4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 #ifdef smp_mb__before_atomic
94 void __smp_mb__before_atomic(void)
96 smp_mb__before_atomic();
98 EXPORT_SYMBOL(__smp_mb__before_atomic);
101 #ifdef smp_mb__after_atomic
102 void __smp_mb__after_atomic(void)
104 smp_mb__after_atomic();
106 EXPORT_SYMBOL(__smp_mb__after_atomic);
109 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
112 ktime_t soft, hard, now;
115 if (hrtimer_active(period_timer))
118 now = hrtimer_cb_get_time(period_timer);
119 hrtimer_forward(period_timer, now, period);
121 soft = hrtimer_get_softexpires(period_timer);
122 hard = hrtimer_get_expires(period_timer);
123 delta = ktime_to_ns(ktime_sub(hard, soft));
124 __hrtimer_start_range_ns(period_timer, soft, delta,
125 HRTIMER_MODE_ABS_PINNED, 0);
129 DEFINE_MUTEX(sched_domains_mutex);
130 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
132 static void update_rq_clock_task(struct rq *rq, s64 delta);
134 void update_rq_clock(struct rq *rq)
138 if (rq->skip_clock_update > 0)
141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
145 update_rq_clock_task(rq, delta);
149 * Debugging: various feature bits
152 #define SCHED_FEAT(name, enabled) \
153 (1UL << __SCHED_FEAT_##name) * enabled |
155 const_debug unsigned int sysctl_sched_features =
156 #include "features.h"
161 #ifdef CONFIG_SCHED_DEBUG
162 #define SCHED_FEAT(name, enabled) \
165 static const char * const sched_feat_names[] = {
166 #include "features.h"
171 static int sched_feat_show(struct seq_file *m, void *v)
175 for (i = 0; i < __SCHED_FEAT_NR; i++) {
176 if (!(sysctl_sched_features & (1UL << i)))
178 seq_printf(m, "%s ", sched_feat_names[i]);
185 #ifdef HAVE_JUMP_LABEL
187 #define jump_label_key__true STATIC_KEY_INIT_TRUE
188 #define jump_label_key__false STATIC_KEY_INIT_FALSE
190 #define SCHED_FEAT(name, enabled) \
191 jump_label_key__##enabled ,
193 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
194 #include "features.h"
199 static void sched_feat_disable(int i)
201 if (static_key_enabled(&sched_feat_keys[i]))
202 static_key_slow_dec(&sched_feat_keys[i]);
205 static void sched_feat_enable(int i)
207 if (!static_key_enabled(&sched_feat_keys[i]))
208 static_key_slow_inc(&sched_feat_keys[i]);
211 static void sched_feat_disable(int i) { };
212 static void sched_feat_enable(int i) { };
213 #endif /* HAVE_JUMP_LABEL */
215 static int sched_feat_set(char *cmp)
220 if (strncmp(cmp, "NO_", 3) == 0) {
225 for (i = 0; i < __SCHED_FEAT_NR; i++) {
226 if (strcmp(cmp, sched_feat_names[i]) == 0) {
228 sysctl_sched_features &= ~(1UL << i);
229 sched_feat_disable(i);
231 sysctl_sched_features |= (1UL << i);
232 sched_feat_enable(i);
242 sched_feat_write(struct file *filp, const char __user *ubuf,
243 size_t cnt, loff_t *ppos)
252 if (copy_from_user(&buf, ubuf, cnt))
258 i = sched_feat_set(cmp);
259 if (i == __SCHED_FEAT_NR)
267 static int sched_feat_open(struct inode *inode, struct file *filp)
269 return single_open(filp, sched_feat_show, NULL);
272 static const struct file_operations sched_feat_fops = {
273 .open = sched_feat_open,
274 .write = sched_feat_write,
277 .release = single_release,
280 static __init int sched_init_debug(void)
282 debugfs_create_file("sched_features", 0644, NULL, NULL,
287 late_initcall(sched_init_debug);
288 #endif /* CONFIG_SCHED_DEBUG */
291 * Number of tasks to iterate in a single balance run.
292 * Limited because this is done with IRQs disabled.
294 const_debug unsigned int sysctl_sched_nr_migrate = 32;
297 * period over which we average the RT time consumption, measured
302 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
305 * period over which we measure -rt task cpu usage in us.
308 unsigned int sysctl_sched_rt_period = 1000000;
310 __read_mostly int scheduler_running;
313 * part of the period that we allow rt tasks to run in us.
316 int sysctl_sched_rt_runtime = 950000;
319 * __task_rq_lock - lock the rq @p resides on.
321 static inline struct rq *__task_rq_lock(struct task_struct *p)
326 lockdep_assert_held(&p->pi_lock);
330 raw_spin_lock(&rq->lock);
331 if (likely(rq == task_rq(p)))
333 raw_spin_unlock(&rq->lock);
338 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
340 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
341 __acquires(p->pi_lock)
347 raw_spin_lock_irqsave(&p->pi_lock, *flags);
349 raw_spin_lock(&rq->lock);
350 if (likely(rq == task_rq(p)))
352 raw_spin_unlock(&rq->lock);
353 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
357 static void __task_rq_unlock(struct rq *rq)
360 raw_spin_unlock(&rq->lock);
364 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
366 __releases(p->pi_lock)
368 raw_spin_unlock(&rq->lock);
369 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
373 * this_rq_lock - lock this runqueue and disable interrupts.
375 static struct rq *this_rq_lock(void)
382 raw_spin_lock(&rq->lock);
387 #ifdef CONFIG_SCHED_HRTICK
389 * Use HR-timers to deliver accurate preemption points.
392 static void hrtick_clear(struct rq *rq)
394 if (hrtimer_active(&rq->hrtick_timer))
395 hrtimer_cancel(&rq->hrtick_timer);
399 * High-resolution timer tick.
400 * Runs from hardirq context with interrupts disabled.
402 static enum hrtimer_restart hrtick(struct hrtimer *timer)
404 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
406 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
408 raw_spin_lock(&rq->lock);
410 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
411 raw_spin_unlock(&rq->lock);
413 return HRTIMER_NORESTART;
418 static int __hrtick_restart(struct rq *rq)
420 struct hrtimer *timer = &rq->hrtick_timer;
421 ktime_t time = hrtimer_get_softexpires(timer);
423 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
427 * called from hardirq (IPI) context
429 static void __hrtick_start(void *arg)
433 raw_spin_lock(&rq->lock);
434 __hrtick_restart(rq);
435 rq->hrtick_csd_pending = 0;
436 raw_spin_unlock(&rq->lock);
440 * Called to set the hrtick timer state.
442 * called with rq->lock held and irqs disabled
444 void hrtick_start(struct rq *rq, u64 delay)
446 struct hrtimer *timer = &rq->hrtick_timer;
447 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
449 hrtimer_set_expires(timer, time);
451 if (rq == this_rq()) {
452 __hrtick_restart(rq);
453 } else if (!rq->hrtick_csd_pending) {
454 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
455 rq->hrtick_csd_pending = 1;
460 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
462 int cpu = (int)(long)hcpu;
465 case CPU_UP_CANCELED:
466 case CPU_UP_CANCELED_FROZEN:
467 case CPU_DOWN_PREPARE:
468 case CPU_DOWN_PREPARE_FROZEN:
470 case CPU_DEAD_FROZEN:
471 hrtick_clear(cpu_rq(cpu));
478 static __init void init_hrtick(void)
480 hotcpu_notifier(hotplug_hrtick, 0);
484 * Called to set the hrtick timer state.
486 * called with rq->lock held and irqs disabled
488 void hrtick_start(struct rq *rq, u64 delay)
490 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
491 HRTIMER_MODE_REL_PINNED, 0);
494 static inline void init_hrtick(void)
497 #endif /* CONFIG_SMP */
499 static void init_rq_hrtick(struct rq *rq)
502 rq->hrtick_csd_pending = 0;
504 rq->hrtick_csd.flags = 0;
505 rq->hrtick_csd.func = __hrtick_start;
506 rq->hrtick_csd.info = rq;
509 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
510 rq->hrtick_timer.function = hrtick;
512 #else /* CONFIG_SCHED_HRTICK */
513 static inline void hrtick_clear(struct rq *rq)
517 static inline void init_rq_hrtick(struct rq *rq)
521 static inline void init_hrtick(void)
524 #endif /* CONFIG_SCHED_HRTICK */
527 * cmpxchg based fetch_or, macro so it works for different integer types
529 #define fetch_or(ptr, val) \
530 ({ typeof(*(ptr)) __old, __val = *(ptr); \
532 __old = cmpxchg((ptr), __val, __val | (val)); \
533 if (__old == __val) \
540 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
542 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
543 * this avoids any races wrt polling state changes and thereby avoids
546 static bool set_nr_and_not_polling(struct task_struct *p)
548 struct thread_info *ti = task_thread_info(p);
549 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
553 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
555 * If this returns true, then the idle task promises to call
556 * sched_ttwu_pending() and reschedule soon.
558 static bool set_nr_if_polling(struct task_struct *p)
560 struct thread_info *ti = task_thread_info(p);
561 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
564 if (!(val & _TIF_POLLING_NRFLAG))
566 if (val & _TIF_NEED_RESCHED)
568 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
577 static bool set_nr_and_not_polling(struct task_struct *p)
579 set_tsk_need_resched(p);
584 static bool set_nr_if_polling(struct task_struct *p)
592 * resched_curr - mark rq's current task 'to be rescheduled now'.
594 * On UP this means the setting of the need_resched flag, on SMP it
595 * might also involve a cross-CPU call to trigger the scheduler on
598 void resched_curr(struct rq *rq)
600 struct task_struct *curr = rq->curr;
603 lockdep_assert_held(&rq->lock);
605 if (test_tsk_need_resched(curr))
610 if (cpu == smp_processor_id()) {
611 set_tsk_need_resched(curr);
612 set_preempt_need_resched();
616 if (set_nr_and_not_polling(curr))
617 smp_send_reschedule(cpu);
619 trace_sched_wake_idle_without_ipi(cpu);
622 void resched_cpu(int cpu)
624 struct rq *rq = cpu_rq(cpu);
627 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
630 raw_spin_unlock_irqrestore(&rq->lock, flags);
634 #ifdef CONFIG_NO_HZ_COMMON
636 * In the semi idle case, use the nearest busy cpu for migrating timers
637 * from an idle cpu. This is good for power-savings.
639 * We don't do similar optimization for completely idle system, as
640 * selecting an idle cpu will add more delays to the timers than intended
641 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
643 int get_nohz_timer_target(int pinned)
645 int cpu = smp_processor_id();
647 struct sched_domain *sd;
649 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
653 for_each_domain(cpu, sd) {
654 for_each_cpu(i, sched_domain_span(sd)) {
666 * When add_timer_on() enqueues a timer into the timer wheel of an
667 * idle CPU then this timer might expire before the next timer event
668 * which is scheduled to wake up that CPU. In case of a completely
669 * idle system the next event might even be infinite time into the
670 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
671 * leaves the inner idle loop so the newly added timer is taken into
672 * account when the CPU goes back to idle and evaluates the timer
673 * wheel for the next timer event.
675 static void wake_up_idle_cpu(int cpu)
677 struct rq *rq = cpu_rq(cpu);
679 if (cpu == smp_processor_id())
682 if (set_nr_and_not_polling(rq->idle))
683 smp_send_reschedule(cpu);
685 trace_sched_wake_idle_without_ipi(cpu);
688 static bool wake_up_full_nohz_cpu(int cpu)
691 * We just need the target to call irq_exit() and re-evaluate
692 * the next tick. The nohz full kick at least implies that.
693 * If needed we can still optimize that later with an
696 if (tick_nohz_full_cpu(cpu)) {
697 if (cpu != smp_processor_id() ||
698 tick_nohz_tick_stopped())
699 tick_nohz_full_kick_cpu(cpu);
706 void wake_up_nohz_cpu(int cpu)
708 if (!wake_up_full_nohz_cpu(cpu))
709 wake_up_idle_cpu(cpu);
712 static inline bool got_nohz_idle_kick(void)
714 int cpu = smp_processor_id();
716 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
719 if (idle_cpu(cpu) && !need_resched())
723 * We can't run Idle Load Balance on this CPU for this time so we
724 * cancel it and clear NOHZ_BALANCE_KICK
726 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
730 #else /* CONFIG_NO_HZ_COMMON */
732 static inline bool got_nohz_idle_kick(void)
737 #endif /* CONFIG_NO_HZ_COMMON */
739 #ifdef CONFIG_NO_HZ_FULL
740 bool sched_can_stop_tick(void)
743 * More than one running task need preemption.
744 * nr_running update is assumed to be visible
745 * after IPI is sent from wakers.
747 if (this_rq()->nr_running > 1)
752 #endif /* CONFIG_NO_HZ_FULL */
754 void sched_avg_update(struct rq *rq)
756 s64 period = sched_avg_period();
758 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
760 * Inline assembly required to prevent the compiler
761 * optimising this loop into a divmod call.
762 * See __iter_div_u64_rem() for another example of this.
764 asm("" : "+rm" (rq->age_stamp));
765 rq->age_stamp += period;
770 #endif /* CONFIG_SMP */
772 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
773 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
775 * Iterate task_group tree rooted at *from, calling @down when first entering a
776 * node and @up when leaving it for the final time.
778 * Caller must hold rcu_lock or sufficient equivalent.
780 int walk_tg_tree_from(struct task_group *from,
781 tg_visitor down, tg_visitor up, void *data)
783 struct task_group *parent, *child;
789 ret = (*down)(parent, data);
792 list_for_each_entry_rcu(child, &parent->children, siblings) {
799 ret = (*up)(parent, data);
800 if (ret || parent == from)
804 parent = parent->parent;
811 int tg_nop(struct task_group *tg, void *data)
817 static void set_load_weight(struct task_struct *p)
819 int prio = p->static_prio - MAX_RT_PRIO;
820 struct load_weight *load = &p->se.load;
823 * SCHED_IDLE tasks get minimal weight:
825 if (p->policy == SCHED_IDLE) {
826 load->weight = scale_load(WEIGHT_IDLEPRIO);
827 load->inv_weight = WMULT_IDLEPRIO;
831 load->weight = scale_load(prio_to_weight[prio]);
832 load->inv_weight = prio_to_wmult[prio];
835 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
838 sched_info_queued(rq, p);
839 p->sched_class->enqueue_task(rq, p, flags);
842 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
845 sched_info_dequeued(rq, p);
846 p->sched_class->dequeue_task(rq, p, flags);
849 void activate_task(struct rq *rq, struct task_struct *p, int flags)
851 if (task_contributes_to_load(p))
852 rq->nr_uninterruptible--;
854 enqueue_task(rq, p, flags);
857 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
859 if (task_contributes_to_load(p))
860 rq->nr_uninterruptible++;
862 dequeue_task(rq, p, flags);
865 static void update_rq_clock_task(struct rq *rq, s64 delta)
868 * In theory, the compile should just see 0 here, and optimize out the call
869 * to sched_rt_avg_update. But I don't trust it...
871 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
872 s64 steal = 0, irq_delta = 0;
874 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
878 * Since irq_time is only updated on {soft,}irq_exit, we might run into
879 * this case when a previous update_rq_clock() happened inside a
882 * When this happens, we stop ->clock_task and only update the
883 * prev_irq_time stamp to account for the part that fit, so that a next
884 * update will consume the rest. This ensures ->clock_task is
887 * It does however cause some slight miss-attribution of {soft,}irq
888 * time, a more accurate solution would be to update the irq_time using
889 * the current rq->clock timestamp, except that would require using
892 if (irq_delta > delta)
895 rq->prev_irq_time += irq_delta;
898 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
899 if (static_key_false((¶virt_steal_rq_enabled))) {
900 steal = paravirt_steal_clock(cpu_of(rq));
901 steal -= rq->prev_steal_time_rq;
903 if (unlikely(steal > delta))
906 rq->prev_steal_time_rq += steal;
911 rq->clock_task += delta;
913 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
914 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
915 sched_rt_avg_update(rq, irq_delta + steal);
919 void sched_set_stop_task(int cpu, struct task_struct *stop)
921 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
922 struct task_struct *old_stop = cpu_rq(cpu)->stop;
926 * Make it appear like a SCHED_FIFO task, its something
927 * userspace knows about and won't get confused about.
929 * Also, it will make PI more or less work without too
930 * much confusion -- but then, stop work should not
931 * rely on PI working anyway.
933 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
935 stop->sched_class = &stop_sched_class;
938 cpu_rq(cpu)->stop = stop;
942 * Reset it back to a normal scheduling class so that
943 * it can die in pieces.
945 old_stop->sched_class = &rt_sched_class;
950 * __normal_prio - return the priority that is based on the static prio
952 static inline int __normal_prio(struct task_struct *p)
954 return p->static_prio;
958 * Calculate the expected normal priority: i.e. priority
959 * without taking RT-inheritance into account. Might be
960 * boosted by interactivity modifiers. Changes upon fork,
961 * setprio syscalls, and whenever the interactivity
962 * estimator recalculates.
964 static inline int normal_prio(struct task_struct *p)
968 if (task_has_dl_policy(p))
969 prio = MAX_DL_PRIO-1;
970 else if (task_has_rt_policy(p))
971 prio = MAX_RT_PRIO-1 - p->rt_priority;
973 prio = __normal_prio(p);
978 * Calculate the current priority, i.e. the priority
979 * taken into account by the scheduler. This value might
980 * be boosted by RT tasks, or might be boosted by
981 * interactivity modifiers. Will be RT if the task got
982 * RT-boosted. If not then it returns p->normal_prio.
984 static int effective_prio(struct task_struct *p)
986 p->normal_prio = normal_prio(p);
988 * If we are RT tasks or we were boosted to RT priority,
989 * keep the priority unchanged. Otherwise, update priority
990 * to the normal priority:
992 if (!rt_prio(p->prio))
993 return p->normal_prio;
998 * task_curr - is this task currently executing on a CPU?
999 * @p: the task in question.
1001 * Return: 1 if the task is currently executing. 0 otherwise.
1003 inline int task_curr(const struct task_struct *p)
1005 return cpu_curr(task_cpu(p)) == p;
1008 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1009 const struct sched_class *prev_class,
1012 if (prev_class != p->sched_class) {
1013 if (prev_class->switched_from)
1014 prev_class->switched_from(rq, p);
1015 p->sched_class->switched_to(rq, p);
1016 } else if (oldprio != p->prio || dl_task(p))
1017 p->sched_class->prio_changed(rq, p, oldprio);
1020 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1022 const struct sched_class *class;
1024 if (p->sched_class == rq->curr->sched_class) {
1025 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1027 for_each_class(class) {
1028 if (class == rq->curr->sched_class)
1030 if (class == p->sched_class) {
1038 * A queue event has occurred, and we're going to schedule. In
1039 * this case, we can save a useless back to back clock update.
1041 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1042 rq->skip_clock_update = 1;
1046 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1048 #ifdef CONFIG_SCHED_DEBUG
1050 * We should never call set_task_cpu() on a blocked task,
1051 * ttwu() will sort out the placement.
1053 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1054 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1056 #ifdef CONFIG_LOCKDEP
1058 * The caller should hold either p->pi_lock or rq->lock, when changing
1059 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1061 * sched_move_task() holds both and thus holding either pins the cgroup,
1064 * Furthermore, all task_rq users should acquire both locks, see
1067 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1068 lockdep_is_held(&task_rq(p)->lock)));
1072 trace_sched_migrate_task(p, new_cpu);
1074 if (task_cpu(p) != new_cpu) {
1075 if (p->sched_class->migrate_task_rq)
1076 p->sched_class->migrate_task_rq(p, new_cpu);
1077 p->se.nr_migrations++;
1078 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1081 __set_task_cpu(p, new_cpu);
1084 static void __migrate_swap_task(struct task_struct *p, int cpu)
1087 struct rq *src_rq, *dst_rq;
1089 src_rq = task_rq(p);
1090 dst_rq = cpu_rq(cpu);
1092 deactivate_task(src_rq, p, 0);
1093 set_task_cpu(p, cpu);
1094 activate_task(dst_rq, p, 0);
1095 check_preempt_curr(dst_rq, p, 0);
1098 * Task isn't running anymore; make it appear like we migrated
1099 * it before it went to sleep. This means on wakeup we make the
1100 * previous cpu our targer instead of where it really is.
1106 struct migration_swap_arg {
1107 struct task_struct *src_task, *dst_task;
1108 int src_cpu, dst_cpu;
1111 static int migrate_swap_stop(void *data)
1113 struct migration_swap_arg *arg = data;
1114 struct rq *src_rq, *dst_rq;
1117 src_rq = cpu_rq(arg->src_cpu);
1118 dst_rq = cpu_rq(arg->dst_cpu);
1120 double_raw_lock(&arg->src_task->pi_lock,
1121 &arg->dst_task->pi_lock);
1122 double_rq_lock(src_rq, dst_rq);
1123 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1126 if (task_cpu(arg->src_task) != arg->src_cpu)
1129 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1132 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1135 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1136 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1141 double_rq_unlock(src_rq, dst_rq);
1142 raw_spin_unlock(&arg->dst_task->pi_lock);
1143 raw_spin_unlock(&arg->src_task->pi_lock);
1149 * Cross migrate two tasks
1151 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1153 struct migration_swap_arg arg;
1156 arg = (struct migration_swap_arg){
1158 .src_cpu = task_cpu(cur),
1160 .dst_cpu = task_cpu(p),
1163 if (arg.src_cpu == arg.dst_cpu)
1167 * These three tests are all lockless; this is OK since all of them
1168 * will be re-checked with proper locks held further down the line.
