4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
96 ktime_t soft, hard, now;
99 if (hrtimer_active(period_timer))
102 now = hrtimer_cb_get_time(period_timer);
103 hrtimer_forward(period_timer, now, period);
105 soft = hrtimer_get_softexpires(period_timer);
106 hard = hrtimer_get_expires(period_timer);
107 delta = ktime_to_ns(ktime_sub(hard, soft));
108 __hrtimer_start_range_ns(period_timer, soft, delta,
109 HRTIMER_MODE_ABS_PINNED, 0);
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
118 void update_rq_clock(struct rq *rq)
122 lockdep_assert_held(&rq->lock);
124 if (rq->clock_skip_update & RQCF_ACT_SKIP)
127 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
131 update_rq_clock_task(rq, delta);
135 * Debugging: various feature bits
138 #define SCHED_FEAT(name, enabled) \
139 (1UL << __SCHED_FEAT_##name) * enabled |
141 const_debug unsigned int sysctl_sched_features =
142 #include "features.h"
147 #ifdef CONFIG_SCHED_DEBUG
148 #define SCHED_FEAT(name, enabled) \
151 static const char * const sched_feat_names[] = {
152 #include "features.h"
157 static int sched_feat_show(struct seq_file *m, void *v)
161 for (i = 0; i < __SCHED_FEAT_NR; i++) {
162 if (!(sysctl_sched_features & (1UL << i)))
164 seq_printf(m, "%s ", sched_feat_names[i]);
171 #ifdef HAVE_JUMP_LABEL
173 #define jump_label_key__true STATIC_KEY_INIT_TRUE
174 #define jump_label_key__false STATIC_KEY_INIT_FALSE
176 #define SCHED_FEAT(name, enabled) \
177 jump_label_key__##enabled ,
179 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
180 #include "features.h"
185 static void sched_feat_disable(int i)
187 if (static_key_enabled(&sched_feat_keys[i]))
188 static_key_slow_dec(&sched_feat_keys[i]);
191 static void sched_feat_enable(int i)
193 if (!static_key_enabled(&sched_feat_keys[i]))
194 static_key_slow_inc(&sched_feat_keys[i]);
197 static void sched_feat_disable(int i) { };
198 static void sched_feat_enable(int i) { };
199 #endif /* HAVE_JUMP_LABEL */
201 static int sched_feat_set(char *cmp)
206 if (strncmp(cmp, "NO_", 3) == 0) {
211 for (i = 0; i < __SCHED_FEAT_NR; i++) {
212 if (strcmp(cmp, sched_feat_names[i]) == 0) {
214 sysctl_sched_features &= ~(1UL << i);
215 sched_feat_disable(i);
217 sysctl_sched_features |= (1UL << i);
218 sched_feat_enable(i);
228 sched_feat_write(struct file *filp, const char __user *ubuf,
229 size_t cnt, loff_t *ppos)
239 if (copy_from_user(&buf, ubuf, cnt))
245 /* Ensure the static_key remains in a consistent state */
246 inode = file_inode(filp);
247 mutex_lock(&inode->i_mutex);
248 i = sched_feat_set(cmp);
249 mutex_unlock(&inode->i_mutex);
250 if (i == __SCHED_FEAT_NR)
258 static int sched_feat_open(struct inode *inode, struct file *filp)
260 return single_open(filp, sched_feat_show, NULL);
263 static const struct file_operations sched_feat_fops = {
264 .open = sched_feat_open,
265 .write = sched_feat_write,
268 .release = single_release,
271 static __init int sched_init_debug(void)
273 debugfs_create_file("sched_features", 0644, NULL, NULL,
278 late_initcall(sched_init_debug);
279 #endif /* CONFIG_SCHED_DEBUG */
282 * Number of tasks to iterate in a single balance run.
283 * Limited because this is done with IRQs disabled.
285 const_debug unsigned int sysctl_sched_nr_migrate = 32;
288 * period over which we average the RT time consumption, measured
293 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
296 * period over which we measure -rt task cpu usage in us.
299 unsigned int sysctl_sched_rt_period = 1000000;
301 __read_mostly int scheduler_running;
304 * part of the period that we allow rt tasks to run in us.
307 int sysctl_sched_rt_runtime = 950000;
309 /* cpus with isolated domains */
310 cpumask_var_t cpu_isolated_map;
313 * this_rq_lock - lock this runqueue and disable interrupts.
315 static struct rq *this_rq_lock(void)
322 raw_spin_lock(&rq->lock);
327 #ifdef CONFIG_SCHED_HRTICK
329 * Use HR-timers to deliver accurate preemption points.
332 static void hrtick_clear(struct rq *rq)
334 if (hrtimer_active(&rq->hrtick_timer))
335 hrtimer_cancel(&rq->hrtick_timer);
339 * High-resolution timer tick.
340 * Runs from hardirq context with interrupts disabled.
342 static enum hrtimer_restart hrtick(struct hrtimer *timer)
344 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
346 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
348 raw_spin_lock(&rq->lock);
350 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
351 raw_spin_unlock(&rq->lock);
353 return HRTIMER_NORESTART;
358 static int __hrtick_restart(struct rq *rq)
360 struct hrtimer *timer = &rq->hrtick_timer;
361 ktime_t time = hrtimer_get_softexpires(timer);
363 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
367 * called from hardirq (IPI) context
369 static void __hrtick_start(void *arg)
373 raw_spin_lock(&rq->lock);
374 __hrtick_restart(rq);
375 rq->hrtick_csd_pending = 0;
376 raw_spin_unlock(&rq->lock);
380 * Called to set the hrtick timer state.
382 * called with rq->lock held and irqs disabled
384 void hrtick_start(struct rq *rq, u64 delay)
386 struct hrtimer *timer = &rq->hrtick_timer;
391 * Don't schedule slices shorter than 10000ns, that just
392 * doesn't make sense and can cause timer DoS.
394 delta = max_t(s64, delay, 10000LL);
395 time = ktime_add_ns(timer->base->get_time(), delta);
397 hrtimer_set_expires(timer, time);
399 if (rq == this_rq()) {
400 __hrtick_restart(rq);
401 } else if (!rq->hrtick_csd_pending) {
402 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
403 rq->hrtick_csd_pending = 1;
408 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
410 int cpu = (int)(long)hcpu;
413 case CPU_UP_CANCELED:
414 case CPU_UP_CANCELED_FROZEN:
415 case CPU_DOWN_PREPARE:
416 case CPU_DOWN_PREPARE_FROZEN:
418 case CPU_DEAD_FROZEN:
419 hrtick_clear(cpu_rq(cpu));
426 static __init void init_hrtick(void)
428 hotcpu_notifier(hotplug_hrtick, 0);
432 * Called to set the hrtick timer state.
434 * called with rq->lock held and irqs disabled
436 void hrtick_start(struct rq *rq, u64 delay)
439 * Don't schedule slices shorter than 10000ns, that just
440 * doesn't make sense. Rely on vruntime for fairness.
442 delay = max_t(u64, delay, 10000LL);
443 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
444 HRTIMER_MODE_REL_PINNED, 0);
447 static inline void init_hrtick(void)
450 #endif /* CONFIG_SMP */
452 static void init_rq_hrtick(struct rq *rq)
455 rq->hrtick_csd_pending = 0;
457 rq->hrtick_csd.flags = 0;
458 rq->hrtick_csd.func = __hrtick_start;
459 rq->hrtick_csd.info = rq;
462 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
463 rq->hrtick_timer.function = hrtick;
465 #else /* CONFIG_SCHED_HRTICK */
466 static inline void hrtick_clear(struct rq *rq)
470 static inline void init_rq_hrtick(struct rq *rq)
474 static inline void init_hrtick(void)
477 #endif /* CONFIG_SCHED_HRTICK */
480 * cmpxchg based fetch_or, macro so it works for different integer types
482 #define fetch_or(ptr, val) \
483 ({ typeof(*(ptr)) __old, __val = *(ptr); \
485 __old = cmpxchg((ptr), __val, __val | (val)); \
486 if (__old == __val) \
493 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
495 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
496 * this avoids any races wrt polling state changes and thereby avoids
499 static bool set_nr_and_not_polling(struct task_struct *p)
501 struct thread_info *ti = task_thread_info(p);
502 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
506 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
508 * If this returns true, then the idle task promises to call
509 * sched_ttwu_pending() and reschedule soon.
511 static bool set_nr_if_polling(struct task_struct *p)
513 struct thread_info *ti = task_thread_info(p);
514 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
517 if (!(val & _TIF_POLLING_NRFLAG))
519 if (val & _TIF_NEED_RESCHED)
521 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
530 static bool set_nr_and_not_polling(struct task_struct *p)
532 set_tsk_need_resched(p);
537 static bool set_nr_if_polling(struct task_struct *p)
545 * resched_curr - mark rq's current task 'to be rescheduled now'.
547 * On UP this means the setting of the need_resched flag, on SMP it
548 * might also involve a cross-CPU call to trigger the scheduler on
551 void resched_curr(struct rq *rq)
553 struct task_struct *curr = rq->curr;
556 lockdep_assert_held(&rq->lock);
558 if (test_tsk_need_resched(curr))
563 if (cpu == smp_processor_id()) {
564 set_tsk_need_resched(curr);
565 set_preempt_need_resched();
569 if (set_nr_and_not_polling(curr))
570 smp_send_reschedule(cpu);
572 trace_sched_wake_idle_without_ipi(cpu);
575 void resched_cpu(int cpu)
577 struct rq *rq = cpu_rq(cpu);
580 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
583 raw_spin_unlock_irqrestore(&rq->lock, flags);
587 #ifdef CONFIG_NO_HZ_COMMON
589 * In the semi idle case, use the nearest busy cpu for migrating timers
590 * from an idle cpu. This is good for power-savings.
592 * We don't do similar optimization for completely idle system, as
593 * selecting an idle cpu will add more delays to the timers than intended
594 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
596 int get_nohz_timer_target(int pinned)
598 int cpu = smp_processor_id();
600 struct sched_domain *sd;
602 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
606 for_each_domain(cpu, sd) {
607 for_each_cpu(i, sched_domain_span(sd)) {
619 * When add_timer_on() enqueues a timer into the timer wheel of an
620 * idle CPU then this timer might expire before the next timer event
621 * which is scheduled to wake up that CPU. In case of a completely
622 * idle system the next event might even be infinite time into the
623 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
624 * leaves the inner idle loop so the newly added timer is taken into
625 * account when the CPU goes back to idle and evaluates the timer
626 * wheel for the next timer event.
628 static void wake_up_idle_cpu(int cpu)
630 struct rq *rq = cpu_rq(cpu);
632 if (cpu == smp_processor_id())
635 if (set_nr_and_not_polling(rq->idle))
636 smp_send_reschedule(cpu);
638 trace_sched_wake_idle_without_ipi(cpu);
641 static bool wake_up_full_nohz_cpu(int cpu)
644 * We just need the target to call irq_exit() and re-evaluate
645 * the next tick. The nohz full kick at least implies that.
646 * If needed we can still optimize that later with an
649 if (tick_nohz_full_cpu(cpu)) {
650 if (cpu != smp_processor_id() ||
651 tick_nohz_tick_stopped())
652 tick_nohz_full_kick_cpu(cpu);
659 void wake_up_nohz_cpu(int cpu)
661 if (!wake_up_full_nohz_cpu(cpu))
662 wake_up_idle_cpu(cpu);
665 static inline bool got_nohz_idle_kick(void)
667 int cpu = smp_processor_id();
669 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
672 if (idle_cpu(cpu) && !need_resched())
676 * We can't run Idle Load Balance on this CPU for this time so we
677 * cancel it and clear NOHZ_BALANCE_KICK
679 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
683 #else /* CONFIG_NO_HZ_COMMON */
685 static inline bool got_nohz_idle_kick(void)
690 #endif /* CONFIG_NO_HZ_COMMON */
692 #ifdef CONFIG_NO_HZ_FULL
693 bool sched_can_stop_tick(void)
696 * FIFO realtime policy runs the highest priority task. Other runnable
697 * tasks are of a lower priority. The scheduler tick does nothing.
699 if (current->policy == SCHED_FIFO)
703 * Round-robin realtime tasks time slice with other tasks at the same
704 * realtime priority. Is this task the only one at this priority?
706 if (current->policy == SCHED_RR) {
707 struct sched_rt_entity *rt_se = ¤t->rt;
709 return rt_se->run_list.prev == rt_se->run_list.next;
713 * More than one running task need preemption.
714 * nr_running update is assumed to be visible
715 * after IPI is sent from wakers.
717 if (this_rq()->nr_running > 1)
722 #endif /* CONFIG_NO_HZ_FULL */
724 void sched_avg_update(struct rq *rq)
726 s64 period = sched_avg_period();
728 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
730 * Inline assembly required to prevent the compiler
731 * optimising this loop into a divmod call.
732 * See __iter_div_u64_rem() for another example of this.
734 asm("" : "+rm" (rq->age_stamp));
735 rq->age_stamp += period;
740 #endif /* CONFIG_SMP */
742 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
743 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
745 * Iterate task_group tree rooted at *from, calling @down when first entering a
746 * node and @up when leaving it for the final time.
748 * Caller must hold rcu_lock or sufficient equivalent.
750 int walk_tg_tree_from(struct task_group *from,
751 tg_visitor down, tg_visitor up, void *data)
753 struct task_group *parent, *child;
759 ret = (*down)(parent, data);
762 list_for_each_entry_rcu(child, &parent->children, siblings) {
769 ret = (*up)(parent, data);
770 if (ret || parent == from)
774 parent = parent->parent;
781 int tg_nop(struct task_group *tg, void *data)
787 static void set_load_weight(struct task_struct *p)
789 int prio = p->static_prio - MAX_RT_PRIO;
790 struct load_weight *load = &p->se.load;
793 * SCHED_IDLE tasks get minimal weight:
795 if (p->policy == SCHED_IDLE) {
796 load->weight = scale_load(WEIGHT_IDLEPRIO);
797 load->inv_weight = WMULT_IDLEPRIO;
801 load->weight = scale_load(prio_to_weight[prio]);
802 load->inv_weight = prio_to_wmult[prio];
805 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
808 sched_info_queued(rq, p);
809 p->sched_class->enqueue_task(rq, p, flags);
812 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
815 sched_info_dequeued(rq, p);
816 p->sched_class->dequeue_task(rq, p, flags);
819 void activate_task(struct rq *rq, struct task_struct *p, int flags)
821 if (task_contributes_to_load(p))
822 rq->nr_uninterruptible--;
824 enqueue_task(rq, p, flags);
827 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
829 if (task_contributes_to_load(p))
830 rq->nr_uninterruptible++;
832 dequeue_task(rq, p, flags);
835 static void update_rq_clock_task(struct rq *rq, s64 delta)
838 * In theory, the compile should just see 0 here, and optimize out the call
839 * to sched_rt_avg_update. But I don't trust it...
841 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
842 s64 steal = 0, irq_delta = 0;
844 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
845 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
848 * Since irq_time is only updated on {soft,}irq_exit, we might run into
849 * this case when a previous update_rq_clock() happened inside a
852 * When this happens, we stop ->clock_task and only update the
853 * prev_irq_time stamp to account for the part that fit, so that a next
854 * update will consume the rest. This ensures ->clock_task is
857 * It does however cause some slight miss-attribution of {soft,}irq
858 * time, a more accurate solution would be to update the irq_time using
859 * the current rq->clock timestamp, except that would require using
862 if (irq_delta > delta)
865 rq->prev_irq_time += irq_delta;
868 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
869 if (static_key_false((¶virt_steal_rq_enabled))) {
870 steal = paravirt_steal_clock(cpu_of(rq));
871 steal -= rq->prev_steal_time_rq;
873 if (unlikely(steal > delta))
876 rq->prev_steal_time_rq += steal;
881 rq->clock_task += delta;
883 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
884 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
885 sched_rt_avg_update(rq, irq_delta + steal);
889 void sched_set_stop_task(int cpu, struct task_struct *stop)
891 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
892 struct task_struct *old_stop = cpu_rq(cpu)->stop;
896 * Make it appear like a SCHED_FIFO task, its something
897 * userspace knows about and won't get confused about.
899 * Also, it will make PI more or less work without too
900 * much confusion -- but then, stop work should not
901 * rely on PI working anyway.
903 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
905 stop->sched_class = &stop_sched_class;
908 cpu_rq(cpu)->stop = stop;
912 * Reset it back to a normal scheduling class so that
913 * it can die in pieces.
915 old_stop->sched_class = &rt_sched_class;
920 * __normal_prio - return the priority that is based on the static prio
922 static inline int __normal_prio(struct task_struct *p)
924 return p->static_prio;
928 * Calculate the expected normal priority: i.e. priority
929 * without taking RT-inheritance into account. Might be
930 * boosted by interactivity modifiers. Changes upon fork,
931 * setprio syscalls, and whenever the interactivity
932 * estimator recalculates.
934 static inline int normal_prio(struct task_struct *p)
938 if (task_has_dl_policy(p))
939 prio = MAX_DL_PRIO-1;
940 else if (task_has_rt_policy(p))
941 prio = MAX_RT_PRIO-1 - p->rt_priority;
943 prio = __normal_prio(p);
948 * Calculate the current priority, i.e. the priority
949 * taken into account by the scheduler. This value might
950 * be boosted by RT tasks, or might be boosted by
951 * interactivity modifiers. Will be RT if the task got
952 * RT-boosted. If not then it returns p->normal_prio.
954 static int effective_prio(struct task_struct *p)
956 p->normal_prio = normal_prio(p);
958 * If we are RT tasks or we were boosted to RT priority,
959 * keep the priority unchanged. Otherwise, update priority
960 * to the normal priority:
962 if (!rt_prio(p->prio))
963 return p->normal_prio;
968 * task_curr - is this task currently executing on a CPU?
969 * @p: the task in question.
971 * Return: 1 if the task is currently executing. 0 otherwise.
973 inline int task_curr(const struct task_struct *p)
975 return cpu_curr(task_cpu(p)) == p;
979 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
981 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
982 const struct sched_class *prev_class,
985 if (prev_class != p->sched_class) {
986 if (prev_class->switched_from)
987 prev_class->switched_from(rq, p);
988 /* Possble rq->lock 'hole'. */
989 p->sched_class->switched_to(rq, p);
990 } else if (oldprio != p->prio || dl_task(p))
991 p->sched_class->prio_changed(rq, p, oldprio);
994 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
996 const struct sched_class *class;
998 if (p->sched_class == rq->curr->sched_class) {
999 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1001 for_each_class(class) {
1002 if (class == rq->curr->sched_class)
1004 if (class == p->sched_class) {
1012 * A queue event has occurred, and we're going to schedule. In
1013 * this case, we can save a useless back to back clock update.
1015 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1016 rq_clock_skip_update(rq, true);
1020 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1022 #ifdef CONFIG_SCHED_DEBUG
1024 * We should never call set_task_cpu() on a blocked task,
1025 * ttwu() will sort out the placement.
1027 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1030 #ifdef CONFIG_LOCKDEP
1032 * The caller should hold either p->pi_lock or rq->lock, when changing
1033 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1035 * sched_move_task() holds both and thus holding either pins the cgroup,
1038 * Furthermore, all task_rq users should acquire both locks, see
1041 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1042 lockdep_is_held(&task_rq(p)->lock)));
1046 trace_sched_migrate_task(p, new_cpu);
1048 if (task_cpu(p) != new_cpu) {
1049 if (p->sched_class->migrate_task_rq)
1050 p->sched_class->migrate_task_rq(p, new_cpu);
1051 p->se.nr_migrations++;
1052 perf_sw_event_sched(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 0);
1055 __set_task_cpu(p, new_cpu);
1058 static void __migrate_swap_task(struct task_struct *p, int cpu)
1060 if (task_on_rq_queued(p)) {
1061 struct rq *src_rq, *dst_rq;
1063 src_rq = task_rq(p);
1064 dst_rq = cpu_rq(cpu);
1066 deactivate_task(src_rq, p, 0);
1067 set_task_cpu(p, cpu);
1068 activate_task(dst_rq, p, 0);
1069 check_preempt_curr(dst_rq, p, 0);
1072 * Task isn't running anymore; make it appear like we migrated
1073 * it before it went to sleep. This means on wakeup we make the
1074 * previous cpu our targer instead of where it really is.
1080 struct migration_swap_arg {
1081 struct task_struct *src_task, *dst_task;
1082 int src_cpu, dst_cpu;
1085 static int migrate_swap_stop(void *data)
1087 struct migration_swap_arg *arg = data;
1088 struct rq *src_rq, *dst_rq;
1091 src_rq = cpu_rq(arg->src_cpu);
1092 dst_rq = cpu_rq(arg->dst_cpu);
1094 double_raw_lock(&arg->src_task->pi_lock,
1095 &arg->dst_task->pi_lock);
1096 double_rq_lock(src_rq, dst_rq);
1097 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1100 if (task_cpu(arg->src_task) != arg->src_cpu)
1103 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1106 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1109 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1110 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1115 double_rq_unlock(src_rq, dst_rq);
1116 raw_spin_unlock(&arg->dst_task->pi_lock);
1117 raw_spin_unlock(&arg->src_task->pi_lock);
1123 * Cross migrate two tasks
1125 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1127 struct migration_swap_arg arg;
1130 arg = (struct migration_swap_arg){
1132 .src_cpu = task_cpu(cur),
1134 .dst_cpu = task_cpu(p),
1137 if (arg.src_cpu == arg.dst_cpu)
1141 * These three tests are all lockless; this is OK since all of them
1142 * will be re-checked with proper locks held further down the line.
1144 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1147 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1150 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1153 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1154 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1160 struct migration_arg {
1161 struct task_struct *task;
1165 static int migration_cpu_stop(void *data);
1168 * wait_task_inactive - wait for a thread to unschedule.
1170 * If @match_state is nonzero, it's the @p->state value just checked and
1171 * not expected to change. If it changes, i.e. @p might have woken up,
1172 * then return zero. When we succeed in waiting for @p to be off its CPU,
1173 * we return a positive number (its total switch count). If a second call
1174 * a short while later returns the same number, the caller can be sure that
1175 * @p has remained unscheduled the whole time.
1177 * The caller must ensure that the task *will* unschedule sometime soon,
1178 * else this function might spin for a *long* time. This function can't
1179 * be called with interrupts off, or it may introduce deadlock with
1180 * smp_call_function() if an IPI is sent by the same process we are
1181 * waiting to become inactive.
