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 if (rq->skip_clock_update > 0)
125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
129 update_rq_clock_task(rq, delta);
133 * Debugging: various feature bits
136 #define SCHED_FEAT(name, enabled) \
137 (1UL << __SCHED_FEAT_##name) * enabled |
139 const_debug unsigned int sysctl_sched_features =
140 #include "features.h"
145 #ifdef CONFIG_SCHED_DEBUG
146 #define SCHED_FEAT(name, enabled) \
149 static const char * const sched_feat_names[] = {
150 #include "features.h"
155 static int sched_feat_show(struct seq_file *m, void *v)
159 for (i = 0; i < __SCHED_FEAT_NR; i++) {
160 if (!(sysctl_sched_features & (1UL << i)))
162 seq_printf(m, "%s ", sched_feat_names[i]);
169 #ifdef HAVE_JUMP_LABEL
171 #define jump_label_key__true STATIC_KEY_INIT_TRUE
172 #define jump_label_key__false STATIC_KEY_INIT_FALSE
174 #define SCHED_FEAT(name, enabled) \
175 jump_label_key__##enabled ,
177 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
178 #include "features.h"
183 static void sched_feat_disable(int i)
185 if (static_key_enabled(&sched_feat_keys[i]))
186 static_key_slow_dec(&sched_feat_keys[i]);
189 static void sched_feat_enable(int i)
191 if (!static_key_enabled(&sched_feat_keys[i]))
192 static_key_slow_inc(&sched_feat_keys[i]);
195 static void sched_feat_disable(int i) { };
196 static void sched_feat_enable(int i) { };
197 #endif /* HAVE_JUMP_LABEL */
199 static int sched_feat_set(char *cmp)
204 if (strncmp(cmp, "NO_", 3) == 0) {
209 for (i = 0; i < __SCHED_FEAT_NR; i++) {
210 if (strcmp(cmp, sched_feat_names[i]) == 0) {
212 sysctl_sched_features &= ~(1UL << i);
213 sched_feat_disable(i);
215 sysctl_sched_features |= (1UL << i);
216 sched_feat_enable(i);
226 sched_feat_write(struct file *filp, const char __user *ubuf,
227 size_t cnt, loff_t *ppos)
237 if (copy_from_user(&buf, ubuf, cnt))
243 /* Ensure the static_key remains in a consistent state */
244 inode = file_inode(filp);
245 mutex_lock(&inode->i_mutex);
246 i = sched_feat_set(cmp);
247 mutex_unlock(&inode->i_mutex);
248 if (i == __SCHED_FEAT_NR)
256 static int sched_feat_open(struct inode *inode, struct file *filp)
258 return single_open(filp, sched_feat_show, NULL);
261 static const struct file_operations sched_feat_fops = {
262 .open = sched_feat_open,
263 .write = sched_feat_write,
266 .release = single_release,
269 static __init int sched_init_debug(void)
271 debugfs_create_file("sched_features", 0644, NULL, NULL,
276 late_initcall(sched_init_debug);
277 #endif /* CONFIG_SCHED_DEBUG */
280 * Number of tasks to iterate in a single balance run.
281 * Limited because this is done with IRQs disabled.
283 const_debug unsigned int sysctl_sched_nr_migrate = 32;
286 * period over which we average the RT time consumption, measured
291 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
294 * period over which we measure -rt task cpu usage in us.
297 unsigned int sysctl_sched_rt_period = 1000000;
299 __read_mostly int scheduler_running;
302 * part of the period that we allow rt tasks to run in us.
305 int sysctl_sched_rt_runtime = 950000;
308 * __task_rq_lock - lock the rq @p resides on.
310 static inline struct rq *__task_rq_lock(struct task_struct *p)
315 lockdep_assert_held(&p->pi_lock);
319 raw_spin_lock(&rq->lock);
320 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
322 raw_spin_unlock(&rq->lock);
324 while (unlikely(task_on_rq_migrating(p)))
330 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
332 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
333 __acquires(p->pi_lock)
339 raw_spin_lock_irqsave(&p->pi_lock, *flags);
341 raw_spin_lock(&rq->lock);
342 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p)))
344 raw_spin_unlock(&rq->lock);
345 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
347 while (unlikely(task_on_rq_migrating(p)))
352 static void __task_rq_unlock(struct rq *rq)
355 raw_spin_unlock(&rq->lock);
359 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
361 __releases(p->pi_lock)
363 raw_spin_unlock(&rq->lock);
364 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
368 * this_rq_lock - lock this runqueue and disable interrupts.
370 static struct rq *this_rq_lock(void)
377 raw_spin_lock(&rq->lock);
382 #ifdef CONFIG_SCHED_HRTICK
384 * Use HR-timers to deliver accurate preemption points.
387 static void hrtick_clear(struct rq *rq)
389 if (hrtimer_active(&rq->hrtick_timer))
390 hrtimer_cancel(&rq->hrtick_timer);
394 * High-resolution timer tick.
395 * Runs from hardirq context with interrupts disabled.
397 static enum hrtimer_restart hrtick(struct hrtimer *timer)
399 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
401 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
403 raw_spin_lock(&rq->lock);
405 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
406 raw_spin_unlock(&rq->lock);
408 return HRTIMER_NORESTART;
413 static int __hrtick_restart(struct rq *rq)
415 struct hrtimer *timer = &rq->hrtick_timer;
416 ktime_t time = hrtimer_get_softexpires(timer);
418 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
422 * called from hardirq (IPI) context
424 static void __hrtick_start(void *arg)
428 raw_spin_lock(&rq->lock);
429 __hrtick_restart(rq);
430 rq->hrtick_csd_pending = 0;
431 raw_spin_unlock(&rq->lock);
435 * Called to set the hrtick timer state.
437 * called with rq->lock held and irqs disabled
439 void hrtick_start(struct rq *rq, u64 delay)
441 struct hrtimer *timer = &rq->hrtick_timer;
446 * Don't schedule slices shorter than 10000ns, that just
447 * doesn't make sense and can cause timer DoS.
449 delta = max_t(s64, delay, 10000LL);
450 time = ktime_add_ns(timer->base->get_time(), delta);
452 hrtimer_set_expires(timer, time);
454 if (rq == this_rq()) {
455 __hrtick_restart(rq);
456 } else if (!rq->hrtick_csd_pending) {
457 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
458 rq->hrtick_csd_pending = 1;
463 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
465 int cpu = (int)(long)hcpu;
468 case CPU_UP_CANCELED:
469 case CPU_UP_CANCELED_FROZEN:
470 case CPU_DOWN_PREPARE:
471 case CPU_DOWN_PREPARE_FROZEN:
473 case CPU_DEAD_FROZEN:
474 hrtick_clear(cpu_rq(cpu));
481 static __init void init_hrtick(void)
483 hotcpu_notifier(hotplug_hrtick, 0);
487 * Called to set the hrtick timer state.
489 * called with rq->lock held and irqs disabled
491 void hrtick_start(struct rq *rq, u64 delay)
493 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
494 HRTIMER_MODE_REL_PINNED, 0);
497 static inline void init_hrtick(void)
500 #endif /* CONFIG_SMP */
502 static void init_rq_hrtick(struct rq *rq)
505 rq->hrtick_csd_pending = 0;
507 rq->hrtick_csd.flags = 0;
508 rq->hrtick_csd.func = __hrtick_start;
509 rq->hrtick_csd.info = rq;
512 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
513 rq->hrtick_timer.function = hrtick;
515 #else /* CONFIG_SCHED_HRTICK */
516 static inline void hrtick_clear(struct rq *rq)
520 static inline void init_rq_hrtick(struct rq *rq)
524 static inline void init_hrtick(void)
527 #endif /* CONFIG_SCHED_HRTICK */
530 * cmpxchg based fetch_or, macro so it works for different integer types
532 #define fetch_or(ptr, val) \
533 ({ typeof(*(ptr)) __old, __val = *(ptr); \
535 __old = cmpxchg((ptr), __val, __val | (val)); \
536 if (__old == __val) \
543 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
545 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
546 * this avoids any races wrt polling state changes and thereby avoids
549 static bool set_nr_and_not_polling(struct task_struct *p)
551 struct thread_info *ti = task_thread_info(p);
552 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
556 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
558 * If this returns true, then the idle task promises to call
559 * sched_ttwu_pending() and reschedule soon.
561 static bool set_nr_if_polling(struct task_struct *p)
563 struct thread_info *ti = task_thread_info(p);
564 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
567 if (!(val & _TIF_POLLING_NRFLAG))
569 if (val & _TIF_NEED_RESCHED)
571 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
580 static bool set_nr_and_not_polling(struct task_struct *p)
582 set_tsk_need_resched(p);
587 static bool set_nr_if_polling(struct task_struct *p)
595 * resched_curr - mark rq's current task 'to be rescheduled now'.
597 * On UP this means the setting of the need_resched flag, on SMP it
598 * might also involve a cross-CPU call to trigger the scheduler on
601 void resched_curr(struct rq *rq)
603 struct task_struct *curr = rq->curr;
606 lockdep_assert_held(&rq->lock);
608 if (test_tsk_need_resched(curr))
613 if (cpu == smp_processor_id()) {
614 set_tsk_need_resched(curr);
615 set_preempt_need_resched();
619 if (set_nr_and_not_polling(curr))
620 smp_send_reschedule(cpu);
622 trace_sched_wake_idle_without_ipi(cpu);
625 void resched_cpu(int cpu)
627 struct rq *rq = cpu_rq(cpu);
630 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
633 raw_spin_unlock_irqrestore(&rq->lock, flags);
637 #ifdef CONFIG_NO_HZ_COMMON
639 * In the semi idle case, use the nearest busy cpu for migrating timers
640 * from an idle cpu. This is good for power-savings.
642 * We don't do similar optimization for completely idle system, as
643 * selecting an idle cpu will add more delays to the timers than intended
644 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
646 int get_nohz_timer_target(int pinned)
648 int cpu = smp_processor_id();
650 struct sched_domain *sd;
652 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
656 for_each_domain(cpu, sd) {
657 for_each_cpu(i, sched_domain_span(sd)) {
669 * When add_timer_on() enqueues a timer into the timer wheel of an
670 * idle CPU then this timer might expire before the next timer event
671 * which is scheduled to wake up that CPU. In case of a completely
672 * idle system the next event might even be infinite time into the
673 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
674 * leaves the inner idle loop so the newly added timer is taken into
675 * account when the CPU goes back to idle and evaluates the timer
676 * wheel for the next timer event.
678 static void wake_up_idle_cpu(int cpu)
680 struct rq *rq = cpu_rq(cpu);
682 if (cpu == smp_processor_id())
685 if (set_nr_and_not_polling(rq->idle))
686 smp_send_reschedule(cpu);
688 trace_sched_wake_idle_without_ipi(cpu);
691 static bool wake_up_full_nohz_cpu(int cpu)
694 * We just need the target to call irq_exit() and re-evaluate
695 * the next tick. The nohz full kick at least implies that.
696 * If needed we can still optimize that later with an
699 if (tick_nohz_full_cpu(cpu)) {
700 if (cpu != smp_processor_id() ||
701 tick_nohz_tick_stopped())
702 tick_nohz_full_kick_cpu(cpu);
709 void wake_up_nohz_cpu(int cpu)
711 if (!wake_up_full_nohz_cpu(cpu))
712 wake_up_idle_cpu(cpu);
715 static inline bool got_nohz_idle_kick(void)
717 int cpu = smp_processor_id();
719 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
722 if (idle_cpu(cpu) && !need_resched())
726 * We can't run Idle Load Balance on this CPU for this time so we
727 * cancel it and clear NOHZ_BALANCE_KICK
729 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
733 #else /* CONFIG_NO_HZ_COMMON */
735 static inline bool got_nohz_idle_kick(void)
740 #endif /* CONFIG_NO_HZ_COMMON */
742 #ifdef CONFIG_NO_HZ_FULL
743 bool sched_can_stop_tick(void)
746 * More than one running task need preemption.
747 * nr_running update is assumed to be visible
748 * after IPI is sent from wakers.
750 if (this_rq()->nr_running > 1)
755 #endif /* CONFIG_NO_HZ_FULL */
757 void sched_avg_update(struct rq *rq)
759 s64 period = sched_avg_period();
761 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
763 * Inline assembly required to prevent the compiler
764 * optimising this loop into a divmod call.
765 * See __iter_div_u64_rem() for another example of this.
767 asm("" : "+rm" (rq->age_stamp));
768 rq->age_stamp += period;
773 #endif /* CONFIG_SMP */
775 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
776 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
778 * Iterate task_group tree rooted at *from, calling @down when first entering a
779 * node and @up when leaving it for the final time.
781 * Caller must hold rcu_lock or sufficient equivalent.
783 int walk_tg_tree_from(struct task_group *from,
784 tg_visitor down, tg_visitor up, void *data)
786 struct task_group *parent, *child;
792 ret = (*down)(parent, data);
795 list_for_each_entry_rcu(child, &parent->children, siblings) {
802 ret = (*up)(parent, data);
803 if (ret || parent == from)
807 parent = parent->parent;
814 int tg_nop(struct task_group *tg, void *data)
820 static void set_load_weight(struct task_struct *p)
822 int prio = p->static_prio - MAX_RT_PRIO;
823 struct load_weight *load = &p->se.load;
826 * SCHED_IDLE tasks get minimal weight:
828 if (p->policy == SCHED_IDLE) {
829 load->weight = scale_load(WEIGHT_IDLEPRIO);
830 load->inv_weight = WMULT_IDLEPRIO;
834 load->weight = scale_load(prio_to_weight[prio]);
835 load->inv_weight = prio_to_wmult[prio];
838 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
841 sched_info_queued(rq, p);
842 p->sched_class->enqueue_task(rq, p, flags);
845 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
848 sched_info_dequeued(rq, p);
849 p->sched_class->dequeue_task(rq, p, flags);
852 void activate_task(struct rq *rq, struct task_struct *p, int flags)
854 if (task_contributes_to_load(p))
855 rq->nr_uninterruptible--;
857 enqueue_task(rq, p, flags);
860 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
862 if (task_contributes_to_load(p))
863 rq->nr_uninterruptible++;
865 dequeue_task(rq, p, flags);
868 static void update_rq_clock_task(struct rq *rq, s64 delta)
871 * In theory, the compile should just see 0 here, and optimize out the call
872 * to sched_rt_avg_update. But I don't trust it...
874 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
875 s64 steal = 0, irq_delta = 0;
877 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
878 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
881 * Since irq_time is only updated on {soft,}irq_exit, we might run into
882 * this case when a previous update_rq_clock() happened inside a
885 * When this happens, we stop ->clock_task and only update the
886 * prev_irq_time stamp to account for the part that fit, so that a next
887 * update will consume the rest. This ensures ->clock_task is
890 * It does however cause some slight miss-attribution of {soft,}irq
891 * time, a more accurate solution would be to update the irq_time using
892 * the current rq->clock timestamp, except that would require using
895 if (irq_delta > delta)
898 rq->prev_irq_time += irq_delta;
901 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
902 if (static_key_false((¶virt_steal_rq_enabled))) {
903 steal = paravirt_steal_clock(cpu_of(rq));
904 steal -= rq->prev_steal_time_rq;
906 if (unlikely(steal > delta))
909 rq->prev_steal_time_rq += steal;
914 rq->clock_task += delta;
916 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
917 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
918 sched_rt_avg_update(rq, irq_delta + steal);
922 void sched_set_stop_task(int cpu, struct task_struct *stop)
924 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
925 struct task_struct *old_stop = cpu_rq(cpu)->stop;
929 * Make it appear like a SCHED_FIFO task, its something
930 * userspace knows about and won't get confused about.
932 * Also, it will make PI more or less work without too
933 * much confusion -- but then, stop work should not
934 * rely on PI working anyway.
936 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
938 stop->sched_class = &stop_sched_class;
941 cpu_rq(cpu)->stop = stop;
945 * Reset it back to a normal scheduling class so that
946 * it can die in pieces.
948 old_stop->sched_class = &rt_sched_class;
953 * __normal_prio - return the priority that is based on the static prio
955 static inline int __normal_prio(struct task_struct *p)
957 return p->static_prio;
961 * Calculate the expected normal priority: i.e. priority
962 * without taking RT-inheritance into account. Might be
963 * boosted by interactivity modifiers. Changes upon fork,
964 * setprio syscalls, and whenever the interactivity
965 * estimator recalculates.
967 static inline int normal_prio(struct task_struct *p)
971 if (task_has_dl_policy(p))
972 prio = MAX_DL_PRIO-1;
973 else if (task_has_rt_policy(p))
974 prio = MAX_RT_PRIO-1 - p->rt_priority;
976 prio = __normal_prio(p);
981 * Calculate the current priority, i.e. the priority
982 * taken into account by the scheduler. This value might
983 * be boosted by RT tasks, or might be boosted by
984 * interactivity modifiers. Will be RT if the task got
985 * RT-boosted. If not then it returns p->normal_prio.
987 static int effective_prio(struct task_struct *p)
989 p->normal_prio = normal_prio(p);
991 * If we are RT tasks or we were boosted to RT priority,
992 * keep the priority unchanged. Otherwise, update priority
993 * to the normal priority:
995 if (!rt_prio(p->prio))
996 return p->normal_prio;
1001 * task_curr - is this task currently executing on a CPU?
1002 * @p: the task in question.
1004 * Return: 1 if the task is currently executing. 0 otherwise.
1006 inline int task_curr(const struct task_struct *p)
1008 return cpu_curr(task_cpu(p)) == p;
1011 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1012 const struct sched_class *prev_class,
1015 if (prev_class != p->sched_class) {
1016 if (prev_class->switched_from)
1017 prev_class->switched_from(rq, p);
1018 p->sched_class->switched_to(rq, p);
1019 } else if (oldprio != p->prio || dl_task(p))
1020 p->sched_class->prio_changed(rq, p, oldprio);
1023 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1025 const struct sched_class *class;
1027 if (p->sched_class == rq->curr->sched_class) {
1028 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1030 for_each_class(class) {
1031 if (class == rq->curr->sched_class)
1033 if (class == p->sched_class) {
1041 * A queue event has occurred, and we're going to schedule. In
1042 * this case, we can save a useless back to back clock update.
1044 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1045 rq->skip_clock_update = 1;
1049 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1051 #ifdef CONFIG_SCHED_DEBUG
1053 * We should never call set_task_cpu() on a blocked task,
1054 * ttwu() will sort out the placement.
1056 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1059 #ifdef CONFIG_LOCKDEP
1061 * The caller should hold either p->pi_lock or rq->lock, when changing
1062 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1064 * sched_move_task() holds both and thus holding either pins the cgroup,
1067 * Furthermore, all task_rq users should acquire both locks, see
1070 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1071 lockdep_is_held(&task_rq(p)->lock)));
1075 trace_sched_migrate_task(p, new_cpu);
1077 if (task_cpu(p) != new_cpu) {
1078 if (p->sched_class->migrate_task_rq)
1079 p->sched_class->migrate_task_rq(p, new_cpu);
1080 p->se.nr_migrations++;
1081 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1084 __set_task_cpu(p, new_cpu);
1087 static void __migrate_swap_task(struct task_struct *p, int cpu)
1089 if (task_on_rq_queued(p)) {
1090 struct rq *src_rq, *dst_rq;
1092 src_rq = task_rq(p);
1093 dst_rq = cpu_rq(cpu);
1095 deactivate_task(src_rq, p, 0);
1096 set_task_cpu(p, cpu);
1097 activate_task(dst_rq, p, 0);
1098 check_preempt_curr(dst_rq, p, 0);
1101 * Task isn't running anymore; make it appear like we migrated
1102 * it before it went to sleep. This means on wakeup we make the
1103 * previous cpu our targer instead of where it really is.
1109 struct migration_swap_arg {
1110 struct task_struct *src_task, *dst_task;
1111 int src_cpu, dst_cpu;
1114 static int migrate_swap_stop(void *data)
1116 struct migration_swap_arg *arg = data;
1117 struct rq *src_rq, *dst_rq;
1120 src_rq = cpu_rq(arg->src_cpu);
1121 dst_rq = cpu_rq(arg->dst_cpu);
1123 double_raw_lock(&arg->src_task->pi_lock,
1124 &arg->dst_task->pi_lock);
1125 double_rq_lock(src_rq, dst_rq);
1126 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1129 if (task_cpu(arg->src_task) != arg->src_cpu)
1132 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1135 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1138 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1139 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1144 double_rq_unlock(src_rq, dst_rq);
1145 raw_spin_unlock(&arg->dst_task->pi_lock);
1146 raw_spin_unlock(&arg->src_task->pi_lock);
1152 * Cross migrate two tasks
1154 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1156 struct migration_swap_arg arg;
1159 arg = (struct migration_swap_arg){
1161 .src_cpu = task_cpu(cur),
1163 .dst_cpu = task_cpu(p),
1166 if (arg.src_cpu == arg.dst_cpu)
1170 * These three tests are all lockless; this is OK since all of them
1171 * will be re-checked with proper locks held further down the line.
1173 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1176 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1179 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1182 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1183 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1189 struct migration_arg {
1190 struct task_struct *task;
1194 static int migration_cpu_stop(void *data);
1197 * wait_task_inactive - wait for a thread to unschedule.
1199 * If @match_state is nonzero, it's the @p->state value just checked and
1200 * not expected to change. If it changes, i.e. @p might have woken up,
1201 * then return zero. When we succeed in waiting for @p to be off its CPU,
1202 * we return a positive number (its total switch count). If a second call
1203 * a short while later returns the same number, the caller can be sure that
1204 * @p has remained unscheduled the whole time.
1206 * The caller must ensure that the task *will* unschedule sometime soon,
1207 * else this function might spin for a *long* time. This function can't
1208 * be called with interrupts off, or it may introduce deadlock with
1209 * smp_call_function() if an IPI is sent by the same process we are
1210 * waiting to become inactive.
