4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
143 update_rq_clock_task(rq, delta);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug unsigned int sysctl_sched_features =
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file *m, void *v)
173 for (i = 0; i < __SCHED_FEAT_NR; i++) {
174 if (!(sysctl_sched_features & (1UL << i)))
176 seq_printf(m, "%s ", sched_feat_names[i]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
192 #include "features.h"
197 static void sched_feat_disable(int i)
199 if (static_key_enabled(&sched_feat_keys[i]))
200 static_key_slow_dec(&sched_feat_keys[i]);
203 static void sched_feat_enable(int i)
205 if (!static_key_enabled(&sched_feat_keys[i]))
206 static_key_slow_inc(&sched_feat_keys[i]);
209 static void sched_feat_disable(int i) { };
210 static void sched_feat_enable(int i) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp)
218 if (strncmp(cmp, "NO_", 3) == 0) {
223 for (i = 0; i < __SCHED_FEAT_NR; i++) {
224 if (strcmp(cmp, sched_feat_names[i]) == 0) {
226 sysctl_sched_features &= ~(1UL << i);
227 sched_feat_disable(i);
229 sysctl_sched_features |= (1UL << i);
230 sched_feat_enable(i);
240 sched_feat_write(struct file *filp, const char __user *ubuf,
241 size_t cnt, loff_t *ppos)
250 if (copy_from_user(&buf, ubuf, cnt))
256 i = sched_feat_set(cmp);
257 if (i == __SCHED_FEAT_NR)
265 static int sched_feat_open(struct inode *inode, struct file *filp)
267 return single_open(filp, sched_feat_show, NULL);
270 static const struct file_operations sched_feat_fops = {
271 .open = sched_feat_open,
272 .write = sched_feat_write,
275 .release = single_release,
278 static __init int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL, NULL,
285 late_initcall(sched_init_debug);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug unsigned int sysctl_sched_nr_migrate = 32;
295 * period over which we average the RT time consumption, measured
300 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period = 1000000;
308 __read_mostly int scheduler_running;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime = 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq *__task_rq_lock(struct task_struct *p)
324 lockdep_assert_held(&p->pi_lock);
328 raw_spin_lock(&rq->lock);
329 if (likely(rq == task_rq(p)))
331 raw_spin_unlock(&rq->lock);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
339 __acquires(p->pi_lock)
345 raw_spin_lock_irqsave(&p->pi_lock, *flags);
347 raw_spin_lock(&rq->lock);
348 if (likely(rq == task_rq(p)))
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 static void __task_rq_unlock(struct rq *rq)
358 raw_spin_unlock(&rq->lock);
362 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
364 __releases(p->pi_lock)
366 raw_spin_unlock(&rq->lock);
367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq *this_rq_lock(void)
380 raw_spin_lock(&rq->lock);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq *rq)
392 if (hrtimer_active(&rq->hrtick_timer))
393 hrtimer_cancel(&rq->hrtick_timer);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart hrtick(struct hrtimer *timer)
402 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
406 raw_spin_lock(&rq->lock);
408 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
409 raw_spin_unlock(&rq->lock);
411 return HRTIMER_NORESTART;
416 static int __hrtick_restart(struct rq *rq)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = hrtimer_get_softexpires(timer);
421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg)
431 raw_spin_lock(&rq->lock);
432 __hrtick_restart(rq);
433 rq->hrtick_csd_pending = 0;
434 raw_spin_unlock(&rq->lock);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq *rq, u64 delay)
444 struct hrtimer *timer = &rq->hrtick_timer;
445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
447 hrtimer_set_expires(timer, time);
449 if (rq == this_rq()) {
450 __hrtick_restart(rq);
451 } else if (!rq->hrtick_csd_pending) {
452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
453 rq->hrtick_csd_pending = 1;
458 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
460 int cpu = (int)(long)hcpu;
463 case CPU_UP_CANCELED:
464 case CPU_UP_CANCELED_FROZEN:
465 case CPU_DOWN_PREPARE:
466 case CPU_DOWN_PREPARE_FROZEN:
468 case CPU_DEAD_FROZEN:
469 hrtick_clear(cpu_rq(cpu));
476 static __init void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq *rq, u64 delay)
488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
489 HRTIMER_MODE_REL_PINNED, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq *rq)
500 rq->hrtick_csd_pending = 0;
502 rq->hrtick_csd.flags = 0;
503 rq->hrtick_csd.func = __hrtick_start;
504 rq->hrtick_csd.info = rq;
507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
508 rq->hrtick_timer.function = hrtick;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq *rq)
515 static inline void init_rq_hrtick(struct rq *rq)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * cmpxchg based fetch_or, macro so it works for different integer types
527 #define fetch_or(ptr, val) \
528 ({ typeof(*(ptr)) __old, __val = *(ptr); \
530 __old = cmpxchg((ptr), __val, __val | (val)); \
531 if (__old == __val) \
538 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
541 * this avoids any races wrt polling state changes and thereby avoids
544 static bool set_nr_and_not_polling(struct task_struct *p)
546 struct thread_info *ti = task_thread_info(p);
547 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
551 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
553 * If this returns true, then the idle task promises to call
554 * sched_ttwu_pending() and reschedule soon.
556 static bool set_nr_if_polling(struct task_struct *p)
558 struct thread_info *ti = task_thread_info(p);
559 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
562 if (!(val & _TIF_POLLING_NRFLAG))
564 if (val & _TIF_NEED_RESCHED)
566 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
575 static bool set_nr_and_not_polling(struct task_struct *p)
577 set_tsk_need_resched(p);
582 static bool set_nr_if_polling(struct task_struct *p)
590 * resched_task - mark a task 'to be rescheduled now'.
592 * On UP this means the setting of the need_resched flag, on SMP it
593 * might also involve a cross-CPU call to trigger the scheduler on
596 void resched_task(struct task_struct *p)
600 lockdep_assert_held(&task_rq(p)->lock);
602 if (test_tsk_need_resched(p))
607 if (cpu == smp_processor_id()) {
608 set_tsk_need_resched(p);
609 set_preempt_need_resched();
613 if (set_nr_and_not_polling(p))
614 smp_send_reschedule(cpu);
616 trace_sched_wake_idle_without_ipi(cpu);
619 void resched_cpu(int cpu)
621 struct rq *rq = cpu_rq(cpu);
624 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
626 resched_task(cpu_curr(cpu));
627 raw_spin_unlock_irqrestore(&rq->lock, flags);
631 #ifdef CONFIG_NO_HZ_COMMON
633 * In the semi idle case, use the nearest busy cpu for migrating timers
634 * from an idle cpu. This is good for power-savings.
636 * We don't do similar optimization for completely idle system, as
637 * selecting an idle cpu will add more delays to the timers than intended
638 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
640 int get_nohz_timer_target(int pinned)
642 int cpu = smp_processor_id();
644 struct sched_domain *sd;
646 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
650 for_each_domain(cpu, sd) {
651 for_each_cpu(i, sched_domain_span(sd)) {
663 * When add_timer_on() enqueues a timer into the timer wheel of an
664 * idle CPU then this timer might expire before the next timer event
665 * which is scheduled to wake up that CPU. In case of a completely
666 * idle system the next event might even be infinite time into the
667 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
668 * leaves the inner idle loop so the newly added timer is taken into
669 * account when the CPU goes back to idle and evaluates the timer
670 * wheel for the next timer event.
672 static void wake_up_idle_cpu(int cpu)
674 struct rq *rq = cpu_rq(cpu);
676 if (cpu == smp_processor_id())
679 if (set_nr_and_not_polling(rq->idle))
680 smp_send_reschedule(cpu);
682 trace_sched_wake_idle_without_ipi(cpu);
685 static bool wake_up_full_nohz_cpu(int cpu)
687 if (tick_nohz_full_cpu(cpu)) {
688 if (cpu != smp_processor_id() ||
689 tick_nohz_tick_stopped())
690 smp_send_reschedule(cpu);
697 void wake_up_nohz_cpu(int cpu)
699 if (!wake_up_full_nohz_cpu(cpu))
700 wake_up_idle_cpu(cpu);
703 static inline bool got_nohz_idle_kick(void)
705 int cpu = smp_processor_id();
707 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
710 if (idle_cpu(cpu) && !need_resched())
714 * We can't run Idle Load Balance on this CPU for this time so we
715 * cancel it and clear NOHZ_BALANCE_KICK
717 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
721 #else /* CONFIG_NO_HZ_COMMON */
723 static inline bool got_nohz_idle_kick(void)
728 #endif /* CONFIG_NO_HZ_COMMON */
730 #ifdef CONFIG_NO_HZ_FULL
731 bool sched_can_stop_tick(void)
737 /* Make sure rq->nr_running update is visible after the IPI */
740 /* More than one running task need preemption */
741 if (rq->nr_running > 1)
746 #endif /* CONFIG_NO_HZ_FULL */
748 void sched_avg_update(struct rq *rq)
750 s64 period = sched_avg_period();
752 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
754 * Inline assembly required to prevent the compiler
755 * optimising this loop into a divmod call.
756 * See __iter_div_u64_rem() for another example of this.
758 asm("" : "+rm" (rq->age_stamp));
759 rq->age_stamp += period;
764 #endif /* CONFIG_SMP */
766 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
767 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
769 * Iterate task_group tree rooted at *from, calling @down when first entering a
770 * node and @up when leaving it for the final time.
772 * Caller must hold rcu_lock or sufficient equivalent.
774 int walk_tg_tree_from(struct task_group *from,
775 tg_visitor down, tg_visitor up, void *data)
777 struct task_group *parent, *child;
783 ret = (*down)(parent, data);
786 list_for_each_entry_rcu(child, &parent->children, siblings) {
793 ret = (*up)(parent, data);
794 if (ret || parent == from)
798 parent = parent->parent;
805 int tg_nop(struct task_group *tg, void *data)
811 static void set_load_weight(struct task_struct *p)
813 int prio = p->static_prio - MAX_RT_PRIO;
814 struct load_weight *load = &p->se.load;
817 * SCHED_IDLE tasks get minimal weight:
819 if (p->policy == SCHED_IDLE) {
820 load->weight = scale_load(WEIGHT_IDLEPRIO);
821 load->inv_weight = WMULT_IDLEPRIO;
825 load->weight = scale_load(prio_to_weight[prio]);
826 load->inv_weight = prio_to_wmult[prio];
829 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
832 sched_info_queued(rq, p);
833 p->sched_class->enqueue_task(rq, p, flags);
836 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
839 sched_info_dequeued(rq, p);
840 p->sched_class->dequeue_task(rq, p, flags);
843 void activate_task(struct rq *rq, struct task_struct *p, int flags)
845 if (task_contributes_to_load(p))
846 rq->nr_uninterruptible--;
848 enqueue_task(rq, p, flags);
851 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
853 if (task_contributes_to_load(p))
854 rq->nr_uninterruptible++;
856 dequeue_task(rq, p, flags);
859 static void update_rq_clock_task(struct rq *rq, s64 delta)
862 * In theory, the compile should just see 0 here, and optimize out the call
863 * to sched_rt_avg_update. But I don't trust it...
865 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
866 s64 steal = 0, irq_delta = 0;
868 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
869 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
872 * Since irq_time is only updated on {soft,}irq_exit, we might run into
873 * this case when a previous update_rq_clock() happened inside a
876 * When this happens, we stop ->clock_task and only update the
877 * prev_irq_time stamp to account for the part that fit, so that a next
878 * update will consume the rest. This ensures ->clock_task is
881 * It does however cause some slight miss-attribution of {soft,}irq
882 * time, a more accurate solution would be to update the irq_time using
883 * the current rq->clock timestamp, except that would require using
886 if (irq_delta > delta)
889 rq->prev_irq_time += irq_delta;
892 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
893 if (static_key_false((¶virt_steal_rq_enabled))) {
894 steal = paravirt_steal_clock(cpu_of(rq));
895 steal -= rq->prev_steal_time_rq;
897 if (unlikely(steal > delta))
900 rq->prev_steal_time_rq += steal;
905 rq->clock_task += delta;
907 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
908 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
909 sched_rt_avg_update(rq, irq_delta + steal);
913 void sched_set_stop_task(int cpu, struct task_struct *stop)
915 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
916 struct task_struct *old_stop = cpu_rq(cpu)->stop;
920 * Make it appear like a SCHED_FIFO task, its something
921 * userspace knows about and won't get confused about.
923 * Also, it will make PI more or less work without too
924 * much confusion -- but then, stop work should not
925 * rely on PI working anyway.
927 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
929 stop->sched_class = &stop_sched_class;
932 cpu_rq(cpu)->stop = stop;
936 * Reset it back to a normal scheduling class so that
937 * it can die in pieces.
939 old_stop->sched_class = &rt_sched_class;
944 * __normal_prio - return the priority that is based on the static prio
946 static inline int __normal_prio(struct task_struct *p)
948 return p->static_prio;
952 * Calculate the expected normal priority: i.e. priority
953 * without taking RT-inheritance into account. Might be
954 * boosted by interactivity modifiers. Changes upon fork,
955 * setprio syscalls, and whenever the interactivity
956 * estimator recalculates.
958 static inline int normal_prio(struct task_struct *p)
962 if (task_has_dl_policy(p))
963 prio = MAX_DL_PRIO-1;
964 else if (task_has_rt_policy(p))
965 prio = MAX_RT_PRIO-1 - p->rt_priority;
967 prio = __normal_prio(p);
972 * Calculate the current priority, i.e. the priority
973 * taken into account by the scheduler. This value might
974 * be boosted by RT tasks, or might be boosted by
975 * interactivity modifiers. Will be RT if the task got
976 * RT-boosted. If not then it returns p->normal_prio.
978 static int effective_prio(struct task_struct *p)
980 p->normal_prio = normal_prio(p);
982 * If we are RT tasks or we were boosted to RT priority,
983 * keep the priority unchanged. Otherwise, update priority
984 * to the normal priority:
986 if (!rt_prio(p->prio))
987 return p->normal_prio;
992 * task_curr - is this task currently executing on a CPU?
993 * @p: the task in question.
995 * Return: 1 if the task is currently executing. 0 otherwise.
997 inline int task_curr(const struct task_struct *p)
999 return cpu_curr(task_cpu(p)) == p;
1002 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1003 const struct sched_class *prev_class,
1006 if (prev_class != p->sched_class) {
1007 if (prev_class->switched_from)
1008 prev_class->switched_from(rq, p);
1009 p->sched_class->switched_to(rq, p);
1010 } else if (oldprio != p->prio || dl_task(p))
1011 p->sched_class->prio_changed(rq, p, oldprio);
1014 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1016 const struct sched_class *class;
1018 if (p->sched_class == rq->curr->sched_class) {
1019 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1021 for_each_class(class) {
1022 if (class == rq->curr->sched_class)
1024 if (class == p->sched_class) {
1025 resched_task(rq->curr);
1032 * A queue event has occurred, and we're going to schedule. In
1033 * this case, we can save a useless back to back clock update.
1035 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1036 rq->skip_clock_update = 1;
1040 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1042 #ifdef CONFIG_SCHED_DEBUG
1044 * We should never call set_task_cpu() on a blocked task,
1045 * ttwu() will sort out the placement.
1047 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1048 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1050 #ifdef CONFIG_LOCKDEP
1052 * The caller should hold either p->pi_lock or rq->lock, when changing
1053 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1055 * sched_move_task() holds both and thus holding either pins the cgroup,
1058 * Furthermore, all task_rq users should acquire both locks, see
1061 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1062 lockdep_is_held(&task_rq(p)->lock)));
1066 trace_sched_migrate_task(p, new_cpu);
1068 if (task_cpu(p) != new_cpu) {
1069 if (p->sched_class->migrate_task_rq)
1070 p->sched_class->migrate_task_rq(p, new_cpu);
1071 p->se.nr_migrations++;
1072 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1075 __set_task_cpu(p, new_cpu);
1078 static void __migrate_swap_task(struct task_struct *p, int cpu)
1081 struct rq *src_rq, *dst_rq;
1083 src_rq = task_rq(p);
1084 dst_rq = cpu_rq(cpu);
1086 deactivate_task(src_rq, p, 0);
1087 set_task_cpu(p, cpu);
1088 activate_task(dst_rq, p, 0);
1089 check_preempt_curr(dst_rq, p, 0);
1092 * Task isn't running anymore; make it appear like we migrated
1093 * it before it went to sleep. This means on wakeup we make the
1094 * previous cpu our targer instead of where it really is.
1100 struct migration_swap_arg {
1101 struct task_struct *src_task, *dst_task;
1102 int src_cpu, dst_cpu;
1105 static int migrate_swap_stop(void *data)
1107 struct migration_swap_arg *arg = data;
1108 struct rq *src_rq, *dst_rq;
1111 src_rq = cpu_rq(arg->src_cpu);
1112 dst_rq = cpu_rq(arg->dst_cpu);
1114 double_raw_lock(&arg->src_task->pi_lock,
1115 &arg->dst_task->pi_lock);
1116 double_rq_lock(src_rq, dst_rq);
1117 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1120 if (task_cpu(arg->src_task) != arg->src_cpu)
1123 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1126 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1129 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1130 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1135 double_rq_unlock(src_rq, dst_rq);
1136 raw_spin_unlock(&arg->dst_task->pi_lock);
1137 raw_spin_unlock(&arg->src_task->pi_lock);
1143 * Cross migrate two tasks
1145 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1147 struct migration_swap_arg arg;
1150 arg = (struct migration_swap_arg){
1152 .src_cpu = task_cpu(cur),
1154 .dst_cpu = task_cpu(p),
1157 if (arg.src_cpu == arg.dst_cpu)
1161 * These three tests are all lockless; this is OK since all of them
1162 * will be re-checked with proper locks held further down the line.
1164 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1167 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1170 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1173 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1174 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1180 struct migration_arg {
1181 struct task_struct *task;
1185 static int migration_cpu_stop(void *data);
1188 * wait_task_inactive - wait for a thread to unschedule.
1190 * If @match_state is nonzero, it's the @p->state value just checked and
1191 * not expected to change. If it changes, i.e. @p might have woken up,
1192 * then return zero. When we succeed in waiting for @p to be off its CPU,
1193 * we return a positive number (its total switch count). If a second call
1194 * a short while later returns the same number, the caller can be sure that
1195 * @p has remained unscheduled the whole time.
1197 * The caller must ensure that the task *will* unschedule sometime soon,
1198 * else this function might spin for a *long* time. This function can't
1199 * be called with interrupts off, or it may introduce deadlock with
1200 * smp_call_function() if an IPI is sent by the same process we are
1201 * waiting to become inactive.
1203 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1205 unsigned long flags;
1212 * We do the initial early heuristics without holding
1213 * any task-queue locks at all. We'll only try to get
1214 * the runqueue lock when things look like they will
1220 * If the task is actively running on another CPU
1221 * still, just relax and busy-wait without holding
1224 * NOTE! Since we don't hold any locks, it's not
1225 * even sure that "rq" stays as the right runqueue!
1226 * But we don't care, since "task_running()" will
1227 * return false if the runqueue has changed and p
1228 * is actually now running somewhere else!
1230 while (task_running(rq, p)) {
1231 if (match_state && unlikely(p->state != match_state))
1237 * Ok, time to look more closely! We need the rq
1238 * lock now, to be *sure*. If we're wrong, we'll
1239 * just go back and repeat.
1241 rq = task_rq_lock(p, &flags);
1242 trace_sched_wait_task(p);
1243 running = task_running(rq, p);
1246 if (!match_state || p->state == match_state)
1247 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1248 task_rq_unlock(rq, p, &flags);
1251 * If it changed from the expected state, bail out now.
1253 if (unlikely(!ncsw))
1257 * Was it really running after all now that we
1258 * checked with the proper locks actually held?
1260 * Oops. Go back and try again..
1262 if (unlikely(running)) {
1268 * It's not enough that it's not actively running,
1269 * it must be off the runqueue _entirely_, and not
1272 * So if it was still runnable (but just not actively
1273 * running right now), it's preempted, and we should
1274 * yield - it could be a while.
1276 if (unlikely(on_rq)) {
1277 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1279 set_current_state(TASK_UNINTERRUPTIBLE);
1280 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1285 * Ahh, all good. It wasn't running, and it wasn't
1286 * runnable, which means that it will never become
1287 * running in the future either. We're all done!
1296 * kick_process - kick a running thread to enter/exit the kernel
1297 * @p: the to-be-kicked thread
1299 * Cause a process which is running on another CPU to enter
1300 * kernel-mode, without any delay. (to get signals handled.)
