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 * this_rq_lock - lock this runqueue and disable interrupts.
312 static struct rq *this_rq_lock(void)
319 raw_spin_lock(&rq->lock);
324 #ifdef CONFIG_SCHED_HRTICK
326 * Use HR-timers to deliver accurate preemption points.
329 static void hrtick_clear(struct rq *rq)
331 if (hrtimer_active(&rq->hrtick_timer))
332 hrtimer_cancel(&rq->hrtick_timer);
336 * High-resolution timer tick.
337 * Runs from hardirq context with interrupts disabled.
339 static enum hrtimer_restart hrtick(struct hrtimer *timer)
341 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
343 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
345 raw_spin_lock(&rq->lock);
347 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
348 raw_spin_unlock(&rq->lock);
350 return HRTIMER_NORESTART;
355 static int __hrtick_restart(struct rq *rq)
357 struct hrtimer *timer = &rq->hrtick_timer;
358 ktime_t time = hrtimer_get_softexpires(timer);
360 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
364 * called from hardirq (IPI) context
366 static void __hrtick_start(void *arg)
370 raw_spin_lock(&rq->lock);
371 __hrtick_restart(rq);
372 rq->hrtick_csd_pending = 0;
373 raw_spin_unlock(&rq->lock);
377 * Called to set the hrtick timer state.
379 * called with rq->lock held and irqs disabled
381 void hrtick_start(struct rq *rq, u64 delay)
383 struct hrtimer *timer = &rq->hrtick_timer;
388 * Don't schedule slices shorter than 10000ns, that just
389 * doesn't make sense and can cause timer DoS.
391 delta = max_t(s64, delay, 10000LL);
392 time = ktime_add_ns(timer->base->get_time(), delta);
394 hrtimer_set_expires(timer, time);
396 if (rq == this_rq()) {
397 __hrtick_restart(rq);
398 } else if (!rq->hrtick_csd_pending) {
399 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
400 rq->hrtick_csd_pending = 1;
405 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
407 int cpu = (int)(long)hcpu;
410 case CPU_UP_CANCELED:
411 case CPU_UP_CANCELED_FROZEN:
412 case CPU_DOWN_PREPARE:
413 case CPU_DOWN_PREPARE_FROZEN:
415 case CPU_DEAD_FROZEN:
416 hrtick_clear(cpu_rq(cpu));
423 static __init void init_hrtick(void)
425 hotcpu_notifier(hotplug_hrtick, 0);
429 * Called to set the hrtick timer state.
431 * called with rq->lock held and irqs disabled
433 void hrtick_start(struct rq *rq, u64 delay)
436 * Don't schedule slices shorter than 10000ns, that just
437 * doesn't make sense. Rely on vruntime for fairness.
439 delay = max_t(u64, delay, 10000LL);
440 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
441 HRTIMER_MODE_REL_PINNED, 0);
444 static inline void init_hrtick(void)
447 #endif /* CONFIG_SMP */
449 static void init_rq_hrtick(struct rq *rq)
452 rq->hrtick_csd_pending = 0;
454 rq->hrtick_csd.flags = 0;
455 rq->hrtick_csd.func = __hrtick_start;
456 rq->hrtick_csd.info = rq;
459 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
460 rq->hrtick_timer.function = hrtick;
462 #else /* CONFIG_SCHED_HRTICK */
463 static inline void hrtick_clear(struct rq *rq)
467 static inline void init_rq_hrtick(struct rq *rq)
471 static inline void init_hrtick(void)
474 #endif /* CONFIG_SCHED_HRTICK */
477 * cmpxchg based fetch_or, macro so it works for different integer types
479 #define fetch_or(ptr, val) \
480 ({ typeof(*(ptr)) __old, __val = *(ptr); \
482 __old = cmpxchg((ptr), __val, __val | (val)); \
483 if (__old == __val) \
490 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
492 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
493 * this avoids any races wrt polling state changes and thereby avoids
496 static bool set_nr_and_not_polling(struct task_struct *p)
498 struct thread_info *ti = task_thread_info(p);
499 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
503 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
505 * If this returns true, then the idle task promises to call
506 * sched_ttwu_pending() and reschedule soon.
508 static bool set_nr_if_polling(struct task_struct *p)
510 struct thread_info *ti = task_thread_info(p);
511 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
514 if (!(val & _TIF_POLLING_NRFLAG))
516 if (val & _TIF_NEED_RESCHED)
518 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
527 static bool set_nr_and_not_polling(struct task_struct *p)
529 set_tsk_need_resched(p);
534 static bool set_nr_if_polling(struct task_struct *p)
542 * resched_curr - mark rq's current task 'to be rescheduled now'.
544 * On UP this means the setting of the need_resched flag, on SMP it
545 * might also involve a cross-CPU call to trigger the scheduler on
548 void resched_curr(struct rq *rq)
550 struct task_struct *curr = rq->curr;
553 lockdep_assert_held(&rq->lock);
555 if (test_tsk_need_resched(curr))
560 if (cpu == smp_processor_id()) {
561 set_tsk_need_resched(curr);
562 set_preempt_need_resched();
566 if (set_nr_and_not_polling(curr))
567 smp_send_reschedule(cpu);
569 trace_sched_wake_idle_without_ipi(cpu);
572 void resched_cpu(int cpu)
574 struct rq *rq = cpu_rq(cpu);
577 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
580 raw_spin_unlock_irqrestore(&rq->lock, flags);
584 #ifdef CONFIG_NO_HZ_COMMON
586 * In the semi idle case, use the nearest busy cpu for migrating timers
587 * from an idle cpu. This is good for power-savings.
589 * We don't do similar optimization for completely idle system, as
590 * selecting an idle cpu will add more delays to the timers than intended
591 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
593 int get_nohz_timer_target(int pinned)
595 int cpu = smp_processor_id();
597 struct sched_domain *sd;
599 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
603 for_each_domain(cpu, sd) {
604 for_each_cpu(i, sched_domain_span(sd)) {
616 * When add_timer_on() enqueues a timer into the timer wheel of an
617 * idle CPU then this timer might expire before the next timer event
618 * which is scheduled to wake up that CPU. In case of a completely
619 * idle system the next event might even be infinite time into the
620 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
621 * leaves the inner idle loop so the newly added timer is taken into
622 * account when the CPU goes back to idle and evaluates the timer
623 * wheel for the next timer event.
625 static void wake_up_idle_cpu(int cpu)
627 struct rq *rq = cpu_rq(cpu);
629 if (cpu == smp_processor_id())
632 if (set_nr_and_not_polling(rq->idle))
633 smp_send_reschedule(cpu);
635 trace_sched_wake_idle_without_ipi(cpu);
638 static bool wake_up_full_nohz_cpu(int cpu)
641 * We just need the target to call irq_exit() and re-evaluate
642 * the next tick. The nohz full kick at least implies that.
643 * If needed we can still optimize that later with an
646 if (tick_nohz_full_cpu(cpu)) {
647 if (cpu != smp_processor_id() ||
648 tick_nohz_tick_stopped())
649 tick_nohz_full_kick_cpu(cpu);
656 void wake_up_nohz_cpu(int cpu)
658 if (!wake_up_full_nohz_cpu(cpu))
659 wake_up_idle_cpu(cpu);
662 static inline bool got_nohz_idle_kick(void)
664 int cpu = smp_processor_id();
666 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
669 if (idle_cpu(cpu) && !need_resched())
673 * We can't run Idle Load Balance on this CPU for this time so we
674 * cancel it and clear NOHZ_BALANCE_KICK
676 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
680 #else /* CONFIG_NO_HZ_COMMON */
682 static inline bool got_nohz_idle_kick(void)
687 #endif /* CONFIG_NO_HZ_COMMON */
689 #ifdef CONFIG_NO_HZ_FULL
690 bool sched_can_stop_tick(void)
693 * More than one running task need preemption.
694 * nr_running update is assumed to be visible
695 * after IPI is sent from wakers.
697 if (this_rq()->nr_running > 1)
702 #endif /* CONFIG_NO_HZ_FULL */
704 void sched_avg_update(struct rq *rq)
706 s64 period = sched_avg_period();
708 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
710 * Inline assembly required to prevent the compiler
711 * optimising this loop into a divmod call.
712 * See __iter_div_u64_rem() for another example of this.
714 asm("" : "+rm" (rq->age_stamp));
715 rq->age_stamp += period;
720 #endif /* CONFIG_SMP */
722 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
723 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
725 * Iterate task_group tree rooted at *from, calling @down when first entering a
726 * node and @up when leaving it for the final time.
728 * Caller must hold rcu_lock or sufficient equivalent.
730 int walk_tg_tree_from(struct task_group *from,
731 tg_visitor down, tg_visitor up, void *data)
733 struct task_group *parent, *child;
739 ret = (*down)(parent, data);
742 list_for_each_entry_rcu(child, &parent->children, siblings) {
749 ret = (*up)(parent, data);
750 if (ret || parent == from)
754 parent = parent->parent;
761 int tg_nop(struct task_group *tg, void *data)
767 static void set_load_weight(struct task_struct *p)
769 int prio = p->static_prio - MAX_RT_PRIO;
770 struct load_weight *load = &p->se.load;
773 * SCHED_IDLE tasks get minimal weight:
775 if (p->policy == SCHED_IDLE) {
776 load->weight = scale_load(WEIGHT_IDLEPRIO);
777 load->inv_weight = WMULT_IDLEPRIO;
781 load->weight = scale_load(prio_to_weight[prio]);
782 load->inv_weight = prio_to_wmult[prio];
785 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
788 sched_info_queued(rq, p);
789 p->sched_class->enqueue_task(rq, p, flags);
792 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
795 sched_info_dequeued(rq, p);
796 p->sched_class->dequeue_task(rq, p, flags);
799 void activate_task(struct rq *rq, struct task_struct *p, int flags)
801 if (task_contributes_to_load(p))
802 rq->nr_uninterruptible--;
804 enqueue_task(rq, p, flags);
807 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
809 if (task_contributes_to_load(p))
810 rq->nr_uninterruptible++;
812 dequeue_task(rq, p, flags);
815 static void update_rq_clock_task(struct rq *rq, s64 delta)
818 * In theory, the compile should just see 0 here, and optimize out the call
819 * to sched_rt_avg_update. But I don't trust it...
821 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
822 s64 steal = 0, irq_delta = 0;
824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
825 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
828 * Since irq_time is only updated on {soft,}irq_exit, we might run into
829 * this case when a previous update_rq_clock() happened inside a
832 * When this happens, we stop ->clock_task and only update the
833 * prev_irq_time stamp to account for the part that fit, so that a next
834 * update will consume the rest. This ensures ->clock_task is
837 * It does however cause some slight miss-attribution of {soft,}irq
838 * time, a more accurate solution would be to update the irq_time using
839 * the current rq->clock timestamp, except that would require using
842 if (irq_delta > delta)
845 rq->prev_irq_time += irq_delta;
848 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
849 if (static_key_false((¶virt_steal_rq_enabled))) {
850 steal = paravirt_steal_clock(cpu_of(rq));
851 steal -= rq->prev_steal_time_rq;
853 if (unlikely(steal > delta))
856 rq->prev_steal_time_rq += steal;
861 rq->clock_task += delta;
863 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
864 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
865 sched_rt_avg_update(rq, irq_delta + steal);
869 void sched_set_stop_task(int cpu, struct task_struct *stop)
871 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
872 struct task_struct *old_stop = cpu_rq(cpu)->stop;
876 * Make it appear like a SCHED_FIFO task, its something
877 * userspace knows about and won't get confused about.
879 * Also, it will make PI more or less work without too
880 * much confusion -- but then, stop work should not
881 * rely on PI working anyway.
883 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
885 stop->sched_class = &stop_sched_class;
888 cpu_rq(cpu)->stop = stop;
892 * Reset it back to a normal scheduling class so that
893 * it can die in pieces.
895 old_stop->sched_class = &rt_sched_class;
900 * __normal_prio - return the priority that is based on the static prio
902 static inline int __normal_prio(struct task_struct *p)
904 return p->static_prio;
908 * Calculate the expected normal priority: i.e. priority
909 * without taking RT-inheritance into account. Might be
910 * boosted by interactivity modifiers. Changes upon fork,
911 * setprio syscalls, and whenever the interactivity
912 * estimator recalculates.
914 static inline int normal_prio(struct task_struct *p)
918 if (task_has_dl_policy(p))
919 prio = MAX_DL_PRIO-1;
920 else if (task_has_rt_policy(p))
921 prio = MAX_RT_PRIO-1 - p->rt_priority;
923 prio = __normal_prio(p);
928 * Calculate the current priority, i.e. the priority
929 * taken into account by the scheduler. This value might
930 * be boosted by RT tasks, or might be boosted by
931 * interactivity modifiers. Will be RT if the task got
932 * RT-boosted. If not then it returns p->normal_prio.
934 static int effective_prio(struct task_struct *p)
936 p->normal_prio = normal_prio(p);
938 * If we are RT tasks or we were boosted to RT priority,
939 * keep the priority unchanged. Otherwise, update priority
940 * to the normal priority:
942 if (!rt_prio(p->prio))
943 return p->normal_prio;
948 * task_curr - is this task currently executing on a CPU?
949 * @p: the task in question.
951 * Return: 1 if the task is currently executing. 0 otherwise.
953 inline int task_curr(const struct task_struct *p)
955 return cpu_curr(task_cpu(p)) == p;
959 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
961 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
962 const struct sched_class *prev_class,
965 if (prev_class != p->sched_class) {
966 if (prev_class->switched_from)
967 prev_class->switched_from(rq, p);
968 /* Possble rq->lock 'hole'. */
969 p->sched_class->switched_to(rq, p);
970 } else if (oldprio != p->prio || dl_task(p))
971 p->sched_class->prio_changed(rq, p, oldprio);
974 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
976 const struct sched_class *class;
978 if (p->sched_class == rq->curr->sched_class) {
979 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
981 for_each_class(class) {
982 if (class == rq->curr->sched_class)
984 if (class == p->sched_class) {
992 * A queue event has occurred, and we're going to schedule. In
993 * this case, we can save a useless back to back clock update.
995 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
996 rq_clock_skip_update(rq, true);
1000 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1002 #ifdef CONFIG_SCHED_DEBUG
1004 * We should never call set_task_cpu() on a blocked task,
1005 * ttwu() will sort out the placement.
1007 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1010 #ifdef CONFIG_LOCKDEP
1012 * The caller should hold either p->pi_lock or rq->lock, when changing
1013 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1015 * sched_move_task() holds both and thus holding either pins the cgroup,
1018 * Furthermore, all task_rq users should acquire both locks, see
1021 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1022 lockdep_is_held(&task_rq(p)->lock)));
1026 trace_sched_migrate_task(p, new_cpu);
1028 if (task_cpu(p) != new_cpu) {
1029 if (p->sched_class->migrate_task_rq)
1030 p->sched_class->migrate_task_rq(p, new_cpu);
1031 p->se.nr_migrations++;
1032 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1035 __set_task_cpu(p, new_cpu);
1038 static void __migrate_swap_task(struct task_struct *p, int cpu)
1040 if (task_on_rq_queued(p)) {
1041 struct rq *src_rq, *dst_rq;
1043 src_rq = task_rq(p);
1044 dst_rq = cpu_rq(cpu);
1046 deactivate_task(src_rq, p, 0);
1047 set_task_cpu(p, cpu);
1048 activate_task(dst_rq, p, 0);
1049 check_preempt_curr(dst_rq, p, 0);
1052 * Task isn't running anymore; make it appear like we migrated
1053 * it before it went to sleep. This means on wakeup we make the
1054 * previous cpu our targer instead of where it really is.
1060 struct migration_swap_arg {
1061 struct task_struct *src_task, *dst_task;
1062 int src_cpu, dst_cpu;
1065 static int migrate_swap_stop(void *data)
1067 struct migration_swap_arg *arg = data;
1068 struct rq *src_rq, *dst_rq;
1071 src_rq = cpu_rq(arg->src_cpu);
1072 dst_rq = cpu_rq(arg->dst_cpu);
1074 double_raw_lock(&arg->src_task->pi_lock,
1075 &arg->dst_task->pi_lock);
1076 double_rq_lock(src_rq, dst_rq);
1077 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1080 if (task_cpu(arg->src_task) != arg->src_cpu)
1083 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1086 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1089 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1090 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1095 double_rq_unlock(src_rq, dst_rq);
1096 raw_spin_unlock(&arg->dst_task->pi_lock);
1097 raw_spin_unlock(&arg->src_task->pi_lock);
1103 * Cross migrate two tasks
1105 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1107 struct migration_swap_arg arg;
1110 arg = (struct migration_swap_arg){
1112 .src_cpu = task_cpu(cur),
1114 .dst_cpu = task_cpu(p),
1117 if (arg.src_cpu == arg.dst_cpu)
1121 * These three tests are all lockless; this is OK since all of them
1122 * will be re-checked with proper locks held further down the line.
1124 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1127 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1130 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1133 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1134 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1140 struct migration_arg {
1141 struct task_struct *task;
1145 static int migration_cpu_stop(void *data);
1148 * wait_task_inactive - wait for a thread to unschedule.
1150 * If @match_state is nonzero, it's the @p->state value just checked and
1151 * not expected to change. If it changes, i.e. @p might have woken up,
1152 * then return zero. When we succeed in waiting for @p to be off its CPU,
1153 * we return a positive number (its total switch count). If a second call
1154 * a short while later returns the same number, the caller can be sure that
1155 * @p has remained unscheduled the whole time.
1157 * The caller must ensure that the task *will* unschedule sometime soon,
1158 * else this function might spin for a *long* time. This function can't
1159 * be called with interrupts off, or it may introduce deadlock with
1160 * smp_call_function() if an IPI is sent by the same process we are
1161 * waiting to become inactive.
1163 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1165 unsigned long flags;
1166 int running, queued;
1172 * We do the initial early heuristics without holding
1173 * any task-queue locks at all. We'll only try to get
1174 * the runqueue lock when things look like they will
1180 * If the task is actively running on another CPU
1181 * still, just relax and busy-wait without holding
1184 * NOTE! Since we don't hold any locks, it's not
1185 * even sure that "rq" stays as the right runqueue!
1186 * But we don't care, since "task_running()" will
1187 * return false if the runqueue has changed and p
1188 * is actually now running somewhere else!
1190 while (task_running(rq, p)) {
1191 if (match_state && unlikely(p->state != match_state))
1197 * Ok, time to look more closely! We need the rq
1198 * lock now, to be *sure*. If we're wrong, we'll
1199 * just go back and repeat.
1201 rq = task_rq_lock(p, &flags);
1202 trace_sched_wait_task(p);
1203 running = task_running(rq, p);
1204 queued = task_on_rq_queued(p);
1206 if (!match_state || p->state == match_state)
1207 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1208 task_rq_unlock(rq, p, &flags);
1211 * If it changed from the expected state, bail out now.
1213 if (unlikely(!ncsw))
1217 * Was it really running after all now that we
1218 * checked with the proper locks actually held?
1220 * Oops. Go back and try again..
1222 if (unlikely(running)) {
1228 * It's not enough that it's not actively running,
1229 * it must be off the runqueue _entirely_, and not
1232 * So if it was still runnable (but just not actively
1233 * running right now), it's preempted, and we should
1234 * yield - it could be a while.
1236 if (unlikely(queued)) {
1237 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1239 set_current_state(TASK_UNINTERRUPTIBLE);
1240 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1245 * Ahh, all good. It wasn't running, and it wasn't
1246 * runnable, which means that it will never become
1247 * running in the future either. We're all done!
1256 * kick_process - kick a running thread to enter/exit the kernel
1257 * @p: the to-be-kicked thread
1259 * Cause a process which is running on another CPU to enter
1260 * kernel-mode, without any delay. (to get signals handled.)
1262 * NOTE: this function doesn't have to take the runqueue lock,
1263 * because all it wants to ensure is that the remote task enters
1264 * the kernel. If the IPI races and the task has been migrated
1265 * to another CPU then no harm is done and the purpose has been
1268 void kick_process(struct task_struct *p)
1274 if ((cpu != smp_processor_id()) && task_curr(p))
1275 smp_send_reschedule(cpu);
1278 EXPORT_SYMBOL_GPL(kick_process);
1279 #endif /* CONFIG_SMP */
1283 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1285 static int select_fallback_rq(int cpu, struct task_struct *p)
1287 int nid = cpu_to_node(cpu);
1288 const struct cpumask *nodemask = NULL;
1289 enum { cpuset, possible, fail } state = cpuset;
1293 * If the node that the cpu is on has been offlined, cpu_to_node()
1294 * will return -1. There is no cpu on the node, and we should
1295 * select the cpu on the other node.
1298 nodemask = cpumask_of_node(nid);
1300 /* Look for allowed, online CPU in same node. */
1301 for_each_cpu(dest_cpu, nodemask) {
1302 if (!cpu_online(dest_cpu))
1304 if (!cpu_active(dest_cpu))
1306 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1312 /* Any allowed, online CPU? */
1313 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1314 if (!cpu_online(dest_cpu))
1316 if (!cpu_active(dest_cpu))
1323 /* No more Mr. Nice Guy. */
1324 cpuset_cpus_allowed_fallback(p);
1329 do_set_cpus_allowed(p, cpu_possible_mask);
1340 if (state != cpuset) {
1342 * Don't tell them about moving exiting tasks or
1343 * kernel threads (both mm NULL), since they never
1346 if (p->mm && printk_ratelimit()) {
1347 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1348 task_pid_nr(p), p->comm, cpu);
1356 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1359 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1361 if (p->nr_cpus_allowed > 1)
1362 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1365 * In order not to call set_task_cpu() on a blocking task we need
1366 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1369 * Since this is common to all placement strategies, this lives here.
