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>
76 #include <asm/switch_to.h>
78 #include <asm/irq_regs.h>
79 #include <asm/mutex.h>
80 #ifdef CONFIG_PARAVIRT
81 #include <asm/paravirt.h>
85 #include "../workqueue_sched.h"
86 #include "../smpboot.h"
88 #define CREATE_TRACE_POINTS
89 #include <trace/events/sched.h>
91 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
94 ktime_t soft, hard, now;
97 if (hrtimer_active(period_timer))
100 now = hrtimer_cb_get_time(period_timer);
101 hrtimer_forward(period_timer, now, period);
103 soft = hrtimer_get_softexpires(period_timer);
104 hard = hrtimer_get_expires(period_timer);
105 delta = ktime_to_ns(ktime_sub(hard, soft));
106 __hrtimer_start_range_ns(period_timer, soft, delta,
107 HRTIMER_MODE_ABS_PINNED, 0);
111 DEFINE_MUTEX(sched_domains_mutex);
112 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
114 static void update_rq_clock_task(struct rq *rq, s64 delta);
116 void update_rq_clock(struct rq *rq)
120 if (rq->skip_clock_update > 0)
123 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
125 update_rq_clock_task(rq, delta);
129 * Debugging: various feature bits
132 #define SCHED_FEAT(name, enabled) \
133 (1UL << __SCHED_FEAT_##name) * enabled |
135 const_debug unsigned int sysctl_sched_features =
136 #include "features.h"
141 #ifdef CONFIG_SCHED_DEBUG
142 #define SCHED_FEAT(name, enabled) \
145 static const char * const sched_feat_names[] = {
146 #include "features.h"
151 static int sched_feat_show(struct seq_file *m, void *v)
155 for (i = 0; i < __SCHED_FEAT_NR; i++) {
156 if (!(sysctl_sched_features & (1UL << i)))
158 seq_printf(m, "%s ", sched_feat_names[i]);
165 #ifdef HAVE_JUMP_LABEL
167 #define jump_label_key__true STATIC_KEY_INIT_TRUE
168 #define jump_label_key__false STATIC_KEY_INIT_FALSE
170 #define SCHED_FEAT(name, enabled) \
171 jump_label_key__##enabled ,
173 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
174 #include "features.h"
179 static void sched_feat_disable(int i)
181 if (static_key_enabled(&sched_feat_keys[i]))
182 static_key_slow_dec(&sched_feat_keys[i]);
185 static void sched_feat_enable(int i)
187 if (!static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_inc(&sched_feat_keys[i]);
191 static void sched_feat_disable(int i) { };
192 static void sched_feat_enable(int i) { };
193 #endif /* HAVE_JUMP_LABEL */
195 static int sched_feat_set(char *cmp)
200 if (strncmp(cmp, "NO_", 3) == 0) {
205 for (i = 0; i < __SCHED_FEAT_NR; i++) {
206 if (strcmp(cmp, sched_feat_names[i]) == 0) {
208 sysctl_sched_features &= ~(1UL << i);
209 sched_feat_disable(i);
211 sysctl_sched_features |= (1UL << i);
212 sched_feat_enable(i);
222 sched_feat_write(struct file *filp, const char __user *ubuf,
223 size_t cnt, loff_t *ppos)
232 if (copy_from_user(&buf, ubuf, cnt))
238 i = sched_feat_set(cmp);
239 if (i == __SCHED_FEAT_NR)
247 static int sched_feat_open(struct inode *inode, struct file *filp)
249 return single_open(filp, sched_feat_show, NULL);
252 static const struct file_operations sched_feat_fops = {
253 .open = sched_feat_open,
254 .write = sched_feat_write,
257 .release = single_release,
260 static __init int sched_init_debug(void)
262 debugfs_create_file("sched_features", 0644, NULL, NULL,
267 late_initcall(sched_init_debug);
268 #endif /* CONFIG_SCHED_DEBUG */
271 * Number of tasks to iterate in a single balance run.
272 * Limited because this is done with IRQs disabled.
274 const_debug unsigned int sysctl_sched_nr_migrate = 32;
277 * period over which we average the RT time consumption, measured
282 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
285 * period over which we measure -rt task cpu usage in us.
288 unsigned int sysctl_sched_rt_period = 1000000;
290 __read_mostly int scheduler_running;
293 * part of the period that we allow rt tasks to run in us.
296 int sysctl_sched_rt_runtime = 950000;
301 * __task_rq_lock - lock the rq @p resides on.
303 static inline struct rq *__task_rq_lock(struct task_struct *p)
308 lockdep_assert_held(&p->pi_lock);
312 raw_spin_lock(&rq->lock);
313 if (likely(rq == task_rq(p)))
315 raw_spin_unlock(&rq->lock);
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
322 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323 __acquires(p->pi_lock)
329 raw_spin_lock_irqsave(&p->pi_lock, *flags);
331 raw_spin_lock(&rq->lock);
332 if (likely(rq == task_rq(p)))
334 raw_spin_unlock(&rq->lock);
335 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
339 static void __task_rq_unlock(struct rq *rq)
342 raw_spin_unlock(&rq->lock);
346 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
348 __releases(p->pi_lock)
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 * this_rq_lock - lock this runqueue and disable interrupts.
357 static struct rq *this_rq_lock(void)
364 raw_spin_lock(&rq->lock);
369 #ifdef CONFIG_SCHED_HRTICK
371 * Use HR-timers to deliver accurate preemption points.
373 * Its all a bit involved since we cannot program an hrt while holding the
374 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
377 * When we get rescheduled we reprogram the hrtick_timer outside of the
381 static void hrtick_clear(struct rq *rq)
383 if (hrtimer_active(&rq->hrtick_timer))
384 hrtimer_cancel(&rq->hrtick_timer);
388 * High-resolution timer tick.
389 * Runs from hardirq context with interrupts disabled.
391 static enum hrtimer_restart hrtick(struct hrtimer *timer)
393 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
395 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
397 raw_spin_lock(&rq->lock);
399 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
400 raw_spin_unlock(&rq->lock);
402 return HRTIMER_NORESTART;
407 * called from hardirq (IPI) context
409 static void __hrtick_start(void *arg)
413 raw_spin_lock(&rq->lock);
414 hrtimer_restart(&rq->hrtick_timer);
415 rq->hrtick_csd_pending = 0;
416 raw_spin_unlock(&rq->lock);
420 * Called to set the hrtick timer state.
422 * called with rq->lock held and irqs disabled
424 void hrtick_start(struct rq *rq, u64 delay)
426 struct hrtimer *timer = &rq->hrtick_timer;
427 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
429 hrtimer_set_expires(timer, time);
431 if (rq == this_rq()) {
432 hrtimer_restart(timer);
433 } else if (!rq->hrtick_csd_pending) {
434 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
435 rq->hrtick_csd_pending = 1;
440 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
442 int cpu = (int)(long)hcpu;
445 case CPU_UP_CANCELED:
446 case CPU_UP_CANCELED_FROZEN:
447 case CPU_DOWN_PREPARE:
448 case CPU_DOWN_PREPARE_FROZEN:
450 case CPU_DEAD_FROZEN:
451 hrtick_clear(cpu_rq(cpu));
458 static __init void init_hrtick(void)
460 hotcpu_notifier(hotplug_hrtick, 0);
464 * Called to set the hrtick timer state.
466 * called with rq->lock held and irqs disabled
468 void hrtick_start(struct rq *rq, u64 delay)
470 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
471 HRTIMER_MODE_REL_PINNED, 0);
474 static inline void init_hrtick(void)
477 #endif /* CONFIG_SMP */
479 static void init_rq_hrtick(struct rq *rq)
482 rq->hrtick_csd_pending = 0;
484 rq->hrtick_csd.flags = 0;
485 rq->hrtick_csd.func = __hrtick_start;
486 rq->hrtick_csd.info = rq;
489 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
490 rq->hrtick_timer.function = hrtick;
492 #else /* CONFIG_SCHED_HRTICK */
493 static inline void hrtick_clear(struct rq *rq)
497 static inline void init_rq_hrtick(struct rq *rq)
501 static inline void init_hrtick(void)
504 #endif /* CONFIG_SCHED_HRTICK */
507 * resched_task - mark a task 'to be rescheduled now'.
509 * On UP this means the setting of the need_resched flag, on SMP it
510 * might also involve a cross-CPU call to trigger the scheduler on
515 #ifndef tsk_is_polling
516 #define tsk_is_polling(t) 0
519 void resched_task(struct task_struct *p)
523 assert_raw_spin_locked(&task_rq(p)->lock);
525 if (test_tsk_need_resched(p))
528 set_tsk_need_resched(p);
531 if (cpu == smp_processor_id())
534 /* NEED_RESCHED must be visible before we test polling */
536 if (!tsk_is_polling(p))
537 smp_send_reschedule(cpu);
540 void resched_cpu(int cpu)
542 struct rq *rq = cpu_rq(cpu);
545 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
547 resched_task(cpu_curr(cpu));
548 raw_spin_unlock_irqrestore(&rq->lock, flags);
553 * In the semi idle case, use the nearest busy cpu for migrating timers
554 * from an idle cpu. This is good for power-savings.
556 * We don't do similar optimization for completely idle system, as
557 * selecting an idle cpu will add more delays to the timers than intended
558 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
560 int get_nohz_timer_target(void)
562 int cpu = smp_processor_id();
564 struct sched_domain *sd;
567 for_each_domain(cpu, sd) {
568 for_each_cpu(i, sched_domain_span(sd)) {
580 * When add_timer_on() enqueues a timer into the timer wheel of an
581 * idle CPU then this timer might expire before the next timer event
582 * which is scheduled to wake up that CPU. In case of a completely
583 * idle system the next event might even be infinite time into the
584 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
585 * leaves the inner idle loop so the newly added timer is taken into
586 * account when the CPU goes back to idle and evaluates the timer
587 * wheel for the next timer event.
589 void wake_up_idle_cpu(int cpu)
591 struct rq *rq = cpu_rq(cpu);
593 if (cpu == smp_processor_id())
597 * This is safe, as this function is called with the timer
598 * wheel base lock of (cpu) held. When the CPU is on the way
599 * to idle and has not yet set rq->curr to idle then it will
600 * be serialized on the timer wheel base lock and take the new
601 * timer into account automatically.
603 if (rq->curr != rq->idle)
607 * We can set TIF_RESCHED on the idle task of the other CPU
608 * lockless. The worst case is that the other CPU runs the
609 * idle task through an additional NOOP schedule()
611 set_tsk_need_resched(rq->idle);
613 /* NEED_RESCHED must be visible before we test polling */
615 if (!tsk_is_polling(rq->idle))
616 smp_send_reschedule(cpu);
619 static inline bool got_nohz_idle_kick(void)
621 int cpu = smp_processor_id();
622 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
625 #else /* CONFIG_NO_HZ */
627 static inline bool got_nohz_idle_kick(void)
632 #endif /* CONFIG_NO_HZ */
634 void sched_avg_update(struct rq *rq)
636 s64 period = sched_avg_period();
638 while ((s64)(rq->clock - rq->age_stamp) > period) {
640 * Inline assembly required to prevent the compiler
641 * optimising this loop into a divmod call.
642 * See __iter_div_u64_rem() for another example of this.
644 asm("" : "+rm" (rq->age_stamp));
645 rq->age_stamp += period;
650 #else /* !CONFIG_SMP */
651 void resched_task(struct task_struct *p)
653 assert_raw_spin_locked(&task_rq(p)->lock);
654 set_tsk_need_resched(p);
656 #endif /* CONFIG_SMP */
658 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
659 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
661 * Iterate task_group tree rooted at *from, calling @down when first entering a
662 * node and @up when leaving it for the final time.
664 * Caller must hold rcu_lock or sufficient equivalent.
666 int walk_tg_tree_from(struct task_group *from,
667 tg_visitor down, tg_visitor up, void *data)
669 struct task_group *parent, *child;
675 ret = (*down)(parent, data);
678 list_for_each_entry_rcu(child, &parent->children, siblings) {
685 ret = (*up)(parent, data);
686 if (ret || parent == from)
690 parent = parent->parent;
697 int tg_nop(struct task_group *tg, void *data)
703 static void set_load_weight(struct task_struct *p)
705 int prio = p->static_prio - MAX_RT_PRIO;
706 struct load_weight *load = &p->se.load;
709 * SCHED_IDLE tasks get minimal weight:
711 if (p->policy == SCHED_IDLE) {
712 load->weight = scale_load(WEIGHT_IDLEPRIO);
713 load->inv_weight = WMULT_IDLEPRIO;
717 load->weight = scale_load(prio_to_weight[prio]);
718 load->inv_weight = prio_to_wmult[prio];
721 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
724 sched_info_queued(p);
725 p->sched_class->enqueue_task(rq, p, flags);
728 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
731 sched_info_dequeued(p);
732 p->sched_class->dequeue_task(rq, p, flags);
735 void activate_task(struct rq *rq, struct task_struct *p, int flags)
737 if (task_contributes_to_load(p))
738 rq->nr_uninterruptible--;
740 enqueue_task(rq, p, flags);
743 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
745 if (task_contributes_to_load(p))
746 rq->nr_uninterruptible++;
748 dequeue_task(rq, p, flags);
751 static void update_rq_clock_task(struct rq *rq, s64 delta)
754 * In theory, the compile should just see 0 here, and optimize out the call
755 * to sched_rt_avg_update. But I don't trust it...
757 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
758 s64 steal = 0, irq_delta = 0;
760 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
761 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
764 * Since irq_time is only updated on {soft,}irq_exit, we might run into
765 * this case when a previous update_rq_clock() happened inside a
768 * When this happens, we stop ->clock_task and only update the
769 * prev_irq_time stamp to account for the part that fit, so that a next
770 * update will consume the rest. This ensures ->clock_task is
773 * It does however cause some slight miss-attribution of {soft,}irq
774 * time, a more accurate solution would be to update the irq_time using
775 * the current rq->clock timestamp, except that would require using
778 if (irq_delta > delta)
781 rq->prev_irq_time += irq_delta;
784 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
785 if (static_key_false((¶virt_steal_rq_enabled))) {
788 steal = paravirt_steal_clock(cpu_of(rq));
789 steal -= rq->prev_steal_time_rq;
791 if (unlikely(steal > delta))
794 st = steal_ticks(steal);
795 steal = st * TICK_NSEC;
797 rq->prev_steal_time_rq += steal;
803 rq->clock_task += delta;
805 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
806 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
807 sched_rt_avg_update(rq, irq_delta + steal);
811 void sched_set_stop_task(int cpu, struct task_struct *stop)
813 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
814 struct task_struct *old_stop = cpu_rq(cpu)->stop;
818 * Make it appear like a SCHED_FIFO task, its something
819 * userspace knows about and won't get confused about.
821 * Also, it will make PI more or less work without too
822 * much confusion -- but then, stop work should not
823 * rely on PI working anyway.
825 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
827 stop->sched_class = &stop_sched_class;
830 cpu_rq(cpu)->stop = stop;
834 * Reset it back to a normal scheduling class so that
835 * it can die in pieces.
837 old_stop->sched_class = &rt_sched_class;
842 * __normal_prio - return the priority that is based on the static prio
844 static inline int __normal_prio(struct task_struct *p)
846 return p->static_prio;
850 * Calculate the expected normal priority: i.e. priority
851 * without taking RT-inheritance into account. Might be
852 * boosted by interactivity modifiers. Changes upon fork,
853 * setprio syscalls, and whenever the interactivity
854 * estimator recalculates.
856 static inline int normal_prio(struct task_struct *p)
860 if (task_has_rt_policy(p))
861 prio = MAX_RT_PRIO-1 - p->rt_priority;
863 prio = __normal_prio(p);
868 * Calculate the current priority, i.e. the priority
869 * taken into account by the scheduler. This value might
870 * be boosted by RT tasks, or might be boosted by
871 * interactivity modifiers. Will be RT if the task got
872 * RT-boosted. If not then it returns p->normal_prio.
874 static int effective_prio(struct task_struct *p)
876 p->normal_prio = normal_prio(p);
878 * If we are RT tasks or we were boosted to RT priority,
879 * keep the priority unchanged. Otherwise, update priority
880 * to the normal priority:
882 if (!rt_prio(p->prio))
883 return p->normal_prio;
888 * task_curr - is this task currently executing on a CPU?
889 * @p: the task in question.
891 inline int task_curr(const struct task_struct *p)
893 return cpu_curr(task_cpu(p)) == p;
896 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
897 const struct sched_class *prev_class,
900 if (prev_class != p->sched_class) {
901 if (prev_class->switched_from)
902 prev_class->switched_from(rq, p);
903 p->sched_class->switched_to(rq, p);
904 } else if (oldprio != p->prio)
905 p->sched_class->prio_changed(rq, p, oldprio);
908 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
910 const struct sched_class *class;
912 if (p->sched_class == rq->curr->sched_class) {
913 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
915 for_each_class(class) {
916 if (class == rq->curr->sched_class)
918 if (class == p->sched_class) {
919 resched_task(rq->curr);
926 * A queue event has occurred, and we're going to schedule. In
927 * this case, we can save a useless back to back clock update.
929 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
930 rq->skip_clock_update = 1;
934 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
936 #ifdef CONFIG_SCHED_DEBUG
938 * We should never call set_task_cpu() on a blocked task,
939 * ttwu() will sort out the placement.
941 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
942 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
944 #ifdef CONFIG_LOCKDEP
946 * The caller should hold either p->pi_lock or rq->lock, when changing
947 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
949 * sched_move_task() holds both and thus holding either pins the cgroup,
952 * Furthermore, all task_rq users should acquire both locks, see
955 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
956 lockdep_is_held(&task_rq(p)->lock)));
960 trace_sched_migrate_task(p, new_cpu);
962 if (task_cpu(p) != new_cpu) {
963 p->se.nr_migrations++;
964 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
967 __set_task_cpu(p, new_cpu);
970 struct migration_arg {
971 struct task_struct *task;
975 static int migration_cpu_stop(void *data);
978 * wait_task_inactive - wait for a thread to unschedule.
980 * If @match_state is nonzero, it's the @p->state value just checked and
981 * not expected to change. If it changes, i.e. @p might have woken up,
982 * then return zero. When we succeed in waiting for @p to be off its CPU,
983 * we return a positive number (its total switch count). If a second call
984 * a short while later returns the same number, the caller can be sure that
985 * @p has remained unscheduled the whole time.
987 * The caller must ensure that the task *will* unschedule sometime soon,
988 * else this function might spin for a *long* time. This function can't
989 * be called with interrupts off, or it may introduce deadlock with
990 * smp_call_function() if an IPI is sent by the same process we are
991 * waiting to become inactive.
993 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1002 * We do the initial early heuristics without holding
1003 * any task-queue locks at all. We'll only try to get
1004 * the runqueue lock when things look like they will
1010 * If the task is actively running on another CPU
1011 * still, just relax and busy-wait without holding
1014 * NOTE! Since we don't hold any locks, it's not
1015 * even sure that "rq" stays as the right runqueue!
1016 * But we don't care, since "task_running()" will
1017 * return false if the runqueue has changed and p
1018 * is actually now running somewhere else!
1020 while (task_running(rq, p)) {
1021 if (match_state && unlikely(p->state != match_state))
1027 * Ok, time to look more closely! We need the rq
1028 * lock now, to be *sure*. If we're wrong, we'll
1029 * just go back and repeat.
1031 rq = task_rq_lock(p, &flags);
1032 trace_sched_wait_task(p);
1033 running = task_running(rq, p);
1036 if (!match_state || p->state == match_state)
1037 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1038 task_rq_unlock(rq, p, &flags);
1041 * If it changed from the expected state, bail out now.
1043 if (unlikely(!ncsw))
1047 * Was it really running after all now that we
1048 * checked with the proper locks actually held?
1050 * Oops. Go back and try again..
1052 if (unlikely(running)) {
1058 * It's not enough that it's not actively running,
1059 * it must be off the runqueue _entirely_, and not
1062 * So if it was still runnable (but just not actively
1063 * running right now), it's preempted, and we should
1064 * yield - it could be a while.
1066 if (unlikely(on_rq)) {
1067 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1069 set_current_state(TASK_UNINTERRUPTIBLE);
1070 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1075 * Ahh, all good. It wasn't running, and it wasn't
1076 * runnable, which means that it will never become
1077 * running in the future either. We're all done!
1086 * kick_process - kick a running thread to enter/exit the kernel
1087 * @p: the to-be-kicked thread
1089 * Cause a process which is running on another CPU to enter
1090 * kernel-mode, without any delay. (to get signals handled.)