1170 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1173 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1176 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1179 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1180 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1186 struct migration_arg {
1187 struct task_struct *task;
1191 static int migration_cpu_stop(void *data);
1194 * wait_task_inactive - wait for a thread to unschedule.
1196 * If @match_state is nonzero, it's the @p->state value just checked and
1197 * not expected to change. If it changes, i.e. @p might have woken up,
1198 * then return zero. When we succeed in waiting for @p to be off its CPU,
1199 * we return a positive number (its total switch count). If a second call
1200 * a short while later returns the same number, the caller can be sure that
1201 * @p has remained unscheduled the whole time.
1203 * The caller must ensure that the task *will* unschedule sometime soon,
1204 * else this function might spin for a *long* time. This function can't
1205 * be called with interrupts off, or it may introduce deadlock with
1206 * smp_call_function() if an IPI is sent by the same process we are
1207 * waiting to become inactive.
1209 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1211 unsigned long flags;
1218 * We do the initial early heuristics without holding
1219 * any task-queue locks at all. We'll only try to get
1220 * the runqueue lock when things look like they will
1226 * If the task is actively running on another CPU
1227 * still, just relax and busy-wait without holding
1230 * NOTE! Since we don't hold any locks, it's not
1231 * even sure that "rq" stays as the right runqueue!
1232 * But we don't care, since "task_running()" will
1233 * return false if the runqueue has changed and p
1234 * is actually now running somewhere else!
1236 while (task_running(rq, p)) {
1237 if (match_state && unlikely(p->state != match_state))
1243 * Ok, time to look more closely! We need the rq
1244 * lock now, to be *sure*. If we're wrong, we'll
1245 * just go back and repeat.
1247 rq = task_rq_lock(p, &flags);
1248 trace_sched_wait_task(p);
1249 running = task_running(rq, p);
1252 if (!match_state || p->state == match_state)
1253 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1254 task_rq_unlock(rq, p, &flags);
1257 * If it changed from the expected state, bail out now.
1259 if (unlikely(!ncsw))
1263 * Was it really running after all now that we
1264 * checked with the proper locks actually held?
1266 * Oops. Go back and try again..
1268 if (unlikely(running)) {
1274 * It's not enough that it's not actively running,
1275 * it must be off the runqueue _entirely_, and not
1278 * So if it was still runnable (but just not actively
1279 * running right now), it's preempted, and we should
1280 * yield - it could be a while.
1282 if (unlikely(on_rq)) {
1283 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1285 set_current_state(TASK_UNINTERRUPTIBLE);
1286 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1291 * Ahh, all good. It wasn't running, and it wasn't
1292 * runnable, which means that it will never become
1293 * running in the future either. We're all done!
1302 * kick_process - kick a running thread to enter/exit the kernel
1303 * @p: the to-be-kicked thread
1305 * Cause a process which is running on another CPU to enter
1306 * kernel-mode, without any delay. (to get signals handled.)
1308 * NOTE: this function doesn't have to take the runqueue lock,
1309 * because all it wants to ensure is that the remote task enters
1310 * the kernel. If the IPI races and the task has been migrated
1311 * to another CPU then no harm is done and the purpose has been
1314 void kick_process(struct task_struct *p)
1320 if ((cpu != smp_processor_id()) && task_curr(p))
1321 smp_send_reschedule(cpu);
1324 EXPORT_SYMBOL_GPL(kick_process);
1325 #endif /* CONFIG_SMP */
1329 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1331 static int select_fallback_rq(int cpu, struct task_struct *p)
1333 int nid = cpu_to_node(cpu);
1334 const struct cpumask *nodemask = NULL;
1335 enum { cpuset, possible, fail } state = cpuset;
1339 * If the node that the cpu is on has been offlined, cpu_to_node()
1340 * will return -1. There is no cpu on the node, and we should
1341 * select the cpu on the other node.
1344 nodemask = cpumask_of_node(nid);
1346 /* Look for allowed, online CPU in same node. */
1347 for_each_cpu(dest_cpu, nodemask) {
1348 if (!cpu_online(dest_cpu))
1350 if (!cpu_active(dest_cpu))
1352 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1358 /* Any allowed, online CPU? */
1359 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1360 if (!cpu_online(dest_cpu))
1362 if (!cpu_active(dest_cpu))
1369 /* No more Mr. Nice Guy. */
1370 cpuset_cpus_allowed_fallback(p);
1375 do_set_cpus_allowed(p, cpu_possible_mask);
1386 if (state != cpuset) {
1388 * Don't tell them about moving exiting tasks or
1389 * kernel threads (both mm NULL), since they never
1392 if (p->mm && printk_ratelimit()) {
1393 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1394 task_pid_nr(p), p->comm, cpu);
1402 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1405 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1407 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1410 * In order not to call set_task_cpu() on a blocking task we need
1411 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1414 * Since this is common to all placement strategies, this lives here.
1416 * [ this allows ->select_task() to simply return task_cpu(p) and
1417 * not worry about this generic constraint ]
1419 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1421 cpu = select_fallback_rq(task_cpu(p), p);
1426 static void update_avg(u64 *avg, u64 sample)
1428 s64 diff = sample - *avg;
1434 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1436 #ifdef CONFIG_SCHEDSTATS
1437 struct rq *rq = this_rq();
1440 int this_cpu = smp_processor_id();
1442 if (cpu == this_cpu) {
1443 schedstat_inc(rq, ttwu_local);
1444 schedstat_inc(p, se.statistics.nr_wakeups_local);
1446 struct sched_domain *sd;
1448 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1450 for_each_domain(this_cpu, sd) {
1451 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1452 schedstat_inc(sd, ttwu_wake_remote);
1459 if (wake_flags & WF_MIGRATED)
1460 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1462 #endif /* CONFIG_SMP */
1464 schedstat_inc(rq, ttwu_count);
1465 schedstat_inc(p, se.statistics.nr_wakeups);
1467 if (wake_flags & WF_SYNC)
1468 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1470 #endif /* CONFIG_SCHEDSTATS */
1473 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1475 activate_task(rq, p, en_flags);
1478 /* if a worker is waking up, notify workqueue */
1479 if (p->flags & PF_WQ_WORKER)
1480 wq_worker_waking_up(p, cpu_of(rq));
1484 * Mark the task runnable and perform wakeup-preemption.
1487 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1489 check_preempt_curr(rq, p, wake_flags);
1490 trace_sched_wakeup(p, true);
1492 p->state = TASK_RUNNING;
1494 if (p->sched_class->task_woken)
1495 p->sched_class->task_woken(rq, p);
1497 if (rq->idle_stamp) {
1498 u64 delta = rq_clock(rq) - rq->idle_stamp;
1499 u64 max = 2*rq->max_idle_balance_cost;
1501 update_avg(&rq->avg_idle, delta);
1503 if (rq->avg_idle > max)
1512 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1515 if (p->sched_contributes_to_load)
1516 rq->nr_uninterruptible--;
1519 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1520 ttwu_do_wakeup(rq, p, wake_flags);
1524 * Called in case the task @p isn't fully descheduled from its runqueue,
1525 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1526 * since all we need to do is flip p->state to TASK_RUNNING, since
1527 * the task is still ->on_rq.
1529 static int ttwu_remote(struct task_struct *p, int wake_flags)
1534 rq = __task_rq_lock(p);
1536 /* check_preempt_curr() may use rq clock */
1537 update_rq_clock(rq);
1538 ttwu_do_wakeup(rq, p, wake_flags);
1541 __task_rq_unlock(rq);
1547 void sched_ttwu_pending(void)
1549 struct rq *rq = this_rq();
1550 struct llist_node *llist = llist_del_all(&rq->wake_list);
1551 struct task_struct *p;
1552 unsigned long flags;
1557 raw_spin_lock_irqsave(&rq->lock, flags);
1560 p = llist_entry(llist, struct task_struct, wake_entry);
1561 llist = llist_next(llist);
1562 ttwu_do_activate(rq, p, 0);
1565 raw_spin_unlock_irqrestore(&rq->lock, flags);
1568 void scheduler_ipi(void)
1571 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1572 * TIF_NEED_RESCHED remotely (for the first time) will also send
1575 preempt_fold_need_resched();
1577 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1581 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1582 * traditionally all their work was done from the interrupt return
1583 * path. Now that we actually do some work, we need to make sure
1586 * Some archs already do call them, luckily irq_enter/exit nest
1589 * Arguably we should visit all archs and update all handlers,
1590 * however a fair share of IPIs are still resched only so this would
1591 * somewhat pessimize the simple resched case.
1594 sched_ttwu_pending();
1597 * Check if someone kicked us for doing the nohz idle load balance.
1599 if (unlikely(got_nohz_idle_kick())) {
1600 this_rq()->idle_balance = 1;
1601 raise_softirq_irqoff(SCHED_SOFTIRQ);
1606 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1608 struct rq *rq = cpu_rq(cpu);
1610 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1611 if (!set_nr_if_polling(rq->idle))
1612 smp_send_reschedule(cpu);
1614 trace_sched_wake_idle_without_ipi(cpu);
1618 bool cpus_share_cache(int this_cpu, int that_cpu)
1620 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1622 #endif /* CONFIG_SMP */
1624 static void ttwu_queue(struct task_struct *p, int cpu)
1626 struct rq *rq = cpu_rq(cpu);
1628 #if defined(CONFIG_SMP)
1629 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1630 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1631 ttwu_queue_remote(p, cpu);
1636 raw_spin_lock(&rq->lock);
1637 ttwu_do_activate(rq, p, 0);
1638 raw_spin_unlock(&rq->lock);
1642 * try_to_wake_up - wake up a thread
1643 * @p: the thread to be awakened
1644 * @state: the mask of task states that can be woken
1645 * @wake_flags: wake modifier flags (WF_*)
1647 * Put it on the run-queue if it's not already there. The "current"
1648 * thread is always on the run-queue (except when the actual
1649 * re-schedule is in progress), and as such you're allowed to do
1650 * the simpler "current->state = TASK_RUNNING" to mark yourself
1651 * runnable without the overhead of this.
1653 * Return: %true if @p was woken up, %false if it was already running.
1654 * or @state didn't match @p's state.
1657 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1659 unsigned long flags;
1660 int cpu, success = 0;
1663 * If we are going to wake up a thread waiting for CONDITION we
1664 * need to ensure that CONDITION=1 done by the caller can not be
1665 * reordered with p->state check below. This pairs with mb() in
1666 * set_current_state() the waiting thread does.
1668 smp_mb__before_spinlock();
1669 raw_spin_lock_irqsave(&p->pi_lock, flags);
1670 if (!(p->state & state))
1673 success = 1; /* we're going to change ->state */
1676 if (p->on_rq && ttwu_remote(p, wake_flags))
1681 * If the owning (remote) cpu is still in the middle of schedule() with
1682 * this task as prev, wait until its done referencing the task.
1687 * Pairs with the smp_wmb() in finish_lock_switch().
1691 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1692 p->state = TASK_WAKING;
1694 if (p->sched_class->task_waking)
1695 p->sched_class->task_waking(p);
1697 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1698 if (task_cpu(p) != cpu) {
1699 wake_flags |= WF_MIGRATED;
1700 set_task_cpu(p, cpu);
1702 #endif /* CONFIG_SMP */
1706 ttwu_stat(p, cpu, wake_flags);
1708 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1714 * try_to_wake_up_local - try to wake up a local task with rq lock held
1715 * @p: the thread to be awakened
1717 * Put @p on the run-queue if it's not already there. The caller must
1718 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1721 static void try_to_wake_up_local(struct task_struct *p)
1723 struct rq *rq = task_rq(p);
1725 if (WARN_ON_ONCE(rq != this_rq()) ||
1726 WARN_ON_ONCE(p == current))
1729 lockdep_assert_held(&rq->lock);
1731 if (!raw_spin_trylock(&p->pi_lock)) {
1732 raw_spin_unlock(&rq->lock);
1733 raw_spin_lock(&p->pi_lock);
1734 raw_spin_lock(&rq->lock);
1737 if (!(p->state & TASK_NORMAL))
1741 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1743 ttwu_do_wakeup(rq, p, 0);
1744 ttwu_stat(p, smp_processor_id(), 0);
1746 raw_spin_unlock(&p->pi_lock);
1750 * wake_up_process - Wake up a specific process
1751 * @p: The process to be woken up.
1753 * Attempt to wake up the nominated process and move it to the set of runnable
1756 * Return: 1 if the process was woken up, 0 if it was already running.
1758 * It may be assumed that this function implies a write memory barrier before
1759 * changing the task state if and only if any tasks are woken up.
1761 int wake_up_process(struct task_struct *p)
1763 WARN_ON(task_is_stopped_or_traced(p));
1764 return try_to_wake_up(p, TASK_NORMAL, 0);
1766 EXPORT_SYMBOL(wake_up_process);
1768 int wake_up_state(struct task_struct *p, unsigned int state)
1770 return try_to_wake_up(p, state, 0);
1774 * Perform scheduler related setup for a newly forked process p.
1775 * p is forked by current.
1777 * __sched_fork() is basic setup used by init_idle() too:
1779 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1784 p->se.exec_start = 0;
1785 p->se.sum_exec_runtime = 0;
1786 p->se.prev_sum_exec_runtime = 0;
1787 p->se.nr_migrations = 0;
1789 INIT_LIST_HEAD(&p->se.group_node);
1791 #ifdef CONFIG_SCHEDSTATS
1792 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1795 RB_CLEAR_NODE(&p->dl.rb_node);
1796 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1797 p->dl.dl_runtime = p->dl.runtime = 0;
1798 p->dl.dl_deadline = p->dl.deadline = 0;
1799 p->dl.dl_period = 0;
1802 INIT_LIST_HEAD(&p->rt.run_list);
1804 #ifdef CONFIG_PREEMPT_NOTIFIERS
1805 INIT_HLIST_HEAD(&p->preempt_notifiers);
1808 #ifdef CONFIG_NUMA_BALANCING
1809 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1810 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1811 p->mm->numa_scan_seq = 0;
1814 if (clone_flags & CLONE_VM)
1815 p->numa_preferred_nid = current->numa_preferred_nid;
1817 p->numa_preferred_nid = -1;
1819 p->node_stamp = 0ULL;
1820 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1821 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1822 p->numa_work.next = &p->numa_work;
1823 p->numa_faults_memory = NULL;
1824 p->numa_faults_buffer_memory = NULL;
1825 p->last_task_numa_placement = 0;
1826 p->last_sum_exec_runtime = 0;
1828 INIT_LIST_HEAD(&p->numa_entry);
1829 p->numa_group = NULL;
1830 #endif /* CONFIG_NUMA_BALANCING */
1833 #ifdef CONFIG_NUMA_BALANCING
1834 #ifdef CONFIG_SCHED_DEBUG
1835 void set_numabalancing_state(bool enabled)
1838 sched_feat_set("NUMA");
1840 sched_feat_set("NO_NUMA");
1843 __read_mostly bool numabalancing_enabled;
1845 void set_numabalancing_state(bool enabled)
1847 numabalancing_enabled = enabled;
1849 #endif /* CONFIG_SCHED_DEBUG */
1851 #ifdef CONFIG_PROC_SYSCTL
1852 int sysctl_numa_balancing(struct ctl_table *table, int write,
1853 void __user *buffer, size_t *lenp, loff_t *ppos)
1857 int state = numabalancing_enabled;
1859 if (write && !capable(CAP_SYS_ADMIN))
1864 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1868 set_numabalancing_state(state);
1875 * fork()/clone()-time setup:
1877 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1879 unsigned long flags;
1880 int cpu = get_cpu();
1882 __sched_fork(clone_flags, p);
1884 * We mark the process as running here. This guarantees that
1885 * nobody will actually run it, and a signal or other external
1886 * event cannot wake it up and insert it on the runqueue either.
1888 p->state = TASK_RUNNING;
1891 * Make sure we do not leak PI boosting priority to the child.
1893 p->prio = current->normal_prio;
1896 * Revert to default priority/policy on fork if requested.
1898 if (unlikely(p->sched_reset_on_fork)) {
1899 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1900 p->policy = SCHED_NORMAL;
1901 p->static_prio = NICE_TO_PRIO(0);
1903 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1904 p->static_prio = NICE_TO_PRIO(0);
1906 p->prio = p->normal_prio = __normal_prio(p);
1910 * We don't need the reset flag anymore after the fork. It has
1911 * fulfilled its duty:
1913 p->sched_reset_on_fork = 0;
1916 if (dl_prio(p->prio)) {
1919 } else if (rt_prio(p->prio)) {
1920 p->sched_class = &rt_sched_class;
1922 p->sched_class = &fair_sched_class;
1925 if (p->sched_class->task_fork)
1926 p->sched_class->task_fork(p);
1929 * The child is not yet in the pid-hash so no cgroup attach races,
1930 * and the cgroup is pinned to this child due to cgroup_fork()
1931 * is ran before sched_fork().
1933 * Silence PROVE_RCU.
1935 raw_spin_lock_irqsave(&p->pi_lock, flags);
1936 set_task_cpu(p, cpu);
1937 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1939 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1940 if (likely(sched_info_on()))
1941 memset(&p->sched_info, 0, sizeof(p->sched_info));
1943 #if defined(CONFIG_SMP)
1946 init_task_preempt_count(p);
1948 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1949 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1956 unsigned long to_ratio(u64 period, u64 runtime)
1958 if (runtime == RUNTIME_INF)
1962 * Doing this here saves a lot of checks in all
1963 * the calling paths, and returning zero seems
1964 * safe for them anyway.
1969 return div64_u64(runtime << 20, period);
1973 inline struct dl_bw *dl_bw_of(int i)
1975 return &cpu_rq(i)->rd->dl_bw;
1978 static inline int dl_bw_cpus(int i)
1980 struct root_domain *rd = cpu_rq(i)->rd;
1983 for_each_cpu_and(i, rd->span, cpu_active_mask)
1989 inline struct dl_bw *dl_bw_of(int i)
1991 return &cpu_rq(i)->dl.dl_bw;
1994 static inline int dl_bw_cpus(int i)
2001 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2003 dl_b->total_bw -= tsk_bw;
2007 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2009 dl_b->total_bw += tsk_bw;
2013 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2015 return dl_b->bw != -1 &&
2016 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2020 * We must be sure that accepting a new task (or allowing changing the
2021 * parameters of an existing one) is consistent with the bandwidth
2022 * constraints. If yes, this function also accordingly updates the currently
2023 * allocated bandwidth to reflect the new situation.
2025 * This function is called while holding p's rq->lock.
2027 static int dl_overflow(struct task_struct *p, int policy,
2028 const struct sched_attr *attr)
2031 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2032 u64 period = attr->sched_period ?: attr->sched_deadline;
2033 u64 runtime = attr->sched_runtime;
2034 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2037 if (new_bw == p->dl.dl_bw)
2041 * Either if a task, enters, leave, or stays -deadline but changes
2042 * its parameters, we may need to update accordingly the total
2043 * allocated bandwidth of the container.
2045 raw_spin_lock(&dl_b->lock);
2046 cpus = dl_bw_cpus(task_cpu(p));
2047 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2048 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2049 __dl_add(dl_b, new_bw);
2051 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2052 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2053 __dl_clear(dl_b, p->dl.dl_bw);
2054 __dl_add(dl_b, new_bw);
2056 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2057 __dl_clear(dl_b, p->dl.dl_bw);
2060 raw_spin_unlock(&dl_b->lock);
2065 extern void init_dl_bw(struct dl_bw *dl_b);
2068 * wake_up_new_task - wake up a newly created task for the first time.
2070 * This function will do some initial scheduler statistics housekeeping
2071 * that must be done for every newly created context, then puts the task
2072 * on the runqueue and wakes it.
2074 void wake_up_new_task(struct task_struct *p)
2076 unsigned long flags;
2079 raw_spin_lock_irqsave(&p->pi_lock, flags);
2082 * Fork balancing, do it here and not earlier because:
2083 * - cpus_allowed can change in the fork path
2084 * - any previously selected cpu might disappear through hotplug
2086 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2089 /* Initialize new task's runnable average */
2090 init_task_runnable_average(p);
2091 rq = __task_rq_lock(p);
2092 activate_task(rq, p, 0);
2094 trace_sched_wakeup_new(p, true);
2095 check_preempt_curr(rq, p, WF_FORK);
2097 if (p->sched_class->task_woken)
2098 p->sched_class->task_woken(rq, p);
2100 task_rq_unlock(rq, p, &flags);
2103 #ifdef CONFIG_PREEMPT_NOTIFIERS
2106 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2107 * @notifier: notifier struct to register
2109 void preempt_notifier_register(struct preempt_notifier *notifier)
2111 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2113 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2116 * preempt_notifier_unregister - no longer interested in preemption notifications
2117 * @notifier: notifier struct to unregister
2119 * This is safe to call from within a preemption notifier.
