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 * FIFO realtime policy runs the highest priority task. Other runnable
694 * tasks are of a lower priority. The scheduler tick does nothing.
696 if (current->policy == SCHED_FIFO)
700 * Round-robin realtime tasks time slice with other tasks at the same
701 * realtime priority. Is this task the only one at this priority?
703 if (current->policy == SCHED_RR) {
704 struct sched_rt_entity *rt_se = ¤t->rt;
706 return rt_se->run_list.prev == rt_se->run_list.next;
710 * More than one running task need preemption.
711 * nr_running update is assumed to be visible
712 * after IPI is sent from wakers.
714 if (this_rq()->nr_running > 1)
719 #endif /* CONFIG_NO_HZ_FULL */
721 void sched_avg_update(struct rq *rq)
723 s64 period = sched_avg_period();
725 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
727 * Inline assembly required to prevent the compiler
728 * optimising this loop into a divmod call.
729 * See __iter_div_u64_rem() for another example of this.
731 asm("" : "+rm" (rq->age_stamp));
732 rq->age_stamp += period;
737 #endif /* CONFIG_SMP */
739 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
740 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
742 * Iterate task_group tree rooted at *from, calling @down when first entering a
743 * node and @up when leaving it for the final time.
745 * Caller must hold rcu_lock or sufficient equivalent.
747 int walk_tg_tree_from(struct task_group *from,
748 tg_visitor down, tg_visitor up, void *data)
750 struct task_group *parent, *child;
756 ret = (*down)(parent, data);
759 list_for_each_entry_rcu(child, &parent->children, siblings) {
766 ret = (*up)(parent, data);
767 if (ret || parent == from)
771 parent = parent->parent;
778 int tg_nop(struct task_group *tg, void *data)
784 static void set_load_weight(struct task_struct *p)
786 int prio = p->static_prio - MAX_RT_PRIO;
787 struct load_weight *load = &p->se.load;
790 * SCHED_IDLE tasks get minimal weight:
792 if (p->policy == SCHED_IDLE) {
793 load->weight = scale_load(WEIGHT_IDLEPRIO);
794 load->inv_weight = WMULT_IDLEPRIO;
798 load->weight = scale_load(prio_to_weight[prio]);
799 load->inv_weight = prio_to_wmult[prio];
802 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
805 sched_info_queued(rq, p);
806 p->sched_class->enqueue_task(rq, p, flags);
809 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
812 sched_info_dequeued(rq, p);
813 p->sched_class->dequeue_task(rq, p, flags);
816 void activate_task(struct rq *rq, struct task_struct *p, int flags)
818 if (task_contributes_to_load(p))
819 rq->nr_uninterruptible--;
821 enqueue_task(rq, p, flags);
824 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
826 if (task_contributes_to_load(p))
827 rq->nr_uninterruptible++;
829 dequeue_task(rq, p, flags);
832 static void update_rq_clock_task(struct rq *rq, s64 delta)
835 * In theory, the compile should just see 0 here, and optimize out the call
836 * to sched_rt_avg_update. But I don't trust it...
838 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
839 s64 steal = 0, irq_delta = 0;
841 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
842 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
845 * Since irq_time is only updated on {soft,}irq_exit, we might run into
846 * this case when a previous update_rq_clock() happened inside a
849 * When this happens, we stop ->clock_task and only update the
850 * prev_irq_time stamp to account for the part that fit, so that a next
851 * update will consume the rest. This ensures ->clock_task is
854 * It does however cause some slight miss-attribution of {soft,}irq
855 * time, a more accurate solution would be to update the irq_time using
856 * the current rq->clock timestamp, except that would require using
859 if (irq_delta > delta)
862 rq->prev_irq_time += irq_delta;
865 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
866 if (static_key_false((¶virt_steal_rq_enabled))) {
867 steal = paravirt_steal_clock(cpu_of(rq));
868 steal -= rq->prev_steal_time_rq;
870 if (unlikely(steal > delta))
873 rq->prev_steal_time_rq += steal;
878 rq->clock_task += delta;
880 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
881 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
882 sched_rt_avg_update(rq, irq_delta + steal);
886 void sched_set_stop_task(int cpu, struct task_struct *stop)
888 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
889 struct task_struct *old_stop = cpu_rq(cpu)->stop;
893 * Make it appear like a SCHED_FIFO task, its something
894 * userspace knows about and won't get confused about.
896 * Also, it will make PI more or less work without too
897 * much confusion -- but then, stop work should not
898 * rely on PI working anyway.
900 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
902 stop->sched_class = &stop_sched_class;
905 cpu_rq(cpu)->stop = stop;
909 * Reset it back to a normal scheduling class so that
910 * it can die in pieces.
912 old_stop->sched_class = &rt_sched_class;
917 * __normal_prio - return the priority that is based on the static prio
919 static inline int __normal_prio(struct task_struct *p)
921 return p->static_prio;
925 * Calculate the expected normal priority: i.e. priority
926 * without taking RT-inheritance into account. Might be
927 * boosted by interactivity modifiers. Changes upon fork,
928 * setprio syscalls, and whenever the interactivity
929 * estimator recalculates.
931 static inline int normal_prio(struct task_struct *p)
935 if (task_has_dl_policy(p))
936 prio = MAX_DL_PRIO-1;
937 else if (task_has_rt_policy(p))
938 prio = MAX_RT_PRIO-1 - p->rt_priority;
940 prio = __normal_prio(p);
945 * Calculate the current priority, i.e. the priority
946 * taken into account by the scheduler. This value might
947 * be boosted by RT tasks, or might be boosted by
948 * interactivity modifiers. Will be RT if the task got
949 * RT-boosted. If not then it returns p->normal_prio.
951 static int effective_prio(struct task_struct *p)
953 p->normal_prio = normal_prio(p);
955 * If we are RT tasks or we were boosted to RT priority,
956 * keep the priority unchanged. Otherwise, update priority
957 * to the normal priority:
959 if (!rt_prio(p->prio))
960 return p->normal_prio;
965 * task_curr - is this task currently executing on a CPU?
966 * @p: the task in question.
968 * Return: 1 if the task is currently executing. 0 otherwise.
970 inline int task_curr(const struct task_struct *p)
972 return cpu_curr(task_cpu(p)) == p;
976 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
978 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
979 const struct sched_class *prev_class,
982 if (prev_class != p->sched_class) {
983 if (prev_class->switched_from)
984 prev_class->switched_from(rq, p);
985 /* Possble rq->lock 'hole'. */
986 p->sched_class->switched_to(rq, p);
987 } else if (oldprio != p->prio || dl_task(p))
988 p->sched_class->prio_changed(rq, p, oldprio);
991 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
993 const struct sched_class *class;
995 if (p->sched_class == rq->curr->sched_class) {
996 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
998 for_each_class(class) {
999 if (class == rq->curr->sched_class)
1001 if (class == p->sched_class) {
1009 * A queue event has occurred, and we're going to schedule. In
1010 * this case, we can save a useless back to back clock update.
1012 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1013 rq_clock_skip_update(rq, true);
1017 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1019 #ifdef CONFIG_SCHED_DEBUG
1021 * We should never call set_task_cpu() on a blocked task,
1022 * ttwu() will sort out the placement.
1024 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1027 #ifdef CONFIG_LOCKDEP
1029 * The caller should hold either p->pi_lock or rq->lock, when changing
1030 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1032 * sched_move_task() holds both and thus holding either pins the cgroup,
1035 * Furthermore, all task_rq users should acquire both locks, see
1038 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1039 lockdep_is_held(&task_rq(p)->lock)));
1043 trace_sched_migrate_task(p, new_cpu);
1045 if (task_cpu(p) != new_cpu) {
1046 if (p->sched_class->migrate_task_rq)
1047 p->sched_class->migrate_task_rq(p, new_cpu);
1048 p->se.nr_migrations++;
1049 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1052 __set_task_cpu(p, new_cpu);
1055 static void __migrate_swap_task(struct task_struct *p, int cpu)
1057 if (task_on_rq_queued(p)) {
1058 struct rq *src_rq, *dst_rq;
1060 src_rq = task_rq(p);
1061 dst_rq = cpu_rq(cpu);
1063 deactivate_task(src_rq, p, 0);
1064 set_task_cpu(p, cpu);
1065 activate_task(dst_rq, p, 0);
1066 check_preempt_curr(dst_rq, p, 0);
1069 * Task isn't running anymore; make it appear like we migrated
1070 * it before it went to sleep. This means on wakeup we make the
1071 * previous cpu our targer instead of where it really is.
1077 struct migration_swap_arg {
1078 struct task_struct *src_task, *dst_task;
1079 int src_cpu, dst_cpu;
1082 static int migrate_swap_stop(void *data)
1084 struct migration_swap_arg *arg = data;
1085 struct rq *src_rq, *dst_rq;
1088 src_rq = cpu_rq(arg->src_cpu);
1089 dst_rq = cpu_rq(arg->dst_cpu);
1091 double_raw_lock(&arg->src_task->pi_lock,
1092 &arg->dst_task->pi_lock);
1093 double_rq_lock(src_rq, dst_rq);
1094 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1097 if (task_cpu(arg->src_task) != arg->src_cpu)
1100 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1103 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1106 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1107 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1112 double_rq_unlock(src_rq, dst_rq);
1113 raw_spin_unlock(&arg->dst_task->pi_lock);
1114 raw_spin_unlock(&arg->src_task->pi_lock);
1120 * Cross migrate two tasks
1122 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1124 struct migration_swap_arg arg;
1127 arg = (struct migration_swap_arg){
1129 .src_cpu = task_cpu(cur),
1131 .dst_cpu = task_cpu(p),
1134 if (arg.src_cpu == arg.dst_cpu)
1138 * These three tests are all lockless; this is OK since all of them
1139 * will be re-checked with proper locks held further down the line.
1141 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1144 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1147 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1150 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1151 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1157 struct migration_arg {
1158 struct task_struct *task;
1162 static int migration_cpu_stop(void *data);
1165 * wait_task_inactive - wait for a thread to unschedule.
1167 * If @match_state is nonzero, it's the @p->state value just checked and
1168 * not expected to change. If it changes, i.e. @p might have woken up,
1169 * then return zero. When we succeed in waiting for @p to be off its CPU,
1170 * we return a positive number (its total switch count). If a second call
1171 * a short while later returns the same number, the caller can be sure that
1172 * @p has remained unscheduled the whole time.
1174 * The caller must ensure that the task *will* unschedule sometime soon,
1175 * else this function might spin for a *long* time. This function can't
1176 * be called with interrupts off, or it may introduce deadlock with
1177 * smp_call_function() if an IPI is sent by the same process we are
1178 * waiting to become inactive.
1180 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1182 unsigned long flags;
1183 int running, queued;
1189 * We do the initial early heuristics without holding
1190 * any task-queue locks at all. We'll only try to get
1191 * the runqueue lock when things look like they will
1197 * If the task is actively running on another CPU
1198 * still, just relax and busy-wait without holding
1201 * NOTE! Since we don't hold any locks, it's not
1202 * even sure that "rq" stays as the right runqueue!
1203 * But we don't care, since "task_running()" will
1204 * return false if the runqueue has changed and p
1205 * is actually now running somewhere else!
1207 while (task_running(rq, p)) {
1208 if (match_state && unlikely(p->state != match_state))
1214 * Ok, time to look more closely! We need the rq
1215 * lock now, to be *sure*. If we're wrong, we'll
1216 * just go back and repeat.
1218 rq = task_rq_lock(p, &flags);
1219 trace_sched_wait_task(p);
1220 running = task_running(rq, p);
1221 queued = task_on_rq_queued(p);
1223 if (!match_state || p->state == match_state)
1224 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1225 task_rq_unlock(rq, p, &flags);
1228 * If it changed from the expected state, bail out now.
1230 if (unlikely(!ncsw))
1234 * Was it really running after all now that we
1235 * checked with the proper locks actually held?
1237 * Oops. Go back and try again..
1239 if (unlikely(running)) {
1245 * It's not enough that it's not actively running,
1246 * it must be off the runqueue _entirely_, and not
1249 * So if it was still runnable (but just not actively
1250 * running right now), it's preempted, and we should
1251 * yield - it could be a while.
1253 if (unlikely(queued)) {
1254 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1256 set_current_state(TASK_UNINTERRUPTIBLE);
1257 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1262 * Ahh, all good. It wasn't running, and it wasn't
1263 * runnable, which means that it will never become
1264 * running in the future either. We're all done!
1273 * kick_process - kick a running thread to enter/exit the kernel
1274 * @p: the to-be-kicked thread
1276 * Cause a process which is running on another CPU to enter
1277 * kernel-mode, without any delay. (to get signals handled.)
1279 * NOTE: this function doesn't have to take the runqueue lock,
1280 * because all it wants to ensure is that the remote task enters
1281 * the kernel. If the IPI races and the task has been migrated
1282 * to another CPU then no harm is done and the purpose has been
1285 void kick_process(struct task_struct *p)
1291 if ((cpu != smp_processor_id()) && task_curr(p))
1292 smp_send_reschedule(cpu);
1295 EXPORT_SYMBOL_GPL(kick_process);
1296 #endif /* CONFIG_SMP */
1300 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1302 static int select_fallback_rq(int cpu, struct task_struct *p)
1304 int nid = cpu_to_node(cpu);
1305 const struct cpumask *nodemask = NULL;
1306 enum { cpuset, possible, fail } state = cpuset;
1310 * If the node that the cpu is on has been offlined, cpu_to_node()
1311 * will return -1. There is no cpu on the node, and we should
1312 * select the cpu on the other node.
1315 nodemask = cpumask_of_node(nid);
1317 /* Look for allowed, online CPU in same node. */
1318 for_each_cpu(dest_cpu, nodemask) {
1319 if (!cpu_online(dest_cpu))
1321 if (!cpu_active(dest_cpu))
1323 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1329 /* Any allowed, online CPU? */
1330 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1331 if (!cpu_online(dest_cpu))
1333 if (!cpu_active(dest_cpu))
1340 /* No more Mr. Nice Guy. */
1341 cpuset_cpus_allowed_fallback(p);
1346 do_set_cpus_allowed(p, cpu_possible_mask);
1357 if (state != cpuset) {
1359 * Don't tell them about moving exiting tasks or
1360 * kernel threads (both mm NULL), since they never
1363 if (p->mm && printk_ratelimit()) {
1364 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1365 task_pid_nr(p), p->comm, cpu);
1373 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1376 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1378 if (p->nr_cpus_allowed > 1)
1379 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1382 * In order not to call set_task_cpu() on a blocking task we need
1383 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1386 * Since this is common to all placement strategies, this lives here.
1388 * [ this allows ->select_task() to simply return task_cpu(p) and
1389 * not worry about this generic constraint ]
1391 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1393 cpu = select_fallback_rq(task_cpu(p), p);
1398 static void update_avg(u64 *avg, u64 sample)
1400 s64 diff = sample - *avg;
1406 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1408 #ifdef CONFIG_SCHEDSTATS
1409 struct rq *rq = this_rq();
1412 int this_cpu = smp_processor_id();
1414 if (cpu == this_cpu) {
1415 schedstat_inc(rq, ttwu_local);
1416 schedstat_inc(p, se.statistics.nr_wakeups_local);
1418 struct sched_domain *sd;
1420 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1422 for_each_domain(this_cpu, sd) {
1423 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1424 schedstat_inc(sd, ttwu_wake_remote);
1431 if (wake_flags & WF_MIGRATED)
1432 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1434 #endif /* CONFIG_SMP */
1436 schedstat_inc(rq, ttwu_count);
1437 schedstat_inc(p, se.statistics.nr_wakeups);
1439 if (wake_flags & WF_SYNC)
1440 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1442 #endif /* CONFIG_SCHEDSTATS */
1445 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1447 activate_task(rq, p, en_flags);
1448 p->on_rq = TASK_ON_RQ_QUEUED;
1450 /* if a worker is waking up, notify workqueue */
1451 if (p->flags & PF_WQ_WORKER)
1452 wq_worker_waking_up(p, cpu_of(rq));
1456 * Mark the task runnable and perform wakeup-preemption.
1459 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1461 check_preempt_curr(rq, p, wake_flags);
1462 trace_sched_wakeup(p, true);
1464 p->state = TASK_RUNNING;
1466 if (p->sched_class->task_woken)
1467 p->sched_class->task_woken(rq, p);
1469 if (rq->idle_stamp) {
1470 u64 delta = rq_clock(rq) - rq->idle_stamp;
1471 u64 max = 2*rq->max_idle_balance_cost;
1473 update_avg(&rq->avg_idle, delta);
1475 if (rq->avg_idle > max)
1484 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1487 if (p->sched_contributes_to_load)
1488 rq->nr_uninterruptible--;
1491 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1492 ttwu_do_wakeup(rq, p, wake_flags);
1496 * Called in case the task @p isn't fully descheduled from its runqueue,
1497 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1498 * since all we need to do is flip p->state to TASK_RUNNING, since
1499 * the task is still ->on_rq.
1501 static int ttwu_remote(struct task_struct *p, int wake_flags)
1506 rq = __task_rq_lock(p);
1507 if (task_on_rq_queued(p)) {
1508 /* check_preempt_curr() may use rq clock */
1509 update_rq_clock(rq);
1510 ttwu_do_wakeup(rq, p, wake_flags);
1513 __task_rq_unlock(rq);
1519 void sched_ttwu_pending(void)
1521 struct rq *rq = this_rq();
1522 struct llist_node *llist = llist_del_all(&rq->wake_list);
1523 struct task_struct *p;
1524 unsigned long flags;
1529 raw_spin_lock_irqsave(&rq->lock, flags);
1532 p = llist_entry(llist, struct task_struct, wake_entry);
1533 llist = llist_next(llist);
1534 ttwu_do_activate(rq, p, 0);
1537 raw_spin_unlock_irqrestore(&rq->lock, flags);
1540 void scheduler_ipi(void)
1543 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1544 * TIF_NEED_RESCHED remotely (for the first time) will also send
1547 preempt_fold_need_resched();
1549 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1553 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1554 * traditionally all their work was done from the interrupt return
1555 * path. Now that we actually do some work, we need to make sure
1558 * Some archs already do call them, luckily irq_enter/exit nest
1561 * Arguably we should visit all archs and update all handlers,
1562 * however a fair share of IPIs are still resched only so this would
1563 * somewhat pessimize the simple resched case.
1566 sched_ttwu_pending();
1569 * Check if someone kicked us for doing the nohz idle load balance.
1571 if (unlikely(got_nohz_idle_kick())) {
1572 this_rq()->idle_balance = 1;
1573 raise_softirq_irqoff(SCHED_SOFTIRQ);
1578 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1580 struct rq *rq = cpu_rq(cpu);
1582 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1583 if (!set_nr_if_polling(rq->idle))
1584 smp_send_reschedule(cpu);
1586 trace_sched_wake_idle_without_ipi(cpu);
1590 void wake_up_if_idle(int cpu)
1592 struct rq *rq = cpu_rq(cpu);
1593 unsigned long flags;
1597 if (!is_idle_task(rcu_dereference(rq->curr)))
1600 if (set_nr_if_polling(rq->idle)) {
1601 trace_sched_wake_idle_without_ipi(cpu);
1603 raw_spin_lock_irqsave(&rq->lock, flags);
1604 if (is_idle_task(rq->curr))
1605 smp_send_reschedule(cpu);
1606 /* Else cpu is not in idle, do nothing here */
1607 raw_spin_unlock_irqrestore(&rq->lock, flags);
1614 bool cpus_share_cache(int this_cpu, int that_cpu)
1616 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1618 #endif /* CONFIG_SMP */
1620 static void ttwu_queue(struct task_struct *p, int cpu)
1622 struct rq *rq = cpu_rq(cpu);
1624 #if defined(CONFIG_SMP)
1625 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1626 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1627 ttwu_queue_remote(p, cpu);
1632 raw_spin_lock(&rq->lock);
1633 ttwu_do_activate(rq, p, 0);
1634 raw_spin_unlock(&rq->lock);
1638 * try_to_wake_up - wake up a thread
1639 * @p: the thread to be awakened
1640 * @state: the mask of task states that can be woken
1641 * @wake_flags: wake modifier flags (WF_*)
1643 * Put it on the run-queue if it's not already there. The "current"
1644 * thread is always on the run-queue (except when the actual
1645 * re-schedule is in progress), and as such you're allowed to do
1646 * the simpler "current->state = TASK_RUNNING" to mark yourself
1647 * runnable without the overhead of this.
1649 * Return: %true if @p was woken up, %false if it was already running.
1650 * or @state didn't match @p's state.
1653 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1655 unsigned long flags;
1656 int cpu, success = 0;
1659 * If we are going to wake up a thread waiting for CONDITION we
1660 * need to ensure that CONDITION=1 done by the caller can not be
1661 * reordered with p->state check below. This pairs with mb() in
1662 * set_current_state() the waiting thread does.
