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.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 assert_raw_spin_locked(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p))
533 smp_send_reschedule(cpu);
536 void resched_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
541 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 resched_task(cpu_curr(cpu));
544 raw_spin_unlock_irqrestore(&rq->lock, flags);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu = smp_processor_id();
560 struct sched_domain *sd;
563 for_each_domain(cpu, sd) {
564 for_each_cpu(i, sched_domain_span(sd)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 static void wake_up_idle_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
589 if (cpu == smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq->curr != rq->idle)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq->idle);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq->idle))
612 smp_send_reschedule(cpu);
615 static bool wake_up_full_nohz_cpu(int cpu)
617 if (tick_nohz_full_cpu(cpu)) {
618 if (cpu != smp_processor_id() ||
619 tick_nohz_tick_stopped())
620 smp_send_reschedule(cpu);
627 void wake_up_nohz_cpu(int cpu)
629 if (!wake_up_full_nohz_cpu(cpu))
630 wake_up_idle_cpu(cpu);
633 static inline bool got_nohz_idle_kick(void)
635 int cpu = smp_processor_id();
637 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
640 if (idle_cpu(cpu) && !need_resched())
644 * We can't run Idle Load Balance on this CPU for this time so we
645 * cancel it and clear NOHZ_BALANCE_KICK
647 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
651 #else /* CONFIG_NO_HZ_COMMON */
653 static inline bool got_nohz_idle_kick(void)
658 #endif /* CONFIG_NO_HZ_COMMON */
660 #ifdef CONFIG_NO_HZ_FULL
661 bool sched_can_stop_tick(void)
667 /* Make sure rq->nr_running update is visible after the IPI */
670 /* More than one running task need preemption */
671 if (rq->nr_running > 1)
676 #endif /* CONFIG_NO_HZ_FULL */
678 void sched_avg_update(struct rq *rq)
680 s64 period = sched_avg_period();
682 while ((s64)(rq->clock - rq->age_stamp) > period) {
684 * Inline assembly required to prevent the compiler
685 * optimising this loop into a divmod call.
686 * See __iter_div_u64_rem() for another example of this.
688 asm("" : "+rm" (rq->age_stamp));
689 rq->age_stamp += period;
694 #else /* !CONFIG_SMP */
695 void resched_task(struct task_struct *p)
697 assert_raw_spin_locked(&task_rq(p)->lock);
698 set_tsk_need_resched(p);
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
713 struct task_group *parent, *child;
719 ret = (*down)(parent, data);
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
734 parent = parent->parent;
741 int tg_nop(struct task_group *tg, void *data)
747 static void set_load_weight(struct task_struct *p)
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p->policy == SCHED_IDLE) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
761 load->weight = scale_load(prio_to_weight[prio]);
762 load->inv_weight = prio_to_wmult[prio];
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
768 sched_info_queued(p);
769 p->sched_class->enqueue_task(rq, p, flags);
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
775 sched_info_dequeued(p);
776 p->sched_class->dequeue_task(rq, p, flags);
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible--;
784 enqueue_task(rq, p, flags);
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
789 if (task_contributes_to_load(p))
790 rq->nr_uninterruptible++;
792 dequeue_task(rq, p, flags);
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal = 0, irq_delta = 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta > delta)
825 rq->prev_irq_time += irq_delta;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled))) {
832 steal = paravirt_steal_clock(cpu_of(rq));
833 steal -= rq->prev_steal_time_rq;
835 if (unlikely(steal > delta))
838 st = steal_ticks(steal);
839 steal = st * TICK_NSEC;
841 rq->prev_steal_time_rq += steal;
847 rq->clock_task += delta;
849 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
850 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
851 sched_rt_avg_update(rq, irq_delta + steal);
855 void sched_set_stop_task(int cpu, struct task_struct *stop)
857 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
858 struct task_struct *old_stop = cpu_rq(cpu)->stop;
862 * Make it appear like a SCHED_FIFO task, its something
863 * userspace knows about and won't get confused about.
865 * Also, it will make PI more or less work without too
866 * much confusion -- but then, stop work should not
867 * rely on PI working anyway.
869 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
871 stop->sched_class = &stop_sched_class;
874 cpu_rq(cpu)->stop = stop;
878 * Reset it back to a normal scheduling class so that
879 * it can die in pieces.
881 old_stop->sched_class = &rt_sched_class;
886 * __normal_prio - return the priority that is based on the static prio
888 static inline int __normal_prio(struct task_struct *p)
890 return p->static_prio;
894 * Calculate the expected normal priority: i.e. priority
895 * without taking RT-inheritance into account. Might be
896 * boosted by interactivity modifiers. Changes upon fork,
897 * setprio syscalls, and whenever the interactivity
898 * estimator recalculates.
900 static inline int normal_prio(struct task_struct *p)
904 if (task_has_rt_policy(p))
905 prio = MAX_RT_PRIO-1 - p->rt_priority;
907 prio = __normal_prio(p);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct *p)
920 p->normal_prio = normal_prio(p);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p->prio))
927 return p->normal_prio;
932 * task_curr - is this task currently executing on a CPU?
933 * @p: the task in question.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio)
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
977 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
979 void register_task_migration_notifier(struct notifier_block *n)
981 atomic_notifier_chain_register(&task_migration_notifier, n);
985 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
987 #ifdef CONFIG_SCHED_DEBUG
989 * We should never call set_task_cpu() on a blocked task,
990 * ttwu() will sort out the placement.
992 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
993 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
995 #ifdef CONFIG_LOCKDEP
997 * The caller should hold either p->pi_lock or rq->lock, when changing
998 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1000 * sched_move_task() holds both and thus holding either pins the cgroup,
1003 * Furthermore, all task_rq users should acquire both locks, see
1006 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1007 lockdep_is_held(&task_rq(p)->lock)));
1011 trace_sched_migrate_task(p, new_cpu);
1013 if (task_cpu(p) != new_cpu) {
1014 struct task_migration_notifier tmn;
1016 if (p->sched_class->migrate_task_rq)
1017 p->sched_class->migrate_task_rq(p, new_cpu);
1018 p->se.nr_migrations++;
1019 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1022 tmn.from_cpu = task_cpu(p);
1023 tmn.to_cpu = new_cpu;
1025 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1028 __set_task_cpu(p, new_cpu);
1031 struct migration_arg {
1032 struct task_struct *task;
1036 static int migration_cpu_stop(void *data);
1039 * wait_task_inactive - wait for a thread to unschedule.
1041 * If @match_state is nonzero, it's the @p->state value just checked and
1042 * not expected to change. If it changes, i.e. @p might have woken up,
1043 * then return zero. When we succeed in waiting for @p to be off its CPU,
1044 * we return a positive number (its total switch count). If a second call
1045 * a short while later returns the same number, the caller can be sure that
1046 * @p has remained unscheduled the whole time.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1056 unsigned long flags;
1063 * We do the initial early heuristics without holding
1064 * any task-queue locks at all. We'll only try to get
1065 * the runqueue lock when things look like they will
1071 * If the task is actively running on another CPU
1072 * still, just relax and busy-wait without holding
1075 * NOTE! Since we don't hold any locks, it's not
1076 * even sure that "rq" stays as the right runqueue!
1077 * But we don't care, since "task_running()" will
1078 * return false if the runqueue has changed and p
1079 * is actually now running somewhere else!
1081 while (task_running(rq, p)) {
1082 if (match_state && unlikely(p->state != match_state))
1088 * Ok, time to look more closely! We need the rq
1089 * lock now, to be *sure*. If we're wrong, we'll
1090 * just go back and repeat.
1092 rq = task_rq_lock(p, &flags);
1093 trace_sched_wait_task(p);
1094 running = task_running(rq, p);
1097 if (!match_state || p->state == match_state)
1098 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1099 task_rq_unlock(rq, p, &flags);
1102 * If it changed from the expected state, bail out now.
1104 if (unlikely(!ncsw))
1108 * Was it really running after all now that we
1109 * checked with the proper locks actually held?
1111 * Oops. Go back and try again..
1113 if (unlikely(running)) {
1119 * It's not enough that it's not actively running,
1120 * it must be off the runqueue _entirely_, and not
1123 * So if it was still runnable (but just not actively
1124 * running right now), it's preempted, and we should
1125 * yield - it could be a while.
1127 if (unlikely(on_rq)) {
1128 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1130 set_current_state(TASK_UNINTERRUPTIBLE);
1131 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1136 * Ahh, all good. It wasn't running, and it wasn't
1137 * runnable, which means that it will never become
1138 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesn't have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct *p)
1165 if ((cpu != smp_processor_id()) && task_curr(p))
1166 smp_send_reschedule(cpu);
1169 EXPORT_SYMBOL_GPL(kick_process);
1170 #endif /* CONFIG_SMP */
1174 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1176 static int select_fallback_rq(int cpu, struct task_struct *p)
1178 int nid = cpu_to_node(cpu);
1179 const struct cpumask *nodemask = NULL;
1180 enum { cpuset, possible, fail } state = cpuset;
1184 * If the node that the cpu is on has been offlined, cpu_to_node()
1185 * will return -1. There is no cpu on the node, and we should
1186 * select the cpu on the other node.
1189 nodemask = cpumask_of_node(nid);
1191 /* Look for allowed, online CPU in same node. */
1192 for_each_cpu(dest_cpu, nodemask) {
1193 if (!cpu_online(dest_cpu))
1195 if (!cpu_active(dest_cpu))
1197 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1203 /* Any allowed, online CPU? */
1204 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1205 if (!cpu_online(dest_cpu))
1207 if (!cpu_active(dest_cpu))
1214 /* No more Mr. Nice Guy. */
1215 cpuset_cpus_allowed_fallback(p);
1220 do_set_cpus_allowed(p, cpu_possible_mask);
1231 if (state != cpuset) {
1233 * Don't tell them about moving exiting tasks or
1234 * kernel threads (both mm NULL), since they never
1237 if (p->mm && printk_ratelimit()) {
1238 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1239 task_pid_nr(p), p->comm, cpu);
1247 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1250 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1252 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1255 * In order not to call set_task_cpu() on a blocking task we need
1256 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1259 * Since this is common to all placement strategies, this lives here.
1261 * [ this allows ->select_task() to simply return task_cpu(p) and
1262 * not worry about this generic constraint ]
1264 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1266 cpu = select_fallback_rq(task_cpu(p), p);
1271 static void update_avg(u64 *avg, u64 sample)
1273 s64 diff = sample - *avg;
1279 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1281 #ifdef CONFIG_SCHEDSTATS
1282 struct rq *rq = this_rq();
1285 int this_cpu = smp_processor_id();
1287 if (cpu == this_cpu) {
1288 schedstat_inc(rq, ttwu_local);
1289 schedstat_inc(p, se.statistics.nr_wakeups_local);
1291 struct sched_domain *sd;
1293 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1295 for_each_domain(this_cpu, sd) {
1296 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1297 schedstat_inc(sd, ttwu_wake_remote);
1304 if (wake_flags & WF_MIGRATED)
1305 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1307 #endif /* CONFIG_SMP */
1309 schedstat_inc(rq, ttwu_count);
1310 schedstat_inc(p, se.statistics.nr_wakeups);
1312 if (wake_flags & WF_SYNC)
1313 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1315 #endif /* CONFIG_SCHEDSTATS */
1318 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1320 activate_task(rq, p, en_flags);
1323 /* if a worker is waking up, notify workqueue */
1324 if (p->flags & PF_WQ_WORKER)
1325 wq_worker_waking_up(p, cpu_of(rq));
1329 * Mark the task runnable and perform wakeup-preemption.
1332 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1334 check_preempt_curr(rq, p, wake_flags);
1335 trace_sched_wakeup(p, true);
1337 p->state = TASK_RUNNING;
1339 if (p->sched_class->task_woken)
1340 p->sched_class->task_woken(rq, p);
1342 if (rq->idle_stamp) {
1343 u64 delta = rq->clock - rq->idle_stamp;
1344 u64 max = 2*sysctl_sched_migration_cost;
1349 update_avg(&rq->avg_idle, delta);
1356 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1359 if (p->sched_contributes_to_load)
1360 rq->nr_uninterruptible--;
1363 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1364 ttwu_do_wakeup(rq, p, wake_flags);
1368 * Called in case the task @p isn't fully descheduled from its runqueue,
1369 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1370 * since all we need to do is flip p->state to TASK_RUNNING, since
1371 * the task is still ->on_rq.
1373 static int ttwu_remote(struct task_struct *p, int wake_flags)
1378 rq = __task_rq_lock(p);
1380 ttwu_do_wakeup(rq, p, wake_flags);
1383 __task_rq_unlock(rq);
1389 static void sched_ttwu_pending(void)
1391 struct rq *rq = this_rq();
1392 struct llist_node *llist = llist_del_all(&rq->wake_list);
1393 struct task_struct *p;
1395 raw_spin_lock(&rq->lock);
1398 p = llist_entry(llist, struct task_struct, wake_entry);
1399 llist = llist_next(llist);
1400 ttwu_do_activate(rq, p, 0);
1403 raw_spin_unlock(&rq->lock);
1406 void scheduler_ipi(void)
1408 if (llist_empty(&this_rq()->wake_list)
1409 && !tick_nohz_full_cpu(smp_processor_id())
1410 && !got_nohz_idle_kick())
1414 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1415 * traditionally all their work was done from the interrupt return
1416 * path. Now that we actually do some work, we need to make sure
1419 * Some archs already do call them, luckily irq_enter/exit nest
1422 * Arguably we should visit all archs and update all handlers,
1423 * however a fair share of IPIs are still resched only so this would
1424 * somewhat pessimize the simple resched case.
1427 tick_nohz_full_check();
1428 sched_ttwu_pending();
1431 * Check if someone kicked us for doing the nohz idle load balance.
1433 if (unlikely(got_nohz_idle_kick())) {
1434 this_rq()->idle_balance = 1;
1435 raise_softirq_irqoff(SCHED_SOFTIRQ);
1440 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1442 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1443 smp_send_reschedule(cpu);
1446 bool cpus_share_cache(int this_cpu, int that_cpu)
1448 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1450 #endif /* CONFIG_SMP */
1452 static void ttwu_queue(struct task_struct *p, int cpu)
1454 struct rq *rq = cpu_rq(cpu);
1456 #if defined(CONFIG_SMP)
1457 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1458 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1459 ttwu_queue_remote(p, cpu);
1464 raw_spin_lock(&rq->lock);
1465 ttwu_do_activate(rq, p, 0);
1466 raw_spin_unlock(&rq->lock);
1470 * try_to_wake_up - wake up a thread
1471 * @p: the thread to be awakened
1472 * @state: the mask of task states that can be woken
1473 * @wake_flags: wake modifier flags (WF_*)
1475 * Put it on the run-queue if it's not already there. The "current"
1476 * thread is always on the run-queue (except when the actual
1477 * re-schedule is in progress), and as such you're allowed to do
1478 * the simpler "current->state = TASK_RUNNING" to mark yourself
1479 * runnable without the overhead of this.
1481 * Returns %true if @p was woken up, %false if it was already running
1482 * or @state didn't match @p's state.
1485 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1487 unsigned long flags;
1488 int cpu, success = 0;
1491 * If we are going to wake up a thread waiting for CONDITION we
1492 * need to ensure that CONDITION=1 done by the caller can not be
1493 * reordered with p->state check below. This pairs with mb() in
1494 * set_current_state() the waiting thread does.
1496 smp_mb__before_spinlock();
1497 raw_spin_lock_irqsave(&p->pi_lock, flags);
1498 if (!(p->state & state))
1501 success = 1; /* we're going to change ->state */
1504 if (p->on_rq && ttwu_remote(p, wake_flags))
1509 * If the owning (remote) cpu is still in the middle of schedule() with
1510 * this task as prev, wait until its done referencing the task.
1515 * Pairs with the smp_wmb() in finish_lock_switch().
1519 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1520 p->state = TASK_WAKING;
1522 if (p->sched_class->task_waking)
1523 p->sched_class->task_waking(p);
1525 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1526 if (task_cpu(p) != cpu) {
1527 wake_flags |= WF_MIGRATED;
1528 set_task_cpu(p, cpu);
1530 #endif /* CONFIG_SMP */
1534 ttwu_stat(p, cpu, wake_flags);
1536 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1542 * try_to_wake_up_local - try to wake up a local task with rq lock held
1543 * @p: the thread to be awakened
1545 * Put @p on the run-queue if it's not already there. The caller must
1546 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1549 static void try_to_wake_up_local(struct task_struct *p)
1551 struct rq *rq = task_rq(p);
1553 if (WARN_ON_ONCE(rq != this_rq()) ||
1554 WARN_ON_ONCE(p == current))
1557 lockdep_assert_held(&rq->lock);
1559 if (!raw_spin_trylock(&p->pi_lock)) {
1560 raw_spin_unlock(&rq->lock);
1561 raw_spin_lock(&p->pi_lock);
1562 raw_spin_lock(&rq->lock);
1565 if (!(p->state & TASK_NORMAL))
1569 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1571 ttwu_do_wakeup(rq, p, 0);
1572 ttwu_stat(p, smp_processor_id(), 0);
1574 raw_spin_unlock(&p->pi_lock);
1578 * wake_up_process - Wake up a specific process
1579 * @p: The process to be woken up.
1581 * Attempt to wake up the nominated process and move it to the set of runnable
1582 * processes. Returns 1 if the process was woken up, 0 if it was already
1585 * It may be assumed that this function implies a write memory barrier before
1586 * changing the task state if and only if any tasks are woken up.
1588 int wake_up_process(struct task_struct *p)
1590 WARN_ON(task_is_stopped_or_traced(p));
1591 return try_to_wake_up(p, TASK_NORMAL, 0);
1593 EXPORT_SYMBOL(wake_up_process);
1595 int wake_up_state(struct task_struct *p, unsigned int state)
1597 return try_to_wake_up(p, state, 0);
1601 * Perform scheduler related setup for a newly forked process p.
1602 * p is forked by current.
1604 * __sched_fork() is basic setup used by init_idle() too:
1606 static void __sched_fork(struct task_struct *p)
1611 p->se.exec_start = 0;
1612 p->se.sum_exec_runtime = 0;
1613 p->se.prev_sum_exec_runtime = 0;
1614 p->se.nr_migrations = 0;
1616 INIT_LIST_HEAD(&p->se.group_node);
1619 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1620 * removed when useful for applications beyond shares distribution (e.g.
1623 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1624 p->se.avg.runnable_avg_period = 0;
1625 p->se.avg.runnable_avg_sum = 0;
1627 #ifdef CONFIG_SCHEDSTATS
1628 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1631 INIT_LIST_HEAD(&p->rt.run_list);
1633 #ifdef CONFIG_PREEMPT_NOTIFIERS
1634 INIT_HLIST_HEAD(&p->preempt_notifiers);
1637 #ifdef CONFIG_NUMA_BALANCING
1638 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1639 p->mm->numa_next_scan = jiffies;
1640 p->mm->numa_next_reset = jiffies;
1641 p->mm->numa_scan_seq = 0;
1644 p->node_stamp = 0ULL;
1645 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1646 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1647 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1648 p->numa_work.next = &p->numa_work;
1649 #endif /* CONFIG_NUMA_BALANCING */
1652 #ifdef CONFIG_NUMA_BALANCING
1653 #ifdef CONFIG_SCHED_DEBUG
1654 void set_numabalancing_state(bool enabled)
1657 sched_feat_set("NUMA");
1659 sched_feat_set("NO_NUMA");
1662 __read_mostly bool numabalancing_enabled;
1664 void set_numabalancing_state(bool enabled)
1666 numabalancing_enabled = enabled;
1668 #endif /* CONFIG_SCHED_DEBUG */
1669 #endif /* CONFIG_NUMA_BALANCING */
1672 * fork()/clone()-time setup:
1674 void sched_fork(struct task_struct *p)
1676 unsigned long flags;
1677 int cpu = get_cpu();
1681 * We mark the process as running here. This guarantees that
1682 * nobody will actually run it, and a signal or other external
1683 * event cannot wake it up and insert it on the runqueue either.
