task_cpu(tsk));
7253 if (unlikely(running))
7254 tsk->sched_class->set_curr_task(rq);
7256 enqueue_task(rq, tsk, 0);
7258 task_rq_unlock(rq, tsk, &flags);
7260 #endif /* CONFIG_CGROUP_SCHED */
7262 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7263 static unsigned long to_ratio(u64 period, u64 runtime)
7265 if (runtime == RUNTIME_INF)
7268 return div64_u64(runtime << 20, period);
7272 #ifdef CONFIG_RT_GROUP_SCHED
7274 * Ensure that the real time constraints are schedulable.
7276 static DEFINE_MUTEX(rt_constraints_mutex);
7278 /* Must be called with tasklist_lock held */
7279 static inline int tg_has_rt_tasks(struct task_group *tg)
7281 struct task_struct *g, *p;
7283 do_each_thread(g, p) {
7284 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7286 } while_each_thread(g, p);
7291 struct rt_schedulable_data {
7292 struct task_group *tg;
7297 static int tg_rt_schedulable(struct task_group *tg, void *data)
7299 struct rt_schedulable_data *d = data;
7300 struct task_group *child;
7301 unsigned long total, sum = 0;
7302 u64 period, runtime;
7304 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7305 runtime = tg->rt_bandwidth.rt_runtime;
7308 period = d->rt_period;
7309 runtime = d->rt_runtime;
7313 * Cannot have more runtime than the period.
7315 if (runtime > period && runtime != RUNTIME_INF)
7319 * Ensure we don't starve existing RT tasks.
7321 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7324 total = to_ratio(period, runtime);
7327 * Nobody can have more than the global setting allows.
7329 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7333 * The sum of our children's runtime should not exceed our own.
7335 list_for_each_entry_rcu(child, &tg->children, siblings) {
7336 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7337 runtime = child->rt_bandwidth.rt_runtime;
7339 if (child == d->tg) {
7340 period = d->rt_period;
7341 runtime = d->rt_runtime;
7344 sum += to_ratio(period, runtime);
7353 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7357 struct rt_schedulable_data data = {
7359 .rt_period = period,
7360 .rt_runtime = runtime,
7364 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7370 static int tg_set_rt_bandwidth(struct task_group *tg,
7371 u64 rt_period, u64 rt_runtime)
7375 mutex_lock(&rt_constraints_mutex);
7376 read_lock(&tasklist_lock);
7377 err = __rt_schedulable(tg, rt_period, rt_runtime);
7381 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7382 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7383 tg->rt_bandwidth.rt_runtime = rt_runtime;
7385 for_each_possible_cpu(i) {
7386 struct rt_rq *rt_rq = tg->rt_rq[i];
7388 raw_spin_lock(&rt_rq->rt_runtime_lock);
7389 rt_rq->rt_runtime = rt_runtime;
7390 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7392 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7394 read_unlock(&tasklist_lock);
7395 mutex_unlock(&rt_constraints_mutex);
7400 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7402 u64 rt_runtime, rt_period;
7404 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7405 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7406 if (rt_runtime_us < 0)
7407 rt_runtime = RUNTIME_INF;
7409 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7412 long sched_group_rt_runtime(struct task_group *tg)
7416 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7419 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7420 do_div(rt_runtime_us, NSEC_PER_USEC);
7421 return rt_runtime_us;
7424 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7426 u64 rt_runtime, rt_period;
7428 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7429 rt_runtime = tg->rt_bandwidth.rt_runtime;
7434 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7437 long sched_group_rt_period(struct task_group *tg)
7441 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7442 do_div(rt_period_us, NSEC_PER_USEC);
7443 return rt_period_us;
7446 static int sched_rt_global_constraints(void)
7448 u64 runtime, period;
7451 if (sysctl_sched_rt_period <= 0)
7454 runtime = global_rt_runtime();
7455 period = global_rt_period();
7458 * Sanity check on the sysctl variables.
7460 if (runtime > period && runtime != RUNTIME_INF)
7463 mutex_lock(&rt_constraints_mutex);
7464 read_lock(&tasklist_lock);
7465 ret = __rt_schedulable(NULL, 0, 0);
7466 read_unlock(&tasklist_lock);
7467 mutex_unlock(&rt_constraints_mutex);
7472 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7474 /* Don't accept realtime tasks when there is no way for them to run */
7475 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7481 #else /* !CONFIG_RT_GROUP_SCHED */
7482 static int sched_rt_global_constraints(void)
7484 unsigned long flags;
7487 if (sysctl_sched_rt_period <= 0)
7491 * There's always some RT tasks in the root group
7492 * -- migration, kstopmachine etc..
