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;
300 * Maximum bandwidth available for all -deadline tasks and groups
301 * (if group scheduling is configured) on each CPU.
305 unsigned int sysctl_sched_dl_period = 1000000;
306 int sysctl_sched_dl_runtime = 50000;
311 * __task_rq_lock - lock the rq @p resides on.
313 static inline struct rq *__task_rq_lock(struct task_struct *p)
318 lockdep_assert_held(&p->pi_lock);
322 raw_spin_lock(&rq->lock);
323 if (likely(rq == task_rq(p)))
325 raw_spin_unlock(&rq->lock);
330 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
332 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
333 __acquires(p->pi_lock)
339 raw_spin_lock_irqsave(&p->pi_lock, *flags);
341 raw_spin_lock(&rq->lock);
342 if (likely(rq == task_rq(p)))
344 raw_spin_unlock(&rq->lock);
345 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
349 static void __task_rq_unlock(struct rq *rq)
352 raw_spin_unlock(&rq->lock);
356 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
358 __releases(p->pi_lock)
360 raw_spin_unlock(&rq->lock);
361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
365 * this_rq_lock - lock this runqueue and disable interrupts.
367 static struct rq *this_rq_lock(void)
374 raw_spin_lock(&rq->lock);
379 #ifdef CONFIG_SCHED_HRTICK
381 * Use HR-timers to deliver accurate preemption points.
384 static void hrtick_clear(struct rq *rq)
386 if (hrtimer_active(&rq->hrtick_timer))
387 hrtimer_cancel(&rq->hrtick_timer);
391 * High-resolution timer tick.
392 * Runs from hardirq context with interrupts disabled.
394 static enum hrtimer_restart hrtick(struct hrtimer *timer)
396 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
398 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
400 raw_spin_lock(&rq->lock);
402 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
403 raw_spin_unlock(&rq->lock);
405 return HRTIMER_NORESTART;
410 static int __hrtick_restart(struct rq *rq)
412 struct hrtimer *timer = &rq->hrtick_timer;
413 ktime_t time = hrtimer_get_softexpires(timer);
415 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
419 * called from hardirq (IPI) context
421 static void __hrtick_start(void *arg)
425 raw_spin_lock(&rq->lock);
426 __hrtick_restart(rq);
427 rq->hrtick_csd_pending = 0;
428 raw_spin_unlock(&rq->lock);
432 * Called to set the hrtick timer state.
434 * called with rq->lock held and irqs disabled
436 void hrtick_start(struct rq *rq, u64 delay)
438 struct hrtimer *timer = &rq->hrtick_timer;
439 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
441 hrtimer_set_expires(timer, time);
443 if (rq == this_rq()) {
444 __hrtick_restart(rq);
445 } else if (!rq->hrtick_csd_pending) {
446 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
447 rq->hrtick_csd_pending = 1;
452 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
454 int cpu = (int)(long)hcpu;
457 case CPU_UP_CANCELED:
458 case CPU_UP_CANCELED_FROZEN:
459 case CPU_DOWN_PREPARE:
460 case CPU_DOWN_PREPARE_FROZEN:
462 case CPU_DEAD_FROZEN:
463 hrtick_clear(cpu_rq(cpu));
470 static __init void init_hrtick(void)
472 hotcpu_notifier(hotplug_hrtick, 0);
476 * Called to set the hrtick timer state.
478 * called with rq->lock held and irqs disabled
480 void hrtick_start(struct rq *rq, u64 delay)
482 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
483 HRTIMER_MODE_REL_PINNED, 0);
486 static inline void init_hrtick(void)
489 #endif /* CONFIG_SMP */
491 static void init_rq_hrtick(struct rq *rq)
494 rq->hrtick_csd_pending = 0;
496 rq->hrtick_csd.flags = 0;
497 rq->hrtick_csd.func = __hrtick_start;
498 rq->hrtick_csd.info = rq;
501 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
502 rq->hrtick_timer.function = hrtick;
504 #else /* CONFIG_SCHED_HRTICK */
505 static inline void hrtick_clear(struct rq *rq)
509 static inline void init_rq_hrtick(struct rq *rq)
513 static inline void init_hrtick(void)
516 #endif /* CONFIG_SCHED_HRTICK */
519 * resched_task - mark a task 'to be rescheduled now'.
521 * On UP this means the setting of the need_resched flag, on SMP it
522 * might also involve a cross-CPU call to trigger the scheduler on
525 void resched_task(struct task_struct *p)
529 lockdep_assert_held(&task_rq(p)->lock);
531 if (test_tsk_need_resched(p))
534 set_tsk_need_resched(p);
537 if (cpu == smp_processor_id()) {
538 set_preempt_need_resched();
542 /* NEED_RESCHED must be visible before we test polling */
544 if (!tsk_is_polling(p))
545 smp_send_reschedule(cpu);
548 void resched_cpu(int cpu)
550 struct rq *rq = cpu_rq(cpu);
553 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
555 resched_task(cpu_curr(cpu));
556 raw_spin_unlock_irqrestore(&rq->lock, flags);
560 #ifdef CONFIG_NO_HZ_COMMON
562 * In the semi idle case, use the nearest busy cpu for migrating timers
563 * from an idle cpu. This is good for power-savings.
565 * We don't do similar optimization for completely idle system, as
566 * selecting an idle cpu will add more delays to the timers than intended
567 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
569 int get_nohz_timer_target(void)
571 int cpu = smp_processor_id();
573 struct sched_domain *sd;
576 for_each_domain(cpu, sd) {
577 for_each_cpu(i, sched_domain_span(sd)) {
589 * When add_timer_on() enqueues a timer into the timer wheel of an
590 * idle CPU then this timer might expire before the next timer event
591 * which is scheduled to wake up that CPU. In case of a completely
592 * idle system the next event might even be infinite time into the
593 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
594 * leaves the inner idle loop so the newly added timer is taken into
595 * account when the CPU goes back to idle and evaluates the timer
596 * wheel for the next timer event.
598 static void wake_up_idle_cpu(int cpu)
600 struct rq *rq = cpu_rq(cpu);
602 if (cpu == smp_processor_id())
606 * This is safe, as this function is called with the timer
607 * wheel base lock of (cpu) held. When the CPU is on the way
608 * to idle and has not yet set rq->curr to idle then it will
609 * be serialized on the timer wheel base lock and take the new
610 * timer into account automatically.
612 if (rq->curr != rq->idle)
616 * We can set TIF_RESCHED on the idle task of the other CPU
617 * lockless. The worst case is that the other CPU runs the
618 * idle task through an additional NOOP schedule()
620 set_tsk_need_resched(rq->idle);
622 /* NEED_RESCHED must be visible before we test polling */
624 if (!tsk_is_polling(rq->idle))
625 smp_send_reschedule(cpu);
628 static bool wake_up_full_nohz_cpu(int cpu)
630 if (tick_nohz_full_cpu(cpu)) {
631 if (cpu != smp_processor_id() ||
632 tick_nohz_tick_stopped())
633 smp_send_reschedule(cpu);
640 void wake_up_nohz_cpu(int cpu)
642 if (!wake_up_full_nohz_cpu(cpu))
643 wake_up_idle_cpu(cpu);
646 static inline bool got_nohz_idle_kick(void)
648 int cpu = smp_processor_id();
650 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
653 if (idle_cpu(cpu) && !need_resched())
657 * We can't run Idle Load Balance on this CPU for this time so we
658 * cancel it and clear NOHZ_BALANCE_KICK
660 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
664 #else /* CONFIG_NO_HZ_COMMON */
666 static inline bool got_nohz_idle_kick(void)
671 #endif /* CONFIG_NO_HZ_COMMON */
673 #ifdef CONFIG_NO_HZ_FULL
674 bool sched_can_stop_tick(void)
680 /* Make sure rq->nr_running update is visible after the IPI */
683 /* More than one running task need preemption */
684 if (rq->nr_running > 1)
689 #endif /* CONFIG_NO_HZ_FULL */
691 void sched_avg_update(struct rq *rq)
693 s64 period = sched_avg_period();
695 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
697 * Inline assembly required to prevent the compiler
698 * optimising this loop into a divmod call.
699 * See __iter_div_u64_rem() for another example of this.
701 asm("" : "+rm" (rq->age_stamp));
702 rq->age_stamp += period;
707 #endif /* CONFIG_SMP */
709 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
710 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
712 * Iterate task_group tree rooted at *from, calling @down when first entering a
713 * node and @up when leaving it for the final time.
715 * Caller must hold rcu_lock or sufficient equivalent.
717 int walk_tg_tree_from(struct task_group *from,
718 tg_visitor down, tg_visitor up, void *data)
720 struct task_group *parent, *child;
726 ret = (*down)(parent, data);
729 list_for_each_entry_rcu(child, &parent->children, siblings) {
736 ret = (*up)(parent, data);
737 if (ret || parent == from)
741 parent = parent->parent;
748 int tg_nop(struct task_group *tg, void *data)
754 static void set_load_weight(struct task_struct *p)
756 int prio = p->static_prio - MAX_RT_PRIO;
757 struct load_weight *load = &p->se.load;
760 * SCHED_IDLE tasks get minimal weight:
762 if (p->policy == SCHED_IDLE) {
763 load->weight = scale_load(WEIGHT_IDLEPRIO);
764 load->inv_weight = WMULT_IDLEPRIO;
768 load->weight = scale_load(prio_to_weight[prio]);
769 load->inv_weight = prio_to_wmult[prio];
772 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
775 sched_info_queued(rq, p);
776 p->sched_class->enqueue_task(rq, p, flags);
779 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
782 sched_info_dequeued(rq, p);
783 p->sched_class->dequeue_task(rq, p, flags);
786 void activate_task(struct rq *rq, struct task_struct *p, int flags)
788 if (task_contributes_to_load(p))
789 rq->nr_uninterruptible--;
791 enqueue_task(rq, p, flags);
794 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
796 if (task_contributes_to_load(p))
797 rq->nr_uninterruptible++;
799 dequeue_task(rq, p, flags);
802 static void update_rq_clock_task(struct rq *rq, s64 delta)
805 * In theory, the compile should just see 0 here, and optimize out the call
806 * to sched_rt_avg_update. But I don't trust it...
808 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
809 s64 steal = 0, irq_delta = 0;
811 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
812 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
815 * Since irq_time is only updated on {soft,}irq_exit, we might run into
816 * this case when a previous update_rq_clock() happened inside a
819 * When this happens, we stop ->clock_task and only update the
820 * prev_irq_time stamp to account for the part that fit, so that a next
821 * update will consume the rest. This ensures ->clock_task is
824 * It does however cause some slight miss-attribution of {soft,}irq
825 * time, a more accurate solution would be to update the irq_time using
826 * the current rq->clock timestamp, except that would require using
829 if (irq_delta > delta)
832 rq->prev_irq_time += irq_delta;
835 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
836 if (static_key_false((¶virt_steal_rq_enabled))) {
839 steal = paravirt_steal_clock(cpu_of(rq));
840 steal -= rq->prev_steal_time_rq;
842 if (unlikely(steal > delta))
845 st = steal_ticks(steal);
846 steal = st * TICK_NSEC;
848 rq->prev_steal_time_rq += steal;
854 rq->clock_task += delta;
856 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
857 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
858 sched_rt_avg_update(rq, irq_delta + steal);
862 void sched_set_stop_task(int cpu, struct task_struct *stop)
864 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
865 struct task_struct *old_stop = cpu_rq(cpu)->stop;
869 * Make it appear like a SCHED_FIFO task, its something
870 * userspace knows about and won't get confused about.
872 * Also, it will make PI more or less work without too
873 * much confusion -- but then, stop work should not
874 * rely on PI working anyway.
876 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
878 stop->sched_class = &stop_sched_class;
881 cpu_rq(cpu)->stop = stop;
885 * Reset it back to a normal scheduling class so that
886 * it can die in pieces.
888 old_stop->sched_class = &rt_sched_class;
893 * __normal_prio - return the priority that is based on the static prio
895 static inline int __normal_prio(struct task_struct *p)
897 return p->static_prio;
901 * Calculate the expected normal priority: i.e. priority
902 * without taking RT-inheritance into account. Might be
903 * boosted by interactivity modifiers. Changes upon fork,
904 * setprio syscalls, and whenever the interactivity
905 * estimator recalculates.
907 static inline int normal_prio(struct task_struct *p)
911 if (task_has_dl_policy(p))
912 prio = MAX_DL_PRIO-1;
913 else if (task_has_rt_policy(p))
914 prio = MAX_RT_PRIO-1 - p->rt_priority;
916 prio = __normal_prio(p);
921 * Calculate the current priority, i.e. the priority
922 * taken into account by the scheduler. This value might
923 * be boosted by RT tasks, or might be boosted by
924 * interactivity modifiers. Will be RT if the task got
925 * RT-boosted. If not then it returns p->normal_prio.
927 static int effective_prio(struct task_struct *p)
929 p->normal_prio = normal_prio(p);
931 * If we are RT tasks or we were boosted to RT priority,
932 * keep the priority unchanged. Otherwise, update priority
933 * to the normal priority:
935 if (!rt_prio(p->prio))
936 return p->normal_prio;
941 * task_curr - is this task currently executing on a CPU?
942 * @p: the task in question.
944 * Return: 1 if the task is currently executing. 0 otherwise.
946 inline int task_curr(const struct task_struct *p)
948 return cpu_curr(task_cpu(p)) == p;
951 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
952 const struct sched_class *prev_class,
955 if (prev_class != p->sched_class) {
956 if (prev_class->switched_from)
957 prev_class->switched_from(rq, p);
958 p->sched_class->switched_to(rq, p);
959 } else if (oldprio != p->prio || dl_task(p))
960 p->sched_class->prio_changed(rq, p, oldprio);
963 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
965 const struct sched_class *class;
967 if (p->sched_class == rq->curr->sched_class) {
968 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
970 for_each_class(class) {
971 if (class == rq->curr->sched_class)
973 if (class == p->sched_class) {
974 resched_task(rq->curr);
981 * A queue event has occurred, and we're going to schedule. In
982 * this case, we can save a useless back to back clock update.
984 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
985 rq->skip_clock_update = 1;
989 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
991 #ifdef CONFIG_SCHED_DEBUG
993 * We should never call set_task_cpu() on a blocked task,
994 * ttwu() will sort out the placement.
996 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
997 !(task_preempt_count(p) & PREEMPT_ACTIVE));
999 #ifdef CONFIG_LOCKDEP
1001 * The caller should hold either p->pi_lock or rq->lock, when changing
1002 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1004 * sched_move_task() holds both and thus holding either pins the cgroup,
1007 * Furthermore, all task_rq users should acquire both locks, see
1010 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1011 lockdep_is_held(&task_rq(p)->lock)));
1015 trace_sched_migrate_task(p, new_cpu);
1017 if (task_cpu(p) != new_cpu) {
1018 if (p->sched_class->migrate_task_rq)
1019 p->sched_class->migrate_task_rq(p, new_cpu);
1020 p->se.nr_migrations++;
1021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1024 __set_task_cpu(p, new_cpu);
1027 static void __migrate_swap_task(struct task_struct *p, int cpu)
1030 struct rq *src_rq, *dst_rq;
1032 src_rq = task_rq(p);
1033 dst_rq = cpu_rq(cpu);
1035 deactivate_task(src_rq, p, 0);
1036 set_task_cpu(p, cpu);
1037 activate_task(dst_rq, p, 0);
1038 check_preempt_curr(dst_rq, p, 0);
1041 * Task isn't running anymore; make it appear like we migrated
1042 * it before it went to sleep. This means on wakeup we make the
1043 * previous cpu our targer instead of where it really is.
1049 struct migration_swap_arg {
1050 struct task_struct *src_task, *dst_task;
1051 int src_cpu, dst_cpu;
1054 static int migrate_swap_stop(void *data)
1056 struct migration_swap_arg *arg = data;
1057 struct rq *src_rq, *dst_rq;
1060 src_rq = cpu_rq(arg->src_cpu);
1061 dst_rq = cpu_rq(arg->dst_cpu);
1063 double_raw_lock(&arg->src_task->pi_lock,
1064 &arg->dst_task->pi_lock);
1065 double_rq_lock(src_rq, dst_rq);
1066 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1069 if (task_cpu(arg->src_task) != arg->src_cpu)
1072 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1075 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1078 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1079 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1084 double_rq_unlock(src_rq, dst_rq);
1085 raw_spin_unlock(&arg->dst_task->pi_lock);
1086 raw_spin_unlock(&arg->src_task->pi_lock);
1092 * Cross migrate two tasks
1094 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1096 struct migration_swap_arg arg;
1099 arg = (struct migration_swap_arg){
1101 .src_cpu = task_cpu(cur),
1103 .dst_cpu = task_cpu(p),
1106 if (arg.src_cpu == arg.dst_cpu)
1110 * These three tests are all lockless; this is OK since all of them
1111 * will be re-checked with proper locks held further down the line.
1113 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1116 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1119 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1122 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1128 struct migration_arg {
1129 struct task_struct *task;
1133 static int migration_cpu_stop(void *data);
1136 * wait_task_inactive - wait for a thread to unschedule.
1138 * If @match_state is nonzero, it's the @p->state value just checked and
1139 * not expected to change. If it changes, i.e. @p might have woken up,
1140 * then return zero. When we succeed in waiting for @p to be off its CPU,
1141 * we return a positive number (its total switch count). If a second call
1142 * a short while later returns the same number, the caller can be sure that
1143 * @p has remained unscheduled the whole time.
1145 * The caller must ensure that the task *will* unschedule sometime soon,
1146 * else this function might spin for a *long* time. This function can't
1147 * be called with interrupts off, or it may introduce deadlock with
1148 * smp_call_function() if an IPI is sent by the same process we are
1149 * waiting to become inactive.
1151 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1153 unsigned long flags;
1160 * We do the initial early heuristics without holding
1161 * any task-queue locks at all. We'll only try to get
1162 * the runqueue lock when things look like they will
1168 * If the task is actively running on another CPU
1169 * still, just relax and busy-wait without holding
1172 * NOTE! Since we don't hold any locks, it's not
1173 * even sure that "rq" stays as the right runqueue!
1174 * But we don't care, since "task_running()" will
1175 * return false if the runqueue has changed and p
1176 * is actually now running somewhere else!
1178 while (task_running(rq, p)) {
1179 if (match_state && unlikely(p->state != match_state))
1185 * Ok, time to look more closely! We need the rq
1186 * lock now, to be *sure*. If we're wrong, we'll
1187 * just go back and repeat.
1189 rq = task_rq_lock(p, &flags);
1190 trace_sched_wait_task(p);
1191 running = task_running(rq, p);
1194 if (!match_state || p->state == match_state)
1195 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1196 task_rq_unlock(rq, p, &flags);
1199 * If it changed from the expected state, bail out now.
1201 if (unlikely(!ncsw))
1205 * Was it really running after all now that we
1206 * checked with the proper locks actually held?
1208 * Oops. Go back and try again..
1210 if (unlikely(running)) {
1216 * It's not enough that it's not actively running,
1217 * it must be off the runqueue _entirely_, and not
1220 * So if it was still runnable (but just not actively
1221 * running right now), it's preempted, and we should
1222 * yield - it could be a while.
1224 if (unlikely(on_rq)) {
1225 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1227 set_current_state(TASK_UNINTERRUPTIBLE);
1228 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1233 * Ahh, all good. It wasn't running, and it wasn't
1234 * runnable, which means that it will never become
1235 * running in the future either. We're all done!
1244 * kick_process - kick a running thread to enter/exit the kernel
1245 * @p: the to-be-kicked thread
1247 * Cause a process which is running on another CPU to enter
1248 * kernel-mode, without any delay. (to get signals handled.)
1250 * NOTE: this function doesn't have to take the runqueue lock,
1251 * because all it wants to ensure is that the remote task enters
1252 * the kernel. If the IPI races and the task has been migrated
1253 * to another CPU then no harm is done and the purpose has been
1256 void kick_process(struct task_struct *p)
1262 if ((cpu != smp_processor_id()) && task_curr(p))
1263 smp_send_reschedule(cpu);
1266 EXPORT_SYMBOL_GPL(kick_process);
1267 #endif /* CONFIG_SMP */
1271 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1273 static int select_fallback_rq(int cpu, struct task_struct *p)
1275 int nid = cpu_to_node(cpu);
1276 const struct cpumask *nodemask = NULL;
1277 enum { cpuset, possible, fail } state = cpuset;
1281 * If the node that the cpu is on has been offlined, cpu_to_node()
1282 * will return -1. There is no cpu on the node, and we should
1283 * select the cpu on the other node.