2121 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2123 hlist_del(¬ifier->link);
2125 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2127 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2129 struct preempt_notifier *notifier;
2131 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2132 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2136 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2137 struct task_struct *next)
2139 struct preempt_notifier *notifier;
2141 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2142 notifier->ops->sched_out(notifier, next);
2145 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2147 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2152 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2153 struct task_struct *next)
2157 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2160 * prepare_task_switch - prepare to switch tasks
2161 * @rq: the runqueue preparing to switch
2162 * @prev: the current task that is being switched out
2163 * @next: the task we are going to switch to.
2165 * This is called with the rq lock held and interrupts off. It must
2166 * be paired with a subsequent finish_task_switch after the context
2169 * prepare_task_switch sets up locking and calls architecture specific
2173 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2174 struct task_struct *next)
2176 trace_sched_switch(prev, next);
2177 sched_info_switch(rq, prev, next);
2178 perf_event_task_sched_out(prev, next);
2179 fire_sched_out_preempt_notifiers(prev, next);
2180 prepare_lock_switch(rq, next);
2181 prepare_arch_switch(next);
2185 * finish_task_switch - clean up after a task-switch
2186 * @rq: runqueue associated with task-switch
2187 * @prev: the thread we just switched away from.
2189 * finish_task_switch must be called after the context switch, paired
2190 * with a prepare_task_switch call before the context switch.
2191 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2192 * and do any other architecture-specific cleanup actions.
2194 * Note that we may have delayed dropping an mm in context_switch(). If
2195 * so, we finish that here outside of the runqueue lock. (Doing it
2196 * with the lock held can cause deadlocks; see schedule() for
2199 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2200 __releases(rq->lock)
2202 struct mm_struct *mm = rq->prev_mm;
2208 * A task struct has one reference for the use as "current".
2209 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2210 * schedule one last time. The schedule call will never return, and
2211 * the scheduled task must drop that reference.
2212 * The test for TASK_DEAD must occur while the runqueue locks are
2213 * still held, otherwise prev could be scheduled on another cpu, die
2214 * there before we look at prev->state, and then the reference would
2216 * Manfred Spraul <manfred@colorfullife.com>
2218 prev_state = prev->state;
2219 vtime_task_switch(prev);
2220 finish_arch_switch(prev);
2221 perf_event_task_sched_in(prev, current);
2222 finish_lock_switch(rq, prev);
2223 finish_arch_post_lock_switch();
2225 fire_sched_in_preempt_notifiers(current);
2228 if (unlikely(prev_state == TASK_DEAD)) {
2229 if (prev->sched_class->task_dead)
2230 prev->sched_class->task_dead(prev);
2233 * Remove function-return probe instances associated with this
2234 * task and put them back on the free list.
2236 kprobe_flush_task(prev);
2237 put_task_struct(prev);
2240 tick_nohz_task_switch(current);
2245 /* rq->lock is NOT held, but preemption is disabled */
2246 static inline void post_schedule(struct rq *rq)
2248 if (rq->post_schedule) {
2249 unsigned long flags;
2251 raw_spin_lock_irqsave(&rq->lock, flags);
2252 if (rq->curr->sched_class->post_schedule)
2253 rq->curr->sched_class->post_schedule(rq);
2254 raw_spin_unlock_irqrestore(&rq->lock, flags);
2256 rq->post_schedule = 0;
2262 static inline void post_schedule(struct rq *rq)
2269 * schedule_tail - first thing a freshly forked thread must call.
2270 * @prev: the thread we just switched away from.
2272 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2273 __releases(rq->lock)
2275 struct rq *rq = this_rq();
2277 finish_task_switch(rq, prev);
2280 * FIXME: do we need to worry about rq being invalidated by the
2285 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2286 /* In this case, finish_task_switch does not reenable preemption */
2289 if (current->set_child_tid)
2290 put_user(task_pid_vnr(current), current->set_child_tid);
2294 * context_switch - switch to the new MM and the new
2295 * thread's register state.
2298 context_switch(struct rq *rq, struct task_struct *prev,
2299 struct task_struct *next)
2301 struct mm_struct *mm, *oldmm;
2303 prepare_task_switch(rq, prev, next);
2306 oldmm = prev->active_mm;
2308 * For paravirt, this is coupled with an exit in switch_to to
2309 * combine the page table reload and the switch backend into
2312 arch_start_context_switch(prev);
2315 next->active_mm = oldmm;
2316 atomic_inc(&oldmm->mm_count);
2317 enter_lazy_tlb(oldmm, next);
2319 switch_mm(oldmm, mm, next);
2322 prev->active_mm = NULL;
2323 rq->prev_mm = oldmm;
2326 * Since the runqueue lock will be released by the next
2327 * task (which is an invalid locking op but in the case
2328 * of the scheduler it's an obvious special-case), so we
2329 * do an early lockdep release here:
2331 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2332 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2335 context_tracking_task_switch(prev, next);
2336 /* Here we just switch the register state and the stack. */
2337 switch_to(prev, next, prev);
2341 * this_rq must be evaluated again because prev may have moved
2342 * CPUs since it called schedule(), thus the 'rq' on its stack
2343 * frame will be invalid.
2345 finish_task_switch(this_rq(), prev);
2349 * nr_running and nr_context_switches:
2351 * externally visible scheduler statistics: current number of runnable
2352 * threads, total number of context switches performed since bootup.
2354 unsigned long nr_running(void)
2356 unsigned long i, sum = 0;
2358 for_each_online_cpu(i)
2359 sum += cpu_rq(i)->nr_running;
2364 unsigned long long nr_context_switches(void)
2367 unsigned long long sum = 0;
2369 for_each_possible_cpu(i)
2370 sum += cpu_rq(i)->nr_switches;
2375 unsigned long nr_iowait(void)
2377 unsigned long i, sum = 0;
2379 for_each_possible_cpu(i)
2380 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2385 unsigned long nr_iowait_cpu(int cpu)
2387 struct rq *this = cpu_rq(cpu);
2388 return atomic_read(&this->nr_iowait);
2394 * sched_exec - execve() is a valuable balancing opportunity, because at
2395 * this point the task has the smallest effective memory and cache footprint.
2397 void sched_exec(void)
2399 struct task_struct *p = current;
2400 unsigned long flags;
2403 raw_spin_lock_irqsave(&p->pi_lock, flags);
2404 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2405 if (dest_cpu == smp_processor_id())
2408 if (likely(cpu_active(dest_cpu))) {
2409 struct migration_arg arg = { p, dest_cpu };
2411 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2412 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2416 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2421 DEFINE_PER_CPU(struct kernel_stat, kstat);
2422 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2424 EXPORT_PER_CPU_SYMBOL(kstat);
2425 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2428 * Return any ns on the sched_clock that have not yet been accounted in
2429 * @p in case that task is currently running.
2431 * Called with task_rq_lock() held on @rq.
2433 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2438 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2439 * project cycles that may never be accounted to this
2440 * thread, breaking clock_gettime().
2442 if (task_current(rq, p) && p->on_rq) {
2443 update_rq_clock(rq);
2444 ns = rq_clock_task(rq) - p->se.exec_start;
2452 unsigned long long task_delta_exec(struct task_struct *p)
2454 unsigned long flags;
2458 rq = task_rq_lock(p, &flags);
2459 ns = do_task_delta_exec(p, rq);
2460 task_rq_unlock(rq, p, &flags);
2466 * Return accounted runtime for the task.
2467 * In case the task is currently running, return the runtime plus current's
2468 * pending runtime that have not been accounted yet.
2470 unsigned long long task_sched_runtime(struct task_struct *p)
2472 unsigned long flags;
2476 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2478 * 64-bit doesn't need locks to atomically read a 64bit value.
2479 * So we have a optimization chance when the task's delta_exec is 0.
2480 * Reading ->on_cpu is racy, but this is ok.
2482 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2483 * If we race with it entering cpu, unaccounted time is 0. This is
2484 * indistinguishable from the read occurring a few cycles earlier.
2485 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2486 * been accounted, so we're correct here as well.
2488 if (!p->on_cpu || !p->on_rq)
2489 return p->se.sum_exec_runtime;
2492 rq = task_rq_lock(p, &flags);
2493 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2494 task_rq_unlock(rq, p, &flags);
2500 * This function gets called by the timer code, with HZ frequency.
2501 * We call it with interrupts disabled.
2503 void scheduler_tick(void)
2505 int cpu = smp_processor_id();
2506 struct rq *rq = cpu_rq(cpu);
2507 struct task_struct *curr = rq->curr;
2511 raw_spin_lock(&rq->lock);
2512 update_rq_clock(rq);
2513 curr->sched_class->task_tick(rq, curr, 0);
2514 update_cpu_load_active(rq);
2515 raw_spin_unlock(&rq->lock);
2517 perf_event_task_tick();
2520 rq->idle_balance = idle_cpu(cpu);
2521 trigger_load_balance(rq);
2523 rq_last_tick_reset(rq);
2526 #ifdef CONFIG_NO_HZ_FULL
2528 * scheduler_tick_max_deferment
2530 * Keep at least one tick per second when a single
2531 * active task is running because the scheduler doesn't
2532 * yet completely support full dynticks environment.
2534 * This makes sure that uptime, CFS vruntime, load
2535 * balancing, etc... continue to move forward, even
2536 * with a very low granularity.
2538 * Return: Maximum deferment in nanoseconds.
2540 u64 scheduler_tick_max_deferment(void)
2542 struct rq *rq = this_rq();
2543 unsigned long next, now = ACCESS_ONCE(jiffies);
2545 next = rq->last_sched_tick + HZ;
2547 if (time_before_eq(next, now))
2550 return jiffies_to_nsecs(next - now);
2554 notrace unsigned long get_parent_ip(unsigned long addr)
2556 if (in_lock_functions(addr)) {
2557 addr = CALLER_ADDR2;
2558 if (in_lock_functions(addr))
2559 addr = CALLER_ADDR3;
2564 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2565 defined(CONFIG_PREEMPT_TRACER))
2567 void preempt_count_add(int val)
2569 #ifdef CONFIG_DEBUG_PREEMPT
2573 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2576 __preempt_count_add(val);
2577 #ifdef CONFIG_DEBUG_PREEMPT
2579 * Spinlock count overflowing soon?
2581 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2584 if (preempt_count() == val) {
2585 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2586 #ifdef CONFIG_DEBUG_PREEMPT
2587 current->preempt_disable_ip = ip;
2589 trace_preempt_off(CALLER_ADDR0, ip);
2592 EXPORT_SYMBOL(preempt_count_add);
2593 NOKPROBE_SYMBOL(preempt_count_add);
2595 void preempt_count_sub(int val)
2597 #ifdef CONFIG_DEBUG_PREEMPT
2601 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2604 * Is the spinlock portion underflowing?
2606 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2607 !(preempt_count() & PREEMPT_MASK)))
2611 if (preempt_count() == val)
2612 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2613 __preempt_count_sub(val);
2615 EXPORT_SYMBOL(preempt_count_sub);
2616 NOKPROBE_SYMBOL(preempt_count_sub);
2621 * Print scheduling while atomic bug:
2623 static noinline void __schedule_bug(struct task_struct *prev)
2625 if (oops_in_progress)
2628 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2629 prev->comm, prev->pid, preempt_count());
2631 debug_show_held_locks(prev);
2633 if (irqs_disabled())
2634 print_irqtrace_events(prev);
2635 #ifdef CONFIG_DEBUG_PREEMPT
2636 if (in_atomic_preempt_off()) {
2637 pr_err("Preemption disabled at:");
2638 print_ip_sym(current->preempt_disable_ip);
2643 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2647 * Various schedule()-time debugging checks and statistics:
2649 static inline void schedule_debug(struct task_struct *prev)
2652 * Test if we are atomic. Since do_exit() needs to call into
2653 * schedule() atomically, we ignore that path. Otherwise whine
2654 * if we are scheduling when we should not.
2656 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2657 __schedule_bug(prev);
2660 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2662 schedstat_inc(this_rq(), sched_count);
2666 * Pick up the highest-prio task:
2668 static inline struct task_struct *
2669 pick_next_task(struct rq *rq, struct task_struct *prev)
2671 const struct sched_class *class = &fair_sched_class;
2672 struct task_struct *p;
2675 * Optimization: we know that if all tasks are in
2676 * the fair class we can call that function directly:
2678 if (likely(prev->sched_class == class &&
2679 rq->nr_running == rq->cfs.h_nr_running)) {
2680 p = fair_sched_class.pick_next_task(rq, prev);
2681 if (unlikely(p == RETRY_TASK))
2684 /* assumes fair_sched_class->next == idle_sched_class */
2686 p = idle_sched_class.pick_next_task(rq, prev);
2692 for_each_class(class) {
2693 p = class->pick_next_task(rq, prev);
2695 if (unlikely(p == RETRY_TASK))
2701 BUG(); /* the idle class will always have a runnable task */
2705 * __schedule() is the main scheduler function.
2707 * The main means of driving the scheduler and thus entering this function are:
2709 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2711 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2712 * paths. For example, see arch/x86/entry_64.S.
2714 * To drive preemption between tasks, the scheduler sets the flag in timer
2715 * interrupt handler scheduler_tick().
2717 * 3. Wakeups don't really cause entry into schedule(). They add a
2718 * task to the run-queue and that's it.
2720 * Now, if the new task added to the run-queue preempts the current
2721 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2722 * called on the nearest possible occasion:
2724 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2726 * - in syscall or exception context, at the next outmost
2727 * preempt_enable(). (this might be as soon as the wake_up()'s
2730 * - in IRQ context, return from interrupt-handler to
2731 * preemptible context
2733 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2736 * - cond_resched() call
2737 * - explicit schedule() call
2738 * - return from syscall or exception to user-space
2739 * - return from interrupt-handler to user-space
2741 static void __sched __schedule(void)
2743 struct task_struct *prev, *next;
2744 unsigned long *switch_count;
2750 cpu = smp_processor_id();
2752 rcu_note_context_switch(cpu);
2755 schedule_debug(prev);
2757 if (sched_feat(HRTICK))
2761 * Make sure that signal_pending_state()->signal_pending() below
2762 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2763 * done by the caller to avoid the race with signal_wake_up().
2765 smp_mb__before_spinlock();
2766 raw_spin_lock_irq(&rq->lock);
2768 switch_count = &prev->nivcsw;
2769 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2770 if (unlikely(signal_pending_state(prev->state, prev))) {
2771 prev->state = TASK_RUNNING;
2773 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2777 * If a worker went to sleep, notify and ask workqueue
2778 * whether it wants to wake up a task to maintain
2781 if (prev->flags & PF_WQ_WORKER) {
2782 struct task_struct *to_wakeup;
2784 to_wakeup = wq_worker_sleeping(prev, cpu);
2786 try_to_wake_up_local(to_wakeup);
2789 switch_count = &prev->nvcsw;
2792 if (prev->on_rq || rq->skip_clock_update < 0)
2793 update_rq_clock(rq);
2795 next = pick_next_task(rq, prev);
2796 clear_tsk_need_resched(prev);
2797 clear_preempt_need_resched();
2798 rq->skip_clock_update = 0;
2800 if (likely(prev != next)) {
2805 context_switch(rq, prev, next); /* unlocks the rq */
2807 * The context switch have flipped the stack from under us
2808 * and restored the local variables which were saved when
2809 * this task called schedule() in the past. prev == current
2810 * is still correct, but it can be moved to another cpu/rq.
2812 cpu = smp_processor_id();
2815 raw_spin_unlock_irq(&rq->lock);
2819 sched_preempt_enable_no_resched();
2824 static inline void sched_submit_work(struct task_struct *tsk)
2826 if (!tsk->state || tsk_is_pi_blocked(tsk))
2829 * If we are going to sleep and we have plugged IO queued,
2830 * make sure to submit it to avoid deadlocks.
2832 if (blk_needs_flush_plug(tsk))
2833 blk_schedule_flush_plug(tsk);
2836 asmlinkage __visible void __sched schedule(void)
2838 struct task_struct *tsk = current;
2840 sched_submit_work(tsk);
2843 EXPORT_SYMBOL(schedule);
2845 #ifdef CONFIG_CONTEXT_TRACKING
2846 asmlinkage __visible void __sched schedule_user(void)
2849 * If we come here after a random call to set_need_resched(),
2850 * or we have been woken up remotely but the IPI has not yet arrived,
2851 * we haven't yet exited the RCU idle mode. Do it here manually until
2852 * we find a better solution.
2861 * schedule_preempt_disabled - called with preemption disabled
2863 * Returns with preemption disabled. Note: preempt_count must be 1
2865 void __sched schedule_preempt_disabled(void)
2867 sched_preempt_enable_no_resched();
2872 #ifdef CONFIG_PREEMPT
2874 * this is the entry point to schedule() from in-kernel preemption
2875 * off of preempt_enable. Kernel preemptions off return from interrupt
2876 * occur there and call schedule directly.
2878 asmlinkage __visible void __sched notrace preempt_schedule(void)
2881 * If there is a non-zero preempt_count or interrupts are disabled,
2882 * we do not want to preempt the current task. Just return..
2884 if (likely(!preemptible()))
2888 __preempt_count_add(PREEMPT_ACTIVE);
2890 __preempt_count_sub(PREEMPT_ACTIVE);
2893 * Check again in case we missed a preemption opportunity
2894 * between schedule and now.
2897 } while (need_resched());
2899 NOKPROBE_SYMBOL(preempt_schedule);
2900 EXPORT_SYMBOL(preempt_schedule);
2901 #endif /* CONFIG_PREEMPT */
2904 * this is the entry point to schedule() from kernel preemption
2905 * off of irq context.
2906 * Note, that this is called and return with irqs disabled. This will
2907 * protect us against recursive calling from irq.
2909 asmlinkage __visible void __sched preempt_schedule_irq(void)
2911 enum ctx_state prev_state;
2913 /* Catch callers which need to be fixed */
2914 BUG_ON(preempt_count() || !irqs_disabled());
2916 prev_state = exception_enter();
2919 __preempt_count_add(PREEMPT_ACTIVE);
2922 local_irq_disable();
2923 __preempt_count_sub(PREEMPT_ACTIVE);
2926 * Check again in case we missed a preemption opportunity
2927 * between schedule and now.
2930 } while (need_resched());
2932 exception_exit(prev_state);
2935 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2938 return try_to_wake_up(curr->private, mode, wake_flags);
2940 EXPORT_SYMBOL(default_wake_function);
2942 #ifdef CONFIG_RT_MUTEXES
2945 * rt_mutex_setprio - set the current priority of a task
2947 * @prio: prio value (kernel-internal form)
2949 * This function changes the 'effective' priority of a task. It does
2950 * not touch ->normal_prio like __setscheduler().
2952 * Used by the rt_mutex code to implement priority inheritance
2953 * logic. Call site only calls if the priority of the task changed.
2955 void rt_mutex_setprio(struct task_struct *p, int prio)
2957 int oldprio, on_rq, running, enqueue_flag = 0;
2959 const struct sched_class *prev_class;
2961 BUG_ON(prio > MAX_PRIO);
2963 rq = __task_rq_lock(p);
2966 * Idle task boosting is a nono in general. There is one
2967 * exception, when PREEMPT_RT and NOHZ is active:
2969 * The idle task calls get_next_timer_interrupt() and holds
2970 * the timer wheel base->lock on the CPU and another CPU wants
2971 * to access the timer (probably to cancel it). We can safely
2972 * ignore the boosting request, as the idle CPU runs this code
2973 * with interrupts disabled and will complete the lock
2974 * protected section without being interrupted. So there is no
2975 * real need to boost.
2977 if (unlikely(p == rq->idle)) {
2978 WARN_ON(p != rq->curr);
2979 WARN_ON(p->pi_blocked_on);
2983 trace_sched_pi_setprio(p, prio);
2985 prev_class = p->sched_class;
2987 running = task_current(rq, p);
2989 dequeue_task(rq, p, 0);
2991 p->sched_class->put_prev_task(rq, p);
2994 * Boosting condition are:
2995 * 1. -rt task is running and holds mutex A
2996 * --> -dl task blocks on mutex A
2998 * 2. -dl task is running and holds mutex A
2999 * --> -dl task blocks on mutex A and could preempt the
3002 if (dl_prio(prio)) {
3003 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3004 if (!dl_prio(p->normal_prio) ||
3005 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3006 p->dl.dl_boosted = 1;
3007 p->dl.dl_throttled = 0;
3008 enqueue_flag = ENQUEUE_REPLENISH;
3010 p->dl.dl_boosted = 0;
3011 p->sched_class = &dl_sched_class;
3012 } else if (rt_prio(prio)) {
3013 if (dl_prio(oldprio))
3014 p->dl.dl_boosted = 0;
3016 enqueue_flag = ENQUEUE_HEAD;
3017 p->sched_class = &rt_sched_class;
3019 if (dl_prio(oldprio))
3020 p->dl.dl_boosted = 0;
3021 p->sched_class = &fair_sched_class;
3027 p->sched_class->set_curr_task(rq);
3029 enqueue_task(rq, p, enqueue_flag);
3031 check_class_changed(rq, p, prev_class, oldprio);
3033 __task_rq_unlock(rq);
3037 void set_user_nice(struct task_struct *p, long nice)
3039 int old_prio, delta, on_rq;
3040 unsigned long flags;
3043 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3046 * We have to be careful, if called from sys_setpriority(),
3047 * the task might be in the middle of scheduling on another CPU.
