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 if (p->nr_cpus_allowed > 1)
1415 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1418 * In order not to call set_task_cpu() on a blocking task we need
1419 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1422 * Since this is common to all placement strategies, this lives here.
1424 * [ this allows ->select_task() to simply return task_cpu(p) and
1425 * not worry about this generic constraint ]
1427 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1429 cpu = select_fallback_rq(task_cpu(p), p);
1434 static void update_avg(u64 *avg, u64 sample)
1436 s64 diff = sample - *avg;
1442 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1444 #ifdef CONFIG_SCHEDSTATS
1445 struct rq *rq = this_rq();
1448 int this_cpu = smp_processor_id();
1450 if (cpu == this_cpu) {
1451 schedstat_inc(rq, ttwu_local);
1452 schedstat_inc(p, se.statistics.nr_wakeups_local);
1454 struct sched_domain *sd;
1456 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1458 for_each_domain(this_cpu, sd) {
1459 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1460 schedstat_inc(sd, ttwu_wake_remote);
1467 if (wake_flags & WF_MIGRATED)
1468 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1470 #endif /* CONFIG_SMP */
1472 schedstat_inc(rq, ttwu_count);
1473 schedstat_inc(p, se.statistics.nr_wakeups);
1475 if (wake_flags & WF_SYNC)
1476 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1478 #endif /* CONFIG_SCHEDSTATS */
1481 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1483 activate_task(rq, p, en_flags);
1484 p->on_rq = TASK_ON_RQ_QUEUED;
1486 /* if a worker is waking up, notify workqueue */
1487 if (p->flags & PF_WQ_WORKER)
1488 wq_worker_waking_up(p, cpu_of(rq));
1492 * Mark the task runnable and perform wakeup-preemption.
1495 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1497 check_preempt_curr(rq, p, wake_flags);
1498 trace_sched_wakeup(p, true);
1500 p->state = TASK_RUNNING;
1502 if (p->sched_class->task_woken)
1503 p->sched_class->task_woken(rq, p);
1505 if (rq->idle_stamp) {
1506 u64 delta = rq_clock(rq) - rq->idle_stamp;
1507 u64 max = 2*rq->max_idle_balance_cost;
1509 update_avg(&rq->avg_idle, delta);
1511 if (rq->avg_idle > max)
1520 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1523 if (p->sched_contributes_to_load)
1524 rq->nr_uninterruptible--;
1527 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1528 ttwu_do_wakeup(rq, p, wake_flags);
1532 * Called in case the task @p isn't fully descheduled from its runqueue,
1533 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1534 * since all we need to do is flip p->state to TASK_RUNNING, since
1535 * the task is still ->on_rq.
1537 static int ttwu_remote(struct task_struct *p, int wake_flags)
1542 rq = __task_rq_lock(p);
1543 if (task_on_rq_queued(p)) {
1544 /* check_preempt_curr() may use rq clock */
1545 update_rq_clock(rq);
1546 ttwu_do_wakeup(rq, p, wake_flags);
1549 __task_rq_unlock(rq);
1555 void sched_ttwu_pending(void)
1557 struct rq *rq = this_rq();
1558 struct llist_node *llist = llist_del_all(&rq->wake_list);
1559 struct task_struct *p;
1560 unsigned long flags;
1565 raw_spin_lock_irqsave(&rq->lock, flags);
1568 p = llist_entry(llist, struct task_struct, wake_entry);
1569 llist = llist_next(llist);
1570 ttwu_do_activate(rq, p, 0);
1573 raw_spin_unlock_irqrestore(&rq->lock, flags);
1576 void scheduler_ipi(void)
1579 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1580 * TIF_NEED_RESCHED remotely (for the first time) will also send
1583 preempt_fold_need_resched();
1585 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1589 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1590 * traditionally all their work was done from the interrupt return
1591 * path. Now that we actually do some work, we need to make sure
1594 * Some archs already do call them, luckily irq_enter/exit nest
1597 * Arguably we should visit all archs and update all handlers,
1598 * however a fair share of IPIs are still resched only so this would
1599 * somewhat pessimize the simple resched case.
1602 sched_ttwu_pending();
1605 * Check if someone kicked us for doing the nohz idle load balance.
1607 if (unlikely(got_nohz_idle_kick())) {
1608 this_rq()->idle_balance = 1;
1609 raise_softirq_irqoff(SCHED_SOFTIRQ);
1614 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1616 struct rq *rq = cpu_rq(cpu);
1618 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1619 if (!set_nr_if_polling(rq->idle))
1620 smp_send_reschedule(cpu);
1622 trace_sched_wake_idle_without_ipi(cpu);
1626 void wake_up_if_idle(int cpu)
1628 struct rq *rq = cpu_rq(cpu);
1629 unsigned long flags;
1633 if (!is_idle_task(rcu_dereference(rq->curr)))
1636 if (set_nr_if_polling(rq->idle)) {
1637 trace_sched_wake_idle_without_ipi(cpu);
1639 raw_spin_lock_irqsave(&rq->lock, flags);
1640 if (is_idle_task(rq->curr))
1641 smp_send_reschedule(cpu);
1642 /* Else cpu is not in idle, do nothing here */
1643 raw_spin_unlock_irqrestore(&rq->lock, flags);
1650 bool cpus_share_cache(int this_cpu, int that_cpu)
1652 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1654 #endif /* CONFIG_SMP */
1656 static void ttwu_queue(struct task_struct *p, int cpu)
1658 struct rq *rq = cpu_rq(cpu);
1660 #if defined(CONFIG_SMP)
1661 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1662 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1663 ttwu_queue_remote(p, cpu);
1668 raw_spin_lock(&rq->lock);
1669 ttwu_do_activate(rq, p, 0);
1670 raw_spin_unlock(&rq->lock);
1674 * try_to_wake_up - wake up a thread
1675 * @p: the thread to be awakened
1676 * @state: the mask of task states that can be woken
1677 * @wake_flags: wake modifier flags (WF_*)
1679 * Put it on the run-queue if it's not already there. The "current"
1680 * thread is always on the run-queue (except when the actual
1681 * re-schedule is in progress), and as such you're allowed to do
1682 * the simpler "current->state = TASK_RUNNING" to mark yourself
1683 * runnable without the overhead of this.
1685 * Return: %true if @p was woken up, %false if it was already running.
1686 * or @state didn't match @p's state.
1689 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1691 unsigned long flags;
1692 int cpu, success = 0;
1695 * If we are going to wake up a thread waiting for CONDITION we
1696 * need to ensure that CONDITION=1 done by the caller can not be
1697 * reordered with p->state check below. This pairs with mb() in
1698 * set_current_state() the waiting thread does.
1700 smp_mb__before_spinlock();
1701 raw_spin_lock_irqsave(&p->pi_lock, flags);
1702 if (!(p->state & state))
1705 success = 1; /* we're going to change ->state */
1708 if (p->on_rq && ttwu_remote(p, wake_flags))
1713 * If the owning (remote) cpu is still in the middle of schedule() with
1714 * this task as prev, wait until its done referencing the task.
1719 * Pairs with the smp_wmb() in finish_lock_switch().
1723 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1724 p->state = TASK_WAKING;
1726 if (p->sched_class->task_waking)
1727 p->sched_class->task_waking(p);
1729 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1730 if (task_cpu(p) != cpu) {
1731 wake_flags |= WF_MIGRATED;
1732 set_task_cpu(p, cpu);
1734 #endif /* CONFIG_SMP */
1738 ttwu_stat(p, cpu, wake_flags);
1740 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1746 * try_to_wake_up_local - try to wake up a local task with rq lock held
1747 * @p: the thread to be awakened
1749 * Put @p on the run-queue if it's not already there. The caller must
1750 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1753 static void try_to_wake_up_local(struct task_struct *p)
1755 struct rq *rq = task_rq(p);
1757 if (WARN_ON_ONCE(rq != this_rq()) ||
1758 WARN_ON_ONCE(p == current))
1761 lockdep_assert_held(&rq->lock);
1763 if (!raw_spin_trylock(&p->pi_lock)) {
1764 raw_spin_unlock(&rq->lock);
1765 raw_spin_lock(&p->pi_lock);
1766 raw_spin_lock(&rq->lock);
1769 if (!(p->state & TASK_NORMAL))
1772 if (!task_on_rq_queued(p))
1773 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1775 ttwu_do_wakeup(rq, p, 0);
1776 ttwu_stat(p, smp_processor_id(), 0);
1778 raw_spin_unlock(&p->pi_lock);
1782 * wake_up_process - Wake up a specific process
1783 * @p: The process to be woken up.
1785 * Attempt to wake up the nominated process and move it to the set of runnable
1788 * Return: 1 if the process was woken up, 0 if it was already running.
1790 * It may be assumed that this function implies a write memory barrier before
1791 * changing the task state if and only if any tasks are woken up.
1793 int wake_up_process(struct task_struct *p)
1795 WARN_ON(task_is_stopped_or_traced(p));
1796 return try_to_wake_up(p, TASK_NORMAL, 0);
1798 EXPORT_SYMBOL(wake_up_process);
1800 int wake_up_state(struct task_struct *p, unsigned int state)
1802 return try_to_wake_up(p, state, 0);
1806 * This function clears the sched_dl_entity static params.
1808 void __dl_clear_params(struct task_struct *p)
1810 struct sched_dl_entity *dl_se = &p->dl;
1812 dl_se->dl_runtime = 0;
1813 dl_se->dl_deadline = 0;
1814 dl_se->dl_period = 0;
1820 * Perform scheduler related setup for a newly forked process p.
1821 * p is forked by current.
1823 * __sched_fork() is basic setup used by init_idle() too:
1825 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1830 p->se.exec_start = 0;
1831 p->se.sum_exec_runtime = 0;
1832 p->se.prev_sum_exec_runtime = 0;
1833 p->se.nr_migrations = 0;
1835 INIT_LIST_HEAD(&p->se.group_node);
1837 #ifdef CONFIG_SCHEDSTATS
1838 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1841 RB_CLEAR_NODE(&p->dl.rb_node);
1842 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1843 __dl_clear_params(p);
1845 INIT_LIST_HEAD(&p->rt.run_list);
1847 #ifdef CONFIG_PREEMPT_NOTIFIERS
1848 INIT_HLIST_HEAD(&p->preempt_notifiers);
1851 #ifdef CONFIG_NUMA_BALANCING
1852 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1853 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1854 p->mm->numa_scan_seq = 0;
1857 if (clone_flags & CLONE_VM)
1858 p->numa_preferred_nid = current->numa_preferred_nid;
1860 p->numa_preferred_nid = -1;
1862 p->node_stamp = 0ULL;
1863 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1864 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1865 p->numa_work.next = &p->numa_work;
1866 p->numa_faults = NULL;
1867 p->last_task_numa_placement = 0;
1868 p->last_sum_exec_runtime = 0;
1870 p->numa_group = NULL;
1871 #endif /* CONFIG_NUMA_BALANCING */
1874 #ifdef CONFIG_NUMA_BALANCING
1875 #ifdef CONFIG_SCHED_DEBUG
1876 void set_numabalancing_state(bool enabled)
1879 sched_feat_set("NUMA");
1881 sched_feat_set("NO_NUMA");
1884 __read_mostly bool numabalancing_enabled;
1886 void set_numabalancing_state(bool enabled)
1888 numabalancing_enabled = enabled;
1890 #endif /* CONFIG_SCHED_DEBUG */
1892 #ifdef CONFIG_PROC_SYSCTL
1893 int sysctl_numa_balancing(struct ctl_table *table, int write,
1894 void __user *buffer, size_t *lenp, loff_t *ppos)
1898 int state = numabalancing_enabled;
1900 if (write && !capable(CAP_SYS_ADMIN))
1905 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1909 set_numabalancing_state(state);
1916 * fork()/clone()-time setup:
1918 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1920 unsigned long flags;
1921 int cpu = get_cpu();
1923 __sched_fork(clone_flags, p);
1925 * We mark the process as running here. This guarantees that
1926 * nobody will actually run it, and a signal or other external
1927 * event cannot wake it up and insert it on the runqueue either.
1929 p->state = TASK_RUNNING;
1932 * Make sure we do not leak PI boosting priority to the child.
1934 p->prio = current->normal_prio;
1937 * Revert to default priority/policy on fork if requested.
1939 if (unlikely(p->sched_reset_on_fork)) {
1940 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1941 p->policy = SCHED_NORMAL;
1942 p->static_prio = NICE_TO_PRIO(0);
1944 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1945 p->static_prio = NICE_TO_PRIO(0);
1947 p->prio = p->normal_prio = __normal_prio(p);
1951 * We don't need the reset flag anymore after the fork. It has
1952 * fulfilled its duty:
1954 p->sched_reset_on_fork = 0;
1957 if (dl_prio(p->prio)) {
1960 } else if (rt_prio(p->prio)) {
1961 p->sched_class = &rt_sched_class;
1963 p->sched_class = &fair_sched_class;
1966 if (p->sched_class->task_fork)
1967 p->sched_class->task_fork(p);
1970 * The child is not yet in the pid-hash so no cgroup attach races,
1971 * and the cgroup is pinned to this child due to cgroup_fork()
1972 * is ran before sched_fork().
1974 * Silence PROVE_RCU.
1976 raw_spin_lock_irqsave(&p->pi_lock, flags);
1977 set_task_cpu(p, cpu);
1978 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1980 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1981 if (likely(sched_info_on()))
1982 memset(&p->sched_info, 0, sizeof(p->sched_info));
1984 #if defined(CONFIG_SMP)
1987 init_task_preempt_count(p);
1989 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1990 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1997 unsigned long to_ratio(u64 period, u64 runtime)
1999 if (runtime == RUNTIME_INF)
2003 * Doing this here saves a lot of checks in all
2004 * the calling paths, and returning zero seems
2005 * safe for them anyway.
2010 return div64_u64(runtime << 20, period);
2014 inline struct dl_bw *dl_bw_of(int i)
2016 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2017 "sched RCU must be held");
2018 return &cpu_rq(i)->rd->dl_bw;
2021 static inline int dl_bw_cpus(int i)
2023 struct root_domain *rd = cpu_rq(i)->rd;
2026 rcu_lockdep_assert(rcu_read_lock_sched_held(),
2027 "sched RCU must be held");
2028 for_each_cpu_and(i, rd->span, cpu_active_mask)
2034 inline struct dl_bw *dl_bw_of(int i)
2036 return &cpu_rq(i)->dl.dl_bw;
2039 static inline int dl_bw_cpus(int i)
2046 * We must be sure that accepting a new task (or allowing changing the
2047 * parameters of an existing one) is consistent with the bandwidth
2048 * constraints. If yes, this function also accordingly updates the currently
2049 * allocated bandwidth to reflect the new situation.
2051 * This function is called while holding p's rq->lock.
2053 static int dl_overflow(struct task_struct *p, int policy,
2054 const struct sched_attr *attr)
2057 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2058 u64 period = attr->sched_period ?: attr->sched_deadline;
2059 u64 runtime = attr->sched_runtime;
2060 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2063 if (new_bw == p->dl.dl_bw)
2067 * Either if a task, enters, leave, or stays -deadline but changes
2068 * its parameters, we may need to update accordingly the total
2069 * allocated bandwidth of the container.
2071 raw_spin_lock(&dl_b->lock);
2072 cpus = dl_bw_cpus(task_cpu(p));
2073 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2074 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2075 __dl_add(dl_b, new_bw);
2077 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2078 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2079 __dl_clear(dl_b, p->dl.dl_bw);
2080 __dl_add(dl_b, new_bw);
2082 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2083 __dl_clear(dl_b, p->dl.dl_bw);
2086 raw_spin_unlock(&dl_b->lock);
2091 extern void init_dl_bw(struct dl_bw *dl_b);
2094 * wake_up_new_task - wake up a newly created task for the first time.
2096 * This function will do some initial scheduler statistics housekeeping
2097 * that must be done for every newly created context, then puts the task
2098 * on the runqueue and wakes it.
2100 void wake_up_new_task(struct task_struct *p)
2102 unsigned long flags;
2105 raw_spin_lock_irqsave(&p->pi_lock, flags);
2108 * Fork balancing, do it here and not earlier because:
2109 * - cpus_allowed can change in the fork path
2110 * - any previously selected cpu might disappear through hotplug
2112 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2115 /* Initialize new task's runnable average */
2116 init_task_runnable_average(p);
2117 rq = __task_rq_lock(p);
2118 activate_task(rq, p, 0);
2119 p->on_rq = TASK_ON_RQ_QUEUED;
2120 trace_sched_wakeup_new(p, true);
2121 check_preempt_curr(rq, p, WF_FORK);
2123 if (p->sched_class->task_woken)
2124 p->sched_class->task_woken(rq, p);
2126 task_rq_unlock(rq, p, &flags);
2129 #ifdef CONFIG_PREEMPT_NOTIFIERS
2132 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2133 * @notifier: notifier struct to register
2135 void preempt_notifier_register(struct preempt_notifier *notifier)
2137 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2139 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2142 * preempt_notifier_unregister - no longer interested in preemption notifications
2143 * @notifier: notifier struct to unregister
2145 * This is safe to call from within a preemption notifier.
2147 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2149 hlist_del(¬ifier->link);
2151 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2153 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2155 struct preempt_notifier *notifier;
2157 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2158 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2162 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2163 struct task_struct *next)
2165 struct preempt_notifier *notifier;
2167 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2168 notifier->ops->sched_out(notifier, next);
2171 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2173 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2178 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2179 struct task_struct *next)
2183 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2186 * prepare_task_switch - prepare to switch tasks
2187 * @rq: the runqueue preparing to switch
2188 * @prev: the current task that is being switched out
2189 * @next: the task we are going to switch to.
2191 * This is called with the rq lock held and interrupts off. It must
2192 * be paired with a subsequent finish_task_switch after the context
2195 * prepare_task_switch sets up locking and calls architecture specific
2199 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2200 struct task_struct *next)
2202 trace_sched_switch(prev, next);
2203 sched_info_switch(rq, prev, next);
2204 perf_event_task_sched_out(prev, next);
2205 fire_sched_out_preempt_notifiers(prev, next);
2206 prepare_lock_switch(rq, next);
2207 prepare_arch_switch(next);
2211 * finish_task_switch - clean up after a task-switch
2212 * @prev: the thread we just switched away from.
2214 * finish_task_switch must be called after the context switch, paired
2215 * with a prepare_task_switch call before the context switch.
2216 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2217 * and do any other architecture-specific cleanup actions.
2219 * Note that we may have delayed dropping an mm in context_switch(). If
2220 * so, we finish that here outside of the runqueue lock. (Doing it
2221 * with the lock held can cause deadlocks; see schedule() for
2224 * The context switch have flipped the stack from under us and restored the
2225 * local variables which were saved when this task called schedule() in the
2226 * past. prev == current is still correct but we need to recalculate this_rq
2227 * because prev may have moved to another CPU.
2229 static struct rq *finish_task_switch(struct task_struct *prev)
2230 __releases(rq->lock)
2232 struct rq *rq = this_rq();
2233 struct mm_struct *mm = rq->prev_mm;
2239 * A task struct has one reference for the use as "current".