1371 * [ this allows ->select_task() to simply return task_cpu(p) and
1372 * not worry about this generic constraint ]
1374 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1376 cpu = select_fallback_rq(task_cpu(p), p);
1381 static void update_avg(u64 *avg, u64 sample)
1383 s64 diff = sample - *avg;
1389 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1391 #ifdef CONFIG_SCHEDSTATS
1392 struct rq *rq = this_rq();
1395 int this_cpu = smp_processor_id();
1397 if (cpu == this_cpu) {
1398 schedstat_inc(rq, ttwu_local);
1399 schedstat_inc(p, se.statistics.nr_wakeups_local);
1401 struct sched_domain *sd;
1403 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1405 for_each_domain(this_cpu, sd) {
1406 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1407 schedstat_inc(sd, ttwu_wake_remote);
1414 if (wake_flags & WF_MIGRATED)
1415 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1417 #endif /* CONFIG_SMP */
1419 schedstat_inc(rq, ttwu_count);
1420 schedstat_inc(p, se.statistics.nr_wakeups);
1422 if (wake_flags & WF_SYNC)
1423 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1425 #endif /* CONFIG_SCHEDSTATS */
1428 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1430 activate_task(rq, p, en_flags);
1431 p->on_rq = TASK_ON_RQ_QUEUED;
1433 /* if a worker is waking up, notify workqueue */
1434 if (p->flags & PF_WQ_WORKER)
1435 wq_worker_waking_up(p, cpu_of(rq));
1439 * Mark the task runnable and perform wakeup-preemption.
1442 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1444 check_preempt_curr(rq, p, wake_flags);
1445 trace_sched_wakeup(p, true);
1447 p->state = TASK_RUNNING;
1449 if (p->sched_class->task_woken)
1450 p->sched_class->task_woken(rq, p);
1452 if (rq->idle_stamp) {
1453 u64 delta = rq_clock(rq) - rq->idle_stamp;
1454 u64 max = 2*rq->max_idle_balance_cost;
1456 update_avg(&rq->avg_idle, delta);
1458 if (rq->avg_idle > max)
1467 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1470 if (p->sched_contributes_to_load)
1471 rq->nr_uninterruptible--;
1474 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1475 ttwu_do_wakeup(rq, p, wake_flags);
1479 * Called in case the task @p isn't fully descheduled from its runqueue,
1480 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1481 * since all we need to do is flip p->state to TASK_RUNNING, since
1482 * the task is still ->on_rq.
1484 static int ttwu_remote(struct task_struct *p, int wake_flags)
1489 rq = __task_rq_lock(p);
1490 if (task_on_rq_queued(p)) {
1491 /* check_preempt_curr() may use rq clock */
1492 update_rq_clock(rq);
1493 ttwu_do_wakeup(rq, p, wake_flags);
1496 __task_rq_unlock(rq);
1502 void sched_ttwu_pending(void)
1504 struct rq *rq = this_rq();
1505 struct llist_node *llist = llist_del_all(&rq->wake_list);
1506 struct task_struct *p;
1507 unsigned long flags;
1512 raw_spin_lock_irqsave(&rq->lock, flags);
1515 p = llist_entry(llist, struct task_struct, wake_entry);
1516 llist = llist_next(llist);
1517 ttwu_do_activate(rq, p, 0);
1520 raw_spin_unlock_irqrestore(&rq->lock, flags);
1523 void scheduler_ipi(void)
1526 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1527 * TIF_NEED_RESCHED remotely (for the first time) will also send
1530 preempt_fold_need_resched();
1532 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1536 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1537 * traditionally all their work was done from the interrupt return
1538 * path. Now that we actually do some work, we need to make sure
1541 * Some archs already do call them, luckily irq_enter/exit nest
1544 * Arguably we should visit all archs and update all handlers,
1545 * however a fair share of IPIs are still resched only so this would
1546 * somewhat pessimize the simple resched case.
1549 sched_ttwu_pending();
1552 * Check if someone kicked us for doing the nohz idle load balance.
1554 if (unlikely(got_nohz_idle_kick())) {
1555 this_rq()->idle_balance = 1;
1556 raise_softirq_irqoff(SCHED_SOFTIRQ);
1561 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1563 struct rq *rq = cpu_rq(cpu);
1565 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1566 if (!set_nr_if_polling(rq->idle))
1567 smp_send_reschedule(cpu);
1569 trace_sched_wake_idle_without_ipi(cpu);
1573 void wake_up_if_idle(int cpu)
1575 struct rq *rq = cpu_rq(cpu);
1576 unsigned long flags;
1580 if (!is_idle_task(rcu_dereference(rq->curr)))
1583 if (set_nr_if_polling(rq->idle)) {
1584 trace_sched_wake_idle_without_ipi(cpu);
1586 raw_spin_lock_irqsave(&rq->lock, flags);
1587 if (is_idle_task(rq->curr))
1588 smp_send_reschedule(cpu);
1589 /* Else cpu is not in idle, do nothing here */
1590 raw_spin_unlock_irqrestore(&rq->lock, flags);
1597 bool cpus_share_cache(int this_cpu, int that_cpu)
1599 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1601 #endif /* CONFIG_SMP */
1603 static void ttwu_queue(struct task_struct *p, int cpu)
1605 struct rq *rq = cpu_rq(cpu);
1607 #if defined(CONFIG_SMP)
1608 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1609 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1610 ttwu_queue_remote(p, cpu);
1615 raw_spin_lock(&rq->lock);
1616 ttwu_do_activate(rq, p, 0);
1617 raw_spin_unlock(&rq->lock);
1621 * try_to_wake_up - wake up a thread
1622 * @p: the thread to be awakened
1623 * @state: the mask of task states that can be woken
1624 * @wake_flags: wake modifier flags (WF_*)
1626 * Put it on the run-queue if it's not already there. The "current"
1627 * thread is always on the run-queue (except when the actual
1628 * re-schedule is in progress), and as such you're allowed to do
1629 * the simpler "current->state = TASK_RUNNING" to mark yourself
1630 * runnable without the overhead of this.
1632 * Return: %true if @p was woken up, %false if it was already running.
1633 * or @state didn't match @p's state.
1636 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1638 unsigned long flags;
1639 int cpu, success = 0;
1642 * If we are going to wake up a thread waiting for CONDITION we
1643 * need to ensure that CONDITION=1 done by the caller can not be
1644 * reordered with p->state check below. This pairs with mb() in
1645 * set_current_state() the waiting thread does.
1647 smp_mb__before_spinlock();
1648 raw_spin_lock_irqsave(&p->pi_lock, flags);
1649 if (!(p->state & state))
1652 success = 1; /* we're going to change ->state */
1655 if (p->on_rq && ttwu_remote(p, wake_flags))
1660 * If the owning (remote) cpu is still in the middle of schedule() with
1661 * this task as prev, wait until its done referencing the task.
1666 * Pairs with the smp_wmb() in finish_lock_switch().
1670 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1671 p->state = TASK_WAKING;
1673 if (p->sched_class->task_waking)
1674 p->sched_class->task_waking(p);
1676 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1677 if (task_cpu(p) != cpu) {
1678 wake_flags |= WF_MIGRATED;
1679 set_task_cpu(p, cpu);
1681 #endif /* CONFIG_SMP */
1685 ttwu_stat(p, cpu, wake_flags);
1687 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1693 * try_to_wake_up_local - try to wake up a local task with rq lock held
1694 * @p: the thread to be awakened
1696 * Put @p on the run-queue if it's not already there. The caller must
1697 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1700 static void try_to_wake_up_local(struct task_struct *p)
1702 struct rq *rq = task_rq(p);
1704 if (WARN_ON_ONCE(rq != this_rq()) ||
1705 WARN_ON_ONCE(p == current))
1708 lockdep_assert_held(&rq->lock);
1710 if (!raw_spin_trylock(&p->pi_lock)) {
1711 raw_spin_unlock(&rq->lock);
1712 raw_spin_lock(&p->pi_lock);
1713 raw_spin_lock(&rq->lock);
1716 if (!(p->state & TASK_NORMAL))
1719 if (!task_on_rq_queued(p))
1720 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1722 ttwu_do_wakeup(rq, p, 0);
1723 ttwu_stat(p, smp_processor_id(), 0);
1725 raw_spin_unlock(&p->pi_lock);
1729 * wake_up_process - Wake up a specific process
1730 * @p: The process to be woken up.
1732 * Attempt to wake up the nominated process and move it to the set of runnable
1735 * Return: 1 if the process was woken up, 0 if it was already running.
1737 * It may be assumed that this function implies a write memory barrier before
1738 * changing the task state if and only if any tasks are woken up.
1740 int wake_up_process(struct task_struct *p)
1742 WARN_ON(task_is_stopped_or_traced(p));
1743 return try_to_wake_up(p, TASK_NORMAL, 0);
1745 EXPORT_SYMBOL(wake_up_process);
1747 int wake_up_state(struct task_struct *p, unsigned int state)
1749 return try_to_wake_up(p, state, 0);
1753 * This function clears the sched_dl_entity static params.
1755 void __dl_clear_params(struct task_struct *p)
1757 struct sched_dl_entity *dl_se = &p->dl;
1759 dl_se->dl_runtime = 0;
1760 dl_se->dl_deadline = 0;
1761 dl_se->dl_period = 0;
1765 dl_se->dl_throttled = 0;
1767 dl_se->dl_yielded = 0;
1771 * Perform scheduler related setup for a newly forked process p.
1772 * p is forked by current.
1774 * __sched_fork() is basic setup used by init_idle() too:
1776 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1781 p->se.exec_start = 0;
1782 p->se.sum_exec_runtime = 0;
1783 p->se.prev_sum_exec_runtime = 0;
1784 p->se.nr_migrations = 0;
1787 p->se.avg.decay_count = 0;
1789 INIT_LIST_HEAD(&p->se.group_node);
1791 #ifdef CONFIG_SCHEDSTATS
1792 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1795 RB_CLEAR_NODE(&p->dl.rb_node);
1796 init_dl_task_timer(&p->dl);
1797 __dl_clear_params(p);
1799 INIT_LIST_HEAD(&p->rt.run_list);
1801 #ifdef CONFIG_PREEMPT_NOTIFIERS
1802 INIT_HLIST_HEAD(&p->preempt_notifiers);
1805 #ifdef CONFIG_NUMA_BALANCING
1806 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1807 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1808 p->mm->numa_scan_seq = 0;
1811 if (clone_flags & CLONE_VM)
1812 p->numa_preferred_nid = current->numa_preferred_nid;
1814 p->numa_preferred_nid = -1;
1816 p->node_stamp = 0ULL;
1817 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1818 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1819 p->numa_work.next = &p->numa_work;
1820 p->numa_faults = NULL;
1821 p->last_task_numa_placement = 0;
1822 p->last_sum_exec_runtime = 0;
1824 p->numa_group = NULL;
1825 #endif /* CONFIG_NUMA_BALANCING */
1828 #ifdef CONFIG_NUMA_BALANCING
1829 #ifdef CONFIG_SCHED_DEBUG
1830 void set_numabalancing_state(bool enabled)
1833 sched_feat_set("NUMA");
1835 sched_feat_set("NO_NUMA");
1838 __read_mostly bool numabalancing_enabled;
1840 void set_numabalancing_state(bool enabled)
1842 numabalancing_enabled = enabled;
1844 #endif /* CONFIG_SCHED_DEBUG */
1846 #ifdef CONFIG_PROC_SYSCTL
1847 int sysctl_numa_balancing(struct ctl_table *table, int write,
1848 void __user *buffer, size_t *lenp, loff_t *ppos)
1852 int state = numabalancing_enabled;
1854 if (write && !capable(CAP_SYS_ADMIN))
1859 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1863 set_numabalancing_state(state);
1870 * fork()/clone()-time setup:
1872 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1874 unsigned long flags;
1875 int cpu = get_cpu();
1877 __sched_fork(clone_flags, p);
1879 * We mark the process as running here. This guarantees that
1880 * nobody will actually run it, and a signal or other external
1881 * event cannot wake it up and insert it on the runqueue either.
1883 p->state = TASK_RUNNING;
1886 * Make sure we do not leak PI boosting priority to the child.
1888 p->prio = current->normal_prio;
1891 * Revert to default priority/policy on fork if requested.
1893 if (unlikely(p->sched_reset_on_fork)) {
1894 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1895 p->policy = SCHED_NORMAL;
1896 p->static_prio = NICE_TO_PRIO(0);
1898 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1899 p->static_prio = NICE_TO_PRIO(0);
1901 p->prio = p->normal_prio = __normal_prio(p);
1905 * We don't need the reset flag anymore after the fork. It has
1906 * fulfilled its duty:
1908 p->sched_reset_on_fork = 0;
1911 if (dl_prio(p->prio)) {
1914 } else if (rt_prio(p->prio)) {
1915 p->sched_class = &rt_sched_class;
1917 p->sched_class = &fair_sched_class;
1920 if (p->sched_class->task_fork)
1921 p->sched_class->task_fork(p);
1924 * The child is not yet in the pid-hash so no cgroup attach races,
1925 * and the cgroup is pinned to this child due to cgroup_fork()
1926 * is ran before sched_fork().
1928 * Silence PROVE_RCU.
1930 raw_spin_lock_irqsave(&p->pi_lock, flags);
1931 set_task_cpu(p, cpu);
1932 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1934 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1935 if (likely(sched_info_on()))
1936 memset(&p->sched_info, 0, sizeof(p->sched_info));
1938 #if defined(CONFIG_SMP)
1941 init_task_preempt_count(p);
1943 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1944 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1951 unsigned long to_ratio(u64 period, u64 runtime)
1953 if (runtime == RUNTIME_INF)
1957 * Doing this here saves a lot of checks in all
1958 * the calling paths, and returning zero seems
1959 * safe for them anyway.
1964 return div64_u64(runtime << 20, period);
1968 inline struct dl_bw *dl_bw_of(int i)
1970 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1971 "sched RCU must be held");
1972 return &cpu_rq(i)->rd->dl_bw;
1975 static inline int dl_bw_cpus(int i)
1977 struct root_domain *rd = cpu_rq(i)->rd;
1980 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1981 "sched RCU must be held");
1982 for_each_cpu_and(i, rd->span, cpu_active_mask)
1988 inline struct dl_bw *dl_bw_of(int i)
1990 return &cpu_rq(i)->dl.dl_bw;
1993 static inline int dl_bw_cpus(int i)
2000 * We must be sure that accepting a new task (or allowing changing the
2001 * parameters of an existing one) is consistent with the bandwidth
2002 * constraints. If yes, this function also accordingly updates the currently
2003 * allocated bandwidth to reflect the new situation.
2005 * This function is called while holding p's rq->lock.
2007 * XXX we should delay bw change until the task's 0-lag point, see
2010 static int dl_overflow(struct task_struct *p, int policy,
2011 const struct sched_attr *attr)
2014 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2015 u64 period = attr->sched_period ?: attr->sched_deadline;
2016 u64 runtime = attr->sched_runtime;
2017 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2020 if (new_bw == p->dl.dl_bw)
2024 * Either if a task, enters, leave, or stays -deadline but changes
2025 * its parameters, we may need to update accordingly the total
2026 * allocated bandwidth of the container.
2028 raw_spin_lock(&dl_b->lock);
2029 cpus = dl_bw_cpus(task_cpu(p));
2030 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2031 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2032 __dl_add(dl_b, new_bw);
2034 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2035 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2036 __dl_clear(dl_b, p->dl.dl_bw);
2037 __dl_add(dl_b, new_bw);
2039 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2040 __dl_clear(dl_b, p->dl.dl_bw);
2043 raw_spin_unlock(&dl_b->lock);
2048 extern void init_dl_bw(struct dl_bw *dl_b);
2051 * wake_up_new_task - wake up a newly created task for the first time.
2053 * This function will do some initial scheduler statistics housekeeping
2054 * that must be done for every newly created context, then puts the task
2055 * on the runqueue and wakes it.
2057 void wake_up_new_task(struct task_struct *p)
2059 unsigned long flags;
2062 raw_spin_lock_irqsave(&p->pi_lock, flags);
2065 * Fork balancing, do it here and not earlier because:
2066 * - cpus_allowed can change in the fork path
2067 * - any previously selected cpu might disappear through hotplug
2069 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2072 /* Initialize new task's runnable average */
2073 init_task_runnable_average(p);
2074 rq = __task_rq_lock(p);
2075 activate_task(rq, p, 0);
2076 p->on_rq = TASK_ON_RQ_QUEUED;
2077 trace_sched_wakeup_new(p, true);
2078 check_preempt_curr(rq, p, WF_FORK);
2080 if (p->sched_class->task_woken)
2081 p->sched_class->task_woken(rq, p);
2083 task_rq_unlock(rq, p, &flags);
2086 #ifdef CONFIG_PREEMPT_NOTIFIERS
2089 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2090 * @notifier: notifier struct to register
2092 void preempt_notifier_register(struct preempt_notifier *notifier)
2094 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2096 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2099 * preempt_notifier_unregister - no longer interested in preemption notifications
2100 * @notifier: notifier struct to unregister
2102 * This is safe to call from within a preemption notifier.
2104 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2106 hlist_del(¬ifier->link);
2108 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2110 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2112 struct preempt_notifier *notifier;
2114 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2115 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2119 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2120 struct task_struct *next)
2122 struct preempt_notifier *notifier;
2124 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2125 notifier->ops->sched_out(notifier, next);
2128 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2135 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2136 struct task_struct *next)
2140 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2143 * prepare_task_switch - prepare to switch tasks
2144 * @rq: the runqueue preparing to switch
2145 * @prev: the current task that is being switched out
2146 * @next: the task we are going to switch to.
2148 * This is called with the rq lock held and interrupts off. It must
2149 * be paired with a subsequent finish_task_switch after the context
2152 * prepare_task_switch sets up locking and calls architecture specific
2156 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2157 struct task_struct *next)
2159 trace_sched_switch(prev, next);
2160 sched_info_switch(rq, prev, next);
2161 perf_event_task_sched_out(prev, next);
2162 fire_sched_out_preempt_notifiers(prev, next);
2163 prepare_lock_switch(rq, next);
2164 prepare_arch_switch(next);
2168 * finish_task_switch - clean up after a task-switch
2169 * @prev: the thread we just switched away from.
2171 * finish_task_switch must be called after the context switch, paired
2172 * with a prepare_task_switch call before the context switch.
2173 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2174 * and do any other architecture-specific cleanup actions.
2176 * Note that we may have delayed dropping an mm in context_switch(). If
2177 * so, we finish that here outside of the runqueue lock. (Doing it
2178 * with the lock held can cause deadlocks; see schedule() for
2181 * The context switch have flipped the stack from under us and restored the
2182 * local variables which were saved when this task called schedule() in the
2183 * past. prev == current is still correct but we need to recalculate this_rq
2184 * because prev may have moved to another CPU.
2186 static struct rq *finish_task_switch(struct task_struct *prev)
2187 __releases(rq->lock)
2189 struct rq *rq = this_rq();
2190 struct mm_struct *mm = rq->prev_mm;
2196 * A task struct has one reference for the use as "current".
2197 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2198 * schedule one last time. The schedule call will never return, and
2199 * the scheduled task must drop that reference.
2200 * The test for TASK_DEAD must occur while the runqueue locks are
2201 * still held, otherwise prev could be scheduled on another cpu, die
2202 * there before we look at prev->state, and then the reference would
2204 * Manfred Spraul <manfred@colorfullife.com>
2206 prev_state = prev->state;
2207 vtime_task_switch(prev);
2208 finish_arch_switch(prev);
2209 perf_event_task_sched_in(prev, current);
2210 finish_lock_switch(rq, prev);
2211 finish_arch_post_lock_switch();
2213 fire_sched_in_preempt_notifiers(current);
2216 if (unlikely(prev_state == TASK_DEAD)) {
2217 if (prev->sched_class->task_dead)
2218 prev->sched_class->task_dead(prev);
2221 * Remove function-return probe instances associated with this
2222 * task and put them back on the free list.
2224 kprobe_flush_task(prev);
2225 put_task_struct(prev);
2228 tick_nohz_task_switch(current);
2234 /* rq->lock is NOT held, but preemption is disabled */
2235 static inline void post_schedule(struct rq *rq)
2237 if (rq->post_schedule) {
2238 unsigned long flags;
2240 raw_spin_lock_irqsave(&rq->lock, flags);
2241 if (rq->curr->sched_class->post_schedule)
2242 rq->curr->sched_class->post_schedule(rq);
2243 raw_spin_unlock_irqrestore(&rq->lock, flags);
2245 rq->post_schedule = 0;
2251 static inline void post_schedule(struct rq *rq)
2258 * schedule_tail - first thing a freshly forked thread must call.
2259 * @prev: the thread we just switched away from.
2261 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2262 __releases(rq->lock)
2266 /* finish_task_switch() drops rq->lock and enables preemtion */
2268 rq = finish_task_switch(prev);
2272 if (current->set_child_tid)
2273 put_user(task_pid_vnr(current), current->set_child_tid);
2277 * context_switch - switch to the new MM and the new thread's register state.
2279 static inline struct rq *
2280 context_switch(struct rq *rq, struct task_struct *prev,
2281 struct task_struct *next)
2283 struct mm_struct *mm, *oldmm;
2285 prepare_task_switch(rq, prev, next);
2288 oldmm = prev->active_mm;
2290 * For paravirt, this is coupled with an exit in switch_to to
2291 * combine the page table reload and the switch backend into
2294 arch_start_context_switch(prev);
2297 next->active_mm = oldmm;
2298 atomic_inc(&oldmm->mm_count);
2299 enter_lazy_tlb(oldmm, next);
2301 switch_mm(oldmm, mm, next);
2304 prev->active_mm = NULL;
2305 rq->prev_mm = oldmm;
2308 * Since the runqueue lock will be released by the next
2309 * task (which is an invalid locking op but in the case
2310 * of the scheduler it's an obvious special-case), so we
2311 * do an early lockdep release here:
2313 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2315 context_tracking_task_switch(prev, next);
2316 /* Here we just switch the register state and the stack. */
2317 switch_to(prev, next, prev);
2320 return finish_task_switch(prev);
2324 * nr_running and nr_context_switches:
2326 * externally visible scheduler statistics: current number of runnable
2327 * threads, total number of context switches performed since bootup.
