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>
77 #include <asm/switch_to.h>
79 #include <asm/irq_regs.h>
80 #include <asm/mutex.h>
81 #ifdef CONFIG_PARAVIRT
82 #include <asm/paravirt.h>
86 #include "../workqueue_internal.h"
87 #include "../smpboot.h"
89 #define CREATE_TRACE_POINTS
90 #include <trace/events/sched.h>
92 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
95 ktime_t soft, hard, now;
98 if (hrtimer_active(period_timer))
101 now = hrtimer_cb_get_time(period_timer);
102 hrtimer_forward(period_timer, now, period);
104 soft = hrtimer_get_softexpires(period_timer);
105 hard = hrtimer_get_expires(period_timer);
106 delta = ktime_to_ns(ktime_sub(hard, soft));
107 __hrtimer_start_range_ns(period_timer, soft, delta,
108 HRTIMER_MODE_ABS_PINNED, 0);
112 DEFINE_MUTEX(sched_domains_mutex);
113 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
115 static void update_rq_clock_task(struct rq *rq, s64 delta);
117 void update_rq_clock(struct rq *rq)
121 if (rq->skip_clock_update > 0)
124 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
126 update_rq_clock_task(rq, delta);
130 * Debugging: various feature bits
133 #define SCHED_FEAT(name, enabled) \
134 (1UL << __SCHED_FEAT_##name) * enabled |
136 const_debug unsigned int sysctl_sched_features =
137 #include "features.h"
142 #ifdef CONFIG_SCHED_DEBUG
143 #define SCHED_FEAT(name, enabled) \
146 static const char * const sched_feat_names[] = {
147 #include "features.h"
152 static int sched_feat_show(struct seq_file *m, void *v)
156 for (i = 0; i < __SCHED_FEAT_NR; i++) {
157 if (!(sysctl_sched_features & (1UL << i)))
159 seq_printf(m, "%s ", sched_feat_names[i]);
166 #ifdef HAVE_JUMP_LABEL
168 #define jump_label_key__true STATIC_KEY_INIT_TRUE
169 #define jump_label_key__false STATIC_KEY_INIT_FALSE
171 #define SCHED_FEAT(name, enabled) \
172 jump_label_key__##enabled ,
174 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
175 #include "features.h"
180 static void sched_feat_disable(int i)
182 if (static_key_enabled(&sched_feat_keys[i]))
183 static_key_slow_dec(&sched_feat_keys[i]);
186 static void sched_feat_enable(int i)
188 if (!static_key_enabled(&sched_feat_keys[i]))
189 static_key_slow_inc(&sched_feat_keys[i]);
192 static void sched_feat_disable(int i) { };
193 static void sched_feat_enable(int i) { };
194 #endif /* HAVE_JUMP_LABEL */
196 static int sched_feat_set(char *cmp)
201 if (strncmp(cmp, "NO_", 3) == 0) {
206 for (i = 0; i < __SCHED_FEAT_NR; i++) {
207 if (strcmp(cmp, sched_feat_names[i]) == 0) {
209 sysctl_sched_features &= ~(1UL << i);
210 sched_feat_disable(i);
212 sysctl_sched_features |= (1UL << i);
213 sched_feat_enable(i);
223 sched_feat_write(struct file *filp, const char __user *ubuf,
224 size_t cnt, loff_t *ppos)
233 if (copy_from_user(&buf, ubuf, cnt))
239 i = sched_feat_set(cmp);
240 if (i == __SCHED_FEAT_NR)
248 static int sched_feat_open(struct inode *inode, struct file *filp)
250 return single_open(filp, sched_feat_show, NULL);
253 static const struct file_operations sched_feat_fops = {
254 .open = sched_feat_open,
255 .write = sched_feat_write,
258 .release = single_release,
261 static __init int sched_init_debug(void)
263 debugfs_create_file("sched_features", 0644, NULL, NULL,
268 late_initcall(sched_init_debug);
269 #endif /* CONFIG_SCHED_DEBUG */
272 * Number of tasks to iterate in a single balance run.
273 * Limited because this is done with IRQs disabled.
275 const_debug unsigned int sysctl_sched_nr_migrate = 32;
278 * period over which we average the RT time consumption, measured
283 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
286 * period over which we measure -rt task cpu usage in us.
289 unsigned int sysctl_sched_rt_period = 1000000;
291 __read_mostly int scheduler_running;
294 * part of the period that we allow rt tasks to run in us.
297 int sysctl_sched_rt_runtime = 950000;
300 * __task_rq_lock - lock the rq @p resides on.
302 static inline struct rq *__task_rq_lock(struct task_struct *p)
307 lockdep_assert_held(&p->pi_lock);
311 raw_spin_lock(&rq->lock);
312 if (likely(rq == task_rq(p)))
314 raw_spin_unlock(&rq->lock);
319 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
321 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
322 __acquires(p->pi_lock)
328 raw_spin_lock_irqsave(&p->pi_lock, *flags);
330 raw_spin_lock(&rq->lock);
331 if (likely(rq == task_rq(p)))
333 raw_spin_unlock(&rq->lock);
334 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
338 static void __task_rq_unlock(struct rq *rq)
341 raw_spin_unlock(&rq->lock);
345 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
347 __releases(p->pi_lock)
349 raw_spin_unlock(&rq->lock);
350 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
354 * this_rq_lock - lock this runqueue and disable interrupts.
356 static struct rq *this_rq_lock(void)
363 raw_spin_lock(&rq->lock);
368 #ifdef CONFIG_SCHED_HRTICK
370 * Use HR-timers to deliver accurate preemption points.
373 static void hrtick_clear(struct rq *rq)
375 if (hrtimer_active(&rq->hrtick_timer))
376 hrtimer_cancel(&rq->hrtick_timer);
380 * High-resolution timer tick.
381 * Runs from hardirq context with interrupts disabled.
383 static enum hrtimer_restart hrtick(struct hrtimer *timer)
385 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
389 raw_spin_lock(&rq->lock);
391 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
392 raw_spin_unlock(&rq->lock);
394 return HRTIMER_NORESTART;
399 static int __hrtick_restart(struct rq *rq)
401 struct hrtimer *timer = &rq->hrtick_timer;
402 ktime_t time = hrtimer_get_softexpires(timer);
404 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
408 * called from hardirq (IPI) context
410 static void __hrtick_start(void *arg)
414 raw_spin_lock(&rq->lock);
415 __hrtick_restart(rq);
416 rq->hrtick_csd_pending = 0;
417 raw_spin_unlock(&rq->lock);
421 * Called to set the hrtick timer state.
423 * called with rq->lock held and irqs disabled
425 void hrtick_start(struct rq *rq, u64 delay)
427 struct hrtimer *timer = &rq->hrtick_timer;
428 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
430 hrtimer_set_expires(timer, time);
432 if (rq == this_rq()) {
433 __hrtick_restart(rq);
434 } else if (!rq->hrtick_csd_pending) {
435 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
436 rq->hrtick_csd_pending = 1;
441 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
443 int cpu = (int)(long)hcpu;
446 case CPU_UP_CANCELED:
447 case CPU_UP_CANCELED_FROZEN:
448 case CPU_DOWN_PREPARE:
449 case CPU_DOWN_PREPARE_FROZEN:
451 case CPU_DEAD_FROZEN:
452 hrtick_clear(cpu_rq(cpu));
459 static __init void init_hrtick(void)
461 hotcpu_notifier(hotplug_hrtick, 0);
465 * Called to set the hrtick timer state.
467 * called with rq->lock held and irqs disabled
469 void hrtick_start(struct rq *rq, u64 delay)
471 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
472 HRTIMER_MODE_REL_PINNED, 0);
475 static inline void init_hrtick(void)
478 #endif /* CONFIG_SMP */
480 static void init_rq_hrtick(struct rq *rq)
483 rq->hrtick_csd_pending = 0;
485 rq->hrtick_csd.flags = 0;
486 rq->hrtick_csd.func = __hrtick_start;
487 rq->hrtick_csd.info = rq;
490 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
491 rq->hrtick_timer.function = hrtick;
493 #else /* CONFIG_SCHED_HRTICK */
494 static inline void hrtick_clear(struct rq *rq)
498 static inline void init_rq_hrtick(struct rq *rq)
502 static inline void init_hrtick(void)
505 #endif /* CONFIG_SCHED_HRTICK */
508 * resched_task - mark a task 'to be rescheduled now'.
510 * On UP this means the setting of the need_resched flag, on SMP it
511 * might also involve a cross-CPU call to trigger the scheduler on
514 void resched_task(struct task_struct *p)
518 lockdep_assert_held(&task_rq(p)->lock);
520 if (test_tsk_need_resched(p))
523 set_tsk_need_resched(p);
526 if (cpu == smp_processor_id()) {
527 set_preempt_need_resched();
531 /* NEED_RESCHED must be visible before we test polling */
533 if (!tsk_is_polling(p))
534 smp_send_reschedule(cpu);
537 void resched_cpu(int cpu)
539 struct rq *rq = cpu_rq(cpu);
542 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
544 resched_task(cpu_curr(cpu));
545 raw_spin_unlock_irqrestore(&rq->lock, flags);
549 #ifdef CONFIG_NO_HZ_COMMON
551 * In the semi idle case, use the nearest busy cpu for migrating timers
552 * from an idle cpu. This is good for power-savings.
554 * We don't do similar optimization for completely idle system, as
555 * selecting an idle cpu will add more delays to the timers than intended
556 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
558 int get_nohz_timer_target(void)
560 int cpu = smp_processor_id();
562 struct sched_domain *sd;
565 for_each_domain(cpu, sd) {
566 for_each_cpu(i, sched_domain_span(sd)) {
578 * When add_timer_on() enqueues a timer into the timer wheel of an
579 * idle CPU then this timer might expire before the next timer event
580 * which is scheduled to wake up that CPU. In case of a completely
581 * idle system the next event might even be infinite time into the
582 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
583 * leaves the inner idle loop so the newly added timer is taken into
584 * account when the CPU goes back to idle and evaluates the timer
585 * wheel for the next timer event.
587 static void wake_up_idle_cpu(int cpu)
589 struct rq *rq = cpu_rq(cpu);
591 if (cpu == smp_processor_id())
595 * This is safe, as this function is called with the timer
596 * wheel base lock of (cpu) held. When the CPU is on the way
597 * to idle and has not yet set rq->curr to idle then it will
598 * be serialized on the timer wheel base lock and take the new
599 * timer into account automatically.
601 if (rq->curr != rq->idle)
605 * We can set TIF_RESCHED on the idle task of the other CPU
606 * lockless. The worst case is that the other CPU runs the
607 * idle task through an additional NOOP schedule()
609 set_tsk_need_resched(rq->idle);
611 /* NEED_RESCHED must be visible before we test polling */
613 if (!tsk_is_polling(rq->idle))
614 smp_send_reschedule(cpu);
617 static bool wake_up_full_nohz_cpu(int cpu)
619 if (tick_nohz_full_cpu(cpu)) {
620 if (cpu != smp_processor_id() ||
621 tick_nohz_tick_stopped())
622 smp_send_reschedule(cpu);
629 void wake_up_nohz_cpu(int cpu)
631 if (!wake_up_full_nohz_cpu(cpu))
632 wake_up_idle_cpu(cpu);
635 static inline bool got_nohz_idle_kick(void)
637 int cpu = smp_processor_id();
639 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
642 if (idle_cpu(cpu) && !need_resched())
646 * We can't run Idle Load Balance on this CPU for this time so we
647 * cancel it and clear NOHZ_BALANCE_KICK
649 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
653 #else /* CONFIG_NO_HZ_COMMON */
655 static inline bool got_nohz_idle_kick(void)
660 #endif /* CONFIG_NO_HZ_COMMON */
662 #ifdef CONFIG_NO_HZ_FULL
663 bool sched_can_stop_tick(void)
669 /* Make sure rq->nr_running update is visible after the IPI */
672 /* More than one running task need preemption */
673 if (rq->nr_running > 1)
678 #endif /* CONFIG_NO_HZ_FULL */
680 void sched_avg_update(struct rq *rq)
682 s64 period = sched_avg_period();
684 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
686 * Inline assembly required to prevent the compiler
687 * optimising this loop into a divmod call.
688 * See __iter_div_u64_rem() for another example of this.
690 asm("" : "+rm" (rq->age_stamp));
691 rq->age_stamp += period;
696 #endif /* CONFIG_SMP */
698 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
699 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
701 * Iterate task_group tree rooted at *from, calling @down when first entering a
702 * node and @up when leaving it for the final time.
704 * Caller must hold rcu_lock or sufficient equivalent.
706 int walk_tg_tree_from(struct task_group *from,
707 tg_visitor down, tg_visitor up, void *data)
709 struct task_group *parent, *child;
715 ret = (*down)(parent, data);
718 list_for_each_entry_rcu(child, &parent->children, siblings) {
725 ret = (*up)(parent, data);
726 if (ret || parent == from)
730 parent = parent->parent;
737 int tg_nop(struct task_group *tg, void *data)
743 static void set_load_weight(struct task_struct *p)
745 int prio = p->static_prio - MAX_RT_PRIO;
746 struct load_weight *load = &p->se.load;
749 * SCHED_IDLE tasks get minimal weight:
751 if (p->policy == SCHED_IDLE) {
752 load->weight = scale_load(WEIGHT_IDLEPRIO);
753 load->inv_weight = WMULT_IDLEPRIO;
757 load->weight = scale_load(prio_to_weight[prio]);
758 load->inv_weight = prio_to_wmult[prio];
761 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
764 sched_info_queued(rq, p);
765 p->sched_class->enqueue_task(rq, p, flags);
768 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
771 sched_info_dequeued(rq, p);
772 p->sched_class->dequeue_task(rq, p, flags);
775 void activate_task(struct rq *rq, struct task_struct *p, int flags)
777 if (task_contributes_to_load(p))
778 rq->nr_uninterruptible--;
780 enqueue_task(rq, p, flags);
783 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
785 if (task_contributes_to_load(p))
786 rq->nr_uninterruptible++;
788 dequeue_task(rq, p, flags);
791 static void update_rq_clock_task(struct rq *rq, s64 delta)
794 * In theory, the compile should just see 0 here, and optimize out the call
795 * to sched_rt_avg_update. But I don't trust it...
797 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
798 s64 steal = 0, irq_delta = 0;
800 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
801 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
804 * Since irq_time is only updated on {soft,}irq_exit, we might run into
805 * this case when a previous update_rq_clock() happened inside a
808 * When this happens, we stop ->clock_task and only update the
809 * prev_irq_time stamp to account for the part that fit, so that a next
810 * update will consume the rest. This ensures ->clock_task is
813 * It does however cause some slight miss-attribution of {soft,}irq
814 * time, a more accurate solution would be to update the irq_time using
815 * the current rq->clock timestamp, except that would require using
818 if (irq_delta > delta)
821 rq->prev_irq_time += irq_delta;
824 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
825 if (static_key_false((¶virt_steal_rq_enabled))) {
828 steal = paravirt_steal_clock(cpu_of(rq));
829 steal -= rq->prev_steal_time_rq;
831 if (unlikely(steal > delta))
834 st = steal_ticks(steal);
835 steal = st * TICK_NSEC;
837 rq->prev_steal_time_rq += steal;
843 rq->clock_task += delta;
845 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
846 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
847 sched_rt_avg_update(rq, irq_delta + steal);
851 void sched_set_stop_task(int cpu, struct task_struct *stop)
853 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
854 struct task_struct *old_stop = cpu_rq(cpu)->stop;
858 * Make it appear like a SCHED_FIFO task, its something
859 * userspace knows about and won't get confused about.
861 * Also, it will make PI more or less work without too
862 * much confusion -- but then, stop work should not
863 * rely on PI working anyway.
865 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
867 stop->sched_class = &stop_sched_class;
870 cpu_rq(cpu)->stop = stop;
874 * Reset it back to a normal scheduling class so that
875 * it can die in pieces.
877 old_stop->sched_class = &rt_sched_class;
882 * __normal_prio - return the priority that is based on the static prio
884 static inline int __normal_prio(struct task_struct *p)
886 return p->static_prio;
890 * Calculate the expected normal priority: i.e. priority
891 * without taking RT-inheritance into account. Might be
892 * boosted by interactivity modifiers. Changes upon fork,
893 * setprio syscalls, and whenever the interactivity
894 * estimator recalculates.
896 static inline int normal_prio(struct task_struct *p)
900 if (task_has_dl_policy(p))
901 prio = MAX_DL_PRIO-1;
902 else if (task_has_rt_policy(p))
903 prio = MAX_RT_PRIO-1 - p->rt_priority;
905 prio = __normal_prio(p);
910 * Calculate the current priority, i.e. the priority
911 * taken into account by the scheduler. This value might
912 * be boosted by RT tasks, or might be boosted by
913 * interactivity modifiers. Will be RT if the task got
914 * RT-boosted. If not then it returns p->normal_prio.
916 static int effective_prio(struct task_struct *p)
918 p->normal_prio = normal_prio(p);
920 * If we are RT tasks or we were boosted to RT priority,
921 * keep the priority unchanged. Otherwise, update priority
922 * to the normal priority:
924 if (!rt_prio(p->prio))
925 return p->normal_prio;
930 * task_curr - is this task currently executing on a CPU?
931 * @p: the task in question.
933 * Return: 1 if the task is currently executing. 0 otherwise.
935 inline int task_curr(const struct task_struct *p)
937 return cpu_curr(task_cpu(p)) == p;
940 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
941 const struct sched_class *prev_class,
944 if (prev_class != p->sched_class) {
945 if (prev_class->switched_from)
946 prev_class->switched_from(rq, p);
947 p->sched_class->switched_to(rq, p);
948 } else if (oldprio != p->prio || dl_task(p))
949 p->sched_class->prio_changed(rq, p, oldprio);
952 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
954 const struct sched_class *class;
956 if (p->sched_class == rq->curr->sched_class) {
957 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
959 for_each_class(class) {
960 if (class == rq->curr->sched_class)
962 if (class == p->sched_class) {
963 resched_task(rq->curr);
970 * A queue event has occurred, and we're going to schedule. In
971 * this case, we can save a useless back to back clock update.
973 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
974 rq->skip_clock_update = 1;
978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
980 #ifdef CONFIG_SCHED_DEBUG
982 * We should never call set_task_cpu() on a blocked task,
983 * ttwu() will sort out the placement.
985 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
986 !(task_preempt_count(p) & PREEMPT_ACTIVE));
988 #ifdef CONFIG_LOCKDEP
990 * The caller should hold either p->pi_lock or rq->lock, when changing
991 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
993 * sched_move_task() holds both and thus holding either pins the cgroup,
996 * Furthermore, all task_rq users should acquire both locks, see
999 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1000 lockdep_is_held(&task_rq(p)->lock)));
1004 trace_sched_migrate_task(p, new_cpu);
1006 if (task_cpu(p) != new_cpu) {
1007 if (p->sched_class->migrate_task_rq)
1008 p->sched_class->migrate_task_rq(p, new_cpu);
1009 p->se.nr_migrations++;
1010 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1013 __set_task_cpu(p, new_cpu);
1016 static void __migrate_swap_task(struct task_struct *p, int cpu)
1019 struct rq *src_rq, *dst_rq;
1021 src_rq = task_rq(p);
1022 dst_rq = cpu_rq(cpu);
1024 deactivate_task(src_rq, p, 0);
1025 set_task_cpu(p, cpu);
1026 activate_task(dst_rq, p, 0);
1027 check_preempt_curr(dst_rq, p, 0);
1030 * Task isn't running anymore; make it appear like we migrated
1031 * it before it went to sleep. This means on wakeup we make the
1032 * previous cpu our targer instead of where it really is.
1038 struct migration_swap_arg {
1039 struct task_struct *src_task, *dst_task;
1040 int src_cpu, dst_cpu;
1043 static int migrate_swap_stop(void *data)
1045 struct migration_swap_arg *arg = data;
1046 struct rq *src_rq, *dst_rq;
1049 src_rq = cpu_rq(arg->src_cpu);
1050 dst_rq = cpu_rq(arg->dst_cpu);
1052 double_raw_lock(&arg->src_task->pi_lock,
1053 &arg->dst_task->pi_lock);
1054 double_rq_lock(src_rq, dst_rq);
1055 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1058 if (task_cpu(arg->src_task) != arg->src_cpu)
1061 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1064 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1067 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1068 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1073 double_rq_unlock(src_rq, dst_rq);
1074 raw_spin_unlock(&arg->dst_task->pi_lock);
1075 raw_spin_unlock(&arg->src_task->pi_lock);
1081 * Cross migrate two tasks
1083 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1085 struct migration_swap_arg arg;
1088 arg = (struct migration_swap_arg){
1090 .src_cpu = task_cpu(cur),
1092 .dst_cpu = task_cpu(p),
1095 if (arg.src_cpu == arg.dst_cpu)
1099 * These three tests are all lockless; this is OK since all of them
1100 * will be re-checked with proper locks held further down the line.
