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 <linux/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>
77 #include <linux/cpufreq_times.h>
79 #include <asm/switch_to.h>
81 #include <asm/irq_regs.h>
82 #include <asm/mutex.h>
83 #ifdef CONFIG_PARAVIRT
84 #include <asm/paravirt.h>
88 #include "../workqueue_internal.h"
89 #include "../smpboot.h"
91 #define CREATE_TRACE_POINTS
92 #include <trace/events/sched.h>
95 DEFINE_MUTEX(sched_domains_mutex);
96 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
98 static void update_rq_clock_task(struct rq *rq, s64 delta);
100 void update_rq_clock(struct rq *rq)
104 lockdep_assert_held(&rq->lock);
106 if (rq->clock_skip_update & RQCF_ACT_SKIP)
109 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
113 update_rq_clock_task(rq, delta);
117 * Debugging: various feature bits
120 #define SCHED_FEAT(name, enabled) \
121 (1UL << __SCHED_FEAT_##name) * enabled |
123 const_debug unsigned int sysctl_sched_features =
124 #include "features.h"
129 #ifdef CONFIG_SCHED_DEBUG
130 #define SCHED_FEAT(name, enabled) \
133 static const char * const sched_feat_names[] = {
134 #include "features.h"
139 static int sched_feat_show(struct seq_file *m, void *v)
143 for (i = 0; i < __SCHED_FEAT_NR; i++) {
144 if (!(sysctl_sched_features & (1UL << i)))
146 seq_printf(m, "%s ", sched_feat_names[i]);
153 #ifdef HAVE_JUMP_LABEL
155 #define jump_label_key__true STATIC_KEY_INIT_TRUE
156 #define jump_label_key__false STATIC_KEY_INIT_FALSE
158 #define SCHED_FEAT(name, enabled) \
159 jump_label_key__##enabled ,
161 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
162 #include "features.h"
167 static void sched_feat_disable(int i)
169 static_key_disable(&sched_feat_keys[i]);
172 static void sched_feat_enable(int i)
174 static_key_enable(&sched_feat_keys[i]);
177 static void sched_feat_disable(int i) { };
178 static void sched_feat_enable(int i) { };
179 #endif /* HAVE_JUMP_LABEL */
181 static int sched_feat_set(char *cmp)
186 if (strncmp(cmp, "NO_", 3) == 0) {
191 for (i = 0; i < __SCHED_FEAT_NR; i++) {
192 if (strcmp(cmp, sched_feat_names[i]) == 0) {
194 sysctl_sched_features &= ~(1UL << i);
195 sched_feat_disable(i);
197 sysctl_sched_features |= (1UL << i);
198 sched_feat_enable(i);
208 sched_feat_write(struct file *filp, const char __user *ubuf,
209 size_t cnt, loff_t *ppos)
219 if (copy_from_user(&buf, ubuf, cnt))
225 /* Ensure the static_key remains in a consistent state */
226 inode = file_inode(filp);
227 mutex_lock(&inode->i_mutex);
228 i = sched_feat_set(cmp);
229 mutex_unlock(&inode->i_mutex);
230 if (i == __SCHED_FEAT_NR)
238 static int sched_feat_open(struct inode *inode, struct file *filp)
240 return single_open(filp, sched_feat_show, NULL);
243 static const struct file_operations sched_feat_fops = {
244 .open = sched_feat_open,
245 .write = sched_feat_write,
248 .release = single_release,
251 static __init int sched_init_debug(void)
253 debugfs_create_file("sched_features", 0644, NULL, NULL,
258 late_initcall(sched_init_debug);
259 #endif /* CONFIG_SCHED_DEBUG */
262 * Number of tasks to iterate in a single balance run.
263 * Limited because this is done with IRQs disabled.
265 const_debug unsigned int sysctl_sched_nr_migrate = 32;
268 * period over which we average the RT time consumption, measured
273 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
276 * period over which we measure -rt task cpu usage in us.
279 unsigned int sysctl_sched_rt_period = 1000000;
281 __read_mostly int scheduler_running;
284 * part of the period that we allow rt tasks to run in us.
287 int sysctl_sched_rt_runtime = 950000;
289 /* cpus with isolated domains */
290 cpumask_var_t cpu_isolated_map;
293 lock_rq_of(struct task_struct *p, unsigned long *flags)
295 return task_rq_lock(p, flags);
299 unlock_rq_of(struct rq *rq, struct task_struct *p, unsigned long *flags)
301 task_rq_unlock(rq, p, flags);
305 * this_rq_lock - lock this runqueue and disable interrupts.
307 static struct rq *this_rq_lock(void)
314 raw_spin_lock(&rq->lock);
319 #ifdef CONFIG_SCHED_HRTICK
321 * Use HR-timers to deliver accurate preemption points.
324 static void hrtick_clear(struct rq *rq)
326 if (hrtimer_active(&rq->hrtick_timer))
327 hrtimer_cancel(&rq->hrtick_timer);
331 * High-resolution timer tick.
332 * Runs from hardirq context with interrupts disabled.
334 static enum hrtimer_restart hrtick(struct hrtimer *timer)
336 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
338 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
340 raw_spin_lock(&rq->lock);
342 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
343 raw_spin_unlock(&rq->lock);
345 return HRTIMER_NORESTART;
350 static void __hrtick_restart(struct rq *rq)
352 struct hrtimer *timer = &rq->hrtick_timer;
354 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
358 * called from hardirq (IPI) context
360 static void __hrtick_start(void *arg)
364 raw_spin_lock(&rq->lock);
365 __hrtick_restart(rq);
366 rq->hrtick_csd_pending = 0;
367 raw_spin_unlock(&rq->lock);
371 * Called to set the hrtick timer state.
373 * called with rq->lock held and irqs disabled
375 void hrtick_start(struct rq *rq, u64 delay)
377 struct hrtimer *timer = &rq->hrtick_timer;
382 * Don't schedule slices shorter than 10000ns, that just
383 * doesn't make sense and can cause timer DoS.
385 delta = max_t(s64, delay, 10000LL);
386 time = ktime_add_ns(timer->base->get_time(), delta);
388 hrtimer_set_expires(timer, time);
390 if (rq == this_rq()) {
391 __hrtick_restart(rq);
392 } else if (!rq->hrtick_csd_pending) {
393 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
394 rq->hrtick_csd_pending = 1;
399 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
401 int cpu = (int)(long)hcpu;
404 case CPU_UP_CANCELED:
405 case CPU_UP_CANCELED_FROZEN:
406 case CPU_DOWN_PREPARE:
407 case CPU_DOWN_PREPARE_FROZEN:
409 case CPU_DEAD_FROZEN:
410 hrtick_clear(cpu_rq(cpu));
417 static __init void init_hrtick(void)
419 hotcpu_notifier(hotplug_hrtick, 0);
423 * Called to set the hrtick timer state.
425 * called with rq->lock held and irqs disabled
427 void hrtick_start(struct rq *rq, u64 delay)
430 * Don't schedule slices shorter than 10000ns, that just
431 * doesn't make sense. Rely on vruntime for fairness.
433 delay = max_t(u64, delay, 10000LL);
434 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
435 HRTIMER_MODE_REL_PINNED);
438 static inline void init_hrtick(void)
441 #endif /* CONFIG_SMP */
443 static void init_rq_hrtick(struct rq *rq)
446 rq->hrtick_csd_pending = 0;
448 rq->hrtick_csd.flags = 0;
449 rq->hrtick_csd.func = __hrtick_start;
450 rq->hrtick_csd.info = rq;
453 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
454 rq->hrtick_timer.function = hrtick;
456 #else /* CONFIG_SCHED_HRTICK */
457 static inline void hrtick_clear(struct rq *rq)
461 static inline void init_rq_hrtick(struct rq *rq)
465 static inline void init_hrtick(void)
468 #endif /* CONFIG_SCHED_HRTICK */
471 * cmpxchg based fetch_or, macro so it works for different integer types
473 #define fetch_or(ptr, val) \
474 ({ typeof(*(ptr)) __old, __val = *(ptr); \
476 __old = cmpxchg((ptr), __val, __val | (val)); \
477 if (__old == __val) \
484 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
486 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
487 * this avoids any races wrt polling state changes and thereby avoids
490 static bool set_nr_and_not_polling(struct task_struct *p)
492 struct thread_info *ti = task_thread_info(p);
493 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
497 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
499 * If this returns true, then the idle task promises to call
500 * sched_ttwu_pending() and reschedule soon.
502 static bool set_nr_if_polling(struct task_struct *p)
504 struct thread_info *ti = task_thread_info(p);
505 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
508 if (!(val & _TIF_POLLING_NRFLAG))
510 if (val & _TIF_NEED_RESCHED)
512 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
521 static bool set_nr_and_not_polling(struct task_struct *p)
523 set_tsk_need_resched(p);
528 static bool set_nr_if_polling(struct task_struct *p)
535 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
537 struct wake_q_node *node = &task->wake_q;
540 * Atomically grab the task, if ->wake_q is !nil already it means
541 * its already queued (either by us or someone else) and will get the
542 * wakeup due to that.
544 * This cmpxchg() implies a full barrier, which pairs with the write
545 * barrier implied by the wakeup in wake_up_list().
547 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
552 get_task_struct(task);
555 * The head is context local, there can be no concurrency.
558 head->lastp = &node->next;
562 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
563 int sibling_count_hint);
565 void wake_up_q(struct wake_q_head *head)
567 struct wake_q_node *node = head->first;
569 while (node != WAKE_Q_TAIL) {
570 struct task_struct *task;
572 task = container_of(node, struct task_struct, wake_q);
574 /* task can safely be re-inserted now */
576 task->wake_q.next = NULL;
579 * try_to_wake_up() implies a wmb() to pair with the queueing
580 * in wake_q_add() so as not to miss wakeups.
582 try_to_wake_up(task, TASK_NORMAL, 0, head->count);
583 put_task_struct(task);
588 * resched_curr - mark rq's current task 'to be rescheduled now'.
590 * On UP this means the setting of the need_resched flag, on SMP it
591 * might also involve a cross-CPU call to trigger the scheduler on
594 void resched_curr(struct rq *rq)
596 struct task_struct *curr = rq->curr;
599 lockdep_assert_held(&rq->lock);
601 if (test_tsk_need_resched(curr))
606 if (cpu == smp_processor_id()) {
607 set_tsk_need_resched(curr);
608 set_preempt_need_resched();
612 if (set_nr_and_not_polling(curr))
613 smp_send_reschedule(cpu);
615 trace_sched_wake_idle_without_ipi(cpu);
618 void resched_cpu(int cpu)
620 struct rq *rq = cpu_rq(cpu);
623 raw_spin_lock_irqsave(&rq->lock, flags);
624 if (cpu_online(cpu) || cpu == smp_processor_id())
626 raw_spin_unlock_irqrestore(&rq->lock, flags);
630 #ifdef CONFIG_NO_HZ_COMMON
632 * In the semi idle case, use the nearest busy cpu for migrating timers
633 * from an idle cpu. This is good for power-savings.
635 * We don't do similar optimization for completely idle system, as
636 * selecting an idle cpu will add more delays to the timers than intended
637 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
639 int get_nohz_timer_target(void)
641 int i, cpu = smp_processor_id();
642 struct sched_domain *sd;
644 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
648 for_each_domain(cpu, sd) {
649 for_each_cpu(i, sched_domain_span(sd)) {
653 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
660 if (!is_housekeeping_cpu(cpu))
661 cpu = housekeeping_any_cpu();
667 * When add_timer_on() enqueues a timer into the timer wheel of an
668 * idle CPU then this timer might expire before the next timer event
669 * which is scheduled to wake up that CPU. In case of a completely
670 * idle system the next event might even be infinite time into the
671 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
672 * leaves the inner idle loop so the newly added timer is taken into
673 * account when the CPU goes back to idle and evaluates the timer
674 * wheel for the next timer event.
676 static void wake_up_idle_cpu(int cpu)
678 struct rq *rq = cpu_rq(cpu);
680 if (cpu == smp_processor_id())
683 if (set_nr_and_not_polling(rq->idle))
684 smp_send_reschedule(cpu);
686 trace_sched_wake_idle_without_ipi(cpu);
689 static bool wake_up_full_nohz_cpu(int cpu)
692 * We just need the target to call irq_exit() and re-evaluate
693 * the next tick. The nohz full kick at least implies that.
694 * If needed we can still optimize that later with an
697 if (tick_nohz_full_cpu(cpu)) {
698 if (cpu != smp_processor_id() ||
699 tick_nohz_tick_stopped())
700 tick_nohz_full_kick_cpu(cpu);
707 void wake_up_nohz_cpu(int cpu)
709 if (!wake_up_full_nohz_cpu(cpu))
710 wake_up_idle_cpu(cpu);
713 static inline bool got_nohz_idle_kick(void)
715 int cpu = smp_processor_id();
717 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
720 if (idle_cpu(cpu) && !need_resched())
724 * We can't run Idle Load Balance on this CPU for this time so we
725 * cancel it and clear NOHZ_BALANCE_KICK
727 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
731 #else /* CONFIG_NO_HZ_COMMON */
733 static inline bool got_nohz_idle_kick(void)
738 #endif /* CONFIG_NO_HZ_COMMON */
740 #ifdef CONFIG_NO_HZ_FULL
741 bool sched_can_stop_tick(void)
744 * FIFO realtime policy runs the highest priority task. Other runnable
745 * tasks are of a lower priority. The scheduler tick does nothing.
747 if (current->policy == SCHED_FIFO)
751 * Round-robin realtime tasks time slice with other tasks at the same
752 * realtime priority. Is this task the only one at this priority?
754 if (current->policy == SCHED_RR) {
755 struct sched_rt_entity *rt_se = ¤t->rt;
757 return rt_se->run_list.prev == rt_se->run_list.next;
761 * More than one running task need preemption.
762 * nr_running update is assumed to be visible
763 * after IPI is sent from wakers.
765 if (this_rq()->nr_running > 1)
770 #endif /* CONFIG_NO_HZ_FULL */
772 void sched_avg_update(struct rq *rq)
774 s64 period = sched_avg_period();
776 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
778 * Inline assembly required to prevent the compiler
779 * optimising this loop into a divmod call.
780 * See __iter_div_u64_rem() for another example of this.
782 asm("" : "+rm" (rq->age_stamp));
783 rq->age_stamp += period;
788 #endif /* CONFIG_SMP */
790 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
791 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
793 * Iterate task_group tree rooted at *from, calling @down when first entering a
794 * node and @up when leaving it for the final time.
796 * Caller must hold rcu_lock or sufficient equivalent.
798 int walk_tg_tree_from(struct task_group *from,
799 tg_visitor down, tg_visitor up, void *data)
801 struct task_group *parent, *child;
807 ret = (*down)(parent, data);
810 list_for_each_entry_rcu(child, &parent->children, siblings) {
817 ret = (*up)(parent, data);
818 if (ret || parent == from)
822 parent = parent->parent;
829 int tg_nop(struct task_group *tg, void *data)
835 static void set_load_weight(struct task_struct *p)
837 int prio = p->static_prio - MAX_RT_PRIO;
838 struct load_weight *load = &p->se.load;
841 * SCHED_IDLE tasks get minimal weight:
843 if (idle_policy(p->policy)) {
844 load->weight = scale_load(WEIGHT_IDLEPRIO);
845 load->inv_weight = WMULT_IDLEPRIO;
849 load->weight = scale_load(prio_to_weight[prio]);
850 load->inv_weight = prio_to_wmult[prio];
853 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
856 if (!(flags & ENQUEUE_RESTORE))
857 sched_info_queued(rq, p);
858 p->sched_class->enqueue_task(rq, p, flags);
861 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
864 if (!(flags & DEQUEUE_SAVE))
865 sched_info_dequeued(rq, p);
866 p->sched_class->dequeue_task(rq, p, flags);
869 void activate_task(struct rq *rq, struct task_struct *p, int flags)
871 if (task_contributes_to_load(p))
872 rq->nr_uninterruptible--;
874 enqueue_task(rq, p, flags);
877 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
879 if (task_contributes_to_load(p))
880 rq->nr_uninterruptible++;
882 dequeue_task(rq, p, flags);
885 static void update_rq_clock_task(struct rq *rq, s64 delta)
888 * In theory, the compile should just see 0 here, and optimize out the call
889 * to sched_rt_avg_update. But I don't trust it...
891 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
892 s64 steal = 0, irq_delta = 0;
894 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
895 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
898 * Since irq_time is only updated on {soft,}irq_exit, we might run into
899 * this case when a previous update_rq_clock() happened inside a
902 * When this happens, we stop ->clock_task and only update the
903 * prev_irq_time stamp to account for the part that fit, so that a next
904 * update will consume the rest. This ensures ->clock_task is
907 * It does however cause some slight miss-attribution of {soft,}irq
908 * time, a more accurate solution would be to update the irq_time using
909 * the current rq->clock timestamp, except that would require using
912 if (irq_delta > delta)
915 rq->prev_irq_time += irq_delta;
918 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
919 if (static_key_false((¶virt_steal_rq_enabled))) {
920 steal = paravirt_steal_clock(cpu_of(rq));
921 steal -= rq->prev_steal_time_rq;
923 if (unlikely(steal > delta))
926 rq->prev_steal_time_rq += steal;
931 rq->clock_task += delta;
933 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
934 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
935 sched_rt_avg_update(rq, irq_delta + steal);
939 void sched_set_stop_task(int cpu, struct task_struct *stop)
941 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
942 struct task_struct *old_stop = cpu_rq(cpu)->stop;
946 * Make it appear like a SCHED_FIFO task, its something
947 * userspace knows about and won't get confused about.