1092 * NOTE: this function doesn't have to take the runqueue lock,
1093 * because all it wants to ensure is that the remote task enters
1094 * the kernel. If the IPI races and the task has been migrated
1095 * to another CPU then no harm is done and the purpose has been
1098 void kick_process(struct task_struct *p)
1104 if ((cpu != smp_processor_id()) && task_curr(p))
1105 smp_send_reschedule(cpu);
1108 EXPORT_SYMBOL_GPL(kick_process);
1109 #endif /* CONFIG_SMP */
1113 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1115 static int select_fallback_rq(int cpu, struct task_struct *p)
1117 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1118 enum { cpuset, possible, fail } state = cpuset;
1121 /* Look for allowed, online CPU in same node. */
1122 for_each_cpu(dest_cpu, nodemask) {
1123 if (!cpu_online(dest_cpu))
1125 if (!cpu_active(dest_cpu))
1127 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1132 /* Any allowed, online CPU? */
1133 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1134 if (!cpu_online(dest_cpu))
1136 if (!cpu_active(dest_cpu))
1143 /* No more Mr. Nice Guy. */
1144 cpuset_cpus_allowed_fallback(p);
1149 do_set_cpus_allowed(p, cpu_possible_mask);
1160 if (state != cpuset) {
1162 * Don't tell them about moving exiting tasks or
1163 * kernel threads (both mm NULL), since they never
1166 if (p->mm && printk_ratelimit()) {
1167 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1168 task_pid_nr(p), p->comm, cpu);
1176 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1179 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1181 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1184 * In order not to call set_task_cpu() on a blocking task we need
1185 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1188 * Since this is common to all placement strategies, this lives here.
1190 * [ this allows ->select_task() to simply return task_cpu(p) and
1191 * not worry about this generic constraint ]
1193 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1195 cpu = select_fallback_rq(task_cpu(p), p);
1200 static void update_avg(u64 *avg, u64 sample)
1202 s64 diff = sample - *avg;
1208 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1210 #ifdef CONFIG_SCHEDSTATS
1211 struct rq *rq = this_rq();
1214 int this_cpu = smp_processor_id();
1216 if (cpu == this_cpu) {
1217 schedstat_inc(rq, ttwu_local);
1218 schedstat_inc(p, se.statistics.nr_wakeups_local);
1220 struct sched_domain *sd;
1222 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1224 for_each_domain(this_cpu, sd) {
1225 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1226 schedstat_inc(sd, ttwu_wake_remote);
1233 if (wake_flags & WF_MIGRATED)
1234 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1236 #endif /* CONFIG_SMP */
1238 schedstat_inc(rq, ttwu_count);
1239 schedstat_inc(p, se.statistics.nr_wakeups);
1241 if (wake_flags & WF_SYNC)
1242 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1244 #endif /* CONFIG_SCHEDSTATS */
1247 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1249 activate_task(rq, p, en_flags);
1252 /* if a worker is waking up, notify workqueue */
1253 if (p->flags & PF_WQ_WORKER)
1254 wq_worker_waking_up(p, cpu_of(rq));
1258 * Mark the task runnable and perform wakeup-preemption.
1261 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1263 trace_sched_wakeup(p, true);
1264 check_preempt_curr(rq, p, wake_flags);
1266 p->state = TASK_RUNNING;
1268 if (p->sched_class->task_woken)
1269 p->sched_class->task_woken(rq, p);
1271 if (rq->idle_stamp) {
1272 u64 delta = rq->clock - rq->idle_stamp;
1273 u64 max = 2*sysctl_sched_migration_cost;
1278 update_avg(&rq->avg_idle, delta);
1285 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1288 if (p->sched_contributes_to_load)
1289 rq->nr_uninterruptible--;
1292 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1293 ttwu_do_wakeup(rq, p, wake_flags);
1297 * Called in case the task @p isn't fully descheduled from its runqueue,
1298 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1299 * since all we need to do is flip p->state to TASK_RUNNING, since
1300 * the task is still ->on_rq.
1302 static int ttwu_remote(struct task_struct *p, int wake_flags)
1307 rq = __task_rq_lock(p);
1309 ttwu_do_wakeup(rq, p, wake_flags);
1312 __task_rq_unlock(rq);
1318 static void sched_ttwu_pending(void)
1320 struct rq *rq = this_rq();
1321 struct llist_node *llist = llist_del_all(&rq->wake_list);
1322 struct task_struct *p;
1324 raw_spin_lock(&rq->lock);
1327 p = llist_entry(llist, struct task_struct, wake_entry);
1328 llist = llist_next(llist);
1329 ttwu_do_activate(rq, p, 0);
1332 raw_spin_unlock(&rq->lock);
1335 void scheduler_ipi(void)
1337 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1341 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1342 * traditionally all their work was done from the interrupt return
1343 * path. Now that we actually do some work, we need to make sure
1346 * Some archs already do call them, luckily irq_enter/exit nest
1349 * Arguably we should visit all archs and update all handlers,
1350 * however a fair share of IPIs are still resched only so this would
1351 * somewhat pessimize the simple resched case.
1354 sched_ttwu_pending();
1357 * Check if someone kicked us for doing the nohz idle load balance.
1359 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1360 this_rq()->idle_balance = 1;
1361 raise_softirq_irqoff(SCHED_SOFTIRQ);
1366 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1368 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1369 smp_send_reschedule(cpu);
1372 bool cpus_share_cache(int this_cpu, int that_cpu)
1374 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1376 #endif /* CONFIG_SMP */
1378 static void ttwu_queue(struct task_struct *p, int cpu)
1380 struct rq *rq = cpu_rq(cpu);
1382 #if defined(CONFIG_SMP)
1383 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1384 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1385 ttwu_queue_remote(p, cpu);
1390 raw_spin_lock(&rq->lock);
1391 ttwu_do_activate(rq, p, 0);
1392 raw_spin_unlock(&rq->lock);
1396 * try_to_wake_up - wake up a thread
1397 * @p: the thread to be awakened
1398 * @state: the mask of task states that can be woken
1399 * @wake_flags: wake modifier flags (WF_*)
1401 * Put it on the run-queue if it's not already there. The "current"
1402 * thread is always on the run-queue (except when the actual
1403 * re-schedule is in progress), and as such you're allowed to do
1404 * the simpler "current->state = TASK_RUNNING" to mark yourself
1405 * runnable without the overhead of this.
1407 * Returns %true if @p was woken up, %false if it was already running
1408 * or @state didn't match @p's state.
1411 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1413 unsigned long flags;
1414 int cpu, success = 0;
1417 raw_spin_lock_irqsave(&p->pi_lock, flags);
1418 if (!(p->state & state))
1421 success = 1; /* we're going to change ->state */
1424 if (p->on_rq && ttwu_remote(p, wake_flags))
1429 * If the owning (remote) cpu is still in the middle of schedule() with
1430 * this task as prev, wait until its done referencing the task.
1435 * Pairs with the smp_wmb() in finish_lock_switch().
1439 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1440 p->state = TASK_WAKING;
1442 if (p->sched_class->task_waking)
1443 p->sched_class->task_waking(p);
1445 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1446 if (task_cpu(p) != cpu) {
1447 wake_flags |= WF_MIGRATED;
1448 set_task_cpu(p, cpu);
1450 #endif /* CONFIG_SMP */
1454 ttwu_stat(p, cpu, wake_flags);
1456 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1462 * try_to_wake_up_local - try to wake up a local task with rq lock held
1463 * @p: the thread to be awakened
1465 * Put @p on the run-queue if it's not already there. The caller must
1466 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1469 static void try_to_wake_up_local(struct task_struct *p)
1471 struct rq *rq = task_rq(p);
1473 BUG_ON(rq != this_rq());
1474 BUG_ON(p == current);
1475 lockdep_assert_held(&rq->lock);
1477 if (!raw_spin_trylock(&p->pi_lock)) {
1478 raw_spin_unlock(&rq->lock);
1479 raw_spin_lock(&p->pi_lock);
1480 raw_spin_lock(&rq->lock);
1483 if (!(p->state & TASK_NORMAL))
1487 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1489 ttwu_do_wakeup(rq, p, 0);
1490 ttwu_stat(p, smp_processor_id(), 0);
1492 raw_spin_unlock(&p->pi_lock);
1496 * wake_up_process - Wake up a specific process
1497 * @p: The process to be woken up.
1499 * Attempt to wake up the nominated process and move it to the set of runnable
1500 * processes. Returns 1 if the process was woken up, 0 if it was already
1503 * It may be assumed that this function implies a write memory barrier before
1504 * changing the task state if and only if any tasks are woken up.
1506 int wake_up_process(struct task_struct *p)
1508 return try_to_wake_up(p, TASK_ALL, 0);
1510 EXPORT_SYMBOL(wake_up_process);
1512 int wake_up_state(struct task_struct *p, unsigned int state)
1514 return try_to_wake_up(p, state, 0);
1518 * Perform scheduler related setup for a newly forked process p.
1519 * p is forked by current.
1521 * __sched_fork() is basic setup used by init_idle() too:
1523 static void __sched_fork(struct task_struct *p)
1528 p->se.exec_start = 0;
1529 p->se.sum_exec_runtime = 0;
1530 p->se.prev_sum_exec_runtime = 0;
1531 p->se.nr_migrations = 0;
1533 INIT_LIST_HEAD(&p->se.group_node);
1535 #ifdef CONFIG_SCHEDSTATS
1536 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1539 INIT_LIST_HEAD(&p->rt.run_list);
1541 #ifdef CONFIG_PREEMPT_NOTIFIERS
1542 INIT_HLIST_HEAD(&p->preempt_notifiers);
1545 #ifdef CONFIG_NUMA_BALANCING
1546 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1547 p->mm->numa_next_scan = jiffies;
1548 p->mm->numa_next_reset = jiffies;
1549 p->mm->numa_scan_seq = 0;
1552 p->node_stamp = 0ULL;
1553 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1554 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1555 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1556 p->numa_work.next = &p->numa_work;
1557 #endif /* CONFIG_NUMA_BALANCING */
1560 #ifdef CONFIG_NUMA_BALANCING
1561 #ifdef CONFIG_SCHED_DEBUG
1562 void set_numabalancing_state(bool enabled)
1565 sched_feat_set("NUMA");
1567 sched_feat_set("NO_NUMA");
1570 __read_mostly bool numabalancing_enabled;
1572 void set_numabalancing_state(bool enabled)
1574 numabalancing_enabled = enabled;
1576 #endif /* CONFIG_SCHED_DEBUG */
1577 #endif /* CONFIG_NUMA_BALANCING */
1580 * fork()/clone()-time setup:
1582 void sched_fork(struct task_struct *p)
1584 unsigned long flags;
1585 int cpu = get_cpu();
1589 * We mark the process as running here. This guarantees that
1590 * nobody will actually run it, and a signal or other external
1591 * event cannot wake it up and insert it on the runqueue either.
1593 p->state = TASK_RUNNING;
1596 * Make sure we do not leak PI boosting priority to the child.
1598 p->prio = current->normal_prio;
1601 * Revert to default priority/policy on fork if requested.
1603 if (unlikely(p->sched_reset_on_fork)) {
1604 if (task_has_rt_policy(p)) {
1605 p->policy = SCHED_NORMAL;
1606 p->static_prio = NICE_TO_PRIO(0);
1608 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1609 p->static_prio = NICE_TO_PRIO(0);
1611 p->prio = p->normal_prio = __normal_prio(p);
1615 * We don't need the reset flag anymore after the fork. It has
1616 * fulfilled its duty:
1618 p->sched_reset_on_fork = 0;
1621 if (!rt_prio(p->prio))
1622 p->sched_class = &fair_sched_class;
1624 if (p->sched_class->task_fork)
1625 p->sched_class->task_fork(p);
1628 * The child is not yet in the pid-hash so no cgroup attach races,
1629 * and the cgroup is pinned to this child due to cgroup_fork()
1630 * is ran before sched_fork().
1632 * Silence PROVE_RCU.
1634 raw_spin_lock_irqsave(&p->pi_lock, flags);
1635 set_task_cpu(p, cpu);
1636 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1639 if (likely(sched_info_on()))
1640 memset(&p->sched_info, 0, sizeof(p->sched_info));
1642 #if defined(CONFIG_SMP)
1645 #ifdef CONFIG_PREEMPT_COUNT
1646 /* Want to start with kernel preemption disabled. */
1647 task_thread_info(p)->preempt_count = 1;
1650 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1657 * wake_up_new_task - wake up a newly created task for the first time.
1659 * This function will do some initial scheduler statistics housekeeping
1660 * that must be done for every newly created context, then puts the task
1661 * on the runqueue and wakes it.
1663 void wake_up_new_task(struct task_struct *p)
1665 unsigned long flags;
1668 raw_spin_lock_irqsave(&p->pi_lock, flags);
1671 * Fork balancing, do it here and not earlier because:
1672 * - cpus_allowed can change in the fork path
1673 * - any previously selected cpu might disappear through hotplug
1675 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1678 rq = __task_rq_lock(p);
1679 activate_task(rq, p, 0);
1681 trace_sched_wakeup_new(p, true);
1682 check_preempt_curr(rq, p, WF_FORK);
1684 if (p->sched_class->task_woken)
1685 p->sched_class->task_woken(rq, p);
1687 task_rq_unlock(rq, p, &flags);
1690 #ifdef CONFIG_PREEMPT_NOTIFIERS
1693 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1694 * @notifier: notifier struct to register
1696 void preempt_notifier_register(struct preempt_notifier *notifier)
1698 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1700 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1703 * preempt_notifier_unregister - no longer interested in preemption notifications
1704 * @notifier: notifier struct to unregister
1706 * This is safe to call from within a preemption notifier.
1708 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1710 hlist_del(¬ifier->link);
1712 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1714 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1716 struct preempt_notifier *notifier;
1717 struct hlist_node *node;
1719 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1720 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1724 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1725 struct task_struct *next)
1727 struct preempt_notifier *notifier;
1728 struct hlist_node *node;
1730 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1731 notifier->ops->sched_out(notifier, next);
1734 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1736 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1741 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1742 struct task_struct *next)
1746 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1749 * prepare_task_switch - prepare to switch tasks
1750 * @rq: the runqueue preparing to switch
1751 * @prev: the current task that is being switched out
1752 * @next: the task we are going to switch to.
1754 * This is called with the rq lock held and interrupts off. It must
1755 * be paired with a subsequent finish_task_switch after the context
1758 * prepare_task_switch sets up locking and calls architecture specific
1762 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1763 struct task_struct *next)
1765 trace_sched_switch(prev, next);
1766 sched_info_switch(prev, next);
1767 perf_event_task_sched_out(prev, next);
1768 fire_sched_out_preempt_notifiers(prev, next);
1769 prepare_lock_switch(rq, next);
1770 prepare_arch_switch(next);
1774 * finish_task_switch - clean up after a task-switch
1775 * @rq: runqueue associated with task-switch
1776 * @prev: the thread we just switched away from.
1778 * finish_task_switch must be called after the context switch, paired
1779 * with a prepare_task_switch call before the context switch.
1780 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1781 * and do any other architecture-specific cleanup actions.
1783 * Note that we may have delayed dropping an mm in context_switch(). If
1784 * so, we finish that here outside of the runqueue lock. (Doing it
1785 * with the lock held can cause deadlocks; see schedule() for
1788 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1789 __releases(rq->lock)
1791 struct mm_struct *mm = rq->prev_mm;
1797 * A task struct has one reference for the use as "current".
1798 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1799 * schedule one last time. The schedule call will never return, and
1800 * the scheduled task must drop that reference.
1801 * The test for TASK_DEAD must occur while the runqueue locks are
1802 * still held, otherwise prev could be scheduled on another cpu, die
1803 * there before we look at prev->state, and then the reference would
1805 * Manfred Spraul <manfred@colorfullife.com>
1807 prev_state = prev->state;
1808 vtime_task_switch(prev);
1809 finish_arch_switch(prev);
1810 perf_event_task_sched_in(prev, current);
1811 finish_lock_switch(rq, prev);
1812 finish_arch_post_lock_switch();
1814 fire_sched_in_preempt_notifiers(current);
1817 if (unlikely(prev_state == TASK_DEAD)) {
1819 * Remove function-return probe instances associated with this
1820 * task and put them back on the free list.
1822 kprobe_flush_task(prev);
1823 put_task_struct(prev);
1829 /* assumes rq->lock is held */
1830 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1832 if (prev->sched_class->pre_schedule)
1833 prev->sched_class->pre_schedule(rq, prev);
1836 /* rq->lock is NOT held, but preemption is disabled */
1837 static inline void post_schedule(struct rq *rq)
1839 if (rq->post_schedule) {
1840 unsigned long flags;
1842 raw_spin_lock_irqsave(&rq->lock, flags);
1843 if (rq->curr->sched_class->post_schedule)
1844 rq->curr->sched_class->post_schedule(rq);
1845 raw_spin_unlock_irqrestore(&rq->lock, flags);
1847 rq->post_schedule = 0;
1853 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1857 static inline void post_schedule(struct rq *rq)
1864 * schedule_tail - first thing a freshly forked thread must call.
1865 * @prev: the thread we just switched away from.
1867 asmlinkage void schedule_tail(struct task_struct *prev)
1868 __releases(rq->lock)
1870 struct rq *rq = this_rq();
1872 finish_task_switch(rq, prev);
1875 * FIXME: do we need to worry about rq being invalidated by the
1880 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1881 /* In this case, finish_task_switch does not reenable preemption */
1884 if (current->set_child_tid)
1885 put_user(task_pid_vnr(current), current->set_child_tid);
1889 * context_switch - switch to the new MM and the new
1890 * thread's register state.
1893 context_switch(struct rq *rq, struct task_struct *prev,
1894 struct task_struct *next)
1896 struct mm_struct *mm, *oldmm;
1898 prepare_task_switch(rq, prev, next);
1901 oldmm = prev->active_mm;
1903 * For paravirt, this is coupled with an exit in switch_to to
1904 * combine the page table reload and the switch backend into
1907 arch_start_context_switch(prev);
1910 next->active_mm = oldmm;
1911 atomic_inc(&oldmm->mm_count);
1912 enter_lazy_tlb(oldmm, next);
1914 switch_mm(oldmm, mm, next);
1917 prev->active_mm = NULL;
1918 rq->prev_mm = oldmm;
1921 * Since the runqueue lock will be released by the next
1922 * task (which is an invalid locking op but in the case
1923 * of the scheduler it's an obvious special-case), so we
1924 * do an early lockdep release here:
1926 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1927 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1930 /* Here we just switch the register state and the stack. */
1931 rcu_switch(prev, next);
1932 switch_to(prev, next, prev);
1936 * this_rq must be evaluated again because prev may have moved
1937 * CPUs since it called schedule(), thus the 'rq' on its stack
1938 * frame will be invalid.
1940 finish_task_switch(this_rq(), prev);
1944 * nr_running, nr_uninterruptible and nr_context_switches:
1946 * externally visible scheduler statistics: current number of runnable
1947 * threads, current number of uninterruptible-sleeping threads, total
1948 * number of context switches performed since bootup.
1950 unsigned long nr_running(void)
1952 unsigned long i, sum = 0;
1954 for_each_online_cpu(i)
1955 sum += cpu_rq(i)->nr_running;
1960 unsigned long nr_uninterruptible(void)
1962 unsigned long i, sum = 0;
1964 for_each_possible_cpu(i)
1965 sum += cpu_rq(i)->nr_uninterruptible;
1968 * Since we read the counters lockless, it might be slightly
1969 * inaccurate. Do not allow it to go below zero though:
1971 if (unlikely((long)sum < 0))
1977 unsigned long long nr_context_switches(void)
1980 unsigned long long sum = 0;
1982 for_each_possible_cpu(i)
1983 sum += cpu_rq(i)->nr_switches;
1988 unsigned long nr_iowait(void)
1990 unsigned long i, sum = 0;
1992 for_each_possible_cpu(i)
1993 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1998 unsigned long nr_iowait_cpu(int cpu)
2000 struct rq *this = cpu_rq(cpu);
2001 return atomic_read(&this->nr_iowait);
2004 unsigned long this_cpu_load(void)
2006 struct rq *this = this_rq();
2007 return this->cpu_load[0];
2012 * Global load-average calculations
2014 * We take a distributed and async approach to calculating the global load-avg
2015 * in order to minimize overhead.
2017 * The global load average is an exponentially decaying average of nr_running +
2018 * nr_uninterruptible.
2020 * Once every LOAD_FREQ:
2023 * for_each_possible_cpu(cpu)
2024 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2026 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2028 * Due to a number of reasons the above turns in the mess below:
2030 * - for_each_possible_cpu() is prohibitively expensive on machines with
2031 * serious number of cpus, therefore we need to take a distributed approach
2032 * to calculating nr_active.
2034 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2035 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2037 * So assuming nr_active := 0 when we start out -- true per definition, we
2038 * can simply take per-cpu deltas and fold those into a global accumulate
2039 * to obtain the same result. See calc_load_fold_active().
2041 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2042 * across the machine, we assume 10 ticks is sufficient time for every
2043 * cpu to have completed this task.
2045 * This places an upper-bound on the IRQ-off latency of the machine. Then
2046 * again, being late doesn't loose the delta, just wrecks the sample.
2048 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2049 * this would add another cross-cpu cacheline miss and atomic operation
2050 * to the wakeup path. Instead we increment on whatever cpu the task ran
2051 * when it went into uninterruptible state and decrement on whatever cpu
2052 * did the wakeup. This means that only the sum of nr_uninterruptible over
2053 * all cpus yields the correct result.