1212 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1214 unsigned long flags;
1215 int running, queued;
1221 * We do the initial early heuristics without holding
1222 * any task-queue locks at all. We'll only try to get
1223 * the runqueue lock when things look like they will
1229 * If the task is actively running on another CPU
1230 * still, just relax and busy-wait without holding
1233 * NOTE! Since we don't hold any locks, it's not
1234 * even sure that "rq" stays as the right runqueue!
1235 * But we don't care, since "task_running()" will
1236 * return false if the runqueue has changed and p
1237 * is actually now running somewhere else!
1239 while (task_running(rq, p)) {
1240 if (match_state && unlikely(p->state != match_state))
1246 * Ok, time to look more closely! We need the rq
1247 * lock now, to be *sure*. If we're wrong, we'll
1248 * just go back and repeat.
1250 rq = task_rq_lock(p, &flags);
1251 trace_sched_wait_task(p);
1252 running = task_running(rq, p);
1253 queued = task_on_rq_queued(p);
1255 if (!match_state || p->state == match_state)
1256 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1257 task_rq_unlock(rq, p, &flags);
1260 * If it changed from the expected state, bail out now.
1262 if (unlikely(!ncsw))
1266 * Was it really running after all now that we
1267 * checked with the proper locks actually held?
1269 * Oops. Go back and try again..
1271 if (unlikely(running)) {
1277 * It's not enough that it's not actively running,
1278 * it must be off the runqueue _entirely_, and not
1281 * So if it was still runnable (but just not actively
1282 * running right now), it's preempted, and we should
1283 * yield - it could be a while.
1285 if (unlikely(queued)) {
1286 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1288 set_current_state(TASK_UNINTERRUPTIBLE);
1289 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1294 * Ahh, all good. It wasn't running, and it wasn't
1295 * runnable, which means that it will never become
1296 * running in the future either. We're all done!
1305 * kick_process - kick a running thread to enter/exit the kernel
1306 * @p: the to-be-kicked thread
1308 * Cause a process which is running on another CPU to enter
1309 * kernel-mode, without any delay. (to get signals handled.)
1311 * NOTE: this function doesn't have to take the runqueue lock,
1312 * because all it wants to ensure is that the remote task enters
1313 * the kernel. If the IPI races and the task has been migrated
1314 * to another CPU then no harm is done and the purpose has been
1317 void kick_process(struct task_struct *p)
1323 if ((cpu != smp_processor_id()) && task_curr(p))
1324 smp_send_reschedule(cpu);
1327 EXPORT_SYMBOL_GPL(kick_process);
1328 #endif /* CONFIG_SMP */
1332 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1334 static int select_fallback_rq(int cpu, struct task_struct *p)
1336 int nid = cpu_to_node(cpu);
1337 const struct cpumask *nodemask = NULL;
1338 enum { cpuset, possible, fail } state = cpuset;
1342 * If the node that the cpu is on has been offlined, cpu_to_node()
1343 * will return -1. There is no cpu on the node, and we should
1344 * select the cpu on the other node.
1347 nodemask = cpumask_of_node(nid);
1349 /* Look for allowed, online CPU in same node. */
1350 for_each_cpu(dest_cpu, nodemask) {
1351 if (!cpu_online(dest_cpu))
1353 if (!cpu_active(dest_cpu))
1355 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1361 /* Any allowed, online CPU? */
1362 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1363 if (!cpu_online(dest_cpu))
1365 if (!cpu_active(dest_cpu))
1372 /* No more Mr. Nice Guy. */
1373 cpuset_cpus_allowed_fallback(p);
1378 do_set_cpus_allowed(p, cpu_possible_mask);
1389 if (state != cpuset) {
1391 * Don't tell them about moving exiting tasks or
1392 * kernel threads (both mm NULL), since they never
1395 if (p->mm && printk_ratelimit()) {
1396 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1397 task_pid_nr(p), p->comm, cpu);
1405 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1408 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1410 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1413 * In order not to call set_task_cpu() on a blocking task we need
1414 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1417 * Since this is common to all placement strategies, this lives here.
1419 * [ this allows ->select_task() to simply return task_cpu(p) and
1420 * not worry about this generic constraint ]
1422 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1424 cpu = select_fallback_rq(task_cpu(p), p);
1429 static void update_avg(u64 *avg, u64 sample)
1431 s64 diff = sample - *avg;
1437 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1439 #ifdef CONFIG_SCHEDSTATS
1440 struct rq *rq = this_rq();
1443 int this_cpu = smp_processor_id();
1445 if (cpu == this_cpu) {
1446 schedstat_inc(rq, ttwu_local);
1447 schedstat_inc(p, se.statistics.nr_wakeups_local);
1449 struct sched_domain *sd;
1451 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1453 for_each_domain(this_cpu, sd) {
1454 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1455 schedstat_inc(sd, ttwu_wake_remote);
1462 if (wake_flags & WF_MIGRATED)
1463 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1465 #endif /* CONFIG_SMP */
1467 schedstat_inc(rq, ttwu_count);
1468 schedstat_inc(p, se.statistics.nr_wakeups);
1470 if (wake_flags & WF_SYNC)
1471 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1473 #endif /* CONFIG_SCHEDSTATS */
1476 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1478 activate_task(rq, p, en_flags);
1479 p->on_rq = TASK_ON_RQ_QUEUED;
1481 /* if a worker is waking up, notify workqueue */
1482 if (p->flags & PF_WQ_WORKER)
1483 wq_worker_waking_up(p, cpu_of(rq));
1487 * Mark the task runnable and perform wakeup-preemption.
1490 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1492 check_preempt_curr(rq, p, wake_flags);
1493 trace_sched_wakeup(p, true);
1495 p->state = TASK_RUNNING;
1497 if (p->sched_class->task_woken)
1498 p->sched_class->task_woken(rq, p);
1500 if (rq->idle_stamp) {
1501 u64 delta = rq_clock(rq) - rq->idle_stamp;
1502 u64 max = 2*rq->max_idle_balance_cost;
1504 update_avg(&rq->avg_idle, delta);
1506 if (rq->avg_idle > max)
1515 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1518 if (p->sched_contributes_to_load)
1519 rq->nr_uninterruptible--;
1522 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1523 ttwu_do_wakeup(rq, p, wake_flags);
1527 * Called in case the task @p isn't fully descheduled from its runqueue,
1528 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1529 * since all we need to do is flip p->state to TASK_RUNNING, since
1530 * the task is still ->on_rq.
1532 static int ttwu_remote(struct task_struct *p, int wake_flags)
1537 rq = __task_rq_lock(p);
1538 if (task_on_rq_queued(p)) {
1539 /* check_preempt_curr() may use rq clock */
1540 update_rq_clock(rq);
1541 ttwu_do_wakeup(rq, p, wake_flags);
1544 __task_rq_unlock(rq);
1550 void sched_ttwu_pending(void)
1552 struct rq *rq = this_rq();
1553 struct llist_node *llist = llist_del_all(&rq->wake_list);
1554 struct task_struct *p;
1555 unsigned long flags;
1560 raw_spin_lock_irqsave(&rq->lock, flags);
1563 p = llist_entry(llist, struct task_struct, wake_entry);
1564 llist = llist_next(llist);
1565 ttwu_do_activate(rq, p, 0);
1568 raw_spin_unlock_irqrestore(&rq->lock, flags);
1571 void scheduler_ipi(void)
1574 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1575 * TIF_NEED_RESCHED remotely (for the first time) will also send
1578 preempt_fold_need_resched();
1580 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1584 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1585 * traditionally all their work was done from the interrupt return
1586 * path. Now that we actually do some work, we need to make sure
1589 * Some archs already do call them, luckily irq_enter/exit nest
1592 * Arguably we should visit all archs and update all handlers,
1593 * however a fair share of IPIs are still resched only so this would
1594 * somewhat pessimize the simple resched case.
1597 sched_ttwu_pending();
1600 * Check if someone kicked us for doing the nohz idle load balance.
1602 if (unlikely(got_nohz_idle_kick())) {
1603 this_rq()->idle_balance = 1;
1604 raise_softirq_irqoff(SCHED_SOFTIRQ);
1609 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1611 struct rq *rq = cpu_rq(cpu);
1613 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1614 if (!set_nr_if_polling(rq->idle))
1615 smp_send_reschedule(cpu);
1617 trace_sched_wake_idle_without_ipi(cpu);
1621 void wake_up_if_idle(int cpu)
1623 struct rq *rq = cpu_rq(cpu);
1624 unsigned long flags;
1626 if (!is_idle_task(rq->curr))
1629 if (set_nr_if_polling(rq->idle)) {
1630 trace_sched_wake_idle_without_ipi(cpu);
1632 raw_spin_lock_irqsave(&rq->lock, flags);
1633 if (is_idle_task(rq->curr))
1634 smp_send_reschedule(cpu);
1635 /* Else cpu is not in idle, do nothing here */
1636 raw_spin_unlock_irqrestore(&rq->lock, flags);
1640 bool cpus_share_cache(int this_cpu, int that_cpu)
1642 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1644 #endif /* CONFIG_SMP */
1646 static void ttwu_queue(struct task_struct *p, int cpu)
1648 struct rq *rq = cpu_rq(cpu);
1650 #if defined(CONFIG_SMP)
1651 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1652 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1653 ttwu_queue_remote(p, cpu);
1658 raw_spin_lock(&rq->lock);
1659 ttwu_do_activate(rq, p, 0);
1660 raw_spin_unlock(&rq->lock);
1664 * try_to_wake_up - wake up a thread
1665 * @p: the thread to be awakened
1666 * @state: the mask of task states that can be woken
1667 * @wake_flags: wake modifier flags (WF_*)
1669 * Put it on the run-queue if it's not already there. The "current"
1670 * thread is always on the run-queue (except when the actual
1671 * re-schedule is in progress), and as such you're allowed to do
1672 * the simpler "current->state = TASK_RUNNING" to mark yourself
1673 * runnable without the overhead of this.
1675 * Return: %true if @p was woken up, %false if it was already running.
1676 * or @state didn't match @p's state.
1679 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1681 unsigned long flags;
1682 int cpu, success = 0;
1685 * If we are going to wake up a thread waiting for CONDITION we
1686 * need to ensure that CONDITION=1 done by the caller can not be
1687 * reordered with p->state check below. This pairs with mb() in
1688 * set_current_state() the waiting thread does.
1690 smp_mb__before_spinlock();
1691 raw_spin_lock_irqsave(&p->pi_lock, flags);
1692 if (!(p->state & state))
1695 success = 1; /* we're going to change ->state */
1698 if (p->on_rq && ttwu_remote(p, wake_flags))
1703 * If the owning (remote) cpu is still in the middle of schedule() with
1704 * this task as prev, wait until its done referencing the task.
1709 * Pairs with the smp_wmb() in finish_lock_switch().
1713 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1714 p->state = TASK_WAKING;
1716 if (p->sched_class->task_waking)
1717 p->sched_class->task_waking(p);
1719 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1720 if (task_cpu(p) != cpu) {
1721 wake_flags |= WF_MIGRATED;
1722 set_task_cpu(p, cpu);
1724 #endif /* CONFIG_SMP */
1728 ttwu_stat(p, cpu, wake_flags);
1730 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1736 * try_to_wake_up_local - try to wake up a local task with rq lock held
1737 * @p: the thread to be awakened
1739 * Put @p on the run-queue if it's not already there. The caller must
1740 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1743 static void try_to_wake_up_local(struct task_struct *p)
1745 struct rq *rq = task_rq(p);
1747 if (WARN_ON_ONCE(rq != this_rq()) ||
1748 WARN_ON_ONCE(p == current))
1751 lockdep_assert_held(&rq->lock);
1753 if (!raw_spin_trylock(&p->pi_lock)) {
1754 raw_spin_unlock(&rq->lock);
1755 raw_spin_lock(&p->pi_lock);
1756 raw_spin_lock(&rq->lock);
1759 if (!(p->state & TASK_NORMAL))
1762 if (!task_on_rq_queued(p))
1763 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1765 ttwu_do_wakeup(rq, p, 0);
1766 ttwu_stat(p, smp_processor_id(), 0);
1768 raw_spin_unlock(&p->pi_lock);
1772 * wake_up_process - Wake up a specific process
1773 * @p: The process to be woken up.
1775 * Attempt to wake up the nominated process and move it to the set of runnable
1778 * Return: 1 if the process was woken up, 0 if it was already running.
1780 * It may be assumed that this function implies a write memory barrier before
1781 * changing the task state if and only if any tasks are woken up.
1783 int wake_up_process(struct task_struct *p)
1785 WARN_ON(task_is_stopped_or_traced(p));
1786 return try_to_wake_up(p, TASK_NORMAL, 0);
1788 EXPORT_SYMBOL(wake_up_process);
1790 int wake_up_state(struct task_struct *p, unsigned int state)
1792 return try_to_wake_up(p, state, 0);
1796 * This function clears the sched_dl_entity static params.
1798 void __dl_clear_params(struct task_struct *p)
1800 struct sched_dl_entity *dl_se = &p->dl;
1802 dl_se->dl_runtime = 0;
1803 dl_se->dl_deadline = 0;
1804 dl_se->dl_period = 0;
1810 * Perform scheduler related setup for a newly forked process p.
1811 * p is forked by current.
1813 * __sched_fork() is basic setup used by init_idle() too:
1815 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1820 p->se.exec_start = 0;
1821 p->se.sum_exec_runtime = 0;
1822 p->se.prev_sum_exec_runtime = 0;
1823 p->se.nr_migrations = 0;
1825 INIT_LIST_HEAD(&p->se.group_node);
1827 #ifdef CONFIG_SCHEDSTATS
1828 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1831 RB_CLEAR_NODE(&p->dl.rb_node);
1832 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1833 __dl_clear_params(p);
1835 INIT_LIST_HEAD(&p->rt.run_list);
1837 #ifdef CONFIG_PREEMPT_NOTIFIERS
1838 INIT_HLIST_HEAD(&p->preempt_notifiers);
1841 #ifdef CONFIG_NUMA_BALANCING
1842 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1843 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1844 p->mm->numa_scan_seq = 0;
1847 if (clone_flags & CLONE_VM)
1848 p->numa_preferred_nid = current->numa_preferred_nid;
1850 p->numa_preferred_nid = -1;
1852 p->node_stamp = 0ULL;
1853 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1854 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1855 p->numa_work.next = &p->numa_work;
1856 p->numa_faults_memory = NULL;
1857 p->numa_faults_buffer_memory = NULL;
1858 p->last_task_numa_placement = 0;
1859 p->last_sum_exec_runtime = 0;
1861 INIT_LIST_HEAD(&p->numa_entry);
1862 p->numa_group = NULL;
1863 #endif /* CONFIG_NUMA_BALANCING */
1866 #ifdef CONFIG_NUMA_BALANCING
1867 #ifdef CONFIG_SCHED_DEBUG
1868 void set_numabalancing_state(bool enabled)
1871 sched_feat_set("NUMA");
1873 sched_feat_set("NO_NUMA");
1876 __read_mostly bool numabalancing_enabled;
1878 void set_numabalancing_state(bool enabled)
1880 numabalancing_enabled = enabled;
1882 #endif /* CONFIG_SCHED_DEBUG */
1884 #ifdef CONFIG_PROC_SYSCTL
1885 int sysctl_numa_balancing(struct ctl_table *table, int write,
1886 void __user *buffer, size_t *lenp, loff_t *ppos)
1890 int state = numabalancing_enabled;
1892 if (write && !capable(CAP_SYS_ADMIN))
1897 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1901 set_numabalancing_state(state);
1908 * fork()/clone()-time setup:
1910 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1912 unsigned long flags;
1913 int cpu = get_cpu();
1915 __sched_fork(clone_flags, p);
1917 * We mark the process as running here. This guarantees that
1918 * nobody will actually run it, and a signal or other external
1919 * event cannot wake it up and insert it on the runqueue either.
1921 p->state = TASK_RUNNING;
1924 * Make sure we do not leak PI boosting priority to the child.
1926 p->prio = current->normal_prio;
1929 * Revert to default priority/policy on fork if requested.
1931 if (unlikely(p->sched_reset_on_fork)) {
1932 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1933 p->policy = SCHED_NORMAL;
1934 p->static_prio = NICE_TO_PRIO(0);
1936 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1937 p->static_prio = NICE_TO_PRIO(0);
1939 p->prio = p->normal_prio = __normal_prio(p);
1943 * We don't need the reset flag anymore after the fork. It has
1944 * fulfilled its duty:
1946 p->sched_reset_on_fork = 0;
1949 if (dl_prio(p->prio)) {
1952 } else if (rt_prio(p->prio)) {
1953 p->sched_class = &rt_sched_class;
1955 p->sched_class = &fair_sched_class;
1958 if (p->sched_class->task_fork)
1959 p->sched_class->task_fork(p);
1962 * The child is not yet in the pid-hash so no cgroup attach races,
1963 * and the cgroup is pinned to this child due to cgroup_fork()
1964 * is ran before sched_fork().
1966 * Silence PROVE_RCU.
1968 raw_spin_lock_irqsave(&p->pi_lock, flags);
1969 set_task_cpu(p, cpu);
1970 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1972 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1973 if (likely(sched_info_on()))
1974 memset(&p->sched_info, 0, sizeof(p->sched_info));
1976 #if defined(CONFIG_SMP)
1979 init_task_preempt_count(p);
1981 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1982 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1989 unsigned long to_ratio(u64 period, u64 runtime)
1991 if (runtime == RUNTIME_INF)
1995 * Doing this here saves a lot of checks in all
1996 * the calling paths, and returning zero seems
1997 * safe for them anyway.
2002 return div64_u64(runtime << 20, period);
2006 inline struct dl_bw *dl_bw_of(int i)
2008 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2009 "sched RCU must be held");
2010 return &cpu_rq(i)->rd->dl_bw;
2013 static inline int dl_bw_cpus(int i)
2015 struct root_domain *rd = cpu_rq(i)->rd;
2018 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2019 "sched RCU must be held");
2020 for_each_cpu_and(i, rd->span, cpu_active_mask)
2026 inline struct dl_bw *dl_bw_of(int i)
2028 return &cpu_rq(i)->dl.dl_bw;
2031 static inline int dl_bw_cpus(int i)
2038 * We must be sure that accepting a new task (or allowing changing the
2039 * parameters of an existing one) is consistent with the bandwidth
2040 * constraints. If yes, this function also accordingly updates the currently
2041 * allocated bandwidth to reflect the new situation.
2043 * This function is called while holding p's rq->lock.
2045 static int dl_overflow(struct task_struct *p, int policy,
2046 const struct sched_attr *attr)
2049 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2050 u64 period = attr->sched_period ?: attr->sched_deadline;
2051 u64 runtime = attr->sched_runtime;
2052 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2055 if (new_bw == p->dl.dl_bw)
2059 * Either if a task, enters, leave, or stays -deadline but changes
2060 * its parameters, we may need to update accordingly the total
2061 * allocated bandwidth of the container.
2063 raw_spin_lock(&dl_b->lock);
2064 cpus = dl_bw_cpus(task_cpu(p));
2065 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2066 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2067 __dl_add(dl_b, new_bw);
2069 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2070 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2071 __dl_clear(dl_b, p->dl.dl_bw);
2072 __dl_add(dl_b, new_bw);
2074 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2075 __dl_clear(dl_b, p->dl.dl_bw);
2078 raw_spin_unlock(&dl_b->lock);
2083 extern void init_dl_bw(struct dl_bw *dl_b);
2086 * wake_up_new_task - wake up a newly created task for the first time.
2088 * This function will do some initial scheduler statistics housekeeping
2089 * that must be done for every newly created context, then puts the task
2090 * on the runqueue and wakes it.
2092 void wake_up_new_task(struct task_struct *p)
2094 unsigned long flags;
2097 raw_spin_lock_irqsave(&p->pi_lock, flags);
2100 * Fork balancing, do it here and not earlier because:
2101 * - cpus_allowed can change in the fork path
2102 * - any previously selected cpu might disappear through hotplug
2104 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2107 /* Initialize new task's runnable average */
2108 init_task_runnable_average(p);
2109 rq = __task_rq_lock(p);
2110 activate_task(rq, p, 0);
2111 p->on_rq = TASK_ON_RQ_QUEUED;
2112 trace_sched_wakeup_new(p, true);
2113 check_preempt_curr(rq, p, WF_FORK);
2115 if (p->sched_class->task_woken)
2116 p->sched_class->task_woken(rq, p);
2118 task_rq_unlock(rq, p, &flags);
2121 #ifdef CONFIG_PREEMPT_NOTIFIERS
2124 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2125 * @notifier: notifier struct to register
2127 void preempt_notifier_register(struct preempt_notifier *notifier)
2129 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2131 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2134 * preempt_notifier_unregister - no longer interested in preemption notifications
2135 * @notifier: notifier struct to unregister
2137 * This is safe to call from within a preemption notifier.
2139 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2141 hlist_del(¬ifier->link);
2143 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2145 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2147 struct preempt_notifier *notifier;
2149 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2150 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2154 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2155 struct task_struct *next)
2157 struct preempt_notifier *notifier;
2159 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2160 notifier->ops->sched_out(notifier, next);
2163 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2165 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2170 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2171 struct task_struct *next)
2175 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2178 * prepare_task_switch - prepare to switch tasks
2179 * @rq: the runqueue preparing to switch
2180 * @prev: the current task that is being switched out
2181 * @next: the task we are going to switch to.
2183 * This is called with the rq lock held and interrupts off. It must
2184 * be paired with a subsequent finish_task_switch after the context
2187 * prepare_task_switch sets up locking and calls architecture specific
2191 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2192 struct task_struct *next)
2194 trace_sched_switch(prev, next);
2195 sched_info_switch(rq, prev, next);
2196 perf_event_task_sched_out(prev, next);
2197 fire_sched_out_preempt_notifiers(prev, next);
2198 prepare_lock_switch(rq, next);
2199 prepare_arch_switch(next);
2203 * finish_task_switch - clean up after a task-switch
2204 * @prev: the thread we just switched away from.
2206 * finish_task_switch must be called after the context switch, paired
2207 * with a prepare_task_switch call before the context switch.
2208 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2209 * and do any other architecture-specific cleanup actions.