2329 unsigned long nr_running(void)
2331 unsigned long i, sum = 0;
2333 for_each_online_cpu(i)
2334 sum += cpu_rq(i)->nr_running;
2340 * Check if only the current task is running on the cpu.
2342 bool single_task_running(void)
2344 if (cpu_rq(smp_processor_id())->nr_running == 1)
2349 EXPORT_SYMBOL(single_task_running);
2351 unsigned long long nr_context_switches(void)
2354 unsigned long long sum = 0;
2356 for_each_possible_cpu(i)
2357 sum += cpu_rq(i)->nr_switches;
2362 unsigned long nr_iowait(void)
2364 unsigned long i, sum = 0;
2366 for_each_possible_cpu(i)
2367 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2372 unsigned long nr_iowait_cpu(int cpu)
2374 struct rq *this = cpu_rq(cpu);
2375 return atomic_read(&this->nr_iowait);
2378 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2380 struct rq *this = this_rq();
2381 *nr_waiters = atomic_read(&this->nr_iowait);
2382 *load = this->cpu_load[0];
2388 * sched_exec - execve() is a valuable balancing opportunity, because at
2389 * this point the task has the smallest effective memory and cache footprint.
2391 void sched_exec(void)
2393 struct task_struct *p = current;
2394 unsigned long flags;
2397 raw_spin_lock_irqsave(&p->pi_lock, flags);
2398 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2399 if (dest_cpu == smp_processor_id())
2402 if (likely(cpu_active(dest_cpu))) {
2403 struct migration_arg arg = { p, dest_cpu };
2405 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2406 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2410 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2415 DEFINE_PER_CPU(struct kernel_stat, kstat);
2416 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2418 EXPORT_PER_CPU_SYMBOL(kstat);
2419 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2422 * Return accounted runtime for the task.
2423 * In case the task is currently running, return the runtime plus current's
2424 * pending runtime that have not been accounted yet.
2426 unsigned long long task_sched_runtime(struct task_struct *p)
2428 unsigned long flags;
2432 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2434 * 64-bit doesn't need locks to atomically read a 64bit value.
2435 * So we have a optimization chance when the task's delta_exec is 0.
2436 * Reading ->on_cpu is racy, but this is ok.
2438 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2439 * If we race with it entering cpu, unaccounted time is 0. This is
2440 * indistinguishable from the read occurring a few cycles earlier.
2441 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2442 * been accounted, so we're correct here as well.
2444 if (!p->on_cpu || !task_on_rq_queued(p))
2445 return p->se.sum_exec_runtime;
2448 rq = task_rq_lock(p, &flags);
2450 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2451 * project cycles that may never be accounted to this
2452 * thread, breaking clock_gettime().
2454 if (task_current(rq, p) && task_on_rq_queued(p)) {
2455 update_rq_clock(rq);
2456 p->sched_class->update_curr(rq);
2458 ns = p->se.sum_exec_runtime;
2459 task_rq_unlock(rq, p, &flags);
2465 * This function gets called by the timer code, with HZ frequency.
2466 * We call it with interrupts disabled.
2468 void scheduler_tick(void)
2470 int cpu = smp_processor_id();
2471 struct rq *rq = cpu_rq(cpu);
2472 struct task_struct *curr = rq->curr;
2476 raw_spin_lock(&rq->lock);
2477 update_rq_clock(rq);
2478 curr->sched_class->task_tick(rq, curr, 0);
2479 update_cpu_load_active(rq);
2480 raw_spin_unlock(&rq->lock);
2482 perf_event_task_tick();
2485 rq->idle_balance = idle_cpu(cpu);
2486 trigger_load_balance(rq);
2488 rq_last_tick_reset(rq);
2491 #ifdef CONFIG_NO_HZ_FULL
2493 * scheduler_tick_max_deferment
2495 * Keep at least one tick per second when a single
2496 * active task is running because the scheduler doesn't
2497 * yet completely support full dynticks environment.
2499 * This makes sure that uptime, CFS vruntime, load
2500 * balancing, etc... continue to move forward, even
2501 * with a very low granularity.
2503 * Return: Maximum deferment in nanoseconds.
2505 u64 scheduler_tick_max_deferment(void)
2507 struct rq *rq = this_rq();
2508 unsigned long next, now = ACCESS_ONCE(jiffies);
2510 next = rq->last_sched_tick + HZ;
2512 if (time_before_eq(next, now))
2515 return jiffies_to_nsecs(next - now);
2519 notrace unsigned long get_parent_ip(unsigned long addr)
2521 if (in_lock_functions(addr)) {
2522 addr = CALLER_ADDR2;
2523 if (in_lock_functions(addr))
2524 addr = CALLER_ADDR3;
2529 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2530 defined(CONFIG_PREEMPT_TRACER))
2532 void preempt_count_add(int val)
2534 #ifdef CONFIG_DEBUG_PREEMPT
2538 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2541 __preempt_count_add(val);
2542 #ifdef CONFIG_DEBUG_PREEMPT
2544 * Spinlock count overflowing soon?
2546 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2549 if (preempt_count() == val) {
2550 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2551 #ifdef CONFIG_DEBUG_PREEMPT
2552 current->preempt_disable_ip = ip;
2554 trace_preempt_off(CALLER_ADDR0, ip);
2557 EXPORT_SYMBOL(preempt_count_add);
2558 NOKPROBE_SYMBOL(preempt_count_add);
2560 void preempt_count_sub(int val)
2562 #ifdef CONFIG_DEBUG_PREEMPT
2566 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2569 * Is the spinlock portion underflowing?
2571 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2572 !(preempt_count() & PREEMPT_MASK)))
2576 if (preempt_count() == val)
2577 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2578 __preempt_count_sub(val);
2580 EXPORT_SYMBOL(preempt_count_sub);
2581 NOKPROBE_SYMBOL(preempt_count_sub);
2586 * Print scheduling while atomic bug:
2588 static noinline void __schedule_bug(struct task_struct *prev)
2590 if (oops_in_progress)
2593 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2594 prev->comm, prev->pid, preempt_count());
2596 debug_show_held_locks(prev);
2598 if (irqs_disabled())
2599 print_irqtrace_events(prev);
2600 #ifdef CONFIG_DEBUG_PREEMPT
2601 if (in_atomic_preempt_off()) {
2602 pr_err("Preemption disabled at:");
2603 print_ip_sym(current->preempt_disable_ip);
2608 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2612 * Various schedule()-time debugging checks and statistics:
2614 static inline void schedule_debug(struct task_struct *prev)
2616 #ifdef CONFIG_SCHED_STACK_END_CHECK
2617 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2620 * Test if we are atomic. Since do_exit() needs to call into
2621 * schedule() atomically, we ignore that path. Otherwise whine
2622 * if we are scheduling when we should not.
2624 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2625 __schedule_bug(prev);
2628 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2630 schedstat_inc(this_rq(), sched_count);
2634 * Pick up the highest-prio task:
2636 static inline struct task_struct *
2637 pick_next_task(struct rq *rq, struct task_struct *prev)
2639 const struct sched_class *class = &fair_sched_class;
2640 struct task_struct *p;
2643 * Optimization: we know that if all tasks are in
2644 * the fair class we can call that function directly:
2646 if (likely(prev->sched_class == class &&
2647 rq->nr_running == rq->cfs.h_nr_running)) {
2648 p = fair_sched_class.pick_next_task(rq, prev);
2649 if (unlikely(p == RETRY_TASK))
2652 /* assumes fair_sched_class->next == idle_sched_class */
2654 p = idle_sched_class.pick_next_task(rq, prev);
2660 for_each_class(class) {
2661 p = class->pick_next_task(rq, prev);
2663 if (unlikely(p == RETRY_TASK))
2669 BUG(); /* the idle class will always have a runnable task */
2673 * __schedule() is the main scheduler function.
2675 * The main means of driving the scheduler and thus entering this function are:
2677 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2679 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2680 * paths. For example, see arch/x86/entry_64.S.
2682 * To drive preemption between tasks, the scheduler sets the flag in timer
2683 * interrupt handler scheduler_tick().
2685 * 3. Wakeups don't really cause entry into schedule(). They add a
2686 * task to the run-queue and that's it.
2688 * Now, if the new task added to the run-queue preempts the current
2689 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2690 * called on the nearest possible occasion:
2692 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2694 * - in syscall or exception context, at the next outmost
2695 * preempt_enable(). (this might be as soon as the wake_up()'s
2698 * - in IRQ context, return from interrupt-handler to
2699 * preemptible context
2701 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2704 * - cond_resched() call
2705 * - explicit schedule() call
2706 * - return from syscall or exception to user-space
2707 * - return from interrupt-handler to user-space
2709 * WARNING: all callers must re-check need_resched() afterward and reschedule
2710 * accordingly in case an event triggered the need for rescheduling (such as
2711 * an interrupt waking up a task) while preemption was disabled in __schedule().
2713 static void __sched __schedule(void)
2715 struct task_struct *prev, *next;
2716 unsigned long *switch_count;
2721 cpu = smp_processor_id();
2723 rcu_note_context_switch();
2726 schedule_debug(prev);
2728 if (sched_feat(HRTICK))
2732 * Make sure that signal_pending_state()->signal_pending() below
2733 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2734 * done by the caller to avoid the race with signal_wake_up().
2736 smp_mb__before_spinlock();
2737 raw_spin_lock_irq(&rq->lock);
2739 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2741 switch_count = &prev->nivcsw;
2742 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2743 if (unlikely(signal_pending_state(prev->state, prev))) {
2744 prev->state = TASK_RUNNING;
2746 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2750 * If a worker went to sleep, notify and ask workqueue
2751 * whether it wants to wake up a task to maintain
2754 if (prev->flags & PF_WQ_WORKER) {
2755 struct task_struct *to_wakeup;
2757 to_wakeup = wq_worker_sleeping(prev, cpu);
2759 try_to_wake_up_local(to_wakeup);
2762 switch_count = &prev->nvcsw;
2765 if (task_on_rq_queued(prev))
2766 update_rq_clock(rq);
2768 next = pick_next_task(rq, prev);
2769 clear_tsk_need_resched(prev);
2770 clear_preempt_need_resched();
2771 rq->clock_skip_update = 0;
2773 if (likely(prev != next)) {
2778 rq = context_switch(rq, prev, next); /* unlocks the rq */
2781 raw_spin_unlock_irq(&rq->lock);
2785 sched_preempt_enable_no_resched();
2788 static inline void sched_submit_work(struct task_struct *tsk)
2790 if (!tsk->state || tsk_is_pi_blocked(tsk))
2793 * If we are going to sleep and we have plugged IO queued,
2794 * make sure to submit it to avoid deadlocks.
2796 if (blk_needs_flush_plug(tsk))
2797 blk_schedule_flush_plug(tsk);
2800 asmlinkage __visible void __sched schedule(void)
2802 struct task_struct *tsk = current;
2804 sched_submit_work(tsk);
2807 } while (need_resched());
2809 EXPORT_SYMBOL(schedule);
2811 #ifdef CONFIG_CONTEXT_TRACKING
2812 asmlinkage __visible void __sched schedule_user(void)
2815 * If we come here after a random call to set_need_resched(),
2816 * or we have been woken up remotely but the IPI has not yet arrived,
2817 * we haven't yet exited the RCU idle mode. Do it here manually until
2818 * we find a better solution.
2820 * NB: There are buggy callers of this function. Ideally we
2821 * should warn if prev_state != IN_USER, but that will trigger
2822 * too frequently to make sense yet.
2824 enum ctx_state prev_state = exception_enter();
2826 exception_exit(prev_state);
2831 * schedule_preempt_disabled - called with preemption disabled
2833 * Returns with preemption disabled. Note: preempt_count must be 1
2835 void __sched schedule_preempt_disabled(void)
2837 sched_preempt_enable_no_resched();
2842 static void __sched notrace preempt_schedule_common(void)
2845 __preempt_count_add(PREEMPT_ACTIVE);
2847 __preempt_count_sub(PREEMPT_ACTIVE);
2850 * Check again in case we missed a preemption opportunity
2851 * between schedule and now.
2854 } while (need_resched());
2857 #ifdef CONFIG_PREEMPT
2859 * this is the entry point to schedule() from in-kernel preemption
2860 * off of preempt_enable. Kernel preemptions off return from interrupt
2861 * occur there and call schedule directly.
2863 asmlinkage __visible void __sched notrace preempt_schedule(void)
2866 * If there is a non-zero preempt_count or interrupts are disabled,
2867 * we do not want to preempt the current task. Just return..
2869 if (likely(!preemptible()))
2872 preempt_schedule_common();
2874 NOKPROBE_SYMBOL(preempt_schedule);
2875 EXPORT_SYMBOL(preempt_schedule);
2877 #ifdef CONFIG_CONTEXT_TRACKING
2879 * preempt_schedule_context - preempt_schedule called by tracing
2881 * The tracing infrastructure uses preempt_enable_notrace to prevent
2882 * recursion and tracing preempt enabling caused by the tracing
2883 * infrastructure itself. But as tracing can happen in areas coming
2884 * from userspace or just about to enter userspace, a preempt enable
2885 * can occur before user_exit() is called. This will cause the scheduler
2886 * to be called when the system is still in usermode.
2888 * To prevent this, the preempt_enable_notrace will use this function
2889 * instead of preempt_schedule() to exit user context if needed before
2890 * calling the scheduler.
2892 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2894 enum ctx_state prev_ctx;
2896 if (likely(!preemptible()))
2900 __preempt_count_add(PREEMPT_ACTIVE);
2902 * Needs preempt disabled in case user_exit() is traced
2903 * and the tracer calls preempt_enable_notrace() causing
2904 * an infinite recursion.
2906 prev_ctx = exception_enter();
2908 exception_exit(prev_ctx);
2910 __preempt_count_sub(PREEMPT_ACTIVE);
2912 } while (need_resched());
2914 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2915 #endif /* CONFIG_CONTEXT_TRACKING */
2917 #endif /* CONFIG_PREEMPT */
2920 * this is the entry point to schedule() from kernel preemption
2921 * off of irq context.
2922 * Note, that this is called and return with irqs disabled. This will
2923 * protect us against recursive calling from irq.
2925 asmlinkage __visible void __sched preempt_schedule_irq(void)
2927 enum ctx_state prev_state;
2929 /* Catch callers which need to be fixed */
2930 BUG_ON(preempt_count() || !irqs_disabled());
2932 prev_state = exception_enter();
2935 __preempt_count_add(PREEMPT_ACTIVE);
2938 local_irq_disable();
2939 __preempt_count_sub(PREEMPT_ACTIVE);
2942 * Check again in case we missed a preemption opportunity
2943 * between schedule and now.
2946 } while (need_resched());
2948 exception_exit(prev_state);
2951 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2954 return try_to_wake_up(curr->private, mode, wake_flags);
2956 EXPORT_SYMBOL(default_wake_function);
2958 #ifdef CONFIG_RT_MUTEXES
2961 * rt_mutex_setprio - set the current priority of a task
2963 * @prio: prio value (kernel-internal form)
2965 * This function changes the 'effective' priority of a task. It does
2966 * not touch ->normal_prio like __setscheduler().
2968 * Used by the rt_mutex code to implement priority inheritance
2969 * logic. Call site only calls if the priority of the task changed.
2971 void rt_mutex_setprio(struct task_struct *p, int prio)
2973 int oldprio, queued, running, enqueue_flag = 0;
2975 const struct sched_class *prev_class;
2977 BUG_ON(prio > MAX_PRIO);
2979 rq = __task_rq_lock(p);
2982 * Idle task boosting is a nono in general. There is one
2983 * exception, when PREEMPT_RT and NOHZ is active:
2985 * The idle task calls get_next_timer_interrupt() and holds
2986 * the timer wheel base->lock on the CPU and another CPU wants
2987 * to access the timer (probably to cancel it). We can safely
2988 * ignore the boosting request, as the idle CPU runs this code
2989 * with interrupts disabled and will complete the lock
2990 * protected section without being interrupted. So there is no
2991 * real need to boost.
2993 if (unlikely(p == rq->idle)) {
2994 WARN_ON(p != rq->curr);
2995 WARN_ON(p->pi_blocked_on);
2999 trace_sched_pi_setprio(p, prio);
3001 prev_class = p->sched_class;
3002 queued = task_on_rq_queued(p);
3003 running = task_current(rq, p);
3005 dequeue_task(rq, p, 0);
3007 put_prev_task(rq, p);
3010 * Boosting condition are:
3011 * 1. -rt task is running and holds mutex A
3012 * --> -dl task blocks on mutex A
3014 * 2. -dl task is running and holds mutex A
3015 * --> -dl task blocks on mutex A and could preempt the
3018 if (dl_prio(prio)) {
3019 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3020 if (!dl_prio(p->normal_prio) ||
3021 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3022 p->dl.dl_boosted = 1;
3023 p->dl.dl_throttled = 0;
3024 enqueue_flag = ENQUEUE_REPLENISH;
3026 p->dl.dl_boosted = 0;
3027 p->sched_class = &dl_sched_class;
3028 } else if (rt_prio(prio)) {
3029 if (dl_prio(oldprio))
3030 p->dl.dl_boosted = 0;
3032 enqueue_flag = ENQUEUE_HEAD;
3033 p->sched_class = &rt_sched_class;
3035 if (dl_prio(oldprio))
3036 p->dl.dl_boosted = 0;
3037 p->sched_class = &fair_sched_class;
3043 p->sched_class->set_curr_task(rq);
3045 enqueue_task(rq, p, enqueue_flag);
3047 check_class_changed(rq, p, prev_class, oldprio);
3049 __task_rq_unlock(rq);
3053 void set_user_nice(struct task_struct *p, long nice)
3055 int old_prio, delta, queued;
3056 unsigned long flags;
3059 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3062 * We have to be careful, if called from sys_setpriority(),
3063 * the task might be in the middle of scheduling on another CPU.
3065 rq = task_rq_lock(p, &flags);
3067 * The RT priorities are set via sched_setscheduler(), but we still
3068 * allow the 'normal' nice value to be set - but as expected
3069 * it wont have any effect on scheduling until the task is
3070 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3072 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3073 p->static_prio = NICE_TO_PRIO(nice);
3076 queued = task_on_rq_queued(p);
3078 dequeue_task(rq, p, 0);
3080 p->static_prio = NICE_TO_PRIO(nice);
3083 p->prio = effective_prio(p);
3084 delta = p->prio - old_prio;
3087 enqueue_task(rq, p, 0);
3089 * If the task increased its priority or is running and
3090 * lowered its priority, then reschedule its CPU:
3092 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3096 task_rq_unlock(rq, p, &flags);
3098 EXPORT_SYMBOL(set_user_nice);
3101 * can_nice - check if a task can reduce its nice value
3105 int can_nice(const struct task_struct *p, const int nice)
3107 /* convert nice value [19,-20] to rlimit style value [1,40] */
3108 int nice_rlim = nice_to_rlimit(nice);
3110 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3111 capable(CAP_SYS_NICE));
3114 #ifdef __ARCH_WANT_SYS_NICE
3117 * sys_nice - change the priority of the current process.
3118 * @increment: priority increment
3120 * sys_setpriority is a more generic, but much slower function that
3121 * does similar things.
3123 SYSCALL_DEFINE1(nice, int, increment)
3128 * Setpriority might change our priority at the same moment.
3129 * We don't have to worry. Conceptually one call occurs first
3130 * and we have a single winner.
3132 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3133 nice = task_nice(current) + increment;
3135 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3136 if (increment < 0 && !can_nice(current, nice))
3139 retval = security_task_setnice(current, nice);
3143 set_user_nice(current, nice);
3150 * task_prio - return the priority value of a given task.
3151 * @p: the task in question.
3153 * Return: The priority value as seen by users in /proc.
3154 * RT tasks are offset by -200. Normal tasks are centered
3155 * around 0, value goes from -16 to +15.
3157 int task_prio(const struct task_struct *p)
3159 return p->prio - MAX_RT_PRIO;
3163 * idle_cpu - is a given cpu idle currently?
3164 * @cpu: the processor in question.
3166 * Return: 1 if the CPU is currently idle. 0 otherwise.
3168 int idle_cpu(int cpu)
3170 struct rq *rq = cpu_rq(cpu);
3172 if (rq->curr != rq->idle)
3179 if (!llist_empty(&rq->wake_list))
3187 * idle_task - return the idle task for a given cpu.
3188 * @cpu: the processor in question.
3190 * Return: The idle task for the cpu @cpu.
3192 struct task_struct *idle_task(int cpu)
3194 return cpu_rq(cpu)->idle;
3198 * find_process_by_pid - find a process with a matching PID value.
3199 * @pid: the pid in question.
3201 * The task of @pid, if found. %NULL otherwise.
3203 static struct task_struct *find_process_by_pid(pid_t pid)
3205 return pid ? find_task_by_vpid(pid) : current;
3209 * This function initializes the sched_dl_entity of a newly becoming
3210 * SCHED_DEADLINE task.
3212 * Only the static values are considered here, the actual runtime and the
3213 * absolute deadline will be properly calculated when the task is enqueued
3214 * for the first time with its new policy.
3217 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3219 struct sched_dl_entity *dl_se = &p->dl;
3221 dl_se->dl_runtime = attr->sched_runtime;
3222 dl_se->dl_deadline = attr->sched_deadline;
3223 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3224 dl_se->flags = attr->sched_flags;
3225 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3228 * Changing the parameters of a task is 'tricky' and we're not doing
3229 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3231 * What we SHOULD do is delay the bandwidth release until the 0-lag
3232 * point. This would include retaining the task_struct until that time
3233 * and change dl_overflow() to not immediately decrement the current
3236 * Instead we retain the current runtime/deadline and let the new
3237 * parameters take effect after the current reservation period lapses.