1664 smp_mb__before_spinlock();
1665 raw_spin_lock_irqsave(&p->pi_lock, flags);
1666 if (!(p->state & state))
1669 success = 1; /* we're going to change ->state */
1672 if (p->on_rq && ttwu_remote(p, wake_flags))
1677 * If the owning (remote) cpu is still in the middle of schedule() with
1678 * this task as prev, wait until its done referencing the task.
1683 * Pairs with the smp_wmb() in finish_lock_switch().
1687 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1688 p->state = TASK_WAKING;
1690 if (p->sched_class->task_waking)
1691 p->sched_class->task_waking(p);
1693 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1694 if (task_cpu(p) != cpu) {
1695 wake_flags |= WF_MIGRATED;
1696 set_task_cpu(p, cpu);
1698 #endif /* CONFIG_SMP */
1702 ttwu_stat(p, cpu, wake_flags);
1704 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1710 * try_to_wake_up_local - try to wake up a local task with rq lock held
1711 * @p: the thread to be awakened
1713 * Put @p on the run-queue if it's not already there. The caller must
1714 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1717 static void try_to_wake_up_local(struct task_struct *p)
1719 struct rq *rq = task_rq(p);
1721 if (WARN_ON_ONCE(rq != this_rq()) ||
1722 WARN_ON_ONCE(p == current))
1725 lockdep_assert_held(&rq->lock);
1727 if (!raw_spin_trylock(&p->pi_lock)) {
1728 raw_spin_unlock(&rq->lock);
1729 raw_spin_lock(&p->pi_lock);
1730 raw_spin_lock(&rq->lock);
1733 if (!(p->state & TASK_NORMAL))
1736 if (!task_on_rq_queued(p))
1737 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1739 ttwu_do_wakeup(rq, p, 0);
1740 ttwu_stat(p, smp_processor_id(), 0);
1742 raw_spin_unlock(&p->pi_lock);
1746 * wake_up_process - Wake up a specific process
1747 * @p: The process to be woken up.
1749 * Attempt to wake up the nominated process and move it to the set of runnable
1752 * Return: 1 if the process was woken up, 0 if it was already running.
1754 * It may be assumed that this function implies a write memory barrier before
1755 * changing the task state if and only if any tasks are woken up.
1757 int wake_up_process(struct task_struct *p)
1759 WARN_ON(task_is_stopped_or_traced(p));
1760 return try_to_wake_up(p, TASK_NORMAL, 0);
1762 EXPORT_SYMBOL(wake_up_process);
1764 int wake_up_state(struct task_struct *p, unsigned int state)
1766 return try_to_wake_up(p, state, 0);
1770 * This function clears the sched_dl_entity static params.
1772 void __dl_clear_params(struct task_struct *p)
1774 struct sched_dl_entity *dl_se = &p->dl;
1776 dl_se->dl_runtime = 0;
1777 dl_se->dl_deadline = 0;
1778 dl_se->dl_period = 0;
1782 dl_se->dl_throttled = 0;
1784 dl_se->dl_yielded = 0;
1788 * Perform scheduler related setup for a newly forked process p.
1789 * p is forked by current.
1791 * __sched_fork() is basic setup used by init_idle() too:
1793 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1798 p->se.exec_start = 0;
1799 p->se.sum_exec_runtime = 0;
1800 p->se.prev_sum_exec_runtime = 0;
1801 p->se.nr_migrations = 0;
1804 p->se.avg.decay_count = 0;
1806 INIT_LIST_HEAD(&p->se.group_node);
1808 #ifdef CONFIG_SCHEDSTATS
1809 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1812 RB_CLEAR_NODE(&p->dl.rb_node);
1813 init_dl_task_timer(&p->dl);
1814 __dl_clear_params(p);
1816 INIT_LIST_HEAD(&p->rt.run_list);
1818 #ifdef CONFIG_PREEMPT_NOTIFIERS
1819 INIT_HLIST_HEAD(&p->preempt_notifiers);
1822 #ifdef CONFIG_NUMA_BALANCING
1823 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1824 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1825 p->mm->numa_scan_seq = 0;
1828 if (clone_flags & CLONE_VM)
1829 p->numa_preferred_nid = current->numa_preferred_nid;
1831 p->numa_preferred_nid = -1;
1833 p->node_stamp = 0ULL;
1834 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1835 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1836 p->numa_work.next = &p->numa_work;
1837 p->numa_faults = NULL;
1838 p->last_task_numa_placement = 0;
1839 p->last_sum_exec_runtime = 0;
1841 p->numa_group = NULL;
1842 #endif /* CONFIG_NUMA_BALANCING */
1845 #ifdef CONFIG_NUMA_BALANCING
1846 #ifdef CONFIG_SCHED_DEBUG
1847 void set_numabalancing_state(bool enabled)
1850 sched_feat_set("NUMA");
1852 sched_feat_set("NO_NUMA");
1855 __read_mostly bool numabalancing_enabled;
1857 void set_numabalancing_state(bool enabled)
1859 numabalancing_enabled = enabled;
1861 #endif /* CONFIG_SCHED_DEBUG */
1863 #ifdef CONFIG_PROC_SYSCTL
1864 int sysctl_numa_balancing(struct ctl_table *table, int write,
1865 void __user *buffer, size_t *lenp, loff_t *ppos)
1869 int state = numabalancing_enabled;
1871 if (write && !capable(CAP_SYS_ADMIN))
1876 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1880 set_numabalancing_state(state);
1887 * fork()/clone()-time setup:
1889 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1891 unsigned long flags;
1892 int cpu = get_cpu();
1894 __sched_fork(clone_flags, p);
1896 * We mark the process as running here. This guarantees that
1897 * nobody will actually run it, and a signal or other external
1898 * event cannot wake it up and insert it on the runqueue either.
1900 p->state = TASK_RUNNING;
1903 * Make sure we do not leak PI boosting priority to the child.
1905 p->prio = current->normal_prio;
1908 * Revert to default priority/policy on fork if requested.
1910 if (unlikely(p->sched_reset_on_fork)) {
1911 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1912 p->policy = SCHED_NORMAL;
1913 p->static_prio = NICE_TO_PRIO(0);
1915 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1916 p->static_prio = NICE_TO_PRIO(0);
1918 p->prio = p->normal_prio = __normal_prio(p);
1922 * We don't need the reset flag anymore after the fork. It has
1923 * fulfilled its duty:
1925 p->sched_reset_on_fork = 0;
1928 if (dl_prio(p->prio)) {
1931 } else if (rt_prio(p->prio)) {
1932 p->sched_class = &rt_sched_class;
1934 p->sched_class = &fair_sched_class;
1937 if (p->sched_class->task_fork)
1938 p->sched_class->task_fork(p);
1941 * The child is not yet in the pid-hash so no cgroup attach races,
1942 * and the cgroup is pinned to this child due to cgroup_fork()
1943 * is ran before sched_fork().
1945 * Silence PROVE_RCU.
1947 raw_spin_lock_irqsave(&p->pi_lock, flags);
1948 set_task_cpu(p, cpu);
1949 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1951 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1952 if (likely(sched_info_on()))
1953 memset(&p->sched_info, 0, sizeof(p->sched_info));
1955 #if defined(CONFIG_SMP)
1958 init_task_preempt_count(p);
1960 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1961 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1968 unsigned long to_ratio(u64 period, u64 runtime)
1970 if (runtime == RUNTIME_INF)
1974 * Doing this here saves a lot of checks in all
1975 * the calling paths, and returning zero seems
1976 * safe for them anyway.
1981 return div64_u64(runtime << 20, period);
1985 inline struct dl_bw *dl_bw_of(int i)
1987 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1988 "sched RCU must be held");
1989 return &cpu_rq(i)->rd->dl_bw;
1992 static inline int dl_bw_cpus(int i)
1994 struct root_domain *rd = cpu_rq(i)->rd;
1997 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1998 "sched RCU must be held");
1999 for_each_cpu_and(i, rd->span, cpu_active_mask)
2005 inline struct dl_bw *dl_bw_of(int i)
2007 return &cpu_rq(i)->dl.dl_bw;
2010 static inline int dl_bw_cpus(int i)
2017 * We must be sure that accepting a new task (or allowing changing the
2018 * parameters of an existing one) is consistent with the bandwidth
2019 * constraints. If yes, this function also accordingly updates the currently
2020 * allocated bandwidth to reflect the new situation.
2022 * This function is called while holding p's rq->lock.
2024 * XXX we should delay bw change until the task's 0-lag point, see
2027 static int dl_overflow(struct task_struct *p, int policy,
2028 const struct sched_attr *attr)
2031 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2032 u64 period = attr->sched_period ?: attr->sched_deadline;
2033 u64 runtime = attr->sched_runtime;
2034 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2037 if (new_bw == p->dl.dl_bw)
2041 * Either if a task, enters, leave, or stays -deadline but changes
2042 * its parameters, we may need to update accordingly the total
2043 * allocated bandwidth of the container.
2045 raw_spin_lock(&dl_b->lock);
2046 cpus = dl_bw_cpus(task_cpu(p));
2047 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2048 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2049 __dl_add(dl_b, new_bw);
2051 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2052 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2053 __dl_clear(dl_b, p->dl.dl_bw);
2054 __dl_add(dl_b, new_bw);
2056 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2057 __dl_clear(dl_b, p->dl.dl_bw);
2060 raw_spin_unlock(&dl_b->lock);
2065 extern void init_dl_bw(struct dl_bw *dl_b);
2068 * wake_up_new_task - wake up a newly created task for the first time.
2070 * This function will do some initial scheduler statistics housekeeping
2071 * that must be done for every newly created context, then puts the task
2072 * on the runqueue and wakes it.
2074 void wake_up_new_task(struct task_struct *p)
2076 unsigned long flags;
2079 raw_spin_lock_irqsave(&p->pi_lock, flags);
2082 * Fork balancing, do it here and not earlier because:
2083 * - cpus_allowed can change in the fork path
2084 * - any previously selected cpu might disappear through hotplug
2086 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2089 /* Initialize new task's runnable average */
2090 init_task_runnable_average(p);
2091 rq = __task_rq_lock(p);
2092 activate_task(rq, p, 0);
2093 p->on_rq = TASK_ON_RQ_QUEUED;
2094 trace_sched_wakeup_new(p, true);
2095 check_preempt_curr(rq, p, WF_FORK);
2097 if (p->sched_class->task_woken)
2098 p->sched_class->task_woken(rq, p);
2100 task_rq_unlock(rq, p, &flags);
2103 #ifdef CONFIG_PREEMPT_NOTIFIERS
2106 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2107 * @notifier: notifier struct to register
2109 void preempt_notifier_register(struct preempt_notifier *notifier)
2111 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2113 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2116 * preempt_notifier_unregister - no longer interested in preemption notifications
2117 * @notifier: notifier struct to unregister
2119 * This is safe to call from within a preemption notifier.
2121 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2123 hlist_del(¬ifier->link);
2125 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2127 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2129 struct preempt_notifier *notifier;
2131 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2132 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2136 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2137 struct task_struct *next)
2139 struct preempt_notifier *notifier;
2141 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2142 notifier->ops->sched_out(notifier, next);
2145 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2147 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2152 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2153 struct task_struct *next)
2157 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2160 * prepare_task_switch - prepare to switch tasks
2161 * @rq: the runqueue preparing to switch
2162 * @prev: the current task that is being switched out
2163 * @next: the task we are going to switch to.
2165 * This is called with the rq lock held and interrupts off. It must
2166 * be paired with a subsequent finish_task_switch after the context
2169 * prepare_task_switch sets up locking and calls architecture specific
2173 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2174 struct task_struct *next)
2176 trace_sched_switch(prev, next);
2177 sched_info_switch(rq, prev, next);
2178 perf_event_task_sched_out(prev, next);
2179 fire_sched_out_preempt_notifiers(prev, next);
2180 prepare_lock_switch(rq, next);
2181 prepare_arch_switch(next);
2185 * finish_task_switch - clean up after a task-switch
2186 * @prev: the thread we just switched away from.
2188 * finish_task_switch must be called after the context switch, paired
2189 * with a prepare_task_switch call before the context switch.
2190 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2191 * and do any other architecture-specific cleanup actions.
2193 * Note that we may have delayed dropping an mm in context_switch(). If
2194 * so, we finish that here outside of the runqueue lock. (Doing it
2195 * with the lock held can cause deadlocks; see schedule() for
2198 * The context switch have flipped the stack from under us and restored the
2199 * local variables which were saved when this task called schedule() in the
2200 * past. prev == current is still correct but we need to recalculate this_rq
2201 * because prev may have moved to another CPU.
2203 static struct rq *finish_task_switch(struct task_struct *prev)
2204 __releases(rq->lock)
2206 struct rq *rq = this_rq();
2207 struct mm_struct *mm = rq->prev_mm;
2213 * A task struct has one reference for the use as "current".
2214 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2215 * schedule one last time. The schedule call will never return, and
2216 * the scheduled task must drop that reference.
2217 * The test for TASK_DEAD must occur while the runqueue locks are
2218 * still held, otherwise prev could be scheduled on another cpu, die
2219 * there before we look at prev->state, and then the reference would
2221 * Manfred Spraul <manfred@colorfullife.com>
2223 prev_state = prev->state;
2224 vtime_task_switch(prev);
2225 finish_arch_switch(prev);
2226 perf_event_task_sched_in(prev, current);
2227 finish_lock_switch(rq, prev);
2228 finish_arch_post_lock_switch();
2230 fire_sched_in_preempt_notifiers(current);
2233 if (unlikely(prev_state == TASK_DEAD)) {
2234 if (prev->sched_class->task_dead)
2235 prev->sched_class->task_dead(prev);
2238 * Remove function-return probe instances associated with this
2239 * task and put them back on the free list.
2241 kprobe_flush_task(prev);
2242 put_task_struct(prev);
2245 tick_nohz_task_switch(current);
2251 /* rq->lock is NOT held, but preemption is disabled */
2252 static inline void post_schedule(struct rq *rq)
2254 if (rq->post_schedule) {
2255 unsigned long flags;
2257 raw_spin_lock_irqsave(&rq->lock, flags);
2258 if (rq->curr->sched_class->post_schedule)
2259 rq->curr->sched_class->post_schedule(rq);
2260 raw_spin_unlock_irqrestore(&rq->lock, flags);
2262 rq->post_schedule = 0;
2268 static inline void post_schedule(struct rq *rq)
2275 * schedule_tail - first thing a freshly forked thread must call.
2276 * @prev: the thread we just switched away from.
2278 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2279 __releases(rq->lock)
2283 /* finish_task_switch() drops rq->lock and enables preemtion */
2285 rq = finish_task_switch(prev);
2289 if (current->set_child_tid)
2290 put_user(task_pid_vnr(current), current->set_child_tid);
2294 * context_switch - switch to the new MM and the new thread's register state.
2296 static inline struct rq *
2297 context_switch(struct rq *rq, struct task_struct *prev,
2298 struct task_struct *next)
2300 struct mm_struct *mm, *oldmm;
2302 prepare_task_switch(rq, prev, next);
2305 oldmm = prev->active_mm;
2307 * For paravirt, this is coupled with an exit in switch_to to
2308 * combine the page table reload and the switch backend into
2311 arch_start_context_switch(prev);
2314 next->active_mm = oldmm;
2315 atomic_inc(&oldmm->mm_count);
2316 enter_lazy_tlb(oldmm, next);
2318 switch_mm(oldmm, mm, next);
2321 prev->active_mm = NULL;
2322 rq->prev_mm = oldmm;
2325 * Since the runqueue lock will be released by the next
2326 * task (which is an invalid locking op but in the case
2327 * of the scheduler it's an obvious special-case), so we
2328 * do an early lockdep release here:
2330 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2332 context_tracking_task_switch(prev, next);
2333 /* Here we just switch the register state and the stack. */
2334 switch_to(prev, next, prev);
2337 return finish_task_switch(prev);
2341 * nr_running and nr_context_switches:
2343 * externally visible scheduler statistics: current number of runnable
2344 * threads, total number of context switches performed since bootup.
2346 unsigned long nr_running(void)
2348 unsigned long i, sum = 0;
2350 for_each_online_cpu(i)
2351 sum += cpu_rq(i)->nr_running;
2357 * Check if only the current task is running on the cpu.
2359 bool single_task_running(void)
2361 if (cpu_rq(smp_processor_id())->nr_running == 1)
2366 EXPORT_SYMBOL(single_task_running);
2368 unsigned long long nr_context_switches(void)
2371 unsigned long long sum = 0;
2373 for_each_possible_cpu(i)
2374 sum += cpu_rq(i)->nr_switches;
2379 unsigned long nr_iowait(void)
2381 unsigned long i, sum = 0;
2383 for_each_possible_cpu(i)
2384 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2389 unsigned long nr_iowait_cpu(int cpu)
2391 struct rq *this = cpu_rq(cpu);
2392 return atomic_read(&this->nr_iowait);
2395 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2397 struct rq *this = this_rq();
2398 *nr_waiters = atomic_read(&this->nr_iowait);
2399 *load = this->cpu_load[0];
2405 * sched_exec - execve() is a valuable balancing opportunity, because at
2406 * this point the task has the smallest effective memory and cache footprint.
2408 void sched_exec(void)
2410 struct task_struct *p = current;
2411 unsigned long flags;
2414 raw_spin_lock_irqsave(&p->pi_lock, flags);
2415 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2416 if (dest_cpu == smp_processor_id())
2419 if (likely(cpu_active(dest_cpu))) {
2420 struct migration_arg arg = { p, dest_cpu };
2422 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2423 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2427 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2432 DEFINE_PER_CPU(struct kernel_stat, kstat);
2433 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2435 EXPORT_PER_CPU_SYMBOL(kstat);
2436 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2439 * Return accounted runtime for the task.
2440 * In case the task is currently running, return the runtime plus current's
2441 * pending runtime that have not been accounted yet.
2443 unsigned long long task_sched_runtime(struct task_struct *p)
2445 unsigned long flags;
2449 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2451 * 64-bit doesn't need locks to atomically read a 64bit value.
2452 * So we have a optimization chance when the task's delta_exec is 0.
2453 * Reading ->on_cpu is racy, but this is ok.
2455 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2456 * If we race with it entering cpu, unaccounted time is 0. This is
2457 * indistinguishable from the read occurring a few cycles earlier.
2458 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2459 * been accounted, so we're correct here as well.
2461 if (!p->on_cpu || !task_on_rq_queued(p))
2462 return p->se.sum_exec_runtime;
2465 rq = task_rq_lock(p, &flags);
2467 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2468 * project cycles that may never be accounted to this
2469 * thread, breaking clock_gettime().
2471 if (task_current(rq, p) && task_on_rq_queued(p)) {
2472 update_rq_clock(rq);
2473 p->sched_class->update_curr(rq);
2475 ns = p->se.sum_exec_runtime;
2476 task_rq_unlock(rq, p, &flags);
2482 * This function gets called by the timer code, with HZ frequency.
2483 * We call it with interrupts disabled.
2485 void scheduler_tick(void)
2487 int cpu = smp_processor_id();
2488 struct rq *rq = cpu_rq(cpu);
2489 struct task_struct *curr = rq->curr;
2493 raw_spin_lock(&rq->lock);
2494 update_rq_clock(rq);
2495 curr->sched_class->task_tick(rq, curr, 0);
2496 update_cpu_load_active(rq);
2497 raw_spin_unlock(&rq->lock);
2499 perf_event_task_tick();
2502 rq->idle_balance = idle_cpu(cpu);
2503 trigger_load_balance(rq);
2505 rq_last_tick_reset(rq);
2508 #ifdef CONFIG_NO_HZ_FULL
2510 * scheduler_tick_max_deferment
2512 * Keep at least one tick per second when a single
2513 * active task is running because the scheduler doesn't
2514 * yet completely support full dynticks environment.
2516 * This makes sure that uptime, CFS vruntime, load
2517 * balancing, etc... continue to move forward, even
2518 * with a very low granularity.
2520 * Return: Maximum deferment in nanoseconds.
2522 u64 scheduler_tick_max_deferment(void)
2524 struct rq *rq = this_rq();
2525 unsigned long next, now = ACCESS_ONCE(jiffies);
2527 next = rq->last_sched_tick + HZ;
2529 if (time_before_eq(next, now))
2532 return jiffies_to_nsecs(next - now);
2536 notrace unsigned long get_parent_ip(unsigned long addr)
2538 if (in_lock_functions(addr)) {
2539 addr = CALLER_ADDR2;
2540 if (in_lock_functions(addr))
2541 addr = CALLER_ADDR3;
2546 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2547 defined(CONFIG_PREEMPT_TRACER))
2549 void preempt_count_add(int val)
2551 #ifdef CONFIG_DEBUG_PREEMPT
2555 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2558 __preempt_count_add(val);
2559 #ifdef CONFIG_DEBUG_PREEMPT
2561 * Spinlock count overflowing soon?
