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;
143 update_rq_clock_task(rq, delta);
147 * Debugging: various feature bits
150 #define SCHED_FEAT(name, enabled) \
151 (1UL << __SCHED_FEAT_##name) * enabled |
153 const_debug unsigned int sysctl_sched_features =
154 #include "features.h"
159 #ifdef CONFIG_SCHED_DEBUG
160 #define SCHED_FEAT(name, enabled) \
163 static const char * const sched_feat_names[] = {
164 #include "features.h"
169 static int sched_feat_show(struct seq_file *m, void *v)
173 for (i = 0; i < __SCHED_FEAT_NR; i++) {
174 if (!(sysctl_sched_features & (1UL << i)))
176 seq_printf(m, "%s ", sched_feat_names[i]);
183 #ifdef HAVE_JUMP_LABEL
185 #define jump_label_key__true STATIC_KEY_INIT_TRUE
186 #define jump_label_key__false STATIC_KEY_INIT_FALSE
188 #define SCHED_FEAT(name, enabled) \
189 jump_label_key__##enabled ,
191 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
192 #include "features.h"
197 static void sched_feat_disable(int i)
199 if (static_key_enabled(&sched_feat_keys[i]))
200 static_key_slow_dec(&sched_feat_keys[i]);
203 static void sched_feat_enable(int i)
205 if (!static_key_enabled(&sched_feat_keys[i]))
206 static_key_slow_inc(&sched_feat_keys[i]);
209 static void sched_feat_disable(int i) { };
210 static void sched_feat_enable(int i) { };
211 #endif /* HAVE_JUMP_LABEL */
213 static int sched_feat_set(char *cmp)
218 if (strncmp(cmp, "NO_", 3) == 0) {
223 for (i = 0; i < __SCHED_FEAT_NR; i++) {
224 if (strcmp(cmp, sched_feat_names[i]) == 0) {
226 sysctl_sched_features &= ~(1UL << i);
227 sched_feat_disable(i);
229 sysctl_sched_features |= (1UL << i);
230 sched_feat_enable(i);
240 sched_feat_write(struct file *filp, const char __user *ubuf,
241 size_t cnt, loff_t *ppos)
250 if (copy_from_user(&buf, ubuf, cnt))
256 i = sched_feat_set(cmp);
257 if (i == __SCHED_FEAT_NR)
265 static int sched_feat_open(struct inode *inode, struct file *filp)
267 return single_open(filp, sched_feat_show, NULL);
270 static const struct file_operations sched_feat_fops = {
271 .open = sched_feat_open,
272 .write = sched_feat_write,
275 .release = single_release,
278 static __init int sched_init_debug(void)
280 debugfs_create_file("sched_features", 0644, NULL, NULL,
285 late_initcall(sched_init_debug);
286 #endif /* CONFIG_SCHED_DEBUG */
289 * Number of tasks to iterate in a single balance run.
290 * Limited because this is done with IRQs disabled.
292 const_debug unsigned int sysctl_sched_nr_migrate = 32;
295 * period over which we average the RT time consumption, measured
300 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
303 * period over which we measure -rt task cpu usage in us.
306 unsigned int sysctl_sched_rt_period = 1000000;
308 __read_mostly int scheduler_running;
311 * part of the period that we allow rt tasks to run in us.
314 int sysctl_sched_rt_runtime = 950000;
317 * __task_rq_lock - lock the rq @p resides on.
319 static inline struct rq *__task_rq_lock(struct task_struct *p)
324 lockdep_assert_held(&p->pi_lock);
328 raw_spin_lock(&rq->lock);
329 if (likely(rq == task_rq(p)))
331 raw_spin_unlock(&rq->lock);
336 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
338 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
339 __acquires(p->pi_lock)
345 raw_spin_lock_irqsave(&p->pi_lock, *flags);
347 raw_spin_lock(&rq->lock);
348 if (likely(rq == task_rq(p)))
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 static void __task_rq_unlock(struct rq *rq)
358 raw_spin_unlock(&rq->lock);
362 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
364 __releases(p->pi_lock)
366 raw_spin_unlock(&rq->lock);
367 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
371 * this_rq_lock - lock this runqueue and disable interrupts.
373 static struct rq *this_rq_lock(void)
380 raw_spin_lock(&rq->lock);
385 #ifdef CONFIG_SCHED_HRTICK
387 * Use HR-timers to deliver accurate preemption points.
390 static void hrtick_clear(struct rq *rq)
392 if (hrtimer_active(&rq->hrtick_timer))
393 hrtimer_cancel(&rq->hrtick_timer);
397 * High-resolution timer tick.
398 * Runs from hardirq context with interrupts disabled.
400 static enum hrtimer_restart hrtick(struct hrtimer *timer)
402 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
404 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
406 raw_spin_lock(&rq->lock);
408 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
409 raw_spin_unlock(&rq->lock);
411 return HRTIMER_NORESTART;
416 static int __hrtick_restart(struct rq *rq)
418 struct hrtimer *timer = &rq->hrtick_timer;
419 ktime_t time = hrtimer_get_softexpires(timer);
421 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
425 * called from hardirq (IPI) context
427 static void __hrtick_start(void *arg)
431 raw_spin_lock(&rq->lock);
432 __hrtick_restart(rq);
433 rq->hrtick_csd_pending = 0;
434 raw_spin_unlock(&rq->lock);
438 * Called to set the hrtick timer state.
440 * called with rq->lock held and irqs disabled
442 void hrtick_start(struct rq *rq, u64 delay)
444 struct hrtimer *timer = &rq->hrtick_timer;
445 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
447 hrtimer_set_expires(timer, time);
449 if (rq == this_rq()) {
450 __hrtick_restart(rq);
451 } else if (!rq->hrtick_csd_pending) {
452 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
453 rq->hrtick_csd_pending = 1;
458 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
460 int cpu = (int)(long)hcpu;
463 case CPU_UP_CANCELED:
464 case CPU_UP_CANCELED_FROZEN:
465 case CPU_DOWN_PREPARE:
466 case CPU_DOWN_PREPARE_FROZEN:
468 case CPU_DEAD_FROZEN:
469 hrtick_clear(cpu_rq(cpu));
476 static __init void init_hrtick(void)
478 hotcpu_notifier(hotplug_hrtick, 0);
482 * Called to set the hrtick timer state.
484 * called with rq->lock held and irqs disabled
486 void hrtick_start(struct rq *rq, u64 delay)
488 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
489 HRTIMER_MODE_REL_PINNED, 0);
492 static inline void init_hrtick(void)
495 #endif /* CONFIG_SMP */
497 static void init_rq_hrtick(struct rq *rq)
500 rq->hrtick_csd_pending = 0;
502 rq->hrtick_csd.flags = 0;
503 rq->hrtick_csd.func = __hrtick_start;
504 rq->hrtick_csd.info = rq;
507 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
508 rq->hrtick_timer.function = hrtick;
510 #else /* CONFIG_SCHED_HRTICK */
511 static inline void hrtick_clear(struct rq *rq)
515 static inline void init_rq_hrtick(struct rq *rq)
519 static inline void init_hrtick(void)
522 #endif /* CONFIG_SCHED_HRTICK */
525 * cmpxchg based fetch_or, macro so it works for different integer types
527 #define fetch_or(ptr, val) \
528 ({ typeof(*(ptr)) __old, __val = *(ptr); \
530 __old = cmpxchg((ptr), __val, __val | (val)); \
531 if (__old == __val) \
538 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
540 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
541 * this avoids any races wrt polling state changes and thereby avoids
544 static bool set_nr_and_not_polling(struct task_struct *p)
546 struct thread_info *ti = task_thread_info(p);
547 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
551 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
553 * If this returns true, then the idle task promises to call
554 * sched_ttwu_pending() and reschedule soon.
556 static bool set_nr_if_polling(struct task_struct *p)
558 struct thread_info *ti = task_thread_info(p);
559 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags);
562 if (!(val & _TIF_POLLING_NRFLAG))
564 if (val & _TIF_NEED_RESCHED)
566 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
575 static bool set_nr_and_not_polling(struct task_struct *p)
577 set_tsk_need_resched(p);
582 static bool set_nr_if_polling(struct task_struct *p)
590 * resched_task - mark a task 'to be rescheduled now'.
592 * On UP this means the setting of the need_resched flag, on SMP it
593 * might also involve a cross-CPU call to trigger the scheduler on
596 void resched_task(struct task_struct *p)
600 lockdep_assert_held(&task_rq(p)->lock);
602 if (test_tsk_need_resched(p))
607 if (cpu == smp_processor_id()) {
608 set_tsk_need_resched(p);
609 set_preempt_need_resched();
613 if (set_nr_and_not_polling(p))
614 smp_send_reschedule(cpu);
616 trace_sched_wake_idle_without_ipi(cpu);
619 void resched_cpu(int cpu)
621 struct rq *rq = cpu_rq(cpu);
624 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
626 resched_task(cpu_curr(cpu));
627 raw_spin_unlock_irqrestore(&rq->lock, flags);
631 #ifdef CONFIG_NO_HZ_COMMON
633 * In the semi idle case, use the nearest busy cpu for migrating timers
634 * from an idle cpu. This is good for power-savings.
636 * We don't do similar optimization for completely idle system, as
637 * selecting an idle cpu will add more delays to the timers than intended
638 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
640 int get_nohz_timer_target(int pinned)
642 int cpu = smp_processor_id();
644 struct sched_domain *sd;
646 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
650 for_each_domain(cpu, sd) {
651 for_each_cpu(i, sched_domain_span(sd)) {
663 * When add_timer_on() enqueues a timer into the timer wheel of an
664 * idle CPU then this timer might expire before the next timer event
665 * which is scheduled to wake up that CPU. In case of a completely
666 * idle system the next event might even be infinite time into the
667 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
668 * leaves the inner idle loop so the newly added timer is taken into
669 * account when the CPU goes back to idle and evaluates the timer
670 * wheel for the next timer event.
672 static void wake_up_idle_cpu(int cpu)
674 struct rq *rq = cpu_rq(cpu);
676 if (cpu == smp_processor_id())
679 if (set_nr_and_not_polling(rq->idle))
680 smp_send_reschedule(cpu);
682 trace_sched_wake_idle_without_ipi(cpu);
685 static bool wake_up_full_nohz_cpu(int cpu)
688 * We just need the target to call irq_exit() and re-evaluate
689 * the next tick. The nohz full kick at least implies that.
690 * If needed we can still optimize that later with an
693 if (tick_nohz_full_cpu(cpu)) {
694 if (cpu != smp_processor_id() ||
695 tick_nohz_tick_stopped())
696 tick_nohz_full_kick_cpu(cpu);
703 void wake_up_nohz_cpu(int cpu)
705 if (!wake_up_full_nohz_cpu(cpu))
706 wake_up_idle_cpu(cpu);
709 static inline bool got_nohz_idle_kick(void)
711 int cpu = smp_processor_id();
713 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
716 if (idle_cpu(cpu) && !need_resched())
720 * We can't run Idle Load Balance on this CPU for this time so we
721 * cancel it and clear NOHZ_BALANCE_KICK
723 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
727 #else /* CONFIG_NO_HZ_COMMON */
729 static inline bool got_nohz_idle_kick(void)
734 #endif /* CONFIG_NO_HZ_COMMON */
736 #ifdef CONFIG_NO_HZ_FULL
737 bool sched_can_stop_tick(void)
743 /* Make sure rq->nr_running update is visible after the IPI */
746 /* More than one running task need preemption */
747 if (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) {
1031 resched_task(rq->curr);
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)
1578 && !tick_nohz_full_cpu(smp_processor_id())
1579 && !got_nohz_idle_kick())
1583 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1584 * traditionally all their work was done from the interrupt return
1585 * path. Now that we actually do some work, we need to make sure
1588 * Some archs already do call them, luckily irq_enter/exit nest
1591 * Arguably we should visit all archs and update all handlers,
1592 * however a fair share of IPIs are still resched only so this would
1593 * somewhat pessimize the simple resched case.
1596 tick_nohz_full_check();
1597 sched_ttwu_pending();
1600 * Check if someone kicked us for doing the nohz idle load balance.
1602 if (unlikely(got_nohz_idle_kick())) {
1603 this_rq()->idle_balance = 1;
1604 raise_softirq_irqoff(SCHED_SOFTIRQ);
1609 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1611 struct rq *rq = cpu_rq(cpu);
1613 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1614 if (!set_nr_if_polling(rq->idle))
1615 smp_send_reschedule(cpu);
1617 trace_sched_wake_idle_without_ipi(cpu);
1621 bool cpus_share_cache(int this_cpu, int that_cpu)
1623 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1625 #endif /* CONFIG_SMP */
1627 static void ttwu_queue(struct task_struct *p, int cpu)
1629 struct rq *rq = cpu_rq(cpu);
1631 #if defined(CONFIG_SMP)
1632 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1633 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1634 ttwu_queue_remote(p, cpu);
1639 raw_spin_lock(&rq->lock);
1640 ttwu_do_activate(rq, p, 0);
1641 raw_spin_unlock(&rq->lock);
1645 * try_to_wake_up - wake up a thread
1646 * @p: the thread to be awakened
1647 * @state: the mask of task states that can be woken
1648 * @wake_flags: wake modifier flags (WF_*)
1650 * Put it on the run-queue if it's not already there. The "current"
1651 * thread is always on the run-queue (except when the actual
1652 * re-schedule is in progress), and as such you're allowed to do
1653 * the simpler "current->state = TASK_RUNNING" to mark yourself
1654 * runnable without the overhead of this.
1656 * Return: %true if @p was woken up, %false if it was already running.
1657 * or @state didn't match @p's state.
1660 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1662 unsigned long flags;
1663 int cpu, success = 0;
1666 * If we are going to wake up a thread waiting for CONDITION we
1667 * need to ensure that CONDITION=1 done by the caller can not be
1668 * reordered with p->state check below. This pairs with mb() in
1669 * set_current_state() the waiting thread does.
1671 smp_mb__before_spinlock();
1672 raw_spin_lock_irqsave(&p->pi_lock, flags);
1673 if (!(p->state & state))
1676 success = 1; /* we're going to change ->state */
1679 if (p->on_rq && ttwu_remote(p, wake_flags))
1684 * If the owning (remote) cpu is still in the middle of schedule() with
1685 * this task as prev, wait until its done referencing the task.
1690 * Pairs with the smp_wmb() in finish_lock_switch().
1694 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1695 p->state = TASK_WAKING;
1697 if (p->sched_class->task_waking)
1698 p->sched_class->task_waking(p);
1700 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1701 if (task_cpu(p) != cpu) {
1702 wake_flags |= WF_MIGRATED;
1703 set_task_cpu(p, cpu);
1705 #endif /* CONFIG_SMP */
1709 ttwu_stat(p, cpu, wake_flags);
1711 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1717 * try_to_wake_up_local - try to wake up a local task with rq lock held
1718 * @p: the thread to be awakened
1720 * Put @p on the run-queue if it's not already there. The caller must
1721 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1724 static void try_to_wake_up_local(struct task_struct *p)
1726 struct rq *rq = task_rq(p);
1728 if (WARN_ON_ONCE(rq != this_rq()) ||
1729 WARN_ON_ONCE(p == current))
1732 lockdep_assert_held(&rq->lock);
1734 if (!raw_spin_trylock(&p->pi_lock)) {
1735 raw_spin_unlock(&rq->lock);
1736 raw_spin_lock(&p->pi_lock);
1737 raw_spin_lock(&rq->lock);
1740 if (!(p->state & TASK_NORMAL))
1744 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1746 ttwu_do_wakeup(rq, p, 0);
1747 ttwu_stat(p, smp_processor_id(), 0);
1749 raw_spin_unlock(&p->pi_lock);
1753 * wake_up_process - Wake up a specific process
1754 * @p: The process to be woken up.
1756 * Attempt to wake up the nominated process and move it to the set of runnable
1759 * Return: 1 if the process was woken up, 0 if it was already running.
1761 * It may be assumed that this function implies a write memory barrier before
1762 * changing the task state if and only if any tasks are woken up.
1764 int wake_up_process(struct task_struct *p)
1766 WARN_ON(task_is_stopped_or_traced(p));
1767 return try_to_wake_up(p, TASK_NORMAL, 0);
1769 EXPORT_SYMBOL(wake_up_process);
1771 int wake_up_state(struct task_struct *p, unsigned int state)
1773 return try_to_wake_up(p, state, 0);
1777 * Perform scheduler related setup for a newly forked process p.
1778 * p is forked by current.
1780 * __sched_fork() is basic setup used by init_idle() too:
1782 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1787 p->se.exec_start = 0;
1788 p->se.sum_exec_runtime = 0;
1789 p->se.prev_sum_exec_runtime = 0;
1790 p->se.nr_migrations = 0;
1792 INIT_LIST_HEAD(&p->se.group_node);
1794 #ifdef CONFIG_SCHEDSTATS
1795 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1798 RB_CLEAR_NODE(&p->dl.rb_node);
1799 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1800 p->dl.dl_runtime = p->dl.runtime = 0;
1801 p->dl.dl_deadline = p->dl.deadline = 0;
1802 p->dl.dl_period = 0;
1805 INIT_LIST_HEAD(&p->rt.run_list);
1807 #ifdef CONFIG_PREEMPT_NOTIFIERS
1808 INIT_HLIST_HEAD(&p->preempt_notifiers);
1811 #ifdef CONFIG_NUMA_BALANCING
1812 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1813 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1814 p->mm->numa_scan_seq = 0;
1817 if (clone_flags & CLONE_VM)
1818 p->numa_preferred_nid = current->numa_preferred_nid;
1820 p->numa_preferred_nid = -1;
1822 p->node_stamp = 0ULL;
1823 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1824 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1825 p->numa_work.next = &p->numa_work;
1826 p->numa_faults_memory = NULL;
1827 p->numa_faults_buffer_memory = NULL;
1828 p->last_task_numa_placement = 0;
1829 p->last_sum_exec_runtime = 0;
1831 INIT_LIST_HEAD(&p->numa_entry);
1832 p->numa_group = NULL;
1833 #endif /* CONFIG_NUMA_BALANCING */
1836 #ifdef CONFIG_NUMA_BALANCING
1837 #ifdef CONFIG_SCHED_DEBUG
1838 void set_numabalancing_state(bool enabled)
1841 sched_feat_set("NUMA");
1843 sched_feat_set("NO_NUMA");
1846 __read_mostly bool numabalancing_enabled;
1848 void set_numabalancing_state(bool enabled)
1850 numabalancing_enabled = enabled;
1852 #endif /* CONFIG_SCHED_DEBUG */
1854 #ifdef CONFIG_PROC_SYSCTL
1855 int sysctl_numa_balancing(struct ctl_table *table, int write,
1856 void __user *buffer, size_t *lenp, loff_t *ppos)
1860 int state = numabalancing_enabled;
1862 if (write && !capable(CAP_SYS_ADMIN))
1867 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1871 set_numabalancing_state(state);
1878 * fork()/clone()-time setup:
1880 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1882 unsigned long flags;
1883 int cpu = get_cpu();
1885 __sched_fork(clone_flags, p);
1887 * We mark the process as running here. This guarantees that
1888 * nobody will actually run it, and a signal or other external
1889 * event cannot wake it up and insert it on the runqueue either.