3238 * This is safe (albeit pessimistic) because the 0-lag point is always
3239 * before the current scheduling deadline.
3241 * We can still have temporary overloads because we do not delay the
3242 * change in bandwidth until that time; so admission control is
3243 * not on the safe side. It does however guarantee tasks will never
3244 * consume more than promised.
3249 * sched_setparam() passes in -1 for its policy, to let the functions
3250 * it calls know not to change it.
3252 #define SETPARAM_POLICY -1
3254 static void __setscheduler_params(struct task_struct *p,
3255 const struct sched_attr *attr)
3257 int policy = attr->sched_policy;
3259 if (policy == SETPARAM_POLICY)
3264 if (dl_policy(policy))
3265 __setparam_dl(p, attr);
3266 else if (fair_policy(policy))
3267 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3270 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3271 * !rt_policy. Always setting this ensures that things like
3272 * getparam()/getattr() don't report silly values for !rt tasks.
3274 p->rt_priority = attr->sched_priority;
3275 p->normal_prio = normal_prio(p);
3279 /* Actually do priority change: must hold pi & rq lock. */
3280 static void __setscheduler(struct rq *rq, struct task_struct *p,
3281 const struct sched_attr *attr)
3283 __setscheduler_params(p, attr);
3286 * If we get here, there was no pi waiters boosting the
3287 * task. It is safe to use the normal prio.
3289 p->prio = normal_prio(p);
3291 if (dl_prio(p->prio))
3292 p->sched_class = &dl_sched_class;
3293 else if (rt_prio(p->prio))
3294 p->sched_class = &rt_sched_class;
3296 p->sched_class = &fair_sched_class;
3300 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3302 struct sched_dl_entity *dl_se = &p->dl;
3304 attr->sched_priority = p->rt_priority;
3305 attr->sched_runtime = dl_se->dl_runtime;
3306 attr->sched_deadline = dl_se->dl_deadline;
3307 attr->sched_period = dl_se->dl_period;
3308 attr->sched_flags = dl_se->flags;
3312 * This function validates the new parameters of a -deadline task.
3313 * We ask for the deadline not being zero, and greater or equal
3314 * than the runtime, as well as the period of being zero or
3315 * greater than deadline. Furthermore, we have to be sure that
3316 * user parameters are above the internal resolution of 1us (we
3317 * check sched_runtime only since it is always the smaller one) and
3318 * below 2^63 ns (we have to check both sched_deadline and
3319 * sched_period, as the latter can be zero).
3322 __checkparam_dl(const struct sched_attr *attr)
3325 if (attr->sched_deadline == 0)
3329 * Since we truncate DL_SCALE bits, make sure we're at least
3332 if (attr->sched_runtime < (1ULL << DL_SCALE))
3336 * Since we use the MSB for wrap-around and sign issues, make
3337 * sure it's not set (mind that period can be equal to zero).
3339 if (attr->sched_deadline & (1ULL << 63) ||
3340 attr->sched_period & (1ULL << 63))
3343 /* runtime <= deadline <= period (if period != 0) */
3344 if ((attr->sched_period != 0 &&
3345 attr->sched_period < attr->sched_deadline) ||
3346 attr->sched_deadline < attr->sched_runtime)
3353 * check the target process has a UID that matches the current process's
3355 static bool check_same_owner(struct task_struct *p)
3357 const struct cred *cred = current_cred(), *pcred;
3361 pcred = __task_cred(p);
3362 match = (uid_eq(cred->euid, pcred->euid) ||
3363 uid_eq(cred->euid, pcred->uid));
3368 static bool dl_param_changed(struct task_struct *p,
3369 const struct sched_attr *attr)
3371 struct sched_dl_entity *dl_se = &p->dl;
3373 if (dl_se->dl_runtime != attr->sched_runtime ||
3374 dl_se->dl_deadline != attr->sched_deadline ||
3375 dl_se->dl_period != attr->sched_period ||
3376 dl_se->flags != attr->sched_flags)
3382 static int __sched_setscheduler(struct task_struct *p,
3383 const struct sched_attr *attr,
3386 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3387 MAX_RT_PRIO - 1 - attr->sched_priority;
3388 int retval, oldprio, oldpolicy = -1, queued, running;
3389 int policy = attr->sched_policy;
3390 unsigned long flags;
3391 const struct sched_class *prev_class;
3395 /* may grab non-irq protected spin_locks */
3396 BUG_ON(in_interrupt());
3398 /* double check policy once rq lock held */
3400 reset_on_fork = p->sched_reset_on_fork;
3401 policy = oldpolicy = p->policy;
3403 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3405 if (policy != SCHED_DEADLINE &&
3406 policy != SCHED_FIFO && policy != SCHED_RR &&
3407 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3408 policy != SCHED_IDLE)
3412 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3416 * Valid priorities for SCHED_FIFO and SCHED_RR are
3417 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3418 * SCHED_BATCH and SCHED_IDLE is 0.
3420 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3421 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3423 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3424 (rt_policy(policy) != (attr->sched_priority != 0)))
3428 * Allow unprivileged RT tasks to decrease priority:
3430 if (user && !capable(CAP_SYS_NICE)) {
3431 if (fair_policy(policy)) {
3432 if (attr->sched_nice < task_nice(p) &&
3433 !can_nice(p, attr->sched_nice))
3437 if (rt_policy(policy)) {
3438 unsigned long rlim_rtprio =
3439 task_rlimit(p, RLIMIT_RTPRIO);
3441 /* can't set/change the rt policy */
3442 if (policy != p->policy && !rlim_rtprio)
3445 /* can't increase priority */
3446 if (attr->sched_priority > p->rt_priority &&
3447 attr->sched_priority > rlim_rtprio)
3452 * Can't set/change SCHED_DEADLINE policy at all for now
3453 * (safest behavior); in the future we would like to allow
3454 * unprivileged DL tasks to increase their relative deadline
3455 * or reduce their runtime (both ways reducing utilization)
3457 if (dl_policy(policy))
3461 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3462 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3464 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3465 if (!can_nice(p, task_nice(p)))
3469 /* can't change other user's priorities */
3470 if (!check_same_owner(p))
3473 /* Normal users shall not reset the sched_reset_on_fork flag */
3474 if (p->sched_reset_on_fork && !reset_on_fork)
3479 retval = security_task_setscheduler(p);
3485 * make sure no PI-waiters arrive (or leave) while we are
3486 * changing the priority of the task:
3488 * To be able to change p->policy safely, the appropriate
3489 * runqueue lock must be held.
3491 rq = task_rq_lock(p, &flags);
3494 * Changing the policy of the stop threads its a very bad idea
3496 if (p == rq->stop) {
3497 task_rq_unlock(rq, p, &flags);
3502 * If not changing anything there's no need to proceed further,
3503 * but store a possible modification of reset_on_fork.
3505 if (unlikely(policy == p->policy)) {
3506 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3508 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3510 if (dl_policy(policy) && dl_param_changed(p, attr))
3513 p->sched_reset_on_fork = reset_on_fork;
3514 task_rq_unlock(rq, p, &flags);
3520 #ifdef CONFIG_RT_GROUP_SCHED
3522 * Do not allow realtime tasks into groups that have no runtime
3525 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3526 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3527 !task_group_is_autogroup(task_group(p))) {
3528 task_rq_unlock(rq, p, &flags);
3533 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3534 cpumask_t *span = rq->rd->span;
3537 * Don't allow tasks with an affinity mask smaller than
3538 * the entire root_domain to become SCHED_DEADLINE. We
3539 * will also fail if there's no bandwidth available.
3541 if (!cpumask_subset(span, &p->cpus_allowed) ||
3542 rq->rd->dl_bw.bw == 0) {
3543 task_rq_unlock(rq, p, &flags);
3550 /* recheck policy now with rq lock held */
3551 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3552 policy = oldpolicy = -1;
3553 task_rq_unlock(rq, p, &flags);
3558 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3559 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3562 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3563 task_rq_unlock(rq, p, &flags);
3567 p->sched_reset_on_fork = reset_on_fork;
3571 * Special case for priority boosted tasks.
3573 * If the new priority is lower or equal (user space view)
3574 * than the current (boosted) priority, we just store the new
3575 * normal parameters and do not touch the scheduler class and
3576 * the runqueue. This will be done when the task deboost
3579 if (rt_mutex_check_prio(p, newprio)) {
3580 __setscheduler_params(p, attr);
3581 task_rq_unlock(rq, p, &flags);
3585 queued = task_on_rq_queued(p);
3586 running = task_current(rq, p);
3588 dequeue_task(rq, p, 0);
3590 put_prev_task(rq, p);
3592 prev_class = p->sched_class;
3593 __setscheduler(rq, p, attr);
3596 p->sched_class->set_curr_task(rq);
3599 * We enqueue to tail when the priority of a task is
3600 * increased (user space view).
3602 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3605 check_class_changed(rq, p, prev_class, oldprio);
3606 task_rq_unlock(rq, p, &flags);
3608 rt_mutex_adjust_pi(p);
3613 static int _sched_setscheduler(struct task_struct *p, int policy,
3614 const struct sched_param *param, bool check)
3616 struct sched_attr attr = {
3617 .sched_policy = policy,
3618 .sched_priority = param->sched_priority,
3619 .sched_nice = PRIO_TO_NICE(p->static_prio),
3622 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3623 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3624 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3625 policy &= ~SCHED_RESET_ON_FORK;
3626 attr.sched_policy = policy;
3629 return __sched_setscheduler(p, &attr, check);
3632 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3633 * @p: the task in question.
3634 * @policy: new policy.
3635 * @param: structure containing the new RT priority.
3637 * Return: 0 on success. An error code otherwise.
3639 * NOTE that the task may be already dead.
3641 int sched_setscheduler(struct task_struct *p, int policy,
3642 const struct sched_param *param)
3644 return _sched_setscheduler(p, policy, param, true);
3646 EXPORT_SYMBOL_GPL(sched_setscheduler);
3648 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3650 return __sched_setscheduler(p, attr, true);
3652 EXPORT_SYMBOL_GPL(sched_setattr);
3655 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3656 * @p: the task in question.
3657 * @policy: new policy.
3658 * @param: structure containing the new RT priority.
3660 * Just like sched_setscheduler, only don't bother checking if the
3661 * current context has permission. For example, this is needed in
3662 * stop_machine(): we create temporary high priority worker threads,
3663 * but our caller might not have that capability.
3665 * Return: 0 on success. An error code otherwise.
3667 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3668 const struct sched_param *param)
3670 return _sched_setscheduler(p, policy, param, false);
3674 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3676 struct sched_param lparam;
3677 struct task_struct *p;
3680 if (!param || pid < 0)
3682 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3687 p = find_process_by_pid(pid);
3689 retval = sched_setscheduler(p, policy, &lparam);
3696 * Mimics kernel/events/core.c perf_copy_attr().
3698 static int sched_copy_attr(struct sched_attr __user *uattr,
3699 struct sched_attr *attr)
3704 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3708 * zero the full structure, so that a short copy will be nice.
3710 memset(attr, 0, sizeof(*attr));
3712 ret = get_user(size, &uattr->size);
3716 if (size > PAGE_SIZE) /* silly large */
3719 if (!size) /* abi compat */
3720 size = SCHED_ATTR_SIZE_VER0;
3722 if (size < SCHED_ATTR_SIZE_VER0)
3726 * If we're handed a bigger struct than we know of,
3727 * ensure all the unknown bits are 0 - i.e. new
3728 * user-space does not rely on any kernel feature
3729 * extensions we dont know about yet.
3731 if (size > sizeof(*attr)) {
3732 unsigned char __user *addr;
3733 unsigned char __user *end;
3736 addr = (void __user *)uattr + sizeof(*attr);
3737 end = (void __user *)uattr + size;
3739 for (; addr < end; addr++) {
3740 ret = get_user(val, addr);
3746 size = sizeof(*attr);
3749 ret = copy_from_user(attr, uattr, size);
3754 * XXX: do we want to be lenient like existing syscalls; or do we want
3755 * to be strict and return an error on out-of-bounds values?
3757 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3762 put_user(sizeof(*attr), &uattr->size);
3767 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3768 * @pid: the pid in question.
3769 * @policy: new policy.
3770 * @param: structure containing the new RT priority.
3772 * Return: 0 on success. An error code otherwise.
3774 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3775 struct sched_param __user *, param)
3777 /* negative values for policy are not valid */
3781 return do_sched_setscheduler(pid, policy, param);
3785 * sys_sched_setparam - set/change the RT priority of a thread
3786 * @pid: the pid in question.
3787 * @param: structure containing the new RT priority.
3789 * Return: 0 on success. An error code otherwise.
3791 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3793 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3797 * sys_sched_setattr - same as above, but with extended sched_attr
3798 * @pid: the pid in question.
3799 * @uattr: structure containing the extended parameters.
3800 * @flags: for future extension.
3802 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3803 unsigned int, flags)
3805 struct sched_attr attr;
3806 struct task_struct *p;
3809 if (!uattr || pid < 0 || flags)
3812 retval = sched_copy_attr(uattr, &attr);
3816 if ((int)attr.sched_policy < 0)
3821 p = find_process_by_pid(pid);
3823 retval = sched_setattr(p, &attr);
3830 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3831 * @pid: the pid in question.
3833 * Return: On success, the policy of the thread. Otherwise, a negative error
3836 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3838 struct task_struct *p;
3846 p = find_process_by_pid(pid);
3848 retval = security_task_getscheduler(p);
3851 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3858 * sys_sched_getparam - get the RT priority of a thread
3859 * @pid: the pid in question.
3860 * @param: structure containing the RT priority.
3862 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3865 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3867 struct sched_param lp = { .sched_priority = 0 };
3868 struct task_struct *p;
3871 if (!param || pid < 0)
3875 p = find_process_by_pid(pid);
3880 retval = security_task_getscheduler(p);
3884 if (task_has_rt_policy(p))
3885 lp.sched_priority = p->rt_priority;
3889 * This one might sleep, we cannot do it with a spinlock held ...
3891 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3900 static int sched_read_attr(struct sched_attr __user *uattr,
3901 struct sched_attr *attr,
3906 if (!access_ok(VERIFY_WRITE, uattr, usize))
3910 * If we're handed a smaller struct than we know of,
3911 * ensure all the unknown bits are 0 - i.e. old
3912 * user-space does not get uncomplete information.
3914 if (usize < sizeof(*attr)) {
3915 unsigned char *addr;
3918 addr = (void *)attr + usize;
3919 end = (void *)attr + sizeof(*attr);
3921 for (; addr < end; addr++) {
3929 ret = copy_to_user(uattr, attr, attr->size);
3937 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3938 * @pid: the pid in question.
3939 * @uattr: structure containing the extended parameters.
3940 * @size: sizeof(attr) for fwd/bwd comp.
3941 * @flags: for future extension.
3943 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3944 unsigned int, size, unsigned int, flags)
3946 struct sched_attr attr = {
3947 .size = sizeof(struct sched_attr),
3949 struct task_struct *p;
3952 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3953 size < SCHED_ATTR_SIZE_VER0 || flags)
3957 p = find_process_by_pid(pid);
3962 retval = security_task_getscheduler(p);
3966 attr.sched_policy = p->policy;
3967 if (p->sched_reset_on_fork)
3968 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3969 if (task_has_dl_policy(p))
3970 __getparam_dl(p, &attr);
3971 else if (task_has_rt_policy(p))
3972 attr.sched_priority = p->rt_priority;
3974 attr.sched_nice = task_nice(p);
3978 retval = sched_read_attr(uattr, &attr, size);
3986 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3988 cpumask_var_t cpus_allowed, new_mask;
3989 struct task_struct *p;
3994 p = find_process_by_pid(pid);
4000 /* Prevent p going away */
4004 if (p->flags & PF_NO_SETAFFINITY) {
4008 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4012 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4014 goto out_free_cpus_allowed;
4017 if (!check_same_owner(p)) {
4019 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4021 goto out_free_new_mask;
4026 retval = security_task_setscheduler(p);
4028 goto out_free_new_mask;
4031 cpuset_cpus_allowed(p, cpus_allowed);
4032 cpumask_and(new_mask, in_mask, cpus_allowed);
4035 * Since bandwidth control happens on root_domain basis,
4036 * if admission test is enabled, we only admit -deadline
4037 * tasks allowed to run on all the CPUs in the task's
4041 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4043 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4046 goto out_free_new_mask;
4052 retval = set_cpus_allowed_ptr(p, new_mask);
4055 cpuset_cpus_allowed(p, cpus_allowed);
4056 if (!cpumask_subset(new_mask, cpus_allowed)) {
4058 * We must have raced with a concurrent cpuset
4059 * update. Just reset the cpus_allowed to the
4060 * cpuset's cpus_allowed
4062 cpumask_copy(new_mask, cpus_allowed);
4067 free_cpumask_var(new_mask);
4068 out_free_cpus_allowed:
4069 free_cpumask_var(cpus_allowed);
4075 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4076 struct cpumask *new_mask)
4078 if (len < cpumask_size())
4079 cpumask_clear(new_mask);
4080 else if (len > cpumask_size())
4081 len = cpumask_size();
4083 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4087 * sys_sched_setaffinity - set the cpu affinity of a process
4088 * @pid: pid of the process
4089 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4090 * @user_mask_ptr: user-space pointer to the new cpu mask
4092 * Return: 0 on success. An error code otherwise.
4094 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4095 unsigned long __user *, user_mask_ptr)
4097 cpumask_var_t new_mask;
4100 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4103 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4105 retval = sched_setaffinity(pid, new_mask);
4106 free_cpumask_var(new_mask);
4110 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4112 struct task_struct *p;
4113 unsigned long flags;
4119 p = find_process_by_pid(pid);
4123 retval = security_task_getscheduler(p);
4127 raw_spin_lock_irqsave(&p->pi_lock, flags);
4128 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4129 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4138 * sys_sched_getaffinity - get the cpu affinity of a process
4139 * @pid: pid of the process
4140 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4141 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4143 * Return: 0 on success. An error code otherwise.
4145 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4146 unsigned long __user *, user_mask_ptr)
4151 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4153 if (len & (sizeof(unsigned long)-1))
4156 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4159 ret = sched_getaffinity(pid, mask);
4161 size_t retlen = min_t(size_t, len, cpumask_size());
4163 if (copy_to_user(user_mask_ptr, mask, retlen))
4168 free_cpumask_var(mask);
4174 * sys_sched_yield - yield the current processor to other threads.
4176 * This function yields the current CPU to other tasks. If there are no
4177 * other threads running on this CPU then this function will return.
4181 SYSCALL_DEFINE0(sched_yield)
4183 struct rq *rq = this_rq_lock();
4185 schedstat_inc(rq, yld_count);
4186 current->sched_class->yield_task(rq);
4189 * Since we are going to call schedule() anyway, there's
4190 * no need to preempt or enable interrupts:
4192 __release(rq->lock);
4193 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4194 do_raw_spin_unlock(&rq->lock);
4195 sched_preempt_enable_no_resched();
4202 int __sched _cond_resched(void)
4204 if (should_resched()) {
4205 preempt_schedule_common();
4210 EXPORT_SYMBOL(_cond_resched);
4213 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4214 * call schedule, and on return reacquire the lock.
4216 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4217 * operations here to prevent schedule() from being called twice (once via
4218 * spin_unlock(), once by hand).
4220 int __cond_resched_lock(spinlock_t *lock)
4222 int resched = should_resched();
4225 lockdep_assert_held(lock);
4227 if (spin_needbreak(lock) || resched) {
4230 preempt_schedule_common();
4238 EXPORT_SYMBOL(__cond_resched_lock);
4240 int __sched __cond_resched_softirq(void)
4242 BUG_ON(!in_softirq());
4244 if (should_resched()) {
4246 preempt_schedule_common();
4252 EXPORT_SYMBOL(__cond_resched_softirq);
4255 * yield - yield the current processor to other threads.
4257 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4259 * The scheduler is at all times free to pick the calling task as the most
4260 * eligible task to run, if removing the yield() call from your code breaks
4261 * it, its already broken.
4263 * Typical broken usage is:
4268 * where one assumes that yield() will let 'the other' process run that will
4269 * make event true. If the current task is a SCHED_FIFO task that will never
4270 * happen. Never use yield() as a progress guarantee!!
4272 * If you want to use yield() to wait for something, use wait_event().
4273 * If you want to use yield() to be 'nice' for others, use cond_resched().
4274 * If you still want to use yield(), do not!
4276 void __sched yield(void)
4278 set_current_state(TASK_RUNNING);
4281 EXPORT_SYMBOL(yield);
4284 * yield_to - yield the current processor to another thread in
4285 * your thread group, or accelerate that thread toward the
4286 * processor it's on.
4288 * @preempt: whether task preemption is allowed or not
4290 * It's the caller's job to ensure that the target task struct
4291 * can't go away on us before we can do any checks.
4294 * true (>0) if we indeed boosted the target task.
4295 * false (0) if we failed to boost the target.
4296 * -ESRCH if there's no task to yield to.
4298 int __sched yield_to(struct task_struct *p, bool preempt)
4300 struct task_struct *curr = current;
4301 struct rq *rq, *p_rq;
4302 unsigned long flags;
4305 local_irq_save(flags);
4311 * If we're the only runnable task on the rq and target rq also
4312 * has only one task, there's absolutely no point in yielding.