2211 * Note that we may have delayed dropping an mm in context_switch(). If
2212 * so, we finish that here outside of the runqueue lock. (Doing it
2213 * with the lock held can cause deadlocks; see schedule() for
2216 * The context switch have flipped the stack from under us and restored the
2217 * local variables which were saved when this task called schedule() in the
2218 * past. prev == current is still correct but we need to recalculate this_rq
2219 * because prev may have moved to another CPU.
2221 static struct rq *finish_task_switch(struct task_struct *prev)
2222 __releases(rq->lock)
2224 struct rq *rq = this_rq();
2225 struct mm_struct *mm = rq->prev_mm;
2231 * A task struct has one reference for the use as "current".
2232 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2233 * schedule one last time. The schedule call will never return, and
2234 * the scheduled task must drop that reference.
2235 * The test for TASK_DEAD must occur while the runqueue locks are
2236 * still held, otherwise prev could be scheduled on another cpu, die
2237 * there before we look at prev->state, and then the reference would
2239 * Manfred Spraul <manfred@colorfullife.com>
2241 prev_state = prev->state;
2242 vtime_task_switch(prev);
2243 finish_arch_switch(prev);
2244 perf_event_task_sched_in(prev, current);
2245 finish_lock_switch(rq, prev);
2246 finish_arch_post_lock_switch();
2248 fire_sched_in_preempt_notifiers(current);
2251 if (unlikely(prev_state == TASK_DEAD)) {
2252 if (prev->sched_class->task_dead)
2253 prev->sched_class->task_dead(prev);
2256 * Remove function-return probe instances associated with this
2257 * task and put them back on the free list.
2259 kprobe_flush_task(prev);
2260 put_task_struct(prev);
2263 tick_nohz_task_switch(current);
2269 /* rq->lock is NOT held, but preemption is disabled */
2270 static inline void post_schedule(struct rq *rq)
2272 if (rq->post_schedule) {
2273 unsigned long flags;
2275 raw_spin_lock_irqsave(&rq->lock, flags);
2276 if (rq->curr->sched_class->post_schedule)
2277 rq->curr->sched_class->post_schedule(rq);
2278 raw_spin_unlock_irqrestore(&rq->lock, flags);
2280 rq->post_schedule = 0;
2286 static inline void post_schedule(struct rq *rq)
2293 * schedule_tail - first thing a freshly forked thread must call.
2294 * @prev: the thread we just switched away from.
2296 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2297 __releases(rq->lock)
2301 /* finish_task_switch() drops rq->lock and enables preemtion */
2303 rq = finish_task_switch(prev);
2307 if (current->set_child_tid)
2308 put_user(task_pid_vnr(current), current->set_child_tid);
2312 * context_switch - switch to the new MM and the new thread's register state.
2314 static inline struct rq *
2315 context_switch(struct rq *rq, struct task_struct *prev,
2316 struct task_struct *next)
2318 struct mm_struct *mm, *oldmm;
2320 prepare_task_switch(rq, prev, next);
2323 oldmm = prev->active_mm;
2325 * For paravirt, this is coupled with an exit in switch_to to
2326 * combine the page table reload and the switch backend into
2329 arch_start_context_switch(prev);
2332 next->active_mm = oldmm;
2333 atomic_inc(&oldmm->mm_count);
2334 enter_lazy_tlb(oldmm, next);
2336 switch_mm(oldmm, mm, next);
2339 prev->active_mm = NULL;
2340 rq->prev_mm = oldmm;
2343 * Since the runqueue lock will be released by the next
2344 * task (which is an invalid locking op but in the case
2345 * of the scheduler it's an obvious special-case), so we
2346 * do an early lockdep release here:
2348 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2350 context_tracking_task_switch(prev, next);
2351 /* Here we just switch the register state and the stack. */
2352 switch_to(prev, next, prev);
2355 return finish_task_switch(prev);
2359 * nr_running and nr_context_switches:
2361 * externally visible scheduler statistics: current number of runnable
2362 * threads, total number of context switches performed since bootup.
2364 unsigned long nr_running(void)
2366 unsigned long i, sum = 0;
2368 for_each_online_cpu(i)
2369 sum += cpu_rq(i)->nr_running;
2375 * Check if only the current task is running on the cpu.
2377 bool single_task_running(void)
2379 if (cpu_rq(smp_processor_id())->nr_running == 1)
2384 EXPORT_SYMBOL(single_task_running);
2386 unsigned long long nr_context_switches(void)
2389 unsigned long long sum = 0;
2391 for_each_possible_cpu(i)
2392 sum += cpu_rq(i)->nr_switches;
2397 unsigned long nr_iowait(void)
2399 unsigned long i, sum = 0;
2401 for_each_possible_cpu(i)
2402 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2407 unsigned long nr_iowait_cpu(int cpu)
2409 struct rq *this = cpu_rq(cpu);
2410 return atomic_read(&this->nr_iowait);
2413 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2415 struct rq *this = this_rq();
2416 *nr_waiters = atomic_read(&this->nr_iowait);
2417 *load = this->cpu_load[0];
2423 * sched_exec - execve() is a valuable balancing opportunity, because at
2424 * this point the task has the smallest effective memory and cache footprint.
2426 void sched_exec(void)
2428 struct task_struct *p = current;
2429 unsigned long flags;
2432 raw_spin_lock_irqsave(&p->pi_lock, flags);
2433 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2434 if (dest_cpu == smp_processor_id())
2437 if (likely(cpu_active(dest_cpu))) {
2438 struct migration_arg arg = { p, dest_cpu };
2440 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2441 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2445 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2450 DEFINE_PER_CPU(struct kernel_stat, kstat);
2451 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2453 EXPORT_PER_CPU_SYMBOL(kstat);
2454 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2457 * Return any ns on the sched_clock that have not yet been accounted in
2458 * @p in case that task is currently running.
2460 * Called with task_rq_lock() held on @rq.
2462 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2467 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2468 * project cycles that may never be accounted to this
2469 * thread, breaking clock_gettime().
2471 if (task_current(rq, p) && task_on_rq_queued(p)) {
2472 update_rq_clock(rq);
2473 ns = rq_clock_task(rq) - p->se.exec_start;
2481 unsigned long long task_delta_exec(struct task_struct *p)
2483 unsigned long flags;
2487 rq = task_rq_lock(p, &flags);
2488 ns = do_task_delta_exec(p, rq);
2489 task_rq_unlock(rq, p, &flags);
2495 * Return accounted runtime for the task.
2496 * In case the task is currently running, return the runtime plus current's
2497 * pending runtime that have not been accounted yet.
2499 unsigned long long task_sched_runtime(struct task_struct *p)
2501 unsigned long flags;
2505 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2507 * 64-bit doesn't need locks to atomically read a 64bit value.
2508 * So we have a optimization chance when the task's delta_exec is 0.
2509 * Reading ->on_cpu is racy, but this is ok.
2511 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2512 * If we race with it entering cpu, unaccounted time is 0. This is
2513 * indistinguishable from the read occurring a few cycles earlier.
2514 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2515 * been accounted, so we're correct here as well.
2517 if (!p->on_cpu || !task_on_rq_queued(p))
2518 return p->se.sum_exec_runtime;
2521 rq = task_rq_lock(p, &flags);
2522 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2523 task_rq_unlock(rq, p, &flags);
2529 * This function gets called by the timer code, with HZ frequency.
2530 * We call it with interrupts disabled.
2532 void scheduler_tick(void)
2534 int cpu = smp_processor_id();
2535 struct rq *rq = cpu_rq(cpu);
2536 struct task_struct *curr = rq->curr;
2540 raw_spin_lock(&rq->lock);
2541 update_rq_clock(rq);
2542 curr->sched_class->task_tick(rq, curr, 0);
2543 update_cpu_load_active(rq);
2544 raw_spin_unlock(&rq->lock);
2546 perf_event_task_tick();
2549 rq->idle_balance = idle_cpu(cpu);
2550 trigger_load_balance(rq);
2552 rq_last_tick_reset(rq);
2555 #ifdef CONFIG_NO_HZ_FULL
2557 * scheduler_tick_max_deferment
2559 * Keep at least one tick per second when a single
2560 * active task is running because the scheduler doesn't
2561 * yet completely support full dynticks environment.
2563 * This makes sure that uptime, CFS vruntime, load
2564 * balancing, etc... continue to move forward, even
2565 * with a very low granularity.
2567 * Return: Maximum deferment in nanoseconds.
2569 u64 scheduler_tick_max_deferment(void)
2571 struct rq *rq = this_rq();
2572 unsigned long next, now = ACCESS_ONCE(jiffies);
2574 next = rq->last_sched_tick + HZ;
2576 if (time_before_eq(next, now))
2579 return jiffies_to_nsecs(next - now);
2583 notrace unsigned long get_parent_ip(unsigned long addr)
2585 if (in_lock_functions(addr)) {
2586 addr = CALLER_ADDR2;
2587 if (in_lock_functions(addr))
2588 addr = CALLER_ADDR3;
2593 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2594 defined(CONFIG_PREEMPT_TRACER))
2596 void preempt_count_add(int val)
2598 #ifdef CONFIG_DEBUG_PREEMPT
2602 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2605 __preempt_count_add(val);
2606 #ifdef CONFIG_DEBUG_PREEMPT
2608 * Spinlock count overflowing soon?
2610 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2613 if (preempt_count() == val) {
2614 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2615 #ifdef CONFIG_DEBUG_PREEMPT
2616 current->preempt_disable_ip = ip;
2618 trace_preempt_off(CALLER_ADDR0, ip);
2621 EXPORT_SYMBOL(preempt_count_add);
2622 NOKPROBE_SYMBOL(preempt_count_add);
2624 void preempt_count_sub(int val)
2626 #ifdef CONFIG_DEBUG_PREEMPT
2630 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2633 * Is the spinlock portion underflowing?
2635 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2636 !(preempt_count() & PREEMPT_MASK)))
2640 if (preempt_count() == val)
2641 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2642 __preempt_count_sub(val);
2644 EXPORT_SYMBOL(preempt_count_sub);
2645 NOKPROBE_SYMBOL(preempt_count_sub);
2650 * Print scheduling while atomic bug:
2652 static noinline void __schedule_bug(struct task_struct *prev)
2654 if (oops_in_progress)
2657 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2658 prev->comm, prev->pid, preempt_count());
2660 debug_show_held_locks(prev);
2662 if (irqs_disabled())
2663 print_irqtrace_events(prev);
2664 #ifdef CONFIG_DEBUG_PREEMPT
2665 if (in_atomic_preempt_off()) {
2666 pr_err("Preemption disabled at:");
2667 print_ip_sym(current->preempt_disable_ip);
2672 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2676 * Various schedule()-time debugging checks and statistics:
2678 static inline void schedule_debug(struct task_struct *prev)
2680 #ifdef CONFIG_SCHED_STACK_END_CHECK
2681 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2684 * Test if we are atomic. Since do_exit() needs to call into
2685 * schedule() atomically, we ignore that path. Otherwise whine
2686 * if we are scheduling when we should not.
2688 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2689 __schedule_bug(prev);
2692 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2694 schedstat_inc(this_rq(), sched_count);
2698 * Pick up the highest-prio task:
2700 static inline struct task_struct *
2701 pick_next_task(struct rq *rq, struct task_struct *prev)
2703 const struct sched_class *class = &fair_sched_class;
2704 struct task_struct *p;
2707 * Optimization: we know that if all tasks are in
2708 * the fair class we can call that function directly:
2710 if (likely(prev->sched_class == class &&
2711 rq->nr_running == rq->cfs.h_nr_running)) {
2712 p = fair_sched_class.pick_next_task(rq, prev);
2713 if (unlikely(p == RETRY_TASK))
2716 /* assumes fair_sched_class->next == idle_sched_class */
2718 p = idle_sched_class.pick_next_task(rq, prev);
2724 for_each_class(class) {
2725 p = class->pick_next_task(rq, prev);
2727 if (unlikely(p == RETRY_TASK))
2733 BUG(); /* the idle class will always have a runnable task */
2737 * __schedule() is the main scheduler function.
2739 * The main means of driving the scheduler and thus entering this function are:
2741 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2743 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2744 * paths. For example, see arch/x86/entry_64.S.
2746 * To drive preemption between tasks, the scheduler sets the flag in timer
2747 * interrupt handler scheduler_tick().
2749 * 3. Wakeups don't really cause entry into schedule(). They add a
2750 * task to the run-queue and that's it.
2752 * Now, if the new task added to the run-queue preempts the current
2753 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2754 * called on the nearest possible occasion:
2756 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2758 * - in syscall or exception context, at the next outmost
2759 * preempt_enable(). (this might be as soon as the wake_up()'s
2762 * - in IRQ context, return from interrupt-handler to
2763 * preemptible context
2765 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2768 * - cond_resched() call
2769 * - explicit schedule() call
2770 * - return from syscall or exception to user-space
2771 * - return from interrupt-handler to user-space
2773 static void __sched __schedule(void)
2775 struct task_struct *prev, *next;
2776 unsigned long *switch_count;
2782 cpu = smp_processor_id();
2784 rcu_note_context_switch(cpu);
2787 schedule_debug(prev);
2789 if (sched_feat(HRTICK))
2793 * Make sure that signal_pending_state()->signal_pending() below
2794 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2795 * done by the caller to avoid the race with signal_wake_up().
2797 smp_mb__before_spinlock();
2798 raw_spin_lock_irq(&rq->lock);
2800 switch_count = &prev->nivcsw;
2801 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2802 if (unlikely(signal_pending_state(prev->state, prev))) {
2803 prev->state = TASK_RUNNING;
2805 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2809 * If a worker went to sleep, notify and ask workqueue
2810 * whether it wants to wake up a task to maintain
2813 if (prev->flags & PF_WQ_WORKER) {
2814 struct task_struct *to_wakeup;
2816 to_wakeup = wq_worker_sleeping(prev, cpu);
2818 try_to_wake_up_local(to_wakeup);
2821 switch_count = &prev->nvcsw;
2824 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2825 update_rq_clock(rq);
2827 next = pick_next_task(rq, prev);
2828 clear_tsk_need_resched(prev);
2829 clear_preempt_need_resched();
2830 rq->skip_clock_update = 0;
2832 if (likely(prev != next)) {
2837 rq = context_switch(rq, prev, next); /* unlocks the rq */
2840 raw_spin_unlock_irq(&rq->lock);
2844 sched_preempt_enable_no_resched();
2849 static inline void sched_submit_work(struct task_struct *tsk)
2851 if (!tsk->state || tsk_is_pi_blocked(tsk))
2854 * If we are going to sleep and we have plugged IO queued,
2855 * make sure to submit it to avoid deadlocks.
2857 if (blk_needs_flush_plug(tsk))
2858 blk_schedule_flush_plug(tsk);
2861 asmlinkage __visible void __sched schedule(void)
2863 struct task_struct *tsk = current;
2865 sched_submit_work(tsk);
2868 EXPORT_SYMBOL(schedule);
2870 #ifdef CONFIG_CONTEXT_TRACKING
2871 asmlinkage __visible void __sched schedule_user(void)
2874 * If we come here after a random call to set_need_resched(),
2875 * or we have been woken up remotely but the IPI has not yet arrived,
2876 * we haven't yet exited the RCU idle mode. Do it here manually until
2877 * we find a better solution.
2886 * schedule_preempt_disabled - called with preemption disabled
2888 * Returns with preemption disabled. Note: preempt_count must be 1
2890 void __sched schedule_preempt_disabled(void)
2892 sched_preempt_enable_no_resched();
2897 #ifdef CONFIG_PREEMPT
2899 * this is the entry point to schedule() from in-kernel preemption
2900 * off of preempt_enable. Kernel preemptions off return from interrupt
2901 * occur there and call schedule directly.
2903 asmlinkage __visible void __sched notrace preempt_schedule(void)
2906 * If there is a non-zero preempt_count or interrupts are disabled,
2907 * we do not want to preempt the current task. Just return..
2909 if (likely(!preemptible()))
2913 __preempt_count_add(PREEMPT_ACTIVE);
2915 __preempt_count_sub(PREEMPT_ACTIVE);
2918 * Check again in case we missed a preemption opportunity
2919 * between schedule and now.
2922 } while (need_resched());
2924 NOKPROBE_SYMBOL(preempt_schedule);
2925 EXPORT_SYMBOL(preempt_schedule);
2927 #ifdef CONFIG_CONTEXT_TRACKING
2929 * preempt_schedule_context - preempt_schedule called by tracing
2931 * The tracing infrastructure uses preempt_enable_notrace to prevent
2932 * recursion and tracing preempt enabling caused by the tracing
2933 * infrastructure itself. But as tracing can happen in areas coming
2934 * from userspace or just about to enter userspace, a preempt enable
2935 * can occur before user_exit() is called. This will cause the scheduler
2936 * to be called when the system is still in usermode.
2938 * To prevent this, the preempt_enable_notrace will use this function
2939 * instead of preempt_schedule() to exit user context if needed before
2940 * calling the scheduler.
2942 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2944 enum ctx_state prev_ctx;
2946 if (likely(!preemptible()))
2950 __preempt_count_add(PREEMPT_ACTIVE);
2952 * Needs preempt disabled in case user_exit() is traced
2953 * and the tracer calls preempt_enable_notrace() causing
2954 * an infinite recursion.
2956 prev_ctx = exception_enter();
2958 exception_exit(prev_ctx);
2960 __preempt_count_sub(PREEMPT_ACTIVE);
2962 } while (need_resched());
2964 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2965 #endif /* CONFIG_CONTEXT_TRACKING */
2967 #endif /* CONFIG_PREEMPT */
2970 * this is the entry point to schedule() from kernel preemption
2971 * off of irq context.
2972 * Note, that this is called and return with irqs disabled. This will
2973 * protect us against recursive calling from irq.
2975 asmlinkage __visible void __sched preempt_schedule_irq(void)
2977 enum ctx_state prev_state;
2979 /* Catch callers which need to be fixed */
2980 BUG_ON(preempt_count() || !irqs_disabled());
2982 prev_state = exception_enter();
2985 __preempt_count_add(PREEMPT_ACTIVE);
2988 local_irq_disable();
2989 __preempt_count_sub(PREEMPT_ACTIVE);
2992 * Check again in case we missed a preemption opportunity
2993 * between schedule and now.
2996 } while (need_resched());
2998 exception_exit(prev_state);
3001 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3004 return try_to_wake_up(curr->private, mode, wake_flags);
3006 EXPORT_SYMBOL(default_wake_function);
3008 #ifdef CONFIG_RT_MUTEXES
3011 * rt_mutex_setprio - set the current priority of a task
3013 * @prio: prio value (kernel-internal form)
3015 * This function changes the 'effective' priority of a task. It does
3016 * not touch ->normal_prio like __setscheduler().
3018 * Used by the rt_mutex code to implement priority inheritance
3019 * logic. Call site only calls if the priority of the task changed.
3021 void rt_mutex_setprio(struct task_struct *p, int prio)
3023 int oldprio, queued, running, enqueue_flag = 0;
3025 const struct sched_class *prev_class;
3027 BUG_ON(prio > MAX_PRIO);
3029 rq = __task_rq_lock(p);
3032 * Idle task boosting is a nono in general. There is one
3033 * exception, when PREEMPT_RT and NOHZ is active:
3035 * The idle task calls get_next_timer_interrupt() and holds
3036 * the timer wheel base->lock on the CPU and another CPU wants
3037 * to access the timer (probably to cancel it). We can safely
3038 * ignore the boosting request, as the idle CPU runs this code
3039 * with interrupts disabled and will complete the lock
3040 * protected section without being interrupted. So there is no
3041 * real need to boost.
3043 if (unlikely(p == rq->idle)) {
3044 WARN_ON(p != rq->curr);
3045 WARN_ON(p->pi_blocked_on);
3049 trace_sched_pi_setprio(p, prio);
3051 prev_class = p->sched_class;
3052 queued = task_on_rq_queued(p);
3053 running = task_current(rq, p);
3055 dequeue_task(rq, p, 0);
3057 put_prev_task(rq, p);
3060 * Boosting condition are:
3061 * 1. -rt task is running and holds mutex A
3062 * --> -dl task blocks on mutex A
3064 * 2. -dl task is running and holds mutex A
3065 * --> -dl task blocks on mutex A and could preempt the
3068 if (dl_prio(prio)) {
3069 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3070 if (!dl_prio(p->normal_prio) ||
3071 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3072 p->dl.dl_boosted = 1;
3073 p->dl.dl_throttled = 0;
3074 enqueue_flag = ENQUEUE_REPLENISH;
3076 p->dl.dl_boosted = 0;
3077 p->sched_class = &dl_sched_class;
3078 } else if (rt_prio(prio)) {
3079 if (dl_prio(oldprio))
3080 p->dl.dl_boosted = 0;
3082 enqueue_flag = ENQUEUE_HEAD;
3083 p->sched_class = &rt_sched_class;
3085 if (dl_prio(oldprio))
3086 p->dl.dl_boosted = 0;
3087 p->sched_class = &fair_sched_class;
3093 p->sched_class->set_curr_task(rq);
3095 enqueue_task(rq, p, enqueue_flag);
3097 check_class_changed(rq, p, prev_class, oldprio);
3099 __task_rq_unlock(rq);
3103 void set_user_nice(struct task_struct *p, long nice)
3105 int old_prio, delta, queued;
3106 unsigned long flags;
3109 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3112 * We have to be careful, if called from sys_setpriority(),
3113 * the task might be in the middle of scheduling on another CPU.
