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;
145 update_rq_clock_task(rq, delta);
149 * Debugging: various feature bits
152 #define SCHED_FEAT(name, enabled) \
153 (1UL << __SCHED_FEAT_##name) * enabled |
155 const_debug unsigned int sysctl_sched_features =
156 #include "features.h"
161 #ifdef CONFIG_SCHED_DEBUG
162 #define SCHED_FEAT(name, enabled) \
165 static const char * const sched_feat_names[] = {
166 #include "features.h"
171 static int sched_feat_show(struct seq_file *m, void *v)
175 for (i = 0; i < __SCHED_FEAT_NR; i++) {
176 if (!(sysctl_sched_features & (1UL << i)))
178 seq_printf(m, "%s ", sched_feat_names[i]);
185 #ifdef HAVE_JUMP_LABEL
187 #define jump_label_key__true STATIC_KEY_INIT_TRUE
188 #define jump_label_key__false STATIC_KEY_INIT_FALSE
190 #define SCHED_FEAT(name, enabled) \
191 jump_label_key__##enabled ,
193 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
194 #include "features.h"
199 static void sched_feat_disable(int i)
201 if (static_key_enabled(&sched_feat_keys[i]))
202 static_key_slow_dec(&sched_feat_keys[i]);
205 static void sched_feat_enable(int i)
207 if (!static_key_enabled(&sched_feat_keys[i]))
208 static_key_slow_inc(&sched_feat_keys[i]);
211 static void sched_feat_disable(int i) { };
212 static void sched_feat_enable(int i) { };
213 #endif /* HAVE_JUMP_LABEL */
215 static int sched_feat_set(char *cmp)
220 if (strncmp(cmp, "NO_", 3) == 0) {
225 for (i = 0; i < __SCHED_FEAT_NR; i++) {
226 if (strcmp(cmp, sched_feat_names[i]) == 0) {
228 sysctl_sched_features &= ~(1UL << i);
229 sched_feat_disable(i);
231 sysctl_sched_features |= (1UL << i);
232 sched_feat_enable(i);
242 sched_feat_write(struct file *filp, const char __user *ubuf,
243 size_t cnt, loff_t *ppos)
253 if (copy_from_user(&buf, ubuf, cnt))
259 /* Ensure the static_key remains in a consistent state */
260 inode = file_inode(filp);
261 mutex_lock(&inode->i_mutex);
262 i = sched_feat_set(cmp);
263 mutex_unlock(&inode->i_mutex);
264 if (i == __SCHED_FEAT_NR)
272 static int sched_feat_open(struct inode *inode, struct file *filp)
274 return single_open(filp, sched_feat_show, NULL);
277 static const struct file_operations sched_feat_fops = {
278 .open = sched_feat_open,
279 .write = sched_feat_write,
282 .release = single_release,
285 static __init int sched_init_debug(void)
287 debugfs_create_file("sched_features", 0644, NULL, NULL,
292 late_initcall(sched_init_debug);
293 #endif /* CONFIG_SCHED_DEBUG */
296 * Number of tasks to iterate in a single balance run.
297 * Limited because this is done with IRQs disabled.
299 const_debug unsigned int sysctl_sched_nr_migrate = 32;
302 * period over which we average the RT time consumption, measured
307 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
310 * period over which we measure -rt task cpu usage in us.
313 unsigned int sysctl_sched_rt_period = 1000000;
315 __read_mostly int scheduler_running;
318 * part of the period that we allow rt tasks to run in us.
321 int sysctl_sched_rt_runtime = 950000;
324 * __task_rq_lock - lock the rq @p resides on.
326 static inline struct rq *__task_rq_lock(struct task_struct *p)
331 lockdep_assert_held(&p->pi_lock);
335 raw_spin_lock(&rq->lock);
336 if (likely(rq == task_rq(p)))
338 raw_spin_unlock(&rq->lock);
343 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
345 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
346 __acquires(p->pi_lock)
352 raw_spin_lock_irqsave(&p->pi_lock, *flags);
354 raw_spin_lock(&rq->lock);
355 if (likely(rq == task_rq(p)))
357 raw_spin_unlock(&rq->lock);
358 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
362 static void __task_rq_unlock(struct rq *rq)
365 raw_spin_unlock(&rq->lock);
369 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
371 __releases(p->pi_lock)
373 raw_spin_unlock(&rq->lock);
374 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
378 * this_rq_lock - lock this runqueue and disable interrupts.
380 static struct rq *this_rq_lock(void)
387 raw_spin_lock(&rq->lock);
392 #ifdef CONFIG_SCHED_HRTICK
394 * Use HR-timers to deliver accurate preemption points.
397 static void hrtick_clear(struct rq *rq)
399 if (hrtimer_active(&rq->hrtick_timer))
400 hrtimer_cancel(&rq->hrtick_timer);
404 * High-resolution timer tick.
405 * Runs from hardirq context with interrupts disabled.
407 static enum hrtimer_restart hrtick(struct hrtimer *timer)
409 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
411 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
413 raw_spin_lock(&rq->lock);
415 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
416 raw_spin_unlock(&rq->lock);
418 return HRTIMER_NORESTART;
423 static int __hrtick_restart(struct rq *rq)
425 struct hrtimer *timer = &rq->hrtick_timer;
426 ktime_t time = hrtimer_get_softexpires(timer);
428 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
432 * called from hardirq (IPI) context
434 static void __hrtick_start(void *arg)
438 raw_spin_lock(&rq->lock);
439 __hrtick_restart(rq);
440 rq->hrtick_csd_pending = 0;
441 raw_spin_unlock(&rq->lock);
445 * Called to set the hrtick timer state.
447 * called with rq->lock held and irqs disabled
449 void hrtick_start(struct rq *rq, u64 delay)
451 struct hrtimer *timer = &rq->hrtick_timer;
452 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
454 hrtimer_set_expires(timer, time);
456 if (rq == this_rq()) {
457 __hrtick_restart(rq);
458 } else if (!rq->hrtick_csd_pending) {
459 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
460 rq->hrtick_csd_pending = 1;
465 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
467 int cpu = (int)(long)hcpu;
470 case CPU_UP_CANCELED:
471 case CPU_UP_CANCELED_FROZEN:
472 case CPU_DOWN_PREPARE:
473 case CPU_DOWN_PREPARE_FROZEN:
475 case CPU_DEAD_FROZEN:
476 hrtick_clear(cpu_rq(cpu));
483 static __init void init_hrtick(void)
485 hotcpu_notifier(hotplug_hrtick, 0);
489 * Called to set the hrtick timer state.
491 * called with rq->lock held and irqs disabled
493 void hrtick_start(struct rq *rq, u64 delay)
495 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
496 HRTIMER_MODE_REL_PINNED, 0);
499 static inline void init_hrtick(void)
502 #endif /* CONFIG_SMP */
504 static void init_rq_hrtick(struct rq *rq)
507 rq->hrtick_csd_pending = 0;
509 rq->hrtick_csd.flags = 0;
510 rq->hrtick_csd.func = __hrtick_start;
511 rq->hrtick_csd.info = rq;
514 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
515 rq->hrtick_timer.function = hrtick;
517 #else /* CONFIG_SCHED_HRTICK */
518 static inline void hrtick_clear(struct rq *rq)
522 static inline void init_rq_hrtick(struct rq *rq)
526 static inline void init_hrtick(void)
529 #endif /* CONFIG_SCHED_HRTICK */
532 * cmpxchg based fetch_or, macro so it works for different integer types
534 #define fetch_or(ptr, val) \
535 ({ typeof(*(ptr)) __old, __val = *(ptr); \
537 __old = cmpxchg((ptr), __val, __val | (val)); \
538 if (__old == __val) \
545 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
547 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
548 * this avoids any races wrt polling state changes and thereby avoids
551 static bool set_nr_and_not_polling(struct task_struct *p)
553 struct thread_info *ti = task_thread_info(p);
554 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
558 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
560 * If this returns true, then the idle task promises to call
561 * sched_ttwu_pending() and reschedule soon.
563 static bool set_nr_if_polling(struct task_struct *p)
565 struct thread_info *ti = task_thread_info(p);
566 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
569 if (!(val & _TIF_POLLING_NRFLAG))
571 if (val & _TIF_NEED_RESCHED)
573 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
582 static bool set_nr_and_not_polling(struct task_struct *p)
584 set_tsk_need_resched(p);
589 static bool set_nr_if_polling(struct task_struct *p)
597 * resched_curr - mark rq's current task 'to be rescheduled now'.
599 * On UP this means the setting of the need_resched flag, on SMP it
600 * might also involve a cross-CPU call to trigger the scheduler on
603 void resched_curr(struct rq *rq)
605 struct task_struct *curr = rq->curr;
608 lockdep_assert_held(&rq->lock);
610 if (test_tsk_need_resched(curr))
615 if (cpu == smp_processor_id()) {
616 set_tsk_need_resched(curr);
617 set_preempt_need_resched();
621 if (set_nr_and_not_polling(curr))
622 smp_send_reschedule(cpu);
624 trace_sched_wake_idle_without_ipi(cpu);
627 void resched_cpu(int cpu)
629 struct rq *rq = cpu_rq(cpu);
632 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
635 raw_spin_unlock_irqrestore(&rq->lock, flags);
639 #ifdef CONFIG_NO_HZ_COMMON
641 * In the semi idle case, use the nearest busy cpu for migrating timers
642 * from an idle cpu. This is good for power-savings.
644 * We don't do similar optimization for completely idle system, as
645 * selecting an idle cpu will add more delays to the timers than intended
646 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
648 int get_nohz_timer_target(int pinned)
650 int cpu = smp_processor_id();
652 struct sched_domain *sd;
654 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
658 for_each_domain(cpu, sd) {
659 for_each_cpu(i, sched_domain_span(sd)) {
671 * When add_timer_on() enqueues a timer into the timer wheel of an
672 * idle CPU then this timer might expire before the next timer event
673 * which is scheduled to wake up that CPU. In case of a completely
674 * idle system the next event might even be infinite time into the
675 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
676 * leaves the inner idle loop so the newly added timer is taken into
677 * account when the CPU goes back to idle and evaluates the timer
678 * wheel for the next timer event.
680 static void wake_up_idle_cpu(int cpu)
682 struct rq *rq = cpu_rq(cpu);
684 if (cpu == smp_processor_id())
687 if (set_nr_and_not_polling(rq->idle))
688 smp_send_reschedule(cpu);
690 trace_sched_wake_idle_without_ipi(cpu);
693 static bool wake_up_full_nohz_cpu(int cpu)
696 * We just need the target to call irq_exit() and re-evaluate
697 * the next tick. The nohz full kick at least implies that.
698 * If needed we can still optimize that later with an
701 if (tick_nohz_full_cpu(cpu)) {
702 if (cpu != smp_processor_id() ||
703 tick_nohz_tick_stopped())
704 tick_nohz_full_kick_cpu(cpu);
711 void wake_up_nohz_cpu(int cpu)
713 if (!wake_up_full_nohz_cpu(cpu))
714 wake_up_idle_cpu(cpu);
717 static inline bool got_nohz_idle_kick(void)
719 int cpu = smp_processor_id();
721 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
724 if (idle_cpu(cpu) && !need_resched())
728 * We can't run Idle Load Balance on this CPU for this time so we
729 * cancel it and clear NOHZ_BALANCE_KICK
731 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
735 #else /* CONFIG_NO_HZ_COMMON */
737 static inline bool got_nohz_idle_kick(void)
742 #endif /* CONFIG_NO_HZ_COMMON */
744 #ifdef CONFIG_NO_HZ_FULL
745 bool sched_can_stop_tick(void)
748 * More than one running task need preemption.
749 * nr_running update is assumed to be visible
750 * after IPI is sent from wakers.
752 if (this_rq()->nr_running > 1)
757 #endif /* CONFIG_NO_HZ_FULL */
759 void sched_avg_update(struct rq *rq)
761 s64 period = sched_avg_period();
763 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
765 * Inline assembly required to prevent the compiler
766 * optimising this loop into a divmod call.
767 * See __iter_div_u64_rem() for another example of this.
769 asm("" : "+rm" (rq->age_stamp));
770 rq->age_stamp += period;
775 #endif /* CONFIG_SMP */
777 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
778 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
780 * Iterate task_group tree rooted at *from, calling @down when first entering a
781 * node and @up when leaving it for the final time.
783 * Caller must hold rcu_lock or sufficient equivalent.
785 int walk_tg_tree_from(struct task_group *from,
786 tg_visitor down, tg_visitor up, void *data)
788 struct task_group *parent, *child;
794 ret = (*down)(parent, data);
797 list_for_each_entry_rcu(child, &parent->children, siblings) {
804 ret = (*up)(parent, data);
805 if (ret || parent == from)
809 parent = parent->parent;
816 int tg_nop(struct task_group *tg, void *data)
822 static void set_load_weight(struct task_struct *p)
824 int prio = p->static_prio - MAX_RT_PRIO;
825 struct load_weight *load = &p->se.load;
828 * SCHED_IDLE tasks get minimal weight:
830 if (p->policy == SCHED_IDLE) {
831 load->weight = scale_load(WEIGHT_IDLEPRIO);
832 load->inv_weight = WMULT_IDLEPRIO;
836 load->weight = scale_load(prio_to_weight[prio]);
837 load->inv_weight = prio_to_wmult[prio];
840 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
843 sched_info_queued(rq, p);
844 p->sched_class->enqueue_task(rq, p, flags);
847 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
850 sched_info_dequeued(rq, p);
851 p->sched_class->dequeue_task(rq, p, flags);
854 void activate_task(struct rq *rq, struct task_struct *p, int flags)
856 if (task_contributes_to_load(p))
857 rq->nr_uninterruptible--;
859 enqueue_task(rq, p, flags);
862 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
864 if (task_contributes_to_load(p))
865 rq->nr_uninterruptible++;
867 dequeue_task(rq, p, flags);
870 static void update_rq_clock_task(struct rq *rq, s64 delta)
873 * In theory, the compile should just see 0 here, and optimize out the call
874 * to sched_rt_avg_update. But I don't trust it...
876 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
877 s64 steal = 0, irq_delta = 0;
879 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
880 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
883 * Since irq_time is only updated on {soft,}irq_exit, we might run into
884 * this case when a previous update_rq_clock() happened inside a
887 * When this happens, we stop ->clock_task and only update the
888 * prev_irq_time stamp to account for the part that fit, so that a next
889 * update will consume the rest. This ensures ->clock_task is
892 * It does however cause some slight miss-attribution of {soft,}irq
893 * time, a more accurate solution would be to update the irq_time using
894 * the current rq->clock timestamp, except that would require using
897 if (irq_delta > delta)
900 rq->prev_irq_time += irq_delta;
903 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
904 if (static_key_false((¶virt_steal_rq_enabled))) {
905 steal = paravirt_steal_clock(cpu_of(rq));
906 steal -= rq->prev_steal_time_rq;
908 if (unlikely(steal > delta))
911 rq->prev_steal_time_rq += steal;
916 rq->clock_task += delta;
918 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
919 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
920 sched_rt_avg_update(rq, irq_delta + steal);
924 void sched_set_stop_task(int cpu, struct task_struct *stop)
926 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
927 struct task_struct *old_stop = cpu_rq(cpu)->stop;
931 * Make it appear like a SCHED_FIFO task, its something
932 * userspace knows about and won't get confused about.
934 * Also, it will make PI more or less work without too
935 * much confusion -- but then, stop work should not
936 * rely on PI working anyway.
938 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
940 stop->sched_class = &stop_sched_class;
943 cpu_rq(cpu)->stop = stop;
947 * Reset it back to a normal scheduling class so that
948 * it can die in pieces.
950 old_stop->sched_class = &rt_sched_class;
955 * __normal_prio - return the priority that is based on the static prio
957 static inline int __normal_prio(struct task_struct *p)
959 return p->static_prio;
963 * Calculate the expected normal priority: i.e. priority
964 * without taking RT-inheritance into account. Might be
965 * boosted by interactivity modifiers. Changes upon fork,
966 * setprio syscalls, and whenever the interactivity
967 * estimator recalculates.
969 static inline int normal_prio(struct task_struct *p)
973 if (task_has_dl_policy(p))
974 prio = MAX_DL_PRIO-1;
975 else if (task_has_rt_policy(p))
976 prio = MAX_RT_PRIO-1 - p->rt_priority;
978 prio = __normal_prio(p);
983 * Calculate the current priority, i.e. the priority
984 * taken into account by the scheduler. This value might
985 * be boosted by RT tasks, or might be boosted by
986 * interactivity modifiers. Will be RT if the task got
987 * RT-boosted. If not then it returns p->normal_prio.
989 static int effective_prio(struct task_struct *p)
991 p->normal_prio = normal_prio(p);
993 * If we are RT tasks or we were boosted to RT priority,
994 * keep the priority unchanged. Otherwise, update priority
995 * to the normal priority:
997 if (!rt_prio(p->prio))
998 return p->normal_prio;
1003 * task_curr - is this task currently executing on a CPU?
1004 * @p: the task in question.
1006 * Return: 1 if the task is currently executing. 0 otherwise.
1008 inline int task_curr(const struct task_struct *p)
1010 return cpu_curr(task_cpu(p)) == p;
1013 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1014 const struct sched_class *prev_class,
1017 if (prev_class != p->sched_class) {
1018 if (prev_class->switched_from)
1019 prev_class->switched_from(rq, p);
1020 p->sched_class->switched_to(rq, p);
1021 } else if (oldprio != p->prio || dl_task(p))
1022 p->sched_class->prio_changed(rq, p, oldprio);
1025 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1027 const struct sched_class *class;
1029 if (p->sched_class == rq->curr->sched_class) {
1030 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1032 for_each_class(class) {
1033 if (class == rq->curr->sched_class)
1035 if (class == p->sched_class) {
1043 * A queue event has occurred, and we're going to schedule. In
1044 * this case, we can save a useless back to back clock update.
1046 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1047 rq->skip_clock_update = 1;
1051 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1053 #ifdef CONFIG_SCHED_DEBUG
1055 * We should never call set_task_cpu() on a blocked task,
1056 * ttwu() will sort out the placement.
1058 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1059 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1061 #ifdef CONFIG_LOCKDEP
1063 * The caller should hold either p->pi_lock or rq->lock, when changing
1064 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1066 * sched_move_task() holds both and thus holding either pins the cgroup,
1069 * Furthermore, all task_rq users should acquire both locks, see
1072 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1073 lockdep_is_held(&task_rq(p)->lock)));
1077 trace_sched_migrate_task(p, new_cpu);
1079 if (task_cpu(p) != new_cpu) {
1080 if (p->sched_class->migrate_task_rq)
1081 p->sched_class->migrate_task_rq(p, new_cpu);
1082 p->se.nr_migrations++;
1083 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1086 __set_task_cpu(p, new_cpu);
1089 static void __migrate_swap_task(struct task_struct *p, int cpu)
1092 struct rq *src_rq, *dst_rq;
1094 src_rq = task_rq(p);
1095 dst_rq = cpu_rq(cpu);
1097 deactivate_task(src_rq, p, 0);
1098 set_task_cpu(p, cpu);
1099 activate_task(dst_rq, p, 0);
1100 check_preempt_curr(dst_rq, p, 0);
1103 * Task isn't running anymore; make it appear like we migrated
1104 * it before it went to sleep. This means on wakeup we make the
1105 * previous cpu our targer instead of where it really is.
1111 struct migration_swap_arg {
1112 struct task_struct *src_task, *dst_task;
1113 int src_cpu, dst_cpu;
1116 static int migrate_swap_stop(void *data)
1118 struct migration_swap_arg *arg = data;
1119 struct rq *src_rq, *dst_rq;
1122 src_rq = cpu_rq(arg->src_cpu);
1123 dst_rq = cpu_rq(arg->dst_cpu);
1125 double_raw_lock(&arg->src_task->pi_lock,
1126 &arg->dst_task->pi_lock);
1127 double_rq_lock(src_rq, dst_rq);
1128 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1131 if (task_cpu(arg->src_task) != arg->src_cpu)
1134 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1137 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1140 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1141 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1146 double_rq_unlock(src_rq, dst_rq);
1147 raw_spin_unlock(&arg->dst_task->pi_lock);
1148 raw_spin_unlock(&arg->src_task->pi_lock);
1154 * Cross migrate two tasks
1156 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1158 struct migration_swap_arg arg;
1161 arg = (struct migration_swap_arg){
1163 .src_cpu = task_cpu(cur),
1165 .dst_cpu = task_cpu(p),
1168 if (arg.src_cpu == arg.dst_cpu)
1172 * These three tests are all lockless; this is OK since all of them
1173 * will be re-checked with proper locks held further down the line.
1175 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1178 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1181 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1184 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1185 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1191 struct migration_arg {
1192 struct task_struct *task;
1196 static int migration_cpu_stop(void *data);
1199 * wait_task_inactive - wait for a thread to unschedule.
1201 * If @match_state is nonzero, it's the @p->state value just checked and
1202 * not expected to change. If it changes, i.e. @p might have woken up,
1203 * then return zero. When we succeed in waiting for @p to be off its CPU,
1204 * we return a positive number (its total switch count). If a second call
1205 * a short while later returns the same number, the caller can be sure that
1206 * @p has remained unscheduled the whole time.
