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 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2258 * 64-bit doesn't need locks to atomically read a 64bit value.
2259 * So we have a optimization chance when the task's delta_exec is 0.
2260 * Reading ->on_cpu is racy, but this is ok.
2262 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2263 * If we race with it entering cpu, unaccounted time is 0. This is
2264 * indistinguishable from the read occurring a few cycles earlier.
2267 return p->se.sum_exec_runtime;
2270 rq = task_rq_lock(p, &flags);
2271 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2272 task_rq_unlock(rq, p, &flags);
2278 * This function gets called by the timer code, with HZ frequency.
2279 * We call it with interrupts disabled.
2281 void scheduler_tick(void)
2283 int cpu = smp_processor_id();
2284 struct rq *rq = cpu_rq(cpu);
2285 struct task_struct *curr = rq->curr;
2289 raw_spin_lock(&rq->lock);
2290 update_rq_clock(rq);
2291 curr->sched_class->task_tick(rq, curr, 0);
2292 update_cpu_load_active(rq);
2293 raw_spin_unlock(&rq->lock);
2295 perf_event_task_tick();
2298 rq->idle_balance = idle_cpu(cpu);
2299 trigger_load_balance(rq, cpu);
2301 rq_last_tick_reset(rq);
2304 #ifdef CONFIG_NO_HZ_FULL
2306 * scheduler_tick_max_deferment
2308 * Keep at least one tick per second when a single
2309 * active task is running because the scheduler doesn't
2310 * yet completely support full dynticks environment.
2312 * This makes sure that uptime, CFS vruntime, load
2313 * balancing, etc... continue to move forward, even
2314 * with a very low granularity.
2316 * Return: Maximum deferment in nanoseconds.
2318 u64 scheduler_tick_max_deferment(void)
2320 struct rq *rq = this_rq();
2321 unsigned long next, now = ACCESS_ONCE(jiffies);
2323 next = rq->last_sched_tick + HZ;
2325 if (time_before_eq(next, now))
2328 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2332 notrace unsigned long get_parent_ip(unsigned long addr)
2334 if (in_lock_functions(addr)) {
2335 addr = CALLER_ADDR2;
2336 if (in_lock_functions(addr))
2337 addr = CALLER_ADDR3;
2342 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2343 defined(CONFIG_PREEMPT_TRACER))
2345 void __kprobes preempt_count_add(int val)
2347 #ifdef CONFIG_DEBUG_PREEMPT
2351 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2354 __preempt_count_add(val);
2355 #ifdef CONFIG_DEBUG_PREEMPT
2357 * Spinlock count overflowing soon?
2359 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2362 if (preempt_count() == val)
2363 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2365 EXPORT_SYMBOL(preempt_count_add);
2367 void __kprobes preempt_count_sub(int val)
2369 #ifdef CONFIG_DEBUG_PREEMPT
2373 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2376 * Is the spinlock portion underflowing?
2378 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2379 !(preempt_count() & PREEMPT_MASK)))
2383 if (preempt_count() == val)
2384 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2385 __preempt_count_sub(val);
2387 EXPORT_SYMBOL(preempt_count_sub);
2392 * Print scheduling while atomic bug:
2394 static noinline void __schedule_bug(struct task_struct *prev)
2396 if (oops_in_progress)
2399 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2400 prev->comm, prev->pid, preempt_count());
2402 debug_show_held_locks(prev);
2404 if (irqs_disabled())
2405 print_irqtrace_events(prev);
2407 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2411 * Various schedule()-time debugging checks and statistics:
2413 static inline void schedule_debug(struct task_struct *prev)
2416 * Test if we are atomic. Since do_exit() needs to call into
2417 * schedule() atomically, we ignore that path for now.
2418 * Otherwise, whine if we are scheduling when we should not be.
2420 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2421 __schedule_bug(prev);
2424 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2426 schedstat_inc(this_rq(), sched_count);
2429 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2431 if (prev->on_rq || rq->skip_clock_update < 0)
2432 update_rq_clock(rq);
2433 prev->sched_class->put_prev_task(rq, prev);
2437 * Pick up the highest-prio task:
2439 static inline struct task_struct *
2440 pick_next_task(struct rq *rq)
2442 const struct sched_class *class;
2443 struct task_struct *p;
2446 * Optimization: we know that if all tasks are in
2447 * the fair class we can call that function directly:
2449 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2450 p = fair_sched_class.pick_next_task(rq);
2455 for_each_class(class) {
2456 p = class->pick_next_task(rq);
2461 BUG(); /* the idle class will always have a runnable task */
2465 * __schedule() is the main scheduler function.
2467 * The main means of driving the scheduler and thus entering this function are:
2469 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2471 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2472 * paths. For example, see arch/x86/entry_64.S.
2474 * To drive preemption between tasks, the scheduler sets the flag in timer
2475 * interrupt handler scheduler_tick().
2477 * 3. Wakeups don't really cause entry into schedule(). They add a
2478 * task to the run-queue and that's it.
2480 * Now, if the new task added to the run-queue preempts the current
2481 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2482 * called on the nearest possible occasion:
2484 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2486 * - in syscall or exception context, at the next outmost
2487 * preempt_enable(). (this might be as soon as the wake_up()'s
2490 * - in IRQ context, return from interrupt-handler to
2491 * preemptible context
2493 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2496 * - cond_resched() call
2497 * - explicit schedule() call
2498 * - return from syscall or exception to user-space
2499 * - return from interrupt-handler to user-space
2501 static void __sched __schedule(void)
2503 struct task_struct *prev, *next;
2504 unsigned long *switch_count;
2510 cpu = smp_processor_id();
2512 rcu_note_context_switch(cpu);
2515 schedule_debug(prev);
2517 if (sched_feat(HRTICK))
2521 * Make sure that signal_pending_state()->signal_pending() below
2522 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2523 * done by the caller to avoid the race with signal_wake_up().
2525 smp_mb__before_spinlock();
2526 raw_spin_lock_irq(&rq->lock);
2528 switch_count = &prev->nivcsw;
2529 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2530 if (unlikely(signal_pending_state(prev->state, prev))) {
2531 prev->state = TASK_RUNNING;
2533 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2537 * If a worker went to sleep, notify and ask workqueue
2538 * whether it wants to wake up a task to maintain
2541 if (prev->flags & PF_WQ_WORKER) {
2542 struct task_struct *to_wakeup;
2544 to_wakeup = wq_worker_sleeping(prev, cpu);
2546 try_to_wake_up_local(to_wakeup);
2549 switch_count = &prev->nvcsw;
2552 pre_schedule(rq, prev);
2554 if (unlikely(!rq->nr_running))
2555 idle_balance(cpu, rq);
2557 put_prev_task(rq, prev);
2558 next = pick_next_task(rq);
2559 clear_tsk_need_resched(prev);
2560 clear_preempt_need_resched();
2561 rq->skip_clock_update = 0;
2563 if (likely(prev != next)) {
2568 context_switch(rq, prev, next); /* unlocks the rq */
2570 * The context switch have flipped the stack from under us
2571 * and restored the local variables which were saved when
2572 * this task called schedule() in the past. prev == current
2573 * is still correct, but it can be moved to another cpu/rq.
2575 cpu = smp_processor_id();
2578 raw_spin_unlock_irq(&rq->lock);
2582 sched_preempt_enable_no_resched();
2587 static inline void sched_submit_work(struct task_struct *tsk)
2589 if (!tsk->state || tsk_is_pi_blocked(tsk))
2592 * If we are going to sleep and we have plugged IO queued,
2593 * make sure to submit it to avoid deadlocks.
2595 if (blk_needs_flush_plug(tsk))
2596 blk_schedule_flush_plug(tsk);
2599 asmlinkage void __sched schedule(void)
2601 struct task_struct *tsk = current;
2603 sched_submit_work(tsk);
2606 EXPORT_SYMBOL(schedule);
2608 #ifdef CONFIG_CONTEXT_TRACKING
2609 asmlinkage void __sched schedule_user(void)
2612 * If we come here after a random call to set_need_resched(),
2613 * or we have been woken up remotely but the IPI has not yet arrived,
2614 * we haven't yet exited the RCU idle mode. Do it here manually until
2615 * we find a better solution.
2624 * schedule_preempt_disabled - called with preemption disabled
2626 * Returns with preemption disabled. Note: preempt_count must be 1
2628 void __sched schedule_preempt_disabled(void)
2630 sched_preempt_enable_no_resched();
2635 #ifdef CONFIG_PREEMPT
2637 * this is the entry point to schedule() from in-kernel preemption
2638 * off of preempt_enable. Kernel preemptions off return from interrupt
2639 * occur there and call schedule directly.
2641 asmlinkage void __sched notrace preempt_schedule(void)
2644 * If there is a non-zero preempt_count or interrupts are disabled,
2645 * we do not want to preempt the current task. Just return..
2647 if (likely(!preemptible()))
2651 __preempt_count_add(PREEMPT_ACTIVE);
2653 __preempt_count_sub(PREEMPT_ACTIVE);
2656 * Check again in case we missed a preemption opportunity
2657 * between schedule and now.
2660 } while (need_resched());
2662 EXPORT_SYMBOL(preempt_schedule);
2663 #endif /* CONFIG_PREEMPT */
2666 * this is the entry point to schedule() from kernel preemption
2667 * off of irq context.
2668 * Note, that this is called and return with irqs disabled. This will
2669 * protect us against recursive calling from irq.
2671 asmlinkage void __sched preempt_schedule_irq(void)
2673 enum ctx_state prev_state;
2675 /* Catch callers which need to be fixed */
2676 BUG_ON(preempt_count() || !irqs_disabled());
2678 prev_state = exception_enter();
2681 __preempt_count_add(PREEMPT_ACTIVE);
2684 local_irq_disable();
2685 __preempt_count_sub(PREEMPT_ACTIVE);
2688 * Check again in case we missed a preemption opportunity
2689 * between schedule and now.
2692 } while (need_resched());
2694 exception_exit(prev_state);
2697 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2700 return try_to_wake_up(curr->private, mode, wake_flags);
2702 EXPORT_SYMBOL(default_wake_function);
2705 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2707 unsigned long flags;
2710 init_waitqueue_entry(&wait, current);
2712 __set_current_state(state);
2714 spin_lock_irqsave(&q->lock, flags);
2715 __add_wait_queue(q, &wait);
2716 spin_unlock(&q->lock);
2717 timeout = schedule_timeout(timeout);
2718 spin_lock_irq(&q->lock);
2719 __remove_wait_queue(q, &wait);
2720 spin_unlock_irqrestore(&q->lock, flags);
2725 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2727 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2729 EXPORT_SYMBOL(interruptible_sleep_on);
2732 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2734 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2736 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2738 void __sched sleep_on(wait_queue_head_t *q)
2740 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2742 EXPORT_SYMBOL(sleep_on);
2744 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2746 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2748 EXPORT_SYMBOL(sleep_on_timeout);
2750 #ifdef CONFIG_RT_MUTEXES
2753 * rt_mutex_setprio - set the current priority of a task
2755 * @prio: prio value (kernel-internal form)
2757 * This function changes the 'effective' priority of a task. It does
2758 * not touch ->normal_prio like __setscheduler().
2760 * Used by the rt_mutex code to implement priority inheritance logic.
2762 void rt_mutex_setprio(struct task_struct *p, int prio)
2764 int oldprio, on_rq, running;
2766 const struct sched_class *prev_class;
2768 BUG_ON(prio < 0 || prio > MAX_PRIO);
2770 rq = __task_rq_lock(p);
2773 * Idle task boosting is a nono in general. There is one
2774 * exception, when PREEMPT_RT and NOHZ is active:
2776 * The idle task calls get_next_timer_interrupt() and holds
2777 * the timer wheel base->lock on the CPU and another CPU wants
2778 * to access the timer (probably to cancel it). We can safely
2779 * ignore the boosting request, as the idle CPU runs this code
2780 * with interrupts disabled and will complete the lock
2781 * protected section without being interrupted. So there is no
2782 * real need to boost.
2784 if (unlikely(p == rq->idle)) {
2785 WARN_ON(p != rq->curr);
2786 WARN_ON(p->pi_blocked_on);
2790 trace_sched_pi_setprio(p, prio);
2792 prev_class = p->sched_class;
2794 running = task_current(rq, p);
2796 dequeue_task(rq, p, 0);
2798 p->sched_class->put_prev_task(rq, p);
2801 p->sched_class = &rt_sched_class;
2803 p->sched_class = &fair_sched_class;
2808 p->sched_class->set_curr_task(rq);
2810 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
2812 check_class_changed(rq, p, prev_class, oldprio);
2814 __task_rq_unlock(rq);
2817 void set_user_nice(struct task_struct *p, long nice)
2819 int old_prio, delta, on_rq;
2820 unsigned long flags;
2823 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2826 * We have to be careful, if called from sys_setpriority(),
2827 * the task might be in the middle of scheduling on another CPU.
2829 rq = task_rq_lock(p, &flags);
2831 * The RT priorities are set via sched_setscheduler(), but we still
2832 * allow the 'normal' nice value to be set - but as expected
2833 * it wont have any effect on scheduling until the task is
2834 * SCHED_FIFO/SCHED_RR:
2836 if (task_has_rt_policy(p)) {
2837 p->static_prio = NICE_TO_PRIO(nice);
2842 dequeue_task(rq, p, 0);
2844 p->static_prio = NICE_TO_PRIO(nice);
2847 p->prio = effective_prio(p);
2848 delta = p->prio - old_prio;
2851 enqueue_task(rq, p, 0);
2853 * If the task increased its priority or is running and
2854 * lowered its priority, then reschedule its CPU:
2856 if (delta < 0 || (delta > 0 && task_running(rq, p)))
2857 resched_task(rq->curr);
2860 task_rq_unlock(rq, p, &flags);
2862 EXPORT_SYMBOL(set_user_nice);
2865 * can_nice - check if a task can reduce its nice value
2869 int can_nice(const struct task_struct *p, const int nice)
2871 /* convert nice value [19,-20] to rlimit style value [1,40] */
2872 int nice_rlim = 20 - nice;
2874 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
2875 capable(CAP_SYS_NICE));
2878 #ifdef __ARCH_WANT_SYS_NICE
2881 * sys_nice - change the priority of the current process.