1302 * NOTE: this function doesn't have to take the runqueue lock,
1303 * because all it wants to ensure is that the remote task enters
1304 * the kernel. If the IPI races and the task has been migrated
1305 * to another CPU then no harm is done and the purpose has been
1308 void kick_process(struct task_struct *p)
1314 if ((cpu != smp_processor_id()) && task_curr(p))
1315 smp_send_reschedule(cpu);
1318 EXPORT_SYMBOL_GPL(kick_process);
1319 #endif /* CONFIG_SMP */
1323 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1325 static int select_fallback_rq(int cpu, struct task_struct *p)
1327 int nid = cpu_to_node(cpu);
1328 const struct cpumask *nodemask = NULL;
1329 enum { cpuset, possible, fail } state = cpuset;
1333 * If the node that the cpu is on has been offlined, cpu_to_node()
1334 * will return -1. There is no cpu on the node, and we should
1335 * select the cpu on the other node.
1338 nodemask = cpumask_of_node(nid);
1340 /* Look for allowed, online CPU in same node. */
1341 for_each_cpu(dest_cpu, nodemask) {
1342 if (!cpu_online(dest_cpu))
1344 if (!cpu_active(dest_cpu))
1346 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1352 /* Any allowed, online CPU? */
1353 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1354 if (!cpu_online(dest_cpu))
1356 if (!cpu_active(dest_cpu))
1363 /* No more Mr. Nice Guy. */
1364 cpuset_cpus_allowed_fallback(p);
1369 do_set_cpus_allowed(p, cpu_possible_mask);
1380 if (state != cpuset) {
1382 * Don't tell them about moving exiting tasks or
1383 * kernel threads (both mm NULL), since they never
1386 if (p->mm && printk_ratelimit()) {
1387 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1388 task_pid_nr(p), p->comm, cpu);
1396 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1399 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1401 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1404 * In order not to call set_task_cpu() on a blocking task we need
1405 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1408 * Since this is common to all placement strategies, this lives here.
1410 * [ this allows ->select_task() to simply return task_cpu(p) and
1411 * not worry about this generic constraint ]
1413 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1415 cpu = select_fallback_rq(task_cpu(p), p);
1420 static void update_avg(u64 *avg, u64 sample)
1422 s64 diff = sample - *avg;
1428 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1430 #ifdef CONFIG_SCHEDSTATS
1431 struct rq *rq = this_rq();
1434 int this_cpu = smp_processor_id();
1436 if (cpu == this_cpu) {
1437 schedstat_inc(rq, ttwu_local);
1438 schedstat_inc(p, se.statistics.nr_wakeups_local);
1440 struct sched_domain *sd;
1442 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1444 for_each_domain(this_cpu, sd) {
1445 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1446 schedstat_inc(sd, ttwu_wake_remote);
1453 if (wake_flags & WF_MIGRATED)
1454 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1456 #endif /* CONFIG_SMP */
1458 schedstat_inc(rq, ttwu_count);
1459 schedstat_inc(p, se.statistics.nr_wakeups);
1461 if (wake_flags & WF_SYNC)
1462 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1464 #endif /* CONFIG_SCHEDSTATS */
1467 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1469 activate_task(rq, p, en_flags);
1472 /* if a worker is waking up, notify workqueue */
1473 if (p->flags & PF_WQ_WORKER)
1474 wq_worker_waking_up(p, cpu_of(rq));
1478 * Mark the task runnable and perform wakeup-preemption.
1481 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1483 check_preempt_curr(rq, p, wake_flags);
1484 trace_sched_wakeup(p, true);
1486 p->state = TASK_RUNNING;
1488 if (p->sched_class->task_woken)
1489 p->sched_class->task_woken(rq, p);
1491 if (rq->idle_stamp) {
1492 u64 delta = rq_clock(rq) - rq->idle_stamp;
1493 u64 max = 2*rq->max_idle_balance_cost;
1495 update_avg(&rq->avg_idle, delta);
1497 if (rq->avg_idle > max)
1506 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1509 if (p->sched_contributes_to_load)
1510 rq->nr_uninterruptible--;
1513 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1514 ttwu_do_wakeup(rq, p, wake_flags);
1518 * Called in case the task @p isn't fully descheduled from its runqueue,
1519 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1520 * since all we need to do is flip p->state to TASK_RUNNING, since
1521 * the task is still ->on_rq.
1523 static int ttwu_remote(struct task_struct *p, int wake_flags)
1528 rq = __task_rq_lock(p);
1530 /* check_preempt_curr() may use rq clock */
1531 update_rq_clock(rq);
1532 ttwu_do_wakeup(rq, p, wake_flags);
1535 __task_rq_unlock(rq);
1541 void sched_ttwu_pending(void)
1543 struct rq *rq = this_rq();
1544 struct llist_node *llist = llist_del_all(&rq->wake_list);
1545 struct task_struct *p;
1546 unsigned long flags;
1551 raw_spin_lock_irqsave(&rq->lock, flags);
1554 p = llist_entry(llist, struct task_struct, wake_entry);
1555 llist = llist_next(llist);
1556 ttwu_do_activate(rq, p, 0);
1559 raw_spin_unlock_irqrestore(&rq->lock, flags);
1562 void scheduler_ipi(void)
1565 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1566 * TIF_NEED_RESCHED remotely (for the first time) will also send
1569 preempt_fold_need_resched();
1571 if (llist_empty(&this_rq()->wake_list)
1572 && !tick_nohz_full_cpu(smp_processor_id())
1573 && !got_nohz_idle_kick())
1577 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1578 * traditionally all their work was done from the interrupt return
1579 * path. Now that we actually do some work, we need to make sure
1582 * Some archs already do call them, luckily irq_enter/exit nest
1585 * Arguably we should visit all archs and update all handlers,
1586 * however a fair share of IPIs are still resched only so this would
1587 * somewhat pessimize the simple resched case.
1590 tick_nohz_full_check();
1591 sched_ttwu_pending();
1594 * Check if someone kicked us for doing the nohz idle load balance.
1596 if (unlikely(got_nohz_idle_kick())) {
1597 this_rq()->idle_balance = 1;
1598 raise_softirq_irqoff(SCHED_SOFTIRQ);
1603 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1605 struct rq *rq = cpu_rq(cpu);
1607 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1608 if (!set_nr_if_polling(rq->idle))
1609 smp_send_reschedule(cpu);
1611 trace_sched_wake_idle_without_ipi(cpu);
1615 bool cpus_share_cache(int this_cpu, int that_cpu)
1617 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1619 #endif /* CONFIG_SMP */
1621 static void ttwu_queue(struct task_struct *p, int cpu)
1623 struct rq *rq = cpu_rq(cpu);
1625 #if defined(CONFIG_SMP)
1626 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1627 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1628 ttwu_queue_remote(p, cpu);
1633 raw_spin_lock(&rq->lock);
1634 ttwu_do_activate(rq, p, 0);
1635 raw_spin_unlock(&rq->lock);
1639 * try_to_wake_up - wake up a thread
1640 * @p: the thread to be awakened
1641 * @state: the mask of task states that can be woken
1642 * @wake_flags: wake modifier flags (WF_*)
1644 * Put it on the run-queue if it's not already there. The "current"
1645 * thread is always on the run-queue (except when the actual
1646 * re-schedule is in progress), and as such you're allowed to do
1647 * the simpler "current->state = TASK_RUNNING" to mark yourself
1648 * runnable without the overhead of this.
1650 * Return: %true if @p was woken up, %false if it was already running.
1651 * or @state didn't match @p's state.
1654 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1656 unsigned long flags;
1657 int cpu, success = 0;
1660 * If we are going to wake up a thread waiting for CONDITION we
1661 * need to ensure that CONDITION=1 done by the caller can not be
1662 * reordered with p->state check below. This pairs with mb() in
1663 * set_current_state() the waiting thread does.
1665 smp_mb__before_spinlock();
1666 raw_spin_lock_irqsave(&p->pi_lock, flags);
1667 if (!(p->state & state))
1670 success = 1; /* we're going to change ->state */
1673 if (p->on_rq && ttwu_remote(p, wake_flags))
1678 * If the owning (remote) cpu is still in the middle of schedule() with
1679 * this task as prev, wait until its done referencing the task.
1684 * Pairs with the smp_wmb() in finish_lock_switch().
1688 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1689 p->state = TASK_WAKING;
1691 if (p->sched_class->task_waking)
1692 p->sched_class->task_waking(p);
1694 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1695 if (task_cpu(p) != cpu) {
1696 wake_flags |= WF_MIGRATED;
1697 set_task_cpu(p, cpu);
1699 #endif /* CONFIG_SMP */
1703 ttwu_stat(p, cpu, wake_flags);
1705 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1711 * try_to_wake_up_local - try to wake up a local task with rq lock held
1712 * @p: the thread to be awakened
1714 * Put @p on the run-queue if it's not already there. The caller must
1715 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1718 static void try_to_wake_up_local(struct task_struct *p)
1720 struct rq *rq = task_rq(p);
1722 if (WARN_ON_ONCE(rq != this_rq()) ||
1723 WARN_ON_ONCE(p == current))
1726 lockdep_assert_held(&rq->lock);
1728 if (!raw_spin_trylock(&p->pi_lock)) {
1729 raw_spin_unlock(&rq->lock);
1730 raw_spin_lock(&p->pi_lock);
1731 raw_spin_lock(&rq->lock);
1734 if (!(p->state & TASK_NORMAL))
1738 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1740 ttwu_do_wakeup(rq, p, 0);
1741 ttwu_stat(p, smp_processor_id(), 0);
1743 raw_spin_unlock(&p->pi_lock);
1747 * wake_up_process - Wake up a specific process
1748 * @p: The process to be woken up.
1750 * Attempt to wake up the nominated process and move it to the set of runnable
1753 * Return: 1 if the process was woken up, 0 if it was already running.
1755 * It may be assumed that this function implies a write memory barrier before
1756 * changing the task state if and only if any tasks are woken up.
1758 int wake_up_process(struct task_struct *p)
1760 WARN_ON(task_is_stopped_or_traced(p));
1761 return try_to_wake_up(p, TASK_NORMAL, 0);
1763 EXPORT_SYMBOL(wake_up_process);
1765 int wake_up_state(struct task_struct *p, unsigned int state)
1767 return try_to_wake_up(p, state, 0);
1771 * Perform scheduler related setup for a newly forked process p.
1772 * p is forked by current.
1774 * __sched_fork() is basic setup used by init_idle() too:
1776 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1781 p->se.exec_start = 0;
1782 p->se.sum_exec_runtime = 0;
1783 p->se.prev_sum_exec_runtime = 0;
1784 p->se.nr_migrations = 0;
1786 INIT_LIST_HEAD(&p->se.group_node);
1788 #ifdef CONFIG_SCHEDSTATS
1789 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1792 RB_CLEAR_NODE(&p->dl.rb_node);
1793 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1794 p->dl.dl_runtime = p->dl.runtime = 0;
1795 p->dl.dl_deadline = p->dl.deadline = 0;
1796 p->dl.dl_period = 0;
1799 INIT_LIST_HEAD(&p->rt.run_list);
1801 #ifdef CONFIG_PREEMPT_NOTIFIERS
1802 INIT_HLIST_HEAD(&p->preempt_notifiers);
1805 #ifdef CONFIG_NUMA_BALANCING
1806 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1807 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1808 p->mm->numa_scan_seq = 0;
1811 if (clone_flags & CLONE_VM)
1812 p->numa_preferred_nid = current->numa_preferred_nid;
1814 p->numa_preferred_nid = -1;
1816 p->node_stamp = 0ULL;
1817 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1818 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1819 p->numa_work.next = &p->numa_work;
1820 p->numa_faults_memory = NULL;
1821 p->numa_faults_buffer_memory = NULL;
1822 p->last_task_numa_placement = 0;
1823 p->last_sum_exec_runtime = 0;
1825 INIT_LIST_HEAD(&p->numa_entry);
1826 p->numa_group = NULL;
1827 #endif /* CONFIG_NUMA_BALANCING */
1830 #ifdef CONFIG_NUMA_BALANCING
1831 #ifdef CONFIG_SCHED_DEBUG
1832 void set_numabalancing_state(bool enabled)
1835 sched_feat_set("NUMA");
1837 sched_feat_set("NO_NUMA");
1840 __read_mostly bool numabalancing_enabled;
1842 void set_numabalancing_state(bool enabled)
1844 numabalancing_enabled = enabled;
1846 #endif /* CONFIG_SCHED_DEBUG */
1848 #ifdef CONFIG_PROC_SYSCTL
1849 int sysctl_numa_balancing(struct ctl_table *table, int write,
1850 void __user *buffer, size_t *lenp, loff_t *ppos)
1854 int state = numabalancing_enabled;
1856 if (write && !capable(CAP_SYS_ADMIN))
1861 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1865 set_numabalancing_state(state);
1872 * fork()/clone()-time setup:
1874 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1876 unsigned long flags;
1877 int cpu = get_cpu();
1879 __sched_fork(clone_flags, p);
1881 * We mark the process as running here. This guarantees that
1882 * nobody will actually run it, and a signal or other external
1883 * event cannot wake it up and insert it on the runqueue either.
1885 p->state = TASK_RUNNING;
1888 * Make sure we do not leak PI boosting priority to the child.
1890 p->prio = current->normal_prio;
1893 * Revert to default priority/policy on fork if requested.
1895 if (unlikely(p->sched_reset_on_fork)) {
1896 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1897 p->policy = SCHED_NORMAL;
1898 p->static_prio = NICE_TO_PRIO(0);
1900 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1901 p->static_prio = NICE_TO_PRIO(0);
1903 p->prio = p->normal_prio = __normal_prio(p);
1907 * We don't need the reset flag anymore after the fork. It has
1908 * fulfilled its duty:
1910 p->sched_reset_on_fork = 0;
1913 if (dl_prio(p->prio)) {
1916 } else if (rt_prio(p->prio)) {
1917 p->sched_class = &rt_sched_class;
1919 p->sched_class = &fair_sched_class;
1922 if (p->sched_class->task_fork)
1923 p->sched_class->task_fork(p);
1926 * The child is not yet in the pid-hash so no cgroup attach races,
1927 * and the cgroup is pinned to this child due to cgroup_fork()
1928 * is ran before sched_fork().
1930 * Silence PROVE_RCU.
1932 raw_spin_lock_irqsave(&p->pi_lock, flags);
1933 set_task_cpu(p, cpu);
1934 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1936 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1937 if (likely(sched_info_on()))
1938 memset(&p->sched_info, 0, sizeof(p->sched_info));
1940 #if defined(CONFIG_SMP)
1943 init_task_preempt_count(p);
1945 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1946 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1953 unsigned long to_ratio(u64 period, u64 runtime)
1955 if (runtime == RUNTIME_INF)
1959 * Doing this here saves a lot of checks in all
1960 * the calling paths, and returning zero seems
1961 * safe for them anyway.
1966 return div64_u64(runtime << 20, period);
1970 inline struct dl_bw *dl_bw_of(int i)
1972 return &cpu_rq(i)->rd->dl_bw;
1975 static inline int dl_bw_cpus(int i)
1977 struct root_domain *rd = cpu_rq(i)->rd;
1980 for_each_cpu_and(i, rd->span, cpu_active_mask)
1986 inline struct dl_bw *dl_bw_of(int i)
1988 return &cpu_rq(i)->dl.dl_bw;
1991 static inline int dl_bw_cpus(int i)
1998 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2000 dl_b->total_bw -= tsk_bw;
2004 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2006 dl_b->total_bw += tsk_bw;
2010 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2012 return dl_b->bw != -1 &&
2013 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2017 * We must be sure that accepting a new task (or allowing changing the
2018 * parameters of an existing one) is consistent with the bandwidth
2019 * constraints. If yes, this function also accordingly updates the currently
2020 * allocated bandwidth to reflect the new situation.
2022 * This function is called while holding p's rq->lock.
2024 static int dl_overflow(struct task_struct *p, int policy,
2025 const struct sched_attr *attr)
2028 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2029 u64 period = attr->sched_period ?: attr->sched_deadline;
2030 u64 runtime = attr->sched_runtime;
2031 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2034 if (new_bw == p->dl.dl_bw)
2038 * Either if a task, enters, leave, or stays -deadline but changes
2039 * its parameters, we may need to update accordingly the total
2040 * allocated bandwidth of the container.
2042 raw_spin_lock(&dl_b->lock);
2043 cpus = dl_bw_cpus(task_cpu(p));
2044 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2045 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2046 __dl_add(dl_b, new_bw);
2048 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2049 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2050 __dl_clear(dl_b, p->dl.dl_bw);
2051 __dl_add(dl_b, new_bw);
2053 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2054 __dl_clear(dl_b, p->dl.dl_bw);
2057 raw_spin_unlock(&dl_b->lock);
2062 extern void init_dl_bw(struct dl_bw *dl_b);
2065 * wake_up_new_task - wake up a newly created task for the first time.
2067 * This function will do some initial scheduler statistics housekeeping
2068 * that must be done for every newly created context, then puts the task
2069 * on the runqueue and wakes it.
2071 void wake_up_new_task(struct task_struct *p)
2073 unsigned long flags;
2076 raw_spin_lock_irqsave(&p->pi_lock, flags);
2079 * Fork balancing, do it here and not earlier because:
2080 * - cpus_allowed can change in the fork path
2081 * - any previously selected cpu might disappear through hotplug
2083 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2086 /* Initialize new task's runnable average */
2087 init_task_runnable_average(p);
2088 rq = __task_rq_lock(p);
2089 activate_task(rq, p, 0);
2091 trace_sched_wakeup_new(p, true);
2092 check_preempt_curr(rq, p, WF_FORK);
2094 if (p->sched_class->task_woken)
2095 p->sched_class->task_woken(rq, p);
2097 task_rq_unlock(rq, p, &flags);
2100 #ifdef CONFIG_PREEMPT_NOTIFIERS
2103 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2104 * @notifier: notifier struct to register
2106 void preempt_notifier_register(struct preempt_notifier *notifier)
2108 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2110 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2113 * preempt_notifier_unregister - no longer interested in preemption notifications
2114 * @notifier: notifier struct to unregister
2116 * This is safe to call from within a preemption notifier.
2118 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2120 hlist_del(¬ifier->link);
2122 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2124 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2126 struct preempt_notifier *notifier;
2128 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2129 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2133 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2134 struct task_struct *next)
2136 struct preempt_notifier *notifier;
2138 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2139 notifier->ops->sched_out(notifier, next);
2142 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2144 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2149 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2150 struct task_struct *next)
2154 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2157 * prepare_task_switch - prepare to switch tasks
2158 * @rq: the runqueue preparing to switch
2159 * @prev: the current task that is being switched out
2160 * @next: the task we are going to switch to.
2162 * This is called with the rq lock held and interrupts off. It must
2163 * be paired with a subsequent finish_task_switch after the context
2166 * prepare_task_switch sets up locking and calls architecture specific
2170 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2171 struct task_struct *next)
2173 trace_sched_switch(prev, next);
2174 sched_info_switch(rq, prev, next);
2175 perf_event_task_sched_out(prev, next);
2176 fire_sched_out_preempt_notifiers(prev, next);
2177 prepare_lock_switch(rq, next);
2178 prepare_arch_switch(next);
2182 * finish_task_switch - clean up after a task-switch
2183 * @rq: runqueue associated with task-switch
2184 * @prev: the thread we just switched away from.
2186 * finish_task_switch must be called after the context switch, paired
2187 * with a prepare_task_switch call before the context switch.
2188 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2189 * and do any other architecture-specific cleanup actions.
2191 * Note that we may have delayed dropping an mm in context_switch(). If
2192 * so, we finish that here outside of the runqueue lock. (Doing it
2193 * with the lock held can cause deadlocks; see schedule() for
2196 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2197 __releases(rq->lock)
2199 struct mm_struct *mm = rq->prev_mm;
2205 * A task struct has one reference for the use as "current".
2206 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2207 * schedule one last time. The schedule call will never return, and
2208 * the scheduled task must drop that reference.
2209 * The test for TASK_DEAD must occur while the runqueue locks are
2210 * still held, otherwise prev could be scheduled on another cpu, die
2211 * there before we look at prev->state, and then the reference would
2213 * Manfred Spraul <manfred@colorfullife.com>
2215 prev_state = prev->state;
2216 vtime_task_switch(prev);
2217 finish_arch_switch(prev);
2218 perf_event_task_sched_in(prev, current);
2219 finish_lock_switch(rq, prev);
2220 finish_arch_post_lock_switch();
2222 fire_sched_in_preempt_notifiers(current);
2225 if (unlikely(prev_state == TASK_DEAD)) {
2226 if (prev->sched_class->task_dead)
2227 prev->sched_class->task_dead(prev);
2230 * Remove function-return probe instances associated with this
2231 * task and put them back on the free list.
