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);
999 static ATOMIC_NOTIFIER_HEAD(task_migration_notifier);
1001 void register_task_migration_notifier(struct notifier_block *n)
1003 atomic_notifier_chain_register(&task_migration_notifier, n);
1007 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1009 #ifdef CONFIG_SCHED_DEBUG
1011 * We should never call set_task_cpu() on a blocked task,
1012 * ttwu() will sort out the placement.
1014 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1017 #ifdef CONFIG_LOCKDEP
1019 * The caller should hold either p->pi_lock or rq->lock, when changing
1020 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1022 * sched_move_task() holds both and thus holding either pins the cgroup,
1025 * Furthermore, all task_rq users should acquire both locks, see
1028 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1029 lockdep_is_held(&task_rq(p)->lock)));
1033 trace_sched_migrate_task(p, new_cpu);
1035 if (task_cpu(p) != new_cpu) {
1036 struct task_migration_notifier tmn;
1038 if (p->sched_class->migrate_task_rq)
1039 p->sched_class->migrate_task_rq(p, new_cpu);
1040 p->se.nr_migrations++;
1041 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1044 tmn.from_cpu = task_cpu(p);
1045 tmn.to_cpu = new_cpu;
1047 atomic_notifier_call_chain(&task_migration_notifier, 0, &tmn);
1050 __set_task_cpu(p, new_cpu);
1053 static void __migrate_swap_task(struct task_struct *p, int cpu)
1055 if (task_on_rq_queued(p)) {
1056 struct rq *src_rq, *dst_rq;
1058 src_rq = task_rq(p);
1059 dst_rq = cpu_rq(cpu);
1061 deactivate_task(src_rq, p, 0);
1062 set_task_cpu(p, cpu);
1063 activate_task(dst_rq, p, 0);
1064 check_preempt_curr(dst_rq, p, 0);
1067 * Task isn't running anymore; make it appear like we migrated
1068 * it before it went to sleep. This means on wakeup we make the
1069 * previous cpu our targer instead of where it really is.
1075 struct migration_swap_arg {
1076 struct task_struct *src_task, *dst_task;
1077 int src_cpu, dst_cpu;
1080 static int migrate_swap_stop(void *data)
1082 struct migration_swap_arg *arg = data;
1083 struct rq *src_rq, *dst_rq;
1086 src_rq = cpu_rq(arg->src_cpu);
1087 dst_rq = cpu_rq(arg->dst_cpu);
1089 double_raw_lock(&arg->src_task->pi_lock,
1090 &arg->dst_task->pi_lock);
1091 double_rq_lock(src_rq, dst_rq);
1092 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1095 if (task_cpu(arg->src_task) != arg->src_cpu)
1098 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1101 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1104 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1105 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1110 double_rq_unlock(src_rq, dst_rq);
1111 raw_spin_unlock(&arg->dst_task->pi_lock);
1112 raw_spin_unlock(&arg->src_task->pi_lock);
1118 * Cross migrate two tasks
1120 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1122 struct migration_swap_arg arg;
1125 arg = (struct migration_swap_arg){
1127 .src_cpu = task_cpu(cur),
1129 .dst_cpu = task_cpu(p),
1132 if (arg.src_cpu == arg.dst_cpu)
1136 * These three tests are all lockless; this is OK since all of them
1137 * will be re-checked with proper locks held further down the line.
1139 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1142 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1145 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1148 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1149 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1155 struct migration_arg {
1156 struct task_struct *task;
1160 static int migration_cpu_stop(void *data);
1163 * wait_task_inactive - wait for a thread to unschedule.
1165 * If @match_state is nonzero, it's the @p->state value just checked and
1166 * not expected to change. If it changes, i.e. @p might have woken up,
1167 * then return zero. When we succeed in waiting for @p to be off its CPU,
1168 * we return a positive number (its total switch count). If a second call
1169 * a short while later returns the same number, the caller can be sure that
1170 * @p has remained unscheduled the whole time.
1172 * The caller must ensure that the task *will* unschedule sometime soon,
1173 * else this function might spin for a *long* time. This function can't
1174 * be called with interrupts off, or it may introduce deadlock with
1175 * smp_call_function() if an IPI is sent by the same process we are
1176 * waiting to become inactive.
1178 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1180 unsigned long flags;
1181 int running, queued;
1187 * We do the initial early heuristics without holding
1188 * any task-queue locks at all. We'll only try to get
1189 * the runqueue lock when things look like they will
1195 * If the task is actively running on another CPU
1196 * still, just relax and busy-wait without holding
1199 * NOTE! Since we don't hold any locks, it's not
1200 * even sure that "rq" stays as the right runqueue!
1201 * But we don't care, since "task_running()" will
1202 * return false if the runqueue has changed and p
1203 * is actually now running somewhere else!
1205 while (task_running(rq, p)) {
1206 if (match_state && unlikely(p->state != match_state))
1212 * Ok, time to look more closely! We need the rq
1213 * lock now, to be *sure*. If we're wrong, we'll
1214 * just go back and repeat.
1216 rq = task_rq_lock(p, &flags);
1217 trace_sched_wait_task(p);
1218 running = task_running(rq, p);
1219 queued = task_on_rq_queued(p);
1221 if (!match_state || p->state == match_state)
1222 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1223 task_rq_unlock(rq, p, &flags);
1226 * If it changed from the expected state, bail out now.
1228 if (unlikely(!ncsw))
1232 * Was it really running after all now that we
1233 * checked with the proper locks actually held?
1235 * Oops. Go back and try again..
1237 if (unlikely(running)) {
1243 * It's not enough that it's not actively running,
1244 * it must be off the runqueue _entirely_, and not
1247 * So if it was still runnable (but just not actively
1248 * running right now), it's preempted, and we should
1249 * yield - it could be a while.
1251 if (unlikely(queued)) {
1252 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1254 set_current_state(TASK_UNINTERRUPTIBLE);
1255 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1260 * Ahh, all good. It wasn't running, and it wasn't
1261 * runnable, which means that it will never become
1262 * running in the future either. We're all done!
1271 * kick_process - kick a running thread to enter/exit the kernel
1272 * @p: the to-be-kicked thread
1274 * Cause a process which is running on another CPU to enter
1275 * kernel-mode, without any delay. (to get signals handled.)
1277 * NOTE: this function doesn't have to take the runqueue lock,
1278 * because all it wants to ensure is that the remote task enters
1279 * the kernel. If the IPI races and the task has been migrated
1280 * to another CPU then no harm is done and the purpose has been
1283 void kick_process(struct task_struct *p)
1289 if ((cpu != smp_processor_id()) && task_curr(p))
1290 smp_send_reschedule(cpu);
1293 EXPORT_SYMBOL_GPL(kick_process);
1294 #endif /* CONFIG_SMP */
1298 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1300 static int select_fallback_rq(int cpu, struct task_struct *p)
1302 int nid = cpu_to_node(cpu);
1303 const struct cpumask *nodemask = NULL;
1304 enum { cpuset, possible, fail } state = cpuset;
1308 * If the node that the cpu is on has been offlined, cpu_to_node()
1309 * will return -1. There is no cpu on the node, and we should
1310 * select the cpu on the other node.
1313 nodemask = cpumask_of_node(nid);
1315 /* Look for allowed, online CPU in same node. */
1316 for_each_cpu(dest_cpu, nodemask) {
1317 if (!cpu_online(dest_cpu))
1319 if (!cpu_active(dest_cpu))
1321 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1327 /* Any allowed, online CPU? */
1328 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1329 if (!cpu_online(dest_cpu))
1331 if (!cpu_active(dest_cpu))
1338 /* No more Mr. Nice Guy. */
1339 cpuset_cpus_allowed_fallback(p);
1344 do_set_cpus_allowed(p, cpu_possible_mask);
1355 if (state != cpuset) {
1357 * Don't tell them about moving exiting tasks or
1358 * kernel threads (both mm NULL), since they never
1361 if (p->mm && printk_ratelimit()) {
1362 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1363 task_pid_nr(p), p->comm, cpu);
1371 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1374 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1376 if (p->nr_cpus_allowed > 1)
1377 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1380 * In order not to call set_task_cpu() on a blocking task we need
1381 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1384 * Since this is common to all placement strategies, this lives here.
1386 * [ this allows ->select_task() to simply return task_cpu(p) and
1387 * not worry about this generic constraint ]
1389 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1391 cpu = select_fallback_rq(task_cpu(p), p);
1396 static void update_avg(u64 *avg, u64 sample)
1398 s64 diff = sample - *avg;
1404 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1406 #ifdef CONFIG_SCHEDSTATS
1407 struct rq *rq = this_rq();
1410 int this_cpu = smp_processor_id();
1412 if (cpu == this_cpu) {
1413 schedstat_inc(rq, ttwu_local);
1414 schedstat_inc(p, se.statistics.nr_wakeups_local);
1416 struct sched_domain *sd;
1418 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1420 for_each_domain(this_cpu, sd) {
1421 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1422 schedstat_inc(sd, ttwu_wake_remote);
1429 if (wake_flags & WF_MIGRATED)
1430 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1432 #endif /* CONFIG_SMP */
1434 schedstat_inc(rq, ttwu_count);
1435 schedstat_inc(p, se.statistics.nr_wakeups);
1437 if (wake_flags & WF_SYNC)
1438 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1440 #endif /* CONFIG_SCHEDSTATS */
1443 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1445 activate_task(rq, p, en_flags);
1446 p->on_rq = TASK_ON_RQ_QUEUED;
1448 /* if a worker is waking up, notify workqueue */
1449 if (p->flags & PF_WQ_WORKER)
1450 wq_worker_waking_up(p, cpu_of(rq));
1454 * Mark the task runnable and perform wakeup-preemption.
1457 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1459 check_preempt_curr(rq, p, wake_flags);
1460 trace_sched_wakeup(p, true);
1462 p->state = TASK_RUNNING;
1464 if (p->sched_class->task_woken)
1465 p->sched_class->task_woken(rq, p);
1467 if (rq->idle_stamp) {
1468 u64 delta = rq_clock(rq) - rq->idle_stamp;
1469 u64 max = 2*rq->max_idle_balance_cost;
1471 update_avg(&rq->avg_idle, delta);
1473 if (rq->avg_idle > max)
1482 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1485 if (p->sched_contributes_to_load)
1486 rq->nr_uninterruptible--;
1489 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1490 ttwu_do_wakeup(rq, p, wake_flags);
1494 * Called in case the task @p isn't fully descheduled from its runqueue,
1495 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1496 * since all we need to do is flip p->state to TASK_RUNNING, since
1497 * the task is still ->on_rq.
1499 static int ttwu_remote(struct task_struct *p, int wake_flags)
1504 rq = __task_rq_lock(p);
1505 if (task_on_rq_queued(p)) {
1506 /* check_preempt_curr() may use rq clock */
1507 update_rq_clock(rq);
1508 ttwu_do_wakeup(rq, p, wake_flags);
1511 __task_rq_unlock(rq);
1517 void sched_ttwu_pending(void)
1519 struct rq *rq = this_rq();
1520 struct llist_node *llist = llist_del_all(&rq->wake_list);
1521 struct task_struct *p;
1522 unsigned long flags;
1527 raw_spin_lock_irqsave(&rq->lock, flags);
1530 p = llist_entry(llist, struct task_struct, wake_entry);
1531 llist = llist_next(llist);
1532 ttwu_do_activate(rq, p, 0);
1535 raw_spin_unlock_irqrestore(&rq->lock, flags);
1538 void scheduler_ipi(void)
1541 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1542 * TIF_NEED_RESCHED remotely (for the first time) will also send
1545 preempt_fold_need_resched();
1547 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1551 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1552 * traditionally all their work was done from the interrupt return
1553 * path. Now that we actually do some work, we need to make sure
1556 * Some archs already do call them, luckily irq_enter/exit nest
1559 * Arguably we should visit all archs and update all handlers,
1560 * however a fair share of IPIs are still resched only so this would
1561 * somewhat pessimize the simple resched case.
1564 sched_ttwu_pending();
1567 * Check if someone kicked us for doing the nohz idle load balance.
1569 if (unlikely(got_nohz_idle_kick())) {
1570 this_rq()->idle_balance = 1;
1571 raise_softirq_irqoff(SCHED_SOFTIRQ);
1576 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1578 struct rq *rq = cpu_rq(cpu);
1580 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1581 if (!set_nr_if_polling(rq->idle))
1582 smp_send_reschedule(cpu);
1584 trace_sched_wake_idle_without_ipi(cpu);
1588 void wake_up_if_idle(int cpu)
1590 struct rq *rq = cpu_rq(cpu);
1591 unsigned long flags;
1595 if (!is_idle_task(rcu_dereference(rq->curr)))
1598 if (set_nr_if_polling(rq->idle)) {
1599 trace_sched_wake_idle_without_ipi(cpu);
1601 raw_spin_lock_irqsave(&rq->lock, flags);
1602 if (is_idle_task(rq->curr))
1603 smp_send_reschedule(cpu);
1604 /* Else cpu is not in idle, do nothing here */
1605 raw_spin_unlock_irqrestore(&rq->lock, flags);
1612 bool cpus_share_cache(int this_cpu, int that_cpu)
1614 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1616 #endif /* CONFIG_SMP */
1618 static void ttwu_queue(struct task_struct *p, int cpu)
1620 struct rq *rq = cpu_rq(cpu);
1622 #if defined(CONFIG_SMP)
1623 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1624 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1625 ttwu_queue_remote(p, cpu);
1630 raw_spin_lock(&rq->lock);
1631 ttwu_do_activate(rq, p, 0);
1632 raw_spin_unlock(&rq->lock);
1636 * try_to_wake_up - wake up a thread
1637 * @p: the thread to be awakened
1638 * @state: the mask of task states that can be woken
1639 * @wake_flags: wake modifier flags (WF_*)
1641 * Put it on the run-queue if it's not already there. The "current"
1642 * thread is always on the run-queue (except when the actual
1643 * re-schedule is in progress), and as such you're allowed to do
1644 * the simpler "current->state = TASK_RUNNING" to mark yourself
1645 * runnable without the overhead of this.
1647 * Return: %true if @p was woken up, %false if it was already running.
1648 * or @state didn't match @p's state.
1651 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1653 unsigned long flags;
1654 int cpu, success = 0;
1657 * If we are going to wake up a thread waiting for CONDITION we
1658 * need to ensure that CONDITION=1 done by the caller can not be
1659 * reordered with p->state check below. This pairs with mb() in
1660 * set_current_state() the waiting thread does.
1662 smp_mb__before_spinlock();
1663 raw_spin_lock_irqsave(&p->pi_lock, flags);
1664 if (!(p->state & state))
1667 success = 1; /* we're going to change ->state */
1670 if (p->on_rq && ttwu_remote(p, wake_flags))
1675 * If the owning (remote) cpu is still in the middle of schedule() with
1676 * this task as prev, wait until its done referencing the task.
1681 * Pairs with the smp_wmb() in finish_lock_switch().
1685 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1686 p->state = TASK_WAKING;
1688 if (p->sched_class->task_waking)
1689 p->sched_class->task_waking(p);
1691 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1692 if (task_cpu(p) != cpu) {
1693 wake_flags |= WF_MIGRATED;
1694 set_task_cpu(p, cpu);
1696 #endif /* CONFIG_SMP */
1700 ttwu_stat(p, cpu, wake_flags);
1702 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1708 * try_to_wake_up_local - try to wake up a local task with rq lock held
1709 * @p: the thread to be awakened
1711 * Put @p on the run-queue if it's not already there. The caller must
1712 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1715 static void try_to_wake_up_local(struct task_struct *p)
1717 struct rq *rq = task_rq(p);
1719 if (WARN_ON_ONCE(rq != this_rq()) ||
1720 WARN_ON_ONCE(p == current))
1723 lockdep_assert_held(&rq->lock);
1725 if (!raw_spin_trylock(&p->pi_lock)) {
1726 raw_spin_unlock(&rq->lock);
1727 raw_spin_lock(&p->pi_lock);
1728 raw_spin_lock(&rq->lock);
1731 if (!(p->state & TASK_NORMAL))
1734 if (!task_on_rq_queued(p))
1735 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1737 ttwu_do_wakeup(rq, p, 0);
1738 ttwu_stat(p, smp_processor_id(), 0);
1740 raw_spin_unlock(&p->pi_lock);
1744 * wake_up_process - Wake up a specific process
1745 * @p: The process to be woken up.
1747 * Attempt to wake up the nominated process and move it to the set of runnable
1750 * Return: 1 if the process was woken up, 0 if it was already running.
1752 * It may be assumed that this function implies a write memory barrier before
1753 * changing the task state if and only if any tasks are woken up.
1755 int wake_up_process(struct task_struct *p)
1757 WARN_ON(task_is_stopped_or_traced(p));
1758 return try_to_wake_up(p, TASK_NORMAL, 0);
1760 EXPORT_SYMBOL(wake_up_process);
1762 int wake_up_state(struct task_struct *p, unsigned int state)
1764 return try_to_wake_up(p, state, 0);
1768 * This function clears the sched_dl_entity static params.
1770 void __dl_clear_params(struct task_struct *p)
1772 struct sched_dl_entity *dl_se = &p->dl;
1774 dl_se->dl_runtime = 0;
1775 dl_se->dl_deadline = 0;
1776 dl_se->dl_period = 0;
1780 dl_se->dl_throttled = 0;
1782 dl_se->dl_yielded = 0;
1786 * Perform scheduler related setup for a newly forked process p.
1787 * p is forked by current.
1789 * __sched_fork() is basic setup used by init_idle() too:
1791 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1796 p->se.exec_start = 0;
1797 p->se.sum_exec_runtime = 0;
1798 p->se.prev_sum_exec_runtime = 0;
1799 p->se.nr_migrations = 0;
1802 p->se.avg.decay_count = 0;
1804 INIT_LIST_HEAD(&p->se.group_node);
1806 #ifdef CONFIG_SCHEDSTATS
1807 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1810 RB_CLEAR_NODE(&p->dl.rb_node);
1811 init_dl_task_timer(&p->dl);
1812 __dl_clear_params(p);
1814 INIT_LIST_HEAD(&p->rt.run_list);
1816 #ifdef CONFIG_PREEMPT_NOTIFIERS
1817 INIT_HLIST_HEAD(&p->preempt_notifiers);
1820 #ifdef CONFIG_NUMA_BALANCING
1821 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1822 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1823 p->mm->numa_scan_seq = 0;
1826 if (clone_flags & CLONE_VM)
1827 p->numa_preferred_nid = current->numa_preferred_nid;
1829 p->numa_preferred_nid = -1;
1831 p->node_stamp = 0ULL;
1832 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1833 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1834 p->numa_work.next = &p->numa_work;
1835 p->numa_faults = NULL;
1836 p->last_task_numa_placement = 0;
1837 p->last_sum_exec_runtime = 0;
1839 p->numa_group = NULL;
1840 #endif /* CONFIG_NUMA_BALANCING */
1843 #ifdef CONFIG_NUMA_BALANCING
1844 #ifdef CONFIG_SCHED_DEBUG
1845 void set_numabalancing_state(bool enabled)
1848 sched_feat_set("NUMA");
1850 sched_feat_set("NO_NUMA");
1853 __read_mostly bool numabalancing_enabled;
1855 void set_numabalancing_state(bool enabled)
1857 numabalancing_enabled = enabled;
1859 #endif /* CONFIG_SCHED_DEBUG */
1861 #ifdef CONFIG_PROC_SYSCTL
1862 int sysctl_numa_balancing(struct ctl_table *table, int write,
1863 void __user *buffer, size_t *lenp, loff_t *ppos)
1867 int state = numabalancing_enabled;
1869 if (write && !capable(CAP_SYS_ADMIN))
1874 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1878 set_numabalancing_state(state);
1885 * fork()/clone()-time setup:
1887 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1889 unsigned long flags;
1890 int cpu = get_cpu();
1892 __sched_fork(clone_flags, p);
1894 * We mark the process as running here. This guarantees that
1895 * nobody will actually run it, and a signal or other external
1896 * event cannot wake it up and insert it on the runqueue either.
1898 p->state = TASK_RUNNING;
1901 * Make sure we do not leak PI boosting priority to the child.