1208 * The caller must ensure that the task *will* unschedule sometime soon,
1209 * else this function might spin for a *long* time. This function can't
1210 * be called with interrupts off, or it may introduce deadlock with
1211 * smp_call_function() if an IPI is sent by the same process we are
1212 * waiting to become inactive.
1214 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1216 unsigned long flags;
1223 * We do the initial early heuristics without holding
1224 * any task-queue locks at all. We'll only try to get
1225 * the runqueue lock when things look like they will
1231 * If the task is actively running on another CPU
1232 * still, just relax and busy-wait without holding
1235 * NOTE! Since we don't hold any locks, it's not
1236 * even sure that "rq" stays as the right runqueue!
1237 * But we don't care, since "task_running()" will
1238 * return false if the runqueue has changed and p
1239 * is actually now running somewhere else!
1241 while (task_running(rq, p)) {
1242 if (match_state && unlikely(p->state != match_state))
1248 * Ok, time to look more closely! We need the rq
1249 * lock now, to be *sure*. If we're wrong, we'll
1250 * just go back and repeat.
1252 rq = task_rq_lock(p, &flags);
1253 trace_sched_wait_task(p);
1254 running = task_running(rq, p);
1257 if (!match_state || p->state == match_state)
1258 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1259 task_rq_unlock(rq, p, &flags);
1262 * If it changed from the expected state, bail out now.
1264 if (unlikely(!ncsw))
1268 * Was it really running after all now that we
1269 * checked with the proper locks actually held?
1271 * Oops. Go back and try again..
1273 if (unlikely(running)) {
1279 * It's not enough that it's not actively running,
1280 * it must be off the runqueue _entirely_, and not
1283 * So if it was still runnable (but just not actively
1284 * running right now), it's preempted, and we should
1285 * yield - it could be a while.
1287 if (unlikely(on_rq)) {
1288 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1290 set_current_state(TASK_UNINTERRUPTIBLE);
1291 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1296 * Ahh, all good. It wasn't running, and it wasn't
1297 * runnable, which means that it will never become
1298 * running in the future either. We're all done!
1307 * kick_process - kick a running thread to enter/exit the kernel
1308 * @p: the to-be-kicked thread
1310 * Cause a process which is running on another CPU to enter
1311 * kernel-mode, without any delay. (to get signals handled.)
1313 * NOTE: this function doesn't have to take the runqueue lock,
1314 * because all it wants to ensure is that the remote task enters
1315 * the kernel. If the IPI races and the task has been migrated
1316 * to another CPU then no harm is done and the purpose has been
1319 void kick_process(struct task_struct *p)
1325 if ((cpu != smp_processor_id()) && task_curr(p))
1326 smp_send_reschedule(cpu);
1329 EXPORT_SYMBOL_GPL(kick_process);
1330 #endif /* CONFIG_SMP */
1334 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1336 static int select_fallback_rq(int cpu, struct task_struct *p)
1338 int nid = cpu_to_node(cpu);
1339 const struct cpumask *nodemask = NULL;
1340 enum { cpuset, possible, fail } state = cpuset;
1344 * If the node that the cpu is on has been offlined, cpu_to_node()
1345 * will return -1. There is no cpu on the node, and we should
1346 * select the cpu on the other node.
1349 nodemask = cpumask_of_node(nid);
1351 /* Look for allowed, online CPU in same node. */
1352 for_each_cpu(dest_cpu, nodemask) {
1353 if (!cpu_online(dest_cpu))
1355 if (!cpu_active(dest_cpu))
1357 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1363 /* Any allowed, online CPU? */
1364 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1365 if (!cpu_online(dest_cpu))
1367 if (!cpu_active(dest_cpu))
1374 /* No more Mr. Nice Guy. */
1375 cpuset_cpus_allowed_fallback(p);
1380 do_set_cpus_allowed(p, cpu_possible_mask);
1391 if (state != cpuset) {
1393 * Don't tell them about moving exiting tasks or
1394 * kernel threads (both mm NULL), since they never
1397 if (p->mm && printk_ratelimit()) {
1398 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1399 task_pid_nr(p), p->comm, cpu);
1407 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1410 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1412 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1415 * In order not to call set_task_cpu() on a blocking task we need
1416 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1419 * Since this is common to all placement strategies, this lives here.
1421 * [ this allows ->select_task() to simply return task_cpu(p) and
1422 * not worry about this generic constraint ]
1424 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1426 cpu = select_fallback_rq(task_cpu(p), p);
1431 static void update_avg(u64 *avg, u64 sample)
1433 s64 diff = sample - *avg;
1439 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1441 #ifdef CONFIG_SCHEDSTATS
1442 struct rq *rq = this_rq();
1445 int this_cpu = smp_processor_id();
1447 if (cpu == this_cpu) {
1448 schedstat_inc(rq, ttwu_local);
1449 schedstat_inc(p, se.statistics.nr_wakeups_local);
1451 struct sched_domain *sd;
1453 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1455 for_each_domain(this_cpu, sd) {
1456 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1457 schedstat_inc(sd, ttwu_wake_remote);
1464 if (wake_flags & WF_MIGRATED)
1465 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1467 #endif /* CONFIG_SMP */
1469 schedstat_inc(rq, ttwu_count);
1470 schedstat_inc(p, se.statistics.nr_wakeups);
1472 if (wake_flags & WF_SYNC)
1473 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1475 #endif /* CONFIG_SCHEDSTATS */
1478 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1480 activate_task(rq, p, en_flags);
1483 /* if a worker is waking up, notify workqueue */
1484 if (p->flags & PF_WQ_WORKER)
1485 wq_worker_waking_up(p, cpu_of(rq));
1489 * Mark the task runnable and perform wakeup-preemption.
1492 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1494 check_preempt_curr(rq, p, wake_flags);
1495 trace_sched_wakeup(p, true);
1497 p->state = TASK_RUNNING;
1499 if (p->sched_class->task_woken)
1500 p->sched_class->task_woken(rq, p);
1502 if (rq->idle_stamp) {
1503 u64 delta = rq_clock(rq) - rq->idle_stamp;
1504 u64 max = 2*rq->max_idle_balance_cost;
1506 update_avg(&rq->avg_idle, delta);
1508 if (rq->avg_idle > max)
1517 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1520 if (p->sched_contributes_to_load)
1521 rq->nr_uninterruptible--;
1524 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1525 ttwu_do_wakeup(rq, p, wake_flags);
1529 * Called in case the task @p isn't fully descheduled from its runqueue,
1530 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1531 * since all we need to do is flip p->state to TASK_RUNNING, since
1532 * the task is still ->on_rq.
1534 static int ttwu_remote(struct task_struct *p, int wake_flags)
1539 rq = __task_rq_lock(p);
1541 /* check_preempt_curr() may use rq clock */
1542 update_rq_clock(rq);
1543 ttwu_do_wakeup(rq, p, wake_flags);
1546 __task_rq_unlock(rq);
1552 void sched_ttwu_pending(void)
1554 struct rq *rq = this_rq();
1555 struct llist_node *llist = llist_del_all(&rq->wake_list);
1556 struct task_struct *p;
1557 unsigned long flags;
1562 raw_spin_lock_irqsave(&rq->lock, flags);
1565 p = llist_entry(llist, struct task_struct, wake_entry);
1566 llist = llist_next(llist);
1567 ttwu_do_activate(rq, p, 0);
1570 raw_spin_unlock_irqrestore(&rq->lock, flags);
1573 void scheduler_ipi(void)
1576 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1577 * TIF_NEED_RESCHED remotely (for the first time) will also send
1580 preempt_fold_need_resched();
1582 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1586 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1587 * traditionally all their work was done from the interrupt return
1588 * path. Now that we actually do some work, we need to make sure
1591 * Some archs already do call them, luckily irq_enter/exit nest
1594 * Arguably we should visit all archs and update all handlers,
1595 * however a fair share of IPIs are still resched only so this would
1596 * somewhat pessimize the simple resched case.
1599 sched_ttwu_pending();
1602 * Check if someone kicked us for doing the nohz idle load balance.
1604 if (unlikely(got_nohz_idle_kick())) {
1605 this_rq()->idle_balance = 1;
1606 raise_softirq_irqoff(SCHED_SOFTIRQ);
1611 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1613 struct rq *rq = cpu_rq(cpu);
1615 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1616 if (!set_nr_if_polling(rq->idle))
1617 smp_send_reschedule(cpu);
1619 trace_sched_wake_idle_without_ipi(cpu);
1623 bool cpus_share_cache(int this_cpu, int that_cpu)
1625 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1627 #endif /* CONFIG_SMP */
1629 static void ttwu_queue(struct task_struct *p, int cpu)
1631 struct rq *rq = cpu_rq(cpu);
1633 #if defined(CONFIG_SMP)
1634 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1635 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1636 ttwu_queue_remote(p, cpu);
1641 raw_spin_lock(&rq->lock);
1642 ttwu_do_activate(rq, p, 0);
1643 raw_spin_unlock(&rq->lock);
1647 * try_to_wake_up - wake up a thread
1648 * @p: the thread to be awakened
1649 * @state: the mask of task states that can be woken
1650 * @wake_flags: wake modifier flags (WF_*)
1652 * Put it on the run-queue if it's not already there. The "current"
1653 * thread is always on the run-queue (except when the actual
1654 * re-schedule is in progress), and as such you're allowed to do
1655 * the simpler "current->state = TASK_RUNNING" to mark yourself
1656 * runnable without the overhead of this.
1658 * Return: %true if @p was woken up, %false if it was already running.
1659 * or @state didn't match @p's state.
1662 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1664 unsigned long flags;
1665 int cpu, success = 0;
1668 * If we are going to wake up a thread waiting for CONDITION we
1669 * need to ensure that CONDITION=1 done by the caller can not be
1670 * reordered with p->state check below. This pairs with mb() in
1671 * set_current_state() the waiting thread does.
1673 smp_mb__before_spinlock();
1674 raw_spin_lock_irqsave(&p->pi_lock, flags);
1675 if (!(p->state & state))
1678 success = 1; /* we're going to change ->state */
1681 if (p->on_rq && ttwu_remote(p, wake_flags))
1686 * If the owning (remote) cpu is still in the middle of schedule() with
1687 * this task as prev, wait until its done referencing the task.
1692 * Pairs with the smp_wmb() in finish_lock_switch().
1696 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1697 p->state = TASK_WAKING;
1699 if (p->sched_class->task_waking)
1700 p->sched_class->task_waking(p);
1702 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1703 if (task_cpu(p) != cpu) {
1704 wake_flags |= WF_MIGRATED;
1705 set_task_cpu(p, cpu);
1707 #endif /* CONFIG_SMP */
1711 ttwu_stat(p, cpu, wake_flags);
1713 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1719 * try_to_wake_up_local - try to wake up a local task with rq lock held
1720 * @p: the thread to be awakened
1722 * Put @p on the run-queue if it's not already there. The caller must
1723 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1726 static void try_to_wake_up_local(struct task_struct *p)
1728 struct rq *rq = task_rq(p);
1730 if (WARN_ON_ONCE(rq != this_rq()) ||
1731 WARN_ON_ONCE(p == current))
1734 lockdep_assert_held(&rq->lock);
1736 if (!raw_spin_trylock(&p->pi_lock)) {
1737 raw_spin_unlock(&rq->lock);
1738 raw_spin_lock(&p->pi_lock);
1739 raw_spin_lock(&rq->lock);
1742 if (!(p->state & TASK_NORMAL))
1746 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1748 ttwu_do_wakeup(rq, p, 0);
1749 ttwu_stat(p, smp_processor_id(), 0);
1751 raw_spin_unlock(&p->pi_lock);
1755 * wake_up_process - Wake up a specific process
1756 * @p: The process to be woken up.
1758 * Attempt to wake up the nominated process and move it to the set of runnable
1761 * Return: 1 if the process was woken up, 0 if it was already running.
1763 * It may be assumed that this function implies a write memory barrier before
1764 * changing the task state if and only if any tasks are woken up.
1766 int wake_up_process(struct task_struct *p)
1768 WARN_ON(task_is_stopped_or_traced(p));
1769 return try_to_wake_up(p, TASK_NORMAL, 0);
1771 EXPORT_SYMBOL(wake_up_process);
1773 int wake_up_state(struct task_struct *p, unsigned int state)
1775 return try_to_wake_up(p, state, 0);
1779 * Perform scheduler related setup for a newly forked process p.
1780 * p is forked by current.
1782 * __sched_fork() is basic setup used by init_idle() too:
1784 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1789 p->se.exec_start = 0;
1790 p->se.sum_exec_runtime = 0;
1791 p->se.prev_sum_exec_runtime = 0;
1792 p->se.nr_migrations = 0;
1794 INIT_LIST_HEAD(&p->se.group_node);
1796 #ifdef CONFIG_SCHEDSTATS
1797 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1800 RB_CLEAR_NODE(&p->dl.rb_node);
1801 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1802 p->dl.dl_runtime = p->dl.runtime = 0;
1803 p->dl.dl_deadline = p->dl.deadline = 0;
1804 p->dl.dl_period = 0;
1807 INIT_LIST_HEAD(&p->rt.run_list);
1809 #ifdef CONFIG_PREEMPT_NOTIFIERS
1810 INIT_HLIST_HEAD(&p->preempt_notifiers);
1813 #ifdef CONFIG_NUMA_BALANCING
1814 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1815 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1816 p->mm->numa_scan_seq = 0;
1819 if (clone_flags & CLONE_VM)
1820 p->numa_preferred_nid = current->numa_preferred_nid;
1822 p->numa_preferred_nid = -1;
1824 p->node_stamp = 0ULL;
1825 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1826 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1827 p->numa_work.next = &p->numa_work;
1828 p->numa_faults_memory = NULL;
1829 p->numa_faults_buffer_memory = NULL;
1830 p->last_task_numa_placement = 0;
1831 p->last_sum_exec_runtime = 0;
1833 INIT_LIST_HEAD(&p->numa_entry);
1834 p->numa_group = NULL;
1835 #endif /* CONFIG_NUMA_BALANCING */
1838 #ifdef CONFIG_NUMA_BALANCING
1839 #ifdef CONFIG_SCHED_DEBUG
1840 void set_numabalancing_state(bool enabled)
1843 sched_feat_set("NUMA");
1845 sched_feat_set("NO_NUMA");
1848 __read_mostly bool numabalancing_enabled;
1850 void set_numabalancing_state(bool enabled)
1852 numabalancing_enabled = enabled;
1854 #endif /* CONFIG_SCHED_DEBUG */
1856 #ifdef CONFIG_PROC_SYSCTL
1857 int sysctl_numa_balancing(struct ctl_table *table, int write,
1858 void __user *buffer, size_t *lenp, loff_t *ppos)
1862 int state = numabalancing_enabled;
1864 if (write && !capable(CAP_SYS_ADMIN))
1869 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1873 set_numabalancing_state(state);
1880 * fork()/clone()-time setup:
1882 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1884 unsigned long flags;
1885 int cpu = get_cpu();
1887 __sched_fork(clone_flags, p);
1889 * We mark the process as running here. This guarantees that
1890 * nobody will actually run it, and a signal or other external
1891 * event cannot wake it up and insert it on the runqueue either.
1893 p->state = TASK_RUNNING;
1896 * Make sure we do not leak PI boosting priority to the child.
1898 p->prio = current->normal_prio;
1901 * Revert to default priority/policy on fork if requested.
1903 if (unlikely(p->sched_reset_on_fork)) {
1904 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1905 p->policy = SCHED_NORMAL;
1906 p->static_prio = NICE_TO_PRIO(0);
1908 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1909 p->static_prio = NICE_TO_PRIO(0);
1911 p->prio = p->normal_prio = __normal_prio(p);
1915 * We don't need the reset flag anymore after the fork. It has
1916 * fulfilled its duty:
1918 p->sched_reset_on_fork = 0;
1921 if (dl_prio(p->prio)) {
1924 } else if (rt_prio(p->prio)) {
1925 p->sched_class = &rt_sched_class;
1927 p->sched_class = &fair_sched_class;
1930 if (p->sched_class->task_fork)
1931 p->sched_class->task_fork(p);
1934 * The child is not yet in the pid-hash so no cgroup attach races,
1935 * and the cgroup is pinned to this child due to cgroup_fork()
1936 * is ran before sched_fork().
1938 * Silence PROVE_RCU.
1940 raw_spin_lock_irqsave(&p->pi_lock, flags);
1941 set_task_cpu(p, cpu);
1942 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1944 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1945 if (likely(sched_info_on()))
1946 memset(&p->sched_info, 0, sizeof(p->sched_info));
1948 #if defined(CONFIG_SMP)
1951 init_task_preempt_count(p);
1953 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1954 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1961 unsigned long to_ratio(u64 period, u64 runtime)
1963 if (runtime == RUNTIME_INF)
1967 * Doing this here saves a lot of checks in all
1968 * the calling paths, and returning zero seems
1969 * safe for them anyway.
1974 return div64_u64(runtime << 20, period);
1978 inline struct dl_bw *dl_bw_of(int i)
1980 return &cpu_rq(i)->rd->dl_bw;
1983 static inline int dl_bw_cpus(int i)
1985 struct root_domain *rd = cpu_rq(i)->rd;
1988 for_each_cpu_and(i, rd->span, cpu_active_mask)
1994 inline struct dl_bw *dl_bw_of(int i)
1996 return &cpu_rq(i)->dl.dl_bw;
1999 static inline int dl_bw_cpus(int i)
2006 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2008 dl_b->total_bw -= tsk_bw;
2012 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2014 dl_b->total_bw += tsk_bw;
2018 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2020 return dl_b->bw != -1 &&
2021 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2025 * We must be sure that accepting a new task (or allowing changing the
2026 * parameters of an existing one) is consistent with the bandwidth
2027 * constraints. If yes, this function also accordingly updates the currently
2028 * allocated bandwidth to reflect the new situation.
2030 * This function is called while holding p's rq->lock.
2032 static int dl_overflow(struct task_struct *p, int policy,
2033 const struct sched_attr *attr)
2036 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2037 u64 period = attr->sched_period ?: attr->sched_deadline;
2038 u64 runtime = attr->sched_runtime;
2039 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2042 if (new_bw == p->dl.dl_bw)
2046 * Either if a task, enters, leave, or stays -deadline but changes
2047 * its parameters, we may need to update accordingly the total
2048 * allocated bandwidth of the container.
2050 raw_spin_lock(&dl_b->lock);
2051 cpus = dl_bw_cpus(task_cpu(p));
2052 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2053 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2054 __dl_add(dl_b, new_bw);
2056 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2057 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2058 __dl_clear(dl_b, p->dl.dl_bw);
2059 __dl_add(dl_b, new_bw);
2061 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2062 __dl_clear(dl_b, p->dl.dl_bw);
2065 raw_spin_unlock(&dl_b->lock);
2070 extern void init_dl_bw(struct dl_bw *dl_b);
2073 * wake_up_new_task - wake up a newly created task for the first time.
2075 * This function will do some initial scheduler statistics housekeeping
2076 * that must be done for every newly created context, then puts the task
2077 * on the runqueue and wakes it.
2079 void wake_up_new_task(struct task_struct *p)
2081 unsigned long flags;
2084 raw_spin_lock_irqsave(&p->pi_lock, flags);
2087 * Fork balancing, do it here and not earlier because:
2088 * - cpus_allowed can change in the fork path
2089 * - any previously selected cpu might disappear through hotplug
2091 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2094 /* Initialize new task's runnable average */
2095 init_task_runnable_average(p);
2096 rq = __task_rq_lock(p);
2097 activate_task(rq, p, 0);
2099 trace_sched_wakeup_new(p, true);
2100 check_preempt_curr(rq, p, WF_FORK);
2102 if (p->sched_class->task_woken)
2103 p->sched_class->task_woken(rq, p);
2105 task_rq_unlock(rq, p, &flags);
2108 #ifdef CONFIG_PREEMPT_NOTIFIERS
2111 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2112 * @notifier: notifier struct to register
2114 void preempt_notifier_register(struct preempt_notifier *notifier)
2116 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2118 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2121 * preempt_notifier_unregister - no longer interested in preemption notifications
2122 * @notifier: notifier struct to unregister
2124 * This is safe to call from within a preemption notifier.
2126 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2128 hlist_del(¬ifier->link);
2130 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2132 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2134 struct preempt_notifier *notifier;
2136 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2137 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2141 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2142 struct task_struct *next)
2144 struct preempt_notifier *notifier;
2146 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2147 notifier->ops->sched_out(notifier, next);
2150 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2152 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2157 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2158 struct task_struct *next)
2162 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2165 * prepare_task_switch - prepare to switch tasks
2166 * @rq: the runqueue preparing to switch
2167 * @prev: the current task that is being switched out
2168 * @next: the task we are going to switch to.
2170 * This is called with the rq lock held and interrupts off. It must
2171 * be paired with a subsequent finish_task_switch after the context
2174 * prepare_task_switch sets up locking and calls architecture specific
2178 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2179 struct task_struct *next)
2181 trace_sched_switch(prev, next);
2182 sched_info_switch(rq, prev, next);
2183 perf_event_task_sched_out(prev, next);
2184 fire_sched_out_preempt_notifiers(prev, next);
2185 prepare_lock_switch(rq, next);
2186 prepare_arch_switch(next);
2190 * finish_task_switch - clean up after a task-switch
2191 * @rq: runqueue associated with task-switch
2192 * @prev: the thread we just switched away from.