7494 if (sysctl_sched_rt_runtime == 0)
7497 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7498 for_each_possible_cpu(i) {
7499 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7501 raw_spin_lock(&rt_rq->rt_runtime_lock);
7502 rt_rq->rt_runtime = global_rt_runtime();
7503 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7505 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7509 #endif /* CONFIG_RT_GROUP_SCHED */
7511 int sched_rt_handler(struct ctl_table *table, int write,
7512 void __user *buffer, size_t *lenp,
7516 int old_period, old_runtime;
7517 static DEFINE_MUTEX(mutex);
7520 old_period = sysctl_sched_rt_period;
7521 old_runtime = sysctl_sched_rt_runtime;
7523 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7525 if (!ret && write) {
7526 ret = sched_rt_global_constraints();
7528 sysctl_sched_rt_period = old_period;
7529 sysctl_sched_rt_runtime = old_runtime;
7531 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7532 def_rt_bandwidth.rt_period =
7533 ns_to_ktime(global_rt_period());
7536 mutex_unlock(&mutex);
7541 #ifdef CONFIG_CGROUP_SCHED
7543 /* return corresponding task_group object of a cgroup */
7544 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7546 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7547 struct task_group, css);
7550 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7552 struct task_group *tg, *parent;
7554 if (!cgrp->parent) {
7555 /* This is early initialization for the top cgroup */
7556 return &root_task_group.css;
7559 parent = cgroup_tg(cgrp->parent);
7560 tg = sched_create_group(parent);
7562 return ERR_PTR(-ENOMEM);
7567 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7569 struct task_group *tg = cgroup_tg(cgrp);
7571 sched_destroy_group(tg);
7574 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7575 struct cgroup_taskset *tset)
7577 struct task_struct *task;
7579 cgroup_taskset_for_each(task, cgrp, tset) {
7580 #ifdef CONFIG_RT_GROUP_SCHED
7581 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7584 /* We don't support RT-tasks being in separate groups */
7585 if (task->sched_class != &fair_sched_class)
7592 static void cpu_cgroup_attach(struct cgroup *cgrp,
7593 struct cgroup_taskset *tset)
7595 struct task_struct *task;
7597 cgroup_taskset_for_each(task, cgrp, tset)
7598 sched_move_task(task);
7602 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7603 struct task_struct *task)
7606 * cgroup_exit() is called in the copy_process() failure path.
7607 * Ignore this case since the task hasn't ran yet, this avoids
7608 * trying to poke a half freed task state from generic code.
7610 if (!(task->flags & PF_EXITING))
7613 sched_move_task(task);
7616 #ifdef CONFIG_FAIR_GROUP_SCHED
7617 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7620 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7623 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7625 struct task_group *tg = cgroup_tg(cgrp);
7627 return (u64) scale_load_down(tg->shares);
7630 #ifdef CONFIG_CFS_BANDWIDTH
7631 static DEFINE_MUTEX(cfs_constraints_mutex);
7633 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7634 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7636 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7638 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7640 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7641 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7643 if (tg == &root_task_group)
7647 * Ensure we have at some amount of bandwidth every period. This is
7648 * to prevent reaching a state of large arrears when throttled via
7649 * entity_tick() resulting in prolonged exit starvation.