1903 p->prio = current->normal_prio;
1906 * Revert to default priority/policy on fork if requested.
1908 if (unlikely(p->sched_reset_on_fork)) {
1909 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1910 p->policy = SCHED_NORMAL;
1911 p->static_prio = NICE_TO_PRIO(0);
1913 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1914 p->static_prio = NICE_TO_PRIO(0);
1916 p->prio = p->normal_prio = __normal_prio(p);
1920 * We don't need the reset flag anymore after the fork. It has
1921 * fulfilled its duty:
1923 p->sched_reset_on_fork = 0;
1926 if (dl_prio(p->prio)) {
1929 } else if (rt_prio(p->prio)) {
1930 p->sched_class = &rt_sched_class;
1932 p->sched_class = &fair_sched_class;
1935 if (p->sched_class->task_fork)
1936 p->sched_class->task_fork(p);
1939 * The child is not yet in the pid-hash so no cgroup attach races,
1940 * and the cgroup is pinned to this child due to cgroup_fork()
1941 * is ran before sched_fork().
1943 * Silence PROVE_RCU.
1945 raw_spin_lock_irqsave(&p->pi_lock, flags);
1946 set_task_cpu(p, cpu);
1947 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1949 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1950 if (likely(sched_info_on()))
1951 memset(&p->sched_info, 0, sizeof(p->sched_info));
1953 #if defined(CONFIG_SMP)
1956 init_task_preempt_count(p);
1958 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1959 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1966 unsigned long to_ratio(u64 period, u64 runtime)
1968 if (runtime == RUNTIME_INF)
1972 * Doing this here saves a lot of checks in all
1973 * the calling paths, and returning zero seems
1974 * safe for them anyway.
1979 return div64_u64(runtime << 20, period);
1983 inline struct dl_bw *dl_bw_of(int i)
1985 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1986 "sched RCU must be held");
1987 return &cpu_rq(i)->rd->dl_bw;
1990 static inline int dl_bw_cpus(int i)
1992 struct root_domain *rd = cpu_rq(i)->rd;
1995 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1996 "sched RCU must be held");
1997 for_each_cpu_and(i, rd->span, cpu_active_mask)
2003 inline struct dl_bw *dl_bw_of(int i)
2005 return &cpu_rq(i)->dl.dl_bw;
2008 static inline int dl_bw_cpus(int i)
2015 * We must be sure that accepting a new task (or allowing changing the
2016 * parameters of an existing one) is consistent with the bandwidth
2017 * constraints. If yes, this function also accordingly updates the currently
2018 * allocated bandwidth to reflect the new situation.
2020 * This function is called while holding p's rq->lock.
2022 * XXX we should delay bw change until the task's 0-lag point, see
2025 static int dl_overflow(struct task_struct *p, int policy,
2026 const struct sched_attr *attr)
2029 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2030 u64 period = attr->sched_period ?: attr->sched_deadline;
2031 u64 runtime = attr->sched_runtime;
2032 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2035 if (new_bw == p->dl.dl_bw)
2039 * Either if a task, enters, leave, or stays -deadline but changes
2040 * its parameters, we may need to update accordingly the total
2041 * allocated bandwidth of the container.
2043 raw_spin_lock(&dl_b->lock);
2044 cpus = dl_bw_cpus(task_cpu(p));
2045 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2046 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2047 __dl_add(dl_b, new_bw);
2049 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2050 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2051 __dl_clear(dl_b, p->dl.dl_bw);
2052 __dl_add(dl_b, new_bw);
2054 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2055 __dl_clear(dl_b, p->dl.dl_bw);
2058 raw_spin_unlock(&dl_b->lock);
2063 extern void init_dl_bw(struct dl_bw *dl_b);
2066 * wake_up_new_task - wake up a newly created task for the first time.
2068 * This function will do some initial scheduler statistics housekeeping
2069 * that must be done for every newly created context, then puts the task
2070 * on the runqueue and wakes it.
2072 void wake_up_new_task(struct task_struct *p)
2074 unsigned long flags;
2077 raw_spin_lock_irqsave(&p->pi_lock, flags);
2080 * Fork balancing, do it here and not earlier because:
2081 * - cpus_allowed can change in the fork path
2082 * - any previously selected cpu might disappear through hotplug
2084 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2087 /* Initialize new task's runnable average */
2088 init_task_runnable_average(p);
2089 rq = __task_rq_lock(p);
2090 activate_task(rq, p, 0);
2091 p->on_rq = TASK_ON_RQ_QUEUED;
2092 trace_sched_wakeup_new(p, true);
2093 check_preempt_curr(rq, p, WF_FORK);
2095 if (p->sched_class->task_woken)
2096 p->sched_class->task_woken(rq, p);
2098 task_rq_unlock(rq, p, &flags);
2101 #ifdef CONFIG_PREEMPT_NOTIFIERS
2104 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2105 * @notifier: notifier struct to register
2107 void preempt_notifier_register(struct preempt_notifier *notifier)
2109 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2111 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2114 * preempt_notifier_unregister - no longer interested in preemption notifications
2115 * @notifier: notifier struct to unregister
2117 * This is safe to call from within a preemption notifier.
2119 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2121 hlist_del(¬ifier->link);
2123 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2125 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2127 struct preempt_notifier *notifier;
2129 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2130 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2134 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2135 struct task_struct *next)
2137 struct preempt_notifier *notifier;
2139 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2140 notifier->ops->sched_out(notifier, next);
2143 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2145 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2150 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2151 struct task_struct *next)
2155 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2158 * prepare_task_switch - prepare to switch tasks
2159 * @rq: the runqueue preparing to switch
2160 * @prev: the current task that is being switched out
2161 * @next: the task we are going to switch to.
2163 * This is called with the rq lock held and interrupts off. It must
2164 * be paired with a subsequent finish_task_switch after the context
2167 * prepare_task_switch sets up locking and calls architecture specific
2171 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2172 struct task_struct *next)
2174 trace_sched_switch(prev, next);
2175 sched_info_switch(rq, prev, next);
2176 perf_event_task_sched_out(prev, next);
2177 fire_sched_out_preempt_notifiers(prev, next);
2178 prepare_lock_switch(rq, next);
2179 prepare_arch_switch(next);
2183 * finish_task_switch - clean up after a task-switch
2184 * @prev: the thread we just switched away from.
2186 * finish_task_switch must be called after the context switch, paired
2187 * with a prepare_task_switch call before the context switch.
2188 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2189 * and do any other architecture-specific cleanup actions.
2191 * Note that we may have delayed dropping an mm in context_switch(). If
2192 * so, we finish that here outside of the runqueue lock. (Doing it
2193 * with the lock held can cause deadlocks; see schedule() for
2196 * The context switch have flipped the stack from under us and restored the
2197 * local variables which were saved when this task called schedule() in the
2198 * past. prev == current is still correct but we need to recalculate this_rq
2199 * because prev may have moved to another CPU.
2201 static struct rq *finish_task_switch(struct task_struct *prev)
2202 __releases(rq->lock)
2204 struct rq *rq = this_rq();
2205 struct mm_struct *mm = rq->prev_mm;
2211 * A task struct has one reference for the use as "current".
2212 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2213 * schedule one last time. The schedule call will never return, and
2214 * the scheduled task must drop that reference.
2215 * The test for TASK_DEAD must occur while the runqueue locks are
2216 * still held, otherwise prev could be scheduled on another cpu, die
2217 * there before we look at prev->state, and then the reference would
2219 * Manfred Spraul <manfred@colorfullife.com>
2221 prev_state = prev->state;
2222 vtime_task_switch(prev);
2223 finish_arch_switch(prev);
2224 perf_event_task_sched_in(prev, current);
2225 finish_lock_switch(rq, prev);
2226 finish_arch_post_lock_switch();
2228 fire_sched_in_preempt_notifiers(current);
2231 if (unlikely(prev_state == TASK_DEAD)) {
2232 if (prev->sched_class->task_dead)
2233 prev->sched_class->task_dead(prev);
2236 * Remove function-return probe instances associated with this
2237 * task and put them back on the free list.
2239 kprobe_flush_task(prev);
2240 put_task_struct(prev);
2243 tick_nohz_task_switch(current);
2249 /* rq->lock is NOT held, but preemption is disabled */
2250 static inline void post_schedule(struct rq *rq)
2252 if (rq->post_schedule) {
2253 unsigned long flags;
2255 raw_spin_lock_irqsave(&rq->lock, flags);
2256 if (rq->curr->sched_class->post_schedule)
2257 rq->curr->sched_class->post_schedule(rq);
2258 raw_spin_unlock_irqrestore(&rq->lock, flags);
2260 rq->post_schedule = 0;
2266 static inline void post_schedule(struct rq *rq)
2273 * schedule_tail - first thing a freshly forked thread must call.
2274 * @prev: the thread we just switched away from.
2276 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2277 __releases(rq->lock)
2281 /* finish_task_switch() drops rq->lock and enables preemtion */
2283 rq = finish_task_switch(prev);
2287 if (current->set_child_tid)
2288 put_user(task_pid_vnr(current), current->set_child_tid);
2292 * context_switch - switch to the new MM and the new thread's register state.
2294 static inline struct rq *
2295 context_switch(struct rq *rq, struct task_struct *prev,
2296 struct task_struct *next)
2298 struct mm_struct *mm, *oldmm;
2300 prepare_task_switch(rq, prev, next);
2303 oldmm = prev->active_mm;
2305 * For paravirt, this is coupled with an exit in switch_to to
2306 * combine the page table reload and the switch backend into
2309 arch_start_context_switch(prev);
2312 next->active_mm = oldmm;
2313 atomic_inc(&oldmm->mm_count);
2314 enter_lazy_tlb(oldmm, next);
2316 switch_mm(oldmm, mm, next);
2319 prev->active_mm = NULL;
2320 rq->prev_mm = oldmm;
2323 * Since the runqueue lock will be released by the next
2324 * task (which is an invalid locking op but in the case
2325 * of the scheduler it's an obvious special-case), so we
2326 * do an early lockdep release here:
2328 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2330 context_tracking_task_switch(prev, next);
2331 /* Here we just switch the register state and the stack. */
2332 switch_to(prev, next, prev);
2335 return finish_task_switch(prev);
2339 * nr_running and nr_context_switches:
2341 * externally visible scheduler statistics: current number of runnable
2342 * threads, total number of context switches performed since bootup.
2344 unsigned long nr_running(void)
2346 unsigned long i, sum = 0;
2348 for_each_online_cpu(i)
2349 sum += cpu_rq(i)->nr_running;
2355 * Check if only the current task is running on the cpu.
2357 bool single_task_running(void)
2359 if (cpu_rq(smp_processor_id())->nr_running == 1)
2364 EXPORT_SYMBOL(single_task_running);
2366 unsigned long long nr_context_switches(void)
2369 unsigned long long sum = 0;
2371 for_each_possible_cpu(i)
2372 sum += cpu_rq(i)->nr_switches;
2377 unsigned long nr_iowait(void)
2379 unsigned long i, sum = 0;
2381 for_each_possible_cpu(i)
2382 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2387 unsigned long nr_iowait_cpu(int cpu)
2389 struct rq *this = cpu_rq(cpu);
2390 return atomic_read(&this->nr_iowait);
2393 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2395 struct rq *this = this_rq();
2396 *nr_waiters = atomic_read(&this->nr_iowait);
2397 *load = this->cpu_load[0];
2403 * sched_exec - execve() is a valuable balancing opportunity, because at
2404 * this point the task has the smallest effective memory and cache footprint.
2406 void sched_exec(void)
2408 struct task_struct *p = current;
2409 unsigned long flags;
2412 raw_spin_lock_irqsave(&p->pi_lock, flags);
2413 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2414 if (dest_cpu == smp_processor_id())
2417 if (likely(cpu_active(dest_cpu))) {
2418 struct migration_arg arg = { p, dest_cpu };
2420 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2421 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2425 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2430 DEFINE_PER_CPU(struct kernel_stat, kstat);
2431 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2433 EXPORT_PER_CPU_SYMBOL(kstat);
2434 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2437 * Return accounted runtime for the task.
2438 * In case the task is currently running, return the runtime plus current's
2439 * pending runtime that have not been accounted yet.
2441 unsigned long long task_sched_runtime(struct task_struct *p)
2443 unsigned long flags;
2447 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2449 * 64-bit doesn't need locks to atomically read a 64bit value.
2450 * So we have a optimization chance when the task's delta_exec is 0.
2451 * Reading ->on_cpu is racy, but this is ok.
2453 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2454 * If we race with it entering cpu, unaccounted time is 0. This is
2455 * indistinguishable from the read occurring a few cycles earlier.
2456 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2457 * been accounted, so we're correct here as well.
2459 if (!p->on_cpu || !task_on_rq_queued(p))
2460 return p->se.sum_exec_runtime;
2463 rq = task_rq_lock(p, &flags);
2465 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2466 * project cycles that may never be accounted to this
2467 * thread, breaking clock_gettime().
2469 if (task_current(rq, p) && task_on_rq_queued(p)) {
2470 update_rq_clock(rq);
2471 p->sched_class->update_curr(rq);
2473 ns = p->se.sum_exec_runtime;
2474 task_rq_unlock(rq, p, &flags);
2480 * This function gets called by the timer code, with HZ frequency.
2481 * We call it with interrupts disabled.
2483 void scheduler_tick(void)
2485 int cpu = smp_processor_id();
2486 struct rq *rq = cpu_rq(cpu);
2487 struct task_struct *curr = rq->curr;
2491 raw_spin_lock(&rq->lock);
2492 update_rq_clock(rq);
2493 curr->sched_class->task_tick(rq, curr, 0);
2494 update_cpu_load_active(rq);
2495 raw_spin_unlock(&rq->lock);
2497 perf_event_task_tick();
2500 rq->idle_balance = idle_cpu(cpu);
2501 trigger_load_balance(rq);
2503 rq_last_tick_reset(rq);
2506 #ifdef CONFIG_NO_HZ_FULL
2508 * scheduler_tick_max_deferment
2510 * Keep at least one tick per second when a single
2511 * active task is running because the scheduler doesn't
2512 * yet completely support full dynticks environment.
2514 * This makes sure that uptime, CFS vruntime, load
2515 * balancing, etc... continue to move forward, even
2516 * with a very low granularity.
2518 * Return: Maximum deferment in nanoseconds.
2520 u64 scheduler_tick_max_deferment(void)
2522 struct rq *rq = this_rq();
2523 unsigned long next, now = ACCESS_ONCE(jiffies);
2525 next = rq->last_sched_tick + HZ;
2527 if (time_before_eq(next, now))
2530 return jiffies_to_nsecs(next - now);
2534 notrace unsigned long get_parent_ip(unsigned long addr)
2536 if (in_lock_functions(addr)) {
2537 addr = CALLER_ADDR2;
2538 if (in_lock_functions(addr))
2539 addr = CALLER_ADDR3;
2544 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2545 defined(CONFIG_PREEMPT_TRACER))
2547 void preempt_count_add(int val)
2549 #ifdef CONFIG_DEBUG_PREEMPT
2553 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2556 __preempt_count_add(val);
2557 #ifdef CONFIG_DEBUG_PREEMPT
2559 * Spinlock count overflowing soon?
2561 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2564 if (preempt_count() == val) {
2565 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2566 #ifdef CONFIG_DEBUG_PREEMPT
2567 current->preempt_disable_ip = ip;
2569 trace_preempt_off(CALLER_ADDR0, ip);
2572 EXPORT_SYMBOL(preempt_count_add);
2573 NOKPROBE_SYMBOL(preempt_count_add);
2575 void preempt_count_sub(int val)
2577 #ifdef CONFIG_DEBUG_PREEMPT
2581 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2584 * Is the spinlock portion underflowing?
2586 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2587 !(preempt_count() & PREEMPT_MASK)))
2591 if (preempt_count() == val)
2592 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2593 __preempt_count_sub(val);
2595 EXPORT_SYMBOL(preempt_count_sub);
2596 NOKPROBE_SYMBOL(preempt_count_sub);
2601 * Print scheduling while atomic bug:
2603 static noinline void __schedule_bug(struct task_struct *prev)
2605 if (oops_in_progress)
2608 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2609 prev->comm, prev->pid, preempt_count());
2611 debug_show_held_locks(prev);
2613 if (irqs_disabled())
2614 print_irqtrace_events(prev);
2615 #ifdef CONFIG_DEBUG_PREEMPT
2616 if (in_atomic_preempt_off()) {
2617 pr_err("Preemption disabled at:");
2618 print_ip_sym(current->preempt_disable_ip);
2623 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2627 * Various schedule()-time debugging checks and statistics:
2629 static inline void schedule_debug(struct task_struct *prev)
2631 #ifdef CONFIG_SCHED_STACK_END_CHECK
2632 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2635 * Test if we are atomic. Since do_exit() needs to call into
2636 * schedule() atomically, we ignore that path. Otherwise whine
2637 * if we are scheduling when we should not.
2639 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2640 __schedule_bug(prev);
2643 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2645 schedstat_inc(this_rq(), sched_count);
2649 * Pick up the highest-prio task:
2651 static inline struct task_struct *
2652 pick_next_task(struct rq *rq, struct task_struct *prev)
2654 const struct sched_class *class = &fair_sched_class;
2655 struct task_struct *p;
2658 * Optimization: we know that if all tasks are in
2659 * the fair class we can call that function directly:
2661 if (likely(prev->sched_class == class &&
2662 rq->nr_running == rq->cfs.h_nr_running)) {
2663 p = fair_sched_class.pick_next_task(rq, prev);
2664 if (unlikely(p == RETRY_TASK))
2667 /* assumes fair_sched_class->next == idle_sched_class */
2669 p = idle_sched_class.pick_next_task(rq, prev);
2675 for_each_class(class) {
2676 p = class->pick_next_task(rq, prev);
2678 if (unlikely(p == RETRY_TASK))
2684 BUG(); /* the idle class will always have a runnable task */
2688 * __schedule() is the main scheduler function.
2690 * The main means of driving the scheduler and thus entering this function are:
2692 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2694 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2695 * paths. For example, see arch/x86/entry_64.S.
2697 * To drive preemption between tasks, the scheduler sets the flag in timer
2698 * interrupt handler scheduler_tick().
2700 * 3. Wakeups don't really cause entry into schedule(). They add a
2701 * task to the run-queue and that's it.
2703 * Now, if the new task added to the run-queue preempts the current
2704 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2705 * called on the nearest possible occasion:
2707 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2709 * - in syscall or exception context, at the next outmost
2710 * preempt_enable(). (this might be as soon as the wake_up()'s
2713 * - in IRQ context, return from interrupt-handler to
2714 * preemptible context
2716 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2719 * - cond_resched() call
2720 * - explicit schedule() call
2721 * - return from syscall or exception to user-space
2722 * - return from interrupt-handler to user-space
2724 * WARNING: all callers must re-check need_resched() afterward and reschedule
2725 * accordingly in case an event triggered the need for rescheduling (such as
2726 * an interrupt waking up a task) while preemption was disabled in __schedule().
2728 static void __sched __schedule(void)
2730 struct task_struct *prev, *next;
2731 unsigned long *switch_count;
2736 cpu = smp_processor_id();
2738 rcu_note_context_switch();
2741 schedule_debug(prev);
2743 if (sched_feat(HRTICK))
2747 * Make sure that signal_pending_state()->signal_pending() below
2748 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2749 * done by the caller to avoid the race with signal_wake_up().
2751 smp_mb__before_spinlock();
2752 raw_spin_lock_irq(&rq->lock);
2754 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2756 switch_count = &prev->nivcsw;
2757 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2758 if (unlikely(signal_pending_state(prev->state, prev))) {
2759 prev->state = TASK_RUNNING;
2761 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2765 * If a worker went to sleep, notify and ask workqueue
2766 * whether it wants to wake up a task to maintain
2769 if (prev->flags & PF_WQ_WORKER) {
2770 struct task_struct *to_wakeup;
2772 to_wakeup = wq_worker_sleeping(prev, cpu);
2774 try_to_wake_up_local(to_wakeup);
2777 switch_count = &prev->nvcsw;
2780 if (task_on_rq_queued(prev))
2781 update_rq_clock(rq);
2783 next = pick_next_task(rq, prev);
2784 clear_tsk_need_resched(prev);
2785 clear_preempt_need_resched();
2786 rq->clock_skip_update = 0;
2788 if (likely(prev != next)) {
2793 rq = context_switch(rq, prev, next); /* unlocks the rq */
2796 raw_spin_unlock_irq(&rq->lock);
2800 sched_preempt_enable_no_resched();
2803 static inline void sched_submit_work(struct task_struct *tsk)
2805 if (!tsk->state || tsk_is_pi_blocked(tsk))
2808 * If we are going to sleep and we have plugged IO queued,
2809 * make sure to submit it to avoid deadlocks.