2233 kprobe_flush_task(prev);
2234 put_task_struct(prev);
2237 tick_nohz_task_switch(current);
2242 /* rq->lock is NOT held, but preemption is disabled */
2243 static inline void post_schedule(struct rq *rq)
2245 if (rq->post_schedule) {
2246 unsigned long flags;
2248 raw_spin_lock_irqsave(&rq->lock, flags);
2249 if (rq->curr->sched_class->post_schedule)
2250 rq->curr->sched_class->post_schedule(rq);
2251 raw_spin_unlock_irqrestore(&rq->lock, flags);
2253 rq->post_schedule = 0;
2259 static inline void post_schedule(struct rq *rq)
2266 * schedule_tail - first thing a freshly forked thread must call.
2267 * @prev: the thread we just switched away from.
2269 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2270 __releases(rq->lock)
2272 struct rq *rq = this_rq();
2274 finish_task_switch(rq, prev);
2277 * FIXME: do we need to worry about rq being invalidated by the
2282 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2283 /* In this case, finish_task_switch does not reenable preemption */
2286 if (current->set_child_tid)
2287 put_user(task_pid_vnr(current), current->set_child_tid);
2291 * context_switch - switch to the new MM and the new
2292 * thread's register state.
2295 context_switch(struct rq *rq, struct task_struct *prev,
2296 struct task_struct *next)
2298 struct mm_struct *mm, *oldmm;
2300 prepare_task_switch(rq, prev, next);
2303 oldmm = prev->active_mm;
2305 * For paravirt, this is coupled with an exit in switch_to to
2306 * combine the page table reload and the switch backend into
2309 arch_start_context_switch(prev);
2312 next->active_mm = oldmm;
2313 atomic_inc(&oldmm->mm_count);
2314 enter_lazy_tlb(oldmm, next);
2316 switch_mm(oldmm, mm, next);
2319 prev->active_mm = NULL;
2320 rq->prev_mm = oldmm;
2323 * Since the runqueue lock will be released by the next
2324 * task (which is an invalid locking op but in the case
2325 * of the scheduler it's an obvious special-case), so we
2326 * do an early lockdep release here:
2328 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2329 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2332 context_tracking_task_switch(prev, next);
2333 /* Here we just switch the register state and the stack. */
2334 switch_to(prev, next, prev);
2338 * this_rq must be evaluated again because prev may have moved
2339 * CPUs since it called schedule(), thus the 'rq' on its stack
2340 * frame will be invalid.
2342 finish_task_switch(this_rq(), prev);
2346 * nr_running and nr_context_switches:
2348 * externally visible scheduler statistics: current number of runnable
2349 * threads, total number of context switches performed since bootup.
2351 unsigned long nr_running(void)
2353 unsigned long i, sum = 0;
2355 for_each_online_cpu(i)
2356 sum += cpu_rq(i)->nr_running;
2361 unsigned long long nr_context_switches(void)
2364 unsigned long long sum = 0;
2366 for_each_possible_cpu(i)
2367 sum += cpu_rq(i)->nr_switches;
2372 unsigned long nr_iowait(void)
2374 unsigned long i, sum = 0;
2376 for_each_possible_cpu(i)
2377 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2382 unsigned long nr_iowait_cpu(int cpu)
2384 struct rq *this = cpu_rq(cpu);
2385 return atomic_read(&this->nr_iowait);
2391 * sched_exec - execve() is a valuable balancing opportunity, because at
2392 * this point the task has the smallest effective memory and cache footprint.
2394 void sched_exec(void)
2396 struct task_struct *p = current;
2397 unsigned long flags;
2400 raw_spin_lock_irqsave(&p->pi_lock, flags);
2401 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2402 if (dest_cpu == smp_processor_id())
2405 if (likely(cpu_active(dest_cpu))) {
2406 struct migration_arg arg = { p, dest_cpu };
2408 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2409 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2413 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2418 DEFINE_PER_CPU(struct kernel_stat, kstat);
2419 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2421 EXPORT_PER_CPU_SYMBOL(kstat);
2422 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2425 * Return any ns on the sched_clock that have not yet been accounted in
2426 * @p in case that task is currently running.
2428 * Called with task_rq_lock() held on @rq.
2430 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2434 if (task_current(rq, p)) {
2435 update_rq_clock(rq);
2436 ns = rq_clock_task(rq) - p->se.exec_start;
2444 unsigned long long task_delta_exec(struct task_struct *p)
2446 unsigned long flags;
2450 rq = task_rq_lock(p, &flags);
2451 ns = do_task_delta_exec(p, rq);
2452 task_rq_unlock(rq, p, &flags);
2458 * Return accounted runtime for the task.
2459 * In case the task is currently running, return the runtime plus current's
2460 * pending runtime that have not been accounted yet.
2462 unsigned long long task_sched_runtime(struct task_struct *p)
2464 unsigned long flags;
2468 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2470 * 64-bit doesn't need locks to atomically read a 64bit value.
2471 * So we have a optimization chance when the task's delta_exec is 0.
2472 * Reading ->on_cpu is racy, but this is ok.
2474 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2475 * If we race with it entering cpu, unaccounted time is 0. This is
2476 * indistinguishable from the read occurring a few cycles earlier.
2479 return p->se.sum_exec_runtime;
2482 rq = task_rq_lock(p, &flags);
2483 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2484 task_rq_unlock(rq, p, &flags);
2490 * This function gets called by the timer code, with HZ frequency.
2491 * We call it with interrupts disabled.
2493 void scheduler_tick(void)
2495 int cpu = smp_processor_id();
2496 struct rq *rq = cpu_rq(cpu);
2497 struct task_struct *curr = rq->curr;
2501 raw_spin_lock(&rq->lock);
2502 update_rq_clock(rq);
2503 curr->sched_class->task_tick(rq, curr, 0);
2504 update_cpu_load_active(rq);
2505 raw_spin_unlock(&rq->lock);
2507 perf_event_task_tick();
2510 rq->idle_balance = idle_cpu(cpu);
2511 trigger_load_balance(rq);
2513 rq_last_tick_reset(rq);
2516 #ifdef CONFIG_NO_HZ_FULL
2518 * scheduler_tick_max_deferment
2520 * Keep at least one tick per second when a single
2521 * active task is running because the scheduler doesn't
2522 * yet completely support full dynticks environment.
2524 * This makes sure that uptime, CFS vruntime, load
2525 * balancing, etc... continue to move forward, even
2526 * with a very low granularity.
2528 * Return: Maximum deferment in nanoseconds.
2530 u64 scheduler_tick_max_deferment(void)
2532 struct rq *rq = this_rq();
2533 unsigned long next, now = ACCESS_ONCE(jiffies);
2535 next = rq->last_sched_tick + HZ;
2537 if (time_before_eq(next, now))
2540 return jiffies_to_nsecs(next - now);
2544 notrace unsigned long get_parent_ip(unsigned long addr)
2546 if (in_lock_functions(addr)) {
2547 addr = CALLER_ADDR2;
2548 if (in_lock_functions(addr))
2549 addr = CALLER_ADDR3;
2554 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2555 defined(CONFIG_PREEMPT_TRACER))
2557 void preempt_count_add(int val)
2559 #ifdef CONFIG_DEBUG_PREEMPT
2563 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2566 __preempt_count_add(val);
2567 #ifdef CONFIG_DEBUG_PREEMPT
2569 * Spinlock count overflowing soon?
2571 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2574 if (preempt_count() == val) {
2575 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2576 #ifdef CONFIG_DEBUG_PREEMPT
2577 current->preempt_disable_ip = ip;
2579 trace_preempt_off(CALLER_ADDR0, ip);
2582 EXPORT_SYMBOL(preempt_count_add);
2583 NOKPROBE_SYMBOL(preempt_count_add);
2585 void preempt_count_sub(int val)
2587 #ifdef CONFIG_DEBUG_PREEMPT
2591 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2594 * Is the spinlock portion underflowing?
2596 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2597 !(preempt_count() & PREEMPT_MASK)))
2601 if (preempt_count() == val)
2602 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2603 __preempt_count_sub(val);
2605 EXPORT_SYMBOL(preempt_count_sub);
2606 NOKPROBE_SYMBOL(preempt_count_sub);
2611 * Print scheduling while atomic bug:
2613 static noinline void __schedule_bug(struct task_struct *prev)
2615 if (oops_in_progress)
2618 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2619 prev->comm, prev->pid, preempt_count());
2621 debug_show_held_locks(prev);
2623 if (irqs_disabled())
2624 print_irqtrace_events(prev);
2625 #ifdef CONFIG_DEBUG_PREEMPT
2626 if (in_atomic_preempt_off()) {
2627 pr_err("Preemption disabled at:");
2628 print_ip_sym(current->preempt_disable_ip);
2633 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2637 * Various schedule()-time debugging checks and statistics:
2639 static inline void schedule_debug(struct task_struct *prev)
2642 * Test if we are atomic. Since do_exit() needs to call into
2643 * schedule() atomically, we ignore that path. Otherwise whine
2644 * if we are scheduling when we should not.
2646 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2647 __schedule_bug(prev);
2650 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2652 schedstat_inc(this_rq(), sched_count);
2656 * Pick up the highest-prio task:
2658 static inline struct task_struct *
2659 pick_next_task(struct rq *rq, struct task_struct *prev)
2661 const struct sched_class *class = &fair_sched_class;
2662 struct task_struct *p;
2665 * Optimization: we know that if all tasks are in
2666 * the fair class we can call that function directly:
2668 if (likely(prev->sched_class == class &&
2669 rq->nr_running == rq->cfs.h_nr_running)) {
2670 p = fair_sched_class.pick_next_task(rq, prev);
2671 if (unlikely(p == RETRY_TASK))
2674 /* assumes fair_sched_class->next == idle_sched_class */
2676 p = idle_sched_class.pick_next_task(rq, prev);
2682 for_each_class(class) {
2683 p = class->pick_next_task(rq, prev);
2685 if (unlikely(p == RETRY_TASK))
2691 BUG(); /* the idle class will always have a runnable task */
2695 * __schedule() is the main scheduler function.
2697 * The main means of driving the scheduler and thus entering this function are:
2699 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2701 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2702 * paths. For example, see arch/x86/entry_64.S.
2704 * To drive preemption between tasks, the scheduler sets the flag in timer
2705 * interrupt handler scheduler_tick().
2707 * 3. Wakeups don't really cause entry into schedule(). They add a
2708 * task to the run-queue and that's it.
2710 * Now, if the new task added to the run-queue preempts the current
2711 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2712 * called on the nearest possible occasion:
2714 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2716 * - in syscall or exception context, at the next outmost
2717 * preempt_enable(). (this might be as soon as the wake_up()'s
2720 * - in IRQ context, return from interrupt-handler to
2721 * preemptible context
2723 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2726 * - cond_resched() call
2727 * - explicit schedule() call
2728 * - return from syscall or exception to user-space
2729 * - return from interrupt-handler to user-space
2731 static void __sched __schedule(void)
2733 struct task_struct *prev, *next;
2734 unsigned long *switch_count;
2740 cpu = smp_processor_id();
2742 rcu_note_context_switch(cpu);
2745 schedule_debug(prev);
2747 if (sched_feat(HRTICK))
2751 * Make sure that signal_pending_state()->signal_pending() below
2752 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2753 * done by the caller to avoid the race with signal_wake_up().
2755 smp_mb__before_spinlock();
2756 raw_spin_lock_irq(&rq->lock);
2758 switch_count = &prev->nivcsw;
2759 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2760 if (unlikely(signal_pending_state(prev->state, prev))) {
2761 prev->state = TASK_RUNNING;
2763 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2767 * If a worker went to sleep, notify and ask workqueue
2768 * whether it wants to wake up a task to maintain
2771 if (prev->flags & PF_WQ_WORKER) {
2772 struct task_struct *to_wakeup;
2774 to_wakeup = wq_worker_sleeping(prev, cpu);
2776 try_to_wake_up_local(to_wakeup);
2779 switch_count = &prev->nvcsw;
2782 if (prev->on_rq || rq->skip_clock_update < 0)
2783 update_rq_clock(rq);
2785 next = pick_next_task(rq, prev);
2786 clear_tsk_need_resched(prev);
2787 clear_preempt_need_resched();
2788 rq->skip_clock_update = 0;
2790 if (likely(prev != next)) {
2795 context_switch(rq, prev, next); /* unlocks the rq */
2797 * The context switch have flipped the stack from under us
2798 * and restored the local variables which were saved when
2799 * this task called schedule() in the past. prev == current
2800 * is still correct, but it can be moved to another cpu/rq.
2802 cpu = smp_processor_id();
2805 raw_spin_unlock_irq(&rq->lock);
2809 sched_preempt_enable_no_resched();
2814 static inline void sched_submit_work(struct task_struct *tsk)
2816 if (!tsk->state || tsk_is_pi_blocked(tsk))
2819 * If we are going to sleep and we have plugged IO queued,
2820 * make sure to submit it to avoid deadlocks.
2822 if (blk_needs_flush_plug(tsk))
2823 blk_schedule_flush_plug(tsk);
2826 asmlinkage __visible void __sched schedule(void)
2828 struct task_struct *tsk = current;
2830 sched_submit_work(tsk);
2833 EXPORT_SYMBOL(schedule);
2835 #ifdef CONFIG_CONTEXT_TRACKING
2836 asmlinkage __visible void __sched schedule_user(void)
2839 * If we come here after a random call to set_need_resched(),
2840 * or we have been woken up remotely but the IPI has not yet arrived,
2841 * we haven't yet exited the RCU idle mode. Do it here manually until
2842 * we find a better solution.
2851 * schedule_preempt_disabled - called with preemption disabled
2853 * Returns with preemption disabled. Note: preempt_count must be 1
2855 void __sched schedule_preempt_disabled(void)
2857 sched_preempt_enable_no_resched();
2862 #ifdef CONFIG_PREEMPT
2864 * this is the entry point to schedule() from in-kernel preemption
2865 * off of preempt_enable. Kernel preemptions off return from interrupt
2866 * occur there and call schedule directly.
2868 asmlinkage __visible void __sched notrace preempt_schedule(void)
2871 * If there is a non-zero preempt_count or interrupts are disabled,
2872 * we do not want to preempt the current task. Just return..
2874 if (likely(!preemptible()))
2878 __preempt_count_add(PREEMPT_ACTIVE);
2880 __preempt_count_sub(PREEMPT_ACTIVE);
2883 * Check again in case we missed a preemption opportunity
2884 * between schedule and now.
2887 } while (need_resched());
2889 NOKPROBE_SYMBOL(preempt_schedule);
2890 EXPORT_SYMBOL(preempt_schedule);
2891 #endif /* CONFIG_PREEMPT */
2894 * this is the entry point to schedule() from kernel preemption
2895 * off of irq context.
2896 * Note, that this is called and return with irqs disabled. This will
2897 * protect us against recursive calling from irq.
2899 asmlinkage __visible void __sched preempt_schedule_irq(void)
2901 enum ctx_state prev_state;
2903 /* Catch callers which need to be fixed */
2904 BUG_ON(preempt_count() || !irqs_disabled());
2906 prev_state = exception_enter();
2909 __preempt_count_add(PREEMPT_ACTIVE);
2912 local_irq_disable();
2913 __preempt_count_sub(PREEMPT_ACTIVE);
2916 * Check again in case we missed a preemption opportunity
2917 * between schedule and now.
2920 } while (need_resched());
2922 exception_exit(prev_state);
2925 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2928 return try_to_wake_up(curr->private, mode, wake_flags);
2930 EXPORT_SYMBOL(default_wake_function);
2932 #ifdef CONFIG_RT_MUTEXES
2935 * rt_mutex_setprio - set the current priority of a task
2937 * @prio: prio value (kernel-internal form)
2939 * This function changes the 'effective' priority of a task. It does
2940 * not touch ->normal_prio like __setscheduler().
2942 * Used by the rt_mutex code to implement priority inheritance
2943 * logic. Call site only calls if the priority of the task changed.
2945 void rt_mutex_setprio(struct task_struct *p, int prio)
2947 int oldprio, on_rq, running, enqueue_flag = 0;
2949 const struct sched_class *prev_class;
2951 BUG_ON(prio > MAX_PRIO);
2953 rq = __task_rq_lock(p);
2956 * Idle task boosting is a nono in general. There is one
2957 * exception, when PREEMPT_RT and NOHZ is active:
2959 * The idle task calls get_next_timer_interrupt() and holds
2960 * the timer wheel base->lock on the CPU and another CPU wants
2961 * to access the timer (probably to cancel it). We can safely
2962 * ignore the boosting request, as the idle CPU runs this code
2963 * with interrupts disabled and will complete the lock
2964 * protected section without being interrupted. So there is no
2965 * real need to boost.
2967 if (unlikely(p == rq->idle)) {
2968 WARN_ON(p != rq->curr);
2969 WARN_ON(p->pi_blocked_on);
2973 trace_sched_pi_setprio(p, prio);
2974 p->pi_top_task = rt_mutex_get_top_task(p);
2976 prev_class = p->sched_class;
2978 running = task_current(rq, p);
2980 dequeue_task(rq, p, 0);
2982 p->sched_class->put_prev_task(rq, p);
2985 * Boosting condition are:
2986 * 1. -rt task is running and holds mutex A
2987 * --> -dl task blocks on mutex A
2989 * 2. -dl task is running and holds mutex A
2990 * --> -dl task blocks on mutex A and could preempt the
2993 if (dl_prio(prio)) {
2994 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2995 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2996 p->dl.dl_boosted = 1;
2997 p->dl.dl_throttled = 0;
2998 enqueue_flag = ENQUEUE_REPLENISH;
3000 p->dl.dl_boosted = 0;
3001 p->sched_class = &dl_sched_class;
3002 } else if (rt_prio(prio)) {
3003 if (dl_prio(oldprio))
3004 p->dl.dl_boosted = 0;
3006 enqueue_flag = ENQUEUE_HEAD;
3007 p->sched_class = &rt_sched_class;
3009 if (dl_prio(oldprio))
3010 p->dl.dl_boosted = 0;
3011 p->sched_class = &fair_sched_class;
3017 p->sched_class->set_curr_task(rq);
3019 enqueue_task(rq, p, enqueue_flag);
3021 check_class_changed(rq, p, prev_class, oldprio);
3023 __task_rq_unlock(rq);
3027 void set_user_nice(struct task_struct *p, long nice)
3029 int old_prio, delta, on_rq;
3030 unsigned long flags;
3033 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3036 * We have to be careful, if called from sys_setpriority(),
3037 * the task might be in the middle of scheduling on another CPU.
3039 rq = task_rq_lock(p, &flags);
3041 * The RT priorities are set via sched_setscheduler(), but we still
3042 * allow the 'normal' nice value to be set - but as expected
3043 * it wont have any effect on scheduling until the task is
3044 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3046 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3047 p->static_prio = NICE_TO_PRIO(nice);
3052 dequeue_task(rq, p, 0);
3054 p->static_prio = NICE_TO_PRIO(nice);
3057 p->prio = effective_prio(p);
3058 delta = p->prio - old_prio;
3061 enqueue_task(rq, p, 0);
3063 * If the task increased its priority or is running and
3064 * lowered its priority, then reschedule its CPU:
3066 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3067 resched_task(rq->curr);
3070 task_rq_unlock(rq, p, &flags);
3072 EXPORT_SYMBOL(set_user_nice);
3075 * can_nice - check if a task can reduce its nice value
3079 int can_nice(const struct task_struct *p, const int nice)
3081 /* convert nice value [19,-20] to rlimit style value [1,40] */
3082 int nice_rlim = nice_to_rlimit(nice);
3084 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3085 capable(CAP_SYS_NICE));
3088 #ifdef __ARCH_WANT_SYS_NICE
3091 * sys_nice - change the priority of the current process.
3092 * @increment: priority increment
3094 * sys_setpriority is a more generic, but much slower function that
3095 * does similar things.
3097 SYSCALL_DEFINE1(nice, int, increment)
3102 * Setpriority might change our priority at the same moment.
3103 * We don't have to worry. Conceptually one call occurs first
3104 * and we have a single winner.
3106 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3107 nice = task_nice(current) + increment;
3109 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3110 if (increment < 0 && !can_nice(current, nice))
3113 retval = security_task_setnice(current, nice);
3117 set_user_nice(current, nice);
3124 * task_prio - return the priority value of a given task.
3125 * @p: the task in question.
3127 * Return: The priority value as seen by users in /proc.
3128 * RT tasks are offset by -200. Normal tasks are centered
3129 * around 0, value goes from -16 to +15.
3131 int task_prio(const struct task_struct *p)
3133 return p->prio - MAX_RT_PRIO;
3137 * idle_cpu - is a given cpu idle currently?
3138 * @cpu: the processor in question.
3140 * Return: 1 if the CPU is currently idle. 0 otherwise.
3142 int idle_cpu(int cpu)
3144 struct rq *rq = cpu_rq(cpu);
3146 if (rq->curr != rq->idle)
3153 if (!llist_empty(&rq->wake_list))
3161 * idle_task - return the idle task for a given cpu.
3162 * @cpu: the processor in question.
3164 * Return: The idle task for the cpu @cpu.
