4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
302 * __task_rq_lock - lock the rq @p resides on.
304 static inline struct rq *__task_rq_lock(struct task_struct *p)
309 lockdep_assert_held(&p->pi_lock);
313 raw_spin_lock(&rq->lock);
314 if (likely(rq == task_rq(p)))
316 raw_spin_unlock(&rq->lock);
321 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
323 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
324 __acquires(p->pi_lock)
330 raw_spin_lock_irqsave(&p->pi_lock, *flags);
332 raw_spin_lock(&rq->lock);
333 if (likely(rq == task_rq(p)))
335 raw_spin_unlock(&rq->lock);
336 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
340 static void __task_rq_unlock(struct rq *rq)
343 raw_spin_unlock(&rq->lock);
347 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
349 __releases(p->pi_lock)
351 raw_spin_unlock(&rq->lock);
352 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
356 * this_rq_lock - lock this runqueue and disable interrupts.
358 static struct rq *this_rq_lock(void)
365 raw_spin_lock(&rq->lock);
370 #ifdef CONFIG_SCHED_HRTICK
372 * Use HR-timers to deliver accurate preemption points.
374 * Its all a bit involved since we cannot program an hrt while holding the
375 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
378 * When we get rescheduled we reprogram the hrtick_timer outside of the
382 static void hrtick_clear(struct rq *rq)
384 if (hrtimer_active(&rq->hrtick_timer))
385 hrtimer_cancel(&rq->hrtick_timer);
389 * High-resolution timer tick.
390 * Runs from hardirq context with interrupts disabled.
392 static enum hrtimer_restart hrtick(struct hrtimer *timer)
394 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
396 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
398 raw_spin_lock(&rq->lock);
400 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
401 raw_spin_unlock(&rq->lock);
403 return HRTIMER_NORESTART;
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 hrtimer_restart(&rq->hrtick_timer);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 hrtimer_restart(timer);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 assert_raw_spin_locked(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id())
530 /* NEED_RESCHED must be visible before we test polling */
532 if (!tsk_is_polling(p))
533 smp_send_reschedule(cpu);
536 void resched_cpu(int cpu)
538 struct rq *rq = cpu_rq(cpu);
541 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
543 resched_task(cpu_curr(cpu));
544 raw_spin_unlock_irqrestore(&rq->lock, flags);
547 #ifdef CONFIG_NO_HZ_COMMON
549 * In the semi idle case, use the nearest busy cpu for migrating timers
550 * from an idle cpu. This is good for power-savings.
552 * We don't do similar optimization for completely idle system, as
553 * selecting an idle cpu will add more delays to the timers than intended
554 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
556 int get_nohz_timer_target(void)
558 int cpu = smp_processor_id();
560 struct sched_domain *sd;
563 for_each_domain(cpu, sd) {
564 for_each_cpu(i, sched_domain_span(sd)) {
576 * When add_timer_on() enqueues a timer into the timer wheel of an
577 * idle CPU then this timer might expire before the next timer event
578 * which is scheduled to wake up that CPU. In case of a completely
579 * idle system the next event might even be infinite time into the
580 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
581 * leaves the inner idle loop so the newly added timer is taken into
582 * account when the CPU goes back to idle and evaluates the timer
583 * wheel for the next timer event.
585 static void wake_up_idle_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
589 if (cpu == smp_processor_id())
593 * This is safe, as this function is called with the timer
594 * wheel base lock of (cpu) held. When the CPU is on the way
595 * to idle and has not yet set rq->curr to idle then it will
596 * be serialized on the timer wheel base lock and take the new
597 * timer into account automatically.
599 if (rq->curr != rq->idle)
603 * We can set TIF_RESCHED on the idle task of the other CPU
604 * lockless. The worst case is that the other CPU runs the
605 * idle task through an additional NOOP schedule()
607 set_tsk_need_resched(rq->idle);
609 /* NEED_RESCHED must be visible before we test polling */
611 if (!tsk_is_polling(rq->idle))
612 smp_send_reschedule(cpu);
615 static bool wake_up_full_nohz_cpu(int cpu)
617 if (tick_nohz_full_cpu(cpu)) {
618 if (cpu != smp_processor_id() ||
619 tick_nohz_tick_stopped())
620 smp_send_reschedule(cpu);
627 void wake_up_nohz_cpu(int cpu)
629 if (!wake_up_full_nohz_cpu(cpu))
630 wake_up_idle_cpu(cpu);
633 static inline bool got_nohz_idle_kick(void)
635 int cpu = smp_processor_id();
637 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
640 if (idle_cpu(cpu) && !need_resched())
644 * We can't run Idle Load Balance on this CPU for this time so we
645 * cancel it and clear NOHZ_BALANCE_KICK
647 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
651 #else /* CONFIG_NO_HZ_COMMON */
653 static inline bool got_nohz_idle_kick(void)
658 #endif /* CONFIG_NO_HZ_COMMON */
660 #ifdef CONFIG_NO_HZ_FULL
661 bool sched_can_stop_tick(void)
667 /* Make sure rq->nr_running update is visible after the IPI */
670 /* More than one running task need preemption */
671 if (rq->nr_running > 1)
676 #endif /* CONFIG_NO_HZ_FULL */
678 void sched_avg_update(struct rq *rq)
680 s64 period = sched_avg_period();
682 while ((s64)(rq->clock - rq->age_stamp) > period) {
684 * Inline assembly required to prevent the compiler
685 * optimising this loop into a divmod call.
686 * See __iter_div_u64_rem() for another example of this.
688 asm("" : "+rm" (rq->age_stamp));
689 rq->age_stamp += period;
694 #else /* !CONFIG_SMP */
695 void resched_task(struct task_struct *p)
697 assert_raw_spin_locked(&task_rq(p)->lock);
698 set_tsk_need_resched(p);
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
713 struct task_group *parent, *child;
719 ret = (*down)(parent, data);
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
734 parent = parent->parent;
741 int tg_nop(struct task_group *tg, void *data)
747 static void set_load_weight(struct task_struct *p)
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p->policy == SCHED_IDLE) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
761 load->weight = scale_load(prio_to_weight[prio]);
762 load->inv_weight = prio_to_wmult[prio];
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
768 sched_info_queued(p);
769 p->sched_class->enqueue_task(rq, p, flags);
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
775 sched_info_dequeued(p);
776 p->sched_class->dequeue_task(rq, p, flags);
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible--;
784 enqueue_task(rq, p, flags);
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
789 if (task_contributes_to_load(p))
790 rq->nr_uninterruptible++;
792 dequeue_task(rq, p, flags);
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal = 0, irq_delta = 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta > delta)
825 rq->prev_irq_time += irq_delta;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled))) {
832 steal = paravirt_steal_clock(cpu_of(rq));
833 steal -= rq->prev_steal_time_rq;
835 if (unlikely(steal > delta))
838 st = steal_ticks(steal);
839 steal = st * TICK_NSEC;
841 rq->prev_steal_time_rq += steal;
847 rq->clock_task += delta;
849 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
850 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
851 sched_rt_avg_update(rq, irq_delta + steal);
855 void sched_set_stop_task(int cpu, struct task_struct *stop)
857 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
858 struct task_struct *old_stop = cpu_rq(cpu)->stop;
862 * Make it appear like a SCHED_FIFO task, its something
863 * userspace knows about and won't get confused about.
865 * Also, it will make PI more or less work without too
866 * much confusion -- but then, stop work should not
867 * rely on PI working anyway.
869 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
871 stop->sched_class = &stop_sched_class;
874 cpu_rq(cpu)->stop = stop;
878 * Reset it back to a normal scheduling class so that
879 * it can die in pieces.
881 old_stop->sched_class = &rt_sched_class;
886 * __normal_prio - return the priority that is based on the static prio
888 static inline int __normal_prio(struct task_struct *p)
890 return p->static_prio;
894 * Calculate the expected normal priority: i.e. priority
895 * without taking RT-inheritance into account. Might be
896 * boosted by interactivity modifiers. Changes upon fork,
897 * setprio syscalls, and whenever the interactivity
898 * estimator recalculates.
900 static inline int normal_prio(struct task_struct *p)
904 if (task_has_rt_policy(p))
905 prio = MAX_RT_PRIO-1 - p->rt_priority;
907 prio = __normal_prio(p);
912 * Calculate the current priority, i.e. the priority
913 * taken into account by the scheduler. This value might
914 * be boosted by RT tasks, or might be boosted by
915 * interactivity modifiers. Will be RT if the task got
916 * RT-boosted. If not then it returns p->normal_prio.
918 static int effective_prio(struct task_struct *p)
920 p->normal_prio = normal_prio(p);
922 * If we are RT tasks or we were boosted to RT priority,
923 * keep the priority unchanged. Otherwise, update priority
924 * to the normal priority:
926 if (!rt_prio(p->prio))
927 return p->normal_prio;
932 * task_curr - is this task currently executing on a CPU?
933 * @p: the task in question.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio)
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
977 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
979 void register_task_migration_notifier(struct notifier_block *n)
981 atomic_notifier_chain_register(&task_migration_notifier, n);
985 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
987 #ifdef CONFIG_SCHED_DEBUG
989 * We should never call set_task_cpu() on a blocked task,
990 * ttwu() will sort out the placement.
992 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
993 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
995 #ifdef CONFIG_LOCKDEP
997 * The caller should hold either p->pi_lock or rq->lock, when changing
998 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1000 * sched_move_task() holds both and thus holding either pins the cgroup,
1003 * Furthermore, all task_rq users should acquire both locks, see
1006 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1007 lockdep_is_held(&task_rq(p)->lock)));
1011 trace_sched_migrate_task(p, new_cpu);
1013 if (task_cpu(p) != new_cpu) {
1014 struct task_migration_notifier tmn;
1016 if (p->sched_class->migrate_task_rq)
1017 p->sched_class->migrate_task_rq(p, new_cpu);
1018 p->se.nr_migrations++;
1019 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1022 tmn.from_cpu = task_cpu(p);
1023 tmn.to_cpu = new_cpu;
1025 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1028 __set_task_cpu(p, new_cpu);
1031 struct migration_arg {
1032 struct task_struct *task;
1036 static int migration_cpu_stop(void *data);
1039 * wait_task_inactive - wait for a thread to unschedule.
1041 * If @match_state is nonzero, it's the @p->state value just checked and
1042 * not expected to change. If it changes, i.e. @p might have woken up,
1043 * then return zero. When we succeed in waiting for @p to be off its CPU,
1044 * we return a positive number (its total switch count). If a second call
1045 * a short while later returns the same number, the caller can be sure that
1046 * @p has remained unscheduled the whole time.
1048 * The caller must ensure that the task *will* unschedule sometime soon,
1049 * else this function might spin for a *long* time. This function can't
1050 * be called with interrupts off, or it may introduce deadlock with
1051 * smp_call_function() if an IPI is sent by the same process we are
1052 * waiting to become inactive.
1054 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1056 unsigned long flags;
1063 * We do the initial early heuristics without holding
1064 * any task-queue locks at all. We'll only try to get
1065 * the runqueue lock when things look like they will
1071 * If the task is actively running on another CPU
1072 * still, just relax and busy-wait without holding
1075 * NOTE! Since we don't hold any locks, it's not
1076 * even sure that "rq" stays as the right runqueue!
1077 * But we don't care, since "task_running()" will
1078 * return false if the runqueue has changed and p
1079 * is actually now running somewhere else!
1081 while (task_running(rq, p)) {
1082 if (match_state && unlikely(p->state != match_state))
1088 * Ok, time to look more closely! We need the rq
1089 * lock now, to be *sure*. If we're wrong, we'll
1090 * just go back and repeat.
1092 rq = task_rq_lock(p, &flags);
1093 trace_sched_wait_task(p);
1094 running = task_running(rq, p);
1097 if (!match_state || p->state == match_state)
1098 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1099 task_rq_unlock(rq, p, &flags);
1102 * If it changed from the expected state, bail out now.
1104 if (unlikely(!ncsw))
1108 * Was it really running after all now that we
1109 * checked with the proper locks actually held?
1111 * Oops. Go back and try again..
1113 if (unlikely(running)) {
1119 * It's not enough that it's not actively running,
1120 * it must be off the runqueue _entirely_, and not
1123 * So if it was still runnable (but just not actively
1124 * running right now), it's preempted, and we should
1125 * yield - it could be a while.
1127 if (unlikely(on_rq)) {
1128 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1130 set_current_state(TASK_UNINTERRUPTIBLE);
1131 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1136 * Ahh, all good. It wasn't running, and it wasn't
1137 * runnable, which means that it will never become
1138 * running in the future either. We're all done!
1147 * kick_process - kick a running thread to enter/exit the kernel
1148 * @p: the to-be-kicked thread
1150 * Cause a process which is running on another CPU to enter
1151 * kernel-mode, without any delay. (to get signals handled.)
1153 * NOTE: this function doesn't have to take the runqueue lock,
1154 * because all it wants to ensure is that the remote task enters
1155 * the kernel. If the IPI races and the task has been migrated
1156 * to another CPU then no harm is done and the purpose has been
1159 void kick_process(struct task_struct *p)
1165 if ((cpu != smp_processor_id()) && task_curr(p))
1166 smp_send_reschedule(cpu);
1169 EXPORT_SYMBOL_GPL(kick_process);
1170 #endif /* CONFIG_SMP */
1174 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1176 static int select_fallback_rq(int cpu, struct task_struct *p)
1178 int nid = cpu_to_node(cpu);
1179 const struct cpumask *nodemask = NULL;
1180 enum { cpuset, possible, fail } state = cpuset;
1184 * If the node that the cpu is on has been offlined, cpu_to_node()
1185 * will return -1. There is no cpu on the node, and we should
1186 * select the cpu on the other node.
1189 nodemask = cpumask_of_node(nid);
1191 /* Look for allowed, online CPU in same node. */
1192 for_each_cpu(dest_cpu, nodemask) {
1193 if (!cpu_online(dest_cpu))
1195 if (!cpu_active(dest_cpu))
1197 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1203 /* Any allowed, online CPU? */
1204 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1205 if (!cpu_online(dest_cpu))
1207 if (!cpu_active(dest_cpu))
1214 /* No more Mr. Nice Guy. */
1215 cpuset_cpus_allowed_fallback(p);
1220 do_set_cpus_allowed(p, cpu_possible_mask);
1231 if (state != cpuset) {
1233 * Don't tell them about moving exiting tasks or
1234 * kernel threads (both mm NULL), since they never
1237 if (p->mm && printk_ratelimit()) {
1238 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1239 task_pid_nr(p), p->comm, cpu);
1247 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1250 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
1252 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
1255 * In order not to call set_task_cpu() on a blocking task we need
1256 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1259 * Since this is common to all placement strategies, this lives here.
1261 * [ this allows ->select_task() to simply return task_cpu(p) and
1262 * not worry about this generic constraint ]
1264 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1266 cpu = select_fallback_rq(task_cpu(p), p);
1271 static void update_avg(u64 *avg, u64 sample)
1273 s64 diff = sample - *avg;
1279 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1281 #ifdef CONFIG_SCHEDSTATS
1282 struct rq *rq = this_rq();
1285 int this_cpu = smp_processor_id();
1287 if (cpu == this_cpu) {
1288 schedstat_inc(rq, ttwu_local);
1289 schedstat_inc(p, se.statistics.nr_wakeups_local);
1291 struct sched_domain *sd;
1293 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1295 for_each_domain(this_cpu, sd) {
1296 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1297 schedstat_inc(sd, ttwu_wake_remote);
1304 if (wake_flags & WF_MIGRATED)
1305 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1307 #endif /* CONFIG_SMP */
1309 schedstat_inc(rq, ttwu_count);
1310 schedstat_inc(p, se.statistics.nr_wakeups);
1312 if (wake_flags & WF_SYNC)
1313 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1315 #endif /* CONFIG_SCHEDSTATS */
1318 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1320 activate_task(rq, p, en_flags);
1323 /* if a worker is waking up, notify workqueue */
1324 if (p->flags & PF_WQ_WORKER)
1325 wq_worker_waking_up(p, cpu_of(rq));
1329 * Mark the task runnable and perform wakeup-preemption.
1332 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1334 check_preempt_curr(rq, p, wake_flags);
1335 trace_sched_wakeup(p, true);
1337 p->state = TASK_RUNNING;
1339 if (p->sched_class->task_woken)
1340 p->sched_class->task_woken(rq, p);
1342 if (rq->idle_stamp) {
1343 u64 delta = rq->clock - rq->idle_stamp;
1344 u64 max = 2*sysctl_sched_migration_cost;
1349 update_avg(&rq->avg_idle, delta);
1356 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1359 if (p->sched_contributes_to_load)
1360 rq->nr_uninterruptible--;
1363 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1364 ttwu_do_wakeup(rq, p, wake_flags);
1368 * Called in case the task @p isn't fully descheduled from its runqueue,
1369 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1370 * since all we need to do is flip p->state to TASK_RUNNING, since
1371 * the task is still ->on_rq.
1373 static int ttwu_remote(struct task_struct *p, int wake_flags)
1378 rq = __task_rq_lock(p);
1380 ttwu_do_wakeup(rq, p, wake_flags);
1383 __task_rq_unlock(rq);
1389 static void sched_ttwu_pending(void)
1391 struct rq *rq = this_rq();
1392 struct llist_node *llist = llist_del_all(&rq->wake_list);
1393 struct task_struct *p;
1395 raw_spin_lock(&rq->lock);
1398 p = llist_entry(llist, struct task_struct, wake_entry);
1399 llist = llist_next(llist);
1400 ttwu_do_activate(rq, p, 0);
1403 raw_spin_unlock(&rq->lock);
1406 void scheduler_ipi(void)
1408 if (llist_empty(&this_rq()->wake_list)
1409 && !tick_nohz_full_cpu(smp_processor_id())
1410 && !got_nohz_idle_kick()
1411 #ifdef CONFIG_SCHED_HMP
1412 && !this_rq()->wake_for_idle_pull
1418 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1419 * traditionally all their work was done from the interrupt return
1420 * path. Now that we actually do some work, we need to make sure
1423 * Some archs already do call them, luckily irq_enter/exit nest
1426 * Arguably we should visit all archs and update all handlers,
1427 * however a fair share of IPIs are still resched only so this would
1428 * somewhat pessimize the simple resched case.
1431 tick_nohz_full_check();
1432 sched_ttwu_pending();
1435 * Check if someone kicked us for doing the nohz idle load balance.
1437 if (unlikely(got_nohz_idle_kick())) {
1438 this_rq()->idle_balance = 1;
1439 raise_softirq_irqoff(SCHED_SOFTIRQ);
1441 #ifdef CONFIG_SCHED_HMP
1442 else if (unlikely(this_rq()->wake_for_idle_pull))
1443 raise_softirq_irqoff(SCHED_SOFTIRQ);
1449 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1451 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1452 smp_send_reschedule(cpu);
1455 bool cpus_share_cache(int this_cpu, int that_cpu)
1457 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1459 #endif /* CONFIG_SMP */
1461 static void ttwu_queue(struct task_struct *p, int cpu)
1463 struct rq *rq = cpu_rq(cpu);
1465 #if defined(CONFIG_SMP)
1466 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1467 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1468 ttwu_queue_remote(p, cpu);
1473 raw_spin_lock(&rq->lock);
1474 ttwu_do_activate(rq, p, 0);
1475 raw_spin_unlock(&rq->lock);
1479 * try_to_wake_up - wake up a thread
1480 * @p: the thread to be awakened
1481 * @state: the mask of task states that can be woken
1482 * @wake_flags: wake modifier flags (WF_*)
1484 * Put it on the run-queue if it's not already there. The "current"
1485 * thread is always on the run-queue (except when the actual
1486 * re-schedule is in progress), and as such you're allowed to do
1487 * the simpler "current->state = TASK_RUNNING" to mark yourself
1488 * runnable without the overhead of this.
1490 * Returns %true if @p was woken up, %false if it was already running
1491 * or @state didn't match @p's state.
1494 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1496 unsigned long flags;
1497 int cpu, success = 0;
1500 raw_spin_lock_irqsave(&p->pi_lock, flags);
1501 if (!(p->state & state))
1504 success = 1; /* we're going to change ->state */
1507 if (p->on_rq && ttwu_remote(p, wake_flags))
1512 * If the owning (remote) cpu is still in the middle of schedule() with
1513 * this task as prev, wait until its done referencing the task.
1518 * Pairs with the smp_wmb() in finish_lock_switch().
