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>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
96 ktime_t soft, hard, now;
99 if (hrtimer_active(period_timer))
102 now = hrtimer_cb_get_time(period_timer);
103 hrtimer_forward(period_timer, now, period);
105 soft = hrtimer_get_softexpires(period_timer);
106 hard = hrtimer_get_expires(period_timer);
107 delta = ktime_to_ns(ktime_sub(hard, soft));
108 __hrtimer_start_range_ns(period_timer, soft, delta,
109 HRTIMER_MODE_ABS_PINNED, 0);
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
118 void update_rq_clock(struct rq *rq)
122 lockdep_assert_held(&rq->lock);
124 if (rq->clock_skip_update & RQCF_ACT_SKIP)
127 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
131 update_rq_clock_task(rq, delta);
135 * Debugging: various feature bits
138 #define SCHED_FEAT(name, enabled) \
139 (1UL << __SCHED_FEAT_##name) * enabled |
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled) \
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
157 static int sched_feat_show(struct seq_file *m, void *v)
161 for (i = 0; i < __SCHED_FEAT_NR; i++) {
162 if (!(sysctl_sched_features & (1UL << i)))
164 seq_printf(m, "%s ", sched_feat_names[i]);
171 #ifdef HAVE_JUMP_LABEL
173 #define jump_label_key__true STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
176 #define SCHED_FEAT(name, enabled) \
177 jump_label_key__##enabled ,
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
185 static void sched_feat_disable(int i)
187 if (static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_dec(&sched_feat_keys[i]);
191 static void sched_feat_enable(int i)
193 if (!static_key_enabled(&sched_feat_keys[i]))
194 static_key_slow_inc(&sched_feat_keys[i]);
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
201 static int sched_feat_set(char *cmp)
206 if (strncmp(cmp, "NO_", 3) == 0) {
211 for (i = 0; i < __SCHED_FEAT_NR; i++) {
212 if (strcmp(cmp, sched_feat_names[i]) == 0) {
214 sysctl_sched_features &= ~(1UL << i);
215 sched_feat_disable(i);
217 sysctl_sched_features |= (1UL << i);
218 sched_feat_enable(i);
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229 size_t cnt, loff_t *ppos)
239 if (copy_from_user(&buf, ubuf, cnt))
245 /* Ensure the static_key remains in a consistent state */
246 inode = file_inode(filp);
247 mutex_lock(&inode->i_mutex);
248 i = sched_feat_set(cmp);
249 mutex_unlock(&inode->i_mutex);
250 if (i == __SCHED_FEAT_NR)
258 static int sched_feat_open(struct inode *inode, struct file *filp)
260 return single_open(filp, sched_feat_show, NULL);
263 static const struct file_operations sched_feat_fops = {
264 .open = sched_feat_open,
265 .write = sched_feat_write,
268 .release = single_release,
271 static __init int sched_init_debug(void)
273 debugfs_create_file("sched_features", 0644, NULL, NULL,
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
282 * Number of tasks to iterate in a single balance run.
283 * Limited because this is done with IRQs disabled.
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
288 * period over which we average the RT time consumption, measured
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
296 * period over which we measure -rt task cpu usage in us.
299 unsigned int sysctl_sched_rt_period = 1000000;
301 __read_mostly int scheduler_running;
304 * part of the period that we allow rt tasks to run in us.
307 int sysctl_sched_rt_runtime = 950000;
310 * __task_rq_lock - lock the rq @p resides on.
312 static inline struct rq *__task_rq_lock(struct task_struct *p)
317 lockdep_assert_held(&p->pi_lock);
321 raw_spin_lock(&rq->lock);
322 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
324 raw_spin_unlock(&rq->lock);
326 while (unlikely(task_on_rq_migrating(p)))
332 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
334 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
335 __acquires(p->pi_lock)
341 raw_spin_lock_irqsave(&p->pi_lock, *flags);
343 raw_spin_lock(&rq->lock);
344 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
346 raw_spin_unlock(&rq->lock);
347 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
349 while (unlikely(task_on_rq_migrating(p)))
354 static void __task_rq_unlock(struct rq *rq)
357 raw_spin_unlock(&rq->lock);
361 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
363 __releases(p->pi_lock)
365 raw_spin_unlock(&rq->lock);
366 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
370 * this_rq_lock - lock this runqueue and disable interrupts.
372 static struct rq *this_rq_lock(void)
379 raw_spin_lock(&rq->lock);
384 #ifdef CONFIG_SCHED_HRTICK
386 * Use HR-timers to deliver accurate preemption points.
389 static void hrtick_clear(struct rq *rq)
391 if (hrtimer_active(&rq->hrtick_timer))
392 hrtimer_cancel(&rq->hrtick_timer);
396 * High-resolution timer tick.
397 * Runs from hardirq context with interrupts disabled.
399 static enum hrtimer_restart hrtick(struct hrtimer *timer)
401 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
403 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
405 raw_spin_lock(&rq->lock);
407 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
408 raw_spin_unlock(&rq->lock);
410 return HRTIMER_NORESTART;
415 static int __hrtick_restart(struct rq *rq)
417 struct hrtimer *timer = &rq->hrtick_timer;
418 ktime_t time = hrtimer_get_softexpires(timer);
420 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
424 * called from hardirq (IPI) context
426 static void __hrtick_start(void *arg)
430 raw_spin_lock(&rq->lock);
431 __hrtick_restart(rq);
432 rq->hrtick_csd_pending = 0;
433 raw_spin_unlock(&rq->lock);
437 * Called to set the hrtick timer state.
439 * called with rq->lock held and irqs disabled
441 void hrtick_start(struct rq *rq, u64 delay)
443 struct hrtimer *timer = &rq->hrtick_timer;
448 * Don't schedule slices shorter than 10000ns, that just
449 * doesn't make sense and can cause timer DoS.
451 delta = max_t(s64, delay, 10000LL);
452 time = ktime_add_ns(timer->base->get_time(), delta);
454 hrtimer_set_expires(timer, time);
456 if (rq == this_rq()) {
457 __hrtick_restart(rq);
458 } else if (!rq->hrtick_csd_pending) {
459 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
460 rq->hrtick_csd_pending = 1;
465 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
467 int cpu = (int)(long)hcpu;
470 case CPU_UP_CANCELED:
471 case CPU_UP_CANCELED_FROZEN:
472 case CPU_DOWN_PREPARE:
473 case CPU_DOWN_PREPARE_FROZEN:
475 case CPU_DEAD_FROZEN:
476 hrtick_clear(cpu_rq(cpu));
483 static __init void init_hrtick(void)
485 hotcpu_notifier(hotplug_hrtick, 0);
489 * Called to set the hrtick timer state.
491 * called with rq->lock held and irqs disabled
493 void hrtick_start(struct rq *rq, u64 delay)
495 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
496 HRTIMER_MODE_REL_PINNED, 0);
499 static inline void init_hrtick(void)
502 #endif /* CONFIG_SMP */
504 static void init_rq_hrtick(struct rq *rq)
507 rq->hrtick_csd_pending = 0;
509 rq->hrtick_csd.flags = 0;
510 rq->hrtick_csd.func = __hrtick_start;
511 rq->hrtick_csd.info = rq;
514 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
515 rq->hrtick_timer.function = hrtick;
517 #else /* CONFIG_SCHED_HRTICK */
518 static inline void hrtick_clear(struct rq *rq)
522 static inline void init_rq_hrtick(struct rq *rq)
526 static inline void init_hrtick(void)
529 #endif /* CONFIG_SCHED_HRTICK */
532 * cmpxchg based fetch_or, macro so it works for different integer types
534 #define fetch_or(ptr, val) \
535 ({ typeof(*(ptr)) __old, __val = *(ptr); \
537 __old = cmpxchg((ptr), __val, __val | (val)); \
538 if (__old == __val) \
545 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
547 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
548 * this avoids any races wrt polling state changes and thereby avoids
551 static bool set_nr_and_not_polling(struct task_struct *p)
553 struct thread_info *ti = task_thread_info(p);
554 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
558 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
560 * If this returns true, then the idle task promises to call
561 * sched_ttwu_pending() and reschedule soon.
563 static bool set_nr_if_polling(struct task_struct *p)
565 struct thread_info *ti = task_thread_info(p);
566 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
569 if (!(val & _TIF_POLLING_NRFLAG))
571 if (val & _TIF_NEED_RESCHED)
573 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
582 static bool set_nr_and_not_polling(struct task_struct *p)
584 set_tsk_need_resched(p);
589 static bool set_nr_if_polling(struct task_struct *p)
597 * resched_curr - mark rq's current task 'to be rescheduled now'.
599 * On UP this means the setting of the need_resched flag, on SMP it
600 * might also involve a cross-CPU call to trigger the scheduler on
603 void resched_curr(struct rq *rq)
605 struct task_struct *curr = rq->curr;
608 lockdep_assert_held(&rq->lock);
610 if (test_tsk_need_resched(curr))
615 if (cpu == smp_processor_id()) {
616 set_tsk_need_resched(curr);
617 set_preempt_need_resched();
621 if (set_nr_and_not_polling(curr))
622 smp_send_reschedule(cpu);
624 trace_sched_wake_idle_without_ipi(cpu);
627 void resched_cpu(int cpu)
629 struct rq *rq = cpu_rq(cpu);
632 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
635 raw_spin_unlock_irqrestore(&rq->lock, flags);
639 #ifdef CONFIG_NO_HZ_COMMON
641 * In the semi idle case, use the nearest busy cpu for migrating timers
642 * from an idle cpu. This is good for power-savings.
644 * We don't do similar optimization for completely idle system, as
645 * selecting an idle cpu will add more delays to the timers than intended
646 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
648 int get_nohz_timer_target(int pinned)
650 int cpu = smp_processor_id();
652 struct sched_domain *sd;
654 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
658 for_each_domain(cpu, sd) {
659 for_each_cpu(i, sched_domain_span(sd)) {
671 * When add_timer_on() enqueues a timer into the timer wheel of an
672 * idle CPU then this timer might expire before the next timer event
673 * which is scheduled to wake up that CPU. In case of a completely
674 * idle system the next event might even be infinite time into the
675 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
676 * leaves the inner idle loop so the newly added timer is taken into
677 * account when the CPU goes back to idle and evaluates the timer
678 * wheel for the next timer event.
680 static void wake_up_idle_cpu(int cpu)
682 struct rq *rq = cpu_rq(cpu);
684 if (cpu == smp_processor_id())
687 if (set_nr_and_not_polling(rq->idle))
688 smp_send_reschedule(cpu);
690 trace_sched_wake_idle_without_ipi(cpu);
693 static bool wake_up_full_nohz_cpu(int cpu)
696 * We just need the target to call irq_exit() and re-evaluate
697 * the next tick. The nohz full kick at least implies that.
698 * If needed we can still optimize that later with an
701 if (tick_nohz_full_cpu(cpu)) {
702 if (cpu != smp_processor_id() ||
703 tick_nohz_tick_stopped())
704 tick_nohz_full_kick_cpu(cpu);
711 void wake_up_nohz_cpu(int cpu)
713 if (!wake_up_full_nohz_cpu(cpu))
714 wake_up_idle_cpu(cpu);
717 static inline bool got_nohz_idle_kick(void)
719 int cpu = smp_processor_id();
721 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
724 if (idle_cpu(cpu) && !need_resched())
728 * We can't run Idle Load Balance on this CPU for this time so we
729 * cancel it and clear NOHZ_BALANCE_KICK
731 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
735 #else /* CONFIG_NO_HZ_COMMON */
737 static inline bool got_nohz_idle_kick(void)
742 #endif /* CONFIG_NO_HZ_COMMON */
744 #ifdef CONFIG_NO_HZ_FULL
745 bool sched_can_stop_tick(void)
748 * More than one running task need preemption.
749 * nr_running update is assumed to be visible
750 * after IPI is sent from wakers.
752 if (this_rq()->nr_running > 1)
757 #endif /* CONFIG_NO_HZ_FULL */
759 void sched_avg_update(struct rq *rq)
761 s64 period = sched_avg_period();
763 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
765 * Inline assembly required to prevent the compiler
766 * optimising this loop into a divmod call.
767 * See __iter_div_u64_rem() for another example of this.
769 asm("" : "+rm" (rq->age_stamp));
770 rq->age_stamp += period;
775 #endif /* CONFIG_SMP */
777 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
778 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
780 * Iterate task_group tree rooted at *from, calling @down when first entering a
781 * node and @up when leaving it for the final time.
783 * Caller must hold rcu_lock or sufficient equivalent.
785 int walk_tg_tree_from(struct task_group *from,
786 tg_visitor down, tg_visitor up, void *data)
788 struct task_group *parent, *child;
794 ret = (*down)(parent, data);
797 list_for_each_entry_rcu(child, &parent->children, siblings) {
804 ret = (*up)(parent, data);
805 if (ret || parent == from)
809 parent = parent->parent;
816 int tg_nop(struct task_group *tg, void *data)
822 static void set_load_weight(struct task_struct *p)
824 int prio = p->static_prio - MAX_RT_PRIO;
825 struct load_weight *load = &p->se.load;
828 * SCHED_IDLE tasks get minimal weight:
830 if (p->policy == SCHED_IDLE) {
831 load->weight = scale_load(WEIGHT_IDLEPRIO);
832 load->inv_weight = WMULT_IDLEPRIO;
836 load->weight = scale_load(prio_to_weight[prio]);
837 load->inv_weight = prio_to_wmult[prio];
840 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
843 sched_info_queued(rq, p);
844 p->sched_class->enqueue_task(rq, p, flags);
847 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
850 sched_info_dequeued(rq, p);
851 p->sched_class->dequeue_task(rq, p, flags);
854 void activate_task(struct rq *rq, struct task_struct *p, int flags)
856 if (task_contributes_to_load(p))
857 rq->nr_uninterruptible--;
859 enqueue_task(rq, p, flags);
862 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
864 if (task_contributes_to_load(p))
865 rq->nr_uninterruptible++;
867 dequeue_task(rq, p, flags);
870 static void update_rq_clock_task(struct rq *rq, s64 delta)
873 * In theory, the compile should just see 0 here, and optimize out the call
874 * to sched_rt_avg_update. But I don't trust it...
876 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
877 s64 steal = 0, irq_delta = 0;
879 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
880 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
883 * Since irq_time is only updated on {soft,}irq_exit, we might run into
884 * this case when a previous update_rq_clock() happened inside a
887 * When this happens, we stop ->clock_task and only update the
888 * prev_irq_time stamp to account for the part that fit, so that a next
889 * update will consume the rest. This ensures ->clock_task is
892 * It does however cause some slight miss-attribution of {soft,}irq
893 * time, a more accurate solution would be to update the irq_time using
894 * the current rq->clock timestamp, except that would require using
897 if (irq_delta > delta)
900 rq->prev_irq_time += irq_delta;
903 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
904 if (static_key_false((¶virt_steal_rq_enabled))) {
905 steal = paravirt_steal_clock(cpu_of(rq));
906 steal -= rq->prev_steal_time_rq;
908 if (unlikely(steal > delta))
911 rq->prev_steal_time_rq += steal;
916 rq->clock_task += delta;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
920 sched_rt_avg_update(rq, irq_delta + steal);
924 void sched_set_stop_task(int cpu, struct task_struct *stop)
926 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
927 struct task_struct *old_stop = cpu_rq(cpu)->stop;
931 * Make it appear like a SCHED_FIFO task, its something
932 * userspace knows about and won't get confused about.
934 * Also, it will make PI more or less work without too
935 * much confusion -- but then, stop work should not
936 * rely on PI working anyway.
938 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
940 stop->sched_class = &stop_sched_class;
943 cpu_rq(cpu)->stop = stop;
947 * Reset it back to a normal scheduling class so that
948 * it can die in pieces.
950 old_stop->sched_class = &rt_sched_class;
955 * __normal_prio - return the priority that is based on the static prio
957 static inline int __normal_prio(struct task_struct *p)
959 return p->static_prio;
963 * Calculate the expected normal priority: i.e. priority
964 * without taking RT-inheritance into account. Might be
965 * boosted by interactivity modifiers. Changes upon fork,
966 * setprio syscalls, and whenever the interactivity
967 * estimator recalculates.
969 static inline int normal_prio(struct task_struct *p)
973 if (task_has_dl_policy(p))
974 prio = MAX_DL_PRIO-1;
975 else if (task_has_rt_policy(p))
976 prio = MAX_RT_PRIO-1 - p->rt_priority;
978 prio = __normal_prio(p);
983 * Calculate the current priority, i.e. the priority
984 * taken into account by the scheduler. This value might
985 * be boosted by RT tasks, or might be boosted by
986 * interactivity modifiers. Will be RT if the task got
987 * RT-boosted. If not then it returns p->normal_prio.
989 static int effective_prio(struct task_struct *p)
991 p->normal_prio = normal_prio(p);
993 * If we are RT tasks or we were boosted to RT priority,
994 * keep the priority unchanged. Otherwise, update priority
995 * to the normal priority:
997 if (!rt_prio(p->prio))
998 return p->normal_prio;
1003 * task_curr - is this task currently executing on a CPU?
1004 * @p: the task in question.
1006 * Return: 1 if the task is currently executing. 0 otherwise.
1008 inline int task_curr(const struct task_struct *p)
1010 return cpu_curr(task_cpu(p)) == p;
1014 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1016 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1017 const struct sched_class *prev_class,
1020 if (prev_class != p->sched_class) {
1021 if (prev_class->switched_from)
1022 prev_class->switched_from(rq, p);
1023 /* Possble rq->lock 'hole'. */
1024 p->sched_class->switched_to(rq, p);
1025 } else if (oldprio != p->prio || dl_task(p))
1026 p->sched_class->prio_changed(rq, p, oldprio);
1029 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1031 const struct sched_class *class;
1033 if (p->sched_class == rq->curr->sched_class) {
1034 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1036 for_each_class(class) {
1037 if (class == rq->curr->sched_class)
1039 if (class == p->sched_class) {
1047 * A queue event has occurred, and we're going to schedule. In
1048 * this case, we can save a useless back to back clock update.
1050 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1051 rq_clock_skip_update(rq, true);
1055 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1057 #ifdef CONFIG_SCHED_DEBUG
1059 * We should never call set_task_cpu() on a blocked task,
1060 * ttwu() will sort out the placement.
1062 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1065 #ifdef CONFIG_LOCKDEP
1067 * The caller should hold either p->pi_lock or rq->lock, when changing
1068 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1070 * sched_move_task() holds both and thus holding either pins the cgroup,
1073 * Furthermore, all task_rq users should acquire both locks, see
1076 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1077 lockdep_is_held(&task_rq(p)->lock)));
1081 trace_sched_migrate_task(p, new_cpu);
1083 if (task_cpu(p) != new_cpu) {
1084 if (p->sched_class->migrate_task_rq)
1085 p->sched_class->migrate_task_rq(p, new_cpu);
1086 p->se.nr_migrations++;
1087 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1090 __set_task_cpu(p, new_cpu);
1093 static void __migrate_swap_task(struct task_struct *p, int cpu)
1095 if (task_on_rq_queued(p)) {
1096 struct rq *src_rq, *dst_rq;
1098 src_rq = task_rq(p);
1099 dst_rq = cpu_rq(cpu);
1101 deactivate_task(src_rq, p, 0);
1102 set_task_cpu(p, cpu);
1103 activate_task(dst_rq, p, 0);
1104 check_preempt_curr(dst_rq, p, 0);
1107 * Task isn't running anymore; make it appear like we migrated
1108 * it before it went to sleep. This means on wakeup we make the
1109 * previous cpu our targer instead of where it really is.
1115 struct migration_swap_arg {
1116 struct task_struct *src_task, *dst_task;
1117 int src_cpu, dst_cpu;
1120 static int migrate_swap_stop(void *data)
1122 struct migration_swap_arg *arg = data;
1123 struct rq *src_rq, *dst_rq;
1126 src_rq = cpu_rq(arg->src_cpu);
1127 dst_rq = cpu_rq(arg->dst_cpu);
1129 double_raw_lock(&arg->src_task->pi_lock,
1130 &arg->dst_task->pi_lock);
1131 double_rq_lock(src_rq, dst_rq);
1132 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1135 if (task_cpu(arg->src_task) != arg->src_cpu)
1138 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1141 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1144 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1145 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1150 double_rq_unlock(src_rq, dst_rq);
1151 raw_spin_unlock(&arg->dst_task->pi_lock);
1152 raw_spin_unlock(&arg->src_task->pi_lock);
1158 * Cross migrate two tasks
1160 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1162 struct migration_swap_arg arg;
1165 arg = (struct migration_swap_arg){
1167 .src_cpu = task_cpu(cur),
1169 .dst_cpu = task_cpu(p),
1172 if (arg.src_cpu == arg.dst_cpu)
1176 * These three tests are all lockless; this is OK since all of them
1177 * will be re-checked with proper locks held further down the line.
1179 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1182 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1185 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1188 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1189 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1195 struct migration_arg {
1196 struct task_struct *task;
1200 static int migration_cpu_stop(void *data);
1203 * wait_task_inactive - wait for a thread to unschedule.
1205 * If @match_state is nonzero, it's the @p->state value just checked and
1206 * not expected to change. If it changes, i.e. @p might have woken up,
1207 * then return zero. When we succeed in waiting for @p to be off its CPU,
1208 * we return a positive number (its total switch count). If a second call
1209 * a short while later returns the same number, the caller can be sure that
1210 * @p has remained unscheduled the whole time.
1212 * The caller must ensure that the task *will* unschedule sometime soon,
1213 * else this function might spin for a *long* time. This function can't
1214 * be called with interrupts off, or it may introduce deadlock with
1215 * smp_call_function() if an IPI is sent by the same process we are
1216 * waiting to become inactive.
1218 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1220 unsigned long flags;
1221 int running, queued;
1227 * We do the initial early heuristics without holding
1228 * any task-queue locks at all. We'll only try to get
1229 * the runqueue lock when things look like they will
1235 * If the task is actively running on another CPU
1236 * still, just relax and busy-wait without holding
1239 * NOTE! Since we don't hold any locks, it's not
1240 * even sure that "rq" stays as the right runqueue!
1241 * But we don't care, since "task_running()" will
1242 * return false if the runqueue has changed and p
1243 * is actually now running somewhere else!
