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 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1112 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1118 struct migration_arg {
1119 struct task_struct *task;
1123 static int migration_cpu_stop(void *data);
1126 * wait_task_inactive - wait for a thread to unschedule.
1128 * If @match_state is nonzero, it's the @p->state value just checked and
1129 * not expected to change. If it changes, i.e. @p might have woken up,
1130 * then return zero. When we succeed in waiting for @p to be off its CPU,
1131 * we return a positive number (its total switch count). If a second call
1132 * a short while later returns the same number, the caller can be sure that
1133 * @p has remained unscheduled the whole time.
1135 * The caller must ensure that the task *will* unschedule sometime soon,
1136 * else this function might spin for a *long* time. This function can't
1137 * be called with interrupts off, or it may introduce deadlock with
1138 * smp_call_function() if an IPI is sent by the same process we are
1139 * waiting to become inactive.
1141 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1143 unsigned long flags;
1150 * We do the initial early heuristics without holding
1151 * any task-queue locks at all. We'll only try to get
1152 * the runqueue lock when things look like they will
1158 * If the task is actively running on another CPU
1159 * still, just relax and busy-wait without holding
1162 * NOTE! Since we don't hold any locks, it's not
1163 * even sure that "rq" stays as the right runqueue!
1164 * But we don't care, since "task_running()" will
1165 * return false if the runqueue has changed and p
1166 * is actually now running somewhere else!
1168 while (task_running(rq, p)) {
1169 if (match_state && unlikely(p->state != match_state))
1175 * Ok, time to look more closely! We need the rq
1176 * lock now, to be *sure*. If we're wrong, we'll
1177 * just go back and repeat.
1179 rq = task_rq_lock(p, &flags);
1180 trace_sched_wait_task(p);
1181 running = task_running(rq, p);
1184 if (!match_state || p->state == match_state)
1185 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1186 task_rq_unlock(rq, p, &flags);
1189 * If it changed from the expected state, bail out now.
1191 if (unlikely(!ncsw))
1195 * Was it really running after all now that we
1196 * checked with the proper locks actually held?
1198 * Oops. Go back and try again..
1200 if (unlikely(running)) {
1206 * It's not enough that it's not actively running,
1207 * it must be off the runqueue _entirely_, and not
1210 * So if it was still runnable (but just not actively
1211 * running right now), it's preempted, and we should
1212 * yield - it could be a while.
1214 if (unlikely(on_rq)) {
1215 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1217 set_current_state(TASK_UNINTERRUPTIBLE);
1218 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1223 * Ahh, all good. It wasn't running, and it wasn't
1224 * runnable, which means that it will never become
1225 * running in the future either. We're all done!
1234 * kick_process - kick a running thread to enter/exit the kernel
1235 * @p: the to-be-kicked thread
1237 * Cause a process which is running on another CPU to enter
1238 * kernel-mode, without any delay. (to get signals handled.)
1240 * NOTE: this function doesn't have to take the runqueue lock,
1241 * because all it wants to ensure is that the remote task enters
1242 * the kernel. If the IPI races and the task has been migrated
1243 * to another CPU then no harm is done and the purpose has been
1246 void kick_process(struct task_struct *p)
1252 if ((cpu != smp_processor_id()) && task_curr(p))
1253 smp_send_reschedule(cpu);
1256 EXPORT_SYMBOL_GPL(kick_process);
1257 #endif /* CONFIG_SMP */
1261 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1263 static int select_fallback_rq(int cpu, struct task_struct *p)
1265 int nid = cpu_to_node(cpu);
1266 const struct cpumask *nodemask = NULL;
1267 enum { cpuset, possible, fail } state = cpuset;
1271 * If the node that the cpu is on has been offlined, cpu_to_node()
1272 * will return -1. There is no cpu on the node, and we should
1273 * select the cpu on the other node.
1276 nodemask = cpumask_of_node(nid);
1278 /* Look for allowed, online CPU in same node. */
1279 for_each_cpu(dest_cpu, nodemask) {
1280 if (!cpu_online(dest_cpu))
1282 if (!cpu_active(dest_cpu))
1284 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1290 /* Any allowed, online CPU? */
1291 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1292 if (!cpu_online(dest_cpu))
1294 if (!cpu_active(dest_cpu))
1301 /* No more Mr. Nice Guy. */
1302 cpuset_cpus_allowed_fallback(p);
1307 do_set_cpus_allowed(p, cpu_possible_mask);
1318 if (state != cpuset) {
1320 * Don't tell them about moving exiting tasks or
1321 * kernel threads (both mm NULL), since they never
1324 if (p->mm && printk_ratelimit()) {
1325 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1326 task_pid_nr(p), p->comm, cpu);
1334 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1337 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1339 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1342 * In order not to call set_task_cpu() on a blocking task we need
1343 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1346 * Since this is common to all placement strategies, this lives here.
1348 * [ this allows ->select_task() to simply return task_cpu(p) and
1349 * not worry about this generic constraint ]
1351 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1353 cpu = select_fallback_rq(task_cpu(p), p);
1358 static void update_avg(u64 *avg, u64 sample)
1360 s64 diff = sample - *avg;
1366 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1368 #ifdef CONFIG_SCHEDSTATS
1369 struct rq *rq = this_rq();
1372 int this_cpu = smp_processor_id();
1374 if (cpu == this_cpu) {
1375 schedstat_inc(rq, ttwu_local);
1376 schedstat_inc(p, se.statistics.nr_wakeups_local);
1378 struct sched_domain *sd;
1380 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1382 for_each_domain(this_cpu, sd) {
1383 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1384 schedstat_inc(sd, ttwu_wake_remote);
1391 if (wake_flags & WF_MIGRATED)
1392 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1394 #endif /* CONFIG_SMP */
1396 schedstat_inc(rq, ttwu_count);
1397 schedstat_inc(p, se.statistics.nr_wakeups);
1399 if (wake_flags & WF_SYNC)
1400 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1402 #endif /* CONFIG_SCHEDSTATS */
1405 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1407 activate_task(rq, p, en_flags);
1410 /* if a worker is waking up, notify workqueue */
1411 if (p->flags & PF_WQ_WORKER)
1412 wq_worker_waking_up(p, cpu_of(rq));
1416 * Mark the task runnable and perform wakeup-preemption.
1419 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1421 check_preempt_curr(rq, p, wake_flags);
1422 trace_sched_wakeup(p, true);
1424 p->state = TASK_RUNNING;
1426 if (p->sched_class->task_woken)
1427 p->sched_class->task_woken(rq, p);
1429 if (rq->idle_stamp) {
1430 u64 delta = rq_clock(rq) - rq->idle_stamp;
1431 u64 max = 2*rq->max_idle_balance_cost;
1433 update_avg(&rq->avg_idle, delta);
1435 if (rq->avg_idle > max)
1444 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1447 if (p->sched_contributes_to_load)
1448 rq->nr_uninterruptible--;
1451 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1452 ttwu_do_wakeup(rq, p, wake_flags);
1456 * Called in case the task @p isn't fully descheduled from its runqueue,
1457 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1458 * since all we need to do is flip p->state to TASK_RUNNING, since
1459 * the task is still ->on_rq.
1461 static int ttwu_remote(struct task_struct *p, int wake_flags)
1466 rq = __task_rq_lock(p);
1468 /* check_preempt_curr() may use rq clock */
1469 update_rq_clock(rq);
1470 ttwu_do_wakeup(rq, p, wake_flags);
1473 __task_rq_unlock(rq);
1479 static void sched_ttwu_pending(void)
1481 struct rq *rq = this_rq();
1482 struct llist_node *llist = llist_del_all(&rq->wake_list);
1483 struct task_struct *p;
1485 raw_spin_lock(&rq->lock);
1488 p = llist_entry(llist, struct task_struct, wake_entry);
1489 llist = llist_next(llist);
1490 ttwu_do_activate(rq, p, 0);
1493 raw_spin_unlock(&rq->lock);
1496 void scheduler_ipi(void)
1499 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1500 * TIF_NEED_RESCHED remotely (for the first time) will also send
1503 preempt_fold_need_resched();
1505 if (llist_empty(&this_rq()->wake_list)
1506 && !tick_nohz_full_cpu(smp_processor_id())
1507 && !got_nohz_idle_kick())
1511 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1512 * traditionally all their work was done from the interrupt return
1513 * path. Now that we actually do some work, we need to make sure
1516 * Some archs already do call them, luckily irq_enter/exit nest
1519 * Arguably we should visit all archs and update all handlers,
1520 * however a fair share of IPIs are still resched only so this would
1521 * somewhat pessimize the simple resched case.
1524 tick_nohz_full_check();
1525 sched_ttwu_pending();
1528 * Check if someone kicked us for doing the nohz idle load balance.
1530 if (unlikely(got_nohz_idle_kick())) {
1531 this_rq()->idle_balance = 1;
1532 raise_softirq_irqoff(SCHED_SOFTIRQ);
1537 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1539 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1540 smp_send_reschedule(cpu);
1543 bool cpus_share_cache(int this_cpu, int that_cpu)
1545 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1547 #endif /* CONFIG_SMP */
1549 static void ttwu_queue(struct task_struct *p, int cpu)
1551 struct rq *rq = cpu_rq(cpu);
1553 #if defined(CONFIG_SMP)
1554 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1555 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1556 ttwu_queue_remote(p, cpu);
1561 raw_spin_lock(&rq->lock);
1562 ttwu_do_activate(rq, p, 0);
1563 raw_spin_unlock(&rq->lock);
1567 * try_to_wake_up - wake up a thread
1568 * @p: the thread to be awakened
1569 * @state: the mask of task states that can be woken
1570 * @wake_flags: wake modifier flags (WF_*)
1572 * Put it on the run-queue if it's not already there. The "current"
1573 * thread is always on the run-queue (except when the actual
1574 * re-schedule is in progress), and as such you're allowed to do
1575 * the simpler "current->state = TASK_RUNNING" to mark yourself
1576 * runnable without the overhead of this.
1578 * Return: %true if @p was woken up, %false if it was already running.
1579 * or @state didn't match @p's state.
1582 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1584 unsigned long flags;
1585 int cpu, success = 0;
1588 * If we are going to wake up a thread waiting for CONDITION we
1589 * need to ensure that CONDITION=1 done by the caller can not be
1590 * reordered with p->state check below. This pairs with mb() in
1591 * set_current_state() the waiting thread does.
1593 smp_mb__before_spinlock();
1594 raw_spin_lock_irqsave(&p->pi_lock, flags);
1595 if (!(p->state & state))
1598 success = 1; /* we're going to change ->state */
1601 if (p->on_rq && ttwu_remote(p, wake_flags))
1606 * If the owning (remote) cpu is still in the middle of schedule() with
1607 * this task as prev, wait until its done referencing the task.
1612 * Pairs with the smp_wmb() in finish_lock_switch().
1616 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1617 p->state = TASK_WAKING;
1619 if (p->sched_class->task_waking)
1620 p->sched_class->task_waking(p);
1622 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1623 if (task_cpu(p) != cpu) {
1624 wake_flags |= WF_MIGRATED;
1625 set_task_cpu(p, cpu);
1627 #endif /* CONFIG_SMP */
1631 ttwu_stat(p, cpu, wake_flags);
1633 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1639 * try_to_wake_up_local - try to wake up a local task with rq lock held
1640 * @p: the thread to be awakened
1642 * Put @p on the run-queue if it's not already there. The caller must
1643 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1646 static void try_to_wake_up_local(struct task_struct *p)
1648 struct rq *rq = task_rq(p);
1650 if (WARN_ON_ONCE(rq != this_rq()) ||
1651 WARN_ON_ONCE(p == current))
1654 lockdep_assert_held(&rq->lock);
1656 if (!raw_spin_trylock(&p->pi_lock)) {
1657 raw_spin_unlock(&rq->lock);
1658 raw_spin_lock(&p->pi_lock);
1659 raw_spin_lock(&rq->lock);
1662 if (!(p->state & TASK_NORMAL))
1666 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1668 ttwu_do_wakeup(rq, p, 0);
1669 ttwu_stat(p, smp_processor_id(), 0);
1671 raw_spin_unlock(&p->pi_lock);
1675 * wake_up_process - Wake up a specific process
1676 * @p: The process to be woken up.
1678 * Attempt to wake up the nominated process and move it to the set of runnable
1681 * Return: 1 if the process was woken up, 0 if it was already running.
1683 * It may be assumed that this function implies a write memory barrier before
1684 * changing the task state if and only if any tasks are woken up.
1686 int wake_up_process(struct task_struct *p)
1688 WARN_ON(task_is_stopped_or_traced(p));
1689 return try_to_wake_up(p, TASK_NORMAL, 0);
1691 EXPORT_SYMBOL(wake_up_process);
1693 int wake_up_state(struct task_struct *p, unsigned int state)
1695 return try_to_wake_up(p, state, 0);
1699 * Perform scheduler related setup for a newly forked process p.
1700 * p is forked by current.
1702 * __sched_fork() is basic setup used by init_idle() too:
1704 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1709 p->se.exec_start = 0;
1710 p->se.sum_exec_runtime = 0;
1711 p->se.prev_sum_exec_runtime = 0;
1712 p->se.nr_migrations = 0;
1714 INIT_LIST_HEAD(&p->se.group_node);
1716 #ifdef CONFIG_SCHEDSTATS
1717 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1720 RB_CLEAR_NODE(&p->dl.rb_node);
1721 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1722 p->dl.dl_runtime = p->dl.runtime = 0;
1723 p->dl.dl_deadline = p->dl.deadline = 0;
1724 p->dl.dl_period = 0;
1727 INIT_LIST_HEAD(&p->rt.run_list);
1729 #ifdef CONFIG_PREEMPT_NOTIFIERS
1730 INIT_HLIST_HEAD(&p->preempt_notifiers);
1733 #ifdef CONFIG_NUMA_BALANCING
1734 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1735 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1736 p->mm->numa_scan_seq = 0;
1739 if (clone_flags & CLONE_VM)
1740 p->numa_preferred_nid = current->numa_preferred_nid;
1742 p->numa_preferred_nid = -1;
1744 p->node_stamp = 0ULL;
1745 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1746 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1747 p->numa_work.next = &p->numa_work;
1748 p->numa_faults = NULL;
1749 p->numa_faults_buffer = NULL;
1751 INIT_LIST_HEAD(&p->numa_entry);
1752 p->numa_group = NULL;
1753 #endif /* CONFIG_NUMA_BALANCING */
1756 #ifdef CONFIG_NUMA_BALANCING
1757 #ifdef CONFIG_SCHED_DEBUG
1758 void set_numabalancing_state(bool enabled)
1761 sched_feat_set("NUMA");
1763 sched_feat_set("NO_NUMA");
1766 __read_mostly bool numabalancing_enabled;
1768 void set_numabalancing_state(bool enabled)
1770 numabalancing_enabled = enabled;
1772 #endif /* CONFIG_SCHED_DEBUG */
1774 #ifdef CONFIG_PROC_SYSCTL
1775 int sysctl_numa_balancing(struct ctl_table *table, int write,
1776 void __user *buffer, size_t *lenp, loff_t *ppos)
1780 int state = numabalancing_enabled;
1782 if (write && !capable(CAP_SYS_ADMIN))
1787 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1791 set_numabalancing_state(state);
1798 * fork()/clone()-time setup:
1800 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1802 unsigned long flags;
1803 int cpu = get_cpu();
1805 __sched_fork(clone_flags, p);
1807 * We mark the process as running here. This guarantees that
1808 * nobody will actually run it, and a signal or other external
1809 * event cannot wake it up and insert it on the runqueue either.
1811 p->state = TASK_RUNNING;
1814 * Make sure we do not leak PI boosting priority to the child.
1816 p->prio = current->normal_prio;
1819 * Revert to default priority/policy on fork if requested.
1821 if (unlikely(p->sched_reset_on_fork)) {
1822 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1823 p->policy = SCHED_NORMAL;
1824 p->static_prio = NICE_TO_PRIO(0);
1826 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1827 p->static_prio = NICE_TO_PRIO(0);
1829 p->prio = p->normal_prio = __normal_prio(p);
1833 * We don't need the reset flag anymore after the fork. It has
1834 * fulfilled its duty:
1836 p->sched_reset_on_fork = 0;
1839 if (dl_prio(p->prio)) {
1842 } else if (rt_prio(p->prio)) {
1843 p->sched_class = &rt_sched_class;
1845 p->sched_class = &fair_sched_class;
1848 if (p->sched_class->task_fork)
1849 p->sched_class->task_fork(p);
1852 * The child is not yet in the pid-hash so no cgroup attach races,
1853 * and the cgroup is pinned to this child due to cgroup_fork()
1854 * is ran before sched_fork().
1856 * Silence PROVE_RCU.
1858 raw_spin_lock_irqsave(&p->pi_lock, flags);
1859 set_task_cpu(p, cpu);
1860 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1862 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1863 if (likely(sched_info_on()))
1864 memset(&p->sched_info, 0, sizeof(p->sched_info));
1866 #if defined(CONFIG_SMP)
1869 init_task_preempt_count(p);
1871 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1872 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1879 unsigned long to_ratio(u64 period, u64 runtime)
1881 if (runtime == RUNTIME_INF)
1885 * Doing this here saves a lot of checks in all
1886 * the calling paths, and returning zero seems
1887 * safe for them anyway.
1892 return div64_u64(runtime << 20, period);
1896 inline struct dl_bw *dl_bw_of(int i)
1898 return &cpu_rq(i)->rd->dl_bw;
1901 static inline int dl_bw_cpus(int i)
1903 struct root_domain *rd = cpu_rq(i)->rd;
1906 for_each_cpu_and(i, rd->span, cpu_active_mask)
1912 inline struct dl_bw *dl_bw_of(int i)
1914 return &cpu_rq(i)->dl.dl_bw;
1917 static inline int dl_bw_cpus(int i)
1924 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1926 dl_b->total_bw -= tsk_bw;
1930 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1932 dl_b->total_bw += tsk_bw;
1936 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1938 return dl_b->bw != -1 &&
1939 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1943 * We must be sure that accepting a new task (or allowing changing the
1944 * parameters of an existing one) is consistent with the bandwidth
1945 * constraints. If yes, this function also accordingly updates the currently
1946 * allocated bandwidth to reflect the new situation.
1948 * This function is called while holding p's rq->lock.
