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 * cmpxchg based fetch_or, macro so it works for different integer types
511 #define fetch_or(ptr, val) \
512 ({ typeof(*(ptr)) __old, __val = *(ptr); \
514 __old = cmpxchg((ptr), __val, __val | (val)); \
515 if (__old == __val) \
522 #ifdef TIF_POLLING_NRFLAG
524 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
525 * this avoids any races wrt polling state changes and thereby avoids
528 static bool set_nr_and_not_polling(struct task_struct *p)
530 struct thread_info *ti = task_thread_info(p);
531 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
534 static bool set_nr_and_not_polling(struct task_struct *p)
536 set_tsk_need_resched(p);
542 * resched_task - mark a task 'to be rescheduled now'.
544 * On UP this means the setting of the need_resched flag, on SMP it
545 * might also involve a cross-CPU call to trigger the scheduler on
548 void resched_task(struct task_struct *p)
552 lockdep_assert_held(&task_rq(p)->lock);
554 if (test_tsk_need_resched(p))
559 if (cpu == smp_processor_id()) {
560 set_tsk_need_resched(p);
561 set_preempt_need_resched();
565 if (set_nr_and_not_polling(p))
566 smp_send_reschedule(cpu);
569 void resched_cpu(int cpu)
571 struct rq *rq = cpu_rq(cpu);
574 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
576 resched_task(cpu_curr(cpu));
577 raw_spin_unlock_irqrestore(&rq->lock, flags);
581 #ifdef CONFIG_NO_HZ_COMMON
583 * In the semi idle case, use the nearest busy cpu for migrating timers
584 * from an idle cpu. This is good for power-savings.
586 * We don't do similar optimization for completely idle system, as
587 * selecting an idle cpu will add more delays to the timers than intended
588 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
590 int get_nohz_timer_target(int pinned)
592 int cpu = smp_processor_id();
594 struct sched_domain *sd;
596 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu))
600 for_each_domain(cpu, sd) {
601 for_each_cpu(i, sched_domain_span(sd)) {
613 * When add_timer_on() enqueues a timer into the timer wheel of an
614 * idle CPU then this timer might expire before the next timer event
615 * which is scheduled to wake up that CPU. In case of a completely
616 * idle system the next event might even be infinite time into the
617 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
618 * leaves the inner idle loop so the newly added timer is taken into
619 * account when the CPU goes back to idle and evaluates the timer
620 * wheel for the next timer event.
622 static void wake_up_idle_cpu(int cpu)
624 struct rq *rq = cpu_rq(cpu);
626 if (cpu == smp_processor_id())
630 * This is safe, as this function is called with the timer
631 * wheel base lock of (cpu) held. When the CPU is on the way
632 * to idle and has not yet set rq->curr to idle then it will
633 * be serialized on the timer wheel base lock and take the new
634 * timer into account automatically.
636 if (rq->curr != rq->idle)
640 * We can set TIF_RESCHED on the idle task of the other CPU
641 * lockless. The worst case is that the other CPU runs the
642 * idle task through an additional NOOP schedule()
644 set_tsk_need_resched(rq->idle);
646 /* NEED_RESCHED must be visible before we test polling */
648 if (!tsk_is_polling(rq->idle))
649 smp_send_reschedule(cpu);
652 static bool wake_up_full_nohz_cpu(int cpu)
654 if (tick_nohz_full_cpu(cpu)) {
655 if (cpu != smp_processor_id() ||
656 tick_nohz_tick_stopped())
657 smp_send_reschedule(cpu);
664 void wake_up_nohz_cpu(int cpu)
666 if (!wake_up_full_nohz_cpu(cpu))
667 wake_up_idle_cpu(cpu);
670 static inline bool got_nohz_idle_kick(void)
672 int cpu = smp_processor_id();
674 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
677 if (idle_cpu(cpu) && !need_resched())
681 * We can't run Idle Load Balance on this CPU for this time so we
682 * cancel it and clear NOHZ_BALANCE_KICK
684 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
688 #else /* CONFIG_NO_HZ_COMMON */
690 static inline bool got_nohz_idle_kick(void)
695 #endif /* CONFIG_NO_HZ_COMMON */
697 #ifdef CONFIG_NO_HZ_FULL
698 bool sched_can_stop_tick(void)
704 /* Make sure rq->nr_running update is visible after the IPI */
707 /* More than one running task need preemption */
708 if (rq->nr_running > 1)
713 #endif /* CONFIG_NO_HZ_FULL */
715 void sched_avg_update(struct rq *rq)
717 s64 period = sched_avg_period();
719 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
721 * Inline assembly required to prevent the compiler
722 * optimising this loop into a divmod call.
723 * See __iter_div_u64_rem() for another example of this.
725 asm("" : "+rm" (rq->age_stamp));
726 rq->age_stamp += period;
731 #endif /* CONFIG_SMP */
733 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
734 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
736 * Iterate task_group tree rooted at *from, calling @down when first entering a
737 * node and @up when leaving it for the final time.
739 * Caller must hold rcu_lock or sufficient equivalent.
741 int walk_tg_tree_from(struct task_group *from,
742 tg_visitor down, tg_visitor up, void *data)
744 struct task_group *parent, *child;
750 ret = (*down)(parent, data);
753 list_for_each_entry_rcu(child, &parent->children, siblings) {
760 ret = (*up)(parent, data);
761 if (ret || parent == from)
765 parent = parent->parent;
772 int tg_nop(struct task_group *tg, void *data)
778 static void set_load_weight(struct task_struct *p)
780 int prio = p->static_prio - MAX_RT_PRIO;
781 struct load_weight *load = &p->se.load;
784 * SCHED_IDLE tasks get minimal weight:
786 if (p->policy == SCHED_IDLE) {
787 load->weight = scale_load(WEIGHT_IDLEPRIO);
788 load->inv_weight = WMULT_IDLEPRIO;
792 load->weight = scale_load(prio_to_weight[prio]);
793 load->inv_weight = prio_to_wmult[prio];
796 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
799 sched_info_queued(rq, p);
800 p->sched_class->enqueue_task(rq, p, flags);
803 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
806 sched_info_dequeued(rq, p);
807 p->sched_class->dequeue_task(rq, p, flags);
810 void activate_task(struct rq *rq, struct task_struct *p, int flags)
812 if (task_contributes_to_load(p))
813 rq->nr_uninterruptible--;
815 enqueue_task(rq, p, flags);
818 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
820 if (task_contributes_to_load(p))
821 rq->nr_uninterruptible++;
823 dequeue_task(rq, p, flags);
826 static void update_rq_clock_task(struct rq *rq, s64 delta)
829 * In theory, the compile should just see 0 here, and optimize out the call
830 * to sched_rt_avg_update. But I don't trust it...
832 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
833 s64 steal = 0, irq_delta = 0;
835 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
836 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
839 * Since irq_time is only updated on {soft,}irq_exit, we might run into
840 * this case when a previous update_rq_clock() happened inside a
843 * When this happens, we stop ->clock_task and only update the
844 * prev_irq_time stamp to account for the part that fit, so that a next
845 * update will consume the rest. This ensures ->clock_task is
848 * It does however cause some slight miss-attribution of {soft,}irq
849 * time, a more accurate solution would be to update the irq_time using
850 * the current rq->clock timestamp, except that would require using
853 if (irq_delta > delta)
856 rq->prev_irq_time += irq_delta;
859 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
860 if (static_key_false((¶virt_steal_rq_enabled))) {
861 steal = paravirt_steal_clock(cpu_of(rq));
862 steal -= rq->prev_steal_time_rq;
864 if (unlikely(steal > delta))
867 rq->prev_steal_time_rq += steal;
872 rq->clock_task += delta;
874 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
875 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
876 sched_rt_avg_update(rq, irq_delta + steal);
880 void sched_set_stop_task(int cpu, struct task_struct *stop)
882 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
883 struct task_struct *old_stop = cpu_rq(cpu)->stop;
887 * Make it appear like a SCHED_FIFO task, its something
888 * userspace knows about and won't get confused about.
890 * Also, it will make PI more or less work without too
891 * much confusion -- but then, stop work should not
892 * rely on PI working anyway.
894 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
896 stop->sched_class = &stop_sched_class;
899 cpu_rq(cpu)->stop = stop;
903 * Reset it back to a normal scheduling class so that
904 * it can die in pieces.
906 old_stop->sched_class = &rt_sched_class;
911 * __normal_prio - return the priority that is based on the static prio
913 static inline int __normal_prio(struct task_struct *p)
915 return p->static_prio;
919 * Calculate the expected normal priority: i.e. priority
920 * without taking RT-inheritance into account. Might be
921 * boosted by interactivity modifiers. Changes upon fork,
922 * setprio syscalls, and whenever the interactivity
923 * estimator recalculates.
925 static inline int normal_prio(struct task_struct *p)
929 if (task_has_dl_policy(p))
930 prio = MAX_DL_PRIO-1;
931 else if (task_has_rt_policy(p))
932 prio = MAX_RT_PRIO-1 - p->rt_priority;
934 prio = __normal_prio(p);
939 * Calculate the current priority, i.e. the priority
940 * taken into account by the scheduler. This value might
941 * be boosted by RT tasks, or might be boosted by
942 * interactivity modifiers. Will be RT if the task got
943 * RT-boosted. If not then it returns p->normal_prio.
945 static int effective_prio(struct task_struct *p)
947 p->normal_prio = normal_prio(p);
949 * If we are RT tasks or we were boosted to RT priority,
950 * keep the priority unchanged. Otherwise, update priority
951 * to the normal priority:
953 if (!rt_prio(p->prio))
954 return p->normal_prio;
959 * task_curr - is this task currently executing on a CPU?
960 * @p: the task in question.
962 * Return: 1 if the task is currently executing. 0 otherwise.
964 inline int task_curr(const struct task_struct *p)
966 return cpu_curr(task_cpu(p)) == p;
969 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
970 const struct sched_class *prev_class,
973 if (prev_class != p->sched_class) {
974 if (prev_class->switched_from)
975 prev_class->switched_from(rq, p);
976 p->sched_class->switched_to(rq, p);
977 } else if (oldprio != p->prio || dl_task(p))
978 p->sched_class->prio_changed(rq, p, oldprio);
981 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
983 const struct sched_class *class;
985 if (p->sched_class == rq->curr->sched_class) {
986 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
988 for_each_class(class) {
989 if (class == rq->curr->sched_class)
991 if (class == p->sched_class) {
992 resched_task(rq->curr);
999 * A queue event has occurred, and we're going to schedule. In
1000 * this case, we can save a useless back to back clock update.
1002 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
1003 rq->skip_clock_update = 1;
1007 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1009 #ifdef CONFIG_SCHED_DEBUG
1011 * We should never call set_task_cpu() on a blocked task,
1012 * ttwu() will sort out the placement.
1014 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1015 !(task_preempt_count(p) & PREEMPT_ACTIVE));
1017 #ifdef CONFIG_LOCKDEP
1019 * The caller should hold either p->pi_lock or rq->lock, when changing
1020 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1022 * sched_move_task() holds both and thus holding either pins the cgroup,
1025 * Furthermore, all task_rq users should acquire both locks, see
1028 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1029 lockdep_is_held(&task_rq(p)->lock)));
1033 trace_sched_migrate_task(p, new_cpu);
1035 if (task_cpu(p) != new_cpu) {
1036 if (p->sched_class->migrate_task_rq)
1037 p->sched_class->migrate_task_rq(p, new_cpu);
1038 p->se.nr_migrations++;
1039 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
1042 __set_task_cpu(p, new_cpu);
1045 static void __migrate_swap_task(struct task_struct *p, int cpu)
1048 struct rq *src_rq, *dst_rq;
1050 src_rq = task_rq(p);
1051 dst_rq = cpu_rq(cpu);
1053 deactivate_task(src_rq, p, 0);
1054 set_task_cpu(p, cpu);
1055 activate_task(dst_rq, p, 0);
1056 check_preempt_curr(dst_rq, p, 0);
1059 * Task isn't running anymore; make it appear like we migrated
1060 * it before it went to sleep. This means on wakeup we make the
1061 * previous cpu our targer instead of where it really is.
1067 struct migration_swap_arg {
1068 struct task_struct *src_task, *dst_task;
1069 int src_cpu, dst_cpu;
1072 static int migrate_swap_stop(void *data)
1074 struct migration_swap_arg *arg = data;
1075 struct rq *src_rq, *dst_rq;
1078 src_rq = cpu_rq(arg->src_cpu);
1079 dst_rq = cpu_rq(arg->dst_cpu);
1081 double_raw_lock(&arg->src_task->pi_lock,
1082 &arg->dst_task->pi_lock);
1083 double_rq_lock(src_rq, dst_rq);
1084 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1087 if (task_cpu(arg->src_task) != arg->src_cpu)
1090 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1093 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1096 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1097 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1102 double_rq_unlock(src_rq, dst_rq);
1103 raw_spin_unlock(&arg->dst_task->pi_lock);
1104 raw_spin_unlock(&arg->src_task->pi_lock);
1110 * Cross migrate two tasks
1112 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1114 struct migration_swap_arg arg;
1117 arg = (struct migration_swap_arg){
1119 .src_cpu = task_cpu(cur),
1121 .dst_cpu = task_cpu(p),
1124 if (arg.src_cpu == arg.dst_cpu)
1128 * These three tests are all lockless; this is OK since all of them
1129 * will be re-checked with proper locks held further down the line.
1131 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1134 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1137 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1140 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1141 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1147 struct migration_arg {
1148 struct task_struct *task;
1152 static int migration_cpu_stop(void *data);
1155 * wait_task_inactive - wait for a thread to unschedule.
1157 * If @match_state is nonzero, it's the @p->state value just checked and
1158 * not expected to change. If it changes, i.e. @p might have woken up,
1159 * then return zero. When we succeed in waiting for @p to be off its CPU,
1160 * we return a positive number (its total switch count). If a second call
1161 * a short while later returns the same number, the caller can be sure that
1162 * @p has remained unscheduled the whole time.
1164 * The caller must ensure that the task *will* unschedule sometime soon,
1165 * else this function might spin for a *long* time. This function can't
1166 * be called with interrupts off, or it may introduce deadlock with
1167 * smp_call_function() if an IPI is sent by the same process we are
1168 * waiting to become inactive.
1170 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1172 unsigned long flags;
1179 * We do the initial early heuristics without holding
1180 * any task-queue locks at all. We'll only try to get
1181 * the runqueue lock when things look like they will
1187 * If the task is actively running on another CPU
1188 * still, just relax and busy-wait without holding
1191 * NOTE! Since we don't hold any locks, it's not
1192 * even sure that "rq" stays as the right runqueue!
1193 * But we don't care, since "task_running()" will
1194 * return false if the runqueue has changed and p
1195 * is actually now running somewhere else!
1197 while (task_running(rq, p)) {
1198 if (match_state && unlikely(p->state != match_state))
1204 * Ok, time to look more closely! We need the rq
1205 * lock now, to be *sure*. If we're wrong, we'll
1206 * just go back and repeat.
1208 rq = task_rq_lock(p, &flags);
1209 trace_sched_wait_task(p);
1210 running = task_running(rq, p);
1213 if (!match_state || p->state == match_state)
1214 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1215 task_rq_unlock(rq, p, &flags);
1218 * If it changed from the expected state, bail out now.
1220 if (unlikely(!ncsw))
1224 * Was it really running after all now that we
1225 * checked with the proper locks actually held?
1227 * Oops. Go back and try again..
1229 if (unlikely(running)) {
1235 * It's not enough that it's not actively running,
1236 * it must be off the runqueue _entirely_, and not
1239 * So if it was still runnable (but just not actively
1240 * running right now), it's preempted, and we should
1241 * yield - it could be a while.
1243 if (unlikely(on_rq)) {
1244 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1246 set_current_state(TASK_UNINTERRUPTIBLE);
1247 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1252 * Ahh, all good. It wasn't running, and it wasn't
1253 * runnable, which means that it will never become
1254 * running in the future either. We're all done!
1263 * kick_process - kick a running thread to enter/exit the kernel
1264 * @p: the to-be-kicked thread
1266 * Cause a process which is running on another CPU to enter
1267 * kernel-mode, without any delay. (to get signals handled.)
1269 * NOTE: this function doesn't have to take the runqueue lock,
1270 * because all it wants to ensure is that the remote task enters
1271 * the kernel. If the IPI races and the task has been migrated
1272 * to another CPU then no harm is done and the purpose has been
1275 void kick_process(struct task_struct *p)
1281 if ((cpu != smp_processor_id()) && task_curr(p))
1282 smp_send_reschedule(cpu);
1285 EXPORT_SYMBOL_GPL(kick_process);
1286 #endif /* CONFIG_SMP */
1290 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1292 static int select_fallback_rq(int cpu, struct task_struct *p)
1294 int nid = cpu_to_node(cpu);
1295 const struct cpumask *nodemask = NULL;
1296 enum { cpuset, possible, fail } state = cpuset;
1300 * If the node that the cpu is on has been offlined, cpu_to_node()
1301 * will return -1. There is no cpu on the node, and we should
1302 * select the cpu on the other node.
1305 nodemask = cpumask_of_node(nid);
1307 /* Look for allowed, online CPU in same node. */
1308 for_each_cpu(dest_cpu, nodemask) {
1309 if (!cpu_online(dest_cpu))
1311 if (!cpu_active(dest_cpu))
1313 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1319 /* Any allowed, online CPU? */
1320 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1321 if (!cpu_online(dest_cpu))
1323 if (!cpu_active(dest_cpu))
1330 /* No more Mr. Nice Guy. */
1331 cpuset_cpus_allowed_fallback(p);
1336 do_set_cpus_allowed(p, cpu_possible_mask);
1347 if (state != cpuset) {
1349 * Don't tell them about moving exiting tasks or
1350 * kernel threads (both mm NULL), since they never
1353 if (p->mm && printk_ratelimit()) {
1354 printk_sched("process %d (%s) no longer affine to cpu%d\n",
1355 task_pid_nr(p), p->comm, cpu);
1363 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1366 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1368 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1371 * In order not to call set_task_cpu() on a blocking task we need
1372 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1375 * Since this is common to all placement strategies, this lives here.
1377 * [ this allows ->select_task() to simply return task_cpu(p) and
1378 * not worry about this generic constraint ]
1380 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1382 cpu = select_fallback_rq(task_cpu(p), p);
1387 static void update_avg(u64 *avg, u64 sample)
1389 s64 diff = sample - *avg;
1395 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1397 #ifdef CONFIG_SCHEDSTATS
1398 struct rq *rq = this_rq();
1401 int this_cpu = smp_processor_id();
1403 if (cpu == this_cpu) {
1404 schedstat_inc(rq, ttwu_local);
1405 schedstat_inc(p, se.statistics.nr_wakeups_local);
1407 struct sched_domain *sd;
1409 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1411 for_each_domain(this_cpu, sd) {
1412 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1413 schedstat_inc(sd, ttwu_wake_remote);
1420 if (wake_flags & WF_MIGRATED)
1421 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1423 #endif /* CONFIG_SMP */
1425 schedstat_inc(rq, ttwu_count);
1426 schedstat_inc(p, se.statistics.nr_wakeups);
1428 if (wake_flags & WF_SYNC)
1429 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1431 #endif /* CONFIG_SCHEDSTATS */
1434 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1436 activate_task(rq, p, en_flags);
1439 /* if a worker is waking up, notify workqueue */
1440 if (p->flags & PF_WQ_WORKER)
1441 wq_worker_waking_up(p, cpu_of(rq));
1445 * Mark the task runnable and perform wakeup-preemption.
1448 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1450 check_preempt_curr(rq, p, wake_flags);
1451 trace_sched_wakeup(p, true);
1453 p->state = TASK_RUNNING;
1455 if (p->sched_class->task_woken)
1456 p->sched_class->task_woken(rq, p);
1458 if (rq->idle_stamp) {
1459 u64 delta = rq_clock(rq) - rq->idle_stamp;
1460 u64 max = 2*rq->max_idle_balance_cost;
1462 update_avg(&rq->avg_idle, delta);
1464 if (rq->avg_idle > max)
1473 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1476 if (p->sched_contributes_to_load)
1477 rq->nr_uninterruptible--;
1480 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1481 ttwu_do_wakeup(rq, p, wake_flags);
1485 * Called in case the task @p isn't fully descheduled from its runqueue,
1486 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1487 * since all we need to do is flip p->state to TASK_RUNNING, since
1488 * the task is still ->on_rq.
