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;
1012 * Can drop rq->lock because from sched_class::switched_from() methods drop it.
1014 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1015 const struct sched_class *prev_class,
1018 if (prev_class != p->sched_class) {
1019 if (prev_class->switched_from)
1020 prev_class->switched_from(rq, p);
1021 /* Possble rq->lock 'hole'. */
1022 p->sched_class->switched_to(rq, p);
1023 } else if (oldprio != p->prio || dl_task(p))
1024 p->sched_class->prio_changed(rq, p, oldprio);
1027 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1029 const struct sched_class *class;
1031 if (p->sched_class == rq->curr->sched_class) {
1032 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1034 for_each_class(class) {
1035 if (class == rq->curr->sched_class)
1037 if (class == p->sched_class) {
1045 * A queue event has occurred, and we're going to schedule. In
1046 * this case, we can save a useless back to back clock update.
1048 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1049 rq->skip_clock_update = 1;
1053 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1055 #ifdef CONFIG_SCHED_DEBUG
1057 * We should never call set_task_cpu() on a blocked task,
1058 * ttwu() will sort out the placement.
1060 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1063 #ifdef CONFIG_LOCKDEP
1065 * The caller should hold either p->pi_lock or rq->lock, when changing
1066 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1068 * sched_move_task() holds both and thus holding either pins the cgroup,
1071 * Furthermore, all task_rq users should acquire both locks, see
1074 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1075 lockdep_is_held(&task_rq(p)->lock)));
1079 trace_sched_migrate_task(p, new_cpu);
1081 if (task_cpu(p) != new_cpu) {
1082 if (p->sched_class->migrate_task_rq)
1083 p->sched_class->migrate_task_rq(p, new_cpu);
1084 p->se.nr_migrations++;
1085 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1088 __set_task_cpu(p, new_cpu);
1091 static void __migrate_swap_task(struct task_struct *p, int cpu)
1093 if (task_on_rq_queued(p)) {
1094 struct rq *src_rq, *dst_rq;
1096 src_rq = task_rq(p);
1097 dst_rq = cpu_rq(cpu);
1099 deactivate_task(src_rq, p, 0);
1100 set_task_cpu(p, cpu);
1101 activate_task(dst_rq, p, 0);
1102 check_preempt_curr(dst_rq, p, 0);
1105 * Task isn't running anymore; make it appear like we migrated
1106 * it before it went to sleep. This means on wakeup we make the
1107 * previous cpu our targer instead of where it really is.
1113 struct migration_swap_arg {
1114 struct task_struct *src_task, *dst_task;
1115 int src_cpu, dst_cpu;
1118 static int migrate_swap_stop(void *data)
1120 struct migration_swap_arg *arg = data;
1121 struct rq *src_rq, *dst_rq;
1124 src_rq = cpu_rq(arg->src_cpu);
1125 dst_rq = cpu_rq(arg->dst_cpu);
1127 double_raw_lock(&arg->src_task->pi_lock,
1128 &arg->dst_task->pi_lock);
1129 double_rq_lock(src_rq, dst_rq);
1130 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1133 if (task_cpu(arg->src_task) != arg->src_cpu)
1136 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1139 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1142 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1143 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1148 double_rq_unlock(src_rq, dst_rq);
1149 raw_spin_unlock(&arg->dst_task->pi_lock);
1150 raw_spin_unlock(&arg->src_task->pi_lock);
1156 * Cross migrate two tasks
1158 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1160 struct migration_swap_arg arg;
1163 arg = (struct migration_swap_arg){
1165 .src_cpu = task_cpu(cur),
1167 .dst_cpu = task_cpu(p),
1170 if (arg.src_cpu == arg.dst_cpu)
1174 * These three tests are all lockless; this is OK since all of them
1175 * will be re-checked with proper locks held further down the line.
1177 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1180 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1183 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1186 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1187 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1193 struct migration_arg {
1194 struct task_struct *task;
1198 static int migration_cpu_stop(void *data);
1201 * wait_task_inactive - wait for a thread to unschedule.
1203 * If @match_state is nonzero, it's the @p->state value just checked and
1204 * not expected to change. If it changes, i.e. @p might have woken up,
1205 * then return zero. When we succeed in waiting for @p to be off its CPU,
1206 * we return a positive number (its total switch count). If a second call
1207 * a short while later returns the same number, the caller can be sure that
1208 * @p has remained unscheduled the whole time.
1210 * The caller must ensure that the task *will* unschedule sometime soon,
1211 * else this function might spin for a *long* time. This function can't
1212 * be called with interrupts off, or it may introduce deadlock with
1213 * smp_call_function() if an IPI is sent by the same process we are
1214 * waiting to become inactive.
1216 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1218 unsigned long flags;
1219 int running, queued;
1225 * We do the initial early heuristics without holding
1226 * any task-queue locks at all. We'll only try to get
1227 * the runqueue lock when things look like they will
1233 * If the task is actively running on another CPU
1234 * still, just relax and busy-wait without holding
1237 * NOTE! Since we don't hold any locks, it's not
1238 * even sure that "rq" stays as the right runqueue!
1239 * But we don't care, since "task_running()" will
1240 * return false if the runqueue has changed and p
1241 * is actually now running somewhere else!
1243 while (task_running(rq, p)) {
1244 if (match_state && unlikely(p->state != match_state))
1250 * Ok, time to look more closely! We need the rq
1251 * lock now, to be *sure*. If we're wrong, we'll
1252 * just go back and repeat.
1254 rq = task_rq_lock(p, &flags);
1255 trace_sched_wait_task(p);
1256 running = task_running(rq, p);
1257 queued = task_on_rq_queued(p);
1259 if (!match_state || p->state == match_state)
1260 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1261 task_rq_unlock(rq, p, &flags);
1264 * If it changed from the expected state, bail out now.
1266 if (unlikely(!ncsw))
1270 * Was it really running after all now that we
1271 * checked with the proper locks actually held?
1273 * Oops. Go back and try again..
1275 if (unlikely(running)) {
1281 * It's not enough that it's not actively running,
1282 * it must be off the runqueue _entirely_, and not
1285 * So if it was still runnable (but just not actively
1286 * running right now), it's preempted, and we should
1287 * yield - it could be a while.
1289 if (unlikely(queued)) {
1290 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1292 set_current_state(TASK_UNINTERRUPTIBLE);
1293 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1298 * Ahh, all good. It wasn't running, and it wasn't
1299 * runnable, which means that it will never become
1300 * running in the future either. We're all done!
1309 * kick_process - kick a running thread to enter/exit the kernel
1310 * @p: the to-be-kicked thread
1312 * Cause a process which is running on another CPU to enter
1313 * kernel-mode, without any delay. (to get signals handled.)
1315 * NOTE: this function doesn't have to take the runqueue lock,
1316 * because all it wants to ensure is that the remote task enters
1317 * the kernel. If the IPI races and the task has been migrated
1318 * to another CPU then no harm is done and the purpose has been
1321 void kick_process(struct task_struct *p)
1327 if ((cpu != smp_processor_id()) && task_curr(p))
1328 smp_send_reschedule(cpu);
1331 EXPORT_SYMBOL_GPL(kick_process);
1332 #endif /* CONFIG_SMP */
1336 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1338 static int select_fallback_rq(int cpu, struct task_struct *p)
1340 int nid = cpu_to_node(cpu);
1341 const struct cpumask *nodemask = NULL;
1342 enum { cpuset, possible, fail } state = cpuset;
1346 * If the node that the cpu is on has been offlined, cpu_to_node()
1347 * will return -1. There is no cpu on the node, and we should
1348 * select the cpu on the other node.
1351 nodemask = cpumask_of_node(nid);
1353 /* Look for allowed, online CPU in same node. */
1354 for_each_cpu(dest_cpu, nodemask) {
1355 if (!cpu_online(dest_cpu))
1357 if (!cpu_active(dest_cpu))
1359 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1365 /* Any allowed, online CPU? */
1366 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1367 if (!cpu_online(dest_cpu))
1369 if (!cpu_active(dest_cpu))
1376 /* No more Mr. Nice Guy. */
1377 cpuset_cpus_allowed_fallback(p);
1382 do_set_cpus_allowed(p, cpu_possible_mask);
1393 if (state != cpuset) {
1395 * Don't tell them about moving exiting tasks or
1396 * kernel threads (both mm NULL), since they never
1399 if (p->mm && printk_ratelimit()) {
1400 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1401 task_pid_nr(p), p->comm, cpu);
1409 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1412 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1414 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1417 * In order not to call set_task_cpu() on a blocking task we need
1418 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1421 * Since this is common to all placement strategies, this lives here.
1423 * [ this allows ->select_task() to simply return task_cpu(p) and
1424 * not worry about this generic constraint ]
1426 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1428 cpu = select_fallback_rq(task_cpu(p), p);
1433 static void update_avg(u64 *avg, u64 sample)
1435 s64 diff = sample - *avg;
1441 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1443 #ifdef CONFIG_SCHEDSTATS
1444 struct rq *rq = this_rq();
1447 int this_cpu = smp_processor_id();
1449 if (cpu == this_cpu) {
1450 schedstat_inc(rq, ttwu_local);
1451 schedstat_inc(p, se.statistics.nr_wakeups_local);
1453 struct sched_domain *sd;
1455 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1457 for_each_domain(this_cpu, sd) {
1458 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1459 schedstat_inc(sd, ttwu_wake_remote);
1466 if (wake_flags & WF_MIGRATED)
1467 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1469 #endif /* CONFIG_SMP */
1471 schedstat_inc(rq, ttwu_count);
1472 schedstat_inc(p, se.statistics.nr_wakeups);
1474 if (wake_flags & WF_SYNC)
1475 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1477 #endif /* CONFIG_SCHEDSTATS */
1480 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1482 activate_task(rq, p, en_flags);
1483 p->on_rq = TASK_ON_RQ_QUEUED;
1485 /* if a worker is waking up, notify workqueue */
1486 if (p->flags & PF_WQ_WORKER)
1487 wq_worker_waking_up(p, cpu_of(rq));
1491 * Mark the task runnable and perform wakeup-preemption.
1494 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1496 check_preempt_curr(rq, p, wake_flags);
1497 trace_sched_wakeup(p, true);
1499 p->state = TASK_RUNNING;
1501 if (p->sched_class->task_woken)
1502 p->sched_class->task_woken(rq, p);
1504 if (rq->idle_stamp) {
1505 u64 delta = rq_clock(rq) - rq->idle_stamp;
1506 u64 max = 2*rq->max_idle_balance_cost;
1508 update_avg(&rq->avg_idle, delta);
1510 if (rq->avg_idle > max)
1519 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1522 if (p->sched_contributes_to_load)
1523 rq->nr_uninterruptible--;
1526 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1527 ttwu_do_wakeup(rq, p, wake_flags);
1531 * Called in case the task @p isn't fully descheduled from its runqueue,
1532 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1533 * since all we need to do is flip p->state to TASK_RUNNING, since
1534 * the task is still ->on_rq.
1536 static int ttwu_remote(struct task_struct *p, int wake_flags)
1541 rq = __task_rq_lock(p);
1542 if (task_on_rq_queued(p)) {
1543 /* check_preempt_curr() may use rq clock */
1544 update_rq_clock(rq);
1545 ttwu_do_wakeup(rq, p, wake_flags);
1548 __task_rq_unlock(rq);
1554 void sched_ttwu_pending(void)
1556 struct rq *rq = this_rq();
1557 struct llist_node *llist = llist_del_all(&rq->wake_list);
1558 struct task_struct *p;
1559 unsigned long flags;
1564 raw_spin_lock_irqsave(&rq->lock, flags);
1567 p = llist_entry(llist, struct task_struct, wake_entry);
1568 llist = llist_next(llist);
1569 ttwu_do_activate(rq, p, 0);
1572 raw_spin_unlock_irqrestore(&rq->lock, flags);
1575 void scheduler_ipi(void)
1578 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1579 * TIF_NEED_RESCHED remotely (for the first time) will also send
1582 preempt_fold_need_resched();
1584 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1588 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1589 * traditionally all their work was done from the interrupt return
1590 * path. Now that we actually do some work, we need to make sure
1593 * Some archs already do call them, luckily irq_enter/exit nest
1596 * Arguably we should visit all archs and update all handlers,
1597 * however a fair share of IPIs are still resched only so this would
1598 * somewhat pessimize the simple resched case.
1601 sched_ttwu_pending();
1604 * Check if someone kicked us for doing the nohz idle load balance.
1606 if (unlikely(got_nohz_idle_kick())) {
1607 this_rq()->idle_balance = 1;
1608 raise_softirq_irqoff(SCHED_SOFTIRQ);
1613 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1615 struct rq *rq = cpu_rq(cpu);
1617 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1618 if (!set_nr_if_polling(rq->idle))
1619 smp_send_reschedule(cpu);
1621 trace_sched_wake_idle_without_ipi(cpu);
1625 void wake_up_if_idle(int cpu)
1627 struct rq *rq = cpu_rq(cpu);
1628 unsigned long flags;
1630 if (!is_idle_task(rq->curr))
1633 if (set_nr_if_polling(rq->idle)) {
1634 trace_sched_wake_idle_without_ipi(cpu);
1636 raw_spin_lock_irqsave(&rq->lock, flags);
1637 if (is_idle_task(rq->curr))
1638 smp_send_reschedule(cpu);
1639 /* Else cpu is not in idle, do nothing here */
1640 raw_spin_unlock_irqrestore(&rq->lock, flags);
1644 bool cpus_share_cache(int this_cpu, int that_cpu)
1646 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1648 #endif /* CONFIG_SMP */
1650 static void ttwu_queue(struct task_struct *p, int cpu)
1652 struct rq *rq = cpu_rq(cpu);
1654 #if defined(CONFIG_SMP)
1655 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1656 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1657 ttwu_queue_remote(p, cpu);
1662 raw_spin_lock(&rq->lock);
1663 ttwu_do_activate(rq, p, 0);
1664 raw_spin_unlock(&rq->lock);
1668 * try_to_wake_up - wake up a thread
1669 * @p: the thread to be awakened
1670 * @state: the mask of task states that can be woken
1671 * @wake_flags: wake modifier flags (WF_*)
1673 * Put it on the run-queue if it's not already there. The "current"
1674 * thread is always on the run-queue (except when the actual
1675 * re-schedule is in progress), and as such you're allowed to do
1676 * the simpler "current->state = TASK_RUNNING" to mark yourself
1677 * runnable without the overhead of this.
1679 * Return: %true if @p was woken up, %false if it was already running.
1680 * or @state didn't match @p's state.
1683 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1685 unsigned long flags;
1686 int cpu, success = 0;
1689 * If we are going to wake up a thread waiting for CONDITION we
1690 * need to ensure that CONDITION=1 done by the caller can not be
1691 * reordered with p->state check below. This pairs with mb() in
1692 * set_current_state() the waiting thread does.
1694 smp_mb__before_spinlock();
1695 raw_spin_lock_irqsave(&p->pi_lock, flags);
1696 if (!(p->state & state))
1699 success = 1; /* we're going to change ->state */
1702 if (p->on_rq && ttwu_remote(p, wake_flags))
1707 * If the owning (remote) cpu is still in the middle of schedule() with
1708 * this task as prev, wait until its done referencing the task.
1713 * Pairs with the smp_wmb() in finish_lock_switch().
1717 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1718 p->state = TASK_WAKING;
1720 if (p->sched_class->task_waking)
1721 p->sched_class->task_waking(p);
1723 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1724 if (task_cpu(p) != cpu) {
1725 wake_flags |= WF_MIGRATED;
1726 set_task_cpu(p, cpu);
1728 #endif /* CONFIG_SMP */
1732 ttwu_stat(p, cpu, wake_flags);
1734 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1740 * try_to_wake_up_local - try to wake up a local task with rq lock held
1741 * @p: the thread to be awakened
1743 * Put @p on the run-queue if it's not already there. The caller must
1744 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1747 static void try_to_wake_up_local(struct task_struct *p)
1749 struct rq *rq = task_rq(p);
1751 if (WARN_ON_ONCE(rq != this_rq()) ||
1752 WARN_ON_ONCE(p == current))
1755 lockdep_assert_held(&rq->lock);
1757 if (!raw_spin_trylock(&p->pi_lock)) {
1758 raw_spin_unlock(&rq->lock);
1759 raw_spin_lock(&p->pi_lock);
1760 raw_spin_lock(&rq->lock);
1763 if (!(p->state & TASK_NORMAL))
1766 if (!task_on_rq_queued(p))
1767 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1769 ttwu_do_wakeup(rq, p, 0);
1770 ttwu_stat(p, smp_processor_id(), 0);
1772 raw_spin_unlock(&p->pi_lock);
1776 * wake_up_process - Wake up a specific process
1777 * @p: The process to be woken up.
1779 * Attempt to wake up the nominated process and move it to the set of runnable
1782 * Return: 1 if the process was woken up, 0 if it was already running.
1784 * It may be assumed that this function implies a write memory barrier before
1785 * changing the task state if and only if any tasks are woken up.
1787 int wake_up_process(struct task_struct *p)
1789 WARN_ON(task_is_stopped_or_traced(p));
1790 return try_to_wake_up(p, TASK_NORMAL, 0);
1792 EXPORT_SYMBOL(wake_up_process);
1794 int wake_up_state(struct task_struct *p, unsigned int state)
1796 return try_to_wake_up(p, state, 0);
1800 * This function clears the sched_dl_entity static params.
1802 void __dl_clear_params(struct task_struct *p)
1804 struct sched_dl_entity *dl_se = &p->dl;
1806 dl_se->dl_runtime = 0;
1807 dl_se->dl_deadline = 0;
1808 dl_se->dl_period = 0;
1814 * Perform scheduler related setup for a newly forked process p.
1815 * p is forked by current.
1817 * __sched_fork() is basic setup used by init_idle() too:
1819 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1824 p->se.exec_start = 0;
1825 p->se.sum_exec_runtime = 0;
1826 p->se.prev_sum_exec_runtime = 0;
1827 p->se.nr_migrations = 0;
1829 INIT_LIST_HEAD(&p->se.group_node);
1831 #ifdef CONFIG_SCHEDSTATS
1832 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1835 RB_CLEAR_NODE(&p->dl.rb_node);
1836 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1837 __dl_clear_params(p);
1839 INIT_LIST_HEAD(&p->rt.run_list);
1841 #ifdef CONFIG_PREEMPT_NOTIFIERS
1842 INIT_HLIST_HEAD(&p->preempt_notifiers);
1845 #ifdef CONFIG_NUMA_BALANCING
1846 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1847 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1848 p->mm->numa_scan_seq = 0;
1851 if (clone_flags & CLONE_VM)
1852 p->numa_preferred_nid = current->numa_preferred_nid;
1854 p->numa_preferred_nid = -1;
1856 p->node_stamp = 0ULL;
1857 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1858 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1859 p->numa_work.next = &p->numa_work;
1860 p->numa_faults_memory = NULL;
1861 p->numa_faults_buffer_memory = NULL;
1862 p->last_task_numa_placement = 0;
1863 p->last_sum_exec_runtime = 0;
1865 INIT_LIST_HEAD(&p->numa_entry);
1866 p->numa_group = NULL;
1867 #endif /* CONFIG_NUMA_BALANCING */
1870 #ifdef CONFIG_NUMA_BALANCING
1871 #ifdef CONFIG_SCHED_DEBUG
1872 void set_numabalancing_state(bool enabled)
1875 sched_feat_set("NUMA");
1877 sched_feat_set("NO_NUMA");
1880 __read_mostly bool numabalancing_enabled;
1882 void set_numabalancing_state(bool enabled)
1884 numabalancing_enabled = enabled;
1886 #endif /* CONFIG_SCHED_DEBUG */
1888 #ifdef CONFIG_PROC_SYSCTL
1889 int sysctl_numa_balancing(struct ctl_table *table, int write,
1890 void __user *buffer, size_t *lenp, loff_t *ppos)
1894 int state = numabalancing_enabled;
1896 if (write && !capable(CAP_SYS_ADMIN))
1901 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1905 set_numabalancing_state(state);
1912 * fork()/clone()-time setup:
1914 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1916 unsigned long flags;
1917 int cpu = get_cpu();
1919 __sched_fork(clone_flags, p);
1921 * We mark the process as running here. This guarantees that
1922 * nobody will actually run it, and a signal or other external
1923 * event cannot wake it up and insert it on the runqueue either.
