4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_sched.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
197 sched_feat_write(struct file *filp, const char __user *ubuf,
198 size_t cnt, loff_t *ppos)
208 if (copy_from_user(&buf, ubuf, cnt))
214 if (strncmp(cmp, "NO_", 3) == 0) {
219 for (i = 0; i < __SCHED_FEAT_NR; i++) {
220 if (strcmp(cmp, sched_feat_names[i]) == 0) {
222 sysctl_sched_features &= ~(1UL << i);
223 sched_feat_disable(i);
225 sysctl_sched_features |= (1UL << i);
226 sched_feat_enable(i);
232 if (i == __SCHED_FEAT_NR)
240 static int sched_feat_open(struct inode *inode, struct file *filp)
242 return single_open(filp, sched_feat_show, NULL);
245 static const struct file_operations sched_feat_fops = {
246 .open = sched_feat_open,
247 .write = sched_feat_write,
250 .release = single_release,
253 static __init int sched_init_debug(void)
255 debugfs_create_file("sched_features", 0644, NULL, NULL,
260 late_initcall(sched_init_debug);
261 #endif /* CONFIG_SCHED_DEBUG */
264 * Number of tasks to iterate in a single balance run.
265 * Limited because this is done with IRQs disabled.
267 const_debug unsigned int sysctl_sched_nr_migrate = 32;
270 * period over which we average the RT time consumption, measured
275 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
278 * period over which we measure -rt task cpu usage in us.
281 unsigned int sysctl_sched_rt_period = 1000000;
283 __read_mostly int scheduler_running;
286 * part of the period that we allow rt tasks to run in us.
289 int sysctl_sched_rt_runtime = 950000;
294 * __task_rq_lock - lock the rq @p resides on.
296 static inline struct rq *__task_rq_lock(struct task_struct *p)
301 lockdep_assert_held(&p->pi_lock);
305 raw_spin_lock(&rq->lock);
306 if (likely(rq == task_rq(p)))
308 raw_spin_unlock(&rq->lock);
313 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
315 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
316 __acquires(p->pi_lock)
322 raw_spin_lock_irqsave(&p->pi_lock, *flags);
324 raw_spin_lock(&rq->lock);
325 if (likely(rq == task_rq(p)))
327 raw_spin_unlock(&rq->lock);
328 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
332 static void __task_rq_unlock(struct rq *rq)
335 raw_spin_unlock(&rq->lock);
339 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
341 __releases(p->pi_lock)
343 raw_spin_unlock(&rq->lock);
344 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
348 * this_rq_lock - lock this runqueue and disable interrupts.
350 static struct rq *this_rq_lock(void)
357 raw_spin_lock(&rq->lock);
362 #ifdef CONFIG_SCHED_HRTICK
364 * Use HR-timers to deliver accurate preemption points.
366 * Its all a bit involved since we cannot program an hrt while holding the
367 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
370 * When we get rescheduled we reprogram the hrtick_timer outside of the
374 static void hrtick_clear(struct rq *rq)
376 if (hrtimer_active(&rq->hrtick_timer))
377 hrtimer_cancel(&rq->hrtick_timer);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
386 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
390 raw_spin_lock(&rq->lock);
392 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393 raw_spin_unlock(&rq->lock);
395 return HRTIMER_NORESTART;
400 * called from hardirq (IPI) context
402 static void __hrtick_start(void *arg)
406 raw_spin_lock(&rq->lock);
407 hrtimer_restart(&rq->hrtick_timer);
408 rq->hrtick_csd_pending = 0;
409 raw_spin_unlock(&rq->lock);
413 * Called to set the hrtick timer state.
415 * called with rq->lock held and irqs disabled
417 void hrtick_start(struct rq *rq, u64 delay)
419 struct hrtimer *timer = &rq->hrtick_timer;
420 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
422 hrtimer_set_expires(timer, time);
424 if (rq == this_rq()) {
425 hrtimer_restart(timer);
426 } else if (!rq->hrtick_csd_pending) {
427 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
428 rq->hrtick_csd_pending = 1;
433 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
435 int cpu = (int)(long)hcpu;
438 case CPU_UP_CANCELED:
439 case CPU_UP_CANCELED_FROZEN:
440 case CPU_DOWN_PREPARE:
441 case CPU_DOWN_PREPARE_FROZEN:
443 case CPU_DEAD_FROZEN:
444 hrtick_clear(cpu_rq(cpu));
451 static __init void init_hrtick(void)
453 hotcpu_notifier(hotplug_hrtick, 0);
457 * Called to set the hrtick timer state.
459 * called with rq->lock held and irqs disabled
461 void hrtick_start(struct rq *rq, u64 delay)
463 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
464 HRTIMER_MODE_REL_PINNED, 0);
467 static inline void init_hrtick(void)
470 #endif /* CONFIG_SMP */
472 static void init_rq_hrtick(struct rq *rq)
475 rq->hrtick_csd_pending = 0;
477 rq->hrtick_csd.flags = 0;
478 rq->hrtick_csd.func = __hrtick_start;
479 rq->hrtick_csd.info = rq;
482 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
483 rq->hrtick_timer.function = hrtick;
485 #else /* CONFIG_SCHED_HRTICK */
486 static inline void hrtick_clear(struct rq *rq)
490 static inline void init_rq_hrtick(struct rq *rq)
494 static inline void init_hrtick(void)
497 #endif /* CONFIG_SCHED_HRTICK */
500 * resched_task - mark a task 'to be rescheduled now'.
502 * On UP this means the setting of the need_resched flag, on SMP it
503 * might also involve a cross-CPU call to trigger the scheduler on
508 #ifndef tsk_is_polling
509 #define tsk_is_polling(t) 0
512 void resched_task(struct task_struct *p)
516 assert_raw_spin_locked(&task_rq(p)->lock);
518 if (test_tsk_need_resched(p))
521 set_tsk_need_resched(p);
524 if (cpu == smp_processor_id())
527 /* NEED_RESCHED must be visible before we test polling */
529 if (!tsk_is_polling(p))
530 smp_send_reschedule(cpu);
533 void resched_cpu(int cpu)
535 struct rq *rq = cpu_rq(cpu);
538 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
540 resched_task(cpu_curr(cpu));
541 raw_spin_unlock_irqrestore(&rq->lock, flags);
546 * In the semi idle case, use the nearest busy cpu for migrating timers
547 * from an idle cpu. This is good for power-savings.
549 * We don't do similar optimization for completely idle system, as
550 * selecting an idle cpu will add more delays to the timers than intended
551 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
553 int get_nohz_timer_target(void)
555 int cpu = smp_processor_id();
557 struct sched_domain *sd;
560 for_each_domain(cpu, sd) {
561 for_each_cpu(i, sched_domain_span(sd)) {
573 * When add_timer_on() enqueues a timer into the timer wheel of an
574 * idle CPU then this timer might expire before the next timer event
575 * which is scheduled to wake up that CPU. In case of a completely
576 * idle system the next event might even be infinite time into the
577 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
578 * leaves the inner idle loop so the newly added timer is taken into
579 * account when the CPU goes back to idle and evaluates the timer
580 * wheel for the next timer event.
582 void wake_up_idle_cpu(int cpu)
584 struct rq *rq = cpu_rq(cpu);
586 if (cpu == smp_processor_id())
590 * This is safe, as this function is called with the timer
591 * wheel base lock of (cpu) held. When the CPU is on the way
592 * to idle and has not yet set rq->curr to idle then it will
593 * be serialized on the timer wheel base lock and take the new
594 * timer into account automatically.
596 if (rq->curr != rq->idle)
600 * We can set TIF_RESCHED on the idle task of the other CPU
601 * lockless. The worst case is that the other CPU runs the
602 * idle task through an additional NOOP schedule()
604 set_tsk_need_resched(rq->idle);
606 /* NEED_RESCHED must be visible before we test polling */
608 if (!tsk_is_polling(rq->idle))
609 smp_send_reschedule(cpu);
612 static inline bool got_nohz_idle_kick(void)
614 int cpu = smp_processor_id();
615 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
618 #else /* CONFIG_NO_HZ */
620 static inline bool got_nohz_idle_kick(void)
625 #endif /* CONFIG_NO_HZ */
627 void sched_avg_update(struct rq *rq)
629 s64 period = sched_avg_period();
631 while ((s64)(rq->clock - rq->age_stamp) > period) {
633 * Inline assembly required to prevent the compiler
634 * optimising this loop into a divmod call.
635 * See __iter_div_u64_rem() for another example of this.
637 asm("" : "+rm" (rq->age_stamp));
638 rq->age_stamp += period;
643 #else /* !CONFIG_SMP */
644 void resched_task(struct task_struct *p)
646 assert_raw_spin_locked(&task_rq(p)->lock);
647 set_tsk_need_resched(p);
649 #endif /* CONFIG_SMP */
651 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
652 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
654 * Iterate task_group tree rooted at *from, calling @down when first entering a
655 * node and @up when leaving it for the final time.
657 * Caller must hold rcu_lock or sufficient equivalent.
659 int walk_tg_tree_from(struct task_group *from,
660 tg_visitor down, tg_visitor up, void *data)
662 struct task_group *parent, *child;
668 ret = (*down)(parent, data);
671 list_for_each_entry_rcu(child, &parent->children, siblings) {
678 ret = (*up)(parent, data);
679 if (ret || parent == from)
683 parent = parent->parent;
690 int tg_nop(struct task_group *tg, void *data)
696 static void set_load_weight(struct task_struct *p)
698 int prio = p->static_prio - MAX_RT_PRIO;
699 struct load_weight *load = &p->se.load;
702 * SCHED_IDLE tasks get minimal weight:
704 if (p->policy == SCHED_IDLE) {
705 load->weight = scale_load(WEIGHT_IDLEPRIO);
706 load->inv_weight = WMULT_IDLEPRIO;
710 load->weight = scale_load(prio_to_weight[prio]);
711 load->inv_weight = prio_to_wmult[prio];
714 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
717 sched_info_queued(p);
718 p->sched_class->enqueue_task(rq, p, flags);
721 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
724 sched_info_dequeued(p);
725 p->sched_class->dequeue_task(rq, p, flags);
728 void activate_task(struct rq *rq, struct task_struct *p, int flags)
730 if (task_contributes_to_load(p))
731 rq->nr_uninterruptible--;
733 enqueue_task(rq, p, flags);
736 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
738 if (task_contributes_to_load(p))
739 rq->nr_uninterruptible++;
741 dequeue_task(rq, p, flags);
744 static void update_rq_clock_task(struct rq *rq, s64 delta)
747 * In theory, the compile should just see 0 here, and optimize out the call
748 * to sched_rt_avg_update. But I don't trust it...
750 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
751 s64 steal = 0, irq_delta = 0;
753 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
754 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
757 * Since irq_time is only updated on {soft,}irq_exit, we might run into
758 * this case when a previous update_rq_clock() happened inside a
761 * When this happens, we stop ->clock_task and only update the
762 * prev_irq_time stamp to account for the part that fit, so that a next
763 * update will consume the rest. This ensures ->clock_task is
766 * It does however cause some slight miss-attribution of {soft,}irq
767 * time, a more accurate solution would be to update the irq_time using
768 * the current rq->clock timestamp, except that would require using
771 if (irq_delta > delta)
774 rq->prev_irq_time += irq_delta;
777 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
778 if (static_key_false((¶virt_steal_rq_enabled))) {
781 steal = paravirt_steal_clock(cpu_of(rq));
782 steal -= rq->prev_steal_time_rq;
784 if (unlikely(steal > delta))
787 st = steal_ticks(steal);
788 steal = st * TICK_NSEC;
790 rq->prev_steal_time_rq += steal;
796 rq->clock_task += delta;
798 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
799 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
800 sched_rt_avg_update(rq, irq_delta + steal);
804 void sched_set_stop_task(int cpu, struct task_struct *stop)
806 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
807 struct task_struct *old_stop = cpu_rq(cpu)->stop;
811 * Make it appear like a SCHED_FIFO task, its something
812 * userspace knows about and won't get confused about.
814 * Also, it will make PI more or less work without too
815 * much confusion -- but then, stop work should not
816 * rely on PI working anyway.
818 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
820 stop->sched_class = &stop_sched_class;
823 cpu_rq(cpu)->stop = stop;
827 * Reset it back to a normal scheduling class so that
828 * it can die in pieces.
830 old_stop->sched_class = &rt_sched_class;
835 * __normal_prio - return the priority that is based on the static prio
837 static inline int __normal_prio(struct task_struct *p)
839 return p->static_prio;
843 * Calculate the expected normal priority: i.e. priority
844 * without taking RT-inheritance into account. Might be
845 * boosted by interactivity modifiers. Changes upon fork,
846 * setprio syscalls, and whenever the interactivity
847 * estimator recalculates.
849 static inline int normal_prio(struct task_struct *p)
853 if (task_has_rt_policy(p))
854 prio = MAX_RT_PRIO-1 - p->rt_priority;
856 prio = __normal_prio(p);
861 * Calculate the current priority, i.e. the priority
862 * taken into account by the scheduler. This value might
863 * be boosted by RT tasks, or might be boosted by
864 * interactivity modifiers. Will be RT if the task got
865 * RT-boosted. If not then it returns p->normal_prio.
867 static int effective_prio(struct task_struct *p)
869 p->normal_prio = normal_prio(p);
871 * If we are RT tasks or we were boosted to RT priority,
872 * keep the priority unchanged. Otherwise, update priority
873 * to the normal priority:
875 if (!rt_prio(p->prio))
876 return p->normal_prio;
881 * task_curr - is this task currently executing on a CPU?
882 * @p: the task in question.
884 inline int task_curr(const struct task_struct *p)
886 return cpu_curr(task_cpu(p)) == p;
889 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
890 const struct sched_class *prev_class,
893 if (prev_class != p->sched_class) {
894 if (prev_class->switched_from)
895 prev_class->switched_from(rq, p);
896 p->sched_class->switched_to(rq, p);
897 } else if (oldprio != p->prio)
898 p->sched_class->prio_changed(rq, p, oldprio);
901 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
903 const struct sched_class *class;
905 if (p->sched_class == rq->curr->sched_class) {
906 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
908 for_each_class(class) {
909 if (class == rq->curr->sched_class)
911 if (class == p->sched_class) {
912 resched_task(rq->curr);
919 * A queue event has occurred, and we're going to schedule. In
920 * this case, we can save a useless back to back clock update.
922 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
923 rq->skip_clock_update = 1;
927 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
929 #ifdef CONFIG_SCHED_DEBUG
931 * We should never call set_task_cpu() on a blocked task,
932 * ttwu() will sort out the placement.
934 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
935 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
937 #ifdef CONFIG_LOCKDEP
939 * The caller should hold either p->pi_lock or rq->lock, when changing
940 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
942 * sched_move_task() holds both and thus holding either pins the cgroup,
945 * Furthermore, all task_rq users should acquire both locks, see
948 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
949 lockdep_is_held(&task_rq(p)->lock)));
953 trace_sched_migrate_task(p, new_cpu);
955 if (task_cpu(p) != new_cpu) {
956 p->se.nr_migrations++;
957 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
960 __set_task_cpu(p, new_cpu);
963 struct migration_arg {
964 struct task_struct *task;
968 static int migration_cpu_stop(void *data);
971 * wait_task_inactive - wait for a thread to unschedule.
973 * If @match_state is nonzero, it's the @p->state value just checked and
974 * not expected to change. If it changes, i.e. @p might have woken up,
975 * then return zero. When we succeed in waiting for @p to be off its CPU,
976 * we return a positive number (its total switch count). If a second call
977 * a short while later returns the same number, the caller can be sure that
978 * @p has remained unscheduled the whole time.
980 * The caller must ensure that the task *will* unschedule sometime soon,
981 * else this function might spin for a *long* time. This function can't
982 * be called with interrupts off, or it may introduce deadlock with
983 * smp_call_function() if an IPI is sent by the same process we are
984 * waiting to become inactive.
986 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
995 * We do the initial early heuristics without holding
996 * any task-queue locks at all. We'll only try to get
997 * the runqueue lock when things look like they will
1003 * If the task is actively running on another CPU
1004 * still, just relax and busy-wait without holding
1007 * NOTE! Since we don't hold any locks, it's not
1008 * even sure that "rq" stays as the right runqueue!
1009 * But we don't care, since "task_running()" will
1010 * return false if the runqueue has changed and p
1011 * is actually now running somewhere else!
1013 while (task_running(rq, p)) {
1014 if (match_state && unlikely(p->state != match_state))
1020 * Ok, time to look more closely! We need the rq
1021 * lock now, to be *sure*. If we're wrong, we'll
1022 * just go back and repeat.
1024 rq = task_rq_lock(p, &flags);
1025 trace_sched_wait_task(p);
1026 running = task_running(rq, p);
1029 if (!match_state || p->state == match_state)
1030 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1031 task_rq_unlock(rq, p, &flags);
1034 * If it changed from the expected state, bail out now.
1036 if (unlikely(!ncsw))
1040 * Was it really running after all now that we
1041 * checked with the proper locks actually held?
1043 * Oops. Go back and try again..
1045 if (unlikely(running)) {
1051 * It's not enough that it's not actively running,
1052 * it must be off the runqueue _entirely_, and not
1055 * So if it was still runnable (but just not actively
1056 * running right now), it's preempted, and we should
1057 * yield - it could be a while.
1059 if (unlikely(on_rq)) {
1060 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1062 set_current_state(TASK_UNINTERRUPTIBLE);
1063 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1068 * Ahh, all good. It wasn't running, and it wasn't
1069 * runnable, which means that it will never become
1070 * running in the future either. We're all done!
1079 * kick_process - kick a running thread to enter/exit the kernel
1080 * @p: the to-be-kicked thread
1082 * Cause a process which is running on another CPU to enter
1083 * kernel-mode, without any delay. (to get signals handled.)
1085 * NOTE: this function doesn't have to take the runqueue lock,
1086 * because all it wants to ensure is that the remote task enters
1087 * the kernel. If the IPI races and the task has been migrated
1088 * to another CPU then no harm is done and the purpose has been
1091 void kick_process(struct task_struct *p)
1097 if ((cpu != smp_processor_id()) && task_curr(p))
1098 smp_send_reschedule(cpu);
1101 EXPORT_SYMBOL_GPL(kick_process);
1102 #endif /* CONFIG_SMP */
1106 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1108 static int select_fallback_rq(int cpu, struct task_struct *p)
1110 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1111 enum { cpuset, possible, fail } state = cpuset;
1114 /* Look for allowed, online CPU in same node. */
1115 for_each_cpu(dest_cpu, nodemask) {
1116 if (!cpu_online(dest_cpu))
1118 if (!cpu_active(dest_cpu))
1120 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1125 /* Any allowed, online CPU? */
1126 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1127 if (!cpu_online(dest_cpu))
1129 if (!cpu_active(dest_cpu))
1136 /* No more Mr. Nice Guy. */
1137 cpuset_cpus_allowed_fallback(p);
1142 do_set_cpus_allowed(p, cpu_possible_mask);
1153 if (state != cpuset) {
1155 * Don't tell them about moving exiting tasks or
1156 * kernel threads (both mm NULL), since they never
1159 if (p->mm && printk_ratelimit()) {
1160 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1161 task_pid_nr(p), p->comm, cpu);
1169 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1172 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1174 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1177 * In order not to call set_task_cpu() on a blocking task we need
1178 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1181 * Since this is common to all placement strategies, this lives here.
1183 * [ this allows ->select_task() to simply return task_cpu(p) and
1184 * not worry about this generic constraint ]
1186 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1188 cpu = select_fallback_rq(task_cpu(p), p);
1193 static void update_avg(u64 *avg, u64 sample)
1195 s64 diff = sample - *avg;
1201 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1203 #ifdef CONFIG_SCHEDSTATS
1204 struct rq *rq = this_rq();
1207 int this_cpu = smp_processor_id();
1209 if (cpu == this_cpu) {
1210 schedstat_inc(rq, ttwu_local);
1211 schedstat_inc(p, se.statistics.nr_wakeups_local);
1213 struct sched_domain *sd;
1215 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1217 for_each_domain(this_cpu, sd) {
1218 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1219 schedstat_inc(sd, ttwu_wake_remote);
1226 if (wake_flags & WF_MIGRATED)
1227 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1229 #endif /* CONFIG_SMP */
1231 schedstat_inc(rq, ttwu_count);
1232 schedstat_inc(p, se.statistics.nr_wakeups);
1234 if (wake_flags & WF_SYNC)
1235 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1237 #endif /* CONFIG_SCHEDSTATS */
1240 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1242 activate_task(rq, p, en_flags);
1245 /* if a worker is waking up, notify workqueue */
1246 if (p->flags & PF_WQ_WORKER)
1247 wq_worker_waking_up(p, cpu_of(rq));
1251 * Mark the task runnable and perform wakeup-preemption.