1490 static int ttwu_remote(struct task_struct *p, int wake_flags)
1495 rq = __task_rq_lock(p);
1497 /* check_preempt_curr() may use rq clock */
1498 update_rq_clock(rq);
1499 ttwu_do_wakeup(rq, p, wake_flags);
1502 __task_rq_unlock(rq);
1508 static void sched_ttwu_pending(void)
1510 struct rq *rq = this_rq();
1511 struct llist_node *llist = llist_del_all(&rq->wake_list);
1512 struct task_struct *p;
1514 raw_spin_lock(&rq->lock);
1517 p = llist_entry(llist, struct task_struct, wake_entry);
1518 llist = llist_next(llist);
1519 ttwu_do_activate(rq, p, 0);
1522 raw_spin_unlock(&rq->lock);
1525 void scheduler_ipi(void)
1528 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1529 * TIF_NEED_RESCHED remotely (for the first time) will also send
1532 preempt_fold_need_resched();
1534 if (llist_empty(&this_rq()->wake_list)
1535 && !tick_nohz_full_cpu(smp_processor_id())
1536 && !got_nohz_idle_kick())
1540 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1541 * traditionally all their work was done from the interrupt return
1542 * path. Now that we actually do some work, we need to make sure
1545 * Some archs already do call them, luckily irq_enter/exit nest
1548 * Arguably we should visit all archs and update all handlers,
1549 * however a fair share of IPIs are still resched only so this would
1550 * somewhat pessimize the simple resched case.
1553 tick_nohz_full_check();
1554 sched_ttwu_pending();
1557 * Check if someone kicked us for doing the nohz idle load balance.
1559 if (unlikely(got_nohz_idle_kick())) {
1560 this_rq()->idle_balance = 1;
1561 raise_softirq_irqoff(SCHED_SOFTIRQ);
1566 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1568 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list))
1569 smp_send_reschedule(cpu);
1572 bool cpus_share_cache(int this_cpu, int that_cpu)
1574 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1576 #endif /* CONFIG_SMP */
1578 static void ttwu_queue(struct task_struct *p, int cpu)
1580 struct rq *rq = cpu_rq(cpu);
1582 #if defined(CONFIG_SMP)
1583 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1584 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1585 ttwu_queue_remote(p, cpu);
1590 raw_spin_lock(&rq->lock);
1591 ttwu_do_activate(rq, p, 0);
1592 raw_spin_unlock(&rq->lock);
1596 * try_to_wake_up - wake up a thread
1597 * @p: the thread to be awakened
1598 * @state: the mask of task states that can be woken
1599 * @wake_flags: wake modifier flags (WF_*)
1601 * Put it on the run-queue if it's not already there. The "current"
1602 * thread is always on the run-queue (except when the actual
1603 * re-schedule is in progress), and as such you're allowed to do
1604 * the simpler "current->state = TASK_RUNNING" to mark yourself
1605 * runnable without the overhead of this.
1607 * Return: %true if @p was woken up, %false if it was already running.
1608 * or @state didn't match @p's state.
1611 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1613 unsigned long flags;
1614 int cpu, success = 0;
1617 * If we are going to wake up a thread waiting for CONDITION we
1618 * need to ensure that CONDITION=1 done by the caller can not be
1619 * reordered with p->state check below. This pairs with mb() in
1620 * set_current_state() the waiting thread does.
1622 smp_mb__before_spinlock();
1623 raw_spin_lock_irqsave(&p->pi_lock, flags);
1624 if (!(p->state & state))
1627 success = 1; /* we're going to change ->state */
1630 if (p->on_rq && ttwu_remote(p, wake_flags))
1635 * If the owning (remote) cpu is still in the middle of schedule() with
1636 * this task as prev, wait until its done referencing the task.
1641 * Pairs with the smp_wmb() in finish_lock_switch().
1645 p->sched_contributes_to_load = !!task_contributes_to_load(p);
1646 p->state = TASK_WAKING;
1648 if (p->sched_class->task_waking)
1649 p->sched_class->task_waking(p);
1651 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
1652 if (task_cpu(p) != cpu) {
1653 wake_flags |= WF_MIGRATED;
1654 set_task_cpu(p, cpu);
1656 #endif /* CONFIG_SMP */
1660 ttwu_stat(p, cpu, wake_flags);
1662 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1668 * try_to_wake_up_local - try to wake up a local task with rq lock held
1669 * @p: the thread to be awakened
1671 * Put @p on the run-queue if it's not already there. The caller must
1672 * ensure that this_rq() is locked, @p is bound to this_rq() and not
1675 static void try_to_wake_up_local(struct task_struct *p)
1677 struct rq *rq = task_rq(p);
1679 if (WARN_ON_ONCE(rq != this_rq()) ||
1680 WARN_ON_ONCE(p == current))
1683 lockdep_assert_held(&rq->lock);
1685 if (!raw_spin_trylock(&p->pi_lock)) {
1686 raw_spin_unlock(&rq->lock);
1687 raw_spin_lock(&p->pi_lock);
1688 raw_spin_lock(&rq->lock);
1691 if (!(p->state & TASK_NORMAL))
1695 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
1697 ttwu_do_wakeup(rq, p, 0);
1698 ttwu_stat(p, smp_processor_id(), 0);
1700 raw_spin_unlock(&p->pi_lock);
1704 * wake_up_process - Wake up a specific process
1705 * @p: The process to be woken up.
1707 * Attempt to wake up the nominated process and move it to the set of runnable
1710 * Return: 1 if the process was woken up, 0 if it was already running.
1712 * It may be assumed that this function implies a write memory barrier before
1713 * changing the task state if and only if any tasks are woken up.
1715 int wake_up_process(struct task_struct *p)
1717 WARN_ON(task_is_stopped_or_traced(p));
1718 return try_to_wake_up(p, TASK_NORMAL, 0);
1720 EXPORT_SYMBOL(wake_up_process);
1722 int wake_up_state(struct task_struct *p, unsigned int state)
1724 return try_to_wake_up(p, state, 0);
1728 * Perform scheduler related setup for a newly forked process p.
1729 * p is forked by current.
1731 * __sched_fork() is basic setup used by init_idle() too:
1733 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
1738 p->se.exec_start = 0;
1739 p->se.sum_exec_runtime = 0;
1740 p->se.prev_sum_exec_runtime = 0;
1741 p->se.nr_migrations = 0;
1743 INIT_LIST_HEAD(&p->se.group_node);
1745 #ifdef CONFIG_SCHEDSTATS
1746 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
1749 RB_CLEAR_NODE(&p->dl.rb_node);
1750 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1751 p->dl.dl_runtime = p->dl.runtime = 0;
1752 p->dl.dl_deadline = p->dl.deadline = 0;
1753 p->dl.dl_period = 0;
1756 INIT_LIST_HEAD(&p->rt.run_list);
1758 #ifdef CONFIG_PREEMPT_NOTIFIERS
1759 INIT_HLIST_HEAD(&p->preempt_notifiers);
1762 #ifdef CONFIG_NUMA_BALANCING
1763 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
1764 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
1765 p->mm->numa_scan_seq = 0;
1768 if (clone_flags & CLONE_VM)
1769 p->numa_preferred_nid = current->numa_preferred_nid;
1771 p->numa_preferred_nid = -1;
1773 p->node_stamp = 0ULL;
1774 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
1775 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
1776 p->numa_work.next = &p->numa_work;
1777 p->numa_faults_memory = NULL;
1778 p->numa_faults_buffer_memory = NULL;
1779 p->last_task_numa_placement = 0;
1780 p->last_sum_exec_runtime = 0;
1782 INIT_LIST_HEAD(&p->numa_entry);
1783 p->numa_group = NULL;
1784 #endif /* CONFIG_NUMA_BALANCING */
1787 #ifdef CONFIG_NUMA_BALANCING
1788 #ifdef CONFIG_SCHED_DEBUG
1789 void set_numabalancing_state(bool enabled)
1792 sched_feat_set("NUMA");
1794 sched_feat_set("NO_NUMA");
1797 __read_mostly bool numabalancing_enabled;
1799 void set_numabalancing_state(bool enabled)
1801 numabalancing_enabled = enabled;
1803 #endif /* CONFIG_SCHED_DEBUG */
1805 #ifdef CONFIG_PROC_SYSCTL
1806 int sysctl_numa_balancing(struct ctl_table *table, int write,
1807 void __user *buffer, size_t *lenp, loff_t *ppos)
1811 int state = numabalancing_enabled;
1813 if (write && !capable(CAP_SYS_ADMIN))
1818 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
1822 set_numabalancing_state(state);
1829 * fork()/clone()-time setup:
1831 int sched_fork(unsigned long clone_flags, struct task_struct *p)
1833 unsigned long flags;
1834 int cpu = get_cpu();
1836 __sched_fork(clone_flags, p);
1838 * We mark the process as running here. This guarantees that
1839 * nobody will actually run it, and a signal or other external
1840 * event cannot wake it up and insert it on the runqueue either.
1842 p->state = TASK_RUNNING;
1845 * Make sure we do not leak PI boosting priority to the child.
1847 p->prio = current->normal_prio;
1850 * Revert to default priority/policy on fork if requested.
1852 if (unlikely(p->sched_reset_on_fork)) {
1853 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
1854 p->policy = SCHED_NORMAL;
1855 p->static_prio = NICE_TO_PRIO(0);
1857 } else if (PRIO_TO_NICE(p->static_prio) < 0)
1858 p->static_prio = NICE_TO_PRIO(0);
1860 p->prio = p->normal_prio = __normal_prio(p);
1864 * We don't need the reset flag anymore after the fork. It has
1865 * fulfilled its duty:
1867 p->sched_reset_on_fork = 0;
1870 if (dl_prio(p->prio)) {
1873 } else if (rt_prio(p->prio)) {
1874 p->sched_class = &rt_sched_class;
1876 p->sched_class = &fair_sched_class;
1879 if (p->sched_class->task_fork)
1880 p->sched_class->task_fork(p);
1883 * The child is not yet in the pid-hash so no cgroup attach races,
1884 * and the cgroup is pinned to this child due to cgroup_fork()
1885 * is ran before sched_fork().
1887 * Silence PROVE_RCU.
1889 raw_spin_lock_irqsave(&p->pi_lock, flags);
1890 set_task_cpu(p, cpu);
1891 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
1893 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1894 if (likely(sched_info_on()))
1895 memset(&p->sched_info, 0, sizeof(p->sched_info));
1897 #if defined(CONFIG_SMP)
1900 init_task_preempt_count(p);
1902 plist_node_init(&p->pushable_tasks, MAX_PRIO);
1903 RB_CLEAR_NODE(&p->pushable_dl_tasks);
1910 unsigned long to_ratio(u64 period, u64 runtime)
1912 if (runtime == RUNTIME_INF)
1916 * Doing this here saves a lot of checks in all
1917 * the calling paths, and returning zero seems
1918 * safe for them anyway.
1923 return div64_u64(runtime << 20, period);
1927 inline struct dl_bw *dl_bw_of(int i)
1929 return &cpu_rq(i)->rd->dl_bw;
1932 static inline int dl_bw_cpus(int i)
1934 struct root_domain *rd = cpu_rq(i)->rd;
1937 for_each_cpu_and(i, rd->span, cpu_active_mask)
1943 inline struct dl_bw *dl_bw_of(int i)
1945 return &cpu_rq(i)->dl.dl_bw;
1948 static inline int dl_bw_cpus(int i)
1955 void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw)
1957 dl_b->total_bw -= tsk_bw;
1961 void __dl_add(struct dl_bw *dl_b, u64 tsk_bw)
1963 dl_b->total_bw += tsk_bw;
1967 bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw)
1969 return dl_b->bw != -1 &&
1970 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw;
1974 * We must be sure that accepting a new task (or allowing changing the
1975 * parameters of an existing one) is consistent with the bandwidth
1976 * constraints. If yes, this function also accordingly updates the currently
1977 * allocated bandwidth to reflect the new situation.
1979 * This function is called while holding p's rq->lock.
1981 static int dl_overflow(struct task_struct *p, int policy,
1982 const struct sched_attr *attr)
1985 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
1986 u64 period = attr->sched_period ?: attr->sched_deadline;
1987 u64 runtime = attr->sched_runtime;
1988 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
1991 if (new_bw == p->dl.dl_bw)
1995 * Either if a task, enters, leave, or stays -deadline but changes
1996 * its parameters, we may need to update accordingly the total
1997 * allocated bandwidth of the container.
1999 raw_spin_lock(&dl_b->lock);
2000 cpus = dl_bw_cpus(task_cpu(p));
2001 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2002 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2003 __dl_add(dl_b, new_bw);
2005 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2006 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2007 __dl_clear(dl_b, p->dl.dl_bw);
2008 __dl_add(dl_b, new_bw);
2010 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2011 __dl_clear(dl_b, p->dl.dl_bw);
2014 raw_spin_unlock(&dl_b->lock);
2019 extern void init_dl_bw(struct dl_bw *dl_b);
2022 * wake_up_new_task - wake up a newly created task for the first time.
2024 * This function will do some initial scheduler statistics housekeeping
2025 * that must be done for every newly created context, then puts the task
2026 * on the runqueue and wakes it.
2028 void wake_up_new_task(struct task_struct *p)
2030 unsigned long flags;
2033 raw_spin_lock_irqsave(&p->pi_lock, flags);
2036 * Fork balancing, do it here and not earlier because:
2037 * - cpus_allowed can change in the fork path
2038 * - any previously selected cpu might disappear through hotplug
2040 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2043 /* Initialize new task's runnable average */
2044 init_task_runnable_average(p);
2045 rq = __task_rq_lock(p);
2046 activate_task(rq, p, 0);
2048 trace_sched_wakeup_new(p, true);
2049 check_preempt_curr(rq, p, WF_FORK);
2051 if (p->sched_class->task_woken)
2052 p->sched_class->task_woken(rq, p);
2054 task_rq_unlock(rq, p, &flags);
2057 #ifdef CONFIG_PREEMPT_NOTIFIERS
2060 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2061 * @notifier: notifier struct to register
2063 void preempt_notifier_register(struct preempt_notifier *notifier)
2065 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2067 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2070 * preempt_notifier_unregister - no longer interested in preemption notifications
2071 * @notifier: notifier struct to unregister
2073 * This is safe to call from within a preemption notifier.
2075 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2077 hlist_del(¬ifier->link);
2079 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2081 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2083 struct preempt_notifier *notifier;
2085 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2086 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2090 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2091 struct task_struct *next)
2093 struct preempt_notifier *notifier;
2095 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2096 notifier->ops->sched_out(notifier, next);
2099 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2101 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2106 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2107 struct task_struct *next)
2111 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2114 * prepare_task_switch - prepare to switch tasks
2115 * @rq: the runqueue preparing to switch
2116 * @prev: the current task that is being switched out
2117 * @next: the task we are going to switch to.
2119 * This is called with the rq lock held and interrupts off. It must
2120 * be paired with a subsequent finish_task_switch after the context
2123 * prepare_task_switch sets up locking and calls architecture specific
2127 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2128 struct task_struct *next)
2130 trace_sched_switch(prev, next);
2131 sched_info_switch(rq, prev, next);
2132 perf_event_task_sched_out(prev, next);
2133 fire_sched_out_preempt_notifiers(prev, next);
2134 prepare_lock_switch(rq, next);
2135 prepare_arch_switch(next);
2139 * finish_task_switch - clean up after a task-switch
2140 * @rq: runqueue associated with task-switch
2141 * @prev: the thread we just switched away from.
2143 * finish_task_switch must be called after the context switch, paired
2144 * with a prepare_task_switch call before the context switch.
2145 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2146 * and do any other architecture-specific cleanup actions.
2148 * Note that we may have delayed dropping an mm in context_switch(). If
2149 * so, we finish that here outside of the runqueue lock. (Doing it
2150 * with the lock held can cause deadlocks; see schedule() for
2153 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2154 __releases(rq->lock)
2156 struct mm_struct *mm = rq->prev_mm;
2162 * A task struct has one reference for the use as "current".
2163 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2164 * schedule one last time. The schedule call will never return, and
2165 * the scheduled task must drop that reference.
2166 * The test for TASK_DEAD must occur while the runqueue locks are
2167 * still held, otherwise prev could be scheduled on another cpu, die
2168 * there before we look at prev->state, and then the reference would
2170 * Manfred Spraul <manfred@colorfullife.com>
2172 prev_state = prev->state;
2173 vtime_task_switch(prev);
2174 finish_arch_switch(prev);
2175 perf_event_task_sched_in(prev, current);
2176 finish_lock_switch(rq, prev);
2177 finish_arch_post_lock_switch();
2179 fire_sched_in_preempt_notifiers(current);
2182 if (unlikely(prev_state == TASK_DEAD)) {
2183 if (prev->sched_class->task_dead)
2184 prev->sched_class->task_dead(prev);
2187 * Remove function-return probe instances associated with this
2188 * task and put them back on the free list.
2190 kprobe_flush_task(prev);
2191 put_task_struct(prev);
2194 tick_nohz_task_switch(current);
2199 /* rq->lock is NOT held, but preemption is disabled */
2200 static inline void post_schedule(struct rq *rq)
2202 if (rq->post_schedule) {
2203 unsigned long flags;
2205 raw_spin_lock_irqsave(&rq->lock, flags);
2206 if (rq->curr->sched_class->post_schedule)
2207 rq->curr->sched_class->post_schedule(rq);
2208 raw_spin_unlock_irqrestore(&rq->lock, flags);
2210 rq->post_schedule = 0;
2216 static inline void post_schedule(struct rq *rq)
2223 * schedule_tail - first thing a freshly forked thread must call.
2224 * @prev: the thread we just switched away from.
2226 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2227 __releases(rq->lock)
2229 struct rq *rq = this_rq();
2231 finish_task_switch(rq, prev);
2234 * FIXME: do we need to worry about rq being invalidated by the
2239 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2240 /* In this case, finish_task_switch does not reenable preemption */
2243 if (current->set_child_tid)
2244 put_user(task_pid_vnr(current), current->set_child_tid);
2248 * context_switch - switch to the new MM and the new
2249 * thread's register state.
2252 context_switch(struct rq *rq, struct task_struct *prev,
2253 struct task_struct *next)
2255 struct mm_struct *mm, *oldmm;
2257 prepare_task_switch(rq, prev, next);
2260 oldmm = prev->active_mm;
2262 * For paravirt, this is coupled with an exit in switch_to to
2263 * combine the page table reload and the switch backend into
2266 arch_start_context_switch(prev);
2269 next->active_mm = oldmm;
2270 atomic_inc(&oldmm->mm_count);
2271 enter_lazy_tlb(oldmm, next);
2273 switch_mm(oldmm, mm, next);
2276 prev->active_mm = NULL;
2277 rq->prev_mm = oldmm;
2280 * Since the runqueue lock will be released by the next
2281 * task (which is an invalid locking op but in the case
2282 * of the scheduler it's an obvious special-case), so we
2283 * do an early lockdep release here:
2285 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2286 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2289 context_tracking_task_switch(prev, next);
2290 /* Here we just switch the register state and the stack. */
2291 switch_to(prev, next, prev);
2295 * this_rq must be evaluated again because prev may have moved
2296 * CPUs since it called schedule(), thus the 'rq' on its stack
2297 * frame will be invalid.
2299 finish_task_switch(this_rq(), prev);
2303 * nr_running and nr_context_switches:
2305 * externally visible scheduler statistics: current number of runnable
2306 * threads, total number of context switches performed since bootup.
2308 unsigned long nr_running(void)
2310 unsigned long i, sum = 0;
2312 for_each_online_cpu(i)
2313 sum += cpu_rq(i)->nr_running;
2318 unsigned long long nr_context_switches(void)
2321 unsigned long long sum = 0;
2323 for_each_possible_cpu(i)
2324 sum += cpu_rq(i)->nr_switches;
2329 unsigned long nr_iowait(void)
2331 unsigned long i, sum = 0;
2333 for_each_possible_cpu(i)
2334 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2339 unsigned long nr_iowait_cpu(int cpu)
2341 struct rq *this = cpu_rq(cpu);
2342 return atomic_read(&this->nr_iowait);
2348 * sched_exec - execve() is a valuable balancing opportunity, because at
2349 * this point the task has the smallest effective memory and cache footprint.
2351 void sched_exec(void)
2353 struct task_struct *p = current;
2354 unsigned long flags;
2357 raw_spin_lock_irqsave(&p->pi_lock, flags);
2358 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2359 if (dest_cpu == smp_processor_id())
2362 if (likely(cpu_active(dest_cpu))) {
2363 struct migration_arg arg = { p, dest_cpu };
2365 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2366 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2370 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2375 DEFINE_PER_CPU(struct kernel_stat, kstat);
2376 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2378 EXPORT_PER_CPU_SYMBOL(kstat);
2379 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2382 * Return any ns on the sched_clock that have not yet been accounted in
2383 * @p in case that task is currently running.
2385 * Called with task_rq_lock() held on @rq.