1685 p->state = TASK_RUNNING;
1688 * Make sure we do not leak PI boosting priority to the child.
1690 p->prio = current->normal_prio;
1693 * Revert to default priority/policy on fork if requested.
1695 if (unlikely(p->sched_reset_on_fork)) {
1696 if (task_has_rt_policy(p)) {
1697 p->policy = SCHED_NORMAL;
1698 p->static_prio = NICE_TO_PRIO(0);
1700 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1701 p->static_prio = NICE_TO_PRIO(0);
1703 p->prio = p->normal_prio = __normal_prio(p);
1707 * We don't need the reset flag anymore after the fork. It has
1708 * fulfilled its duty:
1710 p->sched_reset_on_fork = 0;
1713 if (!rt_prio(p->prio))
1714 p->sched_class = &fair_sched_class;
1716 if (p->sched_class->task_fork)
1717 p->sched_class->task_fork(p);
1720 * The child is not yet in the pid-hash so no cgroup attach races,
1721 * and the cgroup is pinned to this child due to cgroup_fork()
1722 * is ran before sched_fork().
1724 * Silence PROVE_RCU.
1726 raw_spin_lock_irqsave(&p->pi_lock, flags);
1727 set_task_cpu(p, cpu);
1728 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1730 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1731 if (likely(sched_info_on()))
1732 memset(&p->sched_info, 0, sizeof(p->sched_info));
1734 #if defined(CONFIG_SMP)
1737 #ifdef CONFIG_PREEMPT_COUNT
1738 /* Want to start with kernel preemption disabled. */
1739 task_thread_info(p)->preempt_count = 1;
1742 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1749 * wake_up_new_task - wake up a newly created task for the first time.
1751 * This function will do some initial scheduler statistics housekeeping
1752 * that must be done for every newly created context, then puts the task
1753 * on the runqueue and wakes it.
1755 void wake_up_new_task(struct task_struct *p)
1757 unsigned long flags;
1760 raw_spin_lock_irqsave(&p->pi_lock, flags);
1763 * Fork balancing, do it here and not earlier because:
1764 * - cpus_allowed can change in the fork path
1765 * - any previously selected cpu might disappear through hotplug
1767 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1770 rq = __task_rq_lock(p);
1771 activate_task(rq, p, 0);
1773 trace_sched_wakeup_new(p, true);
1774 check_preempt_curr(rq, p, WF_FORK);
1776 if (p->sched_class->task_woken)
1777 p->sched_class->task_woken(rq, p);
1779 task_rq_unlock(rq, p, &flags);
1782 #ifdef CONFIG_PREEMPT_NOTIFIERS
1785 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1786 * @notifier: notifier struct to register
1788 void preempt_notifier_register(struct preempt_notifier *notifier)
1790 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1792 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1795 * preempt_notifier_unregister - no longer interested in preemption notifications
1796 * @notifier: notifier struct to unregister
1798 * This is safe to call from within a preemption notifier.
1800 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1802 hlist_del(¬ifier->link);
1804 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1806 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1808 struct preempt_notifier *notifier;
1810 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1811 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1815 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1816 struct task_struct *next)
1818 struct preempt_notifier *notifier;
1820 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1821 notifier->ops->sched_out(notifier, next);
1824 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1826 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1831 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1832 struct task_struct *next)
1836 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1839 * prepare_task_switch - prepare to switch tasks
1840 * @rq: the runqueue preparing to switch
1841 * @prev: the current task that is being switched out
1842 * @next: the task we are going to switch to.
1844 * This is called with the rq lock held and interrupts off. It must
1845 * be paired with a subsequent finish_task_switch after the context
1848 * prepare_task_switch sets up locking and calls architecture specific
1852 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1853 struct task_struct *next)
1855 trace_sched_switch(prev, next);
1856 sched_info_switch(prev, next);
1857 perf_event_task_sched_out(prev, next);
1858 fire_sched_out_preempt_notifiers(prev, next);
1859 prepare_lock_switch(rq, next);
1860 prepare_arch_switch(next);
1864 * finish_task_switch - clean up after a task-switch
1865 * @rq: runqueue associated with task-switch
1866 * @prev: the thread we just switched away from.
1868 * finish_task_switch must be called after the context switch, paired
1869 * with a prepare_task_switch call before the context switch.
1870 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1871 * and do any other architecture-specific cleanup actions.
1873 * Note that we may have delayed dropping an mm in context_switch(). If
1874 * so, we finish that here outside of the runqueue lock. (Doing it
1875 * with the lock held can cause deadlocks; see schedule() for
1878 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1879 __releases(rq->lock)
1881 struct mm_struct *mm = rq->prev_mm;
1887 * A task struct has one reference for the use as "current".
1888 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1889 * schedule one last time. The schedule call will never return, and
1890 * the scheduled task must drop that reference.
1891 * The test for TASK_DEAD must occur while the runqueue locks are
1892 * still held, otherwise prev could be scheduled on another cpu, die
1893 * there before we look at prev->state, and then the reference would
1895 * Manfred Spraul <manfred@colorfullife.com>
1897 prev_state = prev->state;
1898 vtime_task_switch(prev);
1899 finish_arch_switch(prev);
1900 perf_event_task_sched_in(prev, current);
1901 finish_lock_switch(rq, prev);
1902 finish_arch_post_lock_switch();
1904 fire_sched_in_preempt_notifiers(current);
1907 if (unlikely(prev_state == TASK_DEAD)) {
1909 * Remove function-return probe instances associated with this
1910 * task and put them back on the free list.
1912 kprobe_flush_task(prev);
1913 put_task_struct(prev);
1916 tick_nohz_task_switch(current);
1921 /* assumes rq->lock is held */
1922 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1924 if (prev->sched_class->pre_schedule)
1925 prev->sched_class->pre_schedule(rq, prev);
1928 /* rq->lock is NOT held, but preemption is disabled */
1929 static inline void post_schedule(struct rq *rq)
1931 if (rq->post_schedule) {
1932 unsigned long flags;
1934 raw_spin_lock_irqsave(&rq->lock, flags);
1935 if (rq->curr->sched_class->post_schedule)
1936 rq->curr->sched_class->post_schedule(rq);
1937 raw_spin_unlock_irqrestore(&rq->lock, flags);
1939 rq->post_schedule = 0;
1945 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1949 static inline void post_schedule(struct rq *rq)
1956 * schedule_tail - first thing a freshly forked thread must call.
1957 * @prev: the thread we just switched away from.
1959 asmlinkage void schedule_tail(struct task_struct *prev)
1960 __releases(rq->lock)
1962 struct rq *rq = this_rq();
1964 finish_task_switch(rq, prev);
1967 * FIXME: do we need to worry about rq being invalidated by the
1972 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1973 /* In this case, finish_task_switch does not reenable preemption */
1976 if (current->set_child_tid)
1977 put_user(task_pid_vnr(current), current->set_child_tid);
1981 * context_switch - switch to the new MM and the new
1982 * thread's register state.
1985 context_switch(struct rq *rq, struct task_struct *prev,
1986 struct task_struct *next)
1988 struct mm_struct *mm, *oldmm;
1990 prepare_task_switch(rq, prev, next);
1993 oldmm = prev->active_mm;
1995 * For paravirt, this is coupled with an exit in switch_to to
1996 * combine the page table reload and the switch backend into
1999 arch_start_context_switch(prev);
2002 next->active_mm = oldmm;
2003 atomic_inc(&oldmm->mm_count);
2004 enter_lazy_tlb(oldmm, next);
2006 switch_mm(oldmm, mm, next);
2009 prev->active_mm = NULL;
2010 rq->prev_mm = oldmm;
2013 * Since the runqueue lock will be released by the next
2014 * task (which is an invalid locking op but in the case
2015 * of the scheduler it's an obvious special-case), so we
2016 * do an early lockdep release here:
2018 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2019 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2022 context_tracking_task_switch(prev, next);
2023 /* Here we just switch the register state and the stack. */
2024 switch_to(prev, next, prev);
2028 * this_rq must be evaluated again because prev may have moved
2029 * CPUs since it called schedule(), thus the 'rq' on its stack
2030 * frame will be invalid.
2032 finish_task_switch(this_rq(), prev);
2036 * nr_running and nr_context_switches:
2038 * externally visible scheduler statistics: current number of runnable
2039 * threads, total number of context switches performed since bootup.
2041 unsigned long nr_running(void)
2043 unsigned long i, sum = 0;
2045 for_each_online_cpu(i)
2046 sum += cpu_rq(i)->nr_running;
2051 unsigned long long nr_context_switches(void)
2054 unsigned long long sum = 0;
2056 for_each_possible_cpu(i)
2057 sum += cpu_rq(i)->nr_switches;
2062 unsigned long nr_iowait(void)
2064 unsigned long i, sum = 0;
2066 for_each_possible_cpu(i)
2067 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2072 unsigned long nr_iowait_cpu(int cpu)
2074 struct rq *this = cpu_rq(cpu);
2075 return atomic_read(&this->nr_iowait);
2078 unsigned long this_cpu_load(void)
2080 struct rq *this = this_rq();
2081 return this->cpu_load[0];
2086 * Global load-average calculations
2088 * We take a distributed and async approach to calculating the global load-avg
2089 * in order to minimize overhead.
2091 * The global load average is an exponentially decaying average of nr_running +
2092 * nr_uninterruptible.
2094 * Once every LOAD_FREQ:
2097 * for_each_possible_cpu(cpu)
2098 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2100 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2102 * Due to a number of reasons the above turns in the mess below:
2104 * - for_each_possible_cpu() is prohibitively expensive on machines with
2105 * serious number of cpus, therefore we need to take a distributed approach
2106 * to calculating nr_active.
2108 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2109 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2111 * So assuming nr_active := 0 when we start out -- true per definition, we
2112 * can simply take per-cpu deltas and fold those into a global accumulate
2113 * to obtain the same result. See calc_load_fold_active().
2115 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2116 * across the machine, we assume 10 ticks is sufficient time for every
2117 * cpu to have completed this task.
2119 * This places an upper-bound on the IRQ-off latency of the machine. Then
2120 * again, being late doesn't loose the delta, just wrecks the sample.
2122 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2123 * this would add another cross-cpu cacheline miss and atomic operation
2124 * to the wakeup path. Instead we increment on whatever cpu the task ran
2125 * when it went into uninterruptible state and decrement on whatever cpu
2126 * did the wakeup. This means that only the sum of nr_uninterruptible over
2127 * all cpus yields the correct result.
2129 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2132 /* Variables and functions for calc_load */
2133 static atomic_long_t calc_load_tasks;
2134 static unsigned long calc_load_update;
2135 unsigned long avenrun[3];
2136 EXPORT_SYMBOL(avenrun); /* should be removed */
2139 * get_avenrun - get the load average array
2140 * @loads: pointer to dest load array
2141 * @offset: offset to add
2142 * @shift: shift count to shift the result left
2144 * These values are estimates at best, so no need for locking.
2146 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2148 loads[0] = (avenrun[0] + offset) << shift;
2149 loads[1] = (avenrun[1] + offset) << shift;
2150 loads[2] = (avenrun[2] + offset) << shift;
2153 static long calc_load_fold_active(struct rq *this_rq)
2155 long nr_active, delta = 0;
2157 nr_active = this_rq->nr_running;
2158 nr_active += (long) this_rq->nr_uninterruptible;
2160 if (nr_active != this_rq->calc_load_active) {
2161 delta = nr_active - this_rq->calc_load_active;
2162 this_rq->calc_load_active = nr_active;
2169 * a1 = a0 * e + a * (1 - e)
2171 static unsigned long
2172 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2175 load += active * (FIXED_1 - exp);
2176 load += 1UL << (FSHIFT - 1);
2177 return load >> FSHIFT;
2180 #ifdef CONFIG_NO_HZ_COMMON
2182 * Handle NO_HZ for the global load-average.
2184 * Since the above described distributed algorithm to compute the global
2185 * load-average relies on per-cpu sampling from the tick, it is affected by
2188 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2189 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2190 * when we read the global state.
2192 * Obviously reality has to ruin such a delightfully simple scheme:
2194 * - When we go NO_HZ idle during the window, we can negate our sample
2195 * contribution, causing under-accounting.
2197 * We avoid this by keeping two idle-delta counters and flipping them
2198 * when the window starts, thus separating old and new NO_HZ load.
2200 * The only trick is the slight shift in index flip for read vs write.
2204 * |-|-----------|-|-----------|-|-----------|-|
2205 * r:0 0 1 1 0 0 1 1 0
2206 * w:0 1 1 0 0 1 1 0 0
2208 * This ensures we'll fold the old idle contribution in this window while
2209 * accumlating the new one.
2211 * - When we wake up from NO_HZ idle during the window, we push up our
2212 * contribution, since we effectively move our sample point to a known
2215 * This is solved by pushing the window forward, and thus skipping the
2216 * sample, for this cpu (effectively using the idle-delta for this cpu which
2217 * was in effect at the time the window opened). This also solves the issue
2218 * of having to deal with a cpu having been in NOHZ idle for multiple
2219 * LOAD_FREQ intervals.
2221 * When making the ILB scale, we should try to pull this in as well.
2223 static atomic_long_t calc_load_idle[2];
2224 static int calc_load_idx;
2226 static inline int calc_load_write_idx(void)
2228 int idx = calc_load_idx;
2231 * See calc_global_nohz(), if we observe the new index, we also
2232 * need to observe the new update time.
2237 * If the folding window started, make sure we start writing in the
2240 if (!time_before(jiffies, calc_load_update))
2246 static inline int calc_load_read_idx(void)
2248 return calc_load_idx & 1;
2251 void calc_load_enter_idle(void)
2253 struct rq *this_rq = this_rq();
2257 * We're going into NOHZ mode, if there's any pending delta, fold it
2258 * into the pending idle delta.
2260 delta = calc_load_fold_active(this_rq);
2262 int idx = calc_load_write_idx();
2263 atomic_long_add(delta, &calc_load_idle[idx]);
2267 void calc_load_exit_idle(void)
2269 struct rq *this_rq = this_rq();
2272 * If we're still before the sample window, we're done.
2274 if (time_before(jiffies, this_rq->calc_load_update))
2278 * We woke inside or after the sample window, this means we're already
2279 * accounted through the nohz accounting, so skip the entire deal and
2280 * sync up for the next window.
2282 this_rq->calc_load_update = calc_load_update;
2283 if (time_before(jiffies, this_rq->calc_load_update + 10))
2284 this_rq->calc_load_update += LOAD_FREQ;
2287 static long calc_load_fold_idle(void)
2289 int idx = calc_load_read_idx();
2292 if (atomic_long_read(&calc_load_idle[idx]))
2293 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2299 * fixed_power_int - compute: x^n, in O(log n) time
2301 * @x: base of the power
2302 * @frac_bits: fractional bits of @x
2303 * @n: power to raise @x to.
2305 * By exploiting the relation between the definition of the natural power
2306 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2307 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2308 * (where: n_i \elem {0, 1}, the binary vector representing n),
2309 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2310 * of course trivially computable in O(log_2 n), the length of our binary
2313 static unsigned long
2314 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2316 unsigned long result = 1UL << frac_bits;
2321 result += 1UL << (frac_bits - 1);
2322 result >>= frac_bits;
2328 x += 1UL << (frac_bits - 1);
2336 * a1 = a0 * e + a * (1 - e)
2338 * a2 = a1 * e + a * (1 - e)
2339 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2340 * = a0 * e^2 + a * (1 - e) * (1 + e)
2342 * a3 = a2 * e + a * (1 - e)
2343 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2344 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2348 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2349 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2350 * = a0 * e^n + a * (1 - e^n)
2352 * [1] application of the geometric series:
2355 * S_n := \Sum x^i = -------------
2358 static unsigned long
2359 calc_load_n(unsigned long load, unsigned long exp,
2360 unsigned long active, unsigned int n)
2363 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2367 * NO_HZ can leave us missing all per-cpu ticks calling
2368 * calc_load_account_active(), but since an idle CPU folds its delta into
2369 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2370 * in the pending idle delta if our idle period crossed a load cycle boundary.
2372 * Once we've updated the global active value, we need to apply the exponential
2373 * weights adjusted to the number of cycles missed.
2375 static void calc_global_nohz(void)
2377 long delta, active, n;
2379 if (!time_before(jiffies, calc_load_update + 10)) {
2381 * Catch-up, fold however many we are behind still
2383 delta = jiffies - calc_load_update - 10;
2384 n = 1 + (delta / LOAD_FREQ);
2386 active = atomic_long_read(&calc_load_tasks);
2387 active = active > 0 ? active * FIXED_1 : 0;
2389 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2390 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2391 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2393 calc_load_update += n * LOAD_FREQ;
2397 * Flip the idle index...
2399 * Make sure we first write the new time then flip the index, so that
2400 * calc_load_write_idx() will see the new time when it reads the new
2401 * index, this avoids a double flip messing things up.
2406 #else /* !CONFIG_NO_HZ_COMMON */
2408 static inline long calc_load_fold_idle(void) { return 0; }
2409 static inline void calc_global_nohz(void) { }
2411 #endif /* CONFIG_NO_HZ_COMMON */
2414 * calc_load - update the avenrun load estimates 10 ticks after the
2415 * CPUs have updated calc_load_tasks.
2417 void calc_global_load(unsigned long ticks)
2421 if (time_before(jiffies, calc_load_update + 10))
2425 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2427 delta = calc_load_fold_idle();
2429 atomic_long_add(delta, &calc_load_tasks);
2431 active = atomic_long_read(&calc_load_tasks);
2432 active = active > 0 ? active * FIXED_1 : 0;
2434 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2435 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2436 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2438 calc_load_update += LOAD_FREQ;
2441 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2447 * Called from update_cpu_load() to periodically update this CPU's
2450 static void calc_load_account_active(struct rq *this_rq)
2454 if (time_before(jiffies, this_rq->calc_load_update))
2457 delta = calc_load_fold_active(this_rq);
2459 atomic_long_add(delta, &calc_load_tasks);
2461 this_rq->calc_load_update += LOAD_FREQ;
2465 * End of global load-average stuff
2469 * The exact cpuload at various idx values, calculated at every tick would be
2470 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2472 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2473 * on nth tick when cpu may be busy, then we have:
2474 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2475 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2477 * decay_load_missed() below does efficient calculation of
2478 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2479 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2481 * The calculation is approximated on a 128 point scale.
2482 * degrade_zero_ticks is the number of ticks after which load at any
2483 * particular idx is approximated to be zero.
2484 * degrade_factor is a precomputed table, a row for each load idx.
2485 * Each column corresponds to degradation factor for a power of two ticks,
2486 * based on 128 point scale.
2488 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2489 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2491 * With this power of 2 load factors, we can degrade the load n times
2492 * by looking at 1 bits in n and doing as many mult/shift instead of
2493 * n mult/shifts needed by the exact degradation.
2495 #define DEGRADE_SHIFT 7
2496 static const unsigned char
2497 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2498 static const unsigned char
2499 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2500 {0, 0, 0, 0, 0, 0, 0, 0},
2501 {64, 32, 8, 0, 0, 0, 0, 0},
2502 {96, 72, 40, 12, 1, 0, 0},
2503 {112, 98, 75, 43, 15, 1, 0},
2504 {120, 112, 98, 76, 45, 16, 2} };
2507 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2508 * would be when CPU is idle and so we just decay the old load without
2509 * adding any new load.