2194 * finish_task_switch must be called after the context switch, paired
2195 * with a prepare_task_switch call before the context switch.
2196 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2197 * and do any other architecture-specific cleanup actions.
2199 * Note that we may have delayed dropping an mm in context_switch(). If
2200 * so, we finish that here outside of the runqueue lock. (Doing it
2201 * with the lock held can cause deadlocks; see schedule() for
2204 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2205 __releases(rq->lock)
2207 struct mm_struct *mm = rq->prev_mm;
2213 * A task struct has one reference for the use as "current".
2214 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2215 * schedule one last time. The schedule call will never return, and
2216 * the scheduled task must drop that reference.
2217 * The test for TASK_DEAD must occur while the runqueue locks are
2218 * still held, otherwise prev could be scheduled on another cpu, die
2219 * there before we look at prev->state, and then the reference would
2221 * Manfred Spraul <manfred@colorfullife.com>
2223 prev_state = prev->state;
2224 vtime_task_switch(prev);
2225 finish_arch_switch(prev);
2226 perf_event_task_sched_in(prev, current);
2227 finish_lock_switch(rq, prev);
2228 finish_arch_post_lock_switch();
2230 fire_sched_in_preempt_notifiers(current);
2233 if (unlikely(prev_state == TASK_DEAD)) {
2234 if (prev->sched_class->task_dead)
2235 prev->sched_class->task_dead(prev);
2238 * Remove function-return probe instances associated with this
2239 * task and put them back on the free list.
2241 kprobe_flush_task(prev);
2242 put_task_struct(prev);
2245 tick_nohz_task_switch(current);
2250 /* rq->lock is NOT held, but preemption is disabled */
2251 static inline void post_schedule(struct rq *rq)
2253 if (rq->post_schedule) {
2254 unsigned long flags;
2256 raw_spin_lock_irqsave(&rq->lock, flags);
2257 if (rq->curr->sched_class->post_schedule)
2258 rq->curr->sched_class->post_schedule(rq);
2259 raw_spin_unlock_irqrestore(&rq->lock, flags);
2261 rq->post_schedule = 0;
2267 static inline void post_schedule(struct rq *rq)
2274 * schedule_tail - first thing a freshly forked thread must call.
2275 * @prev: the thread we just switched away from.
2277 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2278 __releases(rq->lock)
2280 struct rq *rq = this_rq();
2282 finish_task_switch(rq, prev);
2285 * FIXME: do we need to worry about rq being invalidated by the
2290 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2291 /* In this case, finish_task_switch does not reenable preemption */
2294 if (current->set_child_tid)
2295 put_user(task_pid_vnr(current), current->set_child_tid);
2299 * context_switch - switch to the new MM and the new
2300 * thread's register state.
2303 context_switch(struct rq *rq, struct task_struct *prev,
2304 struct task_struct *next)
2306 struct mm_struct *mm, *oldmm;
2308 prepare_task_switch(rq, prev, next);
2311 oldmm = prev->active_mm;
2313 * For paravirt, this is coupled with an exit in switch_to to
2314 * combine the page table reload and the switch backend into
2317 arch_start_context_switch(prev);
2320 next->active_mm = oldmm;
2321 atomic_inc(&oldmm->mm_count);
2322 enter_lazy_tlb(oldmm, next);
2324 switch_mm(oldmm, mm, next);
2327 prev->active_mm = NULL;
2328 rq->prev_mm = oldmm;
2331 * Since the runqueue lock will be released by the next
2332 * task (which is an invalid locking op but in the case
2333 * of the scheduler it's an obvious special-case), so we
2334 * do an early lockdep release here:
2336 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2337 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2340 context_tracking_task_switch(prev, next);
2341 /* Here we just switch the register state and the stack. */
2342 switch_to(prev, next, prev);
2346 * this_rq must be evaluated again because prev may have moved
2347 * CPUs since it called schedule(), thus the 'rq' on its stack
2348 * frame will be invalid.
2350 finish_task_switch(this_rq(), prev);
2354 * nr_running and nr_context_switches:
2356 * externally visible scheduler statistics: current number of runnable
2357 * threads, total number of context switches performed since bootup.
2359 unsigned long nr_running(void)
2361 unsigned long i, sum = 0;
2363 for_each_online_cpu(i)
2364 sum += cpu_rq(i)->nr_running;
2369 unsigned long long nr_context_switches(void)
2372 unsigned long long sum = 0;
2374 for_each_possible_cpu(i)
2375 sum += cpu_rq(i)->nr_switches;
2380 unsigned long nr_iowait(void)
2382 unsigned long i, sum = 0;
2384 for_each_possible_cpu(i)
2385 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2390 unsigned long nr_iowait_cpu(int cpu)
2392 struct rq *this = cpu_rq(cpu);
2393 return atomic_read(&this->nr_iowait);
2399 * sched_exec - execve() is a valuable balancing opportunity, because at
2400 * this point the task has the smallest effective memory and cache footprint.
2402 void sched_exec(void)
2404 struct task_struct *p = current;
2405 unsigned long flags;
2408 raw_spin_lock_irqsave(&p->pi_lock, flags);
2409 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2410 if (dest_cpu == smp_processor_id())
2413 if (likely(cpu_active(dest_cpu))) {
2414 struct migration_arg arg = { p, dest_cpu };
2416 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2417 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2421 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2426 DEFINE_PER_CPU(struct kernel_stat, kstat);
2427 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2429 EXPORT_PER_CPU_SYMBOL(kstat);
2430 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2433 * Return any ns on the sched_clock that have not yet been accounted in
2434 * @p in case that task is currently running.
2436 * Called with task_rq_lock() held on @rq.
2438 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2443 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2444 * project cycles that may never be accounted to this
2445 * thread, breaking clock_gettime().
2447 if (task_current(rq, p) && p->on_rq) {
2448 update_rq_clock(rq);
2449 ns = rq_clock_task(rq) - p->se.exec_start;
2457 unsigned long long task_delta_exec(struct task_struct *p)
2459 unsigned long flags;
2463 rq = task_rq_lock(p, &flags);
2464 ns = do_task_delta_exec(p, rq);
2465 task_rq_unlock(rq, p, &flags);
2471 * Return accounted runtime for the task.
2472 * In case the task is currently running, return the runtime plus current's
2473 * pending runtime that have not been accounted yet.
2475 unsigned long long task_sched_runtime(struct task_struct *p)
2477 unsigned long flags;
2481 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2483 * 64-bit doesn't need locks to atomically read a 64bit value.
2484 * So we have a optimization chance when the task's delta_exec is 0.
2485 * Reading ->on_cpu is racy, but this is ok.
2487 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2488 * If we race with it entering cpu, unaccounted time is 0. This is
2489 * indistinguishable from the read occurring a few cycles earlier.
2490 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2491 * been accounted, so we're correct here as well.
2493 if (!p->on_cpu || !p->on_rq)
2494 return p->se.sum_exec_runtime;
2497 rq = task_rq_lock(p, &flags);
2498 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2499 task_rq_unlock(rq, p, &flags);
2505 * This function gets called by the timer code, with HZ frequency.
2506 * We call it with interrupts disabled.
2508 void scheduler_tick(void)
2510 int cpu = smp_processor_id();
2511 struct rq *rq = cpu_rq(cpu);
2512 struct task_struct *curr = rq->curr;
2516 raw_spin_lock(&rq->lock);
2517 update_rq_clock(rq);
2518 curr->sched_class->task_tick(rq, curr, 0);
2519 update_cpu_load_active(rq);
2520 raw_spin_unlock(&rq->lock);
2522 perf_event_task_tick();
2525 rq->idle_balance = idle_cpu(cpu);
2526 trigger_load_balance(rq);
2528 rq_last_tick_reset(rq);
2531 #ifdef CONFIG_NO_HZ_FULL
2533 * scheduler_tick_max_deferment
2535 * Keep at least one tick per second when a single
2536 * active task is running because the scheduler doesn't
2537 * yet completely support full dynticks environment.
2539 * This makes sure that uptime, CFS vruntime, load
2540 * balancing, etc... continue to move forward, even
2541 * with a very low granularity.
2543 * Return: Maximum deferment in nanoseconds.
2545 u64 scheduler_tick_max_deferment(void)
2547 struct rq *rq = this_rq();
2548 unsigned long next, now = ACCESS_ONCE(jiffies);
2550 next = rq->last_sched_tick + HZ;
2552 if (time_before_eq(next, now))
2555 return jiffies_to_nsecs(next - now);
2559 notrace unsigned long get_parent_ip(unsigned long addr)
2561 if (in_lock_functions(addr)) {
2562 addr = CALLER_ADDR2;
2563 if (in_lock_functions(addr))
2564 addr = CALLER_ADDR3;
2569 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2570 defined(CONFIG_PREEMPT_TRACER))
2572 void preempt_count_add(int val)
2574 #ifdef CONFIG_DEBUG_PREEMPT
2578 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2581 __preempt_count_add(val);
2582 #ifdef CONFIG_DEBUG_PREEMPT
2584 * Spinlock count overflowing soon?
2586 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2589 if (preempt_count() == val) {
2590 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2591 #ifdef CONFIG_DEBUG_PREEMPT
2592 current->preempt_disable_ip = ip;
2594 trace_preempt_off(CALLER_ADDR0, ip);
2597 EXPORT_SYMBOL(preempt_count_add);
2598 NOKPROBE_SYMBOL(preempt_count_add);
2600 void preempt_count_sub(int val)
2602 #ifdef CONFIG_DEBUG_PREEMPT
2606 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2609 * Is the spinlock portion underflowing?
2611 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2612 !(preempt_count() & PREEMPT_MASK)))
2616 if (preempt_count() == val)
2617 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2618 __preempt_count_sub(val);
2620 EXPORT_SYMBOL(preempt_count_sub);
2621 NOKPROBE_SYMBOL(preempt_count_sub);
2626 * Print scheduling while atomic bug:
2628 static noinline void __schedule_bug(struct task_struct *prev)
2630 if (oops_in_progress)
2633 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2634 prev->comm, prev->pid, preempt_count());
2636 debug_show_held_locks(prev);
2638 if (irqs_disabled())
2639 print_irqtrace_events(prev);
2640 #ifdef CONFIG_DEBUG_PREEMPT
2641 if (in_atomic_preempt_off()) {
2642 pr_err("Preemption disabled at:");
2643 print_ip_sym(current->preempt_disable_ip);
2648 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2652 * Various schedule()-time debugging checks and statistics:
2654 static inline void schedule_debug(struct task_struct *prev)
2657 * Test if we are atomic. Since do_exit() needs to call into
2658 * schedule() atomically, we ignore that path. Otherwise whine
2659 * if we are scheduling when we should not.
2661 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2662 __schedule_bug(prev);
2665 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2667 schedstat_inc(this_rq(), sched_count);
2671 * Pick up the highest-prio task:
2673 static inline struct task_struct *
2674 pick_next_task(struct rq *rq, struct task_struct *prev)
2676 const struct sched_class *class = &fair_sched_class;
2677 struct task_struct *p;
2680 * Optimization: we know that if all tasks are in
2681 * the fair class we can call that function directly:
2683 if (likely(prev->sched_class == class &&
2684 rq->nr_running == rq->cfs.h_nr_running)) {
2685 p = fair_sched_class.pick_next_task(rq, prev);
2686 if (unlikely(p == RETRY_TASK))
2689 /* assumes fair_sched_class->next == idle_sched_class */
2691 p = idle_sched_class.pick_next_task(rq, prev);
2697 for_each_class(class) {
2698 p = class->pick_next_task(rq, prev);
2700 if (unlikely(p == RETRY_TASK))
2706 BUG(); /* the idle class will always have a runnable task */
2710 * __schedule() is the main scheduler function.
2712 * The main means of driving the scheduler and thus entering this function are:
2714 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2716 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2717 * paths. For example, see arch/x86/entry_64.S.
2719 * To drive preemption between tasks, the scheduler sets the flag in timer
2720 * interrupt handler scheduler_tick().
2722 * 3. Wakeups don't really cause entry into schedule(). They add a
2723 * task to the run-queue and that's it.
2725 * Now, if the new task added to the run-queue preempts the current
2726 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2727 * called on the nearest possible occasion:
2729 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2731 * - in syscall or exception context, at the next outmost
2732 * preempt_enable(). (this might be as soon as the wake_up()'s
2735 * - in IRQ context, return from interrupt-handler to
2736 * preemptible context
2738 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2741 * - cond_resched() call
2742 * - explicit schedule() call
2743 * - return from syscall or exception to user-space
2744 * - return from interrupt-handler to user-space
2746 static void __sched __schedule(void)
2748 struct task_struct *prev, *next;
2749 unsigned long *switch_count;
2755 cpu = smp_processor_id();
2757 rcu_note_context_switch(cpu);
2760 schedule_debug(prev);
2762 if (sched_feat(HRTICK))
2766 * Make sure that signal_pending_state()->signal_pending() below
2767 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2768 * done by the caller to avoid the race with signal_wake_up().
2770 smp_mb__before_spinlock();
2771 raw_spin_lock_irq(&rq->lock);
2773 switch_count = &prev->nivcsw;
2774 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2775 if (unlikely(signal_pending_state(prev->state, prev))) {
2776 prev->state = TASK_RUNNING;
2778 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2782 * If a worker went to sleep, notify and ask workqueue
2783 * whether it wants to wake up a task to maintain
2786 if (prev->flags & PF_WQ_WORKER) {
2787 struct task_struct *to_wakeup;
2789 to_wakeup = wq_worker_sleeping(prev, cpu);
2791 try_to_wake_up_local(to_wakeup);
2794 switch_count = &prev->nvcsw;
2797 if (prev->on_rq || rq->skip_clock_update < 0)
2798 update_rq_clock(rq);
2800 next = pick_next_task(rq, prev);
2801 clear_tsk_need_resched(prev);
2802 clear_preempt_need_resched();
2803 rq->skip_clock_update = 0;
2805 if (likely(prev != next)) {
2810 context_switch(rq, prev, next); /* unlocks the rq */
2812 * The context switch have flipped the stack from under us
2813 * and restored the local variables which were saved when
2814 * this task called schedule() in the past. prev == current
2815 * is still correct, but it can be moved to another cpu/rq.
2817 cpu = smp_processor_id();
2820 raw_spin_unlock_irq(&rq->lock);
2824 sched_preempt_enable_no_resched();
2829 static inline void sched_submit_work(struct task_struct *tsk)
2831 if (!tsk->state || tsk_is_pi_blocked(tsk))
2834 * If we are going to sleep and we have plugged IO queued,
2835 * make sure to submit it to avoid deadlocks.
2837 if (blk_needs_flush_plug(tsk))
2838 blk_schedule_flush_plug(tsk);
2841 asmlinkage __visible void __sched schedule(void)
2843 struct task_struct *tsk = current;
2845 sched_submit_work(tsk);
2848 EXPORT_SYMBOL(schedule);
2850 #ifdef CONFIG_CONTEXT_TRACKING
2851 asmlinkage __visible void __sched schedule_user(void)
2854 * If we come here after a random call to set_need_resched(),
2855 * or we have been woken up remotely but the IPI has not yet arrived,
2856 * we haven't yet exited the RCU idle mode. Do it here manually until
2857 * we find a better solution.
2866 * schedule_preempt_disabled - called with preemption disabled
2868 * Returns with preemption disabled. Note: preempt_count must be 1
2870 void __sched schedule_preempt_disabled(void)
2872 sched_preempt_enable_no_resched();
2877 #ifdef CONFIG_PREEMPT
2879 * this is the entry point to schedule() from in-kernel preemption
2880 * off of preempt_enable. Kernel preemptions off return from interrupt
2881 * occur there and call schedule directly.
2883 asmlinkage __visible void __sched notrace preempt_schedule(void)
2886 * If there is a non-zero preempt_count or interrupts are disabled,
2887 * we do not want to preempt the current task. Just return..
2889 if (likely(!preemptible()))
2893 __preempt_count_add(PREEMPT_ACTIVE);
2895 __preempt_count_sub(PREEMPT_ACTIVE);
2898 * Check again in case we missed a preemption opportunity
2899 * between schedule and now.
2902 } while (need_resched());
2904 NOKPROBE_SYMBOL(preempt_schedule);
2905 EXPORT_SYMBOL(preempt_schedule);
2906 #endif /* CONFIG_PREEMPT */
2909 * this is the entry point to schedule() from kernel preemption
2910 * off of irq context.
2911 * Note, that this is called and return with irqs disabled. This will
2912 * protect us against recursive calling from irq.
2914 asmlinkage __visible void __sched preempt_schedule_irq(void)
2916 enum ctx_state prev_state;
2918 /* Catch callers which need to be fixed */
2919 BUG_ON(preempt_count() || !irqs_disabled());
2921 prev_state = exception_enter();
2924 __preempt_count_add(PREEMPT_ACTIVE);
2927 local_irq_disable();
2928 __preempt_count_sub(PREEMPT_ACTIVE);
2931 * Check again in case we missed a preemption opportunity
2932 * between schedule and now.
2935 } while (need_resched());
2937 exception_exit(prev_state);
2940 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2943 return try_to_wake_up(curr->private, mode, wake_flags);
2945 EXPORT_SYMBOL(default_wake_function);
2947 #ifdef CONFIG_RT_MUTEXES
2950 * rt_mutex_setprio - set the current priority of a task
2952 * @prio: prio value (kernel-internal form)
2954 * This function changes the 'effective' priority of a task. It does
2955 * not touch ->normal_prio like __setscheduler().
2957 * Used by the rt_mutex code to implement priority inheritance
2958 * logic. Call site only calls if the priority of the task changed.
2960 void rt_mutex_setprio(struct task_struct *p, int prio)
2962 int oldprio, on_rq, running, enqueue_flag = 0;
2964 const struct sched_class *prev_class;
2966 BUG_ON(prio > MAX_PRIO);
2968 rq = __task_rq_lock(p);
2971 * Idle task boosting is a nono in general. There is one
2972 * exception, when PREEMPT_RT and NOHZ is active:
2974 * The idle task calls get_next_timer_interrupt() and holds
2975 * the timer wheel base->lock on the CPU and another CPU wants
2976 * to access the timer (probably to cancel it). We can safely
2977 * ignore the boosting request, as the idle CPU runs this code
2978 * with interrupts disabled and will complete the lock
2979 * protected section without being interrupted. So there is no
2980 * real need to boost.
2982 if (unlikely(p == rq->idle)) {
2983 WARN_ON(p != rq->curr);
2984 WARN_ON(p->pi_blocked_on);
2988 trace_sched_pi_setprio(p, prio);
2990 prev_class = p->sched_class;
2992 running = task_current(rq, p);
2994 dequeue_task(rq, p, 0);
2996 p->sched_class->put_prev_task(rq, p);
2999 * Boosting condition are:
3000 * 1. -rt task is running and holds mutex A
3001 * --> -dl task blocks on mutex A
3003 * 2. -dl task is running and holds mutex A
3004 * --> -dl task blocks on mutex A and could preempt the
3007 if (dl_prio(prio)) {
3008 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3009 if (!dl_prio(p->normal_prio) ||
3010 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3011 p->dl.dl_boosted = 1;
3012 p->dl.dl_throttled = 0;
3013 enqueue_flag = ENQUEUE_REPLENISH;
3015 p->dl.dl_boosted = 0;
3016 p->sched_class = &dl_sched_class;
3017 } else if (rt_prio(prio)) {
3018 if (dl_prio(oldprio))
3019 p->dl.dl_boosted = 0;
3021 enqueue_flag = ENQUEUE_HEAD;
3022 p->sched_class = &rt_sched_class;
3024 if (dl_prio(oldprio))
3025 p->dl.dl_boosted = 0;
3026 p->sched_class = &fair_sched_class;
3032 p->sched_class->set_curr_task(rq);
3034 enqueue_task(rq, p, enqueue_flag);
3036 check_class_changed(rq, p, prev_class, oldprio);
3038 __task_rq_unlock(rq);
3042 void set_user_nice(struct task_struct *p, long nice)
3044 int old_prio, delta, on_rq;
3045 unsigned long flags;
3048 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3051 * We have to be careful, if called from sys_setpriority(),
3052 * the task might be in the middle of scheduling on another CPU.
3054 rq = task_rq_lock(p, &flags);
3056 * The RT priorities are set via sched_setscheduler(), but we still
3057 * allow the 'normal' nice value to be set - but as expected
3058 * it wont have any effect on scheduling until the task is
3059 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3061 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3062 p->static_prio = NICE_TO_PRIO(nice);
3067 dequeue_task(rq, p, 0);
3069 p->static_prio = NICE_TO_PRIO(nice);
3072 p->prio = effective_prio(p);
3073 delta = p->prio - old_prio;
3076 enqueue_task(rq, p, 0);
3078 * If the task increased its priority or is running and
3079 * lowered its priority, then reschedule its CPU:
3081 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3085 task_rq_unlock(rq, p, &flags);
3087 EXPORT_SYMBOL(set_user_nice);
3090 * can_nice - check if a task can reduce its nice value
3094 int can_nice(const struct task_struct *p, const int nice)
3096 /* convert nice value [19,-20] to rlimit style value [1,40] */
3097 int nice_rlim = nice_to_rlimit(nice);
3099 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3100 capable(CAP_SYS_NICE));
3103 #ifdef __ARCH_WANT_SYS_NICE
3106 * sys_nice - change the priority of the current process.