1522 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1523 p->state = TASK_WAKING;
1525 if (p->sched_class->task_waking)
1526 p->sched_class->task_waking(p);
1528 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
1529 if (task_cpu(p) != cpu) {
1530 wake_flags |= WF_MIGRATED;
1531 set_task_cpu(p, cpu);
1533 #endif /* CONFIG_SMP */
1537 ttwu_stat(p, cpu, wake_flags);
1539 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1545 * try_to_wake_up_local - try to wake up a local task with rq lock held
1546 * @p: the thread to be awakened
1548 * Put @p on the run-queue if it's not already there. The caller must
1549 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1552 static void try_to_wake_up_local(struct task_struct *p)
1554 struct rq *rq = task_rq(p);
1556 if (WARN_ON_ONCE(rq != this_rq()) ||
1557 WARN_ON_ONCE(p == current))
1560 lockdep_assert_held(&rq->lock);
1562 if (!raw_spin_trylock(&p->pi_lock)) {
1563 raw_spin_unlock(&rq->lock);
1564 raw_spin_lock(&p->pi_lock);
1565 raw_spin_lock(&rq->lock);
1568 if (!(p->state & TASK_NORMAL))
1572 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1574 ttwu_do_wakeup(rq, p, 0);
1575 ttwu_stat(p, smp_processor_id(), 0);
1577 raw_spin_unlock(&p->pi_lock);
1581 * wake_up_process - Wake up a specific process
1582 * @p: The process to be woken up.
1584 * Attempt to wake up the nominated process and move it to the set of runnable
1585 * processes. Returns 1 if the process was woken up, 0 if it was already
1588 * It may be assumed that this function implies a write memory barrier before
1589 * changing the task state if and only if any tasks are woken up.
1591 int wake_up_process(struct task_struct *p)
1593 WARN_ON(task_is_stopped_or_traced(p));
1594 return try_to_wake_up(p, TASK_NORMAL, 0);
1596 EXPORT_SYMBOL(wake_up_process);
1598 int wake_up_state(struct task_struct *p, unsigned int state)
1600 return try_to_wake_up(p, state, 0);
1604 * Perform scheduler related setup for a newly forked process p.
1605 * p is forked by current.
1607 * __sched_fork() is basic setup used by init_idle() too:
1609 static void __sched_fork(struct task_struct *p)
1614 p->se.exec_start = 0;
1615 p->se.sum_exec_runtime = 0;
1616 p->se.prev_sum_exec_runtime = 0;
1617 p->se.nr_migrations = 0;
1619 INIT_LIST_HEAD(&p->se.group_node);
1622 * Load-tracking only depends on SMP, FAIR_GROUP_SCHED dependency below may be
1623 * removed when useful for applications beyond shares distribution (e.g.
1626 #if defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)
1627 p->se.avg.runnable_avg_period = 0;
1628 p->se.avg.runnable_avg_sum = 0;
1629 #ifdef CONFIG_SCHED_HMP
1630 /* keep LOAD_AVG_MAX in sync with fair.c if load avg series is changed */
1631 #define LOAD_AVG_MAX 47742
1632 p->se.avg.hmp_last_up_migration = 0;
1633 p->se.avg.hmp_last_down_migration = 0;
1634 if (hmp_task_should_forkboost(p)) {
1635 p->se.avg.load_avg_ratio = 1023;
1636 p->se.avg.load_avg_contrib =
1637 (1023 * scale_load_down(p->se.load.weight));
1638 p->se.avg.runnable_avg_period = LOAD_AVG_MAX;
1639 p->se.avg.runnable_avg_sum = LOAD_AVG_MAX;
1640 p->se.avg.usage_avg_sum = LOAD_AVG_MAX;
1644 #ifdef CONFIG_SCHEDSTATS
1645 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1648 INIT_LIST_HEAD(&p->rt.run_list);
1650 #ifdef CONFIG_PREEMPT_NOTIFIERS
1651 INIT_HLIST_HEAD(&p->preempt_notifiers);
1654 #ifdef CONFIG_NUMA_BALANCING
1655 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1656 p->mm->numa_next_scan = jiffies;
1657 p->mm->numa_next_reset = jiffies;
1658 p->mm->numa_scan_seq = 0;
1661 p->node_stamp = 0ULL;
1662 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1663 p->numa_migrate_seq = p->mm ? p->mm->numa_scan_seq - 1 : 0;
1664 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1665 p->numa_work.next = &p->numa_work;
1666 #endif /* CONFIG_NUMA_BALANCING */
1669 #ifdef CONFIG_NUMA_BALANCING
1670 #ifdef CONFIG_SCHED_DEBUG
1671 void set_numabalancing_state(bool enabled)
1674 sched_feat_set("NUMA");
1676 sched_feat_set("NO_NUMA");
1679 __read_mostly bool numabalancing_enabled;
1681 void set_numabalancing_state(bool enabled)
1683 numabalancing_enabled = enabled;
1685 #endif /* CONFIG_SCHED_DEBUG */
1686 #endif /* CONFIG_NUMA_BALANCING */
1689 * fork()/clone()-time setup:
1691 void sched_fork(struct task_struct *p)
1693 unsigned long flags;
1694 int cpu = get_cpu();
1698 * We mark the process as running here. This guarantees that
1699 * nobody will actually run it, and a signal or other external
1700 * event cannot wake it up and insert it on the runqueue either.
1702 p->state = TASK_RUNNING;
1705 * Make sure we do not leak PI boosting priority to the child.
1707 p->prio = current->normal_prio;
1710 * Revert to default priority/policy on fork if requested.
1712 if (unlikely(p->sched_reset_on_fork)) {
1713 if (task_has_rt_policy(p)) {
1714 p->policy = SCHED_NORMAL;
1715 p->static_prio = NICE_TO_PRIO(0);
1717 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1718 p->static_prio = NICE_TO_PRIO(0);
1720 p->prio = p->normal_prio = __normal_prio(p);
1724 * We don't need the reset flag anymore after the fork. It has
1725 * fulfilled its duty:
1727 p->sched_reset_on_fork = 0;
1730 if (!rt_prio(p->prio))
1731 p->sched_class = &fair_sched_class;
1733 if (p->sched_class->task_fork)
1734 p->sched_class->task_fork(p);
1737 * The child is not yet in the pid-hash so no cgroup attach races,
1738 * and the cgroup is pinned to this child due to cgroup_fork()
1739 * is ran before sched_fork().
1741 * Silence PROVE_RCU.
1743 raw_spin_lock_irqsave(&p->pi_lock, flags);
1744 set_task_cpu(p, cpu);
1745 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1747 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1748 if (likely(sched_info_on()))
1749 memset(&p->sched_info, 0, sizeof(p->sched_info));
1751 #if defined(CONFIG_SMP)
1754 #ifdef CONFIG_PREEMPT_COUNT
1755 /* Want to start with kernel preemption disabled. */
1756 task_thread_info(p)->preempt_count = 1;
1759 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1766 * wake_up_new_task - wake up a newly created task for the first time.
1768 * This function will do some initial scheduler statistics housekeeping
1769 * that must be done for every newly created context, then puts the task
1770 * on the runqueue and wakes it.
1772 void wake_up_new_task(struct task_struct *p)
1774 unsigned long flags;
1777 raw_spin_lock_irqsave(&p->pi_lock, flags);
1780 * Fork balancing, do it here and not earlier because:
1781 * - cpus_allowed can change in the fork path
1782 * - any previously selected cpu might disappear through hotplug
1784 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
1787 rq = __task_rq_lock(p);
1788 activate_task(rq, p, 0);
1790 trace_sched_wakeup_new(p, true);
1791 check_preempt_curr(rq, p, WF_FORK);
1793 if (p->sched_class->task_woken)
1794 p->sched_class->task_woken(rq, p);
1796 task_rq_unlock(rq, p, &flags);
1799 #ifdef CONFIG_PREEMPT_NOTIFIERS
1802 * preempt_notifier_register - tell me when current is being preempted & rescheduled
1803 * @notifier: notifier struct to register
1805 void preempt_notifier_register(struct preempt_notifier *notifier)
1807 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
1809 EXPORT_SYMBOL_GPL(preempt_notifier_register);
1812 * preempt_notifier_unregister - no longer interested in preemption notifications
1813 * @notifier: notifier struct to unregister
1815 * This is safe to call from within a preemption notifier.
1817 void preempt_notifier_unregister(struct preempt_notifier *notifier)
1819 hlist_del(¬ifier->link);
1821 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
1823 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1825 struct preempt_notifier *notifier;
1827 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1828 notifier->ops->sched_in(notifier, raw_smp_processor_id());
1832 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1833 struct task_struct *next)
1835 struct preempt_notifier *notifier;
1837 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
1838 notifier->ops->sched_out(notifier, next);
1841 #else /* !CONFIG_PREEMPT_NOTIFIERS */
1843 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
1848 fire_sched_out_preempt_notifiers(struct task_struct *curr,
1849 struct task_struct *next)
1853 #endif /* CONFIG_PREEMPT_NOTIFIERS */
1856 * prepare_task_switch - prepare to switch tasks
1857 * @rq: the runqueue preparing to switch
1858 * @prev: the current task that is being switched out
1859 * @next: the task we are going to switch to.
1861 * This is called with the rq lock held and interrupts off. It must
1862 * be paired with a subsequent finish_task_switch after the context
1865 * prepare_task_switch sets up locking and calls architecture specific
1869 prepare_task_switch(struct rq *rq, struct task_struct *prev,
1870 struct task_struct *next)
1872 trace_sched_switch(prev, next);
1873 sched_info_switch(prev, next);
1874 perf_event_task_sched_out(prev, next);
1875 fire_sched_out_preempt_notifiers(prev, next);
1876 prepare_lock_switch(rq, next);
1877 prepare_arch_switch(next);
1881 * finish_task_switch - clean up after a task-switch
1882 * @rq: runqueue associated with task-switch
1883 * @prev: the thread we just switched away from.
1885 * finish_task_switch must be called after the context switch, paired
1886 * with a prepare_task_switch call before the context switch.
1887 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1888 * and do any other architecture-specific cleanup actions.
1890 * Note that we may have delayed dropping an mm in context_switch(). If
1891 * so, we finish that here outside of the runqueue lock. (Doing it
1892 * with the lock held can cause deadlocks; see schedule() for
1895 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
1896 __releases(rq->lock)
1898 struct mm_struct *mm = rq->prev_mm;
1904 * A task struct has one reference for the use as "current".
1905 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1906 * schedule one last time. The schedule call will never return, and
1907 * the scheduled task must drop that reference.
1908 * The test for TASK_DEAD must occur while the runqueue locks are
1909 * still held, otherwise prev could be scheduled on another cpu, die
1910 * there before we look at prev->state, and then the reference would
1912 * Manfred Spraul <manfred@colorfullife.com>
1914 prev_state = prev->state;
1915 vtime_task_switch(prev);
1916 finish_arch_switch(prev);
1917 perf_event_task_sched_in(prev, current);
1918 finish_lock_switch(rq, prev);
1919 finish_arch_post_lock_switch();
1921 fire_sched_in_preempt_notifiers(current);
1924 if (unlikely(prev_state == TASK_DEAD)) {
1926 * Remove function-return probe instances associated with this
1927 * task and put them back on the free list.
1929 kprobe_flush_task(prev);
1930 put_task_struct(prev);
1933 tick_nohz_task_switch(current);
1938 /* assumes rq->lock is held */
1939 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
1941 if (prev->sched_class->pre_schedule)
1942 prev->sched_class->pre_schedule(rq, prev);
1945 /* rq->lock is NOT held, but preemption is disabled */
1946 static inline void post_schedule(struct rq *rq)
1948 if (rq->post_schedule) {
1949 unsigned long flags;
1951 raw_spin_lock_irqsave(&rq->lock, flags);
1952 if (rq->curr->sched_class->post_schedule)
1953 rq->curr->sched_class->post_schedule(rq);
1954 raw_spin_unlock_irqrestore(&rq->lock, flags);
1956 rq->post_schedule = 0;
1962 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
1966 static inline void post_schedule(struct rq *rq)
1973 * schedule_tail - first thing a freshly forked thread must call.
1974 * @prev: the thread we just switched away from.
1976 asmlinkage void schedule_tail(struct task_struct *prev)
1977 __releases(rq->lock)
1979 struct rq *rq = this_rq();
1981 finish_task_switch(rq, prev);
1984 * FIXME: do we need to worry about rq being invalidated by the
1989 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1990 /* In this case, finish_task_switch does not reenable preemption */
1993 if (current->set_child_tid)
1994 put_user(task_pid_vnr(current), current->set_child_tid);
1998 * context_switch - switch to the new MM and the new
1999 * thread's register state.
2002 context_switch(struct rq *rq, struct task_struct *prev,
2003 struct task_struct *next)
2005 struct mm_struct *mm, *oldmm;
2007 prepare_task_switch(rq, prev, next);
2010 oldmm = prev->active_mm;
2012 * For paravirt, this is coupled with an exit in switch_to to
2013 * combine the page table reload and the switch backend into
2016 arch_start_context_switch(prev);
2019 next->active_mm = oldmm;
2020 atomic_inc(&oldmm->mm_count);
2021 enter_lazy_tlb(oldmm, next);
2023 switch_mm(oldmm, mm, next);
2026 prev->active_mm = NULL;
2027 rq->prev_mm = oldmm;
2030 * Since the runqueue lock will be released by the next
2031 * task (which is an invalid locking op but in the case
2032 * of the scheduler it's an obvious special-case), so we
2033 * do an early lockdep release here:
2035 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2036 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2039 context_tracking_task_switch(prev, next);
2040 /* Here we just switch the register state and the stack. */
2041 switch_to(prev, next, prev);
2045 * this_rq must be evaluated again because prev may have moved
2046 * CPUs since it called schedule(), thus the 'rq' on its stack
2047 * frame will be invalid.
2049 finish_task_switch(this_rq(), prev);
2053 * nr_running and nr_context_switches:
2055 * externally visible scheduler statistics: current number of runnable
2056 * threads, total number of context switches performed since bootup.
2058 unsigned long nr_running(void)
2060 unsigned long i, sum = 0;
2062 for_each_online_cpu(i)
2063 sum += cpu_rq(i)->nr_running;
2068 unsigned long long nr_context_switches(void)
2071 unsigned long long sum = 0;
2073 for_each_possible_cpu(i)
2074 sum += cpu_rq(i)->nr_switches;
2079 unsigned long nr_iowait(void)
2081 unsigned long i, sum = 0;
2083 for_each_possible_cpu(i)
2084 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2089 unsigned long nr_iowait_cpu(int cpu)
2091 struct rq *this = cpu_rq(cpu);
2092 return atomic_read(&this->nr_iowait);
2095 unsigned long this_cpu_load(void)
2097 struct rq *this = this_rq();
2098 return this->cpu_load[0];
2103 * Global load-average calculations
2105 * We take a distributed and async approach to calculating the global load-avg
2106 * in order to minimize overhead.
2108 * The global load average is an exponentially decaying average of nr_running +
2109 * nr_uninterruptible.
2111 * Once every LOAD_FREQ:
2114 * for_each_possible_cpu(cpu)
2115 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
2117 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
2119 * Due to a number of reasons the above turns in the mess below:
2121 * - for_each_possible_cpu() is prohibitively expensive on machines with
2122 * serious number of cpus, therefore we need to take a distributed approach
2123 * to calculating nr_active.
2125 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
2126 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
2128 * So assuming nr_active := 0 when we start out -- true per definition, we
2129 * can simply take per-cpu deltas and fold those into a global accumulate
2130 * to obtain the same result. See calc_load_fold_active().
2132 * Furthermore, in order to avoid synchronizing all per-cpu delta folding
2133 * across the machine, we assume 10 ticks is sufficient time for every
2134 * cpu to have completed this task.
2136 * This places an upper-bound on the IRQ-off latency of the machine. Then
2137 * again, being late doesn't loose the delta, just wrecks the sample.
2139 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because
2140 * this would add another cross-cpu cacheline miss and atomic operation
2141 * to the wakeup path. Instead we increment on whatever cpu the task ran
2142 * when it went into uninterruptible state and decrement on whatever cpu
2143 * did the wakeup. This means that only the sum of nr_uninterruptible over
2144 * all cpus yields the correct result.
2146 * This covers the NO_HZ=n code, for extra head-aches, see the comment below.
2149 /* Variables and functions for calc_load */
2150 static atomic_long_t calc_load_tasks;
2151 static unsigned long calc_load_update;
2152 unsigned long avenrun[3];
2153 EXPORT_SYMBOL(avenrun); /* should be removed */
2156 * get_avenrun - get the load average array
2157 * @loads: pointer to dest load array
2158 * @offset: offset to add
2159 * @shift: shift count to shift the result left
2161 * These values are estimates at best, so no need for locking.
2163 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2165 loads[0] = (avenrun[0] + offset) << shift;
2166 loads[1] = (avenrun[1] + offset) << shift;
2167 loads[2] = (avenrun[2] + offset) << shift;
2170 static long calc_load_fold_active(struct rq *this_rq)
2172 long nr_active, delta = 0;
2174 nr_active = this_rq->nr_running;
2175 nr_active += (long) this_rq->nr_uninterruptible;
2177 if (nr_active != this_rq->calc_load_active) {
2178 delta = nr_active - this_rq->calc_load_active;
2179 this_rq->calc_load_active = nr_active;
2186 * a1 = a0 * e + a * (1 - e)
2188 static unsigned long
2189 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2192 load += active * (FIXED_1 - exp);
2193 load += 1UL << (FSHIFT - 1);
2194 return load >> FSHIFT;
2197 #ifdef CONFIG_NO_HZ_COMMON
2199 * Handle NO_HZ for the global load-average.
2201 * Since the above described distributed algorithm to compute the global
2202 * load-average relies on per-cpu sampling from the tick, it is affected by
2205 * The basic idea is to fold the nr_active delta into a global idle-delta upon
2206 * entering NO_HZ state such that we can include this as an 'extra' cpu delta
2207 * when we read the global state.
2209 * Obviously reality has to ruin such a delightfully simple scheme:
2211 * - When we go NO_HZ idle during the window, we can negate our sample
2212 * contribution, causing under-accounting.
2214 * We avoid this by keeping two idle-delta counters and flipping them
2215 * when the window starts, thus separating old and new NO_HZ load.
2217 * The only trick is the slight shift in index flip for read vs write.
2221 * |-|-----------|-|-----------|-|-----------|-|
2222 * r:0 0 1 1 0 0 1 1 0
2223 * w:0 1 1 0 0 1 1 0 0
2225 * This ensures we'll fold the old idle contribution in this window while
2226 * accumlating the new one.
2228 * - When we wake up from NO_HZ idle during the window, we push up our
2229 * contribution, since we effectively move our sample point to a known
2232 * This is solved by pushing the window forward, and thus skipping the
2233 * sample, for this cpu (effectively using the idle-delta for this cpu which
2234 * was in effect at the time the window opened). This also solves the issue
2235 * of having to deal with a cpu having been in NOHZ idle for multiple
2236 * LOAD_FREQ intervals.
2238 * When making the ILB scale, we should try to pull this in as well.
2240 static atomic_long_t calc_load_idle[2];
2241 static int calc_load_idx;
2243 static inline int calc_load_write_idx(void)
2245 int idx = calc_load_idx;
2248 * See calc_global_nohz(), if we observe the new index, we also
2249 * need to observe the new update time.
2254 * If the folding window started, make sure we start writing in the
2257 if (!time_before(jiffies, calc_load_update))
2263 static inline int calc_load_read_idx(void)
2265 return calc_load_idx & 1;
2268 void calc_load_enter_idle(void)
2270 struct rq *this_rq = this_rq();
2274 * We're going into NOHZ mode, if there's any pending delta, fold it
2275 * into the pending idle delta.
2277 delta = calc_load_fold_active(this_rq);
2279 int idx = calc_load_write_idx();
2280 atomic_long_add(delta, &calc_load_idle[idx]);
2284 void calc_load_exit_idle(void)
2286 struct rq *this_rq = this_rq();
2289 * If we're still before the sample window, we're done.
2291 if (time_before(jiffies, this_rq->calc_load_update))
2295 * We woke inside or after the sample window, this means we're already
2296 * accounted through the nohz accounting, so skip the entire deal and
2297 * sync up for the next window.
2299 this_rq->calc_load_update = calc_load_update;
2300 if (time_before(jiffies, this_rq->calc_load_update + 10))
2301 this_rq->calc_load_update += LOAD_FREQ;
2304 static long calc_load_fold_idle(void)
2306 int idx = calc_load_read_idx();
2309 if (atomic_long_read(&calc_load_idle[idx]))
2310 delta = atomic_long_xchg(&calc_load_idle[idx], 0);
2316 * fixed_power_int - compute: x^n, in O(log n) time
2318 * @x: base of the power
2319 * @frac_bits: fractional bits of @x
2320 * @n: power to raise @x to.