1254 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1256 trace_sched_wakeup(p, true);
1257 check_preempt_curr(rq, p, wake_flags);
1259 p->state = TASK_RUNNING;
1261 if (p->sched_class->task_woken)
1262 p->sched_class->task_woken(rq, p);
1264 if (rq->idle_stamp) {
1265 u64 delta = rq->clock - rq->idle_stamp;
1266 u64 max = 2*sysctl_sched_migration_cost;
1271 update_avg(&rq->avg_idle, delta);
1278 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1281 if (p->sched_contributes_to_load)
1282 rq->nr_uninterruptible--;
1285 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1286 ttwu_do_wakeup(rq, p, wake_flags);
1290 * Called in case the task @p isn't fully descheduled from its runqueue,
1291 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1292 * since all we need to do is flip p->state to TASK_RUNNING, since
1293 * the task is still ->on_rq.
1295 static int ttwu_remote(struct task_struct *p, int wake_flags)
1300 rq = __task_rq_lock(p);
1302 ttwu_do_wakeup(rq, p, wake_flags);
1305 __task_rq_unlock(rq);
1311 static void sched_ttwu_pending(void)
1313 struct rq *rq = this_rq();
1314 struct llist_node *llist = llist_del_all(&rq->wake_list);
1315 struct task_struct *p;
1317 raw_spin_lock(&rq->lock);
1320 p = llist_entry(llist, struct task_struct, wake_entry);
1321 llist = llist_next(llist);
1322 ttwu_do_activate(rq, p, 0);
1325 raw_spin_unlock(&rq->lock);
1328 void scheduler_ipi(void)
1330 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1334 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1335 * traditionally all their work was done from the interrupt return
1336 * path. Now that we actually do some work, we need to make sure
1339 * Some archs already do call them, luckily irq_enter/exit nest
1342 * Arguably we should visit all archs and update all handlers,
1343 * however a fair share of IPIs are still resched only so this would
1344 * somewhat pessimize the simple resched case.
1347 sched_ttwu_pending();
1350 * Check if someone kicked us for doing the nohz idle load balance.
1352 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1353 this_rq()->idle_balance = 1;
1354 raise_softirq_irqoff(SCHED_SOFTIRQ);
1359 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1361 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1362 smp_send_reschedule(cpu);
1365 bool cpus_share_cache(int this_cpu, int that_cpu)
1367 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1369 #endif /* CONFIG_SMP */
1371 static void ttwu_queue(struct task_struct *p, int cpu)
1373 struct rq *rq = cpu_rq(cpu);
1375 #if defined(CONFIG_SMP)
1376 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1377 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1378 ttwu_queue_remote(p, cpu);
1383 raw_spin_lock(&rq->lock);
1384 ttwu_do_activate(rq, p, 0);
1385 raw_spin_unlock(&rq->lock);
1389 * try_to_wake_up - wake up a thread
1390 * @p: the thread to be awakened
1391 * @state: the mask of task states that can be woken
1392 * @wake_flags: wake modifier flags (WF_*)
1394 * Put it on the run-queue if it's not already there. The "current"
1395 * thread is always on the run-queue (except when the actual
1396 * re-schedule is in progress), and as such you're allowed to do
1397 * the simpler "current->state = TASK_RUNNING" to mark yourself
1398 * runnable without the overhead of this.
1400 * Returns %true if @p was woken up, %false if it was already running
1401 * or @state didn't match @p's state.
1404 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1406 unsigned long flags;
1407 int cpu, success = 0;
1410 raw_spin_lock_irqsave(&p->pi_lock, flags);
1411 if (!(p->state & state))
1414 success = 1; /* we're going to change ->state */
1417 if (p->on_rq && ttwu_remote(p, wake_flags))
1422 * If the owning (remote) cpu is still in the middle of schedule() with
1423 * this task as prev, wait until its done referencing the task.
1428 * Pairs with the smp_wmb() in finish_lock_switch().
1432 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1433 p->state = TASK_WAKING;
1435 if (p->sched_class->task_waking)
1436 p->sched_class->task_waking(p);
1438 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1439 if (task_cpu(p) != cpu) {
1440 wake_flags |= WF_MIGRATED;
1441 set_task_cpu(p, cpu);
1443 #endif /* CONFIG_SMP */
1447 ttwu_stat(p, cpu, wake_flags);
1449 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1455 * try_to_wake_up_local - try to wake up a local task with rq lock held
1456 * @p: the thread to be awakened
1458 * Put @p on the run-queue if it's not already there. The caller must
1459 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1462 static void try_to_wake_up_local(struct task_struct *p)
1464 struct rq *rq = task_rq(p);
1466 BUG_ON(rq != this_rq());
1467 BUG_ON(p == current);
1468 lockdep_assert_held(&rq->lock);
1470 if (!raw_spin_trylock(&p->pi_lock)) {
1471 raw_spin_unlock(&rq->lock);
1472 raw_spin_lock(&p->pi_lock);
1473 raw_spin_lock(&rq->lock);
1476 if (!(p->state & TASK_NORMAL))
1480 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1482 ttwu_do_wakeup(rq, p, 0);
1483 ttwu_stat(p, smp_processor_id(), 0);
1485 raw_spin_unlock(&p->pi_lock);
1489 * wake_up_process - Wake up a specific process
1490 * @p: The process to be woken up.
1492 * Attempt to wake up the nominated process and move it to the set of runnable
1493 * processes. Returns 1 if the process was woken up, 0 if it was already
1496 * It may be assumed that this function implies a write memory barrier before
1497 * changing the task state if and only if any tasks are woken up.
1499 int wake_up_process(struct task_struct *p)
1501 return try_to_wake_up(p, TASK_ALL, 0);
1503 EXPORT_SYMBOL(wake_up_process);
1505 int wake_up_state(struct task_struct *p, unsigned int state)
1507 return try_to_wake_up(p, state, 0);
1511 * Perform scheduler related setup for a newly forked process p.
1512 * p is forked by current.
1514 * __sched_fork() is basic setup used by init_idle() too:
1516 static void __sched_fork(struct task_struct *p)
1521 p->se.exec_start = 0;
1522 p->se.sum_exec_runtime = 0;
1523 p->se.prev_sum_exec_runtime = 0;
1524 p->se.nr_migrations = 0;
1526 INIT_LIST_HEAD(&p->se.group_node);
1528 #ifdef CONFIG_SCHEDSTATS
1529 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1532 INIT_LIST_HEAD(&p->rt.run_list);
1534 #ifdef CONFIG_PREEMPT_NOTIFIERS
1535 INIT_HLIST_HEAD(&p->preempt_notifiers);
1540 * fork()/clone()-time setup:
1542 void sched_fork(struct task_struct *p)
1544 unsigned long flags;
1545 int cpu = get_cpu();
1549 * We mark the process as running here. This guarantees that
1550 * nobody will actually run it, and a signal or other external
1551 * event cannot wake it up and insert it on the runqueue either.
1553 p->state = TASK_RUNNING;
1556 * Make sure we do not leak PI boosting priority to the child.
1558 p->prio = current->normal_prio;
1561 * Revert to default priority/policy on fork if requested.
1563 if (unlikely(p->sched_reset_on_fork)) {
1564 if (task_has_rt_policy(p)) {
1565 p->policy = SCHED_NORMAL;
1566 p->static_prio = NICE_TO_PRIO(0);
1568 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1569 p->static_prio = NICE_TO_PRIO(0);
1571 p->prio = p->normal_prio = __normal_prio(p);
1575 * We don't need the reset flag anymore after the fork. It has
1576 * fulfilled its duty:
1578 p->sched_reset_on_fork = 0;
1581 if (!rt_prio(p->prio))
1582 p->sched_class = &fair_sched_class;
1584 if (p->sched_class->task_fork)
1585 p->sched_class->task_fork(p);
1588 * The child is not yet in the pid-hash so no cgroup attach races,
1589 * and the cgroup is pinned to this child due to cgroup_fork()
1590 * is ran before sched_fork().
1592 * Silence PROVE_RCU.
1594 raw_spin_lock_irqsave(&p->pi_lock, flags);
1595 set_task_cpu(p, cpu);
1596 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1598 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1599 if (likely(sched_info_on()))
1600 memset(&p->sched_info, 0, sizeof(p->sched_info));
1602 #if defined(CONFIG_SMP)
1605 #ifdef CONFIG_PREEMPT_COUNT
1606 /* Want to start with kernel preemption disabled. */
1607 task_thread_info(p)->preempt_count = 1;
1610 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1617 * wake_up_new_task - wake up a newly created task for the first time.
1619 * This function will do some initial scheduler statistics housekeeping
1620 * that must be done for every newly created context, then puts the task
1621 * on the runqueue and wakes it.
1623 void wake_up_new_task(struct task_struct *p)
1625 unsigned long flags;
1628 raw_spin_lock_irqsave(&p->pi_lock, flags);
1631 * Fork balancing, do it here and not earlier because:
1632 * - cpus_allowed can change in the fork path
1633 * - any previously selected cpu might disappear through hotplug
1635 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1638 rq = __task_rq_lock(p);
1639 activate_task(rq, p, 0);
1641 trace_sched_wakeup_new(p, true);
1642 check_preempt_curr(rq, p, WF_FORK);
1644 if (p->sched_class->task_woken)
1645 p->sched_class->task_woken(rq, p);
1647 task_rq_unlock(rq, p, &flags);
1650 #ifdef CONFIG_PREEMPT_NOTIFIERS
1653 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1654 * @notifier: notifier struct to register
1656 void preempt_notifier_register(struct preempt_notifier *notifier)
1658 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1660 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1663 * preempt_notifier_unregister - no longer interested in preemption notifications
1664 * @notifier: notifier struct to unregister
1666 * This is safe to call from within a preemption notifier.
1668 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1670 hlist_del(¬ifier->link);
1672 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1674 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1676 struct preempt_notifier *notifier;
1677 struct hlist_node *node;
1679 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1680 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1684 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1685 struct task_struct *next)
1687 struct preempt_notifier *notifier;
1688 struct hlist_node *node;
1690 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1691 notifier->ops->sched_out(notifier, next);
1694 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1696 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1701 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1702 struct task_struct *next)
1706 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1709 * prepare_task_switch - prepare to switch tasks
1710 * @rq: the runqueue preparing to switch
1711 * @prev: the current task that is being switched out
1712 * @next: the task we are going to switch to.
1714 * This is called with the rq lock held and interrupts off. It must
1715 * be paired with a subsequent finish_task_switch after the context
1718 * prepare_task_switch sets up locking and calls architecture specific
1722 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1723 struct task_struct *next)
1725 trace_sched_switch(prev, next);
1726 sched_info_switch(prev, next);
1727 perf_event_task_sched_out(prev, next);
1728 fire_sched_out_preempt_notifiers(prev, next);
1729 prepare_lock_switch(rq, next);
1730 prepare_arch_switch(next);
1734 * finish_task_switch - clean up after a task-switch
1735 * @rq: runqueue associated with task-switch
1736 * @prev: the thread we just switched away from.
1738 * finish_task_switch must be called after the context switch, paired
1739 * with a prepare_task_switch call before the context switch.
1740 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1741 * and do any other architecture-specific cleanup actions.
1743 * Note that we may have delayed dropping an mm in context_switch(). If
1744 * so, we finish that here outside of the runqueue lock. (Doing it
1745 * with the lock held can cause deadlocks; see schedule() for
1748 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1749 __releases(rq->lock)
1751 struct mm_struct *mm = rq->prev_mm;
1757 * A task struct has one reference for the use as "current".
1758 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1759 * schedule one last time. The schedule call will never return, and
1760 * the scheduled task must drop that reference.
1761 * The test for TASK_DEAD must occur while the runqueue locks are
1762 * still held, otherwise prev could be scheduled on another cpu, die
1763 * there before we look at prev->state, and then the reference would
1765 * Manfred Spraul <manfred@colorfullife.com>
1767 prev_state = prev->state;
1768 vtime_task_switch(prev);
1769 finish_arch_switch(prev);
1770 perf_event_task_sched_in(prev, current);
1771 finish_lock_switch(rq, prev);
1772 finish_arch_post_lock_switch();
1774 fire_sched_in_preempt_notifiers(current);
1777 if (unlikely(prev_state == TASK_DEAD)) {
1779 * Remove function-return probe instances associated with this
1780 * task and put them back on the free list.
1782 kprobe_flush_task(prev);
1783 put_task_struct(prev);
1789 /* assumes rq->lock is held */
1790 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1792 if (prev->sched_class->pre_schedule)
1793 prev->sched_class->pre_schedule(rq, prev);
1796 /* rq->lock is NOT held, but preemption is disabled */
1797 static inline void post_schedule(struct rq *rq)
1799 if (rq->post_schedule) {
1800 unsigned long flags;
1802 raw_spin_lock_irqsave(&rq->lock, flags);
1803 if (rq->curr->sched_class->post_schedule)
1804 rq->curr->sched_class->post_schedule(rq);
1805 raw_spin_unlock_irqrestore(&rq->lock, flags);
1807 rq->post_schedule = 0;
1813 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1817 static inline void post_schedule(struct rq *rq)
1824 * schedule_tail - first thing a freshly forked thread must call.
1825 * @prev: the thread we just switched away from.
1827 asmlinkage void schedule_tail(struct task_struct *prev)
1828 __releases(rq->lock)
1830 struct rq *rq = this_rq();
1832 finish_task_switch(rq, prev);
1835 * FIXME: do we need to worry about rq being invalidated by the
1840 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1841 /* In this case, finish_task_switch does not reenable preemption */
1844 if (current->set_child_tid)
1845 put_user(task_pid_vnr(current), current->set_child_tid);
1849 * context_switch - switch to the new MM and the new
1850 * thread's register state.
1853 context_switch(struct rq *rq, struct task_struct *prev,
1854 struct task_struct *next)
1856 struct mm_struct *mm, *oldmm;
1858 prepare_task_switch(rq, prev, next);
1861 oldmm = prev->active_mm;
1863 * For paravirt, this is coupled with an exit in switch_to to
1864 * combine the page table reload and the switch backend into
1867 arch_start_context_switch(prev);
1870 next->active_mm = oldmm;
1871 atomic_inc(&oldmm->mm_count);
1872 enter_lazy_tlb(oldmm, next);
1874 switch_mm(oldmm, mm, next);
1877 prev->active_mm = NULL;
1878 rq->prev_mm = oldmm;
1881 * Since the runqueue lock will be released by the next
1882 * task (which is an invalid locking op but in the case
1883 * of the scheduler it's an obvious special-case), so we
1884 * do an early lockdep release here:
1886 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1887 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1890 context_tracking_task_switch(prev, next);
1891 /* Here we just switch the register state and the stack. */
1892 switch_to(prev, next, prev);
1896 * this_rq must be evaluated again because prev may have moved
1897 * CPUs since it called schedule(), thus the 'rq' on its stack
1898 * frame will be invalid.
1900 finish_task_switch(this_rq(), prev);
1904 * nr_running, nr_uninterruptible and nr_context_switches:
1906 * externally visible scheduler statistics: current number of runnable
1907 * threads, current number of uninterruptible-sleeping threads, total
1908 * number of context switches performed since bootup.
1910 unsigned long nr_running(void)
1912 unsigned long i, sum = 0;
1914 for_each_online_cpu(i)
1915 sum += cpu_rq(i)->nr_running;
1920 unsigned long nr_uninterruptible(void)
1922 unsigned long i, sum = 0;
1924 for_each_possible_cpu(i)
1925 sum += cpu_rq(i)->nr_uninterruptible;
1928 * Since we read the counters lockless, it might be slightly
1929 * inaccurate. Do not allow it to go below zero though:
1931 if (unlikely((long)sum < 0))
1937 unsigned long long nr_context_switches(void)
1940 unsigned long long sum = 0;
1942 for_each_possible_cpu(i)
1943 sum += cpu_rq(i)->nr_switches;
1948 unsigned long nr_iowait(void)
1950 unsigned long i, sum = 0;
1952 for_each_possible_cpu(i)
1953 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1958 unsigned long nr_iowait_cpu(int cpu)
1960 struct rq *this = cpu_rq(cpu);
1961 return atomic_read(&this->nr_iowait);
1964 unsigned long this_cpu_load(void)
1966 struct rq *this = this_rq();
1967 return this->cpu_load[0];
1972 * Global load-average calculations
1974 * We take a distributed and async approach to calculating the global load-avg
1975 * in order to minimize overhead.
1977 * The global load average is an exponentially decaying average of nr_running +
1978 * nr_uninterruptible.
1980 * Once every LOAD_FREQ:
1983 * for_each_possible_cpu(cpu)
1984 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1986 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1988 * Due to a number of reasons the above turns in the mess below:
1990 * - for_each_possible_cpu() is prohibitively expensive on machines with
1991 * serious number of cpus, therefore we need to take a distributed approach
1992 * to calculating nr_active.
1994 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
1995 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
1997 * So assuming nr_active := 0 when we start out -- true per definition, we
1998 * can simply take per-cpu deltas and fold those into a global accumulate
1999 * to obtain the same result. See calc_load_fold_active().
2001 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2002 * across the machine, we assume 10 ticks is sufficient time for every
2003 * cpu to have completed this task.
2005 * This places an upper-bound on the IRQ-off latency of the machine. Then
2006 * again, being late doesn't loose the delta, just wrecks the sample.
2008 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2009 * this would add another cross-cpu cacheline miss and atomic operation
2010 * to the wakeup path. Instead we increment on whatever cpu the task ran
2011 * when it went into uninterruptible state and decrement on whatever cpu
2012 * did the wakeup. This means that only the sum of nr_uninterruptible over
2013 * all cpus yields the correct result.
2015 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2018 /* Variables and functions for calc_load */
2019 static atomic_long_t calc_load_tasks;
2020 static unsigned long calc_load_update;
2021 unsigned long avenrun[3];
2022 EXPORT_SYMBOL(avenrun); /* should be removed */
2025 * get_avenrun - get the load average array
2026 * @loads: pointer to dest load array
2027 * @offset: offset to add
2028 * @shift: shift count to shift the result left
2030 * These values are estimates at best, so no need for locking.
2032 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2034 loads[0] = (avenrun[0] + offset) << shift;
2035 loads[1] = (avenrun[1] + offset) << shift;
2036 loads[2] = (avenrun[2] + offset) << shift;
2039 static long calc_load_fold_active(struct rq *this_rq)
2041 long nr_active, delta = 0;
2043 nr_active = this_rq->nr_running;
2044 nr_active += (long) this_rq->nr_uninterruptible;
2046 if (nr_active != this_rq->calc_load_active) {
2047 delta = nr_active - this_rq->calc_load_active;
2048 this_rq->calc_load_active = nr_active;
2055 * a1 = a0 * e + a * (1 - e)
2057 static unsigned long
2058 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2061 load += active * (FIXED_1 - exp);
2062 load += 1UL << (FSHIFT - 1);
2063 return load >> FSHIFT;
2068 * Handle NO_HZ for the global load-average.
2070 * Since the above described distributed algorithm to compute the global
2071 * load-average relies on per-cpu sampling from the tick, it is affected by
2074 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2075 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2076 * when we read the global state.
2078 * Obviously reality has to ruin such a delightfully simple scheme:
2080 * - When we go NO_HZ idle during the window, we can negate our sample
2081 * contribution, causing under-accounting.
2083 * We avoid this by keeping two idle-delta counters and flipping them
2084 * when the window starts, thus separating old and new NO_HZ load.
2086 * The only trick is the slight shift in index flip for read vs write.
2090 * |-|-----------|-|-----------|-|-----------|-|
2091 * r:0 0 1 1 0 0 1 1 0
2092 * w:0 1 1 0 0 1 1 0 0
2094 * This ensures we'll fold the old idle contribution in this window while
2095 * accumlating the new one.
2097 * - When we wake up from NO_HZ idle during the window, we push up our
2098 * contribution, since we effectively move our sample point to a known
2101 * This is solved by pushing the window forward, and thus skipping the
2102 * sample, for this cpu (effectively using the idle-delta for this cpu which
2103 * was in effect at the time the window opened). This also solves the issue
2104 * of having to deal with a cpu having been in NOHZ idle for multiple
2105 * LOAD_FREQ intervals.
2107 * When making the ILB scale, we should try to pull this in as well.
2109 static atomic_long_t calc_load_idle[2];
2110 static int calc_load_idx;
2112 static inline int calc_load_write_idx(void)
2114 int idx = calc_load_idx;
2117 * See calc_global_nohz(), if we observe the new index, we also
2118 * need to observe the new update time.
