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 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
143 update_rq_clock_task(rq, delta);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug unsigned int sysctl_sched_features =
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file *m, void *v)
173 for (i = 0; i < __SCHED_FEAT_NR; i++) {
174 if (!(sysctl_sched_features & (1UL << i)))
176 seq_printf(m, "%s ", sched_feat_names[i]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
192 #include "features.h"
197 static void sched_feat_disable(int i)
199 if (static_key_enabled(&sched_feat_keys[i]))
200 static_key_slow_dec(&sched_feat_keys[i]);
203 static void sched_feat_enable(int i)
205 if (!static_key_enabled(&sched_feat_keys[i]))
206 static_key_slow_inc(&sched_feat_keys[i]);
209 static void sched_feat_disable(int i) { };
210 static void sched_feat_enable(int i) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp)
218 if (strncmp(cmp, "NO_", 3) == 0) {
223 for (i = 0; i < __SCHED_FEAT_NR; i++) {
224 if (strcmp(cmp, sched_feat_names[i]) == 0) {
226 sysctl_sched_features &= ~(1UL << i);
227 sched_feat_disable(i);
229 sysctl_sched_features |= (1UL << i);
230 sched_feat_enable(i);
240 sched_feat_write(struct file *filp, const char __user *ubuf,
241 size_t cnt, loff_t *ppos)
250 if (copy_from_user(&buf, ubuf, cnt))
256 i = sched_feat_set(cmp);
257 if (i == __SCHED_FEAT_NR)
265 static int sched_feat_open(struct inode *inode, struct file *filp)
267 return single_open(filp, sched_feat_show, NULL);
270 static const struct file_operations sched_feat_fops = {
271 .open = sched_feat_open,
272 .write = sched_feat_write,
275 .release = single_release,
278 static __init int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL, NULL,
285 late_initcall(sched_init_debug);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug unsigned int sysctl_sched_nr_migrate = 32;
295 * period over which we average the RT time consumption, measured
300 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period = 1000000;
308 __read_mostly int scheduler_running;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime = 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq *__task_rq_lock(struct task_struct *p)
324 lockdep_assert_held(&p->pi_lock);
328 raw_spin_lock(&rq->lock);
329 if (likely(rq == task_rq(p)))
331 raw_spin_unlock(&rq->lock);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
339 __acquires(p->pi_lock)
345 raw_spin_lock_irqsave(&p->pi_lock, *flags);
347 raw_spin_lock(&rq->lock);
348 if (likely(rq == task_rq(p)))
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 static void __task_rq_unlock(struct rq *rq)
358 raw_spin_unlock(&rq->lock);
362 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
364 __releases(p->pi_lock)
366 raw_spin_unlock(&rq->lock);
367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq *this_rq_lock(void)
380 raw_spin_lock(&rq->lock);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq *rq)
392 if (hrtimer_active(&rq->hrtick_timer))
393 hrtimer_cancel(&rq->hrtick_timer);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart hrtick(struct hrtimer *timer)
402 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
406 raw_spin_lock(&rq->lock);
408 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
409 raw_spin_unlock(&rq->lock);
411 return HRTIMER_NORESTART;
416 static int __hrtick_restart(struct rq *rq)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = hrtimer_get_softexpires(timer);
421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg)
431 raw_spin_lock(&rq->lock);
432 __hrtick_restart(rq);
433 rq->hrtick_csd_pending = 0;
434 raw_spin_unlock(&rq->lock);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq *rq, u64 delay)
444 struct hrtimer *timer = &rq->hrtick_timer;
445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
447 hrtimer_set_expires(timer, time);
449 if (rq == this_rq()) {
450 __hrtick_restart(rq);
451 } else if (!rq->hrtick_csd_pending) {
452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
453 rq->hrtick_csd_pending = 1;
458 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
460 int cpu = (int)(long)hcpu;
463 case CPU_UP_CANCELED:
464 case CPU_UP_CANCELED_FROZEN:
465 case CPU_DOWN_PREPARE:
466 case CPU_DOWN_PREPARE_FROZEN:
468 case CPU_DEAD_FROZEN:
469 hrtick_clear(cpu_rq(cpu));
476 static __init void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq *rq, u64 delay)
488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
489 HRTIMER_MODE_REL_PINNED, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq *rq)
500 rq->hrtick_csd_pending = 0;
502 rq->hrtick_csd.flags = 0;
503 rq->hrtick_csd.func = __hrtick_start;
504 rq->hrtick_csd.info = rq;
507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
508 rq->hrtick_timer.function = hrtick;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq *rq)
515 static inline void init_rq_hrtick(struct rq *rq)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * cmpxchg based fetch_or, macro so it works for different integer types
527 #define fetch_or(ptr, val) \
528 ({ typeof(*(ptr)) __old, __val = *(ptr); \
530 __old = cmpxchg((ptr), __val, __val | (val)); \
531 if (__old == __val) \
538 #ifdef TIF_POLLING_NRFLAG
540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
541 * this avoids any races wrt polling state changes and thereby avoids
544 static bool set_nr_and_not_polling(struct task_struct *p)
546 struct thread_info *ti = task_thread_info(p);
547 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
550 static bool set_nr_and_not_polling(struct task_struct *p)
552 set_tsk_need_resched(p);
558 * resched_task - mark a task 'to be rescheduled now'.
560 * On UP this means the setting of the need_resched flag, on SMP it
561 * might also involve a cross-CPU call to trigger the scheduler on
564 void resched_task(struct task_struct *p)
568 lockdep_assert_held(&task_rq(p)->lock);
570 if (test_tsk_need_resched(p))
575 if (cpu == smp_processor_id()) {
576 set_tsk_need_resched(p);
577 set_preempt_need_resched();
581 if (set_nr_and_not_polling(p))
582 smp_send_reschedule(cpu);
585 void resched_cpu(int cpu)
587 struct rq *rq = cpu_rq(cpu);
590 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
592 resched_task(cpu_curr(cpu));
593 raw_spin_unlock_irqrestore(&rq->lock, flags);
597 #ifdef CONFIG_NO_HZ_COMMON
599 * In the semi idle case, use the nearest busy cpu for migrating timers
600 * from an idle cpu. This is good for power-savings.
602 * We don't do similar optimization for completely idle system, as
603 * selecting an idle cpu will add more delays to the timers than intended
604 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
606 int get_nohz_timer_target(int pinned)
608 int cpu = smp_processor_id();
610 struct sched_domain *sd;
612 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
616 for_each_domain(cpu, sd) {
617 for_each_cpu(i, sched_domain_span(sd)) {
629 * When add_timer_on() enqueues a timer into the timer wheel of an
630 * idle CPU then this timer might expire before the next timer event
631 * which is scheduled to wake up that CPU. In case of a completely
632 * idle system the next event might even be infinite time into the
633 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
634 * leaves the inner idle loop so the newly added timer is taken into
635 * account when the CPU goes back to idle and evaluates the timer
636 * wheel for the next timer event.
638 static void wake_up_idle_cpu(int cpu)
640 struct rq *rq = cpu_rq(cpu);
642 if (cpu == smp_processor_id())
646 * This is safe, as this function is called with the timer
647 * wheel base lock of (cpu) held. When the CPU is on the way
648 * to idle and has not yet set rq->curr to idle then it will
649 * be serialized on the timer wheel base lock and take the new
650 * timer into account automatically.
652 if (rq->curr != rq->idle)
656 * We can set TIF_RESCHED on the idle task of the other CPU
657 * lockless. The worst case is that the other CPU runs the
658 * idle task through an additional NOOP schedule()
660 set_tsk_need_resched(rq->idle);
662 /* NEED_RESCHED must be visible before we test polling */
664 if (!tsk_is_polling(rq->idle))
665 smp_send_reschedule(cpu);
668 static bool wake_up_full_nohz_cpu(int cpu)
670 if (tick_nohz_full_cpu(cpu)) {
671 if (cpu != smp_processor_id() ||
672 tick_nohz_tick_stopped())
673 smp_send_reschedule(cpu);
680 void wake_up_nohz_cpu(int cpu)
682 if (!wake_up_full_nohz_cpu(cpu))
683 wake_up_idle_cpu(cpu);
686 static inline bool got_nohz_idle_kick(void)
688 int cpu = smp_processor_id();
690 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
693 if (idle_cpu(cpu) && !need_resched())
697 * We can't run Idle Load Balance on this CPU for this time so we
698 * cancel it and clear NOHZ_BALANCE_KICK
700 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
704 #else /* CONFIG_NO_HZ_COMMON */
706 static inline bool got_nohz_idle_kick(void)
711 #endif /* CONFIG_NO_HZ_COMMON */
713 #ifdef CONFIG_NO_HZ_FULL
714 bool sched_can_stop_tick(void)
720 /* Make sure rq->nr_running update is visible after the IPI */
723 /* More than one running task need preemption */
724 if (rq->nr_running > 1)
729 #endif /* CONFIG_NO_HZ_FULL */
731 void sched_avg_update(struct rq *rq)
733 s64 period = sched_avg_period();
735 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
737 * Inline assembly required to prevent the compiler
738 * optimising this loop into a divmod call.
739 * See __iter_div_u64_rem() for another example of this.
741 asm("" : "+rm" (rq->age_stamp));
742 rq->age_stamp += period;
747 #endif /* CONFIG_SMP */
749 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
750 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
752 * Iterate task_group tree rooted at *from, calling @down when first entering a
753 * node and @up when leaving it for the final time.
755 * Caller must hold rcu_lock or sufficient equivalent.
757 int walk_tg_tree_from(struct task_group *from,
758 tg_visitor down, tg_visitor up, void *data)
760 struct task_group *parent, *child;
766 ret = (*down)(parent, data);
769 list_for_each_entry_rcu(child, &parent->children, siblings) {
776 ret = (*up)(parent, data);
777 if (ret || parent == from)
781 parent = parent->parent;
788 int tg_nop(struct task_group *tg, void *data)
794 static void set_load_weight(struct task_struct *p)
796 int prio = p->static_prio - MAX_RT_PRIO;
797 struct load_weight *load = &p->se.load;
800 * SCHED_IDLE tasks get minimal weight:
802 if (p->policy == SCHED_IDLE) {
803 load->weight = scale_load(WEIGHT_IDLEPRIO);
804 load->inv_weight = WMULT_IDLEPRIO;
808 load->weight = scale_load(prio_to_weight[prio]);
809 load->inv_weight = prio_to_wmult[prio];
812 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
815 sched_info_queued(rq, p);
816 p->sched_class->enqueue_task(rq, p, flags);
819 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
822 sched_info_dequeued(rq, p);
823 p->sched_class->dequeue_task(rq, p, flags);
826 void activate_task(struct rq *rq, struct task_struct *p, int flags)
828 if (task_contributes_to_load(p))
829 rq->nr_uninterruptible--;
831 enqueue_task(rq, p, flags);
834 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
836 if (task_contributes_to_load(p))
837 rq->nr_uninterruptible++;
839 dequeue_task(rq, p, flags);
842 static void update_rq_clock_task(struct rq *rq, s64 delta)
845 * In theory, the compile should just see 0 here, and optimize out the call
846 * to sched_rt_avg_update. But I don't trust it...
848 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
849 s64 steal = 0, irq_delta = 0;
851 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
852 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
855 * Since irq_time is only updated on {soft,}irq_exit, we might run into
856 * this case when a previous update_rq_clock() happened inside a
859 * When this happens, we stop ->clock_task and only update the
860 * prev_irq_time stamp to account for the part that fit, so that a next
861 * update will consume the rest. This ensures ->clock_task is
864 * It does however cause some slight miss-attribution of {soft,}irq
865 * time, a more accurate solution would be to update the irq_time using
866 * the current rq->clock timestamp, except that would require using
869 if (irq_delta > delta)
872 rq->prev_irq_time += irq_delta;
875 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
876 if (static_key_false((¶virt_steal_rq_enabled))) {
877 steal = paravirt_steal_clock(cpu_of(rq));
878 steal -= rq->prev_steal_time_rq;
880 if (unlikely(steal > delta))
883 rq->prev_steal_time_rq += steal;
888 rq->clock_task += delta;
890 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
891 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
892 sched_rt_avg_update(rq, irq_delta + steal);
896 void sched_set_stop_task(int cpu, struct task_struct *stop)
898 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
899 struct task_struct *old_stop = cpu_rq(cpu)->stop;
903 * Make it appear like a SCHED_FIFO task, its something
904 * userspace knows about and won't get confused about.
906 * Also, it will make PI more or less work without too
907 * much confusion -- but then, stop work should not
908 * rely on PI working anyway.
910 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
912 stop->sched_class = &stop_sched_class;
915 cpu_rq(cpu)->stop = stop;
919 * Reset it back to a normal scheduling class so that
920 * it can die in pieces.
922 old_stop->sched_class = &rt_sched_class;
927 * __normal_prio - return the priority that is based on the static prio
929 static inline int __normal_prio(struct task_struct *p)
931 return p->static_prio;
935 * Calculate the expected normal priority: i.e. priority
936 * without taking RT-inheritance into account. Might be
937 * boosted by interactivity modifiers. Changes upon fork,
938 * setprio syscalls, and whenever the interactivity
939 * estimator recalculates.
941 static inline int normal_prio(struct task_struct *p)
945 if (task_has_dl_policy(p))
946 prio = MAX_DL_PRIO-1;
947 else if (task_has_rt_policy(p))
948 prio = MAX_RT_PRIO-1 - p->rt_priority;
950 prio = __normal_prio(p);
955 * Calculate the current priority, i.e. the priority
956 * taken into account by the scheduler. This value might
957 * be boosted by RT tasks, or might be boosted by
958 * interactivity modifiers. Will be RT if the task got
959 * RT-boosted. If not then it returns p->normal_prio.
961 static int effective_prio(struct task_struct *p)
963 p->normal_prio = normal_prio(p);
965 * If we are RT tasks or we were boosted to RT priority,
966 * keep the priority unchanged. Otherwise, update priority
967 * to the normal priority:
969 if (!rt_prio(p->prio))
970 return p->normal_prio;
975 * task_curr - is this task currently executing on a CPU?
976 * @p: the task in question.
978 * Return: 1 if the task is currently executing. 0 otherwise.
980 inline int task_curr(const struct task_struct *p)
982 return cpu_curr(task_cpu(p)) == p;
985 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
986 const struct sched_class *prev_class,
989 if (prev_class != p->sched_class) {
990 if (prev_class->switched_from)
991 prev_class->switched_from(rq, p);
992 p->sched_class->switched_to(rq, p);
993 } else if (oldprio != p->prio || dl_task(p))
994 p->sched_class->prio_changed(rq, p, oldprio);
997 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
999 const struct sched_class *class;
1001 if (p->sched_class == rq->curr->sched_class) {
1002 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1004 for_each_class(class) {
1005 if (class == rq->curr->sched_class)
1007 if (class == p->sched_class) {
1008 resched_task(rq->curr);
1015 * A queue event has occurred, and we're going to schedule. In
1016 * this case, we can save a useless back to back clock update.
1018 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1019 rq->skip_clock_update = 1;
1023 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1025 #ifdef CONFIG_SCHED_DEBUG
1027 * We should never call set_task_cpu() on a blocked task,
1028 * ttwu() will sort out the placement.
1030 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1031 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1033 #ifdef CONFIG_LOCKDEP
1035 * The caller should hold either p->pi_lock or rq->lock, when changing
1036 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1038 * sched_move_task() holds both and thus holding either pins the cgroup,
1041 * Furthermore, all task_rq users should acquire both locks, see
1044 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1045 lockdep_is_held(&task_rq(p)->lock)));
1049 trace_sched_migrate_task(p, new_cpu);
1051 if (task_cpu(p) != new_cpu) {
1052 if (p->sched_class->migrate_task_rq)
1053 p->sched_class->migrate_task_rq(p, new_cpu);
1054 p->se.nr_migrations++;
1055 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1058 __set_task_cpu(p, new_cpu);
1061 static void __migrate_swap_task(struct task_struct *p, int cpu)
1064 struct rq *src_rq, *dst_rq;
1066 src_rq = task_rq(p);
1067 dst_rq = cpu_rq(cpu);
1069 deactivate_task(src_rq, p, 0);
1070 set_task_cpu(p, cpu);
1071 activate_task(dst_rq, p, 0);
1072 check_preempt_curr(dst_rq, p, 0);
1075 * Task isn't running anymore; make it appear like we migrated
1076 * it before it went to sleep. This means on wakeup we make the
1077 * previous cpu our targer instead of where it really is.
1083 struct migration_swap_arg {
1084 struct task_struct *src_task, *dst_task;
1085 int src_cpu, dst_cpu;
1088 static int migrate_swap_stop(void *data)
1090 struct migration_swap_arg *arg = data;
1091 struct rq *src_rq, *dst_rq;
1094 src_rq = cpu_rq(arg->src_cpu);
1095 dst_rq = cpu_rq(arg->dst_cpu);
1097 double_raw_lock(&arg->src_task->pi_lock,
1098 &arg->dst_task->pi_lock);
1099 double_rq_lock(src_rq, dst_rq);
1100 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1103 if (task_cpu(arg->src_task) != arg->src_cpu)
1106 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1109 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1112 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1113 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1118 double_rq_unlock(src_rq, dst_rq);
1119 raw_spin_unlock(&arg->dst_task->pi_lock);
1120 raw_spin_unlock(&arg->src_task->pi_lock);
1126 * Cross migrate two tasks
1128 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1130 struct migration_swap_arg arg;
1133 arg = (struct migration_swap_arg){
1135 .src_cpu = task_cpu(cur),
1137 .dst_cpu = task_cpu(p),
1140 if (arg.src_cpu == arg.dst_cpu)
1144 * These three tests are all lockless; this is OK since all of them
1145 * will be re-checked with proper locks held further down the line.
1147 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1150 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1153 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1156 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1157 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1163 struct migration_arg {
1164 struct task_struct *task;
1168 static int migration_cpu_stop(void *data);
1171 * wait_task_inactive - wait for a thread to unschedule.
1173 * If @match_state is nonzero, it's the @p->state value just checked and
1174 * not expected to change. If it changes, i.e. @p might have woken up,
1175 * then return zero. When we succeed in waiting for @p to be off its CPU,
1176 * we return a positive number (its total switch count). If a second call
1177 * a short while later returns the same number, the caller can be sure that
1178 * @p has remained unscheduled the whole time.
1180 * The caller must ensure that the task *will* unschedule sometime soon,
1181 * else this function might spin for a *long* time. This function can't
1182 * be called with interrupts off, or it may introduce deadlock with
1183 * smp_call_function() if an IPI is sent by the same process we are
1184 * waiting to become inactive.
1186 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1188 unsigned long flags;
1195 * We do the initial early heuristics without holding
1196 * any task-queue locks at all. We'll only try to get
1197 * the runqueue lock when things look like they will
1203 * If the task is actively running on another CPU
1204 * still, just relax and busy-wait without holding
1207 * NOTE! Since we don't hold any locks, it's not
1208 * even sure that "rq" stays as the right runqueue!
1209 * But we don't care, since "task_running()" will
1210 * return false if the runqueue has changed and p
1211 * is actually now running somewhere else!
1213 while (task_running(rq, p)) {
1214 if (match_state && unlikely(p->state != match_state))
1220 * Ok, time to look more closely! We need the rq
1221 * lock now, to be *sure*. If we're wrong, we'll
1222 * just go back and repeat.
1224 rq = task_rq_lock(p, &flags);
1225 trace_sched_wait_task(p);
1226 running = task_running(rq, p);
1229 if (!match_state || p->state == match_state)
1230 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1231 task_rq_unlock(rq, p, &flags);
1234 * If it changed from the expected state, bail out now.
1236 if (unlikely(!ncsw))
1240 * Was it really running after all now that we
1241 * checked with the proper locks actually held?
1243 * Oops. Go back and try again..
1245 if (unlikely(running)) {
1251 * It's not enough that it's not actively running,
1252 * it must be off the runqueue _entirely_, and not
1255 * So if it was still runnable (but just not actively
1256 * running right now), it's preempted, and we should
1257 * yield - it could be a while.
1259 if (unlikely(on_rq)) {
1260 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1262 set_current_state(TASK_UNINTERRUPTIBLE);
1263 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1268 * Ahh, all good. It wasn't running, and it wasn't
1269 * runnable, which means that it will never become
1270 * running in the future either. We're all done!
1279 * kick_process - kick a running thread to enter/exit the kernel
1280 * @p: the to-be-kicked thread
1282 * Cause a process which is running on another CPU to enter
1283 * kernel-mode, without any delay. (to get signals handled.)