2811 if (blk_needs_flush_plug(tsk))
2812 blk_schedule_flush_plug(tsk);
2815 asmlinkage __visible void __sched schedule(void)
2817 struct task_struct *tsk = current;
2819 sched_submit_work(tsk);
2822 } while (need_resched());
2824 EXPORT_SYMBOL(schedule);
2826 #ifdef CONFIG_CONTEXT_TRACKING
2827 asmlinkage __visible void __sched schedule_user(void)
2830 * If we come here after a random call to set_need_resched(),
2831 * or we have been woken up remotely but the IPI has not yet arrived,
2832 * we haven't yet exited the RCU idle mode. Do it here manually until
2833 * we find a better solution.
2835 * NB: There are buggy callers of this function. Ideally we
2836 * should warn if prev_state != IN_USER, but that will trigger
2837 * too frequently to make sense yet.
2839 enum ctx_state prev_state = exception_enter();
2841 exception_exit(prev_state);
2846 * schedule_preempt_disabled - called with preemption disabled
2848 * Returns with preemption disabled. Note: preempt_count must be 1
2850 void __sched schedule_preempt_disabled(void)
2852 sched_preempt_enable_no_resched();
2857 static void __sched notrace preempt_schedule_common(void)
2860 __preempt_count_add(PREEMPT_ACTIVE);
2862 __preempt_count_sub(PREEMPT_ACTIVE);
2865 * Check again in case we missed a preemption opportunity
2866 * between schedule and now.
2869 } while (need_resched());
2872 #ifdef CONFIG_PREEMPT
2874 * this is the entry point to schedule() from in-kernel preemption
2875 * off of preempt_enable. Kernel preemptions off return from interrupt
2876 * occur there and call schedule directly.
2878 asmlinkage __visible void __sched notrace preempt_schedule(void)
2881 * If there is a non-zero preempt_count or interrupts are disabled,
2882 * we do not want to preempt the current task. Just return..
2884 if (likely(!preemptible()))
2887 preempt_schedule_common();
2889 NOKPROBE_SYMBOL(preempt_schedule);
2890 EXPORT_SYMBOL(preempt_schedule);
2892 #ifdef CONFIG_CONTEXT_TRACKING
2894 * preempt_schedule_context - preempt_schedule called by tracing
2896 * The tracing infrastructure uses preempt_enable_notrace to prevent
2897 * recursion and tracing preempt enabling caused by the tracing
2898 * infrastructure itself. But as tracing can happen in areas coming
2899 * from userspace or just about to enter userspace, a preempt enable
2900 * can occur before user_exit() is called. This will cause the scheduler
2901 * to be called when the system is still in usermode.
2903 * To prevent this, the preempt_enable_notrace will use this function
2904 * instead of preempt_schedule() to exit user context if needed before
2905 * calling the scheduler.
2907 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2909 enum ctx_state prev_ctx;
2911 if (likely(!preemptible()))
2915 __preempt_count_add(PREEMPT_ACTIVE);
2917 * Needs preempt disabled in case user_exit() is traced
2918 * and the tracer calls preempt_enable_notrace() causing
2919 * an infinite recursion.
2921 prev_ctx = exception_enter();
2923 exception_exit(prev_ctx);
2925 __preempt_count_sub(PREEMPT_ACTIVE);
2927 } while (need_resched());
2929 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2930 #endif /* CONFIG_CONTEXT_TRACKING */
2932 #endif /* CONFIG_PREEMPT */
2935 * this is the entry point to schedule() from kernel preemption
2936 * off of irq context.
2937 * Note, that this is called and return with irqs disabled. This will
2938 * protect us against recursive calling from irq.
2940 asmlinkage __visible void __sched preempt_schedule_irq(void)
2942 enum ctx_state prev_state;
2944 /* Catch callers which need to be fixed */
2945 BUG_ON(preempt_count() || !irqs_disabled());
2947 prev_state = exception_enter();
2950 __preempt_count_add(PREEMPT_ACTIVE);
2953 local_irq_disable();
2954 __preempt_count_sub(PREEMPT_ACTIVE);
2957 * Check again in case we missed a preemption opportunity
2958 * between schedule and now.
2961 } while (need_resched());
2963 exception_exit(prev_state);
2966 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2969 return try_to_wake_up(curr->private, mode, wake_flags);
2971 EXPORT_SYMBOL(default_wake_function);
2973 #ifdef CONFIG_RT_MUTEXES
2976 * rt_mutex_setprio - set the current priority of a task
2978 * @prio: prio value (kernel-internal form)
2980 * This function changes the 'effective' priority of a task. It does
2981 * not touch ->normal_prio like __setscheduler().
2983 * Used by the rt_mutex code to implement priority inheritance
2984 * logic. Call site only calls if the priority of the task changed.
2986 void rt_mutex_setprio(struct task_struct *p, int prio)
2988 int oldprio, queued, running, enqueue_flag = 0;
2990 const struct sched_class *prev_class;
2992 BUG_ON(prio > MAX_PRIO);
2994 rq = __task_rq_lock(p);
2997 * Idle task boosting is a nono in general. There is one
2998 * exception, when PREEMPT_RT and NOHZ is active:
3000 * The idle task calls get_next_timer_interrupt() and holds
3001 * the timer wheel base->lock on the CPU and another CPU wants
3002 * to access the timer (probably to cancel it). We can safely
3003 * ignore the boosting request, as the idle CPU runs this code
3004 * with interrupts disabled and will complete the lock
3005 * protected section without being interrupted. So there is no
3006 * real need to boost.
3008 if (unlikely(p == rq->idle)) {
3009 WARN_ON(p != rq->curr);
3010 WARN_ON(p->pi_blocked_on);
3014 trace_sched_pi_setprio(p, prio);
3016 prev_class = p->sched_class;
3017 queued = task_on_rq_queued(p);
3018 running = task_current(rq, p);
3020 dequeue_task(rq, p, 0);
3022 put_prev_task(rq, p);
3025 * Boosting condition are:
3026 * 1. -rt task is running and holds mutex A
3027 * --> -dl task blocks on mutex A
3029 * 2. -dl task is running and holds mutex A
3030 * --> -dl task blocks on mutex A and could preempt the
3033 if (dl_prio(prio)) {
3034 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3035 if (!dl_prio(p->normal_prio) ||
3036 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3037 p->dl.dl_boosted = 1;
3038 p->dl.dl_throttled = 0;
3039 enqueue_flag = ENQUEUE_REPLENISH;
3041 p->dl.dl_boosted = 0;
3042 p->sched_class = &dl_sched_class;
3043 } else if (rt_prio(prio)) {
3044 if (dl_prio(oldprio))
3045 p->dl.dl_boosted = 0;
3047 enqueue_flag = ENQUEUE_HEAD;
3048 p->sched_class = &rt_sched_class;
3050 if (dl_prio(oldprio))
3051 p->dl.dl_boosted = 0;
3052 if (rt_prio(oldprio))
3054 p->sched_class = &fair_sched_class;
3060 p->sched_class->set_curr_task(rq);
3062 enqueue_task(rq, p, enqueue_flag);
3064 check_class_changed(rq, p, prev_class, oldprio);
3066 __task_rq_unlock(rq);
3070 void set_user_nice(struct task_struct *p, long nice)
3072 int old_prio, delta, queued;
3073 unsigned long flags;
3076 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3079 * We have to be careful, if called from sys_setpriority(),
3080 * the task might be in the middle of scheduling on another CPU.
3082 rq = task_rq_lock(p, &flags);
3084 * The RT priorities are set via sched_setscheduler(), but we still
3085 * allow the 'normal' nice value to be set - but as expected
3086 * it wont have any effect on scheduling until the task is
3087 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3089 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3090 p->static_prio = NICE_TO_PRIO(nice);
3093 queued = task_on_rq_queued(p);
3095 dequeue_task(rq, p, 0);
3097 p->static_prio = NICE_TO_PRIO(nice);
3100 p->prio = effective_prio(p);
3101 delta = p->prio - old_prio;
3104 enqueue_task(rq, p, 0);
3106 * If the task increased its priority or is running and
3107 * lowered its priority, then reschedule its CPU:
3109 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3113 task_rq_unlock(rq, p, &flags);
3115 EXPORT_SYMBOL(set_user_nice);
3118 * can_nice - check if a task can reduce its nice value
3122 int can_nice(const struct task_struct *p, const int nice)
3124 /* convert nice value [19,-20] to rlimit style value [1,40] */
3125 int nice_rlim = nice_to_rlimit(nice);
3127 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3128 capable(CAP_SYS_NICE));
3131 #ifdef __ARCH_WANT_SYS_NICE
3134 * sys_nice - change the priority of the current process.
3135 * @increment: priority increment
3137 * sys_setpriority is a more generic, but much slower function that
3138 * does similar things.
3140 SYSCALL_DEFINE1(nice, int, increment)
3145 * Setpriority might change our priority at the same moment.
3146 * We don't have to worry. Conceptually one call occurs first
3147 * and we have a single winner.
3149 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3150 nice = task_nice(current) + increment;
3152 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3153 if (increment < 0 && !can_nice(current, nice))
3156 retval = security_task_setnice(current, nice);
3160 set_user_nice(current, nice);
3167 * task_prio - return the priority value of a given task.
3168 * @p: the task in question.
3170 * Return: The priority value as seen by users in /proc.
3171 * RT tasks are offset by -200. Normal tasks are centered
3172 * around 0, value goes from -16 to +15.
3174 int task_prio(const struct task_struct *p)
3176 return p->prio - MAX_RT_PRIO;
3180 * idle_cpu - is a given cpu idle currently?
3181 * @cpu: the processor in question.
3183 * Return: 1 if the CPU is currently idle. 0 otherwise.
3185 int idle_cpu(int cpu)
3187 struct rq *rq = cpu_rq(cpu);
3189 if (rq->curr != rq->idle)
3196 if (!llist_empty(&rq->wake_list))
3204 * idle_task - return the idle task for a given cpu.
3205 * @cpu: the processor in question.
3207 * Return: The idle task for the cpu @cpu.
3209 struct task_struct *idle_task(int cpu)
3211 return cpu_rq(cpu)->idle;
3215 * find_process_by_pid - find a process with a matching PID value.
3216 * @pid: the pid in question.
3218 * The task of @pid, if found. %NULL otherwise.
3220 static struct task_struct *find_process_by_pid(pid_t pid)
3222 return pid ? find_task_by_vpid(pid) : current;
3226 * This function initializes the sched_dl_entity of a newly becoming
3227 * SCHED_DEADLINE task.
3229 * Only the static values are considered here, the actual runtime and the
3230 * absolute deadline will be properly calculated when the task is enqueued
3231 * for the first time with its new policy.
3234 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3236 struct sched_dl_entity *dl_se = &p->dl;
3238 dl_se->dl_runtime = attr->sched_runtime;
3239 dl_se->dl_deadline = attr->sched_deadline;
3240 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3241 dl_se->flags = attr->sched_flags;
3242 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3245 * Changing the parameters of a task is 'tricky' and we're not doing
3246 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3248 * What we SHOULD do is delay the bandwidth release until the 0-lag
3249 * point. This would include retaining the task_struct until that time
3250 * and change dl_overflow() to not immediately decrement the current
3253 * Instead we retain the current runtime/deadline and let the new
3254 * parameters take effect after the current reservation period lapses.
3255 * This is safe (albeit pessimistic) because the 0-lag point is always
3256 * before the current scheduling deadline.
3258 * We can still have temporary overloads because we do not delay the
3259 * change in bandwidth until that time; so admission control is
3260 * not on the safe side. It does however guarantee tasks will never
3261 * consume more than promised.
3266 * sched_setparam() passes in -1 for its policy, to let the functions
3267 * it calls know not to change it.
3269 #define SETPARAM_POLICY -1
3271 static void __setscheduler_params(struct task_struct *p,
3272 const struct sched_attr *attr)
3274 int policy = attr->sched_policy;
3276 if (policy == SETPARAM_POLICY)
3281 if (dl_policy(policy))
3282 __setparam_dl(p, attr);
3283 else if (fair_policy(policy))
3284 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3287 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3288 * !rt_policy. Always setting this ensures that things like
3289 * getparam()/getattr() don't report silly values for !rt tasks.
3291 p->rt_priority = attr->sched_priority;
3292 p->normal_prio = normal_prio(p);
3296 /* Actually do priority change: must hold pi & rq lock. */
3297 static void __setscheduler(struct rq *rq, struct task_struct *p,
3298 const struct sched_attr *attr)
3300 __setscheduler_params(p, attr);
3303 * If we get here, there was no pi waiters boosting the
3304 * task. It is safe to use the normal prio.
3306 p->prio = normal_prio(p);
3308 if (dl_prio(p->prio))
3309 p->sched_class = &dl_sched_class;
3310 else if (rt_prio(p->prio))
3311 p->sched_class = &rt_sched_class;
3313 p->sched_class = &fair_sched_class;
3317 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3319 struct sched_dl_entity *dl_se = &p->dl;
3321 attr->sched_priority = p->rt_priority;
3322 attr->sched_runtime = dl_se->dl_runtime;
3323 attr->sched_deadline = dl_se->dl_deadline;
3324 attr->sched_period = dl_se->dl_period;
3325 attr->sched_flags = dl_se->flags;
3329 * This function validates the new parameters of a -deadline task.
3330 * We ask for the deadline not being zero, and greater or equal
3331 * than the runtime, as well as the period of being zero or
3332 * greater than deadline. Furthermore, we have to be sure that
3333 * user parameters are above the internal resolution of 1us (we
3334 * check sched_runtime only since it is always the smaller one) and
3335 * below 2^63 ns (we have to check both sched_deadline and
3336 * sched_period, as the latter can be zero).
3339 __checkparam_dl(const struct sched_attr *attr)
3342 if (attr->sched_deadline == 0)
3346 * Since we truncate DL_SCALE bits, make sure we're at least
3349 if (attr->sched_runtime < (1ULL << DL_SCALE))
3353 * Since we use the MSB for wrap-around and sign issues, make
3354 * sure it's not set (mind that period can be equal to zero).
3356 if (attr->sched_deadline & (1ULL << 63) ||
3357 attr->sched_period & (1ULL << 63))
3360 /* runtime <= deadline <= period (if period != 0) */
3361 if ((attr->sched_period != 0 &&
3362 attr->sched_period < attr->sched_deadline) ||
3363 attr->sched_deadline < attr->sched_runtime)
3370 * check the target process has a UID that matches the current process's
3372 static bool check_same_owner(struct task_struct *p)
3374 const struct cred *cred = current_cred(), *pcred;
3378 pcred = __task_cred(p);
3379 match = (uid_eq(cred->euid, pcred->euid) ||
3380 uid_eq(cred->euid, pcred->uid));
3385 static bool dl_param_changed(struct task_struct *p,
3386 const struct sched_attr *attr)
3388 struct sched_dl_entity *dl_se = &p->dl;
3390 if (dl_se->dl_runtime != attr->sched_runtime ||
3391 dl_se->dl_deadline != attr->sched_deadline ||
3392 dl_se->dl_period != attr->sched_period ||
3393 dl_se->flags != attr->sched_flags)
3399 static int __sched_setscheduler(struct task_struct *p,
3400 const struct sched_attr *attr,
3403 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3404 MAX_RT_PRIO - 1 - attr->sched_priority;
3405 int retval, oldprio, oldpolicy = -1, queued, running;
3406 int policy = attr->sched_policy;
3407 unsigned long flags;
3408 const struct sched_class *prev_class;
3412 /* may grab non-irq protected spin_locks */
3413 BUG_ON(in_interrupt());
3415 /* double check policy once rq lock held */
3417 reset_on_fork = p->sched_reset_on_fork;
3418 policy = oldpolicy = p->policy;
3420 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3422 if (policy != SCHED_DEADLINE &&
3423 policy != SCHED_FIFO && policy != SCHED_RR &&
3424 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3425 policy != SCHED_IDLE)
3429 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3433 * Valid priorities for SCHED_FIFO and SCHED_RR are
3434 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3435 * SCHED_BATCH and SCHED_IDLE is 0.
3437 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3438 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3440 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3441 (rt_policy(policy) != (attr->sched_priority != 0)))
3445 * Allow unprivileged RT tasks to decrease priority:
3447 if (user && !capable(CAP_SYS_NICE)) {
3448 if (fair_policy(policy)) {
3449 if (attr->sched_nice < task_nice(p) &&
3450 !can_nice(p, attr->sched_nice))
3454 if (rt_policy(policy)) {
3455 unsigned long rlim_rtprio =
3456 task_rlimit(p, RLIMIT_RTPRIO);
3458 /* can't set/change the rt policy */
3459 if (policy != p->policy && !rlim_rtprio)
3462 /* can't increase priority */
3463 if (attr->sched_priority > p->rt_priority &&
3464 attr->sched_priority > rlim_rtprio)
3469 * Can't set/change SCHED_DEADLINE policy at all for now
3470 * (safest behavior); in the future we would like to allow
3471 * unprivileged DL tasks to increase their relative deadline
3472 * or reduce their runtime (both ways reducing utilization)
3474 if (dl_policy(policy))
3478 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3479 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3481 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3482 if (!can_nice(p, task_nice(p)))
3486 /* can't change other user's priorities */
3487 if (!check_same_owner(p))
3490 /* Normal users shall not reset the sched_reset_on_fork flag */
3491 if (p->sched_reset_on_fork && !reset_on_fork)
3496 retval = security_task_setscheduler(p);
3502 * make sure no PI-waiters arrive (or leave) while we are
3503 * changing the priority of the task:
3505 * To be able to change p->policy safely, the appropriate
3506 * runqueue lock must be held.
3508 rq = task_rq_lock(p, &flags);
3511 * Changing the policy of the stop threads its a very bad idea
3513 if (p == rq->stop) {
3514 task_rq_unlock(rq, p, &flags);
3519 * If not changing anything there's no need to proceed further,
3520 * but store a possible modification of reset_on_fork.
3522 if (unlikely(policy == p->policy)) {
3523 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3525 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3527 if (dl_policy(policy) && dl_param_changed(p, attr))
3530 p->sched_reset_on_fork = reset_on_fork;
3531 task_rq_unlock(rq, p, &flags);
3537 #ifdef CONFIG_RT_GROUP_SCHED
3539 * Do not allow realtime tasks into groups that have no runtime
3542 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3543 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3544 !task_group_is_autogroup(task_group(p))) {
3545 task_rq_unlock(rq, p, &flags);
3550 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3551 cpumask_t *span = rq->rd->span;
3554 * Don't allow tasks with an affinity mask smaller than
3555 * the entire root_domain to become SCHED_DEADLINE. We
3556 * will also fail if there's no bandwidth available.
3558 if (!cpumask_subset(span, &p->cpus_allowed) ||
3559 rq->rd->dl_bw.bw == 0) {
3560 task_rq_unlock(rq, p, &flags);
3567 /* recheck policy now with rq lock held */
3568 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3569 policy = oldpolicy = -1;
3570 task_rq_unlock(rq, p, &flags);
3575 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3576 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3579 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3580 task_rq_unlock(rq, p, &flags);
3584 p->sched_reset_on_fork = reset_on_fork;
3588 * Special case for priority boosted tasks.
3590 * If the new priority is lower or equal (user space view)
3591 * than the current (boosted) priority, we just store the new
3592 * normal parameters and do not touch the scheduler class and
3593 * the runqueue. This will be done when the task deboost
3596 if (rt_mutex_check_prio(p, newprio)) {
3597 __setscheduler_params(p, attr);
3598 task_rq_unlock(rq, p, &flags);
3602 queued = task_on_rq_queued(p);
3603 running = task_current(rq, p);
3605 dequeue_task(rq, p, 0);
3607 put_prev_task(rq, p);
3609 prev_class = p->sched_class;
3610 __setscheduler(rq, p, attr);
3613 p->sched_class->set_curr_task(rq);
3616 * We enqueue to tail when the priority of a task is
3617 * increased (user space view).
