4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
143 update_rq_clock_task(rq, delta);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug unsigned int sysctl_sched_features =
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file *m, void *v)
173 for (i = 0; i < __SCHED_FEAT_NR; i++) {
174 if (!(sysctl_sched_features & (1UL << i)))
176 seq_printf(m, "%s ", sched_feat_names[i]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
192 #include "features.h"
197 static void sched_feat_disable(int i)
199 if (static_key_enabled(&sched_feat_keys[i]))
200 static_key_slow_dec(&sched_feat_keys[i]);
203 static void sched_feat_enable(int i)
205 if (!static_key_enabled(&sched_feat_keys[i]))
206 static_key_slow_inc(&sched_feat_keys[i]);
209 static void sched_feat_disable(int i) { };
210 static void sched_feat_enable(int i) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp)
218 if (strncmp(cmp, "NO_", 3) == 0) {
223 for (i = 0; i < __SCHED_FEAT_NR; i++) {
224 if (strcmp(cmp, sched_feat_names[i]) == 0) {
226 sysctl_sched_features &= ~(1UL << i);
227 sched_feat_disable(i);
229 sysctl_sched_features |= (1UL << i);
230 sched_feat_enable(i);
240 sched_feat_write(struct file *filp, const char __user *ubuf,
241 size_t cnt, loff_t *ppos)
250 if (copy_from_user(&buf, ubuf, cnt))
256 i = sched_feat_set(cmp);
257 if (i == __SCHED_FEAT_NR)
265 static int sched_feat_open(struct inode *inode, struct file *filp)
267 return single_open(filp, sched_feat_show, NULL);
270 static const struct file_operations sched_feat_fops = {
271 .open = sched_feat_open,
272 .write = sched_feat_write,
275 .release = single_release,
278 static __init int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL, NULL,
285 late_initcall(sched_init_debug);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug unsigned int sysctl_sched_nr_migrate = 32;
295 * period over which we average the RT time consumption, measured
300 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period = 1000000;
308 __read_mostly int scheduler_running;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime = 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq *__task_rq_lock(struct task_struct *p)
324 lockdep_assert_held(&p->pi_lock);
328 raw_spin_lock(&rq->lock);
329 if (likely(rq == task_rq(p)))
331 raw_spin_unlock(&rq->lock);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
339 __acquires(p->pi_lock)
345 raw_spin_lock_irqsave(&p->pi_lock, *flags);
347 raw_spin_lock(&rq->lock);
348 if (likely(rq == task_rq(p)))
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 static void __task_rq_unlock(struct rq *rq)
358 raw_spin_unlock(&rq->lock);
362 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
364 __releases(p->pi_lock)
366 raw_spin_unlock(&rq->lock);
367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq *this_rq_lock(void)
380 raw_spin_lock(&rq->lock);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq *rq)
392 if (hrtimer_active(&rq->hrtick_timer))
393 hrtimer_cancel(&rq->hrtick_timer);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart hrtick(struct hrtimer *timer)
402 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
406 raw_spin_lock(&rq->lock);
408 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
409 raw_spin_unlock(&rq->lock);
411 return HRTIMER_NORESTART;
416 static int __hrtick_restart(struct rq *rq)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = hrtimer_get_softexpires(timer);
421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg)
431 raw_spin_lock(&rq->lock);
432 __hrtick_restart(rq);
433 rq->hrtick_csd_pending = 0;
434 raw_spin_unlock(&rq->lock);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq *rq, u64 delay)
444 struct hrtimer *timer = &rq->hrtick_timer;
445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
447 hrtimer_set_expires(timer, time);
449 if (rq == this_rq()) {
450 __hrtick_restart(rq);
451 } else if (!rq->hrtick_csd_pending) {
452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
453 rq->hrtick_csd_pending = 1;
458 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
460 int cpu = (int)(long)hcpu;
463 case CPU_UP_CANCELED:
464 case CPU_UP_CANCELED_FROZEN:
465 case CPU_DOWN_PREPARE:
466 case CPU_DOWN_PREPARE_FROZEN:
468 case CPU_DEAD_FROZEN:
469 hrtick_clear(cpu_rq(cpu));
476 static __init void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq *rq, u64 delay)
488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
489 HRTIMER_MODE_REL_PINNED, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq *rq)
500 rq->hrtick_csd_pending = 0;
502 rq->hrtick_csd.flags = 0;
503 rq->hrtick_csd.func = __hrtick_start;
504 rq->hrtick_csd.info = rq;
507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
508 rq->hrtick_timer.function = hrtick;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq *rq)
515 static inline void init_rq_hrtick(struct rq *rq)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * resched_task - mark a task 'to be rescheduled now'.
527 * On UP this means the setting of the need_resched flag, on SMP it
528 * might also involve a cross-CPU call to trigger the scheduler on
531 void resched_task(struct task_struct *p)
535 lockdep_assert_held(&task_rq(p)->lock);
537 if (test_tsk_need_resched(p))
540 set_tsk_need_resched(p);
543 if (cpu == smp_processor_id()) {
544 set_preempt_need_resched();
548 /* NEED_RESCHED must be visible before we test polling */
550 if (!tsk_is_polling(p))
551 smp_send_reschedule(cpu);
554 void resched_cpu(int cpu)
556 struct rq *rq = cpu_rq(cpu);
559 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
561 resched_task(cpu_curr(cpu));
562 raw_spin_unlock_irqrestore(&rq->lock, flags);
566 #ifdef CONFIG_NO_HZ_COMMON
568 * In the semi idle case, use the nearest busy cpu for migrating timers
569 * from an idle cpu. This is good for power-savings.
571 * We don't do similar optimization for completely idle system, as
572 * selecting an idle cpu will add more delays to the timers than intended
573 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
575 int get_nohz_timer_target(int pinned)
577 int cpu = smp_processor_id();
579 struct sched_domain *sd;
581 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
585 for_each_domain(cpu, sd) {
586 for_each_cpu(i, sched_domain_span(sd)) {
598 * When add_timer_on() enqueues a timer into the timer wheel of an
599 * idle CPU then this timer might expire before the next timer event
600 * which is scheduled to wake up that CPU. In case of a completely
601 * idle system the next event might even be infinite time into the
602 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
603 * leaves the inner idle loop so the newly added timer is taken into
604 * account when the CPU goes back to idle and evaluates the timer
605 * wheel for the next timer event.
607 static void wake_up_idle_cpu(int cpu)
609 struct rq *rq = cpu_rq(cpu);
611 if (cpu == smp_processor_id())
615 * This is safe, as this function is called with the timer
616 * wheel base lock of (cpu) held. When the CPU is on the way
617 * to idle and has not yet set rq->curr to idle then it will
618 * be serialized on the timer wheel base lock and take the new
619 * timer into account automatically.
621 if (rq->curr != rq->idle)
625 * We can set TIF_RESCHED on the idle task of the other CPU
626 * lockless. The worst case is that the other CPU runs the
627 * idle task through an additional NOOP schedule()
629 set_tsk_need_resched(rq->idle);
631 /* NEED_RESCHED must be visible before we test polling */
633 if (!tsk_is_polling(rq->idle))
634 smp_send_reschedule(cpu);
637 static bool wake_up_full_nohz_cpu(int cpu)
639 if (tick_nohz_full_cpu(cpu)) {
640 if (cpu != smp_processor_id() ||
641 tick_nohz_tick_stopped())
642 smp_send_reschedule(cpu);
649 void wake_up_nohz_cpu(int cpu)
651 if (!wake_up_full_nohz_cpu(cpu))
652 wake_up_idle_cpu(cpu);
655 static inline bool got_nohz_idle_kick(void)
657 int cpu = smp_processor_id();
659 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
662 if (idle_cpu(cpu) && !need_resched())
666 * We can't run Idle Load Balance on this CPU for this time so we
667 * cancel it and clear NOHZ_BALANCE_KICK
669 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
673 #else /* CONFIG_NO_HZ_COMMON */
675 static inline bool got_nohz_idle_kick(void)
680 #endif /* CONFIG_NO_HZ_COMMON */
682 #ifdef CONFIG_NO_HZ_FULL
683 bool sched_can_stop_tick(void)
689 /* Make sure rq->nr_running update is visible after the IPI */
692 /* More than one running task need preemption */
693 if (rq->nr_running > 1)
698 #endif /* CONFIG_NO_HZ_FULL */
700 void sched_avg_update(struct rq *rq)
702 s64 period = sched_avg_period();
704 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
706 * Inline assembly required to prevent the compiler
707 * optimising this loop into a divmod call.
708 * See __iter_div_u64_rem() for another example of this.
710 asm("" : "+rm" (rq->age_stamp));
711 rq->age_stamp += period;
716 #endif /* CONFIG_SMP */
718 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
719 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
721 * Iterate task_group tree rooted at *from, calling @down when first entering a
722 * node and @up when leaving it for the final time.
724 * Caller must hold rcu_lock or sufficient equivalent.
726 int walk_tg_tree_from(struct task_group *from,
727 tg_visitor down, tg_visitor up, void *data)
729 struct task_group *parent, *child;
735 ret = (*down)(parent, data);
738 list_for_each_entry_rcu(child, &parent->children, siblings) {
745 ret = (*up)(parent, data);
746 if (ret || parent == from)
750 parent = parent->parent;
757 int tg_nop(struct task_group *tg, void *data)
763 static void set_load_weight(struct task_struct *p)
765 int prio = p->static_prio - MAX_RT_PRIO;
766 struct load_weight *load = &p->se.load;
769 * SCHED_IDLE tasks get minimal weight:
771 if (p->policy == SCHED_IDLE) {
772 load->weight = scale_load(WEIGHT_IDLEPRIO);
773 load->inv_weight = WMULT_IDLEPRIO;
777 load->weight = scale_load(prio_to_weight[prio]);
778 load->inv_weight = prio_to_wmult[prio];
781 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
784 sched_info_queued(rq, p);
785 p->sched_class->enqueue_task(rq, p, flags);
788 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
791 sched_info_dequeued(rq, p);
792 p->sched_class->dequeue_task(rq, p, flags);
795 void activate_task(struct rq *rq, struct task_struct *p, int flags)
797 if (task_contributes_to_load(p))
798 rq->nr_uninterruptible--;
800 enqueue_task(rq, p, flags);
803 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
805 if (task_contributes_to_load(p))
806 rq->nr_uninterruptible++;
808 dequeue_task(rq, p, flags);
811 static void update_rq_clock_task(struct rq *rq, s64 delta)
814 * In theory, the compile should just see 0 here, and optimize out the call
815 * to sched_rt_avg_update. But I don't trust it...
817 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
818 s64 steal = 0, irq_delta = 0;
820 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
821 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
824 * Since irq_time is only updated on {soft,}irq_exit, we might run into
825 * this case when a previous update_rq_clock() happened inside a
828 * When this happens, we stop ->clock_task and only update the
829 * prev_irq_time stamp to account for the part that fit, so that a next
830 * update will consume the rest. This ensures ->clock_task is
833 * It does however cause some slight miss-attribution of {soft,}irq
834 * time, a more accurate solution would be to update the irq_time using
835 * the current rq->clock timestamp, except that would require using
838 if (irq_delta > delta)
841 rq->prev_irq_time += irq_delta;
844 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
845 if (static_key_false((¶virt_steal_rq_enabled))) {
846 steal = paravirt_steal_clock(cpu_of(rq));
847 steal -= rq->prev_steal_time_rq;
849 if (unlikely(steal > delta))
852 rq->prev_steal_time_rq += steal;
857 rq->clock_task += delta;
859 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
860 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
861 sched_rt_avg_update(rq, irq_delta + steal);
865 void sched_set_stop_task(int cpu, struct task_struct *stop)
867 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
868 struct task_struct *old_stop = cpu_rq(cpu)->stop;
872 * Make it appear like a SCHED_FIFO task, its something
873 * userspace knows about and won't get confused about.
875 * Also, it will make PI more or less work without too
876 * much confusion -- but then, stop work should not
877 * rely on PI working anyway.
879 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
881 stop->sched_class = &stop_sched_class;
884 cpu_rq(cpu)->stop = stop;
888 * Reset it back to a normal scheduling class so that
889 * it can die in pieces.
891 old_stop->sched_class = &rt_sched_class;
896 * __normal_prio - return the priority that is based on the static prio
898 static inline int __normal_prio(struct task_struct *p)
900 return p->static_prio;
904 * Calculate the expected normal priority: i.e. priority
905 * without taking RT-inheritance into account. Might be
906 * boosted by interactivity modifiers. Changes upon fork,
907 * setprio syscalls, and whenever the interactivity
908 * estimator recalculates.
910 static inline int normal_prio(struct task_struct *p)
914 if (task_has_dl_policy(p))
915 prio = MAX_DL_PRIO-1;
916 else if (task_has_rt_policy(p))
917 prio = MAX_RT_PRIO-1 - p->rt_priority;
919 prio = __normal_prio(p);
924 * Calculate the current priority, i.e. the priority
925 * taken into account by the scheduler. This value might
926 * be boosted by RT tasks, or might be boosted by
927 * interactivity modifiers. Will be RT if the task got
928 * RT-boosted. If not then it returns p->normal_prio.
930 static int effective_prio(struct task_struct *p)
932 p->normal_prio = normal_prio(p);
934 * If we are RT tasks or we were boosted to RT priority,
935 * keep the priority unchanged. Otherwise, update priority
936 * to the normal priority:
938 if (!rt_prio(p->prio))
939 return p->normal_prio;
944 * task_curr - is this task currently executing on a CPU?
945 * @p: the task in question.
947 * Return: 1 if the task is currently executing. 0 otherwise.
949 inline int task_curr(const struct task_struct *p)
951 return cpu_curr(task_cpu(p)) == p;
954 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
955 const struct sched_class *prev_class,
958 if (prev_class != p->sched_class) {
959 if (prev_class->switched_from)
960 prev_class->switched_from(rq, p);
961 p->sched_class->switched_to(rq, p);
962 } else if (oldprio != p->prio || dl_task(p))
963 p->sched_class->prio_changed(rq, p, oldprio);
966 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
968 const struct sched_class *class;
970 if (p->sched_class == rq->curr->sched_class) {
971 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
973 for_each_class(class) {
974 if (class == rq->curr->sched_class)
976 if (class == p->sched_class) {
977 resched_task(rq->curr);
984 * A queue event has occurred, and we're going to schedule. In
985 * this case, we can save a useless back to back clock update.
987 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
988 rq->skip_clock_update = 1;
992 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
994 #ifdef CONFIG_SCHED_DEBUG
996 * We should never call set_task_cpu() on a blocked task,
997 * ttwu() will sort out the placement.
999 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1000 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1002 #ifdef CONFIG_LOCKDEP
1004 * The caller should hold either p->pi_lock or rq->lock, when changing
1005 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1007 * sched_move_task() holds both and thus holding either pins the cgroup,
1010 * Furthermore, all task_rq users should acquire both locks, see
1013 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1014 lockdep_is_held(&task_rq(p)->lock)));
1018 trace_sched_migrate_task(p, new_cpu);
1020 if (task_cpu(p) != new_cpu) {
1021 if (p->sched_class->migrate_task_rq)
1022 p->sched_class->migrate_task_rq(p, new_cpu);
1023 p->se.nr_migrations++;
1024 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1027 __set_task_cpu(p, new_cpu);
1030 static void __migrate_swap_task(struct task_struct *p, int cpu)
1033 struct rq *src_rq, *dst_rq;
1035 src_rq = task_rq(p);
1036 dst_rq = cpu_rq(cpu);
1038 deactivate_task(src_rq, p, 0);
1039 set_task_cpu(p, cpu);
1040 activate_task(dst_rq, p, 0);
1041 check_preempt_curr(dst_rq, p, 0);
1044 * Task isn't running anymore; make it appear like we migrated
1045 * it before it went to sleep. This means on wakeup we make the
1046 * previous cpu our targer instead of where it really is.
1052 struct migration_swap_arg {
1053 struct task_struct *src_task, *dst_task;
1054 int src_cpu, dst_cpu;
1057 static int migrate_swap_stop(void *data)
1059 struct migration_swap_arg *arg = data;
1060 struct rq *src_rq, *dst_rq;
1063 src_rq = cpu_rq(arg->src_cpu);
1064 dst_rq = cpu_rq(arg->dst_cpu);
1066 double_raw_lock(&arg->src_task->pi_lock,
1067 &arg->dst_task->pi_lock);
1068 double_rq_lock(src_rq, dst_rq);
1069 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1072 if (task_cpu(arg->src_task) != arg->src_cpu)
1075 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1078 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1081 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1082 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1087 double_rq_unlock(src_rq, dst_rq);
1088 raw_spin_unlock(&arg->dst_task->pi_lock);
1089 raw_spin_unlock(&arg->src_task->pi_lock);
1095 * Cross migrate two tasks
1097 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1099 struct migration_swap_arg arg;
1102 arg = (struct migration_swap_arg){
1104 .src_cpu = task_cpu(cur),
1106 .dst_cpu = task_cpu(p),
1109 if (arg.src_cpu == arg.dst_cpu)
1113 * These three tests are all lockless; this is OK since all of them
1114 * will be re-checked with proper locks held further down the line.
1116 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1119 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1122 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1125 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1126 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1132 struct migration_arg {
1133 struct task_struct *task;
1137 static int migration_cpu_stop(void *data);
1140 * wait_task_inactive - wait for a thread to unschedule.
1142 * If @match_state is nonzero, it's the @p->state value just checked and
1143 * not expected to change. If it changes, i.e. @p might have woken up,
1144 * then return zero. When we succeed in waiting for @p to be off its CPU,
1145 * we return a positive number (its total switch count). If a second call
1146 * a short while later returns the same number, the caller can be sure that
1147 * @p has remained unscheduled the whole time.
1149 * The caller must ensure that the task *will* unschedule sometime soon,
1150 * else this function might spin for a *long* time. This function can't
1151 * be called with interrupts off, or it may introduce deadlock with
1152 * smp_call_function() if an IPI is sent by the same process we are
1153 * waiting to become inactive.
1155 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1157 unsigned long flags;
1164 * We do the initial early heuristics without holding
1165 * any task-queue locks at all. We'll only try to get
1166 * the runqueue lock when things look like they will
1172 * If the task is actively running on another CPU
1173 * still, just relax and busy-wait without holding
1176 * NOTE! Since we don't hold any locks, it's not
1177 * even sure that "rq" stays as the right runqueue!
1178 * But we don't care, since "task_running()" will
1179 * return false if the runqueue has changed and p
1180 * is actually now running somewhere else!
1182 while (task_running(rq, p)) {
1183 if (match_state && unlikely(p->state != match_state))
1189 * Ok, time to look more closely! We need the rq
1190 * lock now, to be *sure*. If we're wrong, we'll
1191 * just go back and repeat.
1193 rq = task_rq_lock(p, &flags);
1194 trace_sched_wait_task(p);
1195 running = task_running(rq, p);
1198 if (!match_state || p->state == match_state)
1199 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1200 task_rq_unlock(rq, p, &flags);
1203 * If it changed from the expected state, bail out now.
1205 if (unlikely(!ncsw))
1209 * Was it really running after all now that we
1210 * checked with the proper locks actually held?
1212 * Oops. Go back and try again..
1214 if (unlikely(running)) {
1220 * It's not enough that it's not actively running,
1221 * it must be off the runqueue _entirely_, and not
1224 * So if it was still runnable (but just not actively
1225 * running right now), it's preempted, and we should
1226 * yield - it could be a while.
1228 if (unlikely(on_rq)) {
1229 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1231 set_current_state(TASK_UNINTERRUPTIBLE);
1232 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1237 * Ahh, all good. It wasn't running, and it wasn't
1238 * runnable, which means that it will never become
1239 * running in the future either. We're all done!
1248 * kick_process - kick a running thread to enter/exit the kernel
1249 * @p: the to-be-kicked thread
1251 * Cause a process which is running on another CPU to enter
1252 * kernel-mode, without any delay. (to get signals handled.)
1254 * NOTE: this function doesn't have to take the runqueue lock,
1255 * because all it wants to ensure is that the remote task enters
1256 * the kernel. If the IPI races and the task has been migrated
1257 * to another CPU then no harm is done and the purpose has been
1260 void kick_process(struct task_struct *p)
1266 if ((cpu != smp_processor_id()) && task_curr(p))
1267 smp_send_reschedule(cpu);
1270 EXPORT_SYMBOL_GPL(kick_process);
1271 #endif /* CONFIG_SMP */
1275 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1277 static int select_fallback_rq(int cpu, struct task_struct *p)
1279 int nid = cpu_to_node(cpu);
1280 const struct cpumask *nodemask = NULL;
1281 enum { cpuset, possible, fail } state = cpuset;
1285 * If the node that the cpu is on has been offlined, cpu_to_node()
1286 * will return -1. There is no cpu on the node, and we should
1287 * select the cpu on the other node.