2511 static unsigned long
2512 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2516 if (!missed_updates)
2519 if (missed_updates >= degrade_zero_ticks[idx])
2523 return load >> missed_updates;
2525 while (missed_updates) {
2526 if (missed_updates % 2)
2527 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2529 missed_updates >>= 1;
2536 * Update rq->cpu_load[] statistics. This function is usually called every
2537 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2538 * every tick. We fix it up based on jiffies.
2540 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2541 unsigned long pending_updates)
2545 this_rq->nr_load_updates++;
2547 /* Update our load: */
2548 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2549 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2550 unsigned long old_load, new_load;
2552 /* scale is effectively 1 << i now, and >> i divides by scale */
2554 old_load = this_rq->cpu_load[i];
2555 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2556 new_load = this_load;
2558 * Round up the averaging division if load is increasing. This
2559 * prevents us from getting stuck on 9 if the load is 10, for
2562 if (new_load > old_load)
2563 new_load += scale - 1;
2565 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2568 sched_avg_update(this_rq);
2571 #ifdef CONFIG_NO_HZ_COMMON
2573 * There is no sane way to deal with nohz on smp when using jiffies because the
2574 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2575 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2577 * Therefore we cannot use the delta approach from the regular tick since that
2578 * would seriously skew the load calculation. However we'll make do for those
2579 * updates happening while idle (nohz_idle_balance) or coming out of idle
2580 * (tick_nohz_idle_exit).
2582 * This means we might still be one tick off for nohz periods.
2586 * Called from nohz_idle_balance() to update the load ratings before doing the
2589 void update_idle_cpu_load(struct rq *this_rq)
2591 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2592 unsigned long load = this_rq->load.weight;
2593 unsigned long pending_updates;
2596 * bail if there's load or we're actually up-to-date.
2598 if (load || curr_jiffies == this_rq->last_load_update_tick)
2601 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2602 this_rq->last_load_update_tick = curr_jiffies;
2604 __update_cpu_load(this_rq, load, pending_updates);
2608 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2610 void update_cpu_load_nohz(void)
2612 struct rq *this_rq = this_rq();
2613 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2614 unsigned long pending_updates;
2616 if (curr_jiffies == this_rq->last_load_update_tick)
2619 raw_spin_lock(&this_rq->lock);
2620 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2621 if (pending_updates) {
2622 this_rq->last_load_update_tick = curr_jiffies;
2624 * We were idle, this means load 0, the current load might be
2625 * !0 due to remote wakeups and the sort.
2627 __update_cpu_load(this_rq, 0, pending_updates);
2629 raw_spin_unlock(&this_rq->lock);
2631 #endif /* CONFIG_NO_HZ_COMMON */
2634 * Called from scheduler_tick()
2636 static void update_cpu_load_active(struct rq *this_rq)
2639 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2641 this_rq->last_load_update_tick = jiffies;
2642 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2644 calc_load_account_active(this_rq);
2650 * sched_exec - execve() is a valuable balancing opportunity, because at
2651 * this point the task has the smallest effective memory and cache footprint.
2653 void sched_exec(void)
2655 struct task_struct *p = current;
2656 unsigned long flags;
2659 raw_spin_lock_irqsave(&p->pi_lock, flags);
2660 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2661 if (dest_cpu == smp_processor_id())
2664 if (likely(cpu_active(dest_cpu))) {
2665 struct migration_arg arg = { p, dest_cpu };
2667 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2668 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2672 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2677 DEFINE_PER_CPU(struct kernel_stat, kstat);
2678 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2680 EXPORT_PER_CPU_SYMBOL(kstat);
2681 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2684 * Return any ns on the sched_clock that have not yet been accounted in
2685 * @p in case that task is currently running.
2687 * Called with task_rq_lock() held on @rq.
2689 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2693 if (task_current(rq, p)) {
2694 update_rq_clock(rq);
2695 ns = rq->clock_task - p->se.exec_start;
2703 unsigned long long task_delta_exec(struct task_struct *p)
2705 unsigned long flags;
2709 rq = task_rq_lock(p, &flags);
2710 ns = do_task_delta_exec(p, rq);
2711 task_rq_unlock(rq, p, &flags);
2717 * Return accounted runtime for the task.
2718 * In case the task is currently running, return the runtime plus current's
2719 * pending runtime that have not been accounted yet.
2721 unsigned long long task_sched_runtime(struct task_struct *p)
2723 unsigned long flags;
2727 rq = task_rq_lock(p, &flags);
2728 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2729 task_rq_unlock(rq, p, &flags);
2735 * This function gets called by the timer code, with HZ frequency.
2736 * We call it with interrupts disabled.
2738 void scheduler_tick(void)
2740 int cpu = smp_processor_id();
2741 struct rq *rq = cpu_rq(cpu);
2742 struct task_struct *curr = rq->curr;
2746 raw_spin_lock(&rq->lock);
2747 update_rq_clock(rq);
2748 update_cpu_load_active(rq);
2749 curr->sched_class->task_tick(rq, curr, 0);
2750 raw_spin_unlock(&rq->lock);
2752 perf_event_task_tick();
2755 rq->idle_balance = idle_cpu(cpu);
2756 trigger_load_balance(rq, cpu);
2758 rq_last_tick_reset(rq);
2761 #ifdef CONFIG_NO_HZ_FULL
2763 * scheduler_tick_max_deferment
2765 * Keep at least one tick per second when a single
2766 * active task is running because the scheduler doesn't
2767 * yet completely support full dynticks environment.
2769 * This makes sure that uptime, CFS vruntime, load
2770 * balancing, etc... continue to move forward, even
2771 * with a very low granularity.
2773 u64 scheduler_tick_max_deferment(void)
2775 struct rq *rq = this_rq();
2776 unsigned long next, now = ACCESS_ONCE(jiffies);
2778 next = rq->last_sched_tick + HZ;
2780 if (time_before_eq(next, now))
2783 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2787 notrace unsigned long get_parent_ip(unsigned long addr)
2789 if (in_lock_functions(addr)) {
2790 addr = CALLER_ADDR2;
2791 if (in_lock_functions(addr))
2792 addr = CALLER_ADDR3;
2797 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2798 defined(CONFIG_PREEMPT_TRACER))
2800 void __kprobes add_preempt_count(int val)
2802 #ifdef CONFIG_DEBUG_PREEMPT
2806 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2809 preempt_count() += val;
2810 #ifdef CONFIG_DEBUG_PREEMPT
2812 * Spinlock count overflowing soon?
2814 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2817 if (preempt_count() == val)
2818 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2820 EXPORT_SYMBOL(add_preempt_count);
2822 void __kprobes sub_preempt_count(int val)
2824 #ifdef CONFIG_DEBUG_PREEMPT
2828 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2831 * Is the spinlock portion underflowing?
2833 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2834 !(preempt_count() & PREEMPT_MASK)))
2838 if (preempt_count() == val)
2839 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2840 preempt_count() -= val;
2842 EXPORT_SYMBOL(sub_preempt_count);
2847 * Print scheduling while atomic bug:
2849 static noinline void __schedule_bug(struct task_struct *prev)
2851 if (oops_in_progress)
2854 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2855 prev->comm, prev->pid, preempt_count());
2857 debug_show_held_locks(prev);
2859 if (irqs_disabled())
2860 print_irqtrace_events(prev);
2862 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2866 * Various schedule()-time debugging checks and statistics:
2868 static inline void schedule_debug(struct task_struct *prev)
2871 * Test if we are atomic. Since do_exit() needs to call into
2872 * schedule() atomically, we ignore that path for now.
2873 * Otherwise, whine if we are scheduling when we should not be.
2875 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2876 __schedule_bug(prev);
2879 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2881 schedstat_inc(this_rq(), sched_count);
2884 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2886 if (prev->on_rq || rq->skip_clock_update < 0)
2887 update_rq_clock(rq);
2888 prev->sched_class->put_prev_task(rq, prev);
2892 * Pick up the highest-prio task:
2894 static inline struct task_struct *
2895 pick_next_task(struct rq *rq)
2897 const struct sched_class *class;
2898 struct task_struct *p;
2901 * Optimization: we know that if all tasks are in
2902 * the fair class we can call that function directly:
2904 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2905 p = fair_sched_class.pick_next_task(rq);
2910 for_each_class(class) {
2911 p = class->pick_next_task(rq);
2916 BUG(); /* the idle class will always have a runnable task */
2920 * __schedule() is the main scheduler function.
2922 * The main means of driving the scheduler and thus entering this function are:
2924 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2926 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2927 * paths. For example, see arch/x86/entry_64.S.
2929 * To drive preemption between tasks, the scheduler sets the flag in timer
2930 * interrupt handler scheduler_tick().
2932 * 3. Wakeups don't really cause entry into schedule(). They add a
2933 * task to the run-queue and that's it.
2935 * Now, if the new task added to the run-queue preempts the current
2936 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2937 * called on the nearest possible occasion:
2939 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2941 * - in syscall or exception context, at the next outmost
2942 * preempt_enable(). (this might be as soon as the wake_up()'s
2945 * - in IRQ context, return from interrupt-handler to
2946 * preemptible context
2948 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2951 * - cond_resched() call
2952 * - explicit schedule() call
2953 * - return from syscall or exception to user-space
2954 * - return from interrupt-handler to user-space
2956 static void __sched __schedule(void)
2958 struct task_struct *prev, *next;
2959 unsigned long *switch_count;
2965 cpu = smp_processor_id();
2967 rcu_note_context_switch(cpu);
2970 schedule_debug(prev);
2972 if (sched_feat(HRTICK))
2976 * Make sure that signal_pending_state()->signal_pending() below
2977 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2978 * done by the caller to avoid the race with signal_wake_up().
2980 smp_mb__before_spinlock();
2981 raw_spin_lock_irq(&rq->lock);
2983 switch_count = &prev->nivcsw;
2984 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2985 if (unlikely(signal_pending_state(prev->state, prev))) {
2986 prev->state = TASK_RUNNING;
2988 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2992 * If a worker went to sleep, notify and ask workqueue
2993 * whether it wants to wake up a task to maintain
2996 if (prev->flags & PF_WQ_WORKER) {
2997 struct task_struct *to_wakeup;
2999 to_wakeup = wq_worker_sleeping(prev, cpu);
3001 try_to_wake_up_local(to_wakeup);
3004 switch_count = &prev->nvcsw;
3007 pre_schedule(rq, prev);
3009 if (unlikely(!rq->nr_running))
3010 idle_balance(cpu, rq);
3012 put_prev_task(rq, prev);
3013 next = pick_next_task(rq);
3014 clear_tsk_need_resched(prev);
3015 rq->skip_clock_update = 0;
3017 if (likely(prev != next)) {
3022 context_switch(rq, prev, next); /* unlocks the rq */
3024 * The context switch have flipped the stack from under us
3025 * and restored the local variables which were saved when
3026 * this task called schedule() in the past. prev == current
3027 * is still correct, but it can be moved to another cpu/rq.
3029 cpu = smp_processor_id();
3032 raw_spin_unlock_irq(&rq->lock);
3036 sched_preempt_enable_no_resched();
3041 static inline void sched_submit_work(struct task_struct *tsk)
3043 if (!tsk->state || tsk_is_pi_blocked(tsk))
3046 * If we are going to sleep and we have plugged IO queued,
3047 * make sure to submit it to avoid deadlocks.
3049 if (blk_needs_flush_plug(tsk))
3050 blk_schedule_flush_plug(tsk);
3053 asmlinkage void __sched schedule(void)
3055 struct task_struct *tsk = current;
3057 sched_submit_work(tsk);
3060 EXPORT_SYMBOL(schedule);
3062 #ifdef CONFIG_CONTEXT_TRACKING
3063 asmlinkage void __sched schedule_user(void)
3066 * If we come here after a random call to set_need_resched(),
3067 * or we have been woken up remotely but the IPI has not yet arrived,
3068 * we haven't yet exited the RCU idle mode. Do it here manually until
3069 * we find a better solution.
3078 * schedule_preempt_disabled - called with preemption disabled
3080 * Returns with preemption disabled. Note: preempt_count must be 1
3082 void __sched schedule_preempt_disabled(void)
3084 sched_preempt_enable_no_resched();
3089 #ifdef CONFIG_PREEMPT
3091 * this is the entry point to schedule() from in-kernel preemption
3092 * off of preempt_enable. Kernel preemptions off return from interrupt
3093 * occur there and call schedule directly.
3095 asmlinkage void __sched notrace preempt_schedule(void)
3097 struct thread_info *ti = current_thread_info();
3100 * If there is a non-zero preempt_count or interrupts are disabled,
3101 * we do not want to preempt the current task. Just return..
3103 if (likely(ti->preempt_count || irqs_disabled()))
3107 add_preempt_count_notrace(PREEMPT_ACTIVE);
3109 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3112 * Check again in case we missed a preemption opportunity
3113 * between schedule and now.
3116 } while (need_resched());
3118 EXPORT_SYMBOL(preempt_schedule);
3121 * this is the entry point to schedule() from kernel preemption
3122 * off of irq context.
3123 * Note, that this is called and return with irqs disabled. This will
3124 * protect us against recursive calling from irq.
3126 asmlinkage void __sched preempt_schedule_irq(void)
3128 struct thread_info *ti = current_thread_info();
3129 enum ctx_state prev_state;
3131 /* Catch callers which need to be fixed */
3132 BUG_ON(ti->preempt_count || !irqs_disabled());
3134 prev_state = exception_enter();
3137 add_preempt_count(PREEMPT_ACTIVE);
3140 local_irq_disable();
3141 sub_preempt_count(PREEMPT_ACTIVE);
3144 * Check again in case we missed a preemption opportunity
3145 * between schedule and now.
3148 } while (need_resched());
3150 exception_exit(prev_state);
3153 #endif /* CONFIG_PREEMPT */
3155 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3158 return try_to_wake_up(curr->private, mode, wake_flags);
3160 EXPORT_SYMBOL(default_wake_function);
3163 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3164 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3165 * number) then we wake all the non-exclusive tasks and one exclusive task.
3167 * There are circumstances in which we can try to wake a task which has already
3168 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3169 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3171 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3172 int nr_exclusive, int wake_flags, void *key)
3174 wait_queue_t *curr, *next;
3176 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3177 unsigned flags = curr->flags;
3179 if (curr->func(curr, mode, wake_flags, key) &&
3180 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3186 * __wake_up - wake up threads blocked on a waitqueue.
3188 * @mode: which threads
3189 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3190 * @key: is directly passed to the wakeup function
3192 * It may be assumed that this function implies a write memory barrier before
3193 * changing the task state if and only if any tasks are woken up.
3195 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3196 int nr_exclusive, void *key)
3198 unsigned long flags;
3200 spin_lock_irqsave(&q->lock, flags);
3201 __wake_up_common(q, mode, nr_exclusive, 0, key);
3202 spin_unlock_irqrestore(&q->lock, flags);
3204 EXPORT_SYMBOL(__wake_up);
3207 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3209 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3211 __wake_up_common(q, mode, nr, 0, NULL);
3213 EXPORT_SYMBOL_GPL(__wake_up_locked);
3215 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3217 __wake_up_common(q, mode, 1, 0, key);
3219 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3222 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3224 * @mode: which threads
3225 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3226 * @key: opaque value to be passed to wakeup targets
3228 * The sync wakeup differs that the waker knows that it will schedule
3229 * away soon, so while the target thread will be woken up, it will not
3230 * be migrated to another CPU - ie. the two threads are 'synchronized'
3231 * with each other. This can prevent needless bouncing between CPUs.
3233 * On UP it can prevent extra preemption.
3235 * It may be assumed that this function implies a write memory barrier before
3236 * changing the task state if and only if any tasks are woken up.
3238 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3239 int nr_exclusive, void *key)
3241 unsigned long flags;
3242 int wake_flags = WF_SYNC;
3247 if (unlikely(!nr_exclusive))
3250 spin_lock_irqsave(&q->lock, flags);
3251 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3252 spin_unlock_irqrestore(&q->lock, flags);
3254 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3257 * __wake_up_sync - see __wake_up_sync_key()
3259 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3261 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3263 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3266 * complete: - signals a single thread waiting on this completion
3267 * @x: holds the state of this particular completion
3269 * This will wake up a single thread waiting on this completion. Threads will be
3270 * awakened in the same order in which they were queued.
3272 * See also complete_all(), wait_for_completion() and related routines.
3274 * It may be assumed that this function implies a write memory barrier before
3275 * changing the task state if and only if any tasks are woken up.
3277 void complete(struct completion *x)
3279 unsigned long flags;
3281 spin_lock_irqsave(&x->wait.lock, flags);
3283 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3284 spin_unlock_irqrestore(&x->wait.lock, flags);
3286 EXPORT_SYMBOL(complete);
3289 * complete_all: - signals all threads waiting on this completion
3290 * @x: holds the state of this particular completion
3292 * This will wake up all threads waiting on this particular completion event.
3294 * It may be assumed that this function implies a write memory barrier before
3295 * changing the task state if and only if any tasks are woken up.
3297 void complete_all(struct completion *x)
3299 unsigned long flags;
3301 spin_lock_irqsave(&x->wait.lock, flags);
3302 x->done += UINT_MAX/2;
3303 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3304 spin_unlock_irqrestore(&x->wait.lock, flags);
3306 EXPORT_SYMBOL(complete_all);
3308 static inline long __sched
3309 do_wait_for_common(struct completion *x,
3310 long (*action)(long), long timeout, int state)
3313 DECLARE_WAITQUEUE(wait, current);
3315 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3317 if (signal_pending_state(state, current)) {
3318 timeout = -ERESTARTSYS;
3321 __set_current_state(state);
3322 spin_unlock_irq(&x->wait.lock);
3323 timeout = action(timeout);
3324 spin_lock_irq(&x->wait.lock);
3325 } while (!x->done && timeout);
3326 __remove_wait_queue(&x->wait, &wait);
3331 return timeout ?: 1;
3334 static inline long __sched
3335 __wait_for_common(struct completion *x,
3336 long (*action)(long), long timeout, int state)
3340 spin_lock_irq(&x->wait.lock);
3341 timeout = do_wait_for_common(x, action, timeout, state);
3342 spin_unlock_irq(&x->wait.lock);
3347 wait_for_common(struct completion *x, long timeout, int state)
3349 return __wait_for_common(x, schedule_timeout, timeout, state);
3353 wait_for_common_io(struct completion *x, long timeout, int state)
3355 return __wait_for_common(x, io_schedule_timeout, timeout, state);
3359 * wait_for_completion: - waits for completion of a task
3360 * @x: holds the state of this particular completion
3362 * This waits to be signaled for completion of a specific task. It is NOT
3363 * interruptible and there is no timeout.
3365 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3366 * and interrupt capability. Also see complete().
3368 void __sched wait_for_completion(struct completion *x)
3370 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3372 EXPORT_SYMBOL(wait_for_completion);
3375 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3376 * @x: holds the state of this particular completion
3377 * @timeout: timeout value in jiffies
3379 * This waits for either a completion of a specific task to be signaled or for a
3380 * specified timeout to expire. The timeout is in jiffies. It is not
3383 * The return value is 0 if timed out, and positive (at least 1, or number of
3384 * jiffies left till timeout) if completed.
