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 if (p->sched_class->migrate_task_rq)
957 p->sched_class->migrate_task_rq(p, new_cpu);
958 p->se.nr_migrations++;
959 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
962 __set_task_cpu(p, new_cpu);
965 struct migration_arg {
966 struct task_struct *task;
970 static int migration_cpu_stop(void *data);
973 * wait_task_inactive - wait for a thread to unschedule.
975 * If @match_state is nonzero, it's the @p->state value just checked and
976 * not expected to change. If it changes, i.e. @p might have woken up,
977 * then return zero. When we succeed in waiting for @p to be off its CPU,
978 * we return a positive number (its total switch count). If a second call
979 * a short while later returns the same number, the caller can be sure that
980 * @p has remained unscheduled the whole time.
982 * The caller must ensure that the task *will* unschedule sometime soon,
983 * else this function might spin for a *long* time. This function can't
984 * be called with interrupts off, or it may introduce deadlock with
985 * smp_call_function() if an IPI is sent by the same process we are
986 * waiting to become inactive.
988 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
997 * We do the initial early heuristics without holding
998 * any task-queue locks at all. We'll only try to get
999 * the runqueue lock when things look like they will
1005 * If the task is actively running on another CPU
1006 * still, just relax and busy-wait without holding
1009 * NOTE! Since we don't hold any locks, it's not
1010 * even sure that "rq" stays as the right runqueue!
1011 * But we don't care, since "task_running()" will
1012 * return false if the runqueue has changed and p
1013 * is actually now running somewhere else!
1015 while (task_running(rq, p)) {
1016 if (match_state && unlikely(p->state != match_state))
1022 * Ok, time to look more closely! We need the rq
1023 * lock now, to be *sure*. If we're wrong, we'll
1024 * just go back and repeat.
1026 rq = task_rq_lock(p, &flags);
1027 trace_sched_wait_task(p);
1028 running = task_running(rq, p);
1031 if (!match_state || p->state == match_state)
1032 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1033 task_rq_unlock(rq, p, &flags);
1036 * If it changed from the expected state, bail out now.
1038 if (unlikely(!ncsw))
1042 * Was it really running after all now that we
1043 * checked with the proper locks actually held?
1045 * Oops. Go back and try again..
1047 if (unlikely(running)) {
1053 * It's not enough that it's not actively running,
1054 * it must be off the runqueue _entirely_, and not
1057 * So if it was still runnable (but just not actively
1058 * running right now), it's preempted, and we should
1059 * yield - it could be a while.
1061 if (unlikely(on_rq)) {
1062 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1064 set_current_state(TASK_UNINTERRUPTIBLE);
1065 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1070 * Ahh, all good. It wasn't running, and it wasn't
1071 * runnable, which means that it will never become
1072 * running in the future either. We're all done!
1081 * kick_process - kick a running thread to enter/exit the kernel
1082 * @p: the to-be-kicked thread
1084 * Cause a process which is running on another CPU to enter
1085 * kernel-mode, without any delay. (to get signals handled.)
1087 * NOTE: this function doesn't have to take the runqueue lock,
1088 * because all it wants to ensure is that the remote task enters
1089 * the kernel. If the IPI races and the task has been migrated
1090 * to another CPU then no harm is done and the purpose has been
1093 void kick_process(struct task_struct *p)
1099 if ((cpu != smp_processor_id()) && task_curr(p))
1100 smp_send_reschedule(cpu);
1103 EXPORT_SYMBOL_GPL(kick_process);
1104 #endif /* CONFIG_SMP */
1108 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1110 static int select_fallback_rq(int cpu, struct task_struct *p)
1112 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
1113 enum { cpuset, possible, fail } state = cpuset;
1116 /* Look for allowed, online CPU in same node. */
1117 for_each_cpu(dest_cpu, nodemask) {
1118 if (!cpu_online(dest_cpu))
1120 if (!cpu_active(dest_cpu))
1122 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1127 /* Any allowed, online CPU? */
1128 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1129 if (!cpu_online(dest_cpu))
1131 if (!cpu_active(dest_cpu))
1138 /* No more Mr. Nice Guy. */
1139 cpuset_cpus_allowed_fallback(p);
1144 do_set_cpus_allowed(p, cpu_possible_mask);
1155 if (state != cpuset) {
1157 * Don't tell them about moving exiting tasks or
1158 * kernel threads (both mm NULL), since they never
1161 if (p->mm && printk_ratelimit()) {
1162 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1163 task_pid_nr(p), p->comm, cpu);
1171 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1174 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1176 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1179 * In order not to call set_task_cpu() on a blocking task we need
1180 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1183 * Since this is common to all placement strategies, this lives here.
1185 * [ this allows ->select_task() to simply return task_cpu(p) and
1186 * not worry about this generic constraint ]
1188 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1190 cpu = select_fallback_rq(task_cpu(p), p);
1195 static void update_avg(u64 *avg, u64 sample)
1197 s64 diff = sample - *avg;
1203 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1205 #ifdef CONFIG_SCHEDSTATS
1206 struct rq *rq = this_rq();
1209 int this_cpu = smp_processor_id();
1211 if (cpu == this_cpu) {
1212 schedstat_inc(rq, ttwu_local);
1213 schedstat_inc(p, se.statistics.nr_wakeups_local);
1215 struct sched_domain *sd;
1217 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1219 for_each_domain(this_cpu, sd) {
1220 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1221 schedstat_inc(sd, ttwu_wake_remote);
1228 if (wake_flags & WF_MIGRATED)
1229 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1231 #endif /* CONFIG_SMP */
1233 schedstat_inc(rq, ttwu_count);
1234 schedstat_inc(p, se.statistics.nr_wakeups);
1236 if (wake_flags & WF_SYNC)
1237 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1239 #endif /* CONFIG_SCHEDSTATS */
1242 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1244 activate_task(rq, p, en_flags);
1247 /* if a worker is waking up, notify workqueue */
1248 if (p->flags & PF_WQ_WORKER)
1249 wq_worker_waking_up(p, cpu_of(rq));
1253 * Mark the task runnable and perform wakeup-preemption.
1256 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1258 trace_sched_wakeup(p, true);
1259 check_preempt_curr(rq, p, wake_flags);
1261 p->state = TASK_RUNNING;
1263 if (p->sched_class->task_woken)
1264 p->sched_class->task_woken(rq, p);
1266 if (rq->idle_stamp) {
1267 u64 delta = rq->clock - rq->idle_stamp;
1268 u64 max = 2*sysctl_sched_migration_cost;
1273 update_avg(&rq->avg_idle, delta);
1280 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1283 if (p->sched_contributes_to_load)
1284 rq->nr_uninterruptible--;
1287 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1288 ttwu_do_wakeup(rq, p, wake_flags);
1292 * Called in case the task @p isn't fully descheduled from its runqueue,
1293 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1294 * since all we need to do is flip p->state to TASK_RUNNING, since
1295 * the task is still ->on_rq.
1297 static int ttwu_remote(struct task_struct *p, int wake_flags)
1302 rq = __task_rq_lock(p);
1304 ttwu_do_wakeup(rq, p, wake_flags);
1307 __task_rq_unlock(rq);
1313 static void sched_ttwu_pending(void)
1315 struct rq *rq = this_rq();
1316 struct llist_node *llist = llist_del_all(&rq->wake_list);
1317 struct task_struct *p;
1319 raw_spin_lock(&rq->lock);
1322 p = llist_entry(llist, struct task_struct, wake_entry);
1323 llist = llist_next(llist);
1324 ttwu_do_activate(rq, p, 0);
1327 raw_spin_unlock(&rq->lock);
1330 void scheduler_ipi(void)
1332 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1336 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1337 * traditionally all their work was done from the interrupt return
1338 * path. Now that we actually do some work, we need to make sure
1341 * Some archs already do call them, luckily irq_enter/exit nest
1344 * Arguably we should visit all archs and update all handlers,
1345 * however a fair share of IPIs are still resched only so this would
1346 * somewhat pessimize the simple resched case.
1349 sched_ttwu_pending();
1352 * Check if someone kicked us for doing the nohz idle load balance.
1354 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1355 this_rq()->idle_balance = 1;
1356 raise_softirq_irqoff(SCHED_SOFTIRQ);
1361 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1363 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1364 smp_send_reschedule(cpu);
1367 bool cpus_share_cache(int this_cpu, int that_cpu)
1369 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1371 #endif /* CONFIG_SMP */
1373 static void ttwu_queue(struct task_struct *p, int cpu)
1375 struct rq *rq = cpu_rq(cpu);
1377 #if defined(CONFIG_SMP)
1378 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1379 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1380 ttwu_queue_remote(p, cpu);
1385 raw_spin_lock(&rq->lock);
1386 ttwu_do_activate(rq, p, 0);
1387 raw_spin_unlock(&rq->lock);
1391 * try_to_wake_up - wake up a thread
1392 * @p: the thread to be awakened
1393 * @state: the mask of task states that can be woken
1394 * @wake_flags: wake modifier flags (WF_*)
1396 * Put it on the run-queue if it's not already there. The "current"
1397 * thread is always on the run-queue (except when the actual
1398 * re-schedule is in progress), and as such you're allowed to do
1399 * the simpler "current->state = TASK_RUNNING" to mark yourself
1400 * runnable without the overhead of this.
1402 * Returns %true if @p was woken up, %false if it was already running
1403 * or @state didn't match @p's state.
1406 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1408 unsigned long flags;
1409 int cpu, success = 0;
1412 raw_spin_lock_irqsave(&p->pi_lock, flags);
1413 if (!(p->state & state))
1416 success = 1; /* we're going to change ->state */
1419 if (p->on_rq && ttwu_remote(p, wake_flags))
1424 * If the owning (remote) cpu is still in the middle of schedule() with
1425 * this task as prev, wait until its done referencing the task.
1430 * Pairs with the smp_wmb() in finish_lock_switch().
1434 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1435 p->state = TASK_WAKING;
1437 if (p->sched_class->task_waking)
1438 p->sched_class->task_waking(p);
1440 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1441 if (task_cpu(p) != cpu) {
1442 wake_flags |= WF_MIGRATED;
1443 set_task_cpu(p, cpu);
1445 #endif /* CONFIG_SMP */
1449 ttwu_stat(p, cpu, wake_flags);
1451 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1457 * try_to_wake_up_local - try to wake up a local task with rq lock held
1458 * @p: the thread to be awakened
1460 * Put @p on the run-queue if it's not already there. The caller must
1461 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1464 static void try_to_wake_up_local(struct task_struct *p)
1466 struct rq *rq = task_rq(p);
1468 BUG_ON(rq != this_rq());
1469 BUG_ON(p == current);
1470 lockdep_assert_held(&rq->lock);
1472 if (!raw_spin_trylock(&p->pi_lock)) {
1473 raw_spin_unlock(&rq->lock);
1474 raw_spin_lock(&p->pi_lock);
1475 raw_spin_lock(&rq->lock);
1478 if (!(p->state & TASK_NORMAL))
1482 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1484 ttwu_do_wakeup(rq, p, 0);
1485 ttwu_stat(p, smp_processor_id(), 0);
1487 raw_spin_unlock(&p->pi_lock);
1491 * wake_up_process - Wake up a specific process
1492 * @p: The process to be woken up.
1494 * Attempt to wake up the nominated process and move it to the set of runnable
1495 * processes. Returns 1 if the process was woken up, 0 if it was already
1498 * It may be assumed that this function implies a write memory barrier before
1499 * changing the task state if and only if any tasks are woken up.
1501 int wake_up_process(struct task_struct *p)
1503 return try_to_wake_up(p, TASK_ALL, 0);
1505 EXPORT_SYMBOL(wake_up_process);
1507 int wake_up_state(struct task_struct *p, unsigned int state)
1509 return try_to_wake_up(p, state, 0);
1513 * Perform scheduler related setup for a newly forked process p.
1514 * p is forked by current.
1516 * __sched_fork() is basic setup used by init_idle() too:
1518 static void __sched_fork(struct task_struct *p)
1523 p->se.exec_start = 0;
1524 p->se.sum_exec_runtime = 0;
1525 p->se.prev_sum_exec_runtime = 0;
1526 p->se.nr_migrations = 0;
1528 INIT_LIST_HEAD(&p->se.group_node);
1531 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1532 * removed when useful for applications beyond shares distribution (e.g.
1535 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1536 p->se.avg.runnable_avg_period = 0;
1537 p->se.avg.runnable_avg_sum = 0;
1539 #ifdef CONFIG_SCHEDSTATS
1540 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1543 INIT_LIST_HEAD(&p->rt.run_list);
1545 #ifdef CONFIG_PREEMPT_NOTIFIERS
1546 INIT_HLIST_HEAD(&p->preempt_notifiers);
1551 * fork()/clone()-time setup:
1553 void sched_fork(struct task_struct *p)
1555 unsigned long flags;
1556 int cpu = get_cpu();
1560 * We mark the process as running here. This guarantees that
1561 * nobody will actually run it, and a signal or other external
1562 * event cannot wake it up and insert it on the runqueue either.
1564 p->state = TASK_RUNNING;
1567 * Make sure we do not leak PI boosting priority to the child.
1569 p->prio = current->normal_prio;
1572 * Revert to default priority/policy on fork if requested.
1574 if (unlikely(p->sched_reset_on_fork)) {
1575 if (task_has_rt_policy(p)) {
1576 p->policy = SCHED_NORMAL;
1577 p->static_prio = NICE_TO_PRIO(0);
1579 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1580 p->static_prio = NICE_TO_PRIO(0);
1582 p->prio = p->normal_prio = __normal_prio(p);
1586 * We don't need the reset flag anymore after the fork. It has
1587 * fulfilled its duty:
1589 p->sched_reset_on_fork = 0;
1592 if (!rt_prio(p->prio))
1593 p->sched_class = &fair_sched_class;
1595 if (p->sched_class->task_fork)
1596 p->sched_class->task_fork(p);
1599 * The child is not yet in the pid-hash so no cgroup attach races,
1600 * and the cgroup is pinned to this child due to cgroup_fork()
1601 * is ran before sched_fork().
1603 * Silence PROVE_RCU.
1605 raw_spin_lock_irqsave(&p->pi_lock, flags);
1606 set_task_cpu(p, cpu);
1607 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1609 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1610 if (likely(sched_info_on()))
1611 memset(&p->sched_info, 0, sizeof(p->sched_info));
1613 #if defined(CONFIG_SMP)
1616 #ifdef CONFIG_PREEMPT_COUNT
1617 /* Want to start with kernel preemption disabled. */
1618 task_thread_info(p)->preempt_count = 1;
1621 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1628 * wake_up_new_task - wake up a newly created task for the first time.
1630 * This function will do some initial scheduler statistics housekeeping
1631 * that must be done for every newly created context, then puts the task
1632 * on the runqueue and wakes it.
1634 void wake_up_new_task(struct task_struct *p)
1636 unsigned long flags;
1639 raw_spin_lock_irqsave(&p->pi_lock, flags);
1642 * Fork balancing, do it here and not earlier because:
1643 * - cpus_allowed can change in the fork path
1644 * - any previously selected cpu might disappear through hotplug
1646 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1649 rq = __task_rq_lock(p);
1650 activate_task(rq, p, 0);
1652 trace_sched_wakeup_new(p, true);
1653 check_preempt_curr(rq, p, WF_FORK);
1655 if (p->sched_class->task_woken)
1656 p->sched_class->task_woken(rq, p);
1658 task_rq_unlock(rq, p, &flags);
1661 #ifdef CONFIG_PREEMPT_NOTIFIERS
1664 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1665 * @notifier: notifier struct to register
1667 void preempt_notifier_register(struct preempt_notifier *notifier)
1669 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1671 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1674 * preempt_notifier_unregister - no longer interested in preemption notifications
1675 * @notifier: notifier struct to unregister
1677 * This is safe to call from within a preemption notifier.
1679 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1681 hlist_del(¬ifier->link);
1683 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1685 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
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_in(notifier, raw_smp_processor_id());
1695 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1696 struct task_struct *next)
1698 struct preempt_notifier *notifier;
1699 struct hlist_node *node;
1701 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
1702 notifier->ops->sched_out(notifier, next);
1705 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1707 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1712 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1713 struct task_struct *next)
1717 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1720 * prepare_task_switch - prepare to switch tasks
1721 * @rq: the runqueue preparing to switch
1722 * @prev: the current task that is being switched out
1723 * @next: the task we are going to switch to.
1725 * This is called with the rq lock held and interrupts off. It must
1726 * be paired with a subsequent finish_task_switch after the context
1729 * prepare_task_switch sets up locking and calls architecture specific
1733 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1734 struct task_struct *next)
1736 trace_sched_switch(prev, next);
1737 sched_info_switch(prev, next);
1738 perf_event_task_sched_out(prev, next);
1739 fire_sched_out_preempt_notifiers(prev, next);
1740 prepare_lock_switch(rq, next);
1741 prepare_arch_switch(next);
1745 * finish_task_switch - clean up after a task-switch
1746 * @rq: runqueue associated with task-switch
1747 * @prev: the thread we just switched away from.
1749 * finish_task_switch must be called after the context switch, paired
1750 * with a prepare_task_switch call before the context switch.
1751 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1752 * and do any other architecture-specific cleanup actions.
1754 * Note that we may have delayed dropping an mm in context_switch(). If
1755 * so, we finish that here outside of the runqueue lock. (Doing it
1756 * with the lock held can cause deadlocks; see schedule() for
1759 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1760 __releases(rq->lock)
1762 struct mm_struct *mm = rq->prev_mm;
1768 * A task struct has one reference for the use as "current".
1769 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1770 * schedule one last time. The schedule call will never return, and
1771 * the scheduled task must drop that reference.
1772 * The test for TASK_DEAD must occur while the runqueue locks are
1773 * still held, otherwise prev could be scheduled on another cpu, die
1774 * there before we look at prev->state, and then the reference would
1776 * Manfred Spraul <manfred@colorfullife.com>
1778 prev_state = prev->state;
1779 vtime_task_switch(prev);
1780 finish_arch_switch(prev);
1781 perf_event_task_sched_in(prev, current);
1782 finish_lock_switch(rq, prev);
1783 finish_arch_post_lock_switch();
1785 fire_sched_in_preempt_notifiers(current);
1788 if (unlikely(prev_state == TASK_DEAD)) {
1790 * Remove function-return probe instances associated with this
1791 * task and put them back on the free list.
1793 kprobe_flush_task(prev);
1794 put_task_struct(prev);
1800 /* assumes rq->lock is held */
1801 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1803 if (prev->sched_class->pre_schedule)
1804 prev->sched_class->pre_schedule(rq, prev);
1807 /* rq->lock is NOT held, but preemption is disabled */
1808 static inline void post_schedule(struct rq *rq)
1810 if (rq->post_schedule) {
1811 unsigned long flags;
1813 raw_spin_lock_irqsave(&rq->lock, flags);
1814 if (rq->curr->sched_class->post_schedule)
1815 rq->curr->sched_class->post_schedule(rq);
1816 raw_spin_unlock_irqrestore(&rq->lock, flags);
1818 rq->post_schedule = 0;
1824 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1828 static inline void post_schedule(struct rq *rq)
1835 * schedule_tail - first thing a freshly forked thread must call.