3049 rq = task_rq_lock(p, &flags);
3051 * The RT priorities are set via sched_setscheduler(), but we still
3052 * allow the 'normal' nice value to be set - but as expected
3053 * it wont have any effect on scheduling until the task is
3054 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3056 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3057 p->static_prio = NICE_TO_PRIO(nice);
3062 dequeue_task(rq, p, 0);
3064 p->static_prio = NICE_TO_PRIO(nice);
3067 p->prio = effective_prio(p);
3068 delta = p->prio - old_prio;
3071 enqueue_task(rq, p, 0);
3073 * If the task increased its priority or is running and
3074 * lowered its priority, then reschedule its CPU:
3076 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3080 task_rq_unlock(rq, p, &flags);
3082 EXPORT_SYMBOL(set_user_nice);
3085 * can_nice - check if a task can reduce its nice value
3089 int can_nice(const struct task_struct *p, const int nice)
3091 /* convert nice value [19,-20] to rlimit style value [1,40] */
3092 int nice_rlim = nice_to_rlimit(nice);
3094 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3095 capable(CAP_SYS_NICE));
3098 #ifdef __ARCH_WANT_SYS_NICE
3101 * sys_nice - change the priority of the current process.
3102 * @increment: priority increment
3104 * sys_setpriority is a more generic, but much slower function that
3105 * does similar things.
3107 SYSCALL_DEFINE1(nice, int, increment)
3112 * Setpriority might change our priority at the same moment.
3113 * We don't have to worry. Conceptually one call occurs first
3114 * and we have a single winner.
3116 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3117 nice = task_nice(current) + increment;
3119 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3120 if (increment < 0 && !can_nice(current, nice))
3123 retval = security_task_setnice(current, nice);
3127 set_user_nice(current, nice);
3134 * task_prio - return the priority value of a given task.
3135 * @p: the task in question.
3137 * Return: The priority value as seen by users in /proc.
3138 * RT tasks are offset by -200. Normal tasks are centered
3139 * around 0, value goes from -16 to +15.
3141 int task_prio(const struct task_struct *p)
3143 return p->prio - MAX_RT_PRIO;
3147 * idle_cpu - is a given cpu idle currently?
3148 * @cpu: the processor in question.
3150 * Return: 1 if the CPU is currently idle. 0 otherwise.
3152 int idle_cpu(int cpu)
3154 struct rq *rq = cpu_rq(cpu);
3156 if (rq->curr != rq->idle)
3163 if (!llist_empty(&rq->wake_list))
3171 * idle_task - return the idle task for a given cpu.
3172 * @cpu: the processor in question.
3174 * Return: The idle task for the cpu @cpu.
3176 struct task_struct *idle_task(int cpu)
3178 return cpu_rq(cpu)->idle;
3182 * find_process_by_pid - find a process with a matching PID value.
3183 * @pid: the pid in question.
3185 * The task of @pid, if found. %NULL otherwise.
3187 static struct task_struct *find_process_by_pid(pid_t pid)
3189 return pid ? find_task_by_vpid(pid) : current;
3193 * This function initializes the sched_dl_entity of a newly becoming
3194 * SCHED_DEADLINE task.
3196 * Only the static values are considered here, the actual runtime and the
3197 * absolute deadline will be properly calculated when the task is enqueued
3198 * for the first time with its new policy.
3201 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3203 struct sched_dl_entity *dl_se = &p->dl;
3205 init_dl_task_timer(dl_se);
3206 dl_se->dl_runtime = attr->sched_runtime;
3207 dl_se->dl_deadline = attr->sched_deadline;
3208 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3209 dl_se->flags = attr->sched_flags;
3210 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3211 dl_se->dl_throttled = 0;
3213 dl_se->dl_yielded = 0;
3216 static void __setscheduler_params(struct task_struct *p,
3217 const struct sched_attr *attr)
3219 int policy = attr->sched_policy;
3221 if (policy == -1) /* setparam */
3226 if (dl_policy(policy))
3227 __setparam_dl(p, attr);
3228 else if (fair_policy(policy))
3229 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3232 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3233 * !rt_policy. Always setting this ensures that things like
3234 * getparam()/getattr() don't report silly values for !rt tasks.
3236 p->rt_priority = attr->sched_priority;
3237 p->normal_prio = normal_prio(p);
3241 /* Actually do priority change: must hold pi & rq lock. */
3242 static void __setscheduler(struct rq *rq, struct task_struct *p,
3243 const struct sched_attr *attr)
3245 __setscheduler_params(p, attr);
3248 * If we get here, there was no pi waiters boosting the
3249 * task. It is safe to use the normal prio.
3251 p->prio = normal_prio(p);
3253 if (dl_prio(p->prio))
3254 p->sched_class = &dl_sched_class;
3255 else if (rt_prio(p->prio))
3256 p->sched_class = &rt_sched_class;
3258 p->sched_class = &fair_sched_class;
3262 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3264 struct sched_dl_entity *dl_se = &p->dl;
3266 attr->sched_priority = p->rt_priority;
3267 attr->sched_runtime = dl_se->dl_runtime;
3268 attr->sched_deadline = dl_se->dl_deadline;
3269 attr->sched_period = dl_se->dl_period;
3270 attr->sched_flags = dl_se->flags;
3274 * This function validates the new parameters of a -deadline task.
3275 * We ask for the deadline not being zero, and greater or equal
3276 * than the runtime, as well as the period of being zero or
3277 * greater than deadline. Furthermore, we have to be sure that
3278 * user parameters are above the internal resolution of 1us (we
3279 * check sched_runtime only since it is always the smaller one) and
3280 * below 2^63 ns (we have to check both sched_deadline and
3281 * sched_period, as the latter can be zero).
3284 __checkparam_dl(const struct sched_attr *attr)
3287 if (attr->sched_deadline == 0)
3291 * Since we truncate DL_SCALE bits, make sure we're at least
3294 if (attr->sched_runtime < (1ULL << DL_SCALE))
3298 * Since we use the MSB for wrap-around and sign issues, make
3299 * sure it's not set (mind that period can be equal to zero).
3301 if (attr->sched_deadline & (1ULL << 63) ||
3302 attr->sched_period & (1ULL << 63))
3305 /* runtime <= deadline <= period (if period != 0) */
3306 if ((attr->sched_period != 0 &&
3307 attr->sched_period < attr->sched_deadline) ||
3308 attr->sched_deadline < attr->sched_runtime)
3315 * check the target process has a UID that matches the current process's
3317 static bool check_same_owner(struct task_struct *p)
3319 const struct cred *cred = current_cred(), *pcred;
3323 pcred = __task_cred(p);
3324 match = (uid_eq(cred->euid, pcred->euid) ||
3325 uid_eq(cred->euid, pcred->uid));
3330 static int __sched_setscheduler(struct task_struct *p,
3331 const struct sched_attr *attr,
3334 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3335 MAX_RT_PRIO - 1 - attr->sched_priority;
3336 int retval, oldprio, oldpolicy = -1, on_rq, running;
3337 int policy = attr->sched_policy;
3338 unsigned long flags;
3339 const struct sched_class *prev_class;
3343 /* may grab non-irq protected spin_locks */
3344 BUG_ON(in_interrupt());
3346 /* double check policy once rq lock held */
3348 reset_on_fork = p->sched_reset_on_fork;
3349 policy = oldpolicy = p->policy;
3351 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3353 if (policy != SCHED_DEADLINE &&
3354 policy != SCHED_FIFO && policy != SCHED_RR &&
3355 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3356 policy != SCHED_IDLE)
3360 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3364 * Valid priorities for SCHED_FIFO and SCHED_RR are
3365 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3366 * SCHED_BATCH and SCHED_IDLE is 0.
3368 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3369 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3371 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3372 (rt_policy(policy) != (attr->sched_priority != 0)))
3376 * Allow unprivileged RT tasks to decrease priority:
3378 if (user && !capable(CAP_SYS_NICE)) {
3379 if (fair_policy(policy)) {
3380 if (attr->sched_nice < task_nice(p) &&
3381 !can_nice(p, attr->sched_nice))
3385 if (rt_policy(policy)) {
3386 unsigned long rlim_rtprio =
3387 task_rlimit(p, RLIMIT_RTPRIO);
3389 /* can't set/change the rt policy */
3390 if (policy != p->policy && !rlim_rtprio)
3393 /* can't increase priority */
3394 if (attr->sched_priority > p->rt_priority &&
3395 attr->sched_priority > rlim_rtprio)
3400 * Can't set/change SCHED_DEADLINE policy at all for now
3401 * (safest behavior); in the future we would like to allow
3402 * unprivileged DL tasks to increase their relative deadline
3403 * or reduce their runtime (both ways reducing utilization)
3405 if (dl_policy(policy))
3409 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3410 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3412 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3413 if (!can_nice(p, task_nice(p)))
3417 /* can't change other user's priorities */
3418 if (!check_same_owner(p))
3421 /* Normal users shall not reset the sched_reset_on_fork flag */
3422 if (p->sched_reset_on_fork && !reset_on_fork)
3427 retval = security_task_setscheduler(p);
3433 * make sure no PI-waiters arrive (or leave) while we are
3434 * changing the priority of the task:
3436 * To be able to change p->policy safely, the appropriate
3437 * runqueue lock must be held.
3439 rq = task_rq_lock(p, &flags);
3442 * Changing the policy of the stop threads its a very bad idea
3444 if (p == rq->stop) {
3445 task_rq_unlock(rq, p, &flags);
3450 * If not changing anything there's no need to proceed further,
3451 * but store a possible modification of reset_on_fork.
3453 if (unlikely(policy == p->policy)) {
3454 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3456 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3458 if (dl_policy(policy))
3461 p->sched_reset_on_fork = reset_on_fork;
3462 task_rq_unlock(rq, p, &flags);
3468 #ifdef CONFIG_RT_GROUP_SCHED
3470 * Do not allow realtime tasks into groups that have no runtime
3473 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3474 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3475 !task_group_is_autogroup(task_group(p))) {
3476 task_rq_unlock(rq, p, &flags);
3481 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3482 cpumask_t *span = rq->rd->span;
3485 * Don't allow tasks with an affinity mask smaller than
3486 * the entire root_domain to become SCHED_DEADLINE. We
3487 * will also fail if there's no bandwidth available.
3489 if (!cpumask_subset(span, &p->cpus_allowed) ||
3490 rq->rd->dl_bw.bw == 0) {
3491 task_rq_unlock(rq, p, &flags);
3498 /* recheck policy now with rq lock held */
3499 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3500 policy = oldpolicy = -1;
3501 task_rq_unlock(rq, p, &flags);
3506 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3507 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3510 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3511 task_rq_unlock(rq, p, &flags);
3515 p->sched_reset_on_fork = reset_on_fork;
3519 * Special case for priority boosted tasks.
3521 * If the new priority is lower or equal (user space view)
3522 * than the current (boosted) priority, we just store the new
3523 * normal parameters and do not touch the scheduler class and
3524 * the runqueue. This will be done when the task deboost
3527 if (rt_mutex_check_prio(p, newprio)) {
3528 __setscheduler_params(p, attr);
3529 task_rq_unlock(rq, p, &flags);
3534 running = task_current(rq, p);
3536 dequeue_task(rq, p, 0);
3538 p->sched_class->put_prev_task(rq, p);
3540 prev_class = p->sched_class;
3541 __setscheduler(rq, p, attr);
3544 p->sched_class->set_curr_task(rq);
3547 * We enqueue to tail when the priority of a task is
3548 * increased (user space view).
3550 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3553 check_class_changed(rq, p, prev_class, oldprio);
3554 task_rq_unlock(rq, p, &flags);
3556 rt_mutex_adjust_pi(p);
3561 static int _sched_setscheduler(struct task_struct *p, int policy,
3562 const struct sched_param *param, bool check)
3564 struct sched_attr attr = {
3565 .sched_policy = policy,
3566 .sched_priority = param->sched_priority,
3567 .sched_nice = PRIO_TO_NICE(p->static_prio),
3571 * Fixup the legacy SCHED_RESET_ON_FORK hack
3573 if (policy & SCHED_RESET_ON_FORK) {
3574 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3575 policy &= ~SCHED_RESET_ON_FORK;
3576 attr.sched_policy = policy;
3579 return __sched_setscheduler(p, &attr, check);
3582 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3583 * @p: the task in question.
3584 * @policy: new policy.
3585 * @param: structure containing the new RT priority.
3587 * Return: 0 on success. An error code otherwise.
3589 * NOTE that the task may be already dead.
3591 int sched_setscheduler(struct task_struct *p, int policy,
3592 const struct sched_param *param)
3594 return _sched_setscheduler(p, policy, param, true);
3596 EXPORT_SYMBOL_GPL(sched_setscheduler);
3598 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3600 return __sched_setscheduler(p, attr, true);
3602 EXPORT_SYMBOL_GPL(sched_setattr);
3605 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3606 * @p: the task in question.
3607 * @policy: new policy.
3608 * @param: structure containing the new RT priority.
3610 * Just like sched_setscheduler, only don't bother checking if the
3611 * current context has permission. For example, this is needed in
3612 * stop_machine(): we create temporary high priority worker threads,
3613 * but our caller might not have that capability.
3615 * Return: 0 on success. An error code otherwise.
3617 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3618 const struct sched_param *param)
3620 return _sched_setscheduler(p, policy, param, false);
3624 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3626 struct sched_param lparam;
3627 struct task_struct *p;
3630 if (!param || pid < 0)
3632 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3637 p = find_process_by_pid(pid);
3639 retval = sched_setscheduler(p, policy, &lparam);
3646 * Mimics kernel/events/core.c perf_copy_attr().
3648 static int sched_copy_attr(struct sched_attr __user *uattr,
3649 struct sched_attr *attr)
3654 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3658 * zero the full structure, so that a short copy will be nice.
3660 memset(attr, 0, sizeof(*attr));
3662 ret = get_user(size, &uattr->size);
3666 if (size > PAGE_SIZE) /* silly large */
3669 if (!size) /* abi compat */
3670 size = SCHED_ATTR_SIZE_VER0;
3672 if (size < SCHED_ATTR_SIZE_VER0)
3676 * If we're handed a bigger struct than we know of,
3677 * ensure all the unknown bits are 0 - i.e. new
3678 * user-space does not rely on any kernel feature
3679 * extensions we dont know about yet.
3681 if (size > sizeof(*attr)) {
3682 unsigned char __user *addr;
3683 unsigned char __user *end;
3686 addr = (void __user *)uattr + sizeof(*attr);
3687 end = (void __user *)uattr + size;
3689 for (; addr < end; addr++) {
3690 ret = get_user(val, addr);
3696 size = sizeof(*attr);
3699 ret = copy_from_user(attr, uattr, size);
3704 * XXX: do we want to be lenient like existing syscalls; or do we want
3705 * to be strict and return an error on out-of-bounds values?
3707 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3712 put_user(sizeof(*attr), &uattr->size);
3717 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3718 * @pid: the pid in question.
3719 * @policy: new policy.
3720 * @param: structure containing the new RT priority.
3722 * Return: 0 on success. An error code otherwise.
3724 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3725 struct sched_param __user *, param)
3727 /* negative values for policy are not valid */
3731 return do_sched_setscheduler(pid, policy, param);
3735 * sys_sched_setparam - set/change the RT priority of a thread
3736 * @pid: the pid in question.
3737 * @param: structure containing the new RT priority.
3739 * Return: 0 on success. An error code otherwise.
3741 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3743 return do_sched_setscheduler(pid, -1, param);
3747 * sys_sched_setattr - same as above, but with extended sched_attr
3748 * @pid: the pid in question.
3749 * @uattr: structure containing the extended parameters.
3750 * @flags: for future extension.
3752 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3753 unsigned int, flags)
3755 struct sched_attr attr;
3756 struct task_struct *p;
3759 if (!uattr || pid < 0 || flags)
3762 retval = sched_copy_attr(uattr, &attr);
3766 if ((int)attr.sched_policy < 0)
3771 p = find_process_by_pid(pid);
3773 retval = sched_setattr(p, &attr);
3780 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3781 * @pid: the pid in question.
3783 * Return: On success, the policy of the thread. Otherwise, a negative error
3786 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3788 struct task_struct *p;
3796 p = find_process_by_pid(pid);
3798 retval = security_task_getscheduler(p);
3801 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3808 * sys_sched_getparam - get the RT priority of a thread
3809 * @pid: the pid in question.
3810 * @param: structure containing the RT priority.
3812 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3815 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3817 struct sched_param lp = { .sched_priority = 0 };
3818 struct task_struct *p;
3821 if (!param || pid < 0)
3825 p = find_process_by_pid(pid);
3830 retval = security_task_getscheduler(p);
3834 if (task_has_rt_policy(p))
3835 lp.sched_priority = p->rt_priority;
3839 * This one might sleep, we cannot do it with a spinlock held ...
3841 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3850 static int sched_read_attr(struct sched_attr __user *uattr,
3851 struct sched_attr *attr,
3856 if (!access_ok(VERIFY_WRITE, uattr, usize))
3860 * If we're handed a smaller struct than we know of,
3861 * ensure all the unknown bits are 0 - i.e. old
3862 * user-space does not get uncomplete information.
3864 if (usize < sizeof(*attr)) {
3865 unsigned char *addr;
3868 addr = (void *)attr + usize;
3869 end = (void *)attr + sizeof(*attr);
3871 for (; addr < end; addr++) {
3879 ret = copy_to_user(uattr, attr, attr->size);
3887 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3888 * @pid: the pid in question.
3889 * @uattr: structure containing the extended parameters.
3890 * @size: sizeof(attr) for fwd/bwd comp.
3891 * @flags: for future extension.
3893 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3894 unsigned int, size, unsigned int, flags)
3896 struct sched_attr attr = {
3897 .size = sizeof(struct sched_attr),
3899 struct task_struct *p;
3902 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3903 size < SCHED_ATTR_SIZE_VER0 || flags)
3907 p = find_process_by_pid(pid);
3912 retval = security_task_getscheduler(p);
3916 attr.sched_policy = p->policy;
3917 if (p->sched_reset_on_fork)
3918 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3919 if (task_has_dl_policy(p))
3920 __getparam_dl(p, &attr);
3921 else if (task_has_rt_policy(p))
3922 attr.sched_priority = p->rt_priority;
3924 attr.sched_nice = task_nice(p);
3928 retval = sched_read_attr(uattr, &attr, size);
3936 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3938 cpumask_var_t cpus_allowed, new_mask;
3939 struct task_struct *p;
3944 p = find_process_by_pid(pid);
3950 /* Prevent p going away */
3954 if (p->flags & PF_NO_SETAFFINITY) {
3958 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3962 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3964 goto out_free_cpus_allowed;
3967 if (!check_same_owner(p)) {
3969 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3976 retval = security_task_setscheduler(p);
3981 cpuset_cpus_allowed(p, cpus_allowed);
3982 cpumask_and(new_mask, in_mask, cpus_allowed);
3985 * Since bandwidth control happens on root_domain basis,
3986 * if admission test is enabled, we only admit -deadline
3987 * tasks allowed to run on all the CPUs in the task's
3991 if (task_has_dl_policy(p)) {
3992 const struct cpumask *span = task_rq(p)->rd->span;
3994 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
4001 retval = set_cpus_allowed_ptr(p, new_mask);
4004 cpuset_cpus_allowed(p, cpus_allowed);
4005 if (!cpumask_subset(new_mask, cpus_allowed)) {
4007 * We must have raced with a concurrent cpuset
4008 * update. Just reset the cpus_allowed to the
4009 * cpuset's cpus_allowed
4011 cpumask_copy(new_mask, cpus_allowed);
4016 free_cpumask_var(new_mask);
4017 out_free_cpus_allowed:
4018 free_cpumask_var(cpus_allowed);
4024 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4025 struct cpumask *new_mask)
4027 if (len < cpumask_size())
4028 cpumask_clear(new_mask);
4029 else if (len > cpumask_size())
4030 len = cpumask_size();
4032 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4036 * sys_sched_setaffinity - set the cpu affinity of a process
4037 * @pid: pid of the process
4038 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4039 * @user_mask_ptr: user-space pointer to the new cpu mask
4041 * Return: 0 on success. An error code otherwise.
4043 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4044 unsigned long __user *, user_mask_ptr)
4046 cpumask_var_t new_mask;
4049 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4052 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4054 retval = sched_setaffinity(pid, new_mask);
4055 free_cpumask_var(new_mask);
4059 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4061 struct task_struct *p;
4062 unsigned long flags;
4068 p = find_process_by_pid(pid);
4072 retval = security_task_getscheduler(p);
4076 raw_spin_lock_irqsave(&p->pi_lock, flags);
4077 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4078 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4087 * sys_sched_getaffinity - get the cpu affinity of a process
4088 * @pid: pid of the process
4089 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4090 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4092 * Return: 0 on success. An error code otherwise.
4094 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4095 unsigned long __user *, user_mask_ptr)
4100 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4102 if (len & (sizeof(unsigned long)-1))
4105 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4108 ret = sched_getaffinity(pid, mask);
4110 size_t retlen = min_t(size_t, len, cpumask_size());
4112 if (copy_to_user(user_mask_ptr, mask, retlen))
4117 free_cpumask_var(mask);
4123 * sys_sched_yield - yield the current processor to other threads.
4125 * This function yields the current CPU to other tasks. If there are no
4126 * other threads running on this CPU then this function will return.