2055 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2058 /* Variables and functions for calc_load */
2059 static atomic_long_t calc_load_tasks;
2060 static unsigned long calc_load_update;
2061 unsigned long avenrun[3];
2062 EXPORT_SYMBOL(avenrun); /* should be removed */
2065 * get_avenrun - get the load average array
2066 * @loads: pointer to dest load array
2067 * @offset: offset to add
2068 * @shift: shift count to shift the result left
2070 * These values are estimates at best, so no need for locking.
2072 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2074 loads[0] = (avenrun[0] + offset) << shift;
2075 loads[1] = (avenrun[1] + offset) << shift;
2076 loads[2] = (avenrun[2] + offset) << shift;
2079 static long calc_load_fold_active(struct rq *this_rq)
2081 long nr_active, delta = 0;
2083 nr_active = this_rq->nr_running;
2084 nr_active += (long) this_rq->nr_uninterruptible;
2086 if (nr_active != this_rq->calc_load_active) {
2087 delta = nr_active - this_rq->calc_load_active;
2088 this_rq->calc_load_active = nr_active;
2095 * a1 = a0 * e + a * (1 - e)
2097 static unsigned long
2098 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2101 load += active * (FIXED_1 - exp);
2102 load += 1UL << (FSHIFT - 1);
2103 return load >> FSHIFT;
2108 * Handle NO_HZ for the global load-average.
2110 * Since the above described distributed algorithm to compute the global
2111 * load-average relies on per-cpu sampling from the tick, it is affected by
2114 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2115 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2116 * when we read the global state.
2118 * Obviously reality has to ruin such a delightfully simple scheme:
2120 * - When we go NO_HZ idle during the window, we can negate our sample
2121 * contribution, causing under-accounting.
2123 * We avoid this by keeping two idle-delta counters and flipping them
2124 * when the window starts, thus separating old and new NO_HZ load.
2126 * The only trick is the slight shift in index flip for read vs write.
2130 * |-|-----------|-|-----------|-|-----------|-|
2131 * r:0 0 1 1 0 0 1 1 0
2132 * w:0 1 1 0 0 1 1 0 0
2134 * This ensures we'll fold the old idle contribution in this window while
2135 * accumlating the new one.
2137 * - When we wake up from NO_HZ idle during the window, we push up our
2138 * contribution, since we effectively move our sample point to a known
2141 * This is solved by pushing the window forward, and thus skipping the
2142 * sample, for this cpu (effectively using the idle-delta for this cpu which
2143 * was in effect at the time the window opened). This also solves the issue
2144 * of having to deal with a cpu having been in NOHZ idle for multiple
2145 * LOAD_FREQ intervals.
2147 * When making the ILB scale, we should try to pull this in as well.
2149 static atomic_long_t calc_load_idle[2];
2150 static int calc_load_idx;
2152 static inline int calc_load_write_idx(void)
2154 int idx = calc_load_idx;
2157 * See calc_global_nohz(), if we observe the new index, we also
2158 * need to observe the new update time.
2163 * If the folding window started, make sure we start writing in the
2166 if (!time_before(jiffies, calc_load_update))
2172 static inline int calc_load_read_idx(void)
2174 return calc_load_idx & 1;
2177 void calc_load_enter_idle(void)
2179 struct rq *this_rq = this_rq();
2183 * We're going into NOHZ mode, if there's any pending delta, fold it
2184 * into the pending idle delta.
2186 delta = calc_load_fold_active(this_rq);
2188 int idx = calc_load_write_idx();
2189 atomic_long_add(delta, &calc_load_idle[idx]);
2193 void calc_load_exit_idle(void)
2195 struct rq *this_rq = this_rq();
2198 * If we're still before the sample window, we're done.
2200 if (time_before(jiffies, this_rq->calc_load_update))
2204 * We woke inside or after the sample window, this means we're already
2205 * accounted through the nohz accounting, so skip the entire deal and
2206 * sync up for the next window.
2208 this_rq->calc_load_update = calc_load_update;
2209 if (time_before(jiffies, this_rq->calc_load_update + 10))
2210 this_rq->calc_load_update += LOAD_FREQ;
2213 static long calc_load_fold_idle(void)
2215 int idx = calc_load_read_idx();
2218 if (atomic_long_read(&calc_load_idle[idx]))
2219 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2225 * fixed_power_int - compute: x^n, in O(log n) time
2227 * @x: base of the power
2228 * @frac_bits: fractional bits of @x
2229 * @n: power to raise @x to.
2231 * By exploiting the relation between the definition of the natural power
2232 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2233 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2234 * (where: n_i \elem {0, 1}, the binary vector representing n),
2235 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2236 * of course trivially computable in O(log_2 n), the length of our binary
2239 static unsigned long
2240 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2242 unsigned long result = 1UL << frac_bits;
2247 result += 1UL << (frac_bits - 1);
2248 result >>= frac_bits;
2254 x += 1UL << (frac_bits - 1);
2262 * a1 = a0 * e + a * (1 - e)
2264 * a2 = a1 * e + a * (1 - e)
2265 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2266 * = a0 * e^2 + a * (1 - e) * (1 + e)
2268 * a3 = a2 * e + a * (1 - e)
2269 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2270 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2274 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2275 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2276 * = a0 * e^n + a * (1 - e^n)
2278 * [1] application of the geometric series:
2281 * S_n := \Sum x^i = -------------
2284 static unsigned long
2285 calc_load_n(unsigned long load, unsigned long exp,
2286 unsigned long active, unsigned int n)
2289 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2293 * NO_HZ can leave us missing all per-cpu ticks calling
2294 * calc_load_account_active(), but since an idle CPU folds its delta into
2295 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2296 * in the pending idle delta if our idle period crossed a load cycle boundary.
2298 * Once we've updated the global active value, we need to apply the exponential
2299 * weights adjusted to the number of cycles missed.
2301 static void calc_global_nohz(void)
2303 long delta, active, n;
2305 if (!time_before(jiffies, calc_load_update + 10)) {
2307 * Catch-up, fold however many we are behind still
2309 delta = jiffies - calc_load_update - 10;
2310 n = 1 + (delta / LOAD_FREQ);
2312 active = atomic_long_read(&calc_load_tasks);
2313 active = active > 0 ? active * FIXED_1 : 0;
2315 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2316 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2317 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2319 calc_load_update += n * LOAD_FREQ;
2323 * Flip the idle index...
2325 * Make sure we first write the new time then flip the index, so that
2326 * calc_load_write_idx() will see the new time when it reads the new
2327 * index, this avoids a double flip messing things up.
2332 #else /* !CONFIG_NO_HZ */
2334 static inline long calc_load_fold_idle(void) { return 0; }
2335 static inline void calc_global_nohz(void) { }
2337 #endif /* CONFIG_NO_HZ */
2340 * calc_load - update the avenrun load estimates 10 ticks after the
2341 * CPUs have updated calc_load_tasks.
2343 void calc_global_load(unsigned long ticks)
2347 if (time_before(jiffies, calc_load_update + 10))
2351 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2353 delta = calc_load_fold_idle();
2355 atomic_long_add(delta, &calc_load_tasks);
2357 active = atomic_long_read(&calc_load_tasks);
2358 active = active > 0 ? active * FIXED_1 : 0;
2360 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2361 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2362 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2364 calc_load_update += LOAD_FREQ;
2367 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2373 * Called from update_cpu_load() to periodically update this CPU's
2376 static void calc_load_account_active(struct rq *this_rq)
2380 if (time_before(jiffies, this_rq->calc_load_update))
2383 delta = calc_load_fold_active(this_rq);
2385 atomic_long_add(delta, &calc_load_tasks);
2387 this_rq->calc_load_update += LOAD_FREQ;
2391 * End of global load-average stuff
2395 * The exact cpuload at various idx values, calculated at every tick would be
2396 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2398 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2399 * on nth tick when cpu may be busy, then we have:
2400 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2401 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2403 * decay_load_missed() below does efficient calculation of
2404 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2405 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2407 * The calculation is approximated on a 128 point scale.
2408 * degrade_zero_ticks is the number of ticks after which load at any
2409 * particular idx is approximated to be zero.
2410 * degrade_factor is a precomputed table, a row for each load idx.
2411 * Each column corresponds to degradation factor for a power of two ticks,
2412 * based on 128 point scale.
2414 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2415 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2417 * With this power of 2 load factors, we can degrade the load n times
2418 * by looking at 1 bits in n and doing as many mult/shift instead of
2419 * n mult/shifts needed by the exact degradation.
2421 #define DEGRADE_SHIFT 7
2422 static const unsigned char
2423 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2424 static const unsigned char
2425 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2426 {0, 0, 0, 0, 0, 0, 0, 0},
2427 {64, 32, 8, 0, 0, 0, 0, 0},
2428 {96, 72, 40, 12, 1, 0, 0},
2429 {112, 98, 75, 43, 15, 1, 0},
2430 {120, 112, 98, 76, 45, 16, 2} };
2433 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2434 * would be when CPU is idle and so we just decay the old load without
2435 * adding any new load.
2437 static unsigned long
2438 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2442 if (!missed_updates)
2445 if (missed_updates >= degrade_zero_ticks[idx])
2449 return load >> missed_updates;
2451 while (missed_updates) {
2452 if (missed_updates % 2)
2453 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2455 missed_updates >>= 1;
2462 * Update rq->cpu_load[] statistics. This function is usually called every
2463 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2464 * every tick. We fix it up based on jiffies.
2466 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2467 unsigned long pending_updates)
2471 this_rq->nr_load_updates++;
2473 /* Update our load: */
2474 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2475 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2476 unsigned long old_load, new_load;
2478 /* scale is effectively 1 << i now, and >> i divides by scale */
2480 old_load = this_rq->cpu_load[i];
2481 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2482 new_load = this_load;
2484 * Round up the averaging division if load is increasing. This
2485 * prevents us from getting stuck on 9 if the load is 10, for
2488 if (new_load > old_load)
2489 new_load += scale - 1;
2491 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2494 sched_avg_update(this_rq);
2499 * There is no sane way to deal with nohz on smp when using jiffies because the
2500 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2501 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2503 * Therefore we cannot use the delta approach from the regular tick since that
2504 * would seriously skew the load calculation. However we'll make do for those
2505 * updates happening while idle (nohz_idle_balance) or coming out of idle
2506 * (tick_nohz_idle_exit).
2508 * This means we might still be one tick off for nohz periods.
2512 * Called from nohz_idle_balance() to update the load ratings before doing the
2515 void update_idle_cpu_load(struct rq *this_rq)
2517 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2518 unsigned long load = this_rq->load.weight;
2519 unsigned long pending_updates;
2522 * bail if there's load or we're actually up-to-date.
2524 if (load || curr_jiffies == this_rq->last_load_update_tick)
2527 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2528 this_rq->last_load_update_tick = curr_jiffies;
2530 __update_cpu_load(this_rq, load, pending_updates);
2534 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2536 void update_cpu_load_nohz(void)
2538 struct rq *this_rq = this_rq();
2539 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2540 unsigned long pending_updates;
2542 if (curr_jiffies == this_rq->last_load_update_tick)
2545 raw_spin_lock(&this_rq->lock);
2546 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2547 if (pending_updates) {
2548 this_rq->last_load_update_tick = curr_jiffies;
2550 * We were idle, this means load 0, the current load might be
2551 * !0 due to remote wakeups and the sort.
2553 __update_cpu_load(this_rq, 0, pending_updates);
2555 raw_spin_unlock(&this_rq->lock);
2557 #endif /* CONFIG_NO_HZ */
2560 * Called from scheduler_tick()
2562 static void update_cpu_load_active(struct rq *this_rq)
2565 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2567 this_rq->last_load_update_tick = jiffies;
2568 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2570 calc_load_account_active(this_rq);
2576 * sched_exec - execve() is a valuable balancing opportunity, because at
2577 * this point the task has the smallest effective memory and cache footprint.
2579 void sched_exec(void)
2581 struct task_struct *p = current;
2582 unsigned long flags;
2585 raw_spin_lock_irqsave(&p->pi_lock, flags);
2586 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2587 if (dest_cpu == smp_processor_id())
2590 if (likely(cpu_active(dest_cpu))) {
2591 struct migration_arg arg = { p, dest_cpu };
2593 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2594 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2598 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2603 DEFINE_PER_CPU(struct kernel_stat, kstat);
2604 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2606 EXPORT_PER_CPU_SYMBOL(kstat);
2607 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2610 * Return any ns on the sched_clock that have not yet been accounted in
2611 * @p in case that task is currently running.
2613 * Called with task_rq_lock() held on @rq.
2615 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2619 if (task_current(rq, p)) {
2620 update_rq_clock(rq);
2621 ns = rq->clock_task - p->se.exec_start;
2629 unsigned long long task_delta_exec(struct task_struct *p)
2631 unsigned long flags;
2635 rq = task_rq_lock(p, &flags);
2636 ns = do_task_delta_exec(p, rq);
2637 task_rq_unlock(rq, p, &flags);
2643 * Return accounted runtime for the task.
2644 * In case the task is currently running, return the runtime plus current's
2645 * pending runtime that have not been accounted yet.
2647 unsigned long long task_sched_runtime(struct task_struct *p)
2649 unsigned long flags;
2653 rq = task_rq_lock(p, &flags);
2654 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2655 task_rq_unlock(rq, p, &flags);
2661 * This function gets called by the timer code, with HZ frequency.
2662 * We call it with interrupts disabled.
2664 void scheduler_tick(void)
2666 int cpu = smp_processor_id();
2667 struct rq *rq = cpu_rq(cpu);
2668 struct task_struct *curr = rq->curr;
2672 raw_spin_lock(&rq->lock);
2673 update_rq_clock(rq);
2674 update_cpu_load_active(rq);
2675 curr->sched_class->task_tick(rq, curr, 0);
2676 raw_spin_unlock(&rq->lock);
2678 perf_event_task_tick();
2681 rq->idle_balance = idle_cpu(cpu);
2682 trigger_load_balance(rq, cpu);
2686 notrace unsigned long get_parent_ip(unsigned long addr)
2688 if (in_lock_functions(addr)) {
2689 addr = CALLER_ADDR2;
2690 if (in_lock_functions(addr))
2691 addr = CALLER_ADDR3;
2696 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2697 defined(CONFIG_PREEMPT_TRACER))
2699 void __kprobes add_preempt_count(int val)
2701 #ifdef CONFIG_DEBUG_PREEMPT
2705 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2708 preempt_count() += val;
2709 #ifdef CONFIG_DEBUG_PREEMPT
2711 * Spinlock count overflowing soon?
2713 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2716 if (preempt_count() == val)
2717 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2719 EXPORT_SYMBOL(add_preempt_count);
2721 void __kprobes sub_preempt_count(int val)
2723 #ifdef CONFIG_DEBUG_PREEMPT
2727 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2730 * Is the spinlock portion underflowing?
2732 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2733 !(preempt_count() & PREEMPT_MASK)))
2737 if (preempt_count() == val)
2738 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2739 preempt_count() -= val;
2741 EXPORT_SYMBOL(sub_preempt_count);
2746 * Print scheduling while atomic bug:
2748 static noinline void __schedule_bug(struct task_struct *prev)
2750 if (oops_in_progress)
2753 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2754 prev->comm, prev->pid, preempt_count());
2756 debug_show_held_locks(prev);
2758 if (irqs_disabled())
2759 print_irqtrace_events(prev);
2761 add_taint(TAINT_WARN);
2765 * Various schedule()-time debugging checks and statistics:
2767 static inline void schedule_debug(struct task_struct *prev)
2770 * Test if we are atomic. Since do_exit() needs to call into
2771 * schedule() atomically, we ignore that path for now.
2772 * Otherwise, whine if we are scheduling when we should not be.
2774 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2775 __schedule_bug(prev);
2778 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2780 schedstat_inc(this_rq(), sched_count);
2783 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2785 if (prev->on_rq || rq->skip_clock_update < 0)
2786 update_rq_clock(rq);
2787 prev->sched_class->put_prev_task(rq, prev);
2791 * Pick up the highest-prio task:
2793 static inline struct task_struct *
2794 pick_next_task(struct rq *rq)
2796 const struct sched_class *class;
2797 struct task_struct *p;
2800 * Optimization: we know that if all tasks are in
2801 * the fair class we can call that function directly:
2803 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2804 p = fair_sched_class.pick_next_task(rq);
2809 for_each_class(class) {
2810 p = class->pick_next_task(rq);
2815 BUG(); /* the idle class will always have a runnable task */
2819 * __schedule() is the main scheduler function.
2821 * The main means of driving the scheduler and thus entering this function are:
2823 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2825 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2826 * paths. For example, see arch/x86/entry_64.S.
2828 * To drive preemption between tasks, the scheduler sets the flag in timer
2829 * interrupt handler scheduler_tick().
2831 * 3. Wakeups don't really cause entry into schedule(). They add a
2832 * task to the run-queue and that's it.
2834 * Now, if the new task added to the run-queue preempts the current
2835 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2836 * called on the nearest possible occasion:
2838 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2840 * - in syscall or exception context, at the next outmost
2841 * preempt_enable(). (this might be as soon as the wake_up()'s
2844 * - in IRQ context, return from interrupt-handler to
2845 * preemptible context
2847 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2850 * - cond_resched() call
2851 * - explicit schedule() call
2852 * - return from syscall or exception to user-space
2853 * - return from interrupt-handler to user-space
2855 static void __sched __schedule(void)
2857 struct task_struct *prev, *next;
2858 unsigned long *switch_count;
2864 cpu = smp_processor_id();
2866 rcu_note_context_switch(cpu);
2869 schedule_debug(prev);
2871 if (sched_feat(HRTICK))
2874 raw_spin_lock_irq(&rq->lock);
2876 switch_count = &prev->nivcsw;
2877 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2878 if (unlikely(signal_pending_state(prev->state, prev))) {
2879 prev->state = TASK_RUNNING;
2881 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2885 * If a worker went to sleep, notify and ask workqueue
2886 * whether it wants to wake up a task to maintain
2889 if (prev->flags & PF_WQ_WORKER) {
2890 struct task_struct *to_wakeup;
2892 to_wakeup = wq_worker_sleeping(prev, cpu);
2894 try_to_wake_up_local(to_wakeup);
2897 switch_count = &prev->nvcsw;
2900 pre_schedule(rq, prev);
2902 if (unlikely(!rq->nr_running))
2903 idle_balance(cpu, rq);
2905 put_prev_task(rq, prev);
2906 next = pick_next_task(rq);
2907 clear_tsk_need_resched(prev);
2908 rq->skip_clock_update = 0;
2910 if (likely(prev != next)) {
2915 context_switch(rq, prev, next); /* unlocks the rq */
2917 * The context switch have flipped the stack from under us
2918 * and restored the local variables which were saved when
2919 * this task called schedule() in the past. prev == current
2920 * is still correct, but it can be moved to another cpu/rq.
2922 cpu = smp_processor_id();
2925 raw_spin_unlock_irq(&rq->lock);
2929 sched_preempt_enable_no_resched();
2934 static inline void sched_submit_work(struct task_struct *tsk)
2936 if (!tsk->state || tsk_is_pi_blocked(tsk))
2939 * If we are going to sleep and we have plugged IO queued,
2940 * make sure to submit it to avoid deadlocks.
2942 if (blk_needs_flush_plug(tsk))
2943 blk_schedule_flush_plug(tsk);
2946 asmlinkage void __sched schedule(void)
2948 struct task_struct *tsk = current;
2950 sched_submit_work(tsk);
2953 EXPORT_SYMBOL(schedule);
2955 #ifdef CONFIG_RCU_USER_QS
2956 asmlinkage void __sched schedule_user(void)
2959 * If we come here after a random call to set_need_resched(),
2960 * or we have been woken up remotely but the IPI has not yet arrived,
2961 * we haven't yet exited the RCU idle mode. Do it here manually until
2962 * we find a better solution.
2971 * schedule_preempt_disabled - called with preemption disabled
2973 * Returns with preemption disabled. Note: preempt_count must be 1
2975 void __sched schedule_preempt_disabled(void)
2977 sched_preempt_enable_no_resched();
2982 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2984 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2986 if (lock->owner != owner)
2990 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2991 * lock->owner still matches owner, if that fails, owner might
2992 * point to free()d memory, if it still matches, the rcu_read_lock()
2993 * ensures the memory stays valid.
2997 return owner->on_cpu;
3001 * Look out! "owner" is an entirely speculative pointer
3002 * access and not reliable.
3004 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
3006 if (!sched_feat(OWNER_SPIN))
3010 while (owner_running(lock, owner)) {
3014 arch_mutex_cpu_relax();
3019 * We break out the loop above on need_resched() and when the
3020 * owner changed, which is a sign for heavy contention. Return
3021 * success only when lock->owner is NULL.
3023 return lock->owner == NULL;
3027 #ifdef CONFIG_PREEMPT
3029 * this is the entry point to schedule() from in-kernel preemption
3030 * off of preempt_enable. Kernel preemptions off return from interrupt
3031 * occur there and call schedule directly.