1836 * @prev: the thread we just switched away from.
1838 asmlinkage void schedule_tail(struct task_struct *prev)
1839 __releases(rq->lock)
1841 struct rq *rq = this_rq();
1843 finish_task_switch(rq, prev);
1846 * FIXME: do we need to worry about rq being invalidated by the
1851 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1852 /* In this case, finish_task_switch does not reenable preemption */
1855 if (current->set_child_tid)
1856 put_user(task_pid_vnr(current), current->set_child_tid);
1860 * context_switch - switch to the new MM and the new
1861 * thread's register state.
1864 context_switch(struct rq *rq, struct task_struct *prev,
1865 struct task_struct *next)
1867 struct mm_struct *mm, *oldmm;
1869 prepare_task_switch(rq, prev, next);
1872 oldmm = prev->active_mm;
1874 * For paravirt, this is coupled with an exit in switch_to to
1875 * combine the page table reload and the switch backend into
1878 arch_start_context_switch(prev);
1881 next->active_mm = oldmm;
1882 atomic_inc(&oldmm->mm_count);
1883 enter_lazy_tlb(oldmm, next);
1885 switch_mm(oldmm, mm, next);
1888 prev->active_mm = NULL;
1889 rq->prev_mm = oldmm;
1892 * Since the runqueue lock will be released by the next
1893 * task (which is an invalid locking op but in the case
1894 * of the scheduler it's an obvious special-case), so we
1895 * do an early lockdep release here:
1897 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1898 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
1901 context_tracking_task_switch(prev, next);
1902 /* Here we just switch the register state and the stack. */
1903 switch_to(prev, next, prev);
1907 * this_rq must be evaluated again because prev may have moved
1908 * CPUs since it called schedule(), thus the 'rq' on its stack
1909 * frame will be invalid.
1911 finish_task_switch(this_rq(), prev);
1915 * nr_running, nr_uninterruptible and nr_context_switches:
1917 * externally visible scheduler statistics: current number of runnable
1918 * threads, current number of uninterruptible-sleeping threads, total
1919 * number of context switches performed since bootup.
1921 unsigned long nr_running(void)
1923 unsigned long i, sum = 0;
1925 for_each_online_cpu(i)
1926 sum += cpu_rq(i)->nr_running;
1931 unsigned long nr_uninterruptible(void)
1933 unsigned long i, sum = 0;
1935 for_each_possible_cpu(i)
1936 sum += cpu_rq(i)->nr_uninterruptible;
1939 * Since we read the counters lockless, it might be slightly
1940 * inaccurate. Do not allow it to go below zero though:
1942 if (unlikely((long)sum < 0))
1948 unsigned long long nr_context_switches(void)
1951 unsigned long long sum = 0;
1953 for_each_possible_cpu(i)
1954 sum += cpu_rq(i)->nr_switches;
1959 unsigned long nr_iowait(void)
1961 unsigned long i, sum = 0;
1963 for_each_possible_cpu(i)
1964 sum += atomic_read(&cpu_rq(i)->nr_iowait);
1969 unsigned long nr_iowait_cpu(int cpu)
1971 struct rq *this = cpu_rq(cpu);
1972 return atomic_read(&this->nr_iowait);
1975 unsigned long this_cpu_load(void)
1977 struct rq *this = this_rq();
1978 return this->cpu_load[0];
1983 * Global load-average calculations
1985 * We take a distributed and async approach to calculating the global load-avg
1986 * in order to minimize overhead.
1988 * The global load average is an exponentially decaying average of nr_running +
1989 * nr_uninterruptible.
1991 * Once every LOAD_FREQ:
1994 * for_each_possible_cpu(cpu)
1995 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
1997 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
1999 * Due to a number of reasons the above turns in the mess below:
2001 * - for_each_possible_cpu() is prohibitively expensive on machines with
2002 * serious number of cpus, therefore we need to take a distributed approach
2003 * to calculating nr_active.
2005 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2006 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2008 * So assuming nr_active := 0 when we start out -- true per definition, we
2009 * can simply take per-cpu deltas and fold those into a global accumulate
2010 * to obtain the same result. See calc_load_fold_active().
2012 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2013 * across the machine, we assume 10 ticks is sufficient time for every
2014 * cpu to have completed this task.
2016 * This places an upper-bound on the IRQ-off latency of the machine. Then
2017 * again, being late doesn't loose the delta, just wrecks the sample.
2019 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2020 * this would add another cross-cpu cacheline miss and atomic operation
2021 * to the wakeup path. Instead we increment on whatever cpu the task ran
2022 * when it went into uninterruptible state and decrement on whatever cpu
2023 * did the wakeup. This means that only the sum of nr_uninterruptible over
2024 * all cpus yields the correct result.
2026 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2029 /* Variables and functions for calc_load */
2030 static atomic_long_t calc_load_tasks;
2031 static unsigned long calc_load_update;
2032 unsigned long avenrun[3];
2033 EXPORT_SYMBOL(avenrun); /* should be removed */
2036 * get_avenrun - get the load average array
2037 * @loads: pointer to dest load array
2038 * @offset: offset to add
2039 * @shift: shift count to shift the result left
2041 * These values are estimates at best, so no need for locking.
2043 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2045 loads[0] = (avenrun[0] + offset) << shift;
2046 loads[1] = (avenrun[1] + offset) << shift;
2047 loads[2] = (avenrun[2] + offset) << shift;
2050 static long calc_load_fold_active(struct rq *this_rq)
2052 long nr_active, delta = 0;
2054 nr_active = this_rq->nr_running;
2055 nr_active += (long) this_rq->nr_uninterruptible;
2057 if (nr_active != this_rq->calc_load_active) {
2058 delta = nr_active - this_rq->calc_load_active;
2059 this_rq->calc_load_active = nr_active;
2066 * a1 = a0 * e + a * (1 - e)
2068 static unsigned long
2069 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2072 load += active * (FIXED_1 - exp);
2073 load += 1UL << (FSHIFT - 1);
2074 return load >> FSHIFT;
2079 * Handle NO_HZ for the global load-average.
2081 * Since the above described distributed algorithm to compute the global
2082 * load-average relies on per-cpu sampling from the tick, it is affected by
2085 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2086 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2087 * when we read the global state.
2089 * Obviously reality has to ruin such a delightfully simple scheme:
2091 * - When we go NO_HZ idle during the window, we can negate our sample
2092 * contribution, causing under-accounting.
2094 * We avoid this by keeping two idle-delta counters and flipping them
2095 * when the window starts, thus separating old and new NO_HZ load.
2097 * The only trick is the slight shift in index flip for read vs write.
2101 * |-|-----------|-|-----------|-|-----------|-|
2102 * r:0 0 1 1 0 0 1 1 0
2103 * w:0 1 1 0 0 1 1 0 0
2105 * This ensures we'll fold the old idle contribution in this window while
2106 * accumlating the new one.
2108 * - When we wake up from NO_HZ idle during the window, we push up our
2109 * contribution, since we effectively move our sample point to a known
2112 * This is solved by pushing the window forward, and thus skipping the
2113 * sample, for this cpu (effectively using the idle-delta for this cpu which
2114 * was in effect at the time the window opened). This also solves the issue
2115 * of having to deal with a cpu having been in NOHZ idle for multiple
2116 * LOAD_FREQ intervals.
2118 * When making the ILB scale, we should try to pull this in as well.
2120 static atomic_long_t calc_load_idle[2];
2121 static int calc_load_idx;
2123 static inline int calc_load_write_idx(void)
2125 int idx = calc_load_idx;
2128 * See calc_global_nohz(), if we observe the new index, we also
2129 * need to observe the new update time.
2134 * If the folding window started, make sure we start writing in the
2137 if (!time_before(jiffies, calc_load_update))
2143 static inline int calc_load_read_idx(void)
2145 return calc_load_idx & 1;
2148 void calc_load_enter_idle(void)
2150 struct rq *this_rq = this_rq();
2154 * We're going into NOHZ mode, if there's any pending delta, fold it
2155 * into the pending idle delta.
2157 delta = calc_load_fold_active(this_rq);
2159 int idx = calc_load_write_idx();
2160 atomic_long_add(delta, &calc_load_idle[idx]);
2164 void calc_load_exit_idle(void)
2166 struct rq *this_rq = this_rq();
2169 * If we're still before the sample window, we're done.
2171 if (time_before(jiffies, this_rq->calc_load_update))
2175 * We woke inside or after the sample window, this means we're already
2176 * accounted through the nohz accounting, so skip the entire deal and
2177 * sync up for the next window.
2179 this_rq->calc_load_update = calc_load_update;
2180 if (time_before(jiffies, this_rq->calc_load_update + 10))
2181 this_rq->calc_load_update += LOAD_FREQ;
2184 static long calc_load_fold_idle(void)
2186 int idx = calc_load_read_idx();
2189 if (atomic_long_read(&calc_load_idle[idx]))
2190 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2196 * fixed_power_int - compute: x^n, in O(log n) time
2198 * @x: base of the power
2199 * @frac_bits: fractional bits of @x
2200 * @n: power to raise @x to.
2202 * By exploiting the relation between the definition of the natural power
2203 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2204 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2205 * (where: n_i \elem {0, 1}, the binary vector representing n),
2206 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2207 * of course trivially computable in O(log_2 n), the length of our binary
2210 static unsigned long
2211 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2213 unsigned long result = 1UL << frac_bits;
2218 result += 1UL << (frac_bits - 1);
2219 result >>= frac_bits;
2225 x += 1UL << (frac_bits - 1);
2233 * a1 = a0 * e + a * (1 - e)
2235 * a2 = a1 * e + a * (1 - e)
2236 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2237 * = a0 * e^2 + a * (1 - e) * (1 + e)
2239 * a3 = a2 * e + a * (1 - e)
2240 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2241 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2245 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2246 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2247 * = a0 * e^n + a * (1 - e^n)
2249 * [1] application of the geometric series:
2252 * S_n := \Sum x^i = -------------
2255 static unsigned long
2256 calc_load_n(unsigned long load, unsigned long exp,
2257 unsigned long active, unsigned int n)
2260 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2264 * NO_HZ can leave us missing all per-cpu ticks calling
2265 * calc_load_account_active(), but since an idle CPU folds its delta into
2266 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2267 * in the pending idle delta if our idle period crossed a load cycle boundary.
2269 * Once we've updated the global active value, we need to apply the exponential
2270 * weights adjusted to the number of cycles missed.
2272 static void calc_global_nohz(void)
2274 long delta, active, n;
2276 if (!time_before(jiffies, calc_load_update + 10)) {
2278 * Catch-up, fold however many we are behind still
2280 delta = jiffies - calc_load_update - 10;
2281 n = 1 + (delta / LOAD_FREQ);
2283 active = atomic_long_read(&calc_load_tasks);
2284 active = active > 0 ? active * FIXED_1 : 0;
2286 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2287 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2288 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2290 calc_load_update += n * LOAD_FREQ;
2294 * Flip the idle index...
2296 * Make sure we first write the new time then flip the index, so that
2297 * calc_load_write_idx() will see the new time when it reads the new
2298 * index, this avoids a double flip messing things up.
2303 #else /* !CONFIG_NO_HZ */
2305 static inline long calc_load_fold_idle(void) { return 0; }
2306 static inline void calc_global_nohz(void) { }
2308 #endif /* CONFIG_NO_HZ */
2311 * calc_load - update the avenrun load estimates 10 ticks after the
2312 * CPUs have updated calc_load_tasks.
2314 void calc_global_load(unsigned long ticks)
2318 if (time_before(jiffies, calc_load_update + 10))
2322 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2324 delta = calc_load_fold_idle();
2326 atomic_long_add(delta, &calc_load_tasks);
2328 active = atomic_long_read(&calc_load_tasks);
2329 active = active > 0 ? active * FIXED_1 : 0;
2331 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2332 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2333 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2335 calc_load_update += LOAD_FREQ;
2338 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2344 * Called from update_cpu_load() to periodically update this CPU's
2347 static void calc_load_account_active(struct rq *this_rq)
2351 if (time_before(jiffies, this_rq->calc_load_update))
2354 delta = calc_load_fold_active(this_rq);
2356 atomic_long_add(delta, &calc_load_tasks);
2358 this_rq->calc_load_update += LOAD_FREQ;
2362 * End of global load-average stuff
2366 * The exact cpuload at various idx values, calculated at every tick would be
2367 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2369 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2370 * on nth tick when cpu may be busy, then we have:
2371 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2372 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2374 * decay_load_missed() below does efficient calculation of
2375 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2376 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2378 * The calculation is approximated on a 128 point scale.
2379 * degrade_zero_ticks is the number of ticks after which load at any
2380 * particular idx is approximated to be zero.
2381 * degrade_factor is a precomputed table, a row for each load idx.
2382 * Each column corresponds to degradation factor for a power of two ticks,
2383 * based on 128 point scale.
2385 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2386 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2388 * With this power of 2 load factors, we can degrade the load n times
2389 * by looking at 1 bits in n and doing as many mult/shift instead of
2390 * n mult/shifts needed by the exact degradation.
2392 #define DEGRADE_SHIFT 7
2393 static const unsigned char
2394 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2395 static const unsigned char
2396 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2397 {0, 0, 0, 0, 0, 0, 0, 0},
2398 {64, 32, 8, 0, 0, 0, 0, 0},
2399 {96, 72, 40, 12, 1, 0, 0},
2400 {112, 98, 75, 43, 15, 1, 0},
2401 {120, 112, 98, 76, 45, 16, 2} };
2404 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2405 * would be when CPU is idle and so we just decay the old load without
2406 * adding any new load.
2408 static unsigned long
2409 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2413 if (!missed_updates)
2416 if (missed_updates >= degrade_zero_ticks[idx])
2420 return load >> missed_updates;
2422 while (missed_updates) {
2423 if (missed_updates % 2)
2424 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2426 missed_updates >>= 1;
2433 * Update rq->cpu_load[] statistics. This function is usually called every
2434 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2435 * every tick. We fix it up based on jiffies.
2437 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2438 unsigned long pending_updates)
2442 this_rq->nr_load_updates++;
2444 /* Update our load: */
2445 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2446 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2447 unsigned long old_load, new_load;
2449 /* scale is effectively 1 << i now, and >> i divides by scale */
2451 old_load = this_rq->cpu_load[i];
2452 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2453 new_load = this_load;
2455 * Round up the averaging division if load is increasing. This
2456 * prevents us from getting stuck on 9 if the load is 10, for
2459 if (new_load > old_load)
2460 new_load += scale - 1;
2462 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2465 sched_avg_update(this_rq);
2470 * There is no sane way to deal with nohz on smp when using jiffies because the
2471 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2472 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2474 * Therefore we cannot use the delta approach from the regular tick since that
2475 * would seriously skew the load calculation. However we'll make do for those
2476 * updates happening while idle (nohz_idle_balance) or coming out of idle
2477 * (tick_nohz_idle_exit).
2479 * This means we might still be one tick off for nohz periods.
2483 * Called from nohz_idle_balance() to update the load ratings before doing the
2486 void update_idle_cpu_load(struct rq *this_rq)
2488 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2489 unsigned long load = this_rq->load.weight;
2490 unsigned long pending_updates;
2493 * bail if there's load or we're actually up-to-date.
2495 if (load || curr_jiffies == this_rq->last_load_update_tick)
2498 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2499 this_rq->last_load_update_tick = curr_jiffies;
2501 __update_cpu_load(this_rq, load, pending_updates);
2505 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2507 void update_cpu_load_nohz(void)
2509 struct rq *this_rq = this_rq();
2510 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2511 unsigned long pending_updates;
2513 if (curr_jiffies == this_rq->last_load_update_tick)
2516 raw_spin_lock(&this_rq->lock);
2517 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2518 if (pending_updates) {
2519 this_rq->last_load_update_tick = curr_jiffies;
2521 * We were idle, this means load 0, the current load might be
2522 * !0 due to remote wakeups and the sort.
2524 __update_cpu_load(this_rq, 0, pending_updates);
2526 raw_spin_unlock(&this_rq->lock);
2528 #endif /* CONFIG_NO_HZ */
2531 * Called from scheduler_tick()
2533 static void update_cpu_load_active(struct rq *this_rq)
2536 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2538 this_rq->last_load_update_tick = jiffies;
2539 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2541 calc_load_account_active(this_rq);
2547 * sched_exec - execve() is a valuable balancing opportunity, because at
2548 * this point the task has the smallest effective memory and cache footprint.
2550 void sched_exec(void)
2552 struct task_struct *p = current;
2553 unsigned long flags;
2556 raw_spin_lock_irqsave(&p->pi_lock, flags);
2557 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2558 if (dest_cpu == smp_processor_id())
2561 if (likely(cpu_active(dest_cpu))) {
2562 struct migration_arg arg = { p, dest_cpu };
2564 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2565 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2569 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2574 DEFINE_PER_CPU(struct kernel_stat, kstat);
2575 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2577 EXPORT_PER_CPU_SYMBOL(kstat);
2578 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2581 * Return any ns on the sched_clock that have not yet been accounted in
2582 * @p in case that task is currently running.
2584 * Called with task_rq_lock() held on @rq.
2586 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2590 if (task_current(rq, p)) {
2591 update_rq_clock(rq);
2592 ns = rq->clock_task - p->se.exec_start;
2600 unsigned long long task_delta_exec(struct task_struct *p)
2602 unsigned long flags;
2606 rq = task_rq_lock(p, &flags);
2607 ns = do_task_delta_exec(p, rq);
2608 task_rq_unlock(rq, p, &flags);
2614 * Return accounted runtime for the task.
2615 * In case the task is currently running, return the runtime plus current's
2616 * pending runtime that have not been accounted yet.
2618 unsigned long long task_sched_runtime(struct task_struct *p)
2620 unsigned long flags;
2624 rq = task_rq_lock(p, &flags);
2625 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2626 task_rq_unlock(rq, p, &flags);
2632 * This function gets called by the timer code, with HZ frequency.
2633 * We call it with interrupts disabled.
2635 void scheduler_tick(void)
2637 int cpu = smp_processor_id();
2638 struct rq *rq = cpu_rq(cpu);
2639 struct task_struct *curr = rq->curr;
2643 raw_spin_lock(&rq->lock);
2644 update_rq_clock(rq);
2645 update_cpu_load_active(rq);
2646 curr->sched_class->task_tick(rq, curr, 0);
2647 raw_spin_unlock(&rq->lock);
2649 perf_event_task_tick();
2652 rq->idle_balance = idle_cpu(cpu);
2653 trigger_load_balance(rq, cpu);
2657 notrace unsigned long get_parent_ip(unsigned long addr)
2659 if (in_lock_functions(addr)) {
2660 addr = CALLER_ADDR2;
2661 if (in_lock_functions(addr))
2662 addr = CALLER_ADDR3;
2667 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2668 defined(CONFIG_PREEMPT_TRACER))
2670 void __kprobes add_preempt_count(int val)
2672 #ifdef CONFIG_DEBUG_PREEMPT
2676 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2679 preempt_count() += val;
2680 #ifdef CONFIG_DEBUG_PREEMPT
2682 * Spinlock count overflowing soon?
