4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
93 void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
96 ktime_t soft, hard, now;
99 if (hrtimer_active(period_timer))
102 now = hrtimer_cb_get_time(period_timer);
103 hrtimer_forward(period_timer, now, period);
105 soft = hrtimer_get_softexpires(period_timer);
106 hard = hrtimer_get_expires(period_timer);
107 delta = ktime_to_ns(ktime_sub(hard, soft));
108 __hrtimer_start_range_ns(period_timer, soft, delta,
109 HRTIMER_MODE_ABS_PINNED, 0);
113 DEFINE_MUTEX(sched_domains_mutex);
114 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
116 static void update_rq_clock_task(struct rq *rq, s64 delta);
118 void update_rq_clock(struct rq *rq)
122 if (rq->skip_clock_update > 0)
125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
127 update_rq_clock_task(rq, delta);
131 * Debugging: various feature bits
134 #define SCHED_FEAT(name, enabled) \
135 (1UL << __SCHED_FEAT_##name) * enabled |
137 const_debug unsigned int sysctl_sched_features =
138 #include "features.h"
143 #ifdef CONFIG_SCHED_DEBUG
144 #define SCHED_FEAT(name, enabled) \
147 static const char * const sched_feat_names[] = {
148 #include "features.h"
153 static int sched_feat_show(struct seq_file *m, void *v)
157 for (i = 0; i < __SCHED_FEAT_NR; i++) {
158 if (!(sysctl_sched_features & (1UL << i)))
160 seq_printf(m, "%s ", sched_feat_names[i]);
167 #ifdef HAVE_JUMP_LABEL
169 #define jump_label_key__true STATIC_KEY_INIT_TRUE
170 #define jump_label_key__false STATIC_KEY_INIT_FALSE
172 #define SCHED_FEAT(name, enabled) \
173 jump_label_key__##enabled ,
175 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
176 #include "features.h"
181 static void sched_feat_disable(int i)
183 if (static_key_enabled(&sched_feat_keys[i]))
184 static_key_slow_dec(&sched_feat_keys[i]);
187 static void sched_feat_enable(int i)
189 if (!static_key_enabled(&sched_feat_keys[i]))
190 static_key_slow_inc(&sched_feat_keys[i]);
193 static void sched_feat_disable(int i) { };
194 static void sched_feat_enable(int i) { };
195 #endif /* HAVE_JUMP_LABEL */
197 static int sched_feat_set(char *cmp)
202 if (strncmp(cmp, "NO_", 3) == 0) {
207 for (i = 0; i < __SCHED_FEAT_NR; i++) {
208 if (strcmp(cmp, sched_feat_names[i]) == 0) {
210 sysctl_sched_features &= ~(1UL << i);
211 sched_feat_disable(i);
213 sysctl_sched_features |= (1UL << i);
214 sched_feat_enable(i);
224 sched_feat_write(struct file *filp, const char __user *ubuf,
225 size_t cnt, loff_t *ppos)
234 if (copy_from_user(&buf, ubuf, cnt))
240 i = sched_feat_set(cmp);
241 if (i == __SCHED_FEAT_NR)
249 static int sched_feat_open(struct inode *inode, struct file *filp)
251 return single_open(filp, sched_feat_show, NULL);
254 static const struct file_operations sched_feat_fops = {
255 .open = sched_feat_open,
256 .write = sched_feat_write,
259 .release = single_release,
262 static __init int sched_init_debug(void)
264 debugfs_create_file("sched_features", 0644, NULL, NULL,
269 late_initcall(sched_init_debug);
270 #endif /* CONFIG_SCHED_DEBUG */
273 * Number of tasks to iterate in a single balance run.
274 * Limited because this is done with IRQs disabled.
276 const_debug unsigned int sysctl_sched_nr_migrate = 32;
279 * period over which we average the RT time consumption, measured
284 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
287 * period over which we measure -rt task cpu usage in us.
290 unsigned int sysctl_sched_rt_period = 1000000;
292 __read_mostly int scheduler_running;
295 * part of the period that we allow rt tasks to run in us.
298 int sysctl_sched_rt_runtime = 950000;
301 * __task_rq_lock - lock the rq @p resides on.
303 static inline struct rq *__task_rq_lock(struct task_struct *p)
308 lockdep_assert_held(&p->pi_lock);
312 raw_spin_lock(&rq->lock);
313 if (likely(rq == task_rq(p)))
315 raw_spin_unlock(&rq->lock);
320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
322 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
323 __acquires(p->pi_lock)
329 raw_spin_lock_irqsave(&p->pi_lock, *flags);
331 raw_spin_lock(&rq->lock);
332 if (likely(rq == task_rq(p)))
334 raw_spin_unlock(&rq->lock);
335 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
339 static void __task_rq_unlock(struct rq *rq)
342 raw_spin_unlock(&rq->lock);
346 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
348 __releases(p->pi_lock)
350 raw_spin_unlock(&rq->lock);
351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
355 * this_rq_lock - lock this runqueue and disable interrupts.
357 static struct rq *this_rq_lock(void)
364 raw_spin_lock(&rq->lock);
369 #ifdef CONFIG_SCHED_HRTICK
371 * Use HR-timers to deliver accurate preemption points.
374 static void hrtick_clear(struct rq *rq)
376 if (hrtimer_active(&rq->hrtick_timer))
377 hrtimer_cancel(&rq->hrtick_timer);
381 * High-resolution timer tick.
382 * Runs from hardirq context with interrupts disabled.
384 static enum hrtimer_restart hrtick(struct hrtimer *timer)
386 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
390 raw_spin_lock(&rq->lock);
392 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
393 raw_spin_unlock(&rq->lock);
395 return HRTIMER_NORESTART;
400 static int __hrtick_restart(struct rq *rq)
402 struct hrtimer *timer = &rq->hrtick_timer;
403 ktime_t time = hrtimer_get_softexpires(timer);
405 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0);
409 * called from hardirq (IPI) context
411 static void __hrtick_start(void *arg)
415 raw_spin_lock(&rq->lock);
416 __hrtick_restart(rq);
417 rq->hrtick_csd_pending = 0;
418 raw_spin_unlock(&rq->lock);
422 * Called to set the hrtick timer state.
424 * called with rq->lock held and irqs disabled
426 void hrtick_start(struct rq *rq, u64 delay)
428 struct hrtimer *timer = &rq->hrtick_timer;
429 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
431 hrtimer_set_expires(timer, time);
433 if (rq == this_rq()) {
434 __hrtick_restart(rq);
435 } else if (!rq->hrtick_csd_pending) {
436 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
437 rq->hrtick_csd_pending = 1;
442 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
444 int cpu = (int)(long)hcpu;
447 case CPU_UP_CANCELED:
448 case CPU_UP_CANCELED_FROZEN:
449 case CPU_DOWN_PREPARE:
450 case CPU_DOWN_PREPARE_FROZEN:
452 case CPU_DEAD_FROZEN:
453 hrtick_clear(cpu_rq(cpu));
460 static __init void init_hrtick(void)
462 hotcpu_notifier(hotplug_hrtick, 0);
466 * Called to set the hrtick timer state.
468 * called with rq->lock held and irqs disabled
470 void hrtick_start(struct rq *rq, u64 delay)
472 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
473 HRTIMER_MODE_REL_PINNED, 0);
476 static inline void init_hrtick(void)
479 #endif /* CONFIG_SMP */
481 static void init_rq_hrtick(struct rq *rq)
484 rq->hrtick_csd_pending = 0;
486 rq->hrtick_csd.flags = 0;
487 rq->hrtick_csd.func = __hrtick_start;
488 rq->hrtick_csd.info = rq;
491 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
492 rq->hrtick_timer.function = hrtick;
494 #else /* CONFIG_SCHED_HRTICK */
495 static inline void hrtick_clear(struct rq *rq)
499 static inline void init_rq_hrtick(struct rq *rq)
503 static inline void init_hrtick(void)
506 #endif /* CONFIG_SCHED_HRTICK */
509 * resched_task - mark a task 'to be rescheduled now'.
511 * On UP this means the setting of the need_resched flag, on SMP it
512 * might also involve a cross-CPU call to trigger the scheduler on
515 void resched_task(struct task_struct *p)
519 lockdep_assert_held(&task_rq(p)->lock);
521 if (test_tsk_need_resched(p))
524 set_tsk_need_resched(p);
527 if (cpu == smp_processor_id()) {
528 set_preempt_need_resched();
532 /* NEED_RESCHED must be visible before we test polling */
534 if (!tsk_is_polling(p))
535 smp_send_reschedule(cpu);
538 void resched_cpu(int cpu)
540 struct rq *rq = cpu_rq(cpu);
543 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
545 resched_task(cpu_curr(cpu));
546 raw_spin_unlock_irqrestore(&rq->lock, flags);
550 #ifdef CONFIG_NO_HZ_COMMON
552 * In the semi idle case, use the nearest busy cpu for migrating timers
553 * from an idle cpu. This is good for power-savings.
555 * We don't do similar optimization for completely idle system, as
556 * selecting an idle cpu will add more delays to the timers than intended
557 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
559 int get_nohz_timer_target(int pinned)
561 int cpu = smp_processor_id();
563 struct sched_domain *sd;
565 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
569 for_each_domain(cpu, sd) {
570 for_each_cpu(i, sched_domain_span(sd)) {
582 * When add_timer_on() enqueues a timer into the timer wheel of an
583 * idle CPU then this timer might expire before the next timer event
584 * which is scheduled to wake up that CPU. In case of a completely
585 * idle system the next event might even be infinite time into the
586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
587 * leaves the inner idle loop so the newly added timer is taken into
588 * account when the CPU goes back to idle and evaluates the timer
589 * wheel for the next timer event.
591 static void wake_up_idle_cpu(int cpu)
593 struct rq *rq = cpu_rq(cpu);
595 if (cpu == smp_processor_id())
599 * This is safe, as this function is called with the timer
600 * wheel base lock of (cpu) held. When the CPU is on the way
601 * to idle and has not yet set rq->curr to idle then it will
602 * be serialized on the timer wheel base lock and take the new
603 * timer into account automatically.
605 if (rq->curr != rq->idle)
609 * We can set TIF_RESCHED on the idle task of the other CPU
610 * lockless. The worst case is that the other CPU runs the
611 * idle task through an additional NOOP schedule()
613 set_tsk_need_resched(rq->idle);
615 /* NEED_RESCHED must be visible before we test polling */
617 if (!tsk_is_polling(rq->idle))
618 smp_send_reschedule(cpu);
621 static bool wake_up_full_nohz_cpu(int cpu)
623 if (tick_nohz_full_cpu(cpu)) {
624 if (cpu != smp_processor_id() ||
625 tick_nohz_tick_stopped())
626 smp_send_reschedule(cpu);
633 void wake_up_nohz_cpu(int cpu)
635 if (!wake_up_full_nohz_cpu(cpu))
636 wake_up_idle_cpu(cpu);
639 static inline bool got_nohz_idle_kick(void)
641 int cpu = smp_processor_id();
643 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
646 if (idle_cpu(cpu) && !need_resched())
650 * We can't run Idle Load Balance on this CPU for this time so we
651 * cancel it and clear NOHZ_BALANCE_KICK
653 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
657 #else /* CONFIG_NO_HZ_COMMON */
659 static inline bool got_nohz_idle_kick(void)
664 #endif /* CONFIG_NO_HZ_COMMON */
666 #ifdef CONFIG_NO_HZ_FULL
667 bool sched_can_stop_tick(void)
673 /* Make sure rq->nr_running update is visible after the IPI */
676 /* More than one running task need preemption */
677 if (rq->nr_running > 1)
682 #endif /* CONFIG_NO_HZ_FULL */
684 void sched_avg_update(struct rq *rq)
686 s64 period = sched_avg_period();
688 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
690 * Inline assembly required to prevent the compiler
691 * optimising this loop into a divmod call.
692 * See __iter_div_u64_rem() for another example of this.
694 asm("" : "+rm" (rq->age_stamp));
695 rq->age_stamp += period;
700 #endif /* CONFIG_SMP */
702 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
705 * Iterate task_group tree rooted at *from, calling @down when first entering a
706 * node and @up when leaving it for the final time.
708 * Caller must hold rcu_lock or sufficient equivalent.
710 int walk_tg_tree_from(struct task_group *from,
711 tg_visitor down, tg_visitor up, void *data)
713 struct task_group *parent, *child;
719 ret = (*down)(parent, data);
722 list_for_each_entry_rcu(child, &parent->children, siblings) {
729 ret = (*up)(parent, data);
730 if (ret || parent == from)
734 parent = parent->parent;
741 int tg_nop(struct task_group *tg, void *data)
747 static void set_load_weight(struct task_struct *p)
749 int prio = p->static_prio - MAX_RT_PRIO;
750 struct load_weight *load = &p->se.load;
753 * SCHED_IDLE tasks get minimal weight:
755 if (p->policy == SCHED_IDLE) {
756 load->weight = scale_load(WEIGHT_IDLEPRIO);
757 load->inv_weight = WMULT_IDLEPRIO;
761 load->weight = scale_load(prio_to_weight[prio]);
762 load->inv_weight = prio_to_wmult[prio];
765 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
768 sched_info_queued(rq, p);
769 p->sched_class->enqueue_task(rq, p, flags);
772 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
775 sched_info_dequeued(rq, p);
776 p->sched_class->dequeue_task(rq, p, flags);
779 void activate_task(struct rq *rq, struct task_struct *p, int flags)
781 if (task_contributes_to_load(p))
782 rq->nr_uninterruptible--;
784 enqueue_task(rq, p, flags);
787 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
789 if (task_contributes_to_load(p))
790 rq->nr_uninterruptible++;
792 dequeue_task(rq, p, flags);
795 static void update_rq_clock_task(struct rq *rq, s64 delta)
798 * In theory, the compile should just see 0 here, and optimize out the call
799 * to sched_rt_avg_update. But I don't trust it...
801 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
802 s64 steal = 0, irq_delta = 0;
804 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
808 * Since irq_time is only updated on {soft,}irq_exit, we might run into
809 * this case when a previous update_rq_clock() happened inside a
812 * When this happens, we stop ->clock_task and only update the
813 * prev_irq_time stamp to account for the part that fit, so that a next
814 * update will consume the rest. This ensures ->clock_task is
817 * It does however cause some slight miss-attribution of {soft,}irq
818 * time, a more accurate solution would be to update the irq_time using
819 * the current rq->clock timestamp, except that would require using
822 if (irq_delta > delta)
825 rq->prev_irq_time += irq_delta;
828 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
829 if (static_key_false((¶virt_steal_rq_enabled))) {
830 steal = paravirt_steal_clock(cpu_of(rq));
831 steal -= rq->prev_steal_time_rq;
833 if (unlikely(steal > delta))
836 rq->prev_steal_time_rq += steal;
841 rq->clock_task += delta;
843 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
844 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
845 sched_rt_avg_update(rq, irq_delta + steal);
849 void sched_set_stop_task(int cpu, struct task_struct *stop)
851 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
852 struct task_struct *old_stop = cpu_rq(cpu)->stop;
856 * Make it appear like a SCHED_FIFO task, its something
857 * userspace knows about and won't get confused about.
859 * Also, it will make PI more or less work without too
860 * much confusion -- but then, stop work should not
861 * rely on PI working anyway.
863 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
865 stop->sched_class = &stop_sched_class;
868 cpu_rq(cpu)->stop = stop;
872 * Reset it back to a normal scheduling class so that
873 * it can die in pieces.
875 old_stop->sched_class = &rt_sched_class;
880 * __normal_prio - return the priority that is based on the static prio
882 static inline int __normal_prio(struct task_struct *p)
884 return p->static_prio;
888 * Calculate the expected normal priority: i.e. priority
889 * without taking RT-inheritance into account. Might be
890 * boosted by interactivity modifiers. Changes upon fork,
891 * setprio syscalls, and whenever the interactivity
892 * estimator recalculates.
894 static inline int normal_prio(struct task_struct *p)
898 if (task_has_dl_policy(p))
899 prio = MAX_DL_PRIO-1;
900 else if (task_has_rt_policy(p))
901 prio = MAX_RT_PRIO-1 - p->rt_priority;
903 prio = __normal_prio(p);
908 * Calculate the current priority, i.e. the priority
909 * taken into account by the scheduler. This value might
910 * be boosted by RT tasks, or might be boosted by
911 * interactivity modifiers. Will be RT if the task got
912 * RT-boosted. If not then it returns p->normal_prio.
914 static int effective_prio(struct task_struct *p)
916 p->normal_prio = normal_prio(p);
918 * If we are RT tasks or we were boosted to RT priority,
919 * keep the priority unchanged. Otherwise, update priority
920 * to the normal priority:
922 if (!rt_prio(p->prio))
923 return p->normal_prio;
928 * task_curr - is this task currently executing on a CPU?
929 * @p: the task in question.
931 * Return: 1 if the task is currently executing. 0 otherwise.
933 inline int task_curr(const struct task_struct *p)
935 return cpu_curr(task_cpu(p)) == p;
938 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
939 const struct sched_class *prev_class,
942 if (prev_class != p->sched_class) {
943 if (prev_class->switched_from)
944 prev_class->switched_from(rq, p);
945 p->sched_class->switched_to(rq, p);
946 } else if (oldprio != p->prio || dl_task(p))
947 p->sched_class->prio_changed(rq, p, oldprio);
950 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
952 const struct sched_class *class;
954 if (p->sched_class == rq->curr->sched_class) {
955 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
957 for_each_class(class) {
958 if (class == rq->curr->sched_class)
960 if (class == p->sched_class) {
961 resched_task(rq->curr);
968 * A queue event has occurred, and we're going to schedule. In
969 * this case, we can save a useless back to back clock update.
971 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
972 rq->skip_clock_update = 1;
976 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
978 #ifdef CONFIG_SCHED_DEBUG
980 * We should never call set_task_cpu() on a blocked task,
981 * ttwu() will sort out the placement.
983 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
984 !(task_preempt_count(p) & PREEMPT_ACTIVE));
986 #ifdef CONFIG_LOCKDEP
988 * The caller should hold either p->pi_lock or rq->lock, when changing
989 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
991 * sched_move_task() holds both and thus holding either pins the cgroup,
994 * Furthermore, all task_rq users should acquire both locks, see
997 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
998 lockdep_is_held(&task_rq(p)->lock)));
1002 trace_sched_migrate_task(p, new_cpu);
1004 if (task_cpu(p) != new_cpu) {
1005 if (p->sched_class->migrate_task_rq)
1006 p->sched_class->migrate_task_rq(p, new_cpu);
1007 p->se.nr_migrations++;
1008 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1011 __set_task_cpu(p, new_cpu);
1014 static void __migrate_swap_task(struct task_struct *p, int cpu)
1017 struct rq *src_rq, *dst_rq;
1019 src_rq = task_rq(p);
1020 dst_rq = cpu_rq(cpu);
1022 deactivate_task(src_rq, p, 0);
1023 set_task_cpu(p, cpu);
1024 activate_task(dst_rq, p, 0);
1025 check_preempt_curr(dst_rq, p, 0);
1028 * Task isn't running anymore; make it appear like we migrated
1029 * it before it went to sleep. This means on wakeup we make the
1030 * previous cpu our targer instead of where it really is.
1036 struct migration_swap_arg {
1037 struct task_struct *src_task, *dst_task;
1038 int src_cpu, dst_cpu;
1041 static int migrate_swap_stop(void *data)
1043 struct migration_swap_arg *arg = data;
1044 struct rq *src_rq, *dst_rq;
1047 src_rq = cpu_rq(arg->src_cpu);
1048 dst_rq = cpu_rq(arg->dst_cpu);
1050 double_raw_lock(&arg->src_task->pi_lock,
1051 &arg->dst_task->pi_lock);
1052 double_rq_lock(src_rq, dst_rq);
1053 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1056 if (task_cpu(arg->src_task) != arg->src_cpu)
1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1065 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1066 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1071 double_rq_unlock(src_rq, dst_rq);
1072 raw_spin_unlock(&arg->dst_task->pi_lock);
1073 raw_spin_unlock(&arg->src_task->pi_lock);
1079 * Cross migrate two tasks
1081 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1083 struct migration_swap_arg arg;
1086 arg = (struct migration_swap_arg){
1088 .src_cpu = task_cpu(cur),
1090 .dst_cpu = task_cpu(p),
1093 if (arg.src_cpu == arg.dst_cpu)
1097 * These three tests are all lockless; this is OK since all of them
1098 * will be re-checked with proper locks held further down the line.