949 * Also, it will make PI more or less work without too
950 * much confusion -- but then, stop work should not
951 * rely on PI working anyway.
953 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
955 stop->sched_class = &stop_sched_class;
958 cpu_rq(cpu)->stop = stop;
962 * Reset it back to a normal scheduling class so that
963 * it can die in pieces.
965 old_stop->sched_class = &rt_sched_class;
970 * __normal_prio - return the priority that is based on the static prio
972 static inline int __normal_prio(struct task_struct *p)
974 return p->static_prio;
978 * Calculate the expected normal priority: i.e. priority
979 * without taking RT-inheritance into account. Might be
980 * boosted by interactivity modifiers. Changes upon fork,
981 * setprio syscalls, and whenever the interactivity
982 * estimator recalculates.
984 static inline int normal_prio(struct task_struct *p)
988 if (task_has_dl_policy(p))
989 prio = MAX_DL_PRIO-1;
990 else if (task_has_rt_policy(p))
991 prio = MAX_RT_PRIO-1 - p->rt_priority;
993 prio = __normal_prio(p);
998 * Calculate the current priority, i.e. the priority
999 * taken into account by the scheduler. This value might
1000 * be boosted by RT tasks, or might be boosted by
1001 * interactivity modifiers. Will be RT if the task got
1002 * RT-boosted. If not then it returns p->normal_prio.
1004 static int effective_prio(struct task_struct *p)
1006 p->normal_prio = normal_prio(p);
1008 * If we are RT tasks or we were boosted to RT priority,
1009 * keep the priority unchanged. Otherwise, update priority
1010 * to the normal priority:
1012 if (!rt_prio(p->prio))
1013 return p->normal_prio;
1018 * task_curr - is this task currently executing on a CPU?
1019 * @p: the task in question.
1021 * Return: 1 if the task is currently executing. 0 otherwise.
1023 inline int task_curr(const struct task_struct *p)
1025 return cpu_curr(task_cpu(p)) == p;
1029 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1030 * use the balance_callback list if you want balancing.
1032 * this means any call to check_class_changed() must be followed by a call to
1033 * balance_callback().
1035 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1036 const struct sched_class *prev_class,
1039 if (prev_class != p->sched_class) {
1040 if (prev_class->switched_from)
1041 prev_class->switched_from(rq, p);
1043 p->sched_class->switched_to(rq, p);
1044 } else if (oldprio != p->prio || dl_task(p))
1045 p->sched_class->prio_changed(rq, p, oldprio);
1048 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1050 const struct sched_class *class;
1052 if (p->sched_class == rq->curr->sched_class) {
1053 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1055 for_each_class(class) {
1056 if (class == rq->curr->sched_class)
1058 if (class == p->sched_class) {
1066 * A queue event has occurred, and we're going to schedule. In
1067 * this case, we can save a useless back to back clock update.
1069 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1070 rq_clock_skip_update(rq, true);
1075 * This is how migration works:
1077 * 1) we invoke migration_cpu_stop() on the target CPU using
1079 * 2) stopper starts to run (implicitly forcing the migrated thread
1081 * 3) it checks whether the migrated task is still in the wrong runqueue.
1082 * 4) if it's in the wrong runqueue then the migration thread removes
1083 * it and puts it into the right queue.
1084 * 5) stopper completes and stop_one_cpu() returns and the migration
1089 * move_queued_task - move a queued task to new rq.
1091 * Returns (locked) new rq. Old rq's lock is released.
1093 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1095 lockdep_assert_held(&rq->lock);
1097 dequeue_task(rq, p, 0);
1098 p->on_rq = TASK_ON_RQ_MIGRATING;
1099 double_lock_balance(rq, cpu_rq(new_cpu));
1100 set_task_cpu(p, new_cpu);
1101 double_unlock_balance(rq, cpu_rq(new_cpu));
1102 raw_spin_unlock(&rq->lock);
1104 rq = cpu_rq(new_cpu);
1106 raw_spin_lock(&rq->lock);
1107 BUG_ON(task_cpu(p) != new_cpu);
1108 p->on_rq = TASK_ON_RQ_QUEUED;
1109 enqueue_task(rq, p, 0);
1110 check_preempt_curr(rq, p, 0);
1115 struct migration_arg {
1116 struct task_struct *task;
1121 * Move (not current) task off this cpu, onto dest cpu. We're doing
1122 * this because either it can't run here any more (set_cpus_allowed()
1123 * away from this CPU, or CPU going down), or because we're
1124 * attempting to rebalance this task on exec (sched_exec).
1126 * So we race with normal scheduler movements, but that's OK, as long
1127 * as the task is no longer on this CPU.
1129 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1131 if (unlikely(!cpu_active(dest_cpu)))
1134 /* Affinity changed (again). */
1135 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1138 rq = move_queued_task(rq, p, dest_cpu);
1144 * migration_cpu_stop - this will be executed by a highprio stopper thread
1145 * and performs thread migration by bumping thread off CPU then
1146 * 'pushing' onto another runqueue.
1148 static int migration_cpu_stop(void *data)
1150 struct migration_arg *arg = data;
1151 struct task_struct *p = arg->task;
1152 struct rq *rq = this_rq();
1155 * The original target cpu might have gone down and we might
1156 * be on another cpu but it doesn't matter.
1158 local_irq_disable();
1160 * We need to explicitly wake pending tasks before running
1161 * __migrate_task() such that we will not miss enforcing cpus_allowed
1162 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1164 sched_ttwu_pending();
1166 raw_spin_lock(&p->pi_lock);
1167 raw_spin_lock(&rq->lock);
1169 * If task_rq(p) != rq, it cannot be migrated here, because we're
1170 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1171 * we're holding p->pi_lock.
1173 if (task_rq(p) == rq && task_on_rq_queued(p))
1174 rq = __migrate_task(rq, p, arg->dest_cpu);
1175 raw_spin_unlock(&rq->lock);
1176 raw_spin_unlock(&p->pi_lock);
1183 * sched_class::set_cpus_allowed must do the below, but is not required to
1184 * actually call this function.
1186 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1188 cpumask_copy(&p->cpus_allowed, new_mask);
1189 p->nr_cpus_allowed = cpumask_weight(new_mask);
1192 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1194 struct rq *rq = task_rq(p);
1195 bool queued, running;
1197 lockdep_assert_held(&p->pi_lock);
1199 queued = task_on_rq_queued(p);
1200 running = task_current(rq, p);
1204 * Because __kthread_bind() calls this on blocked tasks without
1207 lockdep_assert_held(&rq->lock);
1208 dequeue_task(rq, p, DEQUEUE_SAVE);
1211 put_prev_task(rq, p);
1213 p->sched_class->set_cpus_allowed(p, new_mask);
1216 p->sched_class->set_curr_task(rq);
1218 enqueue_task(rq, p, ENQUEUE_RESTORE);
1222 * Change a given task's CPU affinity. Migrate the thread to a
1223 * proper CPU and schedule it away if the CPU it's executing on
1224 * is removed from the allowed bitmask.
1226 * NOTE: the caller must have a valid reference to the task, the
1227 * task must not exit() & deallocate itself prematurely. The
1228 * call is not atomic; no spinlocks may be held.
1230 static int __set_cpus_allowed_ptr(struct task_struct *p,
1231 const struct cpumask *new_mask, bool check)
1233 unsigned long flags;
1235 unsigned int dest_cpu;
1238 rq = task_rq_lock(p, &flags);
1241 * Must re-check here, to close a race against __kthread_bind(),
1242 * sched_setaffinity() is not guaranteed to observe the flag.
1244 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1249 if (cpumask_equal(&p->cpus_allowed, new_mask))
1252 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1257 do_set_cpus_allowed(p, new_mask);
1259 /* Can the task run on the task's current CPU? If so, we're done */
1260 if (cpumask_test_cpu(task_cpu(p), new_mask))
1263 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1264 if (task_running(rq, p) || p->state == TASK_WAKING) {
1265 struct migration_arg arg = { p, dest_cpu };
1266 /* Need help from migration thread: drop lock and wait. */
1267 task_rq_unlock(rq, p, &flags);
1268 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1269 tlb_migrate_finish(p->mm);
1271 } else if (task_on_rq_queued(p)) {
1273 * OK, since we're going to drop the lock immediately
1274 * afterwards anyway.
1276 lockdep_unpin_lock(&rq->lock);
1277 rq = move_queued_task(rq, p, dest_cpu);
1278 lockdep_pin_lock(&rq->lock);
1281 task_rq_unlock(rq, p, &flags);
1286 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1288 return __set_cpus_allowed_ptr(p, new_mask, false);
1290 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1292 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1294 #ifdef CONFIG_SCHED_DEBUG
1296 * We should never call set_task_cpu() on a blocked task,
1297 * ttwu() will sort out the placement.
1299 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1302 #ifdef CONFIG_LOCKDEP
1304 * The caller should hold either p->pi_lock or rq->lock, when changing
1305 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1307 * sched_move_task() holds both and thus holding either pins the cgroup,
1310 * Furthermore, all task_rq users should acquire both locks, see
1313 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1314 lockdep_is_held(&task_rq(p)->lock)));
1318 trace_sched_migrate_task(p, new_cpu);
1320 if (task_cpu(p) != new_cpu) {
1321 if (p->sched_class->migrate_task_rq)
1322 p->sched_class->migrate_task_rq(p);
1323 p->se.nr_migrations++;
1324 perf_event_task_migrate(p);
1326 walt_fixup_busy_time(p, new_cpu);
1329 __set_task_cpu(p, new_cpu);
1332 static void __migrate_swap_task(struct task_struct *p, int cpu)
1334 if (task_on_rq_queued(p)) {
1335 struct rq *src_rq, *dst_rq;
1337 src_rq = task_rq(p);
1338 dst_rq = cpu_rq(cpu);
1340 deactivate_task(src_rq, p, 0);
1341 p->on_rq = TASK_ON_RQ_MIGRATING;
1342 set_task_cpu(p, cpu);
1343 p->on_rq = TASK_ON_RQ_QUEUED;
1344 activate_task(dst_rq, p, 0);
1345 check_preempt_curr(dst_rq, p, 0);
1348 * Task isn't running anymore; make it appear like we migrated
1349 * it before it went to sleep. This means on wakeup we make the
1350 * previous cpu our targer instead of where it really is.
1356 struct migration_swap_arg {
1357 struct task_struct *src_task, *dst_task;
1358 int src_cpu, dst_cpu;
1361 static int migrate_swap_stop(void *data)
1363 struct migration_swap_arg *arg = data;
1364 struct rq *src_rq, *dst_rq;
1367 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1370 src_rq = cpu_rq(arg->src_cpu);
1371 dst_rq = cpu_rq(arg->dst_cpu);
1373 double_raw_lock(&arg->src_task->pi_lock,
1374 &arg->dst_task->pi_lock);
1375 double_rq_lock(src_rq, dst_rq);
1377 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1380 if (task_cpu(arg->src_task) != arg->src_cpu)
1383 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1386 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1389 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1390 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1395 double_rq_unlock(src_rq, dst_rq);
1396 raw_spin_unlock(&arg->dst_task->pi_lock);
1397 raw_spin_unlock(&arg->src_task->pi_lock);
1403 * Cross migrate two tasks
1405 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1407 struct migration_swap_arg arg;
1410 arg = (struct migration_swap_arg){
1412 .src_cpu = task_cpu(cur),
1414 .dst_cpu = task_cpu(p),
1417 if (arg.src_cpu == arg.dst_cpu)
1421 * These three tests are all lockless; this is OK since all of them
1422 * will be re-checked with proper locks held further down the line.
1424 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1427 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1430 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1433 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1434 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1441 * wait_task_inactive - wait for a thread to unschedule.
1443 * If @match_state is nonzero, it's the @p->state value just checked and
1444 * not expected to change. If it changes, i.e. @p might have woken up,
1445 * then return zero. When we succeed in waiting for @p to be off its CPU,
1446 * we return a positive number (its total switch count). If a second call
1447 * a short while later returns the same number, the caller can be sure that
1448 * @p has remained unscheduled the whole time.
1450 * The caller must ensure that the task *will* unschedule sometime soon,
1451 * else this function might spin for a *long* time. This function can't
1452 * be called with interrupts off, or it may introduce deadlock with
1453 * smp_call_function() if an IPI is sent by the same process we are
1454 * waiting to become inactive.
1456 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1458 unsigned long flags;
1459 int running, queued;
1465 * We do the initial early heuristics without holding
1466 * any task-queue locks at all. We'll only try to get
1467 * the runqueue lock when things look like they will
1473 * If the task is actively running on another CPU
1474 * still, just relax and busy-wait without holding
1477 * NOTE! Since we don't hold any locks, it's not
1478 * even sure that "rq" stays as the right runqueue!
1479 * But we don't care, since "task_running()" will
1480 * return false if the runqueue has changed and p
1481 * is actually now running somewhere else!
1483 while (task_running(rq, p)) {
1484 if (match_state && unlikely(p->state != match_state))
1490 * Ok, time to look more closely! We need the rq
1491 * lock now, to be *sure*. If we're wrong, we'll
1492 * just go back and repeat.
1494 rq = task_rq_lock(p, &flags);
1495 trace_sched_wait_task(p);
1496 running = task_running(rq, p);
1497 queued = task_on_rq_queued(p);
1499 if (!match_state || p->state == match_state)
1500 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1501 task_rq_unlock(rq, p, &flags);
1504 * If it changed from the expected state, bail out now.
1506 if (unlikely(!ncsw))
1510 * Was it really running after all now that we
1511 * checked with the proper locks actually held?
1513 * Oops. Go back and try again..
1515 if (unlikely(running)) {
1521 * It's not enough that it's not actively running,
1522 * it must be off the runqueue _entirely_, and not
1525 * So if it was still runnable (but just not actively
1526 * running right now), it's preempted, and we should
1527 * yield - it could be a while.
1529 if (unlikely(queued)) {
1530 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1532 set_current_state(TASK_UNINTERRUPTIBLE);
1533 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1538 * Ahh, all good. It wasn't running, and it wasn't
1539 * runnable, which means that it will never become
1540 * running in the future either. We're all done!
1549 * kick_process - kick a running thread to enter/exit the kernel
1550 * @p: the to-be-kicked thread
1552 * Cause a process which is running on another CPU to enter
1553 * kernel-mode, without any delay. (to get signals handled.)
1555 * NOTE: this function doesn't have to take the runqueue lock,
1556 * because all it wants to ensure is that the remote task enters
1557 * the kernel. If the IPI races and the task has been migrated
1558 * to another CPU then no harm is done and the purpose has been
1561 void kick_process(struct task_struct *p)
1567 if ((cpu != smp_processor_id()) && task_curr(p))
1568 smp_send_reschedule(cpu);
1571 EXPORT_SYMBOL_GPL(kick_process);
1574 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1576 static int select_fallback_rq(int cpu, struct task_struct *p)
1578 int nid = cpu_to_node(cpu);
1579 const struct cpumask *nodemask = NULL;
1580 enum { cpuset, possible, fail } state = cpuset;
1584 * If the node that the cpu is on has been offlined, cpu_to_node()
1585 * will return -1. There is no cpu on the node, and we should
1586 * select the cpu on the other node.
1589 nodemask = cpumask_of_node(nid);
1591 /* Look for allowed, online CPU in same node. */
1592 for_each_cpu(dest_cpu, nodemask) {
1593 if (!cpu_online(dest_cpu))
1595 if (!cpu_active(dest_cpu))
1597 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1603 /* Any allowed, online CPU? */
1604 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1605 if (!cpu_online(dest_cpu))
1607 if (!cpu_active(dest_cpu))
1612 /* No more Mr. Nice Guy. */
1615 if (IS_ENABLED(CONFIG_CPUSETS)) {
1616 cpuset_cpus_allowed_fallback(p);
1622 do_set_cpus_allowed(p, cpu_possible_mask);
1633 if (state != cpuset) {
1635 * Don't tell them about moving exiting tasks or
1636 * kernel threads (both mm NULL), since they never
1639 if (p->mm && printk_ratelimit()) {
1640 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1641 task_pid_nr(p), p->comm, cpu);
1649 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1652 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags,
1653 int sibling_count_hint)
1655 lockdep_assert_held(&p->pi_lock);
1657 if (p->nr_cpus_allowed > 1)
1658 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags,
1659 sibling_count_hint);
1662 * In order not to call set_task_cpu() on a blocking task we need
1663 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1666 * Since this is common to all placement strategies, this lives here.
1668 * [ this allows ->select_task() to simply return task_cpu(p) and
1669 * not worry about this generic constraint ]
1671 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1673 cpu = select_fallback_rq(task_cpu(p), p);
1678 static void update_avg(u64 *avg, u64 sample)
1680 s64 diff = sample - *avg;
1686 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1687 const struct cpumask *new_mask, bool check)
1689 return set_cpus_allowed_ptr(p, new_mask);
1692 #endif /* CONFIG_SMP */
1695 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1697 #ifdef CONFIG_SCHEDSTATS
1698 struct rq *rq = this_rq();
1701 int this_cpu = smp_processor_id();
1703 if (cpu == this_cpu) {
1704 schedstat_inc(rq, ttwu_local);
1705 schedstat_inc(p, se.statistics.nr_wakeups_local);
1707 struct sched_domain *sd;
1709 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1711 for_each_domain(this_cpu, sd) {
1712 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1713 schedstat_inc(sd, ttwu_wake_remote);
1720 if (wake_flags & WF_MIGRATED)
1721 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1723 #endif /* CONFIG_SMP */
1725 schedstat_inc(rq, ttwu_count);
1726 schedstat_inc(p, se.statistics.nr_wakeups);
1728 if (wake_flags & WF_SYNC)
1729 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1731 #endif /* CONFIG_SCHEDSTATS */
1734 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1736 activate_task(rq, p, en_flags);
1737 p->on_rq = TASK_ON_RQ_QUEUED;
1739 /* if a worker is waking up, notify workqueue */
1740 if (p->flags & PF_WQ_WORKER)
1741 wq_worker_waking_up(p, cpu_of(rq));
1745 * Mark the task runnable and perform wakeup-preemption.