3107 * @increment: priority increment
3109 * sys_setpriority is a more generic, but much slower function that
3110 * does similar things.
3112 SYSCALL_DEFINE1(nice, int, increment)
3117 * Setpriority might change our priority at the same moment.
3118 * We don't have to worry. Conceptually one call occurs first
3119 * and we have a single winner.
3121 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3122 nice = task_nice(current) + increment;
3124 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3125 if (increment < 0 && !can_nice(current, nice))
3128 retval = security_task_setnice(current, nice);
3132 set_user_nice(current, nice);
3139 * task_prio - return the priority value of a given task.
3140 * @p: the task in question.
3142 * Return: The priority value as seen by users in /proc.
3143 * RT tasks are offset by -200. Normal tasks are centered
3144 * around 0, value goes from -16 to +15.
3146 int task_prio(const struct task_struct *p)
3148 return p->prio - MAX_RT_PRIO;
3152 * idle_cpu - is a given cpu idle currently?
3153 * @cpu: the processor in question.
3155 * Return: 1 if the CPU is currently idle. 0 otherwise.
3157 int idle_cpu(int cpu)
3159 struct rq *rq = cpu_rq(cpu);
3161 if (rq->curr != rq->idle)
3168 if (!llist_empty(&rq->wake_list))
3176 * idle_task - return the idle task for a given cpu.
3177 * @cpu: the processor in question.
3179 * Return: The idle task for the cpu @cpu.
3181 struct task_struct *idle_task(int cpu)
3183 return cpu_rq(cpu)->idle;
3187 * find_process_by_pid - find a process with a matching PID value.
3188 * @pid: the pid in question.
3190 * The task of @pid, if found. %NULL otherwise.
3192 static struct task_struct *find_process_by_pid(pid_t pid)
3194 return pid ? find_task_by_vpid(pid) : current;
3198 * This function initializes the sched_dl_entity of a newly becoming
3199 * SCHED_DEADLINE task.
3201 * Only the static values are considered here, the actual runtime and the
3202 * absolute deadline will be properly calculated when the task is enqueued
3203 * for the first time with its new policy.
3206 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3208 struct sched_dl_entity *dl_se = &p->dl;
3210 init_dl_task_timer(dl_se);
3211 dl_se->dl_runtime = attr->sched_runtime;
3212 dl_se->dl_deadline = attr->sched_deadline;
3213 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3214 dl_se->flags = attr->sched_flags;
3215 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3216 dl_se->dl_throttled = 0;
3218 dl_se->dl_yielded = 0;
3222 * sched_setparam() passes in -1 for its policy, to let the functions
3223 * it calls know not to change it.
3225 #define SETPARAM_POLICY -1
3227 static void __setscheduler_params(struct task_struct *p,
3228 const struct sched_attr *attr)
3230 int policy = attr->sched_policy;
3232 if (policy == SETPARAM_POLICY)
3237 if (dl_policy(policy))
3238 __setparam_dl(p, attr);
3239 else if (fair_policy(policy))
3240 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3243 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3244 * !rt_policy. Always setting this ensures that things like
3245 * getparam()/getattr() don't report silly values for !rt tasks.
3247 p->rt_priority = attr->sched_priority;
3248 p->normal_prio = normal_prio(p);
3252 /* Actually do priority change: must hold pi & rq lock. */
3253 static void __setscheduler(struct rq *rq, struct task_struct *p,
3254 const struct sched_attr *attr)
3256 __setscheduler_params(p, attr);
3259 * If we get here, there was no pi waiters boosting the
3260 * task. It is safe to use the normal prio.
3262 p->prio = normal_prio(p);
3264 if (dl_prio(p->prio))
3265 p->sched_class = &dl_sched_class;
3266 else if (rt_prio(p->prio))
3267 p->sched_class = &rt_sched_class;
3269 p->sched_class = &fair_sched_class;
3273 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3275 struct sched_dl_entity *dl_se = &p->dl;
3277 attr->sched_priority = p->rt_priority;
3278 attr->sched_runtime = dl_se->dl_runtime;
3279 attr->sched_deadline = dl_se->dl_deadline;
3280 attr->sched_period = dl_se->dl_period;
3281 attr->sched_flags = dl_se->flags;
3285 * This function validates the new parameters of a -deadline task.
3286 * We ask for the deadline not being zero, and greater or equal
3287 * than the runtime, as well as the period of being zero or
3288 * greater than deadline. Furthermore, we have to be sure that
3289 * user parameters are above the internal resolution of 1us (we
3290 * check sched_runtime only since it is always the smaller one) and
3291 * below 2^63 ns (we have to check both sched_deadline and
3292 * sched_period, as the latter can be zero).
3295 __checkparam_dl(const struct sched_attr *attr)
3298 if (attr->sched_deadline == 0)
3302 * Since we truncate DL_SCALE bits, make sure we're at least
3305 if (attr->sched_runtime < (1ULL << DL_SCALE))
3309 * Since we use the MSB for wrap-around and sign issues, make
3310 * sure it's not set (mind that period can be equal to zero).
3312 if (attr->sched_deadline & (1ULL << 63) ||
3313 attr->sched_period & (1ULL << 63))
3316 /* runtime <= deadline <= period (if period != 0) */
3317 if ((attr->sched_period != 0 &&
3318 attr->sched_period < attr->sched_deadline) ||
3319 attr->sched_deadline < attr->sched_runtime)
3326 * check the target process has a UID that matches the current process's
3328 static bool check_same_owner(struct task_struct *p)
3330 const struct cred *cred = current_cred(), *pcred;
3334 pcred = __task_cred(p);
3335 match = (uid_eq(cred->euid, pcred->euid) ||
3336 uid_eq(cred->euid, pcred->uid));
3341 static int __sched_setscheduler(struct task_struct *p,
3342 const struct sched_attr *attr,
3345 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3346 MAX_RT_PRIO - 1 - attr->sched_priority;
3347 int retval, oldprio, oldpolicy = -1, on_rq, running;
3348 int policy = attr->sched_policy;
3349 unsigned long flags;
3350 const struct sched_class *prev_class;
3354 /* may grab non-irq protected spin_locks */
3355 BUG_ON(in_interrupt());
3357 /* double check policy once rq lock held */
3359 reset_on_fork = p->sched_reset_on_fork;
3360 policy = oldpolicy = p->policy;
3362 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3364 if (policy != SCHED_DEADLINE &&
3365 policy != SCHED_FIFO && policy != SCHED_RR &&
3366 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3367 policy != SCHED_IDLE)
3371 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3375 * Valid priorities for SCHED_FIFO and SCHED_RR are
3376 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3377 * SCHED_BATCH and SCHED_IDLE is 0.
3379 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3380 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3382 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3383 (rt_policy(policy) != (attr->sched_priority != 0)))
3387 * Allow unprivileged RT tasks to decrease priority:
3389 if (user && !capable(CAP_SYS_NICE)) {
3390 if (fair_policy(policy)) {
3391 if (attr->sched_nice < task_nice(p) &&
3392 !can_nice(p, attr->sched_nice))
3396 if (rt_policy(policy)) {
3397 unsigned long rlim_rtprio =
3398 task_rlimit(p, RLIMIT_RTPRIO);
3400 /* can't set/change the rt policy */
3401 if (policy != p->policy && !rlim_rtprio)
3404 /* can't increase priority */
3405 if (attr->sched_priority > p->rt_priority &&
3406 attr->sched_priority > rlim_rtprio)
3411 * Can't set/change SCHED_DEADLINE policy at all for now
3412 * (safest behavior); in the future we would like to allow
3413 * unprivileged DL tasks to increase their relative deadline
3414 * or reduce their runtime (both ways reducing utilization)
3416 if (dl_policy(policy))
3420 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3421 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3423 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3424 if (!can_nice(p, task_nice(p)))
3428 /* can't change other user's priorities */
3429 if (!check_same_owner(p))
3432 /* Normal users shall not reset the sched_reset_on_fork flag */
3433 if (p->sched_reset_on_fork && !reset_on_fork)
3438 retval = security_task_setscheduler(p);
3444 * make sure no PI-waiters arrive (or leave) while we are
3445 * changing the priority of the task:
3447 * To be able to change p->policy safely, the appropriate
3448 * runqueue lock must be held.
3450 rq = task_rq_lock(p, &flags);
3453 * Changing the policy of the stop threads its a very bad idea
3455 if (p == rq->stop) {
3456 task_rq_unlock(rq, p, &flags);
3461 * If not changing anything there's no need to proceed further,
3462 * but store a possible modification of reset_on_fork.
3464 if (unlikely(policy == p->policy)) {
3465 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3467 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3469 if (dl_policy(policy))
3472 p->sched_reset_on_fork = reset_on_fork;
3473 task_rq_unlock(rq, p, &flags);
3479 #ifdef CONFIG_RT_GROUP_SCHED
3481 * Do not allow realtime tasks into groups that have no runtime
3484 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3485 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3486 !task_group_is_autogroup(task_group(p))) {
3487 task_rq_unlock(rq, p, &flags);
3492 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3493 cpumask_t *span = rq->rd->span;
3496 * Don't allow tasks with an affinity mask smaller than
3497 * the entire root_domain to become SCHED_DEADLINE. We
3498 * will also fail if there's no bandwidth available.
3500 if (!cpumask_subset(span, &p->cpus_allowed) ||
3501 rq->rd->dl_bw.bw == 0) {
3502 task_rq_unlock(rq, p, &flags);
3509 /* recheck policy now with rq lock held */
3510 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3511 policy = oldpolicy = -1;
3512 task_rq_unlock(rq, p, &flags);
3517 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3518 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3521 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3522 task_rq_unlock(rq, p, &flags);
3526 p->sched_reset_on_fork = reset_on_fork;
3530 * Special case for priority boosted tasks.
3532 * If the new priority is lower or equal (user space view)
3533 * than the current (boosted) priority, we just store the new
3534 * normal parameters and do not touch the scheduler class and
3535 * the runqueue. This will be done when the task deboost
3538 if (rt_mutex_check_prio(p, newprio)) {
3539 __setscheduler_params(p, attr);
3540 task_rq_unlock(rq, p, &flags);
3545 running = task_current(rq, p);
3547 dequeue_task(rq, p, 0);
3549 p->sched_class->put_prev_task(rq, p);
3551 prev_class = p->sched_class;
3552 __setscheduler(rq, p, attr);
3555 p->sched_class->set_curr_task(rq);
3558 * We enqueue to tail when the priority of a task is
3559 * increased (user space view).
3561 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3564 check_class_changed(rq, p, prev_class, oldprio);
3565 task_rq_unlock(rq, p, &flags);
3567 rt_mutex_adjust_pi(p);
3572 static int _sched_setscheduler(struct task_struct *p, int policy,
3573 const struct sched_param *param, bool check)
3575 struct sched_attr attr = {
3576 .sched_policy = policy,
3577 .sched_priority = param->sched_priority,
3578 .sched_nice = PRIO_TO_NICE(p->static_prio),
3581 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3582 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3583 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3584 policy &= ~SCHED_RESET_ON_FORK;
3585 attr.sched_policy = policy;
3588 return __sched_setscheduler(p, &attr, check);
3591 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3592 * @p: the task in question.
3593 * @policy: new policy.
3594 * @param: structure containing the new RT priority.
3596 * Return: 0 on success. An error code otherwise.
3598 * NOTE that the task may be already dead.
3600 int sched_setscheduler(struct task_struct *p, int policy,
3601 const struct sched_param *param)
3603 return _sched_setscheduler(p, policy, param, true);
3605 EXPORT_SYMBOL_GPL(sched_setscheduler);
3607 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3609 return __sched_setscheduler(p, attr, true);
3611 EXPORT_SYMBOL_GPL(sched_setattr);
3614 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3615 * @p: the task in question.
3616 * @policy: new policy.
3617 * @param: structure containing the new RT priority.
3619 * Just like sched_setscheduler, only don't bother checking if the
3620 * current context has permission. For example, this is needed in
3621 * stop_machine(): we create temporary high priority worker threads,
3622 * but our caller might not have that capability.
3624 * Return: 0 on success. An error code otherwise.
3626 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3627 const struct sched_param *param)
3629 return _sched_setscheduler(p, policy, param, false);
3633 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3635 struct sched_param lparam;
3636 struct task_struct *p;
3639 if (!param || pid < 0)
3641 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3646 p = find_process_by_pid(pid);
3648 retval = sched_setscheduler(p, policy, &lparam);
3655 * Mimics kernel/events/core.c perf_copy_attr().
3657 static int sched_copy_attr(struct sched_attr __user *uattr,
3658 struct sched_attr *attr)
3663 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3667 * zero the full structure, so that a short copy will be nice.
3669 memset(attr, 0, sizeof(*attr));
3671 ret = get_user(size, &uattr->size);
3675 if (size > PAGE_SIZE) /* silly large */
3678 if (!size) /* abi compat */
3679 size = SCHED_ATTR_SIZE_VER0;
3681 if (size < SCHED_ATTR_SIZE_VER0)
3685 * If we're handed a bigger struct than we know of,
3686 * ensure all the unknown bits are 0 - i.e. new
3687 * user-space does not rely on any kernel feature
3688 * extensions we dont know about yet.
3690 if (size > sizeof(*attr)) {
3691 unsigned char __user *addr;
3692 unsigned char __user *end;
3695 addr = (void __user *)uattr + sizeof(*attr);
3696 end = (void __user *)uattr + size;
3698 for (; addr < end; addr++) {
3699 ret = get_user(val, addr);
3705 size = sizeof(*attr);
3708 ret = copy_from_user(attr, uattr, size);
3713 * XXX: do we want to be lenient like existing syscalls; or do we want
3714 * to be strict and return an error on out-of-bounds values?
3716 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3721 put_user(sizeof(*attr), &uattr->size);
3726 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3727 * @pid: the pid in question.
3728 * @policy: new policy.
3729 * @param: structure containing the new RT priority.
3731 * Return: 0 on success. An error code otherwise.
3733 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3734 struct sched_param __user *, param)
3736 /* negative values for policy are not valid */
3740 return do_sched_setscheduler(pid, policy, param);
3744 * sys_sched_setparam - set/change the RT priority of a thread
3745 * @pid: the pid in question.
3746 * @param: structure containing the new RT priority.
3748 * Return: 0 on success. An error code otherwise.
3750 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3752 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3756 * sys_sched_setattr - same as above, but with extended sched_attr
3757 * @pid: the pid in question.
3758 * @uattr: structure containing the extended parameters.
3759 * @flags: for future extension.
3761 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3762 unsigned int, flags)
3764 struct sched_attr attr;
3765 struct task_struct *p;
3768 if (!uattr || pid < 0 || flags)
3771 retval = sched_copy_attr(uattr, &attr);
3775 if ((int)attr.sched_policy < 0)
3780 p = find_process_by_pid(pid);
3782 retval = sched_setattr(p, &attr);
3789 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3790 * @pid: the pid in question.
3792 * Return: On success, the policy of the thread. Otherwise, a negative error
3795 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3797 struct task_struct *p;
3805 p = find_process_by_pid(pid);
3807 retval = security_task_getscheduler(p);
3810 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3817 * sys_sched_getparam - get the RT priority of a thread
3818 * @pid: the pid in question.
3819 * @param: structure containing the RT priority.
3821 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3824 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3826 struct sched_param lp = { .sched_priority = 0 };
3827 struct task_struct *p;
3830 if (!param || pid < 0)
3834 p = find_process_by_pid(pid);
3839 retval = security_task_getscheduler(p);
3843 if (task_has_rt_policy(p))
3844 lp.sched_priority = p->rt_priority;
3848 * This one might sleep, we cannot do it with a spinlock held ...
3850 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3859 static int sched_read_attr(struct sched_attr __user *uattr,
3860 struct sched_attr *attr,
3865 if (!access_ok(VERIFY_WRITE, uattr, usize))
3869 * If we're handed a smaller struct than we know of,
3870 * ensure all the unknown bits are 0 - i.e. old
3871 * user-space does not get uncomplete information.
3873 if (usize < sizeof(*attr)) {
3874 unsigned char *addr;
3877 addr = (void *)attr + usize;
3878 end = (void *)attr + sizeof(*attr);
3880 for (; addr < end; addr++) {
3888 ret = copy_to_user(uattr, attr, attr->size);
3896 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3897 * @pid: the pid in question.
3898 * @uattr: structure containing the extended parameters.
3899 * @size: sizeof(attr) for fwd/bwd comp.
3900 * @flags: for future extension.
3902 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3903 unsigned int, size, unsigned int, flags)
3905 struct sched_attr attr = {
3906 .size = sizeof(struct sched_attr),
3908 struct task_struct *p;
3911 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3912 size < SCHED_ATTR_SIZE_VER0 || flags)
3916 p = find_process_by_pid(pid);
3921 retval = security_task_getscheduler(p);
3925 attr.sched_policy = p->policy;
3926 if (p->sched_reset_on_fork)
3927 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3928 if (task_has_dl_policy(p))
3929 __getparam_dl(p, &attr);
3930 else if (task_has_rt_policy(p))
3931 attr.sched_priority = p->rt_priority;
3933 attr.sched_nice = task_nice(p);
3937 retval = sched_read_attr(uattr, &attr, size);
3945 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3947 cpumask_var_t cpus_allowed, new_mask;
3948 struct task_struct *p;
3953 p = find_process_by_pid(pid);
3959 /* Prevent p going away */
3963 if (p->flags & PF_NO_SETAFFINITY) {
3967 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3971 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3973 goto out_free_cpus_allowed;
3976 if (!check_same_owner(p)) {
3978 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3985 retval = security_task_setscheduler(p);
3990 cpuset_cpus_allowed(p, cpus_allowed);
3991 cpumask_and(new_mask, in_mask, cpus_allowed);
3994 * Since bandwidth control happens on root_domain basis,
3995 * if admission test is enabled, we only admit -deadline
3996 * tasks allowed to run on all the CPUs in the task's
4000 if (task_has_dl_policy(p)) {
4001 const struct cpumask *span = task_rq(p)->rd->span;
4003 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4010 retval = set_cpus_allowed_ptr(p, new_mask);
4013 cpuset_cpus_allowed(p, cpus_allowed);
4014 if (!cpumask_subset(new_mask, cpus_allowed)) {
4016 * We must have raced with a concurrent cpuset
4017 * update. Just reset the cpus_allowed to the
4018 * cpuset's cpus_allowed
4020 cpumask_copy(new_mask, cpus_allowed);
4025 free_cpumask_var(new_mask);
4026 out_free_cpus_allowed:
4027 free_cpumask_var(cpus_allowed);
4033 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4034 struct cpumask *new_mask)
4036 if (len < cpumask_size())
4037 cpumask_clear(new_mask);
4038 else if (len > cpumask_size())
4039 len = cpumask_size();
4041 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4045 * sys_sched_setaffinity - set the cpu affinity of a process
4046 * @pid: pid of the process
4047 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4048 * @user_mask_ptr: user-space pointer to the new cpu mask
4050 * Return: 0 on success. An error code otherwise.
4052 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4053 unsigned long __user *, user_mask_ptr)
4055 cpumask_var_t new_mask;
4058 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4061 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4063 retval = sched_setaffinity(pid, new_mask);
4064 free_cpumask_var(new_mask);
4068 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4070 struct task_struct *p;
4071 unsigned long flags;
4077 p = find_process_by_pid(pid);
4081 retval = security_task_getscheduler(p);
4085 raw_spin_lock_irqsave(&p->pi_lock, flags);
4086 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4087 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4096 * sys_sched_getaffinity - get the cpu affinity of a process
4097 * @pid: pid of the process
4098 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4099 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4101 * Return: 0 on success. An error code otherwise.
4103 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4104 unsigned long __user *, user_mask_ptr)
4109 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4111 if (len & (sizeof(unsigned long)-1))
4114 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4117 ret = sched_getaffinity(pid, mask);
4119 size_t retlen = min_t(size_t, len, cpumask_size());
4121 if (copy_to_user(user_mask_ptr, mask, retlen))
4126 free_cpumask_var(mask);
4132 * sys_sched_yield - yield the current processor to other threads.
4134 * This function yields the current CPU to other tasks. If there are no
4135 * other threads running on this CPU then this function will return.
4139 SYSCALL_DEFINE0(sched_yield)
4141 struct rq *rq = this_rq_lock();
4143 schedstat_inc(rq, yld_count);
4144 current->sched_class->yield_task(rq);
4147 * Since we are going to call schedule() anyway, there's
4148 * no need to preempt or enable interrupts:
4150 __release(rq->lock);
4151 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4152 do_raw_spin_unlock(&rq->lock);
4153 sched_preempt_enable_no_resched();
4160 static void __cond_resched(void)
4162 __preempt_count_add(PREEMPT_ACTIVE);
4164 __preempt_count_sub(PREEMPT_ACTIVE);
4167 int __sched _cond_resched(void)
4169 if (should_resched()) {
4175 EXPORT_SYMBOL(_cond_resched);
4178 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4179 * call schedule, and on return reacquire the lock.