2882 * @increment: priority increment
2884 * sys_setpriority is a more generic, but much slower function that
2885 * does similar things.
2887 SYSCALL_DEFINE1(nice, int, increment)
2892 * Setpriority might change our priority at the same moment.
2893 * We don't have to worry. Conceptually one call occurs first
2894 * and we have a single winner.
2896 if (increment < -40)
2901 nice = TASK_NICE(current) + increment;
2907 if (increment < 0 && !can_nice(current, nice))
2910 retval = security_task_setnice(current, nice);
2914 set_user_nice(current, nice);
2921 * task_prio - return the priority value of a given task.
2922 * @p: the task in question.
2924 * Return: The priority value as seen by users in /proc.
2925 * RT tasks are offset by -200. Normal tasks are centered
2926 * around 0, value goes from -16 to +15.
2928 int task_prio(const struct task_struct *p)
2930 return p->prio - MAX_RT_PRIO;
2934 * task_nice - return the nice value of a given task.
2935 * @p: the task in question.
2937 * Return: The nice value [ -20 ... 0 ... 19 ].
2939 int task_nice(const struct task_struct *p)
2941 return TASK_NICE(p);
2943 EXPORT_SYMBOL(task_nice);
2946 * idle_cpu - is a given cpu idle currently?
2947 * @cpu: the processor in question.
2949 * Return: 1 if the CPU is currently idle. 0 otherwise.
2951 int idle_cpu(int cpu)
2953 struct rq *rq = cpu_rq(cpu);
2955 if (rq->curr != rq->idle)
2962 if (!llist_empty(&rq->wake_list))
2970 * idle_task - return the idle task for a given cpu.
2971 * @cpu: the processor in question.
2973 * Return: The idle task for the cpu @cpu.
2975 struct task_struct *idle_task(int cpu)
2977 return cpu_rq(cpu)->idle;
2981 * find_process_by_pid - find a process with a matching PID value.
2982 * @pid: the pid in question.
2984 * The task of @pid, if found. %NULL otherwise.
2986 static struct task_struct *find_process_by_pid(pid_t pid)
2988 return pid ? find_task_by_vpid(pid) : current;
2991 /* Actually do priority change: must hold rq lock. */
2993 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
2996 p->rt_priority = prio;
2997 p->normal_prio = normal_prio(p);
2998 /* we are holding p->pi_lock already */
2999 p->prio = rt_mutex_getprio(p);
3000 if (rt_prio(p->prio))
3001 p->sched_class = &rt_sched_class;
3003 p->sched_class = &fair_sched_class;
3008 * check the target process has a UID that matches the current process's
3010 static bool check_same_owner(struct task_struct *p)
3012 const struct cred *cred = current_cred(), *pcred;
3016 pcred = __task_cred(p);
3017 match = (uid_eq(cred->euid, pcred->euid) ||
3018 uid_eq(cred->euid, pcred->uid));
3023 static int __sched_setscheduler(struct task_struct *p, int policy,
3024 const struct sched_param *param, bool user)
3026 int retval, oldprio, oldpolicy = -1, on_rq, running;
3027 unsigned long flags;
3028 const struct sched_class *prev_class;
3032 /* may grab non-irq protected spin_locks */
3033 BUG_ON(in_interrupt());
3035 /* double check policy once rq lock held */
3037 reset_on_fork = p->sched_reset_on_fork;
3038 policy = oldpolicy = p->policy;
3040 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3041 policy &= ~SCHED_RESET_ON_FORK;
3043 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3044 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3045 policy != SCHED_IDLE)
3050 * Valid priorities for SCHED_FIFO and SCHED_RR are
3051 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3052 * SCHED_BATCH and SCHED_IDLE is 0.
3054 if (param->sched_priority < 0 ||
3055 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3056 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3058 if (rt_policy(policy) != (param->sched_priority != 0))
3062 * Allow unprivileged RT tasks to decrease priority:
3064 if (user && !capable(CAP_SYS_NICE)) {
3065 if (rt_policy(policy)) {
3066 unsigned long rlim_rtprio =
3067 task_rlimit(p, RLIMIT_RTPRIO);
3069 /* can't set/change the rt policy */
3070 if (policy != p->policy && !rlim_rtprio)
3073 /* can't increase priority */
3074 if (param->sched_priority > p->rt_priority &&
3075 param->sched_priority > rlim_rtprio)
3080 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3081 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3083 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3084 if (!can_nice(p, TASK_NICE(p)))
3088 /* can't change other user's priorities */
3089 if (!check_same_owner(p))
3092 /* Normal users shall not reset the sched_reset_on_fork flag */
3093 if (p->sched_reset_on_fork && !reset_on_fork)
3098 retval = security_task_setscheduler(p);
3104 * make sure no PI-waiters arrive (or leave) while we are
3105 * changing the priority of the task:
3107 * To be able to change p->policy safely, the appropriate
3108 * runqueue lock must be held.
3110 rq = task_rq_lock(p, &flags);
3113 * Changing the policy of the stop threads its a very bad idea
3115 if (p == rq->stop) {
3116 task_rq_unlock(rq, p, &flags);
3121 * If not changing anything there's no need to proceed further:
3123 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3124 param->sched_priority == p->rt_priority))) {
3125 task_rq_unlock(rq, p, &flags);
3129 #ifdef CONFIG_RT_GROUP_SCHED
3132 * Do not allow realtime tasks into groups that have no runtime
3135 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3136 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3137 !task_group_is_autogroup(task_group(p))) {
3138 task_rq_unlock(rq, p, &flags);
3144 /* recheck policy now with rq lock held */
3145 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3146 policy = oldpolicy = -1;
3147 task_rq_unlock(rq, p, &flags);
3151 running = task_current(rq, p);
3153 dequeue_task(rq, p, 0);
3155 p->sched_class->put_prev_task(rq, p);
3157 p->sched_reset_on_fork = reset_on_fork;
3160 prev_class = p->sched_class;
3161 __setscheduler(rq, p, policy, param->sched_priority);
3164 p->sched_class->set_curr_task(rq);
3166 enqueue_task(rq, p, 0);
3168 check_class_changed(rq, p, prev_class, oldprio);
3169 task_rq_unlock(rq, p, &flags);
3171 rt_mutex_adjust_pi(p);
3177 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3178 * @p: the task in question.
3179 * @policy: new policy.
3180 * @param: structure containing the new RT priority.
3182 * Return: 0 on success. An error code otherwise.
3184 * NOTE that the task may be already dead.
3186 int sched_setscheduler(struct task_struct *p, int policy,
3187 const struct sched_param *param)
3189 return __sched_setscheduler(p, policy, param, true);
3191 EXPORT_SYMBOL_GPL(sched_setscheduler);
3194 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3195 * @p: the task in question.
3196 * @policy: new policy.
3197 * @param: structure containing the new RT priority.
3199 * Just like sched_setscheduler, only don't bother checking if the
3200 * current context has permission. For example, this is needed in
3201 * stop_machine(): we create temporary high priority worker threads,
3202 * but our caller might not have that capability.
3204 * Return: 0 on success. An error code otherwise.
3206 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3207 const struct sched_param *param)
3209 return __sched_setscheduler(p, policy, param, false);
3213 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3215 struct sched_param lparam;
3216 struct task_struct *p;
3219 if (!param || pid < 0)
3221 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3226 p = find_process_by_pid(pid);
3228 retval = sched_setscheduler(p, policy, &lparam);
3235 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3236 * @pid: the pid in question.
3237 * @policy: new policy.
3238 * @param: structure containing the new RT priority.
3240 * Return: 0 on success. An error code otherwise.
3242 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3243 struct sched_param __user *, param)
3245 /* negative values for policy are not valid */
3249 return do_sched_setscheduler(pid, policy, param);
3253 * sys_sched_setparam - set/change the RT priority of a thread
3254 * @pid: the pid in question.
3255 * @param: structure containing the new RT priority.
3257 * Return: 0 on success. An error code otherwise.
3259 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3261 return do_sched_setscheduler(pid, -1, param);
3265 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3266 * @pid: the pid in question.
3268 * Return: On success, the policy of the thread. Otherwise, a negative error
3271 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3273 struct task_struct *p;
3281 p = find_process_by_pid(pid);
3283 retval = security_task_getscheduler(p);
3286 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3293 * sys_sched_getparam - get the RT priority of a thread
3294 * @pid: the pid in question.
3295 * @param: structure containing the RT priority.
3297 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3300 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3302 struct sched_param lp;
3303 struct task_struct *p;
3306 if (!param || pid < 0)
3310 p = find_process_by_pid(pid);
3315 retval = security_task_getscheduler(p);
3319 lp.sched_priority = p->rt_priority;
3323 * This one might sleep, we cannot do it with a spinlock held ...
3325 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3334 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3336 cpumask_var_t cpus_allowed, new_mask;
3337 struct task_struct *p;
3342 p = find_process_by_pid(pid);
3348 /* Prevent p going away */
3352 if (p->flags & PF_NO_SETAFFINITY) {
3356 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3360 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3362 goto out_free_cpus_allowed;
3365 if (!check_same_owner(p)) {
3367 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3374 retval = security_task_setscheduler(p);
3378 cpuset_cpus_allowed(p, cpus_allowed);
3379 cpumask_and(new_mask, in_mask, cpus_allowed);
3381 retval = set_cpus_allowed_ptr(p, new_mask);
3384 cpuset_cpus_allowed(p, cpus_allowed);
3385 if (!cpumask_subset(new_mask, cpus_allowed)) {
3387 * We must have raced with a concurrent cpuset
3388 * update. Just reset the cpus_allowed to the
3389 * cpuset's cpus_allowed
3391 cpumask_copy(new_mask, cpus_allowed);
3396 free_cpumask_var(new_mask);
3397 out_free_cpus_allowed:
3398 free_cpumask_var(cpus_allowed);
3404 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3405 struct cpumask *new_mask)
3407 if (len < cpumask_size())
3408 cpumask_clear(new_mask);
3409 else if (len > cpumask_size())
3410 len = cpumask_size();
3412 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3416 * sys_sched_setaffinity - set the cpu affinity of a process
3417 * @pid: pid of the process
3418 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3419 * @user_mask_ptr: user-space pointer to the new cpu mask
3421 * Return: 0 on success. An error code otherwise.
3423 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3424 unsigned long __user *, user_mask_ptr)
3426 cpumask_var_t new_mask;
3429 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3432 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3434 retval = sched_setaffinity(pid, new_mask);
3435 free_cpumask_var(new_mask);
3439 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3441 struct task_struct *p;
3442 unsigned long flags;
3448 p = find_process_by_pid(pid);
3452 retval = security_task_getscheduler(p);
3456 raw_spin_lock_irqsave(&p->pi_lock, flags);
3457 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3458 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3467 * sys_sched_getaffinity - get the cpu affinity of a process
3468 * @pid: pid of the process
3469 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3470 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3472 * Return: 0 on success. An error code otherwise.
3474 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3475 unsigned long __user *, user_mask_ptr)
3480 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3482 if (len & (sizeof(unsigned long)-1))
3485 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3488 ret = sched_getaffinity(pid, mask);
3490 size_t retlen = min_t(size_t, len, cpumask_size());
3492 if (copy_to_user(user_mask_ptr, mask, retlen))
3497 free_cpumask_var(mask);
3503 * sys_sched_yield - yield the current processor to other threads.
3505 * This function yields the current CPU to other tasks. If there are no
3506 * other threads running on this CPU then this function will return.
3510 SYSCALL_DEFINE0(sched_yield)
3512 struct rq *rq = this_rq_lock();
3514 schedstat_inc(rq, yld_count);
3515 current->sched_class->yield_task(rq);
3518 * Since we are going to call schedule() anyway, there's
3519 * no need to preempt or enable interrupts:
3521 __release(rq->lock);
3522 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3523 do_raw_spin_unlock(&rq->lock);
3524 sched_preempt_enable_no_resched();
3531 static void __cond_resched(void)
3533 __preempt_count_add(PREEMPT_ACTIVE);
3535 __preempt_count_sub(PREEMPT_ACTIVE);
3538 int __sched _cond_resched(void)
3540 if (should_resched()) {
3546 EXPORT_SYMBOL(_cond_resched);
3549 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3550 * call schedule, and on return reacquire the lock.
3552 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3553 * operations here to prevent schedule() from being called twice (once via
3554 * spin_unlock(), once by hand).
3556 int __cond_resched_lock(spinlock_t *lock)
3558 int resched = should_resched();
3561 lockdep_assert_held(lock);
3563 if (spin_needbreak(lock) || resched) {
3574 EXPORT_SYMBOL(__cond_resched_lock);
3576 int __sched __cond_resched_softirq(void)
3578 BUG_ON(!in_softirq());
3580 if (should_resched()) {
3588 EXPORT_SYMBOL(__cond_resched_softirq);
3591 * yield - yield the current processor to other threads.
3593 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3595 * The scheduler is at all times free to pick the calling task as the most
3596 * eligible task to run, if removing the yield() call from your code breaks
3597 * it, its already broken.