1100 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1103 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1106 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1109 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1110 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1116 struct migration_arg {
1117 struct task_struct *task;
1121 static int migration_cpu_stop(void *data);
1124 * wait_task_inactive - wait for a thread to unschedule.
1126 * If @match_state is nonzero, it's the @p->state value just checked and
1127 * not expected to change. If it changes, i.e. @p might have woken up,
1128 * then return zero. When we succeed in waiting for @p to be off its CPU,
1129 * we return a positive number (its total switch count). If a second call
1130 * a short while later returns the same number, the caller can be sure that
1131 * @p has remained unscheduled the whole time.
1133 * The caller must ensure that the task *will* unschedule sometime soon,
1134 * else this function might spin for a *long* time. This function can't
1135 * be called with interrupts off, or it may introduce deadlock with
1136 * smp_call_function() if an IPI is sent by the same process we are
1137 * waiting to become inactive.
1139 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1141 unsigned long flags;
1148 * We do the initial early heuristics without holding
1149 * any task-queue locks at all. We'll only try to get
1150 * the runqueue lock when things look like they will
1156 * If the task is actively running on another CPU
1157 * still, just relax and busy-wait without holding
1160 * NOTE! Since we don't hold any locks, it's not
1161 * even sure that "rq" stays as the right runqueue!
1162 * But we don't care, since "task_running()" will
1163 * return false if the runqueue has changed and p
1164 * is actually now running somewhere else!
1166 while (task_running(rq, p)) {
1167 if (match_state && unlikely(p->state != match_state))
1173 * Ok, time to look more closely! We need the rq
1174 * lock now, to be *sure*. If we're wrong, we'll
1175 * just go back and repeat.
1177 rq = task_rq_lock(p, &flags);
1178 trace_sched_wait_task(p);
1179 running = task_running(rq, p);
1182 if (!match_state || p->state == match_state)
1183 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1184 task_rq_unlock(rq, p, &flags);
1187 * If it changed from the expected state, bail out now.
1189 if (unlikely(!ncsw))
1193 * Was it really running after all now that we
1194 * checked with the proper locks actually held?
1196 * Oops. Go back and try again..
1198 if (unlikely(running)) {
1204 * It's not enough that it's not actively running,
1205 * it must be off the runqueue _entirely_, and not
1208 * So if it was still runnable (but just not actively
1209 * running right now), it's preempted, and we should
1210 * yield - it could be a while.
1212 if (unlikely(on_rq)) {
1213 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1215 set_current_state(TASK_UNINTERRUPTIBLE);
1216 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1221 * Ahh, all good. It wasn't running, and it wasn't
1222 * runnable, which means that it will never become
1223 * running in the future either. We're all done!
1232 * kick_process - kick a running thread to enter/exit the kernel
1233 * @p: the to-be-kicked thread
1235 * Cause a process which is running on another CPU to enter
1236 * kernel-mode, without any delay. (to get signals handled.)
1238 * NOTE: this function doesn't have to take the runqueue lock,
1239 * because all it wants to ensure is that the remote task enters
1240 * the kernel. If the IPI races and the task has been migrated
1241 * to another CPU then no harm is done and the purpose has been
1244 void kick_process(struct task_struct *p)
1250 if ((cpu != smp_processor_id()) && task_curr(p))
1251 smp_send_reschedule(cpu);
1254 EXPORT_SYMBOL_GPL(kick_process);
1255 #endif /* CONFIG_SMP */
1259 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1261 static int select_fallback_rq(int cpu, struct task_struct *p)
1263 int nid = cpu_to_node(cpu);
1264 const struct cpumask *nodemask = NULL;
1265 enum { cpuset, possible, fail } state = cpuset;
1269 * If the node that the cpu is on has been offlined, cpu_to_node()
1270 * will return -1. There is no cpu on the node, and we should
1271 * select the cpu on the other node.
1274 nodemask = cpumask_of_node(nid);
1276 /* Look for allowed, online CPU in same node. */
1277 for_each_cpu(dest_cpu, nodemask) {
1278 if (!cpu_online(dest_cpu))
1280 if (!cpu_active(dest_cpu))
1282 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1288 /* Any allowed, online CPU? */
1289 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1290 if (!cpu_online(dest_cpu))
1292 if (!cpu_active(dest_cpu))
1299 /* No more Mr. Nice Guy. */
1300 cpuset_cpus_allowed_fallback(p);
1305 do_set_cpus_allowed(p, cpu_possible_mask);
1316 if (state != cpuset) {
1318 * Don't tell them about moving exiting tasks or
1319 * kernel threads (both mm NULL), since they never
1322 if (p->mm && printk_ratelimit()) {
1323 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1324 task_pid_nr(p), p->comm, cpu);
1332 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1335 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1337 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1340 * In order not to call set_task_cpu() on a blocking task we need
1341 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1344 * Since this is common to all placement strategies, this lives here.
1346 * [ this allows ->select_task() to simply return task_cpu(p) and
1347 * not worry about this generic constraint ]
1349 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1351 cpu = select_fallback_rq(task_cpu(p), p);
1356 static void update_avg(u64 *avg, u64 sample)
1358 s64 diff = sample - *avg;
1364 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1366 #ifdef CONFIG_SCHEDSTATS
1367 struct rq *rq = this_rq();
1370 int this_cpu = smp_processor_id();
1372 if (cpu == this_cpu) {
1373 schedstat_inc(rq, ttwu_local);
1374 schedstat_inc(p, se.statistics.nr_wakeups_local);
1376 struct sched_domain *sd;
1378 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1380 for_each_domain(this_cpu, sd) {
1381 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1382 schedstat_inc(sd, ttwu_wake_remote);
1389 if (wake_flags & WF_MIGRATED)
1390 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1392 #endif /* CONFIG_SMP */
1394 schedstat_inc(rq, ttwu_count);
1395 schedstat_inc(p, se.statistics.nr_wakeups);
1397 if (wake_flags & WF_SYNC)
1398 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1400 #endif /* CONFIG_SCHEDSTATS */
1403 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1405 activate_task(rq, p, en_flags);
1408 /* if a worker is waking up, notify workqueue */
1409 if (p->flags & PF_WQ_WORKER)
1410 wq_worker_waking_up(p, cpu_of(rq));
1414 * Mark the task runnable and perform wakeup-preemption.
1417 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1419 check_preempt_curr(rq, p, wake_flags);
1420 trace_sched_wakeup(p, true);
1422 p->state = TASK_RUNNING;
1424 if (p->sched_class->task_woken)
1425 p->sched_class->task_woken(rq, p);
1427 if (rq->idle_stamp) {
1428 u64 delta = rq_clock(rq) - rq->idle_stamp;
1429 u64 max = 2*rq->max_idle_balance_cost;
1431 update_avg(&rq->avg_idle, delta);
1433 if (rq->avg_idle > max)
1442 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1445 if (p->sched_contributes_to_load)
1446 rq->nr_uninterruptible--;
1449 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1450 ttwu_do_wakeup(rq, p, wake_flags);
1454 * Called in case the task @p isn't fully descheduled from its runqueue,
1455 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1456 * since all we need to do is flip p->state to TASK_RUNNING, since
1457 * the task is still ->on_rq.
1459 static int ttwu_remote(struct task_struct *p, int wake_flags)
1464 rq = __task_rq_lock(p);
1466 /* check_preempt_curr() may use rq clock */
1467 update_rq_clock(rq);
1468 ttwu_do_wakeup(rq, p, wake_flags);
1471 __task_rq_unlock(rq);
1477 static void sched_ttwu_pending(void)
1479 struct rq *rq = this_rq();
1480 struct llist_node *llist = llist_del_all(&rq->wake_list);
1481 struct task_struct *p;
1483 raw_spin_lock(&rq->lock);
1486 p = llist_entry(llist, struct task_struct, wake_entry);
1487 llist = llist_next(llist);
1488 ttwu_do_activate(rq, p, 0);
1491 raw_spin_unlock(&rq->lock);
1494 void scheduler_ipi(void)
1497 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1498 * TIF_NEED_RESCHED remotely (for the first time) will also send
1501 preempt_fold_need_resched();
1503 if (llist_empty(&this_rq()->wake_list)
1504 && !tick_nohz_full_cpu(smp_processor_id())
1505 && !got_nohz_idle_kick())
1509 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1510 * traditionally all their work was done from the interrupt return
1511 * path. Now that we actually do some work, we need to make sure
1514 * Some archs already do call them, luckily irq_enter/exit nest
1517 * Arguably we should visit all archs and update all handlers,
1518 * however a fair share of IPIs are still resched only so this would
1519 * somewhat pessimize the simple resched case.
1522 tick_nohz_full_check();
1523 sched_ttwu_pending();
1526 * Check if someone kicked us for doing the nohz idle load balance.
1528 if (unlikely(got_nohz_idle_kick())) {
1529 this_rq()->idle_balance = 1;
1530 raise_softirq_irqoff(SCHED_SOFTIRQ);
1535 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1537 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1538 smp_send_reschedule(cpu);
1541 bool cpus_share_cache(int this_cpu, int that_cpu)
1543 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1545 #endif /* CONFIG_SMP */
1547 static void ttwu_queue(struct task_struct *p, int cpu)
1549 struct rq *rq = cpu_rq(cpu);
1551 #if defined(CONFIG_SMP)
1552 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1553 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1554 ttwu_queue_remote(p, cpu);
1559 raw_spin_lock(&rq->lock);
1560 ttwu_do_activate(rq, p, 0);
1561 raw_spin_unlock(&rq->lock);
1565 * try_to_wake_up - wake up a thread
1566 * @p: the thread to be awakened
1567 * @state: the mask of task states that can be woken
1568 * @wake_flags: wake modifier flags (WF_*)
1570 * Put it on the run-queue if it's not already there. The "current"
1571 * thread is always on the run-queue (except when the actual
1572 * re-schedule is in progress), and as such you're allowed to do
1573 * the simpler "current->state = TASK_RUNNING" to mark yourself
1574 * runnable without the overhead of this.
1576 * Return: %true if @p was woken up, %false if it was already running.
1577 * or @state didn't match @p's state.
1580 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1582 unsigned long flags;
1583 int cpu, success = 0;
1586 * If we are going to wake up a thread waiting for CONDITION we
1587 * need to ensure that CONDITION=1 done by the caller can not be
1588 * reordered with p->state check below. This pairs with mb() in
1589 * set_current_state() the waiting thread does.
1591 smp_mb__before_spinlock();
1592 raw_spin_lock_irqsave(&p->pi_lock, flags);
1593 if (!(p->state & state))
1596 success = 1; /* we're going to change ->state */
1599 if (p->on_rq && ttwu_remote(p, wake_flags))
1604 * If the owning (remote) cpu is still in the middle of schedule() with
1605 * this task as prev, wait until its done referencing the task.
1610 * Pairs with the smp_wmb() in finish_lock_switch().
1614 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1615 p->state = TASK_WAKING;
1617 if (p->sched_class->task_waking)
1618 p->sched_class->task_waking(p);
1620 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1621 if (task_cpu(p) != cpu) {
1622 wake_flags |= WF_MIGRATED;
1623 set_task_cpu(p, cpu);
1625 #endif /* CONFIG_SMP */
1629 ttwu_stat(p, cpu, wake_flags);
1631 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1637 * try_to_wake_up_local - try to wake up a local task with rq lock held
1638 * @p: the thread to be awakened
1640 * Put @p on the run-queue if it's not already there. The caller must
1641 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1644 static void try_to_wake_up_local(struct task_struct *p)
1646 struct rq *rq = task_rq(p);
1648 if (WARN_ON_ONCE(rq != this_rq()) ||
1649 WARN_ON_ONCE(p == current))
1652 lockdep_assert_held(&rq->lock);
1654 if (!raw_spin_trylock(&p->pi_lock)) {
1655 raw_spin_unlock(&rq->lock);
1656 raw_spin_lock(&p->pi_lock);
1657 raw_spin_lock(&rq->lock);
1660 if (!(p->state & TASK_NORMAL))
1664 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1666 ttwu_do_wakeup(rq, p, 0);
1667 ttwu_stat(p, smp_processor_id(), 0);
1669 raw_spin_unlock(&p->pi_lock);
1673 * wake_up_process - Wake up a specific process
1674 * @p: The process to be woken up.
1676 * Attempt to wake up the nominated process and move it to the set of runnable
1679 * Return: 1 if the process was woken up, 0 if it was already running.
1681 * It may be assumed that this function implies a write memory barrier before
1682 * changing the task state if and only if any tasks are woken up.
1684 int wake_up_process(struct task_struct *p)
1686 WARN_ON(task_is_stopped_or_traced(p));
1687 return try_to_wake_up(p, TASK_NORMAL, 0);
1689 EXPORT_SYMBOL(wake_up_process);
1691 int wake_up_state(struct task_struct *p, unsigned int state)
1693 return try_to_wake_up(p, state, 0);
1697 * Perform scheduler related setup for a newly forked process p.
1698 * p is forked by current.
1700 * __sched_fork() is basic setup used by init_idle() too:
1702 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1707 p->se.exec_start = 0;
1708 p->se.sum_exec_runtime = 0;
1709 p->se.prev_sum_exec_runtime = 0;
1710 p->se.nr_migrations = 0;
1712 INIT_LIST_HEAD(&p->se.group_node);
1714 #ifdef CONFIG_SCHEDSTATS
1715 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1718 RB_CLEAR_NODE(&p->dl.rb_node);
1719 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1720 p->dl.dl_runtime = p->dl.runtime = 0;
1721 p->dl.dl_deadline = p->dl.deadline = 0;
1722 p->dl.dl_period = 0;
1725 INIT_LIST_HEAD(&p->rt.run_list);
1727 #ifdef CONFIG_PREEMPT_NOTIFIERS
1728 INIT_HLIST_HEAD(&p->preempt_notifiers);
1731 #ifdef CONFIG_NUMA_BALANCING
1732 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1733 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1734 p->mm->numa_scan_seq = 0;
1737 if (clone_flags & CLONE_VM)
1738 p->numa_preferred_nid = current->numa_preferred_nid;
1740 p->numa_preferred_nid = -1;
1742 p->node_stamp = 0ULL;
1743 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1744 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1745 p->numa_work.next = &p->numa_work;
1746 p->numa_faults_memory = NULL;
1747 p->numa_faults_buffer_memory = NULL;
1748 p->last_task_numa_placement = 0;
1749 p->last_sum_exec_runtime = 0;
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 if (prev->sched_class->task_dead)
2153 prev->sched_class->task_dead(prev);
2156 * Remove function-return probe instances associated with this
2157 * task and put them back on the free list.
2159 kprobe_flush_task(prev);
2160 put_task_struct(prev);
2163 tick_nohz_task_switch(current);
2168 /* rq->lock is NOT held, but preemption is disabled */
2169 static inline void post_schedule(struct rq *rq)
2171 if (rq->post_schedule) {
2172 unsigned long flags;
2174 raw_spin_lock_irqsave(&rq->lock, flags);
2175 if (rq->curr->sched_class->post_schedule)
2176 rq->curr->sched_class->post_schedule(rq);
2177 raw_spin_unlock_irqrestore(&rq->lock, flags);
2179 rq->post_schedule = 0;
2185 static inline void post_schedule(struct rq *rq)
2192 * schedule_tail - first thing a freshly forked thread must call.
2193 * @prev: the thread we just switched away from.
2195 asmlinkage void schedule_tail(struct task_struct *prev)
2196 __releases(rq->lock)
2198 struct rq *rq = this_rq();
2200 finish_task_switch(rq, prev);
2203 * FIXME: do we need to worry about rq being invalidated by the
2208 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2209 /* In this case, finish_task_switch does not reenable preemption */
2212 if (current->set_child_tid)
2213 put_user(task_pid_vnr(current), current->set_child_tid);
2217 * context_switch - switch to the new MM and the new
2218 * thread's register state.
2221 context_switch(struct rq *rq, struct task_struct *prev,
2222 struct task_struct *next)
2224 struct mm_struct *mm, *oldmm;
2226 prepare_task_switch(rq, prev, next);
2229 oldmm = prev->active_mm;
2231 * For paravirt, this is coupled with an exit in switch_to to
2232 * combine the page table reload and the switch backend into
2235 arch_start_context_switch(prev);
2238 next->active_mm = oldmm;
2239 atomic_inc(&oldmm->mm_count);
2240 enter_lazy_tlb(oldmm, next);
2242 switch_mm(oldmm, mm, next);
2245 prev->active_mm = NULL;
2246 rq->prev_mm = oldmm;
2249 * Since the runqueue lock will be released by the next
2250 * task (which is an invalid locking op but in the case
2251 * of the scheduler it's an obvious special-case), so we
2252 * do an early lockdep release here:
2254 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2255 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2258 context_tracking_task_switch(prev, next);
2259 /* Here we just switch the register state and the stack. */
2260 switch_to(prev, next, prev);
2264 * this_rq must be evaluated again because prev may have moved
2265 * CPUs since it called schedule(), thus the 'rq' on its stack
2266 * frame will be invalid.
2268 finish_task_switch(this_rq(), prev);
2272 * nr_running and nr_context_switches:
2274 * externally visible scheduler statistics: current number of runnable
2275 * threads, total number of context switches performed since bootup.
2277 unsigned long nr_running(void)
2279 unsigned long i, sum = 0;
2281 for_each_online_cpu(i)
2282 sum += cpu_rq(i)->nr_running;
2287 unsigned long long nr_context_switches(void)
2290 unsigned long long sum = 0;
2292 for_each_possible_cpu(i)
2293 sum += cpu_rq(i)->nr_switches;
2298 unsigned long nr_iowait(void)
2300 unsigned long i, sum = 0;
2302 for_each_possible_cpu(i)
2303 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2308 unsigned long nr_iowait_cpu(int cpu)
2310 struct rq *this = cpu_rq(cpu);
2311 return atomic_read(&this->nr_iowait);
2317 * sched_exec - execve() is a valuable balancing opportunity, because at
2318 * this point the task has the smallest effective memory and cache footprint.
2320 void sched_exec(void)
2322 struct task_struct *p = current;
2323 unsigned long flags;
2326 raw_spin_lock_irqsave(&p->pi_lock, flags);
2327 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2328 if (dest_cpu == smp_processor_id())
2331 if (likely(cpu_active(dest_cpu))) {
2332 struct migration_arg arg = { p, dest_cpu };
2334 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2335 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2339 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2344 DEFINE_PER_CPU(struct kernel_stat, kstat);
2345 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2347 EXPORT_PER_CPU_SYMBOL(kstat);
2348 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2351 * Return any ns on the sched_clock that have not yet been accounted in
2352 * @p in case that task is currently running.
2354 * Called with task_rq_lock() held on @rq.
2356 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2360 if (task_current(rq, p)) {
2361 update_rq_clock(rq);
2362 ns = rq_clock_task(rq) - p->se.exec_start;
2370 unsigned long long task_delta_exec(struct task_struct *p)
2372 unsigned long flags;
2376 rq = task_rq_lock(p, &flags);
2377 ns = do_task_delta_exec(p, rq);
2378 task_rq_unlock(rq, p, &flags);
2384 * Return accounted runtime for the task.
2385 * In case the task is currently running, return the runtime plus current's
2386 * pending runtime that have not been accounted yet.
2388 unsigned long long task_sched_runtime(struct task_struct *p)
2390 unsigned long flags;
2394 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2396 * 64-bit doesn't need locks to atomically read a 64bit value.
2397 * So we have a optimization chance when the task's delta_exec is 0.
2398 * Reading ->on_cpu is racy, but this is ok.