1285 * NOTE: this function doesn't have to take the runqueue lock,
1286 * because all it wants to ensure is that the remote task enters
1287 * the kernel. If the IPI races and the task has been migrated
1288 * to another CPU then no harm is done and the purpose has been
1291 void kick_process(struct task_struct *p)
1297 if ((cpu != smp_processor_id()) && task_curr(p))
1298 smp_send_reschedule(cpu);
1301 EXPORT_SYMBOL_GPL(kick_process);
1302 #endif /* CONFIG_SMP */
1306 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1308 static int select_fallback_rq(int cpu, struct task_struct *p)
1310 int nid = cpu_to_node(cpu);
1311 const struct cpumask *nodemask = NULL;
1312 enum { cpuset, possible, fail } state = cpuset;
1316 * If the node that the cpu is on has been offlined, cpu_to_node()
1317 * will return -1. There is no cpu on the node, and we should
1318 * select the cpu on the other node.
1321 nodemask = cpumask_of_node(nid);
1323 /* Look for allowed, online CPU in same node. */
1324 for_each_cpu(dest_cpu, nodemask) {
1325 if (!cpu_online(dest_cpu))
1327 if (!cpu_active(dest_cpu))
1329 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1335 /* Any allowed, online CPU? */
1336 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1337 if (!cpu_online(dest_cpu))
1339 if (!cpu_active(dest_cpu))
1346 /* No more Mr. Nice Guy. */
1347 cpuset_cpus_allowed_fallback(p);
1352 do_set_cpus_allowed(p, cpu_possible_mask);
1363 if (state != cpuset) {
1365 * Don't tell them about moving exiting tasks or
1366 * kernel threads (both mm NULL), since they never
1369 if (p->mm && printk_ratelimit()) {
1370 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1371 task_pid_nr(p), p->comm, cpu);
1379 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1382 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1384 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1387 * In order not to call set_task_cpu() on a blocking task we need
1388 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1391 * Since this is common to all placement strategies, this lives here.
1393 * [ this allows ->select_task() to simply return task_cpu(p) and
1394 * not worry about this generic constraint ]
1396 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1398 cpu = select_fallback_rq(task_cpu(p), p);
1403 static void update_avg(u64 *avg, u64 sample)
1405 s64 diff = sample - *avg;
1411 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1413 #ifdef CONFIG_SCHEDSTATS
1414 struct rq *rq = this_rq();
1417 int this_cpu = smp_processor_id();
1419 if (cpu == this_cpu) {
1420 schedstat_inc(rq, ttwu_local);
1421 schedstat_inc(p, se.statistics.nr_wakeups_local);
1423 struct sched_domain *sd;
1425 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1427 for_each_domain(this_cpu, sd) {
1428 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1429 schedstat_inc(sd, ttwu_wake_remote);
1436 if (wake_flags & WF_MIGRATED)
1437 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1439 #endif /* CONFIG_SMP */
1441 schedstat_inc(rq, ttwu_count);
1442 schedstat_inc(p, se.statistics.nr_wakeups);
1444 if (wake_flags & WF_SYNC)
1445 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1447 #endif /* CONFIG_SCHEDSTATS */
1450 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1452 activate_task(rq, p, en_flags);
1455 /* if a worker is waking up, notify workqueue */
1456 if (p->flags & PF_WQ_WORKER)
1457 wq_worker_waking_up(p, cpu_of(rq));
1461 * Mark the task runnable and perform wakeup-preemption.
1464 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1466 check_preempt_curr(rq, p, wake_flags);
1467 trace_sched_wakeup(p, true);
1469 p->state = TASK_RUNNING;
1471 if (p->sched_class->task_woken)
1472 p->sched_class->task_woken(rq, p);
1474 if (rq->idle_stamp) {
1475 u64 delta = rq_clock(rq) - rq->idle_stamp;
1476 u64 max = 2*rq->max_idle_balance_cost;
1478 update_avg(&rq->avg_idle, delta);
1480 if (rq->avg_idle > max)
1489 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1492 if (p->sched_contributes_to_load)
1493 rq->nr_uninterruptible--;
1496 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1497 ttwu_do_wakeup(rq, p, wake_flags);
1501 * Called in case the task @p isn't fully descheduled from its runqueue,
1502 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1503 * since all we need to do is flip p->state to TASK_RUNNING, since
1504 * the task is still ->on_rq.
1506 static int ttwu_remote(struct task_struct *p, int wake_flags)
1511 rq = __task_rq_lock(p);
1513 /* check_preempt_curr() may use rq clock */
1514 update_rq_clock(rq);
1515 ttwu_do_wakeup(rq, p, wake_flags);
1518 __task_rq_unlock(rq);
1524 static void sched_ttwu_pending(void)
1526 struct rq *rq = this_rq();
1527 struct llist_node *llist = llist_del_all(&rq->wake_list);
1528 struct task_struct *p;
1530 raw_spin_lock(&rq->lock);
1533 p = llist_entry(llist, struct task_struct, wake_entry);
1534 llist = llist_next(llist);
1535 ttwu_do_activate(rq, p, 0);
1538 raw_spin_unlock(&rq->lock);
1541 void scheduler_ipi(void)
1544 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1545 * TIF_NEED_RESCHED remotely (for the first time) will also send
1548 preempt_fold_need_resched();
1550 if (llist_empty(&this_rq()->wake_list)
1551 && !tick_nohz_full_cpu(smp_processor_id())
1552 && !got_nohz_idle_kick())
1556 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1557 * traditionally all their work was done from the interrupt return
1558 * path. Now that we actually do some work, we need to make sure
1561 * Some archs already do call them, luckily irq_enter/exit nest
1564 * Arguably we should visit all archs and update all handlers,
1565 * however a fair share of IPIs are still resched only so this would
1566 * somewhat pessimize the simple resched case.
1569 tick_nohz_full_check();
1570 sched_ttwu_pending();
1573 * Check if someone kicked us for doing the nohz idle load balance.
1575 if (unlikely(got_nohz_idle_kick())) {
1576 this_rq()->idle_balance = 1;
1577 raise_softirq_irqoff(SCHED_SOFTIRQ);
1582 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1584 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1585 smp_send_reschedule(cpu);
1588 bool cpus_share_cache(int this_cpu, int that_cpu)
1590 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1592 #endif /* CONFIG_SMP */
1594 static void ttwu_queue(struct task_struct *p, int cpu)
1596 struct rq *rq = cpu_rq(cpu);
1598 #if defined(CONFIG_SMP)
1599 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1600 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1601 ttwu_queue_remote(p, cpu);
1606 raw_spin_lock(&rq->lock);
1607 ttwu_do_activate(rq, p, 0);
1608 raw_spin_unlock(&rq->lock);
1612 * try_to_wake_up - wake up a thread
1613 * @p: the thread to be awakened
1614 * @state: the mask of task states that can be woken
1615 * @wake_flags: wake modifier flags (WF_*)
1617 * Put it on the run-queue if it's not already there. The "current"
1618 * thread is always on the run-queue (except when the actual
1619 * re-schedule is in progress), and as such you're allowed to do
1620 * the simpler "current->state = TASK_RUNNING" to mark yourself
1621 * runnable without the overhead of this.
1623 * Return: %true if @p was woken up, %false if it was already running.
1624 * or @state didn't match @p's state.
1627 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1629 unsigned long flags;
1630 int cpu, success = 0;
1633 * If we are going to wake up a thread waiting for CONDITION we
1634 * need to ensure that CONDITION=1 done by the caller can not be
1635 * reordered with p->state check below. This pairs with mb() in
1636 * set_current_state() the waiting thread does.
1638 smp_mb__before_spinlock();
1639 raw_spin_lock_irqsave(&p->pi_lock, flags);
1640 if (!(p->state & state))
1643 success = 1; /* we're going to change ->state */
1646 if (p->on_rq && ttwu_remote(p, wake_flags))
1651 * If the owning (remote) cpu is still in the middle of schedule() with
1652 * this task as prev, wait until its done referencing the task.
1657 * Pairs with the smp_wmb() in finish_lock_switch().
1661 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1662 p->state = TASK_WAKING;
1664 if (p->sched_class->task_waking)
1665 p->sched_class->task_waking(p);
1667 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1668 if (task_cpu(p) != cpu) {
1669 wake_flags |= WF_MIGRATED;
1670 set_task_cpu(p, cpu);
1672 #endif /* CONFIG_SMP */
1676 ttwu_stat(p, cpu, wake_flags);
1678 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1684 * try_to_wake_up_local - try to wake up a local task with rq lock held
1685 * @p: the thread to be awakened
1687 * Put @p on the run-queue if it's not already there. The caller must
1688 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1691 static void try_to_wake_up_local(struct task_struct *p)
1693 struct rq *rq = task_rq(p);
1695 if (WARN_ON_ONCE(rq != this_rq()) ||
1696 WARN_ON_ONCE(p == current))
1699 lockdep_assert_held(&rq->lock);
1701 if (!raw_spin_trylock(&p->pi_lock)) {
1702 raw_spin_unlock(&rq->lock);
1703 raw_spin_lock(&p->pi_lock);
1704 raw_spin_lock(&rq->lock);
1707 if (!(p->state & TASK_NORMAL))
1711 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1713 ttwu_do_wakeup(rq, p, 0);
1714 ttwu_stat(p, smp_processor_id(), 0);
1716 raw_spin_unlock(&p->pi_lock);
1720 * wake_up_process - Wake up a specific process
1721 * @p: The process to be woken up.
1723 * Attempt to wake up the nominated process and move it to the set of runnable
1726 * Return: 1 if the process was woken up, 0 if it was already running.
1728 * It may be assumed that this function implies a write memory barrier before
1729 * changing the task state if and only if any tasks are woken up.
1731 int wake_up_process(struct task_struct *p)
1733 WARN_ON(task_is_stopped_or_traced(p));
1734 return try_to_wake_up(p, TASK_NORMAL, 0);
1736 EXPORT_SYMBOL(wake_up_process);
1738 int wake_up_state(struct task_struct *p, unsigned int state)
1740 return try_to_wake_up(p, state, 0);
1744 * Perform scheduler related setup for a newly forked process p.
1745 * p is forked by current.
1747 * __sched_fork() is basic setup used by init_idle() too:
1749 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1754 p->se.exec_start = 0;
1755 p->se.sum_exec_runtime = 0;
1756 p->se.prev_sum_exec_runtime = 0;
1757 p->se.nr_migrations = 0;
1759 INIT_LIST_HEAD(&p->se.group_node);
1761 #ifdef CONFIG_SCHEDSTATS
1762 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1765 RB_CLEAR_NODE(&p->dl.rb_node);
1766 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1767 p->dl.dl_runtime = p->dl.runtime = 0;
1768 p->dl.dl_deadline = p->dl.deadline = 0;
1769 p->dl.dl_period = 0;
1772 INIT_LIST_HEAD(&p->rt.run_list);
1774 #ifdef CONFIG_PREEMPT_NOTIFIERS
1775 INIT_HLIST_HEAD(&p->preempt_notifiers);
1778 #ifdef CONFIG_NUMA_BALANCING
1779 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1780 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1781 p->mm->numa_scan_seq = 0;
1784 if (clone_flags & CLONE_VM)
1785 p->numa_preferred_nid = current->numa_preferred_nid;
1787 p->numa_preferred_nid = -1;
1789 p->node_stamp = 0ULL;
1790 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1791 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1792 p->numa_work.next = &p->numa_work;
1793 p->numa_faults_memory = NULL;
1794 p->numa_faults_buffer_memory = NULL;
1795 p->last_task_numa_placement = 0;
1796 p->last_sum_exec_runtime = 0;
1798 INIT_LIST_HEAD(&p->numa_entry);
1799 p->numa_group = NULL;
1800 #endif /* CONFIG_NUMA_BALANCING */
1803 #ifdef CONFIG_NUMA_BALANCING
1804 #ifdef CONFIG_SCHED_DEBUG
1805 void set_numabalancing_state(bool enabled)
1808 sched_feat_set("NUMA");
1810 sched_feat_set("NO_NUMA");
1813 __read_mostly bool numabalancing_enabled;
1815 void set_numabalancing_state(bool enabled)
1817 numabalancing_enabled = enabled;
1819 #endif /* CONFIG_SCHED_DEBUG */
1821 #ifdef CONFIG_PROC_SYSCTL
1822 int sysctl_numa_balancing(struct ctl_table *table, int write,
1823 void __user *buffer, size_t *lenp, loff_t *ppos)
1827 int state = numabalancing_enabled;
1829 if (write && !capable(CAP_SYS_ADMIN))
1834 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1838 set_numabalancing_state(state);
1845 * fork()/clone()-time setup:
1847 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1849 unsigned long flags;
1850 int cpu = get_cpu();
1852 __sched_fork(clone_flags, p);
1854 * We mark the process as running here. This guarantees that
1855 * nobody will actually run it, and a signal or other external
1856 * event cannot wake it up and insert it on the runqueue either.
1858 p->state = TASK_RUNNING;
1861 * Make sure we do not leak PI boosting priority to the child.
1863 p->prio = current->normal_prio;
1866 * Revert to default priority/policy on fork if requested.
1868 if (unlikely(p->sched_reset_on_fork)) {
1869 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1870 p->policy = SCHED_NORMAL;
1871 p->static_prio = NICE_TO_PRIO(0);
1873 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1874 p->static_prio = NICE_TO_PRIO(0);
1876 p->prio = p->normal_prio = __normal_prio(p);
1880 * We don't need the reset flag anymore after the fork. It has
1881 * fulfilled its duty:
1883 p->sched_reset_on_fork = 0;
1886 if (dl_prio(p->prio)) {
1889 } else if (rt_prio(p->prio)) {
1890 p->sched_class = &rt_sched_class;
1892 p->sched_class = &fair_sched_class;
1895 if (p->sched_class->task_fork)
1896 p->sched_class->task_fork(p);
1899 * The child is not yet in the pid-hash so no cgroup attach races,
1900 * and the cgroup is pinned to this child due to cgroup_fork()
1901 * is ran before sched_fork().
1903 * Silence PROVE_RCU.
1905 raw_spin_lock_irqsave(&p->pi_lock, flags);
1906 set_task_cpu(p, cpu);
1907 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1909 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1910 if (likely(sched_info_on()))
1911 memset(&p->sched_info, 0, sizeof(p->sched_info));
1913 #if defined(CONFIG_SMP)
1916 init_task_preempt_count(p);
1918 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1919 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1926 unsigned long to_ratio(u64 period, u64 runtime)
1928 if (runtime == RUNTIME_INF)
1932 * Doing this here saves a lot of checks in all
1933 * the calling paths, and returning zero seems
1934 * safe for them anyway.
1939 return div64_u64(runtime << 20, period);
1943 inline struct dl_bw *dl_bw_of(int i)
1945 return &cpu_rq(i)->rd->dl_bw;
1948 static inline int dl_bw_cpus(int i)
1950 struct root_domain *rd = cpu_rq(i)->rd;
1953 for_each_cpu_and(i, rd->span, cpu_active_mask)
1959 inline struct dl_bw *dl_bw_of(int i)
1961 return &cpu_rq(i)->dl.dl_bw;
1964 static inline int dl_bw_cpus(int i)
1971 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1973 dl_b->total_bw -= tsk_bw;
1977 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1979 dl_b->total_bw += tsk_bw;
1983 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1985 return dl_b->bw != -1 &&
1986 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1990 * We must be sure that accepting a new task (or allowing changing the
1991 * parameters of an existing one) is consistent with the bandwidth
1992 * constraints. If yes, this function also accordingly updates the currently
1993 * allocated bandwidth to reflect the new situation.
1995 * This function is called while holding p's rq->lock.
1997 static int dl_overflow(struct task_struct *p, int policy,
1998 const struct sched_attr *attr)
2001 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2002 u64 period = attr->sched_period ?: attr->sched_deadline;
2003 u64 runtime = attr->sched_runtime;
2004 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2007 if (new_bw == p->dl.dl_bw)
2011 * Either if a task, enters, leave, or stays -deadline but changes
2012 * its parameters, we may need to update accordingly the total
2013 * allocated bandwidth of the container.
2015 raw_spin_lock(&dl_b->lock);
2016 cpus = dl_bw_cpus(task_cpu(p));
2017 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2018 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2019 __dl_add(dl_b, new_bw);
2021 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2022 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2023 __dl_clear(dl_b, p->dl.dl_bw);
2024 __dl_add(dl_b, new_bw);
2026 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2027 __dl_clear(dl_b, p->dl.dl_bw);
2030 raw_spin_unlock(&dl_b->lock);
2035 extern void init_dl_bw(struct dl_bw *dl_b);
2038 * wake_up_new_task - wake up a newly created task for the first time.
2040 * This function will do some initial scheduler statistics housekeeping
2041 * that must be done for every newly created context, then puts the task
2042 * on the runqueue and wakes it.
2044 void wake_up_new_task(struct task_struct *p)
2046 unsigned long flags;
2049 raw_spin_lock_irqsave(&p->pi_lock, flags);
2052 * Fork balancing, do it here and not earlier because:
2053 * - cpus_allowed can change in the fork path
2054 * - any previously selected cpu might disappear through hotplug
2056 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2059 /* Initialize new task's runnable average */
2060 init_task_runnable_average(p);
2061 rq = __task_rq_lock(p);
2062 activate_task(rq, p, 0);
2064 trace_sched_wakeup_new(p, true);
2065 check_preempt_curr(rq, p, WF_FORK);
2067 if (p->sched_class->task_woken)
2068 p->sched_class->task_woken(rq, p);
2070 task_rq_unlock(rq, p, &flags);
2073 #ifdef CONFIG_PREEMPT_NOTIFIERS
2076 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2077 * @notifier: notifier struct to register
2079 void preempt_notifier_register(struct preempt_notifier *notifier)
2081 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2083 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2086 * preempt_notifier_unregister - no longer interested in preemption notifications
2087 * @notifier: notifier struct to unregister
2089 * This is safe to call from within a preemption notifier.
2091 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2093 hlist_del(¬ifier->link);
2095 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2097 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2099 struct preempt_notifier *notifier;
2101 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2102 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2106 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2107 struct task_struct *next)
2109 struct preempt_notifier *notifier;
2111 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2112 notifier->ops->sched_out(notifier, next);
2115 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2117 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2122 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2123 struct task_struct *next)
2127 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2130 * prepare_task_switch - prepare to switch tasks
2131 * @rq: the runqueue preparing to switch
2132 * @prev: the current task that is being switched out
2133 * @next: the task we are going to switch to.
2135 * This is called with the rq lock held and interrupts off. It must
2136 * be paired with a subsequent finish_task_switch after the context
2139 * prepare_task_switch sets up locking and calls architecture specific
2143 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2144 struct task_struct *next)
2146 trace_sched_switch(prev, next);
2147 sched_info_switch(rq, prev, next);
2148 perf_event_task_sched_out(prev, next);
2149 fire_sched_out_preempt_notifiers(prev, next);
2150 prepare_lock_switch(rq, next);
2151 prepare_arch_switch(next);
2155 * finish_task_switch - clean up after a task-switch
2156 * @rq: runqueue associated with task-switch
2157 * @prev: the thread we just switched away from.
2159 * finish_task_switch must be called after the context switch, paired
2160 * with a prepare_task_switch call before the context switch.
2161 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2162 * and do any other architecture-specific cleanup actions.
2164 * Note that we may have delayed dropping an mm in context_switch(). If
2165 * so, we finish that here outside of the runqueue lock. (Doing it
2166 * with the lock held can cause deadlocks; see schedule() for
2169 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2170 __releases(rq->lock)
2172 struct mm_struct *mm = rq->prev_mm;
2178 * A task struct has one reference for the use as "current".
2179 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2180 * schedule one last time. The schedule call will never return, and
2181 * the scheduled task must drop that reference.
2182 * The test for TASK_DEAD must occur while the runqueue locks are
2183 * still held, otherwise prev could be scheduled on another cpu, die
2184 * there before we look at prev->state, and then the reference would
2186 * Manfred Spraul <manfred@colorfullife.com>
2188 prev_state = prev->state;
2189 vtime_task_switch(prev);
2190 finish_arch_switch(prev);
2191 perf_event_task_sched_in(prev, current);
2192 finish_lock_switch(rq, prev);
2193 finish_arch_post_lock_switch();
2195 fire_sched_in_preempt_notifiers(current);
2198 if (unlikely(prev_state == TASK_DEAD)) {
2199 if (prev->sched_class->task_dead)
2200 prev->sched_class->task_dead(prev);
2203 * Remove function-return probe instances associated with this
2204 * task and put them back on the free list.