1290 nodemask = cpumask_of_node(nid);
1292 /* Look for allowed, online CPU in same node. */
1293 for_each_cpu(dest_cpu, nodemask) {
1294 if (!cpu_online(dest_cpu))
1296 if (!cpu_active(dest_cpu))
1298 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1304 /* Any allowed, online CPU? */
1305 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1306 if (!cpu_online(dest_cpu))
1308 if (!cpu_active(dest_cpu))
1315 /* No more Mr. Nice Guy. */
1316 cpuset_cpus_allowed_fallback(p);
1321 do_set_cpus_allowed(p, cpu_possible_mask);
1332 if (state != cpuset) {
1334 * Don't tell them about moving exiting tasks or
1335 * kernel threads (both mm NULL), since they never
1338 if (p->mm && printk_ratelimit()) {
1339 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1340 task_pid_nr(p), p->comm, cpu);
1348 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1351 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1353 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1356 * In order not to call set_task_cpu() on a blocking task we need
1357 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1360 * Since this is common to all placement strategies, this lives here.
1362 * [ this allows ->select_task() to simply return task_cpu(p) and
1363 * not worry about this generic constraint ]
1365 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1367 cpu = select_fallback_rq(task_cpu(p), p);
1372 static void update_avg(u64 *avg, u64 sample)
1374 s64 diff = sample - *avg;
1380 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1382 #ifdef CONFIG_SCHEDSTATS
1383 struct rq *rq = this_rq();
1386 int this_cpu = smp_processor_id();
1388 if (cpu == this_cpu) {
1389 schedstat_inc(rq, ttwu_local);
1390 schedstat_inc(p, se.statistics.nr_wakeups_local);
1392 struct sched_domain *sd;
1394 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1396 for_each_domain(this_cpu, sd) {
1397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1398 schedstat_inc(sd, ttwu_wake_remote);
1405 if (wake_flags & WF_MIGRATED)
1406 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1408 #endif /* CONFIG_SMP */
1410 schedstat_inc(rq, ttwu_count);
1411 schedstat_inc(p, se.statistics.nr_wakeups);
1413 if (wake_flags & WF_SYNC)
1414 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1416 #endif /* CONFIG_SCHEDSTATS */
1419 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1421 activate_task(rq, p, en_flags);
1424 /* if a worker is waking up, notify workqueue */
1425 if (p->flags & PF_WQ_WORKER)
1426 wq_worker_waking_up(p, cpu_of(rq));
1430 * Mark the task runnable and perform wakeup-preemption.
1433 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1435 check_preempt_curr(rq, p, wake_flags);
1436 trace_sched_wakeup(p, true);
1438 p->state = TASK_RUNNING;
1440 if (p->sched_class->task_woken)
1441 p->sched_class->task_woken(rq, p);
1443 if (rq->idle_stamp) {
1444 u64 delta = rq_clock(rq) - rq->idle_stamp;
1445 u64 max = 2*rq->max_idle_balance_cost;
1447 update_avg(&rq->avg_idle, delta);
1449 if (rq->avg_idle > max)
1458 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1461 if (p->sched_contributes_to_load)
1462 rq->nr_uninterruptible--;
1465 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1466 ttwu_do_wakeup(rq, p, wake_flags);
1470 * Called in case the task @p isn't fully descheduled from its runqueue,
1471 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1472 * since all we need to do is flip p->state to TASK_RUNNING, since
1473 * the task is still ->on_rq.
1475 static int ttwu_remote(struct task_struct *p, int wake_flags)
1480 rq = __task_rq_lock(p);
1482 /* check_preempt_curr() may use rq clock */
1483 update_rq_clock(rq);
1484 ttwu_do_wakeup(rq, p, wake_flags);
1487 __task_rq_unlock(rq);
1493 static void sched_ttwu_pending(void)
1495 struct rq *rq = this_rq();
1496 struct llist_node *llist = llist_del_all(&rq->wake_list);
1497 struct task_struct *p;
1499 raw_spin_lock(&rq->lock);
1502 p = llist_entry(llist, struct task_struct, wake_entry);
1503 llist = llist_next(llist);
1504 ttwu_do_activate(rq, p, 0);
1507 raw_spin_unlock(&rq->lock);
1510 void scheduler_ipi(void)
1513 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1514 * TIF_NEED_RESCHED remotely (for the first time) will also send
1517 preempt_fold_need_resched();
1519 if (llist_empty(&this_rq()->wake_list)
1520 && !tick_nohz_full_cpu(smp_processor_id())
1521 && !got_nohz_idle_kick())
1525 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1526 * traditionally all their work was done from the interrupt return
1527 * path. Now that we actually do some work, we need to make sure
1530 * Some archs already do call them, luckily irq_enter/exit nest
1533 * Arguably we should visit all archs and update all handlers,
1534 * however a fair share of IPIs are still resched only so this would
1535 * somewhat pessimize the simple resched case.
1538 tick_nohz_full_check();
1539 sched_ttwu_pending();
1542 * Check if someone kicked us for doing the nohz idle load balance.
1544 if (unlikely(got_nohz_idle_kick())) {
1545 this_rq()->idle_balance = 1;
1546 raise_softirq_irqoff(SCHED_SOFTIRQ);
1551 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1553 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1554 smp_send_reschedule(cpu);
1557 bool cpus_share_cache(int this_cpu, int that_cpu)
1559 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1561 #endif /* CONFIG_SMP */
1563 static void ttwu_queue(struct task_struct *p, int cpu)
1565 struct rq *rq = cpu_rq(cpu);
1567 #if defined(CONFIG_SMP)
1568 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1569 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1570 ttwu_queue_remote(p, cpu);
1575 raw_spin_lock(&rq->lock);
1576 ttwu_do_activate(rq, p, 0);
1577 raw_spin_unlock(&rq->lock);
1581 * try_to_wake_up - wake up a thread
1582 * @p: the thread to be awakened
1583 * @state: the mask of task states that can be woken
1584 * @wake_flags: wake modifier flags (WF_*)
1586 * Put it on the run-queue if it's not already there. The "current"
1587 * thread is always on the run-queue (except when the actual
1588 * re-schedule is in progress), and as such you're allowed to do
1589 * the simpler "current->state = TASK_RUNNING" to mark yourself
1590 * runnable without the overhead of this.
1592 * Return: %true if @p was woken up, %false if it was already running.
1593 * or @state didn't match @p's state.
1596 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1598 unsigned long flags;
1599 int cpu, success = 0;
1602 * If we are going to wake up a thread waiting for CONDITION we
1603 * need to ensure that CONDITION=1 done by the caller can not be
1604 * reordered with p->state check below. This pairs with mb() in
1605 * set_current_state() the waiting thread does.
1607 smp_mb__before_spinlock();
1608 raw_spin_lock_irqsave(&p->pi_lock, flags);
1609 if (!(p->state & state))
1612 success = 1; /* we're going to change ->state */
1615 if (p->on_rq && ttwu_remote(p, wake_flags))
1620 * If the owning (remote) cpu is still in the middle of schedule() with
1621 * this task as prev, wait until its done referencing the task.
1626 * Pairs with the smp_wmb() in finish_lock_switch().
1630 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1631 p->state = TASK_WAKING;
1633 if (p->sched_class->task_waking)
1634 p->sched_class->task_waking(p);
1636 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1637 if (task_cpu(p) != cpu) {
1638 wake_flags |= WF_MIGRATED;
1639 set_task_cpu(p, cpu);
1641 #endif /* CONFIG_SMP */
1645 ttwu_stat(p, cpu, wake_flags);
1647 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1653 * try_to_wake_up_local - try to wake up a local task with rq lock held
1654 * @p: the thread to be awakened
1656 * Put @p on the run-queue if it's not already there. The caller must
1657 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1660 static void try_to_wake_up_local(struct task_struct *p)
1662 struct rq *rq = task_rq(p);
1664 if (WARN_ON_ONCE(rq != this_rq()) ||
1665 WARN_ON_ONCE(p == current))
1668 lockdep_assert_held(&rq->lock);
1670 if (!raw_spin_trylock(&p->pi_lock)) {
1671 raw_spin_unlock(&rq->lock);
1672 raw_spin_lock(&p->pi_lock);
1673 raw_spin_lock(&rq->lock);
1676 if (!(p->state & TASK_NORMAL))
1680 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1682 ttwu_do_wakeup(rq, p, 0);
1683 ttwu_stat(p, smp_processor_id(), 0);
1685 raw_spin_unlock(&p->pi_lock);
1689 * wake_up_process - Wake up a specific process
1690 * @p: The process to be woken up.
1692 * Attempt to wake up the nominated process and move it to the set of runnable
1695 * Return: 1 if the process was woken up, 0 if it was already running.
1697 * It may be assumed that this function implies a write memory barrier before
1698 * changing the task state if and only if any tasks are woken up.
1700 int wake_up_process(struct task_struct *p)
1702 WARN_ON(task_is_stopped_or_traced(p));
1703 return try_to_wake_up(p, TASK_NORMAL, 0);
1705 EXPORT_SYMBOL(wake_up_process);
1707 int wake_up_state(struct task_struct *p, unsigned int state)
1709 return try_to_wake_up(p, state, 0);
1713 * Perform scheduler related setup for a newly forked process p.
1714 * p is forked by current.
1716 * __sched_fork() is basic setup used by init_idle() too:
1718 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1723 p->se.exec_start = 0;
1724 p->se.sum_exec_runtime = 0;
1725 p->se.prev_sum_exec_runtime = 0;
1726 p->se.nr_migrations = 0;
1728 INIT_LIST_HEAD(&p->se.group_node);
1730 #ifdef CONFIG_SCHEDSTATS
1731 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1734 RB_CLEAR_NODE(&p->dl.rb_node);
1735 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1736 p->dl.dl_runtime = p->dl.runtime = 0;
1737 p->dl.dl_deadline = p->dl.deadline = 0;
1738 p->dl.dl_period = 0;
1741 INIT_LIST_HEAD(&p->rt.run_list);
1743 #ifdef CONFIG_PREEMPT_NOTIFIERS
1744 INIT_HLIST_HEAD(&p->preempt_notifiers);
1747 #ifdef CONFIG_NUMA_BALANCING
1748 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1749 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1750 p->mm->numa_scan_seq = 0;
1753 if (clone_flags & CLONE_VM)
1754 p->numa_preferred_nid = current->numa_preferred_nid;
1756 p->numa_preferred_nid = -1;
1758 p->node_stamp = 0ULL;
1759 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1760 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1761 p->numa_work.next = &p->numa_work;
1762 p->numa_faults_memory = NULL;
1763 p->numa_faults_buffer_memory = NULL;
1764 p->last_task_numa_placement = 0;
1765 p->last_sum_exec_runtime = 0;
1767 INIT_LIST_HEAD(&p->numa_entry);
1768 p->numa_group = NULL;
1769 #endif /* CONFIG_NUMA_BALANCING */
1772 #ifdef CONFIG_NUMA_BALANCING
1773 #ifdef CONFIG_SCHED_DEBUG
1774 void set_numabalancing_state(bool enabled)
1777 sched_feat_set("NUMA");
1779 sched_feat_set("NO_NUMA");
1782 __read_mostly bool numabalancing_enabled;
1784 void set_numabalancing_state(bool enabled)
1786 numabalancing_enabled = enabled;
1788 #endif /* CONFIG_SCHED_DEBUG */
1790 #ifdef CONFIG_PROC_SYSCTL
1791 int sysctl_numa_balancing(struct ctl_table *table, int write,
1792 void __user *buffer, size_t *lenp, loff_t *ppos)
1796 int state = numabalancing_enabled;
1798 if (write && !capable(CAP_SYS_ADMIN))
1803 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1807 set_numabalancing_state(state);
1814 * fork()/clone()-time setup:
1816 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1818 unsigned long flags;
1819 int cpu = get_cpu();
1821 __sched_fork(clone_flags, p);
1823 * We mark the process as running here. This guarantees that
1824 * nobody will actually run it, and a signal or other external
1825 * event cannot wake it up and insert it on the runqueue either.
1827 p->state = TASK_RUNNING;
1830 * Make sure we do not leak PI boosting priority to the child.
1832 p->prio = current->normal_prio;
1835 * Revert to default priority/policy on fork if requested.
1837 if (unlikely(p->sched_reset_on_fork)) {
1838 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1839 p->policy = SCHED_NORMAL;
1840 p->static_prio = NICE_TO_PRIO(0);
1842 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1843 p->static_prio = NICE_TO_PRIO(0);
1845 p->prio = p->normal_prio = __normal_prio(p);
1849 * We don't need the reset flag anymore after the fork. It has
1850 * fulfilled its duty:
1852 p->sched_reset_on_fork = 0;
1855 if (dl_prio(p->prio)) {
1858 } else if (rt_prio(p->prio)) {
1859 p->sched_class = &rt_sched_class;
1861 p->sched_class = &fair_sched_class;
1864 if (p->sched_class->task_fork)
1865 p->sched_class->task_fork(p);
1868 * The child is not yet in the pid-hash so no cgroup attach races,
1869 * and the cgroup is pinned to this child due to cgroup_fork()
1870 * is ran before sched_fork().
1872 * Silence PROVE_RCU.
1874 raw_spin_lock_irqsave(&p->pi_lock, flags);
1875 set_task_cpu(p, cpu);
1876 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1878 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1879 if (likely(sched_info_on()))
1880 memset(&p->sched_info, 0, sizeof(p->sched_info));
1882 #if defined(CONFIG_SMP)
1885 init_task_preempt_count(p);
1887 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1888 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1895 unsigned long to_ratio(u64 period, u64 runtime)
1897 if (runtime == RUNTIME_INF)
1901 * Doing this here saves a lot of checks in all
1902 * the calling paths, and returning zero seems
1903 * safe for them anyway.
1908 return div64_u64(runtime << 20, period);
1912 inline struct dl_bw *dl_bw_of(int i)
1914 return &cpu_rq(i)->rd->dl_bw;
1917 static inline int dl_bw_cpus(int i)
1919 struct root_domain *rd = cpu_rq(i)->rd;
1922 for_each_cpu_and(i, rd->span, cpu_active_mask)
1928 inline struct dl_bw *dl_bw_of(int i)
1930 return &cpu_rq(i)->dl.dl_bw;
1933 static inline int dl_bw_cpus(int i)
1940 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1942 dl_b->total_bw -= tsk_bw;
1946 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1948 dl_b->total_bw += tsk_bw;
1952 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1954 return dl_b->bw != -1 &&
1955 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1959 * We must be sure that accepting a new task (or allowing changing the
1960 * parameters of an existing one) is consistent with the bandwidth
1961 * constraints. If yes, this function also accordingly updates the currently
1962 * allocated bandwidth to reflect the new situation.
1964 * This function is called while holding p's rq->lock.
1966 static int dl_overflow(struct task_struct *p, int policy,
1967 const struct sched_attr *attr)
1970 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1971 u64 period = attr->sched_period ?: attr->sched_deadline;
1972 u64 runtime = attr->sched_runtime;
1973 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1976 if (new_bw == p->dl.dl_bw)
1980 * Either if a task, enters, leave, or stays -deadline but changes
1981 * its parameters, we may need to update accordingly the total
1982 * allocated bandwidth of the container.
1984 raw_spin_lock(&dl_b->lock);
1985 cpus = dl_bw_cpus(task_cpu(p));
1986 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1987 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1988 __dl_add(dl_b, new_bw);
1990 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1991 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1992 __dl_clear(dl_b, p->dl.dl_bw);
1993 __dl_add(dl_b, new_bw);
1995 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1996 __dl_clear(dl_b, p->dl.dl_bw);
1999 raw_spin_unlock(&dl_b->lock);
2004 extern void init_dl_bw(struct dl_bw *dl_b);
2007 * wake_up_new_task - wake up a newly created task for the first time.
2009 * This function will do some initial scheduler statistics housekeeping
2010 * that must be done for every newly created context, then puts the task
2011 * on the runqueue and wakes it.
2013 void wake_up_new_task(struct task_struct *p)
2015 unsigned long flags;
2018 raw_spin_lock_irqsave(&p->pi_lock, flags);
2021 * Fork balancing, do it here and not earlier because:
2022 * - cpus_allowed can change in the fork path
2023 * - any previously selected cpu might disappear through hotplug
2025 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2028 /* Initialize new task's runnable average */
2029 init_task_runnable_average(p);
2030 rq = __task_rq_lock(p);
2031 activate_task(rq, p, 0);
2033 trace_sched_wakeup_new(p, true);
2034 check_preempt_curr(rq, p, WF_FORK);
2036 if (p->sched_class->task_woken)
2037 p->sched_class->task_woken(rq, p);
2039 task_rq_unlock(rq, p, &flags);
2042 #ifdef CONFIG_PREEMPT_NOTIFIERS
2045 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2046 * @notifier: notifier struct to register
2048 void preempt_notifier_register(struct preempt_notifier *notifier)
2050 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2052 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2055 * preempt_notifier_unregister - no longer interested in preemption notifications
2056 * @notifier: notifier struct to unregister
2058 * This is safe to call from within a preemption notifier.
2060 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2062 hlist_del(¬ifier->link);
2064 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2066 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2068 struct preempt_notifier *notifier;
2070 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2071 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2075 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076 struct task_struct *next)
2078 struct preempt_notifier *notifier;
2080 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2081 notifier->ops->sched_out(notifier, next);
2084 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2086 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2091 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2092 struct task_struct *next)
2096 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2099 * prepare_task_switch - prepare to switch tasks
2100 * @rq: the runqueue preparing to switch
2101 * @prev: the current task that is being switched out
2102 * @next: the task we are going to switch to.
2104 * This is called with the rq lock held and interrupts off. It must
2105 * be paired with a subsequent finish_task_switch after the context
2108 * prepare_task_switch sets up locking and calls architecture specific
2112 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2113 struct task_struct *next)
2115 trace_sched_switch(prev, next);
2116 sched_info_switch(rq, prev, next);
2117 perf_event_task_sched_out(prev, next);
2118 fire_sched_out_preempt_notifiers(prev, next);
2119 prepare_lock_switch(rq, next);
2120 prepare_arch_switch(next);
2124 * finish_task_switch - clean up after a task-switch
2125 * @rq: runqueue associated with task-switch
2126 * @prev: the thread we just switched away from.
2128 * finish_task_switch must be called after the context switch, paired
2129 * with a prepare_task_switch call before the context switch.
2130 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2131 * and do any other architecture-specific cleanup actions.
2133 * Note that we may have delayed dropping an mm in context_switch(). If
2134 * so, we finish that here outside of the runqueue lock. (Doing it
2135 * with the lock held can cause deadlocks; see schedule() for
2138 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2139 __releases(rq->lock)
2141 struct mm_struct *mm = rq->prev_mm;
2147 * A task struct has one reference for the use as "current".
2148 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2149 * schedule one last time. The schedule call will never return, and
2150 * the scheduled task must drop that reference.
2151 * The test for TASK_DEAD must occur while the runqueue locks are
2152 * still held, otherwise prev could be scheduled on another cpu, die
2153 * there before we look at prev->state, and then the reference would
2155 * Manfred Spraul <manfred@colorfullife.com>
2157 prev_state = prev->state;
2158 vtime_task_switch(prev);
2159 finish_arch_switch(prev);
2160 perf_event_task_sched_in(prev, current);
2161 finish_lock_switch(rq, prev);
2162 finish_arch_post_lock_switch();
2164 fire_sched_in_preempt_notifiers(current);
2167 if (unlikely(prev_state == TASK_DEAD)) {
2168 if (prev->sched_class->task_dead)
2169 prev->sched_class->task_dead(prev);
2172 * Remove function-return probe instances associated with this
2173 * task and put them back on the free list.
2175 kprobe_flush_task(prev);
2176 put_task_struct(prev);
2179 tick_nohz_task_switch(current);
2184 /* rq->lock is NOT held, but preemption is disabled */
2185 static inline void post_schedule(struct rq *rq)
2187 if (rq->post_schedule) {
2188 unsigned long flags;
2190 raw_spin_lock_irqsave(&rq->lock, flags);
2191 if (rq->curr->sched_class->post_schedule)
2192 rq->curr->sched_class->post_schedule(rq);
2193 raw_spin_unlock_irqrestore(&rq->lock, flags);
2195 rq->post_schedule = 0;
2201 static inline void post_schedule(struct rq *rq)
2208 * schedule_tail - first thing a freshly forked thread must call.
2209 * @prev: the thread we just switched away from.