4130 SYSCALL_DEFINE0(sched_yield)
4132 struct rq *rq = this_rq_lock();
4134 schedstat_inc(rq, yld_count);
4135 current->sched_class->yield_task(rq);
4138 * Since we are going to call schedule() anyway, there's
4139 * no need to preempt or enable interrupts:
4141 __release(rq->lock);
4142 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4143 do_raw_spin_unlock(&rq->lock);
4144 sched_preempt_enable_no_resched();
4151 static void __cond_resched(void)
4153 __preempt_count_add(PREEMPT_ACTIVE);
4155 __preempt_count_sub(PREEMPT_ACTIVE);
4158 int __sched _cond_resched(void)
4161 if (should_resched()) {
4167 EXPORT_SYMBOL(_cond_resched);
4170 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4171 * call schedule, and on return reacquire the lock.
4173 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4174 * operations here to prevent schedule() from being called twice (once via
4175 * spin_unlock(), once by hand).
4177 int __cond_resched_lock(spinlock_t *lock)
4179 bool need_rcu_resched = rcu_should_resched();
4180 int resched = should_resched();
4183 lockdep_assert_held(lock);
4185 if (spin_needbreak(lock) || resched || need_rcu_resched) {
4189 else if (unlikely(need_rcu_resched))
4198 EXPORT_SYMBOL(__cond_resched_lock);
4200 int __sched __cond_resched_softirq(void)
4202 BUG_ON(!in_softirq());
4204 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4205 if (should_resched()) {
4213 EXPORT_SYMBOL(__cond_resched_softirq);
4216 * yield - yield the current processor to other threads.
4218 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4220 * The scheduler is at all times free to pick the calling task as the most
4221 * eligible task to run, if removing the yield() call from your code breaks
4222 * it, its already broken.
4224 * Typical broken usage is:
4229 * where one assumes that yield() will let 'the other' process run that will
4230 * make event true. If the current task is a SCHED_FIFO task that will never
4231 * happen. Never use yield() as a progress guarantee!!
4233 * If you want to use yield() to wait for something, use wait_event().
4234 * If you want to use yield() to be 'nice' for others, use cond_resched().
4235 * If you still want to use yield(), do not!
4237 void __sched yield(void)
4239 set_current_state(TASK_RUNNING);
4242 EXPORT_SYMBOL(yield);
4245 * yield_to - yield the current processor to another thread in
4246 * your thread group, or accelerate that thread toward the
4247 * processor it's on.
4249 * @preempt: whether task preemption is allowed or not
4251 * It's the caller's job to ensure that the target task struct
4252 * can't go away on us before we can do any checks.
4255 * true (>0) if we indeed boosted the target task.
4256 * false (0) if we failed to boost the target.
4257 * -ESRCH if there's no task to yield to.
4259 int __sched yield_to(struct task_struct *p, bool preempt)
4261 struct task_struct *curr = current;
4262 struct rq *rq, *p_rq;
4263 unsigned long flags;
4266 local_irq_save(flags);
4272 * If we're the only runnable task on the rq and target rq also
4273 * has only one task, there's absolutely no point in yielding.
4275 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4280 double_rq_lock(rq, p_rq);
4281 if (task_rq(p) != p_rq) {
4282 double_rq_unlock(rq, p_rq);
4286 if (!curr->sched_class->yield_to_task)
4289 if (curr->sched_class != p->sched_class)
4292 if (task_running(p_rq, p) || p->state)
4295 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4297 schedstat_inc(rq, yld_count);
4299 * Make p's CPU reschedule; pick_next_entity takes care of
4302 if (preempt && rq != p_rq)
4307 double_rq_unlock(rq, p_rq);
4309 local_irq_restore(flags);
4316 EXPORT_SYMBOL_GPL(yield_to);
4319 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4320 * that process accounting knows that this is a task in IO wait state.
4322 void __sched io_schedule(void)
4324 struct rq *rq = raw_rq();
4326 delayacct_blkio_start();
4327 atomic_inc(&rq->nr_iowait);
4328 blk_flush_plug(current);
4329 current->in_iowait = 1;
4331 current->in_iowait = 0;
4332 atomic_dec(&rq->nr_iowait);
4333 delayacct_blkio_end();
4335 EXPORT_SYMBOL(io_schedule);
4337 long __sched io_schedule_timeout(long timeout)
4339 struct rq *rq = raw_rq();
4342 delayacct_blkio_start();
4343 atomic_inc(&rq->nr_iowait);
4344 blk_flush_plug(current);
4345 current->in_iowait = 1;
4346 ret = schedule_timeout(timeout);
4347 current->in_iowait = 0;
4348 atomic_dec(&rq->nr_iowait);
4349 delayacct_blkio_end();
4354 * sys_sched_get_priority_max - return maximum RT priority.
4355 * @policy: scheduling class.
4357 * Return: On success, this syscall returns the maximum
4358 * rt_priority that can be used by a given scheduling class.
4359 * On failure, a negative error code is returned.
4361 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4368 ret = MAX_USER_RT_PRIO-1;
4370 case SCHED_DEADLINE:
4381 * sys_sched_get_priority_min - return minimum RT priority.
4382 * @policy: scheduling class.
4384 * Return: On success, this syscall returns the minimum
4385 * rt_priority that can be used by a given scheduling class.
4386 * On failure, a negative error code is returned.
4388 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4397 case SCHED_DEADLINE:
4407 * sys_sched_rr_get_interval - return the default timeslice of a process.
4408 * @pid: pid of the process.
4409 * @interval: userspace pointer to the timeslice value.
4411 * this syscall writes the default timeslice value of a given process
4412 * into the user-space timespec buffer. A value of '0' means infinity.
4414 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4417 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4418 struct timespec __user *, interval)
4420 struct task_struct *p;
4421 unsigned int time_slice;
4422 unsigned long flags;
4432 p = find_process_by_pid(pid);
4436 retval = security_task_getscheduler(p);
4440 rq = task_rq_lock(p, &flags);
4442 if (p->sched_class->get_rr_interval)
4443 time_slice = p->sched_class->get_rr_interval(rq, p);
4444 task_rq_unlock(rq, p, &flags);
4447 jiffies_to_timespec(time_slice, &t);
4448 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4456 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4458 void sched_show_task(struct task_struct *p)
4460 unsigned long free = 0;
4464 state = p->state ? __ffs(p->state) + 1 : 0;
4465 printk(KERN_INFO "%-15.15s %c", p->comm,
4466 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4467 #if BITS_PER_LONG == 32
4468 if (state == TASK_RUNNING)
4469 printk(KERN_CONT " running ");
4471 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4473 if (state == TASK_RUNNING)
4474 printk(KERN_CONT " running task ");
4476 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4478 #ifdef CONFIG_DEBUG_STACK_USAGE
4479 free = stack_not_used(p);
4482 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4484 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4485 task_pid_nr(p), ppid,
4486 (unsigned long)task_thread_info(p)->flags);
4488 print_worker_info(KERN_INFO, p);
4489 show_stack(p, NULL);
4492 void show_state_filter(unsigned long state_filter)
4494 struct task_struct *g, *p;
4496 #if BITS_PER_LONG == 32
4498 " task PC stack pid father\n");
4501 " task PC stack pid father\n");
4504 do_each_thread(g, p) {
4506 * reset the NMI-timeout, listing all files on a slow
4507 * console might take a lot of time:
4509 touch_nmi_watchdog();
4510 if (!state_filter || (p->state & state_filter))
4512 } while_each_thread(g, p);
4514 touch_all_softlockup_watchdogs();
4516 #ifdef CONFIG_SCHED_DEBUG
4517 sysrq_sched_debug_show();
4521 * Only show locks if all tasks are dumped:
4524 debug_show_all_locks();
4527 void init_idle_bootup_task(struct task_struct *idle)
4529 idle->sched_class = &idle_sched_class;
4533 * init_idle - set up an idle thread for a given CPU
4534 * @idle: task in question
4535 * @cpu: cpu the idle task belongs to
4537 * NOTE: this function does not set the idle thread's NEED_RESCHED
4538 * flag, to make booting more robust.
4540 void init_idle(struct task_struct *idle, int cpu)
4542 struct rq *rq = cpu_rq(cpu);
4543 unsigned long flags;
4545 raw_spin_lock_irqsave(&rq->lock, flags);
4547 __sched_fork(0, idle);
4548 idle->state = TASK_RUNNING;
4549 idle->se.exec_start = sched_clock();
4551 do_set_cpus_allowed(idle, cpumask_of(cpu));
4553 * We're having a chicken and egg problem, even though we are
4554 * holding rq->lock, the cpu isn't yet set to this cpu so the
4555 * lockdep check in task_group() will fail.
4557 * Similar case to sched_fork(). / Alternatively we could
4558 * use task_rq_lock() here and obtain the other rq->lock.
4563 __set_task_cpu(idle, cpu);
4566 rq->curr = rq->idle = idle;
4568 #if defined(CONFIG_SMP)
4571 raw_spin_unlock_irqrestore(&rq->lock, flags);
4573 /* Set the preempt count _outside_ the spinlocks! */
4574 init_idle_preempt_count(idle, cpu);
4577 * The idle tasks have their own, simple scheduling class:
4579 idle->sched_class = &idle_sched_class;
4580 ftrace_graph_init_idle_task(idle, cpu);
4581 vtime_init_idle(idle, cpu);
4582 #if defined(CONFIG_SMP)
4583 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4588 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4590 if (p->sched_class && p->sched_class->set_cpus_allowed)
4591 p->sched_class->set_cpus_allowed(p, new_mask);
4593 cpumask_copy(&p->cpus_allowed, new_mask);
4594 p->nr_cpus_allowed = cpumask_weight(new_mask);
4598 * This is how migration works:
4600 * 1) we invoke migration_cpu_stop() on the target CPU using
4602 * 2) stopper starts to run (implicitly forcing the migrated thread
4604 * 3) it checks whether the migrated task is still in the wrong runqueue.
4605 * 4) if it's in the wrong runqueue then the migration thread removes
4606 * it and puts it into the right queue.
4607 * 5) stopper completes and stop_one_cpu() returns and the migration
4612 * Change a given task's CPU affinity. Migrate the thread to a
4613 * proper CPU and schedule it away if the CPU it's executing on
4614 * is removed from the allowed bitmask.
4616 * NOTE: the caller must have a valid reference to the task, the
4617 * task must not exit() & deallocate itself prematurely. The
4618 * call is not atomic; no spinlocks may be held.
4620 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4622 unsigned long flags;
4624 unsigned int dest_cpu;
4627 rq = task_rq_lock(p, &flags);
4629 if (cpumask_equal(&p->cpus_allowed, new_mask))
4632 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4637 do_set_cpus_allowed(p, new_mask);
4639 /* Can the task run on the task's current CPU? If so, we're done */
4640 if (cpumask_test_cpu(task_cpu(p), new_mask))
4643 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4645 struct migration_arg arg = { p, dest_cpu };
4646 /* Need help from migration thread: drop lock and wait. */
4647 task_rq_unlock(rq, p, &flags);
4648 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4649 tlb_migrate_finish(p->mm);
4653 task_rq_unlock(rq, p, &flags);
4657 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4660 * Move (not current) task off this cpu, onto dest cpu. We're doing
4661 * this because either it can't run here any more (set_cpus_allowed()
4662 * away from this CPU, or CPU going down), or because we're
4663 * attempting to rebalance this task on exec (sched_exec).
4665 * So we race with normal scheduler movements, but that's OK, as long
4666 * as the task is no longer on this CPU.
4668 * Returns non-zero if task was successfully migrated.
4670 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4672 struct rq *rq_dest, *rq_src;
4675 if (unlikely(!cpu_active(dest_cpu)))
4678 rq_src = cpu_rq(src_cpu);
4679 rq_dest = cpu_rq(dest_cpu);
4681 raw_spin_lock(&p->pi_lock);
4682 double_rq_lock(rq_src, rq_dest);
4683 /* Already moved. */
4684 if (task_cpu(p) != src_cpu)
4686 /* Affinity changed (again). */
4687 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4691 * If we're not on a rq, the next wake-up will ensure we're
4695 dequeue_task(rq_src, p, 0);
4696 set_task_cpu(p, dest_cpu);
4697 enqueue_task(rq_dest, p, 0);
4698 check_preempt_curr(rq_dest, p, 0);
4703 double_rq_unlock(rq_src, rq_dest);
4704 raw_spin_unlock(&p->pi_lock);
4708 #ifdef CONFIG_NUMA_BALANCING
4709 /* Migrate current task p to target_cpu */
4710 int migrate_task_to(struct task_struct *p, int target_cpu)
4712 struct migration_arg arg = { p, target_cpu };
4713 int curr_cpu = task_cpu(p);
4715 if (curr_cpu == target_cpu)
4718 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4721 /* TODO: This is not properly updating schedstats */
4723 trace_sched_move_numa(p, curr_cpu, target_cpu);
4724 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4728 * Requeue a task on a given node and accurately track the number of NUMA
4729 * tasks on the runqueues
4731 void sched_setnuma(struct task_struct *p, int nid)
4734 unsigned long flags;
4735 bool on_rq, running;
4737 rq = task_rq_lock(p, &flags);
4739 running = task_current(rq, p);
4742 dequeue_task(rq, p, 0);
4744 p->sched_class->put_prev_task(rq, p);
4746 p->numa_preferred_nid = nid;
4749 p->sched_class->set_curr_task(rq);
4751 enqueue_task(rq, p, 0);
4752 task_rq_unlock(rq, p, &flags);
4757 * migration_cpu_stop - this will be executed by a highprio stopper thread
4758 * and performs thread migration by bumping thread off CPU then
4759 * 'pushing' onto another runqueue.
4761 static int migration_cpu_stop(void *data)
4763 struct migration_arg *arg = data;
4766 * The original target cpu might have gone down and we might
4767 * be on another cpu but it doesn't matter.
4769 local_irq_disable();
4770 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4775 #ifdef CONFIG_HOTPLUG_CPU
4778 * Ensures that the idle task is using init_mm right before its cpu goes
4781 void idle_task_exit(void)
4783 struct mm_struct *mm = current->active_mm;
4785 BUG_ON(cpu_online(smp_processor_id()));
4787 if (mm != &init_mm) {
4788 switch_mm(mm, &init_mm, current);
4789 finish_arch_post_lock_switch();
4795 * Since this CPU is going 'away' for a while, fold any nr_active delta
4796 * we might have. Assumes we're called after migrate_tasks() so that the
4797 * nr_active count is stable.
4799 * Also see the comment "Global load-average calculations".
4801 static void calc_load_migrate(struct rq *rq)
4803 long delta = calc_load_fold_active(rq);
4805 atomic_long_add(delta, &calc_load_tasks);
4808 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4812 static const struct sched_class fake_sched_class = {
4813 .put_prev_task = put_prev_task_fake,
4816 static struct task_struct fake_task = {
4818 * Avoid pull_{rt,dl}_task()
4820 .prio = MAX_PRIO + 1,
4821 .sched_class = &fake_sched_class,
4825 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4826 * try_to_wake_up()->select_task_rq().
4828 * Called with rq->lock held even though we'er in stop_machine() and
4829 * there's no concurrency possible, we hold the required locks anyway
4830 * because of lock validation efforts.
4832 static void migrate_tasks(unsigned int dead_cpu)
4834 struct rq *rq = cpu_rq(dead_cpu);
4835 struct task_struct *next, *stop = rq->stop;
4839 * Fudge the rq selection such that the below task selection loop
4840 * doesn't get stuck on the currently eligible stop task.
4842 * We're currently inside stop_machine() and the rq is either stuck
4843 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4844 * either way we should never end up calling schedule() until we're
4850 * put_prev_task() and pick_next_task() sched
4851 * class method both need to have an up-to-date
4852 * value of rq->clock[_task]
4854 update_rq_clock(rq);
4858 * There's this thread running, bail when that's the only
4861 if (rq->nr_running == 1)
4864 next = pick_next_task(rq, &fake_task);
4866 next->sched_class->put_prev_task(rq, next);
4868 /* Find suitable destination for @next, with force if needed. */
4869 dest_cpu = select_fallback_rq(dead_cpu, next);
4870 raw_spin_unlock(&rq->lock);
4872 __migrate_task(next, dead_cpu, dest_cpu);
4874 raw_spin_lock(&rq->lock);
4880 #endif /* CONFIG_HOTPLUG_CPU */
4882 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4884 static struct ctl_table sd_ctl_dir[] = {
4886 .procname = "sched_domain",
4892 static struct ctl_table sd_ctl_root[] = {
4894 .procname = "kernel",
4896 .child = sd_ctl_dir,
4901 static struct ctl_table *sd_alloc_ctl_entry(int n)
4903 struct ctl_table *entry =
4904 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4909 static void sd_free_ctl_entry(struct ctl_table **tablep)
4911 struct ctl_table *entry;
4914 * In the intermediate directories, both the child directory and
4915 * procname are dynamically allocated and could fail but the mode
4916 * will always be set. In the lowest directory the names are
4917 * static strings and all have proc handlers.
4919 for (entry = *tablep; entry->mode; entry++) {
4921 sd_free_ctl_entry(&entry->child);
4922 if (entry->proc_handler == NULL)
4923 kfree(entry->procname);
4930 static int min_load_idx = 0;
4931 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4934 set_table_entry(struct ctl_table *entry,
4935 const char *procname, void *data, int maxlen,
4936 umode_t mode, proc_handler *proc_handler,
4939 entry->procname = procname;
4941 entry->maxlen = maxlen;
4943 entry->proc_handler = proc_handler;
4946 entry->extra1 = &min_load_idx;
4947 entry->extra2 = &max_load_idx;
4951 static struct ctl_table *
4952 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4954 struct ctl_table *table = sd_alloc_ctl_entry(14);
4959 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4960 sizeof(long), 0644, proc_doulongvec_minmax, false);
4961 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4962 sizeof(long), 0644, proc_doulongvec_minmax, false);
4963 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4964 sizeof(int), 0644, proc_dointvec_minmax, true);
4965 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4966 sizeof(int), 0644, proc_dointvec_minmax, true);
4967 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4968 sizeof(int), 0644, proc_dointvec_minmax, true);
4969 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4970 sizeof(int), 0644, proc_dointvec_minmax, true);
4971 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4972 sizeof(int), 0644, proc_dointvec_minmax, true);
4973 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4974 sizeof(int), 0644, proc_dointvec_minmax, false);
4975 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4976 sizeof(int), 0644, proc_dointvec_minmax, false);
4977 set_table_entry(&table[9], "cache_nice_tries",
4978 &sd->cache_nice_tries,
4979 sizeof(int), 0644, proc_dointvec_minmax, false);
4980 set_table_entry(&table[10], "flags", &sd->flags,
4981 sizeof(int), 0644, proc_dointvec_minmax, false);
4982 set_table_entry(&table[11], "max_newidle_lb_cost",
4983 &sd->max_newidle_lb_cost,
4984 sizeof(long), 0644, proc_doulongvec_minmax, false);
4985 set_table_entry(&table[12], "name", sd->name,
4986 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4987 /* &table[13] is terminator */
4992 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4994 struct ctl_table *entry, *table;
4995 struct sched_domain *sd;
4996 int domain_num = 0, i;
4999 for_each_domain(cpu, sd)
5001 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5006 for_each_domain(cpu, sd) {
5007 snprintf(buf, 32, "domain%d", i);
5008 entry->procname = kstrdup(buf, GFP_KERNEL);
5010 entry->child = sd_alloc_ctl_domain_table(sd);
5017 static struct ctl_table_header *sd_sysctl_header;
5018 static void register_sched_domain_sysctl(void)
5020 int i, cpu_num = num_possible_cpus();
5021 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5024 WARN_ON(sd_ctl_dir[0].child);
5025 sd_ctl_dir[0].child = entry;
5030 for_each_possible_cpu(i) {
5031 snprintf(buf, 32, "cpu%d", i);
5032 entry->procname = kstrdup(buf, GFP_KERNEL);
5034 entry->child = sd_alloc_ctl_cpu_table(i);
5038 WARN_ON(sd_sysctl_header);
5039 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5042 /* may be called multiple times per register */
5043 static void unregister_sched_domain_sysctl(void)
5045 if (sd_sysctl_header)
5046 unregister_sysctl_table(sd_sysctl_header);
5047 sd_sysctl_header = NULL;
5048 if (sd_ctl_dir[0].child)
5049 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5052 static void register_sched_domain_sysctl(void)
5055 static void unregister_sched_domain_sysctl(void)
5060 static void set_rq_online(struct rq *rq)
5063 const struct sched_class *class;
5065 cpumask_set_cpu(rq->cpu, rq->rd->online);
5068 for_each_class(class) {
5069 if (class->rq_online)
5070 class->rq_online(rq);
5075 static void set_rq_offline(struct rq *rq)
5078 const struct sched_class *class;
5080 for_each_class(class) {
5081 if (class->rq_offline)
5082 class->rq_offline(rq);
5085 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5091 * migration_call - callback that gets triggered when a CPU is added.
5092 * Here we can start up the necessary migration thread for the new CPU.