1245 while (task_running(rq, p)) {
1246 if (match_state && unlikely(p->state != match_state))
1252 * Ok, time to look more closely! We need the rq
1253 * lock now, to be *sure*. If we're wrong, we'll
1254 * just go back and repeat.
1256 rq = task_rq_lock(p, &flags);
1257 trace_sched_wait_task(p);
1258 running = task_running(rq, p);
1259 queued = task_on_rq_queued(p);
1261 if (!match_state || p->state == match_state)
1262 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1263 task_rq_unlock(rq, p, &flags);
1266 * If it changed from the expected state, bail out now.
1268 if (unlikely(!ncsw))
1272 * Was it really running after all now that we
1273 * checked with the proper locks actually held?
1275 * Oops. Go back and try again..
1277 if (unlikely(running)) {
1283 * It's not enough that it's not actively running,
1284 * it must be off the runqueue _entirely_, and not
1287 * So if it was still runnable (but just not actively
1288 * running right now), it's preempted, and we should
1289 * yield - it could be a while.
1291 if (unlikely(queued)) {
1292 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1294 set_current_state(TASK_UNINTERRUPTIBLE);
1295 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1300 * Ahh, all good. It wasn't running, and it wasn't
1301 * runnable, which means that it will never become
1302 * running in the future either. We're all done!
1311 * kick_process - kick a running thread to enter/exit the kernel
1312 * @p: the to-be-kicked thread
1314 * Cause a process which is running on another CPU to enter
1315 * kernel-mode, without any delay. (to get signals handled.)
1317 * NOTE: this function doesn't have to take the runqueue lock,
1318 * because all it wants to ensure is that the remote task enters
1319 * the kernel. If the IPI races and the task has been migrated
1320 * to another CPU then no harm is done and the purpose has been
1323 void kick_process(struct task_struct *p)
1329 if ((cpu != smp_processor_id()) && task_curr(p))
1330 smp_send_reschedule(cpu);
1333 EXPORT_SYMBOL_GPL(kick_process);
1334 #endif /* CONFIG_SMP */
1338 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1340 static int select_fallback_rq(int cpu, struct task_struct *p)
1342 int nid = cpu_to_node(cpu);
1343 const struct cpumask *nodemask = NULL;
1344 enum { cpuset, possible, fail } state = cpuset;
1348 * If the node that the cpu is on has been offlined, cpu_to_node()
1349 * will return -1. There is no cpu on the node, and we should
1350 * select the cpu on the other node.
1353 nodemask = cpumask_of_node(nid);
1355 /* Look for allowed, online CPU in same node. */
1356 for_each_cpu(dest_cpu, nodemask) {
1357 if (!cpu_online(dest_cpu))
1359 if (!cpu_active(dest_cpu))
1361 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1367 /* Any allowed, online CPU? */
1368 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1369 if (!cpu_online(dest_cpu))
1371 if (!cpu_active(dest_cpu))
1378 /* No more Mr. Nice Guy. */
1379 cpuset_cpus_allowed_fallback(p);
1384 do_set_cpus_allowed(p, cpu_possible_mask);
1395 if (state != cpuset) {
1397 * Don't tell them about moving exiting tasks or
1398 * kernel threads (both mm NULL), since they never
1401 if (p->mm && printk_ratelimit()) {
1402 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1403 task_pid_nr(p), p->comm, cpu);
1411 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1414 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1416 if (p->nr_cpus_allowed > 1)
1417 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1420 * In order not to call set_task_cpu() on a blocking task we need
1421 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1424 * Since this is common to all placement strategies, this lives here.
1426 * [ this allows ->select_task() to simply return task_cpu(p) and
1427 * not worry about this generic constraint ]
1429 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1431 cpu = select_fallback_rq(task_cpu(p), p);
1436 static void update_avg(u64 *avg, u64 sample)
1438 s64 diff = sample - *avg;
1444 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1446 #ifdef CONFIG_SCHEDSTATS
1447 struct rq *rq = this_rq();
1450 int this_cpu = smp_processor_id();
1452 if (cpu == this_cpu) {
1453 schedstat_inc(rq, ttwu_local);
1454 schedstat_inc(p, se.statistics.nr_wakeups_local);
1456 struct sched_domain *sd;
1458 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1460 for_each_domain(this_cpu, sd) {
1461 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1462 schedstat_inc(sd, ttwu_wake_remote);
1469 if (wake_flags & WF_MIGRATED)
1470 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1472 #endif /* CONFIG_SMP */
1474 schedstat_inc(rq, ttwu_count);
1475 schedstat_inc(p, se.statistics.nr_wakeups);
1477 if (wake_flags & WF_SYNC)
1478 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1480 #endif /* CONFIG_SCHEDSTATS */
1483 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1485 activate_task(rq, p, en_flags);
1486 p->on_rq = TASK_ON_RQ_QUEUED;
1488 /* if a worker is waking up, notify workqueue */
1489 if (p->flags & PF_WQ_WORKER)
1490 wq_worker_waking_up(p, cpu_of(rq));
1494 * Mark the task runnable and perform wakeup-preemption.
1497 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1499 check_preempt_curr(rq, p, wake_flags);
1500 trace_sched_wakeup(p, true);
1502 p->state = TASK_RUNNING;
1504 if (p->sched_class->task_woken)
1505 p->sched_class->task_woken(rq, p);
1507 if (rq->idle_stamp) {
1508 u64 delta = rq_clock(rq) - rq->idle_stamp;
1509 u64 max = 2*rq->max_idle_balance_cost;
1511 update_avg(&rq->avg_idle, delta);
1513 if (rq->avg_idle > max)
1522 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1525 if (p->sched_contributes_to_load)
1526 rq->nr_uninterruptible--;
1529 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1530 ttwu_do_wakeup(rq, p, wake_flags);
1534 * Called in case the task @p isn't fully descheduled from its runqueue,
1535 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1536 * since all we need to do is flip p->state to TASK_RUNNING, since
1537 * the task is still ->on_rq.
1539 static int ttwu_remote(struct task_struct *p, int wake_flags)
1544 rq = __task_rq_lock(p);
1545 if (task_on_rq_queued(p)) {
1546 /* check_preempt_curr() may use rq clock */
1547 update_rq_clock(rq);
1548 ttwu_do_wakeup(rq, p, wake_flags);
1551 __task_rq_unlock(rq);
1557 void sched_ttwu_pending(void)
1559 struct rq *rq = this_rq();
1560 struct llist_node *llist = llist_del_all(&rq->wake_list);
1561 struct task_struct *p;
1562 unsigned long flags;
1567 raw_spin_lock_irqsave(&rq->lock, flags);
1570 p = llist_entry(llist, struct task_struct, wake_entry);
1571 llist = llist_next(llist);
1572 ttwu_do_activate(rq, p, 0);
1575 raw_spin_unlock_irqrestore(&rq->lock, flags);
1578 void scheduler_ipi(void)
1581 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1582 * TIF_NEED_RESCHED remotely (for the first time) will also send
1585 preempt_fold_need_resched();
1587 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1591 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1592 * traditionally all their work was done from the interrupt return
1593 * path. Now that we actually do some work, we need to make sure
1596 * Some archs already do call them, luckily irq_enter/exit nest
1599 * Arguably we should visit all archs and update all handlers,
1600 * however a fair share of IPIs are still resched only so this would
1601 * somewhat pessimize the simple resched case.
1604 sched_ttwu_pending();
1607 * Check if someone kicked us for doing the nohz idle load balance.
1609 if (unlikely(got_nohz_idle_kick())) {
1610 this_rq()->idle_balance = 1;
1611 raise_softirq_irqoff(SCHED_SOFTIRQ);
1616 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1618 struct rq *rq = cpu_rq(cpu);
1620 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1621 if (!set_nr_if_polling(rq->idle))
1622 smp_send_reschedule(cpu);
1624 trace_sched_wake_idle_without_ipi(cpu);
1628 void wake_up_if_idle(int cpu)
1630 struct rq *rq = cpu_rq(cpu);
1631 unsigned long flags;
1635 if (!is_idle_task(rcu_dereference(rq->curr)))
1638 if (set_nr_if_polling(rq->idle)) {
1639 trace_sched_wake_idle_without_ipi(cpu);
1641 raw_spin_lock_irqsave(&rq->lock, flags);
1642 if (is_idle_task(rq->curr))
1643 smp_send_reschedule(cpu);
1644 /* Else cpu is not in idle, do nothing here */
1645 raw_spin_unlock_irqrestore(&rq->lock, flags);
1652 bool cpus_share_cache(int this_cpu, int that_cpu)
1654 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1656 #endif /* CONFIG_SMP */
1658 static void ttwu_queue(struct task_struct *p, int cpu)
1660 struct rq *rq = cpu_rq(cpu);
1662 #if defined(CONFIG_SMP)
1663 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1664 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1665 ttwu_queue_remote(p, cpu);
1670 raw_spin_lock(&rq->lock);
1671 ttwu_do_activate(rq, p, 0);
1672 raw_spin_unlock(&rq->lock);
1676 * try_to_wake_up - wake up a thread
1677 * @p: the thread to be awakened
1678 * @state: the mask of task states that can be woken
1679 * @wake_flags: wake modifier flags (WF_*)
1681 * Put it on the run-queue if it's not already there. The "current"
1682 * thread is always on the run-queue (except when the actual
1683 * re-schedule is in progress), and as such you're allowed to do
1684 * the simpler "current->state = TASK_RUNNING" to mark yourself
1685 * runnable without the overhead of this.
1687 * Return: %true if @p was woken up, %false if it was already running.
1688 * or @state didn't match @p's state.
1691 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1693 unsigned long flags;
1694 int cpu, success = 0;
1697 * If we are going to wake up a thread waiting for CONDITION we
1698 * need to ensure that CONDITION=1 done by the caller can not be
1699 * reordered with p->state check below. This pairs with mb() in
1700 * set_current_state() the waiting thread does.
1702 smp_mb__before_spinlock();
1703 raw_spin_lock_irqsave(&p->pi_lock, flags);
1704 if (!(p->state & state))
1707 success = 1; /* we're going to change ->state */
1710 if (p->on_rq && ttwu_remote(p, wake_flags))
1715 * If the owning (remote) cpu is still in the middle of schedule() with
1716 * this task as prev, wait until its done referencing the task.
1721 * Pairs with the smp_wmb() in finish_lock_switch().
1725 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1726 p->state = TASK_WAKING;
1728 if (p->sched_class->task_waking)
1729 p->sched_class->task_waking(p);
1731 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1732 if (task_cpu(p) != cpu) {
1733 wake_flags |= WF_MIGRATED;
1734 set_task_cpu(p, cpu);
1736 #endif /* CONFIG_SMP */
1740 ttwu_stat(p, cpu, wake_flags);
1742 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1748 * try_to_wake_up_local - try to wake up a local task with rq lock held
1749 * @p: the thread to be awakened
1751 * Put @p on the run-queue if it's not already there. The caller must
1752 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1755 static void try_to_wake_up_local(struct task_struct *p)
1757 struct rq *rq = task_rq(p);
1759 if (WARN_ON_ONCE(rq != this_rq()) ||
1760 WARN_ON_ONCE(p == current))
1763 lockdep_assert_held(&rq->lock);
1765 if (!raw_spin_trylock(&p->pi_lock)) {
1766 raw_spin_unlock(&rq->lock);
1767 raw_spin_lock(&p->pi_lock);
1768 raw_spin_lock(&rq->lock);
1771 if (!(p->state & TASK_NORMAL))
1774 if (!task_on_rq_queued(p))
1775 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1777 ttwu_do_wakeup(rq, p, 0);
1778 ttwu_stat(p, smp_processor_id(), 0);
1780 raw_spin_unlock(&p->pi_lock);
1784 * wake_up_process - Wake up a specific process
1785 * @p: The process to be woken up.
1787 * Attempt to wake up the nominated process and move it to the set of runnable
1790 * Return: 1 if the process was woken up, 0 if it was already running.
1792 * It may be assumed that this function implies a write memory barrier before
1793 * changing the task state if and only if any tasks are woken up.
1795 int wake_up_process(struct task_struct *p)
1797 WARN_ON(task_is_stopped_or_traced(p));
1798 return try_to_wake_up(p, TASK_NORMAL, 0);
1800 EXPORT_SYMBOL(wake_up_process);
1802 int wake_up_state(struct task_struct *p, unsigned int state)
1804 return try_to_wake_up(p, state, 0);
1808 * This function clears the sched_dl_entity static params.
1810 void __dl_clear_params(struct task_struct *p)
1812 struct sched_dl_entity *dl_se = &p->dl;
1814 dl_se->dl_runtime = 0;
1815 dl_se->dl_deadline = 0;
1816 dl_se->dl_period = 0;
1820 dl_se->dl_throttled = 0;
1822 dl_se->dl_yielded = 0;
1826 * Perform scheduler related setup for a newly forked process p.
1827 * p is forked by current.
1829 * __sched_fork() is basic setup used by init_idle() too:
1831 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1836 p->se.exec_start = 0;
1837 p->se.sum_exec_runtime = 0;
1838 p->se.prev_sum_exec_runtime = 0;
1839 p->se.nr_migrations = 0;
1842 p->se.avg.decay_count = 0;
1844 INIT_LIST_HEAD(&p->se.group_node);
1846 #ifdef CONFIG_SCHEDSTATS
1847 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1850 RB_CLEAR_NODE(&p->dl.rb_node);
1851 init_dl_task_timer(&p->dl);
1852 __dl_clear_params(p);
1854 INIT_LIST_HEAD(&p->rt.run_list);
1856 #ifdef CONFIG_PREEMPT_NOTIFIERS
1857 INIT_HLIST_HEAD(&p->preempt_notifiers);
1860 #ifdef CONFIG_NUMA_BALANCING
1861 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1862 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1863 p->mm->numa_scan_seq = 0;
1866 if (clone_flags & CLONE_VM)
1867 p->numa_preferred_nid = current->numa_preferred_nid;
1869 p->numa_preferred_nid = -1;
1871 p->node_stamp = 0ULL;
1872 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1873 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1874 p->numa_work.next = &p->numa_work;
1875 p->numa_faults = NULL;
1876 p->last_task_numa_placement = 0;
1877 p->last_sum_exec_runtime = 0;
1879 p->numa_group = NULL;
1880 #endif /* CONFIG_NUMA_BALANCING */
1883 #ifdef CONFIG_NUMA_BALANCING
1884 #ifdef CONFIG_SCHED_DEBUG
1885 void set_numabalancing_state(bool enabled)
1888 sched_feat_set("NUMA");
1890 sched_feat_set("NO_NUMA");
1893 __read_mostly bool numabalancing_enabled;
1895 void set_numabalancing_state(bool enabled)
1897 numabalancing_enabled = enabled;
1899 #endif /* CONFIG_SCHED_DEBUG */
1901 #ifdef CONFIG_PROC_SYSCTL
1902 int sysctl_numa_balancing(struct ctl_table *table, int write,
1903 void __user *buffer, size_t *lenp, loff_t *ppos)
1907 int state = numabalancing_enabled;
1909 if (write && !capable(CAP_SYS_ADMIN))
1914 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1918 set_numabalancing_state(state);
1925 * fork()/clone()-time setup:
1927 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1929 unsigned long flags;
1930 int cpu = get_cpu();
1932 __sched_fork(clone_flags, p);
1934 * We mark the process as running here. This guarantees that
1935 * nobody will actually run it, and a signal or other external
1936 * event cannot wake it up and insert it on the runqueue either.
1938 p->state = TASK_RUNNING;
1941 * Make sure we do not leak PI boosting priority to the child.
1943 p->prio = current->normal_prio;
1946 * Revert to default priority/policy on fork if requested.
1948 if (unlikely(p->sched_reset_on_fork)) {
1949 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1950 p->policy = SCHED_NORMAL;
1951 p->static_prio = NICE_TO_PRIO(0);
1953 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1954 p->static_prio = NICE_TO_PRIO(0);
1956 p->prio = p->normal_prio = __normal_prio(p);
1960 * We don't need the reset flag anymore after the fork. It has
1961 * fulfilled its duty:
1963 p->sched_reset_on_fork = 0;
1966 if (dl_prio(p->prio)) {
1969 } else if (rt_prio(p->prio)) {
1970 p->sched_class = &rt_sched_class;
1972 p->sched_class = &fair_sched_class;
1975 if (p->sched_class->task_fork)
1976 p->sched_class->task_fork(p);
1979 * The child is not yet in the pid-hash so no cgroup attach races,
1980 * and the cgroup is pinned to this child due to cgroup_fork()
1981 * is ran before sched_fork().
1983 * Silence PROVE_RCU.
1985 raw_spin_lock_irqsave(&p->pi_lock, flags);
1986 set_task_cpu(p, cpu);
1987 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1989 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1990 if (likely(sched_info_on()))
1991 memset(&p->sched_info, 0, sizeof(p->sched_info));
1993 #if defined(CONFIG_SMP)
1996 init_task_preempt_count(p);
1998 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1999 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2006 unsigned long to_ratio(u64 period, u64 runtime)
2008 if (runtime == RUNTIME_INF)
2012 * Doing this here saves a lot of checks in all
2013 * the calling paths, and returning zero seems
2014 * safe for them anyway.
2019 return div64_u64(runtime << 20, period);
2023 inline struct dl_bw *dl_bw_of(int i)
2025 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2026 "sched RCU must be held");
2027 return &cpu_rq(i)->rd->dl_bw;
2030 static inline int dl_bw_cpus(int i)
2032 struct root_domain *rd = cpu_rq(i)->rd;
2035 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2036 "sched RCU must be held");
2037 for_each_cpu_and(i, rd->span, cpu_active_mask)
2043 inline struct dl_bw *dl_bw_of(int i)
2045 return &cpu_rq(i)->dl.dl_bw;
2048 static inline int dl_bw_cpus(int i)
2055 * We must be sure that accepting a new task (or allowing changing the
2056 * parameters of an existing one) is consistent with the bandwidth
2057 * constraints. If yes, this function also accordingly updates the currently
2058 * allocated bandwidth to reflect the new situation.
2060 * This function is called while holding p's rq->lock.
2062 * XXX we should delay bw change until the task's 0-lag point, see
2065 static int dl_overflow(struct task_struct *p, int policy,
2066 const struct sched_attr *attr)
2069 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2070 u64 period = attr->sched_period ?: attr->sched_deadline;
2071 u64 runtime = attr->sched_runtime;
2072 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2075 if (new_bw == p->dl.dl_bw)
2079 * Either if a task, enters, leave, or stays -deadline but changes
2080 * its parameters, we may need to update accordingly the total
2081 * allocated bandwidth of the container.
2083 raw_spin_lock(&dl_b->lock);
2084 cpus = dl_bw_cpus(task_cpu(p));
2085 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2086 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2087 __dl_add(dl_b, new_bw);
2089 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2090 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2091 __dl_clear(dl_b, p->dl.dl_bw);
2092 __dl_add(dl_b, new_bw);
2094 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2095 __dl_clear(dl_b, p->dl.dl_bw);
2098 raw_spin_unlock(&dl_b->lock);
2103 extern void init_dl_bw(struct dl_bw *dl_b);
2106 * wake_up_new_task - wake up a newly created task for the first time.
2108 * This function will do some initial scheduler statistics housekeeping
2109 * that must be done for every newly created context, then puts the task
2110 * on the runqueue and wakes it.
2112 void wake_up_new_task(struct task_struct *p)
2114 unsigned long flags;
2117 raw_spin_lock_irqsave(&p->pi_lock, flags);
2120 * Fork balancing, do it here and not earlier because:
2121 * - cpus_allowed can change in the fork path
2122 * - any previously selected cpu might disappear through hotplug
2124 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2127 /* Initialize new task's runnable average */
2128 init_task_runnable_average(p);
2129 rq = __task_rq_lock(p);
2130 activate_task(rq, p, 0);
2131 p->on_rq = TASK_ON_RQ_QUEUED;
2132 trace_sched_wakeup_new(p, true);
2133 check_preempt_curr(rq, p, WF_FORK);
2135 if (p->sched_class->task_woken)
2136 p->sched_class->task_woken(rq, p);
2138 task_rq_unlock(rq, p, &flags);
2141 #ifdef CONFIG_PREEMPT_NOTIFIERS
2144 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2145 * @notifier: notifier struct to register
2147 void preempt_notifier_register(struct preempt_notifier *notifier)
2149 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2151 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2154 * preempt_notifier_unregister - no longer interested in preemption notifications
2155 * @notifier: notifier struct to unregister
2157 * This is safe to call from within a preemption notifier.
2159 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2161 hlist_del(¬ifier->link);
2163 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2165 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2167 struct preempt_notifier *notifier;
2169 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2170 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2174 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2175 struct task_struct *next)
2177 struct preempt_notifier *notifier;
2179 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2180 notifier->ops->sched_out(notifier, next);
2183 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2185 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2190 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2191 struct task_struct *next)
2195 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2198 * prepare_task_switch - prepare to switch tasks
2199 * @rq: the runqueue preparing to switch
2200 * @prev: the current task that is being switched out
2201 * @next: the task we are going to switch to.
2203 * This is called with the rq lock held and interrupts off. It must
2204 * be paired with a subsequent finish_task_switch after the context
2207 * prepare_task_switch sets up locking and calls architecture specific
2211 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2212 struct task_struct *next)
2214 trace_sched_switch(prev, next);
2215 sched_info_switch(rq, prev, next);
2216 perf_event_task_sched_out(prev, next);
2217 fire_sched_out_preempt_notifiers(prev, next);
2218 prepare_lock_switch(rq, next);
2219 prepare_arch_switch(next);
2223 * finish_task_switch - clean up after a task-switch
2224 * @prev: the thread we just switched away from.
2226 * finish_task_switch must be called after the context switch, paired
2227 * with a prepare_task_switch call before the context switch.
2228 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2229 * and do any other architecture-specific cleanup actions.
2231 * Note that we may have delayed dropping an mm in context_switch(). If
2232 * so, we finish that here outside of the runqueue lock. (Doing it
2233 * with the lock held can cause deadlocks; see schedule() for
2236 * The context switch have flipped the stack from under us and restored the
2237 * local variables which were saved when this task called schedule() in the
2238 * past. prev == current is still correct but we need to recalculate this_rq
2239 * because prev may have moved to another CPU.