2684 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2687 if (preempt_count() == val)
2688 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2690 EXPORT_SYMBOL(add_preempt_count);
2692 void __kprobes sub_preempt_count(int val)
2694 #ifdef CONFIG_DEBUG_PREEMPT
2698 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2701 * Is the spinlock portion underflowing?
2703 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2704 !(preempt_count() & PREEMPT_MASK)))
2708 if (preempt_count() == val)
2709 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2710 preempt_count() -= val;
2712 EXPORT_SYMBOL(sub_preempt_count);
2717 * Print scheduling while atomic bug:
2719 static noinline void __schedule_bug(struct task_struct *prev)
2721 if (oops_in_progress)
2724 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2725 prev->comm, prev->pid, preempt_count());
2727 debug_show_held_locks(prev);
2729 if (irqs_disabled())
2730 print_irqtrace_events(prev);
2732 add_taint(TAINT_WARN);
2736 * Various schedule()-time debugging checks and statistics:
2738 static inline void schedule_debug(struct task_struct *prev)
2741 * Test if we are atomic. Since do_exit() needs to call into
2742 * schedule() atomically, we ignore that path for now.
2743 * Otherwise, whine if we are scheduling when we should not be.
2745 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2746 __schedule_bug(prev);
2749 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2751 schedstat_inc(this_rq(), sched_count);
2754 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2756 if (prev->on_rq || rq->skip_clock_update < 0)
2757 update_rq_clock(rq);
2758 prev->sched_class->put_prev_task(rq, prev);
2762 * Pick up the highest-prio task:
2764 static inline struct task_struct *
2765 pick_next_task(struct rq *rq)
2767 const struct sched_class *class;
2768 struct task_struct *p;
2771 * Optimization: we know that if all tasks are in
2772 * the fair class we can call that function directly:
2774 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2775 p = fair_sched_class.pick_next_task(rq);
2780 for_each_class(class) {
2781 p = class->pick_next_task(rq);
2786 BUG(); /* the idle class will always have a runnable task */
2790 * __schedule() is the main scheduler function.
2792 * The main means of driving the scheduler and thus entering this function are:
2794 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2796 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2797 * paths. For example, see arch/x86/entry_64.S.
2799 * To drive preemption between tasks, the scheduler sets the flag in timer
2800 * interrupt handler scheduler_tick().
2802 * 3. Wakeups don't really cause entry into schedule(). They add a
2803 * task to the run-queue and that's it.
2805 * Now, if the new task added to the run-queue preempts the current
2806 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2807 * called on the nearest possible occasion:
2809 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2811 * - in syscall or exception context, at the next outmost
2812 * preempt_enable(). (this might be as soon as the wake_up()'s
2815 * - in IRQ context, return from interrupt-handler to
2816 * preemptible context
2818 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2821 * - cond_resched() call
2822 * - explicit schedule() call
2823 * - return from syscall or exception to user-space
2824 * - return from interrupt-handler to user-space
2826 static void __sched __schedule(void)
2828 struct task_struct *prev, *next;
2829 unsigned long *switch_count;
2835 cpu = smp_processor_id();
2837 rcu_note_context_switch(cpu);
2840 schedule_debug(prev);
2842 if (sched_feat(HRTICK))
2845 raw_spin_lock_irq(&rq->lock);
2847 switch_count = &prev->nivcsw;
2848 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2849 if (unlikely(signal_pending_state(prev->state, prev))) {
2850 prev->state = TASK_RUNNING;
2852 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2856 * If a worker went to sleep, notify and ask workqueue
2857 * whether it wants to wake up a task to maintain
2860 if (prev->flags & PF_WQ_WORKER) {
2861 struct task_struct *to_wakeup;
2863 to_wakeup = wq_worker_sleeping(prev, cpu);
2865 try_to_wake_up_local(to_wakeup);
2868 switch_count = &prev->nvcsw;
2871 pre_schedule(rq, prev);
2873 if (unlikely(!rq->nr_running))
2874 idle_balance(cpu, rq);
2876 put_prev_task(rq, prev);
2877 next = pick_next_task(rq);
2878 clear_tsk_need_resched(prev);
2879 rq->skip_clock_update = 0;
2881 if (likely(prev != next)) {
2886 context_switch(rq, prev, next); /* unlocks the rq */
2888 * The context switch have flipped the stack from under us
2889 * and restored the local variables which were saved when
2890 * this task called schedule() in the past. prev == current
2891 * is still correct, but it can be moved to another cpu/rq.
2893 cpu = smp_processor_id();
2896 raw_spin_unlock_irq(&rq->lock);
2900 sched_preempt_enable_no_resched();
2905 static inline void sched_submit_work(struct task_struct *tsk)
2907 if (!tsk->state || tsk_is_pi_blocked(tsk))
2910 * If we are going to sleep and we have plugged IO queued,
2911 * make sure to submit it to avoid deadlocks.
2913 if (blk_needs_flush_plug(tsk))
2914 blk_schedule_flush_plug(tsk);
2917 asmlinkage void __sched schedule(void)
2919 struct task_struct *tsk = current;
2921 sched_submit_work(tsk);
2924 EXPORT_SYMBOL(schedule);
2926 #ifdef CONFIG_CONTEXT_TRACKING
2927 asmlinkage void __sched schedule_user(void)
2930 * If we come here after a random call to set_need_resched(),
2931 * or we have been woken up remotely but the IPI has not yet arrived,
2932 * we haven't yet exited the RCU idle mode. Do it here manually until
2933 * we find a better solution.
2942 * schedule_preempt_disabled - called with preemption disabled
2944 * Returns with preemption disabled. Note: preempt_count must be 1
2946 void __sched schedule_preempt_disabled(void)
2948 sched_preempt_enable_no_resched();
2953 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
2955 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
2957 if (lock->owner != owner)
2961 * Ensure we emit the owner->on_cpu, dereference _after_ checking
2962 * lock->owner still matches owner, if that fails, owner might
2963 * point to free()d memory, if it still matches, the rcu_read_lock()
2964 * ensures the memory stays valid.
2968 return owner->on_cpu;
2972 * Look out! "owner" is an entirely speculative pointer
2973 * access and not reliable.
2975 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
2977 if (!sched_feat(OWNER_SPIN))
2981 while (owner_running(lock, owner)) {
2985 arch_mutex_cpu_relax();
2990 * We break out the loop above on need_resched() and when the
2991 * owner changed, which is a sign for heavy contention. Return
2992 * success only when lock->owner is NULL.
2994 return lock->owner == NULL;
2998 #ifdef CONFIG_PREEMPT
3000 * this is the entry point to schedule() from in-kernel preemption
3001 * off of preempt_enable. Kernel preemptions off return from interrupt
3002 * occur there and call schedule directly.
3004 asmlinkage void __sched notrace preempt_schedule(void)
3006 struct thread_info *ti = current_thread_info();
3009 * If there is a non-zero preempt_count or interrupts are disabled,
3010 * we do not want to preempt the current task. Just return..
3012 if (likely(ti->preempt_count || irqs_disabled()))
3016 add_preempt_count_notrace(PREEMPT_ACTIVE);
3018 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3021 * Check again in case we missed a preemption opportunity
3022 * between schedule and now.
3025 } while (need_resched());
3027 EXPORT_SYMBOL(preempt_schedule);
3030 * this is the entry point to schedule() from kernel preemption
3031 * off of irq context.
3032 * Note, that this is called and return with irqs disabled. This will
3033 * protect us against recursive calling from irq.
3035 asmlinkage void __sched preempt_schedule_irq(void)
3037 struct thread_info *ti = current_thread_info();
3039 /* Catch callers which need to be fixed */
3040 BUG_ON(ti->preempt_count || !irqs_disabled());
3044 add_preempt_count(PREEMPT_ACTIVE);
3047 local_irq_disable();
3048 sub_preempt_count(PREEMPT_ACTIVE);
3051 * Check again in case we missed a preemption opportunity
3052 * between schedule and now.
3055 } while (need_resched());
3058 #endif /* CONFIG_PREEMPT */
3060 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3063 return try_to_wake_up(curr->private, mode, wake_flags);
3065 EXPORT_SYMBOL(default_wake_function);
3068 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3069 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3070 * number) then we wake all the non-exclusive tasks and one exclusive task.
3072 * There are circumstances in which we can try to wake a task which has already
3073 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3074 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3076 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3077 int nr_exclusive, int wake_flags, void *key)
3079 wait_queue_t *curr, *next;
3081 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3082 unsigned flags = curr->flags;
3084 if (curr->func(curr, mode, wake_flags, key) &&
3085 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3091 * __wake_up - wake up threads blocked on a waitqueue.
3093 * @mode: which threads
3094 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3095 * @key: is directly passed to the wakeup function
3097 * It may be assumed that this function implies a write memory barrier before
3098 * changing the task state if and only if any tasks are woken up.
3100 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3101 int nr_exclusive, void *key)
3103 unsigned long flags;
3105 spin_lock_irqsave(&q->lock, flags);
3106 __wake_up_common(q, mode, nr_exclusive, 0, key);
3107 spin_unlock_irqrestore(&q->lock, flags);
3109 EXPORT_SYMBOL(__wake_up);
3112 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3114 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3116 __wake_up_common(q, mode, nr, 0, NULL);
3118 EXPORT_SYMBOL_GPL(__wake_up_locked);
3120 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3122 __wake_up_common(q, mode, 1, 0, key);
3124 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3127 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3129 * @mode: which threads
3130 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3131 * @key: opaque value to be passed to wakeup targets
3133 * The sync wakeup differs that the waker knows that it will schedule
3134 * away soon, so while the target thread will be woken up, it will not
3135 * be migrated to another CPU - ie. the two threads are 'synchronized'
3136 * with each other. This can prevent needless bouncing between CPUs.
3138 * On UP it can prevent extra preemption.
3140 * It may be assumed that this function implies a write memory barrier before
3141 * changing the task state if and only if any tasks are woken up.
3143 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3144 int nr_exclusive, void *key)
3146 unsigned long flags;
3147 int wake_flags = WF_SYNC;
3152 if (unlikely(!nr_exclusive))
3155 spin_lock_irqsave(&q->lock, flags);
3156 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3157 spin_unlock_irqrestore(&q->lock, flags);
3159 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3162 * __wake_up_sync - see __wake_up_sync_key()
3164 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3166 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3168 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3171 * complete: - signals a single thread waiting on this completion
3172 * @x: holds the state of this particular completion
3174 * This will wake up a single thread waiting on this completion. Threads will be
3175 * awakened in the same order in which they were queued.
3177 * See also complete_all(), wait_for_completion() and related routines.
3179 * It may be assumed that this function implies a write memory barrier before
3180 * changing the task state if and only if any tasks are woken up.
3182 void complete(struct completion *x)
3184 unsigned long flags;
3186 spin_lock_irqsave(&x->wait.lock, flags);
3188 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3189 spin_unlock_irqrestore(&x->wait.lock, flags);
3191 EXPORT_SYMBOL(complete);
3194 * complete_all: - signals all threads waiting on this completion
3195 * @x: holds the state of this particular completion
3197 * This will wake up all threads waiting on this particular completion event.
3199 * It may be assumed that this function implies a write memory barrier before
3200 * changing the task state if and only if any tasks are woken up.
3202 void complete_all(struct completion *x)
3204 unsigned long flags;
3206 spin_lock_irqsave(&x->wait.lock, flags);
3207 x->done += UINT_MAX/2;
3208 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3209 spin_unlock_irqrestore(&x->wait.lock, flags);
3211 EXPORT_SYMBOL(complete_all);
3213 static inline long __sched
3214 do_wait_for_common(struct completion *x, long timeout, int state)
3217 DECLARE_WAITQUEUE(wait, current);
3219 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3221 if (signal_pending_state(state, current)) {
3222 timeout = -ERESTARTSYS;
3225 __set_current_state(state);
3226 spin_unlock_irq(&x->wait.lock);
3227 timeout = schedule_timeout(timeout);
3228 spin_lock_irq(&x->wait.lock);
3229 } while (!x->done && timeout);
3230 __remove_wait_queue(&x->wait, &wait);
3235 return timeout ?: 1;
3239 wait_for_common(struct completion *x, long timeout, int state)
3243 spin_lock_irq(&x->wait.lock);
3244 timeout = do_wait_for_common(x, timeout, state);
3245 spin_unlock_irq(&x->wait.lock);
3250 * wait_for_completion: - waits for completion of a task
3251 * @x: holds the state of this particular completion
3253 * This waits to be signaled for completion of a specific task. It is NOT
3254 * interruptible and there is no timeout.
3256 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3257 * and interrupt capability. Also see complete().
3259 void __sched wait_for_completion(struct completion *x)
3261 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3263 EXPORT_SYMBOL(wait_for_completion);
3266 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3267 * @x: holds the state of this particular completion
3268 * @timeout: timeout value in jiffies
3270 * This waits for either a completion of a specific task to be signaled or for a
3271 * specified timeout to expire. The timeout is in jiffies. It is not
3274 * The return value is 0 if timed out, and positive (at least 1, or number of
3275 * jiffies left till timeout) if completed.
3277 unsigned long __sched
3278 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3280 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3282 EXPORT_SYMBOL(wait_for_completion_timeout);
3285 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3286 * @x: holds the state of this particular completion
3288 * This waits for completion of a specific task to be signaled. It is
3291 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3293 int __sched wait_for_completion_interruptible(struct completion *x)
3295 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3296 if (t == -ERESTARTSYS)
3300 EXPORT_SYMBOL(wait_for_completion_interruptible);
3303 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3304 * @x: holds the state of this particular completion
3305 * @timeout: timeout value in jiffies
3307 * This waits for either a completion of a specific task to be signaled or for a
3308 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3310 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3311 * positive (at least 1, or number of jiffies left till timeout) if completed.
3314 wait_for_completion_interruptible_timeout(struct completion *x,
3315 unsigned long timeout)
3317 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3319 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3322 * wait_for_completion_killable: - waits for completion of a task (killable)
3323 * @x: holds the state of this particular completion
3325 * This waits to be signaled for completion of a specific task. It can be
3326 * interrupted by a kill signal.
3328 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3330 int __sched wait_for_completion_killable(struct completion *x)
3332 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3333 if (t == -ERESTARTSYS)
3337 EXPORT_SYMBOL(wait_for_completion_killable);
3340 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3341 * @x: holds the state of this particular completion
3342 * @timeout: timeout value in jiffies
3344 * This waits for either a completion of a specific task to be
3345 * signaled or for a specified timeout to expire. It can be
3346 * interrupted by a kill signal. The timeout is in jiffies.
3348 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3349 * positive (at least 1, or number of jiffies left till timeout) if completed.
3352 wait_for_completion_killable_timeout(struct completion *x,
3353 unsigned long timeout)
3355 return wait_for_common(x, timeout, TASK_KILLABLE);
3357 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3360 * try_wait_for_completion - try to decrement a completion without blocking
3361 * @x: completion structure
3363 * Returns: 0 if a decrement cannot be done without blocking
3364 * 1 if a decrement succeeded.
3366 * If a completion is being used as a counting completion,
3367 * attempt to decrement the counter without blocking. This
3368 * enables us to avoid waiting if the resource the completion
3369 * is protecting is not available.
3371 bool try_wait_for_completion(struct completion *x)
3373 unsigned long flags;
3376 spin_lock_irqsave(&x->wait.lock, flags);
3381 spin_unlock_irqrestore(&x->wait.lock, flags);
3384 EXPORT_SYMBOL(try_wait_for_completion);
3387 * completion_done - Test to see if a completion has any waiters
3388 * @x: completion structure
3390 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3391 * 1 if there are no waiters.
3394 bool completion_done(struct completion *x)
3396 unsigned long flags;
3399 spin_lock_irqsave(&x->wait.lock, flags);
3402 spin_unlock_irqrestore(&x->wait.lock, flags);
3405 EXPORT_SYMBOL(completion_done);
3408 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3410 unsigned long flags;
3413 init_waitqueue_entry(&wait, current);
3415 __set_current_state(state);
3417 spin_lock_irqsave(&q->lock, flags);
3418 __add_wait_queue(q, &wait);
3419 spin_unlock(&q->lock);
3420 timeout = schedule_timeout(timeout);
3421 spin_lock_irq(&q->lock);
3422 __remove_wait_queue(q, &wait);
3423 spin_unlock_irqrestore(&q->lock, flags);
3428 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3430 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3432 EXPORT_SYMBOL(interruptible_sleep_on);
3435 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3437 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3439 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3441 void __sched sleep_on(wait_queue_head_t *q)
3443 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3445 EXPORT_SYMBOL(sleep_on);
3447 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3449 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3451 EXPORT_SYMBOL(sleep_on_timeout);
3453 #ifdef CONFIG_RT_MUTEXES
3456 * rt_mutex_setprio - set the current priority of a task
3458 * @prio: prio value (kernel-internal form)
3460 * This function changes the 'effective' priority of a task. It does
3461 * not touch ->normal_prio like __setscheduler().
3463 * Used by the rt_mutex code to implement priority inheritance logic.
3465 void rt_mutex_setprio(struct task_struct *p, int prio)
3467 int oldprio, on_rq, running;
3469 const struct sched_class *prev_class;
3471 BUG_ON(prio < 0 || prio > MAX_PRIO);
3473 rq = __task_rq_lock(p);
3476 * Idle task boosting is a nono in general. There is one
3477 * exception, when PREEMPT_RT and NOHZ is active:
3479 * The idle task calls get_next_timer_interrupt() and holds
3480 * the timer wheel base->lock on the CPU and another CPU wants
3481 * to access the timer (probably to cancel it). We can safely
3482 * ignore the boosting request, as the idle CPU runs this code
3483 * with interrupts disabled and will complete the lock
3484 * protected section without being interrupted. So there is no
3485 * real need to boost.
3487 if (unlikely(p == rq->idle)) {
3488 WARN_ON(p != rq->curr);
3489 WARN_ON(p->pi_blocked_on);
3493 trace_sched_pi_setprio(p, prio);
3495 prev_class = p->sched_class;
3497 running = task_current(rq, p);
3499 dequeue_task(rq, p, 0);
3501 p->sched_class->put_prev_task(rq, p);
3504 p->sched_class = &rt_sched_class;
3506 p->sched_class = &fair_sched_class;
3511 p->sched_class->set_curr_task(rq);
3513 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3515 check_class_changed(rq, p, prev_class, oldprio);
3517 __task_rq_unlock(rq);
3520 void set_user_nice(struct task_struct *p, long nice)
3522 int old_prio, delta, on_rq;
3523 unsigned long flags;
3526 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3529 * We have to be careful, if called from sys_setpriority(),
3530 * the task might be in the middle of scheduling on another CPU.