2240 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2241 * schedule one last time. The schedule call will never return, and
2242 * the scheduled task must drop that reference.
2243 * The test for TASK_DEAD must occur while the runqueue locks are
2244 * still held, otherwise prev could be scheduled on another cpu, die
2245 * there before we look at prev->state, and then the reference would
2247 * Manfred Spraul <manfred@colorfullife.com>
2249 prev_state = prev->state;
2250 vtime_task_switch(prev);
2251 finish_arch_switch(prev);
2252 perf_event_task_sched_in(prev, current);
2253 finish_lock_switch(rq, prev);
2254 finish_arch_post_lock_switch();
2256 fire_sched_in_preempt_notifiers(current);
2259 if (unlikely(prev_state == TASK_DEAD)) {
2260 if (prev->sched_class->task_dead)
2261 prev->sched_class->task_dead(prev);
2264 * Remove function-return probe instances associated with this
2265 * task and put them back on the free list.
2267 kprobe_flush_task(prev);
2268 put_task_struct(prev);
2271 tick_nohz_task_switch(current);
2277 /* rq->lock is NOT held, but preemption is disabled */
2278 static inline void post_schedule(struct rq *rq)
2280 if (rq->post_schedule) {
2281 unsigned long flags;
2283 raw_spin_lock_irqsave(&rq->lock, flags);
2284 if (rq->curr->sched_class->post_schedule)
2285 rq->curr->sched_class->post_schedule(rq);
2286 raw_spin_unlock_irqrestore(&rq->lock, flags);
2288 rq->post_schedule = 0;
2294 static inline void post_schedule(struct rq *rq)
2301 * schedule_tail - first thing a freshly forked thread must call.
2302 * @prev: the thread we just switched away from.
2304 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2305 __releases(rq->lock)
2309 /* finish_task_switch() drops rq->lock and enables preemtion */
2311 rq = finish_task_switch(prev);
2315 if (current->set_child_tid)
2316 put_user(task_pid_vnr(current), current->set_child_tid);
2320 * context_switch - switch to the new MM and the new thread's register state.
2322 static inline struct rq *
2323 context_switch(struct rq *rq, struct task_struct *prev,
2324 struct task_struct *next)
2326 struct mm_struct *mm, *oldmm;
2328 prepare_task_switch(rq, prev, next);
2331 oldmm = prev->active_mm;
2333 * For paravirt, this is coupled with an exit in switch_to to
2334 * combine the page table reload and the switch backend into
2337 arch_start_context_switch(prev);
2340 next->active_mm = oldmm;
2341 atomic_inc(&oldmm->mm_count);
2342 enter_lazy_tlb(oldmm, next);
2344 switch_mm(oldmm, mm, next);
2347 prev->active_mm = NULL;
2348 rq->prev_mm = oldmm;
2351 * Since the runqueue lock will be released by the next
2352 * task (which is an invalid locking op but in the case
2353 * of the scheduler it's an obvious special-case), so we
2354 * do an early lockdep release here:
2356 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2358 context_tracking_task_switch(prev, next);
2359 /* Here we just switch the register state and the stack. */
2360 switch_to(prev, next, prev);
2363 return finish_task_switch(prev);
2367 * nr_running and nr_context_switches:
2369 * externally visible scheduler statistics: current number of runnable
2370 * threads, total number of context switches performed since bootup.
2372 unsigned long nr_running(void)
2374 unsigned long i, sum = 0;
2376 for_each_online_cpu(i)
2377 sum += cpu_rq(i)->nr_running;
2383 * Check if only the current task is running on the cpu.
2385 bool single_task_running(void)
2387 if (cpu_rq(smp_processor_id())->nr_running == 1)
2392 EXPORT_SYMBOL(single_task_running);
2394 unsigned long long nr_context_switches(void)
2397 unsigned long long sum = 0;
2399 for_each_possible_cpu(i)
2400 sum += cpu_rq(i)->nr_switches;
2405 unsigned long nr_iowait(void)
2407 unsigned long i, sum = 0;
2409 for_each_possible_cpu(i)
2410 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2415 unsigned long nr_iowait_cpu(int cpu)
2417 struct rq *this = cpu_rq(cpu);
2418 return atomic_read(&this->nr_iowait);
2421 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2423 struct rq *this = this_rq();
2424 *nr_waiters = atomic_read(&this->nr_iowait);
2425 *load = this->cpu_load[0];
2431 * sched_exec - execve() is a valuable balancing opportunity, because at
2432 * this point the task has the smallest effective memory and cache footprint.
2434 void sched_exec(void)
2436 struct task_struct *p = current;
2437 unsigned long flags;
2440 raw_spin_lock_irqsave(&p->pi_lock, flags);
2441 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2442 if (dest_cpu == smp_processor_id())
2445 if (likely(cpu_active(dest_cpu))) {
2446 struct migration_arg arg = { p, dest_cpu };
2448 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2449 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2453 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2458 DEFINE_PER_CPU(struct kernel_stat, kstat);
2459 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2461 EXPORT_PER_CPU_SYMBOL(kstat);
2462 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2465 * Return accounted runtime for the task.
2466 * In case the task is currently running, return the runtime plus current's
2467 * pending runtime that have not been accounted yet.
2469 unsigned long long task_sched_runtime(struct task_struct *p)
2471 unsigned long flags;
2475 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2477 * 64-bit doesn't need locks to atomically read a 64bit value.
2478 * So we have a optimization chance when the task's delta_exec is 0.
2479 * Reading ->on_cpu is racy, but this is ok.
2481 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2482 * If we race with it entering cpu, unaccounted time is 0. This is
2483 * indistinguishable from the read occurring a few cycles earlier.
2484 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2485 * been accounted, so we're correct here as well.
2487 if (!p->on_cpu || !task_on_rq_queued(p))
2488 return p->se.sum_exec_runtime;
2491 rq = task_rq_lock(p, &flags);
2493 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2494 * project cycles that may never be accounted to this
2495 * thread, breaking clock_gettime().
2497 if (task_current(rq, p) && task_on_rq_queued(p)) {
2498 update_rq_clock(rq);
2499 p->sched_class->update_curr(rq);
2501 ns = p->se.sum_exec_runtime;
2502 task_rq_unlock(rq, p, &flags);
2508 * This function gets called by the timer code, with HZ frequency.
2509 * We call it with interrupts disabled.
2511 void scheduler_tick(void)
2513 int cpu = smp_processor_id();
2514 struct rq *rq = cpu_rq(cpu);
2515 struct task_struct *curr = rq->curr;
2519 raw_spin_lock(&rq->lock);
2520 update_rq_clock(rq);
2521 curr->sched_class->task_tick(rq, curr, 0);
2522 update_cpu_load_active(rq);
2523 raw_spin_unlock(&rq->lock);
2525 perf_event_task_tick();
2528 rq->idle_balance = idle_cpu(cpu);
2529 trigger_load_balance(rq);
2531 rq_last_tick_reset(rq);
2534 #ifdef CONFIG_NO_HZ_FULL
2536 * scheduler_tick_max_deferment
2538 * Keep at least one tick per second when a single
2539 * active task is running because the scheduler doesn't
2540 * yet completely support full dynticks environment.
2542 * This makes sure that uptime, CFS vruntime, load
2543 * balancing, etc... continue to move forward, even
2544 * with a very low granularity.
2546 * Return: Maximum deferment in nanoseconds.
2548 u64 scheduler_tick_max_deferment(void)
2550 struct rq *rq = this_rq();
2551 unsigned long next, now = ACCESS_ONCE(jiffies);
2553 next = rq->last_sched_tick + HZ;
2555 if (time_before_eq(next, now))
2558 return jiffies_to_nsecs(next - now);
2562 notrace unsigned long get_parent_ip(unsigned long addr)
2564 if (in_lock_functions(addr)) {
2565 addr = CALLER_ADDR2;
2566 if (in_lock_functions(addr))
2567 addr = CALLER_ADDR3;
2572 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2573 defined(CONFIG_PREEMPT_TRACER))
2575 void preempt_count_add(int val)
2577 #ifdef CONFIG_DEBUG_PREEMPT
2581 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2584 __preempt_count_add(val);
2585 #ifdef CONFIG_DEBUG_PREEMPT
2587 * Spinlock count overflowing soon?
2589 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2592 if (preempt_count() == val) {
2593 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2594 #ifdef CONFIG_DEBUG_PREEMPT
2595 current->preempt_disable_ip = ip;
2597 trace_preempt_off(CALLER_ADDR0, ip);
2600 EXPORT_SYMBOL(preempt_count_add);
2601 NOKPROBE_SYMBOL(preempt_count_add);
2603 void preempt_count_sub(int val)
2605 #ifdef CONFIG_DEBUG_PREEMPT
2609 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2612 * Is the spinlock portion underflowing?
2614 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2615 !(preempt_count() & PREEMPT_MASK)))
2619 if (preempt_count() == val)
2620 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2621 __preempt_count_sub(val);
2623 EXPORT_SYMBOL(preempt_count_sub);
2624 NOKPROBE_SYMBOL(preempt_count_sub);
2629 * Print scheduling while atomic bug:
2631 static noinline void __schedule_bug(struct task_struct *prev)
2633 if (oops_in_progress)
2636 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2637 prev->comm, prev->pid, preempt_count());
2639 debug_show_held_locks(prev);
2641 if (irqs_disabled())
2642 print_irqtrace_events(prev);
2643 #ifdef CONFIG_DEBUG_PREEMPT
2644 if (in_atomic_preempt_off()) {
2645 pr_err("Preemption disabled at:");
2646 print_ip_sym(current->preempt_disable_ip);
2651 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2655 * Various schedule()-time debugging checks and statistics:
2657 static inline void schedule_debug(struct task_struct *prev)
2659 #ifdef CONFIG_SCHED_STACK_END_CHECK
2660 BUG_ON(unlikely(task_stack_end_corrupted(prev)));
2663 * Test if we are atomic. Since do_exit() needs to call into
2664 * schedule() atomically, we ignore that path. Otherwise whine
2665 * if we are scheduling when we should not.
2667 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2668 __schedule_bug(prev);
2671 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2673 schedstat_inc(this_rq(), sched_count);
2677 * Pick up the highest-prio task:
2679 static inline struct task_struct *
2680 pick_next_task(struct rq *rq, struct task_struct *prev)
2682 const struct sched_class *class = &fair_sched_class;
2683 struct task_struct *p;
2686 * Optimization: we know that if all tasks are in
2687 * the fair class we can call that function directly:
2689 if (likely(prev->sched_class == class &&
2690 rq->nr_running == rq->cfs.h_nr_running)) {
2691 p = fair_sched_class.pick_next_task(rq, prev);
2692 if (unlikely(p == RETRY_TASK))
2695 /* assumes fair_sched_class->next == idle_sched_class */
2697 p = idle_sched_class.pick_next_task(rq, prev);
2703 for_each_class(class) {
2704 p = class->pick_next_task(rq, prev);
2706 if (unlikely(p == RETRY_TASK))
2712 BUG(); /* the idle class will always have a runnable task */
2716 * __schedule() is the main scheduler function.
2718 * The main means of driving the scheduler and thus entering this function are:
2720 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2722 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2723 * paths. For example, see arch/x86/entry_64.S.
2725 * To drive preemption between tasks, the scheduler sets the flag in timer
2726 * interrupt handler scheduler_tick().
2728 * 3. Wakeups don't really cause entry into schedule(). They add a
2729 * task to the run-queue and that's it.
2731 * Now, if the new task added to the run-queue preempts the current
2732 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2733 * called on the nearest possible occasion:
2735 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2737 * - in syscall or exception context, at the next outmost
2738 * preempt_enable(). (this might be as soon as the wake_up()'s
2741 * - in IRQ context, return from interrupt-handler to
2742 * preemptible context
2744 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2747 * - cond_resched() call
2748 * - explicit schedule() call
2749 * - return from syscall or exception to user-space
2750 * - return from interrupt-handler to user-space
2752 static void __sched __schedule(void)
2754 struct task_struct *prev, *next;
2755 unsigned long *switch_count;
2761 cpu = smp_processor_id();
2763 rcu_note_context_switch();
2766 schedule_debug(prev);
2768 if (sched_feat(HRTICK))
2772 * Make sure that signal_pending_state()->signal_pending() below
2773 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2774 * done by the caller to avoid the race with signal_wake_up().
2776 smp_mb__before_spinlock();
2777 raw_spin_lock_irq(&rq->lock);
2779 switch_count = &prev->nivcsw;
2780 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2781 if (unlikely(signal_pending_state(prev->state, prev))) {
2782 prev->state = TASK_RUNNING;
2784 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2788 * If a worker went to sleep, notify and ask workqueue
2789 * whether it wants to wake up a task to maintain
2792 if (prev->flags & PF_WQ_WORKER) {
2793 struct task_struct *to_wakeup;
2795 to_wakeup = wq_worker_sleeping(prev, cpu);
2797 try_to_wake_up_local(to_wakeup);
2800 switch_count = &prev->nvcsw;
2803 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0)
2804 update_rq_clock(rq);
2806 next = pick_next_task(rq, prev);
2807 clear_tsk_need_resched(prev);
2808 clear_preempt_need_resched();
2809 rq->skip_clock_update = 0;
2811 if (likely(prev != next)) {
2816 rq = context_switch(rq, prev, next); /* unlocks the rq */
2819 raw_spin_unlock_irq(&rq->lock);
2823 sched_preempt_enable_no_resched();
2828 static inline void sched_submit_work(struct task_struct *tsk)
2830 if (!tsk->state || tsk_is_pi_blocked(tsk))
2833 * If we are going to sleep and we have plugged IO queued,
2834 * make sure to submit it to avoid deadlocks.
2836 if (blk_needs_flush_plug(tsk))
2837 blk_schedule_flush_plug(tsk);
2840 asmlinkage __visible void __sched schedule(void)
2842 struct task_struct *tsk = current;
2844 sched_submit_work(tsk);
2847 EXPORT_SYMBOL(schedule);
2849 #ifdef CONFIG_CONTEXT_TRACKING
2850 asmlinkage __visible void __sched schedule_user(void)
2853 * If we come here after a random call to set_need_resched(),
2854 * or we have been woken up remotely but the IPI has not yet arrived,
2855 * we haven't yet exited the RCU idle mode. Do it here manually until
2856 * we find a better solution.
2858 * NB: There are buggy callers of this function. Ideally we
2859 * should warn if prev_state != IN_USER, but that will trigger
2860 * too frequently to make sense yet.
2862 enum ctx_state prev_state = exception_enter();
2864 exception_exit(prev_state);
2869 * schedule_preempt_disabled - called with preemption disabled
2871 * Returns with preemption disabled. Note: preempt_count must be 1
2873 void __sched schedule_preempt_disabled(void)
2875 sched_preempt_enable_no_resched();
2880 #ifdef CONFIG_PREEMPT
2882 * this is the entry point to schedule() from in-kernel preemption
2883 * off of preempt_enable. Kernel preemptions off return from interrupt
2884 * occur there and call schedule directly.
2886 asmlinkage __visible void __sched notrace preempt_schedule(void)
2889 * If there is a non-zero preempt_count or interrupts are disabled,
2890 * we do not want to preempt the current task. Just return..
2892 if (likely(!preemptible()))
2896 __preempt_count_add(PREEMPT_ACTIVE);
2898 __preempt_count_sub(PREEMPT_ACTIVE);
2901 * Check again in case we missed a preemption opportunity
2902 * between schedule and now.
2905 } while (need_resched());
2907 NOKPROBE_SYMBOL(preempt_schedule);
2908 EXPORT_SYMBOL(preempt_schedule);
2910 #ifdef CONFIG_CONTEXT_TRACKING
2912 * preempt_schedule_context - preempt_schedule called by tracing
2914 * The tracing infrastructure uses preempt_enable_notrace to prevent
2915 * recursion and tracing preempt enabling caused by the tracing
2916 * infrastructure itself. But as tracing can happen in areas coming
2917 * from userspace or just about to enter userspace, a preempt enable
2918 * can occur before user_exit() is called. This will cause the scheduler
2919 * to be called when the system is still in usermode.
2921 * To prevent this, the preempt_enable_notrace will use this function
2922 * instead of preempt_schedule() to exit user context if needed before
2923 * calling the scheduler.
2925 asmlinkage __visible void __sched notrace preempt_schedule_context(void)
2927 enum ctx_state prev_ctx;
2929 if (likely(!preemptible()))
2933 __preempt_count_add(PREEMPT_ACTIVE);
2935 * Needs preempt disabled in case user_exit() is traced
2936 * and the tracer calls preempt_enable_notrace() causing
2937 * an infinite recursion.
2939 prev_ctx = exception_enter();
2941 exception_exit(prev_ctx);
2943 __preempt_count_sub(PREEMPT_ACTIVE);
2945 } while (need_resched());
2947 EXPORT_SYMBOL_GPL(preempt_schedule_context);
2948 #endif /* CONFIG_CONTEXT_TRACKING */
2950 #endif /* CONFIG_PREEMPT */
2953 * this is the entry point to schedule() from kernel preemption
2954 * off of irq context.
2955 * Note, that this is called and return with irqs disabled. This will
2956 * protect us against recursive calling from irq.
2958 asmlinkage __visible void __sched preempt_schedule_irq(void)
2960 enum ctx_state prev_state;
2962 /* Catch callers which need to be fixed */
2963 BUG_ON(preempt_count() || !irqs_disabled());
2965 prev_state = exception_enter();
2968 __preempt_count_add(PREEMPT_ACTIVE);
2971 local_irq_disable();
2972 __preempt_count_sub(PREEMPT_ACTIVE);
2975 * Check again in case we missed a preemption opportunity
2976 * between schedule and now.
2979 } while (need_resched());
2981 exception_exit(prev_state);
2984 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2987 return try_to_wake_up(curr->private, mode, wake_flags);
2989 EXPORT_SYMBOL(default_wake_function);
2991 #ifdef CONFIG_RT_MUTEXES
2994 * rt_mutex_setprio - set the current priority of a task
2996 * @prio: prio value (kernel-internal form)
2998 * This function changes the 'effective' priority of a task. It does
2999 * not touch ->normal_prio like __setscheduler().
3001 * Used by the rt_mutex code to implement priority inheritance
3002 * logic. Call site only calls if the priority of the task changed.
3004 void rt_mutex_setprio(struct task_struct *p, int prio)
3006 int oldprio, queued, running, enqueue_flag = 0;
3008 const struct sched_class *prev_class;
3010 BUG_ON(prio > MAX_PRIO);
3012 rq = __task_rq_lock(p);
3015 * Idle task boosting is a nono in general. There is one
3016 * exception, when PREEMPT_RT and NOHZ is active:
3018 * The idle task calls get_next_timer_interrupt() and holds
3019 * the timer wheel base->lock on the CPU and another CPU wants
3020 * to access the timer (probably to cancel it). We can safely
3021 * ignore the boosting request, as the idle CPU runs this code
3022 * with interrupts disabled and will complete the lock
3023 * protected section without being interrupted. So there is no
3024 * real need to boost.
