4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 assert_raw_spin_locked(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p))
533 smp_send_reschedule(cpu);
536 void resched_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
541 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 resched_task(cpu_curr(cpu));
544 raw_spin_unlock_irqrestore(&rq->lock, flags);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu = smp_processor_id();
560 struct sched_domain *sd;
563 for_each_domain(cpu, sd) {
564 for_each_cpu(i, sched_domain_span(sd)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 static void wake_up_idle_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
589 if (cpu == smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq->curr != rq->idle)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq->idle);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq->idle))
612 smp_send_reschedule(cpu);
615 static bool wake_up_full_nohz_cpu(int cpu)
617 if (tick_nohz_full_cpu(cpu)) {
618 if (cpu != smp_processor_id() ||
619 tick_nohz_tick_stopped())
620 smp_send_reschedule(cpu);
627 void wake_up_nohz_cpu(int cpu)
629 if (!wake_up_full_nohz_cpu(cpu))
630 wake_up_idle_cpu(cpu);
633 static inline bool got_nohz_idle_kick(void)
635 int cpu = smp_processor_id();
636 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
639 #else /* CONFIG_NO_HZ_COMMON */
641 static inline bool got_nohz_idle_kick(void)
646 #endif /* CONFIG_NO_HZ_COMMON */
648 #ifdef CONFIG_NO_HZ_FULL
649 bool sched_can_stop_tick(void)
655 /* Make sure rq->nr_running update is visible after the IPI */
658 /* More than one running task need preemption */
659 if (rq->nr_running > 1)
664 #endif /* CONFIG_NO_HZ_FULL */
666 void sched_avg_update(struct rq *rq)
668 s64 period = sched_avg_period();
670 while ((s64)(rq->clock - rq->age_stamp) > period) {
672 * Inline assembly required to prevent the compiler
673 * optimising this loop into a divmod call.
674 * See __iter_div_u64_rem() for another example of this.
676 asm("" : "+rm" (rq->age_stamp));
677 rq->age_stamp += period;
682 #else /* !CONFIG_SMP */
683 void resched_task(struct task_struct *p)
685 assert_raw_spin_locked(&task_rq(p)->lock);
686 set_tsk_need_resched(p);
688 #endif /* CONFIG_SMP */
690 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
691 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
693 * Iterate task_group tree rooted at *from, calling @down when first entering a
694 * node and @up when leaving it for the final time.
696 * Caller must hold rcu_lock or sufficient equivalent.
698 int walk_tg_tree_from(struct task_group *from,
699 tg_visitor down, tg_visitor up, void *data)
701 struct task_group *parent, *child;
707 ret = (*down)(parent, data);
710 list_for_each_entry_rcu(child, &parent->children, siblings) {
717 ret = (*up)(parent, data);
718 if (ret || parent == from)
722 parent = parent->parent;
729 int tg_nop(struct task_group *tg, void *data)
735 static void set_load_weight(struct task_struct *p)
737 int prio = p->static_prio - MAX_RT_PRIO;
738 struct load_weight *load = &p->se.load;
741 * SCHED_IDLE tasks get minimal weight:
743 if (p->policy == SCHED_IDLE) {
744 load->weight = scale_load(WEIGHT_IDLEPRIO);
745 load->inv_weight = WMULT_IDLEPRIO;
749 load->weight = scale_load(prio_to_weight[prio]);
750 load->inv_weight = prio_to_wmult[prio];
753 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
756 sched_info_queued(p);
757 p->sched_class->enqueue_task(rq, p, flags);
760 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
763 sched_info_dequeued(p);
764 p->sched_class->dequeue_task(rq, p, flags);
767 void activate_task(struct rq *rq, struct task_struct *p, int flags)
769 if (task_contributes_to_load(p))
770 rq->nr_uninterruptible--;
772 enqueue_task(rq, p, flags);
775 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
777 if (task_contributes_to_load(p))
778 rq->nr_uninterruptible++;
780 dequeue_task(rq, p, flags);
783 static void update_rq_clock_task(struct rq *rq, s64 delta)
786 * In theory, the compile should just see 0 here, and optimize out the call
787 * to sched_rt_avg_update. But I don't trust it...
789 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
790 s64 steal = 0, irq_delta = 0;
792 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
793 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
796 * Since irq_time is only updated on {soft,}irq_exit, we might run into
797 * this case when a previous update_rq_clock() happened inside a
800 * When this happens, we stop ->clock_task and only update the
801 * prev_irq_time stamp to account for the part that fit, so that a next
802 * update will consume the rest. This ensures ->clock_task is
805 * It does however cause some slight miss-attribution of {soft,}irq
806 * time, a more accurate solution would be to update the irq_time using
807 * the current rq->clock timestamp, except that would require using
810 if (irq_delta > delta)
813 rq->prev_irq_time += irq_delta;
816 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
817 if (static_key_false((¶virt_steal_rq_enabled))) {
820 steal = paravirt_steal_clock(cpu_of(rq));
821 steal -= rq->prev_steal_time_rq;
823 if (unlikely(steal > delta))
826 st = steal_ticks(steal);
827 steal = st * TICK_NSEC;
829 rq->prev_steal_time_rq += steal;
835 rq->clock_task += delta;
837 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
838 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
839 sched_rt_avg_update(rq, irq_delta + steal);
843 void sched_set_stop_task(int cpu, struct task_struct *stop)
845 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
846 struct task_struct *old_stop = cpu_rq(cpu)->stop;
850 * Make it appear like a SCHED_FIFO task, its something
851 * userspace knows about and won't get confused about.
853 * Also, it will make PI more or less work without too
854 * much confusion -- but then, stop work should not
855 * rely on PI working anyway.
857 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
859 stop->sched_class = &stop_sched_class;
862 cpu_rq(cpu)->stop = stop;
866 * Reset it back to a normal scheduling class so that
867 * it can die in pieces.
869 old_stop->sched_class = &rt_sched_class;
874 * __normal_prio - return the priority that is based on the static prio
876 static inline int __normal_prio(struct task_struct *p)
878 return p->static_prio;
882 * Calculate the expected normal priority: i.e. priority
883 * without taking RT-inheritance into account. Might be
884 * boosted by interactivity modifiers. Changes upon fork,
885 * setprio syscalls, and whenever the interactivity
886 * estimator recalculates.
888 static inline int normal_prio(struct task_struct *p)
892 if (task_has_rt_policy(p))
893 prio = MAX_RT_PRIO-1 - p->rt_priority;
895 prio = __normal_prio(p);
900 * Calculate the current priority, i.e. the priority
901 * taken into account by the scheduler. This value might
902 * be boosted by RT tasks, or might be boosted by
903 * interactivity modifiers. Will be RT if the task got
904 * RT-boosted. If not then it returns p->normal_prio.
906 static int effective_prio(struct task_struct *p)
908 p->normal_prio = normal_prio(p);
910 * If we are RT tasks or we were boosted to RT priority,
911 * keep the priority unchanged. Otherwise, update priority
912 * to the normal priority:
914 if (!rt_prio(p->prio))
915 return p->normal_prio;
920 * task_curr - is this task currently executing on a CPU?
921 * @p: the task in question.
923 inline int task_curr(const struct task_struct *p)
925 return cpu_curr(task_cpu(p)) == p;
928 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
929 const struct sched_class *prev_class,
932 if (prev_class != p->sched_class) {
933 if (prev_class->switched_from)
934 prev_class->switched_from(rq, p);
935 p->sched_class->switched_to(rq, p);
936 } else if (oldprio != p->prio)
937 p->sched_class->prio_changed(rq, p, oldprio);
940 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
942 const struct sched_class *class;
944 if (p->sched_class == rq->curr->sched_class) {
945 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
947 for_each_class(class) {
948 if (class == rq->curr->sched_class)
950 if (class == p->sched_class) {
951 resched_task(rq->curr);
958 * A queue event has occurred, and we're going to schedule. In
959 * this case, we can save a useless back to back clock update.
961 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
962 rq->skip_clock_update = 1;
965 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
967 void register_task_migration_notifier(struct notifier_block *n)
969 atomic_notifier_chain_register(&task_migration_notifier, n);
973 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
975 #ifdef CONFIG_SCHED_DEBUG
977 * We should never call set_task_cpu() on a blocked task,
978 * ttwu() will sort out the placement.
980 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
981 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
983 #ifdef CONFIG_LOCKDEP
985 * The caller should hold either p->pi_lock or rq->lock, when changing
986 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
988 * sched_move_task() holds both and thus holding either pins the cgroup,
991 * Furthermore, all task_rq users should acquire both locks, see
994 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
995 lockdep_is_held(&task_rq(p)->lock)));
999 trace_sched_migrate_task(p, new_cpu);
1001 if (task_cpu(p) != new_cpu) {
1002 struct task_migration_notifier tmn;
1004 if (p->sched_class->migrate_task_rq)
1005 p->sched_class->migrate_task_rq(p, new_cpu);
1006 p->se.nr_migrations++;
1007 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1010 tmn.from_cpu = task_cpu(p);
1011 tmn.to_cpu = new_cpu;
1013 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1016 __set_task_cpu(p, new_cpu);
1019 struct migration_arg {
1020 struct task_struct *task;
1024 static int migration_cpu_stop(void *data);
1027 * wait_task_inactive - wait for a thread to unschedule.
1029 * If @match_state is nonzero, it's the @p->state value just checked and
1030 * not expected to change. If it changes, i.e. @p might have woken up,
1031 * then return zero. When we succeed in waiting for @p to be off its CPU,
1032 * we return a positive number (its total switch count). If a second call
1033 * a short while later returns the same number, the caller can be sure that
1034 * @p has remained unscheduled the whole time.
1036 * The caller must ensure that the task *will* unschedule sometime soon,
1037 * else this function might spin for a *long* time. This function can't
1038 * be called with interrupts off, or it may introduce deadlock with
1039 * smp_call_function() if an IPI is sent by the same process we are
1040 * waiting to become inactive.
1042 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1044 unsigned long flags;
1051 * We do the initial early heuristics without holding
1052 * any task-queue locks at all. We'll only try to get
1053 * the runqueue lock when things look like they will
1059 * If the task is actively running on another CPU
1060 * still, just relax and busy-wait without holding
1063 * NOTE! Since we don't hold any locks, it's not
1064 * even sure that "rq" stays as the right runqueue!
1065 * But we don't care, since "task_running()" will
1066 * return false if the runqueue has changed and p
1067 * is actually now running somewhere else!
1069 while (task_running(rq, p)) {
1070 if (match_state && unlikely(p->state != match_state))
1076 * Ok, time to look more closely! We need the rq
1077 * lock now, to be *sure*. If we're wrong, we'll
1078 * just go back and repeat.
1080 rq = task_rq_lock(p, &flags);
1081 trace_sched_wait_task(p);
1082 running = task_running(rq, p);
1085 if (!match_state || p->state == match_state)
1086 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1087 task_rq_unlock(rq, p, &flags);
1090 * If it changed from the expected state, bail out now.
1092 if (unlikely(!ncsw))
1096 * Was it really running after all now that we
1097 * checked with the proper locks actually held?
1099 * Oops. Go back and try again..
1101 if (unlikely(running)) {
1107 * It's not enough that it's not actively running,
1108 * it must be off the runqueue _entirely_, and not
1111 * So if it was still runnable (but just not actively
1112 * running right now), it's preempted, and we should
1113 * yield - it could be a while.
1115 if (unlikely(on_rq)) {
1116 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1118 set_current_state(TASK_UNINTERRUPTIBLE);
1119 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1124 * Ahh, all good. It wasn't running, and it wasn't
1125 * runnable, which means that it will never become
1126 * running in the future either. We're all done!
1135 * kick_process - kick a running thread to enter/exit the kernel
1136 * @p: the to-be-kicked thread
1138 * Cause a process which is running on another CPU to enter
1139 * kernel-mode, without any delay. (to get signals handled.)
1141 * NOTE: this function doesn't have to take the runqueue lock,
1142 * because all it wants to ensure is that the remote task enters
1143 * the kernel. If the IPI races and the task has been migrated
1144 * to another CPU then no harm is done and the purpose has been
1147 void kick_process(struct task_struct *p)
1153 if ((cpu != smp_processor_id()) && task_curr(p))
1154 smp_send_reschedule(cpu);
1157 EXPORT_SYMBOL_GPL(kick_process);
1158 #endif /* CONFIG_SMP */
1162 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1164 static int select_fallback_rq(int cpu, struct task_struct *p)
1166 int nid = cpu_to_node(cpu);
1167 const struct cpumask *nodemask = NULL;
1168 enum { cpuset, possible, fail } state = cpuset;
1172 * If the node that the cpu is on has been offlined, cpu_to_node()
1173 * will return -1. There is no cpu on the node, and we should
1174 * select the cpu on the other node.
1177 nodemask = cpumask_of_node(nid);
1179 /* Look for allowed, online CPU in same node. */
1180 for_each_cpu(dest_cpu, nodemask) {
1181 if (!cpu_online(dest_cpu))
1183 if (!cpu_active(dest_cpu))
1185 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1191 /* Any allowed, online CPU? */
1192 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1193 if (!cpu_online(dest_cpu))
1195 if (!cpu_active(dest_cpu))
1202 /* No more Mr. Nice Guy. */
1203 cpuset_cpus_allowed_fallback(p);
1208 do_set_cpus_allowed(p, cpu_possible_mask);
1219 if (state != cpuset) {
1221 * Don't tell them about moving exiting tasks or
1222 * kernel threads (both mm NULL), since they never
1225 if (p->mm && printk_ratelimit()) {
1226 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1227 task_pid_nr(p), p->comm, cpu);
1235 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1238 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1240 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1243 * In order not to call set_task_cpu() on a blocking task we need
1244 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1247 * Since this is common to all placement strategies, this lives here.
1249 * [ this allows ->select_task() to simply return task_cpu(p) and
1250 * not worry about this generic constraint ]
1252 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1254 cpu = select_fallback_rq(task_cpu(p), p);
1259 static void update_avg(u64 *avg, u64 sample)
1261 s64 diff = sample - *avg;
1267 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1269 #ifdef CONFIG_SCHEDSTATS
1270 struct rq *rq = this_rq();
1273 int this_cpu = smp_processor_id();
1275 if (cpu == this_cpu) {
1276 schedstat_inc(rq, ttwu_local);
1277 schedstat_inc(p, se.statistics.nr_wakeups_local);
1279 struct sched_domain *sd;
1281 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1283 for_each_domain(this_cpu, sd) {
1284 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1285 schedstat_inc(sd, ttwu_wake_remote);
1292 if (wake_flags & WF_MIGRATED)
1293 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1295 #endif /* CONFIG_SMP */
1297 schedstat_inc(rq, ttwu_count);
1298 schedstat_inc(p, se.statistics.nr_wakeups);
1300 if (wake_flags & WF_SYNC)
1301 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1303 #endif /* CONFIG_SCHEDSTATS */
1306 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1308 activate_task(rq, p, en_flags);
1311 /* if a worker is waking up, notify workqueue */
1312 if (p->flags & PF_WQ_WORKER)
1313 wq_worker_waking_up(p, cpu_of(rq));
1317 * Mark the task runnable and perform wakeup-preemption.
1320 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1322 check_preempt_curr(rq, p, wake_flags);
1323 trace_sched_wakeup(p, true);
1325 p->state = TASK_RUNNING;
1327 if (p->sched_class->task_woken)
1328 p->sched_class->task_woken(rq, p);
1330 if (rq->idle_stamp) {
1331 u64 delta = rq->clock - rq->idle_stamp;
1332 u64 max = 2*sysctl_sched_migration_cost;
1337 update_avg(&rq->avg_idle, delta);
1344 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1347 if (p->sched_contributes_to_load)
1348 rq->nr_uninterruptible--;
1351 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1352 ttwu_do_wakeup(rq, p, wake_flags);
1356 * Called in case the task @p isn't fully descheduled from its runqueue,
1357 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1358 * since all we need to do is flip p->state to TASK_RUNNING, since
1359 * the task is still ->on_rq.
1361 static int ttwu_remote(struct task_struct *p, int wake_flags)
1366 rq = __task_rq_lock(p);
1368 ttwu_do_wakeup(rq, p, wake_flags);
1371 __task_rq_unlock(rq);
1377 static void sched_ttwu_pending(void)
1379 struct rq *rq = this_rq();
1380 struct llist_node *llist = llist_del_all(&rq->wake_list);
1381 struct task_struct *p;
1383 raw_spin_lock(&rq->lock);
1386 p = llist_entry(llist, struct task_struct, wake_entry);
1387 llist = llist_next(llist);
1388 ttwu_do_activate(rq, p, 0);
1391 raw_spin_unlock(&rq->lock);
1394 void scheduler_ipi(void)
1396 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()
1397 && !tick_nohz_full_cpu(smp_processor_id()))
1401 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1402 * traditionally all their work was done from the interrupt return
1403 * path. Now that we actually do some work, we need to make sure
1406 * Some archs already do call them, luckily irq_enter/exit nest
1409 * Arguably we should visit all archs and update all handlers,
1410 * however a fair share of IPIs are still resched only so this would
1411 * somewhat pessimize the simple resched case.
1414 tick_nohz_full_check();
1415 sched_ttwu_pending();
1418 * Check if someone kicked us for doing the nohz idle load balance.
1420 if (unlikely(got_nohz_idle_kick() && !need_resched())) {
1421 this_rq()->idle_balance = 1;
1422 raise_softirq_irqoff(SCHED_SOFTIRQ);
1427 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1429 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1430 smp_send_reschedule(cpu);
1433 bool cpus_share_cache(int this_cpu, int that_cpu)
1435 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1437 #endif /* CONFIG_SMP */
1439 static void ttwu_queue(struct task_struct *p, int cpu)
1441 struct rq *rq = cpu_rq(cpu);
1443 #if defined(CONFIG_SMP)
1444 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1445 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1446 ttwu_queue_remote(p, cpu);
1451 raw_spin_lock(&rq->lock);
1452 ttwu_do_activate(rq, p, 0);
1453 raw_spin_unlock(&rq->lock);
1457 * try_to_wake_up - wake up a thread
1458 * @p: the thread to be awakened
1459 * @state: the mask of task states that can be woken
1460 * @wake_flags: wake modifier flags (WF_*)
1462 * Put it on the run-queue if it's not already there. The "current"
1463 * thread is always on the run-queue (except when the actual
1464 * re-schedule is in progress), and as such you're allowed to do
1465 * the simpler "current->state = TASK_RUNNING" to mark yourself
1466 * runnable without the overhead of this.
1468 * Returns %true if @p was woken up, %false if it was already running
1469 * or @state didn't match @p's state.
1472 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1474 unsigned long flags;
1475 int cpu, success = 0;
1478 raw_spin_lock_irqsave(&p->pi_lock, flags);
1479 if (!(p->state & state))
1482 success = 1; /* we're going to change ->state */
1485 if (p->on_rq && ttwu_remote(p, wake_flags))
1490 * If the owning (remote) cpu is still in the middle of schedule() with
1491 * this task as prev, wait until its done referencing the task.
1496 * Pairs with the smp_wmb() in finish_lock_switch().
1500 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1501 p->state = TASK_WAKING;
1503 if (p->sched_class->task_waking)
1504 p->sched_class->task_waking(p);
1506 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1507 if (task_cpu(p) != cpu) {
1508 wake_flags |= WF_MIGRATED;
1509 set_task_cpu(p, cpu);
1511 #endif /* CONFIG_SMP */
1515 ttwu_stat(p, cpu, wake_flags);
1517 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1523 * try_to_wake_up_local - try to wake up a local task with rq lock held
1524 * @p: the thread to be awakened
1526 * Put @p on the run-queue if it's not already there. The caller must
1527 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1530 static void try_to_wake_up_local(struct task_struct *p)
1532 struct rq *rq = task_rq(p);
1534 if (WARN_ON_ONCE(rq != this_rq()) ||
1535 WARN_ON_ONCE(p == current))
1538 lockdep_assert_held(&rq->lock);
1540 if (!raw_spin_trylock(&p->pi_lock)) {
1541 raw_spin_unlock(&rq->lock);
1542 raw_spin_lock(&p->pi_lock);
1543 raw_spin_lock(&rq->lock);
1546 if (!(p->state & TASK_NORMAL))
1550 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1552 ttwu_do_wakeup(rq, p, 0);
1553 ttwu_stat(p, smp_processor_id(), 0);
1555 raw_spin_unlock(&p->pi_lock);
1559 * wake_up_process - Wake up a specific process
1560 * @p: The process to be woken up.