2387 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
2391 if (task_current(rq, p)) {
2392 update_rq_clock(rq);
2393 ns = rq_clock_task(rq) - p->se.exec_start;
2401 unsigned long long task_delta_exec(struct task_struct *p)
2403 unsigned long flags;
2407 rq = task_rq_lock(p, &flags);
2408 ns = do_task_delta_exec(p, rq);
2409 task_rq_unlock(rq, p, &flags);
2415 * Return accounted runtime for the task.
2416 * In case the task is currently running, return the runtime plus current's
2417 * pending runtime that have not been accounted yet.
2419 unsigned long long task_sched_runtime(struct task_struct *p)
2421 unsigned long flags;
2425 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2427 * 64-bit doesn't need locks to atomically read a 64bit value.
2428 * So we have a optimization chance when the task's delta_exec is 0.
2429 * Reading ->on_cpu is racy, but this is ok.
2431 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2432 * If we race with it entering cpu, unaccounted time is 0. This is
2433 * indistinguishable from the read occurring a few cycles earlier.
2436 return p->se.sum_exec_runtime;
2439 rq = task_rq_lock(p, &flags);
2440 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
2441 task_rq_unlock(rq, p, &flags);
2447 * This function gets called by the timer code, with HZ frequency.
2448 * We call it with interrupts disabled.
2450 void scheduler_tick(void)
2452 int cpu = smp_processor_id();
2453 struct rq *rq = cpu_rq(cpu);
2454 struct task_struct *curr = rq->curr;
2458 raw_spin_lock(&rq->lock);
2459 update_rq_clock(rq);
2460 curr->sched_class->task_tick(rq, curr, 0);
2461 update_cpu_load_active(rq);
2462 raw_spin_unlock(&rq->lock);
2464 perf_event_task_tick();
2467 rq->idle_balance = idle_cpu(cpu);
2468 trigger_load_balance(rq);
2470 rq_last_tick_reset(rq);
2473 #ifdef CONFIG_NO_HZ_FULL
2475 * scheduler_tick_max_deferment
2477 * Keep at least one tick per second when a single
2478 * active task is running because the scheduler doesn't
2479 * yet completely support full dynticks environment.
2481 * This makes sure that uptime, CFS vruntime, load
2482 * balancing, etc... continue to move forward, even
2483 * with a very low granularity.
2485 * Return: Maximum deferment in nanoseconds.
2487 u64 scheduler_tick_max_deferment(void)
2489 struct rq *rq = this_rq();
2490 unsigned long next, now = ACCESS_ONCE(jiffies);
2492 next = rq->last_sched_tick + HZ;
2494 if (time_before_eq(next, now))
2497 return jiffies_to_nsecs(next - now);
2501 notrace unsigned long get_parent_ip(unsigned long addr)
2503 if (in_lock_functions(addr)) {
2504 addr = CALLER_ADDR2;
2505 if (in_lock_functions(addr))
2506 addr = CALLER_ADDR3;
2511 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
2512 defined(CONFIG_PREEMPT_TRACER))
2514 void __kprobes preempt_count_add(int val)
2516 #ifdef CONFIG_DEBUG_PREEMPT
2520 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
2523 __preempt_count_add(val);
2524 #ifdef CONFIG_DEBUG_PREEMPT
2526 * Spinlock count overflowing soon?
2528 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
2531 if (preempt_count() == val) {
2532 unsigned long ip = get_parent_ip(CALLER_ADDR1);
2533 #ifdef CONFIG_DEBUG_PREEMPT
2534 current->preempt_disable_ip = ip;
2536 trace_preempt_off(CALLER_ADDR0, ip);
2539 EXPORT_SYMBOL(preempt_count_add);
2541 void __kprobes preempt_count_sub(int val)
2543 #ifdef CONFIG_DEBUG_PREEMPT
2547 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
2550 * Is the spinlock portion underflowing?
2552 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
2553 !(preempt_count() & PREEMPT_MASK)))
2557 if (preempt_count() == val)
2558 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
2559 __preempt_count_sub(val);
2561 EXPORT_SYMBOL(preempt_count_sub);
2566 * Print scheduling while atomic bug:
2568 static noinline void __schedule_bug(struct task_struct *prev)
2570 if (oops_in_progress)
2573 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
2574 prev->comm, prev->pid, preempt_count());
2576 debug_show_held_locks(prev);
2578 if (irqs_disabled())
2579 print_irqtrace_events(prev);
2580 #ifdef CONFIG_DEBUG_PREEMPT
2581 if (in_atomic_preempt_off()) {
2582 pr_err("Preemption disabled at:");
2583 print_ip_sym(current->preempt_disable_ip);
2588 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
2592 * Various schedule()-time debugging checks and statistics:
2594 static inline void schedule_debug(struct task_struct *prev)
2597 * Test if we are atomic. Since do_exit() needs to call into
2598 * schedule() atomically, we ignore that path. Otherwise whine
2599 * if we are scheduling when we should not.
2601 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD))
2602 __schedule_bug(prev);
2605 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
2607 schedstat_inc(this_rq(), sched_count);
2611 * Pick up the highest-prio task:
2613 static inline struct task_struct *
2614 pick_next_task(struct rq *rq, struct task_struct *prev)
2616 const struct sched_class *class = &fair_sched_class;
2617 struct task_struct *p;
2620 * Optimization: we know that if all tasks are in
2621 * the fair class we can call that function directly:
2623 if (likely(prev->sched_class == class &&
2624 rq->nr_running == rq->cfs.h_nr_running)) {
2625 p = fair_sched_class.pick_next_task(rq, prev);
2626 if (unlikely(p == RETRY_TASK))
2629 /* assumes fair_sched_class->next == idle_sched_class */
2631 p = idle_sched_class.pick_next_task(rq, prev);
2637 for_each_class(class) {
2638 p = class->pick_next_task(rq, prev);
2640 if (unlikely(p == RETRY_TASK))
2646 BUG(); /* the idle class will always have a runnable task */
2650 * __schedule() is the main scheduler function.
2652 * The main means of driving the scheduler and thus entering this function are:
2654 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
2656 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
2657 * paths. For example, see arch/x86/entry_64.S.
2659 * To drive preemption between tasks, the scheduler sets the flag in timer
2660 * interrupt handler scheduler_tick().
2662 * 3. Wakeups don't really cause entry into schedule(). They add a
2663 * task to the run-queue and that's it.
2665 * Now, if the new task added to the run-queue preempts the current
2666 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
2667 * called on the nearest possible occasion:
2669 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
2671 * - in syscall or exception context, at the next outmost
2672 * preempt_enable(). (this might be as soon as the wake_up()'s
2675 * - in IRQ context, return from interrupt-handler to
2676 * preemptible context
2678 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
2681 * - cond_resched() call
2682 * - explicit schedule() call
2683 * - return from syscall or exception to user-space
2684 * - return from interrupt-handler to user-space
2686 static void __sched __schedule(void)
2688 struct task_struct *prev, *next;
2689 unsigned long *switch_count;
2695 cpu = smp_processor_id();
2697 rcu_note_context_switch(cpu);
2700 schedule_debug(prev);
2702 if (sched_feat(HRTICK))
2706 * Make sure that signal_pending_state()->signal_pending() below
2707 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
2708 * done by the caller to avoid the race with signal_wake_up().
2710 smp_mb__before_spinlock();
2711 raw_spin_lock_irq(&rq->lock);
2713 switch_count = &prev->nivcsw;
2714 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
2715 if (unlikely(signal_pending_state(prev->state, prev))) {
2716 prev->state = TASK_RUNNING;
2718 deactivate_task(rq, prev, DEQUEUE_SLEEP);
2722 * If a worker went to sleep, notify and ask workqueue
2723 * whether it wants to wake up a task to maintain
2726 if (prev->flags & PF_WQ_WORKER) {
2727 struct task_struct *to_wakeup;
2729 to_wakeup = wq_worker_sleeping(prev, cpu);
2731 try_to_wake_up_local(to_wakeup);
2734 switch_count = &prev->nvcsw;
2737 if (prev->on_rq || rq->skip_clock_update < 0)
2738 update_rq_clock(rq);
2740 next = pick_next_task(rq, prev);
2741 clear_tsk_need_resched(prev);
2742 clear_preempt_need_resched();
2743 rq->skip_clock_update = 0;
2745 if (likely(prev != next)) {
2750 context_switch(rq, prev, next); /* unlocks the rq */
2752 * The context switch have flipped the stack from under us
2753 * and restored the local variables which were saved when
2754 * this task called schedule() in the past. prev == current
2755 * is still correct, but it can be moved to another cpu/rq.
2757 cpu = smp_processor_id();
2760 raw_spin_unlock_irq(&rq->lock);
2764 sched_preempt_enable_no_resched();
2769 static inline void sched_submit_work(struct task_struct *tsk)
2771 if (!tsk->state || tsk_is_pi_blocked(tsk))
2774 * If we are going to sleep and we have plugged IO queued,
2775 * make sure to submit it to avoid deadlocks.
2777 if (blk_needs_flush_plug(tsk))
2778 blk_schedule_flush_plug(tsk);
2781 asmlinkage __visible void __sched schedule(void)
2783 struct task_struct *tsk = current;
2785 sched_submit_work(tsk);
2788 EXPORT_SYMBOL(schedule);
2790 #ifdef CONFIG_CONTEXT_TRACKING
2791 asmlinkage __visible void __sched schedule_user(void)
2794 * If we come here after a random call to set_need_resched(),
2795 * or we have been woken up remotely but the IPI has not yet arrived,
2796 * we haven't yet exited the RCU idle mode. Do it here manually until
2797 * we find a better solution.
2806 * schedule_preempt_disabled - called with preemption disabled
2808 * Returns with preemption disabled. Note: preempt_count must be 1
2810 void __sched schedule_preempt_disabled(void)
2812 sched_preempt_enable_no_resched();
2817 #ifdef CONFIG_PREEMPT
2819 * this is the entry point to schedule() from in-kernel preemption
2820 * off of preempt_enable. Kernel preemptions off return from interrupt
2821 * occur there and call schedule directly.
2823 asmlinkage __visible void __sched notrace preempt_schedule(void)
2826 * If there is a non-zero preempt_count or interrupts are disabled,
2827 * we do not want to preempt the current task. Just return..
2829 if (likely(!preemptible()))
2833 __preempt_count_add(PREEMPT_ACTIVE);
2835 __preempt_count_sub(PREEMPT_ACTIVE);
2838 * Check again in case we missed a preemption opportunity
2839 * between schedule and now.
2842 } while (need_resched());
2844 EXPORT_SYMBOL(preempt_schedule);
2845 #endif /* CONFIG_PREEMPT */
2848 * this is the entry point to schedule() from kernel preemption
2849 * off of irq context.
2850 * Note, that this is called and return with irqs disabled. This will
2851 * protect us against recursive calling from irq.
2853 asmlinkage __visible void __sched preempt_schedule_irq(void)
2855 enum ctx_state prev_state;
2857 /* Catch callers which need to be fixed */
2858 BUG_ON(preempt_count() || !irqs_disabled());
2860 prev_state = exception_enter();
2863 __preempt_count_add(PREEMPT_ACTIVE);
2866 local_irq_disable();
2867 __preempt_count_sub(PREEMPT_ACTIVE);
2870 * Check again in case we missed a preemption opportunity
2871 * between schedule and now.
2874 } while (need_resched());
2876 exception_exit(prev_state);
2879 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
2882 return try_to_wake_up(curr->private, mode, wake_flags);
2884 EXPORT_SYMBOL(default_wake_function);
2886 #ifdef CONFIG_RT_MUTEXES
2889 * rt_mutex_setprio - set the current priority of a task
2891 * @prio: prio value (kernel-internal form)
2893 * This function changes the 'effective' priority of a task. It does
2894 * not touch ->normal_prio like __setscheduler().
2896 * Used by the rt_mutex code to implement priority inheritance
2897 * logic. Call site only calls if the priority of the task changed.
2899 void rt_mutex_setprio(struct task_struct *p, int prio)
2901 int oldprio, on_rq, running, enqueue_flag = 0;
2903 const struct sched_class *prev_class;
2905 BUG_ON(prio > MAX_PRIO);
2907 rq = __task_rq_lock(p);
2910 * Idle task boosting is a nono in general. There is one
2911 * exception, when PREEMPT_RT and NOHZ is active:
2913 * The idle task calls get_next_timer_interrupt() and holds
2914 * the timer wheel base->lock on the CPU and another CPU wants
2915 * to access the timer (probably to cancel it). We can safely
2916 * ignore the boosting request, as the idle CPU runs this code
2917 * with interrupts disabled and will complete the lock
2918 * protected section without being interrupted. So there is no
2919 * real need to boost.
2921 if (unlikely(p == rq->idle)) {
2922 WARN_ON(p != rq->curr);
2923 WARN_ON(p->pi_blocked_on);
2927 trace_sched_pi_setprio(p, prio);
2928 p->pi_top_task = rt_mutex_get_top_task(p);
2930 prev_class = p->sched_class;
2932 running = task_current(rq, p);
2934 dequeue_task(rq, p, 0);
2936 p->sched_class->put_prev_task(rq, p);
2939 * Boosting condition are:
2940 * 1. -rt task is running and holds mutex A
2941 * --> -dl task blocks on mutex A
2943 * 2. -dl task is running and holds mutex A
2944 * --> -dl task blocks on mutex A and could preempt the
2947 if (dl_prio(prio)) {
2948 if (!dl_prio(p->normal_prio) || (p->pi_top_task &&
2949 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) {
2950 p->dl.dl_boosted = 1;
2951 p->dl.dl_throttled = 0;
2952 enqueue_flag = ENQUEUE_REPLENISH;
2954 p->dl.dl_boosted = 0;
2955 p->sched_class = &dl_sched_class;
2956 } else if (rt_prio(prio)) {
2957 if (dl_prio(oldprio))
2958 p->dl.dl_boosted = 0;
2960 enqueue_flag = ENQUEUE_HEAD;
2961 p->sched_class = &rt_sched_class;
2963 if (dl_prio(oldprio))
2964 p->dl.dl_boosted = 0;
2965 p->sched_class = &fair_sched_class;
2971 p->sched_class->set_curr_task(rq);
2973 enqueue_task(rq, p, enqueue_flag);
2975 check_class_changed(rq, p, prev_class, oldprio);
2977 __task_rq_unlock(rq);
2981 void set_user_nice(struct task_struct *p, long nice)
2983 int old_prio, delta, on_rq;
2984 unsigned long flags;
2987 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
2990 * We have to be careful, if called from sys_setpriority(),
2991 * the task might be in the middle of scheduling on another CPU.
2993 rq = task_rq_lock(p, &flags);
2995 * The RT priorities are set via sched_setscheduler(), but we still
2996 * allow the 'normal' nice value to be set - but as expected
2997 * it wont have any effect on scheduling until the task is
2998 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3000 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3001 p->static_prio = NICE_TO_PRIO(nice);
3006 dequeue_task(rq, p, 0);
3008 p->static_prio = NICE_TO_PRIO(nice);
3011 p->prio = effective_prio(p);
3012 delta = p->prio - old_prio;
3015 enqueue_task(rq, p, 0);
3017 * If the task increased its priority or is running and
3018 * lowered its priority, then reschedule its CPU:
3020 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3021 resched_task(rq->curr);
3024 task_rq_unlock(rq, p, &flags);
3026 EXPORT_SYMBOL(set_user_nice);
3029 * can_nice - check if a task can reduce its nice value
3033 int can_nice(const struct task_struct *p, const int nice)
3035 /* convert nice value [19,-20] to rlimit style value [1,40] */
3036 int nice_rlim = 20 - nice;
3038 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3039 capable(CAP_SYS_NICE));
3042 #ifdef __ARCH_WANT_SYS_NICE
3045 * sys_nice - change the priority of the current process.
3046 * @increment: priority increment
3048 * sys_setpriority is a more generic, but much slower function that
3049 * does similar things.
3051 SYSCALL_DEFINE1(nice, int, increment)
3056 * Setpriority might change our priority at the same moment.
3057 * We don't have to worry. Conceptually one call occurs first
3058 * and we have a single winner.
3060 if (increment < -40)
3065 nice = task_nice(current) + increment;
3066 if (nice < MIN_NICE)
3068 if (nice > MAX_NICE)
3071 if (increment < 0 && !can_nice(current, nice))
3074 retval = security_task_setnice(current, nice);
3078 set_user_nice(current, nice);
3085 * task_prio - return the priority value of a given task.
3086 * @p: the task in question.
3088 * Return: The priority value as seen by users in /proc.
3089 * RT tasks are offset by -200. Normal tasks are centered
3090 * around 0, value goes from -16 to +15.
3092 int task_prio(const struct task_struct *p)
3094 return p->prio - MAX_RT_PRIO;
3098 * idle_cpu - is a given cpu idle currently?
3099 * @cpu: the processor in question.
3101 * Return: 1 if the CPU is currently idle. 0 otherwise.
3103 int idle_cpu(int cpu)
3105 struct rq *rq = cpu_rq(cpu);
3107 if (rq->curr != rq->idle)
3114 if (!llist_empty(&rq->wake_list))
3122 * idle_task - return the idle task for a given cpu.
3123 * @cpu: the processor in question.
3125 * Return: The idle task for the cpu @cpu.
3127 struct task_struct *idle_task(int cpu)
3129 return cpu_rq(cpu)->idle;
3133 * find_process_by_pid - find a process with a matching PID value.
3134 * @pid: the pid in question.
3136 * The task of @pid, if found. %NULL otherwise.
3138 static struct task_struct *find_process_by_pid(pid_t pid)
3140 return pid ? find_task_by_vpid(pid) : current;
3144 * This function initializes the sched_dl_entity of a newly becoming
3145 * SCHED_DEADLINE task.
3147 * Only the static values are considered here, the actual runtime and the
3148 * absolute deadline will be properly calculated when the task is enqueued
3149 * for the first time with its new policy.
3152 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3154 struct sched_dl_entity *dl_se = &p->dl;
3156 init_dl_task_timer(dl_se);
3157 dl_se->dl_runtime = attr->sched_runtime;
3158 dl_se->dl_deadline = attr->sched_deadline;
3159 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3160 dl_se->flags = attr->sched_flags;
3161 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3162 dl_se->dl_throttled = 0;
3164 dl_se->dl_yielded = 0;
3167 static void __setscheduler_params(struct task_struct *p,
3168 const struct sched_attr *attr)
3170 int policy = attr->sched_policy;
3172 if (policy == -1) /* setparam */
3177 if (dl_policy(policy))
3178 __setparam_dl(p, attr);
3179 else if (fair_policy(policy))
3180 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3183 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3184 * !rt_policy. Always setting this ensures that things like
3185 * getparam()/getattr() don't report silly values for !rt tasks.
3187 p->rt_priority = attr->sched_priority;
3188 p->normal_prio = normal_prio(p);
3192 /* Actually do priority change: must hold pi & rq lock. */
3193 static void __setscheduler(struct rq *rq, struct task_struct *p,
3194 const struct sched_attr *attr)
3196 __setscheduler_params(p, attr);
3199 * If we get here, there was no pi waiters boosting the
3200 * task. It is safe to use the normal prio.
3202 p->prio = normal_prio(p);
3204 if (dl_prio(p->prio))
3205 p->sched_class = &dl_sched_class;
3206 else if (rt_prio(p->prio))
3207 p->sched_class = &rt_sched_class;
3209 p->sched_class = &fair_sched_class;
3213 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3215 struct sched_dl_entity *dl_se = &p->dl;
3217 attr->sched_priority = p->rt_priority;
3218 attr->sched_runtime = dl_se->dl_runtime;
3219 attr->sched_deadline = dl_se->dl_deadline;
3220 attr->sched_period = dl_se->dl_period;
3221 attr->sched_flags = dl_se->flags;
3225 * This function validates the new parameters of a -deadline task.
3226 * We ask for the deadline not being zero, and greater or equal
3227 * than the runtime, as well as the period of being zero or
3228 * greater than deadline. Furthermore, we have to be sure that
3229 * user parameters are above the internal resolution of 1us (we
3230 * check sched_runtime only since it is always the smaller one) and
3231 * below 2^63 ns (we have to check both sched_deadline and
3232 * sched_period, as the latter can be zero).
3235 __checkparam_dl(const struct sched_attr *attr)
3238 if (attr->sched_deadline == 0)
3242 * Since we truncate DL_SCALE bits, make sure we're at least
3245 if (attr->sched_runtime < (1ULL << DL_SCALE))
3249 * Since we use the MSB for wrap-around and sign issues, make
3250 * sure it's not set (mind that period can be equal to zero).