1748 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1750 check_preempt_curr(rq, p, wake_flags);
1751 p->state = TASK_RUNNING;
1752 trace_sched_wakeup(p);
1755 if (p->sched_class->task_woken) {
1757 * Our task @p is fully woken up and running; so its safe to
1758 * drop the rq->lock, hereafter rq is only used for statistics.
1760 lockdep_unpin_lock(&rq->lock);
1761 p->sched_class->task_woken(rq, p);
1762 lockdep_pin_lock(&rq->lock);
1765 if (rq->idle_stamp) {
1766 u64 delta = rq_clock(rq) - rq->idle_stamp;
1767 u64 max = 2*rq->max_idle_balance_cost;
1769 update_avg(&rq->avg_idle, delta);
1771 if (rq->avg_idle > max)
1780 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1782 lockdep_assert_held(&rq->lock);
1785 if (p->sched_contributes_to_load)
1786 rq->nr_uninterruptible--;
1789 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1790 ttwu_do_wakeup(rq, p, wake_flags);
1794 * Called in case the task @p isn't fully descheduled from its runqueue,
1795 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1796 * since all we need to do is flip p->state to TASK_RUNNING, since
1797 * the task is still ->on_rq.
1799 static int ttwu_remote(struct task_struct *p, int wake_flags)
1804 rq = __task_rq_lock(p);
1805 if (task_on_rq_queued(p)) {
1806 /* check_preempt_curr() may use rq clock */
1807 update_rq_clock(rq);
1808 ttwu_do_wakeup(rq, p, wake_flags);
1811 __task_rq_unlock(rq);
1817 void sched_ttwu_pending(void)
1819 struct rq *rq = this_rq();
1820 struct llist_node *llist = llist_del_all(&rq->wake_list);
1821 struct task_struct *p;
1822 unsigned long flags;
1827 raw_spin_lock_irqsave(&rq->lock, flags);
1828 lockdep_pin_lock(&rq->lock);
1831 p = llist_entry(llist, struct task_struct, wake_entry);
1832 llist = llist_next(llist);
1833 ttwu_do_activate(rq, p, 0);
1836 lockdep_unpin_lock(&rq->lock);
1837 raw_spin_unlock_irqrestore(&rq->lock, flags);
1840 void scheduler_ipi(void)
1843 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1844 * TIF_NEED_RESCHED remotely (for the first time) will also send
1847 preempt_fold_need_resched();
1849 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1853 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1854 * traditionally all their work was done from the interrupt return
1855 * path. Now that we actually do some work, we need to make sure
1858 * Some archs already do call them, luckily irq_enter/exit nest
1861 * Arguably we should visit all archs and update all handlers,
1862 * however a fair share of IPIs are still resched only so this would
1863 * somewhat pessimize the simple resched case.
1866 sched_ttwu_pending();
1869 * Check if someone kicked us for doing the nohz idle load balance.
1871 if (unlikely(got_nohz_idle_kick())) {
1872 this_rq()->idle_balance = 1;
1873 raise_softirq_irqoff(SCHED_SOFTIRQ);
1878 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1880 struct rq *rq = cpu_rq(cpu);
1882 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1883 if (!set_nr_if_polling(rq->idle))
1884 smp_send_reschedule(cpu);
1886 trace_sched_wake_idle_without_ipi(cpu);
1890 void wake_up_if_idle(int cpu)
1892 struct rq *rq = cpu_rq(cpu);
1893 unsigned long flags;
1897 if (!is_idle_task(rcu_dereference(rq->curr)))
1900 if (set_nr_if_polling(rq->idle)) {
1901 trace_sched_wake_idle_without_ipi(cpu);
1903 raw_spin_lock_irqsave(&rq->lock, flags);
1904 if (is_idle_task(rq->curr))
1905 smp_send_reschedule(cpu);
1906 /* Else cpu is not in idle, do nothing here */
1907 raw_spin_unlock_irqrestore(&rq->lock, flags);
1914 bool cpus_share_cache(int this_cpu, int that_cpu)
1916 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1918 #endif /* CONFIG_SMP */
1920 static void ttwu_queue(struct task_struct *p, int cpu)
1922 struct rq *rq = cpu_rq(cpu);
1924 #if defined(CONFIG_SMP)
1925 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1926 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1927 ttwu_queue_remote(p, cpu);
1932 raw_spin_lock(&rq->lock);
1933 lockdep_pin_lock(&rq->lock);
1934 ttwu_do_activate(rq, p, 0);
1935 lockdep_unpin_lock(&rq->lock);
1936 raw_spin_unlock(&rq->lock);
1940 * try_to_wake_up - wake up a thread
1941 * @p: the thread to be awakened
1942 * @state: the mask of task states that can be woken
1943 * @wake_flags: wake modifier flags (WF_*)
1944 * @sibling_count_hint: A hint at the number of threads that are being woken up
1947 * Put it on the run-queue if it's not already there. The "current"
1948 * thread is always on the run-queue (except when the actual
1949 * re-schedule is in progress), and as such you're allowed to do
1950 * the simpler "current->state = TASK_RUNNING" to mark yourself
1951 * runnable without the overhead of this.
1953 * Return: %true if @p was woken up, %false if it was already running.
1954 * or @state didn't match @p's state.
1957 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
1958 int sibling_count_hint)
1960 unsigned long flags;
1961 int cpu, success = 0;
1968 * If we are going to wake up a thread waiting for CONDITION we
1969 * need to ensure that CONDITION=1 done by the caller can not be
1970 * reordered with p->state check below. This pairs with mb() in
1971 * set_current_state() the waiting thread does.
1973 smp_mb__before_spinlock();
1974 raw_spin_lock_irqsave(&p->pi_lock, flags);
1975 if (!(p->state & state))
1978 trace_sched_waking(p);
1980 success = 1; /* we're going to change ->state */
1984 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1985 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1986 * in smp_cond_load_acquire() below.
1988 * sched_ttwu_pending() try_to_wake_up()
1989 * [S] p->on_rq = 1; [L] P->state
1990 * UNLOCK rq->lock -----.
1994 * LOCK rq->lock -----'
1998 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
2000 * Pairs with the UNLOCK+LOCK on rq->lock from the
2001 * last wakeup of our task and the schedule that got our task
2005 if (p->on_rq && ttwu_remote(p, wake_flags))
2010 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2011 * possible to, falsely, observe p->on_cpu == 0.
2013 * One must be running (->on_cpu == 1) in order to remove oneself
2014 * from the runqueue.
2016 * [S] ->on_cpu = 1; [L] ->on_rq
2020 * [S] ->on_rq = 0; [L] ->on_cpu
2022 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2023 * from the consecutive calls to schedule(); the first switching to our
2024 * task, the second putting it to sleep.
2029 * If the owning (remote) cpu is still in the middle of schedule() with
2030 * this task as prev, wait until its done referencing the task.
2035 * Combined with the control dependency above, we have an effective
2036 * smp_load_acquire() without the need for full barriers.
2038 * Pairs with the smp_store_release() in finish_lock_switch().
2040 * This ensures that tasks getting woken will be fully ordered against
2041 * their previous state and preserve Program Order.
2045 rq = cpu_rq(task_cpu(p));
2047 raw_spin_lock(&rq->lock);
2048 wallclock = walt_ktime_clock();
2049 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2050 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2051 raw_spin_unlock(&rq->lock);
2053 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2054 p->state = TASK_WAKING;
2056 if (p->sched_class->task_waking)
2057 p->sched_class->task_waking(p);
2059 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags,
2060 sibling_count_hint);
2061 if (task_cpu(p) != cpu) {
2062 wake_flags |= WF_MIGRATED;
2063 set_task_cpu(p, cpu);
2066 #endif /* CONFIG_SMP */
2070 ttwu_stat(p, cpu, wake_flags);
2072 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2078 * try_to_wake_up_local - try to wake up a local task with rq lock held
2079 * @p: the thread to be awakened
2081 * Put @p on the run-queue if it's not already there. The caller must
2082 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2085 static void try_to_wake_up_local(struct task_struct *p)
2087 struct rq *rq = task_rq(p);
2089 if (WARN_ON_ONCE(rq != this_rq()) ||
2090 WARN_ON_ONCE(p == current))
2093 lockdep_assert_held(&rq->lock);
2095 if (!raw_spin_trylock(&p->pi_lock)) {
2097 * This is OK, because current is on_cpu, which avoids it being
2098 * picked for load-balance and preemption/IRQs are still
2099 * disabled avoiding further scheduler activity on it and we've
2100 * not yet picked a replacement task.
2102 lockdep_unpin_lock(&rq->lock);
2103 raw_spin_unlock(&rq->lock);
2104 raw_spin_lock(&p->pi_lock);
2105 raw_spin_lock(&rq->lock);
2106 lockdep_pin_lock(&rq->lock);
2109 if (!(p->state & TASK_NORMAL))
2112 trace_sched_waking(p);
2114 if (!task_on_rq_queued(p)) {
2115 u64 wallclock = walt_ktime_clock();
2117 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2118 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2119 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2122 ttwu_do_wakeup(rq, p, 0);
2123 ttwu_stat(p, smp_processor_id(), 0);
2125 raw_spin_unlock(&p->pi_lock);
2129 * wake_up_process - Wake up a specific process
2130 * @p: The process to be woken up.
2132 * Attempt to wake up the nominated process and move it to the set of runnable
2135 * Return: 1 if the process was woken up, 0 if it was already running.
2137 * It may be assumed that this function implies a write memory barrier before
2138 * changing the task state if and only if any tasks are woken up.
2140 int wake_up_process(struct task_struct *p)
2142 return try_to_wake_up(p, TASK_NORMAL, 0, 1);
2144 EXPORT_SYMBOL(wake_up_process);
2146 int wake_up_state(struct task_struct *p, unsigned int state)
2148 return try_to_wake_up(p, state, 0, 1);
2152 * This function clears the sched_dl_entity static params.
2154 void __dl_clear_params(struct task_struct *p)
2156 struct sched_dl_entity *dl_se = &p->dl;
2158 dl_se->dl_runtime = 0;
2159 dl_se->dl_deadline = 0;
2160 dl_se->dl_period = 0;
2163 dl_se->dl_density = 0;
2165 dl_se->dl_throttled = 0;
2167 dl_se->dl_yielded = 0;
2171 * Perform scheduler related setup for a newly forked process p.
2172 * p is forked by current.
2174 * __sched_fork() is basic setup used by init_idle() too:
2176 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2181 p->se.exec_start = 0;
2182 p->se.sum_exec_runtime = 0;
2183 p->se.prev_sum_exec_runtime = 0;
2184 p->se.nr_migrations = 0;
2186 #ifdef CONFIG_SCHED_WALT
2187 p->last_sleep_ts = 0;
2190 INIT_LIST_HEAD(&p->se.group_node);
2191 walt_init_new_task_load(p);
2193 #ifdef CONFIG_FAIR_GROUP_SCHED
2194 p->se.cfs_rq = NULL;
2197 #ifdef CONFIG_SCHEDSTATS
2198 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2201 RB_CLEAR_NODE(&p->dl.rb_node);
2202 init_dl_task_timer(&p->dl);
2203 __dl_clear_params(p);
2205 init_rt_schedtune_timer(&p->rt);
2206 INIT_LIST_HEAD(&p->rt.run_list);
2208 #ifdef CONFIG_PREEMPT_NOTIFIERS
2209 INIT_HLIST_HEAD(&p->preempt_notifiers);
2212 #ifdef CONFIG_NUMA_BALANCING
2213 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2214 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2215 p->mm->numa_scan_seq = 0;
2218 if (clone_flags & CLONE_VM)
2219 p->numa_preferred_nid = current->numa_preferred_nid;
2221 p->numa_preferred_nid = -1;
2223 p->node_stamp = 0ULL;
2224 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2225 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2226 p->numa_work.next = &p->numa_work;
2227 p->numa_faults = NULL;
2228 p->last_task_numa_placement = 0;
2229 p->last_sum_exec_runtime = 0;
2231 p->numa_group = NULL;
2232 #endif /* CONFIG_NUMA_BALANCING */
2235 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2237 #ifdef CONFIG_NUMA_BALANCING
2239 void set_numabalancing_state(bool enabled)
2242 static_branch_enable(&sched_numa_balancing);
2244 static_branch_disable(&sched_numa_balancing);
2247 #ifdef CONFIG_PROC_SYSCTL
2248 int sysctl_numa_balancing(struct ctl_table *table, int write,
2249 void __user *buffer, size_t *lenp, loff_t *ppos)
2253 int state = static_branch_likely(&sched_numa_balancing);
2255 if (write && !capable(CAP_SYS_ADMIN))
2260 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2264 set_numabalancing_state(state);
2271 * fork()/clone()-time setup:
2273 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2275 unsigned long flags;
2276 int cpu = get_cpu();
2278 __sched_fork(clone_flags, p);
2280 * We mark the process as NEW here. This guarantees that
2281 * nobody will actually run it, and a signal or other external
2282 * event cannot wake it up and insert it on the runqueue either.
2284 p->state = TASK_NEW;
2287 * Make sure we do not leak PI boosting priority to the child.
2289 p->prio = current->normal_prio;
2292 * Revert to default priority/policy on fork if requested.
2294 if (unlikely(p->sched_reset_on_fork)) {
2295 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2296 p->policy = SCHED_NORMAL;
2297 p->static_prio = NICE_TO_PRIO(0);
2299 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2300 p->static_prio = NICE_TO_PRIO(0);
2302 p->prio = p->normal_prio = __normal_prio(p);
2306 * We don't need the reset flag anymore after the fork. It has
2307 * fulfilled its duty:
2309 p->sched_reset_on_fork = 0;
2312 if (dl_prio(p->prio)) {
2315 } else if (rt_prio(p->prio)) {
2316 p->sched_class = &rt_sched_class;
2318 p->sched_class = &fair_sched_class;
2321 init_entity_runnable_average(&p->se);
2324 * The child is not yet in the pid-hash so no cgroup attach races,
2325 * and the cgroup is pinned to this child due to cgroup_fork()
2326 * is ran before sched_fork().
2328 * Silence PROVE_RCU.
2330 raw_spin_lock_irqsave(&p->pi_lock, flags);
2332 * We're setting the cpu for the first time, we don't migrate,
2333 * so use __set_task_cpu().
2335 __set_task_cpu(p, cpu);
2336 if (p->sched_class->task_fork)
2337 p->sched_class->task_fork(p);
2338 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2340 #ifdef CONFIG_SCHED_INFO
2341 if (likely(sched_info_on()))
2342 memset(&p->sched_info, 0, sizeof(p->sched_info));
2344 #if defined(CONFIG_SMP)
2347 init_task_preempt_count(p);
2349 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2350 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2357 unsigned long to_ratio(u64 period, u64 runtime)
2359 if (runtime == RUNTIME_INF)
2363 * Doing this here saves a lot of checks in all
2364 * the calling paths, and returning zero seems
2365 * safe for them anyway.
2370 return div64_u64(runtime << 20, period);
2374 inline struct dl_bw *dl_bw_of(int i)
2376 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2377 "sched RCU must be held");
2378 return &cpu_rq(i)->rd->dl_bw;
2381 static inline int dl_bw_cpus(int i)
2383 struct root_domain *rd = cpu_rq(i)->rd;
2386 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2387 "sched RCU must be held");
2388 for_each_cpu_and(i, rd->span, cpu_active_mask)
2394 inline struct dl_bw *dl_bw_of(int i)
2396 return &cpu_rq(i)->dl.dl_bw;
2399 static inline int dl_bw_cpus(int i)
2406 * We must be sure that accepting a new task (or allowing changing the
2407 * parameters of an existing one) is consistent with the bandwidth
2408 * constraints. If yes, this function also accordingly updates the currently
2409 * allocated bandwidth to reflect the new situation.
2411 * This function is called while holding p's rq->lock.
2413 * XXX we should delay bw change until the task's 0-lag point, see
2416 static int dl_overflow(struct task_struct *p, int policy,
2417 const struct sched_attr *attr)
2420 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2421 u64 period = attr->sched_period ?: attr->sched_deadline;
2422 u64 runtime = attr->sched_runtime;
2423 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2426 if (new_bw == p->dl.dl_bw)
2430 * Either if a task, enters, leave, or stays -deadline but changes
2431 * its parameters, we may need to update accordingly the total
2432 * allocated bandwidth of the container.
2434 raw_spin_lock(&dl_b->lock);
2435 cpus = dl_bw_cpus(task_cpu(p));
2436 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2437 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2438 __dl_add(dl_b, new_bw);
2440 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2441 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2442 __dl_clear(dl_b, p->dl.dl_bw);
2443 __dl_add(dl_b, new_bw);
2445 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2446 __dl_clear(dl_b, p->dl.dl_bw);
2449 raw_spin_unlock(&dl_b->lock);
2454 extern void init_dl_bw(struct dl_bw *dl_b);
2457 * wake_up_new_task - wake up a newly created task for the first time.