4314 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4319 double_rq_lock(rq, p_rq);
4320 if (task_rq(p) != p_rq) {
4321 double_rq_unlock(rq, p_rq);
4325 if (!curr->sched_class->yield_to_task)
4328 if (curr->sched_class != p->sched_class)
4331 if (task_running(p_rq, p) || p->state)
4334 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4336 schedstat_inc(rq, yld_count);
4338 * Make p's CPU reschedule; pick_next_entity takes care of
4341 if (preempt && rq != p_rq)
4346 double_rq_unlock(rq, p_rq);
4348 local_irq_restore(flags);
4355 EXPORT_SYMBOL_GPL(yield_to);
4358 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4359 * that process accounting knows that this is a task in IO wait state.
4361 long __sched io_schedule_timeout(long timeout)
4363 int old_iowait = current->in_iowait;
4367 current->in_iowait = 1;
4369 blk_schedule_flush_plug(current);
4371 blk_flush_plug(current);
4373 delayacct_blkio_start();
4375 atomic_inc(&rq->nr_iowait);
4376 ret = schedule_timeout(timeout);
4377 current->in_iowait = old_iowait;
4378 atomic_dec(&rq->nr_iowait);
4379 delayacct_blkio_end();
4383 EXPORT_SYMBOL(io_schedule_timeout);
4386 * sys_sched_get_priority_max - return maximum RT priority.
4387 * @policy: scheduling class.
4389 * Return: On success, this syscall returns the maximum
4390 * rt_priority that can be used by a given scheduling class.
4391 * On failure, a negative error code is returned.
4393 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4400 ret = MAX_USER_RT_PRIO-1;
4402 case SCHED_DEADLINE:
4413 * sys_sched_get_priority_min - return minimum RT priority.
4414 * @policy: scheduling class.
4416 * Return: On success, this syscall returns the minimum
4417 * rt_priority that can be used by a given scheduling class.
4418 * On failure, a negative error code is returned.
4420 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4429 case SCHED_DEADLINE:
4439 * sys_sched_rr_get_interval - return the default timeslice of a process.
4440 * @pid: pid of the process.
4441 * @interval: userspace pointer to the timeslice value.
4443 * this syscall writes the default timeslice value of a given process
4444 * into the user-space timespec buffer. A value of '0' means infinity.
4446 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4449 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4450 struct timespec __user *, interval)
4452 struct task_struct *p;
4453 unsigned int time_slice;
4454 unsigned long flags;
4464 p = find_process_by_pid(pid);
4468 retval = security_task_getscheduler(p);
4472 rq = task_rq_lock(p, &flags);
4474 if (p->sched_class->get_rr_interval)
4475 time_slice = p->sched_class->get_rr_interval(rq, p);
4476 task_rq_unlock(rq, p, &flags);
4479 jiffies_to_timespec(time_slice, &t);
4480 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4488 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4490 void sched_show_task(struct task_struct *p)
4492 unsigned long free = 0;
4494 unsigned long state = p->state;
4497 state = __ffs(state) + 1;
4498 printk(KERN_INFO "%-15.15s %c", p->comm,
4499 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4500 #if BITS_PER_LONG == 32
4501 if (state == TASK_RUNNING)
4502 printk(KERN_CONT " running ");
4504 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4506 if (state == TASK_RUNNING)
4507 printk(KERN_CONT " running task ");
4509 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4511 #ifdef CONFIG_DEBUG_STACK_USAGE
4512 free = stack_not_used(p);
4517 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4519 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4520 task_pid_nr(p), ppid,
4521 (unsigned long)task_thread_info(p)->flags);
4523 print_worker_info(KERN_INFO, p);
4524 show_stack(p, NULL);
4527 void show_state_filter(unsigned long state_filter)
4529 struct task_struct *g, *p;
4531 #if BITS_PER_LONG == 32
4533 " task PC stack pid father\n");
4536 " task PC stack pid father\n");
4539 for_each_process_thread(g, p) {
4541 * reset the NMI-timeout, listing all files on a slow
4542 * console might take a lot of time:
4544 touch_nmi_watchdog();
4545 if (!state_filter || (p->state & state_filter))
4549 touch_all_softlockup_watchdogs();
4551 #ifdef CONFIG_SCHED_DEBUG
4552 sysrq_sched_debug_show();
4556 * Only show locks if all tasks are dumped:
4559 debug_show_all_locks();
4562 void init_idle_bootup_task(struct task_struct *idle)
4564 idle->sched_class = &idle_sched_class;
4568 * init_idle - set up an idle thread for a given CPU
4569 * @idle: task in question
4570 * @cpu: cpu the idle task belongs to
4572 * NOTE: this function does not set the idle thread's NEED_RESCHED
4573 * flag, to make booting more robust.
4575 void init_idle(struct task_struct *idle, int cpu)
4577 struct rq *rq = cpu_rq(cpu);
4578 unsigned long flags;
4580 raw_spin_lock_irqsave(&rq->lock, flags);
4582 __sched_fork(0, idle);
4583 idle->state = TASK_RUNNING;
4584 idle->se.exec_start = sched_clock();
4586 do_set_cpus_allowed(idle, cpumask_of(cpu));
4588 * We're having a chicken and egg problem, even though we are
4589 * holding rq->lock, the cpu isn't yet set to this cpu so the
4590 * lockdep check in task_group() will fail.
4592 * Similar case to sched_fork(). / Alternatively we could
4593 * use task_rq_lock() here and obtain the other rq->lock.
4598 __set_task_cpu(idle, cpu);
4601 rq->curr = rq->idle = idle;
4602 idle->on_rq = TASK_ON_RQ_QUEUED;
4603 #if defined(CONFIG_SMP)
4606 raw_spin_unlock_irqrestore(&rq->lock, flags);
4608 /* Set the preempt count _outside_ the spinlocks! */
4609 init_idle_preempt_count(idle, cpu);
4612 * The idle tasks have their own, simple scheduling class:
4614 idle->sched_class = &idle_sched_class;
4615 ftrace_graph_init_idle_task(idle, cpu);
4616 vtime_init_idle(idle, cpu);
4617 #if defined(CONFIG_SMP)
4618 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4622 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4623 const struct cpumask *trial)
4625 int ret = 1, trial_cpus;
4626 struct dl_bw *cur_dl_b;
4627 unsigned long flags;
4629 if (!cpumask_weight(cur))
4632 rcu_read_lock_sched();
4633 cur_dl_b = dl_bw_of(cpumask_any(cur));
4634 trial_cpus = cpumask_weight(trial);
4636 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4637 if (cur_dl_b->bw != -1 &&
4638 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4640 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4641 rcu_read_unlock_sched();
4646 int task_can_attach(struct task_struct *p,
4647 const struct cpumask *cs_cpus_allowed)
4652 * Kthreads which disallow setaffinity shouldn't be moved
4653 * to a new cpuset; we don't want to change their cpu
4654 * affinity and isolating such threads by their set of
4655 * allowed nodes is unnecessary. Thus, cpusets are not
4656 * applicable for such threads. This prevents checking for
4657 * success of set_cpus_allowed_ptr() on all attached tasks
4658 * before cpus_allowed may be changed.
4660 if (p->flags & PF_NO_SETAFFINITY) {
4666 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4668 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4673 unsigned long flags;
4675 rcu_read_lock_sched();
4676 dl_b = dl_bw_of(dest_cpu);
4677 raw_spin_lock_irqsave(&dl_b->lock, flags);
4678 cpus = dl_bw_cpus(dest_cpu);
4679 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4684 * We reserve space for this task in the destination
4685 * root_domain, as we can't fail after this point.
4686 * We will free resources in the source root_domain
4687 * later on (see set_cpus_allowed_dl()).
4689 __dl_add(dl_b, p->dl.dl_bw);
4691 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4692 rcu_read_unlock_sched();
4702 * move_queued_task - move a queued task to new rq.
4704 * Returns (locked) new rq. Old rq's lock is released.
4706 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4708 struct rq *rq = task_rq(p);
4710 lockdep_assert_held(&rq->lock);
4712 dequeue_task(rq, p, 0);
4713 p->on_rq = TASK_ON_RQ_MIGRATING;
4714 set_task_cpu(p, new_cpu);
4715 raw_spin_unlock(&rq->lock);
4717 rq = cpu_rq(new_cpu);
4719 raw_spin_lock(&rq->lock);
4720 BUG_ON(task_cpu(p) != new_cpu);
4721 p->on_rq = TASK_ON_RQ_QUEUED;
4722 enqueue_task(rq, p, 0);
4723 check_preempt_curr(rq, p, 0);
4728 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4730 if (p->sched_class->set_cpus_allowed)
4731 p->sched_class->set_cpus_allowed(p, new_mask);
4733 cpumask_copy(&p->cpus_allowed, new_mask);
4734 p->nr_cpus_allowed = cpumask_weight(new_mask);
4738 * This is how migration works:
4740 * 1) we invoke migration_cpu_stop() on the target CPU using
4742 * 2) stopper starts to run (implicitly forcing the migrated thread
4744 * 3) it checks whether the migrated task is still in the wrong runqueue.
4745 * 4) if it's in the wrong runqueue then the migration thread removes
4746 * it and puts it into the right queue.
4747 * 5) stopper completes and stop_one_cpu() returns and the migration
4752 * Change a given task's CPU affinity. Migrate the thread to a
4753 * proper CPU and schedule it away if the CPU it's executing on
4754 * is removed from the allowed bitmask.
4756 * NOTE: the caller must have a valid reference to the task, the
4757 * task must not exit() & deallocate itself prematurely. The
4758 * call is not atomic; no spinlocks may be held.
4760 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4762 unsigned long flags;
4764 unsigned int dest_cpu;
4767 rq = task_rq_lock(p, &flags);
4769 if (cpumask_equal(&p->cpus_allowed, new_mask))
4772 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4777 do_set_cpus_allowed(p, new_mask);
4779 /* Can the task run on the task's current CPU? If so, we're done */
4780 if (cpumask_test_cpu(task_cpu(p), new_mask))
4783 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4784 if (task_running(rq, p) || p->state == TASK_WAKING) {
4785 struct migration_arg arg = { p, dest_cpu };
4786 /* Need help from migration thread: drop lock and wait. */
4787 task_rq_unlock(rq, p, &flags);
4788 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4789 tlb_migrate_finish(p->mm);
4791 } else if (task_on_rq_queued(p))
4792 rq = move_queued_task(p, dest_cpu);
4794 task_rq_unlock(rq, p, &flags);
4798 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4801 * Move (not current) task off this cpu, onto dest cpu. We're doing
4802 * this because either it can't run here any more (set_cpus_allowed()
4803 * away from this CPU, or CPU going down), or because we're
4804 * attempting to rebalance this task on exec (sched_exec).
4806 * So we race with normal scheduler movements, but that's OK, as long
4807 * as the task is no longer on this CPU.
4809 * Returns non-zero if task was successfully migrated.
4811 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4816 if (unlikely(!cpu_active(dest_cpu)))
4819 rq = cpu_rq(src_cpu);
4821 raw_spin_lock(&p->pi_lock);
4822 raw_spin_lock(&rq->lock);
4823 /* Already moved. */
4824 if (task_cpu(p) != src_cpu)
4827 /* Affinity changed (again). */
4828 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4832 * If we're not on a rq, the next wake-up will ensure we're
4835 if (task_on_rq_queued(p))
4836 rq = move_queued_task(p, dest_cpu);
4840 raw_spin_unlock(&rq->lock);
4841 raw_spin_unlock(&p->pi_lock);
4845 #ifdef CONFIG_NUMA_BALANCING
4846 /* Migrate current task p to target_cpu */
4847 int migrate_task_to(struct task_struct *p, int target_cpu)
4849 struct migration_arg arg = { p, target_cpu };
4850 int curr_cpu = task_cpu(p);
4852 if (curr_cpu == target_cpu)
4855 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4858 /* TODO: This is not properly updating schedstats */
4860 trace_sched_move_numa(p, curr_cpu, target_cpu);
4861 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4865 * Requeue a task on a given node and accurately track the number of NUMA
4866 * tasks on the runqueues
4868 void sched_setnuma(struct task_struct *p, int nid)
4871 unsigned long flags;
4872 bool queued, running;
4874 rq = task_rq_lock(p, &flags);
4875 queued = task_on_rq_queued(p);
4876 running = task_current(rq, p);
4879 dequeue_task(rq, p, 0);
4881 put_prev_task(rq, p);
4883 p->numa_preferred_nid = nid;
4886 p->sched_class->set_curr_task(rq);
4888 enqueue_task(rq, p, 0);
4889 task_rq_unlock(rq, p, &flags);
4894 * migration_cpu_stop - this will be executed by a highprio stopper thread
4895 * and performs thread migration by bumping thread off CPU then
4896 * 'pushing' onto another runqueue.
4898 static int migration_cpu_stop(void *data)
4900 struct migration_arg *arg = data;
4903 * The original target cpu might have gone down and we might
4904 * be on another cpu but it doesn't matter.
4906 local_irq_disable();
4908 * We need to explicitly wake pending tasks before running
4909 * __migrate_task() such that we will not miss enforcing cpus_allowed
4910 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4912 sched_ttwu_pending();
4913 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4918 #ifdef CONFIG_HOTPLUG_CPU
4921 * Ensures that the idle task is using init_mm right before its cpu goes
4924 void idle_task_exit(void)
4926 struct mm_struct *mm = current->active_mm;
4928 BUG_ON(cpu_online(smp_processor_id()));
4930 if (mm != &init_mm) {
4931 switch_mm(mm, &init_mm, current);
4932 finish_arch_post_lock_switch();
4938 * Since this CPU is going 'away' for a while, fold any nr_active delta
4939 * we might have. Assumes we're called after migrate_tasks() so that the
4940 * nr_active count is stable.
4942 * Also see the comment "Global load-average calculations".
4944 static void calc_load_migrate(struct rq *rq)
4946 long delta = calc_load_fold_active(rq);
4948 atomic_long_add(delta, &calc_load_tasks);
4951 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4955 static const struct sched_class fake_sched_class = {
4956 .put_prev_task = put_prev_task_fake,
4959 static struct task_struct fake_task = {
4961 * Avoid pull_{rt,dl}_task()
4963 .prio = MAX_PRIO + 1,
4964 .sched_class = &fake_sched_class,
4968 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4969 * try_to_wake_up()->select_task_rq().
4971 * Called with rq->lock held even though we'er in stop_machine() and
4972 * there's no concurrency possible, we hold the required locks anyway
4973 * because of lock validation efforts.
4975 static void migrate_tasks(unsigned int dead_cpu)
4977 struct rq *rq = cpu_rq(dead_cpu);
4978 struct task_struct *next, *stop = rq->stop;
4982 * Fudge the rq selection such that the below task selection loop
4983 * doesn't get stuck on the currently eligible stop task.
4985 * We're currently inside stop_machine() and the rq is either stuck
4986 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4987 * either way we should never end up calling schedule() until we're
4993 * put_prev_task() and pick_next_task() sched
4994 * class method both need to have an up-to-date
4995 * value of rq->clock[_task]
4997 update_rq_clock(rq);
5001 * There's this thread running, bail when that's the only
5004 if (rq->nr_running == 1)
5007 next = pick_next_task(rq, &fake_task);
5009 next->sched_class->put_prev_task(rq, next);
5011 /* Find suitable destination for @next, with force if needed. */
5012 dest_cpu = select_fallback_rq(dead_cpu, next);
5013 raw_spin_unlock(&rq->lock);
5015 __migrate_task(next, dead_cpu, dest_cpu);
5017 raw_spin_lock(&rq->lock);
5023 #endif /* CONFIG_HOTPLUG_CPU */
5025 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5027 static struct ctl_table sd_ctl_dir[] = {
5029 .procname = "sched_domain",
5035 static struct ctl_table sd_ctl_root[] = {
5037 .procname = "kernel",
5039 .child = sd_ctl_dir,
5044 static struct ctl_table *sd_alloc_ctl_entry(int n)
5046 struct ctl_table *entry =
5047 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5052 static void sd_free_ctl_entry(struct ctl_table **tablep)
5054 struct ctl_table *entry;
5057 * In the intermediate directories, both the child directory and
5058 * procname are dynamically allocated and could fail but the mode
5059 * will always be set. In the lowest directory the names are
5060 * static strings and all have proc handlers.
5062 for (entry = *tablep; entry->mode; entry++) {
5064 sd_free_ctl_entry(&entry->child);
5065 if (entry->proc_handler == NULL)
5066 kfree(entry->procname);
5073 static int min_load_idx = 0;
5074 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5077 set_table_entry(struct ctl_table *entry,
5078 const char *procname, void *data, int maxlen,
5079 umode_t mode, proc_handler *proc_handler,
5082 entry->procname = procname;
5084 entry->maxlen = maxlen;
5086 entry->proc_handler = proc_handler;
5089 entry->extra1 = &min_load_idx;
5090 entry->extra2 = &max_load_idx;
5094 static struct ctl_table *
5095 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5097 struct ctl_table *table = sd_alloc_ctl_entry(14);
5102 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5103 sizeof(long), 0644, proc_doulongvec_minmax, false);
5104 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5105 sizeof(long), 0644, proc_doulongvec_minmax, false);
5106 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5107 sizeof(int), 0644, proc_dointvec_minmax, true);
5108 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5109 sizeof(int), 0644, proc_dointvec_minmax, true);
5110 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5111 sizeof(int), 0644, proc_dointvec_minmax, true);
5112 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5113 sizeof(int), 0644, proc_dointvec_minmax, true);
5114 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5115 sizeof(int), 0644, proc_dointvec_minmax, true);
5116 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5117 sizeof(int), 0644, proc_dointvec_minmax, false);
5118 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5119 sizeof(int), 0644, proc_dointvec_minmax, false);
5120 set_table_entry(&table[9], "cache_nice_tries",
5121 &sd->cache_nice_tries,
5122 sizeof(int), 0644, proc_dointvec_minmax, false);
5123 set_table_entry(&table[10], "flags", &sd->flags,
5124 sizeof(int), 0644, proc_dointvec_minmax, false);
5125 set_table_entry(&table[11], "max_newidle_lb_cost",
5126 &sd->max_newidle_lb_cost,
5127 sizeof(long), 0644, proc_doulongvec_minmax, false);
5128 set_table_entry(&table[12], "name", sd->name,
5129 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5130 /* &table[13] is terminator */
5135 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5137 struct ctl_table *entry, *table;
5138 struct sched_domain *sd;
5139 int domain_num = 0, i;
5142 for_each_domain(cpu, sd)
5144 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5149 for_each_domain(cpu, sd) {
5150 snprintf(buf, 32, "domain%d", i);
5151 entry->procname = kstrdup(buf, GFP_KERNEL);
5153 entry->child = sd_alloc_ctl_domain_table(sd);
5160 static struct ctl_table_header *sd_sysctl_header;
5161 static void register_sched_domain_sysctl(void)
5163 int i, cpu_num = num_possible_cpus();
5164 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5167 WARN_ON(sd_ctl_dir[0].child);
5168 sd_ctl_dir[0].child = entry;
5173 for_each_possible_cpu(i) {
5174 snprintf(buf, 32, "cpu%d", i);
5175 entry->procname = kstrdup(buf, GFP_KERNEL);
5177 entry->child = sd_alloc_ctl_cpu_table(i);
5181 WARN_ON(sd_sysctl_header);
5182 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5185 /* may be called multiple times per register */
5186 static void unregister_sched_domain_sysctl(void)
5188 if (sd_sysctl_header)
5189 unregister_sysctl_table(sd_sysctl_header);
5190 sd_sysctl_header = NULL;
5191 if (sd_ctl_dir[0].child)
5192 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5195 static void register_sched_domain_sysctl(void)
5198 static void unregister_sched_domain_sysctl(void)
5203 static void set_rq_online(struct rq *rq)
5206 const struct sched_class *class;
5208 cpumask_set_cpu(rq->cpu, rq->rd->online);
5211 for_each_class(class) {
5212 if (class->rq_online)
5213 class->rq_online(rq);
5218 static void set_rq_offline(struct rq *rq)
5221 const struct sched_class *class;
5223 for_each_class(class) {
5224 if (class->rq_offline)
5225 class->rq_offline(rq);
5228 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5234 * migration_call - callback that gets triggered when a CPU is added.
5235 * Here we can start up the necessary migration thread for the new CPU.
5238 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5240 int cpu = (long)hcpu;
5241 unsigned long flags;
5242 struct rq *rq = cpu_rq(cpu);
5244 switch (action & ~CPU_TASKS_FROZEN) {
5246 case CPU_UP_PREPARE:
5247 rq->calc_load_update = calc_load_update;
5251 /* Update our root-domain */
5252 raw_spin_lock_irqsave(&rq->lock, flags);
5254 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5258 raw_spin_unlock_irqrestore(&rq->lock, flags);
5261 #ifdef CONFIG_HOTPLUG_CPU
5263 sched_ttwu_pending();
5264 /* Update our root-domain */
5265 raw_spin_lock_irqsave(&rq->lock, flags);
5267 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5271 BUG_ON(rq->nr_running != 1); /* the migration thread */
5272 raw_spin_unlock_irqrestore(&rq->lock, flags);
5276 calc_load_migrate(rq);
5281 update_max_interval();
5287 * Register at high priority so that task migration (migrate_all_tasks)
5288 * happens before everything else. This has to be lower priority than
5289 * the notifier in the perf_event subsystem, though.