3252 if (attr->sched_deadline & (1ULL << 63) ||
3253 attr->sched_period & (1ULL << 63))
3256 /* runtime <= deadline <= period (if period != 0) */
3257 if ((attr->sched_period != 0 &&
3258 attr->sched_period < attr->sched_deadline) ||
3259 attr->sched_deadline < attr->sched_runtime)
3266 * check the target process has a UID that matches the current process's
3268 static bool check_same_owner(struct task_struct *p)
3270 const struct cred *cred = current_cred(), *pcred;
3274 pcred = __task_cred(p);
3275 match = (uid_eq(cred->euid, pcred->euid) ||
3276 uid_eq(cred->euid, pcred->uid));
3281 static int __sched_setscheduler(struct task_struct *p,
3282 const struct sched_attr *attr,
3285 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3286 MAX_RT_PRIO - 1 - attr->sched_priority;
3287 int retval, oldprio, oldpolicy = -1, on_rq, running;
3288 int policy = attr->sched_policy;
3289 unsigned long flags;
3290 const struct sched_class *prev_class;
3294 /* may grab non-irq protected spin_locks */
3295 BUG_ON(in_interrupt());
3297 /* double check policy once rq lock held */
3299 reset_on_fork = p->sched_reset_on_fork;
3300 policy = oldpolicy = p->policy;
3302 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3304 if (policy != SCHED_DEADLINE &&
3305 policy != SCHED_FIFO && policy != SCHED_RR &&
3306 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
3307 policy != SCHED_IDLE)
3311 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3315 * Valid priorities for SCHED_FIFO and SCHED_RR are
3316 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3317 * SCHED_BATCH and SCHED_IDLE is 0.
3319 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3320 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3322 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3323 (rt_policy(policy) != (attr->sched_priority != 0)))
3327 * Allow unprivileged RT tasks to decrease priority:
3329 if (user && !capable(CAP_SYS_NICE)) {
3330 if (fair_policy(policy)) {
3331 if (attr->sched_nice < task_nice(p) &&
3332 !can_nice(p, attr->sched_nice))
3336 if (rt_policy(policy)) {
3337 unsigned long rlim_rtprio =
3338 task_rlimit(p, RLIMIT_RTPRIO);
3340 /* can't set/change the rt policy */
3341 if (policy != p->policy && !rlim_rtprio)
3344 /* can't increase priority */
3345 if (attr->sched_priority > p->rt_priority &&
3346 attr->sched_priority > rlim_rtprio)
3351 * Can't set/change SCHED_DEADLINE policy at all for now
3352 * (safest behavior); in the future we would like to allow
3353 * unprivileged DL tasks to increase their relative deadline
3354 * or reduce their runtime (both ways reducing utilization)
3356 if (dl_policy(policy))
3360 * Treat SCHED_IDLE as nice 20. Only allow a switch to
3361 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
3363 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
3364 if (!can_nice(p, task_nice(p)))
3368 /* can't change other user's priorities */
3369 if (!check_same_owner(p))
3372 /* Normal users shall not reset the sched_reset_on_fork flag */
3373 if (p->sched_reset_on_fork && !reset_on_fork)
3378 retval = security_task_setscheduler(p);
3384 * make sure no PI-waiters arrive (or leave) while we are
3385 * changing the priority of the task:
3387 * To be able to change p->policy safely, the appropriate
3388 * runqueue lock must be held.
3390 rq = task_rq_lock(p, &flags);
3393 * Changing the policy of the stop threads its a very bad idea
3395 if (p == rq->stop) {
3396 task_rq_unlock(rq, p, &flags);
3401 * If not changing anything there's no need to proceed further,
3402 * but store a possible modification of reset_on_fork.
3404 if (unlikely(policy == p->policy)) {
3405 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
3407 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
3409 if (dl_policy(policy))
3412 p->sched_reset_on_fork = reset_on_fork;
3413 task_rq_unlock(rq, p, &flags);
3419 #ifdef CONFIG_RT_GROUP_SCHED
3421 * Do not allow realtime tasks into groups that have no runtime
3424 if (rt_bandwidth_enabled() && rt_policy(policy) &&
3425 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
3426 !task_group_is_autogroup(task_group(p))) {
3427 task_rq_unlock(rq, p, &flags);
3432 if (dl_bandwidth_enabled() && dl_policy(policy)) {
3433 cpumask_t *span = rq->rd->span;
3436 * Don't allow tasks with an affinity mask smaller than
3437 * the entire root_domain to become SCHED_DEADLINE. We
3438 * will also fail if there's no bandwidth available.
3440 if (!cpumask_subset(span, &p->cpus_allowed) ||
3441 rq->rd->dl_bw.bw == 0) {
3442 task_rq_unlock(rq, p, &flags);
3449 /* recheck policy now with rq lock held */
3450 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
3451 policy = oldpolicy = -1;
3452 task_rq_unlock(rq, p, &flags);
3457 * If setscheduling to SCHED_DEADLINE (or changing the parameters
3458 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
3461 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
3462 task_rq_unlock(rq, p, &flags);
3466 p->sched_reset_on_fork = reset_on_fork;
3470 * Special case for priority boosted tasks.
3472 * If the new priority is lower or equal (user space view)
3473 * than the current (boosted) priority, we just store the new
3474 * normal parameters and do not touch the scheduler class and
3475 * the runqueue. This will be done when the task deboost
3478 if (rt_mutex_check_prio(p, newprio)) {
3479 __setscheduler_params(p, attr);
3480 task_rq_unlock(rq, p, &flags);
3485 running = task_current(rq, p);
3487 dequeue_task(rq, p, 0);
3489 p->sched_class->put_prev_task(rq, p);
3491 prev_class = p->sched_class;
3492 __setscheduler(rq, p, attr);
3495 p->sched_class->set_curr_task(rq);
3498 * We enqueue to tail when the priority of a task is
3499 * increased (user space view).
3501 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0);
3504 check_class_changed(rq, p, prev_class, oldprio);
3505 task_rq_unlock(rq, p, &flags);
3507 rt_mutex_adjust_pi(p);
3512 static int _sched_setscheduler(struct task_struct *p, int policy,
3513 const struct sched_param *param, bool check)
3515 struct sched_attr attr = {
3516 .sched_policy = policy,
3517 .sched_priority = param->sched_priority,
3518 .sched_nice = PRIO_TO_NICE(p->static_prio),
3522 * Fixup the legacy SCHED_RESET_ON_FORK hack
3524 if (policy & SCHED_RESET_ON_FORK) {
3525 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3526 policy &= ~SCHED_RESET_ON_FORK;
3527 attr.sched_policy = policy;
3530 return __sched_setscheduler(p, &attr, check);
3533 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
3534 * @p: the task in question.
3535 * @policy: new policy.
3536 * @param: structure containing the new RT priority.
3538 * Return: 0 on success. An error code otherwise.
3540 * NOTE that the task may be already dead.
3542 int sched_setscheduler(struct task_struct *p, int policy,
3543 const struct sched_param *param)
3545 return _sched_setscheduler(p, policy, param, true);
3547 EXPORT_SYMBOL_GPL(sched_setscheduler);
3549 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
3551 return __sched_setscheduler(p, attr, true);
3553 EXPORT_SYMBOL_GPL(sched_setattr);
3556 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
3557 * @p: the task in question.
3558 * @policy: new policy.
3559 * @param: structure containing the new RT priority.
3561 * Just like sched_setscheduler, only don't bother checking if the
3562 * current context has permission. For example, this is needed in
3563 * stop_machine(): we create temporary high priority worker threads,
3564 * but our caller might not have that capability.
3566 * Return: 0 on success. An error code otherwise.
3568 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
3569 const struct sched_param *param)
3571 return _sched_setscheduler(p, policy, param, false);
3575 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
3577 struct sched_param lparam;
3578 struct task_struct *p;
3581 if (!param || pid < 0)
3583 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
3588 p = find_process_by_pid(pid);
3590 retval = sched_setscheduler(p, policy, &lparam);
3597 * Mimics kernel/events/core.c perf_copy_attr().
3599 static int sched_copy_attr(struct sched_attr __user *uattr,
3600 struct sched_attr *attr)
3605 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
3609 * zero the full structure, so that a short copy will be nice.
3611 memset(attr, 0, sizeof(*attr));
3613 ret = get_user(size, &uattr->size);
3617 if (size > PAGE_SIZE) /* silly large */
3620 if (!size) /* abi compat */
3621 size = SCHED_ATTR_SIZE_VER0;
3623 if (size < SCHED_ATTR_SIZE_VER0)
3627 * If we're handed a bigger struct than we know of,
3628 * ensure all the unknown bits are 0 - i.e. new
3629 * user-space does not rely on any kernel feature
3630 * extensions we dont know about yet.
3632 if (size > sizeof(*attr)) {
3633 unsigned char __user *addr;
3634 unsigned char __user *end;
3637 addr = (void __user *)uattr + sizeof(*attr);
3638 end = (void __user *)uattr + size;
3640 for (; addr < end; addr++) {
3641 ret = get_user(val, addr);
3647 size = sizeof(*attr);
3650 ret = copy_from_user(attr, uattr, size);
3655 * XXX: do we want to be lenient like existing syscalls; or do we want
3656 * to be strict and return an error on out-of-bounds values?
3658 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
3663 put_user(sizeof(*attr), &uattr->size);
3668 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
3669 * @pid: the pid in question.
3670 * @policy: new policy.
3671 * @param: structure containing the new RT priority.
3673 * Return: 0 on success. An error code otherwise.
3675 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
3676 struct sched_param __user *, param)
3678 /* negative values for policy are not valid */
3682 return do_sched_setscheduler(pid, policy, param);
3686 * sys_sched_setparam - set/change the RT priority of a thread
3687 * @pid: the pid in question.
3688 * @param: structure containing the new RT priority.
3690 * Return: 0 on success. An error code otherwise.
3692 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
3694 return do_sched_setscheduler(pid, -1, param);
3698 * sys_sched_setattr - same as above, but with extended sched_attr
3699 * @pid: the pid in question.
3700 * @uattr: structure containing the extended parameters.
3701 * @flags: for future extension.
3703 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
3704 unsigned int, flags)
3706 struct sched_attr attr;
3707 struct task_struct *p;
3710 if (!uattr || pid < 0 || flags)
3713 retval = sched_copy_attr(uattr, &attr);
3717 if (attr.sched_policy < 0)
3722 p = find_process_by_pid(pid);
3724 retval = sched_setattr(p, &attr);
3731 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
3732 * @pid: the pid in question.
3734 * Return: On success, the policy of the thread. Otherwise, a negative error
3737 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
3739 struct task_struct *p;
3747 p = find_process_by_pid(pid);
3749 retval = security_task_getscheduler(p);
3752 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
3759 * sys_sched_getparam - get the RT priority of a thread
3760 * @pid: the pid in question.
3761 * @param: structure containing the RT priority.
3763 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
3766 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
3768 struct sched_param lp = { .sched_priority = 0 };
3769 struct task_struct *p;
3772 if (!param || pid < 0)
3776 p = find_process_by_pid(pid);
3781 retval = security_task_getscheduler(p);
3785 if (task_has_rt_policy(p))
3786 lp.sched_priority = p->rt_priority;
3790 * This one might sleep, we cannot do it with a spinlock held ...
3792 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
3801 static int sched_read_attr(struct sched_attr __user *uattr,
3802 struct sched_attr *attr,
3807 if (!access_ok(VERIFY_WRITE, uattr, usize))
3811 * If we're handed a smaller struct than we know of,
3812 * ensure all the unknown bits are 0 - i.e. old
3813 * user-space does not get uncomplete information.
3815 if (usize < sizeof(*attr)) {
3816 unsigned char *addr;
3819 addr = (void *)attr + usize;
3820 end = (void *)attr + sizeof(*attr);
3822 for (; addr < end; addr++) {
3830 ret = copy_to_user(uattr, attr, attr->size);
3838 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
3839 * @pid: the pid in question.
3840 * @uattr: structure containing the extended parameters.
3841 * @size: sizeof(attr) for fwd/bwd comp.
3842 * @flags: for future extension.
3844 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
3845 unsigned int, size, unsigned int, flags)
3847 struct sched_attr attr = {
3848 .size = sizeof(struct sched_attr),
3850 struct task_struct *p;
3853 if (!uattr || pid < 0 || size > PAGE_SIZE ||
3854 size < SCHED_ATTR_SIZE_VER0 || flags)
3858 p = find_process_by_pid(pid);
3863 retval = security_task_getscheduler(p);
3867 attr.sched_policy = p->policy;
3868 if (p->sched_reset_on_fork)
3869 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
3870 if (task_has_dl_policy(p))
3871 __getparam_dl(p, &attr);
3872 else if (task_has_rt_policy(p))
3873 attr.sched_priority = p->rt_priority;
3875 attr.sched_nice = task_nice(p);
3879 retval = sched_read_attr(uattr, &attr, size);
3887 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
3889 cpumask_var_t cpus_allowed, new_mask;
3890 struct task_struct *p;
3895 p = find_process_by_pid(pid);
3901 /* Prevent p going away */
3905 if (p->flags & PF_NO_SETAFFINITY) {
3909 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
3913 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
3915 goto out_free_cpus_allowed;
3918 if (!check_same_owner(p)) {
3920 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
3927 retval = security_task_setscheduler(p);
3932 cpuset_cpus_allowed(p, cpus_allowed);
3933 cpumask_and(new_mask, in_mask, cpus_allowed);
3936 * Since bandwidth control happens on root_domain basis,
3937 * if admission test is enabled, we only admit -deadline
3938 * tasks allowed to run on all the CPUs in the task's
3942 if (task_has_dl_policy(p)) {
3943 const struct cpumask *span = task_rq(p)->rd->span;
3945 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) {
3952 retval = set_cpus_allowed_ptr(p, new_mask);
3955 cpuset_cpus_allowed(p, cpus_allowed);
3956 if (!cpumask_subset(new_mask, cpus_allowed)) {
3958 * We must have raced with a concurrent cpuset
3959 * update. Just reset the cpus_allowed to the
3960 * cpuset's cpus_allowed
3962 cpumask_copy(new_mask, cpus_allowed);
3967 free_cpumask_var(new_mask);
3968 out_free_cpus_allowed:
3969 free_cpumask_var(cpus_allowed);
3975 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
3976 struct cpumask *new_mask)
3978 if (len < cpumask_size())
3979 cpumask_clear(new_mask);
3980 else if (len > cpumask_size())
3981 len = cpumask_size();
3983 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
3987 * sys_sched_setaffinity - set the cpu affinity of a process
3988 * @pid: pid of the process
3989 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
3990 * @user_mask_ptr: user-space pointer to the new cpu mask
3992 * Return: 0 on success. An error code otherwise.
3994 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
3995 unsigned long __user *, user_mask_ptr)
3997 cpumask_var_t new_mask;
4000 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4003 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4005 retval = sched_setaffinity(pid, new_mask);
4006 free_cpumask_var(new_mask);
4010 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4012 struct task_struct *p;
4013 unsigned long flags;
4019 p = find_process_by_pid(pid);
4023 retval = security_task_getscheduler(p);
4027 raw_spin_lock_irqsave(&p->pi_lock, flags);
4028 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4029 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4038 * sys_sched_getaffinity - get the cpu affinity of a process
4039 * @pid: pid of the process
4040 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4041 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4043 * Return: 0 on success. An error code otherwise.
4045 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4046 unsigned long __user *, user_mask_ptr)
4051 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4053 if (len & (sizeof(unsigned long)-1))
4056 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4059 ret = sched_getaffinity(pid, mask);
4061 size_t retlen = min_t(size_t, len, cpumask_size());
4063 if (copy_to_user(user_mask_ptr, mask, retlen))
4068 free_cpumask_var(mask);
4074 * sys_sched_yield - yield the current processor to other threads.
4076 * This function yields the current CPU to other tasks. If there are no
4077 * other threads running on this CPU then this function will return.
4081 SYSCALL_DEFINE0(sched_yield)
4083 struct rq *rq = this_rq_lock();
4085 schedstat_inc(rq, yld_count);
4086 current->sched_class->yield_task(rq);
4089 * Since we are going to call schedule() anyway, there's
4090 * no need to preempt or enable interrupts:
4092 __release(rq->lock);
4093 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4094 do_raw_spin_unlock(&rq->lock);
4095 sched_preempt_enable_no_resched();
4102 static void __cond_resched(void)
4104 __preempt_count_add(PREEMPT_ACTIVE);
4106 __preempt_count_sub(PREEMPT_ACTIVE);
4109 int __sched _cond_resched(void)
4111 if (should_resched()) {
4117 EXPORT_SYMBOL(_cond_resched);
4120 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4121 * call schedule, and on return reacquire the lock.
4123 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4124 * operations here to prevent schedule() from being called twice (once via
4125 * spin_unlock(), once by hand).
4127 int __cond_resched_lock(spinlock_t *lock)
4129 int resched = should_resched();
4132 lockdep_assert_held(lock);
4134 if (spin_needbreak(lock) || resched) {
4145 EXPORT_SYMBOL(__cond_resched_lock);
4147 int __sched __cond_resched_softirq(void)
4149 BUG_ON(!in_softirq());
4151 if (should_resched()) {
4159 EXPORT_SYMBOL(__cond_resched_softirq);
4162 * yield - yield the current processor to other threads.
4164 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4166 * The scheduler is at all times free to pick the calling task as the most
4167 * eligible task to run, if removing the yield() call from your code breaks
4168 * it, its already broken.
4170 * Typical broken usage is:
4175 * where one assumes that yield() will let 'the other' process run that will
4176 * make event true. If the current task is a SCHED_FIFO task that will never
4177 * happen. Never use yield() as a progress guarantee!!
4179 * If you want to use yield() to wait for something, use wait_event().
4180 * If you want to use yield() to be 'nice' for others, use cond_resched().
4181 * If you still want to use yield(), do not!
4183 void __sched yield(void)
4185 set_current_state(TASK_RUNNING);
4188 EXPORT_SYMBOL(yield);
4191 * yield_to - yield the current processor to another thread in
4192 * your thread group, or accelerate that thread toward the
4193 * processor it's on.
4195 * @preempt: whether task preemption is allowed or not
4197 * It's the caller's job to ensure that the target task struct
4198 * can't go away on us before we can do any checks.
4201 * true (>0) if we indeed boosted the target task.
4202 * false (0) if we failed to boost the target.
4203 * -ESRCH if there's no task to yield to.
4205 bool __sched yield_to(struct task_struct *p, bool preempt)
4207 struct task_struct *curr = current;
4208 struct rq *rq, *p_rq;
4209 unsigned long flags;
4212 local_irq_save(flags);
4218 * If we're the only runnable task on the rq and target rq also
4219 * has only one task, there's absolutely no point in yielding.
4221 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4226 double_rq_lock(rq, p_rq);
4227 if (task_rq(p) != p_rq) {
4228 double_rq_unlock(rq, p_rq);
4232 if (!curr->sched_class->yield_to_task)
4235 if (curr->sched_class != p->sched_class)
4238 if (task_running(p_rq, p) || p->state)
4241 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4243 schedstat_inc(rq, yld_count);
4245 * Make p's CPU reschedule; pick_next_entity takes care of
4248 if (preempt && rq != p_rq)
4249 resched_task(p_rq->curr);
4253 double_rq_unlock(rq, p_rq);
4255 local_irq_restore(flags);
4262 EXPORT_SYMBOL_GPL(yield_to);
4265 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4266 * that process accounting knows that this is a task in IO wait state.
4268 void __sched io_schedule(void)
4270 struct rq *rq = raw_rq();
4272 delayacct_blkio_start();
4273 atomic_inc(&rq->nr_iowait);
4274 blk_flush_plug(current);
4275 current->in_iowait = 1;
4277 current->in_iowait = 0;
4278 atomic_dec(&rq->nr_iowait);
4279 delayacct_blkio_end();
4281 EXPORT_SYMBOL(io_schedule);
4283 long __sched io_schedule_timeout(long timeout)
4285 struct rq *rq = raw_rq();
4288 delayacct_blkio_start();
4289 atomic_inc(&rq->nr_iowait);
4290 blk_flush_plug(current);
4291 current->in_iowait = 1;
4292 ret = schedule_timeout(timeout);
4293 current->in_iowait = 0;
4294 atomic_dec(&rq->nr_iowait);
4295 delayacct_blkio_end();
4300 * sys_sched_get_priority_max - return maximum RT priority.
4301 * @policy: scheduling class.
4303 * Return: On success, this syscall returns the maximum
4304 * rt_priority that can be used by a given scheduling class.
4305 * On failure, a negative error code is returned.
4307 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4314 ret = MAX_USER_RT_PRIO-1;
4316 case SCHED_DEADLINE:
4327 * sys_sched_get_priority_min - return minimum RT priority.
4328 * @policy: scheduling class.
4330 * Return: On success, this syscall returns the minimum
4331 * rt_priority that can be used by a given scheduling class.
4332 * On failure, a negative error code is returned.
4334 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4343 case SCHED_DEADLINE:
4353 * sys_sched_rr_get_interval - return the default timeslice of a process.
4354 * @pid: pid of the process.
4355 * @interval: userspace pointer to the timeslice value.