1562 * Attempt to wake up the nominated process and move it to the set of runnable
1563 * processes. Returns 1 if the process was woken up, 0 if it was already
1566 * It may be assumed that this function implies a write memory barrier before
1567 * changing the task state if and only if any tasks are woken up.
1569 int wake_up_process(struct task_struct *p)
1571 WARN_ON(task_is_stopped_or_traced(p));
1572 return try_to_wake_up(p, TASK_NORMAL, 0);
1574 EXPORT_SYMBOL(wake_up_process);
1576 int wake_up_state(struct task_struct *p, unsigned int state)
1578 return try_to_wake_up(p, state, 0);
1582 * Perform scheduler related setup for a newly forked process p.
1583 * p is forked by current.
1585 * __sched_fork() is basic setup used by init_idle() too:
1587 static void __sched_fork(struct task_struct *p)
1592 p->se.exec_start = 0;
1593 p->se.sum_exec_runtime = 0;
1594 p->se.prev_sum_exec_runtime = 0;
1595 p->se.nr_migrations = 0;
1597 INIT_LIST_HEAD(&p->se.group_node);
1600 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1601 * removed when useful for applications beyond shares distribution (e.g.
1604 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1605 p->se.avg.runnable_avg_period = 0;
1606 p->se.avg.runnable_avg_sum = 0;
1608 #ifdef CONFIG_SCHEDSTATS
1609 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1612 INIT_LIST_HEAD(&p->rt.run_list);
1614 #ifdef CONFIG_PREEMPT_NOTIFIERS
1615 INIT_HLIST_HEAD(&p->preempt_notifiers);
1618 #ifdef CONFIG_NUMA_BALANCING
1619 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1620 p->mm->numa_next_scan = jiffies;
1621 p->mm->numa_next_reset = jiffies;
1622 p->mm->numa_scan_seq = 0;
1625 p->node_stamp = 0ULL;
1626 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1627 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1628 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1629 p->numa_work.next = &p->numa_work;
1630 #endif /* CONFIG_NUMA_BALANCING */
1633 #ifdef CONFIG_NUMA_BALANCING
1634 #ifdef CONFIG_SCHED_DEBUG
1635 void set_numabalancing_state(bool enabled)
1638 sched_feat_set("NUMA");
1640 sched_feat_set("NO_NUMA");
1643 __read_mostly bool numabalancing_enabled;
1645 void set_numabalancing_state(bool enabled)
1647 numabalancing_enabled = enabled;
1649 #endif /* CONFIG_SCHED_DEBUG */
1650 #endif /* CONFIG_NUMA_BALANCING */
1653 * fork()/clone()-time setup:
1655 void sched_fork(struct task_struct *p)
1657 unsigned long flags;
1658 int cpu = get_cpu();
1662 * We mark the process as running here. This guarantees that
1663 * nobody will actually run it, and a signal or other external
1664 * event cannot wake it up and insert it on the runqueue either.
1666 p->state = TASK_RUNNING;
1669 * Make sure we do not leak PI boosting priority to the child.
1671 p->prio = current->normal_prio;
1674 * Revert to default priority/policy on fork if requested.
1676 if (unlikely(p->sched_reset_on_fork)) {
1677 if (task_has_rt_policy(p)) {
1678 p->policy = SCHED_NORMAL;
1679 p->static_prio = NICE_TO_PRIO(0);
1681 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1682 p->static_prio = NICE_TO_PRIO(0);
1684 p->prio = p->normal_prio = __normal_prio(p);
1688 * We don't need the reset flag anymore after the fork. It has
1689 * fulfilled its duty:
1691 p->sched_reset_on_fork = 0;
1694 if (!rt_prio(p->prio))
1695 p->sched_class = &fair_sched_class;
1697 if (p->sched_class->task_fork)
1698 p->sched_class->task_fork(p);
1701 * The child is not yet in the pid-hash so no cgroup attach races,
1702 * and the cgroup is pinned to this child due to cgroup_fork()
1703 * is ran before sched_fork().
1705 * Silence PROVE_RCU.
1707 raw_spin_lock_irqsave(&p->pi_lock, flags);
1708 set_task_cpu(p, cpu);
1709 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1711 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1712 if (likely(sched_info_on()))
1713 memset(&p->sched_info, 0, sizeof(p->sched_info));
1715 #if defined(CONFIG_SMP)
1718 #ifdef CONFIG_PREEMPT_COUNT
1719 /* Want to start with kernel preemption disabled. */
1720 task_thread_info(p)->preempt_count = 1;
1723 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1730 * wake_up_new_task - wake up a newly created task for the first time.
1732 * This function will do some initial scheduler statistics housekeeping
1733 * that must be done for every newly created context, then puts the task
1734 * on the runqueue and wakes it.
1736 void wake_up_new_task(struct task_struct *p)
1738 unsigned long flags;
1741 raw_spin_lock_irqsave(&p->pi_lock, flags);
1744 * Fork balancing, do it here and not earlier because:
1745 * - cpus_allowed can change in the fork path
1746 * - any previously selected cpu might disappear through hotplug
1748 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1751 rq = __task_rq_lock(p);
1752 activate_task(rq, p, 0);
1754 trace_sched_wakeup_new(p, true);
1755 check_preempt_curr(rq, p, WF_FORK);
1757 if (p->sched_class->task_woken)
1758 p->sched_class->task_woken(rq, p);
1760 task_rq_unlock(rq, p, &flags);
1763 #ifdef CONFIG_PREEMPT_NOTIFIERS
1766 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1767 * @notifier: notifier struct to register
1769 void preempt_notifier_register(struct preempt_notifier *notifier)
1771 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1773 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1776 * preempt_notifier_unregister - no longer interested in preemption notifications
1777 * @notifier: notifier struct to unregister
1779 * This is safe to call from within a preemption notifier.
1781 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1783 hlist_del(¬ifier->link);
1785 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1787 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1789 struct preempt_notifier *notifier;
1791 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1792 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1796 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1797 struct task_struct *next)
1799 struct preempt_notifier *notifier;
1801 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1802 notifier->ops->sched_out(notifier, next);
1805 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1807 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1812 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1813 struct task_struct *next)
1817 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1820 * prepare_task_switch - prepare to switch tasks
1821 * @rq: the runqueue preparing to switch
1822 * @prev: the current task that is being switched out
1823 * @next: the task we are going to switch to.
1825 * This is called with the rq lock held and interrupts off. It must
1826 * be paired with a subsequent finish_task_switch after the context
1829 * prepare_task_switch sets up locking and calls architecture specific
1833 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1834 struct task_struct *next)
1836 trace_sched_switch(prev, next);
1837 sched_info_switch(prev, next);
1838 perf_event_task_sched_out(prev, next);
1839 fire_sched_out_preempt_notifiers(prev, next);
1840 prepare_lock_switch(rq, next);
1841 prepare_arch_switch(next);
1845 * finish_task_switch - clean up after a task-switch
1846 * @rq: runqueue associated with task-switch
1847 * @prev: the thread we just switched away from.
1849 * finish_task_switch must be called after the context switch, paired
1850 * with a prepare_task_switch call before the context switch.
1851 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1852 * and do any other architecture-specific cleanup actions.
1854 * Note that we may have delayed dropping an mm in context_switch(). If
1855 * so, we finish that here outside of the runqueue lock. (Doing it
1856 * with the lock held can cause deadlocks; see schedule() for
1859 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1860 __releases(rq->lock)
1862 struct mm_struct *mm = rq->prev_mm;
1868 * A task struct has one reference for the use as "current".
1869 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1870 * schedule one last time. The schedule call will never return, and
1871 * the scheduled task must drop that reference.
1872 * The test for TASK_DEAD must occur while the runqueue locks are
1873 * still held, otherwise prev could be scheduled on another cpu, die
1874 * there before we look at prev->state, and then the reference would
1876 * Manfred Spraul <manfred@colorfullife.com>
1878 prev_state = prev->state;
1879 vtime_task_switch(prev);
1880 finish_arch_switch(prev);
1881 perf_event_task_sched_in(prev, current);
1882 finish_lock_switch(rq, prev);
1883 finish_arch_post_lock_switch();
1885 fire_sched_in_preempt_notifiers(current);
1888 if (unlikely(prev_state == TASK_DEAD)) {
1890 * Remove function-return probe instances associated with this
1891 * task and put them back on the free list.
1893 kprobe_flush_task(prev);
1894 put_task_struct(prev);
1897 tick_nohz_task_switch(current);
1902 /* assumes rq->lock is held */
1903 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1905 if (prev->sched_class->pre_schedule)
1906 prev->sched_class->pre_schedule(rq, prev);
1909 /* rq->lock is NOT held, but preemption is disabled */
1910 static inline void post_schedule(struct rq *rq)
1912 if (rq->post_schedule) {
1913 unsigned long flags;
1915 raw_spin_lock_irqsave(&rq->lock, flags);
1916 if (rq->curr->sched_class->post_schedule)
1917 rq->curr->sched_class->post_schedule(rq);
1918 raw_spin_unlock_irqrestore(&rq->lock, flags);
1920 rq->post_schedule = 0;
1926 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1930 static inline void post_schedule(struct rq *rq)
1937 * schedule_tail - first thing a freshly forked thread must call.
1938 * @prev: the thread we just switched away from.
1940 asmlinkage void schedule_tail(struct task_struct *prev)
1941 __releases(rq->lock)
1943 struct rq *rq = this_rq();
1945 finish_task_switch(rq, prev);
1948 * FIXME: do we need to worry about rq being invalidated by the
1953 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1954 /* In this case, finish_task_switch does not reenable preemption */
1957 if (current->set_child_tid)
1958 put_user(task_pid_vnr(current), current->set_child_tid);
1962 * context_switch - switch to the new MM and the new
1963 * thread's register state.
1966 context_switch(struct rq *rq, struct task_struct *prev,
1967 struct task_struct *next)
1969 struct mm_struct *mm, *oldmm;
1971 prepare_task_switch(rq, prev, next);
1974 oldmm = prev->active_mm;
1976 * For paravirt, this is coupled with an exit in switch_to to
1977 * combine the page table reload and the switch backend into
1980 arch_start_context_switch(prev);
1983 next->active_mm = oldmm;
1984 atomic_inc(&oldmm->mm_count);
1985 enter_lazy_tlb(oldmm, next);
1987 switch_mm(oldmm, mm, next);
1990 prev->active_mm = NULL;
1991 rq->prev_mm = oldmm;
1994 * Since the runqueue lock will be released by the next
1995 * task (which is an invalid locking op but in the case
1996 * of the scheduler it's an obvious special-case), so we
1997 * do an early lockdep release here:
1999 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2000 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2003 context_tracking_task_switch(prev, next);
2004 /* Here we just switch the register state and the stack. */
2005 switch_to(prev, next, prev);
2009 * this_rq must be evaluated again because prev may have moved
2010 * CPUs since it called schedule(), thus the 'rq' on its stack
2011 * frame will be invalid.
2013 finish_task_switch(this_rq(), prev);
2017 * nr_running and nr_context_switches:
2019 * externally visible scheduler statistics: current number of runnable
2020 * threads, total number of context switches performed since bootup.
2022 unsigned long nr_running(void)
2024 unsigned long i, sum = 0;
2026 for_each_online_cpu(i)
2027 sum += cpu_rq(i)->nr_running;
2032 unsigned long long nr_context_switches(void)
2035 unsigned long long sum = 0;
2037 for_each_possible_cpu(i)
2038 sum += cpu_rq(i)->nr_switches;
2043 unsigned long nr_iowait(void)
2045 unsigned long i, sum = 0;
2047 for_each_possible_cpu(i)
2048 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2053 unsigned long nr_iowait_cpu(int cpu)
2055 struct rq *this = cpu_rq(cpu);
2056 return atomic_read(&this->nr_iowait);
2059 unsigned long this_cpu_load(void)
2061 struct rq *this = this_rq();
2062 return this->cpu_load[0];
2067 * Global load-average calculations
2069 * We take a distributed and async approach to calculating the global load-avg
2070 * in order to minimize overhead.
2072 * The global load average is an exponentially decaying average of nr_running +
2073 * nr_uninterruptible.
2075 * Once every LOAD_FREQ:
2078 * for_each_possible_cpu(cpu)
2079 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2081 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2083 * Due to a number of reasons the above turns in the mess below:
2085 * - for_each_possible_cpu() is prohibitively expensive on machines with
2086 * serious number of cpus, therefore we need to take a distributed approach
2087 * to calculating nr_active.
2089 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2090 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2092 * So assuming nr_active := 0 when we start out -- true per definition, we
2093 * can simply take per-cpu deltas and fold those into a global accumulate
2094 * to obtain the same result. See calc_load_fold_active().
2096 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2097 * across the machine, we assume 10 ticks is sufficient time for every
2098 * cpu to have completed this task.
2100 * This places an upper-bound on the IRQ-off latency of the machine. Then
2101 * again, being late doesn't loose the delta, just wrecks the sample.
2103 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2104 * this would add another cross-cpu cacheline miss and atomic operation
2105 * to the wakeup path. Instead we increment on whatever cpu the task ran
2106 * when it went into uninterruptible state and decrement on whatever cpu
2107 * did the wakeup. This means that only the sum of nr_uninterruptible over
2108 * all cpus yields the correct result.
2110 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2113 /* Variables and functions for calc_load */
2114 static atomic_long_t calc_load_tasks;
2115 static unsigned long calc_load_update;
2116 unsigned long avenrun[3];
2117 EXPORT_SYMBOL(avenrun); /* should be removed */
2120 * get_avenrun - get the load average array
2121 * @loads: pointer to dest load array
2122 * @offset: offset to add
2123 * @shift: shift count to shift the result left
2125 * These values are estimates at best, so no need for locking.
2127 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2129 loads[0] = (avenrun[0] + offset) << shift;
2130 loads[1] = (avenrun[1] + offset) << shift;
2131 loads[2] = (avenrun[2] + offset) << shift;
2134 static long calc_load_fold_active(struct rq *this_rq)
2136 long nr_active, delta = 0;
2138 nr_active = this_rq->nr_running;
2139 nr_active += (long) this_rq->nr_uninterruptible;
2141 if (nr_active != this_rq->calc_load_active) {
2142 delta = nr_active - this_rq->calc_load_active;
2143 this_rq->calc_load_active = nr_active;
2150 * a1 = a0 * e + a * (1 - e)
2152 static unsigned long
2153 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2156 load += active * (FIXED_1 - exp);
2157 load += 1UL << (FSHIFT - 1);
2158 return load >> FSHIFT;
2161 #ifdef CONFIG_NO_HZ_COMMON
2163 * Handle NO_HZ for the global load-average.
2165 * Since the above described distributed algorithm to compute the global
2166 * load-average relies on per-cpu sampling from the tick, it is affected by
2169 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2170 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2171 * when we read the global state.
2173 * Obviously reality has to ruin such a delightfully simple scheme:
2175 * - When we go NO_HZ idle during the window, we can negate our sample
2176 * contribution, causing under-accounting.
2178 * We avoid this by keeping two idle-delta counters and flipping them
2179 * when the window starts, thus separating old and new NO_HZ load.
2181 * The only trick is the slight shift in index flip for read vs write.
2185 * |-|-----------|-|-----------|-|-----------|-|
2186 * r:0 0 1 1 0 0 1 1 0
2187 * w:0 1 1 0 0 1 1 0 0
2189 * This ensures we'll fold the old idle contribution in this window while
2190 * accumlating the new one.
2192 * - When we wake up from NO_HZ idle during the window, we push up our
2193 * contribution, since we effectively move our sample point to a known
2196 * This is solved by pushing the window forward, and thus skipping the
2197 * sample, for this cpu (effectively using the idle-delta for this cpu which
2198 * was in effect at the time the window opened). This also solves the issue
2199 * of having to deal with a cpu having been in NOHZ idle for multiple
2200 * LOAD_FREQ intervals.
2202 * When making the ILB scale, we should try to pull this in as well.
2204 static atomic_long_t calc_load_idle[2];
2205 static int calc_load_idx;
2207 static inline int calc_load_write_idx(void)
2209 int idx = calc_load_idx;
2212 * See calc_global_nohz(), if we observe the new index, we also
2213 * need to observe the new update time.
2218 * If the folding window started, make sure we start writing in the
2221 if (!time_before(jiffies, calc_load_update))
2227 static inline int calc_load_read_idx(void)
2229 return calc_load_idx & 1;
2232 void calc_load_enter_idle(void)
2234 struct rq *this_rq = this_rq();
2238 * We're going into NOHZ mode, if there's any pending delta, fold it
2239 * into the pending idle delta.
2241 delta = calc_load_fold_active(this_rq);
2243 int idx = calc_load_write_idx();
2244 atomic_long_add(delta, &calc_load_idle[idx]);
2248 void calc_load_exit_idle(void)
2250 struct rq *this_rq = this_rq();
2253 * If we're still before the sample window, we're done.
2255 if (time_before(jiffies, this_rq->calc_load_update))
2259 * We woke inside or after the sample window, this means we're already
2260 * accounted through the nohz accounting, so skip the entire deal and
2261 * sync up for the next window.
2263 this_rq->calc_load_update = calc_load_update;
2264 if (time_before(jiffies, this_rq->calc_load_update + 10))
2265 this_rq->calc_load_update += LOAD_FREQ;
2268 static long calc_load_fold_idle(void)
2270 int idx = calc_load_read_idx();
2273 if (atomic_long_read(&calc_load_idle[idx]))
2274 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2280 * fixed_power_int - compute: x^n, in O(log n) time
2282 * @x: base of the power
2283 * @frac_bits: fractional bits of @x
2284 * @n: power to raise @x to.
2286 * By exploiting the relation between the definition of the natural power
2287 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2288 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2289 * (where: n_i \elem {0, 1}, the binary vector representing n),
2290 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2291 * of course trivially computable in O(log_2 n), the length of our binary
2294 static unsigned long
2295 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2297 unsigned long result = 1UL << frac_bits;
2302 result += 1UL << (frac_bits - 1);
2303 result >>= frac_bits;
2309 x += 1UL << (frac_bits - 1);
2317 * a1 = a0 * e + a * (1 - e)
2319 * a2 = a1 * e + a * (1 - e)
2320 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2321 * = a0 * e^2 + a * (1 - e) * (1 + e)
2323 * a3 = a2 * e + a * (1 - e)
2324 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2325 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2329 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2330 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2331 * = a0 * e^n + a * (1 - e^n)
2333 * [1] application of the geometric series:
2336 * S_n := \Sum x^i = -------------
2339 static unsigned long
2340 calc_load_n(unsigned long load, unsigned long exp,
2341 unsigned long active, unsigned int n)
2344 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2348 * NO_HZ can leave us missing all per-cpu ticks calling
2349 * calc_load_account_active(), but since an idle CPU folds its delta into
2350 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2351 * in the pending idle delta if our idle period crossed a load cycle boundary.
2353 * Once we've updated the global active value, we need to apply the exponential
2354 * weights adjusted to the number of cycles missed.
2356 static void calc_global_nohz(void)
2358 long delta, active, n;
2360 if (!time_before(jiffies, calc_load_update + 10)) {
2362 * Catch-up, fold however many we are behind still
2364 delta = jiffies - calc_load_update - 10;
2365 n = 1 + (delta / LOAD_FREQ);
2367 active = atomic_long_read(&calc_load_tasks);
2368 active = active > 0 ? active * FIXED_1 : 0;
2370 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2371 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2372 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2374 calc_load_update += n * LOAD_FREQ;
2378 * Flip the idle index...