3115 rq = task_rq_lock(p, &flags);
3117 * The RT priorities are set via sched_setscheduler(), but we still
3118 * allow the 'normal' nice value to be set - but as expected
3119 * it wont have any effect on scheduling until the task is
3120 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3122 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3123 p->static_prio = NICE_TO_PRIO(nice);
3126 queued = task_on_rq_queued(p);
3128 dequeue_task(rq, p, 0);
3130 p->static_prio = NICE_TO_PRIO(nice);
3133 p->prio = effective_prio(p);
3134 delta = p->prio - old_prio;
3137 enqueue_task(rq, p, 0);
3139 * If the task increased its priority or is running and
3140 * lowered its priority, then reschedule its CPU:
3142 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3146 task_rq_unlock(rq, p, &flags);
3148 EXPORT_SYMBOL(set_user_nice);
3151 * can_nice - check if a task can reduce its nice value
3155 int can_nice(const struct task_struct *p, const int nice)
3157 /* convert nice value [19,-20] to rlimit style value [1,40] */
3158 int nice_rlim = nice_to_rlimit(nice);
3160 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3161 capable(CAP_SYS_NICE));
3164 #ifdef __ARCH_WANT_SYS_NICE
3167 * sys_nice - change the priority of the current process.
3168 * @increment: priority increment
3170 * sys_setpriority is a more generic, but much slower function that
3171 * does similar things.
3173 SYSCALL_DEFINE1(nice, int, increment)
3178 * Setpriority might change our priority at the same moment.
3179 * We don't have to worry. Conceptually one call occurs first
3180 * and we have a single winner.
3182 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3183 nice = task_nice(current) + increment;
3185 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3186 if (increment < 0 && !can_nice(current, nice))
3189 retval = security_task_setnice(current, nice);
3193 set_user_nice(current, nice);
3200 * task_prio - return the priority value of a given task.
3201 * @p: the task in question.
3203 * Return: The priority value as seen by users in /proc.
3204 * RT tasks are offset by -200. Normal tasks are centered
3205 * around 0, value goes from -16 to +15.
3207 int task_prio(const struct task_struct *p)
3209 return p->prio - MAX_RT_PRIO;
3213 * idle_cpu - is a given cpu idle currently?
3214 * @cpu: the processor in question.
3216 * Return: 1 if the CPU is currently idle. 0 otherwise.
3218 int idle_cpu(int cpu)
3220 struct rq *rq = cpu_rq(cpu);
3222 if (rq->curr != rq->idle)
3229 if (!llist_empty(&rq->wake_list))
3237 * idle_task - return the idle task for a given cpu.
3238 * @cpu: the processor in question.
3240 * Return: The idle task for the cpu @cpu.
3242 struct task_struct *idle_task(int cpu)
3244 return cpu_rq(cpu)->idle;
3248 * find_process_by_pid - find a process with a matching PID value.
3249 * @pid: the pid in question.
3251 * The task of @pid, if found. %NULL otherwise.
3253 static struct task_struct *find_process_by_pid(pid_t pid)
3255 return pid ? find_task_by_vpid(pid) : current;
3259 * This function initializes the sched_dl_entity of a newly becoming
3260 * SCHED_DEADLINE task.
3262 * Only the static values are considered here, the actual runtime and the
3263 * absolute deadline will be properly calculated when the task is enqueued
3264 * for the first time with its new policy.
3267 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3269 struct sched_dl_entity *dl_se = &p->dl;
3271 init_dl_task_timer(dl_se);
3272 dl_se->dl_runtime = attr->sched_runtime;
3273 dl_se->dl_deadline = attr->sched_deadline;
3274 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3275 dl_se->flags = attr->sched_flags;
3276 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3277 dl_se->dl_throttled = 0;
3279 dl_se->dl_yielded = 0;
3283 * sched_setparam() passes in -1 for its policy, to let the functions
3284 * it calls know not to change it.
3286 #define SETPARAM_POLICY -1
3288 static void __setscheduler_params(struct task_struct *p,
3289 const struct sched_attr *attr)
3291 int policy = attr->sched_policy;
3293 if (policy == SETPARAM_POLICY)
3298 if (dl_policy(policy))
3299 __setparam_dl(p, attr);
3300 else if (fair_policy(policy))
3301 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3304 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3305 * !rt_policy. Always setting this ensures that things like
3306 * getparam()/getattr() don't report silly values for !rt tasks.
3308 p->rt_priority = attr->sched_priority;
3309 p->normal_prio = normal_prio(p);
3313 /* Actually do priority change: must hold pi & rq lock. */
3314 static void __setscheduler(struct rq *rq, struct task_struct *p,
3315 const struct sched_attr *attr)
3317 __setscheduler_params(p, attr);
3320 * If we get here, there was no pi waiters boosting the
3321 * task. It is safe to use the normal prio.
3323 p->prio = normal_prio(p);
3325 if (dl_prio(p->prio))
3326 p->sched_class = &dl_sched_class;
3327 else if (rt_prio(p->prio))
3328 p->sched_class = &rt_sched_class;
3330 p->sched_class = &fair_sched_class;
3334 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3336 struct sched_dl_entity *dl_se = &p->dl;
3338 attr->sched_priority = p->rt_priority;
3339 attr->sched_runtime = dl_se->dl_runtime;
3340 attr->sched_deadline = dl_se->dl_deadline;
3341 attr->sched_period = dl_se->dl_period;
3342 attr->sched_flags = dl_se->flags;
3346 * This function validates the new parameters of a -deadline task.
3347 * We ask for the deadline not being zero, and greater or equal
3348 * than the runtime, as well as the period of being zero or
3349 * greater than deadline. Furthermore, we have to be sure that
3350 * user parameters are above the internal resolution of 1us (we
3351 * check sched_runtime only since it is always the smaller one) and
3352 * below 2^63 ns (we have to check both sched_deadline and
3353 * sched_period, as the latter can be zero).
3356 __checkparam_dl(const struct sched_attr *attr)
3359 if (attr->sched_deadline == 0)
3363 * Since we truncate DL_SCALE bits, make sure we're at least
3366 if (attr->sched_runtime < (1ULL << DL_SCALE))
3370 * Since we use the MSB for wrap-around and sign issues, make
3371 * sure it's not set (mind that period can be equal to zero).
3373 if (attr->sched_deadline & (1ULL << 63) ||
3374 attr->sched_period & (1ULL << 63))
3377 /* runtime <= deadline <= period (if period != 0) */
3378 if ((attr->sched_period != 0 &&
3379 attr->sched_period < attr->sched_deadline) ||
3380 attr->sched_deadline < attr->sched_runtime)
3387 * check the target process has a UID that matches the current process's
3389 static bool check_same_owner(struct task_struct *p)
3391 const struct cred *cred = current_cred(), *pcred;
3395 pcred = __task_cred(p);
3396 match = (uid_eq(cred->euid, pcred->euid) ||
3397 uid_eq(cred->euid, pcred->uid));
3402 static int __sched_setscheduler(struct task_struct *p,
3403 const struct sched_attr *attr,
3406 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3407 MAX_RT_PRIO - 1 - attr->sched_priority;
3408 int retval, oldprio, oldpolicy = -1, queued, running;
3409 int policy = attr->sched_policy;
3410 unsigned long flags;
3411 const struct sched_class *prev_class;
3415 /* may grab non-irq protected spin_locks */
3416 BUG_ON(in_interrupt());
3418 /* double check policy once rq lock held */
3420 reset_on_fork = p->sched_reset_on_fork;
3421 policy = oldpolicy = p->policy;
3423 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3425 if (policy != SCHED_DEADLINE &&
3426 policy != SCHED_FIFO && policy != SCHED_RR &&
3427 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3428 policy != SCHED_IDLE)
3432 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3436 * Valid priorities for SCHED_FIFO and SCHED_RR are
3437 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3438 * SCHED_BATCH and SCHED_IDLE is 0.
3440 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3441 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3443 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3444 (rt_policy(policy) != (attr->sched_priority != 0)))
3448 * Allow unprivileged RT tasks to decrease priority:
3450 if (user && !capable(CAP_SYS_NICE)) {
3451 if (fair_policy(policy)) {
3452 if (attr->sched_nice < task_nice(p) &&
3453 !can_nice(p, attr->sched_nice))
3457 if (rt_policy(policy)) {
3458 unsigned long rlim_rtprio =
3459 task_rlimit(p, RLIMIT_RTPRIO);
3461 /* can't set/change the rt policy */
3462 if (policy != p->policy && !rlim_rtprio)
3465 /* can't increase priority */
3466 if (attr->sched_priority > p->rt_priority &&
3467 attr->sched_priority > rlim_rtprio)
3472 * Can't set/change SCHED_DEADLINE policy at all for now
3473 * (safest behavior); in the future we would like to allow
3474 * unprivileged DL tasks to increase their relative deadline
3475 * or reduce their runtime (both ways reducing utilization)
3477 if (dl_policy(policy))
3481 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3482 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3484 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3485 if (!can_nice(p, task_nice(p)))
3489 /* can't change other user's priorities */
3490 if (!check_same_owner(p))
3493 /* Normal users shall not reset the sched_reset_on_fork flag */
3494 if (p->sched_reset_on_fork && !reset_on_fork)
3499 retval = security_task_setscheduler(p);
3505 * make sure no PI-waiters arrive (or leave) while we are
3506 * changing the priority of the task:
3508 * To be able to change p->policy safely, the appropriate
3509 * runqueue lock must be held.
3511 rq = task_rq_lock(p, &flags);
3514 * Changing the policy of the stop threads its a very bad idea
3516 if (p == rq->stop) {
3517 task_rq_unlock(rq, p, &flags);
3522 * If not changing anything there's no need to proceed further,
3523 * but store a possible modification of reset_on_fork.
3525 if (unlikely(policy == p->policy)) {
3526 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3528 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3530 if (dl_policy(policy))
3533 p->sched_reset_on_fork = reset_on_fork;
3534 task_rq_unlock(rq, p, &flags);
3540 #ifdef CONFIG_RT_GROUP_SCHED
3542 * Do not allow realtime tasks into groups that have no runtime
3545 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3546 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3547 !task_group_is_autogroup(task_group(p))) {
3548 task_rq_unlock(rq, p, &flags);
3553 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3554 cpumask_t *span = rq->rd->span;
3557 * Don't allow tasks with an affinity mask smaller than
3558 * the entire root_domain to become SCHED_DEADLINE. We
3559 * will also fail if there's no bandwidth available.
3561 if (!cpumask_subset(span, &p->cpus_allowed) ||
3562 rq->rd->dl_bw.bw == 0) {
3563 task_rq_unlock(rq, p, &flags);
3570 /* recheck policy now with rq lock held */
3571 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3572 policy = oldpolicy = -1;
3573 task_rq_unlock(rq, p, &flags);
3578 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3579 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3582 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3583 task_rq_unlock(rq, p, &flags);
3587 p->sched_reset_on_fork = reset_on_fork;
3591 * Special case for priority boosted tasks.
3593 * If the new priority is lower or equal (user space view)
3594 * than the current (boosted) priority, we just store the new
3595 * normal parameters and do not touch the scheduler class and
3596 * the runqueue. This will be done when the task deboost
3599 if (rt_mutex_check_prio(p, newprio)) {
3600 __setscheduler_params(p, attr);
3601 task_rq_unlock(rq, p, &flags);
3605 queued = task_on_rq_queued(p);
3606 running = task_current(rq, p);
3608 dequeue_task(rq, p, 0);
3610 put_prev_task(rq, p);
3612 prev_class = p->sched_class;
3613 __setscheduler(rq, p, attr);
3616 p->sched_class->set_curr_task(rq);
3619 * We enqueue to tail when the priority of a task is
3620 * increased (user space view).
3622 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3625 check_class_changed(rq, p, prev_class, oldprio);
3626 task_rq_unlock(rq, p, &flags);
3628 rt_mutex_adjust_pi(p);
3633 static int _sched_setscheduler(struct task_struct *p, int policy,
3634 const struct sched_param *param, bool check)
3636 struct sched_attr attr = {
3637 .sched_policy = policy,
3638 .sched_priority = param->sched_priority,
3639 .sched_nice = PRIO_TO_NICE(p->static_prio),
3642 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3643 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3644 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3645 policy &= ~SCHED_RESET_ON_FORK;
3646 attr.sched_policy = policy;
3649 return __sched_setscheduler(p, &attr, check);
3652 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3653 * @p: the task in question.
3654 * @policy: new policy.
3655 * @param: structure containing the new RT priority.
3657 * Return: 0 on success. An error code otherwise.
3659 * NOTE that the task may be already dead.
3661 int sched_setscheduler(struct task_struct *p, int policy,
3662 const struct sched_param *param)
3664 return _sched_setscheduler(p, policy, param, true);
3666 EXPORT_SYMBOL_GPL(sched_setscheduler);
3668 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3670 return __sched_setscheduler(p, attr, true);
3672 EXPORT_SYMBOL_GPL(sched_setattr);
3675 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3676 * @p: the task in question.
3677 * @policy: new policy.
3678 * @param: structure containing the new RT priority.
3680 * Just like sched_setscheduler, only don't bother checking if the
3681 * current context has permission. For example, this is needed in
3682 * stop_machine(): we create temporary high priority worker threads,
3683 * but our caller might not have that capability.
3685 * Return: 0 on success. An error code otherwise.
3687 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3688 const struct sched_param *param)
3690 return _sched_setscheduler(p, policy, param, false);
3694 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3696 struct sched_param lparam;
3697 struct task_struct *p;
3700 if (!param || pid < 0)
3702 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3707 p = find_process_by_pid(pid);
3709 retval = sched_setscheduler(p, policy, &lparam);
3716 * Mimics kernel/events/core.c perf_copy_attr().
3718 static int sched_copy_attr(struct sched_attr __user *uattr,
3719 struct sched_attr *attr)
3724 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3728 * zero the full structure, so that a short copy will be nice.
3730 memset(attr, 0, sizeof(*attr));
3732 ret = get_user(size, &uattr->size);
3736 if (size > PAGE_SIZE) /* silly large */
3739 if (!size) /* abi compat */
3740 size = SCHED_ATTR_SIZE_VER0;
3742 if (size < SCHED_ATTR_SIZE_VER0)
3746 * If we're handed a bigger struct than we know of,
3747 * ensure all the unknown bits are 0 - i.e. new
3748 * user-space does not rely on any kernel feature
3749 * extensions we dont know about yet.
3751 if (size > sizeof(*attr)) {
3752 unsigned char __user *addr;
3753 unsigned char __user *end;
3756 addr = (void __user *)uattr + sizeof(*attr);
3757 end = (void __user *)uattr + size;
3759 for (; addr < end; addr++) {
3760 ret = get_user(val, addr);
3766 size = sizeof(*attr);
3769 ret = copy_from_user(attr, uattr, size);
3774 * XXX: do we want to be lenient like existing syscalls; or do we want
3775 * to be strict and return an error on out-of-bounds values?
3777 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3782 put_user(sizeof(*attr), &uattr->size);
3787 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3788 * @pid: the pid in question.
3789 * @policy: new policy.
3790 * @param: structure containing the new RT priority.
3792 * Return: 0 on success. An error code otherwise.
3794 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3795 struct sched_param __user *, param)
3797 /* negative values for policy are not valid */
3801 return do_sched_setscheduler(pid, policy, param);
3805 * sys_sched_setparam - set/change the RT priority of a thread
3806 * @pid: the pid in question.
3807 * @param: structure containing the new RT priority.
3809 * Return: 0 on success. An error code otherwise.
3811 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3813 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3817 * sys_sched_setattr - same as above, but with extended sched_attr
3818 * @pid: the pid in question.
3819 * @uattr: structure containing the extended parameters.
3820 * @flags: for future extension.
3822 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3823 unsigned int, flags)
3825 struct sched_attr attr;
3826 struct task_struct *p;
3829 if (!uattr || pid < 0 || flags)
3832 retval = sched_copy_attr(uattr, &attr);
3836 if ((int)attr.sched_policy < 0)
3841 p = find_process_by_pid(pid);
3843 retval = sched_setattr(p, &attr);
3850 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3851 * @pid: the pid in question.
3853 * Return: On success, the policy of the thread. Otherwise, a negative error
3856 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3858 struct task_struct *p;
3866 p = find_process_by_pid(pid);
3868 retval = security_task_getscheduler(p);
3871 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3878 * sys_sched_getparam - get the RT priority of a thread
3879 * @pid: the pid in question.
3880 * @param: structure containing the RT priority.
3882 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3885 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3887 struct sched_param lp = { .sched_priority = 0 };
3888 struct task_struct *p;
3891 if (!param || pid < 0)
3895 p = find_process_by_pid(pid);
3900 retval = security_task_getscheduler(p);
3904 if (task_has_rt_policy(p))
3905 lp.sched_priority = p->rt_priority;
3909 * This one might sleep, we cannot do it with a spinlock held ...
3911 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3920 static int sched_read_attr(struct sched_attr __user *uattr,
3921 struct sched_attr *attr,
3926 if (!access_ok(VERIFY_WRITE, uattr, usize))
3930 * If we're handed a smaller struct than we know of,
3931 * ensure all the unknown bits are 0 - i.e. old
3932 * user-space does not get uncomplete information.
3934 if (usize < sizeof(*attr)) {
3935 unsigned char *addr;
3938 addr = (void *)attr + usize;
3939 end = (void *)attr + sizeof(*attr);
3941 for (; addr < end; addr++) {
3949 ret = copy_to_user(uattr, attr, attr->size);
3957 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3958 * @pid: the pid in question.
3959 * @uattr: structure containing the extended parameters.
3960 * @size: sizeof(attr) for fwd/bwd comp.
3961 * @flags: for future extension.
3963 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3964 unsigned int, size, unsigned int, flags)
3966 struct sched_attr attr = {
3967 .size = sizeof(struct sched_attr),
3969 struct task_struct *p;
3972 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3973 size < SCHED_ATTR_SIZE_VER0 || flags)
3977 p = find_process_by_pid(pid);
3982 retval = security_task_getscheduler(p);
3986 attr.sched_policy = p->policy;
3987 if (p->sched_reset_on_fork)
3988 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3989 if (task_has_dl_policy(p))
3990 __getparam_dl(p, &attr);
3991 else if (task_has_rt_policy(p))
3992 attr.sched_priority = p->rt_priority;
3994 attr.sched_nice = task_nice(p);
3998 retval = sched_read_attr(uattr, &attr, size);
4006 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4008 cpumask_var_t cpus_allowed, new_mask;
4009 struct task_struct *p;
4014 p = find_process_by_pid(pid);
4020 /* Prevent p going away */
4024 if (p->flags & PF_NO_SETAFFINITY) {
4028 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4032 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4034 goto out_free_cpus_allowed;
4037 if (!check_same_owner(p)) {
4039 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4041 goto out_free_new_mask;
4046 retval = security_task_setscheduler(p);
4048 goto out_free_new_mask;
4051 cpuset_cpus_allowed(p, cpus_allowed);
4052 cpumask_and(new_mask, in_mask, cpus_allowed);
4055 * Since bandwidth control happens on root_domain basis,
4056 * if admission test is enabled, we only admit -deadline
4057 * tasks allowed to run on all the CPUs in the task's
4061 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4063 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4066 goto out_free_new_mask;
4072 retval = set_cpus_allowed_ptr(p, new_mask);
4075 cpuset_cpus_allowed(p, cpus_allowed);
4076 if (!cpumask_subset(new_mask, cpus_allowed)) {
4078 * We must have raced with a concurrent cpuset
4079 * update. Just reset the cpus_allowed to the
4080 * cpuset's cpus_allowed
4082 cpumask_copy(new_mask, cpus_allowed);
4087 free_cpumask_var(new_mask);
4088 out_free_cpus_allowed:
4089 free_cpumask_var(cpus_allowed);
4095 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4096 struct cpumask *new_mask)
4098 if (len < cpumask_size())
4099 cpumask_clear(new_mask);
4100 else if (len > cpumask_size())
4101 len = cpumask_size();
4103 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4107 * sys_sched_setaffinity - set the cpu affinity of a process
4108 * @pid: pid of the process
4109 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4110 * @user_mask_ptr: user-space pointer to the new cpu mask
4112 * Return: 0 on success. An error code otherwise.
4114 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4115 unsigned long __user *, user_mask_ptr)
4117 cpumask_var_t new_mask;
4120 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4123 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4125 retval = sched_setaffinity(pid, new_mask);
4126 free_cpumask_var(new_mask);
4130 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4132 struct task_struct *p;
4133 unsigned long flags;
4139 p = find_process_by_pid(pid);
4143 retval = security_task_getscheduler(p);
4147 raw_spin_lock_irqsave(&p->pi_lock, flags);
4148 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4149 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4158 * sys_sched_getaffinity - get the cpu affinity of a process
4159 * @pid: pid of the process
4160 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4161 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4163 * Return: 0 on success. An error code otherwise.
4165 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4166 unsigned long __user *, user_mask_ptr)
4171 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4173 if (len & (sizeof(unsigned long)-1))
4176 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4179 ret = sched_getaffinity(pid, mask);
4181 size_t retlen = min_t(size_t, len, cpumask_size());
4183 if (copy_to_user(user_mask_ptr, mask, retlen))
4188 free_cpumask_var(mask);
4194 * sys_sched_yield - yield the current processor to other threads.
4196 * This function yields the current CPU to other tasks. If there are no
4197 * other threads running on this CPU then this function will return.
4201 SYSCALL_DEFINE0(sched_yield)
4203 struct rq *rq = this_rq_lock();
4205 schedstat_inc(rq, yld_count);
4206 current->sched_class->yield_task(rq);
4209 * Since we are going to call schedule() anyway, there's
4210 * no need to preempt or enable interrupts:
4212 __release(rq->lock);
4213 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4214 do_raw_spin_unlock(&rq->lock);
4215 sched_preempt_enable_no_resched();
4222 static void __cond_resched(void)
4224 __preempt_count_add(PREEMPT_ACTIVE);
4226 __preempt_count_sub(PREEMPT_ACTIVE);
4229 int __sched _cond_resched(void)
4231 if (should_resched()) {
4237 EXPORT_SYMBOL(_cond_resched);
4240 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4241 * call schedule, and on return reacquire the lock.
4243 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4244 * operations here to prevent schedule() from being called twice (once via
4245 * spin_unlock(), once by hand).
4247 int __cond_resched_lock(spinlock_t *lock)
4249 int resched = should_resched();
4252 lockdep_assert_held(lock);
4254 if (spin_needbreak(lock) || resched) {
4265 EXPORT_SYMBOL(__cond_resched_lock);
4267 int __sched __cond_resched_softirq(void)
4269 BUG_ON(!in_softirq());
4271 if (should_resched()) {
4279 EXPORT_SYMBOL(__cond_resched_softirq);
4282 * yield - yield the current processor to other threads.