1286 nodemask = cpumask_of_node(nid);
1288 /* Look for allowed, online CPU in same node. */
1289 for_each_cpu(dest_cpu, nodemask) {
1290 if (!cpu_online(dest_cpu))
1292 if (!cpu_active(dest_cpu))
1294 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1300 /* Any allowed, online CPU? */
1301 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1302 if (!cpu_online(dest_cpu))
1304 if (!cpu_active(dest_cpu))
1311 /* No more Mr. Nice Guy. */
1312 cpuset_cpus_allowed_fallback(p);
1317 do_set_cpus_allowed(p, cpu_possible_mask);
1328 if (state != cpuset) {
1330 * Don't tell them about moving exiting tasks or
1331 * kernel threads (both mm NULL), since they never
1334 if (p->mm && printk_ratelimit()) {
1335 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1336 task_pid_nr(p), p->comm, cpu);
1344 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1347 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1349 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1352 * In order not to call set_task_cpu() on a blocking task we need
1353 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1356 * Since this is common to all placement strategies, this lives here.
1358 * [ this allows ->select_task() to simply return task_cpu(p) and
1359 * not worry about this generic constraint ]
1361 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1363 cpu = select_fallback_rq(task_cpu(p), p);
1368 static void update_avg(u64 *avg, u64 sample)
1370 s64 diff = sample - *avg;
1376 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1378 #ifdef CONFIG_SCHEDSTATS
1379 struct rq *rq = this_rq();
1382 int this_cpu = smp_processor_id();
1384 if (cpu == this_cpu) {
1385 schedstat_inc(rq, ttwu_local);
1386 schedstat_inc(p, se.statistics.nr_wakeups_local);
1388 struct sched_domain *sd;
1390 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1392 for_each_domain(this_cpu, sd) {
1393 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1394 schedstat_inc(sd, ttwu_wake_remote);
1401 if (wake_flags & WF_MIGRATED)
1402 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1404 #endif /* CONFIG_SMP */
1406 schedstat_inc(rq, ttwu_count);
1407 schedstat_inc(p, se.statistics.nr_wakeups);
1409 if (wake_flags & WF_SYNC)
1410 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1412 #endif /* CONFIG_SCHEDSTATS */
1415 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1417 activate_task(rq, p, en_flags);
1420 /* if a worker is waking up, notify workqueue */
1421 if (p->flags & PF_WQ_WORKER)
1422 wq_worker_waking_up(p, cpu_of(rq));
1426 * Mark the task runnable and perform wakeup-preemption.
1429 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1431 check_preempt_curr(rq, p, wake_flags);
1432 trace_sched_wakeup(p, true);
1434 p->state = TASK_RUNNING;
1436 if (p->sched_class->task_woken)
1437 p->sched_class->task_woken(rq, p);
1439 if (rq->idle_stamp) {
1440 u64 delta = rq_clock(rq) - rq->idle_stamp;
1441 u64 max = 2*rq->max_idle_balance_cost;
1443 update_avg(&rq->avg_idle, delta);
1445 if (rq->avg_idle > max)
1454 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1457 if (p->sched_contributes_to_load)
1458 rq->nr_uninterruptible--;
1461 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1462 ttwu_do_wakeup(rq, p, wake_flags);
1466 * Called in case the task @p isn't fully descheduled from its runqueue,
1467 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1468 * since all we need to do is flip p->state to TASK_RUNNING, since
1469 * the task is still ->on_rq.
1471 static int ttwu_remote(struct task_struct *p, int wake_flags)
1476 rq = __task_rq_lock(p);
1478 /* check_preempt_curr() may use rq clock */
1479 update_rq_clock(rq);
1480 ttwu_do_wakeup(rq, p, wake_flags);
1483 __task_rq_unlock(rq);
1489 static void sched_ttwu_pending(void)
1491 struct rq *rq = this_rq();
1492 struct llist_node *llist = llist_del_all(&rq->wake_list);
1493 struct task_struct *p;
1495 raw_spin_lock(&rq->lock);
1498 p = llist_entry(llist, struct task_struct, wake_entry);
1499 llist = llist_next(llist);
1500 ttwu_do_activate(rq, p, 0);
1503 raw_spin_unlock(&rq->lock);
1506 void scheduler_ipi(void)
1509 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1510 * TIF_NEED_RESCHED remotely (for the first time) will also send
1513 if (tif_need_resched())
1514 set_preempt_need_resched();
1516 if (llist_empty(&this_rq()->wake_list)
1517 && !tick_nohz_full_cpu(smp_processor_id())
1518 && !got_nohz_idle_kick())
1522 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1523 * traditionally all their work was done from the interrupt return
1524 * path. Now that we actually do some work, we need to make sure
1527 * Some archs already do call them, luckily irq_enter/exit nest
1530 * Arguably we should visit all archs and update all handlers,
1531 * however a fair share of IPIs are still resched only so this would
1532 * somewhat pessimize the simple resched case.
1535 tick_nohz_full_check();
1536 sched_ttwu_pending();
1539 * Check if someone kicked us for doing the nohz idle load balance.
1541 if (unlikely(got_nohz_idle_kick())) {
1542 this_rq()->idle_balance = 1;
1543 raise_softirq_irqoff(SCHED_SOFTIRQ);
1548 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1550 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1551 smp_send_reschedule(cpu);
1554 bool cpus_share_cache(int this_cpu, int that_cpu)
1556 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1558 #endif /* CONFIG_SMP */
1560 static void ttwu_queue(struct task_struct *p, int cpu)
1562 struct rq *rq = cpu_rq(cpu);
1564 #if defined(CONFIG_SMP)
1565 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1566 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1567 ttwu_queue_remote(p, cpu);
1572 raw_spin_lock(&rq->lock);
1573 ttwu_do_activate(rq, p, 0);
1574 raw_spin_unlock(&rq->lock);
1578 * try_to_wake_up - wake up a thread
1579 * @p: the thread to be awakened
1580 * @state: the mask of task states that can be woken
1581 * @wake_flags: wake modifier flags (WF_*)
1583 * Put it on the run-queue if it's not already there. The "current"
1584 * thread is always on the run-queue (except when the actual
1585 * re-schedule is in progress), and as such you're allowed to do
1586 * the simpler "current->state = TASK_RUNNING" to mark yourself
1587 * runnable without the overhead of this.
1589 * Return: %true if @p was woken up, %false if it was already running.
1590 * or @state didn't match @p's state.
1593 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1595 unsigned long flags;
1596 int cpu, success = 0;
1599 * If we are going to wake up a thread waiting for CONDITION we
1600 * need to ensure that CONDITION=1 done by the caller can not be
1601 * reordered with p->state check below. This pairs with mb() in
1602 * set_current_state() the waiting thread does.
1604 smp_mb__before_spinlock();
1605 raw_spin_lock_irqsave(&p->pi_lock, flags);
1606 if (!(p->state & state))
1609 success = 1; /* we're going to change ->state */
1612 if (p->on_rq && ttwu_remote(p, wake_flags))
1617 * If the owning (remote) cpu is still in the middle of schedule() with
1618 * this task as prev, wait until its done referencing the task.
1623 * Pairs with the smp_wmb() in finish_lock_switch().
1627 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1628 p->state = TASK_WAKING;
1630 if (p->sched_class->task_waking)
1631 p->sched_class->task_waking(p);
1633 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1634 if (task_cpu(p) != cpu) {
1635 wake_flags |= WF_MIGRATED;
1636 set_task_cpu(p, cpu);
1638 #endif /* CONFIG_SMP */
1642 ttwu_stat(p, cpu, wake_flags);
1644 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1650 * try_to_wake_up_local - try to wake up a local task with rq lock held
1651 * @p: the thread to be awakened
1653 * Put @p on the run-queue if it's not already there. The caller must
1654 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1657 static void try_to_wake_up_local(struct task_struct *p)
1659 struct rq *rq = task_rq(p);
1661 if (WARN_ON_ONCE(rq != this_rq()) ||
1662 WARN_ON_ONCE(p == current))
1665 lockdep_assert_held(&rq->lock);
1667 if (!raw_spin_trylock(&p->pi_lock)) {
1668 raw_spin_unlock(&rq->lock);
1669 raw_spin_lock(&p->pi_lock);
1670 raw_spin_lock(&rq->lock);
1673 if (!(p->state & TASK_NORMAL))
1677 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1679 ttwu_do_wakeup(rq, p, 0);
1680 ttwu_stat(p, smp_processor_id(), 0);
1682 raw_spin_unlock(&p->pi_lock);
1686 * wake_up_process - Wake up a specific process
1687 * @p: The process to be woken up.
1689 * Attempt to wake up the nominated process and move it to the set of runnable
1692 * Return: 1 if the process was woken up, 0 if it was already running.
1694 * It may be assumed that this function implies a write memory barrier before
1695 * changing the task state if and only if any tasks are woken up.
1697 int wake_up_process(struct task_struct *p)
1699 WARN_ON(task_is_stopped_or_traced(p));
1700 return try_to_wake_up(p, TASK_NORMAL, 0);
1702 EXPORT_SYMBOL(wake_up_process);
1704 int wake_up_state(struct task_struct *p, unsigned int state)
1706 return try_to_wake_up(p, state, 0);
1710 * Perform scheduler related setup for a newly forked process p.
1711 * p is forked by current.
1713 * __sched_fork() is basic setup used by init_idle() too:
1715 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1720 p->se.exec_start = 0;
1721 p->se.sum_exec_runtime = 0;
1722 p->se.prev_sum_exec_runtime = 0;
1723 p->se.nr_migrations = 0;
1725 INIT_LIST_HEAD(&p->se.group_node);
1727 #ifdef CONFIG_SCHEDSTATS
1728 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1731 RB_CLEAR_NODE(&p->dl.rb_node);
1732 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1733 p->dl.dl_runtime = p->dl.runtime = 0;
1734 p->dl.dl_deadline = p->dl.deadline = 0;
1735 p->dl.dl_period = 0;
1738 INIT_LIST_HEAD(&p->rt.run_list);
1740 #ifdef CONFIG_PREEMPT_NOTIFIERS
1741 INIT_HLIST_HEAD(&p->preempt_notifiers);
1744 #ifdef CONFIG_NUMA_BALANCING
1745 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1746 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1747 p->mm->numa_scan_seq = 0;
1750 if (clone_flags & CLONE_VM)
1751 p->numa_preferred_nid = current->numa_preferred_nid;
1753 p->numa_preferred_nid = -1;
1755 p->node_stamp = 0ULL;
1756 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1757 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1758 p->numa_work.next = &p->numa_work;
1759 p->numa_faults = NULL;
1760 p->numa_faults_buffer = NULL;
1762 INIT_LIST_HEAD(&p->numa_entry);
1763 p->numa_group = NULL;
1764 #endif /* CONFIG_NUMA_BALANCING */
1767 #ifdef CONFIG_NUMA_BALANCING
1768 #ifdef CONFIG_SCHED_DEBUG
1769 void set_numabalancing_state(bool enabled)
1772 sched_feat_set("NUMA");
1774 sched_feat_set("NO_NUMA");
1777 __read_mostly bool numabalancing_enabled;
1779 void set_numabalancing_state(bool enabled)
1781 numabalancing_enabled = enabled;
1783 #endif /* CONFIG_SCHED_DEBUG */
1784 #endif /* CONFIG_NUMA_BALANCING */
1787 * fork()/clone()-time setup:
1789 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1791 unsigned long flags;
1792 int cpu = get_cpu();
1794 __sched_fork(clone_flags, p);
1796 * We mark the process as running here. This guarantees that
1797 * nobody will actually run it, and a signal or other external
1798 * event cannot wake it up and insert it on the runqueue either.
1800 p->state = TASK_RUNNING;
1803 * Make sure we do not leak PI boosting priority to the child.
1805 p->prio = current->normal_prio;
1808 * Revert to default priority/policy on fork if requested.
1810 if (unlikely(p->sched_reset_on_fork)) {
1811 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1812 p->policy = SCHED_NORMAL;
1813 p->static_prio = NICE_TO_PRIO(0);
1815 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1816 p->static_prio = NICE_TO_PRIO(0);
1818 p->prio = p->normal_prio = __normal_prio(p);
1822 * We don't need the reset flag anymore after the fork. It has
1823 * fulfilled its duty:
1825 p->sched_reset_on_fork = 0;
1828 if (dl_prio(p->prio)) {
1831 } else if (rt_prio(p->prio)) {
1832 p->sched_class = &rt_sched_class;
1834 p->sched_class = &fair_sched_class;
1837 if (p->sched_class->task_fork)
1838 p->sched_class->task_fork(p);
1841 * The child is not yet in the pid-hash so no cgroup attach races,
1842 * and the cgroup is pinned to this child due to cgroup_fork()
1843 * is ran before sched_fork().
1845 * Silence PROVE_RCU.
1847 raw_spin_lock_irqsave(&p->pi_lock, flags);
1848 set_task_cpu(p, cpu);
1849 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1851 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1852 if (likely(sched_info_on()))
1853 memset(&p->sched_info, 0, sizeof(p->sched_info));
1855 #if defined(CONFIG_SMP)
1858 init_task_preempt_count(p);
1860 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1861 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1868 unsigned long to_ratio(u64 period, u64 runtime)
1870 if (runtime == RUNTIME_INF)
1874 * Doing this here saves a lot of checks in all
1875 * the calling paths, and returning zero seems
1876 * safe for them anyway.
1881 return div64_u64(runtime << 20, period);
1885 inline struct dl_bw *dl_bw_of(int i)
1887 return &cpu_rq(i)->rd->dl_bw;
1890 static inline int __dl_span_weight(struct rq *rq)
1892 return cpumask_weight(rq->rd->span);
1895 inline struct dl_bw *dl_bw_of(int i)
1897 return &cpu_rq(i)->dl.dl_bw;
1900 static inline int __dl_span_weight(struct rq *rq)
1907 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1909 dl_b->total_bw -= tsk_bw;
1913 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1915 dl_b->total_bw += tsk_bw;
1919 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1921 return dl_b->bw != -1 &&
1922 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1926 * We must be sure that accepting a new task (or allowing changing the
1927 * parameters of an existing one) is consistent with the bandwidth
1928 * constraints. If yes, this function also accordingly updates the currently
1929 * allocated bandwidth to reflect the new situation.
1931 * This function is called while holding p's rq->lock.
1933 static int dl_overflow(struct task_struct *p, int policy,
1934 const struct sched_attr *attr)
1937 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1938 u64 period = attr->sched_period;
1939 u64 runtime = attr->sched_runtime;
1940 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1941 int cpus = __dl_span_weight(task_rq(p));
1944 if (new_bw == p->dl.dl_bw)
1948 * Either if a task, enters, leave, or stays -deadline but changes
1949 * its parameters, we may need to update accordingly the total
1950 * allocated bandwidth of the container.
1952 raw_spin_lock(&dl_b->lock);
1953 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1954 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1955 __dl_add(dl_b, new_bw);
1957 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1958 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1959 __dl_clear(dl_b, p->dl.dl_bw);
1960 __dl_add(dl_b, new_bw);
1962 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1963 __dl_clear(dl_b, p->dl.dl_bw);
1966 raw_spin_unlock(&dl_b->lock);
1971 extern void init_dl_bw(struct dl_bw *dl_b);
1974 * wake_up_new_task - wake up a newly created task for the first time.
1976 * This function will do some initial scheduler statistics housekeeping
1977 * that must be done for every newly created context, then puts the task
1978 * on the runqueue and wakes it.
1980 void wake_up_new_task(struct task_struct *p)
1982 unsigned long flags;
1985 raw_spin_lock_irqsave(&p->pi_lock, flags);
1988 * Fork balancing, do it here and not earlier because:
1989 * - cpus_allowed can change in the fork path
1990 * - any previously selected cpu might disappear through hotplug
1992 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1995 /* Initialize new task's runnable average */
1996 init_task_runnable_average(p);
1997 rq = __task_rq_lock(p);
1998 activate_task(rq, p, 0);
2000 trace_sched_wakeup_new(p, true);
2001 check_preempt_curr(rq, p, WF_FORK);
2003 if (p->sched_class->task_woken)
2004 p->sched_class->task_woken(rq, p);
2006 task_rq_unlock(rq, p, &flags);
2009 #ifdef CONFIG_PREEMPT_NOTIFIERS
2012 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2013 * @notifier: notifier struct to register
2015 void preempt_notifier_register(struct preempt_notifier *notifier)
2017 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2019 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2022 * preempt_notifier_unregister - no longer interested in preemption notifications
2023 * @notifier: notifier struct to unregister
2025 * This is safe to call from within a preemption notifier.
2027 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2029 hlist_del(¬ifier->link);
2031 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2033 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2035 struct preempt_notifier *notifier;
2037 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2038 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2042 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2043 struct task_struct *next)
2045 struct preempt_notifier *notifier;
2047 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2048 notifier->ops->sched_out(notifier, next);
2051 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2053 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2058 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2059 struct task_struct *next)
2063 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2066 * prepare_task_switch - prepare to switch tasks
2067 * @rq: the runqueue preparing to switch
2068 * @prev: the current task that is being switched out
2069 * @next: the task we are going to switch to.
2071 * This is called with the rq lock held and interrupts off. It must
2072 * be paired with a subsequent finish_task_switch after the context
2075 * prepare_task_switch sets up locking and calls architecture specific
2079 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2080 struct task_struct *next)
2082 trace_sched_switch(prev, next);
2083 sched_info_switch(rq, prev, next);
2084 perf_event_task_sched_out(prev, next);
2085 fire_sched_out_preempt_notifiers(prev, next);
2086 prepare_lock_switch(rq, next);
2087 prepare_arch_switch(next);
2091 * finish_task_switch - clean up after a task-switch
2092 * @rq: runqueue associated with task-switch
2093 * @prev: the thread we just switched away from.
2095 * finish_task_switch must be called after the context switch, paired
2096 * with a prepare_task_switch call before the context switch.
2097 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2098 * and do any other architecture-specific cleanup actions.
2100 * Note that we may have delayed dropping an mm in context_switch(). If
2101 * so, we finish that here outside of the runqueue lock. (Doing it
2102 * with the lock held can cause deadlocks; see schedule() for
2105 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2106 __releases(rq->lock)
2108 struct mm_struct *mm = rq->prev_mm;
2114 * A task struct has one reference for the use as "current".
2115 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2116 * schedule one last time. The schedule call will never return, and
2117 * the scheduled task must drop that reference.
2118 * The test for TASK_DEAD must occur while the runqueue locks are
2119 * still held, otherwise prev could be scheduled on another cpu, die
2120 * there before we look at prev->state, and then the reference would
2122 * Manfred Spraul <manfred@colorfullife.com>
2124 prev_state = prev->state;
2125 vtime_task_switch(prev);
2126 finish_arch_switch(prev);
2127 perf_event_task_sched_in(prev, current);
2128 finish_lock_switch(rq, prev);
2129 finish_arch_post_lock_switch();
2131 fire_sched_in_preempt_notifiers(current);
2134 if (unlikely(prev_state == TASK_DEAD)) {
2135 task_numa_free(prev);
2137 if (prev->sched_class->task_dead)
2138 prev->sched_class->task_dead(prev);
2141 * Remove function-return probe instances associated with this
2142 * task and put them back on the free list.
2144 kprobe_flush_task(prev);
2145 put_task_struct(prev);
2148 tick_nohz_task_switch(current);
2153 /* assumes rq->lock is held */
2154 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2156 if (prev->sched_class->pre_schedule)
2157 prev->sched_class->pre_schedule(rq, prev);
2160 /* rq->lock is NOT held, but preemption is disabled */
2161 static inline void post_schedule(struct rq *rq)
2163 if (rq->post_schedule) {
2164 unsigned long flags;
2166 raw_spin_lock_irqsave(&rq->lock, flags);
2167 if (rq->curr->sched_class->post_schedule)
2168 rq->curr->sched_class->post_schedule(rq);
2169 raw_spin_unlock_irqrestore(&rq->lock, flags);
2171 rq->post_schedule = 0;
2177 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2181 static inline void post_schedule(struct rq *rq)
2188 * schedule_tail - first thing a freshly forked thread must call.
2189 * @prev: the thread we just switched away from.
2191 asmlinkage void schedule_tail(struct task_struct *prev)
2192 __releases(rq->lock)
2194 struct rq *rq = this_rq();
2196 finish_task_switch(rq, prev);
2199 * FIXME: do we need to worry about rq being invalidated by the
2204 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2205 /* In this case, finish_task_switch does not reenable preemption */
2208 if (current->set_child_tid)
2209 put_user(task_pid_vnr(current), current->set_child_tid);
2213 * context_switch - switch to the new MM and the new
2214 * thread's register state.