1950 static int dl_overflow(struct task_struct *p, int policy,
1951 const struct sched_attr *attr)
1954 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1955 u64 period = attr->sched_period ?: attr->sched_deadline;
1956 u64 runtime = attr->sched_runtime;
1957 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1960 if (new_bw == p->dl.dl_bw)
1964 * Either if a task, enters, leave, or stays -deadline but changes
1965 * its parameters, we may need to update accordingly the total
1966 * allocated bandwidth of the container.
1968 raw_spin_lock(&dl_b->lock);
1969 cpus = dl_bw_cpus(task_cpu(p));
1970 if (dl_policy(policy) && !task_has_dl_policy(p) &&
1971 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
1972 __dl_add(dl_b, new_bw);
1974 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
1975 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
1976 __dl_clear(dl_b, p->dl.dl_bw);
1977 __dl_add(dl_b, new_bw);
1979 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
1980 __dl_clear(dl_b, p->dl.dl_bw);
1983 raw_spin_unlock(&dl_b->lock);
1988 extern void init_dl_bw(struct dl_bw *dl_b);
1991 * wake_up_new_task - wake up a newly created task for the first time.
1993 * This function will do some initial scheduler statistics housekeeping
1994 * that must be done for every newly created context, then puts the task
1995 * on the runqueue and wakes it.
1997 void wake_up_new_task(struct task_struct *p)
1999 unsigned long flags;
2002 raw_spin_lock_irqsave(&p->pi_lock, flags);
2005 * Fork balancing, do it here and not earlier because:
2006 * - cpus_allowed can change in the fork path
2007 * - any previously selected cpu might disappear through hotplug
2009 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2012 /* Initialize new task's runnable average */
2013 init_task_runnable_average(p);
2014 rq = __task_rq_lock(p);
2015 activate_task(rq, p, 0);
2017 trace_sched_wakeup_new(p, true);
2018 check_preempt_curr(rq, p, WF_FORK);
2020 if (p->sched_class->task_woken)
2021 p->sched_class->task_woken(rq, p);
2023 task_rq_unlock(rq, p, &flags);
2026 #ifdef CONFIG_PREEMPT_NOTIFIERS
2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2030 * @notifier: notifier struct to register
2032 void preempt_notifier_register(struct preempt_notifier *notifier)
2034 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2036 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2039 * preempt_notifier_unregister - no longer interested in preemption notifications
2040 * @notifier: notifier struct to unregister
2042 * This is safe to call from within a preemption notifier.
2044 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2046 hlist_del(¬ifier->link);
2048 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2050 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2052 struct preempt_notifier *notifier;
2054 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2055 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2059 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2060 struct task_struct *next)
2062 struct preempt_notifier *notifier;
2064 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2065 notifier->ops->sched_out(notifier, next);
2068 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2070 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2075 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2076 struct task_struct *next)
2080 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2083 * prepare_task_switch - prepare to switch tasks
2084 * @rq: the runqueue preparing to switch
2085 * @prev: the current task that is being switched out
2086 * @next: the task we are going to switch to.
2088 * This is called with the rq lock held and interrupts off. It must
2089 * be paired with a subsequent finish_task_switch after the context
2092 * prepare_task_switch sets up locking and calls architecture specific
2096 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2097 struct task_struct *next)
2099 trace_sched_switch(prev, next);
2100 sched_info_switch(rq, prev, next);
2101 perf_event_task_sched_out(prev, next);
2102 fire_sched_out_preempt_notifiers(prev, next);
2103 prepare_lock_switch(rq, next);
2104 prepare_arch_switch(next);
2108 * finish_task_switch - clean up after a task-switch
2109 * @rq: runqueue associated with task-switch
2110 * @prev: the thread we just switched away from.
2112 * finish_task_switch must be called after the context switch, paired
2113 * with a prepare_task_switch call before the context switch.
2114 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2115 * and do any other architecture-specific cleanup actions.
2117 * Note that we may have delayed dropping an mm in context_switch(). If
2118 * so, we finish that here outside of the runqueue lock. (Doing it
2119 * with the lock held can cause deadlocks; see schedule() for
2122 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2123 __releases(rq->lock)
2125 struct mm_struct *mm = rq->prev_mm;
2131 * A task struct has one reference for the use as "current".
2132 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2133 * schedule one last time. The schedule call will never return, and
2134 * the scheduled task must drop that reference.
2135 * The test for TASK_DEAD must occur while the runqueue locks are
2136 * still held, otherwise prev could be scheduled on another cpu, die
2137 * there before we look at prev->state, and then the reference would
2139 * Manfred Spraul <manfred@colorfullife.com>
2141 prev_state = prev->state;
2142 vtime_task_switch(prev);
2143 finish_arch_switch(prev);
2144 perf_event_task_sched_in(prev, current);
2145 finish_lock_switch(rq, prev);
2146 finish_arch_post_lock_switch();
2148 fire_sched_in_preempt_notifiers(current);
2151 if (unlikely(prev_state == TASK_DEAD)) {
2152 task_numa_free(prev);
2154 if (prev->sched_class->task_dead)
2155 prev->sched_class->task_dead(prev);
2158 * Remove function-return probe instances associated with this
2159 * task and put them back on the free list.
2161 kprobe_flush_task(prev);
2162 put_task_struct(prev);
2165 tick_nohz_task_switch(current);
2170 /* assumes rq->lock is held */
2171 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2173 if (prev->sched_class->pre_schedule)
2174 prev->sched_class->pre_schedule(rq, prev);
2177 /* rq->lock is NOT held, but preemption is disabled */
2178 static inline void post_schedule(struct rq *rq)
2180 if (rq->post_schedule) {
2181 unsigned long flags;
2183 raw_spin_lock_irqsave(&rq->lock, flags);
2184 if (rq->curr->sched_class->post_schedule)
2185 rq->curr->sched_class->post_schedule(rq);
2186 raw_spin_unlock_irqrestore(&rq->lock, flags);
2188 rq->post_schedule = 0;
2194 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2198 static inline void post_schedule(struct rq *rq)
2205 * schedule_tail - first thing a freshly forked thread must call.
2206 * @prev: the thread we just switched away from.
2208 asmlinkage void schedule_tail(struct task_struct *prev)
2209 __releases(rq->lock)
2211 struct rq *rq = this_rq();
2213 finish_task_switch(rq, prev);
2216 * FIXME: do we need to worry about rq being invalidated by the
2221 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2222 /* In this case, finish_task_switch does not reenable preemption */
2225 if (current->set_child_tid)
2226 put_user(task_pid_vnr(current), current->set_child_tid);
2230 * context_switch - switch to the new MM and the new
2231 * thread's register state.
2234 context_switch(struct rq *rq, struct task_struct *prev,
2235 struct task_struct *next)
2237 struct mm_struct *mm, *oldmm;
2239 prepare_task_switch(rq, prev, next);
2242 oldmm = prev->active_mm;
2244 * For paravirt, this is coupled with an exit in switch_to to
2245 * combine the page table reload and the switch backend into
2248 arch_start_context_switch(prev);
2251 next->active_mm = oldmm;
2252 atomic_inc(&oldmm->mm_count);
2253 enter_lazy_tlb(oldmm, next);
2255 switch_mm(oldmm, mm, next);
2258 prev->active_mm = NULL;
2259 rq->prev_mm = oldmm;
2262 * Since the runqueue lock will be released by the next
2263 * task (which is an invalid locking op but in the case
2264 * of the scheduler it's an obvious special-case), so we
2265 * do an early lockdep release here:
2267 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2268 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2271 context_tracking_task_switch(prev, next);
2272 /* Here we just switch the register state and the stack. */
2273 switch_to(prev, next, prev);
2277 * this_rq must be evaluated again because prev may have moved
2278 * CPUs since it called schedule(), thus the 'rq' on its stack
2279 * frame will be invalid.
2281 finish_task_switch(this_rq(), prev);
2285 * nr_running and nr_context_switches:
2287 * externally visible scheduler statistics: current number of runnable
2288 * threads, total number of context switches performed since bootup.
2290 unsigned long nr_running(void)
2292 unsigned long i, sum = 0;
2294 for_each_online_cpu(i)
2295 sum += cpu_rq(i)->nr_running;
2300 unsigned long long nr_context_switches(void)
2303 unsigned long long sum = 0;
2305 for_each_possible_cpu(i)
2306 sum += cpu_rq(i)->nr_switches;
2311 unsigned long nr_iowait(void)
2313 unsigned long i, sum = 0;
2315 for_each_possible_cpu(i)
2316 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2321 unsigned long nr_iowait_cpu(int cpu)
2323 struct rq *this = cpu_rq(cpu);
2324 return atomic_read(&this->nr_iowait);
2330 * sched_exec - execve() is a valuable balancing opportunity, because at
2331 * this point the task has the smallest effective memory and cache footprint.
2333 void sched_exec(void)
2335 struct task_struct *p = current;
2336 unsigned long flags;
2339 raw_spin_lock_irqsave(&p->pi_lock, flags);
2340 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2341 if (dest_cpu == smp_processor_id())
2344 if (likely(cpu_active(dest_cpu))) {
2345 struct migration_arg arg = { p, dest_cpu };
2347 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2348 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2352 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2357 DEFINE_PER_CPU(struct kernel_stat, kstat);
2358 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2360 EXPORT_PER_CPU_SYMBOL(kstat);
2361 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2364 * Return any ns on the sched_clock that have not yet been accounted in
2365 * @p in case that task is currently running.
2367 * Called with task_rq_lock() held on @rq.
2369 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2373 if (task_current(rq, p)) {
2374 update_rq_clock(rq);
2375 ns = rq_clock_task(rq) - p->se.exec_start;
2383 unsigned long long task_delta_exec(struct task_struct *p)
2385 unsigned long flags;
2389 rq = task_rq_lock(p, &flags);
2390 ns = do_task_delta_exec(p, rq);
2391 task_rq_unlock(rq, p, &flags);
2397 * Return accounted runtime for the task.
2398 * In case the task is currently running, return the runtime plus current's
2399 * pending runtime that have not been accounted yet.
2401 unsigned long long task_sched_runtime(struct task_struct *p)
2403 unsigned long flags;
2407 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2409 * 64-bit doesn't need locks to atomically read a 64bit value.
2410 * So we have a optimization chance when the task's delta_exec is 0.
2411 * Reading ->on_cpu is racy, but this is ok.
2413 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2414 * If we race with it entering cpu, unaccounted time is 0. This is
2415 * indistinguishable from the read occurring a few cycles earlier.
2418 return p->se.sum_exec_runtime;
2421 rq = task_rq_lock(p, &flags);
2422 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2423 task_rq_unlock(rq, p, &flags);
2429 * This function gets called by the timer code, with HZ frequency.
2430 * We call it with interrupts disabled.
2432 void scheduler_tick(void)
2434 int cpu = smp_processor_id();
2435 struct rq *rq = cpu_rq(cpu);
2436 struct task_struct *curr = rq->curr;
2440 raw_spin_lock(&rq->lock);
2441 update_rq_clock(rq);
2442 curr->sched_class->task_tick(rq, curr, 0);
2443 update_cpu_load_active(rq);
2444 raw_spin_unlock(&rq->lock);
2446 perf_event_task_tick();
2449 rq->idle_balance = idle_cpu(cpu);
2450 trigger_load_balance(rq);
2452 rq_last_tick_reset(rq);
2455 #ifdef CONFIG_NO_HZ_FULL
2457 * scheduler_tick_max_deferment
2459 * Keep at least one tick per second when a single
2460 * active task is running because the scheduler doesn't
2461 * yet completely support full dynticks environment.
2463 * This makes sure that uptime, CFS vruntime, load
2464 * balancing, etc... continue to move forward, even
2465 * with a very low granularity.
2467 * Return: Maximum deferment in nanoseconds.
2469 u64 scheduler_tick_max_deferment(void)
2471 struct rq *rq = this_rq();
2472 unsigned long next, now = ACCESS_ONCE(jiffies);
2474 next = rq->last_sched_tick + HZ;
2476 if (time_before_eq(next, now))
2479 return jiffies_to_nsecs(next - now);
2483 notrace unsigned long get_parent_ip(unsigned long addr)
2485 if (in_lock_functions(addr)) {
2486 addr = CALLER_ADDR2;
2487 if (in_lock_functions(addr))
2488 addr = CALLER_ADDR3;
2493 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2494 defined(CONFIG_PREEMPT_TRACER))
2496 void __kprobes preempt_count_add(int val)
2498 #ifdef CONFIG_DEBUG_PREEMPT
2502 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2505 __preempt_count_add(val);
2506 #ifdef CONFIG_DEBUG_PREEMPT
2508 * Spinlock count overflowing soon?
2510 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2513 if (preempt_count() == val)
2514 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2516 EXPORT_SYMBOL(preempt_count_add);
2518 void __kprobes preempt_count_sub(int val)
2520 #ifdef CONFIG_DEBUG_PREEMPT
2524 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2527 * Is the spinlock portion underflowing?
2529 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2530 !(preempt_count() & PREEMPT_MASK)))
2534 if (preempt_count() == val)
2535 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2536 __preempt_count_sub(val);
2538 EXPORT_SYMBOL(preempt_count_sub);
2543 * Print scheduling while atomic bug:
2545 static noinline void __schedule_bug(struct task_struct *prev)
2547 if (oops_in_progress)
2550 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2551 prev->comm, prev->pid, preempt_count());
2553 debug_show_held_locks(prev);
2555 if (irqs_disabled())
2556 print_irqtrace_events(prev);
2558 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2562 * Various schedule()-time debugging checks and statistics:
2564 static inline void schedule_debug(struct task_struct *prev)
2567 * Test if we are atomic. Since do_exit() needs to call into
2568 * schedule() atomically, we ignore that path. Otherwise whine
2569 * if we are scheduling when we should not.
2571 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2572 __schedule_bug(prev);
2575 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2577 schedstat_inc(this_rq(), sched_count);
2580 static void put_prev_task(struct rq *rq, struct task_struct *prev)
2582 if (prev->on_rq || rq->skip_clock_update < 0)
2583 update_rq_clock(rq);
2584 prev->sched_class->put_prev_task(rq, prev);
2588 * Pick up the highest-prio task:
2590 static inline struct task_struct *
2591 pick_next_task(struct rq *rq)
2593 const struct sched_class *class;
2594 struct task_struct *p;
2597 * Optimization: we know that if all tasks are in
2598 * the fair class we can call that function directly:
2600 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
2601 p = fair_sched_class.pick_next_task(rq);
2606 for_each_class(class) {
2607 p = class->pick_next_task(rq);
2612 BUG(); /* the idle class will always have a runnable task */
2616 * __schedule() is the main scheduler function.
2618 * The main means of driving the scheduler and thus entering this function are:
2620 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2622 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2623 * paths. For example, see arch/x86/entry_64.S.
2625 * To drive preemption between tasks, the scheduler sets the flag in timer
2626 * interrupt handler scheduler_tick().
2628 * 3. Wakeups don't really cause entry into schedule(). They add a
2629 * task to the run-queue and that's it.
2631 * Now, if the new task added to the run-queue preempts the current
2632 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2633 * called on the nearest possible occasion:
2635 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2637 * - in syscall or exception context, at the next outmost
2638 * preempt_enable(). (this might be as soon as the wake_up()'s
2641 * - in IRQ context, return from interrupt-handler to
2642 * preemptible context
2644 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2647 * - cond_resched() call
2648 * - explicit schedule() call
2649 * - return from syscall or exception to user-space
2650 * - return from interrupt-handler to user-space
2652 static void __sched __schedule(void)
2654 struct task_struct *prev, *next;
2655 unsigned long *switch_count;
2661 cpu = smp_processor_id();
2663 rcu_note_context_switch(cpu);
2666 schedule_debug(prev);
2668 if (sched_feat(HRTICK))
2672 * Make sure that signal_pending_state()->signal_pending() below
2673 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2674 * done by the caller to avoid the race with signal_wake_up().
2676 smp_mb__before_spinlock();
2677 raw_spin_lock_irq(&rq->lock);
2679 switch_count = &prev->nivcsw;
2680 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2681 if (unlikely(signal_pending_state(prev->state, prev))) {
2682 prev->state = TASK_RUNNING;
2684 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2688 * If a worker went to sleep, notify and ask workqueue
2689 * whether it wants to wake up a task to maintain
2692 if (prev->flags & PF_WQ_WORKER) {
2693 struct task_struct *to_wakeup;
2695 to_wakeup = wq_worker_sleeping(prev, cpu);
2697 try_to_wake_up_local(to_wakeup);
2700 switch_count = &prev->nvcsw;
2703 pre_schedule(rq, prev);
2705 if (unlikely(!rq->nr_running))
2706 idle_balance(cpu, rq);
2708 put_prev_task(rq, prev);
2709 next = pick_next_task(rq);
2710 clear_tsk_need_resched(prev);
2711 clear_preempt_need_resched();
2712 rq->skip_clock_update = 0;
2714 if (likely(prev != next)) {
2719 context_switch(rq, prev, next); /* unlocks the rq */
2721 * The context switch have flipped the stack from under us
2722 * and restored the local variables which were saved when
2723 * this task called schedule() in the past. prev == current
2724 * is still correct, but it can be moved to another cpu/rq.
2726 cpu = smp_processor_id();
2729 raw_spin_unlock_irq(&rq->lock);
2733 sched_preempt_enable_no_resched();
2738 static inline void sched_submit_work(struct task_struct *tsk)
2740 if (!tsk->state || tsk_is_pi_blocked(tsk))
2743 * If we are going to sleep and we have plugged IO queued,
2744 * make sure to submit it to avoid deadlocks.
2746 if (blk_needs_flush_plug(tsk))
2747 blk_schedule_flush_plug(tsk);
2750 asmlinkage void __sched schedule(void)
2752 struct task_struct *tsk = current;
2754 sched_submit_work(tsk);
2757 EXPORT_SYMBOL(schedule);
2759 #ifdef CONFIG_CONTEXT_TRACKING
2760 asmlinkage void __sched schedule_user(void)
2763 * If we come here after a random call to set_need_resched(),
2764 * or we have been woken up remotely but the IPI has not yet arrived,
2765 * we haven't yet exited the RCU idle mode. Do it here manually until
2766 * we find a better solution.
2775 * schedule_preempt_disabled - called with preemption disabled
2777 * Returns with preemption disabled. Note: preempt_count must be 1
2779 void __sched schedule_preempt_disabled(void)
2781 sched_preempt_enable_no_resched();
2786 #ifdef CONFIG_PREEMPT
2788 * this is the entry point to schedule() from in-kernel preemption
2789 * off of preempt_enable. Kernel preemptions off return from interrupt
2790 * occur there and call schedule directly.