2400 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2401 * If we race with it entering cpu, unaccounted time is 0. This is
2402 * indistinguishable from the read occurring a few cycles earlier.
2405 return p->se.sum_exec_runtime;
2408 rq = task_rq_lock(p, &flags);
2409 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2410 task_rq_unlock(rq, p, &flags);
2416 * This function gets called by the timer code, with HZ frequency.
2417 * We call it with interrupts disabled.
2419 void scheduler_tick(void)
2421 int cpu = smp_processor_id();
2422 struct rq *rq = cpu_rq(cpu);
2423 struct task_struct *curr = rq->curr;
2427 raw_spin_lock(&rq->lock);
2428 update_rq_clock(rq);
2429 curr->sched_class->task_tick(rq, curr, 0);
2430 update_cpu_load_active(rq);
2431 raw_spin_unlock(&rq->lock);
2433 perf_event_task_tick();
2436 rq->idle_balance = idle_cpu(cpu);
2437 trigger_load_balance(rq);
2439 rq_last_tick_reset(rq);
2442 #ifdef CONFIG_NO_HZ_FULL
2444 * scheduler_tick_max_deferment
2446 * Keep at least one tick per second when a single
2447 * active task is running because the scheduler doesn't
2448 * yet completely support full dynticks environment.
2450 * This makes sure that uptime, CFS vruntime, load
2451 * balancing, etc... continue to move forward, even
2452 * with a very low granularity.
2454 * Return: Maximum deferment in nanoseconds.
2456 u64 scheduler_tick_max_deferment(void)
2458 struct rq *rq = this_rq();
2459 unsigned long next, now = ACCESS_ONCE(jiffies);
2461 next = rq->last_sched_tick + HZ;
2463 if (time_before_eq(next, now))
2466 return jiffies_to_nsecs(next - now);
2470 notrace unsigned long get_parent_ip(unsigned long addr)
2472 if (in_lock_functions(addr)) {
2473 addr = CALLER_ADDR2;
2474 if (in_lock_functions(addr))
2475 addr = CALLER_ADDR3;
2480 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2481 defined(CONFIG_PREEMPT_TRACER))
2483 void __kprobes preempt_count_add(int val)
2485 #ifdef CONFIG_DEBUG_PREEMPT
2489 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2492 __preempt_count_add(val);
2493 #ifdef CONFIG_DEBUG_PREEMPT
2495 * Spinlock count overflowing soon?
2497 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2500 if (preempt_count() == val) {
2501 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2502 #ifdef CONFIG_DEBUG_PREEMPT
2503 current->preempt_disable_ip = ip;
2505 trace_preempt_off(CALLER_ADDR0, ip);
2508 EXPORT_SYMBOL(preempt_count_add);
2510 void __kprobes preempt_count_sub(int val)
2512 #ifdef CONFIG_DEBUG_PREEMPT
2516 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2519 * Is the spinlock portion underflowing?
2521 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2522 !(preempt_count() & PREEMPT_MASK)))
2526 if (preempt_count() == val)
2527 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2528 __preempt_count_sub(val);
2530 EXPORT_SYMBOL(preempt_count_sub);
2535 * Print scheduling while atomic bug:
2537 static noinline void __schedule_bug(struct task_struct *prev)
2539 if (oops_in_progress)
2542 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2543 prev->comm, prev->pid, preempt_count());
2545 debug_show_held_locks(prev);
2547 if (irqs_disabled())
2548 print_irqtrace_events(prev);
2549 #ifdef CONFIG_DEBUG_PREEMPT
2550 if (in_atomic_preempt_off()) {
2551 pr_err("Preemption disabled at:");
2552 print_ip_sym(current->preempt_disable_ip);
2557 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2561 * Various schedule()-time debugging checks and statistics:
2563 static inline void schedule_debug(struct task_struct *prev)
2566 * Test if we are atomic. Since do_exit() needs to call into
2567 * schedule() atomically, we ignore that path. Otherwise whine
2568 * if we are scheduling when we should not.
2570 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2571 __schedule_bug(prev);
2574 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2576 schedstat_inc(this_rq(), sched_count);
2580 * Pick up the highest-prio task:
2582 static inline struct task_struct *
2583 pick_next_task(struct rq *rq, struct task_struct *prev)
2585 const struct sched_class *class = &fair_sched_class;
2586 struct task_struct *p;
2589 * Optimization: we know that if all tasks are in
2590 * the fair class we can call that function directly:
2592 if (likely(prev->sched_class == class &&
2593 rq->nr_running == rq->cfs.h_nr_running)) {
2594 p = fair_sched_class.pick_next_task(rq, prev);
2595 if (likely(p && p != RETRY_TASK))
2600 for_each_class(class) {
2601 p = class->pick_next_task(rq, prev);
2603 if (unlikely(p == RETRY_TASK))
2609 BUG(); /* the idle class will always have a runnable task */
2613 * __schedule() is the main scheduler function.
2615 * The main means of driving the scheduler and thus entering this function are:
2617 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2619 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2620 * paths. For example, see arch/x86/entry_64.S.
2622 * To drive preemption between tasks, the scheduler sets the flag in timer
2623 * interrupt handler scheduler_tick().
2625 * 3. Wakeups don't really cause entry into schedule(). They add a
2626 * task to the run-queue and that's it.
2628 * Now, if the new task added to the run-queue preempts the current
2629 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2630 * called on the nearest possible occasion:
2632 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2634 * - in syscall or exception context, at the next outmost
2635 * preempt_enable(). (this might be as soon as the wake_up()'s
2638 * - in IRQ context, return from interrupt-handler to
2639 * preemptible context
2641 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2644 * - cond_resched() call
2645 * - explicit schedule() call
2646 * - return from syscall or exception to user-space
2647 * - return from interrupt-handler to user-space
2649 static void __sched __schedule(void)
2651 struct task_struct *prev, *next;
2652 unsigned long *switch_count;
2658 cpu = smp_processor_id();
2660 rcu_note_context_switch(cpu);
2663 schedule_debug(prev);
2665 if (sched_feat(HRTICK))
2669 * Make sure that signal_pending_state()->signal_pending() below
2670 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2671 * done by the caller to avoid the race with signal_wake_up().
2673 smp_mb__before_spinlock();
2674 raw_spin_lock_irq(&rq->lock);
2676 switch_count = &prev->nivcsw;
2677 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2678 if (unlikely(signal_pending_state(prev->state, prev))) {
2679 prev->state = TASK_RUNNING;
2681 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2685 * If a worker went to sleep, notify and ask workqueue
2686 * whether it wants to wake up a task to maintain
2689 if (prev->flags & PF_WQ_WORKER) {
2690 struct task_struct *to_wakeup;
2692 to_wakeup = wq_worker_sleeping(prev, cpu);
2694 try_to_wake_up_local(to_wakeup);
2697 switch_count = &prev->nvcsw;
2700 if (prev->on_rq || rq->skip_clock_update < 0)
2701 update_rq_clock(rq);
2703 next = pick_next_task(rq, prev);
2704 clear_tsk_need_resched(prev);
2705 clear_preempt_need_resched();
2706 rq->skip_clock_update = 0;
2708 if (likely(prev != next)) {
2713 context_switch(rq, prev, next); /* unlocks the rq */
2715 * The context switch have flipped the stack from under us
2716 * and restored the local variables which were saved when
2717 * this task called schedule() in the past. prev == current
2718 * is still correct, but it can be moved to another cpu/rq.
2720 cpu = smp_processor_id();
2723 raw_spin_unlock_irq(&rq->lock);
2727 sched_preempt_enable_no_resched();
2732 static inline void sched_submit_work(struct task_struct *tsk)
2734 if (!tsk->state || tsk_is_pi_blocked(tsk))
2737 * If we are going to sleep and we have plugged IO queued,
2738 * make sure to submit it to avoid deadlocks.
2740 if (blk_needs_flush_plug(tsk))
2741 blk_schedule_flush_plug(tsk);
2744 asmlinkage void __sched schedule(void)
2746 struct task_struct *tsk = current;
2748 sched_submit_work(tsk);
2751 EXPORT_SYMBOL(schedule);
2753 #ifdef CONFIG_CONTEXT_TRACKING
2754 asmlinkage void __sched schedule_user(void)
2757 * If we come here after a random call to set_need_resched(),
2758 * or we have been woken up remotely but the IPI has not yet arrived,
2759 * we haven't yet exited the RCU idle mode. Do it here manually until
2760 * we find a better solution.
2769 * schedule_preempt_disabled - called with preemption disabled
2771 * Returns with preemption disabled. Note: preempt_count must be 1
2773 void __sched schedule_preempt_disabled(void)
2775 sched_preempt_enable_no_resched();
2780 #ifdef CONFIG_PREEMPT
2782 * this is the entry point to schedule() from in-kernel preemption
2783 * off of preempt_enable. Kernel preemptions off return from interrupt
2784 * occur there and call schedule directly.
2786 asmlinkage void __sched notrace preempt_schedule(void)
2789 * If there is a non-zero preempt_count or interrupts are disabled,
2790 * we do not want to preempt the current task. Just return..
2792 if (likely(!preemptible()))
2796 __preempt_count_add(PREEMPT_ACTIVE);
2798 __preempt_count_sub(PREEMPT_ACTIVE);
2801 * Check again in case we missed a preemption opportunity
2802 * between schedule and now.
2805 } while (need_resched());
2807 EXPORT_SYMBOL(preempt_schedule);
2808 #endif /* CONFIG_PREEMPT */
2811 * this is the entry point to schedule() from kernel preemption
2812 * off of irq context.
2813 * Note, that this is called and return with irqs disabled. This will
2814 * protect us against recursive calling from irq.
2816 asmlinkage void __sched preempt_schedule_irq(void)
2818 enum ctx_state prev_state;
2820 /* Catch callers which need to be fixed */
2821 BUG_ON(preempt_count() || !irqs_disabled());
2823 prev_state = exception_enter();
2826 __preempt_count_add(PREEMPT_ACTIVE);
2829 local_irq_disable();
2830 __preempt_count_sub(PREEMPT_ACTIVE);
2833 * Check again in case we missed a preemption opportunity
2834 * between schedule and now.
2837 } while (need_resched());
2839 exception_exit(prev_state);
2842 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2845 return try_to_wake_up(curr->private, mode, wake_flags);
2847 EXPORT_SYMBOL(default_wake_function);
2850 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
2852 unsigned long flags;
2855 init_waitqueue_entry(&wait, current);
2857 __set_current_state(state);
2859 spin_lock_irqsave(&q->lock, flags);
2860 __add_wait_queue(q, &wait);
2861 spin_unlock(&q->lock);
2862 timeout = schedule_timeout(timeout);
2863 spin_lock_irq(&q->lock);
2864 __remove_wait_queue(q, &wait);
2865 spin_unlock_irqrestore(&q->lock, flags);
2870 void __sched interruptible_sleep_on(wait_queue_head_t *q)
2872 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2874 EXPORT_SYMBOL(interruptible_sleep_on);
2877 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
2879 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
2881 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
2883 void __sched sleep_on(wait_queue_head_t *q)
2885 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
2887 EXPORT_SYMBOL(sleep_on);
2889 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
2891 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
2893 EXPORT_SYMBOL(sleep_on_timeout);
2895 #ifdef CONFIG_RT_MUTEXES
2898 * rt_mutex_setprio - set the current priority of a task
2900 * @prio: prio value (kernel-internal form)
2902 * This function changes the 'effective' priority of a task. It does
2903 * not touch ->normal_prio like __setscheduler().
2905 * Used by the rt_mutex code to implement priority inheritance
2906 * logic. Call site only calls if the priority of the task changed.
2908 void rt_mutex_setprio(struct task_struct *p, int prio)
2910 int oldprio, on_rq, running, enqueue_flag = 0;
2912 const struct sched_class *prev_class;
2914 BUG_ON(prio > MAX_PRIO);
2916 rq = __task_rq_lock(p);
2919 * Idle task boosting is a nono in general. There is one
2920 * exception, when PREEMPT_RT and NOHZ is active:
2922 * The idle task calls get_next_timer_interrupt() and holds
2923 * the timer wheel base->lock on the CPU and another CPU wants
2924 * to access the timer (probably to cancel it). We can safely
2925 * ignore the boosting request, as the idle CPU runs this code
2926 * with interrupts disabled and will complete the lock
2927 * protected section without being interrupted. So there is no
2928 * real need to boost.
2930 if (unlikely(p == rq->idle)) {
2931 WARN_ON(p != rq->curr);
2932 WARN_ON(p->pi_blocked_on);
2936 trace_sched_pi_setprio(p, prio);
2937 p->pi_top_task = rt_mutex_get_top_task(p);
2939 prev_class = p->sched_class;
2941 running = task_current(rq, p);
2943 dequeue_task(rq, p, 0);
2945 p->sched_class->put_prev_task(rq, p);
2948 * Boosting condition are:
2949 * 1. -rt task is running and holds mutex A
2950 * --> -dl task blocks on mutex A
2952 * 2. -dl task is running and holds mutex A
2953 * --> -dl task blocks on mutex A and could preempt the
2956 if (dl_prio(prio)) {
2957 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2958 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2959 p->dl.dl_boosted = 1;
2960 p->dl.dl_throttled = 0;
2961 enqueue_flag = ENQUEUE_REPLENISH;
2963 p->dl.dl_boosted = 0;
2964 p->sched_class = &dl_sched_class;
2965 } else if (rt_prio(prio)) {
2966 if (dl_prio(oldprio))
2967 p->dl.dl_boosted = 0;
2969 enqueue_flag = ENQUEUE_HEAD;
2970 p->sched_class = &rt_sched_class;
2972 if (dl_prio(oldprio))
2973 p->dl.dl_boosted = 0;
2974 p->sched_class = &fair_sched_class;
2980 p->sched_class->set_curr_task(rq);
2982 enqueue_task(rq, p, enqueue_flag);
2984 check_class_changed(rq, p, prev_class, oldprio);
2986 __task_rq_unlock(rq);
2990 void set_user_nice(struct task_struct *p, long nice)
2992 int old_prio, delta, on_rq;
2993 unsigned long flags;
2996 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2999 * We have to be careful, if called from sys_setpriority(),
3000 * the task might be in the middle of scheduling on another CPU.
3002 rq = task_rq_lock(p, &flags);
3004 * The RT priorities are set via sched_setscheduler(), but we still
3005 * allow the 'normal' nice value to be set - but as expected
3006 * it wont have any effect on scheduling until the task is
3007 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3009 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3010 p->static_prio = NICE_TO_PRIO(nice);
3015 dequeue_task(rq, p, 0);
3017 p->static_prio = NICE_TO_PRIO(nice);
3020 p->prio = effective_prio(p);
3021 delta = p->prio - old_prio;
3024 enqueue_task(rq, p, 0);
3026 * If the task increased its priority or is running and
3027 * lowered its priority, then reschedule its CPU:
3029 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3030 resched_task(rq->curr);
3033 task_rq_unlock(rq, p, &flags);
3035 EXPORT_SYMBOL(set_user_nice);
3038 * can_nice - check if a task can reduce its nice value
3042 int can_nice(const struct task_struct *p, const int nice)
3044 /* convert nice value [19,-20] to rlimit style value [1,40] */
3045 int nice_rlim = 20 - nice;
3047 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3048 capable(CAP_SYS_NICE));
3051 #ifdef __ARCH_WANT_SYS_NICE
3054 * sys_nice - change the priority of the current process.
3055 * @increment: priority increment
3057 * sys_setpriority is a more generic, but much slower function that
3058 * does similar things.
3060 SYSCALL_DEFINE1(nice, int, increment)
3065 * Setpriority might change our priority at the same moment.
3066 * We don't have to worry. Conceptually one call occurs first
3067 * and we have a single winner.
3069 if (increment < -40)
3074 nice = task_nice(current) + increment;
3075 if (nice < MIN_NICE)
3077 if (nice > MAX_NICE)
3080 if (increment < 0 && !can_nice(current, nice))
3083 retval = security_task_setnice(current, nice);
3087 set_user_nice(current, nice);
3094 * task_prio - return the priority value of a given task.
3095 * @p: the task in question.
3097 * Return: The priority value as seen by users in /proc.
3098 * RT tasks are offset by -200. Normal tasks are centered
3099 * around 0, value goes from -16 to +15.
3101 int task_prio(const struct task_struct *p)
3103 return p->prio - MAX_RT_PRIO;
3107 * idle_cpu - is a given cpu idle currently?
3108 * @cpu: the processor in question.
3110 * Return: 1 if the CPU is currently idle. 0 otherwise.
3112 int idle_cpu(int cpu)
3114 struct rq *rq = cpu_rq(cpu);
3116 if (rq->curr != rq->idle)
3123 if (!llist_empty(&rq->wake_list))
3131 * idle_task - return the idle task for a given cpu.
3132 * @cpu: the processor in question.
3134 * Return: The idle task for the cpu @cpu.
3136 struct task_struct *idle_task(int cpu)
3138 return cpu_rq(cpu)->idle;
3142 * find_process_by_pid - find a process with a matching PID value.
3143 * @pid: the pid in question.
3145 * The task of @pid, if found. %NULL otherwise.
3147 static struct task_struct *find_process_by_pid(pid_t pid)
3149 return pid ? find_task_by_vpid(pid) : current;
3153 * This function initializes the sched_dl_entity of a newly becoming
3154 * SCHED_DEADLINE task.
3156 * Only the static values are considered here, the actual runtime and the
3157 * absolute deadline will be properly calculated when the task is enqueued
3158 * for the first time with its new policy.
3161 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3163 struct sched_dl_entity *dl_se = &p->dl;
3165 init_dl_task_timer(dl_se);
3166 dl_se->dl_runtime = attr->sched_runtime;
3167 dl_se->dl_deadline = attr->sched_deadline;
3168 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3169 dl_se->flags = attr->sched_flags;
3170 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3171 dl_se->dl_throttled = 0;
3175 static void __setscheduler_params(struct task_struct *p,
3176 const struct sched_attr *attr)
3178 int policy = attr->sched_policy;
3180 if (policy == -1) /* setparam */
3185 if (dl_policy(policy))
3186 __setparam_dl(p, attr);
3187 else if (fair_policy(policy))
3188 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3191 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3192 * !rt_policy. Always setting this ensures that things like
3193 * getparam()/getattr() don't report silly values for !rt tasks.
3195 p->rt_priority = attr->sched_priority;
3196 p->normal_prio = normal_prio(p);
3200 /* Actually do priority change: must hold pi & rq lock. */
3201 static void __setscheduler(struct rq *rq, struct task_struct *p,
3202 const struct sched_attr *attr)
3204 __setscheduler_params(p, attr);
3207 * If we get here, there was no pi waiters boosting the
3208 * task. It is safe to use the normal prio.
3210 p->prio = normal_prio(p);
3212 if (dl_prio(p->prio))
3213 p->sched_class = &dl_sched_class;
3214 else if (rt_prio(p->prio))
3215 p->sched_class = &rt_sched_class;
3217 p->sched_class = &fair_sched_class;
3221 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3223 struct sched_dl_entity *dl_se = &p->dl;
3225 attr->sched_priority = p->rt_priority;
3226 attr->sched_runtime = dl_se->dl_runtime;
3227 attr->sched_deadline = dl_se->dl_deadline;
3228 attr->sched_period = dl_se->dl_period;
3229 attr->sched_flags = dl_se->flags;
3233 * This function validates the new parameters of a -deadline task.
3234 * We ask for the deadline not being zero, and greater or equal
3235 * than the runtime, as well as the period of being zero or
3236 * greater than deadline. Furthermore, we have to be sure that
3237 * user parameters are above the internal resolution (1us); we
3238 * check sched_runtime only since it is always the smaller one.