3619 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3622 check_class_changed(rq, p, prev_class, oldprio);
3623 task_rq_unlock(rq, p, &flags);
3625 rt_mutex_adjust_pi(p);
3630 static int _sched_setscheduler(struct task_struct *p, int policy,
3631 const struct sched_param *param, bool check)
3633 struct sched_attr attr = {
3634 .sched_policy = policy,
3635 .sched_priority = param->sched_priority,
3636 .sched_nice = PRIO_TO_NICE(p->static_prio),
3639 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3640 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3641 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3642 policy &= ~SCHED_RESET_ON_FORK;
3643 attr.sched_policy = policy;
3646 return __sched_setscheduler(p, &attr, check);
3649 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3650 * @p: the task in question.
3651 * @policy: new policy.
3652 * @param: structure containing the new RT priority.
3654 * Return: 0 on success. An error code otherwise.
3656 * NOTE that the task may be already dead.
3658 int sched_setscheduler(struct task_struct *p, int policy,
3659 const struct sched_param *param)
3661 return _sched_setscheduler(p, policy, param, true);
3663 EXPORT_SYMBOL_GPL(sched_setscheduler);
3665 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3667 return __sched_setscheduler(p, attr, true);
3669 EXPORT_SYMBOL_GPL(sched_setattr);
3672 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3673 * @p: the task in question.
3674 * @policy: new policy.
3675 * @param: structure containing the new RT priority.
3677 * Just like sched_setscheduler, only don't bother checking if the
3678 * current context has permission. For example, this is needed in
3679 * stop_machine(): we create temporary high priority worker threads,
3680 * but our caller might not have that capability.
3682 * Return: 0 on success. An error code otherwise.
3684 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3685 const struct sched_param *param)
3687 return _sched_setscheduler(p, policy, param, false);
3691 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3693 struct sched_param lparam;
3694 struct task_struct *p;
3697 if (!param || pid < 0)
3699 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3704 p = find_process_by_pid(pid);
3706 retval = sched_setscheduler(p, policy, &lparam);
3713 * Mimics kernel/events/core.c perf_copy_attr().
3715 static int sched_copy_attr(struct sched_attr __user *uattr,
3716 struct sched_attr *attr)
3721 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3725 * zero the full structure, so that a short copy will be nice.
3727 memset(attr, 0, sizeof(*attr));
3729 ret = get_user(size, &uattr->size);
3733 if (size > PAGE_SIZE) /* silly large */
3736 if (!size) /* abi compat */
3737 size = SCHED_ATTR_SIZE_VER0;
3739 if (size < SCHED_ATTR_SIZE_VER0)
3743 * If we're handed a bigger struct than we know of,
3744 * ensure all the unknown bits are 0 - i.e. new
3745 * user-space does not rely on any kernel feature
3746 * extensions we dont know about yet.
3748 if (size > sizeof(*attr)) {
3749 unsigned char __user *addr;
3750 unsigned char __user *end;
3753 addr = (void __user *)uattr + sizeof(*attr);
3754 end = (void __user *)uattr + size;
3756 for (; addr < end; addr++) {
3757 ret = get_user(val, addr);
3763 size = sizeof(*attr);
3766 ret = copy_from_user(attr, uattr, size);
3771 * XXX: do we want to be lenient like existing syscalls; or do we want
3772 * to be strict and return an error on out-of-bounds values?
3774 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3779 put_user(sizeof(*attr), &uattr->size);
3784 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3785 * @pid: the pid in question.
3786 * @policy: new policy.
3787 * @param: structure containing the new RT priority.
3789 * Return: 0 on success. An error code otherwise.
3791 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3792 struct sched_param __user *, param)
3794 /* negative values for policy are not valid */
3798 return do_sched_setscheduler(pid, policy, param);
3802 * sys_sched_setparam - set/change the RT priority of a thread
3803 * @pid: the pid in question.
3804 * @param: structure containing the new RT priority.
3806 * Return: 0 on success. An error code otherwise.
3808 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3810 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3814 * sys_sched_setattr - same as above, but with extended sched_attr
3815 * @pid: the pid in question.
3816 * @uattr: structure containing the extended parameters.
3817 * @flags: for future extension.
3819 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3820 unsigned int, flags)
3822 struct sched_attr attr;
3823 struct task_struct *p;
3826 if (!uattr || pid < 0 || flags)
3829 retval = sched_copy_attr(uattr, &attr);
3833 if ((int)attr.sched_policy < 0)
3838 p = find_process_by_pid(pid);
3840 retval = sched_setattr(p, &attr);
3847 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3848 * @pid: the pid in question.
3850 * Return: On success, the policy of the thread. Otherwise, a negative error
3853 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3855 struct task_struct *p;
3863 p = find_process_by_pid(pid);
3865 retval = security_task_getscheduler(p);
3868 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3875 * sys_sched_getparam - get the RT priority of a thread
3876 * @pid: the pid in question.
3877 * @param: structure containing the RT priority.
3879 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3882 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3884 struct sched_param lp = { .sched_priority = 0 };
3885 struct task_struct *p;
3888 if (!param || pid < 0)
3892 p = find_process_by_pid(pid);
3897 retval = security_task_getscheduler(p);
3901 if (task_has_rt_policy(p))
3902 lp.sched_priority = p->rt_priority;
3906 * This one might sleep, we cannot do it with a spinlock held ...
3908 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3917 static int sched_read_attr(struct sched_attr __user *uattr,
3918 struct sched_attr *attr,
3923 if (!access_ok(VERIFY_WRITE, uattr, usize))
3927 * If we're handed a smaller struct than we know of,
3928 * ensure all the unknown bits are 0 - i.e. old
3929 * user-space does not get uncomplete information.
3931 if (usize < sizeof(*attr)) {
3932 unsigned char *addr;
3935 addr = (void *)attr + usize;
3936 end = (void *)attr + sizeof(*attr);
3938 for (; addr < end; addr++) {
3946 ret = copy_to_user(uattr, attr, attr->size);
3954 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3955 * @pid: the pid in question.
3956 * @uattr: structure containing the extended parameters.
3957 * @size: sizeof(attr) for fwd/bwd comp.
3958 * @flags: for future extension.
3960 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3961 unsigned int, size, unsigned int, flags)
3963 struct sched_attr attr = {
3964 .size = sizeof(struct sched_attr),
3966 struct task_struct *p;
3969 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3970 size < SCHED_ATTR_SIZE_VER0 || flags)
3974 p = find_process_by_pid(pid);
3979 retval = security_task_getscheduler(p);
3983 attr.sched_policy = p->policy;
3984 if (p->sched_reset_on_fork)
3985 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3986 if (task_has_dl_policy(p))
3987 __getparam_dl(p, &attr);
3988 else if (task_has_rt_policy(p))
3989 attr.sched_priority = p->rt_priority;
3991 attr.sched_nice = task_nice(p);
3995 retval = sched_read_attr(uattr, &attr, size);
4003 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4005 cpumask_var_t cpus_allowed, new_mask;
4006 struct task_struct *p;
4011 p = find_process_by_pid(pid);
4017 /* Prevent p going away */
4021 if (p->flags & PF_NO_SETAFFINITY) {
4025 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4029 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4031 goto out_free_cpus_allowed;
4034 if (!check_same_owner(p)) {
4036 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4038 goto out_free_new_mask;
4043 retval = security_task_setscheduler(p);
4045 goto out_free_new_mask;
4048 cpuset_cpus_allowed(p, cpus_allowed);
4049 cpumask_and(new_mask, in_mask, cpus_allowed);
4052 * Since bandwidth control happens on root_domain basis,
4053 * if admission test is enabled, we only admit -deadline
4054 * tasks allowed to run on all the CPUs in the task's
4058 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4060 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4063 goto out_free_new_mask;
4069 retval = set_cpus_allowed_ptr(p, new_mask);
4072 cpuset_cpus_allowed(p, cpus_allowed);
4073 if (!cpumask_subset(new_mask, cpus_allowed)) {
4075 * We must have raced with a concurrent cpuset
4076 * update. Just reset the cpus_allowed to the
4077 * cpuset's cpus_allowed
4079 cpumask_copy(new_mask, cpus_allowed);
4084 free_cpumask_var(new_mask);
4085 out_free_cpus_allowed:
4086 free_cpumask_var(cpus_allowed);
4092 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4093 struct cpumask *new_mask)
4095 if (len < cpumask_size())
4096 cpumask_clear(new_mask);
4097 else if (len > cpumask_size())
4098 len = cpumask_size();
4100 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4104 * sys_sched_setaffinity - set the cpu affinity of a process
4105 * @pid: pid of the process
4106 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4107 * @user_mask_ptr: user-space pointer to the new cpu mask
4109 * Return: 0 on success. An error code otherwise.
4111 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4112 unsigned long __user *, user_mask_ptr)
4114 cpumask_var_t new_mask;
4117 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4120 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4122 retval = sched_setaffinity(pid, new_mask);
4123 free_cpumask_var(new_mask);
4127 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4129 struct task_struct *p;
4130 unsigned long flags;
4136 p = find_process_by_pid(pid);
4140 retval = security_task_getscheduler(p);
4144 raw_spin_lock_irqsave(&p->pi_lock, flags);
4145 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4146 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4155 * sys_sched_getaffinity - get the cpu affinity of a process
4156 * @pid: pid of the process
4157 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4158 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4160 * Return: 0 on success. An error code otherwise.
4162 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4163 unsigned long __user *, user_mask_ptr)
4168 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4170 if (len & (sizeof(unsigned long)-1))
4173 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4176 ret = sched_getaffinity(pid, mask);
4178 size_t retlen = min_t(size_t, len, cpumask_size());
4180 if (copy_to_user(user_mask_ptr, mask, retlen))
4185 free_cpumask_var(mask);
4191 * sys_sched_yield - yield the current processor to other threads.
4193 * This function yields the current CPU to other tasks. If there are no
4194 * other threads running on this CPU then this function will return.
4198 SYSCALL_DEFINE0(sched_yield)
4200 struct rq *rq = this_rq_lock();
4202 schedstat_inc(rq, yld_count);
4203 current->sched_class->yield_task(rq);
4206 * Since we are going to call schedule() anyway, there's
4207 * no need to preempt or enable interrupts:
4209 __release(rq->lock);
4210 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4211 do_raw_spin_unlock(&rq->lock);
4212 sched_preempt_enable_no_resched();
4219 int __sched _cond_resched(void)
4221 if (should_resched()) {
4222 preempt_schedule_common();
4227 EXPORT_SYMBOL(_cond_resched);
4230 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4231 * call schedule, and on return reacquire the lock.
4233 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4234 * operations here to prevent schedule() from being called twice (once via
4235 * spin_unlock(), once by hand).
4237 int __cond_resched_lock(spinlock_t *lock)
4239 int resched = should_resched();
4242 lockdep_assert_held(lock);
4244 if (spin_needbreak(lock) || resched) {
4247 preempt_schedule_common();
4255 EXPORT_SYMBOL(__cond_resched_lock);
4257 int __sched __cond_resched_softirq(void)
4259 BUG_ON(!in_softirq());
4261 if (should_resched()) {
4263 preempt_schedule_common();
4269 EXPORT_SYMBOL(__cond_resched_softirq);
4272 * yield - yield the current processor to other threads.
4274 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4276 * The scheduler is at all times free to pick the calling task as the most
4277 * eligible task to run, if removing the yield() call from your code breaks
4278 * it, its already broken.
4280 * Typical broken usage is:
4285 * where one assumes that yield() will let 'the other' process run that will
4286 * make event true. If the current task is a SCHED_FIFO task that will never
4287 * happen. Never use yield() as a progress guarantee!!
4289 * If you want to use yield() to wait for something, use wait_event().
4290 * If you want to use yield() to be 'nice' for others, use cond_resched().
4291 * If you still want to use yield(), do not!
4293 void __sched yield(void)
4295 set_current_state(TASK_RUNNING);
4298 EXPORT_SYMBOL(yield);
4301 * yield_to - yield the current processor to another thread in
4302 * your thread group, or accelerate that thread toward the
4303 * processor it's on.
4305 * @preempt: whether task preemption is allowed or not
4307 * It's the caller's job to ensure that the target task struct
4308 * can't go away on us before we can do any checks.
4311 * true (>0) if we indeed boosted the target task.
4312 * false (0) if we failed to boost the target.
4313 * -ESRCH if there's no task to yield to.
4315 int __sched yield_to(struct task_struct *p, bool preempt)
4317 struct task_struct *curr = current;
4318 struct rq *rq, *p_rq;
4319 unsigned long flags;
4322 local_irq_save(flags);
4328 * If we're the only runnable task on the rq and target rq also
4329 * has only one task, there's absolutely no point in yielding.
4331 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4336 double_rq_lock(rq, p_rq);
4337 if (task_rq(p) != p_rq) {
4338 double_rq_unlock(rq, p_rq);
4342 if (!curr->sched_class->yield_to_task)
4345 if (curr->sched_class != p->sched_class)
4348 if (task_running(p_rq, p) || p->state)
4351 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4353 schedstat_inc(rq, yld_count);
4355 * Make p's CPU reschedule; pick_next_entity takes care of
4358 if (preempt && rq != p_rq)
4363 double_rq_unlock(rq, p_rq);
4365 local_irq_restore(flags);
4372 EXPORT_SYMBOL_GPL(yield_to);
4375 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4376 * that process accounting knows that this is a task in IO wait state.
4378 long __sched io_schedule_timeout(long timeout)
4380 int old_iowait = current->in_iowait;
4384 current->in_iowait = 1;
4386 blk_schedule_flush_plug(current);
4388 blk_flush_plug(current);
4390 delayacct_blkio_start();
4392 atomic_inc(&rq->nr_iowait);
4393 ret = schedule_timeout(timeout);
4394 current->in_iowait = old_iowait;
4395 atomic_dec(&rq->nr_iowait);
4396 delayacct_blkio_end();
4400 EXPORT_SYMBOL(io_schedule_timeout);
4403 * sys_sched_get_priority_max - return maximum RT priority.
4404 * @policy: scheduling class.
4406 * Return: On success, this syscall returns the maximum
4407 * rt_priority that can be used by a given scheduling class.
4408 * On failure, a negative error code is returned.
4410 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4417 ret = MAX_USER_RT_PRIO-1;
4419 case SCHED_DEADLINE:
4430 * sys_sched_get_priority_min - return minimum RT priority.
4431 * @policy: scheduling class.
4433 * Return: On success, this syscall returns the minimum
4434 * rt_priority that can be used by a given scheduling class.
4435 * On failure, a negative error code is returned.
4437 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4446 case SCHED_DEADLINE:
4456 * sys_sched_rr_get_interval - return the default timeslice of a process.
4457 * @pid: pid of the process.
4458 * @interval: userspace pointer to the timeslice value.
4460 * this syscall writes the default timeslice value of a given process
4461 * into the user-space timespec buffer. A value of '0' means infinity.
4463 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4466 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4467 struct timespec __user *, interval)
4469 struct task_struct *p;
4470 unsigned int time_slice;
4471 unsigned long flags;
4481 p = find_process_by_pid(pid);
4485 retval = security_task_getscheduler(p);
4489 rq = task_rq_lock(p, &flags);
4491 if (p->sched_class->get_rr_interval)
4492 time_slice = p->sched_class->get_rr_interval(rq, p);
4493 task_rq_unlock(rq, p, &flags);
4496 jiffies_to_timespec(time_slice, &t);
4497 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4505 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4507 void sched_show_task(struct task_struct *p)
4509 unsigned long free = 0;
4511 unsigned long state = p->state;
4514 state = __ffs(state) + 1;
4515 printk(KERN_INFO "%-15.15s %c", p->comm,
4516 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4517 #if BITS_PER_LONG == 32
4518 if (state == TASK_RUNNING)
4519 printk(KERN_CONT " running ");
4521 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4523 if (state == TASK_RUNNING)
4524 printk(KERN_CONT " running task ");
4526 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4528 #ifdef CONFIG_DEBUG_STACK_USAGE
4529 free = stack_not_used(p);
4534 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4536 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4537 task_pid_nr(p), ppid,
4538 (unsigned long)task_thread_info(p)->flags);
4540 print_worker_info(KERN_INFO, p);
4541 show_stack(p, NULL);
4544 void show_state_filter(unsigned long state_filter)
4546 struct task_struct *g, *p;
4548 #if BITS_PER_LONG == 32
4550 " task PC stack pid father\n");
4553 " task PC stack pid father\n");
4556 for_each_process_thread(g, p) {
4558 * reset the NMI-timeout, listing all files on a slow
4559 * console might take a lot of time:
4561 touch_nmi_watchdog();
4562 if (!state_filter || (p->state & state_filter))
4566 touch_all_softlockup_watchdogs();
4568 #ifdef CONFIG_SCHED_DEBUG
4569 sysrq_sched_debug_show();
4573 * Only show locks if all tasks are dumped:
4576 debug_show_all_locks();
4579 void init_idle_bootup_task(struct task_struct *idle)
4581 idle->sched_class = &idle_sched_class;
4585 * init_idle - set up an idle thread for a given CPU
4586 * @idle: task in question
4587 * @cpu: cpu the idle task belongs to
4589 * NOTE: this function does not set the idle thread's NEED_RESCHED
4590 * flag, to make booting more robust.
4592 void init_idle(struct task_struct *idle, int cpu)
4594 struct rq *rq = cpu_rq(cpu);
4595 unsigned long flags;
4597 raw_spin_lock_irqsave(&rq->lock, flags);
4599 __sched_fork(0, idle);
4600 idle->state = TASK_RUNNING;
4601 idle->se.exec_start = sched_clock();
4603 do_set_cpus_allowed(idle, cpumask_of(cpu));
4605 * We're having a chicken and egg problem, even though we are
4606 * holding rq->lock, the cpu isn't yet set to this cpu so the
4607 * lockdep check in task_group() will fail.
4609 * Similar case to sched_fork(). / Alternatively we could
4610 * use task_rq_lock() here and obtain the other rq->lock.
4615 __set_task_cpu(idle, cpu);
4618 rq->curr = rq->idle = idle;
4619 idle->on_rq = TASK_ON_RQ_QUEUED;
4620 #if defined(CONFIG_SMP)
4623 raw_spin_unlock_irqrestore(&rq->lock, flags);
4625 /* Set the preempt count _outside_ the spinlocks! */
4626 init_idle_preempt_count(idle, cpu);
4629 * The idle tasks have their own, simple scheduling class:
4631 idle->sched_class = &idle_sched_class;
4632 ftrace_graph_init_idle_task(idle, cpu);
4633 vtime_init_idle(idle, cpu);
4634 #if defined(CONFIG_SMP)
4635 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4639 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4640 const struct cpumask *trial)
4642 int ret = 1, trial_cpus;
4643 struct dl_bw *cur_dl_b;
4644 unsigned long flags;
4646 if (!cpumask_weight(cur))
4649 rcu_read_lock_sched();
4650 cur_dl_b = dl_bw_of(cpumask_any(cur));
4651 trial_cpus = cpumask_weight(trial);
4653 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4654 if (cur_dl_b->bw != -1 &&
4655 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4657 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4658 rcu_read_unlock_sched();
4663 int task_can_attach(struct task_struct *p,
4664 const struct cpumask *cs_cpus_allowed)
4669 * Kthreads which disallow setaffinity shouldn't be moved
4670 * to a new cpuset; we don't want to change their cpu
4671 * affinity and isolating such threads by their set of
4672 * allowed nodes is unnecessary. Thus, cpusets are not
4673 * applicable for such threads. This prevents checking for
4674 * success of set_cpus_allowed_ptr() on all attached tasks
4675 * before cpus_allowed may be changed.