3166 struct task_struct *idle_task(int cpu)
3168 return cpu_rq(cpu)->idle;
3172 * find_process_by_pid - find a process with a matching PID value.
3173 * @pid: the pid in question.
3175 * The task of @pid, if found. %NULL otherwise.
3177 static struct task_struct *find_process_by_pid(pid_t pid)
3179 return pid ? find_task_by_vpid(pid) : current;
3183 * This function initializes the sched_dl_entity of a newly becoming
3184 * SCHED_DEADLINE task.
3186 * Only the static values are considered here, the actual runtime and the
3187 * absolute deadline will be properly calculated when the task is enqueued
3188 * for the first time with its new policy.
3191 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3193 struct sched_dl_entity *dl_se = &p->dl;
3195 init_dl_task_timer(dl_se);
3196 dl_se->dl_runtime = attr->sched_runtime;
3197 dl_se->dl_deadline = attr->sched_deadline;
3198 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3199 dl_se->flags = attr->sched_flags;
3200 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3201 dl_se->dl_throttled = 0;
3203 dl_se->dl_yielded = 0;
3206 static void __setscheduler_params(struct task_struct *p,
3207 const struct sched_attr *attr)
3209 int policy = attr->sched_policy;
3211 if (policy == -1) /* setparam */
3216 if (dl_policy(policy))
3217 __setparam_dl(p, attr);
3218 else if (fair_policy(policy))
3219 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3222 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3223 * !rt_policy. Always setting this ensures that things like
3224 * getparam()/getattr() don't report silly values for !rt tasks.
3226 p->rt_priority = attr->sched_priority;
3227 p->normal_prio = normal_prio(p);
3231 /* Actually do priority change: must hold pi & rq lock. */
3232 static void __setscheduler(struct rq *rq, struct task_struct *p,
3233 const struct sched_attr *attr)
3235 __setscheduler_params(p, attr);
3238 * If we get here, there was no pi waiters boosting the
3239 * task. It is safe to use the normal prio.
3241 p->prio = normal_prio(p);
3243 if (dl_prio(p->prio))
3244 p->sched_class = &dl_sched_class;
3245 else if (rt_prio(p->prio))
3246 p->sched_class = &rt_sched_class;
3248 p->sched_class = &fair_sched_class;
3252 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3254 struct sched_dl_entity *dl_se = &p->dl;
3256 attr->sched_priority = p->rt_priority;
3257 attr->sched_runtime = dl_se->dl_runtime;
3258 attr->sched_deadline = dl_se->dl_deadline;
3259 attr->sched_period = dl_se->dl_period;
3260 attr->sched_flags = dl_se->flags;
3264 * This function validates the new parameters of a -deadline task.
3265 * We ask for the deadline not being zero, and greater or equal
3266 * than the runtime, as well as the period of being zero or
3267 * greater than deadline. Furthermore, we have to be sure that
3268 * user parameters are above the internal resolution of 1us (we
3269 * check sched_runtime only since it is always the smaller one) and
3270 * below 2^63 ns (we have to check both sched_deadline and
3271 * sched_period, as the latter can be zero).
3274 __checkparam_dl(const struct sched_attr *attr)
3277 if (attr->sched_deadline == 0)
3281 * Since we truncate DL_SCALE bits, make sure we're at least
3284 if (attr->sched_runtime < (1ULL << DL_SCALE))
3288 * Since we use the MSB for wrap-around and sign issues, make
3289 * sure it's not set (mind that period can be equal to zero).
3291 if (attr->sched_deadline & (1ULL << 63) ||
3292 attr->sched_period & (1ULL << 63))
3295 /* runtime <= deadline <= period (if period != 0) */
3296 if ((attr->sched_period != 0 &&
3297 attr->sched_period < attr->sched_deadline) ||
3298 attr->sched_deadline < attr->sched_runtime)
3305 * check the target process has a UID that matches the current process's
3307 static bool check_same_owner(struct task_struct *p)
3309 const struct cred *cred = current_cred(), *pcred;
3313 pcred = __task_cred(p);
3314 match = (uid_eq(cred->euid, pcred->euid) ||
3315 uid_eq(cred->euid, pcred->uid));
3320 static int __sched_setscheduler(struct task_struct *p,
3321 const struct sched_attr *attr,
3324 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3325 MAX_RT_PRIO - 1 - attr->sched_priority;
3326 int retval, oldprio, oldpolicy = -1, on_rq, running;
3327 int policy = attr->sched_policy;
3328 unsigned long flags;
3329 const struct sched_class *prev_class;
3333 /* may grab non-irq protected spin_locks */
3334 BUG_ON(in_interrupt());
3336 /* double check policy once rq lock held */
3338 reset_on_fork = p->sched_reset_on_fork;
3339 policy = oldpolicy = p->policy;
3341 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3343 if (policy != SCHED_DEADLINE &&
3344 policy != SCHED_FIFO && policy != SCHED_RR &&
3345 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3346 policy != SCHED_IDLE)
3350 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3354 * Valid priorities for SCHED_FIFO and SCHED_RR are
3355 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3356 * SCHED_BATCH and SCHED_IDLE is 0.
3358 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3359 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3361 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3362 (rt_policy(policy) != (attr->sched_priority != 0)))
3366 * Allow unprivileged RT tasks to decrease priority:
3368 if (user && !capable(CAP_SYS_NICE)) {
3369 if (fair_policy(policy)) {
3370 if (attr->sched_nice < task_nice(p) &&
3371 !can_nice(p, attr->sched_nice))
3375 if (rt_policy(policy)) {
3376 unsigned long rlim_rtprio =
3377 task_rlimit(p, RLIMIT_RTPRIO);
3379 /* can't set/change the rt policy */
3380 if (policy != p->policy && !rlim_rtprio)
3383 /* can't increase priority */
3384 if (attr->sched_priority > p->rt_priority &&
3385 attr->sched_priority > rlim_rtprio)
3390 * Can't set/change SCHED_DEADLINE policy at all for now
3391 * (safest behavior); in the future we would like to allow
3392 * unprivileged DL tasks to increase their relative deadline
3393 * or reduce their runtime (both ways reducing utilization)
3395 if (dl_policy(policy))
3399 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3400 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3402 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3403 if (!can_nice(p, task_nice(p)))
3407 /* can't change other user's priorities */
3408 if (!check_same_owner(p))
3411 /* Normal users shall not reset the sched_reset_on_fork flag */
3412 if (p->sched_reset_on_fork && !reset_on_fork)
3417 retval = security_task_setscheduler(p);
3423 * make sure no PI-waiters arrive (or leave) while we are
3424 * changing the priority of the task:
3426 * To be able to change p->policy safely, the appropriate
3427 * runqueue lock must be held.
3429 rq = task_rq_lock(p, &flags);
3432 * Changing the policy of the stop threads its a very bad idea
3434 if (p == rq->stop) {
3435 task_rq_unlock(rq, p, &flags);
3440 * If not changing anything there's no need to proceed further,
3441 * but store a possible modification of reset_on_fork.
3443 if (unlikely(policy == p->policy)) {
3444 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3446 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3448 if (dl_policy(policy))
3451 p->sched_reset_on_fork = reset_on_fork;
3452 task_rq_unlock(rq, p, &flags);
3458 #ifdef CONFIG_RT_GROUP_SCHED
3460 * Do not allow realtime tasks into groups that have no runtime
3463 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3464 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3465 !task_group_is_autogroup(task_group(p))) {
3466 task_rq_unlock(rq, p, &flags);
3471 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3472 cpumask_t *span = rq->rd->span;
3475 * Don't allow tasks with an affinity mask smaller than
3476 * the entire root_domain to become SCHED_DEADLINE. We
3477 * will also fail if there's no bandwidth available.
3479 if (!cpumask_subset(span, &p->cpus_allowed) ||
3480 rq->rd->dl_bw.bw == 0) {
3481 task_rq_unlock(rq, p, &flags);
3488 /* recheck policy now with rq lock held */
3489 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3490 policy = oldpolicy = -1;
3491 task_rq_unlock(rq, p, &flags);
3496 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3497 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3500 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3501 task_rq_unlock(rq, p, &flags);
3505 p->sched_reset_on_fork = reset_on_fork;
3509 * Special case for priority boosted tasks.
3511 * If the new priority is lower or equal (user space view)
3512 * than the current (boosted) priority, we just store the new
3513 * normal parameters and do not touch the scheduler class and
3514 * the runqueue. This will be done when the task deboost
3517 if (rt_mutex_check_prio(p, newprio)) {
3518 __setscheduler_params(p, attr);
3519 task_rq_unlock(rq, p, &flags);
3524 running = task_current(rq, p);
3526 dequeue_task(rq, p, 0);
3528 p->sched_class->put_prev_task(rq, p);
3530 prev_class = p->sched_class;
3531 __setscheduler(rq, p, attr);
3534 p->sched_class->set_curr_task(rq);
3537 * We enqueue to tail when the priority of a task is
3538 * increased (user space view).
3540 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3543 check_class_changed(rq, p, prev_class, oldprio);
3544 task_rq_unlock(rq, p, &flags);
3546 rt_mutex_adjust_pi(p);
3551 static int _sched_setscheduler(struct task_struct *p, int policy,
3552 const struct sched_param *param, bool check)
3554 struct sched_attr attr = {
3555 .sched_policy = policy,
3556 .sched_priority = param->sched_priority,
3557 .sched_nice = PRIO_TO_NICE(p->static_prio),
3561 * Fixup the legacy SCHED_RESET_ON_FORK hack
3563 if (policy & SCHED_RESET_ON_FORK) {
3564 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3565 policy &= ~SCHED_RESET_ON_FORK;
3566 attr.sched_policy = policy;
3569 return __sched_setscheduler(p, &attr, check);
3572 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3573 * @p: the task in question.
3574 * @policy: new policy.
3575 * @param: structure containing the new RT priority.
3577 * Return: 0 on success. An error code otherwise.
3579 * NOTE that the task may be already dead.
3581 int sched_setscheduler(struct task_struct *p, int policy,
3582 const struct sched_param *param)
3584 return _sched_setscheduler(p, policy, param, true);
3586 EXPORT_SYMBOL_GPL(sched_setscheduler);
3588 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3590 return __sched_setscheduler(p, attr, true);
3592 EXPORT_SYMBOL_GPL(sched_setattr);
3595 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3596 * @p: the task in question.
3597 * @policy: new policy.
3598 * @param: structure containing the new RT priority.
3600 * Just like sched_setscheduler, only don't bother checking if the
3601 * current context has permission. For example, this is needed in
3602 * stop_machine(): we create temporary high priority worker threads,
3603 * but our caller might not have that capability.
3605 * Return: 0 on success. An error code otherwise.
3607 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3608 const struct sched_param *param)
3610 return _sched_setscheduler(p, policy, param, false);
3614 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3616 struct sched_param lparam;
3617 struct task_struct *p;
3620 if (!param || pid < 0)
3622 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3627 p = find_process_by_pid(pid);
3629 retval = sched_setscheduler(p, policy, &lparam);
3636 * Mimics kernel/events/core.c perf_copy_attr().
3638 static int sched_copy_attr(struct sched_attr __user *uattr,
3639 struct sched_attr *attr)
3644 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3648 * zero the full structure, so that a short copy will be nice.
3650 memset(attr, 0, sizeof(*attr));
3652 ret = get_user(size, &uattr->size);
3656 if (size > PAGE_SIZE) /* silly large */
3659 if (!size) /* abi compat */
3660 size = SCHED_ATTR_SIZE_VER0;
3662 if (size < SCHED_ATTR_SIZE_VER0)
3666 * If we're handed a bigger struct than we know of,
3667 * ensure all the unknown bits are 0 - i.e. new
3668 * user-space does not rely on any kernel feature
3669 * extensions we dont know about yet.
3671 if (size > sizeof(*attr)) {
3672 unsigned char __user *addr;
3673 unsigned char __user *end;
3676 addr = (void __user *)uattr + sizeof(*attr);
3677 end = (void __user *)uattr + size;
3679 for (; addr < end; addr++) {
3680 ret = get_user(val, addr);
3686 size = sizeof(*attr);
3689 ret = copy_from_user(attr, uattr, size);
3694 * XXX: do we want to be lenient like existing syscalls; or do we want
3695 * to be strict and return an error on out-of-bounds values?
3697 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3702 put_user(sizeof(*attr), &uattr->size);
3707 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3708 * @pid: the pid in question.
3709 * @policy: new policy.
3710 * @param: structure containing the new RT priority.
3712 * Return: 0 on success. An error code otherwise.
3714 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3715 struct sched_param __user *, param)
3717 /* negative values for policy are not valid */
3721 return do_sched_setscheduler(pid, policy, param);
3725 * sys_sched_setparam - set/change the RT priority of a thread
3726 * @pid: the pid in question.
3727 * @param: structure containing the new RT priority.
3729 * Return: 0 on success. An error code otherwise.
3731 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3733 return do_sched_setscheduler(pid, -1, param);
3737 * sys_sched_setattr - same as above, but with extended sched_attr
3738 * @pid: the pid in question.
3739 * @uattr: structure containing the extended parameters.
3740 * @flags: for future extension.
3742 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3743 unsigned int, flags)
3745 struct sched_attr attr;
3746 struct task_struct *p;
3749 if (!uattr || pid < 0 || flags)
3752 retval = sched_copy_attr(uattr, &attr);
3756 if ((int)attr.sched_policy < 0)
3761 p = find_process_by_pid(pid);
3763 retval = sched_setattr(p, &attr);
3770 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3771 * @pid: the pid in question.
3773 * Return: On success, the policy of the thread. Otherwise, a negative error
3776 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3778 struct task_struct *p;
3786 p = find_process_by_pid(pid);
3788 retval = security_task_getscheduler(p);
3791 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3798 * sys_sched_getparam - get the RT priority of a thread
3799 * @pid: the pid in question.
3800 * @param: structure containing the RT priority.
3802 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3805 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3807 struct sched_param lp = { .sched_priority = 0 };
3808 struct task_struct *p;
3811 if (!param || pid < 0)
3815 p = find_process_by_pid(pid);
3820 retval = security_task_getscheduler(p);
3824 if (task_has_rt_policy(p))
3825 lp.sched_priority = p->rt_priority;
3829 * This one might sleep, we cannot do it with a spinlock held ...
3831 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3840 static int sched_read_attr(struct sched_attr __user *uattr,
3841 struct sched_attr *attr,
3846 if (!access_ok(VERIFY_WRITE, uattr, usize))
3850 * If we're handed a smaller struct than we know of,
3851 * ensure all the unknown bits are 0 - i.e. old
3852 * user-space does not get uncomplete information.
3854 if (usize < sizeof(*attr)) {
3855 unsigned char *addr;
3858 addr = (void *)attr + usize;
3859 end = (void *)attr + sizeof(*attr);
3861 for (; addr < end; addr++) {
3869 ret = copy_to_user(uattr, attr, attr->size);
3877 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3878 * @pid: the pid in question.
3879 * @uattr: structure containing the extended parameters.
3880 * @size: sizeof(attr) for fwd/bwd comp.
3881 * @flags: for future extension.
3883 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3884 unsigned int, size, unsigned int, flags)
3886 struct sched_attr attr = {
3887 .size = sizeof(struct sched_attr),
3889 struct task_struct *p;
3892 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3893 size < SCHED_ATTR_SIZE_VER0 || flags)
3897 p = find_process_by_pid(pid);
3902 retval = security_task_getscheduler(p);
3906 attr.sched_policy = p->policy;
3907 if (p->sched_reset_on_fork)
3908 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3909 if (task_has_dl_policy(p))
3910 __getparam_dl(p, &attr);
3911 else if (task_has_rt_policy(p))
3912 attr.sched_priority = p->rt_priority;
3914 attr.sched_nice = task_nice(p);
3918 retval = sched_read_attr(uattr, &attr, size);
3926 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3928 cpumask_var_t cpus_allowed, new_mask;
3929 struct task_struct *p;
3934 p = find_process_by_pid(pid);
3940 /* Prevent p going away */
3944 if (p->flags & PF_NO_SETAFFINITY) {
3948 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3952 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3954 goto out_free_cpus_allowed;
3957 if (!check_same_owner(p)) {
3959 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3966 retval = security_task_setscheduler(p);
3971 cpuset_cpus_allowed(p, cpus_allowed);
3972 cpumask_and(new_mask, in_mask, cpus_allowed);
3975 * Since bandwidth control happens on root_domain basis,
3976 * if admission test is enabled, we only admit -deadline
3977 * tasks allowed to run on all the CPUs in the task's
3981 if (task_has_dl_policy(p)) {
3982 const struct cpumask *span = task_rq(p)->rd->span;
3984 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3991 retval = set_cpus_allowed_ptr(p, new_mask);
3994 cpuset_cpus_allowed(p, cpus_allowed);
3995 if (!cpumask_subset(new_mask, cpus_allowed)) {
3997 * We must have raced with a concurrent cpuset
3998 * update. Just reset the cpus_allowed to the
3999 * cpuset's cpus_allowed
4001 cpumask_copy(new_mask, cpus_allowed);
4006 free_cpumask_var(new_mask);
4007 out_free_cpus_allowed:
4008 free_cpumask_var(cpus_allowed);
4014 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4015 struct cpumask *new_mask)
4017 if (len < cpumask_size())
4018 cpumask_clear(new_mask);
4019 else if (len > cpumask_size())
4020 len = cpumask_size();
4022 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4026 * sys_sched_setaffinity - set the cpu affinity of a process
4027 * @pid: pid of the process
4028 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4029 * @user_mask_ptr: user-space pointer to the new cpu mask
4031 * Return: 0 on success. An error code otherwise.
4033 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4034 unsigned long __user *, user_mask_ptr)
4036 cpumask_var_t new_mask;
4039 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4042 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4044 retval = sched_setaffinity(pid, new_mask);
4045 free_cpumask_var(new_mask);
4049 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4051 struct task_struct *p;
4052 unsigned long flags;
4058 p = find_process_by_pid(pid);
4062 retval = security_task_getscheduler(p);
4066 raw_spin_lock_irqsave(&p->pi_lock, flags);
4067 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4068 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4077 * sys_sched_getaffinity - get the cpu affinity of a process
4078 * @pid: pid of the process
4079 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4080 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4082 * Return: 0 on success. An error code otherwise.
4084 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4085 unsigned long __user *, user_mask_ptr)
4090 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4092 if (len & (sizeof(unsigned long)-1))
4095 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4098 ret = sched_getaffinity(pid, mask);
4100 size_t retlen = min_t(size_t, len, cpumask_size());
4102 if (copy_to_user(user_mask_ptr, mask, retlen))
4107 free_cpumask_var(mask);
4113 * sys_sched_yield - yield the current processor to other threads.
4115 * This function yields the current CPU to other tasks. If there are no
4116 * other threads running on this CPU then this function will return.
4120 SYSCALL_DEFINE0(sched_yield)
4122 struct rq *rq = this_rq_lock();
4124 schedstat_inc(rq, yld_count);
4125 current->sched_class->yield_task(rq);
4128 * Since we are going to call schedule() anyway, there's
4129 * no need to preempt or enable interrupts:
4131 __release(rq->lock);
4132 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4133 do_raw_spin_unlock(&rq->lock);
4134 sched_preempt_enable_no_resched();
4141 static void __cond_resched(void)
4143 __preempt_count_add(PREEMPT_ACTIVE);
4145 __preempt_count_sub(PREEMPT_ACTIVE);
4148 int __sched _cond_resched(void)
4151 if (should_resched()) {
4157 EXPORT_SYMBOL(_cond_resched);
4160 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4161 * call schedule, and on return reacquire the lock.
4163 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4164 * operations here to prevent schedule() from being called twice (once via
4165 * spin_unlock(), once by hand).
4167 int __cond_resched_lock(spinlock_t *lock)
4169 bool need_rcu_resched = rcu_should_resched();
4170 int resched = should_resched();
4173 lockdep_assert_held(lock);
4175 if (spin_needbreak(lock) || resched || need_rcu_resched) {
4179 else if (unlikely(need_rcu_resched))
4188 EXPORT_SYMBOL(__cond_resched_lock);
4190 int __sched __cond_resched_softirq(void)
4192 BUG_ON(!in_softirq());
4194 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4195 if (should_resched()) {
4203 EXPORT_SYMBOL(__cond_resched_softirq);
4206 * yield - yield the current processor to other threads.
4208 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4210 * The scheduler is at all times free to pick the calling task as the most
4211 * eligible task to run, if removing the yield() call from your code breaks
4212 * it, its already broken.
4214 * Typical broken usage is:
4219 * where one assumes that yield() will let 'the other' process run that will
4220 * make event true. If the current task is a SCHED_FIFO task that will never
4221 * happen. Never use yield() as a progress guarantee!!
4223 * If you want to use yield() to wait for something, use wait_event().
4224 * If you want to use yield() to be 'nice' for others, use cond_resched().