2217 context_switch(struct rq *rq, struct task_struct *prev,
2218 struct task_struct *next)
2220 struct mm_struct *mm, *oldmm;
2222 prepare_task_switch(rq, prev, next);
2225 oldmm = prev->active_mm;
2227 * For paravirt, this is coupled with an exit in switch_to to
2228 * combine the page table reload and the switch backend into
2231 arch_start_context_switch(prev);
2234 next->active_mm = oldmm;
2235 atomic_inc(&oldmm->mm_count);
2236 enter_lazy_tlb(oldmm, next);
2238 switch_mm(oldmm, mm, next);
2241 prev->active_mm = NULL;
2242 rq->prev_mm = oldmm;
2245 * Since the runqueue lock will be released by the next
2246 * task (which is an invalid locking op but in the case
2247 * of the scheduler it's an obvious special-case), so we
2248 * do an early lockdep release here:
2250 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2251 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2254 context_tracking_task_switch(prev, next);
2255 /* Here we just switch the register state and the stack. */
2256 switch_to(prev, next, prev);
2260 * this_rq must be evaluated again because prev may have moved
2261 * CPUs since it called schedule(), thus the 'rq' on its stack
2262 * frame will be invalid.
2264 finish_task_switch(this_rq(), prev);
2268 * nr_running and nr_context_switches:
2270 * externally visible scheduler statistics: current number of runnable
2271 * threads, total number of context switches performed since bootup.
2273 unsigned long nr_running(void)
2275 unsigned long i, sum = 0;
2277 for_each_online_cpu(i)
2278 sum += cpu_rq(i)->nr_running;
2283 unsigned long long nr_context_switches(void)
2286 unsigned long long sum = 0;
2288 for_each_possible_cpu(i)
2289 sum += cpu_rq(i)->nr_switches;
2294 unsigned long nr_iowait(void)
2296 unsigned long i, sum = 0;
2298 for_each_possible_cpu(i)
2299 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2304 unsigned long nr_iowait_cpu(int cpu)
2306 struct rq *this = cpu_rq(cpu);
2307 return atomic_read(&this->nr_iowait);
2313 * sched_exec - execve() is a valuable balancing opportunity, because at
2314 * this point the task has the smallest effective memory and cache footprint.
2316 void sched_exec(void)
2318 struct task_struct *p = current;
2319 unsigned long flags;
2322 raw_spin_lock_irqsave(&p->pi_lock, flags);
2323 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2324 if (dest_cpu == smp_processor_id())
2327 if (likely(cpu_active(dest_cpu))) {
2328 struct migration_arg arg = { p, dest_cpu };
2330 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2331 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2335 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 DEFINE_PER_CPU(struct kernel_stat, kstat);
2341 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2343 EXPORT_PER_CPU_SYMBOL(kstat);
2344 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2347 * Return any ns on the sched_clock that have not yet been accounted in
2348 * @p in case that task is currently running.
2350 * Called with task_rq_lock() held on @rq.
2352 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2356 if (task_current(rq, p)) {
2357 update_rq_clock(rq);
2358 ns = rq_clock_task(rq) - p->se.exec_start;
2366 unsigned long long task_delta_exec(struct task_struct *p)
2368 unsigned long flags;
2372 rq = task_rq_lock(p, &flags);
2373 ns = do_task_delta_exec(p, rq);
2374 task_rq_unlock(rq, p, &flags);
2380 * Return accounted runtime for the task.
2381 * In case the task is currently running, return the runtime plus current's
2382 * pending runtime that have not been accounted yet.
2384 unsigned long long task_sched_runtime(struct task_struct *p)
2386 unsigned long flags;
2390 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2392 * 64-bit doesn't need locks to atomically read a 64bit value.
2393 * So we have a optimization chance when the task's delta_exec is 0.
2394 * Reading ->on_cpu is racy, but this is ok.
2396 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2397 * If we race with it entering cpu, unaccounted time is 0. This is
2398 * indistinguishable from the read occurring a few cycles earlier.
2401 return p->se.sum_exec_runtime;
2404 rq = task_rq_lock(p, &flags);
2405 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2406 task_rq_unlock(rq, p, &flags);
2412 * This function gets called by the timer code, with HZ frequency.
2413 * We call it with interrupts disabled.
2415 void scheduler_tick(void)
2417 int cpu = smp_processor_id();
2418 struct rq *rq = cpu_rq(cpu);
2419 struct task_struct *curr = rq->curr;
2423 raw_spin_lock(&rq->lock);
2424 update_rq_clock(rq);
2425 curr->sched_class->task_tick(rq, curr, 0);
2426 update_cpu_load_active(rq);
2427 raw_spin_unlock(&rq->lock);
2429 perf_event_task_tick();
2432 rq->idle_balance = idle_cpu(cpu);
2433 trigger_load_balance(rq, cpu);
2435 rq_last_tick_reset(rq);
2438 #ifdef CONFIG_NO_HZ_FULL
2440 * scheduler_tick_max_deferment
2442 * Keep at least one tick per second when a single
2443 * active task is running because the scheduler doesn't
2444 * yet completely support full dynticks environment.
2446 * This makes sure that uptime, CFS vruntime, load
2447 * balancing, etc... continue to move forward, even
2448 * with a very low granularity.
2450 * Return: Maximum deferment in nanoseconds.
2452 u64 scheduler_tick_max_deferment(void)
2454 struct rq *rq = this_rq();
2455 unsigned long next, now = ACCESS_ONCE(jiffies);
2457 next = rq->last_sched_tick + HZ;
2459 if (time_before_eq(next, now))
2462 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2466 notrace unsigned long get_parent_ip(unsigned long addr)
2468 if (in_lock_functions(addr)) {
2469 addr = CALLER_ADDR2;
2470 if (in_lock_functions(addr))
2471 addr = CALLER_ADDR3;
2476 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2477 defined(CONFIG_PREEMPT_TRACER))
2479 void __kprobes preempt_count_add(int val)
2481 #ifdef CONFIG_DEBUG_PREEMPT
2485 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2488 __preempt_count_add(val);
2489 #ifdef CONFIG_DEBUG_PREEMPT
2491 * Spinlock count overflowing soon?
2493 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2496 if (preempt_count() == val)
2497 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2499 EXPORT_SYMBOL(preempt_count_add);
2501 void __kprobes preempt_count_sub(int val)
2503 #ifdef CONFIG_DEBUG_PREEMPT
2507 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2510 * Is the spinlock portion underflowing?
2512 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2513 !(preempt_count() & PREEMPT_MASK)))
2517 if (preempt_count() == val)
2518 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2519 __preempt_count_sub(val);
2521 EXPORT_SYMBOL(preempt_count_sub);
2526 * Print scheduling while atomic bug:
2528 static noinline void __schedule_bug(struct task_struct *prev)
2530 if (oops_in_progress)
2533 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2534 prev->comm, prev->pid, preempt_count());
2536 debug_show_held_locks(prev);
2538 if (irqs_disabled())
2539 print_irqtrace_events(prev);
2541 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2545 * Various schedule()-time debugging checks and statistics:
2547 static inline void schedule_debug(struct task_struct *prev)
2550 * Test if we are atomic. Since do_exit() needs to call into
2551 * schedule() atomically, we ignore that path. Otherwise whine
2552 * if we are scheduling when we should not.
2554 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2555 __schedule_bug(prev);
2558 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2560 schedstat_inc(this_rq(), sched_count);
2563 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2565 if (prev->on_rq || rq->skip_clock_update < 0)
2566 update_rq_clock(rq);
2567 prev->sched_class->put_prev_task(rq, prev);
2571 * Pick up the highest-prio task:
2573 static inline struct task_struct *
2574 pick_next_task(struct rq *rq)
2576 const struct sched_class *class;
2577 struct task_struct *p;
2580 * Optimization: we know that if all tasks are in
2581 * the fair class we can call that function directly:
2583 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2584 p = fair_sched_class.pick_next_task(rq);
2589 for_each_class(class) {
2590 p = class->pick_next_task(rq);
2595 BUG(); /* the idle class will always have a runnable task */
2599 * __schedule() is the main scheduler function.
2601 * The main means of driving the scheduler and thus entering this function are:
2603 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2605 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2606 * paths. For example, see arch/x86/entry_64.S.
2608 * To drive preemption between tasks, the scheduler sets the flag in timer
2609 * interrupt handler scheduler_tick().
2611 * 3. Wakeups don't really cause entry into schedule(). They add a
2612 * task to the run-queue and that's it.
2614 * Now, if the new task added to the run-queue preempts the current
2615 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2616 * called on the nearest possible occasion:
2618 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2620 * - in syscall or exception context, at the next outmost
2621 * preempt_enable(). (this might be as soon as the wake_up()'s
2624 * - in IRQ context, return from interrupt-handler to
2625 * preemptible context
2627 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2630 * - cond_resched() call
2631 * - explicit schedule() call
2632 * - return from syscall or exception to user-space
2633 * - return from interrupt-handler to user-space
2635 static void __sched __schedule(void)
2637 struct task_struct *prev, *next;
2638 unsigned long *switch_count;
2644 cpu = smp_processor_id();
2646 rcu_note_context_switch(cpu);
2649 schedule_debug(prev);
2651 if (sched_feat(HRTICK))
2655 * Make sure that signal_pending_state()->signal_pending() below
2656 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2657 * done by the caller to avoid the race with signal_wake_up().
2659 smp_mb__before_spinlock();
2660 raw_spin_lock_irq(&rq->lock);
2662 switch_count = &prev->nivcsw;
2663 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2664 if (unlikely(signal_pending_state(prev->state, prev))) {
2665 prev->state = TASK_RUNNING;
2667 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2671 * If a worker went to sleep, notify and ask workqueue
2672 * whether it wants to wake up a task to maintain
2675 if (prev->flags & PF_WQ_WORKER) {
2676 struct task_struct *to_wakeup;
2678 to_wakeup = wq_worker_sleeping(prev, cpu);
2680 try_to_wake_up_local(to_wakeup);
2683 switch_count = &prev->nvcsw;
2686 pre_schedule(rq, prev);
2688 if (unlikely(!rq->nr_running))
2689 idle_balance(cpu, rq);
2691 put_prev_task(rq, prev);
2692 next = pick_next_task(rq);
2693 clear_tsk_need_resched(prev);
2694 clear_preempt_need_resched();
2695 rq->skip_clock_update = 0;
2697 if (likely(prev != next)) {
2702 context_switch(rq, prev, next); /* unlocks the rq */
2704 * The context switch have flipped the stack from under us
2705 * and restored the local variables which were saved when
2706 * this task called schedule() in the past. prev == current
2707 * is still correct, but it can be moved to another cpu/rq.
2709 cpu = smp_processor_id();
2712 raw_spin_unlock_irq(&rq->lock);
2716 sched_preempt_enable_no_resched();
2721 static inline void sched_submit_work(struct task_struct *tsk)
2723 if (!tsk->state || tsk_is_pi_blocked(tsk))
2726 * If we are going to sleep and we have plugged IO queued,
2727 * make sure to submit it to avoid deadlocks.
2729 if (blk_needs_flush_plug(tsk))
2730 blk_schedule_flush_plug(tsk);
2733 asmlinkage void __sched schedule(void)
2735 struct task_struct *tsk = current;
2737 sched_submit_work(tsk);
2740 EXPORT_SYMBOL(schedule);
2742 #ifdef CONFIG_CONTEXT_TRACKING
2743 asmlinkage void __sched schedule_user(void)
2746 * If we come here after a random call to set_need_resched(),
2747 * or we have been woken up remotely but the IPI has not yet arrived,
2748 * we haven't yet exited the RCU idle mode. Do it here manually until
2749 * we find a better solution.
2758 * schedule_preempt_disabled - called with preemption disabled
2760 * Returns with preemption disabled. Note: preempt_count must be 1
2762 void __sched schedule_preempt_disabled(void)
2764 sched_preempt_enable_no_resched();
2769 #ifdef CONFIG_PREEMPT
2771 * this is the entry point to schedule() from in-kernel preemption
2772 * off of preempt_enable. Kernel preemptions off return from interrupt
2773 * occur there and call schedule directly.
2775 asmlinkage void __sched notrace preempt_schedule(void)
2778 * If there is a non-zero preempt_count or interrupts are disabled,
2779 * we do not want to preempt the current task. Just return..
2781 if (likely(!preemptible()))
2785 __preempt_count_add(PREEMPT_ACTIVE);
2787 __preempt_count_sub(PREEMPT_ACTIVE);
2790 * Check again in case we missed a preemption opportunity
2791 * between schedule and now.
2794 } while (need_resched());
2796 EXPORT_SYMBOL(preempt_schedule);
2797 #endif /* CONFIG_PREEMPT */
2800 * this is the entry point to schedule() from kernel preemption
2801 * off of irq context.
2802 * Note, that this is called and return with irqs disabled. This will
2803 * protect us against recursive calling from irq.
2805 asmlinkage void __sched preempt_schedule_irq(void)
2807 enum ctx_state prev_state;
2809 /* Catch callers which need to be fixed */
2810 BUG_ON(preempt_count() || !irqs_disabled());
2812 prev_state = exception_enter();
2815 __preempt_count_add(PREEMPT_ACTIVE);
2818 local_irq_disable();
2819 __preempt_count_sub(PREEMPT_ACTIVE);
2822 * Check again in case we missed a preemption opportunity
2823 * between schedule and now.
2826 } while (need_resched());
2828 exception_exit(prev_state);
2831 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2834 return try_to_wake_up(curr->private, mode, wake_flags);
2836 EXPORT_SYMBOL(default_wake_function);
2839 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2841 unsigned long flags;
2844 init_waitqueue_entry(&wait, current);
2846 __set_current_state(state);
2848 spin_lock_irqsave(&q->lock, flags);
2849 __add_wait_queue(q, &wait);
2850 spin_unlock(&q->lock);
2851 timeout = schedule_timeout(timeout);
2852 spin_lock_irq(&q->lock);
2853 __remove_wait_queue(q, &wait);
2854 spin_unlock_irqrestore(&q->lock, flags);
2859 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2861 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2863 EXPORT_SYMBOL(interruptible_sleep_on);
2866 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2868 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2870 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2872 void __sched sleep_on(wait_queue_head_t *q)
2874 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2876 EXPORT_SYMBOL(sleep_on);
2878 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2880 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2882 EXPORT_SYMBOL(sleep_on_timeout);
2884 #ifdef CONFIG_RT_MUTEXES
2887 * rt_mutex_setprio - set the current priority of a task
2889 * @prio: prio value (kernel-internal form)
2891 * This function changes the 'effective' priority of a task. It does
2892 * not touch ->normal_prio like __setscheduler().
2894 * Used by the rt_mutex code to implement priority inheritance logic.
2896 void rt_mutex_setprio(struct task_struct *p, int prio)
2898 int oldprio, on_rq, running, enqueue_flag = 0;
2900 const struct sched_class *prev_class;
2902 BUG_ON(prio > MAX_PRIO);
2904 rq = __task_rq_lock(p);
2907 * Idle task boosting is a nono in general. There is one
2908 * exception, when PREEMPT_RT and NOHZ is active:
2910 * The idle task calls get_next_timer_interrupt() and holds
2911 * the timer wheel base->lock on the CPU and another CPU wants
2912 * to access the timer (probably to cancel it). We can safely
2913 * ignore the boosting request, as the idle CPU runs this code
2914 * with interrupts disabled and will complete the lock
2915 * protected section without being interrupted. So there is no
2916 * real need to boost.
2918 if (unlikely(p == rq->idle)) {
2919 WARN_ON(p != rq->curr);
2920 WARN_ON(p->pi_blocked_on);
2924 trace_sched_pi_setprio(p, prio);
2925 p->pi_top_task = rt_mutex_get_top_task(p);
2927 prev_class = p->sched_class;
2929 running = task_current(rq, p);
2931 dequeue_task(rq, p, 0);
2933 p->sched_class->put_prev_task(rq, p);
2936 * Boosting condition are:
2937 * 1. -rt task is running and holds mutex A
2938 * --> -dl task blocks on mutex A
2940 * 2. -dl task is running and holds mutex A
2941 * --> -dl task blocks on mutex A and could preempt the
2944 if (dl_prio(prio)) {
2945 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2946 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2947 p->dl.dl_boosted = 1;
2948 p->dl.dl_throttled = 0;
2949 enqueue_flag = ENQUEUE_REPLENISH;
2951 p->dl.dl_boosted = 0;
2952 p->sched_class = &dl_sched_class;
2953 } else if (rt_prio(prio)) {
2954 if (dl_prio(oldprio))
2955 p->dl.dl_boosted = 0;
2957 enqueue_flag = ENQUEUE_HEAD;
2958 p->sched_class = &rt_sched_class;
2960 if (dl_prio(oldprio))
2961 p->dl.dl_boosted = 0;
2962 p->sched_class = &fair_sched_class;
2968 p->sched_class->set_curr_task(rq);
2970 enqueue_task(rq, p, enqueue_flag);
2972 check_class_changed(rq, p, prev_class, oldprio);
2974 __task_rq_unlock(rq);
2978 void set_user_nice(struct task_struct *p, long nice)
2980 int old_prio, delta, on_rq;
2981 unsigned long flags;
2984 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2987 * We have to be careful, if called from sys_setpriority(),
2988 * the task might be in the middle of scheduling on another CPU.
2990 rq = task_rq_lock(p, &flags);
2992 * The RT priorities are set via sched_setscheduler(), but we still
2993 * allow the 'normal' nice value to be set - but as expected
2994 * it wont have any effect on scheduling until the task is
2995 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2997 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2998 p->static_prio = NICE_TO_PRIO(nice);
3003 dequeue_task(rq, p, 0);
3005 p->static_prio = NICE_TO_PRIO(nice);
3008 p->prio = effective_prio(p);
3009 delta = p->prio - old_prio;
3012 enqueue_task(rq, p, 0);
3014 * If the task increased its priority or is running and
3015 * lowered its priority, then reschedule its CPU:
3017 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3018 resched_task(rq->curr);
3021 task_rq_unlock(rq, p, &flags);
3023 EXPORT_SYMBOL(set_user_nice);
3026 * can_nice - check if a task can reduce its nice value
3030 int can_nice(const struct task_struct *p, const int nice)
3032 /* convert nice value [19,-20] to rlimit style value [1,40] */
3033 int nice_rlim = 20 - nice;
3035 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3036 capable(CAP_SYS_NICE));
3039 #ifdef __ARCH_WANT_SYS_NICE
3042 * sys_nice - change the priority of the current process.
3043 * @increment: priority increment
3045 * sys_setpriority is a more generic, but much slower function that
3046 * does similar things.
3048 SYSCALL_DEFINE1(nice, int, increment)
3053 * Setpriority might change our priority at the same moment.
3054 * We don't have to worry. Conceptually one call occurs first
3055 * and we have a single winner.
3057 if (increment < -40)
3062 nice = TASK_NICE(current) + increment;
3068 if (increment < 0 && !can_nice(current, nice))
3071 retval = security_task_setnice(current, nice);
3075 set_user_nice(current, nice);
3082 * task_prio - return the priority value of a given task.
3083 * @p: the task in question.
3085 * Return: The priority value as seen by users in /proc.
3086 * RT tasks are offset by -200. Normal tasks are centered
3087 * around 0, value goes from -16 to +15.
3089 int task_prio(const struct task_struct *p)
3091 return p->prio - MAX_RT_PRIO;
3095 * task_nice - return the nice value of a given task.
3096 * @p: the task in question.
3098 * Return: The nice value [ -20 ... 0 ... 19 ].
3100 int task_nice(const struct task_struct *p)
3102 return TASK_NICE(p);
3104 EXPORT_SYMBOL(task_nice);
3107 * idle_cpu - is a given cpu idle currently?
3108 * @cpu: the processor in question.
3110 * Return: 1 if the CPU is currently idle. 0 otherwise.
3112 int idle_cpu(int cpu)
3114 struct rq *rq = cpu_rq(cpu);
3116 if (rq->curr != rq->idle)
3123 if (!llist_empty(&rq->wake_list))
3131 * idle_task - return the idle task for a given cpu.
3132 * @cpu: the processor in question.
3134 * Return: The idle task for the cpu @cpu.
3136 struct task_struct *idle_task(int cpu)
3138 return cpu_rq(cpu)->idle;
3142 * find_process_by_pid - find a process with a matching PID value.
3143 * @pid: the pid in question.
3145 * The task of @pid, if found. %NULL otherwise.
3147 static struct task_struct *find_process_by_pid(pid_t pid)
3149 return pid ? find_task_by_vpid(pid) : current;
3153 * This function initializes the sched_dl_entity of a newly becoming
3154 * SCHED_DEADLINE task.
3156 * Only the static values are considered here, the actual runtime and the
3157 * absolute deadline will be properly calculated when the task is enqueued
3158 * for the first time with its new policy.