3033 asmlinkage void __sched notrace preempt_schedule(void)
3035 struct thread_info *ti = current_thread_info();
3038 * If there is a non-zero preempt_count or interrupts are disabled,
3039 * we do not want to preempt the current task. Just return..
3041 if (likely(ti->preempt_count || irqs_disabled()))
3045 add_preempt_count_notrace(PREEMPT_ACTIVE);
3047 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3050 * Check again in case we missed a preemption opportunity
3051 * between schedule and now.
3054 } while (need_resched());
3056 EXPORT_SYMBOL(preempt_schedule);
3059 * this is the entry point to schedule() from kernel preemption
3060 * off of irq context.
3061 * Note, that this is called and return with irqs disabled. This will
3062 * protect us against recursive calling from irq.
3064 asmlinkage void __sched preempt_schedule_irq(void)
3066 struct thread_info *ti = current_thread_info();
3068 /* Catch callers which need to be fixed */
3069 BUG_ON(ti->preempt_count || !irqs_disabled());
3073 add_preempt_count(PREEMPT_ACTIVE);
3076 local_irq_disable();
3077 sub_preempt_count(PREEMPT_ACTIVE);
3080 * Check again in case we missed a preemption opportunity
3081 * between schedule and now.
3084 } while (need_resched());
3087 #endif /* CONFIG_PREEMPT */
3089 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3092 return try_to_wake_up(curr->private, mode, wake_flags);
3094 EXPORT_SYMBOL(default_wake_function);
3097 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3098 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3099 * number) then we wake all the non-exclusive tasks and one exclusive task.
3101 * There are circumstances in which we can try to wake a task which has already
3102 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3103 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3105 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3106 int nr_exclusive, int wake_flags, void *key)
3108 wait_queue_t *curr, *next;
3110 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3111 unsigned flags = curr->flags;
3113 if (curr->func(curr, mode, wake_flags, key) &&
3114 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3120 * __wake_up - wake up threads blocked on a waitqueue.
3122 * @mode: which threads
3123 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3124 * @key: is directly passed to the wakeup function
3126 * It may be assumed that this function implies a write memory barrier before
3127 * changing the task state if and only if any tasks are woken up.
3129 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3130 int nr_exclusive, void *key)
3132 unsigned long flags;
3134 spin_lock_irqsave(&q->lock, flags);
3135 __wake_up_common(q, mode, nr_exclusive, 0, key);
3136 spin_unlock_irqrestore(&q->lock, flags);
3138 EXPORT_SYMBOL(__wake_up);
3141 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3143 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3145 __wake_up_common(q, mode, nr, 0, NULL);
3147 EXPORT_SYMBOL_GPL(__wake_up_locked);
3149 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3151 __wake_up_common(q, mode, 1, 0, key);
3153 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3156 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3158 * @mode: which threads
3159 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3160 * @key: opaque value to be passed to wakeup targets
3162 * The sync wakeup differs that the waker knows that it will schedule
3163 * away soon, so while the target thread will be woken up, it will not
3164 * be migrated to another CPU - ie. the two threads are 'synchronized'
3165 * with each other. This can prevent needless bouncing between CPUs.
3167 * On UP it can prevent extra preemption.
3169 * It may be assumed that this function implies a write memory barrier before
3170 * changing the task state if and only if any tasks are woken up.
3172 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3173 int nr_exclusive, void *key)
3175 unsigned long flags;
3176 int wake_flags = WF_SYNC;
3181 if (unlikely(!nr_exclusive))
3184 spin_lock_irqsave(&q->lock, flags);
3185 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3186 spin_unlock_irqrestore(&q->lock, flags);
3188 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3191 * __wake_up_sync - see __wake_up_sync_key()
3193 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3195 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3197 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3200 * complete: - signals a single thread waiting on this completion
3201 * @x: holds the state of this particular completion
3203 * This will wake up a single thread waiting on this completion. Threads will be
3204 * awakened in the same order in which they were queued.
3206 * See also complete_all(), wait_for_completion() and related routines.
3208 * It may be assumed that this function implies a write memory barrier before
3209 * changing the task state if and only if any tasks are woken up.
3211 void complete(struct completion *x)
3213 unsigned long flags;
3215 spin_lock_irqsave(&x->wait.lock, flags);
3217 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3218 spin_unlock_irqrestore(&x->wait.lock, flags);
3220 EXPORT_SYMBOL(complete);
3223 * complete_all: - signals all threads waiting on this completion
3224 * @x: holds the state of this particular completion
3226 * This will wake up all threads waiting on this particular completion event.
3228 * It may be assumed that this function implies a write memory barrier before
3229 * changing the task state if and only if any tasks are woken up.
3231 void complete_all(struct completion *x)
3233 unsigned long flags;
3235 spin_lock_irqsave(&x->wait.lock, flags);
3236 x->done += UINT_MAX/2;
3237 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3238 spin_unlock_irqrestore(&x->wait.lock, flags);
3240 EXPORT_SYMBOL(complete_all);
3242 static inline long __sched
3243 do_wait_for_common(struct completion *x, long timeout, int state)
3246 DECLARE_WAITQUEUE(wait, current);
3248 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3250 if (signal_pending_state(state, current)) {
3251 timeout = -ERESTARTSYS;
3254 __set_current_state(state);
3255 spin_unlock_irq(&x->wait.lock);
3256 timeout = schedule_timeout(timeout);
3257 spin_lock_irq(&x->wait.lock);
3258 } while (!x->done && timeout);
3259 __remove_wait_queue(&x->wait, &wait);
3264 return timeout ?: 1;
3268 wait_for_common(struct completion *x, long timeout, int state)
3272 spin_lock_irq(&x->wait.lock);
3273 timeout = do_wait_for_common(x, timeout, state);
3274 spin_unlock_irq(&x->wait.lock);
3279 * wait_for_completion: - waits for completion of a task
3280 * @x: holds the state of this particular completion
3282 * This waits to be signaled for completion of a specific task. It is NOT
3283 * interruptible and there is no timeout.
3285 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3286 * and interrupt capability. Also see complete().
3288 void __sched wait_for_completion(struct completion *x)
3290 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3292 EXPORT_SYMBOL(wait_for_completion);
3295 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3296 * @x: holds the state of this particular completion
3297 * @timeout: timeout value in jiffies
3299 * This waits for either a completion of a specific task to be signaled or for a
3300 * specified timeout to expire. The timeout is in jiffies. It is not
3303 * The return value is 0 if timed out, and positive (at least 1, or number of
3304 * jiffies left till timeout) if completed.
3306 unsigned long __sched
3307 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3309 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3311 EXPORT_SYMBOL(wait_for_completion_timeout);
3314 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3315 * @x: holds the state of this particular completion
3317 * This waits for completion of a specific task to be signaled. It is
3320 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3322 int __sched wait_for_completion_interruptible(struct completion *x)
3324 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3325 if (t == -ERESTARTSYS)
3329 EXPORT_SYMBOL(wait_for_completion_interruptible);
3332 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3333 * @x: holds the state of this particular completion
3334 * @timeout: timeout value in jiffies
3336 * This waits for either a completion of a specific task to be signaled or for a
3337 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3339 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3340 * positive (at least 1, or number of jiffies left till timeout) if completed.
3343 wait_for_completion_interruptible_timeout(struct completion *x,
3344 unsigned long timeout)
3346 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3348 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3351 * wait_for_completion_killable: - waits for completion of a task (killable)
3352 * @x: holds the state of this particular completion
3354 * This waits to be signaled for completion of a specific task. It can be
3355 * interrupted by a kill signal.
3357 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3359 int __sched wait_for_completion_killable(struct completion *x)
3361 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3362 if (t == -ERESTARTSYS)
3366 EXPORT_SYMBOL(wait_for_completion_killable);
3369 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3370 * @x: holds the state of this particular completion
3371 * @timeout: timeout value in jiffies
3373 * This waits for either a completion of a specific task to be
3374 * signaled or for a specified timeout to expire. It can be
3375 * interrupted by a kill signal. The timeout is in jiffies.
3377 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3378 * positive (at least 1, or number of jiffies left till timeout) if completed.
3381 wait_for_completion_killable_timeout(struct completion *x,
3382 unsigned long timeout)
3384 return wait_for_common(x, timeout, TASK_KILLABLE);
3386 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3389 * try_wait_for_completion - try to decrement a completion without blocking
3390 * @x: completion structure
3392 * Returns: 0 if a decrement cannot be done without blocking
3393 * 1 if a decrement succeeded.
3395 * If a completion is being used as a counting completion,
3396 * attempt to decrement the counter without blocking. This
3397 * enables us to avoid waiting if the resource the completion
3398 * is protecting is not available.
3400 bool try_wait_for_completion(struct completion *x)
3402 unsigned long flags;
3405 spin_lock_irqsave(&x->wait.lock, flags);
3410 spin_unlock_irqrestore(&x->wait.lock, flags);
3413 EXPORT_SYMBOL(try_wait_for_completion);
3416 * completion_done - Test to see if a completion has any waiters
3417 * @x: completion structure
3419 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3420 * 1 if there are no waiters.
3423 bool completion_done(struct completion *x)
3425 unsigned long flags;
3428 spin_lock_irqsave(&x->wait.lock, flags);
3431 spin_unlock_irqrestore(&x->wait.lock, flags);
3434 EXPORT_SYMBOL(completion_done);
3437 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3439 unsigned long flags;
3442 init_waitqueue_entry(&wait, current);
3444 __set_current_state(state);
3446 spin_lock_irqsave(&q->lock, flags);
3447 __add_wait_queue(q, &wait);
3448 spin_unlock(&q->lock);
3449 timeout = schedule_timeout(timeout);
3450 spin_lock_irq(&q->lock);
3451 __remove_wait_queue(q, &wait);
3452 spin_unlock_irqrestore(&q->lock, flags);
3457 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3459 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3461 EXPORT_SYMBOL(interruptible_sleep_on);
3464 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3466 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3468 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3470 void __sched sleep_on(wait_queue_head_t *q)
3472 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3474 EXPORT_SYMBOL(sleep_on);
3476 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3478 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3480 EXPORT_SYMBOL(sleep_on_timeout);
3482 #ifdef CONFIG_RT_MUTEXES
3485 * rt_mutex_setprio - set the current priority of a task
3487 * @prio: prio value (kernel-internal form)
3489 * This function changes the 'effective' priority of a task. It does
3490 * not touch ->normal_prio like __setscheduler().
3492 * Used by the rt_mutex code to implement priority inheritance logic.
3494 void rt_mutex_setprio(struct task_struct *p, int prio)
3496 int oldprio, on_rq, running;
3498 const struct sched_class *prev_class;
3500 BUG_ON(prio < 0 || prio > MAX_PRIO);
3502 rq = __task_rq_lock(p);
3505 * Idle task boosting is a nono in general. There is one
3506 * exception, when PREEMPT_RT and NOHZ is active:
3508 * The idle task calls get_next_timer_interrupt() and holds
3509 * the timer wheel base->lock on the CPU and another CPU wants
3510 * to access the timer (probably to cancel it). We can safely
3511 * ignore the boosting request, as the idle CPU runs this code
3512 * with interrupts disabled and will complete the lock
3513 * protected section without being interrupted. So there is no
3514 * real need to boost.
3516 if (unlikely(p == rq->idle)) {
3517 WARN_ON(p != rq->curr);
3518 WARN_ON(p->pi_blocked_on);
3522 trace_sched_pi_setprio(p, prio);
3524 prev_class = p->sched_class;
3526 running = task_current(rq, p);
3528 dequeue_task(rq, p, 0);
3530 p->sched_class->put_prev_task(rq, p);
3533 p->sched_class = &rt_sched_class;
3535 p->sched_class = &fair_sched_class;
3540 p->sched_class->set_curr_task(rq);
3542 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3544 check_class_changed(rq, p, prev_class, oldprio);
3546 __task_rq_unlock(rq);
3549 void set_user_nice(struct task_struct *p, long nice)
3551 int old_prio, delta, on_rq;
3552 unsigned long flags;
3555 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3558 * We have to be careful, if called from sys_setpriority(),
3559 * the task might be in the middle of scheduling on another CPU.
3561 rq = task_rq_lock(p, &flags);
3563 * The RT priorities are set via sched_setscheduler(), but we still
3564 * allow the 'normal' nice value to be set - but as expected
3565 * it wont have any effect on scheduling until the task is
3566 * SCHED_FIFO/SCHED_RR:
3568 if (task_has_rt_policy(p)) {
3569 p->static_prio = NICE_TO_PRIO(nice);
3574 dequeue_task(rq, p, 0);
3576 p->static_prio = NICE_TO_PRIO(nice);
3579 p->prio = effective_prio(p);
3580 delta = p->prio - old_prio;
3583 enqueue_task(rq, p, 0);
3585 * If the task increased its priority or is running and
3586 * lowered its priority, then reschedule its CPU:
3588 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3589 resched_task(rq->curr);
3592 task_rq_unlock(rq, p, &flags);
3594 EXPORT_SYMBOL(set_user_nice);
3597 * can_nice - check if a task can reduce its nice value
3601 int can_nice(const struct task_struct *p, const int nice)
3603 /* convert nice value [19,-20] to rlimit style value [1,40] */
3604 int nice_rlim = 20 - nice;
3606 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3607 capable(CAP_SYS_NICE));
3610 #ifdef __ARCH_WANT_SYS_NICE
3613 * sys_nice - change the priority of the current process.
3614 * @increment: priority increment
3616 * sys_setpriority is a more generic, but much slower function that
3617 * does similar things.
3619 SYSCALL_DEFINE1(nice, int, increment)
3624 * Setpriority might change our priority at the same moment.
3625 * We don't have to worry. Conceptually one call occurs first
3626 * and we have a single winner.
3628 if (increment < -40)
3633 nice = TASK_NICE(current) + increment;
3639 if (increment < 0 && !can_nice(current, nice))
3642 retval = security_task_setnice(current, nice);
3646 set_user_nice(current, nice);
3653 * task_prio - return the priority value of a given task.
3654 * @p: the task in question.
3656 * This is the priority value as seen by users in /proc.
3657 * RT tasks are offset by -200. Normal tasks are centered
3658 * around 0, value goes from -16 to +15.
3660 int task_prio(const struct task_struct *p)
3662 return p->prio - MAX_RT_PRIO;
3666 * task_nice - return the nice value of a given task.
3667 * @p: the task in question.
3669 int task_nice(const struct task_struct *p)
3671 return TASK_NICE(p);
3673 EXPORT_SYMBOL(task_nice);
3676 * idle_cpu - is a given cpu idle currently?
3677 * @cpu: the processor in question.
3679 int idle_cpu(int cpu)
3681 struct rq *rq = cpu_rq(cpu);
3683 if (rq->curr != rq->idle)
3690 if (!llist_empty(&rq->wake_list))
3698 * idle_task - return the idle task for a given cpu.
3699 * @cpu: the processor in question.
3701 struct task_struct *idle_task(int cpu)
3703 return cpu_rq(cpu)->idle;
3707 * find_process_by_pid - find a process with a matching PID value.
3708 * @pid: the pid in question.
3710 static struct task_struct *find_process_by_pid(pid_t pid)
3712 return pid ? find_task_by_vpid(pid) : current;
3715 /* Actually do priority change: must hold rq lock. */
3717 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3720 p->rt_priority = prio;
3721 p->normal_prio = normal_prio(p);
3722 /* we are holding p->pi_lock already */
3723 p->prio = rt_mutex_getprio(p);
3724 if (rt_prio(p->prio))
3725 p->sched_class = &rt_sched_class;
3727 p->sched_class = &fair_sched_class;
3732 * check the target process has a UID that matches the current process's
3734 static bool check_same_owner(struct task_struct *p)
3736 const struct cred *cred = current_cred(), *pcred;
3740 pcred = __task_cred(p);
3741 match = (uid_eq(cred->euid, pcred->euid) ||
3742 uid_eq(cred->euid, pcred->uid));
3747 static int __sched_setscheduler(struct task_struct *p, int policy,
3748 const struct sched_param *param, bool user)
3750 int retval, oldprio, oldpolicy = -1, on_rq, running;
3751 unsigned long flags;
3752 const struct sched_class *prev_class;
3756 /* may grab non-irq protected spin_locks */
3757 BUG_ON(in_interrupt());
3759 /* double check policy once rq lock held */
3761 reset_on_fork = p->sched_reset_on_fork;
3762 policy = oldpolicy = p->policy;
3764 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3765 policy &= ~SCHED_RESET_ON_FORK;
3767 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3768 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3769 policy != SCHED_IDLE)
3774 * Valid priorities for SCHED_FIFO and SCHED_RR are
3775 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3776 * SCHED_BATCH and SCHED_IDLE is 0.
3778 if (param->sched_priority < 0 ||
3779 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3780 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3782 if (rt_policy(policy) != (param->sched_priority != 0))
3786 * Allow unprivileged RT tasks to decrease priority:
3788 if (user && !capable(CAP_SYS_NICE)) {
3789 if (rt_policy(policy)) {
3790 unsigned long rlim_rtprio =
3791 task_rlimit(p, RLIMIT_RTPRIO);
3793 /* can't set/change the rt policy */
3794 if (policy != p->policy && !rlim_rtprio)
3797 /* can't increase priority */
3798 if (param->sched_priority > p->rt_priority &&
3799 param->sched_priority > rlim_rtprio)
3804 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3805 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3807 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3808 if (!can_nice(p, TASK_NICE(p)))
3812 /* can't change other user's priorities */
3813 if (!check_same_owner(p))
3816 /* Normal users shall not reset the sched_reset_on_fork flag */
3817 if (p->sched_reset_on_fork && !reset_on_fork)
3822 retval = security_task_setscheduler(p);
3828 * make sure no PI-waiters arrive (or leave) while we are
3829 * changing the priority of the task:
3831 * To be able to change p->policy safely, the appropriate
3832 * runqueue lock must be held.
3834 rq = task_rq_lock(p, &flags);
3837 * Changing the policy of the stop threads its a very bad idea
3839 if (p == rq->stop) {
3840 task_rq_unlock(rq, p, &flags);
3845 * If not changing anything there's no need to proceed further:
3847 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3848 param->sched_priority == p->rt_priority))) {
3849 task_rq_unlock(rq, p, &flags);
3853 #ifdef CONFIG_RT_GROUP_SCHED
3856 * Do not allow realtime tasks into groups that have no runtime
3859 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3860 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3861 !task_group_is_autogroup(task_group(p))) {
3862 task_rq_unlock(rq, p, &flags);
3868 /* recheck policy now with rq lock held */
3869 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3870 policy = oldpolicy = -1;
3871 task_rq_unlock(rq, p, &flags);
3875 running = task_current(rq, p);
3877 dequeue_task(rq, p, 0);
3879 p->sched_class->put_prev_task(rq, p);
3881 p->sched_reset_on_fork = reset_on_fork;
3884 prev_class = p->sched_class;
3885 __setscheduler(rq, p, policy, param->sched_priority);
3888 p->sched_class->set_curr_task(rq);
3890 enqueue_task(rq, p, 0);
3892 check_class_changed(rq, p, prev_class, oldprio);
3893 task_rq_unlock(rq, p, &flags);
3895 rt_mutex_adjust_pi(p);
3901 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3902 * @p: the task in question.
3903 * @policy: new policy.
3904 * @param: structure containing the new RT priority.
3906 * NOTE that the task may be already dead.
3908 int sched_setscheduler(struct task_struct *p, int policy,
3909 const struct sched_param *param)
3911 return __sched_setscheduler(p, policy, param, true);
3913 EXPORT_SYMBOL_GPL(sched_setscheduler);
3916 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3917 * @p: the task in question.
3918 * @policy: new policy.
3919 * @param: structure containing the new RT priority.
3921 * Just like sched_setscheduler, only don't bother checking if the
3922 * current context has permission. For example, this is needed in
3923 * stop_machine(): we create temporary high priority worker threads,
3924 * but our caller might not have that capability.
3926 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3927 const struct sched_param *param)
3929 return __sched_setscheduler(p, policy, param, false);
3933 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3935 struct sched_param lparam;
3936 struct task_struct *p;
3939 if (!param || pid < 0)
3941 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3946 p = find_process_by_pid(pid);
3948 retval = sched_setscheduler(p, policy, &lparam);
3955 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3956 * @pid: the pid in question.
3957 * @policy: new policy.
3958 * @param: structure containing the new RT priority.
3960 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3961 struct sched_param __user *, param)
3963 /* negative values for policy are not valid */
3967 return do_sched_setscheduler(pid, policy, param);
3971 * sys_sched_setparam - set/change the RT priority of a thread
3972 * @pid: the pid in question.
3973 * @param: structure containing the new RT priority.
3975 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3977 return do_sched_setscheduler(pid, -1, param);
3981 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3982 * @pid: the pid in question.
3984 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3986 struct task_struct *p;
3994 p = find_process_by_pid(pid);
3996 retval = security_task_getscheduler(p);
3999 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4006 * sys_sched_getparam - get the RT priority of a thread
4007 * @pid: the pid in question.
4008 * @param: structure containing the RT priority.