2241 static struct rq *finish_task_switch(struct task_struct *prev)
2242 __releases(rq->lock)
2244 struct rq *rq = this_rq();
2245 struct mm_struct *mm = rq->prev_mm;
2251 * A task struct has one reference for the use as "current".
2252 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2253 * schedule one last time. The schedule call will never return, and
2254 * the scheduled task must drop that reference.
2255 * The test for TASK_DEAD must occur while the runqueue locks are
2256 * still held, otherwise prev could be scheduled on another cpu, die
2257 * there before we look at prev->state, and then the reference would
2259 * Manfred Spraul <manfred@colorfullife.com>
2261 prev_state = prev->state;
2262 vtime_task_switch(prev);
2263 finish_arch_switch(prev);
2264 perf_event_task_sched_in(prev, current);
2265 finish_lock_switch(rq, prev);
2266 finish_arch_post_lock_switch();
2268 fire_sched_in_preempt_notifiers(current);
2271 if (unlikely(prev_state == TASK_DEAD)) {
2272 if (prev->sched_class->task_dead)
2273 prev->sched_class->task_dead(prev);
2276 * Remove function-return probe instances associated with this
2277 * task and put them back on the free list.
2279 kprobe_flush_task(prev);
2280 put_task_struct(prev);
2283 tick_nohz_task_switch(current);
2289 /* rq->lock is NOT held, but preemption is disabled */
2290 static inline void post_schedule(struct rq *rq)
2292 if (rq->post_schedule) {
2293 unsigned long flags;
2295 raw_spin_lock_irqsave(&rq->lock, flags);
2296 if (rq->curr->sched_class->post_schedule)
2297 rq->curr->sched_class->post_schedule(rq);
2298 raw_spin_unlock_irqrestore(&rq->lock, flags);
2300 rq->post_schedule = 0;
2306 static inline void post_schedule(struct rq *rq)
2313 * schedule_tail - first thing a freshly forked thread must call.
2314 * @prev: the thread we just switched away from.
2316 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2317 __releases(rq->lock)
2321 /* finish_task_switch() drops rq->lock and enables preemtion */
2323 rq = finish_task_switch(prev);
2327 if (current->set_child_tid)
2328 put_user(task_pid_vnr(current), current->set_child_tid);
2332 * context_switch - switch to the new MM and the new thread's register state.
2334 static inline struct rq *
2335 context_switch(struct rq *rq, struct task_struct *prev,
2336 struct task_struct *next)
2338 struct mm_struct *mm, *oldmm;
2340 prepare_task_switch(rq, prev, next);
2343 oldmm = prev->active_mm;
2345 * For paravirt, this is coupled with an exit in switch_to to
2346 * combine the page table reload and the switch backend into
2349 arch_start_context_switch(prev);
2352 next->active_mm = oldmm;
2353 atomic_inc(&oldmm->mm_count);
2354 enter_lazy_tlb(oldmm, next);
2356 switch_mm(oldmm, mm, next);
2359 prev->active_mm = NULL;
2360 rq->prev_mm = oldmm;
2363 * Since the runqueue lock will be released by the next
2364 * task (which is an invalid locking op but in the case
2365 * of the scheduler it's an obvious special-case), so we
2366 * do an early lockdep release here:
2368 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2370 context_tracking_task_switch(prev, next);
2371 /* Here we just switch the register state and the stack. */
2372 switch_to(prev, next, prev);
2375 return finish_task_switch(prev);
2379 * nr_running and nr_context_switches:
2381 * externally visible scheduler statistics: current number of runnable
2382 * threads, total number of context switches performed since bootup.
2384 unsigned long nr_running(void)
2386 unsigned long i, sum = 0;
2388 for_each_online_cpu(i)
2389 sum += cpu_rq(i)->nr_running;
2395 * Check if only the current task is running on the cpu.
2397 bool single_task_running(void)
2399 if (cpu_rq(smp_processor_id())->nr_running == 1)
2404 EXPORT_SYMBOL(single_task_running);
2406 unsigned long long nr_context_switches(void)
2409 unsigned long long sum = 0;
2411 for_each_possible_cpu(i)
2412 sum += cpu_rq(i)->nr_switches;
2417 unsigned long nr_iowait(void)
2419 unsigned long i, sum = 0;
2421 for_each_possible_cpu(i)
2422 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2427 unsigned long nr_iowait_cpu(int cpu)
2429 struct rq *this = cpu_rq(cpu);
2430 return atomic_read(&this->nr_iowait);
2433 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2435 struct rq *this = this_rq();
2436 *nr_waiters = atomic_read(&this->nr_iowait);
2437 *load = this->cpu_load[0];
2443 * sched_exec - execve() is a valuable balancing opportunity, because at
2444 * this point the task has the smallest effective memory and cache footprint.
2446 void sched_exec(void)
2448 struct task_struct *p = current;
2449 unsigned long flags;
2452 raw_spin_lock_irqsave(&p->pi_lock, flags);
2453 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2454 if (dest_cpu == smp_processor_id())
2457 if (likely(cpu_active(dest_cpu))) {
2458 struct migration_arg arg = { p, dest_cpu };
2460 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2461 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2465 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2470 DEFINE_PER_CPU(struct kernel_stat, kstat);
2471 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2473 EXPORT_PER_CPU_SYMBOL(kstat);
2474 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2477 * Return accounted runtime for the task.
2478 * In case the task is currently running, return the runtime plus current's
2479 * pending runtime that have not been accounted yet.
2481 unsigned long long task_sched_runtime(struct task_struct *p)
2483 unsigned long flags;
2487 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2489 * 64-bit doesn't need locks to atomically read a 64bit value.
2490 * So we have a optimization chance when the task's delta_exec is 0.
2491 * Reading ->on_cpu is racy, but this is ok.
2493 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2494 * If we race with it entering cpu, unaccounted time is 0. This is
2495 * indistinguishable from the read occurring a few cycles earlier.
2496 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2497 * been accounted, so we're correct here as well.
2499 if (!p->on_cpu || !task_on_rq_queued(p))
2500 return p->se.sum_exec_runtime;
2503 rq = task_rq_lock(p, &flags);
2505 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2506 * project cycles that may never be accounted to this
2507 * thread, breaking clock_gettime().
2509 if (task_current(rq, p) && task_on_rq_queued(p)) {
2510 update_rq_clock(rq);
2511 p->sched_class->update_curr(rq);
2513 ns = p->se.sum_exec_runtime;
2514 task_rq_unlock(rq, p, &flags);
2520 * This function gets called by the timer code, with HZ frequency.
2521 * We call it with interrupts disabled.
2523 void scheduler_tick(void)
2525 int cpu = smp_processor_id();
2526 struct rq *rq = cpu_rq(cpu);
2527 struct task_struct *curr = rq->curr;
2531 raw_spin_lock(&rq->lock);
2532 update_rq_clock(rq);
2533 curr->sched_class->task_tick(rq, curr, 0);
2534 update_cpu_load_active(rq);
2535 raw_spin_unlock(&rq->lock);
2537 perf_event_task_tick();
2540 rq->idle_balance = idle_cpu(cpu);
2541 trigger_load_balance(rq);
2543 rq_last_tick_reset(rq);
2546 #ifdef CONFIG_NO_HZ_FULL
2548 * scheduler_tick_max_deferment
2550 * Keep at least one tick per second when a single
2551 * active task is running because the scheduler doesn't
2552 * yet completely support full dynticks environment.
2554 * This makes sure that uptime, CFS vruntime, load
2555 * balancing, etc... continue to move forward, even
2556 * with a very low granularity.
2558 * Return: Maximum deferment in nanoseconds.
2560 u64 scheduler_tick_max_deferment(void)
2562 struct rq *rq = this_rq();
2563 unsigned long next, now = ACCESS_ONCE(jiffies);
2565 next = rq->last_sched_tick + HZ;
2567 if (time_before_eq(next, now))
2570 return jiffies_to_nsecs(next - now);
2574 notrace unsigned long get_parent_ip(unsigned long addr)
2576 if (in_lock_functions(addr)) {
2577 addr = CALLER_ADDR2;
2578 if (in_lock_functions(addr))
2579 addr = CALLER_ADDR3;
2584 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2585 defined(CONFIG_PREEMPT_TRACER))
2587 void preempt_count_add(int val)
2589 #ifdef CONFIG_DEBUG_PREEMPT
2593 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2596 __preempt_count_add(val);
2597 #ifdef CONFIG_DEBUG_PREEMPT
2599 * Spinlock count overflowing soon?
2601 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2604 if (preempt_count() == val) {
2605 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2606 #ifdef CONFIG_DEBUG_PREEMPT
2607 current->preempt_disable_ip = ip;
2609 trace_preempt_off(CALLER_ADDR0, ip);
2612 EXPORT_SYMBOL(preempt_count_add);
2613 NOKPROBE_SYMBOL(preempt_count_add);
2615 void preempt_count_sub(int val)
2617 #ifdef CONFIG_DEBUG_PREEMPT
2621 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2624 * Is the spinlock portion underflowing?
2626 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2627 !(preempt_count() & PREEMPT_MASK)))
2631 if (preempt_count() == val)
2632 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2633 __preempt_count_sub(val);
2635 EXPORT_SYMBOL(preempt_count_sub);
2636 NOKPROBE_SYMBOL(preempt_count_sub);
2641 * Print scheduling while atomic bug:
2643 static noinline void __schedule_bug(struct task_struct *prev)
2645 if (oops_in_progress)
2648 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2649 prev->comm, prev->pid, preempt_count());
2651 debug_show_held_locks(prev);
2653 if (irqs_disabled())
2654 print_irqtrace_events(prev);
2655 #ifdef CONFIG_DEBUG_PREEMPT
2656 if (in_atomic_preempt_off()) {
2657 pr_err("Preemption disabled at:");
2658 print_ip_sym(current->preempt_disable_ip);
2663 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2667 * Various schedule()-time debugging checks and statistics:
2669 static inline void schedule_debug(struct task_struct *prev)
2671 #ifdef CONFIG_SCHED_STACK_END_CHECK
2672 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2675 * Test if we are atomic. Since do_exit() needs to call into
2676 * schedule() atomically, we ignore that path. Otherwise whine
2677 * if we are scheduling when we should not.
2679 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2680 __schedule_bug(prev);
2683 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2685 schedstat_inc(this_rq(), sched_count);
2689 * Pick up the highest-prio task:
2691 static inline struct task_struct *
2692 pick_next_task(struct rq *rq, struct task_struct *prev)
2694 const struct sched_class *class = &fair_sched_class;
2695 struct task_struct *p;
2698 * Optimization: we know that if all tasks are in
2699 * the fair class we can call that function directly:
2701 if (likely(prev->sched_class == class &&
2702 rq->nr_running == rq->cfs.h_nr_running)) {
2703 p = fair_sched_class.pick_next_task(rq, prev);
2704 if (unlikely(p == RETRY_TASK))
2707 /* assumes fair_sched_class->next == idle_sched_class */
2709 p = idle_sched_class.pick_next_task(rq, prev);
2715 for_each_class(class) {
2716 p = class->pick_next_task(rq, prev);
2718 if (unlikely(p == RETRY_TASK))
2724 BUG(); /* the idle class will always have a runnable task */
2728 * __schedule() is the main scheduler function.
2730 * The main means of driving the scheduler and thus entering this function are:
2732 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2734 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2735 * paths. For example, see arch/x86/entry_64.S.
2737 * To drive preemption between tasks, the scheduler sets the flag in timer
2738 * interrupt handler scheduler_tick().
2740 * 3. Wakeups don't really cause entry into schedule(). They add a
2741 * task to the run-queue and that's it.
2743 * Now, if the new task added to the run-queue preempts the current
2744 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2745 * called on the nearest possible occasion:
2747 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2749 * - in syscall or exception context, at the next outmost
2750 * preempt_enable(). (this might be as soon as the wake_up()'s
2753 * - in IRQ context, return from interrupt-handler to
2754 * preemptible context
2756 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2759 * - cond_resched() call
2760 * - explicit schedule() call
2761 * - return from syscall or exception to user-space
2762 * - return from interrupt-handler to user-space
2764 static void __sched __schedule(void)
2766 struct task_struct *prev, *next;
2767 unsigned long *switch_count;
2773 cpu = smp_processor_id();
2775 rcu_note_context_switch();
2778 schedule_debug(prev);
2780 if (sched_feat(HRTICK))
2784 * Make sure that signal_pending_state()->signal_pending() below
2785 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2786 * done by the caller to avoid the race with signal_wake_up().
2788 smp_mb__before_spinlock();
2789 raw_spin_lock_irq(&rq->lock);
2791 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2793 switch_count = &prev->nivcsw;
2794 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2795 if (unlikely(signal_pending_state(prev->state, prev))) {
2796 prev->state = TASK_RUNNING;
2798 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2802 * If a worker went to sleep, notify and ask workqueue
2803 * whether it wants to wake up a task to maintain
2806 if (prev->flags & PF_WQ_WORKER) {
2807 struct task_struct *to_wakeup;
2809 to_wakeup = wq_worker_sleeping(prev, cpu);
2811 try_to_wake_up_local(to_wakeup);
2814 switch_count = &prev->nvcsw;
2817 if (task_on_rq_queued(prev))
2818 update_rq_clock(rq);
2820 next = pick_next_task(rq, prev);
2821 clear_tsk_need_resched(prev);
2822 clear_preempt_need_resched();
2823 rq->clock_skip_update = 0;
2825 if (likely(prev != next)) {
2830 rq = context_switch(rq, prev, next); /* unlocks the rq */
2833 raw_spin_unlock_irq(&rq->lock);
2837 sched_preempt_enable_no_resched();
2842 static inline void sched_submit_work(struct task_struct *tsk)
2844 if (!tsk->state || tsk_is_pi_blocked(tsk))
2847 * If we are going to sleep and we have plugged IO queued,
2848 * make sure to submit it to avoid deadlocks.
2850 if (blk_needs_flush_plug(tsk))
2851 blk_schedule_flush_plug(tsk);
2854 asmlinkage __visible void __sched schedule(void)
2856 struct task_struct *tsk = current;
2858 sched_submit_work(tsk);
2861 EXPORT_SYMBOL(schedule);
2863 #ifdef CONFIG_CONTEXT_TRACKING
2864 asmlinkage __visible void __sched schedule_user(void)
2867 * If we come here after a random call to set_need_resched(),
2868 * or we have been woken up remotely but the IPI has not yet arrived,
2869 * we haven't yet exited the RCU idle mode. Do it here manually until
2870 * we find a better solution.
2872 * NB: There are buggy callers of this function. Ideally we
2873 * should warn if prev_state != IN_USER, but that will trigger
2874 * too frequently to make sense yet.
2876 enum ctx_state prev_state = exception_enter();
2878 exception_exit(prev_state);
2883 * schedule_preempt_disabled - called with preemption disabled
2885 * Returns with preemption disabled. Note: preempt_count must be 1
2887 void __sched schedule_preempt_disabled(void)
2889 sched_preempt_enable_no_resched();
2894 static void preempt_schedule_common(void)
2897 __preempt_count_add(PREEMPT_ACTIVE);
2899 __preempt_count_sub(PREEMPT_ACTIVE);
2902 * Check again in case we missed a preemption opportunity
2903 * between schedule and now.
2906 } while (need_resched());
2909 #ifdef CONFIG_PREEMPT
2911 * this is the entry point to schedule() from in-kernel preemption
2912 * off of preempt_enable. Kernel preemptions off return from interrupt
2913 * occur there and call schedule directly.
2915 asmlinkage __visible void __sched notrace preempt_schedule(void)
2918 * If there is a non-zero preempt_count or interrupts are disabled,
2919 * we do not want to preempt the current task. Just return..
2921 if (likely(!preemptible()))
2924 preempt_schedule_common();
2926 NOKPROBE_SYMBOL(preempt_schedule);
2927 EXPORT_SYMBOL(preempt_schedule);
2929 #ifdef CONFIG_CONTEXT_TRACKING
2931 * preempt_schedule_context - preempt_schedule called by tracing
2933 * The tracing infrastructure uses preempt_enable_notrace to prevent
2934 * recursion and tracing preempt enabling caused by the tracing
2935 * infrastructure itself. But as tracing can happen in areas coming
2936 * from userspace or just about to enter userspace, a preempt enable
2937 * can occur before user_exit() is called. This will cause the scheduler
2938 * to be called when the system is still in usermode.
2940 * To prevent this, the preempt_enable_notrace will use this function
2941 * instead of preempt_schedule() to exit user context if needed before
2942 * calling the scheduler.
2944 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2946 enum ctx_state prev_ctx;
2948 if (likely(!preemptible()))
2952 __preempt_count_add(PREEMPT_ACTIVE);
2954 * Needs preempt disabled in case user_exit() is traced
2955 * and the tracer calls preempt_enable_notrace() causing
2956 * an infinite recursion.
2958 prev_ctx = exception_enter();
2960 exception_exit(prev_ctx);
2962 __preempt_count_sub(PREEMPT_ACTIVE);
2964 } while (need_resched());
2966 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2967 #endif /* CONFIG_CONTEXT_TRACKING */
2969 #endif /* CONFIG_PREEMPT */
2972 * this is the entry point to schedule() from kernel preemption
2973 * off of irq context.
2974 * Note, that this is called and return with irqs disabled. This will
2975 * protect us against recursive calling from irq.
2977 asmlinkage __visible void __sched preempt_schedule_irq(void)
2979 enum ctx_state prev_state;
2981 /* Catch callers which need to be fixed */
2982 BUG_ON(preempt_count() || !irqs_disabled());
2984 prev_state = exception_enter();
2987 __preempt_count_add(PREEMPT_ACTIVE);
2990 local_irq_disable();
2991 __preempt_count_sub(PREEMPT_ACTIVE);
2994 * Check again in case we missed a preemption opportunity
2995 * between schedule and now.
2998 } while (need_resched());
3000 exception_exit(prev_state);
3003 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3006 return try_to_wake_up(curr->private, mode, wake_flags);
3008 EXPORT_SYMBOL(default_wake_function);
3010 #ifdef CONFIG_RT_MUTEXES
3013 * rt_mutex_setprio - set the current priority of a task
3015 * @prio: prio value (kernel-internal form)
3017 * This function changes the 'effective' priority of a task. It does
3018 * not touch ->normal_prio like __setscheduler().
3020 * Used by the rt_mutex code to implement priority inheritance
3021 * logic. Call site only calls if the priority of the task changed.
3023 void rt_mutex_setprio(struct task_struct *p, int prio)
3025 int oldprio, queued, running, enqueue_flag = 0;
3027 const struct sched_class *prev_class;
3029 BUG_ON(prio > MAX_PRIO);
3031 rq = __task_rq_lock(p);
3034 * Idle task boosting is a nono in general. There is one
3035 * exception, when PREEMPT_RT and NOHZ is active:
3037 * The idle task calls get_next_timer_interrupt() and holds
3038 * the timer wheel base->lock on the CPU and another CPU wants
3039 * to access the timer (probably to cancel it). We can safely
3040 * ignore the boosting request, as the idle CPU runs this code
3041 * with interrupts disabled and will complete the lock
3042 * protected section without being interrupted. So there is no
3043 * real need to boost.
3045 if (unlikely(p == rq->idle)) {
3046 WARN_ON(p != rq->curr);
3047 WARN_ON(p->pi_blocked_on);
3051 trace_sched_pi_setprio(p, prio);
3053 prev_class = p->sched_class;
3054 queued = task_on_rq_queued(p);
3055 running = task_current(rq, p);
3057 dequeue_task(rq, p, 0);
3059 put_prev_task(rq, p);
3062 * Boosting condition are:
3063 * 1. -rt task is running and holds mutex A
3064 * --> -dl task blocks on mutex A
3066 * 2. -dl task is running and holds mutex A
3067 * --> -dl task blocks on mutex A and could preempt the
3070 if (dl_prio(prio)) {
3071 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3072 if (!dl_prio(p->normal_prio) ||
3073 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3074 p->dl.dl_boosted = 1;
3075 p->dl.dl_throttled = 0;
3076 enqueue_flag = ENQUEUE_REPLENISH;
3078 p->dl.dl_boosted = 0;
3079 p->sched_class = &dl_sched_class;
3080 } else if (rt_prio(prio)) {
3081 if (dl_prio(oldprio))
3082 p->dl.dl_boosted = 0;
3084 enqueue_flag = ENQUEUE_HEAD;
3085 p->sched_class = &rt_sched_class;
3087 if (dl_prio(oldprio))
3088 p->dl.dl_boosted = 0;
3089 p->sched_class = &fair_sched_class;
3095 p->sched_class->set_curr_task(rq);
3097 enqueue_task(rq, p, enqueue_flag);
3099 check_class_changed(rq, p, prev_class, oldprio);
3101 __task_rq_unlock(rq);
3105 void set_user_nice(struct task_struct *p, long nice)
3107 int old_prio, delta, queued;
3108 unsigned long flags;
3111 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3114 * We have to be careful, if called from sys_setpriority(),
3115 * the task might be in the middle of scheduling on another CPU.
3117 rq = task_rq_lock(p, &flags);
3119 * The RT priorities are set via sched_setscheduler(), but we still
3120 * allow the 'normal' nice value to be set - but as expected
3121 * it wont have any effect on scheduling until the task is
3122 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3124 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3125 p->static_prio = NICE_TO_PRIO(nice);
3128 queued = task_on_rq_queued(p);
3130 dequeue_task(rq, p, 0);
3132 p->static_prio = NICE_TO_PRIO(nice);
3135 p->prio = effective_prio(p);
3136 delta = p->prio - old_prio;
3139 enqueue_task(rq, p, 0);
3141 * If the task increased its priority or is running and
3142 * lowered its priority, then reschedule its CPU:
3144 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3148 task_rq_unlock(rq, p, &flags);
3150 EXPORT_SYMBOL(set_user_nice);
3153 * can_nice - check if a task can reduce its nice value
3157 int can_nice(const struct task_struct *p, const int nice)
3159 /* convert nice value [19,-20] to rlimit style value [1,40] */
3160 int nice_rlim = nice_to_rlimit(nice);
3162 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3163 capable(CAP_SYS_NICE));
3166 #ifdef __ARCH_WANT_SYS_NICE
3169 * sys_nice - change the priority of the current process.