4677 if (p->flags & PF_NO_SETAFFINITY) {
4683 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4685 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4690 unsigned long flags;
4692 rcu_read_lock_sched();
4693 dl_b = dl_bw_of(dest_cpu);
4694 raw_spin_lock_irqsave(&dl_b->lock, flags);
4695 cpus = dl_bw_cpus(dest_cpu);
4696 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4701 * We reserve space for this task in the destination
4702 * root_domain, as we can't fail after this point.
4703 * We will free resources in the source root_domain
4704 * later on (see set_cpus_allowed_dl()).
4706 __dl_add(dl_b, p->dl.dl_bw);
4708 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4709 rcu_read_unlock_sched();
4719 * move_queued_task - move a queued task to new rq.
4721 * Returns (locked) new rq. Old rq's lock is released.
4723 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4725 struct rq *rq = task_rq(p);
4727 lockdep_assert_held(&rq->lock);
4729 dequeue_task(rq, p, 0);
4730 p->on_rq = TASK_ON_RQ_MIGRATING;
4731 set_task_cpu(p, new_cpu);
4732 raw_spin_unlock(&rq->lock);
4734 rq = cpu_rq(new_cpu);
4736 raw_spin_lock(&rq->lock);
4737 BUG_ON(task_cpu(p) != new_cpu);
4738 p->on_rq = TASK_ON_RQ_QUEUED;
4739 enqueue_task(rq, p, 0);
4740 check_preempt_curr(rq, p, 0);
4745 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4747 if (p->sched_class->set_cpus_allowed)
4748 p->sched_class->set_cpus_allowed(p, new_mask);
4750 cpumask_copy(&p->cpus_allowed, new_mask);
4751 p->nr_cpus_allowed = cpumask_weight(new_mask);
4755 * This is how migration works:
4757 * 1) we invoke migration_cpu_stop() on the target CPU using
4759 * 2) stopper starts to run (implicitly forcing the migrated thread
4761 * 3) it checks whether the migrated task is still in the wrong runqueue.
4762 * 4) if it's in the wrong runqueue then the migration thread removes
4763 * it and puts it into the right queue.
4764 * 5) stopper completes and stop_one_cpu() returns and the migration
4769 * Change a given task's CPU affinity. Migrate the thread to a
4770 * proper CPU and schedule it away if the CPU it's executing on
4771 * is removed from the allowed bitmask.
4773 * NOTE: the caller must have a valid reference to the task, the
4774 * task must not exit() & deallocate itself prematurely. The
4775 * call is not atomic; no spinlocks may be held.
4777 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4779 unsigned long flags;
4781 unsigned int dest_cpu;
4784 rq = task_rq_lock(p, &flags);
4786 if (cpumask_equal(&p->cpus_allowed, new_mask))
4789 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4794 do_set_cpus_allowed(p, new_mask);
4796 /* Can the task run on the task's current CPU? If so, we're done */
4797 if (cpumask_test_cpu(task_cpu(p), new_mask))
4800 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4801 if (task_running(rq, p) || p->state == TASK_WAKING) {
4802 struct migration_arg arg = { p, dest_cpu };
4803 /* Need help from migration thread: drop lock and wait. */
4804 task_rq_unlock(rq, p, &flags);
4805 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4806 tlb_migrate_finish(p->mm);
4808 } else if (task_on_rq_queued(p))
4809 rq = move_queued_task(p, dest_cpu);
4811 task_rq_unlock(rq, p, &flags);
4815 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4818 * Move (not current) task off this cpu, onto dest cpu. We're doing
4819 * this because either it can't run here any more (set_cpus_allowed()
4820 * away from this CPU, or CPU going down), or because we're
4821 * attempting to rebalance this task on exec (sched_exec).
4823 * So we race with normal scheduler movements, but that's OK, as long
4824 * as the task is no longer on this CPU.
4826 * Returns non-zero if task was successfully migrated.
4828 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4833 if (unlikely(!cpu_active(dest_cpu)))
4836 rq = cpu_rq(src_cpu);
4838 raw_spin_lock(&p->pi_lock);
4839 raw_spin_lock(&rq->lock);
4840 /* Already moved. */
4841 if (task_cpu(p) != src_cpu)
4844 /* Affinity changed (again). */
4845 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4849 * If we're not on a rq, the next wake-up will ensure we're
4852 if (task_on_rq_queued(p))
4853 rq = move_queued_task(p, dest_cpu);
4857 raw_spin_unlock(&rq->lock);
4858 raw_spin_unlock(&p->pi_lock);
4862 #ifdef CONFIG_NUMA_BALANCING
4863 /* Migrate current task p to target_cpu */
4864 int migrate_task_to(struct task_struct *p, int target_cpu)
4866 struct migration_arg arg = { p, target_cpu };
4867 int curr_cpu = task_cpu(p);
4869 if (curr_cpu == target_cpu)
4872 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4875 /* TODO: This is not properly updating schedstats */
4877 trace_sched_move_numa(p, curr_cpu, target_cpu);
4878 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4882 * Requeue a task on a given node and accurately track the number of NUMA
4883 * tasks on the runqueues
4885 void sched_setnuma(struct task_struct *p, int nid)
4888 unsigned long flags;
4889 bool queued, running;
4891 rq = task_rq_lock(p, &flags);
4892 queued = task_on_rq_queued(p);
4893 running = task_current(rq, p);
4896 dequeue_task(rq, p, 0);
4898 put_prev_task(rq, p);
4900 p->numa_preferred_nid = nid;
4903 p->sched_class->set_curr_task(rq);
4905 enqueue_task(rq, p, 0);
4906 task_rq_unlock(rq, p, &flags);
4911 * migration_cpu_stop - this will be executed by a highprio stopper thread
4912 * and performs thread migration by bumping thread off CPU then
4913 * 'pushing' onto another runqueue.
4915 static int migration_cpu_stop(void *data)
4917 struct migration_arg *arg = data;
4920 * The original target cpu might have gone down and we might
4921 * be on another cpu but it doesn't matter.
4923 local_irq_disable();
4925 * We need to explicitly wake pending tasks before running
4926 * __migrate_task() such that we will not miss enforcing cpus_allowed
4927 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4929 sched_ttwu_pending();
4930 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4935 #ifdef CONFIG_HOTPLUG_CPU
4938 * Ensures that the idle task is using init_mm right before its cpu goes
4941 void idle_task_exit(void)
4943 struct mm_struct *mm = current->active_mm;
4945 BUG_ON(cpu_online(smp_processor_id()));
4947 if (mm != &init_mm) {
4948 switch_mm(mm, &init_mm, current);
4949 finish_arch_post_lock_switch();
4955 * Since this CPU is going 'away' for a while, fold any nr_active delta
4956 * we might have. Assumes we're called after migrate_tasks() so that the
4957 * nr_active count is stable.
4959 * Also see the comment "Global load-average calculations".
4961 static void calc_load_migrate(struct rq *rq)
4963 long delta = calc_load_fold_active(rq);
4965 atomic_long_add(delta, &calc_load_tasks);
4968 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4972 static const struct sched_class fake_sched_class = {
4973 .put_prev_task = put_prev_task_fake,
4976 static struct task_struct fake_task = {
4978 * Avoid pull_{rt,dl}_task()
4980 .prio = MAX_PRIO + 1,
4981 .sched_class = &fake_sched_class,
4985 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4986 * try_to_wake_up()->select_task_rq().
4988 * Called with rq->lock held even though we'er in stop_machine() and
4989 * there's no concurrency possible, we hold the required locks anyway
4990 * because of lock validation efforts.
4992 static void migrate_tasks(unsigned int dead_cpu)
4994 struct rq *rq = cpu_rq(dead_cpu);
4995 struct task_struct *next, *stop = rq->stop;
4999 * Fudge the rq selection such that the below task selection loop
5000 * doesn't get stuck on the currently eligible stop task.
5002 * We're currently inside stop_machine() and the rq is either stuck
5003 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5004 * either way we should never end up calling schedule() until we're
5010 * put_prev_task() and pick_next_task() sched
5011 * class method both need to have an up-to-date
5012 * value of rq->clock[_task]
5014 update_rq_clock(rq);
5018 * There's this thread running, bail when that's the only
5021 if (rq->nr_running == 1)
5024 next = pick_next_task(rq, &fake_task);
5026 next->sched_class->put_prev_task(rq, next);
5028 /* Find suitable destination for @next, with force if needed. */
5029 dest_cpu = select_fallback_rq(dead_cpu, next);
5030 raw_spin_unlock(&rq->lock);
5032 __migrate_task(next, dead_cpu, dest_cpu);
5034 raw_spin_lock(&rq->lock);
5040 #endif /* CONFIG_HOTPLUG_CPU */
5042 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5044 static struct ctl_table sd_ctl_dir[] = {
5046 .procname = "sched_domain",
5052 static struct ctl_table sd_ctl_root[] = {
5054 .procname = "kernel",
5056 .child = sd_ctl_dir,
5061 static struct ctl_table *sd_alloc_ctl_entry(int n)
5063 struct ctl_table *entry =
5064 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5069 static void sd_free_ctl_entry(struct ctl_table **tablep)
5071 struct ctl_table *entry;
5074 * In the intermediate directories, both the child directory and
5075 * procname are dynamically allocated and could fail but the mode
5076 * will always be set. In the lowest directory the names are
5077 * static strings and all have proc handlers.
5079 for (entry = *tablep; entry->mode; entry++) {
5081 sd_free_ctl_entry(&entry->child);
5082 if (entry->proc_handler == NULL)
5083 kfree(entry->procname);
5090 static int min_load_idx = 0;
5091 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5094 set_table_entry(struct ctl_table *entry,
5095 const char *procname, void *data, int maxlen,
5096 umode_t mode, proc_handler *proc_handler,
5099 entry->procname = procname;
5101 entry->maxlen = maxlen;
5103 entry->proc_handler = proc_handler;
5106 entry->extra1 = &min_load_idx;
5107 entry->extra2 = &max_load_idx;
5111 static struct ctl_table *
5112 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5114 struct ctl_table *table = sd_alloc_ctl_entry(14);
5119 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5120 sizeof(long), 0644, proc_doulongvec_minmax, false);
5121 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5122 sizeof(long), 0644, proc_doulongvec_minmax, false);
5123 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5124 sizeof(int), 0644, proc_dointvec_minmax, true);
5125 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5126 sizeof(int), 0644, proc_dointvec_minmax, true);
5127 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5128 sizeof(int), 0644, proc_dointvec_minmax, true);
5129 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5130 sizeof(int), 0644, proc_dointvec_minmax, true);
5131 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5132 sizeof(int), 0644, proc_dointvec_minmax, true);
5133 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5134 sizeof(int), 0644, proc_dointvec_minmax, false);
5135 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5136 sizeof(int), 0644, proc_dointvec_minmax, false);
5137 set_table_entry(&table[9], "cache_nice_tries",
5138 &sd->cache_nice_tries,
5139 sizeof(int), 0644, proc_dointvec_minmax, false);
5140 set_table_entry(&table[10], "flags", &sd->flags,
5141 sizeof(int), 0644, proc_dointvec_minmax, false);
5142 set_table_entry(&table[11], "max_newidle_lb_cost",
5143 &sd->max_newidle_lb_cost,
5144 sizeof(long), 0644, proc_doulongvec_minmax, false);
5145 set_table_entry(&table[12], "name", sd->name,
5146 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5147 /* &table[13] is terminator */
5152 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5154 struct ctl_table *entry, *table;
5155 struct sched_domain *sd;
5156 int domain_num = 0, i;
5159 for_each_domain(cpu, sd)
5161 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5166 for_each_domain(cpu, sd) {
5167 snprintf(buf, 32, "domain%d", i);
5168 entry->procname = kstrdup(buf, GFP_KERNEL);
5170 entry->child = sd_alloc_ctl_domain_table(sd);
5177 static struct ctl_table_header *sd_sysctl_header;
5178 static void register_sched_domain_sysctl(void)
5180 int i, cpu_num = num_possible_cpus();
5181 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5184 WARN_ON(sd_ctl_dir[0].child);
5185 sd_ctl_dir[0].child = entry;
5190 for_each_possible_cpu(i) {
5191 snprintf(buf, 32, "cpu%d", i);
5192 entry->procname = kstrdup(buf, GFP_KERNEL);
5194 entry->child = sd_alloc_ctl_cpu_table(i);
5198 WARN_ON(sd_sysctl_header);
5199 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5202 /* may be called multiple times per register */
5203 static void unregister_sched_domain_sysctl(void)
5205 if (sd_sysctl_header)
5206 unregister_sysctl_table(sd_sysctl_header);
5207 sd_sysctl_header = NULL;
5208 if (sd_ctl_dir[0].child)
5209 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5212 static void register_sched_domain_sysctl(void)
5215 static void unregister_sched_domain_sysctl(void)
5220 static void set_rq_online(struct rq *rq)
5223 const struct sched_class *class;
5225 cpumask_set_cpu(rq->cpu, rq->rd->online);
5228 for_each_class(class) {
5229 if (class->rq_online)
5230 class->rq_online(rq);
5235 static void set_rq_offline(struct rq *rq)
5238 const struct sched_class *class;
5240 for_each_class(class) {
5241 if (class->rq_offline)
5242 class->rq_offline(rq);
5245 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5251 * migration_call - callback that gets triggered when a CPU is added.
5252 * Here we can start up the necessary migration thread for the new CPU.
5255 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5257 int cpu = (long)hcpu;
5258 unsigned long flags;
5259 struct rq *rq = cpu_rq(cpu);
5261 switch (action & ~CPU_TASKS_FROZEN) {
5263 case CPU_UP_PREPARE:
5264 rq->calc_load_update = calc_load_update;
5268 /* Update our root-domain */
5269 raw_spin_lock_irqsave(&rq->lock, flags);
5271 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5275 raw_spin_unlock_irqrestore(&rq->lock, flags);
5278 #ifdef CONFIG_HOTPLUG_CPU
5280 sched_ttwu_pending();
5281 /* Update our root-domain */
5282 raw_spin_lock_irqsave(&rq->lock, flags);
5284 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5288 BUG_ON(rq->nr_running != 1); /* the migration thread */
5289 raw_spin_unlock_irqrestore(&rq->lock, flags);
5293 calc_load_migrate(rq);
5298 update_max_interval();
5304 * Register at high priority so that task migration (migrate_all_tasks)
5305 * happens before everything else. This has to be lower priority than
5306 * the notifier in the perf_event subsystem, though.
5308 static struct notifier_block migration_notifier = {
5309 .notifier_call = migration_call,
5310 .priority = CPU_PRI_MIGRATION,
5313 static void __cpuinit set_cpu_rq_start_time(void)
5315 int cpu = smp_processor_id();
5316 struct rq *rq = cpu_rq(cpu);
5317 rq->age_stamp = sched_clock_cpu(cpu);
5320 static int sched_cpu_active(struct notifier_block *nfb,
5321 unsigned long action, void *hcpu)
5323 switch (action & ~CPU_TASKS_FROZEN) {
5325 set_cpu_rq_start_time();
5327 case CPU_DOWN_FAILED:
5328 set_cpu_active((long)hcpu, true);
5335 static int sched_cpu_inactive(struct notifier_block *nfb,
5336 unsigned long action, void *hcpu)
5338 unsigned long flags;
5339 long cpu = (long)hcpu;
5342 switch (action & ~CPU_TASKS_FROZEN) {
5343 case CPU_DOWN_PREPARE:
5344 set_cpu_active(cpu, false);
5346 /* explicitly allow suspend */
5347 if (!(action & CPU_TASKS_FROZEN)) {
5351 rcu_read_lock_sched();
5352 dl_b = dl_bw_of(cpu);
5354 raw_spin_lock_irqsave(&dl_b->lock, flags);
5355 cpus = dl_bw_cpus(cpu);
5356 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5357 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5359 rcu_read_unlock_sched();
5362 return notifier_from_errno(-EBUSY);
5370 static int __init migration_init(void)
5372 void *cpu = (void *)(long)smp_processor_id();
5375 /* Initialize migration for the boot CPU */
5376 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5377 BUG_ON(err == NOTIFY_BAD);
5378 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5379 register_cpu_notifier(&migration_notifier);
5381 /* Register cpu active notifiers */
5382 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5383 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5387 early_initcall(migration_init);
5392 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5394 #ifdef CONFIG_SCHED_DEBUG
5396 static __read_mostly int sched_debug_enabled;
5398 static int __init sched_debug_setup(char *str)
5400 sched_debug_enabled = 1;
5404 early_param("sched_debug", sched_debug_setup);
5406 static inline bool sched_debug(void)
5408 return sched_debug_enabled;
5411 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5412 struct cpumask *groupmask)
5414 struct sched_group *group = sd->groups;
5416 cpumask_clear(groupmask);
5418 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5420 if (!(sd->flags & SD_LOAD_BALANCE)) {
5421 printk("does not load-balance\n");
5423 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5428 printk(KERN_CONT "span %*pbl level %s\n",
5429 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5431 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5432 printk(KERN_ERR "ERROR: domain->span does not contain "
5435 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5436 printk(KERN_ERR "ERROR: domain->groups does not contain"
5440 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5444 printk(KERN_ERR "ERROR: group is NULL\n");
5449 * Even though we initialize ->capacity to something semi-sane,
5450 * we leave capacity_orig unset. This allows us to detect if
5451 * domain iteration is still funny without causing /0 traps.
5453 if (!group->sgc->capacity_orig) {
5454 printk(KERN_CONT "\n");
5455 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5459 if (!cpumask_weight(sched_group_cpus(group))) {
5460 printk(KERN_CONT "\n");
5461 printk(KERN_ERR "ERROR: empty group\n");
5465 if (!(sd->flags & SD_OVERLAP) &&
5466 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5467 printk(KERN_CONT "\n");
5468 printk(KERN_ERR "ERROR: repeated CPUs\n");
5472 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5474 printk(KERN_CONT " %*pbl",
5475 cpumask_pr_args(sched_group_cpus(group)));
5476 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5477 printk(KERN_CONT " (cpu_capacity = %d)",
5478 group->sgc->capacity);
5481 group = group->next;
5482 } while (group != sd->groups);
5483 printk(KERN_CONT "\n");
5485 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5486 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5489 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5490 printk(KERN_ERR "ERROR: parent span is not a superset "
5491 "of domain->span\n");
5495 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5499 if (!sched_debug_enabled)
5503 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5507 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5510 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5518 #else /* !CONFIG_SCHED_DEBUG */
5519 # define sched_domain_debug(sd, cpu) do { } while (0)
5520 static inline bool sched_debug(void)
5524 #endif /* CONFIG_SCHED_DEBUG */
5526 static int sd_degenerate(struct sched_domain *sd)
5528 if (cpumask_weight(sched_domain_span(sd)) == 1)
5531 /* Following flags need at least 2 groups */
5532 if (sd->flags & (SD_LOAD_BALANCE |
5533 SD_BALANCE_NEWIDLE |
5536 SD_SHARE_CPUCAPACITY |
5537 SD_SHARE_PKG_RESOURCES |
5538 SD_SHARE_POWERDOMAIN)) {
5539 if (sd->groups != sd->groups->next)
5543 /* Following flags don't use groups */
5544 if (sd->flags & (SD_WAKE_AFFINE))
5551 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5553 unsigned long cflags = sd->flags, pflags = parent->flags;
5555 if (sd_degenerate(parent))
5558 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5561 /* Flags needing groups don't count if only 1 group in parent */
5562 if (parent->groups == parent->groups->next) {
5563 pflags &= ~(SD_LOAD_BALANCE |
5564 SD_BALANCE_NEWIDLE |
5567 SD_SHARE_CPUCAPACITY |
5568 SD_SHARE_PKG_RESOURCES |
5570 SD_SHARE_POWERDOMAIN);
5571 if (nr_node_ids == 1)
5572 pflags &= ~SD_SERIALIZE;
5574 if (~cflags & pflags)
5580 static void free_rootdomain(struct rcu_head *rcu)
5582 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5584 cpupri_cleanup(&rd->cpupri);
5585 cpudl_cleanup(&rd->cpudl);
5586 free_cpumask_var(rd->dlo_mask);
5587 free_cpumask_var(rd->rto_mask);
5588 free_cpumask_var(rd->online);
5589 free_cpumask_var(rd->span);
5593 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5595 struct root_domain *old_rd = NULL;
5596 unsigned long flags;
5598 raw_spin_lock_irqsave(&rq->lock, flags);
5603 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5606 cpumask_clear_cpu(rq->cpu, old_rd->span);
5609 * If we dont want to free the old_rd yet then
5610 * set old_rd to NULL to skip the freeing later
5613 if (!atomic_dec_and_test(&old_rd->refcount))
5617 atomic_inc(&rd->refcount);
5620 cpumask_set_cpu(rq->cpu, rd->span);
5621 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5624 raw_spin_unlock_irqrestore(&rq->lock, flags);
5627 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5630 static int init_rootdomain(struct root_domain *rd)
5632 memset(rd, 0, sizeof(*rd));
5634 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5636 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5638 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5640 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5643 init_dl_bw(&rd->dl_bw);
5644 if (cpudl_init(&rd->cpudl) != 0)
5647 if (cpupri_init(&rd->cpupri) != 0)
5652 free_cpumask_var(rd->rto_mask);
5654 free_cpumask_var(rd->dlo_mask);
5656 free_cpumask_var(rd->online);
5658 free_cpumask_var(rd->span);
5664 * By default the system creates a single root-domain with all cpus as
5665 * members (mimicking the global state we have today).