3241 __checkparam_dl(const struct sched_attr *attr)
3243 return attr && attr->sched_deadline != 0 &&
3244 (attr->sched_period == 0 ||
3245 (s64)(attr->sched_period - attr->sched_deadline) >= 0) &&
3246 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 &&
3247 attr->sched_runtime >= (2 << (DL_SCALE - 1));
3251 * check the target process has a UID that matches the current process's
3253 static bool check_same_owner(struct task_struct *p)
3255 const struct cred *cred = current_cred(), *pcred;
3259 pcred = __task_cred(p);
3260 match = (uid_eq(cred->euid, pcred->euid) ||
3261 uid_eq(cred->euid, pcred->uid));
3266 static int __sched_setscheduler(struct task_struct *p,
3267 const struct sched_attr *attr,
3270 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3271 MAX_RT_PRIO - 1 - attr->sched_priority;
3272 int retval, oldprio, oldpolicy = -1, on_rq, running;
3273 int policy = attr->sched_policy;
3274 unsigned long flags;
3275 const struct sched_class *prev_class;
3279 /* may grab non-irq protected spin_locks */
3280 BUG_ON(in_interrupt());
3282 /* double check policy once rq lock held */
3284 reset_on_fork = p->sched_reset_on_fork;
3285 policy = oldpolicy = p->policy;
3287 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3289 if (policy != SCHED_DEADLINE &&
3290 policy != SCHED_FIFO && policy != SCHED_RR &&
3291 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3292 policy != SCHED_IDLE)
3296 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3300 * Valid priorities for SCHED_FIFO and SCHED_RR are
3301 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3302 * SCHED_BATCH and SCHED_IDLE is 0.
3304 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3305 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3307 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3308 (rt_policy(policy) != (attr->sched_priority != 0)))
3312 * Allow unprivileged RT tasks to decrease priority:
3314 if (user && !capable(CAP_SYS_NICE)) {
3315 if (fair_policy(policy)) {
3316 if (attr->sched_nice < task_nice(p) &&
3317 !can_nice(p, attr->sched_nice))
3321 if (rt_policy(policy)) {
3322 unsigned long rlim_rtprio =
3323 task_rlimit(p, RLIMIT_RTPRIO);
3325 /* can't set/change the rt policy */
3326 if (policy != p->policy && !rlim_rtprio)
3329 /* can't increase priority */
3330 if (attr->sched_priority > p->rt_priority &&
3331 attr->sched_priority > rlim_rtprio)
3336 * Can't set/change SCHED_DEADLINE policy at all for now
3337 * (safest behavior); in the future we would like to allow
3338 * unprivileged DL tasks to increase their relative deadline
3339 * or reduce their runtime (both ways reducing utilization)
3341 if (dl_policy(policy))
3345 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3346 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3348 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3349 if (!can_nice(p, task_nice(p)))
3353 /* can't change other user's priorities */
3354 if (!check_same_owner(p))
3357 /* Normal users shall not reset the sched_reset_on_fork flag */
3358 if (p->sched_reset_on_fork && !reset_on_fork)
3363 retval = security_task_setscheduler(p);
3369 * make sure no PI-waiters arrive (or leave) while we are
3370 * changing the priority of the task:
3372 * To be able to change p->policy safely, the appropriate
3373 * runqueue lock must be held.
3375 rq = task_rq_lock(p, &flags);
3378 * Changing the policy of the stop threads its a very bad idea
3380 if (p == rq->stop) {
3381 task_rq_unlock(rq, p, &flags);
3386 * If not changing anything there's no need to proceed further,
3387 * but store a possible modification of reset_on_fork.
3389 if (unlikely(policy == p->policy)) {
3390 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3392 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3394 if (dl_policy(policy))
3397 p->sched_reset_on_fork = reset_on_fork;
3398 task_rq_unlock(rq, p, &flags);
3404 #ifdef CONFIG_RT_GROUP_SCHED
3406 * Do not allow realtime tasks into groups that have no runtime
3409 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3410 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3411 !task_group_is_autogroup(task_group(p))) {
3412 task_rq_unlock(rq, p, &flags);
3417 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3418 cpumask_t *span = rq->rd->span;
3421 * Don't allow tasks with an affinity mask smaller than
3422 * the entire root_domain to become SCHED_DEADLINE. We
3423 * will also fail if there's no bandwidth available.
3425 if (!cpumask_subset(span, &p->cpus_allowed) ||
3426 rq->rd->dl_bw.bw == 0) {
3427 task_rq_unlock(rq, p, &flags);
3434 /* recheck policy now with rq lock held */
3435 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3436 policy = oldpolicy = -1;
3437 task_rq_unlock(rq, p, &flags);
3442 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3443 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3446 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3447 task_rq_unlock(rq, p, &flags);
3451 p->sched_reset_on_fork = reset_on_fork;
3455 * Special case for priority boosted tasks.
3457 * If the new priority is lower or equal (user space view)
3458 * than the current (boosted) priority, we just store the new
3459 * normal parameters and do not touch the scheduler class and
3460 * the runqueue. This will be done when the task deboost
3463 if (rt_mutex_check_prio(p, newprio)) {
3464 __setscheduler_params(p, attr);
3465 task_rq_unlock(rq, p, &flags);
3470 running = task_current(rq, p);
3472 dequeue_task(rq, p, 0);
3474 p->sched_class->put_prev_task(rq, p);
3476 prev_class = p->sched_class;
3477 __setscheduler(rq, p, attr);
3480 p->sched_class->set_curr_task(rq);
3483 * We enqueue to tail when the priority of a task is
3484 * increased (user space view).
3486 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3489 check_class_changed(rq, p, prev_class, oldprio);
3490 task_rq_unlock(rq, p, &flags);
3492 rt_mutex_adjust_pi(p);
3497 static int _sched_setscheduler(struct task_struct *p, int policy,
3498 const struct sched_param *param, bool check)
3500 struct sched_attr attr = {
3501 .sched_policy = policy,
3502 .sched_priority = param->sched_priority,
3503 .sched_nice = PRIO_TO_NICE(p->static_prio),
3507 * Fixup the legacy SCHED_RESET_ON_FORK hack
3509 if (policy & SCHED_RESET_ON_FORK) {
3510 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3511 policy &= ~SCHED_RESET_ON_FORK;
3512 attr.sched_policy = policy;
3515 return __sched_setscheduler(p, &attr, check);
3518 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3519 * @p: the task in question.
3520 * @policy: new policy.
3521 * @param: structure containing the new RT priority.
3523 * Return: 0 on success. An error code otherwise.
3525 * NOTE that the task may be already dead.
3527 int sched_setscheduler(struct task_struct *p, int policy,
3528 const struct sched_param *param)
3530 return _sched_setscheduler(p, policy, param, true);
3532 EXPORT_SYMBOL_GPL(sched_setscheduler);
3534 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3536 return __sched_setscheduler(p, attr, true);
3538 EXPORT_SYMBOL_GPL(sched_setattr);
3541 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3542 * @p: the task in question.
3543 * @policy: new policy.
3544 * @param: structure containing the new RT priority.
3546 * Just like sched_setscheduler, only don't bother checking if the
3547 * current context has permission. For example, this is needed in
3548 * stop_machine(): we create temporary high priority worker threads,
3549 * but our caller might not have that capability.
3551 * Return: 0 on success. An error code otherwise.
3553 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3554 const struct sched_param *param)
3556 return _sched_setscheduler(p, policy, param, false);
3560 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3562 struct sched_param lparam;
3563 struct task_struct *p;
3566 if (!param || pid < 0)
3568 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3573 p = find_process_by_pid(pid);
3575 retval = sched_setscheduler(p, policy, &lparam);
3582 * Mimics kernel/events/core.c perf_copy_attr().
3584 static int sched_copy_attr(struct sched_attr __user *uattr,
3585 struct sched_attr *attr)
3590 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3594 * zero the full structure, so that a short copy will be nice.
3596 memset(attr, 0, sizeof(*attr));
3598 ret = get_user(size, &uattr->size);
3602 if (size > PAGE_SIZE) /* silly large */
3605 if (!size) /* abi compat */
3606 size = SCHED_ATTR_SIZE_VER0;
3608 if (size < SCHED_ATTR_SIZE_VER0)
3612 * If we're handed a bigger struct than we know of,
3613 * ensure all the unknown bits are 0 - i.e. new
3614 * user-space does not rely on any kernel feature
3615 * extensions we dont know about yet.
3617 if (size > sizeof(*attr)) {
3618 unsigned char __user *addr;
3619 unsigned char __user *end;
3622 addr = (void __user *)uattr + sizeof(*attr);
3623 end = (void __user *)uattr + size;
3625 for (; addr < end; addr++) {
3626 ret = get_user(val, addr);
3632 size = sizeof(*attr);
3635 ret = copy_from_user(attr, uattr, size);
3640 * XXX: do we want to be lenient like existing syscalls; or do we want
3641 * to be strict and return an error on out-of-bounds values?
3643 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3649 put_user(sizeof(*attr), &uattr->size);
3655 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3656 * @pid: the pid in question.
3657 * @policy: new policy.
3658 * @param: structure containing the new RT priority.
3660 * Return: 0 on success. An error code otherwise.
3662 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3663 struct sched_param __user *, param)
3665 /* negative values for policy are not valid */
3669 return do_sched_setscheduler(pid, policy, param);
3673 * sys_sched_setparam - set/change the RT priority of a thread
3674 * @pid: the pid in question.
3675 * @param: structure containing the new RT priority.
3677 * Return: 0 on success. An error code otherwise.
3679 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3681 return do_sched_setscheduler(pid, -1, param);
3685 * sys_sched_setattr - same as above, but with extended sched_attr
3686 * @pid: the pid in question.
3687 * @uattr: structure containing the extended parameters.
3689 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3690 unsigned int, flags)
3692 struct sched_attr attr;
3693 struct task_struct *p;
3696 if (!uattr || pid < 0 || flags)
3699 if (sched_copy_attr(uattr, &attr))
3704 p = find_process_by_pid(pid);
3706 retval = sched_setattr(p, &attr);
3713 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3714 * @pid: the pid in question.
3716 * Return: On success, the policy of the thread. Otherwise, a negative error
3719 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3721 struct task_struct *p;
3729 p = find_process_by_pid(pid);
3731 retval = security_task_getscheduler(p);
3734 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3741 * sys_sched_getparam - get the RT priority of a thread
3742 * @pid: the pid in question.
3743 * @param: structure containing the RT priority.
3745 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3748 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3750 struct sched_param lp;
3751 struct task_struct *p;
3754 if (!param || pid < 0)
3758 p = find_process_by_pid(pid);
3763 retval = security_task_getscheduler(p);
3767 if (task_has_dl_policy(p)) {
3771 lp.sched_priority = p->rt_priority;
3775 * This one might sleep, we cannot do it with a spinlock held ...
3777 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3786 static int sched_read_attr(struct sched_attr __user *uattr,
3787 struct sched_attr *attr,
3792 if (!access_ok(VERIFY_WRITE, uattr, usize))
3796 * If we're handed a smaller struct than we know of,
3797 * ensure all the unknown bits are 0 - i.e. old
3798 * user-space does not get uncomplete information.
3800 if (usize < sizeof(*attr)) {
3801 unsigned char *addr;
3804 addr = (void *)attr + usize;
3805 end = (void *)attr + sizeof(*attr);
3807 for (; addr < end; addr++) {
3815 ret = copy_to_user(uattr, attr, attr->size);
3828 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3829 * @pid: the pid in question.
3830 * @uattr: structure containing the extended parameters.
3831 * @size: sizeof(attr) for fwd/bwd comp.
3833 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3834 unsigned int, size, unsigned int, flags)
3836 struct sched_attr attr = {
3837 .size = sizeof(struct sched_attr),
3839 struct task_struct *p;
3842 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3843 size < SCHED_ATTR_SIZE_VER0 || flags)
3847 p = find_process_by_pid(pid);
3852 retval = security_task_getscheduler(p);
3856 attr.sched_policy = p->policy;
3857 if (p->sched_reset_on_fork)
3858 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3859 if (task_has_dl_policy(p))
3860 __getparam_dl(p, &attr);
3861 else if (task_has_rt_policy(p))
3862 attr.sched_priority = p->rt_priority;
3864 attr.sched_nice = task_nice(p);
3868 retval = sched_read_attr(uattr, &attr, size);
3876 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3878 cpumask_var_t cpus_allowed, new_mask;
3879 struct task_struct *p;
3884 p = find_process_by_pid(pid);
3890 /* Prevent p going away */
3894 if (p->flags & PF_NO_SETAFFINITY) {
3898 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3902 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3904 goto out_free_cpus_allowed;
3907 if (!check_same_owner(p)) {
3909 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3916 retval = security_task_setscheduler(p);
3921 cpuset_cpus_allowed(p, cpus_allowed);
3922 cpumask_and(new_mask, in_mask, cpus_allowed);
3925 * Since bandwidth control happens on root_domain basis,
3926 * if admission test is enabled, we only admit -deadline
3927 * tasks allowed to run on all the CPUs in the task's
3931 if (task_has_dl_policy(p)) {
3932 const struct cpumask *span = task_rq(p)->rd->span;
3934 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3941 retval = set_cpus_allowed_ptr(p, new_mask);
3944 cpuset_cpus_allowed(p, cpus_allowed);
3945 if (!cpumask_subset(new_mask, cpus_allowed)) {
3947 * We must have raced with a concurrent cpuset
3948 * update. Just reset the cpus_allowed to the
3949 * cpuset's cpus_allowed
3951 cpumask_copy(new_mask, cpus_allowed);
3956 free_cpumask_var(new_mask);
3957 out_free_cpus_allowed:
3958 free_cpumask_var(cpus_allowed);
3964 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3965 struct cpumask *new_mask)
3967 if (len < cpumask_size())
3968 cpumask_clear(new_mask);
3969 else if (len > cpumask_size())
3970 len = cpumask_size();
3972 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3976 * sys_sched_setaffinity - set the cpu affinity of a process
3977 * @pid: pid of the process
3978 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3979 * @user_mask_ptr: user-space pointer to the new cpu mask
3981 * Return: 0 on success. An error code otherwise.
3983 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3984 unsigned long __user *, user_mask_ptr)
3986 cpumask_var_t new_mask;
3989 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
3992 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
3994 retval = sched_setaffinity(pid, new_mask);
3995 free_cpumask_var(new_mask);
3999 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4001 struct task_struct *p;
4002 unsigned long flags;
4008 p = find_process_by_pid(pid);
4012 retval = security_task_getscheduler(p);
4016 raw_spin_lock_irqsave(&p->pi_lock, flags);
4017 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4018 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4027 * sys_sched_getaffinity - get the cpu affinity of a process
4028 * @pid: pid of the process
4029 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4030 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4032 * Return: 0 on success. An error code otherwise.
4034 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4035 unsigned long __user *, user_mask_ptr)
4040 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4042 if (len & (sizeof(unsigned long)-1))
4045 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4048 ret = sched_getaffinity(pid, mask);
4050 size_t retlen = min_t(size_t, len, cpumask_size());
4052 if (copy_to_user(user_mask_ptr, mask, retlen))
4057 free_cpumask_var(mask);
4063 * sys_sched_yield - yield the current processor to other threads.
4065 * This function yields the current CPU to other tasks. If there are no
4066 * other threads running on this CPU then this function will return.
4070 SYSCALL_DEFINE0(sched_yield)
4072 struct rq *rq = this_rq_lock();
4074 schedstat_inc(rq, yld_count);
4075 current->sched_class->yield_task(rq);
4078 * Since we are going to call schedule() anyway, there's
4079 * no need to preempt or enable interrupts:
4081 __release(rq->lock);
4082 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4083 do_raw_spin_unlock(&rq->lock);
4084 sched_preempt_enable_no_resched();
4091 static void __cond_resched(void)
4093 __preempt_count_add(PREEMPT_ACTIVE);
4095 __preempt_count_sub(PREEMPT_ACTIVE);
4098 int __sched _cond_resched(void)
4100 if (should_resched()) {
4106 EXPORT_SYMBOL(_cond_resched);
4109 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4110 * call schedule, and on return reacquire the lock.
4112 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4113 * operations here to prevent schedule() from being called twice (once via
4114 * spin_unlock(), once by hand).
4116 int __cond_resched_lock(spinlock_t *lock)
4118 int resched = should_resched();
4121 lockdep_assert_held(lock);
4123 if (spin_needbreak(lock) || resched) {
4134 EXPORT_SYMBOL(__cond_resched_lock);
4136 int __sched __cond_resched_softirq(void)
4138 BUG_ON(!in_softirq());
4140 if (should_resched()) {
4148 EXPORT_SYMBOL(__cond_resched_softirq);
4151 * yield - yield the current processor to other threads.
4153 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4155 * The scheduler is at all times free to pick the calling task as the most
4156 * eligible task to run, if removing the yield() call from your code breaks
4157 * it, its already broken.
4159 * Typical broken usage is:
4164 * where one assumes that yield() will let 'the other' process run that will
4165 * make event true. If the current task is a SCHED_FIFO task that will never
4166 * happen. Never use yield() as a progress guarantee!!
4168 * If you want to use yield() to wait for something, use wait_event().
4169 * If you want to use yield() to be 'nice' for others, use cond_resched().
4170 * If you still want to use yield(), do not!
4172 void __sched yield(void)
4174 set_current_state(TASK_RUNNING);
4177 EXPORT_SYMBOL(yield);
4180 * yield_to - yield the current processor to another thread in
4181 * your thread group, or accelerate that thread toward the
4182 * processor it's on.
4184 * @preempt: whether task preemption is allowed or not
4186 * It's the caller's job to ensure that the target task struct
4187 * can't go away on us before we can do any checks.
4190 * true (>0) if we indeed boosted the target task.
4191 * false (0) if we failed to boost the target.
4192 * -ESRCH if there's no task to yield to.
4194 bool __sched yield_to(struct task_struct *p, bool preempt)
4196 struct task_struct *curr = current;
4197 struct rq *rq, *p_rq;
4198 unsigned long flags;
4201 local_irq_save(flags);
4207 * If we're the only runnable task on the rq and target rq also
4208 * has only one task, there's absolutely no point in yielding.
4210 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4215 double_rq_lock(rq, p_rq);
4216 if (task_rq(p) != p_rq) {
4217 double_rq_unlock(rq, p_rq);
4221 if (!curr->sched_class->yield_to_task)
4224 if (curr->sched_class != p->sched_class)
4227 if (task_running(p_rq, p) || p->state)
4230 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4232 schedstat_inc(rq, yld_count);
4234 * Make p's CPU reschedule; pick_next_entity takes care of
4237 if (preempt && rq != p_rq)
4238 resched_task(p_rq->curr);
4242 double_rq_unlock(rq, p_rq);
4244 local_irq_restore(flags);
4251 EXPORT_SYMBOL_GPL(yield_to);
4254 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4255 * that process accounting knows that this is a task in IO wait state.
4257 void __sched io_schedule(void)
4259 struct rq *rq = raw_rq();
4261 delayacct_blkio_start();
4262 atomic_inc(&rq->nr_iowait);
4263 blk_flush_plug(current);
4264 current->in_iowait = 1;
4266 current->in_iowait = 0;
4267 atomic_dec(&rq->nr_iowait);
4268 delayacct_blkio_end();
4270 EXPORT_SYMBOL(io_schedule);
4272 long __sched io_schedule_timeout(long timeout)
4274 struct rq *rq = raw_rq();
4277 delayacct_blkio_start();
4278 atomic_inc(&rq->nr_iowait);
4279 blk_flush_plug(current);
4280 current->in_iowait = 1;
4281 ret = schedule_timeout(timeout);
4282 current->in_iowait = 0;
4283 atomic_dec(&rq->nr_iowait);
4284 delayacct_blkio_end();
4289 * sys_sched_get_priority_max - return maximum RT priority.
4290 * @policy: scheduling class.
4292 * Return: On success, this syscall returns the maximum
4293 * rt_priority that can be used by a given scheduling class.
4294 * On failure, a negative error code is returned.
4296 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4303 ret = MAX_USER_RT_PRIO-1;
4305 case SCHED_DEADLINE:
4316 * sys_sched_get_priority_min - return minimum RT priority.
4317 * @policy: scheduling class.
4319 * Return: On success, this syscall returns the minimum
4320 * rt_priority that can be used by a given scheduling class.
4321 * On failure, a negative error code is returned.
4323 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4332 case SCHED_DEADLINE:
4342 * sys_sched_rr_get_interval - return the default timeslice of a process.