1891 p->state = TASK_RUNNING;
1894 * Make sure we do not leak PI boosting priority to the child.
1896 p->prio = current->normal_prio;
1899 * Revert to default priority/policy on fork if requested.
1901 if (unlikely(p->sched_reset_on_fork)) {
1902 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1903 p->policy = SCHED_NORMAL;
1904 p->static_prio = NICE_TO_PRIO(0);
1906 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1907 p->static_prio = NICE_TO_PRIO(0);
1909 p->prio = p->normal_prio = __normal_prio(p);
1913 * We don't need the reset flag anymore after the fork. It has
1914 * fulfilled its duty:
1916 p->sched_reset_on_fork = 0;
1919 if (dl_prio(p->prio)) {
1922 } else if (rt_prio(p->prio)) {
1923 p->sched_class = &rt_sched_class;
1925 p->sched_class = &fair_sched_class;
1928 if (p->sched_class->task_fork)
1929 p->sched_class->task_fork(p);
1932 * The child is not yet in the pid-hash so no cgroup attach races,
1933 * and the cgroup is pinned to this child due to cgroup_fork()
1934 * is ran before sched_fork().
1936 * Silence PROVE_RCU.
1938 raw_spin_lock_irqsave(&p->pi_lock, flags);
1939 set_task_cpu(p, cpu);
1940 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1942 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1943 if (likely(sched_info_on()))
1944 memset(&p->sched_info, 0, sizeof(p->sched_info));
1946 #if defined(CONFIG_SMP)
1949 init_task_preempt_count(p);
1951 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1952 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1959 unsigned long to_ratio(u64 period, u64 runtime)
1961 if (runtime == RUNTIME_INF)
1965 * Doing this here saves a lot of checks in all
1966 * the calling paths, and returning zero seems
1967 * safe for them anyway.
1972 return div64_u64(runtime << 20, period);
1976 inline struct dl_bw *dl_bw_of(int i)
1978 return &cpu_rq(i)->rd->dl_bw;
1981 static inline int dl_bw_cpus(int i)
1983 struct root_domain *rd = cpu_rq(i)->rd;
1986 for_each_cpu_and(i, rd->span, cpu_active_mask)
1992 inline struct dl_bw *dl_bw_of(int i)
1994 return &cpu_rq(i)->dl.dl_bw;
1997 static inline int dl_bw_cpus(int i)
2004 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
2006 dl_b->total_bw -= tsk_bw;
2010 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
2012 dl_b->total_bw += tsk_bw;
2016 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
2018 return dl_b->bw != -1 &&
2019 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
2023 * We must be sure that accepting a new task (or allowing changing the
2024 * parameters of an existing one) is consistent with the bandwidth
2025 * constraints. If yes, this function also accordingly updates the currently
2026 * allocated bandwidth to reflect the new situation.
2028 * This function is called while holding p's rq->lock.
2030 static int dl_overflow(struct task_struct *p, int policy,
2031 const struct sched_attr *attr)
2034 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2035 u64 period = attr->sched_period ?: attr->sched_deadline;
2036 u64 runtime = attr->sched_runtime;
2037 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2040 if (new_bw == p->dl.dl_bw)
2044 * Either if a task, enters, leave, or stays -deadline but changes
2045 * its parameters, we may need to update accordingly the total
2046 * allocated bandwidth of the container.
2048 raw_spin_lock(&dl_b->lock);
2049 cpus = dl_bw_cpus(task_cpu(p));
2050 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2051 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2052 __dl_add(dl_b, new_bw);
2054 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2055 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2056 __dl_clear(dl_b, p->dl.dl_bw);
2057 __dl_add(dl_b, new_bw);
2059 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2060 __dl_clear(dl_b, p->dl.dl_bw);
2063 raw_spin_unlock(&dl_b->lock);
2068 extern void init_dl_bw(struct dl_bw *dl_b);
2071 * wake_up_new_task - wake up a newly created task for the first time.
2073 * This function will do some initial scheduler statistics housekeeping
2074 * that must be done for every newly created context, then puts the task
2075 * on the runqueue and wakes it.
2077 void wake_up_new_task(struct task_struct *p)
2079 unsigned long flags;
2082 raw_spin_lock_irqsave(&p->pi_lock, flags);
2085 * Fork balancing, do it here and not earlier because:
2086 * - cpus_allowed can change in the fork path
2087 * - any previously selected cpu might disappear through hotplug
2089 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2092 /* Initialize new task's runnable average */
2093 init_task_runnable_average(p);
2094 rq = __task_rq_lock(p);
2095 activate_task(rq, p, 0);
2097 trace_sched_wakeup_new(p, true);
2098 check_preempt_curr(rq, p, WF_FORK);
2100 if (p->sched_class->task_woken)
2101 p->sched_class->task_woken(rq, p);
2103 task_rq_unlock(rq, p, &flags);
2106 #ifdef CONFIG_PREEMPT_NOTIFIERS
2109 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2110 * @notifier: notifier struct to register
2112 void preempt_notifier_register(struct preempt_notifier *notifier)
2114 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2116 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2119 * preempt_notifier_unregister - no longer interested in preemption notifications
2120 * @notifier: notifier struct to unregister
2122 * This is safe to call from within a preemption notifier.
2124 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2126 hlist_del(¬ifier->link);
2128 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2130 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2132 struct preempt_notifier *notifier;
2134 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2135 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2139 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2140 struct task_struct *next)
2142 struct preempt_notifier *notifier;
2144 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2145 notifier->ops->sched_out(notifier, next);
2148 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2150 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2155 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2156 struct task_struct *next)
2160 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2163 * prepare_task_switch - prepare to switch tasks
2164 * @rq: the runqueue preparing to switch
2165 * @prev: the current task that is being switched out
2166 * @next: the task we are going to switch to.
2168 * This is called with the rq lock held and interrupts off. It must
2169 * be paired with a subsequent finish_task_switch after the context
2172 * prepare_task_switch sets up locking and calls architecture specific
2176 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2177 struct task_struct *next)
2179 trace_sched_switch(prev, next);
2180 sched_info_switch(rq, prev, next);
2181 perf_event_task_sched_out(prev, next);
2182 fire_sched_out_preempt_notifiers(prev, next);
2183 prepare_lock_switch(rq, next);
2184 prepare_arch_switch(next);
2188 * finish_task_switch - clean up after a task-switch
2189 * @rq: runqueue associated with task-switch
2190 * @prev: the thread we just switched away from.
2192 * finish_task_switch must be called after the context switch, paired
2193 * with a prepare_task_switch call before the context switch.
2194 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2195 * and do any other architecture-specific cleanup actions.
2197 * Note that we may have delayed dropping an mm in context_switch(). If
2198 * so, we finish that here outside of the runqueue lock. (Doing it
2199 * with the lock held can cause deadlocks; see schedule() for
2202 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2203 __releases(rq->lock)
2205 struct mm_struct *mm = rq->prev_mm;
2211 * A task struct has one reference for the use as "current".
2212 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2213 * schedule one last time. The schedule call will never return, and
2214 * the scheduled task must drop that reference.
2215 * The test for TASK_DEAD must occur while the runqueue locks are
2216 * still held, otherwise prev could be scheduled on another cpu, die
2217 * there before we look at prev->state, and then the reference would
2219 * Manfred Spraul <manfred@colorfullife.com>
2221 prev_state = prev->state;
2222 vtime_task_switch(prev);
2223 finish_arch_switch(prev);
2224 perf_event_task_sched_in(prev, current);
2225 finish_lock_switch(rq, prev);
2226 finish_arch_post_lock_switch();
2228 fire_sched_in_preempt_notifiers(current);
2231 if (unlikely(prev_state == TASK_DEAD)) {
2232 if (prev->sched_class->task_dead)
2233 prev->sched_class->task_dead(prev);
2236 * Remove function-return probe instances associated with this
2237 * task and put them back on the free list.
2239 kprobe_flush_task(prev);
2240 put_task_struct(prev);
2243 tick_nohz_task_switch(current);
2248 /* rq->lock is NOT held, but preemption is disabled */
2249 static inline void post_schedule(struct rq *rq)
2251 if (rq->post_schedule) {
2252 unsigned long flags;
2254 raw_spin_lock_irqsave(&rq->lock, flags);
2255 if (rq->curr->sched_class->post_schedule)
2256 rq->curr->sched_class->post_schedule(rq);
2257 raw_spin_unlock_irqrestore(&rq->lock, flags);
2259 rq->post_schedule = 0;
2265 static inline void post_schedule(struct rq *rq)
2272 * schedule_tail - first thing a freshly forked thread must call.
2273 * @prev: the thread we just switched away from.
2275 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2276 __releases(rq->lock)
2278 struct rq *rq = this_rq();
2280 finish_task_switch(rq, prev);
2283 * FIXME: do we need to worry about rq being invalidated by the
2288 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2289 /* In this case, finish_task_switch does not reenable preemption */
2292 if (current->set_child_tid)
2293 put_user(task_pid_vnr(current), current->set_child_tid);
2297 * context_switch - switch to the new MM and the new
2298 * thread's register state.
2301 context_switch(struct rq *rq, struct task_struct *prev,
2302 struct task_struct *next)
2304 struct mm_struct *mm, *oldmm;
2306 prepare_task_switch(rq, prev, next);
2309 oldmm = prev->active_mm;
2311 * For paravirt, this is coupled with an exit in switch_to to
2312 * combine the page table reload and the switch backend into
2315 arch_start_context_switch(prev);
2318 next->active_mm = oldmm;
2319 atomic_inc(&oldmm->mm_count);
2320 enter_lazy_tlb(oldmm, next);
2322 switch_mm(oldmm, mm, next);
2325 prev->active_mm = NULL;
2326 rq->prev_mm = oldmm;
2329 * Since the runqueue lock will be released by the next
2330 * task (which is an invalid locking op but in the case
2331 * of the scheduler it's an obvious special-case), so we
2332 * do an early lockdep release here:
2334 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2335 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2338 context_tracking_task_switch(prev, next);
2339 /* Here we just switch the register state and the stack. */
2340 switch_to(prev, next, prev);
2344 * this_rq must be evaluated again because prev may have moved
2345 * CPUs since it called schedule(), thus the 'rq' on its stack
2346 * frame will be invalid.
2348 finish_task_switch(this_rq(), prev);
2352 * nr_running and nr_context_switches:
2354 * externally visible scheduler statistics: current number of runnable
2355 * threads, total number of context switches performed since bootup.
2357 unsigned long nr_running(void)
2359 unsigned long i, sum = 0;
2361 for_each_online_cpu(i)
2362 sum += cpu_rq(i)->nr_running;
2367 unsigned long long nr_context_switches(void)
2370 unsigned long long sum = 0;
2372 for_each_possible_cpu(i)
2373 sum += cpu_rq(i)->nr_switches;
2378 unsigned long nr_iowait(void)
2380 unsigned long i, sum = 0;
2382 for_each_possible_cpu(i)
2383 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2388 unsigned long nr_iowait_cpu(int cpu)
2390 struct rq *this = cpu_rq(cpu);
2391 return atomic_read(&this->nr_iowait);
2397 * sched_exec - execve() is a valuable balancing opportunity, because at
2398 * this point the task has the smallest effective memory and cache footprint.
2400 void sched_exec(void)
2402 struct task_struct *p = current;
2403 unsigned long flags;
2406 raw_spin_lock_irqsave(&p->pi_lock, flags);
2407 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2408 if (dest_cpu == smp_processor_id())
2411 if (likely(cpu_active(dest_cpu))) {
2412 struct migration_arg arg = { p, dest_cpu };
2414 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2415 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2419 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2424 DEFINE_PER_CPU(struct kernel_stat, kstat);
2425 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2427 EXPORT_PER_CPU_SYMBOL(kstat);
2428 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2431 * Return any ns on the sched_clock that have not yet been accounted in
2432 * @p in case that task is currently running.
2434 * Called with task_rq_lock() held on @rq.
2436 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2440 if (task_current(rq, p)) {
2441 update_rq_clock(rq);
2442 ns = rq_clock_task(rq) - p->se.exec_start;
2450 unsigned long long task_delta_exec(struct task_struct *p)
2452 unsigned long flags;
2456 rq = task_rq_lock(p, &flags);
2457 ns = do_task_delta_exec(p, rq);
2458 task_rq_unlock(rq, p, &flags);
2464 * Return accounted runtime for the task.
2465 * In case the task is currently running, return the runtime plus current's
2466 * pending runtime that have not been accounted yet.
2468 unsigned long long task_sched_runtime(struct task_struct *p)
2470 unsigned long flags;
2474 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2476 * 64-bit doesn't need locks to atomically read a 64bit value.
2477 * So we have a optimization chance when the task's delta_exec is 0.
2478 * Reading ->on_cpu is racy, but this is ok.
2480 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2481 * If we race with it entering cpu, unaccounted time is 0. This is
2482 * indistinguishable from the read occurring a few cycles earlier.
2485 return p->se.sum_exec_runtime;
2488 rq = task_rq_lock(p, &flags);
2489 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2490 task_rq_unlock(rq, p, &flags);
2496 * This function gets called by the timer code, with HZ frequency.
2497 * We call it with interrupts disabled.
2499 void scheduler_tick(void)
2501 int cpu = smp_processor_id();
2502 struct rq *rq = cpu_rq(cpu);
2503 struct task_struct *curr = rq->curr;
2507 raw_spin_lock(&rq->lock);
2508 update_rq_clock(rq);
2509 curr->sched_class->task_tick(rq, curr, 0);
2510 update_cpu_load_active(rq);
2511 raw_spin_unlock(&rq->lock);
2513 perf_event_task_tick();
2516 rq->idle_balance = idle_cpu(cpu);
2517 trigger_load_balance(rq);
2519 rq_last_tick_reset(rq);
2522 #ifdef CONFIG_NO_HZ_FULL
2524 * scheduler_tick_max_deferment
2526 * Keep at least one tick per second when a single
2527 * active task is running because the scheduler doesn't
2528 * yet completely support full dynticks environment.
2530 * This makes sure that uptime, CFS vruntime, load
2531 * balancing, etc... continue to move forward, even
2532 * with a very low granularity.
2534 * Return: Maximum deferment in nanoseconds.
2536 u64 scheduler_tick_max_deferment(void)
2538 struct rq *rq = this_rq();
2539 unsigned long next, now = ACCESS_ONCE(jiffies);
2541 next = rq->last_sched_tick + HZ;
2543 if (time_before_eq(next, now))
2546 return jiffies_to_nsecs(next - now);
2550 notrace unsigned long get_parent_ip(unsigned long addr)
2552 if (in_lock_functions(addr)) {
2553 addr = CALLER_ADDR2;
2554 if (in_lock_functions(addr))
2555 addr = CALLER_ADDR3;
2560 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2561 defined(CONFIG_PREEMPT_TRACER))
2563 void preempt_count_add(int val)
2565 #ifdef CONFIG_DEBUG_PREEMPT
2569 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2572 __preempt_count_add(val);
2573 #ifdef CONFIG_DEBUG_PREEMPT
2575 * Spinlock count overflowing soon?
2577 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2580 if (preempt_count() == val) {
2581 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2582 #ifdef CONFIG_DEBUG_PREEMPT
2583 current->preempt_disable_ip = ip;
2585 trace_preempt_off(CALLER_ADDR0, ip);
2588 EXPORT_SYMBOL(preempt_count_add);
2589 NOKPROBE_SYMBOL(preempt_count_add);
2591 void preempt_count_sub(int val)
2593 #ifdef CONFIG_DEBUG_PREEMPT
2597 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2600 * Is the spinlock portion underflowing?
2602 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2603 !(preempt_count() & PREEMPT_MASK)))
2607 if (preempt_count() == val)
2608 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2609 __preempt_count_sub(val);
2611 EXPORT_SYMBOL(preempt_count_sub);
2612 NOKPROBE_SYMBOL(preempt_count_sub);
2617 * Print scheduling while atomic bug:
2619 static noinline void __schedule_bug(struct task_struct *prev)
2621 if (oops_in_progress)
2624 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2625 prev->comm, prev->pid, preempt_count());
2627 debug_show_held_locks(prev);
2629 if (irqs_disabled())
2630 print_irqtrace_events(prev);
2631 #ifdef CONFIG_DEBUG_PREEMPT
2632 if (in_atomic_preempt_off()) {
2633 pr_err("Preemption disabled at:");
2634 print_ip_sym(current->preempt_disable_ip);
2639 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2643 * Various schedule()-time debugging checks and statistics:
2645 static inline void schedule_debug(struct task_struct *prev)
2648 * Test if we are atomic. Since do_exit() needs to call into
2649 * schedule() atomically, we ignore that path. Otherwise whine
2650 * if we are scheduling when we should not.
2652 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2653 __schedule_bug(prev);
2656 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2658 schedstat_inc(this_rq(), sched_count);
2662 * Pick up the highest-prio task:
2664 static inline struct task_struct *
2665 pick_next_task(struct rq *rq, struct task_struct *prev)
2667 const struct sched_class *class = &fair_sched_class;
2668 struct task_struct *p;
2671 * Optimization: we know that if all tasks are in
2672 * the fair class we can call that function directly:
2674 if (likely(prev->sched_class == class &&
2675 rq->nr_running == rq->cfs.h_nr_running)) {
2676 p = fair_sched_class.pick_next_task(rq, prev);
2677 if (unlikely(p == RETRY_TASK))
2680 /* assumes fair_sched_class->next == idle_sched_class */
2682 p = idle_sched_class.pick_next_task(rq, prev);
2688 for_each_class(class) {
2689 p = class->pick_next_task(rq, prev);
2691 if (unlikely(p == RETRY_TASK))
2697 BUG(); /* the idle class will always have a runnable task */
2701 * __schedule() is the main scheduler function.
2703 * The main means of driving the scheduler and thus entering this function are:
2705 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2707 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2708 * paths. For example, see arch/x86/entry_64.S.
2710 * To drive preemption between tasks, the scheduler sets the flag in timer
2711 * interrupt handler scheduler_tick().
2713 * 3. Wakeups don't really cause entry into schedule(). They add a
2714 * task to the run-queue and that's it.
2716 * Now, if the new task added to the run-queue preempts the current
2717 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2718 * called on the nearest possible occasion:
2720 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2722 * - in syscall or exception context, at the next outmost
2723 * preempt_enable(). (this might be as soon as the wake_up()'s
2726 * - in IRQ context, return from interrupt-handler to
2727 * preemptible context
2729 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2732 * - cond_resched() call
2733 * - explicit schedule() call
2734 * - return from syscall or exception to user-space
2735 * - return from interrupt-handler to user-space
2737 static void __sched __schedule(void)
2739 struct task_struct *prev, *next;
2740 unsigned long *switch_count;
2746 cpu = smp_processor_id();
2748 rcu_note_context_switch(cpu);
2751 schedule_debug(prev);
2753 if (sched_feat(HRTICK))
2757 * Make sure that signal_pending_state()->signal_pending() below
2758 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2759 * done by the caller to avoid the race with signal_wake_up().