3161 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3163 struct sched_dl_entity *dl_se = &p->dl;
3165 init_dl_task_timer(dl_se);
3166 dl_se->dl_runtime = attr->sched_runtime;
3167 dl_se->dl_deadline = attr->sched_deadline;
3168 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3169 dl_se->flags = attr->sched_flags;
3170 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3171 dl_se->dl_throttled = 0;
3175 /* Actually do priority change: must hold pi & rq lock. */
3176 static void __setscheduler(struct rq *rq, struct task_struct *p,
3177 const struct sched_attr *attr)
3179 int policy = attr->sched_policy;
3183 if (dl_policy(policy))
3184 __setparam_dl(p, attr);
3185 else if (rt_policy(policy))
3186 p->rt_priority = attr->sched_priority;
3188 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3190 p->normal_prio = normal_prio(p);
3191 p->prio = rt_mutex_getprio(p);
3193 if (dl_prio(p->prio))
3194 p->sched_class = &dl_sched_class;
3195 else if (rt_prio(p->prio))
3196 p->sched_class = &rt_sched_class;
3198 p->sched_class = &fair_sched_class;
3204 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3206 struct sched_dl_entity *dl_se = &p->dl;
3208 attr->sched_priority = p->rt_priority;
3209 attr->sched_runtime = dl_se->dl_runtime;
3210 attr->sched_deadline = dl_se->dl_deadline;
3211 attr->sched_period = dl_se->dl_period;
3212 attr->sched_flags = dl_se->flags;
3216 * This function validates the new parameters of a -deadline task.
3217 * We ask for the deadline not being zero, and greater or equal
3218 * than the runtime, as well as the period of being zero or
3219 * greater than deadline. Furthermore, we have to be sure that
3220 * user parameters are above the internal resolution (1us); we
3221 * check sched_runtime only since it is always the smaller one.
3224 __checkparam_dl(const struct sched_attr *attr)
3226 return attr && attr->sched_deadline != 0 &&
3227 (attr->sched_period == 0 ||
3228 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3229 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3230 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3234 * check the target process has a UID that matches the current process's
3236 static bool check_same_owner(struct task_struct *p)
3238 const struct cred *cred = current_cred(), *pcred;
3242 pcred = __task_cred(p);
3243 match = (uid_eq(cred->euid, pcred->euid) ||
3244 uid_eq(cred->euid, pcred->uid));
3249 static int __sched_setscheduler(struct task_struct *p,
3250 const struct sched_attr *attr,
3253 int retval, oldprio, oldpolicy = -1, on_rq, running;
3254 int policy = attr->sched_policy;
3255 unsigned long flags;
3256 const struct sched_class *prev_class;
3260 /* may grab non-irq protected spin_locks */
3261 BUG_ON(in_interrupt());
3263 /* double check policy once rq lock held */
3265 reset_on_fork = p->sched_reset_on_fork;
3266 policy = oldpolicy = p->policy;
3268 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3269 policy &= ~SCHED_RESET_ON_FORK;
3271 if (policy != SCHED_DEADLINE &&
3272 policy != SCHED_FIFO && policy != SCHED_RR &&
3273 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3274 policy != SCHED_IDLE)
3279 * Valid priorities for SCHED_FIFO and SCHED_RR are
3280 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3281 * SCHED_BATCH and SCHED_IDLE is 0.
3283 if (attr->sched_priority < 0 ||
3284 (p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3285 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3287 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3288 (rt_policy(policy) != (attr->sched_priority != 0)))
3292 * Allow unprivileged RT tasks to decrease priority:
3294 if (user && !capable(CAP_SYS_NICE)) {
3295 if (fair_policy(policy)) {
3296 if (!can_nice(p, attr->sched_nice))
3300 if (rt_policy(policy)) {
3301 unsigned long rlim_rtprio =
3302 task_rlimit(p, RLIMIT_RTPRIO);
3304 /* can't set/change the rt policy */
3305 if (policy != p->policy && !rlim_rtprio)
3308 /* can't increase priority */
3309 if (attr->sched_priority > p->rt_priority &&
3310 attr->sched_priority > rlim_rtprio)
3315 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3316 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3318 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3319 if (!can_nice(p, TASK_NICE(p)))
3323 /* can't change other user's priorities */
3324 if (!check_same_owner(p))
3327 /* Normal users shall not reset the sched_reset_on_fork flag */
3328 if (p->sched_reset_on_fork && !reset_on_fork)
3333 retval = security_task_setscheduler(p);
3339 * make sure no PI-waiters arrive (or leave) while we are
3340 * changing the priority of the task:
3342 * To be able to change p->policy safely, the appropriate
3343 * runqueue lock must be held.
3345 rq = task_rq_lock(p, &flags);
3348 * Changing the policy of the stop threads its a very bad idea
3350 if (p == rq->stop) {
3351 task_rq_unlock(rq, p, &flags);
3356 * If not changing anything there's no need to proceed further:
3358 if (unlikely(policy == p->policy)) {
3359 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3361 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3363 if (dl_policy(policy))
3366 task_rq_unlock(rq, p, &flags);
3372 #ifdef CONFIG_RT_GROUP_SCHED
3374 * Do not allow realtime tasks into groups that have no runtime
3377 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3378 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3379 !task_group_is_autogroup(task_group(p))) {
3380 task_rq_unlock(rq, p, &flags);
3385 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3386 cpumask_t *span = rq->rd->span;
3387 cpumask_t act_affinity;
3390 * cpus_allowed mask is statically initialized with
3391 * CPU_MASK_ALL, span is instead dynamic. Here we
3392 * compute the "dynamic" affinity of a task.
3394 cpumask_and(&act_affinity, &p->cpus_allowed,
3398 * Don't allow tasks with an affinity mask smaller than
3399 * the entire root_domain to become SCHED_DEADLINE. We
3400 * will also fail if there's no bandwidth available.
3402 if (!cpumask_equal(&act_affinity, span) ||
3403 rq->rd->dl_bw.bw == 0) {
3404 task_rq_unlock(rq, p, &flags);
3411 /* recheck policy now with rq lock held */
3412 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3413 policy = oldpolicy = -1;
3414 task_rq_unlock(rq, p, &flags);
3419 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3420 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3423 if ((dl_policy(policy) || dl_task(p)) &&
3424 dl_overflow(p, policy, attr)) {
3425 task_rq_unlock(rq, p, &flags);
3430 running = task_current(rq, p);
3432 dequeue_task(rq, p, 0);
3434 p->sched_class->put_prev_task(rq, p);
3436 p->sched_reset_on_fork = reset_on_fork;
3439 prev_class = p->sched_class;
3440 __setscheduler(rq, p, attr);
3443 p->sched_class->set_curr_task(rq);
3445 enqueue_task(rq, p, 0);
3447 check_class_changed(rq, p, prev_class, oldprio);
3448 task_rq_unlock(rq, p, &flags);
3450 rt_mutex_adjust_pi(p);
3456 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3457 * @p: the task in question.
3458 * @policy: new policy.
3459 * @param: structure containing the new RT priority.
3461 * Return: 0 on success. An error code otherwise.
3463 * NOTE that the task may be already dead.
3465 int sched_setscheduler(struct task_struct *p, int policy,
3466 const struct sched_param *param)
3468 struct sched_attr attr = {
3469 .sched_policy = policy,
3470 .sched_priority = param->sched_priority
3472 return __sched_setscheduler(p, &attr, true);
3474 EXPORT_SYMBOL_GPL(sched_setscheduler);
3476 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3478 return __sched_setscheduler(p, attr, true);
3480 EXPORT_SYMBOL_GPL(sched_setattr);
3483 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3484 * @p: the task in question.
3485 * @policy: new policy.
3486 * @param: structure containing the new RT priority.
3488 * Just like sched_setscheduler, only don't bother checking if the
3489 * current context has permission. For example, this is needed in
3490 * stop_machine(): we create temporary high priority worker threads,
3491 * but our caller might not have that capability.
3493 * Return: 0 on success. An error code otherwise.
3495 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3496 const struct sched_param *param)
3498 struct sched_attr attr = {
3499 .sched_policy = policy,
3500 .sched_priority = param->sched_priority
3502 return __sched_setscheduler(p, &attr, false);
3506 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3508 struct sched_param lparam;
3509 struct task_struct *p;
3512 if (!param || pid < 0)
3514 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3519 p = find_process_by_pid(pid);
3521 retval = sched_setscheduler(p, policy, &lparam);
3528 * Mimics kernel/events/core.c perf_copy_attr().
3530 static int sched_copy_attr(struct sched_attr __user *uattr,
3531 struct sched_attr *attr)
3536 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3540 * zero the full structure, so that a short copy will be nice.
3542 memset(attr, 0, sizeof(*attr));
3544 ret = get_user(size, &uattr->size);
3548 if (size > PAGE_SIZE) /* silly large */
3551 if (!size) /* abi compat */
3552 size = SCHED_ATTR_SIZE_VER0;
3554 if (size < SCHED_ATTR_SIZE_VER0)
3558 * If we're handed a bigger struct than we know of,
3559 * ensure all the unknown bits are 0 - i.e. new
3560 * user-space does not rely on any kernel feature
3561 * extensions we dont know about yet.
3563 if (size > sizeof(*attr)) {
3564 unsigned char __user *addr;
3565 unsigned char __user *end;
3568 addr = (void __user *)uattr + sizeof(*attr);
3569 end = (void __user *)uattr + size;
3571 for (; addr < end; addr++) {
3572 ret = get_user(val, addr);
3578 size = sizeof(*attr);
3581 ret = copy_from_user(attr, uattr, size);
3586 * XXX: do we want to be lenient like existing syscalls; or do we want
3587 * to be strict and return an error on out-of-bounds values?
3589 attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3595 put_user(sizeof(*attr), &uattr->size);
3601 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3602 * @pid: the pid in question.
3603 * @policy: new policy.
3604 * @param: structure containing the new RT priority.
3606 * Return: 0 on success. An error code otherwise.
3608 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3609 struct sched_param __user *, param)
3611 /* negative values for policy are not valid */
3615 return do_sched_setscheduler(pid, policy, param);
3619 * sys_sched_setparam - set/change the RT priority of a thread
3620 * @pid: the pid in question.
3621 * @param: structure containing the new RT priority.
3623 * Return: 0 on success. An error code otherwise.
3625 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3627 return do_sched_setscheduler(pid, -1, param);
3631 * sys_sched_setattr - same as above, but with extended sched_attr
3632 * @pid: the pid in question.
3633 * @attr: structure containing the extended parameters.
3635 SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr)
3637 struct sched_attr attr;
3638 struct task_struct *p;
3641 if (!uattr || pid < 0)
3644 if (sched_copy_attr(uattr, &attr))
3649 p = find_process_by_pid(pid);
3651 retval = sched_setattr(p, &attr);
3658 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3659 * @pid: the pid in question.
3661 * Return: On success, the policy of the thread. Otherwise, a negative error
3664 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3666 struct task_struct *p;
3674 p = find_process_by_pid(pid);
3676 retval = security_task_getscheduler(p);
3679 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3686 * sys_sched_getparam - get the RT priority of a thread
3687 * @pid: the pid in question.
3688 * @param: structure containing the RT priority.
3690 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3693 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3695 struct sched_param lp;
3696 struct task_struct *p;
3699 if (!param || pid < 0)
3703 p = find_process_by_pid(pid);
3708 retval = security_task_getscheduler(p);
3712 if (task_has_dl_policy(p)) {
3716 lp.sched_priority = p->rt_priority;
3720 * This one might sleep, we cannot do it with a spinlock held ...
3722 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3731 static int sched_read_attr(struct sched_attr __user *uattr,
3732 struct sched_attr *attr,
3737 if (!access_ok(VERIFY_WRITE, uattr, usize))
3741 * If we're handed a smaller struct than we know of,
3742 * ensure all the unknown bits are 0 - i.e. old
3743 * user-space does not get uncomplete information.
3745 if (usize < sizeof(*attr)) {
3746 unsigned char *addr;
3749 addr = (void *)attr + usize;
3750 end = (void *)attr + sizeof(*attr);
3752 for (; addr < end; addr++) {
3760 ret = copy_to_user(uattr, attr, usize);
3773 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3774 * @pid: the pid in question.
3775 * @attr: structure containing the extended parameters.
3776 * @size: sizeof(attr) for fwd/bwd comp.
3778 SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3781 struct sched_attr attr = {
3782 .size = sizeof(struct sched_attr),
3784 struct task_struct *p;
3787 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3788 size < SCHED_ATTR_SIZE_VER0)
3792 p = find_process_by_pid(pid);
3797 retval = security_task_getscheduler(p);
3801 attr.sched_policy = p->policy;
3802 if (task_has_dl_policy(p))
3803 __getparam_dl(p, &attr);
3804 else if (task_has_rt_policy(p))
3805 attr.sched_priority = p->rt_priority;
3807 attr.sched_nice = TASK_NICE(p);
3811 retval = sched_read_attr(uattr, &attr, size);
3819 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3821 cpumask_var_t cpus_allowed, new_mask;
3822 struct task_struct *p;
3827 p = find_process_by_pid(pid);
3833 /* Prevent p going away */
3837 if (p->flags & PF_NO_SETAFFINITY) {
3841 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3845 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3847 goto out_free_cpus_allowed;
3850 if (!check_same_owner(p)) {
3852 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3859 retval = security_task_setscheduler(p);
3864 * Since bandwidth control happens on root_domain basis,
3865 * if admission test is enabled, we only admit -deadline
3866 * tasks allowed to run on all the CPUs in the task's
3870 if (task_has_dl_policy(p)) {
3871 const struct cpumask *span = task_rq(p)->rd->span;
3873 if (dl_bandwidth_enabled() &&
3874 !cpumask_equal(in_mask, span)) {
3881 cpuset_cpus_allowed(p, cpus_allowed);
3882 cpumask_and(new_mask, in_mask, cpus_allowed);
3884 retval = set_cpus_allowed_ptr(p, new_mask);
3887 cpuset_cpus_allowed(p, cpus_allowed);
3888 if (!cpumask_subset(new_mask, cpus_allowed)) {
3890 * We must have raced with a concurrent cpuset
3891 * update. Just reset the cpus_allowed to the
3892 * cpuset's cpus_allowed
3894 cpumask_copy(new_mask, cpus_allowed);
3899 free_cpumask_var(new_mask);
3900 out_free_cpus_allowed:
3901 free_cpumask_var(cpus_allowed);
3907 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3908 struct cpumask *new_mask)
3910 if (len < cpumask_size())
3911 cpumask_clear(new_mask);
3912 else if (len > cpumask_size())
3913 len = cpumask_size();
3915 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3919 * sys_sched_setaffinity - set the cpu affinity of a process
3920 * @pid: pid of the process
3921 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3922 * @user_mask_ptr: user-space pointer to the new cpu mask
3924 * Return: 0 on success. An error code otherwise.
3926 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3927 unsigned long __user *, user_mask_ptr)
3929 cpumask_var_t new_mask;
3932 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3935 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3937 retval = sched_setaffinity(pid, new_mask);
3938 free_cpumask_var(new_mask);
3942 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3944 struct task_struct *p;
3945 unsigned long flags;
3951 p = find_process_by_pid(pid);
3955 retval = security_task_getscheduler(p);
3959 raw_spin_lock_irqsave(&p->pi_lock, flags);
3960 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3961 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3970 * sys_sched_getaffinity - get the cpu affinity of a process
3971 * @pid: pid of the process
3972 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3973 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3975 * Return: 0 on success. An error code otherwise.
3977 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3978 unsigned long __user *, user_mask_ptr)
3983 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3985 if (len & (sizeof(unsigned long)-1))
3988 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3991 ret = sched_getaffinity(pid, mask);
3993 size_t retlen = min_t(size_t, len, cpumask_size());
3995 if (copy_to_user(user_mask_ptr, mask, retlen))
4000 free_cpumask_var(mask);
4006 * sys_sched_yield - yield the current processor to other threads.
4008 * This function yields the current CPU to other tasks. If there are no
4009 * other threads running on this CPU then this function will return.
4013 SYSCALL_DEFINE0(sched_yield)
4015 struct rq *rq = this_rq_lock();
4017 schedstat_inc(rq, yld_count);
4018 current->sched_class->yield_task(rq);
4021 * Since we are going to call schedule() anyway, there's
4022 * no need to preempt or enable interrupts:
4024 __release(rq->lock);
4025 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4026 do_raw_spin_unlock(&rq->lock);
4027 sched_preempt_enable_no_resched();
4034 static void __cond_resched(void)
4036 __preempt_count_add(PREEMPT_ACTIVE);
4038 __preempt_count_sub(PREEMPT_ACTIVE);
4041 int __sched _cond_resched(void)
4043 if (should_resched()) {
4049 EXPORT_SYMBOL(_cond_resched);
4052 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4053 * call schedule, and on return reacquire the lock.
4055 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4056 * operations here to prevent schedule() from being called twice (once via
4057 * spin_unlock(), once by hand).
4059 int __cond_resched_lock(spinlock_t *lock)
4061 int resched = should_resched();
4064 lockdep_assert_held(lock);
4066 if (spin_needbreak(lock) || resched) {
4077 EXPORT_SYMBOL(__cond_resched_lock);
4079 int __sched __cond_resched_softirq(void)
4081 BUG_ON(!in_softirq());
4083 if (should_resched()) {
4091 EXPORT_SYMBOL(__cond_resched_softirq);
4094 * yield - yield the current processor to other threads.
4096 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4098 * The scheduler is at all times free to pick the calling task as the most
4099 * eligible task to run, if removing the yield() call from your code breaks
4100 * it, its already broken.
4102 * Typical broken usage is:
4107 * where one assumes that yield() will let 'the other' process run that will
4108 * make event true. If the current task is a SCHED_FIFO task that will never
4109 * happen. Never use yield() as a progress guarantee!!
4111 * If you want to use yield() to wait for something, use wait_event().
4112 * If you want to use yield() to be 'nice' for others, use cond_resched().
4113 * If you still want to use yield(), do not!
4115 void __sched yield(void)
4117 set_current_state(TASK_RUNNING);
4120 EXPORT_SYMBOL(yield);
4123 * yield_to - yield the current processor to another thread in
4124 * your thread group, or accelerate that thread toward the
4125 * processor it's on.
4127 * @preempt: whether task preemption is allowed or not
4129 * It's the caller's job to ensure that the target task struct
4130 * can't go away on us before we can do any checks.
4133 * true (>0) if we indeed boosted the target task.
4134 * false (0) if we failed to boost the target.
4135 * -ESRCH if there's no task to yield to.
4137 bool __sched yield_to(struct task_struct *p, bool preempt)
4139 struct task_struct *curr = current;
4140 struct rq *rq, *p_rq;
4141 unsigned long flags;
4144 local_irq_save(flags);
4150 * If we're the only runnable task on the rq and target rq also
4151 * has only one task, there's absolutely no point in yielding.
4153 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4158 double_rq_lock(rq, p_rq);
4159 if (task_rq(p) != p_rq) {
4160 double_rq_unlock(rq, p_rq);
4164 if (!curr->sched_class->yield_to_task)
4167 if (curr->sched_class != p->sched_class)
4170 if (task_running(p_rq, p) || p->state)
4173 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4175 schedstat_inc(rq, yld_count);
4177 * Make p's CPU reschedule; pick_next_entity takes care of
4180 if (preempt && rq != p_rq)
4181 resched_task(p_rq->curr);
4185 double_rq_unlock(rq, p_rq);
4187 local_irq_restore(flags);
4194 EXPORT_SYMBOL_GPL(yield_to);
4197 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4198 * that process accounting knows that this is a task in IO wait state.
4200 void __sched io_schedule(void)
4202 struct rq *rq = raw_rq();
4204 delayacct_blkio_start();
4205 atomic_inc(&rq->nr_iowait);
4206 blk_flush_plug(current);
4207 current->in_iowait = 1;
4209 current->in_iowait = 0;
4210 atomic_dec(&rq->nr_iowait);
4211 delayacct_blkio_end();
4213 EXPORT_SYMBOL(io_schedule);
4215 long __sched io_schedule_timeout(long timeout)
4217 struct rq *rq = raw_rq();
4220 delayacct_blkio_start();
4221 atomic_inc(&rq->nr_iowait);
4222 blk_flush_plug(current);
4223 current->in_iowait = 1;
4224 ret = schedule_timeout(timeout);
4225 current->in_iowait = 0;
4226 atomic_dec(&rq->nr_iowait);
4227 delayacct_blkio_end();
4232 * sys_sched_get_priority_max - return maximum RT priority.
4233 * @policy: scheduling class.
4235 * Return: On success, this syscall returns the maximum
4236 * rt_priority that can be used by a given scheduling class.
4237 * On failure, a negative error code is returned.
4239 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4246 ret = MAX_USER_RT_PRIO-1;
4248 case SCHED_DEADLINE:
4259 * sys_sched_get_priority_min - return minimum RT priority.
4260 * @policy: scheduling class.
4262 * Return: On success, this syscall returns the minimum
4263 * rt_priority that can be used by a given scheduling class.
4264 * On failure, a negative error code is returned.
4266 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4275 case SCHED_DEADLINE:
4285 * sys_sched_rr_get_interval - return the default timeslice of a process.