4343 * @pid: pid of the process.
4344 * @interval: userspace pointer to the timeslice value.
4346 * this syscall writes the default timeslice value of a given process
4347 * into the user-space timespec buffer. A value of '0' means infinity.
4349 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4352 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4353 struct timespec __user *, interval)
4355 struct task_struct *p;
4356 unsigned int time_slice;
4357 unsigned long flags;
4367 p = find_process_by_pid(pid);
4371 retval = security_task_getscheduler(p);
4375 rq = task_rq_lock(p, &flags);
4377 if (p->sched_class->get_rr_interval)
4378 time_slice = p->sched_class->get_rr_interval(rq, p);
4379 task_rq_unlock(rq, p, &flags);
4382 jiffies_to_timespec(time_slice, &t);
4383 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4391 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4393 void sched_show_task(struct task_struct *p)
4395 unsigned long free = 0;
4399 state = p->state ? __ffs(p->state) + 1 : 0;
4400 printk(KERN_INFO "%-15.15s %c", p->comm,
4401 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4402 #if BITS_PER_LONG == 32
4403 if (state == TASK_RUNNING)
4404 printk(KERN_CONT " running ");
4406 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4408 if (state == TASK_RUNNING)
4409 printk(KERN_CONT " running task ");
4411 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4413 #ifdef CONFIG_DEBUG_STACK_USAGE
4414 free = stack_not_used(p);
4417 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4419 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4420 task_pid_nr(p), ppid,
4421 (unsigned long)task_thread_info(p)->flags);
4423 print_worker_info(KERN_INFO, p);
4424 show_stack(p, NULL);
4427 void show_state_filter(unsigned long state_filter)
4429 struct task_struct *g, *p;
4431 #if BITS_PER_LONG == 32
4433 " task PC stack pid father\n");
4436 " task PC stack pid father\n");
4439 do_each_thread(g, p) {
4441 * reset the NMI-timeout, listing all files on a slow
4442 * console might take a lot of time:
4444 touch_nmi_watchdog();
4445 if (!state_filter || (p->state & state_filter))
4447 } while_each_thread(g, p);
4449 touch_all_softlockup_watchdogs();
4451 #ifdef CONFIG_SCHED_DEBUG
4452 sysrq_sched_debug_show();
4456 * Only show locks if all tasks are dumped:
4459 debug_show_all_locks();
4462 void init_idle_bootup_task(struct task_struct *idle)
4464 idle->sched_class = &idle_sched_class;
4468 * init_idle - set up an idle thread for a given CPU
4469 * @idle: task in question
4470 * @cpu: cpu the idle task belongs to
4472 * NOTE: this function does not set the idle thread's NEED_RESCHED
4473 * flag, to make booting more robust.
4475 void init_idle(struct task_struct *idle, int cpu)
4477 struct rq *rq = cpu_rq(cpu);
4478 unsigned long flags;
4480 raw_spin_lock_irqsave(&rq->lock, flags);
4482 __sched_fork(0, idle);
4483 idle->state = TASK_RUNNING;
4484 idle->se.exec_start = sched_clock();
4486 do_set_cpus_allowed(idle, cpumask_of(cpu));
4488 * We're having a chicken and egg problem, even though we are
4489 * holding rq->lock, the cpu isn't yet set to this cpu so the
4490 * lockdep check in task_group() will fail.
4492 * Similar case to sched_fork(). / Alternatively we could
4493 * use task_rq_lock() here and obtain the other rq->lock.
4498 __set_task_cpu(idle, cpu);
4501 rq->curr = rq->idle = idle;
4503 #if defined(CONFIG_SMP)
4506 raw_spin_unlock_irqrestore(&rq->lock, flags);
4508 /* Set the preempt count _outside_ the spinlocks! */
4509 init_idle_preempt_count(idle, cpu);
4512 * The idle tasks have their own, simple scheduling class:
4514 idle->sched_class = &idle_sched_class;
4515 ftrace_graph_init_idle_task(idle, cpu);
4516 vtime_init_idle(idle, cpu);
4517 #if defined(CONFIG_SMP)
4518 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4523 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4525 if (p->sched_class && p->sched_class->set_cpus_allowed)
4526 p->sched_class->set_cpus_allowed(p, new_mask);
4528 cpumask_copy(&p->cpus_allowed, new_mask);
4529 p->nr_cpus_allowed = cpumask_weight(new_mask);
4533 * This is how migration works:
4535 * 1) we invoke migration_cpu_stop() on the target CPU using
4537 * 2) stopper starts to run (implicitly forcing the migrated thread
4539 * 3) it checks whether the migrated task is still in the wrong runqueue.
4540 * 4) if it's in the wrong runqueue then the migration thread removes
4541 * it and puts it into the right queue.
4542 * 5) stopper completes and stop_one_cpu() returns and the migration
4547 * Change a given task's CPU affinity. Migrate the thread to a
4548 * proper CPU and schedule it away if the CPU it's executing on
4549 * is removed from the allowed bitmask.
4551 * NOTE: the caller must have a valid reference to the task, the
4552 * task must not exit() & deallocate itself prematurely. The
4553 * call is not atomic; no spinlocks may be held.
4555 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4557 unsigned long flags;
4559 unsigned int dest_cpu;
4562 rq = task_rq_lock(p, &flags);
4564 if (cpumask_equal(&p->cpus_allowed, new_mask))
4567 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4572 do_set_cpus_allowed(p, new_mask);
4574 /* Can the task run on the task's current CPU? If so, we're done */
4575 if (cpumask_test_cpu(task_cpu(p), new_mask))
4578 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4580 struct migration_arg arg = { p, dest_cpu };
4581 /* Need help from migration thread: drop lock and wait. */
4582 task_rq_unlock(rq, p, &flags);
4583 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4584 tlb_migrate_finish(p->mm);
4588 task_rq_unlock(rq, p, &flags);
4592 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4595 * Move (not current) task off this cpu, onto dest cpu. We're doing
4596 * this because either it can't run here any more (set_cpus_allowed()
4597 * away from this CPU, or CPU going down), or because we're
4598 * attempting to rebalance this task on exec (sched_exec).
4600 * So we race with normal scheduler movements, but that's OK, as long
4601 * as the task is no longer on this CPU.
4603 * Returns non-zero if task was successfully migrated.
4605 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4607 struct rq *rq_dest, *rq_src;
4610 if (unlikely(!cpu_active(dest_cpu)))
4613 rq_src = cpu_rq(src_cpu);
4614 rq_dest = cpu_rq(dest_cpu);
4616 raw_spin_lock(&p->pi_lock);
4617 double_rq_lock(rq_src, rq_dest);
4618 /* Already moved. */
4619 if (task_cpu(p) != src_cpu)
4621 /* Affinity changed (again). */
4622 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4626 * If we're not on a rq, the next wake-up will ensure we're
4630 dequeue_task(rq_src, p, 0);
4631 set_task_cpu(p, dest_cpu);
4632 enqueue_task(rq_dest, p, 0);
4633 check_preempt_curr(rq_dest, p, 0);
4638 double_rq_unlock(rq_src, rq_dest);
4639 raw_spin_unlock(&p->pi_lock);
4643 #ifdef CONFIG_NUMA_BALANCING
4644 /* Migrate current task p to target_cpu */
4645 int migrate_task_to(struct task_struct *p, int target_cpu)
4647 struct migration_arg arg = { p, target_cpu };
4648 int curr_cpu = task_cpu(p);
4650 if (curr_cpu == target_cpu)
4653 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4656 /* TODO: This is not properly updating schedstats */
4658 trace_sched_move_numa(p, curr_cpu, target_cpu);
4659 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4663 * Requeue a task on a given node and accurately track the number of NUMA
4664 * tasks on the runqueues
4666 void sched_setnuma(struct task_struct *p, int nid)
4669 unsigned long flags;
4670 bool on_rq, running;
4672 rq = task_rq_lock(p, &flags);
4674 running = task_current(rq, p);
4677 dequeue_task(rq, p, 0);
4679 p->sched_class->put_prev_task(rq, p);
4681 p->numa_preferred_nid = nid;
4684 p->sched_class->set_curr_task(rq);
4686 enqueue_task(rq, p, 0);
4687 task_rq_unlock(rq, p, &flags);
4692 * migration_cpu_stop - this will be executed by a highprio stopper thread
4693 * and performs thread migration by bumping thread off CPU then
4694 * 'pushing' onto another runqueue.
4696 static int migration_cpu_stop(void *data)
4698 struct migration_arg *arg = data;
4701 * The original target cpu might have gone down and we might
4702 * be on another cpu but it doesn't matter.
4704 local_irq_disable();
4705 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4710 #ifdef CONFIG_HOTPLUG_CPU
4713 * Ensures that the idle task is using init_mm right before its cpu goes
4716 void idle_task_exit(void)
4718 struct mm_struct *mm = current->active_mm;
4720 BUG_ON(cpu_online(smp_processor_id()));
4722 if (mm != &init_mm) {
4723 switch_mm(mm, &init_mm, current);
4724 finish_arch_post_lock_switch();
4730 * Since this CPU is going 'away' for a while, fold any nr_active delta
4731 * we might have. Assumes we're called after migrate_tasks() so that the
4732 * nr_active count is stable.
4734 * Also see the comment "Global load-average calculations".
4736 static void calc_load_migrate(struct rq *rq)
4738 long delta = calc_load_fold_active(rq);
4740 atomic_long_add(delta, &calc_load_tasks);
4743 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4747 static const struct sched_class fake_sched_class = {
4748 .put_prev_task = put_prev_task_fake,
4751 static struct task_struct fake_task = {
4753 * Avoid pull_{rt,dl}_task()
4755 .prio = MAX_PRIO + 1,
4756 .sched_class = &fake_sched_class,
4760 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4761 * try_to_wake_up()->select_task_rq().
4763 * Called with rq->lock held even though we'er in stop_machine() and
4764 * there's no concurrency possible, we hold the required locks anyway
4765 * because of lock validation efforts.
4767 static void migrate_tasks(unsigned int dead_cpu)
4769 struct rq *rq = cpu_rq(dead_cpu);
4770 struct task_struct *next, *stop = rq->stop;
4774 * Fudge the rq selection such that the below task selection loop
4775 * doesn't get stuck on the currently eligible stop task.
4777 * We're currently inside stop_machine() and the rq is either stuck
4778 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4779 * either way we should never end up calling schedule() until we're
4785 * put_prev_task() and pick_next_task() sched
4786 * class method both need to have an up-to-date
4787 * value of rq->clock[_task]
4789 update_rq_clock(rq);
4793 * There's this thread running, bail when that's the only
4796 if (rq->nr_running == 1)
4799 next = pick_next_task(rq, &fake_task);
4801 next->sched_class->put_prev_task(rq, next);
4803 /* Find suitable destination for @next, with force if needed. */
4804 dest_cpu = select_fallback_rq(dead_cpu, next);
4805 raw_spin_unlock(&rq->lock);
4807 __migrate_task(next, dead_cpu, dest_cpu);
4809 raw_spin_lock(&rq->lock);
4815 #endif /* CONFIG_HOTPLUG_CPU */
4817 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4819 static struct ctl_table sd_ctl_dir[] = {
4821 .procname = "sched_domain",
4827 static struct ctl_table sd_ctl_root[] = {
4829 .procname = "kernel",
4831 .child = sd_ctl_dir,
4836 static struct ctl_table *sd_alloc_ctl_entry(int n)
4838 struct ctl_table *entry =
4839 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4844 static void sd_free_ctl_entry(struct ctl_table **tablep)
4846 struct ctl_table *entry;
4849 * In the intermediate directories, both the child directory and
4850 * procname are dynamically allocated and could fail but the mode
4851 * will always be set. In the lowest directory the names are
4852 * static strings and all have proc handlers.
4854 for (entry = *tablep; entry->mode; entry++) {
4856 sd_free_ctl_entry(&entry->child);
4857 if (entry->proc_handler == NULL)
4858 kfree(entry->procname);
4865 static int min_load_idx = 0;
4866 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4869 set_table_entry(struct ctl_table *entry,
4870 const char *procname, void *data, int maxlen,
4871 umode_t mode, proc_handler *proc_handler,
4874 entry->procname = procname;
4876 entry->maxlen = maxlen;
4878 entry->proc_handler = proc_handler;
4881 entry->extra1 = &min_load_idx;
4882 entry->extra2 = &max_load_idx;
4886 static struct ctl_table *
4887 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4889 struct ctl_table *table = sd_alloc_ctl_entry(14);
4894 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4895 sizeof(long), 0644, proc_doulongvec_minmax, false);
4896 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4897 sizeof(long), 0644, proc_doulongvec_minmax, false);
4898 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4899 sizeof(int), 0644, proc_dointvec_minmax, true);
4900 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4901 sizeof(int), 0644, proc_dointvec_minmax, true);
4902 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4903 sizeof(int), 0644, proc_dointvec_minmax, true);
4904 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4905 sizeof(int), 0644, proc_dointvec_minmax, true);
4906 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4907 sizeof(int), 0644, proc_dointvec_minmax, true);
4908 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4909 sizeof(int), 0644, proc_dointvec_minmax, false);
4910 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4911 sizeof(int), 0644, proc_dointvec_minmax, false);
4912 set_table_entry(&table[9], "cache_nice_tries",
4913 &sd->cache_nice_tries,
4914 sizeof(int), 0644, proc_dointvec_minmax, false);
4915 set_table_entry(&table[10], "flags", &sd->flags,
4916 sizeof(int), 0644, proc_dointvec_minmax, false);
4917 set_table_entry(&table[11], "max_newidle_lb_cost",
4918 &sd->max_newidle_lb_cost,
4919 sizeof(long), 0644, proc_doulongvec_minmax, false);
4920 set_table_entry(&table[12], "name", sd->name,
4921 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4922 /* &table[13] is terminator */
4927 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4929 struct ctl_table *entry, *table;
4930 struct sched_domain *sd;
4931 int domain_num = 0, i;
4934 for_each_domain(cpu, sd)
4936 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4941 for_each_domain(cpu, sd) {
4942 snprintf(buf, 32, "domain%d", i);
4943 entry->procname = kstrdup(buf, GFP_KERNEL);
4945 entry->child = sd_alloc_ctl_domain_table(sd);
4952 static struct ctl_table_header *sd_sysctl_header;
4953 static void register_sched_domain_sysctl(void)
4955 int i, cpu_num = num_possible_cpus();
4956 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4959 WARN_ON(sd_ctl_dir[0].child);
4960 sd_ctl_dir[0].child = entry;
4965 for_each_possible_cpu(i) {
4966 snprintf(buf, 32, "cpu%d", i);
4967 entry->procname = kstrdup(buf, GFP_KERNEL);
4969 entry->child = sd_alloc_ctl_cpu_table(i);
4973 WARN_ON(sd_sysctl_header);
4974 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4977 /* may be called multiple times per register */
4978 static void unregister_sched_domain_sysctl(void)
4980 if (sd_sysctl_header)
4981 unregister_sysctl_table(sd_sysctl_header);
4982 sd_sysctl_header = NULL;
4983 if (sd_ctl_dir[0].child)
4984 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4987 static void register_sched_domain_sysctl(void)
4990 static void unregister_sched_domain_sysctl(void)
4995 static void set_rq_online(struct rq *rq)
4998 const struct sched_class *class;
5000 cpumask_set_cpu(rq->cpu, rq->rd->online);
5003 for_each_class(class) {
5004 if (class->rq_online)
5005 class->rq_online(rq);
5010 static void set_rq_offline(struct rq *rq)
5013 const struct sched_class *class;
5015 for_each_class(class) {
5016 if (class->rq_offline)
5017 class->rq_offline(rq);
5020 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5026 * migration_call - callback that gets triggered when a CPU is added.
5027 * Here we can start up the necessary migration thread for the new CPU.
5030 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5032 int cpu = (long)hcpu;
5033 unsigned long flags;
5034 struct rq *rq = cpu_rq(cpu);
5036 switch (action & ~CPU_TASKS_FROZEN) {
5038 case CPU_UP_PREPARE:
5039 rq->calc_load_update = calc_load_update;
5043 /* Update our root-domain */
5044 raw_spin_lock_irqsave(&rq->lock, flags);
5046 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5050 raw_spin_unlock_irqrestore(&rq->lock, flags);
5053 #ifdef CONFIG_HOTPLUG_CPU
5055 sched_ttwu_pending();
5056 /* Update our root-domain */
5057 raw_spin_lock_irqsave(&rq->lock, flags);
5059 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5063 BUG_ON(rq->nr_running != 1); /* the migration thread */
5064 raw_spin_unlock_irqrestore(&rq->lock, flags);
5068 calc_load_migrate(rq);
5073 update_max_interval();
5079 * Register at high priority so that task migration (migrate_all_tasks)
5080 * happens before everything else. This has to be lower priority than
5081 * the notifier in the perf_event subsystem, though.
5083 static struct notifier_block migration_notifier = {
5084 .notifier_call = migration_call,
5085 .priority = CPU_PRI_MIGRATION,
5088 static int sched_cpu_active(struct notifier_block *nfb,
5089 unsigned long action, void *hcpu)
5091 switch (action & ~CPU_TASKS_FROZEN) {
5093 case CPU_DOWN_FAILED:
5094 set_cpu_active((long)hcpu, true);
5101 static int sched_cpu_inactive(struct notifier_block *nfb,
5102 unsigned long action, void *hcpu)
5104 unsigned long flags;
5105 long cpu = (long)hcpu;
5107 switch (action & ~CPU_TASKS_FROZEN) {
5108 case CPU_DOWN_PREPARE:
5109 set_cpu_active(cpu, false);
5111 /* explicitly allow suspend */
5112 if (!(action & CPU_TASKS_FROZEN)) {
5113 struct dl_bw *dl_b = dl_bw_of(cpu);
5117 raw_spin_lock_irqsave(&dl_b->lock, flags);
5118 cpus = dl_bw_cpus(cpu);
5119 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5120 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5123 return notifier_from_errno(-EBUSY);
5131 static int __init migration_init(void)
5133 void *cpu = (void *)(long)smp_processor_id();
5136 /* Initialize migration for the boot CPU */
5137 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5138 BUG_ON(err == NOTIFY_BAD);
5139 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5140 register_cpu_notifier(&migration_notifier);
5142 /* Register cpu active notifiers */
5143 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5144 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5148 early_initcall(migration_init);
5153 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5155 #ifdef CONFIG_SCHED_DEBUG
5157 static __read_mostly int sched_debug_enabled;
5159 static int __init sched_debug_setup(char *str)
5161 sched_debug_enabled = 1;
5165 early_param("sched_debug", sched_debug_setup);
5167 static inline bool sched_debug(void)
5169 return sched_debug_enabled;
5172 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5173 struct cpumask *groupmask)
5175 struct sched_group *group = sd->groups;
5178 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5179 cpumask_clear(groupmask);
5181 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5183 if (!(sd->flags & SD_LOAD_BALANCE)) {
5184 printk("does not load-balance\n");
5186 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5191 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5193 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5194 printk(KERN_ERR "ERROR: domain->span does not contain "
5197 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5198 printk(KERN_ERR "ERROR: domain->groups does not contain"
5202 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5206 printk(KERN_ERR "ERROR: group is NULL\n");
5211 * Even though we initialize ->power to something semi-sane,
5212 * we leave power_orig unset. This allows us to detect if
5213 * domain iteration is still funny without causing /0 traps.