2322 * By exploiting the relation between the definition of the natural power
2323 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2324 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2325 * (where: n_i \elem {0, 1}, the binary vector representing n),
2326 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2327 * of course trivially computable in O(log_2 n), the length of our binary
2330 static unsigned long
2331 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
2333 unsigned long result = 1UL << frac_bits;
2338 result += 1UL << (frac_bits - 1);
2339 result >>= frac_bits;
2345 x += 1UL << (frac_bits - 1);
2353 * a1 = a0 * e + a * (1 - e)
2355 * a2 = a1 * e + a * (1 - e)
2356 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2357 * = a0 * e^2 + a * (1 - e) * (1 + e)
2359 * a3 = a2 * e + a * (1 - e)
2360 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2361 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2365 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2366 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2367 * = a0 * e^n + a * (1 - e^n)
2369 * [1] application of the geometric series:
2372 * S_n := \Sum x^i = -------------
2375 static unsigned long
2376 calc_load_n(unsigned long load, unsigned long exp,
2377 unsigned long active, unsigned int n)
2380 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
2384 * NO_HZ can leave us missing all per-cpu ticks calling
2385 * calc_load_account_active(), but since an idle CPU folds its delta into
2386 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
2387 * in the pending idle delta if our idle period crossed a load cycle boundary.
2389 * Once we've updated the global active value, we need to apply the exponential
2390 * weights adjusted to the number of cycles missed.
2392 static void calc_global_nohz(void)
2394 long delta, active, n;
2396 if (!time_before(jiffies, calc_load_update + 10)) {
2398 * Catch-up, fold however many we are behind still
2400 delta = jiffies - calc_load_update - 10;
2401 n = 1 + (delta / LOAD_FREQ);
2403 active = atomic_long_read(&calc_load_tasks);
2404 active = active > 0 ? active * FIXED_1 : 0;
2406 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
2407 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
2408 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
2410 calc_load_update += n * LOAD_FREQ;
2414 * Flip the idle index...
2416 * Make sure we first write the new time then flip the index, so that
2417 * calc_load_write_idx() will see the new time when it reads the new
2418 * index, this avoids a double flip messing things up.
2423 #else /* !CONFIG_NO_HZ_COMMON */
2425 static inline long calc_load_fold_idle(void) { return 0; }
2426 static inline void calc_global_nohz(void) { }
2428 #endif /* CONFIG_NO_HZ_COMMON */
2431 * calc_load - update the avenrun load estimates 10 ticks after the
2432 * CPUs have updated calc_load_tasks.
2434 void calc_global_load(unsigned long ticks)
2438 if (time_before(jiffies, calc_load_update + 10))
2442 * Fold the 'old' idle-delta to include all NO_HZ cpus.
2444 delta = calc_load_fold_idle();
2446 atomic_long_add(delta, &calc_load_tasks);
2448 active = atomic_long_read(&calc_load_tasks);
2449 active = active > 0 ? active * FIXED_1 : 0;
2451 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2452 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2453 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2455 calc_load_update += LOAD_FREQ;
2458 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk.
2464 * Called from update_cpu_load() to periodically update this CPU's
2467 static void calc_load_account_active(struct rq *this_rq)
2471 if (time_before(jiffies, this_rq->calc_load_update))
2474 delta = calc_load_fold_active(this_rq);
2476 atomic_long_add(delta, &calc_load_tasks);
2478 this_rq->calc_load_update += LOAD_FREQ;
2482 * End of global load-average stuff
2486 * The exact cpuload at various idx values, calculated at every tick would be
2487 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
2489 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
2490 * on nth tick when cpu may be busy, then we have:
2491 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2492 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
2494 * decay_load_missed() below does efficient calculation of
2495 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
2496 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
2498 * The calculation is approximated on a 128 point scale.
2499 * degrade_zero_ticks is the number of ticks after which load at any
2500 * particular idx is approximated to be zero.
2501 * degrade_factor is a precomputed table, a row for each load idx.
2502 * Each column corresponds to degradation factor for a power of two ticks,
2503 * based on 128 point scale.
2505 * row 2, col 3 (=12) says that the degradation at load idx 2 after
2506 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
2508 * With this power of 2 load factors, we can degrade the load n times
2509 * by looking at 1 bits in n and doing as many mult/shift instead of
2510 * n mult/shifts needed by the exact degradation.
2512 #define DEGRADE_SHIFT 7
2513 static const unsigned char
2514 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
2515 static const unsigned char
2516 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
2517 {0, 0, 0, 0, 0, 0, 0, 0},
2518 {64, 32, 8, 0, 0, 0, 0, 0},
2519 {96, 72, 40, 12, 1, 0, 0},
2520 {112, 98, 75, 43, 15, 1, 0},
2521 {120, 112, 98, 76, 45, 16, 2} };
2524 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
2525 * would be when CPU is idle and so we just decay the old load without
2526 * adding any new load.
2528 static unsigned long
2529 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
2533 if (!missed_updates)
2536 if (missed_updates >= degrade_zero_ticks[idx])
2540 return load >> missed_updates;
2542 while (missed_updates) {
2543 if (missed_updates % 2)
2544 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
2546 missed_updates >>= 1;
2553 * Update rq->cpu_load[] statistics. This function is usually called every
2554 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
2555 * every tick. We fix it up based on jiffies.
2557 static void __update_cpu_load(struct rq *this_rq, unsigned long this_load,
2558 unsigned long pending_updates)
2562 this_rq->nr_load_updates++;
2564 /* Update our load: */
2565 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
2566 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2567 unsigned long old_load, new_load;
2569 /* scale is effectively 1 << i now, and >> i divides by scale */
2571 old_load = this_rq->cpu_load[i];
2572 old_load = decay_load_missed(old_load, pending_updates - 1, i);
2573 new_load = this_load;
2575 * Round up the averaging division if load is increasing. This
2576 * prevents us from getting stuck on 9 if the load is 10, for
2579 if (new_load > old_load)
2580 new_load += scale - 1;
2582 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
2585 sched_avg_update(this_rq);
2588 #ifdef CONFIG_NO_HZ_COMMON
2590 * There is no sane way to deal with nohz on smp when using jiffies because the
2591 * cpu doing the jiffies update might drift wrt the cpu doing the jiffy reading
2592 * causing off-by-one errors in observed deltas; {0,2} instead of {1,1}.
2594 * Therefore we cannot use the delta approach from the regular tick since that
2595 * would seriously skew the load calculation. However we'll make do for those
2596 * updates happening while idle (nohz_idle_balance) or coming out of idle
2597 * (tick_nohz_idle_exit).
2599 * This means we might still be one tick off for nohz periods.
2603 * Called from nohz_idle_balance() to update the load ratings before doing the
2606 void update_idle_cpu_load(struct rq *this_rq)
2608 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2609 unsigned long load = this_rq->load.weight;
2610 unsigned long pending_updates;
2613 * bail if there's load or we're actually up-to-date.
2615 if (load || curr_jiffies == this_rq->last_load_update_tick)
2618 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2619 this_rq->last_load_update_tick = curr_jiffies;
2621 __update_cpu_load(this_rq, load, pending_updates);
2625 * Called from tick_nohz_idle_exit() -- try and fix up the ticks we missed.
2627 void update_cpu_load_nohz(void)
2629 struct rq *this_rq = this_rq();
2630 unsigned long curr_jiffies = ACCESS_ONCE(jiffies);
2631 unsigned long pending_updates;
2633 if (curr_jiffies == this_rq->last_load_update_tick)
2636 raw_spin_lock(&this_rq->lock);
2637 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
2638 if (pending_updates) {
2639 this_rq->last_load_update_tick = curr_jiffies;
2641 * We were idle, this means load 0, the current load might be
2642 * !0 due to remote wakeups and the sort.
2644 __update_cpu_load(this_rq, 0, pending_updates);
2646 raw_spin_unlock(&this_rq->lock);
2648 #endif /* CONFIG_NO_HZ_COMMON */
2651 * Called from scheduler_tick()
2653 static void update_cpu_load_active(struct rq *this_rq)
2656 * See the mess around update_idle_cpu_load() / update_cpu_load_nohz().
2658 this_rq->last_load_update_tick = jiffies;
2659 __update_cpu_load(this_rq, this_rq->load.weight, 1);
2661 calc_load_account_active(this_rq);
2667 * sched_exec - execve() is a valuable balancing opportunity, because at
2668 * this point the task has the smallest effective memory and cache footprint.
2670 void sched_exec(void)
2672 struct task_struct *p = current;
2673 unsigned long flags;
2676 raw_spin_lock_irqsave(&p->pi_lock, flags);
2677 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
2678 if (dest_cpu == smp_processor_id())
2681 if (likely(cpu_active(dest_cpu))) {
2682 struct migration_arg arg = { p, dest_cpu };
2684 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2685 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2689 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2694 DEFINE_PER_CPU(struct kernel_stat, kstat);
2695 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2697 EXPORT_PER_CPU_SYMBOL(kstat);
2698 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2701 * Return any ns on the sched_clock that have not yet been accounted in
2702 * @p in case that task is currently running.
2704 * Called with task_rq_lock() held on @rq.
2706 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2710 if (task_current(rq, p)) {
2711 update_rq_clock(rq);
2712 ns = rq->clock_task - p->se.exec_start;
2720 unsigned long long task_delta_exec(struct task_struct *p)
2722 unsigned long flags;
2726 rq = task_rq_lock(p, &flags);
2727 ns = do_task_delta_exec(p, rq);
2728 task_rq_unlock(rq, p, &flags);
2734 * Return accounted runtime for the task.
2735 * In case the task is currently running, return the runtime plus current's
2736 * pending runtime that have not been accounted yet.
2738 unsigned long long task_sched_runtime(struct task_struct *p)
2740 unsigned long flags;
2744 rq = task_rq_lock(p, &flags);
2745 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2746 task_rq_unlock(rq, p, &flags);
2752 * This function gets called by the timer code, with HZ frequency.
2753 * We call it with interrupts disabled.
2755 void scheduler_tick(void)
2757 int cpu = smp_processor_id();
2758 struct rq *rq = cpu_rq(cpu);
2759 struct task_struct *curr = rq->curr;
2763 raw_spin_lock(&rq->lock);
2764 update_rq_clock(rq);
2765 update_cpu_load_active(rq);
2766 curr->sched_class->task_tick(rq, curr, 0);
2767 raw_spin_unlock(&rq->lock);
2769 perf_event_task_tick();
2772 rq->idle_balance = idle_cpu(cpu);
2773 trigger_load_balance(rq, cpu);
2775 rq_last_tick_reset(rq);
2778 #ifdef CONFIG_NO_HZ_FULL
2780 * scheduler_tick_max_deferment
2782 * Keep at least one tick per second when a single
2783 * active task is running because the scheduler doesn't
2784 * yet completely support full dynticks environment.
2786 * This makes sure that uptime, CFS vruntime, load
2787 * balancing, etc... continue to move forward, even
2788 * with a very low granularity.
2790 u64 scheduler_tick_max_deferment(void)
2792 struct rq *rq = this_rq();
2793 unsigned long next, now = ACCESS_ONCE(jiffies);
2795 next = rq->last_sched_tick + HZ;
2797 if (time_before_eq(next, now))
2800 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2804 notrace unsigned long get_parent_ip(unsigned long addr)
2806 if (in_lock_functions(addr)) {
2807 addr = CALLER_ADDR2;
2808 if (in_lock_functions(addr))
2809 addr = CALLER_ADDR3;
2814 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2815 defined(CONFIG_PREEMPT_TRACER))
2817 void __kprobes add_preempt_count(int val)
2819 #ifdef CONFIG_DEBUG_PREEMPT
2823 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2826 preempt_count() += val;
2827 #ifdef CONFIG_DEBUG_PREEMPT
2829 * Spinlock count overflowing soon?
2831 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2834 if (preempt_count() == val)
2835 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2837 EXPORT_SYMBOL(add_preempt_count);
2839 void __kprobes sub_preempt_count(int val)
2841 #ifdef CONFIG_DEBUG_PREEMPT
2845 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2848 * Is the spinlock portion underflowing?
2850 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2851 !(preempt_count() & PREEMPT_MASK)))
2855 if (preempt_count() == val)
2856 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2857 preempt_count() -= val;
2859 EXPORT_SYMBOL(sub_preempt_count);
2864 * Print scheduling while atomic bug:
2866 static noinline void __schedule_bug(struct task_struct *prev)
2868 if (oops_in_progress)
2871 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2872 prev->comm, prev->pid, preempt_count());
2874 debug_show_held_locks(prev);
2876 if (irqs_disabled())
2877 print_irqtrace_events(prev);
2879 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2883 * Various schedule()-time debugging checks and statistics:
2885 static inline void schedule_debug(struct task_struct *prev)
2888 * Test if we are atomic. Since do_exit() needs to call into
2889 * schedule() atomically, we ignore that path for now.
2890 * Otherwise, whine if we are scheduling when we should not be.
2892 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
2893 __schedule_bug(prev);
2896 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2898 schedstat_inc(this_rq(), sched_count);
2901 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2903 if (prev->on_rq || rq->skip_clock_update < 0)
2904 update_rq_clock(rq);
2905 prev->sched_class->put_prev_task(rq, prev);
2909 * Pick up the highest-prio task:
2911 static inline struct task_struct *
2912 pick_next_task(struct rq *rq)
2914 const struct sched_class *class;
2915 struct task_struct *p;
2918 * Optimization: we know that if all tasks are in
2919 * the fair class we can call that function directly:
2921 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2922 p = fair_sched_class.pick_next_task(rq);
2927 for_each_class(class) {
2928 p = class->pick_next_task(rq);
2933 BUG(); /* the idle class will always have a runnable task */
2937 * __schedule() is the main scheduler function.
2939 * The main means of driving the scheduler and thus entering this function are:
2941 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2943 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2944 * paths. For example, see arch/x86/entry_64.S.
2946 * To drive preemption between tasks, the scheduler sets the flag in timer
2947 * interrupt handler scheduler_tick().
2949 * 3. Wakeups don't really cause entry into schedule(). They add a
2950 * task to the run-queue and that's it.
2952 * Now, if the new task added to the run-queue preempts the current
2953 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2954 * called on the nearest possible occasion:
2956 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2958 * - in syscall or exception context, at the next outmost
2959 * preempt_enable(). (this might be as soon as the wake_up()'s
2962 * - in IRQ context, return from interrupt-handler to
2963 * preemptible context
2965 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2968 * - cond_resched() call
2969 * - explicit schedule() call
2970 * - return from syscall or exception to user-space
2971 * - return from interrupt-handler to user-space
2973 static void __sched __schedule(void)
2975 struct task_struct *prev, *next;
2976 unsigned long *switch_count;
2982 cpu = smp_processor_id();
2984 rcu_note_context_switch(cpu);
2987 schedule_debug(prev);
2989 if (sched_feat(HRTICK))
2992 raw_spin_lock_irq(&rq->lock);
2994 switch_count = &prev->nivcsw;
2995 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2996 if (unlikely(signal_pending_state(prev->state, prev))) {
2997 prev->state = TASK_RUNNING;
2999 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3003 * If a worker went to sleep, notify and ask workqueue
3004 * whether it wants to wake up a task to maintain
3007 if (prev->flags & PF_WQ_WORKER) {
3008 struct task_struct *to_wakeup;
3010 to_wakeup = wq_worker_sleeping(prev, cpu);
3012 try_to_wake_up_local(to_wakeup);
3015 switch_count = &prev->nvcsw;
3018 pre_schedule(rq, prev);
3020 if (unlikely(!rq->nr_running))
3021 idle_balance(cpu, rq);
3023 put_prev_task(rq, prev);
3024 next = pick_next_task(rq);
3025 clear_tsk_need_resched(prev);
3026 rq->skip_clock_update = 0;
3028 if (likely(prev != next)) {
3033 context_switch(rq, prev, next); /* unlocks the rq */
3035 * The context switch have flipped the stack from under us
3036 * and restored the local variables which were saved when
3037 * this task called schedule() in the past. prev == current
3038 * is still correct, but it can be moved to another cpu/rq.
3040 cpu = smp_processor_id();
3043 raw_spin_unlock_irq(&rq->lock);
3047 sched_preempt_enable_no_resched();
3052 static inline void sched_submit_work(struct task_struct *tsk)
3054 if (!tsk->state || tsk_is_pi_blocked(tsk))
3057 * If we are going to sleep and we have plugged IO queued,
3058 * make sure to submit it to avoid deadlocks.
3060 if (blk_needs_flush_plug(tsk))
3061 blk_schedule_flush_plug(tsk);
3064 asmlinkage void __sched schedule(void)
3066 struct task_struct *tsk = current;
3068 sched_submit_work(tsk);
3071 EXPORT_SYMBOL(schedule);
3073 #ifdef CONFIG_CONTEXT_TRACKING
3074 asmlinkage void __sched schedule_user(void)
3077 * If we come here after a random call to set_need_resched(),
3078 * or we have been woken up remotely but the IPI has not yet arrived,
3079 * we haven't yet exited the RCU idle mode. Do it here manually until
3080 * we find a better solution.
3089 * schedule_preempt_disabled - called with preemption disabled
3091 * Returns with preemption disabled. Note: preempt_count must be 1
3093 void __sched schedule_preempt_disabled(void)
3095 sched_preempt_enable_no_resched();
3100 #ifdef CONFIG_PREEMPT
3102 * this is the entry point to schedule() from in-kernel preemption
3103 * off of preempt_enable. Kernel preemptions off return from interrupt
3104 * occur there and call schedule directly.
3106 asmlinkage void __sched notrace preempt_schedule(void)
3108 struct thread_info *ti = current_thread_info();
3111 * If there is a non-zero preempt_count or interrupts are disabled,
3112 * we do not want to preempt the current task. Just return..
3114 if (likely(ti->preempt_count || irqs_disabled()))
3118 add_preempt_count_notrace(PREEMPT_ACTIVE);
3120 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3123 * Check again in case we missed a preemption opportunity
3124 * between schedule and now.
3127 } while (need_resched());
3129 EXPORT_SYMBOL(preempt_schedule);
3132 * this is the entry point to schedule() from kernel preemption
3133 * off of irq context.
3134 * Note, that this is called and return with irqs disabled. This will
3135 * protect us against recursive calling from irq.
3137 asmlinkage void __sched preempt_schedule_irq(void)
3139 struct thread_info *ti = current_thread_info();
3140 enum ctx_state prev_state;
3142 /* Catch callers which need to be fixed */
3143 BUG_ON(ti->preempt_count || !irqs_disabled());
3145 prev_state = exception_enter();
3148 add_preempt_count(PREEMPT_ACTIVE);
3151 local_irq_disable();
3152 sub_preempt_count(PREEMPT_ACTIVE);
3155 * Check again in case we missed a preemption opportunity
3156 * between schedule and now.
3159 } while (need_resched());
3161 exception_exit(prev_state);
3164 #endif /* CONFIG_PREEMPT */
3166 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3169 return try_to_wake_up(curr->private, mode, wake_flags);
3171 EXPORT_SYMBOL(default_wake_function);
3174 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3175 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3176 * number) then we wake all the non-exclusive tasks and one exclusive task.
3178 * There are circumstances in which we can try to wake a task which has already
3179 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3180 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3182 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3183 int nr_exclusive, int wake_flags, void *key)
3185 wait_queue_t *curr, *next;
3187 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3188 unsigned flags = curr->flags;
3190 if (curr->func(curr, mode, wake_flags, key) &&
3191 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3197 * __wake_up - 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: is directly passed to the wakeup function
3203 * It may be assumed that this function implies a write memory barrier before
3204 * changing the task state if and only if any tasks are woken up.
3206 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3207 int nr_exclusive, void *key)
3209 unsigned long flags;
3211 spin_lock_irqsave(&q->lock, flags);
3212 __wake_up_common(q, mode, nr_exclusive, 0, key);
3213 spin_unlock_irqrestore(&q->lock, flags);
3215 EXPORT_SYMBOL(__wake_up);
3218 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3220 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr)
3222 __wake_up_common(q, mode, nr, 0, NULL);
3224 EXPORT_SYMBOL_GPL(__wake_up_locked);
3226 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3228 __wake_up_common(q, mode, 1, 0, key);
3230 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
3233 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3235 * @mode: which threads
3236 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3237 * @key: opaque value to be passed to wakeup targets
3239 * The sync wakeup differs that the waker knows that it will schedule
3240 * away soon, so while the target thread will be woken up, it will not
3241 * be migrated to another CPU - ie. the two threads are 'synchronized'
3242 * with each other. This can prevent needless bouncing between CPUs.
3244 * On UP it can prevent extra preemption.
3246 * It may be assumed that this function implies a write memory barrier before
3247 * changing the task state if and only if any tasks are woken up.