2380 * Make sure we first write the new time then flip the index, so that
2381 * calc_load_write_idx() will see the new time when it reads the new
2382 * index, this avoids a double flip messing things up.
2387 #else /* !CONFIG_NO_HZ_COMMON */
2389 static inline long calc_load_fold_idle(void) { return 0; }
2390 static inline void calc_global_nohz(void) { }
2392 #endif /* CONFIG_NO_HZ_COMMON */
2395 * calc_load - update the avenrun load estimates 10 ticks after the
2396 * CPUs have updated calc_load_tasks.
2398 void calc_global_load(unsigned long ticks)
2402 if (time_before(jiffies, calc_load_update + 10))
2406 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2408 delta = calc_load_fold_idle();
2410 atomic_long_add(delta, &calc_load_tasks);
2412 active = atomic_long_read(&calc_load_tasks);
2413 active = active > 0 ? active * FIXED_1 : 0;
2415 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2416 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2417 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2419 calc_load_update += LOAD_FREQ;
2422 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2428 * Called from update_cpu_load() to periodically update this CPU's
2431 static void calc_load_account_active(struct rq *this_rq)
2435 if (time_before(jiffies, this_rq->calc_load_update))
2438 delta = calc_load_fold_active(this_rq);
2440 atomic_long_add(delta, &calc_load_tasks);
2442 this_rq->calc_load_update += LOAD_FREQ;
2446 * End of global load-average stuff
2450 * The exact cpuload at various idx values, calculated at every tick would be
2451 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2453 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2454 * on nth tick when cpu may be busy, then we have:
2455 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2456 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2458 * decay_load_missed() below does efficient calculation of
2459 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2460 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2462 * The calculation is approximated on a 128 point scale.
2463 * degrade_zero_ticks is the number of ticks after which load at any
2464 * particular idx is approximated to be zero.
2465 * degrade_factor is a precomputed table, a row for each load idx.
2466 * Each column corresponds to degradation factor for a power of two ticks,
2467 * based on 128 point scale.
2469 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2470 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2472 * With this power of 2 load factors, we can degrade the load n times
2473 * by looking at 1 bits in n and doing as many mult/shift instead of
2474 * n mult/shifts needed by the exact degradation.
2476 #define DEGRADE_SHIFT 7
2477 static const unsigned char
2478 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2479 static const unsigned char
2480 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2481 {0, 0, 0, 0, 0, 0, 0, 0},
2482 {64, 32, 8, 0, 0, 0, 0, 0},
2483 {96, 72, 40, 12, 1, 0, 0},
2484 {112, 98, 75, 43, 15, 1, 0},
2485 {120, 112, 98, 76, 45, 16, 2} };
2488 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2489 * would be when CPU is idle and so we just decay the old load without
2490 * adding any new load.
2492 static unsigned long
2493 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2497 if (!missed_updates)
2500 if (missed_updates >= degrade_zero_ticks[idx])
2504 return load >> missed_updates;
2506 while (missed_updates) {
2507 if (missed_updates % 2)
2508 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2510 missed_updates >>= 1;
2517 * Update rq->cpu_load[] statistics. This function is usually called every
2518 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2519 * every tick. We fix it up based on jiffies.
2521 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2522 unsigned long pending_updates)
2526 this_rq->nr_load_updates++;
2528 /* Update our load: */
2529 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2530 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2531 unsigned long old_load, new_load;
2533 /* scale is effectively 1 << i now, and >> i divides by scale */
2535 old_load = this_rq->cpu_load[i];
2536 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2537 new_load = this_load;
2539 * Round up the averaging division if load is increasing. This
2540 * prevents us from getting stuck on 9 if the load is 10, for
2543 if (new_load > old_load)
2544 new_load += scale - 1;
2546 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2549 sched_avg_update(this_rq);
2552 #ifdef CONFIG_NO_HZ_COMMON
2554 * There is no sane way to deal with nohz on smp when using jiffies because the
2555 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2556 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2558 * Therefore we cannot use the delta approach from the regular tick since that
2559 * would seriously skew the load calculation. However we'll make do for those
2560 * updates happening while idle (nohz_idle_balance) or coming out of idle
2561 * (tick_nohz_idle_exit).
2563 * This means we might still be one tick off for nohz periods.
2567 * Called from nohz_idle_balance() to update the load ratings before doing the
2570 void update_idle_cpu_load(struct rq *this_rq)
2572 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2573 unsigned long load = this_rq->load.weight;
2574 unsigned long pending_updates;
2577 * bail if there's load or we're actually up-to-date.
2579 if (load || curr_jiffies == this_rq->last_load_update_tick)
2582 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2583 this_rq->last_load_update_tick = curr_jiffies;
2585 __update_cpu_load(this_rq, load, pending_updates);
2589 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2591 void update_cpu_load_nohz(void)
2593 struct rq *this_rq = this_rq();
2594 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2595 unsigned long pending_updates;
2597 if (curr_jiffies == this_rq->last_load_update_tick)
2600 raw_spin_lock(&this_rq->lock);
2601 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2602 if (pending_updates) {
2603 this_rq->last_load_update_tick = curr_jiffies;
2605 * We were idle, this means load 0, the current load might be
2606 * !0 due to remote wakeups and the sort.
2608 __update_cpu_load(this_rq, 0, pending_updates);
2610 raw_spin_unlock(&this_rq->lock);
2612 #endif /* CONFIG_NO_HZ_COMMON */
2615 * Called from scheduler_tick()
2617 static void update_cpu_load_active(struct rq *this_rq)
2620 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2622 this_rq->last_load_update_tick = jiffies;
2623 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2625 calc_load_account_active(this_rq);
2631 * sched_exec - execve() is a valuable balancing opportunity, because at
2632 * this point the task has the smallest effective memory and cache footprint.
2634 void sched_exec(void)
2636 struct task_struct *p = current;
2637 unsigned long flags;
2640 raw_spin_lock_irqsave(&p->pi_lock, flags);
2641 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2642 if (dest_cpu == smp_processor_id())
2645 if (likely(cpu_active(dest_cpu))) {
2646 struct migration_arg arg = { p, dest_cpu };
2648 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2649 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2653 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2658 DEFINE_PER_CPU(struct kernel_stat, kstat);
2659 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2661 EXPORT_PER_CPU_SYMBOL(kstat);
2662 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2665 * Return any ns on the sched_clock that have not yet been accounted in
2666 * @p in case that task is currently running.
2668 * Called with task_rq_lock() held on @rq.
2670 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2674 if (task_current(rq, p)) {
2675 update_rq_clock(rq);
2676 ns = rq->clock_task - p->se.exec_start;
2684 unsigned long long task_delta_exec(struct task_struct *p)
2686 unsigned long flags;
2690 rq = task_rq_lock(p, &flags);
2691 ns = do_task_delta_exec(p, rq);
2692 task_rq_unlock(rq, p, &flags);
2698 * Return accounted runtime for the task.
2699 * In case the task is currently running, return the runtime plus current's
2700 * pending runtime that have not been accounted yet.
2702 unsigned long long task_sched_runtime(struct task_struct *p)
2704 unsigned long flags;
2708 rq = task_rq_lock(p, &flags);
2709 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2710 task_rq_unlock(rq, p, &flags);
2716 * This function gets called by the timer code, with HZ frequency.
2717 * We call it with interrupts disabled.
2719 void scheduler_tick(void)
2721 int cpu = smp_processor_id();
2722 struct rq *rq = cpu_rq(cpu);
2723 struct task_struct *curr = rq->curr;
2727 raw_spin_lock(&rq->lock);
2728 update_rq_clock(rq);
2729 update_cpu_load_active(rq);
2730 curr->sched_class->task_tick(rq, curr, 0);
2731 raw_spin_unlock(&rq->lock);
2733 perf_event_task_tick();
2736 rq->idle_balance = idle_cpu(cpu);
2737 trigger_load_balance(rq, cpu);
2739 rq_last_tick_reset(rq);
2742 #ifdef CONFIG_NO_HZ_FULL
2744 * scheduler_tick_max_deferment
2746 * Keep at least one tick per second when a single
2747 * active task is running because the scheduler doesn't
2748 * yet completely support full dynticks environment.
2750 * This makes sure that uptime, CFS vruntime, load
2751 * balancing, etc... continue to move forward, even
2752 * with a very low granularity.
2754 u64 scheduler_tick_max_deferment(void)
2756 struct rq *rq = this_rq();
2757 unsigned long next, now = ACCESS_ONCE(jiffies);
2759 next = rq->last_sched_tick + HZ;
2761 if (time_before_eq(next, now))
2764 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2768 notrace unsigned long get_parent_ip(unsigned long addr)
2770 if (in_lock_functions(addr)) {
2771 addr = CALLER_ADDR2;
2772 if (in_lock_functions(addr))
2773 addr = CALLER_ADDR3;
2778 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2779 defined(CONFIG_PREEMPT_TRACER))
2781 void __kprobes add_preempt_count(int val)
2783 #ifdef CONFIG_DEBUG_PREEMPT
2787 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2790 preempt_count() += val;
2791 #ifdef CONFIG_DEBUG_PREEMPT
2793 * Spinlock count overflowing soon?
2795 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2798 if (preempt_count() == val)
2799 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2801 EXPORT_SYMBOL(add_preempt_count);
2803 void __kprobes sub_preempt_count(int val)
2805 #ifdef CONFIG_DEBUG_PREEMPT
2809 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2812 * Is the spinlock portion underflowing?
2814 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2815 !(preempt_count() & PREEMPT_MASK)))
2819 if (preempt_count() == val)
2820 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2821 preempt_count() -= val;
2823 EXPORT_SYMBOL(sub_preempt_count);
2828 * Print scheduling while atomic bug:
2830 static noinline void __schedule_bug(struct task_struct *prev)
2832 if (oops_in_progress)
2835 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2836 prev->comm, prev->pid, preempt_count());
2838 debug_show_held_locks(prev);
2840 if (irqs_disabled())
2841 print_irqtrace_events(prev);
2843 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2847 * Various schedule()-time debugging checks and statistics:
2849 static inline void schedule_debug(struct task_struct *prev)
2852 * Test if we are atomic. Since do_exit() needs to call into
2853 * schedule() atomically, we ignore that path for now.
2854 * Otherwise, whine if we are scheduling when we should not be.
2856 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2857 __schedule_bug(prev);
2860 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2862 schedstat_inc(this_rq(), sched_count);
2865 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2867 if (prev->on_rq || rq->skip_clock_update < 0)
2868 update_rq_clock(rq);
2869 prev->sched_class->put_prev_task(rq, prev);
2873 * Pick up the highest-prio task:
2875 static inline struct task_struct *
2876 pick_next_task(struct rq *rq)
2878 const struct sched_class *class;
2879 struct task_struct *p;
2882 * Optimization: we know that if all tasks are in
2883 * the fair class we can call that function directly:
2885 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2886 p = fair_sched_class.pick_next_task(rq);
2891 for_each_class(class) {
2892 p = class->pick_next_task(rq);
2897 BUG(); /* the idle class will always have a runnable task */
2901 * __schedule() is the main scheduler function.
2903 * The main means of driving the scheduler and thus entering this function are:
2905 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2907 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2908 * paths. For example, see arch/x86/entry_64.S.
2910 * To drive preemption between tasks, the scheduler sets the flag in timer
2911 * interrupt handler scheduler_tick().
2913 * 3. Wakeups don't really cause entry into schedule(). They add a
2914 * task to the run-queue and that's it.
2916 * Now, if the new task added to the run-queue preempts the current
2917 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2918 * called on the nearest possible occasion:
2920 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2922 * - in syscall or exception context, at the next outmost
2923 * preempt_enable(). (this might be as soon as the wake_up()'s
2926 * - in IRQ context, return from interrupt-handler to
2927 * preemptible context
2929 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2932 * - cond_resched() call
2933 * - explicit schedule() call
2934 * - return from syscall or exception to user-space
2935 * - return from interrupt-handler to user-space
2937 static void __sched __schedule(void)
2939 struct task_struct *prev, *next;
2940 unsigned long *switch_count;
2946 cpu = smp_processor_id();
2948 rcu_note_context_switch(cpu);
2951 schedule_debug(prev);
2953 if (sched_feat(HRTICK))
2956 raw_spin_lock_irq(&rq->lock);
2958 switch_count = &prev->nivcsw;
2959 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2960 if (unlikely(signal_pending_state(prev->state, prev))) {
2961 prev->state = TASK_RUNNING;
2963 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2967 * If a worker went to sleep, notify and ask workqueue
2968 * whether it wants to wake up a task to maintain
2971 if (prev->flags & PF_WQ_WORKER) {
2972 struct task_struct *to_wakeup;
2974 to_wakeup = wq_worker_sleeping(prev, cpu);
2976 try_to_wake_up_local(to_wakeup);
2979 switch_count = &prev->nvcsw;
2982 pre_schedule(rq, prev);
2984 if (unlikely(!rq->nr_running))
2985 idle_balance(cpu, rq);
2987 put_prev_task(rq, prev);
2988 next = pick_next_task(rq);
2989 clear_tsk_need_resched(prev);
2990 rq->skip_clock_update = 0;
2992 if (likely(prev != next)) {
2997 context_switch(rq, prev, next); /* unlocks the rq */
2999 * The context switch have flipped the stack from under us
3000 * and restored the local variables which were saved when
3001 * this task called schedule() in the past. prev == current
3002 * is still correct, but it can be moved to another cpu/rq.
3004 cpu = smp_processor_id();
3007 raw_spin_unlock_irq(&rq->lock);
3011 sched_preempt_enable_no_resched();
3016 static inline void sched_submit_work(struct task_struct *tsk)
3018 if (!tsk->state || tsk_is_pi_blocked(tsk))
3021 * If we are going to sleep and we have plugged IO queued,
3022 * make sure to submit it to avoid deadlocks.
3024 if (blk_needs_flush_plug(tsk))
3025 blk_schedule_flush_plug(tsk);
3028 asmlinkage void __sched schedule(void)
3030 struct task_struct *tsk = current;
3032 sched_submit_work(tsk);
3035 EXPORT_SYMBOL(schedule);
3037 #ifdef CONFIG_CONTEXT_TRACKING
3038 asmlinkage void __sched schedule_user(void)
3041 * If we come here after a random call to set_need_resched(),
3042 * or we have been woken up remotely but the IPI has not yet arrived,
3043 * we haven't yet exited the RCU idle mode. Do it here manually until
3044 * we find a better solution.
3053 * schedule_preempt_disabled - called with preemption disabled
3055 * Returns with preemption disabled. Note: preempt_count must be 1
3057 void __sched schedule_preempt_disabled(void)
3059 sched_preempt_enable_no_resched();
3064 #ifdef CONFIG_PREEMPT
3066 * this is the entry point to schedule() from in-kernel preemption
3067 * off of preempt_enable. Kernel preemptions off return from interrupt
3068 * occur there and call schedule directly.
3070 asmlinkage void __sched notrace preempt_schedule(void)
3072 struct thread_info *ti = current_thread_info();
3075 * If there is a non-zero preempt_count or interrupts are disabled,
3076 * we do not want to preempt the current task. Just return..
3078 if (likely(ti->preempt_count || irqs_disabled()))
3082 add_preempt_count_notrace(PREEMPT_ACTIVE);
3084 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3087 * Check again in case we missed a preemption opportunity
3088 * between schedule and now.
3091 } while (need_resched());
3093 EXPORT_SYMBOL(preempt_schedule);
3096 * this is the entry point to schedule() from kernel preemption
3097 * off of irq context.
3098 * Note, that this is called and return with irqs disabled. This will
3099 * protect us against recursive calling from irq.
3101 asmlinkage void __sched preempt_schedule_irq(void)
3103 struct thread_info *ti = current_thread_info();
3104 enum ctx_state prev_state;
3106 /* Catch callers which need to be fixed */
3107 BUG_ON(ti->preempt_count || !irqs_disabled());
3109 prev_state = exception_enter();
3112 add_preempt_count(PREEMPT_ACTIVE);
3115 local_irq_disable();
3116 sub_preempt_count(PREEMPT_ACTIVE);
3119 * Check again in case we missed a preemption opportunity
3120 * between schedule and now.
3123 } while (need_resched());
3125 exception_exit(prev_state);
3128 #endif /* CONFIG_PREEMPT */
3130 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3133 return try_to_wake_up(curr->private, mode, wake_flags);
3135 EXPORT_SYMBOL(default_wake_function);
3138 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3139 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3140 * number) then we wake all the non-exclusive tasks and one exclusive task.
3142 * There are circumstances in which we can try to wake a task which has already
3143 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3144 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3146 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3147 int nr_exclusive, int wake_flags, void *key)
3149 wait_queue_t *curr, *next;
3151 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3152 unsigned flags = curr->flags;
3154 if (curr->func(curr, mode, wake_flags, key) &&
3155 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3161 * __wake_up - wake up threads blocked on a waitqueue.
3163 * @mode: which threads
3164 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3165 * @key: is directly passed to the wakeup function
3167 * It may be assumed that this function implies a write memory barrier before
3168 * changing the task state if and only if any tasks are woken up.
3170 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3171 int nr_exclusive, void *key)
3173 unsigned long flags;
3175 spin_lock_irqsave(&q->lock, flags);
3176 __wake_up_common(q, mode, nr_exclusive, 0, key);
3177 spin_unlock_irqrestore(&q->lock, flags);
3179 EXPORT_SYMBOL(__wake_up);
3182 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3184 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3186 __wake_up_common(q, mode, nr, 0, NULL);
3188 EXPORT_SYMBOL_GPL(__wake_up_locked);
3190 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3192 __wake_up_common(q, mode, 1, 0, key);
3194 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3197 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3199 * @mode: which threads
3200 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3201 * @key: opaque value to be passed to wakeup targets
3203 * The sync wakeup differs that the waker knows that it will schedule
3204 * away soon, so while the target thread will be woken up, it will not
3205 * be migrated to another CPU - ie. the two threads are 'synchronized'
3206 * with each other. This can prevent needless bouncing between CPUs.
3208 * On UP it can prevent extra preemption.
3210 * It may be assumed that this function implies a write memory barrier before
3211 * changing the task state if and only if any tasks are woken up.
3213 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3214 int nr_exclusive, void *key)
3216 unsigned long flags;
3217 int wake_flags = WF_SYNC;
3222 if (unlikely(!nr_exclusive))
3225 spin_lock_irqsave(&q->lock, flags);
3226 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3227 spin_unlock_irqrestore(&q->lock, flags);
3229 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3232 * __wake_up_sync - see __wake_up_sync_key()
3234 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3236 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3238 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3241 * complete: - signals a single thread waiting on this completion
3242 * @x: holds the state of this particular completion
3244 * This will wake up a single thread waiting on this completion. Threads will be
3245 * awakened in the same order in which they were queued.
3247 * See also complete_all(), wait_for_completion() and related routines.
3249 * It may be assumed that this function implies a write memory barrier before
3250 * changing the task state if and only if any tasks are woken up.
3252 void complete(struct completion *x)
3254 unsigned long flags;
3256 spin_lock_irqsave(&x->wait.lock, flags);
3258 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3259 spin_unlock_irqrestore(&x->wait.lock, flags);
3261 EXPORT_SYMBOL(complete);
3264 * complete_all: - signals all threads waiting on this completion
3265 * @x: holds the state of this particular completion
3267 * This will wake up all threads waiting on this particular completion event.
3269 * It may be assumed that this function implies a write memory barrier before
3270 * changing the task state if and only if any tasks are woken up.