2563 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2566 if (preempt_count() == val) {
2567 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2568 #ifdef CONFIG_DEBUG_PREEMPT
2569 current->preempt_disable_ip = ip;
2571 trace_preempt_off(CALLER_ADDR0, ip);
2574 EXPORT_SYMBOL(preempt_count_add);
2575 NOKPROBE_SYMBOL(preempt_count_add);
2577 void preempt_count_sub(int val)
2579 #ifdef CONFIG_DEBUG_PREEMPT
2583 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2586 * Is the spinlock portion underflowing?
2588 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2589 !(preempt_count() & PREEMPT_MASK)))
2593 if (preempt_count() == val)
2594 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2595 __preempt_count_sub(val);
2597 EXPORT_SYMBOL(preempt_count_sub);
2598 NOKPROBE_SYMBOL(preempt_count_sub);
2603 * Print scheduling while atomic bug:
2605 static noinline void __schedule_bug(struct task_struct *prev)
2607 if (oops_in_progress)
2610 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2611 prev->comm, prev->pid, preempt_count());
2613 debug_show_held_locks(prev);
2615 if (irqs_disabled())
2616 print_irqtrace_events(prev);
2617 #ifdef CONFIG_DEBUG_PREEMPT
2618 if (in_atomic_preempt_off()) {
2619 pr_err("Preemption disabled at:");
2620 print_ip_sym(current->preempt_disable_ip);
2625 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2629 * Various schedule()-time debugging checks and statistics:
2631 static inline void schedule_debug(struct task_struct *prev)
2633 #ifdef CONFIG_SCHED_STACK_END_CHECK
2634 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2637 * Test if we are atomic. Since do_exit() needs to call into
2638 * schedule() atomically, we ignore that path. Otherwise whine
2639 * if we are scheduling when we should not.
2641 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2642 __schedule_bug(prev);
2645 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2647 schedstat_inc(this_rq(), sched_count);
2651 * Pick up the highest-prio task:
2653 static inline struct task_struct *
2654 pick_next_task(struct rq *rq, struct task_struct *prev)
2656 const struct sched_class *class = &fair_sched_class;
2657 struct task_struct *p;
2660 * Optimization: we know that if all tasks are in
2661 * the fair class we can call that function directly:
2663 if (likely(prev->sched_class == class &&
2664 rq->nr_running == rq->cfs.h_nr_running)) {
2665 p = fair_sched_class.pick_next_task(rq, prev);
2666 if (unlikely(p == RETRY_TASK))
2669 /* assumes fair_sched_class->next == idle_sched_class */
2671 p = idle_sched_class.pick_next_task(rq, prev);
2677 for_each_class(class) {
2678 p = class->pick_next_task(rq, prev);
2680 if (unlikely(p == RETRY_TASK))
2686 BUG(); /* the idle class will always have a runnable task */
2690 * __schedule() is the main scheduler function.
2692 * The main means of driving the scheduler and thus entering this function are:
2694 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2696 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2697 * paths. For example, see arch/x86/entry_64.S.
2699 * To drive preemption between tasks, the scheduler sets the flag in timer
2700 * interrupt handler scheduler_tick().
2702 * 3. Wakeups don't really cause entry into schedule(). They add a
2703 * task to the run-queue and that's it.
2705 * Now, if the new task added to the run-queue preempts the current
2706 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2707 * called on the nearest possible occasion:
2709 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2711 * - in syscall or exception context, at the next outmost
2712 * preempt_enable(). (this might be as soon as the wake_up()'s
2715 * - in IRQ context, return from interrupt-handler to
2716 * preemptible context
2718 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2721 * - cond_resched() call
2722 * - explicit schedule() call
2723 * - return from syscall or exception to user-space
2724 * - return from interrupt-handler to user-space
2726 * WARNING: all callers must re-check need_resched() afterward and reschedule
2727 * accordingly in case an event triggered the need for rescheduling (such as
2728 * an interrupt waking up a task) while preemption was disabled in __schedule().
2730 static void __sched __schedule(void)
2732 struct task_struct *prev, *next;
2733 unsigned long *switch_count;
2738 cpu = smp_processor_id();
2740 rcu_note_context_switch();
2743 schedule_debug(prev);
2745 if (sched_feat(HRTICK))
2749 * Make sure that signal_pending_state()->signal_pending() below
2750 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2751 * done by the caller to avoid the race with signal_wake_up().
2753 smp_mb__before_spinlock();
2754 raw_spin_lock_irq(&rq->lock);
2756 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2758 switch_count = &prev->nivcsw;
2759 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2760 if (unlikely(signal_pending_state(prev->state, prev))) {
2761 prev->state = TASK_RUNNING;
2763 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2767 * If a worker went to sleep, notify and ask workqueue
2768 * whether it wants to wake up a task to maintain
2771 if (prev->flags & PF_WQ_WORKER) {
2772 struct task_struct *to_wakeup;
2774 to_wakeup = wq_worker_sleeping(prev, cpu);
2776 try_to_wake_up_local(to_wakeup);
2779 switch_count = &prev->nvcsw;
2782 if (task_on_rq_queued(prev))
2783 update_rq_clock(rq);
2785 next = pick_next_task(rq, prev);
2786 clear_tsk_need_resched(prev);
2787 clear_preempt_need_resched();
2788 rq->clock_skip_update = 0;
2790 if (likely(prev != next)) {
2795 rq = context_switch(rq, prev, next); /* unlocks the rq */
2798 raw_spin_unlock_irq(&rq->lock);
2802 sched_preempt_enable_no_resched();
2805 static inline void sched_submit_work(struct task_struct *tsk)
2807 if (!tsk->state || tsk_is_pi_blocked(tsk))
2810 * If we are going to sleep and we have plugged IO queued,
2811 * make sure to submit it to avoid deadlocks.
2813 if (blk_needs_flush_plug(tsk))
2814 blk_schedule_flush_plug(tsk);
2817 asmlinkage __visible void __sched schedule(void)
2819 struct task_struct *tsk = current;
2821 sched_submit_work(tsk);
2824 } while (need_resched());
2826 EXPORT_SYMBOL(schedule);
2828 #ifdef CONFIG_CONTEXT_TRACKING
2829 asmlinkage __visible void __sched schedule_user(void)
2832 * If we come here after a random call to set_need_resched(),
2833 * or we have been woken up remotely but the IPI has not yet arrived,
2834 * we haven't yet exited the RCU idle mode. Do it here manually until
2835 * we find a better solution.
2837 * NB: There are buggy callers of this function. Ideally we
2838 * should warn if prev_state != IN_USER, but that will trigger
2839 * too frequently to make sense yet.
2841 enum ctx_state prev_state = exception_enter();
2843 exception_exit(prev_state);
2848 * schedule_preempt_disabled - called with preemption disabled
2850 * Returns with preemption disabled. Note: preempt_count must be 1
2852 void __sched schedule_preempt_disabled(void)
2854 sched_preempt_enable_no_resched();
2859 static void __sched notrace preempt_schedule_common(void)
2862 __preempt_count_add(PREEMPT_ACTIVE);
2864 __preempt_count_sub(PREEMPT_ACTIVE);
2867 * Check again in case we missed a preemption opportunity
2868 * between schedule and now.
2871 } while (need_resched());
2874 #ifdef CONFIG_PREEMPT
2876 * this is the entry point to schedule() from in-kernel preemption
2877 * off of preempt_enable. Kernel preemptions off return from interrupt
2878 * occur there and call schedule directly.
2880 asmlinkage __visible void __sched notrace preempt_schedule(void)
2883 * If there is a non-zero preempt_count or interrupts are disabled,
2884 * we do not want to preempt the current task. Just return..
2886 if (likely(!preemptible()))
2889 preempt_schedule_common();
2891 NOKPROBE_SYMBOL(preempt_schedule);
2892 EXPORT_SYMBOL(preempt_schedule);
2894 #ifdef CONFIG_CONTEXT_TRACKING
2896 * preempt_schedule_context - preempt_schedule called by tracing
2898 * The tracing infrastructure uses preempt_enable_notrace to prevent
2899 * recursion and tracing preempt enabling caused by the tracing
2900 * infrastructure itself. But as tracing can happen in areas coming
2901 * from userspace or just about to enter userspace, a preempt enable
2902 * can occur before user_exit() is called. This will cause the scheduler
2903 * to be called when the system is still in usermode.
2905 * To prevent this, the preempt_enable_notrace will use this function
2906 * instead of preempt_schedule() to exit user context if needed before
2907 * calling the scheduler.
2909 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2911 enum ctx_state prev_ctx;
2913 if (likely(!preemptible()))
2917 __preempt_count_add(PREEMPT_ACTIVE);
2919 * Needs preempt disabled in case user_exit() is traced
2920 * and the tracer calls preempt_enable_notrace() causing
2921 * an infinite recursion.
2923 prev_ctx = exception_enter();
2925 exception_exit(prev_ctx);
2927 __preempt_count_sub(PREEMPT_ACTIVE);
2929 } while (need_resched());
2931 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2932 #endif /* CONFIG_CONTEXT_TRACKING */
2934 #endif /* CONFIG_PREEMPT */
2937 * this is the entry point to schedule() from kernel preemption
2938 * off of irq context.
2939 * Note, that this is called and return with irqs disabled. This will
2940 * protect us against recursive calling from irq.
2942 asmlinkage __visible void __sched preempt_schedule_irq(void)
2944 enum ctx_state prev_state;
2946 /* Catch callers which need to be fixed */
2947 BUG_ON(preempt_count() || !irqs_disabled());
2949 prev_state = exception_enter();
2952 __preempt_count_add(PREEMPT_ACTIVE);
2955 local_irq_disable();
2956 __preempt_count_sub(PREEMPT_ACTIVE);
2959 * Check again in case we missed a preemption opportunity
2960 * between schedule and now.
2963 } while (need_resched());
2965 exception_exit(prev_state);
2968 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2971 return try_to_wake_up(curr->private, mode, wake_flags);
2973 EXPORT_SYMBOL(default_wake_function);
2975 #ifdef CONFIG_RT_MUTEXES
2978 * rt_mutex_setprio - set the current priority of a task
2980 * @prio: prio value (kernel-internal form)
2982 * This function changes the 'effective' priority of a task. It does
2983 * not touch ->normal_prio like __setscheduler().
2985 * Used by the rt_mutex code to implement priority inheritance
2986 * logic. Call site only calls if the priority of the task changed.
2988 void rt_mutex_setprio(struct task_struct *p, int prio)
2990 int oldprio, queued, running, enqueue_flag = 0;
2992 const struct sched_class *prev_class;
2994 BUG_ON(prio > MAX_PRIO);
2996 rq = __task_rq_lock(p);
2999 * Idle task boosting is a nono in general. There is one
3000 * exception, when PREEMPT_RT and NOHZ is active:
3002 * The idle task calls get_next_timer_interrupt() and holds
3003 * the timer wheel base->lock on the CPU and another CPU wants
3004 * to access the timer (probably to cancel it). We can safely
3005 * ignore the boosting request, as the idle CPU runs this code
3006 * with interrupts disabled and will complete the lock
3007 * protected section without being interrupted. So there is no
3008 * real need to boost.
3010 if (unlikely(p == rq->idle)) {
3011 WARN_ON(p != rq->curr);
3012 WARN_ON(p->pi_blocked_on);
3016 trace_sched_pi_setprio(p, prio);
3018 prev_class = p->sched_class;
3019 queued = task_on_rq_queued(p);
3020 running = task_current(rq, p);
3022 dequeue_task(rq, p, 0);
3024 put_prev_task(rq, p);
3027 * Boosting condition are:
3028 * 1. -rt task is running and holds mutex A
3029 * --> -dl task blocks on mutex A
3031 * 2. -dl task is running and holds mutex A
3032 * --> -dl task blocks on mutex A and could preempt the
3035 if (dl_prio(prio)) {
3036 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3037 if (!dl_prio(p->normal_prio) ||
3038 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3039 p->dl.dl_boosted = 1;
3040 p->dl.dl_throttled = 0;
3041 enqueue_flag = ENQUEUE_REPLENISH;
3043 p->dl.dl_boosted = 0;
3044 p->sched_class = &dl_sched_class;
3045 } else if (rt_prio(prio)) {
3046 if (dl_prio(oldprio))
3047 p->dl.dl_boosted = 0;
3049 enqueue_flag = ENQUEUE_HEAD;
3050 p->sched_class = &rt_sched_class;
3052 if (dl_prio(oldprio))
3053 p->dl.dl_boosted = 0;
3054 if (rt_prio(oldprio))
3056 p->sched_class = &fair_sched_class;
3062 p->sched_class->set_curr_task(rq);
3064 enqueue_task(rq, p, enqueue_flag);
3066 check_class_changed(rq, p, prev_class, oldprio);
3068 __task_rq_unlock(rq);
3072 void set_user_nice(struct task_struct *p, long nice)
3074 int old_prio, delta, queued;
3075 unsigned long flags;
3078 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3081 * We have to be careful, if called from sys_setpriority(),
3082 * the task might be in the middle of scheduling on another CPU.
3084 rq = task_rq_lock(p, &flags);
3086 * The RT priorities are set via sched_setscheduler(), but we still
3087 * allow the 'normal' nice value to be set - but as expected
3088 * it wont have any effect on scheduling until the task is
3089 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3091 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3092 p->static_prio = NICE_TO_PRIO(nice);
3095 queued = task_on_rq_queued(p);
3097 dequeue_task(rq, p, 0);
3099 p->static_prio = NICE_TO_PRIO(nice);
3102 p->prio = effective_prio(p);
3103 delta = p->prio - old_prio;
3106 enqueue_task(rq, p, 0);
3108 * If the task increased its priority or is running and
3109 * lowered its priority, then reschedule its CPU:
3111 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3115 task_rq_unlock(rq, p, &flags);
3117 EXPORT_SYMBOL(set_user_nice);
3120 * can_nice - check if a task can reduce its nice value
3124 int can_nice(const struct task_struct *p, const int nice)
3126 /* convert nice value [19,-20] to rlimit style value [1,40] */
3127 int nice_rlim = nice_to_rlimit(nice);
3129 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3130 capable(CAP_SYS_NICE));
3133 #ifdef __ARCH_WANT_SYS_NICE
3136 * sys_nice - change the priority of the current process.
3137 * @increment: priority increment
3139 * sys_setpriority is a more generic, but much slower function that
3140 * does similar things.
3142 SYSCALL_DEFINE1(nice, int, increment)
3147 * Setpriority might change our priority at the same moment.
3148 * We don't have to worry. Conceptually one call occurs first
3149 * and we have a single winner.
3151 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3152 nice = task_nice(current) + increment;
3154 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3155 if (increment < 0 && !can_nice(current, nice))
3158 retval = security_task_setnice(current, nice);
3162 set_user_nice(current, nice);
3169 * task_prio - return the priority value of a given task.
3170 * @p: the task in question.
3172 * Return: The priority value as seen by users in /proc.
3173 * RT tasks are offset by -200. Normal tasks are centered
3174 * around 0, value goes from -16 to +15.
3176 int task_prio(const struct task_struct *p)
3178 return p->prio - MAX_RT_PRIO;
3182 * idle_cpu - is a given cpu idle currently?
3183 * @cpu: the processor in question.
3185 * Return: 1 if the CPU is currently idle. 0 otherwise.
3187 int idle_cpu(int cpu)
3189 struct rq *rq = cpu_rq(cpu);
3191 if (rq->curr != rq->idle)
3198 if (!llist_empty(&rq->wake_list))
3206 * idle_task - return the idle task for a given cpu.
3207 * @cpu: the processor in question.
3209 * Return: The idle task for the cpu @cpu.
3211 struct task_struct *idle_task(int cpu)
3213 return cpu_rq(cpu)->idle;
3217 * find_process_by_pid - find a process with a matching PID value.
3218 * @pid: the pid in question.
3220 * The task of @pid, if found. %NULL otherwise.
3222 static struct task_struct *find_process_by_pid(pid_t pid)
3224 return pid ? find_task_by_vpid(pid) : current;
3228 * This function initializes the sched_dl_entity of a newly becoming
3229 * SCHED_DEADLINE task.
3231 * Only the static values are considered here, the actual runtime and the
3232 * absolute deadline will be properly calculated when the task is enqueued
3233 * for the first time with its new policy.
3236 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3238 struct sched_dl_entity *dl_se = &p->dl;
3240 dl_se->dl_runtime = attr->sched_runtime;
3241 dl_se->dl_deadline = attr->sched_deadline;
3242 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3243 dl_se->flags = attr->sched_flags;
3244 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3247 * Changing the parameters of a task is 'tricky' and we're not doing
3248 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3250 * What we SHOULD do is delay the bandwidth release until the 0-lag
3251 * point. This would include retaining the task_struct until that time
3252 * and change dl_overflow() to not immediately decrement the current
3255 * Instead we retain the current runtime/deadline and let the new
3256 * parameters take effect after the current reservation period lapses.
3257 * This is safe (albeit pessimistic) because the 0-lag point is always
3258 * before the current scheduling deadline.
3260 * We can still have temporary overloads because we do not delay the
3261 * change in bandwidth until that time; so admission control is
3262 * not on the safe side. It does however guarantee tasks will never
3263 * consume more than promised.
3268 * sched_setparam() passes in -1 for its policy, to let the functions
3269 * it calls know not to change it.
3271 #define SETPARAM_POLICY -1
3273 static void __setscheduler_params(struct task_struct *p,
3274 const struct sched_attr *attr)
3276 int policy = attr->sched_policy;
3278 if (policy == SETPARAM_POLICY)
3283 if (dl_policy(policy))
3284 __setparam_dl(p, attr);
3285 else if (fair_policy(policy))
3286 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3289 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3290 * !rt_policy. Always setting this ensures that things like
3291 * getparam()/getattr() don't report silly values for !rt tasks.
3293 p->rt_priority = attr->sched_priority;
3294 p->normal_prio = normal_prio(p);
3298 /* Actually do priority change: must hold pi & rq lock. */
3299 static void __setscheduler(struct rq *rq, struct task_struct *p,
3300 const struct sched_attr *attr)
3302 __setscheduler_params(p, attr);
3305 * If we get here, there was no pi waiters boosting the
3306 * task. It is safe to use the normal prio.
3308 p->prio = normal_prio(p);
3310 if (dl_prio(p->prio))
3311 p->sched_class = &dl_sched_class;
3312 else if (rt_prio(p->prio))
3313 p->sched_class = &rt_sched_class;
3315 p->sched_class = &fair_sched_class;
3319 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3321 struct sched_dl_entity *dl_se = &p->dl;
3323 attr->sched_priority = p->rt_priority;
3324 attr->sched_runtime = dl_se->dl_runtime;
3325 attr->sched_deadline = dl_se->dl_deadline;
3326 attr->sched_period = dl_se->dl_period;
3327 attr->sched_flags = dl_se->flags;
3331 * This function validates the new parameters of a -deadline task.
3332 * We ask for the deadline not being zero, and greater or equal
3333 * than the runtime, as well as the period of being zero or
3334 * greater than deadline. Furthermore, we have to be sure that
3335 * user parameters are above the internal resolution of 1us (we
3336 * check sched_runtime only since it is always the smaller one) and
3337 * below 2^63 ns (we have to check both sched_deadline and
3338 * sched_period, as the latter can be zero).
3341 __checkparam_dl(const struct sched_attr *attr)
3344 if (attr->sched_deadline == 0)
3348 * Since we truncate DL_SCALE bits, make sure we're at least
3351 if (attr->sched_runtime < (1ULL << DL_SCALE))
3355 * Since we use the MSB for wrap-around and sign issues, make
3356 * sure it's not set (mind that period can be equal to zero).
3358 if (attr->sched_deadline & (1ULL << 63) ||
3359 attr->sched_period & (1ULL << 63))
3362 /* runtime <= deadline <= period (if period != 0) */
3363 if ((attr->sched_period != 0 &&
3364 attr->sched_period < attr->sched_deadline) ||
3365 attr->sched_deadline < attr->sched_runtime)
3372 * check the target process has a UID that matches the current process's
3374 static bool check_same_owner(struct task_struct *p)
3376 const struct cred *cred = current_cred(), *pcred;
3380 pcred = __task_cred(p);
3381 match = (uid_eq(cred->euid, pcred->euid) ||
3382 uid_eq(cred->euid, pcred->uid));
3387 static bool dl_param_changed(struct task_struct *p,
3388 const struct sched_attr *attr)
3390 struct sched_dl_entity *dl_se = &p->dl;
3392 if (dl_se->dl_runtime != attr->sched_runtime ||
3393 dl_se->dl_deadline != attr->sched_deadline ||
3394 dl_se->dl_period != attr->sched_period ||
3395 dl_se->flags != attr->sched_flags)
3401 static int __sched_setscheduler(struct task_struct *p,
3402 const struct sched_attr *attr,
3405 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3406 MAX_RT_PRIO - 1 - attr->sched_priority;
3407 int retval, oldprio, oldpolicy = -1, queued, running;
3408 int policy = attr->sched_policy;
3409 unsigned long flags;
3410 const struct sched_class *prev_class;
3414 /* may grab non-irq protected spin_locks */
3415 BUG_ON(in_interrupt());
3417 /* double check policy once rq lock held */
3419 reset_on_fork = p->sched_reset_on_fork;
3420 policy = oldpolicy = p->policy;
3422 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3424 if (policy != SCHED_DEADLINE &&
3425 policy != SCHED_FIFO && policy != SCHED_RR &&
3426 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3427 policy != SCHED_IDLE)
3431 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3435 * Valid priorities for SCHED_FIFO and SCHED_RR are
3436 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3437 * SCHED_BATCH and SCHED_IDLE is 0.