3532 rq = task_rq_lock(p, &flags);
3534 * The RT priorities are set via sched_setscheduler(), but we still
3535 * allow the 'normal' nice value to be set - but as expected
3536 * it wont have any effect on scheduling until the task is
3537 * SCHED_FIFO/SCHED_RR:
3539 if (task_has_rt_policy(p)) {
3540 p->static_prio = NICE_TO_PRIO(nice);
3545 dequeue_task(rq, p, 0);
3547 p->static_prio = NICE_TO_PRIO(nice);
3550 p->prio = effective_prio(p);
3551 delta = p->prio - old_prio;
3554 enqueue_task(rq, p, 0);
3556 * If the task increased its priority or is running and
3557 * lowered its priority, then reschedule its CPU:
3559 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3560 resched_task(rq->curr);
3563 task_rq_unlock(rq, p, &flags);
3565 EXPORT_SYMBOL(set_user_nice);
3568 * can_nice - check if a task can reduce its nice value
3572 int can_nice(const struct task_struct *p, const int nice)
3574 /* convert nice value [19,-20] to rlimit style value [1,40] */
3575 int nice_rlim = 20 - nice;
3577 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3578 capable(CAP_SYS_NICE));
3581 #ifdef __ARCH_WANT_SYS_NICE
3584 * sys_nice - change the priority of the current process.
3585 * @increment: priority increment
3587 * sys_setpriority is a more generic, but much slower function that
3588 * does similar things.
3590 SYSCALL_DEFINE1(nice, int, increment)
3595 * Setpriority might change our priority at the same moment.
3596 * We don't have to worry. Conceptually one call occurs first
3597 * and we have a single winner.
3599 if (increment < -40)
3604 nice = TASK_NICE(current) + increment;
3610 if (increment < 0 && !can_nice(current, nice))
3613 retval = security_task_setnice(current, nice);
3617 set_user_nice(current, nice);
3624 * task_prio - return the priority value of a given task.
3625 * @p: the task in question.
3627 * This is the priority value as seen by users in /proc.
3628 * RT tasks are offset by -200. Normal tasks are centered
3629 * around 0, value goes from -16 to +15.
3631 int task_prio(const struct task_struct *p)
3633 return p->prio - MAX_RT_PRIO;
3637 * task_nice - return the nice value of a given task.
3638 * @p: the task in question.
3640 int task_nice(const struct task_struct *p)
3642 return TASK_NICE(p);
3644 EXPORT_SYMBOL(task_nice);
3647 * idle_cpu - is a given cpu idle currently?
3648 * @cpu: the processor in question.
3650 int idle_cpu(int cpu)
3652 struct rq *rq = cpu_rq(cpu);
3654 if (rq->curr != rq->idle)
3661 if (!llist_empty(&rq->wake_list))
3669 * idle_task - return the idle task for a given cpu.
3670 * @cpu: the processor in question.
3672 struct task_struct *idle_task(int cpu)
3674 return cpu_rq(cpu)->idle;
3678 * find_process_by_pid - find a process with a matching PID value.
3679 * @pid: the pid in question.
3681 static struct task_struct *find_process_by_pid(pid_t pid)
3683 return pid ? find_task_by_vpid(pid) : current;
3686 /* Actually do priority change: must hold rq lock. */
3688 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3691 p->rt_priority = prio;
3692 p->normal_prio = normal_prio(p);
3693 /* we are holding p->pi_lock already */
3694 p->prio = rt_mutex_getprio(p);
3695 if (rt_prio(p->prio))
3696 p->sched_class = &rt_sched_class;
3698 p->sched_class = &fair_sched_class;
3703 * check the target process has a UID that matches the current process's
3705 static bool check_same_owner(struct task_struct *p)
3707 const struct cred *cred = current_cred(), *pcred;
3711 pcred = __task_cred(p);
3712 match = (uid_eq(cred->euid, pcred->euid) ||
3713 uid_eq(cred->euid, pcred->uid));
3718 static int __sched_setscheduler(struct task_struct *p, int policy,
3719 const struct sched_param *param, bool user)
3721 int retval, oldprio, oldpolicy = -1, on_rq, running;
3722 unsigned long flags;
3723 const struct sched_class *prev_class;
3727 /* may grab non-irq protected spin_locks */
3728 BUG_ON(in_interrupt());
3730 /* double check policy once rq lock held */
3732 reset_on_fork = p->sched_reset_on_fork;
3733 policy = oldpolicy = p->policy;
3735 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3736 policy &= ~SCHED_RESET_ON_FORK;
3738 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3739 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3740 policy != SCHED_IDLE)
3745 * Valid priorities for SCHED_FIFO and SCHED_RR are
3746 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3747 * SCHED_BATCH and SCHED_IDLE is 0.
3749 if (param->sched_priority < 0 ||
3750 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3751 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3753 if (rt_policy(policy) != (param->sched_priority != 0))
3757 * Allow unprivileged RT tasks to decrease priority:
3759 if (user && !capable(CAP_SYS_NICE)) {
3760 if (rt_policy(policy)) {
3761 unsigned long rlim_rtprio =
3762 task_rlimit(p, RLIMIT_RTPRIO);
3764 /* can't set/change the rt policy */
3765 if (policy != p->policy && !rlim_rtprio)
3768 /* can't increase priority */
3769 if (param->sched_priority > p->rt_priority &&
3770 param->sched_priority > rlim_rtprio)
3775 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3776 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3778 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3779 if (!can_nice(p, TASK_NICE(p)))
3783 /* can't change other user's priorities */
3784 if (!check_same_owner(p))
3787 /* Normal users shall not reset the sched_reset_on_fork flag */
3788 if (p->sched_reset_on_fork && !reset_on_fork)
3793 retval = security_task_setscheduler(p);
3799 * make sure no PI-waiters arrive (or leave) while we are
3800 * changing the priority of the task:
3802 * To be able to change p->policy safely, the appropriate
3803 * runqueue lock must be held.
3805 rq = task_rq_lock(p, &flags);
3808 * Changing the policy of the stop threads its a very bad idea
3810 if (p == rq->stop) {
3811 task_rq_unlock(rq, p, &flags);
3816 * If not changing anything there's no need to proceed further:
3818 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3819 param->sched_priority == p->rt_priority))) {
3820 task_rq_unlock(rq, p, &flags);
3824 #ifdef CONFIG_RT_GROUP_SCHED
3827 * Do not allow realtime tasks into groups that have no runtime
3830 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3831 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3832 !task_group_is_autogroup(task_group(p))) {
3833 task_rq_unlock(rq, p, &flags);
3839 /* recheck policy now with rq lock held */
3840 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3841 policy = oldpolicy = -1;
3842 task_rq_unlock(rq, p, &flags);
3846 running = task_current(rq, p);
3848 dequeue_task(rq, p, 0);
3850 p->sched_class->put_prev_task(rq, p);
3852 p->sched_reset_on_fork = reset_on_fork;
3855 prev_class = p->sched_class;
3856 __setscheduler(rq, p, policy, param->sched_priority);
3859 p->sched_class->set_curr_task(rq);
3861 enqueue_task(rq, p, 0);
3863 check_class_changed(rq, p, prev_class, oldprio);
3864 task_rq_unlock(rq, p, &flags);
3866 rt_mutex_adjust_pi(p);
3872 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3873 * @p: the task in question.
3874 * @policy: new policy.
3875 * @param: structure containing the new RT priority.
3877 * NOTE that the task may be already dead.
3879 int sched_setscheduler(struct task_struct *p, int policy,
3880 const struct sched_param *param)
3882 return __sched_setscheduler(p, policy, param, true);
3884 EXPORT_SYMBOL_GPL(sched_setscheduler);
3887 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3888 * @p: the task in question.
3889 * @policy: new policy.
3890 * @param: structure containing the new RT priority.
3892 * Just like sched_setscheduler, only don't bother checking if the
3893 * current context has permission. For example, this is needed in
3894 * stop_machine(): we create temporary high priority worker threads,
3895 * but our caller might not have that capability.
3897 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3898 const struct sched_param *param)
3900 return __sched_setscheduler(p, policy, param, false);
3904 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3906 struct sched_param lparam;
3907 struct task_struct *p;
3910 if (!param || pid < 0)
3912 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3917 p = find_process_by_pid(pid);
3919 retval = sched_setscheduler(p, policy, &lparam);
3926 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3927 * @pid: the pid in question.
3928 * @policy: new policy.
3929 * @param: structure containing the new RT priority.
3931 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3932 struct sched_param __user *, param)
3934 /* negative values for policy are not valid */
3938 return do_sched_setscheduler(pid, policy, param);
3942 * sys_sched_setparam - set/change the RT priority of a thread
3943 * @pid: the pid in question.
3944 * @param: structure containing the new RT priority.
3946 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3948 return do_sched_setscheduler(pid, -1, param);
3952 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3953 * @pid: the pid in question.
3955 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3957 struct task_struct *p;
3965 p = find_process_by_pid(pid);
3967 retval = security_task_getscheduler(p);
3970 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3977 * sys_sched_getparam - get the RT priority of a thread
3978 * @pid: the pid in question.
3979 * @param: structure containing the RT priority.
3981 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3983 struct sched_param lp;
3984 struct task_struct *p;
3987 if (!param || pid < 0)
3991 p = find_process_by_pid(pid);
3996 retval = security_task_getscheduler(p);
4000 lp.sched_priority = p->rt_priority;
4004 * This one might sleep, we cannot do it with a spinlock held ...
4006 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4015 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4017 cpumask_var_t cpus_allowed, new_mask;
4018 struct task_struct *p;
4024 p = find_process_by_pid(pid);
4031 /* Prevent p going away */
4035 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4039 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4041 goto out_free_cpus_allowed;
4044 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE))
4047 retval = security_task_setscheduler(p);
4051 cpuset_cpus_allowed(p, cpus_allowed);
4052 cpumask_and(new_mask, in_mask, cpus_allowed);
4054 retval = set_cpus_allowed_ptr(p, new_mask);
4057 cpuset_cpus_allowed(p, cpus_allowed);
4058 if (!cpumask_subset(new_mask, cpus_allowed)) {
4060 * We must have raced with a concurrent cpuset
4061 * update. Just reset the cpus_allowed to the
4062 * cpuset's cpus_allowed
4064 cpumask_copy(new_mask, cpus_allowed);
4069 free_cpumask_var(new_mask);
4070 out_free_cpus_allowed:
4071 free_cpumask_var(cpus_allowed);
4078 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4079 struct cpumask *new_mask)
4081 if (len < cpumask_size())
4082 cpumask_clear(new_mask);
4083 else if (len > cpumask_size())
4084 len = cpumask_size();
4086 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4090 * sys_sched_setaffinity - set the cpu affinity of a process
4091 * @pid: pid of the process
4092 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4093 * @user_mask_ptr: user-space pointer to the new cpu mask
4095 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4096 unsigned long __user *, user_mask_ptr)
4098 cpumask_var_t new_mask;
4101 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4104 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4106 retval = sched_setaffinity(pid, new_mask);
4107 free_cpumask_var(new_mask);
4111 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4113 struct task_struct *p;
4114 unsigned long flags;
4121 p = find_process_by_pid(pid);
4125 retval = security_task_getscheduler(p);
4129 raw_spin_lock_irqsave(&p->pi_lock, flags);
4130 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4131 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4141 * sys_sched_getaffinity - get the cpu affinity of a process
4142 * @pid: pid of the process
4143 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4144 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4146 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4147 unsigned long __user *, user_mask_ptr)
4152 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4154 if (len & (sizeof(unsigned long)-1))
4157 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4160 ret = sched_getaffinity(pid, mask);
4162 size_t retlen = min_t(size_t, len, cpumask_size());
4164 if (copy_to_user(user_mask_ptr, mask, retlen))
4169 free_cpumask_var(mask);
4175 * sys_sched_yield - yield the current processor to other threads.
4177 * This function yields the current CPU to other tasks. If there are no
4178 * other threads running on this CPU then this function will return.
4180 SYSCALL_DEFINE0(sched_yield)
4182 struct rq *rq = this_rq_lock();
4184 schedstat_inc(rq, yld_count);
4185 current->sched_class->yield_task(rq);
4188 * Since we are going to call schedule() anyway, there's
4189 * no need to preempt or enable interrupts:
4191 __release(rq->lock);
4192 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4193 do_raw_spin_unlock(&rq->lock);
4194 sched_preempt_enable_no_resched();
4201 static inline int should_resched(void)
4203 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4206 static void __cond_resched(void)
4208 add_preempt_count(PREEMPT_ACTIVE);
4210 sub_preempt_count(PREEMPT_ACTIVE);
4213 int __sched _cond_resched(void)
4215 if (should_resched()) {
4221 EXPORT_SYMBOL(_cond_resched);
4224 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4225 * call schedule, and on return reacquire the lock.
4227 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4228 * operations here to prevent schedule() from being called twice (once via
4229 * spin_unlock(), once by hand).
4231 int __cond_resched_lock(spinlock_t *lock)
4233 int resched = should_resched();
4236 lockdep_assert_held(lock);
4238 if (spin_needbreak(lock) || resched) {
4249 EXPORT_SYMBOL(__cond_resched_lock);
4251 int __sched __cond_resched_softirq(void)
4253 BUG_ON(!in_softirq());
4255 if (should_resched()) {
4263 EXPORT_SYMBOL(__cond_resched_softirq);
4266 * yield - yield the current processor to other threads.
4268 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4270 * The scheduler is at all times free to pick the calling task as the most
4271 * eligible task to run, if removing the yield() call from your code breaks
4272 * it, its already broken.
4274 * Typical broken usage is:
4279 * where one assumes that yield() will let 'the other' process run that will
4280 * make event true. If the current task is a SCHED_FIFO task that will never
4281 * happen. Never use yield() as a progress guarantee!!
4283 * If you want to use yield() to wait for something, use wait_event().
4284 * If you want to use yield() to be 'nice' for others, use cond_resched().
4285 * If you still want to use yield(), do not!
4287 void __sched yield(void)
4289 set_current_state(TASK_RUNNING);
4292 EXPORT_SYMBOL(yield);
4295 * yield_to - yield the current processor to another thread in
4296 * your thread group, or accelerate that thread toward the
4297 * processor it's on.
4299 * @preempt: whether task preemption is allowed or not
4301 * It's the caller's job to ensure that the target task struct
4302 * can't go away on us before we can do any checks.
4304 * Returns true if we indeed boosted the target task.
4306 bool __sched yield_to(struct task_struct *p, bool preempt)
4308 struct task_struct *curr = current;
4309 struct rq *rq, *p_rq;
4310 unsigned long flags;
4313 local_irq_save(flags);
4318 double_rq_lock(rq, p_rq);
4319 while (task_rq(p) != p_rq) {
4320 double_rq_unlock(rq, p_rq);
4324 if (!curr->sched_class->yield_to_task)
4327 if (curr->sched_class != p->sched_class)
4330 if (task_running(p_rq, p) || p->state)
4333 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4335 schedstat_inc(rq, yld_count);
4337 * Make p's CPU reschedule; pick_next_entity takes care of
4340 if (preempt && rq != p_rq)
4341 resched_task(p_rq->curr);
4345 double_rq_unlock(rq, p_rq);
4346 local_irq_restore(flags);
4353 EXPORT_SYMBOL_GPL(yield_to);
4356 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4357 * that process accounting knows that this is a task in IO wait state.
4359 void __sched io_schedule(void)
4361 struct rq *rq = raw_rq();
4363 delayacct_blkio_start();
4364 atomic_inc(&rq->nr_iowait);
4365 blk_flush_plug(current);
4366 current->in_iowait = 1;
4368 current->in_iowait = 0;
4369 atomic_dec(&rq->nr_iowait);
4370 delayacct_blkio_end();
4372 EXPORT_SYMBOL(io_schedule);
4374 long __sched io_schedule_timeout(long timeout)
4376 struct rq *rq = raw_rq();
4379 delayacct_blkio_start();
4380 atomic_inc(&rq->nr_iowait);
4381 blk_flush_plug(current);
4382 current->in_iowait = 1;
4383 ret = schedule_timeout(timeout);
4384 current->in_iowait = 0;
4385 atomic_dec(&rq->nr_iowait);
4386 delayacct_blkio_end();
4391 * sys_sched_get_priority_max - return maximum RT priority.
4392 * @policy: scheduling class.
4394 * this syscall returns the maximum rt_priority that can be used
4395 * by a given scheduling class.
4397 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4404 ret = MAX_USER_RT_PRIO-1;
4416 * sys_sched_get_priority_min - return minimum RT priority.
4417 * @policy: scheduling class.
4419 * this syscall returns the minimum rt_priority that can be used
4420 * by a given scheduling class.
4422 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4440 * sys_sched_rr_get_interval - return the default timeslice of a process.
4441 * @pid: pid of the process.
4442 * @interval: userspace pointer to the timeslice value.
4444 * this syscall writes the default timeslice value of a given process
4445 * into the user-space timespec buffer. A value of '0' means infinity.
4447 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4448 struct timespec __user *, interval)
4450 struct task_struct *p;
4451 unsigned int time_slice;
4452 unsigned long flags;
4462 p = find_process_by_pid(pid);
4466 retval = security_task_getscheduler(p);
4470 rq = task_rq_lock(p, &flags);
4471 time_slice = p->sched_class->get_rr_interval(rq, p);
4472 task_rq_unlock(rq, p, &flags);
4475 jiffies_to_timespec(time_slice, &t);
4476 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4484 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4486 void sched_show_task(struct task_struct *p)
4488 unsigned long free = 0;
4492 state = p->state ? __ffs(p->state) + 1 : 0;
4493 printk(KERN_INFO "%-15.15s %c", p->comm,
4494 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4495 #if BITS_PER_LONG == 32
4496 if (state == TASK_RUNNING)
4497 printk(KERN_CONT " running ");
4499 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4501 if (state == TASK_RUNNING)
4502 printk(KERN_CONT " running task ");
4504 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4506 #ifdef CONFIG_DEBUG_STACK_USAGE
4507 free = stack_not_used(p);
4510 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4512 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4513 task_pid_nr(p), ppid,
4514 (unsigned long)task_thread_info(p)->flags);
4516 show_stack(p, NULL);
4519 void show_state_filter(unsigned long state_filter)
4521 struct task_struct *g, *p;
4523 #if BITS_PER_LONG == 32
4525 " task PC stack pid father\n");
4528 " task PC stack pid father\n");
4531 do_each_thread(g, p) {
4533 * reset the NMI-timeout, listing all files on a slow
4534 * console might take a lot of time:
4536 touch_nmi_watchdog();
4537 if (!state_filter || (p->state & state_filter))
4539 } while_each_thread(g, p);
4541 touch_all_softlockup_watchdogs();
4543 #ifdef CONFIG_SCHED_DEBUG
4544 sysrq_sched_debug_show();
4548 * Only show locks if all tasks are dumped:
4551 debug_show_all_locks();
4554 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4556 idle->sched_class = &idle_sched_class;
4560 * init_idle - set up an idle thread for a given CPU
4561 * @idle: task in question
4562 * @cpu: cpu the idle task belongs to
4564 * NOTE: this function does not set the idle thread's NEED_RESCHED
4565 * flag, to make booting more robust.