5291 static struct notifier_block migration_notifier = {
5292 .notifier_call = migration_call,
5293 .priority = CPU_PRI_MIGRATION,
5296 static void __cpuinit set_cpu_rq_start_time(void)
5298 int cpu = smp_processor_id();
5299 struct rq *rq = cpu_rq(cpu);
5300 rq->age_stamp = sched_clock_cpu(cpu);
5303 static int sched_cpu_active(struct notifier_block *nfb,
5304 unsigned long action, void *hcpu)
5306 switch (action & ~CPU_TASKS_FROZEN) {
5308 set_cpu_rq_start_time();
5310 case CPU_DOWN_FAILED:
5311 set_cpu_active((long)hcpu, true);
5318 static int sched_cpu_inactive(struct notifier_block *nfb,
5319 unsigned long action, void *hcpu)
5321 unsigned long flags;
5322 long cpu = (long)hcpu;
5325 switch (action & ~CPU_TASKS_FROZEN) {
5326 case CPU_DOWN_PREPARE:
5327 set_cpu_active(cpu, false);
5329 /* explicitly allow suspend */
5330 if (!(action & CPU_TASKS_FROZEN)) {
5334 rcu_read_lock_sched();
5335 dl_b = dl_bw_of(cpu);
5337 raw_spin_lock_irqsave(&dl_b->lock, flags);
5338 cpus = dl_bw_cpus(cpu);
5339 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5340 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5342 rcu_read_unlock_sched();
5345 return notifier_from_errno(-EBUSY);
5353 static int __init migration_init(void)
5355 void *cpu = (void *)(long)smp_processor_id();
5358 /* Initialize migration for the boot CPU */
5359 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5360 BUG_ON(err == NOTIFY_BAD);
5361 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5362 register_cpu_notifier(&migration_notifier);
5364 /* Register cpu active notifiers */
5365 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5366 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5370 early_initcall(migration_init);
5375 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5377 #ifdef CONFIG_SCHED_DEBUG
5379 static __read_mostly int sched_debug_enabled;
5381 static int __init sched_debug_setup(char *str)
5383 sched_debug_enabled = 1;
5387 early_param("sched_debug", sched_debug_setup);
5389 static inline bool sched_debug(void)
5391 return sched_debug_enabled;
5394 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5395 struct cpumask *groupmask)
5397 struct sched_group *group = sd->groups;
5399 cpumask_clear(groupmask);
5401 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5403 if (!(sd->flags & SD_LOAD_BALANCE)) {
5404 printk("does not load-balance\n");
5406 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5411 printk(KERN_CONT "span %*pbl level %s\n",
5412 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5414 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5415 printk(KERN_ERR "ERROR: domain->span does not contain "
5418 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5419 printk(KERN_ERR "ERROR: domain->groups does not contain"
5423 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5427 printk(KERN_ERR "ERROR: group is NULL\n");
5432 * Even though we initialize ->capacity to something semi-sane,
5433 * we leave capacity_orig unset. This allows us to detect if
5434 * domain iteration is still funny without causing /0 traps.
5436 if (!group->sgc->capacity_orig) {
5437 printk(KERN_CONT "\n");
5438 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5442 if (!cpumask_weight(sched_group_cpus(group))) {
5443 printk(KERN_CONT "\n");
5444 printk(KERN_ERR "ERROR: empty group\n");
5448 if (!(sd->flags & SD_OVERLAP) &&
5449 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5450 printk(KERN_CONT "\n");
5451 printk(KERN_ERR "ERROR: repeated CPUs\n");
5455 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5457 printk(KERN_CONT " %*pbl",
5458 cpumask_pr_args(sched_group_cpus(group)));
5459 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5460 printk(KERN_CONT " (cpu_capacity = %d)",
5461 group->sgc->capacity);
5464 group = group->next;
5465 } while (group != sd->groups);
5466 printk(KERN_CONT "\n");
5468 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5469 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5472 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5473 printk(KERN_ERR "ERROR: parent span is not a superset "
5474 "of domain->span\n");
5478 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5482 if (!sched_debug_enabled)
5486 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5490 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5493 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5501 #else /* !CONFIG_SCHED_DEBUG */
5502 # define sched_domain_debug(sd, cpu) do { } while (0)
5503 static inline bool sched_debug(void)
5507 #endif /* CONFIG_SCHED_DEBUG */
5509 static int sd_degenerate(struct sched_domain *sd)
5511 if (cpumask_weight(sched_domain_span(sd)) == 1)
5514 /* Following flags need at least 2 groups */
5515 if (sd->flags & (SD_LOAD_BALANCE |
5516 SD_BALANCE_NEWIDLE |
5519 SD_SHARE_CPUCAPACITY |
5520 SD_SHARE_PKG_RESOURCES |
5521 SD_SHARE_POWERDOMAIN)) {
5522 if (sd->groups != sd->groups->next)
5526 /* Following flags don't use groups */
5527 if (sd->flags & (SD_WAKE_AFFINE))
5534 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5536 unsigned long cflags = sd->flags, pflags = parent->flags;
5538 if (sd_degenerate(parent))
5541 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5544 /* Flags needing groups don't count if only 1 group in parent */
5545 if (parent->groups == parent->groups->next) {
5546 pflags &= ~(SD_LOAD_BALANCE |
5547 SD_BALANCE_NEWIDLE |
5550 SD_SHARE_CPUCAPACITY |
5551 SD_SHARE_PKG_RESOURCES |
5553 SD_SHARE_POWERDOMAIN);
5554 if (nr_node_ids == 1)
5555 pflags &= ~SD_SERIALIZE;
5557 if (~cflags & pflags)
5563 static void free_rootdomain(struct rcu_head *rcu)
5565 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5567 cpupri_cleanup(&rd->cpupri);
5568 cpudl_cleanup(&rd->cpudl);
5569 free_cpumask_var(rd->dlo_mask);
5570 free_cpumask_var(rd->rto_mask);
5571 free_cpumask_var(rd->online);
5572 free_cpumask_var(rd->span);
5576 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5578 struct root_domain *old_rd = NULL;
5579 unsigned long flags;
5581 raw_spin_lock_irqsave(&rq->lock, flags);
5586 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5589 cpumask_clear_cpu(rq->cpu, old_rd->span);
5592 * If we dont want to free the old_rd yet then
5593 * set old_rd to NULL to skip the freeing later
5596 if (!atomic_dec_and_test(&old_rd->refcount))
5600 atomic_inc(&rd->refcount);
5603 cpumask_set_cpu(rq->cpu, rd->span);
5604 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5607 raw_spin_unlock_irqrestore(&rq->lock, flags);
5610 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5613 static int init_rootdomain(struct root_domain *rd)
5615 memset(rd, 0, sizeof(*rd));
5617 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5619 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5621 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5623 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5626 init_dl_bw(&rd->dl_bw);
5627 if (cpudl_init(&rd->cpudl) != 0)
5630 if (cpupri_init(&rd->cpupri) != 0)
5635 free_cpumask_var(rd->rto_mask);
5637 free_cpumask_var(rd->dlo_mask);
5639 free_cpumask_var(rd->online);
5641 free_cpumask_var(rd->span);
5647 * By default the system creates a single root-domain with all cpus as
5648 * members (mimicking the global state we have today).
5650 struct root_domain def_root_domain;
5652 static void init_defrootdomain(void)
5654 init_rootdomain(&def_root_domain);
5656 atomic_set(&def_root_domain.refcount, 1);
5659 static struct root_domain *alloc_rootdomain(void)
5661 struct root_domain *rd;
5663 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5667 if (init_rootdomain(rd) != 0) {
5675 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5677 struct sched_group *tmp, *first;
5686 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5691 } while (sg != first);
5694 static void free_sched_domain(struct rcu_head *rcu)
5696 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5699 * If its an overlapping domain it has private groups, iterate and
5702 if (sd->flags & SD_OVERLAP) {
5703 free_sched_groups(sd->groups, 1);
5704 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5705 kfree(sd->groups->sgc);
5711 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5713 call_rcu(&sd->rcu, free_sched_domain);
5716 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5718 for (; sd; sd = sd->parent)
5719 destroy_sched_domain(sd, cpu);
5723 * Keep a special pointer to the highest sched_domain that has
5724 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5725 * allows us to avoid some pointer chasing select_idle_sibling().
5727 * Also keep a unique ID per domain (we use the first cpu number in
5728 * the cpumask of the domain), this allows us to quickly tell if
5729 * two cpus are in the same cache domain, see cpus_share_cache().
5731 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5732 DEFINE_PER_CPU(int, sd_llc_size);
5733 DEFINE_PER_CPU(int, sd_llc_id);
5734 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5735 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5736 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5738 static void update_top_cache_domain(int cpu)
5740 struct sched_domain *sd;
5741 struct sched_domain *busy_sd = NULL;
5745 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5747 id = cpumask_first(sched_domain_span(sd));
5748 size = cpumask_weight(sched_domain_span(sd));
5749 busy_sd = sd->parent; /* sd_busy */
5751 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5753 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5754 per_cpu(sd_llc_size, cpu) = size;
5755 per_cpu(sd_llc_id, cpu) = id;
5757 sd = lowest_flag_domain(cpu, SD_NUMA);
5758 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5760 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5761 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5765 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5766 * hold the hotplug lock.
5769 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5771 struct rq *rq = cpu_rq(cpu);
5772 struct sched_domain *tmp;
5774 /* Remove the sched domains which do not contribute to scheduling. */
5775 for (tmp = sd; tmp; ) {
5776 struct sched_domain *parent = tmp->parent;
5780 if (sd_parent_degenerate(tmp, parent)) {
5781 tmp->parent = parent->parent;
5783 parent->parent->child = tmp;
5785 * Transfer SD_PREFER_SIBLING down in case of a
5786 * degenerate parent; the spans match for this
5787 * so the property transfers.
5789 if (parent->flags & SD_PREFER_SIBLING)
5790 tmp->flags |= SD_PREFER_SIBLING;
5791 destroy_sched_domain(parent, cpu);
5796 if (sd && sd_degenerate(sd)) {
5799 destroy_sched_domain(tmp, cpu);
5804 sched_domain_debug(sd, cpu);
5806 rq_attach_root(rq, rd);
5808 rcu_assign_pointer(rq->sd, sd);
5809 destroy_sched_domains(tmp, cpu);
5811 update_top_cache_domain(cpu);
5814 /* cpus with isolated domains */
5815 static cpumask_var_t cpu_isolated_map;
5817 /* Setup the mask of cpus configured for isolated domains */
5818 static int __init isolated_cpu_setup(char *str)
5820 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5821 cpulist_parse(str, cpu_isolated_map);
5825 __setup("isolcpus=", isolated_cpu_setup);
5828 struct sched_domain ** __percpu sd;
5829 struct root_domain *rd;
5840 * Build an iteration mask that can exclude certain CPUs from the upwards
5843 * Asymmetric node setups can result in situations where the domain tree is of
5844 * unequal depth, make sure to skip domains that already cover the entire
5847 * In that case build_sched_domains() will have terminated the iteration early
5848 * and our sibling sd spans will be empty. Domains should always include the
5849 * cpu they're built on, so check that.
5852 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5854 const struct cpumask *span = sched_domain_span(sd);
5855 struct sd_data *sdd = sd->private;
5856 struct sched_domain *sibling;
5859 for_each_cpu(i, span) {
5860 sibling = *per_cpu_ptr(sdd->sd, i);
5861 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5864 cpumask_set_cpu(i, sched_group_mask(sg));
5869 * Return the canonical balance cpu for this group, this is the first cpu
5870 * of this group that's also in the iteration mask.
5872 int group_balance_cpu(struct sched_group *sg)
5874 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5878 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5880 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5881 const struct cpumask *span = sched_domain_span(sd);
5882 struct cpumask *covered = sched_domains_tmpmask;
5883 struct sd_data *sdd = sd->private;
5884 struct sched_domain *sibling;
5887 cpumask_clear(covered);
5889 for_each_cpu(i, span) {
5890 struct cpumask *sg_span;
5892 if (cpumask_test_cpu(i, covered))
5895 sibling = *per_cpu_ptr(sdd->sd, i);
5897 /* See the comment near build_group_mask(). */
5898 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5901 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5902 GFP_KERNEL, cpu_to_node(cpu));
5907 sg_span = sched_group_cpus(sg);
5909 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5911 cpumask_set_cpu(i, sg_span);
5913 cpumask_or(covered, covered, sg_span);
5915 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5916 if (atomic_inc_return(&sg->sgc->ref) == 1)
5917 build_group_mask(sd, sg);
5920 * Initialize sgc->capacity such that even if we mess up the
5921 * domains and no possible iteration will get us here, we won't
5924 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5925 sg->sgc->capacity_orig = sg->sgc->capacity;
5928 * Make sure the first group of this domain contains the
5929 * canonical balance cpu. Otherwise the sched_domain iteration
5930 * breaks. See update_sg_lb_stats().
5932 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5933 group_balance_cpu(sg) == cpu)
5943 sd->groups = groups;
5948 free_sched_groups(first, 0);
5953 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5955 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5956 struct sched_domain *child = sd->child;
5959 cpu = cpumask_first(sched_domain_span(child));
5962 *sg = *per_cpu_ptr(sdd->sg, cpu);
5963 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5964 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5971 * build_sched_groups will build a circular linked list of the groups
5972 * covered by the given span, and will set each group's ->cpumask correctly,
5973 * and ->cpu_capacity to 0.
5975 * Assumes the sched_domain tree is fully constructed
5978 build_sched_groups(struct sched_domain *sd, int cpu)
5980 struct sched_group *first = NULL, *last = NULL;
5981 struct sd_data *sdd = sd->private;
5982 const struct cpumask *span = sched_domain_span(sd);
5983 struct cpumask *covered;
5986 get_group(cpu, sdd, &sd->groups);
5987 atomic_inc(&sd->groups->ref);
5989 if (cpu != cpumask_first(span))
5992 lockdep_assert_held(&sched_domains_mutex);
5993 covered = sched_domains_tmpmask;
5995 cpumask_clear(covered);
5997 for_each_cpu(i, span) {
5998 struct sched_group *sg;
6001 if (cpumask_test_cpu(i, covered))
6004 group = get_group(i, sdd, &sg);
6005 cpumask_setall(sched_group_mask(sg));
6007 for_each_cpu(j, span) {
6008 if (get_group(j, sdd, NULL) != group)
6011 cpumask_set_cpu(j, covered);
6012 cpumask_set_cpu(j, sched_group_cpus(sg));
6027 * Initialize sched groups cpu_capacity.
6029 * cpu_capacity indicates the capacity of sched group, which is used while
6030 * distributing the load between different sched groups in a sched domain.
6031 * Typically cpu_capacity for all the groups in a sched domain will be same
6032 * unless there are asymmetries in the topology. If there are asymmetries,
6033 * group having more cpu_capacity will pickup more load compared to the
6034 * group having less cpu_capacity.
6036 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6038 struct sched_group *sg = sd->groups;
6043 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6045 } while (sg != sd->groups);
6047 if (cpu != group_balance_cpu(sg))
6050 update_group_capacity(sd, cpu);
6051 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6055 * Initializers for schedule domains
6056 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6059 static int default_relax_domain_level = -1;
6060 int sched_domain_level_max;
6062 static int __init setup_relax_domain_level(char *str)
6064 if (kstrtoint(str, 0, &default_relax_domain_level))
6065 pr_warn("Unable to set relax_domain_level\n");
6069 __setup("relax_domain_level=", setup_relax_domain_level);
6071 static void set_domain_attribute(struct sched_domain *sd,
6072 struct sched_domain_attr *attr)
6076 if (!attr || attr->relax_domain_level < 0) {
6077 if (default_relax_domain_level < 0)
6080 request = default_relax_domain_level;
6082 request = attr->relax_domain_level;
6083 if (request < sd->level) {
6084 /* turn off idle balance on this domain */
6085 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6087 /* turn on idle balance on this domain */
6088 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6092 static void __sdt_free(const struct cpumask *cpu_map);
6093 static int __sdt_alloc(const struct cpumask *cpu_map);
6095 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6096 const struct cpumask *cpu_map)
6100 if (!atomic_read(&d->rd->refcount))
6101 free_rootdomain(&d->rd->rcu); /* fall through */
6103 free_percpu(d->sd); /* fall through */
6105 __sdt_free(cpu_map); /* fall through */
6111 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6112 const struct cpumask *cpu_map)
6114 memset(d, 0, sizeof(*d));
6116 if (__sdt_alloc(cpu_map))
6117 return sa_sd_storage;
6118 d->sd = alloc_percpu(struct sched_domain *);
6120 return sa_sd_storage;
6121 d->rd = alloc_rootdomain();
6124 return sa_rootdomain;
6128 * NULL the sd_data elements we've used to build the sched_domain and
6129 * sched_group structure so that the subsequent __free_domain_allocs()
6130 * will not free the data we're using.
6132 static void claim_allocations(int cpu, struct sched_domain *sd)
6134 struct sd_data *sdd = sd->private;
6136 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6137 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6139 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6140 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6142 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6143 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6147 static int sched_domains_numa_levels;
6148 enum numa_topology_type sched_numa_topology_type;
6149 static int *sched_domains_numa_distance;
6150 int sched_max_numa_distance;
6151 static struct cpumask ***sched_domains_numa_masks;
6152 static int sched_domains_curr_level;
6156 * SD_flags allowed in topology descriptions.
6158 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6159 * SD_SHARE_PKG_RESOURCES - describes shared caches
6160 * SD_NUMA - describes NUMA topologies
6161 * SD_SHARE_POWERDOMAIN - describes shared power domain
6164 * SD_ASYM_PACKING - describes SMT quirks
6166 #define TOPOLOGY_SD_FLAGS \
6167 (SD_SHARE_CPUCAPACITY | \
6168 SD_SHARE_PKG_RESOURCES | \
6171 SD_SHARE_POWERDOMAIN)
6173 static struct sched_domain *
6174 sd_init(struct sched_domain_topology_level *tl, int cpu)
6176 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6177 int sd_weight, sd_flags = 0;
6181 * Ugly hack to pass state to sd_numa_mask()...
6183 sched_domains_curr_level = tl->numa_level;
6186 sd_weight = cpumask_weight(tl->mask(cpu));
6189 sd_flags = (*tl->sd_flags)();
6190 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6191 "wrong sd_flags in topology description\n"))
6192 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6194 *sd = (struct sched_domain){
6195 .min_interval = sd_weight,
6196 .max_interval = 2*sd_weight,
6198 .imbalance_pct = 125,
6200 .cache_nice_tries = 0,
6207 .flags = 1*SD_LOAD_BALANCE
6208 | 1*SD_BALANCE_NEWIDLE
6213 | 0*SD_SHARE_CPUCAPACITY
6214 | 0*SD_SHARE_PKG_RESOURCES
6216 | 0*SD_PREFER_SIBLING
6221 .last_balance = jiffies,
6222 .balance_interval = sd_weight,
6224 .max_newidle_lb_cost = 0,
6225 .next_decay_max_lb_cost = jiffies,
6226 #ifdef CONFIG_SCHED_DEBUG
6232 * Convert topological properties into behaviour.
6235 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6236 sd->imbalance_pct = 110;
6237 sd->smt_gain = 1178; /* ~15% */
6239 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6240 sd->imbalance_pct = 117;
6241 sd->cache_nice_tries = 1;
6245 } else if (sd->flags & SD_NUMA) {
6246 sd->cache_nice_tries = 2;
6250 sd->flags |= SD_SERIALIZE;
6251 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6252 sd->flags &= ~(SD_BALANCE_EXEC |
6259 sd->flags |= SD_PREFER_SIBLING;
6260 sd->cache_nice_tries = 1;
6265 sd->private = &tl->data;
6271 * Topology list, bottom-up.
6273 static struct sched_domain_topology_level default_topology[] = {
6274 #ifdef CONFIG_SCHED_SMT
6275 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6277 #ifdef CONFIG_SCHED_MC
6278 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6280 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6284 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6286 #define for_each_sd_topology(tl) \
6287 for (tl = sched_domain_topology; tl->mask; tl++)
6289 void set_sched_topology(struct sched_domain_topology_level *tl)
6291 sched_domain_topology = tl;
6296 static const struct cpumask *sd_numa_mask(int cpu)
6298 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6301 static void sched_numa_warn(const char *str)
6303 static int done = false;
6311 printk(KERN_WARNING "ERROR: %s\n\n", str);
6313 for (i = 0; i < nr_node_ids; i++) {
6314 printk(KERN_WARNING " ");
6315 for (j = 0; j < nr_node_ids; j++)
6316 printk(KERN_CONT "%02d ", node_distance(i,j));
6317 printk(KERN_CONT "\n");
6319 printk(KERN_WARNING "\n");
6322 bool find_numa_distance(int distance)
6326 if (distance == node_distance(0, 0))
6329 for (i = 0; i < sched_domains_numa_levels; i++) {
6330 if (sched_domains_numa_distance[i] == distance)
6338 * A system can have three types of NUMA topology:
6339 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6340 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6341 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6343 * The difference between a glueless mesh topology and a backplane
6344 * topology lies in whether communication between not directly
6345 * connected nodes goes through intermediary nodes (where programs
6346 * could run), or through backplane controllers. This affects
6347 * placement of programs.
6349 * The type of topology can be discerned with the following tests:
6350 * - If the maximum distance between any nodes is 1 hop, the system
6351 * is directly connected.
6352 * - If for two nodes A and B, located N > 1 hops away from each other,
6353 * there is an intermediary node C, which is < N hops away from both
6354 * nodes A and B, the system is a glueless mesh.
6356 static void init_numa_topology_type(void)
6360 n = sched_max_numa_distance;
6363 sched_numa_topology_type = NUMA_DIRECT;
6365 for_each_online_node(a) {
6366 for_each_online_node(b) {
6367 /* Find two nodes furthest removed from each other. */
6368 if (node_distance(a, b) < n)
6371 /* Is there an intermediary node between a and b? */
6372 for_each_online_node(c) {
6373 if (node_distance(a, c) < n &&
6374 node_distance(b, c) < n) {
6375 sched_numa_topology_type =
6381 sched_numa_topology_type = NUMA_BACKPLANE;
6387 static void sched_init_numa(void)
6389 int next_distance, curr_distance = node_distance(0, 0);
6390 struct sched_domain_topology_level *tl;
6394 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6395 if (!sched_domains_numa_distance)
6399 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6400 * unique distances in the node_distance() table.
6402 * Assumes node_distance(0,j) includes all distances in
6403 * node_distance(i,j) in order to avoid cubic time.
