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_internal.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.
375 static void hrtick_clear(struct rq *rq)
377 if (hrtimer_active(&rq->hrtick_timer))
378 hrtimer_cancel(&rq->hrtick_timer);
382 * High-resolution timer tick.
383 * Runs from hardirq context with interrupts disabled.
385 static enum hrtimer_restart hrtick(struct hrtimer *timer)
387 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
389 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
391 raw_spin_lock(&rq->lock);
393 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
394 raw_spin_unlock(&rq->lock);
396 return HRTIMER_NORESTART;
401 static int __hrtick_restart(struct rq *rq)
403 struct hrtimer *timer = &rq->hrtick_timer;
404 ktime_t time = hrtimer_get_softexpires(timer);
406 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
410 * called from hardirq (IPI) context
412 static void __hrtick_start(void *arg)
416 raw_spin_lock(&rq->lock);
417 __hrtick_restart(rq);
418 rq->hrtick_csd_pending = 0;
419 raw_spin_unlock(&rq->lock);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq *rq, u64 delay)
429 struct hrtimer *timer = &rq->hrtick_timer;
430 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
432 hrtimer_set_expires(timer, time);
434 if (rq == this_rq()) {
435 __hrtick_restart(rq);
436 } else if (!rq->hrtick_csd_pending) {
437 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
438 rq->hrtick_csd_pending = 1;
443 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
445 int cpu = (int)(long)hcpu;
448 case CPU_UP_CANCELED:
449 case CPU_UP_CANCELED_FROZEN:
450 case CPU_DOWN_PREPARE:
451 case CPU_DOWN_PREPARE_FROZEN:
453 case CPU_DEAD_FROZEN:
454 hrtick_clear(cpu_rq(cpu));
461 static __init void init_hrtick(void)
463 hotcpu_notifier(hotplug_hrtick, 0);
467 * Called to set the hrtick timer state.
469 * called with rq->lock held and irqs disabled
471 void hrtick_start(struct rq *rq, u64 delay)
473 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
474 HRTIMER_MODE_REL_PINNED, 0);
477 static inline void init_hrtick(void)
480 #endif /* CONFIG_SMP */
482 static void init_rq_hrtick(struct rq *rq)
485 rq->hrtick_csd_pending = 0;
487 rq->hrtick_csd.flags = 0;
488 rq->hrtick_csd.func = __hrtick_start;
489 rq->hrtick_csd.info = rq;
492 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
493 rq->hrtick_timer.function = hrtick;
495 #else /* CONFIG_SCHED_HRTICK */
496 static inline void hrtick_clear(struct rq *rq)
500 static inline void init_rq_hrtick(struct rq *rq)
504 static inline void init_hrtick(void)
507 #endif /* CONFIG_SCHED_HRTICK */
510 * resched_task - mark a task 'to be rescheduled now'.
512 * On UP this means the setting of the need_resched flag, on SMP it
513 * might also involve a cross-CPU call to trigger the scheduler on
516 void resched_task(struct task_struct *p)
520 lockdep_assert_held(&task_rq(p)->lock);
522 if (test_tsk_need_resched(p))
525 set_tsk_need_resched(p);
528 if (cpu == smp_processor_id()) {
529 set_preempt_need_resched();
533 /* NEED_RESCHED must be visible before we test polling */
535 if (!tsk_is_polling(p))
536 smp_send_reschedule(cpu);
539 void resched_cpu(int cpu)
541 struct rq *rq = cpu_rq(cpu);
544 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
546 resched_task(cpu_curr(cpu));
547 raw_spin_unlock_irqrestore(&rq->lock, flags);
551 #ifdef CONFIG_NO_HZ_COMMON
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu = smp_processor_id();
564 struct sched_domain *sd;
567 for_each_domain(cpu, sd) {
568 for_each_cpu(i, sched_domain_span(sd)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 static void wake_up_idle_cpu(int cpu)
591 struct rq *rq = cpu_rq(cpu);
593 if (cpu == smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq->curr != rq->idle)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq->idle);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq->idle))
616 smp_send_reschedule(cpu);
619 static bool wake_up_full_nohz_cpu(int cpu)
621 if (tick_nohz_full_cpu(cpu)) {
622 if (cpu != smp_processor_id() ||
623 tick_nohz_tick_stopped())
624 smp_send_reschedule(cpu);
631 void wake_up_nohz_cpu(int cpu)
633 if (!wake_up_full_nohz_cpu(cpu))
634 wake_up_idle_cpu(cpu);
637 static inline bool got_nohz_idle_kick(void)
639 int cpu = smp_processor_id();
641 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
644 if (idle_cpu(cpu) && !need_resched())
648 * We can't run Idle Load Balance on this CPU for this time so we
649 * cancel it and clear NOHZ_BALANCE_KICK
651 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
655 #else /* CONFIG_NO_HZ_COMMON */
657 static inline bool got_nohz_idle_kick(void)
662 #endif /* CONFIG_NO_HZ_COMMON */
664 #ifdef CONFIG_NO_HZ_FULL
665 bool sched_can_stop_tick(void)
671 /* Make sure rq->nr_running update is visible after the IPI */
674 /* More than one running task need preemption */
675 if (rq->nr_running > 1)
680 #endif /* CONFIG_NO_HZ_FULL */
682 void sched_avg_update(struct rq *rq)
684 s64 period = sched_avg_period();
686 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
688 * Inline assembly required to prevent the compiler
689 * optimising this loop into a divmod call.
690 * See __iter_div_u64_rem() for another example of this.
692 asm("" : "+rm" (rq->age_stamp));
693 rq->age_stamp += period;
698 #endif /* CONFIG_SMP */
700 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
701 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
703 * Iterate task_group tree rooted at *from, calling @down when first entering a
704 * node and @up when leaving it for the final time.
706 * Caller must hold rcu_lock or sufficient equivalent.
708 int walk_tg_tree_from(struct task_group *from,
709 tg_visitor down, tg_visitor up, void *data)
711 struct task_group *parent, *child;
717 ret = (*down)(parent, data);
720 list_for_each_entry_rcu(child, &parent->children, siblings) {
727 ret = (*up)(parent, data);
728 if (ret || parent == from)
732 parent = parent->parent;
739 int tg_nop(struct task_group *tg, void *data)
745 static void set_load_weight(struct task_struct *p)
747 int prio = p->static_prio - MAX_RT_PRIO;
748 struct load_weight *load = &p->se.load;
751 * SCHED_IDLE tasks get minimal weight:
753 if (p->policy == SCHED_IDLE) {
754 load->weight = scale_load(WEIGHT_IDLEPRIO);
755 load->inv_weight = WMULT_IDLEPRIO;
759 load->weight = scale_load(prio_to_weight[prio]);
760 load->inv_weight = prio_to_wmult[prio];
763 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
766 sched_info_queued(rq, p);
767 p->sched_class->enqueue_task(rq, p, flags);
770 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
773 sched_info_dequeued(rq, p);
774 p->sched_class->dequeue_task(rq, p, flags);
777 void activate_task(struct rq *rq, struct task_struct *p, int flags)
779 if (task_contributes_to_load(p))
780 rq->nr_uninterruptible--;
782 enqueue_task(rq, p, flags);
785 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
787 if (task_contributes_to_load(p))
788 rq->nr_uninterruptible++;
790 dequeue_task(rq, p, flags);
793 static void update_rq_clock_task(struct rq *rq, s64 delta)
796 * In theory, the compile should just see 0 here, and optimize out the call
797 * to sched_rt_avg_update. But I don't trust it...
799 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
800 s64 steal = 0, irq_delta = 0;
802 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
803 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
806 * Since irq_time is only updated on {soft,}irq_exit, we might run into
807 * this case when a previous update_rq_clock() happened inside a
810 * When this happens, we stop ->clock_task and only update the
811 * prev_irq_time stamp to account for the part that fit, so that a next
812 * update will consume the rest. This ensures ->clock_task is
815 * It does however cause some slight miss-attribution of {soft,}irq
816 * time, a more accurate solution would be to update the irq_time using
817 * the current rq->clock timestamp, except that would require using
820 if (irq_delta > delta)
823 rq->prev_irq_time += irq_delta;
826 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
827 if (static_key_false((¶virt_steal_rq_enabled))) {
830 steal = paravirt_steal_clock(cpu_of(rq));
831 steal -= rq->prev_steal_time_rq;
833 if (unlikely(steal > delta))
836 st = steal_ticks(steal);
837 steal = st * TICK_NSEC;
839 rq->prev_steal_time_rq += steal;
845 rq->clock_task += delta;
847 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
848 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
849 sched_rt_avg_update(rq, irq_delta + steal);
853 void sched_set_stop_task(int cpu, struct task_struct *stop)
855 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
856 struct task_struct *old_stop = cpu_rq(cpu)->stop;
860 * Make it appear like a SCHED_FIFO task, its something
861 * userspace knows about and won't get confused about.
863 * Also, it will make PI more or less work without too
864 * much confusion -- but then, stop work should not
865 * rely on PI working anyway.
867 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
869 stop->sched_class = &stop_sched_class;
872 cpu_rq(cpu)->stop = stop;
876 * Reset it back to a normal scheduling class so that
877 * it can die in pieces.
879 old_stop->sched_class = &rt_sched_class;
884 * __normal_prio - return the priority that is based on the static prio
886 static inline int __normal_prio(struct task_struct *p)
888 return p->static_prio;
892 * Calculate the expected normal priority: i.e. priority
893 * without taking RT-inheritance into account. Might be
894 * boosted by interactivity modifiers. Changes upon fork,
895 * setprio syscalls, and whenever the interactivity
896 * estimator recalculates.
898 static inline int normal_prio(struct task_struct *p)
902 if (task_has_rt_policy(p))
903 prio = MAX_RT_PRIO-1 - p->rt_priority;
905 prio = __normal_prio(p);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct *p)
918 p->normal_prio = normal_prio(p);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p->prio))
925 return p->normal_prio;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio)
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
986 !(task_preempt_count(p) & PREEMPT_ACTIVE));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1000 lockdep_is_held(&task_rq(p)->lock)));
1004 trace_sched_migrate_task(p, new_cpu);
1006 if (task_cpu(p) != new_cpu) {
1007 if (p->sched_class->migrate_task_rq)
1008 p->sched_class->migrate_task_rq(p, new_cpu);
1009 p->se.nr_migrations++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1013 __set_task_cpu(p, new_cpu);
1016 static void __migrate_swap_task(struct task_struct *p, int cpu)
1019 struct rq *src_rq, *dst_rq;
1021 src_rq = task_rq(p);
1022 dst_rq = cpu_rq(cpu);
1024 deactivate_task(src_rq, p, 0);
1025 set_task_cpu(p, cpu);
1026 activate_task(dst_rq, p, 0);
1027 check_preempt_curr(dst_rq, p, 0);
1030 * Task isn't running anymore; make it appear like we migrated
1031 * it before it went to sleep. This means on wakeup we make the
1032 * previous cpu our targer instead of where it really is.
1038 struct migration_swap_arg {
1039 struct task_struct *src_task, *dst_task;
1040 int src_cpu, dst_cpu;
1043 static int migrate_swap_stop(void *data)
1045 struct migration_swap_arg *arg = data;
1046 struct rq *src_rq, *dst_rq;
1049 src_rq = cpu_rq(arg->src_cpu);
1050 dst_rq = cpu_rq(arg->dst_cpu);
1052 double_raw_lock(&arg->src_task->pi_lock,
1053 &arg->dst_task->pi_lock);
1054 double_rq_lock(src_rq, dst_rq);
1055 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1058 if (task_cpu(arg->src_task) != arg->src_cpu)
1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1067 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1068 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1073 double_rq_unlock(src_rq, dst_rq);
1074 raw_spin_unlock(&arg->dst_task->pi_lock);
1075 raw_spin_unlock(&arg->src_task->pi_lock);
1081 * Cross migrate two tasks
1083 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1085 struct migration_swap_arg arg;
1088 arg = (struct migration_swap_arg){
1090 .src_cpu = task_cpu(cur),
1092 .dst_cpu = task_cpu(p),
1095 if (arg.src_cpu == arg.dst_cpu)
1099 * These three tests are all lockless; this is OK since all of them
1100 * will be re-checked with proper locks held further down the line.
1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1111 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1117 struct migration_arg {
1118 struct task_struct *task;
1122 static int migration_cpu_stop(void *data);
1125 * wait_task_inactive - wait for a thread to unschedule.
1127 * If @match_state is nonzero, it's the @p->state value just checked and
1128 * not expected to change. If it changes, i.e. @p might have woken up,
1129 * then return zero. When we succeed in waiting for @p to be off its CPU,
1130 * we return a positive number (its total switch count). If a second call
1131 * a short while later returns the same number, the caller can be sure that
1132 * @p has remained unscheduled the whole time.
1134 * The caller must ensure that the task *will* unschedule sometime soon,
1135 * else this function might spin for a *long* time. This function can't
1136 * be called with interrupts off, or it may introduce deadlock with
1137 * smp_call_function() if an IPI is sent by the same process we are
1138 * waiting to become inactive.
1140 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1142 unsigned long flags;
1149 * We do the initial early heuristics without holding
1150 * any task-queue locks at all. We'll only try to get
1151 * the runqueue lock when things look like they will
1157 * If the task is actively running on another CPU
1158 * still, just relax and busy-wait without holding
1161 * NOTE! Since we don't hold any locks, it's not
1162 * even sure that "rq" stays as the right runqueue!
1163 * But we don't care, since "task_running()" will
1164 * return false if the runqueue has changed and p
1165 * is actually now running somewhere else!
1167 while (task_running(rq, p)) {
1168 if (match_state && unlikely(p->state != match_state))
1174 * Ok, time to look more closely! We need the rq
1175 * lock now, to be *sure*. If we're wrong, we'll
1176 * just go back and repeat.
1178 rq = task_rq_lock(p, &flags);
1179 trace_sched_wait_task(p);
1180 running = task_running(rq, p);
1183 if (!match_state || p->state == match_state)
1184 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1185 task_rq_unlock(rq, p, &flags);
1188 * If it changed from the expected state, bail out now.
1190 if (unlikely(!ncsw))
1194 * Was it really running after all now that we
1195 * checked with the proper locks actually held?
1197 * Oops. Go back and try again..
1199 if (unlikely(running)) {
1205 * It's not enough that it's not actively running,
1206 * it must be off the runqueue _entirely_, and not
1209 * So if it was still runnable (but just not actively
1210 * running right now), it's preempted, and we should
1211 * yield - it could be a while.
1213 if (unlikely(on_rq)) {
1214 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1216 set_current_state(TASK_UNINTERRUPTIBLE);
1217 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1222 * Ahh, all good. It wasn't running, and it wasn't
1223 * runnable, which means that it will never become
1224 * running in the future either. We're all done!
1233 * kick_process - kick a running thread to enter/exit the kernel
1234 * @p: the to-be-kicked thread
1236 * Cause a process which is running on another CPU to enter
1237 * kernel-mode, without any delay. (to get signals handled.)
1239 * NOTE: this function doesn't have to take the runqueue lock,
1240 * because all it wants to ensure is that the remote task enters
1241 * the kernel. If the IPI races and the task has been migrated
1242 * to another CPU then no harm is done and the purpose has been
1245 void kick_process(struct task_struct *p)
1251 if ((cpu != smp_processor_id()) && task_curr(p))
1252 smp_send_reschedule(cpu);
1255 EXPORT_SYMBOL_GPL(kick_process);
1256 #endif /* CONFIG_SMP */
1260 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1262 static int select_fallback_rq(int cpu, struct task_struct *p)
1264 int nid = cpu_to_node(cpu);
1265 const struct cpumask *nodemask = NULL;
1266 enum { cpuset, possible, fail } state = cpuset;
1270 * If the node that the cpu is on has been offlined, cpu_to_node()
1271 * will return -1. There is no cpu on the node, and we should
1272 * select the cpu on the other node.
1275 nodemask = cpumask_of_node(nid);
1277 /* Look for allowed, online CPU in same node. */
1278 for_each_cpu(dest_cpu, nodemask) {
1279 if (!cpu_online(dest_cpu))
1281 if (!cpu_active(dest_cpu))
1283 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1289 /* Any allowed, online CPU? */
1290 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1291 if (!cpu_online(dest_cpu))
1293 if (!cpu_active(dest_cpu))
1300 /* No more Mr. Nice Guy. */
1301 cpuset_cpus_allowed_fallback(p);
1306 do_set_cpus_allowed(p, cpu_possible_mask);
1317 if (state != cpuset) {
1319 * Don't tell them about moving exiting tasks or
1320 * kernel threads (both mm NULL), since they never
1323 if (p->mm && printk_ratelimit()) {
1324 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1325 task_pid_nr(p), p->comm, cpu);
1333 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1336 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1338 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1341 * In order not to call set_task_cpu() on a blocking task we need
1342 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1345 * Since this is common to all placement strategies, this lives here.
1347 * [ this allows ->select_task() to simply return task_cpu(p) and
1348 * not worry about this generic constraint ]
1350 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1352 cpu = select_fallback_rq(task_cpu(p), p);
1357 static void update_avg(u64 *avg, u64 sample)
1359 s64 diff = sample - *avg;
1365 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1367 #ifdef CONFIG_SCHEDSTATS
1368 struct rq *rq = this_rq();
1371 int this_cpu = smp_processor_id();
1373 if (cpu == this_cpu) {
1374 schedstat_inc(rq, ttwu_local);
1375 schedstat_inc(p, se.statistics.nr_wakeups_local);
1377 struct sched_domain *sd;
1379 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1381 for_each_domain(this_cpu, sd) {
1382 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1383 schedstat_inc(sd, ttwu_wake_remote);
1390 if (wake_flags & WF_MIGRATED)
1391 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1393 #endif /* CONFIG_SMP */
1395 schedstat_inc(rq, ttwu_count);
1396 schedstat_inc(p, se.statistics.nr_wakeups);
1398 if (wake_flags & WF_SYNC)
1399 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1401 #endif /* CONFIG_SCHEDSTATS */
1404 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1406 activate_task(rq, p, en_flags);
1409 /* if a worker is waking up, notify workqueue */
1410 if (p->flags & PF_WQ_WORKER)
1411 wq_worker_waking_up(p, cpu_of(rq));
1415 * Mark the task runnable and perform wakeup-preemption.