4284 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4286 * The scheduler is at all times free to pick the calling task as the most
4287 * eligible task to run, if removing the yield() call from your code breaks
4288 * it, its already broken.
4290 * Typical broken usage is:
4295 * where one assumes that yield() will let 'the other' process run that will
4296 * make event true. If the current task is a SCHED_FIFO task that will never
4297 * happen. Never use yield() as a progress guarantee!!
4299 * If you want to use yield() to wait for something, use wait_event().
4300 * If you want to use yield() to be 'nice' for others, use cond_resched().
4301 * If you still want to use yield(), do not!
4303 void __sched yield(void)
4305 set_current_state(TASK_RUNNING);
4308 EXPORT_SYMBOL(yield);
4311 * yield_to - yield the current processor to another thread in
4312 * your thread group, or accelerate that thread toward the
4313 * processor it's on.
4315 * @preempt: whether task preemption is allowed or not
4317 * It's the caller's job to ensure that the target task struct
4318 * can't go away on us before we can do any checks.
4321 * true (>0) if we indeed boosted the target task.
4322 * false (0) if we failed to boost the target.
4323 * -ESRCH if there's no task to yield to.
4325 int __sched yield_to(struct task_struct *p, bool preempt)
4327 struct task_struct *curr = current;
4328 struct rq *rq, *p_rq;
4329 unsigned long flags;
4332 local_irq_save(flags);
4338 * If we're the only runnable task on the rq and target rq also
4339 * has only one task, there's absolutely no point in yielding.
4341 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4346 double_rq_lock(rq, p_rq);
4347 if (task_rq(p) != p_rq) {
4348 double_rq_unlock(rq, p_rq);
4352 if (!curr->sched_class->yield_to_task)
4355 if (curr->sched_class != p->sched_class)
4358 if (task_running(p_rq, p) || p->state)
4361 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4363 schedstat_inc(rq, yld_count);
4365 * Make p's CPU reschedule; pick_next_entity takes care of
4368 if (preempt && rq != p_rq)
4373 double_rq_unlock(rq, p_rq);
4375 local_irq_restore(flags);
4382 EXPORT_SYMBOL_GPL(yield_to);
4385 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4386 * that process accounting knows that this is a task in IO wait state.
4388 void __sched io_schedule(void)
4390 struct rq *rq = raw_rq();
4392 delayacct_blkio_start();
4393 atomic_inc(&rq->nr_iowait);
4394 blk_flush_plug(current);
4395 current->in_iowait = 1;
4397 current->in_iowait = 0;
4398 atomic_dec(&rq->nr_iowait);
4399 delayacct_blkio_end();
4401 EXPORT_SYMBOL(io_schedule);
4403 long __sched io_schedule_timeout(long timeout)
4405 struct rq *rq = raw_rq();
4408 delayacct_blkio_start();
4409 atomic_inc(&rq->nr_iowait);
4410 blk_flush_plug(current);
4411 current->in_iowait = 1;
4412 ret = schedule_timeout(timeout);
4413 current->in_iowait = 0;
4414 atomic_dec(&rq->nr_iowait);
4415 delayacct_blkio_end();
4420 * sys_sched_get_priority_max - return maximum RT priority.
4421 * @policy: scheduling class.
4423 * Return: On success, this syscall returns the maximum
4424 * rt_priority that can be used by a given scheduling class.
4425 * On failure, a negative error code is returned.
4427 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4434 ret = MAX_USER_RT_PRIO-1;
4436 case SCHED_DEADLINE:
4447 * sys_sched_get_priority_min - return minimum RT priority.
4448 * @policy: scheduling class.
4450 * Return: On success, this syscall returns the minimum
4451 * rt_priority that can be used by a given scheduling class.
4452 * On failure, a negative error code is returned.
4454 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4463 case SCHED_DEADLINE:
4473 * sys_sched_rr_get_interval - return the default timeslice of a process.
4474 * @pid: pid of the process.
4475 * @interval: userspace pointer to the timeslice value.
4477 * this syscall writes the default timeslice value of a given process
4478 * into the user-space timespec buffer. A value of '0' means infinity.
4480 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4483 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4484 struct timespec __user *, interval)
4486 struct task_struct *p;
4487 unsigned int time_slice;
4488 unsigned long flags;
4498 p = find_process_by_pid(pid);
4502 retval = security_task_getscheduler(p);
4506 rq = task_rq_lock(p, &flags);
4508 if (p->sched_class->get_rr_interval)
4509 time_slice = p->sched_class->get_rr_interval(rq, p);
4510 task_rq_unlock(rq, p, &flags);
4513 jiffies_to_timespec(time_slice, &t);
4514 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4522 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4524 void sched_show_task(struct task_struct *p)
4526 unsigned long free = 0;
4530 state = p->state ? __ffs(p->state) + 1 : 0;
4531 printk(KERN_INFO "%-15.15s %c", p->comm,
4532 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4533 #if BITS_PER_LONG == 32
4534 if (state == TASK_RUNNING)
4535 printk(KERN_CONT " running ");
4537 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4539 if (state == TASK_RUNNING)
4540 printk(KERN_CONT " running task ");
4542 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4544 #ifdef CONFIG_DEBUG_STACK_USAGE
4545 free = stack_not_used(p);
4548 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4550 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4551 task_pid_nr(p), ppid,
4552 (unsigned long)task_thread_info(p)->flags);
4554 print_worker_info(KERN_INFO, p);
4555 show_stack(p, NULL);
4558 void show_state_filter(unsigned long state_filter)
4560 struct task_struct *g, *p;
4562 #if BITS_PER_LONG == 32
4564 " task PC stack pid father\n");
4567 " task PC stack pid father\n");
4570 for_each_process_thread(g, p) {
4572 * reset the NMI-timeout, listing all files on a slow
4573 * console might take a lot of time:
4575 touch_nmi_watchdog();
4576 if (!state_filter || (p->state & state_filter))
4580 touch_all_softlockup_watchdogs();
4582 #ifdef CONFIG_SCHED_DEBUG
4583 sysrq_sched_debug_show();
4587 * Only show locks if all tasks are dumped:
4590 debug_show_all_locks();
4593 void init_idle_bootup_task(struct task_struct *idle)
4595 idle->sched_class = &idle_sched_class;
4599 * init_idle - set up an idle thread for a given CPU
4600 * @idle: task in question
4601 * @cpu: cpu the idle task belongs to
4603 * NOTE: this function does not set the idle thread's NEED_RESCHED
4604 * flag, to make booting more robust.
4606 void init_idle(struct task_struct *idle, int cpu)
4608 struct rq *rq = cpu_rq(cpu);
4609 unsigned long flags;
4611 raw_spin_lock_irqsave(&rq->lock, flags);
4613 __sched_fork(0, idle);
4614 idle->state = TASK_RUNNING;
4615 idle->se.exec_start = sched_clock();
4617 do_set_cpus_allowed(idle, cpumask_of(cpu));
4619 * We're having a chicken and egg problem, even though we are
4620 * holding rq->lock, the cpu isn't yet set to this cpu so the
4621 * lockdep check in task_group() will fail.
4623 * Similar case to sched_fork(). / Alternatively we could
4624 * use task_rq_lock() here and obtain the other rq->lock.
4629 __set_task_cpu(idle, cpu);
4632 rq->curr = rq->idle = idle;
4633 idle->on_rq = TASK_ON_RQ_QUEUED;
4634 #if defined(CONFIG_SMP)
4637 raw_spin_unlock_irqrestore(&rq->lock, flags);
4639 /* Set the preempt count _outside_ the spinlocks! */
4640 init_idle_preempt_count(idle, cpu);
4643 * The idle tasks have their own, simple scheduling class:
4645 idle->sched_class = &idle_sched_class;
4646 ftrace_graph_init_idle_task(idle, cpu);
4647 vtime_init_idle(idle, cpu);
4648 #if defined(CONFIG_SMP)
4649 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4653 int task_can_attach(struct task_struct *p,
4654 const struct cpumask *cs_cpus_allowed)
4659 * Kthreads which disallow setaffinity shouldn't be moved
4660 * to a new cpuset; we don't want to change their cpu
4661 * affinity and isolating such threads by their set of
4662 * allowed nodes is unnecessary. Thus, cpusets are not
4663 * applicable for such threads. This prevents checking for
4664 * success of set_cpus_allowed_ptr() on all attached tasks
4665 * before cpus_allowed may be changed.
4667 if (p->flags & PF_NO_SETAFFINITY) {
4673 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4675 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4677 struct dl_bw *dl_b = dl_bw_of(dest_cpu);
4680 unsigned long flags;
4682 raw_spin_lock_irqsave(&dl_b->lock, flags);
4683 cpus = dl_bw_cpus(dest_cpu);
4684 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4689 * We reserve space for this task in the destination
4690 * root_domain, as we can't fail after this point.
4691 * We will free resources in the source root_domain
4692 * later on (see set_cpus_allowed_dl()).
4694 __dl_add(dl_b, p->dl.dl_bw);
4696 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4706 * move_queued_task - move a queued task to new rq.
4708 * Returns (locked) new rq. Old rq's lock is released.
4710 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4712 struct rq *rq = task_rq(p);
4714 lockdep_assert_held(&rq->lock);
4716 dequeue_task(rq, p, 0);
4717 p->on_rq = TASK_ON_RQ_MIGRATING;
4718 set_task_cpu(p, new_cpu);
4719 raw_spin_unlock(&rq->lock);
4721 rq = cpu_rq(new_cpu);
4723 raw_spin_lock(&rq->lock);
4724 BUG_ON(task_cpu(p) != new_cpu);
4725 p->on_rq = TASK_ON_RQ_QUEUED;
4726 enqueue_task(rq, p, 0);
4727 check_preempt_curr(rq, p, 0);
4732 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4734 if (p->sched_class && p->sched_class->set_cpus_allowed)
4735 p->sched_class->set_cpus_allowed(p, new_mask);
4737 cpumask_copy(&p->cpus_allowed, new_mask);
4738 p->nr_cpus_allowed = cpumask_weight(new_mask);
4742 * This is how migration works:
4744 * 1) we invoke migration_cpu_stop() on the target CPU using
4746 * 2) stopper starts to run (implicitly forcing the migrated thread
4748 * 3) it checks whether the migrated task is still in the wrong runqueue.
4749 * 4) if it's in the wrong runqueue then the migration thread removes
4750 * it and puts it into the right queue.
4751 * 5) stopper completes and stop_one_cpu() returns and the migration
4756 * Change a given task's CPU affinity. Migrate the thread to a
4757 * proper CPU and schedule it away if the CPU it's executing on
4758 * is removed from the allowed bitmask.
4760 * NOTE: the caller must have a valid reference to the task, the
4761 * task must not exit() & deallocate itself prematurely. The
4762 * call is not atomic; no spinlocks may be held.
4764 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4766 unsigned long flags;
4768 unsigned int dest_cpu;
4771 rq = task_rq_lock(p, &flags);
4773 if (cpumask_equal(&p->cpus_allowed, new_mask))
4776 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4781 do_set_cpus_allowed(p, new_mask);
4783 /* Can the task run on the task's current CPU? If so, we're done */
4784 if (cpumask_test_cpu(task_cpu(p), new_mask))
4787 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4788 if (task_running(rq, p) || p->state == TASK_WAKING) {
4789 struct migration_arg arg = { p, dest_cpu };
4790 /* Need help from migration thread: drop lock and wait. */
4791 task_rq_unlock(rq, p, &flags);
4792 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4793 tlb_migrate_finish(p->mm);
4795 } else if (task_on_rq_queued(p))
4796 rq = move_queued_task(p, dest_cpu);
4798 task_rq_unlock(rq, p, &flags);
4802 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4805 * Move (not current) task off this cpu, onto dest cpu. We're doing
4806 * this because either it can't run here any more (set_cpus_allowed()
4807 * away from this CPU, or CPU going down), or because we're
4808 * attempting to rebalance this task on exec (sched_exec).
4810 * So we race with normal scheduler movements, but that's OK, as long
4811 * as the task is no longer on this CPU.
4813 * Returns non-zero if task was successfully migrated.
4815 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4820 if (unlikely(!cpu_active(dest_cpu)))
4823 rq = cpu_rq(src_cpu);
4825 raw_spin_lock(&p->pi_lock);
4826 raw_spin_lock(&rq->lock);
4827 /* Already moved. */
4828 if (task_cpu(p) != src_cpu)
4831 /* Affinity changed (again). */
4832 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4836 * If we're not on a rq, the next wake-up will ensure we're
4839 if (task_on_rq_queued(p))
4840 rq = move_queued_task(p, dest_cpu);
4844 raw_spin_unlock(&rq->lock);
4845 raw_spin_unlock(&p->pi_lock);
4849 #ifdef CONFIG_NUMA_BALANCING
4850 /* Migrate current task p to target_cpu */
4851 int migrate_task_to(struct task_struct *p, int target_cpu)
4853 struct migration_arg arg = { p, target_cpu };
4854 int curr_cpu = task_cpu(p);
4856 if (curr_cpu == target_cpu)
4859 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4862 /* TODO: This is not properly updating schedstats */
4864 trace_sched_move_numa(p, curr_cpu, target_cpu);
4865 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4869 * Requeue a task on a given node and accurately track the number of NUMA
4870 * tasks on the runqueues
4872 void sched_setnuma(struct task_struct *p, int nid)
4875 unsigned long flags;
4876 bool queued, running;
4878 rq = task_rq_lock(p, &flags);
4879 queued = task_on_rq_queued(p);
4880 running = task_current(rq, p);
4883 dequeue_task(rq, p, 0);
4885 put_prev_task(rq, p);
4887 p->numa_preferred_nid = nid;
4890 p->sched_class->set_curr_task(rq);
4892 enqueue_task(rq, p, 0);
4893 task_rq_unlock(rq, p, &flags);
4898 * migration_cpu_stop - this will be executed by a highprio stopper thread
4899 * and performs thread migration by bumping thread off CPU then
4900 * 'pushing' onto another runqueue.
4902 static int migration_cpu_stop(void *data)
4904 struct migration_arg *arg = data;
4907 * The original target cpu might have gone down and we might
4908 * be on another cpu but it doesn't matter.
4910 local_irq_disable();
4912 * We need to explicitly wake pending tasks before running
4913 * __migrate_task() such that we will not miss enforcing cpus_allowed
4914 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4916 sched_ttwu_pending();
4917 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4922 #ifdef CONFIG_HOTPLUG_CPU
4925 * Ensures that the idle task is using init_mm right before its cpu goes
4928 void idle_task_exit(void)
4930 struct mm_struct *mm = current->active_mm;
4932 BUG_ON(cpu_online(smp_processor_id()));
4934 if (mm != &init_mm) {
4935 switch_mm(mm, &init_mm, current);
4936 finish_arch_post_lock_switch();
4942 * Since this CPU is going 'away' for a while, fold any nr_active delta
4943 * we might have. Assumes we're called after migrate_tasks() so that the
4944 * nr_active count is stable.
4946 * Also see the comment "Global load-average calculations".
4948 static void calc_load_migrate(struct rq *rq)
4950 long delta = calc_load_fold_active(rq);
4952 atomic_long_add(delta, &calc_load_tasks);
4955 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4959 static const struct sched_class fake_sched_class = {
4960 .put_prev_task = put_prev_task_fake,
4963 static struct task_struct fake_task = {
4965 * Avoid pull_{rt,dl}_task()
4967 .prio = MAX_PRIO + 1,
4968 .sched_class = &fake_sched_class,
4972 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4973 * try_to_wake_up()->select_task_rq().
4975 * Called with rq->lock held even though we'er in stop_machine() and
4976 * there's no concurrency possible, we hold the required locks anyway
4977 * because of lock validation efforts.
4979 static void migrate_tasks(unsigned int dead_cpu)
4981 struct rq *rq = cpu_rq(dead_cpu);
4982 struct task_struct *next, *stop = rq->stop;
4986 * Fudge the rq selection such that the below task selection loop
4987 * doesn't get stuck on the currently eligible stop task.
4989 * We're currently inside stop_machine() and the rq is either stuck
4990 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4991 * either way we should never end up calling schedule() until we're
4997 * put_prev_task() and pick_next_task() sched
4998 * class method both need to have an up-to-date
4999 * value of rq->clock[_task]
5001 update_rq_clock(rq);
5005 * There's this thread running, bail when that's the only
5008 if (rq->nr_running == 1)
5011 next = pick_next_task(rq, &fake_task);
5013 next->sched_class->put_prev_task(rq, next);
5015 /* Find suitable destination for @next, with force if needed. */
5016 dest_cpu = select_fallback_rq(dead_cpu, next);
5017 raw_spin_unlock(&rq->lock);
5019 __migrate_task(next, dead_cpu, dest_cpu);
5021 raw_spin_lock(&rq->lock);
5027 #endif /* CONFIG_HOTPLUG_CPU */
5029 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5031 static struct ctl_table sd_ctl_dir[] = {
5033 .procname = "sched_domain",
5039 static struct ctl_table sd_ctl_root[] = {
5041 .procname = "kernel",
5043 .child = sd_ctl_dir,
5048 static struct ctl_table *sd_alloc_ctl_entry(int n)
5050 struct ctl_table *entry =
5051 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5056 static void sd_free_ctl_entry(struct ctl_table **tablep)
5058 struct ctl_table *entry;
5061 * In the intermediate directories, both the child directory and
5062 * procname are dynamically allocated and could fail but the mode
5063 * will always be set. In the lowest directory the names are
5064 * static strings and all have proc handlers.
5066 for (entry = *tablep; entry->mode; entry++) {
5068 sd_free_ctl_entry(&entry->child);
5069 if (entry->proc_handler == NULL)
5070 kfree(entry->procname);
5077 static int min_load_idx = 0;
5078 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5081 set_table_entry(struct ctl_table *entry,
5082 const char *procname, void *data, int maxlen,
5083 umode_t mode, proc_handler *proc_handler,
5086 entry->procname = procname;
5088 entry->maxlen = maxlen;
5090 entry->proc_handler = proc_handler;
5093 entry->extra1 = &min_load_idx;
5094 entry->extra2 = &max_load_idx;
5098 static struct ctl_table *
5099 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5101 struct ctl_table *table = sd_alloc_ctl_entry(14);
5106 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5107 sizeof(long), 0644, proc_doulongvec_minmax, false);
5108 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5109 sizeof(long), 0644, proc_doulongvec_minmax, false);
5110 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5111 sizeof(int), 0644, proc_dointvec_minmax, true);
5112 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5113 sizeof(int), 0644, proc_dointvec_minmax, true);
5114 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5115 sizeof(int), 0644, proc_dointvec_minmax, true);
5116 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5117 sizeof(int), 0644, proc_dointvec_minmax, true);
5118 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5119 sizeof(int), 0644, proc_dointvec_minmax, true);
5120 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5121 sizeof(int), 0644, proc_dointvec_minmax, false);
5122 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5123 sizeof(int), 0644, proc_dointvec_minmax, false);
5124 set_table_entry(&table[9], "cache_nice_tries",
5125 &sd->cache_nice_tries,
5126 sizeof(int), 0644, proc_dointvec_minmax, false);
5127 set_table_entry(&table[10], "flags", &sd->flags,
5128 sizeof(int), 0644, proc_dointvec_minmax, false);
5129 set_table_entry(&table[11], "max_newidle_lb_cost",
5130 &sd->max_newidle_lb_cost,
5131 sizeof(long), 0644, proc_doulongvec_minmax, false);
5132 set_table_entry(&table[12], "name", sd->name,
5133 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5134 /* &table[13] is terminator */
5139 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5141 struct ctl_table *entry, *table;
5142 struct sched_domain *sd;
5143 int domain_num = 0, i;
5146 for_each_domain(cpu, sd)
5148 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5153 for_each_domain(cpu, sd) {
5154 snprintf(buf, 32, "domain%d", i);
5155 entry->procname = kstrdup(buf, GFP_KERNEL);
5157 entry->child = sd_alloc_ctl_domain_table(sd);
5164 static struct ctl_table_header *sd_sysctl_header;
5165 static void register_sched_domain_sysctl(void)
5167 int i, cpu_num = num_possible_cpus();
5168 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5171 WARN_ON(sd_ctl_dir[0].child);
5172 sd_ctl_dir[0].child = entry;
5177 for_each_possible_cpu(i) {
5178 snprintf(buf, 32, "cpu%d", i);
5179 entry->procname = kstrdup(buf, GFP_KERNEL);
5181 entry->child = sd_alloc_ctl_cpu_table(i);
5185 WARN_ON(sd_sysctl_header);
5186 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5189 /* may be called multiple times per register */
5190 static void unregister_sched_domain_sysctl(void)
5192 if (sd_sysctl_header)
5193 unregister_sysctl_table(sd_sysctl_header);
5194 sd_sysctl_header = NULL;
5195 if (sd_ctl_dir[0].child)
5196 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5199 static void register_sched_domain_sysctl(void)
5202 static void unregister_sched_domain_sysctl(void)
5207 static void set_rq_online(struct rq *rq)
5210 const struct sched_class *class;
5212 cpumask_set_cpu(rq->cpu, rq->rd->online);
5215 for_each_class(class) {
5216 if (class->rq_online)
5217 class->rq_online(rq);
5222 static void set_rq_offline(struct rq *rq)
5225 const struct sched_class *class;
5227 for_each_class(class) {
5228 if (class->rq_offline)
5229 class->rq_offline(rq);
5232 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5238 * migration_call - callback that gets triggered when a CPU is added.
5239 * Here we can start up the necessary migration thread for the new CPU.