3170 * @increment: priority increment
3172 * sys_setpriority is a more generic, but much slower function that
3173 * does similar things.
3175 SYSCALL_DEFINE1(nice, int, increment)
3180 * Setpriority might change our priority at the same moment.
3181 * We don't have to worry. Conceptually one call occurs first
3182 * and we have a single winner.
3184 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3185 nice = task_nice(current) + increment;
3187 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3188 if (increment < 0 && !can_nice(current, nice))
3191 retval = security_task_setnice(current, nice);
3195 set_user_nice(current, nice);
3202 * task_prio - return the priority value of a given task.
3203 * @p: the task in question.
3205 * Return: The priority value as seen by users in /proc.
3206 * RT tasks are offset by -200. Normal tasks are centered
3207 * around 0, value goes from -16 to +15.
3209 int task_prio(const struct task_struct *p)
3211 return p->prio - MAX_RT_PRIO;
3215 * idle_cpu - is a given cpu idle currently?
3216 * @cpu: the processor in question.
3218 * Return: 1 if the CPU is currently idle. 0 otherwise.
3220 int idle_cpu(int cpu)
3222 struct rq *rq = cpu_rq(cpu);
3224 if (rq->curr != rq->idle)
3231 if (!llist_empty(&rq->wake_list))
3239 * idle_task - return the idle task for a given cpu.
3240 * @cpu: the processor in question.
3242 * Return: The idle task for the cpu @cpu.
3244 struct task_struct *idle_task(int cpu)
3246 return cpu_rq(cpu)->idle;
3250 * find_process_by_pid - find a process with a matching PID value.
3251 * @pid: the pid in question.
3253 * The task of @pid, if found. %NULL otherwise.
3255 static struct task_struct *find_process_by_pid(pid_t pid)
3257 return pid ? find_task_by_vpid(pid) : current;
3261 * This function initializes the sched_dl_entity of a newly becoming
3262 * SCHED_DEADLINE task.
3264 * Only the static values are considered here, the actual runtime and the
3265 * absolute deadline will be properly calculated when the task is enqueued
3266 * for the first time with its new policy.
3269 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3271 struct sched_dl_entity *dl_se = &p->dl;
3273 dl_se->dl_runtime = attr->sched_runtime;
3274 dl_se->dl_deadline = attr->sched_deadline;
3275 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3276 dl_se->flags = attr->sched_flags;
3277 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3280 * Changing the parameters of a task is 'tricky' and we're not doing
3281 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3283 * What we SHOULD do is delay the bandwidth release until the 0-lag
3284 * point. This would include retaining the task_struct until that time
3285 * and change dl_overflow() to not immediately decrement the current
3288 * Instead we retain the current runtime/deadline and let the new
3289 * parameters take effect after the current reservation period lapses.
3290 * This is safe (albeit pessimistic) because the 0-lag point is always
3291 * before the current scheduling deadline.
3293 * We can still have temporary overloads because we do not delay the
3294 * change in bandwidth until that time; so admission control is
3295 * not on the safe side. It does however guarantee tasks will never
3296 * consume more than promised.
3301 * sched_setparam() passes in -1 for its policy, to let the functions
3302 * it calls know not to change it.
3304 #define SETPARAM_POLICY -1
3306 static void __setscheduler_params(struct task_struct *p,
3307 const struct sched_attr *attr)
3309 int policy = attr->sched_policy;
3311 if (policy == SETPARAM_POLICY)
3316 if (dl_policy(policy))
3317 __setparam_dl(p, attr);
3318 else if (fair_policy(policy))
3319 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3322 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3323 * !rt_policy. Always setting this ensures that things like
3324 * getparam()/getattr() don't report silly values for !rt tasks.
3326 p->rt_priority = attr->sched_priority;
3327 p->normal_prio = normal_prio(p);
3331 /* Actually do priority change: must hold pi & rq lock. */
3332 static void __setscheduler(struct rq *rq, struct task_struct *p,
3333 const struct sched_attr *attr)
3335 __setscheduler_params(p, attr);
3338 * If we get here, there was no pi waiters boosting the
3339 * task. It is safe to use the normal prio.
3341 p->prio = normal_prio(p);
3343 if (dl_prio(p->prio))
3344 p->sched_class = &dl_sched_class;
3345 else if (rt_prio(p->prio))
3346 p->sched_class = &rt_sched_class;
3348 p->sched_class = &fair_sched_class;
3352 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3354 struct sched_dl_entity *dl_se = &p->dl;
3356 attr->sched_priority = p->rt_priority;
3357 attr->sched_runtime = dl_se->dl_runtime;
3358 attr->sched_deadline = dl_se->dl_deadline;
3359 attr->sched_period = dl_se->dl_period;
3360 attr->sched_flags = dl_se->flags;
3364 * This function validates the new parameters of a -deadline task.
3365 * We ask for the deadline not being zero, and greater or equal
3366 * than the runtime, as well as the period of being zero or
3367 * greater than deadline. Furthermore, we have to be sure that
3368 * user parameters are above the internal resolution of 1us (we
3369 * check sched_runtime only since it is always the smaller one) and
3370 * below 2^63 ns (we have to check both sched_deadline and
3371 * sched_period, as the latter can be zero).
3374 __checkparam_dl(const struct sched_attr *attr)
3377 if (attr->sched_deadline == 0)
3381 * Since we truncate DL_SCALE bits, make sure we're at least
3384 if (attr->sched_runtime < (1ULL << DL_SCALE))
3388 * Since we use the MSB for wrap-around and sign issues, make
3389 * sure it's not set (mind that period can be equal to zero).
3391 if (attr->sched_deadline & (1ULL << 63) ||
3392 attr->sched_period & (1ULL << 63))
3395 /* runtime <= deadline <= period (if period != 0) */
3396 if ((attr->sched_period != 0 &&
3397 attr->sched_period < attr->sched_deadline) ||
3398 attr->sched_deadline < attr->sched_runtime)
3405 * check the target process has a UID that matches the current process's
3407 static bool check_same_owner(struct task_struct *p)
3409 const struct cred *cred = current_cred(), *pcred;
3413 pcred = __task_cred(p);
3414 match = (uid_eq(cred->euid, pcred->euid) ||
3415 uid_eq(cred->euid, pcred->uid));
3420 static int __sched_setscheduler(struct task_struct *p,
3421 const struct sched_attr *attr,
3424 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3425 MAX_RT_PRIO - 1 - attr->sched_priority;
3426 int retval, oldprio, oldpolicy = -1, queued, running;
3427 int policy = attr->sched_policy;
3428 unsigned long flags;
3429 const struct sched_class *prev_class;
3433 /* may grab non-irq protected spin_locks */
3434 BUG_ON(in_interrupt());
3436 /* double check policy once rq lock held */
3438 reset_on_fork = p->sched_reset_on_fork;
3439 policy = oldpolicy = p->policy;
3441 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3443 if (policy != SCHED_DEADLINE &&
3444 policy != SCHED_FIFO && policy != SCHED_RR &&
3445 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3446 policy != SCHED_IDLE)
3450 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3454 * Valid priorities for SCHED_FIFO and SCHED_RR are
3455 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3456 * SCHED_BATCH and SCHED_IDLE is 0.
3458 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3459 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3461 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3462 (rt_policy(policy) != (attr->sched_priority != 0)))
3466 * Allow unprivileged RT tasks to decrease priority:
3468 if (user && !capable(CAP_SYS_NICE)) {
3469 if (fair_policy(policy)) {
3470 if (attr->sched_nice < task_nice(p) &&
3471 !can_nice(p, attr->sched_nice))
3475 if (rt_policy(policy)) {
3476 unsigned long rlim_rtprio =
3477 task_rlimit(p, RLIMIT_RTPRIO);
3479 /* can't set/change the rt policy */
3480 if (policy != p->policy && !rlim_rtprio)
3483 /* can't increase priority */
3484 if (attr->sched_priority > p->rt_priority &&
3485 attr->sched_priority > rlim_rtprio)
3490 * Can't set/change SCHED_DEADLINE policy at all for now
3491 * (safest behavior); in the future we would like to allow
3492 * unprivileged DL tasks to increase their relative deadline
3493 * or reduce their runtime (both ways reducing utilization)
3495 if (dl_policy(policy))
3499 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3500 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3502 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3503 if (!can_nice(p, task_nice(p)))
3507 /* can't change other user's priorities */
3508 if (!check_same_owner(p))
3511 /* Normal users shall not reset the sched_reset_on_fork flag */
3512 if (p->sched_reset_on_fork && !reset_on_fork)
3517 retval = security_task_setscheduler(p);
3523 * make sure no PI-waiters arrive (or leave) while we are
3524 * changing the priority of the task:
3526 * To be able to change p->policy safely, the appropriate
3527 * runqueue lock must be held.
3529 rq = task_rq_lock(p, &flags);
3532 * Changing the policy of the stop threads its a very bad idea
3534 if (p == rq->stop) {
3535 task_rq_unlock(rq, p, &flags);
3540 * If not changing anything there's no need to proceed further,
3541 * but store a possible modification of reset_on_fork.
3543 if (unlikely(policy == p->policy)) {
3544 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3546 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3548 if (dl_policy(policy))
3551 p->sched_reset_on_fork = reset_on_fork;
3552 task_rq_unlock(rq, p, &flags);
3558 #ifdef CONFIG_RT_GROUP_SCHED
3560 * Do not allow realtime tasks into groups that have no runtime
3563 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3564 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3565 !task_group_is_autogroup(task_group(p))) {
3566 task_rq_unlock(rq, p, &flags);
3571 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3572 cpumask_t *span = rq->rd->span;
3575 * Don't allow tasks with an affinity mask smaller than
3576 * the entire root_domain to become SCHED_DEADLINE. We
3577 * will also fail if there's no bandwidth available.
3579 if (!cpumask_subset(span, &p->cpus_allowed) ||
3580 rq->rd->dl_bw.bw == 0) {
3581 task_rq_unlock(rq, p, &flags);
3588 /* recheck policy now with rq lock held */
3589 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3590 policy = oldpolicy = -1;
3591 task_rq_unlock(rq, p, &flags);
3596 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3597 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3600 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3601 task_rq_unlock(rq, p, &flags);
3605 p->sched_reset_on_fork = reset_on_fork;
3609 * Special case for priority boosted tasks.
3611 * If the new priority is lower or equal (user space view)
3612 * than the current (boosted) priority, we just store the new
3613 * normal parameters and do not touch the scheduler class and
3614 * the runqueue. This will be done when the task deboost
3617 if (rt_mutex_check_prio(p, newprio)) {
3618 __setscheduler_params(p, attr);
3619 task_rq_unlock(rq, p, &flags);
3623 queued = task_on_rq_queued(p);
3624 running = task_current(rq, p);
3626 dequeue_task(rq, p, 0);
3628 put_prev_task(rq, p);
3630 prev_class = p->sched_class;
3631 __setscheduler(rq, p, attr);
3634 p->sched_class->set_curr_task(rq);
3637 * We enqueue to tail when the priority of a task is
3638 * increased (user space view).
3640 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3643 check_class_changed(rq, p, prev_class, oldprio);
3644 task_rq_unlock(rq, p, &flags);
3646 rt_mutex_adjust_pi(p);
3651 static int _sched_setscheduler(struct task_struct *p, int policy,
3652 const struct sched_param *param, bool check)
3654 struct sched_attr attr = {
3655 .sched_policy = policy,
3656 .sched_priority = param->sched_priority,
3657 .sched_nice = PRIO_TO_NICE(p->static_prio),
3660 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3661 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3662 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3663 policy &= ~SCHED_RESET_ON_FORK;
3664 attr.sched_policy = policy;
3667 return __sched_setscheduler(p, &attr, check);
3670 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3671 * @p: the task in question.
3672 * @policy: new policy.
3673 * @param: structure containing the new RT priority.
3675 * Return: 0 on success. An error code otherwise.
3677 * NOTE that the task may be already dead.
3679 int sched_setscheduler(struct task_struct *p, int policy,
3680 const struct sched_param *param)
3682 return _sched_setscheduler(p, policy, param, true);
3684 EXPORT_SYMBOL_GPL(sched_setscheduler);
3686 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3688 return __sched_setscheduler(p, attr, true);
3690 EXPORT_SYMBOL_GPL(sched_setattr);
3693 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3694 * @p: the task in question.
3695 * @policy: new policy.
3696 * @param: structure containing the new RT priority.
3698 * Just like sched_setscheduler, only don't bother checking if the
3699 * current context has permission. For example, this is needed in
3700 * stop_machine(): we create temporary high priority worker threads,
3701 * but our caller might not have that capability.
3703 * Return: 0 on success. An error code otherwise.
3705 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3706 const struct sched_param *param)
3708 return _sched_setscheduler(p, policy, param, false);
3712 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3714 struct sched_param lparam;
3715 struct task_struct *p;
3718 if (!param || pid < 0)
3720 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3725 p = find_process_by_pid(pid);
3727 retval = sched_setscheduler(p, policy, &lparam);
3734 * Mimics kernel/events/core.c perf_copy_attr().
3736 static int sched_copy_attr(struct sched_attr __user *uattr,
3737 struct sched_attr *attr)
3742 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3746 * zero the full structure, so that a short copy will be nice.
3748 memset(attr, 0, sizeof(*attr));
3750 ret = get_user(size, &uattr->size);
3754 if (size > PAGE_SIZE) /* silly large */
3757 if (!size) /* abi compat */
3758 size = SCHED_ATTR_SIZE_VER0;
3760 if (size < SCHED_ATTR_SIZE_VER0)
3764 * If we're handed a bigger struct than we know of,
3765 * ensure all the unknown bits are 0 - i.e. new
3766 * user-space does not rely on any kernel feature
3767 * extensions we dont know about yet.
3769 if (size > sizeof(*attr)) {
3770 unsigned char __user *addr;
3771 unsigned char __user *end;
3774 addr = (void __user *)uattr + sizeof(*attr);
3775 end = (void __user *)uattr + size;
3777 for (; addr < end; addr++) {
3778 ret = get_user(val, addr);
3784 size = sizeof(*attr);
3787 ret = copy_from_user(attr, uattr, size);
3792 * XXX: do we want to be lenient like existing syscalls; or do we want
3793 * to be strict and return an error on out-of-bounds values?
3795 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3800 put_user(sizeof(*attr), &uattr->size);
3805 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3806 * @pid: the pid in question.
3807 * @policy: new policy.
3808 * @param: structure containing the new RT priority.
3810 * Return: 0 on success. An error code otherwise.
3812 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3813 struct sched_param __user *, param)
3815 /* negative values for policy are not valid */
3819 return do_sched_setscheduler(pid, policy, param);
3823 * sys_sched_setparam - set/change the RT priority of a thread
3824 * @pid: the pid in question.
3825 * @param: structure containing the new RT priority.
3827 * Return: 0 on success. An error code otherwise.
3829 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3831 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3835 * sys_sched_setattr - same as above, but with extended sched_attr
3836 * @pid: the pid in question.
3837 * @uattr: structure containing the extended parameters.
3838 * @flags: for future extension.
3840 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3841 unsigned int, flags)
3843 struct sched_attr attr;
3844 struct task_struct *p;
3847 if (!uattr || pid < 0 || flags)
3850 retval = sched_copy_attr(uattr, &attr);
3854 if ((int)attr.sched_policy < 0)
3859 p = find_process_by_pid(pid);
3861 retval = sched_setattr(p, &attr);
3868 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3869 * @pid: the pid in question.
3871 * Return: On success, the policy of the thread. Otherwise, a negative error
3874 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3876 struct task_struct *p;
3884 p = find_process_by_pid(pid);
3886 retval = security_task_getscheduler(p);
3889 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3896 * sys_sched_getparam - get the RT priority of a thread
3897 * @pid: the pid in question.
3898 * @param: structure containing the RT priority.
3900 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3903 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3905 struct sched_param lp = { .sched_priority = 0 };
3906 struct task_struct *p;
3909 if (!param || pid < 0)
3913 p = find_process_by_pid(pid);
3918 retval = security_task_getscheduler(p);
3922 if (task_has_rt_policy(p))
3923 lp.sched_priority = p->rt_priority;
3927 * This one might sleep, we cannot do it with a spinlock held ...
3929 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3938 static int sched_read_attr(struct sched_attr __user *uattr,
3939 struct sched_attr *attr,
3944 if (!access_ok(VERIFY_WRITE, uattr, usize))
3948 * If we're handed a smaller struct than we know of,
3949 * ensure all the unknown bits are 0 - i.e. old
3950 * user-space does not get uncomplete information.
3952 if (usize < sizeof(*attr)) {
3953 unsigned char *addr;
3956 addr = (void *)attr + usize;
3957 end = (void *)attr + sizeof(*attr);
3959 for (; addr < end; addr++) {
3967 ret = copy_to_user(uattr, attr, attr->size);
3975 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3976 * @pid: the pid in question.
3977 * @uattr: structure containing the extended parameters.
3978 * @size: sizeof(attr) for fwd/bwd comp.
3979 * @flags: for future extension.
3981 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3982 unsigned int, size, unsigned int, flags)
3984 struct sched_attr attr = {
3985 .size = sizeof(struct sched_attr),
3987 struct task_struct *p;
3990 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3991 size < SCHED_ATTR_SIZE_VER0 || flags)
3995 p = find_process_by_pid(pid);
4000 retval = security_task_getscheduler(p);
4004 attr.sched_policy = p->policy;
4005 if (p->sched_reset_on_fork)
4006 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4007 if (task_has_dl_policy(p))
4008 __getparam_dl(p, &attr);
4009 else if (task_has_rt_policy(p))
4010 attr.sched_priority = p->rt_priority;
4012 attr.sched_nice = task_nice(p);
4016 retval = sched_read_attr(uattr, &attr, size);
4024 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4026 cpumask_var_t cpus_allowed, new_mask;
4027 struct task_struct *p;
4032 p = find_process_by_pid(pid);
4038 /* Prevent p going away */
4042 if (p->flags & PF_NO_SETAFFINITY) {
4046 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4050 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4052 goto out_free_cpus_allowed;
4055 if (!check_same_owner(p)) {
4057 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4059 goto out_free_new_mask;
4064 retval = security_task_setscheduler(p);
4066 goto out_free_new_mask;
4069 cpuset_cpus_allowed(p, cpus_allowed);
4070 cpumask_and(new_mask, in_mask, cpus_allowed);
4073 * Since bandwidth control happens on root_domain basis,
4074 * if admission test is enabled, we only admit -deadline
4075 * tasks allowed to run on all the CPUs in the task's
4079 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4081 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4084 goto out_free_new_mask;
4090 retval = set_cpus_allowed_ptr(p, new_mask);
4093 cpuset_cpus_allowed(p, cpus_allowed);
4094 if (!cpumask_subset(new_mask, cpus_allowed)) {
4096 * We must have raced with a concurrent cpuset
4097 * update. Just reset the cpus_allowed to the
4098 * cpuset's cpus_allowed
4100 cpumask_copy(new_mask, cpus_allowed);
4105 free_cpumask_var(new_mask);
4106 out_free_cpus_allowed:
4107 free_cpumask_var(cpus_allowed);
4113 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4114 struct cpumask *new_mask)
4116 if (len < cpumask_size())
4117 cpumask_clear(new_mask);
4118 else if (len > cpumask_size())
4119 len = cpumask_size();
4121 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4125 * sys_sched_setaffinity - set the cpu affinity of a process
4126 * @pid: pid of the process
4127 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4128 * @user_mask_ptr: user-space pointer to the new cpu mask
4130 * Return: 0 on success. An error code otherwise.
4132 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4133 unsigned long __user *, user_mask_ptr)
4135 cpumask_var_t new_mask;
4138 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4141 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4143 retval = sched_setaffinity(pid, new_mask);
4144 free_cpumask_var(new_mask);
4148 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4150 struct task_struct *p;
4151 unsigned long flags;
4157 p = find_process_by_pid(pid);
4161 retval = security_task_getscheduler(p);
4165 raw_spin_lock_irqsave(&p->pi_lock, flags);
4166 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4167 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4176 * sys_sched_getaffinity - get the cpu affinity of a process
4177 * @pid: pid of the process
4178 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4179 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4181 * Return: 0 on success. An error code otherwise.
4183 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4184 unsigned long __user *, user_mask_ptr)
4189 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4191 if (len & (sizeof(unsigned long)-1))
4194 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4197 ret = sched_getaffinity(pid, mask);
4199 size_t retlen = min_t(size_t, len, cpumask_size());
4201 if (copy_to_user(user_mask_ptr, mask, retlen))
4206 free_cpumask_var(mask);
4212 * sys_sched_yield - yield the current processor to other threads.
4214 * This function yields the current CPU to other tasks. If there are no
4215 * other threads running on this CPU then this function will return.
4219 SYSCALL_DEFINE0(sched_yield)
4221 struct rq *rq = this_rq_lock();
4223 schedstat_inc(rq, yld_count);
4224 current->sched_class->yield_task(rq);
4227 * Since we are going to call schedule() anyway, there's
4228 * no need to preempt or enable interrupts:
4230 __release(rq->lock);
4231 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4232 do_raw_spin_unlock(&rq->lock);
4233 sched_preempt_enable_no_resched();
4240 int __sched _cond_resched(void)
4242 if (should_resched()) {
4243 preempt_schedule_common();
4248 EXPORT_SYMBOL(_cond_resched);
4251 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4252 * call schedule, and on return reacquire the lock.
4254 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4255 * operations here to prevent schedule() from being called twice (once via
4256 * spin_unlock(), once by hand).
4258 int __cond_resched_lock(spinlock_t *lock)
4260 int resched = should_resched();
4263 lockdep_assert_held(lock);
4265 if (spin_needbreak(lock) || resched) {
4268 preempt_schedule_common();
4276 EXPORT_SYMBOL(__cond_resched_lock);
4278 int __sched __cond_resched_softirq(void)
4280 BUG_ON(!in_softirq());
4282 if (should_resched()) {
4284 preempt_schedule_common();
4290 EXPORT_SYMBOL(__cond_resched_softirq);
4293 * yield - yield the current processor to other threads.
4295 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4297 * The scheduler is at all times free to pick the calling task as the most
4298 * eligible task to run, if removing the yield() call from your code breaks
4299 * it, its already broken.