3272 void complete_all(struct completion *x)
3274 unsigned long flags;
3276 spin_lock_irqsave(&x->wait.lock, flags);
3277 x->done += UINT_MAX/2;
3278 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3279 spin_unlock_irqrestore(&x->wait.lock, flags);
3281 EXPORT_SYMBOL(complete_all);
3283 static inline long __sched
3284 do_wait_for_common(struct completion *x,
3285 long (*action)(long), long timeout, int state)
3288 DECLARE_WAITQUEUE(wait, current);
3290 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3292 if (signal_pending_state(state, current)) {
3293 timeout = -ERESTARTSYS;
3296 __set_current_state(state);
3297 spin_unlock_irq(&x->wait.lock);
3298 timeout = action(timeout);
3299 spin_lock_irq(&x->wait.lock);
3300 } while (!x->done && timeout);
3301 __remove_wait_queue(&x->wait, &wait);
3306 return timeout ?: 1;
3309 static inline long __sched
3310 __wait_for_common(struct completion *x,
3311 long (*action)(long), long timeout, int state)
3315 spin_lock_irq(&x->wait.lock);
3316 timeout = do_wait_for_common(x, action, timeout, state);
3317 spin_unlock_irq(&x->wait.lock);
3322 wait_for_common(struct completion *x, long timeout, int state)
3324 return __wait_for_common(x, schedule_timeout, timeout, state);
3328 wait_for_common_io(struct completion *x, long timeout, int state)
3330 return __wait_for_common(x, io_schedule_timeout, timeout, state);
3334 * wait_for_completion: - waits for completion of a task
3335 * @x: holds the state of this particular completion
3337 * This waits to be signaled for completion of a specific task. It is NOT
3338 * interruptible and there is no timeout.
3340 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3341 * and interrupt capability. Also see complete().
3343 void __sched wait_for_completion(struct completion *x)
3345 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3347 EXPORT_SYMBOL(wait_for_completion);
3350 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3351 * @x: holds the state of this particular completion
3352 * @timeout: timeout value in jiffies
3354 * This waits for either a completion of a specific task to be signaled or for a
3355 * specified timeout to expire. The timeout is in jiffies. It is not
3358 * The return value is 0 if timed out, and positive (at least 1, or number of
3359 * jiffies left till timeout) if completed.
3361 unsigned long __sched
3362 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3364 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3366 EXPORT_SYMBOL(wait_for_completion_timeout);
3369 * wait_for_completion_io: - waits for completion of a task
3370 * @x: holds the state of this particular completion
3372 * This waits to be signaled for completion of a specific task. It is NOT
3373 * interruptible and there is no timeout. The caller is accounted as waiting
3376 void __sched wait_for_completion_io(struct completion *x)
3378 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3380 EXPORT_SYMBOL(wait_for_completion_io);
3383 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3384 * @x: holds the state of this particular completion
3385 * @timeout: timeout value in jiffies
3387 * This waits for either a completion of a specific task to be signaled or for a
3388 * specified timeout to expire. The timeout is in jiffies. It is not
3389 * interruptible. The caller is accounted as waiting for IO.
3391 * The return value is 0 if timed out, and positive (at least 1, or number of
3392 * jiffies left till timeout) if completed.
3394 unsigned long __sched
3395 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
3397 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
3399 EXPORT_SYMBOL(wait_for_completion_io_timeout);
3402 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3403 * @x: holds the state of this particular completion
3405 * This waits for completion of a specific task to be signaled. It is
3408 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3410 int __sched wait_for_completion_interruptible(struct completion *x)
3412 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3413 if (t == -ERESTARTSYS)
3417 EXPORT_SYMBOL(wait_for_completion_interruptible);
3420 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3421 * @x: holds the state of this particular completion
3422 * @timeout: timeout value in jiffies
3424 * This waits for either a completion of a specific task to be signaled or for a
3425 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3427 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3428 * positive (at least 1, or number of jiffies left till timeout) if completed.
3431 wait_for_completion_interruptible_timeout(struct completion *x,
3432 unsigned long timeout)
3434 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3436 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3439 * wait_for_completion_killable: - waits for completion of a task (killable)
3440 * @x: holds the state of this particular completion
3442 * This waits to be signaled for completion of a specific task. It can be
3443 * interrupted by a kill signal.
3445 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3447 int __sched wait_for_completion_killable(struct completion *x)
3449 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3450 if (t == -ERESTARTSYS)
3454 EXPORT_SYMBOL(wait_for_completion_killable);
3457 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3458 * @x: holds the state of this particular completion
3459 * @timeout: timeout value in jiffies
3461 * This waits for either a completion of a specific task to be
3462 * signaled or for a specified timeout to expire. It can be
3463 * interrupted by a kill signal. The timeout is in jiffies.
3465 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3466 * positive (at least 1, or number of jiffies left till timeout) if completed.
3469 wait_for_completion_killable_timeout(struct completion *x,
3470 unsigned long timeout)
3472 return wait_for_common(x, timeout, TASK_KILLABLE);
3474 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3477 * try_wait_for_completion - try to decrement a completion without blocking
3478 * @x: completion structure
3480 * Returns: 0 if a decrement cannot be done without blocking
3481 * 1 if a decrement succeeded.
3483 * If a completion is being used as a counting completion,
3484 * attempt to decrement the counter without blocking. This
3485 * enables us to avoid waiting if the resource the completion
3486 * is protecting is not available.
3488 bool try_wait_for_completion(struct completion *x)
3490 unsigned long flags;
3493 spin_lock_irqsave(&x->wait.lock, flags);
3498 spin_unlock_irqrestore(&x->wait.lock, flags);
3501 EXPORT_SYMBOL(try_wait_for_completion);
3504 * completion_done - Test to see if a completion has any waiters
3505 * @x: completion structure
3507 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3508 * 1 if there are no waiters.
3511 bool completion_done(struct completion *x)
3513 unsigned long flags;
3516 spin_lock_irqsave(&x->wait.lock, flags);
3519 spin_unlock_irqrestore(&x->wait.lock, flags);
3522 EXPORT_SYMBOL(completion_done);
3525 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3527 unsigned long flags;
3530 init_waitqueue_entry(&wait, current);
3532 __set_current_state(state);
3534 spin_lock_irqsave(&q->lock, flags);
3535 __add_wait_queue(q, &wait);
3536 spin_unlock(&q->lock);
3537 timeout = schedule_timeout(timeout);
3538 spin_lock_irq(&q->lock);
3539 __remove_wait_queue(q, &wait);
3540 spin_unlock_irqrestore(&q->lock, flags);
3545 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3547 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3549 EXPORT_SYMBOL(interruptible_sleep_on);
3552 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3554 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3556 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3558 void __sched sleep_on(wait_queue_head_t *q)
3560 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3562 EXPORT_SYMBOL(sleep_on);
3564 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3566 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3568 EXPORT_SYMBOL(sleep_on_timeout);
3570 #ifdef CONFIG_RT_MUTEXES
3573 * rt_mutex_setprio - set the current priority of a task
3575 * @prio: prio value (kernel-internal form)
3577 * This function changes the 'effective' priority of a task. It does
3578 * not touch ->normal_prio like __setscheduler().
3580 * Used by the rt_mutex code to implement priority inheritance logic.
3582 void rt_mutex_setprio(struct task_struct *p, int prio)
3584 int oldprio, on_rq, running;
3586 const struct sched_class *prev_class;
3588 BUG_ON(prio < 0 || prio > MAX_PRIO);
3590 rq = __task_rq_lock(p);
3593 * Idle task boosting is a nono in general. There is one
3594 * exception, when PREEMPT_RT and NOHZ is active:
3596 * The idle task calls get_next_timer_interrupt() and holds
3597 * the timer wheel base->lock on the CPU and another CPU wants
3598 * to access the timer (probably to cancel it). We can safely
3599 * ignore the boosting request, as the idle CPU runs this code
3600 * with interrupts disabled and will complete the lock
3601 * protected section without being interrupted. So there is no
3602 * real need to boost.
3604 if (unlikely(p == rq->idle)) {
3605 WARN_ON(p != rq->curr);
3606 WARN_ON(p->pi_blocked_on);
3610 trace_sched_pi_setprio(p, prio);
3612 prev_class = p->sched_class;
3614 running = task_current(rq, p);
3616 dequeue_task(rq, p, 0);
3618 p->sched_class->put_prev_task(rq, p);
3621 p->sched_class = &rt_sched_class;
3623 p->sched_class = &fair_sched_class;
3628 p->sched_class->set_curr_task(rq);
3630 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3632 check_class_changed(rq, p, prev_class, oldprio);
3634 __task_rq_unlock(rq);
3637 void set_user_nice(struct task_struct *p, long nice)
3639 int old_prio, delta, on_rq;
3640 unsigned long flags;
3643 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3646 * We have to be careful, if called from sys_setpriority(),
3647 * the task might be in the middle of scheduling on another CPU.
3649 rq = task_rq_lock(p, &flags);
3651 * The RT priorities are set via sched_setscheduler(), but we still
3652 * allow the 'normal' nice value to be set - but as expected
3653 * it wont have any effect on scheduling until the task is
3654 * SCHED_FIFO/SCHED_RR:
3656 if (task_has_rt_policy(p)) {
3657 p->static_prio = NICE_TO_PRIO(nice);
3662 dequeue_task(rq, p, 0);
3664 p->static_prio = NICE_TO_PRIO(nice);
3667 p->prio = effective_prio(p);
3668 delta = p->prio - old_prio;
3671 enqueue_task(rq, p, 0);
3673 * If the task increased its priority or is running and
3674 * lowered its priority, then reschedule its CPU:
3676 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3677 resched_task(rq->curr);
3680 task_rq_unlock(rq, p, &flags);
3682 EXPORT_SYMBOL(set_user_nice);
3685 * can_nice - check if a task can reduce its nice value
3689 int can_nice(const struct task_struct *p, const int nice)
3691 /* convert nice value [19,-20] to rlimit style value [1,40] */
3692 int nice_rlim = 20 - nice;
3694 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3695 capable(CAP_SYS_NICE));
3698 #ifdef __ARCH_WANT_SYS_NICE
3701 * sys_nice - change the priority of the current process.
3702 * @increment: priority increment
3704 * sys_setpriority is a more generic, but much slower function that
3705 * does similar things.
3707 SYSCALL_DEFINE1(nice, int, increment)
3712 * Setpriority might change our priority at the same moment.
3713 * We don't have to worry. Conceptually one call occurs first
3714 * and we have a single winner.
3716 if (increment < -40)
3721 nice = TASK_NICE(current) + increment;
3727 if (increment < 0 && !can_nice(current, nice))
3730 retval = security_task_setnice(current, nice);
3734 set_user_nice(current, nice);
3741 * task_prio - return the priority value of a given task.
3742 * @p: the task in question.
3744 * This is the priority value as seen by users in /proc.
3745 * RT tasks are offset by -200. Normal tasks are centered
3746 * around 0, value goes from -16 to +15.
3748 int task_prio(const struct task_struct *p)
3750 return p->prio - MAX_RT_PRIO;
3754 * task_nice - return the nice value of a given task.
3755 * @p: the task in question.
3757 int task_nice(const struct task_struct *p)
3759 return TASK_NICE(p);
3761 EXPORT_SYMBOL(task_nice);
3764 * idle_cpu - is a given cpu idle currently?
3765 * @cpu: the processor in question.
3767 int idle_cpu(int cpu)
3769 struct rq *rq = cpu_rq(cpu);
3771 if (rq->curr != rq->idle)
3778 if (!llist_empty(&rq->wake_list))
3786 * idle_task - return the idle task for a given cpu.
3787 * @cpu: the processor in question.
3789 struct task_struct *idle_task(int cpu)
3791 return cpu_rq(cpu)->idle;
3795 * find_process_by_pid - find a process with a matching PID value.
3796 * @pid: the pid in question.
3798 static struct task_struct *find_process_by_pid(pid_t pid)
3800 return pid ? find_task_by_vpid(pid) : current;
3803 /* Actually do priority change: must hold rq lock. */
3805 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3808 p->rt_priority = prio;
3809 p->normal_prio = normal_prio(p);
3810 /* we are holding p->pi_lock already */
3811 p->prio = rt_mutex_getprio(p);
3812 if (rt_prio(p->prio))
3813 p->sched_class = &rt_sched_class;
3815 p->sched_class = &fair_sched_class;
3820 * check the target process has a UID that matches the current process's
3822 static bool check_same_owner(struct task_struct *p)
3824 const struct cred *cred = current_cred(), *pcred;
3828 pcred = __task_cred(p);
3829 match = (uid_eq(cred->euid, pcred->euid) ||
3830 uid_eq(cred->euid, pcred->uid));
3835 static int __sched_setscheduler(struct task_struct *p, int policy,
3836 const struct sched_param *param, bool user)
3838 int retval, oldprio, oldpolicy = -1, on_rq, running;
3839 unsigned long flags;
3840 const struct sched_class *prev_class;
3844 /* may grab non-irq protected spin_locks */
3845 BUG_ON(in_interrupt());
3847 /* double check policy once rq lock held */
3849 reset_on_fork = p->sched_reset_on_fork;
3850 policy = oldpolicy = p->policy;
3852 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3853 policy &= ~SCHED_RESET_ON_FORK;
3855 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3856 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3857 policy != SCHED_IDLE)
3862 * Valid priorities for SCHED_FIFO and SCHED_RR are
3863 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3864 * SCHED_BATCH and SCHED_IDLE is 0.
3866 if (param->sched_priority < 0 ||
3867 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3868 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3870 if (rt_policy(policy) != (param->sched_priority != 0))
3874 * Allow unprivileged RT tasks to decrease priority:
3876 if (user && !capable(CAP_SYS_NICE)) {
3877 if (rt_policy(policy)) {
3878 unsigned long rlim_rtprio =
3879 task_rlimit(p, RLIMIT_RTPRIO);
3881 /* can't set/change the rt policy */
3882 if (policy != p->policy && !rlim_rtprio)
3885 /* can't increase priority */
3886 if (param->sched_priority > p->rt_priority &&
3887 param->sched_priority > rlim_rtprio)
3892 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3893 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3895 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3896 if (!can_nice(p, TASK_NICE(p)))
3900 /* can't change other user's priorities */
3901 if (!check_same_owner(p))
3904 /* Normal users shall not reset the sched_reset_on_fork flag */
3905 if (p->sched_reset_on_fork && !reset_on_fork)
3910 retval = security_task_setscheduler(p);
3916 * make sure no PI-waiters arrive (or leave) while we are
3917 * changing the priority of the task:
3919 * To be able to change p->policy safely, the appropriate
3920 * runqueue lock must be held.
3922 rq = task_rq_lock(p, &flags);
3925 * Changing the policy of the stop threads its a very bad idea
3927 if (p == rq->stop) {
3928 task_rq_unlock(rq, p, &flags);
3933 * If not changing anything there's no need to proceed further:
3935 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3936 param->sched_priority == p->rt_priority))) {
3937 task_rq_unlock(rq, p, &flags);
3941 #ifdef CONFIG_RT_GROUP_SCHED
3944 * Do not allow realtime tasks into groups that have no runtime
3947 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3948 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3949 !task_group_is_autogroup(task_group(p))) {
3950 task_rq_unlock(rq, p, &flags);
3956 /* recheck policy now with rq lock held */
3957 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3958 policy = oldpolicy = -1;
3959 task_rq_unlock(rq, p, &flags);
3963 running = task_current(rq, p);
3965 dequeue_task(rq, p, 0);
3967 p->sched_class->put_prev_task(rq, p);
3969 p->sched_reset_on_fork = reset_on_fork;
3972 prev_class = p->sched_class;
3973 __setscheduler(rq, p, policy, param->sched_priority);
3976 p->sched_class->set_curr_task(rq);
3978 enqueue_task(rq, p, 0);
3980 check_class_changed(rq, p, prev_class, oldprio);
3981 task_rq_unlock(rq, p, &flags);
3983 rt_mutex_adjust_pi(p);
3989 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3990 * @p: the task in question.
3991 * @policy: new policy.
3992 * @param: structure containing the new RT priority.
3994 * NOTE that the task may be already dead.
3996 int sched_setscheduler(struct task_struct *p, int policy,
3997 const struct sched_param *param)
3999 return __sched_setscheduler(p, policy, param, true);
4001 EXPORT_SYMBOL_GPL(sched_setscheduler);
4004 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4005 * @p: the task in question.
4006 * @policy: new policy.
4007 * @param: structure containing the new RT priority.
4009 * Just like sched_setscheduler, only don't bother checking if the
4010 * current context has permission. For example, this is needed in
4011 * stop_machine(): we create temporary high priority worker threads,
4012 * but our caller might not have that capability.
4014 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4015 const struct sched_param *param)
4017 return __sched_setscheduler(p, policy, param, false);
4021 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4023 struct sched_param lparam;
4024 struct task_struct *p;
4027 if (!param || pid < 0)
4029 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4034 p = find_process_by_pid(pid);
4036 retval = sched_setscheduler(p, policy, &lparam);
4043 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4044 * @pid: the pid in question.
4045 * @policy: new policy.
4046 * @param: structure containing the new RT priority.
4048 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4049 struct sched_param __user *, param)
4051 /* negative values for policy are not valid */
4055 return do_sched_setscheduler(pid, policy, param);
4059 * sys_sched_setparam - set/change the RT priority of a thread
4060 * @pid: the pid in question.
4061 * @param: structure containing the new RT priority.
4063 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4065 return do_sched_setscheduler(pid, -1, param);
4069 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4070 * @pid: the pid in question.
4072 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4074 struct task_struct *p;
4082 p = find_process_by_pid(pid);
4084 retval = security_task_getscheduler(p);
4087 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4094 * sys_sched_getparam - get the RT priority of a thread
4095 * @pid: the pid in question.
4096 * @param: structure containing the RT priority.
4098 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4100 struct sched_param lp;
4101 struct task_struct *p;
4104 if (!param || pid < 0)
4108 p = find_process_by_pid(pid);
4113 retval = security_task_getscheduler(p);
4117 lp.sched_priority = p->rt_priority;
4121 * This one might sleep, we cannot do it with a spinlock held ...
4123 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4132 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4134 cpumask_var_t cpus_allowed, new_mask;
4135 struct task_struct *p;
4141 p = find_process_by_pid(pid);
4148 /* Prevent p going away */
4152 if (p->flags & PF_NO_SETAFFINITY) {
4156 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4160 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4162 goto out_free_cpus_allowed;
4165 if (!check_same_owner(p)) {
4167 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4174 retval = security_task_setscheduler(p);
4178 cpuset_cpus_allowed(p, cpus_allowed);
4179 cpumask_and(new_mask, in_mask, cpus_allowed);
4181 retval = set_cpus_allowed_ptr(p, new_mask);
4184 cpuset_cpus_allowed(p, cpus_allowed);
4185 if (!cpumask_subset(new_mask, cpus_allowed)) {
4187 * We must have raced with a concurrent cpuset
4188 * update. Just reset the cpus_allowed to the
4189 * cpuset's cpus_allowed
4191 cpumask_copy(new_mask, cpus_allowed);
4196 free_cpumask_var(new_mask);
4197 out_free_cpus_allowed:
4198 free_cpumask_var(cpus_allowed);
4205 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4206 struct cpumask *new_mask)
4208 if (len < cpumask_size())
4209 cpumask_clear(new_mask);
4210 else if (len > cpumask_size())
4211 len = cpumask_size();
4213 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4217 * sys_sched_setaffinity - set the cpu affinity of a process
4218 * @pid: pid of the process
4219 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4220 * @user_mask_ptr: user-space pointer to the new cpu mask
4222 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4223 unsigned long __user *, user_mask_ptr)
4225 cpumask_var_t new_mask;
4228 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4231 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4233 retval = sched_setaffinity(pid, new_mask);
4234 free_cpumask_var(new_mask);
4238 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4240 struct task_struct *p;
4241 unsigned long flags;
4248 p = find_process_by_pid(pid);
4252 retval = security_task_getscheduler(p);
4256 raw_spin_lock_irqsave(&p->pi_lock, flags);
4257 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4258 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4268 * sys_sched_getaffinity - get the cpu affinity of a process
4269 * @pid: pid of the process
4270 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4271 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4273 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4274 unsigned long __user *, user_mask_ptr)
4279 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4281 if (len & (sizeof(unsigned long)-1))
4284 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4287 ret = sched_getaffinity(pid, mask);
4289 size_t retlen = min_t(size_t, len, cpumask_size());
4291 if (copy_to_user(user_mask_ptr, mask, retlen))
4296 free_cpumask_var(mask);
4302 * sys_sched_yield - yield the current processor to other threads.