4225 * If you still want to use yield(), do not!
4227 void __sched yield(void)
4229 set_current_state(TASK_RUNNING);
4232 EXPORT_SYMBOL(yield);
4235 * yield_to - yield the current processor to another thread in
4236 * your thread group, or accelerate that thread toward the
4237 * processor it's on.
4239 * @preempt: whether task preemption is allowed or not
4241 * It's the caller's job to ensure that the target task struct
4242 * can't go away on us before we can do any checks.
4245 * true (>0) if we indeed boosted the target task.
4246 * false (0) if we failed to boost the target.
4247 * -ESRCH if there's no task to yield to.
4249 int __sched yield_to(struct task_struct *p, bool preempt)
4251 struct task_struct *curr = current;
4252 struct rq *rq, *p_rq;
4253 unsigned long flags;
4256 local_irq_save(flags);
4262 * If we're the only runnable task on the rq and target rq also
4263 * has only one task, there's absolutely no point in yielding.
4265 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4270 double_rq_lock(rq, p_rq);
4271 if (task_rq(p) != p_rq) {
4272 double_rq_unlock(rq, p_rq);
4276 if (!curr->sched_class->yield_to_task)
4279 if (curr->sched_class != p->sched_class)
4282 if (task_running(p_rq, p) || p->state)
4285 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4287 schedstat_inc(rq, yld_count);
4289 * Make p's CPU reschedule; pick_next_entity takes care of
4292 if (preempt && rq != p_rq)
4293 resched_task(p_rq->curr);
4297 double_rq_unlock(rq, p_rq);
4299 local_irq_restore(flags);
4306 EXPORT_SYMBOL_GPL(yield_to);
4309 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4310 * that process accounting knows that this is a task in IO wait state.
4312 void __sched io_schedule(void)
4314 struct rq *rq = raw_rq();
4316 delayacct_blkio_start();
4317 atomic_inc(&rq->nr_iowait);
4318 blk_flush_plug(current);
4319 current->in_iowait = 1;
4321 current->in_iowait = 0;
4322 atomic_dec(&rq->nr_iowait);
4323 delayacct_blkio_end();
4325 EXPORT_SYMBOL(io_schedule);
4327 long __sched io_schedule_timeout(long timeout)
4329 struct rq *rq = raw_rq();
4332 delayacct_blkio_start();
4333 atomic_inc(&rq->nr_iowait);
4334 blk_flush_plug(current);
4335 current->in_iowait = 1;
4336 ret = schedule_timeout(timeout);
4337 current->in_iowait = 0;
4338 atomic_dec(&rq->nr_iowait);
4339 delayacct_blkio_end();
4344 * sys_sched_get_priority_max - return maximum RT priority.
4345 * @policy: scheduling class.
4347 * Return: On success, this syscall returns the maximum
4348 * rt_priority that can be used by a given scheduling class.
4349 * On failure, a negative error code is returned.
4351 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4358 ret = MAX_USER_RT_PRIO-1;
4360 case SCHED_DEADLINE:
4371 * sys_sched_get_priority_min - return minimum RT priority.
4372 * @policy: scheduling class.
4374 * Return: On success, this syscall returns the minimum
4375 * rt_priority that can be used by a given scheduling class.
4376 * On failure, a negative error code is returned.
4378 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4387 case SCHED_DEADLINE:
4397 * sys_sched_rr_get_interval - return the default timeslice of a process.
4398 * @pid: pid of the process.
4399 * @interval: userspace pointer to the timeslice value.
4401 * this syscall writes the default timeslice value of a given process
4402 * into the user-space timespec buffer. A value of '0' means infinity.
4404 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4407 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4408 struct timespec __user *, interval)
4410 struct task_struct *p;
4411 unsigned int time_slice;
4412 unsigned long flags;
4422 p = find_process_by_pid(pid);
4426 retval = security_task_getscheduler(p);
4430 rq = task_rq_lock(p, &flags);
4432 if (p->sched_class->get_rr_interval)
4433 time_slice = p->sched_class->get_rr_interval(rq, p);
4434 task_rq_unlock(rq, p, &flags);
4437 jiffies_to_timespec(time_slice, &t);
4438 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4446 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4448 void sched_show_task(struct task_struct *p)
4450 unsigned long free = 0;
4454 state = p->state ? __ffs(p->state) + 1 : 0;
4455 printk(KERN_INFO "%-15.15s %c", p->comm,
4456 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4457 #if BITS_PER_LONG == 32
4458 if (state == TASK_RUNNING)
4459 printk(KERN_CONT " running ");
4461 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4463 if (state == TASK_RUNNING)
4464 printk(KERN_CONT " running task ");
4466 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4468 #ifdef CONFIG_DEBUG_STACK_USAGE
4469 free = stack_not_used(p);
4472 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4474 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4475 task_pid_nr(p), ppid,
4476 (unsigned long)task_thread_info(p)->flags);
4478 print_worker_info(KERN_INFO, p);
4479 show_stack(p, NULL);
4482 void show_state_filter(unsigned long state_filter)
4484 struct task_struct *g, *p;
4486 #if BITS_PER_LONG == 32
4488 " task PC stack pid father\n");
4491 " task PC stack pid father\n");
4494 do_each_thread(g, p) {
4496 * reset the NMI-timeout, listing all files on a slow
4497 * console might take a lot of time:
4499 touch_nmi_watchdog();
4500 if (!state_filter || (p->state & state_filter))
4502 } while_each_thread(g, p);
4504 touch_all_softlockup_watchdogs();
4506 #ifdef CONFIG_SCHED_DEBUG
4507 sysrq_sched_debug_show();
4511 * Only show locks if all tasks are dumped:
4514 debug_show_all_locks();
4517 void init_idle_bootup_task(struct task_struct *idle)
4519 idle->sched_class = &idle_sched_class;
4523 * init_idle - set up an idle thread for a given CPU
4524 * @idle: task in question
4525 * @cpu: cpu the idle task belongs to
4527 * NOTE: this function does not set the idle thread's NEED_RESCHED
4528 * flag, to make booting more robust.
4530 void init_idle(struct task_struct *idle, int cpu)
4532 struct rq *rq = cpu_rq(cpu);
4533 unsigned long flags;
4535 raw_spin_lock_irqsave(&rq->lock, flags);
4537 __sched_fork(0, idle);
4538 idle->state = TASK_RUNNING;
4539 idle->se.exec_start = sched_clock();
4541 do_set_cpus_allowed(idle, cpumask_of(cpu));
4543 * We're having a chicken and egg problem, even though we are
4544 * holding rq->lock, the cpu isn't yet set to this cpu so the
4545 * lockdep check in task_group() will fail.
4547 * Similar case to sched_fork(). / Alternatively we could
4548 * use task_rq_lock() here and obtain the other rq->lock.
4553 __set_task_cpu(idle, cpu);
4556 rq->curr = rq->idle = idle;
4558 #if defined(CONFIG_SMP)
4561 raw_spin_unlock_irqrestore(&rq->lock, flags);
4563 /* Set the preempt count _outside_ the spinlocks! */
4564 init_idle_preempt_count(idle, cpu);
4567 * The idle tasks have their own, simple scheduling class:
4569 idle->sched_class = &idle_sched_class;
4570 ftrace_graph_init_idle_task(idle, cpu);
4571 vtime_init_idle(idle, cpu);
4572 #if defined(CONFIG_SMP)
4573 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4578 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4580 if (p->sched_class && p->sched_class->set_cpus_allowed)
4581 p->sched_class->set_cpus_allowed(p, new_mask);
4583 cpumask_copy(&p->cpus_allowed, new_mask);
4584 p->nr_cpus_allowed = cpumask_weight(new_mask);
4588 * This is how migration works:
4590 * 1) we invoke migration_cpu_stop() on the target CPU using
4592 * 2) stopper starts to run (implicitly forcing the migrated thread
4594 * 3) it checks whether the migrated task is still in the wrong runqueue.
4595 * 4) if it's in the wrong runqueue then the migration thread removes
4596 * it and puts it into the right queue.
4597 * 5) stopper completes and stop_one_cpu() returns and the migration
4602 * Change a given task's CPU affinity. Migrate the thread to a
4603 * proper CPU and schedule it away if the CPU it's executing on
4604 * is removed from the allowed bitmask.
4606 * NOTE: the caller must have a valid reference to the task, the
4607 * task must not exit() & deallocate itself prematurely. The
4608 * call is not atomic; no spinlocks may be held.
4610 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4612 unsigned long flags;
4614 unsigned int dest_cpu;
4617 rq = task_rq_lock(p, &flags);
4619 if (cpumask_equal(&p->cpus_allowed, new_mask))
4622 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4627 do_set_cpus_allowed(p, new_mask);
4629 /* Can the task run on the task's current CPU? If so, we're done */
4630 if (cpumask_test_cpu(task_cpu(p), new_mask))
4633 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4635 struct migration_arg arg = { p, dest_cpu };
4636 /* Need help from migration thread: drop lock and wait. */
4637 task_rq_unlock(rq, p, &flags);
4638 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4639 tlb_migrate_finish(p->mm);
4643 task_rq_unlock(rq, p, &flags);
4647 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4650 * Move (not current) task off this cpu, onto dest cpu. We're doing
4651 * this because either it can't run here any more (set_cpus_allowed()
4652 * away from this CPU, or CPU going down), or because we're
4653 * attempting to rebalance this task on exec (sched_exec).
4655 * So we race with normal scheduler movements, but that's OK, as long
4656 * as the task is no longer on this CPU.
4658 * Returns non-zero if task was successfully migrated.
4660 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4662 struct rq *rq_dest, *rq_src;
4665 if (unlikely(!cpu_active(dest_cpu)))
4668 rq_src = cpu_rq(src_cpu);
4669 rq_dest = cpu_rq(dest_cpu);
4671 raw_spin_lock(&p->pi_lock);
4672 double_rq_lock(rq_src, rq_dest);
4673 /* Already moved. */
4674 if (task_cpu(p) != src_cpu)
4676 /* Affinity changed (again). */
4677 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4681 * If we're not on a rq, the next wake-up will ensure we're
4685 dequeue_task(rq_src, p, 0);
4686 set_task_cpu(p, dest_cpu);
4687 enqueue_task(rq_dest, p, 0);
4688 check_preempt_curr(rq_dest, p, 0);
4693 double_rq_unlock(rq_src, rq_dest);
4694 raw_spin_unlock(&p->pi_lock);
4698 #ifdef CONFIG_NUMA_BALANCING
4699 /* Migrate current task p to target_cpu */
4700 int migrate_task_to(struct task_struct *p, int target_cpu)
4702 struct migration_arg arg = { p, target_cpu };
4703 int curr_cpu = task_cpu(p);
4705 if (curr_cpu == target_cpu)
4708 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4711 /* TODO: This is not properly updating schedstats */
4713 trace_sched_move_numa(p, curr_cpu, target_cpu);
4714 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4718 * Requeue a task on a given node and accurately track the number of NUMA
4719 * tasks on the runqueues
4721 void sched_setnuma(struct task_struct *p, int nid)
4724 unsigned long flags;
4725 bool on_rq, running;
4727 rq = task_rq_lock(p, &flags);
4729 running = task_current(rq, p);
4732 dequeue_task(rq, p, 0);
4734 p->sched_class->put_prev_task(rq, p);
4736 p->numa_preferred_nid = nid;
4739 p->sched_class->set_curr_task(rq);
4741 enqueue_task(rq, p, 0);
4742 task_rq_unlock(rq, p, &flags);
4747 * migration_cpu_stop - this will be executed by a highprio stopper thread
4748 * and performs thread migration by bumping thread off CPU then
4749 * 'pushing' onto another runqueue.
4751 static int migration_cpu_stop(void *data)
4753 struct migration_arg *arg = data;
4756 * The original target cpu might have gone down and we might
4757 * be on another cpu but it doesn't matter.
4759 local_irq_disable();
4760 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4765 #ifdef CONFIG_HOTPLUG_CPU
4768 * Ensures that the idle task is using init_mm right before its cpu goes
4771 void idle_task_exit(void)
4773 struct mm_struct *mm = current->active_mm;
4775 BUG_ON(cpu_online(smp_processor_id()));
4777 if (mm != &init_mm) {
4778 switch_mm(mm, &init_mm, current);
4779 finish_arch_post_lock_switch();
4785 * Since this CPU is going 'away' for a while, fold any nr_active delta
4786 * we might have. Assumes we're called after migrate_tasks() so that the
4787 * nr_active count is stable.
4789 * Also see the comment "Global load-average calculations".
4791 static void calc_load_migrate(struct rq *rq)
4793 long delta = calc_load_fold_active(rq);
4795 atomic_long_add(delta, &calc_load_tasks);
4798 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4802 static const struct sched_class fake_sched_class = {
4803 .put_prev_task = put_prev_task_fake,
4806 static struct task_struct fake_task = {
4808 * Avoid pull_{rt,dl}_task()
4810 .prio = MAX_PRIO + 1,
4811 .sched_class = &fake_sched_class,
4815 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4816 * try_to_wake_up()->select_task_rq().
4818 * Called with rq->lock held even though we'er in stop_machine() and
4819 * there's no concurrency possible, we hold the required locks anyway
4820 * because of lock validation efforts.
4822 static void migrate_tasks(unsigned int dead_cpu)
4824 struct rq *rq = cpu_rq(dead_cpu);
4825 struct task_struct *next, *stop = rq->stop;
4829 * Fudge the rq selection such that the below task selection loop
4830 * doesn't get stuck on the currently eligible stop task.
4832 * We're currently inside stop_machine() and the rq is either stuck
4833 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4834 * either way we should never end up calling schedule() until we're
4840 * put_prev_task() and pick_next_task() sched
4841 * class method both need to have an up-to-date
4842 * value of rq->clock[_task]
4844 update_rq_clock(rq);
4848 * There's this thread running, bail when that's the only
4851 if (rq->nr_running == 1)
4854 next = pick_next_task(rq, &fake_task);
4856 next->sched_class->put_prev_task(rq, next);
4858 /* Find suitable destination for @next, with force if needed. */
4859 dest_cpu = select_fallback_rq(dead_cpu, next);
4860 raw_spin_unlock(&rq->lock);
4862 __migrate_task(next, dead_cpu, dest_cpu);
4864 raw_spin_lock(&rq->lock);
4870 #endif /* CONFIG_HOTPLUG_CPU */
4872 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4874 static struct ctl_table sd_ctl_dir[] = {
4876 .procname = "sched_domain",
4882 static struct ctl_table sd_ctl_root[] = {
4884 .procname = "kernel",
4886 .child = sd_ctl_dir,
4891 static struct ctl_table *sd_alloc_ctl_entry(int n)
4893 struct ctl_table *entry =
4894 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4899 static void sd_free_ctl_entry(struct ctl_table **tablep)
4901 struct ctl_table *entry;
4904 * In the intermediate directories, both the child directory and
4905 * procname are dynamically allocated and could fail but the mode
4906 * will always be set. In the lowest directory the names are
4907 * static strings and all have proc handlers.
4909 for (entry = *tablep; entry->mode; entry++) {
4911 sd_free_ctl_entry(&entry->child);
4912 if (entry->proc_handler == NULL)
4913 kfree(entry->procname);
4920 static int min_load_idx = 0;
4921 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4924 set_table_entry(struct ctl_table *entry,
4925 const char *procname, void *data, int maxlen,
4926 umode_t mode, proc_handler *proc_handler,
4929 entry->procname = procname;
4931 entry->maxlen = maxlen;
4933 entry->proc_handler = proc_handler;
4936 entry->extra1 = &min_load_idx;
4937 entry->extra2 = &max_load_idx;
4941 static struct ctl_table *
4942 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4944 struct ctl_table *table = sd_alloc_ctl_entry(14);
4949 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4950 sizeof(long), 0644, proc_doulongvec_minmax, false);
4951 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4952 sizeof(long), 0644, proc_doulongvec_minmax, false);
4953 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4954 sizeof(int), 0644, proc_dointvec_minmax, true);
4955 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4956 sizeof(int), 0644, proc_dointvec_minmax, true);
4957 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4958 sizeof(int), 0644, proc_dointvec_minmax, true);
4959 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4960 sizeof(int), 0644, proc_dointvec_minmax, true);
4961 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4962 sizeof(int), 0644, proc_dointvec_minmax, true);
4963 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4964 sizeof(int), 0644, proc_dointvec_minmax, false);
4965 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4966 sizeof(int), 0644, proc_dointvec_minmax, false);
4967 set_table_entry(&table[9], "cache_nice_tries",
4968 &sd->cache_nice_tries,
4969 sizeof(int), 0644, proc_dointvec_minmax, false);
4970 set_table_entry(&table[10], "flags", &sd->flags,
4971 sizeof(int), 0644, proc_dointvec_minmax, false);
4972 set_table_entry(&table[11], "max_newidle_lb_cost",
4973 &sd->max_newidle_lb_cost,
4974 sizeof(long), 0644, proc_doulongvec_minmax, false);
4975 set_table_entry(&table[12], "name", sd->name,
4976 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4977 /* &table[13] is terminator */
4982 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4984 struct ctl_table *entry, *table;
4985 struct sched_domain *sd;
4986 int domain_num = 0, i;
4989 for_each_domain(cpu, sd)
4991 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4996 for_each_domain(cpu, sd) {
4997 snprintf(buf, 32, "domain%d", i);
4998 entry->procname = kstrdup(buf, GFP_KERNEL);
5000 entry->child = sd_alloc_ctl_domain_table(sd);
5007 static struct ctl_table_header *sd_sysctl_header;
5008 static void register_sched_domain_sysctl(void)
5010 int i, cpu_num = num_possible_cpus();
5011 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5014 WARN_ON(sd_ctl_dir[0].child);
5015 sd_ctl_dir[0].child = entry;
5020 for_each_possible_cpu(i) {
5021 snprintf(buf, 32, "cpu%d", i);
5022 entry->procname = kstrdup(buf, GFP_KERNEL);
5024 entry->child = sd_alloc_ctl_cpu_table(i);
5028 WARN_ON(sd_sysctl_header);
5029 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5032 /* may be called multiple times per register */
5033 static void unregister_sched_domain_sysctl(void)
5035 if (sd_sysctl_header)
5036 unregister_sysctl_table(sd_sysctl_header);
5037 sd_sysctl_header = NULL;
5038 if (sd_ctl_dir[0].child)
5039 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5042 static void register_sched_domain_sysctl(void)
5045 static void unregister_sched_domain_sysctl(void)
5050 static void set_rq_online(struct rq *rq)
5053 const struct sched_class *class;
5055 cpumask_set_cpu(rq->cpu, rq->rd->online);
5058 for_each_class(class) {
5059 if (class->rq_online)
5060 class->rq_online(rq);
5065 static void set_rq_offline(struct rq *rq)
5068 const struct sched_class *class;
5070 for_each_class(class) {
5071 if (class->rq_offline)
5072 class->rq_offline(rq);
5075 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5081 * migration_call - callback that gets triggered when a CPU is added.
5082 * Here we can start up the necessary migration thread for the new CPU.
5085 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5087 int cpu = (long)hcpu;
5088 unsigned long flags;
5089 struct rq *rq = cpu_rq(cpu);
5091 switch (action & ~CPU_TASKS_FROZEN) {
5093 case CPU_UP_PREPARE:
5094 rq->calc_load_update = calc_load_update;
5098 /* Update our root-domain */
5099 raw_spin_lock_irqsave(&rq->lock, flags);
5101 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5105 raw_spin_unlock_irqrestore(&rq->lock, flags);
5108 #ifdef CONFIG_HOTPLUG_CPU
5110 sched_ttwu_pending();
5111 /* Update our root-domain */
5112 raw_spin_lock_irqsave(&rq->lock, flags);
5114 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5118 BUG_ON(rq->nr_running != 1); /* the migration thread */
5119 raw_spin_unlock_irqrestore(&rq->lock, flags);
5123 calc_load_migrate(rq);
5128 update_max_interval();
5134 * Register at high priority so that task migration (migrate_all_tasks)
5135 * happens before everything else. This has to be lower priority than
5136 * the notifier in the perf_event subsystem, though.