3249 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3250 int nr_exclusive, void *key)
3252 unsigned long flags;
3253 int wake_flags = WF_SYNC;
3258 if (unlikely(!nr_exclusive))
3261 spin_lock_irqsave(&q->lock, flags);
3262 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3263 spin_unlock_irqrestore(&q->lock, flags);
3265 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3268 * __wake_up_sync - see __wake_up_sync_key()
3270 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3272 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3274 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3277 * complete: - signals a single thread waiting on this completion
3278 * @x: holds the state of this particular completion
3280 * This will wake up a single thread waiting on this completion. Threads will be
3281 * awakened in the same order in which they were queued.
3283 * See also complete_all(), wait_for_completion() and related routines.
3285 * It may be assumed that this function implies a write memory barrier before
3286 * changing the task state if and only if any tasks are woken up.
3288 void complete(struct completion *x)
3290 unsigned long flags;
3292 spin_lock_irqsave(&x->wait.lock, flags);
3294 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3295 spin_unlock_irqrestore(&x->wait.lock, flags);
3297 EXPORT_SYMBOL(complete);
3300 * complete_all: - signals all threads waiting on this completion
3301 * @x: holds the state of this particular completion
3303 * This will wake up all threads waiting on this particular completion event.
3305 * It may be assumed that this function implies a write memory barrier before
3306 * changing the task state if and only if any tasks are woken up.
3308 void complete_all(struct completion *x)
3310 unsigned long flags;
3312 spin_lock_irqsave(&x->wait.lock, flags);
3313 x->done += UINT_MAX/2;
3314 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3315 spin_unlock_irqrestore(&x->wait.lock, flags);
3317 EXPORT_SYMBOL(complete_all);
3319 static inline long __sched
3320 do_wait_for_common(struct completion *x,
3321 long (*action)(long), long timeout, int state)
3324 DECLARE_WAITQUEUE(wait, current);
3326 __add_wait_queue_tail_exclusive(&x->wait, &wait);
3328 if (signal_pending_state(state, current)) {
3329 timeout = -ERESTARTSYS;
3332 __set_current_state(state);
3333 spin_unlock_irq(&x->wait.lock);
3334 timeout = action(timeout);
3335 spin_lock_irq(&x->wait.lock);
3336 } while (!x->done && timeout);
3337 __remove_wait_queue(&x->wait, &wait);
3342 return timeout ?: 1;
3345 static inline long __sched
3346 __wait_for_common(struct completion *x,
3347 long (*action)(long), long timeout, int state)
3351 spin_lock_irq(&x->wait.lock);
3352 timeout = do_wait_for_common(x, action, timeout, state);
3353 spin_unlock_irq(&x->wait.lock);
3358 wait_for_common(struct completion *x, long timeout, int state)
3360 return __wait_for_common(x, schedule_timeout, timeout, state);
3364 wait_for_common_io(struct completion *x, long timeout, int state)
3366 return __wait_for_common(x, io_schedule_timeout, timeout, state);
3370 * wait_for_completion: - waits for completion of a task
3371 * @x: holds the state of this particular completion
3373 * This waits to be signaled for completion of a specific task. It is NOT
3374 * interruptible and there is no timeout.
3376 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3377 * and interrupt capability. Also see complete().
3379 void __sched wait_for_completion(struct completion *x)
3381 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3383 EXPORT_SYMBOL(wait_for_completion);
3386 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3387 * @x: holds the state of this particular completion
3388 * @timeout: timeout value in jiffies
3390 * This waits for either a completion of a specific task to be signaled or for a
3391 * specified timeout to expire. The timeout is in jiffies. It is not
3394 * The return value is 0 if timed out, and positive (at least 1, or number of
3395 * jiffies left till timeout) if completed.
3397 unsigned long __sched
3398 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
3400 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
3402 EXPORT_SYMBOL(wait_for_completion_timeout);
3405 * wait_for_completion_io: - waits for completion of a task
3406 * @x: holds the state of this particular completion
3408 * This waits to be signaled for completion of a specific task. It is NOT
3409 * interruptible and there is no timeout. The caller is accounted as waiting
3412 void __sched wait_for_completion_io(struct completion *x)
3414 wait_for_common_io(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
3416 EXPORT_SYMBOL(wait_for_completion_io);
3419 * wait_for_completion_io_timeout: - waits for completion of a task (w/timeout)
3420 * @x: holds the state of this particular completion
3421 * @timeout: timeout value in jiffies
3423 * This waits for either a completion of a specific task to be signaled or for a
3424 * specified timeout to expire. The timeout is in jiffies. It is not
3425 * interruptible. The caller is accounted as waiting for IO.
3427 * The return value is 0 if timed out, and positive (at least 1, or number of
3428 * jiffies left till timeout) if completed.
3430 unsigned long __sched
3431 wait_for_completion_io_timeout(struct completion *x, unsigned long timeout)
3433 return wait_for_common_io(x, timeout, TASK_UNINTERRUPTIBLE);
3435 EXPORT_SYMBOL(wait_for_completion_io_timeout);
3438 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3439 * @x: holds the state of this particular completion
3441 * This waits for completion of a specific task to be signaled. It is
3444 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3446 int __sched wait_for_completion_interruptible(struct completion *x)
3448 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
3449 if (t == -ERESTARTSYS)
3453 EXPORT_SYMBOL(wait_for_completion_interruptible);
3456 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
3457 * @x: holds the state of this particular completion
3458 * @timeout: timeout value in jiffies
3460 * This waits for either a completion of a specific task to be signaled or for a
3461 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
3463 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3464 * positive (at least 1, or number of jiffies left till timeout) if completed.
3467 wait_for_completion_interruptible_timeout(struct completion *x,
3468 unsigned long timeout)
3470 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
3472 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
3475 * wait_for_completion_killable: - waits for completion of a task (killable)
3476 * @x: holds the state of this particular completion
3478 * This waits to be signaled for completion of a specific task. It can be
3479 * interrupted by a kill signal.
3481 * The return value is -ERESTARTSYS if interrupted, 0 if completed.
3483 int __sched wait_for_completion_killable(struct completion *x)
3485 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
3486 if (t == -ERESTARTSYS)
3490 EXPORT_SYMBOL(wait_for_completion_killable);
3493 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
3494 * @x: holds the state of this particular completion
3495 * @timeout: timeout value in jiffies
3497 * This waits for either a completion of a specific task to be
3498 * signaled or for a specified timeout to expire. It can be
3499 * interrupted by a kill signal. The timeout is in jiffies.
3501 * The return value is -ERESTARTSYS if interrupted, 0 if timed out,
3502 * positive (at least 1, or number of jiffies left till timeout) if completed.
3505 wait_for_completion_killable_timeout(struct completion *x,
3506 unsigned long timeout)
3508 return wait_for_common(x, timeout, TASK_KILLABLE);
3510 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
3513 * try_wait_for_completion - try to decrement a completion without blocking
3514 * @x: completion structure
3516 * Returns: 0 if a decrement cannot be done without blocking
3517 * 1 if a decrement succeeded.
3519 * If a completion is being used as a counting completion,
3520 * attempt to decrement the counter without blocking. This
3521 * enables us to avoid waiting if the resource the completion
3522 * is protecting is not available.
3524 bool try_wait_for_completion(struct completion *x)
3526 unsigned long flags;
3529 spin_lock_irqsave(&x->wait.lock, flags);
3534 spin_unlock_irqrestore(&x->wait.lock, flags);
3537 EXPORT_SYMBOL(try_wait_for_completion);
3540 * completion_done - Test to see if a completion has any waiters
3541 * @x: completion structure
3543 * Returns: 0 if there are waiters (wait_for_completion() in progress)
3544 * 1 if there are no waiters.
3547 bool completion_done(struct completion *x)
3549 unsigned long flags;
3552 spin_lock_irqsave(&x->wait.lock, flags);
3555 spin_unlock_irqrestore(&x->wait.lock, flags);
3558 EXPORT_SYMBOL(completion_done);
3561 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
3563 unsigned long flags;
3566 init_waitqueue_entry(&wait, current);
3568 __set_current_state(state);
3570 spin_lock_irqsave(&q->lock, flags);
3571 __add_wait_queue(q, &wait);
3572 spin_unlock(&q->lock);
3573 timeout = schedule_timeout(timeout);
3574 spin_lock_irq(&q->lock);
3575 __remove_wait_queue(q, &wait);
3576 spin_unlock_irqrestore(&q->lock, flags);
3581 void __sched interruptible_sleep_on(wait_queue_head_t *q)
3583 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3585 EXPORT_SYMBOL(interruptible_sleep_on);
3588 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
3590 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
3592 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
3594 void __sched sleep_on(wait_queue_head_t *q)
3596 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
3598 EXPORT_SYMBOL(sleep_on);
3600 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
3602 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
3604 EXPORT_SYMBOL(sleep_on_timeout);
3606 #ifdef CONFIG_RT_MUTEXES
3609 * rt_mutex_setprio - set the current priority of a task
3611 * @prio: prio value (kernel-internal form)
3613 * This function changes the 'effective' priority of a task. It does
3614 * not touch ->normal_prio like __setscheduler().
3616 * Used by the rt_mutex code to implement priority inheritance logic.
3618 void rt_mutex_setprio(struct task_struct *p, int prio)
3620 int oldprio, on_rq, running;
3622 const struct sched_class *prev_class;
3624 BUG_ON(prio < 0 || prio > MAX_PRIO);
3626 rq = __task_rq_lock(p);
3629 * Idle task boosting is a nono in general. There is one
3630 * exception, when PREEMPT_RT and NOHZ is active:
3632 * The idle task calls get_next_timer_interrupt() and holds
3633 * the timer wheel base->lock on the CPU and another CPU wants
3634 * to access the timer (probably to cancel it). We can safely
3635 * ignore the boosting request, as the idle CPU runs this code
3636 * with interrupts disabled and will complete the lock
3637 * protected section without being interrupted. So there is no
3638 * real need to boost.
3640 if (unlikely(p == rq->idle)) {
3641 WARN_ON(p != rq->curr);
3642 WARN_ON(p->pi_blocked_on);
3646 trace_sched_pi_setprio(p, prio);
3648 prev_class = p->sched_class;
3650 running = task_current(rq, p);
3652 dequeue_task(rq, p, 0);
3654 p->sched_class->put_prev_task(rq, p);
3657 p->sched_class = &rt_sched_class;
3659 p->sched_class = &fair_sched_class;
3664 p->sched_class->set_curr_task(rq);
3666 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
3668 check_class_changed(rq, p, prev_class, oldprio);
3670 __task_rq_unlock(rq);
3673 void set_user_nice(struct task_struct *p, long nice)
3675 int old_prio, delta, on_rq;
3676 unsigned long flags;
3679 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3682 * We have to be careful, if called from sys_setpriority(),
3683 * the task might be in the middle of scheduling on another CPU.
3685 rq = task_rq_lock(p, &flags);
3687 * The RT priorities are set via sched_setscheduler(), but we still
3688 * allow the 'normal' nice value to be set - but as expected
3689 * it wont have any effect on scheduling until the task is
3690 * SCHED_FIFO/SCHED_RR:
3692 if (task_has_rt_policy(p)) {
3693 p->static_prio = NICE_TO_PRIO(nice);
3698 dequeue_task(rq, p, 0);
3700 p->static_prio = NICE_TO_PRIO(nice);
3703 p->prio = effective_prio(p);
3704 delta = p->prio - old_prio;
3707 enqueue_task(rq, p, 0);
3709 * If the task increased its priority or is running and
3710 * lowered its priority, then reschedule its CPU:
3712 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3713 resched_task(rq->curr);
3716 task_rq_unlock(rq, p, &flags);
3718 EXPORT_SYMBOL(set_user_nice);
3721 * can_nice - check if a task can reduce its nice value
3725 int can_nice(const struct task_struct *p, const int nice)
3727 /* convert nice value [19,-20] to rlimit style value [1,40] */
3728 int nice_rlim = 20 - nice;
3730 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3731 capable(CAP_SYS_NICE));
3734 #ifdef __ARCH_WANT_SYS_NICE
3737 * sys_nice - change the priority of the current process.
3738 * @increment: priority increment
3740 * sys_setpriority is a more generic, but much slower function that
3741 * does similar things.
3743 SYSCALL_DEFINE1(nice, int, increment)
3748 * Setpriority might change our priority at the same moment.
3749 * We don't have to worry. Conceptually one call occurs first
3750 * and we have a single winner.
3752 if (increment < -40)
3757 nice = TASK_NICE(current) + increment;
3763 if (increment < 0 && !can_nice(current, nice))
3766 retval = security_task_setnice(current, nice);
3770 set_user_nice(current, nice);
3777 * task_prio - return the priority value of a given task.
3778 * @p: the task in question.
3780 * This is the priority value as seen by users in /proc.
3781 * RT tasks are offset by -200. Normal tasks are centered
3782 * around 0, value goes from -16 to +15.
3784 int task_prio(const struct task_struct *p)
3786 return p->prio - MAX_RT_PRIO;
3790 * task_nice - return the nice value of a given task.
3791 * @p: the task in question.
3793 int task_nice(const struct task_struct *p)
3795 return TASK_NICE(p);
3797 EXPORT_SYMBOL(task_nice);
3800 * idle_cpu - is a given cpu idle currently?
3801 * @cpu: the processor in question.
3803 int idle_cpu(int cpu)
3805 struct rq *rq = cpu_rq(cpu);
3807 if (rq->curr != rq->idle)
3814 if (!llist_empty(&rq->wake_list))
3822 * idle_task - return the idle task for a given cpu.
3823 * @cpu: the processor in question.
3825 struct task_struct *idle_task(int cpu)
3827 return cpu_rq(cpu)->idle;
3831 * find_process_by_pid - find a process with a matching PID value.
3832 * @pid: the pid in question.
3834 static struct task_struct *find_process_by_pid(pid_t pid)
3836 return pid ? find_task_by_vpid(pid) : current;
3839 extern struct cpumask hmp_slow_cpu_mask;
3841 /* Actually do priority change: must hold rq lock. */
3843 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
3846 p->rt_priority = prio;
3847 p->normal_prio = normal_prio(p);
3848 /* we are holding p->pi_lock already */
3849 p->prio = rt_mutex_getprio(p);
3850 if (rt_prio(p->prio)) {
3851 p->sched_class = &rt_sched_class;
3852 #ifdef CONFIG_SCHED_HMP
3853 if (!cpumask_empty(&hmp_slow_cpu_mask))
3854 if (cpumask_equal(&p->cpus_allowed, cpu_all_mask)) {
3855 p->nr_cpus_allowed =
3856 cpumask_weight(&hmp_slow_cpu_mask);
3857 do_set_cpus_allowed(p, &hmp_slow_cpu_mask);
3862 p->sched_class = &fair_sched_class;
3867 * check the target process has a UID that matches the current process's
3869 static bool check_same_owner(struct task_struct *p)
3871 const struct cred *cred = current_cred(), *pcred;
3875 pcred = __task_cred(p);
3876 match = (uid_eq(cred->euid, pcred->euid) ||
3877 uid_eq(cred->euid, pcred->uid));
3882 static int __sched_setscheduler(struct task_struct *p, int policy,
3883 const struct sched_param *param, bool user)
3885 int retval, oldprio, oldpolicy = -1, on_rq, running;
3886 unsigned long flags;
3887 const struct sched_class *prev_class;
3891 /* may grab non-irq protected spin_locks */
3892 BUG_ON(in_interrupt());
3894 /* double check policy once rq lock held */
3896 reset_on_fork = p->sched_reset_on_fork;
3897 policy = oldpolicy = p->policy;
3899 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3900 policy &= ~SCHED_RESET_ON_FORK;
3902 if (policy != SCHED_FIFO && policy != SCHED_RR &&
3903 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3904 policy != SCHED_IDLE)
3909 * Valid priorities for SCHED_FIFO and SCHED_RR are
3910 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3911 * SCHED_BATCH and SCHED_IDLE is 0.
3913 if (param->sched_priority < 0 ||
3914 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
3915 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
3917 if (rt_policy(policy) != (param->sched_priority != 0))
3921 * Allow unprivileged RT tasks to decrease priority:
3923 if (user && !capable(CAP_SYS_NICE)) {
3924 if (rt_policy(policy)) {
3925 unsigned long rlim_rtprio =
3926 task_rlimit(p, RLIMIT_RTPRIO);
3928 /* can't set/change the rt policy */
3929 if (policy != p->policy && !rlim_rtprio)
3932 /* can't increase priority */
3933 if (param->sched_priority > p->rt_priority &&
3934 param->sched_priority > rlim_rtprio)
3939 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3940 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3942 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3943 if (!can_nice(p, TASK_NICE(p)))
3947 /* can't change other user's priorities */
3948 if (!check_same_owner(p))
3951 /* Normal users shall not reset the sched_reset_on_fork flag */
3952 if (p->sched_reset_on_fork && !reset_on_fork)
3957 retval = security_task_setscheduler(p);
3963 * make sure no PI-waiters arrive (or leave) while we are
3964 * changing the priority of the task:
3966 * To be able to change p->policy safely, the appropriate
3967 * runqueue lock must be held.
3969 rq = task_rq_lock(p, &flags);
3972 * Changing the policy of the stop threads its a very bad idea
3974 if (p == rq->stop) {
3975 task_rq_unlock(rq, p, &flags);
3980 * If not changing anything there's no need to proceed further:
3982 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
3983 param->sched_priority == p->rt_priority))) {
3984 task_rq_unlock(rq, p, &flags);
3988 #ifdef CONFIG_RT_GROUP_SCHED
3991 * Do not allow realtime tasks into groups that have no runtime
3994 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3995 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3996 !task_group_is_autogroup(task_group(p))) {
3997 task_rq_unlock(rq, p, &flags);
4003 /* recheck policy now with rq lock held */
4004 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4005 policy = oldpolicy = -1;
4006 task_rq_unlock(rq, p, &flags);
4010 running = task_current(rq, p);
4012 dequeue_task(rq, p, 0);
4014 p->sched_class->put_prev_task(rq, p);
4016 p->sched_reset_on_fork = reset_on_fork;
4019 prev_class = p->sched_class;
4020 __setscheduler(rq, p, policy, param->sched_priority);
4023 p->sched_class->set_curr_task(rq);
4025 enqueue_task(rq, p, 0);
4027 check_class_changed(rq, p, prev_class, oldprio);
4028 task_rq_unlock(rq, p, &flags);
4030 rt_mutex_adjust_pi(p);
4036 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4037 * @p: the task in question.
4038 * @policy: new policy.
4039 * @param: structure containing the new RT priority.
4041 * NOTE that the task may be already dead.
4043 int sched_setscheduler(struct task_struct *p, int policy,
4044 const struct sched_param *param)
4046 return __sched_setscheduler(p, policy, param, true);
4048 EXPORT_SYMBOL_GPL(sched_setscheduler);
4051 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4052 * @p: the task in question.
4053 * @policy: new policy.
4054 * @param: structure containing the new RT priority.
4056 * Just like sched_setscheduler, only don't bother checking if the
4057 * current context has permission. For example, this is needed in
4058 * stop_machine(): we create temporary high priority worker threads,
4059 * but our caller might not have that capability.
4061 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4062 const struct sched_param *param)
4064 return __sched_setscheduler(p, policy, param, false);
4068 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4070 struct sched_param lparam;
4071 struct task_struct *p;
4074 if (!param || pid < 0)
4076 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4081 p = find_process_by_pid(pid);
4083 retval = sched_setscheduler(p, policy, &lparam);
4090 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4091 * @pid: the pid in question.
4092 * @policy: new policy.
4093 * @param: structure containing the new RT priority.
4095 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4096 struct sched_param __user *, param)
4098 /* negative values for policy are not valid */
4102 return do_sched_setscheduler(pid, policy, param);
4106 * sys_sched_setparam - set/change the RT priority of a thread
4107 * @pid: the pid in question.
4108 * @param: structure containing the new RT priority.
4110 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4112 return do_sched_setscheduler(pid, -1, param);
4116 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4117 * @pid: the pid in question.
4119 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4121 struct task_struct *p;
4129 p = find_process_by_pid(pid);
4131 retval = security_task_getscheduler(p);
4134 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4141 * sys_sched_getparam - get the RT priority of a thread
4142 * @pid: the pid in question.
4143 * @param: structure containing the RT priority.
4145 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4147 struct sched_param lp;
4148 struct task_struct *p;
4151 if (!param || pid < 0)
4155 p = find_process_by_pid(pid);
4160 retval = security_task_getscheduler(p);
4164 lp.sched_priority = p->rt_priority;
4168 * This one might sleep, we cannot do it with a spinlock held ...