4301 * Typical broken usage is:
4306 * where one assumes that yield() will let 'the other' process run that will
4307 * make event true. If the current task is a SCHED_FIFO task that will never
4308 * happen. Never use yield() as a progress guarantee!!
4310 * If you want to use yield() to wait for something, use wait_event().
4311 * If you want to use yield() to be 'nice' for others, use cond_resched().
4312 * If you still want to use yield(), do not!
4314 void __sched yield(void)
4316 set_current_state(TASK_RUNNING);
4319 EXPORT_SYMBOL(yield);
4322 * yield_to - yield the current processor to another thread in
4323 * your thread group, or accelerate that thread toward the
4324 * processor it's on.
4326 * @preempt: whether task preemption is allowed or not
4328 * It's the caller's job to ensure that the target task struct
4329 * can't go away on us before we can do any checks.
4332 * true (>0) if we indeed boosted the target task.
4333 * false (0) if we failed to boost the target.
4334 * -ESRCH if there's no task to yield to.
4336 int __sched yield_to(struct task_struct *p, bool preempt)
4338 struct task_struct *curr = current;
4339 struct rq *rq, *p_rq;
4340 unsigned long flags;
4343 local_irq_save(flags);
4349 * If we're the only runnable task on the rq and target rq also
4350 * has only one task, there's absolutely no point in yielding.
4352 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4357 double_rq_lock(rq, p_rq);
4358 if (task_rq(p) != p_rq) {
4359 double_rq_unlock(rq, p_rq);
4363 if (!curr->sched_class->yield_to_task)
4366 if (curr->sched_class != p->sched_class)
4369 if (task_running(p_rq, p) || p->state)
4372 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4374 schedstat_inc(rq, yld_count);
4376 * Make p's CPU reschedule; pick_next_entity takes care of
4379 if (preempt && rq != p_rq)
4384 double_rq_unlock(rq, p_rq);
4386 local_irq_restore(flags);
4393 EXPORT_SYMBOL_GPL(yield_to);
4396 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4397 * that process accounting knows that this is a task in IO wait state.
4399 void __sched io_schedule(void)
4401 struct rq *rq = raw_rq();
4403 delayacct_blkio_start();
4404 atomic_inc(&rq->nr_iowait);
4405 blk_flush_plug(current);
4406 current->in_iowait = 1;
4408 current->in_iowait = 0;
4409 atomic_dec(&rq->nr_iowait);
4410 delayacct_blkio_end();
4412 EXPORT_SYMBOL(io_schedule);
4414 long __sched io_schedule_timeout(long timeout)
4416 struct rq *rq = raw_rq();
4419 delayacct_blkio_start();
4420 atomic_inc(&rq->nr_iowait);
4421 blk_flush_plug(current);
4422 current->in_iowait = 1;
4423 ret = schedule_timeout(timeout);
4424 current->in_iowait = 0;
4425 atomic_dec(&rq->nr_iowait);
4426 delayacct_blkio_end();
4431 * sys_sched_get_priority_max - return maximum RT priority.
4432 * @policy: scheduling class.
4434 * Return: On success, this syscall returns the maximum
4435 * rt_priority that can be used by a given scheduling class.
4436 * On failure, a negative error code is returned.
4438 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4445 ret = MAX_USER_RT_PRIO-1;
4447 case SCHED_DEADLINE:
4458 * sys_sched_get_priority_min - return minimum RT priority.
4459 * @policy: scheduling class.
4461 * Return: On success, this syscall returns the minimum
4462 * rt_priority that can be used by a given scheduling class.
4463 * On failure, a negative error code is returned.
4465 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4474 case SCHED_DEADLINE:
4484 * sys_sched_rr_get_interval - return the default timeslice of a process.
4485 * @pid: pid of the process.
4486 * @interval: userspace pointer to the timeslice value.
4488 * this syscall writes the default timeslice value of a given process
4489 * into the user-space timespec buffer. A value of '0' means infinity.
4491 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4494 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4495 struct timespec __user *, interval)
4497 struct task_struct *p;
4498 unsigned int time_slice;
4499 unsigned long flags;
4509 p = find_process_by_pid(pid);
4513 retval = security_task_getscheduler(p);
4517 rq = task_rq_lock(p, &flags);
4519 if (p->sched_class->get_rr_interval)
4520 time_slice = p->sched_class->get_rr_interval(rq, p);
4521 task_rq_unlock(rq, p, &flags);
4524 jiffies_to_timespec(time_slice, &t);
4525 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4533 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4535 void sched_show_task(struct task_struct *p)
4537 unsigned long free = 0;
4539 unsigned long state = p->state;
4542 state = __ffs(state) + 1;
4543 printk(KERN_INFO "%-15.15s %c", p->comm,
4544 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4545 #if BITS_PER_LONG == 32
4546 if (state == TASK_RUNNING)
4547 printk(KERN_CONT " running ");
4549 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4551 if (state == TASK_RUNNING)
4552 printk(KERN_CONT " running task ");
4554 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4556 #ifdef CONFIG_DEBUG_STACK_USAGE
4557 free = stack_not_used(p);
4562 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4564 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4565 task_pid_nr(p), ppid,
4566 (unsigned long)task_thread_info(p)->flags);
4568 print_worker_info(KERN_INFO, p);
4569 show_stack(p, NULL);
4572 void show_state_filter(unsigned long state_filter)
4574 struct task_struct *g, *p;
4576 #if BITS_PER_LONG == 32
4578 " task PC stack pid father\n");
4581 " task PC stack pid father\n");
4584 for_each_process_thread(g, p) {
4586 * reset the NMI-timeout, listing all files on a slow
4587 * console might take a lot of time:
4589 touch_nmi_watchdog();
4590 if (!state_filter || (p->state & state_filter))
4594 touch_all_softlockup_watchdogs();
4596 #ifdef CONFIG_SCHED_DEBUG
4597 sysrq_sched_debug_show();
4601 * Only show locks if all tasks are dumped:
4604 debug_show_all_locks();
4607 void init_idle_bootup_task(struct task_struct *idle)
4609 idle->sched_class = &idle_sched_class;
4613 * init_idle - set up an idle thread for a given CPU
4614 * @idle: task in question
4615 * @cpu: cpu the idle task belongs to
4617 * NOTE: this function does not set the idle thread's NEED_RESCHED
4618 * flag, to make booting more robust.
4620 void init_idle(struct task_struct *idle, int cpu)
4622 struct rq *rq = cpu_rq(cpu);
4623 unsigned long flags;
4625 raw_spin_lock_irqsave(&rq->lock, flags);
4627 __sched_fork(0, idle);
4628 idle->state = TASK_RUNNING;
4629 idle->se.exec_start = sched_clock();
4631 do_set_cpus_allowed(idle, cpumask_of(cpu));
4633 * We're having a chicken and egg problem, even though we are
4634 * holding rq->lock, the cpu isn't yet set to this cpu so the
4635 * lockdep check in task_group() will fail.
4637 * Similar case to sched_fork(). / Alternatively we could
4638 * use task_rq_lock() here and obtain the other rq->lock.
4643 __set_task_cpu(idle, cpu);
4646 rq->curr = rq->idle = idle;
4647 idle->on_rq = TASK_ON_RQ_QUEUED;
4648 #if defined(CONFIG_SMP)
4651 raw_spin_unlock_irqrestore(&rq->lock, flags);
4653 /* Set the preempt count _outside_ the spinlocks! */
4654 init_idle_preempt_count(idle, cpu);
4657 * The idle tasks have their own, simple scheduling class:
4659 idle->sched_class = &idle_sched_class;
4660 ftrace_graph_init_idle_task(idle, cpu);
4661 vtime_init_idle(idle, cpu);
4662 #if defined(CONFIG_SMP)
4663 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4667 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4668 const struct cpumask *trial)
4670 int ret = 1, trial_cpus;
4671 struct dl_bw *cur_dl_b;
4672 unsigned long flags;
4674 if (!cpumask_weight(cur))
4677 rcu_read_lock_sched();
4678 cur_dl_b = dl_bw_of(cpumask_any(cur));
4679 trial_cpus = cpumask_weight(trial);
4681 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4682 if (cur_dl_b->bw != -1 &&
4683 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4685 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4686 rcu_read_unlock_sched();
4691 int task_can_attach(struct task_struct *p,
4692 const struct cpumask *cs_cpus_allowed)
4697 * Kthreads which disallow setaffinity shouldn't be moved
4698 * to a new cpuset; we don't want to change their cpu
4699 * affinity and isolating such threads by their set of
4700 * allowed nodes is unnecessary. Thus, cpusets are not
4701 * applicable for such threads. This prevents checking for
4702 * success of set_cpus_allowed_ptr() on all attached tasks
4703 * before cpus_allowed may be changed.
4705 if (p->flags & PF_NO_SETAFFINITY) {
4711 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4713 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4718 unsigned long flags;
4720 rcu_read_lock_sched();
4721 dl_b = dl_bw_of(dest_cpu);
4722 raw_spin_lock_irqsave(&dl_b->lock, flags);
4723 cpus = dl_bw_cpus(dest_cpu);
4724 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4729 * We reserve space for this task in the destination
4730 * root_domain, as we can't fail after this point.
4731 * We will free resources in the source root_domain
4732 * later on (see set_cpus_allowed_dl()).
4734 __dl_add(dl_b, p->dl.dl_bw);
4736 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4737 rcu_read_unlock_sched();
4747 * move_queued_task - move a queued task to new rq.
4749 * Returns (locked) new rq. Old rq's lock is released.
4751 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4753 struct rq *rq = task_rq(p);
4755 lockdep_assert_held(&rq->lock);
4757 dequeue_task(rq, p, 0);
4758 p->on_rq = TASK_ON_RQ_MIGRATING;
4759 set_task_cpu(p, new_cpu);
4760 raw_spin_unlock(&rq->lock);
4762 rq = cpu_rq(new_cpu);
4764 raw_spin_lock(&rq->lock);
4765 BUG_ON(task_cpu(p) != new_cpu);
4766 p->on_rq = TASK_ON_RQ_QUEUED;
4767 enqueue_task(rq, p, 0);
4768 check_preempt_curr(rq, p, 0);
4773 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4775 if (p->sched_class->set_cpus_allowed)
4776 p->sched_class->set_cpus_allowed(p, new_mask);
4778 cpumask_copy(&p->cpus_allowed, new_mask);
4779 p->nr_cpus_allowed = cpumask_weight(new_mask);
4783 * This is how migration works:
4785 * 1) we invoke migration_cpu_stop() on the target CPU using
4787 * 2) stopper starts to run (implicitly forcing the migrated thread
4789 * 3) it checks whether the migrated task is still in the wrong runqueue.
4790 * 4) if it's in the wrong runqueue then the migration thread removes
4791 * it and puts it into the right queue.
4792 * 5) stopper completes and stop_one_cpu() returns and the migration
4797 * Change a given task's CPU affinity. Migrate the thread to a
4798 * proper CPU and schedule it away if the CPU it's executing on
4799 * is removed from the allowed bitmask.
4801 * NOTE: the caller must have a valid reference to the task, the
4802 * task must not exit() & deallocate itself prematurely. The
4803 * call is not atomic; no spinlocks may be held.
4805 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4807 unsigned long flags;
4809 unsigned int dest_cpu;
4812 rq = task_rq_lock(p, &flags);
4814 if (cpumask_equal(&p->cpus_allowed, new_mask))
4817 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4822 do_set_cpus_allowed(p, new_mask);
4824 /* Can the task run on the task's current CPU? If so, we're done */
4825 if (cpumask_test_cpu(task_cpu(p), new_mask))
4828 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4829 if (task_running(rq, p) || p->state == TASK_WAKING) {
4830 struct migration_arg arg = { p, dest_cpu };
4831 /* Need help from migration thread: drop lock and wait. */
4832 task_rq_unlock(rq, p, &flags);
4833 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4834 tlb_migrate_finish(p->mm);
4836 } else if (task_on_rq_queued(p))
4837 rq = move_queued_task(p, dest_cpu);
4839 task_rq_unlock(rq, p, &flags);
4843 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4846 * Move (not current) task off this cpu, onto dest cpu. We're doing
4847 * this because either it can't run here any more (set_cpus_allowed()
4848 * away from this CPU, or CPU going down), or because we're
4849 * attempting to rebalance this task on exec (sched_exec).
4851 * So we race with normal scheduler movements, but that's OK, as long
4852 * as the task is no longer on this CPU.
4854 * Returns non-zero if task was successfully migrated.
4856 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4861 if (unlikely(!cpu_active(dest_cpu)))
4864 rq = cpu_rq(src_cpu);
4866 raw_spin_lock(&p->pi_lock);
4867 raw_spin_lock(&rq->lock);
4868 /* Already moved. */
4869 if (task_cpu(p) != src_cpu)
4872 /* Affinity changed (again). */
4873 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4877 * If we're not on a rq, the next wake-up will ensure we're
4880 if (task_on_rq_queued(p))
4881 rq = move_queued_task(p, dest_cpu);
4885 raw_spin_unlock(&rq->lock);
4886 raw_spin_unlock(&p->pi_lock);
4890 #ifdef CONFIG_NUMA_BALANCING
4891 /* Migrate current task p to target_cpu */
4892 int migrate_task_to(struct task_struct *p, int target_cpu)
4894 struct migration_arg arg = { p, target_cpu };
4895 int curr_cpu = task_cpu(p);
4897 if (curr_cpu == target_cpu)
4900 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4903 /* TODO: This is not properly updating schedstats */
4905 trace_sched_move_numa(p, curr_cpu, target_cpu);
4906 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4910 * Requeue a task on a given node and accurately track the number of NUMA
4911 * tasks on the runqueues
4913 void sched_setnuma(struct task_struct *p, int nid)
4916 unsigned long flags;
4917 bool queued, running;
4919 rq = task_rq_lock(p, &flags);
4920 queued = task_on_rq_queued(p);
4921 running = task_current(rq, p);
4924 dequeue_task(rq, p, 0);
4926 put_prev_task(rq, p);
4928 p->numa_preferred_nid = nid;
4931 p->sched_class->set_curr_task(rq);
4933 enqueue_task(rq, p, 0);
4934 task_rq_unlock(rq, p, &flags);
4939 * migration_cpu_stop - this will be executed by a highprio stopper thread
4940 * and performs thread migration by bumping thread off CPU then
4941 * 'pushing' onto another runqueue.
4943 static int migration_cpu_stop(void *data)
4945 struct migration_arg *arg = data;
4948 * The original target cpu might have gone down and we might
4949 * be on another cpu but it doesn't matter.
4951 local_irq_disable();
4953 * We need to explicitly wake pending tasks before running
4954 * __migrate_task() such that we will not miss enforcing cpus_allowed
4955 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4957 sched_ttwu_pending();
4958 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4963 #ifdef CONFIG_HOTPLUG_CPU
4966 * Ensures that the idle task is using init_mm right before its cpu goes
4969 void idle_task_exit(void)
4971 struct mm_struct *mm = current->active_mm;
4973 BUG_ON(cpu_online(smp_processor_id()));
4975 if (mm != &init_mm) {
4976 switch_mm(mm, &init_mm, current);
4977 finish_arch_post_lock_switch();
4983 * Since this CPU is going 'away' for a while, fold any nr_active delta
4984 * we might have. Assumes we're called after migrate_tasks() so that the
4985 * nr_active count is stable.
4987 * Also see the comment "Global load-average calculations".
4989 static void calc_load_migrate(struct rq *rq)
4991 long delta = calc_load_fold_active(rq);
4993 atomic_long_add(delta, &calc_load_tasks);
4996 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5000 static const struct sched_class fake_sched_class = {
5001 .put_prev_task = put_prev_task_fake,
5004 static struct task_struct fake_task = {
5006 * Avoid pull_{rt,dl}_task()
5008 .prio = MAX_PRIO + 1,
5009 .sched_class = &fake_sched_class,
5013 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5014 * try_to_wake_up()->select_task_rq().
5016 * Called with rq->lock held even though we'er in stop_machine() and
5017 * there's no concurrency possible, we hold the required locks anyway
5018 * because of lock validation efforts.
5020 static void migrate_tasks(unsigned int dead_cpu)
5022 struct rq *rq = cpu_rq(dead_cpu);
5023 struct task_struct *next, *stop = rq->stop;
5027 * Fudge the rq selection such that the below task selection loop
5028 * doesn't get stuck on the currently eligible stop task.
5030 * We're currently inside stop_machine() and the rq is either stuck
5031 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5032 * either way we should never end up calling schedule() until we're
5038 * put_prev_task() and pick_next_task() sched
5039 * class method both need to have an up-to-date
5040 * value of rq->clock[_task]
5042 update_rq_clock(rq);
5046 * There's this thread running, bail when that's the only
5049 if (rq->nr_running == 1)
5052 next = pick_next_task(rq, &fake_task);
5054 next->sched_class->put_prev_task(rq, next);
5056 /* Find suitable destination for @next, with force if needed. */
5057 dest_cpu = select_fallback_rq(dead_cpu, next);
5058 raw_spin_unlock(&rq->lock);
5060 __migrate_task(next, dead_cpu, dest_cpu);
5062 raw_spin_lock(&rq->lock);
5068 #endif /* CONFIG_HOTPLUG_CPU */
5070 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5072 static struct ctl_table sd_ctl_dir[] = {
5074 .procname = "sched_domain",
5080 static struct ctl_table sd_ctl_root[] = {
5082 .procname = "kernel",
5084 .child = sd_ctl_dir,
5089 static struct ctl_table *sd_alloc_ctl_entry(int n)
5091 struct ctl_table *entry =
5092 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5097 static void sd_free_ctl_entry(struct ctl_table **tablep)
5099 struct ctl_table *entry;
5102 * In the intermediate directories, both the child directory and
5103 * procname are dynamically allocated and could fail but the mode
5104 * will always be set. In the lowest directory the names are
5105 * static strings and all have proc handlers.
5107 for (entry = *tablep; entry->mode; entry++) {
5109 sd_free_ctl_entry(&entry->child);
5110 if (entry->proc_handler == NULL)
5111 kfree(entry->procname);
5118 static int min_load_idx = 0;
5119 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5122 set_table_entry(struct ctl_table *entry,
5123 const char *procname, void *data, int maxlen,
5124 umode_t mode, proc_handler *proc_handler,
5127 entry->procname = procname;
5129 entry->maxlen = maxlen;
5131 entry->proc_handler = proc_handler;
5134 entry->extra1 = &min_load_idx;
5135 entry->extra2 = &max_load_idx;
5139 static struct ctl_table *
5140 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5142 struct ctl_table *table = sd_alloc_ctl_entry(14);
5147 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5148 sizeof(long), 0644, proc_doulongvec_minmax, false);
5149 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5150 sizeof(long), 0644, proc_doulongvec_minmax, false);
5151 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5152 sizeof(int), 0644, proc_dointvec_minmax, true);
5153 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5154 sizeof(int), 0644, proc_dointvec_minmax, true);
5155 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5156 sizeof(int), 0644, proc_dointvec_minmax, true);
5157 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5158 sizeof(int), 0644, proc_dointvec_minmax, true);
5159 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5160 sizeof(int), 0644, proc_dointvec_minmax, true);
5161 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5162 sizeof(int), 0644, proc_dointvec_minmax, false);
5163 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5164 sizeof(int), 0644, proc_dointvec_minmax, false);
5165 set_table_entry(&table[9], "cache_nice_tries",
5166 &sd->cache_nice_tries,
5167 sizeof(int), 0644, proc_dointvec_minmax, false);
5168 set_table_entry(&table[10], "flags", &sd->flags,
5169 sizeof(int), 0644, proc_dointvec_minmax, false);
5170 set_table_entry(&table[11], "max_newidle_lb_cost",
5171 &sd->max_newidle_lb_cost,
5172 sizeof(long), 0644, proc_doulongvec_minmax, false);
5173 set_table_entry(&table[12], "name", sd->name,
5174 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5175 /* &table[13] is terminator */
5180 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5182 struct ctl_table *entry, *table;
5183 struct sched_domain *sd;
5184 int domain_num = 0, i;
5187 for_each_domain(cpu, sd)
5189 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5194 for_each_domain(cpu, sd) {
5195 snprintf(buf, 32, "domain%d", i);
5196 entry->procname = kstrdup(buf, GFP_KERNEL);
5198 entry->child = sd_alloc_ctl_domain_table(sd);
5205 static struct ctl_table_header *sd_sysctl_header;
5206 static void register_sched_domain_sysctl(void)
5208 int i, cpu_num = num_possible_cpus();
5209 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5212 WARN_ON(sd_ctl_dir[0].child);
5213 sd_ctl_dir[0].child = entry;
5218 for_each_possible_cpu(i) {
5219 snprintf(buf, 32, "cpu%d", i);
5220 entry->procname = kstrdup(buf, GFP_KERNEL);
5222 entry->child = sd_alloc_ctl_cpu_table(i);
5226 WARN_ON(sd_sysctl_header);
5227 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5230 /* may be called multiple times per register */
5231 static void unregister_sched_domain_sysctl(void)
5233 if (sd_sysctl_header)
5234 unregister_sysctl_table(sd_sysctl_header);
5235 sd_sysctl_header = NULL;
5236 if (sd_ctl_dir[0].child)
5237 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5240 static void register_sched_domain_sysctl(void)
5243 static void unregister_sched_domain_sysctl(void)
5248 static void set_rq_online(struct rq *rq)
5251 const struct sched_class *class;
5253 cpumask_set_cpu(rq->cpu, rq->rd->online);
5256 for_each_class(class) {
5257 if (class->rq_online)
5258 class->rq_online(rq);
5263 static void set_rq_offline(struct rq *rq)
5266 const struct sched_class *class;
5268 for_each_class(class) {
5269 if (class->rq_offline)
5270 class->rq_offline(rq);
5273 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5279 * migration_call - callback that gets triggered when a CPU is added.
5280 * Here we can start up the necessary migration thread for the new CPU.