4181 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4182 * operations here to prevent schedule() from being called twice (once via
4183 * spin_unlock(), once by hand).
4185 int __cond_resched_lock(spinlock_t *lock)
4187 int resched = should_resched();
4190 lockdep_assert_held(lock);
4192 if (spin_needbreak(lock) || resched) {
4203 EXPORT_SYMBOL(__cond_resched_lock);
4205 int __sched __cond_resched_softirq(void)
4207 BUG_ON(!in_softirq());
4209 if (should_resched()) {
4217 EXPORT_SYMBOL(__cond_resched_softirq);
4220 * yield - yield the current processor to other threads.
4222 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4224 * The scheduler is at all times free to pick the calling task as the most
4225 * eligible task to run, if removing the yield() call from your code breaks
4226 * it, its already broken.
4228 * Typical broken usage is:
4233 * where one assumes that yield() will let 'the other' process run that will
4234 * make event true. If the current task is a SCHED_FIFO task that will never
4235 * happen. Never use yield() as a progress guarantee!!
4237 * If you want to use yield() to wait for something, use wait_event().
4238 * If you want to use yield() to be 'nice' for others, use cond_resched().
4239 * If you still want to use yield(), do not!
4241 void __sched yield(void)
4243 set_current_state(TASK_RUNNING);
4246 EXPORT_SYMBOL(yield);
4249 * yield_to - yield the current processor to another thread in
4250 * your thread group, or accelerate that thread toward the
4251 * processor it's on.
4253 * @preempt: whether task preemption is allowed or not
4255 * It's the caller's job to ensure that the target task struct
4256 * can't go away on us before we can do any checks.
4259 * true (>0) if we indeed boosted the target task.
4260 * false (0) if we failed to boost the target.
4261 * -ESRCH if there's no task to yield to.
4263 int __sched yield_to(struct task_struct *p, bool preempt)
4265 struct task_struct *curr = current;
4266 struct rq *rq, *p_rq;
4267 unsigned long flags;
4270 local_irq_save(flags);
4276 * If we're the only runnable task on the rq and target rq also
4277 * has only one task, there's absolutely no point in yielding.
4279 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4284 double_rq_lock(rq, p_rq);
4285 if (task_rq(p) != p_rq) {
4286 double_rq_unlock(rq, p_rq);
4290 if (!curr->sched_class->yield_to_task)
4293 if (curr->sched_class != p->sched_class)
4296 if (task_running(p_rq, p) || p->state)
4299 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4301 schedstat_inc(rq, yld_count);
4303 * Make p's CPU reschedule; pick_next_entity takes care of
4306 if (preempt && rq != p_rq)
4311 double_rq_unlock(rq, p_rq);
4313 local_irq_restore(flags);
4320 EXPORT_SYMBOL_GPL(yield_to);
4323 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4324 * that process accounting knows that this is a task in IO wait state.
4326 void __sched io_schedule(void)
4328 struct rq *rq = raw_rq();
4330 delayacct_blkio_start();
4331 atomic_inc(&rq->nr_iowait);
4332 blk_flush_plug(current);
4333 current->in_iowait = 1;
4335 current->in_iowait = 0;
4336 atomic_dec(&rq->nr_iowait);
4337 delayacct_blkio_end();
4339 EXPORT_SYMBOL(io_schedule);
4341 long __sched io_schedule_timeout(long timeout)
4343 struct rq *rq = raw_rq();
4346 delayacct_blkio_start();
4347 atomic_inc(&rq->nr_iowait);
4348 blk_flush_plug(current);
4349 current->in_iowait = 1;
4350 ret = schedule_timeout(timeout);
4351 current->in_iowait = 0;
4352 atomic_dec(&rq->nr_iowait);
4353 delayacct_blkio_end();
4358 * sys_sched_get_priority_max - return maximum RT priority.
4359 * @policy: scheduling class.
4361 * Return: On success, this syscall returns the maximum
4362 * rt_priority that can be used by a given scheduling class.
4363 * On failure, a negative error code is returned.
4365 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4372 ret = MAX_USER_RT_PRIO-1;
4374 case SCHED_DEADLINE:
4385 * sys_sched_get_priority_min - return minimum RT priority.
4386 * @policy: scheduling class.
4388 * Return: On success, this syscall returns the minimum
4389 * rt_priority that can be used by a given scheduling class.
4390 * On failure, a negative error code is returned.
4392 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4401 case SCHED_DEADLINE:
4411 * sys_sched_rr_get_interval - return the default timeslice of a process.
4412 * @pid: pid of the process.
4413 * @interval: userspace pointer to the timeslice value.
4415 * this syscall writes the default timeslice value of a given process
4416 * into the user-space timespec buffer. A value of '0' means infinity.
4418 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4421 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4422 struct timespec __user *, interval)
4424 struct task_struct *p;
4425 unsigned int time_slice;
4426 unsigned long flags;
4436 p = find_process_by_pid(pid);
4440 retval = security_task_getscheduler(p);
4444 rq = task_rq_lock(p, &flags);
4446 if (p->sched_class->get_rr_interval)
4447 time_slice = p->sched_class->get_rr_interval(rq, p);
4448 task_rq_unlock(rq, p, &flags);
4451 jiffies_to_timespec(time_slice, &t);
4452 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4460 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4462 void sched_show_task(struct task_struct *p)
4464 unsigned long free = 0;
4468 state = p->state ? __ffs(p->state) + 1 : 0;
4469 printk(KERN_INFO "%-15.15s %c", p->comm,
4470 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4471 #if BITS_PER_LONG == 32
4472 if (state == TASK_RUNNING)
4473 printk(KERN_CONT " running ");
4475 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4477 if (state == TASK_RUNNING)
4478 printk(KERN_CONT " running task ");
4480 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4482 #ifdef CONFIG_DEBUG_STACK_USAGE
4483 free = stack_not_used(p);
4486 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4488 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4489 task_pid_nr(p), ppid,
4490 (unsigned long)task_thread_info(p)->flags);
4492 print_worker_info(KERN_INFO, p);
4493 show_stack(p, NULL);
4496 void show_state_filter(unsigned long state_filter)
4498 struct task_struct *g, *p;
4500 #if BITS_PER_LONG == 32
4502 " task PC stack pid father\n");
4505 " task PC stack pid father\n");
4508 do_each_thread(g, p) {
4510 * reset the NMI-timeout, listing all files on a slow
4511 * console might take a lot of time:
4513 touch_nmi_watchdog();
4514 if (!state_filter || (p->state & state_filter))
4516 } while_each_thread(g, p);
4518 touch_all_softlockup_watchdogs();
4520 #ifdef CONFIG_SCHED_DEBUG
4521 sysrq_sched_debug_show();
4525 * Only show locks if all tasks are dumped:
4528 debug_show_all_locks();
4531 void init_idle_bootup_task(struct task_struct *idle)
4533 idle->sched_class = &idle_sched_class;
4537 * init_idle - set up an idle thread for a given CPU
4538 * @idle: task in question
4539 * @cpu: cpu the idle task belongs to
4541 * NOTE: this function does not set the idle thread's NEED_RESCHED
4542 * flag, to make booting more robust.
4544 void init_idle(struct task_struct *idle, int cpu)
4546 struct rq *rq = cpu_rq(cpu);
4547 unsigned long flags;
4549 raw_spin_lock_irqsave(&rq->lock, flags);
4551 __sched_fork(0, idle);
4552 idle->state = TASK_RUNNING;
4553 idle->se.exec_start = sched_clock();
4555 do_set_cpus_allowed(idle, cpumask_of(cpu));
4557 * We're having a chicken and egg problem, even though we are
4558 * holding rq->lock, the cpu isn't yet set to this cpu so the
4559 * lockdep check in task_group() will fail.
4561 * Similar case to sched_fork(). / Alternatively we could
4562 * use task_rq_lock() here and obtain the other rq->lock.
4567 __set_task_cpu(idle, cpu);
4570 rq->curr = rq->idle = idle;
4572 #if defined(CONFIG_SMP)
4575 raw_spin_unlock_irqrestore(&rq->lock, flags);
4577 /* Set the preempt count _outside_ the spinlocks! */
4578 init_idle_preempt_count(idle, cpu);
4581 * The idle tasks have their own, simple scheduling class:
4583 idle->sched_class = &idle_sched_class;
4584 ftrace_graph_init_idle_task(idle, cpu);
4585 vtime_init_idle(idle, cpu);
4586 #if defined(CONFIG_SMP)
4587 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4592 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4594 if (p->sched_class && p->sched_class->set_cpus_allowed)
4595 p->sched_class->set_cpus_allowed(p, new_mask);
4597 cpumask_copy(&p->cpus_allowed, new_mask);
4598 p->nr_cpus_allowed = cpumask_weight(new_mask);
4602 * This is how migration works:
4604 * 1) we invoke migration_cpu_stop() on the target CPU using
4606 * 2) stopper starts to run (implicitly forcing the migrated thread
4608 * 3) it checks whether the migrated task is still in the wrong runqueue.
4609 * 4) if it's in the wrong runqueue then the migration thread removes
4610 * it and puts it into the right queue.
4611 * 5) stopper completes and stop_one_cpu() returns and the migration
4616 * Change a given task's CPU affinity. Migrate the thread to a
4617 * proper CPU and schedule it away if the CPU it's executing on
4618 * is removed from the allowed bitmask.
4620 * NOTE: the caller must have a valid reference to the task, the
4621 * task must not exit() & deallocate itself prematurely. The
4622 * call is not atomic; no spinlocks may be held.
4624 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4626 unsigned long flags;
4628 unsigned int dest_cpu;
4631 rq = task_rq_lock(p, &flags);
4633 if (cpumask_equal(&p->cpus_allowed, new_mask))
4636 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4641 do_set_cpus_allowed(p, new_mask);
4643 /* Can the task run on the task's current CPU? If so, we're done */
4644 if (cpumask_test_cpu(task_cpu(p), new_mask))
4647 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4649 struct migration_arg arg = { p, dest_cpu };
4650 /* Need help from migration thread: drop lock and wait. */
4651 task_rq_unlock(rq, p, &flags);
4652 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4653 tlb_migrate_finish(p->mm);
4657 task_rq_unlock(rq, p, &flags);
4661 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4664 * Move (not current) task off this cpu, onto dest cpu. We're doing
4665 * this because either it can't run here any more (set_cpus_allowed()
4666 * away from this CPU, or CPU going down), or because we're
4667 * attempting to rebalance this task on exec (sched_exec).
4669 * So we race with normal scheduler movements, but that's OK, as long
4670 * as the task is no longer on this CPU.
4672 * Returns non-zero if task was successfully migrated.
4674 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4676 struct rq *rq_dest, *rq_src;
4679 if (unlikely(!cpu_active(dest_cpu)))
4682 rq_src = cpu_rq(src_cpu);
4683 rq_dest = cpu_rq(dest_cpu);
4685 raw_spin_lock(&p->pi_lock);
4686 double_rq_lock(rq_src, rq_dest);
4687 /* Already moved. */
4688 if (task_cpu(p) != src_cpu)
4690 /* Affinity changed (again). */
4691 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4695 * If we're not on a rq, the next wake-up will ensure we're
4699 dequeue_task(rq_src, p, 0);
4700 set_task_cpu(p, dest_cpu);
4701 enqueue_task(rq_dest, p, 0);
4702 check_preempt_curr(rq_dest, p, 0);
4707 double_rq_unlock(rq_src, rq_dest);
4708 raw_spin_unlock(&p->pi_lock);
4712 #ifdef CONFIG_NUMA_BALANCING
4713 /* Migrate current task p to target_cpu */
4714 int migrate_task_to(struct task_struct *p, int target_cpu)
4716 struct migration_arg arg = { p, target_cpu };
4717 int curr_cpu = task_cpu(p);
4719 if (curr_cpu == target_cpu)
4722 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4725 /* TODO: This is not properly updating schedstats */
4727 trace_sched_move_numa(p, curr_cpu, target_cpu);
4728 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4732 * Requeue a task on a given node and accurately track the number of NUMA
4733 * tasks on the runqueues
4735 void sched_setnuma(struct task_struct *p, int nid)
4738 unsigned long flags;
4739 bool on_rq, running;
4741 rq = task_rq_lock(p, &flags);
4743 running = task_current(rq, p);
4746 dequeue_task(rq, p, 0);
4748 p->sched_class->put_prev_task(rq, p);
4750 p->numa_preferred_nid = nid;
4753 p->sched_class->set_curr_task(rq);
4755 enqueue_task(rq, p, 0);
4756 task_rq_unlock(rq, p, &flags);
4761 * migration_cpu_stop - this will be executed by a highprio stopper thread
4762 * and performs thread migration by bumping thread off CPU then
4763 * 'pushing' onto another runqueue.
4765 static int migration_cpu_stop(void *data)
4767 struct migration_arg *arg = data;
4770 * The original target cpu might have gone down and we might
4771 * be on another cpu but it doesn't matter.
4773 local_irq_disable();
4774 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4779 #ifdef CONFIG_HOTPLUG_CPU
4782 * Ensures that the idle task is using init_mm right before its cpu goes
4785 void idle_task_exit(void)
4787 struct mm_struct *mm = current->active_mm;
4789 BUG_ON(cpu_online(smp_processor_id()));
4791 if (mm != &init_mm) {
4792 switch_mm(mm, &init_mm, current);
4793 finish_arch_post_lock_switch();
4799 * Since this CPU is going 'away' for a while, fold any nr_active delta
4800 * we might have. Assumes we're called after migrate_tasks() so that the
4801 * nr_active count is stable.
4803 * Also see the comment "Global load-average calculations".
4805 static void calc_load_migrate(struct rq *rq)
4807 long delta = calc_load_fold_active(rq);
4809 atomic_long_add(delta, &calc_load_tasks);
4812 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4816 static const struct sched_class fake_sched_class = {
4817 .put_prev_task = put_prev_task_fake,
4820 static struct task_struct fake_task = {
4822 * Avoid pull_{rt,dl}_task()
4824 .prio = MAX_PRIO + 1,
4825 .sched_class = &fake_sched_class,
4829 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4830 * try_to_wake_up()->select_task_rq().
4832 * Called with rq->lock held even though we'er in stop_machine() and
4833 * there's no concurrency possible, we hold the required locks anyway
4834 * because of lock validation efforts.
4836 static void migrate_tasks(unsigned int dead_cpu)
4838 struct rq *rq = cpu_rq(dead_cpu);
4839 struct task_struct *next, *stop = rq->stop;
4843 * Fudge the rq selection such that the below task selection loop
4844 * doesn't get stuck on the currently eligible stop task.
4846 * We're currently inside stop_machine() and the rq is either stuck
4847 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4848 * either way we should never end up calling schedule() until we're
4854 * put_prev_task() and pick_next_task() sched
4855 * class method both need to have an up-to-date
4856 * value of rq->clock[_task]
4858 update_rq_clock(rq);
4862 * There's this thread running, bail when that's the only
4865 if (rq->nr_running == 1)
4868 next = pick_next_task(rq, &fake_task);
4870 next->sched_class->put_prev_task(rq, next);
4872 /* Find suitable destination for @next, with force if needed. */
4873 dest_cpu = select_fallback_rq(dead_cpu, next);
4874 raw_spin_unlock(&rq->lock);
4876 __migrate_task(next, dead_cpu, dest_cpu);
4878 raw_spin_lock(&rq->lock);
4884 #endif /* CONFIG_HOTPLUG_CPU */
4886 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4888 static struct ctl_table sd_ctl_dir[] = {
4890 .procname = "sched_domain",
4896 static struct ctl_table sd_ctl_root[] = {
4898 .procname = "kernel",
4900 .child = sd_ctl_dir,
4905 static struct ctl_table *sd_alloc_ctl_entry(int n)
4907 struct ctl_table *entry =
4908 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4913 static void sd_free_ctl_entry(struct ctl_table **tablep)
4915 struct ctl_table *entry;
4918 * In the intermediate directories, both the child directory and
4919 * procname are dynamically allocated and could fail but the mode
4920 * will always be set. In the lowest directory the names are
4921 * static strings and all have proc handlers.
4923 for (entry = *tablep; entry->mode; entry++) {
4925 sd_free_ctl_entry(&entry->child);
4926 if (entry->proc_handler == NULL)
4927 kfree(entry->procname);
4934 static int min_load_idx = 0;
4935 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4938 set_table_entry(struct ctl_table *entry,
4939 const char *procname, void *data, int maxlen,
4940 umode_t mode, proc_handler *proc_handler,
4943 entry->procname = procname;
4945 entry->maxlen = maxlen;
4947 entry->proc_handler = proc_handler;
4950 entry->extra1 = &min_load_idx;
4951 entry->extra2 = &max_load_idx;
4955 static struct ctl_table *
4956 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4958 struct ctl_table *table = sd_alloc_ctl_entry(14);
4963 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4964 sizeof(long), 0644, proc_doulongvec_minmax, false);
4965 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4966 sizeof(long), 0644, proc_doulongvec_minmax, false);
4967 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4968 sizeof(int), 0644, proc_dointvec_minmax, true);
4969 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4970 sizeof(int), 0644, proc_dointvec_minmax, true);
4971 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4972 sizeof(int), 0644, proc_dointvec_minmax, true);
4973 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4974 sizeof(int), 0644, proc_dointvec_minmax, true);
4975 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4976 sizeof(int), 0644, proc_dointvec_minmax, true);
4977 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4978 sizeof(int), 0644, proc_dointvec_minmax, false);
4979 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4980 sizeof(int), 0644, proc_dointvec_minmax, false);
4981 set_table_entry(&table[9], "cache_nice_tries",
4982 &sd->cache_nice_tries,
4983 sizeof(int), 0644, proc_dointvec_minmax, false);
4984 set_table_entry(&table[10], "flags", &sd->flags,
4985 sizeof(int), 0644, proc_dointvec_minmax, false);
4986 set_table_entry(&table[11], "max_newidle_lb_cost",
4987 &sd->max_newidle_lb_cost,
4988 sizeof(long), 0644, proc_doulongvec_minmax, false);
4989 set_table_entry(&table[12], "name", sd->name,
4990 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4991 /* &table[13] is terminator */
4996 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4998 struct ctl_table *entry, *table;
4999 struct sched_domain *sd;
5000 int domain_num = 0, i;
5003 for_each_domain(cpu, sd)
5005 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5010 for_each_domain(cpu, sd) {
5011 snprintf(buf, 32, "domain%d", i);
5012 entry->procname = kstrdup(buf, GFP_KERNEL);
5014 entry->child = sd_alloc_ctl_domain_table(sd);
5021 static struct ctl_table_header *sd_sysctl_header;
5022 static void register_sched_domain_sysctl(void)
5024 int i, cpu_num = num_possible_cpus();
5025 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5028 WARN_ON(sd_ctl_dir[0].child);
5029 sd_ctl_dir[0].child = entry;
5034 for_each_possible_cpu(i) {
5035 snprintf(buf, 32, "cpu%d", i);
5036 entry->procname = kstrdup(buf, GFP_KERNEL);
5038 entry->child = sd_alloc_ctl_cpu_table(i);
5042 WARN_ON(sd_sysctl_header);
5043 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5046 /* may be called multiple times per register */
5047 static void unregister_sched_domain_sysctl(void)
5049 if (sd_sysctl_header)
5050 unregister_sysctl_table(sd_sysctl_header);
5051 sd_sysctl_header = NULL;
5052 if (sd_ctl_dir[0].child)
5053 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5056 static void register_sched_domain_sysctl(void)
5059 static void unregister_sched_domain_sysctl(void)
5064 static void set_rq_online(struct rq *rq)
5067 const struct sched_class *class;
5069 cpumask_set_cpu(rq->cpu, rq->rd->online);
5072 for_each_class(class) {
5073 if (class->rq_online)
5074 class->rq_online(rq);
5079 static void set_rq_offline(struct rq *rq)
5082 const struct sched_class *class;
5084 for_each_class(class) {
5085 if (class->rq_offline)
5086 class->rq_offline(rq);
5089 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5095 * migration_call - callback that gets triggered when a CPU is added.
5096 * Here we can start up the necessary migration thread for the new CPU.
5099 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5101 int cpu = (long)hcpu;
5102 unsigned long flags;
5103 struct rq *rq = cpu_rq(cpu);
5105 switch (action & ~CPU_TASKS_FROZEN) {
5107 case CPU_UP_PREPARE:
5108 rq->calc_load_update = calc_load_update;
5112 /* Update our root-domain */
5113 raw_spin_lock_irqsave(&rq->lock, flags);
5115 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5119 raw_spin_unlock_irqrestore(&rq->lock, flags);
5122 #ifdef CONFIG_HOTPLUG_CPU
5124 sched_ttwu_pending();
5125 /* Update our root-domain */
5126 raw_spin_lock_irqsave(&rq->lock, flags);
5128 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5132 BUG_ON(rq->nr_running != 1); /* the migration thread */
5133 raw_spin_unlock_irqrestore(&rq->lock, flags);
5137 calc_load_migrate(rq);
5142 update_max_interval();
5148 * Register at high priority so that task migration (migrate_all_tasks)
5149 * happens before everything else. This has to be lower priority than
5150 * the notifier in the perf_event subsystem, though.