4567 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4569 struct rq *rq = cpu_rq(cpu);
4570 unsigned long flags;
4572 raw_spin_lock_irqsave(&rq->lock, flags);
4575 idle->state = TASK_RUNNING;
4576 idle->se.exec_start = sched_clock();
4578 do_set_cpus_allowed(idle, cpumask_of(cpu));
4580 * We're having a chicken and egg problem, even though we are
4581 * holding rq->lock, the cpu isn't yet set to this cpu so the
4582 * lockdep check in task_group() will fail.
4584 * Similar case to sched_fork(). / Alternatively we could
4585 * use task_rq_lock() here and obtain the other rq->lock.
4590 __set_task_cpu(idle, cpu);
4593 rq->curr = rq->idle = idle;
4594 #if defined(CONFIG_SMP)
4597 raw_spin_unlock_irqrestore(&rq->lock, flags);
4599 /* Set the preempt count _outside_ the spinlocks! */
4600 task_thread_info(idle)->preempt_count = 0;
4603 * The idle tasks have their own, simple scheduling class:
4605 idle->sched_class = &idle_sched_class;
4606 ftrace_graph_init_idle_task(idle, cpu);
4607 #if defined(CONFIG_SMP)
4608 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4613 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4615 if (p->sched_class && p->sched_class->set_cpus_allowed)
4616 p->sched_class->set_cpus_allowed(p, new_mask);
4618 cpumask_copy(&p->cpus_allowed, new_mask);
4619 p->nr_cpus_allowed = cpumask_weight(new_mask);
4623 * This is how migration works:
4625 * 1) we invoke migration_cpu_stop() on the target CPU using
4627 * 2) stopper starts to run (implicitly forcing the migrated thread
4629 * 3) it checks whether the migrated task is still in the wrong runqueue.
4630 * 4) if it's in the wrong runqueue then the migration thread removes
4631 * it and puts it into the right queue.
4632 * 5) stopper completes and stop_one_cpu() returns and the migration
4637 * Change a given task's CPU affinity. Migrate the thread to a
4638 * proper CPU and schedule it away if the CPU it's executing on
4639 * is removed from the allowed bitmask.
4641 * NOTE: the caller must have a valid reference to the task, the
4642 * task must not exit() & deallocate itself prematurely. The
4643 * call is not atomic; no spinlocks may be held.
4645 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4647 unsigned long flags;
4649 unsigned int dest_cpu;
4652 rq = task_rq_lock(p, &flags);
4654 if (cpumask_equal(&p->cpus_allowed, new_mask))
4657 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4662 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
4667 do_set_cpus_allowed(p, new_mask);
4669 /* Can the task run on the task's current CPU? If so, we're done */
4670 if (cpumask_test_cpu(task_cpu(p), new_mask))
4673 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4675 struct migration_arg arg = { p, dest_cpu };
4676 /* Need help from migration thread: drop lock and wait. */
4677 task_rq_unlock(rq, p, &flags);
4678 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4679 tlb_migrate_finish(p->mm);
4683 task_rq_unlock(rq, p, &flags);
4687 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4690 * Move (not current) task off this cpu, onto dest cpu. We're doing
4691 * this because either it can't run here any more (set_cpus_allowed()
4692 * away from this CPU, or CPU going down), or because we're
4693 * attempting to rebalance this task on exec (sched_exec).
4695 * So we race with normal scheduler movements, but that's OK, as long
4696 * as the task is no longer on this CPU.
4698 * Returns non-zero if task was successfully migrated.
4700 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4702 struct rq *rq_dest, *rq_src;
4705 if (unlikely(!cpu_active(dest_cpu)))
4708 rq_src = cpu_rq(src_cpu);
4709 rq_dest = cpu_rq(dest_cpu);
4711 raw_spin_lock(&p->pi_lock);
4712 double_rq_lock(rq_src, rq_dest);
4713 /* Already moved. */
4714 if (task_cpu(p) != src_cpu)
4716 /* Affinity changed (again). */
4717 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4721 * If we're not on a rq, the next wake-up will ensure we're
4725 dequeue_task(rq_src, p, 0);
4726 set_task_cpu(p, dest_cpu);
4727 enqueue_task(rq_dest, p, 0);
4728 check_preempt_curr(rq_dest, p, 0);
4733 double_rq_unlock(rq_src, rq_dest);
4734 raw_spin_unlock(&p->pi_lock);
4739 * migration_cpu_stop - this will be executed by a highprio stopper thread
4740 * and performs thread migration by bumping thread off CPU then
4741 * 'pushing' onto another runqueue.
4743 static int migration_cpu_stop(void *data)
4745 struct migration_arg *arg = data;
4748 * The original target cpu might have gone down and we might
4749 * be on another cpu but it doesn't matter.
4751 local_irq_disable();
4752 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4757 #ifdef CONFIG_HOTPLUG_CPU
4760 * Ensures that the idle task is using init_mm right before its cpu goes
4763 void idle_task_exit(void)
4765 struct mm_struct *mm = current->active_mm;
4767 BUG_ON(cpu_online(smp_processor_id()));
4770 switch_mm(mm, &init_mm, current);
4775 * Since this CPU is going 'away' for a while, fold any nr_active delta
4776 * we might have. Assumes we're called after migrate_tasks() so that the
4777 * nr_active count is stable.
4779 * Also see the comment "Global load-average calculations".
4781 static void calc_load_migrate(struct rq *rq)
4783 long delta = calc_load_fold_active(rq);
4785 atomic_long_add(delta, &calc_load_tasks);
4789 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4790 * try_to_wake_up()->select_task_rq().
4792 * Called with rq->lock held even though we'er in stop_machine() and
4793 * there's no concurrency possible, we hold the required locks anyway
4794 * because of lock validation efforts.
4796 static void migrate_tasks(unsigned int dead_cpu)
4798 struct rq *rq = cpu_rq(dead_cpu);
4799 struct task_struct *next, *stop = rq->stop;
4803 * Fudge the rq selection such that the below task selection loop
4804 * doesn't get stuck on the currently eligible stop task.
4806 * We're currently inside stop_machine() and the rq is either stuck
4807 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4808 * either way we should never end up calling schedule() until we're
4815 * There's this thread running, bail when that's the only
4818 if (rq->nr_running == 1)
4821 next = pick_next_task(rq);
4823 next->sched_class->put_prev_task(rq, next);
4825 /* Find suitable destination for @next, with force if needed. */
4826 dest_cpu = select_fallback_rq(dead_cpu, next);
4827 raw_spin_unlock(&rq->lock);
4829 __migrate_task(next, dead_cpu, dest_cpu);
4831 raw_spin_lock(&rq->lock);
4837 #endif /* CONFIG_HOTPLUG_CPU */
4839 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4841 static struct ctl_table sd_ctl_dir[] = {
4843 .procname = "sched_domain",
4849 static struct ctl_table sd_ctl_root[] = {
4851 .procname = "kernel",
4853 .child = sd_ctl_dir,
4858 static struct ctl_table *sd_alloc_ctl_entry(int n)
4860 struct ctl_table *entry =
4861 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4866 static void sd_free_ctl_entry(struct ctl_table **tablep)
4868 struct ctl_table *entry;
4871 * In the intermediate directories, both the child directory and
4872 * procname are dynamically allocated and could fail but the mode
4873 * will always be set. In the lowest directory the names are
4874 * static strings and all have proc handlers.
4876 for (entry = *tablep; entry->mode; entry++) {
4878 sd_free_ctl_entry(&entry->child);
4879 if (entry->proc_handler == NULL)
4880 kfree(entry->procname);
4887 static int min_load_idx = 0;
4888 static int max_load_idx = CPU_LOAD_IDX_MAX;
4891 set_table_entry(struct ctl_table *entry,
4892 const char *procname, void *data, int maxlen,
4893 umode_t mode, proc_handler *proc_handler,
4896 entry->procname = procname;
4898 entry->maxlen = maxlen;
4900 entry->proc_handler = proc_handler;
4903 entry->extra1 = &min_load_idx;
4904 entry->extra2 = &max_load_idx;
4908 static struct ctl_table *
4909 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4911 struct ctl_table *table = sd_alloc_ctl_entry(13);
4916 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4917 sizeof(long), 0644, proc_doulongvec_minmax, false);
4918 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4919 sizeof(long), 0644, proc_doulongvec_minmax, false);
4920 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4921 sizeof(int), 0644, proc_dointvec_minmax, true);
4922 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4923 sizeof(int), 0644, proc_dointvec_minmax, true);
4924 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4925 sizeof(int), 0644, proc_dointvec_minmax, true);
4926 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4927 sizeof(int), 0644, proc_dointvec_minmax, true);
4928 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4929 sizeof(int), 0644, proc_dointvec_minmax, true);
4930 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4931 sizeof(int), 0644, proc_dointvec_minmax, false);
4932 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4933 sizeof(int), 0644, proc_dointvec_minmax, false);
4934 set_table_entry(&table[9], "cache_nice_tries",
4935 &sd->cache_nice_tries,
4936 sizeof(int), 0644, proc_dointvec_minmax, false);
4937 set_table_entry(&table[10], "flags", &sd->flags,
4938 sizeof(int), 0644, proc_dointvec_minmax, false);
4939 set_table_entry(&table[11], "name", sd->name,
4940 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4941 /* &table[12] is terminator */
4946 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4948 struct ctl_table *entry, *table;
4949 struct sched_domain *sd;
4950 int domain_num = 0, i;
4953 for_each_domain(cpu, sd)
4955 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4960 for_each_domain(cpu, sd) {
4961 snprintf(buf, 32, "domain%d", i);
4962 entry->procname = kstrdup(buf, GFP_KERNEL);
4964 entry->child = sd_alloc_ctl_domain_table(sd);
4971 static struct ctl_table_header *sd_sysctl_header;
4972 static void register_sched_domain_sysctl(void)
4974 int i, cpu_num = num_possible_cpus();
4975 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4978 WARN_ON(sd_ctl_dir[0].child);
4979 sd_ctl_dir[0].child = entry;
4984 for_each_possible_cpu(i) {
4985 snprintf(buf, 32, "cpu%d", i);
4986 entry->procname = kstrdup(buf, GFP_KERNEL);
4988 entry->child = sd_alloc_ctl_cpu_table(i);
4992 WARN_ON(sd_sysctl_header);
4993 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4996 /* may be called multiple times per register */
4997 static void unregister_sched_domain_sysctl(void)
4999 if (sd_sysctl_header)
5000 unregister_sysctl_table(sd_sysctl_header);
5001 sd_sysctl_header = NULL;
5002 if (sd_ctl_dir[0].child)
5003 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5006 static void register_sched_domain_sysctl(void)
5009 static void unregister_sched_domain_sysctl(void)
5014 static void set_rq_online(struct rq *rq)
5017 const struct sched_class *class;
5019 cpumask_set_cpu(rq->cpu, rq->rd->online);
5022 for_each_class(class) {
5023 if (class->rq_online)
5024 class->rq_online(rq);
5029 static void set_rq_offline(struct rq *rq)
5032 const struct sched_class *class;
5034 for_each_class(class) {
5035 if (class->rq_offline)
5036 class->rq_offline(rq);
5039 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5045 * migration_call - callback that gets triggered when a CPU is added.
5046 * Here we can start up the necessary migration thread for the new CPU.
5048 static int __cpuinit
5049 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5051 int cpu = (long)hcpu;
5052 unsigned long flags;
5053 struct rq *rq = cpu_rq(cpu);
5055 switch (action & ~CPU_TASKS_FROZEN) {
5057 case CPU_UP_PREPARE:
5058 rq->calc_load_update = calc_load_update;
5062 /* Update our root-domain */
5063 raw_spin_lock_irqsave(&rq->lock, flags);
5065 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5069 raw_spin_unlock_irqrestore(&rq->lock, flags);
5072 #ifdef CONFIG_HOTPLUG_CPU
5074 sched_ttwu_pending();
5075 /* Update our root-domain */
5076 raw_spin_lock_irqsave(&rq->lock, flags);
5078 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5082 BUG_ON(rq->nr_running != 1); /* the migration thread */
5083 raw_spin_unlock_irqrestore(&rq->lock, flags);
5087 calc_load_migrate(rq);
5092 update_max_interval();
5098 * Register at high priority so that task migration (migrate_all_tasks)
5099 * happens before everything else. This has to be lower priority than
5100 * the notifier in the perf_event subsystem, though.
5102 static struct notifier_block __cpuinitdata migration_notifier = {
5103 .notifier_call = migration_call,
5104 .priority = CPU_PRI_MIGRATION,
5107 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5108 unsigned long action, void *hcpu)
5110 switch (action & ~CPU_TASKS_FROZEN) {
5112 case CPU_DOWN_FAILED:
5113 set_cpu_active((long)hcpu, true);
5120 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5121 unsigned long action, void *hcpu)
5123 switch (action & ~CPU_TASKS_FROZEN) {
5124 case CPU_DOWN_PREPARE:
5125 set_cpu_active((long)hcpu, false);
5132 static int __init migration_init(void)
5134 void *cpu = (void *)(long)smp_processor_id();
5137 /* Initialize migration for the boot CPU */
5138 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5139 BUG_ON(err == NOTIFY_BAD);
5140 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5141 register_cpu_notifier(&migration_notifier);
5143 /* Register cpu active notifiers */
5144 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5145 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5149 early_initcall(migration_init);
5154 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5156 #ifdef CONFIG_SCHED_DEBUG
5158 static __read_mostly int sched_debug_enabled;
5160 static int __init sched_debug_setup(char *str)
5162 sched_debug_enabled = 1;
5166 early_param("sched_debug", sched_debug_setup);
5168 static inline bool sched_debug(void)
5170 return sched_debug_enabled;
5173 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5174 struct cpumask *groupmask)
5176 struct sched_group *group = sd->groups;
5179 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5180 cpumask_clear(groupmask);
5182 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5184 if (!(sd->flags & SD_LOAD_BALANCE)) {
5185 printk("does not load-balance\n");
5187 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5192 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5194 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5195 printk(KERN_ERR "ERROR: domain->span does not contain "
5198 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5199 printk(KERN_ERR "ERROR: domain->groups does not contain"
5203 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5207 printk(KERN_ERR "ERROR: group is NULL\n");
5212 * Even though we initialize ->power to something semi-sane,
5213 * we leave power_orig unset. This allows us to detect if
5214 * domain iteration is still funny without causing /0 traps.
5216 if (!group->sgp->power_orig) {
5217 printk(KERN_CONT "\n");
5218 printk(KERN_ERR "ERROR: domain->cpu_power not "
5223 if (!cpumask_weight(sched_group_cpus(group))) {
5224 printk(KERN_CONT "\n");
5225 printk(KERN_ERR "ERROR: empty group\n");
5229 if (!(sd->flags & SD_OVERLAP) &&
5230 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5231 printk(KERN_CONT "\n");
5232 printk(KERN_ERR "ERROR: repeated CPUs\n");
5236 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5238 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5240 printk(KERN_CONT " %s", str);
5241 if (group->sgp->power != SCHED_POWER_SCALE) {
5242 printk(KERN_CONT " (cpu_power = %d)",
5246 group = group->next;
5247 } while (group != sd->groups);
5248 printk(KERN_CONT "\n");
5250 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5251 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5254 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5255 printk(KERN_ERR "ERROR: parent span is not a superset "
5256 "of domain->span\n");
5260 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5264 if (!sched_debug_enabled)
5268 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5272 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5275 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5283 #else /* !CONFIG_SCHED_DEBUG */
5284 # define sched_domain_debug(sd, cpu) do { } while (0)
5285 static inline bool sched_debug(void)
5289 #endif /* CONFIG_SCHED_DEBUG */
5291 static int sd_degenerate(struct sched_domain *sd)
5293 if (cpumask_weight(sched_domain_span(sd)) == 1)
5296 /* Following flags need at least 2 groups */
5297 if (sd->flags & (SD_LOAD_BALANCE |
5298 SD_BALANCE_NEWIDLE |
5302 SD_SHARE_PKG_RESOURCES)) {
5303 if (sd->groups != sd->groups->next)
5307 /* Following flags don't use groups */
5308 if (sd->flags & (SD_WAKE_AFFINE))
5315 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5317 unsigned long cflags = sd->flags, pflags = parent->flags;
5319 if (sd_degenerate(parent))
5322 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5325 /* Flags needing groups don't count if only 1 group in parent */
5326 if (parent->groups == parent->groups->next) {
5327 pflags &= ~(SD_LOAD_BALANCE |
5328 SD_BALANCE_NEWIDLE |
5332 SD_SHARE_PKG_RESOURCES);
5333 if (nr_node_ids == 1)
5334 pflags &= ~SD_SERIALIZE;
5336 if (~cflags & pflags)
5342 static void free_rootdomain(struct rcu_head *rcu)
5344 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5346 cpupri_cleanup(&rd->cpupri);
5347 free_cpumask_var(rd->rto_mask);
5348 free_cpumask_var(rd->online);
5349 free_cpumask_var(rd->span);
5353 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5355 struct root_domain *old_rd = NULL;
5356 unsigned long flags;
5358 raw_spin_lock_irqsave(&rq->lock, flags);
5363 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5366 cpumask_clear_cpu(rq->cpu, old_rd->span);
5369 * If we dont want to free the old_rt yet then
5370 * set old_rd to NULL to skip the freeing later
5373 if (!atomic_dec_and_test(&old_rd->refcount))
5377 atomic_inc(&rd->refcount);
5380 cpumask_set_cpu(rq->cpu, rd->span);
5381 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5384 raw_spin_unlock_irqrestore(&rq->lock, flags);
5387 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5390 static int init_rootdomain(struct root_domain *rd)
5392 memset(rd, 0, sizeof(*rd));
5394 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5396 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5398 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5401 if (cpupri_init(&rd->cpupri) != 0)
5406 free_cpumask_var(rd->rto_mask);
5408 free_cpumask_var(rd->online);
5410 free_cpumask_var(rd->span);
5416 * By default the system creates a single root-domain with all cpus as
5417 * members (mimicking the global state we have today).