4357 * this syscall writes the default timeslice value of a given process
4358 * into the user-space timespec buffer. A value of '0' means infinity.
4360 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
4363 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
4364 struct timespec __user *, interval)
4366 struct task_struct *p;
4367 unsigned int time_slice;
4368 unsigned long flags;
4378 p = find_process_by_pid(pid);
4382 retval = security_task_getscheduler(p);
4386 rq = task_rq_lock(p, &flags);
4388 if (p->sched_class->get_rr_interval)
4389 time_slice = p->sched_class->get_rr_interval(rq, p);
4390 task_rq_unlock(rq, p, &flags);
4393 jiffies_to_timespec(time_slice, &t);
4394 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
4402 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
4404 void sched_show_task(struct task_struct *p)
4406 unsigned long free = 0;
4410 state = p->state ? __ffs(p->state) + 1 : 0;
4411 printk(KERN_INFO "%-15.15s %c", p->comm,
4412 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
4413 #if BITS_PER_LONG == 32
4414 if (state == TASK_RUNNING)
4415 printk(KERN_CONT " running ");
4417 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
4419 if (state == TASK_RUNNING)
4420 printk(KERN_CONT " running task ");
4422 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
4424 #ifdef CONFIG_DEBUG_STACK_USAGE
4425 free = stack_not_used(p);
4428 ppid = task_pid_nr(rcu_dereference(p->real_parent));
4430 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
4431 task_pid_nr(p), ppid,
4432 (unsigned long)task_thread_info(p)->flags);
4434 print_worker_info(KERN_INFO, p);
4435 show_stack(p, NULL);
4438 void show_state_filter(unsigned long state_filter)
4440 struct task_struct *g, *p;
4442 #if BITS_PER_LONG == 32
4444 " task PC stack pid father\n");
4447 " task PC stack pid father\n");
4450 do_each_thread(g, p) {
4452 * reset the NMI-timeout, listing all files on a slow
4453 * console might take a lot of time:
4455 touch_nmi_watchdog();
4456 if (!state_filter || (p->state & state_filter))
4458 } while_each_thread(g, p);
4460 touch_all_softlockup_watchdogs();
4462 #ifdef CONFIG_SCHED_DEBUG
4463 sysrq_sched_debug_show();
4467 * Only show locks if all tasks are dumped:
4470 debug_show_all_locks();
4473 void init_idle_bootup_task(struct task_struct *idle)
4475 idle->sched_class = &idle_sched_class;
4479 * init_idle - set up an idle thread for a given CPU
4480 * @idle: task in question
4481 * @cpu: cpu the idle task belongs to
4483 * NOTE: this function does not set the idle thread's NEED_RESCHED
4484 * flag, to make booting more robust.
4486 void init_idle(struct task_struct *idle, int cpu)
4488 struct rq *rq = cpu_rq(cpu);
4489 unsigned long flags;
4491 raw_spin_lock_irqsave(&rq->lock, flags);
4493 __sched_fork(0, idle);
4494 idle->state = TASK_RUNNING;
4495 idle->se.exec_start = sched_clock();
4497 do_set_cpus_allowed(idle, cpumask_of(cpu));
4499 * We're having a chicken and egg problem, even though we are
4500 * holding rq->lock, the cpu isn't yet set to this cpu so the
4501 * lockdep check in task_group() will fail.
4503 * Similar case to sched_fork(). / Alternatively we could
4504 * use task_rq_lock() here and obtain the other rq->lock.
4509 __set_task_cpu(idle, cpu);
4512 rq->curr = rq->idle = idle;
4514 #if defined(CONFIG_SMP)
4517 raw_spin_unlock_irqrestore(&rq->lock, flags);
4519 /* Set the preempt count _outside_ the spinlocks! */
4520 init_idle_preempt_count(idle, cpu);
4523 * The idle tasks have their own, simple scheduling class:
4525 idle->sched_class = &idle_sched_class;
4526 ftrace_graph_init_idle_task(idle, cpu);
4527 vtime_init_idle(idle, cpu);
4528 #if defined(CONFIG_SMP)
4529 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
4534 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
4536 if (p->sched_class && p->sched_class->set_cpus_allowed)
4537 p->sched_class->set_cpus_allowed(p, new_mask);
4539 cpumask_copy(&p->cpus_allowed, new_mask);
4540 p->nr_cpus_allowed = cpumask_weight(new_mask);
4544 * This is how migration works:
4546 * 1) we invoke migration_cpu_stop() on the target CPU using
4548 * 2) stopper starts to run (implicitly forcing the migrated thread
4550 * 3) it checks whether the migrated task is still in the wrong runqueue.
4551 * 4) if it's in the wrong runqueue then the migration thread removes
4552 * it and puts it into the right queue.
4553 * 5) stopper completes and stop_one_cpu() returns and the migration
4558 * Change a given task's CPU affinity. Migrate the thread to a
4559 * proper CPU and schedule it away if the CPU it's executing on
4560 * is removed from the allowed bitmask.
4562 * NOTE: the caller must have a valid reference to the task, the
4563 * task must not exit() & deallocate itself prematurely. The
4564 * call is not atomic; no spinlocks may be held.
4566 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
4568 unsigned long flags;
4570 unsigned int dest_cpu;
4573 rq = task_rq_lock(p, &flags);
4575 if (cpumask_equal(&p->cpus_allowed, new_mask))
4578 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
4583 do_set_cpus_allowed(p, new_mask);
4585 /* Can the task run on the task's current CPU? If so, we're done */
4586 if (cpumask_test_cpu(task_cpu(p), new_mask))
4589 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
4591 struct migration_arg arg = { p, dest_cpu };
4592 /* Need help from migration thread: drop lock and wait. */
4593 task_rq_unlock(rq, p, &flags);
4594 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
4595 tlb_migrate_finish(p->mm);
4599 task_rq_unlock(rq, p, &flags);
4603 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
4606 * Move (not current) task off this cpu, onto dest cpu. We're doing
4607 * this because either it can't run here any more (set_cpus_allowed()
4608 * away from this CPU, or CPU going down), or because we're
4609 * attempting to rebalance this task on exec (sched_exec).
4611 * So we race with normal scheduler movements, but that's OK, as long
4612 * as the task is no longer on this CPU.
4614 * Returns non-zero if task was successfully migrated.
4616 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
4618 struct rq *rq_dest, *rq_src;
4621 if (unlikely(!cpu_active(dest_cpu)))
4624 rq_src = cpu_rq(src_cpu);
4625 rq_dest = cpu_rq(dest_cpu);
4627 raw_spin_lock(&p->pi_lock);
4628 double_rq_lock(rq_src, rq_dest);
4629 /* Already moved. */
4630 if (task_cpu(p) != src_cpu)
4632 /* Affinity changed (again). */
4633 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
4637 * If we're not on a rq, the next wake-up will ensure we're
4641 dequeue_task(rq_src, p, 0);
4642 set_task_cpu(p, dest_cpu);
4643 enqueue_task(rq_dest, p, 0);
4644 check_preempt_curr(rq_dest, p, 0);
4649 double_rq_unlock(rq_src, rq_dest);
4650 raw_spin_unlock(&p->pi_lock);
4654 #ifdef CONFIG_NUMA_BALANCING
4655 /* Migrate current task p to target_cpu */
4656 int migrate_task_to(struct task_struct *p, int target_cpu)
4658 struct migration_arg arg = { p, target_cpu };
4659 int curr_cpu = task_cpu(p);
4661 if (curr_cpu == target_cpu)
4664 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
4667 /* TODO: This is not properly updating schedstats */
4669 trace_sched_move_numa(p, curr_cpu, target_cpu);
4670 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
4674 * Requeue a task on a given node and accurately track the number of NUMA
4675 * tasks on the runqueues
4677 void sched_setnuma(struct task_struct *p, int nid)
4680 unsigned long flags;
4681 bool on_rq, running;
4683 rq = task_rq_lock(p, &flags);
4685 running = task_current(rq, p);
4688 dequeue_task(rq, p, 0);
4690 p->sched_class->put_prev_task(rq, p);
4692 p->numa_preferred_nid = nid;
4695 p->sched_class->set_curr_task(rq);
4697 enqueue_task(rq, p, 0);
4698 task_rq_unlock(rq, p, &flags);
4703 * migration_cpu_stop - this will be executed by a highprio stopper thread
4704 * and performs thread migration by bumping thread off CPU then
4705 * 'pushing' onto another runqueue.
4707 static int migration_cpu_stop(void *data)
4709 struct migration_arg *arg = data;
4712 * The original target cpu might have gone down and we might
4713 * be on another cpu but it doesn't matter.
4715 local_irq_disable();
4716 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
4721 #ifdef CONFIG_HOTPLUG_CPU
4724 * Ensures that the idle task is using init_mm right before its cpu goes
4727 void idle_task_exit(void)
4729 struct mm_struct *mm = current->active_mm;
4731 BUG_ON(cpu_online(smp_processor_id()));
4733 if (mm != &init_mm) {
4734 switch_mm(mm, &init_mm, current);
4735 finish_arch_post_lock_switch();
4741 * Since this CPU is going 'away' for a while, fold any nr_active delta
4742 * we might have. Assumes we're called after migrate_tasks() so that the
4743 * nr_active count is stable.
4745 * Also see the comment "Global load-average calculations".
4747 static void calc_load_migrate(struct rq *rq)
4749 long delta = calc_load_fold_active(rq);
4751 atomic_long_add(delta, &calc_load_tasks);
4754 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
4758 static const struct sched_class fake_sched_class = {
4759 .put_prev_task = put_prev_task_fake,
4762 static struct task_struct fake_task = {
4764 * Avoid pull_{rt,dl}_task()
4766 .prio = MAX_PRIO + 1,
4767 .sched_class = &fake_sched_class,
4771 * Migrate all tasks from the rq, sleeping tasks will be migrated by
4772 * try_to_wake_up()->select_task_rq().
4774 * Called with rq->lock held even though we'er in stop_machine() and
4775 * there's no concurrency possible, we hold the required locks anyway
4776 * because of lock validation efforts.
4778 static void migrate_tasks(unsigned int dead_cpu)
4780 struct rq *rq = cpu_rq(dead_cpu);
4781 struct task_struct *next, *stop = rq->stop;
4785 * Fudge the rq selection such that the below task selection loop
4786 * doesn't get stuck on the currently eligible stop task.
4788 * We're currently inside stop_machine() and the rq is either stuck
4789 * in the stop_machine_cpu_stop() loop, or we're executing this code,
4790 * either way we should never end up calling schedule() until we're
4796 * put_prev_task() and pick_next_task() sched
4797 * class method both need to have an up-to-date
4798 * value of rq->clock[_task]
4800 update_rq_clock(rq);
4804 * There's this thread running, bail when that's the only
4807 if (rq->nr_running == 1)
4810 next = pick_next_task(rq, &fake_task);
4812 next->sched_class->put_prev_task(rq, next);
4814 /* Find suitable destination for @next, with force if needed. */
4815 dest_cpu = select_fallback_rq(dead_cpu, next);
4816 raw_spin_unlock(&rq->lock);
4818 __migrate_task(next, dead_cpu, dest_cpu);
4820 raw_spin_lock(&rq->lock);
4826 #endif /* CONFIG_HOTPLUG_CPU */
4828 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
4830 static struct ctl_table sd_ctl_dir[] = {
4832 .procname = "sched_domain",
4838 static struct ctl_table sd_ctl_root[] = {
4840 .procname = "kernel",
4842 .child = sd_ctl_dir,
4847 static struct ctl_table *sd_alloc_ctl_entry(int n)
4849 struct ctl_table *entry =
4850 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
4855 static void sd_free_ctl_entry(struct ctl_table **tablep)
4857 struct ctl_table *entry;
4860 * In the intermediate directories, both the child directory and
4861 * procname are dynamically allocated and could fail but the mode
4862 * will always be set. In the lowest directory the names are
4863 * static strings and all have proc handlers.
4865 for (entry = *tablep; entry->mode; entry++) {
4867 sd_free_ctl_entry(&entry->child);
4868 if (entry->proc_handler == NULL)
4869 kfree(entry->procname);
4876 static int min_load_idx = 0;
4877 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
4880 set_table_entry(struct ctl_table *entry,
4881 const char *procname, void *data, int maxlen,
4882 umode_t mode, proc_handler *proc_handler,
4885 entry->procname = procname;
4887 entry->maxlen = maxlen;
4889 entry->proc_handler = proc_handler;
4892 entry->extra1 = &min_load_idx;
4893 entry->extra2 = &max_load_idx;
4897 static struct ctl_table *
4898 sd_alloc_ctl_domain_table(struct sched_domain *sd)
4900 struct ctl_table *table = sd_alloc_ctl_entry(14);
4905 set_table_entry(&table[0], "min_interval", &sd->min_interval,
4906 sizeof(long), 0644, proc_doulongvec_minmax, false);
4907 set_table_entry(&table[1], "max_interval", &sd->max_interval,
4908 sizeof(long), 0644, proc_doulongvec_minmax, false);
4909 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
4910 sizeof(int), 0644, proc_dointvec_minmax, true);
4911 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
4912 sizeof(int), 0644, proc_dointvec_minmax, true);
4913 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
4914 sizeof(int), 0644, proc_dointvec_minmax, true);
4915 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
4916 sizeof(int), 0644, proc_dointvec_minmax, true);
4917 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
4918 sizeof(int), 0644, proc_dointvec_minmax, true);
4919 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
4920 sizeof(int), 0644, proc_dointvec_minmax, false);
4921 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
4922 sizeof(int), 0644, proc_dointvec_minmax, false);
4923 set_table_entry(&table[9], "cache_nice_tries",
4924 &sd->cache_nice_tries,
4925 sizeof(int), 0644, proc_dointvec_minmax, false);
4926 set_table_entry(&table[10], "flags", &sd->flags,
4927 sizeof(int), 0644, proc_dointvec_minmax, false);
4928 set_table_entry(&table[11], "max_newidle_lb_cost",
4929 &sd->max_newidle_lb_cost,
4930 sizeof(long), 0644, proc_doulongvec_minmax, false);
4931 set_table_entry(&table[12], "name", sd->name,
4932 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
4933 /* &table[13] is terminator */
4938 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
4940 struct ctl_table *entry, *table;
4941 struct sched_domain *sd;
4942 int domain_num = 0, i;
4945 for_each_domain(cpu, sd)
4947 entry = table = sd_alloc_ctl_entry(domain_num + 1);
4952 for_each_domain(cpu, sd) {
4953 snprintf(buf, 32, "domain%d", i);
4954 entry->procname = kstrdup(buf, GFP_KERNEL);
4956 entry->child = sd_alloc_ctl_domain_table(sd);
4963 static struct ctl_table_header *sd_sysctl_header;
4964 static void register_sched_domain_sysctl(void)
4966 int i, cpu_num = num_possible_cpus();
4967 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
4970 WARN_ON(sd_ctl_dir[0].child);
4971 sd_ctl_dir[0].child = entry;
4976 for_each_possible_cpu(i) {
4977 snprintf(buf, 32, "cpu%d", i);
4978 entry->procname = kstrdup(buf, GFP_KERNEL);
4980 entry->child = sd_alloc_ctl_cpu_table(i);
4984 WARN_ON(sd_sysctl_header);
4985 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
4988 /* may be called multiple times per register */
4989 static void unregister_sched_domain_sysctl(void)
4991 if (sd_sysctl_header)
4992 unregister_sysctl_table(sd_sysctl_header);
4993 sd_sysctl_header = NULL;
4994 if (sd_ctl_dir[0].child)
4995 sd_free_ctl_entry(&sd_ctl_dir[0].child);
4998 static void register_sched_domain_sysctl(void)
5001 static void unregister_sched_domain_sysctl(void)
5006 static void set_rq_online(struct rq *rq)
5009 const struct sched_class *class;
5011 cpumask_set_cpu(rq->cpu, rq->rd->online);
5014 for_each_class(class) {
5015 if (class->rq_online)
5016 class->rq_online(rq);
5021 static void set_rq_offline(struct rq *rq)
5024 const struct sched_class *class;
5026 for_each_class(class) {
5027 if (class->rq_offline)
5028 class->rq_offline(rq);
5031 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5037 * migration_call - callback that gets triggered when a CPU is added.
5038 * Here we can start up the necessary migration thread for the new CPU.
5041 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5043 int cpu = (long)hcpu;
5044 unsigned long flags;
5045 struct rq *rq = cpu_rq(cpu);
5047 switch (action & ~CPU_TASKS_FROZEN) {
5049 case CPU_UP_PREPARE:
5050 rq->calc_load_update = calc_load_update;
5054 /* Update our root-domain */
5055 raw_spin_lock_irqsave(&rq->lock, flags);
5057 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5061 raw_spin_unlock_irqrestore(&rq->lock, flags);
5064 #ifdef CONFIG_HOTPLUG_CPU
5066 sched_ttwu_pending();
5067 /* Update our root-domain */
5068 raw_spin_lock_irqsave(&rq->lock, flags);
5070 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5074 BUG_ON(rq->nr_running != 1); /* the migration thread */
5075 raw_spin_unlock_irqrestore(&rq->lock, flags);
5079 calc_load_migrate(rq);
5084 update_max_interval();
5090 * Register at high priority so that task migration (migrate_all_tasks)
5091 * happens before everything else. This has to be lower priority than
5092 * the notifier in the perf_event subsystem, though.
5094 static struct notifier_block migration_notifier = {
5095 .notifier_call = migration_call,
5096 .priority = CPU_PRI_MIGRATION,
5099 static int sched_cpu_active(struct notifier_block *nfb,
5100 unsigned long action, void *hcpu)
5102 switch (action & ~CPU_TASKS_FROZEN) {
5103 case CPU_DOWN_FAILED:
5104 set_cpu_active((long)hcpu, true);
5111 static int sched_cpu_inactive(struct notifier_block *nfb,
5112 unsigned long action, void *hcpu)
5114 unsigned long flags;
5115 long cpu = (long)hcpu;
5117 switch (action & ~CPU_TASKS_FROZEN) {
5118 case CPU_DOWN_PREPARE:
5119 set_cpu_active(cpu, false);
5121 /* explicitly allow suspend */
5122 if (!(action & CPU_TASKS_FROZEN)) {
5123 struct dl_bw *dl_b = dl_bw_of(cpu);
5127 raw_spin_lock_irqsave(&dl_b->lock, flags);
5128 cpus = dl_bw_cpus(cpu);
5129 overflow = __dl_overflow(dl_b, cpus, 0, 0);
5130 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5133 return notifier_from_errno(-EBUSY);
5141 static int __init migration_init(void)
5143 void *cpu = (void *)(long)smp_processor_id();
5146 /* Initialize migration for the boot CPU */
5147 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5148 BUG_ON(err == NOTIFY_BAD);
5149 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5150 register_cpu_notifier(&migration_notifier);
5152 /* Register cpu active notifiers */
5153 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5154 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5158 early_initcall(migration_init);
5163 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5165 #ifdef CONFIG_SCHED_DEBUG
5167 static __read_mostly int sched_debug_enabled;
5169 static int __init sched_debug_setup(char *str)
5171 sched_debug_enabled = 1;
5175 early_param("sched_debug", sched_debug_setup);
5177 static inline bool sched_debug(void)
5179 return sched_debug_enabled;
5182 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5183 struct cpumask *groupmask)
5185 struct sched_group *group = sd->groups;
5188 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5189 cpumask_clear(groupmask);
5191 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5193 if (!(sd->flags & SD_LOAD_BALANCE)) {
5194 printk("does not load-balance\n");
5196 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5201 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5203 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5204 printk(KERN_ERR "ERROR: domain->span does not contain "
5207 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5208 printk(KERN_ERR "ERROR: domain->groups does not contain"
5212 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5216 printk(KERN_ERR "ERROR: group is NULL\n");
5221 * Even though we initialize ->power to something semi-sane,
5222 * we leave power_orig unset. This allows us to detect if
5223 * domain iteration is still funny without causing /0 traps.