2123 * If the folding window started, make sure we start writing in the
2126 if (!time_before(jiffies, calc_load_update))
2132 static inline int calc_load_read_idx(void)
2134 return calc_load_idx & 1;
2137 void calc_load_enter_idle(void)
2139 struct rq *this_rq = this_rq();
2143 * We're going into NOHZ mode, if there's any pending delta, fold it
2144 * into the pending idle delta.
2146 delta = calc_load_fold_active(this_rq);
2148 int idx = calc_load_write_idx();
2149 atomic_long_add(delta, &calc_load_idle[idx]);
2153 void calc_load_exit_idle(void)
2155 struct rq *this_rq = this_rq();
2158 * If we're still before the sample window, we're done.
2160 if (time_before(jiffies, this_rq->calc_load_update))
2164 * We woke inside or after the sample window, this means we're already
2165 * accounted through the nohz accounting, so skip the entire deal and
2166 * sync up for the next window.
2168 this_rq->calc_load_update = calc_load_update;
2169 if (time_before(jiffies, this_rq->calc_load_update + 10))
2170 this_rq->calc_load_update += LOAD_FREQ;
2173 static long calc_load_fold_idle(void)
2175 int idx = calc_load_read_idx();
2178 if (atomic_long_read(&calc_load_idle[idx]))
2179 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2185 * fixed_power_int - compute: x^n, in O(log n) time
2187 * @x: base of the power
2188 * @frac_bits: fractional bits of @x
2189 * @n: power to raise @x to.
2191 * By exploiting the relation between the definition of the natural power
2192 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2193 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2194 * (where: n_i \elem {0, 1}, the binary vector representing n),
2195 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2196 * of course trivially computable in O(log_2 n), the length of our binary
2199 static unsigned long
2200 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2202 unsigned long result = 1UL << frac_bits;
2207 result += 1UL << (frac_bits - 1);
2208 result >>= frac_bits;
2214 x += 1UL << (frac_bits - 1);
2222 * a1 = a0 * e + a * (1 - e)
2224 * a2 = a1 * e + a * (1 - e)
2225 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2226 * = a0 * e^2 + a * (1 - e) * (1 + e)
2228 * a3 = a2 * e + a * (1 - e)
2229 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2230 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2234 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2235 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2236 * = a0 * e^n + a * (1 - e^n)
2238 * [1] application of the geometric series:
2241 * S_n := \Sum x^i = -------------
2244 static unsigned long
2245 calc_load_n(unsigned long load, unsigned long exp,
2246 unsigned long active, unsigned int n)
2249 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2253 * NO_HZ can leave us missing all per-cpu ticks calling
2254 * calc_load_account_active(), but since an idle CPU folds its delta into
2255 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2256 * in the pending idle delta if our idle period crossed a load cycle boundary.
2258 * Once we've updated the global active value, we need to apply the exponential
2259 * weights adjusted to the number of cycles missed.
2261 static void calc_global_nohz(void)
2263 long delta, active, n;
2265 if (!time_before(jiffies, calc_load_update + 10)) {
2267 * Catch-up, fold however many we are behind still
2269 delta = jiffies - calc_load_update - 10;
2270 n = 1 + (delta / LOAD_FREQ);
2272 active = atomic_long_read(&calc_load_tasks);
2273 active = active > 0 ? active * FIXED_1 : 0;
2275 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2276 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2277 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2279 calc_load_update += n * LOAD_FREQ;
2283 * Flip the idle index...
2285 * Make sure we first write the new time then flip the index, so that
2286 * calc_load_write_idx() will see the new time when it reads the new
2287 * index, this avoids a double flip messing things up.
2292 #else /* !CONFIG_NO_HZ */
2294 static inline long calc_load_fold_idle(void) { return 0; }
2295 static inline void calc_global_nohz(void) { }
2297 #endif /* CONFIG_NO_HZ */
2300 * calc_load - update the avenrun load estimates 10 ticks after the
2301 * CPUs have updated calc_load_tasks.
2303 void calc_global_load(unsigned long ticks)
2307 if (time_before(jiffies, calc_load_update + 10))
2311 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2313 delta = calc_load_fold_idle();
2315 atomic_long_add(delta, &calc_load_tasks);
2317 active = atomic_long_read(&calc_load_tasks);
2318 active = active > 0 ? active * FIXED_1 : 0;
2320 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2321 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2322 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2324 calc_load_update += LOAD_FREQ;
2327 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2333 * Called from update_cpu_load() to periodically update this CPU's
2336 static void calc_load_account_active(struct rq *this_rq)
2340 if (time_before(jiffies, this_rq->calc_load_update))
2343 delta = calc_load_fold_active(this_rq);
2345 atomic_long_add(delta, &calc_load_tasks);
2347 this_rq->calc_load_update += LOAD_FREQ;
2351 * End of global load-average stuff
2355 * The exact cpuload at various idx values, calculated at every tick would be
2356 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2358 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2359 * on nth tick when cpu may be busy, then we have:
2360 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2361 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2363 * decay_load_missed() below does efficient calculation of
2364 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2365 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2367 * The calculation is approximated on a 128 point scale.
2368 * degrade_zero_ticks is the number of ticks after which load at any
2369 * particular idx is approximated to be zero.
2370 * degrade_factor is a precomputed table, a row for each load idx.
2371 * Each column corresponds to degradation factor for a power of two ticks,
2372 * based on 128 point scale.
2374 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2375 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2377 * With this power of 2 load factors, we can degrade the load n times
2378 * by looking at 1 bits in n and doing as many mult/shift instead of
2379 * n mult/shifts needed by the exact degradation.
2381 #define DEGRADE_SHIFT 7
2382 static const unsigned char
2383 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2384 static const unsigned char
2385 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2386 {0, 0, 0, 0, 0, 0, 0, 0},
2387 {64, 32, 8, 0, 0, 0, 0, 0},
2388 {96, 72, 40, 12, 1, 0, 0},
2389 {112, 98, 75, 43, 15, 1, 0},
2390 {120, 112, 98, 76, 45, 16, 2} };
2393 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2394 * would be when CPU is idle and so we just decay the old load without
2395 * adding any new load.
2397 static unsigned long
2398 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2402 if (!missed_updates)
2405 if (missed_updates >= degrade_zero_ticks[idx])
2409 return load >> missed_updates;
2411 while (missed_updates) {
2412 if (missed_updates % 2)
2413 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2415 missed_updates >>= 1;
2422 * Update rq->cpu_load[] statistics. This function is usually called every
2423 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2424 * every tick. We fix it up based on jiffies.
2426 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2427 unsigned long pending_updates)
2431 this_rq->nr_load_updates++;
2433 /* Update our load: */
2434 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2435 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2436 unsigned long old_load, new_load;
2438 /* scale is effectively 1 << i now, and >> i divides by scale */
2440 old_load = this_rq->cpu_load[i];
2441 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2442 new_load = this_load;
2444 * Round up the averaging division if load is increasing. This
2445 * prevents us from getting stuck on 9 if the load is 10, for
2448 if (new_load > old_load)
2449 new_load += scale - 1;
2451 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2454 sched_avg_update(this_rq);
2459 * There is no sane way to deal with nohz on smp when using jiffies because the
2460 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2461 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2463 * Therefore we cannot use the delta approach from the regular tick since that
2464 * would seriously skew the load calculation. However we'll make do for those
2465 * updates happening while idle (nohz_idle_balance) or coming out of idle
2466 * (tick_nohz_idle_exit).
2468 * This means we might still be one tick off for nohz periods.
2472 * Called from nohz_idle_balance() to update the load ratings before doing the
2475 void update_idle_cpu_load(struct rq *this_rq)
2477 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2478 unsigned long load = this_rq->load.weight;
2479 unsigned long pending_updates;
2482 * bail if there's load or we're actually up-to-date.
2484 if (load || curr_jiffies == this_rq->last_load_update_tick)
2487 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2488 this_rq->last_load_update_tick = curr_jiffies;
2490 __update_cpu_load(this_rq, load, pending_updates);
2494 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2496 void update_cpu_load_nohz(void)
2498 struct rq *this_rq = this_rq();
2499 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2500 unsigned long pending_updates;
2502 if (curr_jiffies == this_rq->last_load_update_tick)
2505 raw_spin_lock(&this_rq->lock);
2506 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2507 if (pending_updates) {
2508 this_rq->last_load_update_tick = curr_jiffies;
2510 * We were idle, this means load 0, the current load might be
2511 * !0 due to remote wakeups and the sort.
2513 __update_cpu_load(this_rq, 0, pending_updates);
2515 raw_spin_unlock(&this_rq->lock);
2517 #endif /* CONFIG_NO_HZ */
2520 * Called from scheduler_tick()
2522 static void update_cpu_load_active(struct rq *this_rq)
2525 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2527 this_rq->last_load_update_tick = jiffies;
2528 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2530 calc_load_account_active(this_rq);
2536 * sched_exec - execve() is a valuable balancing opportunity, because at
2537 * this point the task has the smallest effective memory and cache footprint.
2539 void sched_exec(void)
2541 struct task_struct *p = current;
2542 unsigned long flags;
2545 raw_spin_lock_irqsave(&p->pi_lock, flags);
2546 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2547 if (dest_cpu == smp_processor_id())
2550 if (likely(cpu_active(dest_cpu))) {
2551 struct migration_arg arg = { p, dest_cpu };
2553 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2554 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2558 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2563 DEFINE_PER_CPU(struct kernel_stat, kstat);
2564 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2566 EXPORT_PER_CPU_SYMBOL(kstat);
2567 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2570 * Return any ns on the sched_clock that have not yet been accounted in
2571 * @p in case that task is currently running.
2573 * Called with task_rq_lock() held on @rq.
2575 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2579 if (task_current(rq, p)) {
2580 update_rq_clock(rq);
2581 ns = rq->clock_task - p->se.exec_start;
2589 unsigned long long task_delta_exec(struct task_struct *p)
2591 unsigned long flags;
2595 rq = task_rq_lock(p, &flags);
2596 ns = do_task_delta_exec(p, rq);
2597 task_rq_unlock(rq, p, &flags);
2603 * Return accounted runtime for the task.
2604 * In case the task is currently running, return the runtime plus current's
2605 * pending runtime that have not been accounted yet.
2607 unsigned long long task_sched_runtime(struct task_struct *p)
2609 unsigned long flags;
2613 rq = task_rq_lock(p, &flags);
2614 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2615 task_rq_unlock(rq, p, &flags);
2621 * This function gets called by the timer code, with HZ frequency.
2622 * We call it with interrupts disabled.
2624 void scheduler_tick(void)
2626 int cpu = smp_processor_id();
2627 struct rq *rq = cpu_rq(cpu);
2628 struct task_struct *curr = rq->curr;
2632 raw_spin_lock(&rq->lock);
2633 update_rq_clock(rq);
2634 update_cpu_load_active(rq);
2635 curr->sched_class->task_tick(rq, curr, 0);
2636 raw_spin_unlock(&rq->lock);
2638 perf_event_task_tick();
2641 rq->idle_balance = idle_cpu(cpu);
2642 trigger_load_balance(rq, cpu);
2646 notrace unsigned long get_parent_ip(unsigned long addr)
2648 if (in_lock_functions(addr)) {
2649 addr = CALLER_ADDR2;
2650 if (in_lock_functions(addr))
2651 addr = CALLER_ADDR3;
2656 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2657 defined(CONFIG_PREEMPT_TRACER))
2659 void __kprobes add_preempt_count(int val)
2661 #ifdef CONFIG_DEBUG_PREEMPT
2665 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2668 preempt_count() += val;
2669 #ifdef CONFIG_DEBUG_PREEMPT
2671 * Spinlock count overflowing soon?
2673 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2676 if (preempt_count() == val)
2677 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2679 EXPORT_SYMBOL(add_preempt_count);
2681 void __kprobes sub_preempt_count(int val)
2683 #ifdef CONFIG_DEBUG_PREEMPT
2687 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2690 * Is the spinlock portion underflowing?
2692 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2693 !(preempt_count() & PREEMPT_MASK)))
2697 if (preempt_count() == val)
2698 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2699 preempt_count() -= val;
2701 EXPORT_SYMBOL(sub_preempt_count);
2706 * Print scheduling while atomic bug:
2708 static noinline void __schedule_bug(struct task_struct *prev)
2710 if (oops_in_progress)
2713 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2714 prev->comm, prev->pid, preempt_count());
2716 debug_show_held_locks(prev);
2718 if (irqs_disabled())
2719 print_irqtrace_events(prev);
2721 add_taint(TAINT_WARN);
2725 * Various schedule()-time debugging checks and statistics:
2727 static inline void schedule_debug(struct task_struct *prev)
2730 * Test if we are atomic. Since do_exit() needs to call into
2731 * schedule() atomically, we ignore that path for now.
2732 * Otherwise, whine if we are scheduling when we should not be.
2734 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2735 __schedule_bug(prev);
2738 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2740 schedstat_inc(this_rq(), sched_count);
2743 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2745 if (prev->on_rq || rq->skip_clock_update < 0)
2746 update_rq_clock(rq);
2747 prev->sched_class->put_prev_task(rq, prev);
2751 * Pick up the highest-prio task:
2753 static inline struct task_struct *
2754 pick_next_task(struct rq *rq)
2756 const struct sched_class *class;
2757 struct task_struct *p;
2760 * Optimization: we know that if all tasks are in
2761 * the fair class we can call that function directly:
2763 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2764 p = fair_sched_class.pick_next_task(rq);
2769 for_each_class(class) {
2770 p = class->pick_next_task(rq);
2775 BUG(); /* the idle class will always have a runnable task */
2779 * __schedule() is the main scheduler function.
2781 * The main means of driving the scheduler and thus entering this function are:
2783 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2785 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2786 * paths. For example, see arch/x86/entry_64.S.
2788 * To drive preemption between tasks, the scheduler sets the flag in timer
2789 * interrupt handler scheduler_tick().
2791 * 3. Wakeups don't really cause entry into schedule(). They add a
2792 * task to the run-queue and that's it.
2794 * Now, if the new task added to the run-queue preempts the current
2795 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2796 * called on the nearest possible occasion:
2798 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2800 * - in syscall or exception context, at the next outmost
2801 * preempt_enable(). (this might be as soon as the wake_up()'s
2804 * - in IRQ context, return from interrupt-handler to
2805 * preemptible context
2807 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2810 * - cond_resched() call
2811 * - explicit schedule() call
2812 * - return from syscall or exception to user-space
2813 * - return from interrupt-handler to user-space
2815 static void __sched __schedule(void)
2817 struct task_struct *prev, *next;
2818 unsigned long *switch_count;
2824 cpu = smp_processor_id();
2826 rcu_note_context_switch(cpu);
2829 schedule_debug(prev);
2831 if (sched_feat(HRTICK))
2834 raw_spin_lock_irq(&rq->lock);
2836 switch_count = &prev->nivcsw;
2837 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2838 if (unlikely(signal_pending_state(prev->state, prev))) {
2839 prev->state = TASK_RUNNING;
2841 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2845 * If a worker went to sleep, notify and ask workqueue
2846 * whether it wants to wake up a task to maintain
2849 if (prev->flags & PF_WQ_WORKER) {
2850 struct task_struct *to_wakeup;
2852 to_wakeup = wq_worker_sleeping(prev, cpu);
2854 try_to_wake_up_local(to_wakeup);
2857 switch_count = &prev->nvcsw;
2860 pre_schedule(rq, prev);
2862 if (unlikely(!rq->nr_running))
2863 idle_balance(cpu, rq);
2865 put_prev_task(rq, prev);
2866 next = pick_next_task(rq);
2867 clear_tsk_need_resched(prev);
2868 rq->skip_clock_update = 0;
2870 if (likely(prev != next)) {
2875 context_switch(rq, prev, next); /* unlocks the rq */
2877 * The context switch have flipped the stack from under us
2878 * and restored the local variables which were saved when
2879 * this task called schedule() in the past. prev == current
2880 * is still correct, but it can be moved to another cpu/rq.
2882 cpu = smp_processor_id();
2885 raw_spin_unlock_irq(&rq->lock);
2889 sched_preempt_enable_no_resched();
2894 static inline void sched_submit_work(struct task_struct *tsk)
2896 if (!tsk->state || tsk_is_pi_blocked(tsk))
2899 * If we are going to sleep and we have plugged IO queued,
2900 * make sure to submit it to avoid deadlocks.
2902 if (blk_needs_flush_plug(tsk))
2903 blk_schedule_flush_plug(tsk);
2906 asmlinkage void __sched schedule(void)
2908 struct task_struct *tsk = current;
2910 sched_submit_work(tsk);
2913 EXPORT_SYMBOL(schedule);
2915 #ifdef CONFIG_CONTEXT_TRACKING
2916 asmlinkage void __sched schedule_user(void)
2919 * If we come here after a random call to set_need_resched(),
2920 * or we have been woken up remotely but the IPI has not yet arrived,
2921 * we haven't yet exited the RCU idle mode. Do it here manually until
2922 * we find a better solution.
2931 * schedule_preempt_disabled - called with preemption disabled
2933 * Returns with preemption disabled. Note: preempt_count must be 1
2935 void __sched schedule_preempt_disabled(void)
2937 sched_preempt_enable_no_resched();
2942 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2944 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2946 if (lock->owner != owner)
2950 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2951 * lock->owner still matches owner, if that fails, owner might
2952 * point to free()d memory, if it still matches, the rcu_read_lock()
2953 * ensures the memory stays valid.
2957 return owner->on_cpu;
2961 * Look out! "owner" is an entirely speculative pointer
2962 * access and not reliable.
2964 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2966 if (!sched_feat(OWNER_SPIN))
2970 while (owner_running(lock, owner)) {
2974 arch_mutex_cpu_relax();
2979 * We break out the loop above on need_resched() and when the
2980 * owner changed, which is a sign for heavy contention. Return
2981 * success only when lock->owner is NULL.
2983 return lock->owner == NULL;
2987 #ifdef CONFIG_PREEMPT
2989 * this is the entry point to schedule() from in-kernel preemption
2990 * off of preempt_enable. Kernel preemptions off return from interrupt
2991 * occur there and call schedule directly.
2993 asmlinkage void __sched notrace preempt_schedule(void)
2995 struct thread_info *ti = current_thread_info();
2998 * If there is a non-zero preempt_count or interrupts are disabled,
2999 * we do not want to preempt the current task. Just return..
3001 if (likely(ti->preempt_count || irqs_disabled()))
3005 add_preempt_count_notrace(PREEMPT_ACTIVE);
3007 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3010 * Check again in case we missed a preemption opportunity
3011 * between schedule and now.
3014 } while (need_resched());
3016 EXPORT_SYMBOL(preempt_schedule);
3019 * this is the entry point to schedule() from kernel preemption
3020 * off of irq context.
3021 * Note, that this is called and return with irqs disabled. This will
3022 * protect us against recursive calling from irq.
3024 asmlinkage void __sched preempt_schedule_irq(void)
3026 struct thread_info *ti = current_thread_info();
3028 /* Catch callers which need to be fixed */
3029 BUG_ON(ti->preempt_count || !irqs_disabled());
3033 add_preempt_count(PREEMPT_ACTIVE);
3036 local_irq_disable();
3037 sub_preempt_count(PREEMPT_ACTIVE);
3040 * Check again in case we missed a preemption opportunity
3041 * between schedule and now.
3044 } while (need_resched());
3047 #endif /* CONFIG_PREEMPT */
3049 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3052 return try_to_wake_up(curr->private, mode, wake_flags);
3054 EXPORT_SYMBOL(default_wake_function);
3057 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3058 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3059 * number) then we wake all the non-exclusive tasks and one exclusive task.
3061 * There are circumstances in which we can try to wake a task which has already
3062 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3063 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3065 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3066 int nr_exclusive, int wake_flags, void *key)
3068 wait_queue_t *curr, *next;
3070 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3071 unsigned flags = curr->flags;
3073 if (curr->func(curr, mode, wake_flags, key) &&
3074 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3080 * __wake_up - wake up threads blocked on a waitqueue.
3082 * @mode: which threads
3083 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3084 * @key: is directly passed to the wakeup function
3086 * It may be assumed that this function implies a write memory barrier before
3087 * changing the task state if and only if any tasks are woken up.