3386 unsigned long __sched
3387 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3389 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3391 EXPORT_SYMBOL(wait_for_completion_timeout);
3394 * wait_for_completion_io: - waits for completion of a task
3395 * @x: holds the state of this particular completion
3397 * This waits to be signaled for completion of a specific task. It is NOT
3398 * interruptible and there is no timeout. The caller is accounted as waiting
3401 void __sched wait_for_completion_io(struct completion *x)
3403 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3405 EXPORT_SYMBOL(wait_for_completion_io);
3408 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3409 * @x: holds the state of this particular completion
3410 * @timeout: timeout value in jiffies
3412 * This waits for either a completion of a specific task to be signaled or for a
3413 * specified timeout to expire. The timeout is in jiffies. It is not
3414 * interruptible. The caller is accounted as waiting for IO.
3416 * The return value is 0 if timed out, and positive (at least 1, or number of
3417 * jiffies left till timeout) if completed.
3419 unsigned long __sched
3420 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
3422 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
3424 EXPORT_SYMBOL(wait_for_completion_io_timeout);
3427 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3428 * @x: holds the state of this particular completion
3430 * This waits for completion of a specific task to be signaled. It is
3433 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3435 int __sched wait_for_completion_interruptible(struct completion *x)
3437 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3438 if (t == -ERESTARTSYS)
3442 EXPORT_SYMBOL(wait_for_completion_interruptible);
3445 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3446 * @x: holds the state of this particular completion
3447 * @timeout: timeout value in jiffies
3449 * This waits for either a completion of a specific task to be signaled or for a
3450 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3452 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3453 * positive (at least 1, or number of jiffies left till timeout) if completed.
3456 wait_for_completion_interruptible_timeout(struct completion *x,
3457 unsigned long timeout)
3459 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3461 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3464 * wait_for_completion_killable: - waits for completion of a task (killable)
3465 * @x: holds the state of this particular completion
3467 * This waits to be signaled for completion of a specific task. It can be
3468 * interrupted by a kill signal.
3470 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3472 int __sched wait_for_completion_killable(struct completion *x)
3474 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3475 if (t == -ERESTARTSYS)
3479 EXPORT_SYMBOL(wait_for_completion_killable);
3482 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3483 * @x: holds the state of this particular completion
3484 * @timeout: timeout value in jiffies
3486 * This waits for either a completion of a specific task to be
3487 * signaled or for a specified timeout to expire. It can be
3488 * interrupted by a kill signal. The timeout is in jiffies.
3490 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3491 * positive (at least 1, or number of jiffies left till timeout) if completed.
3494 wait_for_completion_killable_timeout(struct completion *x,
3495 unsigned long timeout)
3497 return wait_for_common(x, timeout, TASK_KILLABLE);
3499 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3502 * try_wait_for_completion - try to decrement a completion without blocking
3503 * @x: completion structure
3505 * Returns: 0 if a decrement cannot be done without blocking
3506 * 1 if a decrement succeeded.
3508 * If a completion is being used as a counting completion,
3509 * attempt to decrement the counter without blocking. This
3510 * enables us to avoid waiting if the resource the completion
3511 * is protecting is not available.
3513 bool try_wait_for_completion(struct completion *x)
3515 unsigned long flags;
3518 spin_lock_irqsave(&x->wait.lock, flags);
3523 spin_unlock_irqrestore(&x->wait.lock, flags);
3526 EXPORT_SYMBOL(try_wait_for_completion);
3529 * completion_done - Test to see if a completion has any waiters
3530 * @x: completion structure
3532 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3533 * 1 if there are no waiters.
3536 bool completion_done(struct completion *x)
3538 unsigned long flags;
3541 spin_lock_irqsave(&x->wait.lock, flags);
3544 spin_unlock_irqrestore(&x->wait.lock, flags);
3547 EXPORT_SYMBOL(completion_done);
3550 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3552 unsigned long flags;
3555 init_waitqueue_entry(&wait, current);
3557 __set_current_state(state);
3559 spin_lock_irqsave(&q->lock, flags);
3560 __add_wait_queue(q, &wait);
3561 spin_unlock(&q->lock);
3562 timeout = schedule_timeout(timeout);
3563 spin_lock_irq(&q->lock);
3564 __remove_wait_queue(q, &wait);
3565 spin_unlock_irqrestore(&q->lock, flags);
3570 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3572 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3574 EXPORT_SYMBOL(interruptible_sleep_on);
3577 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3579 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3581 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3583 void __sched sleep_on(wait_queue_head_t *q)
3585 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3587 EXPORT_SYMBOL(sleep_on);
3589 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3591 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3593 EXPORT_SYMBOL(sleep_on_timeout);
3595 #ifdef CONFIG_RT_MUTEXES
3598 * rt_mutex_setprio - set the current priority of a task
3600 * @prio: prio value (kernel-internal form)
3602 * This function changes the 'effective' priority of a task. It does
3603 * not touch ->normal_prio like __setscheduler().
3605 * Used by the rt_mutex code to implement priority inheritance logic.
3607 void rt_mutex_setprio(struct task_struct *p, int prio)
3609 int oldprio, on_rq, running;
3611 const struct sched_class *prev_class;
3613 BUG_ON(prio < 0 || prio > MAX_PRIO);
3615 rq = __task_rq_lock(p);
3618 * Idle task boosting is a nono in general. There is one
3619 * exception, when PREEMPT_RT and NOHZ is active:
3621 * The idle task calls get_next_timer_interrupt() and holds
3622 * the timer wheel base->lock on the CPU and another CPU wants
3623 * to access the timer (probably to cancel it). We can safely
3624 * ignore the boosting request, as the idle CPU runs this code
3625 * with interrupts disabled and will complete the lock
3626 * protected section without being interrupted. So there is no
3627 * real need to boost.
3629 if (unlikely(p == rq->idle)) {
3630 WARN_ON(p != rq->curr);
3631 WARN_ON(p->pi_blocked_on);
3635 trace_sched_pi_setprio(p, prio);
3637 prev_class = p->sched_class;
3639 running = task_current(rq, p);
3641 dequeue_task(rq, p, 0);
3643 p->sched_class->put_prev_task(rq, p);
3646 p->sched_class = &rt_sched_class;
3648 p->sched_class = &fair_sched_class;
3653 p->sched_class->set_curr_task(rq);
3655 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3657 check_class_changed(rq, p, prev_class, oldprio);
3659 __task_rq_unlock(rq);
3662 void set_user_nice(struct task_struct *p, long nice)
3664 int old_prio, delta, on_rq;
3665 unsigned long flags;
3668 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3671 * We have to be careful, if called from sys_setpriority(),
3672 * the task might be in the middle of scheduling on another CPU.
3674 rq = task_rq_lock(p, &flags);
3676 * The RT priorities are set via sched_setscheduler(), but we still
3677 * allow the 'normal' nice value to be set - but as expected
3678 * it wont have any effect on scheduling until the task is
3679 * SCHED_FIFO/SCHED_RR:
3681 if (task_has_rt_policy(p)) {
3682 p->static_prio = NICE_TO_PRIO(nice);
3687 dequeue_task(rq, p, 0);
3689 p->static_prio = NICE_TO_PRIO(nice);
3692 p->prio = effective_prio(p);
3693 delta = p->prio - old_prio;
3696 enqueue_task(rq, p, 0);
3698 * If the task increased its priority or is running and
3699 * lowered its priority, then reschedule its CPU:
3701 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3702 resched_task(rq->curr);
3705 task_rq_unlock(rq, p, &flags);
3707 EXPORT_SYMBOL(set_user_nice);
3710 * can_nice - check if a task can reduce its nice value
3714 int can_nice(const struct task_struct *p, const int nice)
3716 /* convert nice value [19,-20] to rlimit style value [1,40] */
3717 int nice_rlim = 20 - nice;
3719 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3720 capable(CAP_SYS_NICE));
3723 #ifdef __ARCH_WANT_SYS_NICE
3726 * sys_nice - change the priority of the current process.
3727 * @increment: priority increment
3729 * sys_setpriority is a more generic, but much slower function that
3730 * does similar things.
3732 SYSCALL_DEFINE1(nice, int, increment)
3737 * Setpriority might change our priority at the same moment.
3738 * We don't have to worry. Conceptually one call occurs first
3739 * and we have a single winner.
3741 if (increment < -40)
3746 nice = TASK_NICE(current) + increment;
3752 if (increment < 0 && !can_nice(current, nice))
3755 retval = security_task_setnice(current, nice);
3759 set_user_nice(current, nice);
3766 * task_prio - return the priority value of a given task.
3767 * @p: the task in question.
3769 * This is the priority value as seen by users in /proc.
3770 * RT tasks are offset by -200. Normal tasks are centered
3771 * around 0, value goes from -16 to +15.
3773 int task_prio(const struct task_struct *p)
3775 return p->prio - MAX_RT_PRIO;
3779 * task_nice - return the nice value of a given task.
3780 * @p: the task in question.
3782 int task_nice(const struct task_struct *p)
3784 return TASK_NICE(p);
3786 EXPORT_SYMBOL(task_nice);
3789 * idle_cpu - is a given cpu idle currently?
3790 * @cpu: the processor in question.
3792 int idle_cpu(int cpu)
3794 struct rq *rq = cpu_rq(cpu);
3796 if (rq->curr != rq->idle)
3803 if (!llist_empty(&rq->wake_list))
3811 * idle_task - return the idle task for a given cpu.
3812 * @cpu: the processor in question.
3814 struct task_struct *idle_task(int cpu)
3816 return cpu_rq(cpu)->idle;
3820 * find_process_by_pid - find a process with a matching PID value.
3821 * @pid: the pid in question.
3823 static struct task_struct *find_process_by_pid(pid_t pid)
3825 return pid ? find_task_by_vpid(pid) : current;
3828 /* Actually do priority change: must hold rq lock. */
3830 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3833 p->rt_priority = prio;
3834 p->normal_prio = normal_prio(p);
3835 /* we are holding p->pi_lock already */
3836 p->prio = rt_mutex_getprio(p);
3837 if (rt_prio(p->prio))
3838 p->sched_class = &rt_sched_class;
3840 p->sched_class = &fair_sched_class;
3845 * check the target process has a UID that matches the current process's
3847 static bool check_same_owner(struct task_struct *p)
3849 const struct cred *cred = current_cred(), *pcred;
3853 pcred = __task_cred(p);
3854 match = (uid_eq(cred->euid, pcred->euid) ||
3855 uid_eq(cred->euid, pcred->uid));
3860 static int __sched_setscheduler(struct task_struct *p, int policy,
3861 const struct sched_param *param, bool user)
3863 int retval, oldprio, oldpolicy = -1, on_rq, running;
3864 unsigned long flags;
3865 const struct sched_class *prev_class;
3869 /* may grab non-irq protected spin_locks */
3870 BUG_ON(in_interrupt());
3872 /* double check policy once rq lock held */
3874 reset_on_fork = p->sched_reset_on_fork;
3875 policy = oldpolicy = p->policy;
3877 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3878 policy &= ~SCHED_RESET_ON_FORK;
3880 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3881 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3882 policy != SCHED_IDLE)
3887 * Valid priorities for SCHED_FIFO and SCHED_RR are
3888 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3889 * SCHED_BATCH and SCHED_IDLE is 0.
3891 if (param->sched_priority < 0 ||
3892 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3893 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3895 if (rt_policy(policy) != (param->sched_priority != 0))
3899 * Allow unprivileged RT tasks to decrease priority:
3901 if (user && !capable(CAP_SYS_NICE)) {
3902 if (rt_policy(policy)) {
3903 unsigned long rlim_rtprio =
3904 task_rlimit(p, RLIMIT_RTPRIO);
3906 /* can't set/change the rt policy */
3907 if (policy != p->policy && !rlim_rtprio)
3910 /* can't increase priority */
3911 if (param->sched_priority > p->rt_priority &&
3912 param->sched_priority > rlim_rtprio)
3917 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3918 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3920 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3921 if (!can_nice(p, TASK_NICE(p)))
3925 /* can't change other user's priorities */
3926 if (!check_same_owner(p))
3929 /* Normal users shall not reset the sched_reset_on_fork flag */
3930 if (p->sched_reset_on_fork && !reset_on_fork)
3935 retval = security_task_setscheduler(p);
3941 * make sure no PI-waiters arrive (or leave) while we are
3942 * changing the priority of the task:
3944 * To be able to change p->policy safely, the appropriate
3945 * runqueue lock must be held.
3947 rq = task_rq_lock(p, &flags);
3950 * Changing the policy of the stop threads its a very bad idea
3952 if (p == rq->stop) {
3953 task_rq_unlock(rq, p, &flags);
3958 * If not changing anything there's no need to proceed further:
3960 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3961 param->sched_priority == p->rt_priority))) {
3962 task_rq_unlock(rq, p, &flags);
3966 #ifdef CONFIG_RT_GROUP_SCHED
3969 * Do not allow realtime tasks into groups that have no runtime
3972 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3973 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3974 !task_group_is_autogroup(task_group(p))) {
3975 task_rq_unlock(rq, p, &flags);
3981 /* recheck policy now with rq lock held */
3982 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3983 policy = oldpolicy = -1;
3984 task_rq_unlock(rq, p, &flags);
3988 running = task_current(rq, p);
3990 dequeue_task(rq, p, 0);
3992 p->sched_class->put_prev_task(rq, p);
3994 p->sched_reset_on_fork = reset_on_fork;
3997 prev_class = p->sched_class;
3998 __setscheduler(rq, p, policy, param->sched_priority);
4001 p->sched_class->set_curr_task(rq);
4003 enqueue_task(rq, p, 0);
4005 check_class_changed(rq, p, prev_class, oldprio);
4006 task_rq_unlock(rq, p, &flags);
4008 rt_mutex_adjust_pi(p);
4014 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4015 * @p: the task in question.
4016 * @policy: new policy.
4017 * @param: structure containing the new RT priority.
4019 * NOTE that the task may be already dead.
4021 int sched_setscheduler(struct task_struct *p, int policy,
4022 const struct sched_param *param)
4024 return __sched_setscheduler(p, policy, param, true);
4026 EXPORT_SYMBOL_GPL(sched_setscheduler);
4029 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4030 * @p: the task in question.
4031 * @policy: new policy.
4032 * @param: structure containing the new RT priority.
4034 * Just like sched_setscheduler, only don't bother checking if the
4035 * current context has permission. For example, this is needed in
4036 * stop_machine(): we create temporary high priority worker threads,
4037 * but our caller might not have that capability.
4039 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4040 const struct sched_param *param)
4042 return __sched_setscheduler(p, policy, param, false);
4046 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4048 struct sched_param lparam;
4049 struct task_struct *p;
4052 if (!param || pid < 0)
4054 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4059 p = find_process_by_pid(pid);
4061 retval = sched_setscheduler(p, policy, &lparam);
4068 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4069 * @pid: the pid in question.
4070 * @policy: new policy.
4071 * @param: structure containing the new RT priority.
4073 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4074 struct sched_param __user *, param)
4076 /* negative values for policy are not valid */
4080 return do_sched_setscheduler(pid, policy, param);
4084 * sys_sched_setparam - set/change the RT priority of a thread
4085 * @pid: the pid in question.
4086 * @param: structure containing the new RT priority.
4088 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4090 return do_sched_setscheduler(pid, -1, param);
4094 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4095 * @pid: the pid in question.
4097 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4099 struct task_struct *p;
4107 p = find_process_by_pid(pid);
4109 retval = security_task_getscheduler(p);
4112 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4119 * sys_sched_getparam - get the RT priority of a thread
4120 * @pid: the pid in question.
4121 * @param: structure containing the RT priority.
4123 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4125 struct sched_param lp;
4126 struct task_struct *p;
4129 if (!param || pid < 0)
4133 p = find_process_by_pid(pid);
4138 retval = security_task_getscheduler(p);
4142 lp.sched_priority = p->rt_priority;
4146 * This one might sleep, we cannot do it with a spinlock held ...
4148 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4157 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4159 cpumask_var_t cpus_allowed, new_mask;
4160 struct task_struct *p;
4166 p = find_process_by_pid(pid);
4173 /* Prevent p going away */
4177 if (p->flags & PF_NO_SETAFFINITY) {
4181 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4185 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4187 goto out_free_cpus_allowed;
4190 if (!check_same_owner(p)) {
4192 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4199 retval = security_task_setscheduler(p);
4203 cpuset_cpus_allowed(p, cpus_allowed);
4204 cpumask_and(new_mask, in_mask, cpus_allowed);
4206 retval = set_cpus_allowed_ptr(p, new_mask);
4209 cpuset_cpus_allowed(p, cpus_allowed);
4210 if (!cpumask_subset(new_mask, cpus_allowed)) {
4212 * We must have raced with a concurrent cpuset
4213 * update. Just reset the cpus_allowed to the
4214 * cpuset's cpus_allowed
4216 cpumask_copy(new_mask, cpus_allowed);
4221 free_cpumask_var(new_mask);
4222 out_free_cpus_allowed:
4223 free_cpumask_var(cpus_allowed);
4230 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4231 struct cpumask *new_mask)
4233 if (len < cpumask_size())
4234 cpumask_clear(new_mask);
4235 else if (len > cpumask_size())
4236 len = cpumask_size();
4238 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4242 * sys_sched_setaffinity - set the cpu affinity of a process
4243 * @pid: pid of the process
4244 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4245 * @user_mask_ptr: user-space pointer to the new cpu mask
4247 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4248 unsigned long __user *, user_mask_ptr)
4250 cpumask_var_t new_mask;
4253 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4256 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4258 retval = sched_setaffinity(pid, new_mask);
4259 free_cpumask_var(new_mask);
4263 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4265 struct task_struct *p;
4266 unsigned long flags;
4273 p = find_process_by_pid(pid);
4277 retval = security_task_getscheduler(p);
4281 raw_spin_lock_irqsave(&p->pi_lock, flags);
4282 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4283 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4293 * sys_sched_getaffinity - get the cpu affinity of a process
4294 * @pid: pid of the process
4295 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4296 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4298 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4299 unsigned long __user *, user_mask_ptr)
4304 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4306 if (len & (sizeof(unsigned long)-1))
4309 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4312 ret = sched_getaffinity(pid, mask);
4314 size_t retlen = min_t(size_t, len, cpumask_size());
4316 if (copy_to_user(user_mask_ptr, mask, retlen))
4321 free_cpumask_var(mask);
4327 * sys_sched_yield - yield the current processor to other threads.
4329 * This function yields the current CPU to other tasks. If there are no
4330 * other threads running on this CPU then this function will return.
4332 SYSCALL_DEFINE0(sched_yield)
4334 struct rq *rq = this_rq_lock();
4336 schedstat_inc(rq, yld_count);
4337 current->sched_class->yield_task(rq);
4340 * Since we are going to call schedule() anyway, there's
4341 * no need to preempt or enable interrupts:
4343 __release(rq->lock);
4344 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4345 do_raw_spin_unlock(&rq->lock);
4346 sched_preempt_enable_no_resched();
4353 static inline int should_resched(void)
4355 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4358 static void __cond_resched(void)
4360 add_preempt_count(PREEMPT_ACTIVE);
4362 sub_preempt_count(PREEMPT_ACTIVE);
4365 int __sched _cond_resched(void)
4367 if (should_resched()) {
4373 EXPORT_SYMBOL(_cond_resched);
4376 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4377 * call schedule, and on return reacquire the lock.
4379 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4380 * operations here to prevent schedule() from being called twice (once via
4381 * spin_unlock(), once by hand).
4383 int __cond_resched_lock(spinlock_t *lock)
4385 int resched = should_resched();
4388 lockdep_assert_held(lock);
4390 if (spin_needbreak(lock) || resched) {
4401 EXPORT_SYMBOL(__cond_resched_lock);
4403 int __sched __cond_resched_softirq(void)
4405 BUG_ON(!in_softirq());
4407 if (should_resched()) {
4415 EXPORT_SYMBOL(__cond_resched_softirq);
4418 * yield - yield the current processor to other threads.