1102 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1105 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1108 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1111 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1117 struct migration_arg {
1118 struct task_struct *task;
1122 static int migration_cpu_stop(void *data);
1125 * wait_task_inactive - wait for a thread to unschedule.
1127 * If @match_state is nonzero, it's the @p->state value just checked and
1128 * not expected to change. If it changes, i.e. @p might have woken up,
1129 * then return zero. When we succeed in waiting for @p to be off its CPU,
1130 * we return a positive number (its total switch count). If a second call
1131 * a short while later returns the same number, the caller can be sure that
1132 * @p has remained unscheduled the whole time.
1134 * The caller must ensure that the task *will* unschedule sometime soon,
1135 * else this function might spin for a *long* time. This function can't
1136 * be called with interrupts off, or it may introduce deadlock with
1137 * smp_call_function() if an IPI is sent by the same process we are
1138 * waiting to become inactive.
1140 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1142 unsigned long flags;
1149 * We do the initial early heuristics without holding
1150 * any task-queue locks at all. We'll only try to get
1151 * the runqueue lock when things look like they will
1157 * If the task is actively running on another CPU
1158 * still, just relax and busy-wait without holding
1161 * NOTE! Since we don't hold any locks, it's not
1162 * even sure that "rq" stays as the right runqueue!
1163 * But we don't care, since "task_running()" will
1164 * return false if the runqueue has changed and p
1165 * is actually now running somewhere else!
1167 while (task_running(rq, p)) {
1168 if (match_state && unlikely(p->state != match_state))
1174 * Ok, time to look more closely! We need the rq
1175 * lock now, to be *sure*. If we're wrong, we'll
1176 * just go back and repeat.
1178 rq = task_rq_lock(p, &flags);
1179 trace_sched_wait_task(p);
1180 running = task_running(rq, p);
1183 if (!match_state || p->state == match_state)
1184 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1185 task_rq_unlock(rq, p, &flags);
1188 * If it changed from the expected state, bail out now.
1190 if (unlikely(!ncsw))
1194 * Was it really running after all now that we
1195 * checked with the proper locks actually held?
1197 * Oops. Go back and try again..
1199 if (unlikely(running)) {
1205 * It's not enough that it's not actively running,
1206 * it must be off the runqueue _entirely_, and not
1209 * So if it was still runnable (but just not actively
1210 * running right now), it's preempted, and we should
1211 * yield - it could be a while.
1213 if (unlikely(on_rq)) {
1214 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1216 set_current_state(TASK_UNINTERRUPTIBLE);
1217 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1222 * Ahh, all good. It wasn't running, and it wasn't
1223 * runnable, which means that it will never become
1224 * running in the future either. We're all done!
1233 * kick_process - kick a running thread to enter/exit the kernel
1234 * @p: the to-be-kicked thread
1236 * Cause a process which is running on another CPU to enter
1237 * kernel-mode, without any delay. (to get signals handled.)
1239 * NOTE: this function doesn't have to take the runqueue lock,
1240 * because all it wants to ensure is that the remote task enters
1241 * the kernel. If the IPI races and the task has been migrated
1242 * to another CPU then no harm is done and the purpose has been
1245 void kick_process(struct task_struct *p)
1251 if ((cpu != smp_processor_id()) && task_curr(p))
1252 smp_send_reschedule(cpu);
1255 EXPORT_SYMBOL_GPL(kick_process);
1256 #endif /* CONFIG_SMP */
1260 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1262 static int select_fallback_rq(int cpu, struct task_struct *p)
1264 int nid = cpu_to_node(cpu);
1265 const struct cpumask *nodemask = NULL;
1266 enum { cpuset, possible, fail } state = cpuset;
1270 * If the node that the cpu is on has been offlined, cpu_to_node()
1271 * will return -1. There is no cpu on the node, and we should
1272 * select the cpu on the other node.
1275 nodemask = cpumask_of_node(nid);
1277 /* Look for allowed, online CPU in same node. */
1278 for_each_cpu(dest_cpu, nodemask) {
1279 if (!cpu_online(dest_cpu))
1281 if (!cpu_active(dest_cpu))
1283 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1289 /* Any allowed, online CPU? */
1290 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1291 if (!cpu_online(dest_cpu))
1293 if (!cpu_active(dest_cpu))
1300 /* No more Mr. Nice Guy. */
1301 cpuset_cpus_allowed_fallback(p);
1306 do_set_cpus_allowed(p, cpu_possible_mask);
1317 if (state != cpuset) {
1319 * Don't tell them about moving exiting tasks or
1320 * kernel threads (both mm NULL), since they never
1323 if (p->mm && printk_ratelimit()) {
1324 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1325 task_pid_nr(p), p->comm, cpu);
1333 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1336 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1338 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1341 * In order not to call set_task_cpu() on a blocking task we need
1342 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1345 * Since this is common to all placement strategies, this lives here.
1347 * [ this allows ->select_task() to simply return task_cpu(p) and
1348 * not worry about this generic constraint ]
1350 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1352 cpu = select_fallback_rq(task_cpu(p), p);
1357 static void update_avg(u64 *avg, u64 sample)
1359 s64 diff = sample - *avg;
1365 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1367 #ifdef CONFIG_SCHEDSTATS
1368 struct rq *rq = this_rq();
1371 int this_cpu = smp_processor_id();
1373 if (cpu == this_cpu) {
1374 schedstat_inc(rq, ttwu_local);
1375 schedstat_inc(p, se.statistics.nr_wakeups_local);
1377 struct sched_domain *sd;
1379 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1381 for_each_domain(this_cpu, sd) {
1382 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1383 schedstat_inc(sd, ttwu_wake_remote);
1390 if (wake_flags & WF_MIGRATED)
1391 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1393 #endif /* CONFIG_SMP */
1395 schedstat_inc(rq, ttwu_count);
1396 schedstat_inc(p, se.statistics.nr_wakeups);
1398 if (wake_flags & WF_SYNC)
1399 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1401 #endif /* CONFIG_SCHEDSTATS */
1404 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1406 activate_task(rq, p, en_flags);
1409 /* if a worker is waking up, notify workqueue */
1410 if (p->flags & PF_WQ_WORKER)
1411 wq_worker_waking_up(p, cpu_of(rq));
1415 * Mark the task runnable and perform wakeup-preemption.
1418 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1420 check_preempt_curr(rq, p, wake_flags);
1421 trace_sched_wakeup(p, true);
1423 p->state = TASK_RUNNING;
1425 if (p->sched_class->task_woken)
1426 p->sched_class->task_woken(rq, p);
1428 if (rq->idle_stamp) {
1429 u64 delta = rq_clock(rq) - rq->idle_stamp;
1430 u64 max = 2*rq->max_idle_balance_cost;
1432 update_avg(&rq->avg_idle, delta);
1434 if (rq->avg_idle > max)
1443 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1446 if (p->sched_contributes_to_load)
1447 rq->nr_uninterruptible--;
1450 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1451 ttwu_do_wakeup(rq, p, wake_flags);
1455 * Called in case the task @p isn't fully descheduled from its runqueue,
1456 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1457 * since all we need to do is flip p->state to TASK_RUNNING, since
1458 * the task is still ->on_rq.
1460 static int ttwu_remote(struct task_struct *p, int wake_flags)
1465 rq = __task_rq_lock(p);
1467 /* check_preempt_curr() may use rq clock */
1468 update_rq_clock(rq);
1469 ttwu_do_wakeup(rq, p, wake_flags);
1472 __task_rq_unlock(rq);
1478 static void sched_ttwu_pending(void)
1480 struct rq *rq = this_rq();
1481 struct llist_node *llist = llist_del_all(&rq->wake_list);
1482 struct task_struct *p;
1484 raw_spin_lock(&rq->lock);
1487 p = llist_entry(llist, struct task_struct, wake_entry);
1488 llist = llist_next(llist);
1489 ttwu_do_activate(rq, p, 0);
1492 raw_spin_unlock(&rq->lock);
1495 void scheduler_ipi(void)
1498 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1499 * TIF_NEED_RESCHED remotely (for the first time) will also send
1502 preempt_fold_need_resched();
1504 if (llist_empty(&this_rq()->wake_list)
1505 && !tick_nohz_full_cpu(smp_processor_id())
1506 && !got_nohz_idle_kick())
1510 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1511 * traditionally all their work was done from the interrupt return
1512 * path. Now that we actually do some work, we need to make sure
1515 * Some archs already do call them, luckily irq_enter/exit nest
1518 * Arguably we should visit all archs and update all handlers,
1519 * however a fair share of IPIs are still resched only so this would
1520 * somewhat pessimize the simple resched case.
1523 tick_nohz_full_check();
1524 sched_ttwu_pending();
1527 * Check if someone kicked us for doing the nohz idle load balance.
1529 if (unlikely(got_nohz_idle_kick())) {
1530 this_rq()->idle_balance = 1;
1531 raise_softirq_irqoff(SCHED_SOFTIRQ);
1536 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1538 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1539 smp_send_reschedule(cpu);
1542 bool cpus_share_cache(int this_cpu, int that_cpu)
1544 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1546 #endif /* CONFIG_SMP */
1548 static void ttwu_queue(struct task_struct *p, int cpu)
1550 struct rq *rq = cpu_rq(cpu);
1552 #if defined(CONFIG_SMP)
1553 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1554 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1555 ttwu_queue_remote(p, cpu);
1560 raw_spin_lock(&rq->lock);
1561 ttwu_do_activate(rq, p, 0);
1562 raw_spin_unlock(&rq->lock);
1566 * try_to_wake_up - wake up a thread
1567 * @p: the thread to be awakened
1568 * @state: the mask of task states that can be woken
1569 * @wake_flags: wake modifier flags (WF_*)
1571 * Put it on the run-queue if it's not already there. The "current"
1572 * thread is always on the run-queue (except when the actual
1573 * re-schedule is in progress), and as such you're allowed to do
1574 * the simpler "current->state = TASK_RUNNING" to mark yourself
1575 * runnable without the overhead of this.
1577 * Return: %true if @p was woken up, %false if it was already running.
1578 * or @state didn't match @p's state.
1581 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1583 unsigned long flags;
1584 int cpu, success = 0;
1587 * If we are going to wake up a thread waiting for CONDITION we
1588 * need to ensure that CONDITION=1 done by the caller can not be
1589 * reordered with p->state check below. This pairs with mb() in
1590 * set_current_state() the waiting thread does.
1592 smp_mb__before_spinlock();
1593 raw_spin_lock_irqsave(&p->pi_lock, flags);
1594 if (!(p->state & state))
1597 success = 1; /* we're going to change ->state */
1600 if (p->on_rq && ttwu_remote(p, wake_flags))
1605 * If the owning (remote) cpu is still in the middle of schedule() with
1606 * this task as prev, wait until its done referencing the task.
1611 * Pairs with the smp_wmb() in finish_lock_switch().
1615 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1616 p->state = TASK_WAKING;
1618 if (p->sched_class->task_waking)
1619 p->sched_class->task_waking(p);
1621 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1622 if (task_cpu(p) != cpu) {
1623 wake_flags |= WF_MIGRATED;
1624 set_task_cpu(p, cpu);
1626 #endif /* CONFIG_SMP */
1630 ttwu_stat(p, cpu, wake_flags);
1632 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1638 * try_to_wake_up_local - try to wake up a local task with rq lock held
1639 * @p: the thread to be awakened
1641 * Put @p on the run-queue if it's not already there. The caller must
1642 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1645 static void try_to_wake_up_local(struct task_struct *p)
1647 struct rq *rq = task_rq(p);
1649 if (WARN_ON_ONCE(rq != this_rq()) ||
1650 WARN_ON_ONCE(p == current))
1653 lockdep_assert_held(&rq->lock);
1655 if (!raw_spin_trylock(&p->pi_lock)) {
1656 raw_spin_unlock(&rq->lock);
1657 raw_spin_lock(&p->pi_lock);
1658 raw_spin_lock(&rq->lock);
1661 if (!(p->state & TASK_NORMAL))
1665 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1667 ttwu_do_wakeup(rq, p, 0);
1668 ttwu_stat(p, smp_processor_id(), 0);
1670 raw_spin_unlock(&p->pi_lock);
1674 * wake_up_process - Wake up a specific process
1675 * @p: The process to be woken up.
1677 * Attempt to wake up the nominated process and move it to the set of runnable
1680 * Return: 1 if the process was woken up, 0 if it was already running.
1682 * It may be assumed that this function implies a write memory barrier before
1683 * changing the task state if and only if any tasks are woken up.
1685 int wake_up_process(struct task_struct *p)
1687 WARN_ON(task_is_stopped_or_traced(p));
1688 return try_to_wake_up(p, TASK_NORMAL, 0);
1690 EXPORT_SYMBOL(wake_up_process);
1692 int wake_up_state(struct task_struct *p, unsigned int state)
1694 return try_to_wake_up(p, state, 0);
1698 * Perform scheduler related setup for a newly forked process p.
1699 * p is forked by current.
1701 * __sched_fork() is basic setup used by init_idle() too:
1703 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1708 p->se.exec_start = 0;
1709 p->se.sum_exec_runtime = 0;
1710 p->se.prev_sum_exec_runtime = 0;
1711 p->se.nr_migrations = 0;
1713 INIT_LIST_HEAD(&p->se.group_node);
1715 #ifdef CONFIG_SCHEDSTATS
1716 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1719 RB_CLEAR_NODE(&p->dl.rb_node);
1720 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1721 p->dl.dl_runtime = p->dl.runtime = 0;
1722 p->dl.dl_deadline = p->dl.deadline = 0;
1723 p->dl.dl_period = 0;
1726 INIT_LIST_HEAD(&p->rt.run_list);
1728 #ifdef CONFIG_PREEMPT_NOTIFIERS
1729 INIT_HLIST_HEAD(&p->preempt_notifiers);
1732 #ifdef CONFIG_NUMA_BALANCING
1733 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1734 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1735 p->mm->numa_scan_seq = 0;
1738 if (clone_flags & CLONE_VM)
1739 p->numa_preferred_nid = current->numa_preferred_nid;
1741 p->numa_preferred_nid = -1;
1743 p->node_stamp = 0ULL;
1744 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1745 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1746 p->numa_work.next = &p->numa_work;
1747 p->numa_faults = NULL;
1748 p->numa_faults_buffer = NULL;
1750 INIT_LIST_HEAD(&p->numa_entry);
1751 p->numa_group = NULL;
1752 #endif /* CONFIG_NUMA_BALANCING */
1755 #ifdef CONFIG_NUMA_BALANCING
1756 #ifdef CONFIG_SCHED_DEBUG
1757 void set_numabalancing_state(bool enabled)
1760 sched_feat_set("NUMA");
1762 sched_feat_set("NO_NUMA");
1765 __read_mostly bool numabalancing_enabled;
1767 void set_numabalancing_state(bool enabled)
1769 numabalancing_enabled = enabled;
1771 #endif /* CONFIG_SCHED_DEBUG */
1772 #endif /* CONFIG_NUMA_BALANCING */
1775 * fork()/clone()-time setup:
1777 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1779 unsigned long flags;
1780 int cpu = get_cpu();
1782 __sched_fork(clone_flags, p);
1784 * We mark the process as running here. This guarantees that
1785 * nobody will actually run it, and a signal or other external
1786 * event cannot wake it up and insert it on the runqueue either.
1788 p->state = TASK_RUNNING;
1791 * Make sure we do not leak PI boosting priority to the child.
1793 p->prio = current->normal_prio;
1796 * Revert to default priority/policy on fork if requested.
1798 if (unlikely(p->sched_reset_on_fork)) {
1799 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1800 p->policy = SCHED_NORMAL;
1801 p->static_prio = NICE_TO_PRIO(0);
1803 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1804 p->static_prio = NICE_TO_PRIO(0);
1806 p->prio = p->normal_prio = __normal_prio(p);
1810 * We don't need the reset flag anymore after the fork. It has
1811 * fulfilled its duty:
1813 p->sched_reset_on_fork = 0;
1816 if (dl_prio(p->prio)) {
1819 } else if (rt_prio(p->prio)) {
1820 p->sched_class = &rt_sched_class;
1822 p->sched_class = &fair_sched_class;
1825 if (p->sched_class->task_fork)
1826 p->sched_class->task_fork(p);
1829 * The child is not yet in the pid-hash so no cgroup attach races,
1830 * and the cgroup is pinned to this child due to cgroup_fork()
1831 * is ran before sched_fork().
1833 * Silence PROVE_RCU.
1835 raw_spin_lock_irqsave(&p->pi_lock, flags);
1836 set_task_cpu(p, cpu);
1837 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1839 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1840 if (likely(sched_info_on()))
1841 memset(&p->sched_info, 0, sizeof(p->sched_info));
1843 #if defined(CONFIG_SMP)
1846 init_task_preempt_count(p);
1848 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1849 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1856 unsigned long to_ratio(u64 period, u64 runtime)
1858 if (runtime == RUNTIME_INF)
1862 * Doing this here saves a lot of checks in all
1863 * the calling paths, and returning zero seems
1864 * safe for them anyway.
1869 return div64_u64(runtime << 20, period);
1873 inline struct dl_bw *dl_bw_of(int i)
1875 return &cpu_rq(i)->rd->dl_bw;
1878 static inline int dl_bw_cpus(int i)
1880 struct root_domain *rd = cpu_rq(i)->rd;
1883 for_each_cpu_and(i, rd->span, cpu_active_mask)
1889 inline struct dl_bw *dl_bw_of(int i)
1891 return &cpu_rq(i)->dl.dl_bw;
1894 static inline int dl_bw_cpus(int i)
1901 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1903 dl_b->total_bw -= tsk_bw;
1907 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1909 dl_b->total_bw += tsk_bw;
1913 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1915 return dl_b->bw != -1 &&
1916 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1920 * We must be sure that accepting a new task (or allowing changing the
1921 * parameters of an existing one) is consistent with the bandwidth
1922 * constraints. If yes, this function also accordingly updates the currently
1923 * allocated bandwidth to reflect the new situation.
1925 * This function is called while holding p's rq->lock.
1927 static int dl_overflow(struct task_struct *p, int policy,
1928 const struct sched_attr *attr)
1931 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1932 u64 period = attr->sched_period;
1933 u64 runtime = attr->sched_runtime;
1934 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1937 if (new_bw == p->dl.dl_bw)
1941 * Either if a task, enters, leave, or stays -deadline but changes
1942 * its parameters, we may need to update accordingly the total
1943 * allocated bandwidth of the container.
1945 raw_spin_lock(&dl_b->lock);
1946 cpus = dl_bw_cpus(task_cpu(p));
1947 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1948 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1949 __dl_add(dl_b, new_bw);
1951 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1952 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1953 __dl_clear(dl_b, p->dl.dl_bw);
1954 __dl_add(dl_b, new_bw);
1956 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1957 __dl_clear(dl_b, p->dl.dl_bw);
1960 raw_spin_unlock(&dl_b->lock);
1965 extern void init_dl_bw(struct dl_bw *dl_b);
1968 * wake_up_new_task - wake up a newly created task for the first time.
1970 * This function will do some initial scheduler statistics housekeeping
1971 * that must be done for every newly created context, then puts the task
1972 * on the runqueue and wakes it.
1974 void wake_up_new_task(struct task_struct *p)
1976 unsigned long flags;
1979 raw_spin_lock_irqsave(&p->pi_lock, flags);
1982 * Fork balancing, do it here and not earlier because:
1983 * - cpus_allowed can change in the fork path
1984 * - any previously selected cpu might disappear through hotplug
1986 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
1989 /* Initialize new task's runnable average */
1990 init_task_runnable_average(p);
1991 rq = __task_rq_lock(p);
1992 activate_task(rq, p, 0);
1994 trace_sched_wakeup_new(p, true);
1995 check_preempt_curr(rq, p, WF_FORK);
1997 if (p->sched_class->task_woken)
1998 p->sched_class->task_woken(rq, p);
2000 task_rq_unlock(rq, p, &flags);
2003 #ifdef CONFIG_PREEMPT_NOTIFIERS
2006 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2007 * @notifier: notifier struct to register
2009 void preempt_notifier_register(struct preempt_notifier *notifier)
2011 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2013 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2016 * preempt_notifier_unregister - no longer interested in preemption notifications
2017 * @notifier: notifier struct to unregister
2019 * This is safe to call from within a preemption notifier.
2021 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2023 hlist_del(¬ifier->link);
2025 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2027 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2029 struct preempt_notifier *notifier;
2031 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2032 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2036 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2037 struct task_struct *next)
2039 struct preempt_notifier *notifier;
2041 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2042 notifier->ops->sched_out(notifier, next);
2045 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2047 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2052 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2053 struct task_struct *next)
2057 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2060 * prepare_task_switch - prepare to switch tasks
2061 * @rq: the runqueue preparing to switch
2062 * @prev: the current task that is being switched out
2063 * @next: the task we are going to switch to.