3599 * Typical broken usage is:
3604 * where one assumes that yield() will let 'the other' process run that will
3605 * make event true. If the current task is a SCHED_FIFO task that will never
3606 * happen. Never use yield() as a progress guarantee!!
3608 * If you want to use yield() to wait for something, use wait_event().
3609 * If you want to use yield() to be 'nice' for others, use cond_resched().
3610 * If you still want to use yield(), do not!
3612 void __sched yield(void)
3614 set_current_state(TASK_RUNNING);
3617 EXPORT_SYMBOL(yield);
3620 * yield_to - yield the current processor to another thread in
3621 * your thread group, or accelerate that thread toward the
3622 * processor it's on.
3624 * @preempt: whether task preemption is allowed or not
3626 * It's the caller's job to ensure that the target task struct
3627 * can't go away on us before we can do any checks.
3630 * true (>0) if we indeed boosted the target task.
3631 * false (0) if we failed to boost the target.
3632 * -ESRCH if there's no task to yield to.
3634 bool __sched yield_to(struct task_struct *p, bool preempt)
3636 struct task_struct *curr = current;
3637 struct rq *rq, *p_rq;
3638 unsigned long flags;
3641 local_irq_save(flags);
3647 * If we're the only runnable task on the rq and target rq also
3648 * has only one task, there's absolutely no point in yielding.
3650 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3655 double_rq_lock(rq, p_rq);
3656 while (task_rq(p) != p_rq) {
3657 double_rq_unlock(rq, p_rq);
3661 if (!curr->sched_class->yield_to_task)
3664 if (curr->sched_class != p->sched_class)
3667 if (task_running(p_rq, p) || p->state)
3670 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3672 schedstat_inc(rq, yld_count);
3674 * Make p's CPU reschedule; pick_next_entity takes care of
3677 if (preempt && rq != p_rq)
3678 resched_task(p_rq->curr);
3682 double_rq_unlock(rq, p_rq);
3684 local_irq_restore(flags);
3691 EXPORT_SYMBOL_GPL(yield_to);
3694 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3695 * that process accounting knows that this is a task in IO wait state.
3697 void __sched io_schedule(void)
3699 struct rq *rq = raw_rq();
3701 delayacct_blkio_start();
3702 atomic_inc(&rq->nr_iowait);
3703 blk_flush_plug(current);
3704 current->in_iowait = 1;
3706 current->in_iowait = 0;
3707 atomic_dec(&rq->nr_iowait);
3708 delayacct_blkio_end();
3710 EXPORT_SYMBOL(io_schedule);
3712 long __sched io_schedule_timeout(long timeout)
3714 struct rq *rq = raw_rq();
3717 delayacct_blkio_start();
3718 atomic_inc(&rq->nr_iowait);
3719 blk_flush_plug(current);
3720 current->in_iowait = 1;
3721 ret = schedule_timeout(timeout);
3722 current->in_iowait = 0;
3723 atomic_dec(&rq->nr_iowait);
3724 delayacct_blkio_end();
3729 * sys_sched_get_priority_max - return maximum RT priority.
3730 * @policy: scheduling class.
3732 * Return: On success, this syscall returns the maximum
3733 * rt_priority that can be used by a given scheduling class.
3734 * On failure, a negative error code is returned.
3736 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3743 ret = MAX_USER_RT_PRIO-1;
3755 * sys_sched_get_priority_min - return minimum RT priority.
3756 * @policy: scheduling class.
3758 * Return: On success, this syscall returns the minimum
3759 * rt_priority that can be used by a given scheduling class.
3760 * On failure, a negative error code is returned.
3762 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
3780 * sys_sched_rr_get_interval - return the default timeslice of a process.
3781 * @pid: pid of the process.
3782 * @interval: userspace pointer to the timeslice value.
3784 * this syscall writes the default timeslice value of a given process
3785 * into the user-space timespec buffer. A value of '0' means infinity.
3787 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
3790 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
3791 struct timespec __user *, interval)
3793 struct task_struct *p;
3794 unsigned int time_slice;
3795 unsigned long flags;
3805 p = find_process_by_pid(pid);
3809 retval = security_task_getscheduler(p);
3813 rq = task_rq_lock(p, &flags);
3814 time_slice = p->sched_class->get_rr_interval(rq, p);
3815 task_rq_unlock(rq, p, &flags);
3818 jiffies_to_timespec(time_slice, &t);
3819 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
3827 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
3829 void sched_show_task(struct task_struct *p)
3831 unsigned long free = 0;
3835 state = p->state ? __ffs(p->state) + 1 : 0;
3836 printk(KERN_INFO "%-15.15s %c", p->comm,
3837 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
3838 #if BITS_PER_LONG == 32
3839 if (state == TASK_RUNNING)
3840 printk(KERN_CONT " running ");
3842 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
3844 if (state == TASK_RUNNING)
3845 printk(KERN_CONT " running task ");
3847 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
3849 #ifdef CONFIG_DEBUG_STACK_USAGE
3850 free = stack_not_used(p);
3853 ppid = task_pid_nr(rcu_dereference(p->real_parent));
3855 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
3856 task_pid_nr(p), ppid,
3857 (unsigned long)task_thread_info(p)->flags);
3859 print_worker_info(KERN_INFO, p);
3860 show_stack(p, NULL);
3863 void show_state_filter(unsigned long state_filter)
3865 struct task_struct *g, *p;
3867 #if BITS_PER_LONG == 32
3869 " task PC stack pid father\n");
3872 " task PC stack pid father\n");
3875 do_each_thread(g, p) {
3877 * reset the NMI-timeout, listing all files on a slow
3878 * console might take a lot of time:
3880 touch_nmi_watchdog();
3881 if (!state_filter || (p->state & state_filter))
3883 } while_each_thread(g, p);
3885 touch_all_softlockup_watchdogs();
3887 #ifdef CONFIG_SCHED_DEBUG
3888 sysrq_sched_debug_show();
3892 * Only show locks if all tasks are dumped:
3895 debug_show_all_locks();
3898 void init_idle_bootup_task(struct task_struct *idle)
3900 idle->sched_class = &idle_sched_class;
3904 * init_idle - set up an idle thread for a given CPU
3905 * @idle: task in question
3906 * @cpu: cpu the idle task belongs to
3908 * NOTE: this function does not set the idle thread's NEED_RESCHED
3909 * flag, to make booting more robust.
3911 void init_idle(struct task_struct *idle, int cpu)
3913 struct rq *rq = cpu_rq(cpu);
3914 unsigned long flags;
3916 raw_spin_lock_irqsave(&rq->lock, flags);
3918 __sched_fork(0, idle);
3919 idle->state = TASK_RUNNING;
3920 idle->se.exec_start = sched_clock();
3922 do_set_cpus_allowed(idle, cpumask_of(cpu));
3924 * We're having a chicken and egg problem, even though we are
3925 * holding rq->lock, the cpu isn't yet set to this cpu so the
3926 * lockdep check in task_group() will fail.
3928 * Similar case to sched_fork(). / Alternatively we could
3929 * use task_rq_lock() here and obtain the other rq->lock.
3934 __set_task_cpu(idle, cpu);
3937 rq->curr = rq->idle = idle;
3938 #if defined(CONFIG_SMP)
3941 raw_spin_unlock_irqrestore(&rq->lock, flags);
3943 /* Set the preempt count _outside_ the spinlocks! */
3944 init_idle_preempt_count(idle, cpu);
3947 * The idle tasks have their own, simple scheduling class:
3949 idle->sched_class = &idle_sched_class;
3950 ftrace_graph_init_idle_task(idle, cpu);
3951 vtime_init_idle(idle, cpu);
3952 #if defined(CONFIG_SMP)
3953 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
3958 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
3960 if (p->sched_class && p->sched_class->set_cpus_allowed)
3961 p->sched_class->set_cpus_allowed(p, new_mask);
3963 cpumask_copy(&p->cpus_allowed, new_mask);
3964 p->nr_cpus_allowed = cpumask_weight(new_mask);
3968 * This is how migration works:
3970 * 1) we invoke migration_cpu_stop() on the target CPU using
3972 * 2) stopper starts to run (implicitly forcing the migrated thread
3974 * 3) it checks whether the migrated task is still in the wrong runqueue.
3975 * 4) if it's in the wrong runqueue then the migration thread removes
3976 * it and puts it into the right queue.
3977 * 5) stopper completes and stop_one_cpu() returns and the migration
3982 * Change a given task's CPU affinity. Migrate the thread to a
3983 * proper CPU and schedule it away if the CPU it's executing on
3984 * is removed from the allowed bitmask.
3986 * NOTE: the caller must have a valid reference to the task, the
3987 * task must not exit() & deallocate itself prematurely. The
3988 * call is not atomic; no spinlocks may be held.
3990 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
3992 unsigned long flags;
3994 unsigned int dest_cpu;
3997 rq = task_rq_lock(p, &flags);
3999 if (cpumask_equal(&p->cpus_allowed, new_mask))
4002 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4007 do_set_cpus_allowed(p, new_mask);
4009 /* Can the task run on the task's current CPU? If so, we're done */
4010 if (cpumask_test_cpu(task_cpu(p), new_mask))
4013 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4015 struct migration_arg arg = { p, dest_cpu };
4016 /* Need help from migration thread: drop lock and wait. */
4017 task_rq_unlock(rq, p, &flags);
4018 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4019 tlb_migrate_finish(p->mm);
4023 task_rq_unlock(rq, p, &flags);
4027 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4030 * Move (not current) task off this cpu, onto dest cpu. We're doing
4031 * this because either it can't run here any more (set_cpus_allowed()
4032 * away from this CPU, or CPU going down), or because we're
4033 * attempting to rebalance this task on exec (sched_exec).
4035 * So we race with normal scheduler movements, but that's OK, as long
4036 * as the task is no longer on this CPU.
4038 * Returns non-zero if task was successfully migrated.
4040 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4042 struct rq *rq_dest, *rq_src;
4045 if (unlikely(!cpu_active(dest_cpu)))
4048 rq_src = cpu_rq(src_cpu);
4049 rq_dest = cpu_rq(dest_cpu);
4051 raw_spin_lock(&p->pi_lock);
4052 double_rq_lock(rq_src, rq_dest);
4053 /* Already moved. */
4054 if (task_cpu(p) != src_cpu)
4056 /* Affinity changed (again). */
4057 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4061 * If we're not on a rq, the next wake-up will ensure we're
4065 dequeue_task(rq_src, p, 0);
4066 set_task_cpu(p, dest_cpu);
4067 enqueue_task(rq_dest, p, 0);
4068 check_preempt_curr(rq_dest, p, 0);
4073 double_rq_unlock(rq_src, rq_dest);
4074 raw_spin_unlock(&p->pi_lock);
4078 #ifdef CONFIG_NUMA_BALANCING
4079 /* Migrate current task p to target_cpu */
4080 int migrate_task_to(struct task_struct *p, int target_cpu)
4082 struct migration_arg arg = { p, target_cpu };
4083 int curr_cpu = task_cpu(p);
4085 if (curr_cpu == target_cpu)
4088 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4091 /* TODO: This is not properly updating schedstats */
4093 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4097 * Requeue a task on a given node and accurately track the number of NUMA
4098 * tasks on the runqueues
4100 void sched_setnuma(struct task_struct *p, int nid)
4103 unsigned long flags;
4104 bool on_rq, running;
4106 rq = task_rq_lock(p, &flags);
4108 running = task_current(rq, p);
4111 dequeue_task(rq, p, 0);
4113 p->sched_class->put_prev_task(rq, p);
4115 p->numa_preferred_nid = nid;
4118 p->sched_class->set_curr_task(rq);
4120 enqueue_task(rq, p, 0);
4121 task_rq_unlock(rq, p, &flags);
4126 * migration_cpu_stop - this will be executed by a highprio stopper thread
4127 * and performs thread migration by bumping thread off CPU then
4128 * 'pushing' onto another runqueue.
4130 static int migration_cpu_stop(void *data)
4132 struct migration_arg *arg = data;
4135 * The original target cpu might have gone down and we might
4136 * be on another cpu but it doesn't matter.
4138 local_irq_disable();
4139 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4144 #ifdef CONFIG_HOTPLUG_CPU
4147 * Ensures that the idle task is using init_mm right before its cpu goes
4150 void idle_task_exit(void)
4152 struct mm_struct *mm = current->active_mm;
4154 BUG_ON(cpu_online(smp_processor_id()));
4157 switch_mm(mm, &init_mm, current);
4162 * Since this CPU is going 'away' for a while, fold any nr_active delta
4163 * we might have. Assumes we're called after migrate_tasks() so that the
4164 * nr_active count is stable.
4166 * Also see the comment "Global load-average calculations".
4168 static void calc_load_migrate(struct rq *rq)
4170 long delta = calc_load_fold_active(rq);
4172 atomic_long_add(delta, &calc_load_tasks);
4176 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4177 * try_to_wake_up()->select_task_rq().
4179 * Called with rq->lock held even though we'er in stop_machine() and
4180 * there's no concurrency possible, we hold the required locks anyway
4181 * because of lock validation efforts.
4183 static void migrate_tasks(unsigned int dead_cpu)
4185 struct rq *rq = cpu_rq(dead_cpu);
4186 struct task_struct *next, *stop = rq->stop;
4190 * Fudge the rq selection such that the below task selection loop
4191 * doesn't get stuck on the currently eligible stop task.