5242 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5244 int cpu = (long)hcpu;
5245 unsigned long flags;
5246 struct rq *rq = cpu_rq(cpu);
5248 switch (action & ~CPU_TASKS_FROZEN) {
5250 case CPU_UP_PREPARE:
5251 rq->calc_load_update = calc_load_update;
5255 /* Update our root-domain */
5256 raw_spin_lock_irqsave(&rq->lock, flags);
5258 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5262 raw_spin_unlock_irqrestore(&rq->lock, flags);
5265 #ifdef CONFIG_HOTPLUG_CPU
5267 sched_ttwu_pending();
5268 /* Update our root-domain */
5269 raw_spin_lock_irqsave(&rq->lock, flags);
5271 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5275 BUG_ON(rq->nr_running != 1); /* the migration thread */
5276 raw_spin_unlock_irqrestore(&rq->lock, flags);
5280 calc_load_migrate(rq);
5285 update_max_interval();
5291 * Register at high priority so that task migration (migrate_all_tasks)
5292 * happens before everything else. This has to be lower priority than
5293 * the notifier in the perf_event subsystem, though.
5295 static struct notifier_block migration_notifier = {
5296 .notifier_call = migration_call,
5297 .priority = CPU_PRI_MIGRATION,
5300 static void __cpuinit set_cpu_rq_start_time(void)
5302 int cpu = smp_processor_id();
5303 struct rq *rq = cpu_rq(cpu);
5304 rq->age_stamp = sched_clock_cpu(cpu);
5307 static int sched_cpu_active(struct notifier_block *nfb,
5308 unsigned long action, void *hcpu)
5310 switch (action & ~CPU_TASKS_FROZEN) {
5312 set_cpu_rq_start_time();
5314 case CPU_DOWN_FAILED:
5315 set_cpu_active((long)hcpu, true);
5322 static int sched_cpu_inactive(struct notifier_block *nfb,
5323 unsigned long action, void *hcpu)
5325 unsigned long flags;
5326 long cpu = (long)hcpu;
5329 switch (action & ~CPU_TASKS_FROZEN) {
5330 case CPU_DOWN_PREPARE:
5331 set_cpu_active(cpu, false);
5333 /* explicitly allow suspend */
5334 if (!(action & CPU_TASKS_FROZEN)) {
5338 rcu_read_lock_sched();
5339 dl_b = dl_bw_of(cpu);
5341 raw_spin_lock_irqsave(&dl_b->lock, flags);
5342 cpus = dl_bw_cpus(cpu);
5343 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5344 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5346 rcu_read_unlock_sched();
5349 return notifier_from_errno(-EBUSY);
5357 static int __init migration_init(void)
5359 void *cpu = (void *)(long)smp_processor_id();
5362 /* Initialize migration for the boot CPU */
5363 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5364 BUG_ON(err == NOTIFY_BAD);
5365 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5366 register_cpu_notifier(&migration_notifier);
5368 /* Register cpu active notifiers */
5369 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5370 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5374 early_initcall(migration_init);
5379 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5381 #ifdef CONFIG_SCHED_DEBUG
5383 static __read_mostly int sched_debug_enabled;
5385 static int __init sched_debug_setup(char *str)
5387 sched_debug_enabled = 1;
5391 early_param("sched_debug", sched_debug_setup);
5393 static inline bool sched_debug(void)
5395 return sched_debug_enabled;
5398 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5399 struct cpumask *groupmask)
5401 struct sched_group *group = sd->groups;
5404 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5405 cpumask_clear(groupmask);
5407 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5409 if (!(sd->flags & SD_LOAD_BALANCE)) {
5410 printk("does not load-balance\n");
5412 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5417 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5419 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5420 printk(KERN_ERR "ERROR: domain->span does not contain "
5423 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5424 printk(KERN_ERR "ERROR: domain->groups does not contain"
5428 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5432 printk(KERN_ERR "ERROR: group is NULL\n");
5437 * Even though we initialize ->capacity to something semi-sane,
5438 * we leave capacity_orig unset. This allows us to detect if
5439 * domain iteration is still funny without causing /0 traps.
5441 if (!group->sgc->capacity_orig) {
5442 printk(KERN_CONT "\n");
5443 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5447 if (!cpumask_weight(sched_group_cpus(group))) {
5448 printk(KERN_CONT "\n");
5449 printk(KERN_ERR "ERROR: empty group\n");
5453 if (!(sd->flags & SD_OVERLAP) &&
5454 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5455 printk(KERN_CONT "\n");
5456 printk(KERN_ERR "ERROR: repeated CPUs\n");
5460 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5462 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5464 printk(KERN_CONT " %s", str);
5465 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5466 printk(KERN_CONT " (cpu_capacity = %d)",
5467 group->sgc->capacity);
5470 group = group->next;
5471 } while (group != sd->groups);
5472 printk(KERN_CONT "\n");
5474 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5475 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5478 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5479 printk(KERN_ERR "ERROR: parent span is not a superset "
5480 "of domain->span\n");
5484 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5488 if (!sched_debug_enabled)
5492 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5496 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5499 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5507 #else /* !CONFIG_SCHED_DEBUG */
5508 # define sched_domain_debug(sd, cpu) do { } while (0)
5509 static inline bool sched_debug(void)
5513 #endif /* CONFIG_SCHED_DEBUG */
5515 static int sd_degenerate(struct sched_domain *sd)
5517 if (cpumask_weight(sched_domain_span(sd)) == 1)
5520 /* Following flags need at least 2 groups */
5521 if (sd->flags & (SD_LOAD_BALANCE |
5522 SD_BALANCE_NEWIDLE |
5525 SD_SHARE_CPUCAPACITY |
5526 SD_SHARE_PKG_RESOURCES |
5527 SD_SHARE_POWERDOMAIN)) {
5528 if (sd->groups != sd->groups->next)
5532 /* Following flags don't use groups */
5533 if (sd->flags & (SD_WAKE_AFFINE))
5540 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5542 unsigned long cflags = sd->flags, pflags = parent->flags;
5544 if (sd_degenerate(parent))
5547 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5550 /* Flags needing groups don't count if only 1 group in parent */
5551 if (parent->groups == parent->groups->next) {
5552 pflags &= ~(SD_LOAD_BALANCE |
5553 SD_BALANCE_NEWIDLE |
5556 SD_SHARE_CPUCAPACITY |
5557 SD_SHARE_PKG_RESOURCES |
5559 SD_SHARE_POWERDOMAIN);
5560 if (nr_node_ids == 1)
5561 pflags &= ~SD_SERIALIZE;
5563 if (~cflags & pflags)
5569 static void free_rootdomain(struct rcu_head *rcu)
5571 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5573 cpupri_cleanup(&rd->cpupri);
5574 cpudl_cleanup(&rd->cpudl);
5575 free_cpumask_var(rd->dlo_mask);
5576 free_cpumask_var(rd->rto_mask);
5577 free_cpumask_var(rd->online);
5578 free_cpumask_var(rd->span);
5582 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5584 struct root_domain *old_rd = NULL;
5585 unsigned long flags;
5587 raw_spin_lock_irqsave(&rq->lock, flags);
5592 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5595 cpumask_clear_cpu(rq->cpu, old_rd->span);
5598 * If we dont want to free the old_rd yet then
5599 * set old_rd to NULL to skip the freeing later
5602 if (!atomic_dec_and_test(&old_rd->refcount))
5606 atomic_inc(&rd->refcount);
5609 cpumask_set_cpu(rq->cpu, rd->span);
5610 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5613 raw_spin_unlock_irqrestore(&rq->lock, flags);
5616 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5619 static int init_rootdomain(struct root_domain *rd)
5621 memset(rd, 0, sizeof(*rd));
5623 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5625 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5627 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5629 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5632 init_dl_bw(&rd->dl_bw);
5633 if (cpudl_init(&rd->cpudl) != 0)
5636 if (cpupri_init(&rd->cpupri) != 0)
5641 free_cpumask_var(rd->rto_mask);
5643 free_cpumask_var(rd->dlo_mask);
5645 free_cpumask_var(rd->online);
5647 free_cpumask_var(rd->span);
5653 * By default the system creates a single root-domain with all cpus as
5654 * members (mimicking the global state we have today).
5656 struct root_domain def_root_domain;
5658 static void init_defrootdomain(void)
5660 init_rootdomain(&def_root_domain);
5662 atomic_set(&def_root_domain.refcount, 1);
5665 static struct root_domain *alloc_rootdomain(void)
5667 struct root_domain *rd;
5669 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5673 if (init_rootdomain(rd) != 0) {
5681 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5683 struct sched_group *tmp, *first;
5692 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5697 } while (sg != first);
5700 static void free_sched_domain(struct rcu_head *rcu)
5702 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5705 * If its an overlapping domain it has private groups, iterate and
5708 if (sd->flags & SD_OVERLAP) {
5709 free_sched_groups(sd->groups, 1);
5710 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5711 kfree(sd->groups->sgc);
5717 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5719 call_rcu(&sd->rcu, free_sched_domain);
5722 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5724 for (; sd; sd = sd->parent)
5725 destroy_sched_domain(sd, cpu);
5729 * Keep a special pointer to the highest sched_domain that has
5730 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5731 * allows us to avoid some pointer chasing select_idle_sibling().
5733 * Also keep a unique ID per domain (we use the first cpu number in
5734 * the cpumask of the domain), this allows us to quickly tell if
5735 * two cpus are in the same cache domain, see cpus_share_cache().
5737 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5738 DEFINE_PER_CPU(int, sd_llc_size);
5739 DEFINE_PER_CPU(int, sd_llc_id);
5740 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5741 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5742 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5744 static void update_top_cache_domain(int cpu)
5746 struct sched_domain *sd;
5747 struct sched_domain *busy_sd = NULL;
5751 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5753 id = cpumask_first(sched_domain_span(sd));
5754 size = cpumask_weight(sched_domain_span(sd));
5755 busy_sd = sd->parent; /* sd_busy */
5757 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5759 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5760 per_cpu(sd_llc_size, cpu) = size;
5761 per_cpu(sd_llc_id, cpu) = id;
5763 sd = lowest_flag_domain(cpu, SD_NUMA);
5764 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5766 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5767 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5771 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5772 * hold the hotplug lock.
5775 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5777 struct rq *rq = cpu_rq(cpu);
5778 struct sched_domain *tmp;
5780 /* Remove the sched domains which do not contribute to scheduling. */
5781 for (tmp = sd; tmp; ) {
5782 struct sched_domain *parent = tmp->parent;
5786 if (sd_parent_degenerate(tmp, parent)) {
5787 tmp->parent = parent->parent;
5789 parent->parent->child = tmp;
5791 * Transfer SD_PREFER_SIBLING down in case of a
5792 * degenerate parent; the spans match for this
5793 * so the property transfers.
5795 if (parent->flags & SD_PREFER_SIBLING)
5796 tmp->flags |= SD_PREFER_SIBLING;
5797 destroy_sched_domain(parent, cpu);
5802 if (sd && sd_degenerate(sd)) {
5805 destroy_sched_domain(tmp, cpu);
5810 sched_domain_debug(sd, cpu);
5812 rq_attach_root(rq, rd);
5814 rcu_assign_pointer(rq->sd, sd);
5815 destroy_sched_domains(tmp, cpu);
5817 update_top_cache_domain(cpu);
5820 /* cpus with isolated domains */
5821 static cpumask_var_t cpu_isolated_map;
5823 /* Setup the mask of cpus configured for isolated domains */
5824 static int __init isolated_cpu_setup(char *str)
5826 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5827 cpulist_parse(str, cpu_isolated_map);
5831 __setup("isolcpus=", isolated_cpu_setup);
5834 struct sched_domain ** __percpu sd;
5835 struct root_domain *rd;
5846 * Build an iteration mask that can exclude certain CPUs from the upwards
5849 * Asymmetric node setups can result in situations where the domain tree is of
5850 * unequal depth, make sure to skip domains that already cover the entire
5853 * In that case build_sched_domains() will have terminated the iteration early
5854 * and our sibling sd spans will be empty. Domains should always include the
5855 * cpu they're built on, so check that.
5858 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5860 const struct cpumask *span = sched_domain_span(sd);
5861 struct sd_data *sdd = sd->private;
5862 struct sched_domain *sibling;
5865 for_each_cpu(i, span) {
5866 sibling = *per_cpu_ptr(sdd->sd, i);
5867 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5870 cpumask_set_cpu(i, sched_group_mask(sg));
5875 * Return the canonical balance cpu for this group, this is the first cpu
5876 * of this group that's also in the iteration mask.
5878 int group_balance_cpu(struct sched_group *sg)
5880 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5884 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5886 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5887 const struct cpumask *span = sched_domain_span(sd);
5888 struct cpumask *covered = sched_domains_tmpmask;
5889 struct sd_data *sdd = sd->private;
5890 struct sched_domain *sibling;
5893 cpumask_clear(covered);
5895 for_each_cpu(i, span) {
5896 struct cpumask *sg_span;
5898 if (cpumask_test_cpu(i, covered))
5901 sibling = *per_cpu_ptr(sdd->sd, i);
5903 /* See the comment near build_group_mask(). */
5904 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5907 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5908 GFP_KERNEL, cpu_to_node(cpu));
5913 sg_span = sched_group_cpus(sg);
5915 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5917 cpumask_set_cpu(i, sg_span);
5919 cpumask_or(covered, covered, sg_span);
5921 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5922 if (atomic_inc_return(&sg->sgc->ref) == 1)
5923 build_group_mask(sd, sg);
5926 * Initialize sgc->capacity such that even if we mess up the
5927 * domains and no possible iteration will get us here, we won't
5930 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5931 sg->sgc->capacity_orig = sg->sgc->capacity;
5934 * Make sure the first group of this domain contains the
5935 * canonical balance cpu. Otherwise the sched_domain iteration
5936 * breaks. See update_sg_lb_stats().
5938 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5939 group_balance_cpu(sg) == cpu)
5949 sd->groups = groups;
5954 free_sched_groups(first, 0);
5959 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5961 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5962 struct sched_domain *child = sd->child;
5965 cpu = cpumask_first(sched_domain_span(child));
5968 *sg = *per_cpu_ptr(sdd->sg, cpu);
5969 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5970 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5977 * build_sched_groups will build a circular linked list of the groups
5978 * covered by the given span, and will set each group's ->cpumask correctly,
5979 * and ->cpu_capacity to 0.
5981 * Assumes the sched_domain tree is fully constructed
5984 build_sched_groups(struct sched_domain *sd, int cpu)
5986 struct sched_group *first = NULL, *last = NULL;
5987 struct sd_data *sdd = sd->private;
5988 const struct cpumask *span = sched_domain_span(sd);
5989 struct cpumask *covered;
5992 get_group(cpu, sdd, &sd->groups);
5993 atomic_inc(&sd->groups->ref);
5995 if (cpu != cpumask_first(span))
5998 lockdep_assert_held(&sched_domains_mutex);
5999 covered = sched_domains_tmpmask;
6001 cpumask_clear(covered);
6003 for_each_cpu(i, span) {
6004 struct sched_group *sg;
6007 if (cpumask_test_cpu(i, covered))
6010 group = get_group(i, sdd, &sg);
6011 cpumask_setall(sched_group_mask(sg));
6013 for_each_cpu(j, span) {
6014 if (get_group(j, sdd, NULL) != group)
6017 cpumask_set_cpu(j, covered);
6018 cpumask_set_cpu(j, sched_group_cpus(sg));
6033 * Initialize sched groups cpu_capacity.
6035 * cpu_capacity indicates the capacity of sched group, which is used while
6036 * distributing the load between different sched groups in a sched domain.
6037 * Typically cpu_capacity for all the groups in a sched domain will be same
6038 * unless there are asymmetries in the topology. If there are asymmetries,
6039 * group having more cpu_capacity will pickup more load compared to the
6040 * group having less cpu_capacity.
6042 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6044 struct sched_group *sg = sd->groups;
6049 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6051 } while (sg != sd->groups);
6053 if (cpu != group_balance_cpu(sg))
6056 update_group_capacity(sd, cpu);
6057 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6061 * Initializers for schedule domains
6062 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6065 static int default_relax_domain_level = -1;
6066 int sched_domain_level_max;
6068 static int __init setup_relax_domain_level(char *str)
6070 if (kstrtoint(str, 0, &default_relax_domain_level))
6071 pr_warn("Unable to set relax_domain_level\n");
6075 __setup("relax_domain_level=", setup_relax_domain_level);
6077 static void set_domain_attribute(struct sched_domain *sd,
6078 struct sched_domain_attr *attr)
6082 if (!attr || attr->relax_domain_level < 0) {
6083 if (default_relax_domain_level < 0)
6086 request = default_relax_domain_level;
6088 request = attr->relax_domain_level;
6089 if (request < sd->level) {
6090 /* turn off idle balance on this domain */
6091 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6093 /* turn on idle balance on this domain */
6094 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6098 static void __sdt_free(const struct cpumask *cpu_map);
6099 static int __sdt_alloc(const struct cpumask *cpu_map);
6101 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6102 const struct cpumask *cpu_map)
6106 if (!atomic_read(&d->rd->refcount))
6107 free_rootdomain(&d->rd->rcu); /* fall through */
6109 free_percpu(d->sd); /* fall through */
6111 __sdt_free(cpu_map); /* fall through */
6117 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6118 const struct cpumask *cpu_map)
6120 memset(d, 0, sizeof(*d));
6122 if (__sdt_alloc(cpu_map))
6123 return sa_sd_storage;
6124 d->sd = alloc_percpu(struct sched_domain *);
6126 return sa_sd_storage;
6127 d->rd = alloc_rootdomain();
6130 return sa_rootdomain;
6134 * NULL the sd_data elements we've used to build the sched_domain and
6135 * sched_group structure so that the subsequent __free_domain_allocs()
6136 * will not free the data we're using.
6138 static void claim_allocations(int cpu, struct sched_domain *sd)
6140 struct sd_data *sdd = sd->private;
6142 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6143 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6145 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6146 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6148 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6149 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6153 static int sched_domains_numa_levels;
6154 enum numa_topology_type sched_numa_topology_type;
6155 static int *sched_domains_numa_distance;
6156 int sched_max_numa_distance;
6157 static struct cpumask ***sched_domains_numa_masks;
6158 static int sched_domains_curr_level;
6162 * SD_flags allowed in topology descriptions.
6164 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6165 * SD_SHARE_PKG_RESOURCES - describes shared caches
6166 * SD_NUMA - describes NUMA topologies
6167 * SD_SHARE_POWERDOMAIN - describes shared power domain
6170 * SD_ASYM_PACKING - describes SMT quirks
6172 #define TOPOLOGY_SD_FLAGS \
6173 (SD_SHARE_CPUCAPACITY | \
6174 SD_SHARE_PKG_RESOURCES | \
6177 SD_SHARE_POWERDOMAIN)
6179 static struct sched_domain *
6180 sd_init(struct sched_domain_topology_level *tl, int cpu)
6182 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6183 int sd_weight, sd_flags = 0;
6187 * Ugly hack to pass state to sd_numa_mask()...
6189 sched_domains_curr_level = tl->numa_level;
6192 sd_weight = cpumask_weight(tl->mask(cpu));
6195 sd_flags = (*tl->sd_flags)();
6196 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6197 "wrong sd_flags in topology description\n"))
6198 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6200 *sd = (struct sched_domain){
6201 .min_interval = sd_weight,
6202 .max_interval = 2*sd_weight,
6204 .imbalance_pct = 125,
6206 .cache_nice_tries = 0,
6213 .flags = 1*SD_LOAD_BALANCE
6214 | 1*SD_BALANCE_NEWIDLE
6219 | 0*SD_SHARE_CPUCAPACITY
6220 | 0*SD_SHARE_PKG_RESOURCES
6222 | 0*SD_PREFER_SIBLING
6227 .last_balance = jiffies,
6228 .balance_interval = sd_weight,
6230 .max_newidle_lb_cost = 0,
6231 .next_decay_max_lb_cost = jiffies,
6232 #ifdef CONFIG_SCHED_DEBUG
6238 * Convert topological properties into behaviour.
6241 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6242 sd->imbalance_pct = 110;
6243 sd->smt_gain = 1178; /* ~15% */
6245 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6246 sd->imbalance_pct = 117;
6247 sd->cache_nice_tries = 1;
6251 } else if (sd->flags & SD_NUMA) {
6252 sd->cache_nice_tries = 2;
6256 sd->flags |= SD_SERIALIZE;
6257 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6258 sd->flags &= ~(SD_BALANCE_EXEC |
6265 sd->flags |= SD_PREFER_SIBLING;
6266 sd->cache_nice_tries = 1;
6271 sd->private = &tl->data;
6277 * Topology list, bottom-up.
6279 static struct sched_domain_topology_level default_topology[] = {
6280 #ifdef CONFIG_SCHED_SMT
6281 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6283 #ifdef CONFIG_SCHED_MC
6284 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6286 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6290 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6292 #define for_each_sd_topology(tl) \
6293 for (tl = sched_domain_topology; tl->mask; tl++)
6295 void set_sched_topology(struct sched_domain_topology_level *tl)
6297 sched_domain_topology = tl;
6302 static const struct cpumask *sd_numa_mask(int cpu)
6304 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6307 static void sched_numa_warn(const char *str)
6309 static int done = false;
6317 printk(KERN_WARNING "ERROR: %s\n\n", str);
6319 for (i = 0; i < nr_node_ids; i++) {
6320 printk(KERN_WARNING " ");
6321 for (j = 0; j < nr_node_ids; j++)
6322 printk(KERN_CONT "%02d ", node_distance(i,j));
6323 printk(KERN_CONT "\n");
6325 printk(KERN_WARNING "\n");
6328 bool find_numa_distance(int distance)
6332 if (distance == node_distance(0, 0))
6335 for (i = 0; i < sched_domains_numa_levels; i++) {
6336 if (sched_domains_numa_distance[i] == distance)
6344 * A system can have three types of NUMA topology:
6345 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6346 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6347 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6349 * The difference between a glueless mesh topology and a backplane
6350 * topology lies in whether communication between not directly
6351 * connected nodes goes through intermediary nodes (where programs
6352 * could run), or through backplane controllers. This affects
6353 * placement of programs.
6355 * The type of topology can be discerned with the following tests:
6356 * - If the maximum distance between any nodes is 1 hop, the system
6357 * is directly connected.