6405 next_distance = curr_distance;
6406 for (i = 0; i < nr_node_ids; i++) {
6407 for (j = 0; j < nr_node_ids; j++) {
6408 for (k = 0; k < nr_node_ids; k++) {
6409 int distance = node_distance(i, k);
6411 if (distance > curr_distance &&
6412 (distance < next_distance ||
6413 next_distance == curr_distance))
6414 next_distance = distance;
6417 * While not a strong assumption it would be nice to know
6418 * about cases where if node A is connected to B, B is not
6419 * equally connected to A.
6421 if (sched_debug() && node_distance(k, i) != distance)
6422 sched_numa_warn("Node-distance not symmetric");
6424 if (sched_debug() && i && !find_numa_distance(distance))
6425 sched_numa_warn("Node-0 not representative");
6427 if (next_distance != curr_distance) {
6428 sched_domains_numa_distance[level++] = next_distance;
6429 sched_domains_numa_levels = level;
6430 curr_distance = next_distance;
6435 * In case of sched_debug() we verify the above assumption.
6445 * 'level' contains the number of unique distances, excluding the
6446 * identity distance node_distance(i,i).
6448 * The sched_domains_numa_distance[] array includes the actual distance
6453 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6454 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6455 * the array will contain less then 'level' members. This could be
6456 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6457 * in other functions.
6459 * We reset it to 'level' at the end of this function.
6461 sched_domains_numa_levels = 0;
6463 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6464 if (!sched_domains_numa_masks)
6468 * Now for each level, construct a mask per node which contains all
6469 * cpus of nodes that are that many hops away from us.
6471 for (i = 0; i < level; i++) {
6472 sched_domains_numa_masks[i] =
6473 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6474 if (!sched_domains_numa_masks[i])
6477 for (j = 0; j < nr_node_ids; j++) {
6478 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6482 sched_domains_numa_masks[i][j] = mask;
6484 for (k = 0; k < nr_node_ids; k++) {
6485 if (node_distance(j, k) > sched_domains_numa_distance[i])
6488 cpumask_or(mask, mask, cpumask_of_node(k));
6493 /* Compute default topology size */
6494 for (i = 0; sched_domain_topology[i].mask; i++);
6496 tl = kzalloc((i + level + 1) *
6497 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6502 * Copy the default topology bits..
6504 for (i = 0; sched_domain_topology[i].mask; i++)
6505 tl[i] = sched_domain_topology[i];
6508 * .. and append 'j' levels of NUMA goodness.
6510 for (j = 0; j < level; i++, j++) {
6511 tl[i] = (struct sched_domain_topology_level){
6512 .mask = sd_numa_mask,
6513 .sd_flags = cpu_numa_flags,
6514 .flags = SDTL_OVERLAP,
6520 sched_domain_topology = tl;
6522 sched_domains_numa_levels = level;
6523 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6525 init_numa_topology_type();
6528 static void sched_domains_numa_masks_set(int cpu)
6531 int node = cpu_to_node(cpu);
6533 for (i = 0; i < sched_domains_numa_levels; i++) {
6534 for (j = 0; j < nr_node_ids; j++) {
6535 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6536 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6541 static void sched_domains_numa_masks_clear(int cpu)
6544 for (i = 0; i < sched_domains_numa_levels; i++) {
6545 for (j = 0; j < nr_node_ids; j++)
6546 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6551 * Update sched_domains_numa_masks[level][node] array when new cpus
6554 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6555 unsigned long action,
6558 int cpu = (long)hcpu;
6560 switch (action & ~CPU_TASKS_FROZEN) {
6562 sched_domains_numa_masks_set(cpu);
6566 sched_domains_numa_masks_clear(cpu);
6576 static inline void sched_init_numa(void)
6580 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6581 unsigned long action,
6586 #endif /* CONFIG_NUMA */
6588 static int __sdt_alloc(const struct cpumask *cpu_map)
6590 struct sched_domain_topology_level *tl;
6593 for_each_sd_topology(tl) {
6594 struct sd_data *sdd = &tl->data;
6596 sdd->sd = alloc_percpu(struct sched_domain *);
6600 sdd->sg = alloc_percpu(struct sched_group *);
6604 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6608 for_each_cpu(j, cpu_map) {
6609 struct sched_domain *sd;
6610 struct sched_group *sg;
6611 struct sched_group_capacity *sgc;
6613 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6614 GFP_KERNEL, cpu_to_node(j));
6618 *per_cpu_ptr(sdd->sd, j) = sd;
6620 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6621 GFP_KERNEL, cpu_to_node(j));
6627 *per_cpu_ptr(sdd->sg, j) = sg;
6629 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6630 GFP_KERNEL, cpu_to_node(j));
6634 *per_cpu_ptr(sdd->sgc, j) = sgc;
6641 static void __sdt_free(const struct cpumask *cpu_map)
6643 struct sched_domain_topology_level *tl;
6646 for_each_sd_topology(tl) {
6647 struct sd_data *sdd = &tl->data;
6649 for_each_cpu(j, cpu_map) {
6650 struct sched_domain *sd;
6653 sd = *per_cpu_ptr(sdd->sd, j);
6654 if (sd && (sd->flags & SD_OVERLAP))
6655 free_sched_groups(sd->groups, 0);
6656 kfree(*per_cpu_ptr(sdd->sd, j));
6660 kfree(*per_cpu_ptr(sdd->sg, j));
6662 kfree(*per_cpu_ptr(sdd->sgc, j));
6664 free_percpu(sdd->sd);
6666 free_percpu(sdd->sg);
6668 free_percpu(sdd->sgc);
6673 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6674 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6675 struct sched_domain *child, int cpu)
6677 struct sched_domain *sd = sd_init(tl, cpu);
6681 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6683 sd->level = child->level + 1;
6684 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6688 if (!cpumask_subset(sched_domain_span(child),
6689 sched_domain_span(sd))) {
6690 pr_err("BUG: arch topology borken\n");
6691 #ifdef CONFIG_SCHED_DEBUG
6692 pr_err(" the %s domain not a subset of the %s domain\n",
6693 child->name, sd->name);
6695 /* Fixup, ensure @sd has at least @child cpus. */
6696 cpumask_or(sched_domain_span(sd),
6697 sched_domain_span(sd),
6698 sched_domain_span(child));
6702 set_domain_attribute(sd, attr);
6708 * Build sched domains for a given set of cpus and attach the sched domains
6709 * to the individual cpus
6711 static int build_sched_domains(const struct cpumask *cpu_map,
6712 struct sched_domain_attr *attr)
6714 enum s_alloc alloc_state;
6715 struct sched_domain *sd;
6717 int i, ret = -ENOMEM;
6719 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6720 if (alloc_state != sa_rootdomain)
6723 /* Set up domains for cpus specified by the cpu_map. */
6724 for_each_cpu(i, cpu_map) {
6725 struct sched_domain_topology_level *tl;
6728 for_each_sd_topology(tl) {
6729 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6730 if (tl == sched_domain_topology)
6731 *per_cpu_ptr(d.sd, i) = sd;
6732 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6733 sd->flags |= SD_OVERLAP;
6734 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6739 /* Build the groups for the domains */
6740 for_each_cpu(i, cpu_map) {
6741 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6742 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6743 if (sd->flags & SD_OVERLAP) {
6744 if (build_overlap_sched_groups(sd, i))
6747 if (build_sched_groups(sd, i))
6753 /* Calculate CPU capacity for physical packages and nodes */
6754 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6755 if (!cpumask_test_cpu(i, cpu_map))
6758 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6759 claim_allocations(i, sd);
6760 init_sched_groups_capacity(i, sd);
6764 /* Attach the domains */
6766 for_each_cpu(i, cpu_map) {
6767 sd = *per_cpu_ptr(d.sd, i);
6768 cpu_attach_domain(sd, d.rd, i);
6774 __free_domain_allocs(&d, alloc_state, cpu_map);
6778 static cpumask_var_t *doms_cur; /* current sched domains */
6779 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6780 static struct sched_domain_attr *dattr_cur;
6781 /* attribues of custom domains in 'doms_cur' */
6784 * Special case: If a kmalloc of a doms_cur partition (array of
6785 * cpumask) fails, then fallback to a single sched domain,
6786 * as determined by the single cpumask fallback_doms.
6788 static cpumask_var_t fallback_doms;
6791 * arch_update_cpu_topology lets virtualized architectures update the
6792 * cpu core maps. It is supposed to return 1 if the topology changed
6793 * or 0 if it stayed the same.
6795 int __weak arch_update_cpu_topology(void)
6800 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6803 cpumask_var_t *doms;
6805 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6808 for (i = 0; i < ndoms; i++) {
6809 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6810 free_sched_domains(doms, i);
6817 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6820 for (i = 0; i < ndoms; i++)
6821 free_cpumask_var(doms[i]);
6826 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6827 * For now this just excludes isolated cpus, but could be used to
6828 * exclude other special cases in the future.
6830 static int init_sched_domains(const struct cpumask *cpu_map)
6834 arch_update_cpu_topology();
6836 doms_cur = alloc_sched_domains(ndoms_cur);
6838 doms_cur = &fallback_doms;
6839 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6840 err = build_sched_domains(doms_cur[0], NULL);
6841 register_sched_domain_sysctl();
6847 * Detach sched domains from a group of cpus specified in cpu_map
6848 * These cpus will now be attached to the NULL domain
6850 static void detach_destroy_domains(const struct cpumask *cpu_map)
6855 for_each_cpu(i, cpu_map)
6856 cpu_attach_domain(NULL, &def_root_domain, i);
6860 /* handle null as "default" */
6861 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6862 struct sched_domain_attr *new, int idx_new)
6864 struct sched_domain_attr tmp;
6871 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6872 new ? (new + idx_new) : &tmp,
6873 sizeof(struct sched_domain_attr));
6877 * Partition sched domains as specified by the 'ndoms_new'
6878 * cpumasks in the array doms_new[] of cpumasks. This compares
6879 * doms_new[] to the current sched domain partitioning, doms_cur[].
6880 * It destroys each deleted domain and builds each new domain.
6882 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6883 * The masks don't intersect (don't overlap.) We should setup one
6884 * sched domain for each mask. CPUs not in any of the cpumasks will
6885 * not be load balanced. If the same cpumask appears both in the
6886 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6889 * The passed in 'doms_new' should be allocated using
6890 * alloc_sched_domains. This routine takes ownership of it and will
6891 * free_sched_domains it when done with it. If the caller failed the
6892 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6893 * and partition_sched_domains() will fallback to the single partition
6894 * 'fallback_doms', it also forces the domains to be rebuilt.
6896 * If doms_new == NULL it will be replaced with cpu_online_mask.
6897 * ndoms_new == 0 is a special case for destroying existing domains,
6898 * and it will not create the default domain.
6900 * Call with hotplug lock held
6902 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6903 struct sched_domain_attr *dattr_new)
6908 mutex_lock(&sched_domains_mutex);
6910 /* always unregister in case we don't destroy any domains */
6911 unregister_sched_domain_sysctl();
6913 /* Let architecture update cpu core mappings. */
6914 new_topology = arch_update_cpu_topology();
6916 n = doms_new ? ndoms_new : 0;
6918 /* Destroy deleted domains */
6919 for (i = 0; i < ndoms_cur; i++) {
6920 for (j = 0; j < n && !new_topology; j++) {
6921 if (cpumask_equal(doms_cur[i], doms_new[j])
6922 && dattrs_equal(dattr_cur, i, dattr_new, j))
6925 /* no match - a current sched domain not in new doms_new[] */
6926 detach_destroy_domains(doms_cur[i]);
6932 if (doms_new == NULL) {
6934 doms_new = &fallback_doms;
6935 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6936 WARN_ON_ONCE(dattr_new);
6939 /* Build new domains */
6940 for (i = 0; i < ndoms_new; i++) {
6941 for (j = 0; j < n && !new_topology; j++) {
6942 if (cpumask_equal(doms_new[i], doms_cur[j])
6943 && dattrs_equal(dattr_new, i, dattr_cur, j))
6946 /* no match - add a new doms_new */
6947 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6952 /* Remember the new sched domains */
6953 if (doms_cur != &fallback_doms)
6954 free_sched_domains(doms_cur, ndoms_cur);
6955 kfree(dattr_cur); /* kfree(NULL) is safe */
6956 doms_cur = doms_new;
6957 dattr_cur = dattr_new;
6958 ndoms_cur = ndoms_new;
6960 register_sched_domain_sysctl();
6962 mutex_unlock(&sched_domains_mutex);
6965 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6968 * Update cpusets according to cpu_active mask. If cpusets are
6969 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6970 * around partition_sched_domains().
6972 * If we come here as part of a suspend/resume, don't touch cpusets because we
6973 * want to restore it back to its original state upon resume anyway.
6975 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6979 case CPU_ONLINE_FROZEN:
6980 case CPU_DOWN_FAILED_FROZEN:
6983 * num_cpus_frozen tracks how many CPUs are involved in suspend
6984 * resume sequence. As long as this is not the last online
6985 * operation in the resume sequence, just build a single sched
6986 * domain, ignoring cpusets.
6989 if (likely(num_cpus_frozen)) {
6990 partition_sched_domains(1, NULL, NULL);
6995 * This is the last CPU online operation. So fall through and
6996 * restore the original sched domains by considering the
6997 * cpuset configurations.
7001 case CPU_DOWN_FAILED:
7002 cpuset_update_active_cpus(true);
7010 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7014 case CPU_DOWN_PREPARE:
7015 cpuset_update_active_cpus(false);
7017 case CPU_DOWN_PREPARE_FROZEN:
7019 partition_sched_domains(1, NULL, NULL);
7027 void __init sched_init_smp(void)
7029 cpumask_var_t non_isolated_cpus;
7031 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7032 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7037 * There's no userspace yet to cause hotplug operations; hence all the
7038 * cpu masks are stable and all blatant races in the below code cannot
7041 mutex_lock(&sched_domains_mutex);
7042 init_sched_domains(cpu_active_mask);
7043 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7044 if (cpumask_empty(non_isolated_cpus))
7045 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7046 mutex_unlock(&sched_domains_mutex);
7048 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7049 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7050 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7054 /* Move init over to a non-isolated CPU */
7055 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7057 sched_init_granularity();
7058 free_cpumask_var(non_isolated_cpus);
7060 init_sched_rt_class();
7061 init_sched_dl_class();
7064 void __init sched_init_smp(void)
7066 sched_init_granularity();
7068 #endif /* CONFIG_SMP */
7070 const_debug unsigned int sysctl_timer_migration = 1;
7072 int in_sched_functions(unsigned long addr)
7074 return in_lock_functions(addr) ||
7075 (addr >= (unsigned long)__sched_text_start
7076 && addr < (unsigned long)__sched_text_end);
7079 #ifdef CONFIG_CGROUP_SCHED
7081 * Default task group.
7082 * Every task in system belongs to this group at bootup.
7084 struct task_group root_task_group;
7085 LIST_HEAD(task_groups);
7088 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7090 void __init sched_init(void)
7093 unsigned long alloc_size = 0, ptr;
7095 #ifdef CONFIG_FAIR_GROUP_SCHED
7096 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7098 #ifdef CONFIG_RT_GROUP_SCHED
7099 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7102 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7104 #ifdef CONFIG_FAIR_GROUP_SCHED
7105 root_task_group.se = (struct sched_entity **)ptr;
7106 ptr += nr_cpu_ids * sizeof(void **);
7108 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7109 ptr += nr_cpu_ids * sizeof(void **);
7111 #endif /* CONFIG_FAIR_GROUP_SCHED */
7112 #ifdef CONFIG_RT_GROUP_SCHED
7113 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7114 ptr += nr_cpu_ids * sizeof(void **);
7116 root_task_group.rt_rq = (struct rt_rq **)ptr;
7117 ptr += nr_cpu_ids * sizeof(void **);
7119 #endif /* CONFIG_RT_GROUP_SCHED */
7121 #ifdef CONFIG_CPUMASK_OFFSTACK
7122 for_each_possible_cpu(i) {
7123 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7124 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7126 #endif /* CONFIG_CPUMASK_OFFSTACK */
7128 init_rt_bandwidth(&def_rt_bandwidth,
7129 global_rt_period(), global_rt_runtime());
7130 init_dl_bandwidth(&def_dl_bandwidth,
7131 global_rt_period(), global_rt_runtime());
7134 init_defrootdomain();
7137 #ifdef CONFIG_RT_GROUP_SCHED
7138 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7139 global_rt_period(), global_rt_runtime());
7140 #endif /* CONFIG_RT_GROUP_SCHED */
7142 #ifdef CONFIG_CGROUP_SCHED
7143 list_add(&root_task_group.list, &task_groups);
7144 INIT_LIST_HEAD(&root_task_group.children);
7145 INIT_LIST_HEAD(&root_task_group.siblings);
7146 autogroup_init(&init_task);
7148 #endif /* CONFIG_CGROUP_SCHED */
7150 for_each_possible_cpu(i) {
7154 raw_spin_lock_init(&rq->lock);
7156 rq->calc_load_active = 0;
7157 rq->calc_load_update = jiffies + LOAD_FREQ;
7158 init_cfs_rq(&rq->cfs);
7159 init_rt_rq(&rq->rt, rq);
7160 init_dl_rq(&rq->dl, rq);
7161 #ifdef CONFIG_FAIR_GROUP_SCHED
7162 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7163 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7165 * How much cpu bandwidth does root_task_group get?
7167 * In case of task-groups formed thr' the cgroup filesystem, it
7168 * gets 100% of the cpu resources in the system. This overall
7169 * system cpu resource is divided among the tasks of
7170 * root_task_group and its child task-groups in a fair manner,
7171 * based on each entity's (task or task-group's) weight
7172 * (se->load.weight).
7174 * In other words, if root_task_group has 10 tasks of weight
7175 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7176 * then A0's share of the cpu resource is:
7178 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7180 * We achieve this by letting root_task_group's tasks sit
7181 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7183 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7184 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7185 #endif /* CONFIG_FAIR_GROUP_SCHED */
7187 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7188 #ifdef CONFIG_RT_GROUP_SCHED
7189 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7192 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7193 rq->cpu_load[j] = 0;
7195 rq->last_load_update_tick = jiffies;
7200 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7201 rq->post_schedule = 0;
7202 rq->active_balance = 0;
7203 rq->next_balance = jiffies;
7208 rq->avg_idle = 2*sysctl_sched_migration_cost;
7209 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7211 INIT_LIST_HEAD(&rq->cfs_tasks);
7213 rq_attach_root(rq, &def_root_domain);
7214 #ifdef CONFIG_NO_HZ_COMMON
7217 #ifdef CONFIG_NO_HZ_FULL
7218 rq->last_sched_tick = 0;
7222 atomic_set(&rq->nr_iowait, 0);
7225 set_load_weight(&init_task);
7227 #ifdef CONFIG_PREEMPT_NOTIFIERS
7228 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7232 * The boot idle thread does lazy MMU switching as well:
7234 atomic_inc(&init_mm.mm_count);
7235 enter_lazy_tlb(&init_mm, current);
7238 * During early bootup we pretend to be a normal task:
7240 current->sched_class = &fair_sched_class;
7243 * Make us the idle thread. Technically, schedule() should not be
7244 * called from this thread, however somewhere below it might be,
7245 * but because we are the idle thread, we just pick up running again
7246 * when this runqueue becomes "idle".
7248 init_idle(current, smp_processor_id());
7250 calc_load_update = jiffies + LOAD_FREQ;
7253 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7254 /* May be allocated at isolcpus cmdline parse time */
7255 if (cpu_isolated_map == NULL)
7256 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7257 idle_thread_set_boot_cpu();
7258 set_cpu_rq_start_time();
7260 init_sched_fair_class();
7262 scheduler_running = 1;
7265 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7266 static inline int preempt_count_equals(int preempt_offset)
7268 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7270 return (nested == preempt_offset);
7273 void __might_sleep(const char *file, int line, int preempt_offset)
7276 * Blocking primitives will set (and therefore destroy) current->state,
7277 * since we will exit with TASK_RUNNING make sure we enter with it,
7278 * otherwise we will destroy state.
7280 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7281 "do not call blocking ops when !TASK_RUNNING; "
7282 "state=%lx set at [<%p>] %pS\n",
7284 (void *)current->task_state_change,
7285 (void *)current->task_state_change);
7287 ___might_sleep(file, line, preempt_offset);
7289 EXPORT_SYMBOL(__might_sleep);
7291 void ___might_sleep(const char *file, int line, int preempt_offset)
7293 static unsigned long prev_jiffy; /* ratelimiting */
7295 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7296 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7297 !is_idle_task(current)) ||
7298 system_state != SYSTEM_RUNNING || oops_in_progress)
7300 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7302 prev_jiffy = jiffies;
7305 "BUG: sleeping function called from invalid context at %s:%d\n",
7308 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7309 in_atomic(), irqs_disabled(),
7310 current->pid, current->comm);
7312 if (task_stack_end_corrupted(current))
7313 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7315 debug_show_held_locks(current);
7316 if (irqs_disabled())
7317 print_irqtrace_events(current);
7318 #ifdef CONFIG_DEBUG_PREEMPT
7319 if (!preempt_count_equals(preempt_offset)) {
7320 pr_err("Preemption disabled at:");
7321 print_ip_sym(current->preempt_disable_ip);
7327 EXPORT_SYMBOL(___might_sleep);
7330 #ifdef CONFIG_MAGIC_SYSRQ
7331 static void normalize_task(struct rq *rq, struct task_struct *p)
7333 const struct sched_class *prev_class = p->sched_class;
7334 struct sched_attr attr = {
7335 .sched_policy = SCHED_NORMAL,
7337 int old_prio = p->prio;
7340 queued = task_on_rq_queued(p);
7342 dequeue_task(rq, p, 0);
7343 __setscheduler(rq, p, &attr);
7345 enqueue_task(rq, p, 0);
7349 check_class_changed(rq, p, prev_class, old_prio);
7352 void normalize_rt_tasks(void)
7354 struct task_struct *g, *p;
7355 unsigned long flags;
7358 read_lock(&tasklist_lock);
7359 for_each_process_thread(g, p) {
7361 * Only normalize user tasks:
7363 if (p->flags & PF_KTHREAD)
7366 p->se.exec_start = 0;
7367 #ifdef CONFIG_SCHEDSTATS
7368 p->se.statistics.wait_start = 0;
7369 p->se.statistics.sleep_start = 0;
7370 p->se.statistics.block_start = 0;
7373 if (!dl_task(p) && !rt_task(p)) {
7375 * Renice negative nice level userspace
7378 if (task_nice(p) < 0)
7379 set_user_nice(p, 0);
7383 rq = task_rq_lock(p, &flags);
7384 normalize_task(rq, p);
7385 task_rq_unlock(rq, p, &flags);
7387 read_unlock(&tasklist_lock);
7390 #endif /* CONFIG_MAGIC_SYSRQ */
7392 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7394 * These functions are only useful for the IA64 MCA handling, or kdb.