5419 struct root_domain def_root_domain;
5421 static void init_defrootdomain(void)
5423 init_rootdomain(&def_root_domain);
5425 atomic_set(&def_root_domain.refcount, 1);
5428 static struct root_domain *alloc_rootdomain(void)
5430 struct root_domain *rd;
5432 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5436 if (init_rootdomain(rd) != 0) {
5444 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5446 struct sched_group *tmp, *first;
5455 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5460 } while (sg != first);
5463 static void free_sched_domain(struct rcu_head *rcu)
5465 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5468 * If its an overlapping domain it has private groups, iterate and
5471 if (sd->flags & SD_OVERLAP) {
5472 free_sched_groups(sd->groups, 1);
5473 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5474 kfree(sd->groups->sgp);
5480 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5482 call_rcu(&sd->rcu, free_sched_domain);
5485 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5487 for (; sd; sd = sd->parent)
5488 destroy_sched_domain(sd, cpu);
5492 * Keep a special pointer to the highest sched_domain that has
5493 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5494 * allows us to avoid some pointer chasing select_idle_sibling().
5496 * Also keep a unique ID per domain (we use the first cpu number in
5497 * the cpumask of the domain), this allows us to quickly tell if
5498 * two cpus are in the same cache domain, see cpus_share_cache().
5500 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5501 DEFINE_PER_CPU(int, sd_llc_id);
5503 static void update_top_cache_domain(int cpu)
5505 struct sched_domain *sd;
5508 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5510 id = cpumask_first(sched_domain_span(sd));
5512 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5513 per_cpu(sd_llc_id, cpu) = id;
5517 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5518 * hold the hotplug lock.
5521 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5523 struct rq *rq = cpu_rq(cpu);
5524 struct sched_domain *tmp;
5526 /* Remove the sched domains which do not contribute to scheduling. */
5527 for (tmp = sd; tmp; ) {
5528 struct sched_domain *parent = tmp->parent;
5532 if (sd_parent_degenerate(tmp, parent)) {
5533 tmp->parent = parent->parent;
5535 parent->parent->child = tmp;
5536 destroy_sched_domain(parent, cpu);
5541 if (sd && sd_degenerate(sd)) {
5544 destroy_sched_domain(tmp, cpu);
5549 sched_domain_debug(sd, cpu);
5551 rq_attach_root(rq, rd);
5553 rcu_assign_pointer(rq->sd, sd);
5554 destroy_sched_domains(tmp, cpu);
5556 update_top_cache_domain(cpu);
5559 /* cpus with isolated domains */
5560 static cpumask_var_t cpu_isolated_map;
5562 /* Setup the mask of cpus configured for isolated domains */
5563 static int __init isolated_cpu_setup(char *str)
5565 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5566 cpulist_parse(str, cpu_isolated_map);
5570 __setup("isolcpus=", isolated_cpu_setup);
5572 static const struct cpumask *cpu_cpu_mask(int cpu)
5574 return cpumask_of_node(cpu_to_node(cpu));
5578 struct sched_domain **__percpu sd;
5579 struct sched_group **__percpu sg;
5580 struct sched_group_power **__percpu sgp;
5584 struct sched_domain ** __percpu sd;
5585 struct root_domain *rd;
5595 struct sched_domain_topology_level;
5597 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5598 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5600 #define SDTL_OVERLAP 0x01
5602 struct sched_domain_topology_level {
5603 sched_domain_init_f init;
5604 sched_domain_mask_f mask;
5607 struct sd_data data;
5611 * Build an iteration mask that can exclude certain CPUs from the upwards
5614 * Asymmetric node setups can result in situations where the domain tree is of
5615 * unequal depth, make sure to skip domains that already cover the entire
5618 * In that case build_sched_domains() will have terminated the iteration early
5619 * and our sibling sd spans will be empty. Domains should always include the
5620 * cpu they're built on, so check that.
5623 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5625 const struct cpumask *span = sched_domain_span(sd);
5626 struct sd_data *sdd = sd->private;
5627 struct sched_domain *sibling;
5630 for_each_cpu(i, span) {
5631 sibling = *per_cpu_ptr(sdd->sd, i);
5632 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5635 cpumask_set_cpu(i, sched_group_mask(sg));
5640 * Return the canonical balance cpu for this group, this is the first cpu
5641 * of this group that's also in the iteration mask.
5643 int group_balance_cpu(struct sched_group *sg)
5645 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5649 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5651 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5652 const struct cpumask *span = sched_domain_span(sd);
5653 struct cpumask *covered = sched_domains_tmpmask;
5654 struct sd_data *sdd = sd->private;
5655 struct sched_domain *child;
5658 cpumask_clear(covered);
5660 for_each_cpu(i, span) {
5661 struct cpumask *sg_span;
5663 if (cpumask_test_cpu(i, covered))
5666 child = *per_cpu_ptr(sdd->sd, i);
5668 /* See the comment near build_group_mask(). */
5669 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5672 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5673 GFP_KERNEL, cpu_to_node(cpu));
5678 sg_span = sched_group_cpus(sg);
5680 child = child->child;
5681 cpumask_copy(sg_span, sched_domain_span(child));
5683 cpumask_set_cpu(i, sg_span);
5685 cpumask_or(covered, covered, sg_span);
5687 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5688 if (atomic_inc_return(&sg->sgp->ref) == 1)
5689 build_group_mask(sd, sg);
5692 * Initialize sgp->power such that even if we mess up the
5693 * domains and no possible iteration will get us here, we won't
5696 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5699 * Make sure the first group of this domain contains the
5700 * canonical balance cpu. Otherwise the sched_domain iteration
5701 * breaks. See update_sg_lb_stats().
5703 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5704 group_balance_cpu(sg) == cpu)
5714 sd->groups = groups;
5719 free_sched_groups(first, 0);
5724 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5726 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5727 struct sched_domain *child = sd->child;
5730 cpu = cpumask_first(sched_domain_span(child));
5733 *sg = *per_cpu_ptr(sdd->sg, cpu);
5734 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5735 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5742 * build_sched_groups will build a circular linked list of the groups
5743 * covered by the given span, and will set each group's ->cpumask correctly,
5744 * and ->cpu_power to 0.
5746 * Assumes the sched_domain tree is fully constructed
5749 build_sched_groups(struct sched_domain *sd, int cpu)
5751 struct sched_group *first = NULL, *last = NULL;
5752 struct sd_data *sdd = sd->private;
5753 const struct cpumask *span = sched_domain_span(sd);
5754 struct cpumask *covered;
5757 get_group(cpu, sdd, &sd->groups);
5758 atomic_inc(&sd->groups->ref);
5760 if (cpu != cpumask_first(sched_domain_span(sd)))
5763 lockdep_assert_held(&sched_domains_mutex);
5764 covered = sched_domains_tmpmask;
5766 cpumask_clear(covered);
5768 for_each_cpu(i, span) {
5769 struct sched_group *sg;
5770 int group = get_group(i, sdd, &sg);
5773 if (cpumask_test_cpu(i, covered))
5776 cpumask_clear(sched_group_cpus(sg));
5778 cpumask_setall(sched_group_mask(sg));
5780 for_each_cpu(j, span) {
5781 if (get_group(j, sdd, NULL) != group)
5784 cpumask_set_cpu(j, covered);
5785 cpumask_set_cpu(j, sched_group_cpus(sg));
5800 * Initialize sched groups cpu_power.
5802 * cpu_power indicates the capacity of sched group, which is used while
5803 * distributing the load between different sched groups in a sched domain.
5804 * Typically cpu_power for all the groups in a sched domain will be same unless
5805 * there are asymmetries in the topology. If there are asymmetries, group
5806 * having more cpu_power will pickup more load compared to the group having
5809 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5811 struct sched_group *sg = sd->groups;
5813 WARN_ON(!sd || !sg);
5816 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5818 } while (sg != sd->groups);
5820 if (cpu != group_balance_cpu(sg))
5823 update_group_power(sd, cpu);
5824 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5827 int __weak arch_sd_sibling_asym_packing(void)
5829 return 0*SD_ASYM_PACKING;
5833 * Initializers for schedule domains
5834 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5837 #ifdef CONFIG_SCHED_DEBUG
5838 # define SD_INIT_NAME(sd, type) sd->name = #type
5840 # define SD_INIT_NAME(sd, type) do { } while (0)
5843 #define SD_INIT_FUNC(type) \
5844 static noinline struct sched_domain * \
5845 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5847 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5848 *sd = SD_##type##_INIT; \
5849 SD_INIT_NAME(sd, type); \
5850 sd->private = &tl->data; \
5855 #ifdef CONFIG_SCHED_SMT
5856 SD_INIT_FUNC(SIBLING)
5858 #ifdef CONFIG_SCHED_MC
5861 #ifdef CONFIG_SCHED_BOOK
5865 static int default_relax_domain_level = -1;
5866 int sched_domain_level_max;
5868 static int __init setup_relax_domain_level(char *str)
5870 if (kstrtoint(str, 0, &default_relax_domain_level))
5871 pr_warn("Unable to set relax_domain_level\n");
5875 __setup("relax_domain_level=", setup_relax_domain_level);
5877 static void set_domain_attribute(struct sched_domain *sd,
5878 struct sched_domain_attr *attr)
5882 if (!attr || attr->relax_domain_level < 0) {
5883 if (default_relax_domain_level < 0)
5886 request = default_relax_domain_level;
5888 request = attr->relax_domain_level;
5889 if (request < sd->level) {
5890 /* turn off idle balance on this domain */
5891 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5893 /* turn on idle balance on this domain */
5894 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5898 static void __sdt_free(const struct cpumask *cpu_map);
5899 static int __sdt_alloc(const struct cpumask *cpu_map);
5901 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5902 const struct cpumask *cpu_map)
5906 if (!atomic_read(&d->rd->refcount))
5907 free_rootdomain(&d->rd->rcu); /* fall through */
5909 free_percpu(d->sd); /* fall through */
5911 __sdt_free(cpu_map); /* fall through */
5917 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5918 const struct cpumask *cpu_map)
5920 memset(d, 0, sizeof(*d));
5922 if (__sdt_alloc(cpu_map))
5923 return sa_sd_storage;
5924 d->sd = alloc_percpu(struct sched_domain *);
5926 return sa_sd_storage;
5927 d->rd = alloc_rootdomain();
5930 return sa_rootdomain;
5934 * NULL the sd_data elements we've used to build the sched_domain and
5935 * sched_group structure so that the subsequent __free_domain_allocs()
5936 * will not free the data we're using.
5938 static void claim_allocations(int cpu, struct sched_domain *sd)
5940 struct sd_data *sdd = sd->private;
5942 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5943 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5945 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5946 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5948 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5949 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5952 #ifdef CONFIG_SCHED_SMT
5953 static const struct cpumask *cpu_smt_mask(int cpu)
5955 return topology_thread_cpumask(cpu);
5960 * Topology list, bottom-up.
5962 static struct sched_domain_topology_level default_topology[] = {
5963 #ifdef CONFIG_SCHED_SMT
5964 { sd_init_SIBLING, cpu_smt_mask, },
5966 #ifdef CONFIG_SCHED_MC
5967 { sd_init_MC, cpu_coregroup_mask, },
5969 #ifdef CONFIG_SCHED_BOOK
5970 { sd_init_BOOK, cpu_book_mask, },
5972 { sd_init_CPU, cpu_cpu_mask, },
5976 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5980 static int sched_domains_numa_levels;
5981 static int *sched_domains_numa_distance;
5982 static struct cpumask ***sched_domains_numa_masks;
5983 static int sched_domains_curr_level;
5985 static inline int sd_local_flags(int level)
5987 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5990 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5993 static struct sched_domain *
5994 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5996 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5997 int level = tl->numa_level;
5998 int sd_weight = cpumask_weight(
5999 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6001 *sd = (struct sched_domain){
6002 .min_interval = sd_weight,
6003 .max_interval = 2*sd_weight,
6005 .imbalance_pct = 125,
6006 .cache_nice_tries = 2,
6013 .flags = 1*SD_LOAD_BALANCE
6014 | 1*SD_BALANCE_NEWIDLE
6019 | 0*SD_SHARE_CPUPOWER
6020 | 0*SD_SHARE_PKG_RESOURCES
6022 | 0*SD_PREFER_SIBLING
6023 | sd_local_flags(level)
6025 .last_balance = jiffies,
6026 .balance_interval = sd_weight,
6028 SD_INIT_NAME(sd, NUMA);
6029 sd->private = &tl->data;
6032 * Ugly hack to pass state to sd_numa_mask()...
6034 sched_domains_curr_level = tl->numa_level;
6039 static const struct cpumask *sd_numa_mask(int cpu)
6041 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6044 static void sched_numa_warn(const char *str)
6046 static int done = false;
6054 printk(KERN_WARNING "ERROR: %s\n\n", str);
6056 for (i = 0; i < nr_node_ids; i++) {
6057 printk(KERN_WARNING " ");
6058 for (j = 0; j < nr_node_ids; j++)
6059 printk(KERN_CONT "%02d ", node_distance(i,j));
6060 printk(KERN_CONT "\n");
6062 printk(KERN_WARNING "\n");
6065 static bool find_numa_distance(int distance)
6069 if (distance == node_distance(0, 0))
6072 for (i = 0; i < sched_domains_numa_levels; i++) {
6073 if (sched_domains_numa_distance[i] == distance)
6080 static void sched_init_numa(void)
6082 int next_distance, curr_distance = node_distance(0, 0);
6083 struct sched_domain_topology_level *tl;
6087 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6088 if (!sched_domains_numa_distance)
6092 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6093 * unique distances in the node_distance() table.
6095 * Assumes node_distance(0,j) includes all distances in
6096 * node_distance(i,j) in order to avoid cubic time.
6098 next_distance = curr_distance;
6099 for (i = 0; i < nr_node_ids; i++) {
6100 for (j = 0; j < nr_node_ids; j++) {
6101 for (k = 0; k < nr_node_ids; k++) {
6102 int distance = node_distance(i, k);
6104 if (distance > curr_distance &&
6105 (distance < next_distance ||
6106 next_distance == curr_distance))
6107 next_distance = distance;
6110 * While not a strong assumption it would be nice to know
6111 * about cases where if node A is connected to B, B is not
6112 * equally connected to A.
6114 if (sched_debug() && node_distance(k, i) != distance)
6115 sched_numa_warn("Node-distance not symmetric");
6117 if (sched_debug() && i && !find_numa_distance(distance))
6118 sched_numa_warn("Node-0 not representative");
6120 if (next_distance != curr_distance) {
6121 sched_domains_numa_distance[level++] = next_distance;
6122 sched_domains_numa_levels = level;
6123 curr_distance = next_distance;
6128 * In case of sched_debug() we verify the above assumption.
6134 * 'level' contains the number of unique distances, excluding the
6135 * identity distance node_distance(i,i).
6137 * The sched_domains_nume_distance[] array includes the actual distance
6142 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6143 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6144 * the array will contain less then 'level' members. This could be
6145 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6146 * in other functions.
6148 * We reset it to 'level' at the end of this function.
6150 sched_domains_numa_levels = 0;
6152 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6153 if (!sched_domains_numa_masks)
6157 * Now for each level, construct a mask per node which contains all
6158 * cpus of nodes that are that many hops away from us.
6160 for (i = 0; i < level; i++) {
6161 sched_domains_numa_masks[i] =
6162 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6163 if (!sched_domains_numa_masks[i])
6166 for (j = 0; j < nr_node_ids; j++) {
6167 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6171 sched_domains_numa_masks[i][j] = mask;
6173 for (k = 0; k < nr_node_ids; k++) {
6174 if (node_distance(j, k) > sched_domains_numa_distance[i])
6177 cpumask_or(mask, mask, cpumask_of_node(k));
6182 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6183 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6188 * Copy the default topology bits..
6190 for (i = 0; default_topology[i].init; i++)
6191 tl[i] = default_topology[i];
6194 * .. and append 'j' levels of NUMA goodness.
6196 for (j = 0; j < level; i++, j++) {
6197 tl[i] = (struct sched_domain_topology_level){
6198 .init = sd_numa_init,
6199 .mask = sd_numa_mask,
6200 .flags = SDTL_OVERLAP,
6205 sched_domain_topology = tl;
6207 sched_domains_numa_levels = level;
6210 static void sched_domains_numa_masks_set(int cpu)
6213 int node = cpu_to_node(cpu);
6215 for (i = 0; i < sched_domains_numa_levels; i++) {
6216 for (j = 0; j < nr_node_ids; j++) {
6217 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6218 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6223 static void sched_domains_numa_masks_clear(int cpu)
6226 for (i = 0; i < sched_domains_numa_levels; i++) {
6227 for (j = 0; j < nr_node_ids; j++)
6228 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6233 * Update sched_domains_numa_masks[level][node] array when new cpus
6236 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6237 unsigned long action,
6240 int cpu = (long)hcpu;
6242 switch (action & ~CPU_TASKS_FROZEN) {
6244 sched_domains_numa_masks_set(cpu);
6248 sched_domains_numa_masks_clear(cpu);
6258 static inline void sched_init_numa(void)
6262 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6263 unsigned long action,
6268 #endif /* CONFIG_NUMA */
6270 static int __sdt_alloc(const struct cpumask *cpu_map)
6272 struct sched_domain_topology_level *tl;
6275 for (tl = sched_domain_topology; tl->init; tl++) {
6276 struct sd_data *sdd = &tl->data;
6278 sdd->sd = alloc_percpu(struct sched_domain *);
6282 sdd->sg = alloc_percpu(struct sched_group *);
6286 sdd->sgp = alloc_percpu(struct sched_group_power *);
6290 for_each_cpu(j, cpu_map) {
6291 struct sched_domain *sd;
6292 struct sched_group *sg;
6293 struct sched_group_power *sgp;
6295 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6296 GFP_KERNEL, cpu_to_node(j));
6300 *per_cpu_ptr(sdd->sd, j) = sd;
6302 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6303 GFP_KERNEL, cpu_to_node(j));
6309 *per_cpu_ptr(sdd->sg, j) = sg;
6311 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6312 GFP_KERNEL, cpu_to_node(j));
6316 *per_cpu_ptr(sdd->sgp, j) = sgp;
6323 static void __sdt_free(const struct cpumask *cpu_map)
6325 struct sched_domain_topology_level *tl;
6328 for (tl = sched_domain_topology; tl->init; tl++) {
6329 struct sd_data *sdd = &tl->data;
6331 for_each_cpu(j, cpu_map) {
6332 struct sched_domain *sd;
6335 sd = *per_cpu_ptr(sdd->sd, j);
6336 if (sd && (sd->flags & SD_OVERLAP))
6337 free_sched_groups(sd->groups, 0);
6338 kfree(*per_cpu_ptr(sdd->sd, j));
6342 kfree(*per_cpu_ptr(sdd->sg, j));
6344 kfree(*per_cpu_ptr(sdd->sgp, j));
6346 free_percpu(sdd->sd);
6348 free_percpu(sdd->sg);
6350 free_percpu(sdd->sgp);
6355 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6356 struct s_data *d, const struct cpumask *cpu_map,
6357 struct sched_domain_attr *attr, struct sched_domain *child,
6360 struct sched_domain *sd = tl->init(tl, cpu);
6364 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6366 sd->level = child->level + 1;
6367 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6371 set_domain_attribute(sd, attr);
6377 * Build sched domains for a given set of cpus and attach the sched domains
6378 * to the individual cpus
6380 static int build_sched_domains(const struct cpumask *cpu_map,
6381 struct sched_domain_attr *attr)
6383 enum s_alloc alloc_state = sa_none;
6384 struct sched_domain *sd;
6386 int i, ret = -ENOMEM;
6388 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6389 if (alloc_state != sa_rootdomain)
6392 /* Set up domains for cpus specified by the cpu_map. */
6393 for_each_cpu(i, cpu_map) {
6394 struct sched_domain_topology_level *tl;
6397 for (tl = sched_domain_topology; tl->init; tl++) {
6398 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6399 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6400 sd->flags |= SD_OVERLAP;
6401 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6408 *per_cpu_ptr(d.sd, i) = sd;
6411 /* Build the groups for the domains */
6412 for_each_cpu(i, cpu_map) {
6413 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6414 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6415 if (sd->flags & SD_OVERLAP) {
6416 if (build_overlap_sched_groups(sd, i))
6419 if (build_sched_groups(sd, i))
6425 /* Calculate CPU power for physical packages and nodes */
6426 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6427 if (!cpumask_test_cpu(i, cpu_map))
6430 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6431 claim_allocations(i, sd);
6432 init_sched_groups_power(i, sd);
6436 /* Attach the domains */
6438 for_each_cpu(i, cpu_map) {
6439 sd = *per_cpu_ptr(d.sd, i);
6440 cpu_attach_domain(sd, d.rd, i);
6446 __free_domain_allocs(&d, alloc_state, cpu_map);
6450 static cpumask_var_t *doms_cur; /* current sched domains */
6451 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6452 static struct sched_domain_attr *dattr_cur;
6453 /* attribues of custom domains in 'doms_cur' */
6456 * Special case: If a kmalloc of a doms_cur partition (array of
6457 * cpumask) fails, then fallback to a single sched domain,
6458 * as determined by the single cpumask fallback_doms.