4304 * This function yields the current CPU to other tasks. If there are no
4305 * other threads running on this CPU then this function will return.
4307 SYSCALL_DEFINE0(sched_yield)
4309 struct rq *rq = this_rq_lock();
4311 schedstat_inc(rq, yld_count);
4312 current->sched_class->yield_task(rq);
4315 * Since we are going to call schedule() anyway, there's
4316 * no need to preempt or enable interrupts:
4318 __release(rq->lock);
4319 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4320 do_raw_spin_unlock(&rq->lock);
4321 sched_preempt_enable_no_resched();
4328 static inline int should_resched(void)
4330 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4333 static void __cond_resched(void)
4335 add_preempt_count(PREEMPT_ACTIVE);
4337 sub_preempt_count(PREEMPT_ACTIVE);
4340 int __sched _cond_resched(void)
4342 if (should_resched()) {
4348 EXPORT_SYMBOL(_cond_resched);
4351 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4352 * call schedule, and on return reacquire the lock.
4354 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4355 * operations here to prevent schedule() from being called twice (once via
4356 * spin_unlock(), once by hand).
4358 int __cond_resched_lock(spinlock_t *lock)
4360 int resched = should_resched();
4363 lockdep_assert_held(lock);
4365 if (spin_needbreak(lock) || resched) {
4376 EXPORT_SYMBOL(__cond_resched_lock);
4378 int __sched __cond_resched_softirq(void)
4380 BUG_ON(!in_softirq());
4382 if (should_resched()) {
4390 EXPORT_SYMBOL(__cond_resched_softirq);
4393 * yield - yield the current processor to other threads.
4395 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4397 * The scheduler is at all times free to pick the calling task as the most
4398 * eligible task to run, if removing the yield() call from your code breaks
4399 * it, its already broken.
4401 * Typical broken usage is:
4406 * where one assumes that yield() will let 'the other' process run that will
4407 * make event true. If the current task is a SCHED_FIFO task that will never
4408 * happen. Never use yield() as a progress guarantee!!
4410 * If you want to use yield() to wait for something, use wait_event().
4411 * If you want to use yield() to be 'nice' for others, use cond_resched().
4412 * If you still want to use yield(), do not!
4414 void __sched yield(void)
4416 set_current_state(TASK_RUNNING);
4419 EXPORT_SYMBOL(yield);
4422 * yield_to - yield the current processor to another thread in
4423 * your thread group, or accelerate that thread toward the
4424 * processor it's on.
4426 * @preempt: whether task preemption is allowed or not
4428 * It's the caller's job to ensure that the target task struct
4429 * can't go away on us before we can do any checks.
4432 * true (>0) if we indeed boosted the target task.
4433 * false (0) if we failed to boost the target.
4434 * -ESRCH if there's no task to yield to.
4436 bool __sched yield_to(struct task_struct *p, bool preempt)
4438 struct task_struct *curr = current;
4439 struct rq *rq, *p_rq;
4440 unsigned long flags;
4443 local_irq_save(flags);
4449 * If we're the only runnable task on the rq and target rq also
4450 * has only one task, there's absolutely no point in yielding.
4452 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4457 double_rq_lock(rq, p_rq);
4458 while (task_rq(p) != p_rq) {
4459 double_rq_unlock(rq, p_rq);
4463 if (!curr->sched_class->yield_to_task)
4466 if (curr->sched_class != p->sched_class)
4469 if (task_running(p_rq, p) || p->state)
4472 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4474 schedstat_inc(rq, yld_count);
4476 * Make p's CPU reschedule; pick_next_entity takes care of
4479 if (preempt && rq != p_rq)
4480 resched_task(p_rq->curr);
4484 double_rq_unlock(rq, p_rq);
4486 local_irq_restore(flags);
4493 EXPORT_SYMBOL_GPL(yield_to);
4496 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4497 * that process accounting knows that this is a task in IO wait state.
4499 void __sched io_schedule(void)
4501 struct rq *rq = raw_rq();
4503 delayacct_blkio_start();
4504 atomic_inc(&rq->nr_iowait);
4505 blk_flush_plug(current);
4506 current->in_iowait = 1;
4508 current->in_iowait = 0;
4509 atomic_dec(&rq->nr_iowait);
4510 delayacct_blkio_end();
4512 EXPORT_SYMBOL(io_schedule);
4514 long __sched io_schedule_timeout(long timeout)
4516 struct rq *rq = raw_rq();
4519 delayacct_blkio_start();
4520 atomic_inc(&rq->nr_iowait);
4521 blk_flush_plug(current);
4522 current->in_iowait = 1;
4523 ret = schedule_timeout(timeout);
4524 current->in_iowait = 0;
4525 atomic_dec(&rq->nr_iowait);
4526 delayacct_blkio_end();
4531 * sys_sched_get_priority_max - return maximum RT priority.
4532 * @policy: scheduling class.
4534 * this syscall returns the maximum rt_priority that can be used
4535 * by a given scheduling class.
4537 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4544 ret = MAX_USER_RT_PRIO-1;
4556 * sys_sched_get_priority_min - return minimum RT priority.
4557 * @policy: scheduling class.
4559 * this syscall returns the minimum rt_priority that can be used
4560 * by a given scheduling class.
4562 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4580 * sys_sched_rr_get_interval - return the default timeslice of a process.
4581 * @pid: pid of the process.
4582 * @interval: userspace pointer to the timeslice value.
4584 * this syscall writes the default timeslice value of a given process
4585 * into the user-space timespec buffer. A value of '0' means infinity.
4587 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4588 struct timespec __user *, interval)
4590 struct task_struct *p;
4591 unsigned int time_slice;
4592 unsigned long flags;
4602 p = find_process_by_pid(pid);
4606 retval = security_task_getscheduler(p);
4610 rq = task_rq_lock(p, &flags);
4611 time_slice = p->sched_class->get_rr_interval(rq, p);
4612 task_rq_unlock(rq, p, &flags);
4615 jiffies_to_timespec(time_slice, &t);
4616 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4624 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4626 void sched_show_task(struct task_struct *p)
4628 unsigned long free = 0;
4632 state = p->state ? __ffs(p->state) + 1 : 0;
4633 printk(KERN_INFO "%-15.15s %c", p->comm,
4634 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4635 #if BITS_PER_LONG == 32
4636 if (state == TASK_RUNNING)
4637 printk(KERN_CONT " running ");
4639 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4641 if (state == TASK_RUNNING)
4642 printk(KERN_CONT " running task ");
4644 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4646 #ifdef CONFIG_DEBUG_STACK_USAGE
4647 free = stack_not_used(p);
4650 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4652 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4653 task_pid_nr(p), ppid,
4654 (unsigned long)task_thread_info(p)->flags);
4656 show_stack(p, NULL);
4659 void show_state_filter(unsigned long state_filter)
4661 struct task_struct *g, *p;
4663 #if BITS_PER_LONG == 32
4665 " task PC stack pid father\n");
4668 " task PC stack pid father\n");
4671 do_each_thread(g, p) {
4673 * reset the NMI-timeout, listing all files on a slow
4674 * console might take a lot of time:
4676 touch_nmi_watchdog();
4677 if (!state_filter || (p->state & state_filter))
4679 } while_each_thread(g, p);
4681 touch_all_softlockup_watchdogs();
4683 #ifdef CONFIG_SCHED_DEBUG
4684 sysrq_sched_debug_show();
4688 * Only show locks if all tasks are dumped:
4691 debug_show_all_locks();
4694 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4696 idle->sched_class = &idle_sched_class;
4700 * init_idle - set up an idle thread for a given CPU
4701 * @idle: task in question
4702 * @cpu: cpu the idle task belongs to
4704 * NOTE: this function does not set the idle thread's NEED_RESCHED
4705 * flag, to make booting more robust.
4707 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4709 struct rq *rq = cpu_rq(cpu);
4710 unsigned long flags;
4712 raw_spin_lock_irqsave(&rq->lock, flags);
4715 idle->state = TASK_RUNNING;
4716 idle->se.exec_start = sched_clock();
4718 do_set_cpus_allowed(idle, cpumask_of(cpu));
4720 * We're having a chicken and egg problem, even though we are
4721 * holding rq->lock, the cpu isn't yet set to this cpu so the
4722 * lockdep check in task_group() will fail.
4724 * Similar case to sched_fork(). / Alternatively we could
4725 * use task_rq_lock() here and obtain the other rq->lock.
4730 __set_task_cpu(idle, cpu);
4733 rq->curr = rq->idle = idle;
4734 #if defined(CONFIG_SMP)
4737 raw_spin_unlock_irqrestore(&rq->lock, flags);
4739 /* Set the preempt count _outside_ the spinlocks! */
4740 task_thread_info(idle)->preempt_count = 0;
4743 * The idle tasks have their own, simple scheduling class:
4745 idle->sched_class = &idle_sched_class;
4746 ftrace_graph_init_idle_task(idle, cpu);
4747 vtime_init_idle(idle);
4748 #if defined(CONFIG_SMP)
4749 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4754 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4756 if (p->sched_class && p->sched_class->set_cpus_allowed)
4757 p->sched_class->set_cpus_allowed(p, new_mask);
4759 cpumask_copy(&p->cpus_allowed, new_mask);
4760 p->nr_cpus_allowed = cpumask_weight(new_mask);
4764 * This is how migration works:
4766 * 1) we invoke migration_cpu_stop() on the target CPU using
4768 * 2) stopper starts to run (implicitly forcing the migrated thread
4770 * 3) it checks whether the migrated task is still in the wrong runqueue.
4771 * 4) if it's in the wrong runqueue then the migration thread removes
4772 * it and puts it into the right queue.
4773 * 5) stopper completes and stop_one_cpu() returns and the migration
4778 * Change a given task's CPU affinity. Migrate the thread to a
4779 * proper CPU and schedule it away if the CPU it's executing on
4780 * is removed from the allowed bitmask.
4782 * NOTE: the caller must have a valid reference to the task, the
4783 * task must not exit() & deallocate itself prematurely. The
4784 * call is not atomic; no spinlocks may be held.
4786 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4788 unsigned long flags;
4790 unsigned int dest_cpu;
4793 rq = task_rq_lock(p, &flags);
4795 if (cpumask_equal(&p->cpus_allowed, new_mask))
4798 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4803 do_set_cpus_allowed(p, new_mask);
4805 /* Can the task run on the task's current CPU? If so, we're done */
4806 if (cpumask_test_cpu(task_cpu(p), new_mask))
4809 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4811 struct migration_arg arg = { p, dest_cpu };
4812 /* Need help from migration thread: drop lock and wait. */
4813 task_rq_unlock(rq, p, &flags);
4814 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4815 tlb_migrate_finish(p->mm);
4819 task_rq_unlock(rq, p, &flags);
4823 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4826 * Move (not current) task off this cpu, onto dest cpu. We're doing
4827 * this because either it can't run here any more (set_cpus_allowed()
4828 * away from this CPU, or CPU going down), or because we're
4829 * attempting to rebalance this task on exec (sched_exec).
4831 * So we race with normal scheduler movements, but that's OK, as long
4832 * as the task is no longer on this CPU.
4834 * Returns non-zero if task was successfully migrated.
4836 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4838 struct rq *rq_dest, *rq_src;
4841 if (unlikely(!cpu_active(dest_cpu)))
4844 rq_src = cpu_rq(src_cpu);
4845 rq_dest = cpu_rq(dest_cpu);
4847 raw_spin_lock(&p->pi_lock);
4848 double_rq_lock(rq_src, rq_dest);
4849 /* Already moved. */
4850 if (task_cpu(p) != src_cpu)
4852 /* Affinity changed (again). */
4853 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4857 * If we're not on a rq, the next wake-up will ensure we're
4861 dequeue_task(rq_src, p, 0);
4862 set_task_cpu(p, dest_cpu);
4863 enqueue_task(rq_dest, p, 0);
4864 check_preempt_curr(rq_dest, p, 0);
4869 double_rq_unlock(rq_src, rq_dest);
4870 raw_spin_unlock(&p->pi_lock);
4875 * migration_cpu_stop - this will be executed by a highprio stopper thread
4876 * and performs thread migration by bumping thread off CPU then
4877 * 'pushing' onto another runqueue.
4879 static int migration_cpu_stop(void *data)
4881 struct migration_arg *arg = data;
4884 * The original target cpu might have gone down and we might
4885 * be on another cpu but it doesn't matter.
4887 local_irq_disable();
4888 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4893 #ifdef CONFIG_HOTPLUG_CPU
4896 * Ensures that the idle task is using init_mm right before its cpu goes
4899 void idle_task_exit(void)
4901 struct mm_struct *mm = current->active_mm;
4903 BUG_ON(cpu_online(smp_processor_id()));
4906 switch_mm(mm, &init_mm, current);
4911 * Since this CPU is going 'away' for a while, fold any nr_active delta
4912 * we might have. Assumes we're called after migrate_tasks() so that the
4913 * nr_active count is stable.
4915 * Also see the comment "Global load-average calculations".
4917 static void calc_load_migrate(struct rq *rq)
4919 long delta = calc_load_fold_active(rq);
4921 atomic_long_add(delta, &calc_load_tasks);
4925 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4926 * try_to_wake_up()->select_task_rq().
4928 * Called with rq->lock held even though we'er in stop_machine() and
4929 * there's no concurrency possible, we hold the required locks anyway
4930 * because of lock validation efforts.
4932 static void migrate_tasks(unsigned int dead_cpu)
4934 struct rq *rq = cpu_rq(dead_cpu);
4935 struct task_struct *next, *stop = rq->stop;
4939 * Fudge the rq selection such that the below task selection loop
4940 * doesn't get stuck on the currently eligible stop task.
4942 * We're currently inside stop_machine() and the rq is either stuck
4943 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4944 * either way we should never end up calling schedule() until we're
4951 * There's this thread running, bail when that's the only
4954 if (rq->nr_running == 1)
4957 next = pick_next_task(rq);
4959 next->sched_class->put_prev_task(rq, next);
4961 /* Find suitable destination for @next, with force if needed. */
4962 dest_cpu = select_fallback_rq(dead_cpu, next);
4963 raw_spin_unlock(&rq->lock);
4965 __migrate_task(next, dead_cpu, dest_cpu);
4967 raw_spin_lock(&rq->lock);
4973 #endif /* CONFIG_HOTPLUG_CPU */
4975 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4977 static struct ctl_table sd_ctl_dir[] = {
4979 .procname = "sched_domain",
4985 static struct ctl_table sd_ctl_root[] = {
4987 .procname = "kernel",
4989 .child = sd_ctl_dir,
4994 static struct ctl_table *sd_alloc_ctl_entry(int n)
4996 struct ctl_table *entry =
4997 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5002 static void sd_free_ctl_entry(struct ctl_table **tablep)
5004 struct ctl_table *entry;
5007 * In the intermediate directories, both the child directory and
5008 * procname are dynamically allocated and could fail but the mode
5009 * will always be set. In the lowest directory the names are
5010 * static strings and all have proc handlers.
5012 for (entry = *tablep; entry->mode; entry++) {
5014 sd_free_ctl_entry(&entry->child);
5015 if (entry->proc_handler == NULL)
5016 kfree(entry->procname);
5023 static int min_load_idx = 0;
5024 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5027 set_table_entry(struct ctl_table *entry,
5028 const char *procname, void *data, int maxlen,
5029 umode_t mode, proc_handler *proc_handler,
5032 entry->procname = procname;
5034 entry->maxlen = maxlen;
5036 entry->proc_handler = proc_handler;
5039 entry->extra1 = &min_load_idx;
5040 entry->extra2 = &max_load_idx;
5044 static struct ctl_table *
5045 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5047 struct ctl_table *table = sd_alloc_ctl_entry(13);
5052 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5053 sizeof(long), 0644, proc_doulongvec_minmax, false);
5054 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5055 sizeof(long), 0644, proc_doulongvec_minmax, false);
5056 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5057 sizeof(int), 0644, proc_dointvec_minmax, true);
5058 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5059 sizeof(int), 0644, proc_dointvec_minmax, true);
5060 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5061 sizeof(int), 0644, proc_dointvec_minmax, true);
5062 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5063 sizeof(int), 0644, proc_dointvec_minmax, true);
5064 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5065 sizeof(int), 0644, proc_dointvec_minmax, true);
5066 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5067 sizeof(int), 0644, proc_dointvec_minmax, false);
5068 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5069 sizeof(int), 0644, proc_dointvec_minmax, false);
5070 set_table_entry(&table[9], "cache_nice_tries",
5071 &sd->cache_nice_tries,
5072 sizeof(int), 0644, proc_dointvec_minmax, false);
5073 set_table_entry(&table[10], "flags", &sd->flags,
5074 sizeof(int), 0644, proc_dointvec_minmax, false);
5075 set_table_entry(&table[11], "name", sd->name,
5076 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5077 /* &table[12] is terminator */
5082 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5084 struct ctl_table *entry, *table;
5085 struct sched_domain *sd;
5086 int domain_num = 0, i;
5089 for_each_domain(cpu, sd)
5091 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5096 for_each_domain(cpu, sd) {
5097 snprintf(buf, 32, "domain%d", i);
5098 entry->procname = kstrdup(buf, GFP_KERNEL);
5100 entry->child = sd_alloc_ctl_domain_table(sd);
5107 static struct ctl_table_header *sd_sysctl_header;
5108 static void register_sched_domain_sysctl(void)
5110 int i, cpu_num = num_possible_cpus();
5111 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5114 WARN_ON(sd_ctl_dir[0].child);
5115 sd_ctl_dir[0].child = entry;
5120 for_each_possible_cpu(i) {
5121 snprintf(buf, 32, "cpu%d", i);
5122 entry->procname = kstrdup(buf, GFP_KERNEL);
5124 entry->child = sd_alloc_ctl_cpu_table(i);
5128 WARN_ON(sd_sysctl_header);
5129 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5132 /* may be called multiple times per register */
5133 static void unregister_sched_domain_sysctl(void)
5135 if (sd_sysctl_header)
5136 unregister_sysctl_table(sd_sysctl_header);
5137 sd_sysctl_header = NULL;
5138 if (sd_ctl_dir[0].child)
5139 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5142 static void register_sched_domain_sysctl(void)
5145 static void unregister_sched_domain_sysctl(void)
5150 static void set_rq_online(struct rq *rq)
5153 const struct sched_class *class;
5155 cpumask_set_cpu(rq->cpu, rq->rd->online);
5158 for_each_class(class) {
5159 if (class->rq_online)
5160 class->rq_online(rq);
5165 static void set_rq_offline(struct rq *rq)
5168 const struct sched_class *class;
5170 for_each_class(class) {
5171 if (class->rq_offline)
5172 class->rq_offline(rq);
5175 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5181 * migration_call - callback that gets triggered when a CPU is added.
5182 * Here we can start up the necessary migration thread for the new CPU.
5184 static int __cpuinit
5185 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5187 int cpu = (long)hcpu;
5188 unsigned long flags;
5189 struct rq *rq = cpu_rq(cpu);
5191 switch (action & ~CPU_TASKS_FROZEN) {
5193 case CPU_UP_PREPARE:
5194 rq->calc_load_update = calc_load_update;
5198 /* Update our root-domain */
5199 raw_spin_lock_irqsave(&rq->lock, flags);
5201 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5205 raw_spin_unlock_irqrestore(&rq->lock, flags);
5208 #ifdef CONFIG_HOTPLUG_CPU
5210 sched_ttwu_pending();
5211 /* Update our root-domain */
5212 raw_spin_lock_irqsave(&rq->lock, flags);
5214 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5218 BUG_ON(rq->nr_running != 1); /* the migration thread */
5219 raw_spin_unlock_irqrestore(&rq->lock, flags);
5223 calc_load_migrate(rq);
5228 update_max_interval();
5234 * Register at high priority so that task migration (migrate_all_tasks)
5235 * happens before everything else. This has to be lower priority than
5236 * the notifier in the perf_event subsystem, though.