4010 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4012 struct sched_param lp;
4013 struct task_struct *p;
4016 if (!param || pid < 0)
4020 p = find_process_by_pid(pid);
4025 retval = security_task_getscheduler(p);
4029 lp.sched_priority = p->rt_priority;
4033 * This one might sleep, we cannot do it with a spinlock held ...
4035 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4044 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4046 cpumask_var_t cpus_allowed, new_mask;
4047 struct task_struct *p;
4053 p = find_process_by_pid(pid);
4060 /* Prevent p going away */
4064 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4068 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4070 goto out_free_cpus_allowed;
4073 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4076 retval = security_task_setscheduler(p);
4080 cpuset_cpus_allowed(p, cpus_allowed);
4081 cpumask_and(new_mask, in_mask, cpus_allowed);
4083 retval = set_cpus_allowed_ptr(p, new_mask);
4086 cpuset_cpus_allowed(p, cpus_allowed);
4087 if (!cpumask_subset(new_mask, cpus_allowed)) {
4089 * We must have raced with a concurrent cpuset
4090 * update. Just reset the cpus_allowed to the
4091 * cpuset's cpus_allowed
4093 cpumask_copy(new_mask, cpus_allowed);
4098 free_cpumask_var(new_mask);
4099 out_free_cpus_allowed:
4100 free_cpumask_var(cpus_allowed);
4107 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4108 struct cpumask *new_mask)
4110 if (len < cpumask_size())
4111 cpumask_clear(new_mask);
4112 else if (len > cpumask_size())
4113 len = cpumask_size();
4115 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4119 * sys_sched_setaffinity - set the cpu affinity of a process
4120 * @pid: pid of the process
4121 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4122 * @user_mask_ptr: user-space pointer to the new cpu mask
4124 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4125 unsigned long __user *, user_mask_ptr)
4127 cpumask_var_t new_mask;
4130 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4133 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4135 retval = sched_setaffinity(pid, new_mask);
4136 free_cpumask_var(new_mask);
4140 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4142 struct task_struct *p;
4143 unsigned long flags;
4150 p = find_process_by_pid(pid);
4154 retval = security_task_getscheduler(p);
4158 raw_spin_lock_irqsave(&p->pi_lock, flags);
4159 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4160 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4170 * sys_sched_getaffinity - get the cpu affinity of a process
4171 * @pid: pid of the process
4172 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4173 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4175 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4176 unsigned long __user *, user_mask_ptr)
4181 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4183 if (len & (sizeof(unsigned long)-1))
4186 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4189 ret = sched_getaffinity(pid, mask);
4191 size_t retlen = min_t(size_t, len, cpumask_size());
4193 if (copy_to_user(user_mask_ptr, mask, retlen))
4198 free_cpumask_var(mask);
4204 * sys_sched_yield - yield the current processor to other threads.
4206 * This function yields the current CPU to other tasks. If there are no
4207 * other threads running on this CPU then this function will return.
4209 SYSCALL_DEFINE0(sched_yield)
4211 struct rq *rq = this_rq_lock();
4213 schedstat_inc(rq, yld_count);
4214 current->sched_class->yield_task(rq);
4217 * Since we are going to call schedule() anyway, there's
4218 * no need to preempt or enable interrupts:
4220 __release(rq->lock);
4221 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4222 do_raw_spin_unlock(&rq->lock);
4223 sched_preempt_enable_no_resched();
4230 static inline int should_resched(void)
4232 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4235 static void __cond_resched(void)
4237 add_preempt_count(PREEMPT_ACTIVE);
4239 sub_preempt_count(PREEMPT_ACTIVE);
4242 int __sched _cond_resched(void)
4244 if (should_resched()) {
4250 EXPORT_SYMBOL(_cond_resched);
4253 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4254 * call schedule, and on return reacquire the lock.
4256 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4257 * operations here to prevent schedule() from being called twice (once via
4258 * spin_unlock(), once by hand).
4260 int __cond_resched_lock(spinlock_t *lock)
4262 int resched = should_resched();
4265 lockdep_assert_held(lock);
4267 if (spin_needbreak(lock) || resched) {
4278 EXPORT_SYMBOL(__cond_resched_lock);
4280 int __sched __cond_resched_softirq(void)
4282 BUG_ON(!in_softirq());
4284 if (should_resched()) {
4292 EXPORT_SYMBOL(__cond_resched_softirq);
4295 * yield - yield the current processor to other threads.
4297 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4299 * The scheduler is at all times free to pick the calling task as the most
4300 * eligible task to run, if removing the yield() call from your code breaks
4301 * it, its already broken.
4303 * Typical broken usage is:
4308 * where one assumes that yield() will let 'the other' process run that will
4309 * make event true. If the current task is a SCHED_FIFO task that will never
4310 * happen. Never use yield() as a progress guarantee!!
4312 * If you want to use yield() to wait for something, use wait_event().
4313 * If you want to use yield() to be 'nice' for others, use cond_resched().
4314 * If you still want to use yield(), do not!
4316 void __sched yield(void)
4318 set_current_state(TASK_RUNNING);
4321 EXPORT_SYMBOL(yield);
4324 * yield_to - yield the current processor to another thread in
4325 * your thread group, or accelerate that thread toward the
4326 * processor it's on.
4328 * @preempt: whether task preemption is allowed or not
4330 * It's the caller's job to ensure that the target task struct
4331 * can't go away on us before we can do any checks.
4333 * Returns true if we indeed boosted the target task.
4335 bool __sched yield_to(struct task_struct *p, bool preempt)
4337 struct task_struct *curr = current;
4338 struct rq *rq, *p_rq;
4339 unsigned long flags;
4342 local_irq_save(flags);
4347 double_rq_lock(rq, p_rq);
4348 while (task_rq(p) != p_rq) {
4349 double_rq_unlock(rq, p_rq);
4353 if (!curr->sched_class->yield_to_task)
4356 if (curr->sched_class != p->sched_class)
4359 if (task_running(p_rq, p) || p->state)
4362 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4364 schedstat_inc(rq, yld_count);
4366 * Make p's CPU reschedule; pick_next_entity takes care of
4369 if (preempt && rq != p_rq)
4370 resched_task(p_rq->curr);
4374 double_rq_unlock(rq, p_rq);
4375 local_irq_restore(flags);
4382 EXPORT_SYMBOL_GPL(yield_to);
4385 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4386 * that process accounting knows that this is a task in IO wait state.
4388 void __sched io_schedule(void)
4390 struct rq *rq = raw_rq();
4392 delayacct_blkio_start();
4393 atomic_inc(&rq->nr_iowait);
4394 blk_flush_plug(current);
4395 current->in_iowait = 1;
4397 current->in_iowait = 0;
4398 atomic_dec(&rq->nr_iowait);
4399 delayacct_blkio_end();
4401 EXPORT_SYMBOL(io_schedule);
4403 long __sched io_schedule_timeout(long timeout)
4405 struct rq *rq = raw_rq();
4408 delayacct_blkio_start();
4409 atomic_inc(&rq->nr_iowait);
4410 blk_flush_plug(current);
4411 current->in_iowait = 1;
4412 ret = schedule_timeout(timeout);
4413 current->in_iowait = 0;
4414 atomic_dec(&rq->nr_iowait);
4415 delayacct_blkio_end();
4420 * sys_sched_get_priority_max - return maximum RT priority.
4421 * @policy: scheduling class.
4423 * this syscall returns the maximum rt_priority that can be used
4424 * by a given scheduling class.
4426 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4433 ret = MAX_USER_RT_PRIO-1;
4445 * sys_sched_get_priority_min - return minimum RT priority.
4446 * @policy: scheduling class.
4448 * this syscall returns the minimum rt_priority that can be used
4449 * by a given scheduling class.
4451 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4469 * sys_sched_rr_get_interval - return the default timeslice of a process.
4470 * @pid: pid of the process.
4471 * @interval: userspace pointer to the timeslice value.
4473 * this syscall writes the default timeslice value of a given process
4474 * into the user-space timespec buffer. A value of '0' means infinity.
4476 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4477 struct timespec __user *, interval)
4479 struct task_struct *p;
4480 unsigned int time_slice;
4481 unsigned long flags;
4491 p = find_process_by_pid(pid);
4495 retval = security_task_getscheduler(p);
4499 rq = task_rq_lock(p, &flags);
4500 time_slice = p->sched_class->get_rr_interval(rq, p);
4501 task_rq_unlock(rq, p, &flags);
4504 jiffies_to_timespec(time_slice, &t);
4505 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4513 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4515 void sched_show_task(struct task_struct *p)
4517 unsigned long free = 0;
4520 state = p->state ? __ffs(p->state) + 1 : 0;
4521 printk(KERN_INFO "%-15.15s %c", p->comm,
4522 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4523 #if BITS_PER_LONG == 32
4524 if (state == TASK_RUNNING)
4525 printk(KERN_CONT " running ");
4527 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4529 if (state == TASK_RUNNING)
4530 printk(KERN_CONT " running task ");
4532 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4534 #ifdef CONFIG_DEBUG_STACK_USAGE
4535 free = stack_not_used(p);
4537 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4538 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)),
4539 (unsigned long)task_thread_info(p)->flags);
4541 show_stack(p, NULL);
4544 void show_state_filter(unsigned long state_filter)
4546 struct task_struct *g, *p;
4548 #if BITS_PER_LONG == 32
4550 " task PC stack pid father\n");
4553 " task PC stack pid father\n");
4556 do_each_thread(g, p) {
4558 * reset the NMI-timeout, listing all files on a slow
4559 * console might take a lot of time:
4561 touch_nmi_watchdog();
4562 if (!state_filter || (p->state & state_filter))
4564 } while_each_thread(g, p);
4566 touch_all_softlockup_watchdogs();
4568 #ifdef CONFIG_SCHED_DEBUG
4569 sysrq_sched_debug_show();
4573 * Only show locks if all tasks are dumped:
4576 debug_show_all_locks();
4579 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4581 idle->sched_class = &idle_sched_class;
4585 * init_idle - set up an idle thread for a given CPU
4586 * @idle: task in question
4587 * @cpu: cpu the idle task belongs to
4589 * NOTE: this function does not set the idle thread's NEED_RESCHED
4590 * flag, to make booting more robust.
4592 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4594 struct rq *rq = cpu_rq(cpu);
4595 unsigned long flags;
4597 raw_spin_lock_irqsave(&rq->lock, flags);
4600 idle->state = TASK_RUNNING;
4601 idle->se.exec_start = sched_clock();
4603 do_set_cpus_allowed(idle, cpumask_of(cpu));
4605 * We're having a chicken and egg problem, even though we are
4606 * holding rq->lock, the cpu isn't yet set to this cpu so the
4607 * lockdep check in task_group() will fail.
4609 * Similar case to sched_fork(). / Alternatively we could
4610 * use task_rq_lock() here and obtain the other rq->lock.
4615 __set_task_cpu(idle, cpu);
4618 rq->curr = rq->idle = idle;
4619 #if defined(CONFIG_SMP)
4622 raw_spin_unlock_irqrestore(&rq->lock, flags);
4624 /* Set the preempt count _outside_ the spinlocks! */
4625 task_thread_info(idle)->preempt_count = 0;
4628 * The idle tasks have their own, simple scheduling class:
4630 idle->sched_class = &idle_sched_class;
4631 ftrace_graph_init_idle_task(idle, cpu);
4632 #if defined(CONFIG_SMP)
4633 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4638 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4640 if (p->sched_class && p->sched_class->set_cpus_allowed)
4641 p->sched_class->set_cpus_allowed(p, new_mask);
4643 cpumask_copy(&p->cpus_allowed, new_mask);
4644 p->nr_cpus_allowed = cpumask_weight(new_mask);
4648 * This is how migration works:
4650 * 1) we invoke migration_cpu_stop() on the target CPU using
4652 * 2) stopper starts to run (implicitly forcing the migrated thread
4654 * 3) it checks whether the migrated task is still in the wrong runqueue.
4655 * 4) if it's in the wrong runqueue then the migration thread removes
4656 * it and puts it into the right queue.
4657 * 5) stopper completes and stop_one_cpu() returns and the migration
4662 * Change a given task's CPU affinity. Migrate the thread to a
4663 * proper CPU and schedule it away if the CPU it's executing on
4664 * is removed from the allowed bitmask.
4666 * NOTE: the caller must have a valid reference to the task, the
4667 * task must not exit() & deallocate itself prematurely. The
4668 * call is not atomic; no spinlocks may be held.
4670 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4672 unsigned long flags;
4674 unsigned int dest_cpu;
4677 rq = task_rq_lock(p, &flags);
4679 if (cpumask_equal(&p->cpus_allowed, new_mask))
4682 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4687 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4692 do_set_cpus_allowed(p, new_mask);
4694 /* Can the task run on the task's current CPU? If so, we're done */
4695 if (cpumask_test_cpu(task_cpu(p), new_mask))
4698 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4700 struct migration_arg arg = { p, dest_cpu };
4701 /* Need help from migration thread: drop lock and wait. */
4702 task_rq_unlock(rq, p, &flags);
4703 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4704 tlb_migrate_finish(p->mm);
4708 task_rq_unlock(rq, p, &flags);
4712 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4715 * Move (not current) task off this cpu, onto dest cpu. We're doing
4716 * this because either it can't run here any more (set_cpus_allowed()
4717 * away from this CPU, or CPU going down), or because we're
4718 * attempting to rebalance this task on exec (sched_exec).
4720 * So we race with normal scheduler movements, but that's OK, as long
4721 * as the task is no longer on this CPU.
4723 * Returns non-zero if task was successfully migrated.
4725 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4727 struct rq *rq_dest, *rq_src;
4730 if (unlikely(!cpu_active(dest_cpu)))
4733 rq_src = cpu_rq(src_cpu);
4734 rq_dest = cpu_rq(dest_cpu);
4736 raw_spin_lock(&p->pi_lock);
4737 double_rq_lock(rq_src, rq_dest);
4738 /* Already moved. */
4739 if (task_cpu(p) != src_cpu)
4741 /* Affinity changed (again). */
4742 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4746 * If we're not on a rq, the next wake-up will ensure we're
4750 dequeue_task(rq_src, p, 0);
4751 set_task_cpu(p, dest_cpu);
4752 enqueue_task(rq_dest, p, 0);
4753 check_preempt_curr(rq_dest, p, 0);
4758 double_rq_unlock(rq_src, rq_dest);
4759 raw_spin_unlock(&p->pi_lock);
4764 * migration_cpu_stop - this will be executed by a highprio stopper thread
4765 * and performs thread migration by bumping thread off CPU then
4766 * 'pushing' onto another runqueue.
4768 static int migration_cpu_stop(void *data)
4770 struct migration_arg *arg = data;
4773 * The original target cpu might have gone down and we might
4774 * be on another cpu but it doesn't matter.
4776 local_irq_disable();
4777 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4782 #ifdef CONFIG_HOTPLUG_CPU
4785 * Ensures that the idle task is using init_mm right before its cpu goes
4788 void idle_task_exit(void)
4790 struct mm_struct *mm = current->active_mm;
4792 BUG_ON(cpu_online(smp_processor_id()));
4795 switch_mm(mm, &init_mm, current);
4800 * Since this CPU is going 'away' for a while, fold any nr_active delta
4801 * we might have. Assumes we're called after migrate_tasks() so that the
4802 * nr_active count is stable.
4804 * Also see the comment "Global load-average calculations".
4806 static void calc_load_migrate(struct rq *rq)
4808 long delta = calc_load_fold_active(rq);
4810 atomic_long_add(delta, &calc_load_tasks);
4814 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4815 * try_to_wake_up()->select_task_rq().
4817 * Called with rq->lock held even though we'er in stop_machine() and
4818 * there's no concurrency possible, we hold the required locks anyway
4819 * because of lock validation efforts.
4821 static void migrate_tasks(unsigned int dead_cpu)
4823 struct rq *rq = cpu_rq(dead_cpu);
4824 struct task_struct *next, *stop = rq->stop;
4828 * Fudge the rq selection such that the below task selection loop
4829 * doesn't get stuck on the currently eligible stop task.
4831 * We're currently inside stop_machine() and the rq is either stuck
4832 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4833 * either way we should never end up calling schedule() until we're
4840 * There's this thread running, bail when that's the only
4843 if (rq->nr_running == 1)
4846 next = pick_next_task(rq);
4848 next->sched_class->put_prev_task(rq, next);
4850 /* Find suitable destination for @next, with force if needed. */
4851 dest_cpu = select_fallback_rq(dead_cpu, next);
4852 raw_spin_unlock(&rq->lock);
4854 __migrate_task(next, dead_cpu, dest_cpu);
4856 raw_spin_lock(&rq->lock);
4862 #endif /* CONFIG_HOTPLUG_CPU */
4864 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4866 static struct ctl_table sd_ctl_dir[] = {
4868 .procname = "sched_domain",
4874 static struct ctl_table sd_ctl_root[] = {
4876 .procname = "kernel",
4878 .child = sd_ctl_dir,
4883 static struct ctl_table *sd_alloc_ctl_entry(int n)
4885 struct ctl_table *entry =
4886 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4891 static void sd_free_ctl_entry(struct ctl_table **tablep)
4893 struct ctl_table *entry;
4896 * In the intermediate directories, both the child directory and
4897 * procname are dynamically allocated and could fail but the mode
4898 * will always be set. In the lowest directory the names are
4899 * static strings and all have proc handlers.
4901 for (entry = *tablep; entry->mode; entry++) {
4903 sd_free_ctl_entry(&entry->child);
4904 if (entry->proc_handler == NULL)
4905 kfree(entry->procname);
4912 static int min_load_idx = 0;
4913 static int max_load_idx = CPU_LOAD_IDX_MAX;
4916 set_table_entry(struct ctl_table *entry,
4917 const char *procname, void *data, int maxlen,
4918 umode_t mode, proc_handler *proc_handler,
4921 entry->procname = procname;
4923 entry->maxlen = maxlen;
4925 entry->proc_handler = proc_handler;
4928 entry->extra1 = &min_load_idx;
4929 entry->extra2 = &max_load_idx;
4933 static struct ctl_table *
4934 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4936 struct ctl_table *table = sd_alloc_ctl_entry(13);
4941 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4942 sizeof(long), 0644, proc_doulongvec_minmax, false);
4943 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4944 sizeof(long), 0644, proc_doulongvec_minmax, false);
4945 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4946 sizeof(int), 0644, proc_dointvec_minmax, true);
4947 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4948 sizeof(int), 0644, proc_dointvec_minmax, true);
4949 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4950 sizeof(int), 0644, proc_dointvec_minmax, true);
4951 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4952 sizeof(int), 0644, proc_dointvec_minmax, true);
4953 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4954 sizeof(int), 0644, proc_dointvec_minmax, true);
4955 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4956 sizeof(int), 0644, proc_dointvec_minmax, false);
4957 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4958 sizeof(int), 0644, proc_dointvec_minmax, false);
4959 set_table_entry(&table[9], "cache_nice_tries",
4960 &sd->cache_nice_tries,
4961 sizeof(int), 0644, proc_dointvec_minmax, false);
4962 set_table_entry(&table[10], "flags", &sd->flags,
4963 sizeof(int), 0644, proc_dointvec_minmax, false);
4964 set_table_entry(&table[11], "name", sd->name,
4965 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4966 /* &table[12] is terminator */
4971 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4973 struct ctl_table *entry, *table;
4974 struct sched_domain *sd;
4975 int domain_num = 0, i;
4978 for_each_domain(cpu, sd)
4980 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4985 for_each_domain(cpu, sd) {
4986 snprintf(buf, 32, "domain%d", i);
4987 entry->procname = kstrdup(buf, GFP_KERNEL);
4989 entry->child = sd_alloc_ctl_domain_table(sd);
4996 static struct ctl_table_header *sd_sysctl_header;
4997 static void register_sched_domain_sysctl(void)
4999 int i, cpu_num = num_possible_cpus();
5000 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5003 WARN_ON(sd_ctl_dir[0].child);
5004 sd_ctl_dir[0].child = entry;
5009 for_each_possible_cpu(i) {
5010 snprintf(buf, 32, "cpu%d", i);
5011 entry->procname = kstrdup(buf, GFP_KERNEL);
5013 entry->child = sd_alloc_ctl_cpu_table(i);
5017 WARN_ON(sd_sysctl_header);
5018 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5021 /* may be called multiple times per register */
5022 static void unregister_sched_domain_sysctl(void)
5024 if (sd_sysctl_header)
5025 unregister_sysctl_table(sd_sysctl_header);
5026 sd_sysctl_header = NULL;
5027 if (sd_ctl_dir[0].child)
5028 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5031 static void register_sched_domain_sysctl(void)
5034 static void unregister_sched_domain_sysctl(void)
5039 static void set_rq_online(struct rq *rq)
5042 const struct sched_class *class;
5044 cpumask_set_cpu(rq->cpu, rq->rd->online);
5047 for_each_class(class) {
5048 if (class->rq_online)
5049 class->rq_online(rq);
5054 static void set_rq_offline(struct rq *rq)
5057 const struct sched_class *class;
5059 for_each_class(class) {
5060 if (class->rq_offline)
5061 class->rq_offline(rq);
5064 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5070 * migration_call - callback that gets triggered when a CPU is added.
