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
517 void resched_task(struct task_struct *p)
521 assert_raw_spin_locked(&task_rq(p)->lock);
523 if (test_tsk_need_resched(p))
526 set_tsk_need_resched(p);
529 if (cpu == smp_processor_id())
532 /* NEED_RESCHED must be visible before we test polling */
534 if (!tsk_is_polling(p))
535 smp_send_reschedule(cpu);
538 void resched_cpu(int cpu)
540 struct rq *rq = cpu_rq(cpu);
543 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
545 resched_task(cpu_curr(cpu));
546 raw_spin_unlock_irqrestore(&rq->lock, flags);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu = smp_processor_id();
562 struct sched_domain *sd;
565 for_each_domain(cpu, sd) {
566 for_each_cpu(i, sched_domain_span(sd)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu)
589 struct rq *rq = cpu_rq(cpu);
591 if (cpu == smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq->curr != rq->idle)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq->idle);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq->idle))
614 smp_send_reschedule(cpu);
617 static bool wake_up_full_nohz_cpu(int cpu)
619 if (tick_nohz_full_cpu(cpu)) {
620 if (cpu != smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu);
629 void wake_up_nohz_cpu(int cpu)
631 if (!wake_up_full_nohz_cpu(cpu))
632 wake_up_idle_cpu(cpu);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu = smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
642 if (idle_cpu(cpu) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq->nr_running > 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq *rq)
682 s64 period = sched_avg_period();
684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq->age_stamp));
691 rq->age_stamp += period;
696 #else /* !CONFIG_SMP */
697 void resched_task(struct task_struct *p)
699 assert_raw_spin_locked(&task_rq(p)->lock);
700 set_tsk_need_resched(p);
702 #endif /* CONFIG_SMP */
704 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
705 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
707 * Iterate task_group tree rooted at *from, calling @down when first entering a
708 * node and @up when leaving it for the final time.
710 * Caller must hold rcu_lock or sufficient equivalent.
712 int walk_tg_tree_from(struct task_group *from,
713 tg_visitor down, tg_visitor up, void *data)
715 struct task_group *parent, *child;
721 ret = (*down)(parent, data);
724 list_for_each_entry_rcu(child, &parent->children, siblings) {
731 ret = (*up)(parent, data);
732 if (ret || parent == from)
736 parent = parent->parent;
743 int tg_nop(struct task_group *tg, void *data)
749 static void set_load_weight(struct task_struct *p)
751 int prio = p->static_prio - MAX_RT_PRIO;
752 struct load_weight *load = &p->se.load;
755 * SCHED_IDLE tasks get minimal weight:
757 if (p->policy == SCHED_IDLE) {
758 load->weight = scale_load(WEIGHT_IDLEPRIO);
759 load->inv_weight = WMULT_IDLEPRIO;
763 load->weight = scale_load(prio_to_weight[prio]);
764 load->inv_weight = prio_to_wmult[prio];
767 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
770 sched_info_queued(p);
771 p->sched_class->enqueue_task(rq, p, flags);
774 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
777 sched_info_dequeued(p);
778 p->sched_class->dequeue_task(rq, p, flags);
781 void activate_task(struct rq *rq, struct task_struct *p, int flags)
783 if (task_contributes_to_load(p))
784 rq->nr_uninterruptible--;
786 enqueue_task(rq, p, flags);
789 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
791 if (task_contributes_to_load(p))
792 rq->nr_uninterruptible++;
794 dequeue_task(rq, p, flags);
797 static void update_rq_clock_task(struct rq *rq, s64 delta)
800 * In theory, the compile should just see 0 here, and optimize out the call
801 * to sched_rt_avg_update. But I don't trust it...
803 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
804 s64 steal = 0, irq_delta = 0;
806 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
807 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
810 * Since irq_time is only updated on {soft,}irq_exit, we might run into
811 * this case when a previous update_rq_clock() happened inside a
814 * When this happens, we stop ->clock_task and only update the
815 * prev_irq_time stamp to account for the part that fit, so that a next
816 * update will consume the rest. This ensures ->clock_task is
819 * It does however cause some slight miss-attribution of {soft,}irq
820 * time, a more accurate solution would be to update the irq_time using
821 * the current rq->clock timestamp, except that would require using
824 if (irq_delta > delta)
827 rq->prev_irq_time += irq_delta;
830 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
831 if (static_key_false((¶virt_steal_rq_enabled))) {
834 steal = paravirt_steal_clock(cpu_of(rq));
835 steal -= rq->prev_steal_time_rq;
837 if (unlikely(steal > delta))
840 st = steal_ticks(steal);
841 steal = st * TICK_NSEC;
843 rq->prev_steal_time_rq += steal;
849 rq->clock_task += delta;
851 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
852 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
853 sched_rt_avg_update(rq, irq_delta + steal);
857 void sched_set_stop_task(int cpu, struct task_struct *stop)
859 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
860 struct task_struct *old_stop = cpu_rq(cpu)->stop;
864 * Make it appear like a SCHED_FIFO task, its something
865 * userspace knows about and won't get confused about.
867 * Also, it will make PI more or less work without too
868 * much confusion -- but then, stop work should not
869 * rely on PI working anyway.
871 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
873 stop->sched_class = &stop_sched_class;
876 cpu_rq(cpu)->stop = stop;
880 * Reset it back to a normal scheduling class so that
881 * it can die in pieces.
883 old_stop->sched_class = &rt_sched_class;
888 * __normal_prio - return the priority that is based on the static prio
890 static inline int __normal_prio(struct task_struct *p)
892 return p->static_prio;
896 * Calculate the expected normal priority: i.e. priority
897 * without taking RT-inheritance into account. Might be
898 * boosted by interactivity modifiers. Changes upon fork,
899 * setprio syscalls, and whenever the interactivity
900 * estimator recalculates.
902 static inline int normal_prio(struct task_struct *p)
906 if (task_has_rt_policy(p))
907 prio = MAX_RT_PRIO-1 - p->rt_priority;
909 prio = __normal_prio(p);
914 * Calculate the current priority, i.e. the priority
915 * taken into account by the scheduler. This value might
916 * be boosted by RT tasks, or might be boosted by
917 * interactivity modifiers. Will be RT if the task got
918 * RT-boosted. If not then it returns p->normal_prio.
920 static int effective_prio(struct task_struct *p)
922 p->normal_prio = normal_prio(p);
924 * If we are RT tasks or we were boosted to RT priority,
925 * keep the priority unchanged. Otherwise, update priority
926 * to the normal priority:
928 if (!rt_prio(p->prio))
929 return p->normal_prio;
934 * task_curr - is this task currently executing on a CPU?
935 * @p: the task in question.
937 inline int task_curr(const struct task_struct *p)
939 return cpu_curr(task_cpu(p)) == p;
942 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
943 const struct sched_class *prev_class,
946 if (prev_class != p->sched_class) {
947 if (prev_class->switched_from)
948 prev_class->switched_from(rq, p);
949 p->sched_class->switched_to(rq, p);
950 } else if (oldprio != p->prio)
951 p->sched_class->prio_changed(rq, p, oldprio);
954 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
956 const struct sched_class *class;
958 if (p->sched_class == rq->curr->sched_class) {
959 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
961 for_each_class(class) {
962 if (class == rq->curr->sched_class)
964 if (class == p->sched_class) {
965 resched_task(rq->curr);
972 * A queue event has occurred, and we're going to schedule. In
973 * this case, we can save a useless back to back clock update.
975 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
976 rq->skip_clock_update = 1;
980 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
982 #ifdef CONFIG_SCHED_DEBUG
984 * We should never call set_task_cpu() on a blocked task,
985 * ttwu() will sort out the placement.
987 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
988 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
990 #ifdef CONFIG_LOCKDEP
992 * The caller should hold either p->pi_lock or rq->lock, when changing
993 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
995 * sched_move_task() holds both and thus holding either pins the cgroup,
998 * Furthermore, all task_rq users should acquire both locks, see
1001 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1002 lockdep_is_held(&task_rq(p)->lock)));
1006 trace_sched_migrate_task(p, new_cpu);
1008 if (task_cpu(p) != new_cpu) {
1009 if (p->sched_class->migrate_task_rq)
1010 p->sched_class->migrate_task_rq(p, new_cpu);
1011 p->se.nr_migrations++;
1012 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1015 __set_task_cpu(p, new_cpu);
1018 struct migration_arg {
1019 struct task_struct *task;
1023 static int migration_cpu_stop(void *data);
1026 * wait_task_inactive - wait for a thread to unschedule.
1028 * If @match_state is nonzero, it's the @p->state value just checked and
1029 * not expected to change. If it changes, i.e. @p might have woken up,
1030 * then return zero. When we succeed in waiting for @p to be off its CPU,
1031 * we return a positive number (its total switch count). If a second call
1032 * a short while later returns the same number, the caller can be sure that
1033 * @p has remained unscheduled the whole time.
1035 * The caller must ensure that the task *will* unschedule sometime soon,
1036 * else this function might spin for a *long* time. This function can't
1037 * be called with interrupts off, or it may introduce deadlock with
1038 * smp_call_function() if an IPI is sent by the same process we are
1039 * waiting to become inactive.
1041 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1043 unsigned long flags;
1050 * We do the initial early heuristics without holding
1051 * any task-queue locks at all. We'll only try to get
1052 * the runqueue lock when things look like they will
1058 * If the task is actively running on another CPU
1059 * still, just relax and busy-wait without holding
1062 * NOTE! Since we don't hold any locks, it's not
1063 * even sure that "rq" stays as the right runqueue!
1064 * But we don't care, since "task_running()" will
1065 * return false if the runqueue has changed and p
1066 * is actually now running somewhere else!
1068 while (task_running(rq, p)) {
1069 if (match_state && unlikely(p->state != match_state))
1075 * Ok, time to look more closely! We need the rq
1076 * lock now, to be *sure*. If we're wrong, we'll
1077 * just go back and repeat.
1079 rq = task_rq_lock(p, &flags);
1080 trace_sched_wait_task(p);
1081 running = task_running(rq, p);
1084 if (!match_state || p->state == match_state)
1085 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1086 task_rq_unlock(rq, p, &flags);
1089 * If it changed from the expected state, bail out now.
1091 if (unlikely(!ncsw))
1095 * Was it really running after all now that we
1096 * checked with the proper locks actually held?
1098 * Oops. Go back and try again..
1100 if (unlikely(running)) {
1106 * It's not enough that it's not actively running,
1107 * it must be off the runqueue _entirely_, and not
1110 * So if it was still runnable (but just not actively
1111 * running right now), it's preempted, and we should
1112 * yield - it could be a while.
1114 if (unlikely(on_rq)) {
1115 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1117 set_current_state(TASK_UNINTERRUPTIBLE);
1118 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1123 * Ahh, all good. It wasn't running, and it wasn't
1124 * runnable, which means that it will never become
1125 * running in the future either. We're all done!
1134 * kick_process - kick a running thread to enter/exit the kernel
1135 * @p: the to-be-kicked thread
1137 * Cause a process which is running on another CPU to enter
1138 * kernel-mode, without any delay. (to get signals handled.)
1140 * NOTE: this function doesn't have to take the runqueue lock,
1141 * because all it wants to ensure is that the remote task enters
1142 * the kernel. If the IPI races and the task has been migrated
1143 * to another CPU then no harm is done and the purpose has been
1146 void kick_process(struct task_struct *p)
1152 if ((cpu != smp_processor_id()) && task_curr(p))
1153 smp_send_reschedule(cpu);
1156 EXPORT_SYMBOL_GPL(kick_process);
1157 #endif /* CONFIG_SMP */
1161 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1163 static int select_fallback_rq(int cpu, struct task_struct *p)
1165 int nid = cpu_to_node(cpu);
1166 const struct cpumask *nodemask = NULL;
1167 enum { cpuset, possible, fail } state = cpuset;
1171 * If the node that the cpu is on has been offlined, cpu_to_node()
1172 * will return -1. There is no cpu on the node, and we should
1173 * select the cpu on the other node.
1176 nodemask = cpumask_of_node(nid);
1178 /* Look for allowed, online CPU in same node. */
1179 for_each_cpu(dest_cpu, nodemask) {
1180 if (!cpu_online(dest_cpu))
1182 if (!cpu_active(dest_cpu))
1184 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1190 /* Any allowed, online CPU? */
1191 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1192 if (!cpu_online(dest_cpu))
1194 if (!cpu_active(dest_cpu))
1201 /* No more Mr. Nice Guy. */
1202 cpuset_cpus_allowed_fallback(p);
1207 do_set_cpus_allowed(p, cpu_possible_mask);
1218 if (state != cpuset) {
1220 * Don't tell them about moving exiting tasks or
1221 * kernel threads (both mm NULL), since they never
1224 if (p->mm && printk_ratelimit()) {
1225 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1226 task_pid_nr(p), p->comm, cpu);
1234 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1237 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1239 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1242 * In order not to call set_task_cpu() on a blocking task we need
1243 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1246 * Since this is common to all placement strategies, this lives here.
1248 * [ this allows ->select_task() to simply return task_cpu(p) and
1249 * not worry about this generic constraint ]
1251 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1253 cpu = select_fallback_rq(task_cpu(p), p);
1258 static void update_avg(u64 *avg, u64 sample)
1260 s64 diff = sample - *avg;
1266 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1268 #ifdef CONFIG_SCHEDSTATS
1269 struct rq *rq = this_rq();
1272 int this_cpu = smp_processor_id();
1274 if (cpu == this_cpu) {
1275 schedstat_inc(rq, ttwu_local);
1276 schedstat_inc(p, se.statistics.nr_wakeups_local);
1278 struct sched_domain *sd;
1280 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1282 for_each_domain(this_cpu, sd) {
1283 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1284 schedstat_inc(sd, ttwu_wake_remote);
1291 if (wake_flags & WF_MIGRATED)
1292 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1294 #endif /* CONFIG_SMP */
1296 schedstat_inc(rq, ttwu_count);
1297 schedstat_inc(p, se.statistics.nr_wakeups);
1299 if (wake_flags & WF_SYNC)
1300 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1302 #endif /* CONFIG_SCHEDSTATS */
1305 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1307 activate_task(rq, p, en_flags);
1310 /* if a worker is waking up, notify workqueue */
1311 if (p->flags & PF_WQ_WORKER)
1312 wq_worker_waking_up(p, cpu_of(rq));
1316 * Mark the task runnable and perform wakeup-preemption.
1319 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1321 check_preempt_curr(rq, p, wake_flags);
1322 trace_sched_wakeup(p, true);
1324 p->state = TASK_RUNNING;
1326 if (p->sched_class->task_woken)
1327 p->sched_class->task_woken(rq, p);
1329 if (rq->idle_stamp) {
1330 u64 delta = rq_clock(rq) - rq->idle_stamp;
1331 u64 max = 2*sysctl_sched_migration_cost;
1336 update_avg(&rq->avg_idle, delta);
1343 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1346 if (p->sched_contributes_to_load)
1347 rq->nr_uninterruptible--;
1350 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1351 ttwu_do_wakeup(rq, p, wake_flags);
1355 * Called in case the task @p isn't fully descheduled from its runqueue,
1356 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1357 * since all we need to do is flip p->state to TASK_RUNNING, since
1358 * the task is still ->on_rq.
1360 static int ttwu_remote(struct task_struct *p, int wake_flags)
1365 rq = __task_rq_lock(p);
1367 /* check_preempt_curr() may use rq clock */
1368 update_rq_clock(rq);
1369 ttwu_do_wakeup(rq, p, wake_flags);
1372 __task_rq_unlock(rq);
1378 static void sched_ttwu_pending(void)
1380 struct rq *rq = this_rq();
1381 struct llist_node *llist = llist_del_all(&rq->wake_list);
1382 struct task_struct *p;
1384 raw_spin_lock(&rq->lock);
1387 p = llist_entry(llist, struct task_struct, wake_entry);
1388 llist = llist_next(llist);
1389 ttwu_do_activate(rq, p, 0);
1392 raw_spin_unlock(&rq->lock);
1395 void scheduler_ipi(void)
1397 if (llist_empty(&this_rq()->wake_list)
1398 && !tick_nohz_full_cpu(smp_processor_id())
1399 && !got_nohz_idle_kick())
1403 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1404 * traditionally all their work was done from the interrupt return
1405 * path. Now that we actually do some work, we need to make sure
1408 * Some archs already do call them, luckily irq_enter/exit nest
1411 * Arguably we should visit all archs and update all handlers,
1412 * however a fair share of IPIs are still resched only so this would
1413 * somewhat pessimize the simple resched case.
1416 tick_nohz_full_check();
1417 sched_ttwu_pending();
1420 * Check if someone kicked us for doing the nohz idle load balance.
1422 if (unlikely(got_nohz_idle_kick())) {
1423 this_rq()->idle_balance = 1;
1424 raise_softirq_irqoff(SCHED_SOFTIRQ);
1429 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1431 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1432 smp_send_reschedule(cpu);
1435 bool cpus_share_cache(int this_cpu, int that_cpu)
1437 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1439 #endif /* CONFIG_SMP */
1441 static void ttwu_queue(struct task_struct *p, int cpu)
1443 struct rq *rq = cpu_rq(cpu);
1445 #if defined(CONFIG_SMP)
1446 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1447 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1448 ttwu_queue_remote(p, cpu);
1453 raw_spin_lock(&rq->lock);
1454 ttwu_do_activate(rq, p, 0);
1455 raw_spin_unlock(&rq->lock);
1459 * try_to_wake_up - wake up a thread
1460 * @p: the thread to be awakened
1461 * @state: the mask of task states that can be woken
1462 * @wake_flags: wake modifier flags (WF_*)
1464 * Put it on the run-queue if it's not already there. The "current"
1465 * thread is always on the run-queue (except when the actual
1466 * re-schedule is in progress), and as such you're allowed to do
1467 * the simpler "current->state = TASK_RUNNING" to mark yourself
1468 * runnable without the overhead of this.
1470 * Returns %true if @p was woken up, %false if it was already running
1471 * or @state didn't match @p's state.
1474 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1476 unsigned long flags;
1477 int cpu, success = 0;
1480 raw_spin_lock_irqsave(&p->pi_lock, flags);
1481 if (!(p->state & state))
1484 success = 1; /* we're going to change ->state */
1487 if (p->on_rq && ttwu_remote(p, wake_flags))
1492 * If the owning (remote) cpu is still in the middle of schedule() with
1493 * this task as prev, wait until its done referencing the task.