5238 static struct notifier_block __cpuinitdata migration_notifier = {
5239 .notifier_call = migration_call,
5240 .priority = CPU_PRI_MIGRATION,
5243 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5244 unsigned long action, void *hcpu)
5246 switch (action & ~CPU_TASKS_FROZEN) {
5248 case CPU_DOWN_FAILED:
5249 set_cpu_active((long)hcpu, true);
5256 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5257 unsigned long action, void *hcpu)
5259 switch (action & ~CPU_TASKS_FROZEN) {
5260 case CPU_DOWN_PREPARE:
5261 set_cpu_active((long)hcpu, false);
5268 static int __init migration_init(void)
5270 void *cpu = (void *)(long)smp_processor_id();
5273 /* Initialize migration for the boot CPU */
5274 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5275 BUG_ON(err == NOTIFY_BAD);
5276 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5277 register_cpu_notifier(&migration_notifier);
5279 /* Register cpu active notifiers */
5280 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5281 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5285 early_initcall(migration_init);
5290 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5292 #ifdef CONFIG_SCHED_DEBUG
5294 static __read_mostly int sched_debug_enabled;
5296 static int __init sched_debug_setup(char *str)
5298 sched_debug_enabled = 1;
5302 early_param("sched_debug", sched_debug_setup);
5304 static inline bool sched_debug(void)
5306 return sched_debug_enabled;
5309 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5310 struct cpumask *groupmask)
5312 struct sched_group *group = sd->groups;
5315 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5316 cpumask_clear(groupmask);
5318 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5320 if (!(sd->flags & SD_LOAD_BALANCE)) {
5321 printk("does not load-balance\n");
5323 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5328 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5330 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5331 printk(KERN_ERR "ERROR: domain->span does not contain "
5334 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5335 printk(KERN_ERR "ERROR: domain->groups does not contain"
5339 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5343 printk(KERN_ERR "ERROR: group is NULL\n");
5348 * Even though we initialize ->power to something semi-sane,
5349 * we leave power_orig unset. This allows us to detect if
5350 * domain iteration is still funny without causing /0 traps.
5352 if (!group->sgp->power_orig) {
5353 printk(KERN_CONT "\n");
5354 printk(KERN_ERR "ERROR: domain->cpu_power not "
5359 if (!cpumask_weight(sched_group_cpus(group))) {
5360 printk(KERN_CONT "\n");
5361 printk(KERN_ERR "ERROR: empty group\n");
5365 if (!(sd->flags & SD_OVERLAP) &&
5366 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5367 printk(KERN_CONT "\n");
5368 printk(KERN_ERR "ERROR: repeated CPUs\n");
5372 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5374 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5376 printk(KERN_CONT " %s", str);
5377 if (group->sgp->power != SCHED_POWER_SCALE) {
5378 printk(KERN_CONT " (cpu_power = %d)",
5382 group = group->next;
5383 } while (group != sd->groups);
5384 printk(KERN_CONT "\n");
5386 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5387 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5390 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5391 printk(KERN_ERR "ERROR: parent span is not a superset "
5392 "of domain->span\n");
5396 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5400 if (!sched_debug_enabled)
5404 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5408 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5411 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5419 #else /* !CONFIG_SCHED_DEBUG */
5420 # define sched_domain_debug(sd, cpu) do { } while (0)
5421 static inline bool sched_debug(void)
5425 #endif /* CONFIG_SCHED_DEBUG */
5427 static int sd_degenerate(struct sched_domain *sd)
5429 if (cpumask_weight(sched_domain_span(sd)) == 1)
5432 /* Following flags need at least 2 groups */
5433 if (sd->flags & (SD_LOAD_BALANCE |
5434 SD_BALANCE_NEWIDLE |
5438 SD_SHARE_PKG_RESOURCES)) {
5439 if (sd->groups != sd->groups->next)
5443 /* Following flags don't use groups */
5444 if (sd->flags & (SD_WAKE_AFFINE))
5451 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5453 unsigned long cflags = sd->flags, pflags = parent->flags;
5455 if (sd_degenerate(parent))
5458 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5461 /* Flags needing groups don't count if only 1 group in parent */
5462 if (parent->groups == parent->groups->next) {
5463 pflags &= ~(SD_LOAD_BALANCE |
5464 SD_BALANCE_NEWIDLE |
5468 SD_SHARE_PKG_RESOURCES);
5469 if (nr_node_ids == 1)
5470 pflags &= ~SD_SERIALIZE;
5472 if (~cflags & pflags)
5478 static void free_rootdomain(struct rcu_head *rcu)
5480 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5482 cpupri_cleanup(&rd->cpupri);
5483 free_cpumask_var(rd->rto_mask);
5484 free_cpumask_var(rd->online);
5485 free_cpumask_var(rd->span);
5489 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5491 struct root_domain *old_rd = NULL;
5492 unsigned long flags;
5494 raw_spin_lock_irqsave(&rq->lock, flags);
5499 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5502 cpumask_clear_cpu(rq->cpu, old_rd->span);
5505 * If we dont want to free the old_rt yet then
5506 * set old_rd to NULL to skip the freeing later
5509 if (!atomic_dec_and_test(&old_rd->refcount))
5513 atomic_inc(&rd->refcount);
5516 cpumask_set_cpu(rq->cpu, rd->span);
5517 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5520 raw_spin_unlock_irqrestore(&rq->lock, flags);
5523 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5526 static int init_rootdomain(struct root_domain *rd)
5528 memset(rd, 0, sizeof(*rd));
5530 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5532 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5534 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5537 if (cpupri_init(&rd->cpupri) != 0)
5542 free_cpumask_var(rd->rto_mask);
5544 free_cpumask_var(rd->online);
5546 free_cpumask_var(rd->span);
5552 * By default the system creates a single root-domain with all cpus as
5553 * members (mimicking the global state we have today).
5555 struct root_domain def_root_domain;
5557 static void init_defrootdomain(void)
5559 init_rootdomain(&def_root_domain);
5561 atomic_set(&def_root_domain.refcount, 1);
5564 static struct root_domain *alloc_rootdomain(void)
5566 struct root_domain *rd;
5568 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5572 if (init_rootdomain(rd) != 0) {
5580 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5582 struct sched_group *tmp, *first;
5591 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5596 } while (sg != first);
5599 static void free_sched_domain(struct rcu_head *rcu)
5601 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5604 * If its an overlapping domain it has private groups, iterate and
5607 if (sd->flags & SD_OVERLAP) {
5608 free_sched_groups(sd->groups, 1);
5609 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5610 kfree(sd->groups->sgp);
5616 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5618 call_rcu(&sd->rcu, free_sched_domain);
5621 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5623 for (; sd; sd = sd->parent)
5624 destroy_sched_domain(sd, cpu);
5628 * Keep a special pointer to the highest sched_domain that has
5629 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5630 * allows us to avoid some pointer chasing select_idle_sibling().
5632 * Also keep a unique ID per domain (we use the first cpu number in
5633 * the cpumask of the domain), this allows us to quickly tell if
5634 * two cpus are in the same cache domain, see cpus_share_cache().
5636 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5637 DEFINE_PER_CPU(int, sd_llc_id);
5639 static void update_top_cache_domain(int cpu)
5641 struct sched_domain *sd;
5644 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5646 id = cpumask_first(sched_domain_span(sd));
5648 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5649 per_cpu(sd_llc_id, cpu) = id;
5653 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5654 * hold the hotplug lock.
5657 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5659 struct rq *rq = cpu_rq(cpu);
5660 struct sched_domain *tmp;
5662 /* Remove the sched domains which do not contribute to scheduling. */
5663 for (tmp = sd; tmp; ) {
5664 struct sched_domain *parent = tmp->parent;
5668 if (sd_parent_degenerate(tmp, parent)) {
5669 tmp->parent = parent->parent;
5671 parent->parent->child = tmp;
5672 destroy_sched_domain(parent, cpu);
5677 if (sd && sd_degenerate(sd)) {
5680 destroy_sched_domain(tmp, cpu);
5685 sched_domain_debug(sd, cpu);
5687 rq_attach_root(rq, rd);
5689 rcu_assign_pointer(rq->sd, sd);
5690 destroy_sched_domains(tmp, cpu);
5692 update_top_cache_domain(cpu);
5695 /* cpus with isolated domains */
5696 static cpumask_var_t cpu_isolated_map;
5698 /* Setup the mask of cpus configured for isolated domains */
5699 static int __init isolated_cpu_setup(char *str)
5701 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5702 cpulist_parse(str, cpu_isolated_map);
5706 __setup("isolcpus=", isolated_cpu_setup);
5708 static const struct cpumask *cpu_cpu_mask(int cpu)
5710 return cpumask_of_node(cpu_to_node(cpu));
5714 struct sched_domain **__percpu sd;
5715 struct sched_group **__percpu sg;
5716 struct sched_group_power **__percpu sgp;
5720 struct sched_domain ** __percpu sd;
5721 struct root_domain *rd;
5731 struct sched_domain_topology_level;
5733 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5734 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5736 #define SDTL_OVERLAP 0x01
5738 struct sched_domain_topology_level {
5739 sched_domain_init_f init;
5740 sched_domain_mask_f mask;
5743 struct sd_data data;
5747 * Build an iteration mask that can exclude certain CPUs from the upwards
5750 * Asymmetric node setups can result in situations where the domain tree is of
5751 * unequal depth, make sure to skip domains that already cover the entire
5754 * In that case build_sched_domains() will have terminated the iteration early
5755 * and our sibling sd spans will be empty. Domains should always include the
5756 * cpu they're built on, so check that.
5759 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5761 const struct cpumask *span = sched_domain_span(sd);
5762 struct sd_data *sdd = sd->private;
5763 struct sched_domain *sibling;
5766 for_each_cpu(i, span) {
5767 sibling = *per_cpu_ptr(sdd->sd, i);
5768 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5771 cpumask_set_cpu(i, sched_group_mask(sg));
5776 * Return the canonical balance cpu for this group, this is the first cpu
5777 * of this group that's also in the iteration mask.
5779 int group_balance_cpu(struct sched_group *sg)
5781 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5785 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5787 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5788 const struct cpumask *span = sched_domain_span(sd);
5789 struct cpumask *covered = sched_domains_tmpmask;
5790 struct sd_data *sdd = sd->private;
5791 struct sched_domain *child;
5794 cpumask_clear(covered);
5796 for_each_cpu(i, span) {
5797 struct cpumask *sg_span;
5799 if (cpumask_test_cpu(i, covered))
5802 child = *per_cpu_ptr(sdd->sd, i);
5804 /* See the comment near build_group_mask(). */
5805 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5808 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5809 GFP_KERNEL, cpu_to_node(cpu));
5814 sg_span = sched_group_cpus(sg);
5816 child = child->child;
5817 cpumask_copy(sg_span, sched_domain_span(child));
5819 cpumask_set_cpu(i, sg_span);
5821 cpumask_or(covered, covered, sg_span);
5823 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5824 if (atomic_inc_return(&sg->sgp->ref) == 1)
5825 build_group_mask(sd, sg);
5828 * Initialize sgp->power such that even if we mess up the
5829 * domains and no possible iteration will get us here, we won't
5832 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5835 * Make sure the first group of this domain contains the
5836 * canonical balance cpu. Otherwise the sched_domain iteration
5837 * breaks. See update_sg_lb_stats().
5839 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5840 group_balance_cpu(sg) == cpu)
5850 sd->groups = groups;
5855 free_sched_groups(first, 0);
5860 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5862 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5863 struct sched_domain *child = sd->child;
5866 cpu = cpumask_first(sched_domain_span(child));
5869 *sg = *per_cpu_ptr(sdd->sg, cpu);
5870 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5871 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5878 * build_sched_groups will build a circular linked list of the groups
5879 * covered by the given span, and will set each group's ->cpumask correctly,
5880 * and ->cpu_power to 0.
5882 * Assumes the sched_domain tree is fully constructed
5885 build_sched_groups(struct sched_domain *sd, int cpu)
5887 struct sched_group *first = NULL, *last = NULL;
5888 struct sd_data *sdd = sd->private;
5889 const struct cpumask *span = sched_domain_span(sd);
5890 struct cpumask *covered;
5893 get_group(cpu, sdd, &sd->groups);
5894 atomic_inc(&sd->groups->ref);
5896 if (cpu != cpumask_first(sched_domain_span(sd)))
5899 lockdep_assert_held(&sched_domains_mutex);
5900 covered = sched_domains_tmpmask;
5902 cpumask_clear(covered);
5904 for_each_cpu(i, span) {
5905 struct sched_group *sg;
5906 int group = get_group(i, sdd, &sg);
5909 if (cpumask_test_cpu(i, covered))
5912 cpumask_clear(sched_group_cpus(sg));
5914 cpumask_setall(sched_group_mask(sg));
5916 for_each_cpu(j, span) {
5917 if (get_group(j, sdd, NULL) != group)
5920 cpumask_set_cpu(j, covered);
5921 cpumask_set_cpu(j, sched_group_cpus(sg));
5936 * Initialize sched groups cpu_power.
5938 * cpu_power indicates the capacity of sched group, which is used while
5939 * distributing the load between different sched groups in a sched domain.
5940 * Typically cpu_power for all the groups in a sched domain will be same unless
5941 * there are asymmetries in the topology. If there are asymmetries, group
5942 * having more cpu_power will pickup more load compared to the group having
5945 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5947 struct sched_group *sg = sd->groups;
5949 WARN_ON(!sd || !sg);
5952 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5954 } while (sg != sd->groups);
5956 if (cpu != group_balance_cpu(sg))
5959 update_group_power(sd, cpu);
5960 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5963 int __weak arch_sd_sibling_asym_packing(void)
5965 return 0*SD_ASYM_PACKING;
5969 * Initializers for schedule domains
5970 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5973 #ifdef CONFIG_SCHED_DEBUG
5974 # define SD_INIT_NAME(sd, type) sd->name = #type
5976 # define SD_INIT_NAME(sd, type) do { } while (0)
5979 #define SD_INIT_FUNC(type) \
5980 static noinline struct sched_domain * \
5981 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5983 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5984 *sd = SD_##type##_INIT; \
5985 SD_INIT_NAME(sd, type); \
5986 sd->private = &tl->data; \
5991 #ifdef CONFIG_SCHED_SMT
5992 SD_INIT_FUNC(SIBLING)
5994 #ifdef CONFIG_SCHED_MC
5997 #ifdef CONFIG_SCHED_BOOK
6001 static int default_relax_domain_level = -1;
6002 int sched_domain_level_max;
6004 static int __init setup_relax_domain_level(char *str)
6006 if (kstrtoint(str, 0, &default_relax_domain_level))
6007 pr_warn("Unable to set relax_domain_level\n");
6011 __setup("relax_domain_level=", setup_relax_domain_level);
6013 static void set_domain_attribute(struct sched_domain *sd,
6014 struct sched_domain_attr *attr)
6018 if (!attr || attr->relax_domain_level < 0) {
6019 if (default_relax_domain_level < 0)
6022 request = default_relax_domain_level;
6024 request = attr->relax_domain_level;
6025 if (request < sd->level) {
6026 /* turn off idle balance on this domain */
6027 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6029 /* turn on idle balance on this domain */
6030 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6034 static void __sdt_free(const struct cpumask *cpu_map);
6035 static int __sdt_alloc(const struct cpumask *cpu_map);
6037 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6038 const struct cpumask *cpu_map)
6042 if (!atomic_read(&d->rd->refcount))
6043 free_rootdomain(&d->rd->rcu); /* fall through */
6045 free_percpu(d->sd); /* fall through */
6047 __sdt_free(cpu_map); /* fall through */
6053 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6054 const struct cpumask *cpu_map)
6056 memset(d, 0, sizeof(*d));
6058 if (__sdt_alloc(cpu_map))
6059 return sa_sd_storage;
6060 d->sd = alloc_percpu(struct sched_domain *);
6062 return sa_sd_storage;
6063 d->rd = alloc_rootdomain();
6066 return sa_rootdomain;
6070 * NULL the sd_data elements we've used to build the sched_domain and
6071 * sched_group structure so that the subsequent __free_domain_allocs()
6072 * will not free the data we're using.
6074 static void claim_allocations(int cpu, struct sched_domain *sd)
6076 struct sd_data *sdd = sd->private;
6078 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6079 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6081 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6082 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6084 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6085 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6088 #ifdef CONFIG_SCHED_SMT
6089 static const struct cpumask *cpu_smt_mask(int cpu)
6091 return topology_thread_cpumask(cpu);
6096 * Topology list, bottom-up.
6098 static struct sched_domain_topology_level default_topology[] = {
6099 #ifdef CONFIG_SCHED_SMT
6100 { sd_init_SIBLING, cpu_smt_mask, },
6102 #ifdef CONFIG_SCHED_MC
6103 { sd_init_MC, cpu_coregroup_mask, },
6105 #ifdef CONFIG_SCHED_BOOK
6106 { sd_init_BOOK, cpu_book_mask, },
6108 { sd_init_CPU, cpu_cpu_mask, },
6112 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6116 static int sched_domains_numa_levels;
6117 static int *sched_domains_numa_distance;
6118 static struct cpumask ***sched_domains_numa_masks;
6119 static int sched_domains_curr_level;
6121 static inline int sd_local_flags(int level)
6123 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6126 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6129 static struct sched_domain *
6130 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6132 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6133 int level = tl->numa_level;
6134 int sd_weight = cpumask_weight(
6135 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6137 *sd = (struct sched_domain){
6138 .min_interval = sd_weight,
6139 .max_interval = 2*sd_weight,
6141 .imbalance_pct = 125,
6142 .cache_nice_tries = 2,
6149 .flags = 1*SD_LOAD_BALANCE
6150 | 1*SD_BALANCE_NEWIDLE
6155 | 0*SD_SHARE_CPUPOWER
6156 | 0*SD_SHARE_PKG_RESOURCES
6158 | 0*SD_PREFER_SIBLING
6159 | sd_local_flags(level)
6161 .last_balance = jiffies,
6162 .balance_interval = sd_weight,
6164 SD_INIT_NAME(sd, NUMA);
6165 sd->private = &tl->data;
6168 * Ugly hack to pass state to sd_numa_mask()...
6170 sched_domains_curr_level = tl->numa_level;
6175 static const struct cpumask *sd_numa_mask(int cpu)
6177 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6180 static void sched_numa_warn(const char *str)
6182 static int done = false;
6190 printk(KERN_WARNING "ERROR: %s\n\n", str);
6192 for (i = 0; i < nr_node_ids; i++) {
6193 printk(KERN_WARNING " ");
6194 for (j = 0; j < nr_node_ids; j++)
6195 printk(KERN_CONT "%02d ", node_distance(i,j));
6196 printk(KERN_CONT "\n");
6198 printk(KERN_WARNING "\n");
6201 static bool find_numa_distance(int distance)
6205 if (distance == node_distance(0, 0))
6208 for (i = 0; i < sched_domains_numa_levels; i++) {
6209 if (sched_domains_numa_distance[i] == distance)
6216 static void sched_init_numa(void)
6218 int next_distance, curr_distance = node_distance(0, 0);
6219 struct sched_domain_topology_level *tl;
6223 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6224 if (!sched_domains_numa_distance)
6228 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6229 * unique distances in the node_distance() table.
6231 * Assumes node_distance(0,j) includes all distances in
6232 * node_distance(i,j) in order to avoid cubic time.
6234 next_distance = curr_distance;
6235 for (i = 0; i < nr_node_ids; i++) {
6236 for (j = 0; j < nr_node_ids; j++) {
6237 for (k = 0; k < nr_node_ids; k++) {
6238 int distance = node_distance(i, k);
6240 if (distance > curr_distance &&
6241 (distance < next_distance ||
6242 next_distance == curr_distance))
6243 next_distance = distance;
6246 * While not a strong assumption it would be nice to know
6247 * about cases where if node A is connected to B, B is not
6248 * equally connected to A.