2761 smp_mb__before_spinlock();
2762 raw_spin_lock_irq(&rq->lock);
2764 switch_count = &prev->nivcsw;
2765 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2766 if (unlikely(signal_pending_state(prev->state, prev))) {
2767 prev->state = TASK_RUNNING;
2769 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2773 * If a worker went to sleep, notify and ask workqueue
2774 * whether it wants to wake up a task to maintain
2777 if (prev->flags & PF_WQ_WORKER) {
2778 struct task_struct *to_wakeup;
2780 to_wakeup = wq_worker_sleeping(prev, cpu);
2782 try_to_wake_up_local(to_wakeup);
2785 switch_count = &prev->nvcsw;
2788 if (prev->on_rq || rq->skip_clock_update < 0)
2789 update_rq_clock(rq);
2791 next = pick_next_task(rq, prev);
2792 clear_tsk_need_resched(prev);
2793 clear_preempt_need_resched();
2794 rq->skip_clock_update = 0;
2796 if (likely(prev != next)) {
2801 context_switch(rq, prev, next); /* unlocks the rq */
2803 * The context switch have flipped the stack from under us
2804 * and restored the local variables which were saved when
2805 * this task called schedule() in the past. prev == current
2806 * is still correct, but it can be moved to another cpu/rq.
2808 cpu = smp_processor_id();
2811 raw_spin_unlock_irq(&rq->lock);
2815 sched_preempt_enable_no_resched();
2820 static inline void sched_submit_work(struct task_struct *tsk)
2822 if (!tsk->state || tsk_is_pi_blocked(tsk))
2825 * If we are going to sleep and we have plugged IO queued,
2826 * make sure to submit it to avoid deadlocks.
2828 if (blk_needs_flush_plug(tsk))
2829 blk_schedule_flush_plug(tsk);
2832 asmlinkage __visible void __sched schedule(void)
2834 struct task_struct *tsk = current;
2836 sched_submit_work(tsk);
2839 EXPORT_SYMBOL(schedule);
2841 #ifdef CONFIG_CONTEXT_TRACKING
2842 asmlinkage __visible void __sched schedule_user(void)
2845 * If we come here after a random call to set_need_resched(),
2846 * or we have been woken up remotely but the IPI has not yet arrived,
2847 * we haven't yet exited the RCU idle mode. Do it here manually until
2848 * we find a better solution.
2857 * schedule_preempt_disabled - called with preemption disabled
2859 * Returns with preemption disabled. Note: preempt_count must be 1
2861 void __sched schedule_preempt_disabled(void)
2863 sched_preempt_enable_no_resched();
2868 #ifdef CONFIG_PREEMPT
2870 * this is the entry point to schedule() from in-kernel preemption
2871 * off of preempt_enable. Kernel preemptions off return from interrupt
2872 * occur there and call schedule directly.
2874 asmlinkage __visible void __sched notrace preempt_schedule(void)
2877 * If there is a non-zero preempt_count or interrupts are disabled,
2878 * we do not want to preempt the current task. Just return..
2880 if (likely(!preemptible()))
2884 __preempt_count_add(PREEMPT_ACTIVE);
2886 __preempt_count_sub(PREEMPT_ACTIVE);
2889 * Check again in case we missed a preemption opportunity
2890 * between schedule and now.
2893 } while (need_resched());
2895 NOKPROBE_SYMBOL(preempt_schedule);
2896 EXPORT_SYMBOL(preempt_schedule);
2897 #endif /* CONFIG_PREEMPT */
2900 * this is the entry point to schedule() from kernel preemption
2901 * off of irq context.
2902 * Note, that this is called and return with irqs disabled. This will
2903 * protect us against recursive calling from irq.
2905 asmlinkage __visible void __sched preempt_schedule_irq(void)
2907 enum ctx_state prev_state;
2909 /* Catch callers which need to be fixed */
2910 BUG_ON(preempt_count() || !irqs_disabled());
2912 prev_state = exception_enter();
2915 __preempt_count_add(PREEMPT_ACTIVE);
2918 local_irq_disable();
2919 __preempt_count_sub(PREEMPT_ACTIVE);
2922 * Check again in case we missed a preemption opportunity
2923 * between schedule and now.
2926 } while (need_resched());
2928 exception_exit(prev_state);
2931 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2934 return try_to_wake_up(curr->private, mode, wake_flags);
2936 EXPORT_SYMBOL(default_wake_function);
2938 #ifdef CONFIG_RT_MUTEXES
2941 * rt_mutex_setprio - set the current priority of a task
2943 * @prio: prio value (kernel-internal form)
2945 * This function changes the 'effective' priority of a task. It does
2946 * not touch ->normal_prio like __setscheduler().
2948 * Used by the rt_mutex code to implement priority inheritance
2949 * logic. Call site only calls if the priority of the task changed.
2951 void rt_mutex_setprio(struct task_struct *p, int prio)
2953 int oldprio, on_rq, running, enqueue_flag = 0;
2955 const struct sched_class *prev_class;
2957 BUG_ON(prio > MAX_PRIO);
2959 rq = __task_rq_lock(p);
2962 * Idle task boosting is a nono in general. There is one
2963 * exception, when PREEMPT_RT and NOHZ is active:
2965 * The idle task calls get_next_timer_interrupt() and holds
2966 * the timer wheel base->lock on the CPU and another CPU wants
2967 * to access the timer (probably to cancel it). We can safely
2968 * ignore the boosting request, as the idle CPU runs this code
2969 * with interrupts disabled and will complete the lock
2970 * protected section without being interrupted. So there is no
2971 * real need to boost.
2973 if (unlikely(p == rq->idle)) {
2974 WARN_ON(p != rq->curr);
2975 WARN_ON(p->pi_blocked_on);
2979 trace_sched_pi_setprio(p, prio);
2980 p->pi_top_task = rt_mutex_get_top_task(p);
2982 prev_class = p->sched_class;
2984 running = task_current(rq, p);
2986 dequeue_task(rq, p, 0);
2988 p->sched_class->put_prev_task(rq, p);
2991 * Boosting condition are:
2992 * 1. -rt task is running and holds mutex A
2993 * --> -dl task blocks on mutex A
2995 * 2. -dl task is running and holds mutex A
2996 * --> -dl task blocks on mutex A and could preempt the
2999 if (dl_prio(prio)) {
3000 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
3001 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
3002 p->dl.dl_boosted = 1;
3003 p->dl.dl_throttled = 0;
3004 enqueue_flag = ENQUEUE_REPLENISH;
3006 p->dl.dl_boosted = 0;
3007 p->sched_class = &dl_sched_class;
3008 } else if (rt_prio(prio)) {
3009 if (dl_prio(oldprio))
3010 p->dl.dl_boosted = 0;
3012 enqueue_flag = ENQUEUE_HEAD;
3013 p->sched_class = &rt_sched_class;
3015 if (dl_prio(oldprio))
3016 p->dl.dl_boosted = 0;
3017 p->sched_class = &fair_sched_class;
3023 p->sched_class->set_curr_task(rq);
3025 enqueue_task(rq, p, enqueue_flag);
3027 check_class_changed(rq, p, prev_class, oldprio);
3029 __task_rq_unlock(rq);
3033 void set_user_nice(struct task_struct *p, long nice)
3035 int old_prio, delta, on_rq;
3036 unsigned long flags;
3039 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3042 * We have to be careful, if called from sys_setpriority(),
3043 * the task might be in the middle of scheduling on another CPU.
3045 rq = task_rq_lock(p, &flags);
3047 * The RT priorities are set via sched_setscheduler(), but we still
3048 * allow the 'normal' nice value to be set - but as expected
3049 * it wont have any effect on scheduling until the task is
3050 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3052 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3053 p->static_prio = NICE_TO_PRIO(nice);
3058 dequeue_task(rq, p, 0);
3060 p->static_prio = NICE_TO_PRIO(nice);
3063 p->prio = effective_prio(p);
3064 delta = p->prio - old_prio;
3067 enqueue_task(rq, p, 0);
3069 * If the task increased its priority or is running and
3070 * lowered its priority, then reschedule its CPU:
3072 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3073 resched_task(rq->curr);
3076 task_rq_unlock(rq, p, &flags);
3078 EXPORT_SYMBOL(set_user_nice);
3081 * can_nice - check if a task can reduce its nice value
3085 int can_nice(const struct task_struct *p, const int nice)
3087 /* convert nice value [19,-20] to rlimit style value [1,40] */
3088 int nice_rlim = nice_to_rlimit(nice);
3090 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3091 capable(CAP_SYS_NICE));
3094 #ifdef __ARCH_WANT_SYS_NICE
3097 * sys_nice - change the priority of the current process.
3098 * @increment: priority increment
3100 * sys_setpriority is a more generic, but much slower function that
3101 * does similar things.
3103 SYSCALL_DEFINE1(nice, int, increment)
3108 * Setpriority might change our priority at the same moment.
3109 * We don't have to worry. Conceptually one call occurs first
3110 * and we have a single winner.
3112 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3113 nice = task_nice(current) + increment;
3115 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3116 if (increment < 0 && !can_nice(current, nice))
3119 retval = security_task_setnice(current, nice);
3123 set_user_nice(current, nice);
3130 * task_prio - return the priority value of a given task.
3131 * @p: the task in question.
3133 * Return: The priority value as seen by users in /proc.
3134 * RT tasks are offset by -200. Normal tasks are centered
3135 * around 0, value goes from -16 to +15.
3137 int task_prio(const struct task_struct *p)
3139 return p->prio - MAX_RT_PRIO;
3143 * idle_cpu - is a given cpu idle currently?
3144 * @cpu: the processor in question.
3146 * Return: 1 if the CPU is currently idle. 0 otherwise.
3148 int idle_cpu(int cpu)
3150 struct rq *rq = cpu_rq(cpu);
3152 if (rq->curr != rq->idle)
3159 if (!llist_empty(&rq->wake_list))
3167 * idle_task - return the idle task for a given cpu.
3168 * @cpu: the processor in question.
3170 * Return: The idle task for the cpu @cpu.
3172 struct task_struct *idle_task(int cpu)
3174 return cpu_rq(cpu)->idle;
3178 * find_process_by_pid - find a process with a matching PID value.
3179 * @pid: the pid in question.
3181 * The task of @pid, if found. %NULL otherwise.
3183 static struct task_struct *find_process_by_pid(pid_t pid)
3185 return pid ? find_task_by_vpid(pid) : current;
3189 * This function initializes the sched_dl_entity of a newly becoming
3190 * SCHED_DEADLINE task.
3192 * Only the static values are considered here, the actual runtime and the
3193 * absolute deadline will be properly calculated when the task is enqueued
3194 * for the first time with its new policy.
3197 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3199 struct sched_dl_entity *dl_se = &p->dl;
3201 init_dl_task_timer(dl_se);
3202 dl_se->dl_runtime = attr->sched_runtime;
3203 dl_se->dl_deadline = attr->sched_deadline;
3204 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3205 dl_se->flags = attr->sched_flags;
3206 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3207 dl_se->dl_throttled = 0;
3209 dl_se->dl_yielded = 0;
3212 static void __setscheduler_params(struct task_struct *p,
3213 const struct sched_attr *attr)
3215 int policy = attr->sched_policy;
3217 if (policy == -1) /* setparam */
3222 if (dl_policy(policy))
3223 __setparam_dl(p, attr);
3224 else if (fair_policy(policy))
3225 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3228 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3229 * !rt_policy. Always setting this ensures that things like
3230 * getparam()/getattr() don't report silly values for !rt tasks.
3232 p->rt_priority = attr->sched_priority;
3233 p->normal_prio = normal_prio(p);
3237 /* Actually do priority change: must hold pi & rq lock. */
3238 static void __setscheduler(struct rq *rq, struct task_struct *p,
3239 const struct sched_attr *attr)
3241 __setscheduler_params(p, attr);
3244 * If we get here, there was no pi waiters boosting the
3245 * task. It is safe to use the normal prio.
3247 p->prio = normal_prio(p);
3249 if (dl_prio(p->prio))
3250 p->sched_class = &dl_sched_class;
3251 else if (rt_prio(p->prio))
3252 p->sched_class = &rt_sched_class;
3254 p->sched_class = &fair_sched_class;
3258 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3260 struct sched_dl_entity *dl_se = &p->dl;
3262 attr->sched_priority = p->rt_priority;
3263 attr->sched_runtime = dl_se->dl_runtime;
3264 attr->sched_deadline = dl_se->dl_deadline;
3265 attr->sched_period = dl_se->dl_period;
3266 attr->sched_flags = dl_se->flags;
3270 * This function validates the new parameters of a -deadline task.
3271 * We ask for the deadline not being zero, and greater or equal
3272 * than the runtime, as well as the period of being zero or
3273 * greater than deadline. Furthermore, we have to be sure that
3274 * user parameters are above the internal resolution of 1us (we
3275 * check sched_runtime only since it is always the smaller one) and
3276 * below 2^63 ns (we have to check both sched_deadline and
3277 * sched_period, as the latter can be zero).
3280 __checkparam_dl(const struct sched_attr *attr)
3283 if (attr->sched_deadline == 0)
3287 * Since we truncate DL_SCALE bits, make sure we're at least
3290 if (attr->sched_runtime < (1ULL << DL_SCALE))
3294 * Since we use the MSB for wrap-around and sign issues, make
3295 * sure it's not set (mind that period can be equal to zero).
3297 if (attr->sched_deadline & (1ULL << 63) ||
3298 attr->sched_period & (1ULL << 63))
3301 /* runtime <= deadline <= period (if period != 0) */
3302 if ((attr->sched_period != 0 &&
3303 attr->sched_period < attr->sched_deadline) ||
3304 attr->sched_deadline < attr->sched_runtime)
3311 * check the target process has a UID that matches the current process's
3313 static bool check_same_owner(struct task_struct *p)
3315 const struct cred *cred = current_cred(), *pcred;
3319 pcred = __task_cred(p);
3320 match = (uid_eq(cred->euid, pcred->euid) ||
3321 uid_eq(cred->euid, pcred->uid));
3326 static int __sched_setscheduler(struct task_struct *p,
3327 const struct sched_attr *attr,
3330 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3331 MAX_RT_PRIO - 1 - attr->sched_priority;
3332 int retval, oldprio, oldpolicy = -1, on_rq, running;
3333 int policy = attr->sched_policy;
3334 unsigned long flags;
3335 const struct sched_class *prev_class;
3339 /* may grab non-irq protected spin_locks */
3340 BUG_ON(in_interrupt());
3342 /* double check policy once rq lock held */
3344 reset_on_fork = p->sched_reset_on_fork;
3345 policy = oldpolicy = p->policy;
3347 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3349 if (policy != SCHED_DEADLINE &&
3350 policy != SCHED_FIFO && policy != SCHED_RR &&
3351 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3352 policy != SCHED_IDLE)
3356 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3360 * Valid priorities for SCHED_FIFO and SCHED_RR are
3361 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3362 * SCHED_BATCH and SCHED_IDLE is 0.
3364 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3365 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3367 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3368 (rt_policy(policy) != (attr->sched_priority != 0)))
3372 * Allow unprivileged RT tasks to decrease priority:
3374 if (user && !capable(CAP_SYS_NICE)) {
3375 if (fair_policy(policy)) {
3376 if (attr->sched_nice < task_nice(p) &&
3377 !can_nice(p, attr->sched_nice))
3381 if (rt_policy(policy)) {
3382 unsigned long rlim_rtprio =
3383 task_rlimit(p, RLIMIT_RTPRIO);
3385 /* can't set/change the rt policy */
3386 if (policy != p->policy && !rlim_rtprio)
3389 /* can't increase priority */
3390 if (attr->sched_priority > p->rt_priority &&
3391 attr->sched_priority > rlim_rtprio)
3396 * Can't set/change SCHED_DEADLINE policy at all for now
3397 * (safest behavior); in the future we would like to allow
3398 * unprivileged DL tasks to increase their relative deadline
3399 * or reduce their runtime (both ways reducing utilization)
3401 if (dl_policy(policy))
3405 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3406 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3408 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3409 if (!can_nice(p, task_nice(p)))
3413 /* can't change other user's priorities */
3414 if (!check_same_owner(p))
3417 /* Normal users shall not reset the sched_reset_on_fork flag */
3418 if (p->sched_reset_on_fork && !reset_on_fork)
3423 retval = security_task_setscheduler(p);
3429 * make sure no PI-waiters arrive (or leave) while we are
3430 * changing the priority of the task:
3432 * To be able to change p->policy safely, the appropriate
3433 * runqueue lock must be held.
3435 rq = task_rq_lock(p, &flags);
3438 * Changing the policy of the stop threads its a very bad idea
3440 if (p == rq->stop) {
3441 task_rq_unlock(rq, p, &flags);
3446 * If not changing anything there's no need to proceed further,
3447 * but store a possible modification of reset_on_fork.
3449 if (unlikely(policy == p->policy)) {
3450 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3452 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3454 if (dl_policy(policy))
3457 p->sched_reset_on_fork = reset_on_fork;
3458 task_rq_unlock(rq, p, &flags);
3464 #ifdef CONFIG_RT_GROUP_SCHED
3466 * Do not allow realtime tasks into groups that have no runtime
3469 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3470 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3471 !task_group_is_autogroup(task_group(p))) {
3472 task_rq_unlock(rq, p, &flags);
3477 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3478 cpumask_t *span = rq->rd->span;
3481 * Don't allow tasks with an affinity mask smaller than
3482 * the entire root_domain to become SCHED_DEADLINE. We
3483 * will also fail if there's no bandwidth available.
3485 if (!cpumask_subset(span, &p->cpus_allowed) ||
3486 rq->rd->dl_bw.bw == 0) {
3487 task_rq_unlock(rq, p, &flags);
3494 /* recheck policy now with rq lock held */
3495 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3496 policy = oldpolicy = -1;
3497 task_rq_unlock(rq, p, &flags);
3502 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3503 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3506 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3507 task_rq_unlock(rq, p, &flags);
3511 p->sched_reset_on_fork = reset_on_fork;
3515 * Special case for priority boosted tasks.
3517 * If the new priority is lower or equal (user space view)
3518 * than the current (boosted) priority, we just store the new
3519 * normal parameters and do not touch the scheduler class and
3520 * the runqueue. This will be done when the task deboost
3523 if (rt_mutex_check_prio(p, newprio)) {
3524 __setscheduler_params(p, attr);
3525 task_rq_unlock(rq, p, &flags);
3530 running = task_current(rq, p);
3532 dequeue_task(rq, p, 0);
3534 p->sched_class->put_prev_task(rq, p);
3536 prev_class = p->sched_class;
3537 __setscheduler(rq, p, attr);
3540 p->sched_class->set_curr_task(rq);
3543 * We enqueue to tail when the priority of a task is
3544 * increased (user space view).