2792 asmlinkage void __sched notrace preempt_schedule(void)
2795 * If there is a non-zero preempt_count or interrupts are disabled,
2796 * we do not want to preempt the current task. Just return..
2798 if (likely(!preemptible()))
2802 __preempt_count_add(PREEMPT_ACTIVE);
2804 __preempt_count_sub(PREEMPT_ACTIVE);
2807 * Check again in case we missed a preemption opportunity
2808 * between schedule and now.
2811 } while (need_resched());
2813 EXPORT_SYMBOL(preempt_schedule);
2814 #endif /* CONFIG_PREEMPT */
2817 * this is the entry point to schedule() from kernel preemption
2818 * off of irq context.
2819 * Note, that this is called and return with irqs disabled. This will
2820 * protect us against recursive calling from irq.
2822 asmlinkage void __sched preempt_schedule_irq(void)
2824 enum ctx_state prev_state;
2826 /* Catch callers which need to be fixed */
2827 BUG_ON(preempt_count() || !irqs_disabled());
2829 prev_state = exception_enter();
2832 __preempt_count_add(PREEMPT_ACTIVE);
2835 local_irq_disable();
2836 __preempt_count_sub(PREEMPT_ACTIVE);
2839 * Check again in case we missed a preemption opportunity
2840 * between schedule and now.
2843 } while (need_resched());
2845 exception_exit(prev_state);
2848 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2851 return try_to_wake_up(curr->private, mode, wake_flags);
2853 EXPORT_SYMBOL(default_wake_function);
2856 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2858 unsigned long flags;
2861 init_waitqueue_entry(&wait, current);
2863 __set_current_state(state);
2865 spin_lock_irqsave(&q->lock, flags);
2866 __add_wait_queue(q, &wait);
2867 spin_unlock(&q->lock);
2868 timeout = schedule_timeout(timeout);
2869 spin_lock_irq(&q->lock);
2870 __remove_wait_queue(q, &wait);
2871 spin_unlock_irqrestore(&q->lock, flags);
2876 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2878 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2880 EXPORT_SYMBOL(interruptible_sleep_on);
2883 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2885 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2887 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2889 void __sched sleep_on(wait_queue_head_t *q)
2891 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2893 EXPORT_SYMBOL(sleep_on);
2895 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2897 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2899 EXPORT_SYMBOL(sleep_on_timeout);
2901 #ifdef CONFIG_RT_MUTEXES
2904 * rt_mutex_setprio - set the current priority of a task
2906 * @prio: prio value (kernel-internal form)
2908 * This function changes the 'effective' priority of a task. It does
2909 * not touch ->normal_prio like __setscheduler().
2911 * Used by the rt_mutex code to implement priority inheritance logic.
2913 void rt_mutex_setprio(struct task_struct *p, int prio)
2915 int oldprio, on_rq, running, enqueue_flag = 0;
2917 const struct sched_class *prev_class;
2919 BUG_ON(prio > MAX_PRIO);
2921 rq = __task_rq_lock(p);
2924 * Idle task boosting is a nono in general. There is one
2925 * exception, when PREEMPT_RT and NOHZ is active:
2927 * The idle task calls get_next_timer_interrupt() and holds
2928 * the timer wheel base->lock on the CPU and another CPU wants
2929 * to access the timer (probably to cancel it). We can safely
2930 * ignore the boosting request, as the idle CPU runs this code
2931 * with interrupts disabled and will complete the lock
2932 * protected section without being interrupted. So there is no
2933 * real need to boost.
2935 if (unlikely(p == rq->idle)) {
2936 WARN_ON(p != rq->curr);
2937 WARN_ON(p->pi_blocked_on);
2941 trace_sched_pi_setprio(p, prio);
2942 p->pi_top_task = rt_mutex_get_top_task(p);
2944 prev_class = p->sched_class;
2946 running = task_current(rq, p);
2948 dequeue_task(rq, p, 0);
2950 p->sched_class->put_prev_task(rq, p);
2953 * Boosting condition are:
2954 * 1. -rt task is running and holds mutex A
2955 * --> -dl task blocks on mutex A
2957 * 2. -dl task is running and holds mutex A
2958 * --> -dl task blocks on mutex A and could preempt the
2961 if (dl_prio(prio)) {
2962 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2963 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2964 p->dl.dl_boosted = 1;
2965 p->dl.dl_throttled = 0;
2966 enqueue_flag = ENQUEUE_REPLENISH;
2968 p->dl.dl_boosted = 0;
2969 p->sched_class = &dl_sched_class;
2970 } else if (rt_prio(prio)) {
2971 if (dl_prio(oldprio))
2972 p->dl.dl_boosted = 0;
2974 enqueue_flag = ENQUEUE_HEAD;
2975 p->sched_class = &rt_sched_class;
2977 if (dl_prio(oldprio))
2978 p->dl.dl_boosted = 0;
2979 p->sched_class = &fair_sched_class;
2985 p->sched_class->set_curr_task(rq);
2987 enqueue_task(rq, p, enqueue_flag);
2989 check_class_changed(rq, p, prev_class, oldprio);
2991 __task_rq_unlock(rq);
2995 void set_user_nice(struct task_struct *p, long nice)
2997 int old_prio, delta, on_rq;
2998 unsigned long flags;
3001 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
3004 * We have to be careful, if called from sys_setpriority(),
3005 * the task might be in the middle of scheduling on another CPU.
3007 rq = task_rq_lock(p, &flags);
3009 * The RT priorities are set via sched_setscheduler(), but we still
3010 * allow the 'normal' nice value to be set - but as expected
3011 * it wont have any effect on scheduling until the task is
3012 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3014 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3015 p->static_prio = NICE_TO_PRIO(nice);
3020 dequeue_task(rq, p, 0);
3022 p->static_prio = NICE_TO_PRIO(nice);
3025 p->prio = effective_prio(p);
3026 delta = p->prio - old_prio;
3029 enqueue_task(rq, p, 0);
3031 * If the task increased its priority or is running and
3032 * lowered its priority, then reschedule its CPU:
3034 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3035 resched_task(rq->curr);
3038 task_rq_unlock(rq, p, &flags);
3040 EXPORT_SYMBOL(set_user_nice);
3043 * can_nice - check if a task can reduce its nice value
3047 int can_nice(const struct task_struct *p, const int nice)
3049 /* convert nice value [19,-20] to rlimit style value [1,40] */
3050 int nice_rlim = 20 - nice;
3052 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3053 capable(CAP_SYS_NICE));
3056 #ifdef __ARCH_WANT_SYS_NICE
3059 * sys_nice - change the priority of the current process.
3060 * @increment: priority increment
3062 * sys_setpriority is a more generic, but much slower function that
3063 * does similar things.
3065 SYSCALL_DEFINE1(nice, int, increment)
3070 * Setpriority might change our priority at the same moment.
3071 * We don't have to worry. Conceptually one call occurs first
3072 * and we have a single winner.
3074 if (increment < -40)
3079 nice = TASK_NICE(current) + increment;
3085 if (increment < 0 && !can_nice(current, nice))
3088 retval = security_task_setnice(current, nice);
3092 set_user_nice(current, nice);
3099 * task_prio - return the priority value of a given task.
3100 * @p: the task in question.
3102 * Return: The priority value as seen by users in /proc.
3103 * RT tasks are offset by -200. Normal tasks are centered
3104 * around 0, value goes from -16 to +15.
3106 int task_prio(const struct task_struct *p)
3108 return p->prio - MAX_RT_PRIO;
3112 * task_nice - return the nice value of a given task.
3113 * @p: the task in question.
3115 * Return: The nice value [ -20 ... 0 ... 19 ].
3117 int task_nice(const struct task_struct *p)
3119 return TASK_NICE(p);
3121 EXPORT_SYMBOL(task_nice);
3124 * idle_cpu - is a given cpu idle currently?
3125 * @cpu: the processor in question.
3127 * Return: 1 if the CPU is currently idle. 0 otherwise.
3129 int idle_cpu(int cpu)
3131 struct rq *rq = cpu_rq(cpu);
3133 if (rq->curr != rq->idle)
3140 if (!llist_empty(&rq->wake_list))
3148 * idle_task - return the idle task for a given cpu.
3149 * @cpu: the processor in question.
3151 * Return: The idle task for the cpu @cpu.
3153 struct task_struct *idle_task(int cpu)
3155 return cpu_rq(cpu)->idle;
3159 * find_process_by_pid - find a process with a matching PID value.
3160 * @pid: the pid in question.
3162 * The task of @pid, if found. %NULL otherwise.
3164 static struct task_struct *find_process_by_pid(pid_t pid)
3166 return pid ? find_task_by_vpid(pid) : current;
3170 * This function initializes the sched_dl_entity of a newly becoming
3171 * SCHED_DEADLINE task.
3173 * Only the static values are considered here, the actual runtime and the
3174 * absolute deadline will be properly calculated when the task is enqueued
3175 * for the first time with its new policy.
3178 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3180 struct sched_dl_entity *dl_se = &p->dl;
3182 init_dl_task_timer(dl_se);
3183 dl_se->dl_runtime = attr->sched_runtime;
3184 dl_se->dl_deadline = attr->sched_deadline;
3185 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3186 dl_se->flags = attr->sched_flags;
3187 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3188 dl_se->dl_throttled = 0;
3192 /* Actually do priority change: must hold pi & rq lock. */
3193 static void __setscheduler(struct rq *rq, struct task_struct *p,
3194 const struct sched_attr *attr)
3196 int policy = attr->sched_policy;
3198 if (policy == -1) /* setparam */
3203 if (dl_policy(policy))
3204 __setparam_dl(p, attr);
3205 else if (fair_policy(policy))
3206 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3209 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3210 * !rt_policy. Always setting this ensures that things like
3211 * getparam()/getattr() don't report silly values for !rt tasks.
3213 p->rt_priority = attr->sched_priority;
3215 p->normal_prio = normal_prio(p);
3216 p->prio = rt_mutex_getprio(p);
3218 if (dl_prio(p->prio))
3219 p->sched_class = &dl_sched_class;
3220 else if (rt_prio(p->prio))
3221 p->sched_class = &rt_sched_class;
3223 p->sched_class = &fair_sched_class;
3229 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3231 struct sched_dl_entity *dl_se = &p->dl;
3233 attr->sched_priority = p->rt_priority;
3234 attr->sched_runtime = dl_se->dl_runtime;
3235 attr->sched_deadline = dl_se->dl_deadline;
3236 attr->sched_period = dl_se->dl_period;
3237 attr->sched_flags = dl_se->flags;
3241 * This function validates the new parameters of a -deadline task.
3242 * We ask for the deadline not being zero, and greater or equal
3243 * than the runtime, as well as the period of being zero or
3244 * greater than deadline. Furthermore, we have to be sure that
3245 * user parameters are above the internal resolution (1us); we
3246 * check sched_runtime only since it is always the smaller one.
3249 __checkparam_dl(const struct sched_attr *attr)
3251 return attr && attr->sched_deadline != 0 &&
3252 (attr->sched_period == 0 ||
3253 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3254 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3255 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3259 * check the target process has a UID that matches the current process's
3261 static bool check_same_owner(struct task_struct *p)
3263 const struct cred *cred = current_cred(), *pcred;
3267 pcred = __task_cred(p);
3268 match = (uid_eq(cred->euid, pcred->euid) ||
3269 uid_eq(cred->euid, pcred->uid));
3274 static int __sched_setscheduler(struct task_struct *p,
3275 const struct sched_attr *attr,
3278 int retval, oldprio, oldpolicy = -1, on_rq, running;
3279 int policy = attr->sched_policy;
3280 unsigned long flags;
3281 const struct sched_class *prev_class;
3285 /* may grab non-irq protected spin_locks */
3286 BUG_ON(in_interrupt());
3288 /* double check policy once rq lock held */
3290 reset_on_fork = p->sched_reset_on_fork;
3291 policy = oldpolicy = p->policy;
3293 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3295 if (policy != SCHED_DEADLINE &&
3296 policy != SCHED_FIFO && policy != SCHED_RR &&
3297 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3298 policy != SCHED_IDLE)
3302 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3306 * Valid priorities for SCHED_FIFO and SCHED_RR are
3307 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3308 * SCHED_BATCH and SCHED_IDLE is 0.
3310 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3311 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3313 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3314 (rt_policy(policy) != (attr->sched_priority != 0)))
3318 * Allow unprivileged RT tasks to decrease priority:
3320 if (user && !capable(CAP_SYS_NICE)) {
3321 if (fair_policy(policy)) {
3322 if (attr->sched_nice < TASK_NICE(p) &&
3323 !can_nice(p, attr->sched_nice))
3327 if (rt_policy(policy)) {
3328 unsigned long rlim_rtprio =
3329 task_rlimit(p, RLIMIT_RTPRIO);
3331 /* can't set/change the rt policy */
3332 if (policy != p->policy && !rlim_rtprio)
3335 /* can't increase priority */
3336 if (attr->sched_priority > p->rt_priority &&
3337 attr->sched_priority > rlim_rtprio)
3342 * Can't set/change SCHED_DEADLINE policy at all for now
3343 * (safest behavior); in the future we would like to allow
3344 * unprivileged DL tasks to increase their relative deadline
3345 * or reduce their runtime (both ways reducing utilization)
3347 if (dl_policy(policy))
3351 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3352 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3354 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3355 if (!can_nice(p, TASK_NICE(p)))
3359 /* can't change other user's priorities */
3360 if (!check_same_owner(p))
3363 /* Normal users shall not reset the sched_reset_on_fork flag */
3364 if (p->sched_reset_on_fork && !reset_on_fork)
3369 retval = security_task_setscheduler(p);
3375 * make sure no PI-waiters arrive (or leave) while we are
3376 * changing the priority of the task:
3378 * To be able to change p->policy safely, the appropriate
3379 * runqueue lock must be held.
3381 rq = task_rq_lock(p, &flags);
3384 * Changing the policy of the stop threads its a very bad idea
3386 if (p == rq->stop) {
3387 task_rq_unlock(rq, p, &flags);
3392 * If not changing anything there's no need to proceed further:
3394 if (unlikely(policy == p->policy)) {
3395 if (fair_policy(policy) && attr->sched_nice != TASK_NICE(p))
3397 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3399 if (dl_policy(policy))
3402 task_rq_unlock(rq, p, &flags);
3408 #ifdef CONFIG_RT_GROUP_SCHED
3410 * Do not allow realtime tasks into groups that have no runtime
3413 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3414 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3415 !task_group_is_autogroup(task_group(p))) {
3416 task_rq_unlock(rq, p, &flags);
3421 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3422 cpumask_t *span = rq->rd->span;
3425 * Don't allow tasks with an affinity mask smaller than
3426 * the entire root_domain to become SCHED_DEADLINE. We
3427 * will also fail if there's no bandwidth available.
3429 if (!cpumask_subset(span, &p->cpus_allowed) ||
3430 rq->rd->dl_bw.bw == 0) {
3431 task_rq_unlock(rq, p, &flags);
3438 /* recheck policy now with rq lock held */
3439 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3440 policy = oldpolicy = -1;
3441 task_rq_unlock(rq, p, &flags);
3446 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3447 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3450 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3451 task_rq_unlock(rq, p, &flags);
3456 running = task_current(rq, p);
3458 dequeue_task(rq, p, 0);
3460 p->sched_class->put_prev_task(rq, p);
3462 p->sched_reset_on_fork = reset_on_fork;
3465 prev_class = p->sched_class;
3466 __setscheduler(rq, p, attr);
3469 p->sched_class->set_curr_task(rq);
3471 enqueue_task(rq, p, 0);
3473 check_class_changed(rq, p, prev_class, oldprio);
3474 task_rq_unlock(rq, p, &flags);
3476 rt_mutex_adjust_pi(p);
3481 static int _sched_setscheduler(struct task_struct *p, int policy,
3482 const struct sched_param *param, bool check)
3484 struct sched_attr attr = {
3485 .sched_policy = policy,
3486 .sched_priority = param->sched_priority,
3487 .sched_nice = PRIO_TO_NICE(p->static_prio),
3491 * Fixup the legacy SCHED_RESET_ON_FORK hack
3493 if (policy & SCHED_RESET_ON_FORK) {
3494 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3495 policy &= ~SCHED_RESET_ON_FORK;
3496 attr.sched_policy = policy;
3499 return __sched_setscheduler(p, &attr, check);
3502 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3503 * @p: the task in question.
3504 * @policy: new policy.
3505 * @param: structure containing the new RT priority.
3507 * Return: 0 on success. An error code otherwise.
3509 * NOTE that the task may be already dead.
3511 int sched_setscheduler(struct task_struct *p, int policy,
3512 const struct sched_param *param)
3514 return _sched_setscheduler(p, policy, param, true);
3516 EXPORT_SYMBOL_GPL(sched_setscheduler);
3518 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3520 return __sched_setscheduler(p, attr, true);
3522 EXPORT_SYMBOL_GPL(sched_setattr);
3525 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3526 * @p: the task in question.
3527 * @policy: new policy.
3528 * @param: structure containing the new RT priority.
3530 * Just like sched_setscheduler, only don't bother checking if the
3531 * current context has permission. For example, this is needed in
3532 * stop_machine(): we create temporary high priority worker threads,
3533 * but our caller might not have that capability.
3535 * Return: 0 on success. An error code otherwise.
3537 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3538 const struct sched_param *param)
3540 return _sched_setscheduler(p, policy, param, false);
3544 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3546 struct sched_param lparam;
3547 struct task_struct *p;
3550 if (!param || pid < 0)
3552 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3557 p = find_process_by_pid(pid);
3559 retval = sched_setscheduler(p, policy, &lparam);
3566 * Mimics kernel/events/core.c perf_copy_attr().
3568 static int sched_copy_attr(struct sched_attr __user *uattr,
3569 struct sched_attr *attr)
3574 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3578 * zero the full structure, so that a short copy will be nice.
3580 memset(attr, 0, sizeof(*attr));
3582 ret = get_user(size, &uattr->size);
3586 if (size > PAGE_SIZE) /* silly large */
3589 if (!size) /* abi compat */
3590 size = SCHED_ATTR_SIZE_VER0;
3592 if (size < SCHED_ATTR_SIZE_VER0)
3596 * If we're handed a bigger struct than we know of,
3597 * ensure all the unknown bits are 0 - i.e. new
3598 * user-space does not rely on any kernel feature
3599 * extensions we dont know about yet.