4170 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4179 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4181 cpumask_var_t cpus_allowed, new_mask;
4182 struct task_struct *p;
4188 p = find_process_by_pid(pid);
4195 /* Prevent p going away */
4199 if (p->flags & PF_NO_SETAFFINITY) {
4203 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4207 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4209 goto out_free_cpus_allowed;
4212 if (!check_same_owner(p)) {
4214 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4221 retval = security_task_setscheduler(p);
4225 cpuset_cpus_allowed(p, cpus_allowed);
4226 cpumask_and(new_mask, in_mask, cpus_allowed);
4228 retval = set_cpus_allowed_ptr(p, new_mask);
4231 cpuset_cpus_allowed(p, cpus_allowed);
4232 if (!cpumask_subset(new_mask, cpus_allowed)) {
4234 * We must have raced with a concurrent cpuset
4235 * update. Just reset the cpus_allowed to the
4236 * cpuset's cpus_allowed
4238 cpumask_copy(new_mask, cpus_allowed);
4243 free_cpumask_var(new_mask);
4244 out_free_cpus_allowed:
4245 free_cpumask_var(cpus_allowed);
4252 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4253 struct cpumask *new_mask)
4255 if (len < cpumask_size())
4256 cpumask_clear(new_mask);
4257 else if (len > cpumask_size())
4258 len = cpumask_size();
4260 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4264 * sys_sched_setaffinity - set the cpu affinity of a process
4265 * @pid: pid of the process
4266 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4267 * @user_mask_ptr: user-space pointer to the new cpu mask
4269 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4270 unsigned long __user *, user_mask_ptr)
4272 cpumask_var_t new_mask;
4275 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4278 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4280 retval = sched_setaffinity(pid, new_mask);
4281 free_cpumask_var(new_mask);
4285 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4287 struct task_struct *p;
4288 unsigned long flags;
4295 p = find_process_by_pid(pid);
4299 retval = security_task_getscheduler(p);
4303 raw_spin_lock_irqsave(&p->pi_lock, flags);
4304 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4305 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4315 * sys_sched_getaffinity - get the cpu affinity of a process
4316 * @pid: pid of the process
4317 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4318 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4320 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4321 unsigned long __user *, user_mask_ptr)
4326 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4328 if (len & (sizeof(unsigned long)-1))
4331 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4334 ret = sched_getaffinity(pid, mask);
4336 size_t retlen = min_t(size_t, len, cpumask_size());
4338 if (copy_to_user(user_mask_ptr, mask, retlen))
4343 free_cpumask_var(mask);
4349 * sys_sched_yield - yield the current processor to other threads.
4351 * This function yields the current CPU to other tasks. If there are no
4352 * other threads running on this CPU then this function will return.
4354 SYSCALL_DEFINE0(sched_yield)
4356 struct rq *rq = this_rq_lock();
4358 schedstat_inc(rq, yld_count);
4359 current->sched_class->yield_task(rq);
4362 * Since we are going to call schedule() anyway, there's
4363 * no need to preempt or enable interrupts:
4365 __release(rq->lock);
4366 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4367 do_raw_spin_unlock(&rq->lock);
4368 sched_preempt_enable_no_resched();
4375 static inline int should_resched(void)
4377 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4380 static void __cond_resched(void)
4382 add_preempt_count(PREEMPT_ACTIVE);
4384 sub_preempt_count(PREEMPT_ACTIVE);
4387 int __sched _cond_resched(void)
4389 if (should_resched()) {
4395 EXPORT_SYMBOL(_cond_resched);
4398 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4399 * call schedule, and on return reacquire the lock.
4401 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4402 * operations here to prevent schedule() from being called twice (once via
4403 * spin_unlock(), once by hand).
4405 int __cond_resched_lock(spinlock_t *lock)
4407 int resched = should_resched();
4410 lockdep_assert_held(lock);
4412 if (spin_needbreak(lock) || resched) {
4423 EXPORT_SYMBOL(__cond_resched_lock);
4425 int __sched __cond_resched_softirq(void)
4427 BUG_ON(!in_softirq());
4429 if (should_resched()) {
4437 EXPORT_SYMBOL(__cond_resched_softirq);
4440 * yield - yield the current processor to other threads.
4442 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4444 * The scheduler is at all times free to pick the calling task as the most
4445 * eligible task to run, if removing the yield() call from your code breaks
4446 * it, its already broken.
4448 * Typical broken usage is:
4453 * where one assumes that yield() will let 'the other' process run that will
4454 * make event true. If the current task is a SCHED_FIFO task that will never
4455 * happen. Never use yield() as a progress guarantee!!
4457 * If you want to use yield() to wait for something, use wait_event().
4458 * If you want to use yield() to be 'nice' for others, use cond_resched().
4459 * If you still want to use yield(), do not!
4461 void __sched yield(void)
4463 set_current_state(TASK_RUNNING);
4466 EXPORT_SYMBOL(yield);
4469 * yield_to - yield the current processor to another thread in
4470 * your thread group, or accelerate that thread toward the
4471 * processor it's on.
4473 * @preempt: whether task preemption is allowed or not
4475 * It's the caller's job to ensure that the target task struct
4476 * can't go away on us before we can do any checks.
4479 * true (>0) if we indeed boosted the target task.
4480 * false (0) if we failed to boost the target.
4481 * -ESRCH if there's no task to yield to.
4483 bool __sched yield_to(struct task_struct *p, bool preempt)
4485 struct task_struct *curr = current;
4486 struct rq *rq, *p_rq;
4487 unsigned long flags;
4490 local_irq_save(flags);
4496 * If we're the only runnable task on the rq and target rq also
4497 * has only one task, there's absolutely no point in yielding.
4499 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4504 double_rq_lock(rq, p_rq);
4505 while (task_rq(p) != p_rq) {
4506 double_rq_unlock(rq, p_rq);
4510 if (!curr->sched_class->yield_to_task)
4513 if (curr->sched_class != p->sched_class)
4516 if (task_running(p_rq, p) || p->state)
4519 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4521 schedstat_inc(rq, yld_count);
4523 * Make p's CPU reschedule; pick_next_entity takes care of
4526 if (preempt && rq != p_rq)
4527 resched_task(p_rq->curr);
4531 double_rq_unlock(rq, p_rq);
4533 local_irq_restore(flags);
4540 EXPORT_SYMBOL_GPL(yield_to);
4543 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4544 * that process accounting knows that this is a task in IO wait state.
4546 void __sched io_schedule(void)
4548 struct rq *rq = raw_rq();
4550 delayacct_blkio_start();
4551 atomic_inc(&rq->nr_iowait);
4552 blk_flush_plug(current);
4553 current->in_iowait = 1;
4555 current->in_iowait = 0;
4556 atomic_dec(&rq->nr_iowait);
4557 delayacct_blkio_end();
4559 EXPORT_SYMBOL(io_schedule);
4561 long __sched io_schedule_timeout(long timeout)
4563 struct rq *rq = raw_rq();
4566 delayacct_blkio_start();
4567 atomic_inc(&rq->nr_iowait);
4568 blk_flush_plug(current);
4569 current->in_iowait = 1;
4570 ret = schedule_timeout(timeout);
4571 current->in_iowait = 0;
4572 atomic_dec(&rq->nr_iowait);
4573 delayacct_blkio_end();
4578 * sys_sched_get_priority_max - return maximum RT priority.
4579 * @policy: scheduling class.
4581 * this syscall returns the maximum rt_priority that can be used
4582 * by a given scheduling class.
4584 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4591 ret = MAX_USER_RT_PRIO-1;
4603 * sys_sched_get_priority_min - return minimum RT priority.
4604 * @policy: scheduling class.
4606 * this syscall returns the minimum rt_priority that can be used
4607 * by a given scheduling class.
4609 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4627 * sys_sched_rr_get_interval - return the default timeslice of a process.
4628 * @pid: pid of the process.
4629 * @interval: userspace pointer to the timeslice value.
4631 * this syscall writes the default timeslice value of a given process
4632 * into the user-space timespec buffer. A value of '0' means infinity.
4634 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4635 struct timespec __user *, interval)
4637 struct task_struct *p;
4638 unsigned int time_slice;
4639 unsigned long flags;
4649 p = find_process_by_pid(pid);
4653 retval = security_task_getscheduler(p);
4657 rq = task_rq_lock(p, &flags);
4658 time_slice = p->sched_class->get_rr_interval(rq, p);
4659 task_rq_unlock(rq, p, &flags);
4662 jiffies_to_timespec(time_slice, &t);
4663 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4671 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4673 void sched_show_task(struct task_struct *p)
4675 unsigned long free = 0;
4679 state = p->state ? __ffs(p->state) + 1 : 0;
4680 printk(KERN_INFO "%-15.15s %c", p->comm,
4681 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4682 #if BITS_PER_LONG == 32
4683 if (state == TASK_RUNNING)
4684 printk(KERN_CONT " running ");
4686 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4688 if (state == TASK_RUNNING)
4689 printk(KERN_CONT " running task ");
4691 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4693 #ifdef CONFIG_DEBUG_STACK_USAGE
4694 free = stack_not_used(p);
4697 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4699 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4700 task_pid_nr(p), ppid,
4701 (unsigned long)task_thread_info(p)->flags);
4703 print_worker_info(KERN_INFO, p);
4704 show_stack(p, NULL);
4707 void show_state_filter(unsigned long state_filter)
4709 struct task_struct *g, *p;
4711 #if BITS_PER_LONG == 32
4713 " task PC stack pid father\n");
4716 " task PC stack pid father\n");
4719 do_each_thread(g, p) {
4721 * reset the NMI-timeout, listing all files on a slow
4722 * console might take a lot of time:
4724 touch_nmi_watchdog();
4725 if (!state_filter || (p->state & state_filter))
4727 } while_each_thread(g, p);
4729 touch_all_softlockup_watchdogs();
4731 #ifdef CONFIG_SCHED_DEBUG
4732 sysrq_sched_debug_show();
4736 * Only show locks if all tasks are dumped:
4739 debug_show_all_locks();
4742 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
4744 idle->sched_class = &idle_sched_class;
4748 * init_idle - set up an idle thread for a given CPU
4749 * @idle: task in question
4750 * @cpu: cpu the idle task belongs to
4752 * NOTE: this function does not set the idle thread's NEED_RESCHED
4753 * flag, to make booting more robust.
4755 void __cpuinit init_idle(struct task_struct *idle, int cpu)
4757 struct rq *rq = cpu_rq(cpu);
4758 unsigned long flags;
4760 raw_spin_lock_irqsave(&rq->lock, flags);
4763 idle->state = TASK_RUNNING;
4764 idle->se.exec_start = sched_clock();
4766 do_set_cpus_allowed(idle, cpumask_of(cpu));
4768 * We're having a chicken and egg problem, even though we are
4769 * holding rq->lock, the cpu isn't yet set to this cpu so the
4770 * lockdep check in task_group() will fail.
4772 * Similar case to sched_fork(). / Alternatively we could
4773 * use task_rq_lock() here and obtain the other rq->lock.
4778 __set_task_cpu(idle, cpu);
4781 rq->curr = rq->idle = idle;
4782 #if defined(CONFIG_SMP)
4785 raw_spin_unlock_irqrestore(&rq->lock, flags);
4787 /* Set the preempt count _outside_ the spinlocks! */
4788 task_thread_info(idle)->preempt_count = 0;
4791 * The idle tasks have their own, simple scheduling class:
4793 idle->sched_class = &idle_sched_class;
4794 ftrace_graph_init_idle_task(idle, cpu);
4795 vtime_init_idle(idle, cpu);
4796 #if defined(CONFIG_SMP)
4797 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4802 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4804 if (p->sched_class && p->sched_class->set_cpus_allowed)
4805 p->sched_class->set_cpus_allowed(p, new_mask);
4807 cpumask_copy(&p->cpus_allowed, new_mask);
4808 p->nr_cpus_allowed = cpumask_weight(new_mask);
4812 * This is how migration works:
4814 * 1) we invoke migration_cpu_stop() on the target CPU using
4816 * 2) stopper starts to run (implicitly forcing the migrated thread
4818 * 3) it checks whether the migrated task is still in the wrong runqueue.
4819 * 4) if it's in the wrong runqueue then the migration thread removes
4820 * it and puts it into the right queue.
4821 * 5) stopper completes and stop_one_cpu() returns and the migration
4826 * Change a given task's CPU affinity. Migrate the thread to a
4827 * proper CPU and schedule it away if the CPU it's executing on
4828 * is removed from the allowed bitmask.
4830 * NOTE: the caller must have a valid reference to the task, the
4831 * task must not exit() & deallocate itself prematurely. The
4832 * call is not atomic; no spinlocks may be held.
4834 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4836 unsigned long flags;
4838 unsigned int dest_cpu;
4841 rq = task_rq_lock(p, &flags);
4843 if (cpumask_equal(&p->cpus_allowed, new_mask))
4846 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4851 do_set_cpus_allowed(p, new_mask);
4853 /* Can the task run on the task's current CPU? If so, we're done */
4854 if (cpumask_test_cpu(task_cpu(p), new_mask))
4857 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4859 struct migration_arg arg = { p, dest_cpu };
4860 /* Need help from migration thread: drop lock and wait. */
4861 task_rq_unlock(rq, p, &flags);
4862 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4863 tlb_migrate_finish(p->mm);
4867 task_rq_unlock(rq, p, &flags);
4871 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4874 * Move (not current) task off this cpu, onto dest cpu. We're doing
4875 * this because either it can't run here any more (set_cpus_allowed()
4876 * away from this CPU, or CPU going down), or because we're
4877 * attempting to rebalance this task on exec (sched_exec).
4879 * So we race with normal scheduler movements, but that's OK, as long
4880 * as the task is no longer on this CPU.
4882 * Returns non-zero if task was successfully migrated.
4884 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4886 struct rq *rq_dest, *rq_src;
4889 if (unlikely(!cpu_active(dest_cpu)))
4892 rq_src = cpu_rq(src_cpu);
4893 rq_dest = cpu_rq(dest_cpu);
4895 raw_spin_lock(&p->pi_lock);
4896 double_rq_lock(rq_src, rq_dest);
4897 /* Already moved. */
4898 if (task_cpu(p) != src_cpu)
4900 /* Affinity changed (again). */
4901 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4905 * If we're not on a rq, the next wake-up will ensure we're
4909 dequeue_task(rq_src, p, 0);
4910 set_task_cpu(p, dest_cpu);
4911 enqueue_task(rq_dest, p, 0);
4912 check_preempt_curr(rq_dest, p, 0);
4917 double_rq_unlock(rq_src, rq_dest);
4918 raw_spin_unlock(&p->pi_lock);
4923 * migration_cpu_stop - this will be executed by a highprio stopper thread
4924 * and performs thread migration by bumping thread off CPU then
4925 * 'pushing' onto another runqueue.
4927 static int migration_cpu_stop(void *data)
4929 struct migration_arg *arg = data;
4932 * The original target cpu might have gone down and we might
4933 * be on another cpu but it doesn't matter.
4935 local_irq_disable();
4936 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4941 #ifdef CONFIG_HOTPLUG_CPU
4944 * Ensures that the idle task is using init_mm right before its cpu goes
4947 void idle_task_exit(void)
4949 struct mm_struct *mm = current->active_mm;
4951 BUG_ON(cpu_online(smp_processor_id()));
4954 switch_mm(mm, &init_mm, current);
4959 * Since this CPU is going 'away' for a while, fold any nr_active delta
4960 * we might have. Assumes we're called after migrate_tasks() so that the
4961 * nr_active count is stable.
4963 * Also see the comment "Global load-average calculations".
4965 static void calc_load_migrate(struct rq *rq)
4967 long delta = calc_load_fold_active(rq);
4969 atomic_long_add(delta, &calc_load_tasks);
4973 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4974 * try_to_wake_up()->select_task_rq().
4976 * Called with rq->lock held even though we'er in stop_machine() and
4977 * there's no concurrency possible, we hold the required locks anyway
4978 * because of lock validation efforts.
4980 static void migrate_tasks(unsigned int dead_cpu)
4982 struct rq *rq = cpu_rq(dead_cpu);
4983 struct task_struct *next, *stop = rq->stop;
4987 * Fudge the rq selection such that the below task selection loop
4988 * doesn't get stuck on the currently eligible stop task.
4990 * We're currently inside stop_machine() and the rq is either stuck
4991 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4992 * either way we should never end up calling schedule() until we're
4999 * There's this thread running, bail when that's the only
5002 if (rq->nr_running == 1)
5005 next = pick_next_task(rq);
5007 next->sched_class->put_prev_task(rq, next);
5009 /* Find suitable destination for @next, with force if needed. */
5010 dest_cpu = select_fallback_rq(dead_cpu, next);
5011 raw_spin_unlock(&rq->lock);
5013 __migrate_task(next, dead_cpu, dest_cpu);
5015 raw_spin_lock(&rq->lock);
5021 #endif /* CONFIG_HOTPLUG_CPU */
5023 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5025 static struct ctl_table sd_ctl_dir[] = {
5027 .procname = "sched_domain",
5033 static struct ctl_table sd_ctl_root[] = {
5035 .procname = "kernel",
5037 .child = sd_ctl_dir,
5042 static struct ctl_table *sd_alloc_ctl_entry(int n)
5044 struct ctl_table *entry =
5045 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5050 static void sd_free_ctl_entry(struct ctl_table **tablep)
5052 struct ctl_table *entry;
5055 * In the intermediate directories, both the child directory and
5056 * procname are dynamically allocated and could fail but the mode
5057 * will always be set. In the lowest directory the names are
5058 * static strings and all have proc handlers.
5060 for (entry = *tablep; entry->mode; entry++) {
5062 sd_free_ctl_entry(&entry->child);
5063 if (entry->proc_handler == NULL)
5064 kfree(entry->procname);
5071 static int min_load_idx = 0;
5072 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5075 set_table_entry(struct ctl_table *entry,
5076 const char *procname, void *data, int maxlen,
5077 umode_t mode, proc_handler *proc_handler,
5080 entry->procname = procname;
5082 entry->maxlen = maxlen;
5084 entry->proc_handler = proc_handler;
5087 entry->extra1 = &min_load_idx;
5088 entry->extra2 = &max_load_idx;
5092 static struct ctl_table *
5093 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5095 struct ctl_table *table = sd_alloc_ctl_entry(13);
5100 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5101 sizeof(long), 0644, proc_doulongvec_minmax, false);
5102 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5103 sizeof(long), 0644, proc_doulongvec_minmax, false);
5104 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5105 sizeof(int), 0644, proc_dointvec_minmax, true);
5106 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5107 sizeof(int), 0644, proc_dointvec_minmax, true);
5108 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5109 sizeof(int), 0644, proc_dointvec_minmax, true);
5110 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5111 sizeof(int), 0644, proc_dointvec_minmax, true);
5112 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5113 sizeof(int), 0644, proc_dointvec_minmax, true);
5114 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5115 sizeof(int), 0644, proc_dointvec_minmax, false);
5116 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5117 sizeof(int), 0644, proc_dointvec_minmax, false);
5118 set_table_entry(&table[9], "cache_nice_tries",
5119 &sd->cache_nice_tries,
5120 sizeof(int), 0644, proc_dointvec_minmax, false);
5121 set_table_entry(&table[10], "flags", &sd->flags,
5122 sizeof(int), 0644, proc_dointvec_minmax, false);
5123 set_table_entry(&table[11], "name", sd->name,
5124 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5125 /* &table[12] is terminator */
5130 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5132 struct ctl_table *entry, *table;
5133 struct sched_domain *sd;
5134 int domain_num = 0, i;
5137 for_each_domain(cpu, sd)
5139 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5144 for_each_domain(cpu, sd) {
5145 snprintf(buf, 32, "domain%d", i);
5146 entry->procname = kstrdup(buf, GFP_KERNEL);
5148 entry->child = sd_alloc_ctl_domain_table(sd);
5155 static struct ctl_table_header *sd_sysctl_header;
5156 static void register_sched_domain_sysctl(void)
5158 int i, cpu_num = num_possible_cpus();
5159 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5162 WARN_ON(sd_ctl_dir[0].child);
5163 sd_ctl_dir[0].child = entry;
5168 for_each_possible_cpu(i) {
5169 snprintf(buf, 32, "cpu%d", i);
5170 entry->procname = kstrdup(buf, GFP_KERNEL);
5172 entry->child = sd_alloc_ctl_cpu_table(i);
5176 WARN_ON(sd_sysctl_header);
5177 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5180 /* may be called multiple times per register */
5181 static void unregister_sched_domain_sysctl(void)
5183 if (sd_sysctl_header)
5184 unregister_sysctl_table(sd_sysctl_header);
5185 sd_sysctl_header = NULL;
5186 if (sd_ctl_dir[0].child)
5187 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5190 static void register_sched_domain_sysctl(void)
5193 static void unregister_sched_domain_sysctl(void)
5198 static void set_rq_online(struct rq *rq)
5201 const struct sched_class *class;
5203 cpumask_set_cpu(rq->cpu, rq->rd->online);
5206 for_each_class(class) {
5207 if (class->rq_online)
5208 class->rq_online(rq);
5213 static void set_rq_offline(struct rq *rq)
5216 const struct sched_class *class;
5218 for_each_class(class) {
5219 if (class->rq_offline)
5220 class->rq_offline(rq);
5223 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5229 * migration_call - callback that gets triggered when a CPU is added.