3089 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3090 int nr_exclusive, void *key)
3092 unsigned long flags;
3094 spin_lock_irqsave(&q->lock, flags);
3095 __wake_up_common(q, mode, nr_exclusive, 0, key);
3096 spin_unlock_irqrestore(&q->lock, flags);
3098 EXPORT_SYMBOL(__wake_up);
3101 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3103 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3105 __wake_up_common(q, mode, nr, 0, NULL);
3107 EXPORT_SYMBOL_GPL(__wake_up_locked);
3109 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3111 __wake_up_common(q, mode, 1, 0, key);
3113 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3116 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3118 * @mode: which threads
3119 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3120 * @key: opaque value to be passed to wakeup targets
3122 * The sync wakeup differs that the waker knows that it will schedule
3123 * away soon, so while the target thread will be woken up, it will not
3124 * be migrated to another CPU - ie. the two threads are 'synchronized'
3125 * with each other. This can prevent needless bouncing between CPUs.
3127 * On UP it can prevent extra preemption.
3129 * It may be assumed that this function implies a write memory barrier before
3130 * changing the task state if and only if any tasks are woken up.
3132 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3133 int nr_exclusive, void *key)
3135 unsigned long flags;
3136 int wake_flags = WF_SYNC;
3141 if (unlikely(!nr_exclusive))
3144 spin_lock_irqsave(&q->lock, flags);
3145 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3146 spin_unlock_irqrestore(&q->lock, flags);
3148 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3151 * __wake_up_sync - see __wake_up_sync_key()
3153 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3155 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3157 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3160 * complete: - signals a single thread waiting on this completion
3161 * @x: holds the state of this particular completion
3163 * This will wake up a single thread waiting on this completion. Threads will be
3164 * awakened in the same order in which they were queued.
3166 * See also complete_all(), wait_for_completion() and related routines.
3168 * It may be assumed that this function implies a write memory barrier before
3169 * changing the task state if and only if any tasks are woken up.
3171 void complete(struct completion *x)
3173 unsigned long flags;
3175 spin_lock_irqsave(&x->wait.lock, flags);
3177 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3178 spin_unlock_irqrestore(&x->wait.lock, flags);
3180 EXPORT_SYMBOL(complete);
3183 * complete_all: - signals all threads waiting on this completion
3184 * @x: holds the state of this particular completion
3186 * This will wake up all threads waiting on this particular completion event.
3188 * It may be assumed that this function implies a write memory barrier before
3189 * changing the task state if and only if any tasks are woken up.
3191 void complete_all(struct completion *x)
3193 unsigned long flags;
3195 spin_lock_irqsave(&x->wait.lock, flags);
3196 x->done += UINT_MAX/2;
3197 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3198 spin_unlock_irqrestore(&x->wait.lock, flags);
3200 EXPORT_SYMBOL(complete_all);
3202 static inline long __sched
3203 do_wait_for_common(struct completion *x, long timeout, int state)
3206 DECLARE_WAITQUEUE(wait, current);
3208 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3210 if (signal_pending_state(state, current)) {
3211 timeout = -ERESTARTSYS;
3214 __set_current_state(state);
3215 spin_unlock_irq(&x->wait.lock);
3216 timeout = schedule_timeout(timeout);
3217 spin_lock_irq(&x->wait.lock);
3218 } while (!x->done && timeout);
3219 __remove_wait_queue(&x->wait, &wait);
3224 return timeout ?: 1;
3228 wait_for_common(struct completion *x, long timeout, int state)
3232 spin_lock_irq(&x->wait.lock);
3233 timeout = do_wait_for_common(x, timeout, state);
3234 spin_unlock_irq(&x->wait.lock);
3239 * wait_for_completion: - waits for completion of a task
3240 * @x: holds the state of this particular completion
3242 * This waits to be signaled for completion of a specific task. It is NOT
3243 * interruptible and there is no timeout.
3245 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3246 * and interrupt capability. Also see complete().
3248 void __sched wait_for_completion(struct completion *x)
3250 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3252 EXPORT_SYMBOL(wait_for_completion);
3255 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3256 * @x: holds the state of this particular completion
3257 * @timeout: timeout value in jiffies
3259 * This waits for either a completion of a specific task to be signaled or for a
3260 * specified timeout to expire. The timeout is in jiffies. It is not
3263 * The return value is 0 if timed out, and positive (at least 1, or number of
3264 * jiffies left till timeout) if completed.
3266 unsigned long __sched
3267 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3269 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3271 EXPORT_SYMBOL(wait_for_completion_timeout);
3274 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3275 * @x: holds the state of this particular completion
3277 * This waits for completion of a specific task to be signaled. It is
3280 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3282 int __sched wait_for_completion_interruptible(struct completion *x)
3284 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3285 if (t == -ERESTARTSYS)
3289 EXPORT_SYMBOL(wait_for_completion_interruptible);
3292 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3293 * @x: holds the state of this particular completion
3294 * @timeout: timeout value in jiffies
3296 * This waits for either a completion of a specific task to be signaled or for a
3297 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3299 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3300 * positive (at least 1, or number of jiffies left till timeout) if completed.
3303 wait_for_completion_interruptible_timeout(struct completion *x,
3304 unsigned long timeout)
3306 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3308 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3311 * wait_for_completion_killable: - waits for completion of a task (killable)
3312 * @x: holds the state of this particular completion
3314 * This waits to be signaled for completion of a specific task. It can be
3315 * interrupted by a kill signal.
3317 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3319 int __sched wait_for_completion_killable(struct completion *x)
3321 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3322 if (t == -ERESTARTSYS)
3326 EXPORT_SYMBOL(wait_for_completion_killable);
3329 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3330 * @x: holds the state of this particular completion
3331 * @timeout: timeout value in jiffies
3333 * This waits for either a completion of a specific task to be
3334 * signaled or for a specified timeout to expire. It can be
3335 * interrupted by a kill signal. The timeout is in jiffies.
3337 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3338 * positive (at least 1, or number of jiffies left till timeout) if completed.
3341 wait_for_completion_killable_timeout(struct completion *x,
3342 unsigned long timeout)
3344 return wait_for_common(x, timeout, TASK_KILLABLE);
3346 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3349 * try_wait_for_completion - try to decrement a completion without blocking
3350 * @x: completion structure
3352 * Returns: 0 if a decrement cannot be done without blocking
3353 * 1 if a decrement succeeded.
3355 * If a completion is being used as a counting completion,
3356 * attempt to decrement the counter without blocking. This
3357 * enables us to avoid waiting if the resource the completion
3358 * is protecting is not available.
3360 bool try_wait_for_completion(struct completion *x)
3362 unsigned long flags;
3365 spin_lock_irqsave(&x->wait.lock, flags);
3370 spin_unlock_irqrestore(&x->wait.lock, flags);
3373 EXPORT_SYMBOL(try_wait_for_completion);
3376 * completion_done - Test to see if a completion has any waiters
3377 * @x: completion structure
3379 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3380 * 1 if there are no waiters.
3383 bool completion_done(struct completion *x)
3385 unsigned long flags;
3388 spin_lock_irqsave(&x->wait.lock, flags);
3391 spin_unlock_irqrestore(&x->wait.lock, flags);
3394 EXPORT_SYMBOL(completion_done);
3397 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3399 unsigned long flags;
3402 init_waitqueue_entry(&wait, current);
3404 __set_current_state(state);
3406 spin_lock_irqsave(&q->lock, flags);
3407 __add_wait_queue(q, &wait);
3408 spin_unlock(&q->lock);
3409 timeout = schedule_timeout(timeout);
3410 spin_lock_irq(&q->lock);
3411 __remove_wait_queue(q, &wait);
3412 spin_unlock_irqrestore(&q->lock, flags);
3417 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3419 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3421 EXPORT_SYMBOL(interruptible_sleep_on);
3424 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3426 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3428 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3430 void __sched sleep_on(wait_queue_head_t *q)
3432 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3434 EXPORT_SYMBOL(sleep_on);
3436 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3438 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3440 EXPORT_SYMBOL(sleep_on_timeout);
3442 #ifdef CONFIG_RT_MUTEXES
3445 * rt_mutex_setprio - set the current priority of a task
3447 * @prio: prio value (kernel-internal form)
3449 * This function changes the 'effective' priority of a task. It does
3450 * not touch ->normal_prio like __setscheduler().
3452 * Used by the rt_mutex code to implement priority inheritance logic.
3454 void rt_mutex_setprio(struct task_struct *p, int prio)
3456 int oldprio, on_rq, running;
3458 const struct sched_class *prev_class;
3460 BUG_ON(prio < 0 || prio > MAX_PRIO);
3462 rq = __task_rq_lock(p);
3465 * Idle task boosting is a nono in general. There is one
3466 * exception, when PREEMPT_RT and NOHZ is active:
3468 * The idle task calls get_next_timer_interrupt() and holds
3469 * the timer wheel base->lock on the CPU and another CPU wants
3470 * to access the timer (probably to cancel it). We can safely
3471 * ignore the boosting request, as the idle CPU runs this code
3472 * with interrupts disabled and will complete the lock
3473 * protected section without being interrupted. So there is no
3474 * real need to boost.
3476 if (unlikely(p == rq->idle)) {
3477 WARN_ON(p != rq->curr);
3478 WARN_ON(p->pi_blocked_on);
3482 trace_sched_pi_setprio(p, prio);
3484 prev_class = p->sched_class;
3486 running = task_current(rq, p);
3488 dequeue_task(rq, p, 0);
3490 p->sched_class->put_prev_task(rq, p);
3493 p->sched_class = &rt_sched_class;
3495 p->sched_class = &fair_sched_class;
3500 p->sched_class->set_curr_task(rq);
3502 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3504 check_class_changed(rq, p, prev_class, oldprio);
3506 __task_rq_unlock(rq);
3509 void set_user_nice(struct task_struct *p, long nice)
3511 int old_prio, delta, on_rq;
3512 unsigned long flags;
3515 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3518 * We have to be careful, if called from sys_setpriority(),
3519 * the task might be in the middle of scheduling on another CPU.
3521 rq = task_rq_lock(p, &flags);
3523 * The RT priorities are set via sched_setscheduler(), but we still
3524 * allow the 'normal' nice value to be set - but as expected
3525 * it wont have any effect on scheduling until the task is
3526 * SCHED_FIFO/SCHED_RR:
3528 if (task_has_rt_policy(p)) {
3529 p->static_prio = NICE_TO_PRIO(nice);
3534 dequeue_task(rq, p, 0);
3536 p->static_prio = NICE_TO_PRIO(nice);
3539 p->prio = effective_prio(p);
3540 delta = p->prio - old_prio;
3543 enqueue_task(rq, p, 0);
3545 * If the task increased its priority or is running and
3546 * lowered its priority, then reschedule its CPU:
3548 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3549 resched_task(rq->curr);
3552 task_rq_unlock(rq, p, &flags);
3554 EXPORT_SYMBOL(set_user_nice);
3557 * can_nice - check if a task can reduce its nice value
3561 int can_nice(const struct task_struct *p, const int nice)
3563 /* convert nice value [19,-20] to rlimit style value [1,40] */
3564 int nice_rlim = 20 - nice;
3566 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3567 capable(CAP_SYS_NICE));
3570 #ifdef __ARCH_WANT_SYS_NICE
3573 * sys_nice - change the priority of the current process.
3574 * @increment: priority increment
3576 * sys_setpriority is a more generic, but much slower function that
3577 * does similar things.
3579 SYSCALL_DEFINE1(nice, int, increment)
3584 * Setpriority might change our priority at the same moment.
3585 * We don't have to worry. Conceptually one call occurs first
3586 * and we have a single winner.
3588 if (increment < -40)
3593 nice = TASK_NICE(current) + increment;
3599 if (increment < 0 && !can_nice(current, nice))
3602 retval = security_task_setnice(current, nice);
3606 set_user_nice(current, nice);
3613 * task_prio - return the priority value of a given task.
3614 * @p: the task in question.
3616 * This is the priority value as seen by users in /proc.
3617 * RT tasks are offset by -200. Normal tasks are centered
3618 * around 0, value goes from -16 to +15.
3620 int task_prio(const struct task_struct *p)
3622 return p->prio - MAX_RT_PRIO;
3626 * task_nice - return the nice value of a given task.
3627 * @p: the task in question.
3629 int task_nice(const struct task_struct *p)
3631 return TASK_NICE(p);
3633 EXPORT_SYMBOL(task_nice);
3636 * idle_cpu - is a given cpu idle currently?
3637 * @cpu: the processor in question.
3639 int idle_cpu(int cpu)
3641 struct rq *rq = cpu_rq(cpu);
3643 if (rq->curr != rq->idle)
3650 if (!llist_empty(&rq->wake_list))
3658 * idle_task - return the idle task for a given cpu.
3659 * @cpu: the processor in question.
3661 struct task_struct *idle_task(int cpu)
3663 return cpu_rq(cpu)->idle;
3667 * find_process_by_pid - find a process with a matching PID value.
3668 * @pid: the pid in question.
3670 static struct task_struct *find_process_by_pid(pid_t pid)
3672 return pid ? find_task_by_vpid(pid) : current;
3675 /* Actually do priority change: must hold rq lock. */
3677 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3680 p->rt_priority = prio;
3681 p->normal_prio = normal_prio(p);
3682 /* we are holding p->pi_lock already */
3683 p->prio = rt_mutex_getprio(p);
3684 if (rt_prio(p->prio))
3685 p->sched_class = &rt_sched_class;
3687 p->sched_class = &fair_sched_class;
3692 * check the target process has a UID that matches the current process's
3694 static bool check_same_owner(struct task_struct *p)
3696 const struct cred *cred = current_cred(), *pcred;
3700 pcred = __task_cred(p);
3701 match = (uid_eq(cred->euid, pcred->euid) ||
3702 uid_eq(cred->euid, pcred->uid));
3707 static int __sched_setscheduler(struct task_struct *p, int policy,
3708 const struct sched_param *param, bool user)
3710 int retval, oldprio, oldpolicy = -1, on_rq, running;
3711 unsigned long flags;
3712 const struct sched_class *prev_class;
3716 /* may grab non-irq protected spin_locks */
3717 BUG_ON(in_interrupt());
3719 /* double check policy once rq lock held */
3721 reset_on_fork = p->sched_reset_on_fork;
3722 policy = oldpolicy = p->policy;
3724 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3725 policy &= ~SCHED_RESET_ON_FORK;
3727 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3728 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3729 policy != SCHED_IDLE)
3734 * Valid priorities for SCHED_FIFO and SCHED_RR are
3735 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3736 * SCHED_BATCH and SCHED_IDLE is 0.
3738 if (param->sched_priority < 0 ||
3739 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3740 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3742 if (rt_policy(policy) != (param->sched_priority != 0))
3746 * Allow unprivileged RT tasks to decrease priority:
3748 if (user && !capable(CAP_SYS_NICE)) {
3749 if (rt_policy(policy)) {
3750 unsigned long rlim_rtprio =
3751 task_rlimit(p, RLIMIT_RTPRIO);
3753 /* can't set/change the rt policy */
3754 if (policy != p->policy && !rlim_rtprio)
3757 /* can't increase priority */
3758 if (param->sched_priority > p->rt_priority &&
3759 param->sched_priority > rlim_rtprio)
3764 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3765 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3767 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3768 if (!can_nice(p, TASK_NICE(p)))
3772 /* can't change other user's priorities */
3773 if (!check_same_owner(p))
3776 /* Normal users shall not reset the sched_reset_on_fork flag */
3777 if (p->sched_reset_on_fork && !reset_on_fork)
3782 retval = security_task_setscheduler(p);
3788 * make sure no PI-waiters arrive (or leave) while we are
3789 * changing the priority of the task:
3791 * To be able to change p->policy safely, the appropriate
3792 * runqueue lock must be held.
3794 rq = task_rq_lock(p, &flags);
3797 * Changing the policy of the stop threads its a very bad idea
3799 if (p == rq->stop) {
3800 task_rq_unlock(rq, p, &flags);
3805 * If not changing anything there's no need to proceed further:
3807 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3808 param->sched_priority == p->rt_priority))) {
3809 task_rq_unlock(rq, p, &flags);
3813 #ifdef CONFIG_RT_GROUP_SCHED
3816 * Do not allow realtime tasks into groups that have no runtime
3819 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3820 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3821 !task_group_is_autogroup(task_group(p))) {
3822 task_rq_unlock(rq, p, &flags);
3828 /* recheck policy now with rq lock held */
3829 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3830 policy = oldpolicy = -1;
3831 task_rq_unlock(rq, p, &flags);
3835 running = task_current(rq, p);
3837 dequeue_task(rq, p, 0);
3839 p->sched_class->put_prev_task(rq, p);
3841 p->sched_reset_on_fork = reset_on_fork;
3844 prev_class = p->sched_class;
3845 __setscheduler(rq, p, policy, param->sched_priority);
3848 p->sched_class->set_curr_task(rq);
3850 enqueue_task(rq, p, 0);
3852 check_class_changed(rq, p, prev_class, oldprio);
3853 task_rq_unlock(rq, p, &flags);
3855 rt_mutex_adjust_pi(p);
3861 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3862 * @p: the task in question.
3863 * @policy: new policy.
3864 * @param: structure containing the new RT priority.
3866 * NOTE that the task may be already dead.
3868 int sched_setscheduler(struct task_struct *p, int policy,
3869 const struct sched_param *param)
3871 return __sched_setscheduler(p, policy, param, true);
3873 EXPORT_SYMBOL_GPL(sched_setscheduler);
3876 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3877 * @p: the task in question.
3878 * @policy: new policy.
3879 * @param: structure containing the new RT priority.
3881 * Just like sched_setscheduler, only don't bother checking if the
3882 * current context has permission. For example, this is needed in
3883 * stop_machine(): we create temporary high priority worker threads,
3884 * but our caller might not have that capability.
3886 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3887 const struct sched_param *param)
3889 return __sched_setscheduler(p, policy, param, false);
3893 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3895 struct sched_param lparam;
3896 struct task_struct *p;
3899 if (!param || pid < 0)
3901 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3906 p = find_process_by_pid(pid);
3908 retval = sched_setscheduler(p, policy, &lparam);
3915 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3916 * @pid: the pid in question.
3917 * @policy: new policy.
3918 * @param: structure containing the new RT priority.
3920 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3921 struct sched_param __user *, param)
3923 /* negative values for policy are not valid */
3927 return do_sched_setscheduler(pid, policy, param);
3931 * sys_sched_setparam - set/change the RT priority of a thread
3932 * @pid: the pid in question.
3933 * @param: structure containing the new RT priority.
3935 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3937 return do_sched_setscheduler(pid, -1, param);
3941 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3942 * @pid: the pid in question.
3944 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3946 struct task_struct *p;
3954 p = find_process_by_pid(pid);
3956 retval = security_task_getscheduler(p);
3959 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3966 * sys_sched_getparam - get the RT priority of a thread
3967 * @pid: the pid in question.
3968 * @param: structure containing the RT priority.
3970 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3972 struct sched_param lp;
3973 struct task_struct *p;
3976 if (!param || pid < 0)
3980 p = find_process_by_pid(pid);
3985 retval = security_task_getscheduler(p);
3989 lp.sched_priority = p->rt_priority;
3993 * This one might sleep, we cannot do it with a spinlock held ...
3995 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4004 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4006 cpumask_var_t cpus_allowed, new_mask;
4007 struct task_struct *p;
4013 p = find_process_by_pid(pid);
4020 /* Prevent p going away */
4024 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4028 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4030 goto out_free_cpus_allowed;
4033 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4036 retval = security_task_setscheduler(p);
4040 cpuset_cpus_allowed(p, cpus_allowed);
4041 cpumask_and(new_mask, in_mask, cpus_allowed);
4043 retval = set_cpus_allowed_ptr(p, new_mask);
4046 cpuset_cpus_allowed(p, cpus_allowed);
4047 if (!cpumask_subset(new_mask, cpus_allowed)) {
4049 * We must have raced with a concurrent cpuset
4050 * update. Just reset the cpus_allowed to the
4051 * cpuset's cpus_allowed
4053 cpumask_copy(new_mask, cpus_allowed);
4058 free_cpumask_var(new_mask);
4059 out_free_cpus_allowed:
4060 free_cpumask_var(cpus_allowed);
4067 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4068 struct cpumask *new_mask)
4070 if (len < cpumask_size())
4071 cpumask_clear(new_mask);
4072 else if (len > cpumask_size())
4073 len = cpumask_size();
4075 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4079 * sys_sched_setaffinity - set the cpu affinity of a process
4080 * @pid: pid of the process
4081 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4082 * @user_mask_ptr: user-space pointer to the new cpu mask
4084 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4085 unsigned long __user *, user_mask_ptr)
4087 cpumask_var_t new_mask;
4090 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4093 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4095 retval = sched_setaffinity(pid, new_mask);
4096 free_cpumask_var(new_mask);
4100 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4102 struct task_struct *p;
4103 unsigned long flags;
4110 p = find_process_by_pid(pid);
4114 retval = security_task_getscheduler(p);
4118 raw_spin_lock_irqsave(&p->pi_lock, flags);
4119 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4120 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4130 * sys_sched_getaffinity - get the cpu affinity of a process
4131 * @pid: pid of the process
4132 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4133 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4135 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4136 unsigned long __user *, user_mask_ptr)
4141 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4143 if (len & (sizeof(unsigned long)-1))
4146 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4149 ret = sched_getaffinity(pid, mask);
4151 size_t retlen = min_t(size_t, len, cpumask_size());
4153 if (copy_to_user(user_mask_ptr, mask, retlen))
4158 free_cpumask_var(mask);
4164 * sys_sched_yield - yield the current processor to other threads.