3026 if (unlikely(p == rq->idle)) {
3027 WARN_ON(p != rq->curr);
3028 WARN_ON(p->pi_blocked_on);
3032 trace_sched_pi_setprio(p, prio);
3034 prev_class = p->sched_class;
3035 queued = task_on_rq_queued(p);
3036 running = task_current(rq, p);
3038 dequeue_task(rq, p, 0);
3040 put_prev_task(rq, p);
3043 * Boosting condition are:
3044 * 1. -rt task is running and holds mutex A
3045 * --> -dl task blocks on mutex A
3047 * 2. -dl task is running and holds mutex A
3048 * --> -dl task blocks on mutex A and could preempt the
3051 if (dl_prio(prio)) {
3052 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3053 if (!dl_prio(p->normal_prio) ||
3054 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3055 p->dl.dl_boosted = 1;
3056 p->dl.dl_throttled = 0;
3057 enqueue_flag = ENQUEUE_REPLENISH;
3059 p->dl.dl_boosted = 0;
3060 p->sched_class = &dl_sched_class;
3061 } else if (rt_prio(prio)) {
3062 if (dl_prio(oldprio))
3063 p->dl.dl_boosted = 0;
3065 enqueue_flag = ENQUEUE_HEAD;
3066 p->sched_class = &rt_sched_class;
3068 if (dl_prio(oldprio))
3069 p->dl.dl_boosted = 0;
3070 p->sched_class = &fair_sched_class;
3076 p->sched_class->set_curr_task(rq);
3078 enqueue_task(rq, p, enqueue_flag);
3080 check_class_changed(rq, p, prev_class, oldprio);
3082 __task_rq_unlock(rq);
3086 void set_user_nice(struct task_struct *p, long nice)
3088 int old_prio, delta, queued;
3089 unsigned long flags;
3092 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3095 * We have to be careful, if called from sys_setpriority(),
3096 * the task might be in the middle of scheduling on another CPU.
3098 rq = task_rq_lock(p, &flags);
3100 * The RT priorities are set via sched_setscheduler(), but we still
3101 * allow the 'normal' nice value to be set - but as expected
3102 * it wont have any effect on scheduling until the task is
3103 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3105 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3106 p->static_prio = NICE_TO_PRIO(nice);
3109 queued = task_on_rq_queued(p);
3111 dequeue_task(rq, p, 0);
3113 p->static_prio = NICE_TO_PRIO(nice);
3116 p->prio = effective_prio(p);
3117 delta = p->prio - old_prio;
3120 enqueue_task(rq, p, 0);
3122 * If the task increased its priority or is running and
3123 * lowered its priority, then reschedule its CPU:
3125 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3129 task_rq_unlock(rq, p, &flags);
3131 EXPORT_SYMBOL(set_user_nice);
3134 * can_nice - check if a task can reduce its nice value
3138 int can_nice(const struct task_struct *p, const int nice)
3140 /* convert nice value [19,-20] to rlimit style value [1,40] */
3141 int nice_rlim = nice_to_rlimit(nice);
3143 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3144 capable(CAP_SYS_NICE));
3147 #ifdef __ARCH_WANT_SYS_NICE
3150 * sys_nice - change the priority of the current process.
3151 * @increment: priority increment
3153 * sys_setpriority is a more generic, but much slower function that
3154 * does similar things.
3156 SYSCALL_DEFINE1(nice, int, increment)
3161 * Setpriority might change our priority at the same moment.
3162 * We don't have to worry. Conceptually one call occurs first
3163 * and we have a single winner.
3165 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3166 nice = task_nice(current) + increment;
3168 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3169 if (increment < 0 && !can_nice(current, nice))
3172 retval = security_task_setnice(current, nice);
3176 set_user_nice(current, nice);
3183 * task_prio - return the priority value of a given task.
3184 * @p: the task in question.
3186 * Return: The priority value as seen by users in /proc.
3187 * RT tasks are offset by -200. Normal tasks are centered
3188 * around 0, value goes from -16 to +15.
3190 int task_prio(const struct task_struct *p)
3192 return p->prio - MAX_RT_PRIO;
3196 * idle_cpu - is a given cpu idle currently?
3197 * @cpu: the processor in question.
3199 * Return: 1 if the CPU is currently idle. 0 otherwise.
3201 int idle_cpu(int cpu)
3203 struct rq *rq = cpu_rq(cpu);
3205 if (rq->curr != rq->idle)
3212 if (!llist_empty(&rq->wake_list))
3220 * idle_task - return the idle task for a given cpu.
3221 * @cpu: the processor in question.
3223 * Return: The idle task for the cpu @cpu.
3225 struct task_struct *idle_task(int cpu)
3227 return cpu_rq(cpu)->idle;
3231 * find_process_by_pid - find a process with a matching PID value.
3232 * @pid: the pid in question.
3234 * The task of @pid, if found. %NULL otherwise.
3236 static struct task_struct *find_process_by_pid(pid_t pid)
3238 return pid ? find_task_by_vpid(pid) : current;
3242 * This function initializes the sched_dl_entity of a newly becoming
3243 * SCHED_DEADLINE task.
3245 * Only the static values are considered here, the actual runtime and the
3246 * absolute deadline will be properly calculated when the task is enqueued
3247 * for the first time with its new policy.
3250 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3252 struct sched_dl_entity *dl_se = &p->dl;
3254 init_dl_task_timer(dl_se);
3255 dl_se->dl_runtime = attr->sched_runtime;
3256 dl_se->dl_deadline = attr->sched_deadline;
3257 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3258 dl_se->flags = attr->sched_flags;
3259 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3260 dl_se->dl_throttled = 0;
3262 dl_se->dl_yielded = 0;
3266 * sched_setparam() passes in -1 for its policy, to let the functions
3267 * it calls know not to change it.
3269 #define SETPARAM_POLICY -1
3271 static void __setscheduler_params(struct task_struct *p,
3272 const struct sched_attr *attr)
3274 int policy = attr->sched_policy;
3276 if (policy == SETPARAM_POLICY)
3281 if (dl_policy(policy))
3282 __setparam_dl(p, attr);
3283 else if (fair_policy(policy))
3284 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3287 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3288 * !rt_policy. Always setting this ensures that things like
3289 * getparam()/getattr() don't report silly values for !rt tasks.
3291 p->rt_priority = attr->sched_priority;
3292 p->normal_prio = normal_prio(p);
3296 /* Actually do priority change: must hold pi & rq lock. */
3297 static void __setscheduler(struct rq *rq, struct task_struct *p,
3298 const struct sched_attr *attr)
3300 __setscheduler_params(p, attr);
3303 * If we get here, there was no pi waiters boosting the
3304 * task. It is safe to use the normal prio.
3306 p->prio = normal_prio(p);
3308 if (dl_prio(p->prio))
3309 p->sched_class = &dl_sched_class;
3310 else if (rt_prio(p->prio))
3311 p->sched_class = &rt_sched_class;
3313 p->sched_class = &fair_sched_class;
3317 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3319 struct sched_dl_entity *dl_se = &p->dl;
3321 attr->sched_priority = p->rt_priority;
3322 attr->sched_runtime = dl_se->dl_runtime;
3323 attr->sched_deadline = dl_se->dl_deadline;
3324 attr->sched_period = dl_se->dl_period;
3325 attr->sched_flags = dl_se->flags;
3329 * This function validates the new parameters of a -deadline task.
3330 * We ask for the deadline not being zero, and greater or equal
3331 * than the runtime, as well as the period of being zero or
3332 * greater than deadline. Furthermore, we have to be sure that
3333 * user parameters are above the internal resolution of 1us (we
3334 * check sched_runtime only since it is always the smaller one) and
3335 * below 2^63 ns (we have to check both sched_deadline and
3336 * sched_period, as the latter can be zero).
3339 __checkparam_dl(const struct sched_attr *attr)
3342 if (attr->sched_deadline == 0)
3346 * Since we truncate DL_SCALE bits, make sure we're at least
3349 if (attr->sched_runtime < (1ULL << DL_SCALE))
3353 * Since we use the MSB for wrap-around and sign issues, make
3354 * sure it's not set (mind that period can be equal to zero).
3356 if (attr->sched_deadline & (1ULL << 63) ||
3357 attr->sched_period & (1ULL << 63))
3360 /* runtime <= deadline <= period (if period != 0) */
3361 if ((attr->sched_period != 0 &&
3362 attr->sched_period < attr->sched_deadline) ||
3363 attr->sched_deadline < attr->sched_runtime)
3370 * check the target process has a UID that matches the current process's
3372 static bool check_same_owner(struct task_struct *p)
3374 const struct cred *cred = current_cred(), *pcred;
3378 pcred = __task_cred(p);
3379 match = (uid_eq(cred->euid, pcred->euid) ||
3380 uid_eq(cred->euid, pcred->uid));
3385 static int __sched_setscheduler(struct task_struct *p,
3386 const struct sched_attr *attr,
3389 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3390 MAX_RT_PRIO - 1 - attr->sched_priority;
3391 int retval, oldprio, oldpolicy = -1, queued, running;
3392 int policy = attr->sched_policy;
3393 unsigned long flags;
3394 const struct sched_class *prev_class;
3398 /* may grab non-irq protected spin_locks */
3399 BUG_ON(in_interrupt());
3401 /* double check policy once rq lock held */
3403 reset_on_fork = p->sched_reset_on_fork;
3404 policy = oldpolicy = p->policy;
3406 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3408 if (policy != SCHED_DEADLINE &&
3409 policy != SCHED_FIFO && policy != SCHED_RR &&
3410 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3411 policy != SCHED_IDLE)
3415 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3419 * Valid priorities for SCHED_FIFO and SCHED_RR are
3420 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3421 * SCHED_BATCH and SCHED_IDLE is 0.
3423 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3424 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3426 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3427 (rt_policy(policy) != (attr->sched_priority != 0)))
3431 * Allow unprivileged RT tasks to decrease priority:
3433 if (user && !capable(CAP_SYS_NICE)) {
3434 if (fair_policy(policy)) {
3435 if (attr->sched_nice < task_nice(p) &&
3436 !can_nice(p, attr->sched_nice))
3440 if (rt_policy(policy)) {
3441 unsigned long rlim_rtprio =
3442 task_rlimit(p, RLIMIT_RTPRIO);
3444 /* can't set/change the rt policy */
3445 if (policy != p->policy && !rlim_rtprio)
3448 /* can't increase priority */
3449 if (attr->sched_priority > p->rt_priority &&
3450 attr->sched_priority > rlim_rtprio)
3455 * Can't set/change SCHED_DEADLINE policy at all for now
3456 * (safest behavior); in the future we would like to allow
3457 * unprivileged DL tasks to increase their relative deadline
3458 * or reduce their runtime (both ways reducing utilization)
3460 if (dl_policy(policy))
3464 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3465 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3467 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3468 if (!can_nice(p, task_nice(p)))
3472 /* can't change other user's priorities */
3473 if (!check_same_owner(p))
3476 /* Normal users shall not reset the sched_reset_on_fork flag */
3477 if (p->sched_reset_on_fork && !reset_on_fork)
3482 retval = security_task_setscheduler(p);
3488 * make sure no PI-waiters arrive (or leave) while we are
3489 * changing the priority of the task:
3491 * To be able to change p->policy safely, the appropriate
3492 * runqueue lock must be held.
3494 rq = task_rq_lock(p, &flags);
3497 * Changing the policy of the stop threads its a very bad idea
3499 if (p == rq->stop) {
3500 task_rq_unlock(rq, p, &flags);
3505 * If not changing anything there's no need to proceed further,
3506 * but store a possible modification of reset_on_fork.
3508 if (unlikely(policy == p->policy)) {
3509 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3511 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3513 if (dl_policy(policy))
3516 p->sched_reset_on_fork = reset_on_fork;
3517 task_rq_unlock(rq, p, &flags);
3523 #ifdef CONFIG_RT_GROUP_SCHED
3525 * Do not allow realtime tasks into groups that have no runtime
3528 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3529 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3530 !task_group_is_autogroup(task_group(p))) {
3531 task_rq_unlock(rq, p, &flags);
3536 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3537 cpumask_t *span = rq->rd->span;
3540 * Don't allow tasks with an affinity mask smaller than
3541 * the entire root_domain to become SCHED_DEADLINE. We
3542 * will also fail if there's no bandwidth available.
3544 if (!cpumask_subset(span, &p->cpus_allowed) ||
3545 rq->rd->dl_bw.bw == 0) {
3546 task_rq_unlock(rq, p, &flags);
3553 /* recheck policy now with rq lock held */
3554 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3555 policy = oldpolicy = -1;
3556 task_rq_unlock(rq, p, &flags);
3561 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3562 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3565 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3566 task_rq_unlock(rq, p, &flags);
3570 p->sched_reset_on_fork = reset_on_fork;
3574 * Special case for priority boosted tasks.
3576 * If the new priority is lower or equal (user space view)
3577 * than the current (boosted) priority, we just store the new
3578 * normal parameters and do not touch the scheduler class and
3579 * the runqueue. This will be done when the task deboost
3582 if (rt_mutex_check_prio(p, newprio)) {
3583 __setscheduler_params(p, attr);
3584 task_rq_unlock(rq, p, &flags);
3588 queued = task_on_rq_queued(p);
3589 running = task_current(rq, p);
3591 dequeue_task(rq, p, 0);
3593 put_prev_task(rq, p);
3595 prev_class = p->sched_class;
3596 __setscheduler(rq, p, attr);
3599 p->sched_class->set_curr_task(rq);
3602 * We enqueue to tail when the priority of a task is
3603 * increased (user space view).
3605 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3608 check_class_changed(rq, p, prev_class, oldprio);
3609 task_rq_unlock(rq, p, &flags);
3611 rt_mutex_adjust_pi(p);
3616 static int _sched_setscheduler(struct task_struct *p, int policy,
3617 const struct sched_param *param, bool check)
3619 struct sched_attr attr = {
3620 .sched_policy = policy,
3621 .sched_priority = param->sched_priority,
3622 .sched_nice = PRIO_TO_NICE(p->static_prio),
3625 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
3626 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
3627 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3628 policy &= ~SCHED_RESET_ON_FORK;
3629 attr.sched_policy = policy;
3632 return __sched_setscheduler(p, &attr, check);
3635 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3636 * @p: the task in question.
3637 * @policy: new policy.
3638 * @param: structure containing the new RT priority.
3640 * Return: 0 on success. An error code otherwise.
3642 * NOTE that the task may be already dead.
3644 int sched_setscheduler(struct task_struct *p, int policy,
3645 const struct sched_param *param)
3647 return _sched_setscheduler(p, policy, param, true);
3649 EXPORT_SYMBOL_GPL(sched_setscheduler);
3651 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3653 return __sched_setscheduler(p, attr, true);
3655 EXPORT_SYMBOL_GPL(sched_setattr);
3658 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3659 * @p: the task in question.
3660 * @policy: new policy.
3661 * @param: structure containing the new RT priority.
3663 * Just like sched_setscheduler, only don't bother checking if the
3664 * current context has permission. For example, this is needed in
3665 * stop_machine(): we create temporary high priority worker threads,
3666 * but our caller might not have that capability.
3668 * Return: 0 on success. An error code otherwise.
3670 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3671 const struct sched_param *param)
3673 return _sched_setscheduler(p, policy, param, false);
3677 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3679 struct sched_param lparam;
3680 struct task_struct *p;
3683 if (!param || pid < 0)
3685 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3690 p = find_process_by_pid(pid);
3692 retval = sched_setscheduler(p, policy, &lparam);
3699 * Mimics kernel/events/core.c perf_copy_attr().
3701 static int sched_copy_attr(struct sched_attr __user *uattr,
3702 struct sched_attr *attr)
3707 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3711 * zero the full structure, so that a short copy will be nice.
3713 memset(attr, 0, sizeof(*attr));
3715 ret = get_user(size, &uattr->size);
3719 if (size > PAGE_SIZE) /* silly large */
3722 if (!size) /* abi compat */
3723 size = SCHED_ATTR_SIZE_VER0;
3725 if (size < SCHED_ATTR_SIZE_VER0)
3729 * If we're handed a bigger struct than we know of,
3730 * ensure all the unknown bits are 0 - i.e. new
3731 * user-space does not rely on any kernel feature
3732 * extensions we dont know about yet.
3734 if (size > sizeof(*attr)) {
3735 unsigned char __user *addr;
3736 unsigned char __user *end;
3739 addr = (void __user *)uattr + sizeof(*attr);
3740 end = (void __user *)uattr + size;
3742 for (; addr < end; addr++) {
3743 ret = get_user(val, addr);
3749 size = sizeof(*attr);
3752 ret = copy_from_user(attr, uattr, size);
3757 * XXX: do we want to be lenient like existing syscalls; or do we want
3758 * to be strict and return an error on out-of-bounds values?
3760 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3765 put_user(sizeof(*attr), &uattr->size);
3770 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3771 * @pid: the pid in question.
3772 * @policy: new policy.
3773 * @param: structure containing the new RT priority.
3775 * Return: 0 on success. An error code otherwise.
3777 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3778 struct sched_param __user *, param)
3780 /* negative values for policy are not valid */
3784 return do_sched_setscheduler(pid, policy, param);
3788 * sys_sched_setparam - set/change the RT priority of a thread
3789 * @pid: the pid in question.
3790 * @param: structure containing the new RT priority.
3792 * Return: 0 on success. An error code otherwise.
3794 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3796 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
3800 * sys_sched_setattr - same as above, but with extended sched_attr
3801 * @pid: the pid in question.
3802 * @uattr: structure containing the extended parameters.
3803 * @flags: for future extension.
3805 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3806 unsigned int, flags)
3808 struct sched_attr attr;
3809 struct task_struct *p;
3812 if (!uattr || pid < 0 || flags)
3815 retval = sched_copy_attr(uattr, &attr);
3819 if ((int)attr.sched_policy < 0)
3824 p = find_process_by_pid(pid);
3826 retval = sched_setattr(p, &attr);
3833 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3834 * @pid: the pid in question.
3836 * Return: On success, the policy of the thread. Otherwise, a negative error
3839 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3841 struct task_struct *p;
3849 p = find_process_by_pid(pid);
3851 retval = security_task_getscheduler(p);
3854 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3861 * sys_sched_getparam - get the RT priority of a thread
3862 * @pid: the pid in question.
3863 * @param: structure containing the RT priority.
3865 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3868 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3870 struct sched_param lp = { .sched_priority = 0 };
3871 struct task_struct *p;
3874 if (!param || pid < 0)
3878 p = find_process_by_pid(pid);
3883 retval = security_task_getscheduler(p);
3887 if (task_has_rt_policy(p))
3888 lp.sched_priority = p->rt_priority;
3892 * This one might sleep, we cannot do it with a spinlock held ...
3894 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3903 static int sched_read_attr(struct sched_attr __user *uattr,
3904 struct sched_attr *attr,
3909 if (!access_ok(VERIFY_WRITE, uattr, usize))
3913 * If we're handed a smaller struct than we know of,
3914 * ensure all the unknown bits are 0 - i.e. old
3915 * user-space does not get uncomplete information.