5152 static struct notifier_block migration_notifier = {
5153 .notifier_call = migration_call,
5154 .priority = CPU_PRI_MIGRATION,
5157 static void __cpuinit set_cpu_rq_start_time(void)
5159 int cpu = smp_processor_id();
5160 struct rq *rq = cpu_rq(cpu);
5161 rq->age_stamp = sched_clock_cpu(cpu);
5164 static int sched_cpu_active(struct notifier_block *nfb,
5165 unsigned long action, void *hcpu)
5167 switch (action & ~CPU_TASKS_FROZEN) {
5169 set_cpu_rq_start_time();
5171 case CPU_DOWN_FAILED:
5172 set_cpu_active((long)hcpu, true);
5179 static int sched_cpu_inactive(struct notifier_block *nfb,
5180 unsigned long action, void *hcpu)
5182 unsigned long flags;
5183 long cpu = (long)hcpu;
5185 switch (action & ~CPU_TASKS_FROZEN) {
5186 case CPU_DOWN_PREPARE:
5187 set_cpu_active(cpu, false);
5189 /* explicitly allow suspend */
5190 if (!(action & CPU_TASKS_FROZEN)) {
5191 struct dl_bw *dl_b = dl_bw_of(cpu);
5195 raw_spin_lock_irqsave(&dl_b->lock, flags);
5196 cpus = dl_bw_cpus(cpu);
5197 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5198 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5201 return notifier_from_errno(-EBUSY);
5209 static int __init migration_init(void)
5211 void *cpu = (void *)(long)smp_processor_id();
5214 /* Initialize migration for the boot CPU */
5215 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5216 BUG_ON(err == NOTIFY_BAD);
5217 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5218 register_cpu_notifier(&migration_notifier);
5220 /* Register cpu active notifiers */
5221 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5222 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5226 early_initcall(migration_init);
5231 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5233 #ifdef CONFIG_SCHED_DEBUG
5235 static __read_mostly int sched_debug_enabled;
5237 static int __init sched_debug_setup(char *str)
5239 sched_debug_enabled = 1;
5243 early_param("sched_debug", sched_debug_setup);
5245 static inline bool sched_debug(void)
5247 return sched_debug_enabled;
5250 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5251 struct cpumask *groupmask)
5253 struct sched_group *group = sd->groups;
5256 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5257 cpumask_clear(groupmask);
5259 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5261 if (!(sd->flags & SD_LOAD_BALANCE)) {
5262 printk("does not load-balance\n");
5264 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5269 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5271 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5272 printk(KERN_ERR "ERROR: domain->span does not contain "
5275 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5276 printk(KERN_ERR "ERROR: domain->groups does not contain"
5280 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5284 printk(KERN_ERR "ERROR: group is NULL\n");
5289 * Even though we initialize ->capacity to something semi-sane,
5290 * we leave capacity_orig unset. This allows us to detect if
5291 * domain iteration is still funny without causing /0 traps.
5293 if (!group->sgc->capacity_orig) {
5294 printk(KERN_CONT "\n");
5295 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5299 if (!cpumask_weight(sched_group_cpus(group))) {
5300 printk(KERN_CONT "\n");
5301 printk(KERN_ERR "ERROR: empty group\n");
5305 if (!(sd->flags & SD_OVERLAP) &&
5306 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5307 printk(KERN_CONT "\n");
5308 printk(KERN_ERR "ERROR: repeated CPUs\n");
5312 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5314 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5316 printk(KERN_CONT " %s", str);
5317 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5318 printk(KERN_CONT " (cpu_capacity = %d)",
5319 group->sgc->capacity);
5322 group = group->next;
5323 } while (group != sd->groups);
5324 printk(KERN_CONT "\n");
5326 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5327 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5330 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5331 printk(KERN_ERR "ERROR: parent span is not a superset "
5332 "of domain->span\n");
5336 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5340 if (!sched_debug_enabled)
5344 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5348 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5351 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5359 #else /* !CONFIG_SCHED_DEBUG */
5360 # define sched_domain_debug(sd, cpu) do { } while (0)
5361 static inline bool sched_debug(void)
5365 #endif /* CONFIG_SCHED_DEBUG */
5367 static int sd_degenerate(struct sched_domain *sd)
5369 if (cpumask_weight(sched_domain_span(sd)) == 1)
5372 /* Following flags need at least 2 groups */
5373 if (sd->flags & (SD_LOAD_BALANCE |
5374 SD_BALANCE_NEWIDLE |
5377 SD_SHARE_CPUCAPACITY |
5378 SD_SHARE_PKG_RESOURCES |
5379 SD_SHARE_POWERDOMAIN)) {
5380 if (sd->groups != sd->groups->next)
5384 /* Following flags don't use groups */
5385 if (sd->flags & (SD_WAKE_AFFINE))
5392 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5394 unsigned long cflags = sd->flags, pflags = parent->flags;
5396 if (sd_degenerate(parent))
5399 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5402 /* Flags needing groups don't count if only 1 group in parent */
5403 if (parent->groups == parent->groups->next) {
5404 pflags &= ~(SD_LOAD_BALANCE |
5405 SD_BALANCE_NEWIDLE |
5408 SD_SHARE_CPUCAPACITY |
5409 SD_SHARE_PKG_RESOURCES |
5411 SD_SHARE_POWERDOMAIN);
5412 if (nr_node_ids == 1)
5413 pflags &= ~SD_SERIALIZE;
5415 if (~cflags & pflags)
5421 static void free_rootdomain(struct rcu_head *rcu)
5423 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5425 cpupri_cleanup(&rd->cpupri);
5426 cpudl_cleanup(&rd->cpudl);
5427 free_cpumask_var(rd->dlo_mask);
5428 free_cpumask_var(rd->rto_mask);
5429 free_cpumask_var(rd->online);
5430 free_cpumask_var(rd->span);
5434 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5436 struct root_domain *old_rd = NULL;
5437 unsigned long flags;
5439 raw_spin_lock_irqsave(&rq->lock, flags);
5444 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5447 cpumask_clear_cpu(rq->cpu, old_rd->span);
5450 * If we dont want to free the old_rd yet then
5451 * set old_rd to NULL to skip the freeing later
5454 if (!atomic_dec_and_test(&old_rd->refcount))
5458 atomic_inc(&rd->refcount);
5461 cpumask_set_cpu(rq->cpu, rd->span);
5462 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5465 raw_spin_unlock_irqrestore(&rq->lock, flags);
5468 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5471 static int init_rootdomain(struct root_domain *rd)
5473 memset(rd, 0, sizeof(*rd));
5475 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5477 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5479 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5481 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5484 init_dl_bw(&rd->dl_bw);
5485 if (cpudl_init(&rd->cpudl) != 0)
5488 if (cpupri_init(&rd->cpupri) != 0)
5493 free_cpumask_var(rd->rto_mask);
5495 free_cpumask_var(rd->dlo_mask);
5497 free_cpumask_var(rd->online);
5499 free_cpumask_var(rd->span);
5505 * By default the system creates a single root-domain with all cpus as
5506 * members (mimicking the global state we have today).
5508 struct root_domain def_root_domain;
5510 static void init_defrootdomain(void)
5512 init_rootdomain(&def_root_domain);
5514 atomic_set(&def_root_domain.refcount, 1);
5517 static struct root_domain *alloc_rootdomain(void)
5519 struct root_domain *rd;
5521 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5525 if (init_rootdomain(rd) != 0) {
5533 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5535 struct sched_group *tmp, *first;
5544 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5549 } while (sg != first);
5552 static void free_sched_domain(struct rcu_head *rcu)
5554 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5557 * If its an overlapping domain it has private groups, iterate and
5560 if (sd->flags & SD_OVERLAP) {
5561 free_sched_groups(sd->groups, 1);
5562 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5563 kfree(sd->groups->sgc);
5569 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5571 call_rcu(&sd->rcu, free_sched_domain);
5574 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5576 for (; sd; sd = sd->parent)
5577 destroy_sched_domain(sd, cpu);
5581 * Keep a special pointer to the highest sched_domain that has
5582 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5583 * allows us to avoid some pointer chasing select_idle_sibling().
5585 * Also keep a unique ID per domain (we use the first cpu number in
5586 * the cpumask of the domain), this allows us to quickly tell if
5587 * two cpus are in the same cache domain, see cpus_share_cache().
5589 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5590 DEFINE_PER_CPU(int, sd_llc_size);
5591 DEFINE_PER_CPU(int, sd_llc_id);
5592 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5593 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5594 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5596 static void update_top_cache_domain(int cpu)
5598 struct sched_domain *sd;
5599 struct sched_domain *busy_sd = NULL;
5603 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5605 id = cpumask_first(sched_domain_span(sd));
5606 size = cpumask_weight(sched_domain_span(sd));
5607 busy_sd = sd->parent; /* sd_busy */
5609 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5611 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5612 per_cpu(sd_llc_size, cpu) = size;
5613 per_cpu(sd_llc_id, cpu) = id;
5615 sd = lowest_flag_domain(cpu, SD_NUMA);
5616 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5618 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5619 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5623 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5624 * hold the hotplug lock.
5627 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5629 struct rq *rq = cpu_rq(cpu);
5630 struct sched_domain *tmp;
5632 /* Remove the sched domains which do not contribute to scheduling. */
5633 for (tmp = sd; tmp; ) {
5634 struct sched_domain *parent = tmp->parent;
5638 if (sd_parent_degenerate(tmp, parent)) {
5639 tmp->parent = parent->parent;
5641 parent->parent->child = tmp;
5643 * Transfer SD_PREFER_SIBLING down in case of a
5644 * degenerate parent; the spans match for this
5645 * so the property transfers.
5647 if (parent->flags & SD_PREFER_SIBLING)
5648 tmp->flags |= SD_PREFER_SIBLING;
5649 destroy_sched_domain(parent, cpu);
5654 if (sd && sd_degenerate(sd)) {
5657 destroy_sched_domain(tmp, cpu);
5662 sched_domain_debug(sd, cpu);
5664 rq_attach_root(rq, rd);
5666 rcu_assign_pointer(rq->sd, sd);
5667 destroy_sched_domains(tmp, cpu);
5669 update_top_cache_domain(cpu);
5672 /* cpus with isolated domains */
5673 static cpumask_var_t cpu_isolated_map;
5675 /* Setup the mask of cpus configured for isolated domains */
5676 static int __init isolated_cpu_setup(char *str)
5678 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5679 cpulist_parse(str, cpu_isolated_map);
5683 __setup("isolcpus=", isolated_cpu_setup);
5686 struct sched_domain ** __percpu sd;
5687 struct root_domain *rd;
5698 * Build an iteration mask that can exclude certain CPUs from the upwards
5701 * Asymmetric node setups can result in situations where the domain tree is of
5702 * unequal depth, make sure to skip domains that already cover the entire
5705 * In that case build_sched_domains() will have terminated the iteration early
5706 * and our sibling sd spans will be empty. Domains should always include the
5707 * cpu they're built on, so check that.
5710 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5712 const struct cpumask *span = sched_domain_span(sd);
5713 struct sd_data *sdd = sd->private;
5714 struct sched_domain *sibling;
5717 for_each_cpu(i, span) {
5718 sibling = *per_cpu_ptr(sdd->sd, i);
5719 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5722 cpumask_set_cpu(i, sched_group_mask(sg));
5727 * Return the canonical balance cpu for this group, this is the first cpu
5728 * of this group that's also in the iteration mask.
5730 int group_balance_cpu(struct sched_group *sg)
5732 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5736 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5738 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5739 const struct cpumask *span = sched_domain_span(sd);
5740 struct cpumask *covered = sched_domains_tmpmask;
5741 struct sd_data *sdd = sd->private;
5742 struct sched_domain *child;
5745 cpumask_clear(covered);
5747 for_each_cpu(i, span) {
5748 struct cpumask *sg_span;
5750 if (cpumask_test_cpu(i, covered))
5753 child = *per_cpu_ptr(sdd->sd, i);
5755 /* See the comment near build_group_mask(). */
5756 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5759 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5760 GFP_KERNEL, cpu_to_node(cpu));
5765 sg_span = sched_group_cpus(sg);
5767 child = child->child;
5768 cpumask_copy(sg_span, sched_domain_span(child));
5770 cpumask_set_cpu(i, sg_span);
5772 cpumask_or(covered, covered, sg_span);
5774 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5775 if (atomic_inc_return(&sg->sgc->ref) == 1)
5776 build_group_mask(sd, sg);
5779 * Initialize sgc->capacity such that even if we mess up the
5780 * domains and no possible iteration will get us here, we won't
5783 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5784 sg->sgc->capacity_orig = sg->sgc->capacity;
5787 * Make sure the first group of this domain contains the
5788 * canonical balance cpu. Otherwise the sched_domain iteration
5789 * breaks. See update_sg_lb_stats().
5791 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5792 group_balance_cpu(sg) == cpu)
5802 sd->groups = groups;
5807 free_sched_groups(first, 0);
5812 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5814 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5815 struct sched_domain *child = sd->child;
5818 cpu = cpumask_first(sched_domain_span(child));
5821 *sg = *per_cpu_ptr(sdd->sg, cpu);
5822 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5823 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5830 * build_sched_groups will build a circular linked list of the groups
5831 * covered by the given span, and will set each group's ->cpumask correctly,
5832 * and ->cpu_capacity to 0.
5834 * Assumes the sched_domain tree is fully constructed
5837 build_sched_groups(struct sched_domain *sd, int cpu)
5839 struct sched_group *first = NULL, *last = NULL;
5840 struct sd_data *sdd = sd->private;
5841 const struct cpumask *span = sched_domain_span(sd);
5842 struct cpumask *covered;
5845 get_group(cpu, sdd, &sd->groups);
5846 atomic_inc(&sd->groups->ref);
5848 if (cpu != cpumask_first(span))
5851 lockdep_assert_held(&sched_domains_mutex);
5852 covered = sched_domains_tmpmask;
5854 cpumask_clear(covered);
5856 for_each_cpu(i, span) {
5857 struct sched_group *sg;
5860 if (cpumask_test_cpu(i, covered))
5863 group = get_group(i, sdd, &sg);
5864 cpumask_setall(sched_group_mask(sg));
5866 for_each_cpu(j, span) {
5867 if (get_group(j, sdd, NULL) != group)
5870 cpumask_set_cpu(j, covered);
5871 cpumask_set_cpu(j, sched_group_cpus(sg));
5886 * Initialize sched groups cpu_capacity.
5888 * cpu_capacity indicates the capacity of sched group, which is used while
5889 * distributing the load between different sched groups in a sched domain.
5890 * Typically cpu_capacity for all the groups in a sched domain will be same
5891 * unless there are asymmetries in the topology. If there are asymmetries,
5892 * group having more cpu_capacity will pickup more load compared to the
5893 * group having less cpu_capacity.
5895 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5897 struct sched_group *sg = sd->groups;
5902 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5904 } while (sg != sd->groups);
5906 if (cpu != group_balance_cpu(sg))
5909 update_group_capacity(sd, cpu);
5910 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5914 * Initializers for schedule domains
5915 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5918 static int default_relax_domain_level = -1;
5919 int sched_domain_level_max;
5921 static int __init setup_relax_domain_level(char *str)
5923 if (kstrtoint(str, 0, &default_relax_domain_level))
5924 pr_warn("Unable to set relax_domain_level\n");
5928 __setup("relax_domain_level=", setup_relax_domain_level);
5930 static void set_domain_attribute(struct sched_domain *sd,
5931 struct sched_domain_attr *attr)
5935 if (!attr || attr->relax_domain_level < 0) {
5936 if (default_relax_domain_level < 0)
5939 request = default_relax_domain_level;
5941 request = attr->relax_domain_level;
5942 if (request < sd->level) {
5943 /* turn off idle balance on this domain */
5944 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5946 /* turn on idle balance on this domain */
5947 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5951 static void __sdt_free(const struct cpumask *cpu_map);
5952 static int __sdt_alloc(const struct cpumask *cpu_map);
5954 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5955 const struct cpumask *cpu_map)
5959 if (!atomic_read(&d->rd->refcount))
5960 free_rootdomain(&d->rd->rcu); /* fall through */
5962 free_percpu(d->sd); /* fall through */
5964 __sdt_free(cpu_map); /* fall through */
5970 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5971 const struct cpumask *cpu_map)
5973 memset(d, 0, sizeof(*d));
5975 if (__sdt_alloc(cpu_map))
5976 return sa_sd_storage;
5977 d->sd = alloc_percpu(struct sched_domain *);
5979 return sa_sd_storage;
5980 d->rd = alloc_rootdomain();
5983 return sa_rootdomain;
5987 * NULL the sd_data elements we've used to build the sched_domain and
5988 * sched_group structure so that the subsequent __free_domain_allocs()
5989 * will not free the data we're using.
5991 static void claim_allocations(int cpu, struct sched_domain *sd)
5993 struct sd_data *sdd = sd->private;
5995 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5996 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5998 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5999 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6001 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6002 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6006 static int sched_domains_numa_levels;
6007 static int *sched_domains_numa_distance;
6008 static struct cpumask ***sched_domains_numa_masks;
6009 static int sched_domains_curr_level;
6013 * SD_flags allowed in topology descriptions.
6015 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6016 * SD_SHARE_PKG_RESOURCES - describes shared caches
6017 * SD_NUMA - describes NUMA topologies
6018 * SD_SHARE_POWERDOMAIN - describes shared power domain
6021 * SD_ASYM_PACKING - describes SMT quirks
6023 #define TOPOLOGY_SD_FLAGS \
6024 (SD_SHARE_CPUCAPACITY | \
6025 SD_SHARE_PKG_RESOURCES | \
6028 SD_SHARE_POWERDOMAIN)
6030 static struct sched_domain *
6031 sd_init(struct sched_domain_topology_level *tl, int cpu)
6033 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6034 int sd_weight, sd_flags = 0;
6038 * Ugly hack to pass state to sd_numa_mask()...
6040 sched_domains_curr_level = tl->numa_level;
6043 sd_weight = cpumask_weight(tl->mask(cpu));
6046 sd_flags = (*tl->sd_flags)();
6047 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6048 "wrong sd_flags in topology description\n"))
6049 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6051 *sd = (struct sched_domain){
6052 .min_interval = sd_weight,
6053 .max_interval = 2*sd_weight,
6055 .imbalance_pct = 125,
6057 .cache_nice_tries = 0,
6064 .flags = 1*SD_LOAD_BALANCE
6065 | 1*SD_BALANCE_NEWIDLE
6070 | 0*SD_SHARE_CPUCAPACITY
6071 | 0*SD_SHARE_PKG_RESOURCES
6073 | 0*SD_PREFER_SIBLING
6078 .last_balance = jiffies,
6079 .balance_interval = sd_weight,
6081 .max_newidle_lb_cost = 0,
6082 .next_decay_max_lb_cost = jiffies,
6083 #ifdef CONFIG_SCHED_DEBUG
6089 * Convert topological properties into behaviour.
6092 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6093 sd->imbalance_pct = 110;
6094 sd->smt_gain = 1178; /* ~15% */
6096 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6097 sd->imbalance_pct = 117;
6098 sd->cache_nice_tries = 1;
6102 } else if (sd->flags & SD_NUMA) {
6103 sd->cache_nice_tries = 2;
6107 sd->flags |= SD_SERIALIZE;
6108 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6109 sd->flags &= ~(SD_BALANCE_EXEC |
6116 sd->flags |= SD_PREFER_SIBLING;
6117 sd->cache_nice_tries = 1;
6122 sd->private = &tl->data;
6128 * Topology list, bottom-up.
6130 static struct sched_domain_topology_level default_topology[] = {
6131 #ifdef CONFIG_SCHED_SMT
6132 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6134 #ifdef CONFIG_SCHED_MC
6135 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6137 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6141 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6143 #define for_each_sd_topology(tl) \
6144 for (tl = sched_domain_topology; tl->mask; tl++)
6146 void set_sched_topology(struct sched_domain_topology_level *tl)
6148 sched_domain_topology = tl;
6153 static const struct cpumask *sd_numa_mask(int cpu)
6155 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6158 static void sched_numa_warn(const char *str)
6160 static int done = false;
6168 printk(KERN_WARNING "ERROR: %s\n\n", str);
6170 for (i = 0; i < nr_node_ids; i++) {
6171 printk(KERN_WARNING " ");
6172 for (j = 0; j < nr_node_ids; j++)
6173 printk(KERN_CONT "%02d ", node_distance(i,j));
6174 printk(KERN_CONT "\n");
6176 printk(KERN_WARNING "\n");
6179 static bool find_numa_distance(int distance)
6183 if (distance == node_distance(0, 0))
6186 for (i = 0; i < sched_domains_numa_levels; i++) {
6187 if (sched_domains_numa_distance[i] == distance)
6194 static void sched_init_numa(void)
6196 int next_distance, curr_distance = node_distance(0, 0);
6197 struct sched_domain_topology_level *tl;
6201 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6202 if (!sched_domains_numa_distance)
6206 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6207 * unique distances in the node_distance() table.
6209 * Assumes node_distance(0,j) includes all distances in
6210 * node_distance(i,j) in order to avoid cubic time.