5283 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5285 int cpu = (long)hcpu;
5286 unsigned long flags;
5287 struct rq *rq = cpu_rq(cpu);
5289 switch (action & ~CPU_TASKS_FROZEN) {
5291 case CPU_UP_PREPARE:
5292 rq->calc_load_update = calc_load_update;
5296 /* Update our root-domain */
5297 raw_spin_lock_irqsave(&rq->lock, flags);
5299 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5303 raw_spin_unlock_irqrestore(&rq->lock, flags);
5306 #ifdef CONFIG_HOTPLUG_CPU
5308 sched_ttwu_pending();
5309 /* Update our root-domain */
5310 raw_spin_lock_irqsave(&rq->lock, flags);
5312 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5316 BUG_ON(rq->nr_running != 1); /* the migration thread */
5317 raw_spin_unlock_irqrestore(&rq->lock, flags);
5321 calc_load_migrate(rq);
5326 update_max_interval();
5332 * Register at high priority so that task migration (migrate_all_tasks)
5333 * happens before everything else. This has to be lower priority than
5334 * the notifier in the perf_event subsystem, though.
5336 static struct notifier_block migration_notifier = {
5337 .notifier_call = migration_call,
5338 .priority = CPU_PRI_MIGRATION,
5341 static void __cpuinit set_cpu_rq_start_time(void)
5343 int cpu = smp_processor_id();
5344 struct rq *rq = cpu_rq(cpu);
5345 rq->age_stamp = sched_clock_cpu(cpu);
5348 static int sched_cpu_active(struct notifier_block *nfb,
5349 unsigned long action, void *hcpu)
5351 switch (action & ~CPU_TASKS_FROZEN) {
5353 set_cpu_rq_start_time();
5355 case CPU_DOWN_FAILED:
5356 set_cpu_active((long)hcpu, true);
5363 static int sched_cpu_inactive(struct notifier_block *nfb,
5364 unsigned long action, void *hcpu)
5366 unsigned long flags;
5367 long cpu = (long)hcpu;
5370 switch (action & ~CPU_TASKS_FROZEN) {
5371 case CPU_DOWN_PREPARE:
5372 set_cpu_active(cpu, false);
5374 /* explicitly allow suspend */
5375 if (!(action & CPU_TASKS_FROZEN)) {
5379 rcu_read_lock_sched();
5380 dl_b = dl_bw_of(cpu);
5382 raw_spin_lock_irqsave(&dl_b->lock, flags);
5383 cpus = dl_bw_cpus(cpu);
5384 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5385 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5387 rcu_read_unlock_sched();
5390 return notifier_from_errno(-EBUSY);
5398 static int __init migration_init(void)
5400 void *cpu = (void *)(long)smp_processor_id();
5403 /* Initialize migration for the boot CPU */
5404 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5405 BUG_ON(err == NOTIFY_BAD);
5406 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5407 register_cpu_notifier(&migration_notifier);
5409 /* Register cpu active notifiers */
5410 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5411 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5415 early_initcall(migration_init);
5420 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5422 #ifdef CONFIG_SCHED_DEBUG
5424 static __read_mostly int sched_debug_enabled;
5426 static int __init sched_debug_setup(char *str)
5428 sched_debug_enabled = 1;
5432 early_param("sched_debug", sched_debug_setup);
5434 static inline bool sched_debug(void)
5436 return sched_debug_enabled;
5439 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5440 struct cpumask *groupmask)
5442 struct sched_group *group = sd->groups;
5445 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5446 cpumask_clear(groupmask);
5448 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5450 if (!(sd->flags & SD_LOAD_BALANCE)) {
5451 printk("does not load-balance\n");
5453 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5458 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5460 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5461 printk(KERN_ERR "ERROR: domain->span does not contain "
5464 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5465 printk(KERN_ERR "ERROR: domain->groups does not contain"
5469 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5473 printk(KERN_ERR "ERROR: group is NULL\n");
5478 * Even though we initialize ->capacity to something semi-sane,
5479 * we leave capacity_orig unset. This allows us to detect if
5480 * domain iteration is still funny without causing /0 traps.
5482 if (!group->sgc->capacity_orig) {
5483 printk(KERN_CONT "\n");
5484 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5488 if (!cpumask_weight(sched_group_cpus(group))) {
5489 printk(KERN_CONT "\n");
5490 printk(KERN_ERR "ERROR: empty group\n");
5494 if (!(sd->flags & SD_OVERLAP) &&
5495 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5496 printk(KERN_CONT "\n");
5497 printk(KERN_ERR "ERROR: repeated CPUs\n");
5501 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5503 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5505 printk(KERN_CONT " %s", str);
5506 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5507 printk(KERN_CONT " (cpu_capacity = %d)",
5508 group->sgc->capacity);
5511 group = group->next;
5512 } while (group != sd->groups);
5513 printk(KERN_CONT "\n");
5515 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5516 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5519 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5520 printk(KERN_ERR "ERROR: parent span is not a superset "
5521 "of domain->span\n");
5525 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5529 if (!sched_debug_enabled)
5533 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5537 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5540 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5548 #else /* !CONFIG_SCHED_DEBUG */
5549 # define sched_domain_debug(sd, cpu) do { } while (0)
5550 static inline bool sched_debug(void)
5554 #endif /* CONFIG_SCHED_DEBUG */
5556 static int sd_degenerate(struct sched_domain *sd)
5558 if (cpumask_weight(sched_domain_span(sd)) == 1)
5561 /* Following flags need at least 2 groups */
5562 if (sd->flags & (SD_LOAD_BALANCE |
5563 SD_BALANCE_NEWIDLE |
5566 SD_SHARE_CPUCAPACITY |
5567 SD_SHARE_PKG_RESOURCES |
5568 SD_SHARE_POWERDOMAIN)) {
5569 if (sd->groups != sd->groups->next)
5573 /* Following flags don't use groups */
5574 if (sd->flags & (SD_WAKE_AFFINE))
5581 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5583 unsigned long cflags = sd->flags, pflags = parent->flags;
5585 if (sd_degenerate(parent))
5588 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5591 /* Flags needing groups don't count if only 1 group in parent */
5592 if (parent->groups == parent->groups->next) {
5593 pflags &= ~(SD_LOAD_BALANCE |
5594 SD_BALANCE_NEWIDLE |
5597 SD_SHARE_CPUCAPACITY |
5598 SD_SHARE_PKG_RESOURCES |
5600 SD_SHARE_POWERDOMAIN);
5601 if (nr_node_ids == 1)
5602 pflags &= ~SD_SERIALIZE;
5604 if (~cflags & pflags)
5610 static void free_rootdomain(struct rcu_head *rcu)
5612 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5614 cpupri_cleanup(&rd->cpupri);
5615 cpudl_cleanup(&rd->cpudl);
5616 free_cpumask_var(rd->dlo_mask);
5617 free_cpumask_var(rd->rto_mask);
5618 free_cpumask_var(rd->online);
5619 free_cpumask_var(rd->span);
5623 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5625 struct root_domain *old_rd = NULL;
5626 unsigned long flags;
5628 raw_spin_lock_irqsave(&rq->lock, flags);
5633 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5636 cpumask_clear_cpu(rq->cpu, old_rd->span);
5639 * If we dont want to free the old_rd yet then
5640 * set old_rd to NULL to skip the freeing later
5643 if (!atomic_dec_and_test(&old_rd->refcount))
5647 atomic_inc(&rd->refcount);
5650 cpumask_set_cpu(rq->cpu, rd->span);
5651 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5654 raw_spin_unlock_irqrestore(&rq->lock, flags);
5657 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5660 static int init_rootdomain(struct root_domain *rd)
5662 memset(rd, 0, sizeof(*rd));
5664 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5666 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5668 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5670 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5673 init_dl_bw(&rd->dl_bw);
5674 if (cpudl_init(&rd->cpudl) != 0)
5677 if (cpupri_init(&rd->cpupri) != 0)
5682 free_cpumask_var(rd->rto_mask);
5684 free_cpumask_var(rd->dlo_mask);
5686 free_cpumask_var(rd->online);
5688 free_cpumask_var(rd->span);
5694 * By default the system creates a single root-domain with all cpus as
5695 * members (mimicking the global state we have today).
5697 struct root_domain def_root_domain;
5699 static void init_defrootdomain(void)
5701 init_rootdomain(&def_root_domain);
5703 atomic_set(&def_root_domain.refcount, 1);
5706 static struct root_domain *alloc_rootdomain(void)
5708 struct root_domain *rd;
5710 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5714 if (init_rootdomain(rd) != 0) {
5722 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5724 struct sched_group *tmp, *first;
5733 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5738 } while (sg != first);
5741 static void free_sched_domain(struct rcu_head *rcu)
5743 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5746 * If its an overlapping domain it has private groups, iterate and
5749 if (sd->flags & SD_OVERLAP) {
5750 free_sched_groups(sd->groups, 1);
5751 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5752 kfree(sd->groups->sgc);
5758 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5760 call_rcu(&sd->rcu, free_sched_domain);
5763 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5765 for (; sd; sd = sd->parent)
5766 destroy_sched_domain(sd, cpu);
5770 * Keep a special pointer to the highest sched_domain that has
5771 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5772 * allows us to avoid some pointer chasing select_idle_sibling().
5774 * Also keep a unique ID per domain (we use the first cpu number in
5775 * the cpumask of the domain), this allows us to quickly tell if
5776 * two cpus are in the same cache domain, see cpus_share_cache().
5778 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5779 DEFINE_PER_CPU(int, sd_llc_size);
5780 DEFINE_PER_CPU(int, sd_llc_id);
5781 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5782 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5783 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5785 static void update_top_cache_domain(int cpu)
5787 struct sched_domain *sd;
5788 struct sched_domain *busy_sd = NULL;
5792 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5794 id = cpumask_first(sched_domain_span(sd));
5795 size = cpumask_weight(sched_domain_span(sd));
5796 busy_sd = sd->parent; /* sd_busy */
5798 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5800 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5801 per_cpu(sd_llc_size, cpu) = size;
5802 per_cpu(sd_llc_id, cpu) = id;
5804 sd = lowest_flag_domain(cpu, SD_NUMA);
5805 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5807 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5808 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5812 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5813 * hold the hotplug lock.
5816 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5818 struct rq *rq = cpu_rq(cpu);
5819 struct sched_domain *tmp;
5821 /* Remove the sched domains which do not contribute to scheduling. */
5822 for (tmp = sd; tmp; ) {
5823 struct sched_domain *parent = tmp->parent;
5827 if (sd_parent_degenerate(tmp, parent)) {
5828 tmp->parent = parent->parent;
5830 parent->parent->child = tmp;
5832 * Transfer SD_PREFER_SIBLING down in case of a
5833 * degenerate parent; the spans match for this
5834 * so the property transfers.
5836 if (parent->flags & SD_PREFER_SIBLING)
5837 tmp->flags |= SD_PREFER_SIBLING;
5838 destroy_sched_domain(parent, cpu);
5843 if (sd && sd_degenerate(sd)) {
5846 destroy_sched_domain(tmp, cpu);
5851 sched_domain_debug(sd, cpu);
5853 rq_attach_root(rq, rd);
5855 rcu_assign_pointer(rq->sd, sd);
5856 destroy_sched_domains(tmp, cpu);
5858 update_top_cache_domain(cpu);
5861 /* cpus with isolated domains */
5862 static cpumask_var_t cpu_isolated_map;
5864 /* Setup the mask of cpus configured for isolated domains */
5865 static int __init isolated_cpu_setup(char *str)
5867 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5868 cpulist_parse(str, cpu_isolated_map);
5872 __setup("isolcpus=", isolated_cpu_setup);
5875 struct sched_domain ** __percpu sd;
5876 struct root_domain *rd;
5887 * Build an iteration mask that can exclude certain CPUs from the upwards
5890 * Asymmetric node setups can result in situations where the domain tree is of
5891 * unequal depth, make sure to skip domains that already cover the entire
5894 * In that case build_sched_domains() will have terminated the iteration early
5895 * and our sibling sd spans will be empty. Domains should always include the
5896 * cpu they're built on, so check that.
5899 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5901 const struct cpumask *span = sched_domain_span(sd);
5902 struct sd_data *sdd = sd->private;
5903 struct sched_domain *sibling;
5906 for_each_cpu(i, span) {
5907 sibling = *per_cpu_ptr(sdd->sd, i);
5908 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5911 cpumask_set_cpu(i, sched_group_mask(sg));
5916 * Return the canonical balance cpu for this group, this is the first cpu
5917 * of this group that's also in the iteration mask.
5919 int group_balance_cpu(struct sched_group *sg)
5921 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5925 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5927 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5928 const struct cpumask *span = sched_domain_span(sd);
5929 struct cpumask *covered = sched_domains_tmpmask;
5930 struct sd_data *sdd = sd->private;
5931 struct sched_domain *sibling;
5934 cpumask_clear(covered);
5936 for_each_cpu(i, span) {
5937 struct cpumask *sg_span;
5939 if (cpumask_test_cpu(i, covered))
5942 sibling = *per_cpu_ptr(sdd->sd, i);
5944 /* See the comment near build_group_mask(). */
5945 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5948 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5949 GFP_KERNEL, cpu_to_node(cpu));
5954 sg_span = sched_group_cpus(sg);
5956 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5958 cpumask_set_cpu(i, sg_span);
5960 cpumask_or(covered, covered, sg_span);
5962 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5963 if (atomic_inc_return(&sg->sgc->ref) == 1)
5964 build_group_mask(sd, sg);
5967 * Initialize sgc->capacity such that even if we mess up the
5968 * domains and no possible iteration will get us here, we won't
5971 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5972 sg->sgc->capacity_orig = sg->sgc->capacity;
5975 * Make sure the first group of this domain contains the
5976 * canonical balance cpu. Otherwise the sched_domain iteration
5977 * breaks. See update_sg_lb_stats().
5979 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5980 group_balance_cpu(sg) == cpu)
5990 sd->groups = groups;
5995 free_sched_groups(first, 0);
6000 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6002 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6003 struct sched_domain *child = sd->child;
6006 cpu = cpumask_first(sched_domain_span(child));
6009 *sg = *per_cpu_ptr(sdd->sg, cpu);
6010 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6011 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6018 * build_sched_groups will build a circular linked list of the groups
6019 * covered by the given span, and will set each group's ->cpumask correctly,
6020 * and ->cpu_capacity to 0.
6022 * Assumes the sched_domain tree is fully constructed
6025 build_sched_groups(struct sched_domain *sd, int cpu)
6027 struct sched_group *first = NULL, *last = NULL;
6028 struct sd_data *sdd = sd->private;
6029 const struct cpumask *span = sched_domain_span(sd);
6030 struct cpumask *covered;
6033 get_group(cpu, sdd, &sd->groups);
6034 atomic_inc(&sd->groups->ref);
6036 if (cpu != cpumask_first(span))
6039 lockdep_assert_held(&sched_domains_mutex);
6040 covered = sched_domains_tmpmask;
6042 cpumask_clear(covered);
6044 for_each_cpu(i, span) {
6045 struct sched_group *sg;
6048 if (cpumask_test_cpu(i, covered))
6051 group = get_group(i, sdd, &sg);
6052 cpumask_setall(sched_group_mask(sg));
6054 for_each_cpu(j, span) {
6055 if (get_group(j, sdd, NULL) != group)
6058 cpumask_set_cpu(j, covered);
6059 cpumask_set_cpu(j, sched_group_cpus(sg));
6074 * Initialize sched groups cpu_capacity.
6076 * cpu_capacity indicates the capacity of sched group, which is used while
6077 * distributing the load between different sched groups in a sched domain.
6078 * Typically cpu_capacity for all the groups in a sched domain will be same
6079 * unless there are asymmetries in the topology. If there are asymmetries,
6080 * group having more cpu_capacity will pickup more load compared to the
6081 * group having less cpu_capacity.
6083 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6085 struct sched_group *sg = sd->groups;
6090 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6092 } while (sg != sd->groups);
6094 if (cpu != group_balance_cpu(sg))
6097 update_group_capacity(sd, cpu);
6098 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6102 * Initializers for schedule domains
6103 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6106 static int default_relax_domain_level = -1;
6107 int sched_domain_level_max;
6109 static int __init setup_relax_domain_level(char *str)
6111 if (kstrtoint(str, 0, &default_relax_domain_level))
6112 pr_warn("Unable to set relax_domain_level\n");
6116 __setup("relax_domain_level=", setup_relax_domain_level);
6118 static void set_domain_attribute(struct sched_domain *sd,
6119 struct sched_domain_attr *attr)
6123 if (!attr || attr->relax_domain_level < 0) {
6124 if (default_relax_domain_level < 0)
6127 request = default_relax_domain_level;
6129 request = attr->relax_domain_level;
6130 if (request < sd->level) {
6131 /* turn off idle balance on this domain */
6132 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6134 /* turn on idle balance on this domain */
6135 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6139 static void __sdt_free(const struct cpumask *cpu_map);
6140 static int __sdt_alloc(const struct cpumask *cpu_map);
6142 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6143 const struct cpumask *cpu_map)
6147 if (!atomic_read(&d->rd->refcount))
6148 free_rootdomain(&d->rd->rcu); /* fall through */
6150 free_percpu(d->sd); /* fall through */
6152 __sdt_free(cpu_map); /* fall through */
6158 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6159 const struct cpumask *cpu_map)
6161 memset(d, 0, sizeof(*d));
6163 if (__sdt_alloc(cpu_map))
6164 return sa_sd_storage;
6165 d->sd = alloc_percpu(struct sched_domain *);
6167 return sa_sd_storage;
6168 d->rd = alloc_rootdomain();
6171 return sa_rootdomain;
6175 * NULL the sd_data elements we've used to build the sched_domain and
6176 * sched_group structure so that the subsequent __free_domain_allocs()
6177 * will not free the data we're using.
6179 static void claim_allocations(int cpu, struct sched_domain *sd)
6181 struct sd_data *sdd = sd->private;
6183 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6184 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6186 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6187 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6189 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6190 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6194 static int sched_domains_numa_levels;
6195 enum numa_topology_type sched_numa_topology_type;
6196 static int *sched_domains_numa_distance;
6197 int sched_max_numa_distance;
6198 static struct cpumask ***sched_domains_numa_masks;
6199 static int sched_domains_curr_level;
6203 * SD_flags allowed in topology descriptions.
6205 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6206 * SD_SHARE_PKG_RESOURCES - describes shared caches
6207 * SD_NUMA - describes NUMA topologies
6208 * SD_SHARE_POWERDOMAIN - describes shared power domain
6211 * SD_ASYM_PACKING - describes SMT quirks
6213 #define TOPOLOGY_SD_FLAGS \
6214 (SD_SHARE_CPUCAPACITY | \
6215 SD_SHARE_PKG_RESOURCES | \
6218 SD_SHARE_POWERDOMAIN)
6220 static struct sched_domain *
6221 sd_init(struct sched_domain_topology_level *tl, int cpu)
6223 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6224 int sd_weight, sd_flags = 0;
6228 * Ugly hack to pass state to sd_numa_mask()...
6230 sched_domains_curr_level = tl->numa_level;
6233 sd_weight = cpumask_weight(tl->mask(cpu));
6236 sd_flags = (*tl->sd_flags)();
6237 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6238 "wrong sd_flags in topology description\n"))
6239 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6241 *sd = (struct sched_domain){
6242 .min_interval = sd_weight,
6243 .max_interval = 2*sd_weight,
6245 .imbalance_pct = 125,
6247 .cache_nice_tries = 0,
6254 .flags = 1*SD_LOAD_BALANCE
6255 | 1*SD_BALANCE_NEWIDLE
6260 | 0*SD_SHARE_CPUCAPACITY
6261 | 0*SD_SHARE_PKG_RESOURCES
6263 | 0*SD_PREFER_SIBLING
6268 .last_balance = jiffies,
6269 .balance_interval = sd_weight,
6271 .max_newidle_lb_cost = 0,
6272 .next_decay_max_lb_cost = jiffies,
6273 #ifdef CONFIG_SCHED_DEBUG
6279 * Convert topological properties into behaviour.
6282 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6283 sd->imbalance_pct = 110;
6284 sd->smt_gain = 1178; /* ~15% */
6286 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6287 sd->imbalance_pct = 117;
6288 sd->cache_nice_tries = 1;
6292 } else if (sd->flags & SD_NUMA) {
6293 sd->cache_nice_tries = 2;
6297 sd->flags |= SD_SERIALIZE;
6298 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6299 sd->flags &= ~(SD_BALANCE_EXEC |
6306 sd->flags |= SD_PREFER_SIBLING;
6307 sd->cache_nice_tries = 1;
6312 sd->private = &tl->data;
6318 * Topology list, bottom-up.
6320 static struct sched_domain_topology_level default_topology[] = {
6321 #ifdef CONFIG_SCHED_SMT
6322 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6324 #ifdef CONFIG_SCHED_MC
6325 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6327 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6331 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6333 #define for_each_sd_topology(tl) \
6334 for (tl = sched_domain_topology; tl->mask; tl++)
6336 void set_sched_topology(struct sched_domain_topology_level *tl)
6338 sched_domain_topology = tl;
6343 static const struct cpumask *sd_numa_mask(int cpu)
6345 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6348 static void sched_numa_warn(const char *str)
6350 static int done = false;
6358 printk(KERN_WARNING "ERROR: %s\n\n", str);
6360 for (i = 0; i < nr_node_ids; i++) {
6361 printk(KERN_WARNING " ");
6362 for (j = 0; j < nr_node_ids; j++)
6363 printk(KERN_CONT "%02d ", node_distance(i,j));
6364 printk(KERN_CONT "\n");
6366 printk(KERN_WARNING "\n");
6369 bool find_numa_distance(int distance)
6373 if (distance == node_distance(0, 0))
6376 for (i = 0; i < sched_domains_numa_levels; i++) {
6377 if (sched_domains_numa_distance[i] == distance)
6385 * A system can have three types of NUMA topology:
6386 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6387 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6388 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6390 * The difference between a glueless mesh topology and a backplane
6391 * topology lies in whether communication between not directly
6392 * connected nodes goes through intermediary nodes (where programs
6393 * could run), or through backplane controllers. This affects
6394 * placement of programs.
6396 * The type of topology can be discerned with the following tests:
6397 * - If the maximum distance between any nodes is 1 hop, the system
6398 * is directly connected.
6399 * - If for two nodes A and B, located N > 1 hops away from each other,
6400 * there is an intermediary node C, which is < N hops away from both
6401 * nodes A and B, the system is a glueless mesh.