3546 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3549 check_class_changed(rq, p, prev_class, oldprio);
3550 task_rq_unlock(rq, p, &flags);
3552 rt_mutex_adjust_pi(p);
3557 static int _sched_setscheduler(struct task_struct *p, int policy,
3558 const struct sched_param *param, bool check)
3560 struct sched_attr attr = {
3561 .sched_policy = policy,
3562 .sched_priority = param->sched_priority,
3563 .sched_nice = PRIO_TO_NICE(p->static_prio),
3567 * Fixup the legacy SCHED_RESET_ON_FORK hack
3569 if (policy & SCHED_RESET_ON_FORK) {
3570 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3571 policy &= ~SCHED_RESET_ON_FORK;
3572 attr.sched_policy = policy;
3575 return __sched_setscheduler(p, &attr, check);
3578 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3579 * @p: the task in question.
3580 * @policy: new policy.
3581 * @param: structure containing the new RT priority.
3583 * Return: 0 on success. An error code otherwise.
3585 * NOTE that the task may be already dead.
3587 int sched_setscheduler(struct task_struct *p, int policy,
3588 const struct sched_param *param)
3590 return _sched_setscheduler(p, policy, param, true);
3592 EXPORT_SYMBOL_GPL(sched_setscheduler);
3594 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3596 return __sched_setscheduler(p, attr, true);
3598 EXPORT_SYMBOL_GPL(sched_setattr);
3601 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3602 * @p: the task in question.
3603 * @policy: new policy.
3604 * @param: structure containing the new RT priority.
3606 * Just like sched_setscheduler, only don't bother checking if the
3607 * current context has permission. For example, this is needed in
3608 * stop_machine(): we create temporary high priority worker threads,
3609 * but our caller might not have that capability.
3611 * Return: 0 on success. An error code otherwise.
3613 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3614 const struct sched_param *param)
3616 return _sched_setscheduler(p, policy, param, false);
3620 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3622 struct sched_param lparam;
3623 struct task_struct *p;
3626 if (!param || pid < 0)
3628 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3633 p = find_process_by_pid(pid);
3635 retval = sched_setscheduler(p, policy, &lparam);
3642 * Mimics kernel/events/core.c perf_copy_attr().
3644 static int sched_copy_attr(struct sched_attr __user *uattr,
3645 struct sched_attr *attr)
3650 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3654 * zero the full structure, so that a short copy will be nice.
3656 memset(attr, 0, sizeof(*attr));
3658 ret = get_user(size, &uattr->size);
3662 if (size > PAGE_SIZE) /* silly large */
3665 if (!size) /* abi compat */
3666 size = SCHED_ATTR_SIZE_VER0;
3668 if (size < SCHED_ATTR_SIZE_VER0)
3672 * If we're handed a bigger struct than we know of,
3673 * ensure all the unknown bits are 0 - i.e. new
3674 * user-space does not rely on any kernel feature
3675 * extensions we dont know about yet.
3677 if (size > sizeof(*attr)) {
3678 unsigned char __user *addr;
3679 unsigned char __user *end;
3682 addr = (void __user *)uattr + sizeof(*attr);
3683 end = (void __user *)uattr + size;
3685 for (; addr < end; addr++) {
3686 ret = get_user(val, addr);
3692 size = sizeof(*attr);
3695 ret = copy_from_user(attr, uattr, size);
3700 * XXX: do we want to be lenient like existing syscalls; or do we want
3701 * to be strict and return an error on out-of-bounds values?
3703 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3708 put_user(sizeof(*attr), &uattr->size);
3713 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3714 * @pid: the pid in question.
3715 * @policy: new policy.
3716 * @param: structure containing the new RT priority.
3718 * Return: 0 on success. An error code otherwise.
3720 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3721 struct sched_param __user *, param)
3723 /* negative values for policy are not valid */
3727 return do_sched_setscheduler(pid, policy, param);
3731 * sys_sched_setparam - set/change the RT priority of a thread
3732 * @pid: the pid in question.
3733 * @param: structure containing the new RT priority.
3735 * Return: 0 on success. An error code otherwise.
3737 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3739 return do_sched_setscheduler(pid, -1, param);
3743 * sys_sched_setattr - same as above, but with extended sched_attr
3744 * @pid: the pid in question.
3745 * @uattr: structure containing the extended parameters.
3746 * @flags: for future extension.
3748 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3749 unsigned int, flags)
3751 struct sched_attr attr;
3752 struct task_struct *p;
3755 if (!uattr || pid < 0 || flags)
3758 retval = sched_copy_attr(uattr, &attr);
3762 if ((int)attr.sched_policy < 0)
3767 p = find_process_by_pid(pid);
3769 retval = sched_setattr(p, &attr);
3776 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3777 * @pid: the pid in question.
3779 * Return: On success, the policy of the thread. Otherwise, a negative error
3782 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3784 struct task_struct *p;
3792 p = find_process_by_pid(pid);
3794 retval = security_task_getscheduler(p);
3797 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3804 * sys_sched_getparam - get the RT priority of a thread
3805 * @pid: the pid in question.
3806 * @param: structure containing the RT priority.
3808 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3811 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3813 struct sched_param lp = { .sched_priority = 0 };
3814 struct task_struct *p;
3817 if (!param || pid < 0)
3821 p = find_process_by_pid(pid);
3826 retval = security_task_getscheduler(p);
3830 if (task_has_rt_policy(p))
3831 lp.sched_priority = p->rt_priority;
3835 * This one might sleep, we cannot do it with a spinlock held ...
3837 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3846 static int sched_read_attr(struct sched_attr __user *uattr,
3847 struct sched_attr *attr,
3852 if (!access_ok(VERIFY_WRITE, uattr, usize))
3856 * If we're handed a smaller struct than we know of,
3857 * ensure all the unknown bits are 0 - i.e. old
3858 * user-space does not get uncomplete information.
3860 if (usize < sizeof(*attr)) {
3861 unsigned char *addr;
3864 addr = (void *)attr + usize;
3865 end = (void *)attr + sizeof(*attr);
3867 for (; addr < end; addr++) {
3875 ret = copy_to_user(uattr, attr, attr->size);
3883 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3884 * @pid: the pid in question.
3885 * @uattr: structure containing the extended parameters.
3886 * @size: sizeof(attr) for fwd/bwd comp.
3887 * @flags: for future extension.
3889 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3890 unsigned int, size, unsigned int, flags)
3892 struct sched_attr attr = {
3893 .size = sizeof(struct sched_attr),
3895 struct task_struct *p;
3898 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3899 size < SCHED_ATTR_SIZE_VER0 || flags)
3903 p = find_process_by_pid(pid);
3908 retval = security_task_getscheduler(p);
3912 attr.sched_policy = p->policy;
3913 if (p->sched_reset_on_fork)
3914 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3915 if (task_has_dl_policy(p))
3916 __getparam_dl(p, &attr);
3917 else if (task_has_rt_policy(p))
3918 attr.sched_priority = p->rt_priority;
3920 attr.sched_nice = task_nice(p);
3924 retval = sched_read_attr(uattr, &attr, size);
3932 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3934 cpumask_var_t cpus_allowed, new_mask;
3935 struct task_struct *p;
3940 p = find_process_by_pid(pid);
3946 /* Prevent p going away */
3950 if (p->flags & PF_NO_SETAFFINITY) {
3954 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3958 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3960 goto out_free_cpus_allowed;
3963 if (!check_same_owner(p)) {
3965 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3972 retval = security_task_setscheduler(p);
3977 cpuset_cpus_allowed(p, cpus_allowed);
3978 cpumask_and(new_mask, in_mask, cpus_allowed);
3981 * Since bandwidth control happens on root_domain basis,
3982 * if admission test is enabled, we only admit -deadline
3983 * tasks allowed to run on all the CPUs in the task's
3987 if (task_has_dl_policy(p)) {
3988 const struct cpumask *span = task_rq(p)->rd->span;
3990 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3997 retval = set_cpus_allowed_ptr(p, new_mask);
4000 cpuset_cpus_allowed(p, cpus_allowed);
4001 if (!cpumask_subset(new_mask, cpus_allowed)) {
4003 * We must have raced with a concurrent cpuset
4004 * update. Just reset the cpus_allowed to the
4005 * cpuset's cpus_allowed
4007 cpumask_copy(new_mask, cpus_allowed);
4012 free_cpumask_var(new_mask);
4013 out_free_cpus_allowed:
4014 free_cpumask_var(cpus_allowed);
4020 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4021 struct cpumask *new_mask)
4023 if (len < cpumask_size())
4024 cpumask_clear(new_mask);
4025 else if (len > cpumask_size())
4026 len = cpumask_size();
4028 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4032 * sys_sched_setaffinity - set the cpu affinity of a process
4033 * @pid: pid of the process
4034 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4035 * @user_mask_ptr: user-space pointer to the new cpu mask
4037 * Return: 0 on success. An error code otherwise.
4039 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4040 unsigned long __user *, user_mask_ptr)
4042 cpumask_var_t new_mask;
4045 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4048 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4050 retval = sched_setaffinity(pid, new_mask);
4051 free_cpumask_var(new_mask);
4055 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4057 struct task_struct *p;
4058 unsigned long flags;
4064 p = find_process_by_pid(pid);
4068 retval = security_task_getscheduler(p);
4072 raw_spin_lock_irqsave(&p->pi_lock, flags);
4073 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4074 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4083 * sys_sched_getaffinity - get the cpu affinity of a process
4084 * @pid: pid of the process
4085 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4086 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4088 * Return: 0 on success. An error code otherwise.
4090 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4091 unsigned long __user *, user_mask_ptr)
4096 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4098 if (len & (sizeof(unsigned long)-1))
4101 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4104 ret = sched_getaffinity(pid, mask);
4106 size_t retlen = min_t(size_t, len, cpumask_size());
4108 if (copy_to_user(user_mask_ptr, mask, retlen))
4113 free_cpumask_var(mask);
4119 * sys_sched_yield - yield the current processor to other threads.
4121 * This function yields the current CPU to other tasks. If there are no
4122 * other threads running on this CPU then this function will return.
4126 SYSCALL_DEFINE0(sched_yield)
4128 struct rq *rq = this_rq_lock();
4130 schedstat_inc(rq, yld_count);
4131 current->sched_class->yield_task(rq);
4134 * Since we are going to call schedule() anyway, there's
4135 * no need to preempt or enable interrupts:
4137 __release(rq->lock);
4138 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4139 do_raw_spin_unlock(&rq->lock);
4140 sched_preempt_enable_no_resched();
4147 static void __cond_resched(void)
4149 __preempt_count_add(PREEMPT_ACTIVE);
4151 __preempt_count_sub(PREEMPT_ACTIVE);
4154 int __sched _cond_resched(void)
4157 if (should_resched()) {
4163 EXPORT_SYMBOL(_cond_resched);
4166 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4167 * call schedule, and on return reacquire the lock.
4169 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4170 * operations here to prevent schedule() from being called twice (once via
4171 * spin_unlock(), once by hand).
4173 int __cond_resched_lock(spinlock_t *lock)
4175 bool need_rcu_resched = rcu_should_resched();
4176 int resched = should_resched();
4179 lockdep_assert_held(lock);
4181 if (spin_needbreak(lock) || resched || need_rcu_resched) {
4185 else if (unlikely(need_rcu_resched))
4194 EXPORT_SYMBOL(__cond_resched_lock);
4196 int __sched __cond_resched_softirq(void)
4198 BUG_ON(!in_softirq());
4200 rcu_cond_resched(); /* BH disabled OK, just recording QSes. */
4201 if (should_resched()) {
4209 EXPORT_SYMBOL(__cond_resched_softirq);
4212 * yield - yield the current processor to other threads.
4214 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4216 * The scheduler is at all times free to pick the calling task as the most
4217 * eligible task to run, if removing the yield() call from your code breaks
4218 * it, its already broken.
4220 * Typical broken usage is:
4225 * where one assumes that yield() will let 'the other' process run that will
4226 * make event true. If the current task is a SCHED_FIFO task that will never
4227 * happen. Never use yield() as a progress guarantee!!
4229 * If you want to use yield() to wait for something, use wait_event().
4230 * If you want to use yield() to be 'nice' for others, use cond_resched().
4231 * If you still want to use yield(), do not!
4233 void __sched yield(void)
4235 set_current_state(TASK_RUNNING);
4238 EXPORT_SYMBOL(yield);
4241 * yield_to - yield the current processor to another thread in
4242 * your thread group, or accelerate that thread toward the
4243 * processor it's on.
4245 * @preempt: whether task preemption is allowed or not
4247 * It's the caller's job to ensure that the target task struct
4248 * can't go away on us before we can do any checks.
4251 * true (>0) if we indeed boosted the target task.
4252 * false (0) if we failed to boost the target.
4253 * -ESRCH if there's no task to yield to.
4255 int __sched yield_to(struct task_struct *p, bool preempt)
4257 struct task_struct *curr = current;
4258 struct rq *rq, *p_rq;
4259 unsigned long flags;
4262 local_irq_save(flags);
4268 * If we're the only runnable task on the rq and target rq also
4269 * has only one task, there's absolutely no point in yielding.
4271 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4276 double_rq_lock(rq, p_rq);
4277 if (task_rq(p) != p_rq) {
4278 double_rq_unlock(rq, p_rq);
4282 if (!curr->sched_class->yield_to_task)
4285 if (curr->sched_class != p->sched_class)
4288 if (task_running(p_rq, p) || p->state)
4291 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4293 schedstat_inc(rq, yld_count);
4295 * Make p's CPU reschedule; pick_next_entity takes care of
4298 if (preempt && rq != p_rq)
4299 resched_task(p_rq->curr);
4303 double_rq_unlock(rq, p_rq);
4305 local_irq_restore(flags);
4312 EXPORT_SYMBOL_GPL(yield_to);
4315 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4316 * that process accounting knows that this is a task in IO wait state.
4318 void __sched io_schedule(void)
4320 struct rq *rq = raw_rq();
4322 delayacct_blkio_start();
4323 atomic_inc(&rq->nr_iowait);
4324 blk_flush_plug(current);
4325 current->in_iowait = 1;
4327 current->in_iowait = 0;
4328 atomic_dec(&rq->nr_iowait);
4329 delayacct_blkio_end();
4331 EXPORT_SYMBOL(io_schedule);
4333 long __sched io_schedule_timeout(long timeout)
4335 struct rq *rq = raw_rq();
4338 delayacct_blkio_start();
4339 atomic_inc(&rq->nr_iowait);
4340 blk_flush_plug(current);
4341 current->in_iowait = 1;
4342 ret = schedule_timeout(timeout);
4343 current->in_iowait = 0;
4344 atomic_dec(&rq->nr_iowait);
4345 delayacct_blkio_end();
4350 * sys_sched_get_priority_max - return maximum RT priority.
4351 * @policy: scheduling class.
4353 * Return: On success, this syscall returns the maximum
4354 * rt_priority that can be used by a given scheduling class.
4355 * On failure, a negative error code is returned.
4357 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4364 ret = MAX_USER_RT_PRIO-1;
4366 case SCHED_DEADLINE:
4377 * sys_sched_get_priority_min - return minimum RT priority.
4378 * @policy: scheduling class.
4380 * Return: On success, this syscall returns the minimum
4381 * rt_priority that can be used by a given scheduling class.
4382 * On failure, a negative error code is returned.
4384 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4393 case SCHED_DEADLINE:
4403 * sys_sched_rr_get_interval - return the default timeslice of a process.
4404 * @pid: pid of the process.
4405 * @interval: userspace pointer to the timeslice value.
4407 * this syscall writes the default timeslice value of a given process
4408 * into the user-space timespec buffer. A value of '0' means infinity.
4410 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4413 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4414 struct timespec __user *, interval)
4416 struct task_struct *p;
4417 unsigned int time_slice;
4418 unsigned long flags;
4428 p = find_process_by_pid(pid);
4432 retval = security_task_getscheduler(p);
4436 rq = task_rq_lock(p, &flags);
4438 if (p->sched_class->get_rr_interval)
4439 time_slice = p->sched_class->get_rr_interval(rq, p);
4440 task_rq_unlock(rq, p, &flags);
4443 jiffies_to_timespec(time_slice, &t);
4444 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4452 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4454 void sched_show_task(struct task_struct *p)
4456 unsigned long free = 0;
4460 state = p->state ? __ffs(p->state) + 1 : 0;
4461 printk(KERN_INFO "%-15.15s %c", p->comm,
4462 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4463 #if BITS_PER_LONG == 32
4464 if (state == TASK_RUNNING)
4465 printk(KERN_CONT " running ");
4467 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4469 if (state == TASK_RUNNING)
4470 printk(KERN_CONT " running task ");
4472 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4474 #ifdef CONFIG_DEBUG_STACK_USAGE
4475 free = stack_not_used(p);
4478 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4480 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4481 task_pid_nr(p), ppid,
4482 (unsigned long)task_thread_info(p)->flags);
4484 print_worker_info(KERN_INFO, p);
4485 show_stack(p, NULL);
4488 void show_state_filter(unsigned long state_filter)
4490 struct task_struct *g, *p;
4492 #if BITS_PER_LONG == 32
4494 " task PC stack pid father\n");
4497 " task PC stack pid father\n");
4500 do_each_thread(g, p) {
4502 * reset the NMI-timeout, listing all files on a slow
4503 * console might take a lot of time:
4505 touch_nmi_watchdog();
4506 if (!state_filter || (p->state & state_filter))
4508 } while_each_thread(g, p);
4510 touch_all_softlockup_watchdogs();
4512 #ifdef CONFIG_SCHED_DEBUG
4513 sysrq_sched_debug_show();
4517 * Only show locks if all tasks are dumped:
4520 debug_show_all_locks();
4523 void init_idle_bootup_task(struct task_struct *idle)
4525 idle->sched_class = &idle_sched_class;
4529 * init_idle - set up an idle thread for a given CPU
4530 * @idle: task in question
4531 * @cpu: cpu the idle task belongs to
4533 * NOTE: this function does not set the idle thread's NEED_RESCHED
4534 * flag, to make booting more robust.
4536 void init_idle(struct task_struct *idle, int cpu)
4538 struct rq *rq = cpu_rq(cpu);
4539 unsigned long flags;
4541 raw_spin_lock_irqsave(&rq->lock, flags);
4543 __sched_fork(0, idle);
4544 idle->state = TASK_RUNNING;
4545 idle->se.exec_start = sched_clock();
4547 do_set_cpus_allowed(idle, cpumask_of(cpu));
4549 * We're having a chicken and egg problem, even though we are
4550 * holding rq->lock, the cpu isn't yet set to this cpu so the
4551 * lockdep check in task_group() will fail.
4553 * Similar case to sched_fork(). / Alternatively we could
4554 * use task_rq_lock() here and obtain the other rq->lock.
4559 __set_task_cpu(idle, cpu);
4562 rq->curr = rq->idle = idle;
4564 #if defined(CONFIG_SMP)
4567 raw_spin_unlock_irqrestore(&rq->lock, flags);
4569 /* Set the preempt count _outside_ the spinlocks! */
4570 init_idle_preempt_count(idle, cpu);
4573 * The idle tasks have their own, simple scheduling class:
4575 idle->sched_class = &idle_sched_class;
4576 ftrace_graph_init_idle_task(idle, cpu);
4577 vtime_init_idle(idle, cpu);
4578 #if defined(CONFIG_SMP)
4579 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4584 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4586 if (p->sched_class && p->sched_class->set_cpus_allowed)
4587 p->sched_class->set_cpus_allowed(p, new_mask);
4589 cpumask_copy(&p->cpus_allowed, new_mask);
4590 p->nr_cpus_allowed = cpumask_weight(new_mask);
4594 * This is how migration works:
4596 * 1) we invoke migration_cpu_stop() on the target CPU using
4598 * 2) stopper starts to run (implicitly forcing the migrated thread
4600 * 3) it checks whether the migrated task is still in the wrong runqueue.