6250 if (sched_debug() && node_distance(k, i) != distance)
6251 sched_numa_warn("Node-distance not symmetric");
6253 if (sched_debug() && i && !find_numa_distance(distance))
6254 sched_numa_warn("Node-0 not representative");
6256 if (next_distance != curr_distance) {
6257 sched_domains_numa_distance[level++] = next_distance;
6258 sched_domains_numa_levels = level;
6259 curr_distance = next_distance;
6264 * In case of sched_debug() we verify the above assumption.
6270 * 'level' contains the number of unique distances, excluding the
6271 * identity distance node_distance(i,i).
6273 * The sched_domains_numa_distance[] array includes the actual distance
6278 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6279 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6280 * the array will contain less then 'level' members. This could be
6281 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6282 * in other functions.
6284 * We reset it to 'level' at the end of this function.
6286 sched_domains_numa_levels = 0;
6288 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6289 if (!sched_domains_numa_masks)
6293 * Now for each level, construct a mask per node which contains all
6294 * cpus of nodes that are that many hops away from us.
6296 for (i = 0; i < level; i++) {
6297 sched_domains_numa_masks[i] =
6298 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6299 if (!sched_domains_numa_masks[i])
6302 for (j = 0; j < nr_node_ids; j++) {
6303 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6307 sched_domains_numa_masks[i][j] = mask;
6309 for (k = 0; k < nr_node_ids; k++) {
6310 if (node_distance(j, k) > sched_domains_numa_distance[i])
6313 cpumask_or(mask, mask, cpumask_of_node(k));
6318 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6319 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6324 * Copy the default topology bits..
6326 for (i = 0; default_topology[i].init; i++)
6327 tl[i] = default_topology[i];
6330 * .. and append 'j' levels of NUMA goodness.
6332 for (j = 0; j < level; i++, j++) {
6333 tl[i] = (struct sched_domain_topology_level){
6334 .init = sd_numa_init,
6335 .mask = sd_numa_mask,
6336 .flags = SDTL_OVERLAP,
6341 sched_domain_topology = tl;
6343 sched_domains_numa_levels = level;
6346 static void sched_domains_numa_masks_set(int cpu)
6349 int node = cpu_to_node(cpu);
6351 for (i = 0; i < sched_domains_numa_levels; i++) {
6352 for (j = 0; j < nr_node_ids; j++) {
6353 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6354 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6359 static void sched_domains_numa_masks_clear(int cpu)
6362 for (i = 0; i < sched_domains_numa_levels; i++) {
6363 for (j = 0; j < nr_node_ids; j++)
6364 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6369 * Update sched_domains_numa_masks[level][node] array when new cpus
6372 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6373 unsigned long action,
6376 int cpu = (long)hcpu;
6378 switch (action & ~CPU_TASKS_FROZEN) {
6380 sched_domains_numa_masks_set(cpu);
6384 sched_domains_numa_masks_clear(cpu);
6394 static inline void sched_init_numa(void)
6398 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6399 unsigned long action,
6404 #endif /* CONFIG_NUMA */
6406 static int __sdt_alloc(const struct cpumask *cpu_map)
6408 struct sched_domain_topology_level *tl;
6411 for (tl = sched_domain_topology; tl->init; tl++) {
6412 struct sd_data *sdd = &tl->data;
6414 sdd->sd = alloc_percpu(struct sched_domain *);
6418 sdd->sg = alloc_percpu(struct sched_group *);
6422 sdd->sgp = alloc_percpu(struct sched_group_power *);
6426 for_each_cpu(j, cpu_map) {
6427 struct sched_domain *sd;
6428 struct sched_group *sg;
6429 struct sched_group_power *sgp;
6431 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6432 GFP_KERNEL, cpu_to_node(j));
6436 *per_cpu_ptr(sdd->sd, j) = sd;
6438 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6439 GFP_KERNEL, cpu_to_node(j));
6445 *per_cpu_ptr(sdd->sg, j) = sg;
6447 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6448 GFP_KERNEL, cpu_to_node(j));
6452 *per_cpu_ptr(sdd->sgp, j) = sgp;
6459 static void __sdt_free(const struct cpumask *cpu_map)
6461 struct sched_domain_topology_level *tl;
6464 for (tl = sched_domain_topology; tl->init; tl++) {
6465 struct sd_data *sdd = &tl->data;
6467 for_each_cpu(j, cpu_map) {
6468 struct sched_domain *sd;
6471 sd = *per_cpu_ptr(sdd->sd, j);
6472 if (sd && (sd->flags & SD_OVERLAP))
6473 free_sched_groups(sd->groups, 0);
6474 kfree(*per_cpu_ptr(sdd->sd, j));
6478 kfree(*per_cpu_ptr(sdd->sg, j));
6480 kfree(*per_cpu_ptr(sdd->sgp, j));
6482 free_percpu(sdd->sd);
6484 free_percpu(sdd->sg);
6486 free_percpu(sdd->sgp);
6491 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6492 struct s_data *d, const struct cpumask *cpu_map,
6493 struct sched_domain_attr *attr, struct sched_domain *child,
6496 struct sched_domain *sd = tl->init(tl, cpu);
6500 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6502 sd->level = child->level + 1;
6503 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6507 set_domain_attribute(sd, attr);
6513 * Build sched domains for a given set of cpus and attach the sched domains
6514 * to the individual cpus
6516 static int build_sched_domains(const struct cpumask *cpu_map,
6517 struct sched_domain_attr *attr)
6519 enum s_alloc alloc_state = sa_none;
6520 struct sched_domain *sd;
6522 int i, ret = -ENOMEM;
6524 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6525 if (alloc_state != sa_rootdomain)
6528 /* Set up domains for cpus specified by the cpu_map. */
6529 for_each_cpu(i, cpu_map) {
6530 struct sched_domain_topology_level *tl;
6533 for (tl = sched_domain_topology; tl->init; tl++) {
6534 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6535 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6536 sd->flags |= SD_OVERLAP;
6537 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6544 *per_cpu_ptr(d.sd, i) = sd;
6547 /* Build the groups for the domains */
6548 for_each_cpu(i, cpu_map) {
6549 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6550 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6551 if (sd->flags & SD_OVERLAP) {
6552 if (build_overlap_sched_groups(sd, i))
6555 if (build_sched_groups(sd, i))
6561 /* Calculate CPU power for physical packages and nodes */
6562 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6563 if (!cpumask_test_cpu(i, cpu_map))
6566 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6567 claim_allocations(i, sd);
6568 init_sched_groups_power(i, sd);
6572 /* Attach the domains */
6574 for_each_cpu(i, cpu_map) {
6575 sd = *per_cpu_ptr(d.sd, i);
6576 cpu_attach_domain(sd, d.rd, i);
6582 __free_domain_allocs(&d, alloc_state, cpu_map);
6586 static cpumask_var_t *doms_cur; /* current sched domains */
6587 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6588 static struct sched_domain_attr *dattr_cur;
6589 /* attribues of custom domains in 'doms_cur' */
6592 * Special case: If a kmalloc of a doms_cur partition (array of
6593 * cpumask) fails, then fallback to a single sched domain,
6594 * as determined by the single cpumask fallback_doms.
6596 static cpumask_var_t fallback_doms;
6599 * arch_update_cpu_topology lets virtualized architectures update the
6600 * cpu core maps. It is supposed to return 1 if the topology changed
6601 * or 0 if it stayed the same.
6603 int __attribute__((weak)) arch_update_cpu_topology(void)
6608 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6611 cpumask_var_t *doms;
6613 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6616 for (i = 0; i < ndoms; i++) {
6617 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6618 free_sched_domains(doms, i);
6625 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6628 for (i = 0; i < ndoms; i++)
6629 free_cpumask_var(doms[i]);
6634 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6635 * For now this just excludes isolated cpus, but could be used to
6636 * exclude other special cases in the future.
6638 static int init_sched_domains(const struct cpumask *cpu_map)
6642 arch_update_cpu_topology();
6644 doms_cur = alloc_sched_domains(ndoms_cur);
6646 doms_cur = &fallback_doms;
6647 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6648 err = build_sched_domains(doms_cur[0], NULL);
6649 register_sched_domain_sysctl();
6655 * Detach sched domains from a group of cpus specified in cpu_map
6656 * These cpus will now be attached to the NULL domain
6658 static void detach_destroy_domains(const struct cpumask *cpu_map)
6663 for_each_cpu(i, cpu_map)
6664 cpu_attach_domain(NULL, &def_root_domain, i);
6668 /* handle null as "default" */
6669 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6670 struct sched_domain_attr *new, int idx_new)
6672 struct sched_domain_attr tmp;
6679 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6680 new ? (new + idx_new) : &tmp,
6681 sizeof(struct sched_domain_attr));
6685 * Partition sched domains as specified by the 'ndoms_new'
6686 * cpumasks in the array doms_new[] of cpumasks. This compares
6687 * doms_new[] to the current sched domain partitioning, doms_cur[].
6688 * It destroys each deleted domain and builds each new domain.
6690 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6691 * The masks don't intersect (don't overlap.) We should setup one
6692 * sched domain for each mask. CPUs not in any of the cpumasks will
6693 * not be load balanced. If the same cpumask appears both in the
6694 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6697 * The passed in 'doms_new' should be allocated using
6698 * alloc_sched_domains. This routine takes ownership of it and will
6699 * free_sched_domains it when done with it. If the caller failed the
6700 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6701 * and partition_sched_domains() will fallback to the single partition
6702 * 'fallback_doms', it also forces the domains to be rebuilt.
6704 * If doms_new == NULL it will be replaced with cpu_online_mask.
6705 * ndoms_new == 0 is a special case for destroying existing domains,
6706 * and it will not create the default domain.
6708 * Call with hotplug lock held
6710 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6711 struct sched_domain_attr *dattr_new)
6716 mutex_lock(&sched_domains_mutex);
6718 /* always unregister in case we don't destroy any domains */
6719 unregister_sched_domain_sysctl();
6721 /* Let architecture update cpu core mappings. */
6722 new_topology = arch_update_cpu_topology();
6724 n = doms_new ? ndoms_new : 0;
6726 /* Destroy deleted domains */
6727 for (i = 0; i < ndoms_cur; i++) {
6728 for (j = 0; j < n && !new_topology; j++) {
6729 if (cpumask_equal(doms_cur[i], doms_new[j])
6730 && dattrs_equal(dattr_cur, i, dattr_new, j))
6733 /* no match - a current sched domain not in new doms_new[] */
6734 detach_destroy_domains(doms_cur[i]);
6739 if (doms_new == NULL) {
6741 doms_new = &fallback_doms;
6742 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6743 WARN_ON_ONCE(dattr_new);
6746 /* Build new domains */
6747 for (i = 0; i < ndoms_new; i++) {
6748 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6749 if (cpumask_equal(doms_new[i], doms_cur[j])
6750 && dattrs_equal(dattr_new, i, dattr_cur, j))
6753 /* no match - add a new doms_new */
6754 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6759 /* Remember the new sched domains */
6760 if (doms_cur != &fallback_doms)
6761 free_sched_domains(doms_cur, ndoms_cur);
6762 kfree(dattr_cur); /* kfree(NULL) is safe */
6763 doms_cur = doms_new;
6764 dattr_cur = dattr_new;
6765 ndoms_cur = ndoms_new;
6767 register_sched_domain_sysctl();
6769 mutex_unlock(&sched_domains_mutex);
6772 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6775 * Update cpusets according to cpu_active mask. If cpusets are
6776 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6777 * around partition_sched_domains().
6779 * If we come here as part of a suspend/resume, don't touch cpusets because we
6780 * want to restore it back to its original state upon resume anyway.
6782 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6786 case CPU_ONLINE_FROZEN:
6787 case CPU_DOWN_FAILED_FROZEN:
6790 * num_cpus_frozen tracks how many CPUs are involved in suspend
6791 * resume sequence. As long as this is not the last online
6792 * operation in the resume sequence, just build a single sched
6793 * domain, ignoring cpusets.
6796 if (likely(num_cpus_frozen)) {
6797 partition_sched_domains(1, NULL, NULL);
6802 * This is the last CPU online operation. So fall through and
6803 * restore the original sched domains by considering the
6804 * cpuset configurations.
6808 case CPU_DOWN_FAILED:
6809 cpuset_update_active_cpus(true);
6817 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6821 case CPU_DOWN_PREPARE:
6822 cpuset_update_active_cpus(false);
6824 case CPU_DOWN_PREPARE_FROZEN:
6826 partition_sched_domains(1, NULL, NULL);
6834 void __init sched_init_smp(void)
6836 cpumask_var_t non_isolated_cpus;
6838 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6839 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6844 mutex_lock(&sched_domains_mutex);
6845 init_sched_domains(cpu_active_mask);
6846 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6847 if (cpumask_empty(non_isolated_cpus))
6848 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6849 mutex_unlock(&sched_domains_mutex);
6852 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6853 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6854 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6856 /* RT runtime code needs to handle some hotplug events */
6857 hotcpu_notifier(update_runtime, 0);
6861 /* Move init over to a non-isolated CPU */
6862 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6864 sched_init_granularity();
6865 free_cpumask_var(non_isolated_cpus);
6867 init_sched_rt_class();
6870 void __init sched_init_smp(void)
6872 sched_init_granularity();
6874 #endif /* CONFIG_SMP */
6876 const_debug unsigned int sysctl_timer_migration = 1;
6878 int in_sched_functions(unsigned long addr)
6880 return in_lock_functions(addr) ||
6881 (addr >= (unsigned long)__sched_text_start
6882 && addr < (unsigned long)__sched_text_end);
6885 #ifdef CONFIG_CGROUP_SCHED
6887 * Default task group.
6888 * Every task in system belongs to this group at bootup.
6890 struct task_group root_task_group;
6891 LIST_HEAD(task_groups);
6894 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6896 void __init sched_init(void)
6899 unsigned long alloc_size = 0, ptr;
6901 #ifdef CONFIG_FAIR_GROUP_SCHED
6902 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6904 #ifdef CONFIG_RT_GROUP_SCHED
6905 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6907 #ifdef CONFIG_CPUMASK_OFFSTACK
6908 alloc_size += num_possible_cpus() * cpumask_size();
6911 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6913 #ifdef CONFIG_FAIR_GROUP_SCHED
6914 root_task_group.se = (struct sched_entity **)ptr;
6915 ptr += nr_cpu_ids * sizeof(void **);
6917 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6918 ptr += nr_cpu_ids * sizeof(void **);
6920 #endif /* CONFIG_FAIR_GROUP_SCHED */
6921 #ifdef CONFIG_RT_GROUP_SCHED
6922 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6923 ptr += nr_cpu_ids * sizeof(void **);
6925 root_task_group.rt_rq = (struct rt_rq **)ptr;
6926 ptr += nr_cpu_ids * sizeof(void **);
6928 #endif /* CONFIG_RT_GROUP_SCHED */
6929 #ifdef CONFIG_CPUMASK_OFFSTACK
6930 for_each_possible_cpu(i) {
6931 per_cpu(load_balance_mask, i) = (void *)ptr;
6932 ptr += cpumask_size();
6934 #endif /* CONFIG_CPUMASK_OFFSTACK */
6938 init_defrootdomain();
6941 init_rt_bandwidth(&def_rt_bandwidth,
6942 global_rt_period(), global_rt_runtime());
6944 #ifdef CONFIG_RT_GROUP_SCHED
6945 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6946 global_rt_period(), global_rt_runtime());
6947 #endif /* CONFIG_RT_GROUP_SCHED */
6949 #ifdef CONFIG_CGROUP_SCHED
6950 list_add(&root_task_group.list, &task_groups);
6951 INIT_LIST_HEAD(&root_task_group.children);
6952 INIT_LIST_HEAD(&root_task_group.siblings);
6953 autogroup_init(&init_task);
6955 #endif /* CONFIG_CGROUP_SCHED */
6957 for_each_possible_cpu(i) {
6961 raw_spin_lock_init(&rq->lock);
6963 rq->calc_load_active = 0;
6964 rq->calc_load_update = jiffies + LOAD_FREQ;
6965 init_cfs_rq(&rq->cfs);
6966 init_rt_rq(&rq->rt, rq);
6967 #ifdef CONFIG_FAIR_GROUP_SCHED
6968 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6969 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6971 * How much cpu bandwidth does root_task_group get?
6973 * In case of task-groups formed thr' the cgroup filesystem, it
6974 * gets 100% of the cpu resources in the system. This overall
6975 * system cpu resource is divided among the tasks of
6976 * root_task_group and its child task-groups in a fair manner,
6977 * based on each entity's (task or task-group's) weight
6978 * (se->load.weight).
6980 * In other words, if root_task_group has 10 tasks of weight
6981 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6982 * then A0's share of the cpu resource is:
6984 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6986 * We achieve this by letting root_task_group's tasks sit
6987 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6989 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6990 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6991 #endif /* CONFIG_FAIR_GROUP_SCHED */
6993 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6994 #ifdef CONFIG_RT_GROUP_SCHED
6995 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6996 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6999 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7000 rq->cpu_load[j] = 0;
7002 rq->last_load_update_tick = jiffies;
7007 rq->cpu_power = SCHED_POWER_SCALE;
7008 rq->post_schedule = 0;
7009 rq->active_balance = 0;
7010 rq->next_balance = jiffies;
7015 rq->avg_idle = 2*sysctl_sched_migration_cost;
7017 INIT_LIST_HEAD(&rq->cfs_tasks);
7019 rq_attach_root(rq, &def_root_domain);
7020 #ifdef CONFIG_NO_HZ_COMMON
7023 #ifdef CONFIG_NO_HZ_FULL
7024 rq->last_sched_tick = 0;
7028 atomic_set(&rq->nr_iowait, 0);
7031 set_load_weight(&init_task);
7033 #ifdef CONFIG_PREEMPT_NOTIFIERS
7034 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7037 #ifdef CONFIG_RT_MUTEXES
7038 plist_head_init(&init_task.pi_waiters);
7042 * The boot idle thread does lazy MMU switching as well:
7044 atomic_inc(&init_mm.mm_count);
7045 enter_lazy_tlb(&init_mm, current);
7048 * Make us the idle thread. Technically, schedule() should not be
7049 * called from this thread, however somewhere below it might be,
7050 * but because we are the idle thread, we just pick up running again
7051 * when this runqueue becomes "idle".
7053 init_idle(current, smp_processor_id());
7055 calc_load_update = jiffies + LOAD_FREQ;
7058 * During early bootup we pretend to be a normal task:
7060 current->sched_class = &fair_sched_class;
7063 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7064 /* May be allocated at isolcpus cmdline parse time */
7065 if (cpu_isolated_map == NULL)
7066 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7067 idle_thread_set_boot_cpu();
7069 init_sched_fair_class();
7071 scheduler_running = 1;
7074 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7075 static inline int preempt_count_equals(int preempt_offset)
7077 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7079 return (nested == preempt_offset);
7082 void __might_sleep(const char *file, int line, int preempt_offset)
7084 static unsigned long prev_jiffy; /* ratelimiting */
7086 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7087 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7088 system_state != SYSTEM_RUNNING || oops_in_progress)
7090 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7092 prev_jiffy = jiffies;
7095 "BUG: sleeping function called from invalid context at %s:%d\n",
7098 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7099 in_atomic(), irqs_disabled(),
7100 current->pid, current->comm);
7102 debug_show_held_locks(current);
7103 if (irqs_disabled())
7104 print_irqtrace_events(current);
7107 EXPORT_SYMBOL(__might_sleep);
7110 #ifdef CONFIG_MAGIC_SYSRQ
7111 static void normalize_task(struct rq *rq, struct task_struct *p)
7113 const struct sched_class *prev_class = p->sched_class;
7114 int old_prio = p->prio;
7119 dequeue_task(rq, p, 0);
7120 __setscheduler(rq, p, SCHED_NORMAL, 0);
7122 enqueue_task(rq, p, 0);
7123 resched_task(rq->curr);
7126 check_class_changed(rq, p, prev_class, old_prio);
7129 void normalize_rt_tasks(void)
7131 struct task_struct *g, *p;
7132 unsigned long flags;
7135 read_lock_irqsave(&tasklist_lock, flags);
7136 do_each_thread(g, p) {
7138 * Only normalize user tasks:
7143 p->se.exec_start = 0;
7144 #ifdef CONFIG_SCHEDSTATS
7145 p->se.statistics.wait_start = 0;
7146 p->se.statistics.sleep_start = 0;
7147 p->se.statistics.block_start = 0;
7152 * Renice negative nice level userspace
7155 if (TASK_NICE(p) < 0 && p->mm)
7156 set_user_nice(p, 0);
7160 raw_spin_lock(&p->pi_lock);
7161 rq = __task_rq_lock(p);
7163 normalize_task(rq, p);
7165 __task_rq_unlock(rq);
7166 raw_spin_unlock(&p->pi_lock);
7167 } while_each_thread(g, p);
7169 read_unlock_irqrestore(&tasklist_lock, flags);
7172 #endif /* CONFIG_MAGIC_SYSRQ */
7174 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7176 * These functions are only useful for the IA64 MCA handling, or kdb.