3601 if (size > sizeof(*attr)) {
3602 unsigned char __user *addr;
3603 unsigned char __user *end;
3606 addr = (void __user *)uattr + sizeof(*attr);
3607 end = (void __user *)uattr + size;
3609 for (; addr < end; addr++) {
3610 ret = get_user(val, addr);
3616 size = sizeof(*attr);
3619 ret = copy_from_user(attr, uattr, size);
3624 * XXX: do we want to be lenient like existing syscalls; or do we want
3625 * to be strict and return an error on out-of-bounds values?
3627 attr->sched_nice = clamp(attr->sched_nice, -20, 19);
3633 put_user(sizeof(*attr), &uattr->size);
3639 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3640 * @pid: the pid in question.
3641 * @policy: new policy.
3642 * @param: structure containing the new RT priority.
3644 * Return: 0 on success. An error code otherwise.
3646 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3647 struct sched_param __user *, param)
3649 /* negative values for policy are not valid */
3653 return do_sched_setscheduler(pid, policy, param);
3657 * sys_sched_setparam - set/change the RT priority of a thread
3658 * @pid: the pid in question.
3659 * @param: structure containing the new RT priority.
3661 * Return: 0 on success. An error code otherwise.
3663 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3665 return do_sched_setscheduler(pid, -1, param);
3669 * sys_sched_setattr - same as above, but with extended sched_attr
3670 * @pid: the pid in question.
3671 * @uattr: structure containing the extended parameters.
3673 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3674 unsigned int, flags)
3676 struct sched_attr attr;
3677 struct task_struct *p;
3680 if (!uattr || pid < 0 || flags)
3683 retval = sched_copy_attr(uattr, &attr);
3687 if (attr.sched_policy < 0)
3692 p = find_process_by_pid(pid);
3694 retval = sched_setattr(p, &attr);
3701 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3702 * @pid: the pid in question.
3704 * Return: On success, the policy of the thread. Otherwise, a negative error
3707 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3709 struct task_struct *p;
3717 p = find_process_by_pid(pid);
3719 retval = security_task_getscheduler(p);
3722 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3729 * sys_sched_getparam - get the RT priority of a thread
3730 * @pid: the pid in question.
3731 * @param: structure containing the RT priority.
3733 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3736 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3738 struct sched_param lp;
3739 struct task_struct *p;
3742 if (!param || pid < 0)
3746 p = find_process_by_pid(pid);
3751 retval = security_task_getscheduler(p);
3755 if (task_has_dl_policy(p)) {
3759 lp.sched_priority = p->rt_priority;
3763 * This one might sleep, we cannot do it with a spinlock held ...
3765 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3774 static int sched_read_attr(struct sched_attr __user *uattr,
3775 struct sched_attr *attr,
3780 if (!access_ok(VERIFY_WRITE, uattr, usize))
3784 * If we're handed a smaller struct than we know of,
3785 * ensure all the unknown bits are 0 - i.e. old
3786 * user-space does not get uncomplete information.
3788 if (usize < sizeof(*attr)) {
3789 unsigned char *addr;
3792 addr = (void *)attr + usize;
3793 end = (void *)attr + sizeof(*attr);
3795 for (; addr < end; addr++) {
3803 ret = copy_to_user(uattr, attr, attr->size);
3816 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3817 * @pid: the pid in question.
3818 * @uattr: structure containing the extended parameters.
3819 * @size: sizeof(attr) for fwd/bwd comp.
3821 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3822 unsigned int, size, unsigned int, flags)
3824 struct sched_attr attr = {
3825 .size = sizeof(struct sched_attr),
3827 struct task_struct *p;
3830 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3831 size < SCHED_ATTR_SIZE_VER0 || flags)
3835 p = find_process_by_pid(pid);
3840 retval = security_task_getscheduler(p);
3844 attr.sched_policy = p->policy;
3845 if (p->sched_reset_on_fork)
3846 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3847 if (task_has_dl_policy(p))
3848 __getparam_dl(p, &attr);
3849 else if (task_has_rt_policy(p))
3850 attr.sched_priority = p->rt_priority;
3852 attr.sched_nice = TASK_NICE(p);
3856 retval = sched_read_attr(uattr, &attr, size);
3864 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3866 cpumask_var_t cpus_allowed, new_mask;
3867 struct task_struct *p;
3872 p = find_process_by_pid(pid);
3878 /* Prevent p going away */
3882 if (p->flags & PF_NO_SETAFFINITY) {
3886 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3890 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3892 goto out_free_cpus_allowed;
3895 if (!check_same_owner(p)) {
3897 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3904 retval = security_task_setscheduler(p);
3909 cpuset_cpus_allowed(p, cpus_allowed);
3910 cpumask_and(new_mask, in_mask, cpus_allowed);
3913 * Since bandwidth control happens on root_domain basis,
3914 * if admission test is enabled, we only admit -deadline
3915 * tasks allowed to run on all the CPUs in the task's
3919 if (task_has_dl_policy(p)) {
3920 const struct cpumask *span = task_rq(p)->rd->span;
3922 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3929 retval = set_cpus_allowed_ptr(p, new_mask);
3932 cpuset_cpus_allowed(p, cpus_allowed);
3933 if (!cpumask_subset(new_mask, cpus_allowed)) {
3935 * We must have raced with a concurrent cpuset
3936 * update. Just reset the cpus_allowed to the
3937 * cpuset's cpus_allowed
3939 cpumask_copy(new_mask, cpus_allowed);
3944 free_cpumask_var(new_mask);
3945 out_free_cpus_allowed:
3946 free_cpumask_var(cpus_allowed);
3952 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3953 struct cpumask *new_mask)
3955 if (len < cpumask_size())
3956 cpumask_clear(new_mask);
3957 else if (len > cpumask_size())
3958 len = cpumask_size();
3960 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3964 * sys_sched_setaffinity - set the cpu affinity of a process
3965 * @pid: pid of the process
3966 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3967 * @user_mask_ptr: user-space pointer to the new cpu mask
3969 * Return: 0 on success. An error code otherwise.
3971 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3972 unsigned long __user *, user_mask_ptr)
3974 cpumask_var_t new_mask;
3977 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3980 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3982 retval = sched_setaffinity(pid, new_mask);
3983 free_cpumask_var(new_mask);
3987 long sched_getaffinity(pid_t pid, struct cpumask *mask)
3989 struct task_struct *p;
3990 unsigned long flags;
3996 p = find_process_by_pid(pid);
4000 retval = security_task_getscheduler(p);
4004 raw_spin_lock_irqsave(&p->pi_lock, flags);
4005 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4006 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4015 * sys_sched_getaffinity - get the cpu affinity of a process
4016 * @pid: pid of the process
4017 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4018 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4020 * Return: 0 on success. An error code otherwise.
4022 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4023 unsigned long __user *, user_mask_ptr)
4028 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4030 if (len & (sizeof(unsigned long)-1))
4033 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4036 ret = sched_getaffinity(pid, mask);
4038 size_t retlen = min_t(size_t, len, cpumask_size());
4040 if (copy_to_user(user_mask_ptr, mask, retlen))
4045 free_cpumask_var(mask);
4051 * sys_sched_yield - yield the current processor to other threads.
4053 * This function yields the current CPU to other tasks. If there are no
4054 * other threads running on this CPU then this function will return.
4058 SYSCALL_DEFINE0(sched_yield)
4060 struct rq *rq = this_rq_lock();
4062 schedstat_inc(rq, yld_count);
4063 current->sched_class->yield_task(rq);
4066 * Since we are going to call schedule() anyway, there's
4067 * no need to preempt or enable interrupts:
4069 __release(rq->lock);
4070 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4071 do_raw_spin_unlock(&rq->lock);
4072 sched_preempt_enable_no_resched();
4079 static void __cond_resched(void)
4081 __preempt_count_add(PREEMPT_ACTIVE);
4083 __preempt_count_sub(PREEMPT_ACTIVE);
4086 int __sched _cond_resched(void)
4088 if (should_resched()) {
4094 EXPORT_SYMBOL(_cond_resched);
4097 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4098 * call schedule, and on return reacquire the lock.
4100 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4101 * operations here to prevent schedule() from being called twice (once via
4102 * spin_unlock(), once by hand).
4104 int __cond_resched_lock(spinlock_t *lock)
4106 int resched = should_resched();
4109 lockdep_assert_held(lock);
4111 if (spin_needbreak(lock) || resched) {
4122 EXPORT_SYMBOL(__cond_resched_lock);
4124 int __sched __cond_resched_softirq(void)
4126 BUG_ON(!in_softirq());
4128 if (should_resched()) {
4136 EXPORT_SYMBOL(__cond_resched_softirq);
4139 * yield - yield the current processor to other threads.
4141 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4143 * The scheduler is at all times free to pick the calling task as the most
4144 * eligible task to run, if removing the yield() call from your code breaks
4145 * it, its already broken.
4147 * Typical broken usage is:
4152 * where one assumes that yield() will let 'the other' process run that will
4153 * make event true. If the current task is a SCHED_FIFO task that will never
4154 * happen. Never use yield() as a progress guarantee!!
4156 * If you want to use yield() to wait for something, use wait_event().
4157 * If you want to use yield() to be 'nice' for others, use cond_resched().
4158 * If you still want to use yield(), do not!
4160 void __sched yield(void)
4162 set_current_state(TASK_RUNNING);
4165 EXPORT_SYMBOL(yield);
4168 * yield_to - yield the current processor to another thread in
4169 * your thread group, or accelerate that thread toward the
4170 * processor it's on.
4172 * @preempt: whether task preemption is allowed or not
4174 * It's the caller's job to ensure that the target task struct
4175 * can't go away on us before we can do any checks.
4178 * true (>0) if we indeed boosted the target task.
4179 * false (0) if we failed to boost the target.
4180 * -ESRCH if there's no task to yield to.
4182 bool __sched yield_to(struct task_struct *p, bool preempt)
4184 struct task_struct *curr = current;
4185 struct rq *rq, *p_rq;
4186 unsigned long flags;
4189 local_irq_save(flags);
4195 * If we're the only runnable task on the rq and target rq also
4196 * has only one task, there's absolutely no point in yielding.
4198 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4203 double_rq_lock(rq, p_rq);
4204 if (task_rq(p) != p_rq) {
4205 double_rq_unlock(rq, p_rq);
4209 if (!curr->sched_class->yield_to_task)
4212 if (curr->sched_class != p->sched_class)
4215 if (task_running(p_rq, p) || p->state)
4218 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4220 schedstat_inc(rq, yld_count);
4222 * Make p's CPU reschedule; pick_next_entity takes care of
4225 if (preempt && rq != p_rq)
4226 resched_task(p_rq->curr);
4230 double_rq_unlock(rq, p_rq);
4232 local_irq_restore(flags);
4239 EXPORT_SYMBOL_GPL(yield_to);
4242 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4243 * that process accounting knows that this is a task in IO wait state.
4245 void __sched io_schedule(void)
4247 struct rq *rq = raw_rq();
4249 delayacct_blkio_start();
4250 atomic_inc(&rq->nr_iowait);
4251 blk_flush_plug(current);
4252 current->in_iowait = 1;
4254 current->in_iowait = 0;
4255 atomic_dec(&rq->nr_iowait);
4256 delayacct_blkio_end();
4258 EXPORT_SYMBOL(io_schedule);
4260 long __sched io_schedule_timeout(long timeout)
4262 struct rq *rq = raw_rq();
4265 delayacct_blkio_start();
4266 atomic_inc(&rq->nr_iowait);
4267 blk_flush_plug(current);
4268 current->in_iowait = 1;
4269 ret = schedule_timeout(timeout);
4270 current->in_iowait = 0;
4271 atomic_dec(&rq->nr_iowait);
4272 delayacct_blkio_end();
4277 * sys_sched_get_priority_max - return maximum RT priority.
4278 * @policy: scheduling class.
4280 * Return: On success, this syscall returns the maximum
4281 * rt_priority that can be used by a given scheduling class.
4282 * On failure, a negative error code is returned.
4284 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4291 ret = MAX_USER_RT_PRIO-1;
4293 case SCHED_DEADLINE:
4304 * sys_sched_get_priority_min - return minimum RT priority.
4305 * @policy: scheduling class.
4307 * Return: On success, this syscall returns the minimum
4308 * rt_priority that can be used by a given scheduling class.
4309 * On failure, a negative error code is returned.
4311 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4320 case SCHED_DEADLINE:
4330 * sys_sched_rr_get_interval - return the default timeslice of a process.
4331 * @pid: pid of the process.
4332 * @interval: userspace pointer to the timeslice value.
4334 * this syscall writes the default timeslice value of a given process
4335 * into the user-space timespec buffer. A value of '0' means infinity.
4337 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4340 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4341 struct timespec __user *, interval)
4343 struct task_struct *p;
4344 unsigned int time_slice;
4345 unsigned long flags;
4355 p = find_process_by_pid(pid);
4359 retval = security_task_getscheduler(p);
4363 rq = task_rq_lock(p, &flags);
4365 if (p->sched_class->get_rr_interval)
4366 time_slice = p->sched_class->get_rr_interval(rq, p);
4367 task_rq_unlock(rq, p, &flags);
4370 jiffies_to_timespec(time_slice, &t);
4371 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4379 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4381 void sched_show_task(struct task_struct *p)
4383 unsigned long free = 0;
4387 state = p->state ? __ffs(p->state) + 1 : 0;
4388 printk(KERN_INFO "%-15.15s %c", p->comm,
4389 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4390 #if BITS_PER_LONG == 32
4391 if (state == TASK_RUNNING)
4392 printk(KERN_CONT " running ");
4394 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4396 if (state == TASK_RUNNING)
4397 printk(KERN_CONT " running task ");
4399 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4401 #ifdef CONFIG_DEBUG_STACK_USAGE
4402 free = stack_not_used(p);
4405 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4407 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4408 task_pid_nr(p), ppid,
4409 (unsigned long)task_thread_info(p)->flags);
4411 print_worker_info(KERN_INFO, p);
4412 show_stack(p, NULL);
4415 void show_state_filter(unsigned long state_filter)
4417 struct task_struct *g, *p;
4419 #if BITS_PER_LONG == 32
4421 " task PC stack pid father\n");
4424 " task PC stack pid father\n");
4427 do_each_thread(g, p) {
4429 * reset the NMI-timeout, listing all files on a slow
4430 * console might take a lot of time:
4432 touch_nmi_watchdog();
4433 if (!state_filter || (p->state & state_filter))
4435 } while_each_thread(g, p);
4437 touch_all_softlockup_watchdogs();
4439 #ifdef CONFIG_SCHED_DEBUG
4440 sysrq_sched_debug_show();
4444 * Only show locks if all tasks are dumped:
4447 debug_show_all_locks();
4450 void init_idle_bootup_task(struct task_struct *idle)
4452 idle->sched_class = &idle_sched_class;
4456 * init_idle - set up an idle thread for a given CPU
4457 * @idle: task in question
4458 * @cpu: cpu the idle task belongs to
4460 * NOTE: this function does not set the idle thread's NEED_RESCHED
4461 * flag, to make booting more robust.
4463 void init_idle(struct task_struct *idle, int cpu)
4465 struct rq *rq = cpu_rq(cpu);
4466 unsigned long flags;
4468 raw_spin_lock_irqsave(&rq->lock, flags);
4470 __sched_fork(0, idle);
4471 idle->state = TASK_RUNNING;
4472 idle->se.exec_start = sched_clock();
4474 do_set_cpus_allowed(idle, cpumask_of(cpu));
4476 * We're having a chicken and egg problem, even though we are
4477 * holding rq->lock, the cpu isn't yet set to this cpu so the
4478 * lockdep check in task_group() will fail.
4480 * Similar case to sched_fork(). / Alternatively we could
4481 * use task_rq_lock() here and obtain the other rq->lock.
4486 __set_task_cpu(idle, cpu);
4489 rq->curr = rq->idle = idle;
4490 #if defined(CONFIG_SMP)
4493 raw_spin_unlock_irqrestore(&rq->lock, flags);
4495 /* Set the preempt count _outside_ the spinlocks! */
4496 init_idle_preempt_count(idle, cpu);
4499 * The idle tasks have their own, simple scheduling class:
4501 idle->sched_class = &idle_sched_class;
4502 ftrace_graph_init_idle_task(idle, cpu);
4503 vtime_init_idle(idle, cpu);
4504 #if defined(CONFIG_SMP)
4505 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4510 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4512 if (p->sched_class && p->sched_class->set_cpus_allowed)
4513 p->sched_class->set_cpus_allowed(p, new_mask);
4515 cpumask_copy(&p->cpus_allowed, new_mask);
4516 p->nr_cpus_allowed = cpumask_weight(new_mask);
4520 * This is how migration works:
4522 * 1) we invoke migration_cpu_stop() on the target CPU using
4524 * 2) stopper starts to run (implicitly forcing the migrated thread
4526 * 3) it checks whether the migrated task is still in the wrong runqueue.
4527 * 4) if it's in the wrong runqueue then the migration thread removes
4528 * it and puts it into the right queue.
4529 * 5) stopper completes and stop_one_cpu() returns and the migration
4534 * Change a given task's CPU affinity. Migrate the thread to a
4535 * proper CPU and schedule it away if the CPU it's executing on
4536 * is removed from the allowed bitmask.
4538 * NOTE: the caller must have a valid reference to the task, the
4539 * task must not exit() & deallocate itself prematurely. The
4540 * call is not atomic; no spinlocks may be held.
4542 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4544 unsigned long flags;
4546 unsigned int dest_cpu;
4549 rq = task_rq_lock(p, &flags);
4551 if (cpumask_equal(&p->cpus_allowed, new_mask))
4554 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4559 do_set_cpus_allowed(p, new_mask);
4561 /* Can the task run on the task's current CPU? If so, we're done */
4562 if (cpumask_test_cpu(task_cpu(p), new_mask))
4565 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4567 struct migration_arg arg = { p, dest_cpu };
4568 /* Need help from migration thread: drop lock and wait. */
4569 task_rq_unlock(rq, p, &flags);
4570 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4571 tlb_migrate_finish(p->mm);
4575 task_rq_unlock(rq, p, &flags);
4579 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4582 * Move (not current) task off this cpu, onto dest cpu. We're doing
4583 * this because either it can't run here any more (set_cpus_allowed()
4584 * away from this CPU, or CPU going down), or because we're
4585 * attempting to rebalance this task on exec (sched_exec).