4601 * 4) if it's in the wrong runqueue then the migration thread removes
4602 * it and puts it into the right queue.
4603 * 5) stopper completes and stop_one_cpu() returns and the migration
4608 * Change a given task's CPU affinity. Migrate the thread to a
4609 * proper CPU and schedule it away if the CPU it's executing on
4610 * is removed from the allowed bitmask.
4612 * NOTE: the caller must have a valid reference to the task, the
4613 * task must not exit() & deallocate itself prematurely. The
4614 * call is not atomic; no spinlocks may be held.
4616 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4618 unsigned long flags;
4620 unsigned int dest_cpu;
4623 rq = task_rq_lock(p, &flags);
4625 if (cpumask_equal(&p->cpus_allowed, new_mask))
4628 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4633 do_set_cpus_allowed(p, new_mask);
4635 /* Can the task run on the task's current CPU? If so, we're done */
4636 if (cpumask_test_cpu(task_cpu(p), new_mask))
4639 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4641 struct migration_arg arg = { p, dest_cpu };
4642 /* Need help from migration thread: drop lock and wait. */
4643 task_rq_unlock(rq, p, &flags);
4644 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4645 tlb_migrate_finish(p->mm);
4649 task_rq_unlock(rq, p, &flags);
4653 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4656 * Move (not current) task off this cpu, onto dest cpu. We're doing
4657 * this because either it can't run here any more (set_cpus_allowed()
4658 * away from this CPU, or CPU going down), or because we're
4659 * attempting to rebalance this task on exec (sched_exec).
4661 * So we race with normal scheduler movements, but that's OK, as long
4662 * as the task is no longer on this CPU.
4664 * Returns non-zero if task was successfully migrated.
4666 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4668 struct rq *rq_dest, *rq_src;
4671 if (unlikely(!cpu_active(dest_cpu)))
4674 rq_src = cpu_rq(src_cpu);
4675 rq_dest = cpu_rq(dest_cpu);
4677 raw_spin_lock(&p->pi_lock);
4678 double_rq_lock(rq_src, rq_dest);
4679 /* Already moved. */
4680 if (task_cpu(p) != src_cpu)
4682 /* Affinity changed (again). */
4683 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4687 * If we're not on a rq, the next wake-up will ensure we're
4691 dequeue_task(rq_src, p, 0);
4692 set_task_cpu(p, dest_cpu);
4693 enqueue_task(rq_dest, p, 0);
4694 check_preempt_curr(rq_dest, p, 0);
4699 double_rq_unlock(rq_src, rq_dest);
4700 raw_spin_unlock(&p->pi_lock);
4704 #ifdef CONFIG_NUMA_BALANCING
4705 /* Migrate current task p to target_cpu */
4706 int migrate_task_to(struct task_struct *p, int target_cpu)
4708 struct migration_arg arg = { p, target_cpu };
4709 int curr_cpu = task_cpu(p);
4711 if (curr_cpu == target_cpu)
4714 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4717 /* TODO: This is not properly updating schedstats */
4719 trace_sched_move_numa(p, curr_cpu, target_cpu);
4720 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4724 * Requeue a task on a given node and accurately track the number of NUMA
4725 * tasks on the runqueues
4727 void sched_setnuma(struct task_struct *p, int nid)
4730 unsigned long flags;
4731 bool on_rq, running;
4733 rq = task_rq_lock(p, &flags);
4735 running = task_current(rq, p);
4738 dequeue_task(rq, p, 0);
4740 p->sched_class->put_prev_task(rq, p);
4742 p->numa_preferred_nid = nid;
4745 p->sched_class->set_curr_task(rq);
4747 enqueue_task(rq, p, 0);
4748 task_rq_unlock(rq, p, &flags);
4753 * migration_cpu_stop - this will be executed by a highprio stopper thread
4754 * and performs thread migration by bumping thread off CPU then
4755 * 'pushing' onto another runqueue.
4757 static int migration_cpu_stop(void *data)
4759 struct migration_arg *arg = data;
4762 * The original target cpu might have gone down and we might
4763 * be on another cpu but it doesn't matter.
4765 local_irq_disable();
4766 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4771 #ifdef CONFIG_HOTPLUG_CPU
4774 * Ensures that the idle task is using init_mm right before its cpu goes
4777 void idle_task_exit(void)
4779 struct mm_struct *mm = current->active_mm;
4781 BUG_ON(cpu_online(smp_processor_id()));
4783 if (mm != &init_mm) {
4784 switch_mm(mm, &init_mm, current);
4785 finish_arch_post_lock_switch();
4791 * Since this CPU is going 'away' for a while, fold any nr_active delta
4792 * we might have. Assumes we're called after migrate_tasks() so that the
4793 * nr_active count is stable.
4795 * Also see the comment "Global load-average calculations".
4797 static void calc_load_migrate(struct rq *rq)
4799 long delta = calc_load_fold_active(rq);
4801 atomic_long_add(delta, &calc_load_tasks);
4804 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4808 static const struct sched_class fake_sched_class = {
4809 .put_prev_task = put_prev_task_fake,
4812 static struct task_struct fake_task = {
4814 * Avoid pull_{rt,dl}_task()
4816 .prio = MAX_PRIO + 1,
4817 .sched_class = &fake_sched_class,
4821 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4822 * try_to_wake_up()->select_task_rq().
4824 * Called with rq->lock held even though we'er in stop_machine() and
4825 * there's no concurrency possible, we hold the required locks anyway
4826 * because of lock validation efforts.
4828 static void migrate_tasks(unsigned int dead_cpu)
4830 struct rq *rq = cpu_rq(dead_cpu);
4831 struct task_struct *next, *stop = rq->stop;
4835 * Fudge the rq selection such that the below task selection loop
4836 * doesn't get stuck on the currently eligible stop task.
4838 * We're currently inside stop_machine() and the rq is either stuck
4839 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4840 * either way we should never end up calling schedule() until we're
4846 * put_prev_task() and pick_next_task() sched
4847 * class method both need to have an up-to-date
4848 * value of rq->clock[_task]
4850 update_rq_clock(rq);
4854 * There's this thread running, bail when that's the only
4857 if (rq->nr_running == 1)
4860 next = pick_next_task(rq, &fake_task);
4862 next->sched_class->put_prev_task(rq, next);
4864 /* Find suitable destination for @next, with force if needed. */
4865 dest_cpu = select_fallback_rq(dead_cpu, next);
4866 raw_spin_unlock(&rq->lock);
4868 __migrate_task(next, dead_cpu, dest_cpu);
4870 raw_spin_lock(&rq->lock);
4876 #endif /* CONFIG_HOTPLUG_CPU */
4878 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4880 static struct ctl_table sd_ctl_dir[] = {
4882 .procname = "sched_domain",
4888 static struct ctl_table sd_ctl_root[] = {
4890 .procname = "kernel",
4892 .child = sd_ctl_dir,
4897 static struct ctl_table *sd_alloc_ctl_entry(int n)
4899 struct ctl_table *entry =
4900 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4905 static void sd_free_ctl_entry(struct ctl_table **tablep)
4907 struct ctl_table *entry;
4910 * In the intermediate directories, both the child directory and
4911 * procname are dynamically allocated and could fail but the mode
4912 * will always be set. In the lowest directory the names are
4913 * static strings and all have proc handlers.
4915 for (entry = *tablep; entry->mode; entry++) {
4917 sd_free_ctl_entry(&entry->child);
4918 if (entry->proc_handler == NULL)
4919 kfree(entry->procname);
4926 static int min_load_idx = 0;
4927 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4930 set_table_entry(struct ctl_table *entry,
4931 const char *procname, void *data, int maxlen,
4932 umode_t mode, proc_handler *proc_handler,
4935 entry->procname = procname;
4937 entry->maxlen = maxlen;
4939 entry->proc_handler = proc_handler;
4942 entry->extra1 = &min_load_idx;
4943 entry->extra2 = &max_load_idx;
4947 static struct ctl_table *
4948 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4950 struct ctl_table *table = sd_alloc_ctl_entry(14);
4955 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4956 sizeof(long), 0644, proc_doulongvec_minmax, false);
4957 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4958 sizeof(long), 0644, proc_doulongvec_minmax, false);
4959 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4960 sizeof(int), 0644, proc_dointvec_minmax, true);
4961 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4962 sizeof(int), 0644, proc_dointvec_minmax, true);
4963 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4964 sizeof(int), 0644, proc_dointvec_minmax, true);
4965 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4966 sizeof(int), 0644, proc_dointvec_minmax, true);
4967 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4968 sizeof(int), 0644, proc_dointvec_minmax, true);
4969 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4970 sizeof(int), 0644, proc_dointvec_minmax, false);
4971 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4972 sizeof(int), 0644, proc_dointvec_minmax, false);
4973 set_table_entry(&table[9], "cache_nice_tries",
4974 &sd->cache_nice_tries,
4975 sizeof(int), 0644, proc_dointvec_minmax, false);
4976 set_table_entry(&table[10], "flags", &sd->flags,
4977 sizeof(int), 0644, proc_dointvec_minmax, false);
4978 set_table_entry(&table[11], "max_newidle_lb_cost",
4979 &sd->max_newidle_lb_cost,
4980 sizeof(long), 0644, proc_doulongvec_minmax, false);
4981 set_table_entry(&table[12], "name", sd->name,
4982 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4983 /* &table[13] is terminator */
4988 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4990 struct ctl_table *entry, *table;
4991 struct sched_domain *sd;
4992 int domain_num = 0, i;
4995 for_each_domain(cpu, sd)
4997 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5002 for_each_domain(cpu, sd) {
5003 snprintf(buf, 32, "domain%d", i);
5004 entry->procname = kstrdup(buf, GFP_KERNEL);
5006 entry->child = sd_alloc_ctl_domain_table(sd);
5013 static struct ctl_table_header *sd_sysctl_header;
5014 static void register_sched_domain_sysctl(void)
5016 int i, cpu_num = num_possible_cpus();
5017 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5020 WARN_ON(sd_ctl_dir[0].child);
5021 sd_ctl_dir[0].child = entry;
5026 for_each_possible_cpu(i) {
5027 snprintf(buf, 32, "cpu%d", i);
5028 entry->procname = kstrdup(buf, GFP_KERNEL);
5030 entry->child = sd_alloc_ctl_cpu_table(i);
5034 WARN_ON(sd_sysctl_header);
5035 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5038 /* may be called multiple times per register */
5039 static void unregister_sched_domain_sysctl(void)
5041 if (sd_sysctl_header)
5042 unregister_sysctl_table(sd_sysctl_header);
5043 sd_sysctl_header = NULL;
5044 if (sd_ctl_dir[0].child)
5045 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5048 static void register_sched_domain_sysctl(void)
5051 static void unregister_sched_domain_sysctl(void)
5056 static void set_rq_online(struct rq *rq)
5059 const struct sched_class *class;
5061 cpumask_set_cpu(rq->cpu, rq->rd->online);
5064 for_each_class(class) {
5065 if (class->rq_online)
5066 class->rq_online(rq);
5071 static void set_rq_offline(struct rq *rq)
5074 const struct sched_class *class;
5076 for_each_class(class) {
5077 if (class->rq_offline)
5078 class->rq_offline(rq);
5081 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5087 * migration_call - callback that gets triggered when a CPU is added.
5088 * Here we can start up the necessary migration thread for the new CPU.
5091 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5093 int cpu = (long)hcpu;
5094 unsigned long flags;
5095 struct rq *rq = cpu_rq(cpu);
5097 switch (action & ~CPU_TASKS_FROZEN) {
5099 case CPU_UP_PREPARE:
5100 rq->calc_load_update = calc_load_update;
5104 /* Update our root-domain */
5105 raw_spin_lock_irqsave(&rq->lock, flags);
5107 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5111 raw_spin_unlock_irqrestore(&rq->lock, flags);
5114 #ifdef CONFIG_HOTPLUG_CPU
5116 sched_ttwu_pending();
5117 /* Update our root-domain */
5118 raw_spin_lock_irqsave(&rq->lock, flags);
5120 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5124 BUG_ON(rq->nr_running != 1); /* the migration thread */
5125 raw_spin_unlock_irqrestore(&rq->lock, flags);
5129 calc_load_migrate(rq);
5134 update_max_interval();
5140 * Register at high priority so that task migration (migrate_all_tasks)
5141 * happens before everything else. This has to be lower priority than
5142 * the notifier in the perf_event subsystem, though.
5144 static struct notifier_block migration_notifier = {
5145 .notifier_call = migration_call,
5146 .priority = CPU_PRI_MIGRATION,
5149 static void __cpuinit set_cpu_rq_start_time(void)
5151 int cpu = smp_processor_id();
5152 struct rq *rq = cpu_rq(cpu);
5153 rq->age_stamp = sched_clock_cpu(cpu);
5156 static int sched_cpu_active(struct notifier_block *nfb,
5157 unsigned long action, void *hcpu)
5159 switch (action & ~CPU_TASKS_FROZEN) {
5161 set_cpu_rq_start_time();
5163 case CPU_DOWN_FAILED:
5164 set_cpu_active((long)hcpu, true);
5171 static int sched_cpu_inactive(struct notifier_block *nfb,
5172 unsigned long action, void *hcpu)
5174 unsigned long flags;
5175 long cpu = (long)hcpu;
5177 switch (action & ~CPU_TASKS_FROZEN) {
5178 case CPU_DOWN_PREPARE:
5179 set_cpu_active(cpu, false);
5181 /* explicitly allow suspend */
5182 if (!(action & CPU_TASKS_FROZEN)) {
5183 struct dl_bw *dl_b = dl_bw_of(cpu);
5187 raw_spin_lock_irqsave(&dl_b->lock, flags);
5188 cpus = dl_bw_cpus(cpu);
5189 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5190 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5193 return notifier_from_errno(-EBUSY);
5201 static int __init migration_init(void)
5203 void *cpu = (void *)(long)smp_processor_id();
5206 /* Initialize migration for the boot CPU */
5207 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5208 BUG_ON(err == NOTIFY_BAD);
5209 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5210 register_cpu_notifier(&migration_notifier);
5212 /* Register cpu active notifiers */
5213 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5214 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5218 early_initcall(migration_init);
5223 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5225 #ifdef CONFIG_SCHED_DEBUG
5227 static __read_mostly int sched_debug_enabled;
5229 static int __init sched_debug_setup(char *str)
5231 sched_debug_enabled = 1;
5235 early_param("sched_debug", sched_debug_setup);
5237 static inline bool sched_debug(void)
5239 return sched_debug_enabled;
5242 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5243 struct cpumask *groupmask)
5245 struct sched_group *group = sd->groups;
5248 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5249 cpumask_clear(groupmask);
5251 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5253 if (!(sd->flags & SD_LOAD_BALANCE)) {
5254 printk("does not load-balance\n");
5256 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5261 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5263 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5264 printk(KERN_ERR "ERROR: domain->span does not contain "
5267 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5268 printk(KERN_ERR "ERROR: domain->groups does not contain"
5272 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5276 printk(KERN_ERR "ERROR: group is NULL\n");
5281 * Even though we initialize ->capacity to something semi-sane,
5282 * we leave capacity_orig unset. This allows us to detect if
5283 * domain iteration is still funny without causing /0 traps.
5285 if (!group->sgc->capacity_orig) {
5286 printk(KERN_CONT "\n");
5287 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n");
5291 if (!cpumask_weight(sched_group_cpus(group))) {
5292 printk(KERN_CONT "\n");
5293 printk(KERN_ERR "ERROR: empty group\n");
5297 if (!(sd->flags & SD_OVERLAP) &&
5298 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5299 printk(KERN_CONT "\n");
5300 printk(KERN_ERR "ERROR: repeated CPUs\n");
5304 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5306 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5308 printk(KERN_CONT " %s", str);
5309 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5310 printk(KERN_CONT " (cpu_capacity = %d)",
5311 group->sgc->capacity);
5314 group = group->next;
5315 } while (group != sd->groups);
5316 printk(KERN_CONT "\n");
5318 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5319 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5322 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5323 printk(KERN_ERR "ERROR: parent span is not a superset "
5324 "of domain->span\n");
5328 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5332 if (!sched_debug_enabled)
5336 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5340 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5343 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5351 #else /* !CONFIG_SCHED_DEBUG */
5352 # define sched_domain_debug(sd, cpu) do { } while (0)
5353 static inline bool sched_debug(void)
5357 #endif /* CONFIG_SCHED_DEBUG */
5359 static int sd_degenerate(struct sched_domain *sd)
5361 if (cpumask_weight(sched_domain_span(sd)) == 1)
5364 /* Following flags need at least 2 groups */
5365 if (sd->flags & (SD_LOAD_BALANCE |
5366 SD_BALANCE_NEWIDLE |
5369 SD_SHARE_CPUCAPACITY |
5370 SD_SHARE_PKG_RESOURCES |
5371 SD_SHARE_POWERDOMAIN)) {
5372 if (sd->groups != sd->groups->next)
5376 /* Following flags don't use groups */
5377 if (sd->flags & (SD_WAKE_AFFINE))
5384 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5386 unsigned long cflags = sd->flags, pflags = parent->flags;
5388 if (sd_degenerate(parent))
5391 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5394 /* Flags needing groups don't count if only 1 group in parent */
5395 if (parent->groups == parent->groups->next) {
5396 pflags &= ~(SD_LOAD_BALANCE |
5397 SD_BALANCE_NEWIDLE |
5400 SD_SHARE_CPUCAPACITY |
5401 SD_SHARE_PKG_RESOURCES |
5403 SD_SHARE_POWERDOMAIN);
5404 if (nr_node_ids == 1)
5405 pflags &= ~SD_SERIALIZE;
5407 if (~cflags & pflags)
5413 static void free_rootdomain(struct rcu_head *rcu)
5415 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5417 cpupri_cleanup(&rd->cpupri);
5418 cpudl_cleanup(&rd->cpudl);
5419 free_cpumask_var(rd->dlo_mask);
5420 free_cpumask_var(rd->rto_mask);
5421 free_cpumask_var(rd->online);
5422 free_cpumask_var(rd->span);
5426 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5428 struct root_domain *old_rd = NULL;
5429 unsigned long flags;
5431 raw_spin_lock_irqsave(&rq->lock, flags);
5436 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5439 cpumask_clear_cpu(rq->cpu, old_rd->span);
5442 * If we dont want to free the old_rd yet then
5443 * set old_rd to NULL to skip the freeing later
5446 if (!atomic_dec_and_test(&old_rd->refcount))
5450 atomic_inc(&rd->refcount);
5453 cpumask_set_cpu(rq->cpu, rd->span);
5454 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5457 raw_spin_unlock_irqrestore(&rq->lock, flags);
5460 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5463 static int init_rootdomain(struct root_domain *rd)
5465 memset(rd, 0, sizeof(*rd));
5467 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5469 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5471 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5473 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5476 init_dl_bw(&rd->dl_bw);
5477 if (cpudl_init(&rd->cpudl) != 0)
5480 if (cpupri_init(&rd->cpupri) != 0)
5485 free_cpumask_var(rd->rto_mask);
5487 free_cpumask_var(rd->dlo_mask);
5489 free_cpumask_var(rd->online);
5491 free_cpumask_var(rd->span);
5497 * By default the system creates a single root-domain with all cpus as
5498 * members (mimicking the global state we have today).