3439 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3440 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3442 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3443 (rt_policy(policy) != (attr->sched_priority != 0)))
3447 * Allow unprivileged RT tasks to decrease priority:
3449 if (user && !capable(CAP_SYS_NICE)) {
3450 if (fair_policy(policy)) {
3451 if (attr->sched_nice < task_nice(p) &&
3452 !can_nice(p, attr->sched_nice))
3456 if (rt_policy(policy)) {
3457 unsigned long rlim_rtprio =
3458 task_rlimit(p, RLIMIT_RTPRIO);
3460 /* can't set/change the rt policy */
3461 if (policy != p->policy && !rlim_rtprio)
3464 /* can't increase priority */
3465 if (attr->sched_priority > p->rt_priority &&
3466 attr->sched_priority > rlim_rtprio)
3471 * Can't set/change SCHED_DEADLINE policy at all for now
3472 * (safest behavior); in the future we would like to allow
3473 * unprivileged DL tasks to increase their relative deadline
3474 * or reduce their runtime (both ways reducing utilization)
3476 if (dl_policy(policy))
3480 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3481 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3483 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3484 if (!can_nice(p, task_nice(p)))
3488 /* can't change other user's priorities */
3489 if (!check_same_owner(p))
3492 /* Normal users shall not reset the sched_reset_on_fork flag */
3493 if (p->sched_reset_on_fork && !reset_on_fork)
3498 retval = security_task_setscheduler(p);
3504 * make sure no PI-waiters arrive (or leave) while we are
3505 * changing the priority of the task:
3507 * To be able to change p->policy safely, the appropriate
3508 * runqueue lock must be held.
3510 rq = task_rq_lock(p, &flags);
3513 * Changing the policy of the stop threads its a very bad idea
3515 if (p == rq->stop) {
3516 task_rq_unlock(rq, p, &flags);
3521 * If not changing anything there's no need to proceed further,
3522 * but store a possible modification of reset_on_fork.
3524 if (unlikely(policy == p->policy)) {
3525 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3527 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3529 if (dl_policy(policy) && dl_param_changed(p, attr))
3532 p->sched_reset_on_fork = reset_on_fork;
3533 task_rq_unlock(rq, p, &flags);
3539 #ifdef CONFIG_RT_GROUP_SCHED
3541 * Do not allow realtime tasks into groups that have no runtime
3544 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3545 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3546 !task_group_is_autogroup(task_group(p))) {
3547 task_rq_unlock(rq, p, &flags);
3552 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3553 cpumask_t *span = rq->rd->span;
3556 * Don't allow tasks with an affinity mask smaller than
3557 * the entire root_domain to become SCHED_DEADLINE. We
3558 * will also fail if there's no bandwidth available.
3560 if (!cpumask_subset(span, &p->cpus_allowed) ||
3561 rq->rd->dl_bw.bw == 0) {
3562 task_rq_unlock(rq, p, &flags);
3569 /* recheck policy now with rq lock held */
3570 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3571 policy = oldpolicy = -1;
3572 task_rq_unlock(rq, p, &flags);
3577 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3578 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3581 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3582 task_rq_unlock(rq, p, &flags);
3586 p->sched_reset_on_fork = reset_on_fork;
3590 * Special case for priority boosted tasks.
3592 * If the new priority is lower or equal (user space view)
3593 * than the current (boosted) priority, we just store the new
3594 * normal parameters and do not touch the scheduler class and
3595 * the runqueue. This will be done when the task deboost
3598 if (rt_mutex_check_prio(p, newprio)) {
3599 __setscheduler_params(p, attr);
3600 task_rq_unlock(rq, p, &flags);
3604 queued = task_on_rq_queued(p);
3605 running = task_current(rq, p);
3607 dequeue_task(rq, p, 0);
3609 put_prev_task(rq, p);
3611 prev_class = p->sched_class;
3612 __setscheduler(rq, p, attr);
3615 p->sched_class->set_curr_task(rq);
3618 * We enqueue to tail when the priority of a task is
3619 * increased (user space view).
3621 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3624 check_class_changed(rq, p, prev_class, oldprio);
3625 task_rq_unlock(rq, p, &flags);
3627 rt_mutex_adjust_pi(p);
3632 static int _sched_setscheduler(struct task_struct *p, int policy,
3633 const struct sched_param *param, bool check)
3635 struct sched_attr attr = {
3636 .sched_policy = policy,
3637 .sched_priority = param->sched_priority,
3638 .sched_nice = PRIO_TO_NICE(p->static_prio),
3641 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3642 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3643 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3644 policy &= ~SCHED_RESET_ON_FORK;
3645 attr.sched_policy = policy;
3648 return __sched_setscheduler(p, &attr, check);
3651 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3652 * @p: the task in question.
3653 * @policy: new policy.
3654 * @param: structure containing the new RT priority.
3656 * Return: 0 on success. An error code otherwise.
3658 * NOTE that the task may be already dead.
3660 int sched_setscheduler(struct task_struct *p, int policy,
3661 const struct sched_param *param)
3663 return _sched_setscheduler(p, policy, param, true);
3665 EXPORT_SYMBOL_GPL(sched_setscheduler);
3667 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3669 return __sched_setscheduler(p, attr, true);
3671 EXPORT_SYMBOL_GPL(sched_setattr);
3674 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3675 * @p: the task in question.
3676 * @policy: new policy.
3677 * @param: structure containing the new RT priority.
3679 * Just like sched_setscheduler, only don't bother checking if the
3680 * current context has permission. For example, this is needed in
3681 * stop_machine(): we create temporary high priority worker threads,
3682 * but our caller might not have that capability.
3684 * Return: 0 on success. An error code otherwise.
3686 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3687 const struct sched_param *param)
3689 return _sched_setscheduler(p, policy, param, false);
3693 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3695 struct sched_param lparam;
3696 struct task_struct *p;
3699 if (!param || pid < 0)
3701 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3706 p = find_process_by_pid(pid);
3708 retval = sched_setscheduler(p, policy, &lparam);
3715 * Mimics kernel/events/core.c perf_copy_attr().
3717 static int sched_copy_attr(struct sched_attr __user *uattr,
3718 struct sched_attr *attr)
3723 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3727 * zero the full structure, so that a short copy will be nice.
3729 memset(attr, 0, sizeof(*attr));
3731 ret = get_user(size, &uattr->size);
3735 if (size > PAGE_SIZE) /* silly large */
3738 if (!size) /* abi compat */
3739 size = SCHED_ATTR_SIZE_VER0;
3741 if (size < SCHED_ATTR_SIZE_VER0)
3745 * If we're handed a bigger struct than we know of,
3746 * ensure all the unknown bits are 0 - i.e. new
3747 * user-space does not rely on any kernel feature
3748 * extensions we dont know about yet.
3750 if (size > sizeof(*attr)) {
3751 unsigned char __user *addr;
3752 unsigned char __user *end;
3755 addr = (void __user *)uattr + sizeof(*attr);
3756 end = (void __user *)uattr + size;
3758 for (; addr < end; addr++) {
3759 ret = get_user(val, addr);
3765 size = sizeof(*attr);
3768 ret = copy_from_user(attr, uattr, size);
3773 * XXX: do we want to be lenient like existing syscalls; or do we want
3774 * to be strict and return an error on out-of-bounds values?
3776 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3781 put_user(sizeof(*attr), &uattr->size);
3786 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3787 * @pid: the pid in question.
3788 * @policy: new policy.
3789 * @param: structure containing the new RT priority.
3791 * Return: 0 on success. An error code otherwise.
3793 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3794 struct sched_param __user *, param)
3796 /* negative values for policy are not valid */
3800 return do_sched_setscheduler(pid, policy, param);
3804 * sys_sched_setparam - set/change the RT priority of a thread
3805 * @pid: the pid in question.
3806 * @param: structure containing the new RT priority.
3808 * Return: 0 on success. An error code otherwise.
3810 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3812 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3816 * sys_sched_setattr - same as above, but with extended sched_attr
3817 * @pid: the pid in question.
3818 * @uattr: structure containing the extended parameters.
3819 * @flags: for future extension.
3821 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3822 unsigned int, flags)
3824 struct sched_attr attr;
3825 struct task_struct *p;
3828 if (!uattr || pid < 0 || flags)
3831 retval = sched_copy_attr(uattr, &attr);
3835 if ((int)attr.sched_policy < 0)
3840 p = find_process_by_pid(pid);
3842 retval = sched_setattr(p, &attr);
3849 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3850 * @pid: the pid in question.
3852 * Return: On success, the policy of the thread. Otherwise, a negative error
3855 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3857 struct task_struct *p;
3865 p = find_process_by_pid(pid);
3867 retval = security_task_getscheduler(p);
3870 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3877 * sys_sched_getparam - get the RT priority of a thread
3878 * @pid: the pid in question.
3879 * @param: structure containing the RT priority.
3881 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3884 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3886 struct sched_param lp = { .sched_priority = 0 };
3887 struct task_struct *p;
3890 if (!param || pid < 0)
3894 p = find_process_by_pid(pid);
3899 retval = security_task_getscheduler(p);
3903 if (task_has_rt_policy(p))
3904 lp.sched_priority = p->rt_priority;
3908 * This one might sleep, we cannot do it with a spinlock held ...
3910 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3919 static int sched_read_attr(struct sched_attr __user *uattr,
3920 struct sched_attr *attr,
3925 if (!access_ok(VERIFY_WRITE, uattr, usize))
3929 * If we're handed a smaller struct than we know of,
3930 * ensure all the unknown bits are 0 - i.e. old
3931 * user-space does not get uncomplete information.
3933 if (usize < sizeof(*attr)) {
3934 unsigned char *addr;
3937 addr = (void *)attr + usize;
3938 end = (void *)attr + sizeof(*attr);
3940 for (; addr < end; addr++) {
3948 ret = copy_to_user(uattr, attr, attr->size);
3956 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3957 * @pid: the pid in question.
3958 * @uattr: structure containing the extended parameters.
3959 * @size: sizeof(attr) for fwd/bwd comp.
3960 * @flags: for future extension.
3962 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3963 unsigned int, size, unsigned int, flags)
3965 struct sched_attr attr = {
3966 .size = sizeof(struct sched_attr),
3968 struct task_struct *p;
3971 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3972 size < SCHED_ATTR_SIZE_VER0 || flags)
3976 p = find_process_by_pid(pid);
3981 retval = security_task_getscheduler(p);
3985 attr.sched_policy = p->policy;
3986 if (p->sched_reset_on_fork)
3987 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3988 if (task_has_dl_policy(p))
3989 __getparam_dl(p, &attr);
3990 else if (task_has_rt_policy(p))
3991 attr.sched_priority = p->rt_priority;
3993 attr.sched_nice = task_nice(p);
3997 retval = sched_read_attr(uattr, &attr, size);
4005 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4007 cpumask_var_t cpus_allowed, new_mask;
4008 struct task_struct *p;
4013 p = find_process_by_pid(pid);
4019 /* Prevent p going away */
4023 if (p->flags & PF_NO_SETAFFINITY) {
4027 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4031 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4033 goto out_free_cpus_allowed;
4036 if (!check_same_owner(p)) {
4038 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4040 goto out_free_new_mask;
4045 retval = security_task_setscheduler(p);
4047 goto out_free_new_mask;
4050 cpuset_cpus_allowed(p, cpus_allowed);
4051 cpumask_and(new_mask, in_mask, cpus_allowed);
4054 * Since bandwidth control happens on root_domain basis,
4055 * if admission test is enabled, we only admit -deadline
4056 * tasks allowed to run on all the CPUs in the task's
4060 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4062 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4065 goto out_free_new_mask;
4071 retval = set_cpus_allowed_ptr(p, new_mask);
4074 cpuset_cpus_allowed(p, cpus_allowed);
4075 if (!cpumask_subset(new_mask, cpus_allowed)) {
4077 * We must have raced with a concurrent cpuset
4078 * update. Just reset the cpus_allowed to the
4079 * cpuset's cpus_allowed
4081 cpumask_copy(new_mask, cpus_allowed);
4086 free_cpumask_var(new_mask);
4087 out_free_cpus_allowed:
4088 free_cpumask_var(cpus_allowed);
4094 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4095 struct cpumask *new_mask)
4097 if (len < cpumask_size())
4098 cpumask_clear(new_mask);
4099 else if (len > cpumask_size())
4100 len = cpumask_size();
4102 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4106 * sys_sched_setaffinity - set the cpu affinity of a process
4107 * @pid: pid of the process
4108 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4109 * @user_mask_ptr: user-space pointer to the new cpu mask
4111 * Return: 0 on success. An error code otherwise.
4113 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4114 unsigned long __user *, user_mask_ptr)
4116 cpumask_var_t new_mask;
4119 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4122 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4124 retval = sched_setaffinity(pid, new_mask);
4125 free_cpumask_var(new_mask);
4129 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4131 struct task_struct *p;
4132 unsigned long flags;
4138 p = find_process_by_pid(pid);
4142 retval = security_task_getscheduler(p);
4146 raw_spin_lock_irqsave(&p->pi_lock, flags);
4147 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4148 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4157 * sys_sched_getaffinity - get the cpu affinity of a process
4158 * @pid: pid of the process
4159 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4160 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4162 * Return: 0 on success. An error code otherwise.
4164 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4165 unsigned long __user *, user_mask_ptr)
4170 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4172 if (len & (sizeof(unsigned long)-1))
4175 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4178 ret = sched_getaffinity(pid, mask);
4180 size_t retlen = min_t(size_t, len, cpumask_size());
4182 if (copy_to_user(user_mask_ptr, mask, retlen))
4187 free_cpumask_var(mask);
4193 * sys_sched_yield - yield the current processor to other threads.
4195 * This function yields the current CPU to other tasks. If there are no
4196 * other threads running on this CPU then this function will return.
4200 SYSCALL_DEFINE0(sched_yield)
4202 struct rq *rq = this_rq_lock();
4204 schedstat_inc(rq, yld_count);
4205 current->sched_class->yield_task(rq);
4208 * Since we are going to call schedule() anyway, there's
4209 * no need to preempt or enable interrupts:
4211 __release(rq->lock);
4212 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4213 do_raw_spin_unlock(&rq->lock);
4214 sched_preempt_enable_no_resched();
4221 int __sched _cond_resched(void)
4223 if (should_resched()) {
4224 preempt_schedule_common();
4229 EXPORT_SYMBOL(_cond_resched);
4232 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4233 * call schedule, and on return reacquire the lock.
4235 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4236 * operations here to prevent schedule() from being called twice (once via
4237 * spin_unlock(), once by hand).
4239 int __cond_resched_lock(spinlock_t *lock)
4241 int resched = should_resched();
4244 lockdep_assert_held(lock);
4246 if (spin_needbreak(lock) || resched) {
4249 preempt_schedule_common();
4257 EXPORT_SYMBOL(__cond_resched_lock);
4259 int __sched __cond_resched_softirq(void)
4261 BUG_ON(!in_softirq());
4263 if (should_resched()) {
4265 preempt_schedule_common();
4271 EXPORT_SYMBOL(__cond_resched_softirq);
4274 * yield - yield the current processor to other threads.
4276 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4278 * The scheduler is at all times free to pick the calling task as the most
4279 * eligible task to run, if removing the yield() call from your code breaks
4280 * it, its already broken.
4282 * Typical broken usage is:
4287 * where one assumes that yield() will let 'the other' process run that will
4288 * make event true. If the current task is a SCHED_FIFO task that will never
4289 * happen. Never use yield() as a progress guarantee!!
4291 * If you want to use yield() to wait for something, use wait_event().
4292 * If you want to use yield() to be 'nice' for others, use cond_resched().
4293 * If you still want to use yield(), do not!
4295 void __sched yield(void)
4297 set_current_state(TASK_RUNNING);
4300 EXPORT_SYMBOL(yield);
4303 * yield_to - yield the current processor to another thread in
4304 * your thread group, or accelerate that thread toward the
4305 * processor it's on.
4307 * @preempt: whether task preemption is allowed or not
4309 * It's the caller's job to ensure that the target task struct
4310 * can't go away on us before we can do any checks.
4313 * true (>0) if we indeed boosted the target task.
4314 * false (0) if we failed to boost the target.
4315 * -ESRCH if there's no task to yield to.
4317 int __sched yield_to(struct task_struct *p, bool preempt)
4319 struct task_struct *curr = current;
4320 struct rq *rq, *p_rq;
4321 unsigned long flags;
4324 local_irq_save(flags);
4330 * If we're the only runnable task on the rq and target rq also
4331 * has only one task, there's absolutely no point in yielding.
4333 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4338 double_rq_lock(rq, p_rq);
4339 if (task_rq(p) != p_rq) {
4340 double_rq_unlock(rq, p_rq);
4344 if (!curr->sched_class->yield_to_task)
4347 if (curr->sched_class != p->sched_class)
4350 if (task_running(p_rq, p) || p->state)
4353 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4355 schedstat_inc(rq, yld_count);
4357 * Make p's CPU reschedule; pick_next_entity takes care of
4360 if (preempt && rq != p_rq)
4365 double_rq_unlock(rq, p_rq);
4367 local_irq_restore(flags);
4374 EXPORT_SYMBOL_GPL(yield_to);
4377 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4378 * that process accounting knows that this is a task in IO wait state.
4380 long __sched io_schedule_timeout(long timeout)
4382 int old_iowait = current->in_iowait;
4386 current->in_iowait = 1;
4388 blk_schedule_flush_plug(current);
4390 blk_flush_plug(current);
4392 delayacct_blkio_start();
4394 atomic_inc(&rq->nr_iowait);
4395 ret = schedule_timeout(timeout);
4396 current->in_iowait = old_iowait;
4397 atomic_dec(&rq->nr_iowait);
4398 delayacct_blkio_end();
4402 EXPORT_SYMBOL(io_schedule_timeout);
4405 * sys_sched_get_priority_max - return maximum RT priority.
4406 * @policy: scheduling class.
4408 * Return: On success, this syscall returns the maximum
4409 * rt_priority that can be used by a given scheduling class.
4410 * On failure, a negative error code is returned.
4412 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4419 ret = MAX_USER_RT_PRIO-1;
4421 case SCHED_DEADLINE:
4432 * sys_sched_get_priority_min - return minimum RT priority.
4433 * @policy: scheduling class.
4435 * Return: On success, this syscall returns the minimum
4436 * rt_priority that can be used by a given scheduling class.
4437 * On failure, a negative error code is returned.
4439 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4448 case SCHED_DEADLINE:
4458 * sys_sched_rr_get_interval - return the default timeslice of a process.
4459 * @pid: pid of the process.
4460 * @interval: userspace pointer to the timeslice value.
4462 * this syscall writes the default timeslice value of a given process
4463 * into the user-space timespec buffer. A value of '0' means infinity.
4465 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4468 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4469 struct timespec __user *, interval)
4471 struct task_struct *p;
4472 unsigned int time_slice;
4473 unsigned long flags;
4483 p = find_process_by_pid(pid);
4487 retval = security_task_getscheduler(p);
4491 rq = task_rq_lock(p, &flags);
4493 if (p->sched_class->get_rr_interval)
4494 time_slice = p->sched_class->get_rr_interval(rq, p);
4495 task_rq_unlock(rq, p, &flags);
4498 jiffies_to_timespec(time_slice, &t);
4499 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4507 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4509 void sched_show_task(struct task_struct *p)
4511 unsigned long free = 0;
4513 unsigned long state = p->state;
4516 state = __ffs(state) + 1;
4517 printk(KERN_INFO "%-15.15s %c", p->comm,
4518 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4519 #if BITS_PER_LONG == 32
4520 if (state == TASK_RUNNING)
4521 printk(KERN_CONT " running ");
4523 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4525 if (state == TASK_RUNNING)
4526 printk(KERN_CONT " running task ");
4528 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4530 #ifdef CONFIG_DEBUG_STACK_USAGE
4531 free = stack_not_used(p);
4536 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4538 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4539 task_pid_nr(p), ppid,
4540 (unsigned long)task_thread_info(p)->flags);
4542 print_worker_info(KERN_INFO, p);
4543 show_stack(p, NULL);
4546 void show_state_filter(unsigned long state_filter)
4548 struct task_struct *g, *p;
4550 #if BITS_PER_LONG == 32
4552 " task PC stack pid father\n");
4555 " task PC stack pid father\n");
4558 for_each_process_thread(g, p) {
4560 * reset the NMI-timeout, listing all files on a slow
4561 * console might take a lot of time:
4563 touch_nmi_watchdog();
4564 if (!state_filter || (p->state & state_filter))
4568 touch_all_softlockup_watchdogs();
4570 #ifdef CONFIG_SCHED_DEBUG
4571 sysrq_sched_debug_show();
4575 * Only show locks if all tasks are dumped:
4578 debug_show_all_locks();
4581 void init_idle_bootup_task(struct task_struct *idle)
4583 idle->sched_class = &idle_sched_class;
4587 * init_idle - set up an idle thread for a given CPU
4588 * @idle: task in question
4589 * @cpu: cpu the idle task belongs to
4591 * NOTE: this function does not set the idle thread's NEED_RESCHED
4592 * flag, to make booting more robust.