5215 if (!group->sgp->power_orig) {
5216 printk(KERN_CONT "\n");
5217 printk(KERN_ERR "ERROR: domain->cpu_power not "
5222 if (!cpumask_weight(sched_group_cpus(group))) {
5223 printk(KERN_CONT "\n");
5224 printk(KERN_ERR "ERROR: empty group\n");
5228 if (!(sd->flags & SD_OVERLAP) &&
5229 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5230 printk(KERN_CONT "\n");
5231 printk(KERN_ERR "ERROR: repeated CPUs\n");
5235 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5237 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5239 printk(KERN_CONT " %s", str);
5240 if (group->sgp->power != SCHED_POWER_SCALE) {
5241 printk(KERN_CONT " (cpu_power = %d)",
5245 group = group->next;
5246 } while (group != sd->groups);
5247 printk(KERN_CONT "\n");
5249 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5250 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5253 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5254 printk(KERN_ERR "ERROR: parent span is not a superset "
5255 "of domain->span\n");
5259 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5263 if (!sched_debug_enabled)
5267 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5271 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5274 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5282 #else /* !CONFIG_SCHED_DEBUG */
5283 # define sched_domain_debug(sd, cpu) do { } while (0)
5284 static inline bool sched_debug(void)
5288 #endif /* CONFIG_SCHED_DEBUG */
5290 static int sd_degenerate(struct sched_domain *sd)
5292 if (cpumask_weight(sched_domain_span(sd)) == 1)
5295 /* Following flags need at least 2 groups */
5296 if (sd->flags & (SD_LOAD_BALANCE |
5297 SD_BALANCE_NEWIDLE |
5301 SD_SHARE_PKG_RESOURCES)) {
5302 if (sd->groups != sd->groups->next)
5306 /* Following flags don't use groups */
5307 if (sd->flags & (SD_WAKE_AFFINE))
5314 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5316 unsigned long cflags = sd->flags, pflags = parent->flags;
5318 if (sd_degenerate(parent))
5321 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5324 /* Flags needing groups don't count if only 1 group in parent */
5325 if (parent->groups == parent->groups->next) {
5326 pflags &= ~(SD_LOAD_BALANCE |
5327 SD_BALANCE_NEWIDLE |
5331 SD_SHARE_PKG_RESOURCES |
5333 if (nr_node_ids == 1)
5334 pflags &= ~SD_SERIALIZE;
5336 if (~cflags & pflags)
5342 static void free_rootdomain(struct rcu_head *rcu)
5344 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5346 cpupri_cleanup(&rd->cpupri);
5347 cpudl_cleanup(&rd->cpudl);
5348 free_cpumask_var(rd->dlo_mask);
5349 free_cpumask_var(rd->rto_mask);
5350 free_cpumask_var(rd->online);
5351 free_cpumask_var(rd->span);
5355 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5357 struct root_domain *old_rd = NULL;
5358 unsigned long flags;
5360 raw_spin_lock_irqsave(&rq->lock, flags);
5365 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5368 cpumask_clear_cpu(rq->cpu, old_rd->span);
5371 * If we dont want to free the old_rd yet then
5372 * set old_rd to NULL to skip the freeing later
5375 if (!atomic_dec_and_test(&old_rd->refcount))
5379 atomic_inc(&rd->refcount);
5382 cpumask_set_cpu(rq->cpu, rd->span);
5383 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5386 raw_spin_unlock_irqrestore(&rq->lock, flags);
5389 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5392 static int init_rootdomain(struct root_domain *rd)
5394 memset(rd, 0, sizeof(*rd));
5396 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5398 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5400 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5402 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5405 init_dl_bw(&rd->dl_bw);
5406 if (cpudl_init(&rd->cpudl) != 0)
5409 if (cpupri_init(&rd->cpupri) != 0)
5414 free_cpumask_var(rd->rto_mask);
5416 free_cpumask_var(rd->dlo_mask);
5418 free_cpumask_var(rd->online);
5420 free_cpumask_var(rd->span);
5426 * By default the system creates a single root-domain with all cpus as
5427 * members (mimicking the global state we have today).
5429 struct root_domain def_root_domain;
5431 static void init_defrootdomain(void)
5433 init_rootdomain(&def_root_domain);
5435 atomic_set(&def_root_domain.refcount, 1);
5438 static struct root_domain *alloc_rootdomain(void)
5440 struct root_domain *rd;
5442 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5446 if (init_rootdomain(rd) != 0) {
5454 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5456 struct sched_group *tmp, *first;
5465 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5470 } while (sg != first);
5473 static void free_sched_domain(struct rcu_head *rcu)
5475 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5478 * If its an overlapping domain it has private groups, iterate and
5481 if (sd->flags & SD_OVERLAP) {
5482 free_sched_groups(sd->groups, 1);
5483 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5484 kfree(sd->groups->sgp);
5490 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5492 call_rcu(&sd->rcu, free_sched_domain);
5495 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5497 for (; sd; sd = sd->parent)
5498 destroy_sched_domain(sd, cpu);
5502 * Keep a special pointer to the highest sched_domain that has
5503 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5504 * allows us to avoid some pointer chasing select_idle_sibling().
5506 * Also keep a unique ID per domain (we use the first cpu number in
5507 * the cpumask of the domain), this allows us to quickly tell if
5508 * two cpus are in the same cache domain, see cpus_share_cache().
5510 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5511 DEFINE_PER_CPU(int, sd_llc_size);
5512 DEFINE_PER_CPU(int, sd_llc_id);
5513 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5514 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5515 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5517 static void update_top_cache_domain(int cpu)
5519 struct sched_domain *sd;
5520 struct sched_domain *busy_sd = NULL;
5524 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5526 id = cpumask_first(sched_domain_span(sd));
5527 size = cpumask_weight(sched_domain_span(sd));
5528 busy_sd = sd->parent; /* sd_busy */
5530 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5532 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5533 per_cpu(sd_llc_size, cpu) = size;
5534 per_cpu(sd_llc_id, cpu) = id;
5536 sd = lowest_flag_domain(cpu, SD_NUMA);
5537 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5539 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5540 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5544 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5545 * hold the hotplug lock.
5548 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5550 struct rq *rq = cpu_rq(cpu);
5551 struct sched_domain *tmp;
5553 /* Remove the sched domains which do not contribute to scheduling. */
5554 for (tmp = sd; tmp; ) {
5555 struct sched_domain *parent = tmp->parent;
5559 if (sd_parent_degenerate(tmp, parent)) {
5560 tmp->parent = parent->parent;
5562 parent->parent->child = tmp;
5564 * Transfer SD_PREFER_SIBLING down in case of a
5565 * degenerate parent; the spans match for this
5566 * so the property transfers.
5568 if (parent->flags & SD_PREFER_SIBLING)
5569 tmp->flags |= SD_PREFER_SIBLING;
5570 destroy_sched_domain(parent, cpu);
5575 if (sd && sd_degenerate(sd)) {
5578 destroy_sched_domain(tmp, cpu);
5583 sched_domain_debug(sd, cpu);
5585 rq_attach_root(rq, rd);
5587 rcu_assign_pointer(rq->sd, sd);
5588 destroy_sched_domains(tmp, cpu);
5590 update_top_cache_domain(cpu);
5593 /* cpus with isolated domains */
5594 static cpumask_var_t cpu_isolated_map;
5596 /* Setup the mask of cpus configured for isolated domains */
5597 static int __init isolated_cpu_setup(char *str)
5599 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5600 cpulist_parse(str, cpu_isolated_map);
5604 __setup("isolcpus=", isolated_cpu_setup);
5606 static const struct cpumask *cpu_cpu_mask(int cpu)
5608 return cpumask_of_node(cpu_to_node(cpu));
5612 struct sched_domain **__percpu sd;
5613 struct sched_group **__percpu sg;
5614 struct sched_group_power **__percpu sgp;
5618 struct sched_domain ** __percpu sd;
5619 struct root_domain *rd;
5629 struct sched_domain_topology_level;
5631 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
5632 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
5634 #define SDTL_OVERLAP 0x01
5636 struct sched_domain_topology_level {
5637 sched_domain_init_f init;
5638 sched_domain_mask_f mask;
5641 struct sd_data data;
5645 * Build an iteration mask that can exclude certain CPUs from the upwards
5648 * Asymmetric node setups can result in situations where the domain tree is of
5649 * unequal depth, make sure to skip domains that already cover the entire
5652 * In that case build_sched_domains() will have terminated the iteration early
5653 * and our sibling sd spans will be empty. Domains should always include the
5654 * cpu they're built on, so check that.
5657 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5659 const struct cpumask *span = sched_domain_span(sd);
5660 struct sd_data *sdd = sd->private;
5661 struct sched_domain *sibling;
5664 for_each_cpu(i, span) {
5665 sibling = *per_cpu_ptr(sdd->sd, i);
5666 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5669 cpumask_set_cpu(i, sched_group_mask(sg));
5674 * Return the canonical balance cpu for this group, this is the first cpu
5675 * of this group that's also in the iteration mask.
5677 int group_balance_cpu(struct sched_group *sg)
5679 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5683 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5685 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5686 const struct cpumask *span = sched_domain_span(sd);
5687 struct cpumask *covered = sched_domains_tmpmask;
5688 struct sd_data *sdd = sd->private;
5689 struct sched_domain *child;
5692 cpumask_clear(covered);
5694 for_each_cpu(i, span) {
5695 struct cpumask *sg_span;
5697 if (cpumask_test_cpu(i, covered))
5700 child = *per_cpu_ptr(sdd->sd, i);
5702 /* See the comment near build_group_mask(). */
5703 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5706 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5707 GFP_KERNEL, cpu_to_node(cpu));
5712 sg_span = sched_group_cpus(sg);
5714 child = child->child;
5715 cpumask_copy(sg_span, sched_domain_span(child));
5717 cpumask_set_cpu(i, sg_span);
5719 cpumask_or(covered, covered, sg_span);
5721 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5722 if (atomic_inc_return(&sg->sgp->ref) == 1)
5723 build_group_mask(sd, sg);
5726 * Initialize sgp->power such that even if we mess up the
5727 * domains and no possible iteration will get us here, we won't
5730 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5731 sg->sgp->power_orig = sg->sgp->power;
5734 * Make sure the first group of this domain contains the
5735 * canonical balance cpu. Otherwise the sched_domain iteration
5736 * breaks. See update_sg_lb_stats().
5738 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5739 group_balance_cpu(sg) == cpu)
5749 sd->groups = groups;
5754 free_sched_groups(first, 0);
5759 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5761 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5762 struct sched_domain *child = sd->child;
5765 cpu = cpumask_first(sched_domain_span(child));
5768 *sg = *per_cpu_ptr(sdd->sg, cpu);
5769 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5770 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5777 * build_sched_groups will build a circular linked list of the groups
5778 * covered by the given span, and will set each group's ->cpumask correctly,
5779 * and ->cpu_power to 0.
5781 * Assumes the sched_domain tree is fully constructed
5784 build_sched_groups(struct sched_domain *sd, int cpu)
5786 struct sched_group *first = NULL, *last = NULL;
5787 struct sd_data *sdd = sd->private;
5788 const struct cpumask *span = sched_domain_span(sd);
5789 struct cpumask *covered;
5792 get_group(cpu, sdd, &sd->groups);
5793 atomic_inc(&sd->groups->ref);
5795 if (cpu != cpumask_first(span))
5798 lockdep_assert_held(&sched_domains_mutex);
5799 covered = sched_domains_tmpmask;
5801 cpumask_clear(covered);
5803 for_each_cpu(i, span) {
5804 struct sched_group *sg;
5807 if (cpumask_test_cpu(i, covered))
5810 group = get_group(i, sdd, &sg);
5811 cpumask_clear(sched_group_cpus(sg));
5813 cpumask_setall(sched_group_mask(sg));
5815 for_each_cpu(j, span) {
5816 if (get_group(j, sdd, NULL) != group)
5819 cpumask_set_cpu(j, covered);
5820 cpumask_set_cpu(j, sched_group_cpus(sg));
5835 * Initialize sched groups cpu_power.
5837 * cpu_power indicates the capacity of sched group, which is used while
5838 * distributing the load between different sched groups in a sched domain.
5839 * Typically cpu_power for all the groups in a sched domain will be same unless
5840 * there are asymmetries in the topology. If there are asymmetries, group
5841 * having more cpu_power will pickup more load compared to the group having
5844 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5846 struct sched_group *sg = sd->groups;
5851 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5853 } while (sg != sd->groups);
5855 if (cpu != group_balance_cpu(sg))
5858 update_group_power(sd, cpu);
5859 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5862 int __weak arch_sd_sibling_asym_packing(void)
5864 return 0*SD_ASYM_PACKING;
5868 * Initializers for schedule domains
5869 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5872 #ifdef CONFIG_SCHED_DEBUG
5873 # define SD_INIT_NAME(sd, type) sd->name = #type
5875 # define SD_INIT_NAME(sd, type) do { } while (0)
5878 #define SD_INIT_FUNC(type) \
5879 static noinline struct sched_domain * \
5880 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
5882 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
5883 *sd = SD_##type##_INIT; \
5884 SD_INIT_NAME(sd, type); \
5885 sd->private = &tl->data; \
5890 #ifdef CONFIG_SCHED_SMT
5891 SD_INIT_FUNC(SIBLING)
5893 #ifdef CONFIG_SCHED_MC
5896 #ifdef CONFIG_SCHED_BOOK
5900 static int default_relax_domain_level = -1;
5901 int sched_domain_level_max;
5903 static int __init setup_relax_domain_level(char *str)
5905 if (kstrtoint(str, 0, &default_relax_domain_level))
5906 pr_warn("Unable to set relax_domain_level\n");
5910 __setup("relax_domain_level=", setup_relax_domain_level);
5912 static void set_domain_attribute(struct sched_domain *sd,
5913 struct sched_domain_attr *attr)
5917 if (!attr || attr->relax_domain_level < 0) {
5918 if (default_relax_domain_level < 0)
5921 request = default_relax_domain_level;
5923 request = attr->relax_domain_level;
5924 if (request < sd->level) {
5925 /* turn off idle balance on this domain */
5926 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5928 /* turn on idle balance on this domain */
5929 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5933 static void __sdt_free(const struct cpumask *cpu_map);
5934 static int __sdt_alloc(const struct cpumask *cpu_map);
5936 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5937 const struct cpumask *cpu_map)
5941 if (!atomic_read(&d->rd->refcount))
5942 free_rootdomain(&d->rd->rcu); /* fall through */
5944 free_percpu(d->sd); /* fall through */
5946 __sdt_free(cpu_map); /* fall through */
5952 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5953 const struct cpumask *cpu_map)
5955 memset(d, 0, sizeof(*d));
5957 if (__sdt_alloc(cpu_map))
5958 return sa_sd_storage;
5959 d->sd = alloc_percpu(struct sched_domain *);
5961 return sa_sd_storage;
5962 d->rd = alloc_rootdomain();
5965 return sa_rootdomain;
5969 * NULL the sd_data elements we've used to build the sched_domain and
5970 * sched_group structure so that the subsequent __free_domain_allocs()
5971 * will not free the data we're using.
5973 static void claim_allocations(int cpu, struct sched_domain *sd)
5975 struct sd_data *sdd = sd->private;
5977 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5978 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5980 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5981 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5983 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5984 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5987 #ifdef CONFIG_SCHED_SMT
5988 static const struct cpumask *cpu_smt_mask(int cpu)
5990 return topology_thread_cpumask(cpu);
5995 * Topology list, bottom-up.
5997 static struct sched_domain_topology_level default_topology[] = {
5998 #ifdef CONFIG_SCHED_SMT
5999 { sd_init_SIBLING, cpu_smt_mask, },
6001 #ifdef CONFIG_SCHED_MC
6002 { sd_init_MC, cpu_coregroup_mask, },
6004 #ifdef CONFIG_SCHED_BOOK
6005 { sd_init_BOOK, cpu_book_mask, },
6007 { sd_init_CPU, cpu_cpu_mask, },
6011 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
6013 #define for_each_sd_topology(tl) \
6014 for (tl = sched_domain_topology; tl->init; tl++)
6018 static int sched_domains_numa_levels;
6019 static int *sched_domains_numa_distance;
6020 static struct cpumask ***sched_domains_numa_masks;
6021 static int sched_domains_curr_level;
6023 static inline int sd_local_flags(int level)
6025 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE)
6028 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE;
6031 static struct sched_domain *
6032 sd_numa_init(struct sched_domain_topology_level *tl, int cpu)
6034 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6035 int level = tl->numa_level;
6036 int sd_weight = cpumask_weight(
6037 sched_domains_numa_masks[level][cpu_to_node(cpu)]);
6039 *sd = (struct sched_domain){
6040 .min_interval = sd_weight,
6041 .max_interval = 2*sd_weight,
6043 .imbalance_pct = 125,
6044 .cache_nice_tries = 2,
6051 .flags = 1*SD_LOAD_BALANCE
6052 | 1*SD_BALANCE_NEWIDLE
6057 | 0*SD_SHARE_CPUPOWER
6058 | 0*SD_SHARE_PKG_RESOURCES
6060 | 0*SD_PREFER_SIBLING
6062 | sd_local_flags(level)
6064 .last_balance = jiffies,
6065 .balance_interval = sd_weight,
6067 SD_INIT_NAME(sd, NUMA);
6068 sd->private = &tl->data;
6071 * Ugly hack to pass state to sd_numa_mask()...
6073 sched_domains_curr_level = tl->numa_level;
6078 static const struct cpumask *sd_numa_mask(int cpu)
6080 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6083 static void sched_numa_warn(const char *str)
6085 static int done = false;
6093 printk(KERN_WARNING "ERROR: %s\n\n", str);
6095 for (i = 0; i < nr_node_ids; i++) {
6096 printk(KERN_WARNING " ");
6097 for (j = 0; j < nr_node_ids; j++)
6098 printk(KERN_CONT "%02d ", node_distance(i,j));
6099 printk(KERN_CONT "\n");
6101 printk(KERN_WARNING "\n");
6104 static bool find_numa_distance(int distance)
6108 if (distance == node_distance(0, 0))
6111 for (i = 0; i < sched_domains_numa_levels; i++) {
6112 if (sched_domains_numa_distance[i] == distance)
6119 static void sched_init_numa(void)
6121 int next_distance, curr_distance = node_distance(0, 0);
6122 struct sched_domain_topology_level *tl;
6126 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6127 if (!sched_domains_numa_distance)
6131 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6132 * unique distances in the node_distance() table.
6134 * Assumes node_distance(0,j) includes all distances in
6135 * node_distance(i,j) in order to avoid cubic time.
6137 next_distance = curr_distance;
6138 for (i = 0; i < nr_node_ids; i++) {
6139 for (j = 0; j < nr_node_ids; j++) {
6140 for (k = 0; k < nr_node_ids; k++) {
6141 int distance = node_distance(i, k);
6143 if (distance > curr_distance &&
6144 (distance < next_distance ||
6145 next_distance == curr_distance))
6146 next_distance = distance;
6149 * While not a strong assumption it would be nice to know
6150 * about cases where if node A is connected to B, B is not
6151 * equally connected to A.
6153 if (sched_debug() && node_distance(k, i) != distance)
6154 sched_numa_warn("Node-distance not symmetric");
6156 if (sched_debug() && i && !find_numa_distance(distance))
6157 sched_numa_warn("Node-0 not representative");
6159 if (next_distance != curr_distance) {
6160 sched_domains_numa_distance[level++] = next_distance;
6161 sched_domains_numa_levels = level;
6162 curr_distance = next_distance;
6167 * In case of sched_debug() we verify the above assumption.
6173 * 'level' contains the number of unique distances, excluding the
6174 * identity distance node_distance(i,i).
6176 * The sched_domains_numa_distance[] array includes the actual distance
6181 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6182 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6183 * the array will contain less then 'level' members. This could be
6184 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6185 * in other functions.
6187 * We reset it to 'level' at the end of this function.
6189 sched_domains_numa_levels = 0;
6191 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6192 if (!sched_domains_numa_masks)
6196 * Now for each level, construct a mask per node which contains all
6197 * cpus of nodes that are that many hops away from us.
6199 for (i = 0; i < level; i++) {
6200 sched_domains_numa_masks[i] =
6201 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6202 if (!sched_domains_numa_masks[i])
6205 for (j = 0; j < nr_node_ids; j++) {
6206 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6210 sched_domains_numa_masks[i][j] = mask;
6212 for (k = 0; k < nr_node_ids; k++) {
6213 if (node_distance(j, k) > sched_domains_numa_distance[i])
6216 cpumask_or(mask, mask, cpumask_of_node(k));
6221 tl = kzalloc((ARRAY_SIZE(default_topology) + level) *
6222 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6227 * Copy the default topology bits..