5230 * Here we can start up the necessary migration thread for the new CPU.
5232 static int __cpuinit
5233 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5235 int cpu = (long)hcpu;
5236 unsigned long flags;
5237 struct rq *rq = cpu_rq(cpu);
5239 switch (action & ~CPU_TASKS_FROZEN) {
5241 case CPU_UP_PREPARE:
5242 rq->calc_load_update = calc_load_update;
5246 /* Update our root-domain */
5247 raw_spin_lock_irqsave(&rq->lock, flags);
5249 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5253 raw_spin_unlock_irqrestore(&rq->lock, flags);
5256 #ifdef CONFIG_HOTPLUG_CPU
5258 sched_ttwu_pending();
5259 /* Update our root-domain */
5260 raw_spin_lock_irqsave(&rq->lock, flags);
5262 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5266 BUG_ON(rq->nr_running != 1); /* the migration thread */
5267 raw_spin_unlock_irqrestore(&rq->lock, flags);
5271 calc_load_migrate(rq);
5276 update_max_interval();
5282 * Register at high priority so that task migration (migrate_all_tasks)
5283 * happens before everything else. This has to be lower priority than
5284 * the notifier in the perf_event subsystem, though.
5286 static struct notifier_block __cpuinitdata migration_notifier = {
5287 .notifier_call = migration_call,
5288 .priority = CPU_PRI_MIGRATION,
5291 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5292 unsigned long action, void *hcpu)
5294 switch (action & ~CPU_TASKS_FROZEN) {
5296 case CPU_DOWN_FAILED:
5297 set_cpu_active((long)hcpu, true);
5304 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5305 unsigned long action, void *hcpu)
5307 switch (action & ~CPU_TASKS_FROZEN) {
5308 case CPU_DOWN_PREPARE:
5309 set_cpu_active((long)hcpu, false);
5316 static int __init migration_init(void)
5318 void *cpu = (void *)(long)smp_processor_id();
5321 /* Initialize migration for the boot CPU */
5322 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5323 BUG_ON(err == NOTIFY_BAD);
5324 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5325 register_cpu_notifier(&migration_notifier);
5327 /* Register cpu active notifiers */
5328 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5329 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5333 early_initcall(migration_init);
5338 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5340 #ifdef CONFIG_SCHED_DEBUG
5342 static __read_mostly int sched_debug_enabled;
5344 static int __init sched_debug_setup(char *str)
5346 sched_debug_enabled = 1;
5350 early_param("sched_debug", sched_debug_setup);
5352 static inline bool sched_debug(void)
5354 return sched_debug_enabled;
5357 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5358 struct cpumask *groupmask)
5360 struct sched_group *group = sd->groups;
5363 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5364 cpumask_clear(groupmask);
5366 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5368 if (!(sd->flags & SD_LOAD_BALANCE)) {
5369 printk("does not load-balance\n");
5371 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5376 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5378 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5379 printk(KERN_ERR "ERROR: domain->span does not contain "
5382 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5383 printk(KERN_ERR "ERROR: domain->groups does not contain"
5387 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5391 printk(KERN_ERR "ERROR: group is NULL\n");
5396 * Even though we initialize ->power to something semi-sane,
5397 * we leave power_orig unset. This allows us to detect if
5398 * domain iteration is still funny without causing /0 traps.
5400 if (!group->sgp->power_orig) {
5401 printk(KERN_CONT "\n");
5402 printk(KERN_ERR "ERROR: domain->cpu_power not "
5407 if (!cpumask_weight(sched_group_cpus(group))) {
5408 printk(KERN_CONT "\n");
5409 printk(KERN_ERR "ERROR: empty group\n");
5413 if (!(sd->flags & SD_OVERLAP) &&
5414 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5415 printk(KERN_CONT "\n");
5416 printk(KERN_ERR "ERROR: repeated CPUs\n");
5420 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5422 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5424 printk(KERN_CONT " %s", str);
5425 if (group->sgp->power != SCHED_POWER_SCALE) {
5426 printk(KERN_CONT " (cpu_power = %d)",
5430 group = group->next;
5431 } while (group != sd->groups);
5432 printk(KERN_CONT "\n");
5434 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5435 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5438 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5439 printk(KERN_ERR "ERROR: parent span is not a superset "
5440 "of domain->span\n");
5444 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5448 if (!sched_debug_enabled)
5452 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5456 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5459 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5467 #else /* !CONFIG_SCHED_DEBUG */
5468 # define sched_domain_debug(sd, cpu) do { } while (0)
5469 static inline bool sched_debug(void)
5473 #endif /* CONFIG_SCHED_DEBUG */
5475 static int sd_degenerate(struct sched_domain *sd)
5477 if (cpumask_weight(sched_domain_span(sd)) == 1)
5480 /* Following flags need at least 2 groups */
5481 if (sd->flags & (SD_LOAD_BALANCE |
5482 SD_BALANCE_NEWIDLE |
5486 SD_SHARE_PKG_RESOURCES)) {
5487 if (sd->groups != sd->groups->next)
5491 /* Following flags don't use groups */
5492 if (sd->flags & (SD_WAKE_AFFINE))
5499 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5501 unsigned long cflags = sd->flags, pflags = parent->flags;
5503 if (sd_degenerate(parent))
5506 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5509 /* Flags needing groups don't count if only 1 group in parent */
5510 if (parent->groups == parent->groups->next) {
5511 pflags &= ~(SD_LOAD_BALANCE |
5512 SD_BALANCE_NEWIDLE |
5516 SD_SHARE_PKG_RESOURCES);
5517 if (nr_node_ids == 1)
5518 pflags &= ~SD_SERIALIZE;
5520 if (~cflags & pflags)
5526 static void free_rootdomain(struct rcu_head *rcu)
5528 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5530 cpupri_cleanup(&rd->cpupri);
5531 free_cpumask_var(rd->rto_mask);
5532 free_cpumask_var(rd->online);
5533 free_cpumask_var(rd->span);
5537 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5539 struct root_domain *old_rd = NULL;
5540 unsigned long flags;
5542 raw_spin_lock_irqsave(&rq->lock, flags);
5547 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5550 cpumask_clear_cpu(rq->cpu, old_rd->span);
5553 * If we dont want to free the old_rt yet then
5554 * set old_rd to NULL to skip the freeing later
5557 if (!atomic_dec_and_test(&old_rd->refcount))
5561 atomic_inc(&rd->refcount);
5564 cpumask_set_cpu(rq->cpu, rd->span);
5565 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5568 raw_spin_unlock_irqrestore(&rq->lock, flags);
5571 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5574 static int init_rootdomain(struct root_domain *rd)
5576 memset(rd, 0, sizeof(*rd));
5578 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5580 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5582 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5585 if (cpupri_init(&rd->cpupri) != 0)
5590 free_cpumask_var(rd->rto_mask);
5592 free_cpumask_var(rd->online);
5594 free_cpumask_var(rd->span);
5600 * By default the system creates a single root-domain with all cpus as
5601 * members (mimicking the global state we have today).
5603 struct root_domain def_root_domain;
5605 static void init_defrootdomain(void)
5607 init_rootdomain(&def_root_domain);
5609 atomic_set(&def_root_domain.refcount, 1);
5612 static struct root_domain *alloc_rootdomain(void)
5614 struct root_domain *rd;
5616 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5620 if (init_rootdomain(rd) != 0) {
5628 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5630 struct sched_group *tmp, *first;
5639 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5644 } while (sg != first);
5647 static void free_sched_domain(struct rcu_head *rcu)
5649 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5652 * If its an overlapping domain it has private groups, iterate and
5655 if (sd->flags & SD_OVERLAP) {
5656 free_sched_groups(sd->groups, 1);
5657 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5658 kfree(sd->groups->sgp);
5664 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5666 call_rcu(&sd->rcu, free_sched_domain);
5669 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5671 for (; sd; sd = sd->parent)
5672 destroy_sched_domain(sd, cpu);
5676 * Keep a special pointer to the highest sched_domain that has
5677 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5678 * allows us to avoid some pointer chasing select_idle_sibling().
5680 * Also keep a unique ID per domain (we use the first cpu number in
5681 * the cpumask of the domain), this allows us to quickly tell if
5682 * two cpus are in the same cache domain, see cpus_share_cache().
5684 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5685 DEFINE_PER_CPU(int, sd_llc_id);
5687 static void update_top_cache_domain(int cpu)
5689 struct sched_domain *sd;
5692 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5694 id = cpumask_first(sched_domain_span(sd));
5696 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5697 per_cpu(sd_llc_id, cpu) = id;
5701 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5702 * hold the hotplug lock.
5705 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5707 struct rq *rq = cpu_rq(cpu);
5708 struct sched_domain *tmp;
5710 /* Remove the sched domains which do not contribute to scheduling. */
5711 for (tmp = sd; tmp; ) {
5712 struct sched_domain *parent = tmp->parent;
5716 if (sd_parent_degenerate(tmp, parent)) {
5717 tmp->parent = parent->parent;
5719 parent->parent->child = tmp;
5720 destroy_sched_domain(parent, cpu);
5725 if (sd && sd_degenerate(sd)) {
5728 destroy_sched_domain(tmp, cpu);
5733 sched_domain_debug(sd, cpu);
5735 rq_attach_root(rq, rd);
5737 rcu_assign_pointer(rq->sd, sd);
5738 destroy_sched_domains(tmp, cpu);
5740 update_top_cache_domain(cpu);
5743 /* cpus with isolated domains */
5744 static cpumask_var_t cpu_isolated_map;
5746 /* Setup the mask of cpus configured for isolated domains */
5747 static int __init isolated_cpu_setup(char *str)
5749 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5750 cpulist_parse(str, cpu_isolated_map);
5754 __setup("isolcpus=", isolated_cpu_setup);
5756 static const struct cpumask *cpu_cpu_mask(int cpu)
5758 return cpumask_of_node(cpu_to_node(cpu));
5762 struct sched_domain **__percpu sd;
5763 struct sched_group **__percpu sg;
5764 struct sched_group_power **__percpu sgp;
5768 struct sched_domain ** __percpu sd;
5769 struct root_domain *rd;
5779 struct sched_domain_topology_level;
5781 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5782 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5784 #define SDTL_OVERLAP 0x01
5786 struct sched_domain_topology_level {
5787 sched_domain_init_f init;
5788 sched_domain_mask_f mask;
5791 struct sd_data data;
5795 * Build an iteration mask that can exclude certain CPUs from the upwards
5798 * Asymmetric node setups can result in situations where the domain tree is of
5799 * unequal depth, make sure to skip domains that already cover the entire
5802 * In that case build_sched_domains() will have terminated the iteration early
5803 * and our sibling sd spans will be empty. Domains should always include the
5804 * cpu they're built on, so check that.
5807 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5809 const struct cpumask *span = sched_domain_span(sd);
5810 struct sd_data *sdd = sd->private;
5811 struct sched_domain *sibling;
5814 for_each_cpu(i, span) {
5815 sibling = *per_cpu_ptr(sdd->sd, i);
5816 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5819 cpumask_set_cpu(i, sched_group_mask(sg));
5824 * Return the canonical balance cpu for this group, this is the first cpu
5825 * of this group that's also in the iteration mask.
5827 int group_balance_cpu(struct sched_group *sg)
5829 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5833 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5835 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5836 const struct cpumask *span = sched_domain_span(sd);
5837 struct cpumask *covered = sched_domains_tmpmask;
5838 struct sd_data *sdd = sd->private;
5839 struct sched_domain *child;
5842 cpumask_clear(covered);
5844 for_each_cpu(i, span) {
5845 struct cpumask *sg_span;
5847 if (cpumask_test_cpu(i, covered))
5850 child = *per_cpu_ptr(sdd->sd, i);
5852 /* See the comment near build_group_mask(). */
5853 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5856 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5857 GFP_KERNEL, cpu_to_node(cpu));
5862 sg_span = sched_group_cpus(sg);
5864 child = child->child;
5865 cpumask_copy(sg_span, sched_domain_span(child));
5867 cpumask_set_cpu(i, sg_span);
5869 cpumask_or(covered, covered, sg_span);
5871 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5872 if (atomic_inc_return(&sg->sgp->ref) == 1)
5873 build_group_mask(sd, sg);
5876 * Initialize sgp->power such that even if we mess up the
5877 * domains and no possible iteration will get us here, we won't
5880 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5883 * Make sure the first group of this domain contains the
5884 * canonical balance cpu. Otherwise the sched_domain iteration
5885 * breaks. See update_sg_lb_stats().
5887 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5888 group_balance_cpu(sg) == cpu)
5898 sd->groups = groups;
5903 free_sched_groups(first, 0);
5908 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5910 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5911 struct sched_domain *child = sd->child;
5914 cpu = cpumask_first(sched_domain_span(child));
5917 *sg = *per_cpu_ptr(sdd->sg, cpu);
5918 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5919 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5926 * build_sched_groups will build a circular linked list of the groups
5927 * covered by the given span, and will set each group's ->cpumask correctly,
5928 * and ->cpu_power to 0.
5930 * Assumes the sched_domain tree is fully constructed
5933 build_sched_groups(struct sched_domain *sd, int cpu)
5935 struct sched_group *first = NULL, *last = NULL;
5936 struct sd_data *sdd = sd->private;
5937 const struct cpumask *span = sched_domain_span(sd);
5938 struct cpumask *covered;
5941 get_group(cpu, sdd, &sd->groups);
5942 atomic_inc(&sd->groups->ref);
5944 if (cpu != cpumask_first(sched_domain_span(sd)))
5947 lockdep_assert_held(&sched_domains_mutex);
5948 covered = sched_domains_tmpmask;
5950 cpumask_clear(covered);
5952 for_each_cpu(i, span) {
5953 struct sched_group *sg;
5954 int group = get_group(i, sdd, &sg);
5957 if (cpumask_test_cpu(i, covered))
5960 cpumask_clear(sched_group_cpus(sg));
5962 cpumask_setall(sched_group_mask(sg));
5964 for_each_cpu(j, span) {
5965 if (get_group(j, sdd, NULL) != group)
5968 cpumask_set_cpu(j, covered);
5969 cpumask_set_cpu(j, sched_group_cpus(sg));
5984 * Initialize sched groups cpu_power.
5986 * cpu_power indicates the capacity of sched group, which is used while
5987 * distributing the load between different sched groups in a sched domain.
5988 * Typically cpu_power for all the groups in a sched domain will be same unless
5989 * there are asymmetries in the topology. If there are asymmetries, group
5990 * having more cpu_power will pickup more load compared to the group having
5993 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5995 struct sched_group *sg = sd->groups;
5997 WARN_ON(!sd || !sg);
6000 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6002 } while (sg != sd->groups);
6004 if (cpu != group_balance_cpu(sg))
6007 update_group_power(sd, cpu);
6008 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
6011 int __weak arch_sd_sibling_asym_packing(void)
6013 return 0*SD_ASYM_PACKING;
6017 * Initializers for schedule domains
6018 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6021 #ifdef CONFIG_SCHED_DEBUG
6022 # define SD_INIT_NAME(sd, type) sd->name = #type
6024 # define SD_INIT_NAME(sd, type) do { } while (0)
6027 #define SD_INIT_FUNC(type) \
6028 static noinline struct sched_domain * \
6029 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
6031 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
6032 *sd = SD_##type##_INIT; \
6033 SD_INIT_NAME(sd, type); \
6034 sd->private = &tl->data; \
6039 #ifdef CONFIG_SCHED_SMT
6040 SD_INIT_FUNC(SIBLING)
6042 #ifdef CONFIG_SCHED_MC
6045 #ifdef CONFIG_SCHED_BOOK
6049 static int default_relax_domain_level = -1;
6050 int sched_domain_level_max;
6052 static int __init setup_relax_domain_level(char *str)
6054 if (kstrtoint(str, 0, &default_relax_domain_level))
6055 pr_warn("Unable to set relax_domain_level\n");
6059 __setup("relax_domain_level=", setup_relax_domain_level);
6061 static void set_domain_attribute(struct sched_domain *sd,
6062 struct sched_domain_attr *attr)
6066 if (!attr || attr->relax_domain_level < 0) {
6067 if (default_relax_domain_level < 0)
6070 request = default_relax_domain_level;
6072 request = attr->relax_domain_level;
6073 if (request < sd->level) {
6074 /* turn off idle balance on this domain */
6075 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6077 /* turn on idle balance on this domain */
6078 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6082 static void __sdt_free(const struct cpumask *cpu_map);
6083 static int __sdt_alloc(const struct cpumask *cpu_map);
6085 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6086 const struct cpumask *cpu_map)
6090 if (!atomic_read(&d->rd->refcount))
6091 free_rootdomain(&d->rd->rcu); /* fall through */
6093 free_percpu(d->sd); /* fall through */
6095 __sdt_free(cpu_map); /* fall through */
6101 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6102 const struct cpumask *cpu_map)
6104 memset(d, 0, sizeof(*d));
6106 if (__sdt_alloc(cpu_map))
6107 return sa_sd_storage;
6108 d->sd = alloc_percpu(struct sched_domain *);
6110 return sa_sd_storage;
6111 d->rd = alloc_rootdomain();
6114 return sa_rootdomain;
6118 * NULL the sd_data elements we've used to build the sched_domain and
6119 * sched_group structure so that the subsequent __free_domain_allocs()
6120 * will not free the data we're using.
6122 static void claim_allocations(int cpu, struct sched_domain *sd)
6124 struct sd_data *sdd = sd->private;
6126 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6127 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6129 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6130 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6132 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
6133 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
6136 #ifdef CONFIG_SCHED_SMT
6137 static const struct cpumask *cpu_smt_mask(int cpu)
6139 return topology_thread_cpumask(cpu);
6144 * Topology list, bottom-up.
6146 static struct sched_domain_topology_level default_topology[] = {
6147 #ifdef CONFIG_SCHED_SMT
6148 { sd_init_SIBLING, cpu_smt_mask, },
6150 #ifdef CONFIG_SCHED_MC
6151 { sd_init_MC, cpu_coregroup_mask, },
6153 #ifdef CONFIG_SCHED_BOOK
6154 { sd_init_BOOK, cpu_book_mask, },
6156 { sd_init_CPU, cpu_cpu_mask, },
6160 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6164 static int sched_domains_numa_levels;
6165 static int *sched_domains_numa_distance;
6166 static struct cpumask ***sched_domains_numa_masks;
6167 static int sched_domains_curr_level;
6169 static inline int sd_local_flags(int level)
6171 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6174 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6177 static struct sched_domain *
6178 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6180 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6181 int level = tl->numa_level;
6182 int sd_weight = cpumask_weight(
6183 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6185 *sd = (struct sched_domain){
6186 .min_interval = sd_weight,
6187 .max_interval = 2*sd_weight,
6189 .imbalance_pct = 125,
6190 .cache_nice_tries = 2,
6197 .flags = 1*SD_LOAD_BALANCE
6198 | 1*SD_BALANCE_NEWIDLE
6203 | 0*SD_SHARE_CPUPOWER
6204 | 0*SD_SHARE_PKG_RESOURCES
6206 | 0*SD_PREFER_SIBLING
6207 | sd_local_flags(level)
6209 .last_balance = jiffies,
6210 .balance_interval = sd_weight,
6212 SD_INIT_NAME(sd, NUMA);
6213 sd->private = &tl->data;
6216 * Ugly hack to pass state to sd_numa_mask()...
6218 sched_domains_curr_level = tl->numa_level;
6223 static const struct cpumask *sd_numa_mask(int cpu)
6225 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6228 static void sched_numa_warn(const char *str)
6230 static int done = false;
6238 printk(KERN_WARNING "ERROR: %s\n\n", str);
6240 for (i = 0; i < nr_node_ids; i++) {
6241 printk(KERN_WARNING " ");
6242 for (j = 0; j < nr_node_ids; j++)
6243 printk(KERN_CONT "%02d ", node_distance(i,j));
6244 printk(KERN_CONT "\n");
6246 printk(KERN_WARNING "\n");
6249 static bool find_numa_distance(int distance)
6253 if (distance == node_distance(0, 0))
6256 for (i = 0; i < sched_domains_numa_levels; i++) {
6257 if (sched_domains_numa_distance[i] == distance)
6264 static void sched_init_numa(void)
6266 int next_distance, curr_distance = node_distance(0, 0);
6267 struct sched_domain_topology_level *tl;
6271 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6272 if (!sched_domains_numa_distance)
6276 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6277 * unique distances in the node_distance() table.