6212 next_distance = curr_distance;
6213 for (i = 0; i < nr_node_ids; i++) {
6214 for (j = 0; j < nr_node_ids; j++) {
6215 for (k = 0; k < nr_node_ids; k++) {
6216 int distance = node_distance(i, k);
6218 if (distance > curr_distance &&
6219 (distance < next_distance ||
6220 next_distance == curr_distance))
6221 next_distance = distance;
6224 * While not a strong assumption it would be nice to know
6225 * about cases where if node A is connected to B, B is not
6226 * equally connected to A.
6228 if (sched_debug() && node_distance(k, i) != distance)
6229 sched_numa_warn("Node-distance not symmetric");
6231 if (sched_debug() && i && !find_numa_distance(distance))
6232 sched_numa_warn("Node-0 not representative");
6234 if (next_distance != curr_distance) {
6235 sched_domains_numa_distance[level++] = next_distance;
6236 sched_domains_numa_levels = level;
6237 curr_distance = next_distance;
6242 * In case of sched_debug() we verify the above assumption.
6248 * 'level' contains the number of unique distances, excluding the
6249 * identity distance node_distance(i,i).
6251 * The sched_domains_numa_distance[] array includes the actual distance
6256 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6257 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6258 * the array will contain less then 'level' members. This could be
6259 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6260 * in other functions.
6262 * We reset it to 'level' at the end of this function.
6264 sched_domains_numa_levels = 0;
6266 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6267 if (!sched_domains_numa_masks)
6271 * Now for each level, construct a mask per node which contains all
6272 * cpus of nodes that are that many hops away from us.
6274 for (i = 0; i < level; i++) {
6275 sched_domains_numa_masks[i] =
6276 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6277 if (!sched_domains_numa_masks[i])
6280 for (j = 0; j < nr_node_ids; j++) {
6281 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6285 sched_domains_numa_masks[i][j] = mask;
6287 for (k = 0; k < nr_node_ids; k++) {
6288 if (node_distance(j, k) > sched_domains_numa_distance[i])
6291 cpumask_or(mask, mask, cpumask_of_node(k));
6296 /* Compute default topology size */
6297 for (i = 0; sched_domain_topology[i].mask; i++);
6299 tl = kzalloc((i + level + 1) *
6300 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6305 * Copy the default topology bits..
6307 for (i = 0; sched_domain_topology[i].mask; i++)
6308 tl[i] = sched_domain_topology[i];
6311 * .. and append 'j' levels of NUMA goodness.
6313 for (j = 0; j < level; i++, j++) {
6314 tl[i] = (struct sched_domain_topology_level){
6315 .mask = sd_numa_mask,
6316 .sd_flags = cpu_numa_flags,
6317 .flags = SDTL_OVERLAP,
6323 sched_domain_topology = tl;
6325 sched_domains_numa_levels = level;
6328 static void sched_domains_numa_masks_set(int cpu)
6331 int node = cpu_to_node(cpu);
6333 for (i = 0; i < sched_domains_numa_levels; i++) {
6334 for (j = 0; j < nr_node_ids; j++) {
6335 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6336 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6341 static void sched_domains_numa_masks_clear(int cpu)
6344 for (i = 0; i < sched_domains_numa_levels; i++) {
6345 for (j = 0; j < nr_node_ids; j++)
6346 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6351 * Update sched_domains_numa_masks[level][node] array when new cpus
6354 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6355 unsigned long action,
6358 int cpu = (long)hcpu;
6360 switch (action & ~CPU_TASKS_FROZEN) {
6362 sched_domains_numa_masks_set(cpu);
6366 sched_domains_numa_masks_clear(cpu);
6376 static inline void sched_init_numa(void)
6380 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6381 unsigned long action,
6386 #endif /* CONFIG_NUMA */
6388 static int __sdt_alloc(const struct cpumask *cpu_map)
6390 struct sched_domain_topology_level *tl;
6393 for_each_sd_topology(tl) {
6394 struct sd_data *sdd = &tl->data;
6396 sdd->sd = alloc_percpu(struct sched_domain *);
6400 sdd->sg = alloc_percpu(struct sched_group *);
6404 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6408 for_each_cpu(j, cpu_map) {
6409 struct sched_domain *sd;
6410 struct sched_group *sg;
6411 struct sched_group_capacity *sgc;
6413 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6414 GFP_KERNEL, cpu_to_node(j));
6418 *per_cpu_ptr(sdd->sd, j) = sd;
6420 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6421 GFP_KERNEL, cpu_to_node(j));
6427 *per_cpu_ptr(sdd->sg, j) = sg;
6429 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6430 GFP_KERNEL, cpu_to_node(j));
6434 *per_cpu_ptr(sdd->sgc, j) = sgc;
6441 static void __sdt_free(const struct cpumask *cpu_map)
6443 struct sched_domain_topology_level *tl;
6446 for_each_sd_topology(tl) {
6447 struct sd_data *sdd = &tl->data;
6449 for_each_cpu(j, cpu_map) {
6450 struct sched_domain *sd;
6453 sd = *per_cpu_ptr(sdd->sd, j);
6454 if (sd && (sd->flags & SD_OVERLAP))
6455 free_sched_groups(sd->groups, 0);
6456 kfree(*per_cpu_ptr(sdd->sd, j));
6460 kfree(*per_cpu_ptr(sdd->sg, j));
6462 kfree(*per_cpu_ptr(sdd->sgc, j));
6464 free_percpu(sdd->sd);
6466 free_percpu(sdd->sg);
6468 free_percpu(sdd->sgc);
6473 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6474 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6475 struct sched_domain *child, int cpu)
6477 struct sched_domain *sd = sd_init(tl, cpu);
6481 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6483 sd->level = child->level + 1;
6484 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6488 if (!cpumask_subset(sched_domain_span(child),
6489 sched_domain_span(sd))) {
6490 pr_err("BUG: arch topology borken\n");
6491 #ifdef CONFIG_SCHED_DEBUG
6492 pr_err(" the %s domain not a subset of the %s domain\n",
6493 child->name, sd->name);
6495 /* Fixup, ensure @sd has at least @child cpus. */
6496 cpumask_or(sched_domain_span(sd),
6497 sched_domain_span(sd),
6498 sched_domain_span(child));
6502 set_domain_attribute(sd, attr);
6508 * Build sched domains for a given set of cpus and attach the sched domains
6509 * to the individual cpus
6511 static int build_sched_domains(const struct cpumask *cpu_map,
6512 struct sched_domain_attr *attr)
6514 enum s_alloc alloc_state;
6515 struct sched_domain *sd;
6517 int i, ret = -ENOMEM;
6519 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6520 if (alloc_state != sa_rootdomain)
6523 /* Set up domains for cpus specified by the cpu_map. */
6524 for_each_cpu(i, cpu_map) {
6525 struct sched_domain_topology_level *tl;
6528 for_each_sd_topology(tl) {
6529 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6530 if (tl == sched_domain_topology)
6531 *per_cpu_ptr(d.sd, i) = sd;
6532 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6533 sd->flags |= SD_OVERLAP;
6534 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6539 /* Build the groups for the domains */
6540 for_each_cpu(i, cpu_map) {
6541 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6542 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6543 if (sd->flags & SD_OVERLAP) {
6544 if (build_overlap_sched_groups(sd, i))
6547 if (build_sched_groups(sd, i))
6553 /* Calculate CPU capacity for physical packages and nodes */
6554 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6555 if (!cpumask_test_cpu(i, cpu_map))
6558 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6559 claim_allocations(i, sd);
6560 init_sched_groups_capacity(i, sd);
6564 /* Attach the domains */
6566 for_each_cpu(i, cpu_map) {
6567 sd = *per_cpu_ptr(d.sd, i);
6568 cpu_attach_domain(sd, d.rd, i);
6574 __free_domain_allocs(&d, alloc_state, cpu_map);
6578 static cpumask_var_t *doms_cur; /* current sched domains */
6579 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6580 static struct sched_domain_attr *dattr_cur;
6581 /* attribues of custom domains in 'doms_cur' */
6584 * Special case: If a kmalloc of a doms_cur partition (array of
6585 * cpumask) fails, then fallback to a single sched domain,
6586 * as determined by the single cpumask fallback_doms.
6588 static cpumask_var_t fallback_doms;
6591 * arch_update_cpu_topology lets virtualized architectures update the
6592 * cpu core maps. It is supposed to return 1 if the topology changed
6593 * or 0 if it stayed the same.
6595 int __weak arch_update_cpu_topology(void)
6600 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6603 cpumask_var_t *doms;
6605 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6608 for (i = 0; i < ndoms; i++) {
6609 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6610 free_sched_domains(doms, i);
6617 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6620 for (i = 0; i < ndoms; i++)
6621 free_cpumask_var(doms[i]);
6626 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6627 * For now this just excludes isolated cpus, but could be used to
6628 * exclude other special cases in the future.
6630 static int init_sched_domains(const struct cpumask *cpu_map)
6634 arch_update_cpu_topology();
6636 doms_cur = alloc_sched_domains(ndoms_cur);
6638 doms_cur = &fallback_doms;
6639 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6640 err = build_sched_domains(doms_cur[0], NULL);
6641 register_sched_domain_sysctl();
6647 * Detach sched domains from a group of cpus specified in cpu_map
6648 * These cpus will now be attached to the NULL domain
6650 static void detach_destroy_domains(const struct cpumask *cpu_map)
6655 for_each_cpu(i, cpu_map)
6656 cpu_attach_domain(NULL, &def_root_domain, i);
6660 /* handle null as "default" */
6661 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6662 struct sched_domain_attr *new, int idx_new)
6664 struct sched_domain_attr tmp;
6671 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6672 new ? (new + idx_new) : &tmp,
6673 sizeof(struct sched_domain_attr));
6677 * Partition sched domains as specified by the 'ndoms_new'
6678 * cpumasks in the array doms_new[] of cpumasks. This compares
6679 * doms_new[] to the current sched domain partitioning, doms_cur[].
6680 * It destroys each deleted domain and builds each new domain.
6682 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6683 * The masks don't intersect (don't overlap.) We should setup one
6684 * sched domain for each mask. CPUs not in any of the cpumasks will
6685 * not be load balanced. If the same cpumask appears both in the
6686 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6689 * The passed in 'doms_new' should be allocated using
6690 * alloc_sched_domains. This routine takes ownership of it and will
6691 * free_sched_domains it when done with it. If the caller failed the
6692 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6693 * and partition_sched_domains() will fallback to the single partition
6694 * 'fallback_doms', it also forces the domains to be rebuilt.
6696 * If doms_new == NULL it will be replaced with cpu_online_mask.
6697 * ndoms_new == 0 is a special case for destroying existing domains,
6698 * and it will not create the default domain.
6700 * Call with hotplug lock held
6702 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6703 struct sched_domain_attr *dattr_new)
6708 mutex_lock(&sched_domains_mutex);
6710 /* always unregister in case we don't destroy any domains */
6711 unregister_sched_domain_sysctl();
6713 /* Let architecture update cpu core mappings. */
6714 new_topology = arch_update_cpu_topology();
6716 n = doms_new ? ndoms_new : 0;
6718 /* Destroy deleted domains */
6719 for (i = 0; i < ndoms_cur; i++) {
6720 for (j = 0; j < n && !new_topology; j++) {
6721 if (cpumask_equal(doms_cur[i], doms_new[j])
6722 && dattrs_equal(dattr_cur, i, dattr_new, j))
6725 /* no match - a current sched domain not in new doms_new[] */
6726 detach_destroy_domains(doms_cur[i]);
6732 if (doms_new == NULL) {
6734 doms_new = &fallback_doms;
6735 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6736 WARN_ON_ONCE(dattr_new);
6739 /* Build new domains */
6740 for (i = 0; i < ndoms_new; i++) {
6741 for (j = 0; j < n && !new_topology; j++) {
6742 if (cpumask_equal(doms_new[i], doms_cur[j])
6743 && dattrs_equal(dattr_new, i, dattr_cur, j))
6746 /* no match - add a new doms_new */
6747 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6752 /* Remember the new sched domains */
6753 if (doms_cur != &fallback_doms)
6754 free_sched_domains(doms_cur, ndoms_cur);
6755 kfree(dattr_cur); /* kfree(NULL) is safe */
6756 doms_cur = doms_new;
6757 dattr_cur = dattr_new;
6758 ndoms_cur = ndoms_new;
6760 register_sched_domain_sysctl();
6762 mutex_unlock(&sched_domains_mutex);
6765 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6768 * Update cpusets according to cpu_active mask. If cpusets are
6769 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6770 * around partition_sched_domains().
6772 * If we come here as part of a suspend/resume, don't touch cpusets because we
6773 * want to restore it back to its original state upon resume anyway.
6775 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6779 case CPU_ONLINE_FROZEN:
6780 case CPU_DOWN_FAILED_FROZEN:
6783 * num_cpus_frozen tracks how many CPUs are involved in suspend
6784 * resume sequence. As long as this is not the last online
6785 * operation in the resume sequence, just build a single sched
6786 * domain, ignoring cpusets.
6789 if (likely(num_cpus_frozen)) {
6790 partition_sched_domains(1, NULL, NULL);
6795 * This is the last CPU online operation. So fall through and
6796 * restore the original sched domains by considering the
6797 * cpuset configurations.
6801 case CPU_DOWN_FAILED:
6802 cpuset_update_active_cpus(true);
6810 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6814 case CPU_DOWN_PREPARE:
6815 cpuset_update_active_cpus(false);
6817 case CPU_DOWN_PREPARE_FROZEN:
6819 partition_sched_domains(1, NULL, NULL);
6827 void __init sched_init_smp(void)
6829 cpumask_var_t non_isolated_cpus;
6831 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6832 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6837 * There's no userspace yet to cause hotplug operations; hence all the
6838 * cpu masks are stable and all blatant races in the below code cannot
6841 mutex_lock(&sched_domains_mutex);
6842 init_sched_domains(cpu_active_mask);
6843 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6844 if (cpumask_empty(non_isolated_cpus))
6845 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6846 mutex_unlock(&sched_domains_mutex);
6848 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6849 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6850 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6854 /* Move init over to a non-isolated CPU */
6855 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6857 sched_init_granularity();
6858 free_cpumask_var(non_isolated_cpus);
6860 init_sched_rt_class();
6861 init_sched_dl_class();
6864 void __init sched_init_smp(void)
6866 sched_init_granularity();
6868 #endif /* CONFIG_SMP */
6870 const_debug unsigned int sysctl_timer_migration = 1;
6872 int in_sched_functions(unsigned long addr)
6874 return in_lock_functions(addr) ||
6875 (addr >= (unsigned long)__sched_text_start
6876 && addr < (unsigned long)__sched_text_end);
6879 #ifdef CONFIG_CGROUP_SCHED
6881 * Default task group.
6882 * Every task in system belongs to this group at bootup.
6884 struct task_group root_task_group;
6885 LIST_HEAD(task_groups);
6888 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6890 void __init sched_init(void)
6893 unsigned long alloc_size = 0, ptr;
6895 #ifdef CONFIG_FAIR_GROUP_SCHED
6896 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6898 #ifdef CONFIG_RT_GROUP_SCHED
6899 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6901 #ifdef CONFIG_CPUMASK_OFFSTACK
6902 alloc_size += num_possible_cpus() * cpumask_size();
6905 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6907 #ifdef CONFIG_FAIR_GROUP_SCHED
6908 root_task_group.se = (struct sched_entity **)ptr;
6909 ptr += nr_cpu_ids * sizeof(void **);
6911 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6912 ptr += nr_cpu_ids * sizeof(void **);
6914 #endif /* CONFIG_FAIR_GROUP_SCHED */
6915 #ifdef CONFIG_RT_GROUP_SCHED
6916 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6917 ptr += nr_cpu_ids * sizeof(void **);
6919 root_task_group.rt_rq = (struct rt_rq **)ptr;
6920 ptr += nr_cpu_ids * sizeof(void **);
6922 #endif /* CONFIG_RT_GROUP_SCHED */
6923 #ifdef CONFIG_CPUMASK_OFFSTACK
6924 for_each_possible_cpu(i) {
6925 per_cpu(load_balance_mask, i) = (void *)ptr;
6926 ptr += cpumask_size();
6928 #endif /* CONFIG_CPUMASK_OFFSTACK */
6931 init_rt_bandwidth(&def_rt_bandwidth,
6932 global_rt_period(), global_rt_runtime());
6933 init_dl_bandwidth(&def_dl_bandwidth,
6934 global_rt_period(), global_rt_runtime());
6937 init_defrootdomain();
6940 #ifdef CONFIG_RT_GROUP_SCHED
6941 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6942 global_rt_period(), global_rt_runtime());
6943 #endif /* CONFIG_RT_GROUP_SCHED */
6945 #ifdef CONFIG_CGROUP_SCHED
6946 list_add(&root_task_group.list, &task_groups);
6947 INIT_LIST_HEAD(&root_task_group.children);
6948 INIT_LIST_HEAD(&root_task_group.siblings);
6949 autogroup_init(&init_task);
6951 #endif /* CONFIG_CGROUP_SCHED */
6953 for_each_possible_cpu(i) {
6957 raw_spin_lock_init(&rq->lock);
6959 rq->calc_load_active = 0;
6960 rq->calc_load_update = jiffies + LOAD_FREQ;
6961 init_cfs_rq(&rq->cfs);
6962 init_rt_rq(&rq->rt, rq);
6963 init_dl_rq(&rq->dl, rq);
6964 #ifdef CONFIG_FAIR_GROUP_SCHED
6965 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6966 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6968 * How much cpu bandwidth does root_task_group get?
6970 * In case of task-groups formed thr' the cgroup filesystem, it
6971 * gets 100% of the cpu resources in the system. This overall
6972 * system cpu resource is divided among the tasks of
6973 * root_task_group and its child task-groups in a fair manner,
6974 * based on each entity's (task or task-group's) weight
6975 * (se->load.weight).
6977 * In other words, if root_task_group has 10 tasks of weight
6978 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6979 * then A0's share of the cpu resource is:
6981 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6983 * We achieve this by letting root_task_group's tasks sit
6984 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6986 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6987 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6988 #endif /* CONFIG_FAIR_GROUP_SCHED */
6990 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6991 #ifdef CONFIG_RT_GROUP_SCHED
6992 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6995 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6996 rq->cpu_load[j] = 0;
6998 rq->last_load_update_tick = jiffies;
7003 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7004 rq->post_schedule = 0;
7005 rq->active_balance = 0;
7006 rq->next_balance = jiffies;
7011 rq->avg_idle = 2*sysctl_sched_migration_cost;
7012 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7014 INIT_LIST_HEAD(&rq->cfs_tasks);
7016 rq_attach_root(rq, &def_root_domain);
7017 #ifdef CONFIG_NO_HZ_COMMON
7020 #ifdef CONFIG_NO_HZ_FULL
7021 rq->last_sched_tick = 0;
7025 atomic_set(&rq->nr_iowait, 0);
7028 set_load_weight(&init_task);
7030 #ifdef CONFIG_PREEMPT_NOTIFIERS
7031 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7035 * The boot idle thread does lazy MMU switching as well:
7037 atomic_inc(&init_mm.mm_count);
7038 enter_lazy_tlb(&init_mm, current);
7041 * Make us the idle thread. Technically, schedule() should not be
7042 * called from this thread, however somewhere below it might be,
7043 * but because we are the idle thread, we just pick up running again
7044 * when this runqueue becomes "idle".
7046 init_idle(current, smp_processor_id());
7048 calc_load_update = jiffies + LOAD_FREQ;
7051 * During early bootup we pretend to be a normal task:
7053 current->sched_class = &fair_sched_class;
7056 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7057 /* May be allocated at isolcpus cmdline parse time */
7058 if (cpu_isolated_map == NULL)
7059 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7060 idle_thread_set_boot_cpu();
7061 set_cpu_rq_start_time();
7063 init_sched_fair_class();
7065 scheduler_running = 1;
7068 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7069 static inline int preempt_count_equals(int preempt_offset)
7071 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7073 return (nested == preempt_offset);
7076 void __might_sleep(const char *file, int line, int preempt_offset)
7078 static unsigned long prev_jiffy; /* ratelimiting */
7080 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7081 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7082 !is_idle_task(current)) ||
7083 system_state != SYSTEM_RUNNING || oops_in_progress)
7085 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7087 prev_jiffy = jiffies;
7090 "BUG: sleeping function called from invalid context at %s:%d\n",
7093 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7094 in_atomic(), irqs_disabled(),
7095 current->pid, current->comm);
7097 debug_show_held_locks(current);
7098 if (irqs_disabled())
7099 print_irqtrace_events(current);
7100 #ifdef CONFIG_DEBUG_PREEMPT
7101 if (!preempt_count_equals(preempt_offset)) {
7102 pr_err("Preemption disabled at:");
7103 print_ip_sym(current->preempt_disable_ip);
7109 EXPORT_SYMBOL(__might_sleep);
7112 #ifdef CONFIG_MAGIC_SYSRQ
7113 static void normalize_task(struct rq *rq, struct task_struct *p)
7115 const struct sched_class *prev_class = p->sched_class;
7116 struct sched_attr attr = {
7117 .sched_policy = SCHED_NORMAL,
7119 int old_prio = p->prio;
7124 dequeue_task(rq, p, 0);
7125 __setscheduler(rq, p, &attr);
7127 enqueue_task(rq, p, 0);
7131 check_class_changed(rq, p, prev_class, old_prio);
7134 void normalize_rt_tasks(void)
7136 struct task_struct *g, *p;
7137 unsigned long flags;
7140 read_lock_irqsave(&tasklist_lock, flags);
7141 do_each_thread(g, p) {
7143 * Only normalize user tasks:
7148 p->se.exec_start = 0;
7149 #ifdef CONFIG_SCHEDSTATS
7150 p->se.statistics.wait_start = 0;
7151 p->se.statistics.sleep_start = 0;
7152 p->se.statistics.block_start = 0;
7155 if (!dl_task(p) && !rt_task(p)) {
7157 * Renice negative nice level userspace
7160 if (task_nice(p) < 0 && p->mm)
7161 set_user_nice(p, 0);
7165 raw_spin_lock(&p->pi_lock);
7166 rq = __task_rq_lock(p);
7168 normalize_task(rq, p);
7170 __task_rq_unlock(rq);
7171 raw_spin_unlock(&p->pi_lock);
7172 } while_each_thread(g, p);
7174 read_unlock_irqrestore(&tasklist_lock, flags);
7177 #endif /* CONFIG_MAGIC_SYSRQ */
7179 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7181 * These functions are only useful for the IA64 MCA handling, or kdb.