1925 p->state = TASK_RUNNING;
1928 * Make sure we do not leak PI boosting priority to the child.
1930 p->prio = current->normal_prio;
1933 * Revert to default priority/policy on fork if requested.
1935 if (unlikely(p->sched_reset_on_fork)) {
1936 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1937 p->policy = SCHED_NORMAL;
1938 p->static_prio = NICE_TO_PRIO(0);
1940 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1941 p->static_prio = NICE_TO_PRIO(0);
1943 p->prio = p->normal_prio = __normal_prio(p);
1947 * We don't need the reset flag anymore after the fork. It has
1948 * fulfilled its duty:
1950 p->sched_reset_on_fork = 0;
1953 if (dl_prio(p->prio)) {
1956 } else if (rt_prio(p->prio)) {
1957 p->sched_class = &rt_sched_class;
1959 p->sched_class = &fair_sched_class;
1962 if (p->sched_class->task_fork)
1963 p->sched_class->task_fork(p);
1966 * The child is not yet in the pid-hash so no cgroup attach races,
1967 * and the cgroup is pinned to this child due to cgroup_fork()
1968 * is ran before sched_fork().
1970 * Silence PROVE_RCU.
1972 raw_spin_lock_irqsave(&p->pi_lock, flags);
1973 set_task_cpu(p, cpu);
1974 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1976 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1977 if (likely(sched_info_on()))
1978 memset(&p->sched_info, 0, sizeof(p->sched_info));
1980 #if defined(CONFIG_SMP)
1983 init_task_preempt_count(p);
1985 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1986 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1993 unsigned long to_ratio(u64 period, u64 runtime)
1995 if (runtime == RUNTIME_INF)
1999 * Doing this here saves a lot of checks in all
2000 * the calling paths, and returning zero seems
2001 * safe for them anyway.
2006 return div64_u64(runtime << 20, period);
2010 inline struct dl_bw *dl_bw_of(int i)
2012 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2013 "sched RCU must be held");
2014 return &cpu_rq(i)->rd->dl_bw;
2017 static inline int dl_bw_cpus(int i)
2019 struct root_domain *rd = cpu_rq(i)->rd;
2022 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2023 "sched RCU must be held");
2024 for_each_cpu_and(i, rd->span, cpu_active_mask)
2030 inline struct dl_bw *dl_bw_of(int i)
2032 return &cpu_rq(i)->dl.dl_bw;
2035 static inline int dl_bw_cpus(int i)
2042 * We must be sure that accepting a new task (or allowing changing the
2043 * parameters of an existing one) is consistent with the bandwidth
2044 * constraints. If yes, this function also accordingly updates the currently
2045 * allocated bandwidth to reflect the new situation.
2047 * This function is called while holding p's rq->lock.
2049 static int dl_overflow(struct task_struct *p, int policy,
2050 const struct sched_attr *attr)
2053 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2054 u64 period = attr->sched_period ?: attr->sched_deadline;
2055 u64 runtime = attr->sched_runtime;
2056 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2059 if (new_bw == p->dl.dl_bw)
2063 * Either if a task, enters, leave, or stays -deadline but changes
2064 * its parameters, we may need to update accordingly the total
2065 * allocated bandwidth of the container.
2067 raw_spin_lock(&dl_b->lock);
2068 cpus = dl_bw_cpus(task_cpu(p));
2069 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2070 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2071 __dl_add(dl_b, new_bw);
2073 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2074 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2075 __dl_clear(dl_b, p->dl.dl_bw);
2076 __dl_add(dl_b, new_bw);
2078 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2079 __dl_clear(dl_b, p->dl.dl_bw);
2082 raw_spin_unlock(&dl_b->lock);
2087 extern void init_dl_bw(struct dl_bw *dl_b);
2090 * wake_up_new_task - wake up a newly created task for the first time.
2092 * This function will do some initial scheduler statistics housekeeping
2093 * that must be done for every newly created context, then puts the task
2094 * on the runqueue and wakes it.
2096 void wake_up_new_task(struct task_struct *p)
2098 unsigned long flags;
2101 raw_spin_lock_irqsave(&p->pi_lock, flags);
2104 * Fork balancing, do it here and not earlier because:
2105 * - cpus_allowed can change in the fork path
2106 * - any previously selected cpu might disappear through hotplug
2108 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2111 /* Initialize new task's runnable average */
2112 init_task_runnable_average(p);
2113 rq = __task_rq_lock(p);
2114 activate_task(rq, p, 0);
2115 p->on_rq = TASK_ON_RQ_QUEUED;
2116 trace_sched_wakeup_new(p, true);
2117 check_preempt_curr(rq, p, WF_FORK);
2119 if (p->sched_class->task_woken)
2120 p->sched_class->task_woken(rq, p);
2122 task_rq_unlock(rq, p, &flags);
2125 #ifdef CONFIG_PREEMPT_NOTIFIERS
2128 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2129 * @notifier: notifier struct to register
2131 void preempt_notifier_register(struct preempt_notifier *notifier)
2133 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2135 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2138 * preempt_notifier_unregister - no longer interested in preemption notifications
2139 * @notifier: notifier struct to unregister
2141 * This is safe to call from within a preemption notifier.
2143 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2145 hlist_del(¬ifier->link);
2147 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2149 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2151 struct preempt_notifier *notifier;
2153 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2154 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2158 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2159 struct task_struct *next)
2161 struct preempt_notifier *notifier;
2163 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2164 notifier->ops->sched_out(notifier, next);
2167 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2169 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2174 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2175 struct task_struct *next)
2179 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2182 * prepare_task_switch - prepare to switch tasks
2183 * @rq: the runqueue preparing to switch
2184 * @prev: the current task that is being switched out
2185 * @next: the task we are going to switch to.
2187 * This is called with the rq lock held and interrupts off. It must
2188 * be paired with a subsequent finish_task_switch after the context
2191 * prepare_task_switch sets up locking and calls architecture specific
2195 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2196 struct task_struct *next)
2198 trace_sched_switch(prev, next);
2199 sched_info_switch(rq, prev, next);
2200 perf_event_task_sched_out(prev, next);
2201 fire_sched_out_preempt_notifiers(prev, next);
2202 prepare_lock_switch(rq, next);
2203 prepare_arch_switch(next);
2207 * finish_task_switch - clean up after a task-switch
2208 * @prev: the thread we just switched away from.
2210 * finish_task_switch must be called after the context switch, paired
2211 * with a prepare_task_switch call before the context switch.
2212 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2213 * and do any other architecture-specific cleanup actions.
2215 * Note that we may have delayed dropping an mm in context_switch(). If
2216 * so, we finish that here outside of the runqueue lock. (Doing it
2217 * with the lock held can cause deadlocks; see schedule() for
2220 * The context switch have flipped the stack from under us and restored the
2221 * local variables which were saved when this task called schedule() in the
2222 * past. prev == current is still correct but we need to recalculate this_rq
2223 * because prev may have moved to another CPU.
2225 static struct rq *finish_task_switch(struct task_struct *prev)
2226 __releases(rq->lock)
2228 struct rq *rq = this_rq();
2229 struct mm_struct *mm = rq->prev_mm;
2235 * A task struct has one reference for the use as "current".
2236 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2237 * schedule one last time. The schedule call will never return, and
2238 * the scheduled task must drop that reference.
2239 * The test for TASK_DEAD must occur while the runqueue locks are
2240 * still held, otherwise prev could be scheduled on another cpu, die
2241 * there before we look at prev->state, and then the reference would
2243 * Manfred Spraul <manfred@colorfullife.com>
2245 prev_state = prev->state;
2246 vtime_task_switch(prev);
2247 finish_arch_switch(prev);
2248 perf_event_task_sched_in(prev, current);
2249 finish_lock_switch(rq, prev);
2250 finish_arch_post_lock_switch();
2252 fire_sched_in_preempt_notifiers(current);
2255 if (unlikely(prev_state == TASK_DEAD)) {
2256 if (prev->sched_class->task_dead)
2257 prev->sched_class->task_dead(prev);
2260 * Remove function-return probe instances associated with this
2261 * task and put them back on the free list.
2263 kprobe_flush_task(prev);
2264 put_task_struct(prev);
2267 tick_nohz_task_switch(current);
2273 /* rq->lock is NOT held, but preemption is disabled */
2274 static inline void post_schedule(struct rq *rq)
2276 if (rq->post_schedule) {
2277 unsigned long flags;
2279 raw_spin_lock_irqsave(&rq->lock, flags);
2280 if (rq->curr->sched_class->post_schedule)
2281 rq->curr->sched_class->post_schedule(rq);
2282 raw_spin_unlock_irqrestore(&rq->lock, flags);
2284 rq->post_schedule = 0;
2290 static inline void post_schedule(struct rq *rq)
2297 * schedule_tail - first thing a freshly forked thread must call.
2298 * @prev: the thread we just switched away from.
2300 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2301 __releases(rq->lock)
2305 /* finish_task_switch() drops rq->lock and enables preemtion */
2307 rq = finish_task_switch(prev);
2311 if (current->set_child_tid)
2312 put_user(task_pid_vnr(current), current->set_child_tid);
2316 * context_switch - switch to the new MM and the new thread's register state.
2318 static inline struct rq *
2319 context_switch(struct rq *rq, struct task_struct *prev,
2320 struct task_struct *next)
2322 struct mm_struct *mm, *oldmm;
2324 prepare_task_switch(rq, prev, next);
2327 oldmm = prev->active_mm;
2329 * For paravirt, this is coupled with an exit in switch_to to
2330 * combine the page table reload and the switch backend into
2333 arch_start_context_switch(prev);
2336 next->active_mm = oldmm;
2337 atomic_inc(&oldmm->mm_count);
2338 enter_lazy_tlb(oldmm, next);
2340 switch_mm(oldmm, mm, next);
2343 prev->active_mm = NULL;
2344 rq->prev_mm = oldmm;
2347 * Since the runqueue lock will be released by the next
2348 * task (which is an invalid locking op but in the case
2349 * of the scheduler it's an obvious special-case), so we
2350 * do an early lockdep release here:
2352 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2354 context_tracking_task_switch(prev, next);
2355 /* Here we just switch the register state and the stack. */
2356 switch_to(prev, next, prev);
2359 return finish_task_switch(prev);
2363 * nr_running and nr_context_switches:
2365 * externally visible scheduler statistics: current number of runnable
2366 * threads, total number of context switches performed since bootup.
2368 unsigned long nr_running(void)
2370 unsigned long i, sum = 0;
2372 for_each_online_cpu(i)
2373 sum += cpu_rq(i)->nr_running;
2379 * Check if only the current task is running on the cpu.
2381 bool single_task_running(void)
2383 if (cpu_rq(smp_processor_id())->nr_running == 1)
2388 EXPORT_SYMBOL(single_task_running);
2390 unsigned long long nr_context_switches(void)
2393 unsigned long long sum = 0;
2395 for_each_possible_cpu(i)
2396 sum += cpu_rq(i)->nr_switches;
2401 unsigned long nr_iowait(void)
2403 unsigned long i, sum = 0;
2405 for_each_possible_cpu(i)
2406 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2411 unsigned long nr_iowait_cpu(int cpu)
2413 struct rq *this = cpu_rq(cpu);
2414 return atomic_read(&this->nr_iowait);
2417 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2419 struct rq *this = this_rq();
2420 *nr_waiters = atomic_read(&this->nr_iowait);
2421 *load = this->cpu_load[0];
2427 * sched_exec - execve() is a valuable balancing opportunity, because at
2428 * this point the task has the smallest effective memory and cache footprint.
2430 void sched_exec(void)
2432 struct task_struct *p = current;
2433 unsigned long flags;
2436 raw_spin_lock_irqsave(&p->pi_lock, flags);
2437 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2438 if (dest_cpu == smp_processor_id())
2441 if (likely(cpu_active(dest_cpu))) {
2442 struct migration_arg arg = { p, dest_cpu };
2444 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2445 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2449 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2454 DEFINE_PER_CPU(struct kernel_stat, kstat);
2455 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2457 EXPORT_PER_CPU_SYMBOL(kstat);
2458 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2461 * Return any ns on the sched_clock that have not yet been accounted in
2462 * @p in case that task is currently running.
2464 * Called with task_rq_lock() held on @rq.
2466 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2471 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2472 * project cycles that may never be accounted to this
2473 * thread, breaking clock_gettime().
2475 if (task_current(rq, p) && task_on_rq_queued(p)) {
2476 update_rq_clock(rq);
2477 ns = rq_clock_task(rq) - p->se.exec_start;
2485 unsigned long long task_delta_exec(struct task_struct *p)
2487 unsigned long flags;
2491 rq = task_rq_lock(p, &flags);
2492 ns = do_task_delta_exec(p, rq);
2493 task_rq_unlock(rq, p, &flags);
2499 * Return accounted runtime for the task.
2500 * In case the task is currently running, return the runtime plus current's
2501 * pending runtime that have not been accounted yet.
2503 unsigned long long task_sched_runtime(struct task_struct *p)
2505 unsigned long flags;
2509 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2511 * 64-bit doesn't need locks to atomically read a 64bit value.
2512 * So we have a optimization chance when the task's delta_exec is 0.
2513 * Reading ->on_cpu is racy, but this is ok.
2515 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2516 * If we race with it entering cpu, unaccounted time is 0. This is
2517 * indistinguishable from the read occurring a few cycles earlier.
2518 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2519 * been accounted, so we're correct here as well.
2521 if (!p->on_cpu || !task_on_rq_queued(p))
2522 return p->se.sum_exec_runtime;
2525 rq = task_rq_lock(p, &flags);
2526 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2527 task_rq_unlock(rq, p, &flags);
2533 * This function gets called by the timer code, with HZ frequency.
2534 * We call it with interrupts disabled.
2536 void scheduler_tick(void)
2538 int cpu = smp_processor_id();
2539 struct rq *rq = cpu_rq(cpu);
2540 struct task_struct *curr = rq->curr;
2544 raw_spin_lock(&rq->lock);
2545 update_rq_clock(rq);
2546 curr->sched_class->task_tick(rq, curr, 0);
2547 update_cpu_load_active(rq);
2548 raw_spin_unlock(&rq->lock);
2550 perf_event_task_tick();
2553 rq->idle_balance = idle_cpu(cpu);
2554 trigger_load_balance(rq);
2556 rq_last_tick_reset(rq);
2559 #ifdef CONFIG_NO_HZ_FULL
2561 * scheduler_tick_max_deferment
2563 * Keep at least one tick per second when a single
2564 * active task is running because the scheduler doesn't
2565 * yet completely support full dynticks environment.
2567 * This makes sure that uptime, CFS vruntime, load
2568 * balancing, etc... continue to move forward, even
2569 * with a very low granularity.
2571 * Return: Maximum deferment in nanoseconds.
2573 u64 scheduler_tick_max_deferment(void)
2575 struct rq *rq = this_rq();
2576 unsigned long next, now = ACCESS_ONCE(jiffies);
2578 next = rq->last_sched_tick + HZ;
2580 if (time_before_eq(next, now))
2583 return jiffies_to_nsecs(next - now);
2587 notrace unsigned long get_parent_ip(unsigned long addr)
2589 if (in_lock_functions(addr)) {
2590 addr = CALLER_ADDR2;
2591 if (in_lock_functions(addr))
2592 addr = CALLER_ADDR3;
2597 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2598 defined(CONFIG_PREEMPT_TRACER))
2600 void preempt_count_add(int val)
2602 #ifdef CONFIG_DEBUG_PREEMPT
2606 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2609 __preempt_count_add(val);
2610 #ifdef CONFIG_DEBUG_PREEMPT
2612 * Spinlock count overflowing soon?
2614 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2617 if (preempt_count() == val) {
2618 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2619 #ifdef CONFIG_DEBUG_PREEMPT
2620 current->preempt_disable_ip = ip;
2622 trace_preempt_off(CALLER_ADDR0, ip);
2625 EXPORT_SYMBOL(preempt_count_add);
2626 NOKPROBE_SYMBOL(preempt_count_add);
2628 void preempt_count_sub(int val)
2630 #ifdef CONFIG_DEBUG_PREEMPT
2634 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2637 * Is the spinlock portion underflowing?
2639 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2640 !(preempt_count() & PREEMPT_MASK)))
2644 if (preempt_count() == val)
2645 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2646 __preempt_count_sub(val);
2648 EXPORT_SYMBOL(preempt_count_sub);
2649 NOKPROBE_SYMBOL(preempt_count_sub);
2654 * Print scheduling while atomic bug:
2656 static noinline void __schedule_bug(struct task_struct *prev)
2658 if (oops_in_progress)
2661 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2662 prev->comm, prev->pid, preempt_count());
2664 debug_show_held_locks(prev);
2666 if (irqs_disabled())
2667 print_irqtrace_events(prev);
2668 #ifdef CONFIG_DEBUG_PREEMPT
2669 if (in_atomic_preempt_off()) {
2670 pr_err("Preemption disabled at:");
2671 print_ip_sym(current->preempt_disable_ip);
2676 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2680 * Various schedule()-time debugging checks and statistics:
2682 static inline void schedule_debug(struct task_struct *prev)
2684 #ifdef CONFIG_SCHED_STACK_END_CHECK
2685 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2688 * Test if we are atomic. Since do_exit() needs to call into
2689 * schedule() atomically, we ignore that path. Otherwise whine
2690 * if we are scheduling when we should not.
2692 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2693 __schedule_bug(prev);
2696 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2698 schedstat_inc(this_rq(), sched_count);
2702 * Pick up the highest-prio task:
2704 static inline struct task_struct *
2705 pick_next_task(struct rq *rq, struct task_struct *prev)
2707 const struct sched_class *class = &fair_sched_class;
2708 struct task_struct *p;
2711 * Optimization: we know that if all tasks are in
2712 * the fair class we can call that function directly:
2714 if (likely(prev->sched_class == class &&
2715 rq->nr_running == rq->cfs.h_nr_running)) {
2716 p = fair_sched_class.pick_next_task(rq, prev);
2717 if (unlikely(p == RETRY_TASK))
2720 /* assumes fair_sched_class->next == idle_sched_class */
2722 p = idle_sched_class.pick_next_task(rq, prev);
2728 for_each_class(class) {
2729 p = class->pick_next_task(rq, prev);
2731 if (unlikely(p == RETRY_TASK))
2737 BUG(); /* the idle class will always have a runnable task */
2741 * __schedule() is the main scheduler function.
2743 * The main means of driving the scheduler and thus entering this function are:
2745 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2747 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2748 * paths. For example, see arch/x86/entry_64.S.
2750 * To drive preemption between tasks, the scheduler sets the flag in timer
2751 * interrupt handler scheduler_tick().
2753 * 3. Wakeups don't really cause entry into schedule(). They add a
2754 * task to the run-queue and that's it.
2756 * Now, if the new task added to the run-queue preempts the current
2757 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2758 * called on the nearest possible occasion:
2760 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2762 * - in syscall or exception context, at the next outmost
2763 * preempt_enable(). (this might be as soon as the wake_up()'s
2766 * - in IRQ context, return from interrupt-handler to
2767 * preemptible context
2769 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2772 * - cond_resched() call
2773 * - explicit schedule() call
2774 * - return from syscall or exception to user-space
2775 * - return from interrupt-handler to user-space
2777 static void __sched __schedule(void)
2779 struct task_struct *prev, *next;
2780 unsigned long *switch_count;
2786 cpu = smp_processor_id();
2788 rcu_note_context_switch(cpu);
2791 schedule_debug(prev);
2793 if (sched_feat(HRTICK))
2797 * Make sure that signal_pending_state()->signal_pending() below
2798 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2799 * done by the caller to avoid the race with signal_wake_up().