5095 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5097 int cpu = (long)hcpu;
5098 unsigned long flags;
5099 struct rq *rq = cpu_rq(cpu);
5101 switch (action & ~CPU_TASKS_FROZEN) {
5103 case CPU_UP_PREPARE:
5104 rq->calc_load_update = calc_load_update;
5108 /* Update our root-domain */
5109 raw_spin_lock_irqsave(&rq->lock, flags);
5111 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5115 raw_spin_unlock_irqrestore(&rq->lock, flags);
5118 #ifdef CONFIG_HOTPLUG_CPU
5120 sched_ttwu_pending();
5121 /* Update our root-domain */
5122 raw_spin_lock_irqsave(&rq->lock, flags);
5124 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5128 BUG_ON(rq->nr_running != 1); /* the migration thread */
5129 raw_spin_unlock_irqrestore(&rq->lock, flags);
5133 calc_load_migrate(rq);
5138 update_max_interval();
5144 * Register at high priority so that task migration (migrate_all_tasks)
5145 * happens before everything else. This has to be lower priority than
5146 * the notifier in the perf_event subsystem, though.
5148 static struct notifier_block migration_notifier = {
5149 .notifier_call = migration_call,
5150 .priority = CPU_PRI_MIGRATION,
5153 static void __cpuinit set_cpu_rq_start_time(void)
5155 int cpu = smp_processor_id();
5156 struct rq *rq = cpu_rq(cpu);
5157 rq->age_stamp = sched_clock_cpu(cpu);
5160 static int sched_cpu_active(struct notifier_block *nfb,
5161 unsigned long action, void *hcpu)
5163 switch (action & ~CPU_TASKS_FROZEN) {
5165 set_cpu_rq_start_time();
5167 case CPU_DOWN_FAILED:
5168 set_cpu_active((long)hcpu, true);
5175 static int sched_cpu_inactive(struct notifier_block *nfb,
5176 unsigned long action, void *hcpu)
5178 unsigned long flags;
5179 long cpu = (long)hcpu;
5181 switch (action & ~CPU_TASKS_FROZEN) {
5182 case CPU_DOWN_PREPARE:
5183 set_cpu_active(cpu, false);
5185 /* explicitly allow suspend */
5186 if (!(action & CPU_TASKS_FROZEN)) {
5187 struct dl_bw *dl_b = dl_bw_of(cpu);
5191 raw_spin_lock_irqsave(&dl_b->lock, flags);
5192 cpus = dl_bw_cpus(cpu);
5193 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5194 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5197 return notifier_from_errno(-EBUSY);
5205 static int __init migration_init(void)
5207 void *cpu = (void *)(long)smp_processor_id();
5210 /* Initialize migration for the boot CPU */
5211 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5212 BUG_ON(err == NOTIFY_BAD);
5213 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5214 register_cpu_notifier(&migration_notifier);
5216 /* Register cpu active notifiers */
5217 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5218 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5222 early_initcall(migration_init);
5227 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5229 #ifdef CONFIG_SCHED_DEBUG
5231 static __read_mostly int sched_debug_enabled;
5233 static int __init sched_debug_setup(char *str)
5235 sched_debug_enabled = 1;
5239 early_param("sched_debug", sched_debug_setup);
5241 static inline bool sched_debug(void)
5243 return sched_debug_enabled;
5246 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5247 struct cpumask *groupmask)
5249 struct sched_group *group = sd->groups;
5252 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5253 cpumask_clear(groupmask);
5255 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5257 if (!(sd->flags & SD_LOAD_BALANCE)) {
5258 printk("does not load-balance\n");
5260 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5265 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5267 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5268 printk(KERN_ERR "ERROR: domain->span does not contain "
5271 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5272 printk(KERN_ERR "ERROR: domain->groups does not contain"
5276 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5280 printk(KERN_ERR "ERROR: group is NULL\n");
5285 * Even though we initialize ->capacity to something semi-sane,
5286 * we leave capacity_orig unset. This allows us to detect if
5287 * domain iteration is still funny without causing /0 traps.
5289 if (!group->sgc->capacity_orig) {
5290 printk(KERN_CONT "\n");
5291 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5295 if (!cpumask_weight(sched_group_cpus(group))) {
5296 printk(KERN_CONT "\n");
5297 printk(KERN_ERR "ERROR: empty group\n");
5301 if (!(sd->flags & SD_OVERLAP) &&
5302 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5303 printk(KERN_CONT "\n");
5304 printk(KERN_ERR "ERROR: repeated CPUs\n");
5308 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5310 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5312 printk(KERN_CONT " %s", str);
5313 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5314 printk(KERN_CONT " (cpu_capacity = %d)",
5315 group->sgc->capacity);
5318 group = group->next;
5319 } while (group != sd->groups);
5320 printk(KERN_CONT "\n");
5322 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5323 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5326 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5327 printk(KERN_ERR "ERROR: parent span is not a superset "
5328 "of domain->span\n");
5332 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5336 if (!sched_debug_enabled)
5340 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5344 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5347 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5355 #else /* !CONFIG_SCHED_DEBUG */
5356 # define sched_domain_debug(sd, cpu) do { } while (0)
5357 static inline bool sched_debug(void)
5361 #endif /* CONFIG_SCHED_DEBUG */
5363 static int sd_degenerate(struct sched_domain *sd)
5365 if (cpumask_weight(sched_domain_span(sd)) == 1)
5368 /* Following flags need at least 2 groups */
5369 if (sd->flags & (SD_LOAD_BALANCE |
5370 SD_BALANCE_NEWIDLE |
5373 SD_SHARE_CPUCAPACITY |
5374 SD_SHARE_PKG_RESOURCES |
5375 SD_SHARE_POWERDOMAIN)) {
5376 if (sd->groups != sd->groups->next)
5380 /* Following flags don't use groups */
5381 if (sd->flags & (SD_WAKE_AFFINE))
5388 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5390 unsigned long cflags = sd->flags, pflags = parent->flags;
5392 if (sd_degenerate(parent))
5395 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5398 /* Flags needing groups don't count if only 1 group in parent */
5399 if (parent->groups == parent->groups->next) {
5400 pflags &= ~(SD_LOAD_BALANCE |
5401 SD_BALANCE_NEWIDLE |
5404 SD_SHARE_CPUCAPACITY |
5405 SD_SHARE_PKG_RESOURCES |
5407 SD_SHARE_POWERDOMAIN);
5408 if (nr_node_ids == 1)
5409 pflags &= ~SD_SERIALIZE;
5411 if (~cflags & pflags)
5417 static void free_rootdomain(struct rcu_head *rcu)
5419 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5421 cpupri_cleanup(&rd->cpupri);
5422 cpudl_cleanup(&rd->cpudl);
5423 free_cpumask_var(rd->dlo_mask);
5424 free_cpumask_var(rd->rto_mask);
5425 free_cpumask_var(rd->online);
5426 free_cpumask_var(rd->span);
5430 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5432 struct root_domain *old_rd = NULL;
5433 unsigned long flags;
5435 raw_spin_lock_irqsave(&rq->lock, flags);
5440 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5443 cpumask_clear_cpu(rq->cpu, old_rd->span);
5446 * If we dont want to free the old_rd yet then
5447 * set old_rd to NULL to skip the freeing later
5450 if (!atomic_dec_and_test(&old_rd->refcount))
5454 atomic_inc(&rd->refcount);
5457 cpumask_set_cpu(rq->cpu, rd->span);
5458 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5461 raw_spin_unlock_irqrestore(&rq->lock, flags);
5464 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5467 static int init_rootdomain(struct root_domain *rd)
5469 memset(rd, 0, sizeof(*rd));
5471 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5473 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5475 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5477 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5480 init_dl_bw(&rd->dl_bw);
5481 if (cpudl_init(&rd->cpudl) != 0)
5484 if (cpupri_init(&rd->cpupri) != 0)
5489 free_cpumask_var(rd->rto_mask);
5491 free_cpumask_var(rd->dlo_mask);
5493 free_cpumask_var(rd->online);
5495 free_cpumask_var(rd->span);
5501 * By default the system creates a single root-domain with all cpus as
5502 * members (mimicking the global state we have today).
5504 struct root_domain def_root_domain;
5506 static void init_defrootdomain(void)
5508 init_rootdomain(&def_root_domain);
5510 atomic_set(&def_root_domain.refcount, 1);
5513 static struct root_domain *alloc_rootdomain(void)
5515 struct root_domain *rd;
5517 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5521 if (init_rootdomain(rd) != 0) {
5529 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5531 struct sched_group *tmp, *first;
5540 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5545 } while (sg != first);
5548 static void free_sched_domain(struct rcu_head *rcu)
5550 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5553 * If its an overlapping domain it has private groups, iterate and
5556 if (sd->flags & SD_OVERLAP) {
5557 free_sched_groups(sd->groups, 1);
5558 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5559 kfree(sd->groups->sgc);
5565 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5567 call_rcu(&sd->rcu, free_sched_domain);
5570 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5572 for (; sd; sd = sd->parent)
5573 destroy_sched_domain(sd, cpu);
5577 * Keep a special pointer to the highest sched_domain that has
5578 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5579 * allows us to avoid some pointer chasing select_idle_sibling().
5581 * Also keep a unique ID per domain (we use the first cpu number in
5582 * the cpumask of the domain), this allows us to quickly tell if
5583 * two cpus are in the same cache domain, see cpus_share_cache().
5585 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5586 DEFINE_PER_CPU(int, sd_llc_size);
5587 DEFINE_PER_CPU(int, sd_llc_id);
5588 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5589 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5590 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5592 static void update_top_cache_domain(int cpu)
5594 struct sched_domain *sd;
5595 struct sched_domain *busy_sd = NULL;
5599 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5601 id = cpumask_first(sched_domain_span(sd));
5602 size = cpumask_weight(sched_domain_span(sd));
5603 busy_sd = sd->parent; /* sd_busy */
5605 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5607 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5608 per_cpu(sd_llc_size, cpu) = size;
5609 per_cpu(sd_llc_id, cpu) = id;
5611 sd = lowest_flag_domain(cpu, SD_NUMA);
5612 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5614 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5615 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5619 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5620 * hold the hotplug lock.
5623 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5625 struct rq *rq = cpu_rq(cpu);
5626 struct sched_domain *tmp;
5628 /* Remove the sched domains which do not contribute to scheduling. */
5629 for (tmp = sd; tmp; ) {
5630 struct sched_domain *parent = tmp->parent;
5634 if (sd_parent_degenerate(tmp, parent)) {
5635 tmp->parent = parent->parent;
5637 parent->parent->child = tmp;
5639 * Transfer SD_PREFER_SIBLING down in case of a
5640 * degenerate parent; the spans match for this
5641 * so the property transfers.
5643 if (parent->flags & SD_PREFER_SIBLING)
5644 tmp->flags |= SD_PREFER_SIBLING;
5645 destroy_sched_domain(parent, cpu);
5650 if (sd && sd_degenerate(sd)) {
5653 destroy_sched_domain(tmp, cpu);
5658 sched_domain_debug(sd, cpu);
5660 rq_attach_root(rq, rd);
5662 rcu_assign_pointer(rq->sd, sd);
5663 destroy_sched_domains(tmp, cpu);
5665 update_top_cache_domain(cpu);
5668 /* cpus with isolated domains */
5669 static cpumask_var_t cpu_isolated_map;
5671 /* Setup the mask of cpus configured for isolated domains */
5672 static int __init isolated_cpu_setup(char *str)
5674 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5675 cpulist_parse(str, cpu_isolated_map);
5679 __setup("isolcpus=", isolated_cpu_setup);
5682 struct sched_domain ** __percpu sd;
5683 struct root_domain *rd;
5694 * Build an iteration mask that can exclude certain CPUs from the upwards
5697 * Asymmetric node setups can result in situations where the domain tree is of
5698 * unequal depth, make sure to skip domains that already cover the entire
5701 * In that case build_sched_domains() will have terminated the iteration early
5702 * and our sibling sd spans will be empty. Domains should always include the
5703 * cpu they're built on, so check that.
5706 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5708 const struct cpumask *span = sched_domain_span(sd);
5709 struct sd_data *sdd = sd->private;
5710 struct sched_domain *sibling;
5713 for_each_cpu(i, span) {
5714 sibling = *per_cpu_ptr(sdd->sd, i);
5715 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5718 cpumask_set_cpu(i, sched_group_mask(sg));
5723 * Return the canonical balance cpu for this group, this is the first cpu
5724 * of this group that's also in the iteration mask.
5726 int group_balance_cpu(struct sched_group *sg)
5728 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5732 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5734 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5735 const struct cpumask *span = sched_domain_span(sd);
5736 struct cpumask *covered = sched_domains_tmpmask;
5737 struct sd_data *sdd = sd->private;
5738 struct sched_domain *child;
5741 cpumask_clear(covered);
5743 for_each_cpu(i, span) {
5744 struct cpumask *sg_span;
5746 if (cpumask_test_cpu(i, covered))
5749 child = *per_cpu_ptr(sdd->sd, i);
5751 /* See the comment near build_group_mask(). */
5752 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5755 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5756 GFP_KERNEL, cpu_to_node(cpu));
5761 sg_span = sched_group_cpus(sg);
5763 child = child->child;
5764 cpumask_copy(sg_span, sched_domain_span(child));
5766 cpumask_set_cpu(i, sg_span);
5768 cpumask_or(covered, covered, sg_span);
5770 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5771 if (atomic_inc_return(&sg->sgc->ref) == 1)
5772 build_group_mask(sd, sg);
5775 * Initialize sgc->capacity such that even if we mess up the
5776 * domains and no possible iteration will get us here, we won't
5779 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5780 sg->sgc->capacity_orig = sg->sgc->capacity;
5783 * Make sure the first group of this domain contains the
5784 * canonical balance cpu. Otherwise the sched_domain iteration
5785 * breaks. See update_sg_lb_stats().
5787 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5788 group_balance_cpu(sg) == cpu)
5798 sd->groups = groups;
5803 free_sched_groups(first, 0);
5808 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5810 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5811 struct sched_domain *child = sd->child;
5814 cpu = cpumask_first(sched_domain_span(child));
5817 *sg = *per_cpu_ptr(sdd->sg, cpu);
5818 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5819 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5826 * build_sched_groups will build a circular linked list of the groups
5827 * covered by the given span, and will set each group's ->cpumask correctly,
5828 * and ->cpu_capacity to 0.
5830 * Assumes the sched_domain tree is fully constructed
5833 build_sched_groups(struct sched_domain *sd, int cpu)
5835 struct sched_group *first = NULL, *last = NULL;
5836 struct sd_data *sdd = sd->private;
5837 const struct cpumask *span = sched_domain_span(sd);
5838 struct cpumask *covered;
5841 get_group(cpu, sdd, &sd->groups);
5842 atomic_inc(&sd->groups->ref);
5844 if (cpu != cpumask_first(span))
5847 lockdep_assert_held(&sched_domains_mutex);
5848 covered = sched_domains_tmpmask;
5850 cpumask_clear(covered);
5852 for_each_cpu(i, span) {
5853 struct sched_group *sg;
5856 if (cpumask_test_cpu(i, covered))
5859 group = get_group(i, sdd, &sg);
5860 cpumask_setall(sched_group_mask(sg));
5862 for_each_cpu(j, span) {
5863 if (get_group(j, sdd, NULL) != group)
5866 cpumask_set_cpu(j, covered);
5867 cpumask_set_cpu(j, sched_group_cpus(sg));
5882 * Initialize sched groups cpu_capacity.
5884 * cpu_capacity indicates the capacity of sched group, which is used while
5885 * distributing the load between different sched groups in a sched domain.
5886 * Typically cpu_capacity for all the groups in a sched domain will be same
5887 * unless there are asymmetries in the topology. If there are asymmetries,
5888 * group having more cpu_capacity will pickup more load compared to the
5889 * group having less cpu_capacity.
5891 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5893 struct sched_group *sg = sd->groups;
5898 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5900 } while (sg != sd->groups);
5902 if (cpu != group_balance_cpu(sg))
5905 update_group_capacity(sd, cpu);
5906 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5910 * Initializers for schedule domains
5911 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5914 static int default_relax_domain_level = -1;
5915 int sched_domain_level_max;
5917 static int __init setup_relax_domain_level(char *str)
5919 if (kstrtoint(str, 0, &default_relax_domain_level))
5920 pr_warn("Unable to set relax_domain_level\n");
5924 __setup("relax_domain_level=", setup_relax_domain_level);
5926 static void set_domain_attribute(struct sched_domain *sd,
5927 struct sched_domain_attr *attr)
5931 if (!attr || attr->relax_domain_level < 0) {
5932 if (default_relax_domain_level < 0)
5935 request = default_relax_domain_level;
5937 request = attr->relax_domain_level;
5938 if (request < sd->level) {
5939 /* turn off idle balance on this domain */
5940 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5942 /* turn on idle balance on this domain */
5943 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5947 static void __sdt_free(const struct cpumask *cpu_map);
5948 static int __sdt_alloc(const struct cpumask *cpu_map);
5950 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5951 const struct cpumask *cpu_map)
5955 if (!atomic_read(&d->rd->refcount))
5956 free_rootdomain(&d->rd->rcu); /* fall through */
5958 free_percpu(d->sd); /* fall through */
5960 __sdt_free(cpu_map); /* fall through */
5966 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5967 const struct cpumask *cpu_map)
5969 memset(d, 0, sizeof(*d));
5971 if (__sdt_alloc(cpu_map))
5972 return sa_sd_storage;
5973 d->sd = alloc_percpu(struct sched_domain *);
5975 return sa_sd_storage;
5976 d->rd = alloc_rootdomain();
5979 return sa_rootdomain;
5983 * NULL the sd_data elements we've used to build the sched_domain and
5984 * sched_group structure so that the subsequent __free_domain_allocs()
5985 * will not free the data we're using.
5987 static void claim_allocations(int cpu, struct sched_domain *sd)
5989 struct sd_data *sdd = sd->private;
5991 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5992 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5994 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5995 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5997 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
5998 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6002 static int sched_domains_numa_levels;
6003 static int *sched_domains_numa_distance;
6004 static struct cpumask ***sched_domains_numa_masks;
6005 static int sched_domains_curr_level;
6009 * SD_flags allowed in topology descriptions.
6011 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6012 * SD_SHARE_PKG_RESOURCES - describes shared caches
6013 * SD_NUMA - describes NUMA topologies
6014 * SD_SHARE_POWERDOMAIN - describes shared power domain
6017 * SD_ASYM_PACKING - describes SMT quirks
6019 #define TOPOLOGY_SD_FLAGS \
6020 (SD_SHARE_CPUCAPACITY | \
6021 SD_SHARE_PKG_RESOURCES | \
6024 SD_SHARE_POWERDOMAIN)
6026 static struct sched_domain *
6027 sd_init(struct sched_domain_topology_level *tl, int cpu)
6029 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6030 int sd_weight, sd_flags = 0;
6034 * Ugly hack to pass state to sd_numa_mask()...
6036 sched_domains_curr_level = tl->numa_level;
6039 sd_weight = cpumask_weight(tl->mask(cpu));
6042 sd_flags = (*tl->sd_flags)();
6043 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6044 "wrong sd_flags in topology description\n"))
6045 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6047 *sd = (struct sched_domain){
6048 .min_interval = sd_weight,
6049 .max_interval = 2*sd_weight,
6051 .imbalance_pct = 125,
6053 .cache_nice_tries = 0,
6060 .flags = 1*SD_LOAD_BALANCE
6061 | 1*SD_BALANCE_NEWIDLE
6066 | 0*SD_SHARE_CPUCAPACITY
6067 | 0*SD_SHARE_PKG_RESOURCES
6069 | 0*SD_PREFER_SIBLING
6074 .last_balance = jiffies,
6075 .balance_interval = sd_weight,
6077 .max_newidle_lb_cost = 0,
6078 .next_decay_max_lb_cost = jiffies,
6079 #ifdef CONFIG_SCHED_DEBUG
6085 * Convert topological properties into behaviour.
6088 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6089 sd->imbalance_pct = 110;
6090 sd->smt_gain = 1178; /* ~15% */
6092 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6093 sd->imbalance_pct = 117;
6094 sd->cache_nice_tries = 1;
6098 } else if (sd->flags & SD_NUMA) {
6099 sd->cache_nice_tries = 2;
6103 sd->flags |= SD_SERIALIZE;
6104 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6105 sd->flags &= ~(SD_BALANCE_EXEC |
6112 sd->flags |= SD_PREFER_SIBLING;
6113 sd->cache_nice_tries = 1;
6118 sd->private = &tl->data;
6124 * Topology list, bottom-up.
6126 static struct sched_domain_topology_level default_topology[] = {
6127 #ifdef CONFIG_SCHED_SMT
6128 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6130 #ifdef CONFIG_SCHED_MC
6131 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6133 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6137 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6139 #define for_each_sd_topology(tl) \
6140 for (tl = sched_domain_topology; tl->mask; tl++)
6142 void set_sched_topology(struct sched_domain_topology_level *tl)
6144 sched_domain_topology = tl;
6149 static const struct cpumask *sd_numa_mask(int cpu)
6151 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6154 static void sched_numa_warn(const char *str)
6156 static int done = false;
6164 printk(KERN_WARNING "ERROR: %s\n\n", str);
6166 for (i = 0; i < nr_node_ids; i++) {
6167 printk(KERN_WARNING " ");
6168 for (j = 0; j < nr_node_ids; j++)
6169 printk(KERN_CONT "%02d ", node_distance(i,j));
6170 printk(KERN_CONT "\n");
6172 printk(KERN_WARNING "\n");
6175 static bool find_numa_distance(int distance)
6179 if (distance == node_distance(0, 0))
6182 for (i = 0; i < sched_domains_numa_levels; i++) {
6183 if (sched_domains_numa_distance[i] == distance)
6190 static void sched_init_numa(void)
6192 int next_distance, curr_distance = node_distance(0, 0);
6193 struct sched_domain_topology_level *tl;
6197 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6198 if (!sched_domains_numa_distance)
6202 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6203 * unique distances in the node_distance() table.