4420 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4422 * The scheduler is at all times free to pick the calling task as the most
4423 * eligible task to run, if removing the yield() call from your code breaks
4424 * it, its already broken.
4426 * Typical broken usage is:
4431 * where one assumes that yield() will let 'the other' process run that will
4432 * make event true. If the current task is a SCHED_FIFO task that will never
4433 * happen. Never use yield() as a progress guarantee!!
4435 * If you want to use yield() to wait for something, use wait_event().
4436 * If you want to use yield() to be 'nice' for others, use cond_resched().
4437 * If you still want to use yield(), do not!
4439 void __sched yield(void)
4441 set_current_state(TASK_RUNNING);
4444 EXPORT_SYMBOL(yield);
4447 * yield_to - yield the current processor to another thread in
4448 * your thread group, or accelerate that thread toward the
4449 * processor it's on.
4451 * @preempt: whether task preemption is allowed or not
4453 * It's the caller's job to ensure that the target task struct
4454 * can't go away on us before we can do any checks.
4457 * true (>0) if we indeed boosted the target task.
4458 * false (0) if we failed to boost the target.
4459 * -ESRCH if there's no task to yield to.
4461 bool __sched yield_to(struct task_struct *p, bool preempt)
4463 struct task_struct *curr = current;
4464 struct rq *rq, *p_rq;
4465 unsigned long flags;
4468 local_irq_save(flags);
4474 * If we're the only runnable task on the rq and target rq also
4475 * has only one task, there's absolutely no point in yielding.
4477 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4482 double_rq_lock(rq, p_rq);
4483 while (task_rq(p) != p_rq) {
4484 double_rq_unlock(rq, p_rq);
4488 if (!curr->sched_class->yield_to_task)
4491 if (curr->sched_class != p->sched_class)
4494 if (task_running(p_rq, p) || p->state)
4497 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4499 schedstat_inc(rq, yld_count);
4501 * Make p's CPU reschedule; pick_next_entity takes care of
4504 if (preempt && rq != p_rq)
4505 resched_task(p_rq->curr);
4509 double_rq_unlock(rq, p_rq);
4511 local_irq_restore(flags);
4518 EXPORT_SYMBOL_GPL(yield_to);
4521 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4522 * that process accounting knows that this is a task in IO wait state.
4524 void __sched io_schedule(void)
4526 struct rq *rq = raw_rq();
4528 delayacct_blkio_start();
4529 atomic_inc(&rq->nr_iowait);
4530 blk_flush_plug(current);
4531 current->in_iowait = 1;
4533 current->in_iowait = 0;
4534 atomic_dec(&rq->nr_iowait);
4535 delayacct_blkio_end();
4537 EXPORT_SYMBOL(io_schedule);
4539 long __sched io_schedule_timeout(long timeout)
4541 struct rq *rq = raw_rq();
4544 delayacct_blkio_start();
4545 atomic_inc(&rq->nr_iowait);
4546 blk_flush_plug(current);
4547 current->in_iowait = 1;
4548 ret = schedule_timeout(timeout);
4549 current->in_iowait = 0;
4550 atomic_dec(&rq->nr_iowait);
4551 delayacct_blkio_end();
4556 * sys_sched_get_priority_max - return maximum RT priority.
4557 * @policy: scheduling class.
4559 * this syscall returns the maximum rt_priority that can be used
4560 * by a given scheduling class.
4562 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4569 ret = MAX_USER_RT_PRIO-1;
4581 * sys_sched_get_priority_min - return minimum RT priority.
4582 * @policy: scheduling class.
4584 * this syscall returns the minimum rt_priority that can be used
4585 * by a given scheduling class.
4587 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4605 * sys_sched_rr_get_interval - return the default timeslice of a process.
4606 * @pid: pid of the process.
4607 * @interval: userspace pointer to the timeslice value.
4609 * this syscall writes the default timeslice value of a given process
4610 * into the user-space timespec buffer. A value of '0' means infinity.
4612 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4613 struct timespec __user *, interval)
4615 struct task_struct *p;
4616 unsigned int time_slice;
4617 unsigned long flags;
4627 p = find_process_by_pid(pid);
4631 retval = security_task_getscheduler(p);
4635 rq = task_rq_lock(p, &flags);
4636 time_slice = p->sched_class->get_rr_interval(rq, p);
4637 task_rq_unlock(rq, p, &flags);
4640 jiffies_to_timespec(time_slice, &t);
4641 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4649 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4651 void sched_show_task(struct task_struct *p)
4653 unsigned long free = 0;
4657 state = p->state ? __ffs(p->state) + 1 : 0;
4658 printk(KERN_INFO "%-15.15s %c", p->comm,
4659 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4660 #if BITS_PER_LONG == 32
4661 if (state == TASK_RUNNING)
4662 printk(KERN_CONT " running ");
4664 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4666 if (state == TASK_RUNNING)
4667 printk(KERN_CONT " running task ");
4669 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4671 #ifdef CONFIG_DEBUG_STACK_USAGE
4672 free = stack_not_used(p);
4675 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4677 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4678 task_pid_nr(p), ppid,
4679 (unsigned long)task_thread_info(p)->flags);
4681 print_worker_info(KERN_INFO, p);
4682 show_stack(p, NULL);
4685 void show_state_filter(unsigned long state_filter)
4687 struct task_struct *g, *p;
4689 #if BITS_PER_LONG == 32
4691 " task PC stack pid father\n");
4694 " task PC stack pid father\n");
4697 do_each_thread(g, p) {
4699 * reset the NMI-timeout, listing all files on a slow
4700 * console might take a lot of time:
4702 touch_nmi_watchdog();
4703 if (!state_filter || (p->state & state_filter))
4705 } while_each_thread(g, p);
4707 touch_all_softlockup_watchdogs();
4709 #ifdef CONFIG_SCHED_DEBUG
4710 sysrq_sched_debug_show();
4714 * Only show locks if all tasks are dumped:
4717 debug_show_all_locks();
4720 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4722 idle->sched_class = &idle_sched_class;
4726 * init_idle - set up an idle thread for a given CPU
4727 * @idle: task in question
4728 * @cpu: cpu the idle task belongs to
4730 * NOTE: this function does not set the idle thread's NEED_RESCHED
4731 * flag, to make booting more robust.
4733 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4735 struct rq *rq = cpu_rq(cpu);
4736 unsigned long flags;
4738 raw_spin_lock_irqsave(&rq->lock, flags);
4741 idle->state = TASK_RUNNING;
4742 idle->se.exec_start = sched_clock();
4744 do_set_cpus_allowed(idle, cpumask_of(cpu));
4746 * We're having a chicken and egg problem, even though we are
4747 * holding rq->lock, the cpu isn't yet set to this cpu so the
4748 * lockdep check in task_group() will fail.
4750 * Similar case to sched_fork(). / Alternatively we could
4751 * use task_rq_lock() here and obtain the other rq->lock.
4756 __set_task_cpu(idle, cpu);
4759 rq->curr = rq->idle = idle;
4760 #if defined(CONFIG_SMP)
4763 raw_spin_unlock_irqrestore(&rq->lock, flags);
4765 /* Set the preempt count _outside_ the spinlocks! */
4766 task_thread_info(idle)->preempt_count = 0;
4769 * The idle tasks have their own, simple scheduling class:
4771 idle->sched_class = &idle_sched_class;
4772 ftrace_graph_init_idle_task(idle, cpu);
4773 vtime_init_idle(idle, cpu);
4774 #if defined(CONFIG_SMP)
4775 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4780 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4782 if (p->sched_class && p->sched_class->set_cpus_allowed)
4783 p->sched_class->set_cpus_allowed(p, new_mask);
4785 cpumask_copy(&p->cpus_allowed, new_mask);
4786 p->nr_cpus_allowed = cpumask_weight(new_mask);
4790 * This is how migration works:
4792 * 1) we invoke migration_cpu_stop() on the target CPU using
4794 * 2) stopper starts to run (implicitly forcing the migrated thread
4796 * 3) it checks whether the migrated task is still in the wrong runqueue.
4797 * 4) if it's in the wrong runqueue then the migration thread removes
4798 * it and puts it into the right queue.
4799 * 5) stopper completes and stop_one_cpu() returns and the migration
4804 * Change a given task's CPU affinity. Migrate the thread to a
4805 * proper CPU and schedule it away if the CPU it's executing on
4806 * is removed from the allowed bitmask.
4808 * NOTE: the caller must have a valid reference to the task, the
4809 * task must not exit() & deallocate itself prematurely. The
4810 * call is not atomic; no spinlocks may be held.
4812 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4814 unsigned long flags;
4816 unsigned int dest_cpu;
4819 rq = task_rq_lock(p, &flags);
4821 if (cpumask_equal(&p->cpus_allowed, new_mask))
4824 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4829 do_set_cpus_allowed(p, new_mask);
4831 /* Can the task run on the task's current CPU? If so, we're done */
4832 if (cpumask_test_cpu(task_cpu(p), new_mask))
4835 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4837 struct migration_arg arg = { p, dest_cpu };
4838 /* Need help from migration thread: drop lock and wait. */
4839 task_rq_unlock(rq, p, &flags);
4840 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4841 tlb_migrate_finish(p->mm);
4845 task_rq_unlock(rq, p, &flags);
4849 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4852 * Move (not current) task off this cpu, onto dest cpu. We're doing
4853 * this because either it can't run here any more (set_cpus_allowed()
4854 * away from this CPU, or CPU going down), or because we're
4855 * attempting to rebalance this task on exec (sched_exec).
4857 * So we race with normal scheduler movements, but that's OK, as long
4858 * as the task is no longer on this CPU.
4860 * Returns non-zero if task was successfully migrated.
4862 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4864 struct rq *rq_dest, *rq_src;
4867 if (unlikely(!cpu_active(dest_cpu)))
4870 rq_src = cpu_rq(src_cpu);
4871 rq_dest = cpu_rq(dest_cpu);
4873 raw_spin_lock(&p->pi_lock);
4874 double_rq_lock(rq_src, rq_dest);
4875 /* Already moved. */
4876 if (task_cpu(p) != src_cpu)
4878 /* Affinity changed (again). */
4879 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4883 * If we're not on a rq, the next wake-up will ensure we're
4887 dequeue_task(rq_src, p, 0);
4888 set_task_cpu(p, dest_cpu);
4889 enqueue_task(rq_dest, p, 0);
4890 check_preempt_curr(rq_dest, p, 0);
4895 double_rq_unlock(rq_src, rq_dest);
4896 raw_spin_unlock(&p->pi_lock);
4901 * migration_cpu_stop - this will be executed by a highprio stopper thread
4902 * and performs thread migration by bumping thread off CPU then
4903 * 'pushing' onto another runqueue.
4905 static int migration_cpu_stop(void *data)
4907 struct migration_arg *arg = data;
4910 * The original target cpu might have gone down and we might
4911 * be on another cpu but it doesn't matter.
4913 local_irq_disable();
4914 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4919 #ifdef CONFIG_HOTPLUG_CPU
4922 * Ensures that the idle task is using init_mm right before its cpu goes
4925 void idle_task_exit(void)
4927 struct mm_struct *mm = current->active_mm;
4929 BUG_ON(cpu_online(smp_processor_id()));
4932 switch_mm(mm, &init_mm, current);
4937 * Since this CPU is going 'away' for a while, fold any nr_active delta
4938 * we might have. Assumes we're called after migrate_tasks() so that the
4939 * nr_active count is stable.
4941 * Also see the comment "Global load-average calculations".
4943 static void calc_load_migrate(struct rq *rq)
4945 long delta = calc_load_fold_active(rq);
4947 atomic_long_add(delta, &calc_load_tasks);
4951 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4952 * try_to_wake_up()->select_task_rq().
4954 * Called with rq->lock held even though we'er in stop_machine() and
4955 * there's no concurrency possible, we hold the required locks anyway
4956 * because of lock validation efforts.
4958 static void migrate_tasks(unsigned int dead_cpu)
4960 struct rq *rq = cpu_rq(dead_cpu);
4961 struct task_struct *next, *stop = rq->stop;
4965 * Fudge the rq selection such that the below task selection loop
4966 * doesn't get stuck on the currently eligible stop task.
4968 * We're currently inside stop_machine() and the rq is either stuck
4969 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4970 * either way we should never end up calling schedule() until we're
4977 * There's this thread running, bail when that's the only
4980 if (rq->nr_running == 1)
4983 next = pick_next_task(rq);
4985 next->sched_class->put_prev_task(rq, next);
4987 /* Find suitable destination for @next, with force if needed. */
4988 dest_cpu = select_fallback_rq(dead_cpu, next);
4989 raw_spin_unlock(&rq->lock);
4991 __migrate_task(next, dead_cpu, dest_cpu);
4993 raw_spin_lock(&rq->lock);
4999 #endif /* CONFIG_HOTPLUG_CPU */
5001 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5003 static struct ctl_table sd_ctl_dir[] = {
5005 .procname = "sched_domain",
5011 static struct ctl_table sd_ctl_root[] = {
5013 .procname = "kernel",
5015 .child = sd_ctl_dir,
5020 static struct ctl_table *sd_alloc_ctl_entry(int n)
5022 struct ctl_table *entry =
5023 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5028 static void sd_free_ctl_entry(struct ctl_table **tablep)
5030 struct ctl_table *entry;
5033 * In the intermediate directories, both the child directory and
5034 * procname are dynamically allocated and could fail but the mode
5035 * will always be set. In the lowest directory the names are
5036 * static strings and all have proc handlers.
5038 for (entry = *tablep; entry->mode; entry++) {
5040 sd_free_ctl_entry(&entry->child);
5041 if (entry->proc_handler == NULL)
5042 kfree(entry->procname);
5049 static int min_load_idx = 0;
5050 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5053 set_table_entry(struct ctl_table *entry,
5054 const char *procname, void *data, int maxlen,
5055 umode_t mode, proc_handler *proc_handler,
5058 entry->procname = procname;
5060 entry->maxlen = maxlen;
5062 entry->proc_handler = proc_handler;
5065 entry->extra1 = &min_load_idx;
5066 entry->extra2 = &max_load_idx;
5070 static struct ctl_table *
5071 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5073 struct ctl_table *table = sd_alloc_ctl_entry(13);
5078 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5079 sizeof(long), 0644, proc_doulongvec_minmax, false);
5080 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5081 sizeof(long), 0644, proc_doulongvec_minmax, false);
5082 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5083 sizeof(int), 0644, proc_dointvec_minmax, true);
5084 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5085 sizeof(int), 0644, proc_dointvec_minmax, true);
5086 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5087 sizeof(int), 0644, proc_dointvec_minmax, true);
5088 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5089 sizeof(int), 0644, proc_dointvec_minmax, true);
5090 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5091 sizeof(int), 0644, proc_dointvec_minmax, true);
5092 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5093 sizeof(int), 0644, proc_dointvec_minmax, false);
5094 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5095 sizeof(int), 0644, proc_dointvec_minmax, false);
5096 set_table_entry(&table[9], "cache_nice_tries",
5097 &sd->cache_nice_tries,
5098 sizeof(int), 0644, proc_dointvec_minmax, false);
5099 set_table_entry(&table[10], "flags", &sd->flags,
5100 sizeof(int), 0644, proc_dointvec_minmax, false);
5101 set_table_entry(&table[11], "name", sd->name,
5102 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5103 /* &table[12] is terminator */
5108 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5110 struct ctl_table *entry, *table;
5111 struct sched_domain *sd;
5112 int domain_num = 0, i;
5115 for_each_domain(cpu, sd)
5117 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5122 for_each_domain(cpu, sd) {
5123 snprintf(buf, 32, "domain%d", i);
5124 entry->procname = kstrdup(buf, GFP_KERNEL);
5126 entry->child = sd_alloc_ctl_domain_table(sd);
5133 static struct ctl_table_header *sd_sysctl_header;
5134 static void register_sched_domain_sysctl(void)
5136 int i, cpu_num = num_possible_cpus();
5137 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5140 WARN_ON(sd_ctl_dir[0].child);
5141 sd_ctl_dir[0].child = entry;
5146 for_each_possible_cpu(i) {
5147 snprintf(buf, 32, "cpu%d", i);
5148 entry->procname = kstrdup(buf, GFP_KERNEL);
5150 entry->child = sd_alloc_ctl_cpu_table(i);
5154 WARN_ON(sd_sysctl_header);
5155 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5158 /* may be called multiple times per register */
5159 static void unregister_sched_domain_sysctl(void)
5161 if (sd_sysctl_header)
5162 unregister_sysctl_table(sd_sysctl_header);
5163 sd_sysctl_header = NULL;
5164 if (sd_ctl_dir[0].child)
5165 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5168 static void register_sched_domain_sysctl(void)
5171 static void unregister_sched_domain_sysctl(void)
5176 static void set_rq_online(struct rq *rq)
5179 const struct sched_class *class;
5181 cpumask_set_cpu(rq->cpu, rq->rd->online);
5184 for_each_class(class) {
5185 if (class->rq_online)
5186 class->rq_online(rq);
5191 static void set_rq_offline(struct rq *rq)
5194 const struct sched_class *class;
5196 for_each_class(class) {
5197 if (class->rq_offline)
5198 class->rq_offline(rq);
5201 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5207 * migration_call - callback that gets triggered when a CPU is added.
5208 * Here we can start up the necessary migration thread for the new CPU.
5210 static int __cpuinit
5211 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5213 int cpu = (long)hcpu;
5214 unsigned long flags;
5215 struct rq *rq = cpu_rq(cpu);
5217 switch (action & ~CPU_TASKS_FROZEN) {
5219 case CPU_UP_PREPARE:
5220 rq->calc_load_update = calc_load_update;
5224 /* Update our root-domain */
5225 raw_spin_lock_irqsave(&rq->lock, flags);
5227 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5231 raw_spin_unlock_irqrestore(&rq->lock, flags);
5234 #ifdef CONFIG_HOTPLUG_CPU
5236 sched_ttwu_pending();
5237 /* Update our root-domain */
5238 raw_spin_lock_irqsave(&rq->lock, flags);
5240 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5244 BUG_ON(rq->nr_running != 1); /* the migration thread */
5245 raw_spin_unlock_irqrestore(&rq->lock, flags);
5249 calc_load_migrate(rq);
5254 update_max_interval();
5260 * Register at high priority so that task migration (migrate_all_tasks)
5261 * happens before everything else. This has to be lower priority than
5262 * the notifier in the perf_event subsystem, though.
5264 static struct notifier_block __cpuinitdata migration_notifier = {
5265 .notifier_call = migration_call,
5266 .priority = CPU_PRI_MIGRATION,
5269 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5270 unsigned long action, void *hcpu)
5272 switch (action & ~CPU_TASKS_FROZEN) {
5273 case CPU_DOWN_FAILED:
5274 set_cpu_active((long)hcpu, true);
5281 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5282 unsigned long action, void *hcpu)
5284 switch (action & ~CPU_TASKS_FROZEN) {
5285 case CPU_DOWN_PREPARE:
5286 set_cpu_active((long)hcpu, false);
5293 static int __init migration_init(void)
5295 void *cpu = (void *)(long)smp_processor_id();
5298 /* Initialize migration for the boot CPU */
5299 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5300 BUG_ON(err == NOTIFY_BAD);
5301 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5302 register_cpu_notifier(&migration_notifier);
5304 /* Register cpu active notifiers */
5305 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5306 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5310 early_initcall(migration_init);
5315 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5317 #ifdef CONFIG_SCHED_DEBUG
5319 static __read_mostly int sched_debug_enabled;
5321 static int __init sched_debug_setup(char *str)
5323 sched_debug_enabled = 1;
5327 early_param("sched_debug", sched_debug_setup);
5329 static inline bool sched_debug(void)
5331 return sched_debug_enabled;
5334 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5335 struct cpumask *groupmask)
5337 struct sched_group *group = sd->groups;
5340 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5341 cpumask_clear(groupmask);
5343 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5345 if (!(sd->flags & SD_LOAD_BALANCE)) {
5346 printk("does not load-balance\n");
5348 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5353 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5355 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5356 printk(KERN_ERR "ERROR: domain->span does not contain "
5359 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5360 printk(KERN_ERR "ERROR: domain->groups does not contain"
5364 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5368 printk(KERN_ERR "ERROR: group is NULL\n");
5373 * Even though we initialize ->power to something semi-sane,
5374 * we leave power_orig unset. This allows us to detect if
5375 * domain iteration is still funny without causing /0 traps.