4286 * @pid: pid of the process.
4287 * @interval: userspace pointer to the timeslice value.
4289 * this syscall writes the default timeslice value of a given process
4290 * into the user-space timespec buffer. A value of '0' means infinity.
4292 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4295 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4296 struct timespec __user *, interval)
4298 struct task_struct *p;
4299 unsigned int time_slice;
4300 unsigned long flags;
4310 p = find_process_by_pid(pid);
4314 retval = security_task_getscheduler(p);
4318 rq = task_rq_lock(p, &flags);
4319 time_slice = p->sched_class->get_rr_interval(rq, p);
4320 task_rq_unlock(rq, p, &flags);
4323 jiffies_to_timespec(time_slice, &t);
4324 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4332 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4334 void sched_show_task(struct task_struct *p)
4336 unsigned long free = 0;
4340 state = p->state ? __ffs(p->state) + 1 : 0;
4341 printk(KERN_INFO "%-15.15s %c", p->comm,
4342 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4343 #if BITS_PER_LONG == 32
4344 if (state == TASK_RUNNING)
4345 printk(KERN_CONT " running ");
4347 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4349 if (state == TASK_RUNNING)
4350 printk(KERN_CONT " running task ");
4352 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4354 #ifdef CONFIG_DEBUG_STACK_USAGE
4355 free = stack_not_used(p);
4358 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4360 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4361 task_pid_nr(p), ppid,
4362 (unsigned long)task_thread_info(p)->flags);
4364 print_worker_info(KERN_INFO, p);
4365 show_stack(p, NULL);
4368 void show_state_filter(unsigned long state_filter)
4370 struct task_struct *g, *p;
4372 #if BITS_PER_LONG == 32
4374 " task PC stack pid father\n");
4377 " task PC stack pid father\n");
4380 do_each_thread(g, p) {
4382 * reset the NMI-timeout, listing all files on a slow
4383 * console might take a lot of time:
4385 touch_nmi_watchdog();
4386 if (!state_filter || (p->state & state_filter))
4388 } while_each_thread(g, p);
4390 touch_all_softlockup_watchdogs();
4392 #ifdef CONFIG_SCHED_DEBUG
4393 sysrq_sched_debug_show();
4397 * Only show locks if all tasks are dumped:
4400 debug_show_all_locks();
4403 void init_idle_bootup_task(struct task_struct *idle)
4405 idle->sched_class = &idle_sched_class;
4409 * init_idle - set up an idle thread for a given CPU
4410 * @idle: task in question
4411 * @cpu: cpu the idle task belongs to
4413 * NOTE: this function does not set the idle thread's NEED_RESCHED
4414 * flag, to make booting more robust.
4416 void init_idle(struct task_struct *idle, int cpu)
4418 struct rq *rq = cpu_rq(cpu);
4419 unsigned long flags;
4421 raw_spin_lock_irqsave(&rq->lock, flags);
4423 __sched_fork(0, idle);
4424 idle->state = TASK_RUNNING;
4425 idle->se.exec_start = sched_clock();
4427 do_set_cpus_allowed(idle, cpumask_of(cpu));
4429 * We're having a chicken and egg problem, even though we are
4430 * holding rq->lock, the cpu isn't yet set to this cpu so the
4431 * lockdep check in task_group() will fail.
4433 * Similar case to sched_fork(). / Alternatively we could
4434 * use task_rq_lock() here and obtain the other rq->lock.
4439 __set_task_cpu(idle, cpu);
4442 rq->curr = rq->idle = idle;
4443 #if defined(CONFIG_SMP)
4446 raw_spin_unlock_irqrestore(&rq->lock, flags);
4448 /* Set the preempt count _outside_ the spinlocks! */
4449 init_idle_preempt_count(idle, cpu);
4452 * The idle tasks have their own, simple scheduling class:
4454 idle->sched_class = &idle_sched_class;
4455 ftrace_graph_init_idle_task(idle, cpu);
4456 vtime_init_idle(idle, cpu);
4457 #if defined(CONFIG_SMP)
4458 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4463 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4465 if (p->sched_class && p->sched_class->set_cpus_allowed)
4466 p->sched_class->set_cpus_allowed(p, new_mask);
4468 cpumask_copy(&p->cpus_allowed, new_mask);
4469 p->nr_cpus_allowed = cpumask_weight(new_mask);
4473 * This is how migration works:
4475 * 1) we invoke migration_cpu_stop() on the target CPU using
4477 * 2) stopper starts to run (implicitly forcing the migrated thread
4479 * 3) it checks whether the migrated task is still in the wrong runqueue.
4480 * 4) if it's in the wrong runqueue then the migration thread removes
4481 * it and puts it into the right queue.
4482 * 5) stopper completes and stop_one_cpu() returns and the migration
4487 * Change a given task's CPU affinity. Migrate the thread to a
4488 * proper CPU and schedule it away if the CPU it's executing on
4489 * is removed from the allowed bitmask.
4491 * NOTE: the caller must have a valid reference to the task, the
4492 * task must not exit() & deallocate itself prematurely. The
4493 * call is not atomic; no spinlocks may be held.
4495 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4497 unsigned long flags;
4499 unsigned int dest_cpu;
4502 rq = task_rq_lock(p, &flags);
4504 if (cpumask_equal(&p->cpus_allowed, new_mask))
4507 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4512 do_set_cpus_allowed(p, new_mask);
4514 /* Can the task run on the task's current CPU? If so, we're done */
4515 if (cpumask_test_cpu(task_cpu(p), new_mask))
4518 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4520 struct migration_arg arg = { p, dest_cpu };
4521 /* Need help from migration thread: drop lock and wait. */
4522 task_rq_unlock(rq, p, &flags);
4523 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4524 tlb_migrate_finish(p->mm);
4528 task_rq_unlock(rq, p, &flags);
4532 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4535 * When dealing with a -deadline task, we have to check if moving it to
4536 * a new CPU is possible or not. In fact, this is only true iff there
4537 * is enough bandwidth available on such CPU, otherwise we want the
4538 * whole migration progedure to fail over.
4541 bool set_task_cpu_dl(struct task_struct *p, unsigned int cpu)
4543 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
4544 struct dl_bw *cpu_b = dl_bw_of(cpu);
4551 raw_spin_lock(&dl_b->lock);
4552 raw_spin_lock(&cpu_b->lock);
4554 bw = cpu_b->bw * cpumask_weight(cpu_rq(cpu)->rd->span);
4555 if (dl_bandwidth_enabled() &&
4556 bw < cpu_b->total_bw + p->dl.dl_bw) {
4560 dl_b->total_bw -= p->dl.dl_bw;
4561 cpu_b->total_bw += p->dl.dl_bw;
4564 raw_spin_unlock(&cpu_b->lock);
4565 raw_spin_unlock(&dl_b->lock);
4571 * Move (not current) task off this cpu, onto dest cpu. We're doing
4572 * this because either it can't run here any more (set_cpus_allowed()
4573 * away from this CPU, or CPU going down), or because we're
4574 * attempting to rebalance this task on exec (sched_exec).
4576 * So we race with normal scheduler movements, but that's OK, as long
4577 * as the task is no longer on this CPU.
4579 * Returns non-zero if task was successfully migrated.
4581 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4583 struct rq *rq_dest, *rq_src;
4586 if (unlikely(!cpu_active(dest_cpu)))
4589 rq_src = cpu_rq(src_cpu);
4590 rq_dest = cpu_rq(dest_cpu);
4592 raw_spin_lock(&p->pi_lock);
4593 double_rq_lock(rq_src, rq_dest);
4594 /* Already moved. */
4595 if (task_cpu(p) != src_cpu)
4597 /* Affinity changed (again). */
4598 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4602 * If p is -deadline, proceed only if there is enough
4603 * bandwidth available on dest_cpu
4605 if (unlikely(dl_task(p)) && !set_task_cpu_dl(p, dest_cpu))
4609 * If we're not on a rq, the next wake-up will ensure we're
4613 dequeue_task(rq_src, p, 0);
4614 set_task_cpu(p, dest_cpu);
4615 enqueue_task(rq_dest, p, 0);
4616 check_preempt_curr(rq_dest, p, 0);
4621 double_rq_unlock(rq_src, rq_dest);
4622 raw_spin_unlock(&p->pi_lock);
4626 #ifdef CONFIG_NUMA_BALANCING
4627 /* Migrate current task p to target_cpu */
4628 int migrate_task_to(struct task_struct *p, int target_cpu)
4630 struct migration_arg arg = { p, target_cpu };
4631 int curr_cpu = task_cpu(p);
4633 if (curr_cpu == target_cpu)
4636 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4639 /* TODO: This is not properly updating schedstats */
4641 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4645 * Requeue a task on a given node and accurately track the number of NUMA
4646 * tasks on the runqueues
4648 void sched_setnuma(struct task_struct *p, int nid)
4651 unsigned long flags;
4652 bool on_rq, running;
4654 rq = task_rq_lock(p, &flags);
4656 running = task_current(rq, p);
4659 dequeue_task(rq, p, 0);
4661 p->sched_class->put_prev_task(rq, p);
4663 p->numa_preferred_nid = nid;
4666 p->sched_class->set_curr_task(rq);
4668 enqueue_task(rq, p, 0);
4669 task_rq_unlock(rq, p, &flags);
4674 * migration_cpu_stop - this will be executed by a highprio stopper thread
4675 * and performs thread migration by bumping thread off CPU then
4676 * 'pushing' onto another runqueue.
4678 static int migration_cpu_stop(void *data)
4680 struct migration_arg *arg = data;
4683 * The original target cpu might have gone down and we might
4684 * be on another cpu but it doesn't matter.
4686 local_irq_disable();
4687 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4692 #ifdef CONFIG_HOTPLUG_CPU
4695 * Ensures that the idle task is using init_mm right before its cpu goes
4698 void idle_task_exit(void)
4700 struct mm_struct *mm = current->active_mm;
4702 BUG_ON(cpu_online(smp_processor_id()));
4705 switch_mm(mm, &init_mm, current);
4710 * Since this CPU is going 'away' for a while, fold any nr_active delta
4711 * we might have. Assumes we're called after migrate_tasks() so that the
4712 * nr_active count is stable.
4714 * Also see the comment "Global load-average calculations".
4716 static void calc_load_migrate(struct rq *rq)
4718 long delta = calc_load_fold_active(rq);
4720 atomic_long_add(delta, &calc_load_tasks);
4724 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4725 * try_to_wake_up()->select_task_rq().
4727 * Called with rq->lock held even though we'er in stop_machine() and
4728 * there's no concurrency possible, we hold the required locks anyway
4729 * because of lock validation efforts.
4731 static void migrate_tasks(unsigned int dead_cpu)
4733 struct rq *rq = cpu_rq(dead_cpu);
4734 struct task_struct *next, *stop = rq->stop;
4738 * Fudge the rq selection such that the below task selection loop
4739 * doesn't get stuck on the currently eligible stop task.
4741 * We're currently inside stop_machine() and the rq is either stuck
4742 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4743 * either way we should never end up calling schedule() until we're
4749 * put_prev_task() and pick_next_task() sched
4750 * class method both need to have an up-to-date
4751 * value of rq->clock[_task]
4753 update_rq_clock(rq);
4757 * There's this thread running, bail when that's the only
4760 if (rq->nr_running == 1)
4763 next = pick_next_task(rq);
4765 next->sched_class->put_prev_task(rq, next);
4767 /* Find suitable destination for @next, with force if needed. */
4768 dest_cpu = select_fallback_rq(dead_cpu, next);
4769 raw_spin_unlock(&rq->lock);
4771 __migrate_task(next, dead_cpu, dest_cpu);
4773 raw_spin_lock(&rq->lock);
4779 #endif /* CONFIG_HOTPLUG_CPU */
4781 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4783 static struct ctl_table sd_ctl_dir[] = {
4785 .procname = "sched_domain",
4791 static struct ctl_table sd_ctl_root[] = {
4793 .procname = "kernel",
4795 .child = sd_ctl_dir,
4800 static struct ctl_table *sd_alloc_ctl_entry(int n)
4802 struct ctl_table *entry =
4803 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4808 static void sd_free_ctl_entry(struct ctl_table **tablep)
4810 struct ctl_table *entry;
4813 * In the intermediate directories, both the child directory and
4814 * procname are dynamically allocated and could fail but the mode
4815 * will always be set. In the lowest directory the names are
4816 * static strings and all have proc handlers.
4818 for (entry = *tablep; entry->mode; entry++) {
4820 sd_free_ctl_entry(&entry->child);
4821 if (entry->proc_handler == NULL)
4822 kfree(entry->procname);
4829 static int min_load_idx = 0;
4830 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4833 set_table_entry(struct ctl_table *entry,
4834 const char *procname, void *data, int maxlen,
4835 umode_t mode, proc_handler *proc_handler,
4838 entry->procname = procname;
4840 entry->maxlen = maxlen;
4842 entry->proc_handler = proc_handler;
4845 entry->extra1 = &min_load_idx;
4846 entry->extra2 = &max_load_idx;
4850 static struct ctl_table *
4851 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4853 struct ctl_table *table = sd_alloc_ctl_entry(13);
4858 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4859 sizeof(long), 0644, proc_doulongvec_minmax, false);
4860 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4861 sizeof(long), 0644, proc_doulongvec_minmax, false);
4862 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4863 sizeof(int), 0644, proc_dointvec_minmax, true);
4864 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4865 sizeof(int), 0644, proc_dointvec_minmax, true);
4866 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4867 sizeof(int), 0644, proc_dointvec_minmax, true);
4868 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4869 sizeof(int), 0644, proc_dointvec_minmax, true);
4870 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4871 sizeof(int), 0644, proc_dointvec_minmax, true);
4872 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4873 sizeof(int), 0644, proc_dointvec_minmax, false);
4874 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4875 sizeof(int), 0644, proc_dointvec_minmax, false);
4876 set_table_entry(&table[9], "cache_nice_tries",
4877 &sd->cache_nice_tries,
4878 sizeof(int), 0644, proc_dointvec_minmax, false);
4879 set_table_entry(&table[10], "flags", &sd->flags,
4880 sizeof(int), 0644, proc_dointvec_minmax, false);
4881 set_table_entry(&table[11], "name", sd->name,
4882 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4883 /* &table[12] is terminator */
4888 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4890 struct ctl_table *entry, *table;
4891 struct sched_domain *sd;
4892 int domain_num = 0, i;
4895 for_each_domain(cpu, sd)
4897 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4902 for_each_domain(cpu, sd) {
4903 snprintf(buf, 32, "domain%d", i);
4904 entry->procname = kstrdup(buf, GFP_KERNEL);
4906 entry->child = sd_alloc_ctl_domain_table(sd);
4913 static struct ctl_table_header *sd_sysctl_header;
4914 static void register_sched_domain_sysctl(void)
4916 int i, cpu_num = num_possible_cpus();
4917 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4920 WARN_ON(sd_ctl_dir[0].child);
4921 sd_ctl_dir[0].child = entry;
4926 for_each_possible_cpu(i) {
4927 snprintf(buf, 32, "cpu%d", i);
4928 entry->procname = kstrdup(buf, GFP_KERNEL);
4930 entry->child = sd_alloc_ctl_cpu_table(i);
4934 WARN_ON(sd_sysctl_header);
4935 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4938 /* may be called multiple times per register */
4939 static void unregister_sched_domain_sysctl(void)
4941 if (sd_sysctl_header)
4942 unregister_sysctl_table(sd_sysctl_header);
4943 sd_sysctl_header = NULL;
4944 if (sd_ctl_dir[0].child)
4945 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4948 static void register_sched_domain_sysctl(void)
4951 static void unregister_sched_domain_sysctl(void)
4956 static void set_rq_online(struct rq *rq)
4959 const struct sched_class *class;
4961 cpumask_set_cpu(rq->cpu, rq->rd->online);
4964 for_each_class(class) {
4965 if (class->rq_online)
4966 class->rq_online(rq);
4971 static void set_rq_offline(struct rq *rq)
4974 const struct sched_class *class;
4976 for_each_class(class) {
4977 if (class->rq_offline)
4978 class->rq_offline(rq);
4981 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4987 * migration_call - callback that gets triggered when a CPU is added.
4988 * Here we can start up the necessary migration thread for the new CPU.
4991 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4993 int cpu = (long)hcpu;
4994 unsigned long flags;
4995 struct rq *rq = cpu_rq(cpu);
4997 switch (action & ~CPU_TASKS_FROZEN) {
4999 case CPU_UP_PREPARE:
5000 rq->calc_load_update = calc_load_update;
5004 /* Update our root-domain */
5005 raw_spin_lock_irqsave(&rq->lock, flags);
5007 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5011 raw_spin_unlock_irqrestore(&rq->lock, flags);
5014 #ifdef CONFIG_HOTPLUG_CPU
5016 sched_ttwu_pending();
5017 /* Update our root-domain */
5018 raw_spin_lock_irqsave(&rq->lock, flags);
5020 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5024 BUG_ON(rq->nr_running != 1); /* the migration thread */
5025 raw_spin_unlock_irqrestore(&rq->lock, flags);
5029 calc_load_migrate(rq);
5034 update_max_interval();
5040 * Register at high priority so that task migration (migrate_all_tasks)
5041 * happens before everything else. This has to be lower priority than
5042 * the notifier in the perf_event subsystem, though.
5044 static struct notifier_block migration_notifier = {
5045 .notifier_call = migration_call,
5046 .priority = CPU_PRI_MIGRATION,
5049 static int sched_cpu_active(struct notifier_block *nfb,
5050 unsigned long action, void *hcpu)
5052 switch (action & ~CPU_TASKS_FROZEN) {
5054 case CPU_DOWN_FAILED:
5055 set_cpu_active((long)hcpu, true);
5062 static int sched_cpu_inactive(struct notifier_block *nfb,
5063 unsigned long action, void *hcpu)
5065 switch (action & ~CPU_TASKS_FROZEN) {
5066 case CPU_DOWN_PREPARE:
5067 set_cpu_active((long)hcpu, false);
5074 static int __init migration_init(void)
5076 void *cpu = (void *)(long)smp_processor_id();
5079 /* Initialize migration for the boot CPU */
5080 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5081 BUG_ON(err == NOTIFY_BAD);
5082 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5083 register_cpu_notifier(&migration_notifier);
5085 /* Register cpu active notifiers */
5086 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5087 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5091 early_initcall(migration_init);
5096 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5098 #ifdef CONFIG_SCHED_DEBUG
5100 static __read_mostly int sched_debug_enabled;
5102 static int __init sched_debug_setup(char *str)
5104 sched_debug_enabled = 1;
5108 early_param("sched_debug", sched_debug_setup);
5110 static inline bool sched_debug(void)
5112 return sched_debug_enabled;
5115 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5116 struct cpumask *groupmask)
5118 struct sched_group *group = sd->groups;
5121 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5122 cpumask_clear(groupmask);
5124 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5126 if (!(sd->flags & SD_LOAD_BALANCE)) {
5127 printk("does not load-balance\n");
5129 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5134 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5136 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5137 printk(KERN_ERR "ERROR: domain->span does not contain "
5140 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5141 printk(KERN_ERR "ERROR: domain->groups does not contain"
5145 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5149 printk(KERN_ERR "ERROR: group is NULL\n");
5154 * Even though we initialize ->power to something semi-sane,
5155 * we leave power_orig unset. This allows us to detect if
5156 * domain iteration is still funny without causing /0 traps.