4587 * So we race with normal scheduler movements, but that's OK, as long
4588 * as the task is no longer on this CPU.
4590 * Returns non-zero if task was successfully migrated.
4592 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4594 struct rq *rq_dest, *rq_src;
4597 if (unlikely(!cpu_active(dest_cpu)))
4600 rq_src = cpu_rq(src_cpu);
4601 rq_dest = cpu_rq(dest_cpu);
4603 raw_spin_lock(&p->pi_lock);
4604 double_rq_lock(rq_src, rq_dest);
4605 /* Already moved. */
4606 if (task_cpu(p) != src_cpu)
4608 /* Affinity changed (again). */
4609 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4613 * If we're not on a rq, the next wake-up will ensure we're
4617 dequeue_task(rq_src, p, 0);
4618 set_task_cpu(p, dest_cpu);
4619 enqueue_task(rq_dest, p, 0);
4620 check_preempt_curr(rq_dest, p, 0);
4625 double_rq_unlock(rq_src, rq_dest);
4626 raw_spin_unlock(&p->pi_lock);
4630 #ifdef CONFIG_NUMA_BALANCING
4631 /* Migrate current task p to target_cpu */
4632 int migrate_task_to(struct task_struct *p, int target_cpu)
4634 struct migration_arg arg = { p, target_cpu };
4635 int curr_cpu = task_cpu(p);
4637 if (curr_cpu == target_cpu)
4640 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4643 /* TODO: This is not properly updating schedstats */
4645 trace_sched_move_numa(p, curr_cpu, target_cpu);
4646 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4650 * Requeue a task on a given node and accurately track the number of NUMA
4651 * tasks on the runqueues
4653 void sched_setnuma(struct task_struct *p, int nid)
4656 unsigned long flags;
4657 bool on_rq, running;
4659 rq = task_rq_lock(p, &flags);
4661 running = task_current(rq, p);
4664 dequeue_task(rq, p, 0);
4666 p->sched_class->put_prev_task(rq, p);
4668 p->numa_preferred_nid = nid;
4671 p->sched_class->set_curr_task(rq);
4673 enqueue_task(rq, p, 0);
4674 task_rq_unlock(rq, p, &flags);
4679 * migration_cpu_stop - this will be executed by a highprio stopper thread
4680 * and performs thread migration by bumping thread off CPU then
4681 * 'pushing' onto another runqueue.
4683 static int migration_cpu_stop(void *data)
4685 struct migration_arg *arg = data;
4688 * The original target cpu might have gone down and we might
4689 * be on another cpu but it doesn't matter.
4691 local_irq_disable();
4692 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4697 #ifdef CONFIG_HOTPLUG_CPU
4700 * Ensures that the idle task is using init_mm right before its cpu goes
4703 void idle_task_exit(void)
4705 struct mm_struct *mm = current->active_mm;
4707 BUG_ON(cpu_online(smp_processor_id()));
4710 switch_mm(mm, &init_mm, current);
4715 * Since this CPU is going 'away' for a while, fold any nr_active delta
4716 * we might have. Assumes we're called after migrate_tasks() so that the
4717 * nr_active count is stable.
4719 * Also see the comment "Global load-average calculations".
4721 static void calc_load_migrate(struct rq *rq)
4723 long delta = calc_load_fold_active(rq);
4725 atomic_long_add(delta, &calc_load_tasks);
4729 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4730 * try_to_wake_up()->select_task_rq().
4732 * Called with rq->lock held even though we'er in stop_machine() and
4733 * there's no concurrency possible, we hold the required locks anyway
4734 * because of lock validation efforts.
4736 static void migrate_tasks(unsigned int dead_cpu)
4738 struct rq *rq = cpu_rq(dead_cpu);
4739 struct task_struct *next, *stop = rq->stop;
4743 * Fudge the rq selection such that the below task selection loop
4744 * doesn't get stuck on the currently eligible stop task.
4746 * We're currently inside stop_machine() and the rq is either stuck
4747 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4748 * either way we should never end up calling schedule() until we're
4754 * put_prev_task() and pick_next_task() sched
4755 * class method both need to have an up-to-date
4756 * value of rq->clock[_task]
4758 update_rq_clock(rq);
4762 * There's this thread running, bail when that's the only
4765 if (rq->nr_running == 1)
4768 next = pick_next_task(rq);
4770 next->sched_class->put_prev_task(rq, next);
4772 /* Find suitable destination for @next, with force if needed. */
4773 dest_cpu = select_fallback_rq(dead_cpu, next);
4774 raw_spin_unlock(&rq->lock);
4776 __migrate_task(next, dead_cpu, dest_cpu);
4778 raw_spin_lock(&rq->lock);
4784 #endif /* CONFIG_HOTPLUG_CPU */
4786 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4788 static struct ctl_table sd_ctl_dir[] = {
4790 .procname = "sched_domain",
4796 static struct ctl_table sd_ctl_root[] = {
4798 .procname = "kernel",
4800 .child = sd_ctl_dir,
4805 static struct ctl_table *sd_alloc_ctl_entry(int n)
4807 struct ctl_table *entry =
4808 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4813 static void sd_free_ctl_entry(struct ctl_table **tablep)
4815 struct ctl_table *entry;
4818 * In the intermediate directories, both the child directory and
4819 * procname are dynamically allocated and could fail but the mode
4820 * will always be set. In the lowest directory the names are
4821 * static strings and all have proc handlers.
4823 for (entry = *tablep; entry->mode; entry++) {
4825 sd_free_ctl_entry(&entry->child);
4826 if (entry->proc_handler == NULL)
4827 kfree(entry->procname);
4834 static int min_load_idx = 0;
4835 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4838 set_table_entry(struct ctl_table *entry,
4839 const char *procname, void *data, int maxlen,
4840 umode_t mode, proc_handler *proc_handler,
4843 entry->procname = procname;
4845 entry->maxlen = maxlen;
4847 entry->proc_handler = proc_handler;
4850 entry->extra1 = &min_load_idx;
4851 entry->extra2 = &max_load_idx;
4855 static struct ctl_table *
4856 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4858 struct ctl_table *table = sd_alloc_ctl_entry(13);
4863 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4864 sizeof(long), 0644, proc_doulongvec_minmax, false);
4865 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4866 sizeof(long), 0644, proc_doulongvec_minmax, false);
4867 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4868 sizeof(int), 0644, proc_dointvec_minmax, true);
4869 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4870 sizeof(int), 0644, proc_dointvec_minmax, true);
4871 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4872 sizeof(int), 0644, proc_dointvec_minmax, true);
4873 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4874 sizeof(int), 0644, proc_dointvec_minmax, true);
4875 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4876 sizeof(int), 0644, proc_dointvec_minmax, true);
4877 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4878 sizeof(int), 0644, proc_dointvec_minmax, false);
4879 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4880 sizeof(int), 0644, proc_dointvec_minmax, false);
4881 set_table_entry(&table[9], "cache_nice_tries",
4882 &sd->cache_nice_tries,
4883 sizeof(int), 0644, proc_dointvec_minmax, false);
4884 set_table_entry(&table[10], "flags", &sd->flags,
4885 sizeof(int), 0644, proc_dointvec_minmax, false);
4886 set_table_entry(&table[11], "name", sd->name,
4887 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4888 /* &table[12] is terminator */
4893 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4895 struct ctl_table *entry, *table;
4896 struct sched_domain *sd;
4897 int domain_num = 0, i;
4900 for_each_domain(cpu, sd)
4902 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4907 for_each_domain(cpu, sd) {
4908 snprintf(buf, 32, "domain%d", i);
4909 entry->procname = kstrdup(buf, GFP_KERNEL);
4911 entry->child = sd_alloc_ctl_domain_table(sd);
4918 static struct ctl_table_header *sd_sysctl_header;
4919 static void register_sched_domain_sysctl(void)
4921 int i, cpu_num = num_possible_cpus();
4922 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4925 WARN_ON(sd_ctl_dir[0].child);
4926 sd_ctl_dir[0].child = entry;
4931 for_each_possible_cpu(i) {
4932 snprintf(buf, 32, "cpu%d", i);
4933 entry->procname = kstrdup(buf, GFP_KERNEL);
4935 entry->child = sd_alloc_ctl_cpu_table(i);
4939 WARN_ON(sd_sysctl_header);
4940 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4943 /* may be called multiple times per register */
4944 static void unregister_sched_domain_sysctl(void)
4946 if (sd_sysctl_header)
4947 unregister_sysctl_table(sd_sysctl_header);
4948 sd_sysctl_header = NULL;
4949 if (sd_ctl_dir[0].child)
4950 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4953 static void register_sched_domain_sysctl(void)
4956 static void unregister_sched_domain_sysctl(void)
4961 static void set_rq_online(struct rq *rq)
4964 const struct sched_class *class;
4966 cpumask_set_cpu(rq->cpu, rq->rd->online);
4969 for_each_class(class) {
4970 if (class->rq_online)
4971 class->rq_online(rq);
4976 static void set_rq_offline(struct rq *rq)
4979 const struct sched_class *class;
4981 for_each_class(class) {
4982 if (class->rq_offline)
4983 class->rq_offline(rq);
4986 cpumask_clear_cpu(rq->cpu, rq->rd->online);
4992 * migration_call - callback that gets triggered when a CPU is added.
4993 * Here we can start up the necessary migration thread for the new CPU.
4996 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
4998 int cpu = (long)hcpu;
4999 unsigned long flags;
5000 struct rq *rq = cpu_rq(cpu);
5002 switch (action & ~CPU_TASKS_FROZEN) {
5004 case CPU_UP_PREPARE:
5005 rq->calc_load_update = calc_load_update;
5009 /* Update our root-domain */
5010 raw_spin_lock_irqsave(&rq->lock, flags);
5012 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5016 raw_spin_unlock_irqrestore(&rq->lock, flags);
5019 #ifdef CONFIG_HOTPLUG_CPU
5021 sched_ttwu_pending();
5022 /* Update our root-domain */
5023 raw_spin_lock_irqsave(&rq->lock, flags);
5025 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5029 BUG_ON(rq->nr_running != 1); /* the migration thread */
5030 raw_spin_unlock_irqrestore(&rq->lock, flags);
5034 calc_load_migrate(rq);
5039 update_max_interval();
5045 * Register at high priority so that task migration (migrate_all_tasks)
5046 * happens before everything else. This has to be lower priority than
5047 * the notifier in the perf_event subsystem, though.
5049 static struct notifier_block migration_notifier = {
5050 .notifier_call = migration_call,
5051 .priority = CPU_PRI_MIGRATION,
5054 static int sched_cpu_active(struct notifier_block *nfb,
5055 unsigned long action, void *hcpu)
5057 switch (action & ~CPU_TASKS_FROZEN) {
5059 case CPU_DOWN_FAILED:
5060 set_cpu_active((long)hcpu, true);
5067 static int sched_cpu_inactive(struct notifier_block *nfb,
5068 unsigned long action, void *hcpu)
5070 unsigned long flags;
5071 long cpu = (long)hcpu;
5073 switch (action & ~CPU_TASKS_FROZEN) {
5074 case CPU_DOWN_PREPARE:
5075 set_cpu_active(cpu, false);
5077 /* explicitly allow suspend */
5078 if (!(action & CPU_TASKS_FROZEN)) {
5079 struct dl_bw *dl_b = dl_bw_of(cpu);
5083 raw_spin_lock_irqsave(&dl_b->lock, flags);
5084 cpus = dl_bw_cpus(cpu);
5085 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5086 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5089 return notifier_from_errno(-EBUSY);
5097 static int __init migration_init(void)
5099 void *cpu = (void *)(long)smp_processor_id();
5102 /* Initialize migration for the boot CPU */
5103 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5104 BUG_ON(err == NOTIFY_BAD);
5105 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5106 register_cpu_notifier(&migration_notifier);
5108 /* Register cpu active notifiers */
5109 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5110 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5114 early_initcall(migration_init);
5119 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5121 #ifdef CONFIG_SCHED_DEBUG
5123 static __read_mostly int sched_debug_enabled;
5125 static int __init sched_debug_setup(char *str)
5127 sched_debug_enabled = 1;
5131 early_param("sched_debug", sched_debug_setup);
5133 static inline bool sched_debug(void)
5135 return sched_debug_enabled;
5138 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5139 struct cpumask *groupmask)
5141 struct sched_group *group = sd->groups;
5144 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5145 cpumask_clear(groupmask);
5147 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5149 if (!(sd->flags & SD_LOAD_BALANCE)) {
5150 printk("does not load-balance\n");
5152 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5157 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5159 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5160 printk(KERN_ERR "ERROR: domain->span does not contain "
5163 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5164 printk(KERN_ERR "ERROR: domain->groups does not contain"
5168 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5172 printk(KERN_ERR "ERROR: group is NULL\n");
5177 * Even though we initialize ->power to something semi-sane,
5178 * we leave power_orig unset. This allows us to detect if
5179 * domain iteration is still funny without causing /0 traps.
5181 if (!group->sgp->power_orig) {
5182 printk(KERN_CONT "\n");
5183 printk(KERN_ERR "ERROR: domain->cpu_power not "
5188 if (!cpumask_weight(sched_group_cpus(group))) {
5189 printk(KERN_CONT "\n");
5190 printk(KERN_ERR "ERROR: empty group\n");
5194 if (!(sd->flags & SD_OVERLAP) &&
5195 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5196 printk(KERN_CONT "\n");
5197 printk(KERN_ERR "ERROR: repeated CPUs\n");
5201 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5203 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5205 printk(KERN_CONT " %s", str);
5206 if (group->sgp->power != SCHED_POWER_SCALE) {
5207 printk(KERN_CONT " (cpu_power = %d)",
5211 group = group->next;
5212 } while (group != sd->groups);
5213 printk(KERN_CONT "\n");
5215 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5216 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5219 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5220 printk(KERN_ERR "ERROR: parent span is not a superset "
5221 "of domain->span\n");
5225 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5229 if (!sched_debug_enabled)
5233 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5237 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5240 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5248 #else /* !CONFIG_SCHED_DEBUG */
5249 # define sched_domain_debug(sd, cpu) do { } while (0)
5250 static inline bool sched_debug(void)
5254 #endif /* CONFIG_SCHED_DEBUG */
5256 static int sd_degenerate(struct sched_domain *sd)
5258 if (cpumask_weight(sched_domain_span(sd)) == 1)
5261 /* Following flags need at least 2 groups */
5262 if (sd->flags & (SD_LOAD_BALANCE |
5263 SD_BALANCE_NEWIDLE |
5267 SD_SHARE_PKG_RESOURCES)) {
5268 if (sd->groups != sd->groups->next)
5272 /* Following flags don't use groups */
5273 if (sd->flags & (SD_WAKE_AFFINE))
5280 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5282 unsigned long cflags = sd->flags, pflags = parent->flags;
5284 if (sd_degenerate(parent))
5287 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5290 /* Flags needing groups don't count if only 1 group in parent */
5291 if (parent->groups == parent->groups->next) {
5292 pflags &= ~(SD_LOAD_BALANCE |
5293 SD_BALANCE_NEWIDLE |
5297 SD_SHARE_PKG_RESOURCES |
5299 if (nr_node_ids == 1)
5300 pflags &= ~SD_SERIALIZE;
5302 if (~cflags & pflags)
5308 static void free_rootdomain(struct rcu_head *rcu)
5310 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5312 cpupri_cleanup(&rd->cpupri);
5313 cpudl_cleanup(&rd->cpudl);
5314 free_cpumask_var(rd->dlo_mask);
5315 free_cpumask_var(rd->rto_mask);
5316 free_cpumask_var(rd->online);
5317 free_cpumask_var(rd->span);
5321 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5323 struct root_domain *old_rd = NULL;
5324 unsigned long flags;
5326 raw_spin_lock_irqsave(&rq->lock, flags);
5331 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5334 cpumask_clear_cpu(rq->cpu, old_rd->span);
5337 * If we dont want to free the old_rd yet then
5338 * set old_rd to NULL to skip the freeing later
5341 if (!atomic_dec_and_test(&old_rd->refcount))
5345 atomic_inc(&rd->refcount);
5348 cpumask_set_cpu(rq->cpu, rd->span);
5349 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5352 raw_spin_unlock_irqrestore(&rq->lock, flags);
5355 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5358 static int init_rootdomain(struct root_domain *rd)
5360 memset(rd, 0, sizeof(*rd));
5362 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5364 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5366 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5368 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5371 init_dl_bw(&rd->dl_bw);
5372 if (cpudl_init(&rd->cpudl) != 0)
5375 if (cpupri_init(&rd->cpupri) != 0)
5380 free_cpumask_var(rd->rto_mask);
5382 free_cpumask_var(rd->dlo_mask);
5384 free_cpumask_var(rd->online);
5386 free_cpumask_var(rd->span);
5392 * By default the system creates a single root-domain with all cpus as
5393 * members (mimicking the global state we have today).
5395 struct root_domain def_root_domain;
5397 static void init_defrootdomain(void)
5399 init_rootdomain(&def_root_domain);
5401 atomic_set(&def_root_domain.refcount, 1);
5404 static struct root_domain *alloc_rootdomain(void)
5406 struct root_domain *rd;
5408 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5412 if (init_rootdomain(rd) != 0) {
5420 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5422 struct sched_group *tmp, *first;
5431 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5436 } while (sg != first);
5439 static void free_sched_domain(struct rcu_head *rcu)
5441 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5444 * If its an overlapping domain it has private groups, iterate and
5447 if (sd->flags & SD_OVERLAP) {
5448 free_sched_groups(sd->groups, 1);
5449 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5450 kfree(sd->groups->sgp);
5456 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5458 call_rcu(&sd->rcu, free_sched_domain);
5461 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5463 for (; sd; sd = sd->parent)
5464 destroy_sched_domain(sd, cpu);
5468 * Keep a special pointer to the highest sched_domain that has
5469 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5470 * allows us to avoid some pointer chasing select_idle_sibling().
5472 * Also keep a unique ID per domain (we use the first cpu number in
5473 * the cpumask of the domain), this allows us to quickly tell if
5474 * two cpus are in the same cache domain, see cpus_share_cache().