4193 * We're currently inside stop_machine() and the rq is either stuck
4194 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4195 * either way we should never end up calling schedule() until we're
4201 * put_prev_task() and pick_next_task() sched
4202 * class method both need to have an up-to-date
4203 * value of rq->clock[_task]
4205 update_rq_clock(rq);
4209 * There's this thread running, bail when that's the only
4212 if (rq->nr_running == 1)
4215 next = pick_next_task(rq);
4217 next->sched_class->put_prev_task(rq, next);
4219 /* Find suitable destination for @next, with force if needed. */
4220 dest_cpu = select_fallback_rq(dead_cpu, next);
4221 raw_spin_unlock(&rq->lock);
4223 __migrate_task(next, dead_cpu, dest_cpu);
4225 raw_spin_lock(&rq->lock);
4231 #endif /* CONFIG_HOTPLUG_CPU */
4233 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4235 static struct ctl_table sd_ctl_dir[] = {
4237 .procname = "sched_domain",
4243 static struct ctl_table sd_ctl_root[] = {
4245 .procname = "kernel",
4247 .child = sd_ctl_dir,
4252 static struct ctl_table *sd_alloc_ctl_entry(int n)
4254 struct ctl_table *entry =
4255 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4260 static void sd_free_ctl_entry(struct ctl_table **tablep)
4262 struct ctl_table *entry;
4265 * In the intermediate directories, both the child directory and
4266 * procname are dynamically allocated and could fail but the mode
4267 * will always be set. In the lowest directory the names are
4268 * static strings and all have proc handlers.
4270 for (entry = *tablep; entry->mode; entry++) {
4272 sd_free_ctl_entry(&entry->child);
4273 if (entry->proc_handler == NULL)
4274 kfree(entry->procname);
4281 static int min_load_idx = 0;
4282 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4285 set_table_entry(struct ctl_table *entry,
4286 const char *procname, void *data, int maxlen,
4287 umode_t mode, proc_handler *proc_handler,
4290 entry->procname = procname;
4292 entry->maxlen = maxlen;
4294 entry->proc_handler = proc_handler;
4297 entry->extra1 = &min_load_idx;
4298 entry->extra2 = &max_load_idx;
4302 static struct ctl_table *
4303 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4305 struct ctl_table *table = sd_alloc_ctl_entry(13);
4310 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4311 sizeof(long), 0644, proc_doulongvec_minmax, false);
4312 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4313 sizeof(long), 0644, proc_doulongvec_minmax, false);
4314 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4315 sizeof(int), 0644, proc_dointvec_minmax, true);
4316 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4317 sizeof(int), 0644, proc_dointvec_minmax, true);
4318 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4319 sizeof(int), 0644, proc_dointvec_minmax, true);
4320 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4321 sizeof(int), 0644, proc_dointvec_minmax, true);
4322 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4323 sizeof(int), 0644, proc_dointvec_minmax, true);
4324 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4325 sizeof(int), 0644, proc_dointvec_minmax, false);
4326 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4327 sizeof(int), 0644, proc_dointvec_minmax, false);
4328 set_table_entry(&table[9], "cache_nice_tries",
4329 &sd->cache_nice_tries,
4330 sizeof(int), 0644, proc_dointvec_minmax, false);
4331 set_table_entry(&table[10], "flags", &sd->flags,
4332 sizeof(int), 0644, proc_dointvec_minmax, false);
4333 set_table_entry(&table[11], "name", sd->name,
4334 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4335 /* &table[12] is terminator */
4340 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4342 struct ctl_table *entry, *table;
4343 struct sched_domain *sd;
4344 int domain_num = 0, i;
4347 for_each_domain(cpu, sd)
4349 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4354 for_each_domain(cpu, sd) {
4355 snprintf(buf, 32, "domain%d", i);
4356 entry->procname = kstrdup(buf, GFP_KERNEL);
4358 entry->child = sd_alloc_ctl_domain_table(sd);
4365 static struct ctl_table_header *sd_sysctl_header;
4366 static void register_sched_domain_sysctl(void)
4368 int i, cpu_num = num_possible_cpus();
4369 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4372 WARN_ON(sd_ctl_dir[0].child);
4373 sd_ctl_dir[0].child = entry;
4378 for_each_possible_cpu(i) {
4379 snprintf(buf, 32, "cpu%d", i);
4380 entry->procname = kstrdup(buf, GFP_KERNEL);
4382 entry->child = sd_alloc_ctl_cpu_table(i);
4386 WARN_ON(sd_sysctl_header);
4387 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4390 /* may be called multiple times per register */
4391 static void unregister_sched_domain_sysctl(void)
4393 if (sd_sysctl_header)
4394 unregister_sysctl_table(sd_sysctl_header);
4395 sd_sysctl_header = NULL;
4396 if (sd_ctl_dir[0].child)
4397 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4400 static void register_sched_domain_sysctl(void)
4403 static void unregister_sched_domain_sysctl(void)
4408 static void set_rq_online(struct rq *rq)
4411 const struct sched_class *class;
4413 cpumask_set_cpu(rq->cpu, rq->rd->online);
4416 for_each_class(class) {
4417 if (class->rq_online)
4418 class->rq_online(rq);
4423 static void set_rq_offline(struct rq *rq)
4426 const struct sched_class *class;
4428 for_each_class(class) {
4429 if (class->rq_offline)
4430 class->rq_offline(rq);
4433 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4439 * migration_call - callback that gets triggered when a CPU is added.
4440 * Here we can start up the necessary migration thread for the new CPU.
4443 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4445 int cpu = (long)hcpu;
4446 unsigned long flags;
4447 struct rq *rq = cpu_rq(cpu);
4449 switch (action & ~CPU_TASKS_FROZEN) {
4451 case CPU_UP_PREPARE:
4452 rq->calc_load_update = calc_load_update;
4456 /* Update our root-domain */
4457 raw_spin_lock_irqsave(&rq->lock, flags);
4459 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4463 raw_spin_unlock_irqrestore(&rq->lock, flags);
4466 #ifdef CONFIG_HOTPLUG_CPU
4468 sched_ttwu_pending();
4469 /* Update our root-domain */
4470 raw_spin_lock_irqsave(&rq->lock, flags);
4472 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4476 BUG_ON(rq->nr_running != 1); /* the migration thread */
4477 raw_spin_unlock_irqrestore(&rq->lock, flags);
4481 calc_load_migrate(rq);
4486 update_max_interval();
4492 * Register at high priority so that task migration (migrate_all_tasks)
4493 * happens before everything else. This has to be lower priority than
4494 * the notifier in the perf_event subsystem, though.
4496 static struct notifier_block migration_notifier = {
4497 .notifier_call = migration_call,
4498 .priority = CPU_PRI_MIGRATION,
4501 static int sched_cpu_active(struct notifier_block *nfb,
4502 unsigned long action, void *hcpu)
4504 switch (action & ~CPU_TASKS_FROZEN) {
4506 case CPU_DOWN_FAILED:
4507 set_cpu_active((long)hcpu, true);
4514 static int sched_cpu_inactive(struct notifier_block *nfb,
4515 unsigned long action, void *hcpu)
4517 switch (action & ~CPU_TASKS_FROZEN) {
4518 case CPU_DOWN_PREPARE:
4519 set_cpu_active((long)hcpu, false);
4526 static int __init migration_init(void)
4528 void *cpu = (void *)(long)smp_processor_id();
4531 /* Initialize migration for the boot CPU */
4532 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4533 BUG_ON(err == NOTIFY_BAD);
4534 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4535 register_cpu_notifier(&migration_notifier);
4537 /* Register cpu active notifiers */
4538 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4539 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4543 early_initcall(migration_init);
4548 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4550 #ifdef CONFIG_SCHED_DEBUG
4552 static __read_mostly int sched_debug_enabled;
4554 static int __init sched_debug_setup(char *str)
4556 sched_debug_enabled = 1;
4560 early_param("sched_debug", sched_debug_setup);
4562 static inline bool sched_debug(void)
4564 return sched_debug_enabled;
4567 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4568 struct cpumask *groupmask)
4570 struct sched_group *group = sd->groups;
4573 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4574 cpumask_clear(groupmask);
4576 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4578 if (!(sd->flags & SD_LOAD_BALANCE)) {
4579 printk("does not load-balance\n");
4581 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4586 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4588 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4589 printk(KERN_ERR "ERROR: domain->span does not contain "
4592 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4593 printk(KERN_ERR "ERROR: domain->groups does not contain"
4597 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4601 printk(KERN_ERR "ERROR: group is NULL\n");
4606 * Even though we initialize ->power to something semi-sane,
4607 * we leave power_orig unset. This allows us to detect if
4608 * domain iteration is still funny without causing /0 traps.
4610 if (!group->sgp->power_orig) {
4611 printk(KERN_CONT "\n");
4612 printk(KERN_ERR "ERROR: domain->cpu_power not "
4617 if (!cpumask_weight(sched_group_cpus(group))) {
4618 printk(KERN_CONT "\n");
4619 printk(KERN_ERR "ERROR: empty group\n");
4623 if (!(sd->flags & SD_OVERLAP) &&
4624 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4625 printk(KERN_CONT "\n");
4626 printk(KERN_ERR "ERROR: repeated CPUs\n");
4630 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4632 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4634 printk(KERN_CONT " %s", str);
4635 if (group->sgp->power != SCHED_POWER_SCALE) {
4636 printk(KERN_CONT " (cpu_power = %d)",
4640 group = group->next;
4641 } while (group != sd->groups);
4642 printk(KERN_CONT "\n");
4644 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4645 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4648 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4649 printk(KERN_ERR "ERROR: parent span is not a superset "
4650 "of domain->span\n");
4654 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4658 if (!sched_debug_enabled)
4662 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4666 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4669 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4677 #else /* !CONFIG_SCHED_DEBUG */
4678 # define sched_domain_debug(sd, cpu) do { } while (0)
4679 static inline bool sched_debug(void)
4683 #endif /* CONFIG_SCHED_DEBUG */
4685 static int sd_degenerate(struct sched_domain *sd)
4687 if (cpumask_weight(sched_domain_span(sd)) == 1)
4690 /* Following flags need at least 2 groups */
4691 if (sd->flags & (SD_LOAD_BALANCE |
4692 SD_BALANCE_NEWIDLE |
4696 SD_SHARE_PKG_RESOURCES)) {
4697 if (sd->groups != sd->groups->next)
4701 /* Following flags don't use groups */
4702 if (sd->flags & (SD_WAKE_AFFINE))
4709 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4711 unsigned long cflags = sd->flags, pflags = parent->flags;
4713 if (sd_degenerate(parent))
4716 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4719 /* Flags needing groups don't count if only 1 group in parent */
4720 if (parent->groups == parent->groups->next) {
4721 pflags &= ~(SD_LOAD_BALANCE |
4722 SD_BALANCE_NEWIDLE |
4726 SD_SHARE_PKG_RESOURCES |
4728 if (nr_node_ids == 1)
4729 pflags &= ~SD_SERIALIZE;
4731 if (~cflags & pflags)
4737 static void free_rootdomain(struct rcu_head *rcu)
4739 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4741 cpupri_cleanup(&rd->cpupri);
4742 free_cpumask_var(rd->rto_mask);
4743 free_cpumask_var(rd->online);
4744 free_cpumask_var(rd->span);
4748 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4750 struct root_domain *old_rd = NULL;
4751 unsigned long flags;
4753 raw_spin_lock_irqsave(&rq->lock, flags);
4758 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4761 cpumask_clear_cpu(rq->cpu, old_rd->span);
4764 * If we dont want to free the old_rd yet then
4765 * set old_rd to NULL to skip the freeing later
4768 if (!atomic_dec_and_test(&old_rd->refcount))
4772 atomic_inc(&rd->refcount);
4775 cpumask_set_cpu(rq->cpu, rd->span);
4776 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4779 raw_spin_unlock_irqrestore(&rq->lock, flags);
4782 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4785 static int init_rootdomain(struct root_domain *rd)
4787 memset(rd, 0, sizeof(*rd));
4789 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4791 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4793 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4796 if (cpupri_init(&rd->cpupri) != 0)
4801 free_cpumask_var(rd->rto_mask);
4803 free_cpumask_var(rd->online);
4805 free_cpumask_var(rd->span);
4811 * By default the system creates a single root-domain with all cpus as
4812 * members (mimicking the global state we have today).
4814 struct root_domain def_root_domain;
4816 static void init_defrootdomain(void)
4818 init_rootdomain(&def_root_domain);
4820 atomic_set(&def_root_domain.refcount, 1);
4823 static struct root_domain *alloc_rootdomain(void)
4825 struct root_domain *rd;
4827 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
4831 if (init_rootdomain(rd) != 0) {
4839 static void free_sched_groups(struct sched_group *sg, int free_sgp)
4841 struct sched_group *tmp, *first;
4850 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
4855 } while (sg != first);
4858 static void free_sched_domain(struct rcu_head *rcu)
4860 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
4863 * If its an overlapping domain it has private groups, iterate and
4866 if (sd->flags & SD_OVERLAP) {
4867 free_sched_groups(sd->groups, 1);
4868 } else if (atomic_dec_and_test(&sd->groups->ref)) {
4869 kfree(sd->groups->sgp);
4875 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
4877 call_rcu(&sd->rcu, free_sched_domain);
4880 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
4882 for (; sd; sd = sd->parent)
4883 destroy_sched_domain(sd, cpu);
4887 * Keep a special pointer to the highest sched_domain that has
4888 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
4889 * allows us to avoid some pointer chasing select_idle_sibling().
4891 * Also keep a unique ID per domain (we use the first cpu number in
4892 * the cpumask of the domain), this allows us to quickly tell if
4893 * two cpus are in the same cache domain, see cpus_share_cache().