1498 * Pairs with the smp_wmb() in finish_lock_switch().
1502 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1503 p->state = TASK_WAKING;
1505 if (p->sched_class->task_waking)
1506 p->sched_class->task_waking(p);
1508 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1509 if (task_cpu(p) != cpu) {
1510 wake_flags |= WF_MIGRATED;
1511 set_task_cpu(p, cpu);
1513 #endif /* CONFIG_SMP */
1517 ttwu_stat(p, cpu, wake_flags);
1519 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1525 * try_to_wake_up_local - try to wake up a local task with rq lock held
1526 * @p: the thread to be awakened
1528 * Put @p on the run-queue if it's not already there. The caller must
1529 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1532 static void try_to_wake_up_local(struct task_struct *p)
1534 struct rq *rq = task_rq(p);
1536 if (WARN_ON_ONCE(rq != this_rq()) ||
1537 WARN_ON_ONCE(p == current))
1540 lockdep_assert_held(&rq->lock);
1542 if (!raw_spin_trylock(&p->pi_lock)) {
1543 raw_spin_unlock(&rq->lock);
1544 raw_spin_lock(&p->pi_lock);
1545 raw_spin_lock(&rq->lock);
1548 if (!(p->state & TASK_NORMAL))
1552 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1554 ttwu_do_wakeup(rq, p, 0);
1555 ttwu_stat(p, smp_processor_id(), 0);
1557 raw_spin_unlock(&p->pi_lock);
1561 * wake_up_process - Wake up a specific process
1562 * @p: The process to be woken up.
1564 * Attempt to wake up the nominated process and move it to the set of runnable
1565 * processes. Returns 1 if the process was woken up, 0 if it was already
1568 * It may be assumed that this function implies a write memory barrier before
1569 * changing the task state if and only if any tasks are woken up.
1571 int wake_up_process(struct task_struct *p)
1573 WARN_ON(task_is_stopped_or_traced(p));
1574 return try_to_wake_up(p, TASK_NORMAL, 0);
1576 EXPORT_SYMBOL(wake_up_process);
1578 int wake_up_state(struct task_struct *p, unsigned int state)
1580 return try_to_wake_up(p, state, 0);
1584 * Perform scheduler related setup for a newly forked process p.
1585 * p is forked by current.
1587 * __sched_fork() is basic setup used by init_idle() too:
1589 static void __sched_fork(struct task_struct *p)
1594 p->se.exec_start = 0;
1595 p->se.sum_exec_runtime = 0;
1596 p->se.prev_sum_exec_runtime = 0;
1597 p->se.nr_migrations = 0;
1599 INIT_LIST_HEAD(&p->se.group_node);
1601 #ifdef CONFIG_SCHEDSTATS
1602 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1605 INIT_LIST_HEAD(&p->rt.run_list);
1607 #ifdef CONFIG_PREEMPT_NOTIFIERS
1608 INIT_HLIST_HEAD(&p->preempt_notifiers);
1611 #ifdef CONFIG_NUMA_BALANCING
1612 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1613 p->mm->numa_next_scan = jiffies;
1614 p->mm->numa_next_reset = jiffies;
1615 p->mm->numa_scan_seq = 0;
1618 p->node_stamp = 0ULL;
1619 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1620 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1621 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1622 p->numa_work.next = &p->numa_work;
1623 #endif /* CONFIG_NUMA_BALANCING */
1626 #ifdef CONFIG_NUMA_BALANCING
1627 #ifdef CONFIG_SCHED_DEBUG
1628 void set_numabalancing_state(bool enabled)
1631 sched_feat_set("NUMA");
1633 sched_feat_set("NO_NUMA");
1636 __read_mostly bool numabalancing_enabled;
1638 void set_numabalancing_state(bool enabled)
1640 numabalancing_enabled = enabled;
1642 #endif /* CONFIG_SCHED_DEBUG */
1643 #endif /* CONFIG_NUMA_BALANCING */
1646 * fork()/clone()-time setup:
1648 void sched_fork(struct task_struct *p)
1650 unsigned long flags;
1651 int cpu = get_cpu();
1655 * We mark the process as running here. This guarantees that
1656 * nobody will actually run it, and a signal or other external
1657 * event cannot wake it up and insert it on the runqueue either.
1659 p->state = TASK_RUNNING;
1662 * Make sure we do not leak PI boosting priority to the child.
1664 p->prio = current->normal_prio;
1667 * Revert to default priority/policy on fork if requested.
1669 if (unlikely(p->sched_reset_on_fork)) {
1670 if (task_has_rt_policy(p)) {
1671 p->policy = SCHED_NORMAL;
1672 p->static_prio = NICE_TO_PRIO(0);
1674 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1675 p->static_prio = NICE_TO_PRIO(0);
1677 p->prio = p->normal_prio = __normal_prio(p);
1681 * We don't need the reset flag anymore after the fork. It has
1682 * fulfilled its duty:
1684 p->sched_reset_on_fork = 0;
1687 if (!rt_prio(p->prio))
1688 p->sched_class = &fair_sched_class;
1690 if (p->sched_class->task_fork)
1691 p->sched_class->task_fork(p);
1694 * The child is not yet in the pid-hash so no cgroup attach races,
1695 * and the cgroup is pinned to this child due to cgroup_fork()
1696 * is ran before sched_fork().
1698 * Silence PROVE_RCU.
1700 raw_spin_lock_irqsave(&p->pi_lock, flags);
1701 set_task_cpu(p, cpu);
1702 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1704 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1705 if (likely(sched_info_on()))
1706 memset(&p->sched_info, 0, sizeof(p->sched_info));
1708 #if defined(CONFIG_SMP)
1711 #ifdef CONFIG_PREEMPT_COUNT
1712 /* Want to start with kernel preemption disabled. */
1713 task_thread_info(p)->preempt_count = 1;
1716 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1723 * wake_up_new_task - wake up a newly created task for the first time.
1725 * This function will do some initial scheduler statistics housekeeping
1726 * that must be done for every newly created context, then puts the task
1727 * on the runqueue and wakes it.
1729 void wake_up_new_task(struct task_struct *p)
1731 unsigned long flags;
1734 raw_spin_lock_irqsave(&p->pi_lock, flags);
1737 * Fork balancing, do it here and not earlier because:
1738 * - cpus_allowed can change in the fork path
1739 * - any previously selected cpu might disappear through hotplug
1741 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1744 /* Initialize new task's runnable average */
1745 init_task_runnable_average(p);
1746 rq = __task_rq_lock(p);
1747 activate_task(rq, p, 0);
1749 trace_sched_wakeup_new(p, true);
1750 check_preempt_curr(rq, p, WF_FORK);
1752 if (p->sched_class->task_woken)
1753 p->sched_class->task_woken(rq, p);
1755 task_rq_unlock(rq, p, &flags);
1758 #ifdef CONFIG_PREEMPT_NOTIFIERS
1761 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1762 * @notifier: notifier struct to register
1764 void preempt_notifier_register(struct preempt_notifier *notifier)
1766 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1768 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1771 * preempt_notifier_unregister - no longer interested in preemption notifications
1772 * @notifier: notifier struct to unregister
1774 * This is safe to call from within a preemption notifier.
1776 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1778 hlist_del(¬ifier->link);
1780 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1782 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1784 struct preempt_notifier *notifier;
1786 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1787 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1791 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1792 struct task_struct *next)
1794 struct preempt_notifier *notifier;
1796 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1797 notifier->ops->sched_out(notifier, next);
1800 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1802 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1807 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1808 struct task_struct *next)
1812 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1815 * prepare_task_switch - prepare to switch tasks
1816 * @rq: the runqueue preparing to switch
1817 * @prev: the current task that is being switched out
1818 * @next: the task we are going to switch to.
1820 * This is called with the rq lock held and interrupts off. It must
1821 * be paired with a subsequent finish_task_switch after the context
1824 * prepare_task_switch sets up locking and calls architecture specific
1828 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1829 struct task_struct *next)
1831 trace_sched_switch(prev, next);
1832 sched_info_switch(prev, next);
1833 perf_event_task_sched_out(prev, next);
1834 fire_sched_out_preempt_notifiers(prev, next);
1835 prepare_lock_switch(rq, next);
1836 prepare_arch_switch(next);
1840 * finish_task_switch - clean up after a task-switch
1841 * @rq: runqueue associated with task-switch
1842 * @prev: the thread we just switched away from.
1844 * finish_task_switch must be called after the context switch, paired
1845 * with a prepare_task_switch call before the context switch.
1846 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1847 * and do any other architecture-specific cleanup actions.
1849 * Note that we may have delayed dropping an mm in context_switch(). If
1850 * so, we finish that here outside of the runqueue lock. (Doing it
1851 * with the lock held can cause deadlocks; see schedule() for
1854 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1855 __releases(rq->lock)
1857 struct mm_struct *mm = rq->prev_mm;
1863 * A task struct has one reference for the use as "current".
1864 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1865 * schedule one last time. The schedule call will never return, and
1866 * the scheduled task must drop that reference.
1867 * The test for TASK_DEAD must occur while the runqueue locks are
1868 * still held, otherwise prev could be scheduled on another cpu, die
1869 * there before we look at prev->state, and then the reference would
1871 * Manfred Spraul <manfred@colorfullife.com>
1873 prev_state = prev->state;
1874 vtime_task_switch(prev);
1875 finish_arch_switch(prev);
1876 perf_event_task_sched_in(prev, current);
1877 finish_lock_switch(rq, prev);
1878 finish_arch_post_lock_switch();
1880 fire_sched_in_preempt_notifiers(current);
1883 if (unlikely(prev_state == TASK_DEAD)) {
1885 * Remove function-return probe instances associated with this
1886 * task and put them back on the free list.
1888 kprobe_flush_task(prev);
1889 put_task_struct(prev);
1892 tick_nohz_task_switch(current);
1897 /* assumes rq->lock is held */
1898 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1900 if (prev->sched_class->pre_schedule)
1901 prev->sched_class->pre_schedule(rq, prev);
1904 /* rq->lock is NOT held, but preemption is disabled */
1905 static inline void post_schedule(struct rq *rq)
1907 if (rq->post_schedule) {
1908 unsigned long flags;
1910 raw_spin_lock_irqsave(&rq->lock, flags);
1911 if (rq->curr->sched_class->post_schedule)
1912 rq->curr->sched_class->post_schedule(rq);
1913 raw_spin_unlock_irqrestore(&rq->lock, flags);
1915 rq->post_schedule = 0;
1921 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1925 static inline void post_schedule(struct rq *rq)
1932 * schedule_tail - first thing a freshly forked thread must call.
1933 * @prev: the thread we just switched away from.
1935 asmlinkage void schedule_tail(struct task_struct *prev)
1936 __releases(rq->lock)
1938 struct rq *rq = this_rq();
1940 finish_task_switch(rq, prev);
1943 * FIXME: do we need to worry about rq being invalidated by the
1948 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1949 /* In this case, finish_task_switch does not reenable preemption */
1952 if (current->set_child_tid)
1953 put_user(task_pid_vnr(current), current->set_child_tid);
1957 * context_switch - switch to the new MM and the new
1958 * thread's register state.
1961 context_switch(struct rq *rq, struct task_struct *prev,
1962 struct task_struct *next)
1964 struct mm_struct *mm, *oldmm;
1966 prepare_task_switch(rq, prev, next);
1969 oldmm = prev->active_mm;
1971 * For paravirt, this is coupled with an exit in switch_to to
1972 * combine the page table reload and the switch backend into
1975 arch_start_context_switch(prev);
1978 next->active_mm = oldmm;
1979 atomic_inc(&oldmm->mm_count);
1980 enter_lazy_tlb(oldmm, next);
1982 switch_mm(oldmm, mm, next);
1985 prev->active_mm = NULL;
1986 rq->prev_mm = oldmm;
1989 * Since the runqueue lock will be released by the next
1990 * task (which is an invalid locking op but in the case
1991 * of the scheduler it's an obvious special-case), so we
1992 * do an early lockdep release here:
1994 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1995 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1998 context_tracking_task_switch(prev, next);
1999 /* Here we just switch the register state and the stack. */
2000 switch_to(prev, next, prev);
2004 * this_rq must be evaluated again because prev may have moved
2005 * CPUs since it called schedule(), thus the 'rq' on its stack
2006 * frame will be invalid.
2008 finish_task_switch(this_rq(), prev);
2012 * nr_running and nr_context_switches:
2014 * externally visible scheduler statistics: current number of runnable
2015 * threads, total number of context switches performed since bootup.
2017 unsigned long nr_running(void)
2019 unsigned long i, sum = 0;
2021 for_each_online_cpu(i)
2022 sum += cpu_rq(i)->nr_running;
2027 unsigned long long nr_context_switches(void)
2030 unsigned long long sum = 0;
2032 for_each_possible_cpu(i)
2033 sum += cpu_rq(i)->nr_switches;
2038 unsigned long nr_iowait(void)
2040 unsigned long i, sum = 0;
2042 for_each_possible_cpu(i)
2043 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2048 unsigned long nr_iowait_cpu(int cpu)
2050 struct rq *this = cpu_rq(cpu);
2051 return atomic_read(&this->nr_iowait);
2057 * sched_exec - execve() is a valuable balancing opportunity, because at
2058 * this point the task has the smallest effective memory and cache footprint.
2060 void sched_exec(void)
2062 struct task_struct *p = current;
2063 unsigned long flags;
2066 raw_spin_lock_irqsave(&p->pi_lock, flags);
2067 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2068 if (dest_cpu == smp_processor_id())
2071 if (likely(cpu_active(dest_cpu))) {
2072 struct migration_arg arg = { p, dest_cpu };
2074 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2075 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2079 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2084 DEFINE_PER_CPU(struct kernel_stat, kstat);
2085 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2087 EXPORT_PER_CPU_SYMBOL(kstat);
2088 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2091 * Return any ns on the sched_clock that have not yet been accounted in
2092 * @p in case that task is currently running.
2094 * Called with task_rq_lock() held on @rq.
2096 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2100 if (task_current(rq, p)) {
2101 update_rq_clock(rq);
2102 ns = rq_clock_task(rq) - p->se.exec_start;
2110 unsigned long long task_delta_exec(struct task_struct *p)
2112 unsigned long flags;
2116 rq = task_rq_lock(p, &flags);
2117 ns = do_task_delta_exec(p, rq);
2118 task_rq_unlock(rq, p, &flags);
2124 * Return accounted runtime for the task.
2125 * In case the task is currently running, return the runtime plus current's
2126 * pending runtime that have not been accounted yet.
2128 unsigned long long task_sched_runtime(struct task_struct *p)
2130 unsigned long flags;
2134 rq = task_rq_lock(p, &flags);
2135 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2136 task_rq_unlock(rq, p, &flags);
2142 * This function gets called by the timer code, with HZ frequency.
2143 * We call it with interrupts disabled.
2145 void scheduler_tick(void)
2147 int cpu = smp_processor_id();
2148 struct rq *rq = cpu_rq(cpu);
2149 struct task_struct *curr = rq->curr;
2153 raw_spin_lock(&rq->lock);
2154 update_rq_clock(rq);
2155 curr->sched_class->task_tick(rq, curr, 0);
2156 update_cpu_load_active(rq);
2157 raw_spin_unlock(&rq->lock);
2159 perf_event_task_tick();
2162 rq->idle_balance = idle_cpu(cpu);
2163 trigger_load_balance(rq, cpu);
2165 rq_last_tick_reset(rq);
2168 #ifdef CONFIG_NO_HZ_FULL
2170 * scheduler_tick_max_deferment
2172 * Keep at least one tick per second when a single
2173 * active task is running because the scheduler doesn't
2174 * yet completely support full dynticks environment.
2176 * This makes sure that uptime, CFS vruntime, load
2177 * balancing, etc... continue to move forward, even
2178 * with a very low granularity.
2180 u64 scheduler_tick_max_deferment(void)
2182 struct rq *rq = this_rq();
2183 unsigned long next, now = ACCESS_ONCE(jiffies);
2185 next = rq->last_sched_tick + HZ;
2187 if (time_before_eq(next, now))
2190 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2194 notrace unsigned long get_parent_ip(unsigned long addr)
2196 if (in_lock_functions(addr)) {
2197 addr = CALLER_ADDR2;
2198 if (in_lock_functions(addr))
2199 addr = CALLER_ADDR3;
2204 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2205 defined(CONFIG_PREEMPT_TRACER))
2207 void __kprobes add_preempt_count(int val)
2209 #ifdef CONFIG_DEBUG_PREEMPT
2213 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2216 preempt_count() += val;
2217 #ifdef CONFIG_DEBUG_PREEMPT
2219 * Spinlock count overflowing soon?
2221 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2224 if (preempt_count() == val)
2225 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2227 EXPORT_SYMBOL(add_preempt_count);
2229 void __kprobes sub_preempt_count(int val)
2231 #ifdef CONFIG_DEBUG_PREEMPT
2235 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2238 * Is the spinlock portion underflowing?
2240 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2241 !(preempt_count() & PREEMPT_MASK)))
2245 if (preempt_count() == val)
2246 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2247 preempt_count() -= val;
2249 EXPORT_SYMBOL(sub_preempt_count);
2254 * Print scheduling while atomic bug:
2256 static noinline void __schedule_bug(struct task_struct *prev)
2258 if (oops_in_progress)
2261 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2262 prev->comm, prev->pid, preempt_count());
2264 debug_show_held_locks(prev);
2266 if (irqs_disabled())
2267 print_irqtrace_events(prev);
2269 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2273 * Various schedule()-time debugging checks and statistics:
2275 static inline void schedule_debug(struct task_struct *prev)
2278 * Test if we are atomic. Since do_exit() needs to call into
2279 * schedule() atomically, we ignore that path for now.
2280 * Otherwise, whine if we are scheduling when we should not be.
2282 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2283 __schedule_bug(prev);
2286 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2288 schedstat_inc(this_rq(), sched_count);
2291 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2293 if (prev->on_rq || rq->skip_clock_update < 0)
2294 update_rq_clock(rq);
2295 prev->sched_class->put_prev_task(rq, prev);
2299 * Pick up the highest-prio task:
2301 static inline struct task_struct *
2302 pick_next_task(struct rq *rq)
2304 const struct sched_class *class;
2305 struct task_struct *p;
2308 * Optimization: we know that if all tasks are in
2309 * the fair class we can call that function directly:
2311 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2312 p = fair_sched_class.pick_next_task(rq);
2317 for_each_class(class) {
2318 p = class->pick_next_task(rq);
2323 BUG(); /* the idle class will always have a runnable task */
2327 * __schedule() is the main scheduler function.