7183 * They can only be called when the whole system has been
7184 * stopped - every CPU needs to be quiescent, and no scheduling
7185 * activity can take place. Using them for anything else would
7186 * be a serious bug, and as a result, they aren't even visible
7187 * under any other configuration.
7191 * curr_task - return the current task for a given cpu.
7192 * @cpu: the processor in question.
7194 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7196 * Return: The current task for @cpu.
7198 struct task_struct *curr_task(int cpu)
7200 return cpu_curr(cpu);
7203 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7207 * set_curr_task - set the current task for a given cpu.
7208 * @cpu: the processor in question.
7209 * @p: the task pointer to set.
7211 * Description: This function must only be used when non-maskable interrupts
7212 * are serviced on a separate stack. It allows the architecture to switch the
7213 * notion of the current task on a cpu in a non-blocking manner. This function
7214 * must be called with all CPU's synchronized, and interrupts disabled, the
7215 * and caller must save the original value of the current task (see
7216 * curr_task() above) and restore that value before reenabling interrupts and
7217 * re-starting the system.
7219 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7221 void set_curr_task(int cpu, struct task_struct *p)
7228 #ifdef CONFIG_CGROUP_SCHED
7229 /* task_group_lock serializes the addition/removal of task groups */
7230 static DEFINE_SPINLOCK(task_group_lock);
7232 static void free_sched_group(struct task_group *tg)
7234 free_fair_sched_group(tg);
7235 free_rt_sched_group(tg);
7240 /* allocate runqueue etc for a new task group */
7241 struct task_group *sched_create_group(struct task_group *parent)
7243 struct task_group *tg;
7245 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7247 return ERR_PTR(-ENOMEM);
7249 if (!alloc_fair_sched_group(tg, parent))
7252 if (!alloc_rt_sched_group(tg, parent))
7258 free_sched_group(tg);
7259 return ERR_PTR(-ENOMEM);
7262 void sched_online_group(struct task_group *tg, struct task_group *parent)
7264 unsigned long flags;
7266 spin_lock_irqsave(&task_group_lock, flags);
7267 list_add_rcu(&tg->list, &task_groups);
7269 WARN_ON(!parent); /* root should already exist */
7271 tg->parent = parent;
7272 INIT_LIST_HEAD(&tg->children);
7273 list_add_rcu(&tg->siblings, &parent->children);
7274 spin_unlock_irqrestore(&task_group_lock, flags);
7277 /* rcu callback to free various structures associated with a task group */
7278 static void free_sched_group_rcu(struct rcu_head *rhp)
7280 /* now it should be safe to free those cfs_rqs */
7281 free_sched_group(container_of(rhp, struct task_group, rcu));
7284 /* Destroy runqueue etc associated with a task group */
7285 void sched_destroy_group(struct task_group *tg)
7287 /* wait for possible concurrent references to cfs_rqs complete */
7288 call_rcu(&tg->rcu, free_sched_group_rcu);
7291 void sched_offline_group(struct task_group *tg)
7293 unsigned long flags;
7296 /* end participation in shares distribution */
7297 for_each_possible_cpu(i)
7298 unregister_fair_sched_group(tg, i);
7300 spin_lock_irqsave(&task_group_lock, flags);
7301 list_del_rcu(&tg->list);
7302 list_del_rcu(&tg->siblings);
7303 spin_unlock_irqrestore(&task_group_lock, flags);
7306 /* change task's runqueue when it moves between groups.
7307 * The caller of this function should have put the task in its new group
7308 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7309 * reflect its new group.
7311 void sched_move_task(struct task_struct *tsk)
7313 struct task_group *tg;
7315 unsigned long flags;
7318 rq = task_rq_lock(tsk, &flags);
7320 running = task_current(rq, tsk);
7324 dequeue_task(rq, tsk, 0);
7325 if (unlikely(running))
7326 tsk->sched_class->put_prev_task(rq, tsk);
7328 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7329 lockdep_is_held(&tsk->sighand->siglock)),
7330 struct task_group, css);
7331 tg = autogroup_task_group(tsk, tg);
7332 tsk->sched_task_group = tg;
7334 #ifdef CONFIG_FAIR_GROUP_SCHED
7335 if (tsk->sched_class->task_move_group)
7336 tsk->sched_class->task_move_group(tsk, on_rq);
7339 set_task_rq(tsk, task_cpu(tsk));
7341 if (unlikely(running))
7342 tsk->sched_class->set_curr_task(rq);
7344 enqueue_task(rq, tsk, 0);
7346 task_rq_unlock(rq, tsk, &flags);
7348 #endif /* CONFIG_CGROUP_SCHED */
7350 #ifdef CONFIG_RT_GROUP_SCHED
7352 * Ensure that the real time constraints are schedulable.
7354 static DEFINE_MUTEX(rt_constraints_mutex);
7356 /* Must be called with tasklist_lock held */
7357 static inline int tg_has_rt_tasks(struct task_group *tg)
7359 struct task_struct *g, *p;
7361 do_each_thread(g, p) {
7362 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7364 } while_each_thread(g, p);
7369 struct rt_schedulable_data {
7370 struct task_group *tg;
7375 static int tg_rt_schedulable(struct task_group *tg, void *data)
7377 struct rt_schedulable_data *d = data;
7378 struct task_group *child;
7379 unsigned long total, sum = 0;
7380 u64 period, runtime;
7382 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7383 runtime = tg->rt_bandwidth.rt_runtime;
7386 period = d->rt_period;
7387 runtime = d->rt_runtime;
7391 * Cannot have more runtime than the period.
7393 if (runtime > period && runtime != RUNTIME_INF)
7397 * Ensure we don't starve existing RT tasks.
7399 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7402 total = to_ratio(period, runtime);
7405 * Nobody can have more than the global setting allows.
7407 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7411 * The sum of our children's runtime should not exceed our own.
7413 list_for_each_entry_rcu(child, &tg->children, siblings) {
7414 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7415 runtime = child->rt_bandwidth.rt_runtime;
7417 if (child == d->tg) {
7418 period = d->rt_period;
7419 runtime = d->rt_runtime;
7422 sum += to_ratio(period, runtime);
7431 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7435 struct rt_schedulable_data data = {
7437 .rt_period = period,
7438 .rt_runtime = runtime,
7442 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7448 static int tg_set_rt_bandwidth(struct task_group *tg,
7449 u64 rt_period, u64 rt_runtime)
7453 mutex_lock(&rt_constraints_mutex);
7454 read_lock(&tasklist_lock);
7455 err = __rt_schedulable(tg, rt_period, rt_runtime);
7459 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7460 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7461 tg->rt_bandwidth.rt_runtime = rt_runtime;
7463 for_each_possible_cpu(i) {
7464 struct rt_rq *rt_rq = tg->rt_rq[i];
7466 raw_spin_lock(&rt_rq->rt_runtime_lock);
7467 rt_rq->rt_runtime = rt_runtime;
7468 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7470 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7472 read_unlock(&tasklist_lock);
7473 mutex_unlock(&rt_constraints_mutex);
7478 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7480 u64 rt_runtime, rt_period;
7482 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7483 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7484 if (rt_runtime_us < 0)
7485 rt_runtime = RUNTIME_INF;
7487 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7490 static long sched_group_rt_runtime(struct task_group *tg)
7494 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7497 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7498 do_div(rt_runtime_us, NSEC_PER_USEC);
7499 return rt_runtime_us;
7502 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7504 u64 rt_runtime, rt_period;
7506 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7507 rt_runtime = tg->rt_bandwidth.rt_runtime;
7512 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7515 static long sched_group_rt_period(struct task_group *tg)
7519 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7520 do_div(rt_period_us, NSEC_PER_USEC);
7521 return rt_period_us;
7523 #endif /* CONFIG_RT_GROUP_SCHED */
7525 #ifdef CONFIG_RT_GROUP_SCHED
7526 static int sched_rt_global_constraints(void)
7530 mutex_lock(&rt_constraints_mutex);
7531 read_lock(&tasklist_lock);
7532 ret = __rt_schedulable(NULL, 0, 0);
7533 read_unlock(&tasklist_lock);
7534 mutex_unlock(&rt_constraints_mutex);
7539 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7541 /* Don't accept realtime tasks when there is no way for them to run */
7542 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7548 #else /* !CONFIG_RT_GROUP_SCHED */
7549 static int sched_rt_global_constraints(void)
7551 unsigned long flags;
7554 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7555 for_each_possible_cpu(i) {
7556 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7558 raw_spin_lock(&rt_rq->rt_runtime_lock);
7559 rt_rq->rt_runtime = global_rt_runtime();
7560 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7562 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7566 #endif /* CONFIG_RT_GROUP_SCHED */
7568 static int sched_dl_global_constraints(void)
7570 u64 runtime = global_rt_runtime();
7571 u64 period = global_rt_period();
7572 u64 new_bw = to_ratio(period, runtime);
7574 unsigned long flags;
7577 * Here we want to check the bandwidth not being set to some
7578 * value smaller than the currently allocated bandwidth in
7579 * any of the root_domains.
7581 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7582 * cycling on root_domains... Discussion on different/better
7583 * solutions is welcome!
7585 for_each_possible_cpu(cpu) {
7586 struct dl_bw *dl_b = dl_bw_of(cpu);
7588 raw_spin_lock_irqsave(&dl_b->lock, flags);
7589 if (new_bw < dl_b->total_bw)
7591 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7600 static void sched_dl_do_global(void)
7604 unsigned long flags;
7606 def_dl_bandwidth.dl_period = global_rt_period();
7607 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7609 if (global_rt_runtime() != RUNTIME_INF)
7610 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7613 * FIXME: As above...
7615 for_each_possible_cpu(cpu) {
7616 struct dl_bw *dl_b = dl_bw_of(cpu);
7618 raw_spin_lock_irqsave(&dl_b->lock, flags);
7620 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7624 static int sched_rt_global_validate(void)
7626 if (sysctl_sched_rt_period <= 0)
7629 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7630 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7636 static void sched_rt_do_global(void)
7638 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7639 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7642 int sched_rt_handler(struct ctl_table *table, int write,
7643 void __user *buffer, size_t *lenp,
7646 int old_period, old_runtime;
7647 static DEFINE_MUTEX(mutex);
7651 old_period = sysctl_sched_rt_period;
7652 old_runtime = sysctl_sched_rt_runtime;
7654 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7656 if (!ret && write) {
7657 ret = sched_rt_global_validate();
7661 ret = sched_rt_global_constraints();
7665 ret = sched_dl_global_constraints();
7669 sched_rt_do_global();
7670 sched_dl_do_global();
7674 sysctl_sched_rt_period = old_period;
7675 sysctl_sched_rt_runtime = old_runtime;
7677 mutex_unlock(&mutex);
7682 int sched_rr_handler(struct ctl_table *table, int write,
7683 void __user *buffer, size_t *lenp,
7687 static DEFINE_MUTEX(mutex);
7690 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7691 /* make sure that internally we keep jiffies */
7692 /* also, writing zero resets timeslice to default */
7693 if (!ret && write) {
7694 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7695 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7697 mutex_unlock(&mutex);
7701 #ifdef CONFIG_CGROUP_SCHED
7703 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7705 return css ? container_of(css, struct task_group, css) : NULL;
7708 static struct cgroup_subsys_state *
7709 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7711 struct task_group *parent = css_tg(parent_css);
7712 struct task_group *tg;
7715 /* This is early initialization for the top cgroup */
7716 return &root_task_group.css;
7719 tg = sched_create_group(parent);
7721 return ERR_PTR(-ENOMEM);
7726 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7728 struct task_group *tg = css_tg(css);
7729 struct task_group *parent = css_tg(css->parent);
7732 sched_online_group(tg, parent);
7736 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7738 struct task_group *tg = css_tg(css);
7740 sched_destroy_group(tg);
7743 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7745 struct task_group *tg = css_tg(css);
7747 sched_offline_group(tg);
7750 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7751 struct cgroup_taskset *tset)
7753 struct task_struct *task;
7755 cgroup_taskset_for_each(task, tset) {
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 if (!sched_rt_can_attach(css_tg(css), task))
7760 /* We don't support RT-tasks being in separate groups */
7761 if (task->sched_class != &fair_sched_class)
7768 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7769 struct cgroup_taskset *tset)
7771 struct task_struct *task;
7773 cgroup_taskset_for_each(task, tset)
7774 sched_move_task(task);
7777 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7778 struct cgroup_subsys_state *old_css,
7779 struct task_struct *task)
7782 * cgroup_exit() is called in the copy_process() failure path.
7783 * Ignore this case since the task hasn't ran yet, this avoids
7784 * trying to poke a half freed task state from generic code.
7786 if (!(task->flags & PF_EXITING))
7789 sched_move_task(task);
7792 #ifdef CONFIG_FAIR_GROUP_SCHED
7793 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7794 struct cftype *cftype, u64 shareval)
7796 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7799 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7802 struct task_group *tg = css_tg(css);
7804 return (u64) scale_load_down(tg->shares);
7807 #ifdef CONFIG_CFS_BANDWIDTH
7808 static DEFINE_MUTEX(cfs_constraints_mutex);
7810 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7811 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7813 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7815 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7817 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7818 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7820 if (tg == &root_task_group)
7824 * Ensure we have at some amount of bandwidth every period. This is
7825 * to prevent reaching a state of large arrears when throttled via
7826 * entity_tick() resulting in prolonged exit starvation.
7828 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7832 * Likewise, bound things on the otherside by preventing insane quota
7833 * periods. This also allows us to normalize in computing quota
7836 if (period > max_cfs_quota_period)
7840 * Prevent race between setting of cfs_rq->runtime_enabled and
7841 * unthrottle_offline_cfs_rqs().
7844 mutex_lock(&cfs_constraints_mutex);
7845 ret = __cfs_schedulable(tg, period, quota);
7849 runtime_enabled = quota != RUNTIME_INF;
7850 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7852 * If we need to toggle cfs_bandwidth_used, off->on must occur
7853 * before making related changes, and on->off must occur afterwards
7855 if (runtime_enabled && !runtime_was_enabled)
7856 cfs_bandwidth_usage_inc();
7857 raw_spin_lock_irq(&cfs_b->lock);
7858 cfs_b->period = ns_to_ktime(period);
7859 cfs_b->quota = quota;
7861 __refill_cfs_bandwidth_runtime(cfs_b);
7862 /* restart the period timer (if active) to handle new period expiry */
7863 if (runtime_enabled && cfs_b->timer_active) {
7864 /* force a reprogram */
7865 __start_cfs_bandwidth(cfs_b, true);
7867 raw_spin_unlock_irq(&cfs_b->lock);
7869 for_each_online_cpu(i) {
7870 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7871 struct rq *rq = cfs_rq->rq;
7873 raw_spin_lock_irq(&rq->lock);
7874 cfs_rq->runtime_enabled = runtime_enabled;
7875 cfs_rq->runtime_remaining = 0;
7877 if (cfs_rq->throttled)
7878 unthrottle_cfs_rq(cfs_rq);
7879 raw_spin_unlock_irq(&rq->lock);
7881 if (runtime_was_enabled && !runtime_enabled)
7882 cfs_bandwidth_usage_dec();
7884 mutex_unlock(&cfs_constraints_mutex);
7890 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7894 period = ktime_to_ns(tg->cfs_bandwidth.period);
7895 if (cfs_quota_us < 0)
7896 quota = RUNTIME_INF;
7898 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7900 return tg_set_cfs_bandwidth(tg, period, quota);
7903 long tg_get_cfs_quota(struct task_group *tg)
7907 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7910 quota_us = tg->cfs_bandwidth.quota;
7911 do_div(quota_us, NSEC_PER_USEC);
7916 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7920 period = (u64)cfs_period_us * NSEC_PER_USEC;
7921 quota = tg->cfs_bandwidth.quota;
7923 return tg_set_cfs_bandwidth(tg, period, quota);
7926 long tg_get_cfs_period(struct task_group *tg)
7930 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7931 do_div(cfs_period_us, NSEC_PER_USEC);
7933 return cfs_period_us;
7936 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7939 return tg_get_cfs_quota(css_tg(css));
7942 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7943 struct cftype *cftype, s64 cfs_quota_us)
7945 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7948 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7951 return tg_get_cfs_period(css_tg(css));
7954 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7955 struct cftype *cftype, u64 cfs_period_us)
7957 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7960 struct cfs_schedulable_data {
7961 struct task_group *tg;
7966 * normalize group quota/period to be quota/max_period
7967 * note: units are usecs
7969 static u64 normalize_cfs_quota(struct task_group *tg,
7970 struct cfs_schedulable_data *d)
7978 period = tg_get_cfs_period(tg);
7979 quota = tg_get_cfs_quota(tg);
7982 /* note: these should typically be equivalent */
7983 if (quota == RUNTIME_INF || quota == -1)
7986 return to_ratio(period, quota);
7989 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7991 struct cfs_schedulable_data *d = data;
7992 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7993 s64 quota = 0, parent_quota = -1;
7996 quota = RUNTIME_INF;
7998 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8000 quota = normalize_cfs_quota(tg, d);
8001 parent_quota = parent_b->hierarchal_quota;
8004 * ensure max(child_quota) <= parent_quota, inherit when no
8007 if (quota == RUNTIME_INF)
8008 quota = parent_quota;
8009 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8012 cfs_b->hierarchal_quota = quota;
8017 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8020 struct cfs_schedulable_data data = {
8026 if (quota != RUNTIME_INF) {
8027 do_div(data.period, NSEC_PER_USEC);
8028 do_div(data.quota, NSEC_PER_USEC);
8032 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8038 static int cpu_stats_show(struct seq_file *sf, void *v)
8040 struct task_group *tg = css_tg(seq_css(sf));
8041 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8043 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8044 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8045 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8049 #endif /* CONFIG_CFS_BANDWIDTH */
8050 #endif /* CONFIG_FAIR_GROUP_SCHED */
8052 #ifdef CONFIG_RT_GROUP_SCHED
8053 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8054 struct cftype *cft, s64 val)
8056 return sched_group_set_rt_runtime(css_tg(css), val);
8059 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8062 return sched_group_rt_runtime(css_tg(css));
8065 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8066 struct cftype *cftype, u64 rt_period_us)
8068 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8071 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8074 return sched_group_rt_period(css_tg(css));
8076 #endif /* CONFIG_RT_GROUP_SCHED */
8078 static struct cftype cpu_files[] = {
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8082 .read_u64 = cpu_shares_read_u64,
8083 .write_u64 = cpu_shares_write_u64,
8086 #ifdef CONFIG_CFS_BANDWIDTH
8088 .name = "cfs_quota_us",
8089 .read_s64 = cpu_cfs_quota_read_s64,
8090 .write_s64 = cpu_cfs_quota_write_s64,
8093 .name = "cfs_period_us",
8094 .read_u64 = cpu_cfs_period_read_u64,
8095 .write_u64 = cpu_cfs_period_write_u64,
8099 .seq_show = cpu_stats_show,
8102 #ifdef CONFIG_RT_GROUP_SCHED
8104 .name = "rt_runtime_us",
8105 .read_s64 = cpu_rt_runtime_read,
8106 .write_s64 = cpu_rt_runtime_write,
8109 .name = "rt_period_us",
8110 .read_u64 = cpu_rt_period_read_uint,
8111 .write_u64 = cpu_rt_period_write_uint,
8117 struct cgroup_subsys cpu_cgrp_subsys = {
8118 .css_alloc = cpu_cgroup_css_alloc,
8119 .css_free = cpu_cgroup_css_free,
8120 .css_online = cpu_cgroup_css_online,
8121 .css_offline = cpu_cgroup_css_offline,
8122 .can_attach = cpu_cgroup_can_attach,
8123 .attach = cpu_cgroup_attach,
8124 .exit = cpu_cgroup_exit,
8125 .legacy_cftypes = cpu_files,
8129 #endif /* CONFIG_CGROUP_SCHED */
8131 void dump_cpu_task(int cpu)
8133 pr_info("Task dump for CPU %d:\n", cpu);
8134 sched_show_task(cpu_curr(cpu));