5476 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5477 DEFINE_PER_CPU(int, sd_llc_size);
5478 DEFINE_PER_CPU(int, sd_llc_id);
5479 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5480 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5481 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5483 static void update_top_cache_domain(int cpu)
5485 struct sched_domain *sd;
5486 struct sched_domain *busy_sd = NULL;
5490 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5492 id = cpumask_first(sched_domain_span(sd));
5493 size = cpumask_weight(sched_domain_span(sd));
5494 busy_sd = sd->parent; /* sd_busy */
5496 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5498 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5499 per_cpu(sd_llc_size, cpu) = size;
5500 per_cpu(sd_llc_id, cpu) = id;
5502 sd = lowest_flag_domain(cpu, SD_NUMA);
5503 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5505 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5506 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5510 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5511 * hold the hotplug lock.
5514 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5516 struct rq *rq = cpu_rq(cpu);
5517 struct sched_domain *tmp;
5519 /* Remove the sched domains which do not contribute to scheduling. */
5520 for (tmp = sd; tmp; ) {
5521 struct sched_domain *parent = tmp->parent;
5525 if (sd_parent_degenerate(tmp, parent)) {
5526 tmp->parent = parent->parent;
5528 parent->parent->child = tmp;
5530 * Transfer SD_PREFER_SIBLING down in case of a
5531 * degenerate parent; the spans match for this
5532 * so the property transfers.
5534 if (parent->flags & SD_PREFER_SIBLING)
5535 tmp->flags |= SD_PREFER_SIBLING;
5536 destroy_sched_domain(parent, cpu);
5541 if (sd && sd_degenerate(sd)) {
5544 destroy_sched_domain(tmp, cpu);
5549 sched_domain_debug(sd, cpu);
5551 rq_attach_root(rq, rd);
5553 rcu_assign_pointer(rq->sd, sd);
5554 destroy_sched_domains(tmp, cpu);
5556 update_top_cache_domain(cpu);
5559 /* cpus with isolated domains */
5560 static cpumask_var_t cpu_isolated_map;
5562 /* Setup the mask of cpus configured for isolated domains */
5563 static int __init isolated_cpu_setup(char *str)
5565 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5566 cpulist_parse(str, cpu_isolated_map);
5570 __setup("isolcpus=", isolated_cpu_setup);
5572 static const struct cpumask *cpu_cpu_mask(int cpu)
5574 return cpumask_of_node(cpu_to_node(cpu));
5578 struct sched_domain **__percpu sd;
5579 struct sched_group **__percpu sg;
5580 struct sched_group_power **__percpu sgp;
5584 struct sched_domain ** __percpu sd;
5585 struct root_domain *rd;
5595 struct sched_domain_topology_level;
5597 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5598 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5600 #define SDTL_OVERLAP 0x01
5602 struct sched_domain_topology_level {
5603 sched_domain_init_f init;
5604 sched_domain_mask_f mask;
5607 struct sd_data data;
5611 * Build an iteration mask that can exclude certain CPUs from the upwards
5614 * Asymmetric node setups can result in situations where the domain tree is of
5615 * unequal depth, make sure to skip domains that already cover the entire
5618 * In that case build_sched_domains() will have terminated the iteration early
5619 * and our sibling sd spans will be empty. Domains should always include the
5620 * cpu they're built on, so check that.
5623 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5625 const struct cpumask *span = sched_domain_span(sd);
5626 struct sd_data *sdd = sd->private;
5627 struct sched_domain *sibling;
5630 for_each_cpu(i, span) {
5631 sibling = *per_cpu_ptr(sdd->sd, i);
5632 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5635 cpumask_set_cpu(i, sched_group_mask(sg));
5640 * Return the canonical balance cpu for this group, this is the first cpu
5641 * of this group that's also in the iteration mask.
5643 int group_balance_cpu(struct sched_group *sg)
5645 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5649 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5651 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5652 const struct cpumask *span = sched_domain_span(sd);
5653 struct cpumask *covered = sched_domains_tmpmask;
5654 struct sd_data *sdd = sd->private;
5655 struct sched_domain *child;
5658 cpumask_clear(covered);
5660 for_each_cpu(i, span) {
5661 struct cpumask *sg_span;
5663 if (cpumask_test_cpu(i, covered))
5666 child = *per_cpu_ptr(sdd->sd, i);
5668 /* See the comment near build_group_mask(). */
5669 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5672 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5673 GFP_KERNEL, cpu_to_node(cpu));
5678 sg_span = sched_group_cpus(sg);
5680 child = child->child;
5681 cpumask_copy(sg_span, sched_domain_span(child));
5683 cpumask_set_cpu(i, sg_span);
5685 cpumask_or(covered, covered, sg_span);
5687 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5688 if (atomic_inc_return(&sg->sgp->ref) == 1)
5689 build_group_mask(sd, sg);
5692 * Initialize sgp->power such that even if we mess up the
5693 * domains and no possible iteration will get us here, we won't
5696 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5697 sg->sgp->power_orig = sg->sgp->power;
5700 * Make sure the first group of this domain contains the
5701 * canonical balance cpu. Otherwise the sched_domain iteration
5702 * breaks. See update_sg_lb_stats().
5704 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5705 group_balance_cpu(sg) == cpu)
5715 sd->groups = groups;
5720 free_sched_groups(first, 0);
5725 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5727 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5728 struct sched_domain *child = sd->child;
5731 cpu = cpumask_first(sched_domain_span(child));
5734 *sg = *per_cpu_ptr(sdd->sg, cpu);
5735 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5736 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5743 * build_sched_groups will build a circular linked list of the groups
5744 * covered by the given span, and will set each group's ->cpumask correctly,
5745 * and ->cpu_power to 0.
5747 * Assumes the sched_domain tree is fully constructed
5750 build_sched_groups(struct sched_domain *sd, int cpu)
5752 struct sched_group *first = NULL, *last = NULL;
5753 struct sd_data *sdd = sd->private;
5754 const struct cpumask *span = sched_domain_span(sd);
5755 struct cpumask *covered;
5758 get_group(cpu, sdd, &sd->groups);
5759 atomic_inc(&sd->groups->ref);
5761 if (cpu != cpumask_first(span))
5764 lockdep_assert_held(&sched_domains_mutex);
5765 covered = sched_domains_tmpmask;
5767 cpumask_clear(covered);
5769 for_each_cpu(i, span) {
5770 struct sched_group *sg;
5773 if (cpumask_test_cpu(i, covered))
5776 group = get_group(i, sdd, &sg);
5777 cpumask_clear(sched_group_cpus(sg));
5779 cpumask_setall(sched_group_mask(sg));
5781 for_each_cpu(j, span) {
5782 if (get_group(j, sdd, NULL) != group)
5785 cpumask_set_cpu(j, covered);
5786 cpumask_set_cpu(j, sched_group_cpus(sg));
5801 * Initialize sched groups cpu_power.
5803 * cpu_power indicates the capacity of sched group, which is used while
5804 * distributing the load between different sched groups in a sched domain.
5805 * Typically cpu_power for all the groups in a sched domain will be same unless
5806 * there are asymmetries in the topology. If there are asymmetries, group
5807 * having more cpu_power will pickup more load compared to the group having
5810 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5812 struct sched_group *sg = sd->groups;
5817 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5819 } while (sg != sd->groups);
5821 if (cpu != group_balance_cpu(sg))
5824 update_group_power(sd, cpu);
5825 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5828 int __weak arch_sd_sibling_asym_packing(void)
5830 return 0*SD_ASYM_PACKING;
5834 * Initializers for schedule domains
5835 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5838 #ifdef CONFIG_SCHED_DEBUG
5839 # define SD_INIT_NAME(sd, type) sd->name = #type
5841 # define SD_INIT_NAME(sd, type) do { } while (0)
5844 #define SD_INIT_FUNC(type) \
5845 static noinline struct sched_domain * \
5846 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5848 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5849 *sd = SD_##type##_INIT; \
5850 SD_INIT_NAME(sd, type); \
5851 sd->private = &tl->data; \
5856 #ifdef CONFIG_SCHED_SMT
5857 SD_INIT_FUNC(SIBLING)
5859 #ifdef CONFIG_SCHED_MC
5862 #ifdef CONFIG_SCHED_BOOK
5866 static int default_relax_domain_level = -1;
5867 int sched_domain_level_max;
5869 static int __init setup_relax_domain_level(char *str)
5871 if (kstrtoint(str, 0, &default_relax_domain_level))
5872 pr_warn("Unable to set relax_domain_level\n");
5876 __setup("relax_domain_level=", setup_relax_domain_level);
5878 static void set_domain_attribute(struct sched_domain *sd,
5879 struct sched_domain_attr *attr)
5883 if (!attr || attr->relax_domain_level < 0) {
5884 if (default_relax_domain_level < 0)
5887 request = default_relax_domain_level;
5889 request = attr->relax_domain_level;
5890 if (request < sd->level) {
5891 /* turn off idle balance on this domain */
5892 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5894 /* turn on idle balance on this domain */
5895 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5899 static void __sdt_free(const struct cpumask *cpu_map);
5900 static int __sdt_alloc(const struct cpumask *cpu_map);
5902 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5903 const struct cpumask *cpu_map)
5907 if (!atomic_read(&d->rd->refcount))
5908 free_rootdomain(&d->rd->rcu); /* fall through */
5910 free_percpu(d->sd); /* fall through */
5912 __sdt_free(cpu_map); /* fall through */
5918 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5919 const struct cpumask *cpu_map)
5921 memset(d, 0, sizeof(*d));
5923 if (__sdt_alloc(cpu_map))
5924 return sa_sd_storage;
5925 d->sd = alloc_percpu(struct sched_domain *);
5927 return sa_sd_storage;
5928 d->rd = alloc_rootdomain();
5931 return sa_rootdomain;
5935 * NULL the sd_data elements we've used to build the sched_domain and
5936 * sched_group structure so that the subsequent __free_domain_allocs()
5937 * will not free the data we're using.
5939 static void claim_allocations(int cpu, struct sched_domain *sd)
5941 struct sd_data *sdd = sd->private;
5943 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5944 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5946 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5947 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5949 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5950 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5953 #ifdef CONFIG_SCHED_SMT
5954 static const struct cpumask *cpu_smt_mask(int cpu)
5956 return topology_thread_cpumask(cpu);
5961 * Topology list, bottom-up.
5963 static struct sched_domain_topology_level default_topology[] = {
5964 #ifdef CONFIG_SCHED_SMT
5965 { sd_init_SIBLING, cpu_smt_mask, },
5967 #ifdef CONFIG_SCHED_MC
5968 { sd_init_MC, cpu_coregroup_mask, },
5970 #ifdef CONFIG_SCHED_BOOK
5971 { sd_init_BOOK, cpu_book_mask, },
5973 { sd_init_CPU, cpu_cpu_mask, },
5977 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
5979 #define for_each_sd_topology(tl) \
5980 for (tl = sched_domain_topology; tl->init; tl++)
5984 static int sched_domains_numa_levels;
5985 static int *sched_domains_numa_distance;
5986 static struct cpumask ***sched_domains_numa_masks;
5987 static int sched_domains_curr_level;
5989 static inline int sd_local_flags(int level)
5991 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
5994 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
5997 static struct sched_domain *
5998 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6000 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6001 int level = tl->numa_level;
6002 int sd_weight = cpumask_weight(
6003 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6005 *sd = (struct sched_domain){
6006 .min_interval = sd_weight,
6007 .max_interval = 2*sd_weight,
6009 .imbalance_pct = 125,
6010 .cache_nice_tries = 2,
6017 .flags = 1*SD_LOAD_BALANCE
6018 | 1*SD_BALANCE_NEWIDLE
6023 | 0*SD_SHARE_CPUPOWER
6024 | 0*SD_SHARE_PKG_RESOURCES
6026 | 0*SD_PREFER_SIBLING
6028 | sd_local_flags(level)
6030 .last_balance = jiffies,
6031 .balance_interval = sd_weight,
6033 SD_INIT_NAME(sd, NUMA);
6034 sd->private = &tl->data;
6037 * Ugly hack to pass state to sd_numa_mask()...
6039 sched_domains_curr_level = tl->numa_level;
6044 static const struct cpumask *sd_numa_mask(int cpu)
6046 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6049 static void sched_numa_warn(const char *str)
6051 static int done = false;
6059 printk(KERN_WARNING "ERROR: %s\n\n", str);
6061 for (i = 0; i < nr_node_ids; i++) {
6062 printk(KERN_WARNING " ");
6063 for (j = 0; j < nr_node_ids; j++)
6064 printk(KERN_CONT "%02d ", node_distance(i,j));
6065 printk(KERN_CONT "\n");
6067 printk(KERN_WARNING "\n");
6070 static bool find_numa_distance(int distance)
6074 if (distance == node_distance(0, 0))
6077 for (i = 0; i < sched_domains_numa_levels; i++) {
6078 if (sched_domains_numa_distance[i] == distance)
6085 static void sched_init_numa(void)
6087 int next_distance, curr_distance = node_distance(0, 0);
6088 struct sched_domain_topology_level *tl;
6092 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6093 if (!sched_domains_numa_distance)
6097 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6098 * unique distances in the node_distance() table.
6100 * Assumes node_distance(0,j) includes all distances in
6101 * node_distance(i,j) in order to avoid cubic time.
6103 next_distance = curr_distance;
6104 for (i = 0; i < nr_node_ids; i++) {
6105 for (j = 0; j < nr_node_ids; j++) {
6106 for (k = 0; k < nr_node_ids; k++) {
6107 int distance = node_distance(i, k);
6109 if (distance > curr_distance &&
6110 (distance < next_distance ||
6111 next_distance == curr_distance))
6112 next_distance = distance;
6115 * While not a strong assumption it would be nice to know
6116 * about cases where if node A is connected to B, B is not
6117 * equally connected to A.
6119 if (sched_debug() && node_distance(k, i) != distance)
6120 sched_numa_warn("Node-distance not symmetric");
6122 if (sched_debug() && i && !find_numa_distance(distance))
6123 sched_numa_warn("Node-0 not representative");
6125 if (next_distance != curr_distance) {
6126 sched_domains_numa_distance[level++] = next_distance;
6127 sched_domains_numa_levels = level;
6128 curr_distance = next_distance;
6133 * In case of sched_debug() we verify the above assumption.
6139 * 'level' contains the number of unique distances, excluding the
6140 * identity distance node_distance(i,i).
6142 * The sched_domains_numa_distance[] array includes the actual distance
6147 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6148 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6149 * the array will contain less then 'level' members. This could be
6150 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6151 * in other functions.
6153 * We reset it to 'level' at the end of this function.
6155 sched_domains_numa_levels = 0;
6157 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6158 if (!sched_domains_numa_masks)
6162 * Now for each level, construct a mask per node which contains all
6163 * cpus of nodes that are that many hops away from us.
6165 for (i = 0; i < level; i++) {
6166 sched_domains_numa_masks[i] =
6167 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6168 if (!sched_domains_numa_masks[i])
6171 for (j = 0; j < nr_node_ids; j++) {
6172 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6176 sched_domains_numa_masks[i][j] = mask;
6178 for (k = 0; k < nr_node_ids; k++) {
6179 if (node_distance(j, k) > sched_domains_numa_distance[i])
6182 cpumask_or(mask, mask, cpumask_of_node(k));
6187 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6188 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6193 * Copy the default topology bits..
6195 for (i = 0; default_topology[i].init; i++)
6196 tl[i] = default_topology[i];
6199 * .. and append 'j' levels of NUMA goodness.
6201 for (j = 0; j < level; i++, j++) {
6202 tl[i] = (struct sched_domain_topology_level){
6203 .init = sd_numa_init,
6204 .mask = sd_numa_mask,
6205 .flags = SDTL_OVERLAP,
6210 sched_domain_topology = tl;
6212 sched_domains_numa_levels = level;
6215 static void sched_domains_numa_masks_set(int cpu)
6218 int node = cpu_to_node(cpu);
6220 for (i = 0; i < sched_domains_numa_levels; i++) {
6221 for (j = 0; j < nr_node_ids; j++) {
6222 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6223 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6228 static void sched_domains_numa_masks_clear(int cpu)
6231 for (i = 0; i < sched_domains_numa_levels; i++) {
6232 for (j = 0; j < nr_node_ids; j++)
6233 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6238 * Update sched_domains_numa_masks[level][node] array when new cpus
6241 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6242 unsigned long action,
6245 int cpu = (long)hcpu;
6247 switch (action & ~CPU_TASKS_FROZEN) {
6249 sched_domains_numa_masks_set(cpu);
6253 sched_domains_numa_masks_clear(cpu);
6263 static inline void sched_init_numa(void)
6267 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6268 unsigned long action,
6273 #endif /* CONFIG_NUMA */
6275 static int __sdt_alloc(const struct cpumask *cpu_map)
6277 struct sched_domain_topology_level *tl;
6280 for_each_sd_topology(tl) {
6281 struct sd_data *sdd = &tl->data;
6283 sdd->sd = alloc_percpu(struct sched_domain *);
6287 sdd->sg = alloc_percpu(struct sched_group *);
6291 sdd->sgp = alloc_percpu(struct sched_group_power *);
6295 for_each_cpu(j, cpu_map) {
6296 struct sched_domain *sd;
6297 struct sched_group *sg;
6298 struct sched_group_power *sgp;
6300 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6301 GFP_KERNEL, cpu_to_node(j));
6305 *per_cpu_ptr(sdd->sd, j) = sd;
6307 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6308 GFP_KERNEL, cpu_to_node(j));
6314 *per_cpu_ptr(sdd->sg, j) = sg;
6316 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6317 GFP_KERNEL, cpu_to_node(j));
6321 *per_cpu_ptr(sdd->sgp, j) = sgp;
6328 static void __sdt_free(const struct cpumask *cpu_map)
6330 struct sched_domain_topology_level *tl;
6333 for_each_sd_topology(tl) {
6334 struct sd_data *sdd = &tl->data;
6336 for_each_cpu(j, cpu_map) {
6337 struct sched_domain *sd;
6340 sd = *per_cpu_ptr(sdd->sd, j);
6341 if (sd && (sd->flags & SD_OVERLAP))
6342 free_sched_groups(sd->groups, 0);
6343 kfree(*per_cpu_ptr(sdd->sd, j));
6347 kfree(*per_cpu_ptr(sdd->sg, j));
6349 kfree(*per_cpu_ptr(sdd->sgp, j));
6351 free_percpu(sdd->sd);
6353 free_percpu(sdd->sg);
6355 free_percpu(sdd->sgp);
6360 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6361 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6362 struct sched_domain *child, int cpu)
6364 struct sched_domain *sd = tl->init(tl, cpu);
6368 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6370 sd->level = child->level + 1;
6371 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6375 set_domain_attribute(sd, attr);
6381 * Build sched domains for a given set of cpus and attach the sched domains
6382 * to the individual cpus
6384 static int build_sched_domains(const struct cpumask *cpu_map,
6385 struct sched_domain_attr *attr)
6387 enum s_alloc alloc_state;
6388 struct sched_domain *sd;
6390 int i, ret = -ENOMEM;
6392 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6393 if (alloc_state != sa_rootdomain)
6396 /* Set up domains for cpus specified by the cpu_map. */
6397 for_each_cpu(i, cpu_map) {
6398 struct sched_domain_topology_level *tl;
6401 for_each_sd_topology(tl) {
6402 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6403 if (tl == sched_domain_topology)
6404 *per_cpu_ptr(d.sd, i) = sd;
6405 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6406 sd->flags |= SD_OVERLAP;
6407 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6412 /* Build the groups for the domains */
6413 for_each_cpu(i, cpu_map) {
6414 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6415 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6416 if (sd->flags & SD_OVERLAP) {
6417 if (build_overlap_sched_groups(sd, i))
6420 if (build_sched_groups(sd, i))
6426 /* Calculate CPU power for physical packages and nodes */
6427 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6428 if (!cpumask_test_cpu(i, cpu_map))
6431 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6432 claim_allocations(i, sd);
6433 init_sched_groups_power(i, sd);
6437 /* Attach the domains */
6439 for_each_cpu(i, cpu_map) {
6440 sd = *per_cpu_ptr(d.sd, i);
6441 cpu_attach_domain(sd, d.rd, i);
6447 __free_domain_allocs(&d, alloc_state, cpu_map);
6451 static cpumask_var_t *doms_cur; /* current sched domains */
6452 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6453 static struct sched_domain_attr *dattr_cur;
6454 /* attribues of custom domains in 'doms_cur' */
6457 * Special case: If a kmalloc of a doms_cur partition (array of
6458 * cpumask) fails, then fallback to a single sched domain,
6459 * as determined by the single cpumask fallback_doms.