7651 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7655 * Likewise, bound things on the otherside by preventing insane quota
7656 * periods. This also allows us to normalize in computing quota
7659 if (period > max_cfs_quota_period)
7662 mutex_lock(&cfs_constraints_mutex);
7663 ret = __cfs_schedulable(tg, period, quota);
7667 runtime_enabled = quota != RUNTIME_INF;
7668 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7669 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7670 raw_spin_lock_irq(&cfs_b->lock);
7671 cfs_b->period = ns_to_ktime(period);
7672 cfs_b->quota = quota;
7674 __refill_cfs_bandwidth_runtime(cfs_b);
7675 /* restart the period timer (if active) to handle new period expiry */
7676 if (runtime_enabled && cfs_b->timer_active) {
7677 /* force a reprogram */
7678 cfs_b->timer_active = 0;
7679 __start_cfs_bandwidth(cfs_b);
7681 raw_spin_unlock_irq(&cfs_b->lock);
7683 for_each_possible_cpu(i) {
7684 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7685 struct rq *rq = cfs_rq->rq;
7687 raw_spin_lock_irq(&rq->lock);
7688 cfs_rq->runtime_enabled = runtime_enabled;
7689 cfs_rq->runtime_remaining = 0;
7691 if (cfs_rq->throttled)
7692 unthrottle_cfs_rq(cfs_rq);
7693 raw_spin_unlock_irq(&rq->lock);
7696 mutex_unlock(&cfs_constraints_mutex);
7701 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7705 period = ktime_to_ns(tg->cfs_bandwidth.period);
7706 if (cfs_quota_us < 0)
7707 quota = RUNTIME_INF;
7709 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7711 return tg_set_cfs_bandwidth(tg, period, quota);
7714 long tg_get_cfs_quota(struct task_group *tg)
7718 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7721 quota_us = tg->cfs_bandwidth.quota;
7722 do_div(quota_us, NSEC_PER_USEC);
7727 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7731 period = (u64)cfs_period_us * NSEC_PER_USEC;
7732 quota = tg->cfs_bandwidth.quota;
7734 return tg_set_cfs_bandwidth(tg, period, quota);
7737 long tg_get_cfs_period(struct task_group *tg)
7741 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7742 do_div(cfs_period_us, NSEC_PER_USEC);
7744 return cfs_period_us;
7747 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7749 return tg_get_cfs_quota(cgroup_tg(cgrp));
7752 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7755 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7758 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7760 return tg_get_cfs_period(cgroup_tg(cgrp));
7763 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7766 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7769 struct cfs_schedulable_data {
7770 struct task_group *tg;
7775 * normalize group quota/period to be quota/max_period
7776 * note: units are usecs
7778 static u64 normalize_cfs_quota(struct task_group *tg,
7779 struct cfs_schedulable_data *d)
7787 period = tg_get_cfs_period(tg);
7788 quota = tg_get_cfs_quota(tg);
7791 /* note: these should typically be equivalent */
7792 if (quota == RUNTIME_INF || quota == -1)
7795 return to_ratio(period, quota);
7798 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7800 struct cfs_schedulable_data *d = data;
7801 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7802 s64 quota = 0, parent_quota = -1;
7805 quota = RUNTIME_INF;
7807 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7809 quota = normalize_cfs_quota(tg, d);
7810 parent_quota = parent_b->hierarchal_quota;
7813 * ensure max(child_quota) <= parent_quota, inherit when no
7816 if (quota == RUNTIME_INF)
7817 quota = parent_quota;
7818 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7821 cfs_b->hierarchal_quota = quota;
7826 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7829 struct cfs_schedulable_data data = {
7835 if (quota != RUNTIME_INF) {
7836 do_div(data.period, NSEC_PER_USEC);
7837 do_div(data.quota, NSEC_PER_USEC);
7841 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7847 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7848 struct cgroup_map_cb *cb)
7850 struct task_group *tg = cgroup_tg(cgrp);
7851 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7853 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7854 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7855 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7859 #endif /* CONFIG_CFS_BANDWIDTH */
7860 #endif /* CONFIG_FAIR_GROUP_SCHED */
7862 #ifdef CONFIG_RT_GROUP_SCHED
7863 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7866 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7869 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7871 return sched_group_rt_runtime(cgroup_tg(cgrp));
7874 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7877 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7880 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7882 return sched_group_rt_period(cgroup_tg(cgrp));
7884 #endif /* CONFIG_RT_GROUP_SCHED */
7886 static struct cftype cpu_files[] = {
7887 #ifdef CONFIG_FAIR_GROUP_SCHED
7890 .read_u64 = cpu_shares_read_u64,
7891 .write_u64 = cpu_shares_write_u64,
7894 #ifdef CONFIG_CFS_BANDWIDTH
7896 .name = "cfs_quota_us",
7897 .read_s64 = cpu_cfs_quota_read_s64,
7898 .write_s64 = cpu_cfs_quota_write_s64,
7901 .name = "cfs_period_us",
7902 .read_u64 = cpu_cfs_period_read_u64,
7903 .write_u64 = cpu_cfs_period_write_u64,
7907 .read_map = cpu_stats_show,
7910 #ifdef CONFIG_RT_GROUP_SCHED
7912 .name = "rt_runtime_us",
7913 .read_s64 = cpu_rt_runtime_read,
7914 .write_s64 = cpu_rt_runtime_write,
7917 .name = "rt_period_us",
7918 .read_u64 = cpu_rt_period_read_uint,
7919 .write_u64 = cpu_rt_period_write_uint,
7925 struct cgroup_subsys cpu_cgroup_subsys = {
7927 .css_alloc = cpu_cgroup_css_alloc,
7928 .css_free = cpu_cgroup_css_free,
7929 .can_attach = cpu_cgroup_can_attach,
7930 .attach = cpu_cgroup_attach,
7931 .exit = cpu_cgroup_exit,
7932 .subsys_id = cpu_cgroup_subsys_id,
7933 .base_cftypes = cpu_files,
7937 #endif /* CONFIG_CGROUP_SCHED */
7939 #ifdef CONFIG_CGROUP_CPUACCT
7942 * CPU accounting code for task groups.