2211 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2212 __releases(rq->lock)
2214 struct rq *rq = this_rq();
2216 finish_task_switch(rq, prev);
2219 * FIXME: do we need to worry about rq being invalidated by the
2224 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2225 /* In this case, finish_task_switch does not reenable preemption */
2228 if (current->set_child_tid)
2229 put_user(task_pid_vnr(current), current->set_child_tid);
2233 * context_switch - switch to the new MM and the new
2234 * thread's register state.
2237 context_switch(struct rq *rq, struct task_struct *prev,
2238 struct task_struct *next)
2240 struct mm_struct *mm, *oldmm;
2242 prepare_task_switch(rq, prev, next);
2245 oldmm = prev->active_mm;
2247 * For paravirt, this is coupled with an exit in switch_to to
2248 * combine the page table reload and the switch backend into
2251 arch_start_context_switch(prev);
2254 next->active_mm = oldmm;
2255 atomic_inc(&oldmm->mm_count);
2256 enter_lazy_tlb(oldmm, next);
2258 switch_mm(oldmm, mm, next);
2261 prev->active_mm = NULL;
2262 rq->prev_mm = oldmm;
2265 * Since the runqueue lock will be released by the next
2266 * task (which is an invalid locking op but in the case
2267 * of the scheduler it's an obvious special-case), so we
2268 * do an early lockdep release here:
2270 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2271 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2274 context_tracking_task_switch(prev, next);
2275 /* Here we just switch the register state and the stack. */
2276 switch_to(prev, next, prev);
2280 * this_rq must be evaluated again because prev may have moved
2281 * CPUs since it called schedule(), thus the 'rq' on its stack
2282 * frame will be invalid.
2284 finish_task_switch(this_rq(), prev);
2288 * nr_running and nr_context_switches:
2290 * externally visible scheduler statistics: current number of runnable
2291 * threads, total number of context switches performed since bootup.
2293 unsigned long nr_running(void)
2295 unsigned long i, sum = 0;
2297 for_each_online_cpu(i)
2298 sum += cpu_rq(i)->nr_running;
2303 unsigned long long nr_context_switches(void)
2306 unsigned long long sum = 0;
2308 for_each_possible_cpu(i)
2309 sum += cpu_rq(i)->nr_switches;
2314 unsigned long nr_iowait(void)
2316 unsigned long i, sum = 0;
2318 for_each_possible_cpu(i)
2319 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2324 unsigned long nr_iowait_cpu(int cpu)
2326 struct rq *this = cpu_rq(cpu);
2327 return atomic_read(&this->nr_iowait);
2333 * sched_exec - execve() is a valuable balancing opportunity, because at
2334 * this point the task has the smallest effective memory and cache footprint.
2336 void sched_exec(void)
2338 struct task_struct *p = current;
2339 unsigned long flags;
2342 raw_spin_lock_irqsave(&p->pi_lock, flags);
2343 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2344 if (dest_cpu == smp_processor_id())
2347 if (likely(cpu_active(dest_cpu))) {
2348 struct migration_arg arg = { p, dest_cpu };
2350 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2351 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2355 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2360 DEFINE_PER_CPU(struct kernel_stat, kstat);
2361 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2363 EXPORT_PER_CPU_SYMBOL(kstat);
2364 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2367 * Return any ns on the sched_clock that have not yet been accounted in
2368 * @p in case that task is currently running.
2370 * Called with task_rq_lock() held on @rq.
2372 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2376 if (task_current(rq, p)) {
2377 update_rq_clock(rq);
2378 ns = rq_clock_task(rq) - p->se.exec_start;
2386 unsigned long long task_delta_exec(struct task_struct *p)
2388 unsigned long flags;
2392 rq = task_rq_lock(p, &flags);
2393 ns = do_task_delta_exec(p, rq);
2394 task_rq_unlock(rq, p, &flags);
2400 * Return accounted runtime for the task.
2401 * In case the task is currently running, return the runtime plus current's
2402 * pending runtime that have not been accounted yet.
2404 unsigned long long task_sched_runtime(struct task_struct *p)
2406 unsigned long flags;
2410 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2412 * 64-bit doesn't need locks to atomically read a 64bit value.
2413 * So we have a optimization chance when the task's delta_exec is 0.
2414 * Reading ->on_cpu is racy, but this is ok.
2416 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2417 * If we race with it entering cpu, unaccounted time is 0. This is
2418 * indistinguishable from the read occurring a few cycles earlier.
2421 return p->se.sum_exec_runtime;
2424 rq = task_rq_lock(p, &flags);
2425 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2426 task_rq_unlock(rq, p, &flags);
2432 * This function gets called by the timer code, with HZ frequency.
2433 * We call it with interrupts disabled.
2435 void scheduler_tick(void)
2437 int cpu = smp_processor_id();
2438 struct rq *rq = cpu_rq(cpu);
2439 struct task_struct *curr = rq->curr;
2443 raw_spin_lock(&rq->lock);
2444 update_rq_clock(rq);
2445 curr->sched_class->task_tick(rq, curr, 0);
2446 update_cpu_load_active(rq);
2447 raw_spin_unlock(&rq->lock);
2449 perf_event_task_tick();
2452 rq->idle_balance = idle_cpu(cpu);
2453 trigger_load_balance(rq);
2455 rq_last_tick_reset(rq);
2458 #ifdef CONFIG_NO_HZ_FULL
2460 * scheduler_tick_max_deferment
2462 * Keep at least one tick per second when a single
2463 * active task is running because the scheduler doesn't
2464 * yet completely support full dynticks environment.
2466 * This makes sure that uptime, CFS vruntime, load
2467 * balancing, etc... continue to move forward, even
2468 * with a very low granularity.
2470 * Return: Maximum deferment in nanoseconds.
2472 u64 scheduler_tick_max_deferment(void)
2474 struct rq *rq = this_rq();
2475 unsigned long next, now = ACCESS_ONCE(jiffies);
2477 next = rq->last_sched_tick + HZ;
2479 if (time_before_eq(next, now))
2482 return jiffies_to_nsecs(next - now);
2486 notrace unsigned long get_parent_ip(unsigned long addr)
2488 if (in_lock_functions(addr)) {
2489 addr = CALLER_ADDR2;
2490 if (in_lock_functions(addr))
2491 addr = CALLER_ADDR3;
2496 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2497 defined(CONFIG_PREEMPT_TRACER))
2499 void __kprobes preempt_count_add(int val)
2501 #ifdef CONFIG_DEBUG_PREEMPT
2505 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2508 __preempt_count_add(val);
2509 #ifdef CONFIG_DEBUG_PREEMPT
2511 * Spinlock count overflowing soon?
2513 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2516 if (preempt_count() == val) {
2517 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2518 #ifdef CONFIG_DEBUG_PREEMPT
2519 current->preempt_disable_ip = ip;
2521 trace_preempt_off(CALLER_ADDR0, ip);
2524 EXPORT_SYMBOL(preempt_count_add);
2526 void __kprobes preempt_count_sub(int val)
2528 #ifdef CONFIG_DEBUG_PREEMPT
2532 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2535 * Is the spinlock portion underflowing?
2537 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2538 !(preempt_count() & PREEMPT_MASK)))
2542 if (preempt_count() == val)
2543 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2544 __preempt_count_sub(val);
2546 EXPORT_SYMBOL(preempt_count_sub);
2551 * Print scheduling while atomic bug:
2553 static noinline void __schedule_bug(struct task_struct *prev)
2555 if (oops_in_progress)
2558 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2559 prev->comm, prev->pid, preempt_count());
2561 debug_show_held_locks(prev);
2563 if (irqs_disabled())
2564 print_irqtrace_events(prev);
2565 #ifdef CONFIG_DEBUG_PREEMPT
2566 if (in_atomic_preempt_off()) {
2567 pr_err("Preemption disabled at:");
2568 print_ip_sym(current->preempt_disable_ip);
2573 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2577 * Various schedule()-time debugging checks and statistics:
2579 static inline void schedule_debug(struct task_struct *prev)
2582 * Test if we are atomic. Since do_exit() needs to call into
2583 * schedule() atomically, we ignore that path. Otherwise whine
2584 * if we are scheduling when we should not.
2586 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2587 __schedule_bug(prev);
2590 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2592 schedstat_inc(this_rq(), sched_count);
2596 * Pick up the highest-prio task:
2598 static inline struct task_struct *
2599 pick_next_task(struct rq *rq, struct task_struct *prev)
2601 const struct sched_class *class = &fair_sched_class;
2602 struct task_struct *p;
2605 * Optimization: we know that if all tasks are in
2606 * the fair class we can call that function directly:
2608 if (likely(prev->sched_class == class &&
2609 rq->nr_running == rq->cfs.h_nr_running)) {
2610 p = fair_sched_class.pick_next_task(rq, prev);
2611 if (unlikely(p == RETRY_TASK))
2614 /* assumes fair_sched_class->next == idle_sched_class */
2616 p = idle_sched_class.pick_next_task(rq, prev);
2622 for_each_class(class) {
2623 p = class->pick_next_task(rq, prev);
2625 if (unlikely(p == RETRY_TASK))
2631 BUG(); /* the idle class will always have a runnable task */
2635 * __schedule() is the main scheduler function.
2637 * The main means of driving the scheduler and thus entering this function are:
2639 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2641 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2642 * paths. For example, see arch/x86/entry_64.S.
2644 * To drive preemption between tasks, the scheduler sets the flag in timer
2645 * interrupt handler scheduler_tick().
2647 * 3. Wakeups don't really cause entry into schedule(). They add a
2648 * task to the run-queue and that's it.
2650 * Now, if the new task added to the run-queue preempts the current
2651 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2652 * called on the nearest possible occasion:
2654 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2656 * - in syscall or exception context, at the next outmost
2657 * preempt_enable(). (this might be as soon as the wake_up()'s
2660 * - in IRQ context, return from interrupt-handler to
2661 * preemptible context
2663 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2666 * - cond_resched() call
2667 * - explicit schedule() call
2668 * - return from syscall or exception to user-space
2669 * - return from interrupt-handler to user-space
2671 static void __sched __schedule(void)
2673 struct task_struct *prev, *next;
2674 unsigned long *switch_count;
2680 cpu = smp_processor_id();
2682 rcu_note_context_switch(cpu);
2685 schedule_debug(prev);
2687 if (sched_feat(HRTICK))
2691 * Make sure that signal_pending_state()->signal_pending() below
2692 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2693 * done by the caller to avoid the race with signal_wake_up().
2695 smp_mb__before_spinlock();
2696 raw_spin_lock_irq(&rq->lock);
2698 switch_count = &prev->nivcsw;
2699 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2700 if (unlikely(signal_pending_state(prev->state, prev))) {
2701 prev->state = TASK_RUNNING;
2703 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2707 * If a worker went to sleep, notify and ask workqueue
2708 * whether it wants to wake up a task to maintain
2711 if (prev->flags & PF_WQ_WORKER) {
2712 struct task_struct *to_wakeup;
2714 to_wakeup = wq_worker_sleeping(prev, cpu);
2716 try_to_wake_up_local(to_wakeup);
2719 switch_count = &prev->nvcsw;
2722 if (prev->on_rq || rq->skip_clock_update < 0)
2723 update_rq_clock(rq);
2725 next = pick_next_task(rq, prev);
2726 clear_tsk_need_resched(prev);
2727 clear_preempt_need_resched();
2728 rq->skip_clock_update = 0;
2730 if (likely(prev != next)) {
2735 context_switch(rq, prev, next); /* unlocks the rq */
2737 * The context switch have flipped the stack from under us
2738 * and restored the local variables which were saved when
2739 * this task called schedule() in the past. prev == current
2740 * is still correct, but it can be moved to another cpu/rq.
2742 cpu = smp_processor_id();
2745 raw_spin_unlock_irq(&rq->lock);
2749 sched_preempt_enable_no_resched();
2754 static inline void sched_submit_work(struct task_struct *tsk)
2756 if (!tsk->state || tsk_is_pi_blocked(tsk))
2759 * If we are going to sleep and we have plugged IO queued,
2760 * make sure to submit it to avoid deadlocks.
2762 if (blk_needs_flush_plug(tsk))
2763 blk_schedule_flush_plug(tsk);
2766 asmlinkage __visible void __sched schedule(void)
2768 struct task_struct *tsk = current;
2770 sched_submit_work(tsk);
2773 EXPORT_SYMBOL(schedule);
2775 #ifdef CONFIG_CONTEXT_TRACKING
2776 asmlinkage __visible void __sched schedule_user(void)
2779 * If we come here after a random call to set_need_resched(),
2780 * or we have been woken up remotely but the IPI has not yet arrived,
2781 * we haven't yet exited the RCU idle mode. Do it here manually until
2782 * we find a better solution.
2791 * schedule_preempt_disabled - called with preemption disabled
2793 * Returns with preemption disabled. Note: preempt_count must be 1
2795 void __sched schedule_preempt_disabled(void)
2797 sched_preempt_enable_no_resched();
2802 #ifdef CONFIG_PREEMPT
2804 * this is the entry point to schedule() from in-kernel preemption
2805 * off of preempt_enable. Kernel preemptions off return from interrupt
2806 * occur there and call schedule directly.
2808 asmlinkage __visible void __sched notrace preempt_schedule(void)
2811 * If there is a non-zero preempt_count or interrupts are disabled,
2812 * we do not want to preempt the current task. Just return..
2814 if (likely(!preemptible()))
2818 __preempt_count_add(PREEMPT_ACTIVE);
2820 __preempt_count_sub(PREEMPT_ACTIVE);
2823 * Check again in case we missed a preemption opportunity
2824 * between schedule and now.
2827 } while (need_resched());
2829 EXPORT_SYMBOL(preempt_schedule);
2830 #endif /* CONFIG_PREEMPT */
2833 * this is the entry point to schedule() from kernel preemption
2834 * off of irq context.
2835 * Note, that this is called and return with irqs disabled. This will
2836 * protect us against recursive calling from irq.
2838 asmlinkage __visible void __sched preempt_schedule_irq(void)
2840 enum ctx_state prev_state;
2842 /* Catch callers which need to be fixed */
2843 BUG_ON(preempt_count() || !irqs_disabled());
2845 prev_state = exception_enter();
2848 __preempt_count_add(PREEMPT_ACTIVE);
2851 local_irq_disable();
2852 __preempt_count_sub(PREEMPT_ACTIVE);
2855 * Check again in case we missed a preemption opportunity
2856 * between schedule and now.
2859 } while (need_resched());
2861 exception_exit(prev_state);
2864 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2867 return try_to_wake_up(curr->private, mode, wake_flags);
2869 EXPORT_SYMBOL(default_wake_function);
2871 #ifdef CONFIG_RT_MUTEXES
2874 * rt_mutex_setprio - set the current priority of a task
2876 * @prio: prio value (kernel-internal form)
2878 * This function changes the 'effective' priority of a task. It does
2879 * not touch ->normal_prio like __setscheduler().
2881 * Used by the rt_mutex code to implement priority inheritance
2882 * logic. Call site only calls if the priority of the task changed.
2884 void rt_mutex_setprio(struct task_struct *p, int prio)
2886 int oldprio, on_rq, running, enqueue_flag = 0;
2888 const struct sched_class *prev_class;
2890 BUG_ON(prio > MAX_PRIO);
2892 rq = __task_rq_lock(p);
2895 * Idle task boosting is a nono in general. There is one
2896 * exception, when PREEMPT_RT and NOHZ is active:
2898 * The idle task calls get_next_timer_interrupt() and holds
2899 * the timer wheel base->lock on the CPU and another CPU wants
2900 * to access the timer (probably to cancel it). We can safely
2901 * ignore the boosting request, as the idle CPU runs this code
2902 * with interrupts disabled and will complete the lock
2903 * protected section without being interrupted. So there is no
2904 * real need to boost.
2906 if (unlikely(p == rq->idle)) {
2907 WARN_ON(p != rq->curr);
2908 WARN_ON(p->pi_blocked_on);
2912 trace_sched_pi_setprio(p, prio);
2913 p->pi_top_task = rt_mutex_get_top_task(p);
2915 prev_class = p->sched_class;
2917 running = task_current(rq, p);
2919 dequeue_task(rq, p, 0);
2921 p->sched_class->put_prev_task(rq, p);
2924 * Boosting condition are:
2925 * 1. -rt task is running and holds mutex A
2926 * --> -dl task blocks on mutex A
2928 * 2. -dl task is running and holds mutex A
2929 * --> -dl task blocks on mutex A and could preempt the
2932 if (dl_prio(prio)) {
2933 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2934 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2935 p->dl.dl_boosted = 1;
2936 p->dl.dl_throttled = 0;
2937 enqueue_flag = ENQUEUE_REPLENISH;
2939 p->dl.dl_boosted = 0;
2940 p->sched_class = &dl_sched_class;
2941 } else if (rt_prio(prio)) {
2942 if (dl_prio(oldprio))
2943 p->dl.dl_boosted = 0;
2945 enqueue_flag = ENQUEUE_HEAD;
2946 p->sched_class = &rt_sched_class;
2948 if (dl_prio(oldprio))
2949 p->dl.dl_boosted = 0;
2950 p->sched_class = &fair_sched_class;
2956 p->sched_class->set_curr_task(rq);
2958 enqueue_task(rq, p, enqueue_flag);
2960 check_class_changed(rq, p, prev_class, oldprio);
2962 __task_rq_unlock(rq);
2966 void set_user_nice(struct task_struct *p, long nice)
2968 int old_prio, delta, on_rq;
2969 unsigned long flags;
2972 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2975 * We have to be careful, if called from sys_setpriority(),
2976 * the task might be in the middle of scheduling on another CPU.
2978 rq = task_rq_lock(p, &flags);
2980 * The RT priorities are set via sched_setscheduler(), but we still
2981 * allow the 'normal' nice value to be set - but as expected
2982 * it wont have any effect on scheduling until the task is
2983 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2985 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2986 p->static_prio = NICE_TO_PRIO(nice);
2991 dequeue_task(rq, p, 0);
2993 p->static_prio = NICE_TO_PRIO(nice);
2996 p->prio = effective_prio(p);
2997 delta = p->prio - old_prio;
3000 enqueue_task(rq, p, 0);
3002 * If the task increased its priority or is running and
3003 * lowered its priority, then reschedule its CPU:
3005 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3006 resched_task(rq->curr);
3009 task_rq_unlock(rq, p, &flags);
3011 EXPORT_SYMBOL(set_user_nice);
3014 * can_nice - check if a task can reduce its nice value
3018 int can_nice(const struct task_struct *p, const int nice)
3020 /* convert nice value [19,-20] to rlimit style value [1,40] */
3021 int nice_rlim = 20 - nice;
3023 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3024 capable(CAP_SYS_NICE));
3027 #ifdef __ARCH_WANT_SYS_NICE
3030 * sys_nice - change the priority of the current process.
3031 * @increment: priority increment
3033 * sys_setpriority is a more generic, but much slower function that
3034 * does similar things.
3036 SYSCALL_DEFINE1(nice, int, increment)
3041 * Setpriority might change our priority at the same moment.
3042 * We don't have to worry. Conceptually one call occurs first
3043 * and we have a single winner.
3045 if (increment < -40)
3050 nice = task_nice(current) + increment;
3051 if (nice < MIN_NICE)
3053 if (nice > MAX_NICE)
3056 if (increment < 0 && !can_nice(current, nice))
3059 retval = security_task_setnice(current, nice);
3063 set_user_nice(current, nice);
3070 * task_prio - return the priority value of a given task.
3071 * @p: the task in question.
3073 * Return: The priority value as seen by users in /proc.
3074 * RT tasks are offset by -200. Normal tasks are centered
3075 * around 0, value goes from -16 to +15.
3077 int task_prio(const struct task_struct *p)
3079 return p->prio - MAX_RT_PRIO;
3083 * idle_cpu - is a given cpu idle currently?
3084 * @cpu: the processor in question.
3086 * Return: 1 if the CPU is currently idle. 0 otherwise.
3088 int idle_cpu(int cpu)
3090 struct rq *rq = cpu_rq(cpu);
3092 if (rq->curr != rq->idle)
3099 if (!llist_empty(&rq->wake_list))
3107 * idle_task - return the idle task for a given cpu.
3108 * @cpu: the processor in question.
3110 * Return: The idle task for the cpu @cpu.
3112 struct task_struct *idle_task(int cpu)
3114 return cpu_rq(cpu)->idle;
3118 * find_process_by_pid - find a process with a matching PID value.
3119 * @pid: the pid in question.
3121 * The task of @pid, if found. %NULL otherwise.
3123 static struct task_struct *find_process_by_pid(pid_t pid)
3125 return pid ? find_task_by_vpid(pid) : current;
3129 * This function initializes the sched_dl_entity of a newly becoming
3130 * SCHED_DEADLINE task.