5158 if (!group->sgp->power_orig) {
5159 printk(KERN_CONT "\n");
5160 printk(KERN_ERR "ERROR: domain->cpu_power not "
5165 if (!cpumask_weight(sched_group_cpus(group))) {
5166 printk(KERN_CONT "\n");
5167 printk(KERN_ERR "ERROR: empty group\n");
5171 if (!(sd->flags & SD_OVERLAP) &&
5172 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5173 printk(KERN_CONT "\n");
5174 printk(KERN_ERR "ERROR: repeated CPUs\n");
5178 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5180 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5182 printk(KERN_CONT " %s", str);
5183 if (group->sgp->power != SCHED_POWER_SCALE) {
5184 printk(KERN_CONT " (cpu_power = %d)",
5188 group = group->next;
5189 } while (group != sd->groups);
5190 printk(KERN_CONT "\n");
5192 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5193 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5196 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5197 printk(KERN_ERR "ERROR: parent span is not a superset "
5198 "of domain->span\n");
5202 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5206 if (!sched_debug_enabled)
5210 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5214 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5217 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5225 #else /* !CONFIG_SCHED_DEBUG */
5226 # define sched_domain_debug(sd, cpu) do { } while (0)
5227 static inline bool sched_debug(void)
5231 #endif /* CONFIG_SCHED_DEBUG */
5233 static int sd_degenerate(struct sched_domain *sd)
5235 if (cpumask_weight(sched_domain_span(sd)) == 1)
5238 /* Following flags need at least 2 groups */
5239 if (sd->flags & (SD_LOAD_BALANCE |
5240 SD_BALANCE_NEWIDLE |
5244 SD_SHARE_PKG_RESOURCES)) {
5245 if (sd->groups != sd->groups->next)
5249 /* Following flags don't use groups */
5250 if (sd->flags & (SD_WAKE_AFFINE))
5257 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5259 unsigned long cflags = sd->flags, pflags = parent->flags;
5261 if (sd_degenerate(parent))
5264 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5267 /* Flags needing groups don't count if only 1 group in parent */
5268 if (parent->groups == parent->groups->next) {
5269 pflags &= ~(SD_LOAD_BALANCE |
5270 SD_BALANCE_NEWIDLE |
5274 SD_SHARE_PKG_RESOURCES |
5276 if (nr_node_ids == 1)
5277 pflags &= ~SD_SERIALIZE;
5279 if (~cflags & pflags)
5285 static void free_rootdomain(struct rcu_head *rcu)
5287 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5289 cpupri_cleanup(&rd->cpupri);
5290 cpudl_cleanup(&rd->cpudl);
5291 free_cpumask_var(rd->dlo_mask);
5292 free_cpumask_var(rd->rto_mask);
5293 free_cpumask_var(rd->online);
5294 free_cpumask_var(rd->span);
5298 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5300 struct root_domain *old_rd = NULL;
5301 unsigned long flags;
5303 raw_spin_lock_irqsave(&rq->lock, flags);
5308 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5311 cpumask_clear_cpu(rq->cpu, old_rd->span);
5314 * If we dont want to free the old_rd yet then
5315 * set old_rd to NULL to skip the freeing later
5318 if (!atomic_dec_and_test(&old_rd->refcount))
5322 atomic_inc(&rd->refcount);
5325 cpumask_set_cpu(rq->cpu, rd->span);
5326 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5329 raw_spin_unlock_irqrestore(&rq->lock, flags);
5332 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5335 static int init_rootdomain(struct root_domain *rd)
5337 memset(rd, 0, sizeof(*rd));
5339 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5341 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5343 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5345 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5348 init_dl_bw(&rd->dl_bw);
5349 if (cpudl_init(&rd->cpudl) != 0)
5352 if (cpupri_init(&rd->cpupri) != 0)
5357 free_cpumask_var(rd->rto_mask);
5359 free_cpumask_var(rd->dlo_mask);
5361 free_cpumask_var(rd->online);
5363 free_cpumask_var(rd->span);
5369 * By default the system creates a single root-domain with all cpus as
5370 * members (mimicking the global state we have today).
5372 struct root_domain def_root_domain;
5374 static void init_defrootdomain(void)
5376 init_rootdomain(&def_root_domain);
5378 atomic_set(&def_root_domain.refcount, 1);
5381 static struct root_domain *alloc_rootdomain(void)
5383 struct root_domain *rd;
5385 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5389 if (init_rootdomain(rd) != 0) {
5397 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5399 struct sched_group *tmp, *first;
5408 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5413 } while (sg != first);
5416 static void free_sched_domain(struct rcu_head *rcu)
5418 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5421 * If its an overlapping domain it has private groups, iterate and
5424 if (sd->flags & SD_OVERLAP) {
5425 free_sched_groups(sd->groups, 1);
5426 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5427 kfree(sd->groups->sgp);
5433 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5435 call_rcu(&sd->rcu, free_sched_domain);
5438 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5440 for (; sd; sd = sd->parent)
5441 destroy_sched_domain(sd, cpu);
5445 * Keep a special pointer to the highest sched_domain that has
5446 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5447 * allows us to avoid some pointer chasing select_idle_sibling().
5449 * Also keep a unique ID per domain (we use the first cpu number in
5450 * the cpumask of the domain), this allows us to quickly tell if
5451 * two cpus are in the same cache domain, see cpus_share_cache().
5453 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5454 DEFINE_PER_CPU(int, sd_llc_size);
5455 DEFINE_PER_CPU(int, sd_llc_id);
5456 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5457 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5458 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5460 static void update_top_cache_domain(int cpu)
5462 struct sched_domain *sd;
5463 struct sched_domain *busy_sd = NULL;
5467 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5469 id = cpumask_first(sched_domain_span(sd));
5470 size = cpumask_weight(sched_domain_span(sd));
5471 busy_sd = sd->parent; /* sd_busy */
5473 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5475 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5476 per_cpu(sd_llc_size, cpu) = size;
5477 per_cpu(sd_llc_id, cpu) = id;
5479 sd = lowest_flag_domain(cpu, SD_NUMA);
5480 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5482 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5483 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5487 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5488 * hold the hotplug lock.
5491 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5493 struct rq *rq = cpu_rq(cpu);
5494 struct sched_domain *tmp;
5496 /* Remove the sched domains which do not contribute to scheduling. */
5497 for (tmp = sd; tmp; ) {
5498 struct sched_domain *parent = tmp->parent;
5502 if (sd_parent_degenerate(tmp, parent)) {
5503 tmp->parent = parent->parent;
5505 parent->parent->child = tmp;
5507 * Transfer SD_PREFER_SIBLING down in case of a
5508 * degenerate parent; the spans match for this
5509 * so the property transfers.
5511 if (parent->flags & SD_PREFER_SIBLING)
5512 tmp->flags |= SD_PREFER_SIBLING;
5513 destroy_sched_domain(parent, cpu);
5518 if (sd && sd_degenerate(sd)) {
5521 destroy_sched_domain(tmp, cpu);
5526 sched_domain_debug(sd, cpu);
5528 rq_attach_root(rq, rd);
5530 rcu_assign_pointer(rq->sd, sd);
5531 destroy_sched_domains(tmp, cpu);
5533 update_top_cache_domain(cpu);
5536 /* cpus with isolated domains */
5537 static cpumask_var_t cpu_isolated_map;
5539 /* Setup the mask of cpus configured for isolated domains */
5540 static int __init isolated_cpu_setup(char *str)
5542 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5543 cpulist_parse(str, cpu_isolated_map);
5547 __setup("isolcpus=", isolated_cpu_setup);
5549 static const struct cpumask *cpu_cpu_mask(int cpu)
5551 return cpumask_of_node(cpu_to_node(cpu));
5555 struct sched_domain **__percpu sd;
5556 struct sched_group **__percpu sg;
5557 struct sched_group_power **__percpu sgp;
5561 struct sched_domain ** __percpu sd;
5562 struct root_domain *rd;
5572 struct sched_domain_topology_level;
5574 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5575 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5577 #define SDTL_OVERLAP 0x01
5579 struct sched_domain_topology_level {
5580 sched_domain_init_f init;
5581 sched_domain_mask_f mask;
5584 struct sd_data data;
5588 * Build an iteration mask that can exclude certain CPUs from the upwards
5591 * Asymmetric node setups can result in situations where the domain tree is of
5592 * unequal depth, make sure to skip domains that already cover the entire
5595 * In that case build_sched_domains() will have terminated the iteration early
5596 * and our sibling sd spans will be empty. Domains should always include the
5597 * cpu they're built on, so check that.
5600 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5602 const struct cpumask *span = sched_domain_span(sd);
5603 struct sd_data *sdd = sd->private;
5604 struct sched_domain *sibling;
5607 for_each_cpu(i, span) {
5608 sibling = *per_cpu_ptr(sdd->sd, i);
5609 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5612 cpumask_set_cpu(i, sched_group_mask(sg));
5617 * Return the canonical balance cpu for this group, this is the first cpu
5618 * of this group that's also in the iteration mask.
5620 int group_balance_cpu(struct sched_group *sg)
5622 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5626 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5628 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5629 const struct cpumask *span = sched_domain_span(sd);
5630 struct cpumask *covered = sched_domains_tmpmask;
5631 struct sd_data *sdd = sd->private;
5632 struct sched_domain *child;
5635 cpumask_clear(covered);
5637 for_each_cpu(i, span) {
5638 struct cpumask *sg_span;
5640 if (cpumask_test_cpu(i, covered))
5643 child = *per_cpu_ptr(sdd->sd, i);
5645 /* See the comment near build_group_mask(). */
5646 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5649 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5650 GFP_KERNEL, cpu_to_node(cpu));
5655 sg_span = sched_group_cpus(sg);
5657 child = child->child;
5658 cpumask_copy(sg_span, sched_domain_span(child));
5660 cpumask_set_cpu(i, sg_span);
5662 cpumask_or(covered, covered, sg_span);
5664 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5665 if (atomic_inc_return(&sg->sgp->ref) == 1)
5666 build_group_mask(sd, sg);
5669 * Initialize sgp->power such that even if we mess up the
5670 * domains and no possible iteration will get us here, we won't
5673 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5674 sg->sgp->power_orig = sg->sgp->power;
5677 * Make sure the first group of this domain contains the
5678 * canonical balance cpu. Otherwise the sched_domain iteration
5679 * breaks. See update_sg_lb_stats().
5681 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5682 group_balance_cpu(sg) == cpu)
5692 sd->groups = groups;
5697 free_sched_groups(first, 0);
5702 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5704 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5705 struct sched_domain *child = sd->child;
5708 cpu = cpumask_first(sched_domain_span(child));
5711 *sg = *per_cpu_ptr(sdd->sg, cpu);
5712 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5713 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5720 * build_sched_groups will build a circular linked list of the groups
5721 * covered by the given span, and will set each group's ->cpumask correctly,
5722 * and ->cpu_power to 0.
5724 * Assumes the sched_domain tree is fully constructed
5727 build_sched_groups(struct sched_domain *sd, int cpu)
5729 struct sched_group *first = NULL, *last = NULL;
5730 struct sd_data *sdd = sd->private;
5731 const struct cpumask *span = sched_domain_span(sd);
5732 struct cpumask *covered;
5735 get_group(cpu, sdd, &sd->groups);
5736 atomic_inc(&sd->groups->ref);
5738 if (cpu != cpumask_first(span))
5741 lockdep_assert_held(&sched_domains_mutex);
5742 covered = sched_domains_tmpmask;
5744 cpumask_clear(covered);
5746 for_each_cpu(i, span) {
5747 struct sched_group *sg;
5750 if (cpumask_test_cpu(i, covered))
5753 group = get_group(i, sdd, &sg);
5754 cpumask_clear(sched_group_cpus(sg));
5756 cpumask_setall(sched_group_mask(sg));
5758 for_each_cpu(j, span) {
5759 if (get_group(j, sdd, NULL) != group)
5762 cpumask_set_cpu(j, covered);
5763 cpumask_set_cpu(j, sched_group_cpus(sg));
5778 * Initialize sched groups cpu_power.
5780 * cpu_power indicates the capacity of sched group, which is used while
5781 * distributing the load between different sched groups in a sched domain.
5782 * Typically cpu_power for all the groups in a sched domain will be same unless
5783 * there are asymmetries in the topology. If there are asymmetries, group
5784 * having more cpu_power will pickup more load compared to the group having
5787 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5789 struct sched_group *sg = sd->groups;
5794 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5796 } while (sg != sd->groups);
5798 if (cpu != group_balance_cpu(sg))
5801 update_group_power(sd, cpu);
5802 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5805 int __weak arch_sd_sibling_asym_packing(void)
5807 return 0*SD_ASYM_PACKING;
5811 * Initializers for schedule domains
5812 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5815 #ifdef CONFIG_SCHED_DEBUG
5816 # define SD_INIT_NAME(sd, type) sd->name = #type
5818 # define SD_INIT_NAME(sd, type) do { } while (0)
5821 #define SD_INIT_FUNC(type) \
5822 static noinline struct sched_domain * \
5823 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5825 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5826 *sd = SD_##type##_INIT; \
5827 SD_INIT_NAME(sd, type); \
5828 sd->private = &tl->data; \
5833 #ifdef CONFIG_SCHED_SMT
5834 SD_INIT_FUNC(SIBLING)
5836 #ifdef CONFIG_SCHED_MC
5839 #ifdef CONFIG_SCHED_BOOK
5843 static int default_relax_domain_level = -1;
5844 int sched_domain_level_max;
5846 static int __init setup_relax_domain_level(char *str)
5848 if (kstrtoint(str, 0, &default_relax_domain_level))
5849 pr_warn("Unable to set relax_domain_level\n");
5853 __setup("relax_domain_level=", setup_relax_domain_level);
5855 static void set_domain_attribute(struct sched_domain *sd,
5856 struct sched_domain_attr *attr)
5860 if (!attr || attr->relax_domain_level < 0) {
5861 if (default_relax_domain_level < 0)
5864 request = default_relax_domain_level;
5866 request = attr->relax_domain_level;
5867 if (request < sd->level) {
5868 /* turn off idle balance on this domain */
5869 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5871 /* turn on idle balance on this domain */
5872 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5876 static void __sdt_free(const struct cpumask *cpu_map);
5877 static int __sdt_alloc(const struct cpumask *cpu_map);
5879 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5880 const struct cpumask *cpu_map)
5884 if (!atomic_read(&d->rd->refcount))
5885 free_rootdomain(&d->rd->rcu); /* fall through */
5887 free_percpu(d->sd); /* fall through */
5889 __sdt_free(cpu_map); /* fall through */
5895 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5896 const struct cpumask *cpu_map)
5898 memset(d, 0, sizeof(*d));
5900 if (__sdt_alloc(cpu_map))
5901 return sa_sd_storage;
5902 d->sd = alloc_percpu(struct sched_domain *);
5904 return sa_sd_storage;
5905 d->rd = alloc_rootdomain();
5908 return sa_rootdomain;
5912 * NULL the sd_data elements we've used to build the sched_domain and
5913 * sched_group structure so that the subsequent __free_domain_allocs()
5914 * will not free the data we're using.
5916 static void claim_allocations(int cpu, struct sched_domain *sd)
5918 struct sd_data *sdd = sd->private;
5920 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5921 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5923 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5924 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5926 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5927 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5930 #ifdef CONFIG_SCHED_SMT
5931 static const struct cpumask *cpu_smt_mask(int cpu)
5933 return topology_thread_cpumask(cpu);
5938 * Topology list, bottom-up.
5940 static struct sched_domain_topology_level default_topology[] = {
5941 #ifdef CONFIG_SCHED_SMT
5942 { sd_init_SIBLING, cpu_smt_mask, },
5944 #ifdef CONFIG_SCHED_MC
5945 { sd_init_MC, cpu_coregroup_mask, },
5947 #ifdef CONFIG_SCHED_BOOK
5948 { sd_init_BOOK, cpu_book_mask, },
5950 { sd_init_CPU, cpu_cpu_mask, },
5954 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5956 #define for_each_sd_topology(tl) \
5957 for (tl = sched_domain_topology; tl->init; tl++)
5961 static int sched_domains_numa_levels;
5962 static int *sched_domains_numa_distance;
5963 static struct cpumask ***sched_domains_numa_masks;
5964 static int sched_domains_curr_level;
5966 static inline int sd_local_flags(int level)
5968 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5971 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5974 static struct sched_domain *
5975 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5977 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5978 int level = tl->numa_level;
5979 int sd_weight = cpumask_weight(
5980 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5982 *sd = (struct sched_domain){
5983 .min_interval = sd_weight,
5984 .max_interval = 2*sd_weight,
5986 .imbalance_pct = 125,
5987 .cache_nice_tries = 2,
5994 .flags = 1*SD_LOAD_BALANCE
5995 | 1*SD_BALANCE_NEWIDLE
6000 | 0*SD_SHARE_CPUPOWER
6001 | 0*SD_SHARE_PKG_RESOURCES
6003 | 0*SD_PREFER_SIBLING
6005 | sd_local_flags(level)
6007 .last_balance = jiffies,
6008 .balance_interval = sd_weight,
6010 SD_INIT_NAME(sd, NUMA);
6011 sd->private = &tl->data;
6014 * Ugly hack to pass state to sd_numa_mask()...
6016 sched_domains_curr_level = tl->numa_level;
6021 static const struct cpumask *sd_numa_mask(int cpu)
6023 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6026 static void sched_numa_warn(const char *str)
6028 static int done = false;
6036 printk(KERN_WARNING "ERROR: %s\n\n", str);
6038 for (i = 0; i < nr_node_ids; i++) {
6039 printk(KERN_WARNING " ");
6040 for (j = 0; j < nr_node_ids; j++)
6041 printk(KERN_CONT "%02d ", node_distance(i,j));
6042 printk(KERN_CONT "\n");
6044 printk(KERN_WARNING "\n");
6047 static bool find_numa_distance(int distance)
6051 if (distance == node_distance(0, 0))
6054 for (i = 0; i < sched_domains_numa_levels; i++) {
6055 if (sched_domains_numa_distance[i] == distance)
6062 static void sched_init_numa(void)
6064 int next_distance, curr_distance = node_distance(0, 0);
6065 struct sched_domain_topology_level *tl;
6069 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6070 if (!sched_domains_numa_distance)
6074 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6075 * unique distances in the node_distance() table.
6077 * Assumes node_distance(0,j) includes all distances in
6078 * node_distance(i,j) in order to avoid cubic time.
6080 next_distance = curr_distance;
6081 for (i = 0; i < nr_node_ids; i++) {
6082 for (j = 0; j < nr_node_ids; j++) {
6083 for (k = 0; k < nr_node_ids; k++) {
6084 int distance = node_distance(i, k);
6086 if (distance > curr_distance &&
6087 (distance < next_distance ||
6088 next_distance == curr_distance))
6089 next_distance = distance;
6092 * While not a strong assumption it would be nice to know
6093 * about cases where if node A is connected to B, B is not
6094 * equally connected to A.
6096 if (sched_debug() && node_distance(k, i) != distance)
6097 sched_numa_warn("Node-distance not symmetric");
6099 if (sched_debug() && i && !find_numa_distance(distance))
6100 sched_numa_warn("Node-0 not representative");
6102 if (next_distance != curr_distance) {
6103 sched_domains_numa_distance[level++] = next_distance;
6104 sched_domains_numa_levels = level;
6105 curr_distance = next_distance;
6110 * In case of sched_debug() we verify the above assumption.
6116 * 'level' contains the number of unique distances, excluding the
6117 * identity distance node_distance(i,i).
6119 * The sched_domains_numa_distance[] array includes the actual distance
6124 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6125 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6126 * the array will contain less then 'level' members. This could be
6127 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6128 * in other functions.
6130 * We reset it to 'level' at the end of this function.
6132 sched_domains_numa_levels = 0;
6134 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6135 if (!sched_domains_numa_masks)
6139 * Now for each level, construct a mask per node which contains all
6140 * cpus of nodes that are that many hops away from us.
6142 for (i = 0; i < level; i++) {
6143 sched_domains_numa_masks[i] =
6144 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6145 if (!sched_domains_numa_masks[i])
6148 for (j = 0; j < nr_node_ids; j++) {
6149 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6153 sched_domains_numa_masks[i][j] = mask;
6155 for (k = 0; k < nr_node_ids; k++) {
6156 if (node_distance(j, k) > sched_domains_numa_distance[i])
6159 cpumask_or(mask, mask, cpumask_of_node(k));
6164 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6165 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6170 * Copy the default topology bits..
6172 for (i = 0; default_topology[i].init; i++)
6173 tl[i] = default_topology[i];
6176 * .. and append 'j' levels of NUMA goodness.