2329 * The main means of driving the scheduler and thus entering this function are:
2331 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2333 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2334 * paths. For example, see arch/x86/entry_64.S.
2336 * To drive preemption between tasks, the scheduler sets the flag in timer
2337 * interrupt handler scheduler_tick().
2339 * 3. Wakeups don't really cause entry into schedule(). They add a
2340 * task to the run-queue and that's it.
2342 * Now, if the new task added to the run-queue preempts the current
2343 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2344 * called on the nearest possible occasion:
2346 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2348 * - in syscall or exception context, at the next outmost
2349 * preempt_enable(). (this might be as soon as the wake_up()'s
2352 * - in IRQ context, return from interrupt-handler to
2353 * preemptible context
2355 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2358 * - cond_resched() call
2359 * - explicit schedule() call
2360 * - return from syscall or exception to user-space
2361 * - return from interrupt-handler to user-space
2363 static void __sched __schedule(void)
2365 struct task_struct *prev, *next;
2366 unsigned long *switch_count;
2372 cpu = smp_processor_id();
2374 rcu_note_context_switch(cpu);
2377 schedule_debug(prev);
2379 if (sched_feat(HRTICK))
2382 raw_spin_lock_irq(&rq->lock);
2384 switch_count = &prev->nivcsw;
2385 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2386 if (unlikely(signal_pending_state(prev->state, prev))) {
2387 prev->state = TASK_RUNNING;
2389 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2393 * If a worker went to sleep, notify and ask workqueue
2394 * whether it wants to wake up a task to maintain
2397 if (prev->flags & PF_WQ_WORKER) {
2398 struct task_struct *to_wakeup;
2400 to_wakeup = wq_worker_sleeping(prev, cpu);
2402 try_to_wake_up_local(to_wakeup);
2405 switch_count = &prev->nvcsw;
2408 pre_schedule(rq, prev);
2410 if (unlikely(!rq->nr_running))
2411 idle_balance(cpu, rq);
2413 put_prev_task(rq, prev);
2414 next = pick_next_task(rq);
2415 clear_tsk_need_resched(prev);
2416 rq->skip_clock_update = 0;
2418 if (likely(prev != next)) {
2423 context_switch(rq, prev, next); /* unlocks the rq */
2425 * The context switch have flipped the stack from under us
2426 * and restored the local variables which were saved when
2427 * this task called schedule() in the past. prev == current
2428 * is still correct, but it can be moved to another cpu/rq.
2430 cpu = smp_processor_id();
2433 raw_spin_unlock_irq(&rq->lock);
2437 sched_preempt_enable_no_resched();
2442 static inline void sched_submit_work(struct task_struct *tsk)
2444 if (!tsk->state || tsk_is_pi_blocked(tsk))
2447 * If we are going to sleep and we have plugged IO queued,
2448 * make sure to submit it to avoid deadlocks.
2450 if (blk_needs_flush_plug(tsk))
2451 blk_schedule_flush_plug(tsk);
2454 asmlinkage void __sched schedule(void)
2456 struct task_struct *tsk = current;
2458 sched_submit_work(tsk);
2461 EXPORT_SYMBOL(schedule);
2463 #ifdef CONFIG_CONTEXT_TRACKING
2464 asmlinkage void __sched schedule_user(void)
2467 * If we come here after a random call to set_need_resched(),
2468 * or we have been woken up remotely but the IPI has not yet arrived,
2469 * we haven't yet exited the RCU idle mode. Do it here manually until
2470 * we find a better solution.
2479 * schedule_preempt_disabled - called with preemption disabled
2481 * Returns with preemption disabled. Note: preempt_count must be 1
2483 void __sched schedule_preempt_disabled(void)
2485 sched_preempt_enable_no_resched();
2490 #ifdef CONFIG_PREEMPT
2492 * this is the entry point to schedule() from in-kernel preemption
2493 * off of preempt_enable. Kernel preemptions off return from interrupt
2494 * occur there and call schedule directly.
2496 asmlinkage void __sched notrace preempt_schedule(void)
2498 struct thread_info *ti = current_thread_info();
2501 * If there is a non-zero preempt_count or interrupts are disabled,
2502 * we do not want to preempt the current task. Just return..
2504 if (likely(ti->preempt_count || irqs_disabled()))
2508 add_preempt_count_notrace(PREEMPT_ACTIVE);
2510 sub_preempt_count_notrace(PREEMPT_ACTIVE);
2513 * Check again in case we missed a preemption opportunity
2514 * between schedule and now.
2517 } while (need_resched());
2519 EXPORT_SYMBOL(preempt_schedule);
2522 * this is the entry point to schedule() from kernel preemption
2523 * off of irq context.
2524 * Note, that this is called and return with irqs disabled. This will
2525 * protect us against recursive calling from irq.
2527 asmlinkage void __sched preempt_schedule_irq(void)
2529 struct thread_info *ti = current_thread_info();
2530 enum ctx_state prev_state;
2532 /* Catch callers which need to be fixed */
2533 BUG_ON(ti->preempt_count || !irqs_disabled());
2535 prev_state = exception_enter();
2538 add_preempt_count(PREEMPT_ACTIVE);
2541 local_irq_disable();
2542 sub_preempt_count(PREEMPT_ACTIVE);
2545 * Check again in case we missed a preemption opportunity
2546 * between schedule and now.
2549 } while (need_resched());
2551 exception_exit(prev_state);
2554 #endif /* CONFIG_PREEMPT */
2556 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2559 return try_to_wake_up(curr->private, mode, wake_flags);
2561 EXPORT_SYMBOL(default_wake_function);
2564 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
2565 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
2566 * number) then we wake all the non-exclusive tasks and one exclusive task.
2568 * There are circumstances in which we can try to wake a task which has already
2569 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
2570 * zero in this (rare) case, and we handle it by continuing to scan the queue.
2572 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
2573 int nr_exclusive, int wake_flags, void *key)
2575 wait_queue_t *curr, *next;
2577 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
2578 unsigned flags = curr->flags;
2580 if (curr->func(curr, mode, wake_flags, key) &&
2581 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
2587 * __wake_up - wake up threads blocked on a waitqueue.
2589 * @mode: which threads
2590 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2591 * @key: is directly passed to the wakeup function
2593 * It may be assumed that this function implies a write memory barrier before
2594 * changing the task state if and only if any tasks are woken up.
2596 void __wake_up(wait_queue_head_t *q, unsigned int mode,
2597 int nr_exclusive, void *key)
2599 unsigned long flags;
2601 spin_lock_irqsave(&q->lock, flags);
2602 __wake_up_common(q, mode, nr_exclusive, 0, key);
2603 spin_unlock_irqrestore(&q->lock, flags);
2605 EXPORT_SYMBOL(__wake_up);
2608 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
2610 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
2612 __wake_up_common(q, mode, nr, 0, NULL);
2614 EXPORT_SYMBOL_GPL(__wake_up_locked);
2616 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
2618 __wake_up_common(q, mode, 1, 0, key);
2620 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
2623 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
2625 * @mode: which threads
2626 * @nr_exclusive: how many wake-one or wake-many threads to wake up
2627 * @key: opaque value to be passed to wakeup targets
2629 * The sync wakeup differs that the waker knows that it will schedule
2630 * away soon, so while the target thread will be woken up, it will not
2631 * be migrated to another CPU - ie. the two threads are 'synchronized'
2632 * with each other. This can prevent needless bouncing between CPUs.
2634 * On UP it can prevent extra preemption.
2636 * It may be assumed that this function implies a write memory barrier before
2637 * changing the task state if and only if any tasks are woken up.
2639 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
2640 int nr_exclusive, void *key)
2642 unsigned long flags;
2643 int wake_flags = WF_SYNC;
2648 if (unlikely(!nr_exclusive))
2651 spin_lock_irqsave(&q->lock, flags);
2652 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
2653 spin_unlock_irqrestore(&q->lock, flags);
2655 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
2658 * __wake_up_sync - see __wake_up_sync_key()
2660 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
2662 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
2664 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
2667 * complete: - signals a single thread waiting on this completion
2668 * @x: holds the state of this particular completion
2670 * This will wake up a single thread waiting on this completion. Threads will be
2671 * awakened in the same order in which they were queued.
2673 * See also complete_all(), wait_for_completion() and related routines.
2675 * It may be assumed that this function implies a write memory barrier before
2676 * changing the task state if and only if any tasks are woken up.
2678 void complete(struct completion *x)
2680 unsigned long flags;
2682 spin_lock_irqsave(&x->wait.lock, flags);
2684 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
2685 spin_unlock_irqrestore(&x->wait.lock, flags);
2687 EXPORT_SYMBOL(complete);
2690 * complete_all: - signals all threads waiting on this completion
2691 * @x: holds the state of this particular completion
2693 * This will wake up all threads waiting on this particular completion event.
2695 * It may be assumed that this function implies a write memory barrier before
2696 * changing the task state if and only if any tasks are woken up.
2698 void complete_all(struct completion *x)
2700 unsigned long flags;
2702 spin_lock_irqsave(&x->wait.lock, flags);
2703 x->done += UINT_MAX/2;
2704 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
2705 spin_unlock_irqrestore(&x->wait.lock, flags);
2707 EXPORT_SYMBOL(complete_all);
2709 static inline long __sched
2710 do_wait_for_common(struct completion *x,
2711 long (*action)(long), long timeout, int state)
2714 DECLARE_WAITQUEUE(wait, current);
2716 __add_wait_queue_tail_exclusive(&x->wait, &wait);
2718 if (signal_pending_state(state, current)) {
2719 timeout = -ERESTARTSYS;
2722 __set_current_state(state);
2723 spin_unlock_irq(&x->wait.lock);
2724 timeout = action(timeout);
2725 spin_lock_irq(&x->wait.lock);
2726 } while (!x->done && timeout);
2727 __remove_wait_queue(&x->wait, &wait);
2732 return timeout ?: 1;
2735 static inline long __sched
2736 __wait_for_common(struct completion *x,
2737 long (*action)(long), long timeout, int state)
2741 spin_lock_irq(&x->wait.lock);
2742 timeout = do_wait_for_common(x, action, timeout, state);
2743 spin_unlock_irq(&x->wait.lock);
2748 wait_for_common(struct completion *x, long timeout, int state)
2750 return __wait_for_common(x, schedule_timeout, timeout, state);
2754 wait_for_common_io(struct completion *x, long timeout, int state)
2756 return __wait_for_common(x, io_schedule_timeout, timeout, state);
2760 * wait_for_completion: - waits for completion of a task
2761 * @x: holds the state of this particular completion
2763 * This waits to be signaled for completion of a specific task. It is NOT
2764 * interruptible and there is no timeout.
2766 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
2767 * and interrupt capability. Also see complete().
2769 void __sched wait_for_completion(struct completion *x)
2771 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2773 EXPORT_SYMBOL(wait_for_completion);
2776 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
2777 * @x: holds the state of this particular completion
2778 * @timeout: timeout value in jiffies
2780 * This waits for either a completion of a specific task to be signaled or for a
2781 * specified timeout to expire. The timeout is in jiffies. It is not
2784 * The return value is 0 if timed out, and positive (at least 1, or number of
2785 * jiffies left till timeout) if completed.
2787 unsigned long __sched
2788 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
2790 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
2792 EXPORT_SYMBOL(wait_for_completion_timeout);
2795 * wait_for_completion_io: - waits for completion of a task
2796 * @x: holds the state of this particular completion
2798 * This waits to be signaled for completion of a specific task. It is NOT
2799 * interruptible and there is no timeout. The caller is accounted as waiting
2802 void __sched wait_for_completion_io(struct completion *x)
2804 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
2806 EXPORT_SYMBOL(wait_for_completion_io);
2809 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
2810 * @x: holds the state of this particular completion
2811 * @timeout: timeout value in jiffies
2813 * This waits for either a completion of a specific task to be signaled or for a
2814 * specified timeout to expire. The timeout is in jiffies. It is not
2815 * interruptible. The caller is accounted as waiting for IO.
2817 * The return value is 0 if timed out, and positive (at least 1, or number of
2818 * jiffies left till timeout) if completed.
2820 unsigned long __sched
2821 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
2823 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
2825 EXPORT_SYMBOL(wait_for_completion_io_timeout);
2828 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
2829 * @x: holds the state of this particular completion
2831 * This waits for completion of a specific task to be signaled. It is
2834 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2836 int __sched wait_for_completion_interruptible(struct completion *x)
2838 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
2839 if (t == -ERESTARTSYS)
2843 EXPORT_SYMBOL(wait_for_completion_interruptible);
2846 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
2847 * @x: holds the state of this particular completion
2848 * @timeout: timeout value in jiffies
2850 * This waits for either a completion of a specific task to be signaled or for a
2851 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
2853 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2854 * positive (at least 1, or number of jiffies left till timeout) if completed.
2857 wait_for_completion_interruptible_timeout(struct completion *x,
2858 unsigned long timeout)
2860 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
2862 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
2865 * wait_for_completion_killable: - waits for completion of a task (killable)
2866 * @x: holds the state of this particular completion
2868 * This waits to be signaled for completion of a specific task. It can be
2869 * interrupted by a kill signal.
2871 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
2873 int __sched wait_for_completion_killable(struct completion *x)
2875 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
2876 if (t == -ERESTARTSYS)
2880 EXPORT_SYMBOL(wait_for_completion_killable);
2883 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
2884 * @x: holds the state of this particular completion
2885 * @timeout: timeout value in jiffies
2887 * This waits for either a completion of a specific task to be
2888 * signaled or for a specified timeout to expire. It can be
2889 * interrupted by a kill signal. The timeout is in jiffies.
2891 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
2892 * positive (at least 1, or number of jiffies left till timeout) if completed.
2895 wait_for_completion_killable_timeout(struct completion *x,
2896 unsigned long timeout)
2898 return wait_for_common(x, timeout, TASK_KILLABLE);
2900 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
2903 * try_wait_for_completion - try to decrement a completion without blocking
2904 * @x: completion structure
2906 * Returns: 0 if a decrement cannot be done without blocking
2907 * 1 if a decrement succeeded.
2909 * If a completion is being used as a counting completion,
2910 * attempt to decrement the counter without blocking. This
2911 * enables us to avoid waiting if the resource the completion
2912 * is protecting is not available.
2914 bool try_wait_for_completion(struct completion *x)
2916 unsigned long flags;
2919 spin_lock_irqsave(&x->wait.lock, flags);
2924 spin_unlock_irqrestore(&x->wait.lock, flags);
2927 EXPORT_SYMBOL(try_wait_for_completion);
2930 * completion_done - Test to see if a completion has any waiters
2931 * @x: completion structure
2933 * Returns: 0 if there are waiters (wait_for_completion() in progress)
2934 * 1 if there are no waiters.
2937 bool completion_done(struct completion *x)
2939 unsigned long flags;
2942 spin_lock_irqsave(&x->wait.lock, flags);
2945 spin_unlock_irqrestore(&x->wait.lock, flags);
2948 EXPORT_SYMBOL(completion_done);
2951 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2953 unsigned long flags;
2956 init_waitqueue_entry(&wait, current);
2958 __set_current_state(state);
2960 spin_lock_irqsave(&q->lock, flags);
2961 __add_wait_queue(q, &wait);
2962 spin_unlock(&q->lock);
2963 timeout = schedule_timeout(timeout);
2964 spin_lock_irq(&q->lock);
2965 __remove_wait_queue(q, &wait);
2966 spin_unlock_irqrestore(&q->lock, flags);
2971 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2973 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2975 EXPORT_SYMBOL(interruptible_sleep_on);
2978 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2980 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2982 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2984 void __sched sleep_on(wait_queue_head_t *q)
2986 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2988 EXPORT_SYMBOL(sleep_on);
2990 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2992 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2994 EXPORT_SYMBOL(sleep_on_timeout);
2996 #ifdef CONFIG_RT_MUTEXES
2999 * rt_mutex_setprio - set the current priority of a task
3001 * @prio: prio value (kernel-internal form)
3003 * This function changes the 'effective' priority of a task. It does
3004 * not touch ->normal_prio like __setscheduler().
3006 * Used by the rt_mutex code to implement priority inheritance logic.
3008 void rt_mutex_setprio(struct task_struct *p, int prio)
3010 int oldprio, on_rq, running;
3012 const struct sched_class *prev_class;
3014 BUG_ON(prio < 0 || prio > MAX_PRIO);
3016 rq = __task_rq_lock(p);
3019 * Idle task boosting is a nono in general. There is one
3020 * exception, when PREEMPT_RT and NOHZ is active:
3022 * The idle task calls get_next_timer_interrupt() and holds
3023 * the timer wheel base->lock on the CPU and another CPU wants
3024 * to access the timer (probably to cancel it). We can safely
3025 * ignore the boosting request, as the idle CPU runs this code
3026 * with interrupts disabled and will complete the lock
3027 * protected section without being interrupted. So there is no
3028 * real need to boost.
3030 if (unlikely(p == rq->idle)) {
3031 WARN_ON(p != rq->curr);
3032 WARN_ON(p->pi_blocked_on);
3036 trace_sched_pi_setprio(p, prio);
3038 prev_class = p->sched_class;
3040 running = task_current(rq, p);
3042 dequeue_task(rq, p, 0);
3044 p->sched_class->put_prev_task(rq, p);
3047 p->sched_class = &rt_sched_class;
3049 p->sched_class = &fair_sched_class;
3054 p->sched_class->set_curr_task(rq);
3056 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3058 check_class_changed(rq, p, prev_class, oldprio);
3060 __task_rq_unlock(rq);
3063 void set_user_nice(struct task_struct *p, long nice)
3065 int old_prio, delta, on_rq;
3066 unsigned long flags;
3069 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3072 * We have to be careful, if called from sys_setpriority(),
3073 * the task might be in the middle of scheduling on another CPU.