5500 struct root_domain def_root_domain;
5502 static void init_defrootdomain(void)
5504 init_rootdomain(&def_root_domain);
5506 atomic_set(&def_root_domain.refcount, 1);
5509 static struct root_domain *alloc_rootdomain(void)
5511 struct root_domain *rd;
5513 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5517 if (init_rootdomain(rd) != 0) {
5525 static void free_sched_groups(struct sched_group *sg, int free_sgc)
5527 struct sched_group *tmp, *first;
5536 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
5541 } while (sg != first);
5544 static void free_sched_domain(struct rcu_head *rcu)
5546 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5549 * If its an overlapping domain it has private groups, iterate and
5552 if (sd->flags & SD_OVERLAP) {
5553 free_sched_groups(sd->groups, 1);
5554 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5555 kfree(sd->groups->sgc);
5561 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5563 call_rcu(&sd->rcu, free_sched_domain);
5566 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5568 for (; sd; sd = sd->parent)
5569 destroy_sched_domain(sd, cpu);
5573 * Keep a special pointer to the highest sched_domain that has
5574 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5575 * allows us to avoid some pointer chasing select_idle_sibling().
5577 * Also keep a unique ID per domain (we use the first cpu number in
5578 * the cpumask of the domain), this allows us to quickly tell if
5579 * two cpus are in the same cache domain, see cpus_share_cache().
5581 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5582 DEFINE_PER_CPU(int, sd_llc_size);
5583 DEFINE_PER_CPU(int, sd_llc_id);
5584 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5585 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5586 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5588 static void update_top_cache_domain(int cpu)
5590 struct sched_domain *sd;
5591 struct sched_domain *busy_sd = NULL;
5595 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5597 id = cpumask_first(sched_domain_span(sd));
5598 size = cpumask_weight(sched_domain_span(sd));
5599 busy_sd = sd->parent; /* sd_busy */
5601 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5603 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5604 per_cpu(sd_llc_size, cpu) = size;
5605 per_cpu(sd_llc_id, cpu) = id;
5607 sd = lowest_flag_domain(cpu, SD_NUMA);
5608 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5610 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5611 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5615 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5616 * hold the hotplug lock.
5619 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5621 struct rq *rq = cpu_rq(cpu);
5622 struct sched_domain *tmp;
5624 /* Remove the sched domains which do not contribute to scheduling. */
5625 for (tmp = sd; tmp; ) {
5626 struct sched_domain *parent = tmp->parent;
5630 if (sd_parent_degenerate(tmp, parent)) {
5631 tmp->parent = parent->parent;
5633 parent->parent->child = tmp;
5635 * Transfer SD_PREFER_SIBLING down in case of a
5636 * degenerate parent; the spans match for this
5637 * so the property transfers.
5639 if (parent->flags & SD_PREFER_SIBLING)
5640 tmp->flags |= SD_PREFER_SIBLING;
5641 destroy_sched_domain(parent, cpu);
5646 if (sd && sd_degenerate(sd)) {
5649 destroy_sched_domain(tmp, cpu);
5654 sched_domain_debug(sd, cpu);
5656 rq_attach_root(rq, rd);
5658 rcu_assign_pointer(rq->sd, sd);
5659 destroy_sched_domains(tmp, cpu);
5661 update_top_cache_domain(cpu);
5664 /* cpus with isolated domains */
5665 static cpumask_var_t cpu_isolated_map;
5667 /* Setup the mask of cpus configured for isolated domains */
5668 static int __init isolated_cpu_setup(char *str)
5670 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5671 cpulist_parse(str, cpu_isolated_map);
5675 __setup("isolcpus=", isolated_cpu_setup);
5678 struct sched_domain ** __percpu sd;
5679 struct root_domain *rd;
5690 * Build an iteration mask that can exclude certain CPUs from the upwards
5693 * Asymmetric node setups can result in situations where the domain tree is of
5694 * unequal depth, make sure to skip domains that already cover the entire
5697 * In that case build_sched_domains() will have terminated the iteration early
5698 * and our sibling sd spans will be empty. Domains should always include the
5699 * cpu they're built on, so check that.
5702 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5704 const struct cpumask *span = sched_domain_span(sd);
5705 struct sd_data *sdd = sd->private;
5706 struct sched_domain *sibling;
5709 for_each_cpu(i, span) {
5710 sibling = *per_cpu_ptr(sdd->sd, i);
5711 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5714 cpumask_set_cpu(i, sched_group_mask(sg));
5719 * Return the canonical balance cpu for this group, this is the first cpu
5720 * of this group that's also in the iteration mask.
5722 int group_balance_cpu(struct sched_group *sg)
5724 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5728 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5730 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5731 const struct cpumask *span = sched_domain_span(sd);
5732 struct cpumask *covered = sched_domains_tmpmask;
5733 struct sd_data *sdd = sd->private;
5734 struct sched_domain *child;
5737 cpumask_clear(covered);
5739 for_each_cpu(i, span) {
5740 struct cpumask *sg_span;
5742 if (cpumask_test_cpu(i, covered))
5745 child = *per_cpu_ptr(sdd->sd, i);
5747 /* See the comment near build_group_mask(). */
5748 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5751 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5752 GFP_KERNEL, cpu_to_node(cpu));
5757 sg_span = sched_group_cpus(sg);
5759 child = child->child;
5760 cpumask_copy(sg_span, sched_domain_span(child));
5762 cpumask_set_cpu(i, sg_span);
5764 cpumask_or(covered, covered, sg_span);
5766 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
5767 if (atomic_inc_return(&sg->sgc->ref) == 1)
5768 build_group_mask(sd, sg);
5771 * Initialize sgc->capacity such that even if we mess up the
5772 * domains and no possible iteration will get us here, we won't
5775 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
5776 sg->sgc->capacity_orig = sg->sgc->capacity;
5779 * Make sure the first group of this domain contains the
5780 * canonical balance cpu. Otherwise the sched_domain iteration
5781 * breaks. See update_sg_lb_stats().
5783 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5784 group_balance_cpu(sg) == cpu)
5794 sd->groups = groups;
5799 free_sched_groups(first, 0);
5804 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5806 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5807 struct sched_domain *child = sd->child;
5810 cpu = cpumask_first(sched_domain_span(child));
5813 *sg = *per_cpu_ptr(sdd->sg, cpu);
5814 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
5815 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
5822 * build_sched_groups will build a circular linked list of the groups
5823 * covered by the given span, and will set each group's ->cpumask correctly,
5824 * and ->cpu_capacity to 0.
5826 * Assumes the sched_domain tree is fully constructed
5829 build_sched_groups(struct sched_domain *sd, int cpu)
5831 struct sched_group *first = NULL, *last = NULL;
5832 struct sd_data *sdd = sd->private;
5833 const struct cpumask *span = sched_domain_span(sd);
5834 struct cpumask *covered;
5837 get_group(cpu, sdd, &sd->groups);
5838 atomic_inc(&sd->groups->ref);
5840 if (cpu != cpumask_first(span))
5843 lockdep_assert_held(&sched_domains_mutex);
5844 covered = sched_domains_tmpmask;
5846 cpumask_clear(covered);
5848 for_each_cpu(i, span) {
5849 struct sched_group *sg;
5852 if (cpumask_test_cpu(i, covered))
5855 group = get_group(i, sdd, &sg);
5856 cpumask_setall(sched_group_mask(sg));
5858 for_each_cpu(j, span) {
5859 if (get_group(j, sdd, NULL) != group)
5862 cpumask_set_cpu(j, covered);
5863 cpumask_set_cpu(j, sched_group_cpus(sg));
5878 * Initialize sched groups cpu_capacity.
5880 * cpu_capacity indicates the capacity of sched group, which is used while
5881 * distributing the load between different sched groups in a sched domain.
5882 * Typically cpu_capacity for all the groups in a sched domain will be same
5883 * unless there are asymmetries in the topology. If there are asymmetries,
5884 * group having more cpu_capacity will pickup more load compared to the
5885 * group having less cpu_capacity.
5887 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
5889 struct sched_group *sg = sd->groups;
5894 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5896 } while (sg != sd->groups);
5898 if (cpu != group_balance_cpu(sg))
5901 update_group_capacity(sd, cpu);
5902 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
5906 * Initializers for schedule domains
5907 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5910 static int default_relax_domain_level = -1;
5911 int sched_domain_level_max;
5913 static int __init setup_relax_domain_level(char *str)
5915 if (kstrtoint(str, 0, &default_relax_domain_level))
5916 pr_warn("Unable to set relax_domain_level\n");
5920 __setup("relax_domain_level=", setup_relax_domain_level);
5922 static void set_domain_attribute(struct sched_domain *sd,
5923 struct sched_domain_attr *attr)
5927 if (!attr || attr->relax_domain_level < 0) {
5928 if (default_relax_domain_level < 0)
5931 request = default_relax_domain_level;
5933 request = attr->relax_domain_level;
5934 if (request < sd->level) {
5935 /* turn off idle balance on this domain */
5936 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5938 /* turn on idle balance on this domain */
5939 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5943 static void __sdt_free(const struct cpumask *cpu_map);
5944 static int __sdt_alloc(const struct cpumask *cpu_map);
5946 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5947 const struct cpumask *cpu_map)
5951 if (!atomic_read(&d->rd->refcount))
5952 free_rootdomain(&d->rd->rcu); /* fall through */
5954 free_percpu(d->sd); /* fall through */
5956 __sdt_free(cpu_map); /* fall through */
5962 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5963 const struct cpumask *cpu_map)
5965 memset(d, 0, sizeof(*d));
5967 if (__sdt_alloc(cpu_map))
5968 return sa_sd_storage;
5969 d->sd = alloc_percpu(struct sched_domain *);
5971 return sa_sd_storage;
5972 d->rd = alloc_rootdomain();
5975 return sa_rootdomain;
5979 * NULL the sd_data elements we've used to build the sched_domain and
5980 * sched_group structure so that the subsequent __free_domain_allocs()
5981 * will not free the data we're using.
5983 static void claim_allocations(int cpu, struct sched_domain *sd)
5985 struct sd_data *sdd = sd->private;
5987 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5988 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5990 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5991 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5993 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
5994 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
5998 static int sched_domains_numa_levels;
5999 static int *sched_domains_numa_distance;
6000 static struct cpumask ***sched_domains_numa_masks;
6001 static int sched_domains_curr_level;
6005 * SD_flags allowed in topology descriptions.
6007 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6008 * SD_SHARE_PKG_RESOURCES - describes shared caches
6009 * SD_NUMA - describes NUMA topologies
6010 * SD_SHARE_POWERDOMAIN - describes shared power domain
6013 * SD_ASYM_PACKING - describes SMT quirks
6015 #define TOPOLOGY_SD_FLAGS \
6016 (SD_SHARE_CPUCAPACITY | \
6017 SD_SHARE_PKG_RESOURCES | \
6020 SD_SHARE_POWERDOMAIN)
6022 static struct sched_domain *
6023 sd_init(struct sched_domain_topology_level *tl, int cpu)
6025 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6026 int sd_weight, sd_flags = 0;
6030 * Ugly hack to pass state to sd_numa_mask()...
6032 sched_domains_curr_level = tl->numa_level;
6035 sd_weight = cpumask_weight(tl->mask(cpu));
6038 sd_flags = (*tl->sd_flags)();
6039 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6040 "wrong sd_flags in topology description\n"))
6041 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6043 *sd = (struct sched_domain){
6044 .min_interval = sd_weight,
6045 .max_interval = 2*sd_weight,
6047 .imbalance_pct = 125,
6049 .cache_nice_tries = 0,
6056 .flags = 1*SD_LOAD_BALANCE
6057 | 1*SD_BALANCE_NEWIDLE
6062 | 0*SD_SHARE_CPUCAPACITY
6063 | 0*SD_SHARE_PKG_RESOURCES
6065 | 0*SD_PREFER_SIBLING
6070 .last_balance = jiffies,
6071 .balance_interval = sd_weight,
6073 .max_newidle_lb_cost = 0,
6074 .next_decay_max_lb_cost = jiffies,
6075 #ifdef CONFIG_SCHED_DEBUG
6081 * Convert topological properties into behaviour.
6084 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6085 sd->imbalance_pct = 110;
6086 sd->smt_gain = 1178; /* ~15% */
6088 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6089 sd->imbalance_pct = 117;
6090 sd->cache_nice_tries = 1;
6094 } else if (sd->flags & SD_NUMA) {
6095 sd->cache_nice_tries = 2;
6099 sd->flags |= SD_SERIALIZE;
6100 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6101 sd->flags &= ~(SD_BALANCE_EXEC |
6108 sd->flags |= SD_PREFER_SIBLING;
6109 sd->cache_nice_tries = 1;
6114 sd->private = &tl->data;
6120 * Topology list, bottom-up.
6122 static struct sched_domain_topology_level default_topology[] = {
6123 #ifdef CONFIG_SCHED_SMT
6124 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6126 #ifdef CONFIG_SCHED_MC
6127 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6129 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6133 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6135 #define for_each_sd_topology(tl) \
6136 for (tl = sched_domain_topology; tl->mask; tl++)
6138 void set_sched_topology(struct sched_domain_topology_level *tl)
6140 sched_domain_topology = tl;
6145 static const struct cpumask *sd_numa_mask(int cpu)
6147 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6150 static void sched_numa_warn(const char *str)
6152 static int done = false;
6160 printk(KERN_WARNING "ERROR: %s\n\n", str);
6162 for (i = 0; i < nr_node_ids; i++) {
6163 printk(KERN_WARNING " ");
6164 for (j = 0; j < nr_node_ids; j++)
6165 printk(KERN_CONT "%02d ", node_distance(i,j));
6166 printk(KERN_CONT "\n");
6168 printk(KERN_WARNING "\n");
6171 static bool find_numa_distance(int distance)
6175 if (distance == node_distance(0, 0))
6178 for (i = 0; i < sched_domains_numa_levels; i++) {
6179 if (sched_domains_numa_distance[i] == distance)
6186 static void sched_init_numa(void)
6188 int next_distance, curr_distance = node_distance(0, 0);
6189 struct sched_domain_topology_level *tl;
6193 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6194 if (!sched_domains_numa_distance)
6198 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6199 * unique distances in the node_distance() table.
6201 * Assumes node_distance(0,j) includes all distances in
6202 * node_distance(i,j) in order to avoid cubic time.
6204 next_distance = curr_distance;
6205 for (i = 0; i < nr_node_ids; i++) {
6206 for (j = 0; j < nr_node_ids; j++) {
6207 for (k = 0; k < nr_node_ids; k++) {
6208 int distance = node_distance(i, k);
6210 if (distance > curr_distance &&
6211 (distance < next_distance ||
6212 next_distance == curr_distance))
6213 next_distance = distance;
6216 * While not a strong assumption it would be nice to know
6217 * about cases where if node A is connected to B, B is not
6218 * equally connected to A.
6220 if (sched_debug() && node_distance(k, i) != distance)
6221 sched_numa_warn("Node-distance not symmetric");
6223 if (sched_debug() && i && !find_numa_distance(distance))
6224 sched_numa_warn("Node-0 not representative");
6226 if (next_distance != curr_distance) {
6227 sched_domains_numa_distance[level++] = next_distance;
6228 sched_domains_numa_levels = level;
6229 curr_distance = next_distance;
6234 * In case of sched_debug() we verify the above assumption.
6240 * 'level' contains the number of unique distances, excluding the
6241 * identity distance node_distance(i,i).
6243 * The sched_domains_numa_distance[] array includes the actual distance
6248 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6249 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6250 * the array will contain less then 'level' members. This could be
6251 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6252 * in other functions.
6254 * We reset it to 'level' at the end of this function.
6256 sched_domains_numa_levels = 0;
6258 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6259 if (!sched_domains_numa_masks)
6263 * Now for each level, construct a mask per node which contains all
6264 * cpus of nodes that are that many hops away from us.
6266 for (i = 0; i < level; i++) {
6267 sched_domains_numa_masks[i] =
6268 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6269 if (!sched_domains_numa_masks[i])
6272 for (j = 0; j < nr_node_ids; j++) {
6273 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6277 sched_domains_numa_masks[i][j] = mask;
6279 for (k = 0; k < nr_node_ids; k++) {
6280 if (node_distance(j, k) > sched_domains_numa_distance[i])
6283 cpumask_or(mask, mask, cpumask_of_node(k));
6288 /* Compute default topology size */
6289 for (i = 0; sched_domain_topology[i].mask; i++);
6291 tl = kzalloc((i + level + 1) *
6292 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6297 * Copy the default topology bits..
6299 for (i = 0; sched_domain_topology[i].mask; i++)
6300 tl[i] = sched_domain_topology[i];
6303 * .. and append 'j' levels of NUMA goodness.