6358 * - If for two nodes A and B, located N > 1 hops away from each other,
6359 * there is an intermediary node C, which is < N hops away from both
6360 * nodes A and B, the system is a glueless mesh.
6362 static void init_numa_topology_type(void)
6366 n = sched_max_numa_distance;
6369 sched_numa_topology_type = NUMA_DIRECT;
6371 for_each_online_node(a) {
6372 for_each_online_node(b) {
6373 /* Find two nodes furthest removed from each other. */
6374 if (node_distance(a, b) < n)
6377 /* Is there an intermediary node between a and b? */
6378 for_each_online_node(c) {
6379 if (node_distance(a, c) < n &&
6380 node_distance(b, c) < n) {
6381 sched_numa_topology_type =
6387 sched_numa_topology_type = NUMA_BACKPLANE;
6393 static void sched_init_numa(void)
6395 int next_distance, curr_distance = node_distance(0, 0);
6396 struct sched_domain_topology_level *tl;
6400 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6401 if (!sched_domains_numa_distance)
6405 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6406 * unique distances in the node_distance() table.
6408 * Assumes node_distance(0,j) includes all distances in
6409 * node_distance(i,j) in order to avoid cubic time.
6411 next_distance = curr_distance;
6412 for (i = 0; i < nr_node_ids; i++) {
6413 for (j = 0; j < nr_node_ids; j++) {
6414 for (k = 0; k < nr_node_ids; k++) {
6415 int distance = node_distance(i, k);
6417 if (distance > curr_distance &&
6418 (distance < next_distance ||
6419 next_distance == curr_distance))
6420 next_distance = distance;
6423 * While not a strong assumption it would be nice to know
6424 * about cases where if node A is connected to B, B is not
6425 * equally connected to A.
6427 if (sched_debug() && node_distance(k, i) != distance)
6428 sched_numa_warn("Node-distance not symmetric");
6430 if (sched_debug() && i && !find_numa_distance(distance))
6431 sched_numa_warn("Node-0 not representative");
6433 if (next_distance != curr_distance) {
6434 sched_domains_numa_distance[level++] = next_distance;
6435 sched_domains_numa_levels = level;
6436 curr_distance = next_distance;
6441 * In case of sched_debug() we verify the above assumption.
6447 * 'level' contains the number of unique distances, excluding the
6448 * identity distance node_distance(i,i).
6450 * The sched_domains_numa_distance[] array includes the actual distance
6455 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6456 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6457 * the array will contain less then 'level' members. This could be
6458 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6459 * in other functions.
6461 * We reset it to 'level' at the end of this function.
6463 sched_domains_numa_levels = 0;
6465 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6466 if (!sched_domains_numa_masks)
6470 * Now for each level, construct a mask per node which contains all
6471 * cpus of nodes that are that many hops away from us.
6473 for (i = 0; i < level; i++) {
6474 sched_domains_numa_masks[i] =
6475 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6476 if (!sched_domains_numa_masks[i])
6479 for (j = 0; j < nr_node_ids; j++) {
6480 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6484 sched_domains_numa_masks[i][j] = mask;
6486 for (k = 0; k < nr_node_ids; k++) {
6487 if (node_distance(j, k) > sched_domains_numa_distance[i])
6490 cpumask_or(mask, mask, cpumask_of_node(k));
6495 /* Compute default topology size */
6496 for (i = 0; sched_domain_topology[i].mask; i++);
6498 tl = kzalloc((i + level + 1) *
6499 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6504 * Copy the default topology bits..
6506 for (i = 0; sched_domain_topology[i].mask; i++)
6507 tl[i] = sched_domain_topology[i];
6510 * .. and append 'j' levels of NUMA goodness.
6512 for (j = 0; j < level; i++, j++) {
6513 tl[i] = (struct sched_domain_topology_level){
6514 .mask = sd_numa_mask,
6515 .sd_flags = cpu_numa_flags,
6516 .flags = SDTL_OVERLAP,
6522 sched_domain_topology = tl;
6524 sched_domains_numa_levels = level;
6525 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6527 init_numa_topology_type();
6530 static void sched_domains_numa_masks_set(int cpu)
6533 int node = cpu_to_node(cpu);
6535 for (i = 0; i < sched_domains_numa_levels; i++) {
6536 for (j = 0; j < nr_node_ids; j++) {
6537 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6538 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6543 static void sched_domains_numa_masks_clear(int cpu)
6546 for (i = 0; i < sched_domains_numa_levels; i++) {
6547 for (j = 0; j < nr_node_ids; j++)
6548 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6553 * Update sched_domains_numa_masks[level][node] array when new cpus
6556 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6557 unsigned long action,
6560 int cpu = (long)hcpu;
6562 switch (action & ~CPU_TASKS_FROZEN) {
6564 sched_domains_numa_masks_set(cpu);
6568 sched_domains_numa_masks_clear(cpu);
6578 static inline void sched_init_numa(void)
6582 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6583 unsigned long action,
6588 #endif /* CONFIG_NUMA */
6590 static int __sdt_alloc(const struct cpumask *cpu_map)
6592 struct sched_domain_topology_level *tl;
6595 for_each_sd_topology(tl) {
6596 struct sd_data *sdd = &tl->data;
6598 sdd->sd = alloc_percpu(struct sched_domain *);
6602 sdd->sg = alloc_percpu(struct sched_group *);
6606 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6610 for_each_cpu(j, cpu_map) {
6611 struct sched_domain *sd;
6612 struct sched_group *sg;
6613 struct sched_group_capacity *sgc;
6615 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6616 GFP_KERNEL, cpu_to_node(j));
6620 *per_cpu_ptr(sdd->sd, j) = sd;
6622 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6623 GFP_KERNEL, cpu_to_node(j));
6629 *per_cpu_ptr(sdd->sg, j) = sg;
6631 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6632 GFP_KERNEL, cpu_to_node(j));
6636 *per_cpu_ptr(sdd->sgc, j) = sgc;
6643 static void __sdt_free(const struct cpumask *cpu_map)
6645 struct sched_domain_topology_level *tl;
6648 for_each_sd_topology(tl) {
6649 struct sd_data *sdd = &tl->data;
6651 for_each_cpu(j, cpu_map) {
6652 struct sched_domain *sd;
6655 sd = *per_cpu_ptr(sdd->sd, j);
6656 if (sd && (sd->flags & SD_OVERLAP))
6657 free_sched_groups(sd->groups, 0);
6658 kfree(*per_cpu_ptr(sdd->sd, j));
6662 kfree(*per_cpu_ptr(sdd->sg, j));
6664 kfree(*per_cpu_ptr(sdd->sgc, j));
6666 free_percpu(sdd->sd);
6668 free_percpu(sdd->sg);
6670 free_percpu(sdd->sgc);
6675 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6676 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6677 struct sched_domain *child, int cpu)
6679 struct sched_domain *sd = sd_init(tl, cpu);
6683 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6685 sd->level = child->level + 1;
6686 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6690 if (!cpumask_subset(sched_domain_span(child),
6691 sched_domain_span(sd))) {
6692 pr_err("BUG: arch topology borken\n");
6693 #ifdef CONFIG_SCHED_DEBUG
6694 pr_err(" the %s domain not a subset of the %s domain\n",
6695 child->name, sd->name);
6697 /* Fixup, ensure @sd has at least @child cpus. */
6698 cpumask_or(sched_domain_span(sd),
6699 sched_domain_span(sd),
6700 sched_domain_span(child));
6704 set_domain_attribute(sd, attr);
6710 * Build sched domains for a given set of cpus and attach the sched domains
6711 * to the individual cpus
6713 static int build_sched_domains(const struct cpumask *cpu_map,
6714 struct sched_domain_attr *attr)
6716 enum s_alloc alloc_state;
6717 struct sched_domain *sd;
6719 int i, ret = -ENOMEM;
6721 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6722 if (alloc_state != sa_rootdomain)
6725 /* Set up domains for cpus specified by the cpu_map. */
6726 for_each_cpu(i, cpu_map) {
6727 struct sched_domain_topology_level *tl;
6730 for_each_sd_topology(tl) {
6731 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6732 if (tl == sched_domain_topology)
6733 *per_cpu_ptr(d.sd, i) = sd;
6734 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6735 sd->flags |= SD_OVERLAP;
6736 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6741 /* Build the groups for the domains */
6742 for_each_cpu(i, cpu_map) {
6743 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6744 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6745 if (sd->flags & SD_OVERLAP) {
6746 if (build_overlap_sched_groups(sd, i))
6749 if (build_sched_groups(sd, i))
6755 /* Calculate CPU capacity for physical packages and nodes */
6756 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6757 if (!cpumask_test_cpu(i, cpu_map))
6760 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6761 claim_allocations(i, sd);
6762 init_sched_groups_capacity(i, sd);
6766 /* Attach the domains */
6768 for_each_cpu(i, cpu_map) {
6769 sd = *per_cpu_ptr(d.sd, i);
6770 cpu_attach_domain(sd, d.rd, i);
6776 __free_domain_allocs(&d, alloc_state, cpu_map);
6780 static cpumask_var_t *doms_cur; /* current sched domains */
6781 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6782 static struct sched_domain_attr *dattr_cur;
6783 /* attribues of custom domains in 'doms_cur' */
6786 * Special case: If a kmalloc of a doms_cur partition (array of
6787 * cpumask) fails, then fallback to a single sched domain,
6788 * as determined by the single cpumask fallback_doms.
6790 static cpumask_var_t fallback_doms;
6793 * arch_update_cpu_topology lets virtualized architectures update the
6794 * cpu core maps. It is supposed to return 1 if the topology changed
6795 * or 0 if it stayed the same.
6797 int __weak arch_update_cpu_topology(void)
6802 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6805 cpumask_var_t *doms;
6807 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6810 for (i = 0; i < ndoms; i++) {
6811 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6812 free_sched_domains(doms, i);
6819 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6822 for (i = 0; i < ndoms; i++)
6823 free_cpumask_var(doms[i]);
6828 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6829 * For now this just excludes isolated cpus, but could be used to
6830 * exclude other special cases in the future.
6832 static int init_sched_domains(const struct cpumask *cpu_map)
6836 arch_update_cpu_topology();
6838 doms_cur = alloc_sched_domains(ndoms_cur);
6840 doms_cur = &fallback_doms;
6841 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6842 err = build_sched_domains(doms_cur[0], NULL);
6843 register_sched_domain_sysctl();
6849 * Detach sched domains from a group of cpus specified in cpu_map
6850 * These cpus will now be attached to the NULL domain
6852 static void detach_destroy_domains(const struct cpumask *cpu_map)
6857 for_each_cpu(i, cpu_map)
6858 cpu_attach_domain(NULL, &def_root_domain, i);
6862 /* handle null as "default" */
6863 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6864 struct sched_domain_attr *new, int idx_new)
6866 struct sched_domain_attr tmp;
6873 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6874 new ? (new + idx_new) : &tmp,
6875 sizeof(struct sched_domain_attr));
6879 * Partition sched domains as specified by the 'ndoms_new'
6880 * cpumasks in the array doms_new[] of cpumasks. This compares
6881 * doms_new[] to the current sched domain partitioning, doms_cur[].
6882 * It destroys each deleted domain and builds each new domain.
6884 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6885 * The masks don't intersect (don't overlap.) We should setup one
6886 * sched domain for each mask. CPUs not in any of the cpumasks will
6887 * not be load balanced. If the same cpumask appears both in the
6888 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6891 * The passed in 'doms_new' should be allocated using
6892 * alloc_sched_domains. This routine takes ownership of it and will
6893 * free_sched_domains it when done with it. If the caller failed the
6894 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6895 * and partition_sched_domains() will fallback to the single partition
6896 * 'fallback_doms', it also forces the domains to be rebuilt.
6898 * If doms_new == NULL it will be replaced with cpu_online_mask.
6899 * ndoms_new == 0 is a special case for destroying existing domains,
6900 * and it will not create the default domain.
6902 * Call with hotplug lock held
6904 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6905 struct sched_domain_attr *dattr_new)
6910 mutex_lock(&sched_domains_mutex);
6912 /* always unregister in case we don't destroy any domains */
6913 unregister_sched_domain_sysctl();
6915 /* Let architecture update cpu core mappings. */
6916 new_topology = arch_update_cpu_topology();
6918 n = doms_new ? ndoms_new : 0;
6920 /* Destroy deleted domains */
6921 for (i = 0; i < ndoms_cur; i++) {
6922 for (j = 0; j < n && !new_topology; j++) {
6923 if (cpumask_equal(doms_cur[i], doms_new[j])
6924 && dattrs_equal(dattr_cur, i, dattr_new, j))
6927 /* no match - a current sched domain not in new doms_new[] */
6928 detach_destroy_domains(doms_cur[i]);
6934 if (doms_new == NULL) {
6936 doms_new = &fallback_doms;
6937 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6938 WARN_ON_ONCE(dattr_new);
6941 /* Build new domains */
6942 for (i = 0; i < ndoms_new; i++) {
6943 for (j = 0; j < n && !new_topology; j++) {
6944 if (cpumask_equal(doms_new[i], doms_cur[j])
6945 && dattrs_equal(dattr_new, i, dattr_cur, j))
6948 /* no match - add a new doms_new */
6949 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6954 /* Remember the new sched domains */
6955 if (doms_cur != &fallback_doms)
6956 free_sched_domains(doms_cur, ndoms_cur);
6957 kfree(dattr_cur); /* kfree(NULL) is safe */
6958 doms_cur = doms_new;
6959 dattr_cur = dattr_new;
6960 ndoms_cur = ndoms_new;
6962 register_sched_domain_sysctl();
6964 mutex_unlock(&sched_domains_mutex);
6967 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6970 * Update cpusets according to cpu_active mask. If cpusets are
6971 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6972 * around partition_sched_domains().
6974 * If we come here as part of a suspend/resume, don't touch cpusets because we
6975 * want to restore it back to its original state upon resume anyway.
6977 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6981 case CPU_ONLINE_FROZEN:
6982 case CPU_DOWN_FAILED_FROZEN:
6985 * num_cpus_frozen tracks how many CPUs are involved in suspend
6986 * resume sequence. As long as this is not the last online
6987 * operation in the resume sequence, just build a single sched
6988 * domain, ignoring cpusets.
6991 if (likely(num_cpus_frozen)) {
6992 partition_sched_domains(1, NULL, NULL);
6997 * This is the last CPU online operation. So fall through and
6998 * restore the original sched domains by considering the
6999 * cpuset configurations.
7003 case CPU_DOWN_FAILED:
7004 cpuset_update_active_cpus(true);
7012 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7016 case CPU_DOWN_PREPARE:
7017 cpuset_update_active_cpus(false);
7019 case CPU_DOWN_PREPARE_FROZEN:
7021 partition_sched_domains(1, NULL, NULL);
7029 void __init sched_init_smp(void)
7031 cpumask_var_t non_isolated_cpus;
7033 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7034 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7039 * There's no userspace yet to cause hotplug operations; hence all the
7040 * cpu masks are stable and all blatant races in the below code cannot
7043 mutex_lock(&sched_domains_mutex);
7044 init_sched_domains(cpu_active_mask);
7045 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7046 if (cpumask_empty(non_isolated_cpus))
7047 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7048 mutex_unlock(&sched_domains_mutex);
7050 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7051 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7052 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7056 /* Move init over to a non-isolated CPU */
7057 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7059 sched_init_granularity();
7060 free_cpumask_var(non_isolated_cpus);
7062 init_sched_rt_class();
7063 init_sched_dl_class();
7066 void __init sched_init_smp(void)
7068 sched_init_granularity();
7070 #endif /* CONFIG_SMP */
7072 const_debug unsigned int sysctl_timer_migration = 1;
7074 int in_sched_functions(unsigned long addr)
7076 return in_lock_functions(addr) ||
7077 (addr >= (unsigned long)__sched_text_start
7078 && addr < (unsigned long)__sched_text_end);
7081 #ifdef CONFIG_CGROUP_SCHED
7083 * Default task group.
7084 * Every task in system belongs to this group at bootup.
7086 struct task_group root_task_group;
7087 LIST_HEAD(task_groups);
7090 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7092 void __init sched_init(void)
7095 unsigned long alloc_size = 0, ptr;
7097 #ifdef CONFIG_FAIR_GROUP_SCHED
7098 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7100 #ifdef CONFIG_RT_GROUP_SCHED
7101 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7103 #ifdef CONFIG_CPUMASK_OFFSTACK
7104 alloc_size += num_possible_cpus() * cpumask_size();
7107 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7109 #ifdef CONFIG_FAIR_GROUP_SCHED
7110 root_task_group.se = (struct sched_entity **)ptr;
7111 ptr += nr_cpu_ids * sizeof(void **);
7113 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7114 ptr += nr_cpu_ids * sizeof(void **);
7116 #endif /* CONFIG_FAIR_GROUP_SCHED */
7117 #ifdef CONFIG_RT_GROUP_SCHED
7118 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7119 ptr += nr_cpu_ids * sizeof(void **);
7121 root_task_group.rt_rq = (struct rt_rq **)ptr;
7122 ptr += nr_cpu_ids * sizeof(void **);
7124 #endif /* CONFIG_RT_GROUP_SCHED */
7125 #ifdef CONFIG_CPUMASK_OFFSTACK
7126 for_each_possible_cpu(i) {
7127 per_cpu(load_balance_mask, i) = (void *)ptr;
7128 ptr += cpumask_size();
7130 #endif /* CONFIG_CPUMASK_OFFSTACK */
7133 init_rt_bandwidth(&def_rt_bandwidth,
7134 global_rt_period(), global_rt_runtime());
7135 init_dl_bandwidth(&def_dl_bandwidth,
7136 global_rt_period(), global_rt_runtime());
7139 init_defrootdomain();
7142 #ifdef CONFIG_RT_GROUP_SCHED
7143 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7144 global_rt_period(), global_rt_runtime());
7145 #endif /* CONFIG_RT_GROUP_SCHED */
7147 #ifdef CONFIG_CGROUP_SCHED
7148 list_add(&root_task_group.list, &task_groups);
7149 INIT_LIST_HEAD(&root_task_group.children);
7150 INIT_LIST_HEAD(&root_task_group.siblings);
7151 autogroup_init(&init_task);
7153 #endif /* CONFIG_CGROUP_SCHED */
7155 for_each_possible_cpu(i) {
7159 raw_spin_lock_init(&rq->lock);
7161 rq->calc_load_active = 0;
7162 rq->calc_load_update = jiffies + LOAD_FREQ;
7163 init_cfs_rq(&rq->cfs);
7164 init_rt_rq(&rq->rt, rq);
7165 init_dl_rq(&rq->dl, rq);
7166 #ifdef CONFIG_FAIR_GROUP_SCHED
7167 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7168 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7170 * How much cpu bandwidth does root_task_group get?
7172 * In case of task-groups formed thr' the cgroup filesystem, it
7173 * gets 100% of the cpu resources in the system. This overall
7174 * system cpu resource is divided among the tasks of
7175 * root_task_group and its child task-groups in a fair manner,
7176 * based on each entity's (task or task-group's) weight
7177 * (se->load.weight).
7179 * In other words, if root_task_group has 10 tasks of weight
7180 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7181 * then A0's share of the cpu resource is:
7183 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7185 * We achieve this by letting root_task_group's tasks sit
7186 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7188 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7189 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7190 #endif /* CONFIG_FAIR_GROUP_SCHED */
7192 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7193 #ifdef CONFIG_RT_GROUP_SCHED
7194 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7197 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7198 rq->cpu_load[j] = 0;
7200 rq->last_load_update_tick = jiffies;
7205 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7206 rq->post_schedule = 0;
7207 rq->active_balance = 0;
7208 rq->next_balance = jiffies;
7213 rq->avg_idle = 2*sysctl_sched_migration_cost;
7214 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7216 INIT_LIST_HEAD(&rq->cfs_tasks);
7218 rq_attach_root(rq, &def_root_domain);
7219 #ifdef CONFIG_NO_HZ_COMMON
7222 #ifdef CONFIG_NO_HZ_FULL
7223 rq->last_sched_tick = 0;
7227 atomic_set(&rq->nr_iowait, 0);
7230 set_load_weight(&init_task);
7232 #ifdef CONFIG_PREEMPT_NOTIFIERS
7233 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7237 * The boot idle thread does lazy MMU switching as well:
7239 atomic_inc(&init_mm.mm_count);
7240 enter_lazy_tlb(&init_mm, current);
7243 * Make us the idle thread. Technically, schedule() should not be
7244 * called from this thread, however somewhere below it might be,
7245 * but because we are the idle thread, we just pick up running again
7246 * when this runqueue becomes "idle".