2459 * This function will do some initial scheduler statistics housekeeping
2460 * that must be done for every newly created context, then puts the task
2461 * on the runqueue and wakes it.
2463 void wake_up_new_task(struct task_struct *p)
2465 unsigned long flags;
2468 raw_spin_lock_irqsave(&p->pi_lock, flags);
2469 p->state = TASK_RUNNING;
2471 walt_init_new_task_load(p);
2473 /* Initialize new task's runnable average */
2474 init_entity_runnable_average(&p->se);
2477 * Fork balancing, do it here and not earlier because:
2478 * - cpus_allowed can change in the fork path
2479 * - any previously selected cpu might disappear through hotplug
2481 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2482 * as we're not fully set-up yet.
2484 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0, 1));
2486 rq = __task_rq_lock(p);
2487 update_rq_clock(rq);
2488 post_init_entity_util_avg(&p->se);
2490 walt_mark_task_starting(p);
2491 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2492 p->on_rq = TASK_ON_RQ_QUEUED;
2493 trace_sched_wakeup_new(p);
2494 check_preempt_curr(rq, p, WF_FORK);
2496 if (p->sched_class->task_woken) {
2498 * Nothing relies on rq->lock after this, so its fine to
2501 lockdep_unpin_lock(&rq->lock);
2502 p->sched_class->task_woken(rq, p);
2503 lockdep_pin_lock(&rq->lock);
2506 task_rq_unlock(rq, p, &flags);
2509 #ifdef CONFIG_PREEMPT_NOTIFIERS
2511 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2513 void preempt_notifier_inc(void)
2515 static_key_slow_inc(&preempt_notifier_key);
2517 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2519 void preempt_notifier_dec(void)
2521 static_key_slow_dec(&preempt_notifier_key);
2523 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2526 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2527 * @notifier: notifier struct to register
2529 void preempt_notifier_register(struct preempt_notifier *notifier)
2531 if (!static_key_false(&preempt_notifier_key))
2532 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2534 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2536 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2539 * preempt_notifier_unregister - no longer interested in preemption notifications
2540 * @notifier: notifier struct to unregister
2542 * This is *not* safe to call from within a preemption notifier.
2544 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2546 hlist_del(¬ifier->link);
2548 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2550 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2552 struct preempt_notifier *notifier;
2554 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2555 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2558 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2560 if (static_key_false(&preempt_notifier_key))
2561 __fire_sched_in_preempt_notifiers(curr);
2565 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2566 struct task_struct *next)
2568 struct preempt_notifier *notifier;
2570 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2571 notifier->ops->sched_out(notifier, next);
2574 static __always_inline void
2575 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2576 struct task_struct *next)
2578 if (static_key_false(&preempt_notifier_key))
2579 __fire_sched_out_preempt_notifiers(curr, next);
2582 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2584 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2589 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2590 struct task_struct *next)
2594 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2597 * prepare_task_switch - prepare to switch tasks
2598 * @rq: the runqueue preparing to switch
2599 * @prev: the current task that is being switched out
2600 * @next: the task we are going to switch to.
2602 * This is called with the rq lock held and interrupts off. It must
2603 * be paired with a subsequent finish_task_switch after the context
2606 * prepare_task_switch sets up locking and calls architecture specific
2610 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2611 struct task_struct *next)
2613 sched_info_switch(rq, prev, next);
2614 perf_event_task_sched_out(prev, next);
2615 fire_sched_out_preempt_notifiers(prev, next);
2616 prepare_lock_switch(rq, next);
2617 prepare_arch_switch(next);
2621 * finish_task_switch - clean up after a task-switch
2622 * @prev: the thread we just switched away from.
2624 * finish_task_switch must be called after the context switch, paired
2625 * with a prepare_task_switch call before the context switch.
2626 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2627 * and do any other architecture-specific cleanup actions.
2629 * Note that we may have delayed dropping an mm in context_switch(). If
2630 * so, we finish that here outside of the runqueue lock. (Doing it
2631 * with the lock held can cause deadlocks; see schedule() for
2634 * The context switch have flipped the stack from under us and restored the
2635 * local variables which were saved when this task called schedule() in the
2636 * past. prev == current is still correct but we need to recalculate this_rq
2637 * because prev may have moved to another CPU.
2639 static struct rq *finish_task_switch(struct task_struct *prev)
2640 __releases(rq->lock)
2642 struct rq *rq = this_rq();
2643 struct mm_struct *mm = rq->prev_mm;
2647 * The previous task will have left us with a preempt_count of 2
2648 * because it left us after:
2651 * preempt_disable(); // 1
2653 * raw_spin_lock_irq(&rq->lock) // 2
2655 * Also, see FORK_PREEMPT_COUNT.
2657 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2658 "corrupted preempt_count: %s/%d/0x%x\n",
2659 current->comm, current->pid, preempt_count()))
2660 preempt_count_set(FORK_PREEMPT_COUNT);
2665 * A task struct has one reference for the use as "current".
2666 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2667 * schedule one last time. The schedule call will never return, and
2668 * the scheduled task must drop that reference.
2670 * We must observe prev->state before clearing prev->on_cpu (in
2671 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2672 * running on another CPU and we could rave with its RUNNING -> DEAD
2673 * transition, resulting in a double drop.
2675 prev_state = prev->state;
2676 vtime_task_switch(prev);
2677 perf_event_task_sched_in(prev, current);
2678 finish_lock_switch(rq, prev);
2679 finish_arch_post_lock_switch();
2681 fire_sched_in_preempt_notifiers(current);
2684 if (unlikely(prev_state == TASK_DEAD)) {
2685 if (prev->sched_class->task_dead)
2686 prev->sched_class->task_dead(prev);
2689 * Remove function-return probe instances associated with this
2690 * task and put them back on the free list.
2692 kprobe_flush_task(prev);
2693 put_task_struct(prev);
2696 tick_nohz_task_switch();
2702 /* rq->lock is NOT held, but preemption is disabled */
2703 static void __balance_callback(struct rq *rq)
2705 struct callback_head *head, *next;
2706 void (*func)(struct rq *rq);
2707 unsigned long flags;
2709 raw_spin_lock_irqsave(&rq->lock, flags);
2710 head = rq->balance_callback;
2711 rq->balance_callback = NULL;
2713 func = (void (*)(struct rq *))head->func;
2720 raw_spin_unlock_irqrestore(&rq->lock, flags);
2723 static inline void balance_callback(struct rq *rq)
2725 if (unlikely(rq->balance_callback))
2726 __balance_callback(rq);
2731 static inline void balance_callback(struct rq *rq)
2738 * schedule_tail - first thing a freshly forked thread must call.
2739 * @prev: the thread we just switched away from.
2741 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2742 __releases(rq->lock)
2747 * New tasks start with FORK_PREEMPT_COUNT, see there and
2748 * finish_task_switch() for details.
2750 * finish_task_switch() will drop rq->lock() and lower preempt_count
2751 * and the preempt_enable() will end up enabling preemption (on
2752 * PREEMPT_COUNT kernels).
2755 rq = finish_task_switch(prev);
2756 balance_callback(rq);
2759 if (current->set_child_tid)
2760 put_user(task_pid_vnr(current), current->set_child_tid);
2764 * context_switch - switch to the new MM and the new thread's register state.
2766 static inline struct rq *
2767 context_switch(struct rq *rq, struct task_struct *prev,
2768 struct task_struct *next)
2770 struct mm_struct *mm, *oldmm;
2772 prepare_task_switch(rq, prev, next);
2775 oldmm = prev->active_mm;
2777 * For paravirt, this is coupled with an exit in switch_to to
2778 * combine the page table reload and the switch backend into
2781 arch_start_context_switch(prev);
2784 next->active_mm = oldmm;
2785 atomic_inc(&oldmm->mm_count);
2786 enter_lazy_tlb(oldmm, next);
2788 switch_mm_irqs_off(oldmm, mm, next);
2791 prev->active_mm = NULL;
2792 rq->prev_mm = oldmm;
2795 * Since the runqueue lock will be released by the next
2796 * task (which is an invalid locking op but in the case
2797 * of the scheduler it's an obvious special-case), so we
2798 * do an early lockdep release here:
2800 lockdep_unpin_lock(&rq->lock);
2801 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2803 /* Here we just switch the register state and the stack. */
2804 switch_to(prev, next, prev);
2807 return finish_task_switch(prev);
2811 * nr_running and nr_context_switches:
2813 * externally visible scheduler statistics: current number of runnable
2814 * threads, total number of context switches performed since bootup.
2816 unsigned long nr_running(void)
2818 unsigned long i, sum = 0;
2820 for_each_online_cpu(i)
2821 sum += cpu_rq(i)->nr_running;
2827 * Check if only the current task is running on the cpu.
2829 * Caution: this function does not check that the caller has disabled
2830 * preemption, thus the result might have a time-of-check-to-time-of-use
2831 * race. The caller is responsible to use it correctly, for example:
2833 * - from a non-preemptable section (of course)
2835 * - from a thread that is bound to a single CPU
2837 * - in a loop with very short iterations (e.g. a polling loop)
2839 bool single_task_running(void)
2841 return raw_rq()->nr_running == 1;
2843 EXPORT_SYMBOL(single_task_running);
2845 unsigned long long nr_context_switches(void)
2848 unsigned long long sum = 0;
2850 for_each_possible_cpu(i)
2851 sum += cpu_rq(i)->nr_switches;
2856 unsigned long nr_iowait(void)
2858 unsigned long i, sum = 0;
2860 for_each_possible_cpu(i)
2861 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2866 unsigned long nr_iowait_cpu(int cpu)
2868 struct rq *this = cpu_rq(cpu);
2869 return atomic_read(&this->nr_iowait);
2872 #ifdef CONFIG_CPU_QUIET
2873 u64 nr_running_integral(unsigned int cpu)
2875 unsigned int seqcnt;
2879 if (cpu >= nr_cpu_ids)
2885 * Update average to avoid reading stalled value if there were
2886 * no run-queue changes for a long time. On the other hand if
2887 * the changes are happening right now, just read current value
2891 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2892 integral = do_nr_running_integral(q);
2893 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2894 read_seqcount_begin(&q->ave_seqcnt);
2895 integral = q->nr_running_integral;
2902 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2904 struct rq *rq = this_rq();
2905 *nr_waiters = atomic_read(&rq->nr_iowait);
2906 *load = rq->load.weight;
2912 * sched_exec - execve() is a valuable balancing opportunity, because at
2913 * this point the task has the smallest effective memory and cache footprint.
2915 void sched_exec(void)
2917 struct task_struct *p = current;
2918 unsigned long flags;
2921 raw_spin_lock_irqsave(&p->pi_lock, flags);
2922 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0, 1);
2923 if (dest_cpu == smp_processor_id())
2926 if (likely(cpu_active(dest_cpu))) {
2927 struct migration_arg arg = { p, dest_cpu };
2929 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2930 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2934 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2939 DEFINE_PER_CPU(struct kernel_stat, kstat);
2940 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2942 EXPORT_PER_CPU_SYMBOL(kstat);
2943 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2946 * Return accounted runtime for the task.
2947 * In case the task is currently running, return the runtime plus current's
2948 * pending runtime that have not been accounted yet.
2950 unsigned long long task_sched_runtime(struct task_struct *p)
2952 unsigned long flags;
2956 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2958 * 64-bit doesn't need locks to atomically read a 64bit value.
2959 * So we have a optimization chance when the task's delta_exec is 0.
2960 * Reading ->on_cpu is racy, but this is ok.
2962 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2963 * If we race with it entering cpu, unaccounted time is 0. This is
2964 * indistinguishable from the read occurring a few cycles earlier.
2965 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2966 * been accounted, so we're correct here as well.
2968 if (!p->on_cpu || !task_on_rq_queued(p))
2969 return p->se.sum_exec_runtime;
2972 rq = task_rq_lock(p, &flags);
2974 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2975 * project cycles that may never be accounted to this
2976 * thread, breaking clock_gettime().
2978 if (task_current(rq, p) && task_on_rq_queued(p)) {
2979 update_rq_clock(rq);
2980 p->sched_class->update_curr(rq);
2982 ns = p->se.sum_exec_runtime;
2983 task_rq_unlock(rq, p, &flags);
2989 * This function gets called by the timer code, with HZ frequency.
2990 * We call it with interrupts disabled.
2992 void scheduler_tick(void)
2994 int cpu = smp_processor_id();
2995 struct rq *rq = cpu_rq(cpu);
2996 struct task_struct *curr = rq->curr;
3000 raw_spin_lock(&rq->lock);
3001 walt_set_window_start(rq);
3002 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3003 walt_ktime_clock(), 0);
3004 update_rq_clock(rq);
3005 curr->sched_class->task_tick(rq, curr, 0);
3006 update_cpu_load_active(rq);
3007 calc_global_load_tick(rq);
3008 raw_spin_unlock(&rq->lock);
3010 perf_event_task_tick();
3013 rq->idle_balance = idle_cpu(cpu);
3014 trigger_load_balance(rq);
3016 rq_last_tick_reset(rq);
3018 if (curr->sched_class == &fair_sched_class)
3019 check_for_migration(rq, curr);
3022 #ifdef CONFIG_NO_HZ_FULL
3024 * scheduler_tick_max_deferment
3026 * Keep at least one tick per second when a single
3027 * active task is running because the scheduler doesn't
3028 * yet completely support full dynticks environment.
3030 * This makes sure that uptime, CFS vruntime, load
3031 * balancing, etc... continue to move forward, even
3032 * with a very low granularity.
3034 * Return: Maximum deferment in nanoseconds.
3036 u64 scheduler_tick_max_deferment(void)
3038 struct rq *rq = this_rq();
3039 unsigned long next, now = READ_ONCE(jiffies);
3041 next = rq->last_sched_tick + HZ;
3043 if (time_before_eq(next, now))
3046 return jiffies_to_nsecs(next - now);
3050 notrace unsigned long get_parent_ip(unsigned long addr)
3052 if (in_lock_functions(addr)) {
3053 addr = CALLER_ADDR2;
3054 if (in_lock_functions(addr))
3055 addr = CALLER_ADDR3;
3060 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3061 defined(CONFIG_PREEMPT_TRACER))
3063 void preempt_count_add(int val)
3065 #ifdef CONFIG_DEBUG_PREEMPT
3069 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3072 __preempt_count_add(val);
3073 #ifdef CONFIG_DEBUG_PREEMPT
3075 * Spinlock count overflowing soon?
3077 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3080 if (preempt_count() == val) {
3081 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3082 #ifdef CONFIG_DEBUG_PREEMPT
3083 current->preempt_disable_ip = ip;
3085 trace_preempt_off(CALLER_ADDR0, ip);
3088 EXPORT_SYMBOL(preempt_count_add);
3089 NOKPROBE_SYMBOL(preempt_count_add);
3091 void preempt_count_sub(int val)
3093 #ifdef CONFIG_DEBUG_PREEMPT
3097 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3100 * Is the spinlock portion underflowing?
3102 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3103 !(preempt_count() & PREEMPT_MASK)))
3107 if (preempt_count() == val)
3108 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3109 __preempt_count_sub(val);
3111 EXPORT_SYMBOL(preempt_count_sub);
3112 NOKPROBE_SYMBOL(preempt_count_sub);
3117 * Print scheduling while atomic bug:
3119 static noinline void __schedule_bug(struct task_struct *prev)
3121 if (oops_in_progress)
3124 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3125 prev->comm, prev->pid, preempt_count());
3127 debug_show_held_locks(prev);
3129 if (irqs_disabled())
3130 print_irqtrace_events(prev);
3131 #ifdef CONFIG_DEBUG_PREEMPT
3132 if (in_atomic_preempt_off()) {
3133 pr_err("Preemption disabled at:");
3134 print_ip_sym(current->preempt_disable_ip);
3139 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3143 * Various schedule()-time debugging checks and statistics:
3145 static inline void schedule_debug(struct task_struct *prev)
3147 #ifdef CONFIG_SCHED_STACK_END_CHECK
3148 if (task_stack_end_corrupted(prev))
3149 panic("corrupted stack end detected inside scheduler\n");
3152 if (unlikely(in_atomic_preempt_off())) {
3153 __schedule_bug(prev);
3154 preempt_count_set(PREEMPT_DISABLED);
3158 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3160 schedstat_inc(this_rq(), sched_count);
3164 * Pick up the highest-prio task:
3166 static inline struct task_struct *
3167 pick_next_task(struct rq *rq, struct task_struct *prev)
3169 const struct sched_class *class = &fair_sched_class;
3170 struct task_struct *p;
3173 * Optimization: we know that if all tasks are in
3174 * the fair class we can call that function directly:
3176 if (likely(prev->sched_class == class &&
3177 rq->nr_running == rq->cfs.h_nr_running)) {
3178 p = fair_sched_class.pick_next_task(rq, prev);
3179 if (unlikely(p == RETRY_TASK))
3182 /* assumes fair_sched_class->next == idle_sched_class */
3184 p = idle_sched_class.pick_next_task(rq, prev);
3190 for_each_class(class) {
3191 p = class->pick_next_task(rq, prev);
3193 if (unlikely(p == RETRY_TASK))
3199 BUG(); /* the idle class will always have a runnable task */
3203 * __schedule() is the main scheduler function.
3205 * The main means of driving the scheduler and thus entering this function are:
3207 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3209 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3210 * paths. For example, see arch/x86/entry_64.S.
3212 * To drive preemption between tasks, the scheduler sets the flag in timer
3213 * interrupt handler scheduler_tick().
3215 * 3. Wakeups don't really cause entry into schedule(). They add a
3216 * task to the run-queue and that's it.