6178 for (j = 0; j < level; i++, j++) {
6179 tl[i] = (struct sched_domain_topology_level){
6180 .init = sd_numa_init,
6181 .mask = sd_numa_mask,
6182 .flags = SDTL_OVERLAP,
6187 sched_domain_topology = tl;
6189 sched_domains_numa_levels = level;
6192 static void sched_domains_numa_masks_set(int cpu)
6195 int node = cpu_to_node(cpu);
6197 for (i = 0; i < sched_domains_numa_levels; i++) {
6198 for (j = 0; j < nr_node_ids; j++) {
6199 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6200 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6205 static void sched_domains_numa_masks_clear(int cpu)
6208 for (i = 0; i < sched_domains_numa_levels; i++) {
6209 for (j = 0; j < nr_node_ids; j++)
6210 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6215 * Update sched_domains_numa_masks[level][node] array when new cpus
6218 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6219 unsigned long action,
6222 int cpu = (long)hcpu;
6224 switch (action & ~CPU_TASKS_FROZEN) {
6226 sched_domains_numa_masks_set(cpu);
6230 sched_domains_numa_masks_clear(cpu);
6240 static inline void sched_init_numa(void)
6244 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6245 unsigned long action,
6250 #endif /* CONFIG_NUMA */
6252 static int __sdt_alloc(const struct cpumask *cpu_map)
6254 struct sched_domain_topology_level *tl;
6257 for_each_sd_topology(tl) {
6258 struct sd_data *sdd = &tl->data;
6260 sdd->sd = alloc_percpu(struct sched_domain *);
6264 sdd->sg = alloc_percpu(struct sched_group *);
6268 sdd->sgp = alloc_percpu(struct sched_group_power *);
6272 for_each_cpu(j, cpu_map) {
6273 struct sched_domain *sd;
6274 struct sched_group *sg;
6275 struct sched_group_power *sgp;
6277 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6278 GFP_KERNEL, cpu_to_node(j));
6282 *per_cpu_ptr(sdd->sd, j) = sd;
6284 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6285 GFP_KERNEL, cpu_to_node(j));
6291 *per_cpu_ptr(sdd->sg, j) = sg;
6293 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6294 GFP_KERNEL, cpu_to_node(j));
6298 *per_cpu_ptr(sdd->sgp, j) = sgp;
6305 static void __sdt_free(const struct cpumask *cpu_map)
6307 struct sched_domain_topology_level *tl;
6310 for_each_sd_topology(tl) {
6311 struct sd_data *sdd = &tl->data;
6313 for_each_cpu(j, cpu_map) {
6314 struct sched_domain *sd;
6317 sd = *per_cpu_ptr(sdd->sd, j);
6318 if (sd && (sd->flags & SD_OVERLAP))
6319 free_sched_groups(sd->groups, 0);
6320 kfree(*per_cpu_ptr(sdd->sd, j));
6324 kfree(*per_cpu_ptr(sdd->sg, j));
6326 kfree(*per_cpu_ptr(sdd->sgp, j));
6328 free_percpu(sdd->sd);
6330 free_percpu(sdd->sg);
6332 free_percpu(sdd->sgp);
6337 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6338 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6339 struct sched_domain *child, int cpu)
6341 struct sched_domain *sd = tl->init(tl, cpu);
6345 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6347 sd->level = child->level + 1;
6348 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6352 set_domain_attribute(sd, attr);
6358 * Build sched domains for a given set of cpus and attach the sched domains
6359 * to the individual cpus
6361 static int build_sched_domains(const struct cpumask *cpu_map,
6362 struct sched_domain_attr *attr)
6364 enum s_alloc alloc_state;
6365 struct sched_domain *sd;
6367 int i, ret = -ENOMEM;
6369 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6370 if (alloc_state != sa_rootdomain)
6373 /* Set up domains for cpus specified by the cpu_map. */
6374 for_each_cpu(i, cpu_map) {
6375 struct sched_domain_topology_level *tl;
6378 for_each_sd_topology(tl) {
6379 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6380 if (tl == sched_domain_topology)
6381 *per_cpu_ptr(d.sd, i) = sd;
6382 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6383 sd->flags |= SD_OVERLAP;
6384 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6389 /* Build the groups for the domains */
6390 for_each_cpu(i, cpu_map) {
6391 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6392 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6393 if (sd->flags & SD_OVERLAP) {
6394 if (build_overlap_sched_groups(sd, i))
6397 if (build_sched_groups(sd, i))
6403 /* Calculate CPU power for physical packages and nodes */
6404 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6405 if (!cpumask_test_cpu(i, cpu_map))
6408 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6409 claim_allocations(i, sd);
6410 init_sched_groups_power(i, sd);
6414 /* Attach the domains */
6416 for_each_cpu(i, cpu_map) {
6417 sd = *per_cpu_ptr(d.sd, i);
6418 cpu_attach_domain(sd, d.rd, i);
6424 __free_domain_allocs(&d, alloc_state, cpu_map);
6428 static cpumask_var_t *doms_cur; /* current sched domains */
6429 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6430 static struct sched_domain_attr *dattr_cur;
6431 /* attribues of custom domains in 'doms_cur' */
6434 * Special case: If a kmalloc of a doms_cur partition (array of
6435 * cpumask) fails, then fallback to a single sched domain,
6436 * as determined by the single cpumask fallback_doms.
6438 static cpumask_var_t fallback_doms;
6441 * arch_update_cpu_topology lets virtualized architectures update the
6442 * cpu core maps. It is supposed to return 1 if the topology changed
6443 * or 0 if it stayed the same.
6445 int __attribute__((weak)) arch_update_cpu_topology(void)
6450 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6453 cpumask_var_t *doms;
6455 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6458 for (i = 0; i < ndoms; i++) {
6459 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6460 free_sched_domains(doms, i);
6467 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6470 for (i = 0; i < ndoms; i++)
6471 free_cpumask_var(doms[i]);
6476 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6477 * For now this just excludes isolated cpus, but could be used to
6478 * exclude other special cases in the future.
6480 static int init_sched_domains(const struct cpumask *cpu_map)
6484 arch_update_cpu_topology();
6486 doms_cur = alloc_sched_domains(ndoms_cur);
6488 doms_cur = &fallback_doms;
6489 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6490 err = build_sched_domains(doms_cur[0], NULL);
6491 register_sched_domain_sysctl();
6497 * Detach sched domains from a group of cpus specified in cpu_map
6498 * These cpus will now be attached to the NULL domain
6500 static void detach_destroy_domains(const struct cpumask *cpu_map)
6505 for_each_cpu(i, cpu_map)
6506 cpu_attach_domain(NULL, &def_root_domain, i);
6510 /* handle null as "default" */
6511 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6512 struct sched_domain_attr *new, int idx_new)
6514 struct sched_domain_attr tmp;
6521 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6522 new ? (new + idx_new) : &tmp,
6523 sizeof(struct sched_domain_attr));
6527 * Partition sched domains as specified by the 'ndoms_new'
6528 * cpumasks in the array doms_new[] of cpumasks. This compares
6529 * doms_new[] to the current sched domain partitioning, doms_cur[].
6530 * It destroys each deleted domain and builds each new domain.
6532 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6533 * The masks don't intersect (don't overlap.) We should setup one
6534 * sched domain for each mask. CPUs not in any of the cpumasks will
6535 * not be load balanced. If the same cpumask appears both in the
6536 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6539 * The passed in 'doms_new' should be allocated using
6540 * alloc_sched_domains. This routine takes ownership of it and will
6541 * free_sched_domains it when done with it. If the caller failed the
6542 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6543 * and partition_sched_domains() will fallback to the single partition
6544 * 'fallback_doms', it also forces the domains to be rebuilt.
6546 * If doms_new == NULL it will be replaced with cpu_online_mask.
6547 * ndoms_new == 0 is a special case for destroying existing domains,
6548 * and it will not create the default domain.
6550 * Call with hotplug lock held
6552 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6553 struct sched_domain_attr *dattr_new)
6558 mutex_lock(&sched_domains_mutex);
6560 /* always unregister in case we don't destroy any domains */
6561 unregister_sched_domain_sysctl();
6563 /* Let architecture update cpu core mappings. */
6564 new_topology = arch_update_cpu_topology();
6566 n = doms_new ? ndoms_new : 0;
6568 /* Destroy deleted domains */
6569 for (i = 0; i < ndoms_cur; i++) {
6570 for (j = 0; j < n && !new_topology; j++) {
6571 if (cpumask_equal(doms_cur[i], doms_new[j])
6572 && dattrs_equal(dattr_cur, i, dattr_new, j))
6575 /* no match - a current sched domain not in new doms_new[] */
6576 detach_destroy_domains(doms_cur[i]);
6582 if (doms_new == NULL) {
6584 doms_new = &fallback_doms;
6585 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6586 WARN_ON_ONCE(dattr_new);
6589 /* Build new domains */
6590 for (i = 0; i < ndoms_new; i++) {
6591 for (j = 0; j < n && !new_topology; j++) {
6592 if (cpumask_equal(doms_new[i], doms_cur[j])
6593 && dattrs_equal(dattr_new, i, dattr_cur, j))
6596 /* no match - add a new doms_new */
6597 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6602 /* Remember the new sched domains */
6603 if (doms_cur != &fallback_doms)
6604 free_sched_domains(doms_cur, ndoms_cur);
6605 kfree(dattr_cur); /* kfree(NULL) is safe */
6606 doms_cur = doms_new;
6607 dattr_cur = dattr_new;
6608 ndoms_cur = ndoms_new;
6610 register_sched_domain_sysctl();
6612 mutex_unlock(&sched_domains_mutex);
6615 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6618 * Update cpusets according to cpu_active mask. If cpusets are
6619 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6620 * around partition_sched_domains().
6622 * If we come here as part of a suspend/resume, don't touch cpusets because we
6623 * want to restore it back to its original state upon resume anyway.
6625 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6629 case CPU_ONLINE_FROZEN:
6630 case CPU_DOWN_FAILED_FROZEN:
6633 * num_cpus_frozen tracks how many CPUs are involved in suspend
6634 * resume sequence. As long as this is not the last online
6635 * operation in the resume sequence, just build a single sched
6636 * domain, ignoring cpusets.
6639 if (likely(num_cpus_frozen)) {
6640 partition_sched_domains(1, NULL, NULL);
6645 * This is the last CPU online operation. So fall through and
6646 * restore the original sched domains by considering the
6647 * cpuset configurations.
6651 case CPU_DOWN_FAILED:
6652 cpuset_update_active_cpus(true);
6660 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6664 case CPU_DOWN_PREPARE:
6665 cpuset_update_active_cpus(false);
6667 case CPU_DOWN_PREPARE_FROZEN:
6669 partition_sched_domains(1, NULL, NULL);
6677 void __init sched_init_smp(void)
6679 cpumask_var_t non_isolated_cpus;
6681 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6682 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6687 * There's no userspace yet to cause hotplug operations; hence all the
6688 * cpu masks are stable and all blatant races in the below code cannot
6691 mutex_lock(&sched_domains_mutex);
6692 init_sched_domains(cpu_active_mask);
6693 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6694 if (cpumask_empty(non_isolated_cpus))
6695 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6696 mutex_unlock(&sched_domains_mutex);
6698 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6699 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6700 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6704 /* Move init over to a non-isolated CPU */
6705 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6707 sched_init_granularity();
6708 free_cpumask_var(non_isolated_cpus);
6710 init_sched_rt_class();
6711 init_sched_dl_class();
6714 void __init sched_init_smp(void)
6716 sched_init_granularity();
6718 #endif /* CONFIG_SMP */
6720 const_debug unsigned int sysctl_timer_migration = 1;
6722 int in_sched_functions(unsigned long addr)
6724 return in_lock_functions(addr) ||
6725 (addr >= (unsigned long)__sched_text_start
6726 && addr < (unsigned long)__sched_text_end);
6729 #ifdef CONFIG_CGROUP_SCHED
6731 * Default task group.
6732 * Every task in system belongs to this group at bootup.
6734 struct task_group root_task_group;
6735 LIST_HEAD(task_groups);
6738 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6740 void __init sched_init(void)
6743 unsigned long alloc_size = 0, ptr;
6745 #ifdef CONFIG_FAIR_GROUP_SCHED
6746 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6748 #ifdef CONFIG_RT_GROUP_SCHED
6749 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6751 #ifdef CONFIG_CPUMASK_OFFSTACK
6752 alloc_size += num_possible_cpus() * cpumask_size();
6755 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6757 #ifdef CONFIG_FAIR_GROUP_SCHED
6758 root_task_group.se = (struct sched_entity **)ptr;
6759 ptr += nr_cpu_ids * sizeof(void **);
6761 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6762 ptr += nr_cpu_ids * sizeof(void **);
6764 #endif /* CONFIG_FAIR_GROUP_SCHED */
6765 #ifdef CONFIG_RT_GROUP_SCHED
6766 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6767 ptr += nr_cpu_ids * sizeof(void **);
6769 root_task_group.rt_rq = (struct rt_rq **)ptr;
6770 ptr += nr_cpu_ids * sizeof(void **);
6772 #endif /* CONFIG_RT_GROUP_SCHED */
6773 #ifdef CONFIG_CPUMASK_OFFSTACK
6774 for_each_possible_cpu(i) {
6775 per_cpu(load_balance_mask, i) = (void *)ptr;
6776 ptr += cpumask_size();
6778 #endif /* CONFIG_CPUMASK_OFFSTACK */
6781 init_rt_bandwidth(&def_rt_bandwidth,
6782 global_rt_period(), global_rt_runtime());
6783 init_dl_bandwidth(&def_dl_bandwidth,
6784 global_dl_period(), global_dl_runtime());
6787 init_defrootdomain();
6790 #ifdef CONFIG_RT_GROUP_SCHED
6791 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6792 global_rt_period(), global_rt_runtime());
6793 #endif /* CONFIG_RT_GROUP_SCHED */
6795 #ifdef CONFIG_CGROUP_SCHED
6796 list_add(&root_task_group.list, &task_groups);
6797 INIT_LIST_HEAD(&root_task_group.children);
6798 INIT_LIST_HEAD(&root_task_group.siblings);
6799 autogroup_init(&init_task);
6801 #endif /* CONFIG_CGROUP_SCHED */
6803 for_each_possible_cpu(i) {
6807 raw_spin_lock_init(&rq->lock);
6809 rq->calc_load_active = 0;
6810 rq->calc_load_update = jiffies + LOAD_FREQ;
6811 init_cfs_rq(&rq->cfs);
6812 init_rt_rq(&rq->rt, rq);
6813 init_dl_rq(&rq->dl, rq);
6814 #ifdef CONFIG_FAIR_GROUP_SCHED
6815 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6816 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6818 * How much cpu bandwidth does root_task_group get?
6820 * In case of task-groups formed thr' the cgroup filesystem, it
6821 * gets 100% of the cpu resources in the system. This overall
6822 * system cpu resource is divided among the tasks of
6823 * root_task_group and its child task-groups in a fair manner,
6824 * based on each entity's (task or task-group's) weight
6825 * (se->load.weight).
6827 * In other words, if root_task_group has 10 tasks of weight
6828 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6829 * then A0's share of the cpu resource is:
6831 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6833 * We achieve this by letting root_task_group's tasks sit
6834 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6836 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6837 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6838 #endif /* CONFIG_FAIR_GROUP_SCHED */
6840 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6841 #ifdef CONFIG_RT_GROUP_SCHED
6842 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6843 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6846 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6847 rq->cpu_load[j] = 0;
6849 rq->last_load_update_tick = jiffies;
6854 rq->cpu_power = SCHED_POWER_SCALE;
6855 rq->post_schedule = 0;
6856 rq->active_balance = 0;
6857 rq->next_balance = jiffies;
6862 rq->avg_idle = 2*sysctl_sched_migration_cost;
6863 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6865 INIT_LIST_HEAD(&rq->cfs_tasks);
6867 rq_attach_root(rq, &def_root_domain);
6868 #ifdef CONFIG_NO_HZ_COMMON
6871 #ifdef CONFIG_NO_HZ_FULL
6872 rq->last_sched_tick = 0;
6876 atomic_set(&rq->nr_iowait, 0);
6879 set_load_weight(&init_task);
6881 #ifdef CONFIG_PREEMPT_NOTIFIERS
6882 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6886 * The boot idle thread does lazy MMU switching as well:
6888 atomic_inc(&init_mm.mm_count);
6889 enter_lazy_tlb(&init_mm, current);
6892 * Make us the idle thread. Technically, schedule() should not be
6893 * called from this thread, however somewhere below it might be,
6894 * but because we are the idle thread, we just pick up running again
6895 * when this runqueue becomes "idle".
6897 init_idle(current, smp_processor_id());
6899 calc_load_update = jiffies + LOAD_FREQ;
6902 * During early bootup we pretend to be a normal task:
6904 current->sched_class = &fair_sched_class;
6907 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6908 /* May be allocated at isolcpus cmdline parse time */
6909 if (cpu_isolated_map == NULL)
6910 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6911 idle_thread_set_boot_cpu();
6913 init_sched_fair_class();
6915 scheduler_running = 1;
6918 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6919 static inline int preempt_count_equals(int preempt_offset)
6921 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6923 return (nested == preempt_offset);
6926 void __might_sleep(const char *file, int line, int preempt_offset)
6928 static unsigned long prev_jiffy; /* ratelimiting */
6930 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6931 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6932 system_state != SYSTEM_RUNNING || oops_in_progress)
6934 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6936 prev_jiffy = jiffies;
6939 "BUG: sleeping function called from invalid context at %s:%d\n",
6942 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6943 in_atomic(), irqs_disabled(),
6944 current->pid, current->comm);
6946 debug_show_held_locks(current);
6947 if (irqs_disabled())
6948 print_irqtrace_events(current);
6951 EXPORT_SYMBOL(__might_sleep);
6954 #ifdef CONFIG_MAGIC_SYSRQ
6955 static void normalize_task(struct rq *rq, struct task_struct *p)
6957 const struct sched_class *prev_class = p->sched_class;
6958 struct sched_attr attr = {
6959 .sched_policy = SCHED_NORMAL,
6961 int old_prio = p->prio;
6966 dequeue_task(rq, p, 0);
6967 __setscheduler(rq, p, &attr);
6969 enqueue_task(rq, p, 0);
6970 resched_task(rq->curr);
6973 check_class_changed(rq, p, prev_class, old_prio);
6976 void normalize_rt_tasks(void)
6978 struct task_struct *g, *p;
6979 unsigned long flags;
6982 read_lock_irqsave(&tasklist_lock, flags);
6983 do_each_thread(g, p) {
6985 * Only normalize user tasks:
6990 p->se.exec_start = 0;
6991 #ifdef CONFIG_SCHEDSTATS
6992 p->se.statistics.wait_start = 0;
6993 p->se.statistics.sleep_start = 0;
6994 p->se.statistics.block_start = 0;
6997 if (!dl_task(p) && !rt_task(p)) {
6999 * Renice negative nice level userspace
7002 if (TASK_NICE(p) < 0 && p->mm)
7003 set_user_nice(p, 0);
7007 raw_spin_lock(&p->pi_lock);
7008 rq = __task_rq_lock(p);
7010 normalize_task(rq, p);
7012 __task_rq_unlock(rq);
7013 raw_spin_unlock(&p->pi_lock);
7014 } while_each_thread(g, p);
7016 read_unlock_irqrestore(&tasklist_lock, flags);
7019 #endif /* CONFIG_MAGIC_SYSRQ */
7021 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7023 * These functions are only useful for the IA64 MCA handling, or kdb.
7025 * They can only be called when the whole system has been
7026 * stopped - every CPU needs to be quiescent, and no scheduling
7027 * activity can take place. Using them for anything else would
7028 * be a serious bug, and as a result, they aren't even visible
7029 * under any other configuration.
7033 * curr_task - return the current task for a given cpu.
7034 * @cpu: the processor in question.
7036 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7038 * Return: The current task for @cpu.
7040 struct task_struct *curr_task(int cpu)
7042 return cpu_curr(cpu);
7045 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7049 * set_curr_task - set the current task for a given cpu.
7050 * @cpu: the processor in question.
7051 * @p: the task pointer to set.
7053 * Description: This function must only be used when non-maskable interrupts
7054 * are serviced on a separate stack. It allows the architecture to switch the
7055 * notion of the current task on a cpu in a non-blocking manner. This function
7056 * must be called with all CPU's synchronized, and interrupts disabled, the
7057 * and caller must save the original value of the current task (see
7058 * curr_task() above) and restore that value before reenabling interrupts and
7059 * re-starting the system.
7061 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7063 void set_curr_task(int cpu, struct task_struct *p)
7070 #ifdef CONFIG_CGROUP_SCHED
7071 /* task_group_lock serializes the addition/removal of task groups */
7072 static DEFINE_SPINLOCK(task_group_lock);
7074 static void free_sched_group(struct task_group *tg)
7076 free_fair_sched_group(tg);
7077 free_rt_sched_group(tg);
7082 /* allocate runqueue etc for a new task group */
7083 struct task_group *sched_create_group(struct task_group *parent)
7085 struct task_group *tg;
7087 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7089 return ERR_PTR(-ENOMEM);
7091 if (!alloc_fair_sched_group(tg, parent))
7094 if (!alloc_rt_sched_group(tg, parent))
7100 free_sched_group(tg);
7101 return ERR_PTR(-ENOMEM);
7104 void sched_online_group(struct task_group *tg, struct task_group *parent)
7106 unsigned long flags;
7108 spin_lock_irqsave(&task_group_lock, flags);
7109 list_add_rcu(&tg->list, &task_groups);
7111 WARN_ON(!parent); /* root should already exist */
7113 tg->parent = parent;
7114 INIT_LIST_HEAD(&tg->children);
7115 list_add_rcu(&tg->siblings, &parent->children);
7116 spin_unlock_irqrestore(&task_group_lock, flags);
7119 /* rcu callback to free various structures associated with a task group */
7120 static void free_sched_group_rcu(struct rcu_head *rhp)
7122 /* now it should be safe to free those cfs_rqs */
7123 free_sched_group(container_of(rhp, struct task_group, rcu));
7126 /* Destroy runqueue etc associated with a task group */
7127 void sched_destroy_group(struct task_group *tg)
7129 /* wait for possible concurrent references to cfs_rqs complete */
7130 call_rcu(&tg->rcu, free_sched_group_rcu);
7133 void sched_offline_group(struct task_group *tg)
7135 unsigned long flags;
7138 /* end participation in shares distribution */
7139 for_each_possible_cpu(i)
7140 unregister_fair_sched_group(tg, i);
7142 spin_lock_irqsave(&task_group_lock, flags);
7143 list_del_rcu(&tg->list);
7144 list_del_rcu(&tg->siblings);
7145 spin_unlock_irqrestore(&task_group_lock, flags);
7148 /* change task's runqueue when it moves between groups.