1183 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1185 unsigned long flags;
1186 int running, queued;
1192 * We do the initial early heuristics without holding
1193 * any task-queue locks at all. We'll only try to get
1194 * the runqueue lock when things look like they will
1200 * If the task is actively running on another CPU
1201 * still, just relax and busy-wait without holding
1204 * NOTE! Since we don't hold any locks, it's not
1205 * even sure that "rq" stays as the right runqueue!
1206 * But we don't care, since "task_running()" will
1207 * return false if the runqueue has changed and p
1208 * is actually now running somewhere else!
1210 while (task_running(rq, p)) {
1211 if (match_state && unlikely(p->state != match_state))
1217 * Ok, time to look more closely! We need the rq
1218 * lock now, to be *sure*. If we're wrong, we'll
1219 * just go back and repeat.
1221 rq = task_rq_lock(p, &flags);
1222 trace_sched_wait_task(p);
1223 running = task_running(rq, p);
1224 queued = task_on_rq_queued(p);
1226 if (!match_state || p->state == match_state)
1227 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1228 task_rq_unlock(rq, p, &flags);
1231 * If it changed from the expected state, bail out now.
1233 if (unlikely(!ncsw))
1237 * Was it really running after all now that we
1238 * checked with the proper locks actually held?
1240 * Oops. Go back and try again..
1242 if (unlikely(running)) {
1248 * It's not enough that it's not actively running,
1249 * it must be off the runqueue _entirely_, and not
1252 * So if it was still runnable (but just not actively
1253 * running right now), it's preempted, and we should
1254 * yield - it could be a while.
1256 if (unlikely(queued)) {
1257 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1259 set_current_state(TASK_UNINTERRUPTIBLE);
1260 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1265 * Ahh, all good. It wasn't running, and it wasn't
1266 * runnable, which means that it will never become
1267 * running in the future either. We're all done!
1276 * kick_process - kick a running thread to enter/exit the kernel
1277 * @p: the to-be-kicked thread
1279 * Cause a process which is running on another CPU to enter
1280 * kernel-mode, without any delay. (to get signals handled.)
1282 * NOTE: this function doesn't have to take the runqueue lock,
1283 * because all it wants to ensure is that the remote task enters
1284 * the kernel. If the IPI races and the task has been migrated
1285 * to another CPU then no harm is done and the purpose has been
1288 void kick_process(struct task_struct *p)
1294 if ((cpu != smp_processor_id()) && task_curr(p))
1295 smp_send_reschedule(cpu);
1298 EXPORT_SYMBOL_GPL(kick_process);
1299 #endif /* CONFIG_SMP */
1303 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1305 static int select_fallback_rq(int cpu, struct task_struct *p)
1307 int nid = cpu_to_node(cpu);
1308 const struct cpumask *nodemask = NULL;
1309 enum { cpuset, possible, fail } state = cpuset;
1313 * If the node that the cpu is on has been offlined, cpu_to_node()
1314 * will return -1. There is no cpu on the node, and we should
1315 * select the cpu on the other node.
1318 nodemask = cpumask_of_node(nid);
1320 /* Look for allowed, online CPU in same node. */
1321 for_each_cpu(dest_cpu, nodemask) {
1322 if (!cpu_online(dest_cpu))
1324 if (!cpu_active(dest_cpu))
1326 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1332 /* Any allowed, online CPU? */
1333 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1334 if (!cpu_online(dest_cpu))
1336 if (!cpu_active(dest_cpu))
1343 /* No more Mr. Nice Guy. */
1344 cpuset_cpus_allowed_fallback(p);
1349 do_set_cpus_allowed(p, cpu_possible_mask);
1360 if (state != cpuset) {
1362 * Don't tell them about moving exiting tasks or
1363 * kernel threads (both mm NULL), since they never
1366 if (p->mm && printk_ratelimit()) {
1367 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1368 task_pid_nr(p), p->comm, cpu);
1376 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1379 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1381 if (p->nr_cpus_allowed > 1)
1382 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1385 * In order not to call set_task_cpu() on a blocking task we need
1386 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1389 * Since this is common to all placement strategies, this lives here.
1391 * [ this allows ->select_task() to simply return task_cpu(p) and
1392 * not worry about this generic constraint ]
1394 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1396 cpu = select_fallback_rq(task_cpu(p), p);
1401 static void update_avg(u64 *avg, u64 sample)
1403 s64 diff = sample - *avg;
1409 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1411 #ifdef CONFIG_SCHEDSTATS
1412 struct rq *rq = this_rq();
1415 int this_cpu = smp_processor_id();
1417 if (cpu == this_cpu) {
1418 schedstat_inc(rq, ttwu_local);
1419 schedstat_inc(p, se.statistics.nr_wakeups_local);
1421 struct sched_domain *sd;
1423 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1425 for_each_domain(this_cpu, sd) {
1426 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1427 schedstat_inc(sd, ttwu_wake_remote);
1434 if (wake_flags & WF_MIGRATED)
1435 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1437 #endif /* CONFIG_SMP */
1439 schedstat_inc(rq, ttwu_count);
1440 schedstat_inc(p, se.statistics.nr_wakeups);
1442 if (wake_flags & WF_SYNC)
1443 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1445 #endif /* CONFIG_SCHEDSTATS */
1448 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1450 activate_task(rq, p, en_flags);
1451 p->on_rq = TASK_ON_RQ_QUEUED;
1453 /* if a worker is waking up, notify workqueue */
1454 if (p->flags & PF_WQ_WORKER)
1455 wq_worker_waking_up(p, cpu_of(rq));
1459 * Mark the task runnable and perform wakeup-preemption.
1462 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1464 check_preempt_curr(rq, p, wake_flags);
1465 trace_sched_wakeup(p, true);
1467 p->state = TASK_RUNNING;
1469 if (p->sched_class->task_woken)
1470 p->sched_class->task_woken(rq, p);
1472 if (rq->idle_stamp) {
1473 u64 delta = rq_clock(rq) - rq->idle_stamp;
1474 u64 max = 2*rq->max_idle_balance_cost;
1476 update_avg(&rq->avg_idle, delta);
1478 if (rq->avg_idle > max)
1487 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1490 if (p->sched_contributes_to_load)
1491 rq->nr_uninterruptible--;
1494 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1495 ttwu_do_wakeup(rq, p, wake_flags);
1499 * Called in case the task @p isn't fully descheduled from its runqueue,
1500 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1501 * since all we need to do is flip p->state to TASK_RUNNING, since
1502 * the task is still ->on_rq.
1504 static int ttwu_remote(struct task_struct *p, int wake_flags)
1509 rq = __task_rq_lock(p);
1510 if (task_on_rq_queued(p)) {
1511 /* check_preempt_curr() may use rq clock */
1512 update_rq_clock(rq);
1513 ttwu_do_wakeup(rq, p, wake_flags);
1516 __task_rq_unlock(rq);
1522 void sched_ttwu_pending(void)
1524 struct rq *rq = this_rq();
1525 struct llist_node *llist = llist_del_all(&rq->wake_list);
1526 struct task_struct *p;
1527 unsigned long flags;
1532 raw_spin_lock_irqsave(&rq->lock, flags);
1535 p = llist_entry(llist, struct task_struct, wake_entry);
1536 llist = llist_next(llist);
1537 ttwu_do_activate(rq, p, 0);
1540 raw_spin_unlock_irqrestore(&rq->lock, flags);
1543 void scheduler_ipi(void)
1546 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1547 * TIF_NEED_RESCHED remotely (for the first time) will also send
1550 preempt_fold_need_resched();
1552 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1556 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1557 * traditionally all their work was done from the interrupt return
1558 * path. Now that we actually do some work, we need to make sure
1561 * Some archs already do call them, luckily irq_enter/exit nest
1564 * Arguably we should visit all archs and update all handlers,
1565 * however a fair share of IPIs are still resched only so this would
1566 * somewhat pessimize the simple resched case.
1569 sched_ttwu_pending();
1572 * Check if someone kicked us for doing the nohz idle load balance.
1574 if (unlikely(got_nohz_idle_kick())) {
1575 this_rq()->idle_balance = 1;
1576 raise_softirq_irqoff(SCHED_SOFTIRQ);
1581 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1583 struct rq *rq = cpu_rq(cpu);
1585 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1586 if (!set_nr_if_polling(rq->idle))
1587 smp_send_reschedule(cpu);
1589 trace_sched_wake_idle_without_ipi(cpu);
1593 void wake_up_if_idle(int cpu)
1595 struct rq *rq = cpu_rq(cpu);
1596 unsigned long flags;
1600 if (!is_idle_task(rcu_dereference(rq->curr)))
1603 if (set_nr_if_polling(rq->idle)) {
1604 trace_sched_wake_idle_without_ipi(cpu);
1606 raw_spin_lock_irqsave(&rq->lock, flags);
1607 if (is_idle_task(rq->curr))
1608 smp_send_reschedule(cpu);
1609 /* Else cpu is not in idle, do nothing here */
1610 raw_spin_unlock_irqrestore(&rq->lock, flags);
1617 bool cpus_share_cache(int this_cpu, int that_cpu)
1619 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1621 #endif /* CONFIG_SMP */
1623 static void ttwu_queue(struct task_struct *p, int cpu)
1625 struct rq *rq = cpu_rq(cpu);
1627 #if defined(CONFIG_SMP)
1628 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1629 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1630 ttwu_queue_remote(p, cpu);
1635 raw_spin_lock(&rq->lock);
1636 ttwu_do_activate(rq, p, 0);
1637 raw_spin_unlock(&rq->lock);
1641 * try_to_wake_up - wake up a thread
1642 * @p: the thread to be awakened
1643 * @state: the mask of task states that can be woken
1644 * @wake_flags: wake modifier flags (WF_*)
1646 * Put it on the run-queue if it's not already there. The "current"
1647 * thread is always on the run-queue (except when the actual
1648 * re-schedule is in progress), and as such you're allowed to do
1649 * the simpler "current->state = TASK_RUNNING" to mark yourself
1650 * runnable without the overhead of this.
1652 * Return: %true if @p was woken up, %false if it was already running.
1653 * or @state didn't match @p's state.
1656 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1658 unsigned long flags;
1659 int cpu, success = 0;
1662 * If we are going to wake up a thread waiting for CONDITION we
1663 * need to ensure that CONDITION=1 done by the caller can not be
1664 * reordered with p->state check below. This pairs with mb() in
1665 * set_current_state() the waiting thread does.
1667 smp_mb__before_spinlock();
1668 raw_spin_lock_irqsave(&p->pi_lock, flags);
1669 if (!(p->state & state))
1672 success = 1; /* we're going to change ->state */
1675 if (p->on_rq && ttwu_remote(p, wake_flags))
1680 * If the owning (remote) cpu is still in the middle of schedule() with
1681 * this task as prev, wait until its done referencing the task.
1686 * Pairs with the smp_wmb() in finish_lock_switch().
1690 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1691 p->state = TASK_WAKING;
1693 if (p->sched_class->task_waking)
1694 p->sched_class->task_waking(p);
1696 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1697 if (task_cpu(p) != cpu) {
1698 wake_flags |= WF_MIGRATED;
1699 set_task_cpu(p, cpu);
1701 #endif /* CONFIG_SMP */
1705 ttwu_stat(p, cpu, wake_flags);
1707 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1713 * try_to_wake_up_local - try to wake up a local task with rq lock held
1714 * @p: the thread to be awakened
1716 * Put @p on the run-queue if it's not already there. The caller must
1717 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1720 static void try_to_wake_up_local(struct task_struct *p)
1722 struct rq *rq = task_rq(p);
1724 if (WARN_ON_ONCE(rq != this_rq()) ||
1725 WARN_ON_ONCE(p == current))
1728 lockdep_assert_held(&rq->lock);
1730 if (!raw_spin_trylock(&p->pi_lock)) {
1731 raw_spin_unlock(&rq->lock);
1732 raw_spin_lock(&p->pi_lock);
1733 raw_spin_lock(&rq->lock);
1736 if (!(p->state & TASK_NORMAL))
1739 if (!task_on_rq_queued(p))
1740 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1742 ttwu_do_wakeup(rq, p, 0);
1743 ttwu_stat(p, smp_processor_id(), 0);
1745 raw_spin_unlock(&p->pi_lock);
1749 * wake_up_process - Wake up a specific process
1750 * @p: The process to be woken up.
1752 * Attempt to wake up the nominated process and move it to the set of runnable
1755 * Return: 1 if the process was woken up, 0 if it was already running.
1757 * It may be assumed that this function implies a write memory barrier before
1758 * changing the task state if and only if any tasks are woken up.
1760 int wake_up_process(struct task_struct *p)
1762 WARN_ON(task_is_stopped_or_traced(p));
1763 return try_to_wake_up(p, TASK_NORMAL, 0);
1765 EXPORT_SYMBOL(wake_up_process);
1767 int wake_up_state(struct task_struct *p, unsigned int state)
1769 return try_to_wake_up(p, state, 0);
1773 * This function clears the sched_dl_entity static params.
1775 void __dl_clear_params(struct task_struct *p)
1777 struct sched_dl_entity *dl_se = &p->dl;
1779 dl_se->dl_runtime = 0;
1780 dl_se->dl_deadline = 0;
1781 dl_se->dl_period = 0;
1785 dl_se->dl_throttled = 0;
1787 dl_se->dl_yielded = 0;
1791 * Perform scheduler related setup for a newly forked process p.
1792 * p is forked by current.
1794 * __sched_fork() is basic setup used by init_idle() too:
1796 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1801 p->se.exec_start = 0;
1802 p->se.sum_exec_runtime = 0;
1803 p->se.prev_sum_exec_runtime = 0;
1804 p->se.nr_migrations = 0;
1807 p->se.avg.decay_count = 0;
1809 INIT_LIST_HEAD(&p->se.group_node);
1811 #ifdef CONFIG_SCHEDSTATS
1812 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1815 RB_CLEAR_NODE(&p->dl.rb_node);
1816 init_dl_task_timer(&p->dl);
1817 __dl_clear_params(p);
1819 INIT_LIST_HEAD(&p->rt.run_list);
1821 #ifdef CONFIG_PREEMPT_NOTIFIERS
1822 INIT_HLIST_HEAD(&p->preempt_notifiers);
1825 #ifdef CONFIG_NUMA_BALANCING
1826 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1827 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1828 p->mm->numa_scan_seq = 0;
1831 if (clone_flags & CLONE_VM)
1832 p->numa_preferred_nid = current->numa_preferred_nid;
1834 p->numa_preferred_nid = -1;
1836 p->node_stamp = 0ULL;
1837 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1838 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1839 p->numa_work.next = &p->numa_work;
1840 p->numa_faults = NULL;
1841 p->last_task_numa_placement = 0;
1842 p->last_sum_exec_runtime = 0;
1844 p->numa_group = NULL;
1845 #endif /* CONFIG_NUMA_BALANCING */
1848 #ifdef CONFIG_NUMA_BALANCING
1849 #ifdef CONFIG_SCHED_DEBUG
1850 void set_numabalancing_state(bool enabled)
1853 sched_feat_set("NUMA");
1855 sched_feat_set("NO_NUMA");
1858 __read_mostly bool numabalancing_enabled;
1860 void set_numabalancing_state(bool enabled)
1862 numabalancing_enabled = enabled;
1864 #endif /* CONFIG_SCHED_DEBUG */
1866 #ifdef CONFIG_PROC_SYSCTL
1867 int sysctl_numa_balancing(struct ctl_table *table, int write,
1868 void __user *buffer, size_t *lenp, loff_t *ppos)
1872 int state = numabalancing_enabled;
1874 if (write && !capable(CAP_SYS_ADMIN))
1879 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1883 set_numabalancing_state(state);
1890 * fork()/clone()-time setup:
1892 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1894 unsigned long flags;
1895 int cpu = get_cpu();
1897 __sched_fork(clone_flags, p);
1899 * We mark the process as running here. This guarantees that
1900 * nobody will actually run it, and a signal or other external
1901 * event cannot wake it up and insert it on the runqueue either.
1903 p->state = TASK_RUNNING;
1906 * Make sure we do not leak PI boosting priority to the child.
1908 p->prio = current->normal_prio;
1911 * Revert to default priority/policy on fork if requested.
1913 if (unlikely(p->sched_reset_on_fork)) {
1914 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1915 p->policy = SCHED_NORMAL;
1916 p->static_prio = NICE_TO_PRIO(0);
1918 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1919 p->static_prio = NICE_TO_PRIO(0);
1921 p->prio = p->normal_prio = __normal_prio(p);
1925 * We don't need the reset flag anymore after the fork. It has
1926 * fulfilled its duty:
1928 p->sched_reset_on_fork = 0;
1931 if (dl_prio(p->prio)) {
1934 } else if (rt_prio(p->prio)) {
1935 p->sched_class = &rt_sched_class;
1937 p->sched_class = &fair_sched_class;
1940 if (p->sched_class->task_fork)
1941 p->sched_class->task_fork(p);
1944 * The child is not yet in the pid-hash so no cgroup attach races,
1945 * and the cgroup is pinned to this child due to cgroup_fork()
1946 * is ran before sched_fork().
1948 * Silence PROVE_RCU.
1950 raw_spin_lock_irqsave(&p->pi_lock, flags);
1951 set_task_cpu(p, cpu);
1952 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1954 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1955 if (likely(sched_info_on()))
1956 memset(&p->sched_info, 0, sizeof(p->sched_info));
1958 #if defined(CONFIG_SMP)
1961 init_task_preempt_count(p);
1963 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1964 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1971 unsigned long to_ratio(u64 period, u64 runtime)
1973 if (runtime == RUNTIME_INF)
1977 * Doing this here saves a lot of checks in all
1978 * the calling paths, and returning zero seems
1979 * safe for them anyway.
1984 return div64_u64(runtime << 20, period);
1988 inline struct dl_bw *dl_bw_of(int i)
1990 rcu_lockdep_assert(rcu_read_lock_sched_held(),
1991 "sched RCU must be held");
1992 return &cpu_rq(i)->rd->dl_bw;
1995 static inline int dl_bw_cpus(int i)
1997 struct root_domain *rd = cpu_rq(i)->rd;
2000 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2001 "sched RCU must be held");
2002 for_each_cpu_and(i, rd->span, cpu_active_mask)
2008 inline struct dl_bw *dl_bw_of(int i)
2010 return &cpu_rq(i)->dl.dl_bw;
2013 static inline int dl_bw_cpus(int i)
2020 * We must be sure that accepting a new task (or allowing changing the
2021 * parameters of an existing one) is consistent with the bandwidth
2022 * constraints. If yes, this function also accordingly updates the currently
2023 * allocated bandwidth to reflect the new situation.
2025 * This function is called while holding p's rq->lock.
2027 * XXX we should delay bw change until the task's 0-lag point, see
2030 static int dl_overflow(struct task_struct *p, int policy,
2031 const struct sched_attr *attr)
2034 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2035 u64 period = attr->sched_period ?: attr->sched_deadline;
2036 u64 runtime = attr->sched_runtime;
2037 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2040 if (new_bw == p->dl.dl_bw)
2044 * Either if a task, enters, leave, or stays -deadline but changes
2045 * its parameters, we may need to update accordingly the total
2046 * allocated bandwidth of the container.
2048 raw_spin_lock(&dl_b->lock);
2049 cpus = dl_bw_cpus(task_cpu(p));
2050 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2051 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2052 __dl_add(dl_b, new_bw);
2054 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2055 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2056 __dl_clear(dl_b, p->dl.dl_bw);
2057 __dl_add(dl_b, new_bw);
2059 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2060 __dl_clear(dl_b, p->dl.dl_bw);
2063 raw_spin_unlock(&dl_b->lock);
2068 extern void init_dl_bw(struct dl_bw *dl_b);
2071 * wake_up_new_task - wake up a newly created task for the first time.
2073 * This function will do some initial scheduler statistics housekeeping
2074 * that must be done for every newly created context, then puts the task
2075 * on the runqueue and wakes it.
2077 void wake_up_new_task(struct task_struct *p)
2079 unsigned long flags;
2082 raw_spin_lock_irqsave(&p->pi_lock, flags);
2085 * Fork balancing, do it here and not earlier because:
2086 * - cpus_allowed can change in the fork path
2087 * - any previously selected cpu might disappear through hotplug
2089 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2092 /* Initialize new task's runnable average */
2093 init_task_runnable_average(p);
2094 rq = __task_rq_lock(p);
2095 activate_task(rq, p, 0);
2096 p->on_rq = TASK_ON_RQ_QUEUED;
2097 trace_sched_wakeup_new(p, true);
2098 check_preempt_curr(rq, p, WF_FORK);
2100 if (p->sched_class->task_woken)
2101 p->sched_class->task_woken(rq, p);
2103 task_rq_unlock(rq, p, &flags);
2106 #ifdef CONFIG_PREEMPT_NOTIFIERS
2109 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2110 * @notifier: notifier struct to register
2112 void preempt_notifier_register(struct preempt_notifier *notifier)
2114 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2116 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2119 * preempt_notifier_unregister - no longer interested in preemption notifications
2120 * @notifier: notifier struct to unregister
2122 * This is safe to call from within a preemption notifier.
2124 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2126 hlist_del(¬ifier->link);
2128 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2132 struct preempt_notifier *notifier;
2134 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2135 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2139 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2140 struct task_struct *next)
2142 struct preempt_notifier *notifier;
2144 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2145 notifier->ops->sched_out(notifier, next);
2148 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2150 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2155 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2156 struct task_struct *next)
2160 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2163 * prepare_task_switch - prepare to switch tasks
2164 * @rq: the runqueue preparing to switch
2165 * @prev: the current task that is being switched out
2166 * @next: the task we are going to switch to.
2168 * This is called with the rq lock held and interrupts off. It must
2169 * be paired with a subsequent finish_task_switch after the context
2172 * prepare_task_switch sets up locking and calls architecture specific
2176 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2177 struct task_struct *next)
2179 trace_sched_switch(prev, next);
2180 sched_info_switch(rq, prev, next);
2181 perf_event_task_sched_out(prev, next);
2182 fire_sched_out_preempt_notifiers(prev, next);
2183 prepare_lock_switch(rq, next);
2184 prepare_arch_switch(next);
2188 * finish_task_switch - clean up after a task-switch
2189 * @prev: the thread we just switched away from.