5377 if (!group->sgp->power_orig) {
5378 printk(KERN_CONT "\n");
5379 printk(KERN_ERR "ERROR: domain->cpu_power not "
5384 if (!cpumask_weight(sched_group_cpus(group))) {
5385 printk(KERN_CONT "\n");
5386 printk(KERN_ERR "ERROR: empty group\n");
5390 if (!(sd->flags & SD_OVERLAP) &&
5391 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5392 printk(KERN_CONT "\n");
5393 printk(KERN_ERR "ERROR: repeated CPUs\n");
5397 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5399 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5401 printk(KERN_CONT " %s", str);
5402 if (group->sgp->power != SCHED_POWER_SCALE) {
5403 printk(KERN_CONT " (cpu_power = %d)",
5407 group = group->next;
5408 } while (group != sd->groups);
5409 printk(KERN_CONT "\n");
5411 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5412 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5415 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5416 printk(KERN_ERR "ERROR: parent span is not a superset "
5417 "of domain->span\n");
5421 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5425 if (!sched_debug_enabled)
5429 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5433 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5436 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5444 #else /* !CONFIG_SCHED_DEBUG */
5445 # define sched_domain_debug(sd, cpu) do { } while (0)
5446 static inline bool sched_debug(void)
5450 #endif /* CONFIG_SCHED_DEBUG */
5452 static int sd_degenerate(struct sched_domain *sd)
5454 if (cpumask_weight(sched_domain_span(sd)) == 1)
5457 /* Following flags need at least 2 groups */
5458 if (sd->flags & (SD_LOAD_BALANCE |
5459 SD_BALANCE_NEWIDLE |
5463 SD_SHARE_PKG_RESOURCES)) {
5464 if (sd->groups != sd->groups->next)
5468 /* Following flags don't use groups */
5469 if (sd->flags & (SD_WAKE_AFFINE))
5476 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5478 unsigned long cflags = sd->flags, pflags = parent->flags;
5480 if (sd_degenerate(parent))
5483 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5486 /* Flags needing groups don't count if only 1 group in parent */
5487 if (parent->groups == parent->groups->next) {
5488 pflags &= ~(SD_LOAD_BALANCE |
5489 SD_BALANCE_NEWIDLE |
5493 SD_SHARE_PKG_RESOURCES);
5494 if (nr_node_ids == 1)
5495 pflags &= ~SD_SERIALIZE;
5497 if (~cflags & pflags)
5503 static void free_rootdomain(struct rcu_head *rcu)
5505 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5507 cpupri_cleanup(&rd->cpupri);
5508 free_cpumask_var(rd->rto_mask);
5509 free_cpumask_var(rd->online);
5510 free_cpumask_var(rd->span);
5514 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5516 struct root_domain *old_rd = NULL;
5517 unsigned long flags;
5519 raw_spin_lock_irqsave(&rq->lock, flags);
5524 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5527 cpumask_clear_cpu(rq->cpu, old_rd->span);
5530 * If we dont want to free the old_rt yet then
5531 * set old_rd to NULL to skip the freeing later
5534 if (!atomic_dec_and_test(&old_rd->refcount))
5538 atomic_inc(&rd->refcount);
5541 cpumask_set_cpu(rq->cpu, rd->span);
5542 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5545 raw_spin_unlock_irqrestore(&rq->lock, flags);
5548 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5551 static int init_rootdomain(struct root_domain *rd)
5553 memset(rd, 0, sizeof(*rd));
5555 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5557 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5559 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5562 if (cpupri_init(&rd->cpupri) != 0)
5567 free_cpumask_var(rd->rto_mask);
5569 free_cpumask_var(rd->online);
5571 free_cpumask_var(rd->span);
5577 * By default the system creates a single root-domain with all cpus as
5578 * members (mimicking the global state we have today).
5580 struct root_domain def_root_domain;
5582 static void init_defrootdomain(void)
5584 init_rootdomain(&def_root_domain);
5586 atomic_set(&def_root_domain.refcount, 1);
5589 static struct root_domain *alloc_rootdomain(void)
5591 struct root_domain *rd;
5593 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5597 if (init_rootdomain(rd) != 0) {
5605 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5607 struct sched_group *tmp, *first;
5616 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5621 } while (sg != first);
5624 static void free_sched_domain(struct rcu_head *rcu)
5626 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5629 * If its an overlapping domain it has private groups, iterate and
5632 if (sd->flags & SD_OVERLAP) {
5633 free_sched_groups(sd->groups, 1);
5634 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5635 kfree(sd->groups->sgp);
5641 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5643 call_rcu(&sd->rcu, free_sched_domain);
5646 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5648 for (; sd; sd = sd->parent)
5649 destroy_sched_domain(sd, cpu);
5653 * Keep a special pointer to the highest sched_domain that has
5654 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5655 * allows us to avoid some pointer chasing select_idle_sibling().
5657 * Also keep a unique ID per domain (we use the first cpu number in
5658 * the cpumask of the domain), this allows us to quickly tell if
5659 * two cpus are in the same cache domain, see cpus_share_cache().
5661 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5662 DEFINE_PER_CPU(int, sd_llc_id);
5664 static void update_top_cache_domain(int cpu)
5666 struct sched_domain *sd;
5669 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5671 id = cpumask_first(sched_domain_span(sd));
5673 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5674 per_cpu(sd_llc_id, cpu) = id;
5678 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5679 * hold the hotplug lock.
5682 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5684 struct rq *rq = cpu_rq(cpu);
5685 struct sched_domain *tmp;
5687 /* Remove the sched domains which do not contribute to scheduling. */
5688 for (tmp = sd; tmp; ) {
5689 struct sched_domain *parent = tmp->parent;
5693 if (sd_parent_degenerate(tmp, parent)) {
5694 tmp->parent = parent->parent;
5696 parent->parent->child = tmp;
5697 destroy_sched_domain(parent, cpu);
5702 if (sd && sd_degenerate(sd)) {
5705 destroy_sched_domain(tmp, cpu);
5710 sched_domain_debug(sd, cpu);
5712 rq_attach_root(rq, rd);
5714 rcu_assign_pointer(rq->sd, sd);
5715 destroy_sched_domains(tmp, cpu);
5717 update_top_cache_domain(cpu);
5720 /* cpus with isolated domains */
5721 static cpumask_var_t cpu_isolated_map;
5723 /* Setup the mask of cpus configured for isolated domains */
5724 static int __init isolated_cpu_setup(char *str)
5726 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5727 cpulist_parse(str, cpu_isolated_map);
5731 __setup("isolcpus=", isolated_cpu_setup);
5733 static const struct cpumask *cpu_cpu_mask(int cpu)
5735 return cpumask_of_node(cpu_to_node(cpu));
5739 struct sched_domain **__percpu sd;
5740 struct sched_group **__percpu sg;
5741 struct sched_group_power **__percpu sgp;
5745 struct sched_domain ** __percpu sd;
5746 struct root_domain *rd;
5756 struct sched_domain_topology_level;
5758 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5759 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5761 #define SDTL_OVERLAP 0x01
5763 struct sched_domain_topology_level {
5764 sched_domain_init_f init;
5765 sched_domain_mask_f mask;
5768 struct sd_data data;
5772 * Build an iteration mask that can exclude certain CPUs from the upwards
5775 * Asymmetric node setups can result in situations where the domain tree is of
5776 * unequal depth, make sure to skip domains that already cover the entire
5779 * In that case build_sched_domains() will have terminated the iteration early
5780 * and our sibling sd spans will be empty. Domains should always include the
5781 * cpu they're built on, so check that.
5784 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5786 const struct cpumask *span = sched_domain_span(sd);
5787 struct sd_data *sdd = sd->private;
5788 struct sched_domain *sibling;
5791 for_each_cpu(i, span) {
5792 sibling = *per_cpu_ptr(sdd->sd, i);
5793 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5796 cpumask_set_cpu(i, sched_group_mask(sg));
5801 * Return the canonical balance cpu for this group, this is the first cpu
5802 * of this group that's also in the iteration mask.
5804 int group_balance_cpu(struct sched_group *sg)
5806 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5810 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5812 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5813 const struct cpumask *span = sched_domain_span(sd);
5814 struct cpumask *covered = sched_domains_tmpmask;
5815 struct sd_data *sdd = sd->private;
5816 struct sched_domain *child;
5819 cpumask_clear(covered);
5821 for_each_cpu(i, span) {
5822 struct cpumask *sg_span;
5824 if (cpumask_test_cpu(i, covered))
5827 child = *per_cpu_ptr(sdd->sd, i);
5829 /* See the comment near build_group_mask(). */
5830 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5833 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5834 GFP_KERNEL, cpu_to_node(cpu));
5839 sg_span = sched_group_cpus(sg);
5841 child = child->child;
5842 cpumask_copy(sg_span, sched_domain_span(child));
5844 cpumask_set_cpu(i, sg_span);
5846 cpumask_or(covered, covered, sg_span);
5848 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5849 if (atomic_inc_return(&sg->sgp->ref) == 1)
5850 build_group_mask(sd, sg);
5853 * Initialize sgp->power such that even if we mess up the
5854 * domains and no possible iteration will get us here, we won't
5857 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5860 * Make sure the first group of this domain contains the
5861 * canonical balance cpu. Otherwise the sched_domain iteration
5862 * breaks. See update_sg_lb_stats().
5864 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5865 group_balance_cpu(sg) == cpu)
5875 sd->groups = groups;
5880 free_sched_groups(first, 0);
5885 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5887 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5888 struct sched_domain *child = sd->child;
5891 cpu = cpumask_first(sched_domain_span(child));
5894 *sg = *per_cpu_ptr(sdd->sg, cpu);
5895 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5896 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5903 * build_sched_groups will build a circular linked list of the groups
5904 * covered by the given span, and will set each group's ->cpumask correctly,
5905 * and ->cpu_power to 0.
5907 * Assumes the sched_domain tree is fully constructed
5910 build_sched_groups(struct sched_domain *sd, int cpu)
5912 struct sched_group *first = NULL, *last = NULL;
5913 struct sd_data *sdd = sd->private;
5914 const struct cpumask *span = sched_domain_span(sd);
5915 struct cpumask *covered;
5918 get_group(cpu, sdd, &sd->groups);
5919 atomic_inc(&sd->groups->ref);
5921 if (cpu != cpumask_first(sched_domain_span(sd)))
5924 lockdep_assert_held(&sched_domains_mutex);
5925 covered = sched_domains_tmpmask;
5927 cpumask_clear(covered);
5929 for_each_cpu(i, span) {
5930 struct sched_group *sg;
5931 int group = get_group(i, sdd, &sg);
5934 if (cpumask_test_cpu(i, covered))
5937 cpumask_clear(sched_group_cpus(sg));
5939 cpumask_setall(sched_group_mask(sg));
5941 for_each_cpu(j, span) {
5942 if (get_group(j, sdd, NULL) != group)
5945 cpumask_set_cpu(j, covered);
5946 cpumask_set_cpu(j, sched_group_cpus(sg));
5961 * Initialize sched groups cpu_power.
5963 * cpu_power indicates the capacity of sched group, which is used while
5964 * distributing the load between different sched groups in a sched domain.
5965 * Typically cpu_power for all the groups in a sched domain will be same unless
5966 * there are asymmetries in the topology. If there are asymmetries, group
5967 * having more cpu_power will pickup more load compared to the group having
5970 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5972 struct sched_group *sg = sd->groups;
5974 WARN_ON(!sd || !sg);
5977 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5979 } while (sg != sd->groups);
5981 if (cpu != group_balance_cpu(sg))
5984 update_group_power(sd, cpu);
5985 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5988 int __weak arch_sd_sibling_asym_packing(void)
5990 return 0*SD_ASYM_PACKING;
5994 * Initializers for schedule domains
5995 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5998 #ifdef CONFIG_SCHED_DEBUG
5999 # define SD_INIT_NAME(sd, type) sd->name = #type
6001 # define SD_INIT_NAME(sd, type) do { } while (0)
6004 #define SD_INIT_FUNC(type) \
6005 static noinline struct sched_domain * \
6006 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6008 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6009 *sd = SD_##type##_INIT; \
6010 SD_INIT_NAME(sd, type); \
6011 sd->private = &tl->data; \
6016 #ifdef CONFIG_SCHED_SMT
6017 SD_INIT_FUNC(SIBLING)
6019 #ifdef CONFIG_SCHED_MC
6022 #ifdef CONFIG_SCHED_BOOK
6026 static int default_relax_domain_level = -1;
6027 int sched_domain_level_max;
6029 static int __init setup_relax_domain_level(char *str)
6031 if (kstrtoint(str, 0, &default_relax_domain_level))
6032 pr_warn("Unable to set relax_domain_level\n");
6036 __setup("relax_domain_level=", setup_relax_domain_level);
6038 static void set_domain_attribute(struct sched_domain *sd,
6039 struct sched_domain_attr *attr)
6043 if (!attr || attr->relax_domain_level < 0) {
6044 if (default_relax_domain_level < 0)
6047 request = default_relax_domain_level;
6049 request = attr->relax_domain_level;
6050 if (request < sd->level) {
6051 /* turn off idle balance on this domain */
6052 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6054 /* turn on idle balance on this domain */
6055 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6059 static void __sdt_free(const struct cpumask *cpu_map);
6060 static int __sdt_alloc(const struct cpumask *cpu_map);
6062 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6063 const struct cpumask *cpu_map)
6067 if (!atomic_read(&d->rd->refcount))
6068 free_rootdomain(&d->rd->rcu); /* fall through */
6070 free_percpu(d->sd); /* fall through */
6072 __sdt_free(cpu_map); /* fall through */
6078 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6079 const struct cpumask *cpu_map)
6081 memset(d, 0, sizeof(*d));
6083 if (__sdt_alloc(cpu_map))
6084 return sa_sd_storage;
6085 d->sd = alloc_percpu(struct sched_domain *);
6087 return sa_sd_storage;
6088 d->rd = alloc_rootdomain();
6091 return sa_rootdomain;
6095 * NULL the sd_data elements we've used to build the sched_domain and
6096 * sched_group structure so that the subsequent __free_domain_allocs()
6097 * will not free the data we're using.
6099 static void claim_allocations(int cpu, struct sched_domain *sd)
6101 struct sd_data *sdd = sd->private;
6103 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6104 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6106 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6107 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6109 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6110 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6113 #ifdef CONFIG_SCHED_SMT
6114 static const struct cpumask *cpu_smt_mask(int cpu)
6116 return topology_thread_cpumask(cpu);
6121 * Topology list, bottom-up.
6123 static struct sched_domain_topology_level default_topology[] = {
6124 #ifdef CONFIG_SCHED_SMT
6125 { sd_init_SIBLING, cpu_smt_mask, },
6127 #ifdef CONFIG_SCHED_MC
6128 { sd_init_MC, cpu_coregroup_mask, },
6130 #ifdef CONFIG_SCHED_BOOK
6131 { sd_init_BOOK, cpu_book_mask, },
6133 { sd_init_CPU, cpu_cpu_mask, },
6137 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6141 static int sched_domains_numa_levels;
6142 static int *sched_domains_numa_distance;
6143 static struct cpumask ***sched_domains_numa_masks;
6144 static int sched_domains_curr_level;
6146 static inline int sd_local_flags(int level)
6148 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6151 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6154 static struct sched_domain *
6155 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6157 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6158 int level = tl->numa_level;
6159 int sd_weight = cpumask_weight(
6160 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6162 *sd = (struct sched_domain){
6163 .min_interval = sd_weight,
6164 .max_interval = 2*sd_weight,
6166 .imbalance_pct = 125,
6167 .cache_nice_tries = 2,
6174 .flags = 1*SD_LOAD_BALANCE
6175 | 1*SD_BALANCE_NEWIDLE
6180 | 0*SD_SHARE_CPUPOWER
6181 | 0*SD_SHARE_PKG_RESOURCES
6183 | 0*SD_PREFER_SIBLING
6184 | sd_local_flags(level)
6186 .last_balance = jiffies,
6187 .balance_interval = sd_weight,
6189 SD_INIT_NAME(sd, NUMA);
6190 sd->private = &tl->data;
6193 * Ugly hack to pass state to sd_numa_mask()...
6195 sched_domains_curr_level = tl->numa_level;
6200 static const struct cpumask *sd_numa_mask(int cpu)
6202 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6205 static void sched_numa_warn(const char *str)
6207 static int done = false;
6215 printk(KERN_WARNING "ERROR: %s\n\n", str);
6217 for (i = 0; i < nr_node_ids; i++) {
6218 printk(KERN_WARNING " ");
6219 for (j = 0; j < nr_node_ids; j++)
6220 printk(KERN_CONT "%02d ", node_distance(i,j));
6221 printk(KERN_CONT "\n");
6223 printk(KERN_WARNING "\n");
6226 static bool find_numa_distance(int distance)
6230 if (distance == node_distance(0, 0))
6233 for (i = 0; i < sched_domains_numa_levels; i++) {
6234 if (sched_domains_numa_distance[i] == distance)
6241 static void sched_init_numa(void)
6243 int next_distance, curr_distance = node_distance(0, 0);
6244 struct sched_domain_topology_level *tl;
6248 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6249 if (!sched_domains_numa_distance)
6253 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6254 * unique distances in the node_distance() table.
6256 * Assumes node_distance(0,j) includes all distances in
6257 * node_distance(i,j) in order to avoid cubic time.
6259 next_distance = curr_distance;
6260 for (i = 0; i < nr_node_ids; i++) {
6261 for (j = 0; j < nr_node_ids; j++) {
6262 for (k = 0; k < nr_node_ids; k++) {
6263 int distance = node_distance(i, k);
6265 if (distance > curr_distance &&
6266 (distance < next_distance ||
6267 next_distance == curr_distance))
6268 next_distance = distance;
6271 * While not a strong assumption it would be nice to know
6272 * about cases where if node A is connected to B, B is not
6273 * equally connected to A.
6275 if (sched_debug() && node_distance(k, i) != distance)
6276 sched_numa_warn("Node-distance not symmetric");
6278 if (sched_debug() && i && !find_numa_distance(distance))
6279 sched_numa_warn("Node-0 not representative");
6281 if (next_distance != curr_distance) {
6282 sched_domains_numa_distance[level++] = next_distance;
6283 sched_domains_numa_levels = level;
6284 curr_distance = next_distance;
6289 * In case of sched_debug() we verify the above assumption.
6295 * 'level' contains the number of unique distances, excluding the
6296 * identity distance node_distance(i,i).
6298 * The sched_domains_numa_distance[] array includes the actual distance
6303 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6304 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6305 * the array will contain less then 'level' members. This could be
6306 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6307 * in other functions.
6309 * We reset it to 'level' at the end of this function.