6305 for (j = 0; j < level; i++, j++) {
6306 tl[i] = (struct sched_domain_topology_level){
6307 .mask = sd_numa_mask,
6308 .sd_flags = cpu_numa_flags,
6309 .flags = SDTL_OVERLAP,
6315 sched_domain_topology = tl;
6317 sched_domains_numa_levels = level;
6320 static void sched_domains_numa_masks_set(int cpu)
6323 int node = cpu_to_node(cpu);
6325 for (i = 0; i < sched_domains_numa_levels; i++) {
6326 for (j = 0; j < nr_node_ids; j++) {
6327 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6328 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6333 static void sched_domains_numa_masks_clear(int cpu)
6336 for (i = 0; i < sched_domains_numa_levels; i++) {
6337 for (j = 0; j < nr_node_ids; j++)
6338 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6343 * Update sched_domains_numa_masks[level][node] array when new cpus
6346 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6347 unsigned long action,
6350 int cpu = (long)hcpu;
6352 switch (action & ~CPU_TASKS_FROZEN) {
6354 sched_domains_numa_masks_set(cpu);
6358 sched_domains_numa_masks_clear(cpu);
6368 static inline void sched_init_numa(void)
6372 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6373 unsigned long action,
6378 #endif /* CONFIG_NUMA */
6380 static int __sdt_alloc(const struct cpumask *cpu_map)
6382 struct sched_domain_topology_level *tl;
6385 for_each_sd_topology(tl) {
6386 struct sd_data *sdd = &tl->data;
6388 sdd->sd = alloc_percpu(struct sched_domain *);
6392 sdd->sg = alloc_percpu(struct sched_group *);
6396 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
6400 for_each_cpu(j, cpu_map) {
6401 struct sched_domain *sd;
6402 struct sched_group *sg;
6403 struct sched_group_capacity *sgc;
6405 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6406 GFP_KERNEL, cpu_to_node(j));
6410 *per_cpu_ptr(sdd->sd, j) = sd;
6412 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6413 GFP_KERNEL, cpu_to_node(j));
6419 *per_cpu_ptr(sdd->sg, j) = sg;
6421 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
6422 GFP_KERNEL, cpu_to_node(j));
6426 *per_cpu_ptr(sdd->sgc, j) = sgc;
6433 static void __sdt_free(const struct cpumask *cpu_map)
6435 struct sched_domain_topology_level *tl;
6438 for_each_sd_topology(tl) {
6439 struct sd_data *sdd = &tl->data;
6441 for_each_cpu(j, cpu_map) {
6442 struct sched_domain *sd;
6445 sd = *per_cpu_ptr(sdd->sd, j);
6446 if (sd && (sd->flags & SD_OVERLAP))
6447 free_sched_groups(sd->groups, 0);
6448 kfree(*per_cpu_ptr(sdd->sd, j));
6452 kfree(*per_cpu_ptr(sdd->sg, j));
6454 kfree(*per_cpu_ptr(sdd->sgc, j));
6456 free_percpu(sdd->sd);
6458 free_percpu(sdd->sg);
6460 free_percpu(sdd->sgc);
6465 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6466 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6467 struct sched_domain *child, int cpu)
6469 struct sched_domain *sd = sd_init(tl, cpu);
6473 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6475 sd->level = child->level + 1;
6476 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6480 set_domain_attribute(sd, attr);
6486 * Build sched domains for a given set of cpus and attach the sched domains
6487 * to the individual cpus
6489 static int build_sched_domains(const struct cpumask *cpu_map,
6490 struct sched_domain_attr *attr)
6492 enum s_alloc alloc_state;
6493 struct sched_domain *sd;
6495 int i, ret = -ENOMEM;
6497 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6498 if (alloc_state != sa_rootdomain)
6501 /* Set up domains for cpus specified by the cpu_map. */
6502 for_each_cpu(i, cpu_map) {
6503 struct sched_domain_topology_level *tl;
6506 for_each_sd_topology(tl) {
6507 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6508 if (tl == sched_domain_topology)
6509 *per_cpu_ptr(d.sd, i) = sd;
6510 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6511 sd->flags |= SD_OVERLAP;
6512 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6517 /* Build the groups for the domains */
6518 for_each_cpu(i, cpu_map) {
6519 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6520 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6521 if (sd->flags & SD_OVERLAP) {
6522 if (build_overlap_sched_groups(sd, i))
6525 if (build_sched_groups(sd, i))
6531 /* Calculate CPU capacity for physical packages and nodes */
6532 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6533 if (!cpumask_test_cpu(i, cpu_map))
6536 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6537 claim_allocations(i, sd);
6538 init_sched_groups_capacity(i, sd);
6542 /* Attach the domains */
6544 for_each_cpu(i, cpu_map) {
6545 sd = *per_cpu_ptr(d.sd, i);
6546 cpu_attach_domain(sd, d.rd, i);
6552 __free_domain_allocs(&d, alloc_state, cpu_map);
6556 static cpumask_var_t *doms_cur; /* current sched domains */
6557 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6558 static struct sched_domain_attr *dattr_cur;
6559 /* attribues of custom domains in 'doms_cur' */
6562 * Special case: If a kmalloc of a doms_cur partition (array of
6563 * cpumask) fails, then fallback to a single sched domain,
6564 * as determined by the single cpumask fallback_doms.
6566 static cpumask_var_t fallback_doms;
6569 * arch_update_cpu_topology lets virtualized architectures update the
6570 * cpu core maps. It is supposed to return 1 if the topology changed
6571 * or 0 if it stayed the same.
6573 int __weak arch_update_cpu_topology(void)
6578 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6581 cpumask_var_t *doms;
6583 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6586 for (i = 0; i < ndoms; i++) {
6587 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6588 free_sched_domains(doms, i);
6595 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6598 for (i = 0; i < ndoms; i++)
6599 free_cpumask_var(doms[i]);
6604 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6605 * For now this just excludes isolated cpus, but could be used to
6606 * exclude other special cases in the future.
6608 static int init_sched_domains(const struct cpumask *cpu_map)
6612 arch_update_cpu_topology();
6614 doms_cur = alloc_sched_domains(ndoms_cur);
6616 doms_cur = &fallback_doms;
6617 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6618 err = build_sched_domains(doms_cur[0], NULL);
6619 register_sched_domain_sysctl();
6625 * Detach sched domains from a group of cpus specified in cpu_map
6626 * These cpus will now be attached to the NULL domain
6628 static void detach_destroy_domains(const struct cpumask *cpu_map)
6633 for_each_cpu(i, cpu_map)
6634 cpu_attach_domain(NULL, &def_root_domain, i);
6638 /* handle null as "default" */
6639 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6640 struct sched_domain_attr *new, int idx_new)
6642 struct sched_domain_attr tmp;
6649 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6650 new ? (new + idx_new) : &tmp,
6651 sizeof(struct sched_domain_attr));
6655 * Partition sched domains as specified by the 'ndoms_new'
6656 * cpumasks in the array doms_new[] of cpumasks. This compares
6657 * doms_new[] to the current sched domain partitioning, doms_cur[].
6658 * It destroys each deleted domain and builds each new domain.
6660 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6661 * The masks don't intersect (don't overlap.) We should setup one
6662 * sched domain for each mask. CPUs not in any of the cpumasks will
6663 * not be load balanced. If the same cpumask appears both in the
6664 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6667 * The passed in 'doms_new' should be allocated using
6668 * alloc_sched_domains. This routine takes ownership of it and will
6669 * free_sched_domains it when done with it. If the caller failed the
6670 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6671 * and partition_sched_domains() will fallback to the single partition
6672 * 'fallback_doms', it also forces the domains to be rebuilt.
6674 * If doms_new == NULL it will be replaced with cpu_online_mask.
6675 * ndoms_new == 0 is a special case for destroying existing domains,
6676 * and it will not create the default domain.
6678 * Call with hotplug lock held
6680 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6681 struct sched_domain_attr *dattr_new)
6686 mutex_lock(&sched_domains_mutex);
6688 /* always unregister in case we don't destroy any domains */
6689 unregister_sched_domain_sysctl();
6691 /* Let architecture update cpu core mappings. */
6692 new_topology = arch_update_cpu_topology();
6694 n = doms_new ? ndoms_new : 0;
6696 /* Destroy deleted domains */
6697 for (i = 0; i < ndoms_cur; i++) {
6698 for (j = 0; j < n && !new_topology; j++) {
6699 if (cpumask_equal(doms_cur[i], doms_new[j])
6700 && dattrs_equal(dattr_cur, i, dattr_new, j))
6703 /* no match - a current sched domain not in new doms_new[] */
6704 detach_destroy_domains(doms_cur[i]);
6710 if (doms_new == NULL) {
6712 doms_new = &fallback_doms;
6713 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6714 WARN_ON_ONCE(dattr_new);
6717 /* Build new domains */
6718 for (i = 0; i < ndoms_new; i++) {
6719 for (j = 0; j < n && !new_topology; j++) {
6720 if (cpumask_equal(doms_new[i], doms_cur[j])
6721 && dattrs_equal(dattr_new, i, dattr_cur, j))
6724 /* no match - add a new doms_new */
6725 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6730 /* Remember the new sched domains */
6731 if (doms_cur != &fallback_doms)
6732 free_sched_domains(doms_cur, ndoms_cur);
6733 kfree(dattr_cur); /* kfree(NULL) is safe */
6734 doms_cur = doms_new;
6735 dattr_cur = dattr_new;
6736 ndoms_cur = ndoms_new;
6738 register_sched_domain_sysctl();
6740 mutex_unlock(&sched_domains_mutex);
6743 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6746 * Update cpusets according to cpu_active mask. If cpusets are
6747 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6748 * around partition_sched_domains().
6750 * If we come here as part of a suspend/resume, don't touch cpusets because we
6751 * want to restore it back to its original state upon resume anyway.
6753 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6757 case CPU_ONLINE_FROZEN:
6758 case CPU_DOWN_FAILED_FROZEN:
6761 * num_cpus_frozen tracks how many CPUs are involved in suspend
6762 * resume sequence. As long as this is not the last online
6763 * operation in the resume sequence, just build a single sched
6764 * domain, ignoring cpusets.
6767 if (likely(num_cpus_frozen)) {
6768 partition_sched_domains(1, NULL, NULL);
6773 * This is the last CPU online operation. So fall through and
6774 * restore the original sched domains by considering the
6775 * cpuset configurations.
6779 case CPU_DOWN_FAILED:
6780 cpuset_update_active_cpus(true);
6788 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6792 case CPU_DOWN_PREPARE:
6793 cpuset_update_active_cpus(false);
6795 case CPU_DOWN_PREPARE_FROZEN:
6797 partition_sched_domains(1, NULL, NULL);
6805 void __init sched_init_smp(void)
6807 cpumask_var_t non_isolated_cpus;
6809 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6810 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6815 * There's no userspace yet to cause hotplug operations; hence all the
6816 * cpu masks are stable and all blatant races in the below code cannot
6819 mutex_lock(&sched_domains_mutex);
6820 init_sched_domains(cpu_active_mask);
6821 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6822 if (cpumask_empty(non_isolated_cpus))
6823 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6824 mutex_unlock(&sched_domains_mutex);
6826 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6827 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6828 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6832 /* Move init over to a non-isolated CPU */
6833 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6835 sched_init_granularity();
6836 free_cpumask_var(non_isolated_cpus);
6838 init_sched_rt_class();
6839 init_sched_dl_class();
6842 void __init sched_init_smp(void)
6844 sched_init_granularity();
6846 #endif /* CONFIG_SMP */
6848 const_debug unsigned int sysctl_timer_migration = 1;
6850 int in_sched_functions(unsigned long addr)
6852 return in_lock_functions(addr) ||
6853 (addr >= (unsigned long)__sched_text_start
6854 && addr < (unsigned long)__sched_text_end);
6857 #ifdef CONFIG_CGROUP_SCHED
6859 * Default task group.
6860 * Every task in system belongs to this group at bootup.
6862 struct task_group root_task_group;
6863 LIST_HEAD(task_groups);
6866 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6868 void __init sched_init(void)
6871 unsigned long alloc_size = 0, ptr;
6873 #ifdef CONFIG_FAIR_GROUP_SCHED
6874 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6876 #ifdef CONFIG_RT_GROUP_SCHED
6877 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6879 #ifdef CONFIG_CPUMASK_OFFSTACK
6880 alloc_size += num_possible_cpus() * cpumask_size();
6883 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6885 #ifdef CONFIG_FAIR_GROUP_SCHED
6886 root_task_group.se = (struct sched_entity **)ptr;
6887 ptr += nr_cpu_ids * sizeof(void **);
6889 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6890 ptr += nr_cpu_ids * sizeof(void **);
6892 #endif /* CONFIG_FAIR_GROUP_SCHED */
6893 #ifdef CONFIG_RT_GROUP_SCHED
6894 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6895 ptr += nr_cpu_ids * sizeof(void **);
6897 root_task_group.rt_rq = (struct rt_rq **)ptr;
6898 ptr += nr_cpu_ids * sizeof(void **);
6900 #endif /* CONFIG_RT_GROUP_SCHED */
6901 #ifdef CONFIG_CPUMASK_OFFSTACK
6902 for_each_possible_cpu(i) {
6903 per_cpu(load_balance_mask, i) = (void *)ptr;
6904 ptr += cpumask_size();
6906 #endif /* CONFIG_CPUMASK_OFFSTACK */
6909 init_rt_bandwidth(&def_rt_bandwidth,
6910 global_rt_period(), global_rt_runtime());
6911 init_dl_bandwidth(&def_dl_bandwidth,
6912 global_rt_period(), global_rt_runtime());
6915 init_defrootdomain();
6918 #ifdef CONFIG_RT_GROUP_SCHED
6919 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6920 global_rt_period(), global_rt_runtime());
6921 #endif /* CONFIG_RT_GROUP_SCHED */
6923 #ifdef CONFIG_CGROUP_SCHED
6924 list_add(&root_task_group.list, &task_groups);
6925 INIT_LIST_HEAD(&root_task_group.children);
6926 INIT_LIST_HEAD(&root_task_group.siblings);
6927 autogroup_init(&init_task);
6929 #endif /* CONFIG_CGROUP_SCHED */
6931 for_each_possible_cpu(i) {
6935 raw_spin_lock_init(&rq->lock);
6937 rq->calc_load_active = 0;
6938 rq->calc_load_update = jiffies + LOAD_FREQ;
6939 init_cfs_rq(&rq->cfs);
6940 init_rt_rq(&rq->rt, rq);
6941 init_dl_rq(&rq->dl, rq);
6942 #ifdef CONFIG_FAIR_GROUP_SCHED
6943 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6944 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6946 * How much cpu bandwidth does root_task_group get?
6948 * In case of task-groups formed thr' the cgroup filesystem, it
6949 * gets 100% of the cpu resources in the system. This overall
6950 * system cpu resource is divided among the tasks of
6951 * root_task_group and its child task-groups in a fair manner,
6952 * based on each entity's (task or task-group's) weight
6953 * (se->load.weight).
6955 * In other words, if root_task_group has 10 tasks of weight
6956 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6957 * then A0's share of the cpu resource is:
6959 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6961 * We achieve this by letting root_task_group's tasks sit
6962 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6964 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6965 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6966 #endif /* CONFIG_FAIR_GROUP_SCHED */
6968 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6969 #ifdef CONFIG_RT_GROUP_SCHED
6970 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6973 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6974 rq->cpu_load[j] = 0;
6976 rq->last_load_update_tick = jiffies;
6981 rq->cpu_capacity = SCHED_CAPACITY_SCALE;
6982 rq->post_schedule = 0;
6983 rq->active_balance = 0;
6984 rq->next_balance = jiffies;
6989 rq->avg_idle = 2*sysctl_sched_migration_cost;
6990 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6992 INIT_LIST_HEAD(&rq->cfs_tasks);
6994 rq_attach_root(rq, &def_root_domain);
6995 #ifdef CONFIG_NO_HZ_COMMON
6998 #ifdef CONFIG_NO_HZ_FULL
6999 rq->last_sched_tick = 0;
7003 atomic_set(&rq->nr_iowait, 0);
7006 set_load_weight(&init_task);
7008 #ifdef CONFIG_PREEMPT_NOTIFIERS
7009 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7013 * The boot idle thread does lazy MMU switching as well:
7015 atomic_inc(&init_mm.mm_count);
7016 enter_lazy_tlb(&init_mm, current);
7019 * Make us the idle thread. Technically, schedule() should not be
7020 * called from this thread, however somewhere below it might be,
7021 * but because we are the idle thread, we just pick up running again
7022 * when this runqueue becomes "idle".
7024 init_idle(current, smp_processor_id());
7026 calc_load_update = jiffies + LOAD_FREQ;
7029 * During early bootup we pretend to be a normal task:
7031 current->sched_class = &fair_sched_class;
7034 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7035 /* May be allocated at isolcpus cmdline parse time */
7036 if (cpu_isolated_map == NULL)
7037 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7038 idle_thread_set_boot_cpu();
7039 set_cpu_rq_start_time();
7041 init_sched_fair_class();
7043 scheduler_running = 1;
7046 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7047 static inline int preempt_count_equals(int preempt_offset)
7049 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7051 return (nested == preempt_offset);
7054 void __might_sleep(const char *file, int line, int preempt_offset)
7056 static unsigned long prev_jiffy; /* ratelimiting */
7058 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7059 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7060 !is_idle_task(current)) ||
7061 system_state != SYSTEM_RUNNING || oops_in_progress)
7063 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7065 prev_jiffy = jiffies;
7068 "BUG: sleeping function called from invalid context at %s:%d\n",
7071 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7072 in_atomic(), irqs_disabled(),
7073 current->pid, current->comm);
7075 debug_show_held_locks(current);
7076 if (irqs_disabled())
7077 print_irqtrace_events(current);
7078 #ifdef CONFIG_DEBUG_PREEMPT
7079 if (!preempt_count_equals(preempt_offset)) {
7080 pr_err("Preemption disabled at:");
7081 print_ip_sym(current->preempt_disable_ip);
7087 EXPORT_SYMBOL(__might_sleep);
7090 #ifdef CONFIG_MAGIC_SYSRQ
7091 static void normalize_task(struct rq *rq, struct task_struct *p)
7093 const struct sched_class *prev_class = p->sched_class;
7094 struct sched_attr attr = {
7095 .sched_policy = SCHED_NORMAL,
7097 int old_prio = p->prio;
7102 dequeue_task(rq, p, 0);
7103 __setscheduler(rq, p, &attr);
7105 enqueue_task(rq, p, 0);
7106 resched_task(rq->curr);
7109 check_class_changed(rq, p, prev_class, old_prio);
7112 void normalize_rt_tasks(void)
7114 struct task_struct *g, *p;
7115 unsigned long flags;
7118 read_lock_irqsave(&tasklist_lock, flags);
7119 do_each_thread(g, p) {
7121 * Only normalize user tasks:
7126 p->se.exec_start = 0;
7127 #ifdef CONFIG_SCHEDSTATS
7128 p->se.statistics.wait_start = 0;
7129 p->se.statistics.sleep_start = 0;
7130 p->se.statistics.block_start = 0;
7133 if (!dl_task(p) && !rt_task(p)) {
7135 * Renice negative nice level userspace
7138 if (task_nice(p) < 0 && p->mm)
7139 set_user_nice(p, 0);
7143 raw_spin_lock(&p->pi_lock);
7144 rq = __task_rq_lock(p);
7146 normalize_task(rq, p);
7148 __task_rq_unlock(rq);
7149 raw_spin_unlock(&p->pi_lock);
7150 } while_each_thread(g, p);
7152 read_unlock_irqrestore(&tasklist_lock, flags);
7155 #endif /* CONFIG_MAGIC_SYSRQ */
7157 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7159 * These functions are only useful for the IA64 MCA handling, or kdb.
7161 * They can only be called when the whole system has been
7162 * stopped - every CPU needs to be quiescent, and no scheduling
7163 * activity can take place. Using them for anything else would
7164 * be a serious bug, and as a result, they aren't even visible
7165 * under any other configuration.
7169 * curr_task - return the current task for a given cpu.
7170 * @cpu: the processor in question.
7172 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7174 * Return: The current task for @cpu.
7176 struct task_struct *curr_task(int cpu)
7178 return cpu_curr(cpu);
7181 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7185 * set_curr_task - set the current task for a given cpu.
7186 * @cpu: the processor in question.
7187 * @p: the task pointer to set.
7189 * Description: This function must only be used when non-maskable interrupts
7190 * are serviced on a separate stack. It allows the architecture to switch the
7191 * notion of the current task on a cpu in a non-blocking manner. This function
7192 * must be called with all CPU's synchronized, and interrupts disabled, the
7193 * and caller must save the original value of the current task (see
7194 * curr_task() above) and restore that value before reenabling interrupts and
7195 * re-starting the system.