3917 if (usize < sizeof(*attr)) {
3918 unsigned char *addr;
3921 addr = (void *)attr + usize;
3922 end = (void *)attr + sizeof(*attr);
3924 for (; addr < end; addr++) {
3932 ret = copy_to_user(uattr, attr, attr->size);
3940 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3941 * @pid: the pid in question.
3942 * @uattr: structure containing the extended parameters.
3943 * @size: sizeof(attr) for fwd/bwd comp.
3944 * @flags: for future extension.
3946 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3947 unsigned int, size, unsigned int, flags)
3949 struct sched_attr attr = {
3950 .size = sizeof(struct sched_attr),
3952 struct task_struct *p;
3955 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3956 size < SCHED_ATTR_SIZE_VER0 || flags)
3960 p = find_process_by_pid(pid);
3965 retval = security_task_getscheduler(p);
3969 attr.sched_policy = p->policy;
3970 if (p->sched_reset_on_fork)
3971 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3972 if (task_has_dl_policy(p))
3973 __getparam_dl(p, &attr);
3974 else if (task_has_rt_policy(p))
3975 attr.sched_priority = p->rt_priority;
3977 attr.sched_nice = task_nice(p);
3981 retval = sched_read_attr(uattr, &attr, size);
3989 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3991 cpumask_var_t cpus_allowed, new_mask;
3992 struct task_struct *p;
3997 p = find_process_by_pid(pid);
4003 /* Prevent p going away */
4007 if (p->flags & PF_NO_SETAFFINITY) {
4011 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4015 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4017 goto out_free_cpus_allowed;
4020 if (!check_same_owner(p)) {
4022 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4024 goto out_free_new_mask;
4029 retval = security_task_setscheduler(p);
4031 goto out_free_new_mask;
4034 cpuset_cpus_allowed(p, cpus_allowed);
4035 cpumask_and(new_mask, in_mask, cpus_allowed);
4038 * Since bandwidth control happens on root_domain basis,
4039 * if admission test is enabled, we only admit -deadline
4040 * tasks allowed to run on all the CPUs in the task's
4044 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4046 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4049 goto out_free_new_mask;
4055 retval = set_cpus_allowed_ptr(p, new_mask);
4058 cpuset_cpus_allowed(p, cpus_allowed);
4059 if (!cpumask_subset(new_mask, cpus_allowed)) {
4061 * We must have raced with a concurrent cpuset
4062 * update. Just reset the cpus_allowed to the
4063 * cpuset's cpus_allowed
4065 cpumask_copy(new_mask, cpus_allowed);
4070 free_cpumask_var(new_mask);
4071 out_free_cpus_allowed:
4072 free_cpumask_var(cpus_allowed);
4078 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4079 struct cpumask *new_mask)
4081 if (len < cpumask_size())
4082 cpumask_clear(new_mask);
4083 else if (len > cpumask_size())
4084 len = cpumask_size();
4086 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4090 * sys_sched_setaffinity - set the cpu affinity of a process
4091 * @pid: pid of the process
4092 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4093 * @user_mask_ptr: user-space pointer to the new cpu mask
4095 * Return: 0 on success. An error code otherwise.
4097 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4098 unsigned long __user *, user_mask_ptr)
4100 cpumask_var_t new_mask;
4103 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4106 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4108 retval = sched_setaffinity(pid, new_mask);
4109 free_cpumask_var(new_mask);
4113 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4115 struct task_struct *p;
4116 unsigned long flags;
4122 p = find_process_by_pid(pid);
4126 retval = security_task_getscheduler(p);
4130 raw_spin_lock_irqsave(&p->pi_lock, flags);
4131 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4132 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4141 * sys_sched_getaffinity - get the cpu affinity of a process
4142 * @pid: pid of the process
4143 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4144 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4146 * Return: 0 on success. An error code otherwise.
4148 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4149 unsigned long __user *, user_mask_ptr)
4154 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4156 if (len & (sizeof(unsigned long)-1))
4159 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4162 ret = sched_getaffinity(pid, mask);
4164 size_t retlen = min_t(size_t, len, cpumask_size());
4166 if (copy_to_user(user_mask_ptr, mask, retlen))
4171 free_cpumask_var(mask);
4177 * sys_sched_yield - yield the current processor to other threads.
4179 * This function yields the current CPU to other tasks. If there are no
4180 * other threads running on this CPU then this function will return.
4184 SYSCALL_DEFINE0(sched_yield)
4186 struct rq *rq = this_rq_lock();
4188 schedstat_inc(rq, yld_count);
4189 current->sched_class->yield_task(rq);
4192 * Since we are going to call schedule() anyway, there's
4193 * no need to preempt or enable interrupts:
4195 __release(rq->lock);
4196 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4197 do_raw_spin_unlock(&rq->lock);
4198 sched_preempt_enable_no_resched();
4205 static void __cond_resched(void)
4207 __preempt_count_add(PREEMPT_ACTIVE);
4209 __preempt_count_sub(PREEMPT_ACTIVE);
4212 int __sched _cond_resched(void)
4214 if (should_resched()) {
4220 EXPORT_SYMBOL(_cond_resched);
4223 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4224 * call schedule, and on return reacquire the lock.
4226 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4227 * operations here to prevent schedule() from being called twice (once via
4228 * spin_unlock(), once by hand).
4230 int __cond_resched_lock(spinlock_t *lock)
4232 int resched = should_resched();
4235 lockdep_assert_held(lock);
4237 if (spin_needbreak(lock) || resched) {
4248 EXPORT_SYMBOL(__cond_resched_lock);
4250 int __sched __cond_resched_softirq(void)
4252 BUG_ON(!in_softirq());
4254 if (should_resched()) {
4262 EXPORT_SYMBOL(__cond_resched_softirq);
4265 * yield - yield the current processor to other threads.
4267 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4269 * The scheduler is at all times free to pick the calling task as the most
4270 * eligible task to run, if removing the yield() call from your code breaks
4271 * it, its already broken.
4273 * Typical broken usage is:
4278 * where one assumes that yield() will let 'the other' process run that will
4279 * make event true. If the current task is a SCHED_FIFO task that will never
4280 * happen. Never use yield() as a progress guarantee!!
4282 * If you want to use yield() to wait for something, use wait_event().
4283 * If you want to use yield() to be 'nice' for others, use cond_resched().
4284 * If you still want to use yield(), do not!
4286 void __sched yield(void)
4288 set_current_state(TASK_RUNNING);
4291 EXPORT_SYMBOL(yield);
4294 * yield_to - yield the current processor to another thread in
4295 * your thread group, or accelerate that thread toward the
4296 * processor it's on.
4298 * @preempt: whether task preemption is allowed or not
4300 * It's the caller's job to ensure that the target task struct
4301 * can't go away on us before we can do any checks.
4304 * true (>0) if we indeed boosted the target task.
4305 * false (0) if we failed to boost the target.
4306 * -ESRCH if there's no task to yield to.
4308 int __sched yield_to(struct task_struct *p, bool preempt)
4310 struct task_struct *curr = current;
4311 struct rq *rq, *p_rq;
4312 unsigned long flags;
4315 local_irq_save(flags);
4321 * If we're the only runnable task on the rq and target rq also
4322 * has only one task, there's absolutely no point in yielding.
4324 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4329 double_rq_lock(rq, p_rq);
4330 if (task_rq(p) != p_rq) {
4331 double_rq_unlock(rq, p_rq);
4335 if (!curr->sched_class->yield_to_task)
4338 if (curr->sched_class != p->sched_class)
4341 if (task_running(p_rq, p) || p->state)
4344 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4346 schedstat_inc(rq, yld_count);
4348 * Make p's CPU reschedule; pick_next_entity takes care of
4351 if (preempt && rq != p_rq)
4356 double_rq_unlock(rq, p_rq);
4358 local_irq_restore(flags);
4365 EXPORT_SYMBOL_GPL(yield_to);
4368 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4369 * that process accounting knows that this is a task in IO wait state.
4371 void __sched io_schedule(void)
4373 struct rq *rq = raw_rq();
4375 delayacct_blkio_start();
4376 atomic_inc(&rq->nr_iowait);
4377 blk_flush_plug(current);
4378 current->in_iowait = 1;
4380 current->in_iowait = 0;
4381 atomic_dec(&rq->nr_iowait);
4382 delayacct_blkio_end();
4384 EXPORT_SYMBOL(io_schedule);
4386 long __sched io_schedule_timeout(long timeout)
4388 struct rq *rq = raw_rq();
4391 delayacct_blkio_start();
4392 atomic_inc(&rq->nr_iowait);
4393 blk_flush_plug(current);
4394 current->in_iowait = 1;
4395 ret = schedule_timeout(timeout);
4396 current->in_iowait = 0;
4397 atomic_dec(&rq->nr_iowait);
4398 delayacct_blkio_end();
4403 * sys_sched_get_priority_max - return maximum RT priority.
4404 * @policy: scheduling class.
4406 * Return: On success, this syscall returns the maximum
4407 * rt_priority that can be used by a given scheduling class.
4408 * On failure, a negative error code is returned.
4410 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4417 ret = MAX_USER_RT_PRIO-1;
4419 case SCHED_DEADLINE:
4430 * sys_sched_get_priority_min - return minimum RT priority.
4431 * @policy: scheduling class.
4433 * Return: On success, this syscall returns the minimum
4434 * rt_priority that can be used by a given scheduling class.
4435 * On failure, a negative error code is returned.
4437 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4446 case SCHED_DEADLINE:
4456 * sys_sched_rr_get_interval - return the default timeslice of a process.
4457 * @pid: pid of the process.
4458 * @interval: userspace pointer to the timeslice value.
4460 * this syscall writes the default timeslice value of a given process
4461 * into the user-space timespec buffer. A value of '0' means infinity.
4463 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4466 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4467 struct timespec __user *, interval)
4469 struct task_struct *p;
4470 unsigned int time_slice;
4471 unsigned long flags;
4481 p = find_process_by_pid(pid);
4485 retval = security_task_getscheduler(p);
4489 rq = task_rq_lock(p, &flags);
4491 if (p->sched_class->get_rr_interval)
4492 time_slice = p->sched_class->get_rr_interval(rq, p);
4493 task_rq_unlock(rq, p, &flags);
4496 jiffies_to_timespec(time_slice, &t);
4497 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4505 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4507 void sched_show_task(struct task_struct *p)
4509 unsigned long free = 0;
4513 state = p->state ? __ffs(p->state) + 1 : 0;
4514 printk(KERN_INFO "%-15.15s %c", p->comm,
4515 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4516 #if BITS_PER_LONG == 32
4517 if (state == TASK_RUNNING)
4518 printk(KERN_CONT " running ");
4520 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4522 if (state == TASK_RUNNING)
4523 printk(KERN_CONT " running task ");
4525 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4527 #ifdef CONFIG_DEBUG_STACK_USAGE
4528 free = stack_not_used(p);
4531 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4533 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4534 task_pid_nr(p), ppid,
4535 (unsigned long)task_thread_info(p)->flags);
4537 print_worker_info(KERN_INFO, p);
4538 show_stack(p, NULL);
4541 void show_state_filter(unsigned long state_filter)
4543 struct task_struct *g, *p;
4545 #if BITS_PER_LONG == 32
4547 " task PC stack pid father\n");
4550 " task PC stack pid father\n");
4553 for_each_process_thread(g, p) {
4555 * reset the NMI-timeout, listing all files on a slow
4556 * console might take a lot of time:
4558 touch_nmi_watchdog();
4559 if (!state_filter || (p->state & state_filter))
4563 touch_all_softlockup_watchdogs();
4565 #ifdef CONFIG_SCHED_DEBUG
4566 sysrq_sched_debug_show();
4570 * Only show locks if all tasks are dumped:
4573 debug_show_all_locks();
4576 void init_idle_bootup_task(struct task_struct *idle)
4578 idle->sched_class = &idle_sched_class;
4582 * init_idle - set up an idle thread for a given CPU
4583 * @idle: task in question
4584 * @cpu: cpu the idle task belongs to
4586 * NOTE: this function does not set the idle thread's NEED_RESCHED
4587 * flag, to make booting more robust.
4589 void init_idle(struct task_struct *idle, int cpu)
4591 struct rq *rq = cpu_rq(cpu);
4592 unsigned long flags;
4594 raw_spin_lock_irqsave(&rq->lock, flags);
4596 __sched_fork(0, idle);
4597 idle->state = TASK_RUNNING;
4598 idle->se.exec_start = sched_clock();
4600 do_set_cpus_allowed(idle, cpumask_of(cpu));
4602 * We're having a chicken and egg problem, even though we are
4603 * holding rq->lock, the cpu isn't yet set to this cpu so the
4604 * lockdep check in task_group() will fail.
4606 * Similar case to sched_fork(). / Alternatively we could
4607 * use task_rq_lock() here and obtain the other rq->lock.
4612 __set_task_cpu(idle, cpu);
4615 rq->curr = rq->idle = idle;
4616 idle->on_rq = TASK_ON_RQ_QUEUED;
4617 #if defined(CONFIG_SMP)
4620 raw_spin_unlock_irqrestore(&rq->lock, flags);
4622 /* Set the preempt count _outside_ the spinlocks! */
4623 init_idle_preempt_count(idle, cpu);
4626 * The idle tasks have their own, simple scheduling class:
4628 idle->sched_class = &idle_sched_class;
4629 ftrace_graph_init_idle_task(idle, cpu);
4630 vtime_init_idle(idle, cpu);
4631 #if defined(CONFIG_SMP)
4632 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4636 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
4637 const struct cpumask *trial)
4639 int ret = 1, trial_cpus;
4640 struct dl_bw *cur_dl_b;
4641 unsigned long flags;
4643 rcu_read_lock_sched();
4644 cur_dl_b = dl_bw_of(cpumask_any(cur));
4645 trial_cpus = cpumask_weight(trial);
4647 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
4648 if (cur_dl_b->bw != -1 &&
4649 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
4651 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
4652 rcu_read_unlock_sched();
4657 int task_can_attach(struct task_struct *p,
4658 const struct cpumask *cs_cpus_allowed)
4663 * Kthreads which disallow setaffinity shouldn't be moved
4664 * to a new cpuset; we don't want to change their cpu
4665 * affinity and isolating such threads by their set of
4666 * allowed nodes is unnecessary. Thus, cpusets are not
4667 * applicable for such threads. This prevents checking for
4668 * success of set_cpus_allowed_ptr() on all attached tasks
4669 * before cpus_allowed may be changed.
4671 if (p->flags & PF_NO_SETAFFINITY) {
4677 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
4679 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
4684 unsigned long flags;
4686 rcu_read_lock_sched();
4687 dl_b = dl_bw_of(dest_cpu);
4688 raw_spin_lock_irqsave(&dl_b->lock, flags);
4689 cpus = dl_bw_cpus(dest_cpu);
4690 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
4695 * We reserve space for this task in the destination
4696 * root_domain, as we can't fail after this point.
4697 * We will free resources in the source root_domain
4698 * later on (see set_cpus_allowed_dl()).
4700 __dl_add(dl_b, p->dl.dl_bw);
4702 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
4703 rcu_read_unlock_sched();
4713 * move_queued_task - move a queued task to new rq.
4715 * Returns (locked) new rq. Old rq's lock is released.
4717 static struct rq *move_queued_task(struct task_struct *p, int new_cpu)
4719 struct rq *rq = task_rq(p);
4721 lockdep_assert_held(&rq->lock);
4723 dequeue_task(rq, p, 0);
4724 p->on_rq = TASK_ON_RQ_MIGRATING;
4725 set_task_cpu(p, new_cpu);
4726 raw_spin_unlock(&rq->lock);
4728 rq = cpu_rq(new_cpu);
4730 raw_spin_lock(&rq->lock);
4731 BUG_ON(task_cpu(p) != new_cpu);
4732 p->on_rq = TASK_ON_RQ_QUEUED;
4733 enqueue_task(rq, p, 0);
4734 check_preempt_curr(rq, p, 0);
4739 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4741 if (p->sched_class && p->sched_class->set_cpus_allowed)
4742 p->sched_class->set_cpus_allowed(p, new_mask);
4744 cpumask_copy(&p->cpus_allowed, new_mask);
4745 p->nr_cpus_allowed = cpumask_weight(new_mask);
4749 * This is how migration works:
4751 * 1) we invoke migration_cpu_stop() on the target CPU using
4753 * 2) stopper starts to run (implicitly forcing the migrated thread
4755 * 3) it checks whether the migrated task is still in the wrong runqueue.
4756 * 4) if it's in the wrong runqueue then the migration thread removes
4757 * it and puts it into the right queue.
4758 * 5) stopper completes and stop_one_cpu() returns and the migration
4763 * Change a given task's CPU affinity. Migrate the thread to a
4764 * proper CPU and schedule it away if the CPU it's executing on
4765 * is removed from the allowed bitmask.
4767 * NOTE: the caller must have a valid reference to the task, the
4768 * task must not exit() & deallocate itself prematurely. The
4769 * call is not atomic; no spinlocks may be held.
4771 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4773 unsigned long flags;
4775 unsigned int dest_cpu;
4778 rq = task_rq_lock(p, &flags);
4780 if (cpumask_equal(&p->cpus_allowed, new_mask))
4783 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4788 do_set_cpus_allowed(p, new_mask);
4790 /* Can the task run on the task's current CPU? If so, we're done */
4791 if (cpumask_test_cpu(task_cpu(p), new_mask))
4794 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4795 if (task_running(rq, p) || p->state == TASK_WAKING) {
4796 struct migration_arg arg = { p, dest_cpu };
4797 /* Need help from migration thread: drop lock and wait. */
4798 task_rq_unlock(rq, p, &flags);
4799 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4800 tlb_migrate_finish(p->mm);
4802 } else if (task_on_rq_queued(p))
4803 rq = move_queued_task(p, dest_cpu);
4805 task_rq_unlock(rq, p, &flags);
4809 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4812 * Move (not current) task off this cpu, onto dest cpu. We're doing
4813 * this because either it can't run here any more (set_cpus_allowed()
4814 * away from this CPU, or CPU going down), or because we're
4815 * attempting to rebalance this task on exec (sched_exec).
4817 * So we race with normal scheduler movements, but that's OK, as long
4818 * as the task is no longer on this CPU.
4820 * Returns non-zero if task was successfully migrated.