3132 * Only the static values are considered here, the actual runtime and the
3133 * absolute deadline will be properly calculated when the task is enqueued
3134 * for the first time with its new policy.
3137 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3139 struct sched_dl_entity *dl_se = &p->dl;
3141 init_dl_task_timer(dl_se);
3142 dl_se->dl_runtime = attr->sched_runtime;
3143 dl_se->dl_deadline = attr->sched_deadline;
3144 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3145 dl_se->flags = attr->sched_flags;
3146 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3147 dl_se->dl_throttled = 0;
3149 dl_se->dl_yielded = 0;
3152 static void __setscheduler_params(struct task_struct *p,
3153 const struct sched_attr *attr)
3155 int policy = attr->sched_policy;
3157 if (policy == -1) /* setparam */
3162 if (dl_policy(policy))
3163 __setparam_dl(p, attr);
3164 else if (fair_policy(policy))
3165 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3168 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3169 * !rt_policy. Always setting this ensures that things like
3170 * getparam()/getattr() don't report silly values for !rt tasks.
3172 p->rt_priority = attr->sched_priority;
3173 p->normal_prio = normal_prio(p);
3177 /* Actually do priority change: must hold pi & rq lock. */
3178 static void __setscheduler(struct rq *rq, struct task_struct *p,
3179 const struct sched_attr *attr)
3181 __setscheduler_params(p, attr);
3184 * If we get here, there was no pi waiters boosting the
3185 * task. It is safe to use the normal prio.
3187 p->prio = normal_prio(p);
3189 if (dl_prio(p->prio))
3190 p->sched_class = &dl_sched_class;
3191 else if (rt_prio(p->prio))
3192 p->sched_class = &rt_sched_class;
3194 p->sched_class = &fair_sched_class;
3198 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3200 struct sched_dl_entity *dl_se = &p->dl;
3202 attr->sched_priority = p->rt_priority;
3203 attr->sched_runtime = dl_se->dl_runtime;
3204 attr->sched_deadline = dl_se->dl_deadline;
3205 attr->sched_period = dl_se->dl_period;
3206 attr->sched_flags = dl_se->flags;
3210 * This function validates the new parameters of a -deadline task.
3211 * We ask for the deadline not being zero, and greater or equal
3212 * than the runtime, as well as the period of being zero or
3213 * greater than deadline. Furthermore, we have to be sure that
3214 * user parameters are above the internal resolution of 1us (we
3215 * check sched_runtime only since it is always the smaller one) and
3216 * below 2^63 ns (we have to check both sched_deadline and
3217 * sched_period, as the latter can be zero).
3220 __checkparam_dl(const struct sched_attr *attr)
3223 if (attr->sched_deadline == 0)
3227 * Since we truncate DL_SCALE bits, make sure we're at least
3230 if (attr->sched_runtime < (1ULL << DL_SCALE))
3234 * Since we use the MSB for wrap-around and sign issues, make
3235 * sure it's not set (mind that period can be equal to zero).
3237 if (attr->sched_deadline & (1ULL << 63) ||
3238 attr->sched_period & (1ULL << 63))
3241 /* runtime <= deadline <= period (if period != 0) */
3242 if ((attr->sched_period != 0 &&
3243 attr->sched_period < attr->sched_deadline) ||
3244 attr->sched_deadline < attr->sched_runtime)
3251 * check the target process has a UID that matches the current process's
3253 static bool check_same_owner(struct task_struct *p)
3255 const struct cred *cred = current_cred(), *pcred;
3259 pcred = __task_cred(p);
3260 match = (uid_eq(cred->euid, pcred->euid) ||
3261 uid_eq(cred->euid, pcred->uid));
3266 static int __sched_setscheduler(struct task_struct *p,
3267 const struct sched_attr *attr,
3270 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3271 MAX_RT_PRIO - 1 - attr->sched_priority;
3272 int retval, oldprio, oldpolicy = -1, on_rq, running;
3273 int policy = attr->sched_policy;
3274 unsigned long flags;
3275 const struct sched_class *prev_class;
3279 /* may grab non-irq protected spin_locks */
3280 BUG_ON(in_interrupt());
3282 /* double check policy once rq lock held */
3284 reset_on_fork = p->sched_reset_on_fork;
3285 policy = oldpolicy = p->policy;
3287 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3289 if (policy != SCHED_DEADLINE &&
3290 policy != SCHED_FIFO && policy != SCHED_RR &&
3291 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3292 policy != SCHED_IDLE)
3296 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3300 * Valid priorities for SCHED_FIFO and SCHED_RR are
3301 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3302 * SCHED_BATCH and SCHED_IDLE is 0.
3304 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3305 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3307 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3308 (rt_policy(policy) != (attr->sched_priority != 0)))
3312 * Allow unprivileged RT tasks to decrease priority:
3314 if (user && !capable(CAP_SYS_NICE)) {
3315 if (fair_policy(policy)) {
3316 if (attr->sched_nice < task_nice(p) &&
3317 !can_nice(p, attr->sched_nice))
3321 if (rt_policy(policy)) {
3322 unsigned long rlim_rtprio =
3323 task_rlimit(p, RLIMIT_RTPRIO);
3325 /* can't set/change the rt policy */
3326 if (policy != p->policy && !rlim_rtprio)
3329 /* can't increase priority */
3330 if (attr->sched_priority > p->rt_priority &&
3331 attr->sched_priority > rlim_rtprio)
3336 * Can't set/change SCHED_DEADLINE policy at all for now
3337 * (safest behavior); in the future we would like to allow
3338 * unprivileged DL tasks to increase their relative deadline
3339 * or reduce their runtime (both ways reducing utilization)
3341 if (dl_policy(policy))
3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3349 if (!can_nice(p, task_nice(p)))
3353 /* can't change other user's priorities */
3354 if (!check_same_owner(p))
3357 /* Normal users shall not reset the sched_reset_on_fork flag */
3358 if (p->sched_reset_on_fork && !reset_on_fork)
3363 retval = security_task_setscheduler(p);
3369 * make sure no PI-waiters arrive (or leave) while we are
3370 * changing the priority of the task:
3372 * To be able to change p->policy safely, the appropriate
3373 * runqueue lock must be held.
3375 rq = task_rq_lock(p, &flags);
3378 * Changing the policy of the stop threads its a very bad idea
3380 if (p == rq->stop) {
3381 task_rq_unlock(rq, p, &flags);
3386 * If not changing anything there's no need to proceed further,
3387 * but store a possible modification of reset_on_fork.
3389 if (unlikely(policy == p->policy)) {
3390 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3392 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3394 if (dl_policy(policy))
3397 p->sched_reset_on_fork = reset_on_fork;
3398 task_rq_unlock(rq, p, &flags);
3404 #ifdef CONFIG_RT_GROUP_SCHED
3406 * Do not allow realtime tasks into groups that have no runtime
3409 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3410 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3411 !task_group_is_autogroup(task_group(p))) {
3412 task_rq_unlock(rq, p, &flags);
3417 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3418 cpumask_t *span = rq->rd->span;
3421 * Don't allow tasks with an affinity mask smaller than
3422 * the entire root_domain to become SCHED_DEADLINE. We
3423 * will also fail if there's no bandwidth available.
3425 if (!cpumask_subset(span, &p->cpus_allowed) ||
3426 rq->rd->dl_bw.bw == 0) {
3427 task_rq_unlock(rq, p, &flags);
3434 /* recheck policy now with rq lock held */
3435 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3436 policy = oldpolicy = -1;
3437 task_rq_unlock(rq, p, &flags);
3442 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3443 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3446 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3447 task_rq_unlock(rq, p, &flags);
3451 p->sched_reset_on_fork = reset_on_fork;
3455 * Special case for priority boosted tasks.
3457 * If the new priority is lower or equal (user space view)
3458 * than the current (boosted) priority, we just store the new
3459 * normal parameters and do not touch the scheduler class and
3460 * the runqueue. This will be done when the task deboost
3463 if (rt_mutex_check_prio(p, newprio)) {
3464 __setscheduler_params(p, attr);
3465 task_rq_unlock(rq, p, &flags);
3470 running = task_current(rq, p);
3472 dequeue_task(rq, p, 0);
3474 p->sched_class->put_prev_task(rq, p);
3476 prev_class = p->sched_class;
3477 __setscheduler(rq, p, attr);
3480 p->sched_class->set_curr_task(rq);
3483 * We enqueue to tail when the priority of a task is
3484 * increased (user space view).
3486 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3489 check_class_changed(rq, p, prev_class, oldprio);
3490 task_rq_unlock(rq, p, &flags);
3492 rt_mutex_adjust_pi(p);
3497 static int _sched_setscheduler(struct task_struct *p, int policy,
3498 const struct sched_param *param, bool check)
3500 struct sched_attr attr = {
3501 .sched_policy = policy,
3502 .sched_priority = param->sched_priority,
3503 .sched_nice = PRIO_TO_NICE(p->static_prio),
3507 * Fixup the legacy SCHED_RESET_ON_FORK hack
3509 if (policy & SCHED_RESET_ON_FORK) {
3510 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3511 policy &= ~SCHED_RESET_ON_FORK;
3512 attr.sched_policy = policy;
3515 return __sched_setscheduler(p, &attr, check);
3518 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3519 * @p: the task in question.
3520 * @policy: new policy.
3521 * @param: structure containing the new RT priority.
3523 * Return: 0 on success. An error code otherwise.
3525 * NOTE that the task may be already dead.
3527 int sched_setscheduler(struct task_struct *p, int policy,
3528 const struct sched_param *param)
3530 return _sched_setscheduler(p, policy, param, true);
3532 EXPORT_SYMBOL_GPL(sched_setscheduler);
3534 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3536 return __sched_setscheduler(p, attr, true);
3538 EXPORT_SYMBOL_GPL(sched_setattr);
3541 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3542 * @p: the task in question.
3543 * @policy: new policy.
3544 * @param: structure containing the new RT priority.
3546 * Just like sched_setscheduler, only don't bother checking if the
3547 * current context has permission. For example, this is needed in
3548 * stop_machine(): we create temporary high priority worker threads,
3549 * but our caller might not have that capability.
3551 * Return: 0 on success. An error code otherwise.
3553 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3554 const struct sched_param *param)
3556 return _sched_setscheduler(p, policy, param, false);
3560 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3562 struct sched_param lparam;
3563 struct task_struct *p;
3566 if (!param || pid < 0)
3568 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3573 p = find_process_by_pid(pid);
3575 retval = sched_setscheduler(p, policy, &lparam);
3582 * Mimics kernel/events/core.c perf_copy_attr().
3584 static int sched_copy_attr(struct sched_attr __user *uattr,
3585 struct sched_attr *attr)
3590 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3594 * zero the full structure, so that a short copy will be nice.
3596 memset(attr, 0, sizeof(*attr));
3598 ret = get_user(size, &uattr->size);
3602 if (size > PAGE_SIZE) /* silly large */
3605 if (!size) /* abi compat */
3606 size = SCHED_ATTR_SIZE_VER0;
3608 if (size < SCHED_ATTR_SIZE_VER0)
3612 * If we're handed a bigger struct than we know of,
3613 * ensure all the unknown bits are 0 - i.e. new
3614 * user-space does not rely on any kernel feature
3615 * extensions we dont know about yet.
3617 if (size > sizeof(*attr)) {
3618 unsigned char __user *addr;
3619 unsigned char __user *end;
3622 addr = (void __user *)uattr + sizeof(*attr);
3623 end = (void __user *)uattr + size;
3625 for (; addr < end; addr++) {
3626 ret = get_user(val, addr);
3632 size = sizeof(*attr);
3635 ret = copy_from_user(attr, uattr, size);
3640 * XXX: do we want to be lenient like existing syscalls; or do we want
3641 * to be strict and return an error on out-of-bounds values?
3643 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3649 put_user(sizeof(*attr), &uattr->size);
3655 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3656 * @pid: the pid in question.
3657 * @policy: new policy.
3658 * @param: structure containing the new RT priority.
3660 * Return: 0 on success. An error code otherwise.
3662 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3663 struct sched_param __user *, param)
3665 /* negative values for policy are not valid */
3669 return do_sched_setscheduler(pid, policy, param);
3673 * sys_sched_setparam - set/change the RT priority of a thread
3674 * @pid: the pid in question.
3675 * @param: structure containing the new RT priority.
3677 * Return: 0 on success. An error code otherwise.
3679 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3681 return do_sched_setscheduler(pid, -1, param);
3685 * sys_sched_setattr - same as above, but with extended sched_attr
3686 * @pid: the pid in question.
3687 * @uattr: structure containing the extended parameters.
3688 * @flags: for future extension.
3690 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3691 unsigned int, flags)
3693 struct sched_attr attr;
3694 struct task_struct *p;
3697 if (!uattr || pid < 0 || flags)
3700 retval = sched_copy_attr(uattr, &attr);
3704 if (attr.sched_policy < 0)
3709 p = find_process_by_pid(pid);
3711 retval = sched_setattr(p, &attr);
3718 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3719 * @pid: the pid in question.
3721 * Return: On success, the policy of the thread. Otherwise, a negative error
3724 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3726 struct task_struct *p;
3734 p = find_process_by_pid(pid);
3736 retval = security_task_getscheduler(p);
3739 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3746 * sys_sched_getparam - get the RT priority of a thread
3747 * @pid: the pid in question.
3748 * @param: structure containing the RT priority.
3750 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3753 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3755 struct sched_param lp = { .sched_priority = 0 };
3756 struct task_struct *p;
3759 if (!param || pid < 0)
3763 p = find_process_by_pid(pid);
3768 retval = security_task_getscheduler(p);
3772 if (task_has_rt_policy(p))
3773 lp.sched_priority = p->rt_priority;
3777 * This one might sleep, we cannot do it with a spinlock held ...
3779 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3788 static int sched_read_attr(struct sched_attr __user *uattr,
3789 struct sched_attr *attr,
3794 if (!access_ok(VERIFY_WRITE, uattr, usize))
3798 * If we're handed a smaller struct than we know of,
3799 * ensure all the unknown bits are 0 - i.e. old
3800 * user-space does not get uncomplete information.
3802 if (usize < sizeof(*attr)) {
3803 unsigned char *addr;
3806 addr = (void *)attr + usize;
3807 end = (void *)attr + sizeof(*attr);
3809 for (; addr < end; addr++) {
3817 ret = copy_to_user(uattr, attr, attr->size);
3830 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3831 * @pid: the pid in question.
3832 * @uattr: structure containing the extended parameters.
3833 * @size: sizeof(attr) for fwd/bwd comp.
3834 * @flags: for future extension.
3836 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3837 unsigned int, size, unsigned int, flags)
3839 struct sched_attr attr = {
3840 .size = sizeof(struct sched_attr),
3842 struct task_struct *p;
3845 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3846 size < SCHED_ATTR_SIZE_VER0 || flags)
3850 p = find_process_by_pid(pid);
3855 retval = security_task_getscheduler(p);
3859 attr.sched_policy = p->policy;
3860 if (p->sched_reset_on_fork)
3861 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3862 if (task_has_dl_policy(p))
3863 __getparam_dl(p, &attr);
3864 else if (task_has_rt_policy(p))
3865 attr.sched_priority = p->rt_priority;
3867 attr.sched_nice = task_nice(p);
3871 retval = sched_read_attr(uattr, &attr, size);
3879 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3881 cpumask_var_t cpus_allowed, new_mask;
3882 struct task_struct *p;
3887 p = find_process_by_pid(pid);
3893 /* Prevent p going away */
3897 if (p->flags & PF_NO_SETAFFINITY) {
3901 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3905 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3907 goto out_free_cpus_allowed;
3910 if (!check_same_owner(p)) {
3912 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3919 retval = security_task_setscheduler(p);
3924 cpuset_cpus_allowed(p, cpus_allowed);
3925 cpumask_and(new_mask, in_mask, cpus_allowed);
3928 * Since bandwidth control happens on root_domain basis,
3929 * if admission test is enabled, we only admit -deadline
3930 * tasks allowed to run on all the CPUs in the task's
3934 if (task_has_dl_policy(p)) {
3935 const struct cpumask *span = task_rq(p)->rd->span;
3937 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3944 retval = set_cpus_allowed_ptr(p, new_mask);
3947 cpuset_cpus_allowed(p, cpus_allowed);
3948 if (!cpumask_subset(new_mask, cpus_allowed)) {
3950 * We must have raced with a concurrent cpuset
3951 * update. Just reset the cpus_allowed to the
3952 * cpuset's cpus_allowed
3954 cpumask_copy(new_mask, cpus_allowed);
3959 free_cpumask_var(new_mask);
3960 out_free_cpus_allowed:
3961 free_cpumask_var(cpus_allowed);
3967 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3968 struct cpumask *new_mask)
3970 if (len < cpumask_size())
3971 cpumask_clear(new_mask);
3972 else if (len > cpumask_size())
3973 len = cpumask_size();
3975 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3979 * sys_sched_setaffinity - set the cpu affinity of a process
3980 * @pid: pid of the process
3981 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3982 * @user_mask_ptr: user-space pointer to the new cpu mask
3984 * Return: 0 on success. An error code otherwise.
3986 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3987 unsigned long __user *, user_mask_ptr)
3989 cpumask_var_t new_mask;
3992 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3995 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3997 retval = sched_setaffinity(pid, new_mask);
3998 free_cpumask_var(new_mask);
4002 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4004 struct task_struct *p;
4005 unsigned long flags;
4011 p = find_process_by_pid(pid);
4015 retval = security_task_getscheduler(p);
4019 raw_spin_lock_irqsave(&p->pi_lock, flags);
4020 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4021 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4030 * sys_sched_getaffinity - get the cpu affinity of a process
4031 * @pid: pid of the process
4032 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4033 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4035 * Return: 0 on success. An error code otherwise.
4037 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4038 unsigned long __user *, user_mask_ptr)
4043 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4045 if (len & (sizeof(unsigned long)-1))
4048 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4051 ret = sched_getaffinity(pid, mask);
4053 size_t retlen = min_t(size_t, len, cpumask_size());
4055 if (copy_to_user(user_mask_ptr, mask, retlen))
4060 free_cpumask_var(mask);
4066 * sys_sched_yield - yield the current processor to other threads.
4068 * This function yields the current CPU to other tasks. If there are no
4069 * other threads running on this CPU then this function will return.
4073 SYSCALL_DEFINE0(sched_yield)
4075 struct rq *rq = this_rq_lock();
4077 schedstat_inc(rq, yld_count);
4078 current->sched_class->yield_task(rq);
4081 * Since we are going to call schedule() anyway, there's
4082 * no need to preempt or enable interrupts:
4084 __release(rq->lock);
4085 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4086 do_raw_spin_unlock(&rq->lock);
4087 sched_preempt_enable_no_resched();
4094 static void __cond_resched(void)
4096 __preempt_count_add(PREEMPT_ACTIVE);
4098 __preempt_count_sub(PREEMPT_ACTIVE);
4101 int __sched _cond_resched(void)
4104 if (should_resched()) {
4110 EXPORT_SYMBOL(_cond_resched);
4113 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4114 * call schedule, and on return reacquire the lock.
4116 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4117 * operations here to prevent schedule() from being called twice (once via
4118 * spin_unlock(), once by hand).
4120 int __cond_resched_lock(spinlock_t *lock)
4122 bool need_rcu_resched = rcu_should_resched();
4123 int resched = should_resched();
4126 lockdep_assert_held(lock);
4128 if (spin_needbreak(lock) || resched || need_rcu_resched) {
4132 else if (unlikely(need_rcu_resched))
4141 EXPORT_SYMBOL(__cond_resched_lock);
4143 int __sched __cond_resched_softirq(void)
4145 BUG_ON(!in_softirq());
4147 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4148 if (should_resched()) {
4156 EXPORT_SYMBOL(__cond_resched_softirq);
4159 * yield - yield the current processor to other threads.
4161 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4163 * The scheduler is at all times free to pick the calling task as the most
4164 * eligible task to run, if removing the yield() call from your code breaks
4165 * it, its already broken.
4167 * Typical broken usage is:
4172 * where one assumes that yield() will let 'the other' process run that will
4173 * make event true. If the current task is a SCHED_FIFO task that will never
4174 * happen. Never use yield() as a progress guarantee!!
4176 * If you want to use yield() to wait for something, use wait_event().
4177 * If you want to use yield() to be 'nice' for others, use cond_resched().
4178 * If you still want to use yield(), do not!
4180 void __sched yield(void)
4182 set_current_state(TASK_RUNNING);
4185 EXPORT_SYMBOL(yield);
4188 * yield_to - yield the current processor to another thread in
4189 * your thread group, or accelerate that thread toward the
4190 * processor it's on.
4192 * @preempt: whether task preemption is allowed or not
4194 * It's the caller's job to ensure that the target task struct
4195 * can't go away on us before we can do any checks.