3075 rq = task_rq_lock(p, &flags);
3077 * The RT priorities are set via sched_setscheduler(), but we still
3078 * allow the 'normal' nice value to be set - but as expected
3079 * it wont have any effect on scheduling until the task is
3080 * SCHED_FIFO/SCHED_RR:
3082 if (task_has_rt_policy(p)) {
3083 p->static_prio = NICE_TO_PRIO(nice);
3088 dequeue_task(rq, p, 0);
3090 p->static_prio = NICE_TO_PRIO(nice);
3093 p->prio = effective_prio(p);
3094 delta = p->prio - old_prio;
3097 enqueue_task(rq, p, 0);
3099 * If the task increased its priority or is running and
3100 * lowered its priority, then reschedule its CPU:
3102 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3103 resched_task(rq->curr);
3106 task_rq_unlock(rq, p, &flags);
3108 EXPORT_SYMBOL(set_user_nice);
3111 * can_nice - check if a task can reduce its nice value
3115 int can_nice(const struct task_struct *p, const int nice)
3117 /* convert nice value [19,-20] to rlimit style value [1,40] */
3118 int nice_rlim = 20 - nice;
3120 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3121 capable(CAP_SYS_NICE));
3124 #ifdef __ARCH_WANT_SYS_NICE
3127 * sys_nice - change the priority of the current process.
3128 * @increment: priority increment
3130 * sys_setpriority is a more generic, but much slower function that
3131 * does similar things.
3133 SYSCALL_DEFINE1(nice, int, increment)
3138 * Setpriority might change our priority at the same moment.
3139 * We don't have to worry. Conceptually one call occurs first
3140 * and we have a single winner.
3142 if (increment < -40)
3147 nice = TASK_NICE(current) + increment;
3153 if (increment < 0 && !can_nice(current, nice))
3156 retval = security_task_setnice(current, nice);
3160 set_user_nice(current, nice);
3167 * task_prio - return the priority value of a given task.
3168 * @p: the task in question.
3170 * This is the priority value as seen by users in /proc.
3171 * RT tasks are offset by -200. Normal tasks are centered
3172 * around 0, value goes from -16 to +15.
3174 int task_prio(const struct task_struct *p)
3176 return p->prio - MAX_RT_PRIO;
3180 * task_nice - return the nice value of a given task.
3181 * @p: the task in question.
3183 int task_nice(const struct task_struct *p)
3185 return TASK_NICE(p);
3187 EXPORT_SYMBOL(task_nice);
3190 * idle_cpu - is a given cpu idle currently?
3191 * @cpu: the processor in question.
3193 int idle_cpu(int cpu)
3195 struct rq *rq = cpu_rq(cpu);
3197 if (rq->curr != rq->idle)
3204 if (!llist_empty(&rq->wake_list))
3212 * idle_task - return the idle task for a given cpu.
3213 * @cpu: the processor in question.
3215 struct task_struct *idle_task(int cpu)
3217 return cpu_rq(cpu)->idle;
3221 * find_process_by_pid - find a process with a matching PID value.
3222 * @pid: the pid in question.
3224 static struct task_struct *find_process_by_pid(pid_t pid)
3226 return pid ? find_task_by_vpid(pid) : current;
3229 /* Actually do priority change: must hold rq lock. */
3231 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3234 p->rt_priority = prio;
3235 p->normal_prio = normal_prio(p);
3236 /* we are holding p->pi_lock already */
3237 p->prio = rt_mutex_getprio(p);
3238 if (rt_prio(p->prio))
3239 p->sched_class = &rt_sched_class;
3241 p->sched_class = &fair_sched_class;
3246 * check the target process has a UID that matches the current process's
3248 static bool check_same_owner(struct task_struct *p)
3250 const struct cred *cred = current_cred(), *pcred;
3254 pcred = __task_cred(p);
3255 match = (uid_eq(cred->euid, pcred->euid) ||
3256 uid_eq(cred->euid, pcred->uid));
3261 static int __sched_setscheduler(struct task_struct *p, int policy,
3262 const struct sched_param *param, bool user)
3264 int retval, oldprio, oldpolicy = -1, on_rq, running;
3265 unsigned long flags;
3266 const struct sched_class *prev_class;
3270 /* may grab non-irq protected spin_locks */
3271 BUG_ON(in_interrupt());
3273 /* double check policy once rq lock held */
3275 reset_on_fork = p->sched_reset_on_fork;
3276 policy = oldpolicy = p->policy;
3278 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3279 policy &= ~SCHED_RESET_ON_FORK;
3281 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3282 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3283 policy != SCHED_IDLE)
3288 * Valid priorities for SCHED_FIFO and SCHED_RR are
3289 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3290 * SCHED_BATCH and SCHED_IDLE is 0.
3292 if (param->sched_priority < 0 ||
3293 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3294 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3296 if (rt_policy(policy) != (param->sched_priority != 0))
3300 * Allow unprivileged RT tasks to decrease priority:
3302 if (user && !capable(CAP_SYS_NICE)) {
3303 if (rt_policy(policy)) {
3304 unsigned long rlim_rtprio =
3305 task_rlimit(p, RLIMIT_RTPRIO);
3307 /* can't set/change the rt policy */
3308 if (policy != p->policy && !rlim_rtprio)
3311 /* can't increase priority */
3312 if (param->sched_priority > p->rt_priority &&
3313 param->sched_priority > rlim_rtprio)
3318 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3319 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3321 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3322 if (!can_nice(p, TASK_NICE(p)))
3326 /* can't change other user's priorities */
3327 if (!check_same_owner(p))
3330 /* Normal users shall not reset the sched_reset_on_fork flag */
3331 if (p->sched_reset_on_fork && !reset_on_fork)
3336 retval = security_task_setscheduler(p);
3342 * make sure no PI-waiters arrive (or leave) while we are
3343 * changing the priority of the task:
3345 * To be able to change p->policy safely, the appropriate
3346 * runqueue lock must be held.
3348 rq = task_rq_lock(p, &flags);
3351 * Changing the policy of the stop threads its a very bad idea
3353 if (p == rq->stop) {
3354 task_rq_unlock(rq, p, &flags);
3359 * If not changing anything there's no need to proceed further:
3361 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3362 param->sched_priority == p->rt_priority))) {
3363 task_rq_unlock(rq, p, &flags);
3367 #ifdef CONFIG_RT_GROUP_SCHED
3370 * Do not allow realtime tasks into groups that have no runtime
3373 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3374 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3375 !task_group_is_autogroup(task_group(p))) {
3376 task_rq_unlock(rq, p, &flags);
3382 /* recheck policy now with rq lock held */
3383 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3384 policy = oldpolicy = -1;
3385 task_rq_unlock(rq, p, &flags);
3389 running = task_current(rq, p);
3391 dequeue_task(rq, p, 0);
3393 p->sched_class->put_prev_task(rq, p);
3395 p->sched_reset_on_fork = reset_on_fork;
3398 prev_class = p->sched_class;
3399 __setscheduler(rq, p, policy, param->sched_priority);
3402 p->sched_class->set_curr_task(rq);
3404 enqueue_task(rq, p, 0);
3406 check_class_changed(rq, p, prev_class, oldprio);
3407 task_rq_unlock(rq, p, &flags);
3409 rt_mutex_adjust_pi(p);
3415 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3416 * @p: the task in question.
3417 * @policy: new policy.
3418 * @param: structure containing the new RT priority.
3420 * NOTE that the task may be already dead.
3422 int sched_setscheduler(struct task_struct *p, int policy,
3423 const struct sched_param *param)
3425 return __sched_setscheduler(p, policy, param, true);
3427 EXPORT_SYMBOL_GPL(sched_setscheduler);
3430 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3431 * @p: the task in question.
3432 * @policy: new policy.
3433 * @param: structure containing the new RT priority.
3435 * Just like sched_setscheduler, only don't bother checking if the
3436 * current context has permission. For example, this is needed in
3437 * stop_machine(): we create temporary high priority worker threads,
3438 * but our caller might not have that capability.
3440 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3441 const struct sched_param *param)
3443 return __sched_setscheduler(p, policy, param, false);
3447 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3449 struct sched_param lparam;
3450 struct task_struct *p;
3453 if (!param || pid < 0)
3455 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3460 p = find_process_by_pid(pid);
3462 retval = sched_setscheduler(p, policy, &lparam);
3469 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3470 * @pid: the pid in question.
3471 * @policy: new policy.
3472 * @param: structure containing the new RT priority.
3474 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3475 struct sched_param __user *, param)
3477 /* negative values for policy are not valid */
3481 return do_sched_setscheduler(pid, policy, param);
3485 * sys_sched_setparam - set/change the RT priority of a thread
3486 * @pid: the pid in question.
3487 * @param: structure containing the new RT priority.
3489 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3491 return do_sched_setscheduler(pid, -1, param);
3495 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3496 * @pid: the pid in question.
3498 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3500 struct task_struct *p;
3508 p = find_process_by_pid(pid);
3510 retval = security_task_getscheduler(p);
3513 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3520 * sys_sched_getparam - get the RT priority of a thread
3521 * @pid: the pid in question.
3522 * @param: structure containing the RT priority.
3524 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3526 struct sched_param lp;
3527 struct task_struct *p;
3530 if (!param || pid < 0)
3534 p = find_process_by_pid(pid);
3539 retval = security_task_getscheduler(p);
3543 lp.sched_priority = p->rt_priority;
3547 * This one might sleep, we cannot do it with a spinlock held ...
3549 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3558 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3560 cpumask_var_t cpus_allowed, new_mask;
3561 struct task_struct *p;
3567 p = find_process_by_pid(pid);
3574 /* Prevent p going away */
3578 if (p->flags & PF_NO_SETAFFINITY) {
3582 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3586 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3588 goto out_free_cpus_allowed;
3591 if (!check_same_owner(p)) {
3593 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3600 retval = security_task_setscheduler(p);
3604 cpuset_cpus_allowed(p, cpus_allowed);
3605 cpumask_and(new_mask, in_mask, cpus_allowed);
3607 retval = set_cpus_allowed_ptr(p, new_mask);
3610 cpuset_cpus_allowed(p, cpus_allowed);
3611 if (!cpumask_subset(new_mask, cpus_allowed)) {
3613 * We must have raced with a concurrent cpuset
3614 * update. Just reset the cpus_allowed to the
3615 * cpuset's cpus_allowed
3617 cpumask_copy(new_mask, cpus_allowed);
3622 free_cpumask_var(new_mask);
3623 out_free_cpus_allowed:
3624 free_cpumask_var(cpus_allowed);
3631 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3632 struct cpumask *new_mask)
3634 if (len < cpumask_size())
3635 cpumask_clear(new_mask);
3636 else if (len > cpumask_size())
3637 len = cpumask_size();
3639 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3643 * sys_sched_setaffinity - set the cpu affinity of a process
3644 * @pid: pid of the process
3645 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3646 * @user_mask_ptr: user-space pointer to the new cpu mask
3648 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3649 unsigned long __user *, user_mask_ptr)
3651 cpumask_var_t new_mask;
3654 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3657 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3659 retval = sched_setaffinity(pid, new_mask);
3660 free_cpumask_var(new_mask);
3664 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3666 struct task_struct *p;
3667 unsigned long flags;
3674 p = find_process_by_pid(pid);
3678 retval = security_task_getscheduler(p);
3682 raw_spin_lock_irqsave(&p->pi_lock, flags);
3683 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
3684 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3694 * sys_sched_getaffinity - get the cpu affinity of a process
3695 * @pid: pid of the process
3696 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3697 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3699 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3700 unsigned long __user *, user_mask_ptr)
3705 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3707 if (len & (sizeof(unsigned long)-1))
3710 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3713 ret = sched_getaffinity(pid, mask);
3715 size_t retlen = min_t(size_t, len, cpumask_size());
3717 if (copy_to_user(user_mask_ptr, mask, retlen))
3722 free_cpumask_var(mask);
3728 * sys_sched_yield - yield the current processor to other threads.
3730 * This function yields the current CPU to other tasks. If there are no
3731 * other threads running on this CPU then this function will return.
3733 SYSCALL_DEFINE0(sched_yield)
3735 struct rq *rq = this_rq_lock();
3737 schedstat_inc(rq, yld_count);
3738 current->sched_class->yield_task(rq);
3741 * Since we are going to call schedule() anyway, there's
3742 * no need to preempt or enable interrupts:
3744 __release(rq->lock);
3745 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3746 do_raw_spin_unlock(&rq->lock);
3747 sched_preempt_enable_no_resched();
3754 static inline int should_resched(void)
3756 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
3759 static void __cond_resched(void)
3761 add_preempt_count(PREEMPT_ACTIVE);
3763 sub_preempt_count(PREEMPT_ACTIVE);
3766 int __sched _cond_resched(void)
3768 if (should_resched()) {
3774 EXPORT_SYMBOL(_cond_resched);
3777 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
3778 * call schedule, and on return reacquire the lock.
3780 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
3781 * operations here to prevent schedule() from being called twice (once via
3782 * spin_unlock(), once by hand).
3784 int __cond_resched_lock(spinlock_t *lock)
3786 int resched = should_resched();
3789 lockdep_assert_held(lock);
3791 if (spin_needbreak(lock) || resched) {
3802 EXPORT_SYMBOL(__cond_resched_lock);
3804 int __sched __cond_resched_softirq(void)
3806 BUG_ON(!in_softirq());
3808 if (should_resched()) {
3816 EXPORT_SYMBOL(__cond_resched_softirq);
3819 * yield - yield the current processor to other threads.
3821 * Do not ever use this function, there's a 99% chance you're doing it wrong.
3823 * The scheduler is at all times free to pick the calling task as the most
3824 * eligible task to run, if removing the yield() call from your code breaks
3825 * it, its already broken.
3827 * Typical broken usage is:
3832 * where one assumes that yield() will let 'the other' process run that will
3833 * make event true. If the current task is a SCHED_FIFO task that will never
3834 * happen. Never use yield() as a progress guarantee!!
3836 * If you want to use yield() to wait for something, use wait_event().
3837 * If you want to use yield() to be 'nice' for others, use cond_resched().
3838 * If you still want to use yield(), do not!
3840 void __sched yield(void)
3842 set_current_state(TASK_RUNNING);
3845 EXPORT_SYMBOL(yield);
3848 * yield_to - yield the current processor to another thread in
3849 * your thread group, or accelerate that thread toward the
3850 * processor it's on.
3852 * @preempt: whether task preemption is allowed or not
3854 * It's the caller's job to ensure that the target task struct
3855 * can't go away on us before we can do any checks.
3858 * true (>0) if we indeed boosted the target task.
3859 * false (0) if we failed to boost the target.
3860 * -ESRCH if there's no task to yield to.
3862 bool __sched yield_to(struct task_struct *p, bool preempt)
3864 struct task_struct *curr = current;
3865 struct rq *rq, *p_rq;
3866 unsigned long flags;
3869 local_irq_save(flags);
3875 * If we're the only runnable task on the rq and target rq also
3876 * has only one task, there's absolutely no point in yielding.
3878 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
3883 double_rq_lock(rq, p_rq);
3884 while (task_rq(p) != p_rq) {
3885 double_rq_unlock(rq, p_rq);
3889 if (!curr->sched_class->yield_to_task)
3892 if (curr->sched_class != p->sched_class)
3895 if (task_running(p_rq, p) || p->state)
3898 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
3900 schedstat_inc(rq, yld_count);
3902 * Make p's CPU reschedule; pick_next_entity takes care of
3905 if (preempt && rq != p_rq)
3906 resched_task(p_rq->curr);
3910 double_rq_unlock(rq, p_rq);
3912 local_irq_restore(flags);
3919 EXPORT_SYMBOL_GPL(yield_to);
3922 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
3923 * that process accounting knows that this is a task in IO wait state.
3925 void __sched io_schedule(void)
3927 struct rq *rq = raw_rq();
3929 delayacct_blkio_start();
3930 atomic_inc(&rq->nr_iowait);
3931 blk_flush_plug(current);
3932 current->in_iowait = 1;
3934 current->in_iowait = 0;
3935 atomic_dec(&rq->nr_iowait);
3936 delayacct_blkio_end();
3938 EXPORT_SYMBOL(io_schedule);
3940 long __sched io_schedule_timeout(long timeout)
3942 struct rq *rq = raw_rq();
3945 delayacct_blkio_start();
3946 atomic_inc(&rq->nr_iowait);
3947 blk_flush_plug(current);
3948 current->in_iowait = 1;
3949 ret = schedule_timeout(timeout);
3950 current->in_iowait = 0;
3951 atomic_dec(&rq->nr_iowait);
3952 delayacct_blkio_end();
3957 * sys_sched_get_priority_max - return maximum RT priority.
3958 * @policy: scheduling class.
3960 * this syscall returns the maximum rt_priority that can be used
3961 * by a given scheduling class.
3963 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
3970 ret = MAX_USER_RT_PRIO-1;
3982 * sys_sched_get_priority_min - return minimum RT priority.
3983 * @policy: scheduling class.
3985 * this syscall returns the minimum rt_priority that can be used
3986 * by a given scheduling class.
3988 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4006 * sys_sched_rr_get_interval - return the default timeslice of a process.
4007 * @pid: pid of the process.
4008 * @interval: userspace pointer to the timeslice value.
4010 * this syscall writes the default timeslice value of a given process
4011 * into the user-space timespec buffer. A value of '0' means infinity.
4013 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4014 struct timespec __user *, interval)
4016 struct task_struct *p;
4017 unsigned int time_slice;
4018 unsigned long flags;
4028 p = find_process_by_pid(pid);
4032 retval = security_task_getscheduler(p);
4036 rq = task_rq_lock(p, &flags);
4037 time_slice = p->sched_class->get_rr_interval(rq, p);
4038 task_rq_unlock(rq, p, &flags);
4041 jiffies_to_timespec(time_slice, &t);
4042 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4050 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4052 void sched_show_task(struct task_struct *p)
4054 unsigned long free = 0;
4058 state = p->state ? __ffs(p->state) + 1 : 0;
4059 printk(KERN_INFO "%-15.15s %c", p->comm,
4060 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4061 #if BITS_PER_LONG == 32
4062 if (state == TASK_RUNNING)
4063 printk(KERN_CONT " running ");
4065 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4067 if (state == TASK_RUNNING)
4068 printk(KERN_CONT " running task ");
4070 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4072 #ifdef CONFIG_DEBUG_STACK_USAGE
4073 free = stack_not_used(p);
4076 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4078 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4079 task_pid_nr(p), ppid,
4080 (unsigned long)task_thread_info(p)->flags);
4082 print_worker_info(KERN_INFO, p);
4083 show_stack(p, NULL);
4086 void show_state_filter(unsigned long state_filter)
4088 struct task_struct *g, *p;
4090 #if BITS_PER_LONG == 32
4092 " task PC stack pid father\n");
4095 " task PC stack pid father\n");
4098 do_each_thread(g, p) {
4100 * reset the NMI-timeout, listing all files on a slow
4101 * console might take a lot of time:
4103 touch_nmi_watchdog();
4104 if (!state_filter || (p->state & state_filter))
4106 } while_each_thread(g, p);
4108 touch_all_softlockup_watchdogs();
4110 #ifdef CONFIG_SCHED_DEBUG
4111 sysrq_sched_debug_show();
4115 * Only show locks if all tasks are dumped:
4118 debug_show_all_locks();
4121 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4123 idle->sched_class = &idle_sched_class;
4127 * init_idle - set up an idle thread for a given CPU
4128 * @idle: task in question
4129 * @cpu: cpu the idle task belongs to
4131 * NOTE: this function does not set the idle thread's NEED_RESCHED
4132 * flag, to make booting more robust.