4594 void init_idle(struct task_struct *idle, int cpu)
4596 struct rq *rq = cpu_rq(cpu);
4597 unsigned long flags;
4599 raw_spin_lock_irqsave(&rq->lock, flags);
4601 __sched_fork(0, idle);
4602 idle->state = TASK_RUNNING;
4603 idle->se.exec_start = sched_clock();
4605 do_set_cpus_allowed(idle, cpumask_of(cpu));
4607 * We're having a chicken and egg problem, even though we are
4608 * holding rq->lock, the cpu isn't yet set to this cpu so the
4609 * lockdep check in task_group() will fail.
4611 * Similar case to sched_fork(). / Alternatively we could
4612 * use task_rq_lock() here and obtain the other rq->lock.
4617 __set_task_cpu(idle, cpu);
4620 rq->curr = rq->idle = idle;
4621 idle->on_rq = TASK_ON_RQ_QUEUED;
4622 #if defined(CONFIG_SMP)
4625 raw_spin_unlock_irqrestore(&rq->lock, flags);
4627 /* Set the preempt count _outside_ the spinlocks! */
4628 init_idle_preempt_count(idle, cpu);
4631 * The idle tasks have their own, simple scheduling class:
4633 idle->sched_class = &idle_sched_class;
4634 ftrace_graph_init_idle_task(idle, cpu);
4635 vtime_init_idle(idle, cpu);
4636 #if defined(CONFIG_SMP)
4637 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4641 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4642 const struct cpumask *trial)
4644 int ret = 1, trial_cpus;
4645 struct dl_bw *cur_dl_b;
4646 unsigned long flags;
4648 if (!cpumask_weight(cur))
4651 rcu_read_lock_sched();
4652 cur_dl_b = dl_bw_of(cpumask_any(cur));
4653 trial_cpus = cpumask_weight(trial);
4655 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4656 if (cur_dl_b->bw != -1 &&
4657 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4659 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4660 rcu_read_unlock_sched();
4665 int task_can_attach(struct task_struct *p,
4666 const struct cpumask *cs_cpus_allowed)
4671 * Kthreads which disallow setaffinity shouldn't be moved
4672 * to a new cpuset; we don't want to change their cpu
4673 * affinity and isolating such threads by their set of
4674 * allowed nodes is unnecessary. Thus, cpusets are not
4675 * applicable for such threads. This prevents checking for
4676 * success of set_cpus_allowed_ptr() on all attached tasks
4677 * before cpus_allowed may be changed.
4679 if (p->flags & PF_NO_SETAFFINITY) {
4685 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4687 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4692 unsigned long flags;
4694 rcu_read_lock_sched();
4695 dl_b = dl_bw_of(dest_cpu);
4696 raw_spin_lock_irqsave(&dl_b->lock, flags);
4697 cpus = dl_bw_cpus(dest_cpu);
4698 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4703 * We reserve space for this task in the destination
4704 * root_domain, as we can't fail after this point.
4705 * We will free resources in the source root_domain
4706 * later on (see set_cpus_allowed_dl()).
4708 __dl_add(dl_b, p->dl.dl_bw);
4710 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4711 rcu_read_unlock_sched();
4721 * move_queued_task - move a queued task to new rq.
4723 * Returns (locked) new rq. Old rq's lock is released.
4725 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4727 struct rq *rq = task_rq(p);
4729 lockdep_assert_held(&rq->lock);
4731 dequeue_task(rq, p, 0);
4732 p->on_rq = TASK_ON_RQ_MIGRATING;
4733 set_task_cpu(p, new_cpu);
4734 raw_spin_unlock(&rq->lock);
4736 rq = cpu_rq(new_cpu);
4738 raw_spin_lock(&rq->lock);
4739 BUG_ON(task_cpu(p) != new_cpu);
4740 p->on_rq = TASK_ON_RQ_QUEUED;
4741 enqueue_task(rq, p, 0);
4742 check_preempt_curr(rq, p, 0);
4747 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4749 if (p->sched_class->set_cpus_allowed)
4750 p->sched_class->set_cpus_allowed(p, new_mask);
4752 cpumask_copy(&p->cpus_allowed, new_mask);
4753 p->nr_cpus_allowed = cpumask_weight(new_mask);
4757 * This is how migration works:
4759 * 1) we invoke migration_cpu_stop() on the target CPU using
4761 * 2) stopper starts to run (implicitly forcing the migrated thread
4763 * 3) it checks whether the migrated task is still in the wrong runqueue.
4764 * 4) if it's in the wrong runqueue then the migration thread removes
4765 * it and puts it into the right queue.
4766 * 5) stopper completes and stop_one_cpu() returns and the migration
4771 * Change a given task's CPU affinity. Migrate the thread to a
4772 * proper CPU and schedule it away if the CPU it's executing on
4773 * is removed from the allowed bitmask.
4775 * NOTE: the caller must have a valid reference to the task, the
4776 * task must not exit() & deallocate itself prematurely. The
4777 * call is not atomic; no spinlocks may be held.
4779 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4781 unsigned long flags;
4783 unsigned int dest_cpu;
4786 rq = task_rq_lock(p, &flags);
4788 if (cpumask_equal(&p->cpus_allowed, new_mask))
4791 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4796 do_set_cpus_allowed(p, new_mask);
4798 /* Can the task run on the task's current CPU? If so, we're done */
4799 if (cpumask_test_cpu(task_cpu(p), new_mask))
4802 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4803 if (task_running(rq, p) || p->state == TASK_WAKING) {
4804 struct migration_arg arg = { p, dest_cpu };
4805 /* Need help from migration thread: drop lock and wait. */
4806 task_rq_unlock(rq, p, &flags);
4807 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4808 tlb_migrate_finish(p->mm);
4810 } else if (task_on_rq_queued(p))
4811 rq = move_queued_task(p, dest_cpu);
4813 task_rq_unlock(rq, p, &flags);
4817 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4820 * Move (not current) task off this cpu, onto dest cpu. We're doing
4821 * this because either it can't run here any more (set_cpus_allowed()
4822 * away from this CPU, or CPU going down), or because we're
4823 * attempting to rebalance this task on exec (sched_exec).
4825 * So we race with normal scheduler movements, but that's OK, as long
4826 * as the task is no longer on this CPU.
4828 * Returns non-zero if task was successfully migrated.
4830 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4835 if (unlikely(!cpu_active(dest_cpu)))
4838 rq = cpu_rq(src_cpu);
4840 raw_spin_lock(&p->pi_lock);
4841 raw_spin_lock(&rq->lock);
4842 /* Already moved. */
4843 if (task_cpu(p) != src_cpu)
4846 /* Affinity changed (again). */
4847 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4851 * If we're not on a rq, the next wake-up will ensure we're
4854 if (task_on_rq_queued(p))
4855 rq = move_queued_task(p, dest_cpu);
4859 raw_spin_unlock(&rq->lock);
4860 raw_spin_unlock(&p->pi_lock);
4864 #ifdef CONFIG_NUMA_BALANCING
4865 /* Migrate current task p to target_cpu */
4866 int migrate_task_to(struct task_struct *p, int target_cpu)
4868 struct migration_arg arg = { p, target_cpu };
4869 int curr_cpu = task_cpu(p);
4871 if (curr_cpu == target_cpu)
4874 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4877 /* TODO: This is not properly updating schedstats */
4879 trace_sched_move_numa(p, curr_cpu, target_cpu);
4880 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4884 * Requeue a task on a given node and accurately track the number of NUMA
4885 * tasks on the runqueues
4887 void sched_setnuma(struct task_struct *p, int nid)
4890 unsigned long flags;
4891 bool queued, running;
4893 rq = task_rq_lock(p, &flags);
4894 queued = task_on_rq_queued(p);
4895 running = task_current(rq, p);
4898 dequeue_task(rq, p, 0);
4900 put_prev_task(rq, p);
4902 p->numa_preferred_nid = nid;
4905 p->sched_class->set_curr_task(rq);
4907 enqueue_task(rq, p, 0);
4908 task_rq_unlock(rq, p, &flags);
4913 * migration_cpu_stop - this will be executed by a highprio stopper thread
4914 * and performs thread migration by bumping thread off CPU then
4915 * 'pushing' onto another runqueue.
4917 static int migration_cpu_stop(void *data)
4919 struct migration_arg *arg = data;
4922 * The original target cpu might have gone down and we might
4923 * be on another cpu but it doesn't matter.
4925 local_irq_disable();
4927 * We need to explicitly wake pending tasks before running
4928 * __migrate_task() such that we will not miss enforcing cpus_allowed
4929 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4931 sched_ttwu_pending();
4932 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4937 #ifdef CONFIG_HOTPLUG_CPU
4940 * Ensures that the idle task is using init_mm right before its cpu goes
4943 void idle_task_exit(void)
4945 struct mm_struct *mm = current->active_mm;
4947 BUG_ON(cpu_online(smp_processor_id()));
4949 if (mm != &init_mm) {
4950 switch_mm(mm, &init_mm, current);
4951 finish_arch_post_lock_switch();
4957 * Since this CPU is going 'away' for a while, fold any nr_active delta
4958 * we might have. Assumes we're called after migrate_tasks() so that the
4959 * nr_active count is stable.
4961 * Also see the comment "Global load-average calculations".
4963 static void calc_load_migrate(struct rq *rq)
4965 long delta = calc_load_fold_active(rq);
4967 atomic_long_add(delta, &calc_load_tasks);
4970 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4974 static const struct sched_class fake_sched_class = {
4975 .put_prev_task = put_prev_task_fake,
4978 static struct task_struct fake_task = {
4980 * Avoid pull_{rt,dl}_task()
4982 .prio = MAX_PRIO + 1,
4983 .sched_class = &fake_sched_class,
4987 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4988 * try_to_wake_up()->select_task_rq().
4990 * Called with rq->lock held even though we'er in stop_machine() and
4991 * there's no concurrency possible, we hold the required locks anyway
4992 * because of lock validation efforts.
4994 static void migrate_tasks(unsigned int dead_cpu)
4996 struct rq *rq = cpu_rq(dead_cpu);
4997 struct task_struct *next, *stop = rq->stop;
5001 * Fudge the rq selection such that the below task selection loop
5002 * doesn't get stuck on the currently eligible stop task.
5004 * We're currently inside stop_machine() and the rq is either stuck
5005 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5006 * either way we should never end up calling schedule() until we're
5012 * put_prev_task() and pick_next_task() sched
5013 * class method both need to have an up-to-date
5014 * value of rq->clock[_task]
5016 update_rq_clock(rq);
5020 * There's this thread running, bail when that's the only
5023 if (rq->nr_running == 1)
5026 next = pick_next_task(rq, &fake_task);
5028 next->sched_class->put_prev_task(rq, next);
5030 /* Find suitable destination for @next, with force if needed. */
5031 dest_cpu = select_fallback_rq(dead_cpu, next);
5032 raw_spin_unlock(&rq->lock);
5034 __migrate_task(next, dead_cpu, dest_cpu);
5036 raw_spin_lock(&rq->lock);
5042 #endif /* CONFIG_HOTPLUG_CPU */
5044 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5046 static struct ctl_table sd_ctl_dir[] = {
5048 .procname = "sched_domain",
5054 static struct ctl_table sd_ctl_root[] = {
5056 .procname = "kernel",
5058 .child = sd_ctl_dir,
5063 static struct ctl_table *sd_alloc_ctl_entry(int n)
5065 struct ctl_table *entry =
5066 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5071 static void sd_free_ctl_entry(struct ctl_table **tablep)
5073 struct ctl_table *entry;
5076 * In the intermediate directories, both the child directory and
5077 * procname are dynamically allocated and could fail but the mode
5078 * will always be set. In the lowest directory the names are
5079 * static strings and all have proc handlers.
5081 for (entry = *tablep; entry->mode; entry++) {
5083 sd_free_ctl_entry(&entry->child);
5084 if (entry->proc_handler == NULL)
5085 kfree(entry->procname);
5092 static int min_load_idx = 0;
5093 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5096 set_table_entry(struct ctl_table *entry,
5097 const char *procname, void *data, int maxlen,
5098 umode_t mode, proc_handler *proc_handler,
5101 entry->procname = procname;
5103 entry->maxlen = maxlen;
5105 entry->proc_handler = proc_handler;
5108 entry->extra1 = &min_load_idx;
5109 entry->extra2 = &max_load_idx;
5113 static struct ctl_table *
5114 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5116 struct ctl_table *table = sd_alloc_ctl_entry(14);
5121 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5122 sizeof(long), 0644, proc_doulongvec_minmax, false);
5123 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5124 sizeof(long), 0644, proc_doulongvec_minmax, false);
5125 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5126 sizeof(int), 0644, proc_dointvec_minmax, true);
5127 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5128 sizeof(int), 0644, proc_dointvec_minmax, true);
5129 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5130 sizeof(int), 0644, proc_dointvec_minmax, true);
5131 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5132 sizeof(int), 0644, proc_dointvec_minmax, true);
5133 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5134 sizeof(int), 0644, proc_dointvec_minmax, true);
5135 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5136 sizeof(int), 0644, proc_dointvec_minmax, false);
5137 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5138 sizeof(int), 0644, proc_dointvec_minmax, false);
5139 set_table_entry(&table[9], "cache_nice_tries",
5140 &sd->cache_nice_tries,
5141 sizeof(int), 0644, proc_dointvec_minmax, false);
5142 set_table_entry(&table[10], "flags", &sd->flags,
5143 sizeof(int), 0644, proc_dointvec_minmax, false);
5144 set_table_entry(&table[11], "max_newidle_lb_cost",
5145 &sd->max_newidle_lb_cost,
5146 sizeof(long), 0644, proc_doulongvec_minmax, false);
5147 set_table_entry(&table[12], "name", sd->name,
5148 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5149 /* &table[13] is terminator */
5154 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5156 struct ctl_table *entry, *table;
5157 struct sched_domain *sd;
5158 int domain_num = 0, i;
5161 for_each_domain(cpu, sd)
5163 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5168 for_each_domain(cpu, sd) {
5169 snprintf(buf, 32, "domain%d", i);
5170 entry->procname = kstrdup(buf, GFP_KERNEL);
5172 entry->child = sd_alloc_ctl_domain_table(sd);
5179 static struct ctl_table_header *sd_sysctl_header;
5180 static void register_sched_domain_sysctl(void)
5182 int i, cpu_num = num_possible_cpus();
5183 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5186 WARN_ON(sd_ctl_dir[0].child);
5187 sd_ctl_dir[0].child = entry;
5192 for_each_possible_cpu(i) {
5193 snprintf(buf, 32, "cpu%d", i);
5194 entry->procname = kstrdup(buf, GFP_KERNEL);
5196 entry->child = sd_alloc_ctl_cpu_table(i);
5200 WARN_ON(sd_sysctl_header);
5201 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5204 /* may be called multiple times per register */
5205 static void unregister_sched_domain_sysctl(void)
5207 if (sd_sysctl_header)
5208 unregister_sysctl_table(sd_sysctl_header);
5209 sd_sysctl_header = NULL;
5210 if (sd_ctl_dir[0].child)
5211 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5214 static void register_sched_domain_sysctl(void)
5217 static void unregister_sched_domain_sysctl(void)
5222 static void set_rq_online(struct rq *rq)
5225 const struct sched_class *class;
5227 cpumask_set_cpu(rq->cpu, rq->rd->online);
5230 for_each_class(class) {
5231 if (class->rq_online)
5232 class->rq_online(rq);
5237 static void set_rq_offline(struct rq *rq)
5240 const struct sched_class *class;
5242 for_each_class(class) {
5243 if (class->rq_offline)
5244 class->rq_offline(rq);
5247 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5253 * migration_call - callback that gets triggered when a CPU is added.
5254 * Here we can start up the necessary migration thread for the new CPU.
5257 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5259 int cpu = (long)hcpu;
5260 unsigned long flags;
5261 struct rq *rq = cpu_rq(cpu);
5263 switch (action & ~CPU_TASKS_FROZEN) {
5265 case CPU_UP_PREPARE:
5266 rq->calc_load_update = calc_load_update;
5270 /* Update our root-domain */
5271 raw_spin_lock_irqsave(&rq->lock, flags);
5273 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5277 raw_spin_unlock_irqrestore(&rq->lock, flags);
5280 #ifdef CONFIG_HOTPLUG_CPU
5282 sched_ttwu_pending();
5283 /* Update our root-domain */
5284 raw_spin_lock_irqsave(&rq->lock, flags);
5286 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5290 BUG_ON(rq->nr_running != 1); /* the migration thread */
5291 raw_spin_unlock_irqrestore(&rq->lock, flags);
5295 calc_load_migrate(rq);
5300 update_max_interval();
5306 * Register at high priority so that task migration (migrate_all_tasks)
5307 * happens before everything else. This has to be lower priority than
5308 * the notifier in the perf_event subsystem, though.