6205 * Assumes node_distance(0,j) includes all distances in
6206 * node_distance(i,j) in order to avoid cubic time.
6208 next_distance = curr_distance;
6209 for (i = 0; i < nr_node_ids; i++) {
6210 for (j = 0; j < nr_node_ids; j++) {
6211 for (k = 0; k < nr_node_ids; k++) {
6212 int distance = node_distance(i, k);
6214 if (distance > curr_distance &&
6215 (distance < next_distance ||
6216 next_distance == curr_distance))
6217 next_distance = distance;
6220 * While not a strong assumption it would be nice to know
6221 * about cases where if node A is connected to B, B is not
6222 * equally connected to A.
6224 if (sched_debug() && node_distance(k, i) != distance)
6225 sched_numa_warn("Node-distance not symmetric");
6227 if (sched_debug() && i && !find_numa_distance(distance))
6228 sched_numa_warn("Node-0 not representative");
6230 if (next_distance != curr_distance) {
6231 sched_domains_numa_distance[level++] = next_distance;
6232 sched_domains_numa_levels = level;
6233 curr_distance = next_distance;
6238 * In case of sched_debug() we verify the above assumption.
6244 * 'level' contains the number of unique distances, excluding the
6245 * identity distance node_distance(i,i).
6247 * The sched_domains_numa_distance[] array includes the actual distance
6252 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6253 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6254 * the array will contain less then 'level' members. This could be
6255 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6256 * in other functions.
6258 * We reset it to 'level' at the end of this function.
6260 sched_domains_numa_levels = 0;
6262 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6263 if (!sched_domains_numa_masks)
6267 * Now for each level, construct a mask per node which contains all
6268 * cpus of nodes that are that many hops away from us.
6270 for (i = 0; i < level; i++) {
6271 sched_domains_numa_masks[i] =
6272 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6273 if (!sched_domains_numa_masks[i])
6276 for (j = 0; j < nr_node_ids; j++) {
6277 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6281 sched_domains_numa_masks[i][j] = mask;
6283 for (k = 0; k < nr_node_ids; k++) {
6284 if (node_distance(j, k) > sched_domains_numa_distance[i])
6287 cpumask_or(mask, mask, cpumask_of_node(k));
6292 /* Compute default topology size */
6293 for (i = 0; sched_domain_topology[i].mask; i++);
6295 tl = kzalloc((i + level + 1) *
6296 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6301 * Copy the default topology bits..
6303 for (i = 0; sched_domain_topology[i].mask; i++)
6304 tl[i] = sched_domain_topology[i];
6307 * .. and append 'j' levels of NUMA goodness.
6309 for (j = 0; j < level; i++, j++) {
6310 tl[i] = (struct sched_domain_topology_level){
6311 .mask = sd_numa_mask,
6312 .sd_flags = cpu_numa_flags,
6313 .flags = SDTL_OVERLAP,
6319 sched_domain_topology = tl;
6321 sched_domains_numa_levels = level;
6324 static void sched_domains_numa_masks_set(int cpu)
6327 int node = cpu_to_node(cpu);
6329 for (i = 0; i < sched_domains_numa_levels; i++) {
6330 for (j = 0; j < nr_node_ids; j++) {
6331 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6332 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6337 static void sched_domains_numa_masks_clear(int cpu)
6340 for (i = 0; i < sched_domains_numa_levels; i++) {
6341 for (j = 0; j < nr_node_ids; j++)
6342 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6347 * Update sched_domains_numa_masks[level][node] array when new cpus
6350 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6351 unsigned long action,
6354 int cpu = (long)hcpu;
6356 switch (action & ~CPU_TASKS_FROZEN) {
6358 sched_domains_numa_masks_set(cpu);
6362 sched_domains_numa_masks_clear(cpu);
6372 static inline void sched_init_numa(void)
6376 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6377 unsigned long action,
6382 #endif /* CONFIG_NUMA */
6384 static int __sdt_alloc(const struct cpumask *cpu_map)
6386 struct sched_domain_topology_level *tl;
6389 for_each_sd_topology(tl) {
6390 struct sd_data *sdd = &tl->data;
6392 sdd->sd = alloc_percpu(struct sched_domain *);
6396 sdd->sg = alloc_percpu(struct sched_group *);
6400 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6404 for_each_cpu(j, cpu_map) {
6405 struct sched_domain *sd;
6406 struct sched_group *sg;
6407 struct sched_group_capacity *sgc;
6409 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6410 GFP_KERNEL, cpu_to_node(j));
6414 *per_cpu_ptr(sdd->sd, j) = sd;
6416 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6417 GFP_KERNEL, cpu_to_node(j));
6423 *per_cpu_ptr(sdd->sg, j) = sg;
6425 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6426 GFP_KERNEL, cpu_to_node(j));
6430 *per_cpu_ptr(sdd->sgc, j) = sgc;
6437 static void __sdt_free(const struct cpumask *cpu_map)
6439 struct sched_domain_topology_level *tl;
6442 for_each_sd_topology(tl) {
6443 struct sd_data *sdd = &tl->data;
6445 for_each_cpu(j, cpu_map) {
6446 struct sched_domain *sd;
6449 sd = *per_cpu_ptr(sdd->sd, j);
6450 if (sd && (sd->flags & SD_OVERLAP))
6451 free_sched_groups(sd->groups, 0);
6452 kfree(*per_cpu_ptr(sdd->sd, j));
6456 kfree(*per_cpu_ptr(sdd->sg, j));
6458 kfree(*per_cpu_ptr(sdd->sgc, j));
6460 free_percpu(sdd->sd);
6462 free_percpu(sdd->sg);
6464 free_percpu(sdd->sgc);
6469 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6470 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6471 struct sched_domain *child, int cpu)
6473 struct sched_domain *sd = sd_init(tl, cpu);
6477 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6479 sd->level = child->level + 1;
6480 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6484 set_domain_attribute(sd, attr);
6490 * Build sched domains for a given set of cpus and attach the sched domains
6491 * to the individual cpus
6493 static int build_sched_domains(const struct cpumask *cpu_map,
6494 struct sched_domain_attr *attr)
6496 enum s_alloc alloc_state;
6497 struct sched_domain *sd;
6499 int i, ret = -ENOMEM;
6501 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6502 if (alloc_state != sa_rootdomain)
6505 /* Set up domains for cpus specified by the cpu_map. */
6506 for_each_cpu(i, cpu_map) {
6507 struct sched_domain_topology_level *tl;
6510 for_each_sd_topology(tl) {
6511 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6512 if (tl == sched_domain_topology)
6513 *per_cpu_ptr(d.sd, i) = sd;
6514 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6515 sd->flags |= SD_OVERLAP;
6516 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6521 /* Build the groups for the domains */
6522 for_each_cpu(i, cpu_map) {
6523 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6524 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6525 if (sd->flags & SD_OVERLAP) {
6526 if (build_overlap_sched_groups(sd, i))
6529 if (build_sched_groups(sd, i))
6535 /* Calculate CPU capacity for physical packages and nodes */
6536 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6537 if (!cpumask_test_cpu(i, cpu_map))
6540 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6541 claim_allocations(i, sd);
6542 init_sched_groups_capacity(i, sd);
6546 /* Attach the domains */
6548 for_each_cpu(i, cpu_map) {
6549 sd = *per_cpu_ptr(d.sd, i);
6550 cpu_attach_domain(sd, d.rd, i);
6556 __free_domain_allocs(&d, alloc_state, cpu_map);
6560 static cpumask_var_t *doms_cur; /* current sched domains */
6561 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6562 static struct sched_domain_attr *dattr_cur;
6563 /* attribues of custom domains in 'doms_cur' */
6566 * Special case: If a kmalloc of a doms_cur partition (array of
6567 * cpumask) fails, then fallback to a single sched domain,
6568 * as determined by the single cpumask fallback_doms.
6570 static cpumask_var_t fallback_doms;
6573 * arch_update_cpu_topology lets virtualized architectures update the
6574 * cpu core maps. It is supposed to return 1 if the topology changed
6575 * or 0 if it stayed the same.
6577 int __weak arch_update_cpu_topology(void)
6582 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6585 cpumask_var_t *doms;
6587 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6590 for (i = 0; i < ndoms; i++) {
6591 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6592 free_sched_domains(doms, i);
6599 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6602 for (i = 0; i < ndoms; i++)
6603 free_cpumask_var(doms[i]);
6608 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6609 * For now this just excludes isolated cpus, but could be used to
6610 * exclude other special cases in the future.
6612 static int init_sched_domains(const struct cpumask *cpu_map)
6616 arch_update_cpu_topology();
6618 doms_cur = alloc_sched_domains(ndoms_cur);
6620 doms_cur = &fallback_doms;
6621 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6622 err = build_sched_domains(doms_cur[0], NULL);
6623 register_sched_domain_sysctl();
6629 * Detach sched domains from a group of cpus specified in cpu_map
6630 * These cpus will now be attached to the NULL domain
6632 static void detach_destroy_domains(const struct cpumask *cpu_map)
6637 for_each_cpu(i, cpu_map)
6638 cpu_attach_domain(NULL, &def_root_domain, i);
6642 /* handle null as "default" */
6643 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6644 struct sched_domain_attr *new, int idx_new)
6646 struct sched_domain_attr tmp;
6653 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6654 new ? (new + idx_new) : &tmp,
6655 sizeof(struct sched_domain_attr));
6659 * Partition sched domains as specified by the 'ndoms_new'
6660 * cpumasks in the array doms_new[] of cpumasks. This compares
6661 * doms_new[] to the current sched domain partitioning, doms_cur[].
6662 * It destroys each deleted domain and builds each new domain.
6664 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6665 * The masks don't intersect (don't overlap.) We should setup one
6666 * sched domain for each mask. CPUs not in any of the cpumasks will
6667 * not be load balanced. If the same cpumask appears both in the
6668 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6671 * The passed in 'doms_new' should be allocated using
6672 * alloc_sched_domains. This routine takes ownership of it and will
6673 * free_sched_domains it when done with it. If the caller failed the
6674 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6675 * and partition_sched_domains() will fallback to the single partition
6676 * 'fallback_doms', it also forces the domains to be rebuilt.
6678 * If doms_new == NULL it will be replaced with cpu_online_mask.
6679 * ndoms_new == 0 is a special case for destroying existing domains,
6680 * and it will not create the default domain.
6682 * Call with hotplug lock held
6684 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6685 struct sched_domain_attr *dattr_new)
6690 mutex_lock(&sched_domains_mutex);
6692 /* always unregister in case we don't destroy any domains */
6693 unregister_sched_domain_sysctl();
6695 /* Let architecture update cpu core mappings. */
6696 new_topology = arch_update_cpu_topology();
6698 n = doms_new ? ndoms_new : 0;
6700 /* Destroy deleted domains */
6701 for (i = 0; i < ndoms_cur; i++) {
6702 for (j = 0; j < n && !new_topology; j++) {
6703 if (cpumask_equal(doms_cur[i], doms_new[j])
6704 && dattrs_equal(dattr_cur, i, dattr_new, j))
6707 /* no match - a current sched domain not in new doms_new[] */
6708 detach_destroy_domains(doms_cur[i]);
6714 if (doms_new == NULL) {
6716 doms_new = &fallback_doms;
6717 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6718 WARN_ON_ONCE(dattr_new);
6721 /* Build new domains */
6722 for (i = 0; i < ndoms_new; i++) {
6723 for (j = 0; j < n && !new_topology; j++) {
6724 if (cpumask_equal(doms_new[i], doms_cur[j])
6725 && dattrs_equal(dattr_new, i, dattr_cur, j))
6728 /* no match - add a new doms_new */
6729 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6734 /* Remember the new sched domains */
6735 if (doms_cur != &fallback_doms)
6736 free_sched_domains(doms_cur, ndoms_cur);
6737 kfree(dattr_cur); /* kfree(NULL) is safe */
6738 doms_cur = doms_new;
6739 dattr_cur = dattr_new;
6740 ndoms_cur = ndoms_new;
6742 register_sched_domain_sysctl();
6744 mutex_unlock(&sched_domains_mutex);
6747 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6750 * Update cpusets according to cpu_active mask. If cpusets are
6751 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6752 * around partition_sched_domains().
6754 * If we come here as part of a suspend/resume, don't touch cpusets because we
6755 * want to restore it back to its original state upon resume anyway.
6757 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6761 case CPU_ONLINE_FROZEN:
6762 case CPU_DOWN_FAILED_FROZEN:
6765 * num_cpus_frozen tracks how many CPUs are involved in suspend
6766 * resume sequence. As long as this is not the last online
6767 * operation in the resume sequence, just build a single sched
6768 * domain, ignoring cpusets.
6771 if (likely(num_cpus_frozen)) {
6772 partition_sched_domains(1, NULL, NULL);
6777 * This is the last CPU online operation. So fall through and
6778 * restore the original sched domains by considering the
6779 * cpuset configurations.
6783 case CPU_DOWN_FAILED:
6784 cpuset_update_active_cpus(true);
6792 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6796 case CPU_DOWN_PREPARE:
6797 cpuset_update_active_cpus(false);
6799 case CPU_DOWN_PREPARE_FROZEN:
6801 partition_sched_domains(1, NULL, NULL);
6809 void __init sched_init_smp(void)
6811 cpumask_var_t non_isolated_cpus;
6813 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6814 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6819 * There's no userspace yet to cause hotplug operations; hence all the
6820 * cpu masks are stable and all blatant races in the below code cannot
6823 mutex_lock(&sched_domains_mutex);
6824 init_sched_domains(cpu_active_mask);
6825 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6826 if (cpumask_empty(non_isolated_cpus))
6827 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6828 mutex_unlock(&sched_domains_mutex);
6830 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6831 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6832 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6836 /* Move init over to a non-isolated CPU */
6837 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6839 sched_init_granularity();
6840 free_cpumask_var(non_isolated_cpus);
6842 init_sched_rt_class();
6843 init_sched_dl_class();
6846 void __init sched_init_smp(void)
6848 sched_init_granularity();
6850 #endif /* CONFIG_SMP */
6852 const_debug unsigned int sysctl_timer_migration = 1;
6854 int in_sched_functions(unsigned long addr)
6856 return in_lock_functions(addr) ||
6857 (addr >= (unsigned long)__sched_text_start
6858 && addr < (unsigned long)__sched_text_end);
6861 #ifdef CONFIG_CGROUP_SCHED
6863 * Default task group.
6864 * Every task in system belongs to this group at bootup.
6866 struct task_group root_task_group;
6867 LIST_HEAD(task_groups);
6870 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6872 void __init sched_init(void)
6875 unsigned long alloc_size = 0, ptr;
6877 #ifdef CONFIG_FAIR_GROUP_SCHED
6878 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6880 #ifdef CONFIG_RT_GROUP_SCHED
6881 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6883 #ifdef CONFIG_CPUMASK_OFFSTACK
6884 alloc_size += num_possible_cpus() * cpumask_size();
6887 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6889 #ifdef CONFIG_FAIR_GROUP_SCHED
6890 root_task_group.se = (struct sched_entity **)ptr;
6891 ptr += nr_cpu_ids * sizeof(void **);
6893 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6894 ptr += nr_cpu_ids * sizeof(void **);
6896 #endif /* CONFIG_FAIR_GROUP_SCHED */
6897 #ifdef CONFIG_RT_GROUP_SCHED
6898 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6899 ptr += nr_cpu_ids * sizeof(void **);
6901 root_task_group.rt_rq = (struct rt_rq **)ptr;
6902 ptr += nr_cpu_ids * sizeof(void **);
6904 #endif /* CONFIG_RT_GROUP_SCHED */
6905 #ifdef CONFIG_CPUMASK_OFFSTACK
6906 for_each_possible_cpu(i) {
6907 per_cpu(load_balance_mask, i) = (void *)ptr;
6908 ptr += cpumask_size();
6910 #endif /* CONFIG_CPUMASK_OFFSTACK */
6913 init_rt_bandwidth(&def_rt_bandwidth,
6914 global_rt_period(), global_rt_runtime());
6915 init_dl_bandwidth(&def_dl_bandwidth,
6916 global_rt_period(), global_rt_runtime());
6919 init_defrootdomain();
6922 #ifdef CONFIG_RT_GROUP_SCHED
6923 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6924 global_rt_period(), global_rt_runtime());
6925 #endif /* CONFIG_RT_GROUP_SCHED */
6927 #ifdef CONFIG_CGROUP_SCHED
6928 list_add(&root_task_group.list, &task_groups);
6929 INIT_LIST_HEAD(&root_task_group.children);
6930 INIT_LIST_HEAD(&root_task_group.siblings);
6931 autogroup_init(&init_task);
6933 #endif /* CONFIG_CGROUP_SCHED */
6935 for_each_possible_cpu(i) {
6939 raw_spin_lock_init(&rq->lock);
6941 rq->calc_load_active = 0;
6942 rq->calc_load_update = jiffies + LOAD_FREQ;
6943 init_cfs_rq(&rq->cfs);
6944 init_rt_rq(&rq->rt, rq);
6945 init_dl_rq(&rq->dl, rq);
6946 #ifdef CONFIG_FAIR_GROUP_SCHED
6947 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6948 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6950 * How much cpu bandwidth does root_task_group get?
6952 * In case of task-groups formed thr' the cgroup filesystem, it
6953 * gets 100% of the cpu resources in the system. This overall
6954 * system cpu resource is divided among the tasks of
6955 * root_task_group and its child task-groups in a fair manner,
6956 * based on each entity's (task or task-group's) weight
6957 * (se->load.weight).
6959 * In other words, if root_task_group has 10 tasks of weight
6960 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6961 * then A0's share of the cpu resource is:
6963 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6965 * We achieve this by letting root_task_group's tasks sit
6966 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6968 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6969 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6970 #endif /* CONFIG_FAIR_GROUP_SCHED */
6972 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6973 #ifdef CONFIG_RT_GROUP_SCHED
6974 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6977 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6978 rq->cpu_load[j] = 0;
6980 rq->last_load_update_tick = jiffies;
6985 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
6986 rq->post_schedule = 0;
6987 rq->active_balance = 0;
6988 rq->next_balance = jiffies;
6993 rq->avg_idle = 2*sysctl_sched_migration_cost;
6994 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6996 INIT_LIST_HEAD(&rq->cfs_tasks);
6998 rq_attach_root(rq, &def_root_domain);
6999 #ifdef CONFIG_NO_HZ_COMMON
7002 #ifdef CONFIG_NO_HZ_FULL
7003 rq->last_sched_tick = 0;
7007 atomic_set(&rq->nr_iowait, 0);
7010 set_load_weight(&init_task);
7012 #ifdef CONFIG_PREEMPT_NOTIFIERS
7013 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7017 * The boot idle thread does lazy MMU switching as well:
7019 atomic_inc(&init_mm.mm_count);
7020 enter_lazy_tlb(&init_mm, current);
7023 * Make us the idle thread. Technically, schedule() should not be
7024 * called from this thread, however somewhere below it might be,
7025 * but because we are the idle thread, we just pick up running again
7026 * when this runqueue becomes "idle".
7028 init_idle(current, smp_processor_id());
7030 calc_load_update = jiffies + LOAD_FREQ;
7033 * During early bootup we pretend to be a normal task:
7035 current->sched_class = &fair_sched_class;
7038 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7039 /* May be allocated at isolcpus cmdline parse time */
7040 if (cpu_isolated_map == NULL)
7041 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7042 idle_thread_set_boot_cpu();
7043 set_cpu_rq_start_time();
7045 init_sched_fair_class();
7047 scheduler_running = 1;
7050 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7051 static inline int preempt_count_equals(int preempt_offset)
7053 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7055 return (nested == preempt_offset);
7058 void __might_sleep(const char *file, int line, int preempt_offset)
7060 static unsigned long prev_jiffy; /* ratelimiting */
7062 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7063 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7064 !is_idle_task(current)) ||
7065 system_state != SYSTEM_RUNNING || oops_in_progress)
7067 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7069 prev_jiffy = jiffies;
7072 "BUG: sleeping function called from invalid context at %s:%d\n",
7075 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7076 in_atomic(), irqs_disabled(),
7077 current->pid, current->comm);
7079 debug_show_held_locks(current);
7080 if (irqs_disabled())
7081 print_irqtrace_events(current);
7082 #ifdef CONFIG_DEBUG_PREEMPT
7083 if (!preempt_count_equals(preempt_offset)) {
7084 pr_err("Preemption disabled at:");
7085 print_ip_sym(current->preempt_disable_ip);
7091 EXPORT_SYMBOL(__might_sleep);
7094 #ifdef CONFIG_MAGIC_SYSRQ
7095 static void normalize_task(struct rq *rq, struct task_struct *p)
7097 const struct sched_class *prev_class = p->sched_class;
7098 struct sched_attr attr = {
7099 .sched_policy = SCHED_NORMAL,
7101 int old_prio = p->prio;
7106 dequeue_task(rq, p, 0);
7107 __setscheduler(rq, p, &attr);
7109 enqueue_task(rq, p, 0);
7113 check_class_changed(rq, p, prev_class, old_prio);
7116 void normalize_rt_tasks(void)
7118 struct task_struct *g, *p;
7119 unsigned long flags;
7122 read_lock_irqsave(&tasklist_lock, flags);
7123 do_each_thread(g, p) {
7125 * Only normalize user tasks:
7130 p->se.exec_start = 0;
7131 #ifdef CONFIG_SCHEDSTATS
7132 p->se.statistics.wait_start = 0;
7133 p->se.statistics.sleep_start = 0;
7134 p->se.statistics.block_start = 0;
7137 if (!dl_task(p) && !rt_task(p)) {
7139 * Renice negative nice level userspace
7142 if (task_nice(p) < 0 && p->mm)
7143 set_user_nice(p, 0);
7147 raw_spin_lock(&p->pi_lock);
7148 rq = __task_rq_lock(p);
7150 normalize_task(rq, p);
7152 __task_rq_unlock(rq);
7153 raw_spin_unlock(&p->pi_lock);
7154 } while_each_thread(g, p);
7156 read_unlock_irqrestore(&tasklist_lock, flags);
7159 #endif /* CONFIG_MAGIC_SYSRQ */
7161 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7163 * These functions are only useful for the IA64 MCA handling, or kdb.