4895 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
4896 DEFINE_PER_CPU(int, sd_llc_size);
4897 DEFINE_PER_CPU(int, sd_llc_id);
4898 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
4899 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
4900 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
4902 static void update_top_cache_domain(int cpu)
4904 struct sched_domain *sd;
4905 struct sched_domain *busy_sd = NULL;
4909 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
4911 id = cpumask_first(sched_domain_span(sd));
4912 size = cpumask_weight(sched_domain_span(sd));
4913 busy_sd = sd->parent; /* sd_busy */
4915 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
4917 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
4918 per_cpu(sd_llc_size, cpu) = size;
4919 per_cpu(sd_llc_id, cpu) = id;
4921 sd = lowest_flag_domain(cpu, SD_NUMA);
4922 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
4924 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
4925 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
4929 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
4930 * hold the hotplug lock.
4933 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
4935 struct rq *rq = cpu_rq(cpu);
4936 struct sched_domain *tmp;
4938 /* Remove the sched domains which do not contribute to scheduling. */
4939 for (tmp = sd; tmp; ) {
4940 struct sched_domain *parent = tmp->parent;
4944 if (sd_parent_degenerate(tmp, parent)) {
4945 tmp->parent = parent->parent;
4947 parent->parent->child = tmp;
4949 * Transfer SD_PREFER_SIBLING down in case of a
4950 * degenerate parent; the spans match for this
4951 * so the property transfers.
4953 if (parent->flags & SD_PREFER_SIBLING)
4954 tmp->flags |= SD_PREFER_SIBLING;
4955 destroy_sched_domain(parent, cpu);
4960 if (sd && sd_degenerate(sd)) {
4963 destroy_sched_domain(tmp, cpu);
4968 sched_domain_debug(sd, cpu);
4970 rq_attach_root(rq, rd);
4972 rcu_assign_pointer(rq->sd, sd);
4973 destroy_sched_domains(tmp, cpu);
4975 update_top_cache_domain(cpu);
4978 /* cpus with isolated domains */
4979 static cpumask_var_t cpu_isolated_map;
4981 /* Setup the mask of cpus configured for isolated domains */
4982 static int __init isolated_cpu_setup(char *str)
4984 alloc_bootmem_cpumask_var(&cpu_isolated_map);
4985 cpulist_parse(str, cpu_isolated_map);
4989 __setup("isolcpus=", isolated_cpu_setup);
4991 static const struct cpumask *cpu_cpu_mask(int cpu)
4993 return cpumask_of_node(cpu_to_node(cpu));
4997 struct sched_domain **__percpu sd;
4998 struct sched_group **__percpu sg;
4999 struct sched_group_power **__percpu sgp;
5003 struct sched_domain ** __percpu sd;
5004 struct root_domain *rd;
5014 struct sched_domain_topology_level;
5016 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5017 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5019 #define SDTL_OVERLAP 0x01
5021 struct sched_domain_topology_level {
5022 sched_domain_init_f init;
5023 sched_domain_mask_f mask;
5026 struct sd_data data;
5030 * Build an iteration mask that can exclude certain CPUs from the upwards
5033 * Asymmetric node setups can result in situations where the domain tree is of
5034 * unequal depth, make sure to skip domains that already cover the entire
5037 * In that case build_sched_domains() will have terminated the iteration early
5038 * and our sibling sd spans will be empty. Domains should always include the
5039 * cpu they're built on, so check that.
5042 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5044 const struct cpumask *span = sched_domain_span(sd);
5045 struct sd_data *sdd = sd->private;
5046 struct sched_domain *sibling;
5049 for_each_cpu(i, span) {
5050 sibling = *per_cpu_ptr(sdd->sd, i);
5051 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5054 cpumask_set_cpu(i, sched_group_mask(sg));
5059 * Return the canonical balance cpu for this group, this is the first cpu
5060 * of this group that's also in the iteration mask.
5062 int group_balance_cpu(struct sched_group *sg)
5064 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5068 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5070 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5071 const struct cpumask *span = sched_domain_span(sd);
5072 struct cpumask *covered = sched_domains_tmpmask;
5073 struct sd_data *sdd = sd->private;
5074 struct sched_domain *child;
5077 cpumask_clear(covered);
5079 for_each_cpu(i, span) {
5080 struct cpumask *sg_span;
5082 if (cpumask_test_cpu(i, covered))
5085 child = *per_cpu_ptr(sdd->sd, i);
5087 /* See the comment near build_group_mask(). */
5088 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5091 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5092 GFP_KERNEL, cpu_to_node(cpu));
5097 sg_span = sched_group_cpus(sg);
5099 child = child->child;
5100 cpumask_copy(sg_span, sched_domain_span(child));
5102 cpumask_set_cpu(i, sg_span);
5104 cpumask_or(covered, covered, sg_span);
5106 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5107 if (atomic_inc_return(&sg->sgp->ref) == 1)
5108 build_group_mask(sd, sg);
5111 * Initialize sgp->power such that even if we mess up the
5112 * domains and no possible iteration will get us here, we won't
5115 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5116 sg->sgp->power_orig = sg->sgp->power;
5119 * Make sure the first group of this domain contains the
5120 * canonical balance cpu. Otherwise the sched_domain iteration
5121 * breaks. See update_sg_lb_stats().
5123 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5124 group_balance_cpu(sg) == cpu)
5134 sd->groups = groups;
5139 free_sched_groups(first, 0);
5144 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5146 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5147 struct sched_domain *child = sd->child;
5150 cpu = cpumask_first(sched_domain_span(child));
5153 *sg = *per_cpu_ptr(sdd->sg, cpu);
5154 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5155 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5162 * build_sched_groups will build a circular linked list of the groups
5163 * covered by the given span, and will set each group's ->cpumask correctly,
5164 * and ->cpu_power to 0.
5166 * Assumes the sched_domain tree is fully constructed
5169 build_sched_groups(struct sched_domain *sd, int cpu)
5171 struct sched_group *first = NULL, *last = NULL;
5172 struct sd_data *sdd = sd->private;
5173 const struct cpumask *span = sched_domain_span(sd);
5174 struct cpumask *covered;
5177 get_group(cpu, sdd, &sd->groups);
5178 atomic_inc(&sd->groups->ref);
5180 if (cpu != cpumask_first(span))
5183 lockdep_assert_held(&sched_domains_mutex);
5184 covered = sched_domains_tmpmask;
5186 cpumask_clear(covered);
5188 for_each_cpu(i, span) {
5189 struct sched_group *sg;
5192 if (cpumask_test_cpu(i, covered))
5195 group = get_group(i, sdd, &sg);
5196 cpumask_clear(sched_group_cpus(sg));
5198 cpumask_setall(sched_group_mask(sg));
5200 for_each_cpu(j, span) {
5201 if (get_group(j, sdd, NULL) != group)
5204 cpumask_set_cpu(j, covered);
5205 cpumask_set_cpu(j, sched_group_cpus(sg));
5220 * Initialize sched groups cpu_power.
5222 * cpu_power indicates the capacity of sched group, which is used while
5223 * distributing the load between different sched groups in a sched domain.
5224 * Typically cpu_power for all the groups in a sched domain will be same unless
5225 * there are asymmetries in the topology. If there are asymmetries, group
5226 * having more cpu_power will pickup more load compared to the group having
5229 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5231 struct sched_group *sg = sd->groups;
5236 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5238 } while (sg != sd->groups);
5240 if (cpu != group_balance_cpu(sg))
5243 update_group_power(sd, cpu);
5244 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5247 int __weak arch_sd_sibling_asym_packing(void)
5249 return 0*SD_ASYM_PACKING;
5253 * Initializers for schedule domains
5254 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5257 #ifdef CONFIG_SCHED_DEBUG
5258 # define SD_INIT_NAME(sd, type) sd->name = #type
5260 # define SD_INIT_NAME(sd, type) do { } while (0)
5263 #define SD_INIT_FUNC(type) \
5264 static noinline struct sched_domain * \
5265 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5267 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5268 *sd = SD_##type##_INIT; \
5269 SD_INIT_NAME(sd, type); \
5270 sd->private = &tl->data; \
5275 #ifdef CONFIG_SCHED_SMT
5276 SD_INIT_FUNC(SIBLING)
5278 #ifdef CONFIG_SCHED_MC
5281 #ifdef CONFIG_SCHED_BOOK
5285 static int default_relax_domain_level = -1;
5286 int sched_domain_level_max;
5288 static int __init setup_relax_domain_level(char *str)
5290 if (kstrtoint(str, 0, &default_relax_domain_level))
5291 pr_warn("Unable to set relax_domain_level\n");
5295 __setup("relax_domain_level=", setup_relax_domain_level);
5297 static void set_domain_attribute(struct sched_domain *sd,
5298 struct sched_domain_attr *attr)
5302 if (!attr || attr->relax_domain_level < 0) {
5303 if (default_relax_domain_level < 0)
5306 request = default_relax_domain_level;
5308 request = attr->relax_domain_level;
5309 if (request < sd->level) {
5310 /* turn off idle balance on this domain */
5311 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5313 /* turn on idle balance on this domain */
5314 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5318 static void __sdt_free(const struct cpumask *cpu_map);
5319 static int __sdt_alloc(const struct cpumask *cpu_map);
5321 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5322 const struct cpumask *cpu_map)
5326 if (!atomic_read(&d->rd->refcount))
5327 free_rootdomain(&d->rd->rcu); /* fall through */
5329 free_percpu(d->sd); /* fall through */
5331 __sdt_free(cpu_map); /* fall through */
5337 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5338 const struct cpumask *cpu_map)
5340 memset(d, 0, sizeof(*d));
5342 if (__sdt_alloc(cpu_map))
5343 return sa_sd_storage;
5344 d->sd = alloc_percpu(struct sched_domain *);
5346 return sa_sd_storage;
5347 d->rd = alloc_rootdomain();
5350 return sa_rootdomain;
5354 * NULL the sd_data elements we've used to build the sched_domain and
5355 * sched_group structure so that the subsequent __free_domain_allocs()
5356 * will not free the data we're using.
5358 static void claim_allocations(int cpu, struct sched_domain *sd)
5360 struct sd_data *sdd = sd->private;
5362 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5363 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5365 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5366 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5368 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5369 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5372 #ifdef CONFIG_SCHED_SMT
5373 static const struct cpumask *cpu_smt_mask(int cpu)
5375 return topology_thread_cpumask(cpu);
5380 * Topology list, bottom-up.
5382 static struct sched_domain_topology_level default_topology[] = {
5383 #ifdef CONFIG_SCHED_SMT
5384 { sd_init_SIBLING, cpu_smt_mask, },
5386 #ifdef CONFIG_SCHED_MC
5387 { sd_init_MC, cpu_coregroup_mask, },
5389 #ifdef CONFIG_SCHED_BOOK
5390 { sd_init_BOOK, cpu_book_mask, },
5392 { sd_init_CPU, cpu_cpu_mask, },
5396 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5398 #define for_each_sd_topology(tl) \
5399 for (tl = sched_domain_topology; tl->init; tl++)
5403 static int sched_domains_numa_levels;
5404 static int *sched_domains_numa_distance;
5405 static struct cpumask ***sched_domains_numa_masks;
5406 static int sched_domains_curr_level;
5408 static inline int sd_local_flags(int level)
5410 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5413 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5416 static struct sched_domain *
5417 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5419 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5420 int level = tl->numa_level;
5421 int sd_weight = cpumask_weight(
5422 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5424 *sd = (struct sched_domain){
5425 .min_interval = sd_weight,
5426 .max_interval = 2*sd_weight,
5428 .imbalance_pct = 125,
5429 .cache_nice_tries = 2,
5436 .flags = 1*SD_LOAD_BALANCE
5437 | 1*SD_BALANCE_NEWIDLE
5442 | 0*SD_SHARE_CPUPOWER
5443 | 0*SD_SHARE_PKG_RESOURCES
5445 | 0*SD_PREFER_SIBLING
5447 | sd_local_flags(level)
5449 .last_balance = jiffies,
5450 .balance_interval = sd_weight,
5452 SD_INIT_NAME(sd, NUMA);
5453 sd->private = &tl->data;
5456 * Ugly hack to pass state to sd_numa_mask()...
5458 sched_domains_curr_level = tl->numa_level;
5463 static const struct cpumask *sd_numa_mask(int cpu)
5465 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5468 static void sched_numa_warn(const char *str)
5470 static int done = false;
5478 printk(KERN_WARNING "ERROR: %s\n\n", str);
5480 for (i = 0; i < nr_node_ids; i++) {
5481 printk(KERN_WARNING " ");
5482 for (j = 0; j < nr_node_ids; j++)
5483 printk(KERN_CONT "%02d ", node_distance(i,j));
5484 printk(KERN_CONT "\n");
5486 printk(KERN_WARNING "\n");
5489 static bool find_numa_distance(int distance)
5493 if (distance == node_distance(0, 0))
5496 for (i = 0; i < sched_domains_numa_levels; i++) {
5497 if (sched_domains_numa_distance[i] == distance)
5504 static void sched_init_numa(void)
5506 int next_distance, curr_distance = node_distance(0, 0);
5507 struct sched_domain_topology_level *tl;
5511 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5512 if (!sched_domains_numa_distance)
5516 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5517 * unique distances in the node_distance() table.
5519 * Assumes node_distance(0,j) includes all distances in
5520 * node_distance(i,j) in order to avoid cubic time.
5522 next_distance = curr_distance;
5523 for (i = 0; i < nr_node_ids; i++) {
5524 for (j = 0; j < nr_node_ids; j++) {
5525 for (k = 0; k < nr_node_ids; k++) {
5526 int distance = node_distance(i, k);
5528 if (distance > curr_distance &&
5529 (distance < next_distance ||
5530 next_distance == curr_distance))
5531 next_distance = distance;
5534 * While not a strong assumption it would be nice to know
5535 * about cases where if node A is connected to B, B is not
5536 * equally connected to A.