2801 smp_mb__before_spinlock();
2802 raw_spin_lock_irq(&rq->lock);
2804 switch_count = &prev->nivcsw;
2805 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2806 if (unlikely(signal_pending_state(prev->state, prev))) {
2807 prev->state = TASK_RUNNING;
2809 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2813 * If a worker went to sleep, notify and ask workqueue
2814 * whether it wants to wake up a task to maintain
2817 if (prev->flags & PF_WQ_WORKER) {
2818 struct task_struct *to_wakeup;
2820 to_wakeup = wq_worker_sleeping(prev, cpu);
2822 try_to_wake_up_local(to_wakeup);
2825 switch_count = &prev->nvcsw;
2828 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2829 update_rq_clock(rq);
2831 next = pick_next_task(rq, prev);
2832 clear_tsk_need_resched(prev);
2833 clear_preempt_need_resched();
2834 rq->skip_clock_update = 0;
2836 if (likely(prev != next)) {
2841 rq = context_switch(rq, prev, next); /* unlocks the rq */
2844 raw_spin_unlock_irq(&rq->lock);
2848 sched_preempt_enable_no_resched();
2853 static inline void sched_submit_work(struct task_struct *tsk)
2855 if (!tsk->state || tsk_is_pi_blocked(tsk))
2858 * If we are going to sleep and we have plugged IO queued,
2859 * make sure to submit it to avoid deadlocks.
2861 if (blk_needs_flush_plug(tsk))
2862 blk_schedule_flush_plug(tsk);
2865 asmlinkage __visible void __sched schedule(void)
2867 struct task_struct *tsk = current;
2869 sched_submit_work(tsk);
2872 EXPORT_SYMBOL(schedule);
2874 #ifdef CONFIG_CONTEXT_TRACKING
2875 asmlinkage __visible void __sched schedule_user(void)
2878 * If we come here after a random call to set_need_resched(),
2879 * or we have been woken up remotely but the IPI has not yet arrived,
2880 * we haven't yet exited the RCU idle mode. Do it here manually until
2881 * we find a better solution.
2890 * schedule_preempt_disabled - called with preemption disabled
2892 * Returns with preemption disabled. Note: preempt_count must be 1
2894 void __sched schedule_preempt_disabled(void)
2896 sched_preempt_enable_no_resched();
2901 #ifdef CONFIG_PREEMPT
2903 * this is the entry point to schedule() from in-kernel preemption
2904 * off of preempt_enable. Kernel preemptions off return from interrupt
2905 * occur there and call schedule directly.
2907 asmlinkage __visible void __sched notrace preempt_schedule(void)
2910 * If there is a non-zero preempt_count or interrupts are disabled,
2911 * we do not want to preempt the current task. Just return..
2913 if (likely(!preemptible()))
2917 __preempt_count_add(PREEMPT_ACTIVE);
2919 __preempt_count_sub(PREEMPT_ACTIVE);
2922 * Check again in case we missed a preemption opportunity
2923 * between schedule and now.
2926 } while (need_resched());
2928 NOKPROBE_SYMBOL(preempt_schedule);
2929 EXPORT_SYMBOL(preempt_schedule);
2931 #ifdef CONFIG_CONTEXT_TRACKING
2933 * preempt_schedule_context - preempt_schedule called by tracing
2935 * The tracing infrastructure uses preempt_enable_notrace to prevent
2936 * recursion and tracing preempt enabling caused by the tracing
2937 * infrastructure itself. But as tracing can happen in areas coming
2938 * from userspace or just about to enter userspace, a preempt enable
2939 * can occur before user_exit() is called. This will cause the scheduler
2940 * to be called when the system is still in usermode.
2942 * To prevent this, the preempt_enable_notrace will use this function
2943 * instead of preempt_schedule() to exit user context if needed before
2944 * calling the scheduler.
2946 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2948 enum ctx_state prev_ctx;
2950 if (likely(!preemptible()))
2954 __preempt_count_add(PREEMPT_ACTIVE);
2956 * Needs preempt disabled in case user_exit() is traced
2957 * and the tracer calls preempt_enable_notrace() causing
2958 * an infinite recursion.
2960 prev_ctx = exception_enter();
2962 exception_exit(prev_ctx);
2964 __preempt_count_sub(PREEMPT_ACTIVE);
2966 } while (need_resched());
2968 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2969 #endif /* CONFIG_CONTEXT_TRACKING */
2971 #endif /* CONFIG_PREEMPT */
2974 * this is the entry point to schedule() from kernel preemption
2975 * off of irq context.
2976 * Note, that this is called and return with irqs disabled. This will
2977 * protect us against recursive calling from irq.
2979 asmlinkage __visible void __sched preempt_schedule_irq(void)
2981 enum ctx_state prev_state;
2983 /* Catch callers which need to be fixed */
2984 BUG_ON(preempt_count() || !irqs_disabled());
2986 prev_state = exception_enter();
2989 __preempt_count_add(PREEMPT_ACTIVE);
2992 local_irq_disable();
2993 __preempt_count_sub(PREEMPT_ACTIVE);
2996 * Check again in case we missed a preemption opportunity
2997 * between schedule and now.
3000 } while (need_resched());
3002 exception_exit(prev_state);
3005 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3008 return try_to_wake_up(curr->private, mode, wake_flags);
3010 EXPORT_SYMBOL(default_wake_function);
3012 #ifdef CONFIG_RT_MUTEXES
3015 * rt_mutex_setprio - set the current priority of a task
3017 * @prio: prio value (kernel-internal form)
3019 * This function changes the 'effective' priority of a task. It does
3020 * not touch ->normal_prio like __setscheduler().
3022 * Used by the rt_mutex code to implement priority inheritance
3023 * logic. Call site only calls if the priority of the task changed.
3025 void rt_mutex_setprio(struct task_struct *p, int prio)
3027 int oldprio, queued, running, enqueue_flag = 0;
3029 const struct sched_class *prev_class;
3031 BUG_ON(prio > MAX_PRIO);
3033 rq = __task_rq_lock(p);
3036 * Idle task boosting is a nono in general. There is one
3037 * exception, when PREEMPT_RT and NOHZ is active:
3039 * The idle task calls get_next_timer_interrupt() and holds
3040 * the timer wheel base->lock on the CPU and another CPU wants
3041 * to access the timer (probably to cancel it). We can safely
3042 * ignore the boosting request, as the idle CPU runs this code
3043 * with interrupts disabled and will complete the lock
3044 * protected section without being interrupted. So there is no
3045 * real need to boost.
3047 if (unlikely(p == rq->idle)) {
3048 WARN_ON(p != rq->curr);
3049 WARN_ON(p->pi_blocked_on);
3053 trace_sched_pi_setprio(p, prio);
3055 prev_class = p->sched_class;
3056 queued = task_on_rq_queued(p);
3057 running = task_current(rq, p);
3059 dequeue_task(rq, p, 0);
3061 put_prev_task(rq, p);
3064 * Boosting condition are:
3065 * 1. -rt task is running and holds mutex A
3066 * --> -dl task blocks on mutex A
3068 * 2. -dl task is running and holds mutex A
3069 * --> -dl task blocks on mutex A and could preempt the
3072 if (dl_prio(prio)) {
3073 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3074 if (!dl_prio(p->normal_prio) ||
3075 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3076 p->dl.dl_boosted = 1;
3077 p->dl.dl_throttled = 0;
3078 enqueue_flag = ENQUEUE_REPLENISH;
3080 p->dl.dl_boosted = 0;
3081 p->sched_class = &dl_sched_class;
3082 } else if (rt_prio(prio)) {
3083 if (dl_prio(oldprio))
3084 p->dl.dl_boosted = 0;
3086 enqueue_flag = ENQUEUE_HEAD;
3087 p->sched_class = &rt_sched_class;
3089 if (dl_prio(oldprio))
3090 p->dl.dl_boosted = 0;
3091 p->sched_class = &fair_sched_class;
3097 p->sched_class->set_curr_task(rq);
3099 enqueue_task(rq, p, enqueue_flag);
3101 check_class_changed(rq, p, prev_class, oldprio);
3103 __task_rq_unlock(rq);
3107 void set_user_nice(struct task_struct *p, long nice)
3109 int old_prio, delta, queued;
3110 unsigned long flags;
3113 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3116 * We have to be careful, if called from sys_setpriority(),
3117 * the task might be in the middle of scheduling on another CPU.
3119 rq = task_rq_lock(p, &flags);
3121 * The RT priorities are set via sched_setscheduler(), but we still
3122 * allow the 'normal' nice value to be set - but as expected
3123 * it wont have any effect on scheduling until the task is
3124 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3126 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3127 p->static_prio = NICE_TO_PRIO(nice);
3130 queued = task_on_rq_queued(p);
3132 dequeue_task(rq, p, 0);
3134 p->static_prio = NICE_TO_PRIO(nice);
3137 p->prio = effective_prio(p);
3138 delta = p->prio - old_prio;
3141 enqueue_task(rq, p, 0);
3143 * If the task increased its priority or is running and
3144 * lowered its priority, then reschedule its CPU:
3146 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3150 task_rq_unlock(rq, p, &flags);
3152 EXPORT_SYMBOL(set_user_nice);
3155 * can_nice - check if a task can reduce its nice value
3159 int can_nice(const struct task_struct *p, const int nice)
3161 /* convert nice value [19,-20] to rlimit style value [1,40] */
3162 int nice_rlim = nice_to_rlimit(nice);
3164 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3165 capable(CAP_SYS_NICE));
3168 #ifdef __ARCH_WANT_SYS_NICE
3171 * sys_nice - change the priority of the current process.
3172 * @increment: priority increment
3174 * sys_setpriority is a more generic, but much slower function that
3175 * does similar things.
3177 SYSCALL_DEFINE1(nice, int, increment)
3182 * Setpriority might change our priority at the same moment.
3183 * We don't have to worry. Conceptually one call occurs first
3184 * and we have a single winner.
3186 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3187 nice = task_nice(current) + increment;
3189 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3190 if (increment < 0 && !can_nice(current, nice))
3193 retval = security_task_setnice(current, nice);
3197 set_user_nice(current, nice);
3204 * task_prio - return the priority value of a given task.
3205 * @p: the task in question.
3207 * Return: The priority value as seen by users in /proc.
3208 * RT tasks are offset by -200. Normal tasks are centered
3209 * around 0, value goes from -16 to +15.
3211 int task_prio(const struct task_struct *p)
3213 return p->prio - MAX_RT_PRIO;
3217 * idle_cpu - is a given cpu idle currently?
3218 * @cpu: the processor in question.
3220 * Return: 1 if the CPU is currently idle. 0 otherwise.
3222 int idle_cpu(int cpu)
3224 struct rq *rq = cpu_rq(cpu);
3226 if (rq->curr != rq->idle)
3233 if (!llist_empty(&rq->wake_list))
3241 * idle_task - return the idle task for a given cpu.
3242 * @cpu: the processor in question.
3244 * Return: The idle task for the cpu @cpu.
3246 struct task_struct *idle_task(int cpu)
3248 return cpu_rq(cpu)->idle;
3252 * find_process_by_pid - find a process with a matching PID value.
3253 * @pid: the pid in question.
3255 * The task of @pid, if found. %NULL otherwise.
3257 static struct task_struct *find_process_by_pid(pid_t pid)
3259 return pid ? find_task_by_vpid(pid) : current;
3263 * This function initializes the sched_dl_entity of a newly becoming
3264 * SCHED_DEADLINE task.
3266 * Only the static values are considered here, the actual runtime and the
3267 * absolute deadline will be properly calculated when the task is enqueued
3268 * for the first time with its new policy.
3271 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3273 struct sched_dl_entity *dl_se = &p->dl;
3275 init_dl_task_timer(dl_se);
3276 dl_se->dl_runtime = attr->sched_runtime;
3277 dl_se->dl_deadline = attr->sched_deadline;
3278 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3279 dl_se->flags = attr->sched_flags;
3280 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3281 dl_se->dl_throttled = 0;
3283 dl_se->dl_yielded = 0;
3287 * sched_setparam() passes in -1 for its policy, to let the functions
3288 * it calls know not to change it.
3290 #define SETPARAM_POLICY -1
3292 static void __setscheduler_params(struct task_struct *p,
3293 const struct sched_attr *attr)
3295 int policy = attr->sched_policy;
3297 if (policy == SETPARAM_POLICY)
3302 if (dl_policy(policy))
3303 __setparam_dl(p, attr);
3304 else if (fair_policy(policy))
3305 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3308 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3309 * !rt_policy. Always setting this ensures that things like
3310 * getparam()/getattr() don't report silly values for !rt tasks.
3312 p->rt_priority = attr->sched_priority;
3313 p->normal_prio = normal_prio(p);
3317 /* Actually do priority change: must hold pi & rq lock. */
3318 static void __setscheduler(struct rq *rq, struct task_struct *p,
3319 const struct sched_attr *attr)
3321 __setscheduler_params(p, attr);
3324 * If we get here, there was no pi waiters boosting the
3325 * task. It is safe to use the normal prio.
3327 p->prio = normal_prio(p);
3329 if (dl_prio(p->prio))
3330 p->sched_class = &dl_sched_class;
3331 else if (rt_prio(p->prio))
3332 p->sched_class = &rt_sched_class;
3334 p->sched_class = &fair_sched_class;
3338 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3340 struct sched_dl_entity *dl_se = &p->dl;
3342 attr->sched_priority = p->rt_priority;
3343 attr->sched_runtime = dl_se->dl_runtime;
3344 attr->sched_deadline = dl_se->dl_deadline;
3345 attr->sched_period = dl_se->dl_period;
3346 attr->sched_flags = dl_se->flags;
3350 * This function validates the new parameters of a -deadline task.
3351 * We ask for the deadline not being zero, and greater or equal
3352 * than the runtime, as well as the period of being zero or
3353 * greater than deadline. Furthermore, we have to be sure that
3354 * user parameters are above the internal resolution of 1us (we
3355 * check sched_runtime only since it is always the smaller one) and
3356 * below 2^63 ns (we have to check both sched_deadline and
3357 * sched_period, as the latter can be zero).
3360 __checkparam_dl(const struct sched_attr *attr)
3363 if (attr->sched_deadline == 0)
3367 * Since we truncate DL_SCALE bits, make sure we're at least
3370 if (attr->sched_runtime < (1ULL << DL_SCALE))
3374 * Since we use the MSB for wrap-around and sign issues, make
3375 * sure it's not set (mind that period can be equal to zero).
3377 if (attr->sched_deadline & (1ULL << 63) ||
3378 attr->sched_period & (1ULL << 63))
3381 /* runtime <= deadline <= period (if period != 0) */
3382 if ((attr->sched_period != 0 &&
3383 attr->sched_period < attr->sched_deadline) ||
3384 attr->sched_deadline < attr->sched_runtime)
3391 * check the target process has a UID that matches the current process's
3393 static bool check_same_owner(struct task_struct *p)
3395 const struct cred *cred = current_cred(), *pcred;
3399 pcred = __task_cred(p);
3400 match = (uid_eq(cred->euid, pcred->euid) ||
3401 uid_eq(cred->euid, pcred->uid));
3406 static int __sched_setscheduler(struct task_struct *p,
3407 const struct sched_attr *attr,
3410 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3411 MAX_RT_PRIO - 1 - attr->sched_priority;
3412 int retval, oldprio, oldpolicy = -1, queued, running;
3413 int policy = attr->sched_policy;
3414 unsigned long flags;
3415 const struct sched_class *prev_class;
3419 /* may grab non-irq protected spin_locks */
3420 BUG_ON(in_interrupt());
3422 /* double check policy once rq lock held */
3424 reset_on_fork = p->sched_reset_on_fork;
3425 policy = oldpolicy = p->policy;
3427 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3429 if (policy != SCHED_DEADLINE &&
3430 policy != SCHED_FIFO && policy != SCHED_RR &&
3431 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3432 policy != SCHED_IDLE)
3436 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3440 * Valid priorities for SCHED_FIFO and SCHED_RR are
3441 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3442 * SCHED_BATCH and SCHED_IDLE is 0.
3444 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3445 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3447 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3448 (rt_policy(policy) != (attr->sched_priority != 0)))
3452 * Allow unprivileged RT tasks to decrease priority:
3454 if (user && !capable(CAP_SYS_NICE)) {
3455 if (fair_policy(policy)) {
3456 if (attr->sched_nice < task_nice(p) &&
3457 !can_nice(p, attr->sched_nice))
3461 if (rt_policy(policy)) {
3462 unsigned long rlim_rtprio =
3463 task_rlimit(p, RLIMIT_RTPRIO);
3465 /* can't set/change the rt policy */
3466 if (policy != p->policy && !rlim_rtprio)
3469 /* can't increase priority */
3470 if (attr->sched_priority > p->rt_priority &&
3471 attr->sched_priority > rlim_rtprio)
3476 * Can't set/change SCHED_DEADLINE policy at all for now
3477 * (safest behavior); in the future we would like to allow
3478 * unprivileged DL tasks to increase their relative deadline
3479 * or reduce their runtime (both ways reducing utilization)
3481 if (dl_policy(policy))
3485 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3486 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3488 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3489 if (!can_nice(p, task_nice(p)))
3493 /* can't change other user's priorities */
3494 if (!check_same_owner(p))
3497 /* Normal users shall not reset the sched_reset_on_fork flag */
3498 if (p->sched_reset_on_fork && !reset_on_fork)
3503 retval = security_task_setscheduler(p);
3509 * make sure no PI-waiters arrive (or leave) while we are
3510 * changing the priority of the task:
3512 * To be able to change p->policy safely, the appropriate
3513 * runqueue lock must be held.
3515 rq = task_rq_lock(p, &flags);
3518 * Changing the policy of the stop threads its a very bad idea
3520 if (p == rq->stop) {
3521 task_rq_unlock(rq, p, &flags);
3526 * If not changing anything there's no need to proceed further,
3527 * but store a possible modification of reset_on_fork.
3529 if (unlikely(policy == p->policy)) {
3530 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3532 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3534 if (dl_policy(policy))
3537 p->sched_reset_on_fork = reset_on_fork;
3538 task_rq_unlock(rq, p, &flags);
3544 #ifdef CONFIG_RT_GROUP_SCHED
3546 * Do not allow realtime tasks into groups that have no runtime
3549 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3550 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3551 !task_group_is_autogroup(task_group(p))) {
3552 task_rq_unlock(rq, p, &flags);
3557 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3558 cpumask_t *span = rq->rd->span;
3561 * Don't allow tasks with an affinity mask smaller than
3562 * the entire root_domain to become SCHED_DEADLINE. We
3563 * will also fail if there's no bandwidth available.
3565 if (!cpumask_subset(span, &p->cpus_allowed) ||
3566 rq->rd->dl_bw.bw == 0) {
3567 task_rq_unlock(rq, p, &flags);
3574 /* recheck policy now with rq lock held */
3575 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3576 policy = oldpolicy = -1;
3577 task_rq_unlock(rq, p, &flags);
3582 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3583 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3586 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3587 task_rq_unlock(rq, p, &flags);
3591 p->sched_reset_on_fork = reset_on_fork;
3595 * Special case for priority boosted tasks.
3597 * If the new priority is lower or equal (user space view)
3598 * than the current (boosted) priority, we just store the new
3599 * normal parameters and do not touch the scheduler class and
3600 * the runqueue. This will be done when the task deboost
3603 if (rt_mutex_check_prio(p, newprio)) {
3604 __setscheduler_params(p, attr);
3605 task_rq_unlock(rq, p, &flags);
3609 queued = task_on_rq_queued(p);
3610 running = task_current(rq, p);
3612 dequeue_task(rq, p, 0);
3614 put_prev_task(rq, p);
3616 prev_class = p->sched_class;
3617 __setscheduler(rq, p, attr);
3620 p->sched_class->set_curr_task(rq);
3623 * We enqueue to tail when the priority of a task is
3624 * increased (user space view).