5138 static struct notifier_block migration_notifier = {
5139 .notifier_call = migration_call,
5140 .priority = CPU_PRI_MIGRATION,
5143 static void __cpuinit set_cpu_rq_start_time(void)
5145 int cpu = smp_processor_id();
5146 struct rq *rq = cpu_rq(cpu);
5147 rq->age_stamp = sched_clock_cpu(cpu);
5150 static int sched_cpu_active(struct notifier_block *nfb,
5151 unsigned long action, void *hcpu)
5153 switch (action & ~CPU_TASKS_FROZEN) {
5155 set_cpu_rq_start_time();
5157 case CPU_DOWN_FAILED:
5158 set_cpu_active((long)hcpu, true);
5165 static int sched_cpu_inactive(struct notifier_block *nfb,
5166 unsigned long action, void *hcpu)
5168 unsigned long flags;
5169 long cpu = (long)hcpu;
5171 switch (action & ~CPU_TASKS_FROZEN) {
5172 case CPU_DOWN_PREPARE:
5173 set_cpu_active(cpu, false);
5175 /* explicitly allow suspend */
5176 if (!(action & CPU_TASKS_FROZEN)) {
5177 struct dl_bw *dl_b = dl_bw_of(cpu);
5181 raw_spin_lock_irqsave(&dl_b->lock, flags);
5182 cpus = dl_bw_cpus(cpu);
5183 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5184 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5187 return notifier_from_errno(-EBUSY);
5195 static int __init migration_init(void)
5197 void *cpu = (void *)(long)smp_processor_id();
5200 /* Initialize migration for the boot CPU */
5201 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5202 BUG_ON(err == NOTIFY_BAD);
5203 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5204 register_cpu_notifier(&migration_notifier);
5206 /* Register cpu active notifiers */
5207 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5208 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5212 early_initcall(migration_init);
5217 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5219 #ifdef CONFIG_SCHED_DEBUG
5221 static __read_mostly int sched_debug_enabled;
5223 static int __init sched_debug_setup(char *str)
5225 sched_debug_enabled = 1;
5229 early_param("sched_debug", sched_debug_setup);
5231 static inline bool sched_debug(void)
5233 return sched_debug_enabled;
5236 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5237 struct cpumask *groupmask)
5239 struct sched_group *group = sd->groups;
5242 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5243 cpumask_clear(groupmask);
5245 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5247 if (!(sd->flags & SD_LOAD_BALANCE)) {
5248 printk("does not load-balance\n");
5250 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5255 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5257 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5258 printk(KERN_ERR "ERROR: domain->span does not contain "
5261 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5262 printk(KERN_ERR "ERROR: domain->groups does not contain"
5266 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5270 printk(KERN_ERR "ERROR: group is NULL\n");
5275 * Even though we initialize ->capacity to something semi-sane,
5276 * we leave capacity_orig unset. This allows us to detect if
5277 * domain iteration is still funny without causing /0 traps.
5279 if (!group->sgc->capacity_orig) {
5280 printk(KERN_CONT "\n");
5281 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5285 if (!cpumask_weight(sched_group_cpus(group))) {
5286 printk(KERN_CONT "\n");
5287 printk(KERN_ERR "ERROR: empty group\n");
5291 if (!(sd->flags & SD_OVERLAP) &&
5292 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5293 printk(KERN_CONT "\n");
5294 printk(KERN_ERR "ERROR: repeated CPUs\n");
5298 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5300 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5302 printk(KERN_CONT " %s", str);
5303 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5304 printk(KERN_CONT " (cpu_capacity = %d)",
5305 group->sgc->capacity);
5308 group = group->next;
5309 } while (group != sd->groups);
5310 printk(KERN_CONT "\n");
5312 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5313 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5316 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5317 printk(KERN_ERR "ERROR: parent span is not a superset "
5318 "of domain->span\n");
5322 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5326 if (!sched_debug_enabled)
5330 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5334 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5337 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5345 #else /* !CONFIG_SCHED_DEBUG */
5346 # define sched_domain_debug(sd, cpu) do { } while (0)
5347 static inline bool sched_debug(void)
5351 #endif /* CONFIG_SCHED_DEBUG */
5353 static int sd_degenerate(struct sched_domain *sd)
5355 if (cpumask_weight(sched_domain_span(sd)) == 1)
5358 /* Following flags need at least 2 groups */
5359 if (sd->flags & (SD_LOAD_BALANCE |
5360 SD_BALANCE_NEWIDLE |
5363 SD_SHARE_CPUCAPACITY |
5364 SD_SHARE_PKG_RESOURCES |
5365 SD_SHARE_POWERDOMAIN)) {
5366 if (sd->groups != sd->groups->next)
5370 /* Following flags don't use groups */
5371 if (sd->flags & (SD_WAKE_AFFINE))
5378 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5380 unsigned long cflags = sd->flags, pflags = parent->flags;
5382 if (sd_degenerate(parent))
5385 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5388 /* Flags needing groups don't count if only 1 group in parent */
5389 if (parent->groups == parent->groups->next) {
5390 pflags &= ~(SD_LOAD_BALANCE |
5391 SD_BALANCE_NEWIDLE |
5394 SD_SHARE_CPUCAPACITY |
5395 SD_SHARE_PKG_RESOURCES |
5397 SD_SHARE_POWERDOMAIN);
5398 if (nr_node_ids == 1)
5399 pflags &= ~SD_SERIALIZE;
5401 if (~cflags & pflags)
5407 static void free_rootdomain(struct rcu_head *rcu)
5409 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5411 cpupri_cleanup(&rd->cpupri);
5412 cpudl_cleanup(&rd->cpudl);
5413 free_cpumask_var(rd->dlo_mask);
5414 free_cpumask_var(rd->rto_mask);
5415 free_cpumask_var(rd->online);
5416 free_cpumask_var(rd->span);
5420 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5422 struct root_domain *old_rd = NULL;
5423 unsigned long flags;
5425 raw_spin_lock_irqsave(&rq->lock, flags);
5430 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5433 cpumask_clear_cpu(rq->cpu, old_rd->span);
5436 * If we dont want to free the old_rd yet then
5437 * set old_rd to NULL to skip the freeing later
5440 if (!atomic_dec_and_test(&old_rd->refcount))
5444 atomic_inc(&rd->refcount);
5447 cpumask_set_cpu(rq->cpu, rd->span);
5448 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5451 raw_spin_unlock_irqrestore(&rq->lock, flags);
5454 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5457 static int init_rootdomain(struct root_domain *rd)
5459 memset(rd, 0, sizeof(*rd));
5461 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5463 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5465 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5467 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5470 init_dl_bw(&rd->dl_bw);
5471 if (cpudl_init(&rd->cpudl) != 0)
5474 if (cpupri_init(&rd->cpupri) != 0)
5479 free_cpumask_var(rd->rto_mask);
5481 free_cpumask_var(rd->dlo_mask);
5483 free_cpumask_var(rd->online);
5485 free_cpumask_var(rd->span);
5491 * By default the system creates a single root-domain with all cpus as
5492 * members (mimicking the global state we have today).
5494 struct root_domain def_root_domain;
5496 static void init_defrootdomain(void)
5498 init_rootdomain(&def_root_domain);
5500 atomic_set(&def_root_domain.refcount, 1);
5503 static struct root_domain *alloc_rootdomain(void)
5505 struct root_domain *rd;
5507 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5511 if (init_rootdomain(rd) != 0) {
5519 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5521 struct sched_group *tmp, *first;
5530 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5535 } while (sg != first);
5538 static void free_sched_domain(struct rcu_head *rcu)
5540 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5543 * If its an overlapping domain it has private groups, iterate and
5546 if (sd->flags & SD_OVERLAP) {
5547 free_sched_groups(sd->groups, 1);
5548 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5549 kfree(sd->groups->sgc);
5555 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5557 call_rcu(&sd->rcu, free_sched_domain);
5560 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5562 for (; sd; sd = sd->parent)
5563 destroy_sched_domain(sd, cpu);
5567 * Keep a special pointer to the highest sched_domain that has
5568 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5569 * allows us to avoid some pointer chasing select_idle_sibling().
5571 * Also keep a unique ID per domain (we use the first cpu number in
5572 * the cpumask of the domain), this allows us to quickly tell if
5573 * two cpus are in the same cache domain, see cpus_share_cache().
5575 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5576 DEFINE_PER_CPU(int, sd_llc_size);
5577 DEFINE_PER_CPU(int, sd_llc_id);
5578 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5579 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5580 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5582 static void update_top_cache_domain(int cpu)
5584 struct sched_domain *sd;
5585 struct sched_domain *busy_sd = NULL;
5589 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5591 id = cpumask_first(sched_domain_span(sd));
5592 size = cpumask_weight(sched_domain_span(sd));
5593 busy_sd = sd->parent; /* sd_busy */
5595 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5597 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5598 per_cpu(sd_llc_size, cpu) = size;
5599 per_cpu(sd_llc_id, cpu) = id;
5601 sd = lowest_flag_domain(cpu, SD_NUMA);
5602 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5604 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5605 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5609 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5610 * hold the hotplug lock.
5613 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5615 struct rq *rq = cpu_rq(cpu);
5616 struct sched_domain *tmp;
5618 /* Remove the sched domains which do not contribute to scheduling. */
5619 for (tmp = sd; tmp; ) {
5620 struct sched_domain *parent = tmp->parent;
5624 if (sd_parent_degenerate(tmp, parent)) {
5625 tmp->parent = parent->parent;
5627 parent->parent->child = tmp;
5629 * Transfer SD_PREFER_SIBLING down in case of a
5630 * degenerate parent; the spans match for this
5631 * so the property transfers.
5633 if (parent->flags & SD_PREFER_SIBLING)
5634 tmp->flags |= SD_PREFER_SIBLING;
5635 destroy_sched_domain(parent, cpu);
5640 if (sd && sd_degenerate(sd)) {
5643 destroy_sched_domain(tmp, cpu);
5648 sched_domain_debug(sd, cpu);
5650 rq_attach_root(rq, rd);
5652 rcu_assign_pointer(rq->sd, sd);
5653 destroy_sched_domains(tmp, cpu);
5655 update_top_cache_domain(cpu);
5658 /* cpus with isolated domains */
5659 static cpumask_var_t cpu_isolated_map;
5661 /* Setup the mask of cpus configured for isolated domains */
5662 static int __init isolated_cpu_setup(char *str)
5664 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5665 cpulist_parse(str, cpu_isolated_map);
5669 __setup("isolcpus=", isolated_cpu_setup);
5672 struct sched_domain ** __percpu sd;
5673 struct root_domain *rd;
5684 * Build an iteration mask that can exclude certain CPUs from the upwards
5687 * Asymmetric node setups can result in situations where the domain tree is of
5688 * unequal depth, make sure to skip domains that already cover the entire
5691 * In that case build_sched_domains() will have terminated the iteration early
5692 * and our sibling sd spans will be empty. Domains should always include the
5693 * cpu they're built on, so check that.
5696 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5698 const struct cpumask *span = sched_domain_span(sd);
5699 struct sd_data *sdd = sd->private;
5700 struct sched_domain *sibling;
5703 for_each_cpu(i, span) {
5704 sibling = *per_cpu_ptr(sdd->sd, i);
5705 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5708 cpumask_set_cpu(i, sched_group_mask(sg));
5713 * Return the canonical balance cpu for this group, this is the first cpu
5714 * of this group that's also in the iteration mask.
5716 int group_balance_cpu(struct sched_group *sg)
5718 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5722 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5724 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5725 const struct cpumask *span = sched_domain_span(sd);
5726 struct cpumask *covered = sched_domains_tmpmask;
5727 struct sd_data *sdd = sd->private;
5728 struct sched_domain *child;
5731 cpumask_clear(covered);
5733 for_each_cpu(i, span) {
5734 struct cpumask *sg_span;
5736 if (cpumask_test_cpu(i, covered))
5739 child = *per_cpu_ptr(sdd->sd, i);
5741 /* See the comment near build_group_mask(). */
5742 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5745 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5746 GFP_KERNEL, cpu_to_node(cpu));
5751 sg_span = sched_group_cpus(sg);
5753 child = child->child;
5754 cpumask_copy(sg_span, sched_domain_span(child));
5756 cpumask_set_cpu(i, sg_span);
5758 cpumask_or(covered, covered, sg_span);
5760 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5761 if (atomic_inc_return(&sg->sgc->ref) == 1)
5762 build_group_mask(sd, sg);
5765 * Initialize sgc->capacity such that even if we mess up the
5766 * domains and no possible iteration will get us here, we won't
5769 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5770 sg->sgc->capacity_orig = sg->sgc->capacity;
5773 * Make sure the first group of this domain contains the
5774 * canonical balance cpu. Otherwise the sched_domain iteration
5775 * breaks. See update_sg_lb_stats().
5777 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5778 group_balance_cpu(sg) == cpu)
5788 sd->groups = groups;
5793 free_sched_groups(first, 0);
5798 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5800 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5801 struct sched_domain *child = sd->child;
5804 cpu = cpumask_first(sched_domain_span(child));
5807 *sg = *per_cpu_ptr(sdd->sg, cpu);
5808 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5809 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5816 * build_sched_groups will build a circular linked list of the groups
5817 * covered by the given span, and will set each group's ->cpumask correctly,
5818 * and ->cpu_capacity to 0.
5820 * Assumes the sched_domain tree is fully constructed
5823 build_sched_groups(struct sched_domain *sd, int cpu)
5825 struct sched_group *first = NULL, *last = NULL;
5826 struct sd_data *sdd = sd->private;
5827 const struct cpumask *span = sched_domain_span(sd);
5828 struct cpumask *covered;
5831 get_group(cpu, sdd, &sd->groups);
5832 atomic_inc(&sd->groups->ref);
5834 if (cpu != cpumask_first(span))
5837 lockdep_assert_held(&sched_domains_mutex);
5838 covered = sched_domains_tmpmask;
5840 cpumask_clear(covered);
5842 for_each_cpu(i, span) {
5843 struct sched_group *sg;
5846 if (cpumask_test_cpu(i, covered))
5849 group = get_group(i, sdd, &sg);
5850 cpumask_setall(sched_group_mask(sg));
5852 for_each_cpu(j, span) {
5853 if (get_group(j, sdd, NULL) != group)
5856 cpumask_set_cpu(j, covered);
5857 cpumask_set_cpu(j, sched_group_cpus(sg));
5872 * Initialize sched groups cpu_capacity.
5874 * cpu_capacity indicates the capacity of sched group, which is used while
5875 * distributing the load between different sched groups in a sched domain.
5876 * Typically cpu_capacity for all the groups in a sched domain will be same
5877 * unless there are asymmetries in the topology. If there are asymmetries,
5878 * group having more cpu_capacity will pickup more load compared to the
5879 * group having less cpu_capacity.
5881 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5883 struct sched_group *sg = sd->groups;
5888 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5890 } while (sg != sd->groups);
5892 if (cpu != group_balance_cpu(sg))
5895 update_group_capacity(sd, cpu);
5896 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5900 * Initializers for schedule domains
5901 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5904 static int default_relax_domain_level = -1;
5905 int sched_domain_level_max;
5907 static int __init setup_relax_domain_level(char *str)
5909 if (kstrtoint(str, 0, &default_relax_domain_level))
5910 pr_warn("Unable to set relax_domain_level\n");
5914 __setup("relax_domain_level=", setup_relax_domain_level);
5916 static void set_domain_attribute(struct sched_domain *sd,
5917 struct sched_domain_attr *attr)
5921 if (!attr || attr->relax_domain_level < 0) {
5922 if (default_relax_domain_level < 0)
5925 request = default_relax_domain_level;
5927 request = attr->relax_domain_level;
5928 if (request < sd->level) {
5929 /* turn off idle balance on this domain */
5930 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5932 /* turn on idle balance on this domain */
5933 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5937 static void __sdt_free(const struct cpumask *cpu_map);
5938 static int __sdt_alloc(const struct cpumask *cpu_map);
5940 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5941 const struct cpumask *cpu_map)
5945 if (!atomic_read(&d->rd->refcount))
5946 free_rootdomain(&d->rd->rcu); /* fall through */
5948 free_percpu(d->sd); /* fall through */
5950 __sdt_free(cpu_map); /* fall through */
5956 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5957 const struct cpumask *cpu_map)
5959 memset(d, 0, sizeof(*d));
5961 if (__sdt_alloc(cpu_map))
5962 return sa_sd_storage;
5963 d->sd = alloc_percpu(struct sched_domain *);
5965 return sa_sd_storage;
5966 d->rd = alloc_rootdomain();
5969 return sa_rootdomain;
5973 * NULL the sd_data elements we've used to build the sched_domain and
5974 * sched_group structure so that the subsequent __free_domain_allocs()
5975 * will not free the data we're using.
5977 static void claim_allocations(int cpu, struct sched_domain *sd)
5979 struct sd_data *sdd = sd->private;
5981 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5982 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5984 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5985 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5987 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
5988 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
5992 static int sched_domains_numa_levels;
5993 static int *sched_domains_numa_distance;
5994 static struct cpumask ***sched_domains_numa_masks;
5995 static int sched_domains_curr_level;
5999 * SD_flags allowed in topology descriptions.
6001 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6002 * SD_SHARE_PKG_RESOURCES - describes shared caches
6003 * SD_NUMA - describes NUMA topologies
6004 * SD_SHARE_POWERDOMAIN - describes shared power domain
6007 * SD_ASYM_PACKING - describes SMT quirks
6009 #define TOPOLOGY_SD_FLAGS \
6010 (SD_SHARE_CPUCAPACITY | \
6011 SD_SHARE_PKG_RESOURCES | \
6014 SD_SHARE_POWERDOMAIN)
6016 static struct sched_domain *
6017 sd_init(struct sched_domain_topology_level *tl, int cpu)
6019 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6020 int sd_weight, sd_flags = 0;
6024 * Ugly hack to pass state to sd_numa_mask()...
6026 sched_domains_curr_level = tl->numa_level;
6029 sd_weight = cpumask_weight(tl->mask(cpu));
6032 sd_flags = (*tl->sd_flags)();
6033 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6034 "wrong sd_flags in topology description\n"))
6035 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6037 *sd = (struct sched_domain){
6038 .min_interval = sd_weight,
6039 .max_interval = 2*sd_weight,
6041 .imbalance_pct = 125,
6043 .cache_nice_tries = 0,
6050 .flags = 1*SD_LOAD_BALANCE
6051 | 1*SD_BALANCE_NEWIDLE
6056 | 0*SD_SHARE_CPUCAPACITY
6057 | 0*SD_SHARE_PKG_RESOURCES
6059 | 0*SD_PREFER_SIBLING
6064 .last_balance = jiffies,
6065 .balance_interval = sd_weight,
6067 .max_newidle_lb_cost = 0,
6068 .next_decay_max_lb_cost = jiffies,
6069 #ifdef CONFIG_SCHED_DEBUG
6075 * Convert topological properties into behaviour.
6078 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6079 sd->imbalance_pct = 110;
6080 sd->smt_gain = 1178; /* ~15% */
6082 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6083 sd->imbalance_pct = 117;
6084 sd->cache_nice_tries = 1;
6088 } else if (sd->flags & SD_NUMA) {
6089 sd->cache_nice_tries = 2;
6093 sd->flags |= SD_SERIALIZE;
6094 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6095 sd->flags &= ~(SD_BALANCE_EXEC |
6102 sd->flags |= SD_PREFER_SIBLING;
6103 sd->cache_nice_tries = 1;
6108 sd->private = &tl->data;
6114 * Topology list, bottom-up.
6116 static struct sched_domain_topology_level default_topology[] = {
6117 #ifdef CONFIG_SCHED_SMT
6118 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6120 #ifdef CONFIG_SCHED_MC
6121 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6123 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6127 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6129 #define for_each_sd_topology(tl) \
6130 for (tl = sched_domain_topology; tl->mask; tl++)
6132 void set_sched_topology(struct sched_domain_topology_level *tl)
6134 sched_domain_topology = tl;
6139 static const struct cpumask *sd_numa_mask(int cpu)
6141 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6144 static void sched_numa_warn(const char *str)
6146 static int done = false;
6154 printk(KERN_WARNING "ERROR: %s\n\n", str);
6156 for (i = 0; i < nr_node_ids; i++) {
6157 printk(KERN_WARNING " ");
6158 for (j = 0; j < nr_node_ids; j++)
6159 printk(KERN_CONT "%02d ", node_distance(i,j));
6160 printk(KERN_CONT "\n");
6162 printk(KERN_WARNING "\n");
6165 static bool find_numa_distance(int distance)
6169 if (distance == node_distance(0, 0))
6172 for (i = 0; i < sched_domains_numa_levels; i++) {
6173 if (sched_domains_numa_distance[i] == distance)
6180 static void sched_init_numa(void)
6182 int next_distance, curr_distance = node_distance(0, 0);
6183 struct sched_domain_topology_level *tl;
6187 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6188 if (!sched_domains_numa_distance)
6192 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6193 * unique distances in the node_distance() table.
6195 * Assumes node_distance(0,j) includes all distances in
6196 * node_distance(i,j) in order to avoid cubic time.
6198 next_distance = curr_distance;
6199 for (i = 0; i < nr_node_ids; i++) {
6200 for (j = 0; j < nr_node_ids; j++) {
6201 for (k = 0; k < nr_node_ids; k++) {
6202 int distance = node_distance(i, k);
6204 if (distance > curr_distance &&
6205 (distance < next_distance ||
6206 next_distance == curr_distance))
6207 next_distance = distance;
6210 * While not a strong assumption it would be nice to know
6211 * about cases where if node A is connected to B, B is not
6212 * equally connected to A.