7178 * They can only be called when the whole system has been
7179 * stopped - every CPU needs to be quiescent, and no scheduling
7180 * activity can take place. Using them for anything else would
7181 * be a serious bug, and as a result, they aren't even visible
7182 * under any other configuration.
7186 * curr_task - return the current task for a given cpu.
7187 * @cpu: the processor in question.
7189 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7191 struct task_struct *curr_task(int cpu)
7193 return cpu_curr(cpu);
7196 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7200 * set_curr_task - set the current task for a given cpu.
7201 * @cpu: the processor in question.
7202 * @p: the task pointer to set.
7204 * Description: This function must only be used when non-maskable interrupts
7205 * are serviced on a separate stack. It allows the architecture to switch the
7206 * notion of the current task on a cpu in a non-blocking manner. This function
7207 * must be called with all CPU's synchronized, and interrupts disabled, the
7208 * and caller must save the original value of the current task (see
7209 * curr_task() above) and restore that value before reenabling interrupts and
7210 * re-starting the system.
7212 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7214 void set_curr_task(int cpu, struct task_struct *p)
7221 #ifdef CONFIG_CGROUP_SCHED
7222 /* task_group_lock serializes the addition/removal of task groups */
7223 static DEFINE_SPINLOCK(task_group_lock);
7225 static void free_sched_group(struct task_group *tg)
7227 free_fair_sched_group(tg);
7228 free_rt_sched_group(tg);
7233 /* allocate runqueue etc for a new task group */
7234 struct task_group *sched_create_group(struct task_group *parent)
7236 struct task_group *tg;
7238 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7240 return ERR_PTR(-ENOMEM);
7242 if (!alloc_fair_sched_group(tg, parent))
7245 if (!alloc_rt_sched_group(tg, parent))
7251 free_sched_group(tg);
7252 return ERR_PTR(-ENOMEM);
7255 void sched_online_group(struct task_group *tg, struct task_group *parent)
7257 unsigned long flags;
7259 spin_lock_irqsave(&task_group_lock, flags);
7260 list_add_rcu(&tg->list, &task_groups);
7262 WARN_ON(!parent); /* root should already exist */
7264 tg->parent = parent;
7265 INIT_LIST_HEAD(&tg->children);
7266 list_add_rcu(&tg->siblings, &parent->children);
7267 spin_unlock_irqrestore(&task_group_lock, flags);
7270 /* rcu callback to free various structures associated with a task group */
7271 static void free_sched_group_rcu(struct rcu_head *rhp)
7273 /* now it should be safe to free those cfs_rqs */
7274 free_sched_group(container_of(rhp, struct task_group, rcu));
7277 /* Destroy runqueue etc associated with a task group */
7278 void sched_destroy_group(struct task_group *tg)
7280 /* wait for possible concurrent references to cfs_rqs complete */
7281 call_rcu(&tg->rcu, free_sched_group_rcu);
7284 void sched_offline_group(struct task_group *tg)
7286 unsigned long flags;
7289 /* end participation in shares distribution */
7290 for_each_possible_cpu(i)
7291 unregister_fair_sched_group(tg, i);
7293 spin_lock_irqsave(&task_group_lock, flags);
7294 list_del_rcu(&tg->list);
7295 list_del_rcu(&tg->siblings);
7296 spin_unlock_irqrestore(&task_group_lock, flags);
7299 /* change task's runqueue when it moves between groups.
7300 * The caller of this function should have put the task in its new group
7301 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7302 * reflect its new group.
7304 void sched_move_task(struct task_struct *tsk)
7306 struct task_group *tg;
7308 unsigned long flags;
7311 rq = task_rq_lock(tsk, &flags);
7313 running = task_current(rq, tsk);
7317 dequeue_task(rq, tsk, 0);
7318 if (unlikely(running))
7319 tsk->sched_class->put_prev_task(rq, tsk);
7321 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7322 lockdep_is_held(&tsk->sighand->siglock)),
7323 struct task_group, css);
7324 tg = autogroup_task_group(tsk, tg);
7325 tsk->sched_task_group = tg;
7327 #ifdef CONFIG_FAIR_GROUP_SCHED
7328 if (tsk->sched_class->task_move_group)
7329 tsk->sched_class->task_move_group(tsk, on_rq);
7332 set_task_rq(tsk, task_cpu(tsk));
7334 if (unlikely(running))
7335 tsk->sched_class->set_curr_task(rq);
7337 enqueue_task(rq, tsk, 0);
7339 task_rq_unlock(rq, tsk, &flags);
7341 #endif /* CONFIG_CGROUP_SCHED */
7343 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7344 static unsigned long to_ratio(u64 period, u64 runtime)
7346 if (runtime == RUNTIME_INF)
7349 return div64_u64(runtime << 20, period);
7353 #ifdef CONFIG_RT_GROUP_SCHED
7355 * Ensure that the real time constraints are schedulable.
7357 static DEFINE_MUTEX(rt_constraints_mutex);
7359 /* Must be called with tasklist_lock held */
7360 static inline int tg_has_rt_tasks(struct task_group *tg)
7362 struct task_struct *g, *p;
7364 do_each_thread(g, p) {
7365 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7367 } while_each_thread(g, p);
7372 struct rt_schedulable_data {
7373 struct task_group *tg;
7378 static int tg_rt_schedulable(struct task_group *tg, void *data)
7380 struct rt_schedulable_data *d = data;
7381 struct task_group *child;
7382 unsigned long total, sum = 0;
7383 u64 period, runtime;
7385 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7386 runtime = tg->rt_bandwidth.rt_runtime;
7389 period = d->rt_period;
7390 runtime = d->rt_runtime;
7394 * Cannot have more runtime than the period.
7396 if (runtime > period && runtime != RUNTIME_INF)
7400 * Ensure we don't starve existing RT tasks.
7402 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7405 total = to_ratio(period, runtime);
7408 * Nobody can have more than the global setting allows.
7410 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7414 * The sum of our children's runtime should not exceed our own.
7416 list_for_each_entry_rcu(child, &tg->children, siblings) {
7417 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7418 runtime = child->rt_bandwidth.rt_runtime;
7420 if (child == d->tg) {
7421 period = d->rt_period;
7422 runtime = d->rt_runtime;
7425 sum += to_ratio(period, runtime);
7434 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7438 struct rt_schedulable_data data = {
7440 .rt_period = period,
7441 .rt_runtime = runtime,
7445 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7451 static int tg_set_rt_bandwidth(struct task_group *tg,
7452 u64 rt_period, u64 rt_runtime)
7456 mutex_lock(&rt_constraints_mutex);
7457 read_lock(&tasklist_lock);
7458 err = __rt_schedulable(tg, rt_period, rt_runtime);
7462 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7463 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7464 tg->rt_bandwidth.rt_runtime = rt_runtime;
7466 for_each_possible_cpu(i) {
7467 struct rt_rq *rt_rq = tg->rt_rq[i];
7469 raw_spin_lock(&rt_rq->rt_runtime_lock);
7470 rt_rq->rt_runtime = rt_runtime;
7471 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7473 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7475 read_unlock(&tasklist_lock);
7476 mutex_unlock(&rt_constraints_mutex);
7481 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7483 u64 rt_runtime, rt_period;
7485 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7486 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7487 if (rt_runtime_us < 0)
7488 rt_runtime = RUNTIME_INF;
7490 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7493 static long sched_group_rt_runtime(struct task_group *tg)
7497 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7500 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7501 do_div(rt_runtime_us, NSEC_PER_USEC);
7502 return rt_runtime_us;
7505 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7507 u64 rt_runtime, rt_period;
7509 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7510 rt_runtime = tg->rt_bandwidth.rt_runtime;
7515 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7518 static long sched_group_rt_period(struct task_group *tg)
7522 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7523 do_div(rt_period_us, NSEC_PER_USEC);
7524 return rt_period_us;
7527 static int sched_rt_global_constraints(void)
7529 u64 runtime, period;
7532 if (sysctl_sched_rt_period <= 0)
7535 runtime = global_rt_runtime();
7536 period = global_rt_period();
7539 * Sanity check on the sysctl variables.
7541 if (runtime > period && runtime != RUNTIME_INF)
7544 mutex_lock(&rt_constraints_mutex);
7545 read_lock(&tasklist_lock);
7546 ret = __rt_schedulable(NULL, 0, 0);
7547 read_unlock(&tasklist_lock);
7548 mutex_unlock(&rt_constraints_mutex);
7553 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7555 /* Don't accept realtime tasks when there is no way for them to run */
7556 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7562 #else /* !CONFIG_RT_GROUP_SCHED */
7563 static int sched_rt_global_constraints(void)
7565 unsigned long flags;
7568 if (sysctl_sched_rt_period <= 0)
7572 * There's always some RT tasks in the root group
7573 * -- migration, kstopmachine etc..
7575 if (sysctl_sched_rt_runtime == 0)
7578 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7579 for_each_possible_cpu(i) {
7580 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7582 raw_spin_lock(&rt_rq->rt_runtime_lock);
7583 rt_rq->rt_runtime = global_rt_runtime();
7584 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7586 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7590 #endif /* CONFIG_RT_GROUP_SCHED */
7592 int sched_rr_handler(struct ctl_table *table, int write,
7593 void __user *buffer, size_t *lenp,
7597 static DEFINE_MUTEX(mutex);
7600 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7601 /* make sure that internally we keep jiffies */
7602 /* also, writing zero resets timeslice to default */
7603 if (!ret && write) {
7604 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7605 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7607 mutex_unlock(&mutex);
7611 int sched_rt_handler(struct ctl_table *table, int write,
7612 void __user *buffer, size_t *lenp,
7616 int old_period, old_runtime;
7617 static DEFINE_MUTEX(mutex);
7620 old_period = sysctl_sched_rt_period;
7621 old_runtime = sysctl_sched_rt_runtime;
7623 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7625 if (!ret && write) {
7626 ret = sched_rt_global_constraints();
7628 sysctl_sched_rt_period = old_period;
7629 sysctl_sched_rt_runtime = old_runtime;
7631 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7632 def_rt_bandwidth.rt_period =
7633 ns_to_ktime(global_rt_period());
7636 mutex_unlock(&mutex);
7641 #ifdef CONFIG_CGROUP_SCHED
7643 /* return corresponding task_group object of a cgroup */
7644 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7646 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7647 struct task_group, css);
7650 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7652 struct task_group *tg, *parent;
7654 if (!cgrp->parent) {
7655 /* This is early initialization for the top cgroup */
7656 return &root_task_group.css;
7659 parent = cgroup_tg(cgrp->parent);
7660 tg = sched_create_group(parent);
7662 return ERR_PTR(-ENOMEM);
7667 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7669 struct task_group *tg = cgroup_tg(cgrp);
7670 struct task_group *parent;
7675 parent = cgroup_tg(cgrp->parent);
7676 sched_online_group(tg, parent);
7680 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7682 struct task_group *tg = cgroup_tg(cgrp);
7684 sched_destroy_group(tg);
7687 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7689 struct task_group *tg = cgroup_tg(cgrp);
7691 sched_offline_group(tg);
7694 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7695 struct cgroup_taskset *tset)
7697 struct task_struct *task;
7699 cgroup_taskset_for_each(task, cgrp, tset) {
7700 #ifdef CONFIG_RT_GROUP_SCHED
7701 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7704 /* We don't support RT-tasks being in separate groups */
7705 if (task->sched_class != &fair_sched_class)
7712 static void cpu_cgroup_attach(struct cgroup *cgrp,
7713 struct cgroup_taskset *tset)
7715 struct task_struct *task;
7717 cgroup_taskset_for_each(task, cgrp, tset)
7718 sched_move_task(task);
7722 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7723 struct task_struct *task)
7726 * cgroup_exit() is called in the copy_process() failure path.
7727 * Ignore this case since the task hasn't ran yet, this avoids
7728 * trying to poke a half freed task state from generic code.
7730 if (!(task->flags & PF_EXITING))
7733 sched_move_task(task);
7736 #ifdef CONFIG_FAIR_GROUP_SCHED
7737 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7740 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7743 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7745 struct task_group *tg = cgroup_tg(cgrp);
7747 return (u64) scale_load_down(tg->shares);
7750 #ifdef CONFIG_CFS_BANDWIDTH
7751 static DEFINE_MUTEX(cfs_constraints_mutex);
7753 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7754 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7756 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7758 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7760 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7761 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7763 if (tg == &root_task_group)
7767 * Ensure we have at some amount of bandwidth every period. This is
7768 * to prevent reaching a state of large arrears when throttled via
7769 * entity_tick() resulting in prolonged exit starvation.
7771 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7775 * Likewise, bound things on the otherside by preventing insane quota
7776 * periods. This also allows us to normalize in computing quota
7779 if (period > max_cfs_quota_period)
7782 mutex_lock(&cfs_constraints_mutex);
7783 ret = __cfs_schedulable(tg, period, quota);
7787 runtime_enabled = quota != RUNTIME_INF;
7788 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7789 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7790 raw_spin_lock_irq(&cfs_b->lock);
7791 cfs_b->period = ns_to_ktime(period);
7792 cfs_b->quota = quota;
7794 __refill_cfs_bandwidth_runtime(cfs_b);
7795 /* restart the period timer (if active) to handle new period expiry */
7796 if (runtime_enabled && cfs_b->timer_active) {
7797 /* force a reprogram */
7798 cfs_b->timer_active = 0;
7799 __start_cfs_bandwidth(cfs_b);
7801 raw_spin_unlock_irq(&cfs_b->lock);
7803 for_each_possible_cpu(i) {
7804 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7805 struct rq *rq = cfs_rq->rq;
7807 raw_spin_lock_irq(&rq->lock);
7808 cfs_rq->runtime_enabled = runtime_enabled;
7809 cfs_rq->runtime_remaining = 0;
7811 if (cfs_rq->throttled)
7812 unthrottle_cfs_rq(cfs_rq);
7813 raw_spin_unlock_irq(&rq->lock);
7816 mutex_unlock(&cfs_constraints_mutex);
7821 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7825 period = ktime_to_ns(tg->cfs_bandwidth.period);
7826 if (cfs_quota_us < 0)
7827 quota = RUNTIME_INF;
7829 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7831 return tg_set_cfs_bandwidth(tg, period, quota);
7834 long tg_get_cfs_quota(struct task_group *tg)
7838 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7841 quota_us = tg->cfs_bandwidth.quota;
7842 do_div(quota_us, NSEC_PER_USEC);
7847 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7851 period = (u64)cfs_period_us * NSEC_PER_USEC;
7852 quota = tg->cfs_bandwidth.quota;
7854 return tg_set_cfs_bandwidth(tg, period, quota);
7857 long tg_get_cfs_period(struct task_group *tg)
7861 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7862 do_div(cfs_period_us, NSEC_PER_USEC);
7864 return cfs_period_us;
7867 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7869 return tg_get_cfs_quota(cgroup_tg(cgrp));
7872 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7875 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7878 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7880 return tg_get_cfs_period(cgroup_tg(cgrp));
7883 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7886 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7889 struct cfs_schedulable_data {
7890 struct task_group *tg;
7895 * normalize group quota/period to be quota/max_period
7896 * note: units are usecs
7898 static u64 normalize_cfs_quota(struct task_group *tg,
7899 struct cfs_schedulable_data *d)
7907 period = tg_get_cfs_period(tg);
7908 quota = tg_get_cfs_quota(tg);
7911 /* note: these should typically be equivalent */
7912 if (quota == RUNTIME_INF || quota == -1)
7915 return to_ratio(period, quota);
7918 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7920 struct cfs_schedulable_data *d = data;
7921 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7922 s64 quota = 0, parent_quota = -1;
7925 quota = RUNTIME_INF;
7927 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7929 quota = normalize_cfs_quota(tg, d);
7930 parent_quota = parent_b->hierarchal_quota;
7933 * ensure max(child_quota) <= parent_quota, inherit when no
7936 if (quota == RUNTIME_INF)
7937 quota = parent_quota;
7938 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7941 cfs_b->hierarchal_quota = quota;
7946 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7949 struct cfs_schedulable_data data = {
7955 if (quota != RUNTIME_INF) {
7956 do_div(data.period, NSEC_PER_USEC);
7957 do_div(data.quota, NSEC_PER_USEC);
7961 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7967 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
7968 struct cgroup_map_cb *cb)
7970 struct task_group *tg = cgroup_tg(cgrp);
7971 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7973 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7974 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7975 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7979 #endif /* CONFIG_CFS_BANDWIDTH */
7980 #endif /* CONFIG_FAIR_GROUP_SCHED */
7982 #ifdef CONFIG_RT_GROUP_SCHED
7983 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
7986 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
7989 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
7991 return sched_group_rt_runtime(cgroup_tg(cgrp));
7994 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
7997 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8000 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8002 return sched_group_rt_period(cgroup_tg(cgrp));
8004 #endif /* CONFIG_RT_GROUP_SCHED */
8006 static struct cftype cpu_files[] = {
8007 #ifdef CONFIG_FAIR_GROUP_SCHED
8010 .read_u64 = cpu_shares_read_u64,
8011 .write_u64 = cpu_shares_write_u64,
8014 #ifdef CONFIG_CFS_BANDWIDTH
8016 .name = "cfs_quota_us",
8017 .read_s64 = cpu_cfs_quota_read_s64,
8018 .write_s64 = cpu_cfs_quota_write_s64,
8021 .name = "cfs_period_us",
8022 .read_u64 = cpu_cfs_period_read_u64,
8023 .write_u64 = cpu_cfs_period_write_u64,
8027 .read_map = cpu_stats_show,
8030 #ifdef CONFIG_RT_GROUP_SCHED
8032 .name = "rt_runtime_us",
8033 .read_s64 = cpu_rt_runtime_read,
8034 .write_s64 = cpu_rt_runtime_write,
8037 .name = "rt_period_us",
8038 .read_u64 = cpu_rt_period_read_uint,
8039 .write_u64 = cpu_rt_period_write_uint,
8045 struct cgroup_subsys cpu_cgroup_subsys = {
8047 .css_alloc = cpu_cgroup_css_alloc,
8048 .css_free = cpu_cgroup_css_free,
8049 .css_online = cpu_cgroup_css_online,
8050 .css_offline = cpu_cgroup_css_offline,
8051 .can_attach = cpu_cgroup_can_attach,
8052 .attach = cpu_cgroup_attach,
8053 .exit = cpu_cgroup_exit,
8054 .subsys_id = cpu_cgroup_subsys_id,
8055 .base_cftypes = cpu_files,
8059 #endif /* CONFIG_CGROUP_SCHED */
8061 void dump_cpu_task(int cpu)
8063 pr_info("Task dump for CPU %d:\n", cpu);
8064 sched_show_task(cpu_curr(cpu));