3626 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3629 check_class_changed(rq, p, prev_class, oldprio);
3630 task_rq_unlock(rq, p, &flags);
3632 rt_mutex_adjust_pi(p);
3637 static int _sched_setscheduler(struct task_struct *p, int policy,
3638 const struct sched_param *param, bool check)
3640 struct sched_attr attr = {
3641 .sched_policy = policy,
3642 .sched_priority = param->sched_priority,
3643 .sched_nice = PRIO_TO_NICE(p->static_prio),
3646 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3647 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3648 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3649 policy &= ~SCHED_RESET_ON_FORK;
3650 attr.sched_policy = policy;
3653 return __sched_setscheduler(p, &attr, check);
3656 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3657 * @p: the task in question.
3658 * @policy: new policy.
3659 * @param: structure containing the new RT priority.
3661 * Return: 0 on success. An error code otherwise.
3663 * NOTE that the task may be already dead.
3665 int sched_setscheduler(struct task_struct *p, int policy,
3666 const struct sched_param *param)
3668 return _sched_setscheduler(p, policy, param, true);
3670 EXPORT_SYMBOL_GPL(sched_setscheduler);
3672 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3674 return __sched_setscheduler(p, attr, true);
3676 EXPORT_SYMBOL_GPL(sched_setattr);
3679 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3680 * @p: the task in question.
3681 * @policy: new policy.
3682 * @param: structure containing the new RT priority.
3684 * Just like sched_setscheduler, only don't bother checking if the
3685 * current context has permission. For example, this is needed in
3686 * stop_machine(): we create temporary high priority worker threads,
3687 * but our caller might not have that capability.
3689 * Return: 0 on success. An error code otherwise.
3691 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3692 const struct sched_param *param)
3694 return _sched_setscheduler(p, policy, param, false);
3698 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3700 struct sched_param lparam;
3701 struct task_struct *p;
3704 if (!param || pid < 0)
3706 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3711 p = find_process_by_pid(pid);
3713 retval = sched_setscheduler(p, policy, &lparam);
3720 * Mimics kernel/events/core.c perf_copy_attr().
3722 static int sched_copy_attr(struct sched_attr __user *uattr,
3723 struct sched_attr *attr)
3728 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3732 * zero the full structure, so that a short copy will be nice.
3734 memset(attr, 0, sizeof(*attr));
3736 ret = get_user(size, &uattr->size);
3740 if (size > PAGE_SIZE) /* silly large */
3743 if (!size) /* abi compat */
3744 size = SCHED_ATTR_SIZE_VER0;
3746 if (size < SCHED_ATTR_SIZE_VER0)
3750 * If we're handed a bigger struct than we know of,
3751 * ensure all the unknown bits are 0 - i.e. new
3752 * user-space does not rely on any kernel feature
3753 * extensions we dont know about yet.
3755 if (size > sizeof(*attr)) {
3756 unsigned char __user *addr;
3757 unsigned char __user *end;
3760 addr = (void __user *)uattr + sizeof(*attr);
3761 end = (void __user *)uattr + size;
3763 for (; addr < end; addr++) {
3764 ret = get_user(val, addr);
3770 size = sizeof(*attr);
3773 ret = copy_from_user(attr, uattr, size);
3778 * XXX: do we want to be lenient like existing syscalls; or do we want
3779 * to be strict and return an error on out-of-bounds values?
3781 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3786 put_user(sizeof(*attr), &uattr->size);
3791 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3792 * @pid: the pid in question.
3793 * @policy: new policy.
3794 * @param: structure containing the new RT priority.
3796 * Return: 0 on success. An error code otherwise.
3798 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3799 struct sched_param __user *, param)
3801 /* negative values for policy are not valid */
3805 return do_sched_setscheduler(pid, policy, param);
3809 * sys_sched_setparam - set/change the RT priority of a thread
3810 * @pid: the pid in question.
3811 * @param: structure containing the new RT priority.
3813 * Return: 0 on success. An error code otherwise.
3815 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3817 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3821 * sys_sched_setattr - same as above, but with extended sched_attr
3822 * @pid: the pid in question.
3823 * @uattr: structure containing the extended parameters.
3824 * @flags: for future extension.
3826 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3827 unsigned int, flags)
3829 struct sched_attr attr;
3830 struct task_struct *p;
3833 if (!uattr || pid < 0 || flags)
3836 retval = sched_copy_attr(uattr, &attr);
3840 if ((int)attr.sched_policy < 0)
3845 p = find_process_by_pid(pid);
3847 retval = sched_setattr(p, &attr);
3854 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3855 * @pid: the pid in question.
3857 * Return: On success, the policy of the thread. Otherwise, a negative error
3860 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3862 struct task_struct *p;
3870 p = find_process_by_pid(pid);
3872 retval = security_task_getscheduler(p);
3875 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3882 * sys_sched_getparam - get the RT priority of a thread
3883 * @pid: the pid in question.
3884 * @param: structure containing the RT priority.
3886 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3889 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3891 struct sched_param lp = { .sched_priority = 0 };
3892 struct task_struct *p;
3895 if (!param || pid < 0)
3899 p = find_process_by_pid(pid);
3904 retval = security_task_getscheduler(p);
3908 if (task_has_rt_policy(p))
3909 lp.sched_priority = p->rt_priority;
3913 * This one might sleep, we cannot do it with a spinlock held ...
3915 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3924 static int sched_read_attr(struct sched_attr __user *uattr,
3925 struct sched_attr *attr,
3930 if (!access_ok(VERIFY_WRITE, uattr, usize))
3934 * If we're handed a smaller struct than we know of,
3935 * ensure all the unknown bits are 0 - i.e. old
3936 * user-space does not get uncomplete information.
3938 if (usize < sizeof(*attr)) {
3939 unsigned char *addr;
3942 addr = (void *)attr + usize;
3943 end = (void *)attr + sizeof(*attr);
3945 for (; addr < end; addr++) {
3953 ret = copy_to_user(uattr, attr, attr->size);
3961 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3962 * @pid: the pid in question.
3963 * @uattr: structure containing the extended parameters.
3964 * @size: sizeof(attr) for fwd/bwd comp.
3965 * @flags: for future extension.
3967 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3968 unsigned int, size, unsigned int, flags)
3970 struct sched_attr attr = {
3971 .size = sizeof(struct sched_attr),
3973 struct task_struct *p;
3976 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3977 size < SCHED_ATTR_SIZE_VER0 || flags)
3981 p = find_process_by_pid(pid);
3986 retval = security_task_getscheduler(p);
3990 attr.sched_policy = p->policy;
3991 if (p->sched_reset_on_fork)
3992 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3993 if (task_has_dl_policy(p))
3994 __getparam_dl(p, &attr);
3995 else if (task_has_rt_policy(p))
3996 attr.sched_priority = p->rt_priority;
3998 attr.sched_nice = task_nice(p);
4002 retval = sched_read_attr(uattr, &attr, size);
4010 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4012 cpumask_var_t cpus_allowed, new_mask;
4013 struct task_struct *p;
4018 p = find_process_by_pid(pid);
4024 /* Prevent p going away */
4028 if (p->flags & PF_NO_SETAFFINITY) {
4032 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4036 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4038 goto out_free_cpus_allowed;
4041 if (!check_same_owner(p)) {
4043 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4045 goto out_free_new_mask;
4050 retval = security_task_setscheduler(p);
4052 goto out_free_new_mask;
4055 cpuset_cpus_allowed(p, cpus_allowed);
4056 cpumask_and(new_mask, in_mask, cpus_allowed);
4059 * Since bandwidth control happens on root_domain basis,
4060 * if admission test is enabled, we only admit -deadline
4061 * tasks allowed to run on all the CPUs in the task's
4065 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4067 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4070 goto out_free_new_mask;
4076 retval = set_cpus_allowed_ptr(p, new_mask);
4079 cpuset_cpus_allowed(p, cpus_allowed);
4080 if (!cpumask_subset(new_mask, cpus_allowed)) {
4082 * We must have raced with a concurrent cpuset
4083 * update. Just reset the cpus_allowed to the
4084 * cpuset's cpus_allowed
4086 cpumask_copy(new_mask, cpus_allowed);
4091 free_cpumask_var(new_mask);
4092 out_free_cpus_allowed:
4093 free_cpumask_var(cpus_allowed);
4099 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4100 struct cpumask *new_mask)
4102 if (len < cpumask_size())
4103 cpumask_clear(new_mask);
4104 else if (len > cpumask_size())
4105 len = cpumask_size();
4107 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4111 * sys_sched_setaffinity - set the cpu affinity of a process
4112 * @pid: pid of the process
4113 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4114 * @user_mask_ptr: user-space pointer to the new cpu mask
4116 * Return: 0 on success. An error code otherwise.
4118 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4119 unsigned long __user *, user_mask_ptr)
4121 cpumask_var_t new_mask;
4124 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4127 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4129 retval = sched_setaffinity(pid, new_mask);
4130 free_cpumask_var(new_mask);
4134 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4136 struct task_struct *p;
4137 unsigned long flags;
4143 p = find_process_by_pid(pid);
4147 retval = security_task_getscheduler(p);
4151 raw_spin_lock_irqsave(&p->pi_lock, flags);
4152 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4153 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4162 * sys_sched_getaffinity - get the cpu affinity of a process
4163 * @pid: pid of the process
4164 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4165 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4167 * Return: 0 on success. An error code otherwise.
4169 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4170 unsigned long __user *, user_mask_ptr)
4175 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4177 if (len & (sizeof(unsigned long)-1))
4180 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4183 ret = sched_getaffinity(pid, mask);
4185 size_t retlen = min_t(size_t, len, cpumask_size());
4187 if (copy_to_user(user_mask_ptr, mask, retlen))
4192 free_cpumask_var(mask);
4198 * sys_sched_yield - yield the current processor to other threads.
4200 * This function yields the current CPU to other tasks. If there are no
4201 * other threads running on this CPU then this function will return.
4205 SYSCALL_DEFINE0(sched_yield)
4207 struct rq *rq = this_rq_lock();
4209 schedstat_inc(rq, yld_count);
4210 current->sched_class->yield_task(rq);
4213 * Since we are going to call schedule() anyway, there's
4214 * no need to preempt or enable interrupts:
4216 __release(rq->lock);
4217 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4218 do_raw_spin_unlock(&rq->lock);
4219 sched_preempt_enable_no_resched();
4226 static void __cond_resched(void)
4228 __preempt_count_add(PREEMPT_ACTIVE);
4230 __preempt_count_sub(PREEMPT_ACTIVE);
4233 int __sched _cond_resched(void)
4235 if (should_resched()) {
4241 EXPORT_SYMBOL(_cond_resched);
4244 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4245 * call schedule, and on return reacquire the lock.
4247 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4248 * operations here to prevent schedule() from being called twice (once via
4249 * spin_unlock(), once by hand).
4251 int __cond_resched_lock(spinlock_t *lock)
4253 int resched = should_resched();
4256 lockdep_assert_held(lock);
4258 if (spin_needbreak(lock) || resched) {
4269 EXPORT_SYMBOL(__cond_resched_lock);
4271 int __sched __cond_resched_softirq(void)
4273 BUG_ON(!in_softirq());
4275 if (should_resched()) {
4283 EXPORT_SYMBOL(__cond_resched_softirq);
4286 * yield - yield the current processor to other threads.
4288 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4290 * The scheduler is at all times free to pick the calling task as the most
4291 * eligible task to run, if removing the yield() call from your code breaks
4292 * it, its already broken.
4294 * Typical broken usage is:
4299 * where one assumes that yield() will let 'the other' process run that will
4300 * make event true. If the current task is a SCHED_FIFO task that will never
4301 * happen. Never use yield() as a progress guarantee!!
4303 * If you want to use yield() to wait for something, use wait_event().
4304 * If you want to use yield() to be 'nice' for others, use cond_resched().
4305 * If you still want to use yield(), do not!
4307 void __sched yield(void)
4309 set_current_state(TASK_RUNNING);
4312 EXPORT_SYMBOL(yield);
4315 * yield_to - yield the current processor to another thread in
4316 * your thread group, or accelerate that thread toward the
4317 * processor it's on.
4319 * @preempt: whether task preemption is allowed or not
4321 * It's the caller's job to ensure that the target task struct
4322 * can't go away on us before we can do any checks.
4325 * true (>0) if we indeed boosted the target task.
4326 * false (0) if we failed to boost the target.
4327 * -ESRCH if there's no task to yield to.
4329 int __sched yield_to(struct task_struct *p, bool preempt)
4331 struct task_struct *curr = current;
4332 struct rq *rq, *p_rq;
4333 unsigned long flags;
4336 local_irq_save(flags);
4342 * If we're the only runnable task on the rq and target rq also
4343 * has only one task, there's absolutely no point in yielding.
4345 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4350 double_rq_lock(rq, p_rq);
4351 if (task_rq(p) != p_rq) {
4352 double_rq_unlock(rq, p_rq);
4356 if (!curr->sched_class->yield_to_task)
4359 if (curr->sched_class != p->sched_class)
4362 if (task_running(p_rq, p) || p->state)
4365 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4367 schedstat_inc(rq, yld_count);
4369 * Make p's CPU reschedule; pick_next_entity takes care of
4372 if (preempt && rq != p_rq)
4377 double_rq_unlock(rq, p_rq);
4379 local_irq_restore(flags);
4386 EXPORT_SYMBOL_GPL(yield_to);
4389 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4390 * that process accounting knows that this is a task in IO wait state.
4392 void __sched io_schedule(void)
4394 struct rq *rq = raw_rq();
4396 delayacct_blkio_start();
4397 atomic_inc(&rq->nr_iowait);
4398 blk_flush_plug(current);
4399 current->in_iowait = 1;
4401 current->in_iowait = 0;
4402 atomic_dec(&rq->nr_iowait);
4403 delayacct_blkio_end();
4405 EXPORT_SYMBOL(io_schedule);
4407 long __sched io_schedule_timeout(long timeout)
4409 struct rq *rq = raw_rq();
4412 delayacct_blkio_start();
4413 atomic_inc(&rq->nr_iowait);
4414 blk_flush_plug(current);
4415 current->in_iowait = 1;
4416 ret = schedule_timeout(timeout);
4417 current->in_iowait = 0;
4418 atomic_dec(&rq->nr_iowait);
4419 delayacct_blkio_end();
4424 * sys_sched_get_priority_max - return maximum RT priority.
4425 * @policy: scheduling class.
4427 * Return: On success, this syscall returns the maximum
4428 * rt_priority that can be used by a given scheduling class.
4429 * On failure, a negative error code is returned.
4431 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4438 ret = MAX_USER_RT_PRIO-1;
4440 case SCHED_DEADLINE:
4451 * sys_sched_get_priority_min - return minimum RT priority.
4452 * @policy: scheduling class.
4454 * Return: On success, this syscall returns the minimum
4455 * rt_priority that can be used by a given scheduling class.
4456 * On failure, a negative error code is returned.
4458 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4467 case SCHED_DEADLINE:
4477 * sys_sched_rr_get_interval - return the default timeslice of a process.
4478 * @pid: pid of the process.
4479 * @interval: userspace pointer to the timeslice value.
4481 * this syscall writes the default timeslice value of a given process
4482 * into the user-space timespec buffer. A value of '0' means infinity.
4484 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4487 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4488 struct timespec __user *, interval)
4490 struct task_struct *p;
4491 unsigned int time_slice;
4492 unsigned long flags;
4502 p = find_process_by_pid(pid);
4506 retval = security_task_getscheduler(p);
4510 rq = task_rq_lock(p, &flags);
4512 if (p->sched_class->get_rr_interval)
4513 time_slice = p->sched_class->get_rr_interval(rq, p);
4514 task_rq_unlock(rq, p, &flags);
4517 jiffies_to_timespec(time_slice, &t);
4518 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4526 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4528 void sched_show_task(struct task_struct *p)
4530 unsigned long free = 0;
4534 state = p->state ? __ffs(p->state) + 1 : 0;
4535 printk(KERN_INFO "%-15.15s %c", p->comm,
4536 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4537 #if BITS_PER_LONG == 32
4538 if (state == TASK_RUNNING)
4539 printk(KERN_CONT " running ");
4541 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4543 if (state == TASK_RUNNING)
4544 printk(KERN_CONT " running task ");
4546 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4548 #ifdef CONFIG_DEBUG_STACK_USAGE
4549 free = stack_not_used(p);
4552 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4554 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4555 task_pid_nr(p), ppid,
4556 (unsigned long)task_thread_info(p)->flags);
4558 print_worker_info(KERN_INFO, p);
4559 show_stack(p, NULL);
4562 void show_state_filter(unsigned long state_filter)
4564 struct task_struct *g, *p;
4566 #if BITS_PER_LONG == 32
4568 " task PC stack pid father\n");
4571 " task PC stack pid father\n");
4574 for_each_process_thread(g, p) {
4576 * reset the NMI-timeout, listing all files on a slow
4577 * console might take a lot of time:
4579 touch_nmi_watchdog();
4580 if (!state_filter || (p->state & state_filter))
4584 touch_all_softlockup_watchdogs();
4586 #ifdef CONFIG_SCHED_DEBUG
4587 sysrq_sched_debug_show();
4591 * Only show locks if all tasks are dumped:
4594 debug_show_all_locks();
4597 void init_idle_bootup_task(struct task_struct *idle)
4599 idle->sched_class = &idle_sched_class;
4603 * init_idle - set up an idle thread for a given CPU
4604 * @idle: task in question
4605 * @cpu: cpu the idle task belongs to
4607 * NOTE: this function does not set the idle thread's NEED_RESCHED
4608 * flag, to make booting more robust.
4610 void init_idle(struct task_struct *idle, int cpu)
4612 struct rq *rq = cpu_rq(cpu);
4613 unsigned long flags;
4615 raw_spin_lock_irqsave(&rq->lock, flags);
4617 __sched_fork(0, idle);
4618 idle->state = TASK_RUNNING;
4619 idle->se.exec_start = sched_clock();
4621 do_set_cpus_allowed(idle, cpumask_of(cpu));
4623 * We're having a chicken and egg problem, even though we are
4624 * holding rq->lock, the cpu isn't yet set to this cpu so the
4625 * lockdep check in task_group() will fail.
4627 * Similar case to sched_fork(). / Alternatively we could
4628 * use task_rq_lock() here and obtain the other rq->lock.
4633 __set_task_cpu(idle, cpu);
4636 rq->curr = rq->idle = idle;
4637 idle->on_rq = TASK_ON_RQ_QUEUED;
4638 #if defined(CONFIG_SMP)
4641 raw_spin_unlock_irqrestore(&rq->lock, flags);
4643 /* Set the preempt count _outside_ the spinlocks! */
4644 init_idle_preempt_count(idle, cpu);
4647 * The idle tasks have their own, simple scheduling class:
4649 idle->sched_class = &idle_sched_class;
4650 ftrace_graph_init_idle_task(idle, cpu);
4651 vtime_init_idle(idle, cpu);
4652 #if defined(CONFIG_SMP)
4653 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4657 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4658 const struct cpumask *trial)
4660 int ret = 1, trial_cpus;
4661 struct dl_bw *cur_dl_b;
4662 unsigned long flags;
4664 cur_dl_b = dl_bw_of(cpumask_any(cur));
4665 trial_cpus = cpumask_weight(trial);
4667 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4668 if (cur_dl_b->bw != -1 &&
4669 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4671 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4676 int task_can_attach(struct task_struct *p,
4677 const struct cpumask *cs_cpus_allowed)
4682 * Kthreads which disallow setaffinity shouldn't be moved
4683 * to a new cpuset; we don't want to change their cpu
4684 * affinity and isolating such threads by their set of
4685 * allowed nodes is unnecessary. Thus, cpusets are not
4686 * applicable for such threads. This prevents checking for
4687 * success of set_cpus_allowed_ptr() on all attached tasks
4688 * before cpus_allowed may be changed.
4690 if (p->flags & PF_NO_SETAFFINITY) {
4696 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4698 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4700 struct dl_bw *dl_b = dl_bw_of(dest_cpu);
4703 unsigned long flags;
4705 raw_spin_lock_irqsave(&dl_b->lock, flags);
4706 cpus = dl_bw_cpus(dest_cpu);
4707 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4712 * We reserve space for this task in the destination
4713 * root_domain, as we can't fail after this point.