4822 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4827 if (unlikely(!cpu_active(dest_cpu)))
4830 rq = cpu_rq(src_cpu);
4832 raw_spin_lock(&p->pi_lock);
4833 raw_spin_lock(&rq->lock);
4834 /* Already moved. */
4835 if (task_cpu(p) != src_cpu)
4838 /* Affinity changed (again). */
4839 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4843 * If we're not on a rq, the next wake-up will ensure we're
4846 if (task_on_rq_queued(p))
4847 rq = move_queued_task(p, dest_cpu);
4851 raw_spin_unlock(&rq->lock);
4852 raw_spin_unlock(&p->pi_lock);
4856 #ifdef CONFIG_NUMA_BALANCING
4857 /* Migrate current task p to target_cpu */
4858 int migrate_task_to(struct task_struct *p, int target_cpu)
4860 struct migration_arg arg = { p, target_cpu };
4861 int curr_cpu = task_cpu(p);
4863 if (curr_cpu == target_cpu)
4866 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4869 /* TODO: This is not properly updating schedstats */
4871 trace_sched_move_numa(p, curr_cpu, target_cpu);
4872 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4876 * Requeue a task on a given node and accurately track the number of NUMA
4877 * tasks on the runqueues
4879 void sched_setnuma(struct task_struct *p, int nid)
4882 unsigned long flags;
4883 bool queued, running;
4885 rq = task_rq_lock(p, &flags);
4886 queued = task_on_rq_queued(p);
4887 running = task_current(rq, p);
4890 dequeue_task(rq, p, 0);
4892 put_prev_task(rq, p);
4894 p->numa_preferred_nid = nid;
4897 p->sched_class->set_curr_task(rq);
4899 enqueue_task(rq, p, 0);
4900 task_rq_unlock(rq, p, &flags);
4905 * migration_cpu_stop - this will be executed by a highprio stopper thread
4906 * and performs thread migration by bumping thread off CPU then
4907 * 'pushing' onto another runqueue.
4909 static int migration_cpu_stop(void *data)
4911 struct migration_arg *arg = data;
4914 * The original target cpu might have gone down and we might
4915 * be on another cpu but it doesn't matter.
4917 local_irq_disable();
4919 * We need to explicitly wake pending tasks before running
4920 * __migrate_task() such that we will not miss enforcing cpus_allowed
4921 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
4923 sched_ttwu_pending();
4924 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4929 #ifdef CONFIG_HOTPLUG_CPU
4932 * Ensures that the idle task is using init_mm right before its cpu goes
4935 void idle_task_exit(void)
4937 struct mm_struct *mm = current->active_mm;
4939 BUG_ON(cpu_online(smp_processor_id()));
4941 if (mm != &init_mm) {
4942 switch_mm(mm, &init_mm, current);
4943 finish_arch_post_lock_switch();
4949 * Since this CPU is going 'away' for a while, fold any nr_active delta
4950 * we might have. Assumes we're called after migrate_tasks() so that the
4951 * nr_active count is stable.
4953 * Also see the comment "Global load-average calculations".
4955 static void calc_load_migrate(struct rq *rq)
4957 long delta = calc_load_fold_active(rq);
4959 atomic_long_add(delta, &calc_load_tasks);
4962 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4966 static const struct sched_class fake_sched_class = {
4967 .put_prev_task = put_prev_task_fake,
4970 static struct task_struct fake_task = {
4972 * Avoid pull_{rt,dl}_task()
4974 .prio = MAX_PRIO + 1,
4975 .sched_class = &fake_sched_class,
4979 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4980 * try_to_wake_up()->select_task_rq().
4982 * Called with rq->lock held even though we'er in stop_machine() and
4983 * there's no concurrency possible, we hold the required locks anyway
4984 * because of lock validation efforts.
4986 static void migrate_tasks(unsigned int dead_cpu)
4988 struct rq *rq = cpu_rq(dead_cpu);
4989 struct task_struct *next, *stop = rq->stop;
4993 * Fudge the rq selection such that the below task selection loop
4994 * doesn't get stuck on the currently eligible stop task.
4996 * We're currently inside stop_machine() and the rq is either stuck
4997 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4998 * either way we should never end up calling schedule() until we're
5004 * put_prev_task() and pick_next_task() sched
5005 * class method both need to have an up-to-date
5006 * value of rq->clock[_task]
5008 update_rq_clock(rq);
5012 * There's this thread running, bail when that's the only
5015 if (rq->nr_running == 1)
5018 next = pick_next_task(rq, &fake_task);
5020 next->sched_class->put_prev_task(rq, next);
5022 /* Find suitable destination for @next, with force if needed. */
5023 dest_cpu = select_fallback_rq(dead_cpu, next);
5024 raw_spin_unlock(&rq->lock);
5026 __migrate_task(next, dead_cpu, dest_cpu);
5028 raw_spin_lock(&rq->lock);
5034 #endif /* CONFIG_HOTPLUG_CPU */
5036 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5038 static struct ctl_table sd_ctl_dir[] = {
5040 .procname = "sched_domain",
5046 static struct ctl_table sd_ctl_root[] = {
5048 .procname = "kernel",
5050 .child = sd_ctl_dir,
5055 static struct ctl_table *sd_alloc_ctl_entry(int n)
5057 struct ctl_table *entry =
5058 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5063 static void sd_free_ctl_entry(struct ctl_table **tablep)
5065 struct ctl_table *entry;
5068 * In the intermediate directories, both the child directory and
5069 * procname are dynamically allocated and could fail but the mode
5070 * will always be set. In the lowest directory the names are
5071 * static strings and all have proc handlers.
5073 for (entry = *tablep; entry->mode; entry++) {
5075 sd_free_ctl_entry(&entry->child);
5076 if (entry->proc_handler == NULL)
5077 kfree(entry->procname);
5084 static int min_load_idx = 0;
5085 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5088 set_table_entry(struct ctl_table *entry,
5089 const char *procname, void *data, int maxlen,
5090 umode_t mode, proc_handler *proc_handler,
5093 entry->procname = procname;
5095 entry->maxlen = maxlen;
5097 entry->proc_handler = proc_handler;
5100 entry->extra1 = &min_load_idx;
5101 entry->extra2 = &max_load_idx;
5105 static struct ctl_table *
5106 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5108 struct ctl_table *table = sd_alloc_ctl_entry(14);
5113 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5114 sizeof(long), 0644, proc_doulongvec_minmax, false);
5115 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5116 sizeof(long), 0644, proc_doulongvec_minmax, false);
5117 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5118 sizeof(int), 0644, proc_dointvec_minmax, true);
5119 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5120 sizeof(int), 0644, proc_dointvec_minmax, true);
5121 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5122 sizeof(int), 0644, proc_dointvec_minmax, true);
5123 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5124 sizeof(int), 0644, proc_dointvec_minmax, true);
5125 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5126 sizeof(int), 0644, proc_dointvec_minmax, true);
5127 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5128 sizeof(int), 0644, proc_dointvec_minmax, false);
5129 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5130 sizeof(int), 0644, proc_dointvec_minmax, false);
5131 set_table_entry(&table[9], "cache_nice_tries",
5132 &sd->cache_nice_tries,
5133 sizeof(int), 0644, proc_dointvec_minmax, false);
5134 set_table_entry(&table[10], "flags", &sd->flags,
5135 sizeof(int), 0644, proc_dointvec_minmax, false);
5136 set_table_entry(&table[11], "max_newidle_lb_cost",
5137 &sd->max_newidle_lb_cost,
5138 sizeof(long), 0644, proc_doulongvec_minmax, false);
5139 set_table_entry(&table[12], "name", sd->name,
5140 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5141 /* &table[13] is terminator */
5146 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5148 struct ctl_table *entry, *table;
5149 struct sched_domain *sd;
5150 int domain_num = 0, i;
5153 for_each_domain(cpu, sd)
5155 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5160 for_each_domain(cpu, sd) {
5161 snprintf(buf, 32, "domain%d", i);
5162 entry->procname = kstrdup(buf, GFP_KERNEL);
5164 entry->child = sd_alloc_ctl_domain_table(sd);
5171 static struct ctl_table_header *sd_sysctl_header;
5172 static void register_sched_domain_sysctl(void)
5174 int i, cpu_num = num_possible_cpus();
5175 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5178 WARN_ON(sd_ctl_dir[0].child);
5179 sd_ctl_dir[0].child = entry;
5184 for_each_possible_cpu(i) {
5185 snprintf(buf, 32, "cpu%d", i);
5186 entry->procname = kstrdup(buf, GFP_KERNEL);
5188 entry->child = sd_alloc_ctl_cpu_table(i);
5192 WARN_ON(sd_sysctl_header);
5193 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5196 /* may be called multiple times per register */
5197 static void unregister_sched_domain_sysctl(void)
5199 if (sd_sysctl_header)
5200 unregister_sysctl_table(sd_sysctl_header);
5201 sd_sysctl_header = NULL;
5202 if (sd_ctl_dir[0].child)
5203 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5206 static void register_sched_domain_sysctl(void)
5209 static void unregister_sched_domain_sysctl(void)
5214 static void set_rq_online(struct rq *rq)
5217 const struct sched_class *class;
5219 cpumask_set_cpu(rq->cpu, rq->rd->online);
5222 for_each_class(class) {
5223 if (class->rq_online)
5224 class->rq_online(rq);
5229 static void set_rq_offline(struct rq *rq)
5232 const struct sched_class *class;
5234 for_each_class(class) {
5235 if (class->rq_offline)
5236 class->rq_offline(rq);
5239 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5245 * migration_call - callback that gets triggered when a CPU is added.
5246 * Here we can start up the necessary migration thread for the new CPU.
5249 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5251 int cpu = (long)hcpu;
5252 unsigned long flags;
5253 struct rq *rq = cpu_rq(cpu);
5255 switch (action & ~CPU_TASKS_FROZEN) {
5257 case CPU_UP_PREPARE:
5258 rq->calc_load_update = calc_load_update;
5262 /* Update our root-domain */
5263 raw_spin_lock_irqsave(&rq->lock, flags);
5265 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5269 raw_spin_unlock_irqrestore(&rq->lock, flags);
5272 #ifdef CONFIG_HOTPLUG_CPU
5274 sched_ttwu_pending();
5275 /* Update our root-domain */
5276 raw_spin_lock_irqsave(&rq->lock, flags);
5278 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5282 BUG_ON(rq->nr_running != 1); /* the migration thread */
5283 raw_spin_unlock_irqrestore(&rq->lock, flags);
5287 calc_load_migrate(rq);
5292 update_max_interval();
5298 * Register at high priority so that task migration (migrate_all_tasks)
5299 * happens before everything else. This has to be lower priority than
5300 * the notifier in the perf_event subsystem, though.
5302 static struct notifier_block migration_notifier = {
5303 .notifier_call = migration_call,
5304 .priority = CPU_PRI_MIGRATION,
5307 static void __cpuinit set_cpu_rq_start_time(void)
5309 int cpu = smp_processor_id();
5310 struct rq *rq = cpu_rq(cpu);
5311 rq->age_stamp = sched_clock_cpu(cpu);
5314 static int sched_cpu_active(struct notifier_block *nfb,
5315 unsigned long action, void *hcpu)
5317 switch (action & ~CPU_TASKS_FROZEN) {
5319 set_cpu_rq_start_time();
5321 case CPU_DOWN_FAILED:
5322 set_cpu_active((long)hcpu, true);
5329 static int sched_cpu_inactive(struct notifier_block *nfb,
5330 unsigned long action, void *hcpu)
5332 unsigned long flags;
5333 long cpu = (long)hcpu;
5336 switch (action & ~CPU_TASKS_FROZEN) {
5337 case CPU_DOWN_PREPARE:
5338 set_cpu_active(cpu, false);
5340 /* explicitly allow suspend */
5341 if (!(action & CPU_TASKS_FROZEN)) {
5345 rcu_read_lock_sched();
5346 dl_b = dl_bw_of(cpu);
5348 raw_spin_lock_irqsave(&dl_b->lock, flags);
5349 cpus = dl_bw_cpus(cpu);
5350 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5351 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5353 rcu_read_unlock_sched();
5356 return notifier_from_errno(-EBUSY);
5364 static int __init migration_init(void)
5366 void *cpu = (void *)(long)smp_processor_id();
5369 /* Initialize migration for the boot CPU */
5370 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5371 BUG_ON(err == NOTIFY_BAD);
5372 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5373 register_cpu_notifier(&migration_notifier);
5375 /* Register cpu active notifiers */
5376 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5377 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5381 early_initcall(migration_init);
5386 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5388 #ifdef CONFIG_SCHED_DEBUG
5390 static __read_mostly int sched_debug_enabled;
5392 static int __init sched_debug_setup(char *str)
5394 sched_debug_enabled = 1;
5398 early_param("sched_debug", sched_debug_setup);
5400 static inline bool sched_debug(void)
5402 return sched_debug_enabled;
5405 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5406 struct cpumask *groupmask)
5408 struct sched_group *group = sd->groups;
5411 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5412 cpumask_clear(groupmask);
5414 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5416 if (!(sd->flags & SD_LOAD_BALANCE)) {
5417 printk("does not load-balance\n");
5419 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5424 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5426 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5427 printk(KERN_ERR "ERROR: domain->span does not contain "
5430 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5431 printk(KERN_ERR "ERROR: domain->groups does not contain"
5435 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5439 printk(KERN_ERR "ERROR: group is NULL\n");
5444 * Even though we initialize ->capacity to something semi-sane,
5445 * we leave capacity_orig unset. This allows us to detect if
5446 * domain iteration is still funny without causing /0 traps.
5448 if (!group->sgc->capacity_orig) {
5449 printk(KERN_CONT "\n");
5450 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5454 if (!cpumask_weight(sched_group_cpus(group))) {
5455 printk(KERN_CONT "\n");
5456 printk(KERN_ERR "ERROR: empty group\n");
5460 if (!(sd->flags & SD_OVERLAP) &&
5461 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5462 printk(KERN_CONT "\n");
5463 printk(KERN_ERR "ERROR: repeated CPUs\n");
5467 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5469 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5471 printk(KERN_CONT " %s", str);
5472 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5473 printk(KERN_CONT " (cpu_capacity = %d)",
5474 group->sgc->capacity);
5477 group = group->next;
5478 } while (group != sd->groups);
5479 printk(KERN_CONT "\n");
5481 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5482 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5485 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5486 printk(KERN_ERR "ERROR: parent span is not a superset "
5487 "of domain->span\n");
5491 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5495 if (!sched_debug_enabled)
5499 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5503 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5506 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5514 #else /* !CONFIG_SCHED_DEBUG */
5515 # define sched_domain_debug(sd, cpu) do { } while (0)
5516 static inline bool sched_debug(void)
5520 #endif /* CONFIG_SCHED_DEBUG */
5522 static int sd_degenerate(struct sched_domain *sd)
5524 if (cpumask_weight(sched_domain_span(sd)) == 1)
5527 /* Following flags need at least 2 groups */
5528 if (sd->flags & (SD_LOAD_BALANCE |
5529 SD_BALANCE_NEWIDLE |
5532 SD_SHARE_CPUCAPACITY |
5533 SD_SHARE_PKG_RESOURCES |
5534 SD_SHARE_POWERDOMAIN)) {
5535 if (sd->groups != sd->groups->next)
5539 /* Following flags don't use groups */
5540 if (sd->flags & (SD_WAKE_AFFINE))
5547 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5549 unsigned long cflags = sd->flags, pflags = parent->flags;
5551 if (sd_degenerate(parent))
5554 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5557 /* Flags needing groups don't count if only 1 group in parent */
5558 if (parent->groups == parent->groups->next) {
5559 pflags &= ~(SD_LOAD_BALANCE |
5560 SD_BALANCE_NEWIDLE |
5563 SD_SHARE_CPUCAPACITY |
5564 SD_SHARE_PKG_RESOURCES |
5566 SD_SHARE_POWERDOMAIN);
5567 if (nr_node_ids == 1)
5568 pflags &= ~SD_SERIALIZE;
5570 if (~cflags & pflags)
5576 static void free_rootdomain(struct rcu_head *rcu)
5578 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5580 cpupri_cleanup(&rd->cpupri);
5581 cpudl_cleanup(&rd->cpudl);
5582 free_cpumask_var(rd->dlo_mask);
5583 free_cpumask_var(rd->rto_mask);
5584 free_cpumask_var(rd->online);
5585 free_cpumask_var(rd->span);
5589 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5591 struct root_domain *old_rd = NULL;
5592 unsigned long flags;
5594 raw_spin_lock_irqsave(&rq->lock, flags);
5599 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5602 cpumask_clear_cpu(rq->cpu, old_rd->span);
5605 * If we dont want to free the old_rd yet then
5606 * set old_rd to NULL to skip the freeing later
5609 if (!atomic_dec_and_test(&old_rd->refcount))
5613 atomic_inc(&rd->refcount);
5616 cpumask_set_cpu(rq->cpu, rd->span);
5617 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5620 raw_spin_unlock_irqrestore(&rq->lock, flags);
5623 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5626 static int init_rootdomain(struct root_domain *rd)
5628 memset(rd, 0, sizeof(*rd));
5630 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5632 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5634 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5636 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5639 init_dl_bw(&rd->dl_bw);
5640 if (cpudl_init(&rd->cpudl) != 0)
5643 if (cpupri_init(&rd->cpupri) != 0)
5648 free_cpumask_var(rd->rto_mask);
5650 free_cpumask_var(rd->dlo_mask);
5652 free_cpumask_var(rd->online);
5654 free_cpumask_var(rd->span);
5660 * By default the system creates a single root-domain with all cpus as
5661 * members (mimicking the global state we have today).
5663 struct root_domain def_root_domain;
5665 static void init_defrootdomain(void)
5667 init_rootdomain(&def_root_domain);
5669 atomic_set(&def_root_domain.refcount, 1);
5672 static struct root_domain *alloc_rootdomain(void)
5674 struct root_domain *rd;
5676 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5680 if (init_rootdomain(rd) != 0) {
5688 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5690 struct sched_group *tmp, *first;
5699 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5704 } while (sg != first);
5707 static void free_sched_domain(struct rcu_head *rcu)
5709 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5712 * If its an overlapping domain it has private groups, iterate and
5715 if (sd->flags & SD_OVERLAP) {
5716 free_sched_groups(sd->groups, 1);
5717 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5718 kfree(sd->groups->sgc);
5724 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5726 call_rcu(&sd->rcu, free_sched_domain);
5729 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5731 for (; sd; sd = sd->parent)
5732 destroy_sched_domain(sd, cpu);
5736 * Keep a special pointer to the highest sched_domain that has
5737 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5738 * allows us to avoid some pointer chasing select_idle_sibling().
5740 * Also keep a unique ID per domain (we use the first cpu number in
5741 * the cpumask of the domain), this allows us to quickly tell if
5742 * two cpus are in the same cache domain, see cpus_share_cache().
5744 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5745 DEFINE_PER_CPU(int, sd_llc_size);
5746 DEFINE_PER_CPU(int, sd_llc_id);
5747 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5748 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5749 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5751 static void update_top_cache_domain(int cpu)
5753 struct sched_domain *sd;
5754 struct sched_domain *busy_sd = NULL;
5758 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5760 id = cpumask_first(sched_domain_span(sd));
5761 size = cpumask_weight(sched_domain_span(sd));
5762 busy_sd = sd->parent; /* sd_busy */
5764 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5766 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5767 per_cpu(sd_llc_size, cpu) = size;
5768 per_cpu(sd_llc_id, cpu) = id;
5770 sd = lowest_flag_domain(cpu, SD_NUMA);
5771 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5773 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5774 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5778 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5779 * hold the hotplug lock.