3218 * Now, if the new task added to the run-queue preempts the current
3219 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3220 * called on the nearest possible occasion:
3222 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3224 * - in syscall or exception context, at the next outmost
3225 * preempt_enable(). (this might be as soon as the wake_up()'s
3228 * - in IRQ context, return from interrupt-handler to
3229 * preemptible context
3231 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3234 * - cond_resched() call
3235 * - explicit schedule() call
3236 * - return from syscall or exception to user-space
3237 * - return from interrupt-handler to user-space
3239 * WARNING: must be called with preemption disabled!
3241 static void __sched notrace __schedule(bool preempt)
3243 struct task_struct *prev, *next;
3244 unsigned long *switch_count;
3249 cpu = smp_processor_id();
3251 rcu_note_context_switch();
3255 * do_exit() calls schedule() with preemption disabled as an exception;
3256 * however we must fix that up, otherwise the next task will see an
3257 * inconsistent (higher) preempt count.
3259 * It also avoids the below schedule_debug() test from complaining
3262 if (unlikely(prev->state == TASK_DEAD))
3263 preempt_enable_no_resched_notrace();
3265 schedule_debug(prev);
3267 if (sched_feat(HRTICK))
3271 * Make sure that signal_pending_state()->signal_pending() below
3272 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3273 * done by the caller to avoid the race with signal_wake_up().
3275 smp_mb__before_spinlock();
3276 raw_spin_lock_irq(&rq->lock);
3277 lockdep_pin_lock(&rq->lock);
3279 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3281 switch_count = &prev->nivcsw;
3282 if (!preempt && prev->state) {
3283 if (unlikely(signal_pending_state(prev->state, prev))) {
3284 prev->state = TASK_RUNNING;
3286 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3290 * If a worker went to sleep, notify and ask workqueue
3291 * whether it wants to wake up a task to maintain
3294 if (prev->flags & PF_WQ_WORKER) {
3295 struct task_struct *to_wakeup;
3297 to_wakeup = wq_worker_sleeping(prev, cpu);
3299 try_to_wake_up_local(to_wakeup);
3302 switch_count = &prev->nvcsw;
3305 if (task_on_rq_queued(prev))
3306 update_rq_clock(rq);
3308 next = pick_next_task(rq, prev);
3309 wallclock = walt_ktime_clock();
3310 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3311 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3312 clear_tsk_need_resched(prev);
3313 clear_preempt_need_resched();
3314 rq->clock_skip_update = 0;
3316 if (likely(prev != next)) {
3317 #ifdef CONFIG_SCHED_WALT
3319 prev->last_sleep_ts = wallclock;
3325 trace_sched_switch(preempt, prev, next);
3326 rq = context_switch(rq, prev, next); /* unlocks the rq */
3329 lockdep_unpin_lock(&rq->lock);
3330 raw_spin_unlock_irq(&rq->lock);
3333 balance_callback(rq);
3336 static inline void sched_submit_work(struct task_struct *tsk)
3338 if (!tsk->state || tsk_is_pi_blocked(tsk))
3341 * If we are going to sleep and we have plugged IO queued,
3342 * make sure to submit it to avoid deadlocks.
3344 if (blk_needs_flush_plug(tsk))
3345 blk_schedule_flush_plug(tsk);
3348 asmlinkage __visible void __sched schedule(void)
3350 struct task_struct *tsk = current;
3352 sched_submit_work(tsk);
3356 sched_preempt_enable_no_resched();
3357 } while (need_resched());
3359 EXPORT_SYMBOL(schedule);
3361 #ifdef CONFIG_CONTEXT_TRACKING
3362 asmlinkage __visible void __sched schedule_user(void)
3365 * If we come here after a random call to set_need_resched(),
3366 * or we have been woken up remotely but the IPI has not yet arrived,
3367 * we haven't yet exited the RCU idle mode. Do it here manually until
3368 * we find a better solution.
3370 * NB: There are buggy callers of this function. Ideally we
3371 * should warn if prev_state != CONTEXT_USER, but that will trigger
3372 * too frequently to make sense yet.
3374 enum ctx_state prev_state = exception_enter();
3376 exception_exit(prev_state);
3381 * schedule_preempt_disabled - called with preemption disabled
3383 * Returns with preemption disabled. Note: preempt_count must be 1
3385 void __sched schedule_preempt_disabled(void)
3387 sched_preempt_enable_no_resched();
3392 static void __sched notrace preempt_schedule_common(void)
3395 preempt_disable_notrace();
3397 preempt_enable_no_resched_notrace();
3400 * Check again in case we missed a preemption opportunity
3401 * between schedule and now.
3403 } while (need_resched());
3406 #ifdef CONFIG_PREEMPT
3408 * this is the entry point to schedule() from in-kernel preemption
3409 * off of preempt_enable. Kernel preemptions off return from interrupt
3410 * occur there and call schedule directly.
3412 asmlinkage __visible void __sched notrace preempt_schedule(void)
3415 * If there is a non-zero preempt_count or interrupts are disabled,
3416 * we do not want to preempt the current task. Just return..
3418 if (likely(!preemptible()))
3421 preempt_schedule_common();
3423 NOKPROBE_SYMBOL(preempt_schedule);
3424 EXPORT_SYMBOL(preempt_schedule);
3427 * preempt_schedule_notrace - preempt_schedule called by tracing
3429 * The tracing infrastructure uses preempt_enable_notrace to prevent
3430 * recursion and tracing preempt enabling caused by the tracing
3431 * infrastructure itself. But as tracing can happen in areas coming
3432 * from userspace or just about to enter userspace, a preempt enable
3433 * can occur before user_exit() is called. This will cause the scheduler
3434 * to be called when the system is still in usermode.
3436 * To prevent this, the preempt_enable_notrace will use this function
3437 * instead of preempt_schedule() to exit user context if needed before
3438 * calling the scheduler.
3440 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3442 enum ctx_state prev_ctx;
3444 if (likely(!preemptible()))
3448 preempt_disable_notrace();
3450 * Needs preempt disabled in case user_exit() is traced
3451 * and the tracer calls preempt_enable_notrace() causing
3452 * an infinite recursion.
3454 prev_ctx = exception_enter();
3456 exception_exit(prev_ctx);
3458 preempt_enable_no_resched_notrace();
3459 } while (need_resched());
3461 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3463 #endif /* CONFIG_PREEMPT */
3466 * this is the entry point to schedule() from kernel preemption
3467 * off of irq context.
3468 * Note, that this is called and return with irqs disabled. This will
3469 * protect us against recursive calling from irq.
3471 asmlinkage __visible void __sched preempt_schedule_irq(void)
3473 enum ctx_state prev_state;
3475 /* Catch callers which need to be fixed */
3476 BUG_ON(preempt_count() || !irqs_disabled());
3478 prev_state = exception_enter();
3484 local_irq_disable();
3485 sched_preempt_enable_no_resched();
3486 } while (need_resched());
3488 exception_exit(prev_state);
3491 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3494 return try_to_wake_up(curr->private, mode, wake_flags, 1);
3496 EXPORT_SYMBOL(default_wake_function);
3498 #ifdef CONFIG_RT_MUTEXES
3501 * rt_mutex_setprio - set the current priority of a task
3503 * @prio: prio value (kernel-internal form)
3505 * This function changes the 'effective' priority of a task. It does
3506 * not touch ->normal_prio like __setscheduler().
3508 * Used by the rt_mutex code to implement priority inheritance
3509 * logic. Call site only calls if the priority of the task changed.
3511 void rt_mutex_setprio(struct task_struct *p, int prio)
3513 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3515 const struct sched_class *prev_class;
3517 BUG_ON(prio > MAX_PRIO);
3519 rq = __task_rq_lock(p);
3520 update_rq_clock(rq);
3523 * Idle task boosting is a nono in general. There is one
3524 * exception, when PREEMPT_RT and NOHZ is active:
3526 * The idle task calls get_next_timer_interrupt() and holds
3527 * the timer wheel base->lock on the CPU and another CPU wants
3528 * to access the timer (probably to cancel it). We can safely
3529 * ignore the boosting request, as the idle CPU runs this code
3530 * with interrupts disabled and will complete the lock
3531 * protected section without being interrupted. So there is no
3532 * real need to boost.
3534 if (unlikely(p == rq->idle)) {
3535 WARN_ON(p != rq->curr);
3536 WARN_ON(p->pi_blocked_on);
3540 trace_sched_pi_setprio(p, prio);
3542 prev_class = p->sched_class;
3543 queued = task_on_rq_queued(p);
3544 running = task_current(rq, p);
3546 dequeue_task(rq, p, DEQUEUE_SAVE);
3548 put_prev_task(rq, p);
3551 * Boosting condition are:
3552 * 1. -rt task is running and holds mutex A
3553 * --> -dl task blocks on mutex A
3555 * 2. -dl task is running and holds mutex A
3556 * --> -dl task blocks on mutex A and could preempt the
3559 if (dl_prio(prio)) {
3560 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3561 if (!dl_prio(p->normal_prio) ||
3562 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3563 p->dl.dl_boosted = 1;
3564 enqueue_flag |= ENQUEUE_REPLENISH;
3566 p->dl.dl_boosted = 0;
3567 p->sched_class = &dl_sched_class;
3568 } else if (rt_prio(prio)) {
3569 if (dl_prio(oldprio))
3570 p->dl.dl_boosted = 0;
3572 enqueue_flag |= ENQUEUE_HEAD;
3573 p->sched_class = &rt_sched_class;
3575 if (dl_prio(oldprio))
3576 p->dl.dl_boosted = 0;
3577 if (rt_prio(oldprio))
3579 p->sched_class = &fair_sched_class;
3585 p->sched_class->set_curr_task(rq);
3587 enqueue_task(rq, p, enqueue_flag);
3589 check_class_changed(rq, p, prev_class, oldprio);
3591 preempt_disable(); /* avoid rq from going away on us */
3592 __task_rq_unlock(rq);
3594 balance_callback(rq);
3599 void set_user_nice(struct task_struct *p, long nice)
3601 int old_prio, delta, queued;
3602 unsigned long flags;
3605 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3608 * We have to be careful, if called from sys_setpriority(),
3609 * the task might be in the middle of scheduling on another CPU.
3611 rq = task_rq_lock(p, &flags);
3612 update_rq_clock(rq);
3615 * The RT priorities are set via sched_setscheduler(), but we still
3616 * allow the 'normal' nice value to be set - but as expected
3617 * it wont have any effect on scheduling until the task is
3618 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3620 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3621 p->static_prio = NICE_TO_PRIO(nice);
3624 queued = task_on_rq_queued(p);
3626 dequeue_task(rq, p, DEQUEUE_SAVE);
3628 p->static_prio = NICE_TO_PRIO(nice);
3631 p->prio = effective_prio(p);
3632 delta = p->prio - old_prio;
3635 enqueue_task(rq, p, ENQUEUE_RESTORE);
3637 * If the task increased its priority or is running and
3638 * lowered its priority, then reschedule its CPU:
3640 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3644 task_rq_unlock(rq, p, &flags);
3646 EXPORT_SYMBOL(set_user_nice);
3649 * can_nice - check if a task can reduce its nice value
3653 int can_nice(const struct task_struct *p, const int nice)
3655 /* convert nice value [19,-20] to rlimit style value [1,40] */
3656 int nice_rlim = nice_to_rlimit(nice);
3658 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3659 capable(CAP_SYS_NICE));
3662 #ifdef __ARCH_WANT_SYS_NICE
3665 * sys_nice - change the priority of the current process.
3666 * @increment: priority increment
3668 * sys_setpriority is a more generic, but much slower function that
3669 * does similar things.
3671 SYSCALL_DEFINE1(nice, int, increment)
3676 * Setpriority might change our priority at the same moment.
3677 * We don't have to worry. Conceptually one call occurs first
3678 * and we have a single winner.
3680 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3681 nice = task_nice(current) + increment;
3683 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3684 if (increment < 0 && !can_nice(current, nice))
3687 retval = security_task_setnice(current, nice);
3691 set_user_nice(current, nice);
3698 * task_prio - return the priority value of a given task.
3699 * @p: the task in question.
3701 * Return: The priority value as seen by users in /proc.
3702 * RT tasks are offset by -200. Normal tasks are centered
3703 * around 0, value goes from -16 to +15.
3705 int task_prio(const struct task_struct *p)
3707 return p->prio - MAX_RT_PRIO;
3711 * idle_cpu - is a given cpu idle currently?
3712 * @cpu: the processor in question.
3714 * Return: 1 if the CPU is currently idle. 0 otherwise.
3716 int idle_cpu(int cpu)
3718 struct rq *rq = cpu_rq(cpu);
3720 if (rq->curr != rq->idle)
3727 if (!llist_empty(&rq->wake_list))
3735 * idle_task - return the idle task for a given cpu.
3736 * @cpu: the processor in question.
3738 * Return: The idle task for the cpu @cpu.
3740 struct task_struct *idle_task(int cpu)
3742 return cpu_rq(cpu)->idle;
3746 * find_process_by_pid - find a process with a matching PID value.
3747 * @pid: the pid in question.
3749 * The task of @pid, if found. %NULL otherwise.
3751 static struct task_struct *find_process_by_pid(pid_t pid)
3753 return pid ? find_task_by_vpid(pid) : current;
3757 * This function initializes the sched_dl_entity of a newly becoming
3758 * SCHED_DEADLINE task.
3760 * Only the static values are considered here, the actual runtime and the
3761 * absolute deadline will be properly calculated when the task is enqueued
3762 * for the first time with its new policy.
3765 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3767 struct sched_dl_entity *dl_se = &p->dl;
3769 dl_se->dl_runtime = attr->sched_runtime;
3770 dl_se->dl_deadline = attr->sched_deadline;
3771 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3772 dl_se->flags = attr->sched_flags;
3773 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3774 dl_se->dl_density = to_ratio(dl_se->dl_deadline, dl_se->dl_runtime);
3777 * Changing the parameters of a task is 'tricky' and we're not doing
3778 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3780 * What we SHOULD do is delay the bandwidth release until the 0-lag
3781 * point. This would include retaining the task_struct until that time
3782 * and change dl_overflow() to not immediately decrement the current
3785 * Instead we retain the current runtime/deadline and let the new
3786 * parameters take effect after the current reservation period lapses.
3787 * This is safe (albeit pessimistic) because the 0-lag point is always
3788 * before the current scheduling deadline.
3790 * We can still have temporary overloads because we do not delay the
3791 * change in bandwidth until that time; so admission control is
3792 * not on the safe side. It does however guarantee tasks will never
3793 * consume more than promised.
3798 * sched_setparam() passes in -1 for its policy, to let the functions
3799 * it calls know not to change it.
3801 #define SETPARAM_POLICY -1
3803 static void __setscheduler_params(struct task_struct *p,
3804 const struct sched_attr *attr)
3806 int policy = attr->sched_policy;
3808 if (policy == SETPARAM_POLICY)
3813 if (dl_policy(policy))
3814 __setparam_dl(p, attr);
3815 else if (fair_policy(policy))
3816 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3819 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3820 * !rt_policy. Always setting this ensures that things like
3821 * getparam()/getattr() don't report silly values for !rt tasks.
3823 p->rt_priority = attr->sched_priority;
3824 p->normal_prio = normal_prio(p);
3828 /* Actually do priority change: must hold pi & rq lock. */
3829 static void __setscheduler(struct rq *rq, struct task_struct *p,
3830 const struct sched_attr *attr, bool keep_boost)
3832 __setscheduler_params(p, attr);
3835 * Keep a potential priority boosting if called from
3836 * sched_setscheduler().
3839 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3841 p->prio = normal_prio(p);
3843 if (dl_prio(p->prio))
3844 p->sched_class = &dl_sched_class;
3845 else if (rt_prio(p->prio))
3846 p->sched_class = &rt_sched_class;
3848 p->sched_class = &fair_sched_class;
3852 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3854 struct sched_dl_entity *dl_se = &p->dl;
3856 attr->sched_priority = p->rt_priority;
3857 attr->sched_runtime = dl_se->dl_runtime;
3858 attr->sched_deadline = dl_se->dl_deadline;
3859 attr->sched_period = dl_se->dl_period;
3860 attr->sched_flags = dl_se->flags;
3864 * This function validates the new parameters of a -deadline task.
3865 * We ask for the deadline not being zero, and greater or equal
3866 * than the runtime, as well as the period of being zero or
3867 * greater than deadline. Furthermore, we have to be sure that
3868 * user parameters are above the internal resolution of 1us (we
3869 * check sched_runtime only since it is always the smaller one) and
3870 * below 2^63 ns (we have to check both sched_deadline and
3871 * sched_period, as the latter can be zero).
3874 __checkparam_dl(const struct sched_attr *attr)
3877 if (attr->sched_deadline == 0)
3881 * Since we truncate DL_SCALE bits, make sure we're at least
3884 if (attr->sched_runtime < (1ULL << DL_SCALE))
3888 * Since we use the MSB for wrap-around and sign issues, make
3889 * sure it's not set (mind that period can be equal to zero).