4166 * This function yields the current CPU to other tasks. If there are no
4167 * other threads running on this CPU then this function will return.
4169 SYSCALL_DEFINE0(sched_yield)
4171 struct rq *rq = this_rq_lock();
4173 schedstat_inc(rq, yld_count);
4174 current->sched_class->yield_task(rq);
4177 * Since we are going to call schedule() anyway, there's
4178 * no need to preempt or enable interrupts:
4180 __release(rq->lock);
4181 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4182 do_raw_spin_unlock(&rq->lock);
4183 sched_preempt_enable_no_resched();
4190 static inline int should_resched(void)
4192 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4195 static void __cond_resched(void)
4197 add_preempt_count(PREEMPT_ACTIVE);
4199 sub_preempt_count(PREEMPT_ACTIVE);
4202 int __sched _cond_resched(void)
4204 if (should_resched()) {
4210 EXPORT_SYMBOL(_cond_resched);
4213 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4214 * call schedule, and on return reacquire the lock.
4216 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4217 * operations here to prevent schedule() from being called twice (once via
4218 * spin_unlock(), once by hand).
4220 int __cond_resched_lock(spinlock_t *lock)
4222 int resched = should_resched();
4225 lockdep_assert_held(lock);
4227 if (spin_needbreak(lock) || resched) {
4238 EXPORT_SYMBOL(__cond_resched_lock);
4240 int __sched __cond_resched_softirq(void)
4242 BUG_ON(!in_softirq());
4244 if (should_resched()) {
4252 EXPORT_SYMBOL(__cond_resched_softirq);
4255 * yield - yield the current processor to other threads.
4257 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4259 * The scheduler is at all times free to pick the calling task as the most
4260 * eligible task to run, if removing the yield() call from your code breaks
4261 * it, its already broken.
4263 * Typical broken usage is:
4268 * where one assumes that yield() will let 'the other' process run that will
4269 * make event true. If the current task is a SCHED_FIFO task that will never
4270 * happen. Never use yield() as a progress guarantee!!
4272 * If you want to use yield() to wait for something, use wait_event().
4273 * If you want to use yield() to be 'nice' for others, use cond_resched().
4274 * If you still want to use yield(), do not!
4276 void __sched yield(void)
4278 set_current_state(TASK_RUNNING);
4281 EXPORT_SYMBOL(yield);
4284 * yield_to - yield the current processor to another thread in
4285 * your thread group, or accelerate that thread toward the
4286 * processor it's on.
4288 * @preempt: whether task preemption is allowed or not
4290 * It's the caller's job to ensure that the target task struct
4291 * can't go away on us before we can do any checks.
4293 * Returns true if we indeed boosted the target task.
4295 bool __sched yield_to(struct task_struct *p, bool preempt)
4297 struct task_struct *curr = current;
4298 struct rq *rq, *p_rq;
4299 unsigned long flags;
4302 local_irq_save(flags);
4307 double_rq_lock(rq, p_rq);
4308 while (task_rq(p) != p_rq) {
4309 double_rq_unlock(rq, p_rq);
4313 if (!curr->sched_class->yield_to_task)
4316 if (curr->sched_class != p->sched_class)
4319 if (task_running(p_rq, p) || p->state)
4322 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4324 schedstat_inc(rq, yld_count);
4326 * Make p's CPU reschedule; pick_next_entity takes care of
4329 if (preempt && rq != p_rq)
4330 resched_task(p_rq->curr);
4334 double_rq_unlock(rq, p_rq);
4335 local_irq_restore(flags);
4342 EXPORT_SYMBOL_GPL(yield_to);
4345 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4346 * that process accounting knows that this is a task in IO wait state.
4348 void __sched io_schedule(void)
4350 struct rq *rq = raw_rq();
4352 delayacct_blkio_start();
4353 atomic_inc(&rq->nr_iowait);
4354 blk_flush_plug(current);
4355 current->in_iowait = 1;
4357 current->in_iowait = 0;
4358 atomic_dec(&rq->nr_iowait);
4359 delayacct_blkio_end();
4361 EXPORT_SYMBOL(io_schedule);
4363 long __sched io_schedule_timeout(long timeout)
4365 struct rq *rq = raw_rq();
4368 delayacct_blkio_start();
4369 atomic_inc(&rq->nr_iowait);
4370 blk_flush_plug(current);
4371 current->in_iowait = 1;
4372 ret = schedule_timeout(timeout);
4373 current->in_iowait = 0;
4374 atomic_dec(&rq->nr_iowait);
4375 delayacct_blkio_end();
4380 * sys_sched_get_priority_max - return maximum RT priority.
4381 * @policy: scheduling class.
4383 * this syscall returns the maximum rt_priority that can be used
4384 * by a given scheduling class.
4386 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4393 ret = MAX_USER_RT_PRIO-1;
4405 * sys_sched_get_priority_min - return minimum RT priority.
4406 * @policy: scheduling class.
4408 * this syscall returns the minimum rt_priority that can be used
4409 * by a given scheduling class.
4411 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4429 * sys_sched_rr_get_interval - return the default timeslice of a process.
4430 * @pid: pid of the process.
4431 * @interval: userspace pointer to the timeslice value.
4433 * this syscall writes the default timeslice value of a given process
4434 * into the user-space timespec buffer. A value of '0' means infinity.
4436 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4437 struct timespec __user *, interval)
4439 struct task_struct *p;
4440 unsigned int time_slice;
4441 unsigned long flags;
4451 p = find_process_by_pid(pid);
4455 retval = security_task_getscheduler(p);
4459 rq = task_rq_lock(p, &flags);
4460 time_slice = p->sched_class->get_rr_interval(rq, p);
4461 task_rq_unlock(rq, p, &flags);
4464 jiffies_to_timespec(time_slice, &t);
4465 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4473 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4475 void sched_show_task(struct task_struct *p)
4477 unsigned long free = 0;
4481 state = p->state ? __ffs(p->state) + 1 : 0;
4482 printk(KERN_INFO "%-15.15s %c", p->comm,
4483 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4484 #if BITS_PER_LONG == 32
4485 if (state == TASK_RUNNING)
4486 printk(KERN_CONT " running ");
4488 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4490 if (state == TASK_RUNNING)
4491 printk(KERN_CONT " running task ");
4493 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4495 #ifdef CONFIG_DEBUG_STACK_USAGE
4496 free = stack_not_used(p);
4499 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4501 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4502 task_pid_nr(p), ppid,
4503 (unsigned long)task_thread_info(p)->flags);
4505 show_stack(p, NULL);
4508 void show_state_filter(unsigned long state_filter)
4510 struct task_struct *g, *p;
4512 #if BITS_PER_LONG == 32
4514 " task PC stack pid father\n");
4517 " task PC stack pid father\n");
4520 do_each_thread(g, p) {
4522 * reset the NMI-timeout, listing all files on a slow
4523 * console might take a lot of time:
4525 touch_nmi_watchdog();
4526 if (!state_filter || (p->state & state_filter))
4528 } while_each_thread(g, p);
4530 touch_all_softlockup_watchdogs();
4532 #ifdef CONFIG_SCHED_DEBUG
4533 sysrq_sched_debug_show();
4537 * Only show locks if all tasks are dumped:
4540 debug_show_all_locks();
4543 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4545 idle->sched_class = &idle_sched_class;
4549 * init_idle - set up an idle thread for a given CPU
4550 * @idle: task in question
4551 * @cpu: cpu the idle task belongs to
4553 * NOTE: this function does not set the idle thread's NEED_RESCHED
4554 * flag, to make booting more robust.
4556 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4558 struct rq *rq = cpu_rq(cpu);
4559 unsigned long flags;
4561 raw_spin_lock_irqsave(&rq->lock, flags);
4564 idle->state = TASK_RUNNING;
4565 idle->se.exec_start = sched_clock();
4567 do_set_cpus_allowed(idle, cpumask_of(cpu));
4569 * We're having a chicken and egg problem, even though we are
4570 * holding rq->lock, the cpu isn't yet set to this cpu so the
4571 * lockdep check in task_group() will fail.
4573 * Similar case to sched_fork(). / Alternatively we could
4574 * use task_rq_lock() here and obtain the other rq->lock.
4579 __set_task_cpu(idle, cpu);
4582 rq->curr = rq->idle = idle;
4583 #if defined(CONFIG_SMP)
4586 raw_spin_unlock_irqrestore(&rq->lock, flags);
4588 /* Set the preempt count _outside_ the spinlocks! */
4589 task_thread_info(idle)->preempt_count = 0;
4592 * The idle tasks have their own, simple scheduling class:
4594 idle->sched_class = &idle_sched_class;
4595 ftrace_graph_init_idle_task(idle, cpu);
4596 #if defined(CONFIG_SMP)
4597 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4602 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4604 if (p->sched_class && p->sched_class->set_cpus_allowed)
4605 p->sched_class->set_cpus_allowed(p, new_mask);
4607 cpumask_copy(&p->cpus_allowed, new_mask);
4608 p->nr_cpus_allowed = cpumask_weight(new_mask);
4612 * This is how migration works:
4614 * 1) we invoke migration_cpu_stop() on the target CPU using
4616 * 2) stopper starts to run (implicitly forcing the migrated thread
4618 * 3) it checks whether the migrated task is still in the wrong runqueue.
4619 * 4) if it's in the wrong runqueue then the migration thread removes
4620 * it and puts it into the right queue.
4621 * 5) stopper completes and stop_one_cpu() returns and the migration
4626 * Change a given task's CPU affinity. Migrate the thread to a
4627 * proper CPU and schedule it away if the CPU it's executing on
4628 * is removed from the allowed bitmask.
4630 * NOTE: the caller must have a valid reference to the task, the
4631 * task must not exit() & deallocate itself prematurely. The
4632 * call is not atomic; no spinlocks may be held.
4634 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4636 unsigned long flags;
4638 unsigned int dest_cpu;
4641 rq = task_rq_lock(p, &flags);
4643 if (cpumask_equal(&p->cpus_allowed, new_mask))
4646 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4651 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4656 do_set_cpus_allowed(p, new_mask);
4658 /* Can the task run on the task's current CPU? If so, we're done */
4659 if (cpumask_test_cpu(task_cpu(p), new_mask))
4662 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4664 struct migration_arg arg = { p, dest_cpu };
4665 /* Need help from migration thread: drop lock and wait. */
4666 task_rq_unlock(rq, p, &flags);
4667 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4668 tlb_migrate_finish(p->mm);
4672 task_rq_unlock(rq, p, &flags);
4676 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4679 * Move (not current) task off this cpu, onto dest cpu. We're doing
4680 * this because either it can't run here any more (set_cpus_allowed()
4681 * away from this CPU, or CPU going down), or because we're
4682 * attempting to rebalance this task on exec (sched_exec).
4684 * So we race with normal scheduler movements, but that's OK, as long
4685 * as the task is no longer on this CPU.
4687 * Returns non-zero if task was successfully migrated.
4689 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4691 struct rq *rq_dest, *rq_src;
4694 if (unlikely(!cpu_active(dest_cpu)))
4697 rq_src = cpu_rq(src_cpu);
4698 rq_dest = cpu_rq(dest_cpu);
4700 raw_spin_lock(&p->pi_lock);
4701 double_rq_lock(rq_src, rq_dest);
4702 /* Already moved. */
4703 if (task_cpu(p) != src_cpu)
4705 /* Affinity changed (again). */
4706 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4710 * If we're not on a rq, the next wake-up will ensure we're
4714 dequeue_task(rq_src, p, 0);
4715 set_task_cpu(p, dest_cpu);
4716 enqueue_task(rq_dest, p, 0);
4717 check_preempt_curr(rq_dest, p, 0);
4722 double_rq_unlock(rq_src, rq_dest);
4723 raw_spin_unlock(&p->pi_lock);
4728 * migration_cpu_stop - this will be executed by a highprio stopper thread
4729 * and performs thread migration by bumping thread off CPU then
4730 * 'pushing' onto another runqueue.
4732 static int migration_cpu_stop(void *data)
4734 struct migration_arg *arg = data;
4737 * The original target cpu might have gone down and we might
4738 * be on another cpu but it doesn't matter.
4740 local_irq_disable();
4741 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4746 #ifdef CONFIG_HOTPLUG_CPU
4749 * Ensures that the idle task is using init_mm right before its cpu goes
4752 void idle_task_exit(void)
4754 struct mm_struct *mm = current->active_mm;
4756 BUG_ON(cpu_online(smp_processor_id()));
4759 switch_mm(mm, &init_mm, current);
4764 * Since this CPU is going 'away' for a while, fold any nr_active delta
4765 * we might have. Assumes we're called after migrate_tasks() so that the
4766 * nr_active count is stable.
4768 * Also see the comment "Global load-average calculations".
4770 static void calc_load_migrate(struct rq *rq)
4772 long delta = calc_load_fold_active(rq);
4774 atomic_long_add(delta, &calc_load_tasks);
4778 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4779 * try_to_wake_up()->select_task_rq().
4781 * Called with rq->lock held even though we'er in stop_machine() and
4782 * there's no concurrency possible, we hold the required locks anyway
4783 * because of lock validation efforts.
4785 static void migrate_tasks(unsigned int dead_cpu)
4787 struct rq *rq = cpu_rq(dead_cpu);
4788 struct task_struct *next, *stop = rq->stop;
4792 * Fudge the rq selection such that the below task selection loop
4793 * doesn't get stuck on the currently eligible stop task.
4795 * We're currently inside stop_machine() and the rq is either stuck
4796 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4797 * either way we should never end up calling schedule() until we're
4804 * There's this thread running, bail when that's the only
4807 if (rq->nr_running == 1)
4810 next = pick_next_task(rq);
4812 next->sched_class->put_prev_task(rq, next);
4814 /* Find suitable destination for @next, with force if needed. */
4815 dest_cpu = select_fallback_rq(dead_cpu, next);
4816 raw_spin_unlock(&rq->lock);
4818 __migrate_task(next, dead_cpu, dest_cpu);
4820 raw_spin_lock(&rq->lock);
4826 #endif /* CONFIG_HOTPLUG_CPU */
4828 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4830 static struct ctl_table sd_ctl_dir[] = {
4832 .procname = "sched_domain",
4838 static struct ctl_table sd_ctl_root[] = {
4840 .procname = "kernel",
4842 .child = sd_ctl_dir,
4847 static struct ctl_table *sd_alloc_ctl_entry(int n)
4849 struct ctl_table *entry =
4850 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4855 static void sd_free_ctl_entry(struct ctl_table **tablep)
4857 struct ctl_table *entry;
4860 * In the intermediate directories, both the child directory and
4861 * procname are dynamically allocated and could fail but the mode
4862 * will always be set. In the lowest directory the names are
4863 * static strings and all have proc handlers.
4865 for (entry = *tablep; entry->mode; entry++) {
4867 sd_free_ctl_entry(&entry->child);
4868 if (entry->proc_handler == NULL)
4869 kfree(entry->procname);
4876 static int min_load_idx = 0;
4877 static int max_load_idx = CPU_LOAD_IDX_MAX;
4880 set_table_entry(struct ctl_table *entry,
4881 const char *procname, void *data, int maxlen,
4882 umode_t mode, proc_handler *proc_handler,
4885 entry->procname = procname;
4887 entry->maxlen = maxlen;
4889 entry->proc_handler = proc_handler;
4892 entry->extra1 = &min_load_idx;
4893 entry->extra2 = &max_load_idx;
4897 static struct ctl_table *
4898 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4900 struct ctl_table *table = sd_alloc_ctl_entry(13);
4905 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4906 sizeof(long), 0644, proc_doulongvec_minmax, false);
4907 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4908 sizeof(long), 0644, proc_doulongvec_minmax, false);
4909 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4910 sizeof(int), 0644, proc_dointvec_minmax, true);
4911 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4912 sizeof(int), 0644, proc_dointvec_minmax, true);
4913 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4914 sizeof(int), 0644, proc_dointvec_minmax, true);
4915 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4916 sizeof(int), 0644, proc_dointvec_minmax, true);
4917 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4918 sizeof(int), 0644, proc_dointvec_minmax, true);
4919 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4920 sizeof(int), 0644, proc_dointvec_minmax, false);
4921 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4922 sizeof(int), 0644, proc_dointvec_minmax, false);
4923 set_table_entry(&table[9], "cache_nice_tries",
4924 &sd->cache_nice_tries,
4925 sizeof(int), 0644, proc_dointvec_minmax, false);
4926 set_table_entry(&table[10], "flags", &sd->flags,
4927 sizeof(int), 0644, proc_dointvec_minmax, false);
4928 set_table_entry(&table[11], "name", sd->name,
4929 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4930 /* &table[12] is terminator */
4935 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4937 struct ctl_table *entry, *table;
4938 struct sched_domain *sd;
4939 int domain_num = 0, i;
4942 for_each_domain(cpu, sd)
4944 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4949 for_each_domain(cpu, sd) {
4950 snprintf(buf, 32, "domain%d", i);
4951 entry->procname = kstrdup(buf, GFP_KERNEL);
4953 entry->child = sd_alloc_ctl_domain_table(sd);
4960 static struct ctl_table_header *sd_sysctl_header;
4961 static void register_sched_domain_sysctl(void)
4963 int i, cpu_num = num_possible_cpus();
4964 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4967 WARN_ON(sd_ctl_dir[0].child);
4968 sd_ctl_dir[0].child = entry;
4973 for_each_possible_cpu(i) {
4974 snprintf(buf, 32, "cpu%d", i);
4975 entry->procname = kstrdup(buf, GFP_KERNEL);
4977 entry->child = sd_alloc_ctl_cpu_table(i);
4981 WARN_ON(sd_sysctl_header);
4982 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4985 /* may be called multiple times per register */
4986 static void unregister_sched_domain_sysctl(void)
4988 if (sd_sysctl_header)
4989 unregister_sysctl_table(sd_sysctl_header);
4990 sd_sysctl_header = NULL;
4991 if (sd_ctl_dir[0].child)
4992 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4995 static void register_sched_domain_sysctl(void)
4998 static void unregister_sched_domain_sysctl(void)
5003 static void set_rq_online(struct rq *rq)
5006 const struct sched_class *class;
5008 cpumask_set_cpu(rq->cpu, rq->rd->online);
5011 for_each_class(class) {
5012 if (class->rq_online)
5013 class->rq_online(rq);
5018 static void set_rq_offline(struct rq *rq)
5021 const struct sched_class *class;
5023 for_each_class(class) {
5024 if (class->rq_offline)
5025 class->rq_offline(rq);
5028 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5034 * migration_call - callback that gets triggered when a CPU is added.
5035 * Here we can start up the necessary migration thread for the new CPU.
5037 static int __cpuinit
5038 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5040 int cpu = (long)hcpu;
5041 unsigned long flags;
5042 struct rq *rq = cpu_rq(cpu);
5044 switch (action & ~CPU_TASKS_FROZEN) {
5046 case CPU_UP_PREPARE:
5047 rq->calc_load_update = calc_load_update;
5051 /* Update our root-domain */
5052 raw_spin_lock_irqsave(&rq->lock, flags);
5054 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5058 raw_spin_unlock_irqrestore(&rq->lock, flags);
5061 #ifdef CONFIG_HOTPLUG_CPU
5063 sched_ttwu_pending();
5064 /* Update our root-domain */
5065 raw_spin_lock_irqsave(&rq->lock, flags);
5067 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5071 BUG_ON(rq->nr_running != 1); /* the migration thread */
5072 raw_spin_unlock_irqrestore(&rq->lock, flags);
5076 calc_load_migrate(rq);
5081 update_max_interval();
5087 * Register at high priority so that task migration (migrate_all_tasks)
5088 * happens before everything else. This has to be lower priority than
5089 * the notifier in the perf_event subsystem, though.