5310 static struct notifier_block migration_notifier = {
5311 .notifier_call = migration_call,
5312 .priority = CPU_PRI_MIGRATION,
5315 static void __cpuinit set_cpu_rq_start_time(void)
5317 int cpu = smp_processor_id();
5318 struct rq *rq = cpu_rq(cpu);
5319 rq->age_stamp = sched_clock_cpu(cpu);
5322 static int sched_cpu_active(struct notifier_block *nfb,
5323 unsigned long action, void *hcpu)
5325 switch (action & ~CPU_TASKS_FROZEN) {
5327 set_cpu_rq_start_time();
5329 case CPU_DOWN_FAILED:
5330 set_cpu_active((long)hcpu, true);
5337 static int sched_cpu_inactive(struct notifier_block *nfb,
5338 unsigned long action, void *hcpu)
5340 unsigned long flags;
5341 long cpu = (long)hcpu;
5344 switch (action & ~CPU_TASKS_FROZEN) {
5345 case CPU_DOWN_PREPARE:
5346 set_cpu_active(cpu, false);
5348 /* explicitly allow suspend */
5349 if (!(action & CPU_TASKS_FROZEN)) {
5353 rcu_read_lock_sched();
5354 dl_b = dl_bw_of(cpu);
5356 raw_spin_lock_irqsave(&dl_b->lock, flags);
5357 cpus = dl_bw_cpus(cpu);
5358 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5359 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5361 rcu_read_unlock_sched();
5364 return notifier_from_errno(-EBUSY);
5372 static int __init migration_init(void)
5374 void *cpu = (void *)(long)smp_processor_id();
5377 /* Initialize migration for the boot CPU */
5378 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5379 BUG_ON(err == NOTIFY_BAD);
5380 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5381 register_cpu_notifier(&migration_notifier);
5383 /* Register cpu active notifiers */
5384 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5385 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5389 early_initcall(migration_init);
5394 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5396 #ifdef CONFIG_SCHED_DEBUG
5398 static __read_mostly int sched_debug_enabled;
5400 static int __init sched_debug_setup(char *str)
5402 sched_debug_enabled = 1;
5406 early_param("sched_debug", sched_debug_setup);
5408 static inline bool sched_debug(void)
5410 return sched_debug_enabled;
5413 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5414 struct cpumask *groupmask)
5416 struct sched_group *group = sd->groups;
5418 cpumask_clear(groupmask);
5420 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5422 if (!(sd->flags & SD_LOAD_BALANCE)) {
5423 printk("does not load-balance\n");
5425 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5430 printk(KERN_CONT "span %*pbl level %s\n",
5431 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5433 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5434 printk(KERN_ERR "ERROR: domain->span does not contain "
5437 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5438 printk(KERN_ERR "ERROR: domain->groups does not contain"
5442 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5446 printk(KERN_ERR "ERROR: group is NULL\n");
5450 if (!cpumask_weight(sched_group_cpus(group))) {
5451 printk(KERN_CONT "\n");
5452 printk(KERN_ERR "ERROR: empty group\n");
5456 if (!(sd->flags & SD_OVERLAP) &&
5457 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5458 printk(KERN_CONT "\n");
5459 printk(KERN_ERR "ERROR: repeated CPUs\n");
5463 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5465 printk(KERN_CONT " %*pbl",
5466 cpumask_pr_args(sched_group_cpus(group)));
5467 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5468 printk(KERN_CONT " (cpu_capacity = %d)",
5469 group->sgc->capacity);
5472 group = group->next;
5473 } while (group != sd->groups);
5474 printk(KERN_CONT "\n");
5476 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5477 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5480 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5481 printk(KERN_ERR "ERROR: parent span is not a superset "
5482 "of domain->span\n");
5486 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5490 if (!sched_debug_enabled)
5494 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5498 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5501 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5509 #else /* !CONFIG_SCHED_DEBUG */
5510 # define sched_domain_debug(sd, cpu) do { } while (0)
5511 static inline bool sched_debug(void)
5515 #endif /* CONFIG_SCHED_DEBUG */
5517 static int sd_degenerate(struct sched_domain *sd)
5519 if (cpumask_weight(sched_domain_span(sd)) == 1)
5522 /* Following flags need at least 2 groups */
5523 if (sd->flags & (SD_LOAD_BALANCE |
5524 SD_BALANCE_NEWIDLE |
5527 SD_SHARE_CPUCAPACITY |
5528 SD_SHARE_PKG_RESOURCES |
5529 SD_SHARE_POWERDOMAIN)) {
5530 if (sd->groups != sd->groups->next)
5534 /* Following flags don't use groups */
5535 if (sd->flags & (SD_WAKE_AFFINE))
5542 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5544 unsigned long cflags = sd->flags, pflags = parent->flags;
5546 if (sd_degenerate(parent))
5549 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5552 /* Flags needing groups don't count if only 1 group in parent */
5553 if (parent->groups == parent->groups->next) {
5554 pflags &= ~(SD_LOAD_BALANCE |
5555 SD_BALANCE_NEWIDLE |
5558 SD_SHARE_CPUCAPACITY |
5559 SD_SHARE_PKG_RESOURCES |
5561 SD_SHARE_POWERDOMAIN);
5562 if (nr_node_ids == 1)
5563 pflags &= ~SD_SERIALIZE;
5565 if (~cflags & pflags)
5571 static void free_rootdomain(struct rcu_head *rcu)
5573 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5575 cpupri_cleanup(&rd->cpupri);
5576 cpudl_cleanup(&rd->cpudl);
5577 free_cpumask_var(rd->dlo_mask);
5578 free_cpumask_var(rd->rto_mask);
5579 free_cpumask_var(rd->online);
5580 free_cpumask_var(rd->span);
5584 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5586 struct root_domain *old_rd = NULL;
5587 unsigned long flags;
5589 raw_spin_lock_irqsave(&rq->lock, flags);
5594 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5597 cpumask_clear_cpu(rq->cpu, old_rd->span);
5600 * If we dont want to free the old_rd yet then
5601 * set old_rd to NULL to skip the freeing later
5604 if (!atomic_dec_and_test(&old_rd->refcount))
5608 atomic_inc(&rd->refcount);
5611 cpumask_set_cpu(rq->cpu, rd->span);
5612 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5615 raw_spin_unlock_irqrestore(&rq->lock, flags);
5618 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5621 static int init_rootdomain(struct root_domain *rd)
5623 memset(rd, 0, sizeof(*rd));
5625 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5627 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5629 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5631 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5634 init_dl_bw(&rd->dl_bw);
5635 if (cpudl_init(&rd->cpudl) != 0)
5638 if (cpupri_init(&rd->cpupri) != 0)
5643 free_cpumask_var(rd->rto_mask);
5645 free_cpumask_var(rd->dlo_mask);
5647 free_cpumask_var(rd->online);
5649 free_cpumask_var(rd->span);
5655 * By default the system creates a single root-domain with all cpus as
5656 * members (mimicking the global state we have today).
5658 struct root_domain def_root_domain;
5660 static void init_defrootdomain(void)
5662 init_rootdomain(&def_root_domain);
5664 atomic_set(&def_root_domain.refcount, 1);
5667 static struct root_domain *alloc_rootdomain(void)
5669 struct root_domain *rd;
5671 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5675 if (init_rootdomain(rd) != 0) {
5683 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5685 struct sched_group *tmp, *first;
5694 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5699 } while (sg != first);
5702 static void free_sched_domain(struct rcu_head *rcu)
5704 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5707 * If its an overlapping domain it has private groups, iterate and
5710 if (sd->flags & SD_OVERLAP) {
5711 free_sched_groups(sd->groups, 1);
5712 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5713 kfree(sd->groups->sgc);
5719 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5721 call_rcu(&sd->rcu, free_sched_domain);
5724 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5726 for (; sd; sd = sd->parent)
5727 destroy_sched_domain(sd, cpu);
5731 * Keep a special pointer to the highest sched_domain that has
5732 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5733 * allows us to avoid some pointer chasing select_idle_sibling().
5735 * Also keep a unique ID per domain (we use the first cpu number in
5736 * the cpumask of the domain), this allows us to quickly tell if
5737 * two cpus are in the same cache domain, see cpus_share_cache().
5739 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5740 DEFINE_PER_CPU(int, sd_llc_size);
5741 DEFINE_PER_CPU(int, sd_llc_id);
5742 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5743 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5744 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5746 static void update_top_cache_domain(int cpu)
5748 struct sched_domain *sd;
5749 struct sched_domain *busy_sd = NULL;
5753 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5755 id = cpumask_first(sched_domain_span(sd));
5756 size = cpumask_weight(sched_domain_span(sd));
5757 busy_sd = sd->parent; /* sd_busy */
5759 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5761 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5762 per_cpu(sd_llc_size, cpu) = size;
5763 per_cpu(sd_llc_id, cpu) = id;
5765 sd = lowest_flag_domain(cpu, SD_NUMA);
5766 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5768 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5769 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5773 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5774 * hold the hotplug lock.
5777 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5779 struct rq *rq = cpu_rq(cpu);
5780 struct sched_domain *tmp;
5782 /* Remove the sched domains which do not contribute to scheduling. */
5783 for (tmp = sd; tmp; ) {
5784 struct sched_domain *parent = tmp->parent;
5788 if (sd_parent_degenerate(tmp, parent)) {
5789 tmp->parent = parent->parent;
5791 parent->parent->child = tmp;
5793 * Transfer SD_PREFER_SIBLING down in case of a
5794 * degenerate parent; the spans match for this
5795 * so the property transfers.
5797 if (parent->flags & SD_PREFER_SIBLING)
5798 tmp->flags |= SD_PREFER_SIBLING;
5799 destroy_sched_domain(parent, cpu);
5804 if (sd && sd_degenerate(sd)) {
5807 destroy_sched_domain(tmp, cpu);
5812 sched_domain_debug(sd, cpu);
5814 rq_attach_root(rq, rd);
5816 rcu_assign_pointer(rq->sd, sd);
5817 destroy_sched_domains(tmp, cpu);
5819 update_top_cache_domain(cpu);
5822 /* cpus with isolated domains */
5823 static cpumask_var_t cpu_isolated_map;
5825 /* Setup the mask of cpus configured for isolated domains */
5826 static int __init isolated_cpu_setup(char *str)
5828 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5829 cpulist_parse(str, cpu_isolated_map);
5833 __setup("isolcpus=", isolated_cpu_setup);
5836 struct sched_domain ** __percpu sd;
5837 struct root_domain *rd;
5848 * Build an iteration mask that can exclude certain CPUs from the upwards
5851 * Asymmetric node setups can result in situations where the domain tree is of
5852 * unequal depth, make sure to skip domains that already cover the entire
5855 * In that case build_sched_domains() will have terminated the iteration early
5856 * and our sibling sd spans will be empty. Domains should always include the
5857 * cpu they're built on, so check that.
5860 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5862 const struct cpumask *span = sched_domain_span(sd);
5863 struct sd_data *sdd = sd->private;
5864 struct sched_domain *sibling;
5867 for_each_cpu(i, span) {
5868 sibling = *per_cpu_ptr(sdd->sd, i);
5869 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5872 cpumask_set_cpu(i, sched_group_mask(sg));
5877 * Return the canonical balance cpu for this group, this is the first cpu
5878 * of this group that's also in the iteration mask.
5880 int group_balance_cpu(struct sched_group *sg)
5882 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5886 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5888 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5889 const struct cpumask *span = sched_domain_span(sd);
5890 struct cpumask *covered = sched_domains_tmpmask;
5891 struct sd_data *sdd = sd->private;
5892 struct sched_domain *sibling;
5895 cpumask_clear(covered);
5897 for_each_cpu(i, span) {
5898 struct cpumask *sg_span;
5900 if (cpumask_test_cpu(i, covered))
5903 sibling = *per_cpu_ptr(sdd->sd, i);
5905 /* See the comment near build_group_mask(). */
5906 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5909 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5910 GFP_KERNEL, cpu_to_node(cpu));
5915 sg_span = sched_group_cpus(sg);
5917 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5919 cpumask_set_cpu(i, sg_span);
5921 cpumask_or(covered, covered, sg_span);
5923 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5924 if (atomic_inc_return(&sg->sgc->ref) == 1)
5925 build_group_mask(sd, sg);
5928 * Initialize sgc->capacity such that even if we mess up the
5929 * domains and no possible iteration will get us here, we won't
5932 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5935 * Make sure the first group of this domain contains the
5936 * canonical balance cpu. Otherwise the sched_domain iteration
5937 * breaks. See update_sg_lb_stats().
5939 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5940 group_balance_cpu(sg) == cpu)
5950 sd->groups = groups;
5955 free_sched_groups(first, 0);
5960 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5962 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5963 struct sched_domain *child = sd->child;
5966 cpu = cpumask_first(sched_domain_span(child));
5969 *sg = *per_cpu_ptr(sdd->sg, cpu);
5970 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5971 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5978 * build_sched_groups will build a circular linked list of the groups
5979 * covered by the given span, and will set each group's ->cpumask correctly,
5980 * and ->cpu_capacity to 0.
5982 * Assumes the sched_domain tree is fully constructed
5985 build_sched_groups(struct sched_domain *sd, int cpu)
5987 struct sched_group *first = NULL, *last = NULL;
5988 struct sd_data *sdd = sd->private;
5989 const struct cpumask *span = sched_domain_span(sd);
5990 struct cpumask *covered;
5993 get_group(cpu, sdd, &sd->groups);
5994 atomic_inc(&sd->groups->ref);
5996 if (cpu != cpumask_first(span))
5999 lockdep_assert_held(&sched_domains_mutex);
6000 covered = sched_domains_tmpmask;
6002 cpumask_clear(covered);
6004 for_each_cpu(i, span) {
6005 struct sched_group *sg;
6008 if (cpumask_test_cpu(i, covered))
6011 group = get_group(i, sdd, &sg);
6012 cpumask_setall(sched_group_mask(sg));
6014 for_each_cpu(j, span) {
6015 if (get_group(j, sdd, NULL) != group)
6018 cpumask_set_cpu(j, covered);
6019 cpumask_set_cpu(j, sched_group_cpus(sg));
6034 * Initialize sched groups cpu_capacity.
6036 * cpu_capacity indicates the capacity of sched group, which is used while
6037 * distributing the load between different sched groups in a sched domain.
6038 * Typically cpu_capacity for all the groups in a sched domain will be same
6039 * unless there are asymmetries in the topology. If there are asymmetries,
6040 * group having more cpu_capacity will pickup more load compared to the
6041 * group having less cpu_capacity.
6043 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6045 struct sched_group *sg = sd->groups;
6050 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6052 } while (sg != sd->groups);
6054 if (cpu != group_balance_cpu(sg))
6057 update_group_capacity(sd, cpu);
6058 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6062 * Initializers for schedule domains
6063 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6066 static int default_relax_domain_level = -1;
6067 int sched_domain_level_max;
6069 static int __init setup_relax_domain_level(char *str)
6071 if (kstrtoint(str, 0, &default_relax_domain_level))
6072 pr_warn("Unable to set relax_domain_level\n");
6076 __setup("relax_domain_level=", setup_relax_domain_level);
6078 static void set_domain_attribute(struct sched_domain *sd,
6079 struct sched_domain_attr *attr)
6083 if (!attr || attr->relax_domain_level < 0) {
6084 if (default_relax_domain_level < 0)
6087 request = default_relax_domain_level;
6089 request = attr->relax_domain_level;
6090 if (request < sd->level) {
6091 /* turn off idle balance on this domain */
6092 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6094 /* turn on idle balance on this domain */
6095 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6099 static void __sdt_free(const struct cpumask *cpu_map);
6100 static int __sdt_alloc(const struct cpumask *cpu_map);
6102 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6103 const struct cpumask *cpu_map)
6107 if (!atomic_read(&d->rd->refcount))
6108 free_rootdomain(&d->rd->rcu); /* fall through */
6110 free_percpu(d->sd); /* fall through */
6112 __sdt_free(cpu_map); /* fall through */
6118 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6119 const struct cpumask *cpu_map)
6121 memset(d, 0, sizeof(*d));
6123 if (__sdt_alloc(cpu_map))
6124 return sa_sd_storage;
6125 d->sd = alloc_percpu(struct sched_domain *);
6127 return sa_sd_storage;
6128 d->rd = alloc_rootdomain();
6131 return sa_rootdomain;
6135 * NULL the sd_data elements we've used to build the sched_domain and
6136 * sched_group structure so that the subsequent __free_domain_allocs()
6137 * will not free the data we're using.
6139 static void claim_allocations(int cpu, struct sched_domain *sd)
6141 struct sd_data *sdd = sd->private;
6143 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6144 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6146 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6147 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6149 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6150 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6154 static int sched_domains_numa_levels;
6155 enum numa_topology_type sched_numa_topology_type;
6156 static int *sched_domains_numa_distance;
6157 int sched_max_numa_distance;
6158 static struct cpumask ***sched_domains_numa_masks;
6159 static int sched_domains_curr_level;
6163 * SD_flags allowed in topology descriptions.
6165 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6166 * SD_SHARE_PKG_RESOURCES - describes shared caches
6167 * SD_NUMA - describes NUMA topologies
6168 * SD_SHARE_POWERDOMAIN - describes shared power domain
6171 * SD_ASYM_PACKING - describes SMT quirks
6173 #define TOPOLOGY_SD_FLAGS \
6174 (SD_SHARE_CPUCAPACITY | \
6175 SD_SHARE_PKG_RESOURCES | \
6178 SD_SHARE_POWERDOMAIN)
6180 static struct sched_domain *
6181 sd_init(struct sched_domain_topology_level *tl, int cpu)
6183 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6184 int sd_weight, sd_flags = 0;
6188 * Ugly hack to pass state to sd_numa_mask()...
6190 sched_domains_curr_level = tl->numa_level;
6193 sd_weight = cpumask_weight(tl->mask(cpu));
6196 sd_flags = (*tl->sd_flags)();
6197 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6198 "wrong sd_flags in topology description\n"))
6199 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6201 *sd = (struct sched_domain){
6202 .min_interval = sd_weight,
6203 .max_interval = 2*sd_weight,
6205 .imbalance_pct = 125,
6207 .cache_nice_tries = 0,
6214 .flags = 1*SD_LOAD_BALANCE
6215 | 1*SD_BALANCE_NEWIDLE
6220 | 0*SD_SHARE_CPUCAPACITY
6221 | 0*SD_SHARE_PKG_RESOURCES
6223 | 0*SD_PREFER_SIBLING
6228 .last_balance = jiffies,
6229 .balance_interval = sd_weight,
6231 .max_newidle_lb_cost = 0,
6232 .next_decay_max_lb_cost = jiffies,
6233 #ifdef CONFIG_SCHED_DEBUG
6239 * Convert topological properties into behaviour.
6242 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6243 sd->flags |= SD_PREFER_SIBLING;
6244 sd->imbalance_pct = 110;
6245 sd->smt_gain = 1178; /* ~15% */
6247 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6248 sd->imbalance_pct = 117;
6249 sd->cache_nice_tries = 1;
6253 } else if (sd->flags & SD_NUMA) {
6254 sd->cache_nice_tries = 2;
6258 sd->flags |= SD_SERIALIZE;
6259 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6260 sd->flags &= ~(SD_BALANCE_EXEC |
6267 sd->flags |= SD_PREFER_SIBLING;
6268 sd->cache_nice_tries = 1;
6273 sd->private = &tl->data;
6279 * Topology list, bottom-up.
6281 static struct sched_domain_topology_level default_topology[] = {
6282 #ifdef CONFIG_SCHED_SMT
6283 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6285 #ifdef CONFIG_SCHED_MC
6286 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6288 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6292 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6294 #define for_each_sd_topology(tl) \
6295 for (tl = sched_domain_topology; tl->mask; tl++)
6297 void set_sched_topology(struct sched_domain_topology_level *tl)
6299 sched_domain_topology = tl;
6304 static const struct cpumask *sd_numa_mask(int cpu)
6306 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6309 static void sched_numa_warn(const char *str)
6311 static int done = false;
6319 printk(KERN_WARNING "ERROR: %s\n\n", str);
6321 for (i = 0; i < nr_node_ids; i++) {
6322 printk(KERN_WARNING " ");
6323 for (j = 0; j < nr_node_ids; j++)
6324 printk(KERN_CONT "%02d ", node_distance(i,j));
6325 printk(KERN_CONT "\n");
6327 printk(KERN_WARNING "\n");
6330 bool find_numa_distance(int distance)
6334 if (distance == node_distance(0, 0))
6337 for (i = 0; i < sched_domains_numa_levels; i++) {
6338 if (sched_domains_numa_distance[i] == distance)
6346 * A system can have three types of NUMA topology:
6347 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6348 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6349 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6351 * The difference between a glueless mesh topology and a backplane
6352 * topology lies in whether communication between not directly
6353 * connected nodes goes through intermediary nodes (where programs
6354 * could run), or through backplane controllers. This affects
6355 * placement of programs.
6357 * The type of topology can be discerned with the following tests:
6358 * - If the maximum distance between any nodes is 1 hop, the system
6359 * is directly connected.
6360 * - If for two nodes A and B, located N > 1 hops away from each other,
6361 * there is an intermediary node C, which is < N hops away from both
6362 * nodes A and B, the system is a glueless mesh.
6364 static void init_numa_topology_type(void)
6368 n = sched_max_numa_distance;
6371 sched_numa_topology_type = NUMA_DIRECT;
6373 for_each_online_node(a) {
6374 for_each_online_node(b) {
6375 /* Find two nodes furthest removed from each other. */
6376 if (node_distance(a, b) < n)
6379 /* Is there an intermediary node between a and b? */
6380 for_each_online_node(c) {
6381 if (node_distance(a, c) < n &&
6382 node_distance(b, c) < n) {
6383 sched_numa_topology_type =
6389 sched_numa_topology_type = NUMA_BACKPLANE;
6395 static void sched_init_numa(void)
6397 int next_distance, curr_distance = node_distance(0, 0);
6398 struct sched_domain_topology_level *tl;
6402 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6403 if (!sched_domains_numa_distance)
6407 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6408 * unique distances in the node_distance() table.
6410 * Assumes node_distance(0,j) includes all distances in
6411 * node_distance(i,j) in order to avoid cubic time.
6413 next_distance = curr_distance;
6414 for (i = 0; i < nr_node_ids; i++) {
6415 for (j = 0; j < nr_node_ids; j++) {
6416 for (k = 0; k < nr_node_ids; k++) {
6417 int distance = node_distance(i, k);
6419 if (distance > curr_distance &&
6420 (distance < next_distance ||
6421 next_distance == curr_distance))
6422 next_distance = distance;
6425 * While not a strong assumption it would be nice to know
6426 * about cases where if node A is connected to B, B is not
6427 * equally connected to A.
6429 if (sched_debug() && node_distance(k, i) != distance)
6430 sched_numa_warn("Node-distance not symmetric");
6432 if (sched_debug() && i && !find_numa_distance(distance))
6433 sched_numa_warn("Node-0 not representative");
6435 if (next_distance != curr_distance) {
6436 sched_domains_numa_distance[level++] = next_distance;
6437 sched_domains_numa_levels = level;
6438 curr_distance = next_distance;
6443 * In case of sched_debug() we verify the above assumption.
6453 * 'level' contains the number of unique distances, excluding the
6454 * identity distance node_distance(i,i).
6456 * The sched_domains_numa_distance[] array includes the actual distance
6461 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6462 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6463 * the array will contain less then 'level' members. This could be
6464 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6465 * in other functions.
6467 * We reset it to 'level' at the end of this function.
6469 sched_domains_numa_levels = 0;
6471 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6472 if (!sched_domains_numa_masks)
6476 * Now for each level, construct a mask per node which contains all
6477 * cpus of nodes that are that many hops away from us.
6479 for (i = 0; i < level; i++) {
6480 sched_domains_numa_masks[i] =
6481 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6482 if (!sched_domains_numa_masks[i])
6485 for (j = 0; j < nr_node_ids; j++) {
6486 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6490 sched_domains_numa_masks[i][j] = mask;
6492 for (k = 0; k < nr_node_ids; k++) {
6493 if (node_distance(j, k) > sched_domains_numa_distance[i])
6496 cpumask_or(mask, mask, cpumask_of_node(k));
6501 /* Compute default topology size */
6502 for (i = 0; sched_domain_topology[i].mask; i++);
6504 tl = kzalloc((i + level + 1) *
6505 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6510 * Copy the default topology bits..