1418 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1420 check_preempt_curr(rq, p, wake_flags);
1421 trace_sched_wakeup(p, true);
1423 p->state = TASK_RUNNING;
1425 if (p->sched_class->task_woken)
1426 p->sched_class->task_woken(rq, p);
1428 if (rq->idle_stamp) {
1429 u64 delta = rq_clock(rq) - rq->idle_stamp;
1430 u64 max = 2*rq->max_idle_balance_cost;
1432 update_avg(&rq->avg_idle, delta);
1434 if (rq->avg_idle > max)
1443 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1446 if (p->sched_contributes_to_load)
1447 rq->nr_uninterruptible--;
1450 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1451 ttwu_do_wakeup(rq, p, wake_flags);
1455 * Called in case the task @p isn't fully descheduled from its runqueue,
1456 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1457 * since all we need to do is flip p->state to TASK_RUNNING, since
1458 * the task is still ->on_rq.
1460 static int ttwu_remote(struct task_struct *p, int wake_flags)
1465 rq = __task_rq_lock(p);
1467 /* check_preempt_curr() may use rq clock */
1468 update_rq_clock(rq);
1469 ttwu_do_wakeup(rq, p, wake_flags);
1472 __task_rq_unlock(rq);
1478 static void sched_ttwu_pending(void)
1480 struct rq *rq = this_rq();
1481 struct llist_node *llist = llist_del_all(&rq->wake_list);
1482 struct task_struct *p;
1484 raw_spin_lock(&rq->lock);
1487 p = llist_entry(llist, struct task_struct, wake_entry);
1488 llist = llist_next(llist);
1489 ttwu_do_activate(rq, p, 0);
1492 raw_spin_unlock(&rq->lock);
1495 void scheduler_ipi(void)
1498 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1499 * TIF_NEED_RESCHED remotely (for the first time) will also send
1502 if (tif_need_resched())
1503 set_preempt_need_resched();
1505 if (llist_empty(&this_rq()->wake_list)
1506 && !tick_nohz_full_cpu(smp_processor_id())
1507 && !got_nohz_idle_kick())
1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1512 * traditionally all their work was done from the interrupt return
1513 * path. Now that we actually do some work, we need to make sure
1516 * Some archs already do call them, luckily irq_enter/exit nest
1519 * Arguably we should visit all archs and update all handlers,
1520 * however a fair share of IPIs are still resched only so this would
1521 * somewhat pessimize the simple resched case.
1524 tick_nohz_full_check();
1525 sched_ttwu_pending();
1528 * Check if someone kicked us for doing the nohz idle load balance.
1530 if (unlikely(got_nohz_idle_kick())) {
1531 this_rq()->idle_balance = 1;
1532 raise_softirq_irqoff(SCHED_SOFTIRQ);
1537 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1540 smp_send_reschedule(cpu);
1543 bool cpus_share_cache(int this_cpu, int that_cpu)
1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1547 #endif /* CONFIG_SMP */
1549 static void ttwu_queue(struct task_struct *p, int cpu)
1551 struct rq *rq = cpu_rq(cpu);
1553 #if defined(CONFIG_SMP)
1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1556 ttwu_queue_remote(p, cpu);
1561 raw_spin_lock(&rq->lock);
1562 ttwu_do_activate(rq, p, 0);
1563 raw_spin_unlock(&rq->lock);
1567 * try_to_wake_up - wake up a thread
1568 * @p: the thread to be awakened
1569 * @state: the mask of task states that can be woken
1570 * @wake_flags: wake modifier flags (WF_*)
1572 * Put it on the run-queue if it's not already there. The "current"
1573 * thread is always on the run-queue (except when the actual
1574 * re-schedule is in progress), and as such you're allowed to do
1575 * the simpler "current->state = TASK_RUNNING" to mark yourself
1576 * runnable without the overhead of this.
1578 * Return: %true if @p was woken up, %false if it was already running.
1579 * or @state didn't match @p's state.
1582 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1584 unsigned long flags;
1585 int cpu, success = 0;
1588 * If we are going to wake up a thread waiting for CONDITION we
1589 * need to ensure that CONDITION=1 done by the caller can not be
1590 * reordered with p->state check below. This pairs with mb() in
1591 * set_current_state() the waiting thread does.
1593 smp_mb__before_spinlock();
1594 raw_spin_lock_irqsave(&p->pi_lock, flags);
1595 if (!(p->state & state))
1598 success = 1; /* we're going to change ->state */
1601 if (p->on_rq && ttwu_remote(p, wake_flags))
1606 * If the owning (remote) cpu is still in the middle of schedule() with
1607 * this task as prev, wait until its done referencing the task.
1612 * Pairs with the smp_wmb() in finish_lock_switch().
1616 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1617 p->state = TASK_WAKING;
1619 if (p->sched_class->task_waking)
1620 p->sched_class->task_waking(p);
1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1623 if (task_cpu(p) != cpu) {
1624 wake_flags |= WF_MIGRATED;
1625 set_task_cpu(p, cpu);
1627 #endif /* CONFIG_SMP */
1631 ttwu_stat(p, cpu, wake_flags);
1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1639 * try_to_wake_up_local - try to wake up a local task with rq lock held
1640 * @p: the thread to be awakened
1642 * Put @p on the run-queue if it's not already there. The caller must
1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1646 static void try_to_wake_up_local(struct task_struct *p)
1648 struct rq *rq = task_rq(p);
1650 if (WARN_ON_ONCE(rq != this_rq()) ||
1651 WARN_ON_ONCE(p == current))
1654 lockdep_assert_held(&rq->lock);
1656 if (!raw_spin_trylock(&p->pi_lock)) {
1657 raw_spin_unlock(&rq->lock);
1658 raw_spin_lock(&p->pi_lock);
1659 raw_spin_lock(&rq->lock);
1662 if (!(p->state & TASK_NORMAL))
1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1668 ttwu_do_wakeup(rq, p, 0);
1669 ttwu_stat(p, smp_processor_id(), 0);
1671 raw_spin_unlock(&p->pi_lock);
1675 * wake_up_process - Wake up a specific process
1676 * @p: The process to be woken up.
1678 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * Return: 1 if the process was woken up, 0 if it was already running.
1683 * It may be assumed that this function implies a write memory barrier before
1684 * changing the task state if and only if any tasks are woken up.
1686 int wake_up_process(struct task_struct *p)
1688 WARN_ON(task_is_stopped_or_traced(p));
1689 return try_to_wake_up(p, TASK_NORMAL, 0);
1691 EXPORT_SYMBOL(wake_up_process);
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1695 return try_to_wake_up(p, state, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1720 INIT_LIST_HEAD(&p->rt.run_list);
1722 #ifdef CONFIG_PREEMPT_NOTIFIERS
1723 INIT_HLIST_HEAD(&p->preempt_notifiers);
1726 #ifdef CONFIG_NUMA_BALANCING
1727 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1728 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1729 p->mm->numa_scan_seq = 0;
1732 if (clone_flags & CLONE_VM)
1733 p->numa_preferred_nid = current->numa_preferred_nid;
1735 p->numa_preferred_nid = -1;
1737 p->node_stamp = 0ULL;
1738 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1739 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1740 p->numa_work.next = &p->numa_work;
1741 p->numa_faults = NULL;
1742 p->numa_faults_buffer = NULL;
1744 INIT_LIST_HEAD(&p->numa_entry);
1745 p->numa_group = NULL;
1746 #endif /* CONFIG_NUMA_BALANCING */
1749 #ifdef CONFIG_NUMA_BALANCING
1750 #ifdef CONFIG_SCHED_DEBUG
1751 void set_numabalancing_state(bool enabled)
1754 sched_feat_set("NUMA");
1756 sched_feat_set("NO_NUMA");
1759 __read_mostly bool numabalancing_enabled;
1761 void set_numabalancing_state(bool enabled)
1763 numabalancing_enabled = enabled;
1765 #endif /* CONFIG_SCHED_DEBUG */
1766 #endif /* CONFIG_NUMA_BALANCING */
1769 * fork()/clone()-time setup:
1771 void sched_fork(unsigned long clone_flags, struct task_struct *p)
1773 unsigned long flags;
1774 int cpu = get_cpu();
1776 __sched_fork(clone_flags, p);
1778 * We mark the process as running here. This guarantees that
1779 * nobody will actually run it, and a signal or other external
1780 * event cannot wake it up and insert it on the runqueue either.
1782 p->state = TASK_RUNNING;
1785 * Make sure we do not leak PI boosting priority to the child.
1787 p->prio = current->normal_prio;
1790 * Revert to default priority/policy on fork if requested.
1792 if (unlikely(p->sched_reset_on_fork)) {
1793 if (task_has_rt_policy(p)) {
1794 p->policy = SCHED_NORMAL;
1795 p->static_prio = NICE_TO_PRIO(0);
1797 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1798 p->static_prio = NICE_TO_PRIO(0);
1800 p->prio = p->normal_prio = __normal_prio(p);
1804 * We don't need the reset flag anymore after the fork. It has
1805 * fulfilled its duty:
1807 p->sched_reset_on_fork = 0;
1810 if (!rt_prio(p->prio))
1811 p->sched_class = &fair_sched_class;
1813 if (p->sched_class->task_fork)
1814 p->sched_class->task_fork(p);
1817 * The child is not yet in the pid-hash so no cgroup attach races,
1818 * and the cgroup is pinned to this child due to cgroup_fork()
1819 * is ran before sched_fork().
1821 * Silence PROVE_RCU.
1823 raw_spin_lock_irqsave(&p->pi_lock, flags);
1824 set_task_cpu(p, cpu);
1825 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1827 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1828 if (likely(sched_info_on()))
1829 memset(&p->sched_info, 0, sizeof(p->sched_info));
1831 #if defined(CONFIG_SMP)
1834 init_task_preempt_count(p);
1836 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1843 * wake_up_new_task - wake up a newly created task for the first time.
1845 * This function will do some initial scheduler statistics housekeeping
1846 * that must be done for every newly created context, then puts the task
1847 * on the runqueue and wakes it.
1849 void wake_up_new_task(struct task_struct *p)
1851 unsigned long flags;
1854 raw_spin_lock_irqsave(&p->pi_lock, flags);
1857 * Fork balancing, do it here and not earlier because:
1858 * - cpus_allowed can change in the fork path
1859 * - any previously selected cpu might disappear through hotplug
1861 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1864 /* Initialize new task's runnable average */
1865 init_task_runnable_average(p);
1866 rq = __task_rq_lock(p);
1867 activate_task(rq, p, 0);
1869 trace_sched_wakeup_new(p, true);
1870 check_preempt_curr(rq, p, WF_FORK);
1872 if (p->sched_class->task_woken)
1873 p->sched_class->task_woken(rq, p);
1875 task_rq_unlock(rq, p, &flags);
1878 #ifdef CONFIG_PREEMPT_NOTIFIERS
1881 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1882 * @notifier: notifier struct to register
1884 void preempt_notifier_register(struct preempt_notifier *notifier)
1886 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1888 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1891 * preempt_notifier_unregister - no longer interested in preemption notifications
1892 * @notifier: notifier struct to unregister
1894 * This is safe to call from within a preemption notifier.
1896 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1898 hlist_del(¬ifier->link);
1900 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1902 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1904 struct preempt_notifier *notifier;
1906 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1907 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1911 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1912 struct task_struct *next)
1914 struct preempt_notifier *notifier;
1916 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1917 notifier->ops->sched_out(notifier, next);
1920 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1922 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1927 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1928 struct task_struct *next)
1932 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1935 * prepare_task_switch - prepare to switch tasks
1936 * @rq: the runqueue preparing to switch
1937 * @prev: the current task that is being switched out
1938 * @next: the task we are going to switch to.
1940 * This is called with the rq lock held and interrupts off. It must
1941 * be paired with a subsequent finish_task_switch after the context
1944 * prepare_task_switch sets up locking and calls architecture specific
1948 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1949 struct task_struct *next)
1951 trace_sched_switch(prev, next);
1952 sched_info_switch(rq, prev, next);
1953 perf_event_task_sched_out(prev, next);
1954 fire_sched_out_preempt_notifiers(prev, next);
1955 prepare_lock_switch(rq, next);
1956 prepare_arch_switch(next);
1960 * finish_task_switch - clean up after a task-switch
1961 * @rq: runqueue associated with task-switch
1962 * @prev: the thread we just switched away from.
1964 * finish_task_switch must be called after the context switch, paired
1965 * with a prepare_task_switch call before the context switch.
1966 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1967 * and do any other architecture-specific cleanup actions.
1969 * Note that we may have delayed dropping an mm in context_switch(). If
1970 * so, we finish that here outside of the runqueue lock. (Doing it
1971 * with the lock held can cause deadlocks; see schedule() for
1974 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1975 __releases(rq->lock)
1977 struct mm_struct *mm = rq->prev_mm;
1983 * A task struct has one reference for the use as "current".
1984 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1985 * schedule one last time. The schedule call will never return, and
1986 * the scheduled task must drop that reference.
1987 * The test for TASK_DEAD must occur while the runqueue locks are
1988 * still held, otherwise prev could be scheduled on another cpu, die
1989 * there before we look at prev->state, and then the reference would
1991 * Manfred Spraul <manfred@colorfullife.com>
1993 prev_state = prev->state;
1994 vtime_task_switch(prev);
1995 finish_arch_switch(prev);
1996 perf_event_task_sched_in(prev, current);
1997 finish_lock_switch(rq, prev);
1998 finish_arch_post_lock_switch();
2000 fire_sched_in_preempt_notifiers(current);
2003 if (unlikely(prev_state == TASK_DEAD)) {
2004 task_numa_free(prev);
2007 * Remove function-return probe instances associated with this
2008 * task and put them back on the free list.
2010 kprobe_flush_task(prev);
2011 put_task_struct(prev);
2014 tick_nohz_task_switch(current);
2019 /* assumes rq->lock is held */
2020 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2022 if (prev->sched_class->pre_schedule)
2023 prev->sched_class->pre_schedule(rq, prev);
2026 /* rq->lock is NOT held, but preemption is disabled */
2027 static inline void post_schedule(struct rq *rq)
2029 if (rq->post_schedule) {
2030 unsigned long flags;
2032 raw_spin_lock_irqsave(&rq->lock, flags);
2033 if (rq->curr->sched_class->post_schedule)
2034 rq->curr->sched_class->post_schedule(rq);
2035 raw_spin_unlock_irqrestore(&rq->lock, flags);
2037 rq->post_schedule = 0;
2043 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2047 static inline void post_schedule(struct rq *rq)
2054 * schedule_tail - first thing a freshly forked thread must call.
2055 * @prev: the thread we just switched away from.
2057 asmlinkage void schedule_tail(struct task_struct *prev)
2058 __releases(rq->lock)
2060 struct rq *rq = this_rq();
2062 finish_task_switch(rq, prev);
2065 * FIXME: do we need to worry about rq being invalidated by the
2070 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2071 /* In this case, finish_task_switch does not reenable preemption */
2074 if (current->set_child_tid)
2075 put_user(task_pid_vnr(current), current->set_child_tid);
2079 * context_switch - switch to the new MM and the new
2080 * thread's register state.
2083 context_switch(struct rq *rq, struct task_struct *prev,
2084 struct task_struct *next)
2086 struct mm_struct *mm, *oldmm;
2088 prepare_task_switch(rq, prev, next);
2091 oldmm = prev->active_mm;
2093 * For paravirt, this is coupled with an exit in switch_to to
2094 * combine the page table reload and the switch backend into
2097 arch_start_context_switch(prev);
2100 next->active_mm = oldmm;
2101 atomic_inc(&oldmm->mm_count);
2102 enter_lazy_tlb(oldmm, next);
2104 switch_mm(oldmm, mm, next);
2107 prev->active_mm = NULL;
2108 rq->prev_mm = oldmm;
2111 * Since the runqueue lock will be released by the next
2112 * task (which is an invalid locking op but in the case
2113 * of the scheduler it's an obvious special-case), so we
2114 * do an early lockdep release here:
2116 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2117 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2120 context_tracking_task_switch(prev, next);
2121 /* Here we just switch the register state and the stack. */
2122 switch_to(prev, next, prev);
2126 * this_rq must be evaluated again because prev may have moved
2127 * CPUs since it called schedule(), thus the 'rq' on its stack
2128 * frame will be invalid.
2130 finish_task_switch(this_rq(), prev);
2134 * nr_running and nr_context_switches:
2136 * externally visible scheduler statistics: current number of runnable
2137 * threads, total number of context switches performed since bootup.
2139 unsigned long nr_running(void)
2141 unsigned long i, sum = 0;
2143 for_each_online_cpu(i)
2144 sum += cpu_rq(i)->nr_running;
2149 unsigned long long nr_context_switches(void)
2152 unsigned long long sum = 0;
2154 for_each_possible_cpu(i)
2155 sum += cpu_rq(i)->nr_switches;
2160 unsigned long nr_iowait(void)
2162 unsigned long i, sum = 0;
2164 for_each_possible_cpu(i)
2165 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2170 unsigned long nr_iowait_cpu(int cpu)
2172 struct rq *this = cpu_rq(cpu);
2173 return atomic_read(&this->nr_iowait);
2179 * sched_exec - execve() is a valuable balancing opportunity, because at
2180 * this point the task has the smallest effective memory and cache footprint.
2182 void sched_exec(void)
2184 struct task_struct *p = current;
2185 unsigned long flags;
2188 raw_spin_lock_irqsave(&p->pi_lock, flags);
2189 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2190 if (dest_cpu == smp_processor_id())
2193 if (likely(cpu_active(dest_cpu))) {
2194 struct migration_arg arg = { p, dest_cpu };
2196 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2197 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2201 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2206 DEFINE_PER_CPU(struct kernel_stat, kstat);
2207 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2209 EXPORT_PER_CPU_SYMBOL(kstat);
2210 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2213 * Return any ns on the sched_clock that have not yet been accounted in
2214 * @p in case that task is currently running.