4714 * We will free resources in the source root_domain
4715 * later on (see set_cpus_allowed_dl()).
4717 __dl_add(dl_b, p->dl.dl_bw);
4719 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4729 * move_queued_task - move a queued task to new rq.
4731 * Returns (locked) new rq. Old rq's lock is released.
4733 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4735 struct rq *rq = task_rq(p);
4737 lockdep_assert_held(&rq->lock);
4739 dequeue_task(rq, p, 0);
4740 p->on_rq = TASK_ON_RQ_MIGRATING;
4741 set_task_cpu(p, new_cpu);
4742 raw_spin_unlock(&rq->lock);
4744 rq = cpu_rq(new_cpu);
4746 raw_spin_lock(&rq->lock);
4747 BUG_ON(task_cpu(p) != new_cpu);
4748 p->on_rq = TASK_ON_RQ_QUEUED;
4749 enqueue_task(rq, p, 0);
4750 check_preempt_curr(rq, p, 0);
4755 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4757 if (p->sched_class && p->sched_class->set_cpus_allowed)
4758 p->sched_class->set_cpus_allowed(p, new_mask);
4760 cpumask_copy(&p->cpus_allowed, new_mask);
4761 p->nr_cpus_allowed = cpumask_weight(new_mask);
4765 * This is how migration works:
4767 * 1) we invoke migration_cpu_stop() on the target CPU using
4769 * 2) stopper starts to run (implicitly forcing the migrated thread
4771 * 3) it checks whether the migrated task is still in the wrong runqueue.
4772 * 4) if it's in the wrong runqueue then the migration thread removes
4773 * it and puts it into the right queue.
4774 * 5) stopper completes and stop_one_cpu() returns and the migration
4779 * Change a given task's CPU affinity. Migrate the thread to a
4780 * proper CPU and schedule it away if the CPU it's executing on
4781 * is removed from the allowed bitmask.
4783 * NOTE: the caller must have a valid reference to the task, the
4784 * task must not exit() & deallocate itself prematurely. The
4785 * call is not atomic; no spinlocks may be held.
4787 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4789 unsigned long flags;
4791 unsigned int dest_cpu;
4794 rq = task_rq_lock(p, &flags);
4796 if (cpumask_equal(&p->cpus_allowed, new_mask))
4799 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4804 do_set_cpus_allowed(p, new_mask);
4806 /* Can the task run on the task's current CPU? If so, we're done */
4807 if (cpumask_test_cpu(task_cpu(p), new_mask))
4810 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4811 if (task_running(rq, p) || p->state == TASK_WAKING) {
4812 struct migration_arg arg = { p, dest_cpu };
4813 /* Need help from migration thread: drop lock and wait. */
4814 task_rq_unlock(rq, p, &flags);
4815 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4816 tlb_migrate_finish(p->mm);
4818 } else if (task_on_rq_queued(p))
4819 rq = move_queued_task(p, dest_cpu);
4821 task_rq_unlock(rq, p, &flags);
4825 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4828 * Move (not current) task off this cpu, onto dest cpu. We're doing
4829 * this because either it can't run here any more (set_cpus_allowed()
4830 * away from this CPU, or CPU going down), or because we're
4831 * attempting to rebalance this task on exec (sched_exec).
4833 * So we race with normal scheduler movements, but that's OK, as long
4834 * as the task is no longer on this CPU.
4836 * Returns non-zero if task was successfully migrated.
4838 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4843 if (unlikely(!cpu_active(dest_cpu)))
4846 rq = cpu_rq(src_cpu);
4848 raw_spin_lock(&p->pi_lock);
4849 raw_spin_lock(&rq->lock);
4850 /* Already moved. */
4851 if (task_cpu(p) != src_cpu)
4854 /* Affinity changed (again). */
4855 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4859 * If we're not on a rq, the next wake-up will ensure we're
4862 if (task_on_rq_queued(p))
4863 rq = move_queued_task(p, dest_cpu);
4867 raw_spin_unlock(&rq->lock);
4868 raw_spin_unlock(&p->pi_lock);
4872 #ifdef CONFIG_NUMA_BALANCING
4873 /* Migrate current task p to target_cpu */
4874 int migrate_task_to(struct task_struct *p, int target_cpu)
4876 struct migration_arg arg = { p, target_cpu };
4877 int curr_cpu = task_cpu(p);
4879 if (curr_cpu == target_cpu)
4882 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4885 /* TODO: This is not properly updating schedstats */
4887 trace_sched_move_numa(p, curr_cpu, target_cpu);
4888 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4892 * Requeue a task on a given node and accurately track the number of NUMA
4893 * tasks on the runqueues
4895 void sched_setnuma(struct task_struct *p, int nid)
4898 unsigned long flags;
4899 bool queued, running;
4901 rq = task_rq_lock(p, &flags);
4902 queued = task_on_rq_queued(p);
4903 running = task_current(rq, p);
4906 dequeue_task(rq, p, 0);
4908 put_prev_task(rq, p);
4910 p->numa_preferred_nid = nid;
4913 p->sched_class->set_curr_task(rq);
4915 enqueue_task(rq, p, 0);
4916 task_rq_unlock(rq, p, &flags);
4921 * migration_cpu_stop - this will be executed by a highprio stopper thread
4922 * and performs thread migration by bumping thread off CPU then
4923 * 'pushing' onto another runqueue.
4925 static int migration_cpu_stop(void *data)
4927 struct migration_arg *arg = data;
4930 * The original target cpu might have gone down and we might
4931 * be on another cpu but it doesn't matter.
4933 local_irq_disable();
4935 * We need to explicitly wake pending tasks before running
4936 * __migrate_task() such that we will not miss enforcing cpus_allowed
4937 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4939 sched_ttwu_pending();
4940 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4945 #ifdef CONFIG_HOTPLUG_CPU
4948 * Ensures that the idle task is using init_mm right before its cpu goes
4951 void idle_task_exit(void)
4953 struct mm_struct *mm = current->active_mm;
4955 BUG_ON(cpu_online(smp_processor_id()));
4957 if (mm != &init_mm) {
4958 switch_mm(mm, &init_mm, current);
4959 finish_arch_post_lock_switch();
4965 * Since this CPU is going 'away' for a while, fold any nr_active delta
4966 * we might have. Assumes we're called after migrate_tasks() so that the
4967 * nr_active count is stable.
4969 * Also see the comment "Global load-average calculations".
4971 static void calc_load_migrate(struct rq *rq)
4973 long delta = calc_load_fold_active(rq);
4975 atomic_long_add(delta, &calc_load_tasks);
4978 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4982 static const struct sched_class fake_sched_class = {
4983 .put_prev_task = put_prev_task_fake,
4986 static struct task_struct fake_task = {
4988 * Avoid pull_{rt,dl}_task()
4990 .prio = MAX_PRIO + 1,
4991 .sched_class = &fake_sched_class,
4995 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4996 * try_to_wake_up()->select_task_rq().
4998 * Called with rq->lock held even though we'er in stop_machine() and
4999 * there's no concurrency possible, we hold the required locks anyway
5000 * because of lock validation efforts.
5002 static void migrate_tasks(unsigned int dead_cpu)
5004 struct rq *rq = cpu_rq(dead_cpu);
5005 struct task_struct *next, *stop = rq->stop;
5009 * Fudge the rq selection such that the below task selection loop
5010 * doesn't get stuck on the currently eligible stop task.
5012 * We're currently inside stop_machine() and the rq is either stuck
5013 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5014 * either way we should never end up calling schedule() until we're
5020 * put_prev_task() and pick_next_task() sched
5021 * class method both need to have an up-to-date
5022 * value of rq->clock[_task]
5024 update_rq_clock(rq);
5028 * There's this thread running, bail when that's the only
5031 if (rq->nr_running == 1)
5034 next = pick_next_task(rq, &fake_task);
5036 next->sched_class->put_prev_task(rq, next);
5038 /* Find suitable destination for @next, with force if needed. */
5039 dest_cpu = select_fallback_rq(dead_cpu, next);
5040 raw_spin_unlock(&rq->lock);
5042 __migrate_task(next, dead_cpu, dest_cpu);
5044 raw_spin_lock(&rq->lock);
5050 #endif /* CONFIG_HOTPLUG_CPU */
5052 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5054 static struct ctl_table sd_ctl_dir[] = {
5056 .procname = "sched_domain",
5062 static struct ctl_table sd_ctl_root[] = {
5064 .procname = "kernel",
5066 .child = sd_ctl_dir,
5071 static struct ctl_table *sd_alloc_ctl_entry(int n)
5073 struct ctl_table *entry =
5074 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5079 static void sd_free_ctl_entry(struct ctl_table **tablep)
5081 struct ctl_table *entry;
5084 * In the intermediate directories, both the child directory and
5085 * procname are dynamically allocated and could fail but the mode
5086 * will always be set. In the lowest directory the names are
5087 * static strings and all have proc handlers.
5089 for (entry = *tablep; entry->mode; entry++) {
5091 sd_free_ctl_entry(&entry->child);
5092 if (entry->proc_handler == NULL)
5093 kfree(entry->procname);
5100 static int min_load_idx = 0;
5101 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5104 set_table_entry(struct ctl_table *entry,
5105 const char *procname, void *data, int maxlen,
5106 umode_t mode, proc_handler *proc_handler,
5109 entry->procname = procname;
5111 entry->maxlen = maxlen;
5113 entry->proc_handler = proc_handler;
5116 entry->extra1 = &min_load_idx;
5117 entry->extra2 = &max_load_idx;
5121 static struct ctl_table *
5122 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5124 struct ctl_table *table = sd_alloc_ctl_entry(14);
5129 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5130 sizeof(long), 0644, proc_doulongvec_minmax, false);
5131 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5132 sizeof(long), 0644, proc_doulongvec_minmax, false);
5133 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5134 sizeof(int), 0644, proc_dointvec_minmax, true);
5135 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5136 sizeof(int), 0644, proc_dointvec_minmax, true);
5137 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5138 sizeof(int), 0644, proc_dointvec_minmax, true);
5139 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5140 sizeof(int), 0644, proc_dointvec_minmax, true);
5141 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5142 sizeof(int), 0644, proc_dointvec_minmax, true);
5143 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5144 sizeof(int), 0644, proc_dointvec_minmax, false);
5145 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5146 sizeof(int), 0644, proc_dointvec_minmax, false);
5147 set_table_entry(&table[9], "cache_nice_tries",
5148 &sd->cache_nice_tries,
5149 sizeof(int), 0644, proc_dointvec_minmax, false);
5150 set_table_entry(&table[10], "flags", &sd->flags,
5151 sizeof(int), 0644, proc_dointvec_minmax, false);
5152 set_table_entry(&table[11], "max_newidle_lb_cost",
5153 &sd->max_newidle_lb_cost,
5154 sizeof(long), 0644, proc_doulongvec_minmax, false);
5155 set_table_entry(&table[12], "name", sd->name,
5156 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5157 /* &table[13] is terminator */
5162 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5164 struct ctl_table *entry, *table;
5165 struct sched_domain *sd;
5166 int domain_num = 0, i;
5169 for_each_domain(cpu, sd)
5171 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5176 for_each_domain(cpu, sd) {
5177 snprintf(buf, 32, "domain%d", i);
5178 entry->procname = kstrdup(buf, GFP_KERNEL);
5180 entry->child = sd_alloc_ctl_domain_table(sd);
5187 static struct ctl_table_header *sd_sysctl_header;
5188 static void register_sched_domain_sysctl(void)
5190 int i, cpu_num = num_possible_cpus();
5191 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5194 WARN_ON(sd_ctl_dir[0].child);
5195 sd_ctl_dir[0].child = entry;
5200 for_each_possible_cpu(i) {
5201 snprintf(buf, 32, "cpu%d", i);
5202 entry->procname = kstrdup(buf, GFP_KERNEL);
5204 entry->child = sd_alloc_ctl_cpu_table(i);
5208 WARN_ON(sd_sysctl_header);
5209 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5212 /* may be called multiple times per register */
5213 static void unregister_sched_domain_sysctl(void)
5215 if (sd_sysctl_header)
5216 unregister_sysctl_table(sd_sysctl_header);
5217 sd_sysctl_header = NULL;
5218 if (sd_ctl_dir[0].child)
5219 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5222 static void register_sched_domain_sysctl(void)
5225 static void unregister_sched_domain_sysctl(void)
5230 static void set_rq_online(struct rq *rq)
5233 const struct sched_class *class;
5235 cpumask_set_cpu(rq->cpu, rq->rd->online);
5238 for_each_class(class) {
5239 if (class->rq_online)
5240 class->rq_online(rq);
5245 static void set_rq_offline(struct rq *rq)
5248 const struct sched_class *class;
5250 for_each_class(class) {
5251 if (class->rq_offline)
5252 class->rq_offline(rq);
5255 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5261 * migration_call - callback that gets triggered when a CPU is added.
5262 * Here we can start up the necessary migration thread for the new CPU.
5265 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5267 int cpu = (long)hcpu;
5268 unsigned long flags;
5269 struct rq *rq = cpu_rq(cpu);
5271 switch (action & ~CPU_TASKS_FROZEN) {
5273 case CPU_UP_PREPARE:
5274 rq->calc_load_update = calc_load_update;
5278 /* Update our root-domain */
5279 raw_spin_lock_irqsave(&rq->lock, flags);
5281 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5285 raw_spin_unlock_irqrestore(&rq->lock, flags);
5288 #ifdef CONFIG_HOTPLUG_CPU
5290 sched_ttwu_pending();
5291 /* Update our root-domain */
5292 raw_spin_lock_irqsave(&rq->lock, flags);
5294 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5298 BUG_ON(rq->nr_running != 1); /* the migration thread */
5299 raw_spin_unlock_irqrestore(&rq->lock, flags);
5303 calc_load_migrate(rq);
5308 update_max_interval();
5314 * Register at high priority so that task migration (migrate_all_tasks)
5315 * happens before everything else. This has to be lower priority than
5316 * the notifier in the perf_event subsystem, though.
5318 static struct notifier_block migration_notifier = {
5319 .notifier_call = migration_call,
5320 .priority = CPU_PRI_MIGRATION,
5323 static void __cpuinit set_cpu_rq_start_time(void)
5325 int cpu = smp_processor_id();
5326 struct rq *rq = cpu_rq(cpu);
5327 rq->age_stamp = sched_clock_cpu(cpu);
5330 static int sched_cpu_active(struct notifier_block *nfb,
5331 unsigned long action, void *hcpu)
5333 switch (action & ~CPU_TASKS_FROZEN) {
5335 set_cpu_rq_start_time();
5337 case CPU_DOWN_FAILED:
5338 set_cpu_active((long)hcpu, true);
5345 static int sched_cpu_inactive(struct notifier_block *nfb,
5346 unsigned long action, void *hcpu)
5348 unsigned long flags;
5349 long cpu = (long)hcpu;
5352 switch (action & ~CPU_TASKS_FROZEN) {
5353 case CPU_DOWN_PREPARE:
5354 set_cpu_active(cpu, false);
5356 /* explicitly allow suspend */
5357 if (!(action & CPU_TASKS_FROZEN)) {
5361 rcu_read_lock_sched();
5362 dl_b = dl_bw_of(cpu);
5364 raw_spin_lock_irqsave(&dl_b->lock, flags);
5365 cpus = dl_bw_cpus(cpu);
5366 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5367 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5369 rcu_read_unlock_sched();
5372 return notifier_from_errno(-EBUSY);
5380 static int __init migration_init(void)
5382 void *cpu = (void *)(long)smp_processor_id();
5385 /* Initialize migration for the boot CPU */
5386 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5387 BUG_ON(err == NOTIFY_BAD);
5388 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5389 register_cpu_notifier(&migration_notifier);
5391 /* Register cpu active notifiers */
5392 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5393 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5397 early_initcall(migration_init);
5402 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5404 #ifdef CONFIG_SCHED_DEBUG
5406 static __read_mostly int sched_debug_enabled;
5408 static int __init sched_debug_setup(char *str)
5410 sched_debug_enabled = 1;
5414 early_param("sched_debug", sched_debug_setup);
5416 static inline bool sched_debug(void)
5418 return sched_debug_enabled;
5421 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5422 struct cpumask *groupmask)
5424 struct sched_group *group = sd->groups;
5427 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5428 cpumask_clear(groupmask);
5430 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5432 if (!(sd->flags & SD_LOAD_BALANCE)) {
5433 printk("does not load-balance\n");
5435 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5440 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5442 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5443 printk(KERN_ERR "ERROR: domain->span does not contain "
5446 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5447 printk(KERN_ERR "ERROR: domain->groups does not contain"
5451 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5455 printk(KERN_ERR "ERROR: group is NULL\n");
5460 * Even though we initialize ->capacity to something semi-sane,
5461 * we leave capacity_orig unset. This allows us to detect if
5462 * domain iteration is still funny without causing /0 traps.
5464 if (!group->sgc->capacity_orig) {
5465 printk(KERN_CONT "\n");
5466 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5470 if (!cpumask_weight(sched_group_cpus(group))) {
5471 printk(KERN_CONT "\n");
5472 printk(KERN_ERR "ERROR: empty group\n");
5476 if (!(sd->flags & SD_OVERLAP) &&
5477 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5478 printk(KERN_CONT "\n");
5479 printk(KERN_ERR "ERROR: repeated CPUs\n");
5483 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5485 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5487 printk(KERN_CONT " %s", str);
5488 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5489 printk(KERN_CONT " (cpu_capacity = %d)",
5490 group->sgc->capacity);
5493 group = group->next;
5494 } while (group != sd->groups);
5495 printk(KERN_CONT "\n");
5497 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5498 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5501 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5502 printk(KERN_ERR "ERROR: parent span is not a superset "
5503 "of domain->span\n");
5507 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5511 if (!sched_debug_enabled)
5515 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5519 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5522 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5530 #else /* !CONFIG_SCHED_DEBUG */
5531 # define sched_domain_debug(sd, cpu) do { } while (0)
5532 static inline bool sched_debug(void)
5536 #endif /* CONFIG_SCHED_DEBUG */
5538 static int sd_degenerate(struct sched_domain *sd)
5540 if (cpumask_weight(sched_domain_span(sd)) == 1)
5543 /* Following flags need at least 2 groups */
5544 if (sd->flags & (SD_LOAD_BALANCE |
5545 SD_BALANCE_NEWIDLE |
5548 SD_SHARE_CPUCAPACITY |
5549 SD_SHARE_PKG_RESOURCES |
5550 SD_SHARE_POWERDOMAIN)) {
5551 if (sd->groups != sd->groups->next)
5555 /* Following flags don't use groups */
5556 if (sd->flags & (SD_WAKE_AFFINE))
5563 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5565 unsigned long cflags = sd->flags, pflags = parent->flags;
5567 if (sd_degenerate(parent))
5570 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5573 /* Flags needing groups don't count if only 1 group in parent */
5574 if (parent->groups == parent->groups->next) {
5575 pflags &= ~(SD_LOAD_BALANCE |
5576 SD_BALANCE_NEWIDLE |
5579 SD_SHARE_CPUCAPACITY |
5580 SD_SHARE_PKG_RESOURCES |
5582 SD_SHARE_POWERDOMAIN);
5583 if (nr_node_ids == 1)
5584 pflags &= ~SD_SERIALIZE;
5586 if (~cflags & pflags)
5592 static void free_rootdomain(struct rcu_head *rcu)
5594 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5596 cpupri_cleanup(&rd->cpupri);
5597 cpudl_cleanup(&rd->cpudl);
5598 free_cpumask_var(rd->dlo_mask);
5599 free_cpumask_var(rd->rto_mask);
5600 free_cpumask_var(rd->online);
5601 free_cpumask_var(rd->span);
5605 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5607 struct root_domain *old_rd = NULL;
5608 unsigned long flags;
5610 raw_spin_lock_irqsave(&rq->lock, flags);
5615 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5618 cpumask_clear_cpu(rq->cpu, old_rd->span);
5621 * If we dont want to free the old_rd yet then
5622 * set old_rd to NULL to skip the freeing later
5625 if (!atomic_dec_and_test(&old_rd->refcount))
5629 atomic_inc(&rd->refcount);
5632 cpumask_set_cpu(rq->cpu, rd->span);
5633 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5636 raw_spin_unlock_irqrestore(&rq->lock, flags);
5639 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5642 static int init_rootdomain(struct root_domain *rd)
5644 memset(rd, 0, sizeof(*rd));
5646 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5648 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5650 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5652 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5655 init_dl_bw(&rd->dl_bw);
5656 if (cpudl_init(&rd->cpudl) != 0)
5659 if (cpupri_init(&rd->cpupri) != 0)
5664 free_cpumask_var(rd->rto_mask);
5666 free_cpumask_var(rd->dlo_mask);
5668 free_cpumask_var(rd->online);
5670 free_cpumask_var(rd->span);
5676 * By default the system creates a single root-domain with all cpus as
5677 * members (mimicking the global state we have today).