6403 static void init_numa_topology_type(void)
6407 n = sched_max_numa_distance;
6410 sched_numa_topology_type = NUMA_DIRECT;
6412 for_each_online_node(a) {
6413 for_each_online_node(b) {
6414 /* Find two nodes furthest removed from each other. */
6415 if (node_distance(a, b) < n)
6418 /* Is there an intermediary node between a and b? */
6419 for_each_online_node(c) {
6420 if (node_distance(a, c) < n &&
6421 node_distance(b, c) < n) {
6422 sched_numa_topology_type =
6428 sched_numa_topology_type = NUMA_BACKPLANE;
6434 static void sched_init_numa(void)
6436 int next_distance, curr_distance = node_distance(0, 0);
6437 struct sched_domain_topology_level *tl;
6441 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6442 if (!sched_domains_numa_distance)
6446 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6447 * unique distances in the node_distance() table.
6449 * Assumes node_distance(0,j) includes all distances in
6450 * node_distance(i,j) in order to avoid cubic time.
6452 next_distance = curr_distance;
6453 for (i = 0; i < nr_node_ids; i++) {
6454 for (j = 0; j < nr_node_ids; j++) {
6455 for (k = 0; k < nr_node_ids; k++) {
6456 int distance = node_distance(i, k);
6458 if (distance > curr_distance &&
6459 (distance < next_distance ||
6460 next_distance == curr_distance))
6461 next_distance = distance;
6464 * While not a strong assumption it would be nice to know
6465 * about cases where if node A is connected to B, B is not
6466 * equally connected to A.
6468 if (sched_debug() && node_distance(k, i) != distance)
6469 sched_numa_warn("Node-distance not symmetric");
6471 if (sched_debug() && i && !find_numa_distance(distance))
6472 sched_numa_warn("Node-0 not representative");
6474 if (next_distance != curr_distance) {
6475 sched_domains_numa_distance[level++] = next_distance;
6476 sched_domains_numa_levels = level;
6477 curr_distance = next_distance;
6482 * In case of sched_debug() we verify the above assumption.
6492 * 'level' contains the number of unique distances, excluding the
6493 * identity distance node_distance(i,i).
6495 * The sched_domains_numa_distance[] array includes the actual distance
6500 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6501 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6502 * the array will contain less then 'level' members. This could be
6503 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6504 * in other functions.
6506 * We reset it to 'level' at the end of this function.
6508 sched_domains_numa_levels = 0;
6510 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6511 if (!sched_domains_numa_masks)
6515 * Now for each level, construct a mask per node which contains all
6516 * cpus of nodes that are that many hops away from us.
6518 for (i = 0; i < level; i++) {
6519 sched_domains_numa_masks[i] =
6520 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6521 if (!sched_domains_numa_masks[i])
6524 for (j = 0; j < nr_node_ids; j++) {
6525 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6529 sched_domains_numa_masks[i][j] = mask;
6531 for (k = 0; k < nr_node_ids; k++) {
6532 if (node_distance(j, k) > sched_domains_numa_distance[i])
6535 cpumask_or(mask, mask, cpumask_of_node(k));
6540 /* Compute default topology size */
6541 for (i = 0; sched_domain_topology[i].mask; i++);
6543 tl = kzalloc((i + level + 1) *
6544 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6549 * Copy the default topology bits..
6551 for (i = 0; sched_domain_topology[i].mask; i++)
6552 tl[i] = sched_domain_topology[i];
6555 * .. and append 'j' levels of NUMA goodness.
6557 for (j = 0; j < level; i++, j++) {
6558 tl[i] = (struct sched_domain_topology_level){
6559 .mask = sd_numa_mask,
6560 .sd_flags = cpu_numa_flags,
6561 .flags = SDTL_OVERLAP,
6567 sched_domain_topology = tl;
6569 sched_domains_numa_levels = level;
6570 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6572 init_numa_topology_type();
6575 static void sched_domains_numa_masks_set(int cpu)
6578 int node = cpu_to_node(cpu);
6580 for (i = 0; i < sched_domains_numa_levels; i++) {
6581 for (j = 0; j < nr_node_ids; j++) {
6582 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6583 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6588 static void sched_domains_numa_masks_clear(int cpu)
6591 for (i = 0; i < sched_domains_numa_levels; i++) {
6592 for (j = 0; j < nr_node_ids; j++)
6593 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6598 * Update sched_domains_numa_masks[level][node] array when new cpus
6601 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6602 unsigned long action,
6605 int cpu = (long)hcpu;
6607 switch (action & ~CPU_TASKS_FROZEN) {
6609 sched_domains_numa_masks_set(cpu);
6613 sched_domains_numa_masks_clear(cpu);
6623 static inline void sched_init_numa(void)
6627 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6628 unsigned long action,
6633 #endif /* CONFIG_NUMA */
6635 static int __sdt_alloc(const struct cpumask *cpu_map)
6637 struct sched_domain_topology_level *tl;
6640 for_each_sd_topology(tl) {
6641 struct sd_data *sdd = &tl->data;
6643 sdd->sd = alloc_percpu(struct sched_domain *);
6647 sdd->sg = alloc_percpu(struct sched_group *);
6651 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6655 for_each_cpu(j, cpu_map) {
6656 struct sched_domain *sd;
6657 struct sched_group *sg;
6658 struct sched_group_capacity *sgc;
6660 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6661 GFP_KERNEL, cpu_to_node(j));
6665 *per_cpu_ptr(sdd->sd, j) = sd;
6667 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6668 GFP_KERNEL, cpu_to_node(j));
6674 *per_cpu_ptr(sdd->sg, j) = sg;
6676 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6677 GFP_KERNEL, cpu_to_node(j));
6681 *per_cpu_ptr(sdd->sgc, j) = sgc;
6688 static void __sdt_free(const struct cpumask *cpu_map)
6690 struct sched_domain_topology_level *tl;
6693 for_each_sd_topology(tl) {
6694 struct sd_data *sdd = &tl->data;
6696 for_each_cpu(j, cpu_map) {
6697 struct sched_domain *sd;
6700 sd = *per_cpu_ptr(sdd->sd, j);
6701 if (sd && (sd->flags & SD_OVERLAP))
6702 free_sched_groups(sd->groups, 0);
6703 kfree(*per_cpu_ptr(sdd->sd, j));
6707 kfree(*per_cpu_ptr(sdd->sg, j));
6709 kfree(*per_cpu_ptr(sdd->sgc, j));
6711 free_percpu(sdd->sd);
6713 free_percpu(sdd->sg);
6715 free_percpu(sdd->sgc);
6720 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6721 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6722 struct sched_domain *child, int cpu)
6724 struct sched_domain *sd = sd_init(tl, cpu);
6728 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6730 sd->level = child->level + 1;
6731 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6735 if (!cpumask_subset(sched_domain_span(child),
6736 sched_domain_span(sd))) {
6737 pr_err("BUG: arch topology borken\n");
6738 #ifdef CONFIG_SCHED_DEBUG
6739 pr_err(" the %s domain not a subset of the %s domain\n",
6740 child->name, sd->name);
6742 /* Fixup, ensure @sd has at least @child cpus. */
6743 cpumask_or(sched_domain_span(sd),
6744 sched_domain_span(sd),
6745 sched_domain_span(child));
6749 set_domain_attribute(sd, attr);
6755 * Build sched domains for a given set of cpus and attach the sched domains
6756 * to the individual cpus
6758 static int build_sched_domains(const struct cpumask *cpu_map,
6759 struct sched_domain_attr *attr)
6761 enum s_alloc alloc_state;
6762 struct sched_domain *sd;
6764 int i, ret = -ENOMEM;
6766 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6767 if (alloc_state != sa_rootdomain)
6770 /* Set up domains for cpus specified by the cpu_map. */
6771 for_each_cpu(i, cpu_map) {
6772 struct sched_domain_topology_level *tl;
6775 for_each_sd_topology(tl) {
6776 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6777 if (tl == sched_domain_topology)
6778 *per_cpu_ptr(d.sd, i) = sd;
6779 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6780 sd->flags |= SD_OVERLAP;
6781 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6786 /* Build the groups for the domains */
6787 for_each_cpu(i, cpu_map) {
6788 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6789 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6790 if (sd->flags & SD_OVERLAP) {
6791 if (build_overlap_sched_groups(sd, i))
6794 if (build_sched_groups(sd, i))
6800 /* Calculate CPU capacity for physical packages and nodes */
6801 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6802 if (!cpumask_test_cpu(i, cpu_map))
6805 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6806 claim_allocations(i, sd);
6807 init_sched_groups_capacity(i, sd);
6811 /* Attach the domains */
6813 for_each_cpu(i, cpu_map) {
6814 sd = *per_cpu_ptr(d.sd, i);
6815 cpu_attach_domain(sd, d.rd, i);
6821 __free_domain_allocs(&d, alloc_state, cpu_map);
6825 static cpumask_var_t *doms_cur; /* current sched domains */
6826 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6827 static struct sched_domain_attr *dattr_cur;
6828 /* attribues of custom domains in 'doms_cur' */
6831 * Special case: If a kmalloc of a doms_cur partition (array of
6832 * cpumask) fails, then fallback to a single sched domain,
6833 * as determined by the single cpumask fallback_doms.
6835 static cpumask_var_t fallback_doms;
6838 * arch_update_cpu_topology lets virtualized architectures update the
6839 * cpu core maps. It is supposed to return 1 if the topology changed
6840 * or 0 if it stayed the same.
6842 int __weak arch_update_cpu_topology(void)
6847 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6850 cpumask_var_t *doms;
6852 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6855 for (i = 0; i < ndoms; i++) {
6856 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6857 free_sched_domains(doms, i);
6864 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6867 for (i = 0; i < ndoms; i++)
6868 free_cpumask_var(doms[i]);
6873 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6874 * For now this just excludes isolated cpus, but could be used to
6875 * exclude other special cases in the future.
6877 static int init_sched_domains(const struct cpumask *cpu_map)
6881 arch_update_cpu_topology();
6883 doms_cur = alloc_sched_domains(ndoms_cur);
6885 doms_cur = &fallback_doms;
6886 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6887 err = build_sched_domains(doms_cur[0], NULL);
6888 register_sched_domain_sysctl();
6894 * Detach sched domains from a group of cpus specified in cpu_map
6895 * These cpus will now be attached to the NULL domain
6897 static void detach_destroy_domains(const struct cpumask *cpu_map)
6902 for_each_cpu(i, cpu_map)
6903 cpu_attach_domain(NULL, &def_root_domain, i);
6907 /* handle null as "default" */
6908 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6909 struct sched_domain_attr *new, int idx_new)
6911 struct sched_domain_attr tmp;
6918 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6919 new ? (new + idx_new) : &tmp,
6920 sizeof(struct sched_domain_attr));
6924 * Partition sched domains as specified by the 'ndoms_new'
6925 * cpumasks in the array doms_new[] of cpumasks. This compares
6926 * doms_new[] to the current sched domain partitioning, doms_cur[].
6927 * It destroys each deleted domain and builds each new domain.
6929 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6930 * The masks don't intersect (don't overlap.) We should setup one
6931 * sched domain for each mask. CPUs not in any of the cpumasks will
6932 * not be load balanced. If the same cpumask appears both in the
6933 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6936 * The passed in 'doms_new' should be allocated using
6937 * alloc_sched_domains. This routine takes ownership of it and will
6938 * free_sched_domains it when done with it. If the caller failed the
6939 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6940 * and partition_sched_domains() will fallback to the single partition
6941 * 'fallback_doms', it also forces the domains to be rebuilt.
6943 * If doms_new == NULL it will be replaced with cpu_online_mask.
6944 * ndoms_new == 0 is a special case for destroying existing domains,
6945 * and it will not create the default domain.
6947 * Call with hotplug lock held
6949 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6950 struct sched_domain_attr *dattr_new)
6955 mutex_lock(&sched_domains_mutex);
6957 /* always unregister in case we don't destroy any domains */
6958 unregister_sched_domain_sysctl();
6960 /* Let architecture update cpu core mappings. */
6961 new_topology = arch_update_cpu_topology();
6963 n = doms_new ? ndoms_new : 0;
6965 /* Destroy deleted domains */
6966 for (i = 0; i < ndoms_cur; i++) {
6967 for (j = 0; j < n && !new_topology; j++) {
6968 if (cpumask_equal(doms_cur[i], doms_new[j])
6969 && dattrs_equal(dattr_cur, i, dattr_new, j))
6972 /* no match - a current sched domain not in new doms_new[] */
6973 detach_destroy_domains(doms_cur[i]);
6979 if (doms_new == NULL) {
6981 doms_new = &fallback_doms;
6982 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6983 WARN_ON_ONCE(dattr_new);
6986 /* Build new domains */
6987 for (i = 0; i < ndoms_new; i++) {
6988 for (j = 0; j < n && !new_topology; j++) {
6989 if (cpumask_equal(doms_new[i], doms_cur[j])
6990 && dattrs_equal(dattr_new, i, dattr_cur, j))
6993 /* no match - add a new doms_new */
6994 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6999 /* Remember the new sched domains */
7000 if (doms_cur != &fallback_doms)
7001 free_sched_domains(doms_cur, ndoms_cur);
7002 kfree(dattr_cur); /* kfree(NULL) is safe */
7003 doms_cur = doms_new;
7004 dattr_cur = dattr_new;
7005 ndoms_cur = ndoms_new;
7007 register_sched_domain_sysctl();
7009 mutex_unlock(&sched_domains_mutex);
7012 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7015 * Update cpusets according to cpu_active mask. If cpusets are
7016 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7017 * around partition_sched_domains().
7019 * If we come here as part of a suspend/resume, don't touch cpusets because we
7020 * want to restore it back to its original state upon resume anyway.
7022 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7026 case CPU_ONLINE_FROZEN:
7027 case CPU_DOWN_FAILED_FROZEN:
7030 * num_cpus_frozen tracks how many CPUs are involved in suspend
7031 * resume sequence. As long as this is not the last online
7032 * operation in the resume sequence, just build a single sched
7033 * domain, ignoring cpusets.
7036 if (likely(num_cpus_frozen)) {
7037 partition_sched_domains(1, NULL, NULL);
7042 * This is the last CPU online operation. So fall through and
7043 * restore the original sched domains by considering the
7044 * cpuset configurations.
7048 case CPU_DOWN_FAILED:
7049 cpuset_update_active_cpus(true);
7057 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7061 case CPU_DOWN_PREPARE:
7062 cpuset_update_active_cpus(false);
7064 case CPU_DOWN_PREPARE_FROZEN:
7066 partition_sched_domains(1, NULL, NULL);
7074 void __init sched_init_smp(void)
7076 cpumask_var_t non_isolated_cpus;
7078 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7079 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7084 * There's no userspace yet to cause hotplug operations; hence all the
7085 * cpu masks are stable and all blatant races in the below code cannot
7088 mutex_lock(&sched_domains_mutex);
7089 init_sched_domains(cpu_active_mask);
7090 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7091 if (cpumask_empty(non_isolated_cpus))
7092 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7093 mutex_unlock(&sched_domains_mutex);
7095 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7096 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7097 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7101 /* Move init over to a non-isolated CPU */
7102 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7104 sched_init_granularity();
7105 free_cpumask_var(non_isolated_cpus);
7107 init_sched_rt_class();
7108 init_sched_dl_class();
7111 void __init sched_init_smp(void)
7113 sched_init_granularity();
7115 #endif /* CONFIG_SMP */
7117 const_debug unsigned int sysctl_timer_migration = 1;
7119 int in_sched_functions(unsigned long addr)
7121 return in_lock_functions(addr) ||
7122 (addr >= (unsigned long)__sched_text_start
7123 && addr < (unsigned long)__sched_text_end);
7126 #ifdef CONFIG_CGROUP_SCHED
7128 * Default task group.
7129 * Every task in system belongs to this group at bootup.
7131 struct task_group root_task_group;
7132 LIST_HEAD(task_groups);
7135 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7137 void __init sched_init(void)
7140 unsigned long alloc_size = 0, ptr;
7142 #ifdef CONFIG_FAIR_GROUP_SCHED
7143 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7145 #ifdef CONFIG_RT_GROUP_SCHED
7146 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7149 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7151 #ifdef CONFIG_FAIR_GROUP_SCHED
7152 root_task_group.se = (struct sched_entity **)ptr;
7153 ptr += nr_cpu_ids * sizeof(void **);
7155 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7156 ptr += nr_cpu_ids * sizeof(void **);
7158 #endif /* CONFIG_FAIR_GROUP_SCHED */
7159 #ifdef CONFIG_RT_GROUP_SCHED
7160 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7161 ptr += nr_cpu_ids * sizeof(void **);
7163 root_task_group.rt_rq = (struct rt_rq **)ptr;
7164 ptr += nr_cpu_ids * sizeof(void **);
7166 #endif /* CONFIG_RT_GROUP_SCHED */
7168 #ifdef CONFIG_CPUMASK_OFFSTACK
7169 for_each_possible_cpu(i) {
7170 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7171 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7173 #endif /* CONFIG_CPUMASK_OFFSTACK */
7175 init_rt_bandwidth(&def_rt_bandwidth,
7176 global_rt_period(), global_rt_runtime());
7177 init_dl_bandwidth(&def_dl_bandwidth,
7178 global_rt_period(), global_rt_runtime());
7181 init_defrootdomain();
7184 #ifdef CONFIG_RT_GROUP_SCHED
7185 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7186 global_rt_period(), global_rt_runtime());
7187 #endif /* CONFIG_RT_GROUP_SCHED */
7189 #ifdef CONFIG_CGROUP_SCHED
7190 list_add(&root_task_group.list, &task_groups);
7191 INIT_LIST_HEAD(&root_task_group.children);
7192 INIT_LIST_HEAD(&root_task_group.siblings);
7193 autogroup_init(&init_task);
7195 #endif /* CONFIG_CGROUP_SCHED */
7197 for_each_possible_cpu(i) {
7201 raw_spin_lock_init(&rq->lock);
7203 rq->calc_load_active = 0;
7204 rq->calc_load_update = jiffies + LOAD_FREQ;
7205 init_cfs_rq(&rq->cfs);
7206 init_rt_rq(&rq->rt, rq);
7207 init_dl_rq(&rq->dl, rq);
7208 #ifdef CONFIG_FAIR_GROUP_SCHED
7209 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7210 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7212 * How much cpu bandwidth does root_task_group get?
7214 * In case of task-groups formed thr' the cgroup filesystem, it
7215 * gets 100% of the cpu resources in the system. This overall
7216 * system cpu resource is divided among the tasks of
7217 * root_task_group and its child task-groups in a fair manner,
7218 * based on each entity's (task or task-group's) weight
7219 * (se->load.weight).
7221 * In other words, if root_task_group has 10 tasks of weight
7222 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7223 * then A0's share of the cpu resource is:
7225 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7227 * We achieve this by letting root_task_group's tasks sit
7228 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7230 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7231 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7232 #endif /* CONFIG_FAIR_GROUP_SCHED */
7234 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7235 #ifdef CONFIG_RT_GROUP_SCHED
7236 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7239 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7240 rq->cpu_load[j] = 0;
7242 rq->last_load_update_tick = jiffies;
7247 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7248 rq->post_schedule = 0;
7249 rq->active_balance = 0;
7250 rq->next_balance = jiffies;
7255 rq->avg_idle = 2*sysctl_sched_migration_cost;
7256 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7258 INIT_LIST_HEAD(&rq->cfs_tasks);
7260 rq_attach_root(rq, &def_root_domain);
7261 #ifdef CONFIG_NO_HZ_COMMON
7264 #ifdef CONFIG_NO_HZ_FULL
7265 rq->last_sched_tick = 0;
7269 atomic_set(&rq->nr_iowait, 0);
7272 set_load_weight(&init_task);
7274 #ifdef CONFIG_PREEMPT_NOTIFIERS
7275 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7279 * The boot idle thread does lazy MMU switching as well:
7281 atomic_inc(&init_mm.mm_count);
7282 enter_lazy_tlb(&init_mm, current);
7285 * During early bootup we pretend to be a normal task:
7287 current->sched_class = &fair_sched_class;
7290 * Make us the idle thread. Technically, schedule() should not be
7291 * called from this thread, however somewhere below it might be,
7292 * but because we are the idle thread, we just pick up running again
7293 * when this runqueue becomes "idle".
7295 init_idle(current, smp_processor_id());
7297 calc_load_update = jiffies + LOAD_FREQ;
7300 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7301 /* May be allocated at isolcpus cmdline parse time */
7302 if (cpu_isolated_map == NULL)
7303 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7304 idle_thread_set_boot_cpu();
7305 set_cpu_rq_start_time();
7307 init_sched_fair_class();
7309 scheduler_running = 1;
7312 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7313 static inline int preempt_count_equals(int preempt_offset)
7315 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7317 return (nested == preempt_offset);
7320 void __might_sleep(const char *file, int line, int preempt_offset)
7323 * Blocking primitives will set (and therefore destroy) current->state,
7324 * since we will exit with TASK_RUNNING make sure we enter with it,
7325 * otherwise we will destroy state.