2191 * finish_task_switch must be called after the context switch, paired
2192 * with a prepare_task_switch call before the context switch.
2193 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2194 * and do any other architecture-specific cleanup actions.
2196 * Note that we may have delayed dropping an mm in context_switch(). If
2197 * so, we finish that here outside of the runqueue lock. (Doing it
2198 * with the lock held can cause deadlocks; see schedule() for
2201 * The context switch have flipped the stack from under us and restored the
2202 * local variables which were saved when this task called schedule() in the
2203 * past. prev == current is still correct but we need to recalculate this_rq
2204 * because prev may have moved to another CPU.
2206 static struct rq *finish_task_switch(struct task_struct *prev)
2207 __releases(rq->lock)
2209 struct rq *rq = this_rq();
2210 struct mm_struct *mm = rq->prev_mm;
2216 * A task struct has one reference for the use as "current".
2217 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2218 * schedule one last time. The schedule call will never return, and
2219 * the scheduled task must drop that reference.
2220 * The test for TASK_DEAD must occur while the runqueue locks are
2221 * still held, otherwise prev could be scheduled on another cpu, die
2222 * there before we look at prev->state, and then the reference would
2224 * Manfred Spraul <manfred@colorfullife.com>
2226 prev_state = prev->state;
2227 vtime_task_switch(prev);
2228 finish_arch_switch(prev);
2229 perf_event_task_sched_in(prev, current);
2230 finish_lock_switch(rq, prev);
2231 finish_arch_post_lock_switch();
2233 fire_sched_in_preempt_notifiers(current);
2236 if (unlikely(prev_state == TASK_DEAD)) {
2237 if (prev->sched_class->task_dead)
2238 prev->sched_class->task_dead(prev);
2241 * Remove function-return probe instances associated with this
2242 * task and put them back on the free list.
2244 kprobe_flush_task(prev);
2245 put_task_struct(prev);
2248 tick_nohz_task_switch(current);
2254 /* rq->lock is NOT held, but preemption is disabled */
2255 static inline void post_schedule(struct rq *rq)
2257 if (rq->post_schedule) {
2258 unsigned long flags;
2260 raw_spin_lock_irqsave(&rq->lock, flags);
2261 if (rq->curr->sched_class->post_schedule)
2262 rq->curr->sched_class->post_schedule(rq);
2263 raw_spin_unlock_irqrestore(&rq->lock, flags);
2265 rq->post_schedule = 0;
2271 static inline void post_schedule(struct rq *rq)
2278 * schedule_tail - first thing a freshly forked thread must call.
2279 * @prev: the thread we just switched away from.
2281 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2282 __releases(rq->lock)
2286 /* finish_task_switch() drops rq->lock and enables preemtion */
2288 rq = finish_task_switch(prev);
2292 if (current->set_child_tid)
2293 put_user(task_pid_vnr(current), current->set_child_tid);
2297 * context_switch - switch to the new MM and the new thread's register state.
2299 static inline struct rq *
2300 context_switch(struct rq *rq, struct task_struct *prev,
2301 struct task_struct *next)
2303 struct mm_struct *mm, *oldmm;
2305 prepare_task_switch(rq, prev, next);
2308 oldmm = prev->active_mm;
2310 * For paravirt, this is coupled with an exit in switch_to to
2311 * combine the page table reload and the switch backend into
2314 arch_start_context_switch(prev);
2317 next->active_mm = oldmm;
2318 atomic_inc(&oldmm->mm_count);
2319 enter_lazy_tlb(oldmm, next);
2321 switch_mm(oldmm, mm, next);
2324 prev->active_mm = NULL;
2325 rq->prev_mm = oldmm;
2328 * Since the runqueue lock will be released by the next
2329 * task (which is an invalid locking op but in the case
2330 * of the scheduler it's an obvious special-case), so we
2331 * do an early lockdep release here:
2333 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2335 context_tracking_task_switch(prev, next);
2336 /* Here we just switch the register state and the stack. */
2337 switch_to(prev, next, prev);
2340 return finish_task_switch(prev);
2344 * nr_running and nr_context_switches:
2346 * externally visible scheduler statistics: current number of runnable
2347 * threads, total number of context switches performed since bootup.
2349 unsigned long nr_running(void)
2351 unsigned long i, sum = 0;
2353 for_each_online_cpu(i)
2354 sum += cpu_rq(i)->nr_running;
2360 * Check if only the current task is running on the cpu.
2362 bool single_task_running(void)
2364 if (cpu_rq(smp_processor_id())->nr_running == 1)
2369 EXPORT_SYMBOL(single_task_running);
2371 unsigned long long nr_context_switches(void)
2374 unsigned long long sum = 0;
2376 for_each_possible_cpu(i)
2377 sum += cpu_rq(i)->nr_switches;
2382 unsigned long nr_iowait(void)
2384 unsigned long i, sum = 0;
2386 for_each_possible_cpu(i)
2387 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2392 unsigned long nr_iowait_cpu(int cpu)
2394 struct rq *this = cpu_rq(cpu);
2395 return atomic_read(&this->nr_iowait);
2398 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2400 struct rq *this = this_rq();
2401 *nr_waiters = atomic_read(&this->nr_iowait);
2402 *load = this->cpu_load[0];
2408 * sched_exec - execve() is a valuable balancing opportunity, because at
2409 * this point the task has the smallest effective memory and cache footprint.
2411 void sched_exec(void)
2413 struct task_struct *p = current;
2414 unsigned long flags;
2417 raw_spin_lock_irqsave(&p->pi_lock, flags);
2418 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2419 if (dest_cpu == smp_processor_id())
2422 if (likely(cpu_active(dest_cpu))) {
2423 struct migration_arg arg = { p, dest_cpu };
2425 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2426 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2430 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2435 DEFINE_PER_CPU(struct kernel_stat, kstat);
2436 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2438 EXPORT_PER_CPU_SYMBOL(kstat);
2439 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2442 * Return accounted runtime for the task.
2443 * In case the task is currently running, return the runtime plus current's
2444 * pending runtime that have not been accounted yet.
2446 unsigned long long task_sched_runtime(struct task_struct *p)
2448 unsigned long flags;
2452 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2454 * 64-bit doesn't need locks to atomically read a 64bit value.
2455 * So we have a optimization chance when the task's delta_exec is 0.
2456 * Reading ->on_cpu is racy, but this is ok.
2458 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2459 * If we race with it entering cpu, unaccounted time is 0. This is
2460 * indistinguishable from the read occurring a few cycles earlier.
2461 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2462 * been accounted, so we're correct here as well.
2464 if (!p->on_cpu || !task_on_rq_queued(p))
2465 return p->se.sum_exec_runtime;
2468 rq = task_rq_lock(p, &flags);
2470 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2471 * project cycles that may never be accounted to this
2472 * thread, breaking clock_gettime().
2474 if (task_current(rq, p) && task_on_rq_queued(p)) {
2475 update_rq_clock(rq);
2476 p->sched_class->update_curr(rq);
2478 ns = p->se.sum_exec_runtime;
2479 task_rq_unlock(rq, p, &flags);
2485 * This function gets called by the timer code, with HZ frequency.
2486 * We call it with interrupts disabled.
2488 void scheduler_tick(void)
2490 int cpu = smp_processor_id();
2491 struct rq *rq = cpu_rq(cpu);
2492 struct task_struct *curr = rq->curr;
2496 raw_spin_lock(&rq->lock);
2497 update_rq_clock(rq);
2498 curr->sched_class->task_tick(rq, curr, 0);
2499 update_cpu_load_active(rq);
2500 raw_spin_unlock(&rq->lock);
2502 perf_event_task_tick();
2505 rq->idle_balance = idle_cpu(cpu);
2506 trigger_load_balance(rq);
2508 rq_last_tick_reset(rq);
2511 #ifdef CONFIG_NO_HZ_FULL
2513 * scheduler_tick_max_deferment
2515 * Keep at least one tick per second when a single
2516 * active task is running because the scheduler doesn't
2517 * yet completely support full dynticks environment.
2519 * This makes sure that uptime, CFS vruntime, load
2520 * balancing, etc... continue to move forward, even
2521 * with a very low granularity.
2523 * Return: Maximum deferment in nanoseconds.
2525 u64 scheduler_tick_max_deferment(void)
2527 struct rq *rq = this_rq();
2528 unsigned long next, now = ACCESS_ONCE(jiffies);
2530 next = rq->last_sched_tick + HZ;
2532 if (time_before_eq(next, now))
2535 return jiffies_to_nsecs(next - now);
2539 notrace unsigned long get_parent_ip(unsigned long addr)
2541 if (in_lock_functions(addr)) {
2542 addr = CALLER_ADDR2;
2543 if (in_lock_functions(addr))
2544 addr = CALLER_ADDR3;
2549 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2550 defined(CONFIG_PREEMPT_TRACER))
2552 void preempt_count_add(int val)
2554 #ifdef CONFIG_DEBUG_PREEMPT
2558 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2561 __preempt_count_add(val);
2562 #ifdef CONFIG_DEBUG_PREEMPT
2564 * Spinlock count overflowing soon?
2566 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2569 if (preempt_count() == val) {
2570 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2571 #ifdef CONFIG_DEBUG_PREEMPT
2572 current->preempt_disable_ip = ip;
2574 trace_preempt_off(CALLER_ADDR0, ip);
2577 EXPORT_SYMBOL(preempt_count_add);
2578 NOKPROBE_SYMBOL(preempt_count_add);
2580 void preempt_count_sub(int val)
2582 #ifdef CONFIG_DEBUG_PREEMPT
2586 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2589 * Is the spinlock portion underflowing?
2591 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2592 !(preempt_count() & PREEMPT_MASK)))
2596 if (preempt_count() == val)
2597 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2598 __preempt_count_sub(val);
2600 EXPORT_SYMBOL(preempt_count_sub);
2601 NOKPROBE_SYMBOL(preempt_count_sub);
2606 * Print scheduling while atomic bug:
2608 static noinline void __schedule_bug(struct task_struct *prev)
2610 if (oops_in_progress)
2613 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2614 prev->comm, prev->pid, preempt_count());
2616 debug_show_held_locks(prev);
2618 if (irqs_disabled())
2619 print_irqtrace_events(prev);
2620 #ifdef CONFIG_DEBUG_PREEMPT
2621 if (in_atomic_preempt_off()) {
2622 pr_err("Preemption disabled at:");
2623 print_ip_sym(current->preempt_disable_ip);
2628 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2632 * Various schedule()-time debugging checks and statistics:
2634 static inline void schedule_debug(struct task_struct *prev)
2636 #ifdef CONFIG_SCHED_STACK_END_CHECK
2637 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2640 * Test if we are atomic. Since do_exit() needs to call into
2641 * schedule() atomically, we ignore that path. Otherwise whine
2642 * if we are scheduling when we should not.
2644 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2645 __schedule_bug(prev);
2648 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2650 schedstat_inc(this_rq(), sched_count);
2654 * Pick up the highest-prio task:
2656 static inline struct task_struct *
2657 pick_next_task(struct rq *rq, struct task_struct *prev)
2659 const struct sched_class *class = &fair_sched_class;
2660 struct task_struct *p;
2663 * Optimization: we know that if all tasks are in
2664 * the fair class we can call that function directly:
2666 if (likely(prev->sched_class == class &&
2667 rq->nr_running == rq->cfs.h_nr_running)) {
2668 p = fair_sched_class.pick_next_task(rq, prev);
2669 if (unlikely(p == RETRY_TASK))
2672 /* assumes fair_sched_class->next == idle_sched_class */
2674 p = idle_sched_class.pick_next_task(rq, prev);
2680 for_each_class(class) {
2681 p = class->pick_next_task(rq, prev);
2683 if (unlikely(p == RETRY_TASK))
2689 BUG(); /* the idle class will always have a runnable task */
2693 * __schedule() is the main scheduler function.
2695 * The main means of driving the scheduler and thus entering this function are:
2697 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2699 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2700 * paths. For example, see arch/x86/entry_64.S.
2702 * To drive preemption between tasks, the scheduler sets the flag in timer
2703 * interrupt handler scheduler_tick().
2705 * 3. Wakeups don't really cause entry into schedule(). They add a
2706 * task to the run-queue and that's it.
2708 * Now, if the new task added to the run-queue preempts the current
2709 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2710 * called on the nearest possible occasion:
2712 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2714 * - in syscall or exception context, at the next outmost
2715 * preempt_enable(). (this might be as soon as the wake_up()'s
2718 * - in IRQ context, return from interrupt-handler to
2719 * preemptible context
2721 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2724 * - cond_resched() call
2725 * - explicit schedule() call
2726 * - return from syscall or exception to user-space
2727 * - return from interrupt-handler to user-space
2729 * WARNING: all callers must re-check need_resched() afterward and reschedule
2730 * accordingly in case an event triggered the need for rescheduling (such as
2731 * an interrupt waking up a task) while preemption was disabled in __schedule().
2733 static void __sched __schedule(void)
2735 struct task_struct *prev, *next;
2736 unsigned long *switch_count;
2741 cpu = smp_processor_id();
2743 rcu_note_context_switch();
2746 schedule_debug(prev);
2748 if (sched_feat(HRTICK))
2752 * Make sure that signal_pending_state()->signal_pending() below
2753 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2754 * done by the caller to avoid the race with signal_wake_up().
2756 smp_mb__before_spinlock();
2757 raw_spin_lock_irq(&rq->lock);
2759 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
2761 switch_count = &prev->nivcsw;
2762 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2763 if (unlikely(signal_pending_state(prev->state, prev))) {
2764 prev->state = TASK_RUNNING;
2766 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2770 * If a worker went to sleep, notify and ask workqueue
2771 * whether it wants to wake up a task to maintain
2774 if (prev->flags & PF_WQ_WORKER) {
2775 struct task_struct *to_wakeup;
2777 to_wakeup = wq_worker_sleeping(prev, cpu);
2779 try_to_wake_up_local(to_wakeup);
2782 switch_count = &prev->nvcsw;
2785 if (task_on_rq_queued(prev))
2786 update_rq_clock(rq);
2788 next = pick_next_task(rq, prev);
2789 clear_tsk_need_resched(prev);
2790 clear_preempt_need_resched();
2791 rq->clock_skip_update = 0;
2793 if (likely(prev != next)) {
2798 rq = context_switch(rq, prev, next); /* unlocks the rq */
2801 raw_spin_unlock_irq(&rq->lock);
2805 sched_preempt_enable_no_resched();
2808 static inline void sched_submit_work(struct task_struct *tsk)
2810 if (!tsk->state || tsk_is_pi_blocked(tsk))
2813 * If we are going to sleep and we have plugged IO queued,
2814 * make sure to submit it to avoid deadlocks.
2816 if (blk_needs_flush_plug(tsk))
2817 blk_schedule_flush_plug(tsk);
2820 asmlinkage __visible void __sched schedule(void)
2822 struct task_struct *tsk = current;
2824 sched_submit_work(tsk);
2827 } while (need_resched());
2829 EXPORT_SYMBOL(schedule);
2831 #ifdef CONFIG_CONTEXT_TRACKING
2832 asmlinkage __visible void __sched schedule_user(void)
2835 * If we come here after a random call to set_need_resched(),
2836 * or we have been woken up remotely but the IPI has not yet arrived,
2837 * we haven't yet exited the RCU idle mode. Do it here manually until
2838 * we find a better solution.
2840 * NB: There are buggy callers of this function. Ideally we
2841 * should warn if prev_state != CONTEXT_USER, but that will trigger
2842 * too frequently to make sense yet.
2844 enum ctx_state prev_state = exception_enter();
2846 exception_exit(prev_state);
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 static void __sched notrace preempt_schedule_common(void)
2865 __preempt_count_add(PREEMPT_ACTIVE);
2867 __preempt_count_sub(PREEMPT_ACTIVE);
2870 * Check again in case we missed a preemption opportunity
2871 * between schedule and now.
2874 } while (need_resched());
2877 #ifdef CONFIG_PREEMPT
2879 * this is the entry point to schedule() from in-kernel preemption
2880 * off of preempt_enable. Kernel preemptions off return from interrupt
2881 * occur there and call schedule directly.
2883 asmlinkage __visible void __sched notrace preempt_schedule(void)
2886 * If there is a non-zero preempt_count or interrupts are disabled,
2887 * we do not want to preempt the current task. Just return..
2889 if (likely(!preemptible()))
2892 preempt_schedule_common();
2894 NOKPROBE_SYMBOL(preempt_schedule);
2895 EXPORT_SYMBOL(preempt_schedule);
2897 #ifdef CONFIG_CONTEXT_TRACKING
2899 * preempt_schedule_context - preempt_schedule called by tracing
2901 * The tracing infrastructure uses preempt_enable_notrace to prevent
2902 * recursion and tracing preempt enabling caused by the tracing
2903 * infrastructure itself. But as tracing can happen in areas coming
2904 * from userspace or just about to enter userspace, a preempt enable
2905 * can occur before user_exit() is called. This will cause the scheduler
2906 * to be called when the system is still in usermode.
2908 * To prevent this, the preempt_enable_notrace will use this function
2909 * instead of preempt_schedule() to exit user context if needed before
2910 * calling the scheduler.
2912 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2914 enum ctx_state prev_ctx;
2916 if (likely(!preemptible()))
2920 __preempt_count_add(PREEMPT_ACTIVE);
2922 * Needs preempt disabled in case user_exit() is traced
2923 * and the tracer calls preempt_enable_notrace() causing
2924 * an infinite recursion.
2926 prev_ctx = exception_enter();
2928 exception_exit(prev_ctx);
2930 __preempt_count_sub(PREEMPT_ACTIVE);
2932 } while (need_resched());
2934 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2935 #endif /* CONFIG_CONTEXT_TRACKING */
2937 #endif /* CONFIG_PREEMPT */
2940 * this is the entry point to schedule() from kernel preemption
2941 * off of irq context.
2942 * Note, that this is called and return with irqs disabled. This will
2943 * protect us against recursive calling from irq.
2945 asmlinkage __visible void __sched preempt_schedule_irq(void)
2947 enum ctx_state prev_state;
2949 /* Catch callers which need to be fixed */
2950 BUG_ON(preempt_count() || !irqs_disabled());
2952 prev_state = exception_enter();
2955 __preempt_count_add(PREEMPT_ACTIVE);
2958 local_irq_disable();
2959 __preempt_count_sub(PREEMPT_ACTIVE);
2962 * Check again in case we missed a preemption opportunity
2963 * between schedule and now.
2966 } while (need_resched());
2968 exception_exit(prev_state);
2971 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2974 return try_to_wake_up(curr->private, mode, wake_flags);
2976 EXPORT_SYMBOL(default_wake_function);
2978 #ifdef CONFIG_RT_MUTEXES
2981 * rt_mutex_setprio - set the current priority of a task
2983 * @prio: prio value (kernel-internal form)
2985 * This function changes the 'effective' priority of a task. It does
2986 * not touch ->normal_prio like __setscheduler().
2988 * Used by the rt_mutex code to implement priority inheritance
2989 * logic. Call site only calls if the priority of the task changed.
2991 void rt_mutex_setprio(struct task_struct *p, int prio)
2993 int oldprio, queued, running, enqueue_flag = 0;
2995 const struct sched_class *prev_class;
2997 BUG_ON(prio > MAX_PRIO);
2999 rq = __task_rq_lock(p);
3002 * Idle task boosting is a nono in general. There is one
3003 * exception, when PREEMPT_RT and NOHZ is active:
3005 * The idle task calls get_next_timer_interrupt() and holds
3006 * the timer wheel base->lock on the CPU and another CPU wants
3007 * to access the timer (probably to cancel it). We can safely
3008 * ignore the boosting request, as the idle CPU runs this code
3009 * with interrupts disabled and will complete the lock
3010 * protected section without being interrupted. So there is no
3011 * real need to boost.
3013 if (unlikely(p == rq->idle)) {
3014 WARN_ON(p != rq->curr);
3015 WARN_ON(p->pi_blocked_on);
3019 trace_sched_pi_setprio(p, prio);
3021 prev_class = p->sched_class;
3022 queued = task_on_rq_queued(p);
3023 running = task_current(rq, p);
3025 dequeue_task(rq, p, 0);
3027 put_prev_task(rq, p);
3030 * Boosting condition are:
3031 * 1. -rt task is running and holds mutex A
3032 * --> -dl task blocks on mutex A
3034 * 2. -dl task is running and holds mutex A
3035 * --> -dl task blocks on mutex A and could preempt the
3038 if (dl_prio(prio)) {
3039 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3040 if (!dl_prio(p->normal_prio) ||
3041 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3042 p->dl.dl_boosted = 1;
3043 p->dl.dl_throttled = 0;
3044 enqueue_flag = ENQUEUE_REPLENISH;
3046 p->dl.dl_boosted = 0;
3047 p->sched_class = &dl_sched_class;
3048 } else if (rt_prio(prio)) {
3049 if (dl_prio(oldprio))
3050 p->dl.dl_boosted = 0;
3052 enqueue_flag = ENQUEUE_HEAD;
3053 p->sched_class = &rt_sched_class;
3055 if (dl_prio(oldprio))
3056 p->dl.dl_boosted = 0;
3057 if (rt_prio(oldprio))
3059 p->sched_class = &fair_sched_class;
3065 p->sched_class->set_curr_task(rq);
3067 enqueue_task(rq, p, enqueue_flag);
3069 check_class_changed(rq, p, prev_class, oldprio);
3071 __task_rq_unlock(rq);
3075 void set_user_nice(struct task_struct *p, long nice)
3077 int old_prio, delta, queued;
3078 unsigned long flags;
3081 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3084 * We have to be careful, if called from sys_setpriority(),
3085 * the task might be in the middle of scheduling on another CPU.
3087 rq = task_rq_lock(p, &flags);
3089 * The RT priorities are set via sched_setscheduler(), but we still
3090 * allow the 'normal' nice value to be set - but as expected
3091 * it wont have any effect on scheduling until the task is
3092 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3094 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3095 p->static_prio = NICE_TO_PRIO(nice);
3098 queued = task_on_rq_queued(p);
3100 dequeue_task(rq, p, 0);
3102 p->static_prio = NICE_TO_PRIO(nice);
3105 p->prio = effective_prio(p);
3106 delta = p->prio - old_prio;
3109 enqueue_task(rq, p, 0);
3111 * If the task increased its priority or is running and
3112 * lowered its priority, then reschedule its CPU:
3114 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3118 task_rq_unlock(rq, p, &flags);
3120 EXPORT_SYMBOL(set_user_nice);
3123 * can_nice - check if a task can reduce its nice value
3127 int can_nice(const struct task_struct *p, const int nice)
3129 /* convert nice value [19,-20] to rlimit style value [1,40] */
3130 int nice_rlim = nice_to_rlimit(nice);
3132 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3133 capable(CAP_SYS_NICE));
3136 #ifdef __ARCH_WANT_SYS_NICE
3139 * sys_nice - change the priority of the current process.