5679 struct root_domain def_root_domain;
5681 static void init_defrootdomain(void)
5683 init_rootdomain(&def_root_domain);
5685 atomic_set(&def_root_domain.refcount, 1);
5688 static struct root_domain *alloc_rootdomain(void)
5690 struct root_domain *rd;
5692 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5696 if (init_rootdomain(rd) != 0) {
5704 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5706 struct sched_group *tmp, *first;
5715 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5720 } while (sg != first);
5723 static void free_sched_domain(struct rcu_head *rcu)
5725 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5728 * If its an overlapping domain it has private groups, iterate and
5731 if (sd->flags & SD_OVERLAP) {
5732 free_sched_groups(sd->groups, 1);
5733 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5734 kfree(sd->groups->sgc);
5740 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5742 call_rcu(&sd->rcu, free_sched_domain);
5745 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5747 for (; sd; sd = sd->parent)
5748 destroy_sched_domain(sd, cpu);
5752 * Keep a special pointer to the highest sched_domain that has
5753 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5754 * allows us to avoid some pointer chasing select_idle_sibling().
5756 * Also keep a unique ID per domain (we use the first cpu number in
5757 * the cpumask of the domain), this allows us to quickly tell if
5758 * two cpus are in the same cache domain, see cpus_share_cache().
5760 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5761 DEFINE_PER_CPU(int, sd_llc_size);
5762 DEFINE_PER_CPU(int, sd_llc_id);
5763 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5764 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5765 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5767 static void update_top_cache_domain(int cpu)
5769 struct sched_domain *sd;
5770 struct sched_domain *busy_sd = NULL;
5774 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5776 id = cpumask_first(sched_domain_span(sd));
5777 size = cpumask_weight(sched_domain_span(sd));
5778 busy_sd = sd->parent; /* sd_busy */
5780 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5782 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5783 per_cpu(sd_llc_size, cpu) = size;
5784 per_cpu(sd_llc_id, cpu) = id;
5786 sd = lowest_flag_domain(cpu, SD_NUMA);
5787 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5789 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5790 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5794 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5795 * hold the hotplug lock.
5798 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5800 struct rq *rq = cpu_rq(cpu);
5801 struct sched_domain *tmp;
5803 /* Remove the sched domains which do not contribute to scheduling. */
5804 for (tmp = sd; tmp; ) {
5805 struct sched_domain *parent = tmp->parent;
5809 if (sd_parent_degenerate(tmp, parent)) {
5810 tmp->parent = parent->parent;
5812 parent->parent->child = tmp;
5814 * Transfer SD_PREFER_SIBLING down in case of a
5815 * degenerate parent; the spans match for this
5816 * so the property transfers.
5818 if (parent->flags & SD_PREFER_SIBLING)
5819 tmp->flags |= SD_PREFER_SIBLING;
5820 destroy_sched_domain(parent, cpu);
5825 if (sd && sd_degenerate(sd)) {
5828 destroy_sched_domain(tmp, cpu);
5833 sched_domain_debug(sd, cpu);
5835 rq_attach_root(rq, rd);
5837 rcu_assign_pointer(rq->sd, sd);
5838 destroy_sched_domains(tmp, cpu);
5840 update_top_cache_domain(cpu);
5843 /* cpus with isolated domains */
5844 static cpumask_var_t cpu_isolated_map;
5846 /* Setup the mask of cpus configured for isolated domains */
5847 static int __init isolated_cpu_setup(char *str)
5849 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5850 cpulist_parse(str, cpu_isolated_map);
5854 __setup("isolcpus=", isolated_cpu_setup);
5857 struct sched_domain ** __percpu sd;
5858 struct root_domain *rd;
5869 * Build an iteration mask that can exclude certain CPUs from the upwards
5872 * Asymmetric node setups can result in situations where the domain tree is of
5873 * unequal depth, make sure to skip domains that already cover the entire
5876 * In that case build_sched_domains() will have terminated the iteration early
5877 * and our sibling sd spans will be empty. Domains should always include the
5878 * cpu they're built on, so check that.
5881 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5883 const struct cpumask *span = sched_domain_span(sd);
5884 struct sd_data *sdd = sd->private;
5885 struct sched_domain *sibling;
5888 for_each_cpu(i, span) {
5889 sibling = *per_cpu_ptr(sdd->sd, i);
5890 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5893 cpumask_set_cpu(i, sched_group_mask(sg));
5898 * Return the canonical balance cpu for this group, this is the first cpu
5899 * of this group that's also in the iteration mask.
5901 int group_balance_cpu(struct sched_group *sg)
5903 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5907 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5909 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5910 const struct cpumask *span = sched_domain_span(sd);
5911 struct cpumask *covered = sched_domains_tmpmask;
5912 struct sd_data *sdd = sd->private;
5913 struct sched_domain *sibling;
5916 cpumask_clear(covered);
5918 for_each_cpu(i, span) {
5919 struct cpumask *sg_span;
5921 if (cpumask_test_cpu(i, covered))
5924 sibling = *per_cpu_ptr(sdd->sd, i);
5926 /* See the comment near build_group_mask(). */
5927 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5930 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5931 GFP_KERNEL, cpu_to_node(cpu));
5936 sg_span = sched_group_cpus(sg);
5938 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5940 cpumask_set_cpu(i, sg_span);
5942 cpumask_or(covered, covered, sg_span);
5944 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5945 if (atomic_inc_return(&sg->sgc->ref) == 1)
5946 build_group_mask(sd, sg);
5949 * Initialize sgc->capacity such that even if we mess up the
5950 * domains and no possible iteration will get us here, we won't
5953 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5954 sg->sgc->capacity_orig = sg->sgc->capacity;
5957 * Make sure the first group of this domain contains the
5958 * canonical balance cpu. Otherwise the sched_domain iteration
5959 * breaks. See update_sg_lb_stats().
5961 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5962 group_balance_cpu(sg) == cpu)
5972 sd->groups = groups;
5977 free_sched_groups(first, 0);
5982 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5984 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5985 struct sched_domain *child = sd->child;
5988 cpu = cpumask_first(sched_domain_span(child));
5991 *sg = *per_cpu_ptr(sdd->sg, cpu);
5992 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5993 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6000 * build_sched_groups will build a circular linked list of the groups
6001 * covered by the given span, and will set each group's ->cpumask correctly,
6002 * and ->cpu_capacity to 0.
6004 * Assumes the sched_domain tree is fully constructed
6007 build_sched_groups(struct sched_domain *sd, int cpu)
6009 struct sched_group *first = NULL, *last = NULL;
6010 struct sd_data *sdd = sd->private;
6011 const struct cpumask *span = sched_domain_span(sd);
6012 struct cpumask *covered;
6015 get_group(cpu, sdd, &sd->groups);
6016 atomic_inc(&sd->groups->ref);
6018 if (cpu != cpumask_first(span))
6021 lockdep_assert_held(&sched_domains_mutex);
6022 covered = sched_domains_tmpmask;
6024 cpumask_clear(covered);
6026 for_each_cpu(i, span) {
6027 struct sched_group *sg;
6030 if (cpumask_test_cpu(i, covered))
6033 group = get_group(i, sdd, &sg);
6034 cpumask_setall(sched_group_mask(sg));
6036 for_each_cpu(j, span) {
6037 if (get_group(j, sdd, NULL) != group)
6040 cpumask_set_cpu(j, covered);
6041 cpumask_set_cpu(j, sched_group_cpus(sg));
6056 * Initialize sched groups cpu_capacity.
6058 * cpu_capacity indicates the capacity of sched group, which is used while
6059 * distributing the load between different sched groups in a sched domain.
6060 * Typically cpu_capacity for all the groups in a sched domain will be same
6061 * unless there are asymmetries in the topology. If there are asymmetries,
6062 * group having more cpu_capacity will pickup more load compared to the
6063 * group having less cpu_capacity.
6065 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6067 struct sched_group *sg = sd->groups;
6072 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6074 } while (sg != sd->groups);
6076 if (cpu != group_balance_cpu(sg))
6079 update_group_capacity(sd, cpu);
6080 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6084 * Initializers for schedule domains
6085 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6088 static int default_relax_domain_level = -1;
6089 int sched_domain_level_max;
6091 static int __init setup_relax_domain_level(char *str)
6093 if (kstrtoint(str, 0, &default_relax_domain_level))
6094 pr_warn("Unable to set relax_domain_level\n");
6098 __setup("relax_domain_level=", setup_relax_domain_level);
6100 static void set_domain_attribute(struct sched_domain *sd,
6101 struct sched_domain_attr *attr)
6105 if (!attr || attr->relax_domain_level < 0) {
6106 if (default_relax_domain_level < 0)
6109 request = default_relax_domain_level;
6111 request = attr->relax_domain_level;
6112 if (request < sd->level) {
6113 /* turn off idle balance on this domain */
6114 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6116 /* turn on idle balance on this domain */
6117 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6121 static void __sdt_free(const struct cpumask *cpu_map);
6122 static int __sdt_alloc(const struct cpumask *cpu_map);
6124 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6125 const struct cpumask *cpu_map)
6129 if (!atomic_read(&d->rd->refcount))
6130 free_rootdomain(&d->rd->rcu); /* fall through */
6132 free_percpu(d->sd); /* fall through */
6134 __sdt_free(cpu_map); /* fall through */
6140 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6141 const struct cpumask *cpu_map)
6143 memset(d, 0, sizeof(*d));
6145 if (__sdt_alloc(cpu_map))
6146 return sa_sd_storage;
6147 d->sd = alloc_percpu(struct sched_domain *);
6149 return sa_sd_storage;
6150 d->rd = alloc_rootdomain();
6153 return sa_rootdomain;
6157 * NULL the sd_data elements we've used to build the sched_domain and
6158 * sched_group structure so that the subsequent __free_domain_allocs()
6159 * will not free the data we're using.
6161 static void claim_allocations(int cpu, struct sched_domain *sd)
6163 struct sd_data *sdd = sd->private;
6165 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6166 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6168 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6169 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6171 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6172 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6176 static int sched_domains_numa_levels;
6177 enum numa_topology_type sched_numa_topology_type;
6178 static int *sched_domains_numa_distance;
6179 int sched_max_numa_distance;
6180 static struct cpumask ***sched_domains_numa_masks;
6181 static int sched_domains_curr_level;
6185 * SD_flags allowed in topology descriptions.
6187 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6188 * SD_SHARE_PKG_RESOURCES - describes shared caches
6189 * SD_NUMA - describes NUMA topologies
6190 * SD_SHARE_POWERDOMAIN - describes shared power domain
6193 * SD_ASYM_PACKING - describes SMT quirks
6195 #define TOPOLOGY_SD_FLAGS \
6196 (SD_SHARE_CPUCAPACITY | \
6197 SD_SHARE_PKG_RESOURCES | \
6200 SD_SHARE_POWERDOMAIN)
6202 static struct sched_domain *
6203 sd_init(struct sched_domain_topology_level *tl, int cpu)
6205 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6206 int sd_weight, sd_flags = 0;
6210 * Ugly hack to pass state to sd_numa_mask()...
6212 sched_domains_curr_level = tl->numa_level;
6215 sd_weight = cpumask_weight(tl->mask(cpu));
6218 sd_flags = (*tl->sd_flags)();
6219 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6220 "wrong sd_flags in topology description\n"))
6221 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6223 *sd = (struct sched_domain){
6224 .min_interval = sd_weight,
6225 .max_interval = 2*sd_weight,
6227 .imbalance_pct = 125,
6229 .cache_nice_tries = 0,
6236 .flags = 1*SD_LOAD_BALANCE
6237 | 1*SD_BALANCE_NEWIDLE
6242 | 0*SD_SHARE_CPUCAPACITY
6243 | 0*SD_SHARE_PKG_RESOURCES
6245 | 0*SD_PREFER_SIBLING
6250 .last_balance = jiffies,
6251 .balance_interval = sd_weight,
6253 .max_newidle_lb_cost = 0,
6254 .next_decay_max_lb_cost = jiffies,
6255 #ifdef CONFIG_SCHED_DEBUG
6261 * Convert topological properties into behaviour.
6264 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6265 sd->imbalance_pct = 110;
6266 sd->smt_gain = 1178; /* ~15% */
6268 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6269 sd->imbalance_pct = 117;
6270 sd->cache_nice_tries = 1;
6274 } else if (sd->flags & SD_NUMA) {
6275 sd->cache_nice_tries = 2;
6279 sd->flags |= SD_SERIALIZE;
6280 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6281 sd->flags &= ~(SD_BALANCE_EXEC |
6288 sd->flags |= SD_PREFER_SIBLING;
6289 sd->cache_nice_tries = 1;
6294 sd->private = &tl->data;
6300 * Topology list, bottom-up.
6302 static struct sched_domain_topology_level default_topology[] = {
6303 #ifdef CONFIG_SCHED_SMT
6304 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6306 #ifdef CONFIG_SCHED_MC
6307 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6309 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6313 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6315 #define for_each_sd_topology(tl) \
6316 for (tl = sched_domain_topology; tl->mask; tl++)
6318 void set_sched_topology(struct sched_domain_topology_level *tl)
6320 sched_domain_topology = tl;
6325 static const struct cpumask *sd_numa_mask(int cpu)
6327 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6330 static void sched_numa_warn(const char *str)
6332 static int done = false;
6340 printk(KERN_WARNING "ERROR: %s\n\n", str);
6342 for (i = 0; i < nr_node_ids; i++) {
6343 printk(KERN_WARNING " ");
6344 for (j = 0; j < nr_node_ids; j++)
6345 printk(KERN_CONT "%02d ", node_distance(i,j));
6346 printk(KERN_CONT "\n");
6348 printk(KERN_WARNING "\n");
6351 bool find_numa_distance(int distance)
6355 if (distance == node_distance(0, 0))
6358 for (i = 0; i < sched_domains_numa_levels; i++) {
6359 if (sched_domains_numa_distance[i] == distance)
6367 * A system can have three types of NUMA topology:
6368 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6369 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6370 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6372 * The difference between a glueless mesh topology and a backplane
6373 * topology lies in whether communication between not directly
6374 * connected nodes goes through intermediary nodes (where programs
6375 * could run), or through backplane controllers. This affects
6376 * placement of programs.
6378 * The type of topology can be discerned with the following tests:
6379 * - If the maximum distance between any nodes is 1 hop, the system
6380 * is directly connected.
6381 * - If for two nodes A and B, located N > 1 hops away from each other,
6382 * there is an intermediary node C, which is < N hops away from both
6383 * nodes A and B, the system is a glueless mesh.
6385 static void init_numa_topology_type(void)
6389 n = sched_max_numa_distance;
6392 sched_numa_topology_type = NUMA_DIRECT;
6394 for_each_online_node(a) {
6395 for_each_online_node(b) {
6396 /* Find two nodes furthest removed from each other. */
6397 if (node_distance(a, b) < n)
6400 /* Is there an intermediary node between a and b? */
6401 for_each_online_node(c) {
6402 if (node_distance(a, c) < n &&
6403 node_distance(b, c) < n) {
6404 sched_numa_topology_type =
6410 sched_numa_topology_type = NUMA_BACKPLANE;
6416 static void sched_init_numa(void)
6418 int next_distance, curr_distance = node_distance(0, 0);
6419 struct sched_domain_topology_level *tl;
6423 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6424 if (!sched_domains_numa_distance)
6428 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6429 * unique distances in the node_distance() table.
6431 * Assumes node_distance(0,j) includes all distances in
6432 * node_distance(i,j) in order to avoid cubic time.
6434 next_distance = curr_distance;
6435 for (i = 0; i < nr_node_ids; i++) {
6436 for (j = 0; j < nr_node_ids; j++) {
6437 for (k = 0; k < nr_node_ids; k++) {
6438 int distance = node_distance(i, k);
6440 if (distance > curr_distance &&
6441 (distance < next_distance ||
6442 next_distance == curr_distance))
6443 next_distance = distance;
6446 * While not a strong assumption it would be nice to know
6447 * about cases where if node A is connected to B, B is not
6448 * equally connected to A.
6450 if (sched_debug() && node_distance(k, i) != distance)
6451 sched_numa_warn("Node-distance not symmetric");
6453 if (sched_debug() && i && !find_numa_distance(distance))
6454 sched_numa_warn("Node-0 not representative");
6456 if (next_distance != curr_distance) {
6457 sched_domains_numa_distance[level++] = next_distance;
6458 sched_domains_numa_levels = level;
6459 curr_distance = next_distance;
6464 * In case of sched_debug() we verify the above assumption.
6470 * 'level' contains the number of unique distances, excluding the
6471 * identity distance node_distance(i,i).
6473 * The sched_domains_numa_distance[] array includes the actual distance
6478 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6479 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6480 * the array will contain less then 'level' members. This could be
6481 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6482 * in other functions.
6484 * We reset it to 'level' at the end of this function.
6486 sched_domains_numa_levels = 0;
6488 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6489 if (!sched_domains_numa_masks)
6493 * Now for each level, construct a mask per node which contains all
6494 * cpus of nodes that are that many hops away from us.
6496 for (i = 0; i < level; i++) {
6497 sched_domains_numa_masks[i] =
6498 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6499 if (!sched_domains_numa_masks[i])
6502 for (j = 0; j < nr_node_ids; j++) {
6503 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6507 sched_domains_numa_masks[i][j] = mask;
6509 for (k = 0; k < nr_node_ids; k++) {
6510 if (node_distance(j, k) > sched_domains_numa_distance[i])
6513 cpumask_or(mask, mask, cpumask_of_node(k));
6518 /* Compute default topology size */
6519 for (i = 0; sched_domain_topology[i].mask; i++);
6521 tl = kzalloc((i + level + 1) *
6522 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6527 * Copy the default topology bits..
6529 for (i = 0; sched_domain_topology[i].mask; i++)
6530 tl[i] = sched_domain_topology[i];
6533 * .. and append 'j' levels of NUMA goodness.