2216 * Called with task_rq_lock() held on @rq.
2218 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2222 if (task_current(rq, p)) {
2223 update_rq_clock(rq);
2224 ns = rq_clock_task(rq) - p->se.exec_start;
2232 unsigned long long task_delta_exec(struct task_struct *p)
2234 unsigned long flags;
2238 rq = task_rq_lock(p, &flags);
2239 ns = do_task_delta_exec(p, rq);
2240 task_rq_unlock(rq, p, &flags);
2246 * Return accounted runtime for the task.
2247 * In case the task is currently running, return the runtime plus current's
2248 * pending runtime that have not been accounted yet.
2250 unsigned long long task_sched_runtime(struct task_struct *p)
2252 unsigned long flags;
2256 rq = task_rq_lock(p, &flags);
2257 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2258 task_rq_unlock(rq, p, &flags);
2264 * This function gets called by the timer code, with HZ frequency.
2265 * We call it with interrupts disabled.
2267 void scheduler_tick(void)
2269 int cpu = smp_processor_id();
2270 struct rq *rq = cpu_rq(cpu);
2271 struct task_struct *curr = rq->curr;
2275 raw_spin_lock(&rq->lock);
2276 update_rq_clock(rq);
2277 curr->sched_class->task_tick(rq, curr, 0);
2278 update_cpu_load_active(rq);
2279 raw_spin_unlock(&rq->lock);
2281 perf_event_task_tick();
2284 rq->idle_balance = idle_cpu(cpu);
2285 trigger_load_balance(rq, cpu);
2287 rq_last_tick_reset(rq);
2290 #ifdef CONFIG_NO_HZ_FULL
2292 * scheduler_tick_max_deferment
2294 * Keep at least one tick per second when a single
2295 * active task is running because the scheduler doesn't
2296 * yet completely support full dynticks environment.
2298 * This makes sure that uptime, CFS vruntime, load
2299 * balancing, etc... continue to move forward, even
2300 * with a very low granularity.
2302 * Return: Maximum deferment in nanoseconds.
2304 u64 scheduler_tick_max_deferment(void)
2306 struct rq *rq = this_rq();
2307 unsigned long next, now = ACCESS_ONCE(jiffies);
2309 next = rq->last_sched_tick + HZ;
2311 if (time_before_eq(next, now))
2314 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2318 notrace unsigned long get_parent_ip(unsigned long addr)
2320 if (in_lock_functions(addr)) {
2321 addr = CALLER_ADDR2;
2322 if (in_lock_functions(addr))
2323 addr = CALLER_ADDR3;
2328 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2329 defined(CONFIG_PREEMPT_TRACER))
2331 void __kprobes preempt_count_add(int val)
2333 #ifdef CONFIG_DEBUG_PREEMPT
2337 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2340 __preempt_count_add(val);
2341 #ifdef CONFIG_DEBUG_PREEMPT
2343 * Spinlock count overflowing soon?
2345 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2348 if (preempt_count() == val)
2349 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2351 EXPORT_SYMBOL(preempt_count_add);
2353 void __kprobes preempt_count_sub(int val)
2355 #ifdef CONFIG_DEBUG_PREEMPT
2359 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2362 * Is the spinlock portion underflowing?
2364 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2365 !(preempt_count() & PREEMPT_MASK)))
2369 if (preempt_count() == val)
2370 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2371 __preempt_count_sub(val);
2373 EXPORT_SYMBOL(preempt_count_sub);
2378 * Print scheduling while atomic bug:
2380 static noinline void __schedule_bug(struct task_struct *prev)
2382 if (oops_in_progress)
2385 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2386 prev->comm, prev->pid, preempt_count());
2388 debug_show_held_locks(prev);
2390 if (irqs_disabled())
2391 print_irqtrace_events(prev);
2393 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2397 * Various schedule()-time debugging checks and statistics:
2399 static inline void schedule_debug(struct task_struct *prev)
2402 * Test if we are atomic. Since do_exit() needs to call into
2403 * schedule() atomically, we ignore that path for now.
2404 * Otherwise, whine if we are scheduling when we should not be.
2406 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2407 __schedule_bug(prev);
2410 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2412 schedstat_inc(this_rq(), sched_count);
2415 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2417 if (prev->on_rq || rq->skip_clock_update < 0)
2418 update_rq_clock(rq);
2419 prev->sched_class->put_prev_task(rq, prev);
2423 * Pick up the highest-prio task:
2425 static inline struct task_struct *
2426 pick_next_task(struct rq *rq)
2428 const struct sched_class *class;
2429 struct task_struct *p;
2432 * Optimization: we know that if all tasks are in
2433 * the fair class we can call that function directly:
2435 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2436 p = fair_sched_class.pick_next_task(rq);
2441 for_each_class(class) {
2442 p = class->pick_next_task(rq);
2447 BUG(); /* the idle class will always have a runnable task */
2451 * __schedule() is the main scheduler function.
2453 * The main means of driving the scheduler and thus entering this function are:
2455 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2457 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2458 * paths. For example, see arch/x86/entry_64.S.
2460 * To drive preemption between tasks, the scheduler sets the flag in timer
2461 * interrupt handler scheduler_tick().
2463 * 3. Wakeups don't really cause entry into schedule(). They add a
2464 * task to the run-queue and that's it.
2466 * Now, if the new task added to the run-queue preempts the current
2467 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2468 * called on the nearest possible occasion:
2470 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2472 * - in syscall or exception context, at the next outmost
2473 * preempt_enable(). (this might be as soon as the wake_up()'s
2476 * - in IRQ context, return from interrupt-handler to
2477 * preemptible context
2479 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2482 * - cond_resched() call
2483 * - explicit schedule() call
2484 * - return from syscall or exception to user-space
2485 * - return from interrupt-handler to user-space
2487 static void __sched __schedule(void)
2489 struct task_struct *prev, *next;
2490 unsigned long *switch_count;
2496 cpu = smp_processor_id();
2498 rcu_note_context_switch(cpu);
2501 schedule_debug(prev);
2503 if (sched_feat(HRTICK))
2507 * Make sure that signal_pending_state()->signal_pending() below
2508 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2509 * done by the caller to avoid the race with signal_wake_up().
2511 smp_mb__before_spinlock();
2512 raw_spin_lock_irq(&rq->lock);
2514 switch_count = &prev->nivcsw;
2515 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2516 if (unlikely(signal_pending_state(prev->state, prev))) {
2517 prev->state = TASK_RUNNING;
2519 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2523 * If a worker went to sleep, notify and ask workqueue
2524 * whether it wants to wake up a task to maintain
2527 if (prev->flags & PF_WQ_WORKER) {
2528 struct task_struct *to_wakeup;
2530 to_wakeup = wq_worker_sleeping(prev, cpu);
2532 try_to_wake_up_local(to_wakeup);
2535 switch_count = &prev->nvcsw;
2538 pre_schedule(rq, prev);
2540 if (unlikely(!rq->nr_running))
2541 idle_balance(cpu, rq);
2543 put_prev_task(rq, prev);
2544 next = pick_next_task(rq);
2545 clear_tsk_need_resched(prev);
2546 clear_preempt_need_resched();
2547 rq->skip_clock_update = 0;
2549 if (likely(prev != next)) {
2554 context_switch(rq, prev, next); /* unlocks the rq */
2556 * The context switch have flipped the stack from under us
2557 * and restored the local variables which were saved when
2558 * this task called schedule() in the past. prev == current
2559 * is still correct, but it can be moved to another cpu/rq.
2561 cpu = smp_processor_id();
2564 raw_spin_unlock_irq(&rq->lock);
2568 sched_preempt_enable_no_resched();
2573 static inline void sched_submit_work(struct task_struct *tsk)
2575 if (!tsk->state || tsk_is_pi_blocked(tsk))
2578 * If we are going to sleep and we have plugged IO queued,
2579 * make sure to submit it to avoid deadlocks.
2581 if (blk_needs_flush_plug(tsk))
2582 blk_schedule_flush_plug(tsk);
2585 asmlinkage void __sched schedule(void)
2587 struct task_struct *tsk = current;
2589 sched_submit_work(tsk);
2592 EXPORT_SYMBOL(schedule);
2594 #ifdef CONFIG_CONTEXT_TRACKING
2595 asmlinkage void __sched schedule_user(void)
2598 * If we come here after a random call to set_need_resched(),
2599 * or we have been woken up remotely but the IPI has not yet arrived,
2600 * we haven't yet exited the RCU idle mode. Do it here manually until
2601 * we find a better solution.
2610 * schedule_preempt_disabled - called with preemption disabled
2612 * Returns with preemption disabled. Note: preempt_count must be 1
2614 void __sched schedule_preempt_disabled(void)
2616 sched_preempt_enable_no_resched();
2621 #ifdef CONFIG_PREEMPT
2623 * this is the entry point to schedule() from in-kernel preemption
2624 * off of preempt_enable. Kernel preemptions off return from interrupt
2625 * occur there and call schedule directly.
2627 asmlinkage void __sched notrace preempt_schedule(void)
2630 * If there is a non-zero preempt_count or interrupts are disabled,
2631 * we do not want to preempt the current task. Just return..
2633 if (likely(!preemptible()))
2637 __preempt_count_add(PREEMPT_ACTIVE);
2639 __preempt_count_sub(PREEMPT_ACTIVE);
2642 * Check again in case we missed a preemption opportunity
2643 * between schedule and now.
2646 } while (need_resched());
2648 EXPORT_SYMBOL(preempt_schedule);
2651 * this is the entry point to schedule() from kernel preemption
2652 * off of irq context.
2653 * Note, that this is called and return with irqs disabled. This will
2654 * protect us against recursive calling from irq.
2656 asmlinkage void __sched preempt_schedule_irq(void)
2658 enum ctx_state prev_state;
2660 /* Catch callers which need to be fixed */
2661 BUG_ON(preempt_count() || !irqs_disabled());
2663 prev_state = exception_enter();
2666 __preempt_count_add(PREEMPT_ACTIVE);
2669 local_irq_disable();
2670 __preempt_count_sub(PREEMPT_ACTIVE);
2673 * Check again in case we missed a preemption opportunity
2674 * between schedule and now.
2677 } while (need_resched());
2679 exception_exit(prev_state);
2682 #endif /* CONFIG_PREEMPT */
2684 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2687 return try_to_wake_up(curr->private, mode, wake_flags);
2689 EXPORT_SYMBOL(default_wake_function);
2692 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2694 unsigned long flags;
2697 init_waitqueue_entry(&wait, current);
2699 __set_current_state(state);
2701 spin_lock_irqsave(&q->lock, flags);
2702 __add_wait_queue(q, &wait);
2703 spin_unlock(&q->lock);
2704 timeout = schedule_timeout(timeout);
2705 spin_lock_irq(&q->lock);
2706 __remove_wait_queue(q, &wait);
2707 spin_unlock_irqrestore(&q->lock, flags);
2712 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2714 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2716 EXPORT_SYMBOL(interruptible_sleep_on);
2719 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2721 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2723 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2725 void __sched sleep_on(wait_queue_head_t *q)
2727 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2729 EXPORT_SYMBOL(sleep_on);
2731 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2733 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2735 EXPORT_SYMBOL(sleep_on_timeout);
2737 #ifdef CONFIG_RT_MUTEXES
2740 * rt_mutex_setprio - set the current priority of a task
2742 * @prio: prio value (kernel-internal form)
2744 * This function changes the 'effective' priority of a task. It does
2745 * not touch ->normal_prio like __setscheduler().
2747 * Used by the rt_mutex code to implement priority inheritance logic.
2749 void rt_mutex_setprio(struct task_struct *p, int prio)
2751 int oldprio, on_rq, running;
2753 const struct sched_class *prev_class;
2755 BUG_ON(prio < 0 || prio > MAX_PRIO);
2757 rq = __task_rq_lock(p);
2760 * Idle task boosting is a nono in general. There is one
2761 * exception, when PREEMPT_RT and NOHZ is active:
2763 * The idle task calls get_next_timer_interrupt() and holds
2764 * the timer wheel base->lock on the CPU and another CPU wants
2765 * to access the timer (probably to cancel it). We can safely
2766 * ignore the boosting request, as the idle CPU runs this code
2767 * with interrupts disabled and will complete the lock
2768 * protected section without being interrupted. So there is no
2769 * real need to boost.
2771 if (unlikely(p == rq->idle)) {
2772 WARN_ON(p != rq->curr);
2773 WARN_ON(p->pi_blocked_on);
2777 trace_sched_pi_setprio(p, prio);
2779 prev_class = p->sched_class;
2781 running = task_current(rq, p);
2783 dequeue_task(rq, p, 0);
2785 p->sched_class->put_prev_task(rq, p);
2788 p->sched_class = &rt_sched_class;
2790 p->sched_class = &fair_sched_class;
2795 p->sched_class->set_curr_task(rq);
2797 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
2799 check_class_changed(rq, p, prev_class, oldprio);
2801 __task_rq_unlock(rq);
2804 void set_user_nice(struct task_struct *p, long nice)
2806 int old_prio, delta, on_rq;
2807 unsigned long flags;
2810 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2813 * We have to be careful, if called from sys_setpriority(),
2814 * the task might be in the middle of scheduling on another CPU.
2816 rq = task_rq_lock(p, &flags);
2818 * The RT priorities are set via sched_setscheduler(), but we still
2819 * allow the 'normal' nice value to be set - but as expected
2820 * it wont have any effect on scheduling until the task is
2821 * SCHED_FIFO/SCHED_RR:
2823 if (task_has_rt_policy(p)) {
2824 p->static_prio = NICE_TO_PRIO(nice);
2829 dequeue_task(rq, p, 0);
2831 p->static_prio = NICE_TO_PRIO(nice);
2834 p->prio = effective_prio(p);
2835 delta = p->prio - old_prio;
2838 enqueue_task(rq, p, 0);
2840 * If the task increased its priority or is running and
2841 * lowered its priority, then reschedule its CPU:
2843 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2844 resched_task(rq->curr);
2847 task_rq_unlock(rq, p, &flags);
2849 EXPORT_SYMBOL(set_user_nice);
2852 * can_nice - check if a task can reduce its nice value
2856 int can_nice(const struct task_struct *p, const int nice)
2858 /* convert nice value [19,-20] to rlimit style value [1,40] */
2859 int nice_rlim = 20 - nice;
2861 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
2862 capable(CAP_SYS_NICE));
2865 #ifdef __ARCH_WANT_SYS_NICE
2868 * sys_nice - change the priority of the current process.
2869 * @increment: priority increment
2871 * sys_setpriority is a more generic, but much slower function that
2872 * does similar things.
2874 SYSCALL_DEFINE1(nice, int, increment)
2879 * Setpriority might change our priority at the same moment.
2880 * We don't have to worry. Conceptually one call occurs first
2881 * and we have a single winner.
2883 if (increment < -40)
2888 nice = TASK_NICE(current) + increment;
2894 if (increment < 0 && !can_nice(current, nice))
2897 retval = security_task_setnice(current, nice);
2901 set_user_nice(current, nice);
2908 * task_prio - return the priority value of a given task.
2909 * @p: the task in question.
2911 * Return: The priority value as seen by users in /proc.
2912 * RT tasks are offset by -200. Normal tasks are centered
2913 * around 0, value goes from -16 to +15.
2915 int task_prio(const struct task_struct *p)
2917 return p->prio - MAX_RT_PRIO;
2921 * task_nice - return the nice value of a given task.
2922 * @p: the task in question.
2924 * Return: The nice value [ -20 ... 0 ... 19 ].
2926 int task_nice(const struct task_struct *p)
2928 return TASK_NICE(p);
2930 EXPORT_SYMBOL(task_nice);
2933 * idle_cpu - is a given cpu idle currently?
2934 * @cpu: the processor in question.
2936 * Return: 1 if the CPU is currently idle. 0 otherwise.
2938 int idle_cpu(int cpu)
2940 struct rq *rq = cpu_rq(cpu);
2942 if (rq->curr != rq->idle)
2949 if (!llist_empty(&rq->wake_list))
2957 * idle_task - return the idle task for a given cpu.
2958 * @cpu: the processor in question.
2960 * Return: The idle task for the cpu @cpu.
2962 struct task_struct *idle_task(int cpu)
2964 return cpu_rq(cpu)->idle;
2968 * find_process_by_pid - find a process with a matching PID value.
2969 * @pid: the pid in question.
2971 * The task of @pid, if found. %NULL otherwise.
2973 static struct task_struct *find_process_by_pid(pid_t pid)
2975 return pid ? find_task_by_vpid(pid) : current;
2978 /* Actually do priority change: must hold rq lock. */
2980 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
2983 p->rt_priority = prio;
2984 p->normal_prio = normal_prio(p);
2985 /* we are holding p->pi_lock already */
2986 p->prio = rt_mutex_getprio(p);
2987 if (rt_prio(p->prio))
2988 p->sched_class = &rt_sched_class;
2990 p->sched_class = &fair_sched_class;
2995 * check the target process has a UID that matches the current process's
2997 static bool check_same_owner(struct task_struct *p)
2999 const struct cred *cred = current_cred(), *pcred;
3003 pcred = __task_cred(p);
3004 match = (uid_eq(cred->euid, pcred->euid) ||
3005 uid_eq(cred->euid, pcred->uid));
3010 static int __sched_setscheduler(struct task_struct *p, int policy,
3011 const struct sched_param *param, bool user)
3013 int retval, oldprio, oldpolicy = -1, on_rq, running;
3014 unsigned long flags;
3015 const struct sched_class *prev_class;
3019 /* may grab non-irq protected spin_locks */
3020 BUG_ON(in_interrupt());
3022 /* double check policy once rq lock held */
3024 reset_on_fork = p->sched_reset_on_fork;
3025 policy = oldpolicy = p->policy;
3027 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3028 policy &= ~SCHED_RESET_ON_FORK;
3030 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3031 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3032 policy != SCHED_IDLE)
3037 * Valid priorities for SCHED_FIFO and SCHED_RR are
3038 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3039 * SCHED_BATCH and SCHED_IDLE is 0.