6461 static cpumask_var_t fallback_doms;
6464 * arch_update_cpu_topology lets virtualized architectures update the
6465 * cpu core maps. It is supposed to return 1 if the topology changed
6466 * or 0 if it stayed the same.
6468 int __attribute__((weak)) arch_update_cpu_topology(void)
6473 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6476 cpumask_var_t *doms;
6478 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6481 for (i = 0; i < ndoms; i++) {
6482 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6483 free_sched_domains(doms, i);
6490 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6493 for (i = 0; i < ndoms; i++)
6494 free_cpumask_var(doms[i]);
6499 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6500 * For now this just excludes isolated cpus, but could be used to
6501 * exclude other special cases in the future.
6503 static int init_sched_domains(const struct cpumask *cpu_map)
6507 arch_update_cpu_topology();
6509 doms_cur = alloc_sched_domains(ndoms_cur);
6511 doms_cur = &fallback_doms;
6512 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6513 err = build_sched_domains(doms_cur[0], NULL);
6514 register_sched_domain_sysctl();
6520 * Detach sched domains from a group of cpus specified in cpu_map
6521 * These cpus will now be attached to the NULL domain
6523 static void detach_destroy_domains(const struct cpumask *cpu_map)
6528 for_each_cpu(i, cpu_map)
6529 cpu_attach_domain(NULL, &def_root_domain, i);
6533 /* handle null as "default" */
6534 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6535 struct sched_domain_attr *new, int idx_new)
6537 struct sched_domain_attr tmp;
6544 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6545 new ? (new + idx_new) : &tmp,
6546 sizeof(struct sched_domain_attr));
6550 * Partition sched domains as specified by the 'ndoms_new'
6551 * cpumasks in the array doms_new[] of cpumasks. This compares
6552 * doms_new[] to the current sched domain partitioning, doms_cur[].
6553 * It destroys each deleted domain and builds each new domain.
6555 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6556 * The masks don't intersect (don't overlap.) We should setup one
6557 * sched domain for each mask. CPUs not in any of the cpumasks will
6558 * not be load balanced. If the same cpumask appears both in the
6559 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6562 * The passed in 'doms_new' should be allocated using
6563 * alloc_sched_domains. This routine takes ownership of it and will
6564 * free_sched_domains it when done with it. If the caller failed the
6565 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6566 * and partition_sched_domains() will fallback to the single partition
6567 * 'fallback_doms', it also forces the domains to be rebuilt.
6569 * If doms_new == NULL it will be replaced with cpu_online_mask.
6570 * ndoms_new == 0 is a special case for destroying existing domains,
6571 * and it will not create the default domain.
6573 * Call with hotplug lock held
6575 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6576 struct sched_domain_attr *dattr_new)
6581 mutex_lock(&sched_domains_mutex);
6583 /* always unregister in case we don't destroy any domains */
6584 unregister_sched_domain_sysctl();
6586 /* Let architecture update cpu core mappings. */
6587 new_topology = arch_update_cpu_topology();
6589 n = doms_new ? ndoms_new : 0;
6591 /* Destroy deleted domains */
6592 for (i = 0; i < ndoms_cur; i++) {
6593 for (j = 0; j < n && !new_topology; j++) {
6594 if (cpumask_equal(doms_cur[i], doms_new[j])
6595 && dattrs_equal(dattr_cur, i, dattr_new, j))
6598 /* no match - a current sched domain not in new doms_new[] */
6599 detach_destroy_domains(doms_cur[i]);
6605 if (doms_new == NULL) {
6607 doms_new = &fallback_doms;
6608 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6609 WARN_ON_ONCE(dattr_new);
6612 /* Build new domains */
6613 for (i = 0; i < ndoms_new; i++) {
6614 for (j = 0; j < n && !new_topology; j++) {
6615 if (cpumask_equal(doms_new[i], doms_cur[j])
6616 && dattrs_equal(dattr_new, i, dattr_cur, j))
6619 /* no match - add a new doms_new */
6620 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6625 /* Remember the new sched domains */
6626 if (doms_cur != &fallback_doms)
6627 free_sched_domains(doms_cur, ndoms_cur);
6628 kfree(dattr_cur); /* kfree(NULL) is safe */
6629 doms_cur = doms_new;
6630 dattr_cur = dattr_new;
6631 ndoms_cur = ndoms_new;
6633 register_sched_domain_sysctl();
6635 mutex_unlock(&sched_domains_mutex);
6638 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6641 * Update cpusets according to cpu_active mask. If cpusets are
6642 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6643 * around partition_sched_domains().
6645 * If we come here as part of a suspend/resume, don't touch cpusets because we
6646 * want to restore it back to its original state upon resume anyway.
6648 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6652 case CPU_ONLINE_FROZEN:
6653 case CPU_DOWN_FAILED_FROZEN:
6656 * num_cpus_frozen tracks how many CPUs are involved in suspend
6657 * resume sequence. As long as this is not the last online
6658 * operation in the resume sequence, just build a single sched
6659 * domain, ignoring cpusets.
6662 if (likely(num_cpus_frozen)) {
6663 partition_sched_domains(1, NULL, NULL);
6668 * This is the last CPU online operation. So fall through and
6669 * restore the original sched domains by considering the
6670 * cpuset configurations.
6674 case CPU_DOWN_FAILED:
6675 cpuset_update_active_cpus(true);
6683 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6687 case CPU_DOWN_PREPARE:
6688 cpuset_update_active_cpus(false);
6690 case CPU_DOWN_PREPARE_FROZEN:
6692 partition_sched_domains(1, NULL, NULL);
6700 void __init sched_init_smp(void)
6702 cpumask_var_t non_isolated_cpus;
6704 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6705 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6710 * There's no userspace yet to cause hotplug operations; hence all the
6711 * cpu masks are stable and all blatant races in the below code cannot
6714 mutex_lock(&sched_domains_mutex);
6715 init_sched_domains(cpu_active_mask);
6716 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6717 if (cpumask_empty(non_isolated_cpus))
6718 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6719 mutex_unlock(&sched_domains_mutex);
6721 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6722 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6723 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6727 /* Move init over to a non-isolated CPU */
6728 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6730 sched_init_granularity();
6731 free_cpumask_var(non_isolated_cpus);
6733 init_sched_rt_class();
6734 init_sched_dl_class();
6737 void __init sched_init_smp(void)
6739 sched_init_granularity();
6741 #endif /* CONFIG_SMP */
6743 const_debug unsigned int sysctl_timer_migration = 1;
6745 int in_sched_functions(unsigned long addr)
6747 return in_lock_functions(addr) ||
6748 (addr >= (unsigned long)__sched_text_start
6749 && addr < (unsigned long)__sched_text_end);
6752 #ifdef CONFIG_CGROUP_SCHED
6754 * Default task group.
6755 * Every task in system belongs to this group at bootup.
6757 struct task_group root_task_group;
6758 LIST_HEAD(task_groups);
6761 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6763 void __init sched_init(void)
6766 unsigned long alloc_size = 0, ptr;
6768 #ifdef CONFIG_FAIR_GROUP_SCHED
6769 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6771 #ifdef CONFIG_RT_GROUP_SCHED
6772 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6774 #ifdef CONFIG_CPUMASK_OFFSTACK
6775 alloc_size += num_possible_cpus() * cpumask_size();
6778 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6780 #ifdef CONFIG_FAIR_GROUP_SCHED
6781 root_task_group.se = (struct sched_entity **)ptr;
6782 ptr += nr_cpu_ids * sizeof(void **);
6784 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6785 ptr += nr_cpu_ids * sizeof(void **);
6787 #endif /* CONFIG_FAIR_GROUP_SCHED */
6788 #ifdef CONFIG_RT_GROUP_SCHED
6789 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6790 ptr += nr_cpu_ids * sizeof(void **);
6792 root_task_group.rt_rq = (struct rt_rq **)ptr;
6793 ptr += nr_cpu_ids * sizeof(void **);
6795 #endif /* CONFIG_RT_GROUP_SCHED */
6796 #ifdef CONFIG_CPUMASK_OFFSTACK
6797 for_each_possible_cpu(i) {
6798 per_cpu(load_balance_mask, i) = (void *)ptr;
6799 ptr += cpumask_size();
6801 #endif /* CONFIG_CPUMASK_OFFSTACK */
6804 init_rt_bandwidth(&def_rt_bandwidth,
6805 global_rt_period(), global_rt_runtime());
6806 init_dl_bandwidth(&def_dl_bandwidth,
6807 global_rt_period(), global_rt_runtime());
6810 init_defrootdomain();
6813 #ifdef CONFIG_RT_GROUP_SCHED
6814 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6815 global_rt_period(), global_rt_runtime());
6816 #endif /* CONFIG_RT_GROUP_SCHED */
6818 #ifdef CONFIG_CGROUP_SCHED
6819 list_add(&root_task_group.list, &task_groups);
6820 INIT_LIST_HEAD(&root_task_group.children);
6821 INIT_LIST_HEAD(&root_task_group.siblings);
6822 autogroup_init(&init_task);
6824 #endif /* CONFIG_CGROUP_SCHED */
6826 for_each_possible_cpu(i) {
6830 raw_spin_lock_init(&rq->lock);
6832 rq->calc_load_active = 0;
6833 rq->calc_load_update = jiffies + LOAD_FREQ;
6834 init_cfs_rq(&rq->cfs);
6835 init_rt_rq(&rq->rt, rq);
6836 init_dl_rq(&rq->dl, rq);
6837 #ifdef CONFIG_FAIR_GROUP_SCHED
6838 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6839 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6841 * How much cpu bandwidth does root_task_group get?
6843 * In case of task-groups formed thr' the cgroup filesystem, it
6844 * gets 100% of the cpu resources in the system. This overall
6845 * system cpu resource is divided among the tasks of
6846 * root_task_group and its child task-groups in a fair manner,
6847 * based on each entity's (task or task-group's) weight
6848 * (se->load.weight).
6850 * In other words, if root_task_group has 10 tasks of weight
6851 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6852 * then A0's share of the cpu resource is:
6854 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6856 * We achieve this by letting root_task_group's tasks sit
6857 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6859 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6860 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6861 #endif /* CONFIG_FAIR_GROUP_SCHED */
6863 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6864 #ifdef CONFIG_RT_GROUP_SCHED
6865 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
6866 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6869 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6870 rq->cpu_load[j] = 0;
6872 rq->last_load_update_tick = jiffies;
6877 rq->cpu_power = SCHED_POWER_SCALE;
6878 rq->post_schedule = 0;
6879 rq->active_balance = 0;
6880 rq->next_balance = jiffies;
6885 rq->avg_idle = 2*sysctl_sched_migration_cost;
6886 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6888 INIT_LIST_HEAD(&rq->cfs_tasks);
6890 rq_attach_root(rq, &def_root_domain);
6891 #ifdef CONFIG_NO_HZ_COMMON
6894 #ifdef CONFIG_NO_HZ_FULL
6895 rq->last_sched_tick = 0;
6899 atomic_set(&rq->nr_iowait, 0);
6902 set_load_weight(&init_task);
6904 #ifdef CONFIG_PREEMPT_NOTIFIERS
6905 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6909 * The boot idle thread does lazy MMU switching as well:
6911 atomic_inc(&init_mm.mm_count);
6912 enter_lazy_tlb(&init_mm, current);
6915 * Make us the idle thread. Technically, schedule() should not be
6916 * called from this thread, however somewhere below it might be,
6917 * but because we are the idle thread, we just pick up running again
6918 * when this runqueue becomes "idle".
6920 init_idle(current, smp_processor_id());
6922 calc_load_update = jiffies + LOAD_FREQ;
6925 * During early bootup we pretend to be a normal task:
6927 current->sched_class = &fair_sched_class;
6930 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6931 /* May be allocated at isolcpus cmdline parse time */
6932 if (cpu_isolated_map == NULL)
6933 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6934 idle_thread_set_boot_cpu();
6936 init_sched_fair_class();
6938 scheduler_running = 1;
6941 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6942 static inline int preempt_count_equals(int preempt_offset)
6944 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6946 return (nested == preempt_offset);
6949 void __might_sleep(const char *file, int line, int preempt_offset)
6951 static unsigned long prev_jiffy; /* ratelimiting */
6953 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6954 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
6955 system_state != SYSTEM_RUNNING || oops_in_progress)
6957 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6959 prev_jiffy = jiffies;
6962 "BUG: sleeping function called from invalid context at %s:%d\n",
6965 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6966 in_atomic(), irqs_disabled(),
6967 current->pid, current->comm);
6969 debug_show_held_locks(current);
6970 if (irqs_disabled())
6971 print_irqtrace_events(current);
6974 EXPORT_SYMBOL(__might_sleep);
6977 #ifdef CONFIG_MAGIC_SYSRQ
6978 static void normalize_task(struct rq *rq, struct task_struct *p)
6980 const struct sched_class *prev_class = p->sched_class;
6981 struct sched_attr attr = {
6982 .sched_policy = SCHED_NORMAL,
6984 int old_prio = p->prio;
6989 dequeue_task(rq, p, 0);
6990 __setscheduler(rq, p, &attr);
6992 enqueue_task(rq, p, 0);
6993 resched_task(rq->curr);
6996 check_class_changed(rq, p, prev_class, old_prio);
6999 void normalize_rt_tasks(void)
7001 struct task_struct *g, *p;
7002 unsigned long flags;
7005 read_lock_irqsave(&tasklist_lock, flags);
7006 do_each_thread(g, p) {
7008 * Only normalize user tasks:
7013 p->se.exec_start = 0;
7014 #ifdef CONFIG_SCHEDSTATS
7015 p->se.statistics.wait_start = 0;
7016 p->se.statistics.sleep_start = 0;
7017 p->se.statistics.block_start = 0;
7020 if (!dl_task(p) && !rt_task(p)) {
7022 * Renice negative nice level userspace
7025 if (TASK_NICE(p) < 0 && p->mm)
7026 set_user_nice(p, 0);
7030 raw_spin_lock(&p->pi_lock);
7031 rq = __task_rq_lock(p);
7033 normalize_task(rq, p);
7035 __task_rq_unlock(rq);
7036 raw_spin_unlock(&p->pi_lock);
7037 } while_each_thread(g, p);
7039 read_unlock_irqrestore(&tasklist_lock, flags);
7042 #endif /* CONFIG_MAGIC_SYSRQ */
7044 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7046 * These functions are only useful for the IA64 MCA handling, or kdb.
7048 * They can only be called when the whole system has been
7049 * stopped - every CPU needs to be quiescent, and no scheduling
7050 * activity can take place. Using them for anything else would
7051 * be a serious bug, and as a result, they aren't even visible
7052 * under any other configuration.
7056 * curr_task - return the current task for a given cpu.
7057 * @cpu: the processor in question.
7059 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7061 * Return: The current task for @cpu.
7063 struct task_struct *curr_task(int cpu)
7065 return cpu_curr(cpu);
7068 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7072 * set_curr_task - set the current task for a given cpu.
7073 * @cpu: the processor in question.
7074 * @p: the task pointer to set.
7076 * Description: This function must only be used when non-maskable interrupts
7077 * are serviced on a separate stack. It allows the architecture to switch the
7078 * notion of the current task on a cpu in a non-blocking manner. This function
7079 * must be called with all CPU's synchronized, and interrupts disabled, the
7080 * and caller must save the original value of the current task (see
7081 * curr_task() above) and restore that value before reenabling interrupts and
7082 * re-starting the system.