6214 if (sched_debug() && node_distance(k, i) != distance)
6215 sched_numa_warn("Node-distance not symmetric");
6217 if (sched_debug() && i && !find_numa_distance(distance))
6218 sched_numa_warn("Node-0 not representative");
6220 if (next_distance != curr_distance) {
6221 sched_domains_numa_distance[level++] = next_distance;
6222 sched_domains_numa_levels = level;
6223 curr_distance = next_distance;
6228 * In case of sched_debug() we verify the above assumption.
6234 * 'level' contains the number of unique distances, excluding the
6235 * identity distance node_distance(i,i).
6237 * The sched_domains_numa_distance[] array includes the actual distance
6242 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6243 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6244 * the array will contain less then 'level' members. This could be
6245 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6246 * in other functions.
6248 * We reset it to 'level' at the end of this function.
6250 sched_domains_numa_levels = 0;
6252 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6253 if (!sched_domains_numa_masks)
6257 * Now for each level, construct a mask per node which contains all
6258 * cpus of nodes that are that many hops away from us.
6260 for (i = 0; i < level; i++) {
6261 sched_domains_numa_masks[i] =
6262 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6263 if (!sched_domains_numa_masks[i])
6266 for (j = 0; j < nr_node_ids; j++) {
6267 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6271 sched_domains_numa_masks[i][j] = mask;
6273 for (k = 0; k < nr_node_ids; k++) {
6274 if (node_distance(j, k) > sched_domains_numa_distance[i])
6277 cpumask_or(mask, mask, cpumask_of_node(k));
6282 /* Compute default topology size */
6283 for (i = 0; sched_domain_topology[i].mask; i++);
6285 tl = kzalloc((i + level + 1) *
6286 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6291 * Copy the default topology bits..
6293 for (i = 0; sched_domain_topology[i].mask; i++)
6294 tl[i] = sched_domain_topology[i];
6297 * .. and append 'j' levels of NUMA goodness.
6299 for (j = 0; j < level; i++, j++) {
6300 tl[i] = (struct sched_domain_topology_level){
6301 .mask = sd_numa_mask,
6302 .sd_flags = cpu_numa_flags,
6303 .flags = SDTL_OVERLAP,
6309 sched_domain_topology = tl;
6311 sched_domains_numa_levels = level;
6314 static void sched_domains_numa_masks_set(int cpu)
6317 int node = cpu_to_node(cpu);
6319 for (i = 0; i < sched_domains_numa_levels; i++) {
6320 for (j = 0; j < nr_node_ids; j++) {
6321 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6322 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6327 static void sched_domains_numa_masks_clear(int cpu)
6330 for (i = 0; i < sched_domains_numa_levels; i++) {
6331 for (j = 0; j < nr_node_ids; j++)
6332 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6337 * Update sched_domains_numa_masks[level][node] array when new cpus
6340 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6341 unsigned long action,
6344 int cpu = (long)hcpu;
6346 switch (action & ~CPU_TASKS_FROZEN) {
6348 sched_domains_numa_masks_set(cpu);
6352 sched_domains_numa_masks_clear(cpu);
6362 static inline void sched_init_numa(void)
6366 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6367 unsigned long action,
6372 #endif /* CONFIG_NUMA */
6374 static int __sdt_alloc(const struct cpumask *cpu_map)
6376 struct sched_domain_topology_level *tl;
6379 for_each_sd_topology(tl) {
6380 struct sd_data *sdd = &tl->data;
6382 sdd->sd = alloc_percpu(struct sched_domain *);
6386 sdd->sg = alloc_percpu(struct sched_group *);
6390 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6394 for_each_cpu(j, cpu_map) {
6395 struct sched_domain *sd;
6396 struct sched_group *sg;
6397 struct sched_group_capacity *sgc;
6399 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6400 GFP_KERNEL, cpu_to_node(j));
6404 *per_cpu_ptr(sdd->sd, j) = sd;
6406 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6407 GFP_KERNEL, cpu_to_node(j));
6413 *per_cpu_ptr(sdd->sg, j) = sg;
6415 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6416 GFP_KERNEL, cpu_to_node(j));
6420 *per_cpu_ptr(sdd->sgc, j) = sgc;
6427 static void __sdt_free(const struct cpumask *cpu_map)
6429 struct sched_domain_topology_level *tl;
6432 for_each_sd_topology(tl) {
6433 struct sd_data *sdd = &tl->data;
6435 for_each_cpu(j, cpu_map) {
6436 struct sched_domain *sd;
6439 sd = *per_cpu_ptr(sdd->sd, j);
6440 if (sd && (sd->flags & SD_OVERLAP))
6441 free_sched_groups(sd->groups, 0);
6442 kfree(*per_cpu_ptr(sdd->sd, j));
6446 kfree(*per_cpu_ptr(sdd->sg, j));
6448 kfree(*per_cpu_ptr(sdd->sgc, j));
6450 free_percpu(sdd->sd);
6452 free_percpu(sdd->sg);
6454 free_percpu(sdd->sgc);
6459 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6460 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6461 struct sched_domain *child, int cpu)
6463 struct sched_domain *sd = sd_init(tl, cpu);
6467 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6469 sd->level = child->level + 1;
6470 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6474 set_domain_attribute(sd, attr);
6480 * Build sched domains for a given set of cpus and attach the sched domains
6481 * to the individual cpus
6483 static int build_sched_domains(const struct cpumask *cpu_map,
6484 struct sched_domain_attr *attr)
6486 enum s_alloc alloc_state;
6487 struct sched_domain *sd;
6489 int i, ret = -ENOMEM;
6491 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6492 if (alloc_state != sa_rootdomain)
6495 /* Set up domains for cpus specified by the cpu_map. */
6496 for_each_cpu(i, cpu_map) {
6497 struct sched_domain_topology_level *tl;
6500 for_each_sd_topology(tl) {
6501 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6502 if (tl == sched_domain_topology)
6503 *per_cpu_ptr(d.sd, i) = sd;
6504 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6505 sd->flags |= SD_OVERLAP;
6506 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6511 /* Build the groups for the domains */
6512 for_each_cpu(i, cpu_map) {
6513 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6514 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6515 if (sd->flags & SD_OVERLAP) {
6516 if (build_overlap_sched_groups(sd, i))
6519 if (build_sched_groups(sd, i))
6525 /* Calculate CPU capacity for physical packages and nodes */
6526 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6527 if (!cpumask_test_cpu(i, cpu_map))
6530 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6531 claim_allocations(i, sd);
6532 init_sched_groups_capacity(i, sd);
6536 /* Attach the domains */
6538 for_each_cpu(i, cpu_map) {
6539 sd = *per_cpu_ptr(d.sd, i);
6540 cpu_attach_domain(sd, d.rd, i);
6546 __free_domain_allocs(&d, alloc_state, cpu_map);
6550 static cpumask_var_t *doms_cur; /* current sched domains */
6551 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6552 static struct sched_domain_attr *dattr_cur;
6553 /* attribues of custom domains in 'doms_cur' */
6556 * Special case: If a kmalloc of a doms_cur partition (array of
6557 * cpumask) fails, then fallback to a single sched domain,
6558 * as determined by the single cpumask fallback_doms.
6560 static cpumask_var_t fallback_doms;
6563 * arch_update_cpu_topology lets virtualized architectures update the
6564 * cpu core maps. It is supposed to return 1 if the topology changed
6565 * or 0 if it stayed the same.
6567 int __weak arch_update_cpu_topology(void)
6572 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6575 cpumask_var_t *doms;
6577 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6580 for (i = 0; i < ndoms; i++) {
6581 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6582 free_sched_domains(doms, i);
6589 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6592 for (i = 0; i < ndoms; i++)
6593 free_cpumask_var(doms[i]);
6598 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6599 * For now this just excludes isolated cpus, but could be used to
6600 * exclude other special cases in the future.
6602 static int init_sched_domains(const struct cpumask *cpu_map)
6606 arch_update_cpu_topology();
6608 doms_cur = alloc_sched_domains(ndoms_cur);
6610 doms_cur = &fallback_doms;
6611 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6612 err = build_sched_domains(doms_cur[0], NULL);
6613 register_sched_domain_sysctl();
6619 * Detach sched domains from a group of cpus specified in cpu_map
6620 * These cpus will now be attached to the NULL domain
6622 static void detach_destroy_domains(const struct cpumask *cpu_map)
6627 for_each_cpu(i, cpu_map)
6628 cpu_attach_domain(NULL, &def_root_domain, i);
6632 /* handle null as "default" */
6633 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6634 struct sched_domain_attr *new, int idx_new)
6636 struct sched_domain_attr tmp;
6643 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6644 new ? (new + idx_new) : &tmp,
6645 sizeof(struct sched_domain_attr));
6649 * Partition sched domains as specified by the 'ndoms_new'
6650 * cpumasks in the array doms_new[] of cpumasks. This compares
6651 * doms_new[] to the current sched domain partitioning, doms_cur[].
6652 * It destroys each deleted domain and builds each new domain.
6654 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6655 * The masks don't intersect (don't overlap.) We should setup one
6656 * sched domain for each mask. CPUs not in any of the cpumasks will
6657 * not be load balanced. If the same cpumask appears both in the
6658 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6661 * The passed in 'doms_new' should be allocated using
6662 * alloc_sched_domains. This routine takes ownership of it and will
6663 * free_sched_domains it when done with it. If the caller failed the
6664 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6665 * and partition_sched_domains() will fallback to the single partition
6666 * 'fallback_doms', it also forces the domains to be rebuilt.
6668 * If doms_new == NULL it will be replaced with cpu_online_mask.
6669 * ndoms_new == 0 is a special case for destroying existing domains,
6670 * and it will not create the default domain.
6672 * Call with hotplug lock held
6674 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6675 struct sched_domain_attr *dattr_new)
6680 mutex_lock(&sched_domains_mutex);
6682 /* always unregister in case we don't destroy any domains */
6683 unregister_sched_domain_sysctl();
6685 /* Let architecture update cpu core mappings. */
6686 new_topology = arch_update_cpu_topology();
6688 n = doms_new ? ndoms_new : 0;
6690 /* Destroy deleted domains */
6691 for (i = 0; i < ndoms_cur; i++) {
6692 for (j = 0; j < n && !new_topology; j++) {
6693 if (cpumask_equal(doms_cur[i], doms_new[j])
6694 && dattrs_equal(dattr_cur, i, dattr_new, j))
6697 /* no match - a current sched domain not in new doms_new[] */
6698 detach_destroy_domains(doms_cur[i]);
6704 if (doms_new == NULL) {
6706 doms_new = &fallback_doms;
6707 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6708 WARN_ON_ONCE(dattr_new);
6711 /* Build new domains */
6712 for (i = 0; i < ndoms_new; i++) {
6713 for (j = 0; j < n && !new_topology; j++) {
6714 if (cpumask_equal(doms_new[i], doms_cur[j])
6715 && dattrs_equal(dattr_new, i, dattr_cur, j))
6718 /* no match - add a new doms_new */
6719 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6724 /* Remember the new sched domains */
6725 if (doms_cur != &fallback_doms)
6726 free_sched_domains(doms_cur, ndoms_cur);
6727 kfree(dattr_cur); /* kfree(NULL) is safe */
6728 doms_cur = doms_new;
6729 dattr_cur = dattr_new;
6730 ndoms_cur = ndoms_new;
6732 register_sched_domain_sysctl();
6734 mutex_unlock(&sched_domains_mutex);
6737 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6740 * Update cpusets according to cpu_active mask. If cpusets are
6741 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6742 * around partition_sched_domains().
6744 * If we come here as part of a suspend/resume, don't touch cpusets because we
6745 * want to restore it back to its original state upon resume anyway.
6747 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6751 case CPU_ONLINE_FROZEN:
6752 case CPU_DOWN_FAILED_FROZEN:
6755 * num_cpus_frozen tracks how many CPUs are involved in suspend
6756 * resume sequence. As long as this is not the last online
6757 * operation in the resume sequence, just build a single sched
6758 * domain, ignoring cpusets.
6761 if (likely(num_cpus_frozen)) {
6762 partition_sched_domains(1, NULL, NULL);
6767 * This is the last CPU online operation. So fall through and
6768 * restore the original sched domains by considering the
6769 * cpuset configurations.
6773 case CPU_DOWN_FAILED:
6774 cpuset_update_active_cpus(true);
6782 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6786 case CPU_DOWN_PREPARE:
6787 cpuset_update_active_cpus(false);
6789 case CPU_DOWN_PREPARE_FROZEN:
6791 partition_sched_domains(1, NULL, NULL);
6799 void __init sched_init_smp(void)
6801 cpumask_var_t non_isolated_cpus;
6803 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6804 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6809 * There's no userspace yet to cause hotplug operations; hence all the
6810 * cpu masks are stable and all blatant races in the below code cannot
6813 mutex_lock(&sched_domains_mutex);
6814 init_sched_domains(cpu_active_mask);
6815 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6816 if (cpumask_empty(non_isolated_cpus))
6817 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6818 mutex_unlock(&sched_domains_mutex);
6820 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6821 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6822 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6826 /* Move init over to a non-isolated CPU */
6827 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6829 sched_init_granularity();
6830 free_cpumask_var(non_isolated_cpus);
6832 init_sched_rt_class();
6833 init_sched_dl_class();
6836 void __init sched_init_smp(void)
6838 sched_init_granularity();
6840 #endif /* CONFIG_SMP */
6842 const_debug unsigned int sysctl_timer_migration = 1;
6844 int in_sched_functions(unsigned long addr)
6846 return in_lock_functions(addr) ||
6847 (addr >= (unsigned long)__sched_text_start
6848 && addr < (unsigned long)__sched_text_end);
6851 #ifdef CONFIG_CGROUP_SCHED
6853 * Default task group.
6854 * Every task in system belongs to this group at bootup.
6856 struct task_group root_task_group;
6857 LIST_HEAD(task_groups);
6860 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6862 void __init sched_init(void)
6865 unsigned long alloc_size = 0, ptr;
6867 #ifdef CONFIG_FAIR_GROUP_SCHED
6868 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6870 #ifdef CONFIG_RT_GROUP_SCHED
6871 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6873 #ifdef CONFIG_CPUMASK_OFFSTACK
6874 alloc_size += num_possible_cpus() * cpumask_size();
6877 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6879 #ifdef CONFIG_FAIR_GROUP_SCHED
6880 root_task_group.se = (struct sched_entity **)ptr;
6881 ptr += nr_cpu_ids * sizeof(void **);
6883 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6884 ptr += nr_cpu_ids * sizeof(void **);
6886 #endif /* CONFIG_FAIR_GROUP_SCHED */
6887 #ifdef CONFIG_RT_GROUP_SCHED
6888 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6889 ptr += nr_cpu_ids * sizeof(void **);
6891 root_task_group.rt_rq = (struct rt_rq **)ptr;
6892 ptr += nr_cpu_ids * sizeof(void **);
6894 #endif /* CONFIG_RT_GROUP_SCHED */
6895 #ifdef CONFIG_CPUMASK_OFFSTACK
6896 for_each_possible_cpu(i) {
6897 per_cpu(load_balance_mask, i) = (void *)ptr;
6898 ptr += cpumask_size();
6900 #endif /* CONFIG_CPUMASK_OFFSTACK */
6903 init_rt_bandwidth(&def_rt_bandwidth,
6904 global_rt_period(), global_rt_runtime());
6905 init_dl_bandwidth(&def_dl_bandwidth,
6906 global_rt_period(), global_rt_runtime());
6909 init_defrootdomain();
6912 #ifdef CONFIG_RT_GROUP_SCHED
6913 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6914 global_rt_period(), global_rt_runtime());
6915 #endif /* CONFIG_RT_GROUP_SCHED */
6917 #ifdef CONFIG_CGROUP_SCHED
6918 list_add(&root_task_group.list, &task_groups);
6919 INIT_LIST_HEAD(&root_task_group.children);
6920 INIT_LIST_HEAD(&root_task_group.siblings);
6921 autogroup_init(&init_task);
6923 #endif /* CONFIG_CGROUP_SCHED */
6925 for_each_possible_cpu(i) {
6929 raw_spin_lock_init(&rq->lock);
6931 rq->calc_load_active = 0;
6932 rq->calc_load_update = jiffies + LOAD_FREQ;
6933 init_cfs_rq(&rq->cfs);
6934 init_rt_rq(&rq->rt, rq);
6935 init_dl_rq(&rq->dl, rq);
6936 #ifdef CONFIG_FAIR_GROUP_SCHED
6937 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6938 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6940 * How much cpu bandwidth does root_task_group get?
6942 * In case of task-groups formed thr' the cgroup filesystem, it
6943 * gets 100% of the cpu resources in the system. This overall
6944 * system cpu resource is divided among the tasks of
6945 * root_task_group and its child task-groups in a fair manner,
6946 * based on each entity's (task or task-group's) weight
6947 * (se->load.weight).
6949 * In other words, if root_task_group has 10 tasks of weight
6950 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6951 * then A0's share of the cpu resource is:
6953 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6955 * We achieve this by letting root_task_group's tasks sit
6956 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6958 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6959 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6960 #endif /* CONFIG_FAIR_GROUP_SCHED */
6962 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6963 #ifdef CONFIG_RT_GROUP_SCHED
6964 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6967 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6968 rq->cpu_load[j] = 0;
6970 rq->last_load_update_tick = jiffies;
6975 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
6976 rq->post_schedule = 0;
6977 rq->active_balance = 0;
6978 rq->next_balance = jiffies;
6983 rq->avg_idle = 2*sysctl_sched_migration_cost;
6984 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6986 INIT_LIST_HEAD(&rq->cfs_tasks);
6988 rq_attach_root(rq, &def_root_domain);
6989 #ifdef CONFIG_NO_HZ_COMMON
6992 #ifdef CONFIG_NO_HZ_FULL
6993 rq->last_sched_tick = 0;
6997 atomic_set(&rq->nr_iowait, 0);
7000 set_load_weight(&init_task);
7002 #ifdef CONFIG_PREEMPT_NOTIFIERS
7003 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7007 * The boot idle thread does lazy MMU switching as well:
7009 atomic_inc(&init_mm.mm_count);
7010 enter_lazy_tlb(&init_mm, current);
7013 * Make us the idle thread. Technically, schedule() should not be
7014 * called from this thread, however somewhere below it might be,
7015 * but because we are the idle thread, we just pick up running again
7016 * when this runqueue becomes "idle".
7018 init_idle(current, smp_processor_id());
7020 calc_load_update = jiffies + LOAD_FREQ;
7023 * During early bootup we pretend to be a normal task:
7025 current->sched_class = &fair_sched_class;
7028 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7029 /* May be allocated at isolcpus cmdline parse time */
7030 if (cpu_isolated_map == NULL)
7031 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7032 idle_thread_set_boot_cpu();
7033 set_cpu_rq_start_time();
7035 init_sched_fair_class();
7037 scheduler_running = 1;
7040 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7041 static inline int preempt_count_equals(int preempt_offset)
7043 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7045 return (nested == preempt_offset);
7048 void __might_sleep(const char *file, int line, int preempt_offset)
7050 static unsigned long prev_jiffy; /* ratelimiting */
7052 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7053 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7054 !is_idle_task(current)) ||
7055 system_state != SYSTEM_RUNNING || oops_in_progress)
7057 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7059 prev_jiffy = jiffies;
7062 "BUG: sleeping function called from invalid context at %s:%d\n",
7065 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7066 in_atomic(), irqs_disabled(),
7067 current->pid, current->comm);
7069 debug_show_held_locks(current);
7070 if (irqs_disabled())
7071 print_irqtrace_events(current);
7072 #ifdef CONFIG_DEBUG_PREEMPT
7073 if (!preempt_count_equals(preempt_offset)) {
7074 pr_err("Preemption disabled at:");
7075 print_ip_sym(current->preempt_disable_ip);
7081 EXPORT_SYMBOL(__might_sleep);
7084 #ifdef CONFIG_MAGIC_SYSRQ
7085 static void normalize_task(struct rq *rq, struct task_struct *p)
7087 const struct sched_class *prev_class = p->sched_class;
7088 struct sched_attr attr = {
7089 .sched_policy = SCHED_NORMAL,
7091 int old_prio = p->prio;
7096 dequeue_task(rq, p, 0);
7097 __setscheduler(rq, p, &attr);
7099 enqueue_task(rq, p, 0);
7100 resched_task(rq->curr);
7103 check_class_changed(rq, p, prev_class, old_prio);
7106 void normalize_rt_tasks(void)
7108 struct task_struct *g, *p;
7109 unsigned long flags;
7112 read_lock_irqsave(&tasklist_lock, flags);
7113 do_each_thread(g, p) {
7115 * Only normalize user tasks:
7120 p->se.exec_start = 0;
7121 #ifdef CONFIG_SCHEDSTATS
7122 p->se.statistics.wait_start = 0;
7123 p->se.statistics.sleep_start = 0;
7124 p->se.statistics.block_start = 0;
7127 if (!dl_task(p) && !rt_task(p)) {
7129 * Renice negative nice level userspace
7132 if (task_nice(p) < 0 && p->mm)
7133 set_user_nice(p, 0);
7137 raw_spin_lock(&p->pi_lock);
7138 rq = __task_rq_lock(p);
7140 normalize_task(rq, p);
7142 __task_rq_unlock(rq);
7143 raw_spin_unlock(&p->pi_lock);
7144 } while_each_thread(g, p);
7146 read_unlock_irqrestore(&tasklist_lock, flags);
7149 #endif /* CONFIG_MAGIC_SYSRQ */
7151 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7153 * These functions are only useful for the IA64 MCA handling, or kdb.