5091 static struct notifier_block __cpuinitdata migration_notifier = {
5092 .notifier_call = migration_call,
5093 .priority = CPU_PRI_MIGRATION,
5096 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5097 unsigned long action, void *hcpu)
5099 switch (action & ~CPU_TASKS_FROZEN) {
5101 case CPU_DOWN_FAILED:
5102 set_cpu_active((long)hcpu, true);
5109 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5110 unsigned long action, void *hcpu)
5112 switch (action & ~CPU_TASKS_FROZEN) {
5113 case CPU_DOWN_PREPARE:
5114 set_cpu_active((long)hcpu, false);
5121 static int __init migration_init(void)
5123 void *cpu = (void *)(long)smp_processor_id();
5126 /* Initialize migration for the boot CPU */
5127 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5128 BUG_ON(err == NOTIFY_BAD);
5129 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5130 register_cpu_notifier(&migration_notifier);
5132 /* Register cpu active notifiers */
5133 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5134 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5138 early_initcall(migration_init);
5143 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5145 #ifdef CONFIG_SCHED_DEBUG
5147 static __read_mostly int sched_debug_enabled;
5149 static int __init sched_debug_setup(char *str)
5151 sched_debug_enabled = 1;
5155 early_param("sched_debug", sched_debug_setup);
5157 static inline bool sched_debug(void)
5159 return sched_debug_enabled;
5162 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5163 struct cpumask *groupmask)
5165 struct sched_group *group = sd->groups;
5168 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5169 cpumask_clear(groupmask);
5171 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5173 if (!(sd->flags & SD_LOAD_BALANCE)) {
5174 printk("does not load-balance\n");
5176 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5181 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5183 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5184 printk(KERN_ERR "ERROR: domain->span does not contain "
5187 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5188 printk(KERN_ERR "ERROR: domain->groups does not contain"
5192 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5196 printk(KERN_ERR "ERROR: group is NULL\n");
5201 * Even though we initialize ->power to something semi-sane,
5202 * we leave power_orig unset. This allows us to detect if
5203 * domain iteration is still funny without causing /0 traps.
5205 if (!group->sgp->power_orig) {
5206 printk(KERN_CONT "\n");
5207 printk(KERN_ERR "ERROR: domain->cpu_power not "
5212 if (!cpumask_weight(sched_group_cpus(group))) {
5213 printk(KERN_CONT "\n");
5214 printk(KERN_ERR "ERROR: empty group\n");
5218 if (!(sd->flags & SD_OVERLAP) &&
5219 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5220 printk(KERN_CONT "\n");
5221 printk(KERN_ERR "ERROR: repeated CPUs\n");
5225 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5227 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5229 printk(KERN_CONT " %s", str);
5230 if (group->sgp->power != SCHED_POWER_SCALE) {
5231 printk(KERN_CONT " (cpu_power = %d)",
5235 group = group->next;
5236 } while (group != sd->groups);
5237 printk(KERN_CONT "\n");
5239 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5240 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5243 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5244 printk(KERN_ERR "ERROR: parent span is not a superset "
5245 "of domain->span\n");
5249 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5253 if (!sched_debug_enabled)
5257 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5261 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5264 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5272 #else /* !CONFIG_SCHED_DEBUG */
5273 # define sched_domain_debug(sd, cpu) do { } while (0)
5274 static inline bool sched_debug(void)
5278 #endif /* CONFIG_SCHED_DEBUG */
5280 static int sd_degenerate(struct sched_domain *sd)
5282 if (cpumask_weight(sched_domain_span(sd)) == 1)
5285 /* Following flags need at least 2 groups */
5286 if (sd->flags & (SD_LOAD_BALANCE |
5287 SD_BALANCE_NEWIDLE |
5291 SD_SHARE_PKG_RESOURCES)) {
5292 if (sd->groups != sd->groups->next)
5296 /* Following flags don't use groups */
5297 if (sd->flags & (SD_WAKE_AFFINE))
5304 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5306 unsigned long cflags = sd->flags, pflags = parent->flags;
5308 if (sd_degenerate(parent))
5311 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5314 /* Flags needing groups don't count if only 1 group in parent */
5315 if (parent->groups == parent->groups->next) {
5316 pflags &= ~(SD_LOAD_BALANCE |
5317 SD_BALANCE_NEWIDLE |
5321 SD_SHARE_PKG_RESOURCES);
5322 if (nr_node_ids == 1)
5323 pflags &= ~SD_SERIALIZE;
5325 if (~cflags & pflags)
5331 static void free_rootdomain(struct rcu_head *rcu)
5333 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5335 cpupri_cleanup(&rd->cpupri);
5336 free_cpumask_var(rd->rto_mask);
5337 free_cpumask_var(rd->online);
5338 free_cpumask_var(rd->span);
5342 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5344 struct root_domain *old_rd = NULL;
5345 unsigned long flags;
5347 raw_spin_lock_irqsave(&rq->lock, flags);
5352 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5355 cpumask_clear_cpu(rq->cpu, old_rd->span);
5358 * If we dont want to free the old_rt yet then
5359 * set old_rd to NULL to skip the freeing later
5362 if (!atomic_dec_and_test(&old_rd->refcount))
5366 atomic_inc(&rd->refcount);
5369 cpumask_set_cpu(rq->cpu, rd->span);
5370 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5373 raw_spin_unlock_irqrestore(&rq->lock, flags);
5376 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5379 static int init_rootdomain(struct root_domain *rd)
5381 memset(rd, 0, sizeof(*rd));
5383 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5385 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5387 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5390 if (cpupri_init(&rd->cpupri) != 0)
5395 free_cpumask_var(rd->rto_mask);
5397 free_cpumask_var(rd->online);
5399 free_cpumask_var(rd->span);
5405 * By default the system creates a single root-domain with all cpus as
5406 * members (mimicking the global state we have today).
5408 struct root_domain def_root_domain;
5410 static void init_defrootdomain(void)
5412 init_rootdomain(&def_root_domain);
5414 atomic_set(&def_root_domain.refcount, 1);
5417 static struct root_domain *alloc_rootdomain(void)
5419 struct root_domain *rd;
5421 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5425 if (init_rootdomain(rd) != 0) {
5433 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5435 struct sched_group *tmp, *first;
5444 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5449 } while (sg != first);
5452 static void free_sched_domain(struct rcu_head *rcu)
5454 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5457 * If its an overlapping domain it has private groups, iterate and
5460 if (sd->flags & SD_OVERLAP) {
5461 free_sched_groups(sd->groups, 1);
5462 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5463 kfree(sd->groups->sgp);
5469 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5471 call_rcu(&sd->rcu, free_sched_domain);
5474 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5476 for (; sd; sd = sd->parent)
5477 destroy_sched_domain(sd, cpu);
5481 * Keep a special pointer to the highest sched_domain that has
5482 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5483 * allows us to avoid some pointer chasing select_idle_sibling().
5485 * Also keep a unique ID per domain (we use the first cpu number in
5486 * the cpumask of the domain), this allows us to quickly tell if
5487 * two cpus are in the same cache domain, see cpus_share_cache().
5489 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5490 DEFINE_PER_CPU(int, sd_llc_id);
5492 static void update_top_cache_domain(int cpu)
5494 struct sched_domain *sd;
5497 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5499 id = cpumask_first(sched_domain_span(sd));
5501 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5502 per_cpu(sd_llc_id, cpu) = id;
5506 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5507 * hold the hotplug lock.
5510 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5512 struct rq *rq = cpu_rq(cpu);
5513 struct sched_domain *tmp;
5515 /* Remove the sched domains which do not contribute to scheduling. */
5516 for (tmp = sd; tmp; ) {
5517 struct sched_domain *parent = tmp->parent;
5521 if (sd_parent_degenerate(tmp, parent)) {
5522 tmp->parent = parent->parent;
5524 parent->parent->child = tmp;
5525 destroy_sched_domain(parent, cpu);
5530 if (sd && sd_degenerate(sd)) {
5533 destroy_sched_domain(tmp, cpu);
5538 sched_domain_debug(sd, cpu);
5540 rq_attach_root(rq, rd);
5542 rcu_assign_pointer(rq->sd, sd);
5543 destroy_sched_domains(tmp, cpu);
5545 update_top_cache_domain(cpu);
5548 /* cpus with isolated domains */
5549 static cpumask_var_t cpu_isolated_map;
5551 /* Setup the mask of cpus configured for isolated domains */
5552 static int __init isolated_cpu_setup(char *str)
5554 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5555 cpulist_parse(str, cpu_isolated_map);
5559 __setup("isolcpus=", isolated_cpu_setup);
5561 static const struct cpumask *cpu_cpu_mask(int cpu)
5563 return cpumask_of_node(cpu_to_node(cpu));
5567 struct sched_domain **__percpu sd;
5568 struct sched_group **__percpu sg;
5569 struct sched_group_power **__percpu sgp;
5573 struct sched_domain ** __percpu sd;
5574 struct root_domain *rd;
5584 struct sched_domain_topology_level;
5586 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5587 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5589 #define SDTL_OVERLAP 0x01
5591 struct sched_domain_topology_level {
5592 sched_domain_init_f init;
5593 sched_domain_mask_f mask;
5596 struct sd_data data;
5600 * Build an iteration mask that can exclude certain CPUs from the upwards
5603 * Asymmetric node setups can result in situations where the domain tree is of
5604 * unequal depth, make sure to skip domains that already cover the entire
5607 * In that case build_sched_domains() will have terminated the iteration early
5608 * and our sibling sd spans will be empty. Domains should always include the
5609 * cpu they're built on, so check that.
5612 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5614 const struct cpumask *span = sched_domain_span(sd);
5615 struct sd_data *sdd = sd->private;
5616 struct sched_domain *sibling;
5619 for_each_cpu(i, span) {
5620 sibling = *per_cpu_ptr(sdd->sd, i);
5621 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5624 cpumask_set_cpu(i, sched_group_mask(sg));
5629 * Return the canonical balance cpu for this group, this is the first cpu
5630 * of this group that's also in the iteration mask.
5632 int group_balance_cpu(struct sched_group *sg)
5634 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5638 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5640 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5641 const struct cpumask *span = sched_domain_span(sd);
5642 struct cpumask *covered = sched_domains_tmpmask;
5643 struct sd_data *sdd = sd->private;
5644 struct sched_domain *child;
5647 cpumask_clear(covered);
5649 for_each_cpu(i, span) {
5650 struct cpumask *sg_span;
5652 if (cpumask_test_cpu(i, covered))
5655 child = *per_cpu_ptr(sdd->sd, i);
5657 /* See the comment near build_group_mask(). */
5658 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5661 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5662 GFP_KERNEL, cpu_to_node(cpu));
5667 sg_span = sched_group_cpus(sg);
5669 child = child->child;
5670 cpumask_copy(sg_span, sched_domain_span(child));
5672 cpumask_set_cpu(i, sg_span);
5674 cpumask_or(covered, covered, sg_span);
5676 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5677 if (atomic_inc_return(&sg->sgp->ref) == 1)
5678 build_group_mask(sd, sg);
5681 * Initialize sgp->power such that even if we mess up the
5682 * domains and no possible iteration will get us here, we won't
5685 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5688 * Make sure the first group of this domain contains the
5689 * canonical balance cpu. Otherwise the sched_domain iteration
5690 * breaks. See update_sg_lb_stats().
5692 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5693 group_balance_cpu(sg) == cpu)
5703 sd->groups = groups;
5708 free_sched_groups(first, 0);
5713 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5715 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5716 struct sched_domain *child = sd->child;
5719 cpu = cpumask_first(sched_domain_span(child));
5722 *sg = *per_cpu_ptr(sdd->sg, cpu);
5723 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5724 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5731 * build_sched_groups will build a circular linked list of the groups
5732 * covered by the given span, and will set each group's ->cpumask correctly,
5733 * and ->cpu_power to 0.
5735 * Assumes the sched_domain tree is fully constructed
5738 build_sched_groups(struct sched_domain *sd, int cpu)
5740 struct sched_group *first = NULL, *last = NULL;
5741 struct sd_data *sdd = sd->private;
5742 const struct cpumask *span = sched_domain_span(sd);
5743 struct cpumask *covered;
5746 get_group(cpu, sdd, &sd->groups);
5747 atomic_inc(&sd->groups->ref);
5749 if (cpu != cpumask_first(sched_domain_span(sd)))
5752 lockdep_assert_held(&sched_domains_mutex);
5753 covered = sched_domains_tmpmask;
5755 cpumask_clear(covered);
5757 for_each_cpu(i, span) {
5758 struct sched_group *sg;
5759 int group = get_group(i, sdd, &sg);
5762 if (cpumask_test_cpu(i, covered))
5765 cpumask_clear(sched_group_cpus(sg));
5767 cpumask_setall(sched_group_mask(sg));
5769 for_each_cpu(j, span) {
5770 if (get_group(j, sdd, NULL) != group)
5773 cpumask_set_cpu(j, covered);
5774 cpumask_set_cpu(j, sched_group_cpus(sg));
5789 * Initialize sched groups cpu_power.
5791 * cpu_power indicates the capacity of sched group, which is used while
5792 * distributing the load between different sched groups in a sched domain.
5793 * Typically cpu_power for all the groups in a sched domain will be same unless
5794 * there are asymmetries in the topology. If there are asymmetries, group
5795 * having more cpu_power will pickup more load compared to the group having
5798 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5800 struct sched_group *sg = sd->groups;
5802 WARN_ON(!sd || !sg);
5805 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5807 } while (sg != sd->groups);
5809 if (cpu != group_balance_cpu(sg))
5812 update_group_power(sd, cpu);
5813 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5816 int __weak arch_sd_sibling_asym_packing(void)
5818 return 0*SD_ASYM_PACKING;
5822 * Initializers for schedule domains
5823 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5826 #ifdef CONFIG_SCHED_DEBUG
5827 # define SD_INIT_NAME(sd, type) sd->name = #type
5829 # define SD_INIT_NAME(sd, type) do { } while (0)
5832 #define SD_INIT_FUNC(type) \
5833 static noinline struct sched_domain * \
5834 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5836 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5837 *sd = SD_##type##_INIT; \
5838 SD_INIT_NAME(sd, type); \
5839 sd->private = &tl->data; \
5844 #ifdef CONFIG_SCHED_SMT
5845 SD_INIT_FUNC(SIBLING)
5847 #ifdef CONFIG_SCHED_MC
5850 #ifdef CONFIG_SCHED_BOOK
5854 static int default_relax_domain_level = -1;
5855 int sched_domain_level_max;
5857 static int __init setup_relax_domain_level(char *str)
5859 if (kstrtoint(str, 0, &default_relax_domain_level))
5860 pr_warn("Unable to set relax_domain_level\n");
5864 __setup("relax_domain_level=", setup_relax_domain_level);
5866 static void set_domain_attribute(struct sched_domain *sd,
5867 struct sched_domain_attr *attr)
5871 if (!attr || attr->relax_domain_level < 0) {
5872 if (default_relax_domain_level < 0)
5875 request = default_relax_domain_level;
5877 request = attr->relax_domain_level;
5878 if (request < sd->level) {
5879 /* turn off idle balance on this domain */
5880 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5882 /* turn on idle balance on this domain */
5883 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5887 static void __sdt_free(const struct cpumask *cpu_map);
5888 static int __sdt_alloc(const struct cpumask *cpu_map);
5890 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5891 const struct cpumask *cpu_map)
5895 if (!atomic_read(&d->rd->refcount))
5896 free_rootdomain(&d->rd->rcu); /* fall through */
5898 free_percpu(d->sd); /* fall through */
5900 __sdt_free(cpu_map); /* fall through */
5906 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5907 const struct cpumask *cpu_map)
5909 memset(d, 0, sizeof(*d));
5911 if (__sdt_alloc(cpu_map))
5912 return sa_sd_storage;
5913 d->sd = alloc_percpu(struct sched_domain *);
5915 return sa_sd_storage;
5916 d->rd = alloc_rootdomain();
5919 return sa_rootdomain;
5923 * NULL the sd_data elements we've used to build the sched_domain and
5924 * sched_group structure so that the subsequent __free_domain_allocs()
5925 * will not free the data we're using.
5927 static void claim_allocations(int cpu, struct sched_domain *sd)
5929 struct sd_data *sdd = sd->private;
5931 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5932 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5934 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5935 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5937 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5938 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5941 #ifdef CONFIG_SCHED_SMT
5942 static const struct cpumask *cpu_smt_mask(int cpu)
5944 return topology_thread_cpumask(cpu);
5949 * Topology list, bottom-up.
5951 static struct sched_domain_topology_level default_topology[] = {
5952 #ifdef CONFIG_SCHED_SMT
5953 { sd_init_SIBLING, cpu_smt_mask, },
5955 #ifdef CONFIG_SCHED_MC
5956 { sd_init_MC, cpu_coregroup_mask, },
5958 #ifdef CONFIG_SCHED_BOOK
5959 { sd_init_BOOK, cpu_book_mask, },
5961 { sd_init_CPU, cpu_cpu_mask, },
5965 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5969 static int sched_domains_numa_levels;
5970 static int *sched_domains_numa_distance;
5971 static struct cpumask ***sched_domains_numa_masks;
5972 static int sched_domains_curr_level;
5974 static inline int sd_local_flags(int level)
5976 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5979 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5982 static struct sched_domain *
5983 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5985 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5986 int level = tl->numa_level;
5987 int sd_weight = cpumask_weight(
5988 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5990 *sd = (struct sched_domain){
5991 .min_interval = sd_weight,
5992 .max_interval = 2*sd_weight,
5994 .imbalance_pct = 125,
5995 .cache_nice_tries = 2,
6002 .flags = 1*SD_LOAD_BALANCE
6003 | 1*SD_BALANCE_NEWIDLE
6008 | 0*SD_SHARE_CPUPOWER
6009 | 0*SD_SHARE_PKG_RESOURCES
6011 | 0*SD_PREFER_SIBLING
6012 | sd_local_flags(level)
6014 .last_balance = jiffies,
6015 .balance_interval = sd_weight,
6017 SD_INIT_NAME(sd, NUMA);
6018 sd->private = &tl->data;
6021 * Ugly hack to pass state to sd_numa_mask()...
6023 sched_domains_curr_level = tl->numa_level;
6028 static const struct cpumask *sd_numa_mask(int cpu)
6030 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6033 static void sched_numa_warn(const char *str)
6035 static int done = false;
6043 printk(KERN_WARNING "ERROR: %s\n\n", str);
6045 for (i = 0; i < nr_node_ids; i++) {
6046 printk(KERN_WARNING " ");
6047 for (j = 0; j < nr_node_ids; j++)
6048 printk(KERN_CONT "%02d ", node_distance(i,j));
6049 printk(KERN_CONT "\n");
6051 printk(KERN_WARNING "\n");
6054 static bool find_numa_distance(int distance)
6058 if (distance == node_distance(0, 0))
6061 for (i = 0; i < sched_domains_numa_levels; i++) {
6062 if (sched_domains_numa_distance[i] == distance)
6069 static void sched_init_numa(void)
6071 int next_distance, curr_distance = node_distance(0, 0);
6072 struct sched_domain_topology_level *tl;
6076 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6077 if (!sched_domains_numa_distance)
6081 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6082 * unique distances in the node_distance() table.
6084 * Assumes node_distance(0,j) includes all distances in
6085 * node_distance(i,j) in order to avoid cubic time.
6087 next_distance = curr_distance;
6088 for (i = 0; i < nr_node_ids; i++) {
6089 for (j = 0; j < nr_node_ids; j++) {
6090 for (k = 0; k < nr_node_ids; k++) {
6091 int distance = node_distance(i, k);
6093 if (distance > curr_distance &&
6094 (distance < next_distance ||
6095 next_distance == curr_distance))
6096 next_distance = distance;
6099 * While not a strong assumption it would be nice to know
6100 * about cases where if node A is connected to B, B is not
6101 * equally connected to A.
6103 if (sched_debug() && node_distance(k, i) != distance)
6104 sched_numa_warn("Node-distance not symmetric");
6106 if (sched_debug() && i && !find_numa_distance(distance))
6107 sched_numa_warn("Node-0 not representative");
6109 if (next_distance != curr_distance) {
6110 sched_domains_numa_distance[level++] = next_distance;
6111 sched_domains_numa_levels = level;
6112 curr_distance = next_distance;
6117 * In case of sched_debug() we verify the above assumption.
6123 * 'level' contains the number of unique distances, excluding the
6124 * identity distance node_distance(i,i).
6126 * The sched_domains_nume_distance[] array includes the actual distance
6131 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6132 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6133 * the array will contain less then 'level' members. This could be
6134 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6135 * in other functions.
6137 * We reset it to 'level' at the end of this function.
6139 sched_domains_numa_levels = 0;
6141 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6142 if (!sched_domains_numa_masks)
6146 * Now for each level, construct a mask per node which contains all
6147 * cpus of nodes that are that many hops away from us.