3140 * @increment: priority increment
3142 * sys_setpriority is a more generic, but much slower function that
3143 * does similar things.
3145 SYSCALL_DEFINE1(nice, int, increment)
3150 * Setpriority might change our priority at the same moment.
3151 * We don't have to worry. Conceptually one call occurs first
3152 * and we have a single winner.
3154 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3155 nice = task_nice(current) + increment;
3157 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3158 if (increment < 0 && !can_nice(current, nice))
3161 retval = security_task_setnice(current, nice);
3165 set_user_nice(current, nice);
3172 * task_prio - return the priority value of a given task.
3173 * @p: the task in question.
3175 * Return: The priority value as seen by users in /proc.
3176 * RT tasks are offset by -200. Normal tasks are centered
3177 * around 0, value goes from -16 to +15.
3179 int task_prio(const struct task_struct *p)
3181 return p->prio - MAX_RT_PRIO;
3185 * idle_cpu - is a given cpu idle currently?
3186 * @cpu: the processor in question.
3188 * Return: 1 if the CPU is currently idle. 0 otherwise.
3190 int idle_cpu(int cpu)
3192 struct rq *rq = cpu_rq(cpu);
3194 if (rq->curr != rq->idle)
3201 if (!llist_empty(&rq->wake_list))
3209 * idle_task - return the idle task for a given cpu.
3210 * @cpu: the processor in question.
3212 * Return: The idle task for the cpu @cpu.
3214 struct task_struct *idle_task(int cpu)
3216 return cpu_rq(cpu)->idle;
3220 * find_process_by_pid - find a process with a matching PID value.
3221 * @pid: the pid in question.
3223 * The task of @pid, if found. %NULL otherwise.
3225 static struct task_struct *find_process_by_pid(pid_t pid)
3227 return pid ? find_task_by_vpid(pid) : current;
3231 * This function initializes the sched_dl_entity of a newly becoming
3232 * SCHED_DEADLINE task.
3234 * Only the static values are considered here, the actual runtime and the
3235 * absolute deadline will be properly calculated when the task is enqueued
3236 * for the first time with its new policy.
3239 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3241 struct sched_dl_entity *dl_se = &p->dl;
3243 dl_se->dl_runtime = attr->sched_runtime;
3244 dl_se->dl_deadline = attr->sched_deadline;
3245 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3246 dl_se->flags = attr->sched_flags;
3247 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3250 * Changing the parameters of a task is 'tricky' and we're not doing
3251 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3253 * What we SHOULD do is delay the bandwidth release until the 0-lag
3254 * point. This would include retaining the task_struct until that time
3255 * and change dl_overflow() to not immediately decrement the current
3258 * Instead we retain the current runtime/deadline and let the new
3259 * parameters take effect after the current reservation period lapses.
3260 * This is safe (albeit pessimistic) because the 0-lag point is always
3261 * before the current scheduling deadline.
3263 * We can still have temporary overloads because we do not delay the
3264 * change in bandwidth until that time; so admission control is
3265 * not on the safe side. It does however guarantee tasks will never
3266 * consume more than promised.
3271 * sched_setparam() passes in -1 for its policy, to let the functions
3272 * it calls know not to change it.
3274 #define SETPARAM_POLICY -1
3276 static void __setscheduler_params(struct task_struct *p,
3277 const struct sched_attr *attr)
3279 int policy = attr->sched_policy;
3281 if (policy == SETPARAM_POLICY)
3286 if (dl_policy(policy))
3287 __setparam_dl(p, attr);
3288 else if (fair_policy(policy))
3289 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3292 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3293 * !rt_policy. Always setting this ensures that things like
3294 * getparam()/getattr() don't report silly values for !rt tasks.
3296 p->rt_priority = attr->sched_priority;
3297 p->normal_prio = normal_prio(p);
3301 /* Actually do priority change: must hold pi & rq lock. */
3302 static void __setscheduler(struct rq *rq, struct task_struct *p,
3303 const struct sched_attr *attr)
3305 __setscheduler_params(p, attr);
3308 * If we get here, there was no pi waiters boosting the
3309 * task. It is safe to use the normal prio.
3311 p->prio = normal_prio(p);
3313 if (dl_prio(p->prio))
3314 p->sched_class = &dl_sched_class;
3315 else if (rt_prio(p->prio))
3316 p->sched_class = &rt_sched_class;
3318 p->sched_class = &fair_sched_class;
3322 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3324 struct sched_dl_entity *dl_se = &p->dl;
3326 attr->sched_priority = p->rt_priority;
3327 attr->sched_runtime = dl_se->dl_runtime;
3328 attr->sched_deadline = dl_se->dl_deadline;
3329 attr->sched_period = dl_se->dl_period;
3330 attr->sched_flags = dl_se->flags;
3334 * This function validates the new parameters of a -deadline task.
3335 * We ask for the deadline not being zero, and greater or equal
3336 * than the runtime, as well as the period of being zero or
3337 * greater than deadline. Furthermore, we have to be sure that
3338 * user parameters are above the internal resolution of 1us (we
3339 * check sched_runtime only since it is always the smaller one) and
3340 * below 2^63 ns (we have to check both sched_deadline and
3341 * sched_period, as the latter can be zero).
3344 __checkparam_dl(const struct sched_attr *attr)
3347 if (attr->sched_deadline == 0)
3351 * Since we truncate DL_SCALE bits, make sure we're at least
3354 if (attr->sched_runtime < (1ULL << DL_SCALE))
3358 * Since we use the MSB for wrap-around and sign issues, make
3359 * sure it's not set (mind that period can be equal to zero).
3361 if (attr->sched_deadline & (1ULL << 63) ||
3362 attr->sched_period & (1ULL << 63))
3365 /* runtime <= deadline <= period (if period != 0) */
3366 if ((attr->sched_period != 0 &&
3367 attr->sched_period < attr->sched_deadline) ||
3368 attr->sched_deadline < attr->sched_runtime)
3375 * check the target process has a UID that matches the current process's
3377 static bool check_same_owner(struct task_struct *p)
3379 const struct cred *cred = current_cred(), *pcred;
3383 pcred = __task_cred(p);
3384 match = (uid_eq(cred->euid, pcred->euid) ||
3385 uid_eq(cred->euid, pcred->uid));
3390 static bool dl_param_changed(struct task_struct *p,
3391 const struct sched_attr *attr)
3393 struct sched_dl_entity *dl_se = &p->dl;
3395 if (dl_se->dl_runtime != attr->sched_runtime ||
3396 dl_se->dl_deadline != attr->sched_deadline ||
3397 dl_se->dl_period != attr->sched_period ||
3398 dl_se->flags != attr->sched_flags)
3404 static int __sched_setscheduler(struct task_struct *p,
3405 const struct sched_attr *attr,
3408 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3409 MAX_RT_PRIO - 1 - attr->sched_priority;
3410 int retval, oldprio, oldpolicy = -1, queued, running;
3411 int policy = attr->sched_policy;
3412 unsigned long flags;
3413 const struct sched_class *prev_class;
3417 /* may grab non-irq protected spin_locks */
3418 BUG_ON(in_interrupt());
3420 /* double check policy once rq lock held */
3422 reset_on_fork = p->sched_reset_on_fork;
3423 policy = oldpolicy = p->policy;
3425 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3427 if (policy != SCHED_DEADLINE &&
3428 policy != SCHED_FIFO && policy != SCHED_RR &&
3429 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3430 policy != SCHED_IDLE)
3434 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3438 * Valid priorities for SCHED_FIFO and SCHED_RR are
3439 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3440 * SCHED_BATCH and SCHED_IDLE is 0.
3442 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3443 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3445 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3446 (rt_policy(policy) != (attr->sched_priority != 0)))
3450 * Allow unprivileged RT tasks to decrease priority:
3452 if (user && !capable(CAP_SYS_NICE)) {
3453 if (fair_policy(policy)) {
3454 if (attr->sched_nice < task_nice(p) &&
3455 !can_nice(p, attr->sched_nice))
3459 if (rt_policy(policy)) {
3460 unsigned long rlim_rtprio =
3461 task_rlimit(p, RLIMIT_RTPRIO);
3463 /* can't set/change the rt policy */
3464 if (policy != p->policy && !rlim_rtprio)
3467 /* can't increase priority */
3468 if (attr->sched_priority > p->rt_priority &&
3469 attr->sched_priority > rlim_rtprio)
3474 * Can't set/change SCHED_DEADLINE policy at all for now
3475 * (safest behavior); in the future we would like to allow
3476 * unprivileged DL tasks to increase their relative deadline
3477 * or reduce their runtime (both ways reducing utilization)
3479 if (dl_policy(policy))
3483 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3484 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3486 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3487 if (!can_nice(p, task_nice(p)))
3491 /* can't change other user's priorities */
3492 if (!check_same_owner(p))
3495 /* Normal users shall not reset the sched_reset_on_fork flag */
3496 if (p->sched_reset_on_fork && !reset_on_fork)
3501 retval = security_task_setscheduler(p);
3507 * make sure no PI-waiters arrive (or leave) while we are
3508 * changing the priority of the task:
3510 * To be able to change p->policy safely, the appropriate
3511 * runqueue lock must be held.
3513 rq = task_rq_lock(p, &flags);
3516 * Changing the policy of the stop threads its a very bad idea
3518 if (p == rq->stop) {
3519 task_rq_unlock(rq, p, &flags);
3524 * If not changing anything there's no need to proceed further,
3525 * but store a possible modification of reset_on_fork.
3527 if (unlikely(policy == p->policy)) {
3528 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3530 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3532 if (dl_policy(policy) && dl_param_changed(p, attr))
3535 p->sched_reset_on_fork = reset_on_fork;
3536 task_rq_unlock(rq, p, &flags);
3542 #ifdef CONFIG_RT_GROUP_SCHED
3544 * Do not allow realtime tasks into groups that have no runtime
3547 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3548 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3549 !task_group_is_autogroup(task_group(p))) {
3550 task_rq_unlock(rq, p, &flags);
3555 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3556 cpumask_t *span = rq->rd->span;
3559 * Don't allow tasks with an affinity mask smaller than
3560 * the entire root_domain to become SCHED_DEADLINE. We
3561 * will also fail if there's no bandwidth available.
3563 if (!cpumask_subset(span, &p->cpus_allowed) ||
3564 rq->rd->dl_bw.bw == 0) {
3565 task_rq_unlock(rq, p, &flags);
3572 /* recheck policy now with rq lock held */
3573 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3574 policy = oldpolicy = -1;
3575 task_rq_unlock(rq, p, &flags);
3580 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3581 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3584 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3585 task_rq_unlock(rq, p, &flags);
3589 p->sched_reset_on_fork = reset_on_fork;
3593 * Special case for priority boosted tasks.
3595 * If the new priority is lower or equal (user space view)
3596 * than the current (boosted) priority, we just store the new
3597 * normal parameters and do not touch the scheduler class and
3598 * the runqueue. This will be done when the task deboost
3601 if (rt_mutex_check_prio(p, newprio)) {
3602 __setscheduler_params(p, attr);
3603 task_rq_unlock(rq, p, &flags);
3607 queued = task_on_rq_queued(p);
3608 running = task_current(rq, p);
3610 dequeue_task(rq, p, 0);
3612 put_prev_task(rq, p);
3614 prev_class = p->sched_class;
3615 __setscheduler(rq, p, attr);
3618 p->sched_class->set_curr_task(rq);
3621 * We enqueue to tail when the priority of a task is
3622 * increased (user space view).
3624 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3627 check_class_changed(rq, p, prev_class, oldprio);
3628 task_rq_unlock(rq, p, &flags);
3630 rt_mutex_adjust_pi(p);
3635 static int _sched_setscheduler(struct task_struct *p, int policy,
3636 const struct sched_param *param, bool check)
3638 struct sched_attr attr = {
3639 .sched_policy = policy,
3640 .sched_priority = param->sched_priority,
3641 .sched_nice = PRIO_TO_NICE(p->static_prio),
3644 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3645 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3646 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3647 policy &= ~SCHED_RESET_ON_FORK;
3648 attr.sched_policy = policy;
3651 return __sched_setscheduler(p, &attr, check);
3654 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3655 * @p: the task in question.
3656 * @policy: new policy.
3657 * @param: structure containing the new RT priority.
3659 * Return: 0 on success. An error code otherwise.
3661 * NOTE that the task may be already dead.
3663 int sched_setscheduler(struct task_struct *p, int policy,
3664 const struct sched_param *param)
3666 return _sched_setscheduler(p, policy, param, true);
3668 EXPORT_SYMBOL_GPL(sched_setscheduler);
3670 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3672 return __sched_setscheduler(p, attr, true);
3674 EXPORT_SYMBOL_GPL(sched_setattr);
3677 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3678 * @p: the task in question.
3679 * @policy: new policy.
3680 * @param: structure containing the new RT priority.
3682 * Just like sched_setscheduler, only don't bother checking if the
3683 * current context has permission. For example, this is needed in
3684 * stop_machine(): we create temporary high priority worker threads,
3685 * but our caller might not have that capability.
3687 * Return: 0 on success. An error code otherwise.
3689 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3690 const struct sched_param *param)
3692 return _sched_setscheduler(p, policy, param, false);
3696 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3698 struct sched_param lparam;
3699 struct task_struct *p;
3702 if (!param || pid < 0)
3704 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3709 p = find_process_by_pid(pid);
3711 retval = sched_setscheduler(p, policy, &lparam);
3718 * Mimics kernel/events/core.c perf_copy_attr().
3720 static int sched_copy_attr(struct sched_attr __user *uattr,
3721 struct sched_attr *attr)
3726 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3730 * zero the full structure, so that a short copy will be nice.
3732 memset(attr, 0, sizeof(*attr));
3734 ret = get_user(size, &uattr->size);
3738 if (size > PAGE_SIZE) /* silly large */
3741 if (!size) /* abi compat */
3742 size = SCHED_ATTR_SIZE_VER0;
3744 if (size < SCHED_ATTR_SIZE_VER0)
3748 * If we're handed a bigger struct than we know of,
3749 * ensure all the unknown bits are 0 - i.e. new
3750 * user-space does not rely on any kernel feature
3751 * extensions we dont know about yet.
3753 if (size > sizeof(*attr)) {
3754 unsigned char __user *addr;
3755 unsigned char __user *end;
3758 addr = (void __user *)uattr + sizeof(*attr);
3759 end = (void __user *)uattr + size;
3761 for (; addr < end; addr++) {
3762 ret = get_user(val, addr);
3768 size = sizeof(*attr);
3771 ret = copy_from_user(attr, uattr, size);
3776 * XXX: do we want to be lenient like existing syscalls; or do we want
3777 * to be strict and return an error on out-of-bounds values?
3779 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3784 put_user(sizeof(*attr), &uattr->size);
3789 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3790 * @pid: the pid in question.
3791 * @policy: new policy.
3792 * @param: structure containing the new RT priority.
3794 * Return: 0 on success. An error code otherwise.
3796 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3797 struct sched_param __user *, param)
3799 /* negative values for policy are not valid */
3803 return do_sched_setscheduler(pid, policy, param);
3807 * sys_sched_setparam - set/change the RT priority of a thread
3808 * @pid: the pid in question.
3809 * @param: structure containing the new RT priority.
3811 * Return: 0 on success. An error code otherwise.
3813 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3815 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3819 * sys_sched_setattr - same as above, but with extended sched_attr
3820 * @pid: the pid in question.
3821 * @uattr: structure containing the extended parameters.
3822 * @flags: for future extension.
3824 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3825 unsigned int, flags)
3827 struct sched_attr attr;
3828 struct task_struct *p;
3831 if (!uattr || pid < 0 || flags)
3834 retval = sched_copy_attr(uattr, &attr);
3838 if ((int)attr.sched_policy < 0)
3843 p = find_process_by_pid(pid);
3845 retval = sched_setattr(p, &attr);
3852 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3853 * @pid: the pid in question.
3855 * Return: On success, the policy of the thread. Otherwise, a negative error
3858 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3860 struct task_struct *p;
3868 p = find_process_by_pid(pid);
3870 retval = security_task_getscheduler(p);
3873 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3880 * sys_sched_getparam - get the RT priority of a thread
3881 * @pid: the pid in question.
3882 * @param: structure containing the RT priority.
3884 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3887 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3889 struct sched_param lp = { .sched_priority = 0 };
3890 struct task_struct *p;
3893 if (!param || pid < 0)
3897 p = find_process_by_pid(pid);
3902 retval = security_task_getscheduler(p);
3906 if (task_has_rt_policy(p))
3907 lp.sched_priority = p->rt_priority;
3911 * This one might sleep, we cannot do it with a spinlock held ...
3913 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3922 static int sched_read_attr(struct sched_attr __user *uattr,
3923 struct sched_attr *attr,
3928 if (!access_ok(VERIFY_WRITE, uattr, usize))
3932 * If we're handed a smaller struct than we know of,
3933 * ensure all the unknown bits are 0 - i.e. old
3934 * user-space does not get uncomplete information.
3936 if (usize < sizeof(*attr)) {
3937 unsigned char *addr;
3940 addr = (void *)attr + usize;
3941 end = (void *)attr + sizeof(*attr);
3943 for (; addr < end; addr++) {
3951 ret = copy_to_user(uattr, attr, attr->size);
3959 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3960 * @pid: the pid in question.
3961 * @uattr: structure containing the extended parameters.
3962 * @size: sizeof(attr) for fwd/bwd comp.
3963 * @flags: for future extension.
3965 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3966 unsigned int, size, unsigned int, flags)
3968 struct sched_attr attr = {
3969 .size = sizeof(struct sched_attr),
3971 struct task_struct *p;
3974 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3975 size < SCHED_ATTR_SIZE_VER0 || flags)
3979 p = find_process_by_pid(pid);
3984 retval = security_task_getscheduler(p);
3988 attr.sched_policy = p->policy;
3989 if (p->sched_reset_on_fork)
3990 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3991 if (task_has_dl_policy(p))
3992 __getparam_dl(p, &attr);
3993 else if (task_has_rt_policy(p))
3994 attr.sched_priority = p->rt_priority;
3996 attr.sched_nice = task_nice(p);
4000 retval = sched_read_attr(uattr, &attr, size);
4008 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4010 cpumask_var_t cpus_allowed, new_mask;
4011 struct task_struct *p;
4016 p = find_process_by_pid(pid);
4022 /* Prevent p going away */
4026 if (p->flags & PF_NO_SETAFFINITY) {
4030 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4034 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4036 goto out_free_cpus_allowed;
4039 if (!check_same_owner(p)) {
4041 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4043 goto out_free_new_mask;
4048 retval = security_task_setscheduler(p);
4050 goto out_free_new_mask;
4053 cpuset_cpus_allowed(p, cpus_allowed);
4054 cpumask_and(new_mask, in_mask, cpus_allowed);
4057 * Since bandwidth control happens on root_domain basis,
4058 * if admission test is enabled, we only admit -deadline
4059 * tasks allowed to run on all the CPUs in the task's
4063 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4065 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4068 goto out_free_new_mask;
4074 retval = set_cpus_allowed_ptr(p, new_mask);
4077 cpuset_cpus_allowed(p, cpus_allowed);
4078 if (!cpumask_subset(new_mask, cpus_allowed)) {
4080 * We must have raced with a concurrent cpuset
4081 * update. Just reset the cpus_allowed to the
4082 * cpuset's cpus_allowed
4084 cpumask_copy(new_mask, cpus_allowed);
4089 free_cpumask_var(new_mask);
4090 out_free_cpus_allowed:
4091 free_cpumask_var(cpus_allowed);
4097 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4098 struct cpumask *new_mask)
4100 if (len < cpumask_size())
4101 cpumask_clear(new_mask);
4102 else if (len > cpumask_size())
4103 len = cpumask_size();
4105 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4109 * sys_sched_setaffinity - set the cpu affinity of a process
4110 * @pid: pid of the process
4111 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4112 * @user_mask_ptr: user-space pointer to the new cpu mask
4114 * Return: 0 on success. An error code otherwise.
4116 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4117 unsigned long __user *, user_mask_ptr)
4119 cpumask_var_t new_mask;
4122 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4125 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4127 retval = sched_setaffinity(pid, new_mask);
4128 free_cpumask_var(new_mask);
4132 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4134 struct task_struct *p;
4135 unsigned long flags;
4141 p = find_process_by_pid(pid);
4145 retval = security_task_getscheduler(p);
4149 raw_spin_lock_irqsave(&p->pi_lock, flags);
4150 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4151 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4160 * sys_sched_getaffinity - get the cpu affinity of a process
4161 * @pid: pid of the process
4162 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4163 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4165 * Return: 0 on success. An error code otherwise.
4167 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4168 unsigned long __user *, user_mask_ptr)
4173 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4175 if (len & (sizeof(unsigned long)-1))
4178 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4181 ret = sched_getaffinity(pid, mask);
4183 size_t retlen = min_t(size_t, len, cpumask_size());
4185 if (copy_to_user(user_mask_ptr, mask, retlen))
4190 free_cpumask_var(mask);
4196 * sys_sched_yield - yield the current processor to other threads.
4198 * This function yields the current CPU to other tasks. If there are no
4199 * other threads running on this CPU then this function will return.
4203 SYSCALL_DEFINE0(sched_yield)
4205 struct rq *rq = this_rq_lock();
4207 schedstat_inc(rq, yld_count);
4208 current->sched_class->yield_task(rq);
4211 * Since we are going to call schedule() anyway, there's
4212 * no need to preempt or enable interrupts:
4214 __release(rq->lock);
4215 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4216 do_raw_spin_unlock(&rq->lock);
4217 sched_preempt_enable_no_resched();
4224 int __sched _cond_resched(void)
4226 if (should_resched()) {
4227 preempt_schedule_common();
4232 EXPORT_SYMBOL(_cond_resched);
4235 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4236 * call schedule, and on return reacquire the lock.