5782 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5784 struct rq *rq = cpu_rq(cpu);
5785 struct sched_domain *tmp;
5787 /* Remove the sched domains which do not contribute to scheduling. */
5788 for (tmp = sd; tmp; ) {
5789 struct sched_domain *parent = tmp->parent;
5793 if (sd_parent_degenerate(tmp, parent)) {
5794 tmp->parent = parent->parent;
5796 parent->parent->child = tmp;
5798 * Transfer SD_PREFER_SIBLING down in case of a
5799 * degenerate parent; the spans match for this
5800 * so the property transfers.
5802 if (parent->flags & SD_PREFER_SIBLING)
5803 tmp->flags |= SD_PREFER_SIBLING;
5804 destroy_sched_domain(parent, cpu);
5809 if (sd && sd_degenerate(sd)) {
5812 destroy_sched_domain(tmp, cpu);
5817 sched_domain_debug(sd, cpu);
5819 rq_attach_root(rq, rd);
5821 rcu_assign_pointer(rq->sd, sd);
5822 destroy_sched_domains(tmp, cpu);
5824 update_top_cache_domain(cpu);
5827 /* cpus with isolated domains */
5828 static cpumask_var_t cpu_isolated_map;
5830 /* Setup the mask of cpus configured for isolated domains */
5831 static int __init isolated_cpu_setup(char *str)
5833 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5834 cpulist_parse(str, cpu_isolated_map);
5838 __setup("isolcpus=", isolated_cpu_setup);
5841 struct sched_domain ** __percpu sd;
5842 struct root_domain *rd;
5853 * Build an iteration mask that can exclude certain CPUs from the upwards
5856 * Asymmetric node setups can result in situations where the domain tree is of
5857 * unequal depth, make sure to skip domains that already cover the entire
5860 * In that case build_sched_domains() will have terminated the iteration early
5861 * and our sibling sd spans will be empty. Domains should always include the
5862 * cpu they're built on, so check that.
5865 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5867 const struct cpumask *span = sched_domain_span(sd);
5868 struct sd_data *sdd = sd->private;
5869 struct sched_domain *sibling;
5872 for_each_cpu(i, span) {
5873 sibling = *per_cpu_ptr(sdd->sd, i);
5874 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5877 cpumask_set_cpu(i, sched_group_mask(sg));
5882 * Return the canonical balance cpu for this group, this is the first cpu
5883 * of this group that's also in the iteration mask.
5885 int group_balance_cpu(struct sched_group *sg)
5887 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5891 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5893 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5894 const struct cpumask *span = sched_domain_span(sd);
5895 struct cpumask *covered = sched_domains_tmpmask;
5896 struct sd_data *sdd = sd->private;
5897 struct sched_domain *sibling;
5900 cpumask_clear(covered);
5902 for_each_cpu(i, span) {
5903 struct cpumask *sg_span;
5905 if (cpumask_test_cpu(i, covered))
5908 sibling = *per_cpu_ptr(sdd->sd, i);
5910 /* See the comment near build_group_mask(). */
5911 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5914 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5915 GFP_KERNEL, cpu_to_node(cpu));
5920 sg_span = sched_group_cpus(sg);
5922 cpumask_copy(sg_span, sched_domain_span(sibling->child));
5924 cpumask_set_cpu(i, sg_span);
5926 cpumask_or(covered, covered, sg_span);
5928 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5929 if (atomic_inc_return(&sg->sgc->ref) == 1)
5930 build_group_mask(sd, sg);
5933 * Initialize sgc->capacity such that even if we mess up the
5934 * domains and no possible iteration will get us here, we won't
5937 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5938 sg->sgc->capacity_orig = sg->sgc->capacity;
5941 * Make sure the first group of this domain contains the
5942 * canonical balance cpu. Otherwise the sched_domain iteration
5943 * breaks. See update_sg_lb_stats().
5945 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5946 group_balance_cpu(sg) == cpu)
5956 sd->groups = groups;
5961 free_sched_groups(first, 0);
5966 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5968 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5969 struct sched_domain *child = sd->child;
5972 cpu = cpumask_first(sched_domain_span(child));
5975 *sg = *per_cpu_ptr(sdd->sg, cpu);
5976 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5977 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5984 * build_sched_groups will build a circular linked list of the groups
5985 * covered by the given span, and will set each group's ->cpumask correctly,
5986 * and ->cpu_capacity to 0.
5988 * Assumes the sched_domain tree is fully constructed
5991 build_sched_groups(struct sched_domain *sd, int cpu)
5993 struct sched_group *first = NULL, *last = NULL;
5994 struct sd_data *sdd = sd->private;
5995 const struct cpumask *span = sched_domain_span(sd);
5996 struct cpumask *covered;
5999 get_group(cpu, sdd, &sd->groups);
6000 atomic_inc(&sd->groups->ref);
6002 if (cpu != cpumask_first(span))
6005 lockdep_assert_held(&sched_domains_mutex);
6006 covered = sched_domains_tmpmask;
6008 cpumask_clear(covered);
6010 for_each_cpu(i, span) {
6011 struct sched_group *sg;
6014 if (cpumask_test_cpu(i, covered))
6017 group = get_group(i, sdd, &sg);
6018 cpumask_setall(sched_group_mask(sg));
6020 for_each_cpu(j, span) {
6021 if (get_group(j, sdd, NULL) != group)
6024 cpumask_set_cpu(j, covered);
6025 cpumask_set_cpu(j, sched_group_cpus(sg));
6040 * Initialize sched groups cpu_capacity.
6042 * cpu_capacity indicates the capacity of sched group, which is used while
6043 * distributing the load between different sched groups in a sched domain.
6044 * Typically cpu_capacity for all the groups in a sched domain will be same
6045 * unless there are asymmetries in the topology. If there are asymmetries,
6046 * group having more cpu_capacity will pickup more load compared to the
6047 * group having less cpu_capacity.
6049 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6051 struct sched_group *sg = sd->groups;
6056 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6058 } while (sg != sd->groups);
6060 if (cpu != group_balance_cpu(sg))
6063 update_group_capacity(sd, cpu);
6064 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6068 * Initializers for schedule domains
6069 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6072 static int default_relax_domain_level = -1;
6073 int sched_domain_level_max;
6075 static int __init setup_relax_domain_level(char *str)
6077 if (kstrtoint(str, 0, &default_relax_domain_level))
6078 pr_warn("Unable to set relax_domain_level\n");
6082 __setup("relax_domain_level=", setup_relax_domain_level);
6084 static void set_domain_attribute(struct sched_domain *sd,
6085 struct sched_domain_attr *attr)
6089 if (!attr || attr->relax_domain_level < 0) {
6090 if (default_relax_domain_level < 0)
6093 request = default_relax_domain_level;
6095 request = attr->relax_domain_level;
6096 if (request < sd->level) {
6097 /* turn off idle balance on this domain */
6098 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6100 /* turn on idle balance on this domain */
6101 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6105 static void __sdt_free(const struct cpumask *cpu_map);
6106 static int __sdt_alloc(const struct cpumask *cpu_map);
6108 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6109 const struct cpumask *cpu_map)
6113 if (!atomic_read(&d->rd->refcount))
6114 free_rootdomain(&d->rd->rcu); /* fall through */
6116 free_percpu(d->sd); /* fall through */
6118 __sdt_free(cpu_map); /* fall through */
6124 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6125 const struct cpumask *cpu_map)
6127 memset(d, 0, sizeof(*d));
6129 if (__sdt_alloc(cpu_map))
6130 return sa_sd_storage;
6131 d->sd = alloc_percpu(struct sched_domain *);
6133 return sa_sd_storage;
6134 d->rd = alloc_rootdomain();
6137 return sa_rootdomain;
6141 * NULL the sd_data elements we've used to build the sched_domain and
6142 * sched_group structure so that the subsequent __free_domain_allocs()
6143 * will not free the data we're using.
6145 static void claim_allocations(int cpu, struct sched_domain *sd)
6147 struct sd_data *sdd = sd->private;
6149 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6150 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6152 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6153 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6155 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6156 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6160 static int sched_domains_numa_levels;
6161 enum numa_topology_type sched_numa_topology_type;
6162 static int *sched_domains_numa_distance;
6163 int sched_max_numa_distance;
6164 static struct cpumask ***sched_domains_numa_masks;
6165 static int sched_domains_curr_level;
6169 * SD_flags allowed in topology descriptions.
6171 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6172 * SD_SHARE_PKG_RESOURCES - describes shared caches
6173 * SD_NUMA - describes NUMA topologies
6174 * SD_SHARE_POWERDOMAIN - describes shared power domain
6177 * SD_ASYM_PACKING - describes SMT quirks
6179 #define TOPOLOGY_SD_FLAGS \
6180 (SD_SHARE_CPUCAPACITY | \
6181 SD_SHARE_PKG_RESOURCES | \
6184 SD_SHARE_POWERDOMAIN)
6186 static struct sched_domain *
6187 sd_init(struct sched_domain_topology_level *tl, int cpu)
6189 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6190 int sd_weight, sd_flags = 0;
6194 * Ugly hack to pass state to sd_numa_mask()...
6196 sched_domains_curr_level = tl->numa_level;
6199 sd_weight = cpumask_weight(tl->mask(cpu));
6202 sd_flags = (*tl->sd_flags)();
6203 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6204 "wrong sd_flags in topology description\n"))
6205 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6207 *sd = (struct sched_domain){
6208 .min_interval = sd_weight,
6209 .max_interval = 2*sd_weight,
6211 .imbalance_pct = 125,
6213 .cache_nice_tries = 0,
6220 .flags = 1*SD_LOAD_BALANCE
6221 | 1*SD_BALANCE_NEWIDLE
6226 | 0*SD_SHARE_CPUCAPACITY
6227 | 0*SD_SHARE_PKG_RESOURCES
6229 | 0*SD_PREFER_SIBLING
6234 .last_balance = jiffies,
6235 .balance_interval = sd_weight,
6237 .max_newidle_lb_cost = 0,
6238 .next_decay_max_lb_cost = jiffies,
6239 #ifdef CONFIG_SCHED_DEBUG
6245 * Convert topological properties into behaviour.
6248 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6249 sd->imbalance_pct = 110;
6250 sd->smt_gain = 1178; /* ~15% */
6252 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6253 sd->imbalance_pct = 117;
6254 sd->cache_nice_tries = 1;
6258 } else if (sd->flags & SD_NUMA) {
6259 sd->cache_nice_tries = 2;
6263 sd->flags |= SD_SERIALIZE;
6264 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6265 sd->flags &= ~(SD_BALANCE_EXEC |
6272 sd->flags |= SD_PREFER_SIBLING;
6273 sd->cache_nice_tries = 1;
6278 sd->private = &tl->data;
6284 * Topology list, bottom-up.
6286 static struct sched_domain_topology_level default_topology[] = {
6287 #ifdef CONFIG_SCHED_SMT
6288 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6290 #ifdef CONFIG_SCHED_MC
6291 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6293 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6297 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6299 #define for_each_sd_topology(tl) \
6300 for (tl = sched_domain_topology; tl->mask; tl++)
6302 void set_sched_topology(struct sched_domain_topology_level *tl)
6304 sched_domain_topology = tl;
6309 static const struct cpumask *sd_numa_mask(int cpu)
6311 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6314 static void sched_numa_warn(const char *str)
6316 static int done = false;
6324 printk(KERN_WARNING "ERROR: %s\n\n", str);
6326 for (i = 0; i < nr_node_ids; i++) {
6327 printk(KERN_WARNING " ");
6328 for (j = 0; j < nr_node_ids; j++)
6329 printk(KERN_CONT "%02d ", node_distance(i,j));
6330 printk(KERN_CONT "\n");
6332 printk(KERN_WARNING "\n");
6335 bool find_numa_distance(int distance)
6339 if (distance == node_distance(0, 0))
6342 for (i = 0; i < sched_domains_numa_levels; i++) {
6343 if (sched_domains_numa_distance[i] == distance)
6351 * A system can have three types of NUMA topology:
6352 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6353 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6354 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6356 * The difference between a glueless mesh topology and a backplane
6357 * topology lies in whether communication between not directly
6358 * connected nodes goes through intermediary nodes (where programs
6359 * could run), or through backplane controllers. This affects
6360 * placement of programs.
6362 * The type of topology can be discerned with the following tests:
6363 * - If the maximum distance between any nodes is 1 hop, the system
6364 * is directly connected.
6365 * - If for two nodes A and B, located N > 1 hops away from each other,
6366 * there is an intermediary node C, which is < N hops away from both
6367 * nodes A and B, the system is a glueless mesh.
6369 static void init_numa_topology_type(void)
6373 n = sched_max_numa_distance;
6376 sched_numa_topology_type = NUMA_DIRECT;
6378 for_each_online_node(a) {
6379 for_each_online_node(b) {
6380 /* Find two nodes furthest removed from each other. */
6381 if (node_distance(a, b) < n)
6384 /* Is there an intermediary node between a and b? */
6385 for_each_online_node(c) {
6386 if (node_distance(a, c) < n &&
6387 node_distance(b, c) < n) {
6388 sched_numa_topology_type =
6394 sched_numa_topology_type = NUMA_BACKPLANE;
6400 static void sched_init_numa(void)
6402 int next_distance, curr_distance = node_distance(0, 0);
6403 struct sched_domain_topology_level *tl;
6407 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6408 if (!sched_domains_numa_distance)
6412 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6413 * unique distances in the node_distance() table.
6415 * Assumes node_distance(0,j) includes all distances in
6416 * node_distance(i,j) in order to avoid cubic time.
6418 next_distance = curr_distance;
6419 for (i = 0; i < nr_node_ids; i++) {
6420 for (j = 0; j < nr_node_ids; j++) {
6421 for (k = 0; k < nr_node_ids; k++) {
6422 int distance = node_distance(i, k);
6424 if (distance > curr_distance &&
6425 (distance < next_distance ||
6426 next_distance == curr_distance))
6427 next_distance = distance;
6430 * While not a strong assumption it would be nice to know
6431 * about cases where if node A is connected to B, B is not
6432 * equally connected to A.
6434 if (sched_debug() && node_distance(k, i) != distance)
6435 sched_numa_warn("Node-distance not symmetric");
6437 if (sched_debug() && i && !find_numa_distance(distance))
6438 sched_numa_warn("Node-0 not representative");
6440 if (next_distance != curr_distance) {
6441 sched_domains_numa_distance[level++] = next_distance;
6442 sched_domains_numa_levels = level;
6443 curr_distance = next_distance;
6448 * In case of sched_debug() we verify the above assumption.
6458 * 'level' contains the number of unique distances, excluding the
6459 * identity distance node_distance(i,i).
6461 * The sched_domains_numa_distance[] array includes the actual distance
6466 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6467 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6468 * the array will contain less then 'level' members. This could be
6469 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6470 * in other functions.
6472 * We reset it to 'level' at the end of this function.
6474 sched_domains_numa_levels = 0;
6476 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6477 if (!sched_domains_numa_masks)
6481 * Now for each level, construct a mask per node which contains all
6482 * cpus of nodes that are that many hops away from us.
6484 for (i = 0; i < level; i++) {
6485 sched_domains_numa_masks[i] =
6486 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6487 if (!sched_domains_numa_masks[i])
6490 for (j = 0; j < nr_node_ids; j++) {
6491 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6495 sched_domains_numa_masks[i][j] = mask;
6497 for (k = 0; k < nr_node_ids; k++) {
6498 if (node_distance(j, k) > sched_domains_numa_distance[i])
6501 cpumask_or(mask, mask, cpumask_of_node(k));
6506 /* Compute default topology size */
6507 for (i = 0; sched_domain_topology[i].mask; i++);
6509 tl = kzalloc((i + level + 1) *
6510 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6515 * Copy the default topology bits..
6517 for (i = 0; sched_domain_topology[i].mask; i++)
6518 tl[i] = sched_domain_topology[i];
6521 * .. and append 'j' levels of NUMA goodness.
6523 for (j = 0; j < level; i++, j++) {
6524 tl[i] = (struct sched_domain_topology_level){
6525 .mask = sd_numa_mask,
6526 .sd_flags = cpu_numa_flags,
6527 .flags = SDTL_OVERLAP,
6533 sched_domain_topology = tl;
6535 sched_domains_numa_levels = level;
6536 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
6538 init_numa_topology_type();
6541 static void sched_domains_numa_masks_set(int cpu)
6544 int node = cpu_to_node(cpu);
6546 for (i = 0; i < sched_domains_numa_levels; i++) {
6547 for (j = 0; j < nr_node_ids; j++) {
6548 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6549 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6554 static void sched_domains_numa_masks_clear(int cpu)
6557 for (i = 0; i < sched_domains_numa_levels; i++) {
6558 for (j = 0; j < nr_node_ids; j++)
6559 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6564 * Update sched_domains_numa_masks[level][node] array when new cpus
6567 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6568 unsigned long action,
6571 int cpu = (long)hcpu;
6573 switch (action & ~CPU_TASKS_FROZEN) {
6575 sched_domains_numa_masks_set(cpu);
6579 sched_domains_numa_masks_clear(cpu);
6589 static inline void sched_init_numa(void)
6593 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6594 unsigned long action,
6599 #endif /* CONFIG_NUMA */
6601 static int __sdt_alloc(const struct cpumask *cpu_map)
6603 struct sched_domain_topology_level *tl;
6606 for_each_sd_topology(tl) {
6607 struct sd_data *sdd = &tl->data;
6609 sdd->sd = alloc_percpu(struct sched_domain *);
6613 sdd->sg = alloc_percpu(struct sched_group *);
6617 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6621 for_each_cpu(j, cpu_map) {
6622 struct sched_domain *sd;
6623 struct sched_group *sg;
6624 struct sched_group_capacity *sgc;
6626 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6627 GFP_KERNEL, cpu_to_node(j));
6631 *per_cpu_ptr(sdd->sd, j) = sd;
6633 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6634 GFP_KERNEL, cpu_to_node(j));
6640 *per_cpu_ptr(sdd->sg, j) = sg;
6642 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6643 GFP_KERNEL, cpu_to_node(j));
6647 *per_cpu_ptr(sdd->sgc, j) = sgc;
6654 static void __sdt_free(const struct cpumask *cpu_map)
6656 struct sched_domain_topology_level *tl;
6659 for_each_sd_topology(tl) {
6660 struct sd_data *sdd = &tl->data;
6662 for_each_cpu(j, cpu_map) {
6663 struct sched_domain *sd;
6666 sd = *per_cpu_ptr(sdd->sd, j);
6667 if (sd && (sd->flags & SD_OVERLAP))
6668 free_sched_groups(sd->groups, 0);
6669 kfree(*per_cpu_ptr(sdd->sd, j));
6673 kfree(*per_cpu_ptr(sdd->sg, j));
6675 kfree(*per_cpu_ptr(sdd->sgc, j));
6677 free_percpu(sdd->sd);
6679 free_percpu(sdd->sg);
6681 free_percpu(sdd->sgc);
6686 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6687 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6688 struct sched_domain *child, int cpu)
6690 struct sched_domain *sd = sd_init(tl, cpu);
6694 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6696 sd->level = child->level + 1;
6697 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6701 if (!cpumask_subset(sched_domain_span(child),
6702 sched_domain_span(sd))) {
6703 pr_err("BUG: arch topology borken\n");
6704 #ifdef CONFIG_SCHED_DEBUG
6705 pr_err(" the %s domain not a subset of the %s domain\n",
6706 child->name, sd->name);
6708 /* Fixup, ensure @sd has at least @child cpus. */
6709 cpumask_or(sched_domain_span(sd),
6710 sched_domain_span(sd),
6711 sched_domain_span(child));
6715 set_domain_attribute(sd, attr);
6721 * Build sched domains for a given set of cpus and attach the sched domains
6722 * to the individual cpus
6724 static int build_sched_domains(const struct cpumask *cpu_map,
6725 struct sched_domain_attr *attr)
6727 enum s_alloc alloc_state;
6728 struct sched_domain *sd;
6730 int i, ret = -ENOMEM;
6732 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6733 if (alloc_state != sa_rootdomain)
6736 /* Set up domains for cpus specified by the cpu_map. */
6737 for_each_cpu(i, cpu_map) {
6738 struct sched_domain_topology_level *tl;
6741 for_each_sd_topology(tl) {
6742 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6743 if (tl == sched_domain_topology)
6744 *per_cpu_ptr(d.sd, i) = sd;
6745 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6746 sd->flags |= SD_OVERLAP;
6747 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6752 /* Build the groups for the domains */
6753 for_each_cpu(i, cpu_map) {
6754 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6755 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6756 if (sd->flags & SD_OVERLAP) {
6757 if (build_overlap_sched_groups(sd, i))
6760 if (build_sched_groups(sd, i))
6766 /* Calculate CPU capacity for physical packages and nodes */
6767 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6768 if (!cpumask_test_cpu(i, cpu_map))
6771 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6772 claim_allocations(i, sd);
6773 init_sched_groups_capacity(i, sd);
6777 /* Attach the domains */
6779 for_each_cpu(i, cpu_map) {
6780 sd = *per_cpu_ptr(d.sd, i);
6781 cpu_attach_domain(sd, d.rd, i);
6787 __free_domain_allocs(&d, alloc_state, cpu_map);
6791 static cpumask_var_t *doms_cur; /* current sched domains */
6792 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6793 static struct sched_domain_attr *dattr_cur;
6794 /* attribues of custom domains in 'doms_cur' */
6797 * Special case: If a kmalloc of a doms_cur partition (array of
6798 * cpumask) fails, then fallback to a single sched domain,
6799 * as determined by the single cpumask fallback_doms.