6279 * Assumes node_distance(0,j) includes all distances in
6280 * node_distance(i,j) in order to avoid cubic time.
6282 next_distance = curr_distance;
6283 for (i = 0; i < nr_node_ids; i++) {
6284 for (j = 0; j < nr_node_ids; j++) {
6285 for (k = 0; k < nr_node_ids; k++) {
6286 int distance = node_distance(i, k);
6288 if (distance > curr_distance &&
6289 (distance < next_distance ||
6290 next_distance == curr_distance))
6291 next_distance = distance;
6294 * While not a strong assumption it would be nice to know
6295 * about cases where if node A is connected to B, B is not
6296 * equally connected to A.
6298 if (sched_debug() && node_distance(k, i) != distance)
6299 sched_numa_warn("Node-distance not symmetric");
6301 if (sched_debug() && i && !find_numa_distance(distance))
6302 sched_numa_warn("Node-0 not representative");
6304 if (next_distance != curr_distance) {
6305 sched_domains_numa_distance[level++] = next_distance;
6306 sched_domains_numa_levels = level;
6307 curr_distance = next_distance;
6312 * In case of sched_debug() we verify the above assumption.
6318 * 'level' contains the number of unique distances, excluding the
6319 * identity distance node_distance(i,i).
6321 * The sched_domains_numa_distance[] array includes the actual distance
6326 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6327 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6328 * the array will contain less then 'level' members. This could be
6329 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6330 * in other functions.
6332 * We reset it to 'level' at the end of this function.
6334 sched_domains_numa_levels = 0;
6336 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6337 if (!sched_domains_numa_masks)
6341 * Now for each level, construct a mask per node which contains all
6342 * cpus of nodes that are that many hops away from us.
6344 for (i = 0; i < level; i++) {
6345 sched_domains_numa_masks[i] =
6346 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6347 if (!sched_domains_numa_masks[i])
6350 for (j = 0; j < nr_node_ids; j++) {
6351 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6355 sched_domains_numa_masks[i][j] = mask;
6357 for (k = 0; k < nr_node_ids; k++) {
6358 if (node_distance(j, k) > sched_domains_numa_distance[i])
6361 cpumask_or(mask, mask, cpumask_of_node(k));
6366 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6367 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6372 * Copy the default topology bits..
6374 for (i = 0; default_topology[i].init; i++)
6375 tl[i] = default_topology[i];
6378 * .. and append 'j' levels of NUMA goodness.
6380 for (j = 0; j < level; i++, j++) {
6381 tl[i] = (struct sched_domain_topology_level){
6382 .init = sd_numa_init,
6383 .mask = sd_numa_mask,
6384 .flags = SDTL_OVERLAP,
6389 sched_domain_topology = tl;
6391 sched_domains_numa_levels = level;
6394 static void sched_domains_numa_masks_set(int cpu)
6397 int node = cpu_to_node(cpu);
6399 for (i = 0; i < sched_domains_numa_levels; i++) {
6400 for (j = 0; j < nr_node_ids; j++) {
6401 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6402 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6407 static void sched_domains_numa_masks_clear(int cpu)
6410 for (i = 0; i < sched_domains_numa_levels; i++) {
6411 for (j = 0; j < nr_node_ids; j++)
6412 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6417 * Update sched_domains_numa_masks[level][node] array when new cpus
6420 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6421 unsigned long action,
6424 int cpu = (long)hcpu;
6426 switch (action & ~CPU_TASKS_FROZEN) {
6428 sched_domains_numa_masks_set(cpu);
6432 sched_domains_numa_masks_clear(cpu);
6442 static inline void sched_init_numa(void)
6446 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6447 unsigned long action,
6452 #endif /* CONFIG_NUMA */
6454 static int __sdt_alloc(const struct cpumask *cpu_map)
6456 struct sched_domain_topology_level *tl;
6459 for (tl = sched_domain_topology; tl->init; tl++) {
6460 struct sd_data *sdd = &tl->data;
6462 sdd->sd = alloc_percpu(struct sched_domain *);
6466 sdd->sg = alloc_percpu(struct sched_group *);
6470 sdd->sgp = alloc_percpu(struct sched_group_power *);
6474 for_each_cpu(j, cpu_map) {
6475 struct sched_domain *sd;
6476 struct sched_group *sg;
6477 struct sched_group_power *sgp;
6479 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6480 GFP_KERNEL, cpu_to_node(j));
6484 *per_cpu_ptr(sdd->sd, j) = sd;
6486 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6487 GFP_KERNEL, cpu_to_node(j));
6493 *per_cpu_ptr(sdd->sg, j) = sg;
6495 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6496 GFP_KERNEL, cpu_to_node(j));
6500 *per_cpu_ptr(sdd->sgp, j) = sgp;
6507 static void __sdt_free(const struct cpumask *cpu_map)
6509 struct sched_domain_topology_level *tl;
6512 for (tl = sched_domain_topology; tl->init; tl++) {
6513 struct sd_data *sdd = &tl->data;
6515 for_each_cpu(j, cpu_map) {
6516 struct sched_domain *sd;
6519 sd = *per_cpu_ptr(sdd->sd, j);
6520 if (sd && (sd->flags & SD_OVERLAP))
6521 free_sched_groups(sd->groups, 0);
6522 kfree(*per_cpu_ptr(sdd->sd, j));
6526 kfree(*per_cpu_ptr(sdd->sg, j));
6528 kfree(*per_cpu_ptr(sdd->sgp, j));
6530 free_percpu(sdd->sd);
6532 free_percpu(sdd->sg);
6534 free_percpu(sdd->sgp);
6539 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6540 struct s_data *d, const struct cpumask *cpu_map,
6541 struct sched_domain_attr *attr, struct sched_domain *child,
6544 struct sched_domain *sd = tl->init(tl, cpu);
6548 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6550 sd->level = child->level + 1;
6551 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6555 set_domain_attribute(sd, attr);
6561 * Build sched domains for a given set of cpus and attach the sched domains
6562 * to the individual cpus
6564 static int build_sched_domains(const struct cpumask *cpu_map,
6565 struct sched_domain_attr *attr)
6567 enum s_alloc alloc_state = sa_none;
6568 struct sched_domain *sd;
6570 int i, ret = -ENOMEM;
6572 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6573 if (alloc_state != sa_rootdomain)
6576 /* Set up domains for cpus specified by the cpu_map. */
6577 for_each_cpu(i, cpu_map) {
6578 struct sched_domain_topology_level *tl;
6581 for (tl = sched_domain_topology; tl->init; tl++) {
6582 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
6583 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6584 sd->flags |= SD_OVERLAP;
6585 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6592 *per_cpu_ptr(d.sd, i) = sd;
6595 /* Build the groups for the domains */
6596 for_each_cpu(i, cpu_map) {
6597 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6598 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6599 if (sd->flags & SD_OVERLAP) {
6600 if (build_overlap_sched_groups(sd, i))
6603 if (build_sched_groups(sd, i))
6609 /* Calculate CPU power for physical packages and nodes */
6610 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6611 if (!cpumask_test_cpu(i, cpu_map))
6614 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6615 claim_allocations(i, sd);
6616 init_sched_groups_power(i, sd);
6620 /* Attach the domains */
6622 for_each_cpu(i, cpu_map) {
6623 sd = *per_cpu_ptr(d.sd, i);
6624 cpu_attach_domain(sd, d.rd, i);
6630 __free_domain_allocs(&d, alloc_state, cpu_map);
6634 static cpumask_var_t *doms_cur; /* current sched domains */
6635 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6636 static struct sched_domain_attr *dattr_cur;
6637 /* attribues of custom domains in 'doms_cur' */
6640 * Special case: If a kmalloc of a doms_cur partition (array of
6641 * cpumask) fails, then fallback to a single sched domain,
6642 * as determined by the single cpumask fallback_doms.
6644 static cpumask_var_t fallback_doms;
6647 * arch_update_cpu_topology lets virtualized architectures update the
6648 * cpu core maps. It is supposed to return 1 if the topology changed
6649 * or 0 if it stayed the same.
6651 int __attribute__((weak)) arch_update_cpu_topology(void)
6656 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6659 cpumask_var_t *doms;
6661 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6664 for (i = 0; i < ndoms; i++) {
6665 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6666 free_sched_domains(doms, i);
6673 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6676 for (i = 0; i < ndoms; i++)
6677 free_cpumask_var(doms[i]);
6682 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6683 * For now this just excludes isolated cpus, but could be used to
6684 * exclude other special cases in the future.
6686 static int init_sched_domains(const struct cpumask *cpu_map)
6690 arch_update_cpu_topology();
6692 doms_cur = alloc_sched_domains(ndoms_cur);
6694 doms_cur = &fallback_doms;
6695 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6696 err = build_sched_domains(doms_cur[0], NULL);
6697 register_sched_domain_sysctl();
6703 * Detach sched domains from a group of cpus specified in cpu_map
6704 * These cpus will now be attached to the NULL domain
6706 static void detach_destroy_domains(const struct cpumask *cpu_map)
6711 for_each_cpu(i, cpu_map)
6712 cpu_attach_domain(NULL, &def_root_domain, i);
6716 /* handle null as "default" */
6717 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6718 struct sched_domain_attr *new, int idx_new)
6720 struct sched_domain_attr tmp;
6727 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6728 new ? (new + idx_new) : &tmp,
6729 sizeof(struct sched_domain_attr));
6733 * Partition sched domains as specified by the 'ndoms_new'
6734 * cpumasks in the array doms_new[] of cpumasks. This compares
6735 * doms_new[] to the current sched domain partitioning, doms_cur[].
6736 * It destroys each deleted domain and builds each new domain.
6738 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6739 * The masks don't intersect (don't overlap.) We should setup one
6740 * sched domain for each mask. CPUs not in any of the cpumasks will
6741 * not be load balanced. If the same cpumask appears both in the
6742 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6745 * The passed in 'doms_new' should be allocated using
6746 * alloc_sched_domains. This routine takes ownership of it and will
6747 * free_sched_domains it when done with it. If the caller failed the
6748 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6749 * and partition_sched_domains() will fallback to the single partition
6750 * 'fallback_doms', it also forces the domains to be rebuilt.
6752 * If doms_new == NULL it will be replaced with cpu_online_mask.
6753 * ndoms_new == 0 is a special case for destroying existing domains,
6754 * and it will not create the default domain.
6756 * Call with hotplug lock held
6758 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6759 struct sched_domain_attr *dattr_new)
6764 mutex_lock(&sched_domains_mutex);
6766 /* always unregister in case we don't destroy any domains */
6767 unregister_sched_domain_sysctl();
6769 /* Let architecture update cpu core mappings. */
6770 new_topology = arch_update_cpu_topology();
6772 n = doms_new ? ndoms_new : 0;
6774 /* Destroy deleted domains */
6775 for (i = 0; i < ndoms_cur; i++) {
6776 for (j = 0; j < n && !new_topology; j++) {
6777 if (cpumask_equal(doms_cur[i], doms_new[j])
6778 && dattrs_equal(dattr_cur, i, dattr_new, j))
6781 /* no match - a current sched domain not in new doms_new[] */
6782 detach_destroy_domains(doms_cur[i]);
6787 if (doms_new == NULL) {
6789 doms_new = &fallback_doms;
6790 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6791 WARN_ON_ONCE(dattr_new);
6794 /* Build new domains */
6795 for (i = 0; i < ndoms_new; i++) {
6796 for (j = 0; j < ndoms_cur && !new_topology; j++) {
6797 if (cpumask_equal(doms_new[i], doms_cur[j])
6798 && dattrs_equal(dattr_new, i, dattr_cur, j))
6801 /* no match - add a new doms_new */
6802 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6807 /* Remember the new sched domains */
6808 if (doms_cur != &fallback_doms)
6809 free_sched_domains(doms_cur, ndoms_cur);
6810 kfree(dattr_cur); /* kfree(NULL) is safe */
6811 doms_cur = doms_new;
6812 dattr_cur = dattr_new;
6813 ndoms_cur = ndoms_new;
6815 register_sched_domain_sysctl();
6817 mutex_unlock(&sched_domains_mutex);
6820 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6823 * Update cpusets according to cpu_active mask. If cpusets are
6824 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6825 * around partition_sched_domains().
6827 * If we come here as part of a suspend/resume, don't touch cpusets because we
6828 * want to restore it back to its original state upon resume anyway.
6830 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6834 case CPU_ONLINE_FROZEN:
6835 case CPU_DOWN_FAILED_FROZEN:
6838 * num_cpus_frozen tracks how many CPUs are involved in suspend
6839 * resume sequence. As long as this is not the last online
6840 * operation in the resume sequence, just build a single sched
6841 * domain, ignoring cpusets.
6844 if (likely(num_cpus_frozen)) {
6845 partition_sched_domains(1, NULL, NULL);
6850 * This is the last CPU online operation. So fall through and
6851 * restore the original sched domains by considering the
6852 * cpuset configurations.
6856 case CPU_DOWN_FAILED:
6857 cpuset_update_active_cpus(true);
6865 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6869 case CPU_DOWN_PREPARE:
6870 cpuset_update_active_cpus(false);
6872 case CPU_DOWN_PREPARE_FROZEN:
6874 partition_sched_domains(1, NULL, NULL);
6882 void __init sched_init_smp(void)
6884 cpumask_var_t non_isolated_cpus;
6886 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6887 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6892 mutex_lock(&sched_domains_mutex);
6893 init_sched_domains(cpu_active_mask);
6894 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6895 if (cpumask_empty(non_isolated_cpus))
6896 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6897 mutex_unlock(&sched_domains_mutex);
6900 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6901 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6902 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6904 /* RT runtime code needs to handle some hotplug events */
6905 hotcpu_notifier(update_runtime, 0);
6909 /* Move init over to a non-isolated CPU */
6910 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6912 sched_init_granularity();
6913 free_cpumask_var(non_isolated_cpus);
6915 init_sched_rt_class();
6918 void __init sched_init_smp(void)
6920 sched_init_granularity();
6922 #endif /* CONFIG_SMP */
6924 const_debug unsigned int sysctl_timer_migration = 1;
6926 int in_sched_functions(unsigned long addr)
6928 return in_lock_functions(addr) ||
6929 (addr >= (unsigned long)__sched_text_start
6930 && addr < (unsigned long)__sched_text_end);
6933 #ifdef CONFIG_CGROUP_SCHED
6935 * Default task group.
6936 * Every task in system belongs to this group at bootup.
6938 struct task_group root_task_group;
6939 LIST_HEAD(task_groups);
6942 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6944 void __init sched_init(void)
6947 unsigned long alloc_size = 0, ptr;
6949 #ifdef CONFIG_FAIR_GROUP_SCHED
6950 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6952 #ifdef CONFIG_RT_GROUP_SCHED
6953 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6955 #ifdef CONFIG_CPUMASK_OFFSTACK
6956 alloc_size += num_possible_cpus() * cpumask_size();
6959 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6961 #ifdef CONFIG_FAIR_GROUP_SCHED
6962 root_task_group.se = (struct sched_entity **)ptr;
6963 ptr += nr_cpu_ids * sizeof(void **);
6965 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6966 ptr += nr_cpu_ids * sizeof(void **);
6968 #endif /* CONFIG_FAIR_GROUP_SCHED */
6969 #ifdef CONFIG_RT_GROUP_SCHED
6970 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6971 ptr += nr_cpu_ids * sizeof(void **);
6973 root_task_group.rt_rq = (struct rt_rq **)ptr;
6974 ptr += nr_cpu_ids * sizeof(void **);
6976 #endif /* CONFIG_RT_GROUP_SCHED */
6977 #ifdef CONFIG_CPUMASK_OFFSTACK
6978 for_each_possible_cpu(i) {
6979 per_cpu(load_balance_mask, i) = (void *)ptr;
6980 ptr += cpumask_size();
6982 #endif /* CONFIG_CPUMASK_OFFSTACK */
6986 init_defrootdomain();
6989 init_rt_bandwidth(&def_rt_bandwidth,
6990 global_rt_period(), global_rt_runtime());
6992 #ifdef CONFIG_RT_GROUP_SCHED
6993 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6994 global_rt_period(), global_rt_runtime());
6995 #endif /* CONFIG_RT_GROUP_SCHED */
6997 #ifdef CONFIG_CGROUP_SCHED
6998 list_add(&root_task_group.list, &task_groups);
6999 INIT_LIST_HEAD(&root_task_group.children);
7000 INIT_LIST_HEAD(&root_task_group.siblings);
7001 autogroup_init(&init_task);
7003 #endif /* CONFIG_CGROUP_SCHED */
7005 for_each_possible_cpu(i) {
7009 raw_spin_lock_init(&rq->lock);
7011 rq->calc_load_active = 0;
7012 rq->calc_load_update = jiffies + LOAD_FREQ;
7013 init_cfs_rq(&rq->cfs);
7014 init_rt_rq(&rq->rt, rq);
7015 #ifdef CONFIG_FAIR_GROUP_SCHED
7016 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7017 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7019 * How much cpu bandwidth does root_task_group get?
7021 * In case of task-groups formed thr' the cgroup filesystem, it
7022 * gets 100% of the cpu resources in the system. This overall
7023 * system cpu resource is divided among the tasks of
7024 * root_task_group and its child task-groups in a fair manner,
7025 * based on each entity's (task or task-group's) weight
7026 * (se->load.weight).
7028 * In other words, if root_task_group has 10 tasks of weight
7029 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7030 * then A0's share of the cpu resource is:
7032 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7034 * We achieve this by letting root_task_group's tasks sit
7035 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7037 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7038 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7039 #endif /* CONFIG_FAIR_GROUP_SCHED */
7041 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7042 #ifdef CONFIG_RT_GROUP_SCHED
7043 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7044 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7047 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7048 rq->cpu_load[j] = 0;
7050 rq->last_load_update_tick = jiffies;
7055 rq->cpu_power = SCHED_POWER_SCALE;
7056 rq->post_schedule = 0;
7057 rq->active_balance = 0;
7058 rq->next_balance = jiffies;
7063 rq->avg_idle = 2*sysctl_sched_migration_cost;
7065 INIT_LIST_HEAD(&rq->cfs_tasks);
7067 rq_attach_root(rq, &def_root_domain);
7068 #ifdef CONFIG_NO_HZ_COMMON
7071 #ifdef CONFIG_NO_HZ_FULL
7072 rq->last_sched_tick = 0;
7076 atomic_set(&rq->nr_iowait, 0);
7079 set_load_weight(&init_task);
7081 #ifdef CONFIG_PREEMPT_NOTIFIERS
7082 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7085 #ifdef CONFIG_RT_MUTEXES
7086 plist_head_init(&init_task.pi_waiters);
7090 * The boot idle thread does lazy MMU switching as well:
7092 atomic_inc(&init_mm.mm_count);
7093 enter_lazy_tlb(&init_mm, current);
7096 * Make us the idle thread. Technically, schedule() should not be
7097 * called from this thread, however somewhere below it might be,
7098 * but because we are the idle thread, we just pick up running again
7099 * when this runqueue becomes "idle".
7101 init_idle(current, smp_processor_id());
7103 calc_load_update = jiffies + LOAD_FREQ;
7106 * During early bootup we pretend to be a normal task:
7108 current->sched_class = &fair_sched_class;
7111 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7112 /* May be allocated at isolcpus cmdline parse time */
7113 if (cpu_isolated_map == NULL)
7114 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7115 idle_thread_set_boot_cpu();
7117 init_sched_fair_class();
7119 scheduler_running = 1;
7122 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7123 static inline int preempt_count_equals(int preempt_offset)
7125 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7127 return (nested == preempt_offset);
7130 void __might_sleep(const char *file, int line, int preempt_offset)
7132 static unsigned long prev_jiffy; /* ratelimiting */
7134 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7135 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7136 system_state != SYSTEM_RUNNING || oops_in_progress)
7138 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7140 prev_jiffy = jiffies;
7143 "BUG: sleeping function called from invalid context at %s:%d\n",
7146 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7147 in_atomic(), irqs_disabled(),
7148 current->pid, current->comm);
7150 debug_show_held_locks(current);
7151 if (irqs_disabled())
7152 print_irqtrace_events(current);
7155 EXPORT_SYMBOL(__might_sleep);
7158 #ifdef CONFIG_MAGIC_SYSRQ
7159 static void normalize_task(struct rq *rq, struct task_struct *p)
7161 const struct sched_class *prev_class = p->sched_class;
7162 int old_prio = p->prio;
7167 dequeue_task(rq, p, 0);
7168 __setscheduler(rq, p, SCHED_NORMAL, 0);
7170 enqueue_task(rq, p, 0);
7171 resched_task(rq->curr);
7174 check_class_changed(rq, p, prev_class, old_prio);
7177 void normalize_rt_tasks(void)
7179 struct task_struct *g, *p;
7180 unsigned long flags;
7183 read_lock_irqsave(&tasklist_lock, flags);
7184 do_each_thread(g, p) {
7186 * Only normalize user tasks:
7191 p->se.exec_start = 0;
7192 #ifdef CONFIG_SCHEDSTATS
7193 p->se.statistics.wait_start = 0;
7194 p->se.statistics.sleep_start = 0;
7195 p->se.statistics.block_start = 0;
7200 * Renice negative nice level userspace
7203 if (TASK_NICE(p) < 0 && p->mm)
7204 set_user_nice(p, 0);
7208 raw_spin_lock(&p->pi_lock);
7209 rq = __task_rq_lock(p);
7211 normalize_task(rq, p);
7213 __task_rq_unlock(rq);
7214 raw_spin_unlock(&p->pi_lock);
7215 } while_each_thread(g, p);
7217 read_unlock_irqrestore(&tasklist_lock, flags);
7220 #endif /* CONFIG_MAGIC_SYSRQ */
7222 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7224 * These functions are only useful for the IA64 MCA handling, or kdb.