6229 for (i = 0; default_topology[i].init; i++)
6230 tl[i] = default_topology[i];
6233 * .. and append 'j' levels of NUMA goodness.
6235 for (j = 0; j < level; i++, j++) {
6236 tl[i] = (struct sched_domain_topology_level){
6237 .init = sd_numa_init,
6238 .mask = sd_numa_mask,
6239 .flags = SDTL_OVERLAP,
6244 sched_domain_topology = tl;
6246 sched_domains_numa_levels = level;
6249 static void sched_domains_numa_masks_set(int cpu)
6252 int node = cpu_to_node(cpu);
6254 for (i = 0; i < sched_domains_numa_levels; i++) {
6255 for (j = 0; j < nr_node_ids; j++) {
6256 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6257 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6262 static void sched_domains_numa_masks_clear(int cpu)
6265 for (i = 0; i < sched_domains_numa_levels; i++) {
6266 for (j = 0; j < nr_node_ids; j++)
6267 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6272 * Update sched_domains_numa_masks[level][node] array when new cpus
6275 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6276 unsigned long action,
6279 int cpu = (long)hcpu;
6281 switch (action & ~CPU_TASKS_FROZEN) {
6283 sched_domains_numa_masks_set(cpu);
6287 sched_domains_numa_masks_clear(cpu);
6297 static inline void sched_init_numa(void)
6301 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6302 unsigned long action,
6307 #endif /* CONFIG_NUMA */
6309 static int __sdt_alloc(const struct cpumask *cpu_map)
6311 struct sched_domain_topology_level *tl;
6314 for_each_sd_topology(tl) {
6315 struct sd_data *sdd = &tl->data;
6317 sdd->sd = alloc_percpu(struct sched_domain *);
6321 sdd->sg = alloc_percpu(struct sched_group *);
6325 sdd->sgp = alloc_percpu(struct sched_group_power *);
6329 for_each_cpu(j, cpu_map) {
6330 struct sched_domain *sd;
6331 struct sched_group *sg;
6332 struct sched_group_power *sgp;
6334 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6335 GFP_KERNEL, cpu_to_node(j));
6339 *per_cpu_ptr(sdd->sd, j) = sd;
6341 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6342 GFP_KERNEL, cpu_to_node(j));
6348 *per_cpu_ptr(sdd->sg, j) = sg;
6350 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6351 GFP_KERNEL, cpu_to_node(j));
6355 *per_cpu_ptr(sdd->sgp, j) = sgp;
6362 static void __sdt_free(const struct cpumask *cpu_map)
6364 struct sched_domain_topology_level *tl;
6367 for_each_sd_topology(tl) {
6368 struct sd_data *sdd = &tl->data;
6370 for_each_cpu(j, cpu_map) {
6371 struct sched_domain *sd;
6374 sd = *per_cpu_ptr(sdd->sd, j);
6375 if (sd && (sd->flags & SD_OVERLAP))
6376 free_sched_groups(sd->groups, 0);
6377 kfree(*per_cpu_ptr(sdd->sd, j));
6381 kfree(*per_cpu_ptr(sdd->sg, j));
6383 kfree(*per_cpu_ptr(sdd->sgp, j));
6385 free_percpu(sdd->sd);
6387 free_percpu(sdd->sg);
6389 free_percpu(sdd->sgp);
6394 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6395 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6396 struct sched_domain *child, int cpu)
6398 struct sched_domain *sd = tl->init(tl, cpu);
6402 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6404 sd->level = child->level + 1;
6405 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6409 set_domain_attribute(sd, attr);
6415 * Build sched domains for a given set of cpus and attach the sched domains
6416 * to the individual cpus
6418 static int build_sched_domains(const struct cpumask *cpu_map,
6419 struct sched_domain_attr *attr)
6421 enum s_alloc alloc_state;
6422 struct sched_domain *sd;
6424 int i, ret = -ENOMEM;
6426 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6427 if (alloc_state != sa_rootdomain)
6430 /* Set up domains for cpus specified by the cpu_map. */
6431 for_each_cpu(i, cpu_map) {
6432 struct sched_domain_topology_level *tl;
6435 for_each_sd_topology(tl) {
6436 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6437 if (tl == sched_domain_topology)
6438 *per_cpu_ptr(d.sd, i) = sd;
6439 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6440 sd->flags |= SD_OVERLAP;
6441 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6446 /* Build the groups for the domains */
6447 for_each_cpu(i, cpu_map) {
6448 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6449 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6450 if (sd->flags & SD_OVERLAP) {
6451 if (build_overlap_sched_groups(sd, i))
6454 if (build_sched_groups(sd, i))
6460 /* Calculate CPU power for physical packages and nodes */
6461 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6462 if (!cpumask_test_cpu(i, cpu_map))
6465 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6466 claim_allocations(i, sd);
6467 init_sched_groups_power(i, sd);
6471 /* Attach the domains */
6473 for_each_cpu(i, cpu_map) {
6474 sd = *per_cpu_ptr(d.sd, i);
6475 cpu_attach_domain(sd, d.rd, i);
6481 __free_domain_allocs(&d, alloc_state, cpu_map);
6485 static cpumask_var_t *doms_cur; /* current sched domains */
6486 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6487 static struct sched_domain_attr *dattr_cur;
6488 /* attribues of custom domains in 'doms_cur' */
6491 * Special case: If a kmalloc of a doms_cur partition (array of
6492 * cpumask) fails, then fallback to a single sched domain,
6493 * as determined by the single cpumask fallback_doms.
6495 static cpumask_var_t fallback_doms;
6498 * arch_update_cpu_topology lets virtualized architectures update the
6499 * cpu core maps. It is supposed to return 1 if the topology changed
6500 * or 0 if it stayed the same.
6502 int __weak arch_update_cpu_topology(void)
6507 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6510 cpumask_var_t *doms;
6512 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6515 for (i = 0; i < ndoms; i++) {
6516 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6517 free_sched_domains(doms, i);
6524 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6527 for (i = 0; i < ndoms; i++)
6528 free_cpumask_var(doms[i]);
6533 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6534 * For now this just excludes isolated cpus, but could be used to
6535 * exclude other special cases in the future.
6537 static int init_sched_domains(const struct cpumask *cpu_map)
6541 arch_update_cpu_topology();
6543 doms_cur = alloc_sched_domains(ndoms_cur);
6545 doms_cur = &fallback_doms;
6546 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6547 err = build_sched_domains(doms_cur[0], NULL);
6548 register_sched_domain_sysctl();
6554 * Detach sched domains from a group of cpus specified in cpu_map
6555 * These cpus will now be attached to the NULL domain
6557 static void detach_destroy_domains(const struct cpumask *cpu_map)
6562 for_each_cpu(i, cpu_map)
6563 cpu_attach_domain(NULL, &def_root_domain, i);
6567 /* handle null as "default" */
6568 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6569 struct sched_domain_attr *new, int idx_new)
6571 struct sched_domain_attr tmp;
6578 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6579 new ? (new + idx_new) : &tmp,
6580 sizeof(struct sched_domain_attr));
6584 * Partition sched domains as specified by the 'ndoms_new'
6585 * cpumasks in the array doms_new[] of cpumasks. This compares
6586 * doms_new[] to the current sched domain partitioning, doms_cur[].
6587 * It destroys each deleted domain and builds each new domain.
6589 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6590 * The masks don't intersect (don't overlap.) We should setup one
6591 * sched domain for each mask. CPUs not in any of the cpumasks will
6592 * not be load balanced. If the same cpumask appears both in the
6593 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6596 * The passed in 'doms_new' should be allocated using
6597 * alloc_sched_domains. This routine takes ownership of it and will
6598 * free_sched_domains it when done with it. If the caller failed the
6599 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6600 * and partition_sched_domains() will fallback to the single partition
6601 * 'fallback_doms', it also forces the domains to be rebuilt.
6603 * If doms_new == NULL it will be replaced with cpu_online_mask.
6604 * ndoms_new == 0 is a special case for destroying existing domains,
6605 * and it will not create the default domain.
6607 * Call with hotplug lock held
6609 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6610 struct sched_domain_attr *dattr_new)
6615 mutex_lock(&sched_domains_mutex);
6617 /* always unregister in case we don't destroy any domains */
6618 unregister_sched_domain_sysctl();
6620 /* Let architecture update cpu core mappings. */
6621 new_topology = arch_update_cpu_topology();
6623 n = doms_new ? ndoms_new : 0;
6625 /* Destroy deleted domains */
6626 for (i = 0; i < ndoms_cur; i++) {
6627 for (j = 0; j < n && !new_topology; j++) {
6628 if (cpumask_equal(doms_cur[i], doms_new[j])
6629 && dattrs_equal(dattr_cur, i, dattr_new, j))
6632 /* no match - a current sched domain not in new doms_new[] */
6633 detach_destroy_domains(doms_cur[i]);
6639 if (doms_new == NULL) {
6641 doms_new = &fallback_doms;
6642 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6643 WARN_ON_ONCE(dattr_new);
6646 /* Build new domains */
6647 for (i = 0; i < ndoms_new; i++) {
6648 for (j = 0; j < n && !new_topology; j++) {
6649 if (cpumask_equal(doms_new[i], doms_cur[j])
6650 && dattrs_equal(dattr_new, i, dattr_cur, j))
6653 /* no match - add a new doms_new */
6654 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6659 /* Remember the new sched domains */
6660 if (doms_cur != &fallback_doms)
6661 free_sched_domains(doms_cur, ndoms_cur);
6662 kfree(dattr_cur); /* kfree(NULL) is safe */
6663 doms_cur = doms_new;
6664 dattr_cur = dattr_new;
6665 ndoms_cur = ndoms_new;
6667 register_sched_domain_sysctl();
6669 mutex_unlock(&sched_domains_mutex);
6672 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6675 * Update cpusets according to cpu_active mask. If cpusets are
6676 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6677 * around partition_sched_domains().
6679 * If we come here as part of a suspend/resume, don't touch cpusets because we
6680 * want to restore it back to its original state upon resume anyway.
6682 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6686 case CPU_ONLINE_FROZEN:
6687 case CPU_DOWN_FAILED_FROZEN:
6690 * num_cpus_frozen tracks how many CPUs are involved in suspend
6691 * resume sequence. As long as this is not the last online
6692 * operation in the resume sequence, just build a single sched
6693 * domain, ignoring cpusets.
6696 if (likely(num_cpus_frozen)) {
6697 partition_sched_domains(1, NULL, NULL);
6702 * This is the last CPU online operation. So fall through and
6703 * restore the original sched domains by considering the
6704 * cpuset configurations.
6708 case CPU_DOWN_FAILED:
6709 cpuset_update_active_cpus(true);
6717 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6721 case CPU_DOWN_PREPARE:
6722 cpuset_update_active_cpus(false);
6724 case CPU_DOWN_PREPARE_FROZEN:
6726 partition_sched_domains(1, NULL, NULL);
6734 void __init sched_init_smp(void)
6736 cpumask_var_t non_isolated_cpus;
6738 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6739 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6744 * There's no userspace yet to cause hotplug operations; hence all the
6745 * cpu masks are stable and all blatant races in the below code cannot
6748 mutex_lock(&sched_domains_mutex);
6749 init_sched_domains(cpu_active_mask);
6750 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6751 if (cpumask_empty(non_isolated_cpus))
6752 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6753 mutex_unlock(&sched_domains_mutex);
6755 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6756 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6757 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6761 /* Move init over to a non-isolated CPU */
6762 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6764 sched_init_granularity();
6765 free_cpumask_var(non_isolated_cpus);
6767 init_sched_rt_class();
6768 init_sched_dl_class();
6771 void __init sched_init_smp(void)
6773 sched_init_granularity();
6775 #endif /* CONFIG_SMP */
6777 const_debug unsigned int sysctl_timer_migration = 1;
6779 int in_sched_functions(unsigned long addr)
6781 return in_lock_functions(addr) ||
6782 (addr >= (unsigned long)__sched_text_start
6783 && addr < (unsigned long)__sched_text_end);
6786 #ifdef CONFIG_CGROUP_SCHED
6788 * Default task group.
6789 * Every task in system belongs to this group at bootup.
6791 struct task_group root_task_group;
6792 LIST_HEAD(task_groups);
6795 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6797 void __init sched_init(void)
6800 unsigned long alloc_size = 0, ptr;
6802 #ifdef CONFIG_FAIR_GROUP_SCHED
6803 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6805 #ifdef CONFIG_RT_GROUP_SCHED
6806 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6808 #ifdef CONFIG_CPUMASK_OFFSTACK
6809 alloc_size += num_possible_cpus() * cpumask_size();
6812 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6814 #ifdef CONFIG_FAIR_GROUP_SCHED
6815 root_task_group.se = (struct sched_entity **)ptr;
6816 ptr += nr_cpu_ids * sizeof(void **);
6818 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6819 ptr += nr_cpu_ids * sizeof(void **);
6821 #endif /* CONFIG_FAIR_GROUP_SCHED */
6822 #ifdef CONFIG_RT_GROUP_SCHED
6823 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6824 ptr += nr_cpu_ids * sizeof(void **);
6826 root_task_group.rt_rq = (struct rt_rq **)ptr;
6827 ptr += nr_cpu_ids * sizeof(void **);
6829 #endif /* CONFIG_RT_GROUP_SCHED */
6830 #ifdef CONFIG_CPUMASK_OFFSTACK
6831 for_each_possible_cpu(i) {
6832 per_cpu(load_balance_mask, i) = (void *)ptr;
6833 ptr += cpumask_size();
6835 #endif /* CONFIG_CPUMASK_OFFSTACK */
6838 init_rt_bandwidth(&def_rt_bandwidth,
6839 global_rt_period(), global_rt_runtime());
6840 init_dl_bandwidth(&def_dl_bandwidth,
6841 global_rt_period(), global_rt_runtime());
6844 init_defrootdomain();
6847 #ifdef CONFIG_RT_GROUP_SCHED
6848 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6849 global_rt_period(), global_rt_runtime());
6850 #endif /* CONFIG_RT_GROUP_SCHED */
6852 #ifdef CONFIG_CGROUP_SCHED
6853 list_add(&root_task_group.list, &task_groups);
6854 INIT_LIST_HEAD(&root_task_group.children);
6855 INIT_LIST_HEAD(&root_task_group.siblings);
6856 autogroup_init(&init_task);
6858 #endif /* CONFIG_CGROUP_SCHED */
6860 for_each_possible_cpu(i) {
6864 raw_spin_lock_init(&rq->lock);
6866 rq->calc_load_active = 0;
6867 rq->calc_load_update = jiffies + LOAD_FREQ;
6868 init_cfs_rq(&rq->cfs);
6869 init_rt_rq(&rq->rt, rq);
6870 init_dl_rq(&rq->dl, rq);
6871 #ifdef CONFIG_FAIR_GROUP_SCHED
6872 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6873 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6875 * How much cpu bandwidth does root_task_group get?
6877 * In case of task-groups formed thr' the cgroup filesystem, it
6878 * gets 100% of the cpu resources in the system. This overall
6879 * system cpu resource is divided among the tasks of
6880 * root_task_group and its child task-groups in a fair manner,
6881 * based on each entity's (task or task-group's) weight
6882 * (se->load.weight).
6884 * In other words, if root_task_group has 10 tasks of weight
6885 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6886 * then A0's share of the cpu resource is:
6888 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6890 * We achieve this by letting root_task_group's tasks sit
6891 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6893 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6894 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6895 #endif /* CONFIG_FAIR_GROUP_SCHED */
6897 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6898 #ifdef CONFIG_RT_GROUP_SCHED
6899 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6902 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6903 rq->cpu_load[j] = 0;
6905 rq->last_load_update_tick = jiffies;
6910 rq->cpu_power = SCHED_POWER_SCALE;
6911 rq->post_schedule = 0;
6912 rq->active_balance = 0;
6913 rq->next_balance = jiffies;
6918 rq->avg_idle = 2*sysctl_sched_migration_cost;
6919 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6921 INIT_LIST_HEAD(&rq->cfs_tasks);
6923 rq_attach_root(rq, &def_root_domain);
6924 #ifdef CONFIG_NO_HZ_COMMON
6927 #ifdef CONFIG_NO_HZ_FULL
6928 rq->last_sched_tick = 0;
6932 atomic_set(&rq->nr_iowait, 0);
6935 set_load_weight(&init_task);
6937 #ifdef CONFIG_PREEMPT_NOTIFIERS
6938 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6942 * The boot idle thread does lazy MMU switching as well:
6944 atomic_inc(&init_mm.mm_count);
6945 enter_lazy_tlb(&init_mm, current);
6948 * Make us the idle thread. Technically, schedule() should not be
6949 * called from this thread, however somewhere below it might be,
6950 * but because we are the idle thread, we just pick up running again
6951 * when this runqueue becomes "idle".
6953 init_idle(current, smp_processor_id());
6955 calc_load_update = jiffies + LOAD_FREQ;
6958 * During early bootup we pretend to be a normal task:
6960 current->sched_class = &fair_sched_class;
6963 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6964 /* May be allocated at isolcpus cmdline parse time */
6965 if (cpu_isolated_map == NULL)
6966 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6967 idle_thread_set_boot_cpu();
6969 init_sched_fair_class();
6971 scheduler_running = 1;
6974 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6975 static inline int preempt_count_equals(int preempt_offset)
6977 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6979 return (nested == preempt_offset);
6982 void __might_sleep(const char *file, int line, int preempt_offset)
6984 static unsigned long prev_jiffy; /* ratelimiting */
6986 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6987 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
6988 !is_idle_task(current)) ||
6989 system_state != SYSTEM_RUNNING || oops_in_progress)
6991 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
6993 prev_jiffy = jiffies;
6996 "BUG: sleeping function called from invalid context at %s:%d\n",
6999 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7000 in_atomic(), irqs_disabled(),
7001 current->pid, current->comm);
7003 debug_show_held_locks(current);
7004 if (irqs_disabled())
7005 print_irqtrace_events(current);
7006 #ifdef CONFIG_DEBUG_PREEMPT
7007 if (!preempt_count_equals(preempt_offset)) {
7008 pr_err("Preemption disabled at:");
7009 print_ip_sym(current->preempt_disable_ip);
7015 EXPORT_SYMBOL(__might_sleep);
7018 #ifdef CONFIG_MAGIC_SYSRQ
7019 static void normalize_task(struct rq *rq, struct task_struct *p)
7021 const struct sched_class *prev_class = p->sched_class;
7022 struct sched_attr attr = {
7023 .sched_policy = SCHED_NORMAL,
7025 int old_prio = p->prio;
7030 dequeue_task(rq, p, 0);
7031 __setscheduler(rq, p, &attr);
7033 enqueue_task(rq, p, 0);
7034 resched_task(rq->curr);
7037 check_class_changed(rq, p, prev_class, old_prio);
7040 void normalize_rt_tasks(void)
7042 struct task_struct *g, *p;
7043 unsigned long flags;
7046 read_lock_irqsave(&tasklist_lock, flags);
7047 do_each_thread(g, p) {
7049 * Only normalize user tasks:
7054 p->se.exec_start = 0;
7055 #ifdef CONFIG_SCHEDSTATS
7056 p->se.statistics.wait_start = 0;
7057 p->se.statistics.sleep_start = 0;
7058 p->se.statistics.block_start = 0;
7061 if (!dl_task(p) && !rt_task(p)) {
7063 * Renice negative nice level userspace
7066 if (task_nice(p) < 0 && p->mm)
7067 set_user_nice(p, 0);
7071 raw_spin_lock(&p->pi_lock);
7072 rq = __task_rq_lock(p);
7074 normalize_task(rq, p);
7076 __task_rq_unlock(rq);
7077 raw_spin_unlock(&p->pi_lock);
7078 } while_each_thread(g, p);
7080 read_unlock_irqrestore(&tasklist_lock, flags);
7083 #endif /* CONFIG_MAGIC_SYSRQ */
7085 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7087 * These functions are only useful for the IA64 MCA handling, or kdb.