4198 * true (>0) if we indeed boosted the target task.
4199 * false (0) if we failed to boost the target.
4200 * -ESRCH if there's no task to yield to.
4202 bool __sched yield_to(struct task_struct *p, bool preempt)
4204 struct task_struct *curr = current;
4205 struct rq *rq, *p_rq;
4206 unsigned long flags;
4209 local_irq_save(flags);
4215 * If we're the only runnable task on the rq and target rq also
4216 * has only one task, there's absolutely no point in yielding.
4218 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4223 double_rq_lock(rq, p_rq);
4224 if (task_rq(p) != p_rq) {
4225 double_rq_unlock(rq, p_rq);
4229 if (!curr->sched_class->yield_to_task)
4232 if (curr->sched_class != p->sched_class)
4235 if (task_running(p_rq, p) || p->state)
4238 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4240 schedstat_inc(rq, yld_count);
4242 * Make p's CPU reschedule; pick_next_entity takes care of
4245 if (preempt && rq != p_rq)
4246 resched_task(p_rq->curr);
4250 double_rq_unlock(rq, p_rq);
4252 local_irq_restore(flags);
4259 EXPORT_SYMBOL_GPL(yield_to);
4262 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4263 * that process accounting knows that this is a task in IO wait state.
4265 void __sched io_schedule(void)
4267 struct rq *rq = raw_rq();
4269 delayacct_blkio_start();
4270 atomic_inc(&rq->nr_iowait);
4271 blk_flush_plug(current);
4272 current->in_iowait = 1;
4274 current->in_iowait = 0;
4275 atomic_dec(&rq->nr_iowait);
4276 delayacct_blkio_end();
4278 EXPORT_SYMBOL(io_schedule);
4280 long __sched io_schedule_timeout(long timeout)
4282 struct rq *rq = raw_rq();
4285 delayacct_blkio_start();
4286 atomic_inc(&rq->nr_iowait);
4287 blk_flush_plug(current);
4288 current->in_iowait = 1;
4289 ret = schedule_timeout(timeout);
4290 current->in_iowait = 0;
4291 atomic_dec(&rq->nr_iowait);
4292 delayacct_blkio_end();
4297 * sys_sched_get_priority_max - return maximum RT priority.
4298 * @policy: scheduling class.
4300 * Return: On success, this syscall returns the maximum
4301 * rt_priority that can be used by a given scheduling class.
4302 * On failure, a negative error code is returned.
4304 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4311 ret = MAX_USER_RT_PRIO-1;
4313 case SCHED_DEADLINE:
4324 * sys_sched_get_priority_min - return minimum RT priority.
4325 * @policy: scheduling class.
4327 * Return: On success, this syscall returns the minimum
4328 * rt_priority that can be used by a given scheduling class.
4329 * On failure, a negative error code is returned.
4331 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4340 case SCHED_DEADLINE:
4350 * sys_sched_rr_get_interval - return the default timeslice of a process.
4351 * @pid: pid of the process.
4352 * @interval: userspace pointer to the timeslice value.
4354 * this syscall writes the default timeslice value of a given process
4355 * into the user-space timespec buffer. A value of '0' means infinity.
4357 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4360 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4361 struct timespec __user *, interval)
4363 struct task_struct *p;
4364 unsigned int time_slice;
4365 unsigned long flags;
4375 p = find_process_by_pid(pid);
4379 retval = security_task_getscheduler(p);
4383 rq = task_rq_lock(p, &flags);
4385 if (p->sched_class->get_rr_interval)
4386 time_slice = p->sched_class->get_rr_interval(rq, p);
4387 task_rq_unlock(rq, p, &flags);
4390 jiffies_to_timespec(time_slice, &t);
4391 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4399 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4401 void sched_show_task(struct task_struct *p)
4403 unsigned long free = 0;
4407 state = p->state ? __ffs(p->state) + 1 : 0;
4408 printk(KERN_INFO "%-15.15s %c", p->comm,
4409 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4410 #if BITS_PER_LONG == 32
4411 if (state == TASK_RUNNING)
4412 printk(KERN_CONT " running ");
4414 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4416 if (state == TASK_RUNNING)
4417 printk(KERN_CONT " running task ");
4419 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4421 #ifdef CONFIG_DEBUG_STACK_USAGE
4422 free = stack_not_used(p);
4425 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4427 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4428 task_pid_nr(p), ppid,
4429 (unsigned long)task_thread_info(p)->flags);
4431 print_worker_info(KERN_INFO, p);
4432 show_stack(p, NULL);
4435 void show_state_filter(unsigned long state_filter)
4437 struct task_struct *g, *p;
4439 #if BITS_PER_LONG == 32
4441 " task PC stack pid father\n");
4444 " task PC stack pid father\n");
4447 do_each_thread(g, p) {
4449 * reset the NMI-timeout, listing all files on a slow
4450 * console might take a lot of time:
4452 touch_nmi_watchdog();
4453 if (!state_filter || (p->state & state_filter))
4455 } while_each_thread(g, p);
4457 touch_all_softlockup_watchdogs();
4459 #ifdef CONFIG_SCHED_DEBUG
4460 sysrq_sched_debug_show();
4464 * Only show locks if all tasks are dumped:
4467 debug_show_all_locks();
4470 void init_idle_bootup_task(struct task_struct *idle)
4472 idle->sched_class = &idle_sched_class;
4476 * init_idle - set up an idle thread for a given CPU
4477 * @idle: task in question
4478 * @cpu: cpu the idle task belongs to
4480 * NOTE: this function does not set the idle thread's NEED_RESCHED
4481 * flag, to make booting more robust.
4483 void init_idle(struct task_struct *idle, int cpu)
4485 struct rq *rq = cpu_rq(cpu);
4486 unsigned long flags;
4488 raw_spin_lock_irqsave(&rq->lock, flags);
4490 __sched_fork(0, idle);
4491 idle->state = TASK_RUNNING;
4492 idle->se.exec_start = sched_clock();
4494 do_set_cpus_allowed(idle, cpumask_of(cpu));
4496 * We're having a chicken and egg problem, even though we are
4497 * holding rq->lock, the cpu isn't yet set to this cpu so the
4498 * lockdep check in task_group() will fail.
4500 * Similar case to sched_fork(). / Alternatively we could
4501 * use task_rq_lock() here and obtain the other rq->lock.
4506 __set_task_cpu(idle, cpu);
4509 rq->curr = rq->idle = idle;
4511 #if defined(CONFIG_SMP)
4514 raw_spin_unlock_irqrestore(&rq->lock, flags);
4516 /* Set the preempt count _outside_ the spinlocks! */
4517 init_idle_preempt_count(idle, cpu);
4520 * The idle tasks have their own, simple scheduling class:
4522 idle->sched_class = &idle_sched_class;
4523 ftrace_graph_init_idle_task(idle, cpu);
4524 vtime_init_idle(idle, cpu);
4525 #if defined(CONFIG_SMP)
4526 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4531 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4533 if (p->sched_class && p->sched_class->set_cpus_allowed)
4534 p->sched_class->set_cpus_allowed(p, new_mask);
4536 cpumask_copy(&p->cpus_allowed, new_mask);
4537 p->nr_cpus_allowed = cpumask_weight(new_mask);
4541 * This is how migration works:
4543 * 1) we invoke migration_cpu_stop() on the target CPU using
4545 * 2) stopper starts to run (implicitly forcing the migrated thread
4547 * 3) it checks whether the migrated task is still in the wrong runqueue.
4548 * 4) if it's in the wrong runqueue then the migration thread removes
4549 * it and puts it into the right queue.
4550 * 5) stopper completes and stop_one_cpu() returns and the migration
4555 * Change a given task's CPU affinity. Migrate the thread to a
4556 * proper CPU and schedule it away if the CPU it's executing on
4557 * is removed from the allowed bitmask.
4559 * NOTE: the caller must have a valid reference to the task, the
4560 * task must not exit() & deallocate itself prematurely. The
4561 * call is not atomic; no spinlocks may be held.
4563 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4565 unsigned long flags;
4567 unsigned int dest_cpu;
4570 rq = task_rq_lock(p, &flags);
4572 if (cpumask_equal(&p->cpus_allowed, new_mask))
4575 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4580 do_set_cpus_allowed(p, new_mask);
4582 /* Can the task run on the task's current CPU? If so, we're done */
4583 if (cpumask_test_cpu(task_cpu(p), new_mask))
4586 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4588 struct migration_arg arg = { p, dest_cpu };
4589 /* Need help from migration thread: drop lock and wait. */
4590 task_rq_unlock(rq, p, &flags);
4591 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4592 tlb_migrate_finish(p->mm);
4596 task_rq_unlock(rq, p, &flags);
4600 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4603 * Move (not current) task off this cpu, onto dest cpu. We're doing
4604 * this because either it can't run here any more (set_cpus_allowed()
4605 * away from this CPU, or CPU going down), or because we're
4606 * attempting to rebalance this task on exec (sched_exec).
4608 * So we race with normal scheduler movements, but that's OK, as long
4609 * as the task is no longer on this CPU.
4611 * Returns non-zero if task was successfully migrated.
4613 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4615 struct rq *rq_dest, *rq_src;
4618 if (unlikely(!cpu_active(dest_cpu)))
4621 rq_src = cpu_rq(src_cpu);
4622 rq_dest = cpu_rq(dest_cpu);
4624 raw_spin_lock(&p->pi_lock);
4625 double_rq_lock(rq_src, rq_dest);
4626 /* Already moved. */
4627 if (task_cpu(p) != src_cpu)
4629 /* Affinity changed (again). */
4630 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4634 * If we're not on a rq, the next wake-up will ensure we're
4638 dequeue_task(rq_src, p, 0);
4639 set_task_cpu(p, dest_cpu);
4640 enqueue_task(rq_dest, p, 0);
4641 check_preempt_curr(rq_dest, p, 0);
4646 double_rq_unlock(rq_src, rq_dest);
4647 raw_spin_unlock(&p->pi_lock);
4651 #ifdef CONFIG_NUMA_BALANCING
4652 /* Migrate current task p to target_cpu */
4653 int migrate_task_to(struct task_struct *p, int target_cpu)
4655 struct migration_arg arg = { p, target_cpu };
4656 int curr_cpu = task_cpu(p);
4658 if (curr_cpu == target_cpu)
4661 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4664 /* TODO: This is not properly updating schedstats */
4666 trace_sched_move_numa(p, curr_cpu, target_cpu);
4667 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4671 * Requeue a task on a given node and accurately track the number of NUMA
4672 * tasks on the runqueues
4674 void sched_setnuma(struct task_struct *p, int nid)
4677 unsigned long flags;
4678 bool on_rq, running;
4680 rq = task_rq_lock(p, &flags);
4682 running = task_current(rq, p);
4685 dequeue_task(rq, p, 0);
4687 p->sched_class->put_prev_task(rq, p);
4689 p->numa_preferred_nid = nid;
4692 p->sched_class->set_curr_task(rq);
4694 enqueue_task(rq, p, 0);
4695 task_rq_unlock(rq, p, &flags);
4700 * migration_cpu_stop - this will be executed by a highprio stopper thread
4701 * and performs thread migration by bumping thread off CPU then
4702 * 'pushing' onto another runqueue.
4704 static int migration_cpu_stop(void *data)
4706 struct migration_arg *arg = data;
4709 * The original target cpu might have gone down and we might
4710 * be on another cpu but it doesn't matter.
4712 local_irq_disable();
4713 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4718 #ifdef CONFIG_HOTPLUG_CPU
4721 * Ensures that the idle task is using init_mm right before its cpu goes
4724 void idle_task_exit(void)
4726 struct mm_struct *mm = current->active_mm;
4728 BUG_ON(cpu_online(smp_processor_id()));
4730 if (mm != &init_mm) {
4731 switch_mm(mm, &init_mm, current);
4732 finish_arch_post_lock_switch();
4738 * Since this CPU is going 'away' for a while, fold any nr_active delta
4739 * we might have. Assumes we're called after migrate_tasks() so that the
4740 * nr_active count is stable.
4742 * Also see the comment "Global load-average calculations".
4744 static void calc_load_migrate(struct rq *rq)
4746 long delta = calc_load_fold_active(rq);
4748 atomic_long_add(delta, &calc_load_tasks);
4751 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4755 static const struct sched_class fake_sched_class = {
4756 .put_prev_task = put_prev_task_fake,
4759 static struct task_struct fake_task = {
4761 * Avoid pull_{rt,dl}_task()
4763 .prio = MAX_PRIO + 1,
4764 .sched_class = &fake_sched_class,
4768 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4769 * try_to_wake_up()->select_task_rq().
4771 * Called with rq->lock held even though we'er in stop_machine() and
4772 * there's no concurrency possible, we hold the required locks anyway
4773 * because of lock validation efforts.
4775 static void migrate_tasks(unsigned int dead_cpu)
4777 struct rq *rq = cpu_rq(dead_cpu);
4778 struct task_struct *next, *stop = rq->stop;
4782 * Fudge the rq selection such that the below task selection loop
4783 * doesn't get stuck on the currently eligible stop task.
4785 * We're currently inside stop_machine() and the rq is either stuck
4786 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4787 * either way we should never end up calling schedule() until we're
4793 * put_prev_task() and pick_next_task() sched
4794 * class method both need to have an up-to-date
4795 * value of rq->clock[_task]
4797 update_rq_clock(rq);
4801 * There's this thread running, bail when that's the only
4804 if (rq->nr_running == 1)
4807 next = pick_next_task(rq, &fake_task);
4809 next->sched_class->put_prev_task(rq, next);
4811 /* Find suitable destination for @next, with force if needed. */
4812 dest_cpu = select_fallback_rq(dead_cpu, next);
4813 raw_spin_unlock(&rq->lock);
4815 __migrate_task(next, dead_cpu, dest_cpu);
4817 raw_spin_lock(&rq->lock);
4823 #endif /* CONFIG_HOTPLUG_CPU */
4825 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4827 static struct ctl_table sd_ctl_dir[] = {
4829 .procname = "sched_domain",
4835 static struct ctl_table sd_ctl_root[] = {
4837 .procname = "kernel",
4839 .child = sd_ctl_dir,
4844 static struct ctl_table *sd_alloc_ctl_entry(int n)
4846 struct ctl_table *entry =
4847 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4852 static void sd_free_ctl_entry(struct ctl_table **tablep)
4854 struct ctl_table *entry;
4857 * In the intermediate directories, both the child directory and
4858 * procname are dynamically allocated and could fail but the mode
4859 * will always be set. In the lowest directory the names are
4860 * static strings and all have proc handlers.
4862 for (entry = *tablep; entry->mode; entry++) {
4864 sd_free_ctl_entry(&entry->child);
4865 if (entry->proc_handler == NULL)
4866 kfree(entry->procname);
4873 static int min_load_idx = 0;
4874 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4877 set_table_entry(struct ctl_table *entry,
4878 const char *procname, void *data, int maxlen,
4879 umode_t mode, proc_handler *proc_handler,
4882 entry->procname = procname;
4884 entry->maxlen = maxlen;
4886 entry->proc_handler = proc_handler;
4889 entry->extra1 = &min_load_idx;
4890 entry->extra2 = &max_load_idx;
4894 static struct ctl_table *
4895 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4897 struct ctl_table *table = sd_alloc_ctl_entry(14);
4902 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4903 sizeof(long), 0644, proc_doulongvec_minmax, false);
4904 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4905 sizeof(long), 0644, proc_doulongvec_minmax, false);
4906 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4907 sizeof(int), 0644, proc_dointvec_minmax, true);
4908 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4909 sizeof(int), 0644, proc_dointvec_minmax, true);
4910 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4911 sizeof(int), 0644, proc_dointvec_minmax, true);
4912 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4913 sizeof(int), 0644, proc_dointvec_minmax, true);
4914 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4915 sizeof(int), 0644, proc_dointvec_minmax, true);
4916 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4917 sizeof(int), 0644, proc_dointvec_minmax, false);
4918 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4919 sizeof(int), 0644, proc_dointvec_minmax, false);
4920 set_table_entry(&table[9], "cache_nice_tries",
4921 &sd->cache_nice_tries,
4922 sizeof(int), 0644, proc_dointvec_minmax, false);
4923 set_table_entry(&table[10], "flags", &sd->flags,
4924 sizeof(int), 0644, proc_dointvec_minmax, false);
4925 set_table_entry(&table[11], "max_newidle_lb_cost",
4926 &sd->max_newidle_lb_cost,
4927 sizeof(long), 0644, proc_doulongvec_minmax, false);
4928 set_table_entry(&table[12], "name", sd->name,
4929 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4930 /* &table[13] is terminator */
4935 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4937 struct ctl_table *entry, *table;
4938 struct sched_domain *sd;
4939 int domain_num = 0, i;
4942 for_each_domain(cpu, sd)
4944 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4949 for_each_domain(cpu, sd) {
4950 snprintf(buf, 32, "domain%d", i);
4951 entry->procname = kstrdup(buf, GFP_KERNEL);
4953 entry->child = sd_alloc_ctl_domain_table(sd);
4960 static struct ctl_table_header *sd_sysctl_header;
4961 static void register_sched_domain_sysctl(void)
4963 int i, cpu_num = num_possible_cpus();
4964 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4967 WARN_ON(sd_ctl_dir[0].child);
4968 sd_ctl_dir[0].child = entry;
4973 for_each_possible_cpu(i) {
4974 snprintf(buf, 32, "cpu%d", i);
4975 entry->procname = kstrdup(buf, GFP_KERNEL);
4977 entry->child = sd_alloc_ctl_cpu_table(i);
4981 WARN_ON(sd_sysctl_header);
4982 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4985 /* may be called multiple times per register */
4986 static void unregister_sched_domain_sysctl(void)
4988 if (sd_sysctl_header)
4989 unregister_sysctl_table(sd_sysctl_header);
4990 sd_sysctl_header = NULL;
4991 if (sd_ctl_dir[0].child)
4992 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4995 static void register_sched_domain_sysctl(void)
4998 static void unregister_sched_domain_sysctl(void)
5003 static void set_rq_online(struct rq *rq)
5006 const struct sched_class *class;
5008 cpumask_set_cpu(rq->cpu, rq->rd->online);
5011 for_each_class(class) {
5012 if (class->rq_online)
5013 class->rq_online(rq);
5018 static void set_rq_offline(struct rq *rq)
5021 const struct sched_class *class;
5023 for_each_class(class) {
5024 if (class->rq_offline)
5025 class->rq_offline(rq);
5028 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5034 * migration_call - callback that gets triggered when a CPU is added.
5035 * Here we can start up the necessary migration thread for the new CPU.
5038 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5040 int cpu = (long)hcpu;
5041 unsigned long flags;
5042 struct rq *rq = cpu_rq(cpu);
5044 switch (action & ~CPU_TASKS_FROZEN) {
5046 case CPU_UP_PREPARE:
5047 rq->calc_load_update = calc_load_update;
5051 /* Update our root-domain */
5052 raw_spin_lock_irqsave(&rq->lock, flags);
5054 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5058 raw_spin_unlock_irqrestore(&rq->lock, flags);
5061 #ifdef CONFIG_HOTPLUG_CPU
5063 sched_ttwu_pending();
5064 /* Update our root-domain */
5065 raw_spin_lock_irqsave(&rq->lock, flags);
5067 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5071 BUG_ON(rq->nr_running != 1); /* the migration thread */
5072 raw_spin_unlock_irqrestore(&rq->lock, flags);
5076 calc_load_migrate(rq);
5081 update_max_interval();
5087 * Register at high priority so that task migration (migrate_all_tasks)
5088 * happens before everything else. This has to be lower priority than
5089 * the notifier in the perf_event subsystem, though.