3041 if (param->sched_priority < 0 ||
3042 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3043 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3045 if (rt_policy(policy) != (param->sched_priority != 0))
3049 * Allow unprivileged RT tasks to decrease priority:
3051 if (user && !capable(CAP_SYS_NICE)) {
3052 if (rt_policy(policy)) {
3053 unsigned long rlim_rtprio =
3054 task_rlimit(p, RLIMIT_RTPRIO);
3056 /* can't set/change the rt policy */
3057 if (policy != p->policy && !rlim_rtprio)
3060 /* can't increase priority */
3061 if (param->sched_priority > p->rt_priority &&
3062 param->sched_priority > rlim_rtprio)
3067 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3068 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3070 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3071 if (!can_nice(p, TASK_NICE(p)))
3075 /* can't change other user's priorities */
3076 if (!check_same_owner(p))
3079 /* Normal users shall not reset the sched_reset_on_fork flag */
3080 if (p->sched_reset_on_fork && !reset_on_fork)
3085 retval = security_task_setscheduler(p);
3091 * make sure no PI-waiters arrive (or leave) while we are
3092 * changing the priority of the task:
3094 * To be able to change p->policy safely, the appropriate
3095 * runqueue lock must be held.
3097 rq = task_rq_lock(p, &flags);
3100 * Changing the policy of the stop threads its a very bad idea
3102 if (p == rq->stop) {
3103 task_rq_unlock(rq, p, &flags);
3108 * If not changing anything there's no need to proceed further:
3110 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3111 param->sched_priority == p->rt_priority))) {
3112 task_rq_unlock(rq, p, &flags);
3116 #ifdef CONFIG_RT_GROUP_SCHED
3119 * Do not allow realtime tasks into groups that have no runtime
3122 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3123 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3124 !task_group_is_autogroup(task_group(p))) {
3125 task_rq_unlock(rq, p, &flags);
3131 /* recheck policy now with rq lock held */
3132 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3133 policy = oldpolicy = -1;
3134 task_rq_unlock(rq, p, &flags);
3138 running = task_current(rq, p);
3140 dequeue_task(rq, p, 0);
3142 p->sched_class->put_prev_task(rq, p);
3144 p->sched_reset_on_fork = reset_on_fork;
3147 prev_class = p->sched_class;
3148 __setscheduler(rq, p, policy, param->sched_priority);
3151 p->sched_class->set_curr_task(rq);
3153 enqueue_task(rq, p, 0);
3155 check_class_changed(rq, p, prev_class, oldprio);
3156 task_rq_unlock(rq, p, &flags);
3158 rt_mutex_adjust_pi(p);
3164 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3165 * @p: the task in question.
3166 * @policy: new policy.
3167 * @param: structure containing the new RT priority.
3169 * Return: 0 on success. An error code otherwise.
3171 * NOTE that the task may be already dead.
3173 int sched_setscheduler(struct task_struct *p, int policy,
3174 const struct sched_param *param)
3176 return __sched_setscheduler(p, policy, param, true);
3178 EXPORT_SYMBOL_GPL(sched_setscheduler);
3181 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3182 * @p: the task in question.
3183 * @policy: new policy.
3184 * @param: structure containing the new RT priority.
3186 * Just like sched_setscheduler, only don't bother checking if the
3187 * current context has permission. For example, this is needed in
3188 * stop_machine(): we create temporary high priority worker threads,
3189 * but our caller might not have that capability.
3191 * Return: 0 on success. An error code otherwise.
3193 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3194 const struct sched_param *param)
3196 return __sched_setscheduler(p, policy, param, false);
3200 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3202 struct sched_param lparam;
3203 struct task_struct *p;
3206 if (!param || pid < 0)
3208 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3213 p = find_process_by_pid(pid);
3215 retval = sched_setscheduler(p, policy, &lparam);
3222 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3223 * @pid: the pid in question.
3224 * @policy: new policy.
3225 * @param: structure containing the new RT priority.
3227 * Return: 0 on success. An error code otherwise.
3229 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3230 struct sched_param __user *, param)
3232 /* negative values for policy are not valid */
3236 return do_sched_setscheduler(pid, policy, param);
3240 * sys_sched_setparam - set/change the RT priority of a thread
3241 * @pid: the pid in question.
3242 * @param: structure containing the new RT priority.
3244 * Return: 0 on success. An error code otherwise.
3246 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3248 return do_sched_setscheduler(pid, -1, param);
3252 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3253 * @pid: the pid in question.
3255 * Return: On success, the policy of the thread. Otherwise, a negative error
3258 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3260 struct task_struct *p;
3268 p = find_process_by_pid(pid);
3270 retval = security_task_getscheduler(p);
3273 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3280 * sys_sched_getparam - get the RT priority of a thread
3281 * @pid: the pid in question.
3282 * @param: structure containing the RT priority.
3284 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3287 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3289 struct sched_param lp;
3290 struct task_struct *p;
3293 if (!param || pid < 0)
3297 p = find_process_by_pid(pid);
3302 retval = security_task_getscheduler(p);
3306 lp.sched_priority = p->rt_priority;
3310 * This one might sleep, we cannot do it with a spinlock held ...
3312 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3321 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3323 cpumask_var_t cpus_allowed, new_mask;
3324 struct task_struct *p;
3329 p = find_process_by_pid(pid);
3335 /* Prevent p going away */
3339 if (p->flags & PF_NO_SETAFFINITY) {
3343 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3347 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3349 goto out_free_cpus_allowed;
3352 if (!check_same_owner(p)) {
3354 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3361 retval = security_task_setscheduler(p);
3365 cpuset_cpus_allowed(p, cpus_allowed);
3366 cpumask_and(new_mask, in_mask, cpus_allowed);
3368 retval = set_cpus_allowed_ptr(p, new_mask);
3371 cpuset_cpus_allowed(p, cpus_allowed);
3372 if (!cpumask_subset(new_mask, cpus_allowed)) {
3374 * We must have raced with a concurrent cpuset
3375 * update. Just reset the cpus_allowed to the
3376 * cpuset's cpus_allowed
3378 cpumask_copy(new_mask, cpus_allowed);
3383 free_cpumask_var(new_mask);
3384 out_free_cpus_allowed:
3385 free_cpumask_var(cpus_allowed);
3391 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3392 struct cpumask *new_mask)
3394 if (len < cpumask_size())
3395 cpumask_clear(new_mask);
3396 else if (len > cpumask_size())
3397 len = cpumask_size();
3399 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3403 * sys_sched_setaffinity - set the cpu affinity of a process
3404 * @pid: pid of the process
3405 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3406 * @user_mask_ptr: user-space pointer to the new cpu mask
3408 * Return: 0 on success. An error code otherwise.
3410 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3411 unsigned long __user *, user_mask_ptr)
3413 cpumask_var_t new_mask;
3416 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3419 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3421 retval = sched_setaffinity(pid, new_mask);
3422 free_cpumask_var(new_mask);
3426 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3428 struct task_struct *p;
3429 unsigned long flags;
3435 p = find_process_by_pid(pid);
3439 retval = security_task_getscheduler(p);
3443 raw_spin_lock_irqsave(&p->pi_lock, flags);
3444 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3445 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3454 * sys_sched_getaffinity - get the cpu affinity of a process
3455 * @pid: pid of the process
3456 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3457 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3459 * Return: 0 on success. An error code otherwise.
3461 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3462 unsigned long __user *, user_mask_ptr)
3467 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3469 if (len & (sizeof(unsigned long)-1))
3472 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3475 ret = sched_getaffinity(pid, mask);
3477 size_t retlen = min_t(size_t, len, cpumask_size());
3479 if (copy_to_user(user_mask_ptr, mask, retlen))
3484 free_cpumask_var(mask);
3490 * sys_sched_yield - yield the current processor to other threads.
3492 * This function yields the current CPU to other tasks. If there are no
3493 * other threads running on this CPU then this function will return.
3497 SYSCALL_DEFINE0(sched_yield)
3499 struct rq *rq = this_rq_lock();
3501 schedstat_inc(rq, yld_count);
3502 current->sched_class->yield_task(rq);
3505 * Since we are going to call schedule() anyway, there's
3506 * no need to preempt or enable interrupts:
3508 __release(rq->lock);
3509 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3510 do_raw_spin_unlock(&rq->lock);
3511 sched_preempt_enable_no_resched();
3518 static void __cond_resched(void)
3520 __preempt_count_add(PREEMPT_ACTIVE);
3522 __preempt_count_sub(PREEMPT_ACTIVE);
3525 int __sched _cond_resched(void)
3527 if (should_resched()) {
3533 EXPORT_SYMBOL(_cond_resched);
3536 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3537 * call schedule, and on return reacquire the lock.
3539 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3540 * operations here to prevent schedule() from being called twice (once via
3541 * spin_unlock(), once by hand).
3543 int __cond_resched_lock(spinlock_t *lock)
3545 int resched = should_resched();
3548 lockdep_assert_held(lock);
3550 if (spin_needbreak(lock) || resched) {
3561 EXPORT_SYMBOL(__cond_resched_lock);
3563 int __sched __cond_resched_softirq(void)
3565 BUG_ON(!in_softirq());
3567 if (should_resched()) {
3575 EXPORT_SYMBOL(__cond_resched_softirq);
3578 * yield - yield the current processor to other threads.
3580 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3582 * The scheduler is at all times free to pick the calling task as the most
3583 * eligible task to run, if removing the yield() call from your code breaks
3584 * it, its already broken.
3586 * Typical broken usage is:
3591 * where one assumes that yield() will let 'the other' process run that will
3592 * make event true. If the current task is a SCHED_FIFO task that will never
3593 * happen. Never use yield() as a progress guarantee!!
3595 * If you want to use yield() to wait for something, use wait_event().
3596 * If you want to use yield() to be 'nice' for others, use cond_resched().
3597 * If you still want to use yield(), do not!
3599 void __sched yield(void)
3601 set_current_state(TASK_RUNNING);
3604 EXPORT_SYMBOL(yield);
3607 * yield_to - yield the current processor to another thread in
3608 * your thread group, or accelerate that thread toward the
3609 * processor it's on.
3611 * @preempt: whether task preemption is allowed or not
3613 * It's the caller's job to ensure that the target task struct
3614 * can't go away on us before we can do any checks.
3617 * true (>0) if we indeed boosted the target task.
3618 * false (0) if we failed to boost the target.
3619 * -ESRCH if there's no task to yield to.
3621 bool __sched yield_to(struct task_struct *p, bool preempt)
3623 struct task_struct *curr = current;
3624 struct rq *rq, *p_rq;
3625 unsigned long flags;
3628 local_irq_save(flags);
3634 * If we're the only runnable task on the rq and target rq also
3635 * has only one task, there's absolutely no point in yielding.
3637 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3642 double_rq_lock(rq, p_rq);
3643 while (task_rq(p) != p_rq) {
3644 double_rq_unlock(rq, p_rq);
3648 if (!curr->sched_class->yield_to_task)
3651 if (curr->sched_class != p->sched_class)
3654 if (task_running(p_rq, p) || p->state)
3657 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3659 schedstat_inc(rq, yld_count);
3661 * Make p's CPU reschedule; pick_next_entity takes care of
3664 if (preempt && rq != p_rq)
3665 resched_task(p_rq->curr);
3669 double_rq_unlock(rq, p_rq);
3671 local_irq_restore(flags);
3678 EXPORT_SYMBOL_GPL(yield_to);
3681 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3682 * that process accounting knows that this is a task in IO wait state.
3684 void __sched io_schedule(void)
3686 struct rq *rq = raw_rq();
3688 delayacct_blkio_start();
3689 atomic_inc(&rq->nr_iowait);
3690 blk_flush_plug(current);
3691 current->in_iowait = 1;
3693 current->in_iowait = 0;
3694 atomic_dec(&rq->nr_iowait);
3695 delayacct_blkio_end();
3697 EXPORT_SYMBOL(io_schedule);
3699 long __sched io_schedule_timeout(long timeout)
3701 struct rq *rq = raw_rq();
3704 delayacct_blkio_start();
3705 atomic_inc(&rq->nr_iowait);
3706 blk_flush_plug(current);
3707 current->in_iowait = 1;
3708 ret = schedule_timeout(timeout);
3709 current->in_iowait = 0;
3710 atomic_dec(&rq->nr_iowait);
3711 delayacct_blkio_end();
3716 * sys_sched_get_priority_max - return maximum RT priority.
3717 * @policy: scheduling class.
3719 * Return: On success, this syscall returns the maximum
3720 * rt_priority that can be used by a given scheduling class.
3721 * On failure, a negative error code is returned.
3723 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3730 ret = MAX_USER_RT_PRIO-1;
3742 * sys_sched_get_priority_min - return minimum RT priority.
3743 * @policy: scheduling class.
3745 * Return: On success, this syscall returns the minimum
3746 * rt_priority that can be used by a given scheduling class.
3747 * On failure, a negative error code is returned.
3749 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
3767 * sys_sched_rr_get_interval - return the default timeslice of a process.
3768 * @pid: pid of the process.
3769 * @interval: userspace pointer to the timeslice value.
3771 * this syscall writes the default timeslice value of a given process
3772 * into the user-space timespec buffer. A value of '0' means infinity.
3774 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
3777 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
3778 struct timespec __user *, interval)
3780 struct task_struct *p;
3781 unsigned int time_slice;
3782 unsigned long flags;
3792 p = find_process_by_pid(pid);
3796 retval = security_task_getscheduler(p);
3800 rq = task_rq_lock(p, &flags);
3801 time_slice = p->sched_class->get_rr_interval(rq, p);
3802 task_rq_unlock(rq, p, &flags);
3805 jiffies_to_timespec(time_slice, &t);
3806 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
3814 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
3816 void sched_show_task(struct task_struct *p)
3818 unsigned long free = 0;
3822 state = p->state ? __ffs(p->state) + 1 : 0;
3823 printk(KERN_INFO "%-15.15s %c", p->comm,
3824 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
3825 #if BITS_PER_LONG == 32
3826 if (state == TASK_RUNNING)
3827 printk(KERN_CONT " running ");
3829 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
3831 if (state == TASK_RUNNING)
3832 printk(KERN_CONT " running task ");
3834 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
3836 #ifdef CONFIG_DEBUG_STACK_USAGE
3837 free = stack_not_used(p);
3840 ppid = task_pid_nr(rcu_dereference(p->real_parent));
3842 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
3843 task_pid_nr(p), ppid,
3844 (unsigned long)task_thread_info(p)->flags);
3846 print_worker_info(KERN_INFO, p);
3847 show_stack(p, NULL);
3850 void show_state_filter(unsigned long state_filter)
3852 struct task_struct *g, *p;
3854 #if BITS_PER_LONG == 32
3856 " task PC stack pid father\n");
3859 " task PC stack pid father\n");
3862 do_each_thread(g, p) {
3864 * reset the NMI-timeout, listing all files on a slow
3865 * console might take a lot of time:
3867 touch_nmi_watchdog();
3868 if (!state_filter || (p->state & state_filter))
3870 } while_each_thread(g, p);
3872 touch_all_softlockup_watchdogs();
3874 #ifdef CONFIG_SCHED_DEBUG
3875 sysrq_sched_debug_show();
3879 * Only show locks if all tasks are dumped:
3882 debug_show_all_locks();
3885 void init_idle_bootup_task(struct task_struct *idle)
3887 idle->sched_class = &idle_sched_class;
3891 * init_idle - set up an idle thread for a given CPU
3892 * @idle: task in question
3893 * @cpu: cpu the idle task belongs to
3895 * NOTE: this function does not set the idle thread's NEED_RESCHED
3896 * flag, to make booting more robust.
3898 void init_idle(struct task_struct *idle, int cpu)
3900 struct rq *rq = cpu_rq(cpu);
3901 unsigned long flags;
3903 raw_spin_lock_irqsave(&rq->lock, flags);
3905 __sched_fork(0, idle);
3906 idle->state = TASK_RUNNING;
3907 idle->se.exec_start = sched_clock();
3909 do_set_cpus_allowed(idle, cpumask_of(cpu));
3911 * We're having a chicken and egg problem, even though we are
3912 * holding rq->lock, the cpu isn't yet set to this cpu so the
3913 * lockdep check in task_group() will fail.
3915 * Similar case to sched_fork(). / Alternatively we could
3916 * use task_rq_lock() here and obtain the other rq->lock.
3921 __set_task_cpu(idle, cpu);
3924 rq->curr = rq->idle = idle;
3925 #if defined(CONFIG_SMP)
3928 raw_spin_unlock_irqrestore(&rq->lock, flags);
3930 /* Set the preempt count _outside_ the spinlocks! */
3931 init_idle_preempt_count(idle, cpu);
3934 * The idle tasks have their own, simple scheduling class:
3936 idle->sched_class = &idle_sched_class;
3937 ftrace_graph_init_idle_task(idle, cpu);
3938 vtime_init_idle(idle, cpu);
3939 #if defined(CONFIG_SMP)
3940 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
3945 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
3947 if (p->sched_class && p->sched_class->set_cpus_allowed)
3948 p->sched_class->set_cpus_allowed(p, new_mask);
3950 cpumask_copy(&p->cpus_allowed, new_mask);
3951 p->nr_cpus_allowed = cpumask_weight(new_mask);
3955 * This is how migration works:
3957 * 1) we invoke migration_cpu_stop() on the target CPU using
3959 * 2) stopper starts to run (implicitly forcing the migrated thread
3961 * 3) it checks whether the migrated task is still in the wrong runqueue.
3962 * 4) if it's in the wrong runqueue then the migration thread removes
3963 * it and puts it into the right queue.
3964 * 5) stopper completes and stop_one_cpu() returns and the migration
3969 * Change a given task's CPU affinity. Migrate the thread to a
3970 * proper CPU and schedule it away if the CPU it's executing on
3971 * is removed from the allowed bitmask.
3973 * NOTE: the caller must have a valid reference to the task, the
3974 * task must not exit() & deallocate itself prematurely. The
3975 * call is not atomic; no spinlocks may be held.