5225 if (!group->sgp->power_orig) {
5226 printk(KERN_CONT "\n");
5227 printk(KERN_ERR "ERROR: domain->cpu_power not "
5232 if (!cpumask_weight(sched_group_cpus(group))) {
5233 printk(KERN_CONT "\n");
5234 printk(KERN_ERR "ERROR: empty group\n");
5238 if (!(sd->flags & SD_OVERLAP) &&
5239 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5240 printk(KERN_CONT "\n");
5241 printk(KERN_ERR "ERROR: repeated CPUs\n");
5245 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5247 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
5249 printk(KERN_CONT " %s", str);
5250 if (group->sgp->power != SCHED_POWER_SCALE) {
5251 printk(KERN_CONT " (cpu_power = %d)",
5255 group = group->next;
5256 } while (group != sd->groups);
5257 printk(KERN_CONT "\n");
5259 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5260 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5263 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5264 printk(KERN_ERR "ERROR: parent span is not a superset "
5265 "of domain->span\n");
5269 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5273 if (!sched_debug_enabled)
5277 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5281 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5284 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5292 #else /* !CONFIG_SCHED_DEBUG */
5293 # define sched_domain_debug(sd, cpu) do { } while (0)
5294 static inline bool sched_debug(void)
5298 #endif /* CONFIG_SCHED_DEBUG */
5300 static int sd_degenerate(struct sched_domain *sd)
5302 if (cpumask_weight(sched_domain_span(sd)) == 1)
5305 /* Following flags need at least 2 groups */
5306 if (sd->flags & (SD_LOAD_BALANCE |
5307 SD_BALANCE_NEWIDLE |
5311 SD_SHARE_PKG_RESOURCES |
5312 SD_SHARE_POWERDOMAIN)) {
5313 if (sd->groups != sd->groups->next)
5317 /* Following flags don't use groups */
5318 if (sd->flags & (SD_WAKE_AFFINE))
5325 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
5327 unsigned long cflags = sd->flags, pflags = parent->flags;
5329 if (sd_degenerate(parent))
5332 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
5335 /* Flags needing groups don't count if only 1 group in parent */
5336 if (parent->groups == parent->groups->next) {
5337 pflags &= ~(SD_LOAD_BALANCE |
5338 SD_BALANCE_NEWIDLE |
5342 SD_SHARE_PKG_RESOURCES |
5344 SD_SHARE_POWERDOMAIN);
5345 if (nr_node_ids == 1)
5346 pflags &= ~SD_SERIALIZE;
5348 if (~cflags & pflags)
5354 static void free_rootdomain(struct rcu_head *rcu)
5356 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
5358 cpupri_cleanup(&rd->cpupri);
5359 cpudl_cleanup(&rd->cpudl);
5360 free_cpumask_var(rd->dlo_mask);
5361 free_cpumask_var(rd->rto_mask);
5362 free_cpumask_var(rd->online);
5363 free_cpumask_var(rd->span);
5367 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
5369 struct root_domain *old_rd = NULL;
5370 unsigned long flags;
5372 raw_spin_lock_irqsave(&rq->lock, flags);
5377 if (cpumask_test_cpu(rq->cpu, old_rd->online))
5380 cpumask_clear_cpu(rq->cpu, old_rd->span);
5383 * If we dont want to free the old_rd yet then
5384 * set old_rd to NULL to skip the freeing later
5387 if (!atomic_dec_and_test(&old_rd->refcount))
5391 atomic_inc(&rd->refcount);
5394 cpumask_set_cpu(rq->cpu, rd->span);
5395 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
5398 raw_spin_unlock_irqrestore(&rq->lock, flags);
5401 call_rcu_sched(&old_rd->rcu, free_rootdomain);
5404 static int init_rootdomain(struct root_domain *rd)
5406 memset(rd, 0, sizeof(*rd));
5408 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
5410 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
5412 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
5414 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
5417 init_dl_bw(&rd->dl_bw);
5418 if (cpudl_init(&rd->cpudl) != 0)
5421 if (cpupri_init(&rd->cpupri) != 0)
5426 free_cpumask_var(rd->rto_mask);
5428 free_cpumask_var(rd->dlo_mask);
5430 free_cpumask_var(rd->online);
5432 free_cpumask_var(rd->span);
5438 * By default the system creates a single root-domain with all cpus as
5439 * members (mimicking the global state we have today).
5441 struct root_domain def_root_domain;
5443 static void init_defrootdomain(void)
5445 init_rootdomain(&def_root_domain);
5447 atomic_set(&def_root_domain.refcount, 1);
5450 static struct root_domain *alloc_rootdomain(void)
5452 struct root_domain *rd;
5454 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
5458 if (init_rootdomain(rd) != 0) {
5466 static void free_sched_groups(struct sched_group *sg, int free_sgp)
5468 struct sched_group *tmp, *first;
5477 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
5482 } while (sg != first);
5485 static void free_sched_domain(struct rcu_head *rcu)
5487 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
5490 * If its an overlapping domain it has private groups, iterate and
5493 if (sd->flags & SD_OVERLAP) {
5494 free_sched_groups(sd->groups, 1);
5495 } else if (atomic_dec_and_test(&sd->groups->ref)) {
5496 kfree(sd->groups->sgp);
5502 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
5504 call_rcu(&sd->rcu, free_sched_domain);
5507 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
5509 for (; sd; sd = sd->parent)
5510 destroy_sched_domain(sd, cpu);
5514 * Keep a special pointer to the highest sched_domain that has
5515 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
5516 * allows us to avoid some pointer chasing select_idle_sibling().
5518 * Also keep a unique ID per domain (we use the first cpu number in
5519 * the cpumask of the domain), this allows us to quickly tell if
5520 * two cpus are in the same cache domain, see cpus_share_cache().
5522 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
5523 DEFINE_PER_CPU(int, sd_llc_size);
5524 DEFINE_PER_CPU(int, sd_llc_id);
5525 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
5526 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
5527 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
5529 static void update_top_cache_domain(int cpu)
5531 struct sched_domain *sd;
5532 struct sched_domain *busy_sd = NULL;
5536 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
5538 id = cpumask_first(sched_domain_span(sd));
5539 size = cpumask_weight(sched_domain_span(sd));
5540 busy_sd = sd->parent; /* sd_busy */
5542 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
5544 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
5545 per_cpu(sd_llc_size, cpu) = size;
5546 per_cpu(sd_llc_id, cpu) = id;
5548 sd = lowest_flag_domain(cpu, SD_NUMA);
5549 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
5551 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
5552 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
5556 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5557 * hold the hotplug lock.
5560 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
5562 struct rq *rq = cpu_rq(cpu);
5563 struct sched_domain *tmp;
5565 /* Remove the sched domains which do not contribute to scheduling. */
5566 for (tmp = sd; tmp; ) {
5567 struct sched_domain *parent = tmp->parent;
5571 if (sd_parent_degenerate(tmp, parent)) {
5572 tmp->parent = parent->parent;
5574 parent->parent->child = tmp;
5576 * Transfer SD_PREFER_SIBLING down in case of a
5577 * degenerate parent; the spans match for this
5578 * so the property transfers.
5580 if (parent->flags & SD_PREFER_SIBLING)
5581 tmp->flags |= SD_PREFER_SIBLING;
5582 destroy_sched_domain(parent, cpu);
5587 if (sd && sd_degenerate(sd)) {
5590 destroy_sched_domain(tmp, cpu);
5595 sched_domain_debug(sd, cpu);
5597 rq_attach_root(rq, rd);
5599 rcu_assign_pointer(rq->sd, sd);
5600 destroy_sched_domains(tmp, cpu);
5602 update_top_cache_domain(cpu);
5605 /* cpus with isolated domains */
5606 static cpumask_var_t cpu_isolated_map;
5608 /* Setup the mask of cpus configured for isolated domains */
5609 static int __init isolated_cpu_setup(char *str)
5611 alloc_bootmem_cpumask_var(&cpu_isolated_map);
5612 cpulist_parse(str, cpu_isolated_map);
5616 __setup("isolcpus=", isolated_cpu_setup);
5619 struct sched_domain ** __percpu sd;
5620 struct root_domain *rd;
5631 * Build an iteration mask that can exclude certain CPUs from the upwards
5634 * Asymmetric node setups can result in situations where the domain tree is of
5635 * unequal depth, make sure to skip domains that already cover the entire
5638 * In that case build_sched_domains() will have terminated the iteration early
5639 * and our sibling sd spans will be empty. Domains should always include the
5640 * cpu they're built on, so check that.
5643 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
5645 const struct cpumask *span = sched_domain_span(sd);
5646 struct sd_data *sdd = sd->private;
5647 struct sched_domain *sibling;
5650 for_each_cpu(i, span) {
5651 sibling = *per_cpu_ptr(sdd->sd, i);
5652 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
5655 cpumask_set_cpu(i, sched_group_mask(sg));
5660 * Return the canonical balance cpu for this group, this is the first cpu
5661 * of this group that's also in the iteration mask.
5663 int group_balance_cpu(struct sched_group *sg)
5665 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
5669 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
5671 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
5672 const struct cpumask *span = sched_domain_span(sd);
5673 struct cpumask *covered = sched_domains_tmpmask;
5674 struct sd_data *sdd = sd->private;
5675 struct sched_domain *child;
5678 cpumask_clear(covered);
5680 for_each_cpu(i, span) {
5681 struct cpumask *sg_span;
5683 if (cpumask_test_cpu(i, covered))
5686 child = *per_cpu_ptr(sdd->sd, i);
5688 /* See the comment near build_group_mask(). */
5689 if (!cpumask_test_cpu(i, sched_domain_span(child)))
5692 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
5693 GFP_KERNEL, cpu_to_node(cpu));
5698 sg_span = sched_group_cpus(sg);
5700 child = child->child;
5701 cpumask_copy(sg_span, sched_domain_span(child));
5703 cpumask_set_cpu(i, sg_span);
5705 cpumask_or(covered, covered, sg_span);
5707 sg->sgp = *per_cpu_ptr(sdd->sgp, i);
5708 if (atomic_inc_return(&sg->sgp->ref) == 1)
5709 build_group_mask(sd, sg);
5712 * Initialize sgp->power such that even if we mess up the
5713 * domains and no possible iteration will get us here, we won't
5716 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span);
5717 sg->sgp->power_orig = sg->sgp->power;
5720 * Make sure the first group of this domain contains the
5721 * canonical balance cpu. Otherwise the sched_domain iteration
5722 * breaks. See update_sg_lb_stats().
5724 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
5725 group_balance_cpu(sg) == cpu)
5735 sd->groups = groups;
5740 free_sched_groups(first, 0);
5745 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
5747 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
5748 struct sched_domain *child = sd->child;
5751 cpu = cpumask_first(sched_domain_span(child));
5754 *sg = *per_cpu_ptr(sdd->sg, cpu);
5755 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
5756 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
5763 * build_sched_groups will build a circular linked list of the groups
5764 * covered by the given span, and will set each group's ->cpumask correctly,
5765 * and ->cpu_power to 0.
5767 * Assumes the sched_domain tree is fully constructed
5770 build_sched_groups(struct sched_domain *sd, int cpu)
5772 struct sched_group *first = NULL, *last = NULL;
5773 struct sd_data *sdd = sd->private;
5774 const struct cpumask *span = sched_domain_span(sd);
5775 struct cpumask *covered;
5778 get_group(cpu, sdd, &sd->groups);
5779 atomic_inc(&sd->groups->ref);
5781 if (cpu != cpumask_first(span))
5784 lockdep_assert_held(&sched_domains_mutex);
5785 covered = sched_domains_tmpmask;
5787 cpumask_clear(covered);
5789 for_each_cpu(i, span) {
5790 struct sched_group *sg;
5793 if (cpumask_test_cpu(i, covered))
5796 group = get_group(i, sdd, &sg);
5797 cpumask_setall(sched_group_mask(sg));
5799 for_each_cpu(j, span) {
5800 if (get_group(j, sdd, NULL) != group)
5803 cpumask_set_cpu(j, covered);
5804 cpumask_set_cpu(j, sched_group_cpus(sg));
5819 * Initialize sched groups cpu_power.
5821 * cpu_power indicates the capacity of sched group, which is used while
5822 * distributing the load between different sched groups in a sched domain.
5823 * Typically cpu_power for all the groups in a sched domain will be same unless
5824 * there are asymmetries in the topology. If there are asymmetries, group
5825 * having more cpu_power will pickup more load compared to the group having
5828 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
5830 struct sched_group *sg = sd->groups;
5835 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
5837 } while (sg != sd->groups);
5839 if (cpu != group_balance_cpu(sg))
5842 update_group_power(sd, cpu);
5843 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight);
5847 * Initializers for schedule domains
5848 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
5851 static int default_relax_domain_level = -1;
5852 int sched_domain_level_max;
5854 static int __init setup_relax_domain_level(char *str)
5856 if (kstrtoint(str, 0, &default_relax_domain_level))
5857 pr_warn("Unable to set relax_domain_level\n");
5861 __setup("relax_domain_level=", setup_relax_domain_level);
5863 static void set_domain_attribute(struct sched_domain *sd,
5864 struct sched_domain_attr *attr)
5868 if (!attr || attr->relax_domain_level < 0) {
5869 if (default_relax_domain_level < 0)
5872 request = default_relax_domain_level;
5874 request = attr->relax_domain_level;
5875 if (request < sd->level) {
5876 /* turn off idle balance on this domain */
5877 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5879 /* turn on idle balance on this domain */
5880 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
5884 static void __sdt_free(const struct cpumask *cpu_map);
5885 static int __sdt_alloc(const struct cpumask *cpu_map);
5887 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
5888 const struct cpumask *cpu_map)
5892 if (!atomic_read(&d->rd->refcount))
5893 free_rootdomain(&d->rd->rcu); /* fall through */
5895 free_percpu(d->sd); /* fall through */
5897 __sdt_free(cpu_map); /* fall through */
5903 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
5904 const struct cpumask *cpu_map)
5906 memset(d, 0, sizeof(*d));
5908 if (__sdt_alloc(cpu_map))
5909 return sa_sd_storage;
5910 d->sd = alloc_percpu(struct sched_domain *);
5912 return sa_sd_storage;
5913 d->rd = alloc_rootdomain();
5916 return sa_rootdomain;
5920 * NULL the sd_data elements we've used to build the sched_domain and
5921 * sched_group structure so that the subsequent __free_domain_allocs()
5922 * will not free the data we're using.
5924 static void claim_allocations(int cpu, struct sched_domain *sd)
5926 struct sd_data *sdd = sd->private;
5928 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
5929 *per_cpu_ptr(sdd->sd, cpu) = NULL;
5931 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
5932 *per_cpu_ptr(sdd->sg, cpu) = NULL;
5934 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
5935 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
5939 static int sched_domains_numa_levels;
5940 static int *sched_domains_numa_distance;
5941 static struct cpumask ***sched_domains_numa_masks;
5942 static int sched_domains_curr_level;
5946 * SD_flags allowed in topology descriptions.
5948 * SD_SHARE_CPUPOWER - describes SMT topologies
5949 * SD_SHARE_PKG_RESOURCES - describes shared caches
5950 * SD_NUMA - describes NUMA topologies
5951 * SD_SHARE_POWERDOMAIN - describes shared power domain
5954 * SD_ASYM_PACKING - describes SMT quirks
5956 #define TOPOLOGY_SD_FLAGS \
5957 (SD_SHARE_CPUPOWER | \
5958 SD_SHARE_PKG_RESOURCES | \
5961 SD_SHARE_POWERDOMAIN)
5963 static struct sched_domain *
5964 sd_init(struct sched_domain_topology_level *tl, int cpu)
5966 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
5967 int sd_weight, sd_flags = 0;
5971 * Ugly hack to pass state to sd_numa_mask()...
5973 sched_domains_curr_level = tl->numa_level;
5976 sd_weight = cpumask_weight(tl->mask(cpu));
5979 sd_flags = (*tl->sd_flags)();
5980 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
5981 "wrong sd_flags in topology description\n"))
5982 sd_flags &= ~TOPOLOGY_SD_FLAGS;
5984 *sd = (struct sched_domain){
5985 .min_interval = sd_weight,
5986 .max_interval = 2*sd_weight,
5988 .imbalance_pct = 125,
5990 .cache_nice_tries = 0,
5997 .flags = 1*SD_LOAD_BALANCE
5998 | 1*SD_BALANCE_NEWIDLE
6003 | 0*SD_SHARE_CPUPOWER
6004 | 0*SD_SHARE_PKG_RESOURCES
6006 | 0*SD_PREFER_SIBLING
6011 .last_balance = jiffies,
6012 .balance_interval = sd_weight,
6014 .max_newidle_lb_cost = 0,
6015 .next_decay_max_lb_cost = jiffies,
6016 #ifdef CONFIG_SCHED_DEBUG
6022 * Convert topological properties into behaviour.
6025 if (sd->flags & SD_SHARE_CPUPOWER) {
6026 sd->imbalance_pct = 110;
6027 sd->smt_gain = 1178; /* ~15% */
6029 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6030 sd->imbalance_pct = 117;
6031 sd->cache_nice_tries = 1;
6035 } else if (sd->flags & SD_NUMA) {
6036 sd->cache_nice_tries = 2;
6040 sd->flags |= SD_SERIALIZE;
6041 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6042 sd->flags &= ~(SD_BALANCE_EXEC |
6049 sd->flags |= SD_PREFER_SIBLING;
6050 sd->cache_nice_tries = 1;
6055 sd->private = &tl->data;
6061 * Topology list, bottom-up.
6063 static struct sched_domain_topology_level default_topology[] = {
6064 #ifdef CONFIG_SCHED_SMT
6065 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6067 #ifdef CONFIG_SCHED_MC
6068 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6070 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6074 struct sched_domain_topology_level *sched_domain_topology = default_topology;
6076 #define for_each_sd_topology(tl) \
6077 for (tl = sched_domain_topology; tl->mask; tl++)
6079 void set_sched_topology(struct sched_domain_topology_level *tl)
6081 sched_domain_topology = tl;
6086 static const struct cpumask *sd_numa_mask(int cpu)
6088 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6091 static void sched_numa_warn(const char *str)
6093 static int done = false;
6101 printk(KERN_WARNING "ERROR: %s\n\n", str);
6103 for (i = 0; i < nr_node_ids; i++) {
6104 printk(KERN_WARNING " ");
6105 for (j = 0; j < nr_node_ids; j++)
6106 printk(KERN_CONT "%02d ", node_distance(i,j));
6107 printk(KERN_CONT "\n");
6109 printk(KERN_WARNING "\n");
6112 static bool find_numa_distance(int distance)
6116 if (distance == node_distance(0, 0))
6119 for (i = 0; i < sched_domains_numa_levels; i++) {
6120 if (sched_domains_numa_distance[i] == distance)
6127 static void sched_init_numa(void)
6129 int next_distance, curr_distance = node_distance(0, 0);
6130 struct sched_domain_topology_level *tl;
6134 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6135 if (!sched_domains_numa_distance)
6139 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6140 * unique distances in the node_distance() table.
6142 * Assumes node_distance(0,j) includes all distances in
6143 * node_distance(i,j) in order to avoid cubic time.
6145 next_distance = curr_distance;
6146 for (i = 0; i < nr_node_ids; i++) {
6147 for (j = 0; j < nr_node_ids; j++) {
6148 for (k = 0; k < nr_node_ids; k++) {
6149 int distance = node_distance(i, k);
6151 if (distance > curr_distance &&
6152 (distance < next_distance ||
6153 next_distance == curr_distance))
6154 next_distance = distance;
6157 * While not a strong assumption it would be nice to know
6158 * about cases where if node A is connected to B, B is not
6159 * equally connected to A.
6161 if (sched_debug() && node_distance(k, i) != distance)
6162 sched_numa_warn("Node-distance not symmetric");
6164 if (sched_debug() && i && !find_numa_distance(distance))
6165 sched_numa_warn("Node-0 not representative");
6167 if (next_distance != curr_distance) {
6168 sched_domains_numa_distance[level++] = next_distance;
6169 sched_domains_numa_levels = level;
6170 curr_distance = next_distance;
6175 * In case of sched_debug() we verify the above assumption.
6181 * 'level' contains the number of unique distances, excluding the
6182 * identity distance node_distance(i,i).
6184 * The sched_domains_numa_distance[] array includes the actual distance
6189 * Here, we should temporarily reset sched_domains_numa_levels to 0.
6190 * If it fails to allocate memory for array sched_domains_numa_masks[][],
6191 * the array will contain less then 'level' members. This could be
6192 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
6193 * in other functions.
6195 * We reset it to 'level' at the end of this function.
6197 sched_domains_numa_levels = 0;
6199 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
6200 if (!sched_domains_numa_masks)
6204 * Now for each level, construct a mask per node which contains all
6205 * cpus of nodes that are that many hops away from us.
6207 for (i = 0; i < level; i++) {
6208 sched_domains_numa_masks[i] =
6209 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
6210 if (!sched_domains_numa_masks[i])
6213 for (j = 0; j < nr_node_ids; j++) {
6214 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
6218 sched_domains_numa_masks[i][j] = mask;
6220 for (k = 0; k < nr_node_ids; k++) {
6221 if (node_distance(j, k) > sched_domains_numa_distance[i])
6224 cpumask_or(mask, mask, cpumask_of_node(k));
6229 /* Compute default topology size */
6230 for (i = 0; sched_domain_topology[i].mask; i++);
6232 tl = kzalloc((i + level + 1) *
6233 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
6238 * Copy the default topology bits..
6240 for (i = 0; sched_domain_topology[i].mask; i++)
6241 tl[i] = sched_domain_topology[i];
6244 * .. and append 'j' levels of NUMA goodness.