2206 kprobe_flush_task(prev);
2207 put_task_struct(prev);
2210 tick_nohz_task_switch(current);
2215 /* rq->lock is NOT held, but preemption is disabled */
2216 static inline void post_schedule(struct rq *rq)
2218 if (rq->post_schedule) {
2219 unsigned long flags;
2221 raw_spin_lock_irqsave(&rq->lock, flags);
2222 if (rq->curr->sched_class->post_schedule)
2223 rq->curr->sched_class->post_schedule(rq);
2224 raw_spin_unlock_irqrestore(&rq->lock, flags);
2226 rq->post_schedule = 0;
2232 static inline void post_schedule(struct rq *rq)
2239 * schedule_tail - first thing a freshly forked thread must call.
2240 * @prev: the thread we just switched away from.
2242 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2243 __releases(rq->lock)
2245 struct rq *rq = this_rq();
2247 finish_task_switch(rq, prev);
2250 * FIXME: do we need to worry about rq being invalidated by the
2255 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2256 /* In this case, finish_task_switch does not reenable preemption */
2259 if (current->set_child_tid)
2260 put_user(task_pid_vnr(current), current->set_child_tid);
2264 * context_switch - switch to the new MM and the new
2265 * thread's register state.
2268 context_switch(struct rq *rq, struct task_struct *prev,
2269 struct task_struct *next)
2271 struct mm_struct *mm, *oldmm;
2273 prepare_task_switch(rq, prev, next);
2276 oldmm = prev->active_mm;
2278 * For paravirt, this is coupled with an exit in switch_to to
2279 * combine the page table reload and the switch backend into
2282 arch_start_context_switch(prev);
2285 next->active_mm = oldmm;
2286 atomic_inc(&oldmm->mm_count);
2287 enter_lazy_tlb(oldmm, next);
2289 switch_mm(oldmm, mm, next);
2292 prev->active_mm = NULL;
2293 rq->prev_mm = oldmm;
2296 * Since the runqueue lock will be released by the next
2297 * task (which is an invalid locking op but in the case
2298 * of the scheduler it's an obvious special-case), so we
2299 * do an early lockdep release here:
2301 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2302 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2305 context_tracking_task_switch(prev, next);
2306 /* Here we just switch the register state and the stack. */
2307 switch_to(prev, next, prev);
2311 * this_rq must be evaluated again because prev may have moved
2312 * CPUs since it called schedule(), thus the 'rq' on its stack
2313 * frame will be invalid.
2315 finish_task_switch(this_rq(), prev);
2319 * nr_running and nr_context_switches:
2321 * externally visible scheduler statistics: current number of runnable
2322 * threads, total number of context switches performed since bootup.
2324 unsigned long nr_running(void)
2326 unsigned long i, sum = 0;
2328 for_each_online_cpu(i)
2329 sum += cpu_rq(i)->nr_running;
2334 unsigned long long nr_context_switches(void)
2337 unsigned long long sum = 0;
2339 for_each_possible_cpu(i)
2340 sum += cpu_rq(i)->nr_switches;
2345 unsigned long nr_iowait(void)
2347 unsigned long i, sum = 0;
2349 for_each_possible_cpu(i)
2350 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2355 unsigned long nr_iowait_cpu(int cpu)
2357 struct rq *this = cpu_rq(cpu);
2358 return atomic_read(&this->nr_iowait);
2364 * sched_exec - execve() is a valuable balancing opportunity, because at
2365 * this point the task has the smallest effective memory and cache footprint.
2367 void sched_exec(void)
2369 struct task_struct *p = current;
2370 unsigned long flags;
2373 raw_spin_lock_irqsave(&p->pi_lock, flags);
2374 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2375 if (dest_cpu == smp_processor_id())
2378 if (likely(cpu_active(dest_cpu))) {
2379 struct migration_arg arg = { p, dest_cpu };
2381 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2382 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2386 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2391 DEFINE_PER_CPU(struct kernel_stat, kstat);
2392 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2394 EXPORT_PER_CPU_SYMBOL(kstat);
2395 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2398 * Return any ns on the sched_clock that have not yet been accounted in
2399 * @p in case that task is currently running.
2401 * Called with task_rq_lock() held on @rq.
2403 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2407 if (task_current(rq, p)) {
2408 update_rq_clock(rq);
2409 ns = rq_clock_task(rq) - p->se.exec_start;
2417 unsigned long long task_delta_exec(struct task_struct *p)
2419 unsigned long flags;
2423 rq = task_rq_lock(p, &flags);
2424 ns = do_task_delta_exec(p, rq);
2425 task_rq_unlock(rq, p, &flags);
2431 * Return accounted runtime for the task.
2432 * In case the task is currently running, return the runtime plus current's
2433 * pending runtime that have not been accounted yet.
2435 unsigned long long task_sched_runtime(struct task_struct *p)
2437 unsigned long flags;
2441 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2443 * 64-bit doesn't need locks to atomically read a 64bit value.
2444 * So we have a optimization chance when the task's delta_exec is 0.
2445 * Reading ->on_cpu is racy, but this is ok.
2447 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2448 * If we race with it entering cpu, unaccounted time is 0. This is
2449 * indistinguishable from the read occurring a few cycles earlier.
2452 return p->se.sum_exec_runtime;
2455 rq = task_rq_lock(p, &flags);
2456 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2457 task_rq_unlock(rq, p, &flags);
2463 * This function gets called by the timer code, with HZ frequency.
2464 * We call it with interrupts disabled.
2466 void scheduler_tick(void)
2468 int cpu = smp_processor_id();
2469 struct rq *rq = cpu_rq(cpu);
2470 struct task_struct *curr = rq->curr;
2474 raw_spin_lock(&rq->lock);
2475 update_rq_clock(rq);
2476 curr->sched_class->task_tick(rq, curr, 0);
2477 update_cpu_load_active(rq);
2478 raw_spin_unlock(&rq->lock);
2480 perf_event_task_tick();
2483 rq->idle_balance = idle_cpu(cpu);
2484 trigger_load_balance(rq);
2486 rq_last_tick_reset(rq);
2489 #ifdef CONFIG_NO_HZ_FULL
2491 * scheduler_tick_max_deferment
2493 * Keep at least one tick per second when a single
2494 * active task is running because the scheduler doesn't
2495 * yet completely support full dynticks environment.
2497 * This makes sure that uptime, CFS vruntime, load
2498 * balancing, etc... continue to move forward, even
2499 * with a very low granularity.
2501 * Return: Maximum deferment in nanoseconds.
2503 u64 scheduler_tick_max_deferment(void)
2505 struct rq *rq = this_rq();
2506 unsigned long next, now = ACCESS_ONCE(jiffies);
2508 next = rq->last_sched_tick + HZ;
2510 if (time_before_eq(next, now))
2513 return jiffies_to_nsecs(next - now);
2517 notrace unsigned long get_parent_ip(unsigned long addr)
2519 if (in_lock_functions(addr)) {
2520 addr = CALLER_ADDR2;
2521 if (in_lock_functions(addr))
2522 addr = CALLER_ADDR3;
2527 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2528 defined(CONFIG_PREEMPT_TRACER))
2530 void preempt_count_add(int val)
2532 #ifdef CONFIG_DEBUG_PREEMPT
2536 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2539 __preempt_count_add(val);
2540 #ifdef CONFIG_DEBUG_PREEMPT
2542 * Spinlock count overflowing soon?
2544 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2547 if (preempt_count() == val) {
2548 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550 current->preempt_disable_ip = ip;
2552 trace_preempt_off(CALLER_ADDR0, ip);
2555 EXPORT_SYMBOL(preempt_count_add);
2556 NOKPROBE_SYMBOL(preempt_count_add);
2558 void preempt_count_sub(int val)
2560 #ifdef CONFIG_DEBUG_PREEMPT
2564 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2567 * Is the spinlock portion underflowing?
2569 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2570 !(preempt_count() & PREEMPT_MASK)))
2574 if (preempt_count() == val)
2575 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2576 __preempt_count_sub(val);
2578 EXPORT_SYMBOL(preempt_count_sub);
2579 NOKPROBE_SYMBOL(preempt_count_sub);
2584 * Print scheduling while atomic bug:
2586 static noinline void __schedule_bug(struct task_struct *prev)
2588 if (oops_in_progress)
2591 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2592 prev->comm, prev->pid, preempt_count());
2594 debug_show_held_locks(prev);
2596 if (irqs_disabled())
2597 print_irqtrace_events(prev);
2598 #ifdef CONFIG_DEBUG_PREEMPT
2599 if (in_atomic_preempt_off()) {
2600 pr_err("Preemption disabled at:");
2601 print_ip_sym(current->preempt_disable_ip);
2606 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2610 * Various schedule()-time debugging checks and statistics:
2612 static inline void schedule_debug(struct task_struct *prev)
2615 * Test if we are atomic. Since do_exit() needs to call into
2616 * schedule() atomically, we ignore that path. Otherwise whine
2617 * if we are scheduling when we should not.
2619 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2620 __schedule_bug(prev);
2623 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2625 schedstat_inc(this_rq(), sched_count);
2629 * Pick up the highest-prio task:
2631 static inline struct task_struct *
2632 pick_next_task(struct rq *rq, struct task_struct *prev)
2634 const struct sched_class *class = &fair_sched_class;
2635 struct task_struct *p;
2638 * Optimization: we know that if all tasks are in
2639 * the fair class we can call that function directly:
2641 if (likely(prev->sched_class == class &&
2642 rq->nr_running == rq->cfs.h_nr_running)) {
2643 p = fair_sched_class.pick_next_task(rq, prev);
2644 if (unlikely(p == RETRY_TASK))
2647 /* assumes fair_sched_class->next == idle_sched_class */
2649 p = idle_sched_class.pick_next_task(rq, prev);
2655 for_each_class(class) {
2656 p = class->pick_next_task(rq, prev);
2658 if (unlikely(p == RETRY_TASK))
2664 BUG(); /* the idle class will always have a runnable task */
2668 * __schedule() is the main scheduler function.
2670 * The main means of driving the scheduler and thus entering this function are:
2672 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2674 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2675 * paths. For example, see arch/x86/entry_64.S.
2677 * To drive preemption between tasks, the scheduler sets the flag in timer
2678 * interrupt handler scheduler_tick().
2680 * 3. Wakeups don't really cause entry into schedule(). They add a
2681 * task to the run-queue and that's it.
2683 * Now, if the new task added to the run-queue preempts the current
2684 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2685 * called on the nearest possible occasion:
2687 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2689 * - in syscall or exception context, at the next outmost
2690 * preempt_enable(). (this might be as soon as the wake_up()'s
2693 * - in IRQ context, return from interrupt-handler to
2694 * preemptible context
2696 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2699 * - cond_resched() call
2700 * - explicit schedule() call
2701 * - return from syscall or exception to user-space
2702 * - return from interrupt-handler to user-space
2704 static void __sched __schedule(void)
2706 struct task_struct *prev, *next;
2707 unsigned long *switch_count;
2713 cpu = smp_processor_id();
2715 rcu_note_context_switch(cpu);
2718 schedule_debug(prev);
2720 if (sched_feat(HRTICK))
2724 * Make sure that signal_pending_state()->signal_pending() below
2725 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2726 * done by the caller to avoid the race with signal_wake_up().
2728 smp_mb__before_spinlock();
2729 raw_spin_lock_irq(&rq->lock);
2731 switch_count = &prev->nivcsw;
2732 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2733 if (unlikely(signal_pending_state(prev->state, prev))) {
2734 prev->state = TASK_RUNNING;
2736 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2740 * If a worker went to sleep, notify and ask workqueue
2741 * whether it wants to wake up a task to maintain
2744 if (prev->flags & PF_WQ_WORKER) {
2745 struct task_struct *to_wakeup;
2747 to_wakeup = wq_worker_sleeping(prev, cpu);
2749 try_to_wake_up_local(to_wakeup);
2752 switch_count = &prev->nvcsw;
2755 if (prev->on_rq || rq->skip_clock_update < 0)
2756 update_rq_clock(rq);
2758 next = pick_next_task(rq, prev);
2759 clear_tsk_need_resched(prev);
2760 clear_preempt_need_resched();
2761 rq->skip_clock_update = 0;
2763 if (likely(prev != next)) {
2768 context_switch(rq, prev, next); /* unlocks the rq */
2770 * The context switch have flipped the stack from under us
2771 * and restored the local variables which were saved when
2772 * this task called schedule() in the past. prev == current
2773 * is still correct, but it can be moved to another cpu/rq.
2775 cpu = smp_processor_id();
2778 raw_spin_unlock_irq(&rq->lock);
2782 sched_preempt_enable_no_resched();
2787 static inline void sched_submit_work(struct task_struct *tsk)
2789 if (!tsk->state || tsk_is_pi_blocked(tsk))
2792 * If we are going to sleep and we have plugged IO queued,
2793 * make sure to submit it to avoid deadlocks.
2795 if (blk_needs_flush_plug(tsk))
2796 blk_schedule_flush_plug(tsk);
2799 asmlinkage __visible void __sched schedule(void)
2801 struct task_struct *tsk = current;
2803 sched_submit_work(tsk);
2806 EXPORT_SYMBOL(schedule);
2808 #ifdef CONFIG_CONTEXT_TRACKING
2809 asmlinkage __visible void __sched schedule_user(void)
2812 * If we come here after a random call to set_need_resched(),
2813 * or we have been woken up remotely but the IPI has not yet arrived,
2814 * we haven't yet exited the RCU idle mode. Do it here manually until
2815 * we find a better solution.
2824 * schedule_preempt_disabled - called with preemption disabled
2826 * Returns with preemption disabled. Note: preempt_count must be 1
2828 void __sched schedule_preempt_disabled(void)
2830 sched_preempt_enable_no_resched();
2835 #ifdef CONFIG_PREEMPT
2837 * this is the entry point to schedule() from in-kernel preemption
2838 * off of preempt_enable. Kernel preemptions off return from interrupt
2839 * occur there and call schedule directly.
2841 asmlinkage __visible void __sched notrace preempt_schedule(void)
2844 * If there is a non-zero preempt_count or interrupts are disabled,
2845 * we do not want to preempt the current task. Just return..
2847 if (likely(!preemptible()))
2851 __preempt_count_add(PREEMPT_ACTIVE);
2853 __preempt_count_sub(PREEMPT_ACTIVE);
2856 * Check again in case we missed a preemption opportunity
2857 * between schedule and now.
2860 } while (need_resched());
2862 NOKPROBE_SYMBOL(preempt_schedule);
2863 EXPORT_SYMBOL(preempt_schedule);
2864 #endif /* CONFIG_PREEMPT */
2867 * this is the entry point to schedule() from kernel preemption
2868 * off of irq context.
2869 * Note, that this is called and return with irqs disabled. This will
2870 * protect us against recursive calling from irq.
2872 asmlinkage __visible void __sched preempt_schedule_irq(void)
2874 enum ctx_state prev_state;
2876 /* Catch callers which need to be fixed */
2877 BUG_ON(preempt_count() || !irqs_disabled());
2879 prev_state = exception_enter();
2882 __preempt_count_add(PREEMPT_ACTIVE);
2885 local_irq_disable();
2886 __preempt_count_sub(PREEMPT_ACTIVE);
2889 * Check again in case we missed a preemption opportunity
2890 * between schedule and now.
2893 } while (need_resched());
2895 exception_exit(prev_state);
2898 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2901 return try_to_wake_up(curr->private, mode, wake_flags);
2903 EXPORT_SYMBOL(default_wake_function);
2905 #ifdef CONFIG_RT_MUTEXES
2908 * rt_mutex_setprio - set the current priority of a task
2910 * @prio: prio value (kernel-internal form)
2912 * This function changes the 'effective' priority of a task. It does
2913 * not touch ->normal_prio like __setscheduler().
2915 * Used by the rt_mutex code to implement priority inheritance
2916 * logic. Call site only calls if the priority of the task changed.
2918 void rt_mutex_setprio(struct task_struct *p, int prio)
2920 int oldprio, on_rq, running, enqueue_flag = 0;
2922 const struct sched_class *prev_class;
2924 BUG_ON(prio > MAX_PRIO);
2926 rq = __task_rq_lock(p);
2929 * Idle task boosting is a nono in general. There is one
2930 * exception, when PREEMPT_RT and NOHZ is active:
2932 * The idle task calls get_next_timer_interrupt() and holds
2933 * the timer wheel base->lock on the CPU and another CPU wants
2934 * to access the timer (probably to cancel it). We can safely
2935 * ignore the boosting request, as the idle CPU runs this code
2936 * with interrupts disabled and will complete the lock
2937 * protected section without being interrupted. So there is no
2938 * real need to boost.
2940 if (unlikely(p == rq->idle)) {
2941 WARN_ON(p != rq->curr);
2942 WARN_ON(p->pi_blocked_on);
2946 trace_sched_pi_setprio(p, prio);
2947 p->pi_top_task = rt_mutex_get_top_task(p);
2949 prev_class = p->sched_class;
2951 running = task_current(rq, p);
2953 dequeue_task(rq, p, 0);
2955 p->sched_class->put_prev_task(rq, p);
2958 * Boosting condition are:
2959 * 1. -rt task is running and holds mutex A
2960 * --> -dl task blocks on mutex A
2962 * 2. -dl task is running and holds mutex A
2963 * --> -dl task blocks on mutex A and could preempt the
2966 if (dl_prio(prio)) {
2967 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2968 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2969 p->dl.dl_boosted = 1;
2970 p->dl.dl_throttled = 0;
2971 enqueue_flag = ENQUEUE_REPLENISH;
2973 p->dl.dl_boosted = 0;
2974 p->sched_class = &dl_sched_class;
2975 } else if (rt_prio(prio)) {
2976 if (dl_prio(oldprio))
2977 p->dl.dl_boosted = 0;
2979 enqueue_flag = ENQUEUE_HEAD;
2980 p->sched_class = &rt_sched_class;
2982 if (dl_prio(oldprio))
2983 p->dl.dl_boosted = 0;
2984 p->sched_class = &fair_sched_class;
2990 p->sched_class->set_curr_task(rq);
2992 enqueue_task(rq, p, enqueue_flag);
2994 check_class_changed(rq, p, prev_class, oldprio);
2996 __task_rq_unlock(rq);
3000 void set_user_nice(struct task_struct *p, long nice)
3002 int old_prio, delta, on_rq;
3003 unsigned long flags;
3006 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3009 * We have to be careful, if called from sys_setpriority(),
3010 * the task might be in the middle of scheduling on another CPU.
3012 rq = task_rq_lock(p, &flags);
3014 * The RT priorities are set via sched_setscheduler(), but we still
3015 * allow the 'normal' nice value to be set - but as expected
3016 * it wont have any effect on scheduling until the task is
3017 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3019 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3020 p->static_prio = NICE_TO_PRIO(nice);
3025 dequeue_task(rq, p, 0);
3027 p->static_prio = NICE_TO_PRIO(nice);
3030 p->prio = effective_prio(p);
3031 delta = p->prio - old_prio;
3034 enqueue_task(rq, p, 0);
3036 * If the task increased its priority or is running and
3037 * lowered its priority, then reschedule its CPU:
3039 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3040 resched_task(rq->curr);
3043 task_rq_unlock(rq, p, &flags);
3045 EXPORT_SYMBOL(set_user_nice);
3048 * can_nice - check if a task can reduce its nice value
3052 int can_nice(const struct task_struct *p, const int nice)
3054 /* convert nice value [19,-20] to rlimit style value [1,40] */
3055 int nice_rlim = 20 - nice;
3057 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3058 capable(CAP_SYS_NICE));
3061 #ifdef __ARCH_WANT_SYS_NICE
3064 * sys_nice - change the priority of the current process.
3065 * @increment: priority increment
3067 * sys_setpriority is a more generic, but much slower function that
3068 * does similar things.
3070 SYSCALL_DEFINE1(nice, int, increment)
3075 * Setpriority might change our priority at the same moment.
3076 * We don't have to worry. Conceptually one call occurs first
3077 * and we have a single winner.
3079 if (increment < -40)
3084 nice = task_nice(current) + increment;
3085 if (nice < MIN_NICE)
3087 if (nice > MAX_NICE)
3090 if (increment < 0 && !can_nice(current, nice))
3093 retval = security_task_setnice(current, nice);
3097 set_user_nice(current, nice);
3104 * task_prio - return the priority value of a given task.
3105 * @p: the task in question.
3107 * Return: The priority value as seen by users in /proc.
3108 * RT tasks are offset by -200. Normal tasks are centered
3109 * around 0, value goes from -16 to +15.
3111 int task_prio(const struct task_struct *p)
3113 return p->prio - MAX_RT_PRIO;
3117 * idle_cpu - is a given cpu idle currently?
3118 * @cpu: the processor in question.
3120 * Return: 1 if the CPU is currently idle. 0 otherwise.
3122 int idle_cpu(int cpu)
3124 struct rq *rq = cpu_rq(cpu);
3126 if (rq->curr != rq->idle)
3133 if (!llist_empty(&rq->wake_list))
3141 * idle_task - return the idle task for a given cpu.