5091 static struct notifier_block migration_notifier = {
5092 .notifier_call = migration_call,
5093 .priority = CPU_PRI_MIGRATION,
5096 static int sched_cpu_active(struct notifier_block *nfb,
5097 unsigned long action, void *hcpu)
5099 switch (action & ~CPU_TASKS_FROZEN) {
5100 case CPU_DOWN_FAILED:
5101 set_cpu_active((long)hcpu, true);
5108 static int sched_cpu_inactive(struct notifier_block *nfb,
5109 unsigned long action, void *hcpu)
5111 unsigned long flags;
5112 long cpu = (long)hcpu;
5114 switch (action & ~CPU_TASKS_FROZEN) {
5115 case CPU_DOWN_PREPARE:
5116 set_cpu_active(cpu, false);
5118 /* explicitly allow suspend */
5119 if (!(action & CPU_TASKS_FROZEN)) {
5120 struct dl_bw *dl_b = dl_bw_of(cpu);
5124 raw_spin_lock_irqsave(&dl_b->lock, flags);
5125 cpus = dl_bw_cpus(cpu);
5126 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5127 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5130 return notifier_from_errno(-EBUSY);
5138 static int __init migration_init(void)
5140 void *cpu = (void *)(long)smp_processor_id();
5143 /* Initialize migration for the boot CPU */
5144 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5145 BUG_ON(err == NOTIFY_BAD);
5146 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5147 register_cpu_notifier(&migration_notifier);
5149 /* Register cpu active notifiers */
5150 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5151 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5155 early_initcall(migration_init);
5160 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5162 #ifdef CONFIG_SCHED_DEBUG
5164 static __read_mostly int sched_debug_enabled;
5166 static int __init sched_debug_setup(char *str)
5168 sched_debug_enabled = 1;
5172 early_param("sched_debug", sched_debug_setup);
5174 static inline bool sched_debug(void)
5176 return sched_debug_enabled;
5179 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5180 struct cpumask *groupmask)
5182 struct sched_group *group = sd->groups;
5185 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5186 cpumask_clear(groupmask);
5188 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5190 if (!(sd->flags & SD_LOAD_BALANCE)) {
5191 printk("does not load-balance\n");
5193 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5198 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5200 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5201 printk(KERN_ERR "ERROR: domain->span does not contain "
5204 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5205 printk(KERN_ERR "ERROR: domain->groups does not contain"
5209 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5213 printk(KERN_ERR "ERROR: group is NULL\n");
5218 * Even though we initialize ->power to something semi-sane,
5219 * we leave power_orig unset. This allows us to detect if
5220 * domain iteration is still funny without causing /0 traps.
5222 if (!group->sgp->power_orig) {
5223 printk(KERN_CONT "\n");
5224 printk(KERN_ERR "ERROR: domain->cpu_power not "
5229 if (!cpumask_weight(sched_group_cpus(group))) {
5230 printk(KERN_CONT "\n");
5231 printk(KERN_ERR "ERROR: empty group\n");
5235 if (!(sd->flags & SD_OVERLAP) &&
5236 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5237 printk(KERN_CONT "\n");
5238 printk(KERN_ERR "ERROR: repeated CPUs\n");
5242 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5244 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5246 printk(KERN_CONT " %s", str);
5247 if (group->sgp->power != SCHED_POWER_SCALE) {
5248 printk(KERN_CONT " (cpu_power = %d)",
5252 group = group->next;
5253 } while (group != sd->groups);
5254 printk(KERN_CONT "\n");
5256 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5257 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5260 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5261 printk(KERN_ERR "ERROR: parent span is not a superset "
5262 "of domain->span\n");
5266 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5270 if (!sched_debug_enabled)
5274 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5278 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5281 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5289 #else /* !CONFIG_SCHED_DEBUG */
5290 # define sched_domain_debug(sd, cpu) do { } while (0)
5291 static inline bool sched_debug(void)
5295 #endif /* CONFIG_SCHED_DEBUG */
5297 static int sd_degenerate(struct sched_domain *sd)
5299 if (cpumask_weight(sched_domain_span(sd)) == 1)
5302 /* Following flags need at least 2 groups */
5303 if (sd->flags & (SD_LOAD_BALANCE |
5304 SD_BALANCE_NEWIDLE |
5308 SD_SHARE_PKG_RESOURCES)) {
5309 if (sd->groups != sd->groups->next)
5313 /* Following flags don't use groups */
5314 if (sd->flags & (SD_WAKE_AFFINE))
5321 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5323 unsigned long cflags = sd->flags, pflags = parent->flags;
5325 if (sd_degenerate(parent))
5328 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5331 /* Flags needing groups don't count if only 1 group in parent */
5332 if (parent->groups == parent->groups->next) {
5333 pflags &= ~(SD_LOAD_BALANCE |
5334 SD_BALANCE_NEWIDLE |
5338 SD_SHARE_PKG_RESOURCES |
5340 if (nr_node_ids == 1)
5341 pflags &= ~SD_SERIALIZE;
5343 if (~cflags & pflags)
5349 static void free_rootdomain(struct rcu_head *rcu)
5351 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5353 cpupri_cleanup(&rd->cpupri);
5354 cpudl_cleanup(&rd->cpudl);
5355 free_cpumask_var(rd->dlo_mask);
5356 free_cpumask_var(rd->rto_mask);
5357 free_cpumask_var(rd->online);
5358 free_cpumask_var(rd->span);
5362 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5364 struct root_domain *old_rd = NULL;
5365 unsigned long flags;
5367 raw_spin_lock_irqsave(&rq->lock, flags);
5372 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5375 cpumask_clear_cpu(rq->cpu, old_rd->span);
5378 * If we dont want to free the old_rd yet then
5379 * set old_rd to NULL to skip the freeing later
5382 if (!atomic_dec_and_test(&old_rd->refcount))
5386 atomic_inc(&rd->refcount);
5389 cpumask_set_cpu(rq->cpu, rd->span);
5390 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5393 raw_spin_unlock_irqrestore(&rq->lock, flags);
5396 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5399 static int init_rootdomain(struct root_domain *rd)
5401 memset(rd, 0, sizeof(*rd));
5403 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5405 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5407 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5409 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5412 init_dl_bw(&rd->dl_bw);
5413 if (cpudl_init(&rd->cpudl) != 0)
5416 if (cpupri_init(&rd->cpupri) != 0)
5421 free_cpumask_var(rd->rto_mask);
5423 free_cpumask_var(rd->dlo_mask);
5425 free_cpumask_var(rd->online);
5427 free_cpumask_var(rd->span);
5433 * By default the system creates a single root-domain with all cpus as
5434 * members (mimicking the global state we have today).
5436 struct root_domain def_root_domain;
5438 static void init_defrootdomain(void)
5440 init_rootdomain(&def_root_domain);
5442 atomic_set(&def_root_domain.refcount, 1);
5445 static struct root_domain *alloc_rootdomain(void)
5447 struct root_domain *rd;
5449 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5453 if (init_rootdomain(rd) != 0) {
5461 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5463 struct sched_group *tmp, *first;
5472 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5477 } while (sg != first);
5480 static void free_sched_domain(struct rcu_head *rcu)
5482 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5485 * If its an overlapping domain it has private groups, iterate and
5488 if (sd->flags & SD_OVERLAP) {
5489 free_sched_groups(sd->groups, 1);
5490 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5491 kfree(sd->groups->sgp);
5497 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5499 call_rcu(&sd->rcu, free_sched_domain);
5502 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5504 for (; sd; sd = sd->parent)
5505 destroy_sched_domain(sd, cpu);
5509 * Keep a special pointer to the highest sched_domain that has
5510 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5511 * allows us to avoid some pointer chasing select_idle_sibling().
5513 * Also keep a unique ID per domain (we use the first cpu number in
5514 * the cpumask of the domain), this allows us to quickly tell if
5515 * two cpus are in the same cache domain, see cpus_share_cache().
5517 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5518 DEFINE_PER_CPU(int, sd_llc_size);
5519 DEFINE_PER_CPU(int, sd_llc_id);
5520 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5521 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5522 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5524 static void update_top_cache_domain(int cpu)
5526 struct sched_domain *sd;
5527 struct sched_domain *busy_sd = NULL;
5531 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5533 id = cpumask_first(sched_domain_span(sd));
5534 size = cpumask_weight(sched_domain_span(sd));
5535 busy_sd = sd->parent; /* sd_busy */
5537 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5539 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5540 per_cpu(sd_llc_size, cpu) = size;
5541 per_cpu(sd_llc_id, cpu) = id;
5543 sd = lowest_flag_domain(cpu, SD_NUMA);
5544 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5546 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5547 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5551 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5552 * hold the hotplug lock.
5555 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5557 struct rq *rq = cpu_rq(cpu);
5558 struct sched_domain *tmp;
5560 /* Remove the sched domains which do not contribute to scheduling. */
5561 for (tmp = sd; tmp; ) {
5562 struct sched_domain *parent = tmp->parent;
5566 if (sd_parent_degenerate(tmp, parent)) {
5567 tmp->parent = parent->parent;
5569 parent->parent->child = tmp;
5571 * Transfer SD_PREFER_SIBLING down in case of a
5572 * degenerate parent; the spans match for this
5573 * so the property transfers.
5575 if (parent->flags & SD_PREFER_SIBLING)
5576 tmp->flags |= SD_PREFER_SIBLING;
5577 destroy_sched_domain(parent, cpu);
5582 if (sd && sd_degenerate(sd)) {
5585 destroy_sched_domain(tmp, cpu);
5590 sched_domain_debug(sd, cpu);
5592 rq_attach_root(rq, rd);
5594 rcu_assign_pointer(rq->sd, sd);
5595 destroy_sched_domains(tmp, cpu);
5597 update_top_cache_domain(cpu);
5600 /* cpus with isolated domains */
5601 static cpumask_var_t cpu_isolated_map;
5603 /* Setup the mask of cpus configured for isolated domains */
5604 static int __init isolated_cpu_setup(char *str)
5606 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5607 cpulist_parse(str, cpu_isolated_map);
5611 __setup("isolcpus=", isolated_cpu_setup);
5613 static const struct cpumask *cpu_cpu_mask(int cpu)
5615 return cpumask_of_node(cpu_to_node(cpu));
5619 struct sched_domain **__percpu sd;
5620 struct sched_group **__percpu sg;
5621 struct sched_group_power **__percpu sgp;
5625 struct sched_domain ** __percpu sd;
5626 struct root_domain *rd;
5636 struct sched_domain_topology_level;
5638 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5639 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5641 #define SDTL_OVERLAP 0x01
5643 struct sched_domain_topology_level {
5644 sched_domain_init_f init;
5645 sched_domain_mask_f mask;
5648 struct sd_data data;
5652 * Build an iteration mask that can exclude certain CPUs from the upwards
5655 * Asymmetric node setups can result in situations where the domain tree is of
5656 * unequal depth, make sure to skip domains that already cover the entire
5659 * In that case build_sched_domains() will have terminated the iteration early
5660 * and our sibling sd spans will be empty. Domains should always include the
5661 * cpu they're built on, so check that.
5664 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5666 const struct cpumask *span = sched_domain_span(sd);
5667 struct sd_data *sdd = sd->private;
5668 struct sched_domain *sibling;
5671 for_each_cpu(i, span) {
5672 sibling = *per_cpu_ptr(sdd->sd, i);
5673 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5676 cpumask_set_cpu(i, sched_group_mask(sg));
5681 * Return the canonical balance cpu for this group, this is the first cpu
5682 * of this group that's also in the iteration mask.
5684 int group_balance_cpu(struct sched_group *sg)
5686 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5690 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5692 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5693 const struct cpumask *span = sched_domain_span(sd);
5694 struct cpumask *covered = sched_domains_tmpmask;
5695 struct sd_data *sdd = sd->private;
5696 struct sched_domain *child;
5699 cpumask_clear(covered);
5701 for_each_cpu(i, span) {
5702 struct cpumask *sg_span;
5704 if (cpumask_test_cpu(i, covered))
5707 child = *per_cpu_ptr(sdd->sd, i);
5709 /* See the comment near build_group_mask(). */
5710 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5713 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5714 GFP_KERNEL, cpu_to_node(cpu));
5719 sg_span = sched_group_cpus(sg);
5721 child = child->child;
5722 cpumask_copy(sg_span, sched_domain_span(child));
5724 cpumask_set_cpu(i, sg_span);
5726 cpumask_or(covered, covered, sg_span);
5728 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5729 if (atomic_inc_return(&sg->sgp->ref) == 1)
5730 build_group_mask(sd, sg);
5733 * Initialize sgp->power such that even if we mess up the
5734 * domains and no possible iteration will get us here, we won't
5737 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5738 sg->sgp->power_orig = sg->sgp->power;
5741 * Make sure the first group of this domain contains the
5742 * canonical balance cpu. Otherwise the sched_domain iteration
5743 * breaks. See update_sg_lb_stats().
5745 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5746 group_balance_cpu(sg) == cpu)
5756 sd->groups = groups;
5761 free_sched_groups(first, 0);
5766 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5768 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5769 struct sched_domain *child = sd->child;
5772 cpu = cpumask_first(sched_domain_span(child));
5775 *sg = *per_cpu_ptr(sdd->sg, cpu);
5776 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5777 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5784 * build_sched_groups will build a circular linked list of the groups
5785 * covered by the given span, and will set each group's ->cpumask correctly,
5786 * and ->cpu_power to 0.
5788 * Assumes the sched_domain tree is fully constructed
5791 build_sched_groups(struct sched_domain *sd, int cpu)
5793 struct sched_group *first = NULL, *last = NULL;
5794 struct sd_data *sdd = sd->private;
5795 const struct cpumask *span = sched_domain_span(sd);
5796 struct cpumask *covered;
5799 get_group(cpu, sdd, &sd->groups);
5800 atomic_inc(&sd->groups->ref);
5802 if (cpu != cpumask_first(span))
5805 lockdep_assert_held(&sched_domains_mutex);
5806 covered = sched_domains_tmpmask;
5808 cpumask_clear(covered);
5810 for_each_cpu(i, span) {
5811 struct sched_group *sg;
5814 if (cpumask_test_cpu(i, covered))
5817 group = get_group(i, sdd, &sg);
5818 cpumask_clear(sched_group_cpus(sg));
5820 cpumask_setall(sched_group_mask(sg));
5822 for_each_cpu(j, span) {
5823 if (get_group(j, sdd, NULL) != group)
5826 cpumask_set_cpu(j, covered);
5827 cpumask_set_cpu(j, sched_group_cpus(sg));
5842 * Initialize sched groups cpu_power.
5844 * cpu_power indicates the capacity of sched group, which is used while
5845 * distributing the load between different sched groups in a sched domain.
5846 * Typically cpu_power for all the groups in a sched domain will be same unless
5847 * there are asymmetries in the topology. If there are asymmetries, group
5848 * having more cpu_power will pickup more load compared to the group having
5851 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5853 struct sched_group *sg = sd->groups;
5858 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5860 } while (sg != sd->groups);
5862 if (cpu != group_balance_cpu(sg))
5865 update_group_power(sd, cpu);
5866 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5869 int __weak arch_sd_sibling_asym_packing(void)
5871 return 0*SD_ASYM_PACKING;
5875 * Initializers for schedule domains
5876 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5879 #ifdef CONFIG_SCHED_DEBUG
5880 # define SD_INIT_NAME(sd, type) sd->name = #type
5882 # define SD_INIT_NAME(sd, type) do { } while (0)
5885 #define SD_INIT_FUNC(type) \
5886 static noinline struct sched_domain * \
5887 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5889 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5890 *sd = SD_##type##_INIT; \
5891 SD_INIT_NAME(sd, type); \
5892 sd->private = &tl->data; \
5897 #ifdef CONFIG_SCHED_SMT
5898 SD_INIT_FUNC(SIBLING)
5900 #ifdef CONFIG_SCHED_MC
5903 #ifdef CONFIG_SCHED_BOOK
5907 static int default_relax_domain_level = -1;
5908 int sched_domain_level_max;
5910 static int __init setup_relax_domain_level(char *str)
5912 if (kstrtoint(str, 0, &default_relax_domain_level))
5913 pr_warn("Unable to set relax_domain_level\n");
5917 __setup("relax_domain_level=", setup_relax_domain_level);
5919 static void set_domain_attribute(struct sched_domain *sd,
5920 struct sched_domain_attr *attr)
5924 if (!attr || attr->relax_domain_level < 0) {
5925 if (default_relax_domain_level < 0)
5928 request = default_relax_domain_level;
5930 request = attr->relax_domain_level;
5931 if (request < sd->level) {
5932 /* turn off idle balance on this domain */
5933 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5935 /* turn on idle balance on this domain */
5936 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5940 static void __sdt_free(const struct cpumask *cpu_map);
5941 static int __sdt_alloc(const struct cpumask *cpu_map);
5943 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5944 const struct cpumask *cpu_map)
5948 if (!atomic_read(&d->rd->refcount))
5949 free_rootdomain(&d->rd->rcu); /* fall through */
5951 free_percpu(d->sd); /* fall through */
5953 __sdt_free(cpu_map); /* fall through */
5959 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5960 const struct cpumask *cpu_map)
5962 memset(d, 0, sizeof(*d));
5964 if (__sdt_alloc(cpu_map))
5965 return sa_sd_storage;
5966 d->sd = alloc_percpu(struct sched_domain *);
5968 return sa_sd_storage;
5969 d->rd = alloc_rootdomain();
5972 return sa_rootdomain;
5976 * NULL the sd_data elements we've used to build the sched_domain and
5977 * sched_group structure so that the subsequent __free_domain_allocs()
5978 * will not free the data we're using.
5980 static void claim_allocations(int cpu, struct sched_domain *sd)
5982 struct sd_data *sdd = sd->private;
5984 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5985 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5987 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5988 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5990 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5991 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5994 #ifdef CONFIG_SCHED_SMT
5995 static const struct cpumask *cpu_smt_mask(int cpu)
5997 return topology_thread_cpumask(cpu);
6002 * Topology list, bottom-up.
6004 static struct sched_domain_topology_level default_topology[] = {
6005 #ifdef CONFIG_SCHED_SMT
6006 { sd_init_SIBLING, cpu_smt_mask, },
6008 #ifdef CONFIG_SCHED_MC
6009 { sd_init_MC, cpu_coregroup_mask, },
6011 #ifdef CONFIG_SCHED_BOOK
6012 { sd_init_BOOK, cpu_book_mask, },
6014 { sd_init_CPU, cpu_cpu_mask, },
6018 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6020 #define for_each_sd_topology(tl) \
6021 for (tl = sched_domain_topology; tl->init; tl++)
6025 static int sched_domains_numa_levels;
6026 static int *sched_domains_numa_distance;
6027 static struct cpumask ***sched_domains_numa_masks;
6028 static int sched_domains_curr_level;
6030 static inline int sd_local_flags(int level)
6032 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6035 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6038 static struct sched_domain *
6039 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6041 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6042 int level = tl->numa_level;
6043 int sd_weight = cpumask_weight(
6044 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6046 *sd = (struct sched_domain){
6047 .min_interval = sd_weight,
6048 .max_interval = 2*sd_weight,
6050 .imbalance_pct = 125,
6051 .cache_nice_tries = 2,
6058 .flags = 1*SD_LOAD_BALANCE
6059 | 1*SD_BALANCE_NEWIDLE
6064 | 0*SD_SHARE_CPUPOWER
6065 | 0*SD_SHARE_PKG_RESOURCES
6067 | 0*SD_PREFER_SIBLING
6069 | sd_local_flags(level)
6071 .last_balance = jiffies,
6072 .balance_interval = sd_weight,
6073 .max_newidle_lb_cost = 0,
6074 .next_decay_max_lb_cost = jiffies,
6076 SD_INIT_NAME(sd, NUMA);
6077 sd->private = &tl->data;
6080 * Ugly hack to pass state to sd_numa_mask()...
6082 sched_domains_curr_level = tl->numa_level;
6087 static const struct cpumask *sd_numa_mask(int cpu)
6089 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6092 static void sched_numa_warn(const char *str)
6094 static int done = false;
6102 printk(KERN_WARNING "ERROR: %s\n\n", str);
6104 for (i = 0; i < nr_node_ids; i++) {
6105 printk(KERN_WARNING " ");
6106 for (j = 0; j < nr_node_ids; j++)
6107 printk(KERN_CONT "%02d ", node_distance(i,j));
6108 printk(KERN_CONT "\n");
6110 printk(KERN_WARNING "\n");
6113 static bool find_numa_distance(int distance)
6117 if (distance == node_distance(0, 0))
6120 for (i = 0; i < sched_domains_numa_levels; i++) {
6121 if (sched_domains_numa_distance[i] == distance)
6128 static void sched_init_numa(void)
6130 int next_distance, curr_distance = node_distance(0, 0);
6131 struct sched_domain_topology_level *tl;
6135 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6136 if (!sched_domains_numa_distance)
6140 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6141 * unique distances in the node_distance() table.
6143 * Assumes node_distance(0,j) includes all distances in
6144 * node_distance(i,j) in order to avoid cubic time.
6146 next_distance = curr_distance;
6147 for (i = 0; i < nr_node_ids; i++) {
6148 for (j = 0; j < nr_node_ids; j++) {
6149 for (k = 0; k < nr_node_ids; k++) {
6150 int distance = node_distance(i, k);
6152 if (distance > curr_distance &&
6153 (distance < next_distance ||
6154 next_distance == curr_distance))
6155 next_distance = distance;
6158 * While not a strong assumption it would be nice to know
6159 * about cases where if node A is connected to B, B is not
6160 * equally connected to A.