7165 * They can only be called when the whole system has been
7166 * stopped - every CPU needs to be quiescent, and no scheduling
7167 * activity can take place. Using them for anything else would
7168 * be a serious bug, and as a result, they aren't even visible
7169 * under any other configuration.
7173 * curr_task - return the current task for a given cpu.
7174 * @cpu: the processor in question.
7176 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7178 * Return: The current task for @cpu.
7180 struct task_struct *curr_task(int cpu)
7182 return cpu_curr(cpu);
7185 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7189 * set_curr_task - set the current task for a given cpu.
7190 * @cpu: the processor in question.
7191 * @p: the task pointer to set.
7193 * Description: This function must only be used when non-maskable interrupts
7194 * are serviced on a separate stack. It allows the architecture to switch the
7195 * notion of the current task on a cpu in a non-blocking manner. This function
7196 * must be called with all CPU's synchronized, and interrupts disabled, the
7197 * and caller must save the original value of the current task (see
7198 * curr_task() above) and restore that value before reenabling interrupts and
7199 * re-starting the system.
7201 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7203 void set_curr_task(int cpu, struct task_struct *p)
7210 #ifdef CONFIG_CGROUP_SCHED
7211 /* task_group_lock serializes the addition/removal of task groups */
7212 static DEFINE_SPINLOCK(task_group_lock);
7214 static void free_sched_group(struct task_group *tg)
7216 free_fair_sched_group(tg);
7217 free_rt_sched_group(tg);
7222 /* allocate runqueue etc for a new task group */
7223 struct task_group *sched_create_group(struct task_group *parent)
7225 struct task_group *tg;
7227 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7229 return ERR_PTR(-ENOMEM);
7231 if (!alloc_fair_sched_group(tg, parent))
7234 if (!alloc_rt_sched_group(tg, parent))
7240 free_sched_group(tg);
7241 return ERR_PTR(-ENOMEM);
7244 void sched_online_group(struct task_group *tg, struct task_group *parent)
7246 unsigned long flags;
7248 spin_lock_irqsave(&task_group_lock, flags);
7249 list_add_rcu(&tg->list, &task_groups);
7251 WARN_ON(!parent); /* root should already exist */
7253 tg->parent = parent;
7254 INIT_LIST_HEAD(&tg->children);
7255 list_add_rcu(&tg->siblings, &parent->children);
7256 spin_unlock_irqrestore(&task_group_lock, flags);
7259 /* rcu callback to free various structures associated with a task group */
7260 static void free_sched_group_rcu(struct rcu_head *rhp)
7262 /* now it should be safe to free those cfs_rqs */
7263 free_sched_group(container_of(rhp, struct task_group, rcu));
7266 /* Destroy runqueue etc associated with a task group */
7267 void sched_destroy_group(struct task_group *tg)
7269 /* wait for possible concurrent references to cfs_rqs complete */
7270 call_rcu(&tg->rcu, free_sched_group_rcu);
7273 void sched_offline_group(struct task_group *tg)
7275 unsigned long flags;
7278 /* end participation in shares distribution */
7279 for_each_possible_cpu(i)
7280 unregister_fair_sched_group(tg, i);
7282 spin_lock_irqsave(&task_group_lock, flags);
7283 list_del_rcu(&tg->list);
7284 list_del_rcu(&tg->siblings);
7285 spin_unlock_irqrestore(&task_group_lock, flags);
7288 /* change task's runqueue when it moves between groups.
7289 * The caller of this function should have put the task in its new group
7290 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7291 * reflect its new group.
7293 void sched_move_task(struct task_struct *tsk)
7295 struct task_group *tg;
7297 unsigned long flags;
7300 rq = task_rq_lock(tsk, &flags);
7302 running = task_current(rq, tsk);
7306 dequeue_task(rq, tsk, 0);
7307 if (unlikely(running))
7308 tsk->sched_class->put_prev_task(rq, tsk);
7310 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7311 lockdep_is_held(&tsk->sighand->siglock)),
7312 struct task_group, css);
7313 tg = autogroup_task_group(tsk, tg);
7314 tsk->sched_task_group = tg;
7316 #ifdef CONFIG_FAIR_GROUP_SCHED
7317 if (tsk->sched_class->task_move_group)
7318 tsk->sched_class->task_move_group(tsk, on_rq);
7321 set_task_rq(tsk, task_cpu(tsk));
7323 if (unlikely(running))
7324 tsk->sched_class->set_curr_task(rq);
7326 enqueue_task(rq, tsk, 0);
7328 task_rq_unlock(rq, tsk, &flags);
7330 #endif /* CONFIG_CGROUP_SCHED */
7332 #ifdef CONFIG_RT_GROUP_SCHED
7334 * Ensure that the real time constraints are schedulable.
7336 static DEFINE_MUTEX(rt_constraints_mutex);
7338 /* Must be called with tasklist_lock held */
7339 static inline int tg_has_rt_tasks(struct task_group *tg)
7341 struct task_struct *g, *p;
7343 do_each_thread(g, p) {
7344 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7346 } while_each_thread(g, p);
7351 struct rt_schedulable_data {
7352 struct task_group *tg;
7357 static int tg_rt_schedulable(struct task_group *tg, void *data)
7359 struct rt_schedulable_data *d = data;
7360 struct task_group *child;
7361 unsigned long total, sum = 0;
7362 u64 period, runtime;
7364 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7365 runtime = tg->rt_bandwidth.rt_runtime;
7368 period = d->rt_period;
7369 runtime = d->rt_runtime;
7373 * Cannot have more runtime than the period.
7375 if (runtime > period && runtime != RUNTIME_INF)
7379 * Ensure we don't starve existing RT tasks.
7381 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7384 total = to_ratio(period, runtime);
7387 * Nobody can have more than the global setting allows.
7389 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7393 * The sum of our children's runtime should not exceed our own.
7395 list_for_each_entry_rcu(child, &tg->children, siblings) {
7396 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7397 runtime = child->rt_bandwidth.rt_runtime;
7399 if (child == d->tg) {
7400 period = d->rt_period;
7401 runtime = d->rt_runtime;
7404 sum += to_ratio(period, runtime);
7413 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7417 struct rt_schedulable_data data = {
7419 .rt_period = period,
7420 .rt_runtime = runtime,
7424 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7430 static int tg_set_rt_bandwidth(struct task_group *tg,
7431 u64 rt_period, u64 rt_runtime)
7435 mutex_lock(&rt_constraints_mutex);
7436 read_lock(&tasklist_lock);
7437 err = __rt_schedulable(tg, rt_period, rt_runtime);
7441 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7442 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7443 tg->rt_bandwidth.rt_runtime = rt_runtime;
7445 for_each_possible_cpu(i) {
7446 struct rt_rq *rt_rq = tg->rt_rq[i];
7448 raw_spin_lock(&rt_rq->rt_runtime_lock);
7449 rt_rq->rt_runtime = rt_runtime;
7450 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7452 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7454 read_unlock(&tasklist_lock);
7455 mutex_unlock(&rt_constraints_mutex);
7460 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7462 u64 rt_runtime, rt_period;
7464 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7465 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7466 if (rt_runtime_us < 0)
7467 rt_runtime = RUNTIME_INF;
7469 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7472 static long sched_group_rt_runtime(struct task_group *tg)
7476 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7479 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7480 do_div(rt_runtime_us, NSEC_PER_USEC);
7481 return rt_runtime_us;
7484 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7486 u64 rt_runtime, rt_period;
7488 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7489 rt_runtime = tg->rt_bandwidth.rt_runtime;
7494 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7497 static long sched_group_rt_period(struct task_group *tg)
7501 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7502 do_div(rt_period_us, NSEC_PER_USEC);
7503 return rt_period_us;
7505 #endif /* CONFIG_RT_GROUP_SCHED */
7507 #ifdef CONFIG_RT_GROUP_SCHED
7508 static int sched_rt_global_constraints(void)
7512 mutex_lock(&rt_constraints_mutex);
7513 read_lock(&tasklist_lock);
7514 ret = __rt_schedulable(NULL, 0, 0);
7515 read_unlock(&tasklist_lock);
7516 mutex_unlock(&rt_constraints_mutex);
7521 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7523 /* Don't accept realtime tasks when there is no way for them to run */
7524 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7530 #else /* !CONFIG_RT_GROUP_SCHED */
7531 static int sched_rt_global_constraints(void)
7533 unsigned long flags;
7536 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7537 for_each_possible_cpu(i) {
7538 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7540 raw_spin_lock(&rt_rq->rt_runtime_lock);
7541 rt_rq->rt_runtime = global_rt_runtime();
7542 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7544 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7548 #endif /* CONFIG_RT_GROUP_SCHED */
7550 static int sched_dl_global_constraints(void)
7552 u64 runtime = global_rt_runtime();
7553 u64 period = global_rt_period();
7554 u64 new_bw = to_ratio(period, runtime);
7556 unsigned long flags;
7559 * Here we want to check the bandwidth not being set to some
7560 * value smaller than the currently allocated bandwidth in
7561 * any of the root_domains.
7563 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7564 * cycling on root_domains... Discussion on different/better
7565 * solutions is welcome!
7567 for_each_possible_cpu(cpu) {
7568 struct dl_bw *dl_b = dl_bw_of(cpu);
7570 raw_spin_lock_irqsave(&dl_b->lock, flags);
7571 if (new_bw < dl_b->total_bw)
7573 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7582 static void sched_dl_do_global(void)
7586 unsigned long flags;
7588 def_dl_bandwidth.dl_period = global_rt_period();
7589 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7591 if (global_rt_runtime() != RUNTIME_INF)
7592 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7595 * FIXME: As above...
7597 for_each_possible_cpu(cpu) {
7598 struct dl_bw *dl_b = dl_bw_of(cpu);
7600 raw_spin_lock_irqsave(&dl_b->lock, flags);
7602 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7606 static int sched_rt_global_validate(void)
7608 if (sysctl_sched_rt_period <= 0)
7611 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7612 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7618 static void sched_rt_do_global(void)
7620 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7621 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7624 int sched_rt_handler(struct ctl_table *table, int write,
7625 void __user *buffer, size_t *lenp,
7628 int old_period, old_runtime;
7629 static DEFINE_MUTEX(mutex);
7633 old_period = sysctl_sched_rt_period;
7634 old_runtime = sysctl_sched_rt_runtime;
7636 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7638 if (!ret && write) {
7639 ret = sched_rt_global_validate();
7643 ret = sched_rt_global_constraints();
7647 ret = sched_dl_global_constraints();
7651 sched_rt_do_global();
7652 sched_dl_do_global();
7656 sysctl_sched_rt_period = old_period;
7657 sysctl_sched_rt_runtime = old_runtime;
7659 mutex_unlock(&mutex);
7664 int sched_rr_handler(struct ctl_table *table, int write,
7665 void __user *buffer, size_t *lenp,
7669 static DEFINE_MUTEX(mutex);
7672 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7673 /* make sure that internally we keep jiffies */
7674 /* also, writing zero resets timeslice to default */
7675 if (!ret && write) {
7676 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7677 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7679 mutex_unlock(&mutex);
7683 #ifdef CONFIG_CGROUP_SCHED
7685 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7687 return css ? container_of(css, struct task_group, css) : NULL;
7690 static struct cgroup_subsys_state *
7691 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7693 struct task_group *parent = css_tg(parent_css);
7694 struct task_group *tg;
7697 /* This is early initialization for the top cgroup */
7698 return &root_task_group.css;
7701 tg = sched_create_group(parent);
7703 return ERR_PTR(-ENOMEM);
7708 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7710 struct task_group *tg = css_tg(css);
7711 struct task_group *parent = css_tg(css->parent);
7714 sched_online_group(tg, parent);
7718 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7720 struct task_group *tg = css_tg(css);
7722 sched_destroy_group(tg);
7725 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7727 struct task_group *tg = css_tg(css);
7729 sched_offline_group(tg);
7732 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7733 struct cgroup_taskset *tset)
7735 struct task_struct *task;
7737 cgroup_taskset_for_each(task, tset) {
7738 #ifdef CONFIG_RT_GROUP_SCHED
7739 if (!sched_rt_can_attach(css_tg(css), task))
7742 /* We don't support RT-tasks being in separate groups */
7743 if (task->sched_class != &fair_sched_class)
7750 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7751 struct cgroup_taskset *tset)
7753 struct task_struct *task;
7755 cgroup_taskset_for_each(task, tset)
7756 sched_move_task(task);
7759 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7760 struct cgroup_subsys_state *old_css,
7761 struct task_struct *task)
7764 * cgroup_exit() is called in the copy_process() failure path.
7765 * Ignore this case since the task hasn't ran yet, this avoids
7766 * trying to poke a half freed task state from generic code.
7768 if (!(task->flags & PF_EXITING))
7771 sched_move_task(task);
7774 #ifdef CONFIG_FAIR_GROUP_SCHED
7775 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7776 struct cftype *cftype, u64 shareval)
7778 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7781 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7784 struct task_group *tg = css_tg(css);
7786 return (u64) scale_load_down(tg->shares);
7789 #ifdef CONFIG_CFS_BANDWIDTH
7790 static DEFINE_MUTEX(cfs_constraints_mutex);
7792 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7793 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7795 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7797 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7799 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7800 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7802 if (tg == &root_task_group)
7806 * Ensure we have at some amount of bandwidth every period. This is
7807 * to prevent reaching a state of large arrears when throttled via
7808 * entity_tick() resulting in prolonged exit starvation.
7810 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7814 * Likewise, bound things on the otherside by preventing insane quota
7815 * periods. This also allows us to normalize in computing quota
7818 if (period > max_cfs_quota_period)
7822 * Prevent race between setting of cfs_rq->runtime_enabled and
7823 * unthrottle_offline_cfs_rqs().
7826 mutex_lock(&cfs_constraints_mutex);
7827 ret = __cfs_schedulable(tg, period, quota);
7831 runtime_enabled = quota != RUNTIME_INF;
7832 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7834 * If we need to toggle cfs_bandwidth_used, off->on must occur
7835 * before making related changes, and on->off must occur afterwards
7837 if (runtime_enabled && !runtime_was_enabled)
7838 cfs_bandwidth_usage_inc();
7839 raw_spin_lock_irq(&cfs_b->lock);
7840 cfs_b->period = ns_to_ktime(period);
7841 cfs_b->quota = quota;
7843 __refill_cfs_bandwidth_runtime(cfs_b);
7844 /* restart the period timer (if active) to handle new period expiry */
7845 if (runtime_enabled && cfs_b->timer_active) {
7846 /* force a reprogram */
7847 __start_cfs_bandwidth(cfs_b, true);
7849 raw_spin_unlock_irq(&cfs_b->lock);
7851 for_each_online_cpu(i) {
7852 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7853 struct rq *rq = cfs_rq->rq;
7855 raw_spin_lock_irq(&rq->lock);
7856 cfs_rq->runtime_enabled = runtime_enabled;
7857 cfs_rq->runtime_remaining = 0;
7859 if (cfs_rq->throttled)
7860 unthrottle_cfs_rq(cfs_rq);
7861 raw_spin_unlock_irq(&rq->lock);
7863 if (runtime_was_enabled && !runtime_enabled)
7864 cfs_bandwidth_usage_dec();
7866 mutex_unlock(&cfs_constraints_mutex);
7872 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7876 period = ktime_to_ns(tg->cfs_bandwidth.period);
7877 if (cfs_quota_us < 0)
7878 quota = RUNTIME_INF;
7880 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7882 return tg_set_cfs_bandwidth(tg, period, quota);
7885 long tg_get_cfs_quota(struct task_group *tg)
7889 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7892 quota_us = tg->cfs_bandwidth.quota;
7893 do_div(quota_us, NSEC_PER_USEC);
7898 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7902 period = (u64)cfs_period_us * NSEC_PER_USEC;
7903 quota = tg->cfs_bandwidth.quota;
7905 return tg_set_cfs_bandwidth(tg, period, quota);
7908 long tg_get_cfs_period(struct task_group *tg)
7912 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7913 do_div(cfs_period_us, NSEC_PER_USEC);
7915 return cfs_period_us;
7918 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7921 return tg_get_cfs_quota(css_tg(css));
7924 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7925 struct cftype *cftype, s64 cfs_quota_us)
7927 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7930 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7933 return tg_get_cfs_period(css_tg(css));
7936 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7937 struct cftype *cftype, u64 cfs_period_us)
7939 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7942 struct cfs_schedulable_data {
7943 struct task_group *tg;
7948 * normalize group quota/period to be quota/max_period
7949 * note: units are usecs
7951 static u64 normalize_cfs_quota(struct task_group *tg,
7952 struct cfs_schedulable_data *d)
7960 period = tg_get_cfs_period(tg);
7961 quota = tg_get_cfs_quota(tg);
7964 /* note: these should typically be equivalent */
7965 if (quota == RUNTIME_INF || quota == -1)
7968 return to_ratio(period, quota);
7971 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7973 struct cfs_schedulable_data *d = data;
7974 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7975 s64 quota = 0, parent_quota = -1;
7978 quota = RUNTIME_INF;
7980 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7982 quota = normalize_cfs_quota(tg, d);
7983 parent_quota = parent_b->hierarchal_quota;
7986 * ensure max(child_quota) <= parent_quota, inherit when no
7989 if (quota == RUNTIME_INF)
7990 quota = parent_quota;
7991 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7994 cfs_b->hierarchal_quota = quota;
7999 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8002 struct cfs_schedulable_data data = {
8008 if (quota != RUNTIME_INF) {
8009 do_div(data.period, NSEC_PER_USEC);
8010 do_div(data.quota, NSEC_PER_USEC);
8014 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8020 static int cpu_stats_show(struct seq_file *sf, void *v)
8022 struct task_group *tg = css_tg(seq_css(sf));
8023 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8025 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8026 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8027 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8031 #endif /* CONFIG_CFS_BANDWIDTH */
8032 #endif /* CONFIG_FAIR_GROUP_SCHED */
8034 #ifdef CONFIG_RT_GROUP_SCHED
8035 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8036 struct cftype *cft, s64 val)
8038 return sched_group_set_rt_runtime(css_tg(css), val);
8041 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8044 return sched_group_rt_runtime(css_tg(css));
8047 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8048 struct cftype *cftype, u64 rt_period_us)
8050 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8053 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8056 return sched_group_rt_period(css_tg(css));
8058 #endif /* CONFIG_RT_GROUP_SCHED */
8060 static struct cftype cpu_files[] = {
8061 #ifdef CONFIG_FAIR_GROUP_SCHED
8064 .read_u64 = cpu_shares_read_u64,
8065 .write_u64 = cpu_shares_write_u64,
8068 #ifdef CONFIG_CFS_BANDWIDTH
8070 .name = "cfs_quota_us",
8071 .read_s64 = cpu_cfs_quota_read_s64,
8072 .write_s64 = cpu_cfs_quota_write_s64,
8075 .name = "cfs_period_us",
8076 .read_u64 = cpu_cfs_period_read_u64,
8077 .write_u64 = cpu_cfs_period_write_u64,
8081 .seq_show = cpu_stats_show,
8084 #ifdef CONFIG_RT_GROUP_SCHED
8086 .name = "rt_runtime_us",
8087 .read_s64 = cpu_rt_runtime_read,
8088 .write_s64 = cpu_rt_runtime_write,
8091 .name = "rt_period_us",
8092 .read_u64 = cpu_rt_period_read_uint,
8093 .write_u64 = cpu_rt_period_write_uint,
8099 struct cgroup_subsys cpu_cgrp_subsys = {
8100 .css_alloc = cpu_cgroup_css_alloc,
8101 .css_free = cpu_cgroup_css_free,
8102 .css_online = cpu_cgroup_css_online,
8103 .css_offline = cpu_cgroup_css_offline,
8104 .can_attach = cpu_cgroup_can_attach,
8105 .attach = cpu_cgroup_attach,
8106 .exit = cpu_cgroup_exit,
8107 .base_cftypes = cpu_files,
8111 #endif /* CONFIG_CGROUP_SCHED */
8113 void dump_cpu_task(int cpu)
8115 pr_info("Task dump for CPU %d:\n", cpu);
8116 sched_show_task(cpu_curr(cpu));