3142 * @cpu: the processor in question.
3144 * Return: The idle task for the cpu @cpu.
3146 struct task_struct *idle_task(int cpu)
3148 return cpu_rq(cpu)->idle;
3152 * find_process_by_pid - find a process with a matching PID value.
3153 * @pid: the pid in question.
3155 * The task of @pid, if found. %NULL otherwise.
3157 static struct task_struct *find_process_by_pid(pid_t pid)
3159 return pid ? find_task_by_vpid(pid) : current;
3163 * This function initializes the sched_dl_entity of a newly becoming
3164 * SCHED_DEADLINE task.
3166 * Only the static values are considered here, the actual runtime and the
3167 * absolute deadline will be properly calculated when the task is enqueued
3168 * for the first time with its new policy.
3171 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3173 struct sched_dl_entity *dl_se = &p->dl;
3175 init_dl_task_timer(dl_se);
3176 dl_se->dl_runtime = attr->sched_runtime;
3177 dl_se->dl_deadline = attr->sched_deadline;
3178 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3179 dl_se->flags = attr->sched_flags;
3180 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3181 dl_se->dl_throttled = 0;
3183 dl_se->dl_yielded = 0;
3186 static void __setscheduler_params(struct task_struct *p,
3187 const struct sched_attr *attr)
3189 int policy = attr->sched_policy;
3191 if (policy == -1) /* setparam */
3196 if (dl_policy(policy))
3197 __setparam_dl(p, attr);
3198 else if (fair_policy(policy))
3199 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3202 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3203 * !rt_policy. Always setting this ensures that things like
3204 * getparam()/getattr() don't report silly values for !rt tasks.
3206 p->rt_priority = attr->sched_priority;
3207 p->normal_prio = normal_prio(p);
3211 /* Actually do priority change: must hold pi & rq lock. */
3212 static void __setscheduler(struct rq *rq, struct task_struct *p,
3213 const struct sched_attr *attr)
3215 __setscheduler_params(p, attr);
3218 * If we get here, there was no pi waiters boosting the
3219 * task. It is safe to use the normal prio.
3221 p->prio = normal_prio(p);
3223 if (dl_prio(p->prio))
3224 p->sched_class = &dl_sched_class;
3225 else if (rt_prio(p->prio))
3226 p->sched_class = &rt_sched_class;
3228 p->sched_class = &fair_sched_class;
3232 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3234 struct sched_dl_entity *dl_se = &p->dl;
3236 attr->sched_priority = p->rt_priority;
3237 attr->sched_runtime = dl_se->dl_runtime;
3238 attr->sched_deadline = dl_se->dl_deadline;
3239 attr->sched_period = dl_se->dl_period;
3240 attr->sched_flags = dl_se->flags;
3244 * This function validates the new parameters of a -deadline task.
3245 * We ask for the deadline not being zero, and greater or equal
3246 * than the runtime, as well as the period of being zero or
3247 * greater than deadline. Furthermore, we have to be sure that
3248 * user parameters are above the internal resolution (1us); we
3249 * check sched_runtime only since it is always the smaller one.
3252 __checkparam_dl(const struct sched_attr *attr)
3254 return attr && attr->sched_deadline != 0 &&
3255 (attr->sched_period == 0 ||
3256 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3257 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3258 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3262 * check the target process has a UID that matches the current process's
3264 static bool check_same_owner(struct task_struct *p)
3266 const struct cred *cred = current_cred(), *pcred;
3270 pcred = __task_cred(p);
3271 match = (uid_eq(cred->euid, pcred->euid) ||
3272 uid_eq(cred->euid, pcred->uid));
3277 static int __sched_setscheduler(struct task_struct *p,
3278 const struct sched_attr *attr,
3281 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3282 MAX_RT_PRIO - 1 - attr->sched_priority;
3283 int retval, oldprio, oldpolicy = -1, on_rq, running;
3284 int policy = attr->sched_policy;
3285 unsigned long flags;
3286 const struct sched_class *prev_class;
3290 /* may grab non-irq protected spin_locks */
3291 BUG_ON(in_interrupt());
3293 /* double check policy once rq lock held */
3295 reset_on_fork = p->sched_reset_on_fork;
3296 policy = oldpolicy = p->policy;
3298 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3300 if (policy != SCHED_DEADLINE &&
3301 policy != SCHED_FIFO && policy != SCHED_RR &&
3302 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3303 policy != SCHED_IDLE)
3307 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3311 * Valid priorities for SCHED_FIFO and SCHED_RR are
3312 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3313 * SCHED_BATCH and SCHED_IDLE is 0.
3315 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3316 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3318 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3319 (rt_policy(policy) != (attr->sched_priority != 0)))
3323 * Allow unprivileged RT tasks to decrease priority:
3325 if (user && !capable(CAP_SYS_NICE)) {
3326 if (fair_policy(policy)) {
3327 if (attr->sched_nice < task_nice(p) &&
3328 !can_nice(p, attr->sched_nice))
3332 if (rt_policy(policy)) {
3333 unsigned long rlim_rtprio =
3334 task_rlimit(p, RLIMIT_RTPRIO);
3336 /* can't set/change the rt policy */
3337 if (policy != p->policy && !rlim_rtprio)
3340 /* can't increase priority */
3341 if (attr->sched_priority > p->rt_priority &&
3342 attr->sched_priority > rlim_rtprio)
3347 * Can't set/change SCHED_DEADLINE policy at all for now
3348 * (safest behavior); in the future we would like to allow
3349 * unprivileged DL tasks to increase their relative deadline
3350 * or reduce their runtime (both ways reducing utilization)
3352 if (dl_policy(policy))
3356 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3357 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3359 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3360 if (!can_nice(p, task_nice(p)))
3364 /* can't change other user's priorities */
3365 if (!check_same_owner(p))
3368 /* Normal users shall not reset the sched_reset_on_fork flag */
3369 if (p->sched_reset_on_fork && !reset_on_fork)
3374 retval = security_task_setscheduler(p);
3380 * make sure no PI-waiters arrive (or leave) while we are
3381 * changing the priority of the task:
3383 * To be able to change p->policy safely, the appropriate
3384 * runqueue lock must be held.
3386 rq = task_rq_lock(p, &flags);
3389 * Changing the policy of the stop threads its a very bad idea
3391 if (p == rq->stop) {
3392 task_rq_unlock(rq, p, &flags);
3397 * If not changing anything there's no need to proceed further,
3398 * but store a possible modification of reset_on_fork.
3400 if (unlikely(policy == p->policy)) {
3401 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3403 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3405 if (dl_policy(policy))
3408 p->sched_reset_on_fork = reset_on_fork;
3409 task_rq_unlock(rq, p, &flags);
3415 #ifdef CONFIG_RT_GROUP_SCHED
3417 * Do not allow realtime tasks into groups that have no runtime
3420 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3421 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3422 !task_group_is_autogroup(task_group(p))) {
3423 task_rq_unlock(rq, p, &flags);
3428 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3429 cpumask_t *span = rq->rd->span;
3432 * Don't allow tasks with an affinity mask smaller than
3433 * the entire root_domain to become SCHED_DEADLINE. We
3434 * will also fail if there's no bandwidth available.
3436 if (!cpumask_subset(span, &p->cpus_allowed) ||
3437 rq->rd->dl_bw.bw == 0) {
3438 task_rq_unlock(rq, p, &flags);
3445 /* recheck policy now with rq lock held */
3446 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3447 policy = oldpolicy = -1;
3448 task_rq_unlock(rq, p, &flags);
3453 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3454 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3457 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3458 task_rq_unlock(rq, p, &flags);
3462 p->sched_reset_on_fork = reset_on_fork;
3466 * Special case for priority boosted tasks.
3468 * If the new priority is lower or equal (user space view)
3469 * than the current (boosted) priority, we just store the new
3470 * normal parameters and do not touch the scheduler class and
3471 * the runqueue. This will be done when the task deboost
3474 if (rt_mutex_check_prio(p, newprio)) {
3475 __setscheduler_params(p, attr);
3476 task_rq_unlock(rq, p, &flags);
3481 running = task_current(rq, p);
3483 dequeue_task(rq, p, 0);
3485 p->sched_class->put_prev_task(rq, p);
3487 prev_class = p->sched_class;
3488 __setscheduler(rq, p, attr);
3491 p->sched_class->set_curr_task(rq);
3494 * We enqueue to tail when the priority of a task is
3495 * increased (user space view).
3497 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3500 check_class_changed(rq, p, prev_class, oldprio);
3501 task_rq_unlock(rq, p, &flags);
3503 rt_mutex_adjust_pi(p);
3508 static int _sched_setscheduler(struct task_struct *p, int policy,
3509 const struct sched_param *param, bool check)
3511 struct sched_attr attr = {
3512 .sched_policy = policy,
3513 .sched_priority = param->sched_priority,
3514 .sched_nice = PRIO_TO_NICE(p->static_prio),
3518 * Fixup the legacy SCHED_RESET_ON_FORK hack
3520 if (policy & SCHED_RESET_ON_FORK) {
3521 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3522 policy &= ~SCHED_RESET_ON_FORK;
3523 attr.sched_policy = policy;
3526 return __sched_setscheduler(p, &attr, check);
3529 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3530 * @p: the task in question.
3531 * @policy: new policy.
3532 * @param: structure containing the new RT priority.
3534 * Return: 0 on success. An error code otherwise.
3536 * NOTE that the task may be already dead.
3538 int sched_setscheduler(struct task_struct *p, int policy,
3539 const struct sched_param *param)
3541 return _sched_setscheduler(p, policy, param, true);
3543 EXPORT_SYMBOL_GPL(sched_setscheduler);
3545 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3547 return __sched_setscheduler(p, attr, true);
3549 EXPORT_SYMBOL_GPL(sched_setattr);
3552 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3553 * @p: the task in question.
3554 * @policy: new policy.
3555 * @param: structure containing the new RT priority.
3557 * Just like sched_setscheduler, only don't bother checking if the
3558 * current context has permission. For example, this is needed in
3559 * stop_machine(): we create temporary high priority worker threads,
3560 * but our caller might not have that capability.
3562 * Return: 0 on success. An error code otherwise.
3564 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3565 const struct sched_param *param)
3567 return _sched_setscheduler(p, policy, param, false);
3571 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3573 struct sched_param lparam;
3574 struct task_struct *p;
3577 if (!param || pid < 0)
3579 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3584 p = find_process_by_pid(pid);
3586 retval = sched_setscheduler(p, policy, &lparam);
3593 * Mimics kernel/events/core.c perf_copy_attr().
3595 static int sched_copy_attr(struct sched_attr __user *uattr,
3596 struct sched_attr *attr)
3601 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3605 * zero the full structure, so that a short copy will be nice.
3607 memset(attr, 0, sizeof(*attr));
3609 ret = get_user(size, &uattr->size);
3613 if (size > PAGE_SIZE) /* silly large */
3616 if (!size) /* abi compat */
3617 size = SCHED_ATTR_SIZE_VER0;
3619 if (size < SCHED_ATTR_SIZE_VER0)
3623 * If we're handed a bigger struct than we know of,
3624 * ensure all the unknown bits are 0 - i.e. new
3625 * user-space does not rely on any kernel feature
3626 * extensions we dont know about yet.
3628 if (size > sizeof(*attr)) {
3629 unsigned char __user *addr;
3630 unsigned char __user *end;
3633 addr = (void __user *)uattr + sizeof(*attr);
3634 end = (void __user *)uattr + size;
3636 for (; addr < end; addr++) {
3637 ret = get_user(val, addr);
3643 size = sizeof(*attr);
3646 ret = copy_from_user(attr, uattr, size);
3651 * XXX: do we want to be lenient like existing syscalls; or do we want
3652 * to be strict and return an error on out-of-bounds values?
3654 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3660 put_user(sizeof(*attr), &uattr->size);
3666 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3667 * @pid: the pid in question.
3668 * @policy: new policy.
3669 * @param: structure containing the new RT priority.
3671 * Return: 0 on success. An error code otherwise.
3673 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3674 struct sched_param __user *, param)
3676 /* negative values for policy are not valid */
3680 return do_sched_setscheduler(pid, policy, param);
3684 * sys_sched_setparam - set/change the RT priority of a thread
3685 * @pid: the pid in question.
3686 * @param: structure containing the new RT priority.
3688 * Return: 0 on success. An error code otherwise.
3690 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3692 return do_sched_setscheduler(pid, -1, param);
3696 * sys_sched_setattr - same as above, but with extended sched_attr
3697 * @pid: the pid in question.
3698 * @uattr: structure containing the extended parameters.
3699 * @flags: for future extension.
3701 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3702 unsigned int, flags)
3704 struct sched_attr attr;
3705 struct task_struct *p;
3708 if (!uattr || pid < 0 || flags)
3711 if (sched_copy_attr(uattr, &attr))
3716 p = find_process_by_pid(pid);
3718 retval = sched_setattr(p, &attr);
3725 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3726 * @pid: the pid in question.
3728 * Return: On success, the policy of the thread. Otherwise, a negative error
3731 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3733 struct task_struct *p;
3741 p = find_process_by_pid(pid);
3743 retval = security_task_getscheduler(p);
3746 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3753 * sys_sched_getparam - get the RT priority of a thread
3754 * @pid: the pid in question.
3755 * @param: structure containing the RT priority.
3757 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3760 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3762 struct sched_param lp;
3763 struct task_struct *p;
3766 if (!param || pid < 0)
3770 p = find_process_by_pid(pid);
3775 retval = security_task_getscheduler(p);
3779 if (task_has_dl_policy(p)) {
3783 lp.sched_priority = p->rt_priority;
3787 * This one might sleep, we cannot do it with a spinlock held ...
3789 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3798 static int sched_read_attr(struct sched_attr __user *uattr,
3799 struct sched_attr *attr,
3804 if (!access_ok(VERIFY_WRITE, uattr, usize))
3808 * If we're handed a smaller struct than we know of,
3809 * ensure all the unknown bits are 0 - i.e. old
3810 * user-space does not get uncomplete information.
3812 if (usize < sizeof(*attr)) {
3813 unsigned char *addr;
3816 addr = (void *)attr + usize;
3817 end = (void *)attr + sizeof(*attr);
3819 for (; addr < end; addr++) {
3827 ret = copy_to_user(uattr, attr, attr->size);
3840 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3841 * @pid: the pid in question.
3842 * @uattr: structure containing the extended parameters.
3843 * @size: sizeof(attr) for fwd/bwd comp.
3844 * @flags: for future extension.
3846 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3847 unsigned int, size, unsigned int, flags)
3849 struct sched_attr attr = {
3850 .size = sizeof(struct sched_attr),
3852 struct task_struct *p;
3855 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3856 size < SCHED_ATTR_SIZE_VER0 || flags)
3860 p = find_process_by_pid(pid);
3865 retval = security_task_getscheduler(p);
3869 attr.sched_policy = p->policy;
3870 if (p->sched_reset_on_fork)
3871 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3872 if (task_has_dl_policy(p))
3873 __getparam_dl(p, &attr);
3874 else if (task_has_rt_policy(p))
3875 attr.sched_priority = p->rt_priority;
3877 attr.sched_nice = task_nice(p);
3881 retval = sched_read_attr(uattr, &attr, size);
3889 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3891 cpumask_var_t cpus_allowed, new_mask;
3892 struct task_struct *p;
3897 p = find_process_by_pid(pid);
3903 /* Prevent p going away */
3907 if (p->flags & PF_NO_SETAFFINITY) {
3911 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3915 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3917 goto out_free_cpus_allowed;
3920 if (!check_same_owner(p)) {
3922 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3929 retval = security_task_setscheduler(p);
3934 cpuset_cpus_allowed(p, cpus_allowed);
3935 cpumask_and(new_mask, in_mask, cpus_allowed);
3938 * Since bandwidth control happens on root_domain basis,
3939 * if admission test is enabled, we only admit -deadline
3940 * tasks allowed to run on all the CPUs in the task's
3944 if (task_has_dl_policy(p)) {
3945 const struct cpumask *span = task_rq(p)->rd->span;
3947 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3954 retval = set_cpus_allowed_ptr(p, new_mask);
3957 cpuset_cpus_allowed(p, cpus_allowed);
3958 if (!cpumask_subset(new_mask, cpus_allowed)) {
3960 * We must have raced with a concurrent cpuset
3961 * update. Just reset the cpus_allowed to the
3962 * cpuset's cpus_allowed
3964 cpumask_copy(new_mask, cpus_allowed);
3969 free_cpumask_var(new_mask);
3970 out_free_cpus_allowed:
3971 free_cpumask_var(cpus_allowed);
3977 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3978 struct cpumask *new_mask)
3980 if (len < cpumask_size())
3981 cpumask_clear(new_mask);
3982 else if (len > cpumask_size())
3983 len = cpumask_size();
3985 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3989 * sys_sched_setaffinity - set the cpu affinity of a process
3990 * @pid: pid of the process
3991 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3992 * @user_mask_ptr: user-space pointer to the new cpu mask
3994 * Return: 0 on success. An error code otherwise.
3996 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3997 unsigned long __user *, user_mask_ptr)
3999 cpumask_var_t new_mask;
4002 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4005 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4007 retval = sched_setaffinity(pid, new_mask);
4008 free_cpumask_var(new_mask);
4012 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4014 struct task_struct *p;
4015 unsigned long flags;
4021 p = find_process_by_pid(pid);
4025 retval = security_task_getscheduler(p);
4029 raw_spin_lock_irqsave(&p->pi_lock, flags);
4030 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4031 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4040 * sys_sched_getaffinity - get the cpu affinity of a process
4041 * @pid: pid of the process
4042 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4043 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4045 * Return: 0 on success. An error code otherwise.
4047 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4048 unsigned long __user *, user_mask_ptr)
4053 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4055 if (len & (sizeof(unsigned long)-1))
4058 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4061 ret = sched_getaffinity(pid, mask);
4063 size_t retlen = min_t(size_t, len, cpumask_size());
4065 if (copy_to_user(user_mask_ptr, mask, retlen))
4070 free_cpumask_var(mask);
4076 * sys_sched_yield - yield the current processor to other threads.
4078 * This function yields the current CPU to other tasks. If there are no
4079 * other threads running on this CPU then this function will return.
4083 SYSCALL_DEFINE0(sched_yield)
4085 struct rq *rq = this_rq_lock();
4087 schedstat_inc(rq, yld_count);
4088 current->sched_class->yield_task(rq);
4091 * Since we are going to call schedule() anyway, there's
4092 * no need to preempt or enable interrupts:
4094 __release(rq->lock);
4095 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4096 do_raw_spin_unlock(&rq->lock);
4097 sched_preempt_enable_no_resched();
4104 static void __cond_resched(void)
4106 __preempt_count_add(PREEMPT_ACTIVE);
4108 __preempt_count_sub(PREEMPT_ACTIVE);
4111 int __sched _cond_resched(void)
4114 if (should_resched()) {
4120 EXPORT_SYMBOL(_cond_resched);
4123 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4124 * call schedule, and on return reacquire the lock.
4126 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4127 * operations here to prevent schedule() from being called twice (once via
4128 * spin_unlock(), once by hand).
4130 int __cond_resched_lock(spinlock_t *lock)
4132 bool need_rcu_resched = rcu_should_resched();
4133 int resched = should_resched();
4136 lockdep_assert_held(lock);
4138 if (spin_needbreak(lock) || resched || need_rcu_resched) {
4142 else if (unlikely(need_rcu_resched))
4151 EXPORT_SYMBOL(__cond_resched_lock);
4153 int __sched __cond_resched_softirq(void)
4155 BUG_ON(!in_softirq());
4157 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4158 if (should_resched()) {
4166 EXPORT_SYMBOL(__cond_resched_softirq);
4169 * yield - yield the current processor to other threads.
4171 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4173 * The scheduler is at all times free to pick the calling task as the most
4174 * eligible task to run, if removing the yield() call from your code breaks
4175 * it, its already broken.
4177 * Typical broken usage is:
4182 * where one assumes that yield() will let 'the other' process run that will
4183 * make event true. If the current task is a SCHED_FIFO task that will never
4184 * happen. Never use yield() as a progress guarantee!!
4186 * If you want to use yield() to wait for something, use wait_event().
4187 * If you want to use yield() to be 'nice' for others, use cond_resched().
4188 * If you still want to use yield(), do not!
4190 void __sched yield(void)
4192 set_current_state(TASK_RUNNING);
4195 EXPORT_SYMBOL(yield);
4198 * yield_to - yield the current processor to another thread in
4199 * your thread group, or accelerate that thread toward the
4200 * processor it's on.
4202 * @preempt: whether task preemption is allowed or not
4204 * It's the caller's job to ensure that the target task struct
4205 * can't go away on us before we can do any checks.