6246 for (j = 0; j < level; i++, j++) {
6247 tl[i] = (struct sched_domain_topology_level){
6248 .mask = sd_numa_mask,
6249 .sd_flags = cpu_numa_flags,
6250 .flags = SDTL_OVERLAP,
6256 sched_domain_topology = tl;
6258 sched_domains_numa_levels = level;
6261 static void sched_domains_numa_masks_set(int cpu)
6264 int node = cpu_to_node(cpu);
6266 for (i = 0; i < sched_domains_numa_levels; i++) {
6267 for (j = 0; j < nr_node_ids; j++) {
6268 if (node_distance(j, node) <= sched_domains_numa_distance[i])
6269 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
6274 static void sched_domains_numa_masks_clear(int cpu)
6277 for (i = 0; i < sched_domains_numa_levels; i++) {
6278 for (j = 0; j < nr_node_ids; j++)
6279 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
6284 * Update sched_domains_numa_masks[level][node] array when new cpus
6287 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6288 unsigned long action,
6291 int cpu = (long)hcpu;
6293 switch (action & ~CPU_TASKS_FROZEN) {
6295 sched_domains_numa_masks_set(cpu);
6299 sched_domains_numa_masks_clear(cpu);
6309 static inline void sched_init_numa(void)
6313 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
6314 unsigned long action,
6319 #endif /* CONFIG_NUMA */
6321 static int __sdt_alloc(const struct cpumask *cpu_map)
6323 struct sched_domain_topology_level *tl;
6326 for_each_sd_topology(tl) {
6327 struct sd_data *sdd = &tl->data;
6329 sdd->sd = alloc_percpu(struct sched_domain *);
6333 sdd->sg = alloc_percpu(struct sched_group *);
6337 sdd->sgp = alloc_percpu(struct sched_group_power *);
6341 for_each_cpu(j, cpu_map) {
6342 struct sched_domain *sd;
6343 struct sched_group *sg;
6344 struct sched_group_power *sgp;
6346 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
6347 GFP_KERNEL, cpu_to_node(j));
6351 *per_cpu_ptr(sdd->sd, j) = sd;
6353 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6354 GFP_KERNEL, cpu_to_node(j));
6360 *per_cpu_ptr(sdd->sg, j) = sg;
6362 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(),
6363 GFP_KERNEL, cpu_to_node(j));
6367 *per_cpu_ptr(sdd->sgp, j) = sgp;
6374 static void __sdt_free(const struct cpumask *cpu_map)
6376 struct sched_domain_topology_level *tl;
6379 for_each_sd_topology(tl) {
6380 struct sd_data *sdd = &tl->data;
6382 for_each_cpu(j, cpu_map) {
6383 struct sched_domain *sd;
6386 sd = *per_cpu_ptr(sdd->sd, j);
6387 if (sd && (sd->flags & SD_OVERLAP))
6388 free_sched_groups(sd->groups, 0);
6389 kfree(*per_cpu_ptr(sdd->sd, j));
6393 kfree(*per_cpu_ptr(sdd->sg, j));
6395 kfree(*per_cpu_ptr(sdd->sgp, j));
6397 free_percpu(sdd->sd);
6399 free_percpu(sdd->sg);
6401 free_percpu(sdd->sgp);
6406 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
6407 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6408 struct sched_domain *child, int cpu)
6410 struct sched_domain *sd = sd_init(tl, cpu);
6414 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
6416 sd->level = child->level + 1;
6417 sched_domain_level_max = max(sched_domain_level_max, sd->level);
6421 set_domain_attribute(sd, attr);
6427 * Build sched domains for a given set of cpus and attach the sched domains
6428 * to the individual cpus
6430 static int build_sched_domains(const struct cpumask *cpu_map,
6431 struct sched_domain_attr *attr)
6433 enum s_alloc alloc_state;
6434 struct sched_domain *sd;
6436 int i, ret = -ENOMEM;
6438 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
6439 if (alloc_state != sa_rootdomain)
6442 /* Set up domains for cpus specified by the cpu_map. */
6443 for_each_cpu(i, cpu_map) {
6444 struct sched_domain_topology_level *tl;
6447 for_each_sd_topology(tl) {
6448 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
6449 if (tl == sched_domain_topology)
6450 *per_cpu_ptr(d.sd, i) = sd;
6451 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
6452 sd->flags |= SD_OVERLAP;
6453 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
6458 /* Build the groups for the domains */
6459 for_each_cpu(i, cpu_map) {
6460 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6461 sd->span_weight = cpumask_weight(sched_domain_span(sd));
6462 if (sd->flags & SD_OVERLAP) {
6463 if (build_overlap_sched_groups(sd, i))
6466 if (build_sched_groups(sd, i))
6472 /* Calculate CPU power for physical packages and nodes */
6473 for (i = nr_cpumask_bits-1; i >= 0; i--) {
6474 if (!cpumask_test_cpu(i, cpu_map))
6477 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
6478 claim_allocations(i, sd);
6479 init_sched_groups_power(i, sd);
6483 /* Attach the domains */
6485 for_each_cpu(i, cpu_map) {
6486 sd = *per_cpu_ptr(d.sd, i);
6487 cpu_attach_domain(sd, d.rd, i);
6493 __free_domain_allocs(&d, alloc_state, cpu_map);
6497 static cpumask_var_t *doms_cur; /* current sched domains */
6498 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
6499 static struct sched_domain_attr *dattr_cur;
6500 /* attribues of custom domains in 'doms_cur' */
6503 * Special case: If a kmalloc of a doms_cur partition (array of
6504 * cpumask) fails, then fallback to a single sched domain,
6505 * as determined by the single cpumask fallback_doms.
6507 static cpumask_var_t fallback_doms;
6510 * arch_update_cpu_topology lets virtualized architectures update the
6511 * cpu core maps. It is supposed to return 1 if the topology changed
6512 * or 0 if it stayed the same.
6514 int __weak arch_update_cpu_topology(void)
6519 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
6522 cpumask_var_t *doms;
6524 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
6527 for (i = 0; i < ndoms; i++) {
6528 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
6529 free_sched_domains(doms, i);
6536 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
6539 for (i = 0; i < ndoms; i++)
6540 free_cpumask_var(doms[i]);
6545 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6546 * For now this just excludes isolated cpus, but could be used to
6547 * exclude other special cases in the future.
6549 static int init_sched_domains(const struct cpumask *cpu_map)
6553 arch_update_cpu_topology();
6555 doms_cur = alloc_sched_domains(ndoms_cur);
6557 doms_cur = &fallback_doms;
6558 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
6559 err = build_sched_domains(doms_cur[0], NULL);
6560 register_sched_domain_sysctl();
6566 * Detach sched domains from a group of cpus specified in cpu_map
6567 * These cpus will now be attached to the NULL domain
6569 static void detach_destroy_domains(const struct cpumask *cpu_map)
6574 for_each_cpu(i, cpu_map)
6575 cpu_attach_domain(NULL, &def_root_domain, i);
6579 /* handle null as "default" */
6580 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
6581 struct sched_domain_attr *new, int idx_new)
6583 struct sched_domain_attr tmp;
6590 return !memcmp(cur ? (cur + idx_cur) : &tmp,
6591 new ? (new + idx_new) : &tmp,
6592 sizeof(struct sched_domain_attr));
6596 * Partition sched domains as specified by the 'ndoms_new'
6597 * cpumasks in the array doms_new[] of cpumasks. This compares
6598 * doms_new[] to the current sched domain partitioning, doms_cur[].
6599 * It destroys each deleted domain and builds each new domain.
6601 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
6602 * The masks don't intersect (don't overlap.) We should setup one
6603 * sched domain for each mask. CPUs not in any of the cpumasks will
6604 * not be load balanced. If the same cpumask appears both in the
6605 * current 'doms_cur' domains and in the new 'doms_new', we can leave
6608 * The passed in 'doms_new' should be allocated using
6609 * alloc_sched_domains. This routine takes ownership of it and will
6610 * free_sched_domains it when done with it. If the caller failed the
6611 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
6612 * and partition_sched_domains() will fallback to the single partition
6613 * 'fallback_doms', it also forces the domains to be rebuilt.
6615 * If doms_new == NULL it will be replaced with cpu_online_mask.
6616 * ndoms_new == 0 is a special case for destroying existing domains,
6617 * and it will not create the default domain.
6619 * Call with hotplug lock held
6621 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
6622 struct sched_domain_attr *dattr_new)
6627 mutex_lock(&sched_domains_mutex);
6629 /* always unregister in case we don't destroy any domains */
6630 unregister_sched_domain_sysctl();
6632 /* Let architecture update cpu core mappings. */
6633 new_topology = arch_update_cpu_topology();
6635 n = doms_new ? ndoms_new : 0;
6637 /* Destroy deleted domains */
6638 for (i = 0; i < ndoms_cur; i++) {
6639 for (j = 0; j < n && !new_topology; j++) {
6640 if (cpumask_equal(doms_cur[i], doms_new[j])
6641 && dattrs_equal(dattr_cur, i, dattr_new, j))
6644 /* no match - a current sched domain not in new doms_new[] */
6645 detach_destroy_domains(doms_cur[i]);
6651 if (doms_new == NULL) {
6653 doms_new = &fallback_doms;
6654 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
6655 WARN_ON_ONCE(dattr_new);
6658 /* Build new domains */
6659 for (i = 0; i < ndoms_new; i++) {
6660 for (j = 0; j < n && !new_topology; j++) {
6661 if (cpumask_equal(doms_new[i], doms_cur[j])
6662 && dattrs_equal(dattr_new, i, dattr_cur, j))
6665 /* no match - add a new doms_new */
6666 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
6671 /* Remember the new sched domains */
6672 if (doms_cur != &fallback_doms)
6673 free_sched_domains(doms_cur, ndoms_cur);
6674 kfree(dattr_cur); /* kfree(NULL) is safe */
6675 doms_cur = doms_new;
6676 dattr_cur = dattr_new;
6677 ndoms_cur = ndoms_new;
6679 register_sched_domain_sysctl();
6681 mutex_unlock(&sched_domains_mutex);
6684 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
6687 * Update cpusets according to cpu_active mask. If cpusets are
6688 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
6689 * around partition_sched_domains().
6691 * If we come here as part of a suspend/resume, don't touch cpusets because we
6692 * want to restore it back to its original state upon resume anyway.
6694 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
6698 case CPU_ONLINE_FROZEN:
6699 case CPU_DOWN_FAILED_FROZEN:
6702 * num_cpus_frozen tracks how many CPUs are involved in suspend
6703 * resume sequence. As long as this is not the last online
6704 * operation in the resume sequence, just build a single sched
6705 * domain, ignoring cpusets.
6708 if (likely(num_cpus_frozen)) {
6709 partition_sched_domains(1, NULL, NULL);
6714 * This is the last CPU online operation. So fall through and
6715 * restore the original sched domains by considering the
6716 * cpuset configurations.
6720 case CPU_DOWN_FAILED:
6721 cpuset_update_active_cpus(true);
6729 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
6733 case CPU_DOWN_PREPARE:
6734 cpuset_update_active_cpus(false);
6736 case CPU_DOWN_PREPARE_FROZEN:
6738 partition_sched_domains(1, NULL, NULL);
6746 void __init sched_init_smp(void)
6748 cpumask_var_t non_isolated_cpus;
6750 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
6751 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
6756 * There's no userspace yet to cause hotplug operations; hence all the
6757 * cpu masks are stable and all blatant races in the below code cannot
6760 mutex_lock(&sched_domains_mutex);
6761 init_sched_domains(cpu_active_mask);
6762 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
6763 if (cpumask_empty(non_isolated_cpus))
6764 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
6765 mutex_unlock(&sched_domains_mutex);
6767 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
6768 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
6769 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
6773 /* Move init over to a non-isolated CPU */
6774 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
6776 sched_init_granularity();
6777 free_cpumask_var(non_isolated_cpus);
6779 init_sched_rt_class();
6780 init_sched_dl_class();
6783 void __init sched_init_smp(void)
6785 sched_init_granularity();
6787 #endif /* CONFIG_SMP */
6789 const_debug unsigned int sysctl_timer_migration = 1;
6791 int in_sched_functions(unsigned long addr)
6793 return in_lock_functions(addr) ||
6794 (addr >= (unsigned long)__sched_text_start
6795 && addr < (unsigned long)__sched_text_end);
6798 #ifdef CONFIG_CGROUP_SCHED
6800 * Default task group.
6801 * Every task in system belongs to this group at bootup.
6803 struct task_group root_task_group;
6804 LIST_HEAD(task_groups);
6807 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
6809 void __init sched_init(void)
6812 unsigned long alloc_size = 0, ptr;
6814 #ifdef CONFIG_FAIR_GROUP_SCHED
6815 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6817 #ifdef CONFIG_RT_GROUP_SCHED
6818 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
6820 #ifdef CONFIG_CPUMASK_OFFSTACK
6821 alloc_size += num_possible_cpus() * cpumask_size();
6824 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
6826 #ifdef CONFIG_FAIR_GROUP_SCHED
6827 root_task_group.se = (struct sched_entity **)ptr;
6828 ptr += nr_cpu_ids * sizeof(void **);
6830 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
6831 ptr += nr_cpu_ids * sizeof(void **);
6833 #endif /* CONFIG_FAIR_GROUP_SCHED */
6834 #ifdef CONFIG_RT_GROUP_SCHED
6835 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
6836 ptr += nr_cpu_ids * sizeof(void **);
6838 root_task_group.rt_rq = (struct rt_rq **)ptr;
6839 ptr += nr_cpu_ids * sizeof(void **);
6841 #endif /* CONFIG_RT_GROUP_SCHED */
6842 #ifdef CONFIG_CPUMASK_OFFSTACK
6843 for_each_possible_cpu(i) {
6844 per_cpu(load_balance_mask, i) = (void *)ptr;
6845 ptr += cpumask_size();
6847 #endif /* CONFIG_CPUMASK_OFFSTACK */
6850 init_rt_bandwidth(&def_rt_bandwidth,
6851 global_rt_period(), global_rt_runtime());
6852 init_dl_bandwidth(&def_dl_bandwidth,
6853 global_rt_period(), global_rt_runtime());
6856 init_defrootdomain();
6859 #ifdef CONFIG_RT_GROUP_SCHED
6860 init_rt_bandwidth(&root_task_group.rt_bandwidth,
6861 global_rt_period(), global_rt_runtime());
6862 #endif /* CONFIG_RT_GROUP_SCHED */
6864 #ifdef CONFIG_CGROUP_SCHED
6865 list_add(&root_task_group.list, &task_groups);
6866 INIT_LIST_HEAD(&root_task_group.children);
6867 INIT_LIST_HEAD(&root_task_group.siblings);
6868 autogroup_init(&init_task);
6870 #endif /* CONFIG_CGROUP_SCHED */
6872 for_each_possible_cpu(i) {
6876 raw_spin_lock_init(&rq->lock);
6878 rq->calc_load_active = 0;
6879 rq->calc_load_update = jiffies + LOAD_FREQ;
6880 init_cfs_rq(&rq->cfs);
6881 init_rt_rq(&rq->rt, rq);
6882 init_dl_rq(&rq->dl, rq);
6883 #ifdef CONFIG_FAIR_GROUP_SCHED
6884 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
6885 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
6887 * How much cpu bandwidth does root_task_group get?
6889 * In case of task-groups formed thr' the cgroup filesystem, it
6890 * gets 100% of the cpu resources in the system. This overall
6891 * system cpu resource is divided among the tasks of
6892 * root_task_group and its child task-groups in a fair manner,
6893 * based on each entity's (task or task-group's) weight
6894 * (se->load.weight).
6896 * In other words, if root_task_group has 10 tasks of weight
6897 * 1024) and two child groups A0 and A1 (of weight 1024 each),
6898 * then A0's share of the cpu resource is:
6900 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
6902 * We achieve this by letting root_task_group's tasks sit
6903 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
6905 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
6906 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
6907 #endif /* CONFIG_FAIR_GROUP_SCHED */
6909 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
6910 #ifdef CONFIG_RT_GROUP_SCHED
6911 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
6914 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
6915 rq->cpu_load[j] = 0;
6917 rq->last_load_update_tick = jiffies;
6922 rq->cpu_power = SCHED_POWER_SCALE;
6923 rq->post_schedule = 0;
6924 rq->active_balance = 0;
6925 rq->next_balance = jiffies;
6930 rq->avg_idle = 2*sysctl_sched_migration_cost;
6931 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
6933 INIT_LIST_HEAD(&rq->cfs_tasks);
6935 rq_attach_root(rq, &def_root_domain);
6936 #ifdef CONFIG_NO_HZ_COMMON
6939 #ifdef CONFIG_NO_HZ_FULL
6940 rq->last_sched_tick = 0;
6944 atomic_set(&rq->nr_iowait, 0);
6947 set_load_weight(&init_task);
6949 #ifdef CONFIG_PREEMPT_NOTIFIERS
6950 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
6954 * The boot idle thread does lazy MMU switching as well:
6956 atomic_inc(&init_mm.mm_count);
6957 enter_lazy_tlb(&init_mm, current);
6960 * Make us the idle thread. Technically, schedule() should not be
6961 * called from this thread, however somewhere below it might be,
6962 * but because we are the idle thread, we just pick up running again
6963 * when this runqueue becomes "idle".
6965 init_idle(current, smp_processor_id());
6967 calc_load_update = jiffies + LOAD_FREQ;
6970 * During early bootup we pretend to be a normal task:
6972 current->sched_class = &fair_sched_class;
6975 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
6976 /* May be allocated at isolcpus cmdline parse time */
6977 if (cpu_isolated_map == NULL)
6978 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
6979 idle_thread_set_boot_cpu();
6981 init_sched_fair_class();
6983 scheduler_running = 1;
6986 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
6987 static inline int preempt_count_equals(int preempt_offset)
6989 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
6991 return (nested == preempt_offset);
6994 void __might_sleep(const char *file, int line, int preempt_offset)
6996 static unsigned long prev_jiffy; /* ratelimiting */
6998 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
6999 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7000 !is_idle_task(current)) ||
7001 system_state != SYSTEM_RUNNING || oops_in_progress)
7003 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7005 prev_jiffy = jiffies;
7008 "BUG: sleeping function called from invalid context at %s:%d\n",
7011 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7012 in_atomic(), irqs_disabled(),
7013 current->pid, current->comm);
7015 debug_show_held_locks(current);
7016 if (irqs_disabled())
7017 print_irqtrace_events(current);
7018 #ifdef CONFIG_DEBUG_PREEMPT
7019 if (!preempt_count_equals(preempt_offset)) {
7020 pr_err("Preemption disabled at:");
7021 print_ip_sym(current->preempt_disable_ip);
7027 EXPORT_SYMBOL(__might_sleep);
7030 #ifdef CONFIG_MAGIC_SYSRQ
7031 static void normalize_task(struct rq *rq, struct task_struct *p)
7033 const struct sched_class *prev_class = p->sched_class;
7034 struct sched_attr attr = {
7035 .sched_policy = SCHED_NORMAL,
7037 int old_prio = p->prio;
7042 dequeue_task(rq, p, 0);
7043 __setscheduler(rq, p, &attr);
7045 enqueue_task(rq, p, 0);
7046 resched_task(rq->curr);
7049 check_class_changed(rq, p, prev_class, old_prio);
7052 void normalize_rt_tasks(void)
7054 struct task_struct *g, *p;
7055 unsigned long flags;
7058 read_lock_irqsave(&tasklist_lock, flags);
7059 do_each_thread(g, p) {
7061 * Only normalize user tasks:
7066 p->se.exec_start = 0;
7067 #ifdef CONFIG_SCHEDSTATS
7068 p->se.statistics.wait_start = 0;
7069 p->se.statistics.sleep_start = 0;
7070 p->se.statistics.block_start = 0;
7073 if (!dl_task(p) && !rt_task(p)) {
7075 * Renice negative nice level userspace
7078 if (task_nice(p) < 0 && p->mm)
7079 set_user_nice(p, 0);
7083 raw_spin_lock(&p->pi_lock);
7084 rq = __task_rq_lock(p);
7086 normalize_task(rq, p);
7088 __task_rq_unlock(rq);
7089 raw_spin_unlock(&p->pi_lock);
7090 } while_each_thread(g, p);
7092 read_unlock_irqrestore(&tasklist_lock, flags);
7095 #endif /* CONFIG_MAGIC_SYSRQ */
7097 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7099 * These functions are only useful for the IA64 MCA handling, or kdb.
7101 * They can only be called when the whole system has been
7102 * stopped - every CPU needs to be quiescent, and no scheduling
7103 * activity can take place. Using them for anything else would
7104 * be a serious bug, and as a result, they aren't even visible
7105 * under any other configuration.
7109 * curr_task - return the current task for a given cpu.