6535 for (j = 0; j < level; i++, j++) {
6536 tl[i] = (struct sched_domain_topology_level){
6537 .mask = sd_numa_mask,
6538 .sd_flags = cpu_numa_flags,
6539 .flags = SDTL_OVERLAP,
6545 sched_domain_topology = tl;
6547 sched_domains_numa_levels = level;
6548 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6550 init_numa_topology_type();
6553 static void sched_domains_numa_masks_set(int cpu)
6556 int node = cpu_to_node(cpu);
6558 for (i = 0; i < sched_domains_numa_levels; i++) {
6559 for (j = 0; j < nr_node_ids; j++) {
6560 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6561 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6566 static void sched_domains_numa_masks_clear(int cpu)
6569 for (i = 0; i < sched_domains_numa_levels; i++) {
6570 for (j = 0; j < nr_node_ids; j++)
6571 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6576 * Update sched_domains_numa_masks[level][node] array when new cpus
6579 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6580 unsigned long action,
6583 int cpu = (long)hcpu;
6585 switch (action & ~CPU_TASKS_FROZEN) {
6587 sched_domains_numa_masks_set(cpu);
6591 sched_domains_numa_masks_clear(cpu);
6601 static inline void sched_init_numa(void)
6605 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6606 unsigned long action,
6611 #endif /* CONFIG_NUMA */
6613 static int __sdt_alloc(const struct cpumask *cpu_map)
6615 struct sched_domain_topology_level *tl;
6618 for_each_sd_topology(tl) {
6619 struct sd_data *sdd = &tl->data;
6621 sdd->sd = alloc_percpu(struct sched_domain *);
6625 sdd->sg = alloc_percpu(struct sched_group *);
6629 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6633 for_each_cpu(j, cpu_map) {
6634 struct sched_domain *sd;
6635 struct sched_group *sg;
6636 struct sched_group_capacity *sgc;
6638 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6639 GFP_KERNEL, cpu_to_node(j));
6643 *per_cpu_ptr(sdd->sd, j) = sd;
6645 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6646 GFP_KERNEL, cpu_to_node(j));
6652 *per_cpu_ptr(sdd->sg, j) = sg;
6654 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6655 GFP_KERNEL, cpu_to_node(j));
6659 *per_cpu_ptr(sdd->sgc, j) = sgc;
6666 static void __sdt_free(const struct cpumask *cpu_map)
6668 struct sched_domain_topology_level *tl;
6671 for_each_sd_topology(tl) {
6672 struct sd_data *sdd = &tl->data;
6674 for_each_cpu(j, cpu_map) {
6675 struct sched_domain *sd;
6678 sd = *per_cpu_ptr(sdd->sd, j);
6679 if (sd && (sd->flags & SD_OVERLAP))
6680 free_sched_groups(sd->groups, 0);
6681 kfree(*per_cpu_ptr(sdd->sd, j));
6685 kfree(*per_cpu_ptr(sdd->sg, j));
6687 kfree(*per_cpu_ptr(sdd->sgc, j));
6689 free_percpu(sdd->sd);
6691 free_percpu(sdd->sg);
6693 free_percpu(sdd->sgc);
6698 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6699 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6700 struct sched_domain *child, int cpu)
6702 struct sched_domain *sd = sd_init(tl, cpu);
6706 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6708 sd->level = child->level + 1;
6709 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6713 if (!cpumask_subset(sched_domain_span(child),
6714 sched_domain_span(sd))) {
6715 pr_err("BUG: arch topology borken\n");
6716 #ifdef CONFIG_SCHED_DEBUG
6717 pr_err(" the %s domain not a subset of the %s domain\n",
6718 child->name, sd->name);
6720 /* Fixup, ensure @sd has at least @child cpus. */
6721 cpumask_or(sched_domain_span(sd),
6722 sched_domain_span(sd),
6723 sched_domain_span(child));
6727 set_domain_attribute(sd, attr);
6733 * Build sched domains for a given set of cpus and attach the sched domains
6734 * to the individual cpus
6736 static int build_sched_domains(const struct cpumask *cpu_map,
6737 struct sched_domain_attr *attr)
6739 enum s_alloc alloc_state;
6740 struct sched_domain *sd;
6742 int i, ret = -ENOMEM;
6744 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6745 if (alloc_state != sa_rootdomain)
6748 /* Set up domains for cpus specified by the cpu_map. */
6749 for_each_cpu(i, cpu_map) {
6750 struct sched_domain_topology_level *tl;
6753 for_each_sd_topology(tl) {
6754 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6755 if (tl == sched_domain_topology)
6756 *per_cpu_ptr(d.sd, i) = sd;
6757 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6758 sd->flags |= SD_OVERLAP;
6759 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6764 /* Build the groups for the domains */
6765 for_each_cpu(i, cpu_map) {
6766 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6767 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6768 if (sd->flags & SD_OVERLAP) {
6769 if (build_overlap_sched_groups(sd, i))
6772 if (build_sched_groups(sd, i))
6778 /* Calculate CPU capacity for physical packages and nodes */
6779 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6780 if (!cpumask_test_cpu(i, cpu_map))
6783 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6784 claim_allocations(i, sd);
6785 init_sched_groups_capacity(i, sd);
6789 /* Attach the domains */
6791 for_each_cpu(i, cpu_map) {
6792 sd = *per_cpu_ptr(d.sd, i);
6793 cpu_attach_domain(sd, d.rd, i);
6799 __free_domain_allocs(&d, alloc_state, cpu_map);
6803 static cpumask_var_t *doms_cur; /* current sched domains */
6804 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6805 static struct sched_domain_attr *dattr_cur;
6806 /* attribues of custom domains in 'doms_cur' */
6809 * Special case: If a kmalloc of a doms_cur partition (array of
6810 * cpumask) fails, then fallback to a single sched domain,
6811 * as determined by the single cpumask fallback_doms.
6813 static cpumask_var_t fallback_doms;
6816 * arch_update_cpu_topology lets virtualized architectures update the
6817 * cpu core maps. It is supposed to return 1 if the topology changed
6818 * or 0 if it stayed the same.
6820 int __weak arch_update_cpu_topology(void)
6825 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6828 cpumask_var_t *doms;
6830 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6833 for (i = 0; i < ndoms; i++) {
6834 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6835 free_sched_domains(doms, i);
6842 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6845 for (i = 0; i < ndoms; i++)
6846 free_cpumask_var(doms[i]);
6851 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6852 * For now this just excludes isolated cpus, but could be used to
6853 * exclude other special cases in the future.
6855 static int init_sched_domains(const struct cpumask *cpu_map)
6859 arch_update_cpu_topology();
6861 doms_cur = alloc_sched_domains(ndoms_cur);
6863 doms_cur = &fallback_doms;
6864 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6865 err = build_sched_domains(doms_cur[0], NULL);
6866 register_sched_domain_sysctl();
6872 * Detach sched domains from a group of cpus specified in cpu_map
6873 * These cpus will now be attached to the NULL domain
6875 static void detach_destroy_domains(const struct cpumask *cpu_map)
6880 for_each_cpu(i, cpu_map)
6881 cpu_attach_domain(NULL, &def_root_domain, i);
6885 /* handle null as "default" */
6886 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6887 struct sched_domain_attr *new, int idx_new)
6889 struct sched_domain_attr tmp;
6896 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6897 new ? (new + idx_new) : &tmp,
6898 sizeof(struct sched_domain_attr));
6902 * Partition sched domains as specified by the 'ndoms_new'
6903 * cpumasks in the array doms_new[] of cpumasks. This compares
6904 * doms_new[] to the current sched domain partitioning, doms_cur[].
6905 * It destroys each deleted domain and builds each new domain.
6907 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6908 * The masks don't intersect (don't overlap.) We should setup one
6909 * sched domain for each mask. CPUs not in any of the cpumasks will
6910 * not be load balanced. If the same cpumask appears both in the
6911 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6914 * The passed in 'doms_new' should be allocated using
6915 * alloc_sched_domains. This routine takes ownership of it and will
6916 * free_sched_domains it when done with it. If the caller failed the
6917 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6918 * and partition_sched_domains() will fallback to the single partition
6919 * 'fallback_doms', it also forces the domains to be rebuilt.
6921 * If doms_new == NULL it will be replaced with cpu_online_mask.
6922 * ndoms_new == 0 is a special case for destroying existing domains,
6923 * and it will not create the default domain.
6925 * Call with hotplug lock held
6927 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6928 struct sched_domain_attr *dattr_new)
6933 mutex_lock(&sched_domains_mutex);
6935 /* always unregister in case we don't destroy any domains */
6936 unregister_sched_domain_sysctl();
6938 /* Let architecture update cpu core mappings. */
6939 new_topology = arch_update_cpu_topology();
6941 n = doms_new ? ndoms_new : 0;
6943 /* Destroy deleted domains */
6944 for (i = 0; i < ndoms_cur; i++) {
6945 for (j = 0; j < n && !new_topology; j++) {
6946 if (cpumask_equal(doms_cur[i], doms_new[j])
6947 && dattrs_equal(dattr_cur, i, dattr_new, j))
6950 /* no match - a current sched domain not in new doms_new[] */
6951 detach_destroy_domains(doms_cur[i]);
6957 if (doms_new == NULL) {
6959 doms_new = &fallback_doms;
6960 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6961 WARN_ON_ONCE(dattr_new);
6964 /* Build new domains */
6965 for (i = 0; i < ndoms_new; i++) {
6966 for (j = 0; j < n && !new_topology; j++) {
6967 if (cpumask_equal(doms_new[i], doms_cur[j])
6968 && dattrs_equal(dattr_new, i, dattr_cur, j))
6971 /* no match - add a new doms_new */
6972 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6977 /* Remember the new sched domains */
6978 if (doms_cur != &fallback_doms)
6979 free_sched_domains(doms_cur, ndoms_cur);
6980 kfree(dattr_cur); /* kfree(NULL) is safe */
6981 doms_cur = doms_new;
6982 dattr_cur = dattr_new;
6983 ndoms_cur = ndoms_new;
6985 register_sched_domain_sysctl();
6987 mutex_unlock(&sched_domains_mutex);
6990 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6993 * Update cpusets according to cpu_active mask. If cpusets are
6994 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6995 * around partition_sched_domains().
6997 * If we come here as part of a suspend/resume, don't touch cpusets because we
6998 * want to restore it back to its original state upon resume anyway.
7000 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7004 case CPU_ONLINE_FROZEN:
7005 case CPU_DOWN_FAILED_FROZEN:
7008 * num_cpus_frozen tracks how many CPUs are involved in suspend
7009 * resume sequence. As long as this is not the last online
7010 * operation in the resume sequence, just build a single sched
7011 * domain, ignoring cpusets.
7014 if (likely(num_cpus_frozen)) {
7015 partition_sched_domains(1, NULL, NULL);
7020 * This is the last CPU online operation. So fall through and
7021 * restore the original sched domains by considering the
7022 * cpuset configurations.
7026 case CPU_DOWN_FAILED:
7027 cpuset_update_active_cpus(true);
7035 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7039 case CPU_DOWN_PREPARE:
7040 cpuset_update_active_cpus(false);
7042 case CPU_DOWN_PREPARE_FROZEN:
7044 partition_sched_domains(1, NULL, NULL);
7052 void __init sched_init_smp(void)
7054 cpumask_var_t non_isolated_cpus;
7056 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7057 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7062 * There's no userspace yet to cause hotplug operations; hence all the
7063 * cpu masks are stable and all blatant races in the below code cannot
7066 mutex_lock(&sched_domains_mutex);
7067 init_sched_domains(cpu_active_mask);
7068 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7069 if (cpumask_empty(non_isolated_cpus))
7070 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7071 mutex_unlock(&sched_domains_mutex);
7073 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7074 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7075 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7079 /* Move init over to a non-isolated CPU */
7080 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7082 sched_init_granularity();
7083 free_cpumask_var(non_isolated_cpus);
7085 init_sched_rt_class();
7086 init_sched_dl_class();
7089 void __init sched_init_smp(void)
7091 sched_init_granularity();
7093 #endif /* CONFIG_SMP */
7095 const_debug unsigned int sysctl_timer_migration = 1;
7097 int in_sched_functions(unsigned long addr)
7099 return in_lock_functions(addr) ||
7100 (addr >= (unsigned long)__sched_text_start
7101 && addr < (unsigned long)__sched_text_end);
7104 #ifdef CONFIG_CGROUP_SCHED
7106 * Default task group.
7107 * Every task in system belongs to this group at bootup.
7109 struct task_group root_task_group;
7110 LIST_HEAD(task_groups);
7113 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7115 void __init sched_init(void)
7118 unsigned long alloc_size = 0, ptr;
7120 #ifdef CONFIG_FAIR_GROUP_SCHED
7121 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7123 #ifdef CONFIG_RT_GROUP_SCHED
7124 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7126 #ifdef CONFIG_CPUMASK_OFFSTACK
7127 alloc_size += num_possible_cpus() * cpumask_size();
7130 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7132 #ifdef CONFIG_FAIR_GROUP_SCHED
7133 root_task_group.se = (struct sched_entity **)ptr;
7134 ptr += nr_cpu_ids * sizeof(void **);
7136 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7137 ptr += nr_cpu_ids * sizeof(void **);
7139 #endif /* CONFIG_FAIR_GROUP_SCHED */
7140 #ifdef CONFIG_RT_GROUP_SCHED
7141 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7142 ptr += nr_cpu_ids * sizeof(void **);
7144 root_task_group.rt_rq = (struct rt_rq **)ptr;
7145 ptr += nr_cpu_ids * sizeof(void **);
7147 #endif /* CONFIG_RT_GROUP_SCHED */
7148 #ifdef CONFIG_CPUMASK_OFFSTACK
7149 for_each_possible_cpu(i) {
7150 per_cpu(load_balance_mask, i) = (void *)ptr;
7151 ptr += cpumask_size();
7153 #endif /* CONFIG_CPUMASK_OFFSTACK */
7156 init_rt_bandwidth(&def_rt_bandwidth,
7157 global_rt_period(), global_rt_runtime());
7158 init_dl_bandwidth(&def_dl_bandwidth,
7159 global_rt_period(), global_rt_runtime());
7162 init_defrootdomain();
7165 #ifdef CONFIG_RT_GROUP_SCHED
7166 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7167 global_rt_period(), global_rt_runtime());
7168 #endif /* CONFIG_RT_GROUP_SCHED */
7170 #ifdef CONFIG_CGROUP_SCHED
7171 list_add(&root_task_group.list, &task_groups);
7172 INIT_LIST_HEAD(&root_task_group.children);
7173 INIT_LIST_HEAD(&root_task_group.siblings);
7174 autogroup_init(&init_task);
7176 #endif /* CONFIG_CGROUP_SCHED */
7178 for_each_possible_cpu(i) {
7182 raw_spin_lock_init(&rq->lock);
7184 rq->calc_load_active = 0;
7185 rq->calc_load_update = jiffies + LOAD_FREQ;
7186 init_cfs_rq(&rq->cfs);
7187 init_rt_rq(&rq->rt, rq);
7188 init_dl_rq(&rq->dl, rq);
7189 #ifdef CONFIG_FAIR_GROUP_SCHED
7190 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7191 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7193 * How much cpu bandwidth does root_task_group get?
7195 * In case of task-groups formed thr' the cgroup filesystem, it
7196 * gets 100% of the cpu resources in the system. This overall
7197 * system cpu resource is divided among the tasks of
7198 * root_task_group and its child task-groups in a fair manner,
7199 * based on each entity's (task or task-group's) weight
7200 * (se->load.weight).
7202 * In other words, if root_task_group has 10 tasks of weight
7203 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7204 * then A0's share of the cpu resource is:
7206 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7208 * We achieve this by letting root_task_group's tasks sit
7209 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7211 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7212 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7213 #endif /* CONFIG_FAIR_GROUP_SCHED */
7215 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7216 #ifdef CONFIG_RT_GROUP_SCHED
7217 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7220 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7221 rq->cpu_load[j] = 0;
7223 rq->last_load_update_tick = jiffies;
7228 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7229 rq->post_schedule = 0;
7230 rq->active_balance = 0;
7231 rq->next_balance = jiffies;
7236 rq->avg_idle = 2*sysctl_sched_migration_cost;
7237 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7239 INIT_LIST_HEAD(&rq->cfs_tasks);
7241 rq_attach_root(rq, &def_root_domain);
7242 #ifdef CONFIG_NO_HZ_COMMON
7245 #ifdef CONFIG_NO_HZ_FULL
7246 rq->last_sched_tick = 0;
7250 atomic_set(&rq->nr_iowait, 0);
7253 set_load_weight(&init_task);
7255 #ifdef CONFIG_PREEMPT_NOTIFIERS
7256 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7260 * The boot idle thread does lazy MMU switching as well:
7262 atomic_inc(&init_mm.mm_count);
7263 enter_lazy_tlb(&init_mm, current);
7266 * Make us the idle thread. Technically, schedule() should not be
7267 * called from this thread, however somewhere below it might be,
7268 * but because we are the idle thread, we just pick up running again
7269 * when this runqueue becomes "idle".
7271 init_idle(current, smp_processor_id());
7273 calc_load_update = jiffies + LOAD_FREQ;
7276 * During early bootup we pretend to be a normal task:
7278 current->sched_class = &fair_sched_class;
7281 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7282 /* May be allocated at isolcpus cmdline parse time */
7283 if (cpu_isolated_map == NULL)
7284 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7285 idle_thread_set_boot_cpu();
7286 set_cpu_rq_start_time();
7288 init_sched_fair_class();
7290 scheduler_running = 1;
7293 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7294 static inline int preempt_count_equals(int preempt_offset)
7296 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7298 return (nested == preempt_offset);
7301 void __might_sleep(const char *file, int line, int preempt_offset)
7304 * Blocking primitives will set (and therefore destroy) current->state,
7305 * since we will exit with TASK_RUNNING make sure we enter with it,
7306 * otherwise we will destroy state.
7308 if (WARN_ONCE(current->state != TASK_RUNNING,
7309 "do not call blocking ops when !TASK_RUNNING; "
7310 "state=%lx set at [<%p>] %pS\n",
7312 (void *)current->task_state_change,
7313 (void *)current->task_state_change))
7314 __set_current_state(TASK_RUNNING);
7316 ___might_sleep(file, line, preempt_offset);
7318 EXPORT_SYMBOL(__might_sleep);
7320 void ___might_sleep(const char *file, int line, int preempt_offset)
7322 static unsigned long prev_jiffy; /* ratelimiting */
7324 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7325 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7326 !is_idle_task(current)) ||
7327 system_state != SYSTEM_RUNNING || oops_in_progress)
7329 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7331 prev_jiffy = jiffies;
7334 "BUG: sleeping function called from invalid context at %s:%d\n",
7337 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7338 in_atomic(), irqs_disabled(),
7339 current->pid, current->comm);
7341 debug_show_held_locks(current);
7342 if (irqs_disabled())
7343 print_irqtrace_events(current);
7344 #ifdef CONFIG_DEBUG_PREEMPT
7345 if (!preempt_count_equals(preempt_offset)) {
7346 pr_err("Preemption disabled at:");
7347 print_ip_sym(current->preempt_disable_ip);
7353 EXPORT_SYMBOL(___might_sleep);
7356 #ifdef CONFIG_MAGIC_SYSRQ
7357 static void normalize_task(struct rq *rq, struct task_struct *p)
7359 const struct sched_class *prev_class = p->sched_class;
7360 struct sched_attr attr = {
7361 .sched_policy = SCHED_NORMAL,
7363 int old_prio = p->prio;
7366 queued = task_on_rq_queued(p);
7368 dequeue_task(rq, p, 0);
7369 __setscheduler(rq, p, &attr);
7371 enqueue_task(rq, p, 0);
7375 check_class_changed(rq, p, prev_class, old_prio);
7378 void normalize_rt_tasks(void)
7380 struct task_struct *g, *p;
7381 unsigned long flags;
7384 read_lock(&tasklist_lock);
7385 for_each_process_thread(g, p) {
7387 * Only normalize user tasks:
7389 if (p->flags & PF_KTHREAD)
7392 p->se.exec_start = 0;
7393 #ifdef CONFIG_SCHEDSTATS
7394 p->se.statistics.wait_start = 0;
7395 p->se.statistics.sleep_start = 0;
7396 p->se.statistics.block_start = 0;
7399 if (!dl_task(p) && !rt_task(p)) {
7401 * Renice negative nice level userspace
7404 if (task_nice(p) < 0)
7405 set_user_nice(p, 0);
7409 rq = task_rq_lock(p, &flags);
7410 normalize_task(rq, p);
7411 task_rq_unlock(rq, p, &flags);
7413 read_unlock(&tasklist_lock);
7416 #endif /* CONFIG_MAGIC_SYSRQ */
7418 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7420 * These functions are only useful for the IA64 MCA handling, or kdb.
7422 * They can only be called when the whole system has been
7423 * stopped - every CPU needs to be quiescent, and no scheduling
7424 * activity can take place. Using them for anything else would
7425 * be a serious bug, and as a result, they aren't even visible
7426 * under any other configuration.
7430 * curr_task - return the current task for a given cpu.
7431 * @cpu: the processor in question.
7433 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7435 * Return: The current task for @cpu.