3891 if (attr->sched_deadline & (1ULL << 63) ||
3892 attr->sched_period & (1ULL << 63))
3895 /* runtime <= deadline <= period (if period != 0) */
3896 if ((attr->sched_period != 0 &&
3897 attr->sched_period < attr->sched_deadline) ||
3898 attr->sched_deadline < attr->sched_runtime)
3905 * check the target process has a UID that matches the current process's
3907 static bool check_same_owner(struct task_struct *p)
3909 const struct cred *cred = current_cred(), *pcred;
3913 pcred = __task_cred(p);
3914 match = (uid_eq(cred->euid, pcred->euid) ||
3915 uid_eq(cred->euid, pcred->uid));
3920 static bool dl_param_changed(struct task_struct *p,
3921 const struct sched_attr *attr)
3923 struct sched_dl_entity *dl_se = &p->dl;
3925 if (dl_se->dl_runtime != attr->sched_runtime ||
3926 dl_se->dl_deadline != attr->sched_deadline ||
3927 dl_se->dl_period != attr->sched_period ||
3928 dl_se->flags != attr->sched_flags)
3934 static int __sched_setscheduler(struct task_struct *p,
3935 const struct sched_attr *attr,
3938 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3939 MAX_RT_PRIO - 1 - attr->sched_priority;
3940 int retval, oldprio, oldpolicy = -1, queued, running;
3941 int new_effective_prio, policy = attr->sched_policy;
3942 unsigned long flags;
3943 const struct sched_class *prev_class;
3947 /* may grab non-irq protected spin_locks */
3948 BUG_ON(in_interrupt());
3950 /* double check policy once rq lock held */
3952 reset_on_fork = p->sched_reset_on_fork;
3953 policy = oldpolicy = p->policy;
3955 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3957 if (!valid_policy(policy))
3961 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3965 * Valid priorities for SCHED_FIFO and SCHED_RR are
3966 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3967 * SCHED_BATCH and SCHED_IDLE is 0.
3969 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3970 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3972 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3973 (rt_policy(policy) != (attr->sched_priority != 0)))
3977 * Allow unprivileged RT tasks to decrease priority:
3979 if (user && !capable(CAP_SYS_NICE)) {
3980 if (fair_policy(policy)) {
3981 if (attr->sched_nice < task_nice(p) &&
3982 !can_nice(p, attr->sched_nice))
3986 if (rt_policy(policy)) {
3987 unsigned long rlim_rtprio =
3988 task_rlimit(p, RLIMIT_RTPRIO);
3990 /* can't set/change the rt policy */
3991 if (policy != p->policy && !rlim_rtprio)
3994 /* can't increase priority */
3995 if (attr->sched_priority > p->rt_priority &&
3996 attr->sched_priority > rlim_rtprio)
4001 * Can't set/change SCHED_DEADLINE policy at all for now
4002 * (safest behavior); in the future we would like to allow
4003 * unprivileged DL tasks to increase their relative deadline
4004 * or reduce their runtime (both ways reducing utilization)
4006 if (dl_policy(policy))
4010 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4011 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4013 if (idle_policy(p->policy) && !idle_policy(policy)) {
4014 if (!can_nice(p, task_nice(p)))
4018 /* can't change other user's priorities */
4019 if (!check_same_owner(p))
4022 /* Normal users shall not reset the sched_reset_on_fork flag */
4023 if (p->sched_reset_on_fork && !reset_on_fork)
4028 retval = security_task_setscheduler(p);
4034 * make sure no PI-waiters arrive (or leave) while we are
4035 * changing the priority of the task:
4037 * To be able to change p->policy safely, the appropriate
4038 * runqueue lock must be held.
4040 rq = task_rq_lock(p, &flags);
4041 update_rq_clock(rq);
4044 * Changing the policy of the stop threads its a very bad idea
4046 if (p == rq->stop) {
4047 task_rq_unlock(rq, p, &flags);
4052 * If not changing anything there's no need to proceed further,
4053 * but store a possible modification of reset_on_fork.
4055 if (unlikely(policy == p->policy)) {
4056 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4058 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4060 if (dl_policy(policy) && dl_param_changed(p, attr))
4063 p->sched_reset_on_fork = reset_on_fork;
4064 task_rq_unlock(rq, p, &flags);
4070 #ifdef CONFIG_RT_GROUP_SCHED
4072 * Do not allow realtime tasks into groups that have no runtime
4075 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4076 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4077 !task_group_is_autogroup(task_group(p))) {
4078 task_rq_unlock(rq, p, &flags);
4083 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4084 cpumask_t *span = rq->rd->span;
4087 * Don't allow tasks with an affinity mask smaller than
4088 * the entire root_domain to become SCHED_DEADLINE. We
4089 * will also fail if there's no bandwidth available.
4091 if (!cpumask_subset(span, &p->cpus_allowed) ||
4092 rq->rd->dl_bw.bw == 0) {
4093 task_rq_unlock(rq, p, &flags);
4100 /* recheck policy now with rq lock held */
4101 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4102 policy = oldpolicy = -1;
4103 task_rq_unlock(rq, p, &flags);
4108 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4109 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4112 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4113 task_rq_unlock(rq, p, &flags);
4117 p->sched_reset_on_fork = reset_on_fork;
4122 * Take priority boosted tasks into account. If the new
4123 * effective priority is unchanged, we just store the new
4124 * normal parameters and do not touch the scheduler class and
4125 * the runqueue. This will be done when the task deboost
4128 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4129 if (new_effective_prio == oldprio) {
4130 __setscheduler_params(p, attr);
4131 task_rq_unlock(rq, p, &flags);
4136 queued = task_on_rq_queued(p);
4137 running = task_current(rq, p);
4139 dequeue_task(rq, p, DEQUEUE_SAVE);
4141 put_prev_task(rq, p);
4143 prev_class = p->sched_class;
4144 __setscheduler(rq, p, attr, pi);
4147 p->sched_class->set_curr_task(rq);
4149 int enqueue_flags = ENQUEUE_RESTORE;
4151 * We enqueue to tail when the priority of a task is
4152 * increased (user space view).
4154 if (oldprio <= p->prio)
4155 enqueue_flags |= ENQUEUE_HEAD;
4157 enqueue_task(rq, p, enqueue_flags);
4160 check_class_changed(rq, p, prev_class, oldprio);
4161 preempt_disable(); /* avoid rq from going away on us */
4162 task_rq_unlock(rq, p, &flags);
4165 rt_mutex_adjust_pi(p);
4168 * Run balance callbacks after we've adjusted the PI chain.
4170 balance_callback(rq);
4176 static int _sched_setscheduler(struct task_struct *p, int policy,
4177 const struct sched_param *param, bool check)
4179 struct sched_attr attr = {
4180 .sched_policy = policy,
4181 .sched_priority = param->sched_priority,
4182 .sched_nice = PRIO_TO_NICE(p->static_prio),
4185 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4186 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4187 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4188 policy &= ~SCHED_RESET_ON_FORK;
4189 attr.sched_policy = policy;
4192 return __sched_setscheduler(p, &attr, check, true);
4195 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4196 * @p: the task in question.
4197 * @policy: new policy.
4198 * @param: structure containing the new RT priority.
4200 * Return: 0 on success. An error code otherwise.
4202 * NOTE that the task may be already dead.
4204 int sched_setscheduler(struct task_struct *p, int policy,
4205 const struct sched_param *param)
4207 return _sched_setscheduler(p, policy, param, true);
4209 EXPORT_SYMBOL_GPL(sched_setscheduler);
4211 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4213 return __sched_setscheduler(p, attr, true, true);
4215 EXPORT_SYMBOL_GPL(sched_setattr);
4218 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4219 * @p: the task in question.
4220 * @policy: new policy.
4221 * @param: structure containing the new RT priority.
4223 * Just like sched_setscheduler, only don't bother checking if the
4224 * current context has permission. For example, this is needed in
4225 * stop_machine(): we create temporary high priority worker threads,
4226 * but our caller might not have that capability.
4228 * Return: 0 on success. An error code otherwise.
4230 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4231 const struct sched_param *param)
4233 return _sched_setscheduler(p, policy, param, false);
4235 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4238 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4240 struct sched_param lparam;
4241 struct task_struct *p;
4244 if (!param || pid < 0)
4246 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4251 p = find_process_by_pid(pid);
4253 retval = sched_setscheduler(p, policy, &lparam);
4260 * Mimics kernel/events/core.c perf_copy_attr().
4262 static int sched_copy_attr(struct sched_attr __user *uattr,
4263 struct sched_attr *attr)
4268 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4272 * zero the full structure, so that a short copy will be nice.
4274 memset(attr, 0, sizeof(*attr));
4276 ret = get_user(size, &uattr->size);
4280 if (size > PAGE_SIZE) /* silly large */
4283 if (!size) /* abi compat */
4284 size = SCHED_ATTR_SIZE_VER0;
4286 if (size < SCHED_ATTR_SIZE_VER0)
4290 * If we're handed a bigger struct than we know of,
4291 * ensure all the unknown bits are 0 - i.e. new
4292 * user-space does not rely on any kernel feature
4293 * extensions we dont know about yet.
4295 if (size > sizeof(*attr)) {
4296 unsigned char __user *addr;
4297 unsigned char __user *end;
4300 addr = (void __user *)uattr + sizeof(*attr);
4301 end = (void __user *)uattr + size;
4303 for (; addr < end; addr++) {
4304 ret = get_user(val, addr);
4310 size = sizeof(*attr);
4313 ret = copy_from_user(attr, uattr, size);
4318 * XXX: do we want to be lenient like existing syscalls; or do we want
4319 * to be strict and return an error on out-of-bounds values?
4321 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4326 put_user(sizeof(*attr), &uattr->size);
4331 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4332 * @pid: the pid in question.
4333 * @policy: new policy.
4334 * @param: structure containing the new RT priority.
4336 * Return: 0 on success. An error code otherwise.
4338 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4339 struct sched_param __user *, param)
4341 /* negative values for policy are not valid */
4345 return do_sched_setscheduler(pid, policy, param);
4349 * sys_sched_setparam - set/change the RT priority of a thread
4350 * @pid: the pid in question.
4351 * @param: structure containing the new RT priority.
4353 * Return: 0 on success. An error code otherwise.
4355 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4357 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4361 * sys_sched_setattr - same as above, but with extended sched_attr
4362 * @pid: the pid in question.
4363 * @uattr: structure containing the extended parameters.
4364 * @flags: for future extension.
4366 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4367 unsigned int, flags)
4369 struct sched_attr attr;
4370 struct task_struct *p;
4373 if (!uattr || pid < 0 || flags)
4376 retval = sched_copy_attr(uattr, &attr);
4380 if ((int)attr.sched_policy < 0)
4385 p = find_process_by_pid(pid);
4387 retval = sched_setattr(p, &attr);
4394 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4395 * @pid: the pid in question.
4397 * Return: On success, the policy of the thread. Otherwise, a negative error
4400 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4402 struct task_struct *p;
4410 p = find_process_by_pid(pid);
4412 retval = security_task_getscheduler(p);
4415 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4422 * sys_sched_getparam - get the RT priority of a thread
4423 * @pid: the pid in question.
4424 * @param: structure containing the RT priority.
4426 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4429 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4431 struct sched_param lp = { .sched_priority = 0 };
4432 struct task_struct *p;
4435 if (!param || pid < 0)
4439 p = find_process_by_pid(pid);
4444 retval = security_task_getscheduler(p);
4448 if (task_has_rt_policy(p))
4449 lp.sched_priority = p->rt_priority;
4453 * This one might sleep, we cannot do it with a spinlock held ...
4455 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4464 static int sched_read_attr(struct sched_attr __user *uattr,
4465 struct sched_attr *attr,
4470 if (!access_ok(VERIFY_WRITE, uattr, usize))
4474 * If we're handed a smaller struct than we know of,
4475 * ensure all the unknown bits are 0 - i.e. old
4476 * user-space does not get uncomplete information.
4478 if (usize < sizeof(*attr)) {
4479 unsigned char *addr;
4482 addr = (void *)attr + usize;
4483 end = (void *)attr + sizeof(*attr);
4485 for (; addr < end; addr++) {
4493 ret = copy_to_user(uattr, attr, attr->size);
4501 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4502 * @pid: the pid in question.
4503 * @uattr: structure containing the extended parameters.
4504 * @size: sizeof(attr) for fwd/bwd comp.
4505 * @flags: for future extension.
4507 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4508 unsigned int, size, unsigned int, flags)
4510 struct sched_attr attr = {
4511 .size = sizeof(struct sched_attr),
4513 struct task_struct *p;
4516 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4517 size < SCHED_ATTR_SIZE_VER0 || flags)
4521 p = find_process_by_pid(pid);
4526 retval = security_task_getscheduler(p);
4530 attr.sched_policy = p->policy;
4531 if (p->sched_reset_on_fork)
4532 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4533 if (task_has_dl_policy(p))
4534 __getparam_dl(p, &attr);
4535 else if (task_has_rt_policy(p))
4536 attr.sched_priority = p->rt_priority;
4538 attr.sched_nice = task_nice(p);
4542 retval = sched_read_attr(uattr, &attr, size);
4550 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4552 cpumask_var_t cpus_allowed, new_mask;
4553 struct task_struct *p;
4558 p = find_process_by_pid(pid);
4564 /* Prevent p going away */
4568 if (p->flags & PF_NO_SETAFFINITY) {
4572 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4576 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4578 goto out_free_cpus_allowed;
4581 if (!check_same_owner(p)) {
4583 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4585 goto out_free_new_mask;
4590 retval = security_task_setscheduler(p);
4592 goto out_free_new_mask;
4595 cpuset_cpus_allowed(p, cpus_allowed);
4596 cpumask_and(new_mask, in_mask, cpus_allowed);
4599 * Since bandwidth control happens on root_domain basis,
4600 * if admission test is enabled, we only admit -deadline
4601 * tasks allowed to run on all the CPUs in the task's
4605 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4607 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4610 goto out_free_new_mask;
4616 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4619 cpuset_cpus_allowed(p, cpus_allowed);
4620 if (!cpumask_subset(new_mask, cpus_allowed)) {
4622 * We must have raced with a concurrent cpuset
4623 * update. Just reset the cpus_allowed to the
4624 * cpuset's cpus_allowed
4626 cpumask_copy(new_mask, cpus_allowed);
4631 free_cpumask_var(new_mask);
4632 out_free_cpus_allowed:
4633 free_cpumask_var(cpus_allowed);
4639 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4640 struct cpumask *new_mask)
4642 if (len < cpumask_size())
4643 cpumask_clear(new_mask);
4644 else if (len > cpumask_size())
4645 len = cpumask_size();
4647 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4651 * sys_sched_setaffinity - set the cpu affinity of a process
4652 * @pid: pid of the process
4653 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4654 * @user_mask_ptr: user-space pointer to the new cpu mask
4656 * Return: 0 on success. An error code otherwise.
4658 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4659 unsigned long __user *, user_mask_ptr)
4661 cpumask_var_t new_mask;
4664 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4667 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4669 retval = sched_setaffinity(pid, new_mask);
4670 free_cpumask_var(new_mask);
4674 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4676 struct task_struct *p;
4677 unsigned long flags;
4683 p = find_process_by_pid(pid);
4687 retval = security_task_getscheduler(p);
4691 raw_spin_lock_irqsave(&p->pi_lock, flags);
4692 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4693 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4702 * sys_sched_getaffinity - get the cpu affinity of a process
4703 * @pid: pid of the process
4704 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4705 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4707 * Return: 0 on success. An error code otherwise.
4709 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4710 unsigned long __user *, user_mask_ptr)
4715 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4717 if (len & (sizeof(unsigned long)-1))
4720 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4723 ret = sched_getaffinity(pid, mask);
4725 size_t retlen = min_t(size_t, len, cpumask_size());
4727 if (copy_to_user(user_mask_ptr, mask, retlen))
4732 free_cpumask_var(mask);
4738 * sys_sched_yield - yield the current processor to other threads.
4740 * This function yields the current CPU to other tasks. If there are no
4741 * other threads running on this CPU then this function will return.
4745 SYSCALL_DEFINE0(sched_yield)
4747 struct rq *rq = this_rq_lock();
4749 schedstat_inc(rq, yld_count);
4750 current->sched_class->yield_task(rq);
4753 * Since we are going to call schedule() anyway, there's
4754 * no need to preempt or enable interrupts:
4756 __release(rq->lock);
4757 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4758 do_raw_spin_unlock(&rq->lock);
4759 sched_preempt_enable_no_resched();
4766 int __sched _cond_resched(void)
4768 if (should_resched(0)) {
4769 preempt_schedule_common();
4774 EXPORT_SYMBOL(_cond_resched);
4777 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4778 * call schedule, and on return reacquire the lock.
4780 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4781 * operations here to prevent schedule() from being called twice (once via
4782 * spin_unlock(), once by hand).
4784 int __cond_resched_lock(spinlock_t *lock)
4786 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4789 lockdep_assert_held(lock);
4791 if (spin_needbreak(lock) || resched) {
4794 preempt_schedule_common();
4802 EXPORT_SYMBOL(__cond_resched_lock);
4804 int __sched __cond_resched_softirq(void)
4806 BUG_ON(!in_softirq());
4808 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4810 preempt_schedule_common();
4816 EXPORT_SYMBOL(__cond_resched_softirq);
4819 * yield - yield the current processor to other threads.
4821 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4823 * The scheduler is at all times free to pick the calling task as the most
4824 * eligible task to run, if removing the yield() call from your code breaks
4825 * it, its already broken.
4827 * Typical broken usage is:
4832 * where one assumes that yield() will let 'the other' process run that will
4833 * make event true. If the current task is a SCHED_FIFO task that will never
4834 * happen. Never use yield() as a progress guarantee!!
4836 * If you want to use yield() to wait for something, use wait_event().
4837 * If you want to use yield() to be 'nice' for others, use cond_resched().
4838 * If you still want to use yield(), do not!
4840 void __sched yield(void)
4842 set_current_state(TASK_RUNNING);
4845 EXPORT_SYMBOL(yield);
4848 * yield_to - yield the current processor to another thread in
4849 * your thread group, or accelerate that thread toward the
4850 * processor it's on.