5667 struct root_domain def_root_domain;
5669 static void init_defrootdomain(void)
5671 init_rootdomain(&def_root_domain);
5673 atomic_set(&def_root_domain.refcount, 1);
5676 static struct root_domain *alloc_rootdomain(void)
5678 struct root_domain *rd;
5680 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5684 if (init_rootdomain(rd) != 0) {
5692 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5694 struct sched_group *tmp, *first;
5703 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5708 } while (sg != first);
5711 static void free_sched_domain(struct rcu_head *rcu)
5713 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5716 * If its an overlapping domain it has private groups, iterate and
5719 if (sd->flags & SD_OVERLAP) {
5720 free_sched_groups(sd->groups, 1);
5721 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5722 kfree(sd->groups->sgc);
5728 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5730 call_rcu(&sd->rcu, free_sched_domain);
5733 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5735 for (; sd; sd = sd->parent)
5736 destroy_sched_domain(sd, cpu);
5740 * Keep a special pointer to the highest sched_domain that has
5741 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5742 * allows us to avoid some pointer chasing select_idle_sibling().
5744 * Also keep a unique ID per domain (we use the first cpu number in
5745 * the cpumask of the domain), this allows us to quickly tell if
5746 * two cpus are in the same cache domain, see cpus_share_cache().
5748 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5749 DEFINE_PER_CPU(int, sd_llc_size);
5750 DEFINE_PER_CPU(int, sd_llc_id);
5751 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5752 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5753 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5755 static void update_top_cache_domain(int cpu)
5757 struct sched_domain *sd;
5758 struct sched_domain *busy_sd = NULL;
5762 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5764 id = cpumask_first(sched_domain_span(sd));
5765 size = cpumask_weight(sched_domain_span(sd));
5766 busy_sd = sd->parent; /* sd_busy */
5768 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5770 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5771 per_cpu(sd_llc_size, cpu) = size;
5772 per_cpu(sd_llc_id, cpu) = id;
5774 sd = lowest_flag_domain(cpu, SD_NUMA);
5775 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5777 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5778 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5782 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5783 * hold the hotplug lock.
5786 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5788 struct rq *rq = cpu_rq(cpu);
5789 struct sched_domain *tmp;
5791 /* Remove the sched domains which do not contribute to scheduling. */
5792 for (tmp = sd; tmp; ) {
5793 struct sched_domain *parent = tmp->parent;
5797 if (sd_parent_degenerate(tmp, parent)) {
5798 tmp->parent = parent->parent;
5800 parent->parent->child = tmp;
5802 * Transfer SD_PREFER_SIBLING down in case of a
5803 * degenerate parent; the spans match for this
5804 * so the property transfers.
5806 if (parent->flags & SD_PREFER_SIBLING)
5807 tmp->flags |= SD_PREFER_SIBLING;
5808 destroy_sched_domain(parent, cpu);
5813 if (sd && sd_degenerate(sd)) {
5816 destroy_sched_domain(tmp, cpu);
5821 sched_domain_debug(sd, cpu);
5823 rq_attach_root(rq, rd);
5825 rcu_assign_pointer(rq->sd, sd);
5826 destroy_sched_domains(tmp, cpu);
5828 update_top_cache_domain(cpu);
5831 /* cpus with isolated domains */
5832 static cpumask_var_t cpu_isolated_map;
5834 /* Setup the mask of cpus configured for isolated domains */
5835 static int __init isolated_cpu_setup(char *str)
5837 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5838 cpulist_parse(str, cpu_isolated_map);
5842 __setup("isolcpus=", isolated_cpu_setup);
5845 struct sched_domain ** __percpu sd;
5846 struct root_domain *rd;
5857 * Build an iteration mask that can exclude certain CPUs from the upwards
5860 * Asymmetric node setups can result in situations where the domain tree is of
5861 * unequal depth, make sure to skip domains that already cover the entire
5864 * In that case build_sched_domains() will have terminated the iteration early
5865 * and our sibling sd spans will be empty. Domains should always include the
5866 * cpu they're built on, so check that.
5869 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5871 const struct cpumask *span = sched_domain_span(sd);
5872 struct sd_data *sdd = sd->private;
5873 struct sched_domain *sibling;
5876 for_each_cpu(i, span) {
5877 sibling = *per_cpu_ptr(sdd->sd, i);
5878 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5881 cpumask_set_cpu(i, sched_group_mask(sg));
5886 * Return the canonical balance cpu for this group, this is the first cpu
5887 * of this group that's also in the iteration mask.
5889 int group_balance_cpu(struct sched_group *sg)
5891 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5895 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5897 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5898 const struct cpumask *span = sched_domain_span(sd);
5899 struct cpumask *covered = sched_domains_tmpmask;
5900 struct sd_data *sdd = sd->private;
5901 struct sched_domain *sibling;
5904 cpumask_clear(covered);
5906 for_each_cpu(i, span) {
5907 struct cpumask *sg_span;
5909 if (cpumask_test_cpu(i, covered))
5912 sibling = *per_cpu_ptr(sdd->sd, i);
5914 /* See the comment near build_group_mask(). */
5915 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5918 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5919 GFP_KERNEL, cpu_to_node(cpu));
5924 sg_span = sched_group_cpus(sg);
5926 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5928 cpumask_set_cpu(i, sg_span);
5930 cpumask_or(covered, covered, sg_span);
5932 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5933 if (atomic_inc_return(&sg->sgc->ref) == 1)
5934 build_group_mask(sd, sg);
5937 * Initialize sgc->capacity such that even if we mess up the
5938 * domains and no possible iteration will get us here, we won't
5941 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5942 sg->sgc->capacity_orig = sg->sgc->capacity;
5945 * Make sure the first group of this domain contains the
5946 * canonical balance cpu. Otherwise the sched_domain iteration
5947 * breaks. See update_sg_lb_stats().
5949 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5950 group_balance_cpu(sg) == cpu)
5960 sd->groups = groups;
5965 free_sched_groups(first, 0);
5970 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5972 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5973 struct sched_domain *child = sd->child;
5976 cpu = cpumask_first(sched_domain_span(child));
5979 *sg = *per_cpu_ptr(sdd->sg, cpu);
5980 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5981 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5988 * build_sched_groups will build a circular linked list of the groups
5989 * covered by the given span, and will set each group's ->cpumask correctly,
5990 * and ->cpu_capacity to 0.
5992 * Assumes the sched_domain tree is fully constructed
5995 build_sched_groups(struct sched_domain *sd, int cpu)
5997 struct sched_group *first = NULL, *last = NULL;
5998 struct sd_data *sdd = sd->private;
5999 const struct cpumask *span = sched_domain_span(sd);
6000 struct cpumask *covered;
6003 get_group(cpu, sdd, &sd->groups);
6004 atomic_inc(&sd->groups->ref);
6006 if (cpu != cpumask_first(span))
6009 lockdep_assert_held(&sched_domains_mutex);
6010 covered = sched_domains_tmpmask;
6012 cpumask_clear(covered);
6014 for_each_cpu(i, span) {
6015 struct sched_group *sg;
6018 if (cpumask_test_cpu(i, covered))
6021 group = get_group(i, sdd, &sg);
6022 cpumask_setall(sched_group_mask(sg));
6024 for_each_cpu(j, span) {
6025 if (get_group(j, sdd, NULL) != group)
6028 cpumask_set_cpu(j, covered);
6029 cpumask_set_cpu(j, sched_group_cpus(sg));
6044 * Initialize sched groups cpu_capacity.
6046 * cpu_capacity indicates the capacity of sched group, which is used while
6047 * distributing the load between different sched groups in a sched domain.
6048 * Typically cpu_capacity for all the groups in a sched domain will be same
6049 * unless there are asymmetries in the topology. If there are asymmetries,
6050 * group having more cpu_capacity will pickup more load compared to the
6051 * group having less cpu_capacity.
6053 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6055 struct sched_group *sg = sd->groups;
6060 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6062 } while (sg != sd->groups);
6064 if (cpu != group_balance_cpu(sg))
6067 update_group_capacity(sd, cpu);
6068 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6072 * Initializers for schedule domains
6073 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6076 static int default_relax_domain_level = -1;
6077 int sched_domain_level_max;
6079 static int __init setup_relax_domain_level(char *str)
6081 if (kstrtoint(str, 0, &default_relax_domain_level))
6082 pr_warn("Unable to set relax_domain_level\n");
6086 __setup("relax_domain_level=", setup_relax_domain_level);
6088 static void set_domain_attribute(struct sched_domain *sd,
6089 struct sched_domain_attr *attr)
6093 if (!attr || attr->relax_domain_level < 0) {
6094 if (default_relax_domain_level < 0)
6097 request = default_relax_domain_level;
6099 request = attr->relax_domain_level;
6100 if (request < sd->level) {
6101 /* turn off idle balance on this domain */
6102 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6104 /* turn on idle balance on this domain */
6105 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6109 static void __sdt_free(const struct cpumask *cpu_map);
6110 static int __sdt_alloc(const struct cpumask *cpu_map);
6112 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6113 const struct cpumask *cpu_map)
6117 if (!atomic_read(&d->rd->refcount))
6118 free_rootdomain(&d->rd->rcu); /* fall through */
6120 free_percpu(d->sd); /* fall through */
6122 __sdt_free(cpu_map); /* fall through */
6128 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6129 const struct cpumask *cpu_map)
6131 memset(d, 0, sizeof(*d));
6133 if (__sdt_alloc(cpu_map))
6134 return sa_sd_storage;
6135 d->sd = alloc_percpu(struct sched_domain *);
6137 return sa_sd_storage;
6138 d->rd = alloc_rootdomain();
6141 return sa_rootdomain;
6145 * NULL the sd_data elements we've used to build the sched_domain and
6146 * sched_group structure so that the subsequent __free_domain_allocs()
6147 * will not free the data we're using.
6149 static void claim_allocations(int cpu, struct sched_domain *sd)
6151 struct sd_data *sdd = sd->private;
6153 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6154 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6156 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6157 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6159 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6160 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6164 static int sched_domains_numa_levels;
6165 enum numa_topology_type sched_numa_topology_type;
6166 static int *sched_domains_numa_distance;
6167 int sched_max_numa_distance;
6168 static struct cpumask ***sched_domains_numa_masks;
6169 static int sched_domains_curr_level;
6173 * SD_flags allowed in topology descriptions.
6175 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6176 * SD_SHARE_PKG_RESOURCES - describes shared caches
6177 * SD_NUMA - describes NUMA topologies
6178 * SD_SHARE_POWERDOMAIN - describes shared power domain
6181 * SD_ASYM_PACKING - describes SMT quirks
6183 #define TOPOLOGY_SD_FLAGS \
6184 (SD_SHARE_CPUCAPACITY | \
6185 SD_SHARE_PKG_RESOURCES | \
6188 SD_SHARE_POWERDOMAIN)
6190 static struct sched_domain *
6191 sd_init(struct sched_domain_topology_level *tl, int cpu)
6193 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6194 int sd_weight, sd_flags = 0;
6198 * Ugly hack to pass state to sd_numa_mask()...
6200 sched_domains_curr_level = tl->numa_level;
6203 sd_weight = cpumask_weight(tl->mask(cpu));
6206 sd_flags = (*tl->sd_flags)();
6207 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6208 "wrong sd_flags in topology description\n"))
6209 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6211 *sd = (struct sched_domain){
6212 .min_interval = sd_weight,
6213 .max_interval = 2*sd_weight,
6215 .imbalance_pct = 125,
6217 .cache_nice_tries = 0,
6224 .flags = 1*SD_LOAD_BALANCE
6225 | 1*SD_BALANCE_NEWIDLE
6230 | 0*SD_SHARE_CPUCAPACITY
6231 | 0*SD_SHARE_PKG_RESOURCES
6233 | 0*SD_PREFER_SIBLING
6238 .last_balance = jiffies,
6239 .balance_interval = sd_weight,
6241 .max_newidle_lb_cost = 0,
6242 .next_decay_max_lb_cost = jiffies,
6243 #ifdef CONFIG_SCHED_DEBUG
6249 * Convert topological properties into behaviour.
6252 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6253 sd->imbalance_pct = 110;
6254 sd->smt_gain = 1178; /* ~15% */
6256 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6257 sd->imbalance_pct = 117;
6258 sd->cache_nice_tries = 1;
6262 } else if (sd->flags & SD_NUMA) {
6263 sd->cache_nice_tries = 2;
6267 sd->flags |= SD_SERIALIZE;
6268 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6269 sd->flags &= ~(SD_BALANCE_EXEC |
6276 sd->flags |= SD_PREFER_SIBLING;
6277 sd->cache_nice_tries = 1;
6282 sd->private = &tl->data;
6288 * Topology list, bottom-up.
6290 static struct sched_domain_topology_level default_topology[] = {
6291 #ifdef CONFIG_SCHED_SMT
6292 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6294 #ifdef CONFIG_SCHED_MC
6295 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6297 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6301 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6303 #define for_each_sd_topology(tl) \
6304 for (tl = sched_domain_topology; tl->mask; tl++)
6306 void set_sched_topology(struct sched_domain_topology_level *tl)
6308 sched_domain_topology = tl;
6313 static const struct cpumask *sd_numa_mask(int cpu)
6315 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6318 static void sched_numa_warn(const char *str)
6320 static int done = false;
6328 printk(KERN_WARNING "ERROR: %s\n\n", str);
6330 for (i = 0; i < nr_node_ids; i++) {
6331 printk(KERN_WARNING " ");
6332 for (j = 0; j < nr_node_ids; j++)
6333 printk(KERN_CONT "%02d ", node_distance(i,j));
6334 printk(KERN_CONT "\n");
6336 printk(KERN_WARNING "\n");
6339 bool find_numa_distance(int distance)
6343 if (distance == node_distance(0, 0))
6346 for (i = 0; i < sched_domains_numa_levels; i++) {
6347 if (sched_domains_numa_distance[i] == distance)
6355 * A system can have three types of NUMA topology:
6356 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6357 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6358 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6360 * The difference between a glueless mesh topology and a backplane
6361 * topology lies in whether communication between not directly
6362 * connected nodes goes through intermediary nodes (where programs
6363 * could run), or through backplane controllers. This affects
6364 * placement of programs.
6366 * The type of topology can be discerned with the following tests:
6367 * - If the maximum distance between any nodes is 1 hop, the system
6368 * is directly connected.
6369 * - If for two nodes A and B, located N > 1 hops away from each other,
6370 * there is an intermediary node C, which is < N hops away from both
6371 * nodes A and B, the system is a glueless mesh.
6373 static void init_numa_topology_type(void)
6377 n = sched_max_numa_distance;
6380 sched_numa_topology_type = NUMA_DIRECT;
6382 for_each_online_node(a) {
6383 for_each_online_node(b) {
6384 /* Find two nodes furthest removed from each other. */
6385 if (node_distance(a, b) < n)
6388 /* Is there an intermediary node between a and b? */
6389 for_each_online_node(c) {
6390 if (node_distance(a, c) < n &&
6391 node_distance(b, c) < n) {
6392 sched_numa_topology_type =
6398 sched_numa_topology_type = NUMA_BACKPLANE;
6404 static void sched_init_numa(void)
6406 int next_distance, curr_distance = node_distance(0, 0);
6407 struct sched_domain_topology_level *tl;
6411 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6412 if (!sched_domains_numa_distance)
6416 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6417 * unique distances in the node_distance() table.
6419 * Assumes node_distance(0,j) includes all distances in
6420 * node_distance(i,j) in order to avoid cubic time.
6422 next_distance = curr_distance;
6423 for (i = 0; i < nr_node_ids; i++) {
6424 for (j = 0; j < nr_node_ids; j++) {
6425 for (k = 0; k < nr_node_ids; k++) {
6426 int distance = node_distance(i, k);
6428 if (distance > curr_distance &&
6429 (distance < next_distance ||
6430 next_distance == curr_distance))
6431 next_distance = distance;
6434 * While not a strong assumption it would be nice to know
6435 * about cases where if node A is connected to B, B is not
6436 * equally connected to A.
6438 if (sched_debug() && node_distance(k, i) != distance)
6439 sched_numa_warn("Node-distance not symmetric");
6441 if (sched_debug() && i && !find_numa_distance(distance))
6442 sched_numa_warn("Node-0 not representative");
6444 if (next_distance != curr_distance) {
6445 sched_domains_numa_distance[level++] = next_distance;
6446 sched_domains_numa_levels = level;
6447 curr_distance = next_distance;
6452 * In case of sched_debug() we verify the above assumption.
6462 * 'level' contains the number of unique distances, excluding the
6463 * identity distance node_distance(i,i).
6465 * The sched_domains_numa_distance[] array includes the actual distance
6470 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6471 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6472 * the array will contain less then 'level' members. This could be
6473 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6474 * in other functions.
6476 * We reset it to 'level' at the end of this function.
6478 sched_domains_numa_levels = 0;
6480 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6481 if (!sched_domains_numa_masks)
6485 * Now for each level, construct a mask per node which contains all
6486 * cpus of nodes that are that many hops away from us.
6488 for (i = 0; i < level; i++) {
6489 sched_domains_numa_masks[i] =
6490 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6491 if (!sched_domains_numa_masks[i])
6494 for (j = 0; j < nr_node_ids; j++) {
6495 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6499 sched_domains_numa_masks[i][j] = mask;
6501 for (k = 0; k < nr_node_ids; k++) {
6502 if (node_distance(j, k) > sched_domains_numa_distance[i])
6505 cpumask_or(mask, mask, cpumask_of_node(k));
6510 /* Compute default topology size */
6511 for (i = 0; sched_domain_topology[i].mask; i++);
6513 tl = kzalloc((i + level + 1) *
6514 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6519 * Copy the default topology bits..
6521 for (i = 0; sched_domain_topology[i].mask; i++)
6522 tl[i] = sched_domain_topology[i];
6525 * .. and append 'j' levels of NUMA goodness.