4134 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4136 struct rq *rq = cpu_rq(cpu);
4137 unsigned long flags;
4139 raw_spin_lock_irqsave(&rq->lock, flags);
4142 idle->state = TASK_RUNNING;
4143 idle->se.exec_start = sched_clock();
4145 do_set_cpus_allowed(idle, cpumask_of(cpu));
4147 * We're having a chicken and egg problem, even though we are
4148 * holding rq->lock, the cpu isn't yet set to this cpu so the
4149 * lockdep check in task_group() will fail.
4151 * Similar case to sched_fork(). / Alternatively we could
4152 * use task_rq_lock() here and obtain the other rq->lock.
4157 __set_task_cpu(idle, cpu);
4160 rq->curr = rq->idle = idle;
4161 #if defined(CONFIG_SMP)
4164 raw_spin_unlock_irqrestore(&rq->lock, flags);
4166 /* Set the preempt count _outside_ the spinlocks! */
4167 task_thread_info(idle)->preempt_count = 0;
4170 * The idle tasks have their own, simple scheduling class:
4172 idle->sched_class = &idle_sched_class;
4173 ftrace_graph_init_idle_task(idle, cpu);
4174 vtime_init_idle(idle, cpu);
4175 #if defined(CONFIG_SMP)
4176 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4181 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4183 if (p->sched_class && p->sched_class->set_cpus_allowed)
4184 p->sched_class->set_cpus_allowed(p, new_mask);
4186 cpumask_copy(&p->cpus_allowed, new_mask);
4187 p->nr_cpus_allowed = cpumask_weight(new_mask);
4191 * This is how migration works:
4193 * 1) we invoke migration_cpu_stop() on the target CPU using
4195 * 2) stopper starts to run (implicitly forcing the migrated thread
4197 * 3) it checks whether the migrated task is still in the wrong runqueue.
4198 * 4) if it's in the wrong runqueue then the migration thread removes
4199 * it and puts it into the right queue.
4200 * 5) stopper completes and stop_one_cpu() returns and the migration
4205 * Change a given task's CPU affinity. Migrate the thread to a
4206 * proper CPU and schedule it away if the CPU it's executing on
4207 * is removed from the allowed bitmask.
4209 * NOTE: the caller must have a valid reference to the task, the
4210 * task must not exit() & deallocate itself prematurely. The
4211 * call is not atomic; no spinlocks may be held.
4213 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4215 unsigned long flags;
4217 unsigned int dest_cpu;
4220 rq = task_rq_lock(p, &flags);
4222 if (cpumask_equal(&p->cpus_allowed, new_mask))
4225 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4230 do_set_cpus_allowed(p, new_mask);
4232 /* Can the task run on the task's current CPU? If so, we're done */
4233 if (cpumask_test_cpu(task_cpu(p), new_mask))
4236 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4238 struct migration_arg arg = { p, dest_cpu };
4239 /* Need help from migration thread: drop lock and wait. */
4240 task_rq_unlock(rq, p, &flags);
4241 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4242 tlb_migrate_finish(p->mm);
4246 task_rq_unlock(rq, p, &flags);
4250 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4253 * Move (not current) task off this cpu, onto dest cpu. We're doing
4254 * this because either it can't run here any more (set_cpus_allowed()
4255 * away from this CPU, or CPU going down), or because we're
4256 * attempting to rebalance this task on exec (sched_exec).
4258 * So we race with normal scheduler movements, but that's OK, as long
4259 * as the task is no longer on this CPU.
4261 * Returns non-zero if task was successfully migrated.
4263 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4265 struct rq *rq_dest, *rq_src;
4268 if (unlikely(!cpu_active(dest_cpu)))
4271 rq_src = cpu_rq(src_cpu);
4272 rq_dest = cpu_rq(dest_cpu);
4274 raw_spin_lock(&p->pi_lock);
4275 double_rq_lock(rq_src, rq_dest);
4276 /* Already moved. */
4277 if (task_cpu(p) != src_cpu)
4279 /* Affinity changed (again). */
4280 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4284 * If we're not on a rq, the next wake-up will ensure we're
4288 dequeue_task(rq_src, p, 0);
4289 set_task_cpu(p, dest_cpu);
4290 enqueue_task(rq_dest, p, 0);
4291 check_preempt_curr(rq_dest, p, 0);
4296 double_rq_unlock(rq_src, rq_dest);
4297 raw_spin_unlock(&p->pi_lock);
4302 * migration_cpu_stop - this will be executed by a highprio stopper thread
4303 * and performs thread migration by bumping thread off CPU then
4304 * 'pushing' onto another runqueue.
4306 static int migration_cpu_stop(void *data)
4308 struct migration_arg *arg = data;
4311 * The original target cpu might have gone down and we might
4312 * be on another cpu but it doesn't matter.
4314 local_irq_disable();
4315 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4320 #ifdef CONFIG_HOTPLUG_CPU
4323 * Ensures that the idle task is using init_mm right before its cpu goes
4326 void idle_task_exit(void)
4328 struct mm_struct *mm = current->active_mm;
4330 BUG_ON(cpu_online(smp_processor_id()));
4333 switch_mm(mm, &init_mm, current);
4338 * Since this CPU is going 'away' for a while, fold any nr_active delta
4339 * we might have. Assumes we're called after migrate_tasks() so that the
4340 * nr_active count is stable.
4342 * Also see the comment "Global load-average calculations".
4344 static void calc_load_migrate(struct rq *rq)
4346 long delta = calc_load_fold_active(rq);
4348 atomic_long_add(delta, &calc_load_tasks);
4352 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4353 * try_to_wake_up()->select_task_rq().
4355 * Called with rq->lock held even though we'er in stop_machine() and
4356 * there's no concurrency possible, we hold the required locks anyway
4357 * because of lock validation efforts.
4359 static void migrate_tasks(unsigned int dead_cpu)
4361 struct rq *rq = cpu_rq(dead_cpu);
4362 struct task_struct *next, *stop = rq->stop;
4366 * Fudge the rq selection such that the below task selection loop
4367 * doesn't get stuck on the currently eligible stop task.
4369 * We're currently inside stop_machine() and the rq is either stuck
4370 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4371 * either way we should never end up calling schedule() until we're
4377 * put_prev_task() and pick_next_task() sched
4378 * class method both need to have an up-to-date
4379 * value of rq->clock[_task]
4381 update_rq_clock(rq);
4385 * There's this thread running, bail when that's the only
4388 if (rq->nr_running == 1)
4391 next = pick_next_task(rq);
4393 next->sched_class->put_prev_task(rq, next);
4395 /* Find suitable destination for @next, with force if needed. */
4396 dest_cpu = select_fallback_rq(dead_cpu, next);
4397 raw_spin_unlock(&rq->lock);
4399 __migrate_task(next, dead_cpu, dest_cpu);
4401 raw_spin_lock(&rq->lock);
4407 #endif /* CONFIG_HOTPLUG_CPU */
4409 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4411 static struct ctl_table sd_ctl_dir[] = {
4413 .procname = "sched_domain",
4419 static struct ctl_table sd_ctl_root[] = {
4421 .procname = "kernel",
4423 .child = sd_ctl_dir,
4428 static struct ctl_table *sd_alloc_ctl_entry(int n)
4430 struct ctl_table *entry =
4431 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4436 static void sd_free_ctl_entry(struct ctl_table **tablep)
4438 struct ctl_table *entry;
4441 * In the intermediate directories, both the child directory and
4442 * procname are dynamically allocated and could fail but the mode
4443 * will always be set. In the lowest directory the names are
4444 * static strings and all have proc handlers.
4446 for (entry = *tablep; entry->mode; entry++) {
4448 sd_free_ctl_entry(&entry->child);
4449 if (entry->proc_handler == NULL)
4450 kfree(entry->procname);
4457 static int min_load_idx = 0;
4458 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4461 set_table_entry(struct ctl_table *entry,
4462 const char *procname, void *data, int maxlen,
4463 umode_t mode, proc_handler *proc_handler,
4466 entry->procname = procname;
4468 entry->maxlen = maxlen;
4470 entry->proc_handler = proc_handler;
4473 entry->extra1 = &min_load_idx;
4474 entry->extra2 = &max_load_idx;
4478 static struct ctl_table *
4479 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4481 struct ctl_table *table = sd_alloc_ctl_entry(13);
4486 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4487 sizeof(long), 0644, proc_doulongvec_minmax, false);
4488 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4489 sizeof(long), 0644, proc_doulongvec_minmax, false);
4490 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4491 sizeof(int), 0644, proc_dointvec_minmax, true);
4492 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4493 sizeof(int), 0644, proc_dointvec_minmax, true);
4494 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4495 sizeof(int), 0644, proc_dointvec_minmax, true);
4496 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4497 sizeof(int), 0644, proc_dointvec_minmax, true);
4498 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4499 sizeof(int), 0644, proc_dointvec_minmax, true);
4500 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4501 sizeof(int), 0644, proc_dointvec_minmax, false);
4502 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4503 sizeof(int), 0644, proc_dointvec_minmax, false);
4504 set_table_entry(&table[9], "cache_nice_tries",
4505 &sd->cache_nice_tries,
4506 sizeof(int), 0644, proc_dointvec_minmax, false);
4507 set_table_entry(&table[10], "flags", &sd->flags,
4508 sizeof(int), 0644, proc_dointvec_minmax, false);
4509 set_table_entry(&table[11], "name", sd->name,
4510 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4511 /* &table[12] is terminator */
4516 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4518 struct ctl_table *entry, *table;
4519 struct sched_domain *sd;
4520 int domain_num = 0, i;
4523 for_each_domain(cpu, sd)
4525 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4530 for_each_domain(cpu, sd) {
4531 snprintf(buf, 32, "domain%d", i);
4532 entry->procname = kstrdup(buf, GFP_KERNEL);
4534 entry->child = sd_alloc_ctl_domain_table(sd);
4541 static struct ctl_table_header *sd_sysctl_header;
4542 static void register_sched_domain_sysctl(void)
4544 int i, cpu_num = num_possible_cpus();
4545 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4548 WARN_ON(sd_ctl_dir[0].child);
4549 sd_ctl_dir[0].child = entry;
4554 for_each_possible_cpu(i) {
4555 snprintf(buf, 32, "cpu%d", i);
4556 entry->procname = kstrdup(buf, GFP_KERNEL);
4558 entry->child = sd_alloc_ctl_cpu_table(i);
4562 WARN_ON(sd_sysctl_header);
4563 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4566 /* may be called multiple times per register */
4567 static void unregister_sched_domain_sysctl(void)
4569 if (sd_sysctl_header)
4570 unregister_sysctl_table(sd_sysctl_header);
4571 sd_sysctl_header = NULL;
4572 if (sd_ctl_dir[0].child)
4573 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4576 static void register_sched_domain_sysctl(void)
4579 static void unregister_sched_domain_sysctl(void)
4584 static void set_rq_online(struct rq *rq)
4587 const struct sched_class *class;
4589 cpumask_set_cpu(rq->cpu, rq->rd->online);
4592 for_each_class(class) {
4593 if (class->rq_online)
4594 class->rq_online(rq);
4599 static void set_rq_offline(struct rq *rq)
4602 const struct sched_class *class;
4604 for_each_class(class) {
4605 if (class->rq_offline)
4606 class->rq_offline(rq);
4609 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4615 * migration_call - callback that gets triggered when a CPU is added.
4616 * Here we can start up the necessary migration thread for the new CPU.
4618 static int __cpuinit
4619 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4621 int cpu = (long)hcpu;
4622 unsigned long flags;
4623 struct rq *rq = cpu_rq(cpu);
4625 switch (action & ~CPU_TASKS_FROZEN) {
4627 case CPU_UP_PREPARE:
4628 rq->calc_load_update = calc_load_update;
4632 /* Update our root-domain */
4633 raw_spin_lock_irqsave(&rq->lock, flags);
4635 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4639 raw_spin_unlock_irqrestore(&rq->lock, flags);
4642 #ifdef CONFIG_HOTPLUG_CPU
4644 sched_ttwu_pending();
4645 /* Update our root-domain */
4646 raw_spin_lock_irqsave(&rq->lock, flags);
4648 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4652 BUG_ON(rq->nr_running != 1); /* the migration thread */
4653 raw_spin_unlock_irqrestore(&rq->lock, flags);
4657 calc_load_migrate(rq);
4662 update_max_interval();
4668 * Register at high priority so that task migration (migrate_all_tasks)
4669 * happens before everything else. This has to be lower priority than
4670 * the notifier in the perf_event subsystem, though.
4672 static struct notifier_block __cpuinitdata migration_notifier = {
4673 .notifier_call = migration_call,
4674 .priority = CPU_PRI_MIGRATION,
4677 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
4678 unsigned long action, void *hcpu)
4680 switch (action & ~CPU_TASKS_FROZEN) {
4682 case CPU_DOWN_FAILED:
4683 set_cpu_active((long)hcpu, true);
4690 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
4691 unsigned long action, void *hcpu)
4693 switch (action & ~CPU_TASKS_FROZEN) {
4694 case CPU_DOWN_PREPARE:
4695 set_cpu_active((long)hcpu, false);
4702 static int __init migration_init(void)
4704 void *cpu = (void *)(long)smp_processor_id();
4707 /* Initialize migration for the boot CPU */
4708 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
4709 BUG_ON(err == NOTIFY_BAD);
4710 migration_call(&migration_notifier, CPU_ONLINE, cpu);
4711 register_cpu_notifier(&migration_notifier);
4713 /* Register cpu active notifiers */
4714 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
4715 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
4719 early_initcall(migration_init);
4724 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
4726 #ifdef CONFIG_SCHED_DEBUG
4728 static __read_mostly int sched_debug_enabled;
4730 static int __init sched_debug_setup(char *str)
4732 sched_debug_enabled = 1;
4736 early_param("sched_debug", sched_debug_setup);
4738 static inline bool sched_debug(void)
4740 return sched_debug_enabled;
4743 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
4744 struct cpumask *groupmask)
4746 struct sched_group *group = sd->groups;
4749 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
4750 cpumask_clear(groupmask);
4752 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
4754 if (!(sd->flags & SD_LOAD_BALANCE)) {
4755 printk("does not load-balance\n");
4757 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
4762 printk(KERN_CONT "span %s level %s\n", str, sd->name);
4764 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
4765 printk(KERN_ERR "ERROR: domain->span does not contain "
4768 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
4769 printk(KERN_ERR "ERROR: domain->groups does not contain"
4773 printk(KERN_DEBUG "%*s groups:", level + 1, "");
4777 printk(KERN_ERR "ERROR: group is NULL\n");
4782 * Even though we initialize ->power to something semi-sane,
4783 * we leave power_orig unset. This allows us to detect if
4784 * domain iteration is still funny without causing /0 traps.
4786 if (!group->sgp->power_orig) {
4787 printk(KERN_CONT "\n");
4788 printk(KERN_ERR "ERROR: domain->cpu_power not "
4793 if (!cpumask_weight(sched_group_cpus(group))) {
4794 printk(KERN_CONT "\n");
4795 printk(KERN_ERR "ERROR: empty group\n");
4799 if (!(sd->flags & SD_OVERLAP) &&
4800 cpumask_intersects(groupmask, sched_group_cpus(group))) {
4801 printk(KERN_CONT "\n");
4802 printk(KERN_ERR "ERROR: repeated CPUs\n");
4806 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
4808 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
4810 printk(KERN_CONT " %s", str);
4811 if (group->sgp->power != SCHED_POWER_SCALE) {
4812 printk(KERN_CONT " (cpu_power = %d)",
4816 group = group->next;
4817 } while (group != sd->groups);
4818 printk(KERN_CONT "\n");
4820 if (!cpumask_equal(sched_domain_span(sd), groupmask))
4821 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
4824 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
4825 printk(KERN_ERR "ERROR: parent span is not a superset "
4826 "of domain->span\n");
4830 static void sched_domain_debug(struct sched_domain *sd, int cpu)
4834 if (!sched_debug_enabled)
4838 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
4842 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
4845 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
4853 #else /* !CONFIG_SCHED_DEBUG */
4854 # define sched_domain_debug(sd, cpu) do { } while (0)
4855 static inline bool sched_debug(void)
4859 #endif /* CONFIG_SCHED_DEBUG */
4861 static int sd_degenerate(struct sched_domain *sd)
4863 if (cpumask_weight(sched_domain_span(sd)) == 1)
4866 /* Following flags need at least 2 groups */
4867 if (sd->flags & (SD_LOAD_BALANCE |
4868 SD_BALANCE_NEWIDLE |
4872 SD_SHARE_PKG_RESOURCES)) {
4873 if (sd->groups != sd->groups->next)
4877 /* Following flags don't use groups */
4878 if (sd->flags & (SD_WAKE_AFFINE))
4885 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
4887 unsigned long cflags = sd->flags, pflags = parent->flags;
4889 if (sd_degenerate(parent))
4892 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
4895 /* Flags needing groups don't count if only 1 group in parent */
4896 if (parent->groups == parent->groups->next) {
4897 pflags &= ~(SD_LOAD_BALANCE |
4898 SD_BALANCE_NEWIDLE |
4902 SD_SHARE_PKG_RESOURCES);
4903 if (nr_node_ids == 1)
4904 pflags &= ~SD_SERIALIZE;
4906 if (~cflags & pflags)
4912 static void free_rootdomain(struct rcu_head *rcu)
4914 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
4916 cpupri_cleanup(&rd->cpupri);
4917 free_cpumask_var(rd->rto_mask);
4918 free_cpumask_var(rd->online);
4919 free_cpumask_var(rd->span);
4923 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
4925 struct root_domain *old_rd = NULL;
4926 unsigned long flags;
4928 raw_spin_lock_irqsave(&rq->lock, flags);
4933 if (cpumask_test_cpu(rq->cpu, old_rd->online))
4936 cpumask_clear_cpu(rq->cpu, old_rd->span);
4939 * If we dont want to free the old_rt yet then
4940 * set old_rd to NULL to skip the freeing later
4943 if (!atomic_dec_and_test(&old_rd->refcount))
4947 atomic_inc(&rd->refcount);
4950 cpumask_set_cpu(rq->cpu, rd->span);
4951 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
4954 raw_spin_unlock_irqrestore(&rq->lock, flags);
4957 call_rcu_sched(&old_rd->rcu, free_rootdomain);
4960 static int init_rootdomain(struct root_domain *rd)
4962 memset(rd, 0, sizeof(*rd));
4964 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
4966 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
4968 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
4971 if (cpupri_init(&rd->cpupri) != 0)
4976 free_cpumask_var(rd->rto_mask);
4978 free_cpumask_var(rd->online);
4980 free_cpumask_var(rd->span);
4986 * By default the system creates a single root-domain with all cpus as
4987 * members (mimicking the global state we have today).