4238 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4239 * operations here to prevent schedule() from being called twice (once via
4240 * spin_unlock(), once by hand).
4242 int __cond_resched_lock(spinlock_t *lock)
4244 int resched = should_resched();
4247 lockdep_assert_held(lock);
4249 if (spin_needbreak(lock) || resched) {
4252 preempt_schedule_common();
4260 EXPORT_SYMBOL(__cond_resched_lock);
4262 int __sched __cond_resched_softirq(void)
4264 BUG_ON(!in_softirq());
4266 if (should_resched()) {
4268 preempt_schedule_common();
4274 EXPORT_SYMBOL(__cond_resched_softirq);
4277 * yield - yield the current processor to other threads.
4279 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4281 * The scheduler is at all times free to pick the calling task as the most
4282 * eligible task to run, if removing the yield() call from your code breaks
4283 * it, its already broken.
4285 * Typical broken usage is:
4290 * where one assumes that yield() will let 'the other' process run that will
4291 * make event true. If the current task is a SCHED_FIFO task that will never
4292 * happen. Never use yield() as a progress guarantee!!
4294 * If you want to use yield() to wait for something, use wait_event().
4295 * If you want to use yield() to be 'nice' for others, use cond_resched().
4296 * If you still want to use yield(), do not!
4298 void __sched yield(void)
4300 set_current_state(TASK_RUNNING);
4303 EXPORT_SYMBOL(yield);
4306 * yield_to - yield the current processor to another thread in
4307 * your thread group, or accelerate that thread toward the
4308 * processor it's on.
4310 * @preempt: whether task preemption is allowed or not
4312 * It's the caller's job to ensure that the target task struct
4313 * can't go away on us before we can do any checks.
4316 * true (>0) if we indeed boosted the target task.
4317 * false (0) if we failed to boost the target.
4318 * -ESRCH if there's no task to yield to.
4320 int __sched yield_to(struct task_struct *p, bool preempt)
4322 struct task_struct *curr = current;
4323 struct rq *rq, *p_rq;
4324 unsigned long flags;
4327 local_irq_save(flags);
4333 * If we're the only runnable task on the rq and target rq also
4334 * has only one task, there's absolutely no point in yielding.
4336 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4341 double_rq_lock(rq, p_rq);
4342 if (task_rq(p) != p_rq) {
4343 double_rq_unlock(rq, p_rq);
4347 if (!curr->sched_class->yield_to_task)
4350 if (curr->sched_class != p->sched_class)
4353 if (task_running(p_rq, p) || p->state)
4356 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4358 schedstat_inc(rq, yld_count);
4360 * Make p's CPU reschedule; pick_next_entity takes care of
4363 if (preempt && rq != p_rq)
4368 double_rq_unlock(rq, p_rq);
4370 local_irq_restore(flags);
4377 EXPORT_SYMBOL_GPL(yield_to);
4380 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4381 * that process accounting knows that this is a task in IO wait state.
4383 long __sched io_schedule_timeout(long timeout)
4385 int old_iowait = current->in_iowait;
4389 current->in_iowait = 1;
4391 blk_schedule_flush_plug(current);
4393 blk_flush_plug(current);
4395 delayacct_blkio_start();
4397 atomic_inc(&rq->nr_iowait);
4398 ret = schedule_timeout(timeout);
4399 current->in_iowait = old_iowait;
4400 atomic_dec(&rq->nr_iowait);
4401 delayacct_blkio_end();
4405 EXPORT_SYMBOL(io_schedule_timeout);
4408 * sys_sched_get_priority_max - return maximum RT priority.
4409 * @policy: scheduling class.
4411 * Return: On success, this syscall returns the maximum
4412 * rt_priority that can be used by a given scheduling class.
4413 * On failure, a negative error code is returned.
4415 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4422 ret = MAX_USER_RT_PRIO-1;
4424 case SCHED_DEADLINE:
4435 * sys_sched_get_priority_min - return minimum RT priority.
4436 * @policy: scheduling class.
4438 * Return: On success, this syscall returns the minimum
4439 * rt_priority that can be used by a given scheduling class.
4440 * On failure, a negative error code is returned.
4442 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4451 case SCHED_DEADLINE:
4461 * sys_sched_rr_get_interval - return the default timeslice of a process.
4462 * @pid: pid of the process.
4463 * @interval: userspace pointer to the timeslice value.
4465 * this syscall writes the default timeslice value of a given process
4466 * into the user-space timespec buffer. A value of '0' means infinity.
4468 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4471 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4472 struct timespec __user *, interval)
4474 struct task_struct *p;
4475 unsigned int time_slice;
4476 unsigned long flags;
4486 p = find_process_by_pid(pid);
4490 retval = security_task_getscheduler(p);
4494 rq = task_rq_lock(p, &flags);
4496 if (p->sched_class->get_rr_interval)
4497 time_slice = p->sched_class->get_rr_interval(rq, p);
4498 task_rq_unlock(rq, p, &flags);
4501 jiffies_to_timespec(time_slice, &t);
4502 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4510 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4512 void sched_show_task(struct task_struct *p)
4514 unsigned long free = 0;
4516 unsigned long state = p->state;
4519 state = __ffs(state) + 1;
4520 printk(KERN_INFO "%-15.15s %c", p->comm,
4521 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4522 #if BITS_PER_LONG == 32
4523 if (state == TASK_RUNNING)
4524 printk(KERN_CONT " running ");
4526 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4528 if (state == TASK_RUNNING)
4529 printk(KERN_CONT " running task ");
4531 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4533 #ifdef CONFIG_DEBUG_STACK_USAGE
4534 free = stack_not_used(p);
4539 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4541 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4542 task_pid_nr(p), ppid,
4543 (unsigned long)task_thread_info(p)->flags);
4545 print_worker_info(KERN_INFO, p);
4546 show_stack(p, NULL);
4549 void show_state_filter(unsigned long state_filter)
4551 struct task_struct *g, *p;
4553 #if BITS_PER_LONG == 32
4555 " task PC stack pid father\n");
4558 " task PC stack pid father\n");
4561 for_each_process_thread(g, p) {
4563 * reset the NMI-timeout, listing all files on a slow
4564 * console might take a lot of time:
4566 touch_nmi_watchdog();
4567 if (!state_filter || (p->state & state_filter))
4571 touch_all_softlockup_watchdogs();
4573 #ifdef CONFIG_SCHED_DEBUG
4574 sysrq_sched_debug_show();
4578 * Only show locks if all tasks are dumped:
4581 debug_show_all_locks();
4584 void init_idle_bootup_task(struct task_struct *idle)
4586 idle->sched_class = &idle_sched_class;
4590 * init_idle - set up an idle thread for a given CPU
4591 * @idle: task in question
4592 * @cpu: cpu the idle task belongs to
4594 * NOTE: this function does not set the idle thread's NEED_RESCHED
4595 * flag, to make booting more robust.
4597 void init_idle(struct task_struct *idle, int cpu)
4599 struct rq *rq = cpu_rq(cpu);
4600 unsigned long flags;
4602 raw_spin_lock_irqsave(&rq->lock, flags);
4604 __sched_fork(0, idle);
4605 idle->state = TASK_RUNNING;
4606 idle->se.exec_start = sched_clock();
4608 do_set_cpus_allowed(idle, cpumask_of(cpu));
4610 * We're having a chicken and egg problem, even though we are
4611 * holding rq->lock, the cpu isn't yet set to this cpu so the
4612 * lockdep check in task_group() will fail.
4614 * Similar case to sched_fork(). / Alternatively we could
4615 * use task_rq_lock() here and obtain the other rq->lock.
4620 __set_task_cpu(idle, cpu);
4623 rq->curr = rq->idle = idle;
4624 idle->on_rq = TASK_ON_RQ_QUEUED;
4625 #if defined(CONFIG_SMP)
4628 raw_spin_unlock_irqrestore(&rq->lock, flags);
4630 /* Set the preempt count _outside_ the spinlocks! */
4631 init_idle_preempt_count(idle, cpu);
4634 * The idle tasks have their own, simple scheduling class:
4636 idle->sched_class = &idle_sched_class;
4637 ftrace_graph_init_idle_task(idle, cpu);
4638 vtime_init_idle(idle, cpu);
4639 #if defined(CONFIG_SMP)
4640 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4644 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4645 const struct cpumask *trial)
4647 int ret = 1, trial_cpus;
4648 struct dl_bw *cur_dl_b;
4649 unsigned long flags;
4651 if (!cpumask_weight(cur))
4654 rcu_read_lock_sched();
4655 cur_dl_b = dl_bw_of(cpumask_any(cur));
4656 trial_cpus = cpumask_weight(trial);
4658 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4659 if (cur_dl_b->bw != -1 &&
4660 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4662 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4663 rcu_read_unlock_sched();
4668 int task_can_attach(struct task_struct *p,
4669 const struct cpumask *cs_cpus_allowed)
4674 * Kthreads which disallow setaffinity shouldn't be moved
4675 * to a new cpuset; we don't want to change their cpu
4676 * affinity and isolating such threads by their set of
4677 * allowed nodes is unnecessary. Thus, cpusets are not
4678 * applicable for such threads. This prevents checking for
4679 * success of set_cpus_allowed_ptr() on all attached tasks
4680 * before cpus_allowed may be changed.
4682 if (p->flags & PF_NO_SETAFFINITY) {
4688 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4690 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4695 unsigned long flags;
4697 rcu_read_lock_sched();
4698 dl_b = dl_bw_of(dest_cpu);
4699 raw_spin_lock_irqsave(&dl_b->lock, flags);
4700 cpus = dl_bw_cpus(dest_cpu);
4701 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4706 * We reserve space for this task in the destination
4707 * root_domain, as we can't fail after this point.
4708 * We will free resources in the source root_domain
4709 * later on (see set_cpus_allowed_dl()).
4711 __dl_add(dl_b, p->dl.dl_bw);
4713 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4714 rcu_read_unlock_sched();
4724 * move_queued_task - move a queued task to new rq.
4726 * Returns (locked) new rq. Old rq's lock is released.
4728 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4730 struct rq *rq = task_rq(p);
4732 lockdep_assert_held(&rq->lock);
4734 dequeue_task(rq, p, 0);
4735 p->on_rq = TASK_ON_RQ_MIGRATING;
4736 set_task_cpu(p, new_cpu);
4737 raw_spin_unlock(&rq->lock);
4739 rq = cpu_rq(new_cpu);
4741 raw_spin_lock(&rq->lock);
4742 BUG_ON(task_cpu(p) != new_cpu);
4743 p->on_rq = TASK_ON_RQ_QUEUED;
4744 enqueue_task(rq, p, 0);
4745 check_preempt_curr(rq, p, 0);
4750 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4752 if (p->sched_class->set_cpus_allowed)
4753 p->sched_class->set_cpus_allowed(p, new_mask);
4755 cpumask_copy(&p->cpus_allowed, new_mask);
4756 p->nr_cpus_allowed = cpumask_weight(new_mask);
4760 * This is how migration works:
4762 * 1) we invoke migration_cpu_stop() on the target CPU using
4764 * 2) stopper starts to run (implicitly forcing the migrated thread
4766 * 3) it checks whether the migrated task is still in the wrong runqueue.
4767 * 4) if it's in the wrong runqueue then the migration thread removes
4768 * it and puts it into the right queue.
4769 * 5) stopper completes and stop_one_cpu() returns and the migration
4774 * Change a given task's CPU affinity. Migrate the thread to a
4775 * proper CPU and schedule it away if the CPU it's executing on
4776 * is removed from the allowed bitmask.
4778 * NOTE: the caller must have a valid reference to the task, the
4779 * task must not exit() & deallocate itself prematurely. The
4780 * call is not atomic; no spinlocks may be held.
4782 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4784 unsigned long flags;
4786 unsigned int dest_cpu;
4789 rq = task_rq_lock(p, &flags);
4791 if (cpumask_equal(&p->cpus_allowed, new_mask))
4794 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4799 do_set_cpus_allowed(p, new_mask);
4801 /* Can the task run on the task's current CPU? If so, we're done */
4802 if (cpumask_test_cpu(task_cpu(p), new_mask))
4805 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4806 if (task_running(rq, p) || p->state == TASK_WAKING) {
4807 struct migration_arg arg = { p, dest_cpu };
4808 /* Need help from migration thread: drop lock and wait. */
4809 task_rq_unlock(rq, p, &flags);
4810 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4811 tlb_migrate_finish(p->mm);
4813 } else if (task_on_rq_queued(p))
4814 rq = move_queued_task(p, dest_cpu);
4816 task_rq_unlock(rq, p, &flags);
4820 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4823 * Move (not current) task off this cpu, onto dest cpu. We're doing
4824 * this because either it can't run here any more (set_cpus_allowed()
4825 * away from this CPU, or CPU going down), or because we're
4826 * attempting to rebalance this task on exec (sched_exec).
4828 * So we race with normal scheduler movements, but that's OK, as long
4829 * as the task is no longer on this CPU.
4831 * Returns non-zero if task was successfully migrated.
4833 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4838 if (unlikely(!cpu_active(dest_cpu)))
4841 rq = cpu_rq(src_cpu);
4843 raw_spin_lock(&p->pi_lock);
4844 raw_spin_lock(&rq->lock);
4845 /* Already moved. */
4846 if (task_cpu(p) != src_cpu)
4849 /* Affinity changed (again). */
4850 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4854 * If we're not on a rq, the next wake-up will ensure we're
4857 if (task_on_rq_queued(p))
4858 rq = move_queued_task(p, dest_cpu);
4862 raw_spin_unlock(&rq->lock);
4863 raw_spin_unlock(&p->pi_lock);
4867 #ifdef CONFIG_NUMA_BALANCING
4868 /* Migrate current task p to target_cpu */
4869 int migrate_task_to(struct task_struct *p, int target_cpu)
4871 struct migration_arg arg = { p, target_cpu };
4872 int curr_cpu = task_cpu(p);
4874 if (curr_cpu == target_cpu)
4877 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4880 /* TODO: This is not properly updating schedstats */
4882 trace_sched_move_numa(p, curr_cpu, target_cpu);
4883 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4887 * Requeue a task on a given node and accurately track the number of NUMA
4888 * tasks on the runqueues
4890 void sched_setnuma(struct task_struct *p, int nid)
4893 unsigned long flags;
4894 bool queued, running;
4896 rq = task_rq_lock(p, &flags);
4897 queued = task_on_rq_queued(p);
4898 running = task_current(rq, p);
4901 dequeue_task(rq, p, 0);
4903 put_prev_task(rq, p);
4905 p->numa_preferred_nid = nid;
4908 p->sched_class->set_curr_task(rq);
4910 enqueue_task(rq, p, 0);
4911 task_rq_unlock(rq, p, &flags);
4916 * migration_cpu_stop - this will be executed by a highprio stopper thread
4917 * and performs thread migration by bumping thread off CPU then
4918 * 'pushing' onto another runqueue.
4920 static int migration_cpu_stop(void *data)
4922 struct migration_arg *arg = data;
4925 * The original target cpu might have gone down and we might
4926 * be on another cpu but it doesn't matter.
4928 local_irq_disable();
4930 * We need to explicitly wake pending tasks before running
4931 * __migrate_task() such that we will not miss enforcing cpus_allowed
4932 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4934 sched_ttwu_pending();
4935 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4940 #ifdef CONFIG_HOTPLUG_CPU
4943 * Ensures that the idle task is using init_mm right before its cpu goes
4946 void idle_task_exit(void)
4948 struct mm_struct *mm = current->active_mm;
4950 BUG_ON(cpu_online(smp_processor_id()));
4952 if (mm != &init_mm) {
4953 switch_mm(mm, &init_mm, current);
4954 finish_arch_post_lock_switch();
4960 * Since this CPU is going 'away' for a while, fold any nr_active delta
4961 * we might have. Assumes we're called after migrate_tasks() so that the
4962 * nr_active count is stable.
4964 * Also see the comment "Global load-average calculations".
4966 static void calc_load_migrate(struct rq *rq)
4968 long delta = calc_load_fold_active(rq);
4970 atomic_long_add(delta, &calc_load_tasks);
4973 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4977 static const struct sched_class fake_sched_class = {
4978 .put_prev_task = put_prev_task_fake,
4981 static struct task_struct fake_task = {
4983 * Avoid pull_{rt,dl}_task()
4985 .prio = MAX_PRIO + 1,
4986 .sched_class = &fake_sched_class,
4990 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4991 * try_to_wake_up()->select_task_rq().
4993 * Called with rq->lock held even though we'er in stop_machine() and
4994 * there's no concurrency possible, we hold the required locks anyway
4995 * because of lock validation efforts.
4997 static void migrate_tasks(unsigned int dead_cpu)
4999 struct rq *rq = cpu_rq(dead_cpu);
5000 struct task_struct *next, *stop = rq->stop;
5004 * Fudge the rq selection such that the below task selection loop
5005 * doesn't get stuck on the currently eligible stop task.
5007 * We're currently inside stop_machine() and the rq is either stuck
5008 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5009 * either way we should never end up calling schedule() until we're
5015 * put_prev_task() and pick_next_task() sched
5016 * class method both need to have an up-to-date
5017 * value of rq->clock[_task]
5019 update_rq_clock(rq);
5023 * There's this thread running, bail when that's the only
5026 if (rq->nr_running == 1)
5029 next = pick_next_task(rq, &fake_task);
5031 next->sched_class->put_prev_task(rq, next);
5033 /* Find suitable destination for @next, with force if needed. */
5034 dest_cpu = select_fallback_rq(dead_cpu, next);
5035 raw_spin_unlock(&rq->lock);
5037 __migrate_task(next, dead_cpu, dest_cpu);
5039 raw_spin_lock(&rq->lock);
5045 #endif /* CONFIG_HOTPLUG_CPU */
5047 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5049 static struct ctl_table sd_ctl_dir[] = {
5051 .procname = "sched_domain",
5057 static struct ctl_table sd_ctl_root[] = {
5059 .procname = "kernel",
5061 .child = sd_ctl_dir,
5066 static struct ctl_table *sd_alloc_ctl_entry(int n)
5068 struct ctl_table *entry =
5069 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5074 static void sd_free_ctl_entry(struct ctl_table **tablep)
5076 struct ctl_table *entry;
5079 * In the intermediate directories, both the child directory and
5080 * procname are dynamically allocated and could fail but the mode
5081 * will always be set. In the lowest directory the names are
5082 * static strings and all have proc handlers.
5084 for (entry = *tablep; entry->mode; entry++) {
5086 sd_free_ctl_entry(&entry->child);
5087 if (entry->proc_handler == NULL)
5088 kfree(entry->procname);
5095 static int min_load_idx = 0;
5096 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5099 set_table_entry(struct ctl_table *entry,
5100 const char *procname, void *data, int maxlen,
5101 umode_t mode, proc_handler *proc_handler,
5104 entry->procname = procname;
5106 entry->maxlen = maxlen;
5108 entry->proc_handler = proc_handler;
5111 entry->extra1 = &min_load_idx;
5112 entry->extra2 = &max_load_idx;
5116 static struct ctl_table *
5117 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5119 struct ctl_table *table = sd_alloc_ctl_entry(14);
5124 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5125 sizeof(long), 0644, proc_doulongvec_minmax, false);
5126 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5127 sizeof(long), 0644, proc_doulongvec_minmax, false);
5128 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5129 sizeof(int), 0644, proc_dointvec_minmax, true);
5130 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5131 sizeof(int), 0644, proc_dointvec_minmax, true);
5132 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5133 sizeof(int), 0644, proc_dointvec_minmax, true);
5134 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5135 sizeof(int), 0644, proc_dointvec_minmax, true);
5136 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5137 sizeof(int), 0644, proc_dointvec_minmax, true);
5138 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5139 sizeof(int), 0644, proc_dointvec_minmax, false);
5140 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5141 sizeof(int), 0644, proc_dointvec_minmax, false);
5142 set_table_entry(&table[9], "cache_nice_tries",
5143 &sd->cache_nice_tries,
5144 sizeof(int), 0644, proc_dointvec_minmax, false);
5145 set_table_entry(&table[10], "flags", &sd->flags,
5146 sizeof(int), 0644, proc_dointvec_minmax, false);
5147 set_table_entry(&table[11], "max_newidle_lb_cost",
5148 &sd->max_newidle_lb_cost,
5149 sizeof(long), 0644, proc_doulongvec_minmax, false);
5150 set_table_entry(&table[12], "name", sd->name,
5151 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5152 /* &table[13] is terminator */
5157 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5159 struct ctl_table *entry, *table;
5160 struct sched_domain *sd;
5161 int domain_num = 0, i;
5164 for_each_domain(cpu, sd)
5166 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5171 for_each_domain(cpu, sd) {
5172 snprintf(buf, 32, "domain%d", i);
5173 entry->procname = kstrdup(buf, GFP_KERNEL);
5175 entry->child = sd_alloc_ctl_domain_table(sd);
5182 static struct ctl_table_header *sd_sysctl_header;
5183 static void register_sched_domain_sysctl(void)
5185 int i, cpu_num = num_possible_cpus();
5186 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5189 WARN_ON(sd_ctl_dir[0].child);
5190 sd_ctl_dir[0].child = entry;
5195 for_each_possible_cpu(i) {
5196 snprintf(buf, 32, "cpu%d", i);
5197 entry->procname = kstrdup(buf, GFP_KERNEL);
5199 entry->child = sd_alloc_ctl_cpu_table(i);
5203 WARN_ON(sd_sysctl_header);
5204 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5207 /* may be called multiple times per register */
5208 static void unregister_sched_domain_sysctl(void)
5210 if (sd_sysctl_header)
5211 unregister_sysctl_table(sd_sysctl_header);
5212 sd_sysctl_header = NULL;
5213 if (sd_ctl_dir[0].child)
5214 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5217 static void register_sched_domain_sysctl(void)
5220 static void unregister_sched_domain_sysctl(void)
5225 static void set_rq_online(struct rq *rq)
5228 const struct sched_class *class;
5230 cpumask_set_cpu(rq->cpu, rq->rd->online);
5233 for_each_class(class) {
5234 if (class->rq_online)
5235 class->rq_online(rq);
5240 static void set_rq_offline(struct rq *rq)
5243 const struct sched_class *class;
5245 for_each_class(class) {
5246 if (class->rq_offline)
5247 class->rq_offline(rq);
5250 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5256 * migration_call - callback that gets triggered when a CPU is added.