6801 static cpumask_var_t fallback_doms;
6804 * arch_update_cpu_topology lets virtualized architectures update the
6805 * cpu core maps. It is supposed to return 1 if the topology changed
6806 * or 0 if it stayed the same.
6808 int __weak arch_update_cpu_topology(void)
6813 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6816 cpumask_var_t *doms;
6818 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6821 for (i = 0; i < ndoms; i++) {
6822 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6823 free_sched_domains(doms, i);
6830 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6833 for (i = 0; i < ndoms; i++)
6834 free_cpumask_var(doms[i]);
6839 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6840 * For now this just excludes isolated cpus, but could be used to
6841 * exclude other special cases in the future.
6843 static int init_sched_domains(const struct cpumask *cpu_map)
6847 arch_update_cpu_topology();
6849 doms_cur = alloc_sched_domains(ndoms_cur);
6851 doms_cur = &fallback_doms;
6852 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6853 err = build_sched_domains(doms_cur[0], NULL);
6854 register_sched_domain_sysctl();
6860 * Detach sched domains from a group of cpus specified in cpu_map
6861 * These cpus will now be attached to the NULL domain
6863 static void detach_destroy_domains(const struct cpumask *cpu_map)
6868 for_each_cpu(i, cpu_map)
6869 cpu_attach_domain(NULL, &def_root_domain, i);
6873 /* handle null as "default" */
6874 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6875 struct sched_domain_attr *new, int idx_new)
6877 struct sched_domain_attr tmp;
6884 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6885 new ? (new + idx_new) : &tmp,
6886 sizeof(struct sched_domain_attr));
6890 * Partition sched domains as specified by the 'ndoms_new'
6891 * cpumasks in the array doms_new[] of cpumasks. This compares
6892 * doms_new[] to the current sched domain partitioning, doms_cur[].
6893 * It destroys each deleted domain and builds each new domain.
6895 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6896 * The masks don't intersect (don't overlap.) We should setup one
6897 * sched domain for each mask. CPUs not in any of the cpumasks will
6898 * not be load balanced. If the same cpumask appears both in the
6899 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6902 * The passed in 'doms_new' should be allocated using
6903 * alloc_sched_domains. This routine takes ownership of it and will
6904 * free_sched_domains it when done with it. If the caller failed the
6905 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6906 * and partition_sched_domains() will fallback to the single partition
6907 * 'fallback_doms', it also forces the domains to be rebuilt.
6909 * If doms_new == NULL it will be replaced with cpu_online_mask.
6910 * ndoms_new == 0 is a special case for destroying existing domains,
6911 * and it will not create the default domain.
6913 * Call with hotplug lock held
6915 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6916 struct sched_domain_attr *dattr_new)
6921 mutex_lock(&sched_domains_mutex);
6923 /* always unregister in case we don't destroy any domains */
6924 unregister_sched_domain_sysctl();
6926 /* Let architecture update cpu core mappings. */
6927 new_topology = arch_update_cpu_topology();
6929 n = doms_new ? ndoms_new : 0;
6931 /* Destroy deleted domains */
6932 for (i = 0; i < ndoms_cur; i++) {
6933 for (j = 0; j < n && !new_topology; j++) {
6934 if (cpumask_equal(doms_cur[i], doms_new[j])
6935 && dattrs_equal(dattr_cur, i, dattr_new, j))
6938 /* no match - a current sched domain not in new doms_new[] */
6939 detach_destroy_domains(doms_cur[i]);
6945 if (doms_new == NULL) {
6947 doms_new = &fallback_doms;
6948 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6949 WARN_ON_ONCE(dattr_new);
6952 /* Build new domains */
6953 for (i = 0; i < ndoms_new; i++) {
6954 for (j = 0; j < n && !new_topology; j++) {
6955 if (cpumask_equal(doms_new[i], doms_cur[j])
6956 && dattrs_equal(dattr_new, i, dattr_cur, j))
6959 /* no match - add a new doms_new */
6960 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6965 /* Remember the new sched domains */
6966 if (doms_cur != &fallback_doms)
6967 free_sched_domains(doms_cur, ndoms_cur);
6968 kfree(dattr_cur); /* kfree(NULL) is safe */
6969 doms_cur = doms_new;
6970 dattr_cur = dattr_new;
6971 ndoms_cur = ndoms_new;
6973 register_sched_domain_sysctl();
6975 mutex_unlock(&sched_domains_mutex);
6978 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6981 * Update cpusets according to cpu_active mask. If cpusets are
6982 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6983 * around partition_sched_domains().
6985 * If we come here as part of a suspend/resume, don't touch cpusets because we
6986 * want to restore it back to its original state upon resume anyway.
6988 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6992 case CPU_ONLINE_FROZEN:
6993 case CPU_DOWN_FAILED_FROZEN:
6996 * num_cpus_frozen tracks how many CPUs are involved in suspend
6997 * resume sequence. As long as this is not the last online
6998 * operation in the resume sequence, just build a single sched
6999 * domain, ignoring cpusets.
7002 if (likely(num_cpus_frozen)) {
7003 partition_sched_domains(1, NULL, NULL);
7008 * This is the last CPU online operation. So fall through and
7009 * restore the original sched domains by considering the
7010 * cpuset configurations.
7014 case CPU_DOWN_FAILED:
7015 cpuset_update_active_cpus(true);
7023 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7027 case CPU_DOWN_PREPARE:
7028 cpuset_update_active_cpus(false);
7030 case CPU_DOWN_PREPARE_FROZEN:
7032 partition_sched_domains(1, NULL, NULL);
7040 void __init sched_init_smp(void)
7042 cpumask_var_t non_isolated_cpus;
7044 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7045 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7050 * There's no userspace yet to cause hotplug operations; hence all the
7051 * cpu masks are stable and all blatant races in the below code cannot
7054 mutex_lock(&sched_domains_mutex);
7055 init_sched_domains(cpu_active_mask);
7056 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7057 if (cpumask_empty(non_isolated_cpus))
7058 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7059 mutex_unlock(&sched_domains_mutex);
7061 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7062 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7063 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7067 /* Move init over to a non-isolated CPU */
7068 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7070 sched_init_granularity();
7071 free_cpumask_var(non_isolated_cpus);
7073 init_sched_rt_class();
7074 init_sched_dl_class();
7077 void __init sched_init_smp(void)
7079 sched_init_granularity();
7081 #endif /* CONFIG_SMP */
7083 const_debug unsigned int sysctl_timer_migration = 1;
7085 int in_sched_functions(unsigned long addr)
7087 return in_lock_functions(addr) ||
7088 (addr >= (unsigned long)__sched_text_start
7089 && addr < (unsigned long)__sched_text_end);
7092 #ifdef CONFIG_CGROUP_SCHED
7094 * Default task group.
7095 * Every task in system belongs to this group at bootup.
7097 struct task_group root_task_group;
7098 LIST_HEAD(task_groups);
7101 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7103 void __init sched_init(void)
7106 unsigned long alloc_size = 0, ptr;
7108 #ifdef CONFIG_FAIR_GROUP_SCHED
7109 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7111 #ifdef CONFIG_RT_GROUP_SCHED
7112 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7114 #ifdef CONFIG_CPUMASK_OFFSTACK
7115 alloc_size += num_possible_cpus() * cpumask_size();
7118 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7120 #ifdef CONFIG_FAIR_GROUP_SCHED
7121 root_task_group.se = (struct sched_entity **)ptr;
7122 ptr += nr_cpu_ids * sizeof(void **);
7124 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7125 ptr += nr_cpu_ids * sizeof(void **);
7127 #endif /* CONFIG_FAIR_GROUP_SCHED */
7128 #ifdef CONFIG_RT_GROUP_SCHED
7129 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7130 ptr += nr_cpu_ids * sizeof(void **);
7132 root_task_group.rt_rq = (struct rt_rq **)ptr;
7133 ptr += nr_cpu_ids * sizeof(void **);
7135 #endif /* CONFIG_RT_GROUP_SCHED */
7136 #ifdef CONFIG_CPUMASK_OFFSTACK
7137 for_each_possible_cpu(i) {
7138 per_cpu(load_balance_mask, i) = (void *)ptr;
7139 ptr += cpumask_size();
7141 #endif /* CONFIG_CPUMASK_OFFSTACK */
7144 init_rt_bandwidth(&def_rt_bandwidth,
7145 global_rt_period(), global_rt_runtime());
7146 init_dl_bandwidth(&def_dl_bandwidth,
7147 global_rt_period(), global_rt_runtime());
7150 init_defrootdomain();
7153 #ifdef CONFIG_RT_GROUP_SCHED
7154 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7155 global_rt_period(), global_rt_runtime());
7156 #endif /* CONFIG_RT_GROUP_SCHED */
7158 #ifdef CONFIG_CGROUP_SCHED
7159 list_add(&root_task_group.list, &task_groups);
7160 INIT_LIST_HEAD(&root_task_group.children);
7161 INIT_LIST_HEAD(&root_task_group.siblings);
7162 autogroup_init(&init_task);
7164 #endif /* CONFIG_CGROUP_SCHED */
7166 for_each_possible_cpu(i) {
7170 raw_spin_lock_init(&rq->lock);
7172 rq->calc_load_active = 0;
7173 rq->calc_load_update = jiffies + LOAD_FREQ;
7174 init_cfs_rq(&rq->cfs);
7175 init_rt_rq(&rq->rt, rq);
7176 init_dl_rq(&rq->dl, rq);
7177 #ifdef CONFIG_FAIR_GROUP_SCHED
7178 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7179 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7181 * How much cpu bandwidth does root_task_group get?
7183 * In case of task-groups formed thr' the cgroup filesystem, it
7184 * gets 100% of the cpu resources in the system. This overall
7185 * system cpu resource is divided among the tasks of
7186 * root_task_group and its child task-groups in a fair manner,
7187 * based on each entity's (task or task-group's) weight
7188 * (se->load.weight).
7190 * In other words, if root_task_group has 10 tasks of weight
7191 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7192 * then A0's share of the cpu resource is:
7194 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7196 * We achieve this by letting root_task_group's tasks sit
7197 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7199 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7200 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7201 #endif /* CONFIG_FAIR_GROUP_SCHED */
7203 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7204 #ifdef CONFIG_RT_GROUP_SCHED
7205 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7208 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7209 rq->cpu_load[j] = 0;
7211 rq->last_load_update_tick = jiffies;
7216 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
7217 rq->post_schedule = 0;
7218 rq->active_balance = 0;
7219 rq->next_balance = jiffies;
7224 rq->avg_idle = 2*sysctl_sched_migration_cost;
7225 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7227 INIT_LIST_HEAD(&rq->cfs_tasks);
7229 rq_attach_root(rq, &def_root_domain);
7230 #ifdef CONFIG_NO_HZ_COMMON
7233 #ifdef CONFIG_NO_HZ_FULL
7234 rq->last_sched_tick = 0;
7238 atomic_set(&rq->nr_iowait, 0);
7241 set_load_weight(&init_task);
7243 #ifdef CONFIG_PREEMPT_NOTIFIERS
7244 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7248 * The boot idle thread does lazy MMU switching as well:
7250 atomic_inc(&init_mm.mm_count);
7251 enter_lazy_tlb(&init_mm, current);
7254 * Make us the idle thread. Technically, schedule() should not be
7255 * called from this thread, however somewhere below it might be,
7256 * but because we are the idle thread, we just pick up running again
7257 * when this runqueue becomes "idle".
7259 init_idle(current, smp_processor_id());
7261 calc_load_update = jiffies + LOAD_FREQ;
7264 * During early bootup we pretend to be a normal task:
7266 current->sched_class = &fair_sched_class;
7269 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7270 /* May be allocated at isolcpus cmdline parse time */
7271 if (cpu_isolated_map == NULL)
7272 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7273 idle_thread_set_boot_cpu();
7274 set_cpu_rq_start_time();
7276 init_sched_fair_class();
7278 scheduler_running = 1;
7281 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7282 static inline int preempt_count_equals(int preempt_offset)
7284 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7286 return (nested == preempt_offset);
7289 void __might_sleep(const char *file, int line, int preempt_offset)
7292 * Blocking primitives will set (and therefore destroy) current->state,
7293 * since we will exit with TASK_RUNNING make sure we enter with it,
7294 * otherwise we will destroy state.
7296 if (WARN_ONCE(current->state != TASK_RUNNING,
7297 "do not call blocking ops when !TASK_RUNNING; "
7298 "state=%lx set at [<%p>] %pS\n",
7300 (void *)current->task_state_change,
7301 (void *)current->task_state_change))
7302 __set_current_state(TASK_RUNNING);
7304 ___might_sleep(file, line, preempt_offset);
7306 EXPORT_SYMBOL(__might_sleep);
7308 void ___might_sleep(const char *file, int line, int preempt_offset)
7310 static unsigned long prev_jiffy; /* ratelimiting */
7312 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7313 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7314 !is_idle_task(current)) ||
7315 system_state != SYSTEM_RUNNING || oops_in_progress)
7317 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7319 prev_jiffy = jiffies;
7322 "BUG: sleeping function called from invalid context at %s:%d\n",
7325 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7326 in_atomic(), irqs_disabled(),
7327 current->pid, current->comm);
7329 debug_show_held_locks(current);
7330 if (irqs_disabled())
7331 print_irqtrace_events(current);
7332 #ifdef CONFIG_DEBUG_PREEMPT
7333 if (!preempt_count_equals(preempt_offset)) {
7334 pr_err("Preemption disabled at:");
7335 print_ip_sym(current->preempt_disable_ip);
7341 EXPORT_SYMBOL(___might_sleep);
7344 #ifdef CONFIG_MAGIC_SYSRQ
7345 static void normalize_task(struct rq *rq, struct task_struct *p)
7347 const struct sched_class *prev_class = p->sched_class;
7348 struct sched_attr attr = {
7349 .sched_policy = SCHED_NORMAL,
7351 int old_prio = p->prio;
7354 queued = task_on_rq_queued(p);
7356 dequeue_task(rq, p, 0);
7357 __setscheduler(rq, p, &attr);
7359 enqueue_task(rq, p, 0);
7363 check_class_changed(rq, p, prev_class, old_prio);
7366 void normalize_rt_tasks(void)
7368 struct task_struct *g, *p;
7369 unsigned long flags;
7372 read_lock(&tasklist_lock);
7373 for_each_process_thread(g, p) {
7375 * Only normalize user tasks:
7377 if (p->flags & PF_KTHREAD)
7380 p->se.exec_start = 0;
7381 #ifdef CONFIG_SCHEDSTATS
7382 p->se.statistics.wait_start = 0;
7383 p->se.statistics.sleep_start = 0;
7384 p->se.statistics.block_start = 0;
7387 if (!dl_task(p) && !rt_task(p)) {
7389 * Renice negative nice level userspace
7392 if (task_nice(p) < 0)
7393 set_user_nice(p, 0);
7397 rq = task_rq_lock(p, &flags);
7398 normalize_task(rq, p);
7399 task_rq_unlock(rq, p, &flags);
7401 read_unlock(&tasklist_lock);
7404 #endif /* CONFIG_MAGIC_SYSRQ */
7406 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7408 * These functions are only useful for the IA64 MCA handling, or kdb.
7410 * They can only be called when the whole system has been
7411 * stopped - every CPU needs to be quiescent, and no scheduling
7412 * activity can take place. Using them for anything else would
7413 * be a serious bug, and as a result, they aren't even visible
7414 * under any other configuration.
7418 * curr_task - return the current task for a given cpu.
7419 * @cpu: the processor in question.
7421 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7423 * Return: The current task for @cpu.
7425 struct task_struct *curr_task(int cpu)
7427 return cpu_curr(cpu);
7430 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7434 * set_curr_task - set the current task for a given cpu.
7435 * @cpu: the processor in question.
7436 * @p: the task pointer to set.
7438 * Description: This function must only be used when non-maskable interrupts
7439 * are serviced on a separate stack. It allows the architecture to switch the
7440 * notion of the current task on a cpu in a non-blocking manner. This function
7441 * must be called with all CPU's synchronized, and interrupts disabled, the
7442 * and caller must save the original value of the current task (see
7443 * curr_task() above) and restore that value before reenabling interrupts and
7444 * re-starting the system.