4989 struct root_domain def_root_domain;
4991 static void init_defrootdomain(void)
4993 init_rootdomain(&def_root_domain);
4995 atomic_set(&def_root_domain.refcount, 1);
4998 static struct root_domain *alloc_rootdomain(void)
5000 struct root_domain *rd;
5002 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5006 if (init_rootdomain(rd) != 0) {
5014 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5016 struct sched_group *tmp, *first;
5025 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5030 } while (sg != first);
5033 static void free_sched_domain(struct rcu_head *rcu)
5035 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5038 * If its an overlapping domain it has private groups, iterate and
5041 if (sd->flags & SD_OVERLAP) {
5042 free_sched_groups(sd->groups, 1);
5043 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5044 kfree(sd->groups->sgp);
5050 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5052 call_rcu(&sd->rcu, free_sched_domain);
5055 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5057 for (; sd; sd = sd->parent)
5058 destroy_sched_domain(sd, cpu);
5062 * Keep a special pointer to the highest sched_domain that has
5063 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5064 * allows us to avoid some pointer chasing select_idle_sibling().
5066 * Also keep a unique ID per domain (we use the first cpu number in
5067 * the cpumask of the domain), this allows us to quickly tell if
5068 * two cpus are in the same cache domain, see cpus_share_cache().
5070 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5071 DEFINE_PER_CPU(int, sd_llc_id);
5073 static void update_top_cache_domain(int cpu)
5075 struct sched_domain *sd;
5078 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5080 id = cpumask_first(sched_domain_span(sd));
5082 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5083 per_cpu(sd_llc_id, cpu) = id;
5087 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5088 * hold the hotplug lock.
5091 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5093 struct rq *rq = cpu_rq(cpu);
5094 struct sched_domain *tmp;
5096 /* Remove the sched domains which do not contribute to scheduling. */
5097 for (tmp = sd; tmp; ) {
5098 struct sched_domain *parent = tmp->parent;
5102 if (sd_parent_degenerate(tmp, parent)) {
5103 tmp->parent = parent->parent;
5105 parent->parent->child = tmp;
5106 destroy_sched_domain(parent, cpu);
5111 if (sd && sd_degenerate(sd)) {
5114 destroy_sched_domain(tmp, cpu);
5119 sched_domain_debug(sd, cpu);
5121 rq_attach_root(rq, rd);
5123 rcu_assign_pointer(rq->sd, sd);
5124 destroy_sched_domains(tmp, cpu);
5126 update_top_cache_domain(cpu);
5129 /* cpus with isolated domains */
5130 static cpumask_var_t cpu_isolated_map;
5132 /* Setup the mask of cpus configured for isolated domains */
5133 static int __init isolated_cpu_setup(char *str)
5135 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5136 cpulist_parse(str, cpu_isolated_map);
5140 __setup("isolcpus=", isolated_cpu_setup);
5142 static const struct cpumask *cpu_cpu_mask(int cpu)
5144 return cpumask_of_node(cpu_to_node(cpu));
5148 struct sched_domain **__percpu sd;
5149 struct sched_group **__percpu sg;
5150 struct sched_group_power **__percpu sgp;
5154 struct sched_domain ** __percpu sd;
5155 struct root_domain *rd;
5165 struct sched_domain_topology_level;
5167 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5168 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5170 #define SDTL_OVERLAP 0x01
5172 struct sched_domain_topology_level {
5173 sched_domain_init_f init;
5174 sched_domain_mask_f mask;
5177 struct sd_data data;
5181 * Build an iteration mask that can exclude certain CPUs from the upwards
5184 * Asymmetric node setups can result in situations where the domain tree is of
5185 * unequal depth, make sure to skip domains that already cover the entire
5188 * In that case build_sched_domains() will have terminated the iteration early
5189 * and our sibling sd spans will be empty. Domains should always include the
5190 * cpu they're built on, so check that.
5193 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5195 const struct cpumask *span = sched_domain_span(sd);
5196 struct sd_data *sdd = sd->private;
5197 struct sched_domain *sibling;
5200 for_each_cpu(i, span) {
5201 sibling = *per_cpu_ptr(sdd->sd, i);
5202 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5205 cpumask_set_cpu(i, sched_group_mask(sg));
5210 * Return the canonical balance cpu for this group, this is the first cpu
5211 * of this group that's also in the iteration mask.
5213 int group_balance_cpu(struct sched_group *sg)
5215 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5219 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5221 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5222 const struct cpumask *span = sched_domain_span(sd);
5223 struct cpumask *covered = sched_domains_tmpmask;
5224 struct sd_data *sdd = sd->private;
5225 struct sched_domain *child;
5228 cpumask_clear(covered);
5230 for_each_cpu(i, span) {
5231 struct cpumask *sg_span;
5233 if (cpumask_test_cpu(i, covered))
5236 child = *per_cpu_ptr(sdd->sd, i);
5238 /* See the comment near build_group_mask(). */
5239 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5242 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5243 GFP_KERNEL, cpu_to_node(cpu));
5248 sg_span = sched_group_cpus(sg);
5250 child = child->child;
5251 cpumask_copy(sg_span, sched_domain_span(child));
5253 cpumask_set_cpu(i, sg_span);
5255 cpumask_or(covered, covered, sg_span);
5257 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5258 if (atomic_inc_return(&sg->sgp->ref) == 1)
5259 build_group_mask(sd, sg);
5262 * Initialize sgp->power such that even if we mess up the
5263 * domains and no possible iteration will get us here, we won't
5266 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5269 * Make sure the first group of this domain contains the
5270 * canonical balance cpu. Otherwise the sched_domain iteration
5271 * breaks. See update_sg_lb_stats().
5273 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5274 group_balance_cpu(sg) == cpu)
5284 sd->groups = groups;
5289 free_sched_groups(first, 0);
5294 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5296 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5297 struct sched_domain *child = sd->child;
5300 cpu = cpumask_first(sched_domain_span(child));
5303 *sg = *per_cpu_ptr(sdd->sg, cpu);
5304 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5305 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5312 * build_sched_groups will build a circular linked list of the groups
5313 * covered by the given span, and will set each group's ->cpumask correctly,
5314 * and ->cpu_power to 0.
5316 * Assumes the sched_domain tree is fully constructed
5319 build_sched_groups(struct sched_domain *sd, int cpu)
5321 struct sched_group *first = NULL, *last = NULL;
5322 struct sd_data *sdd = sd->private;
5323 const struct cpumask *span = sched_domain_span(sd);
5324 struct cpumask *covered;
5327 get_group(cpu, sdd, &sd->groups);
5328 atomic_inc(&sd->groups->ref);
5330 if (cpu != cpumask_first(span))
5333 lockdep_assert_held(&sched_domains_mutex);
5334 covered = sched_domains_tmpmask;
5336 cpumask_clear(covered);
5338 for_each_cpu(i, span) {
5339 struct sched_group *sg;
5342 if (cpumask_test_cpu(i, covered))
5345 group = get_group(i, sdd, &sg);
5346 cpumask_clear(sched_group_cpus(sg));
5348 cpumask_setall(sched_group_mask(sg));
5350 for_each_cpu(j, span) {
5351 if (get_group(j, sdd, NULL) != group)
5354 cpumask_set_cpu(j, covered);
5355 cpumask_set_cpu(j, sched_group_cpus(sg));
5370 * Initialize sched groups cpu_power.
5372 * cpu_power indicates the capacity of sched group, which is used while
5373 * distributing the load between different sched groups in a sched domain.
5374 * Typically cpu_power for all the groups in a sched domain will be same unless
5375 * there are asymmetries in the topology. If there are asymmetries, group
5376 * having more cpu_power will pickup more load compared to the group having
5379 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5381 struct sched_group *sg = sd->groups;
5386 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5388 } while (sg != sd->groups);
5390 if (cpu != group_balance_cpu(sg))
5393 update_group_power(sd, cpu);
5394 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5397 int __weak arch_sd_sibling_asym_packing(void)
5399 return 0*SD_ASYM_PACKING;
5403 * Initializers for schedule domains
5404 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5407 #ifdef CONFIG_SCHED_DEBUG
5408 # define SD_INIT_NAME(sd, type) sd->name = #type
5410 # define SD_INIT_NAME(sd, type) do { } while (0)
5413 #define SD_INIT_FUNC(type) \
5414 static noinline struct sched_domain * \
5415 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5417 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5418 *sd = SD_##type##_INIT; \
5419 SD_INIT_NAME(sd, type); \
5420 sd->private = &tl->data; \
5425 #ifdef CONFIG_SCHED_SMT
5426 SD_INIT_FUNC(SIBLING)
5428 #ifdef CONFIG_SCHED_MC
5431 #ifdef CONFIG_SCHED_BOOK
5435 static int default_relax_domain_level = -1;
5436 int sched_domain_level_max;
5438 static int __init setup_relax_domain_level(char *str)
5440 if (kstrtoint(str, 0, &default_relax_domain_level))
5441 pr_warn("Unable to set relax_domain_level\n");
5445 __setup("relax_domain_level=", setup_relax_domain_level);
5447 static void set_domain_attribute(struct sched_domain *sd,
5448 struct sched_domain_attr *attr)
5452 if (!attr || attr->relax_domain_level < 0) {
5453 if (default_relax_domain_level < 0)
5456 request = default_relax_domain_level;
5458 request = attr->relax_domain_level;
5459 if (request < sd->level) {
5460 /* turn off idle balance on this domain */
5461 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5463 /* turn on idle balance on this domain */
5464 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5468 static void __sdt_free(const struct cpumask *cpu_map);
5469 static int __sdt_alloc(const struct cpumask *cpu_map);
5471 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5472 const struct cpumask *cpu_map)
5476 if (!atomic_read(&d->rd->refcount))
5477 free_rootdomain(&d->rd->rcu); /* fall through */
5479 free_percpu(d->sd); /* fall through */
5481 __sdt_free(cpu_map); /* fall through */
5487 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5488 const struct cpumask *cpu_map)
5490 memset(d, 0, sizeof(*d));
5492 if (__sdt_alloc(cpu_map))
5493 return sa_sd_storage;
5494 d->sd = alloc_percpu(struct sched_domain *);
5496 return sa_sd_storage;
5497 d->rd = alloc_rootdomain();
5500 return sa_rootdomain;
5504 * NULL the sd_data elements we've used to build the sched_domain and
5505 * sched_group structure so that the subsequent __free_domain_allocs()
5506 * will not free the data we're using.
5508 static void claim_allocations(int cpu, struct sched_domain *sd)
5510 struct sd_data *sdd = sd->private;
5512 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5513 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5515 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5516 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5518 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5519 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5522 #ifdef CONFIG_SCHED_SMT
5523 static const struct cpumask *cpu_smt_mask(int cpu)
5525 return topology_thread_cpumask(cpu);
5530 * Topology list, bottom-up.
5532 static struct sched_domain_topology_level default_topology[] = {
5533 #ifdef CONFIG_SCHED_SMT
5534 { sd_init_SIBLING, cpu_smt_mask, },
5536 #ifdef CONFIG_SCHED_MC
5537 { sd_init_MC, cpu_coregroup_mask, },
5539 #ifdef CONFIG_SCHED_BOOK
5540 { sd_init_BOOK, cpu_book_mask, },
5542 { sd_init_CPU, cpu_cpu_mask, },
5546 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5548 #define for_each_sd_topology(tl) \
5549 for (tl = sched_domain_topology; tl->init; tl++)
5553 static int sched_domains_numa_levels;
5554 static int *sched_domains_numa_distance;
5555 static struct cpumask ***sched_domains_numa_masks;
5556 static int sched_domains_curr_level;
5558 static inline int sd_local_flags(int level)
5560 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5563 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5566 static struct sched_domain *
5567 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5569 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5570 int level = tl->numa_level;
5571 int sd_weight = cpumask_weight(
5572 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5574 *sd = (struct sched_domain){
5575 .min_interval = sd_weight,
5576 .max_interval = 2*sd_weight,
5578 .imbalance_pct = 125,
5579 .cache_nice_tries = 2,
5586 .flags = 1*SD_LOAD_BALANCE
5587 | 1*SD_BALANCE_NEWIDLE
5592 | 0*SD_SHARE_CPUPOWER
5593 | 0*SD_SHARE_PKG_RESOURCES
5595 | 0*SD_PREFER_SIBLING
5596 | sd_local_flags(level)
5598 .last_balance = jiffies,
5599 .balance_interval = sd_weight,
5601 SD_INIT_NAME(sd, NUMA);
5602 sd->private = &tl->data;
5605 * Ugly hack to pass state to sd_numa_mask()...
5607 sched_domains_curr_level = tl->numa_level;
5612 static const struct cpumask *sd_numa_mask(int cpu)
5614 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5617 static void sched_numa_warn(const char *str)
5619 static int done = false;
5627 printk(KERN_WARNING "ERROR: %s\n\n", str);
5629 for (i = 0; i < nr_node_ids; i++) {
5630 printk(KERN_WARNING " ");
5631 for (j = 0; j < nr_node_ids; j++)
5632 printk(KERN_CONT "%02d ", node_distance(i,j));
5633 printk(KERN_CONT "\n");
5635 printk(KERN_WARNING "\n");
5638 static bool find_numa_distance(int distance)
5642 if (distance == node_distance(0, 0))
5645 for (i = 0; i < sched_domains_numa_levels; i++) {
5646 if (sched_domains_numa_distance[i] == distance)
5653 static void sched_init_numa(void)
5655 int next_distance, curr_distance = node_distance(0, 0);
5656 struct sched_domain_topology_level *tl;
5660 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
5661 if (!sched_domains_numa_distance)
5665 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
5666 * unique distances in the node_distance() table.
5668 * Assumes node_distance(0,j) includes all distances in
5669 * node_distance(i,j) in order to avoid cubic time.
5671 next_distance = curr_distance;
5672 for (i = 0; i < nr_node_ids; i++) {
5673 for (j = 0; j < nr_node_ids; j++) {
5674 for (k = 0; k < nr_node_ids; k++) {
5675 int distance = node_distance(i, k);
5677 if (distance > curr_distance &&
5678 (distance < next_distance ||
5679 next_distance == curr_distance))
5680 next_distance = distance;
5683 * While not a strong assumption it would be nice to know
5684 * about cases where if node A is connected to B, B is not
5685 * equally connected to A.
5687 if (sched_debug() && node_distance(k, i) != distance)
5688 sched_numa_warn("Node-distance not symmetric");
5690 if (sched_debug() && i && !find_numa_distance(distance))
5691 sched_numa_warn("Node-0 not representative");
5693 if (next_distance != curr_distance) {
5694 sched_domains_numa_distance[level++] = next_distance;
5695 sched_domains_numa_levels = level;
5696 curr_distance = next_distance;
5701 * In case of sched_debug() we verify the above assumption.
5707 * 'level' contains the number of unique distances, excluding the
5708 * identity distance node_distance(i,i).
5710 * The sched_domains_numa_distance[] array includes the actual distance
5715 * Here, we should temporarily reset sched_domains_numa_levels to 0.
5716 * If it fails to allocate memory for array sched_domains_numa_masks[][],
5717 * the array will contain less then 'level' members. This could be
5718 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
5719 * in other functions.
5721 * We reset it to 'level' at the end of this function.
5723 sched_domains_numa_levels = 0;
5725 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
5726 if (!sched_domains_numa_masks)
5730 * Now for each level, construct a mask per node which contains all
5731 * cpus of nodes that are that many hops away from us.
5733 for (i = 0; i < level; i++) {
5734 sched_domains_numa_masks[i] =
5735 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
5736 if (!sched_domains_numa_masks[i])
5739 for (j = 0; j < nr_node_ids; j++) {
5740 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
5744 sched_domains_numa_masks[i][j] = mask;
5746 for (k = 0; k < nr_node_ids; k++) {
5747 if (node_distance(j, k) > sched_domains_numa_distance[i])
5750 cpumask_or(mask, mask, cpumask_of_node(k));
5755 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
5756 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
5761 * Copy the default topology bits..
5763 for (i = 0; default_topology[i].init; i++)
5764 tl[i] = default_topology[i];
5767 * .. and append 'j' levels of NUMA goodness.