3977 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3979 unsigned long flags;
3981 unsigned int dest_cpu;
3984 rq = task_rq_lock(p, &flags);
3986 if (cpumask_equal(&p->cpus_allowed, new_mask))
3989 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
3994 do_set_cpus_allowed(p, new_mask);
3996 /* Can the task run on the task's current CPU? If so, we're done */
3997 if (cpumask_test_cpu(task_cpu(p), new_mask))
4000 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4002 struct migration_arg arg = { p, dest_cpu };
4003 /* Need help from migration thread: drop lock and wait. */
4004 task_rq_unlock(rq, p, &flags);
4005 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4006 tlb_migrate_finish(p->mm);
4010 task_rq_unlock(rq, p, &flags);
4014 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4017 * Move (not current) task off this cpu, onto dest cpu. We're doing
4018 * this because either it can't run here any more (set_cpus_allowed()
4019 * away from this CPU, or CPU going down), or because we're
4020 * attempting to rebalance this task on exec (sched_exec).
4022 * So we race with normal scheduler movements, but that's OK, as long
4023 * as the task is no longer on this CPU.
4025 * Returns non-zero if task was successfully migrated.
4027 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4029 struct rq *rq_dest, *rq_src;
4032 if (unlikely(!cpu_active(dest_cpu)))
4035 rq_src = cpu_rq(src_cpu);
4036 rq_dest = cpu_rq(dest_cpu);
4038 raw_spin_lock(&p->pi_lock);
4039 double_rq_lock(rq_src, rq_dest);
4040 /* Already moved. */
4041 if (task_cpu(p) != src_cpu)
4043 /* Affinity changed (again). */
4044 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4048 * If we're not on a rq, the next wake-up will ensure we're
4052 dequeue_task(rq_src, p, 0);
4053 set_task_cpu(p, dest_cpu);
4054 enqueue_task(rq_dest, p, 0);
4055 check_preempt_curr(rq_dest, p, 0);
4060 double_rq_unlock(rq_src, rq_dest);
4061 raw_spin_unlock(&p->pi_lock);
4065 #ifdef CONFIG_NUMA_BALANCING
4066 /* Migrate current task p to target_cpu */
4067 int migrate_task_to(struct task_struct *p, int target_cpu)
4069 struct migration_arg arg = { p, target_cpu };
4070 int curr_cpu = task_cpu(p);
4072 if (curr_cpu == target_cpu)
4075 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4078 /* TODO: This is not properly updating schedstats */
4080 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4084 * Requeue a task on a given node and accurately track the number of NUMA
4085 * tasks on the runqueues
4087 void sched_setnuma(struct task_struct *p, int nid)
4090 unsigned long flags;
4091 bool on_rq, running;
4093 rq = task_rq_lock(p, &flags);
4095 running = task_current(rq, p);
4098 dequeue_task(rq, p, 0);
4100 p->sched_class->put_prev_task(rq, p);
4102 p->numa_preferred_nid = nid;
4105 p->sched_class->set_curr_task(rq);
4107 enqueue_task(rq, p, 0);
4108 task_rq_unlock(rq, p, &flags);
4113 * migration_cpu_stop - this will be executed by a highprio stopper thread
4114 * and performs thread migration by bumping thread off CPU then
4115 * 'pushing' onto another runqueue.
4117 static int migration_cpu_stop(void *data)
4119 struct migration_arg *arg = data;
4122 * The original target cpu might have gone down and we might
4123 * be on another cpu but it doesn't matter.
4125 local_irq_disable();
4126 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4131 #ifdef CONFIG_HOTPLUG_CPU
4134 * Ensures that the idle task is using init_mm right before its cpu goes
4137 void idle_task_exit(void)
4139 struct mm_struct *mm = current->active_mm;
4141 BUG_ON(cpu_online(smp_processor_id()));
4144 switch_mm(mm, &init_mm, current);
4149 * Since this CPU is going 'away' for a while, fold any nr_active delta
4150 * we might have. Assumes we're called after migrate_tasks() so that the
4151 * nr_active count is stable.
4153 * Also see the comment "Global load-average calculations".
4155 static void calc_load_migrate(struct rq *rq)
4157 long delta = calc_load_fold_active(rq);
4159 atomic_long_add(delta, &calc_load_tasks);
4163 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4164 * try_to_wake_up()->select_task_rq().
4166 * Called with rq->lock held even though we'er in stop_machine() and
4167 * there's no concurrency possible, we hold the required locks anyway
4168 * because of lock validation efforts.
4170 static void migrate_tasks(unsigned int dead_cpu)
4172 struct rq *rq = cpu_rq(dead_cpu);
4173 struct task_struct *next, *stop = rq->stop;
4177 * Fudge the rq selection such that the below task selection loop
4178 * doesn't get stuck on the currently eligible stop task.
4180 * We're currently inside stop_machine() and the rq is either stuck
4181 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4182 * either way we should never end up calling schedule() until we're
4188 * put_prev_task() and pick_next_task() sched
4189 * class method both need to have an up-to-date
4190 * value of rq->clock[_task]
4192 update_rq_clock(rq);
4196 * There's this thread running, bail when that's the only
4199 if (rq->nr_running == 1)
4202 next = pick_next_task(rq);
4204 next->sched_class->put_prev_task(rq, next);
4206 /* Find suitable destination for @next, with force if needed. */
4207 dest_cpu = select_fallback_rq(dead_cpu, next);
4208 raw_spin_unlock(&rq->lock);
4210 __migrate_task(next, dead_cpu, dest_cpu);
4212 raw_spin_lock(&rq->lock);
4218 #endif /* CONFIG_HOTPLUG_CPU */
4220 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4222 static struct ctl_table sd_ctl_dir[] = {
4224 .procname = "sched_domain",
4230 static struct ctl_table sd_ctl_root[] = {
4232 .procname = "kernel",
4234 .child = sd_ctl_dir,
4239 static struct ctl_table *sd_alloc_ctl_entry(int n)
4241 struct ctl_table *entry =
4242 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4247 static void sd_free_ctl_entry(struct ctl_table **tablep)
4249 struct ctl_table *entry;
4252 * In the intermediate directories, both the child directory and
4253 * procname are dynamically allocated and could fail but the mode
4254 * will always be set. In the lowest directory the names are
4255 * static strings and all have proc handlers.
4257 for (entry = *tablep; entry->mode; entry++) {
4259 sd_free_ctl_entry(&entry->child);
4260 if (entry->proc_handler == NULL)
4261 kfree(entry->procname);
4268 static int min_load_idx = 0;
4269 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4272 set_table_entry(struct ctl_table *entry,
4273 const char *procname, void *data, int maxlen,
4274 umode_t mode, proc_handler *proc_handler,
4277 entry->procname = procname;
4279 entry->maxlen = maxlen;
4281 entry->proc_handler = proc_handler;
4284 entry->extra1 = &min_load_idx;
4285 entry->extra2 = &max_load_idx;
4289 static struct ctl_table *
4290 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4292 struct ctl_table *table = sd_alloc_ctl_entry(13);
4297 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4298 sizeof(long), 0644, proc_doulongvec_minmax, false);
4299 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4300 sizeof(long), 0644, proc_doulongvec_minmax, false);
4301 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4302 sizeof(int), 0644, proc_dointvec_minmax, true);
4303 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4304 sizeof(int), 0644, proc_dointvec_minmax, true);
4305 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4306 sizeof(int), 0644, proc_dointvec_minmax, true);
4307 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4308 sizeof(int), 0644, proc_dointvec_minmax, true);
4309 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4310 sizeof(int), 0644, proc_dointvec_minmax, true);
4311 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4312 sizeof(int), 0644, proc_dointvec_minmax, false);
4313 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4314 sizeof(int), 0644, proc_dointvec_minmax, false);
4315 set_table_entry(&table[9], "cache_nice_tries",
4316 &sd->cache_nice_tries,
4317 sizeof(int), 0644, proc_dointvec_minmax, false);
4318 set_table_entry(&table[10], "flags", &sd->flags,
4319 sizeof(int), 0644, proc_dointvec_minmax, false);
4320 set_table_entry(&table[11], "name", sd->name,
4321 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4322 /* &table[12] is terminator */
4327 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4329 struct ctl_table *entry, *table;
4330 struct sched_domain *sd;
4331 int domain_num = 0, i;
4334 for_each_domain(cpu, sd)
4336 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4341 for_each_domain(cpu, sd) {
4342 snprintf(buf, 32, "domain%d", i);
4343 entry->procname = kstrdup(buf, GFP_KERNEL);
4345 entry->child = sd_alloc_ctl_domain_table(sd);
4352 static struct ctl_table_header *sd_sysctl_header;
4353 static void register_sched_domain_sysctl(void)
4355 int i, cpu_num = num_possible_cpus();
4356 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4359 WARN_ON(sd_ctl_dir[0].child);
4360 sd_ctl_dir[0].child = entry;
4365 for_each_possible_cpu(i) {
4366 snprintf(buf, 32, "cpu%d", i);
4367 entry->procname = kstrdup(buf, GFP_KERNEL);
4369 entry->child = sd_alloc_ctl_cpu_table(i);
4373 WARN_ON(sd_sysctl_header);
4374 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4377 /* may be called multiple times per register */
4378 static void unregister_sched_domain_sysctl(void)
4380 if (sd_sysctl_header)
4381 unregister_sysctl_table(sd_sysctl_header);
4382 sd_sysctl_header = NULL;
4383 if (sd_ctl_dir[0].child)
4384 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4387 static void register_sched_domain_sysctl(void)
4390 static void unregister_sched_domain_sysctl(void)
4395 static void set_rq_online(struct rq *rq)
4398 const struct sched_class *class;
4400 cpumask_set_cpu(rq->cpu, rq->rd->online);
4403 for_each_class(class) {
4404 if (class->rq_online)
4405 class->rq_online(rq);
4410 static void set_rq_offline(struct rq *rq)
4413 const struct sched_class *class;
4415 for_each_class(class) {
4416 if (class->rq_offline)
4417 class->rq_offline(rq);
4420 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4426 * migration_call - callback that gets triggered when a CPU is added.
4427 * Here we can start up the necessary migration thread for the new CPU.
4430 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4432 int cpu = (long)hcpu;
4433 unsigned long flags;
4434 struct rq *rq = cpu_rq(cpu);
4436 switch (action & ~CPU_TASKS_FROZEN) {
4438 case CPU_UP_PREPARE:
4439 rq->calc_load_update = calc_load_update;
4443 /* Update our root-domain */
4444 raw_spin_lock_irqsave(&rq->lock, flags);
4446 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4450 raw_spin_unlock_irqrestore(&rq->lock, flags);
4453 #ifdef CONFIG_HOTPLUG_CPU
4455 sched_ttwu_pending();
4456 /* Update our root-domain */
4457 raw_spin_lock_irqsave(&rq->lock, flags);
4459 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4463 BUG_ON(rq->nr_running != 1); /* the migration thread */
4464 raw_spin_unlock_irqrestore(&rq->lock, flags);
4468 calc_load_migrate(rq);
4473 update_max_interval();
4479 * Register at high priority so that task migration (migrate_all_tasks)
4480 * happens before everything else. This has to be lower priority than
4481 * the notifier in the perf_event subsystem, though.
4483 static struct notifier_block migration_notifier = {
4484 .notifier_call = migration_call,
4485 .priority = CPU_PRI_MIGRATION,
4488 static int sched_cpu_active(struct notifier_block *nfb,
4489 unsigned long action, void *hcpu)
4491 switch (action & ~CPU_TASKS_FROZEN) {
4493 case CPU_DOWN_FAILED:
4494 set_cpu_active((long)hcpu, true);
4501 static int sched_cpu_inactive(struct notifier_block *nfb,
4502 unsigned long action, void *hcpu)
4504 switch (action & ~CPU_TASKS_FROZEN) {
4505 case CPU_DOWN_PREPARE:
4506 set_cpu_active((long)hcpu, false);
4513 static int __init migration_init(void)
4515 void *cpu = (void *)(long)smp_processor_id();
4518 /* Initialize migration for the boot CPU */
4519 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4520 BUG_ON(err == NOTIFY_BAD);
4521 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4522 register_cpu_notifier(&migration_notifier);
4524 /* Register cpu active notifiers */
4525 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4526 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4530 early_initcall(migration_init);
4535 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4537 #ifdef CONFIG_SCHED_DEBUG
4539 static __read_mostly int sched_debug_enabled;
4541 static int __init sched_debug_setup(char *str)
4543 sched_debug_enabled = 1;
4547 early_param("sched_debug", sched_debug_setup);
4549 static inline bool sched_debug(void)
4551 return sched_debug_enabled;
4554 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4555 struct cpumask *groupmask)
4557 struct sched_group *group = sd->groups;
4560 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4561 cpumask_clear(groupmask);
4563 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4565 if (!(sd->flags & SD_LOAD_BALANCE)) {
4566 printk("does not load-balance\n");
4568 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4573 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4575 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4576 printk(KERN_ERR "ERROR: domain->span does not contain "
4579 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4580 printk(KERN_ERR "ERROR: domain->groups does not contain"
4584 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4588 printk(KERN_ERR "ERROR: group is NULL\n");
4593 * Even though we initialize ->power to something semi-sane,
4594 * we leave power_orig unset. This allows us to detect if
4595 * domain iteration is still funny without causing /0 traps.
4597 if (!group->sgp->power_orig) {
4598 printk(KERN_CONT "\n");
4599 printk(KERN_ERR "ERROR: domain->cpu_power not "
4604 if (!cpumask_weight(sched_group_cpus(group))) {
4605 printk(KERN_CONT "\n");
4606 printk(KERN_ERR "ERROR: empty group\n");
4610 if (!(sd->flags & SD_OVERLAP) &&
4611 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4612 printk(KERN_CONT "\n");
4613 printk(KERN_ERR "ERROR: repeated CPUs\n");
4617 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4619 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4621 printk(KERN_CONT " %s", str);
4622 if (group->sgp->power != SCHED_POWER_SCALE) {
4623 printk(KERN_CONT " (cpu_power = %d)",
4627 group = group->next;
4628 } while (group != sd->groups);
4629 printk(KERN_CONT "\n");
4631 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4632 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4635 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4636 printk(KERN_ERR "ERROR: parent span is not a superset "
4637 "of domain->span\n");
4641 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4645 if (!sched_debug_enabled)
4649 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4653 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4656 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4664 #else /* !CONFIG_SCHED_DEBUG */
4665 # define sched_domain_debug(sd, cpu) do { } while (0)
4666 static inline bool sched_debug(void)
4670 #endif /* CONFIG_SCHED_DEBUG */
4672 static int sd_degenerate(struct sched_domain *sd)
4674 if (cpumask_weight(sched_domain_span(sd)) == 1)
4677 /* Following flags need at least 2 groups */
4678 if (sd->flags & (SD_LOAD_BALANCE |
4679 SD_BALANCE_NEWIDLE |
4683 SD_SHARE_PKG_RESOURCES)) {
4684 if (sd->groups != sd->groups->next)
4688 /* Following flags don't use groups */
4689 if (sd->flags & (SD_WAKE_AFFINE))
4696 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4698 unsigned long cflags = sd->flags, pflags = parent->flags;
4700 if (sd_degenerate(parent))
4703 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4706 /* Flags needing groups don't count if only 1 group in parent */
4707 if (parent->groups == parent->groups->next) {
4708 pflags &= ~(SD_LOAD_BALANCE |
4709 SD_BALANCE_NEWIDLE |
4713 SD_SHARE_PKG_RESOURCES |
4715 if (nr_node_ids == 1)
4716 pflags &= ~SD_SERIALIZE;
4718 if (~cflags & pflags)
4724 static void free_rootdomain(struct rcu_head *rcu)
4726 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4728 cpupri_cleanup(&rd->cpupri);
4729 free_cpumask_var(rd->rto_mask);
4730 free_cpumask_var(rd->online);
4731 free_cpumask_var(rd->span);
4735 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4737 struct root_domain *old_rd = NULL;
4738 unsigned long flags;
4740 raw_spin_lock_irqsave(&rq->lock, flags);
4745 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4748 cpumask_clear_cpu(rq->cpu, old_rd->span);
4751 * If we dont want to free the old_rt yet then
4752 * set old_rd to NULL to skip the freeing later
4755 if (!atomic_dec_and_test(&old_rd->refcount))
4759 atomic_inc(&rd->refcount);
4762 cpumask_set_cpu(rq->cpu, rd->span);
4763 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4766 raw_spin_unlock_irqrestore(&rq->lock, flags);
4769 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4772 static int init_rootdomain(struct root_domain *rd)
4774 memset(rd, 0, sizeof(*rd));
4776 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4778 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4780 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4783 if (cpupri_init(&rd->cpupri) != 0)
4788 free_cpumask_var(rd->rto_mask);
4790 free_cpumask_var(rd->online);
4792 free_cpumask_var(rd->span);
4798 * By default the system creates a single root-domain with all cpus as
4799 * members (mimicking the global state we have today).
4801 struct root_domain def_root_domain;
4803 static void init_defrootdomain(void)
4805 init_rootdomain(&def_root_domain);
4807 atomic_set(&def_root_domain.refcount, 1);
4810 static struct root_domain *alloc_rootdomain(void)
4812 struct root_domain *rd;
4814 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
4818 if (init_rootdomain(rd) != 0) {
4826 static void free_sched_groups(struct sched_group *sg, int free_sgp)
4828 struct sched_group *tmp, *first;
4837 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
4842 } while (sg != first);
4845 static void free_sched_domain(struct rcu_head *rcu)
4847 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
4850 * If its an overlapping domain it has private groups, iterate and
4853 if (sd->flags & SD_OVERLAP) {
4854 free_sched_groups(sd->groups, 1);
4855 } else if (atomic_dec_and_test(&sd->groups->ref)) {
4856 kfree(sd->groups->sgp);
4862 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
4864 call_rcu(&sd->rcu, free_sched_domain);
4867 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
4869 for (; sd; sd = sd->parent)
4870 destroy_sched_domain(sd, cpu);
4874 * Keep a special pointer to the highest sched_domain that has
4875 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
4876 * allows us to avoid some pointer chasing select_idle_sibling().
4878 * Also keep a unique ID per domain (we use the first cpu number in
4879 * the cpumask of the domain), this allows us to quickly tell if
4880 * two cpus are in the same cache domain, see cpus_share_cache().