6460 static cpumask_var_t fallback_doms;
6463 * arch_update_cpu_topology lets virtualized architectures update the
6464 * cpu core maps. It is supposed to return 1 if the topology changed
6465 * or 0 if it stayed the same.
6467 int __attribute__((weak)) arch_update_cpu_topology(void)
6472 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6475 cpumask_var_t *doms;
6477 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6480 for (i = 0; i < ndoms; i++) {
6481 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6482 free_sched_domains(doms, i);
6489 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6492 for (i = 0; i < ndoms; i++)
6493 free_cpumask_var(doms[i]);
6498 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6499 * For now this just excludes isolated cpus, but could be used to
6500 * exclude other special cases in the future.
6502 static int init_sched_domains(const struct cpumask *cpu_map)
6506 arch_update_cpu_topology();
6508 doms_cur = alloc_sched_domains(ndoms_cur);
6510 doms_cur = &fallback_doms;
6511 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6512 err = build_sched_domains(doms_cur[0], NULL);
6513 register_sched_domain_sysctl();
6519 * Detach sched domains from a group of cpus specified in cpu_map
6520 * These cpus will now be attached to the NULL domain
6522 static void detach_destroy_domains(const struct cpumask *cpu_map)
6527 for_each_cpu(i, cpu_map)
6528 cpu_attach_domain(NULL, &def_root_domain, i);
6532 /* handle null as "default" */
6533 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6534 struct sched_domain_attr *new, int idx_new)
6536 struct sched_domain_attr tmp;
6543 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6544 new ? (new + idx_new) : &tmp,
6545 sizeof(struct sched_domain_attr));
6549 * Partition sched domains as specified by the 'ndoms_new'
6550 * cpumasks in the array doms_new[] of cpumasks. This compares
6551 * doms_new[] to the current sched domain partitioning, doms_cur[].
6552 * It destroys each deleted domain and builds each new domain.
6554 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6555 * The masks don't intersect (don't overlap.) We should setup one
6556 * sched domain for each mask. CPUs not in any of the cpumasks will
6557 * not be load balanced. If the same cpumask appears both in the
6558 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6561 * The passed in 'doms_new' should be allocated using
6562 * alloc_sched_domains. This routine takes ownership of it and will
6563 * free_sched_domains it when done with it. If the caller failed the
6564 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6565 * and partition_sched_domains() will fallback to the single partition
6566 * 'fallback_doms', it also forces the domains to be rebuilt.
6568 * If doms_new == NULL it will be replaced with cpu_online_mask.
6569 * ndoms_new == 0 is a special case for destroying existing domains,
6570 * and it will not create the default domain.
6572 * Call with hotplug lock held
6574 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6575 struct sched_domain_attr *dattr_new)
6580 mutex_lock(&sched_domains_mutex);
6582 /* always unregister in case we don't destroy any domains */
6583 unregister_sched_domain_sysctl();
6585 /* Let architecture update cpu core mappings. */
6586 new_topology = arch_update_cpu_topology();
6588 n = doms_new ? ndoms_new : 0;
6590 /* Destroy deleted domains */
6591 for (i = 0; i < ndoms_cur; i++) {
6592 for (j = 0; j < n && !new_topology; j++) {
6593 if (cpumask_equal(doms_cur[i], doms_new[j])
6594 && dattrs_equal(dattr_cur, i, dattr_new, j))
6597 /* no match - a current sched domain not in new doms_new[] */
6598 detach_destroy_domains(doms_cur[i]);
6603 if (doms_new == NULL) {
6605 doms_new = &fallback_doms;
6606 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6607 WARN_ON_ONCE(dattr_new);
6610 /* Build new domains */
6611 for (i = 0; i < ndoms_new; i++) {
6612 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6613 if (cpumask_equal(doms_new[i], doms_cur[j])
6614 && dattrs_equal(dattr_new, i, dattr_cur, j))
6617 /* no match - add a new doms_new */
6618 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6623 /* Remember the new sched domains */
6624 if (doms_cur != &fallback_doms)
6625 free_sched_domains(doms_cur, ndoms_cur);
6626 kfree(dattr_cur); /* kfree(NULL) is safe */
6627 doms_cur = doms_new;
6628 dattr_cur = dattr_new;
6629 ndoms_cur = ndoms_new;
6631 register_sched_domain_sysctl();
6633 mutex_unlock(&sched_domains_mutex);
6636 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6639 * Update cpusets according to cpu_active mask. If cpusets are
6640 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6641 * around partition_sched_domains().
6643 * If we come here as part of a suspend/resume, don't touch cpusets because we
6644 * want to restore it back to its original state upon resume anyway.
6646 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6650 case CPU_ONLINE_FROZEN:
6651 case CPU_DOWN_FAILED_FROZEN:
6654 * num_cpus_frozen tracks how many CPUs are involved in suspend
6655 * resume sequence. As long as this is not the last online
6656 * operation in the resume sequence, just build a single sched
6657 * domain, ignoring cpusets.
6660 if (likely(num_cpus_frozen)) {
6661 partition_sched_domains(1, NULL, NULL);
6666 * This is the last CPU online operation. So fall through and
6667 * restore the original sched domains by considering the
6668 * cpuset configurations.
6672 case CPU_DOWN_FAILED:
6673 cpuset_update_active_cpus(true);
6681 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6685 case CPU_DOWN_PREPARE:
6686 cpuset_update_active_cpus(false);
6688 case CPU_DOWN_PREPARE_FROZEN:
6690 partition_sched_domains(1, NULL, NULL);
6698 void __init sched_init_smp(void)
6700 cpumask_var_t non_isolated_cpus;
6702 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6703 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6708 mutex_lock(&sched_domains_mutex);
6709 init_sched_domains(cpu_active_mask);
6710 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6711 if (cpumask_empty(non_isolated_cpus))
6712 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6713 mutex_unlock(&sched_domains_mutex);
6716 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6717 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6718 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6720 /* RT runtime code needs to handle some hotplug events */
6721 hotcpu_notifier(update_runtime, 0);
6725 /* Move init over to a non-isolated CPU */
6726 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6728 sched_init_granularity();
6729 free_cpumask_var(non_isolated_cpus);
6731 init_sched_rt_class();
6734 void __init sched_init_smp(void)
6736 sched_init_granularity();
6738 #endif /* CONFIG_SMP */
6740 const_debug unsigned int sysctl_timer_migration = 1;
6742 int in_sched_functions(unsigned long addr)
6744 return in_lock_functions(addr) ||
6745 (addr >= (unsigned long)__sched_text_start
6746 && addr < (unsigned long)__sched_text_end);
6749 #ifdef CONFIG_CGROUP_SCHED
6750 struct task_group root_task_group;
6751 LIST_HEAD(task_groups);
6754 DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
6756 void __init sched_init(void)
6759 unsigned long alloc_size = 0, ptr;
6761 #ifdef CONFIG_FAIR_GROUP_SCHED
6762 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6764 #ifdef CONFIG_RT_GROUP_SCHED
6765 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6767 #ifdef CONFIG_CPUMASK_OFFSTACK
6768 alloc_size += num_possible_cpus() * cpumask_size();
6771 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6773 #ifdef CONFIG_FAIR_GROUP_SCHED
6774 root_task_group.se = (struct sched_entity **)ptr;
6775 ptr += nr_cpu_ids * sizeof(void **);
6777 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6778 ptr += nr_cpu_ids * sizeof(void **);
6780 #endif /* CONFIG_FAIR_GROUP_SCHED */
6781 #ifdef CONFIG_RT_GROUP_SCHED
6782 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6783 ptr += nr_cpu_ids * sizeof(void **);
6785 root_task_group.rt_rq = (struct rt_rq **)ptr;
6786 ptr += nr_cpu_ids * sizeof(void **);
6788 #endif /* CONFIG_RT_GROUP_SCHED */
6789 #ifdef CONFIG_CPUMASK_OFFSTACK
6790 for_each_possible_cpu(i) {
6791 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
6792 ptr += cpumask_size();
6794 #endif /* CONFIG_CPUMASK_OFFSTACK */
6798 init_defrootdomain();
6801 init_rt_bandwidth(&def_rt_bandwidth,
6802 global_rt_period(), global_rt_runtime());
6804 #ifdef CONFIG_RT_GROUP_SCHED
6805 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6806 global_rt_period(), global_rt_runtime());
6807 #endif /* CONFIG_RT_GROUP_SCHED */
6809 #ifdef CONFIG_CGROUP_SCHED
6810 list_add(&root_task_group.list, &task_groups);
6811 INIT_LIST_HEAD(&root_task_group.children);
6812 INIT_LIST_HEAD(&root_task_group.siblings);
6813 autogroup_init(&init_task);
6815 #endif /* CONFIG_CGROUP_SCHED */
6817 #ifdef CONFIG_CGROUP_CPUACCT
6818 root_cpuacct.cpustat = &kernel_cpustat;
6819 root_cpuacct.cpuusage = alloc_percpu(u64);
6820 /* Too early, not expected to fail */
6821 BUG_ON(!root_cpuacct.cpuusage);
6823 for_each_possible_cpu(i) {
6827 raw_spin_lock_init(&rq->lock);
6829 rq->calc_load_active = 0;
6830 rq->calc_load_update = jiffies + LOAD_FREQ;
6831 init_cfs_rq(&rq->cfs);
6832 init_rt_rq(&rq->rt, rq);
6833 #ifdef CONFIG_FAIR_GROUP_SCHED
6834 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6835 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6837 * How much cpu bandwidth does root_task_group get?
6839 * In case of task-groups formed thr' the cgroup filesystem, it
6840 * gets 100% of the cpu resources in the system. This overall
6841 * system cpu resource is divided among the tasks of
6842 * root_task_group and its child task-groups in a fair manner,
6843 * based on each entity's (task or task-group's) weight
6844 * (se->load.weight).
6846 * In other words, if root_task_group has 10 tasks of weight
6847 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6848 * then A0's share of the cpu resource is:
6850 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6852 * We achieve this by letting root_task_group's tasks sit
6853 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6855 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6856 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6857 #endif /* CONFIG_FAIR_GROUP_SCHED */
6859 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6860 #ifdef CONFIG_RT_GROUP_SCHED
6861 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6862 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6865 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6866 rq->cpu_load[j] = 0;
6868 rq->last_load_update_tick = jiffies;
6873 rq->cpu_power = SCHED_POWER_SCALE;
6874 rq->post_schedule = 0;
6875 rq->active_balance = 0;
6876 rq->next_balance = jiffies;
6881 rq->avg_idle = 2*sysctl_sched_migration_cost;
6883 INIT_LIST_HEAD(&rq->cfs_tasks);
6885 rq_attach_root(rq, &def_root_domain);
6891 atomic_set(&rq->nr_iowait, 0);
6894 set_load_weight(&init_task);
6896 #ifdef CONFIG_PREEMPT_NOTIFIERS
6897 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6900 #ifdef CONFIG_RT_MUTEXES
6901 plist_head_init(&init_task.pi_waiters);
6905 * The boot idle thread does lazy MMU switching as well:
6907 atomic_inc(&init_mm.mm_count);
6908 enter_lazy_tlb(&init_mm, current);
6911 * Make us the idle thread. Technically, schedule() should not be
6912 * called from this thread, however somewhere below it might be,
6913 * but because we are the idle thread, we just pick up running again
6914 * when this runqueue becomes "idle".
6916 init_idle(current, smp_processor_id());
6918 calc_load_update = jiffies + LOAD_FREQ;
6921 * During early bootup we pretend to be a normal task:
6923 current->sched_class = &fair_sched_class;
6926 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6927 /* May be allocated at isolcpus cmdline parse time */
6928 if (cpu_isolated_map == NULL)
6929 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6930 idle_thread_set_boot_cpu();
6932 init_sched_fair_class();
6934 scheduler_running = 1;
6937 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6938 static inline int preempt_count_equals(int preempt_offset)
6940 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6942 return (nested == preempt_offset);
6945 void __might_sleep(const char *file, int line, int preempt_offset)
6947 static unsigned long prev_jiffy; /* ratelimiting */
6949 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6950 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6951 system_state != SYSTEM_RUNNING || oops_in_progress)
6953 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6955 prev_jiffy = jiffies;
6958 "BUG: sleeping function called from invalid context at %s:%d\n",
6961 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6962 in_atomic(), irqs_disabled(),
6963 current->pid, current->comm);
6965 debug_show_held_locks(current);
6966 if (irqs_disabled())
6967 print_irqtrace_events(current);
6970 EXPORT_SYMBOL(__might_sleep);
6973 #ifdef CONFIG_MAGIC_SYSRQ
6974 static void normalize_task(struct rq *rq, struct task_struct *p)
6976 const struct sched_class *prev_class = p->sched_class;
6977 int old_prio = p->prio;
6982 dequeue_task(rq, p, 0);
6983 __setscheduler(rq, p, SCHED_NORMAL, 0);
6985 enqueue_task(rq, p, 0);
6986 resched_task(rq->curr);
6989 check_class_changed(rq, p, prev_class, old_prio);
6992 void normalize_rt_tasks(void)
6994 struct task_struct *g, *p;
6995 unsigned long flags;
6998 read_lock_irqsave(&tasklist_lock, flags);
6999 do_each_thread(g, p) {
7001 * Only normalize user tasks:
7006 p->se.exec_start = 0;
7007 #ifdef CONFIG_SCHEDSTATS
7008 p->se.statistics.wait_start = 0;
7009 p->se.statistics.sleep_start = 0;
7010 p->se.statistics.block_start = 0;
7015 * Renice negative nice level userspace
7018 if (TASK_NICE(p) < 0 && p->mm)
7019 set_user_nice(p, 0);
7023 raw_spin_lock(&p->pi_lock);
7024 rq = __task_rq_lock(p);
7026 normalize_task(rq, p);
7028 __task_rq_unlock(rq);
7029 raw_spin_unlock(&p->pi_lock);
7030 } while_each_thread(g, p);
7032 read_unlock_irqrestore(&tasklist_lock, flags);
7035 #endif /* CONFIG_MAGIC_SYSRQ */
7037 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7039 * These functions are only useful for the IA64 MCA handling, or kdb.
7041 * They can only be called when the whole system has been
7042 * stopped - every CPU needs to be quiescent, and no scheduling
7043 * activity can take place. Using them for anything else would
7044 * be a serious bug, and as a result, they aren't even visible
7045 * under any other configuration.
7049 * curr_task - return the current task for a given cpu.
7050 * @cpu: the processor in question.
7052 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7054 struct task_struct *curr_task(int cpu)
7056 return cpu_curr(cpu);
7059 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7063 * set_curr_task - set the current task for a given cpu.
7064 * @cpu: the processor in question.
7065 * @p: the task pointer to set.
7067 * Description: This function must only be used when non-maskable interrupts
7068 * are serviced on a separate stack. It allows the architecture to switch the
7069 * notion of the current task on a cpu in a non-blocking manner. This function
7070 * must be called with all CPU's synchronized, and interrupts disabled, the
7071 * and caller must save the original value of the current task (see
7072 * curr_task() above) and restore that value before reenabling interrupts and
7073 * re-starting the system.
7075 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7077 void set_curr_task(int cpu, struct task_struct *p)
7084 #ifdef CONFIG_CGROUP_SCHED
7085 /* task_group_lock serializes the addition/removal of task groups */
7086 static DEFINE_SPINLOCK(task_group_lock);
7088 static void free_sched_group(struct task_group *tg)
7090 free_fair_sched_group(tg);
7091 free_rt_sched_group(tg);
7096 /* allocate runqueue etc for a new task group */
7097 struct task_group *sched_create_group(struct task_group *parent)
7099 struct task_group *tg;
7100 unsigned long flags;
7102 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7104 return ERR_PTR(-ENOMEM);
7106 if (!alloc_fair_sched_group(tg, parent))
7109 if (!alloc_rt_sched_group(tg, parent))
7112 spin_lock_irqsave(&task_group_lock, flags);
7113 list_add_rcu(&tg->list, &task_groups);
7115 WARN_ON(!parent); /* root should already exist */
7117 tg->parent = parent;
7118 INIT_LIST_HEAD(&tg->children);
7119 list_add_rcu(&tg->siblings, &parent->children);
7120 spin_unlock_irqrestore(&task_group_lock, flags);
7125 free_sched_group(tg);
7126 return ERR_PTR(-ENOMEM);
7129 /* rcu callback to free various structures associated with a task group */
7130 static void free_sched_group_rcu(struct rcu_head *rhp)
7132 /* now it should be safe to free those cfs_rqs */
7133 free_sched_group(container_of(rhp, struct task_group, rcu));
7136 /* Destroy runqueue etc associated with a task group */
7137 void sched_destroy_group(struct task_group *tg)
7139 unsigned long flags;
7142 /* end participation in shares distribution */
7143 for_each_possible_cpu(i)
7144 unregister_fair_sched_group(tg, i);
7146 spin_lock_irqsave(&task_group_lock, flags);
7147 list_del_rcu(&tg->list);
7148 list_del_rcu(&tg->siblings);
7149 spin_unlock_irqrestore(&task_group_lock, flags);
7151 /* wait for possible concurrent references to cfs_rqs complete */
7152 call_rcu(&tg->rcu, free_sched_group_rcu);
7155 /* change task's runqueue when it moves between groups.