6527 for (j = 0; j < level; i++, j++) {
6528 tl[i] = (struct sched_domain_topology_level){
6529 .mask = sd_numa_mask,
6530 .sd_flags = cpu_numa_flags,
6531 .flags = SDTL_OVERLAP,
6537 sched_domain_topology = tl;
6539 sched_domains_numa_levels = level;
6540 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6542 init_numa_topology_type();
6545 static void sched_domains_numa_masks_set(int cpu)
6548 int node = cpu_to_node(cpu);
6550 for (i = 0; i < sched_domains_numa_levels; i++) {
6551 for (j = 0; j < nr_node_ids; j++) {
6552 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6553 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6558 static void sched_domains_numa_masks_clear(int cpu)
6561 for (i = 0; i < sched_domains_numa_levels; i++) {
6562 for (j = 0; j < nr_node_ids; j++)
6563 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6568 * Update sched_domains_numa_masks[level][node] array when new cpus
6571 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6572 unsigned long action,
6575 int cpu = (long)hcpu;
6577 switch (action & ~CPU_TASKS_FROZEN) {
6579 sched_domains_numa_masks_set(cpu);
6583 sched_domains_numa_masks_clear(cpu);
6593 static inline void sched_init_numa(void)
6597 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6598 unsigned long action,
6603 #endif /* CONFIG_NUMA */
6605 static int __sdt_alloc(const struct cpumask *cpu_map)
6607 struct sched_domain_topology_level *tl;
6610 for_each_sd_topology(tl) {
6611 struct sd_data *sdd = &tl->data;
6613 sdd->sd = alloc_percpu(struct sched_domain *);
6617 sdd->sg = alloc_percpu(struct sched_group *);
6621 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6625 for_each_cpu(j, cpu_map) {
6626 struct sched_domain *sd;
6627 struct sched_group *sg;
6628 struct sched_group_capacity *sgc;
6630 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6631 GFP_KERNEL, cpu_to_node(j));
6635 *per_cpu_ptr(sdd->sd, j) = sd;
6637 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6638 GFP_KERNEL, cpu_to_node(j));
6644 *per_cpu_ptr(sdd->sg, j) = sg;
6646 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6647 GFP_KERNEL, cpu_to_node(j));
6651 *per_cpu_ptr(sdd->sgc, j) = sgc;
6658 static void __sdt_free(const struct cpumask *cpu_map)
6660 struct sched_domain_topology_level *tl;
6663 for_each_sd_topology(tl) {
6664 struct sd_data *sdd = &tl->data;
6666 for_each_cpu(j, cpu_map) {
6667 struct sched_domain *sd;
6670 sd = *per_cpu_ptr(sdd->sd, j);
6671 if (sd && (sd->flags & SD_OVERLAP))
6672 free_sched_groups(sd->groups, 0);
6673 kfree(*per_cpu_ptr(sdd->sd, j));
6677 kfree(*per_cpu_ptr(sdd->sg, j));
6679 kfree(*per_cpu_ptr(sdd->sgc, j));
6681 free_percpu(sdd->sd);
6683 free_percpu(sdd->sg);
6685 free_percpu(sdd->sgc);
6690 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6691 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6692 struct sched_domain *child, int cpu)
6694 struct sched_domain *sd = sd_init(tl, cpu);
6698 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6700 sd->level = child->level + 1;
6701 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6705 if (!cpumask_subset(sched_domain_span(child),
6706 sched_domain_span(sd))) {
6707 pr_err("BUG: arch topology borken\n");
6708 #ifdef CONFIG_SCHED_DEBUG
6709 pr_err(" the %s domain not a subset of the %s domain\n",
6710 child->name, sd->name);
6712 /* Fixup, ensure @sd has at least @child cpus. */
6713 cpumask_or(sched_domain_span(sd),
6714 sched_domain_span(sd),
6715 sched_domain_span(child));
6719 set_domain_attribute(sd, attr);
6725 * Build sched domains for a given set of cpus and attach the sched domains
6726 * to the individual cpus
6728 static int build_sched_domains(const struct cpumask *cpu_map,
6729 struct sched_domain_attr *attr)
6731 enum s_alloc alloc_state;
6732 struct sched_domain *sd;
6734 int i, ret = -ENOMEM;
6736 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6737 if (alloc_state != sa_rootdomain)
6740 /* Set up domains for cpus specified by the cpu_map. */
6741 for_each_cpu(i, cpu_map) {
6742 struct sched_domain_topology_level *tl;
6745 for_each_sd_topology(tl) {
6746 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6747 if (tl == sched_domain_topology)
6748 *per_cpu_ptr(d.sd, i) = sd;
6749 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6750 sd->flags |= SD_OVERLAP;
6751 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6756 /* Build the groups for the domains */
6757 for_each_cpu(i, cpu_map) {
6758 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6759 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6760 if (sd->flags & SD_OVERLAP) {
6761 if (build_overlap_sched_groups(sd, i))
6764 if (build_sched_groups(sd, i))
6770 /* Calculate CPU capacity for physical packages and nodes */
6771 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6772 if (!cpumask_test_cpu(i, cpu_map))
6775 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6776 claim_allocations(i, sd);
6777 init_sched_groups_capacity(i, sd);
6781 /* Attach the domains */
6783 for_each_cpu(i, cpu_map) {
6784 sd = *per_cpu_ptr(d.sd, i);
6785 cpu_attach_domain(sd, d.rd, i);
6791 __free_domain_allocs(&d, alloc_state, cpu_map);
6795 static cpumask_var_t *doms_cur; /* current sched domains */
6796 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6797 static struct sched_domain_attr *dattr_cur;
6798 /* attribues of custom domains in 'doms_cur' */
6801 * Special case: If a kmalloc of a doms_cur partition (array of
6802 * cpumask) fails, then fallback to a single sched domain,
6803 * as determined by the single cpumask fallback_doms.
6805 static cpumask_var_t fallback_doms;
6808 * arch_update_cpu_topology lets virtualized architectures update the
6809 * cpu core maps. It is supposed to return 1 if the topology changed
6810 * or 0 if it stayed the same.
6812 int __weak arch_update_cpu_topology(void)
6817 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6820 cpumask_var_t *doms;
6822 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6825 for (i = 0; i < ndoms; i++) {
6826 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6827 free_sched_domains(doms, i);
6834 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6837 for (i = 0; i < ndoms; i++)
6838 free_cpumask_var(doms[i]);
6843 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6844 * For now this just excludes isolated cpus, but could be used to
6845 * exclude other special cases in the future.
6847 static int init_sched_domains(const struct cpumask *cpu_map)
6851 arch_update_cpu_topology();
6853 doms_cur = alloc_sched_domains(ndoms_cur);
6855 doms_cur = &fallback_doms;
6856 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6857 err = build_sched_domains(doms_cur[0], NULL);
6858 register_sched_domain_sysctl();
6864 * Detach sched domains from a group of cpus specified in cpu_map
6865 * These cpus will now be attached to the NULL domain
6867 static void detach_destroy_domains(const struct cpumask *cpu_map)
6872 for_each_cpu(i, cpu_map)
6873 cpu_attach_domain(NULL, &def_root_domain, i);
6877 /* handle null as "default" */
6878 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6879 struct sched_domain_attr *new, int idx_new)
6881 struct sched_domain_attr tmp;
6888 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6889 new ? (new + idx_new) : &tmp,
6890 sizeof(struct sched_domain_attr));
6894 * Partition sched domains as specified by the 'ndoms_new'
6895 * cpumasks in the array doms_new[] of cpumasks. This compares
6896 * doms_new[] to the current sched domain partitioning, doms_cur[].
6897 * It destroys each deleted domain and builds each new domain.
6899 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6900 * The masks don't intersect (don't overlap.) We should setup one
6901 * sched domain for each mask. CPUs not in any of the cpumasks will
6902 * not be load balanced. If the same cpumask appears both in the
6903 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6906 * The passed in 'doms_new' should be allocated using
6907 * alloc_sched_domains. This routine takes ownership of it and will
6908 * free_sched_domains it when done with it. If the caller failed the
6909 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6910 * and partition_sched_domains() will fallback to the single partition
6911 * 'fallback_doms', it also forces the domains to be rebuilt.
6913 * If doms_new == NULL it will be replaced with cpu_online_mask.
6914 * ndoms_new == 0 is a special case for destroying existing domains,
6915 * and it will not create the default domain.
6917 * Call with hotplug lock held
6919 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6920 struct sched_domain_attr *dattr_new)
6925 mutex_lock(&sched_domains_mutex);
6927 /* always unregister in case we don't destroy any domains */
6928 unregister_sched_domain_sysctl();
6930 /* Let architecture update cpu core mappings. */
6931 new_topology = arch_update_cpu_topology();
6933 n = doms_new ? ndoms_new : 0;
6935 /* Destroy deleted domains */
6936 for (i = 0; i < ndoms_cur; i++) {
6937 for (j = 0; j < n && !new_topology; j++) {
6938 if (cpumask_equal(doms_cur[i], doms_new[j])
6939 && dattrs_equal(dattr_cur, i, dattr_new, j))
6942 /* no match - a current sched domain not in new doms_new[] */
6943 detach_destroy_domains(doms_cur[i]);
6949 if (doms_new == NULL) {
6951 doms_new = &fallback_doms;
6952 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6953 WARN_ON_ONCE(dattr_new);
6956 /* Build new domains */
6957 for (i = 0; i < ndoms_new; i++) {
6958 for (j = 0; j < n && !new_topology; j++) {
6959 if (cpumask_equal(doms_new[i], doms_cur[j])
6960 && dattrs_equal(dattr_new, i, dattr_cur, j))
6963 /* no match - add a new doms_new */
6964 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6969 /* Remember the new sched domains */
6970 if (doms_cur != &fallback_doms)
6971 free_sched_domains(doms_cur, ndoms_cur);
6972 kfree(dattr_cur); /* kfree(NULL) is safe */
6973 doms_cur = doms_new;
6974 dattr_cur = dattr_new;
6975 ndoms_cur = ndoms_new;
6977 register_sched_domain_sysctl();
6979 mutex_unlock(&sched_domains_mutex);
6982 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6985 * Update cpusets according to cpu_active mask. If cpusets are
6986 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6987 * around partition_sched_domains().
6989 * If we come here as part of a suspend/resume, don't touch cpusets because we
6990 * want to restore it back to its original state upon resume anyway.
6992 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6996 case CPU_ONLINE_FROZEN:
6997 case CPU_DOWN_FAILED_FROZEN:
7000 * num_cpus_frozen tracks how many CPUs are involved in suspend
7001 * resume sequence. As long as this is not the last online
7002 * operation in the resume sequence, just build a single sched
7003 * domain, ignoring cpusets.
7006 if (likely(num_cpus_frozen)) {
7007 partition_sched_domains(1, NULL, NULL);
7012 * This is the last CPU online operation. So fall through and
7013 * restore the original sched domains by considering the
7014 * cpuset configurations.
7018 case CPU_DOWN_FAILED:
7019 cpuset_update_active_cpus(true);
7027 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7031 case CPU_DOWN_PREPARE:
7032 cpuset_update_active_cpus(false);
7034 case CPU_DOWN_PREPARE_FROZEN:
7036 partition_sched_domains(1, NULL, NULL);
7044 void __init sched_init_smp(void)
7046 cpumask_var_t non_isolated_cpus;
7048 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7049 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7054 * There's no userspace yet to cause hotplug operations; hence all the
7055 * cpu masks are stable and all blatant races in the below code cannot
7058 mutex_lock(&sched_domains_mutex);
7059 init_sched_domains(cpu_active_mask);
7060 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7061 if (cpumask_empty(non_isolated_cpus))
7062 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7063 mutex_unlock(&sched_domains_mutex);
7065 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7066 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7067 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7071 /* Move init over to a non-isolated CPU */
7072 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7074 sched_init_granularity();
7075 free_cpumask_var(non_isolated_cpus);
7077 init_sched_rt_class();
7078 init_sched_dl_class();
7081 void __init sched_init_smp(void)
7083 sched_init_granularity();
7085 #endif /* CONFIG_SMP */
7087 const_debug unsigned int sysctl_timer_migration = 1;
7089 int in_sched_functions(unsigned long addr)
7091 return in_lock_functions(addr) ||
7092 (addr >= (unsigned long)__sched_text_start
7093 && addr < (unsigned long)__sched_text_end);
7096 #ifdef CONFIG_CGROUP_SCHED
7098 * Default task group.
7099 * Every task in system belongs to this group at bootup.
7101 struct task_group root_task_group;
7102 LIST_HEAD(task_groups);
7105 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7107 void __init sched_init(void)
7110 unsigned long alloc_size = 0, ptr;
7112 #ifdef CONFIG_FAIR_GROUP_SCHED
7113 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7115 #ifdef CONFIG_RT_GROUP_SCHED
7116 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7119 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7121 #ifdef CONFIG_FAIR_GROUP_SCHED
7122 root_task_group.se = (struct sched_entity **)ptr;
7123 ptr += nr_cpu_ids * sizeof(void **);
7125 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7126 ptr += nr_cpu_ids * sizeof(void **);
7128 #endif /* CONFIG_FAIR_GROUP_SCHED */
7129 #ifdef CONFIG_RT_GROUP_SCHED
7130 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7131 ptr += nr_cpu_ids * sizeof(void **);
7133 root_task_group.rt_rq = (struct rt_rq **)ptr;
7134 ptr += nr_cpu_ids * sizeof(void **);
7136 #endif /* CONFIG_RT_GROUP_SCHED */
7138 #ifdef CONFIG_CPUMASK_OFFSTACK
7139 for_each_possible_cpu(i) {
7140 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7141 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7143 #endif /* CONFIG_CPUMASK_OFFSTACK */
7145 init_rt_bandwidth(&def_rt_bandwidth,
7146 global_rt_period(), global_rt_runtime());
7147 init_dl_bandwidth(&def_dl_bandwidth,
7148 global_rt_period(), global_rt_runtime());
7151 init_defrootdomain();
7154 #ifdef CONFIG_RT_GROUP_SCHED
7155 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7156 global_rt_period(), global_rt_runtime());
7157 #endif /* CONFIG_RT_GROUP_SCHED */
7159 #ifdef CONFIG_CGROUP_SCHED
7160 list_add(&root_task_group.list, &task_groups);
7161 INIT_LIST_HEAD(&root_task_group.children);
7162 INIT_LIST_HEAD(&root_task_group.siblings);
7163 autogroup_init(&init_task);
7165 #endif /* CONFIG_CGROUP_SCHED */
7167 for_each_possible_cpu(i) {
7171 raw_spin_lock_init(&rq->lock);
7173 rq->calc_load_active = 0;
7174 rq->calc_load_update = jiffies + LOAD_FREQ;
7175 init_cfs_rq(&rq->cfs);
7176 init_rt_rq(&rq->rt, rq);
7177 init_dl_rq(&rq->dl, rq);
7178 #ifdef CONFIG_FAIR_GROUP_SCHED
7179 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7180 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7182 * How much cpu bandwidth does root_task_group get?
7184 * In case of task-groups formed thr' the cgroup filesystem, it
7185 * gets 100% of the cpu resources in the system. This overall
7186 * system cpu resource is divided among the tasks of
7187 * root_task_group and its child task-groups in a fair manner,
7188 * based on each entity's (task or task-group's) weight
7189 * (se->load.weight).
7191 * In other words, if root_task_group has 10 tasks of weight
7192 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7193 * then A0's share of the cpu resource is:
7195 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7197 * We achieve this by letting root_task_group's tasks sit
7198 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7200 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7201 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7202 #endif /* CONFIG_FAIR_GROUP_SCHED */
7204 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7205 #ifdef CONFIG_RT_GROUP_SCHED
7206 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7209 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7210 rq->cpu_load[j] = 0;
7212 rq->last_load_update_tick = jiffies;
7217 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7218 rq->post_schedule = 0;
7219 rq->active_balance = 0;
7220 rq->next_balance = jiffies;
7225 rq->avg_idle = 2*sysctl_sched_migration_cost;
7226 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7228 INIT_LIST_HEAD(&rq->cfs_tasks);
7230 rq_attach_root(rq, &def_root_domain);
7231 #ifdef CONFIG_NO_HZ_COMMON
7234 #ifdef CONFIG_NO_HZ_FULL
7235 rq->last_sched_tick = 0;
7239 atomic_set(&rq->nr_iowait, 0);
7242 set_load_weight(&init_task);
7244 #ifdef CONFIG_PREEMPT_NOTIFIERS
7245 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7249 * The boot idle thread does lazy MMU switching as well:
7251 atomic_inc(&init_mm.mm_count);
7252 enter_lazy_tlb(&init_mm, current);
7255 * During early bootup we pretend to be a normal task:
7257 current->sched_class = &fair_sched_class;
7260 * Make us the idle thread. Technically, schedule() should not be
7261 * called from this thread, however somewhere below it might be,
7262 * but because we are the idle thread, we just pick up running again
7263 * when this runqueue becomes "idle".
7265 init_idle(current, smp_processor_id());
7267 calc_load_update = jiffies + LOAD_FREQ;
7270 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7271 /* May be allocated at isolcpus cmdline parse time */
7272 if (cpu_isolated_map == NULL)
7273 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7274 idle_thread_set_boot_cpu();
7275 set_cpu_rq_start_time();
7277 init_sched_fair_class();
7279 scheduler_running = 1;
7282 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7283 static inline int preempt_count_equals(int preempt_offset)
7285 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7287 return (nested == preempt_offset);
7290 void __might_sleep(const char *file, int line, int preempt_offset)
7293 * Blocking primitives will set (and therefore destroy) current->state,
7294 * since we will exit with TASK_RUNNING make sure we enter with it,
7295 * otherwise we will destroy state.
7297 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7298 "do not call blocking ops when !TASK_RUNNING; "
7299 "state=%lx set at [<%p>] %pS\n",
7301 (void *)current->task_state_change,
7302 (void *)current->task_state_change);
7304 ___might_sleep(file, line, preempt_offset);
7306 EXPORT_SYMBOL(__might_sleep);
7308 void ___might_sleep(const char *file, int line, int preempt_offset)
7310 static unsigned long prev_jiffy; /* ratelimiting */
7312 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7313 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7314 !is_idle_task(current)) ||
7315 system_state != SYSTEM_RUNNING || oops_in_progress)
7317 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7319 prev_jiffy = jiffies;
7322 "BUG: sleeping function called from invalid context at %s:%d\n",
7325 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7326 in_atomic(), irqs_disabled(),
7327 current->pid, current->comm);
7329 if (task_stack_end_corrupted(current))
7330 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7332 debug_show_held_locks(current);
7333 if (irqs_disabled())
7334 print_irqtrace_events(current);
7335 #ifdef CONFIG_DEBUG_PREEMPT
7336 if (!preempt_count_equals(preempt_offset)) {
7337 pr_err("Preemption disabled at:");
7338 print_ip_sym(current->preempt_disable_ip);
7344 EXPORT_SYMBOL(___might_sleep);
7347 #ifdef CONFIG_MAGIC_SYSRQ
7348 static void normalize_task(struct rq *rq, struct task_struct *p)
7350 const struct sched_class *prev_class = p->sched_class;
7351 struct sched_attr attr = {
7352 .sched_policy = SCHED_NORMAL,
7354 int old_prio = p->prio;
7357 queued = task_on_rq_queued(p);
7359 dequeue_task(rq, p, 0);
7360 __setscheduler(rq, p, &attr);
7362 enqueue_task(rq, p, 0);
7366 check_class_changed(rq, p, prev_class, old_prio);
7369 void normalize_rt_tasks(void)
7371 struct task_struct *g, *p;
7372 unsigned long flags;
7375 read_lock(&tasklist_lock);
7376 for_each_process_thread(g, p) {
7378 * Only normalize user tasks:
7380 if (p->flags & PF_KTHREAD)
7383 p->se.exec_start = 0;
7384 #ifdef CONFIG_SCHEDSTATS
7385 p->se.statistics.wait_start = 0;
7386 p->se.statistics.sleep_start = 0;
7387 p->se.statistics.block_start = 0;
7390 if (!dl_task(p) && !rt_task(p)) {
7392 * Renice negative nice level userspace
7395 if (task_nice(p) < 0)
7396 set_user_nice(p, 0);
7400 rq = task_rq_lock(p, &flags);
7401 normalize_task(rq, p);
7402 task_rq_unlock(rq, p, &flags);
7404 read_unlock(&tasklist_lock);
7407 #endif /* CONFIG_MAGIC_SYSRQ */
7409 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7411 * These functions are only useful for the IA64 MCA handling, or kdb.
7413 * They can only be called when the whole system has been
7414 * stopped - every CPU needs to be quiescent, and no scheduling
7415 * activity can take place. Using them for anything else would
7416 * be a serious bug, and as a result, they aren't even visible
7417 * under any other configuration.
7421 * curr_task - return the current task for a given cpu.
7422 * @cpu: the processor in question.
7424 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7426 * Return: The current task for @cpu.
7428 struct task_struct *curr_task(int cpu)
7430 return cpu_curr(cpu);
7433 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7437 * set_curr_task - set the current task for a given cpu.
7438 * @cpu: the processor in question.
7439 * @p: the task pointer to set.