4208 * true (>0) if we indeed boosted the target task.
4209 * false (0) if we failed to boost the target.
4210 * -ESRCH if there's no task to yield to.
4212 bool __sched yield_to(struct task_struct *p, bool preempt)
4214 struct task_struct *curr = current;
4215 struct rq *rq, *p_rq;
4216 unsigned long flags;
4219 local_irq_save(flags);
4225 * If we're the only runnable task on the rq and target rq also
4226 * has only one task, there's absolutely no point in yielding.
4228 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4233 double_rq_lock(rq, p_rq);
4234 if (task_rq(p) != p_rq) {
4235 double_rq_unlock(rq, p_rq);
4239 if (!curr->sched_class->yield_to_task)
4242 if (curr->sched_class != p->sched_class)
4245 if (task_running(p_rq, p) || p->state)
4248 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4250 schedstat_inc(rq, yld_count);
4252 * Make p's CPU reschedule; pick_next_entity takes care of
4255 if (preempt && rq != p_rq)
4256 resched_task(p_rq->curr);
4260 double_rq_unlock(rq, p_rq);
4262 local_irq_restore(flags);
4269 EXPORT_SYMBOL_GPL(yield_to);
4272 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4273 * that process accounting knows that this is a task in IO wait state.
4275 void __sched io_schedule(void)
4277 struct rq *rq = raw_rq();
4279 delayacct_blkio_start();
4280 atomic_inc(&rq->nr_iowait);
4281 blk_flush_plug(current);
4282 current->in_iowait = 1;
4284 current->in_iowait = 0;
4285 atomic_dec(&rq->nr_iowait);
4286 delayacct_blkio_end();
4288 EXPORT_SYMBOL(io_schedule);
4290 long __sched io_schedule_timeout(long timeout)
4292 struct rq *rq = raw_rq();
4295 delayacct_blkio_start();
4296 atomic_inc(&rq->nr_iowait);
4297 blk_flush_plug(current);
4298 current->in_iowait = 1;
4299 ret = schedule_timeout(timeout);
4300 current->in_iowait = 0;
4301 atomic_dec(&rq->nr_iowait);
4302 delayacct_blkio_end();
4307 * sys_sched_get_priority_max - return maximum RT priority.
4308 * @policy: scheduling class.
4310 * Return: On success, this syscall returns the maximum
4311 * rt_priority that can be used by a given scheduling class.
4312 * On failure, a negative error code is returned.
4314 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4321 ret = MAX_USER_RT_PRIO-1;
4323 case SCHED_DEADLINE:
4334 * sys_sched_get_priority_min - return minimum RT priority.
4335 * @policy: scheduling class.
4337 * Return: On success, this syscall returns the minimum
4338 * rt_priority that can be used by a given scheduling class.
4339 * On failure, a negative error code is returned.
4341 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4350 case SCHED_DEADLINE:
4360 * sys_sched_rr_get_interval - return the default timeslice of a process.
4361 * @pid: pid of the process.
4362 * @interval: userspace pointer to the timeslice value.
4364 * this syscall writes the default timeslice value of a given process
4365 * into the user-space timespec buffer. A value of '0' means infinity.
4367 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4370 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4371 struct timespec __user *, interval)
4373 struct task_struct *p;
4374 unsigned int time_slice;
4375 unsigned long flags;
4385 p = find_process_by_pid(pid);
4389 retval = security_task_getscheduler(p);
4393 rq = task_rq_lock(p, &flags);
4395 if (p->sched_class->get_rr_interval)
4396 time_slice = p->sched_class->get_rr_interval(rq, p);
4397 task_rq_unlock(rq, p, &flags);
4400 jiffies_to_timespec(time_slice, &t);
4401 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4409 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4411 void sched_show_task(struct task_struct *p)
4413 unsigned long free = 0;
4417 state = p->state ? __ffs(p->state) + 1 : 0;
4418 printk(KERN_INFO "%-15.15s %c", p->comm,
4419 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4420 #if BITS_PER_LONG == 32
4421 if (state == TASK_RUNNING)
4422 printk(KERN_CONT " running ");
4424 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4426 if (state == TASK_RUNNING)
4427 printk(KERN_CONT " running task ");
4429 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4431 #ifdef CONFIG_DEBUG_STACK_USAGE
4432 free = stack_not_used(p);
4435 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4437 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4438 task_pid_nr(p), ppid,
4439 (unsigned long)task_thread_info(p)->flags);
4441 print_worker_info(KERN_INFO, p);
4442 show_stack(p, NULL);
4445 void show_state_filter(unsigned long state_filter)
4447 struct task_struct *g, *p;
4449 #if BITS_PER_LONG == 32
4451 " task PC stack pid father\n");
4454 " task PC stack pid father\n");
4457 do_each_thread(g, p) {
4459 * reset the NMI-timeout, listing all files on a slow
4460 * console might take a lot of time:
4462 touch_nmi_watchdog();
4463 if (!state_filter || (p->state & state_filter))
4465 } while_each_thread(g, p);
4467 touch_all_softlockup_watchdogs();
4469 #ifdef CONFIG_SCHED_DEBUG
4470 sysrq_sched_debug_show();
4474 * Only show locks if all tasks are dumped:
4477 debug_show_all_locks();
4480 void init_idle_bootup_task(struct task_struct *idle)
4482 idle->sched_class = &idle_sched_class;
4486 * init_idle - set up an idle thread for a given CPU
4487 * @idle: task in question
4488 * @cpu: cpu the idle task belongs to
4490 * NOTE: this function does not set the idle thread's NEED_RESCHED
4491 * flag, to make booting more robust.
4493 void init_idle(struct task_struct *idle, int cpu)
4495 struct rq *rq = cpu_rq(cpu);
4496 unsigned long flags;
4498 raw_spin_lock_irqsave(&rq->lock, flags);
4500 __sched_fork(0, idle);
4501 idle->state = TASK_RUNNING;
4502 idle->se.exec_start = sched_clock();
4504 do_set_cpus_allowed(idle, cpumask_of(cpu));
4506 * We're having a chicken and egg problem, even though we are
4507 * holding rq->lock, the cpu isn't yet set to this cpu so the
4508 * lockdep check in task_group() will fail.
4510 * Similar case to sched_fork(). / Alternatively we could
4511 * use task_rq_lock() here and obtain the other rq->lock.
4516 __set_task_cpu(idle, cpu);
4519 rq->curr = rq->idle = idle;
4521 #if defined(CONFIG_SMP)
4524 raw_spin_unlock_irqrestore(&rq->lock, flags);
4526 /* Set the preempt count _outside_ the spinlocks! */
4527 init_idle_preempt_count(idle, cpu);
4530 * The idle tasks have their own, simple scheduling class:
4532 idle->sched_class = &idle_sched_class;
4533 ftrace_graph_init_idle_task(idle, cpu);
4534 vtime_init_idle(idle, cpu);
4535 #if defined(CONFIG_SMP)
4536 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4541 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4543 if (p->sched_class && p->sched_class->set_cpus_allowed)
4544 p->sched_class->set_cpus_allowed(p, new_mask);
4546 cpumask_copy(&p->cpus_allowed, new_mask);
4547 p->nr_cpus_allowed = cpumask_weight(new_mask);
4551 * This is how migration works:
4553 * 1) we invoke migration_cpu_stop() on the target CPU using
4555 * 2) stopper starts to run (implicitly forcing the migrated thread
4557 * 3) it checks whether the migrated task is still in the wrong runqueue.
4558 * 4) if it's in the wrong runqueue then the migration thread removes
4559 * it and puts it into the right queue.
4560 * 5) stopper completes and stop_one_cpu() returns and the migration
4565 * Change a given task's CPU affinity. Migrate the thread to a
4566 * proper CPU and schedule it away if the CPU it's executing on
4567 * is removed from the allowed bitmask.
4569 * NOTE: the caller must have a valid reference to the task, the
4570 * task must not exit() & deallocate itself prematurely. The
4571 * call is not atomic; no spinlocks may be held.
4573 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4575 unsigned long flags;
4577 unsigned int dest_cpu;
4580 rq = task_rq_lock(p, &flags);
4582 if (cpumask_equal(&p->cpus_allowed, new_mask))
4585 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4590 do_set_cpus_allowed(p, new_mask);
4592 /* Can the task run on the task's current CPU? If so, we're done */
4593 if (cpumask_test_cpu(task_cpu(p), new_mask))
4596 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4598 struct migration_arg arg = { p, dest_cpu };
4599 /* Need help from migration thread: drop lock and wait. */
4600 task_rq_unlock(rq, p, &flags);
4601 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4602 tlb_migrate_finish(p->mm);
4606 task_rq_unlock(rq, p, &flags);
4610 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4613 * Move (not current) task off this cpu, onto dest cpu. We're doing
4614 * this because either it can't run here any more (set_cpus_allowed()
4615 * away from this CPU, or CPU going down), or because we're
4616 * attempting to rebalance this task on exec (sched_exec).
4618 * So we race with normal scheduler movements, but that's OK, as long
4619 * as the task is no longer on this CPU.
4621 * Returns non-zero if task was successfully migrated.
4623 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4625 struct rq *rq_dest, *rq_src;
4628 if (unlikely(!cpu_active(dest_cpu)))
4631 rq_src = cpu_rq(src_cpu);
4632 rq_dest = cpu_rq(dest_cpu);
4634 raw_spin_lock(&p->pi_lock);
4635 double_rq_lock(rq_src, rq_dest);
4636 /* Already moved. */
4637 if (task_cpu(p) != src_cpu)
4639 /* Affinity changed (again). */
4640 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4644 * If we're not on a rq, the next wake-up will ensure we're
4648 dequeue_task(rq_src, p, 0);
4649 set_task_cpu(p, dest_cpu);
4650 enqueue_task(rq_dest, p, 0);
4651 check_preempt_curr(rq_dest, p, 0);
4656 double_rq_unlock(rq_src, rq_dest);
4657 raw_spin_unlock(&p->pi_lock);
4661 #ifdef CONFIG_NUMA_BALANCING
4662 /* Migrate current task p to target_cpu */
4663 int migrate_task_to(struct task_struct *p, int target_cpu)
4665 struct migration_arg arg = { p, target_cpu };
4666 int curr_cpu = task_cpu(p);
4668 if (curr_cpu == target_cpu)
4671 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4674 /* TODO: This is not properly updating schedstats */
4676 trace_sched_move_numa(p, curr_cpu, target_cpu);
4677 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4681 * Requeue a task on a given node and accurately track the number of NUMA
4682 * tasks on the runqueues
4684 void sched_setnuma(struct task_struct *p, int nid)
4687 unsigned long flags;
4688 bool on_rq, running;
4690 rq = task_rq_lock(p, &flags);
4692 running = task_current(rq, p);
4695 dequeue_task(rq, p, 0);
4697 p->sched_class->put_prev_task(rq, p);
4699 p->numa_preferred_nid = nid;
4702 p->sched_class->set_curr_task(rq);
4704 enqueue_task(rq, p, 0);
4705 task_rq_unlock(rq, p, &flags);
4710 * migration_cpu_stop - this will be executed by a highprio stopper thread
4711 * and performs thread migration by bumping thread off CPU then
4712 * 'pushing' onto another runqueue.
4714 static int migration_cpu_stop(void *data)
4716 struct migration_arg *arg = data;
4719 * The original target cpu might have gone down and we might
4720 * be on another cpu but it doesn't matter.
4722 local_irq_disable();
4723 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4728 #ifdef CONFIG_HOTPLUG_CPU
4731 * Ensures that the idle task is using init_mm right before its cpu goes
4734 void idle_task_exit(void)
4736 struct mm_struct *mm = current->active_mm;
4738 BUG_ON(cpu_online(smp_processor_id()));
4740 if (mm != &init_mm) {
4741 switch_mm(mm, &init_mm, current);
4742 finish_arch_post_lock_switch();
4748 * Since this CPU is going 'away' for a while, fold any nr_active delta
4749 * we might have. Assumes we're called after migrate_tasks() so that the
4750 * nr_active count is stable.
4752 * Also see the comment "Global load-average calculations".
4754 static void calc_load_migrate(struct rq *rq)
4756 long delta = calc_load_fold_active(rq);
4758 atomic_long_add(delta, &calc_load_tasks);
4761 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4765 static const struct sched_class fake_sched_class = {
4766 .put_prev_task = put_prev_task_fake,
4769 static struct task_struct fake_task = {
4771 * Avoid pull_{rt,dl}_task()
4773 .prio = MAX_PRIO + 1,
4774 .sched_class = &fake_sched_class,
4778 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4779 * try_to_wake_up()->select_task_rq().
4781 * Called with rq->lock held even though we'er in stop_machine() and
4782 * there's no concurrency possible, we hold the required locks anyway
4783 * because of lock validation efforts.
4785 static void migrate_tasks(unsigned int dead_cpu)
4787 struct rq *rq = cpu_rq(dead_cpu);
4788 struct task_struct *next, *stop = rq->stop;
4792 * Fudge the rq selection such that the below task selection loop
4793 * doesn't get stuck on the currently eligible stop task.
4795 * We're currently inside stop_machine() and the rq is either stuck
4796 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4797 * either way we should never end up calling schedule() until we're
4803 * put_prev_task() and pick_next_task() sched
4804 * class method both need to have an up-to-date
4805 * value of rq->clock[_task]
4807 update_rq_clock(rq);
4811 * There's this thread running, bail when that's the only
4814 if (rq->nr_running == 1)
4817 next = pick_next_task(rq, &fake_task);
4819 next->sched_class->put_prev_task(rq, next);
4821 /* Find suitable destination for @next, with force if needed. */
4822 dest_cpu = select_fallback_rq(dead_cpu, next);
4823 raw_spin_unlock(&rq->lock);
4825 __migrate_task(next, dead_cpu, dest_cpu);
4827 raw_spin_lock(&rq->lock);
4833 #endif /* CONFIG_HOTPLUG_CPU */
4835 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4837 static struct ctl_table sd_ctl_dir[] = {
4839 .procname = "sched_domain",
4845 static struct ctl_table sd_ctl_root[] = {
4847 .procname = "kernel",
4849 .child = sd_ctl_dir,
4854 static struct ctl_table *sd_alloc_ctl_entry(int n)
4856 struct ctl_table *entry =
4857 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4862 static void sd_free_ctl_entry(struct ctl_table **tablep)
4864 struct ctl_table *entry;
4867 * In the intermediate directories, both the child directory and
4868 * procname are dynamically allocated and could fail but the mode
4869 * will always be set. In the lowest directory the names are
4870 * static strings and all have proc handlers.
4872 for (entry = *tablep; entry->mode; entry++) {
4874 sd_free_ctl_entry(&entry->child);
4875 if (entry->proc_handler == NULL)
4876 kfree(entry->procname);
4883 static int min_load_idx = 0;
4884 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4887 set_table_entry(struct ctl_table *entry,
4888 const char *procname, void *data, int maxlen,
4889 umode_t mode, proc_handler *proc_handler,
4892 entry->procname = procname;
4894 entry->maxlen = maxlen;
4896 entry->proc_handler = proc_handler;
4899 entry->extra1 = &min_load_idx;
4900 entry->extra2 = &max_load_idx;
4904 static struct ctl_table *
4905 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4907 struct ctl_table *table = sd_alloc_ctl_entry(14);
4912 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4913 sizeof(long), 0644, proc_doulongvec_minmax, false);
4914 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4915 sizeof(long), 0644, proc_doulongvec_minmax, false);
4916 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4917 sizeof(int), 0644, proc_dointvec_minmax, true);
4918 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4919 sizeof(int), 0644, proc_dointvec_minmax, true);
4920 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4921 sizeof(int), 0644, proc_dointvec_minmax, true);
4922 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4923 sizeof(int), 0644, proc_dointvec_minmax, true);
4924 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4925 sizeof(int), 0644, proc_dointvec_minmax, true);
4926 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4927 sizeof(int), 0644, proc_dointvec_minmax, false);
4928 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4929 sizeof(int), 0644, proc_dointvec_minmax, false);
4930 set_table_entry(&table[9], "cache_nice_tries",
4931 &sd->cache_nice_tries,
4932 sizeof(int), 0644, proc_dointvec_minmax, false);
4933 set_table_entry(&table[10], "flags", &sd->flags,
4934 sizeof(int), 0644, proc_dointvec_minmax, false);
4935 set_table_entry(&table[11], "max_newidle_lb_cost",
4936 &sd->max_newidle_lb_cost,
4937 sizeof(long), 0644, proc_doulongvec_minmax, false);
4938 set_table_entry(&table[12], "name", sd->name,
4939 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4940 /* &table[13] is terminator */
4945 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4947 struct ctl_table *entry, *table;
4948 struct sched_domain *sd;
4949 int domain_num = 0, i;
4952 for_each_domain(cpu, sd)
4954 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4959 for_each_domain(cpu, sd) {
4960 snprintf(buf, 32, "domain%d", i);
4961 entry->procname = kstrdup(buf, GFP_KERNEL);
4963 entry->child = sd_alloc_ctl_domain_table(sd);
4970 static struct ctl_table_header *sd_sysctl_header;
4971 static void register_sched_domain_sysctl(void)
4973 int i, cpu_num = num_possible_cpus();
4974 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4977 WARN_ON(sd_ctl_dir[0].child);
4978 sd_ctl_dir[0].child = entry;
4983 for_each_possible_cpu(i) {
4984 snprintf(buf, 32, "cpu%d", i);
4985 entry->procname = kstrdup(buf, GFP_KERNEL);
4987 entry->child = sd_alloc_ctl_cpu_table(i);
4991 WARN_ON(sd_sysctl_header);
4992 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4995 /* may be called multiple times per register */
4996 static void unregister_sched_domain_sysctl(void)
4998 if (sd_sysctl_header)
4999 unregister_sysctl_table(sd_sysctl_header);
5000 sd_sysctl_header = NULL;
5001 if (sd_ctl_dir[0].child)
5002 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5005 static void register_sched_domain_sysctl(void)
5008 static void unregister_sched_domain_sysctl(void)
5013 static void set_rq_online(struct rq *rq)
5016 const struct sched_class *class;
5018 cpumask_set_cpu(rq->cpu, rq->rd->online);
5021 for_each_class(class) {
5022 if (class->rq_online)
5023 class->rq_online(rq);
5028 static void set_rq_offline(struct rq *rq)
5031 const struct sched_class *class;
5033 for_each_class(class) {
5034 if (class->rq_offline)
5035 class->rq_offline(rq);
5038 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5044 * migration_call - callback that gets triggered when a CPU is added.
5045 * Here we can start up the necessary migration thread for the new CPU.
5048 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5050 int cpu = (long)hcpu;
5051 unsigned long flags;
5052 struct rq *rq = cpu_rq(cpu);
5054 switch (action & ~CPU_TASKS_FROZEN) {
5056 case CPU_UP_PREPARE:
5057 rq->calc_load_update = calc_load_update;
5061 /* Update our root-domain */
5062 raw_spin_lock_irqsave(&rq->lock, flags);
5064 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5068 raw_spin_unlock_irqrestore(&rq->lock, flags);
5071 #ifdef CONFIG_HOTPLUG_CPU
5073 sched_ttwu_pending();
5074 /* Update our root-domain */
5075 raw_spin_lock_irqsave(&rq->lock, flags);
5077 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5081 BUG_ON(rq->nr_running != 1); /* the migration thread */
5082 raw_spin_unlock_irqrestore(&rq->lock, flags);
5086 calc_load_migrate(rq);
5091 update_max_interval();
5097 * Register at high priority so that task migration (migrate_all_tasks)
5098 * happens before everything else. This has to be lower priority than
5099 * the notifier in the perf_event subsystem, though.