4852 * @preempt: whether task preemption is allowed or not
4854 * It's the caller's job to ensure that the target task struct
4855 * can't go away on us before we can do any checks.
4858 * true (>0) if we indeed boosted the target task.
4859 * false (0) if we failed to boost the target.
4860 * -ESRCH if there's no task to yield to.
4862 int __sched yield_to(struct task_struct *p, bool preempt)
4864 struct task_struct *curr = current;
4865 struct rq *rq, *p_rq;
4866 unsigned long flags;
4869 local_irq_save(flags);
4875 * If we're the only runnable task on the rq and target rq also
4876 * has only one task, there's absolutely no point in yielding.
4878 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4883 double_rq_lock(rq, p_rq);
4884 if (task_rq(p) != p_rq) {
4885 double_rq_unlock(rq, p_rq);
4889 if (!curr->sched_class->yield_to_task)
4892 if (curr->sched_class != p->sched_class)
4895 if (task_running(p_rq, p) || p->state)
4898 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4900 schedstat_inc(rq, yld_count);
4902 * Make p's CPU reschedule; pick_next_entity takes care of
4905 if (preempt && rq != p_rq)
4910 double_rq_unlock(rq, p_rq);
4912 local_irq_restore(flags);
4919 EXPORT_SYMBOL_GPL(yield_to);
4922 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4923 * that process accounting knows that this is a task in IO wait state.
4925 long __sched io_schedule_timeout(long timeout)
4927 int old_iowait = current->in_iowait;
4931 current->in_iowait = 1;
4932 blk_schedule_flush_plug(current);
4934 delayacct_blkio_start();
4936 atomic_inc(&rq->nr_iowait);
4937 ret = schedule_timeout(timeout);
4938 current->in_iowait = old_iowait;
4939 atomic_dec(&rq->nr_iowait);
4940 delayacct_blkio_end();
4944 EXPORT_SYMBOL(io_schedule_timeout);
4947 * sys_sched_get_priority_max - return maximum RT priority.
4948 * @policy: scheduling class.
4950 * Return: On success, this syscall returns the maximum
4951 * rt_priority that can be used by a given scheduling class.
4952 * On failure, a negative error code is returned.
4954 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4961 ret = MAX_USER_RT_PRIO-1;
4963 case SCHED_DEADLINE:
4974 * sys_sched_get_priority_min - return minimum RT priority.
4975 * @policy: scheduling class.
4977 * Return: On success, this syscall returns the minimum
4978 * rt_priority that can be used by a given scheduling class.
4979 * On failure, a negative error code is returned.
4981 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4990 case SCHED_DEADLINE:
5000 * sys_sched_rr_get_interval - return the default timeslice of a process.
5001 * @pid: pid of the process.
5002 * @interval: userspace pointer to the timeslice value.
5004 * this syscall writes the default timeslice value of a given process
5005 * into the user-space timespec buffer. A value of '0' means infinity.
5007 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5010 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5011 struct timespec __user *, interval)
5013 struct task_struct *p;
5014 unsigned int time_slice;
5015 unsigned long flags;
5025 p = find_process_by_pid(pid);
5029 retval = security_task_getscheduler(p);
5033 rq = task_rq_lock(p, &flags);
5035 if (p->sched_class->get_rr_interval)
5036 time_slice = p->sched_class->get_rr_interval(rq, p);
5037 task_rq_unlock(rq, p, &flags);
5040 jiffies_to_timespec(time_slice, &t);
5041 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5049 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5051 void sched_show_task(struct task_struct *p)
5053 unsigned long free = 0;
5055 unsigned long state = p->state;
5058 state = __ffs(state) + 1;
5059 printk(KERN_INFO "%-15.15s %c", p->comm,
5060 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5061 #if BITS_PER_LONG == 32
5062 if (state == TASK_RUNNING)
5063 printk(KERN_CONT " running ");
5065 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5067 if (state == TASK_RUNNING)
5068 printk(KERN_CONT " running task ");
5070 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5072 #ifdef CONFIG_DEBUG_STACK_USAGE
5073 free = stack_not_used(p);
5078 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5080 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5081 task_pid_nr(p), ppid,
5082 (unsigned long)task_thread_info(p)->flags);
5084 print_worker_info(KERN_INFO, p);
5085 show_stack(p, NULL);
5088 void show_state_filter(unsigned long state_filter)
5090 struct task_struct *g, *p;
5092 #if BITS_PER_LONG == 32
5094 " task PC stack pid father\n");
5097 " task PC stack pid father\n");
5100 for_each_process_thread(g, p) {
5102 * reset the NMI-timeout, listing all files on a slow
5103 * console might take a lot of time:
5104 * Also, reset softlockup watchdogs on all CPUs, because
5105 * another CPU might be blocked waiting for us to process
5108 touch_nmi_watchdog();
5109 touch_all_softlockup_watchdogs();
5110 if (!state_filter || (p->state & state_filter))
5114 #ifdef CONFIG_SCHED_DEBUG
5115 sysrq_sched_debug_show();
5119 * Only show locks if all tasks are dumped:
5122 debug_show_all_locks();
5125 void init_idle_bootup_task(struct task_struct *idle)
5127 idle->sched_class = &idle_sched_class;
5131 * init_idle - set up an idle thread for a given CPU
5132 * @idle: task in question
5133 * @cpu: cpu the idle task belongs to
5135 * NOTE: this function does not set the idle thread's NEED_RESCHED
5136 * flag, to make booting more robust.
5138 void init_idle(struct task_struct *idle, int cpu)
5140 struct rq *rq = cpu_rq(cpu);
5141 unsigned long flags;
5143 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5144 raw_spin_lock(&rq->lock);
5146 __sched_fork(0, idle);
5148 idle->state = TASK_RUNNING;
5149 idle->se.exec_start = sched_clock();
5153 * Its possible that init_idle() gets called multiple times on a task,
5154 * in that case do_set_cpus_allowed() will not do the right thing.
5156 * And since this is boot we can forgo the serialization.
5158 set_cpus_allowed_common(idle, cpumask_of(cpu));
5161 * We're having a chicken and egg problem, even though we are
5162 * holding rq->lock, the cpu isn't yet set to this cpu so the
5163 * lockdep check in task_group() will fail.
5165 * Similar case to sched_fork(). / Alternatively we could
5166 * use task_rq_lock() here and obtain the other rq->lock.
5171 __set_task_cpu(idle, cpu);
5174 rq->curr = rq->idle = idle;
5175 idle->on_rq = TASK_ON_RQ_QUEUED;
5179 raw_spin_unlock(&rq->lock);
5180 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5182 /* Set the preempt count _outside_ the spinlocks! */
5183 init_idle_preempt_count(idle, cpu);
5186 * The idle tasks have their own, simple scheduling class:
5188 idle->sched_class = &idle_sched_class;
5189 ftrace_graph_init_idle_task(idle, cpu);
5190 vtime_init_idle(idle, cpu);
5192 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5196 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5197 const struct cpumask *trial)
5199 int ret = 1, trial_cpus;
5200 struct dl_bw *cur_dl_b;
5201 unsigned long flags;
5203 if (!cpumask_weight(cur))
5206 rcu_read_lock_sched();
5207 cur_dl_b = dl_bw_of(cpumask_any(cur));
5208 trial_cpus = cpumask_weight(trial);
5210 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5211 if (cur_dl_b->bw != -1 &&
5212 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5214 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5215 rcu_read_unlock_sched();
5220 int task_can_attach(struct task_struct *p,
5221 const struct cpumask *cs_cpus_allowed)
5226 * Kthreads which disallow setaffinity shouldn't be moved
5227 * to a new cpuset; we don't want to change their cpu
5228 * affinity and isolating such threads by their set of
5229 * allowed nodes is unnecessary. Thus, cpusets are not
5230 * applicable for such threads. This prevents checking for
5231 * success of set_cpus_allowed_ptr() on all attached tasks
5232 * before cpus_allowed may be changed.
5234 if (p->flags & PF_NO_SETAFFINITY) {
5240 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5242 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5247 unsigned long flags;
5249 rcu_read_lock_sched();
5250 dl_b = dl_bw_of(dest_cpu);
5251 raw_spin_lock_irqsave(&dl_b->lock, flags);
5252 cpus = dl_bw_cpus(dest_cpu);
5253 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5258 * We reserve space for this task in the destination
5259 * root_domain, as we can't fail after this point.
5260 * We will free resources in the source root_domain
5261 * later on (see set_cpus_allowed_dl()).
5263 __dl_add(dl_b, p->dl.dl_bw);
5265 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5266 rcu_read_unlock_sched();
5276 #ifdef CONFIG_NUMA_BALANCING
5277 /* Migrate current task p to target_cpu */
5278 int migrate_task_to(struct task_struct *p, int target_cpu)
5280 struct migration_arg arg = { p, target_cpu };
5281 int curr_cpu = task_cpu(p);
5283 if (curr_cpu == target_cpu)
5286 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5289 /* TODO: This is not properly updating schedstats */
5291 trace_sched_move_numa(p, curr_cpu, target_cpu);
5292 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5296 * Requeue a task on a given node and accurately track the number of NUMA
5297 * tasks on the runqueues
5299 void sched_setnuma(struct task_struct *p, int nid)
5302 unsigned long flags;
5303 bool queued, running;
5305 rq = task_rq_lock(p, &flags);
5306 queued = task_on_rq_queued(p);
5307 running = task_current(rq, p);
5310 dequeue_task(rq, p, DEQUEUE_SAVE);
5312 put_prev_task(rq, p);
5314 p->numa_preferred_nid = nid;
5317 p->sched_class->set_curr_task(rq);
5319 enqueue_task(rq, p, ENQUEUE_RESTORE);
5320 task_rq_unlock(rq, p, &flags);
5322 #endif /* CONFIG_NUMA_BALANCING */
5324 #ifdef CONFIG_HOTPLUG_CPU
5326 * Ensures that the idle task is using init_mm right before its cpu goes
5329 void idle_task_exit(void)
5331 struct mm_struct *mm = current->active_mm;
5333 BUG_ON(cpu_online(smp_processor_id()));
5335 if (mm != &init_mm) {
5336 switch_mm(mm, &init_mm, current);
5337 finish_arch_post_lock_switch();
5343 * Since this CPU is going 'away' for a while, fold any nr_active delta
5344 * we might have. Assumes we're called after migrate_tasks() so that the
5345 * nr_active count is stable.
5347 * Also see the comment "Global load-average calculations".
5349 static void calc_load_migrate(struct rq *rq)
5351 long delta = calc_load_fold_active(rq);
5353 atomic_long_add(delta, &calc_load_tasks);
5356 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5360 static const struct sched_class fake_sched_class = {
5361 .put_prev_task = put_prev_task_fake,
5364 static struct task_struct fake_task = {
5366 * Avoid pull_{rt,dl}_task()
5368 .prio = MAX_PRIO + 1,
5369 .sched_class = &fake_sched_class,
5373 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5374 * try_to_wake_up()->select_task_rq().
5376 * Called with rq->lock held even though we'er in stop_machine() and
5377 * there's no concurrency possible, we hold the required locks anyway
5378 * because of lock validation efforts.
5380 static void migrate_tasks(struct rq *dead_rq)
5382 struct rq *rq = dead_rq;
5383 struct task_struct *next, *stop = rq->stop;
5387 * Fudge the rq selection such that the below task selection loop
5388 * doesn't get stuck on the currently eligible stop task.
5390 * We're currently inside stop_machine() and the rq is either stuck
5391 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5392 * either way we should never end up calling schedule() until we're
5398 * put_prev_task() and pick_next_task() sched
5399 * class method both need to have an up-to-date
5400 * value of rq->clock[_task]
5402 update_rq_clock(rq);
5406 * There's this thread running, bail when that's the only
5409 if (rq->nr_running == 1)
5413 * pick_next_task assumes pinned rq->lock.
5415 lockdep_pin_lock(&rq->lock);
5416 next = pick_next_task(rq, &fake_task);
5418 next->sched_class->put_prev_task(rq, next);
5421 * Rules for changing task_struct::cpus_allowed are holding
5422 * both pi_lock and rq->lock, such that holding either
5423 * stabilizes the mask.
5425 * Drop rq->lock is not quite as disastrous as it usually is
5426 * because !cpu_active at this point, which means load-balance
5427 * will not interfere. Also, stop-machine.
5429 lockdep_unpin_lock(&rq->lock);
5430 raw_spin_unlock(&rq->lock);
5431 raw_spin_lock(&next->pi_lock);
5432 raw_spin_lock(&rq->lock);
5435 * Since we're inside stop-machine, _nothing_ should have
5436 * changed the task, WARN if weird stuff happened, because in
5437 * that case the above rq->lock drop is a fail too.
5439 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5440 raw_spin_unlock(&next->pi_lock);
5444 /* Find suitable destination for @next, with force if needed. */
5445 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5447 rq = __migrate_task(rq, next, dest_cpu);
5448 if (rq != dead_rq) {
5449 raw_spin_unlock(&rq->lock);
5451 raw_spin_lock(&rq->lock);
5453 raw_spin_unlock(&next->pi_lock);
5458 #endif /* CONFIG_HOTPLUG_CPU */
5460 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5462 static struct ctl_table sd_ctl_dir[] = {
5464 .procname = "sched_domain",
5470 static struct ctl_table sd_ctl_root[] = {
5472 .procname = "kernel",
5474 .child = sd_ctl_dir,
5479 static struct ctl_table *sd_alloc_ctl_entry(int n)
5481 struct ctl_table *entry =
5482 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5487 static void sd_free_ctl_entry(struct ctl_table **tablep)
5489 struct ctl_table *entry;
5492 * In the intermediate directories, both the child directory and
5493 * procname are dynamically allocated and could fail but the mode
5494 * will always be set. In the lowest directory the names are
5495 * static strings and all have proc handlers.
5497 for (entry = *tablep; entry->mode; entry++) {
5499 sd_free_ctl_entry(&entry->child);
5500 if (entry->proc_handler == NULL)
5501 kfree(entry->procname);
5508 static int min_load_idx = 0;
5509 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5512 set_table_entry(struct ctl_table *entry,
5513 const char *procname, void *data, int maxlen,
5514 umode_t mode, proc_handler *proc_handler,
5517 entry->procname = procname;
5519 entry->maxlen = maxlen;
5521 entry->proc_handler = proc_handler;
5524 entry->extra1 = &min_load_idx;
5525 entry->extra2 = &max_load_idx;
5529 static struct ctl_table *
5530 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5532 struct ctl_table *table = sd_alloc_ctl_entry(5);
5537 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5538 sizeof(int), 0644, proc_dointvec_minmax, false);
5539 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5540 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5541 proc_doulongvec_minmax, false);
5542 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5543 sizeof(int), 0644, proc_dointvec_minmax, false);
5544 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5545 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5546 proc_doulongvec_minmax, false);
5551 static struct ctl_table *
5552 sd_alloc_ctl_group_table(struct sched_group *sg)
5554 struct ctl_table *table = sd_alloc_ctl_entry(2);
5559 table->procname = kstrdup("energy", GFP_KERNEL);
5561 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5566 static struct ctl_table *
5567 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5569 struct ctl_table *table;
5570 unsigned int nr_entries = 14;
5573 struct sched_group *sg = sd->groups;
5578 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5580 nr_entries += nr_sgs;
5583 table = sd_alloc_ctl_entry(nr_entries);
5588 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5589 sizeof(long), 0644, proc_doulongvec_minmax, false);
5590 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5591 sizeof(long), 0644, proc_doulongvec_minmax, false);
5592 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5593 sizeof(int), 0644, proc_dointvec_minmax, true);
5594 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5595 sizeof(int), 0644, proc_dointvec_minmax, true);
5596 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5597 sizeof(int), 0644, proc_dointvec_minmax, true);
5598 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5599 sizeof(int), 0644, proc_dointvec_minmax, true);
5600 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5601 sizeof(int), 0644, proc_dointvec_minmax, true);
5602 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5603 sizeof(int), 0644, proc_dointvec_minmax, false);
5604 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5605 sizeof(int), 0644, proc_dointvec_minmax, false);
5606 set_table_entry(&table[9], "cache_nice_tries",
5607 &sd->cache_nice_tries,
5608 sizeof(int), 0644, proc_dointvec_minmax, false);
5609 set_table_entry(&table[10], "flags", &sd->flags,
5610 sizeof(int), 0644, proc_dointvec_minmax, false);
5611 set_table_entry(&table[11], "max_newidle_lb_cost",
5612 &sd->max_newidle_lb_cost,
5613 sizeof(long), 0644, proc_doulongvec_minmax, false);
5614 set_table_entry(&table[12], "name", sd->name,
5615 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5619 struct ctl_table *entry = &table[13];
5622 snprintf(buf, 32, "group%d", i);
5623 entry->procname = kstrdup(buf, GFP_KERNEL);
5625 entry->child = sd_alloc_ctl_group_table(sg);
5626 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5628 /* &table[nr_entries-1] is terminator */
5633 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5635 struct ctl_table *entry, *table;
5636 struct sched_domain *sd;
5637 int domain_num = 0, i;
5640 for_each_domain(cpu, sd)
5642 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5647 for_each_domain(cpu, sd) {
5648 snprintf(buf, 32, "domain%d", i);
5649 entry->procname = kstrdup(buf, GFP_KERNEL);
5651 entry->child = sd_alloc_ctl_domain_table(sd);
5658 static struct ctl_table_header *sd_sysctl_header;
5659 static void register_sched_domain_sysctl(void)
5661 int i, cpu_num = num_possible_cpus();
5662 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5665 WARN_ON(sd_ctl_dir[0].child);
5666 sd_ctl_dir[0].child = entry;
5671 for_each_possible_cpu(i) {
5672 snprintf(buf, 32, "cpu%d", i);
5673 entry->procname = kstrdup(buf, GFP_KERNEL);
5675 entry->child = sd_alloc_ctl_cpu_table(i);
5679 WARN_ON(sd_sysctl_header);
5680 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5683 /* may be called multiple times per register */
5684 static void unregister_sched_domain_sysctl(void)
5686 unregister_sysctl_table(sd_sysctl_header);
5687 sd_sysctl_header = NULL;
5688 if (sd_ctl_dir[0].child)
5689 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5692 static void register_sched_domain_sysctl(void)
5695 static void unregister_sched_domain_sysctl(void)
5698 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5700 static void set_rq_online(struct rq *rq)
5703 const struct sched_class *class;
5705 cpumask_set_cpu(rq->cpu, rq->rd->online);
5708 for_each_class(class) {
5709 if (class->rq_online)
5710 class->rq_online(rq);
5715 static void set_rq_offline(struct rq *rq)
5718 const struct sched_class *class;
5720 for_each_class(class) {
5721 if (class->rq_offline)
5722 class->rq_offline(rq);
5725 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5731 * migration_call - callback that gets triggered when a CPU is added.