7441 * Description: This function must only be used when non-maskable interrupts
7442 * are serviced on a separate stack. It allows the architecture to switch the
7443 * notion of the current task on a cpu in a non-blocking manner. This function
7444 * must be called with all CPU's synchronized, and interrupts disabled, the
7445 * and caller must save the original value of the current task (see
7446 * curr_task() above) and restore that value before reenabling interrupts and
7447 * re-starting the system.
7449 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7451 void set_curr_task(int cpu, struct task_struct *p)
7458 #ifdef CONFIG_CGROUP_SCHED
7459 /* task_group_lock serializes the addition/removal of task groups */
7460 static DEFINE_SPINLOCK(task_group_lock);
7462 static void free_sched_group(struct task_group *tg)
7464 free_fair_sched_group(tg);
7465 free_rt_sched_group(tg);
7470 /* allocate runqueue etc for a new task group */
7471 struct task_group *sched_create_group(struct task_group *parent)
7473 struct task_group *tg;
7475 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7477 return ERR_PTR(-ENOMEM);
7479 if (!alloc_fair_sched_group(tg, parent))
7482 if (!alloc_rt_sched_group(tg, parent))
7488 free_sched_group(tg);
7489 return ERR_PTR(-ENOMEM);
7492 void sched_online_group(struct task_group *tg, struct task_group *parent)
7494 unsigned long flags;
7496 spin_lock_irqsave(&task_group_lock, flags);
7497 list_add_rcu(&tg->list, &task_groups);
7499 WARN_ON(!parent); /* root should already exist */
7501 tg->parent = parent;
7502 INIT_LIST_HEAD(&tg->children);
7503 list_add_rcu(&tg->siblings, &parent->children);
7504 spin_unlock_irqrestore(&task_group_lock, flags);
7507 /* rcu callback to free various structures associated with a task group */
7508 static void free_sched_group_rcu(struct rcu_head *rhp)
7510 /* now it should be safe to free those cfs_rqs */
7511 free_sched_group(container_of(rhp, struct task_group, rcu));
7514 /* Destroy runqueue etc associated with a task group */
7515 void sched_destroy_group(struct task_group *tg)
7517 /* wait for possible concurrent references to cfs_rqs complete */
7518 call_rcu(&tg->rcu, free_sched_group_rcu);
7521 void sched_offline_group(struct task_group *tg)
7523 unsigned long flags;
7526 /* end participation in shares distribution */
7527 for_each_possible_cpu(i)
7528 unregister_fair_sched_group(tg, i);
7530 spin_lock_irqsave(&task_group_lock, flags);
7531 list_del_rcu(&tg->list);
7532 list_del_rcu(&tg->siblings);
7533 spin_unlock_irqrestore(&task_group_lock, flags);
7536 /* change task's runqueue when it moves between groups.
7537 * The caller of this function should have put the task in its new group
7538 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7539 * reflect its new group.
7541 void sched_move_task(struct task_struct *tsk)
7543 struct task_group *tg;
7544 int queued, running;
7545 unsigned long flags;
7548 rq = task_rq_lock(tsk, &flags);
7550 running = task_current(rq, tsk);
7551 queued = task_on_rq_queued(tsk);
7554 dequeue_task(rq, tsk, 0);
7555 if (unlikely(running))
7556 put_prev_task(rq, tsk);
7559 * All callers are synchronized by task_rq_lock(); we do not use RCU
7560 * which is pointless here. Thus, we pass "true" to task_css_check()
7561 * to prevent lockdep warnings.
7563 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7564 struct task_group, css);
7565 tg = autogroup_task_group(tsk, tg);
7566 tsk->sched_task_group = tg;
7568 #ifdef CONFIG_FAIR_GROUP_SCHED
7569 if (tsk->sched_class->task_move_group)
7570 tsk->sched_class->task_move_group(tsk, queued);
7573 set_task_rq(tsk, task_cpu(tsk));
7575 if (unlikely(running))
7576 tsk->sched_class->set_curr_task(rq);
7578 enqueue_task(rq, tsk, 0);
7580 task_rq_unlock(rq, tsk, &flags);
7582 #endif /* CONFIG_CGROUP_SCHED */
7584 #ifdef CONFIG_RT_GROUP_SCHED
7586 * Ensure that the real time constraints are schedulable.
7588 static DEFINE_MUTEX(rt_constraints_mutex);
7590 /* Must be called with tasklist_lock held */
7591 static inline int tg_has_rt_tasks(struct task_group *tg)
7593 struct task_struct *g, *p;
7596 * Autogroups do not have RT tasks; see autogroup_create().
7598 if (task_group_is_autogroup(tg))
7601 for_each_process_thread(g, p) {
7602 if (rt_task(p) && task_group(p) == tg)
7609 struct rt_schedulable_data {
7610 struct task_group *tg;
7615 static int tg_rt_schedulable(struct task_group *tg, void *data)
7617 struct rt_schedulable_data *d = data;
7618 struct task_group *child;
7619 unsigned long total, sum = 0;
7620 u64 period, runtime;
7622 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7623 runtime = tg->rt_bandwidth.rt_runtime;
7626 period = d->rt_period;
7627 runtime = d->rt_runtime;
7631 * Cannot have more runtime than the period.
7633 if (runtime > period && runtime != RUNTIME_INF)
7637 * Ensure we don't starve existing RT tasks.
7639 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7642 total = to_ratio(period, runtime);
7645 * Nobody can have more than the global setting allows.
7647 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7651 * The sum of our children's runtime should not exceed our own.
7653 list_for_each_entry_rcu(child, &tg->children, siblings) {
7654 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7655 runtime = child->rt_bandwidth.rt_runtime;
7657 if (child == d->tg) {
7658 period = d->rt_period;
7659 runtime = d->rt_runtime;
7662 sum += to_ratio(period, runtime);
7671 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7675 struct rt_schedulable_data data = {
7677 .rt_period = period,
7678 .rt_runtime = runtime,
7682 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7688 static int tg_set_rt_bandwidth(struct task_group *tg,
7689 u64 rt_period, u64 rt_runtime)
7694 * Disallowing the root group RT runtime is BAD, it would disallow the
7695 * kernel creating (and or operating) RT threads.
7697 if (tg == &root_task_group && rt_runtime == 0)
7700 /* No period doesn't make any sense. */
7704 mutex_lock(&rt_constraints_mutex);
7705 read_lock(&tasklist_lock);
7706 err = __rt_schedulable(tg, rt_period, rt_runtime);
7710 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7711 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7712 tg->rt_bandwidth.rt_runtime = rt_runtime;
7714 for_each_possible_cpu(i) {
7715 struct rt_rq *rt_rq = tg->rt_rq[i];
7717 raw_spin_lock(&rt_rq->rt_runtime_lock);
7718 rt_rq->rt_runtime = rt_runtime;
7719 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7721 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7723 read_unlock(&tasklist_lock);
7724 mutex_unlock(&rt_constraints_mutex);
7729 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7731 u64 rt_runtime, rt_period;
7733 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7734 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7735 if (rt_runtime_us < 0)
7736 rt_runtime = RUNTIME_INF;
7738 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7741 static long sched_group_rt_runtime(struct task_group *tg)
7745 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7748 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7749 do_div(rt_runtime_us, NSEC_PER_USEC);
7750 return rt_runtime_us;
7753 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7755 u64 rt_runtime, rt_period;
7757 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7758 rt_runtime = tg->rt_bandwidth.rt_runtime;
7760 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7763 static long sched_group_rt_period(struct task_group *tg)
7767 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7768 do_div(rt_period_us, NSEC_PER_USEC);
7769 return rt_period_us;
7771 #endif /* CONFIG_RT_GROUP_SCHED */
7773 #ifdef CONFIG_RT_GROUP_SCHED
7774 static int sched_rt_global_constraints(void)
7778 mutex_lock(&rt_constraints_mutex);
7779 read_lock(&tasklist_lock);
7780 ret = __rt_schedulable(NULL, 0, 0);
7781 read_unlock(&tasklist_lock);
7782 mutex_unlock(&rt_constraints_mutex);
7787 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7789 /* Don't accept realtime tasks when there is no way for them to run */
7790 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7796 #else /* !CONFIG_RT_GROUP_SCHED */
7797 static int sched_rt_global_constraints(void)
7799 unsigned long flags;
7802 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7803 for_each_possible_cpu(i) {
7804 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7806 raw_spin_lock(&rt_rq->rt_runtime_lock);
7807 rt_rq->rt_runtime = global_rt_runtime();
7808 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7810 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7814 #endif /* CONFIG_RT_GROUP_SCHED */
7816 static int sched_dl_global_constraints(void)
7818 u64 runtime = global_rt_runtime();
7819 u64 period = global_rt_period();
7820 u64 new_bw = to_ratio(period, runtime);
7823 unsigned long flags;
7826 * Here we want to check the bandwidth not being set to some
7827 * value smaller than the currently allocated bandwidth in
7828 * any of the root_domains.
7830 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7831 * cycling on root_domains... Discussion on different/better
7832 * solutions is welcome!
7834 for_each_possible_cpu(cpu) {
7835 rcu_read_lock_sched();
7836 dl_b = dl_bw_of(cpu);
7838 raw_spin_lock_irqsave(&dl_b->lock, flags);
7839 if (new_bw < dl_b->total_bw)
7841 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7843 rcu_read_unlock_sched();
7852 static void sched_dl_do_global(void)
7857 unsigned long flags;
7859 def_dl_bandwidth.dl_period = global_rt_period();
7860 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7862 if (global_rt_runtime() != RUNTIME_INF)
7863 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7866 * FIXME: As above...
7868 for_each_possible_cpu(cpu) {
7869 rcu_read_lock_sched();
7870 dl_b = dl_bw_of(cpu);
7872 raw_spin_lock_irqsave(&dl_b->lock, flags);
7874 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7876 rcu_read_unlock_sched();
7880 static int sched_rt_global_validate(void)
7882 if (sysctl_sched_rt_period <= 0)
7885 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7886 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7892 static void sched_rt_do_global(void)
7894 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7895 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7898 int sched_rt_handler(struct ctl_table *table, int write,
7899 void __user *buffer, size_t *lenp,
7902 int old_period, old_runtime;
7903 static DEFINE_MUTEX(mutex);
7907 old_period = sysctl_sched_rt_period;
7908 old_runtime = sysctl_sched_rt_runtime;
7910 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7912 if (!ret && write) {
7913 ret = sched_rt_global_validate();
7917 ret = sched_rt_global_constraints();
7921 ret = sched_dl_global_constraints();
7925 sched_rt_do_global();
7926 sched_dl_do_global();
7930 sysctl_sched_rt_period = old_period;
7931 sysctl_sched_rt_runtime = old_runtime;
7933 mutex_unlock(&mutex);
7938 int sched_rr_handler(struct ctl_table *table, int write,
7939 void __user *buffer, size_t *lenp,
7943 static DEFINE_MUTEX(mutex);
7946 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7947 /* make sure that internally we keep jiffies */
7948 /* also, writing zero resets timeslice to default */
7949 if (!ret && write) {
7950 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7951 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7953 mutex_unlock(&mutex);
7957 #ifdef CONFIG_CGROUP_SCHED
7959 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7961 return css ? container_of(css, struct task_group, css) : NULL;
7964 static struct cgroup_subsys_state *
7965 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7967 struct task_group *parent = css_tg(parent_css);
7968 struct task_group *tg;
7971 /* This is early initialization for the top cgroup */
7972 return &root_task_group.css;
7975 tg = sched_create_group(parent);
7977 return ERR_PTR(-ENOMEM);
7982 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7984 struct task_group *tg = css_tg(css);
7985 struct task_group *parent = css_tg(css->parent);
7988 sched_online_group(tg, parent);
7992 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7994 struct task_group *tg = css_tg(css);
7996 sched_destroy_group(tg);
7999 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
8001 struct task_group *tg = css_tg(css);
8003 sched_offline_group(tg);
8006 static void cpu_cgroup_fork(struct task_struct *task)
8008 sched_move_task(task);
8011 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8012 struct cgroup_taskset *tset)
8014 struct task_struct *task;
8016 cgroup_taskset_for_each(task, tset) {
8017 #ifdef CONFIG_RT_GROUP_SCHED
8018 if (!sched_rt_can_attach(css_tg(css), task))
8021 /* We don't support RT-tasks being in separate groups */
8022 if (task->sched_class != &fair_sched_class)
8029 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8030 struct cgroup_taskset *tset)
8032 struct task_struct *task;
8034 cgroup_taskset_for_each(task, tset)
8035 sched_move_task(task);
8038 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8039 struct cgroup_subsys_state *old_css,
8040 struct task_struct *task)
8043 * cgroup_exit() is called in the copy_process() failure path.
8044 * Ignore this case since the task hasn't ran yet, this avoids
8045 * trying to poke a half freed task state from generic code.
8047 if (!(task->flags & PF_EXITING))
8050 sched_move_task(task);
8053 #ifdef CONFIG_FAIR_GROUP_SCHED
8054 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8055 struct cftype *cftype, u64 shareval)
8057 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8060 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8063 struct task_group *tg = css_tg(css);
8065 return (u64) scale_load_down(tg->shares);
8068 #ifdef CONFIG_CFS_BANDWIDTH
8069 static DEFINE_MUTEX(cfs_constraints_mutex);
8071 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8072 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8074 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8076 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8078 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8079 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8081 if (tg == &root_task_group)
8085 * Ensure we have at some amount of bandwidth every period. This is
8086 * to prevent reaching a state of large arrears when throttled via
8087 * entity_tick() resulting in prolonged exit starvation.
8089 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8093 * Likewise, bound things on the otherside by preventing insane quota
8094 * periods. This also allows us to normalize in computing quota
8097 if (period > max_cfs_quota_period)
8101 * Prevent race between setting of cfs_rq->runtime_enabled and
8102 * unthrottle_offline_cfs_rqs().
8105 mutex_lock(&cfs_constraints_mutex);
8106 ret = __cfs_schedulable(tg, period, quota);
8110 runtime_enabled = quota != RUNTIME_INF;
8111 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8113 * If we need to toggle cfs_bandwidth_used, off->on must occur
8114 * before making related changes, and on->off must occur afterwards
8116 if (runtime_enabled && !runtime_was_enabled)
8117 cfs_bandwidth_usage_inc();
8118 raw_spin_lock_irq(&cfs_b->lock);
8119 cfs_b->period = ns_to_ktime(period);
8120 cfs_b->quota = quota;
8122 __refill_cfs_bandwidth_runtime(cfs_b);
8123 /* restart the period timer (if active) to handle new period expiry */
8124 if (runtime_enabled && cfs_b->timer_active) {
8125 /* force a reprogram */
8126 __start_cfs_bandwidth(cfs_b, true);
8128 raw_spin_unlock_irq(&cfs_b->lock);
8130 for_each_online_cpu(i) {
8131 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8132 struct rq *rq = cfs_rq->rq;
8134 raw_spin_lock_irq(&rq->lock);
8135 cfs_rq->runtime_enabled = runtime_enabled;
8136 cfs_rq->runtime_remaining = 0;
8138 if (cfs_rq->throttled)
8139 unthrottle_cfs_rq(cfs_rq);
8140 raw_spin_unlock_irq(&rq->lock);
8142 if (runtime_was_enabled && !runtime_enabled)
8143 cfs_bandwidth_usage_dec();
8145 mutex_unlock(&cfs_constraints_mutex);
8151 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8155 period = ktime_to_ns(tg->cfs_bandwidth.period);
8156 if (cfs_quota_us < 0)
8157 quota = RUNTIME_INF;
8159 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8161 return tg_set_cfs_bandwidth(tg, period, quota);
8164 long tg_get_cfs_quota(struct task_group *tg)
8168 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8171 quota_us = tg->cfs_bandwidth.quota;
8172 do_div(quota_us, NSEC_PER_USEC);
8177 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8181 period = (u64)cfs_period_us * NSEC_PER_USEC;
8182 quota = tg->cfs_bandwidth.quota;
8184 return tg_set_cfs_bandwidth(tg, period, quota);
8187 long tg_get_cfs_period(struct task_group *tg)
8191 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8192 do_div(cfs_period_us, NSEC_PER_USEC);
8194 return cfs_period_us;
8197 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8200 return tg_get_cfs_quota(css_tg(css));
8203 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8204 struct cftype *cftype, s64 cfs_quota_us)
8206 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8209 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8212 return tg_get_cfs_period(css_tg(css));
8215 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8216 struct cftype *cftype, u64 cfs_period_us)
8218 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8221 struct cfs_schedulable_data {
8222 struct task_group *tg;
8227 * normalize group quota/period to be quota/max_period
8228 * note: units are usecs
8230 static u64 normalize_cfs_quota(struct task_group *tg,
8231 struct cfs_schedulable_data *d)
8239 period = tg_get_cfs_period(tg);
8240 quota = tg_get_cfs_quota(tg);
8243 /* note: these should typically be equivalent */
8244 if (quota == RUNTIME_INF || quota == -1)
8247 return to_ratio(period, quota);
8250 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8252 struct cfs_schedulable_data *d = data;
8253 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8254 s64 quota = 0, parent_quota = -1;
8257 quota = RUNTIME_INF;
8259 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8261 quota = normalize_cfs_quota(tg, d);
8262 parent_quota = parent_b->hierarchical_quota;
8265 * ensure max(child_quota) <= parent_quota, inherit when no
8268 if (quota == RUNTIME_INF)
8269 quota = parent_quota;
8270 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8273 cfs_b->hierarchical_quota = quota;
8278 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8281 struct cfs_schedulable_data data = {
8287 if (quota != RUNTIME_INF) {
8288 do_div(data.period, NSEC_PER_USEC);
8289 do_div(data.quota, NSEC_PER_USEC);
8293 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8299 static int cpu_stats_show(struct seq_file *sf, void *v)
8301 struct task_group *tg = css_tg(seq_css(sf));
8302 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8304 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8305 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8306 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8310 #endif /* CONFIG_CFS_BANDWIDTH */
8311 #endif /* CONFIG_FAIR_GROUP_SCHED */
8313 #ifdef CONFIG_RT_GROUP_SCHED
8314 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8315 struct cftype *cft, s64 val)
8317 return sched_group_set_rt_runtime(css_tg(css), val);
8320 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8323 return sched_group_rt_runtime(css_tg(css));
8326 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8327 struct cftype *cftype, u64 rt_period_us)
8329 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8332 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8335 return sched_group_rt_period(css_tg(css));
8337 #endif /* CONFIG_RT_GROUP_SCHED */
8339 static struct cftype cpu_files[] = {
8340 #ifdef CONFIG_FAIR_GROUP_SCHED
8343 .read_u64 = cpu_shares_read_u64,
8344 .write_u64 = cpu_shares_write_u64,
8347 #ifdef CONFIG_CFS_BANDWIDTH
8349 .name = "cfs_quota_us",
8350 .read_s64 = cpu_cfs_quota_read_s64,
8351 .write_s64 = cpu_cfs_quota_write_s64,
8354 .name = "cfs_period_us",
8355 .read_u64 = cpu_cfs_period_read_u64,
8356 .write_u64 = cpu_cfs_period_write_u64,
8360 .seq_show = cpu_stats_show,
8363 #ifdef CONFIG_RT_GROUP_SCHED
8365 .name = "rt_runtime_us",
8366 .read_s64 = cpu_rt_runtime_read,
8367 .write_s64 = cpu_rt_runtime_write,
8370 .name = "rt_period_us",
8371 .read_u64 = cpu_rt_period_read_uint,
8372 .write_u64 = cpu_rt_period_write_uint,
8378 struct cgroup_subsys cpu_cgrp_subsys = {
8379 .css_alloc = cpu_cgroup_css_alloc,
8380 .css_free = cpu_cgroup_css_free,
8381 .css_online = cpu_cgroup_css_online,
8382 .css_offline = cpu_cgroup_css_offline,
8383 .fork = cpu_cgroup_fork,
8384 .can_attach = cpu_cgroup_can_attach,
8385 .attach = cpu_cgroup_attach,
8386 .exit = cpu_cgroup_exit,
8387 .legacy_cftypes = cpu_files,
8391 #endif /* CONFIG_CGROUP_SCHED */
8393 void dump_cpu_task(int cpu)
8395 pr_info("Task dump for CPU %d:\n", cpu);
8396 sched_show_task(cpu_curr(cpu));