2065 * This is called with the rq lock held and interrupts off. It must
2066 * be paired with a subsequent finish_task_switch after the context
2069 * prepare_task_switch sets up locking and calls architecture specific
2073 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2074 struct task_struct *next)
2076 trace_sched_switch(prev, next);
2077 sched_info_switch(rq, prev, next);
2078 perf_event_task_sched_out(prev, next);
2079 fire_sched_out_preempt_notifiers(prev, next);
2080 prepare_lock_switch(rq, next);
2081 prepare_arch_switch(next);
2085 * finish_task_switch - clean up after a task-switch
2086 * @rq: runqueue associated with task-switch
2087 * @prev: the thread we just switched away from.
2089 * finish_task_switch must be called after the context switch, paired
2090 * with a prepare_task_switch call before the context switch.
2091 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2092 * and do any other architecture-specific cleanup actions.
2094 * Note that we may have delayed dropping an mm in context_switch(). If
2095 * so, we finish that here outside of the runqueue lock. (Doing it
2096 * with the lock held can cause deadlocks; see schedule() for
2099 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2100 __releases(rq->lock)
2102 struct mm_struct *mm = rq->prev_mm;
2108 * A task struct has one reference for the use as "current".
2109 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2110 * schedule one last time. The schedule call will never return, and
2111 * the scheduled task must drop that reference.
2112 * The test for TASK_DEAD must occur while the runqueue locks are
2113 * still held, otherwise prev could be scheduled on another cpu, die
2114 * there before we look at prev->state, and then the reference would
2116 * Manfred Spraul <manfred@colorfullife.com>
2118 prev_state = prev->state;
2119 vtime_task_switch(prev);
2120 finish_arch_switch(prev);
2121 perf_event_task_sched_in(prev, current);
2122 finish_lock_switch(rq, prev);
2123 finish_arch_post_lock_switch();
2125 fire_sched_in_preempt_notifiers(current);
2128 if (unlikely(prev_state == TASK_DEAD)) {
2129 task_numa_free(prev);
2131 if (prev->sched_class->task_dead)
2132 prev->sched_class->task_dead(prev);
2135 * Remove function-return probe instances associated with this
2136 * task and put them back on the free list.
2138 kprobe_flush_task(prev);
2139 put_task_struct(prev);
2142 tick_nohz_task_switch(current);
2147 /* assumes rq->lock is held */
2148 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2150 if (prev->sched_class->pre_schedule)
2151 prev->sched_class->pre_schedule(rq, prev);
2154 /* rq->lock is NOT held, but preemption is disabled */
2155 static inline void post_schedule(struct rq *rq)
2157 if (rq->post_schedule) {
2158 unsigned long flags;
2160 raw_spin_lock_irqsave(&rq->lock, flags);
2161 if (rq->curr->sched_class->post_schedule)
2162 rq->curr->sched_class->post_schedule(rq);
2163 raw_spin_unlock_irqrestore(&rq->lock, flags);
2165 rq->post_schedule = 0;
2171 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2175 static inline void post_schedule(struct rq *rq)
2182 * schedule_tail - first thing a freshly forked thread must call.
2183 * @prev: the thread we just switched away from.
2185 asmlinkage void schedule_tail(struct task_struct *prev)
2186 __releases(rq->lock)
2188 struct rq *rq = this_rq();
2190 finish_task_switch(rq, prev);
2193 * FIXME: do we need to worry about rq being invalidated by the
2198 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2199 /* In this case, finish_task_switch does not reenable preemption */
2202 if (current->set_child_tid)
2203 put_user(task_pid_vnr(current), current->set_child_tid);
2207 * context_switch - switch to the new MM and the new
2208 * thread's register state.
2211 context_switch(struct rq *rq, struct task_struct *prev,
2212 struct task_struct *next)
2214 struct mm_struct *mm, *oldmm;
2216 prepare_task_switch(rq, prev, next);
2219 oldmm = prev->active_mm;
2221 * For paravirt, this is coupled with an exit in switch_to to
2222 * combine the page table reload and the switch backend into
2225 arch_start_context_switch(prev);
2228 next->active_mm = oldmm;
2229 atomic_inc(&oldmm->mm_count);
2230 enter_lazy_tlb(oldmm, next);
2232 switch_mm(oldmm, mm, next);
2235 prev->active_mm = NULL;
2236 rq->prev_mm = oldmm;
2239 * Since the runqueue lock will be released by the next
2240 * task (which is an invalid locking op but in the case
2241 * of the scheduler it's an obvious special-case), so we
2242 * do an early lockdep release here:
2244 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2245 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2248 context_tracking_task_switch(prev, next);
2249 /* Here we just switch the register state and the stack. */
2250 switch_to(prev, next, prev);
2254 * this_rq must be evaluated again because prev may have moved
2255 * CPUs since it called schedule(), thus the 'rq' on its stack
2256 * frame will be invalid.
2258 finish_task_switch(this_rq(), prev);
2262 * nr_running and nr_context_switches:
2264 * externally visible scheduler statistics: current number of runnable
2265 * threads, total number of context switches performed since bootup.
2267 unsigned long nr_running(void)
2269 unsigned long i, sum = 0;
2271 for_each_online_cpu(i)
2272 sum += cpu_rq(i)->nr_running;
2277 unsigned long long nr_context_switches(void)
2280 unsigned long long sum = 0;
2282 for_each_possible_cpu(i)
2283 sum += cpu_rq(i)->nr_switches;
2288 unsigned long nr_iowait(void)
2290 unsigned long i, sum = 0;
2292 for_each_possible_cpu(i)
2293 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2298 unsigned long nr_iowait_cpu(int cpu)
2300 struct rq *this = cpu_rq(cpu);
2301 return atomic_read(&this->nr_iowait);
2307 * sched_exec - execve() is a valuable balancing opportunity, because at
2308 * this point the task has the smallest effective memory and cache footprint.
2310 void sched_exec(void)
2312 struct task_struct *p = current;
2313 unsigned long flags;
2316 raw_spin_lock_irqsave(&p->pi_lock, flags);
2317 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2318 if (dest_cpu == smp_processor_id())
2321 if (likely(cpu_active(dest_cpu))) {
2322 struct migration_arg arg = { p, dest_cpu };
2324 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2325 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2329 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2334 DEFINE_PER_CPU(struct kernel_stat, kstat);
2335 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2337 EXPORT_PER_CPU_SYMBOL(kstat);
2338 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2341 * Return any ns on the sched_clock that have not yet been accounted in
2342 * @p in case that task is currently running.
2344 * Called with task_rq_lock() held on @rq.
2346 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2350 if (task_current(rq, p)) {
2351 update_rq_clock(rq);
2352 ns = rq_clock_task(rq) - p->se.exec_start;
2360 unsigned long long task_delta_exec(struct task_struct *p)
2362 unsigned long flags;
2366 rq = task_rq_lock(p, &flags);
2367 ns = do_task_delta_exec(p, rq);
2368 task_rq_unlock(rq, p, &flags);
2374 * Return accounted runtime for the task.
2375 * In case the task is currently running, return the runtime plus current's
2376 * pending runtime that have not been accounted yet.
2378 unsigned long long task_sched_runtime(struct task_struct *p)
2380 unsigned long flags;
2384 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2386 * 64-bit doesn't need locks to atomically read a 64bit value.
2387 * So we have a optimization chance when the task's delta_exec is 0.
2388 * Reading ->on_cpu is racy, but this is ok.
2390 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2391 * If we race with it entering cpu, unaccounted time is 0. This is
2392 * indistinguishable from the read occurring a few cycles earlier.
2395 return p->se.sum_exec_runtime;
2398 rq = task_rq_lock(p, &flags);
2399 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2400 task_rq_unlock(rq, p, &flags);
2406 * This function gets called by the timer code, with HZ frequency.
2407 * We call it with interrupts disabled.
2409 void scheduler_tick(void)
2411 int cpu = smp_processor_id();
2412 struct rq *rq = cpu_rq(cpu);
2413 struct task_struct *curr = rq->curr;
2417 raw_spin_lock(&rq->lock);
2418 update_rq_clock(rq);
2419 curr->sched_class->task_tick(rq, curr, 0);
2420 update_cpu_load_active(rq);
2421 raw_spin_unlock(&rq->lock);
2423 perf_event_task_tick();
2426 rq->idle_balance = idle_cpu(cpu);
2427 trigger_load_balance(rq);
2429 rq_last_tick_reset(rq);
2432 #ifdef CONFIG_NO_HZ_FULL
2434 * scheduler_tick_max_deferment
2436 * Keep at least one tick per second when a single
2437 * active task is running because the scheduler doesn't
2438 * yet completely support full dynticks environment.
2440 * This makes sure that uptime, CFS vruntime, load
2441 * balancing, etc... continue to move forward, even
2442 * with a very low granularity.
2444 * Return: Maximum deferment in nanoseconds.
2446 u64 scheduler_tick_max_deferment(void)
2448 struct rq *rq = this_rq();
2449 unsigned long next, now = ACCESS_ONCE(jiffies);
2451 next = rq->last_sched_tick + HZ;
2453 if (time_before_eq(next, now))
2456 return jiffies_to_usecs(next - now) * NSEC_PER_USEC;
2460 notrace unsigned long get_parent_ip(unsigned long addr)
2462 if (in_lock_functions(addr)) {
2463 addr = CALLER_ADDR2;
2464 if (in_lock_functions(addr))
2465 addr = CALLER_ADDR3;
2470 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2471 defined(CONFIG_PREEMPT_TRACER))
2473 void __kprobes preempt_count_add(int val)
2475 #ifdef CONFIG_DEBUG_PREEMPT
2479 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2482 __preempt_count_add(val);
2483 #ifdef CONFIG_DEBUG_PREEMPT
2485 * Spinlock count overflowing soon?
2487 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2490 if (preempt_count() == val)
2491 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2493 EXPORT_SYMBOL(preempt_count_add);
2495 void __kprobes preempt_count_sub(int val)
2497 #ifdef CONFIG_DEBUG_PREEMPT
2501 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2504 * Is the spinlock portion underflowing?
2506 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2507 !(preempt_count() & PREEMPT_MASK)))
2511 if (preempt_count() == val)
2512 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2513 __preempt_count_sub(val);
2515 EXPORT_SYMBOL(preempt_count_sub);
2520 * Print scheduling while atomic bug:
2522 static noinline void __schedule_bug(struct task_struct *prev)
2524 if (oops_in_progress)
2527 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2528 prev->comm, prev->pid, preempt_count());
2530 debug_show_held_locks(prev);
2532 if (irqs_disabled())
2533 print_irqtrace_events(prev);
2535 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2539 * Various schedule()-time debugging checks and statistics:
2541 static inline void schedule_debug(struct task_struct *prev)
2544 * Test if we are atomic. Since do_exit() needs to call into
2545 * schedule() atomically, we ignore that path. Otherwise whine
2546 * if we are scheduling when we should not.
2548 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2549 __schedule_bug(prev);
2552 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2554 schedstat_inc(this_rq(), sched_count);
2557 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2559 if (prev->on_rq || rq->skip_clock_update < 0)
2560 update_rq_clock(rq);
2561 prev->sched_class->put_prev_task(rq, prev);
2565 * Pick up the highest-prio task:
2567 static inline struct task_struct *
2568 pick_next_task(struct rq *rq)
2570 const struct sched_class *class;
2571 struct task_struct *p;
2574 * Optimization: we know that if all tasks are in
2575 * the fair class we can call that function directly:
2577 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2578 p = fair_sched_class.pick_next_task(rq);
2583 for_each_class(class) {
2584 p = class->pick_next_task(rq);
2589 BUG(); /* the idle class will always have a runnable task */
2593 * __schedule() is the main scheduler function.
2595 * The main means of driving the scheduler and thus entering this function are:
2597 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2599 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2600 * paths. For example, see arch/x86/entry_64.S.
2602 * To drive preemption between tasks, the scheduler sets the flag in timer
2603 * interrupt handler scheduler_tick().
2605 * 3. Wakeups don't really cause entry into schedule(). They add a
2606 * task to the run-queue and that's it.
2608 * Now, if the new task added to the run-queue preempts the current
2609 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2610 * called on the nearest possible occasion:
2612 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2614 * - in syscall or exception context, at the next outmost
2615 * preempt_enable(). (this might be as soon as the wake_up()'s
2618 * - in IRQ context, return from interrupt-handler to
2619 * preemptible context
2621 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2624 * - cond_resched() call
2625 * - explicit schedule() call
2626 * - return from syscall or exception to user-space
2627 * - return from interrupt-handler to user-space
2629 static void __sched __schedule(void)
2631 struct task_struct *prev, *next;
2632 unsigned long *switch_count;
2638 cpu = smp_processor_id();
2640 rcu_note_context_switch(cpu);
2643 schedule_debug(prev);
2645 if (sched_feat(HRTICK))
2649 * Make sure that signal_pending_state()->signal_pending() below
2650 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2651 * done by the caller to avoid the race with signal_wake_up().
2653 smp_mb__before_spinlock();
2654 raw_spin_lock_irq(&rq->lock);
2656 switch_count = &prev->nivcsw;
2657 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2658 if (unlikely(signal_pending_state(prev->state, prev))) {
2659 prev->state = TASK_RUNNING;
2661 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2665 * If a worker went to sleep, notify and ask workqueue
2666 * whether it wants to wake up a task to maintain
2669 if (prev->flags & PF_WQ_WORKER) {
2670 struct task_struct *to_wakeup;
2672 to_wakeup = wq_worker_sleeping(prev, cpu);
2674 try_to_wake_up_local(to_wakeup);
2677 switch_count = &prev->nvcsw;
2680 pre_schedule(rq, prev);
2682 if (unlikely(!rq->nr_running))
2683 idle_balance(cpu, rq);
2685 put_prev_task(rq, prev);
2686 next = pick_next_task(rq);
2687 clear_tsk_need_resched(prev);
2688 clear_preempt_need_resched();
2689 rq->skip_clock_update = 0;
2691 if (likely(prev != next)) {
2696 context_switch(rq, prev, next); /* unlocks the rq */
2698 * The context switch have flipped the stack from under us
2699 * and restored the local variables which were saved when
2700 * this task called schedule() in the past. prev == current
2701 * is still correct, but it can be moved to another cpu/rq.
2703 cpu = smp_processor_id();
2706 raw_spin_unlock_irq(&rq->lock);
2710 sched_preempt_enable_no_resched();
2715 static inline void sched_submit_work(struct task_struct *tsk)
2717 if (!tsk->state || tsk_is_pi_blocked(tsk))
2720 * If we are going to sleep and we have plugged IO queued,
2721 * make sure to submit it to avoid deadlocks.
2723 if (blk_needs_flush_plug(tsk))
2724 blk_schedule_flush_plug(tsk);
2727 asmlinkage void __sched schedule(void)
2729 struct task_struct *tsk = current;
2731 sched_submit_work(tsk);
2734 EXPORT_SYMBOL(schedule);
2736 #ifdef CONFIG_CONTEXT_TRACKING
2737 asmlinkage void __sched schedule_user(void)
2740 * If we come here after a random call to set_need_resched(),
2741 * or we have been woken up remotely but the IPI has not yet arrived,
2742 * we haven't yet exited the RCU idle mode. Do it here manually until
2743 * we find a better solution.
2752 * schedule_preempt_disabled - called with preemption disabled
2754 * Returns with preemption disabled. Note: preempt_count must be 1
2756 void __sched schedule_preempt_disabled(void)
2758 sched_preempt_enable_no_resched();
2763 #ifdef CONFIG_PREEMPT
2765 * this is the entry point to schedule() from in-kernel preemption
2766 * off of preempt_enable. Kernel preemptions off return from interrupt
2767 * occur there and call schedule directly.
2769 asmlinkage void __sched notrace preempt_schedule(void)
2772 * If there is a non-zero preempt_count or interrupts are disabled,
2773 * we do not want to preempt the current task. Just return..
2775 if (likely(!preemptible()))
2779 __preempt_count_add(PREEMPT_ACTIVE);
2781 __preempt_count_sub(PREEMPT_ACTIVE);
2784 * Check again in case we missed a preemption opportunity
2785 * between schedule and now.
2788 } while (need_resched());
2790 EXPORT_SYMBOL(preempt_schedule);
2791 #endif /* CONFIG_PREEMPT */
2794 * this is the entry point to schedule() from kernel preemption
2795 * off of irq context.
2796 * Note, that this is called and return with irqs disabled. This will
2797 * protect us against recursive calling from irq.
2799 asmlinkage void __sched preempt_schedule_irq(void)
2801 enum ctx_state prev_state;
2803 /* Catch callers which need to be fixed */
2804 BUG_ON(preempt_count() || !irqs_disabled());
2806 prev_state = exception_enter();
2809 __preempt_count_add(PREEMPT_ACTIVE);
2812 local_irq_disable();
2813 __preempt_count_sub(PREEMPT_ACTIVE);
2816 * Check again in case we missed a preemption opportunity
2817 * between schedule and now.
2820 } while (need_resched());
2822 exception_exit(prev_state);
2825 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2828 return try_to_wake_up(curr->private, mode, wake_flags);
2830 EXPORT_SYMBOL(default_wake_function);
2833 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2835 unsigned long flags;
2838 init_waitqueue_entry(&wait, current);
2840 __set_current_state(state);
2842 spin_lock_irqsave(&q->lock, flags);
2843 __add_wait_queue(q, &wait);
2844 spin_unlock(&q->lock);
2845 timeout = schedule_timeout(timeout);
2846 spin_lock_irq(&q->lock);
2847 __remove_wait_queue(q, &wait);
2848 spin_unlock_irqrestore(&q->lock, flags);
2853 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2855 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2857 EXPORT_SYMBOL(interruptible_sleep_on);
2860 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2862 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2864 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2866 void __sched sleep_on(wait_queue_head_t *q)
2868 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2870 EXPORT_SYMBOL(sleep_on);
2872 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2874 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2876 EXPORT_SYMBOL(sleep_on_timeout);
2878 #ifdef CONFIG_RT_MUTEXES
2881 * rt_mutex_setprio - set the current priority of a task
2883 * @prio: prio value (kernel-internal form)
2885 * This function changes the 'effective' priority of a task. It does
2886 * not touch ->normal_prio like __setscheduler().
2888 * Used by the rt_mutex code to implement priority inheritance logic.
2890 void rt_mutex_setprio(struct task_struct *p, int prio)
2892 int oldprio, on_rq, running, enqueue_flag = 0;
2894 const struct sched_class *prev_class;
2896 BUG_ON(prio > MAX_PRIO);
2898 rq = __task_rq_lock(p);
2901 * Idle task boosting is a nono in general. There is one
2902 * exception, when PREEMPT_RT and NOHZ is active:
2904 * The idle task calls get_next_timer_interrupt() and holds
2905 * the timer wheel base->lock on the CPU and another CPU wants
2906 * to access the timer (probably to cancel it). We can safely
2907 * ignore the boosting request, as the idle CPU runs this code
2908 * with interrupts disabled and will complete the lock
2909 * protected section without being interrupted. So there is no
2910 * real need to boost.
2912 if (unlikely(p == rq->idle)) {
2913 WARN_ON(p != rq->curr);
2914 WARN_ON(p->pi_blocked_on);
2918 trace_sched_pi_setprio(p, prio);
2919 p->pi_top_task = rt_mutex_get_top_task(p);
2921 prev_class = p->sched_class;
2923 running = task_current(rq, p);
2925 dequeue_task(rq, p, 0);
2927 p->sched_class->put_prev_task(rq, p);
2930 * Boosting condition are:
2931 * 1. -rt task is running and holds mutex A
2932 * --> -dl task blocks on mutex A
2934 * 2. -dl task is running and holds mutex A
2935 * --> -dl task blocks on mutex A and could preempt the
2938 if (dl_prio(prio)) {
2939 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2940 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2941 p->dl.dl_boosted = 1;
2942 p->dl.dl_throttled = 0;
2943 enqueue_flag = ENQUEUE_REPLENISH;
2945 p->dl.dl_boosted = 0;
2946 p->sched_class = &dl_sched_class;
2947 } else if (rt_prio(prio)) {
2948 if (dl_prio(oldprio))
2949 p->dl.dl_boosted = 0;
2951 enqueue_flag = ENQUEUE_HEAD;
2952 p->sched_class = &rt_sched_class;
2954 if (dl_prio(oldprio))
2955 p->dl.dl_boosted = 0;
2956 p->sched_class = &fair_sched_class;
2962 p->sched_class->set_curr_task(rq);
2964 enqueue_task(rq, p, enqueue_flag);
2966 check_class_changed(rq, p, prev_class, oldprio);
2968 __task_rq_unlock(rq);
2972 void set_user_nice(struct task_struct *p, long nice)
2974 int old_prio, delta, on_rq;
2975 unsigned long flags;
2978 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
2981 * We have to be careful, if called from sys_setpriority(),
2982 * the task might be in the middle of scheduling on another CPU.