7944 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7945 * (balbir@in.ibm.com).
7948 struct cpuacct root_cpuacct;
7950 /* create a new cpu accounting group */
7951 static struct cgroup_subsys_state *cpuacct_css_alloc(struct cgroup *cgrp)
7956 return &root_cpuacct.css;
7958 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7962 ca->cpuusage = alloc_percpu(u64);
7966 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7968 goto out_free_cpuusage;
7973 free_percpu(ca->cpuusage);
7977 return ERR_PTR(-ENOMEM);
7980 /* destroy an existing cpu accounting group */
7981 static void cpuacct_css_free(struct cgroup *cgrp)
7983 struct cpuacct *ca = cgroup_ca(cgrp);
7985 free_percpu(ca->cpustat);
7986 free_percpu(ca->cpuusage);
7990 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7992 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7995 #ifndef CONFIG_64BIT
7997 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7999 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8001 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8009 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8011 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8013 #ifndef CONFIG_64BIT
8015 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8017 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8019 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8025 /* return total cpu usage (in nanoseconds) of a group */
8026 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8028 struct cpuacct *ca = cgroup_ca(cgrp);
8029 u64 totalcpuusage = 0;
8032 for_each_present_cpu(i)
8033 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8035 return totalcpuusage;
8038 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8041 struct cpuacct *ca = cgroup_ca(cgrp);
8050 for_each_present_cpu(i)
8051 cpuacct_cpuusage_write(ca, i, 0);
8057 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8060 struct cpuacct *ca = cgroup_ca(cgroup);
8064 for_each_present_cpu(i) {
8065 percpu = cpuacct_cpuusage_read(ca, i);
8066 seq_printf(m, "%llu ", (unsigned long long) percpu);
8068 seq_printf(m, "\n");
8072 static const char *cpuacct_stat_desc[] = {
8073 [CPUACCT_STAT_USER] = "user",
8074 [CPUACCT_STAT_SYSTEM] = "system",
8077 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8078 struct cgroup_map_cb *cb)
8080 struct cpuacct *ca = cgroup_ca(cgrp);
8084 for_each_online_cpu(cpu) {
8085 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8086 val += kcpustat->cpustat[CPUTIME_USER];
8087 val += kcpustat->cpustat[CPUTIME_NICE];
8089 val = cputime64_to_clock_t(val);
8090 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8093 for_each_online_cpu(cpu) {
8094 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8095 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8096 val += kcpustat->cpustat[CPUTIME_IRQ];
8097 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8100 val = cputime64_to_clock_t(val);
8101 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8106 static struct cftype files[] = {
8109 .read_u64 = cpuusage_read,
8110 .write_u64 = cpuusage_write,
8113 .name = "usage_percpu",
8114 .read_seq_string = cpuacct_percpu_seq_read,
8118 .read_map = cpuacct_stats_show,
8124 * charge this task's execution time to its accounting group.
8126 * called with rq->lock held.
8128 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8133 if (unlikely(!cpuacct_subsys.active))
8136 cpu = task_cpu(tsk);
8142 for (; ca; ca = parent_ca(ca)) {
8143 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8144 *cpuusage += cputime;
8150 struct cgroup_subsys cpuacct_subsys = {
8152 .css_alloc = cpuacct_css_alloc,
8153 .css_free = cpuacct_css_free,
8154 .subsys_id = cpuacct_subsys_id,
8155 .base_cftypes = files,
8157 #endif /* CONFIG_CGROUP_CPUACCT */
8159 void dump_cpu_task(int cpu)
8161 pr_info("Task dump for CPU %d:\n", cpu);
8162 sched_show_task(cpu_curr(cpu));