5071 * Here we can start up the necessary migration thread for the new CPU.
5073 static int __cpuinit
5074 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5076 int cpu = (long)hcpu;
5077 unsigned long flags;
5078 struct rq *rq = cpu_rq(cpu);
5080 switch (action & ~CPU_TASKS_FROZEN) {
5082 case CPU_UP_PREPARE:
5083 rq->calc_load_update = calc_load_update;
5087 /* Update our root-domain */
5088 raw_spin_lock_irqsave(&rq->lock, flags);
5090 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5094 raw_spin_unlock_irqrestore(&rq->lock, flags);
5097 #ifdef CONFIG_HOTPLUG_CPU
5099 sched_ttwu_pending();
5100 /* Update our root-domain */
5101 raw_spin_lock_irqsave(&rq->lock, flags);
5103 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5107 BUG_ON(rq->nr_running != 1); /* the migration thread */
5108 raw_spin_unlock_irqrestore(&rq->lock, flags);
5112 calc_load_migrate(rq);
5117 update_max_interval();
5123 * Register at high priority so that task migration (migrate_all_tasks)
5124 * happens before everything else. This has to be lower priority than
5125 * the notifier in the perf_event subsystem, though.
5127 static struct notifier_block __cpuinitdata migration_notifier = {
5128 .notifier_call = migration_call,
5129 .priority = CPU_PRI_MIGRATION,
5132 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5133 unsigned long action, void *hcpu)
5135 switch (action & ~CPU_TASKS_FROZEN) {
5137 case CPU_DOWN_FAILED:
5138 set_cpu_active((long)hcpu, true);
5145 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5146 unsigned long action, void *hcpu)
5148 switch (action & ~CPU_TASKS_FROZEN) {
5149 case CPU_DOWN_PREPARE:
5150 set_cpu_active((long)hcpu, false);
5157 static int __init migration_init(void)
5159 void *cpu = (void *)(long)smp_processor_id();
5162 /* Initialize migration for the boot CPU */
5163 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5164 BUG_ON(err == NOTIFY_BAD);
5165 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5166 register_cpu_notifier(&migration_notifier);
5168 /* Register cpu active notifiers */
5169 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5170 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5174 early_initcall(migration_init);
5179 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5181 #ifdef CONFIG_SCHED_DEBUG
5183 static __read_mostly int sched_debug_enabled;
5185 static int __init sched_debug_setup(char *str)
5187 sched_debug_enabled = 1;
5191 early_param("sched_debug", sched_debug_setup);
5193 static inline bool sched_debug(void)
5195 return sched_debug_enabled;
5198 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5199 struct cpumask *groupmask)
5201 struct sched_group *group = sd->groups;
5204 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5205 cpumask_clear(groupmask);
5207 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5209 if (!(sd->flags & SD_LOAD_BALANCE)) {
5210 printk("does not load-balance\n");
5212 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5217 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5219 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5220 printk(KERN_ERR "ERROR: domain->span does not contain "
5223 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5224 printk(KERN_ERR "ERROR: domain->groups does not contain"
5228 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5232 printk(KERN_ERR "ERROR: group is NULL\n");
5237 * Even though we initialize ->power to something semi-sane,
5238 * we leave power_orig unset. This allows us to detect if
5239 * domain iteration is still funny without causing /0 traps.
5241 if (!group->sgp->power_orig) {
5242 printk(KERN_CONT "\n");
5243 printk(KERN_ERR "ERROR: domain->cpu_power not "
5248 if (!cpumask_weight(sched_group_cpus(group))) {
5249 printk(KERN_CONT "\n");
5250 printk(KERN_ERR "ERROR: empty group\n");
5254 if (!(sd->flags & SD_OVERLAP) &&
5255 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5256 printk(KERN_CONT "\n");
5257 printk(KERN_ERR "ERROR: repeated CPUs\n");
5261 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5263 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5265 printk(KERN_CONT " %s", str);
5266 if (group->sgp->power != SCHED_POWER_SCALE) {
5267 printk(KERN_CONT " (cpu_power = %d)",
5271 group = group->next;
5272 } while (group != sd->groups);
5273 printk(KERN_CONT "\n");
5275 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5276 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5279 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5280 printk(KERN_ERR "ERROR: parent span is not a superset "
5281 "of domain->span\n");
5285 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5289 if (!sched_debug_enabled)
5293 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5297 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5300 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5308 #else /* !CONFIG_SCHED_DEBUG */
5309 # define sched_domain_debug(sd, cpu) do { } while (0)
5310 static inline bool sched_debug(void)
5314 #endif /* CONFIG_SCHED_DEBUG */
5316 static int sd_degenerate(struct sched_domain *sd)
5318 if (cpumask_weight(sched_domain_span(sd)) == 1)
5321 /* Following flags need at least 2 groups */
5322 if (sd->flags & (SD_LOAD_BALANCE |
5323 SD_BALANCE_NEWIDLE |
5327 SD_SHARE_PKG_RESOURCES)) {
5328 if (sd->groups != sd->groups->next)
5332 /* Following flags don't use groups */
5333 if (sd->flags & (SD_WAKE_AFFINE))
5340 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5342 unsigned long cflags = sd->flags, pflags = parent->flags;
5344 if (sd_degenerate(parent))
5347 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5350 /* Flags needing groups don't count if only 1 group in parent */
5351 if (parent->groups == parent->groups->next) {
5352 pflags &= ~(SD_LOAD_BALANCE |
5353 SD_BALANCE_NEWIDLE |
5357 SD_SHARE_PKG_RESOURCES);
5358 if (nr_node_ids == 1)
5359 pflags &= ~SD_SERIALIZE;
5361 if (~cflags & pflags)
5367 static void free_rootdomain(struct rcu_head *rcu)
5369 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5371 cpupri_cleanup(&rd->cpupri);
5372 free_cpumask_var(rd->rto_mask);
5373 free_cpumask_var(rd->online);
5374 free_cpumask_var(rd->span);
5378 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5380 struct root_domain *old_rd = NULL;
5381 unsigned long flags;
5383 raw_spin_lock_irqsave(&rq->lock, flags);
5388 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5391 cpumask_clear_cpu(rq->cpu, old_rd->span);
5394 * If we dont want to free the old_rt yet then
5395 * set old_rd to NULL to skip the freeing later
5398 if (!atomic_dec_and_test(&old_rd->refcount))
5402 atomic_inc(&rd->refcount);
5405 cpumask_set_cpu(rq->cpu, rd->span);
5406 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5409 raw_spin_unlock_irqrestore(&rq->lock, flags);
5412 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5415 static int init_rootdomain(struct root_domain *rd)
5417 memset(rd, 0, sizeof(*rd));
5419 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5421 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5423 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5426 if (cpupri_init(&rd->cpupri) != 0)
5431 free_cpumask_var(rd->rto_mask);
5433 free_cpumask_var(rd->online);
5435 free_cpumask_var(rd->span);
5441 * By default the system creates a single root-domain with all cpus as
5442 * members (mimicking the global state we have today).
5444 struct root_domain def_root_domain;
5446 static void init_defrootdomain(void)
5448 init_rootdomain(&def_root_domain);
5450 atomic_set(&def_root_domain.refcount, 1);
5453 static struct root_domain *alloc_rootdomain(void)
5455 struct root_domain *rd;
5457 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5461 if (init_rootdomain(rd) != 0) {
5469 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5471 struct sched_group *tmp, *first;
5480 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5485 } while (sg != first);
5488 static void free_sched_domain(struct rcu_head *rcu)
5490 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5493 * If its an overlapping domain it has private groups, iterate and
5496 if (sd->flags & SD_OVERLAP) {
5497 free_sched_groups(sd->groups, 1);
5498 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5499 kfree(sd->groups->sgp);
5505 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5507 call_rcu(&sd->rcu, free_sched_domain);
5510 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5512 for (; sd; sd = sd->parent)
5513 destroy_sched_domain(sd, cpu);
5517 * Keep a special pointer to the highest sched_domain that has
5518 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5519 * allows us to avoid some pointer chasing select_idle_sibling().
5521 * Also keep a unique ID per domain (we use the first cpu number in
5522 * the cpumask of the domain), this allows us to quickly tell if
5523 * two cpus are in the same cache domain, see cpus_share_cache().
5525 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5526 DEFINE_PER_CPU(int, sd_llc_id);
5528 static void update_top_cache_domain(int cpu)
5530 struct sched_domain *sd;
5533 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5535 id = cpumask_first(sched_domain_span(sd));
5537 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5538 per_cpu(sd_llc_id, cpu) = id;
5542 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5543 * hold the hotplug lock.
5546 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5548 struct rq *rq = cpu_rq(cpu);
5549 struct sched_domain *tmp;
5551 /* Remove the sched domains which do not contribute to scheduling. */
5552 for (tmp = sd; tmp; ) {
5553 struct sched_domain *parent = tmp->parent;
5557 if (sd_parent_degenerate(tmp, parent)) {
5558 tmp->parent = parent->parent;
5560 parent->parent->child = tmp;
5561 destroy_sched_domain(parent, cpu);
5566 if (sd && sd_degenerate(sd)) {
5569 destroy_sched_domain(tmp, cpu);
5574 sched_domain_debug(sd, cpu);
5576 rq_attach_root(rq, rd);
5578 rcu_assign_pointer(rq->sd, sd);
5579 destroy_sched_domains(tmp, cpu);
5581 update_top_cache_domain(cpu);
5584 /* cpus with isolated domains */
5585 static cpumask_var_t cpu_isolated_map;
5587 /* Setup the mask of cpus configured for isolated domains */
5588 static int __init isolated_cpu_setup(char *str)
5590 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5591 cpulist_parse(str, cpu_isolated_map);
5595 __setup("isolcpus=", isolated_cpu_setup);
5597 static const struct cpumask *cpu_cpu_mask(int cpu)
5599 return cpumask_of_node(cpu_to_node(cpu));
5603 struct sched_domain **__percpu sd;
5604 struct sched_group **__percpu sg;
5605 struct sched_group_power **__percpu sgp;
5609 struct sched_domain ** __percpu sd;
5610 struct root_domain *rd;
5620 struct sched_domain_topology_level;
5622 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5623 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5625 #define SDTL_OVERLAP 0x01
5627 struct sched_domain_topology_level {
5628 sched_domain_init_f init;
5629 sched_domain_mask_f mask;
5632 struct sd_data data;
5636 * Build an iteration mask that can exclude certain CPUs from the upwards
5639 * Asymmetric node setups can result in situations where the domain tree is of
5640 * unequal depth, make sure to skip domains that already cover the entire
5643 * In that case build_sched_domains() will have terminated the iteration early
5644 * and our sibling sd spans will be empty. Domains should always include the
5645 * cpu they're built on, so check that.
5648 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5650 const struct cpumask *span = sched_domain_span(sd);
5651 struct sd_data *sdd = sd->private;
5652 struct sched_domain *sibling;
5655 for_each_cpu(i, span) {
5656 sibling = *per_cpu_ptr(sdd->sd, i);
5657 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5660 cpumask_set_cpu(i, sched_group_mask(sg));
5665 * Return the canonical balance cpu for this group, this is the first cpu
5666 * of this group that's also in the iteration mask.
5668 int group_balance_cpu(struct sched_group *sg)
5670 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5674 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5676 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5677 const struct cpumask *span = sched_domain_span(sd);
5678 struct cpumask *covered = sched_domains_tmpmask;
5679 struct sd_data *sdd = sd->private;
5680 struct sched_domain *child;
5683 cpumask_clear(covered);
5685 for_each_cpu(i, span) {
5686 struct cpumask *sg_span;
5688 if (cpumask_test_cpu(i, covered))
5691 child = *per_cpu_ptr(sdd->sd, i);
5693 /* See the comment near build_group_mask(). */
5694 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5697 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5698 GFP_KERNEL, cpu_to_node(cpu));
5703 sg_span = sched_group_cpus(sg);
5705 child = child->child;
5706 cpumask_copy(sg_span, sched_domain_span(child));
5708 cpumask_set_cpu(i, sg_span);
5710 cpumask_or(covered, covered, sg_span);
5712 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5713 if (atomic_inc_return(&sg->sgp->ref) == 1)
5714 build_group_mask(sd, sg);
5717 * Initialize sgp->power such that even if we mess up the
5718 * domains and no possible iteration will get us here, we won't
5721 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5724 * Make sure the first group of this domain contains the
5725 * canonical balance cpu. Otherwise the sched_domain iteration
5726 * breaks. See update_sg_lb_stats().
5728 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5729 group_balance_cpu(sg) == cpu)
5739 sd->groups = groups;
5744 free_sched_groups(first, 0);
5749 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5751 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5752 struct sched_domain *child = sd->child;
5755 cpu = cpumask_first(sched_domain_span(child));
5758 *sg = *per_cpu_ptr(sdd->sg, cpu);
5759 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5760 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5767 * build_sched_groups will build a circular linked list of the groups
5768 * covered by the given span, and will set each group's ->cpumask correctly,
5769 * and ->cpu_power to 0.
5771 * Assumes the sched_domain tree is fully constructed
5774 build_sched_groups(struct sched_domain *sd, int cpu)
5776 struct sched_group *first = NULL, *last = NULL;
5777 struct sd_data *sdd = sd->private;
5778 const struct cpumask *span = sched_domain_span(sd);
5779 struct cpumask *covered;
5782 get_group(cpu, sdd, &sd->groups);
5783 atomic_inc(&sd->groups->ref);
5785 if (cpu != cpumask_first(sched_domain_span(sd)))
5788 lockdep_assert_held(&sched_domains_mutex);
5789 covered = sched_domains_tmpmask;
5791 cpumask_clear(covered);
5793 for_each_cpu(i, span) {
5794 struct sched_group *sg;
5795 int group = get_group(i, sdd, &sg);
5798 if (cpumask_test_cpu(i, covered))
5801 cpumask_clear(sched_group_cpus(sg));
5803 cpumask_setall(sched_group_mask(sg));
5805 for_each_cpu(j, span) {
5806 if (get_group(j, sdd, NULL) != group)
5809 cpumask_set_cpu(j, covered);
5810 cpumask_set_cpu(j, sched_group_cpus(sg));
5825 * Initialize sched groups cpu_power.
5827 * cpu_power indicates the capacity of sched group, which is used while
5828 * distributing the load between different sched groups in a sched domain.
5829 * Typically cpu_power for all the groups in a sched domain will be same unless
5830 * there are asymmetries in the topology. If there are asymmetries, group
5831 * having more cpu_power will pickup more load compared to the group having
5834 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5836 struct sched_group *sg = sd->groups;
5838 WARN_ON(!sd || !sg);
5841 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5843 } while (sg != sd->groups);
5845 if (cpu != group_balance_cpu(sg))
5848 update_group_power(sd, cpu);
5849 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5852 int __weak arch_sd_sibling_asym_packing(void)
5854 return 0*SD_ASYM_PACKING;
5858 * Initializers for schedule domains
5859 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5862 #ifdef CONFIG_SCHED_DEBUG
5863 # define SD_INIT_NAME(sd, type) sd->name = #type
5865 # define SD_INIT_NAME(sd, type) do { } while (0)
5868 #define SD_INIT_FUNC(type) \
5869 static noinline struct sched_domain * \
5870 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5872 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5873 *sd = SD_##type##_INIT; \
5874 SD_INIT_NAME(sd, type); \
5875 sd->private = &tl->data; \
5880 #ifdef CONFIG_SCHED_SMT
5881 SD_INIT_FUNC(SIBLING)
5883 #ifdef CONFIG_SCHED_MC
5886 #ifdef CONFIG_SCHED_BOOK
5890 static int default_relax_domain_level = -1;
5891 int sched_domain_level_max;
5893 static int __init setup_relax_domain_level(char *str)
5895 if (kstrtoint(str, 0, &default_relax_domain_level))
5896 pr_warn("Unable to set relax_domain_level\n");
5900 __setup("relax_domain_level=", setup_relax_domain_level);
5902 static void set_domain_attribute(struct sched_domain *sd,
5903 struct sched_domain_attr *attr)
5907 if (!attr || attr->relax_domain_level < 0) {
5908 if (default_relax_domain_level < 0)
5911 request = default_relax_domain_level;
5913 request = attr->relax_domain_level;
5914 if (request < sd->level) {
5915 /* turn off idle balance on this domain */
5916 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5918 /* turn on idle balance on this domain */
5919 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5923 static void __sdt_free(const struct cpumask *cpu_map);
5924 static int __sdt_alloc(const struct cpumask *cpu_map);
5926 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5927 const struct cpumask *cpu_map)
5931 if (!atomic_read(&d->rd->refcount))
5932 free_rootdomain(&d->rd->rcu); /* fall through */
5934 free_percpu(d->sd); /* fall through */
5936 __sdt_free(cpu_map); /* fall through */
5942 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5943 const struct cpumask *cpu_map)
5945 memset(d, 0, sizeof(*d));
5947 if (__sdt_alloc(cpu_map))
5948 return sa_sd_storage;
5949 d->sd = alloc_percpu(struct sched_domain *);
5951 return sa_sd_storage;
5952 d->rd = alloc_rootdomain();
5955 return sa_rootdomain;
5959 * NULL the sd_data elements we've used to build the sched_domain and
5960 * sched_group structure so that the subsequent __free_domain_allocs()
5961 * will not free the data we're using.
5963 static void claim_allocations(int cpu, struct sched_domain *sd)
5965 struct sd_data *sdd = sd->private;
5967 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5968 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5970 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5971 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5973 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5974 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5977 #ifdef CONFIG_SCHED_SMT
5978 static const struct cpumask *cpu_smt_mask(int cpu)
5980 return topology_thread_cpumask(cpu);
5985 * Topology list, bottom-up.
5987 static struct sched_domain_topology_level default_topology[] = {
5988 #ifdef CONFIG_SCHED_SMT
5989 { sd_init_SIBLING, cpu_smt_mask, },
5991 #ifdef CONFIG_SCHED_MC
5992 { sd_init_MC, cpu_coregroup_mask, },
5994 #ifdef CONFIG_SCHED_BOOK
5995 { sd_init_BOOK, cpu_book_mask, },
5997 { sd_init_CPU, cpu_cpu_mask, },
6001 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6005 static int sched_domains_numa_levels;
6006 static int *sched_domains_numa_distance;
6007 static struct cpumask ***sched_domains_numa_masks;
6008 static int sched_domains_curr_level;
6010 static inline int sd_local_flags(int level)
6012 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6015 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6018 static struct sched_domain *
6019 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6021 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6022 int level = tl->numa_level;
6023 int sd_weight = cpumask_weight(
6024 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6026 *sd = (struct sched_domain){
6027 .min_interval = sd_weight,
6028 .max_interval = 2*sd_weight,
6030 .imbalance_pct = 125,
6031 .cache_nice_tries = 2,
6038 .flags = 1*SD_LOAD_BALANCE
6039 | 1*SD_BALANCE_NEWIDLE
6044 | 0*SD_SHARE_CPUPOWER
6045 | 0*SD_SHARE_PKG_RESOURCES
6047 | 0*SD_PREFER_SIBLING
6048 | sd_local_flags(level)
6050 .last_balance = jiffies,
6051 .balance_interval = sd_weight,
6053 SD_INIT_NAME(sd, NUMA);
6054 sd->private = &tl->data;
6057 * Ugly hack to pass state to sd_numa_mask()...
6059 sched_domains_curr_level = tl->numa_level;
6064 static const struct cpumask *sd_numa_mask(int cpu)
6066 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6069 static void sched_numa_warn(const char *str)
6071 static int done = false;
6079 printk(KERN_WARNING "ERROR: %s\n\n", str);
6081 for (i = 0; i < nr_node_ids; i++) {
6082 printk(KERN_WARNING " ");
6083 for (j = 0; j < nr_node_ids; j++)
6084 printk(KERN_CONT "%02d ", node_distance(i,j));
6085 printk(KERN_CONT "\n");
6087 printk(KERN_WARNING "\n");
6090 static bool find_numa_distance(int distance)
6094 if (distance == node_distance(0, 0))
6097 for (i = 0; i < sched_domains_numa_levels; i++) {
6098 if (sched_domains_numa_distance[i] == distance)
6105 static void sched_init_numa(void)
6107 int next_distance, curr_distance = node_distance(0, 0);
6108 struct sched_domain_topology_level *tl;
6112 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6113 if (!sched_domains_numa_distance)
6117 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6118 * unique distances in the node_distance() table.
6120 * Assumes node_distance(0,j) includes all distances in
6121 * node_distance(i,j) in order to avoid cubic time.
6123 next_distance = curr_distance;
6124 for (i = 0; i < nr_node_ids; i++) {
6125 for (j = 0; j < nr_node_ids; j++) {
6126 for (k = 0; k < nr_node_ids; k++) {
6127 int distance = node_distance(i, k);
6129 if (distance > curr_distance &&
6130 (distance < next_distance ||
6131 next_distance == curr_distance))
6132 next_distance = distance;
6135 * While not a strong assumption it would be nice to know
6136 * about cases where if node A is connected to B, B is not
6137 * equally connected to A.