2984 rq = task_rq_lock(p, &flags);
2986 * The RT priorities are set via sched_setscheduler(), but we still
2987 * allow the 'normal' nice value to be set - but as expected
2988 * it wont have any effect on scheduling until the task is
2989 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
2991 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2992 p->static_prio = NICE_TO_PRIO(nice);
2997 dequeue_task(rq, p, 0);
2999 p->static_prio = NICE_TO_PRIO(nice);
3002 p->prio = effective_prio(p);
3003 delta = p->prio - old_prio;
3006 enqueue_task(rq, p, 0);
3008 * If the task increased its priority or is running and
3009 * lowered its priority, then reschedule its CPU:
3011 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3012 resched_task(rq->curr);
3015 task_rq_unlock(rq, p, &flags);
3017 EXPORT_SYMBOL(set_user_nice);
3020 * can_nice - check if a task can reduce its nice value
3024 int can_nice(const struct task_struct *p, const int nice)
3026 /* convert nice value [19,-20] to rlimit style value [1,40] */
3027 int nice_rlim = 20 - nice;
3029 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3030 capable(CAP_SYS_NICE));
3033 #ifdef __ARCH_WANT_SYS_NICE
3036 * sys_nice - change the priority of the current process.
3037 * @increment: priority increment
3039 * sys_setpriority is a more generic, but much slower function that
3040 * does similar things.
3042 SYSCALL_DEFINE1(nice, int, increment)
3047 * Setpriority might change our priority at the same moment.
3048 * We don't have to worry. Conceptually one call occurs first
3049 * and we have a single winner.
3051 if (increment < -40)
3056 nice = TASK_NICE(current) + increment;
3062 if (increment < 0 && !can_nice(current, nice))
3065 retval = security_task_setnice(current, nice);
3069 set_user_nice(current, nice);
3076 * task_prio - return the priority value of a given task.
3077 * @p: the task in question.
3079 * Return: The priority value as seen by users in /proc.
3080 * RT tasks are offset by -200. Normal tasks are centered
3081 * around 0, value goes from -16 to +15.
3083 int task_prio(const struct task_struct *p)
3085 return p->prio - MAX_RT_PRIO;
3089 * task_nice - return the nice value of a given task.
3090 * @p: the task in question.
3092 * Return: The nice value [ -20 ... 0 ... 19 ].
3094 int task_nice(const struct task_struct *p)
3096 return TASK_NICE(p);
3098 EXPORT_SYMBOL(task_nice);
3101 * idle_cpu - is a given cpu idle currently?
3102 * @cpu: the processor in question.
3104 * Return: 1 if the CPU is currently idle. 0 otherwise.
3106 int idle_cpu(int cpu)
3108 struct rq *rq = cpu_rq(cpu);
3110 if (rq->curr != rq->idle)
3117 if (!llist_empty(&rq->wake_list))
3125 * idle_task - return the idle task for a given cpu.
3126 * @cpu: the processor in question.
3128 * Return: The idle task for the cpu @cpu.
3130 struct task_struct *idle_task(int cpu)
3132 return cpu_rq(cpu)->idle;
3136 * find_process_by_pid - find a process with a matching PID value.
3137 * @pid: the pid in question.
3139 * The task of @pid, if found. %NULL otherwise.
3141 static struct task_struct *find_process_by_pid(pid_t pid)
3143 return pid ? find_task_by_vpid(pid) : current;
3147 * This function initializes the sched_dl_entity of a newly becoming
3148 * SCHED_DEADLINE task.
3150 * Only the static values are considered here, the actual runtime and the
3151 * absolute deadline will be properly calculated when the task is enqueued
3152 * for the first time with its new policy.
3155 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3157 struct sched_dl_entity *dl_se = &p->dl;
3159 init_dl_task_timer(dl_se);
3160 dl_se->dl_runtime = attr->sched_runtime;
3161 dl_se->dl_deadline = attr->sched_deadline;
3162 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3163 dl_se->flags = attr->sched_flags;
3164 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3165 dl_se->dl_throttled = 0;
3169 /* Actually do priority change: must hold pi & rq lock. */
3170 static void __setscheduler(struct rq *rq, struct task_struct *p,
3171 const struct sched_attr *attr)
3173 int policy = attr->sched_policy;
3177 if (dl_policy(policy))
3178 __setparam_dl(p, attr);
3179 else if (rt_policy(policy))
3180 p->rt_priority = attr->sched_priority;
3182 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3184 p->normal_prio = normal_prio(p);
3185 p->prio = rt_mutex_getprio(p);
3187 if (dl_prio(p->prio))
3188 p->sched_class = &dl_sched_class;
3189 else if (rt_prio(p->prio))
3190 p->sched_class = &rt_sched_class;
3192 p->sched_class = &fair_sched_class;
3198 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3200 struct sched_dl_entity *dl_se = &p->dl;
3202 attr->sched_priority = p->rt_priority;
3203 attr->sched_runtime = dl_se->dl_runtime;
3204 attr->sched_deadline = dl_se->dl_deadline;
3205 attr->sched_period = dl_se->dl_period;
3206 attr->sched_flags = dl_se->flags;
3210 * This function validates the new parameters of a -deadline task.
3211 * We ask for the deadline not being zero, and greater or equal
3212 * than the runtime, as well as the period of being zero or
3213 * greater than deadline. Furthermore, we have to be sure that
3214 * user parameters are above the internal resolution (1us); we
3215 * check sched_runtime only since it is always the smaller one.
3218 __checkparam_dl(const struct sched_attr *attr)
3220 return attr && attr->sched_deadline != 0 &&
3221 (attr->sched_period == 0 ||
3222 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3223 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3224 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3228 * check the target process has a UID that matches the current process's
3230 static bool check_same_owner(struct task_struct *p)
3232 const struct cred *cred = current_cred(), *pcred;
3236 pcred = __task_cred(p);
3237 match = (uid_eq(cred->euid, pcred->euid) ||
3238 uid_eq(cred->euid, pcred->uid));
3243 static int __sched_setscheduler(struct task_struct *p,
3244 const struct sched_attr *attr,
3247 int retval, oldprio, oldpolicy = -1, on_rq, running;
3248 int policy = attr->sched_policy;
3249 unsigned long flags;
3250 const struct sched_class *prev_class;
3254 /* may grab non-irq protected spin_locks */
3255 BUG_ON(in_interrupt());
3257 /* double check policy once rq lock held */
3259 reset_on_fork = p->sched_reset_on_fork;
3260 policy = oldpolicy = p->policy;
3262 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
3263 policy &= ~SCHED_RESET_ON_FORK;
3265 if (policy != SCHED_DEADLINE &&
3266 policy != SCHED_FIFO && policy != SCHED_RR &&
3267 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3268 policy != SCHED_IDLE)
3273 * Valid priorities for SCHED_FIFO and SCHED_RR are
3274 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3275 * SCHED_BATCH and SCHED_IDLE is 0.
3277 if (attr->sched_priority < 0 ||
3278 (p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3279 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3281 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3282 (rt_policy(policy) != (attr->sched_priority != 0)))
3286 * Allow unprivileged RT tasks to decrease priority:
3288 if (user && !capable(CAP_SYS_NICE)) {
3289 if (fair_policy(policy)) {
3290 if (!can_nice(p, attr->sched_nice))
3294 if (rt_policy(policy)) {
3295 unsigned long rlim_rtprio =
3296 task_rlimit(p, RLIMIT_RTPRIO);
3298 /* can't set/change the rt policy */
3299 if (policy != p->policy && !rlim_rtprio)
3302 /* can't increase priority */
3303 if (attr->sched_priority > p->rt_priority &&
3304 attr->sched_priority > rlim_rtprio)
3309 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3310 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3312 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3313 if (!can_nice(p, TASK_NICE(p)))
3317 /* can't change other user's priorities */
3318 if (!check_same_owner(p))
3321 /* Normal users shall not reset the sched_reset_on_fork flag */
3322 if (p->sched_reset_on_fork && !reset_on_fork)
3327 retval = security_task_setscheduler(p);
3333 * make sure no PI-waiters arrive (or leave) while we are
3334 * changing the priority of the task:
3336 * To be able to change p->policy safely, the appropriate
3337 * runqueue lock must be held.
3339 rq = task_rq_lock(p, &flags);
3342 * Changing the policy of the stop threads its a very bad idea
3344 if (p == rq->stop) {
3345 task_rq_unlock(rq, p, &flags);
3350 * If not changing anything there's no need to proceed further:
3352 if (unlikely(policy == p->policy)) {
3353 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3355 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3357 if (dl_policy(policy))
3360 task_rq_unlock(rq, p, &flags);
3366 #ifdef CONFIG_RT_GROUP_SCHED
3368 * Do not allow realtime tasks into groups that have no runtime
3371 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3372 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3373 !task_group_is_autogroup(task_group(p))) {
3374 task_rq_unlock(rq, p, &flags);
3379 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3380 cpumask_t *span = rq->rd->span;
3383 * Don't allow tasks with an affinity mask smaller than
3384 * the entire root_domain to become SCHED_DEADLINE. We
3385 * will also fail if there's no bandwidth available.
3387 if (!cpumask_subset(span, &p->cpus_allowed) ||
3388 rq->rd->dl_bw.bw == 0) {
3389 task_rq_unlock(rq, p, &flags);
3396 /* recheck policy now with rq lock held */
3397 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3398 policy = oldpolicy = -1;
3399 task_rq_unlock(rq, p, &flags);
3404 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3405 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3408 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3409 task_rq_unlock(rq, p, &flags);
3414 running = task_current(rq, p);
3416 dequeue_task(rq, p, 0);
3418 p->sched_class->put_prev_task(rq, p);
3420 p->sched_reset_on_fork = reset_on_fork;
3423 prev_class = p->sched_class;
3424 __setscheduler(rq, p, attr);
3427 p->sched_class->set_curr_task(rq);
3429 enqueue_task(rq, p, 0);
3431 check_class_changed(rq, p, prev_class, oldprio);
3432 task_rq_unlock(rq, p, &flags);
3434 rt_mutex_adjust_pi(p);
3440 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3441 * @p: the task in question.
3442 * @policy: new policy.
3443 * @param: structure containing the new RT priority.
3445 * Return: 0 on success. An error code otherwise.
3447 * NOTE that the task may be already dead.
3449 int sched_setscheduler(struct task_struct *p, int policy,
3450 const struct sched_param *param)
3452 struct sched_attr attr = {
3453 .sched_policy = policy,
3454 .sched_priority = param->sched_priority,
3455 .sched_nice = PRIO_TO_NICE(p->static_prio),
3457 return __sched_setscheduler(p, &attr, true);
3459 EXPORT_SYMBOL_GPL(sched_setscheduler);
3461 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3463 return __sched_setscheduler(p, attr, true);
3465 EXPORT_SYMBOL_GPL(sched_setattr);
3468 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3469 * @p: the task in question.
3470 * @policy: new policy.
3471 * @param: structure containing the new RT priority.
3473 * Just like sched_setscheduler, only don't bother checking if the
3474 * current context has permission. For example, this is needed in
3475 * stop_machine(): we create temporary high priority worker threads,
3476 * but our caller might not have that capability.
3478 * Return: 0 on success. An error code otherwise.
3480 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3481 const struct sched_param *param)
3483 struct sched_attr attr = {
3484 .sched_policy = policy,
3485 .sched_priority = param->sched_priority,
3486 .sched_nice = PRIO_TO_NICE(p->static_prio),
3488 return __sched_setscheduler(p, &attr, false);
3492 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3494 struct sched_param lparam;
3495 struct task_struct *p;
3498 if (!param || pid < 0)
3500 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3505 p = find_process_by_pid(pid);
3507 retval = sched_setscheduler(p, policy, &lparam);
3514 * Mimics kernel/events/core.c perf_copy_attr().
3516 static int sched_copy_attr(struct sched_attr __user *uattr,
3517 struct sched_attr *attr)
3522 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3526 * zero the full structure, so that a short copy will be nice.
3528 memset(attr, 0, sizeof(*attr));
3530 ret = get_user(size, &uattr->size);
3534 if (size > PAGE_SIZE) /* silly large */
3537 if (!size) /* abi compat */
3538 size = SCHED_ATTR_SIZE_VER0;
3540 if (size < SCHED_ATTR_SIZE_VER0)
3544 * If we're handed a bigger struct than we know of,
3545 * ensure all the unknown bits are 0 - i.e. new
3546 * user-space does not rely on any kernel feature
3547 * extensions we dont know about yet.
3549 if (size > sizeof(*attr)) {
3550 unsigned char __user *addr;
3551 unsigned char __user *end;
3554 addr = (void __user *)uattr + sizeof(*attr);
3555 end = (void __user *)uattr + size;
3557 for (; addr < end; addr++) {
3558 ret = get_user(val, addr);
3564 size = sizeof(*attr);
3567 ret = copy_from_user(attr, uattr, size);
3572 * XXX: do we want to be lenient like existing syscalls; or do we want
3573 * to be strict and return an error on out-of-bounds values?
3575 attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3581 put_user(sizeof(*attr), &uattr->size);
3587 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3588 * @pid: the pid in question.
3589 * @policy: new policy.
3590 * @param: structure containing the new RT priority.
3592 * Return: 0 on success. An error code otherwise.
3594 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3595 struct sched_param __user *, param)
3597 /* negative values for policy are not valid */
3601 return do_sched_setscheduler(pid, policy, param);
3605 * sys_sched_setparam - set/change the RT priority of a thread
3606 * @pid: the pid in question.
3607 * @param: structure containing the new RT priority.
3609 * Return: 0 on success. An error code otherwise.
3611 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3613 return do_sched_setscheduler(pid, -1, param);
3617 * sys_sched_setattr - same as above, but with extended sched_attr
3618 * @pid: the pid in question.
3619 * @uattr: structure containing the extended parameters.
3621 SYSCALL_DEFINE2(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr)
3623 struct sched_attr attr;
3624 struct task_struct *p;
3627 if (!uattr || pid < 0)
3630 if (sched_copy_attr(uattr, &attr))
3635 p = find_process_by_pid(pid);
3637 retval = sched_setattr(p, &attr);
3644 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3645 * @pid: the pid in question.
3647 * Return: On success, the policy of the thread. Otherwise, a negative error
3650 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3652 struct task_struct *p;
3660 p = find_process_by_pid(pid);
3662 retval = security_task_getscheduler(p);
3665 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3672 * sys_sched_getparam - get the RT priority of a thread
3673 * @pid: the pid in question.
3674 * @param: structure containing the RT priority.
3676 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3679 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3681 struct sched_param lp;
3682 struct task_struct *p;
3685 if (!param || pid < 0)
3689 p = find_process_by_pid(pid);
3694 retval = security_task_getscheduler(p);
3698 if (task_has_dl_policy(p)) {
3702 lp.sched_priority = p->rt_priority;
3706 * This one might sleep, we cannot do it with a spinlock held ...
3708 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3717 static int sched_read_attr(struct sched_attr __user *uattr,
3718 struct sched_attr *attr,
3723 if (!access_ok(VERIFY_WRITE, uattr, usize))
3727 * If we're handed a smaller struct than we know of,
3728 * ensure all the unknown bits are 0 - i.e. old
3729 * user-space does not get uncomplete information.
3731 if (usize < sizeof(*attr)) {
3732 unsigned char *addr;
3735 addr = (void *)attr + usize;
3736 end = (void *)attr + sizeof(*attr);
3738 for (; addr < end; addr++) {
3746 ret = copy_to_user(uattr, attr, usize);
3759 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3760 * @pid: the pid in question.
3761 * @uattr: structure containing the extended parameters.
3762 * @size: sizeof(attr) for fwd/bwd comp.
3764 SYSCALL_DEFINE3(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3767 struct sched_attr attr = {
3768 .size = sizeof(struct sched_attr),
3770 struct task_struct *p;
3773 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3774 size < SCHED_ATTR_SIZE_VER0)
3778 p = find_process_by_pid(pid);
3783 retval = security_task_getscheduler(p);
3787 attr.sched_policy = p->policy;
3788 if (task_has_dl_policy(p))
3789 __getparam_dl(p, &attr);
3790 else if (task_has_rt_policy(p))
3791 attr.sched_priority = p->rt_priority;
3793 attr.sched_nice = TASK_NICE(p);
3797 retval = sched_read_attr(uattr, &attr, size);
3805 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3807 cpumask_var_t cpus_allowed, new_mask;
3808 struct task_struct *p;
3813 p = find_process_by_pid(pid);
3819 /* Prevent p going away */
3823 if (p->flags & PF_NO_SETAFFINITY) {
3827 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3831 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3833 goto out_free_cpus_allowed;
3836 if (!check_same_owner(p)) {
3838 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3845 retval = security_task_setscheduler(p);
3850 cpuset_cpus_allowed(p, cpus_allowed);
3851 cpumask_and(new_mask, in_mask, cpus_allowed);
3854 * Since bandwidth control happens on root_domain basis,
3855 * if admission test is enabled, we only admit -deadline
3856 * tasks allowed to run on all the CPUs in the task's
3860 if (task_has_dl_policy(p)) {
3861 const struct cpumask *span = task_rq(p)->rd->span;
3863 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3870 retval = set_cpus_allowed_ptr(p, new_mask);
3873 cpuset_cpus_allowed(p, cpus_allowed);
3874 if (!cpumask_subset(new_mask, cpus_allowed)) {
3876 * We must have raced with a concurrent cpuset
3877 * update. Just reset the cpus_allowed to the
3878 * cpuset's cpus_allowed
3880 cpumask_copy(new_mask, cpus_allowed);
3885 free_cpumask_var(new_mask);
3886 out_free_cpus_allowed:
3887 free_cpumask_var(cpus_allowed);
3893 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3894 struct cpumask *new_mask)
3896 if (len < cpumask_size())
3897 cpumask_clear(new_mask);
3898 else if (len > cpumask_size())
3899 len = cpumask_size();
3901 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3905 * sys_sched_setaffinity - set the cpu affinity of a process
3906 * @pid: pid of the process
3907 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3908 * @user_mask_ptr: user-space pointer to the new cpu mask
3910 * Return: 0 on success. An error code otherwise.
3912 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3913 unsigned long __user *, user_mask_ptr)
3915 cpumask_var_t new_mask;
3918 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3921 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3923 retval = sched_setaffinity(pid, new_mask);
3924 free_cpumask_var(new_mask);
3928 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3930 struct task_struct *p;
3931 unsigned long flags;
3937 p = find_process_by_pid(pid);
3941 retval = security_task_getscheduler(p);
3945 raw_spin_lock_irqsave(&p->pi_lock, flags);
3946 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
3947 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3956 * sys_sched_getaffinity - get the cpu affinity of a process
3957 * @pid: pid of the process
3958 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3959 * @user_mask_ptr: user-space pointer to hold the current cpu mask
3961 * Return: 0 on success. An error code otherwise.
3963 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
3964 unsigned long __user *, user_mask_ptr)
3969 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
3971 if (len & (sizeof(unsigned long)-1))
3974 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
3977 ret = sched_getaffinity(pid, mask);
3979 size_t retlen = min_t(size_t, len, cpumask_size());
3981 if (copy_to_user(user_mask_ptr, mask, retlen))
3986 free_cpumask_var(mask);
3992 * sys_sched_yield - yield the current processor to other threads.
3994 * This function yields the current CPU to other tasks. If there are no
3995 * other threads running on this CPU then this function will return.
3999 SYSCALL_DEFINE0(sched_yield)
4001 struct rq *rq = this_rq_lock();
4003 schedstat_inc(rq, yld_count);
4004 current->sched_class->yield_task(rq);
4007 * Since we are going to call schedule() anyway, there's
4008 * no need to preempt or enable interrupts:
4010 __release(rq->lock);
4011 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4012 do_raw_spin_unlock(&rq->lock);
4013 sched_preempt_enable_no_resched();
4020 static void __cond_resched(void)
4022 __preempt_count_add(PREEMPT_ACTIVE);
4024 __preempt_count_sub(PREEMPT_ACTIVE);
4027 int __sched _cond_resched(void)
4029 if (should_resched()) {
4035 EXPORT_SYMBOL(_cond_resched);
4038 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4039 * call schedule, and on return reacquire the lock.
4041 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4042 * operations here to prevent schedule() from being called twice (once via
4043 * spin_unlock(), once by hand).
4045 int __cond_resched_lock(spinlock_t *lock)
4047 int resched = should_resched();
4050 lockdep_assert_held(lock);
4052 if (spin_needbreak(lock) || resched) {
4063 EXPORT_SYMBOL(__cond_resched_lock);
4065 int __sched __cond_resched_softirq(void)
4067 BUG_ON(!in_softirq());
4069 if (should_resched()) {
4077 EXPORT_SYMBOL(__cond_resched_softirq);
4080 * yield - yield the current processor to other threads.