6149 for (i = 0; i < level; i++) {
6150 sched_domains_numa_masks[i] =
6151 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6152 if (!sched_domains_numa_masks[i])
6155 for (j = 0; j < nr_node_ids; j++) {
6156 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6160 sched_domains_numa_masks[i][j] = mask;
6162 for (k = 0; k < nr_node_ids; k++) {
6163 if (node_distance(j, k) > sched_domains_numa_distance[i])
6166 cpumask_or(mask, mask, cpumask_of_node(k));
6171 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6172 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6177 * Copy the default topology bits..
6179 for (i = 0; default_topology[i].init; i++)
6180 tl[i] = default_topology[i];
6183 * .. and append 'j' levels of NUMA goodness.
6185 for (j = 0; j < level; i++, j++) {
6186 tl[i] = (struct sched_domain_topology_level){
6187 .init = sd_numa_init,
6188 .mask = sd_numa_mask,
6189 .flags = SDTL_OVERLAP,
6194 sched_domain_topology = tl;
6196 sched_domains_numa_levels = level;
6199 static void sched_domains_numa_masks_set(int cpu)
6202 int node = cpu_to_node(cpu);
6204 for (i = 0; i < sched_domains_numa_levels; i++) {
6205 for (j = 0; j < nr_node_ids; j++) {
6206 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6207 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6212 static void sched_domains_numa_masks_clear(int cpu)
6215 for (i = 0; i < sched_domains_numa_levels; i++) {
6216 for (j = 0; j < nr_node_ids; j++)
6217 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6222 * Update sched_domains_numa_masks[level][node] array when new cpus
6225 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6226 unsigned long action,
6229 int cpu = (long)hcpu;
6231 switch (action & ~CPU_TASKS_FROZEN) {
6233 sched_domains_numa_masks_set(cpu);
6237 sched_domains_numa_masks_clear(cpu);
6247 static inline void sched_init_numa(void)
6251 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6252 unsigned long action,
6257 #endif /* CONFIG_NUMA */
6259 static int __sdt_alloc(const struct cpumask *cpu_map)
6261 struct sched_domain_topology_level *tl;
6264 for (tl = sched_domain_topology; tl->init; tl++) {
6265 struct sd_data *sdd = &tl->data;
6267 sdd->sd = alloc_percpu(struct sched_domain *);
6271 sdd->sg = alloc_percpu(struct sched_group *);
6275 sdd->sgp = alloc_percpu(struct sched_group_power *);
6279 for_each_cpu(j, cpu_map) {
6280 struct sched_domain *sd;
6281 struct sched_group *sg;
6282 struct sched_group_power *sgp;
6284 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6285 GFP_KERNEL, cpu_to_node(j));
6289 *per_cpu_ptr(sdd->sd, j) = sd;
6291 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6292 GFP_KERNEL, cpu_to_node(j));
6298 *per_cpu_ptr(sdd->sg, j) = sg;
6300 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6301 GFP_KERNEL, cpu_to_node(j));
6305 *per_cpu_ptr(sdd->sgp, j) = sgp;
6312 static void __sdt_free(const struct cpumask *cpu_map)
6314 struct sched_domain_topology_level *tl;
6317 for (tl = sched_domain_topology; tl->init; tl++) {
6318 struct sd_data *sdd = &tl->data;
6320 for_each_cpu(j, cpu_map) {
6321 struct sched_domain *sd;
6324 sd = *per_cpu_ptr(sdd->sd, j);
6325 if (sd && (sd->flags & SD_OVERLAP))
6326 free_sched_groups(sd->groups, 0);
6327 kfree(*per_cpu_ptr(sdd->sd, j));
6331 kfree(*per_cpu_ptr(sdd->sg, j));
6333 kfree(*per_cpu_ptr(sdd->sgp, j));
6335 free_percpu(sdd->sd);
6337 free_percpu(sdd->sg);
6339 free_percpu(sdd->sgp);
6344 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6345 struct s_data *d, const struct cpumask *cpu_map,
6346 struct sched_domain_attr *attr, struct sched_domain *child,
6349 struct sched_domain *sd = tl->init(tl, cpu);
6353 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6355 sd->level = child->level + 1;
6356 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6360 set_domain_attribute(sd, attr);
6366 * Build sched domains for a given set of cpus and attach the sched domains
6367 * to the individual cpus
6369 static int build_sched_domains(const struct cpumask *cpu_map,
6370 struct sched_domain_attr *attr)
6372 enum s_alloc alloc_state = sa_none;
6373 struct sched_domain *sd;
6375 int i, ret = -ENOMEM;
6377 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6378 if (alloc_state != sa_rootdomain)
6381 /* Set up domains for cpus specified by the cpu_map. */
6382 for_each_cpu(i, cpu_map) {
6383 struct sched_domain_topology_level *tl;
6386 for (tl = sched_domain_topology; tl->init; tl++) {
6387 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6388 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6389 sd->flags |= SD_OVERLAP;
6390 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6397 *per_cpu_ptr(d.sd, i) = sd;
6400 /* Build the groups for the domains */
6401 for_each_cpu(i, cpu_map) {
6402 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6403 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6404 if (sd->flags & SD_OVERLAP) {
6405 if (build_overlap_sched_groups(sd, i))
6408 if (build_sched_groups(sd, i))
6414 /* Calculate CPU power for physical packages and nodes */
6415 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6416 if (!cpumask_test_cpu(i, cpu_map))
6419 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6420 claim_allocations(i, sd);
6421 init_sched_groups_power(i, sd);
6425 /* Attach the domains */
6427 for_each_cpu(i, cpu_map) {
6428 sd = *per_cpu_ptr(d.sd, i);
6429 cpu_attach_domain(sd, d.rd, i);
6435 __free_domain_allocs(&d, alloc_state, cpu_map);
6439 static cpumask_var_t *doms_cur; /* current sched domains */
6440 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6441 static struct sched_domain_attr *dattr_cur;
6442 /* attribues of custom domains in 'doms_cur' */
6445 * Special case: If a kmalloc of a doms_cur partition (array of
6446 * cpumask) fails, then fallback to a single sched domain,
6447 * as determined by the single cpumask fallback_doms.
6449 static cpumask_var_t fallback_doms;
6452 * arch_update_cpu_topology lets virtualized architectures update the
6453 * cpu core maps. It is supposed to return 1 if the topology changed
6454 * or 0 if it stayed the same.
6456 int __attribute__((weak)) arch_update_cpu_topology(void)
6461 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6464 cpumask_var_t *doms;
6466 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6469 for (i = 0; i < ndoms; i++) {
6470 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6471 free_sched_domains(doms, i);
6478 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6481 for (i = 0; i < ndoms; i++)
6482 free_cpumask_var(doms[i]);
6487 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6488 * For now this just excludes isolated cpus, but could be used to
6489 * exclude other special cases in the future.
6491 static int init_sched_domains(const struct cpumask *cpu_map)
6495 arch_update_cpu_topology();
6497 doms_cur = alloc_sched_domains(ndoms_cur);
6499 doms_cur = &fallback_doms;
6500 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6501 err = build_sched_domains(doms_cur[0], NULL);
6502 register_sched_domain_sysctl();
6508 * Detach sched domains from a group of cpus specified in cpu_map
6509 * These cpus will now be attached to the NULL domain
6511 static void detach_destroy_domains(const struct cpumask *cpu_map)
6516 for_each_cpu(i, cpu_map)
6517 cpu_attach_domain(NULL, &def_root_domain, i);
6521 /* handle null as "default" */
6522 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6523 struct sched_domain_attr *new, int idx_new)
6525 struct sched_domain_attr tmp;
6532 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6533 new ? (new + idx_new) : &tmp,
6534 sizeof(struct sched_domain_attr));
6538 * Partition sched domains as specified by the 'ndoms_new'
6539 * cpumasks in the array doms_new[] of cpumasks. This compares
6540 * doms_new[] to the current sched domain partitioning, doms_cur[].
6541 * It destroys each deleted domain and builds each new domain.
6543 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6544 * The masks don't intersect (don't overlap.) We should setup one
6545 * sched domain for each mask. CPUs not in any of the cpumasks will
6546 * not be load balanced. If the same cpumask appears both in the
6547 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6550 * The passed in 'doms_new' should be allocated using
6551 * alloc_sched_domains. This routine takes ownership of it and will
6552 * free_sched_domains it when done with it. If the caller failed the
6553 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6554 * and partition_sched_domains() will fallback to the single partition
6555 * 'fallback_doms', it also forces the domains to be rebuilt.
6557 * If doms_new == NULL it will be replaced with cpu_online_mask.
6558 * ndoms_new == 0 is a special case for destroying existing domains,
6559 * and it will not create the default domain.
6561 * Call with hotplug lock held
6563 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6564 struct sched_domain_attr *dattr_new)
6569 mutex_lock(&sched_domains_mutex);
6571 /* always unregister in case we don't destroy any domains */
6572 unregister_sched_domain_sysctl();
6574 /* Let architecture update cpu core mappings. */
6575 new_topology = arch_update_cpu_topology();
6577 n = doms_new ? ndoms_new : 0;
6579 /* Destroy deleted domains */
6580 for (i = 0; i < ndoms_cur; i++) {
6581 for (j = 0; j < n && !new_topology; j++) {
6582 if (cpumask_equal(doms_cur[i], doms_new[j])
6583 && dattrs_equal(dattr_cur, i, dattr_new, j))
6586 /* no match - a current sched domain not in new doms_new[] */
6587 detach_destroy_domains(doms_cur[i]);
6592 if (doms_new == NULL) {
6594 doms_new = &fallback_doms;
6595 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6596 WARN_ON_ONCE(dattr_new);
6599 /* Build new domains */
6600 for (i = 0; i < ndoms_new; i++) {
6601 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6602 if (cpumask_equal(doms_new[i], doms_cur[j])
6603 && dattrs_equal(dattr_new, i, dattr_cur, j))
6606 /* no match - add a new doms_new */
6607 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6612 /* Remember the new sched domains */
6613 if (doms_cur != &fallback_doms)
6614 free_sched_domains(doms_cur, ndoms_cur);
6615 kfree(dattr_cur); /* kfree(NULL) is safe */
6616 doms_cur = doms_new;
6617 dattr_cur = dattr_new;
6618 ndoms_cur = ndoms_new;
6620 register_sched_domain_sysctl();
6622 mutex_unlock(&sched_domains_mutex);
6625 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6628 * Update cpusets according to cpu_active mask. If cpusets are
6629 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6630 * around partition_sched_domains().
6632 * If we come here as part of a suspend/resume, don't touch cpusets because we
6633 * want to restore it back to its original state upon resume anyway.
6635 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6639 case CPU_ONLINE_FROZEN:
6640 case CPU_DOWN_FAILED_FROZEN:
6643 * num_cpus_frozen tracks how many CPUs are involved in suspend
6644 * resume sequence. As long as this is not the last online
6645 * operation in the resume sequence, just build a single sched
6646 * domain, ignoring cpusets.
6649 if (likely(num_cpus_frozen)) {
6650 partition_sched_domains(1, NULL, NULL);
6655 * This is the last CPU online operation. So fall through and
6656 * restore the original sched domains by considering the
6657 * cpuset configurations.
6661 case CPU_DOWN_FAILED:
6662 cpuset_update_active_cpus(true);
6670 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6674 case CPU_DOWN_PREPARE:
6675 cpuset_update_active_cpus(false);
6677 case CPU_DOWN_PREPARE_FROZEN:
6679 partition_sched_domains(1, NULL, NULL);
6687 void __init sched_init_smp(void)
6689 cpumask_var_t non_isolated_cpus;
6691 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6692 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6697 mutex_lock(&sched_domains_mutex);
6698 init_sched_domains(cpu_active_mask);
6699 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6700 if (cpumask_empty(non_isolated_cpus))
6701 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6702 mutex_unlock(&sched_domains_mutex);
6705 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6706 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6707 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6709 /* RT runtime code needs to handle some hotplug events */
6710 hotcpu_notifier(update_runtime, 0);
6714 /* Move init over to a non-isolated CPU */
6715 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6717 sched_init_granularity();
6718 free_cpumask_var(non_isolated_cpus);
6720 init_sched_rt_class();
6723 void __init sched_init_smp(void)
6725 sched_init_granularity();
6727 #endif /* CONFIG_SMP */
6729 const_debug unsigned int sysctl_timer_migration = 1;
6731 int in_sched_functions(unsigned long addr)
6733 return in_lock_functions(addr) ||
6734 (addr >= (unsigned long)__sched_text_start
6735 && addr < (unsigned long)__sched_text_end);
6738 #ifdef CONFIG_CGROUP_SCHED
6739 struct task_group root_task_group;
6740 LIST_HEAD(task_groups);
6743 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6745 void __init sched_init(void)
6748 unsigned long alloc_size = 0, ptr;
6750 #ifdef CONFIG_FAIR_GROUP_SCHED
6751 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6753 #ifdef CONFIG_RT_GROUP_SCHED
6754 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6756 #ifdef CONFIG_CPUMASK_OFFSTACK
6757 alloc_size += num_possible_cpus() * cpumask_size();
6760 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6762 #ifdef CONFIG_FAIR_GROUP_SCHED
6763 root_task_group.se = (struct sched_entity **)ptr;
6764 ptr += nr_cpu_ids * sizeof(void **);
6766 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6767 ptr += nr_cpu_ids * sizeof(void **);
6769 #endif /* CONFIG_FAIR_GROUP_SCHED */
6770 #ifdef CONFIG_RT_GROUP_SCHED
6771 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6772 ptr += nr_cpu_ids * sizeof(void **);
6774 root_task_group.rt_rq = (struct rt_rq **)ptr;
6775 ptr += nr_cpu_ids * sizeof(void **);
6777 #endif /* CONFIG_RT_GROUP_SCHED */
6778 #ifdef CONFIG_CPUMASK_OFFSTACK
6779 for_each_possible_cpu(i) {
6780 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6781 ptr += cpumask_size();
6783 #endif /* CONFIG_CPUMASK_OFFSTACK */
6787 init_defrootdomain();
6790 init_rt_bandwidth(&def_rt_bandwidth,
6791 global_rt_period(), global_rt_runtime());
6793 #ifdef CONFIG_RT_GROUP_SCHED
6794 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6795 global_rt_period(), global_rt_runtime());
6796 #endif /* CONFIG_RT_GROUP_SCHED */
6798 #ifdef CONFIG_CGROUP_SCHED
6799 list_add(&root_task_group.list, &task_groups);
6800 INIT_LIST_HEAD(&root_task_group.children);
6801 INIT_LIST_HEAD(&root_task_group.siblings);
6802 autogroup_init(&init_task);
6804 #endif /* CONFIG_CGROUP_SCHED */
6806 #ifdef CONFIG_CGROUP_CPUACCT
6807 root_cpuacct.cpustat = &kernel_cpustat;
6808 root_cpuacct.cpuusage = alloc_percpu(u64);
6809 /* Too early, not expected to fail */
6810 BUG_ON(!root_cpuacct.cpuusage);
6812 for_each_possible_cpu(i) {
6816 raw_spin_lock_init(&rq->lock);
6818 rq->calc_load_active = 0;
6819 rq->calc_load_update = jiffies + LOAD_FREQ;
6820 init_cfs_rq(&rq->cfs);
6821 init_rt_rq(&rq->rt, rq);
6822 #ifdef CONFIG_FAIR_GROUP_SCHED
6823 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6824 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6826 * How much cpu bandwidth does root_task_group get?
6828 * In case of task-groups formed thr' the cgroup filesystem, it
6829 * gets 100% of the cpu resources in the system. This overall
6830 * system cpu resource is divided among the tasks of
6831 * root_task_group and its child task-groups in a fair manner,
6832 * based on each entity's (task or task-group's) weight
6833 * (se->load.weight).
6835 * In other words, if root_task_group has 10 tasks of weight
6836 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6837 * then A0's share of the cpu resource is:
6839 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6841 * We achieve this by letting root_task_group's tasks sit
6842 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6844 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6845 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6846 #endif /* CONFIG_FAIR_GROUP_SCHED */
6848 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6849 #ifdef CONFIG_RT_GROUP_SCHED
6850 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6851 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6854 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6855 rq->cpu_load[j] = 0;
6857 rq->last_load_update_tick = jiffies;
6862 rq->cpu_power = SCHED_POWER_SCALE;
6863 rq->post_schedule = 0;
6864 rq->active_balance = 0;
6865 rq->next_balance = jiffies;
6870 rq->avg_idle = 2*sysctl_sched_migration_cost;
6872 INIT_LIST_HEAD(&rq->cfs_tasks);
6874 rq_attach_root(rq, &def_root_domain);
6880 atomic_set(&rq->nr_iowait, 0);
6883 set_load_weight(&init_task);
6885 #ifdef CONFIG_PREEMPT_NOTIFIERS
6886 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6889 #ifdef CONFIG_RT_MUTEXES
6890 plist_head_init(&init_task.pi_waiters);
6894 * The boot idle thread does lazy MMU switching as well:
6896 atomic_inc(&init_mm.mm_count);
6897 enter_lazy_tlb(&init_mm, current);
6900 * Make us the idle thread. Technically, schedule() should not be
6901 * called from this thread, however somewhere below it might be,
6902 * but because we are the idle thread, we just pick up running again
6903 * when this runqueue becomes "idle".
6905 init_idle(current, smp_processor_id());
6907 calc_load_update = jiffies + LOAD_FREQ;
6910 * During early bootup we pretend to be a normal task:
6912 current->sched_class = &fair_sched_class;
6915 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6916 /* May be allocated at isolcpus cmdline parse time */
6917 if (cpu_isolated_map == NULL)
6918 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6919 idle_thread_set_boot_cpu();
6921 init_sched_fair_class();
6923 scheduler_running = 1;
6926 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6927 static inline int preempt_count_equals(int preempt_offset)
6929 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6931 return (nested == preempt_offset);
6934 void __might_sleep(const char *file, int line, int preempt_offset)
6936 static unsigned long prev_jiffy; /* ratelimiting */
6938 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6939 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6940 system_state != SYSTEM_RUNNING || oops_in_progress)
6942 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6944 prev_jiffy = jiffies;
6947 "BUG: sleeping function called from invalid context at %s:%d\n",
6950 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6951 in_atomic(), irqs_disabled(),
6952 current->pid, current->comm);
6954 debug_show_held_locks(current);
6955 if (irqs_disabled())
6956 print_irqtrace_events(current);
6959 EXPORT_SYMBOL(__might_sleep);
6962 #ifdef CONFIG_MAGIC_SYSRQ
6963 static void normalize_task(struct rq *rq, struct task_struct *p)
6965 const struct sched_class *prev_class = p->sched_class;
6966 int old_prio = p->prio;
6971 dequeue_task(rq, p, 0);
6972 __setscheduler(rq, p, SCHED_NORMAL, 0);
6974 enqueue_task(rq, p, 0);
6975 resched_task(rq->curr);
6978 check_class_changed(rq, p, prev_class, old_prio);
6981 void normalize_rt_tasks(void)
6983 struct task_struct *g, *p;
6984 unsigned long flags;
6987 read_lock_irqsave(&tasklist_lock, flags);
6988 do_each_thread(g, p) {
6990 * Only normalize user tasks:
6995 p->se.exec_start = 0;
6996 #ifdef CONFIG_SCHEDSTATS
6997 p->se.statistics.wait_start = 0;
6998 p->se.statistics.sleep_start = 0;
6999 p->se.statistics.block_start = 0;
7004 * Renice negative nice level userspace
7007 if (TASK_NICE(p) < 0 && p->mm)
7008 set_user_nice(p, 0);
7012 raw_spin_lock(&p->pi_lock);
7013 rq = __task_rq_lock(p);
7015 normalize_task(rq, p);
7017 __task_rq_unlock(rq);
7018 raw_spin_unlock(&p->pi_lock);
7019 } while_each_thread(g, p);
7021 read_unlock_irqrestore(&tasklist_lock, flags);
7024 #endif /* CONFIG_MAGIC_SYSRQ */
7026 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7028 * These functions are only useful for the IA64 MCA handling, or kdb.
7030 * They can only be called when the whole system has been
7031 * stopped - every CPU needs to be quiescent, and no scheduling
7032 * activity can take place. Using them for anything else would
7033 * be a serious bug, and as a result, they aren't even visible
7034 * under any other configuration.
7038 * curr_task - return the current task for a given cpu.
7039 * @cpu: the processor in question.
7041 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7043 struct task_struct *curr_task(int cpu)
7045 return cpu_curr(cpu);
7048 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7052 * set_curr_task - set the current task for a given cpu.