4882 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
4883 DEFINE_PER_CPU(int, sd_llc_size);
4884 DEFINE_PER_CPU(int, sd_llc_id);
4885 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
4886 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
4887 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
4889 static void update_top_cache_domain(int cpu)
4891 struct sched_domain *sd;
4895 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
4897 id = cpumask_first(sched_domain_span(sd));
4898 size = cpumask_weight(sched_domain_span(sd));
4899 rcu_assign_pointer(per_cpu(sd_busy, cpu), sd->parent);
4902 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
4903 per_cpu(sd_llc_size, cpu) = size;
4904 per_cpu(sd_llc_id, cpu) = id;
4906 sd = lowest_flag_domain(cpu, SD_NUMA);
4907 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
4909 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
4910 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
4914 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4915 * hold the hotplug lock.
4918 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
4920 struct rq *rq = cpu_rq(cpu);
4921 struct sched_domain *tmp;
4923 /* Remove the sched domains which do not contribute to scheduling. */
4924 for (tmp = sd; tmp; ) {
4925 struct sched_domain *parent = tmp->parent;
4929 if (sd_parent_degenerate(tmp, parent)) {
4930 tmp->parent = parent->parent;
4932 parent->parent->child = tmp;
4934 * Transfer SD_PREFER_SIBLING down in case of a
4935 * degenerate parent; the spans match for this
4936 * so the property transfers.
4938 if (parent->flags & SD_PREFER_SIBLING)
4939 tmp->flags |= SD_PREFER_SIBLING;
4940 destroy_sched_domain(parent, cpu);
4945 if (sd && sd_degenerate(sd)) {
4948 destroy_sched_domain(tmp, cpu);
4953 sched_domain_debug(sd, cpu);
4955 rq_attach_root(rq, rd);
4957 rcu_assign_pointer(rq->sd, sd);
4958 destroy_sched_domains(tmp, cpu);
4960 update_top_cache_domain(cpu);
4963 /* cpus with isolated domains */
4964 static cpumask_var_t cpu_isolated_map;
4966 /* Setup the mask of cpus configured for isolated domains */
4967 static int __init isolated_cpu_setup(char *str)
4969 alloc_bootmem_cpumask_var(&cpu_isolated_map);
4970 cpulist_parse(str, cpu_isolated_map);
4974 __setup("isolcpus=", isolated_cpu_setup);
4976 static const struct cpumask *cpu_cpu_mask(int cpu)
4978 return cpumask_of_node(cpu_to_node(cpu));
4982 struct sched_domain **__percpu sd;
4983 struct sched_group **__percpu sg;
4984 struct sched_group_power **__percpu sgp;
4988 struct sched_domain ** __percpu sd;
4989 struct root_domain *rd;
4999 struct sched_domain_topology_level;
5001 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5002 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5004 #define SDTL_OVERLAP 0x01
5006 struct sched_domain_topology_level {
5007 sched_domain_init_f init;
5008 sched_domain_mask_f mask;
5011 struct sd_data data;
5015 * Build an iteration mask that can exclude certain CPUs from the upwards
5018 * Asymmetric node setups can result in situations where the domain tree is of
5019 * unequal depth, make sure to skip domains that already cover the entire
5022 * In that case build_sched_domains() will have terminated the iteration early
5023 * and our sibling sd spans will be empty. Domains should always include the
5024 * cpu they're built on, so check that.
5027 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5029 const struct cpumask *span = sched_domain_span(sd);
5030 struct sd_data *sdd = sd->private;
5031 struct sched_domain *sibling;
5034 for_each_cpu(i, span) {
5035 sibling = *per_cpu_ptr(sdd->sd, i);
5036 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5039 cpumask_set_cpu(i, sched_group_mask(sg));
5044 * Return the canonical balance cpu for this group, this is the first cpu
5045 * of this group that's also in the iteration mask.
5047 int group_balance_cpu(struct sched_group *sg)
5049 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5053 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5055 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5056 const struct cpumask *span = sched_domain_span(sd);
5057 struct cpumask *covered = sched_domains_tmpmask;
5058 struct sd_data *sdd = sd->private;
5059 struct sched_domain *child;
5062 cpumask_clear(covered);
5064 for_each_cpu(i, span) {
5065 struct cpumask *sg_span;
5067 if (cpumask_test_cpu(i, covered))
5070 child = *per_cpu_ptr(sdd->sd, i);
5072 /* See the comment near build_group_mask(). */
5073 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5076 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5077 GFP_KERNEL, cpu_to_node(cpu));
5082 sg_span = sched_group_cpus(sg);
5084 child = child->child;
5085 cpumask_copy(sg_span, sched_domain_span(child));
5087 cpumask_set_cpu(i, sg_span);
5089 cpumask_or(covered, covered, sg_span);
5091 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5092 if (atomic_inc_return(&sg->sgp->ref) == 1)
5093 build_group_mask(sd, sg);
5096 * Initialize sgp->power such that even if we mess up the
5097 * domains and no possible iteration will get us here, we won't
5100 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5103 * Make sure the first group of this domain contains the
5104 * canonical balance cpu. Otherwise the sched_domain iteration
5105 * breaks. See update_sg_lb_stats().
5107 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5108 group_balance_cpu(sg) == cpu)
5118 sd->groups = groups;
5123 free_sched_groups(first, 0);
5128 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5130 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5131 struct sched_domain *child = sd->child;
5134 cpu = cpumask_first(sched_domain_span(child));
5137 *sg = *per_cpu_ptr(sdd->sg, cpu);
5138 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5139 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5146 * build_sched_groups will build a circular linked list of the groups
5147 * covered by the given span, and will set each group's ->cpumask correctly,
5148 * and ->cpu_power to 0.
5150 * Assumes the sched_domain tree is fully constructed
5153 build_sched_groups(struct sched_domain *sd, int cpu)
5155 struct sched_group *first = NULL, *last = NULL;
5156 struct sd_data *sdd = sd->private;
5157 const struct cpumask *span = sched_domain_span(sd);
5158 struct cpumask *covered;
5161 get_group(cpu, sdd, &sd->groups);
5162 atomic_inc(&sd->groups->ref);
5164 if (cpu != cpumask_first(span))
5167 lockdep_assert_held(&sched_domains_mutex);
5168 covered = sched_domains_tmpmask;
5170 cpumask_clear(covered);
5172 for_each_cpu(i, span) {
5173 struct sched_group *sg;
5176 if (cpumask_test_cpu(i, covered))
5179 group = get_group(i, sdd, &sg);
5180 cpumask_clear(sched_group_cpus(sg));
5182 cpumask_setall(sched_group_mask(sg));
5184 for_each_cpu(j, span) {
5185 if (get_group(j, sdd, NULL) != group)
5188 cpumask_set_cpu(j, covered);
5189 cpumask_set_cpu(j, sched_group_cpus(sg));
5204 * Initialize sched groups cpu_power.
5206 * cpu_power indicates the capacity of sched group, which is used while
5207 * distributing the load between different sched groups in a sched domain.
5208 * Typically cpu_power for all the groups in a sched domain will be same unless
5209 * there are asymmetries in the topology. If there are asymmetries, group
5210 * having more cpu_power will pickup more load compared to the group having
5213 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5215 struct sched_group *sg = sd->groups;
5220 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5222 } while (sg != sd->groups);
5224 if (cpu != group_balance_cpu(sg))
5227 update_group_power(sd, cpu);
5228 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5231 int __weak arch_sd_sibling_asym_packing(void)
5233 return 0*SD_ASYM_PACKING;
5237 * Initializers for schedule domains
5238 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5241 #ifdef CONFIG_SCHED_DEBUG
5242 # define SD_INIT_NAME(sd, type) sd->name = #type
5244 # define SD_INIT_NAME(sd, type) do { } while (0)
5247 #define SD_INIT_FUNC(type) \
5248 static noinline struct sched_domain * \
5249 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5251 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5252 *sd = SD_##type##_INIT; \
5253 SD_INIT_NAME(sd, type); \
5254 sd->private = &tl->data; \
5259 #ifdef CONFIG_SCHED_SMT
5260 SD_INIT_FUNC(SIBLING)
5262 #ifdef CONFIG_SCHED_MC
5265 #ifdef CONFIG_SCHED_BOOK
5269 static int default_relax_domain_level = -1;
5270 int sched_domain_level_max;
5272 static int __init setup_relax_domain_level(char *str)
5274 if (kstrtoint(str, 0, &default_relax_domain_level))
5275 pr_warn("Unable to set relax_domain_level\n");
5279 __setup("relax_domain_level=", setup_relax_domain_level);
5281 static void set_domain_attribute(struct sched_domain *sd,
5282 struct sched_domain_attr *attr)
5286 if (!attr || attr->relax_domain_level < 0) {
5287 if (default_relax_domain_level < 0)
5290 request = default_relax_domain_level;
5292 request = attr->relax_domain_level;
5293 if (request < sd->level) {
5294 /* turn off idle balance on this domain */
5295 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5297 /* turn on idle balance on this domain */
5298 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5302 static void __sdt_free(const struct cpumask *cpu_map);
5303 static int __sdt_alloc(const struct cpumask *cpu_map);
5305 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5306 const struct cpumask *cpu_map)
5310 if (!atomic_read(&d->rd->refcount))
5311 free_rootdomain(&d->rd->rcu); /* fall through */
5313 free_percpu(d->sd); /* fall through */
5315 __sdt_free(cpu_map); /* fall through */
5321 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5322 const struct cpumask *cpu_map)
5324 memset(d, 0, sizeof(*d));
5326 if (__sdt_alloc(cpu_map))
5327 return sa_sd_storage;
5328 d->sd = alloc_percpu(struct sched_domain *);
5330 return sa_sd_storage;
5331 d->rd = alloc_rootdomain();
5334 return sa_rootdomain;
5338 * NULL the sd_data elements we've used to build the sched_domain and
5339 * sched_group structure so that the subsequent __free_domain_allocs()
5340 * will not free the data we're using.
5342 static void claim_allocations(int cpu, struct sched_domain *sd)
5344 struct sd_data *sdd = sd->private;
5346 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5347 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5349 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5350 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5352 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5353 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5356 #ifdef CONFIG_SCHED_SMT
5357 static const struct cpumask *cpu_smt_mask(int cpu)
5359 return topology_thread_cpumask(cpu);
5364 * Topology list, bottom-up.
5366 static struct sched_domain_topology_level default_topology[] = {
5367 #ifdef CONFIG_SCHED_SMT
5368 { sd_init_SIBLING, cpu_smt_mask, },
5370 #ifdef CONFIG_SCHED_MC
5371 { sd_init_MC, cpu_coregroup_mask, },
5373 #ifdef CONFIG_SCHED_BOOK
5374 { sd_init_BOOK, cpu_book_mask, },
5376 { sd_init_CPU, cpu_cpu_mask, },
5380 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5382 #define for_each_sd_topology(tl) \
5383 for (tl = sched_domain_topology; tl->init; tl++)
5387 static int sched_domains_numa_levels;
5388 static int *sched_domains_numa_distance;
5389 static struct cpumask ***sched_domains_numa_masks;
5390 static int sched_domains_curr_level;
5392 static inline int sd_local_flags(int level)
5394 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5397 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5400 static struct sched_domain *
5401 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5403 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5404 int level = tl->numa_level;
5405 int sd_weight = cpumask_weight(
5406 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5408 *sd = (struct sched_domain){
5409 .min_interval = sd_weight,
5410 .max_interval = 2*sd_weight,
5412 .imbalance_pct = 125,
5413 .cache_nice_tries = 2,
5420 .flags = 1*SD_LOAD_BALANCE
5421 | 1*SD_BALANCE_NEWIDLE
5426 | 0*SD_SHARE_CPUPOWER
5427 | 0*SD_SHARE_PKG_RESOURCES
5429 | 0*SD_PREFER_SIBLING
5431 | sd_local_flags(level)
5433 .last_balance = jiffies,
5434 .balance_interval = sd_weight,
5436 SD_INIT_NAME(sd, NUMA);
5437 sd->private = &tl->data;
5440 * Ugly hack to pass state to sd_numa_mask()...
5442 sched_domains_curr_level = tl->numa_level;
5447 static const struct cpumask *sd_numa_mask(int cpu)
5449 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5452 static void sched_numa_warn(const char *str)
5454 static int done = false;
5462 printk(KERN_WARNING "ERROR: %s\n\n", str);
5464 for (i = 0; i < nr_node_ids; i++) {
5465 printk(KERN_WARNING " ");
5466 for (j = 0; j < nr_node_ids; j++)
5467 printk(KERN_CONT "%02d ", node_distance(i,j));
5468 printk(KERN_CONT "\n");
5470 printk(KERN_WARNING "\n");
5473 static bool find_numa_distance(int distance)
5477 if (distance == node_distance(0, 0))
5480 for (i = 0; i < sched_domains_numa_levels; i++) {
5481 if (sched_domains_numa_distance[i] == distance)
5488 static void sched_init_numa(void)
5490 int next_distance, curr_distance = node_distance(0, 0);
5491 struct sched_domain_topology_level *tl;
5495 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5496 if (!sched_domains_numa_distance)
5500 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5501 * unique distances in the node_distance() table.
5503 * Assumes node_distance(0,j) includes all distances in
5504 * node_distance(i,j) in order to avoid cubic time.
5506 next_distance = curr_distance;
5507 for (i = 0; i < nr_node_ids; i++) {
5508 for (j = 0; j < nr_node_ids; j++) {
5509 for (k = 0; k < nr_node_ids; k++) {
5510 int distance = node_distance(i, k);
5512 if (distance > curr_distance &&
5513 (distance < next_distance ||
5514 next_distance == curr_distance))
5515 next_distance = distance;
5518 * While not a strong assumption it would be nice to know
5519 * about cases where if node A is connected to B, B is not
5520 * equally connected to A.
5522 if (sched_debug() && node_distance(k, i) != distance)
5523 sched_numa_warn("Node-distance not symmetric");
5525 if (sched_debug() && i && !find_numa_distance(distance))
5526 sched_numa_warn("Node-0 not representative");
5528 if (next_distance != curr_distance) {
5529 sched_domains_numa_distance[level++] = next_distance;
5530 sched_domains_numa_levels = level;
5531 curr_distance = next_distance;
5536 * In case of sched_debug() we verify the above assumption.
5542 * 'level' contains the number of unique distances, excluding the
5543 * identity distance node_distance(i,i).
5545 * The sched_domains_numa_distance[] array includes the actual distance
5550 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5551 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5552 * the array will contain less then 'level' members. This could be
5553 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5554 * in other functions.
5556 * We reset it to 'level' at the end of this function.
5558 sched_domains_numa_levels = 0;
5560 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5561 if (!sched_domains_numa_masks)
5565 * Now for each level, construct a mask per node which contains all
5566 * cpus of nodes that are that many hops away from us.
5568 for (i = 0; i < level; i++) {
5569 sched_domains_numa_masks[i] =
5570 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5571 if (!sched_domains_numa_masks[i])
5574 for (j = 0; j < nr_node_ids; j++) {
5575 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5579 sched_domains_numa_masks[i][j] = mask;
5581 for (k = 0; k < nr_node_ids; k++) {
5582 if (node_distance(j, k) > sched_domains_numa_distance[i])
5585 cpumask_or(mask, mask, cpumask_of_node(k));
5590 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5591 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5596 * Copy the default topology bits..
5598 for (i = 0; default_topology[i].init; i++)
5599 tl[i] = default_topology[i];
5602 * .. and append 'j' levels of NUMA goodness.