4082 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4084 * The scheduler is at all times free to pick the calling task as the most
4085 * eligible task to run, if removing the yield() call from your code breaks
4086 * it, its already broken.
4088 * Typical broken usage is:
4093 * where one assumes that yield() will let 'the other' process run that will
4094 * make event true. If the current task is a SCHED_FIFO task that will never
4095 * happen. Never use yield() as a progress guarantee!!
4097 * If you want to use yield() to wait for something, use wait_event().
4098 * If you want to use yield() to be 'nice' for others, use cond_resched().
4099 * If you still want to use yield(), do not!
4101 void __sched yield(void)
4103 set_current_state(TASK_RUNNING);
4106 EXPORT_SYMBOL(yield);
4109 * yield_to - yield the current processor to another thread in
4110 * your thread group, or accelerate that thread toward the
4111 * processor it's on.
4113 * @preempt: whether task preemption is allowed or not
4115 * It's the caller's job to ensure that the target task struct
4116 * can't go away on us before we can do any checks.
4119 * true (>0) if we indeed boosted the target task.
4120 * false (0) if we failed to boost the target.
4121 * -ESRCH if there's no task to yield to.
4123 bool __sched yield_to(struct task_struct *p, bool preempt)
4125 struct task_struct *curr = current;
4126 struct rq *rq, *p_rq;
4127 unsigned long flags;
4130 local_irq_save(flags);
4136 * If we're the only runnable task on the rq and target rq also
4137 * has only one task, there's absolutely no point in yielding.
4139 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4144 double_rq_lock(rq, p_rq);
4145 if (task_rq(p) != p_rq) {
4146 double_rq_unlock(rq, p_rq);
4150 if (!curr->sched_class->yield_to_task)
4153 if (curr->sched_class != p->sched_class)
4156 if (task_running(p_rq, p) || p->state)
4159 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4161 schedstat_inc(rq, yld_count);
4163 * Make p's CPU reschedule; pick_next_entity takes care of
4166 if (preempt && rq != p_rq)
4167 resched_task(p_rq->curr);
4171 double_rq_unlock(rq, p_rq);
4173 local_irq_restore(flags);
4180 EXPORT_SYMBOL_GPL(yield_to);
4183 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4184 * that process accounting knows that this is a task in IO wait state.
4186 void __sched io_schedule(void)
4188 struct rq *rq = raw_rq();
4190 delayacct_blkio_start();
4191 atomic_inc(&rq->nr_iowait);
4192 blk_flush_plug(current);
4193 current->in_iowait = 1;
4195 current->in_iowait = 0;
4196 atomic_dec(&rq->nr_iowait);
4197 delayacct_blkio_end();
4199 EXPORT_SYMBOL(io_schedule);
4201 long __sched io_schedule_timeout(long timeout)
4203 struct rq *rq = raw_rq();
4206 delayacct_blkio_start();
4207 atomic_inc(&rq->nr_iowait);
4208 blk_flush_plug(current);
4209 current->in_iowait = 1;
4210 ret = schedule_timeout(timeout);
4211 current->in_iowait = 0;
4212 atomic_dec(&rq->nr_iowait);
4213 delayacct_blkio_end();
4218 * sys_sched_get_priority_max - return maximum RT priority.
4219 * @policy: scheduling class.
4221 * Return: On success, this syscall returns the maximum
4222 * rt_priority that can be used by a given scheduling class.
4223 * On failure, a negative error code is returned.
4225 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4232 ret = MAX_USER_RT_PRIO-1;
4234 case SCHED_DEADLINE:
4245 * sys_sched_get_priority_min - return minimum RT priority.
4246 * @policy: scheduling class.
4248 * Return: On success, this syscall returns the minimum
4249 * rt_priority that can be used by a given scheduling class.
4250 * On failure, a negative error code is returned.
4252 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4261 case SCHED_DEADLINE:
4271 * sys_sched_rr_get_interval - return the default timeslice of a process.
4272 * @pid: pid of the process.
4273 * @interval: userspace pointer to the timeslice value.
4275 * this syscall writes the default timeslice value of a given process
4276 * into the user-space timespec buffer. A value of '0' means infinity.
4278 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4281 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4282 struct timespec __user *, interval)
4284 struct task_struct *p;
4285 unsigned int time_slice;
4286 unsigned long flags;
4296 p = find_process_by_pid(pid);
4300 retval = security_task_getscheduler(p);
4304 rq = task_rq_lock(p, &flags);
4305 time_slice = p->sched_class->get_rr_interval(rq, p);
4306 task_rq_unlock(rq, p, &flags);
4309 jiffies_to_timespec(time_slice, &t);
4310 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4318 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4320 void sched_show_task(struct task_struct *p)
4322 unsigned long free = 0;
4326 state = p->state ? __ffs(p->state) + 1 : 0;
4327 printk(KERN_INFO "%-15.15s %c", p->comm,
4328 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4329 #if BITS_PER_LONG == 32
4330 if (state == TASK_RUNNING)
4331 printk(KERN_CONT " running ");
4333 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4335 if (state == TASK_RUNNING)
4336 printk(KERN_CONT " running task ");
4338 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4340 #ifdef CONFIG_DEBUG_STACK_USAGE
4341 free = stack_not_used(p);
4344 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4346 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4347 task_pid_nr(p), ppid,
4348 (unsigned long)task_thread_info(p)->flags);
4350 print_worker_info(KERN_INFO, p);
4351 show_stack(p, NULL);
4354 void show_state_filter(unsigned long state_filter)
4356 struct task_struct *g, *p;
4358 #if BITS_PER_LONG == 32
4360 " task PC stack pid father\n");
4363 " task PC stack pid father\n");
4366 do_each_thread(g, p) {
4368 * reset the NMI-timeout, listing all files on a slow
4369 * console might take a lot of time:
4371 touch_nmi_watchdog();
4372 if (!state_filter || (p->state & state_filter))
4374 } while_each_thread(g, p);
4376 touch_all_softlockup_watchdogs();
4378 #ifdef CONFIG_SCHED_DEBUG
4379 sysrq_sched_debug_show();
4383 * Only show locks if all tasks are dumped:
4386 debug_show_all_locks();
4389 void init_idle_bootup_task(struct task_struct *idle)
4391 idle->sched_class = &idle_sched_class;
4395 * init_idle - set up an idle thread for a given CPU
4396 * @idle: task in question
4397 * @cpu: cpu the idle task belongs to
4399 * NOTE: this function does not set the idle thread's NEED_RESCHED
4400 * flag, to make booting more robust.
4402 void init_idle(struct task_struct *idle, int cpu)
4404 struct rq *rq = cpu_rq(cpu);
4405 unsigned long flags;
4407 raw_spin_lock_irqsave(&rq->lock, flags);
4409 __sched_fork(0, idle);
4410 idle->state = TASK_RUNNING;
4411 idle->se.exec_start = sched_clock();
4413 do_set_cpus_allowed(idle, cpumask_of(cpu));
4415 * We're having a chicken and egg problem, even though we are
4416 * holding rq->lock, the cpu isn't yet set to this cpu so the
4417 * lockdep check in task_group() will fail.
4419 * Similar case to sched_fork(). / Alternatively we could
4420 * use task_rq_lock() here and obtain the other rq->lock.
4425 __set_task_cpu(idle, cpu);
4428 rq->curr = rq->idle = idle;
4429 #if defined(CONFIG_SMP)
4432 raw_spin_unlock_irqrestore(&rq->lock, flags);
4434 /* Set the preempt count _outside_ the spinlocks! */
4435 init_idle_preempt_count(idle, cpu);
4438 * The idle tasks have their own, simple scheduling class:
4440 idle->sched_class = &idle_sched_class;
4441 ftrace_graph_init_idle_task(idle, cpu);
4442 vtime_init_idle(idle, cpu);
4443 #if defined(CONFIG_SMP)
4444 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4449 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4451 if (p->sched_class && p->sched_class->set_cpus_allowed)
4452 p->sched_class->set_cpus_allowed(p, new_mask);
4454 cpumask_copy(&p->cpus_allowed, new_mask);
4455 p->nr_cpus_allowed = cpumask_weight(new_mask);
4459 * This is how migration works:
4461 * 1) we invoke migration_cpu_stop() on the target CPU using
4463 * 2) stopper starts to run (implicitly forcing the migrated thread
4465 * 3) it checks whether the migrated task is still in the wrong runqueue.
4466 * 4) if it's in the wrong runqueue then the migration thread removes
4467 * it and puts it into the right queue.
4468 * 5) stopper completes and stop_one_cpu() returns and the migration
4473 * Change a given task's CPU affinity. Migrate the thread to a
4474 * proper CPU and schedule it away if the CPU it's executing on
4475 * is removed from the allowed bitmask.
4477 * NOTE: the caller must have a valid reference to the task, the
4478 * task must not exit() & deallocate itself prematurely. The
4479 * call is not atomic; no spinlocks may be held.
4481 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4483 unsigned long flags;
4485 unsigned int dest_cpu;
4488 rq = task_rq_lock(p, &flags);
4490 if (cpumask_equal(&p->cpus_allowed, new_mask))
4493 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4498 do_set_cpus_allowed(p, new_mask);
4500 /* Can the task run on the task's current CPU? If so, we're done */
4501 if (cpumask_test_cpu(task_cpu(p), new_mask))
4504 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4506 struct migration_arg arg = { p, dest_cpu };
4507 /* Need help from migration thread: drop lock and wait. */
4508 task_rq_unlock(rq, p, &flags);
4509 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4510 tlb_migrate_finish(p->mm);
4514 task_rq_unlock(rq, p, &flags);
4518 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4521 * Move (not current) task off this cpu, onto dest cpu. We're doing
4522 * this because either it can't run here any more (set_cpus_allowed()
4523 * away from this CPU, or CPU going down), or because we're
4524 * attempting to rebalance this task on exec (sched_exec).
4526 * So we race with normal scheduler movements, but that's OK, as long
4527 * as the task is no longer on this CPU.
4529 * Returns non-zero if task was successfully migrated.
4531 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4533 struct rq *rq_dest, *rq_src;
4536 if (unlikely(!cpu_active(dest_cpu)))
4539 rq_src = cpu_rq(src_cpu);
4540 rq_dest = cpu_rq(dest_cpu);
4542 raw_spin_lock(&p->pi_lock);
4543 double_rq_lock(rq_src, rq_dest);
4544 /* Already moved. */
4545 if (task_cpu(p) != src_cpu)
4547 /* Affinity changed (again). */
4548 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4552 * If we're not on a rq, the next wake-up will ensure we're
4556 dequeue_task(rq_src, p, 0);
4557 set_task_cpu(p, dest_cpu);
4558 enqueue_task(rq_dest, p, 0);
4559 check_preempt_curr(rq_dest, p, 0);
4564 double_rq_unlock(rq_src, rq_dest);
4565 raw_spin_unlock(&p->pi_lock);
4569 #ifdef CONFIG_NUMA_BALANCING
4570 /* Migrate current task p to target_cpu */
4571 int migrate_task_to(struct task_struct *p, int target_cpu)
4573 struct migration_arg arg = { p, target_cpu };
4574 int curr_cpu = task_cpu(p);
4576 if (curr_cpu == target_cpu)
4579 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4582 /* TODO: This is not properly updating schedstats */
4584 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4588 * Requeue a task on a given node and accurately track the number of NUMA
4589 * tasks on the runqueues
4591 void sched_setnuma(struct task_struct *p, int nid)
4594 unsigned long flags;
4595 bool on_rq, running;
4597 rq = task_rq_lock(p, &flags);
4599 running = task_current(rq, p);
4602 dequeue_task(rq, p, 0);
4604 p->sched_class->put_prev_task(rq, p);
4606 p->numa_preferred_nid = nid;
4609 p->sched_class->set_curr_task(rq);
4611 enqueue_task(rq, p, 0);
4612 task_rq_unlock(rq, p, &flags);
4617 * migration_cpu_stop - this will be executed by a highprio stopper thread
4618 * and performs thread migration by bumping thread off CPU then
4619 * 'pushing' onto another runqueue.
4621 static int migration_cpu_stop(void *data)
4623 struct migration_arg *arg = data;
4626 * The original target cpu might have gone down and we might
4627 * be on another cpu but it doesn't matter.
4629 local_irq_disable();
4630 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4635 #ifdef CONFIG_HOTPLUG_CPU
4638 * Ensures that the idle task is using init_mm right before its cpu goes
4641 void idle_task_exit(void)
4643 struct mm_struct *mm = current->active_mm;
4645 BUG_ON(cpu_online(smp_processor_id()));
4648 switch_mm(mm, &init_mm, current);
4653 * Since this CPU is going 'away' for a while, fold any nr_active delta
4654 * we might have. Assumes we're called after migrate_tasks() so that the
4655 * nr_active count is stable.
4657 * Also see the comment "Global load-average calculations".
4659 static void calc_load_migrate(struct rq *rq)
4661 long delta = calc_load_fold_active(rq);
4663 atomic_long_add(delta, &calc_load_tasks);
4667 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4668 * try_to_wake_up()->select_task_rq().
4670 * Called with rq->lock held even though we'er in stop_machine() and
4671 * there's no concurrency possible, we hold the required locks anyway
4672 * because of lock validation efforts.
4674 static void migrate_tasks(unsigned int dead_cpu)
4676 struct rq *rq = cpu_rq(dead_cpu);
4677 struct task_struct *next, *stop = rq->stop;
4681 * Fudge the rq selection such that the below task selection loop
4682 * doesn't get stuck on the currently eligible stop task.
4684 * We're currently inside stop_machine() and the rq is either stuck
4685 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4686 * either way we should never end up calling schedule() until we're
4692 * put_prev_task() and pick_next_task() sched
4693 * class method both need to have an up-to-date
4694 * value of rq->clock[_task]
4696 update_rq_clock(rq);
4700 * There's this thread running, bail when that's the only
4703 if (rq->nr_running == 1)
4706 next = pick_next_task(rq);
4708 next->sched_class->put_prev_task(rq, next);
4710 /* Find suitable destination for @next, with force if needed. */
4711 dest_cpu = select_fallback_rq(dead_cpu, next);
4712 raw_spin_unlock(&rq->lock);
4714 __migrate_task(next, dead_cpu, dest_cpu);
4716 raw_spin_lock(&rq->lock);
4722 #endif /* CONFIG_HOTPLUG_CPU */
4724 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4726 static struct ctl_table sd_ctl_dir[] = {
4728 .procname = "sched_domain",
4734 static struct ctl_table sd_ctl_root[] = {
4736 .procname = "kernel",
4738 .child = sd_ctl_dir,
4743 static struct ctl_table *sd_alloc_ctl_entry(int n)
4745 struct ctl_table *entry =
4746 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4751 static void sd_free_ctl_entry(struct ctl_table **tablep)
4753 struct ctl_table *entry;
4756 * In the intermediate directories, both the child directory and
4757 * procname are dynamically allocated and could fail but the mode
4758 * will always be set. In the lowest directory the names are
4759 * static strings and all have proc handlers.
4761 for (entry = *tablep; entry->mode; entry++) {
4763 sd_free_ctl_entry(&entry->child);
4764 if (entry->proc_handler == NULL)
4765 kfree(entry->procname);
4772 static int min_load_idx = 0;
4773 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4776 set_table_entry(struct ctl_table *entry,
4777 const char *procname, void *data, int maxlen,
4778 umode_t mode, proc_handler *proc_handler,
4781 entry->procname = procname;
4783 entry->maxlen = maxlen;
4785 entry->proc_handler = proc_handler;
4788 entry->extra1 = &min_load_idx;
4789 entry->extra2 = &max_load_idx;
4793 static struct ctl_table *
4794 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4796 struct ctl_table *table = sd_alloc_ctl_entry(13);
4801 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4802 sizeof(long), 0644, proc_doulongvec_minmax, false);
4803 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4804 sizeof(long), 0644, proc_doulongvec_minmax, false);
4805 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4806 sizeof(int), 0644, proc_dointvec_minmax, true);
4807 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4808 sizeof(int), 0644, proc_dointvec_minmax, true);
4809 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4810 sizeof(int), 0644, proc_dointvec_minmax, true);
4811 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4812 sizeof(int), 0644, proc_dointvec_minmax, true);
4813 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4814 sizeof(int), 0644, proc_dointvec_minmax, true);
4815 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4816 sizeof(int), 0644, proc_dointvec_minmax, false);
4817 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4818 sizeof(int), 0644, proc_dointvec_minmax, false);
4819 set_table_entry(&table[9], "cache_nice_tries",
4820 &sd->cache_nice_tries,
4821 sizeof(int), 0644, proc_dointvec_minmax, false);
4822 set_table_entry(&table[10], "flags", &sd->flags,
4823 sizeof(int), 0644, proc_dointvec_minmax, false);
4824 set_table_entry(&table[11], "name", sd->name,
4825 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4826 /* &table[12] is terminator */
4831 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4833 struct ctl_table *entry, *table;
4834 struct sched_domain *sd;
4835 int domain_num = 0, i;
4838 for_each_domain(cpu, sd)
4840 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4845 for_each_domain(cpu, sd) {
4846 snprintf(buf, 32, "domain%d", i);
4847 entry->procname = kstrdup(buf, GFP_KERNEL);
4849 entry->child = sd_alloc_ctl_domain_table(sd);
4856 static struct ctl_table_header *sd_sysctl_header;
4857 static void register_sched_domain_sysctl(void)
4859 int i, cpu_num = num_possible_cpus();
4860 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4863 WARN_ON(sd_ctl_dir[0].child);
4864 sd_ctl_dir[0].child = entry;
4869 for_each_possible_cpu(i) {
4870 snprintf(buf, 32, "cpu%d", i);
4871 entry->procname = kstrdup(buf, GFP_KERNEL);
4873 entry->child = sd_alloc_ctl_cpu_table(i);
4877 WARN_ON(sd_sysctl_header);
4878 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4881 /* may be called multiple times per register */
4882 static void unregister_sched_domain_sysctl(void)
4884 if (sd_sysctl_header)
4885 unregister_sysctl_table(sd_sysctl_header);
4886 sd_sysctl_header = NULL;
4887 if (sd_ctl_dir[0].child)
4888 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4891 static void register_sched_domain_sysctl(void)
4894 static void unregister_sched_domain_sysctl(void)
4899 static void set_rq_online(struct rq *rq)
4902 const struct sched_class *class;
4904 cpumask_set_cpu(rq->cpu, rq->rd->online);
4907 for_each_class(class) {
4908 if (class->rq_online)
4909 class->rq_online(rq);
4914 static void set_rq_offline(struct rq *rq)
4917 const struct sched_class *class;
4919 for_each_class(class) {
4920 if (class->rq_offline)
4921 class->rq_offline(rq);
4924 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4930 * migration_call - callback that gets triggered when a CPU is added.
4931 * Here we can start up the necessary migration thread for the new CPU.
4934 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4936 int cpu = (long)hcpu;
4937 unsigned long flags;
4938 struct rq *rq = cpu_rq(cpu);
4940 switch (action & ~CPU_TASKS_FROZEN) {
4942 case CPU_UP_PREPARE:
4943 rq->calc_load_update = calc_load_update;
4947 /* Update our root-domain */
4948 raw_spin_lock_irqsave(&rq->lock, flags);
4950 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4954 raw_spin_unlock_irqrestore(&rq->lock, flags);
4957 #ifdef CONFIG_HOTPLUG_CPU
4959 sched_ttwu_pending();
4960 /* Update our root-domain */
4961 raw_spin_lock_irqsave(&rq->lock, flags);
4963 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
4967 BUG_ON(rq->nr_running != 1); /* the migration thread */
4968 raw_spin_unlock_irqrestore(&rq->lock, flags);
4972 calc_load_migrate(rq);
4977 update_max_interval();
4983 * Register at high priority so that task migration (migrate_all_tasks)
4984 * happens before everything else. This has to be lower priority than
4985 * the notifier in the perf_event subsystem, though.