7226 * They can only be called when the whole system has been
7227 * stopped - every CPU needs to be quiescent, and no scheduling
7228 * activity can take place. Using them for anything else would
7229 * be a serious bug, and as a result, they aren't even visible
7230 * under any other configuration.
7234 * curr_task - return the current task for a given cpu.
7235 * @cpu: the processor in question.
7237 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7239 struct task_struct *curr_task(int cpu)
7241 return cpu_curr(cpu);
7244 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7248 * set_curr_task - set the current task for a given cpu.
7249 * @cpu: the processor in question.
7250 * @p: the task pointer to set.
7252 * Description: This function must only be used when non-maskable interrupts
7253 * are serviced on a separate stack. It allows the architecture to switch the
7254 * notion of the current task on a cpu in a non-blocking manner. This function
7255 * must be called with all CPU's synchronized, and interrupts disabled, the
7256 * and caller must save the original value of the current task (see
7257 * curr_task() above) and restore that value before reenabling interrupts and
7258 * re-starting the system.
7260 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7262 void set_curr_task(int cpu, struct task_struct *p)
7269 #ifdef CONFIG_CGROUP_SCHED
7270 /* task_group_lock serializes the addition/removal of task groups */
7271 static DEFINE_SPINLOCK(task_group_lock);
7273 static void free_sched_group(struct task_group *tg)
7275 free_fair_sched_group(tg);
7276 free_rt_sched_group(tg);
7281 /* allocate runqueue etc for a new task group */
7282 struct task_group *sched_create_group(struct task_group *parent)
7284 struct task_group *tg;
7286 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7288 return ERR_PTR(-ENOMEM);
7290 if (!alloc_fair_sched_group(tg, parent))
7293 if (!alloc_rt_sched_group(tg, parent))
7299 free_sched_group(tg);
7300 return ERR_PTR(-ENOMEM);
7303 void sched_online_group(struct task_group *tg, struct task_group *parent)
7305 unsigned long flags;
7307 spin_lock_irqsave(&task_group_lock, flags);
7308 list_add_rcu(&tg->list, &task_groups);
7310 WARN_ON(!parent); /* root should already exist */
7312 tg->parent = parent;
7313 INIT_LIST_HEAD(&tg->children);
7314 list_add_rcu(&tg->siblings, &parent->children);
7315 spin_unlock_irqrestore(&task_group_lock, flags);
7318 /* rcu callback to free various structures associated with a task group */
7319 static void free_sched_group_rcu(struct rcu_head *rhp)
7321 /* now it should be safe to free those cfs_rqs */
7322 free_sched_group(container_of(rhp, struct task_group, rcu));
7325 /* Destroy runqueue etc associated with a task group */
7326 void sched_destroy_group(struct task_group *tg)
7328 /* wait for possible concurrent references to cfs_rqs complete */
7329 call_rcu(&tg->rcu, free_sched_group_rcu);
7332 void sched_offline_group(struct task_group *tg)
7334 unsigned long flags;
7337 /* end participation in shares distribution */
7338 for_each_possible_cpu(i)
7339 unregister_fair_sched_group(tg, i);
7341 spin_lock_irqsave(&task_group_lock, flags);
7342 list_del_rcu(&tg->list);
7343 list_del_rcu(&tg->siblings);
7344 spin_unlock_irqrestore(&task_group_lock, flags);
7347 /* change task's runqueue when it moves between groups.
7348 * The caller of this function should have put the task in its new group
7349 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7350 * reflect its new group.
7352 void sched_move_task(struct task_struct *tsk)
7354 struct task_group *tg;
7356 unsigned long flags;
7359 rq = task_rq_lock(tsk, &flags);
7361 running = task_current(rq, tsk);
7365 dequeue_task(rq, tsk, 0);
7366 if (unlikely(running))
7367 tsk->sched_class->put_prev_task(rq, tsk);
7369 tg = container_of(task_subsys_state_check(tsk, cpu_cgroup_subsys_id,
7370 lockdep_is_held(&tsk->sighand->siglock)),
7371 struct task_group, css);
7372 tg = autogroup_task_group(tsk, tg);
7373 tsk->sched_task_group = tg;
7375 #ifdef CONFIG_FAIR_GROUP_SCHED
7376 if (tsk->sched_class->task_move_group)
7377 tsk->sched_class->task_move_group(tsk, on_rq);
7380 set_task_rq(tsk, task_cpu(tsk));
7382 if (unlikely(running))
7383 tsk->sched_class->set_curr_task(rq);
7385 enqueue_task(rq, tsk, 0);
7387 task_rq_unlock(rq, tsk, &flags);
7389 #endif /* CONFIG_CGROUP_SCHED */
7391 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
7392 static unsigned long to_ratio(u64 period, u64 runtime)
7394 if (runtime == RUNTIME_INF)
7397 return div64_u64(runtime << 20, period);
7401 #ifdef CONFIG_RT_GROUP_SCHED
7403 * Ensure that the real time constraints are schedulable.
7405 static DEFINE_MUTEX(rt_constraints_mutex);
7407 /* Must be called with tasklist_lock held */
7408 static inline int tg_has_rt_tasks(struct task_group *tg)
7410 struct task_struct *g, *p;
7412 do_each_thread(g, p) {
7413 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7415 } while_each_thread(g, p);
7420 struct rt_schedulable_data {
7421 struct task_group *tg;
7426 static int tg_rt_schedulable(struct task_group *tg, void *data)
7428 struct rt_schedulable_data *d = data;
7429 struct task_group *child;
7430 unsigned long total, sum = 0;
7431 u64 period, runtime;
7433 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7434 runtime = tg->rt_bandwidth.rt_runtime;
7437 period = d->rt_period;
7438 runtime = d->rt_runtime;
7442 * Cannot have more runtime than the period.
7444 if (runtime > period && runtime != RUNTIME_INF)
7448 * Ensure we don't starve existing RT tasks.
7450 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7453 total = to_ratio(period, runtime);
7456 * Nobody can have more than the global setting allows.
7458 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7462 * The sum of our children's runtime should not exceed our own.
7464 list_for_each_entry_rcu(child, &tg->children, siblings) {
7465 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7466 runtime = child->rt_bandwidth.rt_runtime;
7468 if (child == d->tg) {
7469 period = d->rt_period;
7470 runtime = d->rt_runtime;
7473 sum += to_ratio(period, runtime);
7482 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7486 struct rt_schedulable_data data = {
7488 .rt_period = period,
7489 .rt_runtime = runtime,
7493 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7499 static int tg_set_rt_bandwidth(struct task_group *tg,
7500 u64 rt_period, u64 rt_runtime)
7504 mutex_lock(&rt_constraints_mutex);
7505 read_lock(&tasklist_lock);
7506 err = __rt_schedulable(tg, rt_period, rt_runtime);
7510 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7511 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7512 tg->rt_bandwidth.rt_runtime = rt_runtime;
7514 for_each_possible_cpu(i) {
7515 struct rt_rq *rt_rq = tg->rt_rq[i];
7517 raw_spin_lock(&rt_rq->rt_runtime_lock);
7518 rt_rq->rt_runtime = rt_runtime;
7519 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7521 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7523 read_unlock(&tasklist_lock);
7524 mutex_unlock(&rt_constraints_mutex);
7529 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7531 u64 rt_runtime, rt_period;
7533 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7534 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7535 if (rt_runtime_us < 0)
7536 rt_runtime = RUNTIME_INF;
7538 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7541 static long sched_group_rt_runtime(struct task_group *tg)
7545 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7548 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7549 do_div(rt_runtime_us, NSEC_PER_USEC);
7550 return rt_runtime_us;
7553 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7555 u64 rt_runtime, rt_period;
7557 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7558 rt_runtime = tg->rt_bandwidth.rt_runtime;
7563 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7566 static long sched_group_rt_period(struct task_group *tg)
7570 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7571 do_div(rt_period_us, NSEC_PER_USEC);
7572 return rt_period_us;
7575 static int sched_rt_global_constraints(void)
7577 u64 runtime, period;
7580 if (sysctl_sched_rt_period <= 0)
7583 runtime = global_rt_runtime();
7584 period = global_rt_period();
7587 * Sanity check on the sysctl variables.
7589 if (runtime > period && runtime != RUNTIME_INF)
7592 mutex_lock(&rt_constraints_mutex);
7593 read_lock(&tasklist_lock);
7594 ret = __rt_schedulable(NULL, 0, 0);
7595 read_unlock(&tasklist_lock);
7596 mutex_unlock(&rt_constraints_mutex);
7601 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7603 /* Don't accept realtime tasks when there is no way for them to run */
7604 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7610 #else /* !CONFIG_RT_GROUP_SCHED */
7611 static int sched_rt_global_constraints(void)
7613 unsigned long flags;
7616 if (sysctl_sched_rt_period <= 0)
7620 * There's always some RT tasks in the root group
7621 * -- migration, kstopmachine etc..
7623 if (sysctl_sched_rt_runtime == 0)
7626 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7627 for_each_possible_cpu(i) {
7628 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7630 raw_spin_lock(&rt_rq->rt_runtime_lock);
7631 rt_rq->rt_runtime = global_rt_runtime();
7632 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7634 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7638 #endif /* CONFIG_RT_GROUP_SCHED */
7640 int sched_rr_handler(struct ctl_table *table, int write,
7641 void __user *buffer, size_t *lenp,
7645 static DEFINE_MUTEX(mutex);
7648 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7649 /* make sure that internally we keep jiffies */
7650 /* also, writing zero resets timeslice to default */
7651 if (!ret && write) {
7652 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7653 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7655 mutex_unlock(&mutex);
7659 int sched_rt_handler(struct ctl_table *table, int write,
7660 void __user *buffer, size_t *lenp,
7664 int old_period, old_runtime;
7665 static DEFINE_MUTEX(mutex);
7668 old_period = sysctl_sched_rt_period;
7669 old_runtime = sysctl_sched_rt_runtime;
7671 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7673 if (!ret && write) {
7674 ret = sched_rt_global_constraints();
7676 sysctl_sched_rt_period = old_period;
7677 sysctl_sched_rt_runtime = old_runtime;
7679 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7680 def_rt_bandwidth.rt_period =
7681 ns_to_ktime(global_rt_period());
7684 mutex_unlock(&mutex);
7689 #ifdef CONFIG_CGROUP_SCHED
7691 /* return corresponding task_group object of a cgroup */
7692 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
7694 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
7695 struct task_group, css);
7698 static struct cgroup_subsys_state *cpu_cgroup_css_alloc(struct cgroup *cgrp)
7700 struct task_group *tg, *parent;
7702 if (!cgrp->parent) {
7703 /* This is early initialization for the top cgroup */
7704 return &root_task_group.css;
7707 parent = cgroup_tg(cgrp->parent);
7708 tg = sched_create_group(parent);
7710 return ERR_PTR(-ENOMEM);
7715 static int cpu_cgroup_css_online(struct cgroup *cgrp)
7717 struct task_group *tg = cgroup_tg(cgrp);
7718 struct task_group *parent;
7723 parent = cgroup_tg(cgrp->parent);
7724 sched_online_group(tg, parent);
7728 static void cpu_cgroup_css_free(struct cgroup *cgrp)
7730 struct task_group *tg = cgroup_tg(cgrp);
7732 sched_destroy_group(tg);
7735 static void cpu_cgroup_css_offline(struct cgroup *cgrp)
7737 struct task_group *tg = cgroup_tg(cgrp);
7739 sched_offline_group(tg);
7742 static int cpu_cgroup_can_attach(struct cgroup *cgrp,
7743 struct cgroup_taskset *tset)
7745 struct task_struct *task;
7747 cgroup_taskset_for_each(task, cgrp, tset) {
7748 #ifdef CONFIG_RT_GROUP_SCHED
7749 if (!sched_rt_can_attach(cgroup_tg(cgrp), task))
7752 /* We don't support RT-tasks being in separate groups */
7753 if (task->sched_class != &fair_sched_class)
7760 static void cpu_cgroup_attach(struct cgroup *cgrp,
7761 struct cgroup_taskset *tset)
7763 struct task_struct *task;
7765 cgroup_taskset_for_each(task, cgrp, tset)
7766 sched_move_task(task);
7770 cpu_cgroup_exit(struct cgroup *cgrp, struct cgroup *old_cgrp,
7771 struct task_struct *task)
7774 * cgroup_exit() is called in the copy_process() failure path.
7775 * Ignore this case since the task hasn't ran yet, this avoids
7776 * trying to poke a half freed task state from generic code.
7778 if (!(task->flags & PF_EXITING))
7781 sched_move_task(task);
7784 #ifdef CONFIG_FAIR_GROUP_SCHED
7785 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7788 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
7791 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
7793 struct task_group *tg = cgroup_tg(cgrp);
7795 return (u64) scale_load_down(tg->shares);
7798 #ifdef CONFIG_CFS_BANDWIDTH
7799 static DEFINE_MUTEX(cfs_constraints_mutex);
7801 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7802 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7804 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7806 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7808 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7809 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7811 if (tg == &root_task_group)
7815 * Ensure we have at some amount of bandwidth every period. This is
7816 * to prevent reaching a state of large arrears when throttled via
7817 * entity_tick() resulting in prolonged exit starvation.
7819 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7823 * Likewise, bound things on the otherside by preventing insane quota
7824 * periods. This also allows us to normalize in computing quota
7827 if (period > max_cfs_quota_period)
7830 mutex_lock(&cfs_constraints_mutex);
7831 ret = __cfs_schedulable(tg, period, quota);
7835 runtime_enabled = quota != RUNTIME_INF;
7836 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7837 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled);
7838 raw_spin_lock_irq(&cfs_b->lock);
7839 cfs_b->period = ns_to_ktime(period);
7840 cfs_b->quota = quota;
7842 __refill_cfs_bandwidth_runtime(cfs_b);
7843 /* restart the period timer (if active) to handle new period expiry */
7844 if (runtime_enabled && cfs_b->timer_active) {
7845 /* force a reprogram */
7846 cfs_b->timer_active = 0;
7847 __start_cfs_bandwidth(cfs_b);
7849 raw_spin_unlock_irq(&cfs_b->lock);
7851 for_each_possible_cpu(i) {
7852 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7853 struct rq *rq = cfs_rq->rq;
7855 raw_spin_lock_irq(&rq->lock);
7856 cfs_rq->runtime_enabled = runtime_enabled;
7857 cfs_rq->runtime_remaining = 0;
7859 if (cfs_rq->throttled)
7860 unthrottle_cfs_rq(cfs_rq);
7861 raw_spin_unlock_irq(&rq->lock);
7864 mutex_unlock(&cfs_constraints_mutex);
7869 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7873 period = ktime_to_ns(tg->cfs_bandwidth.period);
7874 if (cfs_quota_us < 0)
7875 quota = RUNTIME_INF;
7877 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7879 return tg_set_cfs_bandwidth(tg, period, quota);
7882 long tg_get_cfs_quota(struct task_group *tg)
7886 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7889 quota_us = tg->cfs_bandwidth.quota;
7890 do_div(quota_us, NSEC_PER_USEC);
7895 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7899 period = (u64)cfs_period_us * NSEC_PER_USEC;
7900 quota = tg->cfs_bandwidth.quota;
7902 return tg_set_cfs_bandwidth(tg, period, quota);
7905 long tg_get_cfs_period(struct task_group *tg)
7909 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7910 do_div(cfs_period_us, NSEC_PER_USEC);
7912 return cfs_period_us;
7915 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
7917 return tg_get_cfs_quota(cgroup_tg(cgrp));
7920 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
7923 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
7926 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
7928 return tg_get_cfs_period(cgroup_tg(cgrp));
7931 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
7934 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
7937 struct cfs_schedulable_data {
7938 struct task_group *tg;
7943 * normalize group quota/period to be quota/max_period
7944 * note: units are usecs
7946 static u64 normalize_cfs_quota(struct task_group *tg,
7947 struct cfs_schedulable_data *d)
7955 period = tg_get_cfs_period(tg);
7956 quota = tg_get_cfs_quota(tg);
7959 /* note: these should typically be equivalent */
7960 if (quota == RUNTIME_INF || quota == -1)
7963 return to_ratio(period, quota);
7966 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7968 struct cfs_schedulable_data *d = data;
7969 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7970 s64 quota = 0, parent_quota = -1;
7973 quota = RUNTIME_INF;
7975 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7977 quota = normalize_cfs_quota(tg, d);
7978 parent_quota = parent_b->hierarchal_quota;
7981 * ensure max(child_quota) <= parent_quota, inherit when no
7984 if (quota == RUNTIME_INF)
7985 quota = parent_quota;
7986 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7989 cfs_b->hierarchal_quota = quota;
7994 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7997 struct cfs_schedulable_data data = {
8003 if (quota != RUNTIME_INF) {
8004 do_div(data.period, NSEC_PER_USEC);
8005 do_div(data.quota, NSEC_PER_USEC);
8009 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8015 static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft,
8016 struct cgroup_map_cb *cb)
8018 struct task_group *tg = cgroup_tg(cgrp);
8019 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8021 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
8022 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
8023 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8027 #endif /* CONFIG_CFS_BANDWIDTH */
8028 #endif /* CONFIG_FAIR_GROUP_SCHED */
8030 #ifdef CONFIG_RT_GROUP_SCHED
8031 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8034 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8037 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8039 return sched_group_rt_runtime(cgroup_tg(cgrp));
8042 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8045 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8048 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8050 return sched_group_rt_period(cgroup_tg(cgrp));
8052 #endif /* CONFIG_RT_GROUP_SCHED */
8054 static struct cftype cpu_files[] = {
8055 #ifdef CONFIG_FAIR_GROUP_SCHED
8058 .read_u64 = cpu_shares_read_u64,
8059 .write_u64 = cpu_shares_write_u64,
8062 #ifdef CONFIG_CFS_BANDWIDTH
8064 .name = "cfs_quota_us",
8065 .read_s64 = cpu_cfs_quota_read_s64,
8066 .write_s64 = cpu_cfs_quota_write_s64,
8069 .name = "cfs_period_us",
8070 .read_u64 = cpu_cfs_period_read_u64,
8071 .write_u64 = cpu_cfs_period_write_u64,
8075 .read_map = cpu_stats_show,
8078 #ifdef CONFIG_RT_GROUP_SCHED
8080 .name = "rt_runtime_us",
8081 .read_s64 = cpu_rt_runtime_read,
8082 .write_s64 = cpu_rt_runtime_write,
8085 .name = "rt_period_us",
8086 .read_u64 = cpu_rt_period_read_uint,
8087 .write_u64 = cpu_rt_period_write_uint,
8093 struct cgroup_subsys cpu_cgroup_subsys = {
8095 .css_alloc = cpu_cgroup_css_alloc,
8096 .css_free = cpu_cgroup_css_free,
8097 .css_online = cpu_cgroup_css_online,
8098 .css_offline = cpu_cgroup_css_offline,
8099 .can_attach = cpu_cgroup_can_attach,
8100 .attach = cpu_cgroup_attach,
8101 .exit = cpu_cgroup_exit,
8102 .subsys_id = cpu_cgroup_subsys_id,
8103 .base_cftypes = cpu_files,
8107 #endif /* CONFIG_CGROUP_SCHED */
8109 void dump_cpu_task(int cpu)
8111 pr_info("Task dump for CPU %d:\n", cpu);
8112 sched_show_task(cpu_curr(cpu));