6139 if (sched_debug() && node_distance(k, i) != distance)
6140 sched_numa_warn("Node-distance not symmetric");
6142 if (sched_debug() && i && !find_numa_distance(distance))
6143 sched_numa_warn("Node-0 not representative");
6145 if (next_distance != curr_distance) {
6146 sched_domains_numa_distance[level++] = next_distance;
6147 sched_domains_numa_levels = level;
6148 curr_distance = next_distance;
6153 * In case of sched_debug() we verify the above assumption.
6159 * 'level' contains the number of unique distances, excluding the
6160 * identity distance node_distance(i,i).
6162 * The sched_domains_nume_distance[] array includes the actual distance
6167 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6168 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6169 * the array will contain less then 'level' members. This could be
6170 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6171 * in other functions.
6173 * We reset it to 'level' at the end of this function.
6175 sched_domains_numa_levels = 0;
6177 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6178 if (!sched_domains_numa_masks)
6182 * Now for each level, construct a mask per node which contains all
6183 * cpus of nodes that are that many hops away from us.
6185 for (i = 0; i < level; i++) {
6186 sched_domains_numa_masks[i] =
6187 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6188 if (!sched_domains_numa_masks[i])
6191 for (j = 0; j < nr_node_ids; j++) {
6192 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6196 sched_domains_numa_masks[i][j] = mask;
6198 for (k = 0; k < nr_node_ids; k++) {
6199 if (node_distance(j, k) > sched_domains_numa_distance[i])
6202 cpumask_or(mask, mask, cpumask_of_node(k));
6207 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6208 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6213 * Copy the default topology bits..
6215 for (i = 0; default_topology[i].init; i++)
6216 tl[i] = default_topology[i];
6219 * .. and append 'j' levels of NUMA goodness.
6221 for (j = 0; j < level; i++, j++) {
6222 tl[i] = (struct sched_domain_topology_level){
6223 .init = sd_numa_init,
6224 .mask = sd_numa_mask,
6225 .flags = SDTL_OVERLAP,
6230 sched_domain_topology = tl;
6232 sched_domains_numa_levels = level;
6235 static void sched_domains_numa_masks_set(int cpu)
6238 int node = cpu_to_node(cpu);
6240 for (i = 0; i < sched_domains_numa_levels; i++) {
6241 for (j = 0; j < nr_node_ids; j++) {
6242 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6243 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6248 static void sched_domains_numa_masks_clear(int cpu)
6251 for (i = 0; i < sched_domains_numa_levels; i++) {
6252 for (j = 0; j < nr_node_ids; j++)
6253 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6258 * Update sched_domains_numa_masks[level][node] array when new cpus
6261 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6262 unsigned long action,
6265 int cpu = (long)hcpu;
6267 switch (action & ~CPU_TASKS_FROZEN) {
6269 sched_domains_numa_masks_set(cpu);
6273 sched_domains_numa_masks_clear(cpu);
6283 static inline void sched_init_numa(void)
6287 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6288 unsigned long action,
6293 #endif /* CONFIG_NUMA */
6295 static int __sdt_alloc(const struct cpumask *cpu_map)
6297 struct sched_domain_topology_level *tl;
6300 for (tl = sched_domain_topology; tl->init; tl++) {
6301 struct sd_data *sdd = &tl->data;
6303 sdd->sd = alloc_percpu(struct sched_domain *);
6307 sdd->sg = alloc_percpu(struct sched_group *);
6311 sdd->sgp = alloc_percpu(struct sched_group_power *);
6315 for_each_cpu(j, cpu_map) {
6316 struct sched_domain *sd;
6317 struct sched_group *sg;
6318 struct sched_group_power *sgp;
6320 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6321 GFP_KERNEL, cpu_to_node(j));
6325 *per_cpu_ptr(sdd->sd, j) = sd;
6327 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6328 GFP_KERNEL, cpu_to_node(j));
6334 *per_cpu_ptr(sdd->sg, j) = sg;
6336 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6337 GFP_KERNEL, cpu_to_node(j));
6341 *per_cpu_ptr(sdd->sgp, j) = sgp;
6348 static void __sdt_free(const struct cpumask *cpu_map)
6350 struct sched_domain_topology_level *tl;
6353 for (tl = sched_domain_topology; tl->init; tl++) {
6354 struct sd_data *sdd = &tl->data;
6356 for_each_cpu(j, cpu_map) {
6357 struct sched_domain *sd;
6360 sd = *per_cpu_ptr(sdd->sd, j);
6361 if (sd && (sd->flags & SD_OVERLAP))
6362 free_sched_groups(sd->groups, 0);
6363 kfree(*per_cpu_ptr(sdd->sd, j));
6367 kfree(*per_cpu_ptr(sdd->sg, j));
6369 kfree(*per_cpu_ptr(sdd->sgp, j));
6371 free_percpu(sdd->sd);
6373 free_percpu(sdd->sg);
6375 free_percpu(sdd->sgp);
6380 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6381 struct s_data *d, const struct cpumask *cpu_map,
6382 struct sched_domain_attr *attr, struct sched_domain *child,
6385 struct sched_domain *sd = tl->init(tl, cpu);
6389 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6391 sd->level = child->level + 1;
6392 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6396 set_domain_attribute(sd, attr);
6402 * Build sched domains for a given set of cpus and attach the sched domains
6403 * to the individual cpus
6405 static int build_sched_domains(const struct cpumask *cpu_map,
6406 struct sched_domain_attr *attr)
6408 enum s_alloc alloc_state = sa_none;
6409 struct sched_domain *sd;
6411 int i, ret = -ENOMEM;
6413 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6414 if (alloc_state != sa_rootdomain)
6417 /* Set up domains for cpus specified by the cpu_map. */
6418 for_each_cpu(i, cpu_map) {
6419 struct sched_domain_topology_level *tl;
6422 for (tl = sched_domain_topology; tl->init; tl++) {
6423 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6424 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6425 sd->flags |= SD_OVERLAP;
6426 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6433 *per_cpu_ptr(d.sd, i) = sd;
6436 /* Build the groups for the domains */
6437 for_each_cpu(i, cpu_map) {
6438 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6439 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6440 if (sd->flags & SD_OVERLAP) {
6441 if (build_overlap_sched_groups(sd, i))
6444 if (build_sched_groups(sd, i))
6450 /* Calculate CPU power for physical packages and nodes */
6451 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6452 if (!cpumask_test_cpu(i, cpu_map))
6455 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6456 claim_allocations(i, sd);
6457 init_sched_groups_power(i, sd);
6461 /* Attach the domains */
6463 for_each_cpu(i, cpu_map) {
6464 sd = *per_cpu_ptr(d.sd, i);
6465 cpu_attach_domain(sd, d.rd, i);
6471 __free_domain_allocs(&d, alloc_state, cpu_map);
6475 static cpumask_var_t *doms_cur; /* current sched domains */
6476 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6477 static struct sched_domain_attr *dattr_cur;
6478 /* attribues of custom domains in 'doms_cur' */
6481 * Special case: If a kmalloc of a doms_cur partition (array of
6482 * cpumask) fails, then fallback to a single sched domain,
6483 * as determined by the single cpumask fallback_doms.
6485 static cpumask_var_t fallback_doms;
6488 * arch_update_cpu_topology lets virtualized architectures update the
6489 * cpu core maps. It is supposed to return 1 if the topology changed
6490 * or 0 if it stayed the same.
6492 int __attribute__((weak)) arch_update_cpu_topology(void)
6497 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6500 cpumask_var_t *doms;
6502 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6505 for (i = 0; i < ndoms; i++) {
6506 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6507 free_sched_domains(doms, i);
6514 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6517 for (i = 0; i < ndoms; i++)
6518 free_cpumask_var(doms[i]);
6523 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6524 * For now this just excludes isolated cpus, but could be used to
6525 * exclude other special cases in the future.
6527 static int init_sched_domains(const struct cpumask *cpu_map)
6531 arch_update_cpu_topology();
6533 doms_cur = alloc_sched_domains(ndoms_cur);
6535 doms_cur = &fallback_doms;
6536 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6537 err = build_sched_domains(doms_cur[0], NULL);
6538 register_sched_domain_sysctl();
6544 * Detach sched domains from a group of cpus specified in cpu_map
6545 * These cpus will now be attached to the NULL domain
6547 static void detach_destroy_domains(const struct cpumask *cpu_map)
6552 for_each_cpu(i, cpu_map)
6553 cpu_attach_domain(NULL, &def_root_domain, i);
6557 /* handle null as "default" */
6558 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6559 struct sched_domain_attr *new, int idx_new)
6561 struct sched_domain_attr tmp;
6568 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6569 new ? (new + idx_new) : &tmp,
6570 sizeof(struct sched_domain_attr));
6574 * Partition sched domains as specified by the 'ndoms_new'
6575 * cpumasks in the array doms_new[] of cpumasks. This compares
6576 * doms_new[] to the current sched domain partitioning, doms_cur[].
6577 * It destroys each deleted domain and builds each new domain.
6579 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6580 * The masks don't intersect (don't overlap.) We should setup one
6581 * sched domain for each mask. CPUs not in any of the cpumasks will
6582 * not be load balanced. If the same cpumask appears both in the
6583 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6586 * The passed in 'doms_new' should be allocated using
6587 * alloc_sched_domains. This routine takes ownership of it and will
6588 * free_sched_domains it when done with it. If the caller failed the
6589 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6590 * and partition_sched_domains() will fallback to the single partition
6591 * 'fallback_doms', it also forces the domains to be rebuilt.
6593 * If doms_new == NULL it will be replaced with cpu_online_mask.
6594 * ndoms_new == 0 is a special case for destroying existing domains,
6595 * and it will not create the default domain.
6597 * Call with hotplug lock held
6599 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6600 struct sched_domain_attr *dattr_new)
6605 mutex_lock(&sched_domains_mutex);
6607 /* always unregister in case we don't destroy any domains */
6608 unregister_sched_domain_sysctl();
6610 /* Let architecture update cpu core mappings. */
6611 new_topology = arch_update_cpu_topology();
6613 n = doms_new ? ndoms_new : 0;
6615 /* Destroy deleted domains */
6616 for (i = 0; i < ndoms_cur; i++) {
6617 for (j = 0; j < n && !new_topology; j++) {
6618 if (cpumask_equal(doms_cur[i], doms_new[j])
6619 && dattrs_equal(dattr_cur, i, dattr_new, j))
6622 /* no match - a current sched domain not in new doms_new[] */
6623 detach_destroy_domains(doms_cur[i]);
6628 if (doms_new == NULL) {
6630 doms_new = &fallback_doms;
6631 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6632 WARN_ON_ONCE(dattr_new);
6635 /* Build new domains */
6636 for (i = 0; i < ndoms_new; i++) {
6637 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6638 if (cpumask_equal(doms_new[i], doms_cur[j])
6639 && dattrs_equal(dattr_new, i, dattr_cur, j))
6642 /* no match - add a new doms_new */
6643 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6648 /* Remember the new sched domains */
6649 if (doms_cur != &fallback_doms)
6650 free_sched_domains(doms_cur, ndoms_cur);
6651 kfree(dattr_cur); /* kfree(NULL) is safe */
6652 doms_cur = doms_new;
6653 dattr_cur = dattr_new;
6654 ndoms_cur = ndoms_new;
6656 register_sched_domain_sysctl();
6658 mutex_unlock(&sched_domains_mutex);
6661 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6664 * Update cpusets according to cpu_active mask. If cpusets are
6665 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6666 * around partition_sched_domains().
6668 * If we come here as part of a suspend/resume, don't touch cpusets because we
6669 * want to restore it back to its original state upon resume anyway.
6671 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6675 case CPU_ONLINE_FROZEN:
6676 case CPU_DOWN_FAILED_FROZEN:
6679 * num_cpus_frozen tracks how many CPUs are involved in suspend
6680 * resume sequence. As long as this is not the last online
6681 * operation in the resume sequence, just build a single sched
6682 * domain, ignoring cpusets.
6685 if (likely(num_cpus_frozen)) {
6686 partition_sched_domains(1, NULL, NULL);
6691 * This is the last CPU online operation. So fall through and
6692 * restore the original sched domains by considering the
6693 * cpuset configurations.
6697 case CPU_DOWN_FAILED:
6698 cpuset_update_active_cpus(true);
6706 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6710 case CPU_DOWN_PREPARE:
6711 cpuset_update_active_cpus(false);
6713 case CPU_DOWN_PREPARE_FROZEN:
6715 partition_sched_domains(1, NULL, NULL);
6723 void __init sched_init_smp(void)
6725 cpumask_var_t non_isolated_cpus;
6727 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6728 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6733 mutex_lock(&sched_domains_mutex);
6734 init_sched_domains(cpu_active_mask);
6735 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6736 if (cpumask_empty(non_isolated_cpus))
6737 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6738 mutex_unlock(&sched_domains_mutex);
6741 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6742 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6743 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6745 /* RT runtime code needs to handle some hotplug events */
6746 hotcpu_notifier(update_runtime, 0);
6750 /* Move init over to a non-isolated CPU */
6751 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6753 sched_init_granularity();
6754 free_cpumask_var(non_isolated_cpus);
6756 init_sched_rt_class();
6759 void __init sched_init_smp(void)
6761 sched_init_granularity();
6763 #endif /* CONFIG_SMP */
6765 const_debug unsigned int sysctl_timer_migration = 1;
6767 int in_sched_functions(unsigned long addr)
6769 return in_lock_functions(addr) ||
6770 (addr >= (unsigned long)__sched_text_start
6771 && addr < (unsigned long)__sched_text_end);
6774 #ifdef CONFIG_CGROUP_SCHED
6775 struct task_group root_task_group;
6776 LIST_HEAD(task_groups);
6779 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6781 void __init sched_init(void)
6784 unsigned long alloc_size = 0, ptr;
6786 #ifdef CONFIG_FAIR_GROUP_SCHED
6787 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6789 #ifdef CONFIG_RT_GROUP_SCHED
6790 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6792 #ifdef CONFIG_CPUMASK_OFFSTACK
6793 alloc_size += num_possible_cpus() * cpumask_size();
6796 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6798 #ifdef CONFIG_FAIR_GROUP_SCHED
6799 root_task_group.se = (struct sched_entity **)ptr;
6800 ptr += nr_cpu_ids * sizeof(void **);
6802 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6803 ptr += nr_cpu_ids * sizeof(void **);
6805 #endif /* CONFIG_FAIR_GROUP_SCHED */
6806 #ifdef CONFIG_RT_GROUP_SCHED
6807 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6808 ptr += nr_cpu_ids * sizeof(void **);
6810 root_task_group.rt_rq = (struct rt_rq **)ptr;
6811 ptr += nr_cpu_ids * sizeof(void **);
6813 #endif /* CONFIG_RT_GROUP_SCHED */
6814 #ifdef CONFIG_CPUMASK_OFFSTACK
6815 for_each_possible_cpu(i) {
6816 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6817 ptr += cpumask_size();
6819 #endif /* CONFIG_CPUMASK_OFFSTACK */
6823 init_defrootdomain();
6826 init_rt_bandwidth(&def_rt_bandwidth,
6827 global_rt_period(), global_rt_runtime());
6829 #ifdef CONFIG_RT_GROUP_SCHED
6830 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6831 global_rt_period(), global_rt_runtime());
6832 #endif /* CONFIG_RT_GROUP_SCHED */
6834 #ifdef CONFIG_CGROUP_SCHED
6835 list_add(&root_task_group.list, &task_groups);
6836 INIT_LIST_HEAD(&root_task_group.children);
6837 INIT_LIST_HEAD(&root_task_group.siblings);
6838 autogroup_init(&init_task);
6840 #endif /* CONFIG_CGROUP_SCHED */
6842 #ifdef CONFIG_CGROUP_CPUACCT
6843 root_cpuacct.cpustat = &kernel_cpustat;
6844 root_cpuacct.cpuusage = alloc_percpu(u64);
6845 /* Too early, not expected to fail */
6846 BUG_ON(!root_cpuacct.cpuusage);
6848 for_each_possible_cpu(i) {
6852 raw_spin_lock_init(&rq->lock);
6854 rq->calc_load_active = 0;
6855 rq->calc_load_update = jiffies + LOAD_FREQ;
6856 init_cfs_rq(&rq->cfs);
6857 init_rt_rq(&rq->rt, rq);
6858 #ifdef CONFIG_FAIR_GROUP_SCHED
6859 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6860 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6862 * How much cpu bandwidth does root_task_group get?
6864 * In case of task-groups formed thr' the cgroup filesystem, it
6865 * gets 100% of the cpu resources in the system. This overall
6866 * system cpu resource is divided among the tasks of
6867 * root_task_group and its child task-groups in a fair manner,
6868 * based on each entity's (task or task-group's) weight
6869 * (se->load.weight).
6871 * In other words, if root_task_group has 10 tasks of weight
6872 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6873 * then A0's share of the cpu resource is:
6875 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6877 * We achieve this by letting root_task_group's tasks sit
6878 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6880 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6881 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6882 #endif /* CONFIG_FAIR_GROUP_SCHED */
6884 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6885 #ifdef CONFIG_RT_GROUP_SCHED
6886 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6887 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6890 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6891 rq->cpu_load[j] = 0;
6893 rq->last_load_update_tick = jiffies;
6898 rq->cpu_power = SCHED_POWER_SCALE;
6899 rq->post_schedule = 0;
6900 rq->active_balance = 0;
6901 rq->next_balance = jiffies;
6906 rq->avg_idle = 2*sysctl_sched_migration_cost;
6908 INIT_LIST_HEAD(&rq->cfs_tasks);
6910 rq_attach_root(rq, &def_root_domain);
6916 atomic_set(&rq->nr_iowait, 0);
6919 set_load_weight(&init_task);
6921 #ifdef CONFIG_PREEMPT_NOTIFIERS
6922 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6925 #ifdef CONFIG_RT_MUTEXES
6926 plist_head_init(&init_task.pi_waiters);
6930 * The boot idle thread does lazy MMU switching as well:
6932 atomic_inc(&init_mm.mm_count);
6933 enter_lazy_tlb(&init_mm, current);
6936 * Make us the idle thread. Technically, schedule() should not be
6937 * called from this thread, however somewhere below it might be,
6938 * but because we are the idle thread, we just pick up running again
6939 * when this runqueue becomes "idle".
6941 init_idle(current, smp_processor_id());
6943 calc_load_update = jiffies + LOAD_FREQ;
6946 * During early bootup we pretend to be a normal task:
6948 current->sched_class = &fair_sched_class;
6951 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6952 /* May be allocated at isolcpus cmdline parse time */
6953 if (cpu_isolated_map == NULL)
6954 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6955 idle_thread_set_boot_cpu();
6957 init_sched_fair_class();
6959 scheduler_running = 1;
6962 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6963 static inline int preempt_count_equals(int preempt_offset)
6965 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6967 return (nested == preempt_offset);
6970 void __might_sleep(const char *file, int line, int preempt_offset)
6972 static unsigned long prev_jiffy; /* ratelimiting */
6974 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6975 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6976 system_state != SYSTEM_RUNNING || oops_in_progress)
6978 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6980 prev_jiffy = jiffies;
6983 "BUG: sleeping function called from invalid context at %s:%d\n",
6986 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6987 in_atomic(), irqs_disabled(),
6988 current->pid, current->comm);
6990 debug_show_held_locks(current);
6991 if (irqs_disabled())
6992 print_irqtrace_events(current);
6995 EXPORT_SYMBOL(__might_sleep);
6998 #ifdef CONFIG_MAGIC_SYSRQ
6999 static void normalize_task(struct rq *rq, struct task_struct *p)
7001 const struct sched_class *prev_class = p->sched_class;
7002 int old_prio = p->prio;
7007 dequeue_task(rq, p, 0);
7008 __setscheduler(rq, p, SCHED_NORMAL, 0);
7010 enqueue_task(rq, p, 0);
7011 resched_task(rq->curr);
7014 check_class_changed(rq, p, prev_class, old_prio);
7017 void normalize_rt_tasks(void)
7019 struct task_struct *g, *p;
7020 unsigned long flags;
7023 read_lock_irqsave(&tasklist_lock, flags);
7024 do_each_thread(g, p) {
7026 * Only normalize user tasks:
7031 p->se.exec_start = 0;
7032 #ifdef CONFIG_SCHEDSTATS
7033 p->se.statistics.wait_start = 0;
7034 p->se.statistics.sleep_start = 0;
7035 p->se.statistics.block_start = 0;
7040 * Renice negative nice level userspace
7043 if (TASK_NICE(p) < 0 && p->mm)
7044 set_user_nice(p, 0);
7048 raw_spin_lock(&p->pi_lock);
7049 rq = __task_rq_lock(p);
7051 normalize_task(rq, p);
7053 __task_rq_unlock(rq);
7054 raw_spin_unlock(&p->pi_lock);
7055 } while_each_thread(g, p);
7057 read_unlock_irqrestore(&tasklist_lock, flags);
7060 #endif /* CONFIG_MAGIC_SYSRQ */
7062 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7064 * These functions are only useful for the IA64 MCA handling, or kdb.