6311 sched_domains_numa_levels = 0;
6313 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6314 if (!sched_domains_numa_masks)
6318 * Now for each level, construct a mask per node which contains all
6319 * cpus of nodes that are that many hops away from us.
6321 for (i = 0; i < level; i++) {
6322 sched_domains_numa_masks[i] =
6323 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6324 if (!sched_domains_numa_masks[i])
6327 for (j = 0; j < nr_node_ids; j++) {
6328 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6332 sched_domains_numa_masks[i][j] = mask;
6334 for (k = 0; k < nr_node_ids; k++) {
6335 if (node_distance(j, k) > sched_domains_numa_distance[i])
6338 cpumask_or(mask, mask, cpumask_of_node(k));
6343 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6344 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6349 * Copy the default topology bits..
6351 for (i = 0; default_topology[i].init; i++)
6352 tl[i] = default_topology[i];
6355 * .. and append 'j' levels of NUMA goodness.
6357 for (j = 0; j < level; i++, j++) {
6358 tl[i] = (struct sched_domain_topology_level){
6359 .init = sd_numa_init,
6360 .mask = sd_numa_mask,
6361 .flags = SDTL_OVERLAP,
6366 sched_domain_topology = tl;
6368 sched_domains_numa_levels = level;
6371 static void sched_domains_numa_masks_set(int cpu)
6374 int node = cpu_to_node(cpu);
6376 for (i = 0; i < sched_domains_numa_levels; i++) {
6377 for (j = 0; j < nr_node_ids; j++) {
6378 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6379 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6384 static void sched_domains_numa_masks_clear(int cpu)
6387 for (i = 0; i < sched_domains_numa_levels; i++) {
6388 for (j = 0; j < nr_node_ids; j++)
6389 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6394 * Update sched_domains_numa_masks[level][node] array when new cpus
6397 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6398 unsigned long action,
6401 int cpu = (long)hcpu;
6403 switch (action & ~CPU_TASKS_FROZEN) {
6405 sched_domains_numa_masks_set(cpu);
6409 sched_domains_numa_masks_clear(cpu);
6419 static inline void sched_init_numa(void)
6423 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6424 unsigned long action,
6429 #endif /* CONFIG_NUMA */
6431 static int __sdt_alloc(const struct cpumask *cpu_map)
6433 struct sched_domain_topology_level *tl;
6436 for (tl = sched_domain_topology; tl->init; tl++) {
6437 struct sd_data *sdd = &tl->data;
6439 sdd->sd = alloc_percpu(struct sched_domain *);
6443 sdd->sg = alloc_percpu(struct sched_group *);
6447 sdd->sgp = alloc_percpu(struct sched_group_power *);
6451 for_each_cpu(j, cpu_map) {
6452 struct sched_domain *sd;
6453 struct sched_group *sg;
6454 struct sched_group_power *sgp;
6456 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6457 GFP_KERNEL, cpu_to_node(j));
6461 *per_cpu_ptr(sdd->sd, j) = sd;
6463 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6464 GFP_KERNEL, cpu_to_node(j));
6470 *per_cpu_ptr(sdd->sg, j) = sg;
6472 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6473 GFP_KERNEL, cpu_to_node(j));
6477 *per_cpu_ptr(sdd->sgp, j) = sgp;
6484 static void __sdt_free(const struct cpumask *cpu_map)
6486 struct sched_domain_topology_level *tl;
6489 for (tl = sched_domain_topology; tl->init; tl++) {
6490 struct sd_data *sdd = &tl->data;
6492 for_each_cpu(j, cpu_map) {
6493 struct sched_domain *sd;
6496 sd = *per_cpu_ptr(sdd->sd, j);
6497 if (sd && (sd->flags & SD_OVERLAP))
6498 free_sched_groups(sd->groups, 0);
6499 kfree(*per_cpu_ptr(sdd->sd, j));
6503 kfree(*per_cpu_ptr(sdd->sg, j));
6505 kfree(*per_cpu_ptr(sdd->sgp, j));
6507 free_percpu(sdd->sd);
6509 free_percpu(sdd->sg);
6511 free_percpu(sdd->sgp);
6516 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6517 struct s_data *d, const struct cpumask *cpu_map,
6518 struct sched_domain_attr *attr, struct sched_domain *child,
6521 struct sched_domain *sd = tl->init(tl, cpu);
6525 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6527 sd->level = child->level + 1;
6528 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6532 set_domain_attribute(sd, attr);
6538 * Build sched domains for a given set of cpus and attach the sched domains
6539 * to the individual cpus
6541 static int build_sched_domains(const struct cpumask *cpu_map,
6542 struct sched_domain_attr *attr)
6544 enum s_alloc alloc_state = sa_none;
6545 struct sched_domain *sd;
6547 int i, ret = -ENOMEM;
6549 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6550 if (alloc_state != sa_rootdomain)
6553 /* Set up domains for cpus specified by the cpu_map. */
6554 for_each_cpu(i, cpu_map) {
6555 struct sched_domain_topology_level *tl;
6558 for (tl = sched_domain_topology; tl->init; tl++) {
6559 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6560 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6561 sd->flags |= SD_OVERLAP;
6562 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6569 *per_cpu_ptr(d.sd, i) = sd;
6572 /* Build the groups for the domains */
6573 for_each_cpu(i, cpu_map) {
6574 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6575 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6576 if (sd->flags & SD_OVERLAP) {
6577 if (build_overlap_sched_groups(sd, i))
6580 if (build_sched_groups(sd, i))
6586 /* Calculate CPU power for physical packages and nodes */
6587 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6588 if (!cpumask_test_cpu(i, cpu_map))
6591 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6592 claim_allocations(i, sd);
6593 init_sched_groups_power(i, sd);
6597 /* Attach the domains */
6599 for_each_cpu(i, cpu_map) {
6600 sd = *per_cpu_ptr(d.sd, i);
6601 cpu_attach_domain(sd, d.rd, i);
6607 __free_domain_allocs(&d, alloc_state, cpu_map);
6611 static cpumask_var_t *doms_cur; /* current sched domains */
6612 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6613 static struct sched_domain_attr *dattr_cur;
6614 /* attribues of custom domains in 'doms_cur' */
6617 * Special case: If a kmalloc of a doms_cur partition (array of
6618 * cpumask) fails, then fallback to a single sched domain,
6619 * as determined by the single cpumask fallback_doms.
6621 static cpumask_var_t fallback_doms;
6624 * arch_update_cpu_topology lets virtualized architectures update the
6625 * cpu core maps. It is supposed to return 1 if the topology changed
6626 * or 0 if it stayed the same.
6628 int __attribute__((weak)) arch_update_cpu_topology(void)
6633 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6636 cpumask_var_t *doms;
6638 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6641 for (i = 0; i < ndoms; i++) {
6642 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6643 free_sched_domains(doms, i);
6650 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6653 for (i = 0; i < ndoms; i++)
6654 free_cpumask_var(doms[i]);
6659 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6660 * For now this just excludes isolated cpus, but could be used to
6661 * exclude other special cases in the future.
6663 static int init_sched_domains(const struct cpumask *cpu_map)
6667 arch_update_cpu_topology();
6669 doms_cur = alloc_sched_domains(ndoms_cur);
6671 doms_cur = &fallback_doms;
6672 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6673 err = build_sched_domains(doms_cur[0], NULL);
6674 register_sched_domain_sysctl();
6680 * Detach sched domains from a group of cpus specified in cpu_map
6681 * These cpus will now be attached to the NULL domain
6683 static void detach_destroy_domains(const struct cpumask *cpu_map)
6688 for_each_cpu(i, cpu_map)
6689 cpu_attach_domain(NULL, &def_root_domain, i);
6693 /* handle null as "default" */
6694 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6695 struct sched_domain_attr *new, int idx_new)
6697 struct sched_domain_attr tmp;
6704 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6705 new ? (new + idx_new) : &tmp,
6706 sizeof(struct sched_domain_attr));
6710 * Partition sched domains as specified by the 'ndoms_new'
6711 * cpumasks in the array doms_new[] of cpumasks. This compares
6712 * doms_new[] to the current sched domain partitioning, doms_cur[].
6713 * It destroys each deleted domain and builds each new domain.
6715 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6716 * The masks don't intersect (don't overlap.) We should setup one
6717 * sched domain for each mask. CPUs not in any of the cpumasks will
6718 * not be load balanced. If the same cpumask appears both in the
6719 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6722 * The passed in 'doms_new' should be allocated using
6723 * alloc_sched_domains. This routine takes ownership of it and will
6724 * free_sched_domains it when done with it. If the caller failed the
6725 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6726 * and partition_sched_domains() will fallback to the single partition
6727 * 'fallback_doms', it also forces the domains to be rebuilt.
6729 * If doms_new == NULL it will be replaced with cpu_online_mask.
6730 * ndoms_new == 0 is a special case for destroying existing domains,
6731 * and it will not create the default domain.
6733 * Call with hotplug lock held
6735 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6736 struct sched_domain_attr *dattr_new)
6741 mutex_lock(&sched_domains_mutex);
6743 /* always unregister in case we don't destroy any domains */
6744 unregister_sched_domain_sysctl();
6746 /* Let architecture update cpu core mappings. */
6747 new_topology = arch_update_cpu_topology();
6749 n = doms_new ? ndoms_new : 0;
6751 /* Destroy deleted domains */
6752 for (i = 0; i < ndoms_cur; i++) {
6753 for (j = 0; j < n && !new_topology; j++) {
6754 if (cpumask_equal(doms_cur[i], doms_new[j])
6755 && dattrs_equal(dattr_cur, i, dattr_new, j))
6758 /* no match - a current sched domain not in new doms_new[] */
6759 detach_destroy_domains(doms_cur[i]);
6764 if (doms_new == NULL) {
6766 doms_new = &fallback_doms;
6767 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6768 WARN_ON_ONCE(dattr_new);
6771 /* Build new domains */
6772 for (i = 0; i < ndoms_new; i++) {
6773 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6774 if (cpumask_equal(doms_new[i], doms_cur[j])
6775 && dattrs_equal(dattr_new, i, dattr_cur, j))
6778 /* no match - add a new doms_new */
6779 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6784 /* Remember the new sched domains */
6785 if (doms_cur != &fallback_doms)
6786 free_sched_domains(doms_cur, ndoms_cur);
6787 kfree(dattr_cur); /* kfree(NULL) is safe */
6788 doms_cur = doms_new;
6789 dattr_cur = dattr_new;
6790 ndoms_cur = ndoms_new;
6792 register_sched_domain_sysctl();
6794 mutex_unlock(&sched_domains_mutex);
6797 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6800 * Update cpusets according to cpu_active mask. If cpusets are
6801 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6802 * around partition_sched_domains().
6804 * If we come here as part of a suspend/resume, don't touch cpusets because we
6805 * want to restore it back to its original state upon resume anyway.
6807 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6811 case CPU_ONLINE_FROZEN:
6812 case CPU_DOWN_FAILED_FROZEN:
6815 * num_cpus_frozen tracks how many CPUs are involved in suspend
6816 * resume sequence. As long as this is not the last online
6817 * operation in the resume sequence, just build a single sched
6818 * domain, ignoring cpusets.
6821 if (likely(num_cpus_frozen)) {
6822 partition_sched_domains(1, NULL, NULL);
6827 * This is the last CPU online operation. So fall through and
6828 * restore the original sched domains by considering the
6829 * cpuset configurations.
6833 case CPU_DOWN_FAILED:
6834 cpuset_update_active_cpus(true);
6842 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6846 case CPU_DOWN_PREPARE:
6847 cpuset_update_active_cpus(false);
6849 case CPU_DOWN_PREPARE_FROZEN:
6851 partition_sched_domains(1, NULL, NULL);
6859 void __init sched_init_smp(void)
6861 cpumask_var_t non_isolated_cpus;
6863 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6864 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6869 mutex_lock(&sched_domains_mutex);
6870 init_sched_domains(cpu_active_mask);
6871 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6872 if (cpumask_empty(non_isolated_cpus))
6873 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6874 mutex_unlock(&sched_domains_mutex);
6877 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6878 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6879 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6881 /* RT runtime code needs to handle some hotplug events */
6882 hotcpu_notifier(update_runtime, 0);
6886 /* Move init over to a non-isolated CPU */
6887 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6889 sched_init_granularity();
6890 free_cpumask_var(non_isolated_cpus);
6892 init_sched_rt_class();
6895 void __init sched_init_smp(void)
6897 sched_init_granularity();
6899 #endif /* CONFIG_SMP */
6901 const_debug unsigned int sysctl_timer_migration = 1;
6903 int in_sched_functions(unsigned long addr)
6905 return in_lock_functions(addr) ||
6906 (addr >= (unsigned long)__sched_text_start
6907 && addr < (unsigned long)__sched_text_end);
6910 #ifdef CONFIG_CGROUP_SCHED
6912 * Default task group.
6913 * Every task in system belongs to this group at bootup.
6915 struct task_group root_task_group;
6916 LIST_HEAD(task_groups);
6919 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6921 void __init sched_init(void)
6924 unsigned long alloc_size = 0, ptr;
6926 #ifdef CONFIG_FAIR_GROUP_SCHED
6927 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6929 #ifdef CONFIG_RT_GROUP_SCHED
6930 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6932 #ifdef CONFIG_CPUMASK_OFFSTACK
6933 alloc_size += num_possible_cpus() * cpumask_size();
6936 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6938 #ifdef CONFIG_FAIR_GROUP_SCHED
6939 root_task_group.se = (struct sched_entity **)ptr;
6940 ptr += nr_cpu_ids * sizeof(void **);
6942 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6943 ptr += nr_cpu_ids * sizeof(void **);
6945 #endif /* CONFIG_FAIR_GROUP_SCHED */
6946 #ifdef CONFIG_RT_GROUP_SCHED
6947 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6948 ptr += nr_cpu_ids * sizeof(void **);
6950 root_task_group.rt_rq = (struct rt_rq **)ptr;
6951 ptr += nr_cpu_ids * sizeof(void **);
6953 #endif /* CONFIG_RT_GROUP_SCHED */
6954 #ifdef CONFIG_CPUMASK_OFFSTACK
6955 for_each_possible_cpu(i) {
6956 per_cpu(load_balance_mask, i) = (void *)ptr;
6957 ptr += cpumask_size();
6959 #endif /* CONFIG_CPUMASK_OFFSTACK */
6963 init_defrootdomain();
6966 init_rt_bandwidth(&def_rt_bandwidth,
6967 global_rt_period(), global_rt_runtime());
6969 #ifdef CONFIG_RT_GROUP_SCHED
6970 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6971 global_rt_period(), global_rt_runtime());
6972 #endif /* CONFIG_RT_GROUP_SCHED */
6974 #ifdef CONFIG_CGROUP_SCHED
6975 list_add(&root_task_group.list, &task_groups);
6976 INIT_LIST_HEAD(&root_task_group.children);
6977 INIT_LIST_HEAD(&root_task_group.siblings);
6978 autogroup_init(&init_task);
6980 #endif /* CONFIG_CGROUP_SCHED */
6982 for_each_possible_cpu(i) {
6986 raw_spin_lock_init(&rq->lock);
6988 rq->calc_load_active = 0;
6989 rq->calc_load_update = jiffies + LOAD_FREQ;
6990 init_cfs_rq(&rq->cfs);
6991 init_rt_rq(&rq->rt, rq);
6992 #ifdef CONFIG_FAIR_GROUP_SCHED
6993 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6994 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6996 * How much cpu bandwidth does root_task_group get?
6998 * In case of task-groups formed thr' the cgroup filesystem, it
6999 * gets 100% of the cpu resources in the system. This overall
7000 * system cpu resource is divided among the tasks of
7001 * root_task_group and its child task-groups in a fair manner,
7002 * based on each entity's (task or task-group's) weight
7003 * (se->load.weight).
7005 * In other words, if root_task_group has 10 tasks of weight
7006 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7007 * then A0's share of the cpu resource is:
7009 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7011 * We achieve this by letting root_task_group's tasks sit
7012 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7014 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7015 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7016 #endif /* CONFIG_FAIR_GROUP_SCHED */
7018 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7019 #ifdef CONFIG_RT_GROUP_SCHED
7020 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7021 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7024 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7025 rq->cpu_load[j] = 0;
7027 rq->last_load_update_tick = jiffies;
7032 rq->cpu_power = SCHED_POWER_SCALE;
7033 rq->post_schedule = 0;
7034 rq->active_balance = 0;
7035 rq->next_balance = jiffies;
7040 rq->avg_idle = 2*sysctl_sched_migration_cost;
7042 INIT_LIST_HEAD(&rq->cfs_tasks);
7044 rq_attach_root(rq, &def_root_domain);
7045 #ifdef CONFIG_NO_HZ_COMMON
7048 #ifdef CONFIG_NO_HZ_FULL
7049 rq->last_sched_tick = 0;
7053 atomic_set(&rq->nr_iowait, 0);
7056 set_load_weight(&init_task);
7058 #ifdef CONFIG_PREEMPT_NOTIFIERS
7059 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7062 #ifdef CONFIG_RT_MUTEXES
7063 plist_head_init(&init_task.pi_waiters);
7067 * The boot idle thread does lazy MMU switching as well:
7069 atomic_inc(&init_mm.mm_count);
7070 enter_lazy_tlb(&init_mm, current);
7073 * Make us the idle thread. Technically, schedule() should not be
7074 * called from this thread, however somewhere below it might be,
7075 * but because we are the idle thread, we just pick up running again
7076 * when this runqueue becomes "idle".
7078 init_idle(current, smp_processor_id());
7080 calc_load_update = jiffies + LOAD_FREQ;
7083 * During early bootup we pretend to be a normal task:
7085 current->sched_class = &fair_sched_class;
7088 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7089 /* May be allocated at isolcpus cmdline parse time */
7090 if (cpu_isolated_map == NULL)
7091 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7092 idle_thread_set_boot_cpu();
7094 init_sched_fair_class();
7096 scheduler_running = 1;
7099 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7100 static inline int preempt_count_equals(int preempt_offset)
7102 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7104 return (nested == preempt_offset);
7107 void __might_sleep(const char *file, int line, int preempt_offset)
7109 static unsigned long prev_jiffy; /* ratelimiting */
7111 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7112 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7113 system_state != SYSTEM_RUNNING || oops_in_progress)
7115 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7117 prev_jiffy = jiffies;
7120 "BUG: sleeping function called from invalid context at %s:%d\n",
7123 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7124 in_atomic(), irqs_disabled(),
7125 current->pid, current->comm);
7127 debug_show_held_locks(current);
7128 if (irqs_disabled())
7129 print_irqtrace_events(current);
7132 EXPORT_SYMBOL(__might_sleep);
7135 #ifdef CONFIG_MAGIC_SYSRQ
7136 static void normalize_task(struct rq *rq, struct task_struct *p)
7138 const struct sched_class *prev_class = p->sched_class;
7139 int old_prio = p->prio;
7144 dequeue_task(rq, p, 0);
7145 __setscheduler(rq, p, SCHED_NORMAL, 0);
7147 enqueue_task(rq, p, 0);
7148 resched_task(rq->curr);
7151 check_class_changed(rq, p, prev_class, old_prio);
7154 void normalize_rt_tasks(void)
7156 struct task_struct *g, *p;
7157 unsigned long flags;
7160 read_lock_irqsave(&tasklist_lock, flags);
7161 do_each_thread(g, p) {
7163 * Only normalize user tasks:
7168 p->se.exec_start = 0;
7169 #ifdef CONFIG_SCHEDSTATS
7170 p->se.statistics.wait_start = 0;
7171 p->se.statistics.sleep_start = 0;
7172 p->se.statistics.block_start = 0;
7177 * Renice negative nice level userspace
7180 if (TASK_NICE(p) < 0 && p->mm)
7181 set_user_nice(p, 0);
7185 raw_spin_lock(&p->pi_lock);
7186 rq = __task_rq_lock(p);
7188 normalize_task(rq, p);
7190 __task_rq_unlock(rq);
7191 raw_spin_unlock(&p->pi_lock);
7192 } while_each_thread(g, p);
7194 read_unlock_irqrestore(&tasklist_lock, flags);
7197 #endif /* CONFIG_MAGIC_SYSRQ */
7199 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7201 * These functions are only useful for the IA64 MCA handling, or kdb.