7197 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7199 void set_curr_task(int cpu, struct task_struct *p)
7206 #ifdef CONFIG_CGROUP_SCHED
7207 /* task_group_lock serializes the addition/removal of task groups */
7208 static DEFINE_SPINLOCK(task_group_lock);
7210 static void free_sched_group(struct task_group *tg)
7212 free_fair_sched_group(tg);
7213 free_rt_sched_group(tg);
7218 /* allocate runqueue etc for a new task group */
7219 struct task_group *sched_create_group(struct task_group *parent)
7221 struct task_group *tg;
7223 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7225 return ERR_PTR(-ENOMEM);
7227 if (!alloc_fair_sched_group(tg, parent))
7230 if (!alloc_rt_sched_group(tg, parent))
7236 free_sched_group(tg);
7237 return ERR_PTR(-ENOMEM);
7240 void sched_online_group(struct task_group *tg, struct task_group *parent)
7242 unsigned long flags;
7244 spin_lock_irqsave(&task_group_lock, flags);
7245 list_add_rcu(&tg->list, &task_groups);
7247 WARN_ON(!parent); /* root should already exist */
7249 tg->parent = parent;
7250 INIT_LIST_HEAD(&tg->children);
7251 list_add_rcu(&tg->siblings, &parent->children);
7252 spin_unlock_irqrestore(&task_group_lock, flags);
7255 /* rcu callback to free various structures associated with a task group */
7256 static void free_sched_group_rcu(struct rcu_head *rhp)
7258 /* now it should be safe to free those cfs_rqs */
7259 free_sched_group(container_of(rhp, struct task_group, rcu));
7262 /* Destroy runqueue etc associated with a task group */
7263 void sched_destroy_group(struct task_group *tg)
7265 /* wait for possible concurrent references to cfs_rqs complete */
7266 call_rcu(&tg->rcu, free_sched_group_rcu);
7269 void sched_offline_group(struct task_group *tg)
7271 unsigned long flags;
7274 /* end participation in shares distribution */
7275 for_each_possible_cpu(i)
7276 unregister_fair_sched_group(tg, i);
7278 spin_lock_irqsave(&task_group_lock, flags);
7279 list_del_rcu(&tg->list);
7280 list_del_rcu(&tg->siblings);
7281 spin_unlock_irqrestore(&task_group_lock, flags);
7284 /* change task's runqueue when it moves between groups.
7285 * The caller of this function should have put the task in its new group
7286 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7287 * reflect its new group.
7289 void sched_move_task(struct task_struct *tsk)
7291 struct task_group *tg;
7293 unsigned long flags;
7296 rq = task_rq_lock(tsk, &flags);
7298 running = task_current(rq, tsk);
7302 dequeue_task(rq, tsk, 0);
7303 if (unlikely(running))
7304 tsk->sched_class->put_prev_task(rq, tsk);
7306 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7307 lockdep_is_held(&tsk->sighand->siglock)),
7308 struct task_group, css);
7309 tg = autogroup_task_group(tsk, tg);
7310 tsk->sched_task_group = tg;
7312 #ifdef CONFIG_FAIR_GROUP_SCHED
7313 if (tsk->sched_class->task_move_group)
7314 tsk->sched_class->task_move_group(tsk, on_rq);
7317 set_task_rq(tsk, task_cpu(tsk));
7319 if (unlikely(running))
7320 tsk->sched_class->set_curr_task(rq);
7322 enqueue_task(rq, tsk, 0);
7324 task_rq_unlock(rq, tsk, &flags);
7326 #endif /* CONFIG_CGROUP_SCHED */
7328 #ifdef CONFIG_RT_GROUP_SCHED
7330 * Ensure that the real time constraints are schedulable.
7332 static DEFINE_MUTEX(rt_constraints_mutex);
7334 /* Must be called with tasklist_lock held */
7335 static inline int tg_has_rt_tasks(struct task_group *tg)
7337 struct task_struct *g, *p;
7339 do_each_thread(g, p) {
7340 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7342 } while_each_thread(g, p);
7347 struct rt_schedulable_data {
7348 struct task_group *tg;
7353 static int tg_rt_schedulable(struct task_group *tg, void *data)
7355 struct rt_schedulable_data *d = data;
7356 struct task_group *child;
7357 unsigned long total, sum = 0;
7358 u64 period, runtime;
7360 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7361 runtime = tg->rt_bandwidth.rt_runtime;
7364 period = d->rt_period;
7365 runtime = d->rt_runtime;
7369 * Cannot have more runtime than the period.
7371 if (runtime > period && runtime != RUNTIME_INF)
7375 * Ensure we don't starve existing RT tasks.
7377 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7380 total = to_ratio(period, runtime);
7383 * Nobody can have more than the global setting allows.
7385 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7389 * The sum of our children's runtime should not exceed our own.
7391 list_for_each_entry_rcu(child, &tg->children, siblings) {
7392 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7393 runtime = child->rt_bandwidth.rt_runtime;
7395 if (child == d->tg) {
7396 period = d->rt_period;
7397 runtime = d->rt_runtime;
7400 sum += to_ratio(period, runtime);
7409 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7413 struct rt_schedulable_data data = {
7415 .rt_period = period,
7416 .rt_runtime = runtime,
7420 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7426 static int tg_set_rt_bandwidth(struct task_group *tg,
7427 u64 rt_period, u64 rt_runtime)
7431 mutex_lock(&rt_constraints_mutex);
7432 read_lock(&tasklist_lock);
7433 err = __rt_schedulable(tg, rt_period, rt_runtime);
7437 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7438 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7439 tg->rt_bandwidth.rt_runtime = rt_runtime;
7441 for_each_possible_cpu(i) {
7442 struct rt_rq *rt_rq = tg->rt_rq[i];
7444 raw_spin_lock(&rt_rq->rt_runtime_lock);
7445 rt_rq->rt_runtime = rt_runtime;
7446 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7448 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7450 read_unlock(&tasklist_lock);
7451 mutex_unlock(&rt_constraints_mutex);
7456 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7458 u64 rt_runtime, rt_period;
7460 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7461 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7462 if (rt_runtime_us < 0)
7463 rt_runtime = RUNTIME_INF;
7465 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7468 static long sched_group_rt_runtime(struct task_group *tg)
7472 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7475 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7476 do_div(rt_runtime_us, NSEC_PER_USEC);
7477 return rt_runtime_us;
7480 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7482 u64 rt_runtime, rt_period;
7484 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7485 rt_runtime = tg->rt_bandwidth.rt_runtime;
7490 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7493 static long sched_group_rt_period(struct task_group *tg)
7497 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7498 do_div(rt_period_us, NSEC_PER_USEC);
7499 return rt_period_us;
7501 #endif /* CONFIG_RT_GROUP_SCHED */
7503 #ifdef CONFIG_RT_GROUP_SCHED
7504 static int sched_rt_global_constraints(void)
7508 mutex_lock(&rt_constraints_mutex);
7509 read_lock(&tasklist_lock);
7510 ret = __rt_schedulable(NULL, 0, 0);
7511 read_unlock(&tasklist_lock);
7512 mutex_unlock(&rt_constraints_mutex);
7517 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7519 /* Don't accept realtime tasks when there is no way for them to run */
7520 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7526 #else /* !CONFIG_RT_GROUP_SCHED */
7527 static int sched_rt_global_constraints(void)
7529 unsigned long flags;
7532 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7533 for_each_possible_cpu(i) {
7534 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7536 raw_spin_lock(&rt_rq->rt_runtime_lock);
7537 rt_rq->rt_runtime = global_rt_runtime();
7538 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7540 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7544 #endif /* CONFIG_RT_GROUP_SCHED */
7546 static int sched_dl_global_constraints(void)
7548 u64 runtime = global_rt_runtime();
7549 u64 period = global_rt_period();
7550 u64 new_bw = to_ratio(period, runtime);
7552 unsigned long flags;
7555 * Here we want to check the bandwidth not being set to some
7556 * value smaller than the currently allocated bandwidth in
7557 * any of the root_domains.
7559 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7560 * cycling on root_domains... Discussion on different/better
7561 * solutions is welcome!
7563 for_each_possible_cpu(cpu) {
7564 struct dl_bw *dl_b = dl_bw_of(cpu);
7566 raw_spin_lock_irqsave(&dl_b->lock, flags);
7567 if (new_bw < dl_b->total_bw)
7569 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7578 static void sched_dl_do_global(void)
7582 unsigned long flags;
7584 def_dl_bandwidth.dl_period = global_rt_period();
7585 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7587 if (global_rt_runtime() != RUNTIME_INF)
7588 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7591 * FIXME: As above...
7593 for_each_possible_cpu(cpu) {
7594 struct dl_bw *dl_b = dl_bw_of(cpu);
7596 raw_spin_lock_irqsave(&dl_b->lock, flags);
7598 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7602 static int sched_rt_global_validate(void)
7604 if (sysctl_sched_rt_period <= 0)
7607 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7608 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7614 static void sched_rt_do_global(void)
7616 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7617 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7620 int sched_rt_handler(struct ctl_table *table, int write,
7621 void __user *buffer, size_t *lenp,
7624 int old_period, old_runtime;
7625 static DEFINE_MUTEX(mutex);
7629 old_period = sysctl_sched_rt_period;
7630 old_runtime = sysctl_sched_rt_runtime;
7632 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7634 if (!ret && write) {
7635 ret = sched_rt_global_validate();
7639 ret = sched_rt_global_constraints();
7643 ret = sched_dl_global_constraints();
7647 sched_rt_do_global();
7648 sched_dl_do_global();
7652 sysctl_sched_rt_period = old_period;
7653 sysctl_sched_rt_runtime = old_runtime;
7655 mutex_unlock(&mutex);
7660 int sched_rr_handler(struct ctl_table *table, int write,
7661 void __user *buffer, size_t *lenp,
7665 static DEFINE_MUTEX(mutex);
7668 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7669 /* make sure that internally we keep jiffies */
7670 /* also, writing zero resets timeslice to default */
7671 if (!ret && write) {
7672 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7673 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7675 mutex_unlock(&mutex);
7679 #ifdef CONFIG_CGROUP_SCHED
7681 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7683 return css ? container_of(css, struct task_group, css) : NULL;
7686 static struct cgroup_subsys_state *
7687 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7689 struct task_group *parent = css_tg(parent_css);
7690 struct task_group *tg;
7693 /* This is early initialization for the top cgroup */
7694 return &root_task_group.css;
7697 tg = sched_create_group(parent);
7699 return ERR_PTR(-ENOMEM);
7704 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7706 struct task_group *tg = css_tg(css);
7707 struct task_group *parent = css_tg(css->parent);
7710 sched_online_group(tg, parent);
7714 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7716 struct task_group *tg = css_tg(css);
7718 sched_destroy_group(tg);
7721 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7723 struct task_group *tg = css_tg(css);
7725 sched_offline_group(tg);
7728 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7729 struct cgroup_taskset *tset)
7731 struct task_struct *task;
7733 cgroup_taskset_for_each(task, tset) {
7734 #ifdef CONFIG_RT_GROUP_SCHED
7735 if (!sched_rt_can_attach(css_tg(css), task))
7738 /* We don't support RT-tasks being in separate groups */
7739 if (task->sched_class != &fair_sched_class)
7746 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7747 struct cgroup_taskset *tset)
7749 struct task_struct *task;
7751 cgroup_taskset_for_each(task, tset)
7752 sched_move_task(task);
7755 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7756 struct cgroup_subsys_state *old_css,
7757 struct task_struct *task)
7760 * cgroup_exit() is called in the copy_process() failure path.
7761 * Ignore this case since the task hasn't ran yet, this avoids
7762 * trying to poke a half freed task state from generic code.
7764 if (!(task->flags & PF_EXITING))
7767 sched_move_task(task);
7770 #ifdef CONFIG_FAIR_GROUP_SCHED
7771 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7772 struct cftype *cftype, u64 shareval)
7774 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7777 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7780 struct task_group *tg = css_tg(css);
7782 return (u64) scale_load_down(tg->shares);
7785 #ifdef CONFIG_CFS_BANDWIDTH
7786 static DEFINE_MUTEX(cfs_constraints_mutex);
7788 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7789 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7791 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7793 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7795 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7796 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7798 if (tg == &root_task_group)
7802 * Ensure we have at some amount of bandwidth every period. This is
7803 * to prevent reaching a state of large arrears when throttled via
7804 * entity_tick() resulting in prolonged exit starvation.
7806 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7810 * Likewise, bound things on the otherside by preventing insane quota
7811 * periods. This also allows us to normalize in computing quota
7814 if (period > max_cfs_quota_period)
7817 mutex_lock(&cfs_constraints_mutex);
7818 ret = __cfs_schedulable(tg, period, quota);
7822 runtime_enabled = quota != RUNTIME_INF;
7823 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7825 * If we need to toggle cfs_bandwidth_used, off->on must occur
7826 * before making related changes, and on->off must occur afterwards
7828 if (runtime_enabled && !runtime_was_enabled)
7829 cfs_bandwidth_usage_inc();
7830 raw_spin_lock_irq(&cfs_b->lock);
7831 cfs_b->period = ns_to_ktime(period);
7832 cfs_b->quota = quota;
7834 __refill_cfs_bandwidth_runtime(cfs_b);
7835 /* restart the period timer (if active) to handle new period expiry */
7836 if (runtime_enabled && cfs_b->timer_active) {
7837 /* force a reprogram */
7838 __start_cfs_bandwidth(cfs_b, true);
7840 raw_spin_unlock_irq(&cfs_b->lock);
7842 for_each_possible_cpu(i) {
7843 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7844 struct rq *rq = cfs_rq->rq;
7846 raw_spin_lock_irq(&rq->lock);
7847 cfs_rq->runtime_enabled = runtime_enabled;
7848 cfs_rq->runtime_remaining = 0;
7850 if (cfs_rq->throttled)
7851 unthrottle_cfs_rq(cfs_rq);
7852 raw_spin_unlock_irq(&rq->lock);
7854 if (runtime_was_enabled && !runtime_enabled)
7855 cfs_bandwidth_usage_dec();
7857 mutex_unlock(&cfs_constraints_mutex);
7862 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7866 period = ktime_to_ns(tg->cfs_bandwidth.period);
7867 if (cfs_quota_us < 0)
7868 quota = RUNTIME_INF;
7870 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7872 return tg_set_cfs_bandwidth(tg, period, quota);
7875 long tg_get_cfs_quota(struct task_group *tg)
7879 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7882 quota_us = tg->cfs_bandwidth.quota;
7883 do_div(quota_us, NSEC_PER_USEC);
7888 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7892 period = (u64)cfs_period_us * NSEC_PER_USEC;
7893 quota = tg->cfs_bandwidth.quota;
7895 return tg_set_cfs_bandwidth(tg, period, quota);
7898 long tg_get_cfs_period(struct task_group *tg)
7902 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7903 do_div(cfs_period_us, NSEC_PER_USEC);
7905 return cfs_period_us;
7908 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7911 return tg_get_cfs_quota(css_tg(css));
7914 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7915 struct cftype *cftype, s64 cfs_quota_us)
7917 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7920 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7923 return tg_get_cfs_period(css_tg(css));
7926 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7927 struct cftype *cftype, u64 cfs_period_us)
7929 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7932 struct cfs_schedulable_data {
7933 struct task_group *tg;
7938 * normalize group quota/period to be quota/max_period
7939 * note: units are usecs
7941 static u64 normalize_cfs_quota(struct task_group *tg,
7942 struct cfs_schedulable_data *d)
7950 period = tg_get_cfs_period(tg);
7951 quota = tg_get_cfs_quota(tg);
7954 /* note: these should typically be equivalent */
7955 if (quota == RUNTIME_INF || quota == -1)
7958 return to_ratio(period, quota);
7961 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7963 struct cfs_schedulable_data *d = data;
7964 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7965 s64 quota = 0, parent_quota = -1;
7968 quota = RUNTIME_INF;
7970 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7972 quota = normalize_cfs_quota(tg, d);
7973 parent_quota = parent_b->hierarchal_quota;
7976 * ensure max(child_quota) <= parent_quota, inherit when no
7979 if (quota == RUNTIME_INF)
7980 quota = parent_quota;
7981 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7984 cfs_b->hierarchal_quota = quota;
7989 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7992 struct cfs_schedulable_data data = {
7998 if (quota != RUNTIME_INF) {
7999 do_div(data.period, NSEC_PER_USEC);
8000 do_div(data.quota, NSEC_PER_USEC);
8004 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8010 static int cpu_stats_show(struct seq_file *sf, void *v)
8012 struct task_group *tg = css_tg(seq_css(sf));
8013 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8015 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8016 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8017 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8021 #endif /* CONFIG_CFS_BANDWIDTH */
8022 #endif /* CONFIG_FAIR_GROUP_SCHED */
8024 #ifdef CONFIG_RT_GROUP_SCHED
8025 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8026 struct cftype *cft, s64 val)
8028 return sched_group_set_rt_runtime(css_tg(css), val);
8031 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8034 return sched_group_rt_runtime(css_tg(css));
8037 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8038 struct cftype *cftype, u64 rt_period_us)
8040 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8043 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8046 return sched_group_rt_period(css_tg(css));
8048 #endif /* CONFIG_RT_GROUP_SCHED */
8050 static struct cftype cpu_files[] = {
8051 #ifdef CONFIG_FAIR_GROUP_SCHED
8054 .read_u64 = cpu_shares_read_u64,
8055 .write_u64 = cpu_shares_write_u64,
8058 #ifdef CONFIG_CFS_BANDWIDTH
8060 .name = "cfs_quota_us",
8061 .read_s64 = cpu_cfs_quota_read_s64,
8062 .write_s64 = cpu_cfs_quota_write_s64,
8065 .name = "cfs_period_us",
8066 .read_u64 = cpu_cfs_period_read_u64,
8067 .write_u64 = cpu_cfs_period_write_u64,
8071 .seq_show = cpu_stats_show,
8074 #ifdef CONFIG_RT_GROUP_SCHED
8076 .name = "rt_runtime_us",
8077 .read_s64 = cpu_rt_runtime_read,
8078 .write_s64 = cpu_rt_runtime_write,
8081 .name = "rt_period_us",
8082 .read_u64 = cpu_rt_period_read_uint,
8083 .write_u64 = cpu_rt_period_write_uint,
8089 struct cgroup_subsys cpu_cgrp_subsys = {
8090 .css_alloc = cpu_cgroup_css_alloc,
8091 .css_free = cpu_cgroup_css_free,
8092 .css_online = cpu_cgroup_css_online,
8093 .css_offline = cpu_cgroup_css_offline,
8094 .can_attach = cpu_cgroup_can_attach,
8095 .attach = cpu_cgroup_attach,
8096 .exit = cpu_cgroup_exit,
8097 .base_cftypes = cpu_files,
8101 #endif /* CONFIG_CGROUP_SCHED */
8103 void dump_cpu_task(int cpu)
8105 pr_info("Task dump for CPU %d:\n", cpu);
8106 sched_show_task(cpu_curr(cpu));