7248 init_idle(current, smp_processor_id());
7250 calc_load_update = jiffies + LOAD_FREQ;
7253 * During early bootup we pretend to be a normal task:
7255 current->sched_class = &fair_sched_class;
7258 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7259 /* May be allocated at isolcpus cmdline parse time */
7260 if (cpu_isolated_map == NULL)
7261 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7262 idle_thread_set_boot_cpu();
7263 set_cpu_rq_start_time();
7265 init_sched_fair_class();
7267 scheduler_running = 1;
7270 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7271 static inline int preempt_count_equals(int preempt_offset)
7273 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7275 return (nested == preempt_offset);
7278 void __might_sleep(const char *file, int line, int preempt_offset)
7280 static unsigned long prev_jiffy; /* ratelimiting */
7282 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7283 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7284 !is_idle_task(current)) ||
7285 system_state != SYSTEM_RUNNING || oops_in_progress)
7287 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7289 prev_jiffy = jiffies;
7292 "BUG: sleeping function called from invalid context at %s:%d\n",
7295 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7296 in_atomic(), irqs_disabled(),
7297 current->pid, current->comm);
7299 debug_show_held_locks(current);
7300 if (irqs_disabled())
7301 print_irqtrace_events(current);
7302 #ifdef CONFIG_DEBUG_PREEMPT
7303 if (!preempt_count_equals(preempt_offset)) {
7304 pr_err("Preemption disabled at:");
7305 print_ip_sym(current->preempt_disable_ip);
7311 EXPORT_SYMBOL(__might_sleep);
7314 #ifdef CONFIG_MAGIC_SYSRQ
7315 static void normalize_task(struct rq *rq, struct task_struct *p)
7317 const struct sched_class *prev_class = p->sched_class;
7318 struct sched_attr attr = {
7319 .sched_policy = SCHED_NORMAL,
7321 int old_prio = p->prio;
7324 queued = task_on_rq_queued(p);
7326 dequeue_task(rq, p, 0);
7327 __setscheduler(rq, p, &attr);
7329 enqueue_task(rq, p, 0);
7333 check_class_changed(rq, p, prev_class, old_prio);
7336 void normalize_rt_tasks(void)
7338 struct task_struct *g, *p;
7339 unsigned long flags;
7342 read_lock(&tasklist_lock);
7343 for_each_process_thread(g, p) {
7345 * Only normalize user tasks:
7347 if (p->flags & PF_KTHREAD)
7350 p->se.exec_start = 0;
7351 #ifdef CONFIG_SCHEDSTATS
7352 p->se.statistics.wait_start = 0;
7353 p->se.statistics.sleep_start = 0;
7354 p->se.statistics.block_start = 0;
7357 if (!dl_task(p) && !rt_task(p)) {
7359 * Renice negative nice level userspace
7362 if (task_nice(p) < 0)
7363 set_user_nice(p, 0);
7367 rq = task_rq_lock(p, &flags);
7368 normalize_task(rq, p);
7369 task_rq_unlock(rq, p, &flags);
7371 read_unlock(&tasklist_lock);
7374 #endif /* CONFIG_MAGIC_SYSRQ */
7376 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7378 * These functions are only useful for the IA64 MCA handling, or kdb.
7380 * They can only be called when the whole system has been
7381 * stopped - every CPU needs to be quiescent, and no scheduling
7382 * activity can take place. Using them for anything else would
7383 * be a serious bug, and as a result, they aren't even visible
7384 * under any other configuration.
7388 * curr_task - return the current task for a given cpu.
7389 * @cpu: the processor in question.
7391 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7393 * Return: The current task for @cpu.
7395 struct task_struct *curr_task(int cpu)
7397 return cpu_curr(cpu);
7400 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7404 * set_curr_task - set the current task for a given cpu.
7405 * @cpu: the processor in question.
7406 * @p: the task pointer to set.
7408 * Description: This function must only be used when non-maskable interrupts
7409 * are serviced on a separate stack. It allows the architecture to switch the
7410 * notion of the current task on a cpu in a non-blocking manner. This function
7411 * must be called with all CPU's synchronized, and interrupts disabled, the
7412 * and caller must save the original value of the current task (see
7413 * curr_task() above) and restore that value before reenabling interrupts and
7414 * re-starting the system.
7416 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7418 void set_curr_task(int cpu, struct task_struct *p)
7425 #ifdef CONFIG_CGROUP_SCHED
7426 /* task_group_lock serializes the addition/removal of task groups */
7427 static DEFINE_SPINLOCK(task_group_lock);
7429 static void free_sched_group(struct task_group *tg)
7431 free_fair_sched_group(tg);
7432 free_rt_sched_group(tg);
7437 /* allocate runqueue etc for a new task group */
7438 struct task_group *sched_create_group(struct task_group *parent)
7440 struct task_group *tg;
7442 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7444 return ERR_PTR(-ENOMEM);
7446 if (!alloc_fair_sched_group(tg, parent))
7449 if (!alloc_rt_sched_group(tg, parent))
7455 free_sched_group(tg);
7456 return ERR_PTR(-ENOMEM);
7459 void sched_online_group(struct task_group *tg, struct task_group *parent)
7461 unsigned long flags;
7463 spin_lock_irqsave(&task_group_lock, flags);
7464 list_add_rcu(&tg->list, &task_groups);
7466 WARN_ON(!parent); /* root should already exist */
7468 tg->parent = parent;
7469 INIT_LIST_HEAD(&tg->children);
7470 list_add_rcu(&tg->siblings, &parent->children);
7471 spin_unlock_irqrestore(&task_group_lock, flags);
7474 /* rcu callback to free various structures associated with a task group */
7475 static void free_sched_group_rcu(struct rcu_head *rhp)
7477 /* now it should be safe to free those cfs_rqs */
7478 free_sched_group(container_of(rhp, struct task_group, rcu));
7481 /* Destroy runqueue etc associated with a task group */
7482 void sched_destroy_group(struct task_group *tg)
7484 /* wait for possible concurrent references to cfs_rqs complete */
7485 call_rcu(&tg->rcu, free_sched_group_rcu);
7488 void sched_offline_group(struct task_group *tg)
7490 unsigned long flags;
7493 /* end participation in shares distribution */
7494 for_each_possible_cpu(i)
7495 unregister_fair_sched_group(tg, i);
7497 spin_lock_irqsave(&task_group_lock, flags);
7498 list_del_rcu(&tg->list);
7499 list_del_rcu(&tg->siblings);
7500 spin_unlock_irqrestore(&task_group_lock, flags);
7503 /* change task's runqueue when it moves between groups.
7504 * The caller of this function should have put the task in its new group
7505 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7506 * reflect its new group.
7508 void sched_move_task(struct task_struct *tsk)
7510 struct task_group *tg;
7511 int queued, running;
7512 unsigned long flags;
7515 rq = task_rq_lock(tsk, &flags);
7517 running = task_current(rq, tsk);
7518 queued = task_on_rq_queued(tsk);
7521 dequeue_task(rq, tsk, 0);
7522 if (unlikely(running))
7523 put_prev_task(rq, tsk);
7525 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7526 lockdep_is_held(&tsk->sighand->siglock)),
7527 struct task_group, css);
7528 tg = autogroup_task_group(tsk, tg);
7529 tsk->sched_task_group = tg;
7531 #ifdef CONFIG_FAIR_GROUP_SCHED
7532 if (tsk->sched_class->task_move_group)
7533 tsk->sched_class->task_move_group(tsk, queued);
7536 set_task_rq(tsk, task_cpu(tsk));
7538 if (unlikely(running))
7539 tsk->sched_class->set_curr_task(rq);
7541 enqueue_task(rq, tsk, 0);
7543 task_rq_unlock(rq, tsk, &flags);
7545 #endif /* CONFIG_CGROUP_SCHED */
7547 #ifdef CONFIG_RT_GROUP_SCHED
7549 * Ensure that the real time constraints are schedulable.
7551 static DEFINE_MUTEX(rt_constraints_mutex);
7553 /* Must be called with tasklist_lock held */
7554 static inline int tg_has_rt_tasks(struct task_group *tg)
7556 struct task_struct *g, *p;
7558 for_each_process_thread(g, p) {
7559 if (rt_task(p) && task_group(p) == tg)
7566 struct rt_schedulable_data {
7567 struct task_group *tg;
7572 static int tg_rt_schedulable(struct task_group *tg, void *data)
7574 struct rt_schedulable_data *d = data;
7575 struct task_group *child;
7576 unsigned long total, sum = 0;
7577 u64 period, runtime;
7579 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7580 runtime = tg->rt_bandwidth.rt_runtime;
7583 period = d->rt_period;
7584 runtime = d->rt_runtime;
7588 * Cannot have more runtime than the period.
7590 if (runtime > period && runtime != RUNTIME_INF)
7594 * Ensure we don't starve existing RT tasks.
7596 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7599 total = to_ratio(period, runtime);
7602 * Nobody can have more than the global setting allows.
7604 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7608 * The sum of our children's runtime should not exceed our own.
7610 list_for_each_entry_rcu(child, &tg->children, siblings) {
7611 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7612 runtime = child->rt_bandwidth.rt_runtime;
7614 if (child == d->tg) {
7615 period = d->rt_period;
7616 runtime = d->rt_runtime;
7619 sum += to_ratio(period, runtime);
7628 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7632 struct rt_schedulable_data data = {
7634 .rt_period = period,
7635 .rt_runtime = runtime,
7639 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7645 static int tg_set_rt_bandwidth(struct task_group *tg,
7646 u64 rt_period, u64 rt_runtime)
7650 mutex_lock(&rt_constraints_mutex);
7651 read_lock(&tasklist_lock);
7652 err = __rt_schedulable(tg, rt_period, rt_runtime);
7656 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7657 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7658 tg->rt_bandwidth.rt_runtime = rt_runtime;
7660 for_each_possible_cpu(i) {
7661 struct rt_rq *rt_rq = tg->rt_rq[i];
7663 raw_spin_lock(&rt_rq->rt_runtime_lock);
7664 rt_rq->rt_runtime = rt_runtime;
7665 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7667 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7669 read_unlock(&tasklist_lock);
7670 mutex_unlock(&rt_constraints_mutex);
7675 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7677 u64 rt_runtime, rt_period;
7679 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7680 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7681 if (rt_runtime_us < 0)
7682 rt_runtime = RUNTIME_INF;
7684 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7687 static long sched_group_rt_runtime(struct task_group *tg)
7691 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7694 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7695 do_div(rt_runtime_us, NSEC_PER_USEC);
7696 return rt_runtime_us;
7699 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7701 u64 rt_runtime, rt_period;
7703 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7704 rt_runtime = tg->rt_bandwidth.rt_runtime;
7709 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7712 static long sched_group_rt_period(struct task_group *tg)
7716 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7717 do_div(rt_period_us, NSEC_PER_USEC);
7718 return rt_period_us;
7720 #endif /* CONFIG_RT_GROUP_SCHED */
7722 #ifdef CONFIG_RT_GROUP_SCHED
7723 static int sched_rt_global_constraints(void)
7727 mutex_lock(&rt_constraints_mutex);
7728 read_lock(&tasklist_lock);
7729 ret = __rt_schedulable(NULL, 0, 0);
7730 read_unlock(&tasklist_lock);
7731 mutex_unlock(&rt_constraints_mutex);
7736 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7738 /* Don't accept realtime tasks when there is no way for them to run */
7739 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7745 #else /* !CONFIG_RT_GROUP_SCHED */
7746 static int sched_rt_global_constraints(void)
7748 unsigned long flags;
7751 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7752 for_each_possible_cpu(i) {
7753 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7755 raw_spin_lock(&rt_rq->rt_runtime_lock);
7756 rt_rq->rt_runtime = global_rt_runtime();
7757 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7759 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7763 #endif /* CONFIG_RT_GROUP_SCHED */
7765 static int sched_dl_global_constraints(void)
7767 u64 runtime = global_rt_runtime();
7768 u64 period = global_rt_period();
7769 u64 new_bw = to_ratio(period, runtime);
7772 unsigned long flags;
7775 * Here we want to check the bandwidth not being set to some
7776 * value smaller than the currently allocated bandwidth in
7777 * any of the root_domains.
7779 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7780 * cycling on root_domains... Discussion on different/better
7781 * solutions is welcome!
7783 for_each_possible_cpu(cpu) {
7784 rcu_read_lock_sched();
7785 dl_b = dl_bw_of(cpu);
7787 raw_spin_lock_irqsave(&dl_b->lock, flags);
7788 if (new_bw < dl_b->total_bw)
7790 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7792 rcu_read_unlock_sched();
7801 static void sched_dl_do_global(void)
7806 unsigned long flags;
7808 def_dl_bandwidth.dl_period = global_rt_period();
7809 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7811 if (global_rt_runtime() != RUNTIME_INF)
7812 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7815 * FIXME: As above...
7817 for_each_possible_cpu(cpu) {
7818 rcu_read_lock_sched();
7819 dl_b = dl_bw_of(cpu);
7821 raw_spin_lock_irqsave(&dl_b->lock, flags);
7823 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7825 rcu_read_unlock_sched();
7829 static int sched_rt_global_validate(void)
7831 if (sysctl_sched_rt_period <= 0)
7834 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7835 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7841 static void sched_rt_do_global(void)
7843 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7844 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7847 int sched_rt_handler(struct ctl_table *table, int write,
7848 void __user *buffer, size_t *lenp,
7851 int old_period, old_runtime;
7852 static DEFINE_MUTEX(mutex);
7856 old_period = sysctl_sched_rt_period;
7857 old_runtime = sysctl_sched_rt_runtime;
7859 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7861 if (!ret && write) {
7862 ret = sched_rt_global_validate();
7866 ret = sched_rt_global_constraints();
7870 ret = sched_dl_global_constraints();
7874 sched_rt_do_global();
7875 sched_dl_do_global();
7879 sysctl_sched_rt_period = old_period;
7880 sysctl_sched_rt_runtime = old_runtime;
7882 mutex_unlock(&mutex);
7887 int sched_rr_handler(struct ctl_table *table, int write,
7888 void __user *buffer, size_t *lenp,
7892 static DEFINE_MUTEX(mutex);
7895 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7896 /* make sure that internally we keep jiffies */
7897 /* also, writing zero resets timeslice to default */
7898 if (!ret && write) {
7899 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7900 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7902 mutex_unlock(&mutex);
7906 #ifdef CONFIG_CGROUP_SCHED
7908 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7910 return css ? container_of(css, struct task_group, css) : NULL;
7913 static struct cgroup_subsys_state *
7914 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7916 struct task_group *parent = css_tg(parent_css);
7917 struct task_group *tg;
7920 /* This is early initialization for the top cgroup */
7921 return &root_task_group.css;
7924 tg = sched_create_group(parent);
7926 return ERR_PTR(-ENOMEM);
7931 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7933 struct task_group *tg = css_tg(css);
7934 struct task_group *parent = css_tg(css->parent);
7937 sched_online_group(tg, parent);
7941 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7943 struct task_group *tg = css_tg(css);
7945 sched_destroy_group(tg);
7948 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7950 struct task_group *tg = css_tg(css);
7952 sched_offline_group(tg);
7955 static void cpu_cgroup_fork(struct task_struct *task)
7957 sched_move_task(task);
7960 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7961 struct cgroup_taskset *tset)
7963 struct task_struct *task;
7965 cgroup_taskset_for_each(task, tset) {
7966 #ifdef CONFIG_RT_GROUP_SCHED
7967 if (!sched_rt_can_attach(css_tg(css), task))
7970 /* We don't support RT-tasks being in separate groups */
7971 if (task->sched_class != &fair_sched_class)
7978 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7979 struct cgroup_taskset *tset)
7981 struct task_struct *task;
7983 cgroup_taskset_for_each(task, tset)
7984 sched_move_task(task);
7987 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7988 struct cgroup_subsys_state *old_css,
7989 struct task_struct *task)
7992 * cgroup_exit() is called in the copy_process() failure path.
7993 * Ignore this case since the task hasn't ran yet, this avoids
7994 * trying to poke a half freed task state from generic code.
7996 if (!(task->flags & PF_EXITING))
7999 sched_move_task(task);
8002 #ifdef CONFIG_FAIR_GROUP_SCHED
8003 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8004 struct cftype *cftype, u64 shareval)
8006 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8009 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8012 struct task_group *tg = css_tg(css);
8014 return (u64) scale_load_down(tg->shares);
8017 #ifdef CONFIG_CFS_BANDWIDTH
8018 static DEFINE_MUTEX(cfs_constraints_mutex);
8020 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8021 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8023 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8025 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8027 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8028 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8030 if (tg == &root_task_group)
8034 * Ensure we have at some amount of bandwidth every period. This is
8035 * to prevent reaching a state of large arrears when throttled via
8036 * entity_tick() resulting in prolonged exit starvation.
8038 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8042 * Likewise, bound things on the otherside by preventing insane quota
8043 * periods. This also allows us to normalize in computing quota
8046 if (period > max_cfs_quota_period)
8050 * Prevent race between setting of cfs_rq->runtime_enabled and
8051 * unthrottle_offline_cfs_rqs().
8054 mutex_lock(&cfs_constraints_mutex);
8055 ret = __cfs_schedulable(tg, period, quota);
8059 runtime_enabled = quota != RUNTIME_INF;
8060 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8062 * If we need to toggle cfs_bandwidth_used, off->on must occur
8063 * before making related changes, and on->off must occur afterwards
8065 if (runtime_enabled && !runtime_was_enabled)
8066 cfs_bandwidth_usage_inc();
8067 raw_spin_lock_irq(&cfs_b->lock);
8068 cfs_b->period = ns_to_ktime(period);
8069 cfs_b->quota = quota;
8071 __refill_cfs_bandwidth_runtime(cfs_b);
8072 /* restart the period timer (if active) to handle new period expiry */
8073 if (runtime_enabled && cfs_b->timer_active) {
8074 /* force a reprogram */
8075 __start_cfs_bandwidth(cfs_b, true);
8077 raw_spin_unlock_irq(&cfs_b->lock);
8079 for_each_online_cpu(i) {
8080 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8081 struct rq *rq = cfs_rq->rq;
8083 raw_spin_lock_irq(&rq->lock);
8084 cfs_rq->runtime_enabled = runtime_enabled;
8085 cfs_rq->runtime_remaining = 0;
8087 if (cfs_rq->throttled)
8088 unthrottle_cfs_rq(cfs_rq);
8089 raw_spin_unlock_irq(&rq->lock);
8091 if (runtime_was_enabled && !runtime_enabled)
8092 cfs_bandwidth_usage_dec();
8094 mutex_unlock(&cfs_constraints_mutex);
8100 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8104 period = ktime_to_ns(tg->cfs_bandwidth.period);
8105 if (cfs_quota_us < 0)
8106 quota = RUNTIME_INF;
8108 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8110 return tg_set_cfs_bandwidth(tg, period, quota);
8113 long tg_get_cfs_quota(struct task_group *tg)
8117 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8120 quota_us = tg->cfs_bandwidth.quota;
8121 do_div(quota_us, NSEC_PER_USEC);
8126 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8130 period = (u64)cfs_period_us * NSEC_PER_USEC;
8131 quota = tg->cfs_bandwidth.quota;
8133 return tg_set_cfs_bandwidth(tg, period, quota);
8136 long tg_get_cfs_period(struct task_group *tg)
8140 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8141 do_div(cfs_period_us, NSEC_PER_USEC);
8143 return cfs_period_us;
8146 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8149 return tg_get_cfs_quota(css_tg(css));
8152 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8153 struct cftype *cftype, s64 cfs_quota_us)
8155 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8158 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8161 return tg_get_cfs_period(css_tg(css));
8164 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8165 struct cftype *cftype, u64 cfs_period_us)
8167 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8170 struct cfs_schedulable_data {
8171 struct task_group *tg;
8176 * normalize group quota/period to be quota/max_period
8177 * note: units are usecs
8179 static u64 normalize_cfs_quota(struct task_group *tg,
8180 struct cfs_schedulable_data *d)
8188 period = tg_get_cfs_period(tg);
8189 quota = tg_get_cfs_quota(tg);
8192 /* note: these should typically be equivalent */
8193 if (quota == RUNTIME_INF || quota == -1)
8196 return to_ratio(period, quota);
8199 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8201 struct cfs_schedulable_data *d = data;
8202 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8203 s64 quota = 0, parent_quota = -1;
8206 quota = RUNTIME_INF;
8208 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8210 quota = normalize_cfs_quota(tg, d);
8211 parent_quota = parent_b->hierarchical_quota;
8214 * ensure max(child_quota) <= parent_quota, inherit when no
8217 if (quota == RUNTIME_INF)
8218 quota = parent_quota;
8219 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8222 cfs_b->hierarchical_quota = quota;
8227 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8230 struct cfs_schedulable_data data = {
8236 if (quota != RUNTIME_INF) {
8237 do_div(data.period, NSEC_PER_USEC);
8238 do_div(data.quota, NSEC_PER_USEC);
8242 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8248 static int cpu_stats_show(struct seq_file *sf, void *v)
8250 struct task_group *tg = css_tg(seq_css(sf));
8251 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8253 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8254 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8255 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8259 #endif /* CONFIG_CFS_BANDWIDTH */
8260 #endif /* CONFIG_FAIR_GROUP_SCHED */
8262 #ifdef CONFIG_RT_GROUP_SCHED
8263 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8264 struct cftype *cft, s64 val)
8266 return sched_group_set_rt_runtime(css_tg(css), val);
8269 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8272 return sched_group_rt_runtime(css_tg(css));
8275 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8276 struct cftype *cftype, u64 rt_period_us)
8278 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8281 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8284 return sched_group_rt_period(css_tg(css));
8286 #endif /* CONFIG_RT_GROUP_SCHED */
8288 static struct cftype cpu_files[] = {
8289 #ifdef CONFIG_FAIR_GROUP_SCHED
8292 .read_u64 = cpu_shares_read_u64,
8293 .write_u64 = cpu_shares_write_u64,
8296 #ifdef CONFIG_CFS_BANDWIDTH
8298 .name = "cfs_quota_us",
8299 .read_s64 = cpu_cfs_quota_read_s64,
8300 .write_s64 = cpu_cfs_quota_write_s64,
8303 .name = "cfs_period_us",
8304 .read_u64 = cpu_cfs_period_read_u64,
8305 .write_u64 = cpu_cfs_period_write_u64,
8309 .seq_show = cpu_stats_show,
8312 #ifdef CONFIG_RT_GROUP_SCHED
8314 .name = "rt_runtime_us",
8315 .read_s64 = cpu_rt_runtime_read,
8316 .write_s64 = cpu_rt_runtime_write,
8319 .name = "rt_period_us",
8320 .read_u64 = cpu_rt_period_read_uint,
8321 .write_u64 = cpu_rt_period_write_uint,
8327 struct cgroup_subsys cpu_cgrp_subsys = {
8328 .css_alloc = cpu_cgroup_css_alloc,
8329 .css_free = cpu_cgroup_css_free,
8330 .css_online = cpu_cgroup_css_online,
8331 .css_offline = cpu_cgroup_css_offline,
8332 .fork = cpu_cgroup_fork,
8333 .can_attach = cpu_cgroup_can_attach,
8334 .attach = cpu_cgroup_attach,
8335 .exit = cpu_cgroup_exit,
8336 .legacy_cftypes = cpu_files,
8340 #endif /* CONFIG_CGROUP_SCHED */
8342 void dump_cpu_task(int cpu)
8344 pr_info("Task dump for CPU %d:\n", cpu);
8345 sched_show_task(cpu_curr(cpu));