7149 * The caller of this function should have put the task in its new group
7150 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7151 * reflect its new group.
7153 void sched_move_task(struct task_struct *tsk)
7155 struct task_group *tg;
7157 unsigned long flags;
7160 rq = task_rq_lock(tsk, &flags);
7162 running = task_current(rq, tsk);
7166 dequeue_task(rq, tsk, 0);
7167 if (unlikely(running))
7168 tsk->sched_class->put_prev_task(rq, tsk);
7170 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
7171 lockdep_is_held(&tsk->sighand->siglock)),
7172 struct task_group, css);
7173 tg = autogroup_task_group(tsk, tg);
7174 tsk->sched_task_group = tg;
7176 #ifdef CONFIG_FAIR_GROUP_SCHED
7177 if (tsk->sched_class->task_move_group)
7178 tsk->sched_class->task_move_group(tsk, on_rq);
7181 set_task_rq(tsk, task_cpu(tsk));
7183 if (unlikely(running))
7184 tsk->sched_class->set_curr_task(rq);
7186 enqueue_task(rq, tsk, 0);
7188 task_rq_unlock(rq, tsk, &flags);
7190 #endif /* CONFIG_CGROUP_SCHED */
7192 #ifdef CONFIG_RT_GROUP_SCHED
7194 * Ensure that the real time constraints are schedulable.
7196 static DEFINE_MUTEX(rt_constraints_mutex);
7198 /* Must be called with tasklist_lock held */
7199 static inline int tg_has_rt_tasks(struct task_group *tg)
7201 struct task_struct *g, *p;
7203 do_each_thread(g, p) {
7204 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7206 } while_each_thread(g, p);
7211 struct rt_schedulable_data {
7212 struct task_group *tg;
7217 static int tg_rt_schedulable(struct task_group *tg, void *data)
7219 struct rt_schedulable_data *d = data;
7220 struct task_group *child;
7221 unsigned long total, sum = 0;
7222 u64 period, runtime;
7224 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7225 runtime = tg->rt_bandwidth.rt_runtime;
7228 period = d->rt_period;
7229 runtime = d->rt_runtime;
7233 * Cannot have more runtime than the period.
7235 if (runtime > period && runtime != RUNTIME_INF)
7239 * Ensure we don't starve existing RT tasks.
7241 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7244 total = to_ratio(period, runtime);
7247 * Nobody can have more than the global setting allows.
7249 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7253 * The sum of our children's runtime should not exceed our own.
7255 list_for_each_entry_rcu(child, &tg->children, siblings) {
7256 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7257 runtime = child->rt_bandwidth.rt_runtime;
7259 if (child == d->tg) {
7260 period = d->rt_period;
7261 runtime = d->rt_runtime;
7264 sum += to_ratio(period, runtime);
7273 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7277 struct rt_schedulable_data data = {
7279 .rt_period = period,
7280 .rt_runtime = runtime,
7284 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7290 static int tg_set_rt_bandwidth(struct task_group *tg,
7291 u64 rt_period, u64 rt_runtime)
7295 mutex_lock(&rt_constraints_mutex);
7296 read_lock(&tasklist_lock);
7297 err = __rt_schedulable(tg, rt_period, rt_runtime);
7301 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7302 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7303 tg->rt_bandwidth.rt_runtime = rt_runtime;
7305 for_each_possible_cpu(i) {
7306 struct rt_rq *rt_rq = tg->rt_rq[i];
7308 raw_spin_lock(&rt_rq->rt_runtime_lock);
7309 rt_rq->rt_runtime = rt_runtime;
7310 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7312 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7314 read_unlock(&tasklist_lock);
7315 mutex_unlock(&rt_constraints_mutex);
7320 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7322 u64 rt_runtime, rt_period;
7324 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7325 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7326 if (rt_runtime_us < 0)
7327 rt_runtime = RUNTIME_INF;
7329 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7332 static long sched_group_rt_runtime(struct task_group *tg)
7336 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7339 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7340 do_div(rt_runtime_us, NSEC_PER_USEC);
7341 return rt_runtime_us;
7344 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7346 u64 rt_runtime, rt_period;
7348 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7349 rt_runtime = tg->rt_bandwidth.rt_runtime;
7354 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7357 static long sched_group_rt_period(struct task_group *tg)
7361 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7362 do_div(rt_period_us, NSEC_PER_USEC);
7363 return rt_period_us;
7365 #endif /* CONFIG_RT_GROUP_SCHED */
7368 * Coupling of -rt and -deadline bandwidth.
7370 * Here we check if the new -rt bandwidth value is consistent
7371 * with the system settings for the bandwidth available
7372 * to -deadline tasks.
7374 * IOW, we want to enforce that
7376 * rt_bandwidth + dl_bandwidth <= 100%
7380 static bool __sched_rt_dl_global_constraints(u64 rt_bw)
7382 unsigned long flags;
7386 raw_spin_lock_irqsave(&def_dl_bandwidth.dl_runtime_lock, flags);
7387 if (global_rt_runtime() == RUNTIME_INF ||
7388 global_dl_runtime() == RUNTIME_INF) {
7393 dl_bw = to_ratio(def_dl_bandwidth.dl_period,
7394 def_dl_bandwidth.dl_runtime);
7396 ret = rt_bw + dl_bw <= to_ratio(RUNTIME_INF, RUNTIME_INF);
7398 raw_spin_unlock_irqrestore(&def_dl_bandwidth.dl_runtime_lock, flags);
7403 #ifdef CONFIG_RT_GROUP_SCHED
7404 static int sched_rt_global_constraints(void)
7406 u64 runtime, period, bw;
7409 if (sysctl_sched_rt_period <= 0)
7412 runtime = global_rt_runtime();
7413 period = global_rt_period();
7416 * Sanity check on the sysctl variables.
7418 if (runtime > period && runtime != RUNTIME_INF)
7421 bw = to_ratio(period, runtime);
7422 if (!__sched_rt_dl_global_constraints(bw))
7425 mutex_lock(&rt_constraints_mutex);
7426 read_lock(&tasklist_lock);
7427 ret = __rt_schedulable(NULL, 0, 0);
7428 read_unlock(&tasklist_lock);
7429 mutex_unlock(&rt_constraints_mutex);
7434 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7436 /* Don't accept realtime tasks when there is no way for them to run */
7437 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7443 #else /* !CONFIG_RT_GROUP_SCHED */
7444 static int sched_rt_global_constraints(void)
7446 unsigned long flags;
7450 if (sysctl_sched_rt_period <= 0)
7453 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7454 bw = to_ratio(global_rt_period(), global_rt_runtime());
7455 if (!__sched_rt_dl_global_constraints(bw)) {
7460 for_each_possible_cpu(i) {
7461 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7463 raw_spin_lock(&rt_rq->rt_runtime_lock);
7464 rt_rq->rt_runtime = global_rt_runtime();
7465 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7468 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7472 #endif /* CONFIG_RT_GROUP_SCHED */
7475 * Coupling of -dl and -rt bandwidth.
7477 * Here we check, while setting the system wide bandwidth available
7478 * for -dl tasks and groups, if the new values are consistent with
7479 * the system settings for the bandwidth available to -rt entities.
7481 * IOW, we want to enforce that
7483 * rt_bandwidth + dl_bandwidth <= 100%
7487 static bool __sched_dl_rt_global_constraints(u64 dl_bw)
7492 raw_spin_lock(&def_rt_bandwidth.rt_runtime_lock);
7493 if (global_dl_runtime() == RUNTIME_INF ||
7494 global_rt_runtime() == RUNTIME_INF) {
7499 rt_bw = to_ratio(ktime_to_ns(def_rt_bandwidth.rt_period),
7500 def_rt_bandwidth.rt_runtime);
7502 ret = rt_bw + dl_bw <= to_ratio(RUNTIME_INF, RUNTIME_INF);
7504 raw_spin_unlock(&def_rt_bandwidth.rt_runtime_lock);
7509 static bool __sched_dl_global_constraints(u64 runtime, u64 period)
7511 if (!period || (runtime != RUNTIME_INF && runtime > period))
7517 static int sched_dl_global_constraints(void)
7519 u64 runtime = global_dl_runtime();
7520 u64 period = global_dl_period();
7521 u64 new_bw = to_ratio(period, runtime);
7524 ret = __sched_dl_global_constraints(runtime, period);
7528 if (!__sched_dl_rt_global_constraints(new_bw))
7532 * Here we want to check the bandwidth not being set to some
7533 * value smaller than the currently allocated bandwidth in
7534 * any of the root_domains.
7536 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7537 * cycling on root_domains... Discussion on different/better
7538 * solutions is welcome!
7540 for_each_possible_cpu(i) {
7541 struct dl_bw *dl_b = dl_bw_of(i);
7543 raw_spin_lock(&dl_b->lock);
7544 if (new_bw < dl_b->total_bw) {
7545 raw_spin_unlock(&dl_b->lock);
7548 raw_spin_unlock(&dl_b->lock);
7554 int sched_rr_handler(struct ctl_table *table, int write,
7555 void __user *buffer, size_t *lenp,
7559 static DEFINE_MUTEX(mutex);
7562 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7563 /* make sure that internally we keep jiffies */
7564 /* also, writing zero resets timeslice to default */
7565 if (!ret && write) {
7566 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7567 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7569 mutex_unlock(&mutex);
7573 int sched_rt_handler(struct ctl_table *table, int write,
7574 void __user *buffer, size_t *lenp,
7578 int old_period, old_runtime;
7579 static DEFINE_MUTEX(mutex);
7582 old_period = sysctl_sched_rt_period;
7583 old_runtime = sysctl_sched_rt_runtime;
7585 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7587 if (!ret && write) {
7588 ret = sched_rt_global_constraints();
7590 sysctl_sched_rt_period = old_period;
7591 sysctl_sched_rt_runtime = old_runtime;
7593 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7594 def_rt_bandwidth.rt_period =
7595 ns_to_ktime(global_rt_period());
7598 mutex_unlock(&mutex);
7603 int sched_dl_handler(struct ctl_table *table, int write,
7604 void __user *buffer, size_t *lenp,
7608 int old_period, old_runtime;
7609 static DEFINE_MUTEX(mutex);
7610 unsigned long flags;
7613 old_period = sysctl_sched_dl_period;
7614 old_runtime = sysctl_sched_dl_runtime;
7616 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7618 if (!ret && write) {
7619 raw_spin_lock_irqsave(&def_dl_bandwidth.dl_runtime_lock,
7622 ret = sched_dl_global_constraints();
7624 sysctl_sched_dl_period = old_period;
7625 sysctl_sched_dl_runtime = old_runtime;
7630 def_dl_bandwidth.dl_period = global_dl_period();
7631 def_dl_bandwidth.dl_runtime = global_dl_runtime();
7632 if (global_dl_runtime() == RUNTIME_INF)
7635 new_bw = to_ratio(global_dl_period(),
7636 global_dl_runtime());
7638 * FIXME: As above...
7640 for_each_possible_cpu(i) {
7641 struct dl_bw *dl_b = dl_bw_of(i);
7643 raw_spin_lock(&dl_b->lock);
7645 raw_spin_unlock(&dl_b->lock);
7649 raw_spin_unlock_irqrestore(&def_dl_bandwidth.dl_runtime_lock,
7652 mutex_unlock(&mutex);
7657 #ifdef CONFIG_CGROUP_SCHED
7659 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7661 return css ? container_of(css, struct task_group, css) : NULL;
7664 static struct cgroup_subsys_state *
7665 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7667 struct task_group *parent = css_tg(parent_css);
7668 struct task_group *tg;
7671 /* This is early initialization for the top cgroup */
7672 return &root_task_group.css;
7675 tg = sched_create_group(parent);
7677 return ERR_PTR(-ENOMEM);
7682 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7684 struct task_group *tg = css_tg(css);
7685 struct task_group *parent = css_tg(css_parent(css));
7688 sched_online_group(tg, parent);
7692 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7694 struct task_group *tg = css_tg(css);
7696 sched_destroy_group(tg);
7699 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7701 struct task_group *tg = css_tg(css);
7703 sched_offline_group(tg);
7706 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7707 struct cgroup_taskset *tset)
7709 struct task_struct *task;
7711 cgroup_taskset_for_each(task, css, tset) {
7712 #ifdef CONFIG_RT_GROUP_SCHED
7713 if (!sched_rt_can_attach(css_tg(css), task))
7716 /* We don't support RT-tasks being in separate groups */
7717 if (task->sched_class != &fair_sched_class)
7724 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7725 struct cgroup_taskset *tset)
7727 struct task_struct *task;
7729 cgroup_taskset_for_each(task, css, tset)
7730 sched_move_task(task);
7733 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7734 struct cgroup_subsys_state *old_css,
7735 struct task_struct *task)
7738 * cgroup_exit() is called in the copy_process() failure path.
7739 * Ignore this case since the task hasn't ran yet, this avoids
7740 * trying to poke a half freed task state from generic code.
7742 if (!(task->flags & PF_EXITING))
7745 sched_move_task(task);
7748 #ifdef CONFIG_FAIR_GROUP_SCHED
7749 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7750 struct cftype *cftype, u64 shareval)
7752 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7755 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7758 struct task_group *tg = css_tg(css);
7760 return (u64) scale_load_down(tg->shares);
7763 #ifdef CONFIG_CFS_BANDWIDTH
7764 static DEFINE_MUTEX(cfs_constraints_mutex);
7766 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7767 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7769 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7771 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7773 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7774 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7776 if (tg == &root_task_group)
7780 * Ensure we have at some amount of bandwidth every period. This is
7781 * to prevent reaching a state of large arrears when throttled via
7782 * entity_tick() resulting in prolonged exit starvation.
7784 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7788 * Likewise, bound things on the otherside by preventing insane quota
7789 * periods. This also allows us to normalize in computing quota
7792 if (period > max_cfs_quota_period)
7795 mutex_lock(&cfs_constraints_mutex);
7796 ret = __cfs_schedulable(tg, period, quota);
7800 runtime_enabled = quota != RUNTIME_INF;
7801 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7803 * If we need to toggle cfs_bandwidth_used, off->on must occur
7804 * before making related changes, and on->off must occur afterwards
7806 if (runtime_enabled && !runtime_was_enabled)
7807 cfs_bandwidth_usage_inc();
7808 raw_spin_lock_irq(&cfs_b->lock);
7809 cfs_b->period = ns_to_ktime(period);
7810 cfs_b->quota = quota;
7812 __refill_cfs_bandwidth_runtime(cfs_b);
7813 /* restart the period timer (if active) to handle new period expiry */
7814 if (runtime_enabled && cfs_b->timer_active) {
7815 /* force a reprogram */
7816 cfs_b->timer_active = 0;
7817 __start_cfs_bandwidth(cfs_b);
7819 raw_spin_unlock_irq(&cfs_b->lock);
7821 for_each_possible_cpu(i) {
7822 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7823 struct rq *rq = cfs_rq->rq;
7825 raw_spin_lock_irq(&rq->lock);
7826 cfs_rq->runtime_enabled = runtime_enabled;
7827 cfs_rq->runtime_remaining = 0;
7829 if (cfs_rq->throttled)
7830 unthrottle_cfs_rq(cfs_rq);
7831 raw_spin_unlock_irq(&rq->lock);
7833 if (runtime_was_enabled && !runtime_enabled)
7834 cfs_bandwidth_usage_dec();
7836 mutex_unlock(&cfs_constraints_mutex);
7841 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7845 period = ktime_to_ns(tg->cfs_bandwidth.period);
7846 if (cfs_quota_us < 0)
7847 quota = RUNTIME_INF;
7849 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7851 return tg_set_cfs_bandwidth(tg, period, quota);
7854 long tg_get_cfs_quota(struct task_group *tg)
7858 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7861 quota_us = tg->cfs_bandwidth.quota;
7862 do_div(quota_us, NSEC_PER_USEC);
7867 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7871 period = (u64)cfs_period_us * NSEC_PER_USEC;
7872 quota = tg->cfs_bandwidth.quota;
7874 return tg_set_cfs_bandwidth(tg, period, quota);
7877 long tg_get_cfs_period(struct task_group *tg)
7881 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7882 do_div(cfs_period_us, NSEC_PER_USEC);
7884 return cfs_period_us;
7887 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7890 return tg_get_cfs_quota(css_tg(css));
7893 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7894 struct cftype *cftype, s64 cfs_quota_us)
7896 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7899 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7902 return tg_get_cfs_period(css_tg(css));
7905 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7906 struct cftype *cftype, u64 cfs_period_us)
7908 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7911 struct cfs_schedulable_data {
7912 struct task_group *tg;
7917 * normalize group quota/period to be quota/max_period
7918 * note: units are usecs
7920 static u64 normalize_cfs_quota(struct task_group *tg,
7921 struct cfs_schedulable_data *d)
7929 period = tg_get_cfs_period(tg);
7930 quota = tg_get_cfs_quota(tg);
7933 /* note: these should typically be equivalent */
7934 if (quota == RUNTIME_INF || quota == -1)
7937 return to_ratio(period, quota);
7940 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7942 struct cfs_schedulable_data *d = data;
7943 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7944 s64 quota = 0, parent_quota = -1;
7947 quota = RUNTIME_INF;
7949 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7951 quota = normalize_cfs_quota(tg, d);
7952 parent_quota = parent_b->hierarchal_quota;
7955 * ensure max(child_quota) <= parent_quota, inherit when no
7958 if (quota == RUNTIME_INF)
7959 quota = parent_quota;
7960 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7963 cfs_b->hierarchal_quota = quota;
7968 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7971 struct cfs_schedulable_data data = {
7977 if (quota != RUNTIME_INF) {
7978 do_div(data.period, NSEC_PER_USEC);
7979 do_div(data.quota, NSEC_PER_USEC);
7983 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7989 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7990 struct cgroup_map_cb *cb)
7992 struct task_group *tg = css_tg(css);
7993 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7995 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7996 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7997 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
8001 #endif /* CONFIG_CFS_BANDWIDTH */
8002 #endif /* CONFIG_FAIR_GROUP_SCHED */
8004 #ifdef CONFIG_RT_GROUP_SCHED
8005 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8006 struct cftype *cft, s64 val)
8008 return sched_group_set_rt_runtime(css_tg(css), val);
8011 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8014 return sched_group_rt_runtime(css_tg(css));
8017 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8018 struct cftype *cftype, u64 rt_period_us)
8020 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8023 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8026 return sched_group_rt_period(css_tg(css));
8028 #endif /* CONFIG_RT_GROUP_SCHED */
8030 static struct cftype cpu_files[] = {
8031 #ifdef CONFIG_FAIR_GROUP_SCHED
8034 .read_u64 = cpu_shares_read_u64,
8035 .write_u64 = cpu_shares_write_u64,
8038 #ifdef CONFIG_CFS_BANDWIDTH
8040 .name = "cfs_quota_us",
8041 .read_s64 = cpu_cfs_quota_read_s64,
8042 .write_s64 = cpu_cfs_quota_write_s64,
8045 .name = "cfs_period_us",
8046 .read_u64 = cpu_cfs_period_read_u64,
8047 .write_u64 = cpu_cfs_period_write_u64,
8051 .read_map = cpu_stats_show,
8054 #ifdef CONFIG_RT_GROUP_SCHED
8056 .name = "rt_runtime_us",
8057 .read_s64 = cpu_rt_runtime_read,
8058 .write_s64 = cpu_rt_runtime_write,
8061 .name = "rt_period_us",
8062 .read_u64 = cpu_rt_period_read_uint,
8063 .write_u64 = cpu_rt_period_write_uint,
8069 struct cgroup_subsys cpu_cgroup_subsys = {
8071 .css_alloc = cpu_cgroup_css_alloc,
8072 .css_free = cpu_cgroup_css_free,
8073 .css_online = cpu_cgroup_css_online,
8074 .css_offline = cpu_cgroup_css_offline,
8075 .can_attach = cpu_cgroup_can_attach,
8076 .attach = cpu_cgroup_attach,
8077 .exit = cpu_cgroup_exit,
8078 .subsys_id = cpu_cgroup_subsys_id,
8079 .base_cftypes = cpu_files,
8083 #endif /* CONFIG_CGROUP_SCHED */
8085 void dump_cpu_task(int cpu)
8087 pr_info("Task dump for CPU %d:\n", cpu);
8088 sched_show_task(cpu_curr(cpu));