5538 if (sched_debug() && node_distance(k, i) != distance)
5539 sched_numa_warn("Node-distance not symmetric");
5541 if (sched_debug() && i && !find_numa_distance(distance))
5542 sched_numa_warn("Node-0 not representative");
5544 if (next_distance != curr_distance) {
5545 sched_domains_numa_distance[level++] = next_distance;
5546 sched_domains_numa_levels = level;
5547 curr_distance = next_distance;
5552 * In case of sched_debug() we verify the above assumption.
5558 * 'level' contains the number of unique distances, excluding the
5559 * identity distance node_distance(i,i).
5561 * The sched_domains_numa_distance[] array includes the actual distance
5566 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5567 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5568 * the array will contain less then 'level' members. This could be
5569 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5570 * in other functions.
5572 * We reset it to 'level' at the end of this function.
5574 sched_domains_numa_levels = 0;
5576 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5577 if (!sched_domains_numa_masks)
5581 * Now for each level, construct a mask per node which contains all
5582 * cpus of nodes that are that many hops away from us.
5584 for (i = 0; i < level; i++) {
5585 sched_domains_numa_masks[i] =
5586 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5587 if (!sched_domains_numa_masks[i])
5590 for (j = 0; j < nr_node_ids; j++) {
5591 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5595 sched_domains_numa_masks[i][j] = mask;
5597 for (k = 0; k < nr_node_ids; k++) {
5598 if (node_distance(j, k) > sched_domains_numa_distance[i])
5601 cpumask_or(mask, mask, cpumask_of_node(k));
5606 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5607 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5612 * Copy the default topology bits..
5614 for (i = 0; default_topology[i].init; i++)
5615 tl[i] = default_topology[i];
5618 * .. and append 'j' levels of NUMA goodness.
5620 for (j = 0; j < level; i++, j++) {
5621 tl[i] = (struct sched_domain_topology_level){
5622 .init = sd_numa_init,
5623 .mask = sd_numa_mask,
5624 .flags = SDTL_OVERLAP,
5629 sched_domain_topology = tl;
5631 sched_domains_numa_levels = level;
5634 static void sched_domains_numa_masks_set(int cpu)
5637 int node = cpu_to_node(cpu);
5639 for (i = 0; i < sched_domains_numa_levels; i++) {
5640 for (j = 0; j < nr_node_ids; j++) {
5641 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5642 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5647 static void sched_domains_numa_masks_clear(int cpu)
5650 for (i = 0; i < sched_domains_numa_levels; i++) {
5651 for (j = 0; j < nr_node_ids; j++)
5652 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5657 * Update sched_domains_numa_masks[level][node] array when new cpus
5660 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5661 unsigned long action,
5664 int cpu = (long)hcpu;
5666 switch (action & ~CPU_TASKS_FROZEN) {
5668 sched_domains_numa_masks_set(cpu);
5672 sched_domains_numa_masks_clear(cpu);
5682 static inline void sched_init_numa(void)
5686 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5687 unsigned long action,
5692 #endif /* CONFIG_NUMA */
5694 static int __sdt_alloc(const struct cpumask *cpu_map)
5696 struct sched_domain_topology_level *tl;
5699 for_each_sd_topology(tl) {
5700 struct sd_data *sdd = &tl->data;
5702 sdd->sd = alloc_percpu(struct sched_domain *);
5706 sdd->sg = alloc_percpu(struct sched_group *);
5710 sdd->sgp = alloc_percpu(struct sched_group_power *);
5714 for_each_cpu(j, cpu_map) {
5715 struct sched_domain *sd;
5716 struct sched_group *sg;
5717 struct sched_group_power *sgp;
5719 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5720 GFP_KERNEL, cpu_to_node(j));
5724 *per_cpu_ptr(sdd->sd, j) = sd;
5726 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5727 GFP_KERNEL, cpu_to_node(j));
5733 *per_cpu_ptr(sdd->sg, j) = sg;
5735 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5736 GFP_KERNEL, cpu_to_node(j));
5740 *per_cpu_ptr(sdd->sgp, j) = sgp;
5747 static void __sdt_free(const struct cpumask *cpu_map)
5749 struct sched_domain_topology_level *tl;
5752 for_each_sd_topology(tl) {
5753 struct sd_data *sdd = &tl->data;
5755 for_each_cpu(j, cpu_map) {
5756 struct sched_domain *sd;
5759 sd = *per_cpu_ptr(sdd->sd, j);
5760 if (sd && (sd->flags & SD_OVERLAP))
5761 free_sched_groups(sd->groups, 0);
5762 kfree(*per_cpu_ptr(sdd->sd, j));
5766 kfree(*per_cpu_ptr(sdd->sg, j));
5768 kfree(*per_cpu_ptr(sdd->sgp, j));
5770 free_percpu(sdd->sd);
5772 free_percpu(sdd->sg);
5774 free_percpu(sdd->sgp);
5779 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5780 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5781 struct sched_domain *child, int cpu)
5783 struct sched_domain *sd = tl->init(tl, cpu);
5787 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5789 sd->level = child->level + 1;
5790 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5794 set_domain_attribute(sd, attr);
5800 * Build sched domains for a given set of cpus and attach the sched domains
5801 * to the individual cpus
5803 static int build_sched_domains(const struct cpumask *cpu_map,
5804 struct sched_domain_attr *attr)
5806 enum s_alloc alloc_state;
5807 struct sched_domain *sd;
5809 int i, ret = -ENOMEM;
5811 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5812 if (alloc_state != sa_rootdomain)
5815 /* Set up domains for cpus specified by the cpu_map. */
5816 for_each_cpu(i, cpu_map) {
5817 struct sched_domain_topology_level *tl;
5820 for_each_sd_topology(tl) {
5821 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
5822 if (tl == sched_domain_topology)
5823 *per_cpu_ptr(d.sd, i) = sd;
5824 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
5825 sd->flags |= SD_OVERLAP;
5826 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
5831 /* Build the groups for the domains */
5832 for_each_cpu(i, cpu_map) {
5833 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5834 sd->span_weight = cpumask_weight(sched_domain_span(sd));
5835 if (sd->flags & SD_OVERLAP) {
5836 if (build_overlap_sched_groups(sd, i))
5839 if (build_sched_groups(sd, i))
5845 /* Calculate CPU power for physical packages and nodes */
5846 for (i = nr_cpumask_bits-1; i >= 0; i--) {
5847 if (!cpumask_test_cpu(i, cpu_map))
5850 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5851 claim_allocations(i, sd);
5852 init_sched_groups_power(i, sd);
5856 /* Attach the domains */
5858 for_each_cpu(i, cpu_map) {
5859 sd = *per_cpu_ptr(d.sd, i);
5860 cpu_attach_domain(sd, d.rd, i);
5866 __free_domain_allocs(&d, alloc_state, cpu_map);
5870 static cpumask_var_t *doms_cur; /* current sched domains */
5871 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
5872 static struct sched_domain_attr *dattr_cur;
5873 /* attribues of custom domains in 'doms_cur' */
5876 * Special case: If a kmalloc of a doms_cur partition (array of
5877 * cpumask) fails, then fallback to a single sched domain,
5878 * as determined by the single cpumask fallback_doms.
5880 static cpumask_var_t fallback_doms;
5883 * arch_update_cpu_topology lets virtualized architectures update the
5884 * cpu core maps. It is supposed to return 1 if the topology changed
5885 * or 0 if it stayed the same.
5887 int __attribute__((weak)) arch_update_cpu_topology(void)
5892 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
5895 cpumask_var_t *doms;
5897 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
5900 for (i = 0; i < ndoms; i++) {
5901 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
5902 free_sched_domains(doms, i);
5909 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
5912 for (i = 0; i < ndoms; i++)
5913 free_cpumask_var(doms[i]);
5918 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
5919 * For now this just excludes isolated cpus, but could be used to
5920 * exclude other special cases in the future.
5922 static int init_sched_domains(const struct cpumask *cpu_map)
5926 arch_update_cpu_topology();
5928 doms_cur = alloc_sched_domains(ndoms_cur);
5930 doms_cur = &fallback_doms;
5931 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
5932 err = build_sched_domains(doms_cur[0], NULL);
5933 register_sched_domain_sysctl();
5939 * Detach sched domains from a group of cpus specified in cpu_map
5940 * These cpus will now be attached to the NULL domain
5942 static void detach_destroy_domains(const struct cpumask *cpu_map)
5947 for_each_cpu(i, cpu_map)
5948 cpu_attach_domain(NULL, &def_root_domain, i);
5952 /* handle null as "default" */
5953 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
5954 struct sched_domain_attr *new, int idx_new)
5956 struct sched_domain_attr tmp;
5963 return !memcmp(cur ? (cur + idx_cur) : &tmp,
5964 new ? (new + idx_new) : &tmp,
5965 sizeof(struct sched_domain_attr));
5969 * Partition sched domains as specified by the 'ndoms_new'
5970 * cpumasks in the array doms_new[] of cpumasks. This compares
5971 * doms_new[] to the current sched domain partitioning, doms_cur[].
5972 * It destroys each deleted domain and builds each new domain.
5974 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
5975 * The masks don't intersect (don't overlap.) We should setup one
5976 * sched domain for each mask. CPUs not in any of the cpumasks will
5977 * not be load balanced. If the same cpumask appears both in the
5978 * current 'doms_cur' domains and in the new 'doms_new', we can leave
5981 * The passed in 'doms_new' should be allocated using
5982 * alloc_sched_domains. This routine takes ownership of it and will
5983 * free_sched_domains it when done with it. If the caller failed the
5984 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
5985 * and partition_sched_domains() will fallback to the single partition
5986 * 'fallback_doms', it also forces the domains to be rebuilt.
5988 * If doms_new == NULL it will be replaced with cpu_online_mask.
5989 * ndoms_new == 0 is a special case for destroying existing domains,
5990 * and it will not create the default domain.
5992 * Call with hotplug lock held
5994 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
5995 struct sched_domain_attr *dattr_new)
6000 mutex_lock(&sched_domains_mutex);
6002 /* always unregister in case we don't destroy any domains */
6003 unregister_sched_domain_sysctl();
6005 /* Let architecture update cpu core mappings. */
6006 new_topology = arch_update_cpu_topology();
6008 n = doms_new ? ndoms_new : 0;
6010 /* Destroy deleted domains */
6011 for (i = 0; i < ndoms_cur; i++) {
6012 for (j = 0; j < n && !new_topology; j++) {
6013 if (cpumask_equal(doms_cur[i], doms_new[j])
6014 && dattrs_equal(dattr_cur, i, dattr_new, j))
6017 /* no match - a current sched domain not in new doms_new[] */
6018 detach_destroy_domains(doms_cur[i]);
6024 if (doms_new == NULL) {
6026 doms_new = &fallback_doms;
6027 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6028 WARN_ON_ONCE(dattr_new);
6031 /* Build new domains */
6032 for (i = 0; i < ndoms_new; i++) {
6033 for (j = 0; j < n && !new_topology; j++) {
6034 if (cpumask_equal(doms_new[i], doms_cur[j])
6035 && dattrs_equal(dattr_new, i, dattr_cur, j))
6038 /* no match - add a new doms_new */
6039 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6044 /* Remember the new sched domains */
6045 if (doms_cur != &fallback_doms)
6046 free_sched_domains(doms_cur, ndoms_cur);
6047 kfree(dattr_cur); /* kfree(NULL) is safe */
6048 doms_cur = doms_new;
6049 dattr_cur = dattr_new;
6050 ndoms_cur = ndoms_new;
6052 register_sched_domain_sysctl();
6054 mutex_unlock(&sched_domains_mutex);
6057 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6060 * Update cpusets according to cpu_active mask. If cpusets are
6061 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6062 * around partition_sched_domains().
6064 * If we come here as part of a suspend/resume, don't touch cpusets because we
6065 * want to restore it back to its original state upon resume anyway.
6067 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6071 case CPU_ONLINE_FROZEN:
6072 case CPU_DOWN_FAILED_FROZEN:
6075 * num_cpus_frozen tracks how many CPUs are involved in suspend
6076 * resume sequence. As long as this is not the last online
6077 * operation in the resume sequence, just build a single sched
6078 * domain, ignoring cpusets.
6081 if (likely(num_cpus_frozen)) {
6082 partition_sched_domains(1, NULL, NULL);
6087 * This is the last CPU online operation. So fall through and
6088 * restore the original sched domains by considering the
6089 * cpuset configurations.
6093 case CPU_DOWN_FAILED:
6094 cpuset_update_active_cpus(true);
6102 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6106 case CPU_DOWN_PREPARE:
6107 cpuset_update_active_cpus(false);
6109 case CPU_DOWN_PREPARE_FROZEN:
6111 partition_sched_domains(1, NULL, NULL);
6119 void __init sched_init_smp(void)
6121 cpumask_var_t non_isolated_cpus;
6123 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6124 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6129 * There's no userspace yet to cause hotplug operations; hence all the
6130 * cpu masks are stable and all blatant races in the below code cannot
6133 mutex_lock(&sched_domains_mutex);
6134 init_sched_domains(cpu_active_mask);
6135 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6136 if (cpumask_empty(non_isolated_cpus))
6137 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6138 mutex_unlock(&sched_domains_mutex);
6140 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6141 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6142 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6146 /* Move init over to a non-isolated CPU */
6147 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6149 sched_init_granularity();
6150 free_cpumask_var(non_isolated_cpus);
6152 init_sched_rt_class();
6155 void __init sched_init_smp(void)
6157 sched_init_granularity();
6159 #endif /* CONFIG_SMP */
6161 const_debug unsigned int sysctl_timer_migration = 1;
6163 int in_sched_functions(unsigned long addr)
6165 return in_lock_functions(addr) ||
6166 (addr >= (unsigned long)__sched_text_start
6167 && addr < (unsigned long)__sched_text_end);
6170 #ifdef CONFIG_CGROUP_SCHED
6172 * Default task group.