7089 * They can only be called when the whole system has been
7090 * stopped - every CPU needs to be quiescent, and no scheduling
7091 * activity can take place. Using them for anything else would
7092 * be a serious bug, and as a result, they aren't even visible
7093 * under any other configuration.
7097 * curr_task - return the current task for a given cpu.
7098 * @cpu: the processor in question.
7100 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7102 * Return: The current task for @cpu.
7104 struct task_struct *curr_task(int cpu)
7106 return cpu_curr(cpu);
7109 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7113 * set_curr_task - set the current task for a given cpu.
7114 * @cpu: the processor in question.
7115 * @p: the task pointer to set.
7117 * Description: This function must only be used when non-maskable interrupts
7118 * are serviced on a separate stack. It allows the architecture to switch the
7119 * notion of the current task on a cpu in a non-blocking manner. This function
7120 * must be called with all CPU's synchronized, and interrupts disabled, the
7121 * and caller must save the original value of the current task (see
7122 * curr_task() above) and restore that value before reenabling interrupts and
7123 * re-starting the system.
7125 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7127 void set_curr_task(int cpu, struct task_struct *p)
7134 #ifdef CONFIG_CGROUP_SCHED
7135 /* task_group_lock serializes the addition/removal of task groups */
7136 static DEFINE_SPINLOCK(task_group_lock);
7138 static void free_sched_group(struct task_group *tg)
7140 free_fair_sched_group(tg);
7141 free_rt_sched_group(tg);
7146 /* allocate runqueue etc for a new task group */
7147 struct task_group *sched_create_group(struct task_group *parent)
7149 struct task_group *tg;
7151 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7153 return ERR_PTR(-ENOMEM);
7155 if (!alloc_fair_sched_group(tg, parent))
7158 if (!alloc_rt_sched_group(tg, parent))
7164 free_sched_group(tg);
7165 return ERR_PTR(-ENOMEM);
7168 void sched_online_group(struct task_group *tg, struct task_group *parent)
7170 unsigned long flags;
7172 spin_lock_irqsave(&task_group_lock, flags);
7173 list_add_rcu(&tg->list, &task_groups);
7175 WARN_ON(!parent); /* root should already exist */
7177 tg->parent = parent;
7178 INIT_LIST_HEAD(&tg->children);
7179 list_add_rcu(&tg->siblings, &parent->children);
7180 spin_unlock_irqrestore(&task_group_lock, flags);
7183 /* rcu callback to free various structures associated with a task group */
7184 static void free_sched_group_rcu(struct rcu_head *rhp)
7186 /* now it should be safe to free those cfs_rqs */
7187 free_sched_group(container_of(rhp, struct task_group, rcu));
7190 /* Destroy runqueue etc associated with a task group */
7191 void sched_destroy_group(struct task_group *tg)
7193 /* wait for possible concurrent references to cfs_rqs complete */
7194 call_rcu(&tg->rcu, free_sched_group_rcu);
7197 void sched_offline_group(struct task_group *tg)
7199 unsigned long flags;
7202 /* end participation in shares distribution */
7203 for_each_possible_cpu(i)
7204 unregister_fair_sched_group(tg, i);
7206 spin_lock_irqsave(&task_group_lock, flags);
7207 list_del_rcu(&tg->list);
7208 list_del_rcu(&tg->siblings);
7209 spin_unlock_irqrestore(&task_group_lock, flags);
7212 /* change task's runqueue when it moves between groups.
7213 * The caller of this function should have put the task in its new group
7214 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7215 * reflect its new group.
7217 void sched_move_task(struct task_struct *tsk)
7219 struct task_group *tg;
7221 unsigned long flags;
7224 rq = task_rq_lock(tsk, &flags);
7226 running = task_current(rq, tsk);
7230 dequeue_task(rq, tsk, 0);
7231 if (unlikely(running))
7232 tsk->sched_class->put_prev_task(rq, tsk);
7234 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7235 lockdep_is_held(&tsk->sighand->siglock)),
7236 struct task_group, css);
7237 tg = autogroup_task_group(tsk, tg);
7238 tsk->sched_task_group = tg;
7240 #ifdef CONFIG_FAIR_GROUP_SCHED
7241 if (tsk->sched_class->task_move_group)
7242 tsk->sched_class->task_move_group(tsk, on_rq);
7245 set_task_rq(tsk, task_cpu(tsk));
7247 if (unlikely(running))
7248 tsk->sched_class->set_curr_task(rq);
7250 enqueue_task(rq, tsk, 0);
7252 task_rq_unlock(rq, tsk, &flags);
7254 #endif /* CONFIG_CGROUP_SCHED */
7256 #ifdef CONFIG_RT_GROUP_SCHED
7258 * Ensure that the real time constraints are schedulable.
7260 static DEFINE_MUTEX(rt_constraints_mutex);
7262 /* Must be called with tasklist_lock held */
7263 static inline int tg_has_rt_tasks(struct task_group *tg)
7265 struct task_struct *g, *p;
7267 do_each_thread(g, p) {
7268 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7270 } while_each_thread(g, p);
7275 struct rt_schedulable_data {
7276 struct task_group *tg;
7281 static int tg_rt_schedulable(struct task_group *tg, void *data)
7283 struct rt_schedulable_data *d = data;
7284 struct task_group *child;
7285 unsigned long total, sum = 0;
7286 u64 period, runtime;
7288 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7289 runtime = tg->rt_bandwidth.rt_runtime;
7292 period = d->rt_period;
7293 runtime = d->rt_runtime;
7297 * Cannot have more runtime than the period.
7299 if (runtime > period && runtime != RUNTIME_INF)
7303 * Ensure we don't starve existing RT tasks.
7305 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7308 total = to_ratio(period, runtime);
7311 * Nobody can have more than the global setting allows.
7313 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7317 * The sum of our children's runtime should not exceed our own.
7319 list_for_each_entry_rcu(child, &tg->children, siblings) {
7320 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7321 runtime = child->rt_bandwidth.rt_runtime;
7323 if (child == d->tg) {
7324 period = d->rt_period;
7325 runtime = d->rt_runtime;
7328 sum += to_ratio(period, runtime);
7337 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7341 struct rt_schedulable_data data = {
7343 .rt_period = period,
7344 .rt_runtime = runtime,
7348 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7354 static int tg_set_rt_bandwidth(struct task_group *tg,
7355 u64 rt_period, u64 rt_runtime)
7359 mutex_lock(&rt_constraints_mutex);
7360 read_lock(&tasklist_lock);
7361 err = __rt_schedulable(tg, rt_period, rt_runtime);
7365 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7366 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7367 tg->rt_bandwidth.rt_runtime = rt_runtime;
7369 for_each_possible_cpu(i) {
7370 struct rt_rq *rt_rq = tg->rt_rq[i];
7372 raw_spin_lock(&rt_rq->rt_runtime_lock);
7373 rt_rq->rt_runtime = rt_runtime;
7374 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7376 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7378 read_unlock(&tasklist_lock);
7379 mutex_unlock(&rt_constraints_mutex);
7384 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7386 u64 rt_runtime, rt_period;
7388 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7389 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7390 if (rt_runtime_us < 0)
7391 rt_runtime = RUNTIME_INF;
7393 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7396 static long sched_group_rt_runtime(struct task_group *tg)
7400 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7403 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7404 do_div(rt_runtime_us, NSEC_PER_USEC);
7405 return rt_runtime_us;
7408 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7410 u64 rt_runtime, rt_period;
7412 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7413 rt_runtime = tg->rt_bandwidth.rt_runtime;
7418 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7421 static long sched_group_rt_period(struct task_group *tg)
7425 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7426 do_div(rt_period_us, NSEC_PER_USEC);
7427 return rt_period_us;
7429 #endif /* CONFIG_RT_GROUP_SCHED */
7431 #ifdef CONFIG_RT_GROUP_SCHED
7432 static int sched_rt_global_constraints(void)
7436 mutex_lock(&rt_constraints_mutex);
7437 read_lock(&tasklist_lock);
7438 ret = __rt_schedulable(NULL, 0, 0);
7439 read_unlock(&tasklist_lock);
7440 mutex_unlock(&rt_constraints_mutex);
7445 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7447 /* Don't accept realtime tasks when there is no way for them to run */
7448 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7454 #else /* !CONFIG_RT_GROUP_SCHED */
7455 static int sched_rt_global_constraints(void)
7457 unsigned long flags;
7460 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7461 for_each_possible_cpu(i) {
7462 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7464 raw_spin_lock(&rt_rq->rt_runtime_lock);
7465 rt_rq->rt_runtime = global_rt_runtime();
7466 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7468 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7472 #endif /* CONFIG_RT_GROUP_SCHED */
7474 static int sched_dl_global_constraints(void)
7476 u64 runtime = global_rt_runtime();
7477 u64 period = global_rt_period();
7478 u64 new_bw = to_ratio(period, runtime);
7480 unsigned long flags;
7483 * Here we want to check the bandwidth not being set to some
7484 * value smaller than the currently allocated bandwidth in
7485 * any of the root_domains.
7487 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7488 * cycling on root_domains... Discussion on different/better
7489 * solutions is welcome!
7491 for_each_possible_cpu(cpu) {
7492 struct dl_bw *dl_b = dl_bw_of(cpu);
7494 raw_spin_lock_irqsave(&dl_b->lock, flags);
7495 if (new_bw < dl_b->total_bw)
7497 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7506 static void sched_dl_do_global(void)
7510 unsigned long flags;
7512 def_dl_bandwidth.dl_period = global_rt_period();
7513 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7515 if (global_rt_runtime() != RUNTIME_INF)
7516 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7519 * FIXME: As above...
7521 for_each_possible_cpu(cpu) {
7522 struct dl_bw *dl_b = dl_bw_of(cpu);
7524 raw_spin_lock_irqsave(&dl_b->lock, flags);
7526 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7530 static int sched_rt_global_validate(void)
7532 if (sysctl_sched_rt_period <= 0)
7535 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7536 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7542 static void sched_rt_do_global(void)
7544 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7545 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7548 int sched_rt_handler(struct ctl_table *table, int write,
7549 void __user *buffer, size_t *lenp,
7552 int old_period, old_runtime;
7553 static DEFINE_MUTEX(mutex);
7557 old_period = sysctl_sched_rt_period;
7558 old_runtime = sysctl_sched_rt_runtime;
7560 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7562 if (!ret && write) {
7563 ret = sched_rt_global_validate();
7567 ret = sched_rt_global_constraints();
7571 ret = sched_dl_global_constraints();
7575 sched_rt_do_global();
7576 sched_dl_do_global();
7580 sysctl_sched_rt_period = old_period;
7581 sysctl_sched_rt_runtime = old_runtime;
7583 mutex_unlock(&mutex);
7588 int sched_rr_handler(struct ctl_table *table, int write,
7589 void __user *buffer, size_t *lenp,
7593 static DEFINE_MUTEX(mutex);
7596 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7597 /* make sure that internally we keep jiffies */
7598 /* also, writing zero resets timeslice to default */
7599 if (!ret && write) {
7600 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7601 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7603 mutex_unlock(&mutex);
7607 #ifdef CONFIG_CGROUP_SCHED
7609 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7611 return css ? container_of(css, struct task_group, css) : NULL;
7614 static struct cgroup_subsys_state *
7615 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7617 struct task_group *parent = css_tg(parent_css);
7618 struct task_group *tg;
7621 /* This is early initialization for the top cgroup */
7622 return &root_task_group.css;
7625 tg = sched_create_group(parent);
7627 return ERR_PTR(-ENOMEM);
7632 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7634 struct task_group *tg = css_tg(css);
7635 struct task_group *parent = css_tg(css_parent(css));
7638 sched_online_group(tg, parent);
7642 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7644 struct task_group *tg = css_tg(css);
7646 sched_destroy_group(tg);
7649 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7651 struct task_group *tg = css_tg(css);
7653 sched_offline_group(tg);
7656 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7657 struct cgroup_taskset *tset)
7659 struct task_struct *task;
7661 cgroup_taskset_for_each(task, tset) {
7662 #ifdef CONFIG_RT_GROUP_SCHED
7663 if (!sched_rt_can_attach(css_tg(css), task))
7666 /* We don't support RT-tasks being in separate groups */
7667 if (task->sched_class != &fair_sched_class)
7674 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7675 struct cgroup_taskset *tset)
7677 struct task_struct *task;
7679 cgroup_taskset_for_each(task, tset)
7680 sched_move_task(task);
7683 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7684 struct cgroup_subsys_state *old_css,
7685 struct task_struct *task)
7688 * cgroup_exit() is called in the copy_process() failure path.
7689 * Ignore this case since the task hasn't ran yet, this avoids
7690 * trying to poke a half freed task state from generic code.
7692 if (!(task->flags & PF_EXITING))
7695 sched_move_task(task);
7698 #ifdef CONFIG_FAIR_GROUP_SCHED
7699 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7700 struct cftype *cftype, u64 shareval)
7702 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7705 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7708 struct task_group *tg = css_tg(css);
7710 return (u64) scale_load_down(tg->shares);
7713 #ifdef CONFIG_CFS_BANDWIDTH
7714 static DEFINE_MUTEX(cfs_constraints_mutex);
7716 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7717 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7719 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7721 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7723 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7724 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7726 if (tg == &root_task_group)
7730 * Ensure we have at some amount of bandwidth every period. This is
7731 * to prevent reaching a state of large arrears when throttled via
7732 * entity_tick() resulting in prolonged exit starvation.
7734 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7738 * Likewise, bound things on the otherside by preventing insane quota
7739 * periods. This also allows us to normalize in computing quota
7742 if (period > max_cfs_quota_period)
7745 mutex_lock(&cfs_constraints_mutex);
7746 ret = __cfs_schedulable(tg, period, quota);
7750 runtime_enabled = quota != RUNTIME_INF;
7751 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7753 * If we need to toggle cfs_bandwidth_used, off->on must occur
7754 * before making related changes, and on->off must occur afterwards
7756 if (runtime_enabled && !runtime_was_enabled)
7757 cfs_bandwidth_usage_inc();
7758 raw_spin_lock_irq(&cfs_b->lock);
7759 cfs_b->period = ns_to_ktime(period);
7760 cfs_b->quota = quota;
7762 __refill_cfs_bandwidth_runtime(cfs_b);
7763 /* restart the period timer (if active) to handle new period expiry */
7764 if (runtime_enabled && cfs_b->timer_active) {
7765 /* force a reprogram */
7766 cfs_b->timer_active = 0;
7767 __start_cfs_bandwidth(cfs_b);
7769 raw_spin_unlock_irq(&cfs_b->lock);
7771 for_each_possible_cpu(i) {
7772 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7773 struct rq *rq = cfs_rq->rq;
7775 raw_spin_lock_irq(&rq->lock);
7776 cfs_rq->runtime_enabled = runtime_enabled;
7777 cfs_rq->runtime_remaining = 0;
7779 if (cfs_rq->throttled)
7780 unthrottle_cfs_rq(cfs_rq);
7781 raw_spin_unlock_irq(&rq->lock);
7783 if (runtime_was_enabled && !runtime_enabled)
7784 cfs_bandwidth_usage_dec();
7786 mutex_unlock(&cfs_constraints_mutex);
7791 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7795 period = ktime_to_ns(tg->cfs_bandwidth.period);
7796 if (cfs_quota_us < 0)
7797 quota = RUNTIME_INF;
7799 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7801 return tg_set_cfs_bandwidth(tg, period, quota);
7804 long tg_get_cfs_quota(struct task_group *tg)
7808 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7811 quota_us = tg->cfs_bandwidth.quota;
7812 do_div(quota_us, NSEC_PER_USEC);
7817 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7821 period = (u64)cfs_period_us * NSEC_PER_USEC;
7822 quota = tg->cfs_bandwidth.quota;
7824 return tg_set_cfs_bandwidth(tg, period, quota);
7827 long tg_get_cfs_period(struct task_group *tg)
7831 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7832 do_div(cfs_period_us, NSEC_PER_USEC);
7834 return cfs_period_us;
7837 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7840 return tg_get_cfs_quota(css_tg(css));
7843 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7844 struct cftype *cftype, s64 cfs_quota_us)
7846 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7849 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7852 return tg_get_cfs_period(css_tg(css));
7855 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7856 struct cftype *cftype, u64 cfs_period_us)
7858 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7861 struct cfs_schedulable_data {
7862 struct task_group *tg;
7867 * normalize group quota/period to be quota/max_period
7868 * note: units are usecs
7870 static u64 normalize_cfs_quota(struct task_group *tg,
7871 struct cfs_schedulable_data *d)
7879 period = tg_get_cfs_period(tg);
7880 quota = tg_get_cfs_quota(tg);
7883 /* note: these should typically be equivalent */
7884 if (quota == RUNTIME_INF || quota == -1)
7887 return to_ratio(period, quota);
7890 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7892 struct cfs_schedulable_data *d = data;
7893 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7894 s64 quota = 0, parent_quota = -1;
7897 quota = RUNTIME_INF;
7899 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7901 quota = normalize_cfs_quota(tg, d);
7902 parent_quota = parent_b->hierarchal_quota;
7905 * ensure max(child_quota) <= parent_quota, inherit when no
7908 if (quota == RUNTIME_INF)
7909 quota = parent_quota;
7910 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7913 cfs_b->hierarchal_quota = quota;
7918 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7921 struct cfs_schedulable_data data = {
7927 if (quota != RUNTIME_INF) {
7928 do_div(data.period, NSEC_PER_USEC);
7929 do_div(data.quota, NSEC_PER_USEC);
7933 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7939 static int cpu_stats_show(struct seq_file *sf, void *v)
7941 struct task_group *tg = css_tg(seq_css(sf));
7942 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7944 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7945 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7946 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7950 #endif /* CONFIG_CFS_BANDWIDTH */
7951 #endif /* CONFIG_FAIR_GROUP_SCHED */
7953 #ifdef CONFIG_RT_GROUP_SCHED
7954 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7955 struct cftype *cft, s64 val)
7957 return sched_group_set_rt_runtime(css_tg(css), val);
7960 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7963 return sched_group_rt_runtime(css_tg(css));
7966 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7967 struct cftype *cftype, u64 rt_period_us)
7969 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7972 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7975 return sched_group_rt_period(css_tg(css));
7977 #endif /* CONFIG_RT_GROUP_SCHED */
7979 static struct cftype cpu_files[] = {
7980 #ifdef CONFIG_FAIR_GROUP_SCHED
7983 .read_u64 = cpu_shares_read_u64,
7984 .write_u64 = cpu_shares_write_u64,
7987 #ifdef CONFIG_CFS_BANDWIDTH
7989 .name = "cfs_quota_us",
7990 .read_s64 = cpu_cfs_quota_read_s64,
7991 .write_s64 = cpu_cfs_quota_write_s64,
7994 .name = "cfs_period_us",
7995 .read_u64 = cpu_cfs_period_read_u64,
7996 .write_u64 = cpu_cfs_period_write_u64,
8000 .seq_show = cpu_stats_show,
8003 #ifdef CONFIG_RT_GROUP_SCHED
8005 .name = "rt_runtime_us",
8006 .read_s64 = cpu_rt_runtime_read,
8007 .write_s64 = cpu_rt_runtime_write,
8010 .name = "rt_period_us",
8011 .read_u64 = cpu_rt_period_read_uint,
8012 .write_u64 = cpu_rt_period_write_uint,
8018 struct cgroup_subsys cpu_cgrp_subsys = {
8019 .css_alloc = cpu_cgroup_css_alloc,
8020 .css_free = cpu_cgroup_css_free,
8021 .css_online = cpu_cgroup_css_online,
8022 .css_offline = cpu_cgroup_css_offline,
8023 .can_attach = cpu_cgroup_can_attach,
8024 .attach = cpu_cgroup_attach,
8025 .exit = cpu_cgroup_exit,
8026 .base_cftypes = cpu_files,
8030 #endif /* CONFIG_CGROUP_SCHED */
8032 void dump_cpu_task(int cpu)
8034 pr_info("Task dump for CPU %d:\n", cpu);
8035 sched_show_task(cpu_curr(cpu));