7396 * They can only be called when the whole system has been
7397 * stopped - every CPU needs to be quiescent, and no scheduling
7398 * activity can take place. Using them for anything else would
7399 * be a serious bug, and as a result, they aren't even visible
7400 * under any other configuration.
7404 * curr_task - return the current task for a given cpu.
7405 * @cpu: the processor in question.
7407 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7409 * Return: The current task for @cpu.
7411 struct task_struct *curr_task(int cpu)
7413 return cpu_curr(cpu);
7416 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7420 * set_curr_task - set the current task for a given cpu.
7421 * @cpu: the processor in question.
7422 * @p: the task pointer to set.
7424 * Description: This function must only be used when non-maskable interrupts
7425 * are serviced on a separate stack. It allows the architecture to switch the
7426 * notion of the current task on a cpu in a non-blocking manner. This function
7427 * must be called with all CPU's synchronized, and interrupts disabled, the
7428 * and caller must save the original value of the current task (see
7429 * curr_task() above) and restore that value before reenabling interrupts and
7430 * re-starting the system.
7432 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7434 void set_curr_task(int cpu, struct task_struct *p)
7441 #ifdef CONFIG_CGROUP_SCHED
7442 /* task_group_lock serializes the addition/removal of task groups */
7443 static DEFINE_SPINLOCK(task_group_lock);
7445 static void free_sched_group(struct task_group *tg)
7447 free_fair_sched_group(tg);
7448 free_rt_sched_group(tg);
7453 /* allocate runqueue etc for a new task group */
7454 struct task_group *sched_create_group(struct task_group *parent)
7456 struct task_group *tg;
7458 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7460 return ERR_PTR(-ENOMEM);
7462 if (!alloc_fair_sched_group(tg, parent))
7465 if (!alloc_rt_sched_group(tg, parent))
7471 free_sched_group(tg);
7472 return ERR_PTR(-ENOMEM);
7475 void sched_online_group(struct task_group *tg, struct task_group *parent)
7477 unsigned long flags;
7479 spin_lock_irqsave(&task_group_lock, flags);
7480 list_add_rcu(&tg->list, &task_groups);
7482 WARN_ON(!parent); /* root should already exist */
7484 tg->parent = parent;
7485 INIT_LIST_HEAD(&tg->children);
7486 list_add_rcu(&tg->siblings, &parent->children);
7487 spin_unlock_irqrestore(&task_group_lock, flags);
7490 /* rcu callback to free various structures associated with a task group */
7491 static void free_sched_group_rcu(struct rcu_head *rhp)
7493 /* now it should be safe to free those cfs_rqs */
7494 free_sched_group(container_of(rhp, struct task_group, rcu));
7497 /* Destroy runqueue etc associated with a task group */
7498 void sched_destroy_group(struct task_group *tg)
7500 /* wait for possible concurrent references to cfs_rqs complete */
7501 call_rcu(&tg->rcu, free_sched_group_rcu);
7504 void sched_offline_group(struct task_group *tg)
7506 unsigned long flags;
7509 /* end participation in shares distribution */
7510 for_each_possible_cpu(i)
7511 unregister_fair_sched_group(tg, i);
7513 spin_lock_irqsave(&task_group_lock, flags);
7514 list_del_rcu(&tg->list);
7515 list_del_rcu(&tg->siblings);
7516 spin_unlock_irqrestore(&task_group_lock, flags);
7519 /* change task's runqueue when it moves between groups.
7520 * The caller of this function should have put the task in its new group
7521 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7522 * reflect its new group.
7524 void sched_move_task(struct task_struct *tsk)
7526 struct task_group *tg;
7527 int queued, running;
7528 unsigned long flags;
7531 rq = task_rq_lock(tsk, &flags);
7533 running = task_current(rq, tsk);
7534 queued = task_on_rq_queued(tsk);
7537 dequeue_task(rq, tsk, 0);
7538 if (unlikely(running))
7539 put_prev_task(rq, tsk);
7542 * All callers are synchronized by task_rq_lock(); we do not use RCU
7543 * which is pointless here. Thus, we pass "true" to task_css_check()
7544 * to prevent lockdep warnings.
7546 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7547 struct task_group, css);
7548 tg = autogroup_task_group(tsk, tg);
7549 tsk->sched_task_group = tg;
7551 #ifdef CONFIG_FAIR_GROUP_SCHED
7552 if (tsk->sched_class->task_move_group)
7553 tsk->sched_class->task_move_group(tsk, queued);
7556 set_task_rq(tsk, task_cpu(tsk));
7558 if (unlikely(running))
7559 tsk->sched_class->set_curr_task(rq);
7561 enqueue_task(rq, tsk, 0);
7563 task_rq_unlock(rq, tsk, &flags);
7565 #endif /* CONFIG_CGROUP_SCHED */
7567 #ifdef CONFIG_RT_GROUP_SCHED
7569 * Ensure that the real time constraints are schedulable.
7571 static DEFINE_MUTEX(rt_constraints_mutex);
7573 /* Must be called with tasklist_lock held */
7574 static inline int tg_has_rt_tasks(struct task_group *tg)
7576 struct task_struct *g, *p;
7579 * Autogroups do not have RT tasks; see autogroup_create().
7581 if (task_group_is_autogroup(tg))
7584 for_each_process_thread(g, p) {
7585 if (rt_task(p) && task_group(p) == tg)
7592 struct rt_schedulable_data {
7593 struct task_group *tg;
7598 static int tg_rt_schedulable(struct task_group *tg, void *data)
7600 struct rt_schedulable_data *d = data;
7601 struct task_group *child;
7602 unsigned long total, sum = 0;
7603 u64 period, runtime;
7605 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7606 runtime = tg->rt_bandwidth.rt_runtime;
7609 period = d->rt_period;
7610 runtime = d->rt_runtime;
7614 * Cannot have more runtime than the period.
7616 if (runtime > period && runtime != RUNTIME_INF)
7620 * Ensure we don't starve existing RT tasks.
7622 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7625 total = to_ratio(period, runtime);
7628 * Nobody can have more than the global setting allows.
7630 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7634 * The sum of our children's runtime should not exceed our own.
7636 list_for_each_entry_rcu(child, &tg->children, siblings) {
7637 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7638 runtime = child->rt_bandwidth.rt_runtime;
7640 if (child == d->tg) {
7641 period = d->rt_period;
7642 runtime = d->rt_runtime;
7645 sum += to_ratio(period, runtime);
7654 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7658 struct rt_schedulable_data data = {
7660 .rt_period = period,
7661 .rt_runtime = runtime,
7665 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7671 static int tg_set_rt_bandwidth(struct task_group *tg,
7672 u64 rt_period, u64 rt_runtime)
7677 * Disallowing the root group RT runtime is BAD, it would disallow the
7678 * kernel creating (and or operating) RT threads.
7680 if (tg == &root_task_group && rt_runtime == 0)
7683 /* No period doesn't make any sense. */
7687 mutex_lock(&rt_constraints_mutex);
7688 read_lock(&tasklist_lock);
7689 err = __rt_schedulable(tg, rt_period, rt_runtime);
7693 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7694 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7695 tg->rt_bandwidth.rt_runtime = rt_runtime;
7697 for_each_possible_cpu(i) {
7698 struct rt_rq *rt_rq = tg->rt_rq[i];
7700 raw_spin_lock(&rt_rq->rt_runtime_lock);
7701 rt_rq->rt_runtime = rt_runtime;
7702 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7704 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7706 read_unlock(&tasklist_lock);
7707 mutex_unlock(&rt_constraints_mutex);
7712 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7714 u64 rt_runtime, rt_period;
7716 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7717 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7718 if (rt_runtime_us < 0)
7719 rt_runtime = RUNTIME_INF;
7721 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7724 static long sched_group_rt_runtime(struct task_group *tg)
7728 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7731 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7732 do_div(rt_runtime_us, NSEC_PER_USEC);
7733 return rt_runtime_us;
7736 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7738 u64 rt_runtime, rt_period;
7740 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7741 rt_runtime = tg->rt_bandwidth.rt_runtime;
7743 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7746 static long sched_group_rt_period(struct task_group *tg)
7750 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7751 do_div(rt_period_us, NSEC_PER_USEC);
7752 return rt_period_us;
7754 #endif /* CONFIG_RT_GROUP_SCHED */
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 static int sched_rt_global_constraints(void)
7761 mutex_lock(&rt_constraints_mutex);
7762 read_lock(&tasklist_lock);
7763 ret = __rt_schedulable(NULL, 0, 0);
7764 read_unlock(&tasklist_lock);
7765 mutex_unlock(&rt_constraints_mutex);
7770 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7772 /* Don't accept realtime tasks when there is no way for them to run */
7773 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7779 #else /* !CONFIG_RT_GROUP_SCHED */
7780 static int sched_rt_global_constraints(void)
7782 unsigned long flags;
7785 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7786 for_each_possible_cpu(i) {
7787 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7789 raw_spin_lock(&rt_rq->rt_runtime_lock);
7790 rt_rq->rt_runtime = global_rt_runtime();
7791 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7793 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7797 #endif /* CONFIG_RT_GROUP_SCHED */
7799 static int sched_dl_global_constraints(void)
7801 u64 runtime = global_rt_runtime();
7802 u64 period = global_rt_period();
7803 u64 new_bw = to_ratio(period, runtime);
7806 unsigned long flags;
7809 * Here we want to check the bandwidth not being set to some
7810 * value smaller than the currently allocated bandwidth in
7811 * any of the root_domains.
7813 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7814 * cycling on root_domains... Discussion on different/better
7815 * solutions is welcome!
7817 for_each_possible_cpu(cpu) {
7818 rcu_read_lock_sched();
7819 dl_b = dl_bw_of(cpu);
7821 raw_spin_lock_irqsave(&dl_b->lock, flags);
7822 if (new_bw < dl_b->total_bw)
7824 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7826 rcu_read_unlock_sched();
7835 static void sched_dl_do_global(void)
7840 unsigned long flags;
7842 def_dl_bandwidth.dl_period = global_rt_period();
7843 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7845 if (global_rt_runtime() != RUNTIME_INF)
7846 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7849 * FIXME: As above...
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);
7857 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7859 rcu_read_unlock_sched();
7863 static int sched_rt_global_validate(void)
7865 if (sysctl_sched_rt_period <= 0)
7868 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7869 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7875 static void sched_rt_do_global(void)
7877 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7878 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7881 int sched_rt_handler(struct ctl_table *table, int write,
7882 void __user *buffer, size_t *lenp,
7885 int old_period, old_runtime;
7886 static DEFINE_MUTEX(mutex);
7890 old_period = sysctl_sched_rt_period;
7891 old_runtime = sysctl_sched_rt_runtime;
7893 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7895 if (!ret && write) {
7896 ret = sched_rt_global_validate();
7900 ret = sched_rt_global_constraints();
7904 ret = sched_dl_global_constraints();
7908 sched_rt_do_global();
7909 sched_dl_do_global();
7913 sysctl_sched_rt_period = old_period;
7914 sysctl_sched_rt_runtime = old_runtime;
7916 mutex_unlock(&mutex);
7921 int sched_rr_handler(struct ctl_table *table, int write,
7922 void __user *buffer, size_t *lenp,
7926 static DEFINE_MUTEX(mutex);
7929 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7930 /* make sure that internally we keep jiffies */
7931 /* also, writing zero resets timeslice to default */
7932 if (!ret && write) {
7933 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7934 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7936 mutex_unlock(&mutex);
7940 #ifdef CONFIG_CGROUP_SCHED
7942 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7944 return css ? container_of(css, struct task_group, css) : NULL;
7947 static struct cgroup_subsys_state *
7948 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7950 struct task_group *parent = css_tg(parent_css);
7951 struct task_group *tg;
7954 /* This is early initialization for the top cgroup */
7955 return &root_task_group.css;
7958 tg = sched_create_group(parent);
7960 return ERR_PTR(-ENOMEM);
7965 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7967 struct task_group *tg = css_tg(css);
7968 struct task_group *parent = css_tg(css->parent);
7971 sched_online_group(tg, parent);
7975 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7977 struct task_group *tg = css_tg(css);
7979 sched_destroy_group(tg);
7982 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7984 struct task_group *tg = css_tg(css);
7986 sched_offline_group(tg);
7989 static void cpu_cgroup_fork(struct task_struct *task)
7991 sched_move_task(task);
7994 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7995 struct cgroup_taskset *tset)
7997 struct task_struct *task;
7999 cgroup_taskset_for_each(task, tset) {
8000 #ifdef CONFIG_RT_GROUP_SCHED
8001 if (!sched_rt_can_attach(css_tg(css), task))
8004 /* We don't support RT-tasks being in separate groups */
8005 if (task->sched_class != &fair_sched_class)
8012 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8013 struct cgroup_taskset *tset)
8015 struct task_struct *task;
8017 cgroup_taskset_for_each(task, tset)
8018 sched_move_task(task);
8021 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8022 struct cgroup_subsys_state *old_css,
8023 struct task_struct *task)
8026 * cgroup_exit() is called in the copy_process() failure path.
8027 * Ignore this case since the task hasn't ran yet, this avoids
8028 * trying to poke a half freed task state from generic code.
8030 if (!(task->flags & PF_EXITING))
8033 sched_move_task(task);
8036 #ifdef CONFIG_FAIR_GROUP_SCHED
8037 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8038 struct cftype *cftype, u64 shareval)
8040 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8043 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8046 struct task_group *tg = css_tg(css);
8048 return (u64) scale_load_down(tg->shares);
8051 #ifdef CONFIG_CFS_BANDWIDTH
8052 static DEFINE_MUTEX(cfs_constraints_mutex);
8054 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8055 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8057 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8059 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8061 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8062 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8064 if (tg == &root_task_group)
8068 * Ensure we have at some amount of bandwidth every period. This is
8069 * to prevent reaching a state of large arrears when throttled via
8070 * entity_tick() resulting in prolonged exit starvation.
8072 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8076 * Likewise, bound things on the otherside by preventing insane quota
8077 * periods. This also allows us to normalize in computing quota
8080 if (period > max_cfs_quota_period)
8084 * Prevent race between setting of cfs_rq->runtime_enabled and
8085 * unthrottle_offline_cfs_rqs().
8088 mutex_lock(&cfs_constraints_mutex);
8089 ret = __cfs_schedulable(tg, period, quota);
8093 runtime_enabled = quota != RUNTIME_INF;
8094 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8096 * If we need to toggle cfs_bandwidth_used, off->on must occur
8097 * before making related changes, and on->off must occur afterwards
8099 if (runtime_enabled && !runtime_was_enabled)
8100 cfs_bandwidth_usage_inc();
8101 raw_spin_lock_irq(&cfs_b->lock);
8102 cfs_b->period = ns_to_ktime(period);
8103 cfs_b->quota = quota;
8105 __refill_cfs_bandwidth_runtime(cfs_b);
8106 /* restart the period timer (if active) to handle new period expiry */
8107 if (runtime_enabled && cfs_b->timer_active) {
8108 /* force a reprogram */
8109 __start_cfs_bandwidth(cfs_b, true);
8111 raw_spin_unlock_irq(&cfs_b->lock);
8113 for_each_online_cpu(i) {
8114 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8115 struct rq *rq = cfs_rq->rq;
8117 raw_spin_lock_irq(&rq->lock);
8118 cfs_rq->runtime_enabled = runtime_enabled;
8119 cfs_rq->runtime_remaining = 0;
8121 if (cfs_rq->throttled)
8122 unthrottle_cfs_rq(cfs_rq);
8123 raw_spin_unlock_irq(&rq->lock);
8125 if (runtime_was_enabled && !runtime_enabled)
8126 cfs_bandwidth_usage_dec();
8128 mutex_unlock(&cfs_constraints_mutex);
8134 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8138 period = ktime_to_ns(tg->cfs_bandwidth.period);
8139 if (cfs_quota_us < 0)
8140 quota = RUNTIME_INF;
8142 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8144 return tg_set_cfs_bandwidth(tg, period, quota);
8147 long tg_get_cfs_quota(struct task_group *tg)
8151 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8154 quota_us = tg->cfs_bandwidth.quota;
8155 do_div(quota_us, NSEC_PER_USEC);
8160 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8164 period = (u64)cfs_period_us * NSEC_PER_USEC;
8165 quota = tg->cfs_bandwidth.quota;
8167 return tg_set_cfs_bandwidth(tg, period, quota);
8170 long tg_get_cfs_period(struct task_group *tg)
8174 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8175 do_div(cfs_period_us, NSEC_PER_USEC);
8177 return cfs_period_us;
8180 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8183 return tg_get_cfs_quota(css_tg(css));
8186 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8187 struct cftype *cftype, s64 cfs_quota_us)
8189 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8192 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8195 return tg_get_cfs_period(css_tg(css));
8198 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8199 struct cftype *cftype, u64 cfs_period_us)
8201 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8204 struct cfs_schedulable_data {
8205 struct task_group *tg;
8210 * normalize group quota/period to be quota/max_period
8211 * note: units are usecs
8213 static u64 normalize_cfs_quota(struct task_group *tg,
8214 struct cfs_schedulable_data *d)
8222 period = tg_get_cfs_period(tg);
8223 quota = tg_get_cfs_quota(tg);
8226 /* note: these should typically be equivalent */
8227 if (quota == RUNTIME_INF || quota == -1)
8230 return to_ratio(period, quota);
8233 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8235 struct cfs_schedulable_data *d = data;
8236 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8237 s64 quota = 0, parent_quota = -1;
8240 quota = RUNTIME_INF;
8242 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8244 quota = normalize_cfs_quota(tg, d);
8245 parent_quota = parent_b->hierarchical_quota;
8248 * ensure max(child_quota) <= parent_quota, inherit when no
8251 if (quota == RUNTIME_INF)
8252 quota = parent_quota;
8253 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8256 cfs_b->hierarchical_quota = quota;
8261 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8264 struct cfs_schedulable_data data = {
8270 if (quota != RUNTIME_INF) {
8271 do_div(data.period, NSEC_PER_USEC);
8272 do_div(data.quota, NSEC_PER_USEC);
8276 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8282 static int cpu_stats_show(struct seq_file *sf, void *v)
8284 struct task_group *tg = css_tg(seq_css(sf));
8285 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8287 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8288 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8289 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8293 #endif /* CONFIG_CFS_BANDWIDTH */
8294 #endif /* CONFIG_FAIR_GROUP_SCHED */
8296 #ifdef CONFIG_RT_GROUP_SCHED
8297 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8298 struct cftype *cft, s64 val)
8300 return sched_group_set_rt_runtime(css_tg(css), val);
8303 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8306 return sched_group_rt_runtime(css_tg(css));
8309 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8310 struct cftype *cftype, u64 rt_period_us)
8312 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8315 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8318 return sched_group_rt_period(css_tg(css));
8320 #endif /* CONFIG_RT_GROUP_SCHED */
8322 static struct cftype cpu_files[] = {
8323 #ifdef CONFIG_FAIR_GROUP_SCHED
8326 .read_u64 = cpu_shares_read_u64,
8327 .write_u64 = cpu_shares_write_u64,
8330 #ifdef CONFIG_CFS_BANDWIDTH
8332 .name = "cfs_quota_us",
8333 .read_s64 = cpu_cfs_quota_read_s64,
8334 .write_s64 = cpu_cfs_quota_write_s64,
8337 .name = "cfs_period_us",
8338 .read_u64 = cpu_cfs_period_read_u64,
8339 .write_u64 = cpu_cfs_period_write_u64,
8343 .seq_show = cpu_stats_show,
8346 #ifdef CONFIG_RT_GROUP_SCHED
8348 .name = "rt_runtime_us",
8349 .read_s64 = cpu_rt_runtime_read,
8350 .write_s64 = cpu_rt_runtime_write,
8353 .name = "rt_period_us",
8354 .read_u64 = cpu_rt_period_read_uint,
8355 .write_u64 = cpu_rt_period_write_uint,
8361 struct cgroup_subsys cpu_cgrp_subsys = {
8362 .css_alloc = cpu_cgroup_css_alloc,
8363 .css_free = cpu_cgroup_css_free,
8364 .css_online = cpu_cgroup_css_online,
8365 .css_offline = cpu_cgroup_css_offline,
8366 .fork = cpu_cgroup_fork,
8367 .can_attach = cpu_cgroup_can_attach,
8368 .attach = cpu_cgroup_attach,
8369 .exit = cpu_cgroup_exit,
8370 .legacy_cftypes = cpu_files,
8374 #endif /* CONFIG_CGROUP_SCHED */
8376 void dump_cpu_task(int cpu)
8378 pr_info("Task dump for CPU %d:\n", cpu);
8379 sched_show_task(cpu_curr(cpu));