7446 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7448 void set_curr_task(int cpu, struct task_struct *p)
7455 #ifdef CONFIG_CGROUP_SCHED
7456 /* task_group_lock serializes the addition/removal of task groups */
7457 static DEFINE_SPINLOCK(task_group_lock);
7459 static void free_sched_group(struct task_group *tg)
7461 free_fair_sched_group(tg);
7462 free_rt_sched_group(tg);
7467 /* allocate runqueue etc for a new task group */
7468 struct task_group *sched_create_group(struct task_group *parent)
7470 struct task_group *tg;
7472 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7474 return ERR_PTR(-ENOMEM);
7476 if (!alloc_fair_sched_group(tg, parent))
7479 if (!alloc_rt_sched_group(tg, parent))
7485 free_sched_group(tg);
7486 return ERR_PTR(-ENOMEM);
7489 void sched_online_group(struct task_group *tg, struct task_group *parent)
7491 unsigned long flags;
7493 spin_lock_irqsave(&task_group_lock, flags);
7494 list_add_rcu(&tg->list, &task_groups);
7496 WARN_ON(!parent); /* root should already exist */
7498 tg->parent = parent;
7499 INIT_LIST_HEAD(&tg->children);
7500 list_add_rcu(&tg->siblings, &parent->children);
7501 spin_unlock_irqrestore(&task_group_lock, flags);
7504 /* rcu callback to free various structures associated with a task group */
7505 static void free_sched_group_rcu(struct rcu_head *rhp)
7507 /* now it should be safe to free those cfs_rqs */
7508 free_sched_group(container_of(rhp, struct task_group, rcu));
7511 /* Destroy runqueue etc associated with a task group */
7512 void sched_destroy_group(struct task_group *tg)
7514 /* wait for possible concurrent references to cfs_rqs complete */
7515 call_rcu(&tg->rcu, free_sched_group_rcu);
7518 void sched_offline_group(struct task_group *tg)
7520 unsigned long flags;
7523 /* end participation in shares distribution */
7524 for_each_possible_cpu(i)
7525 unregister_fair_sched_group(tg, i);
7527 spin_lock_irqsave(&task_group_lock, flags);
7528 list_del_rcu(&tg->list);
7529 list_del_rcu(&tg->siblings);
7530 spin_unlock_irqrestore(&task_group_lock, flags);
7533 /* change task's runqueue when it moves between groups.
7534 * The caller of this function should have put the task in its new group
7535 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7536 * reflect its new group.
7538 void sched_move_task(struct task_struct *tsk)
7540 struct task_group *tg;
7541 int queued, running;
7542 unsigned long flags;
7545 rq = task_rq_lock(tsk, &flags);
7547 running = task_current(rq, tsk);
7548 queued = task_on_rq_queued(tsk);
7551 dequeue_task(rq, tsk, 0);
7552 if (unlikely(running))
7553 put_prev_task(rq, tsk);
7556 * All callers are synchronized by task_rq_lock(); we do not use RCU
7557 * which is pointless here. Thus, we pass "true" to task_css_check()
7558 * to prevent lockdep warnings.
7560 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
7561 struct task_group, css);
7562 tg = autogroup_task_group(tsk, tg);
7563 tsk->sched_task_group = tg;
7565 #ifdef CONFIG_FAIR_GROUP_SCHED
7566 if (tsk->sched_class->task_move_group)
7567 tsk->sched_class->task_move_group(tsk, queued);
7570 set_task_rq(tsk, task_cpu(tsk));
7572 if (unlikely(running))
7573 tsk->sched_class->set_curr_task(rq);
7575 enqueue_task(rq, tsk, 0);
7577 task_rq_unlock(rq, tsk, &flags);
7579 #endif /* CONFIG_CGROUP_SCHED */
7581 #ifdef CONFIG_RT_GROUP_SCHED
7583 * Ensure that the real time constraints are schedulable.
7585 static DEFINE_MUTEX(rt_constraints_mutex);
7587 /* Must be called with tasklist_lock held */
7588 static inline int tg_has_rt_tasks(struct task_group *tg)
7590 struct task_struct *g, *p;
7592 for_each_process_thread(g, p) {
7593 if (rt_task(p) && task_group(p) == tg)
7600 struct rt_schedulable_data {
7601 struct task_group *tg;
7606 static int tg_rt_schedulable(struct task_group *tg, void *data)
7608 struct rt_schedulable_data *d = data;
7609 struct task_group *child;
7610 unsigned long total, sum = 0;
7611 u64 period, runtime;
7613 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7614 runtime = tg->rt_bandwidth.rt_runtime;
7617 period = d->rt_period;
7618 runtime = d->rt_runtime;
7622 * Cannot have more runtime than the period.
7624 if (runtime > period && runtime != RUNTIME_INF)
7628 * Ensure we don't starve existing RT tasks.
7630 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7633 total = to_ratio(period, runtime);
7636 * Nobody can have more than the global setting allows.
7638 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7642 * The sum of our children's runtime should not exceed our own.
7644 list_for_each_entry_rcu(child, &tg->children, siblings) {
7645 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7646 runtime = child->rt_bandwidth.rt_runtime;
7648 if (child == d->tg) {
7649 period = d->rt_period;
7650 runtime = d->rt_runtime;
7653 sum += to_ratio(period, runtime);
7662 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7666 struct rt_schedulable_data data = {
7668 .rt_period = period,
7669 .rt_runtime = runtime,
7673 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7679 static int tg_set_rt_bandwidth(struct task_group *tg,
7680 u64 rt_period, u64 rt_runtime)
7684 mutex_lock(&rt_constraints_mutex);
7685 read_lock(&tasklist_lock);
7686 err = __rt_schedulable(tg, rt_period, rt_runtime);
7690 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7691 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7692 tg->rt_bandwidth.rt_runtime = rt_runtime;
7694 for_each_possible_cpu(i) {
7695 struct rt_rq *rt_rq = tg->rt_rq[i];
7697 raw_spin_lock(&rt_rq->rt_runtime_lock);
7698 rt_rq->rt_runtime = rt_runtime;
7699 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7701 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7703 read_unlock(&tasklist_lock);
7704 mutex_unlock(&rt_constraints_mutex);
7709 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7711 u64 rt_runtime, rt_period;
7713 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7714 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7715 if (rt_runtime_us < 0)
7716 rt_runtime = RUNTIME_INF;
7718 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7721 static long sched_group_rt_runtime(struct task_group *tg)
7725 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7728 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7729 do_div(rt_runtime_us, NSEC_PER_USEC);
7730 return rt_runtime_us;
7733 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7735 u64 rt_runtime, rt_period;
7737 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7738 rt_runtime = tg->rt_bandwidth.rt_runtime;
7743 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7746 static long sched_group_rt_period(struct task_group *tg)
7750 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7751 do_div(rt_period_us, NSEC_PER_USEC);
7752 return rt_period_us;
7754 #endif /* CONFIG_RT_GROUP_SCHED */
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 static int sched_rt_global_constraints(void)
7761 mutex_lock(&rt_constraints_mutex);
7762 read_lock(&tasklist_lock);
7763 ret = __rt_schedulable(NULL, 0, 0);
7764 read_unlock(&tasklist_lock);
7765 mutex_unlock(&rt_constraints_mutex);
7770 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7772 /* Don't accept realtime tasks when there is no way for them to run */
7773 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7779 #else /* !CONFIG_RT_GROUP_SCHED */
7780 static int sched_rt_global_constraints(void)
7782 unsigned long flags;
7785 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7786 for_each_possible_cpu(i) {
7787 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7789 raw_spin_lock(&rt_rq->rt_runtime_lock);
7790 rt_rq->rt_runtime = global_rt_runtime();
7791 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7793 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7797 #endif /* CONFIG_RT_GROUP_SCHED */
7799 static int sched_dl_global_constraints(void)
7801 u64 runtime = global_rt_runtime();
7802 u64 period = global_rt_period();
7803 u64 new_bw = to_ratio(period, runtime);
7806 unsigned long flags;
7809 * Here we want to check the bandwidth not being set to some
7810 * value smaller than the currently allocated bandwidth in
7811 * any of the root_domains.
7813 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7814 * cycling on root_domains... Discussion on different/better
7815 * solutions is welcome!
7817 for_each_possible_cpu(cpu) {
7818 rcu_read_lock_sched();
7819 dl_b = dl_bw_of(cpu);
7821 raw_spin_lock_irqsave(&dl_b->lock, flags);
7822 if (new_bw < dl_b->total_bw)
7824 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7826 rcu_read_unlock_sched();
7835 static void sched_dl_do_global(void)
7840 unsigned long flags;
7842 def_dl_bandwidth.dl_period = global_rt_period();
7843 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7845 if (global_rt_runtime() != RUNTIME_INF)
7846 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7849 * FIXME: As above...
7851 for_each_possible_cpu(cpu) {
7852 rcu_read_lock_sched();
7853 dl_b = dl_bw_of(cpu);
7855 raw_spin_lock_irqsave(&dl_b->lock, flags);
7857 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7859 rcu_read_unlock_sched();
7863 static int sched_rt_global_validate(void)
7865 if (sysctl_sched_rt_period <= 0)
7868 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7869 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7875 static void sched_rt_do_global(void)
7877 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7878 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7881 int sched_rt_handler(struct ctl_table *table, int write,
7882 void __user *buffer, size_t *lenp,
7885 int old_period, old_runtime;
7886 static DEFINE_MUTEX(mutex);
7890 old_period = sysctl_sched_rt_period;
7891 old_runtime = sysctl_sched_rt_runtime;
7893 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7895 if (!ret && write) {
7896 ret = sched_rt_global_validate();
7900 ret = sched_rt_global_constraints();
7904 ret = sched_dl_global_constraints();
7908 sched_rt_do_global();
7909 sched_dl_do_global();
7913 sysctl_sched_rt_period = old_period;
7914 sysctl_sched_rt_runtime = old_runtime;
7916 mutex_unlock(&mutex);
7921 int sched_rr_handler(struct ctl_table *table, int write,
7922 void __user *buffer, size_t *lenp,
7926 static DEFINE_MUTEX(mutex);
7929 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7930 /* make sure that internally we keep jiffies */
7931 /* also, writing zero resets timeslice to default */
7932 if (!ret && write) {
7933 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7934 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7936 mutex_unlock(&mutex);
7940 #ifdef CONFIG_CGROUP_SCHED
7942 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7944 return css ? container_of(css, struct task_group, css) : NULL;
7947 static struct cgroup_subsys_state *
7948 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7950 struct task_group *parent = css_tg(parent_css);
7951 struct task_group *tg;
7954 /* This is early initialization for the top cgroup */
7955 return &root_task_group.css;
7958 tg = sched_create_group(parent);
7960 return ERR_PTR(-ENOMEM);
7965 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7967 struct task_group *tg = css_tg(css);
7968 struct task_group *parent = css_tg(css->parent);
7971 sched_online_group(tg, parent);
7975 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7977 struct task_group *tg = css_tg(css);
7979 sched_destroy_group(tg);
7982 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7984 struct task_group *tg = css_tg(css);
7986 sched_offline_group(tg);
7989 static void cpu_cgroup_fork(struct task_struct *task)
7991 sched_move_task(task);
7994 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7995 struct cgroup_taskset *tset)
7997 struct task_struct *task;
7999 cgroup_taskset_for_each(task, tset) {
8000 #ifdef CONFIG_RT_GROUP_SCHED
8001 if (!sched_rt_can_attach(css_tg(css), task))
8004 /* We don't support RT-tasks being in separate groups */
8005 if (task->sched_class != &fair_sched_class)
8012 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
8013 struct cgroup_taskset *tset)
8015 struct task_struct *task;
8017 cgroup_taskset_for_each(task, tset)
8018 sched_move_task(task);
8021 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
8022 struct cgroup_subsys_state *old_css,
8023 struct task_struct *task)
8026 * cgroup_exit() is called in the copy_process() failure path.
8027 * Ignore this case since the task hasn't ran yet, this avoids
8028 * trying to poke a half freed task state from generic code.
8030 if (!(task->flags & PF_EXITING))
8033 sched_move_task(task);
8036 #ifdef CONFIG_FAIR_GROUP_SCHED
8037 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8038 struct cftype *cftype, u64 shareval)
8040 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8043 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8046 struct task_group *tg = css_tg(css);
8048 return (u64) scale_load_down(tg->shares);
8051 #ifdef CONFIG_CFS_BANDWIDTH
8052 static DEFINE_MUTEX(cfs_constraints_mutex);
8054 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8055 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8057 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8059 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8061 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8062 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8064 if (tg == &root_task_group)
8068 * Ensure we have at some amount of bandwidth every period. This is
8069 * to prevent reaching a state of large arrears when throttled via
8070 * entity_tick() resulting in prolonged exit starvation.
8072 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8076 * Likewise, bound things on the otherside by preventing insane quota
8077 * periods. This also allows us to normalize in computing quota
8080 if (period > max_cfs_quota_period)
8084 * Prevent race between setting of cfs_rq->runtime_enabled and
8085 * unthrottle_offline_cfs_rqs().
8088 mutex_lock(&cfs_constraints_mutex);
8089 ret = __cfs_schedulable(tg, period, quota);
8093 runtime_enabled = quota != RUNTIME_INF;
8094 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8096 * If we need to toggle cfs_bandwidth_used, off->on must occur
8097 * before making related changes, and on->off must occur afterwards
8099 if (runtime_enabled && !runtime_was_enabled)
8100 cfs_bandwidth_usage_inc();
8101 raw_spin_lock_irq(&cfs_b->lock);
8102 cfs_b->period = ns_to_ktime(period);
8103 cfs_b->quota = quota;
8105 __refill_cfs_bandwidth_runtime(cfs_b);
8106 /* restart the period timer (if active) to handle new period expiry */
8107 if (runtime_enabled && cfs_b->timer_active) {
8108 /* force a reprogram */
8109 __start_cfs_bandwidth(cfs_b, true);
8111 raw_spin_unlock_irq(&cfs_b->lock);
8113 for_each_online_cpu(i) {
8114 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8115 struct rq *rq = cfs_rq->rq;
8117 raw_spin_lock_irq(&rq->lock);
8118 cfs_rq->runtime_enabled = runtime_enabled;
8119 cfs_rq->runtime_remaining = 0;
8121 if (cfs_rq->throttled)
8122 unthrottle_cfs_rq(cfs_rq);
8123 raw_spin_unlock_irq(&rq->lock);
8125 if (runtime_was_enabled && !runtime_enabled)
8126 cfs_bandwidth_usage_dec();
8128 mutex_unlock(&cfs_constraints_mutex);
8134 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8138 period = ktime_to_ns(tg->cfs_bandwidth.period);
8139 if (cfs_quota_us < 0)
8140 quota = RUNTIME_INF;
8142 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8144 return tg_set_cfs_bandwidth(tg, period, quota);
8147 long tg_get_cfs_quota(struct task_group *tg)
8151 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8154 quota_us = tg->cfs_bandwidth.quota;
8155 do_div(quota_us, NSEC_PER_USEC);
8160 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8164 period = (u64)cfs_period_us * NSEC_PER_USEC;
8165 quota = tg->cfs_bandwidth.quota;
8167 return tg_set_cfs_bandwidth(tg, period, quota);
8170 long tg_get_cfs_period(struct task_group *tg)
8174 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8175 do_div(cfs_period_us, NSEC_PER_USEC);
8177 return cfs_period_us;
8180 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8183 return tg_get_cfs_quota(css_tg(css));
8186 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8187 struct cftype *cftype, s64 cfs_quota_us)
8189 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8192 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8195 return tg_get_cfs_period(css_tg(css));
8198 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8199 struct cftype *cftype, u64 cfs_period_us)
8201 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8204 struct cfs_schedulable_data {
8205 struct task_group *tg;
8210 * normalize group quota/period to be quota/max_period
8211 * note: units are usecs
8213 static u64 normalize_cfs_quota(struct task_group *tg,
8214 struct cfs_schedulable_data *d)
8222 period = tg_get_cfs_period(tg);
8223 quota = tg_get_cfs_quota(tg);
8226 /* note: these should typically be equivalent */
8227 if (quota == RUNTIME_INF || quota == -1)
8230 return to_ratio(period, quota);
8233 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8235 struct cfs_schedulable_data *d = data;
8236 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8237 s64 quota = 0, parent_quota = -1;
8240 quota = RUNTIME_INF;
8242 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8244 quota = normalize_cfs_quota(tg, d);
8245 parent_quota = parent_b->hierarchical_quota;
8248 * ensure max(child_quota) <= parent_quota, inherit when no
8251 if (quota == RUNTIME_INF)
8252 quota = parent_quota;
8253 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8256 cfs_b->hierarchical_quota = quota;
8261 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8264 struct cfs_schedulable_data data = {
8270 if (quota != RUNTIME_INF) {
8271 do_div(data.period, NSEC_PER_USEC);
8272 do_div(data.quota, NSEC_PER_USEC);
8276 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8282 static int cpu_stats_show(struct seq_file *sf, void *v)
8284 struct task_group *tg = css_tg(seq_css(sf));
8285 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8287 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8288 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8289 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8293 #endif /* CONFIG_CFS_BANDWIDTH */
8294 #endif /* CONFIG_FAIR_GROUP_SCHED */
8296 #ifdef CONFIG_RT_GROUP_SCHED
8297 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8298 struct cftype *cft, s64 val)
8300 return sched_group_set_rt_runtime(css_tg(css), val);
8303 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8306 return sched_group_rt_runtime(css_tg(css));
8309 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8310 struct cftype *cftype, u64 rt_period_us)
8312 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8315 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8318 return sched_group_rt_period(css_tg(css));
8320 #endif /* CONFIG_RT_GROUP_SCHED */
8322 static struct cftype cpu_files[] = {
8323 #ifdef CONFIG_FAIR_GROUP_SCHED
8326 .read_u64 = cpu_shares_read_u64,
8327 .write_u64 = cpu_shares_write_u64,
8330 #ifdef CONFIG_CFS_BANDWIDTH
8332 .name = "cfs_quota_us",
8333 .read_s64 = cpu_cfs_quota_read_s64,
8334 .write_s64 = cpu_cfs_quota_write_s64,
8337 .name = "cfs_period_us",
8338 .read_u64 = cpu_cfs_period_read_u64,
8339 .write_u64 = cpu_cfs_period_write_u64,
8343 .seq_show = cpu_stats_show,
8346 #ifdef CONFIG_RT_GROUP_SCHED
8348 .name = "rt_runtime_us",
8349 .read_s64 = cpu_rt_runtime_read,
8350 .write_s64 = cpu_rt_runtime_write,
8353 .name = "rt_period_us",
8354 .read_u64 = cpu_rt_period_read_uint,
8355 .write_u64 = cpu_rt_period_write_uint,
8361 struct cgroup_subsys cpu_cgrp_subsys = {
8362 .css_alloc = cpu_cgroup_css_alloc,
8363 .css_free = cpu_cgroup_css_free,
8364 .css_online = cpu_cgroup_css_online,
8365 .css_offline = cpu_cgroup_css_offline,
8366 .fork = cpu_cgroup_fork,
8367 .can_attach = cpu_cgroup_can_attach,
8368 .attach = cpu_cgroup_attach,
8369 .exit = cpu_cgroup_exit,
8370 .legacy_cftypes = cpu_files,
8374 #endif /* CONFIG_CGROUP_SCHED */
8376 void dump_cpu_task(int cpu)
8378 pr_info("Task dump for CPU %d:\n", cpu);
8379 sched_show_task(cpu_curr(cpu));