4987 static struct notifier_block migration_notifier = {
4988 .notifier_call = migration_call,
4989 .priority = CPU_PRI_MIGRATION,
4992 static int sched_cpu_active(struct notifier_block *nfb,
4993 unsigned long action, void *hcpu)
4995 switch (action & ~CPU_TASKS_FROZEN) {
4997 case CPU_DOWN_FAILED:
4998 set_cpu_active((long)hcpu, true);
5005 static int sched_cpu_inactive(struct notifier_block *nfb,
5006 unsigned long action, void *hcpu)
5008 unsigned long flags;
5009 long cpu = (long)hcpu;
5011 switch (action & ~CPU_TASKS_FROZEN) {
5012 case CPU_DOWN_PREPARE:
5013 set_cpu_active(cpu, false);
5015 /* explicitly allow suspend */
5016 if (!(action & CPU_TASKS_FROZEN)) {
5017 struct dl_bw *dl_b = dl_bw_of(cpu);
5021 raw_spin_lock_irqsave(&dl_b->lock, flags);
5022 cpus = dl_bw_cpus(cpu);
5023 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5024 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5027 return notifier_from_errno(-EBUSY);
5035 static int __init migration_init(void)
5037 void *cpu = (void *)(long)smp_processor_id();
5040 /* Initialize migration for the boot CPU */
5041 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5042 BUG_ON(err == NOTIFY_BAD);
5043 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5044 register_cpu_notifier(&migration_notifier);
5046 /* Register cpu active notifiers */
5047 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5048 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5052 early_initcall(migration_init);
5057 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5059 #ifdef CONFIG_SCHED_DEBUG
5061 static __read_mostly int sched_debug_enabled;
5063 static int __init sched_debug_setup(char *str)
5065 sched_debug_enabled = 1;
5069 early_param("sched_debug", sched_debug_setup);
5071 static inline bool sched_debug(void)
5073 return sched_debug_enabled;
5076 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5077 struct cpumask *groupmask)
5079 struct sched_group *group = sd->groups;
5082 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5083 cpumask_clear(groupmask);
5085 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5087 if (!(sd->flags & SD_LOAD_BALANCE)) {
5088 printk("does not load-balance\n");
5090 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5095 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5097 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5098 printk(KERN_ERR "ERROR: domain->span does not contain "
5101 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5102 printk(KERN_ERR "ERROR: domain->groups does not contain"
5106 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5110 printk(KERN_ERR "ERROR: group is NULL\n");
5115 * Even though we initialize ->power to something semi-sane,
5116 * we leave power_orig unset. This allows us to detect if
5117 * domain iteration is still funny without causing /0 traps.
5119 if (!group->sgp->power_orig) {
5120 printk(KERN_CONT "\n");
5121 printk(KERN_ERR "ERROR: domain->cpu_power not "
5126 if (!cpumask_weight(sched_group_cpus(group))) {
5127 printk(KERN_CONT "\n");
5128 printk(KERN_ERR "ERROR: empty group\n");
5132 if (!(sd->flags & SD_OVERLAP) &&
5133 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5134 printk(KERN_CONT "\n");
5135 printk(KERN_ERR "ERROR: repeated CPUs\n");
5139 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5141 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5143 printk(KERN_CONT " %s", str);
5144 if (group->sgp->power != SCHED_POWER_SCALE) {
5145 printk(KERN_CONT " (cpu_power = %d)",
5149 group = group->next;
5150 } while (group != sd->groups);
5151 printk(KERN_CONT "\n");
5153 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5154 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5157 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5158 printk(KERN_ERR "ERROR: parent span is not a superset "
5159 "of domain->span\n");
5163 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5167 if (!sched_debug_enabled)
5171 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5175 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5178 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5186 #else /* !CONFIG_SCHED_DEBUG */
5187 # define sched_domain_debug(sd, cpu) do { } while (0)
5188 static inline bool sched_debug(void)
5192 #endif /* CONFIG_SCHED_DEBUG */
5194 static int sd_degenerate(struct sched_domain *sd)
5196 if (cpumask_weight(sched_domain_span(sd)) == 1)
5199 /* Following flags need at least 2 groups */
5200 if (sd->flags & (SD_LOAD_BALANCE |
5201 SD_BALANCE_NEWIDLE |
5205 SD_SHARE_PKG_RESOURCES)) {
5206 if (sd->groups != sd->groups->next)
5210 /* Following flags don't use groups */
5211 if (sd->flags & (SD_WAKE_AFFINE))
5218 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5220 unsigned long cflags = sd->flags, pflags = parent->flags;
5222 if (sd_degenerate(parent))
5225 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5228 /* Flags needing groups don't count if only 1 group in parent */
5229 if (parent->groups == parent->groups->next) {
5230 pflags &= ~(SD_LOAD_BALANCE |
5231 SD_BALANCE_NEWIDLE |
5235 SD_SHARE_PKG_RESOURCES |
5237 if (nr_node_ids == 1)
5238 pflags &= ~SD_SERIALIZE;
5240 if (~cflags & pflags)
5246 static void free_rootdomain(struct rcu_head *rcu)
5248 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5250 cpupri_cleanup(&rd->cpupri);
5251 cpudl_cleanup(&rd->cpudl);
5252 free_cpumask_var(rd->dlo_mask);
5253 free_cpumask_var(rd->rto_mask);
5254 free_cpumask_var(rd->online);
5255 free_cpumask_var(rd->span);
5259 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5261 struct root_domain *old_rd = NULL;
5262 unsigned long flags;
5264 raw_spin_lock_irqsave(&rq->lock, flags);
5269 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5272 cpumask_clear_cpu(rq->cpu, old_rd->span);
5275 * If we dont want to free the old_rd yet then
5276 * set old_rd to NULL to skip the freeing later
5279 if (!atomic_dec_and_test(&old_rd->refcount))
5283 atomic_inc(&rd->refcount);
5286 cpumask_set_cpu(rq->cpu, rd->span);
5287 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5290 raw_spin_unlock_irqrestore(&rq->lock, flags);
5293 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5296 static int init_rootdomain(struct root_domain *rd)
5298 memset(rd, 0, sizeof(*rd));
5300 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5302 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5304 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5306 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5309 init_dl_bw(&rd->dl_bw);
5310 if (cpudl_init(&rd->cpudl) != 0)
5313 if (cpupri_init(&rd->cpupri) != 0)
5318 free_cpumask_var(rd->rto_mask);
5320 free_cpumask_var(rd->dlo_mask);
5322 free_cpumask_var(rd->online);
5324 free_cpumask_var(rd->span);
5330 * By default the system creates a single root-domain with all cpus as
5331 * members (mimicking the global state we have today).
5333 struct root_domain def_root_domain;
5335 static void init_defrootdomain(void)
5337 init_rootdomain(&def_root_domain);
5339 atomic_set(&def_root_domain.refcount, 1);
5342 static struct root_domain *alloc_rootdomain(void)
5344 struct root_domain *rd;
5346 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5350 if (init_rootdomain(rd) != 0) {
5358 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5360 struct sched_group *tmp, *first;
5369 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5374 } while (sg != first);
5377 static void free_sched_domain(struct rcu_head *rcu)
5379 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5382 * If its an overlapping domain it has private groups, iterate and
5385 if (sd->flags & SD_OVERLAP) {
5386 free_sched_groups(sd->groups, 1);
5387 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5388 kfree(sd->groups->sgp);
5394 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5396 call_rcu(&sd->rcu, free_sched_domain);
5399 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5401 for (; sd; sd = sd->parent)
5402 destroy_sched_domain(sd, cpu);
5406 * Keep a special pointer to the highest sched_domain that has
5407 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5408 * allows us to avoid some pointer chasing select_idle_sibling().
5410 * Also keep a unique ID per domain (we use the first cpu number in
5411 * the cpumask of the domain), this allows us to quickly tell if
5412 * two cpus are in the same cache domain, see cpus_share_cache().
5414 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5415 DEFINE_PER_CPU(int, sd_llc_size);
5416 DEFINE_PER_CPU(int, sd_llc_id);
5417 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5418 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5419 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5421 static void update_top_cache_domain(int cpu)
5423 struct sched_domain *sd;
5424 struct sched_domain *busy_sd = NULL;
5428 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5430 id = cpumask_first(sched_domain_span(sd));
5431 size = cpumask_weight(sched_domain_span(sd));
5432 busy_sd = sd->parent; /* sd_busy */
5434 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5436 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5437 per_cpu(sd_llc_size, cpu) = size;
5438 per_cpu(sd_llc_id, cpu) = id;
5440 sd = lowest_flag_domain(cpu, SD_NUMA);
5441 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5443 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5444 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5448 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5449 * hold the hotplug lock.
5452 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5454 struct rq *rq = cpu_rq(cpu);
5455 struct sched_domain *tmp;
5457 /* Remove the sched domains which do not contribute to scheduling. */
5458 for (tmp = sd; tmp; ) {
5459 struct sched_domain *parent = tmp->parent;
5463 if (sd_parent_degenerate(tmp, parent)) {
5464 tmp->parent = parent->parent;
5466 parent->parent->child = tmp;
5468 * Transfer SD_PREFER_SIBLING down in case of a
5469 * degenerate parent; the spans match for this
5470 * so the property transfers.
5472 if (parent->flags & SD_PREFER_SIBLING)
5473 tmp->flags |= SD_PREFER_SIBLING;
5474 destroy_sched_domain(parent, cpu);
5479 if (sd && sd_degenerate(sd)) {
5482 destroy_sched_domain(tmp, cpu);
5487 sched_domain_debug(sd, cpu);
5489 rq_attach_root(rq, rd);
5491 rcu_assign_pointer(rq->sd, sd);
5492 destroy_sched_domains(tmp, cpu);
5494 update_top_cache_domain(cpu);
5497 /* cpus with isolated domains */
5498 static cpumask_var_t cpu_isolated_map;
5500 /* Setup the mask of cpus configured for isolated domains */
5501 static int __init isolated_cpu_setup(char *str)
5503 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5504 cpulist_parse(str, cpu_isolated_map);
5508 __setup("isolcpus=", isolated_cpu_setup);
5510 static const struct cpumask *cpu_cpu_mask(int cpu)
5512 return cpumask_of_node(cpu_to_node(cpu));
5516 struct sched_domain **__percpu sd;
5517 struct sched_group **__percpu sg;
5518 struct sched_group_power **__percpu sgp;
5522 struct sched_domain ** __percpu sd;
5523 struct root_domain *rd;
5533 struct sched_domain_topology_level;
5535 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5536 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5538 #define SDTL_OVERLAP 0x01
5540 struct sched_domain_topology_level {
5541 sched_domain_init_f init;
5542 sched_domain_mask_f mask;
5545 struct sd_data data;
5549 * Build an iteration mask that can exclude certain CPUs from the upwards
5552 * Asymmetric node setups can result in situations where the domain tree is of
5553 * unequal depth, make sure to skip domains that already cover the entire
5556 * In that case build_sched_domains() will have terminated the iteration early
5557 * and our sibling sd spans will be empty. Domains should always include the
5558 * cpu they're built on, so check that.
5561 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5563 const struct cpumask *span = sched_domain_span(sd);
5564 struct sd_data *sdd = sd->private;
5565 struct sched_domain *sibling;
5568 for_each_cpu(i, span) {
5569 sibling = *per_cpu_ptr(sdd->sd, i);
5570 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5573 cpumask_set_cpu(i, sched_group_mask(sg));
5578 * Return the canonical balance cpu for this group, this is the first cpu
5579 * of this group that's also in the iteration mask.
5581 int group_balance_cpu(struct sched_group *sg)
5583 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5587 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5589 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5590 const struct cpumask *span = sched_domain_span(sd);
5591 struct cpumask *covered = sched_domains_tmpmask;
5592 struct sd_data *sdd = sd->private;
5593 struct sched_domain *child;
5596 cpumask_clear(covered);
5598 for_each_cpu(i, span) {
5599 struct cpumask *sg_span;
5601 if (cpumask_test_cpu(i, covered))
5604 child = *per_cpu_ptr(sdd->sd, i);
5606 /* See the comment near build_group_mask(). */
5607 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5610 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5611 GFP_KERNEL, cpu_to_node(cpu));
5616 sg_span = sched_group_cpus(sg);
5618 child = child->child;
5619 cpumask_copy(sg_span, sched_domain_span(child));
5621 cpumask_set_cpu(i, sg_span);
5623 cpumask_or(covered, covered, sg_span);
5625 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5626 if (atomic_inc_return(&sg->sgp->ref) == 1)
5627 build_group_mask(sd, sg);
5630 * Initialize sgp->power such that even if we mess up the
5631 * domains and no possible iteration will get us here, we won't
5634 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5635 sg->sgp->power_orig = sg->sgp->power;
5638 * Make sure the first group of this domain contains the
5639 * canonical balance cpu. Otherwise the sched_domain iteration
5640 * breaks. See update_sg_lb_stats().
5642 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5643 group_balance_cpu(sg) == cpu)
5653 sd->groups = groups;
5658 free_sched_groups(first, 0);
5663 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5665 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5666 struct sched_domain *child = sd->child;
5669 cpu = cpumask_first(sched_domain_span(child));
5672 *sg = *per_cpu_ptr(sdd->sg, cpu);
5673 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5674 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5681 * build_sched_groups will build a circular linked list of the groups
5682 * covered by the given span, and will set each group's ->cpumask correctly,
5683 * and ->cpu_power to 0.
5685 * Assumes the sched_domain tree is fully constructed
5688 build_sched_groups(struct sched_domain *sd, int cpu)
5690 struct sched_group *first = NULL, *last = NULL;
5691 struct sd_data *sdd = sd->private;
5692 const struct cpumask *span = sched_domain_span(sd);
5693 struct cpumask *covered;
5696 get_group(cpu, sdd, &sd->groups);
5697 atomic_inc(&sd->groups->ref);
5699 if (cpu != cpumask_first(span))
5702 lockdep_assert_held(&sched_domains_mutex);
5703 covered = sched_domains_tmpmask;
5705 cpumask_clear(covered);
5707 for_each_cpu(i, span) {
5708 struct sched_group *sg;
5711 if (cpumask_test_cpu(i, covered))
5714 group = get_group(i, sdd, &sg);
5715 cpumask_clear(sched_group_cpus(sg));
5717 cpumask_setall(sched_group_mask(sg));
5719 for_each_cpu(j, span) {
5720 if (get_group(j, sdd, NULL) != group)
5723 cpumask_set_cpu(j, covered);
5724 cpumask_set_cpu(j, sched_group_cpus(sg));
5739 * Initialize sched groups cpu_power.
5741 * cpu_power indicates the capacity of sched group, which is used while
5742 * distributing the load between different sched groups in a sched domain.
5743 * Typically cpu_power for all the groups in a sched domain will be same unless
5744 * there are asymmetries in the topology. If there are asymmetries, group
5745 * having more cpu_power will pickup more load compared to the group having
5748 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5750 struct sched_group *sg = sd->groups;
5755 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5757 } while (sg != sd->groups);
5759 if (cpu != group_balance_cpu(sg))
5762 update_group_power(sd, cpu);
5763 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5766 int __weak arch_sd_sibling_asym_packing(void)
5768 return 0*SD_ASYM_PACKING;
5772 * Initializers for schedule domains
5773 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5776 #ifdef CONFIG_SCHED_DEBUG
5777 # define SD_INIT_NAME(sd, type) sd->name = #type
5779 # define SD_INIT_NAME(sd, type) do { } while (0)
5782 #define SD_INIT_FUNC(type) \
5783 static noinline struct sched_domain * \
5784 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5786 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5787 *sd = SD_##type##_INIT; \
5788 SD_INIT_NAME(sd, type); \
5789 sd->private = &tl->data; \
5794 #ifdef CONFIG_SCHED_SMT
5795 SD_INIT_FUNC(SIBLING)
5797 #ifdef CONFIG_SCHED_MC
5800 #ifdef CONFIG_SCHED_BOOK
5804 static int default_relax_domain_level = -1;
5805 int sched_domain_level_max;
5807 static int __init setup_relax_domain_level(char *str)
5809 if (kstrtoint(str, 0, &default_relax_domain_level))
5810 pr_warn("Unable to set relax_domain_level\n");
5814 __setup("relax_domain_level=", setup_relax_domain_level);
5816 static void set_domain_attribute(struct sched_domain *sd,
5817 struct sched_domain_attr *attr)
5821 if (!attr || attr->relax_domain_level < 0) {
5822 if (default_relax_domain_level < 0)
5825 request = default_relax_domain_level;
5827 request = attr->relax_domain_level;
5828 if (request < sd->level) {
5829 /* turn off idle balance on this domain */
5830 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5832 /* turn on idle balance on this domain */
5833 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5837 static void __sdt_free(const struct cpumask *cpu_map);
5838 static int __sdt_alloc(const struct cpumask *cpu_map);
5840 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5841 const struct cpumask *cpu_map)
5845 if (!atomic_read(&d->rd->refcount))
5846 free_rootdomain(&d->rd->rcu); /* fall through */
5848 free_percpu(d->sd); /* fall through */
5850 __sdt_free(cpu_map); /* fall through */
5856 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5857 const struct cpumask *cpu_map)
5859 memset(d, 0, sizeof(*d));
5861 if (__sdt_alloc(cpu_map))
5862 return sa_sd_storage;
5863 d->sd = alloc_percpu(struct sched_domain *);
5865 return sa_sd_storage;
5866 d->rd = alloc_rootdomain();
5869 return sa_rootdomain;
5873 * NULL the sd_data elements we've used to build the sched_domain and
5874 * sched_group structure so that the subsequent __free_domain_allocs()
5875 * will not free the data we're using.
5877 static void claim_allocations(int cpu, struct sched_domain *sd)
5879 struct sd_data *sdd = sd->private;
5881 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5882 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5884 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5885 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5887 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5888 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5891 #ifdef CONFIG_SCHED_SMT
5892 static const struct cpumask *cpu_smt_mask(int cpu)
5894 return topology_thread_cpumask(cpu);
5899 * Topology list, bottom-up.
5901 static struct sched_domain_topology_level default_topology[] = {
5902 #ifdef CONFIG_SCHED_SMT
5903 { sd_init_SIBLING, cpu_smt_mask, },
5905 #ifdef CONFIG_SCHED_MC
5906 { sd_init_MC, cpu_coregroup_mask, },
5908 #ifdef CONFIG_SCHED_BOOK
5909 { sd_init_BOOK, cpu_book_mask, },
5911 { sd_init_CPU, cpu_cpu_mask, },
5915 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5917 #define for_each_sd_topology(tl) \
5918 for (tl = sched_domain_topology; tl->init; tl++)
5922 static int sched_domains_numa_levels;
5923 static int *sched_domains_numa_distance;
5924 static struct cpumask ***sched_domains_numa_masks;
5925 static int sched_domains_curr_level;
5927 static inline int sd_local_flags(int level)
5929 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5932 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5935 static struct sched_domain *
5936 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
5938 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5939 int level = tl->numa_level;
5940 int sd_weight = cpumask_weight(
5941 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
5943 *sd = (struct sched_domain){
5944 .min_interval = sd_weight,
5945 .max_interval = 2*sd_weight,
5947 .imbalance_pct = 125,
5948 .cache_nice_tries = 2,
5955 .flags = 1*SD_LOAD_BALANCE
5956 | 1*SD_BALANCE_NEWIDLE
5961 | 0*SD_SHARE_CPUPOWER
5962 | 0*SD_SHARE_PKG_RESOURCES
5964 | 0*SD_PREFER_SIBLING
5966 | sd_local_flags(level)
5968 .last_balance = jiffies,
5969 .balance_interval = sd_weight,
5971 SD_INIT_NAME(sd, NUMA);
5972 sd->private = &tl->data;
5975 * Ugly hack to pass state to sd_numa_mask()...
5977 sched_domains_curr_level = tl->numa_level;
5982 static const struct cpumask *sd_numa_mask(int cpu)
5984 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
5987 static void sched_numa_warn(const char *str)
5989 static int done = false;
5997 printk(KERN_WARNING "ERROR: %s\n\n", str);
5999 for (i = 0; i < nr_node_ids; i++) {
6000 printk(KERN_WARNING " ");
6001 for (j = 0; j < nr_node_ids; j++)
6002 printk(KERN_CONT "%02d ", node_distance(i,j));
6003 printk(KERN_CONT "\n");
6005 printk(KERN_WARNING "\n");
6008 static bool find_numa_distance(int distance)
6012 if (distance == node_distance(0, 0))
6015 for (i = 0; i < sched_domains_numa_levels; i++) {
6016 if (sched_domains_numa_distance[i] == distance)
6023 static void sched_init_numa(void)
6025 int next_distance, curr_distance = node_distance(0, 0);
6026 struct sched_domain_topology_level *tl;
6030 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6031 if (!sched_domains_numa_distance)
6035 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6036 * unique distances in the node_distance() table.
6038 * Assumes node_distance(0,j) includes all distances in
6039 * node_distance(i,j) in order to avoid cubic time.
6041 next_distance = curr_distance;
6042 for (i = 0; i < nr_node_ids; i++) {
6043 for (j = 0; j < nr_node_ids; j++) {
6044 for (k = 0; k < nr_node_ids; k++) {
6045 int distance = node_distance(i, k);
6047 if (distance > curr_distance &&
6048 (distance < next_distance ||
6049 next_distance == curr_distance))
6050 next_distance = distance;
6053 * While not a strong assumption it would be nice to know
6054 * about cases where if node A is connected to B, B is not
6055 * equally connected to A.