7053 * @cpu: the processor in question.
7054 * @p: the task pointer to set.
7056 * Description: This function must only be used when non-maskable interrupts
7057 * are serviced on a separate stack. It allows the architecture to switch the
7058 * notion of the current task on a cpu in a non-blocking manner. This function
7059 * must be called with all CPU's synchronized, and interrupts disabled, the
7060 * and caller must save the original value of the current task (see
7061 * curr_task() above) and restore that value before reenabling interrupts and
7062 * re-starting the system.
7064 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7066 void set_curr_task(int cpu, struct task_struct *p)
7073 #ifdef CONFIG_CGROUP_SCHED
7074 /* task_group_lock serializes the addition/removal of task groups */
7075 static DEFINE_SPINLOCK(task_group_lock);
7077 static void free_sched_group(struct task_group *tg)
7079 free_fair_sched_group(tg);
7080 free_rt_sched_group(tg);
7085 /* allocate runqueue etc for a new task group */
7086 struct task_group *sched_create_group(struct task_group *parent)
7088 struct task_group *tg;
7089 unsigned long flags;
7091 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7093 return ERR_PTR(-ENOMEM);
7095 if (!alloc_fair_sched_group(tg, parent))
7098 if (!alloc_rt_sched_group(tg, parent))
7101 spin_lock_irqsave(&task_group_lock, flags);
7102 list_add_rcu(&tg->list, &task_groups);
7104 WARN_ON(!parent); /* root should already exist */
7106 tg->parent = parent;
7107 INIT_LIST_HEAD(&tg->children);
7108 list_add_rcu(&tg->siblings, &parent->children);
7109 spin_unlock_irqrestore(&task_group_lock, flags);
7114 free_sched_group(tg);
7115 return ERR_PTR(-ENOMEM);
7118 /* rcu callback to free various structures associated with a task group */
7119 static void free_sched_group_rcu(struct rcu_head *rhp)
7121 /* now it should be safe to free those cfs_rqs */
7122 free_sched_group(container_of(rhp, struct task_group, rcu));
7125 /* Destroy runqueue etc associated with a task group */
7126 void sched_destroy_group(struct task_group *tg)
7128 unsigned long flags;
7131 /* end participation in shares distribution */
7132 for_each_possible_cpu(i)
7133 unregister_fair_sched_group(tg, i);
7135 spin_lock_irqsave(&task_group_lock, flags);
7136 list_del_rcu(&tg->list);
7137 list_del_rcu(&tg->siblings);
7138 spin_unlock_irqrestore(&task_group_lock, flags);
7140 /* wait for possible concurrent references to cfs_rqs complete */
7141 call_rcu(&tg->rcu, free_sched_group_rcu);
7144 /* change task's runqueue when it moves between groups.
7145 * The caller of this function should have put the task in its new group
7146 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7147 * reflect its new group.
7149 void sched_move_task(struct task_struct *tsk)
7151 struct task_group *tg;
7153 unsigned long flags;
7156 rq = task_rq_lock(tsk, &flags);
7158 running = task_current(rq, tsk);
7162 dequeue_task(rq, tsk, 0);
7163 if (unlikely(running))
7164 tsk->sched_class->put_prev_task(rq, tsk);
7166 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7167 lockdep_is_held(&tsk->sighand->siglock)),
7168 struct task_group, css);
7169 tg = autogroup_task_group(tsk, tg);
7170 tsk->sched_task_group = tg;
7172 #ifdef CONFIG_FAIR_GROUP_SCHED
7173 if (tsk->sched_class->task_move_group)
7174 tsk->sched_class->task_move_group(tsk, on_rq);
7177 set_task_rq(tsk, task_cpu(tsk));
7179 if (unlikely(running))
7180 tsk->sched_class->set_curr_task(rq);
7182 enqueue_task(rq, tsk, 0);
7184 task_rq_unlock(rq, tsk, &flags);
7186 #endif /* CONFIG_CGROUP_SCHED */
7188 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7189 static unsigned long to_ratio(u64 period, u64 runtime)
7191 if (runtime == RUNTIME_INF)
7194 return div64_u64(runtime << 20, period);
7198 #ifdef CONFIG_RT_GROUP_SCHED
7200 * Ensure that the real time constraints are schedulable.
7202 static DEFINE_MUTEX(rt_constraints_mutex);
7204 /* Must be called with tasklist_lock held */
7205 static inline int tg_has_rt_tasks(struct task_group *tg)
7207 struct task_struct *g, *p;
7209 do_each_thread(g, p) {
7210 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7212 } while_each_thread(g, p);
7217 struct rt_schedulable_data {
7218 struct task_group *tg;
7223 static int tg_rt_schedulable(struct task_group *tg, void *data)
7225 struct rt_schedulable_data *d = data;
7226 struct task_group *child;
7227 unsigned long total, sum = 0;
7228 u64 period, runtime;
7230 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7231 runtime = tg->rt_bandwidth.rt_runtime;
7234 period = d->rt_period;
7235 runtime = d->rt_runtime;
7239 * Cannot have more runtime than the period.
7241 if (runtime > period && runtime != RUNTIME_INF)
7245 * Ensure we don't starve existing RT tasks.
7247 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7250 total = to_ratio(period, runtime);
7253 * Nobody can have more than the global setting allows.
7255 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7259 * The sum of our children's runtime should not exceed our own.
7261 list_for_each_entry_rcu(child, &tg->children, siblings) {
7262 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7263 runtime = child->rt_bandwidth.rt_runtime;
7265 if (child == d->tg) {
7266 period = d->rt_period;
7267 runtime = d->rt_runtime;
7270 sum += to_ratio(period, runtime);
7279 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7283 struct rt_schedulable_data data = {
7285 .rt_period = period,
7286 .rt_runtime = runtime,
7290 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7296 static int tg_set_rt_bandwidth(struct task_group *tg,
7297 u64 rt_period, u64 rt_runtime)
7301 mutex_lock(&rt_constraints_mutex);
7302 read_lock(&tasklist_lock);
7303 err = __rt_schedulable(tg, rt_period, rt_runtime);
7307 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7308 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7309 tg->rt_bandwidth.rt_runtime = rt_runtime;
7311 for_each_possible_cpu(i) {
7312 struct rt_rq *rt_rq = tg->rt_rq[i];
7314 raw_spin_lock(&rt_rq->rt_runtime_lock);
7315 rt_rq->rt_runtime = rt_runtime;
7316 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7318 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7320 read_unlock(&tasklist_lock);
7321 mutex_unlock(&rt_constraints_mutex);
7326 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7328 u64 rt_runtime, rt_period;
7330 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7331 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7332 if (rt_runtime_us < 0)
7333 rt_runtime = RUNTIME_INF;
7335 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7338 long sched_group_rt_runtime(struct task_group *tg)
7342 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7345 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7346 do_div(rt_runtime_us, NSEC_PER_USEC);
7347 return rt_runtime_us;
7350 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7352 u64 rt_runtime, rt_period;
7354 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7355 rt_runtime = tg->rt_bandwidth.rt_runtime;
7360 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7363 long sched_group_rt_period(struct task_group *tg)
7367 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7368 do_div(rt_period_us, NSEC_PER_USEC);
7369 return rt_period_us;
7372 static int sched_rt_global_constraints(void)
7374 u64 runtime, period;
7377 if (sysctl_sched_rt_period <= 0)
7380 runtime = global_rt_runtime();
7381 period = global_rt_period();
7384 * Sanity check on the sysctl variables.
7386 if (runtime > period && runtime != RUNTIME_INF)
7389 mutex_lock(&rt_constraints_mutex);
7390 read_lock(&tasklist_lock);
7391 ret = __rt_schedulable(NULL, 0, 0);
7392 read_unlock(&tasklist_lock);
7393 mutex_unlock(&rt_constraints_mutex);
7398 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7400 /* Don't accept realtime tasks when there is no way for them to run */
7401 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7407 #else /* !CONFIG_RT_GROUP_SCHED */
7408 static int sched_rt_global_constraints(void)
7410 unsigned long flags;
7413 if (sysctl_sched_rt_period <= 0)
7417 * There's always some RT tasks in the root group
7418 * -- migration, kstopmachine etc..
7420 if (sysctl_sched_rt_runtime == 0)
7423 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7424 for_each_possible_cpu(i) {
7425 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7427 raw_spin_lock(&rt_rq->rt_runtime_lock);
7428 rt_rq->rt_runtime = global_rt_runtime();
7429 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7431 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7435 #endif /* CONFIG_RT_GROUP_SCHED */
7437 int sched_rt_handler(struct ctl_table *table, int write,
7438 void __user *buffer, size_t *lenp,
7442 int old_period, old_runtime;
7443 static DEFINE_MUTEX(mutex);
7446 old_period = sysctl_sched_rt_period;
7447 old_runtime = sysctl_sched_rt_runtime;
7449 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7451 if (!ret && write) {
7452 ret = sched_rt_global_constraints();
7454 sysctl_sched_rt_period = old_period;
7455 sysctl_sched_rt_runtime = old_runtime;
7457 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7458 def_rt_bandwidth.rt_period =
7459 ns_to_ktime(global_rt_period());
7462 mutex_unlock(&mutex);
7467 #ifdef CONFIG_CGROUP_SCHED
7469 /* return corresponding task_group object of a cgroup */
7470 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7472 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7473 struct task_group, css);
7476 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7478 struct task_group *tg, *parent;
7480 if (!cgrp->parent) {
7481 /* This is early initialization for the top cgroup */
7482 return &root_task_group.css;
7485 parent = cgroup_tg(cgrp->parent);
7486 tg = sched_create_group(parent);
7488 return ERR_PTR(-ENOMEM);
7493 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7495 struct task_group *tg = cgroup_tg(cgrp);
7497 sched_destroy_group(tg);
7500 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7501 struct cgroup_taskset *tset)
7503 struct task_struct *task;
7505 cgroup_taskset_for_each(task, cgrp, tset) {
7506 #ifdef CONFIG_RT_GROUP_SCHED
7507 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7510 /* We don't support RT-tasks being in separate groups */
7511 if (task->sched_class != &fair_sched_class)
7518 static void cpu_cgroup_attach(struct cgroup *cgrp,
7519 struct cgroup_taskset *tset)
7521 struct task_struct *task;
7523 cgroup_taskset_for_each(task, cgrp, tset)
7524 sched_move_task(task);
7528 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7529 struct task_struct *task)
7532 * cgroup_exit() is called in the copy_process() failure path.
7533 * Ignore this case since the task hasn't ran yet, this avoids
7534 * trying to poke a half freed task state from generic code.
7536 if (!(task->flags & PF_EXITING))
7539 sched_move_task(task);
7542 #ifdef CONFIG_FAIR_GROUP_SCHED
7543 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7546 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7549 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7551 struct task_group *tg = cgroup_tg(cgrp);
7553 return (u64) scale_load_down(tg->shares);
7556 #ifdef CONFIG_CFS_BANDWIDTH
7557 static DEFINE_MUTEX(cfs_constraints_mutex);
7559 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7560 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7562 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7564 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7566 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7567 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7569 if (tg == &root_task_group)
7573 * Ensure we have at some amount of bandwidth every period. This is
7574 * to prevent reaching a state of large arrears when throttled via
7575 * entity_tick() resulting in prolonged exit starvation.
7577 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7581 * Likewise, bound things on the otherside by preventing insane quota
7582 * periods. This also allows us to normalize in computing quota
7585 if (period > max_cfs_quota_period)
7588 mutex_lock(&cfs_constraints_mutex);
7589 ret = __cfs_schedulable(tg, period, quota);
7593 runtime_enabled = quota != RUNTIME_INF;
7594 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7595 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7596 raw_spin_lock_irq(&cfs_b->lock);
7597 cfs_b->period = ns_to_ktime(period);
7598 cfs_b->quota = quota;
7600 __refill_cfs_bandwidth_runtime(cfs_b);
7601 /* restart the period timer (if active) to handle new period expiry */
7602 if (runtime_enabled && cfs_b->timer_active) {
7603 /* force a reprogram */
7604 cfs_b->timer_active = 0;
7605 __start_cfs_bandwidth(cfs_b);
7607 raw_spin_unlock_irq(&cfs_b->lock);
7609 for_each_possible_cpu(i) {
7610 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7611 struct rq *rq = cfs_rq->rq;
7613 raw_spin_lock_irq(&rq->lock);
7614 cfs_rq->runtime_enabled = runtime_enabled;
7615 cfs_rq->runtime_remaining = 0;
7617 if (cfs_rq->throttled)
7618 unthrottle_cfs_rq(cfs_rq);
7619 raw_spin_unlock_irq(&rq->lock);
7622 mutex_unlock(&cfs_constraints_mutex);
7627 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7631 period = ktime_to_ns(tg->cfs_bandwidth.period);
7632 if (cfs_quota_us < 0)
7633 quota = RUNTIME_INF;
7635 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7637 return tg_set_cfs_bandwidth(tg, period, quota);
7640 long tg_get_cfs_quota(struct task_group *tg)
7644 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7647 quota_us = tg->cfs_bandwidth.quota;
7648 do_div(quota_us, NSEC_PER_USEC);
7653 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7657 period = (u64)cfs_period_us * NSEC_PER_USEC;
7658 quota = tg->cfs_bandwidth.quota;
7660 return tg_set_cfs_bandwidth(tg, period, quota);
7663 long tg_get_cfs_period(struct task_group *tg)
7667 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7668 do_div(cfs_period_us, NSEC_PER_USEC);
7670 return cfs_period_us;
7673 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7675 return tg_get_cfs_quota(cgroup_tg(cgrp));
7678 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7681 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7684 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7686 return tg_get_cfs_period(cgroup_tg(cgrp));
7689 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7692 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7695 struct cfs_schedulable_data {
7696 struct task_group *tg;
7701 * normalize group quota/period to be quota/max_period
7702 * note: units are usecs
7704 static u64 normalize_cfs_quota(struct task_group *tg,
7705 struct cfs_schedulable_data *d)
7713 period = tg_get_cfs_period(tg);
7714 quota = tg_get_cfs_quota(tg);
7717 /* note: these should typically be equivalent */
7718 if (quota == RUNTIME_INF || quota == -1)
7721 return to_ratio(period, quota);
7724 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7726 struct cfs_schedulable_data *d = data;
7727 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7728 s64 quota = 0, parent_quota = -1;
7731 quota = RUNTIME_INF;
7733 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7735 quota = normalize_cfs_quota(tg, d);
7736 parent_quota = parent_b->hierarchal_quota;
7739 * ensure max(child_quota) <= parent_quota, inherit when no
7742 if (quota == RUNTIME_INF)
7743 quota = parent_quota;
7744 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7747 cfs_b->hierarchal_quota = quota;
7752 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7755 struct cfs_schedulable_data data = {
7761 if (quota != RUNTIME_INF) {
7762 do_div(data.period, NSEC_PER_USEC);
7763 do_div(data.quota, NSEC_PER_USEC);
7767 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7773 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7774 struct cgroup_map_cb *cb)
7776 struct task_group *tg = cgroup_tg(cgrp);
7777 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7779 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7780 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7781 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7785 #endif /* CONFIG_CFS_BANDWIDTH */
7786 #endif /* CONFIG_FAIR_GROUP_SCHED */
7788 #ifdef CONFIG_RT_GROUP_SCHED
7789 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7792 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7795 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7797 return sched_group_rt_runtime(cgroup_tg(cgrp));
7800 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7803 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7806 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7808 return sched_group_rt_period(cgroup_tg(cgrp));
7810 #endif /* CONFIG_RT_GROUP_SCHED */
7812 static struct cftype cpu_files[] = {
7813 #ifdef CONFIG_FAIR_GROUP_SCHED
7816 .read_u64 = cpu_shares_read_u64,
7817 .write_u64 = cpu_shares_write_u64,
7820 #ifdef CONFIG_CFS_BANDWIDTH
7822 .name = "cfs_quota_us",
7823 .read_s64 = cpu_cfs_quota_read_s64,
7824 .write_s64 = cpu_cfs_quota_write_s64,
7827 .name = "cfs_period_us",
7828 .read_u64 = cpu_cfs_period_read_u64,
7829 .write_u64 = cpu_cfs_period_write_u64,
7833 .read_map = cpu_stats_show,
7836 #ifdef CONFIG_RT_GROUP_SCHED
7838 .name = "rt_runtime_us",
7839 .read_s64 = cpu_rt_runtime_read,
7840 .write_s64 = cpu_rt_runtime_write,
7843 .name = "rt_period_us",
7844 .read_u64 = cpu_rt_period_read_uint,
7845 .write_u64 = cpu_rt_period_write_uint,
7851 struct cgroup_subsys cpu_cgroup_subsys = {
7853 .create = cpu_cgroup_create,
7854 .destroy = cpu_cgroup_destroy,
7855 .can_attach = cpu_cgroup_can_attach,
7856 .attach = cpu_cgroup_attach,
7857 .exit = cpu_cgroup_exit,
7858 .subsys_id = cpu_cgroup_subsys_id,
7859 .base_cftypes = cpu_files,
7863 #endif /* CONFIG_CGROUP_SCHED */
7865 #ifdef CONFIG_CGROUP_CPUACCT
7868 * CPU accounting code for task groups.
7870 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7871 * (balbir@in.ibm.com).
7874 struct cpuacct root_cpuacct;
7876 /* create a new cpu accounting group */
7877 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7882 return &root_cpuacct.css;
7884 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7888 ca->cpuusage = alloc_percpu(u64);
7892 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7894 goto out_free_cpuusage;
7899 free_percpu(ca->cpuusage);
7903 return ERR_PTR(-ENOMEM);
7906 /* destroy an existing cpu accounting group */
7907 static void cpuacct_destroy(struct cgroup *cgrp)
7909 struct cpuacct *ca = cgroup_ca(cgrp);
7911 free_percpu(ca->cpustat);
7912 free_percpu(ca->cpuusage);
7916 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7918 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7921 #ifndef CONFIG_64BIT
7923 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7925 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7927 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7935 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7937 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7939 #ifndef CONFIG_64BIT
7941 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7943 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7945 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7951 /* return total cpu usage (in nanoseconds) of a group */
7952 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7954 struct cpuacct *ca = cgroup_ca(cgrp);
7955 u64 totalcpuusage = 0;
7958 for_each_present_cpu(i)
7959 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7961 return totalcpuusage;
7964 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7967 struct cpuacct *ca = cgroup_ca(cgrp);
7976 for_each_present_cpu(i)
7977 cpuacct_cpuusage_write(ca, i, 0);
7983 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7986 struct cpuacct *ca = cgroup_ca(cgroup);
7990 for_each_present_cpu(i) {
7991 percpu = cpuacct_cpuusage_read(ca, i);
7992 seq_printf(m, "%llu ", (unsigned long long) percpu);
7994 seq_printf(m, "\n");
7998 static const char *cpuacct_stat_desc[] = {
7999 [CPUACCT_STAT_USER] = "user",
8000 [CPUACCT_STAT_SYSTEM] = "system",
8003 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8004 struct cgroup_map_cb *cb)
8006 struct cpuacct *ca = cgroup_ca(cgrp);
8010 for_each_online_cpu(cpu) {
8011 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8012 val += kcpustat->cpustat[CPUTIME_USER];
8013 val += kcpustat->cpustat[CPUTIME_NICE];
8015 val = cputime64_to_clock_t(val);
8016 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8019 for_each_online_cpu(cpu) {
8020 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8021 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8022 val += kcpustat->cpustat[CPUTIME_IRQ];
8023 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8026 val = cputime64_to_clock_t(val);
8027 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8032 static struct cftype files[] = {
8035 .read_u64 = cpuusage_read,
8036 .write_u64 = cpuusage_write,
8039 .name = "usage_percpu",
8040 .read_seq_string = cpuacct_percpu_seq_read,
8044 .read_map = cpuacct_stats_show,
8050 * charge this task's execution time to its accounting group.
8052 * called with rq->lock held.
8054 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8059 if (unlikely(!cpuacct_subsys.active))
8062 cpu = task_cpu(tsk);
8068 for (; ca; ca = parent_ca(ca)) {
8069 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8070 *cpuusage += cputime;
8076 struct cgroup_subsys cpuacct_subsys = {
8078 .create = cpuacct_create,
8079 .destroy = cpuacct_destroy,
8080 .subsys_id = cpuacct_subsys_id,
8081 .base_cftypes = files,
8083 #endif /* CONFIG_CGROUP_CPUACCT */
8085 void dump_cpu_task(int cpu)
8087 pr_info("Task dump for CPU %d:\n", cpu);
8088 sched_show_task(cpu_curr(cpu));