7066 * They can only be called when the whole system has been
7067 * stopped - every CPU needs to be quiescent, and no scheduling
7068 * activity can take place. Using them for anything else would
7069 * be a serious bug, and as a result, they aren't even visible
7070 * under any other configuration.
7074 * curr_task - return the current task for a given cpu.
7075 * @cpu: the processor in question.
7077 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7079 struct task_struct *curr_task(int cpu)
7081 return cpu_curr(cpu);
7084 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7088 * set_curr_task - set the current task for a given cpu.
7089 * @cpu: the processor in question.
7090 * @p: the task pointer to set.
7092 * Description: This function must only be used when non-maskable interrupts
7093 * are serviced on a separate stack. It allows the architecture to switch the
7094 * notion of the current task on a cpu in a non-blocking manner. This function
7095 * must be called with all CPU's synchronized, and interrupts disabled, the
7096 * and caller must save the original value of the current task (see
7097 * curr_task() above) and restore that value before reenabling interrupts and
7098 * re-starting the system.
7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7102 void set_curr_task(int cpu, struct task_struct *p)
7109 #ifdef CONFIG_CGROUP_SCHED
7110 /* task_group_lock serializes the addition/removal of task groups */
7111 static DEFINE_SPINLOCK(task_group_lock);
7113 static void free_sched_group(struct task_group *tg)
7115 free_fair_sched_group(tg);
7116 free_rt_sched_group(tg);
7121 /* allocate runqueue etc for a new task group */
7122 struct task_group *sched_create_group(struct task_group *parent)
7124 struct task_group *tg;
7125 unsigned long flags;
7127 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7129 return ERR_PTR(-ENOMEM);
7131 if (!alloc_fair_sched_group(tg, parent))
7134 if (!alloc_rt_sched_group(tg, parent))
7137 spin_lock_irqsave(&task_group_lock, flags);
7138 list_add_rcu(&tg->list, &task_groups);
7140 WARN_ON(!parent); /* root should already exist */
7142 tg->parent = parent;
7143 INIT_LIST_HEAD(&tg->children);
7144 list_add_rcu(&tg->siblings, &parent->children);
7145 spin_unlock_irqrestore(&task_group_lock, flags);
7150 free_sched_group(tg);
7151 return ERR_PTR(-ENOMEM);
7154 /* rcu callback to free various structures associated with a task group */
7155 static void free_sched_group_rcu(struct rcu_head *rhp)
7157 /* now it should be safe to free those cfs_rqs */
7158 free_sched_group(container_of(rhp, struct task_group, rcu));
7161 /* Destroy runqueue etc associated with a task group */
7162 void sched_destroy_group(struct task_group *tg)
7164 unsigned long flags;
7167 /* end participation in shares distribution */
7168 for_each_possible_cpu(i)
7169 unregister_fair_sched_group(tg, i);
7171 spin_lock_irqsave(&task_group_lock, flags);
7172 list_del_rcu(&tg->list);
7173 list_del_rcu(&tg->siblings);
7174 spin_unlock_irqrestore(&task_group_lock, flags);
7176 /* wait for possible concurrent references to cfs_rqs complete */
7177 call_rcu(&tg->rcu, free_sched_group_rcu);
7180 /* change task's runqueue when it moves between groups.
7181 * The caller of this function should have put the task in its new group
7182 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7183 * reflect its new group.
7185 void sched_move_task(struct task_struct *tsk)
7187 struct task_group *tg;
7189 unsigned long flags;
7192 rq = task_rq_lock(tsk, &flags);
7194 running = task_current(rq, tsk);
7198 dequeue_task(rq, tsk, 0);
7199 if (unlikely(running))
7200 tsk->sched_class->put_prev_task(rq, tsk);
7202 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7203 lockdep_is_held(&tsk->sighand->siglock)),
7204 struct task_group, css);
7205 tg = autogroup_task_group(tsk, tg);
7206 tsk->sched_task_group = tg;
7208 #ifdef CONFIG_FAIR_GROUP_SCHED
7209 if (tsk->sched_class->task_move_group)
7210 tsk->sched_class->task_move_group(tsk, on_rq);
7213 set_task_rq(tsk, task_cpu(tsk));
7215 if (unlikely(running))
7216 tsk->sched_class->set_curr_task(rq);
7218 enqueue_task(rq, tsk, 0);
7220 task_rq_unlock(rq, tsk, &flags);
7222 #endif /* CONFIG_CGROUP_SCHED */
7224 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7225 static unsigned long to_ratio(u64 period, u64 runtime)
7227 if (runtime == RUNTIME_INF)
7230 return div64_u64(runtime << 20, period);
7234 #ifdef CONFIG_RT_GROUP_SCHED
7236 * Ensure that the real time constraints are schedulable.
7238 static DEFINE_MUTEX(rt_constraints_mutex);
7240 /* Must be called with tasklist_lock held */
7241 static inline int tg_has_rt_tasks(struct task_group *tg)
7243 struct task_struct *g, *p;
7245 do_each_thread(g, p) {
7246 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7248 } while_each_thread(g, p);
7253 struct rt_schedulable_data {
7254 struct task_group *tg;
7259 static int tg_rt_schedulable(struct task_group *tg, void *data)
7261 struct rt_schedulable_data *d = data;
7262 struct task_group *child;
7263 unsigned long total, sum = 0;
7264 u64 period, runtime;
7266 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7267 runtime = tg->rt_bandwidth.rt_runtime;
7270 period = d->rt_period;
7271 runtime = d->rt_runtime;
7275 * Cannot have more runtime than the period.
7277 if (runtime > period && runtime != RUNTIME_INF)
7281 * Ensure we don't starve existing RT tasks.
7283 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7286 total = to_ratio(period, runtime);
7289 * Nobody can have more than the global setting allows.
7291 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7295 * The sum of our children's runtime should not exceed our own.
7297 list_for_each_entry_rcu(child, &tg->children, siblings) {
7298 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7299 runtime = child->rt_bandwidth.rt_runtime;
7301 if (child == d->tg) {
7302 period = d->rt_period;
7303 runtime = d->rt_runtime;
7306 sum += to_ratio(period, runtime);
7315 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7319 struct rt_schedulable_data data = {
7321 .rt_period = period,
7322 .rt_runtime = runtime,
7326 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7332 static int tg_set_rt_bandwidth(struct task_group *tg,
7333 u64 rt_period, u64 rt_runtime)
7337 mutex_lock(&rt_constraints_mutex);
7338 read_lock(&tasklist_lock);
7339 err = __rt_schedulable(tg, rt_period, rt_runtime);
7343 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7344 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7345 tg->rt_bandwidth.rt_runtime = rt_runtime;
7347 for_each_possible_cpu(i) {
7348 struct rt_rq *rt_rq = tg->rt_rq[i];
7350 raw_spin_lock(&rt_rq->rt_runtime_lock);
7351 rt_rq->rt_runtime = rt_runtime;
7352 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7354 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7356 read_unlock(&tasklist_lock);
7357 mutex_unlock(&rt_constraints_mutex);
7362 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7364 u64 rt_runtime, rt_period;
7366 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7367 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7368 if (rt_runtime_us < 0)
7369 rt_runtime = RUNTIME_INF;
7371 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7374 long sched_group_rt_runtime(struct task_group *tg)
7378 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7381 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7382 do_div(rt_runtime_us, NSEC_PER_USEC);
7383 return rt_runtime_us;
7386 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7388 u64 rt_runtime, rt_period;
7390 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7391 rt_runtime = tg->rt_bandwidth.rt_runtime;
7396 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7399 long sched_group_rt_period(struct task_group *tg)
7403 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7404 do_div(rt_period_us, NSEC_PER_USEC);
7405 return rt_period_us;
7408 static int sched_rt_global_constraints(void)
7410 u64 runtime, period;
7413 if (sysctl_sched_rt_period <= 0)
7416 runtime = global_rt_runtime();
7417 period = global_rt_period();
7420 * Sanity check on the sysctl variables.
7422 if (runtime > period && runtime != RUNTIME_INF)
7425 mutex_lock(&rt_constraints_mutex);
7426 read_lock(&tasklist_lock);
7427 ret = __rt_schedulable(NULL, 0, 0);
7428 read_unlock(&tasklist_lock);
7429 mutex_unlock(&rt_constraints_mutex);
7434 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7436 /* Don't accept realtime tasks when there is no way for them to run */
7437 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7443 #else /* !CONFIG_RT_GROUP_SCHED */
7444 static int sched_rt_global_constraints(void)
7446 unsigned long flags;
7449 if (sysctl_sched_rt_period <= 0)
7453 * There's always some RT tasks in the root group
7454 * -- migration, kstopmachine etc..
7456 if (sysctl_sched_rt_runtime == 0)
7459 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7460 for_each_possible_cpu(i) {
7461 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7463 raw_spin_lock(&rt_rq->rt_runtime_lock);
7464 rt_rq->rt_runtime = global_rt_runtime();
7465 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7467 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7471 #endif /* CONFIG_RT_GROUP_SCHED */
7473 int sched_rt_handler(struct ctl_table *table, int write,
7474 void __user *buffer, size_t *lenp,
7478 int old_period, old_runtime;
7479 static DEFINE_MUTEX(mutex);
7482 old_period = sysctl_sched_rt_period;
7483 old_runtime = sysctl_sched_rt_runtime;
7485 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7487 if (!ret && write) {
7488 ret = sched_rt_global_constraints();
7490 sysctl_sched_rt_period = old_period;
7491 sysctl_sched_rt_runtime = old_runtime;
7493 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7494 def_rt_bandwidth.rt_period =
7495 ns_to_ktime(global_rt_period());
7498 mutex_unlock(&mutex);
7503 #ifdef CONFIG_CGROUP_SCHED
7505 /* return corresponding task_group object of a cgroup */
7506 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7508 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7509 struct task_group, css);
7512 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7514 struct task_group *tg, *parent;
7516 if (!cgrp->parent) {
7517 /* This is early initialization for the top cgroup */
7518 return &root_task_group.css;
7521 parent = cgroup_tg(cgrp->parent);
7522 tg = sched_create_group(parent);
7524 return ERR_PTR(-ENOMEM);
7529 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7531 struct task_group *tg = cgroup_tg(cgrp);
7533 sched_destroy_group(tg);
7536 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7537 struct cgroup_taskset *tset)
7539 struct task_struct *task;
7541 cgroup_taskset_for_each(task, cgrp, tset) {
7542 #ifdef CONFIG_RT_GROUP_SCHED
7543 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7546 /* We don't support RT-tasks being in separate groups */
7547 if (task->sched_class != &fair_sched_class)
7554 static void cpu_cgroup_attach(struct cgroup *cgrp,
7555 struct cgroup_taskset *tset)
7557 struct task_struct *task;
7559 cgroup_taskset_for_each(task, cgrp, tset)
7560 sched_move_task(task);
7564 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7565 struct task_struct *task)
7568 * cgroup_exit() is called in the copy_process() failure path.
7569 * Ignore this case since the task hasn't ran yet, this avoids
7570 * trying to poke a half freed task state from generic code.
7572 if (!(task->flags & PF_EXITING))
7575 sched_move_task(task);
7578 #ifdef CONFIG_FAIR_GROUP_SCHED
7579 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7582 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7585 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7587 struct task_group *tg = cgroup_tg(cgrp);
7589 return (u64) scale_load_down(tg->shares);
7592 #ifdef CONFIG_CFS_BANDWIDTH
7593 static DEFINE_MUTEX(cfs_constraints_mutex);
7595 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7596 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7598 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7600 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7602 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7603 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7605 if (tg == &root_task_group)
7609 * Ensure we have at some amount of bandwidth every period. This is
7610 * to prevent reaching a state of large arrears when throttled via
7611 * entity_tick() resulting in prolonged exit starvation.
7613 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7617 * Likewise, bound things on the otherside by preventing insane quota
7618 * periods. This also allows us to normalize in computing quota
7621 if (period > max_cfs_quota_period)
7624 mutex_lock(&cfs_constraints_mutex);
7625 ret = __cfs_schedulable(tg, period, quota);
7629 runtime_enabled = quota != RUNTIME_INF;
7630 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7631 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7632 raw_spin_lock_irq(&cfs_b->lock);
7633 cfs_b->period = ns_to_ktime(period);
7634 cfs_b->quota = quota;
7636 __refill_cfs_bandwidth_runtime(cfs_b);
7637 /* restart the period timer (if active) to handle new period expiry */
7638 if (runtime_enabled && cfs_b->timer_active) {
7639 /* force a reprogram */
7640 cfs_b->timer_active = 0;
7641 __start_cfs_bandwidth(cfs_b);
7643 raw_spin_unlock_irq(&cfs_b->lock);
7645 for_each_possible_cpu(i) {
7646 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7647 struct rq *rq = cfs_rq->rq;
7649 raw_spin_lock_irq(&rq->lock);
7650 cfs_rq->runtime_enabled = runtime_enabled;
7651 cfs_rq->runtime_remaining = 0;
7653 if (cfs_rq->throttled)
7654 unthrottle_cfs_rq(cfs_rq);
7655 raw_spin_unlock_irq(&rq->lock);
7658 mutex_unlock(&cfs_constraints_mutex);
7663 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7667 period = ktime_to_ns(tg->cfs_bandwidth.period);
7668 if (cfs_quota_us < 0)
7669 quota = RUNTIME_INF;
7671 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7673 return tg_set_cfs_bandwidth(tg, period, quota);
7676 long tg_get_cfs_quota(struct task_group *tg)
7680 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7683 quota_us = tg->cfs_bandwidth.quota;
7684 do_div(quota_us, NSEC_PER_USEC);
7689 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7693 period = (u64)cfs_period_us * NSEC_PER_USEC;
7694 quota = tg->cfs_bandwidth.quota;
7696 return tg_set_cfs_bandwidth(tg, period, quota);
7699 long tg_get_cfs_period(struct task_group *tg)
7703 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7704 do_div(cfs_period_us, NSEC_PER_USEC);
7706 return cfs_period_us;
7709 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7711 return tg_get_cfs_quota(cgroup_tg(cgrp));
7714 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7717 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7720 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7722 return tg_get_cfs_period(cgroup_tg(cgrp));
7725 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7728 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7731 struct cfs_schedulable_data {
7732 struct task_group *tg;
7737 * normalize group quota/period to be quota/max_period
7738 * note: units are usecs
7740 static u64 normalize_cfs_quota(struct task_group *tg,
7741 struct cfs_schedulable_data *d)
7749 period = tg_get_cfs_period(tg);
7750 quota = tg_get_cfs_quota(tg);
7753 /* note: these should typically be equivalent */
7754 if (quota == RUNTIME_INF || quota == -1)
7757 return to_ratio(period, quota);
7760 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7762 struct cfs_schedulable_data *d = data;
7763 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7764 s64 quota = 0, parent_quota = -1;
7767 quota = RUNTIME_INF;
7769 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7771 quota = normalize_cfs_quota(tg, d);
7772 parent_quota = parent_b->hierarchal_quota;
7775 * ensure max(child_quota) <= parent_quota, inherit when no
7778 if (quota == RUNTIME_INF)
7779 quota = parent_quota;
7780 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7783 cfs_b->hierarchal_quota = quota;
7788 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7791 struct cfs_schedulable_data data = {
7797 if (quota != RUNTIME_INF) {
7798 do_div(data.period, NSEC_PER_USEC);
7799 do_div(data.quota, NSEC_PER_USEC);
7803 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7809 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7810 struct cgroup_map_cb *cb)
7812 struct task_group *tg = cgroup_tg(cgrp);
7813 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7815 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7816 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7817 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7821 #endif /* CONFIG_CFS_BANDWIDTH */
7822 #endif /* CONFIG_FAIR_GROUP_SCHED */
7824 #ifdef CONFIG_RT_GROUP_SCHED
7825 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7828 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7831 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7833 return sched_group_rt_runtime(cgroup_tg(cgrp));
7836 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7839 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7842 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7844 return sched_group_rt_period(cgroup_tg(cgrp));
7846 #endif /* CONFIG_RT_GROUP_SCHED */
7848 static struct cftype cpu_files[] = {
7849 #ifdef CONFIG_FAIR_GROUP_SCHED
7852 .read_u64 = cpu_shares_read_u64,
7853 .write_u64 = cpu_shares_write_u64,
7856 #ifdef CONFIG_CFS_BANDWIDTH
7858 .name = "cfs_quota_us",
7859 .read_s64 = cpu_cfs_quota_read_s64,
7860 .write_s64 = cpu_cfs_quota_write_s64,
7863 .name = "cfs_period_us",
7864 .read_u64 = cpu_cfs_period_read_u64,
7865 .write_u64 = cpu_cfs_period_write_u64,
7869 .read_map = cpu_stats_show,
7872 #ifdef CONFIG_RT_GROUP_SCHED
7874 .name = "rt_runtime_us",
7875 .read_s64 = cpu_rt_runtime_read,
7876 .write_s64 = cpu_rt_runtime_write,
7879 .name = "rt_period_us",
7880 .read_u64 = cpu_rt_period_read_uint,
7881 .write_u64 = cpu_rt_period_write_uint,
7887 struct cgroup_subsys cpu_cgroup_subsys = {
7889 .create = cpu_cgroup_create,
7890 .destroy = cpu_cgroup_destroy,
7891 .can_attach = cpu_cgroup_can_attach,
7892 .attach = cpu_cgroup_attach,
7893 .exit = cpu_cgroup_exit,
7894 .subsys_id = cpu_cgroup_subsys_id,
7895 .base_cftypes = cpu_files,
7899 #endif /* CONFIG_CGROUP_SCHED */
7901 #ifdef CONFIG_CGROUP_CPUACCT
7904 * CPU accounting code for task groups.
7906 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7907 * (balbir@in.ibm.com).
7910 struct cpuacct root_cpuacct;
7912 /* create a new cpu accounting group */
7913 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7918 return &root_cpuacct.css;
7920 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7924 ca->cpuusage = alloc_percpu(u64);
7928 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7930 goto out_free_cpuusage;
7935 free_percpu(ca->cpuusage);
7939 return ERR_PTR(-ENOMEM);
7942 /* destroy an existing cpu accounting group */
7943 static void cpuacct_destroy(struct cgroup *cgrp)
7945 struct cpuacct *ca = cgroup_ca(cgrp);
7947 free_percpu(ca->cpustat);
7948 free_percpu(ca->cpuusage);
7952 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7954 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7957 #ifndef CONFIG_64BIT
7959 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7961 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7963 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7971 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7973 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7975 #ifndef CONFIG_64BIT
7977 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7979 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7981 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7987 /* return total cpu usage (in nanoseconds) of a group */
7988 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7990 struct cpuacct *ca = cgroup_ca(cgrp);
7991 u64 totalcpuusage = 0;
7994 for_each_present_cpu(i)
7995 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7997 return totalcpuusage;
8000 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8003 struct cpuacct *ca = cgroup_ca(cgrp);
8012 for_each_present_cpu(i)
8013 cpuacct_cpuusage_write(ca, i, 0);
8019 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8022 struct cpuacct *ca = cgroup_ca(cgroup);
8026 for_each_present_cpu(i) {
8027 percpu = cpuacct_cpuusage_read(ca, i);
8028 seq_printf(m, "%llu ", (unsigned long long) percpu);
8030 seq_printf(m, "\n");
8034 static const char *cpuacct_stat_desc[] = {
8035 [CPUACCT_STAT_USER] = "user",
8036 [CPUACCT_STAT_SYSTEM] = "system",
8039 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8040 struct cgroup_map_cb *cb)
8042 struct cpuacct *ca = cgroup_ca(cgrp);
8046 for_each_online_cpu(cpu) {
8047 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8048 val += kcpustat->cpustat[CPUTIME_USER];
8049 val += kcpustat->cpustat[CPUTIME_NICE];
8051 val = cputime64_to_clock_t(val);
8052 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8055 for_each_online_cpu(cpu) {
8056 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8057 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8058 val += kcpustat->cpustat[CPUTIME_IRQ];
8059 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8062 val = cputime64_to_clock_t(val);
8063 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8068 static struct cftype files[] = {
8071 .read_u64 = cpuusage_read,
8072 .write_u64 = cpuusage_write,
8075 .name = "usage_percpu",
8076 .read_seq_string = cpuacct_percpu_seq_read,
8080 .read_map = cpuacct_stats_show,
8086 * charge this task's execution time to its accounting group.
8088 * called with rq->lock held.
8090 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8095 if (unlikely(!cpuacct_subsys.active))
8098 cpu = task_cpu(tsk);
8104 for (; ca; ca = parent_ca(ca)) {
8105 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8106 *cpuusage += cputime;
8112 struct cgroup_subsys cpuacct_subsys = {
8114 .create = cpuacct_create,
8115 .destroy = cpuacct_destroy,
8116 .subsys_id = cpuacct_subsys_id,
8117 .base_cftypes = files,
8119 #endif /* CONFIG_CGROUP_CPUACCT */