5101 static struct notifier_block migration_notifier = {
5102 .notifier_call = migration_call,
5103 .priority = CPU_PRI_MIGRATION,
5106 static int sched_cpu_active(struct notifier_block *nfb,
5107 unsigned long action, void *hcpu)
5109 switch (action & ~CPU_TASKS_FROZEN) {
5111 case CPU_DOWN_FAILED:
5112 set_cpu_active((long)hcpu, true);
5119 static int sched_cpu_inactive(struct notifier_block *nfb,
5120 unsigned long action, void *hcpu)
5122 unsigned long flags;
5123 long cpu = (long)hcpu;
5125 switch (action & ~CPU_TASKS_FROZEN) {
5126 case CPU_DOWN_PREPARE:
5127 set_cpu_active(cpu, false);
5129 /* explicitly allow suspend */
5130 if (!(action & CPU_TASKS_FROZEN)) {
5131 struct dl_bw *dl_b = dl_bw_of(cpu);
5135 raw_spin_lock_irqsave(&dl_b->lock, flags);
5136 cpus = dl_bw_cpus(cpu);
5137 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5138 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5141 return notifier_from_errno(-EBUSY);
5149 static int __init migration_init(void)
5151 void *cpu = (void *)(long)smp_processor_id();
5154 /* Initialize migration for the boot CPU */
5155 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5156 BUG_ON(err == NOTIFY_BAD);
5157 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5158 register_cpu_notifier(&migration_notifier);
5160 /* Register cpu active notifiers */
5161 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5162 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5166 early_initcall(migration_init);
5171 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5173 #ifdef CONFIG_SCHED_DEBUG
5175 static __read_mostly int sched_debug_enabled;
5177 static int __init sched_debug_setup(char *str)
5179 sched_debug_enabled = 1;
5183 early_param("sched_debug", sched_debug_setup);
5185 static inline bool sched_debug(void)
5187 return sched_debug_enabled;
5190 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5191 struct cpumask *groupmask)
5193 struct sched_group *group = sd->groups;
5196 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5197 cpumask_clear(groupmask);
5199 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5201 if (!(sd->flags & SD_LOAD_BALANCE)) {
5202 printk("does not load-balance\n");
5204 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5209 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5211 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5212 printk(KERN_ERR "ERROR: domain->span does not contain "
5215 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5216 printk(KERN_ERR "ERROR: domain->groups does not contain"
5220 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5224 printk(KERN_ERR "ERROR: group is NULL\n");
5229 * Even though we initialize ->power to something semi-sane,
5230 * we leave power_orig unset. This allows us to detect if
5231 * domain iteration is still funny without causing /0 traps.
5233 if (!group->sgp->power_orig) {
5234 printk(KERN_CONT "\n");
5235 printk(KERN_ERR "ERROR: domain->cpu_power not "
5240 if (!cpumask_weight(sched_group_cpus(group))) {
5241 printk(KERN_CONT "\n");
5242 printk(KERN_ERR "ERROR: empty group\n");
5246 if (!(sd->flags & SD_OVERLAP) &&
5247 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5248 printk(KERN_CONT "\n");
5249 printk(KERN_ERR "ERROR: repeated CPUs\n");
5253 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5255 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5257 printk(KERN_CONT " %s", str);
5258 if (group->sgp->power != SCHED_POWER_SCALE) {
5259 printk(KERN_CONT " (cpu_power = %d)",
5263 group = group->next;
5264 } while (group != sd->groups);
5265 printk(KERN_CONT "\n");
5267 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5268 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5271 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5272 printk(KERN_ERR "ERROR: parent span is not a superset "
5273 "of domain->span\n");
5277 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5281 if (!sched_debug_enabled)
5285 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5289 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5292 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5300 #else /* !CONFIG_SCHED_DEBUG */
5301 # define sched_domain_debug(sd, cpu) do { } while (0)
5302 static inline bool sched_debug(void)
5306 #endif /* CONFIG_SCHED_DEBUG */
5308 static int sd_degenerate(struct sched_domain *sd)
5310 if (cpumask_weight(sched_domain_span(sd)) == 1)
5313 /* Following flags need at least 2 groups */
5314 if (sd->flags & (SD_LOAD_BALANCE |
5315 SD_BALANCE_NEWIDLE |
5319 SD_SHARE_PKG_RESOURCES |
5320 SD_SHARE_POWERDOMAIN)) {
5321 if (sd->groups != sd->groups->next)
5325 /* Following flags don't use groups */
5326 if (sd->flags & (SD_WAKE_AFFINE))
5333 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5335 unsigned long cflags = sd->flags, pflags = parent->flags;
5337 if (sd_degenerate(parent))
5340 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5343 /* Flags needing groups don't count if only 1 group in parent */
5344 if (parent->groups == parent->groups->next) {
5345 pflags &= ~(SD_LOAD_BALANCE |
5346 SD_BALANCE_NEWIDLE |
5350 SD_SHARE_PKG_RESOURCES |
5352 SD_SHARE_POWERDOMAIN);
5353 if (nr_node_ids == 1)
5354 pflags &= ~SD_SERIALIZE;
5356 if (~cflags & pflags)
5362 static void free_rootdomain(struct rcu_head *rcu)
5364 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5366 cpupri_cleanup(&rd->cpupri);
5367 cpudl_cleanup(&rd->cpudl);
5368 free_cpumask_var(rd->dlo_mask);
5369 free_cpumask_var(rd->rto_mask);
5370 free_cpumask_var(rd->online);
5371 free_cpumask_var(rd->span);
5375 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5377 struct root_domain *old_rd = NULL;
5378 unsigned long flags;
5380 raw_spin_lock_irqsave(&rq->lock, flags);
5385 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5388 cpumask_clear_cpu(rq->cpu, old_rd->span);
5391 * If we dont want to free the old_rd yet then
5392 * set old_rd to NULL to skip the freeing later
5395 if (!atomic_dec_and_test(&old_rd->refcount))
5399 atomic_inc(&rd->refcount);
5402 cpumask_set_cpu(rq->cpu, rd->span);
5403 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5406 raw_spin_unlock_irqrestore(&rq->lock, flags);
5409 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5412 static int init_rootdomain(struct root_domain *rd)
5414 memset(rd, 0, sizeof(*rd));
5416 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5418 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5420 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5422 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5425 init_dl_bw(&rd->dl_bw);
5426 if (cpudl_init(&rd->cpudl) != 0)
5429 if (cpupri_init(&rd->cpupri) != 0)
5434 free_cpumask_var(rd->rto_mask);
5436 free_cpumask_var(rd->dlo_mask);
5438 free_cpumask_var(rd->online);
5440 free_cpumask_var(rd->span);
5446 * By default the system creates a single root-domain with all cpus as
5447 * members (mimicking the global state we have today).
5449 struct root_domain def_root_domain;
5451 static void init_defrootdomain(void)
5453 init_rootdomain(&def_root_domain);
5455 atomic_set(&def_root_domain.refcount, 1);
5458 static struct root_domain *alloc_rootdomain(void)
5460 struct root_domain *rd;
5462 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5466 if (init_rootdomain(rd) != 0) {
5474 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5476 struct sched_group *tmp, *first;
5485 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5490 } while (sg != first);
5493 static void free_sched_domain(struct rcu_head *rcu)
5495 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5498 * If its an overlapping domain it has private groups, iterate and
5501 if (sd->flags & SD_OVERLAP) {
5502 free_sched_groups(sd->groups, 1);
5503 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5504 kfree(sd->groups->sgp);
5510 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5512 call_rcu(&sd->rcu, free_sched_domain);
5515 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5517 for (; sd; sd = sd->parent)
5518 destroy_sched_domain(sd, cpu);
5522 * Keep a special pointer to the highest sched_domain that has
5523 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5524 * allows us to avoid some pointer chasing select_idle_sibling().
5526 * Also keep a unique ID per domain (we use the first cpu number in
5527 * the cpumask of the domain), this allows us to quickly tell if
5528 * two cpus are in the same cache domain, see cpus_share_cache().
5530 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5531 DEFINE_PER_CPU(int, sd_llc_size);
5532 DEFINE_PER_CPU(int, sd_llc_id);
5533 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5534 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5535 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5537 static void update_top_cache_domain(int cpu)
5539 struct sched_domain *sd;
5540 struct sched_domain *busy_sd = NULL;
5544 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5546 id = cpumask_first(sched_domain_span(sd));
5547 size = cpumask_weight(sched_domain_span(sd));
5548 busy_sd = sd->parent; /* sd_busy */
5550 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5552 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5553 per_cpu(sd_llc_size, cpu) = size;
5554 per_cpu(sd_llc_id, cpu) = id;
5556 sd = lowest_flag_domain(cpu, SD_NUMA);
5557 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5559 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5560 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5564 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5565 * hold the hotplug lock.
5568 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5570 struct rq *rq = cpu_rq(cpu);
5571 struct sched_domain *tmp;
5573 /* Remove the sched domains which do not contribute to scheduling. */
5574 for (tmp = sd; tmp; ) {
5575 struct sched_domain *parent = tmp->parent;
5579 if (sd_parent_degenerate(tmp, parent)) {
5580 tmp->parent = parent->parent;
5582 parent->parent->child = tmp;
5584 * Transfer SD_PREFER_SIBLING down in case of a
5585 * degenerate parent; the spans match for this
5586 * so the property transfers.
5588 if (parent->flags & SD_PREFER_SIBLING)
5589 tmp->flags |= SD_PREFER_SIBLING;
5590 destroy_sched_domain(parent, cpu);
5595 if (sd && sd_degenerate(sd)) {
5598 destroy_sched_domain(tmp, cpu);
5603 sched_domain_debug(sd, cpu);
5605 rq_attach_root(rq, rd);
5607 rcu_assign_pointer(rq->sd, sd);
5608 destroy_sched_domains(tmp, cpu);
5610 update_top_cache_domain(cpu);
5613 /* cpus with isolated domains */
5614 static cpumask_var_t cpu_isolated_map;
5616 /* Setup the mask of cpus configured for isolated domains */
5617 static int __init isolated_cpu_setup(char *str)
5619 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5620 cpulist_parse(str, cpu_isolated_map);
5624 __setup("isolcpus=", isolated_cpu_setup);
5627 struct sched_domain ** __percpu sd;
5628 struct root_domain *rd;
5639 * Build an iteration mask that can exclude certain CPUs from the upwards
5642 * Asymmetric node setups can result in situations where the domain tree is of
5643 * unequal depth, make sure to skip domains that already cover the entire
5646 * In that case build_sched_domains() will have terminated the iteration early
5647 * and our sibling sd spans will be empty. Domains should always include the
5648 * cpu they're built on, so check that.
5651 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5653 const struct cpumask *span = sched_domain_span(sd);
5654 struct sd_data *sdd = sd->private;
5655 struct sched_domain *sibling;
5658 for_each_cpu(i, span) {
5659 sibling = *per_cpu_ptr(sdd->sd, i);
5660 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5663 cpumask_set_cpu(i, sched_group_mask(sg));
5668 * Return the canonical balance cpu for this group, this is the first cpu
5669 * of this group that's also in the iteration mask.
5671 int group_balance_cpu(struct sched_group *sg)
5673 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5677 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5679 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5680 const struct cpumask *span = sched_domain_span(sd);
5681 struct cpumask *covered = sched_domains_tmpmask;
5682 struct sd_data *sdd = sd->private;
5683 struct sched_domain *child;
5686 cpumask_clear(covered);
5688 for_each_cpu(i, span) {
5689 struct cpumask *sg_span;
5691 if (cpumask_test_cpu(i, covered))
5694 child = *per_cpu_ptr(sdd->sd, i);
5696 /* See the comment near build_group_mask(). */
5697 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5700 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5701 GFP_KERNEL, cpu_to_node(cpu));
5706 sg_span = sched_group_cpus(sg);
5708 child = child->child;
5709 cpumask_copy(sg_span, sched_domain_span(child));
5711 cpumask_set_cpu(i, sg_span);
5713 cpumask_or(covered, covered, sg_span);
5715 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5716 if (atomic_inc_return(&sg->sgp->ref) == 1)
5717 build_group_mask(sd, sg);
5720 * Initialize sgp->power such that even if we mess up the
5721 * domains and no possible iteration will get us here, we won't
5724 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5725 sg->sgp->power_orig = sg->sgp->power;
5728 * Make sure the first group of this domain contains the
5729 * canonical balance cpu. Otherwise the sched_domain iteration
5730 * breaks. See update_sg_lb_stats().
5732 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5733 group_balance_cpu(sg) == cpu)
5743 sd->groups = groups;
5748 free_sched_groups(first, 0);
5753 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5755 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5756 struct sched_domain *child = sd->child;
5759 cpu = cpumask_first(sched_domain_span(child));
5762 *sg = *per_cpu_ptr(sdd->sg, cpu);
5763 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5764 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5771 * build_sched_groups will build a circular linked list of the groups
5772 * covered by the given span, and will set each group's ->cpumask correctly,
5773 * and ->cpu_power to 0.
5775 * Assumes the sched_domain tree is fully constructed
5778 build_sched_groups(struct sched_domain *sd, int cpu)
5780 struct sched_group *first = NULL, *last = NULL;
5781 struct sd_data *sdd = sd->private;
5782 const struct cpumask *span = sched_domain_span(sd);
5783 struct cpumask *covered;
5786 get_group(cpu, sdd, &sd->groups);
5787 atomic_inc(&sd->groups->ref);
5789 if (cpu != cpumask_first(span))
5792 lockdep_assert_held(&sched_domains_mutex);
5793 covered = sched_domains_tmpmask;
5795 cpumask_clear(covered);
5797 for_each_cpu(i, span) {
5798 struct sched_group *sg;
5801 if (cpumask_test_cpu(i, covered))
5804 group = get_group(i, sdd, &sg);
5805 cpumask_clear(sched_group_cpus(sg));
5807 cpumask_setall(sched_group_mask(sg));
5809 for_each_cpu(j, span) {
5810 if (get_group(j, sdd, NULL) != group)
5813 cpumask_set_cpu(j, covered);
5814 cpumask_set_cpu(j, sched_group_cpus(sg));
5829 * Initialize sched groups cpu_power.
5831 * cpu_power indicates the capacity of sched group, which is used while
5832 * distributing the load between different sched groups in a sched domain.
5833 * Typically cpu_power for all the groups in a sched domain will be same unless
5834 * there are asymmetries in the topology. If there are asymmetries, group
5835 * having more cpu_power will pickup more load compared to the group having
5838 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5840 struct sched_group *sg = sd->groups;
5845 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5847 } while (sg != sd->groups);
5849 if (cpu != group_balance_cpu(sg))
5852 update_group_power(sd, cpu);
5853 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5857 * Initializers for schedule domains
5858 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5861 static int default_relax_domain_level = -1;
5862 int sched_domain_level_max;
5864 static int __init setup_relax_domain_level(char *str)
5866 if (kstrtoint(str, 0, &default_relax_domain_level))
5867 pr_warn("Unable to set relax_domain_level\n");
5871 __setup("relax_domain_level=", setup_relax_domain_level);
5873 static void set_domain_attribute(struct sched_domain *sd,
5874 struct sched_domain_attr *attr)
5878 if (!attr || attr->relax_domain_level < 0) {
5879 if (default_relax_domain_level < 0)
5882 request = default_relax_domain_level;
5884 request = attr->relax_domain_level;
5885 if (request < sd->level) {
5886 /* turn off idle balance on this domain */
5887 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5889 /* turn on idle balance on this domain */
5890 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5894 static void __sdt_free(const struct cpumask *cpu_map);
5895 static int __sdt_alloc(const struct cpumask *cpu_map);
5897 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5898 const struct cpumask *cpu_map)
5902 if (!atomic_read(&d->rd->refcount))
5903 free_rootdomain(&d->rd->rcu); /* fall through */
5905 free_percpu(d->sd); /* fall through */
5907 __sdt_free(cpu_map); /* fall through */
5913 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5914 const struct cpumask *cpu_map)
5916 memset(d, 0, sizeof(*d));
5918 if (__sdt_alloc(cpu_map))
5919 return sa_sd_storage;
5920 d->sd = alloc_percpu(struct sched_domain *);
5922 return sa_sd_storage;
5923 d->rd = alloc_rootdomain();
5926 return sa_rootdomain;
5930 * NULL the sd_data elements we've used to build the sched_domain and
5931 * sched_group structure so that the subsequent __free_domain_allocs()
5932 * will not free the data we're using.
5934 static void claim_allocations(int cpu, struct sched_domain *sd)
5936 struct sd_data *sdd = sd->private;
5938 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5939 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5941 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5942 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5944 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5945 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5949 static int sched_domains_numa_levels;
5950 static int *sched_domains_numa_distance;
5951 static struct cpumask ***sched_domains_numa_masks;
5952 static int sched_domains_curr_level;
5956 * SD_flags allowed in topology descriptions.
5958 * SD_SHARE_CPUPOWER - describes SMT topologies
5959 * SD_SHARE_PKG_RESOURCES - describes shared caches
5960 * SD_NUMA - describes NUMA topologies
5961 * SD_SHARE_POWERDOMAIN - describes shared power domain
5964 * SD_ASYM_PACKING - describes SMT quirks
5966 #define TOPOLOGY_SD_FLAGS \
5967 (SD_SHARE_CPUPOWER | \
5968 SD_SHARE_PKG_RESOURCES | \
5971 SD_SHARE_POWERDOMAIN)
5973 static struct sched_domain *
5974 sd_init(struct sched_domain_topology_level *tl, int cpu)
5976 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5977 int sd_weight, sd_flags = 0;
5981 * Ugly hack to pass state to sd_numa_mask()...
5983 sched_domains_curr_level = tl->numa_level;
5986 sd_weight = cpumask_weight(tl->mask(cpu));
5989 sd_flags = (*tl->sd_flags)();
5990 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
5991 "wrong sd_flags in topology description\n"))
5992 sd_flags &= ~TOPOLOGY_SD_FLAGS;
5994 *sd = (struct sched_domain){
5995 .min_interval = sd_weight,
5996 .max_interval = 2*sd_weight,
5998 .imbalance_pct = 125,
6000 .cache_nice_tries = 0,
6007 .flags = 1*SD_LOAD_BALANCE
6008 | 1*SD_BALANCE_NEWIDLE
6013 | 0*SD_SHARE_CPUPOWER
6014 | 0*SD_SHARE_PKG_RESOURCES
6016 | 0*SD_PREFER_SIBLING
6021 .last_balance = jiffies,
6022 .balance_interval = sd_weight,
6024 .max_newidle_lb_cost = 0,
6025 .next_decay_max_lb_cost = jiffies,
6026 #ifdef CONFIG_SCHED_DEBUG
6032 * Convert topological properties into behaviour.
6035 if (sd->flags & SD_SHARE_CPUPOWER) {
6036 sd->imbalance_pct = 110;
6037 sd->smt_gain = 1178; /* ~15% */
6039 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6040 sd->imbalance_pct = 117;
6041 sd->cache_nice_tries = 1;
6045 } else if (sd->flags & SD_NUMA) {
6046 sd->cache_nice_tries = 2;
6050 sd->flags |= SD_SERIALIZE;
6051 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6052 sd->flags &= ~(SD_BALANCE_EXEC |
6059 sd->flags |= SD_PREFER_SIBLING;
6060 sd->cache_nice_tries = 1;
6065 sd->private = &tl->data;
6071 * Topology list, bottom-up.
6073 static struct sched_domain_topology_level default_topology[] = {
6074 #ifdef CONFIG_SCHED_SMT
6075 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6077 #ifdef CONFIG_SCHED_MC
6078 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6080 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6084 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6086 #define for_each_sd_topology(tl) \
6087 for (tl = sched_domain_topology; tl->mask; tl++)
6089 void set_sched_topology(struct sched_domain_topology_level *tl)
6091 sched_domain_topology = tl;
6096 static const struct cpumask *sd_numa_mask(int cpu)
6098 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6101 static void sched_numa_warn(const char *str)
6103 static int done = false;
6111 printk(KERN_WARNING "ERROR: %s\n\n", str);
6113 for (i = 0; i < nr_node_ids; i++) {
6114 printk(KERN_WARNING " ");
6115 for (j = 0; j < nr_node_ids; j++)
6116 printk(KERN_CONT "%02d ", node_distance(i,j));
6117 printk(KERN_CONT "\n");
6119 printk(KERN_WARNING "\n");
6122 static bool find_numa_distance(int distance)
6126 if (distance == node_distance(0, 0))
6129 for (i = 0; i < sched_domains_numa_levels; i++) {
6130 if (sched_domains_numa_distance[i] == distance)
6137 static void sched_init_numa(void)
6139 int next_distance, curr_distance = node_distance(0, 0);
6140 struct sched_domain_topology_level *tl;
6144 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6145 if (!sched_domains_numa_distance)
6149 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6150 * unique distances in the node_distance() table.
6152 * Assumes node_distance(0,j) includes all distances in
6153 * node_distance(i,j) in order to avoid cubic time.
6155 next_distance = curr_distance;
6156 for (i = 0; i < nr_node_ids; i++) {
6157 for (j = 0; j < nr_node_ids; j++) {
6158 for (k = 0; k < nr_node_ids; k++) {
6159 int distance = node_distance(i, k);
6161 if (distance > curr_distance &&
6162 (distance < next_distance ||
6163 next_distance == curr_distance))
6164 next_distance = distance;
6167 * While not a strong assumption it would be nice to know
6168 * about cases where if node A is connected to B, B is not
6169 * equally connected to A.
6171 if (sched_debug() && node_distance(k, i) != distance)
6172 sched_numa_warn("Node-distance not symmetric");
6174 if (sched_debug() && i && !find_numa_distance(distance))
6175 sched_numa_warn("Node-0 not representative");
6177 if (next_distance != curr_distance) {
6178 sched_domains_numa_distance[level++] = next_distance;
6179 sched_domains_numa_levels = level;
6180 curr_distance = next_distance;
6185 * In case of sched_debug() we verify the above assumption.