5257 * Here we can start up the necessary migration thread for the new CPU.
5260 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5262 int cpu = (long)hcpu;
5263 unsigned long flags;
5264 struct rq *rq = cpu_rq(cpu);
5266 switch (action & ~CPU_TASKS_FROZEN) {
5268 case CPU_UP_PREPARE:
5269 rq->calc_load_update = calc_load_update;
5273 /* Update our root-domain */
5274 raw_spin_lock_irqsave(&rq->lock, flags);
5276 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5280 raw_spin_unlock_irqrestore(&rq->lock, flags);
5283 #ifdef CONFIG_HOTPLUG_CPU
5285 sched_ttwu_pending();
5286 /* Update our root-domain */
5287 raw_spin_lock_irqsave(&rq->lock, flags);
5289 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5293 BUG_ON(rq->nr_running != 1); /* the migration thread */
5294 raw_spin_unlock_irqrestore(&rq->lock, flags);
5298 calc_load_migrate(rq);
5303 update_max_interval();
5309 * Register at high priority so that task migration (migrate_all_tasks)
5310 * happens before everything else. This has to be lower priority than
5311 * the notifier in the perf_event subsystem, though.
5313 static struct notifier_block migration_notifier = {
5314 .notifier_call = migration_call,
5315 .priority = CPU_PRI_MIGRATION,
5318 static void __cpuinit set_cpu_rq_start_time(void)
5320 int cpu = smp_processor_id();
5321 struct rq *rq = cpu_rq(cpu);
5322 rq->age_stamp = sched_clock_cpu(cpu);
5325 static int sched_cpu_active(struct notifier_block *nfb,
5326 unsigned long action, void *hcpu)
5328 switch (action & ~CPU_TASKS_FROZEN) {
5330 set_cpu_rq_start_time();
5332 case CPU_DOWN_FAILED:
5333 set_cpu_active((long)hcpu, true);
5340 static int sched_cpu_inactive(struct notifier_block *nfb,
5341 unsigned long action, void *hcpu)
5343 switch (action & ~CPU_TASKS_FROZEN) {
5344 case CPU_DOWN_PREPARE:
5345 set_cpu_active((long)hcpu, false);
5352 static int __init migration_init(void)
5354 void *cpu = (void *)(long)smp_processor_id();
5357 /* Initialize migration for the boot CPU */
5358 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5359 BUG_ON(err == NOTIFY_BAD);
5360 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5361 register_cpu_notifier(&migration_notifier);
5363 /* Register cpu active notifiers */
5364 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5365 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5369 early_initcall(migration_init);
5374 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5376 #ifdef CONFIG_SCHED_DEBUG
5378 static __read_mostly int sched_debug_enabled;
5380 static int __init sched_debug_setup(char *str)
5382 sched_debug_enabled = 1;
5386 early_param("sched_debug", sched_debug_setup);
5388 static inline bool sched_debug(void)
5390 return sched_debug_enabled;
5393 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5394 struct cpumask *groupmask)
5396 struct sched_group *group = sd->groups;
5398 cpumask_clear(groupmask);
5400 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5402 if (!(sd->flags & SD_LOAD_BALANCE)) {
5403 printk("does not load-balance\n");
5405 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5410 printk(KERN_CONT "span %*pbl level %s\n",
5411 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5413 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5414 printk(KERN_ERR "ERROR: domain->span does not contain "
5417 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5418 printk(KERN_ERR "ERROR: domain->groups does not contain"
5422 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5426 printk(KERN_ERR "ERROR: group is NULL\n");
5430 if (!cpumask_weight(sched_group_cpus(group))) {
5431 printk(KERN_CONT "\n");
5432 printk(KERN_ERR "ERROR: empty group\n");
5436 if (!(sd->flags & SD_OVERLAP) &&
5437 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5438 printk(KERN_CONT "\n");
5439 printk(KERN_ERR "ERROR: repeated CPUs\n");
5443 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5445 printk(KERN_CONT " %*pbl",
5446 cpumask_pr_args(sched_group_cpus(group)));
5447 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5448 printk(KERN_CONT " (cpu_capacity = %d)",
5449 group->sgc->capacity);
5452 group = group->next;
5453 } while (group != sd->groups);
5454 printk(KERN_CONT "\n");
5456 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5457 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5460 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5461 printk(KERN_ERR "ERROR: parent span is not a superset "
5462 "of domain->span\n");
5466 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5470 if (!sched_debug_enabled)
5474 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5478 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5481 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5489 #else /* !CONFIG_SCHED_DEBUG */
5490 # define sched_domain_debug(sd, cpu) do { } while (0)
5491 static inline bool sched_debug(void)
5495 #endif /* CONFIG_SCHED_DEBUG */
5497 static int sd_degenerate(struct sched_domain *sd)
5499 if (cpumask_weight(sched_domain_span(sd)) == 1)
5502 /* Following flags need at least 2 groups */
5503 if (sd->flags & (SD_LOAD_BALANCE |
5504 SD_BALANCE_NEWIDLE |
5507 SD_SHARE_CPUCAPACITY |
5508 SD_SHARE_PKG_RESOURCES |
5509 SD_SHARE_POWERDOMAIN)) {
5510 if (sd->groups != sd->groups->next)
5514 /* Following flags don't use groups */
5515 if (sd->flags & (SD_WAKE_AFFINE))
5522 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5524 unsigned long cflags = sd->flags, pflags = parent->flags;
5526 if (sd_degenerate(parent))
5529 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5532 /* Flags needing groups don't count if only 1 group in parent */
5533 if (parent->groups == parent->groups->next) {
5534 pflags &= ~(SD_LOAD_BALANCE |
5535 SD_BALANCE_NEWIDLE |
5538 SD_SHARE_CPUCAPACITY |
5539 SD_SHARE_PKG_RESOURCES |
5541 SD_SHARE_POWERDOMAIN);
5542 if (nr_node_ids == 1)
5543 pflags &= ~SD_SERIALIZE;
5545 if (~cflags & pflags)
5551 static void free_rootdomain(struct rcu_head *rcu)
5553 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5555 cpupri_cleanup(&rd->cpupri);
5556 cpudl_cleanup(&rd->cpudl);
5557 free_cpumask_var(rd->dlo_mask);
5558 free_cpumask_var(rd->rto_mask);
5559 free_cpumask_var(rd->online);
5560 free_cpumask_var(rd->span);
5564 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5566 struct root_domain *old_rd = NULL;
5567 unsigned long flags;
5569 raw_spin_lock_irqsave(&rq->lock, flags);
5574 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5577 cpumask_clear_cpu(rq->cpu, old_rd->span);
5580 * If we dont want to free the old_rd yet then
5581 * set old_rd to NULL to skip the freeing later
5584 if (!atomic_dec_and_test(&old_rd->refcount))
5588 atomic_inc(&rd->refcount);
5591 cpumask_set_cpu(rq->cpu, rd->span);
5592 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5595 raw_spin_unlock_irqrestore(&rq->lock, flags);
5598 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5601 static int init_rootdomain(struct root_domain *rd)
5603 memset(rd, 0, sizeof(*rd));
5605 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5607 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5609 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5611 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5614 init_dl_bw(&rd->dl_bw);
5615 if (cpudl_init(&rd->cpudl) != 0)
5618 if (cpupri_init(&rd->cpupri) != 0)
5623 free_cpumask_var(rd->rto_mask);
5625 free_cpumask_var(rd->dlo_mask);
5627 free_cpumask_var(rd->online);
5629 free_cpumask_var(rd->span);
5635 * By default the system creates a single root-domain with all cpus as
5636 * members (mimicking the global state we have today).
5638 struct root_domain def_root_domain;
5640 static void init_defrootdomain(void)
5642 init_rootdomain(&def_root_domain);
5644 atomic_set(&def_root_domain.refcount, 1);
5647 static struct root_domain *alloc_rootdomain(void)
5649 struct root_domain *rd;
5651 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5655 if (init_rootdomain(rd) != 0) {
5663 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5665 struct sched_group *tmp, *first;
5674 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5679 } while (sg != first);
5682 static void free_sched_domain(struct rcu_head *rcu)
5684 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5687 * If its an overlapping domain it has private groups, iterate and
5690 if (sd->flags & SD_OVERLAP) {
5691 free_sched_groups(sd->groups, 1);
5692 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5693 kfree(sd->groups->sgc);
5699 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5701 call_rcu(&sd->rcu, free_sched_domain);
5704 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5706 for (; sd; sd = sd->parent)
5707 destroy_sched_domain(sd, cpu);
5711 * Keep a special pointer to the highest sched_domain that has
5712 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5713 * allows us to avoid some pointer chasing select_idle_sibling().
5715 * Also keep a unique ID per domain (we use the first cpu number in
5716 * the cpumask of the domain), this allows us to quickly tell if
5717 * two cpus are in the same cache domain, see cpus_share_cache().
5719 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5720 DEFINE_PER_CPU(int, sd_llc_size);
5721 DEFINE_PER_CPU(int, sd_llc_id);
5722 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5723 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5724 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5726 static void update_top_cache_domain(int cpu)
5728 struct sched_domain *sd;
5729 struct sched_domain *busy_sd = NULL;
5733 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5735 id = cpumask_first(sched_domain_span(sd));
5736 size = cpumask_weight(sched_domain_span(sd));
5737 busy_sd = sd->parent; /* sd_busy */
5739 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5741 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5742 per_cpu(sd_llc_size, cpu) = size;
5743 per_cpu(sd_llc_id, cpu) = id;
5745 sd = lowest_flag_domain(cpu, SD_NUMA);
5746 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5748 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5749 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5753 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5754 * hold the hotplug lock.
5757 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5759 struct rq *rq = cpu_rq(cpu);
5760 struct sched_domain *tmp;
5762 /* Remove the sched domains which do not contribute to scheduling. */
5763 for (tmp = sd; tmp; ) {
5764 struct sched_domain *parent = tmp->parent;
5768 if (sd_parent_degenerate(tmp, parent)) {
5769 tmp->parent = parent->parent;
5771 parent->parent->child = tmp;
5773 * Transfer SD_PREFER_SIBLING down in case of a
5774 * degenerate parent; the spans match for this
5775 * so the property transfers.
5777 if (parent->flags & SD_PREFER_SIBLING)
5778 tmp->flags |= SD_PREFER_SIBLING;
5779 destroy_sched_domain(parent, cpu);
5784 if (sd && sd_degenerate(sd)) {
5787 destroy_sched_domain(tmp, cpu);
5792 sched_domain_debug(sd, cpu);
5794 rq_attach_root(rq, rd);
5796 rcu_assign_pointer(rq->sd, sd);
5797 destroy_sched_domains(tmp, cpu);
5799 update_top_cache_domain(cpu);
5802 /* Setup the mask of cpus configured for isolated domains */
5803 static int __init isolated_cpu_setup(char *str)
5805 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5806 cpulist_parse(str, cpu_isolated_map);
5810 __setup("isolcpus=", isolated_cpu_setup);
5813 struct sched_domain ** __percpu sd;
5814 struct root_domain *rd;
5825 * Build an iteration mask that can exclude certain CPUs from the upwards
5828 * Asymmetric node setups can result in situations where the domain tree is of
5829 * unequal depth, make sure to skip domains that already cover the entire
5832 * In that case build_sched_domains() will have terminated the iteration early
5833 * and our sibling sd spans will be empty. Domains should always include the
5834 * cpu they're built on, so check that.
5837 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5839 const struct cpumask *span = sched_domain_span(sd);
5840 struct sd_data *sdd = sd->private;
5841 struct sched_domain *sibling;
5844 for_each_cpu(i, span) {
5845 sibling = *per_cpu_ptr(sdd->sd, i);
5846 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5849 cpumask_set_cpu(i, sched_group_mask(sg));
5854 * Return the canonical balance cpu for this group, this is the first cpu
5855 * of this group that's also in the iteration mask.
5857 int group_balance_cpu(struct sched_group *sg)
5859 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5863 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5865 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5866 const struct cpumask *span = sched_domain_span(sd);
5867 struct cpumask *covered = sched_domains_tmpmask;
5868 struct sd_data *sdd = sd->private;
5869 struct sched_domain *sibling;
5872 cpumask_clear(covered);
5874 for_each_cpu(i, span) {
5875 struct cpumask *sg_span;
5877 if (cpumask_test_cpu(i, covered))
5880 sibling = *per_cpu_ptr(sdd->sd, i);
5882 /* See the comment near build_group_mask(). */
5883 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5886 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5887 GFP_KERNEL, cpu_to_node(cpu));
5892 sg_span = sched_group_cpus(sg);
5894 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5896 cpumask_set_cpu(i, sg_span);
5898 cpumask_or(covered, covered, sg_span);
5900 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5901 if (atomic_inc_return(&sg->sgc->ref) == 1)
5902 build_group_mask(sd, sg);
5905 * Initialize sgc->capacity such that even if we mess up the
5906 * domains and no possible iteration will get us here, we won't
5909 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5912 * Make sure the first group of this domain contains the
5913 * canonical balance cpu. Otherwise the sched_domain iteration
5914 * breaks. See update_sg_lb_stats().
5916 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5917 group_balance_cpu(sg) == cpu)
5927 sd->groups = groups;
5932 free_sched_groups(first, 0);
5937 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5939 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5940 struct sched_domain *child = sd->child;
5943 cpu = cpumask_first(sched_domain_span(child));
5946 *sg = *per_cpu_ptr(sdd->sg, cpu);
5947 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5948 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5955 * build_sched_groups will build a circular linked list of the groups
5956 * covered by the given span, and will set each group's ->cpumask correctly,
5957 * and ->cpu_capacity to 0.
5959 * Assumes the sched_domain tree is fully constructed
5962 build_sched_groups(struct sched_domain *sd, int cpu)
5964 struct sched_group *first = NULL, *last = NULL;
5965 struct sd_data *sdd = sd->private;
5966 const struct cpumask *span = sched_domain_span(sd);
5967 struct cpumask *covered;
5970 get_group(cpu, sdd, &sd->groups);
5971 atomic_inc(&sd->groups->ref);
5973 if (cpu != cpumask_first(span))
5976 lockdep_assert_held(&sched_domains_mutex);
5977 covered = sched_domains_tmpmask;
5979 cpumask_clear(covered);
5981 for_each_cpu(i, span) {
5982 struct sched_group *sg;
5985 if (cpumask_test_cpu(i, covered))
5988 group = get_group(i, sdd, &sg);
5989 cpumask_setall(sched_group_mask(sg));
5991 for_each_cpu(j, span) {
5992 if (get_group(j, sdd, NULL) != group)
5995 cpumask_set_cpu(j, covered);
5996 cpumask_set_cpu(j, sched_group_cpus(sg));
6011 * Initialize sched groups cpu_capacity.
6013 * cpu_capacity indicates the capacity of sched group, which is used while
6014 * distributing the load between different sched groups in a sched domain.
6015 * Typically cpu_capacity for all the groups in a sched domain will be same
6016 * unless there are asymmetries in the topology. If there are asymmetries,
6017 * group having more cpu_capacity will pickup more load compared to the
6018 * group having less cpu_capacity.
6020 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6022 struct sched_group *sg = sd->groups;
6027 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6029 } while (sg != sd->groups);
6031 if (cpu != group_balance_cpu(sg))
6034 update_group_capacity(sd, cpu);
6035 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6039 * Initializers for schedule domains
6040 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6043 static int default_relax_domain_level = -1;
6044 int sched_domain_level_max;
6046 static int __init setup_relax_domain_level(char *str)
6048 if (kstrtoint(str, 0, &default_relax_domain_level))
6049 pr_warn("Unable to set relax_domain_level\n");
6053 __setup("relax_domain_level=", setup_relax_domain_level);
6055 static void set_domain_attribute(struct sched_domain *sd,
6056 struct sched_domain_attr *attr)
6060 if (!attr || attr->relax_domain_level < 0) {
6061 if (default_relax_domain_level < 0)
6064 request = default_relax_domain_level;
6066 request = attr->relax_domain_level;
6067 if (request < sd->level) {
6068 /* turn off idle balance on this domain */
6069 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6071 /* turn on idle balance on this domain */
6072 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6076 static void __sdt_free(const struct cpumask *cpu_map);
6077 static int __sdt_alloc(const struct cpumask *cpu_map);
6079 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6080 const struct cpumask *cpu_map)
6084 if (!atomic_read(&d->rd->refcount))
6085 free_rootdomain(&d->rd->rcu); /* fall through */
6087 free_percpu(d->sd); /* fall through */
6089 __sdt_free(cpu_map); /* fall through */
6095 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6096 const struct cpumask *cpu_map)
6098 memset(d, 0, sizeof(*d));
6100 if (__sdt_alloc(cpu_map))
6101 return sa_sd_storage;
6102 d->sd = alloc_percpu(struct sched_domain *);
6104 return sa_sd_storage;
6105 d->rd = alloc_rootdomain();
6108 return sa_rootdomain;
6112 * NULL the sd_data elements we've used to build the sched_domain and
6113 * sched_group structure so that the subsequent __free_domain_allocs()
6114 * will not free the data we're using.
6116 static void claim_allocations(int cpu, struct sched_domain *sd)
6118 struct sd_data *sdd = sd->private;
6120 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6121 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6123 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6124 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6126 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6127 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6131 static int sched_domains_numa_levels;
6132 enum numa_topology_type sched_numa_topology_type;
6133 static int *sched_domains_numa_distance;
6134 int sched_max_numa_distance;
6135 static struct cpumask ***sched_domains_numa_masks;
6136 static int sched_domains_curr_level;
6140 * SD_flags allowed in topology descriptions.
6142 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6143 * SD_SHARE_PKG_RESOURCES - describes shared caches
6144 * SD_NUMA - describes NUMA topologies
6145 * SD_SHARE_POWERDOMAIN - describes shared power domain
6148 * SD_ASYM_PACKING - describes SMT quirks
6150 #define TOPOLOGY_SD_FLAGS \
6151 (SD_SHARE_CPUCAPACITY | \
6152 SD_SHARE_PKG_RESOURCES | \
6155 SD_SHARE_POWERDOMAIN)
6157 static struct sched_domain *
6158 sd_init(struct sched_domain_topology_level *tl, int cpu)
6160 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6161 int sd_weight, sd_flags = 0;
6165 * Ugly hack to pass state to sd_numa_mask()...
6167 sched_domains_curr_level = tl->numa_level;
6170 sd_weight = cpumask_weight(tl->mask(cpu));
6173 sd_flags = (*tl->sd_flags)();
6174 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6175 "wrong sd_flags in topology description\n"))
6176 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6178 *sd = (struct sched_domain){
6179 .min_interval = sd_weight,
6180 .max_interval = 2*sd_weight,
6182 .imbalance_pct = 125,
6184 .cache_nice_tries = 0,
6191 .flags = 1*SD_LOAD_BALANCE
6192 | 1*SD_BALANCE_NEWIDLE
6197 | 0*SD_SHARE_CPUCAPACITY
6198 | 0*SD_SHARE_PKG_RESOURCES
6200 | 0*SD_PREFER_SIBLING
6205 .last_balance = jiffies,
6206 .balance_interval = sd_weight,
6208 .max_newidle_lb_cost = 0,
6209 .next_decay_max_lb_cost = jiffies,
6210 #ifdef CONFIG_SCHED_DEBUG
6216 * Convert topological properties into behaviour.
6219 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6220 sd->flags |= SD_PREFER_SIBLING;
6221 sd->imbalance_pct = 110;
6222 sd->smt_gain = 1178; /* ~15% */
6224 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6225 sd->imbalance_pct = 117;
6226 sd->cache_nice_tries = 1;
6230 } else if (sd->flags & SD_NUMA) {
6231 sd->cache_nice_tries = 2;
6235 sd->flags |= SD_SERIALIZE;
6236 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6237 sd->flags &= ~(SD_BALANCE_EXEC |
6244 sd->flags |= SD_PREFER_SIBLING;
6245 sd->cache_nice_tries = 1;
6250 sd->private = &tl->data;
6256 * Topology list, bottom-up.
6258 static struct sched_domain_topology_level default_topology[] = {
6259 #ifdef CONFIG_SCHED_SMT
6260 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6262 #ifdef CONFIG_SCHED_MC
6263 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6265 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6269 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6271 #define for_each_sd_topology(tl) \
6272 for (tl = sched_domain_topology; tl->mask; tl++)
6274 void set_sched_topology(struct sched_domain_topology_level *tl)
6276 sched_domain_topology = tl;
6281 static const struct cpumask *sd_numa_mask(int cpu)
6283 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6286 static void sched_numa_warn(const char *str)
6288 static int done = false;
6296 printk(KERN_WARNING "ERROR: %s\n\n", str);
6298 for (i = 0; i < nr_node_ids; i++) {
6299 printk(KERN_WARNING " ");
6300 for (j = 0; j < nr_node_ids; j++)
6301 printk(KERN_CONT "%02d ", node_distance(i,j));
6302 printk(KERN_CONT "\n");
6304 printk(KERN_WARNING "\n");
6307 bool find_numa_distance(int distance)
6311 if (distance == node_distance(0, 0))
6314 for (i = 0; i < sched_domains_numa_levels; i++) {
6315 if (sched_domains_numa_distance[i] == distance)
6323 * A system can have three types of NUMA topology:
6324 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6325 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6326 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6328 * The difference between a glueless mesh topology and a backplane
6329 * topology lies in whether communication between not directly
6330 * connected nodes goes through intermediary nodes (where programs
6331 * could run), or through backplane controllers. This affects
6332 * placement of programs.
6334 * The type of topology can be discerned with the following tests:
6335 * - If the maximum distance between any nodes is 1 hop, the system
6336 * is directly connected.
6337 * - If for two nodes A and B, located N > 1 hops away from each other,
6338 * there is an intermediary node C, which is < N hops away from both
6339 * nodes A and B, the system is a glueless mesh.
6341 static void init_numa_topology_type(void)
6345 n = sched_max_numa_distance;
6348 sched_numa_topology_type = NUMA_DIRECT;
6350 for_each_online_node(a) {
6351 for_each_online_node(b) {
6352 /* Find two nodes furthest removed from each other. */
6353 if (node_distance(a, b) < n)
6356 /* Is there an intermediary node between a and b? */
6357 for_each_online_node(c) {
6358 if (node_distance(a, c) < n &&
6359 node_distance(b, c) < n) {
6360 sched_numa_topology_type =
6366 sched_numa_topology_type = NUMA_BACKPLANE;
6372 static void sched_init_numa(void)
6374 int next_distance, curr_distance = node_distance(0, 0);
6375 struct sched_domain_topology_level *tl;
6379 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6380 if (!sched_domains_numa_distance)
6384 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6385 * unique distances in the node_distance() table.