7110 * @cpu: the processor in question.
7112 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7114 * Return: The current task for @cpu.
7116 struct task_struct *curr_task(int cpu)
7118 return cpu_curr(cpu);
7121 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7125 * set_curr_task - set the current task for a given cpu.
7126 * @cpu: the processor in question.
7127 * @p: the task pointer to set.
7129 * Description: This function must only be used when non-maskable interrupts
7130 * are serviced on a separate stack. It allows the architecture to switch the
7131 * notion of the current task on a cpu in a non-blocking manner. This function
7132 * must be called with all CPU's synchronized, and interrupts disabled, the
7133 * and caller must save the original value of the current task (see
7134 * curr_task() above) and restore that value before reenabling interrupts and
7135 * re-starting the system.
7137 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7139 void set_curr_task(int cpu, struct task_struct *p)
7146 #ifdef CONFIG_CGROUP_SCHED
7147 /* task_group_lock serializes the addition/removal of task groups */
7148 static DEFINE_SPINLOCK(task_group_lock);
7150 static void free_sched_group(struct task_group *tg)
7152 free_fair_sched_group(tg);
7153 free_rt_sched_group(tg);
7158 /* allocate runqueue etc for a new task group */
7159 struct task_group *sched_create_group(struct task_group *parent)
7161 struct task_group *tg;
7163 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
7165 return ERR_PTR(-ENOMEM);
7167 if (!alloc_fair_sched_group(tg, parent))
7170 if (!alloc_rt_sched_group(tg, parent))
7176 free_sched_group(tg);
7177 return ERR_PTR(-ENOMEM);
7180 void sched_online_group(struct task_group *tg, struct task_group *parent)
7182 unsigned long flags;
7184 spin_lock_irqsave(&task_group_lock, flags);
7185 list_add_rcu(&tg->list, &task_groups);
7187 WARN_ON(!parent); /* root should already exist */
7189 tg->parent = parent;
7190 INIT_LIST_HEAD(&tg->children);
7191 list_add_rcu(&tg->siblings, &parent->children);
7192 spin_unlock_irqrestore(&task_group_lock, flags);
7195 /* rcu callback to free various structures associated with a task group */
7196 static void free_sched_group_rcu(struct rcu_head *rhp)
7198 /* now it should be safe to free those cfs_rqs */
7199 free_sched_group(container_of(rhp, struct task_group, rcu));
7202 /* Destroy runqueue etc associated with a task group */
7203 void sched_destroy_group(struct task_group *tg)
7205 /* wait for possible concurrent references to cfs_rqs complete */
7206 call_rcu(&tg->rcu, free_sched_group_rcu);
7209 void sched_offline_group(struct task_group *tg)
7211 unsigned long flags;
7214 /* end participation in shares distribution */
7215 for_each_possible_cpu(i)
7216 unregister_fair_sched_group(tg, i);
7218 spin_lock_irqsave(&task_group_lock, flags);
7219 list_del_rcu(&tg->list);
7220 list_del_rcu(&tg->siblings);
7221 spin_unlock_irqrestore(&task_group_lock, flags);
7224 /* change task's runqueue when it moves between groups.
7225 * The caller of this function should have put the task in its new group
7226 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
7227 * reflect its new group.
7229 void sched_move_task(struct task_struct *tsk)
7231 struct task_group *tg;
7233 unsigned long flags;
7236 rq = task_rq_lock(tsk, &flags);
7238 running = task_current(rq, tsk);
7242 dequeue_task(rq, tsk, 0);
7243 if (unlikely(running))
7244 tsk->sched_class->put_prev_task(rq, tsk);
7246 tg = container_of(task_css_check(tsk, cpu_cgrp_id,
7247 lockdep_is_held(&tsk->sighand->siglock)),
7248 struct task_group, css);
7249 tg = autogroup_task_group(tsk, tg);
7250 tsk->sched_task_group = tg;
7252 #ifdef CONFIG_FAIR_GROUP_SCHED
7253 if (tsk->sched_class->task_move_group)
7254 tsk->sched_class->task_move_group(tsk, on_rq);
7257 set_task_rq(tsk, task_cpu(tsk));
7259 if (unlikely(running))
7260 tsk->sched_class->set_curr_task(rq);
7262 enqueue_task(rq, tsk, 0);
7264 task_rq_unlock(rq, tsk, &flags);
7266 #endif /* CONFIG_CGROUP_SCHED */
7268 #ifdef CONFIG_RT_GROUP_SCHED
7270 * Ensure that the real time constraints are schedulable.
7272 static DEFINE_MUTEX(rt_constraints_mutex);
7274 /* Must be called with tasklist_lock held */
7275 static inline int tg_has_rt_tasks(struct task_group *tg)
7277 struct task_struct *g, *p;
7279 do_each_thread(g, p) {
7280 if (rt_task(p) && task_rq(p)->rt.tg == tg)
7282 } while_each_thread(g, p);
7287 struct rt_schedulable_data {
7288 struct task_group *tg;
7293 static int tg_rt_schedulable(struct task_group *tg, void *data)
7295 struct rt_schedulable_data *d = data;
7296 struct task_group *child;
7297 unsigned long total, sum = 0;
7298 u64 period, runtime;
7300 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7301 runtime = tg->rt_bandwidth.rt_runtime;
7304 period = d->rt_period;
7305 runtime = d->rt_runtime;
7309 * Cannot have more runtime than the period.
7311 if (runtime > period && runtime != RUNTIME_INF)
7315 * Ensure we don't starve existing RT tasks.
7317 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
7320 total = to_ratio(period, runtime);
7323 * Nobody can have more than the global setting allows.
7325 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
7329 * The sum of our children's runtime should not exceed our own.
7331 list_for_each_entry_rcu(child, &tg->children, siblings) {
7332 period = ktime_to_ns(child->rt_bandwidth.rt_period);
7333 runtime = child->rt_bandwidth.rt_runtime;
7335 if (child == d->tg) {
7336 period = d->rt_period;
7337 runtime = d->rt_runtime;
7340 sum += to_ratio(period, runtime);
7349 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
7353 struct rt_schedulable_data data = {
7355 .rt_period = period,
7356 .rt_runtime = runtime,
7360 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
7366 static int tg_set_rt_bandwidth(struct task_group *tg,
7367 u64 rt_period, u64 rt_runtime)
7371 mutex_lock(&rt_constraints_mutex);
7372 read_lock(&tasklist_lock);
7373 err = __rt_schedulable(tg, rt_period, rt_runtime);
7377 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7378 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
7379 tg->rt_bandwidth.rt_runtime = rt_runtime;
7381 for_each_possible_cpu(i) {
7382 struct rt_rq *rt_rq = tg->rt_rq[i];
7384 raw_spin_lock(&rt_rq->rt_runtime_lock);
7385 rt_rq->rt_runtime = rt_runtime;
7386 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7388 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
7390 read_unlock(&tasklist_lock);
7391 mutex_unlock(&rt_constraints_mutex);
7396 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
7398 u64 rt_runtime, rt_period;
7400 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
7401 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
7402 if (rt_runtime_us < 0)
7403 rt_runtime = RUNTIME_INF;
7405 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7408 static long sched_group_rt_runtime(struct task_group *tg)
7412 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
7415 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
7416 do_div(rt_runtime_us, NSEC_PER_USEC);
7417 return rt_runtime_us;
7420 static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
7422 u64 rt_runtime, rt_period;
7424 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
7425 rt_runtime = tg->rt_bandwidth.rt_runtime;
7430 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
7433 static long sched_group_rt_period(struct task_group *tg)
7437 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
7438 do_div(rt_period_us, NSEC_PER_USEC);
7439 return rt_period_us;
7441 #endif /* CONFIG_RT_GROUP_SCHED */
7443 #ifdef CONFIG_RT_GROUP_SCHED
7444 static int sched_rt_global_constraints(void)
7448 mutex_lock(&rt_constraints_mutex);
7449 read_lock(&tasklist_lock);
7450 ret = __rt_schedulable(NULL, 0, 0);
7451 read_unlock(&tasklist_lock);
7452 mutex_unlock(&rt_constraints_mutex);
7457 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
7459 /* Don't accept realtime tasks when there is no way for them to run */
7460 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
7466 #else /* !CONFIG_RT_GROUP_SCHED */
7467 static int sched_rt_global_constraints(void)
7469 unsigned long flags;
7472 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
7473 for_each_possible_cpu(i) {
7474 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
7476 raw_spin_lock(&rt_rq->rt_runtime_lock);
7477 rt_rq->rt_runtime = global_rt_runtime();
7478 raw_spin_unlock(&rt_rq->rt_runtime_lock);
7480 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
7484 #endif /* CONFIG_RT_GROUP_SCHED */
7486 static int sched_dl_global_constraints(void)
7488 u64 runtime = global_rt_runtime();
7489 u64 period = global_rt_period();
7490 u64 new_bw = to_ratio(period, runtime);
7492 unsigned long flags;
7495 * Here we want to check the bandwidth not being set to some
7496 * value smaller than the currently allocated bandwidth in
7497 * any of the root_domains.
7499 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
7500 * cycling on root_domains... Discussion on different/better
7501 * solutions is welcome!
7503 for_each_possible_cpu(cpu) {
7504 struct dl_bw *dl_b = dl_bw_of(cpu);
7506 raw_spin_lock_irqsave(&dl_b->lock, flags);
7507 if (new_bw < dl_b->total_bw)
7509 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7518 static void sched_dl_do_global(void)
7522 unsigned long flags;
7524 def_dl_bandwidth.dl_period = global_rt_period();
7525 def_dl_bandwidth.dl_runtime = global_rt_runtime();
7527 if (global_rt_runtime() != RUNTIME_INF)
7528 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
7531 * FIXME: As above...
7533 for_each_possible_cpu(cpu) {
7534 struct dl_bw *dl_b = dl_bw_of(cpu);
7536 raw_spin_lock_irqsave(&dl_b->lock, flags);
7538 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7542 static int sched_rt_global_validate(void)
7544 if (sysctl_sched_rt_period <= 0)
7547 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
7548 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
7554 static void sched_rt_do_global(void)
7556 def_rt_bandwidth.rt_runtime = global_rt_runtime();
7557 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
7560 int sched_rt_handler(struct ctl_table *table, int write,
7561 void __user *buffer, size_t *lenp,
7564 int old_period, old_runtime;
7565 static DEFINE_MUTEX(mutex);
7569 old_period = sysctl_sched_rt_period;
7570 old_runtime = sysctl_sched_rt_runtime;
7572 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7574 if (!ret && write) {
7575 ret = sched_rt_global_validate();
7579 ret = sched_rt_global_constraints();
7583 ret = sched_dl_global_constraints();
7587 sched_rt_do_global();
7588 sched_dl_do_global();
7592 sysctl_sched_rt_period = old_period;
7593 sysctl_sched_rt_runtime = old_runtime;
7595 mutex_unlock(&mutex);
7600 int sched_rr_handler(struct ctl_table *table, int write,
7601 void __user *buffer, size_t *lenp,
7605 static DEFINE_MUTEX(mutex);
7608 ret = proc_dointvec(table, write, buffer, lenp, ppos);
7609 /* make sure that internally we keep jiffies */
7610 /* also, writing zero resets timeslice to default */
7611 if (!ret && write) {
7612 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
7613 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
7615 mutex_unlock(&mutex);
7619 #ifdef CONFIG_CGROUP_SCHED
7621 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
7623 return css ? container_of(css, struct task_group, css) : NULL;
7626 static struct cgroup_subsys_state *
7627 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
7629 struct task_group *parent = css_tg(parent_css);
7630 struct task_group *tg;
7633 /* This is early initialization for the top cgroup */
7634 return &root_task_group.css;
7637 tg = sched_create_group(parent);
7639 return ERR_PTR(-ENOMEM);
7644 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
7646 struct task_group *tg = css_tg(css);
7647 struct task_group *parent = css_tg(css_parent(css));
7650 sched_online_group(tg, parent);
7654 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
7656 struct task_group *tg = css_tg(css);
7658 sched_destroy_group(tg);
7661 static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css)
7663 struct task_group *tg = css_tg(css);
7665 sched_offline_group(tg);
7668 static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css,
7669 struct cgroup_taskset *tset)
7671 struct task_struct *task;
7673 cgroup_taskset_for_each(task, tset) {
7674 #ifdef CONFIG_RT_GROUP_SCHED
7675 if (!sched_rt_can_attach(css_tg(css), task))
7678 /* We don't support RT-tasks being in separate groups */
7679 if (task->sched_class != &fair_sched_class)
7686 static void cpu_cgroup_attach(struct cgroup_subsys_state *css,
7687 struct cgroup_taskset *tset)
7689 struct task_struct *task;
7691 cgroup_taskset_for_each(task, tset)
7692 sched_move_task(task);
7695 static void cpu_cgroup_exit(struct cgroup_subsys_state *css,
7696 struct cgroup_subsys_state *old_css,
7697 struct task_struct *task)
7700 * cgroup_exit() is called in the copy_process() failure path.
7701 * Ignore this case since the task hasn't ran yet, this avoids
7702 * trying to poke a half freed task state from generic code.
7704 if (!(task->flags & PF_EXITING))
7707 sched_move_task(task);
7710 #ifdef CONFIG_FAIR_GROUP_SCHED
7711 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
7712 struct cftype *cftype, u64 shareval)
7714 return sched_group_set_shares(css_tg(css), scale_load(shareval));
7717 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
7720 struct task_group *tg = css_tg(css);
7722 return (u64) scale_load_down(tg->shares);
7725 #ifdef CONFIG_CFS_BANDWIDTH
7726 static DEFINE_MUTEX(cfs_constraints_mutex);
7728 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
7729 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
7731 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
7733 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
7735 int i, ret = 0, runtime_enabled, runtime_was_enabled;
7736 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7738 if (tg == &root_task_group)
7742 * Ensure we have at some amount of bandwidth every period. This is
7743 * to prevent reaching a state of large arrears when throttled via
7744 * entity_tick() resulting in prolonged exit starvation.
7746 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
7750 * Likewise, bound things on the otherside by preventing insane quota
7751 * periods. This also allows us to normalize in computing quota
7754 if (period > max_cfs_quota_period)
7757 mutex_lock(&cfs_constraints_mutex);
7758 ret = __cfs_schedulable(tg, period, quota);
7762 runtime_enabled = quota != RUNTIME_INF;
7763 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
7765 * If we need to toggle cfs_bandwidth_used, off->on must occur
7766 * before making related changes, and on->off must occur afterwards
7768 if (runtime_enabled && !runtime_was_enabled)
7769 cfs_bandwidth_usage_inc();
7770 raw_spin_lock_irq(&cfs_b->lock);
7771 cfs_b->period = ns_to_ktime(period);
7772 cfs_b->quota = quota;
7774 __refill_cfs_bandwidth_runtime(cfs_b);
7775 /* restart the period timer (if active) to handle new period expiry */
7776 if (runtime_enabled && cfs_b->timer_active) {
7777 /* force a reprogram */
7778 cfs_b->timer_active = 0;
7779 __start_cfs_bandwidth(cfs_b);
7781 raw_spin_unlock_irq(&cfs_b->lock);
7783 for_each_possible_cpu(i) {
7784 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
7785 struct rq *rq = cfs_rq->rq;
7787 raw_spin_lock_irq(&rq->lock);
7788 cfs_rq->runtime_enabled = runtime_enabled;
7789 cfs_rq->runtime_remaining = 0;
7791 if (cfs_rq->throttled)
7792 unthrottle_cfs_rq(cfs_rq);
7793 raw_spin_unlock_irq(&rq->lock);
7795 if (runtime_was_enabled && !runtime_enabled)
7796 cfs_bandwidth_usage_dec();
7798 mutex_unlock(&cfs_constraints_mutex);
7803 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
7807 period = ktime_to_ns(tg->cfs_bandwidth.period);
7808 if (cfs_quota_us < 0)
7809 quota = RUNTIME_INF;
7811 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
7813 return tg_set_cfs_bandwidth(tg, period, quota);
7816 long tg_get_cfs_quota(struct task_group *tg)
7820 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
7823 quota_us = tg->cfs_bandwidth.quota;
7824 do_div(quota_us, NSEC_PER_USEC);
7829 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
7833 period = (u64)cfs_period_us * NSEC_PER_USEC;
7834 quota = tg->cfs_bandwidth.quota;
7836 return tg_set_cfs_bandwidth(tg, period, quota);
7839 long tg_get_cfs_period(struct task_group *tg)
7843 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
7844 do_div(cfs_period_us, NSEC_PER_USEC);
7846 return cfs_period_us;
7849 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
7852 return tg_get_cfs_quota(css_tg(css));
7855 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
7856 struct cftype *cftype, s64 cfs_quota_us)
7858 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
7861 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
7864 return tg_get_cfs_period(css_tg(css));
7867 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
7868 struct cftype *cftype, u64 cfs_period_us)
7870 return tg_set_cfs_period(css_tg(css), cfs_period_us);
7873 struct cfs_schedulable_data {
7874 struct task_group *tg;
7879 * normalize group quota/period to be quota/max_period
7880 * note: units are usecs
7882 static u64 normalize_cfs_quota(struct task_group *tg,
7883 struct cfs_schedulable_data *d)
7891 period = tg_get_cfs_period(tg);
7892 quota = tg_get_cfs_quota(tg);
7895 /* note: these should typically be equivalent */
7896 if (quota == RUNTIME_INF || quota == -1)
7899 return to_ratio(period, quota);
7902 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
7904 struct cfs_schedulable_data *d = data;
7905 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7906 s64 quota = 0, parent_quota = -1;
7909 quota = RUNTIME_INF;
7911 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
7913 quota = normalize_cfs_quota(tg, d);
7914 parent_quota = parent_b->hierarchal_quota;
7917 * ensure max(child_quota) <= parent_quota, inherit when no
7920 if (quota == RUNTIME_INF)
7921 quota = parent_quota;
7922 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
7925 cfs_b->hierarchal_quota = quota;
7930 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
7933 struct cfs_schedulable_data data = {
7939 if (quota != RUNTIME_INF) {
7940 do_div(data.period, NSEC_PER_USEC);
7941 do_div(data.quota, NSEC_PER_USEC);
7945 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
7951 static int cpu_stats_show(struct seq_file *sf, void *v)
7953 struct task_group *tg = css_tg(seq_css(sf));
7954 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
7956 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
7957 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
7958 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
7962 #endif /* CONFIG_CFS_BANDWIDTH */
7963 #endif /* CONFIG_FAIR_GROUP_SCHED */
7965 #ifdef CONFIG_RT_GROUP_SCHED
7966 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
7967 struct cftype *cft, s64 val)
7969 return sched_group_set_rt_runtime(css_tg(css), val);
7972 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
7975 return sched_group_rt_runtime(css_tg(css));
7978 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
7979 struct cftype *cftype, u64 rt_period_us)
7981 return sched_group_set_rt_period(css_tg(css), rt_period_us);
7984 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
7987 return sched_group_rt_period(css_tg(css));
7989 #endif /* CONFIG_RT_GROUP_SCHED */
7991 static struct cftype cpu_files[] = {
7992 #ifdef CONFIG_FAIR_GROUP_SCHED
7995 .read_u64 = cpu_shares_read_u64,
7996 .write_u64 = cpu_shares_write_u64,
7999 #ifdef CONFIG_CFS_BANDWIDTH
8001 .name = "cfs_quota_us",
8002 .read_s64 = cpu_cfs_quota_read_s64,
8003 .write_s64 = cpu_cfs_quota_write_s64,
8006 .name = "cfs_period_us",
8007 .read_u64 = cpu_cfs_period_read_u64,
8008 .write_u64 = cpu_cfs_period_write_u64,
8012 .seq_show = cpu_stats_show,
8015 #ifdef CONFIG_RT_GROUP_SCHED
8017 .name = "rt_runtime_us",
8018 .read_s64 = cpu_rt_runtime_read,
8019 .write_s64 = cpu_rt_runtime_write,
8022 .name = "rt_period_us",
8023 .read_u64 = cpu_rt_period_read_uint,
8024 .write_u64 = cpu_rt_period_write_uint,
8030 struct cgroup_subsys cpu_cgrp_subsys = {
8031 .css_alloc = cpu_cgroup_css_alloc,
8032 .css_free = cpu_cgroup_css_free,
8033 .css_online = cpu_cgroup_css_online,
8034 .css_offline = cpu_cgroup_css_offline,
8035 .can_attach = cpu_cgroup_can_attach,
8036 .attach = cpu_cgroup_attach,
8037 .exit = cpu_cgroup_exit,
8038 .base_cftypes = cpu_files,
8042 #endif /* CONFIG_CGROUP_SCHED */
8044 void dump_cpu_task(int cpu)
8046 pr_info("Task dump for CPU %d:\n", cpu);
8047 sched_show_task(cpu_curr(cpu));