5604 for (j = 0; j < level; i++, j++) {
5605 tl[i] = (struct sched_domain_topology_level){
5606 .init = sd_numa_init,
5607 .mask = sd_numa_mask,
5608 .flags = SDTL_OVERLAP,
5613 sched_domain_topology = tl;
5615 sched_domains_numa_levels = level;
5618 static void sched_domains_numa_masks_set(int cpu)
5621 int node = cpu_to_node(cpu);
5623 for (i = 0; i < sched_domains_numa_levels; i++) {
5624 for (j = 0; j < nr_node_ids; j++) {
5625 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5626 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5631 static void sched_domains_numa_masks_clear(int cpu)
5634 for (i = 0; i < sched_domains_numa_levels; i++) {
5635 for (j = 0; j < nr_node_ids; j++)
5636 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5641 * Update sched_domains_numa_masks[level][node] array when new cpus
5644 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5645 unsigned long action,
5648 int cpu = (long)hcpu;
5650 switch (action & ~CPU_TASKS_FROZEN) {
5652 sched_domains_numa_masks_set(cpu);
5656 sched_domains_numa_masks_clear(cpu);
5666 static inline void sched_init_numa(void)
5670 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5671 unsigned long action,
5676 #endif /* CONFIG_NUMA */
5678 static int __sdt_alloc(const struct cpumask *cpu_map)
5680 struct sched_domain_topology_level *tl;
5683 for_each_sd_topology(tl) {
5684 struct sd_data *sdd = &tl->data;
5686 sdd->sd = alloc_percpu(struct sched_domain *);
5690 sdd->sg = alloc_percpu(struct sched_group *);
5694 sdd->sgp = alloc_percpu(struct sched_group_power *);
5698 for_each_cpu(j, cpu_map) {
5699 struct sched_domain *sd;
5700 struct sched_group *sg;
5701 struct sched_group_power *sgp;
5703 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5704 GFP_KERNEL, cpu_to_node(j));
5708 *per_cpu_ptr(sdd->sd, j) = sd;
5710 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5711 GFP_KERNEL, cpu_to_node(j));
5717 *per_cpu_ptr(sdd->sg, j) = sg;
5719 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5720 GFP_KERNEL, cpu_to_node(j));
5724 *per_cpu_ptr(sdd->sgp, j) = sgp;
5731 static void __sdt_free(const struct cpumask *cpu_map)
5733 struct sched_domain_topology_level *tl;
5736 for_each_sd_topology(tl) {
5737 struct sd_data *sdd = &tl->data;
5739 for_each_cpu(j, cpu_map) {
5740 struct sched_domain *sd;
5743 sd = *per_cpu_ptr(sdd->sd, j);
5744 if (sd && (sd->flags & SD_OVERLAP))
5745 free_sched_groups(sd->groups, 0);
5746 kfree(*per_cpu_ptr(sdd->sd, j));
5750 kfree(*per_cpu_ptr(sdd->sg, j));
5752 kfree(*per_cpu_ptr(sdd->sgp, j));
5754 free_percpu(sdd->sd);
5756 free_percpu(sdd->sg);
5758 free_percpu(sdd->sgp);
5763 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5764 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5765 struct sched_domain *child, int cpu)
5767 struct sched_domain *sd = tl->init(tl, cpu);
5771 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5773 sd->level = child->level + 1;
5774 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5778 set_domain_attribute(sd, attr);
5784 * Build sched domains for a given set of cpus and attach the sched domains
5785 * to the individual cpus
5787 static int build_sched_domains(const struct cpumask *cpu_map,
5788 struct sched_domain_attr *attr)
5790 enum s_alloc alloc_state;
5791 struct sched_domain *sd;
5793 int i, ret = -ENOMEM;
5795 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5796 if (alloc_state != sa_rootdomain)
5799 /* Set up domains for cpus specified by the cpu_map. */
5800 for_each_cpu(i, cpu_map) {
5801 struct sched_domain_topology_level *tl;
5804 for_each_sd_topology(tl) {
5805 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
5806 if (tl == sched_domain_topology)
5807 *per_cpu_ptr(d.sd, i) = sd;
5808 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
5809 sd->flags |= SD_OVERLAP;
5810 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
5815 /* Build the groups for the domains */
5816 for_each_cpu(i, cpu_map) {
5817 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5818 sd->span_weight = cpumask_weight(sched_domain_span(sd));
5819 if (sd->flags & SD_OVERLAP) {
5820 if (build_overlap_sched_groups(sd, i))
5823 if (build_sched_groups(sd, i))
5829 /* Calculate CPU power for physical packages and nodes */
5830 for (i = nr_cpumask_bits-1; i >= 0; i--) {
5831 if (!cpumask_test_cpu(i, cpu_map))
5834 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5835 claim_allocations(i, sd);
5836 init_sched_groups_power(i, sd);
5840 /* Attach the domains */
5842 for_each_cpu(i, cpu_map) {
5843 sd = *per_cpu_ptr(d.sd, i);
5844 cpu_attach_domain(sd, d.rd, i);
5850 __free_domain_allocs(&d, alloc_state, cpu_map);
5854 static cpumask_var_t *doms_cur; /* current sched domains */
5855 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
5856 static struct sched_domain_attr *dattr_cur;
5857 /* attribues of custom domains in 'doms_cur' */
5860 * Special case: If a kmalloc of a doms_cur partition (array of
5861 * cpumask) fails, then fallback to a single sched domain,
5862 * as determined by the single cpumask fallback_doms.
5864 static cpumask_var_t fallback_doms;
5867 * arch_update_cpu_topology lets virtualized architectures update the
5868 * cpu core maps. It is supposed to return 1 if the topology changed
5869 * or 0 if it stayed the same.
5871 int __attribute__((weak)) arch_update_cpu_topology(void)
5876 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
5879 cpumask_var_t *doms;
5881 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
5884 for (i = 0; i < ndoms; i++) {
5885 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
5886 free_sched_domains(doms, i);
5893 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
5896 for (i = 0; i < ndoms; i++)
5897 free_cpumask_var(doms[i]);
5902 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5903 * For now this just excludes isolated cpus, but could be used to
5904 * exclude other special cases in the future.
5906 static int init_sched_domains(const struct cpumask *cpu_map)
5910 arch_update_cpu_topology();
5912 doms_cur = alloc_sched_domains(ndoms_cur);
5914 doms_cur = &fallback_doms;
5915 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
5916 err = build_sched_domains(doms_cur[0], NULL);
5917 register_sched_domain_sysctl();
5923 * Detach sched domains from a group of cpus specified in cpu_map
5924 * These cpus will now be attached to the NULL domain
5926 static void detach_destroy_domains(const struct cpumask *cpu_map)
5931 for_each_cpu(i, cpu_map)
5932 cpu_attach_domain(NULL, &def_root_domain, i);
5936 /* handle null as "default" */
5937 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
5938 struct sched_domain_attr *new, int idx_new)
5940 struct sched_domain_attr tmp;
5947 return !memcmp(cur ? (cur + idx_cur) : &tmp,
5948 new ? (new + idx_new) : &tmp,
5949 sizeof(struct sched_domain_attr));
5953 * Partition sched domains as specified by the 'ndoms_new'
5954 * cpumasks in the array doms_new[] of cpumasks. This compares
5955 * doms_new[] to the current sched domain partitioning, doms_cur[].
5956 * It destroys each deleted domain and builds each new domain.
5958 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
5959 * The masks don't intersect (don't overlap.) We should setup one
5960 * sched domain for each mask. CPUs not in any of the cpumasks will
5961 * not be load balanced. If the same cpumask appears both in the
5962 * current 'doms_cur' domains and in the new 'doms_new', we can leave
5965 * The passed in 'doms_new' should be allocated using
5966 * alloc_sched_domains. This routine takes ownership of it and will
5967 * free_sched_domains it when done with it. If the caller failed the
5968 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
5969 * and partition_sched_domains() will fallback to the single partition
5970 * 'fallback_doms', it also forces the domains to be rebuilt.
5972 * If doms_new == NULL it will be replaced with cpu_online_mask.
5973 * ndoms_new == 0 is a special case for destroying existing domains,
5974 * and it will not create the default domain.
5976 * Call with hotplug lock held
5978 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
5979 struct sched_domain_attr *dattr_new)
5984 mutex_lock(&sched_domains_mutex);
5986 /* always unregister in case we don't destroy any domains */
5987 unregister_sched_domain_sysctl();
5989 /* Let architecture update cpu core mappings. */
5990 new_topology = arch_update_cpu_topology();
5992 n = doms_new ? ndoms_new : 0;
5994 /* Destroy deleted domains */
5995 for (i = 0; i < ndoms_cur; i++) {
5996 for (j = 0; j < n && !new_topology; j++) {
5997 if (cpumask_equal(doms_cur[i], doms_new[j])
5998 && dattrs_equal(dattr_cur, i, dattr_new, j))
6001 /* no match - a current sched domain not in new doms_new[] */
6002 detach_destroy_domains(doms_cur[i]);
6008 if (doms_new == NULL) {
6010 doms_new = &fallback_doms;
6011 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6012 WARN_ON_ONCE(dattr_new);
6015 /* Build new domains */
6016 for (i = 0; i < ndoms_new; i++) {
6017 for (j = 0; j < n && !new_topology; j++) {
6018 if (cpumask_equal(doms_new[i], doms_cur[j])
6019 && dattrs_equal(dattr_new, i, dattr_cur, j))
6022 /* no match - add a new doms_new */
6023 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6028 /* Remember the new sched domains */
6029 if (doms_cur != &fallback_doms)
6030 free_sched_domains(doms_cur, ndoms_cur);
6031 kfree(dattr_cur); /* kfree(NULL) is safe */
6032 doms_cur = doms_new;
6033 dattr_cur = dattr_new;
6034 ndoms_cur = ndoms_new;
6036 register_sched_domain_sysctl();
6038 mutex_unlock(&sched_domains_mutex);
6041 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6044 * Update cpusets according to cpu_active mask. If cpusets are
6045 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6046 * around partition_sched_domains().
6048 * If we come here as part of a suspend/resume, don't touch cpusets because we
6049 * want to restore it back to its original state upon resume anyway.
6051 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6055 case CPU_ONLINE_FROZEN:
6056 case CPU_DOWN_FAILED_FROZEN:
6059 * num_cpus_frozen tracks how many CPUs are involved in suspend
6060 * resume sequence. As long as this is not the last online
6061 * operation in the resume sequence, just build a single sched
6062 * domain, ignoring cpusets.
6065 if (likely(num_cpus_frozen)) {
6066 partition_sched_domains(1, NULL, NULL);
6071 * This is the last CPU online operation. So fall through and
6072 * restore the original sched domains by considering the
6073 * cpuset configurations.
6077 case CPU_DOWN_FAILED:
6078 cpuset_update_active_cpus(true);
6086 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6090 case CPU_DOWN_PREPARE:
6091 cpuset_update_active_cpus(false);
6093 case CPU_DOWN_PREPARE_FROZEN:
6095 partition_sched_domains(1, NULL, NULL);
6103 void __init sched_init_smp(void)
6105 cpumask_var_t non_isolated_cpus;
6107 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6108 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6113 * There's no userspace yet to cause hotplug operations; hence all the
6114 * cpu masks are stable and all blatant races in the below code cannot
6117 mutex_lock(&sched_domains_mutex);
6118 init_sched_domains(cpu_active_mask);
6119 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6120 if (cpumask_empty(non_isolated_cpus))
6121 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6122 mutex_unlock(&sched_domains_mutex);
6124 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6125 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6126 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6130 /* Move init over to a non-isolated CPU */
6131 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6133 sched_init_granularity();
6134 free_cpumask_var(non_isolated_cpus);
6136 init_sched_rt_class();
6139 void __init sched_init_smp(void)
6141 sched_init_granularity();
6143 #endif /* CONFIG_SMP */
6145 const_debug unsigned int sysctl_timer_migration = 1;
6147 int in_sched_functions(unsigned long addr)
6149 return in_lock_functions(addr) ||
6150 (addr >= (unsigned long)__sched_text_start
6151 && addr < (unsigned long)__sched_text_end);
6154 #ifdef CONFIG_CGROUP_SCHED
6156 * Default task group.
6157 * Every task in system belongs to this group at bootup.
6159 struct task_group root_task_group;
6160 LIST_HEAD(task_groups);
6163 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6165 void __init sched_init(void)
6168 unsigned long alloc_size = 0, ptr;
6170 #ifdef CONFIG_FAIR_GROUP_SCHED
6171 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6173 #ifdef CONFIG_RT_GROUP_SCHED
6174 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6176 #ifdef CONFIG_CPUMASK_OFFSTACK
6177 alloc_size += num_possible_cpus() * cpumask_size();
6180 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6182 #ifdef CONFIG_FAIR_GROUP_SCHED
6183 root_task_group.se = (struct sched_entity **)ptr;
6184 ptr += nr_cpu_ids * sizeof(void **);
6186 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6187 ptr += nr_cpu_ids * sizeof(void **);
6189 #endif /* CONFIG_FAIR_GROUP_SCHED */
6190 #ifdef CONFIG_RT_GROUP_SCHED
6191 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6192 ptr += nr_cpu_ids * sizeof(void **);
6194 root_task_group.rt_rq = (struct rt_rq **)ptr;
6195 ptr += nr_cpu_ids * sizeof(void **);
6197 #endif /* CONFIG_RT_GROUP_SCHED */
6198 #ifdef CONFIG_CPUMASK_OFFSTACK
6199 for_each_possible_cpu(i) {
6200 per_cpu(load_balance_mask, i) = (void *)ptr;
6201 ptr += cpumask_size();
6203 #endif /* CONFIG_CPUMASK_OFFSTACK */
6207 init_defrootdomain();
6210 init_rt_bandwidth(&def_rt_bandwidth,
6211 global_rt_period(), global_rt_runtime());
6213 #ifdef CONFIG_RT_GROUP_SCHED
6214 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6215 global_rt_period(), global_rt_runtime());
6216 #endif /* CONFIG_RT_GROUP_SCHED */
6218 #ifdef CONFIG_CGROUP_SCHED
6219 list_add(&root_task_group.list, &task_groups);
6220 INIT_LIST_HEAD(&root_task_group.children);
6221 INIT_LIST_HEAD(&root_task_group.siblings);
6222 autogroup_init(&init_task);
6224 #endif /* CONFIG_CGROUP_SCHED */
6226 for_each_possible_cpu(i) {
6230 raw_spin_lock_init(&rq->lock);
6232 rq->calc_load_active = 0;
6233 rq->calc_load_update = jiffies + LOAD_FREQ;
6234 init_cfs_rq(&rq->cfs);
6235 init_rt_rq(&rq->rt, rq);
6236 #ifdef CONFIG_FAIR_GROUP_SCHED
6237 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6238 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6240 * How much cpu bandwidth does root_task_group get?
6242 * In case of task-groups formed thr' the cgroup filesystem, it
6243 * gets 100% of the cpu resources in the system. This overall
6244 * system cpu resource is divided among the tasks of
6245 * root_task_group and its child task-groups in a fair manner,
6246 * based on each entity's (task or task-group's) weight
6247 * (se->load.weight).
6249 * In other words, if root_task_group has 10 tasks of weight
6250 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6251 * then A0's share of the cpu resource is:
6253 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6255 * We achieve this by letting root_task_group's tasks sit
6256 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6258 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6259 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6260 #endif /* CONFIG_FAIR_GROUP_SCHED */
6262 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6263 #ifdef CONFIG_RT_GROUP_SCHED
6264 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6265 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6268 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6269 rq->cpu_load[j] = 0;
6271 rq->last_load_update_tick = jiffies;
6276 rq->cpu_power = SCHED_POWER_SCALE;
6277 rq->post_schedule = 0;
6278 rq->active_balance = 0;
6279 rq->next_balance = jiffies;
6284 rq->avg_idle = 2*sysctl_sched_migration_cost;
6285 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6287 INIT_LIST_HEAD(&rq->cfs_tasks);
6289 rq_attach_root(rq, &def_root_domain);
6290 #ifdef CONFIG_NO_HZ_COMMON
6293 #ifdef CONFIG_NO_HZ_FULL
6294 rq->last_sched_tick = 0;
6298 atomic_set(&rq->nr_iowait, 0);
6301 set_load_weight(&init_task);
6303 #ifdef CONFIG_PREEMPT_NOTIFIERS
6304 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6307 #ifdef CONFIG_RT_MUTEXES
6308 plist_head_init(&init_task.pi_waiters);
6312 * The boot idle thread does lazy MMU switching as well:
6314 atomic_inc(&init_mm.mm_count);
6315 enter_lazy_tlb(&init_mm, current);
6318 * Make us the idle thread. Technically, schedule() should not be
6319 * called from this thread, however somewhere below it might be,
6320 * but because we are the idle thread, we just pick up running again
6321 * when this runqueue becomes "idle".
6323 init_idle(current, smp_processor_id());
6325 calc_load_update = jiffies + LOAD_FREQ;
6328 * During early bootup we pretend to be a normal task:
6330 current->sched_class = &fair_sched_class;
6333 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6334 /* May be allocated at isolcpus cmdline parse time */
6335 if (cpu_isolated_map == NULL)
6336 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6337 idle_thread_set_boot_cpu();
6339 init_sched_fair_class();
6341 scheduler_running = 1;
6344 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6345 static inline int preempt_count_equals(int preempt_offset)
6347 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6349 return (nested == preempt_offset);
6352 void __might_sleep(const char *file, int line, int preempt_offset)
6354 static unsigned long prev_jiffy; /* ratelimiting */
6356 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6357 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6358 system_state != SYSTEM_RUNNING || oops_in_progress)
6360 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6362 prev_jiffy = jiffies;
6365 "BUG: sleeping function called from invalid context at %s:%d\n",
6368 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6369 in_atomic(), irqs_disabled(),
6370 current->pid, current->comm);
6372 debug_show_held_locks(current);
6373 if (irqs_disabled())
6374 print_irqtrace_events(current);
6377 EXPORT_SYMBOL(__might_sleep);
6380 #ifdef CONFIG_MAGIC_SYSRQ
6381 static void normalize_task(struct rq *rq, struct task_struct *p)
6383 const struct sched_class *prev_class = p->sched_class;
6384 int old_prio = p->prio;
6389 dequeue_task(rq, p, 0);
6390 __setscheduler(rq, p, SCHED_NORMAL, 0);
6392 enqueue_task(rq, p, 0);
6393 resched_task(rq->curr);
6396 check_class_changed(rq, p, prev_class, old_prio);
6399 void normalize_rt_tasks(void)
6401 struct task_struct *g, *p;
6402 unsigned long flags;
6405 read_lock_irqsave(&tasklist_lock, flags);
6406 do_each_thread(g, p) {
6408 * Only normalize user tasks:
6413 p->se.exec_start = 0;
6414 #ifdef CONFIG_SCHEDSTATS
6415 p->se.statistics.wait_start = 0;
6416 p->se.statistics.sleep_start = 0;
6417 p->se.statistics.block_start = 0;
6422 * Renice negative nice level userspace
6425 if (TASK_NICE(p) < 0 && p->mm)
6426 set_user_nice(p, 0);
6430 raw_spin_lock(&p->pi_lock);
6431 rq = __task_rq_lock(p);
6433 normalize_task(rq, p);
6435 __task_rq_unlock(rq);
6436 raw_spin_unlock(&p->pi_lock);
6437 } while_each_thread(g, p);
6439 read_unlock_irqrestore(&tasklist_lock, flags);
6442 #endif /* CONFIG_MAGIC_SYSRQ */
6444 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6446 * These functions are only useful for the IA64 MCA handling, or kdb.
6448 * They can only be called when the whole system has been
6449 * stopped - every CPU needs to be quiescent, and no scheduling
6450 * activity can take place. Using them for anything else would
6451 * be a serious bug, and as a result, they aren't even visible
6452 * under any other configuration.
6456 * curr_task - return the current task for a given cpu.
6457 * @cpu: the processor in question.
6459 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6461 * Return: The current task for @cpu.
6463 struct task_struct *curr_task(int cpu)
6465 return cpu_curr(cpu);
6468 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6472 * set_curr_task - set the current task for a given cpu.
6473 * @cpu: the processor in question.
6474 * @p: the task pointer to set.
6476 * Description: This function must only be used when non-maskable interrupts
6477 * are serviced on a separate stack. It allows the architecture to switch the
6478 * notion of the current task on a cpu in a non-blocking manner. This function
6479 * must be called with all CPU's synchronized, and interrupts disabled, the
6480 * and caller must save the original value of the current task (see
6481 * curr_task() above) and restore that value before reenabling interrupts and
6482 * re-starting the system.