6512 for (i = 0; sched_domain_topology[i].mask; i++)
6513 tl[i] = sched_domain_topology[i];
6516 * .. and append 'j' levels of NUMA goodness.
6518 for (j = 0; j < level; i++, j++) {
6519 tl[i] = (struct sched_domain_topology_level){
6520 .mask = sd_numa_mask,
6521 .sd_flags = cpu_numa_flags,
6522 .flags = SDTL_OVERLAP,
6528 sched_domain_topology = tl;
6530 sched_domains_numa_levels = level;
6531 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6533 init_numa_topology_type();
6536 static void sched_domains_numa_masks_set(int cpu)
6539 int node = cpu_to_node(cpu);
6541 for (i = 0; i < sched_domains_numa_levels; i++) {
6542 for (j = 0; j < nr_node_ids; j++) {
6543 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6544 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6549 static void sched_domains_numa_masks_clear(int cpu)
6552 for (i = 0; i < sched_domains_numa_levels; i++) {
6553 for (j = 0; j < nr_node_ids; j++)
6554 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6559 * Update sched_domains_numa_masks[level][node] array when new cpus
6562 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6563 unsigned long action,
6566 int cpu = (long)hcpu;
6568 switch (action & ~CPU_TASKS_FROZEN) {
6570 sched_domains_numa_masks_set(cpu);
6574 sched_domains_numa_masks_clear(cpu);
6584 static inline void sched_init_numa(void)
6588 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6589 unsigned long action,
6594 #endif /* CONFIG_NUMA */
6596 static int __sdt_alloc(const struct cpumask *cpu_map)
6598 struct sched_domain_topology_level *tl;
6601 for_each_sd_topology(tl) {
6602 struct sd_data *sdd = &tl->data;
6604 sdd->sd = alloc_percpu(struct sched_domain *);
6608 sdd->sg = alloc_percpu(struct sched_group *);
6612 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6616 for_each_cpu(j, cpu_map) {
6617 struct sched_domain *sd;
6618 struct sched_group *sg;
6619 struct sched_group_capacity *sgc;
6621 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6622 GFP_KERNEL, cpu_to_node(j));
6626 *per_cpu_ptr(sdd->sd, j) = sd;
6628 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6629 GFP_KERNEL, cpu_to_node(j));
6635 *per_cpu_ptr(sdd->sg, j) = sg;
6637 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6638 GFP_KERNEL, cpu_to_node(j));
6642 *per_cpu_ptr(sdd->sgc, j) = sgc;
6649 static void __sdt_free(const struct cpumask *cpu_map)
6651 struct sched_domain_topology_level *tl;
6654 for_each_sd_topology(tl) {
6655 struct sd_data *sdd = &tl->data;
6657 for_each_cpu(j, cpu_map) {
6658 struct sched_domain *sd;
6661 sd = *per_cpu_ptr(sdd->sd, j);
6662 if (sd && (sd->flags & SD_OVERLAP))
6663 free_sched_groups(sd->groups, 0);
6664 kfree(*per_cpu_ptr(sdd->sd, j));
6668 kfree(*per_cpu_ptr(sdd->sg, j));
6670 kfree(*per_cpu_ptr(sdd->sgc, j));
6672 free_percpu(sdd->sd);
6674 free_percpu(sdd->sg);
6676 free_percpu(sdd->sgc);
6681 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6682 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6683 struct sched_domain *child, int cpu)
6685 struct sched_domain *sd = sd_init(tl, cpu);
6689 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6691 sd->level = child->level + 1;
6692 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6696 if (!cpumask_subset(sched_domain_span(child),
6697 sched_domain_span(sd))) {
6698 pr_err("BUG: arch topology borken\n");
6699 #ifdef CONFIG_SCHED_DEBUG
6700 pr_err(" the %s domain not a subset of the %s domain\n",
6701 child->name, sd->name);
6703 /* Fixup, ensure @sd has at least @child cpus. */
6704 cpumask_or(sched_domain_span(sd),
6705 sched_domain_span(sd),
6706 sched_domain_span(child));
6710 set_domain_attribute(sd, attr);
6716 * Build sched domains for a given set of cpus and attach the sched domains
6717 * to the individual cpus
6719 static int build_sched_domains(const struct cpumask *cpu_map,
6720 struct sched_domain_attr *attr)
6722 enum s_alloc alloc_state;
6723 struct sched_domain *sd;
6725 int i, ret = -ENOMEM;
6727 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6728 if (alloc_state != sa_rootdomain)
6731 /* Set up domains for cpus specified by the cpu_map. */
6732 for_each_cpu(i, cpu_map) {
6733 struct sched_domain_topology_level *tl;
6736 for_each_sd_topology(tl) {
6737 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6738 if (tl == sched_domain_topology)
6739 *per_cpu_ptr(d.sd, i) = sd;
6740 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6741 sd->flags |= SD_OVERLAP;
6742 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6747 /* Build the groups for the domains */
6748 for_each_cpu(i, cpu_map) {
6749 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6750 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6751 if (sd->flags & SD_OVERLAP) {
6752 if (build_overlap_sched_groups(sd, i))
6755 if (build_sched_groups(sd, i))
6761 /* Calculate CPU capacity for physical packages and nodes */
6762 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6763 if (!cpumask_test_cpu(i, cpu_map))
6766 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6767 claim_allocations(i, sd);
6768 init_sched_groups_capacity(i, sd);
6772 /* Attach the domains */
6774 for_each_cpu(i, cpu_map) {
6775 sd = *per_cpu_ptr(d.sd, i);
6776 cpu_attach_domain(sd, d.rd, i);
6782 __free_domain_allocs(&d, alloc_state, cpu_map);
6786 static cpumask_var_t *doms_cur; /* current sched domains */
6787 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6788 static struct sched_domain_attr *dattr_cur;
6789 /* attribues of custom domains in 'doms_cur' */
6792 * Special case: If a kmalloc of a doms_cur partition (array of
6793 * cpumask) fails, then fallback to a single sched domain,
6794 * as determined by the single cpumask fallback_doms.
6796 static cpumask_var_t fallback_doms;
6799 * arch_update_cpu_topology lets virtualized architectures update the
6800 * cpu core maps. It is supposed to return 1 if the topology changed
6801 * or 0 if it stayed the same.
6803 int __weak arch_update_cpu_topology(void)
6808 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6811 cpumask_var_t *doms;
6813 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6816 for (i = 0; i < ndoms; i++) {
6817 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6818 free_sched_domains(doms, i);
6825 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6828 for (i = 0; i < ndoms; i++)
6829 free_cpumask_var(doms[i]);
6834 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6835 * For now this just excludes isolated cpus, but could be used to
6836 * exclude other special cases in the future.
6838 static int init_sched_domains(const struct cpumask *cpu_map)
6842 arch_update_cpu_topology();
6844 doms_cur = alloc_sched_domains(ndoms_cur);
6846 doms_cur = &fallback_doms;
6847 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6848 err = build_sched_domains(doms_cur[0], NULL);
6849 register_sched_domain_sysctl();
6855 * Detach sched domains from a group of cpus specified in cpu_map
6856 * These cpus will now be attached to the NULL domain
6858 static void detach_destroy_domains(const struct cpumask *cpu_map)
6863 for_each_cpu(i, cpu_map)
6864 cpu_attach_domain(NULL, &def_root_domain, i);
6868 /* handle null as "default" */
6869 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6870 struct sched_domain_attr *new, int idx_new)
6872 struct sched_domain_attr tmp;
6879 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6880 new ? (new + idx_new) : &tmp,
6881 sizeof(struct sched_domain_attr));
6885 * Partition sched domains as specified by the 'ndoms_new'
6886 * cpumasks in the array doms_new[] of cpumasks. This compares
6887 * doms_new[] to the current sched domain partitioning, doms_cur[].
6888 * It destroys each deleted domain and builds each new domain.
6890 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6891 * The masks don't intersect (don't overlap.) We should setup one
6892 * sched domain for each mask. CPUs not in any of the cpumasks will
6893 * not be load balanced. If the same cpumask appears both in the
6894 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6897 * The passed in 'doms_new' should be allocated using
6898 * alloc_sched_domains. This routine takes ownership of it and will
6899 * free_sched_domains it when done with it. If the caller failed the
6900 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6901 * and partition_sched_domains() will fallback to the single partition
6902 * 'fallback_doms', it also forces the domains to be rebuilt.
6904 * If doms_new == NULL it will be replaced with cpu_online_mask.
6905 * ndoms_new == 0 is a special case for destroying existing domains,
6906 * and it will not create the default domain.
6908 * Call with hotplug lock held
6910 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6911 struct sched_domain_attr *dattr_new)
6916 mutex_lock(&sched_domains_mutex);
6918 /* always unregister in case we don't destroy any domains */
6919 unregister_sched_domain_sysctl();
6921 /* Let architecture update cpu core mappings. */
6922 new_topology = arch_update_cpu_topology();
6924 n = doms_new ? ndoms_new : 0;
6926 /* Destroy deleted domains */
6927 for (i = 0; i < ndoms_cur; i++) {
6928 for (j = 0; j < n && !new_topology; j++) {
6929 if (cpumask_equal(doms_cur[i], doms_new[j])
6930 && dattrs_equal(dattr_cur, i, dattr_new, j))
6933 /* no match - a current sched domain not in new doms_new[] */
6934 detach_destroy_domains(doms_cur[i]);
6940 if (doms_new == NULL) {
6942 doms_new = &fallback_doms;
6943 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6944 WARN_ON_ONCE(dattr_new);
6947 /* Build new domains */
6948 for (i = 0; i < ndoms_new; i++) {
6949 for (j = 0; j < n && !new_topology; j++) {
6950 if (cpumask_equal(doms_new[i], doms_cur[j])
6951 && dattrs_equal(dattr_new, i, dattr_cur, j))
6954 /* no match - add a new doms_new */
6955 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6960 /* Remember the new sched domains */
6961 if (doms_cur != &fallback_doms)
6962 free_sched_domains(doms_cur, ndoms_cur);
6963 kfree(dattr_cur); /* kfree(NULL) is safe */
6964 doms_cur = doms_new;
6965 dattr_cur = dattr_new;
6966 ndoms_cur = ndoms_new;
6968 register_sched_domain_sysctl();
6970 mutex_unlock(&sched_domains_mutex);
6973 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6976 * Update cpusets according to cpu_active mask. If cpusets are
6977 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6978 * around partition_sched_domains().
6980 * If we come here as part of a suspend/resume, don't touch cpusets because we
6981 * want to restore it back to its original state upon resume anyway.
6983 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6987 case CPU_ONLINE_FROZEN:
6988 case CPU_DOWN_FAILED_FROZEN:
6991 * num_cpus_frozen tracks how many CPUs are involved in suspend
6992 * resume sequence. As long as this is not the last online
6993 * operation in the resume sequence, just build a single sched
6994 * domain, ignoring cpusets.
6997 if (likely(num_cpus_frozen)) {
6998 partition_sched_domains(1, NULL, NULL);
7003 * This is the last CPU online operation. So fall through and
7004 * restore the original sched domains by considering the
7005 * cpuset configurations.
7009 case CPU_DOWN_FAILED:
7010 cpuset_update_active_cpus(true);
7018 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7022 case CPU_DOWN_PREPARE:
7023 cpuset_update_active_cpus(false);
7025 case CPU_DOWN_PREPARE_FROZEN:
7027 partition_sched_domains(1, NULL, NULL);
7035 void __init sched_init_smp(void)
7037 cpumask_var_t non_isolated_cpus;
7039 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7040 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7045 * There's no userspace yet to cause hotplug operations; hence all the
7046 * cpu masks are stable and all blatant races in the below code cannot
7049 mutex_lock(&sched_domains_mutex);
7050 init_sched_domains(cpu_active_mask);
7051 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7052 if (cpumask_empty(non_isolated_cpus))
7053 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7054 mutex_unlock(&sched_domains_mutex);
7056 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7057 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7058 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7062 /* Move init over to a non-isolated CPU */
7063 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7065 sched_init_granularity();
7066 free_cpumask_var(non_isolated_cpus);
7068 init_sched_rt_class();
7069 init_sched_dl_class();
7072 void __init sched_init_smp(void)
7074 sched_init_granularity();
7076 #endif /* CONFIG_SMP */
7078 const_debug unsigned int sysctl_timer_migration = 1;
7080 int in_sched_functions(unsigned long addr)
7082 return in_lock_functions(addr) ||
7083 (addr >= (unsigned long)__sched_text_start
7084 && addr < (unsigned long)__sched_text_end);
7087 #ifdef CONFIG_CGROUP_SCHED
7089 * Default task group.
7090 * Every task in system belongs to this group at bootup.
7092 struct task_group root_task_group;
7093 LIST_HEAD(task_groups);
7096 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7098 void __init sched_init(void)
7101 unsigned long alloc_size = 0, ptr;
7103 #ifdef CONFIG_FAIR_GROUP_SCHED
7104 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7106 #ifdef CONFIG_RT_GROUP_SCHED
7107 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7110 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7112 #ifdef CONFIG_FAIR_GROUP_SCHED
7113 root_task_group.se = (struct sched_entity **)ptr;
7114 ptr += nr_cpu_ids * sizeof(void **);
7116 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7117 ptr += nr_cpu_ids * sizeof(void **);
7119 #endif /* CONFIG_FAIR_GROUP_SCHED */
7120 #ifdef CONFIG_RT_GROUP_SCHED
7121 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7122 ptr += nr_cpu_ids * sizeof(void **);
7124 root_task_group.rt_rq = (struct rt_rq **)ptr;
7125 ptr += nr_cpu_ids * sizeof(void **);
7127 #endif /* CONFIG_RT_GROUP_SCHED */
7129 #ifdef CONFIG_CPUMASK_OFFSTACK
7130 for_each_possible_cpu(i) {
7131 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7132 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7134 #endif /* CONFIG_CPUMASK_OFFSTACK */
7136 init_rt_bandwidth(&def_rt_bandwidth,
7137 global_rt_period(), global_rt_runtime());
7138 init_dl_bandwidth(&def_dl_bandwidth,
7139 global_rt_period(), global_rt_runtime());
7142 init_defrootdomain();
7145 #ifdef CONFIG_RT_GROUP_SCHED
7146 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7147 global_rt_period(), global_rt_runtime());
7148 #endif /* CONFIG_RT_GROUP_SCHED */
7150 #ifdef CONFIG_CGROUP_SCHED
7151 list_add(&root_task_group.list, &task_groups);
7152 INIT_LIST_HEAD(&root_task_group.children);
7153 INIT_LIST_HEAD(&root_task_group.siblings);
7154 autogroup_init(&init_task);
7156 #endif /* CONFIG_CGROUP_SCHED */
7158 for_each_possible_cpu(i) {
7162 raw_spin_lock_init(&rq->lock);
7164 rq->calc_load_active = 0;
7165 rq->calc_load_update = jiffies + LOAD_FREQ;
7166 init_cfs_rq(&rq->cfs);
7167 init_rt_rq(&rq->rt, rq);
7168 init_dl_rq(&rq->dl, rq);
7169 #ifdef CONFIG_FAIR_GROUP_SCHED
7170 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7171 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7173 * How much cpu bandwidth does root_task_group get?
7175 * In case of task-groups formed thr' the cgroup filesystem, it
7176 * gets 100% of the cpu resources in the system. This overall
7177 * system cpu resource is divided among the tasks of
7178 * root_task_group and its child task-groups in a fair manner,
7179 * based on each entity's (task or task-group's) weight
7180 * (se->load.weight).
7182 * In other words, if root_task_group has 10 tasks of weight
7183 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7184 * then A0's share of the cpu resource is:
7186 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7188 * We achieve this by letting root_task_group's tasks sit
7189 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7191 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7192 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7193 #endif /* CONFIG_FAIR_GROUP_SCHED */
7195 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7196 #ifdef CONFIG_RT_GROUP_SCHED
7197 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7200 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7201 rq->cpu_load[j] = 0;
7203 rq->last_load_update_tick = jiffies;
7208 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7209 rq->post_schedule = 0;
7210 rq->active_balance = 0;
7211 rq->next_balance = jiffies;
7216 rq->avg_idle = 2*sysctl_sched_migration_cost;
7217 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7219 INIT_LIST_HEAD(&rq->cfs_tasks);
7221 rq_attach_root(rq, &def_root_domain);
7222 #ifdef CONFIG_NO_HZ_COMMON
7225 #ifdef CONFIG_NO_HZ_FULL
7226 rq->last_sched_tick = 0;
7230 atomic_set(&rq->nr_iowait, 0);
7233 set_load_weight(&init_task);
7235 #ifdef CONFIG_PREEMPT_NOTIFIERS
7236 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7240 * The boot idle thread does lazy MMU switching as well:
7242 atomic_inc(&init_mm.mm_count);
7243 enter_lazy_tlb(&init_mm, current);
7246 * During early bootup we pretend to be a normal task:
7248 current->sched_class = &fair_sched_class;
7251 * Make us the idle thread. Technically, schedule() should not be
7252 * called from this thread, however somewhere below it might be,
7253 * but because we are the idle thread, we just pick up running again
7254 * when this runqueue becomes "idle".
7256 init_idle(current, smp_processor_id());
7258 calc_load_update = jiffies + LOAD_FREQ;
7261 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7262 /* May be allocated at isolcpus cmdline parse time */
7263 if (cpu_isolated_map == NULL)
7264 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7265 idle_thread_set_boot_cpu();
7266 set_cpu_rq_start_time();
7268 init_sched_fair_class();
7270 scheduler_running = 1;
7273 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7274 static inline int preempt_count_equals(int preempt_offset)
7276 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7278 return (nested == preempt_offset);
7281 void __might_sleep(const char *file, int line, int preempt_offset)
7284 * Blocking primitives will set (and therefore destroy) current->state,
7285 * since we will exit with TASK_RUNNING make sure we enter with it,
7286 * otherwise we will destroy state.
7288 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7289 "do not call blocking ops when !TASK_RUNNING; "
7290 "state=%lx set at [<%p>] %pS\n",
7292 (void *)current->task_state_change,
7293 (void *)current->task_state_change);
7295 ___might_sleep(file, line, preempt_offset);
7297 EXPORT_SYMBOL(__might_sleep);
7299 void ___might_sleep(const char *file, int line, int preempt_offset)
7301 static unsigned long prev_jiffy; /* ratelimiting */
7303 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7304 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7305 !is_idle_task(current)) ||
7306 system_state != SYSTEM_RUNNING || oops_in_progress)
7308 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7310 prev_jiffy = jiffies;
7313 "BUG: sleeping function called from invalid context at %s:%d\n",
7316 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7317 in_atomic(), irqs_disabled(),
7318 current->pid, current->comm);
7320 if (task_stack_end_corrupted(current))
7321 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7323 debug_show_held_locks(current);
7324 if (irqs_disabled())
7325 print_irqtrace_events(current);
7326 #ifdef CONFIG_DEBUG_PREEMPT
7327 if (!preempt_count_equals(preempt_offset)) {
7328 pr_err("Preemption disabled at:");
7329 print_ip_sym(current->preempt_disable_ip);
7335 EXPORT_SYMBOL(___might_sleep);
7338 #ifdef CONFIG_MAGIC_SYSRQ
7339 static void normalize_task(struct rq *rq, struct task_struct *p)
7341 const struct sched_class *prev_class = p->sched_class;
7342 struct sched_attr attr = {
7343 .sched_policy = SCHED_NORMAL,
7345 int old_prio = p->prio;
7348 queued = task_on_rq_queued(p);
7350 dequeue_task(rq, p, 0);
7351 __setscheduler(rq, p, &attr);
7353 enqueue_task(rq, p, 0);
7357 check_class_changed(rq, p, prev_class, old_prio);
7360 void normalize_rt_tasks(void)
7362 struct task_struct *g, *p;
7363 unsigned long flags;
7366 read_lock(&tasklist_lock);
7367 for_each_process_thread(g, p) {
7369 * Only normalize user tasks:
7371 if (p->flags & PF_KTHREAD)
7374 p->se.exec_start = 0;
7375 #ifdef CONFIG_SCHEDSTATS
7376 p->se.statistics.wait_start = 0;
7377 p->se.statistics.sleep_start = 0;
7378 p->se.statistics.block_start = 0;
7381 if (!dl_task(p) && !rt_task(p)) {
7383 * Renice negative nice level userspace
7386 if (task_nice(p) < 0)
7387 set_user_nice(p, 0);
7391 rq = task_rq_lock(p, &flags);
7392 normalize_task(rq, p);
7393 task_rq_unlock(rq, p, &flags);
7395 read_unlock(&tasklist_lock);
7398 #endif /* CONFIG_MAGIC_SYSRQ */
7400 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7402 * These functions are only useful for the IA64 MCA handling, or kdb.