6162 if (sched_debug() && node_distance(k, i) != distance)
6163 sched_numa_warn("Node-distance not symmetric");
6165 if (sched_debug() && i && !find_numa_distance(distance))
6166 sched_numa_warn("Node-0 not representative");
6168 if (next_distance != curr_distance) {
6169 sched_domains_numa_distance[level++] = next_distance;
6170 sched_domains_numa_levels = level;
6171 curr_distance = next_distance;
6176 * In case of sched_debug() we verify the above assumption.
6182 * 'level' contains the number of unique distances, excluding the
6183 * identity distance node_distance(i,i).
6185 * The sched_domains_numa_distance[] array includes the actual distance
6190 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6191 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6192 * the array will contain less then 'level' members. This could be
6193 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6194 * in other functions.
6196 * We reset it to 'level' at the end of this function.
6198 sched_domains_numa_levels = 0;
6200 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6201 if (!sched_domains_numa_masks)
6205 * Now for each level, construct a mask per node which contains all
6206 * cpus of nodes that are that many hops away from us.
6208 for (i = 0; i < level; i++) {
6209 sched_domains_numa_masks[i] =
6210 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6211 if (!sched_domains_numa_masks[i])
6214 for (j = 0; j < nr_node_ids; j++) {
6215 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6219 sched_domains_numa_masks[i][j] = mask;
6221 for (k = 0; k < nr_node_ids; k++) {
6222 if (node_distance(j, k) > sched_domains_numa_distance[i])
6225 cpumask_or(mask, mask, cpumask_of_node(k));
6230 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6231 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6236 * Copy the default topology bits..
6238 for (i = 0; default_topology[i].init; i++)
6239 tl[i] = default_topology[i];
6242 * .. and append 'j' levels of NUMA goodness.
6244 for (j = 0; j < level; i++, j++) {
6245 tl[i] = (struct sched_domain_topology_level){
6246 .init = sd_numa_init,
6247 .mask = sd_numa_mask,
6248 .flags = SDTL_OVERLAP,
6253 sched_domain_topology = tl;
6255 sched_domains_numa_levels = level;
6258 static void sched_domains_numa_masks_set(int cpu)
6261 int node = cpu_to_node(cpu);
6263 for (i = 0; i < sched_domains_numa_levels; i++) {
6264 for (j = 0; j < nr_node_ids; j++) {
6265 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6266 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6271 static void sched_domains_numa_masks_clear(int cpu)
6274 for (i = 0; i < sched_domains_numa_levels; i++) {
6275 for (j = 0; j < nr_node_ids; j++)
6276 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6281 * Update sched_domains_numa_masks[level][node] array when new cpus
6284 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6285 unsigned long action,
6288 int cpu = (long)hcpu;
6290 switch (action & ~CPU_TASKS_FROZEN) {
6292 sched_domains_numa_masks_set(cpu);
6296 sched_domains_numa_masks_clear(cpu);
6306 static inline void sched_init_numa(void)
6310 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6311 unsigned long action,
6316 #endif /* CONFIG_NUMA */
6318 static int __sdt_alloc(const struct cpumask *cpu_map)
6320 struct sched_domain_topology_level *tl;
6323 for_each_sd_topology(tl) {
6324 struct sd_data *sdd = &tl->data;
6326 sdd->sd = alloc_percpu(struct sched_domain *);
6330 sdd->sg = alloc_percpu(struct sched_group *);
6334 sdd->sgp = alloc_percpu(struct sched_group_power *);
6338 for_each_cpu(j, cpu_map) {
6339 struct sched_domain *sd;
6340 struct sched_group *sg;
6341 struct sched_group_power *sgp;
6343 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6344 GFP_KERNEL, cpu_to_node(j));
6348 *per_cpu_ptr(sdd->sd, j) = sd;
6350 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6351 GFP_KERNEL, cpu_to_node(j));
6357 *per_cpu_ptr(sdd->sg, j) = sg;
6359 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6360 GFP_KERNEL, cpu_to_node(j));
6364 *per_cpu_ptr(sdd->sgp, j) = sgp;
6371 static void __sdt_free(const struct cpumask *cpu_map)
6373 struct sched_domain_topology_level *tl;
6376 for_each_sd_topology(tl) {
6377 struct sd_data *sdd = &tl->data;
6379 for_each_cpu(j, cpu_map) {
6380 struct sched_domain *sd;
6383 sd = *per_cpu_ptr(sdd->sd, j);
6384 if (sd && (sd->flags & SD_OVERLAP))
6385 free_sched_groups(sd->groups, 0);
6386 kfree(*per_cpu_ptr(sdd->sd, j));
6390 kfree(*per_cpu_ptr(sdd->sg, j));
6392 kfree(*per_cpu_ptr(sdd->sgp, j));
6394 free_percpu(sdd->sd);
6396 free_percpu(sdd->sg);
6398 free_percpu(sdd->sgp);
6403 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6404 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6405 struct sched_domain *child, int cpu)
6407 struct sched_domain *sd = tl->init(tl, cpu);
6411 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6413 sd->level = child->level + 1;
6414 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6418 set_domain_attribute(sd, attr);
6424 * Build sched domains for a given set of cpus and attach the sched domains
6425 * to the individual cpus
6427 static int build_sched_domains(const struct cpumask *cpu_map,
6428 struct sched_domain_attr *attr)
6430 enum s_alloc alloc_state;
6431 struct sched_domain *sd;
6433 int i, ret = -ENOMEM;
6435 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6436 if (alloc_state != sa_rootdomain)
6439 /* Set up domains for cpus specified by the cpu_map. */
6440 for_each_cpu(i, cpu_map) {
6441 struct sched_domain_topology_level *tl;
6444 for_each_sd_topology(tl) {
6445 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6446 if (tl == sched_domain_topology)
6447 *per_cpu_ptr(d.sd, i) = sd;
6448 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6449 sd->flags |= SD_OVERLAP;
6450 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6455 /* Build the groups for the domains */
6456 for_each_cpu(i, cpu_map) {
6457 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6458 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6459 if (sd->flags & SD_OVERLAP) {
6460 if (build_overlap_sched_groups(sd, i))
6463 if (build_sched_groups(sd, i))
6469 /* Calculate CPU power for physical packages and nodes */
6470 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6471 if (!cpumask_test_cpu(i, cpu_map))
6474 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6475 claim_allocations(i, sd);
6476 init_sched_groups_power(i, sd);
6480 /* Attach the domains */
6482 for_each_cpu(i, cpu_map) {
6483 sd = *per_cpu_ptr(d.sd, i);
6484 cpu_attach_domain(sd, d.rd, i);
6490 __free_domain_allocs(&d, alloc_state, cpu_map);
6494 static cpumask_var_t *doms_cur; /* current sched domains */
6495 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6496 static struct sched_domain_attr *dattr_cur;
6497 /* attribues of custom domains in 'doms_cur' */
6500 * Special case: If a kmalloc of a doms_cur partition (array of
6501 * cpumask) fails, then fallback to a single sched domain,
6502 * as determined by the single cpumask fallback_doms.
6504 static cpumask_var_t fallback_doms;
6507 * arch_update_cpu_topology lets virtualized architectures update the
6508 * cpu core maps. It is supposed to return 1 if the topology changed
6509 * or 0 if it stayed the same.
6511 int __weak arch_update_cpu_topology(void)
6516 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6519 cpumask_var_t *doms;
6521 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6524 for (i = 0; i < ndoms; i++) {
6525 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6526 free_sched_domains(doms, i);
6533 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6536 for (i = 0; i < ndoms; i++)
6537 free_cpumask_var(doms[i]);
6542 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6543 * For now this just excludes isolated cpus, but could be used to
6544 * exclude other special cases in the future.
6546 static int init_sched_domains(const struct cpumask *cpu_map)
6550 arch_update_cpu_topology();
6552 doms_cur = alloc_sched_domains(ndoms_cur);
6554 doms_cur = &fallback_doms;
6555 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6556 err = build_sched_domains(doms_cur[0], NULL);
6557 register_sched_domain_sysctl();
6563 * Detach sched domains from a group of cpus specified in cpu_map
6564 * These cpus will now be attached to the NULL domain
6566 static void detach_destroy_domains(const struct cpumask *cpu_map)
6571 for_each_cpu(i, cpu_map)
6572 cpu_attach_domain(NULL, &def_root_domain, i);
6576 /* handle null as "default" */
6577 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6578 struct sched_domain_attr *new, int idx_new)
6580 struct sched_domain_attr tmp;
6587 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6588 new ? (new + idx_new) : &tmp,
6589 sizeof(struct sched_domain_attr));
6593 * Partition sched domains as specified by the 'ndoms_new'
6594 * cpumasks in the array doms_new[] of cpumasks. This compares
6595 * doms_new[] to the current sched domain partitioning, doms_cur[].
6596 * It destroys each deleted domain and builds each new domain.
6598 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6599 * The masks don't intersect (don't overlap.) We should setup one
6600 * sched domain for each mask. CPUs not in any of the cpumasks will
6601 * not be load balanced. If the same cpumask appears both in the
6602 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6605 * The passed in 'doms_new' should be allocated using
6606 * alloc_sched_domains. This routine takes ownership of it and will
6607 * free_sched_domains it when done with it. If the caller failed the
6608 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6609 * and partition_sched_domains() will fallback to the single partition
6610 * 'fallback_doms', it also forces the domains to be rebuilt.
6612 * If doms_new == NULL it will be replaced with cpu_online_mask.
6613 * ndoms_new == 0 is a special case for destroying existing domains,
6614 * and it will not create the default domain.
6616 * Call with hotplug lock held
6618 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6619 struct sched_domain_attr *dattr_new)
6624 mutex_lock(&sched_domains_mutex);
6626 /* always unregister in case we don't destroy any domains */
6627 unregister_sched_domain_sysctl();
6629 /* Let architecture update cpu core mappings. */
6630 new_topology = arch_update_cpu_topology();
6632 n = doms_new ? ndoms_new : 0;
6634 /* Destroy deleted domains */
6635 for (i = 0; i < ndoms_cur; i++) {
6636 for (j = 0; j < n && !new_topology; j++) {
6637 if (cpumask_equal(doms_cur[i], doms_new[j])
6638 && dattrs_equal(dattr_cur, i, dattr_new, j))
6641 /* no match - a current sched domain not in new doms_new[] */
6642 detach_destroy_domains(doms_cur[i]);
6648 if (doms_new == NULL) {
6650 doms_new = &fallback_doms;
6651 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6652 WARN_ON_ONCE(dattr_new);
6655 /* Build new domains */
6656 for (i = 0; i < ndoms_new; i++) {
6657 for (j = 0; j < n && !new_topology; j++) {
6658 if (cpumask_equal(doms_new[i], doms_cur[j])
6659 && dattrs_equal(dattr_new, i, dattr_cur, j))
6662 /* no match - add a new doms_new */
6663 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6668 /* Remember the new sched domains */
6669 if (doms_cur != &fallback_doms)
6670 free_sched_domains(doms_cur, ndoms_cur);
6671 kfree(dattr_cur); /* kfree(NULL) is safe */
6672 doms_cur = doms_new;
6673 dattr_cur = dattr_new;
6674 ndoms_cur = ndoms_new;
6676 register_sched_domain_sysctl();
6678 mutex_unlock(&sched_domains_mutex);
6681 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6684 * Update cpusets according to cpu_active mask. If cpusets are
6685 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6686 * around partition_sched_domains().
6688 * If we come here as part of a suspend/resume, don't touch cpusets because we
6689 * want to restore it back to its original state upon resume anyway.
6691 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6695 case CPU_ONLINE_FROZEN:
6696 case CPU_DOWN_FAILED_FROZEN:
6699 * num_cpus_frozen tracks how many CPUs are involved in suspend
6700 * resume sequence. As long as this is not the last online
6701 * operation in the resume sequence, just build a single sched
6702 * domain, ignoring cpusets.
6705 if (likely(num_cpus_frozen)) {
6706 partition_sched_domains(1, NULL, NULL);
6711 * This is the last CPU online operation. So fall through and
6712 * restore the original sched domains by considering the
6713 * cpuset configurations.
6717 case CPU_DOWN_FAILED:
6718 cpuset_update_active_cpus(true);
6726 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6730 case CPU_DOWN_PREPARE:
6731 cpuset_update_active_cpus(false);
6733 case CPU_DOWN_PREPARE_FROZEN:
6735 partition_sched_domains(1, NULL, NULL);
6743 void __init sched_init_smp(void)
6745 cpumask_var_t non_isolated_cpus;
6747 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6748 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6753 * There's no userspace yet to cause hotplug operations; hence all the
6754 * cpu masks are stable and all blatant races in the below code cannot
6757 mutex_lock(&sched_domains_mutex);
6758 init_sched_domains(cpu_active_mask);
6759 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6760 if (cpumask_empty(non_isolated_cpus))
6761 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6762 mutex_unlock(&sched_domains_mutex);
6764 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6765 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6766 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6770 /* Move init over to a non-isolated CPU */
6771 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6773 sched_init_granularity();
6774 free_cpumask_var(non_isolated_cpus);
6776 init_sched_rt_class();
6777 init_sched_dl_class();
6780 void __init sched_init_smp(void)
6782 sched_init_granularity();
6784 #endif /* CONFIG_SMP */
6786 const_debug unsigned int sysctl_timer_migration = 1;
6788 int in_sched_functions(unsigned long addr)
6790 return in_lock_functions(addr) ||
6791 (addr >= (unsigned long)__sched_text_start
6792 && addr < (unsigned long)__sched_text_end);
6795 #ifdef CONFIG_CGROUP_SCHED
6797 * Default task group.
6798 * Every task in system belongs to this group at bootup.
6800 struct task_group root_task_group;
6801 LIST_HEAD(task_groups);
6804 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6806 void __init sched_init(void)
6809 unsigned long alloc_size = 0, ptr;
6811 #ifdef CONFIG_FAIR_GROUP_SCHED
6812 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6814 #ifdef CONFIG_RT_GROUP_SCHED
6815 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6817 #ifdef CONFIG_CPUMASK_OFFSTACK
6818 alloc_size += num_possible_cpus() * cpumask_size();
6821 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6823 #ifdef CONFIG_FAIR_GROUP_SCHED
6824 root_task_group.se = (struct sched_entity **)ptr;
6825 ptr += nr_cpu_ids * sizeof(void **);
6827 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6828 ptr += nr_cpu_ids * sizeof(void **);
6830 #endif /* CONFIG_FAIR_GROUP_SCHED */
6831 #ifdef CONFIG_RT_GROUP_SCHED
6832 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6833 ptr += nr_cpu_ids * sizeof(void **);
6835 root_task_group.rt_rq = (struct rt_rq **)ptr;
6836 ptr += nr_cpu_ids * sizeof(void **);
6838 #endif /* CONFIG_RT_GROUP_SCHED */
6839 #ifdef CONFIG_CPUMASK_OFFSTACK
6840 for_each_possible_cpu(i) {
6841 per_cpu(load_balance_mask, i) = (void *)ptr;
6842 ptr += cpumask_size();
6844 #endif /* CONFIG_CPUMASK_OFFSTACK */
6847 init_rt_bandwidth(&def_rt_bandwidth,
6848 global_rt_period(), global_rt_runtime());
6849 init_dl_bandwidth(&def_dl_bandwidth,
6850 global_rt_period(), global_rt_runtime());
6853 init_defrootdomain();
6856 #ifdef CONFIG_RT_GROUP_SCHED
6857 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6858 global_rt_period(), global_rt_runtime());
6859 #endif /* CONFIG_RT_GROUP_SCHED */
6861 #ifdef CONFIG_CGROUP_SCHED
6862 list_add(&root_task_group.list, &task_groups);
6863 INIT_LIST_HEAD(&root_task_group.children);
6864 INIT_LIST_HEAD(&root_task_group.siblings);
6865 autogroup_init(&init_task);
6867 #endif /* CONFIG_CGROUP_SCHED */
6869 for_each_possible_cpu(i) {
6873 raw_spin_lock_init(&rq->lock);
6875 rq->calc_load_active = 0;
6876 rq->calc_load_update = jiffies + LOAD_FREQ;
6877 init_cfs_rq(&rq->cfs);
6878 init_rt_rq(&rq->rt, rq);
6879 init_dl_rq(&rq->dl, rq);
6880 #ifdef CONFIG_FAIR_GROUP_SCHED
6881 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6882 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6884 * How much cpu bandwidth does root_task_group get?
6886 * In case of task-groups formed thr' the cgroup filesystem, it
6887 * gets 100% of the cpu resources in the system. This overall
6888 * system cpu resource is divided among the tasks of
6889 * root_task_group and its child task-groups in a fair manner,
6890 * based on each entity's (task or task-group's) weight
6891 * (se->load.weight).
6893 * In other words, if root_task_group has 10 tasks of weight
6894 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6895 * then A0's share of the cpu resource is:
6897 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6899 * We achieve this by letting root_task_group's tasks sit
6900 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6902 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6903 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6904 #endif /* CONFIG_FAIR_GROUP_SCHED */
6906 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6907 #ifdef CONFIG_RT_GROUP_SCHED
6908 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6911 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6912 rq->cpu_load[j] = 0;
6914 rq->last_load_update_tick = jiffies;
6919 rq->cpu_power = SCHED_POWER_SCALE;
6920 rq->post_schedule = 0;
6921 rq->active_balance = 0;
6922 rq->next_balance = jiffies;
6927 rq->avg_idle = 2*sysctl_sched_migration_cost;
6928 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6930 INIT_LIST_HEAD(&rq->cfs_tasks);
6932 rq_attach_root(rq, &def_root_domain);
6933 #ifdef CONFIG_NO_HZ_COMMON
6936 #ifdef CONFIG_NO_HZ_FULL
6937 rq->last_sched_tick = 0;
6941 atomic_set(&rq->nr_iowait, 0);
6944 set_load_weight(&init_task);
6946 #ifdef CONFIG_PREEMPT_NOTIFIERS
6947 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6951 * The boot idle thread does lazy MMU switching as well:
6953 atomic_inc(&init_mm.mm_count);
6954 enter_lazy_tlb(&init_mm, current);
6957 * Make us the idle thread. Technically, schedule() should not be
6958 * called from this thread, however somewhere below it might be,
6959 * but because we are the idle thread, we just pick up running again
6960 * when this runqueue becomes "idle".
6962 init_idle(current, smp_processor_id());
6964 calc_load_update = jiffies + LOAD_FREQ;
6967 * During early bootup we pretend to be a normal task:
6969 current->sched_class = &fair_sched_class;
6972 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6973 /* May be allocated at isolcpus cmdline parse time */
6974 if (cpu_isolated_map == NULL)
6975 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6976 idle_thread_set_boot_cpu();
6978 init_sched_fair_class();
6980 scheduler_running = 1;
6983 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6984 static inline int preempt_count_equals(int preempt_offset)
6986 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6988 return (nested == preempt_offset);
6991 void __might_sleep(const char *file, int line, int preempt_offset)
6993 static unsigned long prev_jiffy; /* ratelimiting */
6995 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6996 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6997 !is_idle_task(current)) ||
6998 system_state != SYSTEM_RUNNING || oops_in_progress)
7000 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7002 prev_jiffy = jiffies;
7005 "BUG: sleeping function called from invalid context at %s:%d\n",
7008 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7009 in_atomic(), irqs_disabled(),
7010 current->pid, current->comm);
7012 debug_show_held_locks(current);
7013 if (irqs_disabled())
7014 print_irqtrace_events(current);
7015 #ifdef CONFIG_DEBUG_PREEMPT
7016 if (!preempt_count_equals(preempt_offset)) {
7017 pr_err("Preemption disabled at:");
7018 print_ip_sym(current->preempt_disable_ip);
7024 EXPORT_SYMBOL(__might_sleep);
7027 #ifdef CONFIG_MAGIC_SYSRQ
7028 static void normalize_task(struct rq *rq, struct task_struct *p)
7030 const struct sched_class *prev_class = p->sched_class;
7031 struct sched_attr attr = {
7032 .sched_policy = SCHED_NORMAL,
7034 int old_prio = p->prio;
7039 dequeue_task(rq, p, 0);
7040 __setscheduler(rq, p, &attr);
7042 enqueue_task(rq, p, 0);
7043 resched_task(rq->curr);
7046 check_class_changed(rq, p, prev_class, old_prio);
7049 void normalize_rt_tasks(void)
7051 struct task_struct *g, *p;
7052 unsigned long flags;
7055 read_lock_irqsave(&tasklist_lock, flags);
7056 do_each_thread(g, p) {
7058 * Only normalize user tasks:
7063 p->se.exec_start = 0;
7064 #ifdef CONFIG_SCHEDSTATS
7065 p->se.statistics.wait_start = 0;
7066 p->se.statistics.sleep_start = 0;
7067 p->se.statistics.block_start = 0;
7070 if (!dl_task(p) && !rt_task(p)) {
7072 * Renice negative nice level userspace
7075 if (task_nice(p) < 0 && p->mm)
7076 set_user_nice(p, 0);
7080 raw_spin_lock(&p->pi_lock);
7081 rq = __task_rq_lock(p);
7083 normalize_task(rq, p);
7085 __task_rq_unlock(rq);
7086 raw_spin_unlock(&p->pi_lock);
7087 } while_each_thread(g, p);
7089 read_unlock_irqrestore(&tasklist_lock, flags);
7092 #endif /* CONFIG_MAGIC_SYSRQ */
7094 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7096 * These functions are only useful for the IA64 MCA handling, or kdb.