7327 if (WARN_ONCE(current->state != TASK_RUNNING,
7328 "do not call blocking ops when !TASK_RUNNING; "
7329 "state=%lx set at [<%p>] %pS\n",
7331 (void *)current->task_state_change,
7332 (void *)current->task_state_change))
7333 __set_current_state(TASK_RUNNING);
7335 ___might_sleep(file, line, preempt_offset);
7337 EXPORT_SYMBOL(__might_sleep);
7339 void ___might_sleep(const char *file, int line, int preempt_offset)
7341 static unsigned long prev_jiffy; /* ratelimiting */
7343 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7344 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7345 !is_idle_task(current)) ||
7346 system_state != SYSTEM_RUNNING || oops_in_progress)
7348 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7350 prev_jiffy = jiffies;
7353 "BUG: sleeping function called from invalid context at %s:%d\n",
7356 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7357 in_atomic(), irqs_disabled(),
7358 current->pid, current->comm);
7360 if (task_stack_end_corrupted(current))
7361 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7363 debug_show_held_locks(current);
7364 if (irqs_disabled())
7365 print_irqtrace_events(current);
7366 #ifdef CONFIG_DEBUG_PREEMPT
7367 if (!preempt_count_equals(preempt_offset)) {
7368 pr_err("Preemption disabled at:");
7369 print_ip_sym(current->preempt_disable_ip);
7375 EXPORT_SYMBOL(___might_sleep);
7378 #ifdef CONFIG_MAGIC_SYSRQ
7379 static void normalize_task(struct rq *rq, struct task_struct *p)
7381 const struct sched_class *prev_class = p->sched_class;
7382 struct sched_attr attr = {
7383 .sched_policy = SCHED_NORMAL,
7385 int old_prio = p->prio;
7388 queued = task_on_rq_queued(p);
7390 dequeue_task(rq, p, 0);
7391 __setscheduler(rq, p, &attr);
7393 enqueue_task(rq, p, 0);
7397 check_class_changed(rq, p, prev_class, old_prio);
7400 void normalize_rt_tasks(void)
7402 struct task_struct *g, *p;
7403 unsigned long flags;
7406 read_lock(&tasklist_lock);
7407 for_each_process_thread(g, p) {
7409 * Only normalize user tasks:
7411 if (p->flags & PF_KTHREAD)
7414 p->se.exec_start = 0;
7415 #ifdef CONFIG_SCHEDSTATS
7416 p->se.statistics.wait_start = 0;
7417 p->se.statistics.sleep_start = 0;
7418 p->se.statistics.block_start = 0;
7421 if (!dl_task(p) && !rt_task(p)) {
7423 * Renice negative nice level userspace
7426 if (task_nice(p) < 0)
7427 set_user_nice(p, 0);
7431 rq = task_rq_lock(p, &flags);
7432 normalize_task(rq, p);
7433 task_rq_unlock(rq, p, &flags);
7435 read_unlock(&tasklist_lock);
7438 #endif /* CONFIG_MAGIC_SYSRQ */
7440 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7442 * These functions are only useful for the IA64 MCA handling, or kdb.
7444 * They can only be called when the whole system has been
7445 * stopped - every CPU needs to be quiescent, and no scheduling
7446 * activity can take place. Using them for anything else would
7447 * be a serious bug, and as a result, they aren't even visible
7448 * under any other configuration.
7452 * curr_task - return the current task for a given cpu.
7453 * @cpu: the processor in question.
7455 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7457 * Return: The current task for @cpu.
7459 struct task_struct *curr_task(int cpu)
7461 return cpu_curr(cpu);
7464 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7468 * set_curr_task - set the current task for a given cpu.
7469 * @cpu: the processor in question.
7470 * @p: the task pointer to set.
7472 * Description: This function must only be used when non-maskable interrupts
7473 * are serviced on a separate stack. It allows the architecture to switch the
7474 * notion of the current task on a cpu in a non-blocking manner. This function
7475 * must be called with all CPU's synchronized, and interrupts disabled, the
7476 * and caller must save the original value of the current task (see
7477 * curr_task() above) and restore that value before reenabling interrupts and
7478 * re-starting the system.
7480 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7482 void set_curr_task(int cpu, struct task_struct *p)
7489 #ifdef CONFIG_CGROUP_SCHED
7490 /* task_group_lock serializes the addition/removal of task groups */
7491 static DEFINE_SPINLOCK(task_group_lock);
7493 static void free_sched_group(struct task_group *tg)
7495 free_fair_sched_group(tg);
7496 free_rt_sched_group(tg);
7501 /* allocate runqueue etc for a new task group */
7502 struct task_group *sched_create_group(struct task_group *parent)
7504 struct task_group *tg;
7506 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7508 return ERR_PTR(-ENOMEM);
7510 if (!alloc_fair_sched_group(tg, parent))
7513 if (!alloc_rt_sched_group(tg, parent))
7519 free_sched_group(tg);
7520 return ERR_PTR(-ENOMEM);
7523 void sched_online_group(struct task_group *tg, struct task_group *parent)
7525 unsigned long flags;
7527 spin_lock_irqsave(&task_group_lock, flags);
7528 list_add_rcu(&tg->list, &task_groups);
7530 WARN_ON(!parent); /* root should already exist */
7532 tg->parent = parent;
7533 INIT_LIST_HEAD(&tg->children);
7534 list_add_rcu(&tg->siblings, &parent->children);
7535 spin_unlock_irqrestore(&task_group_lock, flags);
7538 /* rcu callback to free various structures associated with a task group */
7539 static void free_sched_group_rcu(struct rcu_head *rhp)
7541 /* now it should be safe to free those cfs_rqs */
7542 free_sched_group(container_of(rhp, struct task_group, rcu));
7545 /* Destroy runqueue etc associated with a task group */
7546 void sched_destroy_group(struct task_group *tg)
7548 /* wait for possible concurrent references to cfs_rqs complete */
7549 call_rcu(&tg->rcu, free_sched_group_rcu);
7552 void sched_offline_group(struct task_group *tg)
7554 unsigned long flags;
7557 /* end participation in shares distribution */
7558 for_each_possible_cpu(i)
7559 unregister_fair_sched_group(tg, i);
7561 spin_lock_irqsave(&task_group_lock, flags);
7562 list_del_rcu(&tg->list);
7563 list_del_rcu(&tg->siblings);
7564 spin_unlock_irqrestore(&task_group_lock, flags);
7567 /* change task's runqueue when it moves between groups.
7568 * The caller of this function should have put the task in its new group
7569 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7570 * reflect its new group.
7572 void sched_move_task(struct task_struct *tsk)
7574 struct task_group *tg;
7575 int queued, running;
7576 unsigned long flags;
7579 rq = task_rq_lock(tsk, &flags);
7581 running = task_current(rq, tsk);
7582 queued = task_on_rq_queued(tsk);
7585 dequeue_task(rq, tsk, 0);
7586 if (unlikely(running))
7587 put_prev_task(rq, tsk);
7590 * All callers are synchronized by task_rq_lock(); we do not use RCU
7591 * which is pointless here. Thus, we pass "true" to task_css_check()
7592 * to prevent lockdep warnings.
7594 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7595 struct task_group, css);
7596 tg = autogroup_task_group(tsk, tg);
7597 tsk->sched_task_group = tg;
7599 #ifdef CONFIG_FAIR_GROUP_SCHED
7600 if (tsk->sched_class->task_move_group)
7601 tsk->sched_class->task_move_group(tsk, queued);
7604 set_task_rq(tsk, task_cpu(tsk));
7606 if (unlikely(running))
7607 tsk->sched_class->set_curr_task(rq);
7609 enqueue_task(rq, tsk, 0);
7611 task_rq_unlock(rq, tsk, &flags);
7613 #endif /* CONFIG_CGROUP_SCHED */
7615 #ifdef CONFIG_RT_GROUP_SCHED
7617 * Ensure that the real time constraints are schedulable.
7619 static DEFINE_MUTEX(rt_constraints_mutex);
7621 /* Must be called with tasklist_lock held */
7622 static inline int tg_has_rt_tasks(struct task_group *tg)
7624 struct task_struct *g, *p;
7626 for_each_process_thread(g, p) {
7627 if (rt_task(p) && task_group(p) == tg)
7634 struct rt_schedulable_data {
7635 struct task_group *tg;
7640 static int tg_rt_schedulable(struct task_group *tg, void *data)
7642 struct rt_schedulable_data *d = data;
7643 struct task_group *child;
7644 unsigned long total, sum = 0;
7645 u64 period, runtime;
7647 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7648 runtime = tg->rt_bandwidth.rt_runtime;
7651 period = d->rt_period;
7652 runtime = d->rt_runtime;
7656 * Cannot have more runtime than the period.
7658 if (runtime > period && runtime != RUNTIME_INF)
7662 * Ensure we don't starve existing RT tasks.
7664 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7667 total = to_ratio(period, runtime);
7670 * Nobody can have more than the global setting allows.
7672 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7676 * The sum of our children's runtime should not exceed our own.
7678 list_for_each_entry_rcu(child, &tg->children, siblings) {
7679 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7680 runtime = child->rt_bandwidth.rt_runtime;
7682 if (child == d->tg) {
7683 period = d->rt_period;
7684 runtime = d->rt_runtime;
7687 sum += to_ratio(period, runtime);
7696 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7700 struct rt_schedulable_data data = {
7702 .rt_period = period,
7703 .rt_runtime = runtime,
7707 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7713 static int tg_set_rt_bandwidth(struct task_group *tg,
7714 u64 rt_period, u64 rt_runtime)
7718 mutex_lock(&rt_constraints_mutex);
7719 read_lock(&tasklist_lock);
7720 err = __rt_schedulable(tg, rt_period, rt_runtime);
7724 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7725 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7726 tg->rt_bandwidth.rt_runtime = rt_runtime;
7728 for_each_possible_cpu(i) {
7729 struct rt_rq *rt_rq = tg->rt_rq[i];
7731 raw_spin_lock(&rt_rq->rt_runtime_lock);
7732 rt_rq->rt_runtime = rt_runtime;
7733 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7735 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7737 read_unlock(&tasklist_lock);
7738 mutex_unlock(&rt_constraints_mutex);
7743 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7745 u64 rt_runtime, rt_period;
7747 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7748 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7749 if (rt_runtime_us < 0)
7750 rt_runtime = RUNTIME_INF;
7752 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7755 static long sched_group_rt_runtime(struct task_group *tg)
7759 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7762 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7763 do_div(rt_runtime_us, NSEC_PER_USEC);
7764 return rt_runtime_us;
7767 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7769 u64 rt_runtime, rt_period;
7771 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7772 rt_runtime = tg->rt_bandwidth.rt_runtime;
7777 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7780 static long sched_group_rt_period(struct task_group *tg)
7784 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7785 do_div(rt_period_us, NSEC_PER_USEC);
7786 return rt_period_us;
7788 #endif /* CONFIG_RT_GROUP_SCHED */
7790 #ifdef CONFIG_RT_GROUP_SCHED
7791 static int sched_rt_global_constraints(void)
7795 mutex_lock(&rt_constraints_mutex);
7796 read_lock(&tasklist_lock);
7797 ret = __rt_schedulable(NULL, 0, 0);
7798 read_unlock(&tasklist_lock);
7799 mutex_unlock(&rt_constraints_mutex);
7804 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7806 /* Don't accept realtime tasks when there is no way for them to run */
7807 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7813 #else /* !CONFIG_RT_GROUP_SCHED */
7814 static int sched_rt_global_constraints(void)
7816 unsigned long flags;
7819 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7820 for_each_possible_cpu(i) {
7821 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7823 raw_spin_lock(&rt_rq->rt_runtime_lock);
7824 rt_rq->rt_runtime = global_rt_runtime();
7825 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7827 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7831 #endif /* CONFIG_RT_GROUP_SCHED */
7833 static int sched_dl_global_constraints(void)
7835 u64 runtime = global_rt_runtime();
7836 u64 period = global_rt_period();
7837 u64 new_bw = to_ratio(period, runtime);
7840 unsigned long flags;
7843 * Here we want to check the bandwidth not being set to some
7844 * value smaller than the currently allocated bandwidth in
7845 * any of the root_domains.
7847 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7848 * cycling on root_domains... Discussion on different/better
7849 * solutions is welcome!
7851 for_each_possible_cpu(cpu) {
7852 rcu_read_lock_sched();
7853 dl_b = dl_bw_of(cpu);
7855 raw_spin_lock_irqsave(&dl_b->lock, flags);
7856 if (new_bw < dl_b->total_bw)
7858 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7860 rcu_read_unlock_sched();
7869 static void sched_dl_do_global(void)
7874 unsigned long flags;
7876 def_dl_bandwidth.dl_period = global_rt_period();
7877 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7879 if (global_rt_runtime() != RUNTIME_INF)
7880 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7883 * FIXME: As above...
7885 for_each_possible_cpu(cpu) {
7886 rcu_read_lock_sched();
7887 dl_b = dl_bw_of(cpu);
7889 raw_spin_lock_irqsave(&dl_b->lock, flags);
7891 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7893 rcu_read_unlock_sched();
7897 static int sched_rt_global_validate(void)
7899 if (sysctl_sched_rt_period <= 0)
7902 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7903 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7909 static void sched_rt_do_global(void)
7911 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7912 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7915 int sched_rt_handler(struct ctl_table *table, int write,
7916 void __user *buffer, size_t *lenp,
7919 int old_period, old_runtime;
7920 static DEFINE_MUTEX(mutex);
7924 old_period = sysctl_sched_rt_period;
7925 old_runtime = sysctl_sched_rt_runtime;
7927 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7929 if (!ret && write) {
7930 ret = sched_rt_global_validate();
7934 ret = sched_rt_global_constraints();
7938 ret = sched_dl_global_constraints();
7942 sched_rt_do_global();
7943 sched_dl_do_global();
7947 sysctl_sched_rt_period = old_period;
7948 sysctl_sched_rt_runtime = old_runtime;
7950 mutex_unlock(&mutex);
7955 int sched_rr_handler(struct ctl_table *table, int write,
7956 void __user *buffer, size_t *lenp,
7960 static DEFINE_MUTEX(mutex);
7963 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7964 /* make sure that internally we keep jiffies */
7965 /* also, writing zero resets timeslice to default */
7966 if (!ret && write) {
7967 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7968 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7970 mutex_unlock(&mutex);
7974 #ifdef CONFIG_CGROUP_SCHED
7976 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7978 return css ? container_of(css, struct task_group, css) : NULL;
7981 static struct cgroup_subsys_state *
7982 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7984 struct task_group *parent = css_tg(parent_css);
7985 struct task_group *tg;
7988 /* This is early initialization for the top cgroup */
7989 return &root_task_group.css;
7992 tg = sched_create_group(parent);
7994 return ERR_PTR(-ENOMEM);
7999 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8001 struct task_group *tg = css_tg(css);
8002 struct task_group *parent = css_tg(css->parent);
8005 sched_online_group(tg, parent);
8009 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8011 struct task_group *tg = css_tg(css);
8013 sched_destroy_group(tg);
8016 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8018 struct task_group *tg = css_tg(css);
8020 sched_offline_group(tg);
8023 static void cpu_cgroup_fork(struct task_struct *task)
8025 sched_move_task(task);
8028 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8029 struct cgroup_taskset *tset)
8031 struct task_struct *task;
8033 cgroup_taskset_for_each(task, tset) {
8034 #ifdef CONFIG_RT_GROUP_SCHED
8035 if (!sched_rt_can_attach(css_tg(css), task))
8038 /* We don't support RT-tasks being in separate groups */
8039 if (task->sched_class != &fair_sched_class)
8046 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8047 struct cgroup_taskset *tset)
8049 struct task_struct *task;
8051 cgroup_taskset_for_each(task, tset)
8052 sched_move_task(task);
8055 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8056 struct cgroup_subsys_state *old_css,
8057 struct task_struct *task)
8060 * cgroup_exit() is called in the copy_process() failure path.
8061 * Ignore this case since the task hasn't ran yet, this avoids
8062 * trying to poke a half freed task state from generic code.
8064 if (!(task->flags & PF_EXITING))
8067 sched_move_task(task);
8070 #ifdef CONFIG_FAIR_GROUP_SCHED
8071 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8072 struct cftype *cftype, u64 shareval)
8074 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8077 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8080 struct task_group *tg = css_tg(css);
8082 return (u64) scale_load_down(tg->shares);
8085 #ifdef CONFIG_CFS_BANDWIDTH
8086 static DEFINE_MUTEX(cfs_constraints_mutex);
8088 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8089 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8091 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8093 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8095 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8096 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8098 if (tg == &root_task_group)
8102 * Ensure we have at some amount of bandwidth every period. This is
8103 * to prevent reaching a state of large arrears when throttled via
8104 * entity_tick() resulting in prolonged exit starvation.
8106 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8110 * Likewise, bound things on the otherside by preventing insane quota
8111 * periods. This also allows us to normalize in computing quota
8114 if (period > max_cfs_quota_period)
8118 * Prevent race between setting of cfs_rq->runtime_enabled and
8119 * unthrottle_offline_cfs_rqs().
8122 mutex_lock(&cfs_constraints_mutex);
8123 ret = __cfs_schedulable(tg, period, quota);
8127 runtime_enabled = quota != RUNTIME_INF;
8128 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8130 * If we need to toggle cfs_bandwidth_used, off->on must occur
8131 * before making related changes, and on->off must occur afterwards
8133 if (runtime_enabled && !runtime_was_enabled)
8134 cfs_bandwidth_usage_inc();
8135 raw_spin_lock_irq(&cfs_b->lock);
8136 cfs_b->period = ns_to_ktime(period);
8137 cfs_b->quota = quota;
8139 __refill_cfs_bandwidth_runtime(cfs_b);
8140 /* restart the period timer (if active) to handle new period expiry */
8141 if (runtime_enabled && cfs_b->timer_active) {
8142 /* force a reprogram */
8143 __start_cfs_bandwidth(cfs_b, true);
8145 raw_spin_unlock_irq(&cfs_b->lock);
8147 for_each_online_cpu(i) {
8148 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8149 struct rq *rq = cfs_rq->rq;
8151 raw_spin_lock_irq(&rq->lock);
8152 cfs_rq->runtime_enabled = runtime_enabled;
8153 cfs_rq->runtime_remaining = 0;
8155 if (cfs_rq->throttled)
8156 unthrottle_cfs_rq(cfs_rq);
8157 raw_spin_unlock_irq(&rq->lock);
8159 if (runtime_was_enabled && !runtime_enabled)
8160 cfs_bandwidth_usage_dec();
8162 mutex_unlock(&cfs_constraints_mutex);
8168 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8172 period = ktime_to_ns(tg->cfs_bandwidth.period);
8173 if (cfs_quota_us < 0)
8174 quota = RUNTIME_INF;
8176 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8178 return tg_set_cfs_bandwidth(tg, period, quota);
8181 long tg_get_cfs_quota(struct task_group *tg)
8185 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8188 quota_us = tg->cfs_bandwidth.quota;
8189 do_div(quota_us, NSEC_PER_USEC);
8194 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8198 period = (u64)cfs_period_us * NSEC_PER_USEC;
8199 quota = tg->cfs_bandwidth.quota;
8201 return tg_set_cfs_bandwidth(tg, period, quota);
8204 long tg_get_cfs_period(struct task_group *tg)
8208 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8209 do_div(cfs_period_us, NSEC_PER_USEC);
8211 return cfs_period_us;
8214 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8217 return tg_get_cfs_quota(css_tg(css));
8220 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8221 struct cftype *cftype, s64 cfs_quota_us)
8223 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8226 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8229 return tg_get_cfs_period(css_tg(css));
8232 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8233 struct cftype *cftype, u64 cfs_period_us)
8235 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8238 struct cfs_schedulable_data {
8239 struct task_group *tg;
8244 * normalize group quota/period to be quota/max_period
8245 * note: units are usecs
8247 static u64 normalize_cfs_quota(struct task_group *tg,
8248 struct cfs_schedulable_data *d)
8256 period = tg_get_cfs_period(tg);
8257 quota = tg_get_cfs_quota(tg);
8260 /* note: these should typically be equivalent */
8261 if (quota == RUNTIME_INF || quota == -1)
8264 return to_ratio(period, quota);
8267 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8269 struct cfs_schedulable_data *d = data;
8270 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8271 s64 quota = 0, parent_quota = -1;
8274 quota = RUNTIME_INF;
8276 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8278 quota = normalize_cfs_quota(tg, d);
8279 parent_quota = parent_b->hierarchical_quota;
8282 * ensure max(child_quota) <= parent_quota, inherit when no
8285 if (quota == RUNTIME_INF)
8286 quota = parent_quota;
8287 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8290 cfs_b->hierarchical_quota = quota;
8295 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8298 struct cfs_schedulable_data data = {
8304 if (quota != RUNTIME_INF) {
8305 do_div(data.period, NSEC_PER_USEC);
8306 do_div(data.quota, NSEC_PER_USEC);
8310 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8316 static int cpu_stats_show(struct seq_file *sf, void *v)
8318 struct task_group *tg = css_tg(seq_css(sf));
8319 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8321 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8322 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8323 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8327 #endif /* CONFIG_CFS_BANDWIDTH */
8328 #endif /* CONFIG_FAIR_GROUP_SCHED */
8330 #ifdef CONFIG_RT_GROUP_SCHED
8331 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8332 struct cftype *cft, s64 val)
8334 return sched_group_set_rt_runtime(css_tg(css), val);
8337 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8340 return sched_group_rt_runtime(css_tg(css));
8343 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8344 struct cftype *cftype, u64 rt_period_us)
8346 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8349 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8352 return sched_group_rt_period(css_tg(css));
8354 #endif /* CONFIG_RT_GROUP_SCHED */
8356 static struct cftype cpu_files[] = {
8357 #ifdef CONFIG_FAIR_GROUP_SCHED
8360 .read_u64 = cpu_shares_read_u64,
8361 .write_u64 = cpu_shares_write_u64,
8364 #ifdef CONFIG_CFS_BANDWIDTH
8366 .name = "cfs_quota_us",
8367 .read_s64 = cpu_cfs_quota_read_s64,
8368 .write_s64 = cpu_cfs_quota_write_s64,
8371 .name = "cfs_period_us",
8372 .read_u64 = cpu_cfs_period_read_u64,
8373 .write_u64 = cpu_cfs_period_write_u64,
8377 .seq_show = cpu_stats_show,
8380 #ifdef CONFIG_RT_GROUP_SCHED
8382 .name = "rt_runtime_us",
8383 .read_s64 = cpu_rt_runtime_read,
8384 .write_s64 = cpu_rt_runtime_write,
8387 .name = "rt_period_us",
8388 .read_u64 = cpu_rt_period_read_uint,
8389 .write_u64 = cpu_rt_period_write_uint,
8395 struct cgroup_subsys cpu_cgrp_subsys = {
8396 .css_alloc = cpu_cgroup_css_alloc,
8397 .css_free = cpu_cgroup_css_free,
8398 .css_online = cpu_cgroup_css_online,
8399 .css_offline = cpu_cgroup_css_offline,
8400 .fork = cpu_cgroup_fork,
8401 .can_attach = cpu_cgroup_can_attach,
8402 .attach = cpu_cgroup_attach,
8403 .exit = cpu_cgroup_exit,
8404 .legacy_cftypes = cpu_files,
8408 #endif /* CONFIG_CGROUP_SCHED */
8410 void dump_cpu_task(int cpu)
8412 pr_info("Task dump for CPU %d:\n", cpu);
8413 sched_show_task(cpu_curr(cpu));