7404 * They can only be called when the whole system has been
7405 * stopped - every CPU needs to be quiescent, and no scheduling
7406 * activity can take place. Using them for anything else would
7407 * be a serious bug, and as a result, they aren't even visible
7408 * under any other configuration.
7412 * curr_task - return the current task for a given cpu.
7413 * @cpu: the processor in question.
7415 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7417 * Return: The current task for @cpu.
7419 struct task_struct *curr_task(int cpu)
7421 return cpu_curr(cpu);
7424 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7428 * set_curr_task - set the current task for a given cpu.
7429 * @cpu: the processor in question.
7430 * @p: the task pointer to set.
7432 * Description: This function must only be used when non-maskable interrupts
7433 * are serviced on a separate stack. It allows the architecture to switch the
7434 * notion of the current task on a cpu in a non-blocking manner. This function
7435 * must be called with all CPU's synchronized, and interrupts disabled, the
7436 * and caller must save the original value of the current task (see
7437 * curr_task() above) and restore that value before reenabling interrupts and
7438 * re-starting the system.
7440 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7442 void set_curr_task(int cpu, struct task_struct *p)
7449 #ifdef CONFIG_CGROUP_SCHED
7450 /* task_group_lock serializes the addition/removal of task groups */
7451 static DEFINE_SPINLOCK(task_group_lock);
7453 static void free_sched_group(struct task_group *tg)
7455 free_fair_sched_group(tg);
7456 free_rt_sched_group(tg);
7461 /* allocate runqueue etc for a new task group */
7462 struct task_group *sched_create_group(struct task_group *parent)
7464 struct task_group *tg;
7466 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7468 return ERR_PTR(-ENOMEM);
7470 if (!alloc_fair_sched_group(tg, parent))
7473 if (!alloc_rt_sched_group(tg, parent))
7479 free_sched_group(tg);
7480 return ERR_PTR(-ENOMEM);
7483 void sched_online_group(struct task_group *tg, struct task_group *parent)
7485 unsigned long flags;
7487 spin_lock_irqsave(&task_group_lock, flags);
7488 list_add_rcu(&tg->list, &task_groups);
7490 WARN_ON(!parent); /* root should already exist */
7492 tg->parent = parent;
7493 INIT_LIST_HEAD(&tg->children);
7494 list_add_rcu(&tg->siblings, &parent->children);
7495 spin_unlock_irqrestore(&task_group_lock, flags);
7498 /* rcu callback to free various structures associated with a task group */
7499 static void free_sched_group_rcu(struct rcu_head *rhp)
7501 /* now it should be safe to free those cfs_rqs */
7502 free_sched_group(container_of(rhp, struct task_group, rcu));
7505 /* Destroy runqueue etc associated with a task group */
7506 void sched_destroy_group(struct task_group *tg)
7508 /* wait for possible concurrent references to cfs_rqs complete */
7509 call_rcu(&tg->rcu, free_sched_group_rcu);
7512 void sched_offline_group(struct task_group *tg)
7514 unsigned long flags;
7517 /* end participation in shares distribution */
7518 for_each_possible_cpu(i)
7519 unregister_fair_sched_group(tg, i);
7521 spin_lock_irqsave(&task_group_lock, flags);
7522 list_del_rcu(&tg->list);
7523 list_del_rcu(&tg->siblings);
7524 spin_unlock_irqrestore(&task_group_lock, flags);
7527 /* change task's runqueue when it moves between groups.
7528 * The caller of this function should have put the task in its new group
7529 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7530 * reflect its new group.
7532 void sched_move_task(struct task_struct *tsk)
7534 struct task_group *tg;
7535 int queued, running;
7536 unsigned long flags;
7539 rq = task_rq_lock(tsk, &flags);
7541 running = task_current(rq, tsk);
7542 queued = task_on_rq_queued(tsk);
7545 dequeue_task(rq, tsk, 0);
7546 if (unlikely(running))
7547 put_prev_task(rq, tsk);
7550 * All callers are synchronized by task_rq_lock(); we do not use RCU
7551 * which is pointless here. Thus, we pass "true" to task_css_check()
7552 * to prevent lockdep warnings.
7554 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7555 struct task_group, css);
7556 tg = autogroup_task_group(tsk, tg);
7557 tsk->sched_task_group = tg;
7559 #ifdef CONFIG_FAIR_GROUP_SCHED
7560 if (tsk->sched_class->task_move_group)
7561 tsk->sched_class->task_move_group(tsk, queued);
7564 set_task_rq(tsk, task_cpu(tsk));
7566 if (unlikely(running))
7567 tsk->sched_class->set_curr_task(rq);
7569 enqueue_task(rq, tsk, 0);
7571 task_rq_unlock(rq, tsk, &flags);
7573 #endif /* CONFIG_CGROUP_SCHED */
7575 #ifdef CONFIG_RT_GROUP_SCHED
7577 * Ensure that the real time constraints are schedulable.
7579 static DEFINE_MUTEX(rt_constraints_mutex);
7581 /* Must be called with tasklist_lock held */
7582 static inline int tg_has_rt_tasks(struct task_group *tg)
7584 struct task_struct *g, *p;
7587 * Autogroups do not have RT tasks; see autogroup_create().
7589 if (task_group_is_autogroup(tg))
7592 for_each_process_thread(g, p) {
7593 if (rt_task(p) && task_group(p) == tg)
7600 struct rt_schedulable_data {
7601 struct task_group *tg;
7606 static int tg_rt_schedulable(struct task_group *tg, void *data)
7608 struct rt_schedulable_data *d = data;
7609 struct task_group *child;
7610 unsigned long total, sum = 0;
7611 u64 period, runtime;
7613 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7614 runtime = tg->rt_bandwidth.rt_runtime;
7617 period = d->rt_period;
7618 runtime = d->rt_runtime;
7622 * Cannot have more runtime than the period.
7624 if (runtime > period && runtime != RUNTIME_INF)
7628 * Ensure we don't starve existing RT tasks.
7630 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7633 total = to_ratio(period, runtime);
7636 * Nobody can have more than the global setting allows.
7638 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7642 * The sum of our children's runtime should not exceed our own.
7644 list_for_each_entry_rcu(child, &tg->children, siblings) {
7645 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7646 runtime = child->rt_bandwidth.rt_runtime;
7648 if (child == d->tg) {
7649 period = d->rt_period;
7650 runtime = d->rt_runtime;
7653 sum += to_ratio(period, runtime);
7662 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7666 struct rt_schedulable_data data = {
7668 .rt_period = period,
7669 .rt_runtime = runtime,
7673 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7679 static int tg_set_rt_bandwidth(struct task_group *tg,
7680 u64 rt_period, u64 rt_runtime)
7685 * Disallowing the root group RT runtime is BAD, it would disallow the
7686 * kernel creating (and or operating) RT threads.
7688 if (tg == &root_task_group && rt_runtime == 0)
7691 /* No period doesn't make any sense. */
7695 mutex_lock(&rt_constraints_mutex);
7696 read_lock(&tasklist_lock);
7697 err = __rt_schedulable(tg, rt_period, rt_runtime);
7701 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7702 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7703 tg->rt_bandwidth.rt_runtime = rt_runtime;
7705 for_each_possible_cpu(i) {
7706 struct rt_rq *rt_rq = tg->rt_rq[i];
7708 raw_spin_lock(&rt_rq->rt_runtime_lock);
7709 rt_rq->rt_runtime = rt_runtime;
7710 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7712 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7714 read_unlock(&tasklist_lock);
7715 mutex_unlock(&rt_constraints_mutex);
7720 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7722 u64 rt_runtime, rt_period;
7724 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7725 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7726 if (rt_runtime_us < 0)
7727 rt_runtime = RUNTIME_INF;
7729 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7732 static long sched_group_rt_runtime(struct task_group *tg)
7736 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7739 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7740 do_div(rt_runtime_us, NSEC_PER_USEC);
7741 return rt_runtime_us;
7744 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7746 u64 rt_runtime, rt_period;
7748 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7749 rt_runtime = tg->rt_bandwidth.rt_runtime;
7751 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7754 static long sched_group_rt_period(struct task_group *tg)
7758 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7759 do_div(rt_period_us, NSEC_PER_USEC);
7760 return rt_period_us;
7762 #endif /* CONFIG_RT_GROUP_SCHED */
7764 #ifdef CONFIG_RT_GROUP_SCHED
7765 static int sched_rt_global_constraints(void)
7769 mutex_lock(&rt_constraints_mutex);
7770 read_lock(&tasklist_lock);
7771 ret = __rt_schedulable(NULL, 0, 0);
7772 read_unlock(&tasklist_lock);
7773 mutex_unlock(&rt_constraints_mutex);
7778 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7780 /* Don't accept realtime tasks when there is no way for them to run */
7781 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7787 #else /* !CONFIG_RT_GROUP_SCHED */
7788 static int sched_rt_global_constraints(void)
7790 unsigned long flags;
7793 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7794 for_each_possible_cpu(i) {
7795 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7797 raw_spin_lock(&rt_rq->rt_runtime_lock);
7798 rt_rq->rt_runtime = global_rt_runtime();
7799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7801 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7805 #endif /* CONFIG_RT_GROUP_SCHED */
7807 static int sched_dl_global_validate(void)
7809 u64 runtime = global_rt_runtime();
7810 u64 period = global_rt_period();
7811 u64 new_bw = to_ratio(period, runtime);
7814 unsigned long flags;
7817 * Here we want to check the bandwidth not being set to some
7818 * value smaller than the currently allocated bandwidth in
7819 * any of the root_domains.
7821 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7822 * cycling on root_domains... Discussion on different/better
7823 * solutions is welcome!
7825 for_each_possible_cpu(cpu) {
7826 rcu_read_lock_sched();
7827 dl_b = dl_bw_of(cpu);
7829 raw_spin_lock_irqsave(&dl_b->lock, flags);
7830 if (new_bw < dl_b->total_bw)
7832 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7834 rcu_read_unlock_sched();
7843 static void sched_dl_do_global(void)
7848 unsigned long flags;
7850 def_dl_bandwidth.dl_period = global_rt_period();
7851 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7853 if (global_rt_runtime() != RUNTIME_INF)
7854 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7857 * FIXME: As above...
7859 for_each_possible_cpu(cpu) {
7860 rcu_read_lock_sched();
7861 dl_b = dl_bw_of(cpu);
7863 raw_spin_lock_irqsave(&dl_b->lock, flags);
7865 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7867 rcu_read_unlock_sched();
7871 static int sched_rt_global_validate(void)
7873 if (sysctl_sched_rt_period <= 0)
7876 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7877 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7883 static void sched_rt_do_global(void)
7885 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7886 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7889 int sched_rt_handler(struct ctl_table *table, int write,
7890 void __user *buffer, size_t *lenp,
7893 int old_period, old_runtime;
7894 static DEFINE_MUTEX(mutex);
7898 old_period = sysctl_sched_rt_period;
7899 old_runtime = sysctl_sched_rt_runtime;
7901 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7903 if (!ret && write) {
7904 ret = sched_rt_global_validate();
7908 ret = sched_dl_global_validate();
7912 ret = sched_rt_global_constraints();
7916 sched_rt_do_global();
7917 sched_dl_do_global();
7921 sysctl_sched_rt_period = old_period;
7922 sysctl_sched_rt_runtime = old_runtime;
7924 mutex_unlock(&mutex);
7929 int sched_rr_handler(struct ctl_table *table, int write,
7930 void __user *buffer, size_t *lenp,
7934 static DEFINE_MUTEX(mutex);
7937 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7938 /* make sure that internally we keep jiffies */
7939 /* also, writing zero resets timeslice to default */
7940 if (!ret && write) {
7941 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7942 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7944 mutex_unlock(&mutex);
7948 #ifdef CONFIG_CGROUP_SCHED
7950 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7952 return css ? container_of(css, struct task_group, css) : NULL;
7955 static struct cgroup_subsys_state *
7956 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7958 struct task_group *parent = css_tg(parent_css);
7959 struct task_group *tg;
7962 /* This is early initialization for the top cgroup */
7963 return &root_task_group.css;
7966 tg = sched_create_group(parent);
7968 return ERR_PTR(-ENOMEM);
7973 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7975 struct task_group *tg = css_tg(css);
7976 struct task_group *parent = css_tg(css->parent);
7979 sched_online_group(tg, parent);
7983 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7985 struct task_group *tg = css_tg(css);
7987 sched_destroy_group(tg);
7990 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7992 struct task_group *tg = css_tg(css);
7994 sched_offline_group(tg);
7997 static void cpu_cgroup_fork(struct task_struct *task)
7999 sched_move_task(task);
8002 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8003 struct cgroup_taskset *tset)
8005 struct task_struct *task;
8007 cgroup_taskset_for_each(task, tset) {
8008 #ifdef CONFIG_RT_GROUP_SCHED
8009 if (!sched_rt_can_attach(css_tg(css), task))
8012 /* We don't support RT-tasks being in separate groups */
8013 if (task->sched_class != &fair_sched_class)
8020 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8021 struct cgroup_taskset *tset)
8023 struct task_struct *task;
8025 cgroup_taskset_for_each(task, tset)
8026 sched_move_task(task);
8029 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8030 struct cgroup_subsys_state *old_css,
8031 struct task_struct *task)
8034 * cgroup_exit() is called in the copy_process() failure path.
8035 * Ignore this case since the task hasn't ran yet, this avoids
8036 * trying to poke a half freed task state from generic code.
8038 if (!(task->flags & PF_EXITING))
8041 sched_move_task(task);
8044 #ifdef CONFIG_FAIR_GROUP_SCHED
8045 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8046 struct cftype *cftype, u64 shareval)
8048 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8051 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8054 struct task_group *tg = css_tg(css);
8056 return (u64) scale_load_down(tg->shares);
8059 #ifdef CONFIG_CFS_BANDWIDTH
8060 static DEFINE_MUTEX(cfs_constraints_mutex);
8062 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8063 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8065 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8067 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8069 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8070 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8072 if (tg == &root_task_group)
8076 * Ensure we have at some amount of bandwidth every period. This is
8077 * to prevent reaching a state of large arrears when throttled via
8078 * entity_tick() resulting in prolonged exit starvation.
8080 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8084 * Likewise, bound things on the otherside by preventing insane quota
8085 * periods. This also allows us to normalize in computing quota
8088 if (period > max_cfs_quota_period)
8092 * Prevent race between setting of cfs_rq->runtime_enabled and
8093 * unthrottle_offline_cfs_rqs().
8096 mutex_lock(&cfs_constraints_mutex);
8097 ret = __cfs_schedulable(tg, period, quota);
8101 runtime_enabled = quota != RUNTIME_INF;
8102 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8104 * If we need to toggle cfs_bandwidth_used, off->on must occur
8105 * before making related changes, and on->off must occur afterwards
8107 if (runtime_enabled && !runtime_was_enabled)
8108 cfs_bandwidth_usage_inc();
8109 raw_spin_lock_irq(&cfs_b->lock);
8110 cfs_b->period = ns_to_ktime(period);
8111 cfs_b->quota = quota;
8113 __refill_cfs_bandwidth_runtime(cfs_b);
8114 /* restart the period timer (if active) to handle new period expiry */
8115 if (runtime_enabled && cfs_b->timer_active) {
8116 /* force a reprogram */
8117 __start_cfs_bandwidth(cfs_b, true);
8119 raw_spin_unlock_irq(&cfs_b->lock);
8121 for_each_online_cpu(i) {
8122 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8123 struct rq *rq = cfs_rq->rq;
8125 raw_spin_lock_irq(&rq->lock);
8126 cfs_rq->runtime_enabled = runtime_enabled;
8127 cfs_rq->runtime_remaining = 0;
8129 if (cfs_rq->throttled)
8130 unthrottle_cfs_rq(cfs_rq);
8131 raw_spin_unlock_irq(&rq->lock);
8133 if (runtime_was_enabled && !runtime_enabled)
8134 cfs_bandwidth_usage_dec();
8136 mutex_unlock(&cfs_constraints_mutex);
8142 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8146 period = ktime_to_ns(tg->cfs_bandwidth.period);
8147 if (cfs_quota_us < 0)
8148 quota = RUNTIME_INF;
8150 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8152 return tg_set_cfs_bandwidth(tg, period, quota);
8155 long tg_get_cfs_quota(struct task_group *tg)
8159 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8162 quota_us = tg->cfs_bandwidth.quota;
8163 do_div(quota_us, NSEC_PER_USEC);
8168 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8172 period = (u64)cfs_period_us * NSEC_PER_USEC;
8173 quota = tg->cfs_bandwidth.quota;
8175 return tg_set_cfs_bandwidth(tg, period, quota);
8178 long tg_get_cfs_period(struct task_group *tg)
8182 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8183 do_div(cfs_period_us, NSEC_PER_USEC);
8185 return cfs_period_us;
8188 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8191 return tg_get_cfs_quota(css_tg(css));
8194 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8195 struct cftype *cftype, s64 cfs_quota_us)
8197 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8200 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8203 return tg_get_cfs_period(css_tg(css));
8206 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8207 struct cftype *cftype, u64 cfs_period_us)
8209 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8212 struct cfs_schedulable_data {
8213 struct task_group *tg;
8218 * normalize group quota/period to be quota/max_period
8219 * note: units are usecs
8221 static u64 normalize_cfs_quota(struct task_group *tg,
8222 struct cfs_schedulable_data *d)
8230 period = tg_get_cfs_period(tg);
8231 quota = tg_get_cfs_quota(tg);
8234 /* note: these should typically be equivalent */
8235 if (quota == RUNTIME_INF || quota == -1)
8238 return to_ratio(period, quota);
8241 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8243 struct cfs_schedulable_data *d = data;
8244 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8245 s64 quota = 0, parent_quota = -1;
8248 quota = RUNTIME_INF;
8250 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8252 quota = normalize_cfs_quota(tg, d);
8253 parent_quota = parent_b->hierarchical_quota;
8256 * ensure max(child_quota) <= parent_quota, inherit when no
8259 if (quota == RUNTIME_INF)
8260 quota = parent_quota;
8261 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8264 cfs_b->hierarchical_quota = quota;
8269 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8272 struct cfs_schedulable_data data = {
8278 if (quota != RUNTIME_INF) {
8279 do_div(data.period, NSEC_PER_USEC);
8280 do_div(data.quota, NSEC_PER_USEC);
8284 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8290 static int cpu_stats_show(struct seq_file *sf, void *v)
8292 struct task_group *tg = css_tg(seq_css(sf));
8293 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8295 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8296 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8297 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8301 #endif /* CONFIG_CFS_BANDWIDTH */
8302 #endif /* CONFIG_FAIR_GROUP_SCHED */
8304 #ifdef CONFIG_RT_GROUP_SCHED
8305 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8306 struct cftype *cft, s64 val)
8308 return sched_group_set_rt_runtime(css_tg(css), val);
8311 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8314 return sched_group_rt_runtime(css_tg(css));
8317 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8318 struct cftype *cftype, u64 rt_period_us)
8320 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8323 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8326 return sched_group_rt_period(css_tg(css));
8328 #endif /* CONFIG_RT_GROUP_SCHED */
8330 static struct cftype cpu_files[] = {
8331 #ifdef CONFIG_FAIR_GROUP_SCHED
8334 .read_u64 = cpu_shares_read_u64,
8335 .write_u64 = cpu_shares_write_u64,
8338 #ifdef CONFIG_CFS_BANDWIDTH
8340 .name = "cfs_quota_us",
8341 .read_s64 = cpu_cfs_quota_read_s64,
8342 .write_s64 = cpu_cfs_quota_write_s64,
8345 .name = "cfs_period_us",
8346 .read_u64 = cpu_cfs_period_read_u64,
8347 .write_u64 = cpu_cfs_period_write_u64,
8351 .seq_show = cpu_stats_show,
8354 #ifdef CONFIG_RT_GROUP_SCHED
8356 .name = "rt_runtime_us",
8357 .read_s64 = cpu_rt_runtime_read,
8358 .write_s64 = cpu_rt_runtime_write,
8361 .name = "rt_period_us",
8362 .read_u64 = cpu_rt_period_read_uint,
8363 .write_u64 = cpu_rt_period_write_uint,
8369 struct cgroup_subsys cpu_cgrp_subsys = {
8370 .css_alloc = cpu_cgroup_css_alloc,
8371 .css_free = cpu_cgroup_css_free,
8372 .css_online = cpu_cgroup_css_online,
8373 .css_offline = cpu_cgroup_css_offline,
8374 .fork = cpu_cgroup_fork,
8375 .can_attach = cpu_cgroup_can_attach,
8376 .attach = cpu_cgroup_attach,
8377 .exit = cpu_cgroup_exit,
8378 .legacy_cftypes = cpu_files,
8382 #endif /* CONFIG_CGROUP_SCHED */
8384 void dump_cpu_task(int cpu)
8386 pr_info("Task dump for CPU %d:\n", cpu);
8387 sched_show_task(cpu_curr(cpu));