6173 * Every task in system belongs to this group at bootup.
6175 struct task_group root_task_group;
6176 LIST_HEAD(task_groups);
6179 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6181 void __init sched_init(void)
6184 unsigned long alloc_size = 0, ptr;
6186 #ifdef CONFIG_FAIR_GROUP_SCHED
6187 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6189 #ifdef CONFIG_RT_GROUP_SCHED
6190 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6192 #ifdef CONFIG_CPUMASK_OFFSTACK
6193 alloc_size += num_possible_cpus() * cpumask_size();
6196 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6198 #ifdef CONFIG_FAIR_GROUP_SCHED
6199 root_task_group.se = (struct sched_entity **)ptr;
6200 ptr += nr_cpu_ids * sizeof(void **);
6202 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6203 ptr += nr_cpu_ids * sizeof(void **);
6205 #endif /* CONFIG_FAIR_GROUP_SCHED */
6206 #ifdef CONFIG_RT_GROUP_SCHED
6207 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6208 ptr += nr_cpu_ids * sizeof(void **);
6210 root_task_group.rt_rq = (struct rt_rq **)ptr;
6211 ptr += nr_cpu_ids * sizeof(void **);
6213 #endif /* CONFIG_RT_GROUP_SCHED */
6214 #ifdef CONFIG_CPUMASK_OFFSTACK
6215 for_each_possible_cpu(i) {
6216 per_cpu(load_balance_mask, i) = (void *)ptr;
6217 ptr += cpumask_size();
6219 #endif /* CONFIG_CPUMASK_OFFSTACK */
6223 init_defrootdomain();
6226 init_rt_bandwidth(&def_rt_bandwidth,
6227 global_rt_period(), global_rt_runtime());
6229 #ifdef CONFIG_RT_GROUP_SCHED
6230 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6231 global_rt_period(), global_rt_runtime());
6232 #endif /* CONFIG_RT_GROUP_SCHED */
6234 #ifdef CONFIG_CGROUP_SCHED
6235 list_add(&root_task_group.list, &task_groups);
6236 INIT_LIST_HEAD(&root_task_group.children);
6237 INIT_LIST_HEAD(&root_task_group.siblings);
6238 autogroup_init(&init_task);
6240 #endif /* CONFIG_CGROUP_SCHED */
6242 for_each_possible_cpu(i) {
6246 raw_spin_lock_init(&rq->lock);
6248 rq->calc_load_active = 0;
6249 rq->calc_load_update = jiffies + LOAD_FREQ;
6250 init_cfs_rq(&rq->cfs);
6251 init_rt_rq(&rq->rt, rq);
6252 #ifdef CONFIG_FAIR_GROUP_SCHED
6253 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6254 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6256 * How much cpu bandwidth does root_task_group get?
6258 * In case of task-groups formed thr' the cgroup filesystem, it
6259 * gets 100% of the cpu resources in the system. This overall
6260 * system cpu resource is divided among the tasks of
6261 * root_task_group and its child task-groups in a fair manner,
6262 * based on each entity's (task or task-group's) weight
6263 * (se->load.weight).
6265 * In other words, if root_task_group has 10 tasks of weight
6266 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6267 * then A0's share of the cpu resource is:
6269 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6271 * We achieve this by letting root_task_group's tasks sit
6272 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6274 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6275 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6276 #endif /* CONFIG_FAIR_GROUP_SCHED */
6278 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6279 #ifdef CONFIG_RT_GROUP_SCHED
6280 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6281 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6284 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6285 rq->cpu_load[j] = 0;
6287 rq->last_load_update_tick = jiffies;
6292 rq->cpu_power = SCHED_POWER_SCALE;
6293 rq->post_schedule = 0;
6294 rq->active_balance = 0;
6295 rq->next_balance = jiffies;
6300 rq->avg_idle = 2*sysctl_sched_migration_cost;
6301 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6303 INIT_LIST_HEAD(&rq->cfs_tasks);
6305 rq_attach_root(rq, &def_root_domain);
6306 #ifdef CONFIG_NO_HZ_COMMON
6309 #ifdef CONFIG_NO_HZ_FULL
6310 rq->last_sched_tick = 0;
6314 atomic_set(&rq->nr_iowait, 0);
6317 set_load_weight(&init_task);
6319 #ifdef CONFIG_PREEMPT_NOTIFIERS
6320 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6323 #ifdef CONFIG_RT_MUTEXES
6324 plist_head_init(&init_task.pi_waiters);
6328 * The boot idle thread does lazy MMU switching as well:
6330 atomic_inc(&init_mm.mm_count);
6331 enter_lazy_tlb(&init_mm, current);
6334 * Make us the idle thread. Technically, schedule() should not be
6335 * called from this thread, however somewhere below it might be,
6336 * but because we are the idle thread, we just pick up running again
6337 * when this runqueue becomes "idle".
6339 init_idle(current, smp_processor_id());
6341 calc_load_update = jiffies + LOAD_FREQ;
6344 * During early bootup we pretend to be a normal task:
6346 current->sched_class = &fair_sched_class;
6349 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6350 /* May be allocated at isolcpus cmdline parse time */
6351 if (cpu_isolated_map == NULL)
6352 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6353 idle_thread_set_boot_cpu();
6355 init_sched_fair_class();
6357 scheduler_running = 1;
6360 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6361 static inline int preempt_count_equals(int preempt_offset)
6363 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6365 return (nested == preempt_offset);
6368 void __might_sleep(const char *file, int line, int preempt_offset)
6370 static unsigned long prev_jiffy; /* ratelimiting */
6372 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6373 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6374 system_state != SYSTEM_RUNNING || oops_in_progress)
6376 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6378 prev_jiffy = jiffies;
6381 "BUG: sleeping function called from invalid context at %s:%d\n",
6384 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6385 in_atomic(), irqs_disabled(),
6386 current->pid, current->comm);
6388 debug_show_held_locks(current);
6389 if (irqs_disabled())
6390 print_irqtrace_events(current);
6393 EXPORT_SYMBOL(__might_sleep);
6396 #ifdef CONFIG_MAGIC_SYSRQ
6397 static void normalize_task(struct rq *rq, struct task_struct *p)
6399 const struct sched_class *prev_class = p->sched_class;
6400 int old_prio = p->prio;
6405 dequeue_task(rq, p, 0);
6406 __setscheduler(rq, p, SCHED_NORMAL, 0);
6408 enqueue_task(rq, p, 0);
6409 resched_task(rq->curr);
6412 check_class_changed(rq, p, prev_class, old_prio);
6415 void normalize_rt_tasks(void)
6417 struct task_struct *g, *p;
6418 unsigned long flags;
6421 read_lock_irqsave(&tasklist_lock, flags);
6422 do_each_thread(g, p) {
6424 * Only normalize user tasks:
6429 p->se.exec_start = 0;
6430 #ifdef CONFIG_SCHEDSTATS
6431 p->se.statistics.wait_start = 0;
6432 p->se.statistics.sleep_start = 0;
6433 p->se.statistics.block_start = 0;
6438 * Renice negative nice level userspace
6441 if (TASK_NICE(p) < 0 && p->mm)
6442 set_user_nice(p, 0);
6446 raw_spin_lock(&p->pi_lock);
6447 rq = __task_rq_lock(p);
6449 normalize_task(rq, p);
6451 __task_rq_unlock(rq);
6452 raw_spin_unlock(&p->pi_lock);
6453 } while_each_thread(g, p);
6455 read_unlock_irqrestore(&tasklist_lock, flags);
6458 #endif /* CONFIG_MAGIC_SYSRQ */
6460 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6462 * These functions are only useful for the IA64 MCA handling, or kdb.
6464 * They can only be called when the whole system has been
6465 * stopped - every CPU needs to be quiescent, and no scheduling
6466 * activity can take place. Using them for anything else would
6467 * be a serious bug, and as a result, they aren't even visible
6468 * under any other configuration.
6472 * curr_task - return the current task for a given cpu.
6473 * @cpu: the processor in question.
6475 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6477 * Return: The current task for @cpu.
6479 struct task_struct *curr_task(int cpu)
6481 return cpu_curr(cpu);
6484 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6488 * set_curr_task - set the current task for a given cpu.
6489 * @cpu: the processor in question.
6490 * @p: the task pointer to set.
6492 * Description: This function must only be used when non-maskable interrupts
6493 * are serviced on a separate stack. It allows the architecture to switch the
6494 * notion of the current task on a cpu in a non-blocking manner. This function
6495 * must be called with all CPU's synchronized, and interrupts disabled, the
6496 * and caller must save the original value of the current task (see
6497 * curr_task() above) and restore that value before reenabling interrupts and
6498 * re-starting the system.
6500 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6502 void set_curr_task(int cpu, struct task_struct *p)
6509 #ifdef CONFIG_CGROUP_SCHED
6510 /* task_group_lock serializes the addition/removal of task groups */
6511 static DEFINE_SPINLOCK(task_group_lock);
6513 static void free_sched_group(struct task_group *tg)
6515 free_fair_sched_group(tg);
6516 free_rt_sched_group(tg);
6521 /* allocate runqueue etc for a new task group */
6522 struct task_group *sched_create_group(struct task_group *parent)
6524 struct task_group *tg;
6526 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6528 return ERR_PTR(-ENOMEM);
6530 if (!alloc_fair_sched_group(tg, parent))
6533 if (!alloc_rt_sched_group(tg, parent))
6539 free_sched_group(tg);
6540 return ERR_PTR(-ENOMEM);
6543 void sched_online_group(struct task_group *tg, struct task_group *parent)
6545 unsigned long flags;
6547 spin_lock_irqsave(&task_group_lock, flags);
6548 list_add_rcu(&tg->list, &task_groups);
6550 WARN_ON(!parent); /* root should already exist */
6552 tg->parent = parent;
6553 INIT_LIST_HEAD(&tg->children);
6554 list_add_rcu(&tg->siblings, &parent->children);
6555 spin_unlock_irqrestore(&task_group_lock, flags);
6558 /* rcu callback to free various structures associated with a task group */
6559 static void free_sched_group_rcu(struct rcu_head *rhp)
6561 /* now it should be safe to free those cfs_rqs */
6562 free_sched_group(container_of(rhp, struct task_group, rcu));
6565 /* Destroy runqueue etc associated with a task group */
6566 void sched_destroy_group(struct task_group *tg)
6568 /* wait for possible concurrent references to cfs_rqs complete */
6569 call_rcu(&tg->rcu, free_sched_group_rcu);
6572 void sched_offline_group(struct task_group *tg)
6574 unsigned long flags;
6577 /* end participation in shares distribution */
6578 for_each_possible_cpu(i)
6579 unregister_fair_sched_group(tg, i);
6581 spin_lock_irqsave(&task_group_lock, flags);
6582 list_del_rcu(&tg->list);
6583 list_del_rcu(&tg->siblings);
6584 spin_unlock_irqrestore(&task_group_lock, flags);
6587 /* change task's runqueue when it moves between groups.
6588 * The caller of this function should have put the task in its new group
6589 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6590 * reflect its new group.
6592 void sched_move_task(struct task_struct *tsk)
6594 struct task_group *tg;
6596 unsigned long flags;
6599 rq = task_rq_lock(tsk, &flags);
6601 running = task_current(rq, tsk);
6605 dequeue_task(rq, tsk, 0);
6606 if (unlikely(running))
6607 tsk->sched_class->put_prev_task(rq, tsk);
6609 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
6610 lockdep_is_held(&tsk->sighand->siglock)),
6611 struct task_group, css);
6612 tg = autogroup_task_group(tsk, tg);
6613 tsk->sched_task_group = tg;
6615 #ifdef CONFIG_FAIR_GROUP_SCHED
6616 if (tsk->sched_class->task_move_group)
6617 tsk->sched_class->task_move_group(tsk, on_rq);
6620 set_task_rq(tsk, task_cpu(tsk));
6622 if (unlikely(running))
6623 tsk->sched_class->set_curr_task(rq);
6625 enqueue_task(rq, tsk, 0);
6627 task_rq_unlock(rq, tsk, &flags);
6629 #endif /* CONFIG_CGROUP_SCHED */
6631 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6632 static unsigned long to_ratio(u64 period, u64 runtime)
6634 if (runtime == RUNTIME_INF)
6637 return div64_u64(runtime << 20, period);
6641 #ifdef CONFIG_RT_GROUP_SCHED
6643 * Ensure that the real time constraints are schedulable.
6645 static DEFINE_MUTEX(rt_constraints_mutex);
6647 /* Must be called with tasklist_lock held */
6648 static inline int tg_has_rt_tasks(struct task_group *tg)
6650 struct task_struct *g, *p;
6652 do_each_thread(g, p) {
6653 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6655 } while_each_thread(g, p);
6660 struct rt_schedulable_data {
6661 struct task_group *tg;
6666 static int tg_rt_schedulable(struct task_group *tg, void *data)
6668 struct rt_schedulable_data *d = data;
6669 struct task_group *child;
6670 unsigned long total, sum = 0;
6671 u64 period, runtime;
6673 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6674 runtime = tg->rt_bandwidth.rt_runtime;
6677 period = d->rt_period;
6678 runtime = d->rt_runtime;
6682 * Cannot have more runtime than the period.
6684 if (runtime > period && runtime != RUNTIME_INF)
6688 * Ensure we don't starve existing RT tasks.
6690 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6693 total = to_ratio(period, runtime);
6696 * Nobody can have more than the global setting allows.