6387 * Assumes node_distance(0,j) includes all distances in
6388 * node_distance(i,j) in order to avoid cubic time.
6390 next_distance = curr_distance;
6391 for (i = 0; i < nr_node_ids; i++) {
6392 for (j = 0; j < nr_node_ids; j++) {
6393 for (k = 0; k < nr_node_ids; k++) {
6394 int distance = node_distance(i, k);
6396 if (distance > curr_distance &&
6397 (distance < next_distance ||
6398 next_distance == curr_distance))
6399 next_distance = distance;
6402 * While not a strong assumption it would be nice to know
6403 * about cases where if node A is connected to B, B is not
6404 * equally connected to A.
6406 if (sched_debug() && node_distance(k, i) != distance)
6407 sched_numa_warn("Node-distance not symmetric");
6409 if (sched_debug() && i && !find_numa_distance(distance))
6410 sched_numa_warn("Node-0 not representative");
6412 if (next_distance != curr_distance) {
6413 sched_domains_numa_distance[level++] = next_distance;
6414 sched_domains_numa_levels = level;
6415 curr_distance = next_distance;
6420 * In case of sched_debug() we verify the above assumption.
6430 * 'level' contains the number of unique distances, excluding the
6431 * identity distance node_distance(i,i).
6433 * The sched_domains_numa_distance[] array includes the actual distance
6438 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6439 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6440 * the array will contain less then 'level' members. This could be
6441 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6442 * in other functions.
6444 * We reset it to 'level' at the end of this function.
6446 sched_domains_numa_levels = 0;
6448 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6449 if (!sched_domains_numa_masks)
6453 * Now for each level, construct a mask per node which contains all
6454 * cpus of nodes that are that many hops away from us.
6456 for (i = 0; i < level; i++) {
6457 sched_domains_numa_masks[i] =
6458 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6459 if (!sched_domains_numa_masks[i])
6462 for (j = 0; j < nr_node_ids; j++) {
6463 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6467 sched_domains_numa_masks[i][j] = mask;
6469 for (k = 0; k < nr_node_ids; k++) {
6470 if (node_distance(j, k) > sched_domains_numa_distance[i])
6473 cpumask_or(mask, mask, cpumask_of_node(k));
6478 /* Compute default topology size */
6479 for (i = 0; sched_domain_topology[i].mask; i++);
6481 tl = kzalloc((i + level + 1) *
6482 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6487 * Copy the default topology bits..
6489 for (i = 0; sched_domain_topology[i].mask; i++)
6490 tl[i] = sched_domain_topology[i];
6493 * .. and append 'j' levels of NUMA goodness.
6495 for (j = 0; j < level; i++, j++) {
6496 tl[i] = (struct sched_domain_topology_level){
6497 .mask = sd_numa_mask,
6498 .sd_flags = cpu_numa_flags,
6499 .flags = SDTL_OVERLAP,
6505 sched_domain_topology = tl;
6507 sched_domains_numa_levels = level;
6508 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6510 init_numa_topology_type();
6513 static void sched_domains_numa_masks_set(int cpu)
6516 int node = cpu_to_node(cpu);
6518 for (i = 0; i < sched_domains_numa_levels; i++) {
6519 for (j = 0; j < nr_node_ids; j++) {
6520 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6521 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6526 static void sched_domains_numa_masks_clear(int cpu)
6529 for (i = 0; i < sched_domains_numa_levels; i++) {
6530 for (j = 0; j < nr_node_ids; j++)
6531 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6536 * Update sched_domains_numa_masks[level][node] array when new cpus
6539 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6540 unsigned long action,
6543 int cpu = (long)hcpu;
6545 switch (action & ~CPU_TASKS_FROZEN) {
6547 sched_domains_numa_masks_set(cpu);
6551 sched_domains_numa_masks_clear(cpu);
6561 static inline void sched_init_numa(void)
6565 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6566 unsigned long action,
6571 #endif /* CONFIG_NUMA */
6573 static int __sdt_alloc(const struct cpumask *cpu_map)
6575 struct sched_domain_topology_level *tl;
6578 for_each_sd_topology(tl) {
6579 struct sd_data *sdd = &tl->data;
6581 sdd->sd = alloc_percpu(struct sched_domain *);
6585 sdd->sg = alloc_percpu(struct sched_group *);
6589 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6593 for_each_cpu(j, cpu_map) {
6594 struct sched_domain *sd;
6595 struct sched_group *sg;
6596 struct sched_group_capacity *sgc;
6598 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6599 GFP_KERNEL, cpu_to_node(j));
6603 *per_cpu_ptr(sdd->sd, j) = sd;
6605 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6606 GFP_KERNEL, cpu_to_node(j));
6612 *per_cpu_ptr(sdd->sg, j) = sg;
6614 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6615 GFP_KERNEL, cpu_to_node(j));
6619 *per_cpu_ptr(sdd->sgc, j) = sgc;
6626 static void __sdt_free(const struct cpumask *cpu_map)
6628 struct sched_domain_topology_level *tl;
6631 for_each_sd_topology(tl) {
6632 struct sd_data *sdd = &tl->data;
6634 for_each_cpu(j, cpu_map) {
6635 struct sched_domain *sd;
6638 sd = *per_cpu_ptr(sdd->sd, j);
6639 if (sd && (sd->flags & SD_OVERLAP))
6640 free_sched_groups(sd->groups, 0);
6641 kfree(*per_cpu_ptr(sdd->sd, j));
6645 kfree(*per_cpu_ptr(sdd->sg, j));
6647 kfree(*per_cpu_ptr(sdd->sgc, j));
6649 free_percpu(sdd->sd);
6651 free_percpu(sdd->sg);
6653 free_percpu(sdd->sgc);
6658 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6659 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6660 struct sched_domain *child, int cpu)
6662 struct sched_domain *sd = sd_init(tl, cpu);
6666 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6668 sd->level = child->level + 1;
6669 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6673 if (!cpumask_subset(sched_domain_span(child),
6674 sched_domain_span(sd))) {
6675 pr_err("BUG: arch topology borken\n");
6676 #ifdef CONFIG_SCHED_DEBUG
6677 pr_err(" the %s domain not a subset of the %s domain\n",
6678 child->name, sd->name);
6680 /* Fixup, ensure @sd has at least @child cpus. */
6681 cpumask_or(sched_domain_span(sd),
6682 sched_domain_span(sd),
6683 sched_domain_span(child));
6687 set_domain_attribute(sd, attr);
6693 * Build sched domains for a given set of cpus and attach the sched domains
6694 * to the individual cpus
6696 static int build_sched_domains(const struct cpumask *cpu_map,
6697 struct sched_domain_attr *attr)
6699 enum s_alloc alloc_state;
6700 struct sched_domain *sd;
6702 int i, ret = -ENOMEM;
6704 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6705 if (alloc_state != sa_rootdomain)
6708 /* Set up domains for cpus specified by the cpu_map. */
6709 for_each_cpu(i, cpu_map) {
6710 struct sched_domain_topology_level *tl;
6713 for_each_sd_topology(tl) {
6714 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6715 if (tl == sched_domain_topology)
6716 *per_cpu_ptr(d.sd, i) = sd;
6717 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6718 sd->flags |= SD_OVERLAP;
6719 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6724 /* Build the groups for the domains */
6725 for_each_cpu(i, cpu_map) {
6726 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6727 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6728 if (sd->flags & SD_OVERLAP) {
6729 if (build_overlap_sched_groups(sd, i))
6732 if (build_sched_groups(sd, i))
6738 /* Calculate CPU capacity for physical packages and nodes */
6739 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6740 if (!cpumask_test_cpu(i, cpu_map))
6743 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6744 claim_allocations(i, sd);
6745 init_sched_groups_capacity(i, sd);
6749 /* Attach the domains */
6751 for_each_cpu(i, cpu_map) {
6752 sd = *per_cpu_ptr(d.sd, i);
6753 cpu_attach_domain(sd, d.rd, i);
6759 __free_domain_allocs(&d, alloc_state, cpu_map);
6763 static cpumask_var_t *doms_cur; /* current sched domains */
6764 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6765 static struct sched_domain_attr *dattr_cur;
6766 /* attribues of custom domains in 'doms_cur' */
6769 * Special case: If a kmalloc of a doms_cur partition (array of
6770 * cpumask) fails, then fallback to a single sched domain,
6771 * as determined by the single cpumask fallback_doms.
6773 static cpumask_var_t fallback_doms;
6776 * arch_update_cpu_topology lets virtualized architectures update the
6777 * cpu core maps. It is supposed to return 1 if the topology changed
6778 * or 0 if it stayed the same.
6780 int __weak arch_update_cpu_topology(void)
6785 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6788 cpumask_var_t *doms;
6790 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6793 for (i = 0; i < ndoms; i++) {
6794 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6795 free_sched_domains(doms, i);
6802 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6805 for (i = 0; i < ndoms; i++)
6806 free_cpumask_var(doms[i]);
6811 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6812 * For now this just excludes isolated cpus, but could be used to
6813 * exclude other special cases in the future.
6815 static int init_sched_domains(const struct cpumask *cpu_map)
6819 arch_update_cpu_topology();
6821 doms_cur = alloc_sched_domains(ndoms_cur);
6823 doms_cur = &fallback_doms;
6824 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6825 err = build_sched_domains(doms_cur[0], NULL);
6826 register_sched_domain_sysctl();
6832 * Detach sched domains from a group of cpus specified in cpu_map
6833 * These cpus will now be attached to the NULL domain
6835 static void detach_destroy_domains(const struct cpumask *cpu_map)
6840 for_each_cpu(i, cpu_map)
6841 cpu_attach_domain(NULL, &def_root_domain, i);
6845 /* handle null as "default" */
6846 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6847 struct sched_domain_attr *new, int idx_new)
6849 struct sched_domain_attr tmp;
6856 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6857 new ? (new + idx_new) : &tmp,
6858 sizeof(struct sched_domain_attr));
6862 * Partition sched domains as specified by the 'ndoms_new'
6863 * cpumasks in the array doms_new[] of cpumasks. This compares
6864 * doms_new[] to the current sched domain partitioning, doms_cur[].
6865 * It destroys each deleted domain and builds each new domain.
6867 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6868 * The masks don't intersect (don't overlap.) We should setup one
6869 * sched domain for each mask. CPUs not in any of the cpumasks will
6870 * not be load balanced. If the same cpumask appears both in the
6871 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6874 * The passed in 'doms_new' should be allocated using
6875 * alloc_sched_domains. This routine takes ownership of it and will
6876 * free_sched_domains it when done with it. If the caller failed the
6877 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6878 * and partition_sched_domains() will fallback to the single partition
6879 * 'fallback_doms', it also forces the domains to be rebuilt.
6881 * If doms_new == NULL it will be replaced with cpu_online_mask.
6882 * ndoms_new == 0 is a special case for destroying existing domains,
6883 * and it will not create the default domain.
6885 * Call with hotplug lock held
6887 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6888 struct sched_domain_attr *dattr_new)
6893 mutex_lock(&sched_domains_mutex);
6895 /* always unregister in case we don't destroy any domains */
6896 unregister_sched_domain_sysctl();
6898 /* Let architecture update cpu core mappings. */
6899 new_topology = arch_update_cpu_topology();
6901 n = doms_new ? ndoms_new : 0;
6903 /* Destroy deleted domains */
6904 for (i = 0; i < ndoms_cur; i++) {
6905 for (j = 0; j < n && !new_topology; j++) {
6906 if (cpumask_equal(doms_cur[i], doms_new[j])
6907 && dattrs_equal(dattr_cur, i, dattr_new, j))
6910 /* no match - a current sched domain not in new doms_new[] */
6911 detach_destroy_domains(doms_cur[i]);
6917 if (doms_new == NULL) {
6919 doms_new = &fallback_doms;
6920 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6921 WARN_ON_ONCE(dattr_new);
6924 /* Build new domains */
6925 for (i = 0; i < ndoms_new; i++) {
6926 for (j = 0; j < n && !new_topology; j++) {
6927 if (cpumask_equal(doms_new[i], doms_cur[j])
6928 && dattrs_equal(dattr_new, i, dattr_cur, j))
6931 /* no match - add a new doms_new */
6932 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6937 /* Remember the new sched domains */
6938 if (doms_cur != &fallback_doms)
6939 free_sched_domains(doms_cur, ndoms_cur);
6940 kfree(dattr_cur); /* kfree(NULL) is safe */
6941 doms_cur = doms_new;
6942 dattr_cur = dattr_new;
6943 ndoms_cur = ndoms_new;
6945 register_sched_domain_sysctl();
6947 mutex_unlock(&sched_domains_mutex);
6950 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6953 * Update cpusets according to cpu_active mask. If cpusets are
6954 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6955 * around partition_sched_domains().
6957 * If we come here as part of a suspend/resume, don't touch cpusets because we
6958 * want to restore it back to its original state upon resume anyway.
6960 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6964 case CPU_ONLINE_FROZEN:
6965 case CPU_DOWN_FAILED_FROZEN:
6968 * num_cpus_frozen tracks how many CPUs are involved in suspend
6969 * resume sequence. As long as this is not the last online
6970 * operation in the resume sequence, just build a single sched
6971 * domain, ignoring cpusets.
6974 if (likely(num_cpus_frozen)) {
6975 partition_sched_domains(1, NULL, NULL);
6980 * This is the last CPU online operation. So fall through and
6981 * restore the original sched domains by considering the
6982 * cpuset configurations.
6986 cpuset_update_active_cpus(true);
6994 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6997 unsigned long flags;
6998 long cpu = (long)hcpu;
7001 switch (action & ~CPU_TASKS_FROZEN) {
7002 case CPU_DOWN_PREPARE:
7003 /* explicitly allow suspend */
7004 if (!(action & CPU_TASKS_FROZEN)) {
7008 rcu_read_lock_sched();
7009 dl_b = dl_bw_of(cpu);
7011 raw_spin_lock_irqsave(&dl_b->lock, flags);
7012 cpus = dl_bw_cpus(cpu);
7013 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7014 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7016 rcu_read_unlock_sched();
7019 return notifier_from_errno(-EBUSY);
7021 cpuset_update_active_cpus(false);
7023 case CPU_DOWN_PREPARE_FROZEN:
7025 partition_sched_domains(1, NULL, NULL);
7033 void __init sched_init_smp(void)
7035 cpumask_var_t non_isolated_cpus;
7037 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7038 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7043 * There's no userspace yet to cause hotplug operations; hence all the
7044 * cpu masks are stable and all blatant races in the below code cannot
7047 mutex_lock(&sched_domains_mutex);
7048 init_sched_domains(cpu_active_mask);
7049 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7050 if (cpumask_empty(non_isolated_cpus))
7051 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7052 mutex_unlock(&sched_domains_mutex);
7054 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7055 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7056 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7060 /* Move init over to a non-isolated CPU */
7061 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7063 sched_init_granularity();
7064 free_cpumask_var(non_isolated_cpus);
7066 init_sched_rt_class();
7067 init_sched_dl_class();
7070 void __init sched_init_smp(void)
7072 sched_init_granularity();
7074 #endif /* CONFIG_SMP */
7076 const_debug unsigned int sysctl_timer_migration = 1;
7078 int in_sched_functions(unsigned long addr)
7080 return in_lock_functions(addr) ||
7081 (addr >= (unsigned long)__sched_text_start
7082 && addr < (unsigned long)__sched_text_end);
7085 #ifdef CONFIG_CGROUP_SCHED
7087 * Default task group.
7088 * Every task in system belongs to this group at bootup.
7090 struct task_group root_task_group;
7091 LIST_HEAD(task_groups);
7094 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7096 void __init sched_init(void)
7099 unsigned long alloc_size = 0, ptr;
7101 #ifdef CONFIG_FAIR_GROUP_SCHED
7102 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7104 #ifdef CONFIG_RT_GROUP_SCHED
7105 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7108 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7110 #ifdef CONFIG_FAIR_GROUP_SCHED
7111 root_task_group.se = (struct sched_entity **)ptr;
7112 ptr += nr_cpu_ids * sizeof(void **);
7114 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7115 ptr += nr_cpu_ids * sizeof(void **);
7117 #endif /* CONFIG_FAIR_GROUP_SCHED */
7118 #ifdef CONFIG_RT_GROUP_SCHED
7119 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7120 ptr += nr_cpu_ids * sizeof(void **);
7122 root_task_group.rt_rq = (struct rt_rq **)ptr;
7123 ptr += nr_cpu_ids * sizeof(void **);
7125 #endif /* CONFIG_RT_GROUP_SCHED */
7127 #ifdef CONFIG_CPUMASK_OFFSTACK
7128 for_each_possible_cpu(i) {
7129 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7130 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7132 #endif /* CONFIG_CPUMASK_OFFSTACK */
7134 init_rt_bandwidth(&def_rt_bandwidth,
7135 global_rt_period(), global_rt_runtime());
7136 init_dl_bandwidth(&def_dl_bandwidth,
7137 global_rt_period(), global_rt_runtime());
7140 init_defrootdomain();
7143 #ifdef CONFIG_RT_GROUP_SCHED
7144 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7145 global_rt_period(), global_rt_runtime());
7146 #endif /* CONFIG_RT_GROUP_SCHED */
7148 #ifdef CONFIG_CGROUP_SCHED
7149 list_add(&root_task_group.list, &task_groups);
7150 INIT_LIST_HEAD(&root_task_group.children);
7151 INIT_LIST_HEAD(&root_task_group.siblings);
7152 autogroup_init(&init_task);
7154 #endif /* CONFIG_CGROUP_SCHED */
7156 for_each_possible_cpu(i) {
7160 raw_spin_lock_init(&rq->lock);
7162 rq->calc_load_active = 0;
7163 rq->calc_load_update = jiffies + LOAD_FREQ;
7164 init_cfs_rq(&rq->cfs);
7165 init_rt_rq(&rq->rt);
7166 init_dl_rq(&rq->dl);
7167 #ifdef CONFIG_FAIR_GROUP_SCHED
7168 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7169 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7171 * How much cpu bandwidth does root_task_group get?
7173 * In case of task-groups formed thr' the cgroup filesystem, it
7174 * gets 100% of the cpu resources in the system. This overall
7175 * system cpu resource is divided among the tasks of
7176 * root_task_group and its child task-groups in a fair manner,
7177 * based on each entity's (task or task-group's) weight
7178 * (se->load.weight).
7180 * In other words, if root_task_group has 10 tasks of weight
7181 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7182 * then A0's share of the cpu resource is:
7184 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7186 * We achieve this by letting root_task_group's tasks sit
7187 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7189 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7190 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7191 #endif /* CONFIG_FAIR_GROUP_SCHED */
7193 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7194 #ifdef CONFIG_RT_GROUP_SCHED
7195 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7198 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7199 rq->cpu_load[j] = 0;
7201 rq->last_load_update_tick = jiffies;
7206 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7207 rq->post_schedule = 0;
7208 rq->active_balance = 0;
7209 rq->next_balance = jiffies;
7214 rq->avg_idle = 2*sysctl_sched_migration_cost;
7215 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7217 INIT_LIST_HEAD(&rq->cfs_tasks);
7219 rq_attach_root(rq, &def_root_domain);
7220 #ifdef CONFIG_NO_HZ_COMMON
7223 #ifdef CONFIG_NO_HZ_FULL
7224 rq->last_sched_tick = 0;
7228 atomic_set(&rq->nr_iowait, 0);
7231 set_load_weight(&init_task);
7233 #ifdef CONFIG_PREEMPT_NOTIFIERS
7234 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7238 * The boot idle thread does lazy MMU switching as well:
7240 atomic_inc(&init_mm.mm_count);
7241 enter_lazy_tlb(&init_mm, current);
7244 * During early bootup we pretend to be a normal task:
7246 current->sched_class = &fair_sched_class;
7249 * Make us the idle thread. Technically, schedule() should not be
7250 * called from this thread, however somewhere below it might be,
7251 * but because we are the idle thread, we just pick up running again
7252 * when this runqueue becomes "idle".
7254 init_idle(current, smp_processor_id());
7256 calc_load_update = jiffies + LOAD_FREQ;
7259 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7260 /* May be allocated at isolcpus cmdline parse time */
7261 if (cpu_isolated_map == NULL)
7262 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7263 idle_thread_set_boot_cpu();
7264 set_cpu_rq_start_time();
7266 init_sched_fair_class();
7268 scheduler_running = 1;
7271 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7272 static inline int preempt_count_equals(int preempt_offset)
7274 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7276 return (nested == preempt_offset);
7279 void __might_sleep(const char *file, int line, int preempt_offset)
7282 * Blocking primitives will set (and therefore destroy) current->state,
7283 * since we will exit with TASK_RUNNING make sure we enter with it,
7284 * otherwise we will destroy state.