7437 struct task_struct *curr_task(int cpu)
7439 return cpu_curr(cpu);
7442 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7446 * set_curr_task - set the current task for a given cpu.
7447 * @cpu: the processor in question.
7448 * @p: the task pointer to set.
7450 * Description: This function must only be used when non-maskable interrupts
7451 * are serviced on a separate stack. It allows the architecture to switch the
7452 * notion of the current task on a cpu in a non-blocking manner. This function
7453 * must be called with all CPU's synchronized, and interrupts disabled, the
7454 * and caller must save the original value of the current task (see
7455 * curr_task() above) and restore that value before reenabling interrupts and
7456 * re-starting the system.
7458 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7460 void set_curr_task(int cpu, struct task_struct *p)
7467 #ifdef CONFIG_CGROUP_SCHED
7468 /* task_group_lock serializes the addition/removal of task groups */
7469 static DEFINE_SPINLOCK(task_group_lock);
7471 static void free_sched_group(struct task_group *tg)
7473 free_fair_sched_group(tg);
7474 free_rt_sched_group(tg);
7479 /* allocate runqueue etc for a new task group */
7480 struct task_group *sched_create_group(struct task_group *parent)
7482 struct task_group *tg;
7484 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7486 return ERR_PTR(-ENOMEM);
7488 if (!alloc_fair_sched_group(tg, parent))
7491 if (!alloc_rt_sched_group(tg, parent))
7497 free_sched_group(tg);
7498 return ERR_PTR(-ENOMEM);
7501 void sched_online_group(struct task_group *tg, struct task_group *parent)
7503 unsigned long flags;
7505 spin_lock_irqsave(&task_group_lock, flags);
7506 list_add_rcu(&tg->list, &task_groups);
7508 WARN_ON(!parent); /* root should already exist */
7510 tg->parent = parent;
7511 INIT_LIST_HEAD(&tg->children);
7512 list_add_rcu(&tg->siblings, &parent->children);
7513 spin_unlock_irqrestore(&task_group_lock, flags);
7516 /* rcu callback to free various structures associated with a task group */
7517 static void free_sched_group_rcu(struct rcu_head *rhp)
7519 /* now it should be safe to free those cfs_rqs */
7520 free_sched_group(container_of(rhp, struct task_group, rcu));
7523 /* Destroy runqueue etc associated with a task group */
7524 void sched_destroy_group(struct task_group *tg)
7526 /* wait for possible concurrent references to cfs_rqs complete */
7527 call_rcu(&tg->rcu, free_sched_group_rcu);
7530 void sched_offline_group(struct task_group *tg)
7532 unsigned long flags;
7535 /* end participation in shares distribution */
7536 for_each_possible_cpu(i)
7537 unregister_fair_sched_group(tg, i);
7539 spin_lock_irqsave(&task_group_lock, flags);
7540 list_del_rcu(&tg->list);
7541 list_del_rcu(&tg->siblings);
7542 spin_unlock_irqrestore(&task_group_lock, flags);
7545 /* change task's runqueue when it moves between groups.
7546 * The caller of this function should have put the task in its new group
7547 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7548 * reflect its new group.
7550 void sched_move_task(struct task_struct *tsk)
7552 struct task_group *tg;
7553 int queued, running;
7554 unsigned long flags;
7557 rq = task_rq_lock(tsk, &flags);
7559 running = task_current(rq, tsk);
7560 queued = task_on_rq_queued(tsk);
7563 dequeue_task(rq, tsk, 0);
7564 if (unlikely(running))
7565 put_prev_task(rq, tsk);
7567 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7568 lockdep_is_held(&tsk->sighand->siglock)),
7569 struct task_group, css);
7570 tg = autogroup_task_group(tsk, tg);
7571 tsk->sched_task_group = tg;
7573 #ifdef CONFIG_FAIR_GROUP_SCHED
7574 if (tsk->sched_class->task_move_group)
7575 tsk->sched_class->task_move_group(tsk, queued);
7578 set_task_rq(tsk, task_cpu(tsk));
7580 if (unlikely(running))
7581 tsk->sched_class->set_curr_task(rq);
7583 enqueue_task(rq, tsk, 0);
7585 task_rq_unlock(rq, tsk, &flags);
7587 #endif /* CONFIG_CGROUP_SCHED */
7589 #ifdef CONFIG_RT_GROUP_SCHED
7591 * Ensure that the real time constraints are schedulable.
7593 static DEFINE_MUTEX(rt_constraints_mutex);
7595 /* Must be called with tasklist_lock held */
7596 static inline int tg_has_rt_tasks(struct task_group *tg)
7598 struct task_struct *g, *p;
7600 for_each_process_thread(g, p) {
7601 if (rt_task(p) && task_group(p) == tg)
7608 struct rt_schedulable_data {
7609 struct task_group *tg;
7614 static int tg_rt_schedulable(struct task_group *tg, void *data)
7616 struct rt_schedulable_data *d = data;
7617 struct task_group *child;
7618 unsigned long total, sum = 0;
7619 u64 period, runtime;
7621 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7622 runtime = tg->rt_bandwidth.rt_runtime;
7625 period = d->rt_period;
7626 runtime = d->rt_runtime;
7630 * Cannot have more runtime than the period.
7632 if (runtime > period && runtime != RUNTIME_INF)
7636 * Ensure we don't starve existing RT tasks.
7638 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7641 total = to_ratio(period, runtime);
7644 * Nobody can have more than the global setting allows.
7646 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7650 * The sum of our children's runtime should not exceed our own.
7652 list_for_each_entry_rcu(child, &tg->children, siblings) {
7653 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7654 runtime = child->rt_bandwidth.rt_runtime;
7656 if (child == d->tg) {
7657 period = d->rt_period;
7658 runtime = d->rt_runtime;
7661 sum += to_ratio(period, runtime);
7670 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7674 struct rt_schedulable_data data = {
7676 .rt_period = period,
7677 .rt_runtime = runtime,
7681 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7687 static int tg_set_rt_bandwidth(struct task_group *tg,
7688 u64 rt_period, u64 rt_runtime)
7692 mutex_lock(&rt_constraints_mutex);
7693 read_lock(&tasklist_lock);
7694 err = __rt_schedulable(tg, rt_period, rt_runtime);
7698 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7699 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7700 tg->rt_bandwidth.rt_runtime = rt_runtime;
7702 for_each_possible_cpu(i) {
7703 struct rt_rq *rt_rq = tg->rt_rq[i];
7705 raw_spin_lock(&rt_rq->rt_runtime_lock);
7706 rt_rq->rt_runtime = rt_runtime;
7707 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7709 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7711 read_unlock(&tasklist_lock);
7712 mutex_unlock(&rt_constraints_mutex);
7717 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7719 u64 rt_runtime, rt_period;
7721 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7722 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7723 if (rt_runtime_us < 0)
7724 rt_runtime = RUNTIME_INF;
7726 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7729 static long sched_group_rt_runtime(struct task_group *tg)
7733 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7736 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7737 do_div(rt_runtime_us, NSEC_PER_USEC);
7738 return rt_runtime_us;
7741 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7743 u64 rt_runtime, rt_period;
7745 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7746 rt_runtime = tg->rt_bandwidth.rt_runtime;
7751 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7754 static long sched_group_rt_period(struct task_group *tg)
7758 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7759 do_div(rt_period_us, NSEC_PER_USEC);
7760 return rt_period_us;
7762 #endif /* CONFIG_RT_GROUP_SCHED */
7764 #ifdef CONFIG_RT_GROUP_SCHED
7765 static int sched_rt_global_constraints(void)
7769 mutex_lock(&rt_constraints_mutex);
7770 read_lock(&tasklist_lock);
7771 ret = __rt_schedulable(NULL, 0, 0);
7772 read_unlock(&tasklist_lock);
7773 mutex_unlock(&rt_constraints_mutex);
7778 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7780 /* Don't accept realtime tasks when there is no way for them to run */
7781 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7787 #else /* !CONFIG_RT_GROUP_SCHED */
7788 static int sched_rt_global_constraints(void)
7790 unsigned long flags;
7793 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7794 for_each_possible_cpu(i) {
7795 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7797 raw_spin_lock(&rt_rq->rt_runtime_lock);
7798 rt_rq->rt_runtime = global_rt_runtime();
7799 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7801 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7805 #endif /* CONFIG_RT_GROUP_SCHED */
7807 static int sched_dl_global_constraints(void)
7809 u64 runtime = global_rt_runtime();
7810 u64 period = global_rt_period();
7811 u64 new_bw = to_ratio(period, runtime);
7814 unsigned long flags;
7817 * Here we want to check the bandwidth not being set to some
7818 * value smaller than the currently allocated bandwidth in
7819 * any of the root_domains.
7821 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7822 * cycling on root_domains... Discussion on different/better
7823 * solutions is welcome!
7825 for_each_possible_cpu(cpu) {
7826 rcu_read_lock_sched();
7827 dl_b = dl_bw_of(cpu);
7829 raw_spin_lock_irqsave(&dl_b->lock, flags);
7830 if (new_bw < dl_b->total_bw)
7832 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7834 rcu_read_unlock_sched();
7843 static void sched_dl_do_global(void)
7848 unsigned long flags;
7850 def_dl_bandwidth.dl_period = global_rt_period();
7851 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7853 if (global_rt_runtime() != RUNTIME_INF)
7854 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7857 * FIXME: As above...
7859 for_each_possible_cpu(cpu) {
7860 rcu_read_lock_sched();
7861 dl_b = dl_bw_of(cpu);
7863 raw_spin_lock_irqsave(&dl_b->lock, flags);
7865 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7867 rcu_read_unlock_sched();
7871 static int sched_rt_global_validate(void)
7873 if (sysctl_sched_rt_period <= 0)
7876 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7877 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7883 static void sched_rt_do_global(void)
7885 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7886 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7889 int sched_rt_handler(struct ctl_table *table, int write,
7890 void __user *buffer, size_t *lenp,
7893 int old_period, old_runtime;
7894 static DEFINE_MUTEX(mutex);
7898 old_period = sysctl_sched_rt_period;
7899 old_runtime = sysctl_sched_rt_runtime;
7901 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7903 if (!ret && write) {
7904 ret = sched_rt_global_validate();
7908 ret = sched_rt_global_constraints();
7912 ret = sched_dl_global_constraints();
7916 sched_rt_do_global();
7917 sched_dl_do_global();
7921 sysctl_sched_rt_period = old_period;
7922 sysctl_sched_rt_runtime = old_runtime;
7924 mutex_unlock(&mutex);
7929 int sched_rr_handler(struct ctl_table *table, int write,
7930 void __user *buffer, size_t *lenp,
7934 static DEFINE_MUTEX(mutex);
7937 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7938 /* make sure that internally we keep jiffies */
7939 /* also, writing zero resets timeslice to default */
7940 if (!ret && write) {
7941 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7942 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7944 mutex_unlock(&mutex);
7948 #ifdef CONFIG_CGROUP_SCHED
7950 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7952 return css ? container_of(css, struct task_group, css) : NULL;
7955 static struct cgroup_subsys_state *
7956 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7958 struct task_group *parent = css_tg(parent_css);
7959 struct task_group *tg;
7962 /* This is early initialization for the top cgroup */
7963 return &root_task_group.css;
7966 tg = sched_create_group(parent);
7968 return ERR_PTR(-ENOMEM);
7973 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7975 struct task_group *tg = css_tg(css);
7976 struct task_group *parent = css_tg(css->parent);
7979 sched_online_group(tg, parent);
7983 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7985 struct task_group *tg = css_tg(css);
7987 sched_destroy_group(tg);
7990 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7992 struct task_group *tg = css_tg(css);
7994 sched_offline_group(tg);
7997 static void cpu_cgroup_fork(struct task_struct *task)
7999 sched_move_task(task);
8002 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
8003 struct cgroup_taskset *tset)
8005 struct task_struct *task;
8007 cgroup_taskset_for_each(task, tset) {
8008 #ifdef CONFIG_RT_GROUP_SCHED
8009 if (!sched_rt_can_attach(css_tg(css), task))
8012 /* We don't support RT-tasks being in separate groups */
8013 if (task->sched_class != &fair_sched_class)
8020 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8021 struct cgroup_taskset *tset)
8023 struct task_struct *task;
8025 cgroup_taskset_for_each(task, tset)
8026 sched_move_task(task);
8029 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8030 struct cgroup_subsys_state *old_css,
8031 struct task_struct *task)
8034 * cgroup_exit() is called in the copy_process() failure path.
8035 * Ignore this case since the task hasn't ran yet, this avoids
8036 * trying to poke a half freed task state from generic code.
8038 if (!(task->flags & PF_EXITING))
8041 sched_move_task(task);
8044 #ifdef CONFIG_FAIR_GROUP_SCHED
8045 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8046 struct cftype *cftype, u64 shareval)
8048 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8051 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8054 struct task_group *tg = css_tg(css);
8056 return (u64) scale_load_down(tg->shares);
8059 #ifdef CONFIG_CFS_BANDWIDTH
8060 static DEFINE_MUTEX(cfs_constraints_mutex);
8062 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8063 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8065 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8067 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8069 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8070 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8072 if (tg == &root_task_group)
8076 * Ensure we have at some amount of bandwidth every period. This is
8077 * to prevent reaching a state of large arrears when throttled via
8078 * entity_tick() resulting in prolonged exit starvation.
8080 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8084 * Likewise, bound things on the otherside by preventing insane quota
8085 * periods. This also allows us to normalize in computing quota
8088 if (period > max_cfs_quota_period)
8092 * Prevent race between setting of cfs_rq->runtime_enabled and
8093 * unthrottle_offline_cfs_rqs().
8096 mutex_lock(&cfs_constraints_mutex);
8097 ret = __cfs_schedulable(tg, period, quota);
8101 runtime_enabled = quota != RUNTIME_INF;
8102 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8104 * If we need to toggle cfs_bandwidth_used, off->on must occur
8105 * before making related changes, and on->off must occur afterwards
8107 if (runtime_enabled && !runtime_was_enabled)
8108 cfs_bandwidth_usage_inc();
8109 raw_spin_lock_irq(&cfs_b->lock);
8110 cfs_b->period = ns_to_ktime(period);
8111 cfs_b->quota = quota;
8113 __refill_cfs_bandwidth_runtime(cfs_b);
8114 /* restart the period timer (if active) to handle new period expiry */
8115 if (runtime_enabled && cfs_b->timer_active) {
8116 /* force a reprogram */
8117 __start_cfs_bandwidth(cfs_b, true);
8119 raw_spin_unlock_irq(&cfs_b->lock);
8121 for_each_online_cpu(i) {
8122 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8123 struct rq *rq = cfs_rq->rq;
8125 raw_spin_lock_irq(&rq->lock);
8126 cfs_rq->runtime_enabled = runtime_enabled;
8127 cfs_rq->runtime_remaining = 0;
8129 if (cfs_rq->throttled)
8130 unthrottle_cfs_rq(cfs_rq);
8131 raw_spin_unlock_irq(&rq->lock);
8133 if (runtime_was_enabled && !runtime_enabled)
8134 cfs_bandwidth_usage_dec();
8136 mutex_unlock(&cfs_constraints_mutex);
8142 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8146 period = ktime_to_ns(tg->cfs_bandwidth.period);
8147 if (cfs_quota_us < 0)
8148 quota = RUNTIME_INF;
8150 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8152 return tg_set_cfs_bandwidth(tg, period, quota);
8155 long tg_get_cfs_quota(struct task_group *tg)
8159 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8162 quota_us = tg->cfs_bandwidth.quota;
8163 do_div(quota_us, NSEC_PER_USEC);
8168 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8172 period = (u64)cfs_period_us * NSEC_PER_USEC;
8173 quota = tg->cfs_bandwidth.quota;
8175 return tg_set_cfs_bandwidth(tg, period, quota);
8178 long tg_get_cfs_period(struct task_group *tg)
8182 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8183 do_div(cfs_period_us, NSEC_PER_USEC);
8185 return cfs_period_us;
8188 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8191 return tg_get_cfs_quota(css_tg(css));
8194 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8195 struct cftype *cftype, s64 cfs_quota_us)
8197 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8200 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8203 return tg_get_cfs_period(css_tg(css));
8206 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8207 struct cftype *cftype, u64 cfs_period_us)
8209 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8212 struct cfs_schedulable_data {
8213 struct task_group *tg;
8218 * normalize group quota/period to be quota/max_period
8219 * note: units are usecs
8221 static u64 normalize_cfs_quota(struct task_group *tg,
8222 struct cfs_schedulable_data *d)
8230 period = tg_get_cfs_period(tg);
8231 quota = tg_get_cfs_quota(tg);
8234 /* note: these should typically be equivalent */
8235 if (quota == RUNTIME_INF || quota == -1)
8238 return to_ratio(period, quota);
8241 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8243 struct cfs_schedulable_data *d = data;
8244 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8245 s64 quota = 0, parent_quota = -1;
8248 quota = RUNTIME_INF;
8250 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8252 quota = normalize_cfs_quota(tg, d);
8253 parent_quota = parent_b->hierarchical_quota;
8256 * ensure max(child_quota) <= parent_quota, inherit when no
8259 if (quota == RUNTIME_INF)
8260 quota = parent_quota;
8261 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8264 cfs_b->hierarchical_quota = quota;
8269 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8272 struct cfs_schedulable_data data = {
8278 if (quota != RUNTIME_INF) {
8279 do_div(data.period, NSEC_PER_USEC);
8280 do_div(data.quota, NSEC_PER_USEC);
8284 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8290 static int cpu_stats_show(struct seq_file *sf, void *v)
8292 struct task_group *tg = css_tg(seq_css(sf));
8293 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8295 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8296 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8297 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8301 #endif /* CONFIG_CFS_BANDWIDTH */
8302 #endif /* CONFIG_FAIR_GROUP_SCHED */
8304 #ifdef CONFIG_RT_GROUP_SCHED
8305 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8306 struct cftype *cft, s64 val)
8308 return sched_group_set_rt_runtime(css_tg(css), val);
8311 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8314 return sched_group_rt_runtime(css_tg(css));
8317 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8318 struct cftype *cftype, u64 rt_period_us)
8320 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8323 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8326 return sched_group_rt_period(css_tg(css));
8328 #endif /* CONFIG_RT_GROUP_SCHED */
8330 static struct cftype cpu_files[] = {
8331 #ifdef CONFIG_FAIR_GROUP_SCHED
8334 .read_u64 = cpu_shares_read_u64,
8335 .write_u64 = cpu_shares_write_u64,
8338 #ifdef CONFIG_CFS_BANDWIDTH
8340 .name = "cfs_quota_us",
8341 .read_s64 = cpu_cfs_quota_read_s64,
8342 .write_s64 = cpu_cfs_quota_write_s64,
8345 .name = "cfs_period_us",
8346 .read_u64 = cpu_cfs_period_read_u64,
8347 .write_u64 = cpu_cfs_period_write_u64,
8351 .seq_show = cpu_stats_show,
8354 #ifdef CONFIG_RT_GROUP_SCHED
8356 .name = "rt_runtime_us",
8357 .read_s64 = cpu_rt_runtime_read,
8358 .write_s64 = cpu_rt_runtime_write,
8361 .name = "rt_period_us",
8362 .read_u64 = cpu_rt_period_read_uint,
8363 .write_u64 = cpu_rt_period_write_uint,
8369 struct cgroup_subsys cpu_cgrp_subsys = {
8370 .css_alloc = cpu_cgroup_css_alloc,
8371 .css_free = cpu_cgroup_css_free,
8372 .css_online = cpu_cgroup_css_online,
8373 .css_offline = cpu_cgroup_css_offline,
8374 .fork = cpu_cgroup_fork,
8375 .can_attach = cpu_cgroup_can_attach,
8376 .attach = cpu_cgroup_attach,
8377 .exit = cpu_cgroup_exit,
8378 .legacy_cftypes = cpu_files,
8382 #endif /* CONFIG_CGROUP_SCHED */
8384 void dump_cpu_task(int cpu)
8386 pr_info("Task dump for CPU %d:\n", cpu);
8387 sched_show_task(cpu_curr(cpu));