7084 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7086 void set_curr_task(int cpu, struct task_struct *p)
7093 #ifdef CONFIG_CGROUP_SCHED
7094 /* task_group_lock serializes the addition/removal of task groups */
7095 static DEFINE_SPINLOCK(task_group_lock);
7097 static void free_sched_group(struct task_group *tg)
7099 free_fair_sched_group(tg);
7100 free_rt_sched_group(tg);
7105 /* allocate runqueue etc for a new task group */
7106 struct task_group *sched_create_group(struct task_group *parent)
7108 struct task_group *tg;
7110 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7112 return ERR_PTR(-ENOMEM);
7114 if (!alloc_fair_sched_group(tg, parent))
7117 if (!alloc_rt_sched_group(tg, parent))
7123 free_sched_group(tg);
7124 return ERR_PTR(-ENOMEM);
7127 void sched_online_group(struct task_group *tg, struct task_group *parent)
7129 unsigned long flags;
7131 spin_lock_irqsave(&task_group_lock, flags);
7132 list_add_rcu(&tg->list, &task_groups);
7134 WARN_ON(!parent); /* root should already exist */
7136 tg->parent = parent;
7137 INIT_LIST_HEAD(&tg->children);
7138 list_add_rcu(&tg->siblings, &parent->children);
7139 spin_unlock_irqrestore(&task_group_lock, flags);
7142 /* rcu callback to free various structures associated with a task group */
7143 static void free_sched_group_rcu(struct rcu_head *rhp)
7145 /* now it should be safe to free those cfs_rqs */
7146 free_sched_group(container_of(rhp, struct task_group, rcu));
7149 /* Destroy runqueue etc associated with a task group */
7150 void sched_destroy_group(struct task_group *tg)
7152 /* wait for possible concurrent references to cfs_rqs complete */
7153 call_rcu(&tg->rcu, free_sched_group_rcu);
7156 void sched_offline_group(struct task_group *tg)
7158 unsigned long flags;
7161 /* end participation in shares distribution */
7162 for_each_possible_cpu(i)
7163 unregister_fair_sched_group(tg, i);
7165 spin_lock_irqsave(&task_group_lock, flags);
7166 list_del_rcu(&tg->list);
7167 list_del_rcu(&tg->siblings);
7168 spin_unlock_irqrestore(&task_group_lock, flags);
7171 /* change task's runqueue when it moves between groups.
7172 * The caller of this function should have put the task in its new group
7173 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7174 * reflect its new group.
7176 void sched_move_task(struct task_struct *tsk)
7178 struct task_group *tg;
7180 unsigned long flags;
7183 rq = task_rq_lock(tsk, &flags);
7185 running = task_current(rq, tsk);
7189 dequeue_task(rq, tsk, 0);
7190 if (unlikely(running))
7191 tsk->sched_class->put_prev_task(rq, tsk);
7193 tg = container_of(task_css_check(tsk, cpu_cgroup_subsys_id,
7194 lockdep_is_held(&tsk->sighand->siglock)),
7195 struct task_group, css);
7196 tg = autogroup_task_group(tsk, tg);
7197 tsk->sched_task_group = tg;
7199 #ifdef CONFIG_FAIR_GROUP_SCHED
7200 if (tsk->sched_class->task_move_group)
7201 tsk->sched_class->task_move_group(tsk, on_rq);
7204 set_task_rq(tsk, task_cpu(tsk));
7206 if (unlikely(running))
7207 tsk->sched_class->set_curr_task(rq);
7209 enqueue_task(rq, tsk, 0);
7211 task_rq_unlock(rq, tsk, &flags);
7213 #endif /* CONFIG_CGROUP_SCHED */
7215 #ifdef CONFIG_RT_GROUP_SCHED
7217 * Ensure that the real time constraints are schedulable.
7219 static DEFINE_MUTEX(rt_constraints_mutex);
7221 /* Must be called with tasklist_lock held */
7222 static inline int tg_has_rt_tasks(struct task_group *tg)
7224 struct task_struct *g, *p;
7226 do_each_thread(g, p) {
7227 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7229 } while_each_thread(g, p);
7234 struct rt_schedulable_data {
7235 struct task_group *tg;
7240 static int tg_rt_schedulable(struct task_group *tg, void *data)
7242 struct rt_schedulable_data *d = data;
7243 struct task_group *child;
7244 unsigned long total, sum = 0;
7245 u64 period, runtime;
7247 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7248 runtime = tg->rt_bandwidth.rt_runtime;
7251 period = d->rt_period;
7252 runtime = d->rt_runtime;
7256 * Cannot have more runtime than the period.
7258 if (runtime > period && runtime != RUNTIME_INF)
7262 * Ensure we don't starve existing RT tasks.
7264 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7267 total = to_ratio(period, runtime);
7270 * Nobody can have more than the global setting allows.
7272 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7276 * The sum of our children's runtime should not exceed our own.
7278 list_for_each_entry_rcu(child, &tg->children, siblings) {
7279 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7280 runtime = child->rt_bandwidth.rt_runtime;
7282 if (child == d->tg) {
7283 period = d->rt_period;
7284 runtime = d->rt_runtime;
7287 sum += to_ratio(period, runtime);
7296 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7300 struct rt_schedulable_data data = {
7302 .rt_period = period,
7303 .rt_runtime = runtime,
7307 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7313 static int tg_set_rt_bandwidth(struct task_group *tg,
7314 u64 rt_period, u64 rt_runtime)
7318 mutex_lock(&rt_constraints_mutex);
7319 read_lock(&tasklist_lock);
7320 err = __rt_schedulable(tg, rt_period, rt_runtime);
7324 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7325 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7326 tg->rt_bandwidth.rt_runtime = rt_runtime;
7328 for_each_possible_cpu(i) {
7329 struct rt_rq *rt_rq = tg->rt_rq[i];
7331 raw_spin_lock(&rt_rq->rt_runtime_lock);
7332 rt_rq->rt_runtime = rt_runtime;
7333 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7335 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7337 read_unlock(&tasklist_lock);
7338 mutex_unlock(&rt_constraints_mutex);
7343 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7345 u64 rt_runtime, rt_period;
7347 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7348 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7349 if (rt_runtime_us < 0)
7350 rt_runtime = RUNTIME_INF;
7352 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7355 static long sched_group_rt_runtime(struct task_group *tg)
7359 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7362 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7363 do_div(rt_runtime_us, NSEC_PER_USEC);
7364 return rt_runtime_us;
7367 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7369 u64 rt_runtime, rt_period;
7371 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7372 rt_runtime = tg->rt_bandwidth.rt_runtime;
7377 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7380 static long sched_group_rt_period(struct task_group *tg)
7384 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7385 do_div(rt_period_us, NSEC_PER_USEC);
7386 return rt_period_us;
7388 #endif /* CONFIG_RT_GROUP_SCHED */
7390 #ifdef CONFIG_RT_GROUP_SCHED
7391 static int sched_rt_global_constraints(void)
7395 mutex_lock(&rt_constraints_mutex);
7396 read_lock(&tasklist_lock);
7397 ret = __rt_schedulable(NULL, 0, 0);
7398 read_unlock(&tasklist_lock);
7399 mutex_unlock(&rt_constraints_mutex);
7404 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7406 /* Don't accept realtime tasks when there is no way for them to run */
7407 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7413 #else /* !CONFIG_RT_GROUP_SCHED */
7414 static int sched_rt_global_constraints(void)
7416 unsigned long flags;
7419 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7420 for_each_possible_cpu(i) {
7421 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7423 raw_spin_lock(&rt_rq->rt_runtime_lock);
7424 rt_rq->rt_runtime = global_rt_runtime();
7425 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7427 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7431 #endif /* CONFIG_RT_GROUP_SCHED */
7433 static int sched_dl_global_constraints(void)
7435 u64 runtime = global_rt_runtime();
7436 u64 period = global_rt_period();
7437 u64 new_bw = to_ratio(period, runtime);
7439 unsigned long flags;
7442 * Here we want to check the bandwidth not being set to some
7443 * value smaller than the currently allocated bandwidth in
7444 * any of the root_domains.
7446 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7447 * cycling on root_domains... Discussion on different/better
7448 * solutions is welcome!
7450 for_each_possible_cpu(cpu) {
7451 struct dl_bw *dl_b = dl_bw_of(cpu);
7453 raw_spin_lock_irqsave(&dl_b->lock, flags);
7454 if (new_bw < dl_b->total_bw)
7456 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7465 static void sched_dl_do_global(void)
7469 unsigned long flags;
7471 def_dl_bandwidth.dl_period = global_rt_period();
7472 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7474 if (global_rt_runtime() != RUNTIME_INF)
7475 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7478 * FIXME: As above...
7480 for_each_possible_cpu(cpu) {
7481 struct dl_bw *dl_b = dl_bw_of(cpu);
7483 raw_spin_lock_irqsave(&dl_b->lock, flags);
7485 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7489 static int sched_rt_global_validate(void)
7491 if (sysctl_sched_rt_period <= 0)
7494 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7495 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7501 static void sched_rt_do_global(void)
7503 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7504 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7507 int sched_rt_handler(struct ctl_table *table, int write,
7508 void __user *buffer, size_t *lenp,
7511 int old_period, old_runtime;
7512 static DEFINE_MUTEX(mutex);
7516 old_period = sysctl_sched_rt_period;
7517 old_runtime = sysctl_sched_rt_runtime;
7519 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7521 if (!ret && write) {
7522 ret = sched_rt_global_validate();
7526 ret = sched_rt_global_constraints();
7530 ret = sched_dl_global_constraints();
7534 sched_rt_do_global();
7535 sched_dl_do_global();
7539 sysctl_sched_rt_period = old_period;
7540 sysctl_sched_rt_runtime = old_runtime;
7542 mutex_unlock(&mutex);
7547 int sched_rr_handler(struct ctl_table *table, int write,
7548 void __user *buffer, size_t *lenp,
7552 static DEFINE_MUTEX(mutex);
7555 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7556 /* make sure that internally we keep jiffies */
7557 /* also, writing zero resets timeslice to default */
7558 if (!ret && write) {
7559 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7560 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7562 mutex_unlock(&mutex);
7566 #ifdef CONFIG_CGROUP_SCHED
7568 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7570 return css ? container_of(css, struct task_group, css) : NULL;
7573 static struct cgroup_subsys_state *
7574 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7576 struct task_group *parent = css_tg(parent_css);
7577 struct task_group *tg;
7580 /* This is early initialization for the top cgroup */
7581 return &root_task_group.css;
7584 tg = sched_create_group(parent);
7586 return ERR_PTR(-ENOMEM);
7591 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7593 struct task_group *tg = css_tg(css);
7594 struct task_group *parent = css_tg(css_parent(css));
7597 sched_online_group(tg, parent);
7601 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7603 struct task_group *tg = css_tg(css);
7605 sched_destroy_group(tg);
7608 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7610 struct task_group *tg = css_tg(css);
7612 sched_offline_group(tg);
7615 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7616 struct cgroup_taskset *tset)
7618 struct task_struct *task;
7620 cgroup_taskset_for_each(task, css, tset) {
7621 #ifdef CONFIG_RT_GROUP_SCHED
7622 if (!sched_rt_can_attach(css_tg(css), task))
7625 /* We don't support RT-tasks being in separate groups */
7626 if (task->sched_class != &fair_sched_class)
7633 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7634 struct cgroup_taskset *tset)
7636 struct task_struct *task;
7638 cgroup_taskset_for_each(task, css, tset)
7639 sched_move_task(task);
7642 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7643 struct cgroup_subsys_state *old_css,
7644 struct task_struct *task)
7647 * cgroup_exit() is called in the copy_process() failure path.
7648 * Ignore this case since the task hasn't ran yet, this avoids
7649 * trying to poke a half freed task state from generic code.
7651 if (!(task->flags & PF_EXITING))
7654 sched_move_task(task);
7657 #ifdef CONFIG_FAIR_GROUP_SCHED
7658 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7659 struct cftype *cftype, u64 shareval)
7661 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7664 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7667 struct task_group *tg = css_tg(css);
7669 return (u64) scale_load_down(tg->shares);
7672 #ifdef CONFIG_CFS_BANDWIDTH
7673 static DEFINE_MUTEX(cfs_constraints_mutex);
7675 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7676 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7678 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7680 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7682 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7683 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7685 if (tg == &root_task_group)
7689 * Ensure we have at some amount of bandwidth every period. This is
7690 * to prevent reaching a state of large arrears when throttled via
7691 * entity_tick() resulting in prolonged exit starvation.
7693 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7697 * Likewise, bound things on the otherside by preventing insane quota
7698 * periods. This also allows us to normalize in computing quota
7701 if (period > max_cfs_quota_period)
7704 mutex_lock(&cfs_constraints_mutex);
7705 ret = __cfs_schedulable(tg, period, quota);
7709 runtime_enabled = quota != RUNTIME_INF;
7710 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7712 * If we need to toggle cfs_bandwidth_used, off->on must occur
7713 * before making related changes, and on->off must occur afterwards
7715 if (runtime_enabled && !runtime_was_enabled)
7716 cfs_bandwidth_usage_inc();
7717 raw_spin_lock_irq(&cfs_b->lock);
7718 cfs_b->period = ns_to_ktime(period);
7719 cfs_b->quota = quota;
7721 __refill_cfs_bandwidth_runtime(cfs_b);
7722 /* restart the period timer (if active) to handle new period expiry */
7723 if (runtime_enabled && cfs_b->timer_active) {
7724 /* force a reprogram */
7725 cfs_b->timer_active = 0;
7726 __start_cfs_bandwidth(cfs_b);
7728 raw_spin_unlock_irq(&cfs_b->lock);
7730 for_each_possible_cpu(i) {
7731 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7732 struct rq *rq = cfs_rq->rq;
7734 raw_spin_lock_irq(&rq->lock);
7735 cfs_rq->runtime_enabled = runtime_enabled;
7736 cfs_rq->runtime_remaining = 0;
7738 if (cfs_rq->throttled)
7739 unthrottle_cfs_rq(cfs_rq);
7740 raw_spin_unlock_irq(&rq->lock);
7742 if (runtime_was_enabled && !runtime_enabled)
7743 cfs_bandwidth_usage_dec();
7745 mutex_unlock(&cfs_constraints_mutex);
7750 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7754 period = ktime_to_ns(tg->cfs_bandwidth.period);
7755 if (cfs_quota_us < 0)
7756 quota = RUNTIME_INF;
7758 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7760 return tg_set_cfs_bandwidth(tg, period, quota);
7763 long tg_get_cfs_quota(struct task_group *tg)
7767 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7770 quota_us = tg->cfs_bandwidth.quota;
7771 do_div(quota_us, NSEC_PER_USEC);
7776 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7780 period = (u64)cfs_period_us * NSEC_PER_USEC;
7781 quota = tg->cfs_bandwidth.quota;
7783 return tg_set_cfs_bandwidth(tg, period, quota);
7786 long tg_get_cfs_period(struct task_group *tg)
7790 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7791 do_div(cfs_period_us, NSEC_PER_USEC);
7793 return cfs_period_us;
7796 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7799 return tg_get_cfs_quota(css_tg(css));
7802 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7803 struct cftype *cftype, s64 cfs_quota_us)
7805 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7808 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7811 return tg_get_cfs_period(css_tg(css));
7814 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7815 struct cftype *cftype, u64 cfs_period_us)
7817 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7820 struct cfs_schedulable_data {
7821 struct task_group *tg;
7826 * normalize group quota/period to be quota/max_period
7827 * note: units are usecs
7829 static u64 normalize_cfs_quota(struct task_group *tg,
7830 struct cfs_schedulable_data *d)
7838 period = tg_get_cfs_period(tg);
7839 quota = tg_get_cfs_quota(tg);
7842 /* note: these should typically be equivalent */
7843 if (quota == RUNTIME_INF || quota == -1)
7846 return to_ratio(period, quota);
7849 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7851 struct cfs_schedulable_data *d = data;
7852 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7853 s64 quota = 0, parent_quota = -1;
7856 quota = RUNTIME_INF;
7858 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7860 quota = normalize_cfs_quota(tg, d);
7861 parent_quota = parent_b->hierarchal_quota;
7864 * ensure max(child_quota) <= parent_quota, inherit when no
7867 if (quota == RUNTIME_INF)
7868 quota = parent_quota;
7869 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7872 cfs_b->hierarchal_quota = quota;
7877 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7880 struct cfs_schedulable_data data = {
7886 if (quota != RUNTIME_INF) {
7887 do_div(data.period, NSEC_PER_USEC);
7888 do_div(data.quota, NSEC_PER_USEC);
7892 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7898 static int cpu_stats_show(struct seq_file *sf, void *v)
7900 struct task_group *tg = css_tg(seq_css(sf));
7901 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7903 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7904 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7905 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7909 #endif /* CONFIG_CFS_BANDWIDTH */
7910 #endif /* CONFIG_FAIR_GROUP_SCHED */
7912 #ifdef CONFIG_RT_GROUP_SCHED
7913 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7914 struct cftype *cft, s64 val)
7916 return sched_group_set_rt_runtime(css_tg(css), val);
7919 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7922 return sched_group_rt_runtime(css_tg(css));
7925 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7926 struct cftype *cftype, u64 rt_period_us)
7928 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7931 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7934 return sched_group_rt_period(css_tg(css));
7936 #endif /* CONFIG_RT_GROUP_SCHED */
7938 static struct cftype cpu_files[] = {
7939 #ifdef CONFIG_FAIR_GROUP_SCHED
7942 .read_u64 = cpu_shares_read_u64,
7943 .write_u64 = cpu_shares_write_u64,
7946 #ifdef CONFIG_CFS_BANDWIDTH
7948 .name = "cfs_quota_us",
7949 .read_s64 = cpu_cfs_quota_read_s64,
7950 .write_s64 = cpu_cfs_quota_write_s64,
7953 .name = "cfs_period_us",
7954 .read_u64 = cpu_cfs_period_read_u64,
7955 .write_u64 = cpu_cfs_period_write_u64,
7959 .seq_show = cpu_stats_show,
7962 #ifdef CONFIG_RT_GROUP_SCHED
7964 .name = "rt_runtime_us",
7965 .read_s64 = cpu_rt_runtime_read,
7966 .write_s64 = cpu_rt_runtime_write,
7969 .name = "rt_period_us",
7970 .read_u64 = cpu_rt_period_read_uint,
7971 .write_u64 = cpu_rt_period_write_uint,
7977 struct cgroup_subsys cpu_cgroup_subsys = {
7979 .css_alloc = cpu_cgroup_css_alloc,
7980 .css_free = cpu_cgroup_css_free,
7981 .css_online = cpu_cgroup_css_online,
7982 .css_offline = cpu_cgroup_css_offline,
7983 .can_attach = cpu_cgroup_can_attach,
7984 .attach = cpu_cgroup_attach,
7985 .exit = cpu_cgroup_exit,
7986 .subsys_id = cpu_cgroup_subsys_id,
7987 .base_cftypes = cpu_files,
7991 #endif /* CONFIG_CGROUP_SCHED */
7993 void dump_cpu_task(int cpu)
7995 pr_info("Task dump for CPU %d:\n", cpu);
7996 sched_show_task(cpu_curr(cpu));