7286 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7287 "do not call blocking ops when !TASK_RUNNING; "
7288 "state=%lx set at [<%p>] %pS\n",
7290 (void *)current->task_state_change,
7291 (void *)current->task_state_change);
7293 ___might_sleep(file, line, preempt_offset);
7295 EXPORT_SYMBOL(__might_sleep);
7297 void ___might_sleep(const char *file, int line, int preempt_offset)
7299 static unsigned long prev_jiffy; /* ratelimiting */
7301 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7302 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7303 !is_idle_task(current)) ||
7304 system_state != SYSTEM_RUNNING || oops_in_progress)
7306 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7308 prev_jiffy = jiffies;
7311 "BUG: sleeping function called from invalid context at %s:%d\n",
7314 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7315 in_atomic(), irqs_disabled(),
7316 current->pid, current->comm);
7318 if (task_stack_end_corrupted(current))
7319 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7321 debug_show_held_locks(current);
7322 if (irqs_disabled())
7323 print_irqtrace_events(current);
7324 #ifdef CONFIG_DEBUG_PREEMPT
7325 if (!preempt_count_equals(preempt_offset)) {
7326 pr_err("Preemption disabled at:");
7327 print_ip_sym(current->preempt_disable_ip);
7333 EXPORT_SYMBOL(___might_sleep);
7336 #ifdef CONFIG_MAGIC_SYSRQ
7337 static void normalize_task(struct rq *rq, struct task_struct *p)
7339 const struct sched_class *prev_class = p->sched_class;
7340 struct sched_attr attr = {
7341 .sched_policy = SCHED_NORMAL,
7343 int old_prio = p->prio;
7346 queued = task_on_rq_queued(p);
7348 dequeue_task(rq, p, 0);
7349 __setscheduler(rq, p, &attr);
7351 enqueue_task(rq, p, 0);
7355 check_class_changed(rq, p, prev_class, old_prio);
7358 void normalize_rt_tasks(void)
7360 struct task_struct *g, *p;
7361 unsigned long flags;
7364 read_lock(&tasklist_lock);
7365 for_each_process_thread(g, p) {
7367 * Only normalize user tasks:
7369 if (p->flags & PF_KTHREAD)
7372 p->se.exec_start = 0;
7373 #ifdef CONFIG_SCHEDSTATS
7374 p->se.statistics.wait_start = 0;
7375 p->se.statistics.sleep_start = 0;
7376 p->se.statistics.block_start = 0;
7379 if (!dl_task(p) && !rt_task(p)) {
7381 * Renice negative nice level userspace
7384 if (task_nice(p) < 0)
7385 set_user_nice(p, 0);
7389 rq = task_rq_lock(p, &flags);
7390 normalize_task(rq, p);
7391 task_rq_unlock(rq, p, &flags);
7393 read_unlock(&tasklist_lock);
7396 #endif /* CONFIG_MAGIC_SYSRQ */
7398 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7400 * These functions are only useful for the IA64 MCA handling, or kdb.
7402 * They can only be called when the whole system has been
7403 * stopped - every CPU needs to be quiescent, and no scheduling
7404 * activity can take place. Using them for anything else would
7405 * be a serious bug, and as a result, they aren't even visible
7406 * under any other configuration.
7410 * curr_task - return the current task for a given cpu.
7411 * @cpu: the processor in question.
7413 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7415 * Return: The current task for @cpu.
7417 struct task_struct *curr_task(int cpu)
7419 return cpu_curr(cpu);
7422 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7426 * set_curr_task - set the current task for a given cpu.
7427 * @cpu: the processor in question.
7428 * @p: the task pointer to set.
7430 * Description: This function must only be used when non-maskable interrupts
7431 * are serviced on a separate stack. It allows the architecture to switch the
7432 * notion of the current task on a cpu in a non-blocking manner. This function
7433 * must be called with all CPU's synchronized, and interrupts disabled, the
7434 * and caller must save the original value of the current task (see
7435 * curr_task() above) and restore that value before reenabling interrupts and
7436 * re-starting the system.
7438 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7440 void set_curr_task(int cpu, struct task_struct *p)
7447 #ifdef CONFIG_CGROUP_SCHED
7448 /* task_group_lock serializes the addition/removal of task groups */
7449 static DEFINE_SPINLOCK(task_group_lock);
7451 static void free_sched_group(struct task_group *tg)
7453 free_fair_sched_group(tg);
7454 free_rt_sched_group(tg);
7459 /* allocate runqueue etc for a new task group */
7460 struct task_group *sched_create_group(struct task_group *parent)
7462 struct task_group *tg;
7464 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7466 return ERR_PTR(-ENOMEM);
7468 if (!alloc_fair_sched_group(tg, parent))
7471 if (!alloc_rt_sched_group(tg, parent))
7477 free_sched_group(tg);
7478 return ERR_PTR(-ENOMEM);
7481 void sched_online_group(struct task_group *tg, struct task_group *parent)
7483 unsigned long flags;
7485 spin_lock_irqsave(&task_group_lock, flags);
7486 list_add_rcu(&tg->list, &task_groups);
7488 WARN_ON(!parent); /* root should already exist */
7490 tg->parent = parent;
7491 INIT_LIST_HEAD(&tg->children);
7492 list_add_rcu(&tg->siblings, &parent->children);
7493 spin_unlock_irqrestore(&task_group_lock, flags);
7496 /* rcu callback to free various structures associated with a task group */
7497 static void free_sched_group_rcu(struct rcu_head *rhp)
7499 /* now it should be safe to free those cfs_rqs */
7500 free_sched_group(container_of(rhp, struct task_group, rcu));
7503 /* Destroy runqueue etc associated with a task group */
7504 void sched_destroy_group(struct task_group *tg)
7506 /* wait for possible concurrent references to cfs_rqs complete */
7507 call_rcu(&tg->rcu, free_sched_group_rcu);
7510 void sched_offline_group(struct task_group *tg)
7512 unsigned long flags;
7515 /* end participation in shares distribution */
7516 for_each_possible_cpu(i)
7517 unregister_fair_sched_group(tg, i);
7519 spin_lock_irqsave(&task_group_lock, flags);
7520 list_del_rcu(&tg->list);
7521 list_del_rcu(&tg->siblings);
7522 spin_unlock_irqrestore(&task_group_lock, flags);
7525 /* change task's runqueue when it moves between groups.
7526 * The caller of this function should have put the task in its new group
7527 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7528 * reflect its new group.
7530 void sched_move_task(struct task_struct *tsk)
7532 struct task_group *tg;
7533 int queued, running;
7534 unsigned long flags;
7537 rq = task_rq_lock(tsk, &flags);
7539 running = task_current(rq, tsk);
7540 queued = task_on_rq_queued(tsk);
7543 dequeue_task(rq, tsk, 0);
7544 if (unlikely(running))
7545 put_prev_task(rq, tsk);
7548 * All callers are synchronized by task_rq_lock(); we do not use RCU
7549 * which is pointless here. Thus, we pass "true" to task_css_check()
7550 * to prevent lockdep warnings.
7552 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7553 struct task_group, css);
7554 tg = autogroup_task_group(tsk, tg);
7555 tsk->sched_task_group = tg;
7557 #ifdef CONFIG_FAIR_GROUP_SCHED
7558 if (tsk->sched_class->task_move_group)
7559 tsk->sched_class->task_move_group(tsk, queued);
7562 set_task_rq(tsk, task_cpu(tsk));
7564 if (unlikely(running))
7565 tsk->sched_class->set_curr_task(rq);
7567 enqueue_task(rq, tsk, 0);
7569 task_rq_unlock(rq, tsk, &flags);
7571 #endif /* CONFIG_CGROUP_SCHED */
7573 #ifdef CONFIG_RT_GROUP_SCHED
7575 * Ensure that the real time constraints are schedulable.
7577 static DEFINE_MUTEX(rt_constraints_mutex);
7579 /* Must be called with tasklist_lock held */
7580 static inline int tg_has_rt_tasks(struct task_group *tg)
7582 struct task_struct *g, *p;
7585 * Autogroups do not have RT tasks; see autogroup_create().
7587 if (task_group_is_autogroup(tg))
7590 for_each_process_thread(g, p) {
7591 if (rt_task(p) && task_group(p) == tg)
7598 struct rt_schedulable_data {
7599 struct task_group *tg;
7604 static int tg_rt_schedulable(struct task_group *tg, void *data)
7606 struct rt_schedulable_data *d = data;
7607 struct task_group *child;
7608 unsigned long total, sum = 0;
7609 u64 period, runtime;
7611 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7612 runtime = tg->rt_bandwidth.rt_runtime;
7615 period = d->rt_period;
7616 runtime = d->rt_runtime;
7620 * Cannot have more runtime than the period.
7622 if (runtime > period && runtime != RUNTIME_INF)
7626 * Ensure we don't starve existing RT tasks.
7628 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7631 total = to_ratio(period, runtime);
7634 * Nobody can have more than the global setting allows.
7636 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7640 * The sum of our children's runtime should not exceed our own.
7642 list_for_each_entry_rcu(child, &tg->children, siblings) {
7643 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7644 runtime = child->rt_bandwidth.rt_runtime;
7646 if (child == d->tg) {
7647 period = d->rt_period;
7648 runtime = d->rt_runtime;
7651 sum += to_ratio(period, runtime);
7660 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7664 struct rt_schedulable_data data = {
7666 .rt_period = period,
7667 .rt_runtime = runtime,
7671 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7677 static int tg_set_rt_bandwidth(struct task_group *tg,
7678 u64 rt_period, u64 rt_runtime)
7683 * Disallowing the root group RT runtime is BAD, it would disallow the
7684 * kernel creating (and or operating) RT threads.
7686 if (tg == &root_task_group && rt_runtime == 0)
7689 /* No period doesn't make any sense. */
7693 mutex_lock(&rt_constraints_mutex);
7694 read_lock(&tasklist_lock);
7695 err = __rt_schedulable(tg, rt_period, rt_runtime);
7699 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7700 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7701 tg->rt_bandwidth.rt_runtime = rt_runtime;
7703 for_each_possible_cpu(i) {
7704 struct rt_rq *rt_rq = tg->rt_rq[i];
7706 raw_spin_lock(&rt_rq->rt_runtime_lock);
7707 rt_rq->rt_runtime = rt_runtime;
7708 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7710 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7712 read_unlock(&tasklist_lock);
7713 mutex_unlock(&rt_constraints_mutex);
7718 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7720 u64 rt_runtime, rt_period;
7722 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7723 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7724 if (rt_runtime_us < 0)
7725 rt_runtime = RUNTIME_INF;
7727 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7730 static long sched_group_rt_runtime(struct task_group *tg)
7734 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7737 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7738 do_div(rt_runtime_us, NSEC_PER_USEC);
7739 return rt_runtime_us;
7742 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7744 u64 rt_runtime, rt_period;
7746 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7747 rt_runtime = tg->rt_bandwidth.rt_runtime;
7749 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7752 static long sched_group_rt_period(struct task_group *tg)
7756 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7757 do_div(rt_period_us, NSEC_PER_USEC);
7758 return rt_period_us;
7760 #endif /* CONFIG_RT_GROUP_SCHED */
7762 #ifdef CONFIG_RT_GROUP_SCHED
7763 static int sched_rt_global_constraints(void)
7767 mutex_lock(&rt_constraints_mutex);
7768 read_lock(&tasklist_lock);
7769 ret = __rt_schedulable(NULL, 0, 0);
7770 read_unlock(&tasklist_lock);
7771 mutex_unlock(&rt_constraints_mutex);
7776 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7778 /* Don't accept realtime tasks when there is no way for them to run */
7779 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7785 #else /* !CONFIG_RT_GROUP_SCHED */
7786 static int sched_rt_global_constraints(void)
7788 unsigned long flags;
7791 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7792 for_each_possible_cpu(i) {
7793 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7795 raw_spin_lock(&rt_rq->rt_runtime_lock);
7796 rt_rq->rt_runtime = global_rt_runtime();
7797 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7799 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7803 #endif /* CONFIG_RT_GROUP_SCHED */
7805 static int sched_dl_global_validate(void)
7807 u64 runtime = global_rt_runtime();
7808 u64 period = global_rt_period();
7809 u64 new_bw = to_ratio(period, runtime);
7812 unsigned long flags;
7815 * Here we want to check the bandwidth not being set to some
7816 * value smaller than the currently allocated bandwidth in
7817 * any of the root_domains.
7819 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7820 * cycling on root_domains... Discussion on different/better
7821 * solutions is welcome!
7823 for_each_possible_cpu(cpu) {
7824 rcu_read_lock_sched();
7825 dl_b = dl_bw_of(cpu);
7827 raw_spin_lock_irqsave(&dl_b->lock, flags);
7828 if (new_bw < dl_b->total_bw)
7830 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7832 rcu_read_unlock_sched();
7841 static void sched_dl_do_global(void)
7846 unsigned long flags;
7848 def_dl_bandwidth.dl_period = global_rt_period();
7849 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7851 if (global_rt_runtime() != RUNTIME_INF)
7852 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7855 * FIXME: As above...
7857 for_each_possible_cpu(cpu) {
7858 rcu_read_lock_sched();
7859 dl_b = dl_bw_of(cpu);
7861 raw_spin_lock_irqsave(&dl_b->lock, flags);
7863 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7865 rcu_read_unlock_sched();
7869 static int sched_rt_global_validate(void)
7871 if (sysctl_sched_rt_period <= 0)
7874 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7875 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7881 static void sched_rt_do_global(void)
7883 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7884 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7887 int sched_rt_handler(struct ctl_table *table, int write,
7888 void __user *buffer, size_t *lenp,
7891 int old_period, old_runtime;
7892 static DEFINE_MUTEX(mutex);
7896 old_period = sysctl_sched_rt_period;
7897 old_runtime = sysctl_sched_rt_runtime;
7899 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7901 if (!ret && write) {
7902 ret = sched_rt_global_validate();
7906 ret = sched_dl_global_validate();
7910 ret = sched_rt_global_constraints();
7914 sched_rt_do_global();
7915 sched_dl_do_global();
7919 sysctl_sched_rt_period = old_period;
7920 sysctl_sched_rt_runtime = old_runtime;
7922 mutex_unlock(&mutex);
7927 int sched_rr_handler(struct ctl_table *table, int write,
7928 void __user *buffer, size_t *lenp,
7932 static DEFINE_MUTEX(mutex);
7935 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7936 /* make sure that internally we keep jiffies */
7937 /* also, writing zero resets timeslice to default */
7938 if (!ret && write) {
7939 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7940 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7942 mutex_unlock(&mutex);
7946 #ifdef CONFIG_CGROUP_SCHED
7948 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7950 return css ? container_of(css, struct task_group, css) : NULL;
7953 static struct cgroup_subsys_state *
7954 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7956 struct task_group *parent = css_tg(parent_css);
7957 struct task_group *tg;
7960 /* This is early initialization for the top cgroup */
7961 return &root_task_group.css;
7964 tg = sched_create_group(parent);
7966 return ERR_PTR(-ENOMEM);
7971 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7973 struct task_group *tg = css_tg(css);
7974 struct task_group *parent = css_tg(css->parent);
7977 sched_online_group(tg, parent);
7981 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7983 struct task_group *tg = css_tg(css);
7985 sched_destroy_group(tg);
7988 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7990 struct task_group *tg = css_tg(css);
7992 sched_offline_group(tg);
7995 static void cpu_cgroup_fork(struct task_struct *task)
7997 sched_move_task(task);
8000 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8001 struct cgroup_taskset *tset)
8003 struct task_struct *task;
8005 cgroup_taskset_for_each(task, tset) {
8006 #ifdef CONFIG_RT_GROUP_SCHED
8007 if (!sched_rt_can_attach(css_tg(css), task))
8010 /* We don't support RT-tasks being in separate groups */
8011 if (task->sched_class != &fair_sched_class)
8018 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8019 struct cgroup_taskset *tset)
8021 struct task_struct *task;
8023 cgroup_taskset_for_each(task, tset)
8024 sched_move_task(task);
8027 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8028 struct cgroup_subsys_state *old_css,
8029 struct task_struct *task)
8032 * cgroup_exit() is called in the copy_process() failure path.
8033 * Ignore this case since the task hasn't ran yet, this avoids
8034 * trying to poke a half freed task state from generic code.
8036 if (!(task->flags & PF_EXITING))
8039 sched_move_task(task);
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8044 struct cftype *cftype, u64 shareval)
8046 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8049 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8052 struct task_group *tg = css_tg(css);
8054 return (u64) scale_load_down(tg->shares);
8057 #ifdef CONFIG_CFS_BANDWIDTH
8058 static DEFINE_MUTEX(cfs_constraints_mutex);
8060 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8061 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8063 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8065 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8067 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8068 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8070 if (tg == &root_task_group)
8074 * Ensure we have at some amount of bandwidth every period. This is
8075 * to prevent reaching a state of large arrears when throttled via
8076 * entity_tick() resulting in prolonged exit starvation.
8078 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8082 * Likewise, bound things on the otherside by preventing insane quota
8083 * periods. This also allows us to normalize in computing quota
8086 if (period > max_cfs_quota_period)
8090 * Prevent race between setting of cfs_rq->runtime_enabled and
8091 * unthrottle_offline_cfs_rqs().
8094 mutex_lock(&cfs_constraints_mutex);
8095 ret = __cfs_schedulable(tg, period, quota);
8099 runtime_enabled = quota != RUNTIME_INF;
8100 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8102 * If we need to toggle cfs_bandwidth_used, off->on must occur
8103 * before making related changes, and on->off must occur afterwards
8105 if (runtime_enabled && !runtime_was_enabled)
8106 cfs_bandwidth_usage_inc();
8107 raw_spin_lock_irq(&cfs_b->lock);
8108 cfs_b->period = ns_to_ktime(period);
8109 cfs_b->quota = quota;
8111 __refill_cfs_bandwidth_runtime(cfs_b);
8112 /* restart the period timer (if active) to handle new period expiry */
8113 if (runtime_enabled && cfs_b->timer_active) {
8114 /* force a reprogram */
8115 __start_cfs_bandwidth(cfs_b, true);
8117 raw_spin_unlock_irq(&cfs_b->lock);
8119 for_each_online_cpu(i) {
8120 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8121 struct rq *rq = cfs_rq->rq;
8123 raw_spin_lock_irq(&rq->lock);
8124 cfs_rq->runtime_enabled = runtime_enabled;
8125 cfs_rq->runtime_remaining = 0;
8127 if (cfs_rq->throttled)
8128 unthrottle_cfs_rq(cfs_rq);
8129 raw_spin_unlock_irq(&rq->lock);
8131 if (runtime_was_enabled && !runtime_enabled)
8132 cfs_bandwidth_usage_dec();
8134 mutex_unlock(&cfs_constraints_mutex);
8140 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8144 period = ktime_to_ns(tg->cfs_bandwidth.period);
8145 if (cfs_quota_us < 0)
8146 quota = RUNTIME_INF;
8148 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8150 return tg_set_cfs_bandwidth(tg, period, quota);
8153 long tg_get_cfs_quota(struct task_group *tg)
8157 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8160 quota_us = tg->cfs_bandwidth.quota;
8161 do_div(quota_us, NSEC_PER_USEC);
8166 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8170 period = (u64)cfs_period_us * NSEC_PER_USEC;
8171 quota = tg->cfs_bandwidth.quota;
8173 return tg_set_cfs_bandwidth(tg, period, quota);
8176 long tg_get_cfs_period(struct task_group *tg)
8180 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8181 do_div(cfs_period_us, NSEC_PER_USEC);
8183 return cfs_period_us;
8186 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8189 return tg_get_cfs_quota(css_tg(css));
8192 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8193 struct cftype *cftype, s64 cfs_quota_us)
8195 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8198 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8201 return tg_get_cfs_period(css_tg(css));
8204 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8205 struct cftype *cftype, u64 cfs_period_us)
8207 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8210 struct cfs_schedulable_data {
8211 struct task_group *tg;
8216 * normalize group quota/period to be quota/max_period
8217 * note: units are usecs
8219 static u64 normalize_cfs_quota(struct task_group *tg,
8220 struct cfs_schedulable_data *d)
8228 period = tg_get_cfs_period(tg);
8229 quota = tg_get_cfs_quota(tg);
8232 /* note: these should typically be equivalent */
8233 if (quota == RUNTIME_INF || quota == -1)
8236 return to_ratio(period, quota);
8239 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8241 struct cfs_schedulable_data *d = data;
8242 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8243 s64 quota = 0, parent_quota = -1;
8246 quota = RUNTIME_INF;
8248 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8250 quota = normalize_cfs_quota(tg, d);
8251 parent_quota = parent_b->hierarchical_quota;
8254 * ensure max(child_quota) <= parent_quota, inherit when no
8257 if (quota == RUNTIME_INF)
8258 quota = parent_quota;
8259 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8262 cfs_b->hierarchical_quota = quota;
8267 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8270 struct cfs_schedulable_data data = {
8276 if (quota != RUNTIME_INF) {
8277 do_div(data.period, NSEC_PER_USEC);
8278 do_div(data.quota, NSEC_PER_USEC);
8282 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8288 static int cpu_stats_show(struct seq_file *sf, void *v)
8290 struct task_group *tg = css_tg(seq_css(sf));
8291 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8293 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8294 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8295 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8299 #endif /* CONFIG_CFS_BANDWIDTH */
8300 #endif /* CONFIG_FAIR_GROUP_SCHED */
8302 #ifdef CONFIG_RT_GROUP_SCHED
8303 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8304 struct cftype *cft, s64 val)
8306 return sched_group_set_rt_runtime(css_tg(css), val);
8309 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8312 return sched_group_rt_runtime(css_tg(css));
8315 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8316 struct cftype *cftype, u64 rt_period_us)
8318 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8321 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8324 return sched_group_rt_period(css_tg(css));
8326 #endif /* CONFIG_RT_GROUP_SCHED */
8328 static struct cftype cpu_files[] = {
8329 #ifdef CONFIG_FAIR_GROUP_SCHED
8332 .read_u64 = cpu_shares_read_u64,
8333 .write_u64 = cpu_shares_write_u64,
8336 #ifdef CONFIG_CFS_BANDWIDTH
8338 .name = "cfs_quota_us",
8339 .read_s64 = cpu_cfs_quota_read_s64,
8340 .write_s64 = cpu_cfs_quota_write_s64,
8343 .name = "cfs_period_us",
8344 .read_u64 = cpu_cfs_period_read_u64,
8345 .write_u64 = cpu_cfs_period_write_u64,
8349 .seq_show = cpu_stats_show,
8352 #ifdef CONFIG_RT_GROUP_SCHED
8354 .name = "rt_runtime_us",
8355 .read_s64 = cpu_rt_runtime_read,
8356 .write_s64 = cpu_rt_runtime_write,
8359 .name = "rt_period_us",
8360 .read_u64 = cpu_rt_period_read_uint,
8361 .write_u64 = cpu_rt_period_write_uint,
8367 struct cgroup_subsys cpu_cgrp_subsys = {
8368 .css_alloc = cpu_cgroup_css_alloc,
8369 .css_free = cpu_cgroup_css_free,
8370 .css_online = cpu_cgroup_css_online,
8371 .css_offline = cpu_cgroup_css_offline,
8372 .fork = cpu_cgroup_fork,
8373 .can_attach = cpu_cgroup_can_attach,
8374 .attach = cpu_cgroup_attach,
8375 .exit = cpu_cgroup_exit,
8376 .legacy_cftypes = cpu_files,
8380 #endif /* CONFIG_CGROUP_SCHED */
8382 void dump_cpu_task(int cpu)
8384 pr_info("Task dump for CPU %d:\n", cpu);
8385 sched_show_task(cpu_curr(cpu));