6698 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6702 * The sum of our children's runtime should not exceed our own.
6704 list_for_each_entry_rcu(child, &tg->children, siblings) {
6705 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6706 runtime = child->rt_bandwidth.rt_runtime;
6708 if (child == d->tg) {
6709 period = d->rt_period;
6710 runtime = d->rt_runtime;
6713 sum += to_ratio(period, runtime);
6722 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6726 struct rt_schedulable_data data = {
6728 .rt_period = period,
6729 .rt_runtime = runtime,
6733 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6739 static int tg_set_rt_bandwidth(struct task_group *tg,
6740 u64 rt_period, u64 rt_runtime)
6744 mutex_lock(&rt_constraints_mutex);
6745 read_lock(&tasklist_lock);
6746 err = __rt_schedulable(tg, rt_period, rt_runtime);
6750 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6751 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6752 tg->rt_bandwidth.rt_runtime = rt_runtime;
6754 for_each_possible_cpu(i) {
6755 struct rt_rq *rt_rq = tg->rt_rq[i];
6757 raw_spin_lock(&rt_rq->rt_runtime_lock);
6758 rt_rq->rt_runtime = rt_runtime;
6759 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6761 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6763 read_unlock(&tasklist_lock);
6764 mutex_unlock(&rt_constraints_mutex);
6769 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6771 u64 rt_runtime, rt_period;
6773 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6774 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6775 if (rt_runtime_us < 0)
6776 rt_runtime = RUNTIME_INF;
6778 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6781 static long sched_group_rt_runtime(struct task_group *tg)
6785 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6788 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6789 do_div(rt_runtime_us, NSEC_PER_USEC);
6790 return rt_runtime_us;
6793 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6795 u64 rt_runtime, rt_period;
6797 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6798 rt_runtime = tg->rt_bandwidth.rt_runtime;
6803 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6806 static long sched_group_rt_period(struct task_group *tg)
6810 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6811 do_div(rt_period_us, NSEC_PER_USEC);
6812 return rt_period_us;
6815 static int sched_rt_global_constraints(void)
6817 u64 runtime, period;
6820 if (sysctl_sched_rt_period <= 0)
6823 runtime = global_rt_runtime();
6824 period = global_rt_period();
6827 * Sanity check on the sysctl variables.
6829 if (runtime > period && runtime != RUNTIME_INF)
6832 mutex_lock(&rt_constraints_mutex);
6833 read_lock(&tasklist_lock);
6834 ret = __rt_schedulable(NULL, 0, 0);
6835 read_unlock(&tasklist_lock);
6836 mutex_unlock(&rt_constraints_mutex);
6841 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6843 /* Don't accept realtime tasks when there is no way for them to run */
6844 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6850 #else /* !CONFIG_RT_GROUP_SCHED */
6851 static int sched_rt_global_constraints(void)
6853 unsigned long flags;
6856 if (sysctl_sched_rt_period <= 0)
6860 * There's always some RT tasks in the root group
6861 * -- migration, kstopmachine etc..
6863 if (sysctl_sched_rt_runtime == 0)
6866 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
6867 for_each_possible_cpu(i) {
6868 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
6870 raw_spin_lock(&rt_rq->rt_runtime_lock);
6871 rt_rq->rt_runtime = global_rt_runtime();
6872 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6874 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
6878 #endif /* CONFIG_RT_GROUP_SCHED */
6880 int sched_rr_handler(struct ctl_table *table, int write,
6881 void __user *buffer, size_t *lenp,
6885 static DEFINE_MUTEX(mutex);
6888 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6889 /* make sure that internally we keep jiffies */
6890 /* also, writing zero resets timeslice to default */
6891 if (!ret && write) {
6892 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
6893 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
6895 mutex_unlock(&mutex);
6899 int sched_rt_handler(struct ctl_table *table, int write,
6900 void __user *buffer, size_t *lenp,
6904 int old_period, old_runtime;
6905 static DEFINE_MUTEX(mutex);
6908 old_period = sysctl_sched_rt_period;
6909 old_runtime = sysctl_sched_rt_runtime;
6911 ret = proc_dointvec(table, write, buffer, lenp, ppos);
6913 if (!ret && write) {
6914 ret = sched_rt_global_constraints();
6916 sysctl_sched_rt_period = old_period;
6917 sysctl_sched_rt_runtime = old_runtime;
6919 def_rt_bandwidth.rt_runtime = global_rt_runtime();
6920 def_rt_bandwidth.rt_period =
6921 ns_to_ktime(global_rt_period());
6924 mutex_unlock(&mutex);
6929 #ifdef CONFIG_CGROUP_SCHED
6931 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
6933 return css ? container_of(css, struct task_group, css) : NULL;
6936 static struct cgroup_subsys_state *
6937 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
6939 struct task_group *parent = css_tg(parent_css);
6940 struct task_group *tg;
6943 /* This is early initialization for the top cgroup */
6944 return &root_task_group.css;
6947 tg = sched_create_group(parent);
6949 return ERR_PTR(-ENOMEM);
6954 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
6956 struct task_group *tg = css_tg(css);
6957 struct task_group *parent = css_tg(css_parent(css));
6960 sched_online_group(tg, parent);
6964 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
6966 struct task_group *tg = css_tg(css);
6968 sched_destroy_group(tg);
6971 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
6973 struct task_group *tg = css_tg(css);
6975 sched_offline_group(tg);
6978 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
6979 struct cgroup_taskset *tset)
6981 struct task_struct *task;
6983 cgroup_taskset_for_each(task, css, tset) {
6984 #ifdef CONFIG_RT_GROUP_SCHED
6985 if (!sched_rt_can_attach(css_tg(css), task))
6988 /* We don't support RT-tasks being in separate groups */
6989 if (task->sched_class != &fair_sched_class)
6996 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
6997 struct cgroup_taskset *tset)
6999 struct task_struct *task;
7001 cgroup_taskset_for_each(task, css, tset)
7002 sched_move_task(task);
7005 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7006 struct cgroup_subsys_state *old_css,
7007 struct task_struct *task)
7010 * cgroup_exit() is called in the copy_process() failure path.
7011 * Ignore this case since the task hasn't ran yet, this avoids
7012 * trying to poke a half freed task state from generic code.
7014 if (!(task->flags & PF_EXITING))
7017 sched_move_task(task);
7020 #ifdef CONFIG_FAIR_GROUP_SCHED
7021 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7022 struct cftype *cftype, u64 shareval)
7024 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7027 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7030 struct task_group *tg = css_tg(css);
7032 return (u64) scale_load_down(tg->shares);
7035 #ifdef CONFIG_CFS_BANDWIDTH
7036 static DEFINE_MUTEX(cfs_constraints_mutex);
7038 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7039 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7041 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7043 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7045 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7046 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7048 if (tg == &root_task_group)
7052 * Ensure we have at some amount of bandwidth every period. This is
7053 * to prevent reaching a state of large arrears when throttled via
7054 * entity_tick() resulting in prolonged exit starvation.
7056 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7060 * Likewise, bound things on the otherside by preventing insane quota
7061 * periods. This also allows us to normalize in computing quota
7064 if (period > max_cfs_quota_period)
7067 mutex_lock(&cfs_constraints_mutex);
7068 ret = __cfs_schedulable(tg, period, quota);
7072 runtime_enabled = quota != RUNTIME_INF;
7073 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7075 * If we need to toggle cfs_bandwidth_used, off->on must occur
7076 * before making related changes, and on->off must occur afterwards
7078 if (runtime_enabled && !runtime_was_enabled)
7079 cfs_bandwidth_usage_inc();
7080 raw_spin_lock_irq(&cfs_b->lock);
7081 cfs_b->period = ns_to_ktime(period);
7082 cfs_b->quota = quota;
7084 __refill_cfs_bandwidth_runtime(cfs_b);
7085 /* restart the period timer (if active) to handle new period expiry */
7086 if (runtime_enabled && cfs_b->timer_active) {
7087 /* force a reprogram */
7088 cfs_b->timer_active = 0;
7089 __start_cfs_bandwidth(cfs_b);
7091 raw_spin_unlock_irq(&cfs_b->lock);
7093 for_each_possible_cpu(i) {
7094 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7095 struct rq *rq = cfs_rq->rq;
7097 raw_spin_lock_irq(&rq->lock);
7098 cfs_rq->runtime_enabled = runtime_enabled;
7099 cfs_rq->runtime_remaining = 0;
7101 if (cfs_rq->throttled)
7102 unthrottle_cfs_rq(cfs_rq);
7103 raw_spin_unlock_irq(&rq->lock);
7105 if (runtime_was_enabled && !runtime_enabled)
7106 cfs_bandwidth_usage_dec();
7108 mutex_unlock(&cfs_constraints_mutex);
7113 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7117 period = ktime_to_ns(tg->cfs_bandwidth.period);
7118 if (cfs_quota_us < 0)
7119 quota = RUNTIME_INF;
7121 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7123 return tg_set_cfs_bandwidth(tg, period, quota);
7126 long tg_get_cfs_quota(struct task_group *tg)
7130 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7133 quota_us = tg->cfs_bandwidth.quota;
7134 do_div(quota_us, NSEC_PER_USEC);
7139 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7143 period = (u64)cfs_period_us * NSEC_PER_USEC;
7144 quota = tg->cfs_bandwidth.quota;
7146 return tg_set_cfs_bandwidth(tg, period, quota);
7149 long tg_get_cfs_period(struct task_group *tg)
7153 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7154 do_div(cfs_period_us, NSEC_PER_USEC);
7156 return cfs_period_us;
7159 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7162 return tg_get_cfs_quota(css_tg(css));
7165 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7166 struct cftype *cftype, s64 cfs_quota_us)
7168 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7171 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7174 return tg_get_cfs_period(css_tg(css));
7177 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7178 struct cftype *cftype, u64 cfs_period_us)
7180 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7183 struct cfs_schedulable_data {
7184 struct task_group *tg;
7189 * normalize group quota/period to be quota/max_period
7190 * note: units are usecs
7192 static u64 normalize_cfs_quota(struct task_group *tg,
7193 struct cfs_schedulable_data *d)
7201 period = tg_get_cfs_period(tg);
7202 quota = tg_get_cfs_quota(tg);
7205 /* note: these should typically be equivalent */
7206 if (quota == RUNTIME_INF || quota == -1)
7209 return to_ratio(period, quota);
7212 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7214 struct cfs_schedulable_data *d = data;
7215 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7216 s64 quota = 0, parent_quota = -1;
7219 quota = RUNTIME_INF;
7221 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7223 quota = normalize_cfs_quota(tg, d);
7224 parent_quota = parent_b->hierarchal_quota;
7227 * ensure max(child_quota) <= parent_quota, inherit when no
7230 if (quota == RUNTIME_INF)
7231 quota = parent_quota;
7232 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7235 cfs_b->hierarchal_quota = quota;
7240 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7243 struct cfs_schedulable_data data = {
7249 if (quota != RUNTIME_INF) {
7250 do_div(data.period, NSEC_PER_USEC);
7251 do_div(data.quota, NSEC_PER_USEC);
7255 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7261 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7262 struct cgroup_map_cb *cb)
7264 struct task_group *tg = css_tg(css);
7265 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7267 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7268 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7269 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7273 #endif /* CONFIG_CFS_BANDWIDTH */
7274 #endif /* CONFIG_FAIR_GROUP_SCHED */
7276 #ifdef CONFIG_RT_GROUP_SCHED
7277 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7278 struct cftype *cft, s64 val)
7280 return sched_group_set_rt_runtime(css_tg(css), val);
7283 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7286 return sched_group_rt_runtime(css_tg(css));
7289 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7290 struct cftype *cftype, u64 rt_period_us)
7292 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7295 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7298 return sched_group_rt_period(css_tg(css));
7300 #endif /* CONFIG_RT_GROUP_SCHED */
7302 static struct cftype cpu_files[] = {
7303 #ifdef CONFIG_FAIR_GROUP_SCHED
7306 .read_u64 = cpu_shares_read_u64,
7307 .write_u64 = cpu_shares_write_u64,
7310 #ifdef CONFIG_CFS_BANDWIDTH
7312 .name = "cfs_quota_us",
7313 .read_s64 = cpu_cfs_quota_read_s64,
7314 .write_s64 = cpu_cfs_quota_write_s64,
7317 .name = "cfs_period_us",
7318 .read_u64 = cpu_cfs_period_read_u64,
7319 .write_u64 = cpu_cfs_period_write_u64,
7323 .read_map = cpu_stats_show,
7326 #ifdef CONFIG_RT_GROUP_SCHED
7328 .name = "rt_runtime_us",
7329 .read_s64 = cpu_rt_runtime_read,
7330 .write_s64 = cpu_rt_runtime_write,
7333 .name = "rt_period_us",
7334 .read_u64 = cpu_rt_period_read_uint,
7335 .write_u64 = cpu_rt_period_write_uint,
7341 struct cgroup_subsys cpu_cgroup_subsys = {
7343 .css_alloc = cpu_cgroup_css_alloc,
7344 .css_free = cpu_cgroup_css_free,
7345 .css_online = cpu_cgroup_css_online,
7346 .css_offline = cpu_cgroup_css_offline,
7347 .can_attach = cpu_cgroup_can_attach,
7348 .attach = cpu_cgroup_attach,
7349 .exit = cpu_cgroup_exit,
7350 .subsys_id = cpu_cgroup_subsys_id,
7351 .base_cftypes = cpu_files,
7355 #endif /* CONFIG_CGROUP_SCHED */
7357 void dump_cpu_task(int cpu)
7359 pr_info("Task dump for CPU %d:\n", cpu);
7360 sched_show_task(cpu_curr(cpu));