6191 * 'level' contains the number of unique distances, excluding the
6192 * identity distance node_distance(i,i).
6194 * The sched_domains_numa_distance[] array includes the actual distance
6199 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6200 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6201 * the array will contain less then 'level' members. This could be
6202 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6203 * in other functions.
6205 * We reset it to 'level' at the end of this function.
6207 sched_domains_numa_levels = 0;
6209 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6210 if (!sched_domains_numa_masks)
6214 * Now for each level, construct a mask per node which contains all
6215 * cpus of nodes that are that many hops away from us.
6217 for (i = 0; i < level; i++) {
6218 sched_domains_numa_masks[i] =
6219 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6220 if (!sched_domains_numa_masks[i])
6223 for (j = 0; j < nr_node_ids; j++) {
6224 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6228 sched_domains_numa_masks[i][j] = mask;
6230 for (k = 0; k < nr_node_ids; k++) {
6231 if (node_distance(j, k) > sched_domains_numa_distance[i])
6234 cpumask_or(mask, mask, cpumask_of_node(k));
6239 /* Compute default topology size */
6240 for (i = 0; sched_domain_topology[i].mask; i++);
6242 tl = kzalloc((i + level) *
6243 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6248 * Copy the default topology bits..
6250 for (i = 0; sched_domain_topology[i].mask; i++)
6251 tl[i] = sched_domain_topology[i];
6254 * .. and append 'j' levels of NUMA goodness.
6256 for (j = 0; j < level; i++, j++) {
6257 tl[i] = (struct sched_domain_topology_level){
6258 .mask = sd_numa_mask,
6259 .sd_flags = cpu_numa_flags,
6260 .flags = SDTL_OVERLAP,
6266 sched_domain_topology = tl;
6268 sched_domains_numa_levels = level;
6271 static void sched_domains_numa_masks_set(int cpu)
6274 int node = cpu_to_node(cpu);
6276 for (i = 0; i < sched_domains_numa_levels; i++) {
6277 for (j = 0; j < nr_node_ids; j++) {
6278 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6279 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6284 static void sched_domains_numa_masks_clear(int cpu)
6287 for (i = 0; i < sched_domains_numa_levels; i++) {
6288 for (j = 0; j < nr_node_ids; j++)
6289 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6294 * Update sched_domains_numa_masks[level][node] array when new cpus
6297 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6298 unsigned long action,
6301 int cpu = (long)hcpu;
6303 switch (action & ~CPU_TASKS_FROZEN) {
6305 sched_domains_numa_masks_set(cpu);
6309 sched_domains_numa_masks_clear(cpu);
6319 static inline void sched_init_numa(void)
6323 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6324 unsigned long action,
6329 #endif /* CONFIG_NUMA */
6331 static int __sdt_alloc(const struct cpumask *cpu_map)
6333 struct sched_domain_topology_level *tl;
6336 for_each_sd_topology(tl) {
6337 struct sd_data *sdd = &tl->data;
6339 sdd->sd = alloc_percpu(struct sched_domain *);
6343 sdd->sg = alloc_percpu(struct sched_group *);
6347 sdd->sgp = alloc_percpu(struct sched_group_power *);
6351 for_each_cpu(j, cpu_map) {
6352 struct sched_domain *sd;
6353 struct sched_group *sg;
6354 struct sched_group_power *sgp;
6356 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6357 GFP_KERNEL, cpu_to_node(j));
6361 *per_cpu_ptr(sdd->sd, j) = sd;
6363 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6364 GFP_KERNEL, cpu_to_node(j));
6370 *per_cpu_ptr(sdd->sg, j) = sg;
6372 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6373 GFP_KERNEL, cpu_to_node(j));
6377 *per_cpu_ptr(sdd->sgp, j) = sgp;
6384 static void __sdt_free(const struct cpumask *cpu_map)
6386 struct sched_domain_topology_level *tl;
6389 for_each_sd_topology(tl) {
6390 struct sd_data *sdd = &tl->data;
6392 for_each_cpu(j, cpu_map) {
6393 struct sched_domain *sd;
6396 sd = *per_cpu_ptr(sdd->sd, j);
6397 if (sd && (sd->flags & SD_OVERLAP))
6398 free_sched_groups(sd->groups, 0);
6399 kfree(*per_cpu_ptr(sdd->sd, j));
6403 kfree(*per_cpu_ptr(sdd->sg, j));
6405 kfree(*per_cpu_ptr(sdd->sgp, j));
6407 free_percpu(sdd->sd);
6409 free_percpu(sdd->sg);
6411 free_percpu(sdd->sgp);
6416 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6417 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6418 struct sched_domain *child, int cpu)
6420 struct sched_domain *sd = sd_init(tl, cpu);
6424 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6426 sd->level = child->level + 1;
6427 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6431 set_domain_attribute(sd, attr);
6437 * Build sched domains for a given set of cpus and attach the sched domains
6438 * to the individual cpus
6440 static int build_sched_domains(const struct cpumask *cpu_map,
6441 struct sched_domain_attr *attr)
6443 enum s_alloc alloc_state;
6444 struct sched_domain *sd;
6446 int i, ret = -ENOMEM;
6448 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6449 if (alloc_state != sa_rootdomain)
6452 /* Set up domains for cpus specified by the cpu_map. */
6453 for_each_cpu(i, cpu_map) {
6454 struct sched_domain_topology_level *tl;
6457 for_each_sd_topology(tl) {
6458 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6459 if (tl == sched_domain_topology)
6460 *per_cpu_ptr(d.sd, i) = sd;
6461 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6462 sd->flags |= SD_OVERLAP;
6463 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6468 /* Build the groups for the domains */
6469 for_each_cpu(i, cpu_map) {
6470 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6471 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6472 if (sd->flags & SD_OVERLAP) {
6473 if (build_overlap_sched_groups(sd, i))
6476 if (build_sched_groups(sd, i))
6482 /* Calculate CPU power for physical packages and nodes */
6483 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6484 if (!cpumask_test_cpu(i, cpu_map))
6487 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6488 claim_allocations(i, sd);
6489 init_sched_groups_power(i, sd);
6493 /* Attach the domains */
6495 for_each_cpu(i, cpu_map) {
6496 sd = *per_cpu_ptr(d.sd, i);
6497 cpu_attach_domain(sd, d.rd, i);
6503 __free_domain_allocs(&d, alloc_state, cpu_map);
6507 static cpumask_var_t *doms_cur; /* current sched domains */
6508 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6509 static struct sched_domain_attr *dattr_cur;
6510 /* attribues of custom domains in 'doms_cur' */
6513 * Special case: If a kmalloc of a doms_cur partition (array of
6514 * cpumask) fails, then fallback to a single sched domain,
6515 * as determined by the single cpumask fallback_doms.
6517 static cpumask_var_t fallback_doms;
6520 * arch_update_cpu_topology lets virtualized architectures update the
6521 * cpu core maps. It is supposed to return 1 if the topology changed
6522 * or 0 if it stayed the same.
6524 int __weak arch_update_cpu_topology(void)
6529 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6532 cpumask_var_t *doms;
6534 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6537 for (i = 0; i < ndoms; i++) {
6538 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6539 free_sched_domains(doms, i);
6546 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6549 for (i = 0; i < ndoms; i++)
6550 free_cpumask_var(doms[i]);
6555 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6556 * For now this just excludes isolated cpus, but could be used to
6557 * exclude other special cases in the future.
6559 static int init_sched_domains(const struct cpumask *cpu_map)
6563 arch_update_cpu_topology();
6565 doms_cur = alloc_sched_domains(ndoms_cur);
6567 doms_cur = &fallback_doms;
6568 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6569 err = build_sched_domains(doms_cur[0], NULL);
6570 register_sched_domain_sysctl();
6576 * Detach sched domains from a group of cpus specified in cpu_map
6577 * These cpus will now be attached to the NULL domain
6579 static void detach_destroy_domains(const struct cpumask *cpu_map)
6584 for_each_cpu(i, cpu_map)
6585 cpu_attach_domain(NULL, &def_root_domain, i);
6589 /* handle null as "default" */
6590 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6591 struct sched_domain_attr *new, int idx_new)
6593 struct sched_domain_attr tmp;
6600 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6601 new ? (new + idx_new) : &tmp,
6602 sizeof(struct sched_domain_attr));
6606 * Partition sched domains as specified by the 'ndoms_new'
6607 * cpumasks in the array doms_new[] of cpumasks. This compares
6608 * doms_new[] to the current sched domain partitioning, doms_cur[].
6609 * It destroys each deleted domain and builds each new domain.
6611 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6612 * The masks don't intersect (don't overlap.) We should setup one
6613 * sched domain for each mask. CPUs not in any of the cpumasks will
6614 * not be load balanced. If the same cpumask appears both in the
6615 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6618 * The passed in 'doms_new' should be allocated using
6619 * alloc_sched_domains. This routine takes ownership of it and will
6620 * free_sched_domains it when done with it. If the caller failed the
6621 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6622 * and partition_sched_domains() will fallback to the single partition
6623 * 'fallback_doms', it also forces the domains to be rebuilt.
6625 * If doms_new == NULL it will be replaced with cpu_online_mask.
6626 * ndoms_new == 0 is a special case for destroying existing domains,
6627 * and it will not create the default domain.
6629 * Call with hotplug lock held
6631 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6632 struct sched_domain_attr *dattr_new)
6637 mutex_lock(&sched_domains_mutex);
6639 /* always unregister in case we don't destroy any domains */
6640 unregister_sched_domain_sysctl();
6642 /* Let architecture update cpu core mappings. */
6643 new_topology = arch_update_cpu_topology();
6645 n = doms_new ? ndoms_new : 0;
6647 /* Destroy deleted domains */
6648 for (i = 0; i < ndoms_cur; i++) {
6649 for (j = 0; j < n && !new_topology; j++) {
6650 if (cpumask_equal(doms_cur[i], doms_new[j])
6651 && dattrs_equal(dattr_cur, i, dattr_new, j))
6654 /* no match - a current sched domain not in new doms_new[] */
6655 detach_destroy_domains(doms_cur[i]);
6661 if (doms_new == NULL) {
6663 doms_new = &fallback_doms;
6664 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6665 WARN_ON_ONCE(dattr_new);
6668 /* Build new domains */
6669 for (i = 0; i < ndoms_new; i++) {
6670 for (j = 0; j < n && !new_topology; j++) {
6671 if (cpumask_equal(doms_new[i], doms_cur[j])
6672 && dattrs_equal(dattr_new, i, dattr_cur, j))
6675 /* no match - add a new doms_new */
6676 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6681 /* Remember the new sched domains */
6682 if (doms_cur != &fallback_doms)
6683 free_sched_domains(doms_cur, ndoms_cur);
6684 kfree(dattr_cur); /* kfree(NULL) is safe */
6685 doms_cur = doms_new;
6686 dattr_cur = dattr_new;
6687 ndoms_cur = ndoms_new;
6689 register_sched_domain_sysctl();
6691 mutex_unlock(&sched_domains_mutex);
6694 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6697 * Update cpusets according to cpu_active mask. If cpusets are
6698 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6699 * around partition_sched_domains().
6701 * If we come here as part of a suspend/resume, don't touch cpusets because we
6702 * want to restore it back to its original state upon resume anyway.
6704 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6708 case CPU_ONLINE_FROZEN:
6709 case CPU_DOWN_FAILED_FROZEN:
6712 * num_cpus_frozen tracks how many CPUs are involved in suspend
6713 * resume sequence. As long as this is not the last online
6714 * operation in the resume sequence, just build a single sched
6715 * domain, ignoring cpusets.
6718 if (likely(num_cpus_frozen)) {
6719 partition_sched_domains(1, NULL, NULL);
6724 * This is the last CPU online operation. So fall through and
6725 * restore the original sched domains by considering the
6726 * cpuset configurations.
6730 case CPU_DOWN_FAILED:
6731 cpuset_update_active_cpus(true);
6739 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6743 case CPU_DOWN_PREPARE:
6744 cpuset_update_active_cpus(false);
6746 case CPU_DOWN_PREPARE_FROZEN:
6748 partition_sched_domains(1, NULL, NULL);
6756 void __init sched_init_smp(void)
6758 cpumask_var_t non_isolated_cpus;
6760 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6761 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6766 * There's no userspace yet to cause hotplug operations; hence all the
6767 * cpu masks are stable and all blatant races in the below code cannot
6770 mutex_lock(&sched_domains_mutex);
6771 init_sched_domains(cpu_active_mask);
6772 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6773 if (cpumask_empty(non_isolated_cpus))
6774 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6775 mutex_unlock(&sched_domains_mutex);
6777 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6778 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6779 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6783 /* Move init over to a non-isolated CPU */
6784 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6786 sched_init_granularity();
6787 free_cpumask_var(non_isolated_cpus);
6789 init_sched_rt_class();
6790 init_sched_dl_class();
6793 void __init sched_init_smp(void)
6795 sched_init_granularity();
6797 #endif /* CONFIG_SMP */
6799 const_debug unsigned int sysctl_timer_migration = 1;
6801 int in_sched_functions(unsigned long addr)
6803 return in_lock_functions(addr) ||
6804 (addr >= (unsigned long)__sched_text_start
6805 && addr < (unsigned long)__sched_text_end);
6808 #ifdef CONFIG_CGROUP_SCHED
6810 * Default task group.
6811 * Every task in system belongs to this group at bootup.
6813 struct task_group root_task_group;
6814 LIST_HEAD(task_groups);
6817 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6819 void __init sched_init(void)
6822 unsigned long alloc_size = 0, ptr;
6824 #ifdef CONFIG_FAIR_GROUP_SCHED
6825 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6827 #ifdef CONFIG_RT_GROUP_SCHED
6828 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6830 #ifdef CONFIG_CPUMASK_OFFSTACK
6831 alloc_size += num_possible_cpus() * cpumask_size();
6834 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6836 #ifdef CONFIG_FAIR_GROUP_SCHED
6837 root_task_group.se = (struct sched_entity **)ptr;
6838 ptr += nr_cpu_ids * sizeof(void **);
6840 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6841 ptr += nr_cpu_ids * sizeof(void **);
6843 #endif /* CONFIG_FAIR_GROUP_SCHED */
6844 #ifdef CONFIG_RT_GROUP_SCHED
6845 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6846 ptr += nr_cpu_ids * sizeof(void **);
6848 root_task_group.rt_rq = (struct rt_rq **)ptr;
6849 ptr += nr_cpu_ids * sizeof(void **);
6851 #endif /* CONFIG_RT_GROUP_SCHED */
6852 #ifdef CONFIG_CPUMASK_OFFSTACK
6853 for_each_possible_cpu(i) {
6854 per_cpu(load_balance_mask, i) = (void *)ptr;
6855 ptr += cpumask_size();
6857 #endif /* CONFIG_CPUMASK_OFFSTACK */
6860 init_rt_bandwidth(&def_rt_bandwidth,
6861 global_rt_period(), global_rt_runtime());
6862 init_dl_bandwidth(&def_dl_bandwidth,
6863 global_rt_period(), global_rt_runtime());
6866 init_defrootdomain();
6869 #ifdef CONFIG_RT_GROUP_SCHED
6870 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6871 global_rt_period(), global_rt_runtime());
6872 #endif /* CONFIG_RT_GROUP_SCHED */
6874 #ifdef CONFIG_CGROUP_SCHED
6875 list_add(&root_task_group.list, &task_groups);
6876 INIT_LIST_HEAD(&root_task_group.children);
6877 INIT_LIST_HEAD(&root_task_group.siblings);
6878 autogroup_init(&init_task);
6880 #endif /* CONFIG_CGROUP_SCHED */
6882 for_each_possible_cpu(i) {
6886 raw_spin_lock_init(&rq->lock);
6888 rq->calc_load_active = 0;
6889 rq->calc_load_update = jiffies + LOAD_FREQ;
6890 init_cfs_rq(&rq->cfs);
6891 init_rt_rq(&rq->rt, rq);
6892 init_dl_rq(&rq->dl, rq);
6893 #ifdef CONFIG_FAIR_GROUP_SCHED
6894 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6895 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6897 * How much cpu bandwidth does root_task_group get?
6899 * In case of task-groups formed thr' the cgroup filesystem, it
6900 * gets 100% of the cpu resources in the system. This overall
6901 * system cpu resource is divided among the tasks of
6902 * root_task_group and its child task-groups in a fair manner,
6903 * based on each entity's (task or task-group's) weight
6904 * (se->load.weight).
6906 * In other words, if root_task_group has 10 tasks of weight
6907 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6908 * then A0's share of the cpu resource is:
6910 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6912 * We achieve this by letting root_task_group's tasks sit
6913 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6915 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6916 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6917 #endif /* CONFIG_FAIR_GROUP_SCHED */
6919 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6920 #ifdef CONFIG_RT_GROUP_SCHED
6921 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6924 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6925 rq->cpu_load[j] = 0;
6927 rq->last_load_update_tick = jiffies;
6932 rq->cpu_power = SCHED_POWER_SCALE;
6933 rq->post_schedule = 0;
6934 rq->active_balance = 0;
6935 rq->next_balance = jiffies;
6940 rq->avg_idle = 2*sysctl_sched_migration_cost;
6941 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6943 INIT_LIST_HEAD(&rq->cfs_tasks);
6945 rq_attach_root(rq, &def_root_domain);
6946 #ifdef CONFIG_NO_HZ_COMMON
6949 #ifdef CONFIG_NO_HZ_FULL
6950 rq->last_sched_tick = 0;
6954 atomic_set(&rq->nr_iowait, 0);
6957 set_load_weight(&init_task);
6959 #ifdef CONFIG_PREEMPT_NOTIFIERS
6960 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6964 * The boot idle thread does lazy MMU switching as well:
6966 atomic_inc(&init_mm.mm_count);
6967 enter_lazy_tlb(&init_mm, current);
6970 * Make us the idle thread. Technically, schedule() should not be
6971 * called from this thread, however somewhere below it might be,
6972 * but because we are the idle thread, we just pick up running again
6973 * when this runqueue becomes "idle".
6975 init_idle(current, smp_processor_id());
6977 calc_load_update = jiffies + LOAD_FREQ;
6980 * During early bootup we pretend to be a normal task:
6982 current->sched_class = &fair_sched_class;
6985 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6986 /* May be allocated at isolcpus cmdline parse time */
6987 if (cpu_isolated_map == NULL)
6988 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6989 idle_thread_set_boot_cpu();
6991 init_sched_fair_class();
6993 scheduler_running = 1;
6996 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6997 static inline int preempt_count_equals(int preempt_offset)
6999 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7001 return (nested == preempt_offset);
7004 void __might_sleep(const char *file, int line, int preempt_offset)
7006 static unsigned long prev_jiffy; /* ratelimiting */
7008 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7009 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7010 !is_idle_task(current)) ||
7011 system_state != SYSTEM_RUNNING || oops_in_progress)
7013 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7015 prev_jiffy = jiffies;
7018 "BUG: sleeping function called from invalid context at %s:%d\n",
7021 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7022 in_atomic(), irqs_disabled(),
7023 current->pid, current->comm);
7025 debug_show_held_locks(current);
7026 if (irqs_disabled())
7027 print_irqtrace_events(current);
7028 #ifdef CONFIG_DEBUG_PREEMPT
7029 if (!preempt_count_equals(preempt_offset)) {
7030 pr_err("Preemption disabled at:");
7031 print_ip_sym(current->preempt_disable_ip);
7037 EXPORT_SYMBOL(__might_sleep);
7040 #ifdef CONFIG_MAGIC_SYSRQ
7041 static void normalize_task(struct rq *rq, struct task_struct *p)
7043 const struct sched_class *prev_class = p->sched_class;
7044 struct sched_attr attr = {
7045 .sched_policy = SCHED_NORMAL,
7047 int old_prio = p->prio;
7052 dequeue_task(rq, p, 0);
7053 __setscheduler(rq, p, &attr);
7055 enqueue_task(rq, p, 0);
7056 resched_task(rq->curr);
7059 check_class_changed(rq, p, prev_class, old_prio);
7062 void normalize_rt_tasks(void)
7064 struct task_struct *g, *p;
7065 unsigned long flags;
7068 read_lock_irqsave(&tasklist_lock, flags);
7069 do_each_thread(g, p) {
7071 * Only normalize user tasks:
7076 p->se.exec_start = 0;
7077 #ifdef CONFIG_SCHEDSTATS
7078 p->se.statistics.wait_start = 0;
7079 p->se.statistics.sleep_start = 0;
7080 p->se.statistics.block_start = 0;
7083 if (!dl_task(p) && !rt_task(p)) {
7085 * Renice negative nice level userspace
7088 if (task_nice(p) < 0 && p->mm)
7089 set_user_nice(p, 0);
7093 raw_spin_lock(&p->pi_lock);
7094 rq = __task_rq_lock(p);
7096 normalize_task(rq, p);
7098 __task_rq_unlock(rq);
7099 raw_spin_unlock(&p->pi_lock);
7100 } while_each_thread(g, p);
7102 read_unlock_irqrestore(&tasklist_lock, flags);
7105 #endif /* CONFIG_MAGIC_SYSRQ */
7107 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7109 * These functions are only useful for the IA64 MCA handling, or kdb.