5769 for (j = 0; j < level; i++, j++) {
5770 tl[i] = (struct sched_domain_topology_level){
5771 .init = sd_numa_init,
5772 .mask = sd_numa_mask,
5773 .flags = SDTL_OVERLAP,
5778 sched_domain_topology = tl;
5780 sched_domains_numa_levels = level;
5783 static void sched_domains_numa_masks_set(int cpu)
5786 int node = cpu_to_node(cpu);
5788 for (i = 0; i < sched_domains_numa_levels; i++) {
5789 for (j = 0; j < nr_node_ids; j++) {
5790 if (node_distance(j, node) <= sched_domains_numa_distance[i])
5791 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
5796 static void sched_domains_numa_masks_clear(int cpu)
5799 for (i = 0; i < sched_domains_numa_levels; i++) {
5800 for (j = 0; j < nr_node_ids; j++)
5801 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
5806 * Update sched_domains_numa_masks[level][node] array when new cpus
5809 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5810 unsigned long action,
5813 int cpu = (long)hcpu;
5815 switch (action & ~CPU_TASKS_FROZEN) {
5817 sched_domains_numa_masks_set(cpu);
5821 sched_domains_numa_masks_clear(cpu);
5831 static inline void sched_init_numa(void)
5835 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
5836 unsigned long action,
5841 #endif /* CONFIG_NUMA */
5843 static int __sdt_alloc(const struct cpumask *cpu_map)
5845 struct sched_domain_topology_level *tl;
5848 for_each_sd_topology(tl) {
5849 struct sd_data *sdd = &tl->data;
5851 sdd->sd = alloc_percpu(struct sched_domain *);
5855 sdd->sg = alloc_percpu(struct sched_group *);
5859 sdd->sgp = alloc_percpu(struct sched_group_power *);
5863 for_each_cpu(j, cpu_map) {
5864 struct sched_domain *sd;
5865 struct sched_group *sg;
5866 struct sched_group_power *sgp;
5868 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
5869 GFP_KERNEL, cpu_to_node(j));
5873 *per_cpu_ptr(sdd->sd, j) = sd;
5875 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5876 GFP_KERNEL, cpu_to_node(j));
5882 *per_cpu_ptr(sdd->sg, j) = sg;
5884 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
5885 GFP_KERNEL, cpu_to_node(j));
5889 *per_cpu_ptr(sdd->sgp, j) = sgp;
5896 static void __sdt_free(const struct cpumask *cpu_map)
5898 struct sched_domain_topology_level *tl;
5901 for_each_sd_topology(tl) {
5902 struct sd_data *sdd = &tl->data;
5904 for_each_cpu(j, cpu_map) {
5905 struct sched_domain *sd;
5908 sd = *per_cpu_ptr(sdd->sd, j);
5909 if (sd && (sd->flags & SD_OVERLAP))
5910 free_sched_groups(sd->groups, 0);
5911 kfree(*per_cpu_ptr(sdd->sd, j));
5915 kfree(*per_cpu_ptr(sdd->sg, j));
5917 kfree(*per_cpu_ptr(sdd->sgp, j));
5919 free_percpu(sdd->sd);
5921 free_percpu(sdd->sg);
5923 free_percpu(sdd->sgp);
5928 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
5929 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
5930 struct sched_domain *child, int cpu)
5932 struct sched_domain *sd = tl->init(tl, cpu);
5936 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
5938 sd->level = child->level + 1;
5939 sched_domain_level_max = max(sched_domain_level_max, sd->level);
5943 set_domain_attribute(sd, attr);
5949 * Build sched domains for a given set of cpus and attach the sched domains
5950 * to the individual cpus
5952 static int build_sched_domains(const struct cpumask *cpu_map,
5953 struct sched_domain_attr *attr)
5955 enum s_alloc alloc_state;
5956 struct sched_domain *sd;
5958 int i, ret = -ENOMEM;
5960 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
5961 if (alloc_state != sa_rootdomain)
5964 /* Set up domains for cpus specified by the cpu_map. */
5965 for_each_cpu(i, cpu_map) {
5966 struct sched_domain_topology_level *tl;
5969 for_each_sd_topology(tl) {
5970 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
5971 if (tl == sched_domain_topology)
5972 *per_cpu_ptr(d.sd, i) = sd;
5973 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
5974 sd->flags |= SD_OVERLAP;
5975 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
5980 /* Build the groups for the domains */
5981 for_each_cpu(i, cpu_map) {
5982 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
5983 sd->span_weight = cpumask_weight(sched_domain_span(sd));
5984 if (sd->flags & SD_OVERLAP) {
5985 if (build_overlap_sched_groups(sd, i))
5988 if (build_sched_groups(sd, i))
5994 /* Calculate CPU power for physical packages and nodes */
5995 for (i = nr_cpumask_bits-1; i >= 0; i--) {
5996 if (!cpumask_test_cpu(i, cpu_map))
5999 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6000 claim_allocations(i, sd);
6001 init_sched_groups_power(i, sd);
6005 /* Attach the domains */
6007 for_each_cpu(i, cpu_map) {
6008 sd = *per_cpu_ptr(d.sd, i);
6009 cpu_attach_domain(sd, d.rd, i);
6015 __free_domain_allocs(&d, alloc_state, cpu_map);
6019 static cpumask_var_t *doms_cur; /* current sched domains */
6020 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6021 static struct sched_domain_attr *dattr_cur;
6022 /* attribues of custom domains in 'doms_cur' */
6025 * Special case: If a kmalloc of a doms_cur partition (array of
6026 * cpumask) fails, then fallback to a single sched domain,
6027 * as determined by the single cpumask fallback_doms.
6029 static cpumask_var_t fallback_doms;
6032 * arch_update_cpu_topology lets virtualized architectures update the
6033 * cpu core maps. It is supposed to return 1 if the topology changed
6034 * or 0 if it stayed the same.
6036 int __attribute__((weak)) arch_update_cpu_topology(void)
6041 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6044 cpumask_var_t *doms;
6046 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6049 for (i = 0; i < ndoms; i++) {
6050 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6051 free_sched_domains(doms, i);
6058 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6061 for (i = 0; i < ndoms; i++)
6062 free_cpumask_var(doms[i]);
6067 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6068 * For now this just excludes isolated cpus, but could be used to
6069 * exclude other special cases in the future.
6071 static int init_sched_domains(const struct cpumask *cpu_map)
6075 arch_update_cpu_topology();
6077 doms_cur = alloc_sched_domains(ndoms_cur);
6079 doms_cur = &fallback_doms;
6080 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6081 err = build_sched_domains(doms_cur[0], NULL);
6082 register_sched_domain_sysctl();
6088 * Detach sched domains from a group of cpus specified in cpu_map
6089 * These cpus will now be attached to the NULL domain
6091 static void detach_destroy_domains(const struct cpumask *cpu_map)
6096 for_each_cpu(i, cpu_map)
6097 cpu_attach_domain(NULL, &def_root_domain, i);
6101 /* handle null as "default" */
6102 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6103 struct sched_domain_attr *new, int idx_new)
6105 struct sched_domain_attr tmp;
6112 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6113 new ? (new + idx_new) : &tmp,
6114 sizeof(struct sched_domain_attr));
6118 * Partition sched domains as specified by the 'ndoms_new'
6119 * cpumasks in the array doms_new[] of cpumasks. This compares
6120 * doms_new[] to the current sched domain partitioning, doms_cur[].
6121 * It destroys each deleted domain and builds each new domain.
6123 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6124 * The masks don't intersect (don't overlap.) We should setup one
6125 * sched domain for each mask. CPUs not in any of the cpumasks will
6126 * not be load balanced. If the same cpumask appears both in the
6127 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6130 * The passed in 'doms_new' should be allocated using
6131 * alloc_sched_domains. This routine takes ownership of it and will
6132 * free_sched_domains it when done with it. If the caller failed the
6133 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6134 * and partition_sched_domains() will fallback to the single partition
6135 * 'fallback_doms', it also forces the domains to be rebuilt.
6137 * If doms_new == NULL it will be replaced with cpu_online_mask.
6138 * ndoms_new == 0 is a special case for destroying existing domains,
6139 * and it will not create the default domain.
6141 * Call with hotplug lock held
6143 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6144 struct sched_domain_attr *dattr_new)
6149 mutex_lock(&sched_domains_mutex);
6151 /* always unregister in case we don't destroy any domains */
6152 unregister_sched_domain_sysctl();
6154 /* Let architecture update cpu core mappings. */
6155 new_topology = arch_update_cpu_topology();
6157 n = doms_new ? ndoms_new : 0;
6159 /* Destroy deleted domains */
6160 for (i = 0; i < ndoms_cur; i++) {
6161 for (j = 0; j < n && !new_topology; j++) {
6162 if (cpumask_equal(doms_cur[i], doms_new[j])
6163 && dattrs_equal(dattr_cur, i, dattr_new, j))
6166 /* no match - a current sched domain not in new doms_new[] */
6167 detach_destroy_domains(doms_cur[i]);
6172 if (doms_new == NULL) {
6174 doms_new = &fallback_doms;
6175 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6176 WARN_ON_ONCE(dattr_new);
6179 /* Build new domains */
6180 for (i = 0; i < ndoms_new; i++) {
6181 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6182 if (cpumask_equal(doms_new[i], doms_cur[j])
6183 && dattrs_equal(dattr_new, i, dattr_cur, j))
6186 /* no match - add a new doms_new */
6187 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6192 /* Remember the new sched domains */
6193 if (doms_cur != &fallback_doms)
6194 free_sched_domains(doms_cur, ndoms_cur);
6195 kfree(dattr_cur); /* kfree(NULL) is safe */
6196 doms_cur = doms_new;
6197 dattr_cur = dattr_new;
6198 ndoms_cur = ndoms_new;
6200 register_sched_domain_sysctl();
6202 mutex_unlock(&sched_domains_mutex);
6205 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6208 * Update cpusets according to cpu_active mask. If cpusets are
6209 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6210 * around partition_sched_domains().
6212 * If we come here as part of a suspend/resume, don't touch cpusets because we
6213 * want to restore it back to its original state upon resume anyway.
6215 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6219 case CPU_ONLINE_FROZEN:
6220 case CPU_DOWN_FAILED_FROZEN:
6223 * num_cpus_frozen tracks how many CPUs are involved in suspend
6224 * resume sequence. As long as this is not the last online
6225 * operation in the resume sequence, just build a single sched
6226 * domain, ignoring cpusets.
6229 if (likely(num_cpus_frozen)) {
6230 partition_sched_domains(1, NULL, NULL);
6235 * This is the last CPU online operation. So fall through and
6236 * restore the original sched domains by considering the
6237 * cpuset configurations.
6241 case CPU_DOWN_FAILED:
6242 cpuset_update_active_cpus(true);
6250 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6254 case CPU_DOWN_PREPARE:
6255 cpuset_update_active_cpus(false);
6257 case CPU_DOWN_PREPARE_FROZEN:
6259 partition_sched_domains(1, NULL, NULL);
6267 void __init sched_init_smp(void)
6269 cpumask_var_t non_isolated_cpus;
6271 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6272 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6277 mutex_lock(&sched_domains_mutex);
6278 init_sched_domains(cpu_active_mask);
6279 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6280 if (cpumask_empty(non_isolated_cpus))
6281 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6282 mutex_unlock(&sched_domains_mutex);
6285 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6286 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6287 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6291 /* Move init over to a non-isolated CPU */
6292 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6294 sched_init_granularity();
6295 free_cpumask_var(non_isolated_cpus);
6297 init_sched_rt_class();
6300 void __init sched_init_smp(void)
6302 sched_init_granularity();
6304 #endif /* CONFIG_SMP */
6306 const_debug unsigned int sysctl_timer_migration = 1;
6308 int in_sched_functions(unsigned long addr)
6310 return in_lock_functions(addr) ||
6311 (addr >= (unsigned long)__sched_text_start
6312 && addr < (unsigned long)__sched_text_end);
6315 #ifdef CONFIG_CGROUP_SCHED
6317 * Default task group.
6318 * Every task in system belongs to this group at bootup.
6320 struct task_group root_task_group;
6321 LIST_HEAD(task_groups);
6324 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6326 void __init sched_init(void)
6329 unsigned long alloc_size = 0, ptr;
6331 #ifdef CONFIG_FAIR_GROUP_SCHED
6332 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6334 #ifdef CONFIG_RT_GROUP_SCHED
6335 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6337 #ifdef CONFIG_CPUMASK_OFFSTACK
6338 alloc_size += num_possible_cpus() * cpumask_size();
6341 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6343 #ifdef CONFIG_FAIR_GROUP_SCHED
6344 root_task_group.se = (struct sched_entity **)ptr;
6345 ptr += nr_cpu_ids * sizeof(void **);
6347 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6348 ptr += nr_cpu_ids * sizeof(void **);
6350 #endif /* CONFIG_FAIR_GROUP_SCHED */
6351 #ifdef CONFIG_RT_GROUP_SCHED
6352 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6353 ptr += nr_cpu_ids * sizeof(void **);
6355 root_task_group.rt_rq = (struct rt_rq **)ptr;
6356 ptr += nr_cpu_ids * sizeof(void **);
6358 #endif /* CONFIG_RT_GROUP_SCHED */
6359 #ifdef CONFIG_CPUMASK_OFFSTACK
6360 for_each_possible_cpu(i) {
6361 per_cpu(load_balance_mask, i) = (void *)ptr;
6362 ptr += cpumask_size();
6364 #endif /* CONFIG_CPUMASK_OFFSTACK */
6368 init_defrootdomain();
6371 init_rt_bandwidth(&def_rt_bandwidth,
6372 global_rt_period(), global_rt_runtime());
6374 #ifdef CONFIG_RT_GROUP_SCHED
6375 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6376 global_rt_period(), global_rt_runtime());
6377 #endif /* CONFIG_RT_GROUP_SCHED */
6379 #ifdef CONFIG_CGROUP_SCHED
6380 list_add(&root_task_group.list, &task_groups);
6381 INIT_LIST_HEAD(&root_task_group.children);
6382 INIT_LIST_HEAD(&root_task_group.siblings);
6383 autogroup_init(&init_task);
6385 #endif /* CONFIG_CGROUP_SCHED */
6387 for_each_possible_cpu(i) {
6391 raw_spin_lock_init(&rq->lock);
6393 rq->calc_load_active = 0;
6394 rq->calc_load_update = jiffies + LOAD_FREQ;
6395 init_cfs_rq(&rq->cfs);
6396 init_rt_rq(&rq->rt, rq);
6397 #ifdef CONFIG_FAIR_GROUP_SCHED
6398 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6399 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6401 * How much cpu bandwidth does root_task_group get?
6403 * In case of task-groups formed thr' the cgroup filesystem, it
6404 * gets 100% of the cpu resources in the system. This overall
6405 * system cpu resource is divided among the tasks of
6406 * root_task_group and its child task-groups in a fair manner,
6407 * based on each entity's (task or task-group's) weight
6408 * (se->load.weight).
6410 * In other words, if root_task_group has 10 tasks of weight
6411 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6412 * then A0's share of the cpu resource is:
6414 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6416 * We achieve this by letting root_task_group's tasks sit
6417 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6419 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6420 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6421 #endif /* CONFIG_FAIR_GROUP_SCHED */
6423 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6424 #ifdef CONFIG_RT_GROUP_SCHED
6425 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6426 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6429 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6430 rq->cpu_load[j] = 0;
6432 rq->last_load_update_tick = jiffies;
6437 rq->cpu_power = SCHED_POWER_SCALE;
6438 rq->post_schedule = 0;
6439 rq->active_balance = 0;
6440 rq->next_balance = jiffies;
6445 rq->avg_idle = 2*sysctl_sched_migration_cost;
6447 INIT_LIST_HEAD(&rq->cfs_tasks);
6449 rq_attach_root(rq, &def_root_domain);
6450 #ifdef CONFIG_NO_HZ_COMMON
6453 #ifdef CONFIG_NO_HZ_FULL
6454 rq->last_sched_tick = 0;
6458 atomic_set(&rq->nr_iowait, 0);
6461 set_load_weight(&init_task);
6463 #ifdef CONFIG_PREEMPT_NOTIFIERS
6464 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6467 #ifdef CONFIG_RT_MUTEXES
6468 plist_head_init(&init_task.pi_waiters);
6472 * The boot idle thread does lazy MMU switching as well:
6474 atomic_inc(&init_mm.mm_count);
6475 enter_lazy_tlb(&init_mm, current);
6478 * Make us the idle thread. Technically, schedule() should not be
6479 * called from this thread, however somewhere below it might be,
6480 * but because we are the idle thread, we just pick up running again
6481 * when this runqueue becomes "idle".
6483 init_idle(current, smp_processor_id());
6485 calc_load_update = jiffies + LOAD_FREQ;
6488 * During early bootup we pretend to be a normal task:
6490 current->sched_class = &fair_sched_class;
6493 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6494 /* May be allocated at isolcpus cmdline parse time */
6495 if (cpu_isolated_map == NULL)
6496 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6497 idle_thread_set_boot_cpu();
6499 init_sched_fair_class();
6501 scheduler_running = 1;
6504 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6505 static inline int preempt_count_equals(int preempt_offset)
6507 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6509 return (nested == preempt_offset);
6512 void __might_sleep(const char *file, int line, int preempt_offset)
6514 static unsigned long prev_jiffy; /* ratelimiting */
6516 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6517 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6518 system_state != SYSTEM_RUNNING || oops_in_progress)
6520 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6522 prev_jiffy = jiffies;
6525 "BUG: sleeping function called from invalid context at %s:%d\n",
6528 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6529 in_atomic(), irqs_disabled(),
6530 current->pid, current->comm);
6532 debug_show_held_locks(current);
6533 if (irqs_disabled())
6534 print_irqtrace_events(current);
6537 EXPORT_SYMBOL(__might_sleep);
6540 #ifdef CONFIG_MAGIC_SYSRQ
6541 static void normalize_task(struct rq *rq, struct task_struct *p)
6543 const struct sched_class *prev_class = p->sched_class;
6544 int old_prio = p->prio;
6549 dequeue_task(rq, p, 0);
6550 __setscheduler(rq, p, SCHED_NORMAL, 0);
6552 enqueue_task(rq, p, 0);
6553 resched_task(rq->curr);
6556 check_class_changed(rq, p, prev_class, old_prio);
6559 void normalize_rt_tasks(void)
6561 struct task_struct *g, *p;
6562 unsigned long flags;
6565 read_lock_irqsave(&tasklist_lock, flags);
6566 do_each_thread(g, p) {
6568 * Only normalize user tasks:
6573 p->se.exec_start = 0;
6574 #ifdef CONFIG_SCHEDSTATS
6575 p->se.statistics.wait_start = 0;
6576 p->se.statistics.sleep_start = 0;
6577 p->se.statistics.block_start = 0;
6582 * Renice negative nice level userspace
6585 if (TASK_NICE(p) < 0 && p->mm)
6586 set_user_nice(p, 0);
6590 raw_spin_lock(&p->pi_lock);
6591 rq = __task_rq_lock(p);
6593 normalize_task(rq, p);
6595 __task_rq_unlock(rq);
6596 raw_spin_unlock(&p->pi_lock);
6597 } while_each_thread(g, p);
6599 read_unlock_irqrestore(&tasklist_lock, flags);
6602 #endif /* CONFIG_MAGIC_SYSRQ */
6604 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6606 * These functions are only useful for the IA64 MCA handling, or kdb.
6608 * They can only be called when the whole system has been
6609 * stopped - every CPU needs to be quiescent, and no scheduling
6610 * activity can take place. Using them for anything else would
6611 * be a serious bug, and as a result, they aren't even visible
6612 * under any other configuration.
6616 * curr_task - return the current task for a given cpu.
6617 * @cpu: the processor in question.
6619 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6621 struct task_struct *curr_task(int cpu)
6623 return cpu_curr(cpu);
6626 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
6630 * set_curr_task - set the current task for a given cpu.