6057 if (sched_debug() && node_distance(k, i) != distance)
6058 sched_numa_warn("Node-distance not symmetric");
6060 if (sched_debug() && i && !find_numa_distance(distance))
6061 sched_numa_warn("Node-0 not representative");
6063 if (next_distance != curr_distance) {
6064 sched_domains_numa_distance[level++] = next_distance;
6065 sched_domains_numa_levels = level;
6066 curr_distance = next_distance;
6071 * In case of sched_debug() we verify the above assumption.
6077 * 'level' contains the number of unique distances, excluding the
6078 * identity distance node_distance(i,i).
6080 * The sched_domains_numa_distance[] array includes the actual distance
6085 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6086 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6087 * the array will contain less then 'level' members. This could be
6088 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6089 * in other functions.
6091 * We reset it to 'level' at the end of this function.
6093 sched_domains_numa_levels = 0;
6095 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6096 if (!sched_domains_numa_masks)
6100 * Now for each level, construct a mask per node which contains all
6101 * cpus of nodes that are that many hops away from us.
6103 for (i = 0; i < level; i++) {
6104 sched_domains_numa_masks[i] =
6105 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6106 if (!sched_domains_numa_masks[i])
6109 for (j = 0; j < nr_node_ids; j++) {
6110 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6114 sched_domains_numa_masks[i][j] = mask;
6116 for (k = 0; k < nr_node_ids; k++) {
6117 if (node_distance(j, k) > sched_domains_numa_distance[i])
6120 cpumask_or(mask, mask, cpumask_of_node(k));
6125 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6126 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6131 * Copy the default topology bits..
6133 for (i = 0; default_topology[i].init; i++)
6134 tl[i] = default_topology[i];
6137 * .. and append 'j' levels of NUMA goodness.
6139 for (j = 0; j < level; i++, j++) {
6140 tl[i] = (struct sched_domain_topology_level){
6141 .init = sd_numa_init,
6142 .mask = sd_numa_mask,
6143 .flags = SDTL_OVERLAP,
6148 sched_domain_topology = tl;
6150 sched_domains_numa_levels = level;
6153 static void sched_domains_numa_masks_set(int cpu)
6156 int node = cpu_to_node(cpu);
6158 for (i = 0; i < sched_domains_numa_levels; i++) {
6159 for (j = 0; j < nr_node_ids; j++) {
6160 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6161 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6166 static void sched_domains_numa_masks_clear(int cpu)
6169 for (i = 0; i < sched_domains_numa_levels; i++) {
6170 for (j = 0; j < nr_node_ids; j++)
6171 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6176 * Update sched_domains_numa_masks[level][node] array when new cpus
6179 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6180 unsigned long action,
6183 int cpu = (long)hcpu;
6185 switch (action & ~CPU_TASKS_FROZEN) {
6187 sched_domains_numa_masks_set(cpu);
6191 sched_domains_numa_masks_clear(cpu);
6201 static inline void sched_init_numa(void)
6205 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6206 unsigned long action,
6211 #endif /* CONFIG_NUMA */
6213 static int __sdt_alloc(const struct cpumask *cpu_map)
6215 struct sched_domain_topology_level *tl;
6218 for_each_sd_topology(tl) {
6219 struct sd_data *sdd = &tl->data;
6221 sdd->sd = alloc_percpu(struct sched_domain *);
6225 sdd->sg = alloc_percpu(struct sched_group *);
6229 sdd->sgp = alloc_percpu(struct sched_group_power *);
6233 for_each_cpu(j, cpu_map) {
6234 struct sched_domain *sd;
6235 struct sched_group *sg;
6236 struct sched_group_power *sgp;
6238 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6239 GFP_KERNEL, cpu_to_node(j));
6243 *per_cpu_ptr(sdd->sd, j) = sd;
6245 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6246 GFP_KERNEL, cpu_to_node(j));
6252 *per_cpu_ptr(sdd->sg, j) = sg;
6254 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6255 GFP_KERNEL, cpu_to_node(j));
6259 *per_cpu_ptr(sdd->sgp, j) = sgp;
6266 static void __sdt_free(const struct cpumask *cpu_map)
6268 struct sched_domain_topology_level *tl;
6271 for_each_sd_topology(tl) {
6272 struct sd_data *sdd = &tl->data;
6274 for_each_cpu(j, cpu_map) {
6275 struct sched_domain *sd;
6278 sd = *per_cpu_ptr(sdd->sd, j);
6279 if (sd && (sd->flags & SD_OVERLAP))
6280 free_sched_groups(sd->groups, 0);
6281 kfree(*per_cpu_ptr(sdd->sd, j));
6285 kfree(*per_cpu_ptr(sdd->sg, j));
6287 kfree(*per_cpu_ptr(sdd->sgp, j));
6289 free_percpu(sdd->sd);
6291 free_percpu(sdd->sg);
6293 free_percpu(sdd->sgp);
6298 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6299 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6300 struct sched_domain *child, int cpu)
6302 struct sched_domain *sd = tl->init(tl, cpu);
6306 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6308 sd->level = child->level + 1;
6309 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6313 set_domain_attribute(sd, attr);
6319 * Build sched domains for a given set of cpus and attach the sched domains
6320 * to the individual cpus
6322 static int build_sched_domains(const struct cpumask *cpu_map,
6323 struct sched_domain_attr *attr)
6325 enum s_alloc alloc_state;
6326 struct sched_domain *sd;
6328 int i, ret = -ENOMEM;
6330 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6331 if (alloc_state != sa_rootdomain)
6334 /* Set up domains for cpus specified by the cpu_map. */
6335 for_each_cpu(i, cpu_map) {
6336 struct sched_domain_topology_level *tl;
6339 for_each_sd_topology(tl) {
6340 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6341 if (tl == sched_domain_topology)
6342 *per_cpu_ptr(d.sd, i) = sd;
6343 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6344 sd->flags |= SD_OVERLAP;
6345 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6350 /* Build the groups for the domains */
6351 for_each_cpu(i, cpu_map) {
6352 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6353 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6354 if (sd->flags & SD_OVERLAP) {
6355 if (build_overlap_sched_groups(sd, i))
6358 if (build_sched_groups(sd, i))
6364 /* Calculate CPU power for physical packages and nodes */
6365 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6366 if (!cpumask_test_cpu(i, cpu_map))
6369 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6370 claim_allocations(i, sd);
6371 init_sched_groups_power(i, sd);
6375 /* Attach the domains */
6377 for_each_cpu(i, cpu_map) {
6378 sd = *per_cpu_ptr(d.sd, i);
6379 cpu_attach_domain(sd, d.rd, i);
6385 __free_domain_allocs(&d, alloc_state, cpu_map);
6389 static cpumask_var_t *doms_cur; /* current sched domains */
6390 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6391 static struct sched_domain_attr *dattr_cur;
6392 /* attribues of custom domains in 'doms_cur' */
6395 * Special case: If a kmalloc of a doms_cur partition (array of
6396 * cpumask) fails, then fallback to a single sched domain,
6397 * as determined by the single cpumask fallback_doms.
6399 static cpumask_var_t fallback_doms;
6402 * arch_update_cpu_topology lets virtualized architectures update the
6403 * cpu core maps. It is supposed to return 1 if the topology changed
6404 * or 0 if it stayed the same.
6406 int __attribute__((weak)) arch_update_cpu_topology(void)
6411 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6414 cpumask_var_t *doms;
6416 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6419 for (i = 0; i < ndoms; i++) {
6420 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6421 free_sched_domains(doms, i);
6428 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6431 for (i = 0; i < ndoms; i++)
6432 free_cpumask_var(doms[i]);
6437 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6438 * For now this just excludes isolated cpus, but could be used to
6439 * exclude other special cases in the future.
6441 static int init_sched_domains(const struct cpumask *cpu_map)
6445 arch_update_cpu_topology();
6447 doms_cur = alloc_sched_domains(ndoms_cur);
6449 doms_cur = &fallback_doms;
6450 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6451 err = build_sched_domains(doms_cur[0], NULL);
6452 register_sched_domain_sysctl();
6458 * Detach sched domains from a group of cpus specified in cpu_map
6459 * These cpus will now be attached to the NULL domain
6461 static void detach_destroy_domains(const struct cpumask *cpu_map)
6466 for_each_cpu(i, cpu_map)
6467 cpu_attach_domain(NULL, &def_root_domain, i);
6471 /* handle null as "default" */
6472 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6473 struct sched_domain_attr *new, int idx_new)
6475 struct sched_domain_attr tmp;
6482 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6483 new ? (new + idx_new) : &tmp,
6484 sizeof(struct sched_domain_attr));
6488 * Partition sched domains as specified by the 'ndoms_new'
6489 * cpumasks in the array doms_new[] of cpumasks. This compares
6490 * doms_new[] to the current sched domain partitioning, doms_cur[].
6491 * It destroys each deleted domain and builds each new domain.
6493 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6494 * The masks don't intersect (don't overlap.) We should setup one
6495 * sched domain for each mask. CPUs not in any of the cpumasks will
6496 * not be load balanced. If the same cpumask appears both in the
6497 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6500 * The passed in 'doms_new' should be allocated using
6501 * alloc_sched_domains. This routine takes ownership of it and will
6502 * free_sched_domains it when done with it. If the caller failed the
6503 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6504 * and partition_sched_domains() will fallback to the single partition
6505 * 'fallback_doms', it also forces the domains to be rebuilt.
6507 * If doms_new == NULL it will be replaced with cpu_online_mask.
6508 * ndoms_new == 0 is a special case for destroying existing domains,
6509 * and it will not create the default domain.
6511 * Call with hotplug lock held
6513 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6514 struct sched_domain_attr *dattr_new)
6519 mutex_lock(&sched_domains_mutex);
6521 /* always unregister in case we don't destroy any domains */
6522 unregister_sched_domain_sysctl();
6524 /* Let architecture update cpu core mappings. */
6525 new_topology = arch_update_cpu_topology();
6527 n = doms_new ? ndoms_new : 0;
6529 /* Destroy deleted domains */
6530 for (i = 0; i < ndoms_cur; i++) {
6531 for (j = 0; j < n && !new_topology; j++) {
6532 if (cpumask_equal(doms_cur[i], doms_new[j])
6533 && dattrs_equal(dattr_cur, i, dattr_new, j))
6536 /* no match - a current sched domain not in new doms_new[] */
6537 detach_destroy_domains(doms_cur[i]);
6543 if (doms_new == NULL) {
6545 doms_new = &fallback_doms;
6546 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6547 WARN_ON_ONCE(dattr_new);
6550 /* Build new domains */
6551 for (i = 0; i < ndoms_new; i++) {
6552 for (j = 0; j < n && !new_topology; j++) {
6553 if (cpumask_equal(doms_new[i], doms_cur[j])
6554 && dattrs_equal(dattr_new, i, dattr_cur, j))
6557 /* no match - add a new doms_new */
6558 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6563 /* Remember the new sched domains */
6564 if (doms_cur != &fallback_doms)
6565 free_sched_domains(doms_cur, ndoms_cur);
6566 kfree(dattr_cur); /* kfree(NULL) is safe */
6567 doms_cur = doms_new;
6568 dattr_cur = dattr_new;
6569 ndoms_cur = ndoms_new;
6571 register_sched_domain_sysctl();
6573 mutex_unlock(&sched_domains_mutex);
6576 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6579 * Update cpusets according to cpu_active mask. If cpusets are
6580 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6581 * around partition_sched_domains().
6583 * If we come here as part of a suspend/resume, don't touch cpusets because we
6584 * want to restore it back to its original state upon resume anyway.
6586 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6590 case CPU_ONLINE_FROZEN:
6591 case CPU_DOWN_FAILED_FROZEN:
6594 * num_cpus_frozen tracks how many CPUs are involved in suspend
6595 * resume sequence. As long as this is not the last online
6596 * operation in the resume sequence, just build a single sched
6597 * domain, ignoring cpusets.
6600 if (likely(num_cpus_frozen)) {
6601 partition_sched_domains(1, NULL, NULL);
6606 * This is the last CPU online operation. So fall through and
6607 * restore the original sched domains by considering the
6608 * cpuset configurations.
6612 case CPU_DOWN_FAILED:
6613 cpuset_update_active_cpus(true);
6621 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6625 case CPU_DOWN_PREPARE:
6626 cpuset_update_active_cpus(false);
6628 case CPU_DOWN_PREPARE_FROZEN:
6630 partition_sched_domains(1, NULL, NULL);
6638 void __init sched_init_smp(void)
6640 cpumask_var_t non_isolated_cpus;
6642 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6643 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6648 * There's no userspace yet to cause hotplug operations; hence all the
6649 * cpu masks are stable and all blatant races in the below code cannot
6652 mutex_lock(&sched_domains_mutex);
6653 init_sched_domains(cpu_active_mask);
6654 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6655 if (cpumask_empty(non_isolated_cpus))
6656 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6657 mutex_unlock(&sched_domains_mutex);
6659 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6660 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6661 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6665 /* Move init over to a non-isolated CPU */
6666 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6668 sched_init_granularity();
6669 free_cpumask_var(non_isolated_cpus);
6671 init_sched_rt_class();
6672 init_sched_dl_class();
6675 void __init sched_init_smp(void)
6677 sched_init_granularity();
6679 #endif /* CONFIG_SMP */
6681 const_debug unsigned int sysctl_timer_migration = 1;
6683 int in_sched_functions(unsigned long addr)
6685 return in_lock_functions(addr) ||
6686 (addr >= (unsigned long)__sched_text_start
6687 && addr < (unsigned long)__sched_text_end);
6690 #ifdef CONFIG_CGROUP_SCHED
6692 * Default task group.
6693 * Every task in system belongs to this group at bootup.
6695 struct task_group root_task_group;
6696 LIST_HEAD(task_groups);
6699 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6701 void __init sched_init(void)
6704 unsigned long alloc_size = 0, ptr;
6706 #ifdef CONFIG_FAIR_GROUP_SCHED
6707 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6709 #ifdef CONFIG_RT_GROUP_SCHED
6710 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6712 #ifdef CONFIG_CPUMASK_OFFSTACK
6713 alloc_size += num_possible_cpus() * cpumask_size();
6716 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6718 #ifdef CONFIG_FAIR_GROUP_SCHED
6719 root_task_group.se = (struct sched_entity **)ptr;
6720 ptr += nr_cpu_ids * sizeof(void **);
6722 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6723 ptr += nr_cpu_ids * sizeof(void **);
6725 #endif /* CONFIG_FAIR_GROUP_SCHED */
6726 #ifdef CONFIG_RT_GROUP_SCHED
6727 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6728 ptr += nr_cpu_ids * sizeof(void **);
6730 root_task_group.rt_rq = (struct rt_rq **)ptr;
6731 ptr += nr_cpu_ids * sizeof(void **);
6733 #endif /* CONFIG_RT_GROUP_SCHED */
6734 #ifdef CONFIG_CPUMASK_OFFSTACK
6735 for_each_possible_cpu(i) {
6736 per_cpu(load_balance_mask, i) = (void *)ptr;
6737 ptr += cpumask_size();
6739 #endif /* CONFIG_CPUMASK_OFFSTACK */
6742 init_rt_bandwidth(&def_rt_bandwidth,
6743 global_rt_period(), global_rt_runtime());
6744 init_dl_bandwidth(&def_dl_bandwidth,
6745 global_rt_period(), global_rt_runtime());
6748 init_defrootdomain();
6751 #ifdef CONFIG_RT_GROUP_SCHED
6752 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6753 global_rt_period(), global_rt_runtime());
6754 #endif /* CONFIG_RT_GROUP_SCHED */
6756 #ifdef CONFIG_CGROUP_SCHED
6757 list_add(&root_task_group.list, &task_groups);
6758 INIT_LIST_HEAD(&root_task_group.children);
6759 INIT_LIST_HEAD(&root_task_group.siblings);
6760 autogroup_init(&init_task);
6762 #endif /* CONFIG_CGROUP_SCHED */
6764 for_each_possible_cpu(i) {
6768 raw_spin_lock_init(&rq->lock);
6770 rq->calc_load_active = 0;
6771 rq->calc_load_update = jiffies + LOAD_FREQ;
6772 init_cfs_rq(&rq->cfs);
6773 init_rt_rq(&rq->rt, rq);
6774 init_dl_rq(&rq->dl, rq);
6775 #ifdef CONFIG_FAIR_GROUP_SCHED
6776 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6777 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6779 * How much cpu bandwidth does root_task_group get?
6781 * In case of task-groups formed thr' the cgroup filesystem, it
6782 * gets 100% of the cpu resources in the system. This overall
6783 * system cpu resource is divided among the tasks of
6784 * root_task_group and its child task-groups in a fair manner,
6785 * based on each entity's (task or task-group's) weight
6786 * (se->load.weight).
6788 * In other words, if root_task_group has 10 tasks of weight
6789 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6790 * then A0's share of the cpu resource is:
6792 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6794 * We achieve this by letting root_task_group's tasks sit
6795 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6797 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6798 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6799 #endif /* CONFIG_FAIR_GROUP_SCHED */
6801 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6802 #ifdef CONFIG_RT_GROUP_SCHED
6803 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6804 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6807 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6808 rq->cpu_load[j] = 0;
6810 rq->last_load_update_tick = jiffies;
6815 rq->cpu_power = SCHED_POWER_SCALE;
6816 rq->post_schedule = 0;
6817 rq->active_balance = 0;
6818 rq->next_balance = jiffies;
6823 rq->avg_idle = 2*sysctl_sched_migration_cost;
6824 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6826 INIT_LIST_HEAD(&rq->cfs_tasks);
6828 rq_attach_root(rq, &def_root_domain);
6829 #ifdef CONFIG_NO_HZ_COMMON
6832 #ifdef CONFIG_NO_HZ_FULL
6833 rq->last_sched_tick = 0;
6837 atomic_set(&rq->nr_iowait, 0);
6840 set_load_weight(&init_task);
6842 #ifdef CONFIG_PREEMPT_NOTIFIERS
6843 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6847 * The boot idle thread does lazy MMU switching as well:
6849 atomic_inc(&init_mm.mm_count);
6850 enter_lazy_tlb(&init_mm, current);
6853 * Make us the idle thread. Technically, schedule() should not be
6854 * called from this thread, however somewhere below it might be,
6855 * but because we are the idle thread, we just pick up running again
6856 * when this runqueue becomes "idle".
6858 init_idle(current, smp_processor_id());
6860 calc_load_update = jiffies + LOAD_FREQ;
6863 * During early bootup we pretend to be a normal task:
6865 current->sched_class = &fair_sched_class;
6868 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6869 /* May be allocated at isolcpus cmdline parse time */
6870 if (cpu_isolated_map == NULL)
6871 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6872 idle_thread_set_boot_cpu();
6874 init_sched_fair_class();
6876 scheduler_running = 1;
6879 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6880 static inline int preempt_count_equals(int preempt_offset)
6882 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6884 return (nested == preempt_offset);
6887 void __might_sleep(const char *file, int line, int preempt_offset)
6889 static unsigned long prev_jiffy; /* ratelimiting */
6891 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6892 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6893 system_state != SYSTEM_RUNNING || oops_in_progress)
6895 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6897 prev_jiffy = jiffies;
6900 "BUG: sleeping function called from invalid context at %s:%d\n",
6903 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6904 in_atomic(), irqs_disabled(),
6905 current->pid, current->comm);
6907 debug_show_held_locks(current);
6908 if (irqs_disabled())
6909 print_irqtrace_events(current);
6912 EXPORT_SYMBOL(__might_sleep);
6915 #ifdef CONFIG_MAGIC_SYSRQ
6916 static void normalize_task(struct rq *rq, struct task_struct *p)
6918 const struct sched_class *prev_class = p->sched_class;
6919 struct sched_attr attr = {
6920 .sched_policy = SCHED_NORMAL,
6922 int old_prio = p->prio;
6927 dequeue_task(rq, p, 0);
6928 __setscheduler(rq, p, &attr);
6930 enqueue_task(rq, p, 0);
6931 resched_task(rq->curr);
6934 check_class_changed(rq, p, prev_class, old_prio);
6937 void normalize_rt_tasks(void)
6939 struct task_struct *g, *p;
6940 unsigned long flags;
6943 read_lock_irqsave(&tasklist_lock, flags);
6944 do_each_thread(g, p) {
6946 * Only normalize user tasks:
6951 p->se.exec_start = 0;
6952 #ifdef CONFIG_SCHEDSTATS
6953 p->se.statistics.wait_start = 0;
6954 p->se.statistics.sleep_start = 0;
6955 p->se.statistics.block_start = 0;
6958 if (!dl_task(p) && !rt_task(p)) {
6960 * Renice negative nice level userspace
6963 if (TASK_NICE(p) < 0 && p->mm)
6964 set_user_nice(p, 0);
6968 raw_spin_lock(&p->pi_lock);
6969 rq = __task_rq_lock(p);
6971 normalize_task(rq, p);
6973 __task_rq_unlock(rq);
6974 raw_spin_unlock(&p->pi_lock);
6975 } while_each_thread(g, p);
6977 read_unlock_irqrestore(&tasklist_lock, flags);
6980 #endif /* CONFIG_MAGIC_SYSRQ */
6982 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
6984 * These functions are only useful for the IA64 MCA handling, or kdb.