7155 * They can only be called when the whole system has been
7156 * stopped - every CPU needs to be quiescent, and no scheduling
7157 * activity can take place. Using them for anything else would
7158 * be a serious bug, and as a result, they aren't even visible
7159 * under any other configuration.
7163 * curr_task - return the current task for a given cpu.
7164 * @cpu: the processor in question.
7166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7168 * Return: The current task for @cpu.
7170 struct task_struct *curr_task(int cpu)
7172 return cpu_curr(cpu);
7175 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7179 * set_curr_task - set the current task for a given cpu.
7180 * @cpu: the processor in question.
7181 * @p: the task pointer to set.
7183 * Description: This function must only be used when non-maskable interrupts
7184 * are serviced on a separate stack. It allows the architecture to switch the
7185 * notion of the current task on a cpu in a non-blocking manner. This function
7186 * must be called with all CPU's synchronized, and interrupts disabled, the
7187 * and caller must save the original value of the current task (see
7188 * curr_task() above) and restore that value before reenabling interrupts and
7189 * re-starting the system.
7191 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7193 void set_curr_task(int cpu, struct task_struct *p)
7200 #ifdef CONFIG_CGROUP_SCHED
7201 /* task_group_lock serializes the addition/removal of task groups */
7202 static DEFINE_SPINLOCK(task_group_lock);
7204 static void free_sched_group(struct task_group *tg)
7206 free_fair_sched_group(tg);
7207 free_rt_sched_group(tg);
7212 /* allocate runqueue etc for a new task group */
7213 struct task_group *sched_create_group(struct task_group *parent)
7215 struct task_group *tg;
7217 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7219 return ERR_PTR(-ENOMEM);
7221 if (!alloc_fair_sched_group(tg, parent))
7224 if (!alloc_rt_sched_group(tg, parent))
7230 free_sched_group(tg);
7231 return ERR_PTR(-ENOMEM);
7234 void sched_online_group(struct task_group *tg, struct task_group *parent)
7236 unsigned long flags;
7238 spin_lock_irqsave(&task_group_lock, flags);
7239 list_add_rcu(&tg->list, &task_groups);
7241 WARN_ON(!parent); /* root should already exist */
7243 tg->parent = parent;
7244 INIT_LIST_HEAD(&tg->children);
7245 list_add_rcu(&tg->siblings, &parent->children);
7246 spin_unlock_irqrestore(&task_group_lock, flags);
7249 /* rcu callback to free various structures associated with a task group */
7250 static void free_sched_group_rcu(struct rcu_head *rhp)
7252 /* now it should be safe to free those cfs_rqs */
7253 free_sched_group(container_of(rhp, struct task_group, rcu));
7256 /* Destroy runqueue etc associated with a task group */
7257 void sched_destroy_group(struct task_group *tg)
7259 /* wait for possible concurrent references to cfs_rqs complete */
7260 call_rcu(&tg->rcu, free_sched_group_rcu);
7263 void sched_offline_group(struct task_group *tg)
7265 unsigned long flags;
7268 /* end participation in shares distribution */
7269 for_each_possible_cpu(i)
7270 unregister_fair_sched_group(tg, i);
7272 spin_lock_irqsave(&task_group_lock, flags);
7273 list_del_rcu(&tg->list);
7274 list_del_rcu(&tg->siblings);
7275 spin_unlock_irqrestore(&task_group_lock, flags);
7278 /* change task's runqueue when it moves between groups.
7279 * The caller of this function should have put the task in its new group
7280 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7281 * reflect its new group.
7283 void sched_move_task(struct task_struct *tsk)
7285 struct task_group *tg;
7287 unsigned long flags;
7290 rq = task_rq_lock(tsk, &flags);
7292 running = task_current(rq, tsk);
7296 dequeue_task(rq, tsk, 0);
7297 if (unlikely(running))
7298 tsk->sched_class->put_prev_task(rq, tsk);
7300 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7301 lockdep_is_held(&tsk->sighand->siglock)),
7302 struct task_group, css);
7303 tg = autogroup_task_group(tsk, tg);
7304 tsk->sched_task_group = tg;
7306 #ifdef CONFIG_FAIR_GROUP_SCHED
7307 if (tsk->sched_class->task_move_group)
7308 tsk->sched_class->task_move_group(tsk, on_rq);
7311 set_task_rq(tsk, task_cpu(tsk));
7313 if (unlikely(running))
7314 tsk->sched_class->set_curr_task(rq);
7316 enqueue_task(rq, tsk, 0);
7318 task_rq_unlock(rq, tsk, &flags);
7320 #endif /* CONFIG_CGROUP_SCHED */
7322 #ifdef CONFIG_RT_GROUP_SCHED
7324 * Ensure that the real time constraints are schedulable.
7326 static DEFINE_MUTEX(rt_constraints_mutex);
7328 /* Must be called with tasklist_lock held */
7329 static inline int tg_has_rt_tasks(struct task_group *tg)
7331 struct task_struct *g, *p;
7333 do_each_thread(g, p) {
7334 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7336 } while_each_thread(g, p);
7341 struct rt_schedulable_data {
7342 struct task_group *tg;
7347 static int tg_rt_schedulable(struct task_group *tg, void *data)
7349 struct rt_schedulable_data *d = data;
7350 struct task_group *child;
7351 unsigned long total, sum = 0;
7352 u64 period, runtime;
7354 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7355 runtime = tg->rt_bandwidth.rt_runtime;
7358 period = d->rt_period;
7359 runtime = d->rt_runtime;
7363 * Cannot have more runtime than the period.
7365 if (runtime > period && runtime != RUNTIME_INF)
7369 * Ensure we don't starve existing RT tasks.
7371 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7374 total = to_ratio(period, runtime);
7377 * Nobody can have more than the global setting allows.
7379 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7383 * The sum of our children's runtime should not exceed our own.
7385 list_for_each_entry_rcu(child, &tg->children, siblings) {
7386 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7387 runtime = child->rt_bandwidth.rt_runtime;
7389 if (child == d->tg) {
7390 period = d->rt_period;
7391 runtime = d->rt_runtime;
7394 sum += to_ratio(period, runtime);
7403 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7407 struct rt_schedulable_data data = {
7409 .rt_period = period,
7410 .rt_runtime = runtime,
7414 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7420 static int tg_set_rt_bandwidth(struct task_group *tg,
7421 u64 rt_period, u64 rt_runtime)
7425 mutex_lock(&rt_constraints_mutex);
7426 read_lock(&tasklist_lock);
7427 err = __rt_schedulable(tg, rt_period, rt_runtime);
7431 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7432 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7433 tg->rt_bandwidth.rt_runtime = rt_runtime;
7435 for_each_possible_cpu(i) {
7436 struct rt_rq *rt_rq = tg->rt_rq[i];
7438 raw_spin_lock(&rt_rq->rt_runtime_lock);
7439 rt_rq->rt_runtime = rt_runtime;
7440 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7442 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7444 read_unlock(&tasklist_lock);
7445 mutex_unlock(&rt_constraints_mutex);
7450 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7452 u64 rt_runtime, rt_period;
7454 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7455 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7456 if (rt_runtime_us < 0)
7457 rt_runtime = RUNTIME_INF;
7459 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7462 static long sched_group_rt_runtime(struct task_group *tg)
7466 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7469 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7470 do_div(rt_runtime_us, NSEC_PER_USEC);
7471 return rt_runtime_us;
7474 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7476 u64 rt_runtime, rt_period;
7478 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7479 rt_runtime = tg->rt_bandwidth.rt_runtime;
7484 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7487 static long sched_group_rt_period(struct task_group *tg)
7491 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7492 do_div(rt_period_us, NSEC_PER_USEC);
7493 return rt_period_us;
7495 #endif /* CONFIG_RT_GROUP_SCHED */
7497 #ifdef CONFIG_RT_GROUP_SCHED
7498 static int sched_rt_global_constraints(void)
7502 mutex_lock(&rt_constraints_mutex);
7503 read_lock(&tasklist_lock);
7504 ret = __rt_schedulable(NULL, 0, 0);
7505 read_unlock(&tasklist_lock);
7506 mutex_unlock(&rt_constraints_mutex);
7511 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7513 /* Don't accept realtime tasks when there is no way for them to run */
7514 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7520 #else /* !CONFIG_RT_GROUP_SCHED */
7521 static int sched_rt_global_constraints(void)
7523 unsigned long flags;
7526 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7527 for_each_possible_cpu(i) {
7528 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7530 raw_spin_lock(&rt_rq->rt_runtime_lock);
7531 rt_rq->rt_runtime = global_rt_runtime();
7532 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7534 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7538 #endif /* CONFIG_RT_GROUP_SCHED */
7540 static int sched_dl_global_constraints(void)
7542 u64 runtime = global_rt_runtime();
7543 u64 period = global_rt_period();
7544 u64 new_bw = to_ratio(period, runtime);
7546 unsigned long flags;
7549 * Here we want to check the bandwidth not being set to some
7550 * value smaller than the currently allocated bandwidth in
7551 * any of the root_domains.
7553 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7554 * cycling on root_domains... Discussion on different/better
7555 * solutions is welcome!
7557 for_each_possible_cpu(cpu) {
7558 struct dl_bw *dl_b = dl_bw_of(cpu);
7560 raw_spin_lock_irqsave(&dl_b->lock, flags);
7561 if (new_bw < dl_b->total_bw)
7563 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7572 static void sched_dl_do_global(void)
7576 unsigned long flags;
7578 def_dl_bandwidth.dl_period = global_rt_period();
7579 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7581 if (global_rt_runtime() != RUNTIME_INF)
7582 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7585 * FIXME: As above...
7587 for_each_possible_cpu(cpu) {
7588 struct dl_bw *dl_b = dl_bw_of(cpu);
7590 raw_spin_lock_irqsave(&dl_b->lock, flags);
7592 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7596 static int sched_rt_global_validate(void)
7598 if (sysctl_sched_rt_period <= 0)
7601 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7602 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7608 static void sched_rt_do_global(void)
7610 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7611 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7614 int sched_rt_handler(struct ctl_table *table, int write,
7615 void __user *buffer, size_t *lenp,
7618 int old_period, old_runtime;
7619 static DEFINE_MUTEX(mutex);
7623 old_period = sysctl_sched_rt_period;
7624 old_runtime = sysctl_sched_rt_runtime;
7626 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7628 if (!ret && write) {
7629 ret = sched_rt_global_validate();
7633 ret = sched_rt_global_constraints();
7637 ret = sched_dl_global_constraints();
7641 sched_rt_do_global();
7642 sched_dl_do_global();
7646 sysctl_sched_rt_period = old_period;
7647 sysctl_sched_rt_runtime = old_runtime;
7649 mutex_unlock(&mutex);
7654 int sched_rr_handler(struct ctl_table *table, int write,
7655 void __user *buffer, size_t *lenp,
7659 static DEFINE_MUTEX(mutex);
7662 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7663 /* make sure that internally we keep jiffies */
7664 /* also, writing zero resets timeslice to default */
7665 if (!ret && write) {
7666 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7667 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7669 mutex_unlock(&mutex);
7673 #ifdef CONFIG_CGROUP_SCHED
7675 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7677 return css ? container_of(css, struct task_group, css) : NULL;
7680 static struct cgroup_subsys_state *
7681 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7683 struct task_group *parent = css_tg(parent_css);
7684 struct task_group *tg;
7687 /* This is early initialization for the top cgroup */
7688 return &root_task_group.css;
7691 tg = sched_create_group(parent);
7693 return ERR_PTR(-ENOMEM);
7698 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7700 struct task_group *tg = css_tg(css);
7701 struct task_group *parent = css_tg(css->parent);
7704 sched_online_group(tg, parent);
7708 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7710 struct task_group *tg = css_tg(css);
7712 sched_destroy_group(tg);
7715 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7717 struct task_group *tg = css_tg(css);
7719 sched_offline_group(tg);
7722 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7723 struct cgroup_taskset *tset)
7725 struct task_struct *task;
7727 cgroup_taskset_for_each(task, tset) {
7728 #ifdef CONFIG_RT_GROUP_SCHED
7729 if (!sched_rt_can_attach(css_tg(css), task))
7732 /* We don't support RT-tasks being in separate groups */
7733 if (task->sched_class != &fair_sched_class)
7740 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7741 struct cgroup_taskset *tset)
7743 struct task_struct *task;
7745 cgroup_taskset_for_each(task, tset)
7746 sched_move_task(task);
7749 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7750 struct cgroup_subsys_state *old_css,
7751 struct task_struct *task)
7754 * cgroup_exit() is called in the copy_process() failure path.
7755 * Ignore this case since the task hasn't ran yet, this avoids
7756 * trying to poke a half freed task state from generic code.
7758 if (!(task->flags & PF_EXITING))
7761 sched_move_task(task);
7764 #ifdef CONFIG_FAIR_GROUP_SCHED
7765 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7766 struct cftype *cftype, u64 shareval)
7768 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7771 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7774 struct task_group *tg = css_tg(css);
7776 return (u64) scale_load_down(tg->shares);
7779 #ifdef CONFIG_CFS_BANDWIDTH
7780 static DEFINE_MUTEX(cfs_constraints_mutex);
7782 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7783 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7785 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7787 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7789 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7790 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7792 if (tg == &root_task_group)
7796 * Ensure we have at some amount of bandwidth every period. This is
7797 * to prevent reaching a state of large arrears when throttled via
7798 * entity_tick() resulting in prolonged exit starvation.
7800 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7804 * Likewise, bound things on the otherside by preventing insane quota
7805 * periods. This also allows us to normalize in computing quota
7808 if (period > max_cfs_quota_period)
7811 mutex_lock(&cfs_constraints_mutex);
7812 ret = __cfs_schedulable(tg, period, quota);
7816 runtime_enabled = quota != RUNTIME_INF;
7817 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7819 * If we need to toggle cfs_bandwidth_used, off->on must occur
7820 * before making related changes, and on->off must occur afterwards
7822 if (runtime_enabled && !runtime_was_enabled)
7823 cfs_bandwidth_usage_inc();
7824 raw_spin_lock_irq(&cfs_b->lock);
7825 cfs_b->period = ns_to_ktime(period);
7826 cfs_b->quota = quota;
7828 __refill_cfs_bandwidth_runtime(cfs_b);
7829 /* restart the period timer (if active) to handle new period expiry */
7830 if (runtime_enabled && cfs_b->timer_active) {
7831 /* force a reprogram */
7832 __start_cfs_bandwidth(cfs_b, true);
7834 raw_spin_unlock_irq(&cfs_b->lock);
7836 for_each_possible_cpu(i) {
7837 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7838 struct rq *rq = cfs_rq->rq;
7840 raw_spin_lock_irq(&rq->lock);
7841 cfs_rq->runtime_enabled = runtime_enabled;
7842 cfs_rq->runtime_remaining = 0;
7844 if (cfs_rq->throttled)
7845 unthrottle_cfs_rq(cfs_rq);
7846 raw_spin_unlock_irq(&rq->lock);
7848 if (runtime_was_enabled && !runtime_enabled)
7849 cfs_bandwidth_usage_dec();
7851 mutex_unlock(&cfs_constraints_mutex);
7856 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7860 period = ktime_to_ns(tg->cfs_bandwidth.period);
7861 if (cfs_quota_us < 0)
7862 quota = RUNTIME_INF;
7864 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7866 return tg_set_cfs_bandwidth(tg, period, quota);
7869 long tg_get_cfs_quota(struct task_group *tg)
7873 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7876 quota_us = tg->cfs_bandwidth.quota;
7877 do_div(quota_us, NSEC_PER_USEC);
7882 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7886 period = (u64)cfs_period_us * NSEC_PER_USEC;
7887 quota = tg->cfs_bandwidth.quota;
7889 return tg_set_cfs_bandwidth(tg, period, quota);
7892 long tg_get_cfs_period(struct task_group *tg)
7896 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7897 do_div(cfs_period_us, NSEC_PER_USEC);
7899 return cfs_period_us;
7902 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7905 return tg_get_cfs_quota(css_tg(css));
7908 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7909 struct cftype *cftype, s64 cfs_quota_us)
7911 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7914 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7917 return tg_get_cfs_period(css_tg(css));
7920 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7921 struct cftype *cftype, u64 cfs_period_us)
7923 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7926 struct cfs_schedulable_data {
7927 struct task_group *tg;
7932 * normalize group quota/period to be quota/max_period
7933 * note: units are usecs
7935 static u64 normalize_cfs_quota(struct task_group *tg,
7936 struct cfs_schedulable_data *d)
7944 period = tg_get_cfs_period(tg);
7945 quota = tg_get_cfs_quota(tg);
7948 /* note: these should typically be equivalent */
7949 if (quota == RUNTIME_INF || quota == -1)
7952 return to_ratio(period, quota);
7955 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7957 struct cfs_schedulable_data *d = data;
7958 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7959 s64 quota = 0, parent_quota = -1;
7962 quota = RUNTIME_INF;
7964 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7966 quota = normalize_cfs_quota(tg, d);
7967 parent_quota = parent_b->hierarchal_quota;
7970 * ensure max(child_quota) <= parent_quota, inherit when no
7973 if (quota == RUNTIME_INF)
7974 quota = parent_quota;
7975 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7978 cfs_b->hierarchal_quota = quota;
7983 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7986 struct cfs_schedulable_data data = {
7992 if (quota != RUNTIME_INF) {
7993 do_div(data.period, NSEC_PER_USEC);
7994 do_div(data.quota, NSEC_PER_USEC);
7998 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8004 static int cpu_stats_show(struct seq_file *sf, void *v)
8006 struct task_group *tg = css_tg(seq_css(sf));
8007 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8009 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8010 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8011 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8015 #endif /* CONFIG_CFS_BANDWIDTH */
8016 #endif /* CONFIG_FAIR_GROUP_SCHED */
8018 #ifdef CONFIG_RT_GROUP_SCHED
8019 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8020 struct cftype *cft, s64 val)
8022 return sched_group_set_rt_runtime(css_tg(css), val);
8025 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8028 return sched_group_rt_runtime(css_tg(css));
8031 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8032 struct cftype *cftype, u64 rt_period_us)
8034 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8037 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8040 return sched_group_rt_period(css_tg(css));
8042 #endif /* CONFIG_RT_GROUP_SCHED */
8044 static struct cftype cpu_files[] = {
8045 #ifdef CONFIG_FAIR_GROUP_SCHED
8048 .read_u64 = cpu_shares_read_u64,
8049 .write_u64 = cpu_shares_write_u64,
8052 #ifdef CONFIG_CFS_BANDWIDTH
8054 .name = "cfs_quota_us",
8055 .read_s64 = cpu_cfs_quota_read_s64,
8056 .write_s64 = cpu_cfs_quota_write_s64,
8059 .name = "cfs_period_us",
8060 .read_u64 = cpu_cfs_period_read_u64,
8061 .write_u64 = cpu_cfs_period_write_u64,
8065 .seq_show = cpu_stats_show,
8068 #ifdef CONFIG_RT_GROUP_SCHED
8070 .name = "rt_runtime_us",
8071 .read_s64 = cpu_rt_runtime_read,
8072 .write_s64 = cpu_rt_runtime_write,
8075 .name = "rt_period_us",
8076 .read_u64 = cpu_rt_period_read_uint,
8077 .write_u64 = cpu_rt_period_write_uint,
8083 struct cgroup_subsys cpu_cgrp_subsys = {
8084 .css_alloc = cpu_cgroup_css_alloc,
8085 .css_free = cpu_cgroup_css_free,
8086 .css_online = cpu_cgroup_css_online,
8087 .css_offline = cpu_cgroup_css_offline,
8088 .can_attach = cpu_cgroup_can_attach,
8089 .attach = cpu_cgroup_attach,
8090 .exit = cpu_cgroup_exit,
8091 .base_cftypes = cpu_files,
8095 #endif /* CONFIG_CGROUP_SCHED */
8097 void dump_cpu_task(int cpu)
8099 pr_info("Task dump for CPU %d:\n", cpu);
8100 sched_show_task(cpu_curr(cpu));