6631 * @cpu: the processor in question.
6632 * @p: the task pointer to set.
6634 * Description: This function must only be used when non-maskable interrupts
6635 * are serviced on a separate stack. It allows the architecture to switch the
6636 * notion of the current task on a cpu in a non-blocking manner. This function
6637 * must be called with all CPU's synchronized, and interrupts disabled, the
6638 * and caller must save the original value of the current task (see
6639 * curr_task() above) and restore that value before reenabling interrupts and
6640 * re-starting the system.
6642 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6644 void set_curr_task(int cpu, struct task_struct *p)
6651 #ifdef CONFIG_CGROUP_SCHED
6652 /* task_group_lock serializes the addition/removal of task groups */
6653 static DEFINE_SPINLOCK(task_group_lock);
6655 static void free_sched_group(struct task_group *tg)
6657 free_fair_sched_group(tg);
6658 free_rt_sched_group(tg);
6663 /* allocate runqueue etc for a new task group */
6664 struct task_group *sched_create_group(struct task_group *parent)
6666 struct task_group *tg;
6668 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
6670 return ERR_PTR(-ENOMEM);
6672 if (!alloc_fair_sched_group(tg, parent))
6675 if (!alloc_rt_sched_group(tg, parent))
6681 free_sched_group(tg);
6682 return ERR_PTR(-ENOMEM);
6685 void sched_online_group(struct task_group *tg, struct task_group *parent)
6687 unsigned long flags;
6689 spin_lock_irqsave(&task_group_lock, flags);
6690 list_add_rcu(&tg->list, &task_groups);
6692 WARN_ON(!parent); /* root should already exist */
6694 tg->parent = parent;
6695 INIT_LIST_HEAD(&tg->children);
6696 list_add_rcu(&tg->siblings, &parent->children);
6697 spin_unlock_irqrestore(&task_group_lock, flags);
6700 /* rcu callback to free various structures associated with a task group */
6701 static void free_sched_group_rcu(struct rcu_head *rhp)
6703 /* now it should be safe to free those cfs_rqs */
6704 free_sched_group(container_of(rhp, struct task_group, rcu));
6707 /* Destroy runqueue etc associated with a task group */
6708 void sched_destroy_group(struct task_group *tg)
6710 /* wait for possible concurrent references to cfs_rqs complete */
6711 call_rcu(&tg->rcu, free_sched_group_rcu);
6714 void sched_offline_group(struct task_group *tg)
6716 unsigned long flags;
6719 /* end participation in shares distribution */
6720 for_each_possible_cpu(i)
6721 unregister_fair_sched_group(tg, i);
6723 spin_lock_irqsave(&task_group_lock, flags);
6724 list_del_rcu(&tg->list);
6725 list_del_rcu(&tg->siblings);
6726 spin_unlock_irqrestore(&task_group_lock, flags);
6729 /* change task's runqueue when it moves between groups.
6730 * The caller of this function should have put the task in its new group
6731 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
6732 * reflect its new group.
6734 void sched_move_task(struct task_struct *tsk)
6736 struct task_group *tg;
6738 unsigned long flags;
6741 rq = task_rq_lock(tsk, &flags);
6743 running = task_current(rq, tsk);
6747 dequeue_task(rq, tsk, 0);
6748 if (unlikely(running))
6749 tsk->sched_class->put_prev_task(rq, tsk);
6751 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
6752 lockdep_is_held(&tsk->sighand->siglock)),
6753 struct task_group, css);
6754 tg = autogroup_task_group(tsk, tg);
6755 tsk->sched_task_group = tg;
6757 #ifdef CONFIG_FAIR_GROUP_SCHED
6758 if (tsk->sched_class->task_move_group)
6759 tsk->sched_class->task_move_group(tsk, on_rq);
6762 set_task_rq(tsk, task_cpu(tsk));
6764 if (unlikely(running))
6765 tsk->sched_class->set_curr_task(rq);
6767 enqueue_task(rq, tsk, 0);
6769 task_rq_unlock(rq, tsk, &flags);
6771 #endif /* CONFIG_CGROUP_SCHED */
6773 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
6774 static unsigned long to_ratio(u64 period, u64 runtime)
6776 if (runtime == RUNTIME_INF)
6779 return div64_u64(runtime << 20, period);
6783 #ifdef CONFIG_RT_GROUP_SCHED
6785 * Ensure that the real time constraints are schedulable.
6787 static DEFINE_MUTEX(rt_constraints_mutex);
6789 /* Must be called with tasklist_lock held */
6790 static inline int tg_has_rt_tasks(struct task_group *tg)
6792 struct task_struct *g, *p;
6794 do_each_thread(g, p) {
6795 if (rt_task(p) && task_rq(p)->rt.tg == tg)
6797 } while_each_thread(g, p);
6802 struct rt_schedulable_data {
6803 struct task_group *tg;
6808 static int tg_rt_schedulable(struct task_group *tg, void *data)
6810 struct rt_schedulable_data *d = data;
6811 struct task_group *child;
6812 unsigned long total, sum = 0;
6813 u64 period, runtime;
6815 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6816 runtime = tg->rt_bandwidth.rt_runtime;
6819 period = d->rt_period;
6820 runtime = d->rt_runtime;
6824 * Cannot have more runtime than the period.
6826 if (runtime > period && runtime != RUNTIME_INF)
6830 * Ensure we don't starve existing RT tasks.
6832 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
6835 total = to_ratio(period, runtime);
6838 * Nobody can have more than the global setting allows.
6840 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
6844 * The sum of our children's runtime should not exceed our own.
6846 list_for_each_entry_rcu(child, &tg->children, siblings) {
6847 period = ktime_to_ns(child->rt_bandwidth.rt_period);
6848 runtime = child->rt_bandwidth.rt_runtime;
6850 if (child == d->tg) {
6851 period = d->rt_period;
6852 runtime = d->rt_runtime;
6855 sum += to_ratio(period, runtime);
6864 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
6868 struct rt_schedulable_data data = {
6870 .rt_period = period,
6871 .rt_runtime = runtime,
6875 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
6881 static int tg_set_rt_bandwidth(struct task_group *tg,
6882 u64 rt_period, u64 rt_runtime)
6886 mutex_lock(&rt_constraints_mutex);
6887 read_lock(&tasklist_lock);
6888 err = __rt_schedulable(tg, rt_period, rt_runtime);
6892 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6893 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
6894 tg->rt_bandwidth.rt_runtime = rt_runtime;
6896 for_each_possible_cpu(i) {
6897 struct rt_rq *rt_rq = tg->rt_rq[i];
6899 raw_spin_lock(&rt_rq->rt_runtime_lock);
6900 rt_rq->rt_runtime = rt_runtime;
6901 raw_spin_unlock(&rt_rq->rt_runtime_lock);
6903 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
6905 read_unlock(&tasklist_lock);
6906 mutex_unlock(&rt_constraints_mutex);
6911 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
6913 u64 rt_runtime, rt_period;
6915 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
6916 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
6917 if (rt_runtime_us < 0)
6918 rt_runtime = RUNTIME_INF;
6920 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6923 static long sched_group_rt_runtime(struct task_group *tg)
6927 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
6930 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
6931 do_div(rt_runtime_us, NSEC_PER_USEC);
6932 return rt_runtime_us;
6935 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
6937 u64 rt_runtime, rt_period;
6939 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
6940 rt_runtime = tg->rt_bandwidth.rt_runtime;
6945 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
6948 static long sched_group_rt_period(struct task_group *tg)
6952 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
6953 do_div(rt_period_us, NSEC_PER_USEC);
6954 return rt_period_us;
6957 static int sched_rt_global_constraints(void)
6959 u64 runtime, period;
6962 if (sysctl_sched_rt_period <= 0)
6965 runtime = global_rt_runtime();
6966 period = global_rt_period();
6969 * Sanity check on the sysctl variables.
6971 if (runtime > period && runtime != RUNTIME_INF)
6974 mutex_lock(&rt_constraints_mutex);
6975 read_lock(&tasklist_lock);
6976 ret = __rt_schedulable(NULL, 0, 0);
6977 read_unlock(&tasklist_lock);
6978 mutex_unlock(&rt_constraints_mutex);
6983 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
6985 /* Don't accept realtime tasks when there is no way for them to run */
6986 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
6992 #else /* !CONFIG_RT_GROUP_SCHED */
6993 static int sched_rt_global_constraints(void)
6995 unsigned long flags;
6998 if (sysctl_sched_rt_period <= 0)
7002 * There's always some RT tasks in the root group
7003 * -- migration, kstopmachine etc..
7005 if (sysctl_sched_rt_runtime == 0)
7008 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7009 for_each_possible_cpu(i) {
7010 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7012 raw_spin_lock(&rt_rq->rt_runtime_lock);
7013 rt_rq->rt_runtime = global_rt_runtime();
7014 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7016 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7020 #endif /* CONFIG_RT_GROUP_SCHED */
7022 int sched_rr_handler(struct ctl_table *table, int write,
7023 void __user *buffer, size_t *lenp,
7027 static DEFINE_MUTEX(mutex);
7030 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7031 /* make sure that internally we keep jiffies */
7032 /* also, writing zero resets timeslice to default */
7033 if (!ret && write) {
7034 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7035 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7037 mutex_unlock(&mutex);
7041 int sched_rt_handler(struct ctl_table *table, int write,
7042 void __user *buffer, size_t *lenp,
7046 int old_period, old_runtime;
7047 static DEFINE_MUTEX(mutex);
7050 old_period = sysctl_sched_rt_period;
7051 old_runtime = sysctl_sched_rt_runtime;
7053 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7055 if (!ret && write) {
7056 ret = sched_rt_global_constraints();
7058 sysctl_sched_rt_period = old_period;
7059 sysctl_sched_rt_runtime = old_runtime;
7061 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7062 def_rt_bandwidth.rt_period =
7063 ns_to_ktime(global_rt_period());
7066 mutex_unlock(&mutex);
7071 #ifdef CONFIG_CGROUP_SCHED
7073 /* return corresponding task_group object of a cgroup */
7074 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7076 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7077 struct task_group, css);
7080 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7082 struct task_group *tg, *parent;
7084 if (!cgrp->parent) {
7085 /* This is early initialization for the top cgroup */
7086 return &root_task_group.css;
7089 parent = cgroup_tg(cgrp->parent);
7090 tg = sched_create_group(parent);
7092 return ERR_PTR(-ENOMEM);
7097 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7099 struct task_group *tg = cgroup_tg(cgrp);
7100 struct task_group *parent;
7105 parent = cgroup_tg(cgrp->parent);
7106 sched_online_group(tg, parent);
7110 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7112 struct task_group *tg = cgroup_tg(cgrp);
7114 sched_destroy_group(tg);
7117 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7119 struct task_group *tg = cgroup_tg(cgrp);
7121 sched_offline_group(tg);
7124 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7125 struct cgroup_taskset *tset)
7127 struct task_struct *task;
7129 cgroup_taskset_for_each(task, cgrp, tset) {
7130 #ifdef CONFIG_RT_GROUP_SCHED
7131 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7134 /* We don't support RT-tasks being in separate groups */
7135 if (task->sched_class != &fair_sched_class)
7142 static void cpu_cgroup_attach(struct cgroup *cgrp,
7143 struct cgroup_taskset *tset)
7145 struct task_struct *task;
7147 cgroup_taskset_for_each(task, cgrp, tset)
7148 sched_move_task(task);
7152 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7153 struct task_struct *task)
7156 * cgroup_exit() is called in the copy_process() failure path.
7157 * Ignore this case since the task hasn't ran yet, this avoids
7158 * trying to poke a half freed task state from generic code.
7160 if (!(task->flags & PF_EXITING))
7163 sched_move_task(task);
7166 #ifdef CONFIG_FAIR_GROUP_SCHED
7167 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7170 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7173 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7175 struct task_group *tg = cgroup_tg(cgrp);
7177 return (u64) scale_load_down(tg->shares);
7180 #ifdef CONFIG_CFS_BANDWIDTH
7181 static DEFINE_MUTEX(cfs_constraints_mutex);
7183 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7184 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7186 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7188 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7190 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7191 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7193 if (tg == &root_task_group)
7197 * Ensure we have at some amount of bandwidth every period. This is
7198 * to prevent reaching a state of large arrears when throttled via
7199 * entity_tick() resulting in prolonged exit starvation.
7201 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7205 * Likewise, bound things on the otherside by preventing insane quota
7206 * periods. This also allows us to normalize in computing quota
7209 if (period > max_cfs_quota_period)
7212 mutex_lock(&cfs_constraints_mutex);
7213 ret = __cfs_schedulable(tg, period, quota);
7217 runtime_enabled = quota != RUNTIME_INF;
7218 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7219 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7220 raw_spin_lock_irq(&cfs_b->lock);
7221 cfs_b->period = ns_to_ktime(period);
7222 cfs_b->quota = quota;
7224 __refill_cfs_bandwidth_runtime(cfs_b);
7225 /* restart the period timer (if active) to handle new period expiry */
7226 if (runtime_enabled && cfs_b->timer_active) {
7227 /* force a reprogram */
7228 cfs_b->timer_active = 0;
7229 __start_cfs_bandwidth(cfs_b);
7231 raw_spin_unlock_irq(&cfs_b->lock);
7233 for_each_possible_cpu(i) {
7234 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7235 struct rq *rq = cfs_rq->rq;
7237 raw_spin_lock_irq(&rq->lock);
7238 cfs_rq->runtime_enabled = runtime_enabled;
7239 cfs_rq->runtime_remaining = 0;
7241 if (cfs_rq->throttled)
7242 unthrottle_cfs_rq(cfs_rq);
7243 raw_spin_unlock_irq(&rq->lock);
7246 mutex_unlock(&cfs_constraints_mutex);
7251 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7255 period = ktime_to_ns(tg->cfs_bandwidth.period);
7256 if (cfs_quota_us < 0)
7257 quota = RUNTIME_INF;
7259 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7261 return tg_set_cfs_bandwidth(tg, period, quota);
7264 long tg_get_cfs_quota(struct task_group *tg)
7268 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7271 quota_us = tg->cfs_bandwidth.quota;
7272 do_div(quota_us, NSEC_PER_USEC);
7277 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7281 period = (u64)cfs_period_us * NSEC_PER_USEC;
7282 quota = tg->cfs_bandwidth.quota;
7284 return tg_set_cfs_bandwidth(tg, period, quota);
7287 long tg_get_cfs_period(struct task_group *tg)
7291 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7292 do_div(cfs_period_us, NSEC_PER_USEC);
7294 return cfs_period_us;
7297 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7299 return tg_get_cfs_quota(cgroup_tg(cgrp));
7302 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7305 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7308 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7310 return tg_get_cfs_period(cgroup_tg(cgrp));
7313 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7316 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7319 struct cfs_schedulable_data {
7320 struct task_group *tg;
7325 * normalize group quota/period to be quota/max_period
7326 * note: units are usecs
7328 static u64 normalize_cfs_quota(struct task_group *tg,
7329 struct cfs_schedulable_data *d)
7337 period = tg_get_cfs_period(tg);
7338 quota = tg_get_cfs_quota(tg);
7341 /* note: these should typically be equivalent */
7342 if (quota == RUNTIME_INF || quota == -1)
7345 return to_ratio(period, quota);
7348 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7350 struct cfs_schedulable_data *d = data;
7351 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7352 s64 quota = 0, parent_quota = -1;
7355 quota = RUNTIME_INF;
7357 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7359 quota = normalize_cfs_quota(tg, d);
7360 parent_quota = parent_b->hierarchal_quota;
7363 * ensure max(child_quota) <= parent_quota, inherit when no
7366 if (quota == RUNTIME_INF)
7367 quota = parent_quota;
7368 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7371 cfs_b->hierarchal_quota = quota;
7376 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7379 struct cfs_schedulable_data data = {
7385 if (quota != RUNTIME_INF) {
7386 do_div(data.period, NSEC_PER_USEC);
7387 do_div(data.quota, NSEC_PER_USEC);
7391 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7397 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7398 struct cgroup_map_cb *cb)
7400 struct task_group *tg = cgroup_tg(cgrp);
7401 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7403 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7404 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7405 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7409 #endif /* CONFIG_CFS_BANDWIDTH */
7410 #endif /* CONFIG_FAIR_GROUP_SCHED */
7412 #ifdef CONFIG_RT_GROUP_SCHED
7413 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7416 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7419 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7421 return sched_group_rt_runtime(cgroup_tg(cgrp));
7424 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7427 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7430 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7432 return sched_group_rt_period(cgroup_tg(cgrp));
7434 #endif /* CONFIG_RT_GROUP_SCHED */
7436 static struct cftype cpu_files[] = {
7437 #ifdef CONFIG_FAIR_GROUP_SCHED
7440 .read_u64 = cpu_shares_read_u64,
7441 .write_u64 = cpu_shares_write_u64,
7444 #ifdef CONFIG_CFS_BANDWIDTH
7446 .name = "cfs_quota_us",
7447 .read_s64 = cpu_cfs_quota_read_s64,
7448 .write_s64 = cpu_cfs_quota_write_s64,
7451 .name = "cfs_period_us",
7452 .read_u64 = cpu_cfs_period_read_u64,
7453 .write_u64 = cpu_cfs_period_write_u64,
7457 .read_map = cpu_stats_show,
7460 #ifdef CONFIG_RT_GROUP_SCHED
7462 .name = "rt_runtime_us",
7463 .read_s64 = cpu_rt_runtime_read,
7464 .write_s64 = cpu_rt_runtime_write,
7467 .name = "rt_period_us",
7468 .read_u64 = cpu_rt_period_read_uint,
7469 .write_u64 = cpu_rt_period_write_uint,
7475 struct cgroup_subsys cpu_cgroup_subsys = {
7477 .css_alloc = cpu_cgroup_css_alloc,
7478 .css_free = cpu_cgroup_css_free,
7479 .css_online = cpu_cgroup_css_online,
7480 .css_offline = cpu_cgroup_css_offline,
7481 .can_attach = cpu_cgroup_can_attach,
7482 .attach = cpu_cgroup_attach,
7483 .exit = cpu_cgroup_exit,
7484 .subsys_id = cpu_cgroup_subsys_id,
7485 .base_cftypes = cpu_files,
7489 #endif /* CONFIG_CGROUP_SCHED */
7491 void dump_cpu_task(int cpu)
7493 pr_info("Task dump for CPU %d:\n", cpu);
7494 sched_show_task(cpu_curr(cpu));