7098 * They can only be called when the whole system has been
7099 * stopped - every CPU needs to be quiescent, and no scheduling
7100 * activity can take place. Using them for anything else would
7101 * be a serious bug, and as a result, they aren't even visible
7102 * under any other configuration.
7106 * curr_task - return the current task for a given cpu.
7107 * @cpu: the processor in question.
7109 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7111 * Return: The current task for @cpu.
7113 struct task_struct *curr_task(int cpu)
7115 return cpu_curr(cpu);
7118 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7122 * set_curr_task - set the current task for a given cpu.
7123 * @cpu: the processor in question.
7124 * @p: the task pointer to set.
7126 * Description: This function must only be used when non-maskable interrupts
7127 * are serviced on a separate stack. It allows the architecture to switch the
7128 * notion of the current task on a cpu in a non-blocking manner. This function
7129 * must be called with all CPU's synchronized, and interrupts disabled, the
7130 * and caller must save the original value of the current task (see
7131 * curr_task() above) and restore that value before reenabling interrupts and
7132 * re-starting the system.
7134 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7136 void set_curr_task(int cpu, struct task_struct *p)
7143 #ifdef CONFIG_CGROUP_SCHED
7144 /* task_group_lock serializes the addition/removal of task groups */
7145 static DEFINE_SPINLOCK(task_group_lock);
7147 static void free_sched_group(struct task_group *tg)
7149 free_fair_sched_group(tg);
7150 free_rt_sched_group(tg);
7155 /* allocate runqueue etc for a new task group */
7156 struct task_group *sched_create_group(struct task_group *parent)
7158 struct task_group *tg;
7160 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7162 return ERR_PTR(-ENOMEM);
7164 if (!alloc_fair_sched_group(tg, parent))
7167 if (!alloc_rt_sched_group(tg, parent))
7173 free_sched_group(tg);
7174 return ERR_PTR(-ENOMEM);
7177 void sched_online_group(struct task_group *tg, struct task_group *parent)
7179 unsigned long flags;
7181 spin_lock_irqsave(&task_group_lock, flags);
7182 list_add_rcu(&tg->list, &task_groups);
7184 WARN_ON(!parent); /* root should already exist */
7186 tg->parent = parent;
7187 INIT_LIST_HEAD(&tg->children);
7188 list_add_rcu(&tg->siblings, &parent->children);
7189 spin_unlock_irqrestore(&task_group_lock, flags);
7192 /* rcu callback to free various structures associated with a task group */
7193 static void free_sched_group_rcu(struct rcu_head *rhp)
7195 /* now it should be safe to free those cfs_rqs */
7196 free_sched_group(container_of(rhp, struct task_group, rcu));
7199 /* Destroy runqueue etc associated with a task group */
7200 void sched_destroy_group(struct task_group *tg)
7202 /* wait for possible concurrent references to cfs_rqs complete */
7203 call_rcu(&tg->rcu, free_sched_group_rcu);
7206 void sched_offline_group(struct task_group *tg)
7208 unsigned long flags;
7211 /* end participation in shares distribution */
7212 for_each_possible_cpu(i)
7213 unregister_fair_sched_group(tg, i);
7215 spin_lock_irqsave(&task_group_lock, flags);
7216 list_del_rcu(&tg->list);
7217 list_del_rcu(&tg->siblings);
7218 spin_unlock_irqrestore(&task_group_lock, flags);
7221 /* change task's runqueue when it moves between groups.
7222 * The caller of this function should have put the task in its new group
7223 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7224 * reflect its new group.
7226 void sched_move_task(struct task_struct *tsk)
7228 struct task_group *tg;
7230 unsigned long flags;
7233 rq = task_rq_lock(tsk, &flags);
7235 running = task_current(rq, tsk);
7239 dequeue_task(rq, tsk, 0);
7240 if (unlikely(running))
7241 tsk->sched_class->put_prev_task(rq, tsk);
7243 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7244 lockdep_is_held(&tsk->sighand->siglock)),
7245 struct task_group, css);
7246 tg = autogroup_task_group(tsk, tg);
7247 tsk->sched_task_group = tg;
7249 #ifdef CONFIG_FAIR_GROUP_SCHED
7250 if (tsk->sched_class->task_move_group)
7251 tsk->sched_class->task_move_group(tsk, on_rq);
7254 set_task_rq(tsk, task_cpu(tsk));
7256 if (unlikely(running))
7257 tsk->sched_class->set_curr_task(rq);
7259 enqueue_task(rq, tsk, 0);
7261 task_rq_unlock(rq, tsk, &flags);
7263 #endif /* CONFIG_CGROUP_SCHED */
7265 #ifdef CONFIG_RT_GROUP_SCHED
7267 * Ensure that the real time constraints are schedulable.
7269 static DEFINE_MUTEX(rt_constraints_mutex);
7271 /* Must be called with tasklist_lock held */
7272 static inline int tg_has_rt_tasks(struct task_group *tg)
7274 struct task_struct *g, *p;
7276 do_each_thread(g, p) {
7277 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7279 } while_each_thread(g, p);
7284 struct rt_schedulable_data {
7285 struct task_group *tg;
7290 static int tg_rt_schedulable(struct task_group *tg, void *data)
7292 struct rt_schedulable_data *d = data;
7293 struct task_group *child;
7294 unsigned long total, sum = 0;
7295 u64 period, runtime;
7297 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7298 runtime = tg->rt_bandwidth.rt_runtime;
7301 period = d->rt_period;
7302 runtime = d->rt_runtime;
7306 * Cannot have more runtime than the period.
7308 if (runtime > period && runtime != RUNTIME_INF)
7312 * Ensure we don't starve existing RT tasks.
7314 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7317 total = to_ratio(period, runtime);
7320 * Nobody can have more than the global setting allows.
7322 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7326 * The sum of our children's runtime should not exceed our own.
7328 list_for_each_entry_rcu(child, &tg->children, siblings) {
7329 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7330 runtime = child->rt_bandwidth.rt_runtime;
7332 if (child == d->tg) {
7333 period = d->rt_period;
7334 runtime = d->rt_runtime;
7337 sum += to_ratio(period, runtime);
7346 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7350 struct rt_schedulable_data data = {
7352 .rt_period = period,
7353 .rt_runtime = runtime,
7357 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7363 static int tg_set_rt_bandwidth(struct task_group *tg,
7364 u64 rt_period, u64 rt_runtime)
7368 mutex_lock(&rt_constraints_mutex);
7369 read_lock(&tasklist_lock);
7370 err = __rt_schedulable(tg, rt_period, rt_runtime);
7374 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7375 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7376 tg->rt_bandwidth.rt_runtime = rt_runtime;
7378 for_each_possible_cpu(i) {
7379 struct rt_rq *rt_rq = tg->rt_rq[i];
7381 raw_spin_lock(&rt_rq->rt_runtime_lock);
7382 rt_rq->rt_runtime = rt_runtime;
7383 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7385 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7387 read_unlock(&tasklist_lock);
7388 mutex_unlock(&rt_constraints_mutex);
7393 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7395 u64 rt_runtime, rt_period;
7397 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7398 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7399 if (rt_runtime_us < 0)
7400 rt_runtime = RUNTIME_INF;
7402 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7405 static long sched_group_rt_runtime(struct task_group *tg)
7409 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7412 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7413 do_div(rt_runtime_us, NSEC_PER_USEC);
7414 return rt_runtime_us;
7417 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7419 u64 rt_runtime, rt_period;
7421 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7422 rt_runtime = tg->rt_bandwidth.rt_runtime;
7427 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7430 static long sched_group_rt_period(struct task_group *tg)
7434 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7435 do_div(rt_period_us, NSEC_PER_USEC);
7436 return rt_period_us;
7438 #endif /* CONFIG_RT_GROUP_SCHED */
7440 #ifdef CONFIG_RT_GROUP_SCHED
7441 static int sched_rt_global_constraints(void)
7445 mutex_lock(&rt_constraints_mutex);
7446 read_lock(&tasklist_lock);
7447 ret = __rt_schedulable(NULL, 0, 0);
7448 read_unlock(&tasklist_lock);
7449 mutex_unlock(&rt_constraints_mutex);
7454 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7456 /* Don't accept realtime tasks when there is no way for them to run */
7457 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7463 #else /* !CONFIG_RT_GROUP_SCHED */
7464 static int sched_rt_global_constraints(void)
7466 unsigned long flags;
7469 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7470 for_each_possible_cpu(i) {
7471 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7473 raw_spin_lock(&rt_rq->rt_runtime_lock);
7474 rt_rq->rt_runtime = global_rt_runtime();
7475 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7477 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7481 #endif /* CONFIG_RT_GROUP_SCHED */
7483 static int sched_dl_global_constraints(void)
7485 u64 runtime = global_rt_runtime();
7486 u64 period = global_rt_period();
7487 u64 new_bw = to_ratio(period, runtime);
7489 unsigned long flags;
7492 * Here we want to check the bandwidth not being set to some
7493 * value smaller than the currently allocated bandwidth in
7494 * any of the root_domains.
7496 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7497 * cycling on root_domains... Discussion on different/better
7498 * solutions is welcome!
7500 for_each_possible_cpu(cpu) {
7501 struct dl_bw *dl_b = dl_bw_of(cpu);
7503 raw_spin_lock_irqsave(&dl_b->lock, flags);
7504 if (new_bw < dl_b->total_bw)
7506 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7515 static void sched_dl_do_global(void)
7519 unsigned long flags;
7521 def_dl_bandwidth.dl_period = global_rt_period();
7522 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7524 if (global_rt_runtime() != RUNTIME_INF)
7525 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7528 * FIXME: As above...
7530 for_each_possible_cpu(cpu) {
7531 struct dl_bw *dl_b = dl_bw_of(cpu);
7533 raw_spin_lock_irqsave(&dl_b->lock, flags);
7535 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7539 static int sched_rt_global_validate(void)
7541 if (sysctl_sched_rt_period <= 0)
7544 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7545 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7551 static void sched_rt_do_global(void)
7553 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7554 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7557 int sched_rt_handler(struct ctl_table *table, int write,
7558 void __user *buffer, size_t *lenp,
7561 int old_period, old_runtime;
7562 static DEFINE_MUTEX(mutex);
7566 old_period = sysctl_sched_rt_period;
7567 old_runtime = sysctl_sched_rt_runtime;
7569 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7571 if (!ret && write) {
7572 ret = sched_rt_global_validate();
7576 ret = sched_rt_global_constraints();
7580 ret = sched_dl_global_constraints();
7584 sched_rt_do_global();
7585 sched_dl_do_global();
7589 sysctl_sched_rt_period = old_period;
7590 sysctl_sched_rt_runtime = old_runtime;
7592 mutex_unlock(&mutex);
7597 int sched_rr_handler(struct ctl_table *table, int write,
7598 void __user *buffer, size_t *lenp,
7602 static DEFINE_MUTEX(mutex);
7605 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7606 /* make sure that internally we keep jiffies */
7607 /* also, writing zero resets timeslice to default */
7608 if (!ret && write) {
7609 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7610 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7612 mutex_unlock(&mutex);
7616 #ifdef CONFIG_CGROUP_SCHED
7618 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7620 return css ? container_of(css, struct task_group, css) : NULL;
7623 static struct cgroup_subsys_state *
7624 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7626 struct task_group *parent = css_tg(parent_css);
7627 struct task_group *tg;
7630 /* This is early initialization for the top cgroup */
7631 return &root_task_group.css;
7634 tg = sched_create_group(parent);
7636 return ERR_PTR(-ENOMEM);
7641 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7643 struct task_group *tg = css_tg(css);
7644 struct task_group *parent = css_tg(css_parent(css));
7647 sched_online_group(tg, parent);
7651 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7653 struct task_group *tg = css_tg(css);
7655 sched_destroy_group(tg);
7658 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7660 struct task_group *tg = css_tg(css);
7662 sched_offline_group(tg);
7665 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7666 struct cgroup_taskset *tset)
7668 struct task_struct *task;
7670 cgroup_taskset_for_each(task, tset) {
7671 #ifdef CONFIG_RT_GROUP_SCHED
7672 if (!sched_rt_can_attach(css_tg(css), task))
7675 /* We don't support RT-tasks being in separate groups */
7676 if (task->sched_class != &fair_sched_class)
7683 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7684 struct cgroup_taskset *tset)
7686 struct task_struct *task;
7688 cgroup_taskset_for_each(task, tset)
7689 sched_move_task(task);
7692 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7693 struct cgroup_subsys_state *old_css,
7694 struct task_struct *task)
7697 * cgroup_exit() is called in the copy_process() failure path.
7698 * Ignore this case since the task hasn't ran yet, this avoids
7699 * trying to poke a half freed task state from generic code.
7701 if (!(task->flags & PF_EXITING))
7704 sched_move_task(task);
7707 #ifdef CONFIG_FAIR_GROUP_SCHED
7708 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7709 struct cftype *cftype, u64 shareval)
7711 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7714 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7717 struct task_group *tg = css_tg(css);
7719 return (u64) scale_load_down(tg->shares);
7722 #ifdef CONFIG_CFS_BANDWIDTH
7723 static DEFINE_MUTEX(cfs_constraints_mutex);
7725 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7726 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7728 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7730 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7732 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7733 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7735 if (tg == &root_task_group)
7739 * Ensure we have at some amount of bandwidth every period. This is
7740 * to prevent reaching a state of large arrears when throttled via
7741 * entity_tick() resulting in prolonged exit starvation.
7743 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7747 * Likewise, bound things on the otherside by preventing insane quota
7748 * periods. This also allows us to normalize in computing quota
7751 if (period > max_cfs_quota_period)
7754 mutex_lock(&cfs_constraints_mutex);
7755 ret = __cfs_schedulable(tg, period, quota);
7759 runtime_enabled = quota != RUNTIME_INF;
7760 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7762 * If we need to toggle cfs_bandwidth_used, off->on must occur
7763 * before making related changes, and on->off must occur afterwards
7765 if (runtime_enabled && !runtime_was_enabled)
7766 cfs_bandwidth_usage_inc();
7767 raw_spin_lock_irq(&cfs_b->lock);
7768 cfs_b->period = ns_to_ktime(period);
7769 cfs_b->quota = quota;
7771 __refill_cfs_bandwidth_runtime(cfs_b);
7772 /* restart the period timer (if active) to handle new period expiry */
7773 if (runtime_enabled && cfs_b->timer_active) {
7774 /* force a reprogram */
7775 cfs_b->timer_active = 0;
7776 __start_cfs_bandwidth(cfs_b);
7778 raw_spin_unlock_irq(&cfs_b->lock);
7780 for_each_possible_cpu(i) {
7781 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7782 struct rq *rq = cfs_rq->rq;
7784 raw_spin_lock_irq(&rq->lock);
7785 cfs_rq->runtime_enabled = runtime_enabled;
7786 cfs_rq->runtime_remaining = 0;
7788 if (cfs_rq->throttled)
7789 unthrottle_cfs_rq(cfs_rq);
7790 raw_spin_unlock_irq(&rq->lock);
7792 if (runtime_was_enabled && !runtime_enabled)
7793 cfs_bandwidth_usage_dec();
7795 mutex_unlock(&cfs_constraints_mutex);
7800 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7804 period = ktime_to_ns(tg->cfs_bandwidth.period);
7805 if (cfs_quota_us < 0)
7806 quota = RUNTIME_INF;
7808 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7810 return tg_set_cfs_bandwidth(tg, period, quota);
7813 long tg_get_cfs_quota(struct task_group *tg)
7817 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7820 quota_us = tg->cfs_bandwidth.quota;
7821 do_div(quota_us, NSEC_PER_USEC);
7826 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7830 period = (u64)cfs_period_us * NSEC_PER_USEC;
7831 quota = tg->cfs_bandwidth.quota;
7833 return tg_set_cfs_bandwidth(tg, period, quota);
7836 long tg_get_cfs_period(struct task_group *tg)
7840 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7841 do_div(cfs_period_us, NSEC_PER_USEC);
7843 return cfs_period_us;
7846 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7849 return tg_get_cfs_quota(css_tg(css));
7852 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7853 struct cftype *cftype, s64 cfs_quota_us)
7855 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7858 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7861 return tg_get_cfs_period(css_tg(css));
7864 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7865 struct cftype *cftype, u64 cfs_period_us)
7867 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7870 struct cfs_schedulable_data {
7871 struct task_group *tg;
7876 * normalize group quota/period to be quota/max_period
7877 * note: units are usecs
7879 static u64 normalize_cfs_quota(struct task_group *tg,
7880 struct cfs_schedulable_data *d)
7888 period = tg_get_cfs_period(tg);
7889 quota = tg_get_cfs_quota(tg);
7892 /* note: these should typically be equivalent */
7893 if (quota == RUNTIME_INF || quota == -1)
7896 return to_ratio(period, quota);
7899 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7901 struct cfs_schedulable_data *d = data;
7902 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7903 s64 quota = 0, parent_quota = -1;
7906 quota = RUNTIME_INF;
7908 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7910 quota = normalize_cfs_quota(tg, d);
7911 parent_quota = parent_b->hierarchal_quota;
7914 * ensure max(child_quota) <= parent_quota, inherit when no
7917 if (quota == RUNTIME_INF)
7918 quota = parent_quota;
7919 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7922 cfs_b->hierarchal_quota = quota;
7927 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7930 struct cfs_schedulable_data data = {
7936 if (quota != RUNTIME_INF) {
7937 do_div(data.period, NSEC_PER_USEC);
7938 do_div(data.quota, NSEC_PER_USEC);
7942 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7948 static int cpu_stats_show(struct seq_file *sf, void *v)
7950 struct task_group *tg = css_tg(seq_css(sf));
7951 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7953 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7954 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7955 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7959 #endif /* CONFIG_CFS_BANDWIDTH */
7960 #endif /* CONFIG_FAIR_GROUP_SCHED */
7962 #ifdef CONFIG_RT_GROUP_SCHED
7963 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7964 struct cftype *cft, s64 val)
7966 return sched_group_set_rt_runtime(css_tg(css), val);
7969 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7972 return sched_group_rt_runtime(css_tg(css));
7975 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7976 struct cftype *cftype, u64 rt_period_us)
7978 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7981 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7984 return sched_group_rt_period(css_tg(css));
7986 #endif /* CONFIG_RT_GROUP_SCHED */
7988 static struct cftype cpu_files[] = {
7989 #ifdef CONFIG_FAIR_GROUP_SCHED
7992 .read_u64 = cpu_shares_read_u64,
7993 .write_u64 = cpu_shares_write_u64,
7996 #ifdef CONFIG_CFS_BANDWIDTH
7998 .name = "cfs_quota_us",
7999 .read_s64 = cpu_cfs_quota_read_s64,
8000 .write_s64 = cpu_cfs_quota_write_s64,
8003 .name = "cfs_period_us",
8004 .read_u64 = cpu_cfs_period_read_u64,
8005 .write_u64 = cpu_cfs_period_write_u64,
8009 .seq_show = cpu_stats_show,
8012 #ifdef CONFIG_RT_GROUP_SCHED
8014 .name = "rt_runtime_us",
8015 .read_s64 = cpu_rt_runtime_read,
8016 .write_s64 = cpu_rt_runtime_write,
8019 .name = "rt_period_us",
8020 .read_u64 = cpu_rt_period_read_uint,
8021 .write_u64 = cpu_rt_period_write_uint,
8027 struct cgroup_subsys cpu_cgrp_subsys = {
8028 .css_alloc = cpu_cgroup_css_alloc,
8029 .css_free = cpu_cgroup_css_free,
8030 .css_online = cpu_cgroup_css_online,
8031 .css_offline = cpu_cgroup_css_offline,
8032 .can_attach = cpu_cgroup_can_attach,
8033 .attach = cpu_cgroup_attach,
8034 .exit = cpu_cgroup_exit,
8035 .base_cftypes = cpu_files,
8039 #endif /* CONFIG_CGROUP_SCHED */
8041 void dump_cpu_task(int cpu)
8043 pr_info("Task dump for CPU %d:\n", cpu);
8044 sched_show_task(cpu_curr(cpu));