6484 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6486 void set_curr_task(int cpu, struct task_struct *p)
6493 #ifdef CONFIG_CGROUP_SCHED
6494 /* task_group_lock serializes the addition/removal of task groups */
6495 static DEFINE_SPINLOCK(task_group_lock);
6497 static void free_sched_group(struct task_group *tg)
6499 free_fair_sched_group(tg);
6500 free_rt_sched_group(tg);
6505 /* allocate runqueue etc for a new task group */
6506 struct task_group *sched_create_group(struct task_group *parent)
6508 struct task_group *tg;
6510 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6512 return ERR_PTR(-ENOMEM);
6514 if (!alloc_fair_sched_group(tg, parent))
6517 if (!alloc_rt_sched_group(tg, parent))
6523 free_sched_group(tg);
6524 return ERR_PTR(-ENOMEM);
6527 void sched_online_group(struct task_group *tg, struct task_group *parent)
6529 unsigned long flags;
6531 spin_lock_irqsave(&task_group_lock, flags);
6532 list_add_rcu(&tg->list, &task_groups);
6534 WARN_ON(!parent); /* root should already exist */
6536 tg->parent = parent;
6537 INIT_LIST_HEAD(&tg->children);
6538 list_add_rcu(&tg->siblings, &parent->children);
6539 spin_unlock_irqrestore(&task_group_lock, flags);
6542 /* rcu callback to free various structures associated with a task group */
6543 static void free_sched_group_rcu(struct rcu_head *rhp)
6545 /* now it should be safe to free those cfs_rqs */
6546 free_sched_group(container_of(rhp, struct task_group, rcu));
6549 /* Destroy runqueue etc associated with a task group */
6550 void sched_destroy_group(struct task_group *tg)
6552 /* wait for possible concurrent references to cfs_rqs complete */
6553 call_rcu(&tg->rcu, free_sched_group_rcu);
6556 void sched_offline_group(struct task_group *tg)
6558 unsigned long flags;
6561 /* end participation in shares distribution */
6562 for_each_possible_cpu(i)
6563 unregister_fair_sched_group(tg, i);
6565 spin_lock_irqsave(&task_group_lock, flags);
6566 list_del_rcu(&tg->list);
6567 list_del_rcu(&tg->siblings);
6568 spin_unlock_irqrestore(&task_group_lock, flags);
6571 /* change task's runqueue when it moves between groups.
6572 * The caller of this function should have put the task in its new group
6573 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6574 * reflect its new group.
6576 void sched_move_task(struct task_struct *tsk)
6578 struct task_group *tg;
6580 unsigned long flags;
6583 rq = task_rq_lock(tsk, &flags);
6585 running = task_current(rq, tsk);
6589 dequeue_task(rq, tsk, 0);
6590 if (unlikely(running))
6591 tsk->sched_class->put_prev_task(rq, tsk);
6593 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6594 lockdep_is_held(&tsk->sighand->siglock)),
6595 struct task_group, css);
6596 tg = autogroup_task_group(tsk, tg);
6597 tsk->sched_task_group = tg;
6599 #ifdef CONFIG_FAIR_GROUP_SCHED
6600 if (tsk->sched_class->task_move_group)
6601 tsk->sched_class->task_move_group(tsk, on_rq);
6604 set_task_rq(tsk, task_cpu(tsk));
6606 if (unlikely(running))
6607 tsk->sched_class->set_curr_task(rq);
6609 enqueue_task(rq, tsk, 0);
6611 task_rq_unlock(rq, tsk, &flags);
6613 #endif /* CONFIG_CGROUP_SCHED */
6615 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6616 static unsigned long to_ratio(u64 period, u64 runtime)
6618 if (runtime == RUNTIME_INF)
6621 return div64_u64(runtime << 20, period);
6625 #ifdef CONFIG_RT_GROUP_SCHED
6627 * Ensure that the real time constraints are schedulable.
6629 static DEFINE_MUTEX(rt_constraints_mutex);
6631 /* Must be called with tasklist_lock held */
6632 static inline int tg_has_rt_tasks(struct task_group *tg)
6634 struct task_struct *g, *p;
6636 do_each_thread(g, p) {
6637 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6639 } while_each_thread(g, p);
6644 struct rt_schedulable_data {
6645 struct task_group *tg;
6650 static int tg_rt_schedulable(struct task_group *tg, void *data)
6652 struct rt_schedulable_data *d = data;
6653 struct task_group *child;
6654 unsigned long total, sum = 0;
6655 u64 period, runtime;
6657 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6658 runtime = tg->rt_bandwidth.rt_runtime;
6661 period = d->rt_period;
6662 runtime = d->rt_runtime;
6666 * Cannot have more runtime than the period.
6668 if (runtime > period && runtime != RUNTIME_INF)
6672 * Ensure we don't starve existing RT tasks.
6674 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6677 total = to_ratio(period, runtime);
6680 * Nobody can have more than the global setting allows.
6682 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6686 * The sum of our children's runtime should not exceed our own.
6688 list_for_each_entry_rcu(child, &tg->children, siblings) {
6689 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6690 runtime = child->rt_bandwidth.rt_runtime;
6692 if (child == d->tg) {
6693 period = d->rt_period;
6694 runtime = d->rt_runtime;
6697 sum += to_ratio(period, runtime);
6706 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6710 struct rt_schedulable_data data = {
6712 .rt_period = period,
6713 .rt_runtime = runtime,
6717 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6723 static int tg_set_rt_bandwidth(struct task_group *tg,
6724 u64 rt_period, u64 rt_runtime)
6728 mutex_lock(&rt_constraints_mutex);
6729 read_lock(&tasklist_lock);
6730 err = __rt_schedulable(tg, rt_period, rt_runtime);
6734 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6735 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6736 tg->rt_bandwidth.rt_runtime = rt_runtime;
6738 for_each_possible_cpu(i) {
6739 struct rt_rq *rt_rq = tg->rt_rq[i];
6741 raw_spin_lock(&rt_rq->rt_runtime_lock);
6742 rt_rq->rt_runtime = rt_runtime;
6743 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6745 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6747 read_unlock(&tasklist_lock);
6748 mutex_unlock(&rt_constraints_mutex);
6753 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6755 u64 rt_runtime, rt_period;
6757 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6758 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6759 if (rt_runtime_us < 0)
6760 rt_runtime = RUNTIME_INF;
6762 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6765 static long sched_group_rt_runtime(struct task_group *tg)
6769 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6772 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6773 do_div(rt_runtime_us, NSEC_PER_USEC);
6774 return rt_runtime_us;
6777 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6779 u64 rt_runtime, rt_period;
6781 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6782 rt_runtime = tg->rt_bandwidth.rt_runtime;
6787 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6790 static long sched_group_rt_period(struct task_group *tg)
6794 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6795 do_div(rt_period_us, NSEC_PER_USEC);
6796 return rt_period_us;
6799 static int sched_rt_global_constraints(void)
6801 u64 runtime, period;
6804 if (sysctl_sched_rt_period <= 0)
6807 runtime = global_rt_runtime();
6808 period = global_rt_period();
6811 * Sanity check on the sysctl variables.
6813 if (runtime > period && runtime != RUNTIME_INF)
6816 mutex_lock(&rt_constraints_mutex);
6817 read_lock(&tasklist_lock);
6818 ret = __rt_schedulable(NULL, 0, 0);
6819 read_unlock(&tasklist_lock);
6820 mutex_unlock(&rt_constraints_mutex);
6825 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6827 /* Don't accept realtime tasks when there is no way for them to run */
6828 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6834 #else /* !CONFIG_RT_GROUP_SCHED */
6835 static int sched_rt_global_constraints(void)
6837 unsigned long flags;
6840 if (sysctl_sched_rt_period <= 0)
6844 * There's always some RT tasks in the root group
6845 * -- migration, kstopmachine etc..
6847 if (sysctl_sched_rt_runtime == 0)
6850 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6851 for_each_possible_cpu(i) {
6852 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6854 raw_spin_lock(&rt_rq->rt_runtime_lock);
6855 rt_rq->rt_runtime = global_rt_runtime();
6856 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6858 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6862 #endif /* CONFIG_RT_GROUP_SCHED */
6864 int sched_rr_handler(struct ctl_table *table, int write,
6865 void __user *buffer, size_t *lenp,
6869 static DEFINE_MUTEX(mutex);
6872 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6873 /* make sure that internally we keep jiffies */
6874 /* also, writing zero resets timeslice to default */
6875 if (!ret && write) {
6876 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
6877 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
6879 mutex_unlock(&mutex);
6883 int sched_rt_handler(struct ctl_table *table, int write,
6884 void __user *buffer, size_t *lenp,
6888 int old_period, old_runtime;
6889 static DEFINE_MUTEX(mutex);
6892 old_period = sysctl_sched_rt_period;
6893 old_runtime = sysctl_sched_rt_runtime;
6895 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6897 if (!ret && write) {
6898 ret = sched_rt_global_constraints();
6900 sysctl_sched_rt_period = old_period;
6901 sysctl_sched_rt_runtime = old_runtime;
6903 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6904 def_rt_bandwidth.rt_period =
6905 ns_to_ktime(global_rt_period());
6908 mutex_unlock(&mutex);
6913 #ifdef CONFIG_CGROUP_SCHED
6915 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6917 return css ? container_of(css, struct task_group, css) : NULL;
6920 static struct cgroup_subsys_state *
6921 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6923 struct task_group *parent = css_tg(parent_css);
6924 struct task_group *tg;
6927 /* This is early initialization for the top cgroup */
6928 return &root_task_group.css;
6931 tg = sched_create_group(parent);
6933 return ERR_PTR(-ENOMEM);
6938 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6940 struct task_group *tg = css_tg(css);
6941 struct task_group *parent = css_tg(css_parent(css));
6944 sched_online_group(tg, parent);
6948 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6950 struct task_group *tg = css_tg(css);
6952 sched_destroy_group(tg);
6955 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
6957 struct task_group *tg = css_tg(css);
6959 sched_offline_group(tg);
6962 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
6963 struct cgroup_taskset *tset)
6965 struct task_struct *task;
6967 cgroup_taskset_for_each(task, css, tset) {
6968 #ifdef CONFIG_RT_GROUP_SCHED
6969 if (!sched_rt_can_attach(css_tg(css), task))
6972 /* We don't support RT-tasks being in separate groups */
6973 if (task->sched_class != &fair_sched_class)
6980 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
6981 struct cgroup_taskset *tset)
6983 struct task_struct *task;
6985 cgroup_taskset_for_each(task, css, tset)
6986 sched_move_task(task);
6989 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
6990 struct cgroup_subsys_state *old_css,
6991 struct task_struct *task)
6994 * cgroup_exit() is called in the copy_process() failure path.
6995 * Ignore this case since the task hasn't ran yet, this avoids
6996 * trying to poke a half freed task state from generic code.
6998 if (!(task->flags & PF_EXITING))
7001 sched_move_task(task);
7004 #ifdef CONFIG_FAIR_GROUP_SCHED
7005 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7006 struct cftype *cftype, u64 shareval)
7008 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7011 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7014 struct task_group *tg = css_tg(css);
7016 return (u64) scale_load_down(tg->shares);
7019 #ifdef CONFIG_CFS_BANDWIDTH
7020 static DEFINE_MUTEX(cfs_constraints_mutex);
7022 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7023 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7025 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7027 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7029 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7030 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7032 if (tg == &root_task_group)
7036 * Ensure we have at some amount of bandwidth every period. This is
7037 * to prevent reaching a state of large arrears when throttled via
7038 * entity_tick() resulting in prolonged exit starvation.
7040 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7044 * Likewise, bound things on the otherside by preventing insane quota
7045 * periods. This also allows us to normalize in computing quota
7048 if (period > max_cfs_quota_period)
7051 mutex_lock(&cfs_constraints_mutex);
7052 ret = __cfs_schedulable(tg, period, quota);
7056 runtime_enabled = quota != RUNTIME_INF;
7057 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7059 * If we need to toggle cfs_bandwidth_used, off->on must occur
7060 * before making related changes, and on->off must occur afterwards
7062 if (runtime_enabled && !runtime_was_enabled)
7063 cfs_bandwidth_usage_inc();
7064 raw_spin_lock_irq(&cfs_b->lock);
7065 cfs_b->period = ns_to_ktime(period);
7066 cfs_b->quota = quota;
7068 __refill_cfs_bandwidth_runtime(cfs_b);
7069 /* restart the period timer (if active) to handle new period expiry */
7070 if (runtime_enabled && cfs_b->timer_active) {
7071 /* force a reprogram */
7072 cfs_b->timer_active = 0;
7073 __start_cfs_bandwidth(cfs_b);
7075 raw_spin_unlock_irq(&cfs_b->lock);
7077 for_each_possible_cpu(i) {
7078 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7079 struct rq *rq = cfs_rq->rq;
7081 raw_spin_lock_irq(&rq->lock);
7082 cfs_rq->runtime_enabled = runtime_enabled;
7083 cfs_rq->runtime_remaining = 0;
7085 if (cfs_rq->throttled)
7086 unthrottle_cfs_rq(cfs_rq);
7087 raw_spin_unlock_irq(&rq->lock);
7089 if (runtime_was_enabled && !runtime_enabled)
7090 cfs_bandwidth_usage_dec();
7092 mutex_unlock(&cfs_constraints_mutex);
7097 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7101 period = ktime_to_ns(tg->cfs_bandwidth.period);
7102 if (cfs_quota_us < 0)
7103 quota = RUNTIME_INF;
7105 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7107 return tg_set_cfs_bandwidth(tg, period, quota);
7110 long tg_get_cfs_quota(struct task_group *tg)
7114 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7117 quota_us = tg->cfs_bandwidth.quota;
7118 do_div(quota_us, NSEC_PER_USEC);
7123 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7127 period = (u64)cfs_period_us * NSEC_PER_USEC;
7128 quota = tg->cfs_bandwidth.quota;
7130 return tg_set_cfs_bandwidth(tg, period, quota);
7133 long tg_get_cfs_period(struct task_group *tg)
7137 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7138 do_div(cfs_period_us, NSEC_PER_USEC);
7140 return cfs_period_us;
7143 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7146 return tg_get_cfs_quota(css_tg(css));
7149 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7150 struct cftype *cftype, s64 cfs_quota_us)
7152 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7155 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7158 return tg_get_cfs_period(css_tg(css));
7161 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7162 struct cftype *cftype, u64 cfs_period_us)
7164 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7167 struct cfs_schedulable_data {
7168 struct task_group *tg;
7173 * normalize group quota/period to be quota/max_period
7174 * note: units are usecs
7176 static u64 normalize_cfs_quota(struct task_group *tg,
7177 struct cfs_schedulable_data *d)
7185 period = tg_get_cfs_period(tg);
7186 quota = tg_get_cfs_quota(tg);
7189 /* note: these should typically be equivalent */
7190 if (quota == RUNTIME_INF || quota == -1)
7193 return to_ratio(period, quota);
7196 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7198 struct cfs_schedulable_data *d = data;
7199 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7200 s64 quota = 0, parent_quota = -1;
7203 quota = RUNTIME_INF;
7205 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7207 quota = normalize_cfs_quota(tg, d);
7208 parent_quota = parent_b->hierarchal_quota;
7211 * ensure max(child_quota) <= parent_quota, inherit when no
7214 if (quota == RUNTIME_INF)
7215 quota = parent_quota;
7216 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7219 cfs_b->hierarchal_quota = quota;
7224 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7227 struct cfs_schedulable_data data = {
7233 if (quota != RUNTIME_INF) {
7234 do_div(data.period, NSEC_PER_USEC);
7235 do_div(data.quota, NSEC_PER_USEC);
7239 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7245 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7246 struct cgroup_map_cb *cb)
7248 struct task_group *tg = css_tg(css);
7249 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7251 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7252 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7253 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7257 #endif /* CONFIG_CFS_BANDWIDTH */
7258 #endif /* CONFIG_FAIR_GROUP_SCHED */
7260 #ifdef CONFIG_RT_GROUP_SCHED
7261 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7262 struct cftype *cft, s64 val)
7264 return sched_group_set_rt_runtime(css_tg(css), val);
7267 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7270 return sched_group_rt_runtime(css_tg(css));
7273 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7274 struct cftype *cftype, u64 rt_period_us)
7276 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7279 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7282 return sched_group_rt_period(css_tg(css));
7284 #endif /* CONFIG_RT_GROUP_SCHED */
7286 static struct cftype cpu_files[] = {
7287 #ifdef CONFIG_FAIR_GROUP_SCHED
7290 .read_u64 = cpu_shares_read_u64,
7291 .write_u64 = cpu_shares_write_u64,
7294 #ifdef CONFIG_CFS_BANDWIDTH
7296 .name = "cfs_quota_us",
7297 .read_s64 = cpu_cfs_quota_read_s64,
7298 .write_s64 = cpu_cfs_quota_write_s64,
7301 .name = "cfs_period_us",
7302 .read_u64 = cpu_cfs_period_read_u64,
7303 .write_u64 = cpu_cfs_period_write_u64,
7307 .read_map = cpu_stats_show,
7310 #ifdef CONFIG_RT_GROUP_SCHED
7312 .name = "rt_runtime_us",
7313 .read_s64 = cpu_rt_runtime_read,
7314 .write_s64 = cpu_rt_runtime_write,
7317 .name = "rt_period_us",
7318 .read_u64 = cpu_rt_period_read_uint,
7319 .write_u64 = cpu_rt_period_write_uint,
7325 struct cgroup_subsys cpu_cgroup_subsys = {
7327 .css_alloc = cpu_cgroup_css_alloc,
7328 .css_free = cpu_cgroup_css_free,
7329 .css_online = cpu_cgroup_css_online,
7330 .css_offline = cpu_cgroup_css_offline,
7331 .can_attach = cpu_cgroup_can_attach,
7332 .attach = cpu_cgroup_attach,
7333 .exit = cpu_cgroup_exit,
7334 .subsys_id = cpu_cgroup_subsys_id,
7335 .base_cftypes = cpu_files,
7339 #endif /* CONFIG_CGROUP_SCHED */
7341 void dump_cpu_task(int cpu)
7343 pr_info("Task dump for CPU %d:\n", cpu);
7344 sched_show_task(cpu_curr(cpu));