6986 * They can only be called when the whole system has been
6987 * stopped - every CPU needs to be quiescent, and no scheduling
6988 * activity can take place. Using them for anything else would
6989 * be a serious bug, and as a result, they aren't even visible
6990 * under any other configuration.
6994 * curr_task - return the current task for a given cpu.
6995 * @cpu: the processor in question.
6997 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
6999 * Return: The current task for @cpu.
7001 struct task_struct *curr_task(int cpu)
7003 return cpu_curr(cpu);
7006 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7010 * set_curr_task - set the current task for a given cpu.
7011 * @cpu: the processor in question.
7012 * @p: the task pointer to set.
7014 * Description: This function must only be used when non-maskable interrupts
7015 * are serviced on a separate stack. It allows the architecture to switch the
7016 * notion of the current task on a cpu in a non-blocking manner. This function
7017 * must be called with all CPU's synchronized, and interrupts disabled, the
7018 * and caller must save the original value of the current task (see
7019 * curr_task() above) and restore that value before reenabling interrupts and
7020 * re-starting the system.
7022 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7024 void set_curr_task(int cpu, struct task_struct *p)
7031 #ifdef CONFIG_CGROUP_SCHED
7032 /* task_group_lock serializes the addition/removal of task groups */
7033 static DEFINE_SPINLOCK(task_group_lock);
7035 static void free_sched_group(struct task_group *tg)
7037 free_fair_sched_group(tg);
7038 free_rt_sched_group(tg);
7043 /* allocate runqueue etc for a new task group */
7044 struct task_group *sched_create_group(struct task_group *parent)
7046 struct task_group *tg;
7048 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7050 return ERR_PTR(-ENOMEM);
7052 if (!alloc_fair_sched_group(tg, parent))
7055 if (!alloc_rt_sched_group(tg, parent))
7061 free_sched_group(tg);
7062 return ERR_PTR(-ENOMEM);
7065 void sched_online_group(struct task_group *tg, struct task_group *parent)
7067 unsigned long flags;
7069 spin_lock_irqsave(&task_group_lock, flags);
7070 list_add_rcu(&tg->list, &task_groups);
7072 WARN_ON(!parent); /* root should already exist */
7074 tg->parent = parent;
7075 INIT_LIST_HEAD(&tg->children);
7076 list_add_rcu(&tg->siblings, &parent->children);
7077 spin_unlock_irqrestore(&task_group_lock, flags);
7080 /* rcu callback to free various structures associated with a task group */
7081 static void free_sched_group_rcu(struct rcu_head *rhp)
7083 /* now it should be safe to free those cfs_rqs */
7084 free_sched_group(container_of(rhp, struct task_group, rcu));
7087 /* Destroy runqueue etc associated with a task group */
7088 void sched_destroy_group(struct task_group *tg)
7090 /* wait for possible concurrent references to cfs_rqs complete */
7091 call_rcu(&tg->rcu, free_sched_group_rcu);
7094 void sched_offline_group(struct task_group *tg)
7096 unsigned long flags;
7099 /* end participation in shares distribution */
7100 for_each_possible_cpu(i)
7101 unregister_fair_sched_group(tg, i);
7103 spin_lock_irqsave(&task_group_lock, flags);
7104 list_del_rcu(&tg->list);
7105 list_del_rcu(&tg->siblings);
7106 spin_unlock_irqrestore(&task_group_lock, flags);
7109 /* change task's runqueue when it moves between groups.
7110 * The caller of this function should have put the task in its new group
7111 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7112 * reflect its new group.
7114 void sched_move_task(struct task_struct *tsk)
7116 struct task_group *tg;
7118 unsigned long flags;
7121 rq = task_rq_lock(tsk, &flags);
7123 running = task_current(rq, tsk);
7127 dequeue_task(rq, tsk, 0);
7128 if (unlikely(running))
7129 tsk->sched_class->put_prev_task(rq, tsk);
7131 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
7132 lockdep_is_held(&tsk->sighand->siglock)),
7133 struct task_group, css);
7134 tg = autogroup_task_group(tsk, tg);
7135 tsk->sched_task_group = tg;
7137 #ifdef CONFIG_FAIR_GROUP_SCHED
7138 if (tsk->sched_class->task_move_group)
7139 tsk->sched_class->task_move_group(tsk, on_rq);
7142 set_task_rq(tsk, task_cpu(tsk));
7144 if (unlikely(running))
7145 tsk->sched_class->set_curr_task(rq);
7147 enqueue_task(rq, tsk, 0);
7149 task_rq_unlock(rq, tsk, &flags);
7151 #endif /* CONFIG_CGROUP_SCHED */
7153 #ifdef CONFIG_RT_GROUP_SCHED
7155 * Ensure that the real time constraints are schedulable.
7157 static DEFINE_MUTEX(rt_constraints_mutex);
7159 /* Must be called with tasklist_lock held */
7160 static inline int tg_has_rt_tasks(struct task_group *tg)
7162 struct task_struct *g, *p;
7164 do_each_thread(g, p) {
7165 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7167 } while_each_thread(g, p);
7172 struct rt_schedulable_data {
7173 struct task_group *tg;
7178 static int tg_rt_schedulable(struct task_group *tg, void *data)
7180 struct rt_schedulable_data *d = data;
7181 struct task_group *child;
7182 unsigned long total, sum = 0;
7183 u64 period, runtime;
7185 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7186 runtime = tg->rt_bandwidth.rt_runtime;
7189 period = d->rt_period;
7190 runtime = d->rt_runtime;
7194 * Cannot have more runtime than the period.
7196 if (runtime > period && runtime != RUNTIME_INF)
7200 * Ensure we don't starve existing RT tasks.
7202 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7205 total = to_ratio(period, runtime);
7208 * Nobody can have more than the global setting allows.
7210 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7214 * The sum of our children's runtime should not exceed our own.
7216 list_for_each_entry_rcu(child, &tg->children, siblings) {
7217 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7218 runtime = child->rt_bandwidth.rt_runtime;
7220 if (child == d->tg) {
7221 period = d->rt_period;
7222 runtime = d->rt_runtime;
7225 sum += to_ratio(period, runtime);
7234 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7238 struct rt_schedulable_data data = {
7240 .rt_period = period,
7241 .rt_runtime = runtime,
7245 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7251 static int tg_set_rt_bandwidth(struct task_group *tg,
7252 u64 rt_period, u64 rt_runtime)
7256 mutex_lock(&rt_constraints_mutex);
7257 read_lock(&tasklist_lock);
7258 err = __rt_schedulable(tg, rt_period, rt_runtime);
7262 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7263 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7264 tg->rt_bandwidth.rt_runtime = rt_runtime;
7266 for_each_possible_cpu(i) {
7267 struct rt_rq *rt_rq = tg->rt_rq[i];
7269 raw_spin_lock(&rt_rq->rt_runtime_lock);
7270 rt_rq->rt_runtime = rt_runtime;
7271 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7273 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7275 read_unlock(&tasklist_lock);
7276 mutex_unlock(&rt_constraints_mutex);
7281 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7283 u64 rt_runtime, rt_period;
7285 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7286 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7287 if (rt_runtime_us < 0)
7288 rt_runtime = RUNTIME_INF;
7290 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7293 static long sched_group_rt_runtime(struct task_group *tg)
7297 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7300 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7301 do_div(rt_runtime_us, NSEC_PER_USEC);
7302 return rt_runtime_us;
7305 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7307 u64 rt_runtime, rt_period;
7309 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7310 rt_runtime = tg->rt_bandwidth.rt_runtime;
7315 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7318 static long sched_group_rt_period(struct task_group *tg)
7322 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7323 do_div(rt_period_us, NSEC_PER_USEC);
7324 return rt_period_us;
7326 #endif /* CONFIG_RT_GROUP_SCHED */
7328 #ifdef CONFIG_RT_GROUP_SCHED
7329 static int sched_rt_global_constraints(void)
7333 mutex_lock(&rt_constraints_mutex);
7334 read_lock(&tasklist_lock);
7335 ret = __rt_schedulable(NULL, 0, 0);
7336 read_unlock(&tasklist_lock);
7337 mutex_unlock(&rt_constraints_mutex);
7342 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7344 /* Don't accept realtime tasks when there is no way for them to run */
7345 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7351 #else /* !CONFIG_RT_GROUP_SCHED */
7352 static int sched_rt_global_constraints(void)
7354 unsigned long flags;
7357 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7358 for_each_possible_cpu(i) {
7359 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7361 raw_spin_lock(&rt_rq->rt_runtime_lock);
7362 rt_rq->rt_runtime = global_rt_runtime();
7363 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7365 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7369 #endif /* CONFIG_RT_GROUP_SCHED */
7371 static int sched_dl_global_constraints(void)
7373 u64 runtime = global_rt_runtime();
7374 u64 period = global_rt_period();
7375 u64 new_bw = to_ratio(period, runtime);
7379 * Here we want to check the bandwidth not being set to some
7380 * value smaller than the currently allocated bandwidth in
7381 * any of the root_domains.
7383 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7384 * cycling on root_domains... Discussion on different/better
7385 * solutions is welcome!
7387 for_each_possible_cpu(cpu) {
7388 struct dl_bw *dl_b = dl_bw_of(cpu);
7390 raw_spin_lock(&dl_b->lock);
7391 if (new_bw < dl_b->total_bw)
7393 raw_spin_unlock(&dl_b->lock);
7402 static void sched_dl_do_global(void)
7407 def_dl_bandwidth.dl_period = global_rt_period();
7408 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7410 if (global_rt_runtime() != RUNTIME_INF)
7411 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7414 * FIXME: As above...
7416 for_each_possible_cpu(cpu) {
7417 struct dl_bw *dl_b = dl_bw_of(cpu);
7419 raw_spin_lock(&dl_b->lock);
7421 raw_spin_unlock(&dl_b->lock);
7425 static int sched_rt_global_validate(void)
7427 if (sysctl_sched_rt_period <= 0)
7430 if (sysctl_sched_rt_runtime > sysctl_sched_rt_period)
7436 static void sched_rt_do_global(void)
7438 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7439 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7442 int sched_rt_handler(struct ctl_table *table, int write,
7443 void __user *buffer, size_t *lenp,
7446 int old_period, old_runtime;
7447 static DEFINE_MUTEX(mutex);
7451 old_period = sysctl_sched_rt_period;
7452 old_runtime = sysctl_sched_rt_runtime;
7454 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7456 if (!ret && write) {
7457 ret = sched_rt_global_validate();
7461 ret = sched_rt_global_constraints();
7465 ret = sched_dl_global_constraints();
7469 sched_rt_do_global();
7470 sched_dl_do_global();
7474 sysctl_sched_rt_period = old_period;
7475 sysctl_sched_rt_runtime = old_runtime;
7477 mutex_unlock(&mutex);
7482 int sched_rr_handler(struct ctl_table *table, int write,
7483 void __user *buffer, size_t *lenp,
7487 static DEFINE_MUTEX(mutex);
7490 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7491 /* make sure that internally we keep jiffies */
7492 /* also, writing zero resets timeslice to default */
7493 if (!ret && write) {
7494 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7495 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7497 mutex_unlock(&mutex);
7501 #ifdef CONFIG_CGROUP_SCHED
7503 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7505 return css ? container_of(css, struct task_group, css) : NULL;
7508 static struct cgroup_subsys_state *
7509 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7511 struct task_group *parent = css_tg(parent_css);
7512 struct task_group *tg;
7515 /* This is early initialization for the top cgroup */
7516 return &root_task_group.css;
7519 tg = sched_create_group(parent);
7521 return ERR_PTR(-ENOMEM);
7526 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7528 struct task_group *tg = css_tg(css);
7529 struct task_group *parent = css_tg(css_parent(css));
7532 sched_online_group(tg, parent);
7536 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7538 struct task_group *tg = css_tg(css);
7540 sched_destroy_group(tg);
7543 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7545 struct task_group *tg = css_tg(css);
7547 sched_offline_group(tg);
7550 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7551 struct cgroup_taskset *tset)
7553 struct task_struct *task;
7555 cgroup_taskset_for_each(task, css, tset) {
7556 #ifdef CONFIG_RT_GROUP_SCHED
7557 if (!sched_rt_can_attach(css_tg(css), task))
7560 /* We don't support RT-tasks being in separate groups */
7561 if (task->sched_class != &fair_sched_class)
7568 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7569 struct cgroup_taskset *tset)
7571 struct task_struct *task;
7573 cgroup_taskset_for_each(task, css, tset)
7574 sched_move_task(task);
7577 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7578 struct cgroup_subsys_state *old_css,
7579 struct task_struct *task)
7582 * cgroup_exit() is called in the copy_process() failure path.
7583 * Ignore this case since the task hasn't ran yet, this avoids
7584 * trying to poke a half freed task state from generic code.
7586 if (!(task->flags & PF_EXITING))
7589 sched_move_task(task);
7592 #ifdef CONFIG_FAIR_GROUP_SCHED
7593 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7594 struct cftype *cftype, u64 shareval)
7596 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7599 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7602 struct task_group *tg = css_tg(css);
7604 return (u64) scale_load_down(tg->shares);
7607 #ifdef CONFIG_CFS_BANDWIDTH
7608 static DEFINE_MUTEX(cfs_constraints_mutex);
7610 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7611 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7613 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7615 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7617 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7618 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7620 if (tg == &root_task_group)
7624 * Ensure we have at some amount of bandwidth every period. This is
7625 * to prevent reaching a state of large arrears when throttled via
7626 * entity_tick() resulting in prolonged exit starvation.
7628 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7632 * Likewise, bound things on the otherside by preventing insane quota
7633 * periods. This also allows us to normalize in computing quota
7636 if (period > max_cfs_quota_period)
7639 mutex_lock(&cfs_constraints_mutex);
7640 ret = __cfs_schedulable(tg, period, quota);
7644 runtime_enabled = quota != RUNTIME_INF;
7645 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7647 * If we need to toggle cfs_bandwidth_used, off->on must occur
7648 * before making related changes, and on->off must occur afterwards
7650 if (runtime_enabled && !runtime_was_enabled)
7651 cfs_bandwidth_usage_inc();
7652 raw_spin_lock_irq(&cfs_b->lock);
7653 cfs_b->period = ns_to_ktime(period);
7654 cfs_b->quota = quota;
7656 __refill_cfs_bandwidth_runtime(cfs_b);
7657 /* restart the period timer (if active) to handle new period expiry */
7658 if (runtime_enabled && cfs_b->timer_active) {
7659 /* force a reprogram */
7660 cfs_b->timer_active = 0;
7661 __start_cfs_bandwidth(cfs_b);
7663 raw_spin_unlock_irq(&cfs_b->lock);
7665 for_each_possible_cpu(i) {
7666 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7667 struct rq *rq = cfs_rq->rq;
7669 raw_spin_lock_irq(&rq->lock);
7670 cfs_rq->runtime_enabled = runtime_enabled;
7671 cfs_rq->runtime_remaining = 0;
7673 if (cfs_rq->throttled)
7674 unthrottle_cfs_rq(cfs_rq);
7675 raw_spin_unlock_irq(&rq->lock);
7677 if (runtime_was_enabled && !runtime_enabled)
7678 cfs_bandwidth_usage_dec();
7680 mutex_unlock(&cfs_constraints_mutex);
7685 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7689 period = ktime_to_ns(tg->cfs_bandwidth.period);
7690 if (cfs_quota_us < 0)
7691 quota = RUNTIME_INF;
7693 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7695 return tg_set_cfs_bandwidth(tg, period, quota);
7698 long tg_get_cfs_quota(struct task_group *tg)
7702 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7705 quota_us = tg->cfs_bandwidth.quota;
7706 do_div(quota_us, NSEC_PER_USEC);
7711 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7715 period = (u64)cfs_period_us * NSEC_PER_USEC;
7716 quota = tg->cfs_bandwidth.quota;
7718 return tg_set_cfs_bandwidth(tg, period, quota);
7721 long tg_get_cfs_period(struct task_group *tg)
7725 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7726 do_div(cfs_period_us, NSEC_PER_USEC);
7728 return cfs_period_us;
7731 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7734 return tg_get_cfs_quota(css_tg(css));
7737 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7738 struct cftype *cftype, s64 cfs_quota_us)
7740 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7743 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7746 return tg_get_cfs_period(css_tg(css));
7749 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7750 struct cftype *cftype, u64 cfs_period_us)
7752 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7755 struct cfs_schedulable_data {
7756 struct task_group *tg;
7761 * normalize group quota/period to be quota/max_period
7762 * note: units are usecs
7764 static u64 normalize_cfs_quota(struct task_group *tg,
7765 struct cfs_schedulable_data *d)
7773 period = tg_get_cfs_period(tg);
7774 quota = tg_get_cfs_quota(tg);
7777 /* note: these should typically be equivalent */
7778 if (quota == RUNTIME_INF || quota == -1)
7781 return to_ratio(period, quota);
7784 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7786 struct cfs_schedulable_data *d = data;
7787 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7788 s64 quota = 0, parent_quota = -1;
7791 quota = RUNTIME_INF;
7793 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7795 quota = normalize_cfs_quota(tg, d);
7796 parent_quota = parent_b->hierarchal_quota;
7799 * ensure max(child_quota) <= parent_quota, inherit when no
7802 if (quota == RUNTIME_INF)
7803 quota = parent_quota;
7804 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7807 cfs_b->hierarchal_quota = quota;
7812 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7815 struct cfs_schedulable_data data = {
7821 if (quota != RUNTIME_INF) {
7822 do_div(data.period, NSEC_PER_USEC);
7823 do_div(data.quota, NSEC_PER_USEC);
7827 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7833 static int cpu_stats_show(struct cgroup_subsys_state *css, struct cftype *cft,
7834 struct cgroup_map_cb *cb)
7836 struct task_group *tg = css_tg(css);
7837 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7839 cb->fill(cb, "nr_periods", cfs_b->nr_periods);
7840 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled);
7841 cb->fill(cb, "throttled_time", cfs_b->throttled_time);
7845 #endif /* CONFIG_CFS_BANDWIDTH */
7846 #endif /* CONFIG_FAIR_GROUP_SCHED */
7848 #ifdef CONFIG_RT_GROUP_SCHED
7849 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7850 struct cftype *cft, s64 val)
7852 return sched_group_set_rt_runtime(css_tg(css), val);
7855 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7858 return sched_group_rt_runtime(css_tg(css));
7861 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7862 struct cftype *cftype, u64 rt_period_us)
7864 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7867 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7870 return sched_group_rt_period(css_tg(css));
7872 #endif /* CONFIG_RT_GROUP_SCHED */
7874 static struct cftype cpu_files[] = {
7875 #ifdef CONFIG_FAIR_GROUP_SCHED
7878 .read_u64 = cpu_shares_read_u64,
7879 .write_u64 = cpu_shares_write_u64,
7882 #ifdef CONFIG_CFS_BANDWIDTH
7884 .name = "cfs_quota_us",
7885 .read_s64 = cpu_cfs_quota_read_s64,
7886 .write_s64 = cpu_cfs_quota_write_s64,
7889 .name = "cfs_period_us",
7890 .read_u64 = cpu_cfs_period_read_u64,
7891 .write_u64 = cpu_cfs_period_write_u64,
7895 .read_map = cpu_stats_show,
7898 #ifdef CONFIG_RT_GROUP_SCHED
7900 .name = "rt_runtime_us",
7901 .read_s64 = cpu_rt_runtime_read,
7902 .write_s64 = cpu_rt_runtime_write,
7905 .name = "rt_period_us",
7906 .read_u64 = cpu_rt_period_read_uint,
7907 .write_u64 = cpu_rt_period_write_uint,
7913 struct cgroup_subsys cpu_cgroup_subsys = {
7915 .css_alloc = cpu_cgroup_css_alloc,
7916 .css_free = cpu_cgroup_css_free,
7917 .css_online = cpu_cgroup_css_online,
7918 .css_offline = cpu_cgroup_css_offline,
7919 .can_attach = cpu_cgroup_can_attach,
7920 .attach = cpu_cgroup_attach,
7921 .exit = cpu_cgroup_exit,
7922 .subsys_id = cpu_cgroup_subsys_id,
7923 .base_cftypes = cpu_files,
7927 #endif /* CONFIG_CGROUP_SCHED */
7929 void dump_cpu_task(int cpu)
7931 pr_info("Task dump for CPU %d:\n", cpu);
7932 sched_show_task(cpu_curr(cpu));