7111 * They can only be called when the whole system has been
7112 * stopped - every CPU needs to be quiescent, and no scheduling
7113 * activity can take place. Using them for anything else would
7114 * be a serious bug, and as a result, they aren't even visible
7115 * under any other configuration.
7119 * curr_task - return the current task for a given cpu.
7120 * @cpu: the processor in question.
7122 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7124 * Return: The current task for @cpu.
7126 struct task_struct *curr_task(int cpu)
7128 return cpu_curr(cpu);
7131 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7135 * set_curr_task - set the current task for a given cpu.
7136 * @cpu: the processor in question.
7137 * @p: the task pointer to set.
7139 * Description: This function must only be used when non-maskable interrupts
7140 * are serviced on a separate stack. It allows the architecture to switch the
7141 * notion of the current task on a cpu in a non-blocking manner. This function
7142 * must be called with all CPU's synchronized, and interrupts disabled, the
7143 * and caller must save the original value of the current task (see
7144 * curr_task() above) and restore that value before reenabling interrupts and
7145 * re-starting the system.
7147 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7149 void set_curr_task(int cpu, struct task_struct *p)
7156 #ifdef CONFIG_CGROUP_SCHED
7157 /* task_group_lock serializes the addition/removal of task groups */
7158 static DEFINE_SPINLOCK(task_group_lock);
7160 static void free_sched_group(struct task_group *tg)
7162 free_fair_sched_group(tg);
7163 free_rt_sched_group(tg);
7168 /* allocate runqueue etc for a new task group */
7169 struct task_group *sched_create_group(struct task_group *parent)
7171 struct task_group *tg;
7173 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7175 return ERR_PTR(-ENOMEM);
7177 if (!alloc_fair_sched_group(tg, parent))
7180 if (!alloc_rt_sched_group(tg, parent))
7186 free_sched_group(tg);
7187 return ERR_PTR(-ENOMEM);
7190 void sched_online_group(struct task_group *tg, struct task_group *parent)
7192 unsigned long flags;
7194 spin_lock_irqsave(&task_group_lock, flags);
7195 list_add_rcu(&tg->list, &task_groups);
7197 WARN_ON(!parent); /* root should already exist */
7199 tg->parent = parent;
7200 INIT_LIST_HEAD(&tg->children);
7201 list_add_rcu(&tg->siblings, &parent->children);
7202 spin_unlock_irqrestore(&task_group_lock, flags);
7205 /* rcu callback to free various structures associated with a task group */
7206 static void free_sched_group_rcu(struct rcu_head *rhp)
7208 /* now it should be safe to free those cfs_rqs */
7209 free_sched_group(container_of(rhp, struct task_group, rcu));
7212 /* Destroy runqueue etc associated with a task group */
7213 void sched_destroy_group(struct task_group *tg)
7215 /* wait for possible concurrent references to cfs_rqs complete */
7216 call_rcu(&tg->rcu, free_sched_group_rcu);
7219 void sched_offline_group(struct task_group *tg)
7221 unsigned long flags;
7224 /* end participation in shares distribution */
7225 for_each_possible_cpu(i)
7226 unregister_fair_sched_group(tg, i);
7228 spin_lock_irqsave(&task_group_lock, flags);
7229 list_del_rcu(&tg->list);
7230 list_del_rcu(&tg->siblings);
7231 spin_unlock_irqrestore(&task_group_lock, flags);
7234 /* change task's runqueue when it moves between groups.
7235 * The caller of this function should have put the task in its new group
7236 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7237 * reflect its new group.
7239 void sched_move_task(struct task_struct *tsk)
7241 struct task_group *tg;
7243 unsigned long flags;
7246 rq = task_rq_lock(tsk, &flags);
7248 running = task_current(rq, tsk);
7252 dequeue_task(rq, tsk, 0);
7253 if (unlikely(running))
7254 tsk->sched_class->put_prev_task(rq, tsk);
7256 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7257 lockdep_is_held(&tsk->sighand->siglock)),
7258 struct task_group, css);
7259 tg = autogroup_task_group(tsk, tg);
7260 tsk->sched_task_group = tg;
7262 #ifdef CONFIG_FAIR_GROUP_SCHED
7263 if (tsk->sched_class->task_move_group)
7264 tsk->sched_class->task_move_group(tsk, on_rq);
7267 set_task_rq(tsk, task_cpu(tsk));
7269 if (unlikely(running))
7270 tsk->sched_class->set_curr_task(rq);
7272 enqueue_task(rq, tsk, 0);
7274 task_rq_unlock(rq, tsk, &flags);
7276 #endif /* CONFIG_CGROUP_SCHED */
7278 #ifdef CONFIG_RT_GROUP_SCHED
7280 * Ensure that the real time constraints are schedulable.
7282 static DEFINE_MUTEX(rt_constraints_mutex);
7284 /* Must be called with tasklist_lock held */
7285 static inline int tg_has_rt_tasks(struct task_group *tg)
7287 struct task_struct *g, *p;
7289 do_each_thread(g, p) {
7290 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7292 } while_each_thread(g, p);
7297 struct rt_schedulable_data {
7298 struct task_group *tg;
7303 static int tg_rt_schedulable(struct task_group *tg, void *data)
7305 struct rt_schedulable_data *d = data;
7306 struct task_group *child;
7307 unsigned long total, sum = 0;
7308 u64 period, runtime;
7310 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7311 runtime = tg->rt_bandwidth.rt_runtime;
7314 period = d->rt_period;
7315 runtime = d->rt_runtime;
7319 * Cannot have more runtime than the period.
7321 if (runtime > period && runtime != RUNTIME_INF)
7325 * Ensure we don't starve existing RT tasks.
7327 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7330 total = to_ratio(period, runtime);
7333 * Nobody can have more than the global setting allows.
7335 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7339 * The sum of our children's runtime should not exceed our own.
7341 list_for_each_entry_rcu(child, &tg->children, siblings) {
7342 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7343 runtime = child->rt_bandwidth.rt_runtime;
7345 if (child == d->tg) {
7346 period = d->rt_period;
7347 runtime = d->rt_runtime;
7350 sum += to_ratio(period, runtime);
7359 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7363 struct rt_schedulable_data data = {
7365 .rt_period = period,
7366 .rt_runtime = runtime,
7370 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7376 static int tg_set_rt_bandwidth(struct task_group *tg,
7377 u64 rt_period, u64 rt_runtime)
7381 mutex_lock(&rt_constraints_mutex);
7382 read_lock(&tasklist_lock);
7383 err = __rt_schedulable(tg, rt_period, rt_runtime);
7387 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7388 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7389 tg->rt_bandwidth.rt_runtime = rt_runtime;
7391 for_each_possible_cpu(i) {
7392 struct rt_rq *rt_rq = tg->rt_rq[i];
7394 raw_spin_lock(&rt_rq->rt_runtime_lock);
7395 rt_rq->rt_runtime = rt_runtime;
7396 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7398 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7400 read_unlock(&tasklist_lock);
7401 mutex_unlock(&rt_constraints_mutex);
7406 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7408 u64 rt_runtime, rt_period;
7410 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7411 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7412 if (rt_runtime_us < 0)
7413 rt_runtime = RUNTIME_INF;
7415 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7418 static long sched_group_rt_runtime(struct task_group *tg)
7422 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7425 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7426 do_div(rt_runtime_us, NSEC_PER_USEC);
7427 return rt_runtime_us;
7430 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7432 u64 rt_runtime, rt_period;
7434 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7435 rt_runtime = tg->rt_bandwidth.rt_runtime;
7440 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7443 static long sched_group_rt_period(struct task_group *tg)
7447 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7448 do_div(rt_period_us, NSEC_PER_USEC);
7449 return rt_period_us;
7451 #endif /* CONFIG_RT_GROUP_SCHED */
7453 #ifdef CONFIG_RT_GROUP_SCHED
7454 static int sched_rt_global_constraints(void)
7458 mutex_lock(&rt_constraints_mutex);
7459 read_lock(&tasklist_lock);
7460 ret = __rt_schedulable(NULL, 0, 0);
7461 read_unlock(&tasklist_lock);
7462 mutex_unlock(&rt_constraints_mutex);
7467 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7469 /* Don't accept realtime tasks when there is no way for them to run */
7470 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7476 #else /* !CONFIG_RT_GROUP_SCHED */
7477 static int sched_rt_global_constraints(void)
7479 unsigned long flags;
7482 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7483 for_each_possible_cpu(i) {
7484 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7486 raw_spin_lock(&rt_rq->rt_runtime_lock);
7487 rt_rq->rt_runtime = global_rt_runtime();
7488 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7490 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7494 #endif /* CONFIG_RT_GROUP_SCHED */
7496 static int sched_dl_global_constraints(void)
7498 u64 runtime = global_rt_runtime();
7499 u64 period = global_rt_period();
7500 u64 new_bw = to_ratio(period, runtime);
7502 unsigned long flags;
7505 * Here we want to check the bandwidth not being set to some
7506 * value smaller than the currently allocated bandwidth in
7507 * any of the root_domains.
7509 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7510 * cycling on root_domains... Discussion on different/better
7511 * solutions is welcome!
7513 for_each_possible_cpu(cpu) {
7514 struct dl_bw *dl_b = dl_bw_of(cpu);
7516 raw_spin_lock_irqsave(&dl_b->lock, flags);
7517 if (new_bw < dl_b->total_bw)
7519 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7528 static void sched_dl_do_global(void)
7532 unsigned long flags;
7534 def_dl_bandwidth.dl_period = global_rt_period();
7535 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7537 if (global_rt_runtime() != RUNTIME_INF)
7538 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7541 * FIXME: As above...
7543 for_each_possible_cpu(cpu) {
7544 struct dl_bw *dl_b = dl_bw_of(cpu);
7546 raw_spin_lock_irqsave(&dl_b->lock, flags);
7548 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7552 static int sched_rt_global_validate(void)
7554 if (sysctl_sched_rt_period <= 0)
7557 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7558 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7564 static void sched_rt_do_global(void)
7566 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7567 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7570 int sched_rt_handler(struct ctl_table *table, int write,
7571 void __user *buffer, size_t *lenp,
7574 int old_period, old_runtime;
7575 static DEFINE_MUTEX(mutex);
7579 old_period = sysctl_sched_rt_period;
7580 old_runtime = sysctl_sched_rt_runtime;
7582 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7584 if (!ret && write) {
7585 ret = sched_rt_global_validate();
7589 ret = sched_rt_global_constraints();
7593 ret = sched_dl_global_constraints();
7597 sched_rt_do_global();
7598 sched_dl_do_global();
7602 sysctl_sched_rt_period = old_period;
7603 sysctl_sched_rt_runtime = old_runtime;
7605 mutex_unlock(&mutex);
7610 int sched_rr_handler(struct ctl_table *table, int write,
7611 void __user *buffer, size_t *lenp,
7615 static DEFINE_MUTEX(mutex);
7618 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7619 /* make sure that internally we keep jiffies */
7620 /* also, writing zero resets timeslice to default */
7621 if (!ret && write) {
7622 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7623 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7625 mutex_unlock(&mutex);
7629 #ifdef CONFIG_CGROUP_SCHED
7631 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7633 return css ? container_of(css, struct task_group, css) : NULL;
7636 static struct cgroup_subsys_state *
7637 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7639 struct task_group *parent = css_tg(parent_css);
7640 struct task_group *tg;
7643 /* This is early initialization for the top cgroup */
7644 return &root_task_group.css;
7647 tg = sched_create_group(parent);
7649 return ERR_PTR(-ENOMEM);
7654 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7656 struct task_group *tg = css_tg(css);
7657 struct task_group *parent = css_tg(css->parent);
7660 sched_online_group(tg, parent);
7664 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7666 struct task_group *tg = css_tg(css);
7668 sched_destroy_group(tg);
7671 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7673 struct task_group *tg = css_tg(css);
7675 sched_offline_group(tg);
7678 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7679 struct cgroup_taskset *tset)
7681 struct task_struct *task;
7683 cgroup_taskset_for_each(task, tset) {
7684 #ifdef CONFIG_RT_GROUP_SCHED
7685 if (!sched_rt_can_attach(css_tg(css), task))
7688 /* We don't support RT-tasks being in separate groups */
7689 if (task->sched_class != &fair_sched_class)
7696 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7697 struct cgroup_taskset *tset)
7699 struct task_struct *task;
7701 cgroup_taskset_for_each(task, tset)
7702 sched_move_task(task);
7705 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7706 struct cgroup_subsys_state *old_css,
7707 struct task_struct *task)
7710 * cgroup_exit() is called in the copy_process() failure path.
7711 * Ignore this case since the task hasn't ran yet, this avoids
7712 * trying to poke a half freed task state from generic code.
7714 if (!(task->flags & PF_EXITING))
7717 sched_move_task(task);
7720 #ifdef CONFIG_FAIR_GROUP_SCHED
7721 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7722 struct cftype *cftype, u64 shareval)
7724 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7727 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7730 struct task_group *tg = css_tg(css);
7732 return (u64) scale_load_down(tg->shares);
7735 #ifdef CONFIG_CFS_BANDWIDTH
7736 static DEFINE_MUTEX(cfs_constraints_mutex);
7738 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7739 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7741 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7743 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7745 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7746 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7748 if (tg == &root_task_group)
7752 * Ensure we have at some amount of bandwidth every period. This is
7753 * to prevent reaching a state of large arrears when throttled via
7754 * entity_tick() resulting in prolonged exit starvation.
7756 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7760 * Likewise, bound things on the otherside by preventing insane quota
7761 * periods. This also allows us to normalize in computing quota
7764 if (period > max_cfs_quota_period)
7767 mutex_lock(&cfs_constraints_mutex);
7768 ret = __cfs_schedulable(tg, period, quota);
7772 runtime_enabled = quota != RUNTIME_INF;
7773 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7775 * If we need to toggle cfs_bandwidth_used, off->on must occur
7776 * before making related changes, and on->off must occur afterwards
7778 if (runtime_enabled && !runtime_was_enabled)
7779 cfs_bandwidth_usage_inc();
7780 raw_spin_lock_irq(&cfs_b->lock);
7781 cfs_b->period = ns_to_ktime(period);
7782 cfs_b->quota = quota;
7784 __refill_cfs_bandwidth_runtime(cfs_b);
7785 /* restart the period timer (if active) to handle new period expiry */
7786 if (runtime_enabled && cfs_b->timer_active) {
7787 /* force a reprogram */
7788 cfs_b->timer_active = 0;
7789 __start_cfs_bandwidth(cfs_b);
7791 raw_spin_unlock_irq(&cfs_b->lock);
7793 for_each_possible_cpu(i) {
7794 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7795 struct rq *rq = cfs_rq->rq;
7797 raw_spin_lock_irq(&rq->lock);
7798 cfs_rq->runtime_enabled = runtime_enabled;
7799 cfs_rq->runtime_remaining = 0;
7801 if (cfs_rq->throttled)
7802 unthrottle_cfs_rq(cfs_rq);
7803 raw_spin_unlock_irq(&rq->lock);
7805 if (runtime_was_enabled && !runtime_enabled)
7806 cfs_bandwidth_usage_dec();
7808 mutex_unlock(&cfs_constraints_mutex);
7813 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7817 period = ktime_to_ns(tg->cfs_bandwidth.period);
7818 if (cfs_quota_us < 0)
7819 quota = RUNTIME_INF;
7821 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7823 return tg_set_cfs_bandwidth(tg, period, quota);
7826 long tg_get_cfs_quota(struct task_group *tg)
7830 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7833 quota_us = tg->cfs_bandwidth.quota;
7834 do_div(quota_us, NSEC_PER_USEC);
7839 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7843 period = (u64)cfs_period_us * NSEC_PER_USEC;
7844 quota = tg->cfs_bandwidth.quota;
7846 return tg_set_cfs_bandwidth(tg, period, quota);
7849 long tg_get_cfs_period(struct task_group *tg)
7853 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7854 do_div(cfs_period_us, NSEC_PER_USEC);
7856 return cfs_period_us;
7859 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7862 return tg_get_cfs_quota(css_tg(css));
7865 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7866 struct cftype *cftype, s64 cfs_quota_us)
7868 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7871 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7874 return tg_get_cfs_period(css_tg(css));
7877 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7878 struct cftype *cftype, u64 cfs_period_us)
7880 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7883 struct cfs_schedulable_data {
7884 struct task_group *tg;
7889 * normalize group quota/period to be quota/max_period
7890 * note: units are usecs
7892 static u64 normalize_cfs_quota(struct task_group *tg,
7893 struct cfs_schedulable_data *d)
7901 period = tg_get_cfs_period(tg);
7902 quota = tg_get_cfs_quota(tg);
7905 /* note: these should typically be equivalent */
7906 if (quota == RUNTIME_INF || quota == -1)
7909 return to_ratio(period, quota);
7912 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7914 struct cfs_schedulable_data *d = data;
7915 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7916 s64 quota = 0, parent_quota = -1;
7919 quota = RUNTIME_INF;
7921 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7923 quota = normalize_cfs_quota(tg, d);
7924 parent_quota = parent_b->hierarchal_quota;
7927 * ensure max(child_quota) <= parent_quota, inherit when no
7930 if (quota == RUNTIME_INF)
7931 quota = parent_quota;
7932 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7935 cfs_b->hierarchal_quota = quota;
7940 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7943 struct cfs_schedulable_data data = {
7949 if (quota != RUNTIME_INF) {
7950 do_div(data.period, NSEC_PER_USEC);
7951 do_div(data.quota, NSEC_PER_USEC);
7955 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7961 static int cpu_stats_show(struct seq_file *sf, void *v)
7963 struct task_group *tg = css_tg(seq_css(sf));
7964 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7966 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7967 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7968 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7972 #endif /* CONFIG_CFS_BANDWIDTH */
7973 #endif /* CONFIG_FAIR_GROUP_SCHED */
7975 #ifdef CONFIG_RT_GROUP_SCHED
7976 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7977 struct cftype *cft, s64 val)
7979 return sched_group_set_rt_runtime(css_tg(css), val);
7982 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7985 return sched_group_rt_runtime(css_tg(css));
7988 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7989 struct cftype *cftype, u64 rt_period_us)
7991 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7994 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7997 return sched_group_rt_period(css_tg(css));
7999 #endif /* CONFIG_RT_GROUP_SCHED */
8001 static struct cftype cpu_files[] = {
8002 #ifdef CONFIG_FAIR_GROUP_SCHED
8005 .read_u64 = cpu_shares_read_u64,
8006 .write_u64 = cpu_shares_write_u64,
8009 #ifdef CONFIG_CFS_BANDWIDTH
8011 .name = "cfs_quota_us",
8012 .read_s64 = cpu_cfs_quota_read_s64,
8013 .write_s64 = cpu_cfs_quota_write_s64,
8016 .name = "cfs_period_us",
8017 .read_u64 = cpu_cfs_period_read_u64,
8018 .write_u64 = cpu_cfs_period_write_u64,
8022 .seq_show = cpu_stats_show,
8025 #ifdef CONFIG_RT_GROUP_SCHED
8027 .name = "rt_runtime_us",
8028 .read_s64 = cpu_rt_runtime_read,
8029 .write_s64 = cpu_rt_runtime_write,
8032 .name = "rt_period_us",
8033 .read_u64 = cpu_rt_period_read_uint,
8034 .write_u64 = cpu_rt_period_write_uint,
8040 struct cgroup_subsys cpu_cgrp_subsys = {
8041 .css_alloc = cpu_cgroup_css_alloc,
8042 .css_free = cpu_cgroup_css_free,
8043 .css_online = cpu_cgroup_css_online,
8044 .css_offline = cpu_cgroup_css_offline,
8045 .can_attach = cpu_cgroup_can_attach,
8046 .attach = cpu_cgroup_attach,
8047 .exit = cpu_cgroup_exit,
8048 .base_cftypes = cpu_files,
8052 #endif /* CONFIG_CGROUP_SCHED */
8054 void dump_cpu_task(int cpu)
8056 pr_info("Task dump for CPU %d:\n", cpu);
8057 sched_show_task(cpu_curr(cpu));