7203 * They can only be called when the whole system has been
7204 * stopped - every CPU needs to be quiescent, and no scheduling
7205 * activity can take place. Using them for anything else would
7206 * be a serious bug, and as a result, they aren't even visible
7207 * under any other configuration.
7211 * curr_task - return the current task for a given cpu.
7212 * @cpu: the processor in question.
7214 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7216 struct task_struct *curr_task(int cpu)
7218 return cpu_curr(cpu);
7221 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7225 * set_curr_task - set the current task for a given cpu.
7226 * @cpu: the processor in question.
7227 * @p: the task pointer to set.
7229 * Description: This function must only be used when non-maskable interrupts
7230 * are serviced on a separate stack. It allows the architecture to switch the
7231 * notion of the current task on a cpu in a non-blocking manner. This function
7232 * must be called with all CPU's synchronized, and interrupts disabled, the
7233 * and caller must save the original value of the current task (see
7234 * curr_task() above) and restore that value before reenabling interrupts and
7235 * re-starting the system.
7237 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7239 void set_curr_task(int cpu, struct task_struct *p)
7246 #ifdef CONFIG_CGROUP_SCHED
7247 /* task_group_lock serializes the addition/removal of task groups */
7248 static DEFINE_SPINLOCK(task_group_lock);
7250 static void free_sched_group(struct task_group *tg)
7252 free_fair_sched_group(tg);
7253 free_rt_sched_group(tg);
7258 /* allocate runqueue etc for a new task group */
7259 struct task_group *sched_create_group(struct task_group *parent)
7261 struct task_group *tg;
7263 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7265 return ERR_PTR(-ENOMEM);
7267 if (!alloc_fair_sched_group(tg, parent))
7270 if (!alloc_rt_sched_group(tg, parent))
7276 free_sched_group(tg);
7277 return ERR_PTR(-ENOMEM);
7280 void sched_online_group(struct task_group *tg, struct task_group *parent)
7282 unsigned long flags;
7284 spin_lock_irqsave(&task_group_lock, flags);
7285 list_add_rcu(&tg->list, &task_groups);
7287 WARN_ON(!parent); /* root should already exist */
7289 tg->parent = parent;
7290 INIT_LIST_HEAD(&tg->children);
7291 list_add_rcu(&tg->siblings, &parent->children);
7292 spin_unlock_irqrestore(&task_group_lock, flags);
7295 /* rcu callback to free various structures associated with a task group */
7296 static void free_sched_group_rcu(struct rcu_head *rhp)
7298 /* now it should be safe to free those cfs_rqs */
7299 free_sched_group(container_of(rhp, struct task_group, rcu));
7302 /* Destroy runqueue etc associated with a task group */
7303 void sched_destroy_group(struct task_group *tg)
7305 /* wait for possible concurrent references to cfs_rqs complete */
7306 call_rcu(&tg->rcu, free_sched_group_rcu);
7309 void sched_offline_group(struct task_group *tg)
7311 unsigned long flags;
7314 /* end participation in shares distribution */
7315 for_each_possible_cpu(i)
7316 unregister_fair_sched_group(tg, i);
7318 spin_lock_irqsave(&task_group_lock, flags);
7319 list_del_rcu(&tg->list);
7320 list_del_rcu(&tg->siblings);
7321 spin_unlock_irqrestore(&task_group_lock, flags);
7324 /* change task's runqueue when it moves between groups.
7325 * The caller of this function should have put the task in its new group
7326 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7327 * reflect its new group.
7329 void sched_move_task(struct task_struct *tsk)
7331 struct task_group *tg;
7333 unsigned long flags;
7336 rq = task_rq_lock(tsk, &flags);
7338 running = task_current(rq, tsk);
7342 dequeue_task(rq, tsk, 0);
7343 if (unlikely(running))
7344 tsk->sched_class->put_prev_task(rq, tsk);
7346 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7347 lockdep_is_held(&tsk->sighand->siglock)),
7348 struct task_group, css);
7349 tg = autogroup_task_group(tsk, tg);
7350 tsk->sched_task_group = tg;
7352 #ifdef CONFIG_FAIR_GROUP_SCHED
7353 if (tsk->sched_class->task_move_group)
7354 tsk->sched_class->task_move_group(tsk, on_rq);
7357 set_task_rq(tsk, task_cpu(tsk));
7359 if (unlikely(running))
7360 tsk->sched_class->set_curr_task(rq);
7362 enqueue_task(rq, tsk, 0);
7364 task_rq_unlock(rq, tsk, &flags);
7366 #endif /* CONFIG_CGROUP_SCHED */
7368 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7369 static unsigned long to_ratio(u64 period, u64 runtime)
7371 if (runtime == RUNTIME_INF)
7374 return div64_u64(runtime << 20, period);
7378 #ifdef CONFIG_RT_GROUP_SCHED
7380 * Ensure that the real time constraints are schedulable.
7382 static DEFINE_MUTEX(rt_constraints_mutex);
7384 /* Must be called with tasklist_lock held */
7385 static inline int tg_has_rt_tasks(struct task_group *tg)
7387 struct task_struct *g, *p;
7389 do_each_thread(g, p) {
7390 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7392 } while_each_thread(g, p);
7397 struct rt_schedulable_data {
7398 struct task_group *tg;
7403 static int tg_rt_schedulable(struct task_group *tg, void *data)
7405 struct rt_schedulable_data *d = data;
7406 struct task_group *child;
7407 unsigned long total, sum = 0;
7408 u64 period, runtime;
7410 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7411 runtime = tg->rt_bandwidth.rt_runtime;
7414 period = d->rt_period;
7415 runtime = d->rt_runtime;
7419 * Cannot have more runtime than the period.
7421 if (runtime > period && runtime != RUNTIME_INF)
7425 * Ensure we don't starve existing RT tasks.
7427 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7430 total = to_ratio(period, runtime);
7433 * Nobody can have more than the global setting allows.
7435 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7439 * The sum of our children's runtime should not exceed our own.
7441 list_for_each_entry_rcu(child, &tg->children, siblings) {
7442 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7443 runtime = child->rt_bandwidth.rt_runtime;
7445 if (child == d->tg) {
7446 period = d->rt_period;
7447 runtime = d->rt_runtime;
7450 sum += to_ratio(period, runtime);
7459 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7463 struct rt_schedulable_data data = {
7465 .rt_period = period,
7466 .rt_runtime = runtime,
7470 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7476 static int tg_set_rt_bandwidth(struct task_group *tg,
7477 u64 rt_period, u64 rt_runtime)
7481 mutex_lock(&rt_constraints_mutex);
7482 read_lock(&tasklist_lock);
7483 err = __rt_schedulable(tg, rt_period, rt_runtime);
7487 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7488 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7489 tg->rt_bandwidth.rt_runtime = rt_runtime;
7491 for_each_possible_cpu(i) {
7492 struct rt_rq *rt_rq = tg->rt_rq[i];
7494 raw_spin_lock(&rt_rq->rt_runtime_lock);
7495 rt_rq->rt_runtime = rt_runtime;
7496 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7498 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7500 read_unlock(&tasklist_lock);
7501 mutex_unlock(&rt_constraints_mutex);
7506 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7508 u64 rt_runtime, rt_period;
7510 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7511 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7512 if (rt_runtime_us < 0)
7513 rt_runtime = RUNTIME_INF;
7515 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7518 static long sched_group_rt_runtime(struct task_group *tg)
7522 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7525 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7526 do_div(rt_runtime_us, NSEC_PER_USEC);
7527 return rt_runtime_us;
7530 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7532 u64 rt_runtime, rt_period;
7534 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7535 rt_runtime = tg->rt_bandwidth.rt_runtime;
7540 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7543 static long sched_group_rt_period(struct task_group *tg)
7547 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7548 do_div(rt_period_us, NSEC_PER_USEC);
7549 return rt_period_us;
7552 static int sched_rt_global_constraints(void)
7554 u64 runtime, period;
7557 if (sysctl_sched_rt_period <= 0)
7560 runtime = global_rt_runtime();
7561 period = global_rt_period();
7564 * Sanity check on the sysctl variables.
7566 if (runtime > period && runtime != RUNTIME_INF)
7569 mutex_lock(&rt_constraints_mutex);
7570 read_lock(&tasklist_lock);
7571 ret = __rt_schedulable(NULL, 0, 0);
7572 read_unlock(&tasklist_lock);
7573 mutex_unlock(&rt_constraints_mutex);
7578 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7580 /* Don't accept realtime tasks when there is no way for them to run */
7581 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7587 #else /* !CONFIG_RT_GROUP_SCHED */
7588 static int sched_rt_global_constraints(void)
7590 unsigned long flags;
7593 if (sysctl_sched_rt_period <= 0)
7597 * There's always some RT tasks in the root group
7598 * -- migration, kstopmachine etc..
7600 if (sysctl_sched_rt_runtime == 0)
7603 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7604 for_each_possible_cpu(i) {
7605 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7607 raw_spin_lock(&rt_rq->rt_runtime_lock);
7608 rt_rq->rt_runtime = global_rt_runtime();
7609 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7611 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7615 #endif /* CONFIG_RT_GROUP_SCHED */
7617 int sched_rr_handler(struct ctl_table *table, int write,
7618 void __user *buffer, size_t *lenp,
7622 static DEFINE_MUTEX(mutex);
7625 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7626 /* make sure that internally we keep jiffies */
7627 /* also, writing zero resets timeslice to default */
7628 if (!ret && write) {
7629 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7630 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7632 mutex_unlock(&mutex);
7636 int sched_rt_handler(struct ctl_table *table, int write,
7637 void __user *buffer, size_t *lenp,
7641 int old_period, old_runtime;
7642 static DEFINE_MUTEX(mutex);
7645 old_period = sysctl_sched_rt_period;
7646 old_runtime = sysctl_sched_rt_runtime;
7648 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7650 if (!ret && write) {
7651 ret = sched_rt_global_constraints();
7653 sysctl_sched_rt_period = old_period;
7654 sysctl_sched_rt_runtime = old_runtime;
7656 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7657 def_rt_bandwidth.rt_period =
7658 ns_to_ktime(global_rt_period());
7661 mutex_unlock(&mutex);
7666 #ifdef CONFIG_CGROUP_SCHED
7668 /* return corresponding task_group object of a cgroup */
7669 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7671 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7672 struct task_group, css);
7675 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7677 struct task_group *tg, *parent;
7679 if (!cgrp->parent) {
7680 /* This is early initialization for the top cgroup */
7681 return &root_task_group.css;
7684 parent = cgroup_tg(cgrp->parent);
7685 tg = sched_create_group(parent);
7687 return ERR_PTR(-ENOMEM);
7692 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7694 struct task_group *tg = cgroup_tg(cgrp);
7695 struct task_group *parent;
7700 parent = cgroup_tg(cgrp->parent);
7701 sched_online_group(tg, parent);
7705 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7707 struct task_group *tg = cgroup_tg(cgrp);
7709 sched_destroy_group(tg);
7712 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7714 struct task_group *tg = cgroup_tg(cgrp);
7716 sched_offline_group(tg);
7719 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7720 struct cgroup_taskset *tset)
7722 struct task_struct *task;
7724 cgroup_taskset_for_each(task, cgrp, tset) {
7725 #ifdef CONFIG_RT_GROUP_SCHED
7726 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7729 /* We don't support RT-tasks being in separate groups */
7730 if (task->sched_class != &fair_sched_class)
7737 static void cpu_cgroup_attach(struct cgroup *cgrp,
7738 struct cgroup_taskset *tset)
7740 struct task_struct *task;
7742 cgroup_taskset_for_each(task, cgrp, tset)
7743 sched_move_task(task);
7747 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7748 struct task_struct *task)
7751 * cgroup_exit() is called in the copy_process() failure path.
7752 * Ignore this case since the task hasn't ran yet, this avoids
7753 * trying to poke a half freed task state from generic code.
7755 if (!(task->flags & PF_EXITING))
7758 sched_move_task(task);
7761 #ifdef CONFIG_FAIR_GROUP_SCHED
7762 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7765 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7768 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7770 struct task_group *tg = cgroup_tg(cgrp);
7772 return (u64) scale_load_down(tg->shares);
7775 #ifdef CONFIG_CFS_BANDWIDTH
7776 static DEFINE_MUTEX(cfs_constraints_mutex);
7778 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7779 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7781 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7783 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7785 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7786 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7788 if (tg == &root_task_group)
7792 * Ensure we have at some amount of bandwidth every period. This is
7793 * to prevent reaching a state of large arrears when throttled via
7794 * entity_tick() resulting in prolonged exit starvation.
7796 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7800 * Likewise, bound things on the otherside by preventing insane quota
7801 * periods. This also allows us to normalize in computing quota
7804 if (period > max_cfs_quota_period)
7807 mutex_lock(&cfs_constraints_mutex);
7808 ret = __cfs_schedulable(tg, period, quota);
7812 runtime_enabled = quota != RUNTIME_INF;
7813 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7815 * If we need to toggle cfs_bandwidth_used, off->on must occur
7816 * before making related changes, and on->off must occur afterwards
7818 if (runtime_enabled && !runtime_was_enabled)
7819 cfs_bandwidth_usage_inc();
7820 raw_spin_lock_irq(&cfs_b->lock);
7821 cfs_b->period = ns_to_ktime(period);
7822 cfs_b->quota = quota;
7824 __refill_cfs_bandwidth_runtime(cfs_b);
7825 /* restart the period timer (if active) to handle new period expiry */
7826 if (runtime_enabled && cfs_b->timer_active) {
7827 /* force a reprogram */
7828 cfs_b->timer_active = 0;
7829 __start_cfs_bandwidth(cfs_b);
7831 raw_spin_unlock_irq(&cfs_b->lock);
7833 for_each_possible_cpu(i) {
7834 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7835 struct rq *rq = cfs_rq->rq;
7837 raw_spin_lock_irq(&rq->lock);
7838 cfs_rq->runtime_enabled = runtime_enabled;
7839 cfs_rq->runtime_remaining = 0;
7841 if (cfs_rq->throttled)
7842 unthrottle_cfs_rq(cfs_rq);
7843 raw_spin_unlock_irq(&rq->lock);
7845 if (runtime_was_enabled && !runtime_enabled)
7846 cfs_bandwidth_usage_dec();
7848 mutex_unlock(&cfs_constraints_mutex);
7853 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7857 period = ktime_to_ns(tg->cfs_bandwidth.period);
7858 if (cfs_quota_us < 0)
7859 quota = RUNTIME_INF;
7861 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7863 return tg_set_cfs_bandwidth(tg, period, quota);
7866 long tg_get_cfs_quota(struct task_group *tg)
7870 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7873 quota_us = tg->cfs_bandwidth.quota;
7874 do_div(quota_us, NSEC_PER_USEC);
7879 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7883 period = (u64)cfs_period_us * NSEC_PER_USEC;
7884 quota = tg->cfs_bandwidth.quota;
7886 return tg_set_cfs_bandwidth(tg, period, quota);
7889 long tg_get_cfs_period(struct task_group *tg)
7893 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7894 do_div(cfs_period_us, NSEC_PER_USEC);
7896 return cfs_period_us;
7899 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7901 return tg_get_cfs_quota(cgroup_tg(cgrp));
7904 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7907 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7910 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7912 return tg_get_cfs_period(cgroup_tg(cgrp));
7915 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7918 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7921 struct cfs_schedulable_data {
7922 struct task_group *tg;
7927 * normalize group quota/period to be quota/max_period
7928 * note: units are usecs
7930 static u64 normalize_cfs_quota(struct task_group *tg,
7931 struct cfs_schedulable_data *d)
7939 period = tg_get_cfs_period(tg);
7940 quota = tg_get_cfs_quota(tg);
7943 /* note: these should typically be equivalent */
7944 if (quota == RUNTIME_INF || quota == -1)
7947 return to_ratio(period, quota);
7950 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7952 struct cfs_schedulable_data *d = data;
7953 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7954 s64 quota = 0, parent_quota = -1;
7957 quota = RUNTIME_INF;
7959 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7961 quota = normalize_cfs_quota(tg, d);
7962 parent_quota = parent_b->hierarchal_quota;
7965 * ensure max(child_quota) <= parent_quota, inherit when no
7968 if (quota == RUNTIME_INF)
7969 quota = parent_quota;
7970 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7973 cfs_b->hierarchal_quota = quota;
7978 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7981 struct cfs_schedulable_data data = {
7987 if (quota != RUNTIME_INF) {
7988 do_div(data.period, NSEC_PER_USEC);
7989 do_div(data.quota, NSEC_PER_USEC);
7993 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7999 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8000 struct cgroup_map_cb *cb)
8002 struct task_group *tg = cgroup_tg(cgrp);
8003 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8005 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8006 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8007 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8011 #endif /* CONFIG_CFS_BANDWIDTH */
8012 #endif /* CONFIG_FAIR_GROUP_SCHED */
8014 #ifdef CONFIG_RT_GROUP_SCHED
8015 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8018 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8021 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8023 return sched_group_rt_runtime(cgroup_tg(cgrp));
8026 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8029 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8032 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8034 return sched_group_rt_period(cgroup_tg(cgrp));
8036 #endif /* CONFIG_RT_GROUP_SCHED */
8038 static struct cftype cpu_files[] = {
8039 #ifdef CONFIG_FAIR_GROUP_SCHED
8042 .read_u64 = cpu_shares_read_u64,
8043 .write_u64 = cpu_shares_write_u64,
8046 #ifdef CONFIG_CFS_BANDWIDTH
8048 .name = "cfs_quota_us",
8049 .read_s64 = cpu_cfs_quota_read_s64,
8050 .write_s64 = cpu_cfs_quota_write_s64,
8053 .name = "cfs_period_us",
8054 .read_u64 = cpu_cfs_period_read_u64,
8055 .write_u64 = cpu_cfs_period_write_u64,
8059 .read_map = cpu_stats_show,
8062 #ifdef CONFIG_RT_GROUP_SCHED
8064 .name = "rt_runtime_us",
8065 .read_s64 = cpu_rt_runtime_read,
8066 .write_s64 = cpu_rt_runtime_write,
8069 .name = "rt_period_us",
8070 .read_u64 = cpu_rt_period_read_uint,
8071 .write_u64 = cpu_rt_period_write_uint,
8077 struct cgroup_subsys cpu_cgroup_subsys = {
8079 .css_alloc = cpu_cgroup_css_alloc,
8080 .css_free = cpu_cgroup_css_free,
8081 .css_online = cpu_cgroup_css_online,
8082 .css_offline = cpu_cgroup_css_offline,
8083 .can_attach = cpu_cgroup_can_attach,
8084 .attach = cpu_cgroup_attach,
8085 .exit = cpu_cgroup_exit,
8086 .subsys_id = cpu_cgroup_subsys_id,
8087 .base_cftypes = cpu_files,
8091 #endif /* CONFIG_CGROUP_SCHED */
8093 void dump_cpu_task(int cpu)
8095 pr_info("Task dump for CPU %d:\n", cpu);
8096 sched_show_task(cpu_curr(cpu));