7156 * The caller of this function should have put the task in its new group
7157 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7158 * reflect its new group.
7160 void sched_move_task(struct task_struct *tsk)
7162 struct task_group *tg;
7164 unsigned long flags;
7167 rq = task_rq_lock(tsk, &flags);
7169 running = task_current(rq, tsk);
7173 dequeue_task(rq, tsk, 0);
7174 if (unlikely(running))
7175 tsk->sched_class->put_prev_task(rq, tsk);
7177 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7178 lockdep_is_held(&tsk->sighand->siglock)),
7179 struct task_group, css);
7180 tg = autogroup_task_group(tsk, tg);
7181 tsk->sched_task_group = tg;
7183 #ifdef CONFIG_FAIR_GROUP_SCHED
7184 if (tsk->sched_class->task_move_group)
7185 tsk->sched_class->task_move_group(tsk, on_rq);
7188 set_task_rq(tsk, task_cpu(tsk));
7190 if (unlikely(running))
7191 tsk->sched_class->set_curr_task(rq);
7193 enqueue_task(rq, tsk, 0);
7195 task_rq_unlock(rq, tsk, &flags);
7197 #endif /* CONFIG_CGROUP_SCHED */
7199 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7200 static unsigned long to_ratio(u64 period, u64 runtime)
7202 if (runtime == RUNTIME_INF)
7205 return div64_u64(runtime << 20, period);
7209 #ifdef CONFIG_RT_GROUP_SCHED
7211 * Ensure that the real time constraints are schedulable.
7213 static DEFINE_MUTEX(rt_constraints_mutex);
7215 /* Must be called with tasklist_lock held */
7216 static inline int tg_has_rt_tasks(struct task_group *tg)
7218 struct task_struct *g, *p;
7220 do_each_thread(g, p) {
7221 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7223 } while_each_thread(g, p);
7228 struct rt_schedulable_data {
7229 struct task_group *tg;
7234 static int tg_rt_schedulable(struct task_group *tg, void *data)
7236 struct rt_schedulable_data *d = data;
7237 struct task_group *child;
7238 unsigned long total, sum = 0;
7239 u64 period, runtime;
7241 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7242 runtime = tg->rt_bandwidth.rt_runtime;
7245 period = d->rt_period;
7246 runtime = d->rt_runtime;
7250 * Cannot have more runtime than the period.
7252 if (runtime > period && runtime != RUNTIME_INF)
7256 * Ensure we don't starve existing RT tasks.
7258 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7261 total = to_ratio(period, runtime);
7264 * Nobody can have more than the global setting allows.
7266 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7270 * The sum of our children's runtime should not exceed our own.
7272 list_for_each_entry_rcu(child, &tg->children, siblings) {
7273 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7274 runtime = child->rt_bandwidth.rt_runtime;
7276 if (child == d->tg) {
7277 period = d->rt_period;
7278 runtime = d->rt_runtime;
7281 sum += to_ratio(period, runtime);
7290 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7294 struct rt_schedulable_data data = {
7296 .rt_period = period,
7297 .rt_runtime = runtime,
7301 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7307 static int tg_set_rt_bandwidth(struct task_group *tg,
7308 u64 rt_period, u64 rt_runtime)
7312 mutex_lock(&rt_constraints_mutex);
7313 read_lock(&tasklist_lock);
7314 err = __rt_schedulable(tg, rt_period, rt_runtime);
7318 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7319 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7320 tg->rt_bandwidth.rt_runtime = rt_runtime;
7322 for_each_possible_cpu(i) {
7323 struct rt_rq *rt_rq = tg->rt_rq[i];
7325 raw_spin_lock(&rt_rq->rt_runtime_lock);
7326 rt_rq->rt_runtime = rt_runtime;
7327 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7329 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7331 read_unlock(&tasklist_lock);
7332 mutex_unlock(&rt_constraints_mutex);
7337 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7339 u64 rt_runtime, rt_period;
7341 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7342 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7343 if (rt_runtime_us < 0)
7344 rt_runtime = RUNTIME_INF;
7346 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7349 long sched_group_rt_runtime(struct task_group *tg)
7353 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7356 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7357 do_div(rt_runtime_us, NSEC_PER_USEC);
7358 return rt_runtime_us;
7361 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7363 u64 rt_runtime, rt_period;
7365 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7366 rt_runtime = tg->rt_bandwidth.rt_runtime;
7371 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7374 long sched_group_rt_period(struct task_group *tg)
7378 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7379 do_div(rt_period_us, NSEC_PER_USEC);
7380 return rt_period_us;
7383 static int sched_rt_global_constraints(void)
7385 u64 runtime, period;
7388 if (sysctl_sched_rt_period <= 0)
7391 runtime = global_rt_runtime();
7392 period = global_rt_period();
7395 * Sanity check on the sysctl variables.
7397 if (runtime > period && runtime != RUNTIME_INF)
7400 mutex_lock(&rt_constraints_mutex);
7401 read_lock(&tasklist_lock);
7402 ret = __rt_schedulable(NULL, 0, 0);
7403 read_unlock(&tasklist_lock);
7404 mutex_unlock(&rt_constraints_mutex);
7409 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7411 /* Don't accept realtime tasks when there is no way for them to run */
7412 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7418 #else /* !CONFIG_RT_GROUP_SCHED */
7419 static int sched_rt_global_constraints(void)
7421 unsigned long flags;
7424 if (sysctl_sched_rt_period <= 0)
7428 * There's always some RT tasks in the root group
7429 * -- migration, kstopmachine etc..
7431 if (sysctl_sched_rt_runtime == 0)
7434 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7435 for_each_possible_cpu(i) {
7436 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7438 raw_spin_lock(&rt_rq->rt_runtime_lock);
7439 rt_rq->rt_runtime = global_rt_runtime();
7440 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7442 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7446 #endif /* CONFIG_RT_GROUP_SCHED */
7448 int sched_rt_handler(struct ctl_table *table, int write,
7449 void __user *buffer, size_t *lenp,
7453 int old_period, old_runtime;
7454 static DEFINE_MUTEX(mutex);
7457 old_period = sysctl_sched_rt_period;
7458 old_runtime = sysctl_sched_rt_runtime;
7460 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7462 if (!ret && write) {
7463 ret = sched_rt_global_constraints();
7465 sysctl_sched_rt_period = old_period;
7466 sysctl_sched_rt_runtime = old_runtime;
7468 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7469 def_rt_bandwidth.rt_period =
7470 ns_to_ktime(global_rt_period());
7473 mutex_unlock(&mutex);
7478 #ifdef CONFIG_CGROUP_SCHED
7480 /* return corresponding task_group object of a cgroup */
7481 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7483 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7484 struct task_group, css);
7487 static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp)
7489 struct task_group *tg, *parent;
7491 if (!cgrp->parent) {
7492 /* This is early initialization for the top cgroup */
7493 return &root_task_group.css;
7496 parent = cgroup_tg(cgrp->parent);
7497 tg = sched_create_group(parent);
7499 return ERR_PTR(-ENOMEM);
7504 static void cpu_cgroup_destroy(struct cgroup *cgrp)
7506 struct task_group *tg = cgroup_tg(cgrp);
7508 sched_destroy_group(tg);
7511 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7512 struct cgroup_taskset *tset)
7514 struct task_struct *task;
7516 cgroup_taskset_for_each(task, cgrp, tset) {
7517 #ifdef CONFIG_RT_GROUP_SCHED
7518 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7521 /* We don't support RT-tasks being in separate groups */
7522 if (task->sched_class != &fair_sched_class)
7529 static void cpu_cgroup_attach(struct cgroup *cgrp,
7530 struct cgroup_taskset *tset)
7532 struct task_struct *task;
7534 cgroup_taskset_for_each(task, cgrp, tset)
7535 sched_move_task(task);
7539 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7540 struct task_struct *task)
7543 * cgroup_exit() is called in the copy_process() failure path.
7544 * Ignore this case since the task hasn't ran yet, this avoids
7545 * trying to poke a half freed task state from generic code.
7547 if (!(task->flags & PF_EXITING))
7550 sched_move_task(task);
7553 #ifdef CONFIG_FAIR_GROUP_SCHED
7554 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7557 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7560 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7562 struct task_group *tg = cgroup_tg(cgrp);
7564 return (u64) scale_load_down(tg->shares);
7567 #ifdef CONFIG_CFS_BANDWIDTH
7568 static DEFINE_MUTEX(cfs_constraints_mutex);
7570 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7571 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7573 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7575 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7577 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7578 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7580 if (tg == &root_task_group)
7584 * Ensure we have at some amount of bandwidth every period. This is
7585 * to prevent reaching a state of large arrears when throttled via
7586 * entity_tick() resulting in prolonged exit starvation.
7588 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7592 * Likewise, bound things on the otherside by preventing insane quota
7593 * periods. This also allows us to normalize in computing quota
7596 if (period > max_cfs_quota_period)
7599 mutex_lock(&cfs_constraints_mutex);
7600 ret = __cfs_schedulable(tg, period, quota);
7604 runtime_enabled = quota != RUNTIME_INF;
7605 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7606 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7607 raw_spin_lock_irq(&cfs_b->lock);
7608 cfs_b->period = ns_to_ktime(period);
7609 cfs_b->quota = quota;
7611 __refill_cfs_bandwidth_runtime(cfs_b);
7612 /* restart the period timer (if active) to handle new period expiry */
7613 if (runtime_enabled && cfs_b->timer_active) {
7614 /* force a reprogram */
7615 cfs_b->timer_active = 0;
7616 __start_cfs_bandwidth(cfs_b);
7618 raw_spin_unlock_irq(&cfs_b->lock);
7620 for_each_possible_cpu(i) {
7621 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7622 struct rq *rq = cfs_rq->rq;
7624 raw_spin_lock_irq(&rq->lock);
7625 cfs_rq->runtime_enabled = runtime_enabled;
7626 cfs_rq->runtime_remaining = 0;
7628 if (cfs_rq->throttled)
7629 unthrottle_cfs_rq(cfs_rq);
7630 raw_spin_unlock_irq(&rq->lock);
7633 mutex_unlock(&cfs_constraints_mutex);
7638 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7642 period = ktime_to_ns(tg->cfs_bandwidth.period);
7643 if (cfs_quota_us < 0)
7644 quota = RUNTIME_INF;
7646 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7648 return tg_set_cfs_bandwidth(tg, period, quota);
7651 long tg_get_cfs_quota(struct task_group *tg)
7655 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7658 quota_us = tg->cfs_bandwidth.quota;
7659 do_div(quota_us, NSEC_PER_USEC);
7664 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7668 period = (u64)cfs_period_us * NSEC_PER_USEC;
7669 quota = tg->cfs_bandwidth.quota;
7671 return tg_set_cfs_bandwidth(tg, period, quota);
7674 long tg_get_cfs_period(struct task_group *tg)
7678 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7679 do_div(cfs_period_us, NSEC_PER_USEC);
7681 return cfs_period_us;
7684 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7686 return tg_get_cfs_quota(cgroup_tg(cgrp));
7689 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7692 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7695 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7697 return tg_get_cfs_period(cgroup_tg(cgrp));
7700 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7703 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7706 struct cfs_schedulable_data {
7707 struct task_group *tg;
7712 * normalize group quota/period to be quota/max_period
7713 * note: units are usecs
7715 static u64 normalize_cfs_quota(struct task_group *tg,
7716 struct cfs_schedulable_data *d)
7724 period = tg_get_cfs_period(tg);
7725 quota = tg_get_cfs_quota(tg);
7728 /* note: these should typically be equivalent */
7729 if (quota == RUNTIME_INF || quota == -1)
7732 return to_ratio(period, quota);
7735 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7737 struct cfs_schedulable_data *d = data;
7738 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7739 s64 quota = 0, parent_quota = -1;
7742 quota = RUNTIME_INF;
7744 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7746 quota = normalize_cfs_quota(tg, d);
7747 parent_quota = parent_b->hierarchal_quota;
7750 * ensure max(child_quota) <= parent_quota, inherit when no
7753 if (quota == RUNTIME_INF)
7754 quota = parent_quota;
7755 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7758 cfs_b->hierarchal_quota = quota;
7763 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7766 struct cfs_schedulable_data data = {
7772 if (quota != RUNTIME_INF) {
7773 do_div(data.period, NSEC_PER_USEC);
7774 do_div(data.quota, NSEC_PER_USEC);
7778 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7784 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7785 struct cgroup_map_cb *cb)
7787 struct task_group *tg = cgroup_tg(cgrp);
7788 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7790 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7791 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7792 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7796 #endif /* CONFIG_CFS_BANDWIDTH */
7797 #endif /* CONFIG_FAIR_GROUP_SCHED */
7799 #ifdef CONFIG_RT_GROUP_SCHED
7800 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7803 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7806 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7808 return sched_group_rt_runtime(cgroup_tg(cgrp));
7811 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7814 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
7817 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
7819 return sched_group_rt_period(cgroup_tg(cgrp));
7821 #endif /* CONFIG_RT_GROUP_SCHED */
7823 static struct cftype cpu_files[] = {
7824 #ifdef CONFIG_FAIR_GROUP_SCHED
7827 .read_u64 = cpu_shares_read_u64,
7828 .write_u64 = cpu_shares_write_u64,
7831 #ifdef CONFIG_CFS_BANDWIDTH
7833 .name = "cfs_quota_us",
7834 .read_s64 = cpu_cfs_quota_read_s64,
7835 .write_s64 = cpu_cfs_quota_write_s64,
7838 .name = "cfs_period_us",
7839 .read_u64 = cpu_cfs_period_read_u64,
7840 .write_u64 = cpu_cfs_period_write_u64,
7844 .read_map = cpu_stats_show,
7847 #ifdef CONFIG_RT_GROUP_SCHED
7849 .name = "rt_runtime_us",
7850 .read_s64 = cpu_rt_runtime_read,
7851 .write_s64 = cpu_rt_runtime_write,
7854 .name = "rt_period_us",
7855 .read_u64 = cpu_rt_period_read_uint,
7856 .write_u64 = cpu_rt_period_write_uint,
7862 struct cgroup_subsys cpu_cgroup_subsys = {
7864 .create = cpu_cgroup_create,
7865 .destroy = cpu_cgroup_destroy,
7866 .can_attach = cpu_cgroup_can_attach,
7867 .attach = cpu_cgroup_attach,
7868 .exit = cpu_cgroup_exit,
7869 .subsys_id = cpu_cgroup_subsys_id,
7870 .base_cftypes = cpu_files,
7874 #endif /* CONFIG_CGROUP_SCHED */
7876 #ifdef CONFIG_CGROUP_CPUACCT
7879 * CPU accounting code for task groups.
7881 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
7882 * (balbir@in.ibm.com).
7885 struct cpuacct root_cpuacct;
7887 /* create a new cpu accounting group */
7888 static struct cgroup_subsys_state *cpuacct_create(struct cgroup *cgrp)
7893 return &root_cpuacct.css;
7895 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
7899 ca->cpuusage = alloc_percpu(u64);
7903 ca->cpustat = alloc_percpu(struct kernel_cpustat);
7905 goto out_free_cpuusage;
7910 free_percpu(ca->cpuusage);
7914 return ERR_PTR(-ENOMEM);
7917 /* destroy an existing cpu accounting group */
7918 static void cpuacct_destroy(struct cgroup *cgrp)
7920 struct cpuacct *ca = cgroup_ca(cgrp);
7922 free_percpu(ca->cpustat);
7923 free_percpu(ca->cpuusage);
7927 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
7929 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7932 #ifndef CONFIG_64BIT
7934 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
7936 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7938 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7946 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
7948 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
7950 #ifndef CONFIG_64BIT
7952 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
7954 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
7956 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
7962 /* return total cpu usage (in nanoseconds) of a group */
7963 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
7965 struct cpuacct *ca = cgroup_ca(cgrp);
7966 u64 totalcpuusage = 0;
7969 for_each_present_cpu(i)
7970 totalcpuusage += cpuacct_cpuusage_read(ca, i);
7972 return totalcpuusage;
7975 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
7978 struct cpuacct *ca = cgroup_ca(cgrp);
7987 for_each_present_cpu(i)
7988 cpuacct_cpuusage_write(ca, i, 0);
7994 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
7997 struct cpuacct *ca = cgroup_ca(cgroup);
8001 for_each_present_cpu(i) {
8002 percpu = cpuacct_cpuusage_read(ca, i);
8003 seq_printf(m, "%llu ", (unsigned long long) percpu);
8005 seq_printf(m, "\n");
8009 static const char *cpuacct_stat_desc[] = {
8010 [CPUACCT_STAT_USER] = "user",
8011 [CPUACCT_STAT_SYSTEM] = "system",
8014 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8015 struct cgroup_map_cb *cb)
8017 struct cpuacct *ca = cgroup_ca(cgrp);
8021 for_each_online_cpu(cpu) {
8022 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8023 val += kcpustat->cpustat[CPUTIME_USER];
8024 val += kcpustat->cpustat[CPUTIME_NICE];
8026 val = cputime64_to_clock_t(val);
8027 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val);
8030 for_each_online_cpu(cpu) {
8031 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu);
8032 val += kcpustat->cpustat[CPUTIME_SYSTEM];
8033 val += kcpustat->cpustat[CPUTIME_IRQ];
8034 val += kcpustat->cpustat[CPUTIME_SOFTIRQ];
8037 val = cputime64_to_clock_t(val);
8038 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val);
8043 static struct cftype files[] = {
8046 .read_u64 = cpuusage_read,
8047 .write_u64 = cpuusage_write,
8050 .name = "usage_percpu",
8051 .read_seq_string = cpuacct_percpu_seq_read,
8055 .read_map = cpuacct_stats_show,
8061 * charge this task's execution time to its accounting group.
8063 * called with rq->lock held.
8065 void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8070 if (unlikely(!cpuacct_subsys.active))
8073 cpu = task_cpu(tsk);
8079 for (; ca; ca = parent_ca(ca)) {
8080 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8081 *cpuusage += cputime;
8087 struct cgroup_subsys cpuacct_subsys = {
8089 .create = cpuacct_create,
8090 .destroy = cpuacct_destroy,
8091 .subsys_id = cpuacct_subsys_id,
8092 .base_cftypes = files,
8094 #endif /* CONFIG_CGROUP_CPUACCT */
8096 void dump_cpu_task(int cpu)
8098 pr_info("Task dump for CPU %d:\n", cpu);
8099 sched_show_task(cpu_curr(cpu));