5732 * Here we can start up the necessary migration thread for the new CPU.
5735 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5737 int cpu = (long)hcpu;
5738 unsigned long flags;
5739 struct rq *rq = cpu_rq(cpu);
5741 switch (action & ~CPU_TASKS_FROZEN) {
5743 case CPU_UP_PREPARE:
5744 raw_spin_lock_irqsave(&rq->lock, flags);
5745 walt_set_window_start(rq);
5746 raw_spin_unlock_irqrestore(&rq->lock, flags);
5747 rq->calc_load_update = calc_load_update;
5751 /* Update our root-domain */
5752 raw_spin_lock_irqsave(&rq->lock, flags);
5754 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5758 raw_spin_unlock_irqrestore(&rq->lock, flags);
5761 #ifdef CONFIG_HOTPLUG_CPU
5763 sched_ttwu_pending();
5764 /* Update our root-domain */
5765 raw_spin_lock_irqsave(&rq->lock, flags);
5766 walt_migrate_sync_cpu(cpu);
5768 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5772 BUG_ON(rq->nr_running != 1); /* the migration thread */
5773 raw_spin_unlock_irqrestore(&rq->lock, flags);
5777 calc_load_migrate(rq);
5782 update_max_interval();
5788 * Register at high priority so that task migration (migrate_all_tasks)
5789 * happens before everything else. This has to be lower priority than
5790 * the notifier in the perf_event subsystem, though.
5792 static struct notifier_block migration_notifier = {
5793 .notifier_call = migration_call,
5794 .priority = CPU_PRI_MIGRATION,
5797 static void set_cpu_rq_start_time(void)
5799 int cpu = smp_processor_id();
5800 struct rq *rq = cpu_rq(cpu);
5801 rq->age_stamp = sched_clock_cpu(cpu);
5804 static int sched_cpu_active(struct notifier_block *nfb,
5805 unsigned long action, void *hcpu)
5807 int cpu = (long)hcpu;
5809 switch (action & ~CPU_TASKS_FROZEN) {
5811 set_cpu_rq_start_time();
5816 * At this point a starting CPU has marked itself as online via
5817 * set_cpu_online(). But it might not yet have marked itself
5818 * as active, which is essential from here on.
5820 set_cpu_active(cpu, true);
5821 stop_machine_unpark(cpu);
5824 case CPU_DOWN_FAILED:
5825 set_cpu_active(cpu, true);
5833 static int sched_cpu_inactive(struct notifier_block *nfb,
5834 unsigned long action, void *hcpu)
5836 switch (action & ~CPU_TASKS_FROZEN) {
5837 case CPU_DOWN_PREPARE:
5838 set_cpu_active((long)hcpu, false);
5845 static int __init migration_init(void)
5847 void *cpu = (void *)(long)smp_processor_id();
5850 /* Initialize migration for the boot CPU */
5851 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5852 BUG_ON(err == NOTIFY_BAD);
5853 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5854 register_cpu_notifier(&migration_notifier);
5856 /* Register cpu active notifiers */
5857 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5858 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5862 early_initcall(migration_init);
5864 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5866 #ifdef CONFIG_SCHED_DEBUG
5868 static __read_mostly int sched_debug_enabled;
5870 static int __init sched_debug_setup(char *str)
5872 sched_debug_enabled = 1;
5876 early_param("sched_debug", sched_debug_setup);
5878 static inline bool sched_debug(void)
5880 return sched_debug_enabled;
5883 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5884 struct cpumask *groupmask)
5886 struct sched_group *group = sd->groups;
5888 cpumask_clear(groupmask);
5890 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5892 if (!(sd->flags & SD_LOAD_BALANCE)) {
5893 printk("does not load-balance\n");
5897 printk(KERN_CONT "span %*pbl level %s\n",
5898 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5900 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5901 printk(KERN_ERR "ERROR: domain->span does not contain "
5904 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5905 printk(KERN_ERR "ERROR: domain->groups does not contain"
5909 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5913 printk(KERN_ERR "ERROR: group is NULL\n");
5917 if (!cpumask_weight(sched_group_cpus(group))) {
5918 printk(KERN_CONT "\n");
5919 printk(KERN_ERR "ERROR: empty group\n");
5923 if (!(sd->flags & SD_OVERLAP) &&
5924 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5925 printk(KERN_CONT "\n");
5926 printk(KERN_ERR "ERROR: repeated CPUs\n");
5930 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5932 printk(KERN_CONT " %*pbl",
5933 cpumask_pr_args(sched_group_cpus(group)));
5934 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5935 printk(KERN_CONT " (cpu_capacity = %lu)",
5936 group->sgc->capacity);
5939 group = group->next;
5940 } while (group != sd->groups);
5941 printk(KERN_CONT "\n");
5943 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5944 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5947 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5948 printk(KERN_ERR "ERROR: parent span is not a superset "
5949 "of domain->span\n");
5953 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5957 if (!sched_debug_enabled)
5961 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5965 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5968 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5976 #else /* !CONFIG_SCHED_DEBUG */
5977 # define sched_domain_debug(sd, cpu) do { } while (0)
5978 static inline bool sched_debug(void)
5982 #endif /* CONFIG_SCHED_DEBUG */
5984 static int sd_degenerate(struct sched_domain *sd)
5986 if (cpumask_weight(sched_domain_span(sd)) == 1) {
5987 if (sd->groups->sge)
5988 sd->flags &= ~SD_LOAD_BALANCE;
5993 /* Following flags need at least 2 groups */
5994 if (sd->flags & (SD_LOAD_BALANCE |
5995 SD_BALANCE_NEWIDLE |
5998 SD_SHARE_CPUCAPACITY |
5999 SD_ASYM_CPUCAPACITY |
6000 SD_SHARE_PKG_RESOURCES |
6001 SD_SHARE_POWERDOMAIN |
6002 SD_SHARE_CAP_STATES)) {
6003 if (sd->groups != sd->groups->next)
6007 /* Following flags don't use groups */
6008 if (sd->flags & (SD_WAKE_AFFINE))
6015 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6017 unsigned long cflags = sd->flags, pflags = parent->flags;
6019 if (sd_degenerate(parent))
6022 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6025 /* Flags needing groups don't count if only 1 group in parent */
6026 if (parent->groups == parent->groups->next) {
6027 pflags &= ~(SD_LOAD_BALANCE |
6028 SD_BALANCE_NEWIDLE |
6031 SD_ASYM_CPUCAPACITY |
6032 SD_SHARE_CPUCAPACITY |
6033 SD_SHARE_PKG_RESOURCES |
6035 SD_SHARE_POWERDOMAIN |
6036 SD_SHARE_CAP_STATES);
6037 if (parent->groups->sge) {
6038 parent->flags &= ~SD_LOAD_BALANCE;
6041 if (nr_node_ids == 1)
6042 pflags &= ~SD_SERIALIZE;
6044 if (~cflags & pflags)
6050 static void free_rootdomain(struct rcu_head *rcu)
6052 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6054 cpupri_cleanup(&rd->cpupri);
6055 cpudl_cleanup(&rd->cpudl);
6056 free_cpumask_var(rd->dlo_mask);
6057 free_cpumask_var(rd->rto_mask);
6058 free_cpumask_var(rd->online);
6059 free_cpumask_var(rd->span);
6063 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6065 struct root_domain *old_rd = NULL;
6066 unsigned long flags;
6068 raw_spin_lock_irqsave(&rq->lock, flags);
6073 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6076 cpumask_clear_cpu(rq->cpu, old_rd->span);
6079 * If we dont want to free the old_rd yet then
6080 * set old_rd to NULL to skip the freeing later
6083 if (!atomic_dec_and_test(&old_rd->refcount))
6087 atomic_inc(&rd->refcount);
6090 cpumask_set_cpu(rq->cpu, rd->span);
6091 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6094 raw_spin_unlock_irqrestore(&rq->lock, flags);
6097 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6100 void sched_get_rd(struct root_domain *rd)
6102 atomic_inc(&rd->refcount);
6105 void sched_put_rd(struct root_domain *rd)
6107 if (!atomic_dec_and_test(&rd->refcount))
6110 call_rcu_sched(&rd->rcu, free_rootdomain);
6113 static int init_rootdomain(struct root_domain *rd)
6115 memset(rd, 0, sizeof(*rd));
6117 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6119 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6121 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6123 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6126 #ifdef HAVE_RT_PUSH_IPI
6128 raw_spin_lock_init(&rd->rto_lock);
6129 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
6132 init_dl_bw(&rd->dl_bw);
6133 if (cpudl_init(&rd->cpudl) != 0)
6136 if (cpupri_init(&rd->cpupri) != 0)
6139 init_max_cpu_capacity(&rd->max_cpu_capacity);
6141 rd->max_cap_orig_cpu = rd->min_cap_orig_cpu = -1;
6146 free_cpumask_var(rd->rto_mask);
6148 free_cpumask_var(rd->dlo_mask);
6150 free_cpumask_var(rd->online);
6152 free_cpumask_var(rd->span);
6158 * By default the system creates a single root-domain with all cpus as
6159 * members (mimicking the global state we have today).
6161 struct root_domain def_root_domain;
6163 static void init_defrootdomain(void)
6165 init_rootdomain(&def_root_domain);
6167 atomic_set(&def_root_domain.refcount, 1);
6170 static struct root_domain *alloc_rootdomain(void)
6172 struct root_domain *rd;
6174 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6178 if (init_rootdomain(rd) != 0) {
6186 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6188 struct sched_group *tmp, *first;
6197 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6202 } while (sg != first);
6205 static void free_sched_domain(struct rcu_head *rcu)
6207 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6210 * If its an overlapping domain it has private groups, iterate and
6213 if (sd->flags & SD_OVERLAP) {
6214 free_sched_groups(sd->groups, 1);
6215 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6216 kfree(sd->groups->sgc);
6222 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6224 call_rcu(&sd->rcu, free_sched_domain);
6227 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6229 for (; sd; sd = sd->parent)
6230 destroy_sched_domain(sd, cpu);
6234 * Keep a special pointer to the highest sched_domain that has
6235 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6236 * allows us to avoid some pointer chasing select_idle_sibling().
6238 * Also keep a unique ID per domain (we use the first cpu number in
6239 * the cpumask of the domain), this allows us to quickly tell if
6240 * two cpus are in the same cache domain, see cpus_share_cache().
6242 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6243 DEFINE_PER_CPU(int, sd_llc_size);
6244 DEFINE_PER_CPU(int, sd_llc_id);
6245 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6246 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6247 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6248 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6249 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6251 static void update_top_cache_domain(int cpu)
6253 struct sched_domain *sd;
6254 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6258 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6260 id = cpumask_first(sched_domain_span(sd));
6261 size = cpumask_weight(sched_domain_span(sd));
6262 busy_sd = sd->parent; /* sd_busy */
6264 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6266 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6267 per_cpu(sd_llc_size, cpu) = size;
6268 per_cpu(sd_llc_id, cpu) = id;
6270 sd = lowest_flag_domain(cpu, SD_NUMA);
6271 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6273 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6274 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6276 for_each_domain(cpu, sd) {
6277 if (sd->groups->sge)
6282 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6284 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6285 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6289 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6290 * hold the hotplug lock.
6293 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6295 struct rq *rq = cpu_rq(cpu);
6296 struct sched_domain *tmp;
6298 /* Remove the sched domains which do not contribute to scheduling. */
6299 for (tmp = sd; tmp; ) {
6300 struct sched_domain *parent = tmp->parent;
6304 if (sd_parent_degenerate(tmp, parent)) {
6305 tmp->parent = parent->parent;
6307 parent->parent->child = tmp;
6309 * Transfer SD_PREFER_SIBLING down in case of a
6310 * degenerate parent; the spans match for this
6311 * so the property transfers.
6313 if (parent->flags & SD_PREFER_SIBLING)
6314 tmp->flags |= SD_PREFER_SIBLING;
6315 destroy_sched_domain(parent, cpu);
6320 if (sd && sd_degenerate(sd)) {
6323 destroy_sched_domain(tmp, cpu);
6328 sched_domain_debug(sd, cpu);
6330 rq_attach_root(rq, rd);
6332 rcu_assign_pointer(rq->sd, sd);
6333 destroy_sched_domains(tmp, cpu);
6335 update_top_cache_domain(cpu);
6338 /* Setup the mask of cpus configured for isolated domains */
6339 static int __init isolated_cpu_setup(char *str)
6341 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6342 cpulist_parse(str, cpu_isolated_map);
6346 __setup("isolcpus=", isolated_cpu_setup);
6349 struct sched_domain ** __percpu sd;
6350 struct root_domain *rd;
6361 * Build an iteration mask that can exclude certain CPUs from the upwards
6364 * Only CPUs that can arrive at this group should be considered to continue
6367 * Asymmetric node setups can result in situations where the domain tree is of
6368 * unequal depth, make sure to skip domains that already cover the entire
6371 * In that case build_sched_domains() will have terminated the iteration early
6372 * and our sibling sd spans will be empty. Domains should always include the
6373 * cpu they're built on, so check that.
6376 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6378 const struct cpumask *sg_span = sched_group_cpus(sg);
6379 struct sd_data *sdd = sd->private;
6380 struct sched_domain *sibling;
6383 for_each_cpu(i, sg_span) {
6384 sibling = *per_cpu_ptr(sdd->sd, i);
6387 * Can happen in the asymmetric case, where these siblings are
6388 * unused. The mask will not be empty because those CPUs that
6389 * do have the top domain _should_ span the domain.
6391 if (!sibling->child)
6394 /* If we would not end up here, we can't continue from here */
6395 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6398 cpumask_set_cpu(i, sched_group_mask(sg));
6401 /* We must not have empty masks here */
6402 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6406 * Return the canonical balance cpu for this group, this is the first cpu
6407 * of this group that's also in the iteration mask.
6409 int group_balance_cpu(struct sched_group *sg)
6411 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6415 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6417 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6418 const struct cpumask *span = sched_domain_span(sd);
6419 struct cpumask *covered = sched_domains_tmpmask;
6420 struct sd_data *sdd = sd->private;
6421 struct sched_domain *sibling;
6424 cpumask_clear(covered);
6426 for_each_cpu(i, span) {
6427 struct cpumask *sg_span;
6429 if (cpumask_test_cpu(i, covered))
6432 sibling = *per_cpu_ptr(sdd->sd, i);
6434 /* See the comment near build_group_mask(). */
6435 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6438 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6439 GFP_KERNEL, cpu_to_node(cpu));
6444 sg_span = sched_group_cpus(sg);
6446 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6448 cpumask_set_cpu(i, sg_span);
6450 cpumask_or(covered, covered, sg_span);
6452 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6453 if (atomic_inc_return(&sg->sgc->ref) == 1)
6454 build_group_mask(sd, sg);
6457 * Initialize sgc->capacity such that even if we mess up the
6458 * domains and no possible iteration will get us here, we won't
6461 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6462 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6463 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6466 * Make sure the first group of this domain contains the
6467 * canonical balance cpu. Otherwise the sched_domain iteration
6468 * breaks. See update_sg_lb_stats().
6470 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6471 group_balance_cpu(sg) == cpu)
6481 sd->groups = groups;
6486 free_sched_groups(first, 0);
6491 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6493 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6494 struct sched_domain *child = sd->child;
6497 cpu = cpumask_first(sched_domain_span(child));
6500 *sg = *per_cpu_ptr(sdd->sg, cpu);
6501 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6502 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6509 * build_sched_groups will build a circular linked list of the groups
6510 * covered by the given span, and will set each group's ->cpumask correctly,
6511 * and ->cpu_capacity to 0.
6513 * Assumes the sched_domain tree is fully constructed
6516 build_sched_groups(struct sched_domain *sd, int cpu)
6518 struct sched_group *first = NULL, *last = NULL;
6519 struct sd_data *sdd = sd->private;
6520 const struct cpumask *span = sched_domain_span(sd);
6521 struct cpumask *covered;
6524 get_group(cpu, sdd, &sd->groups);
6525 atomic_inc(&sd->groups->ref);
6527 if (cpu != cpumask_first(span))
6530 lockdep_assert_held(&sched_domains_mutex);
6531 covered = sched_domains_tmpmask;
6533 cpumask_clear(covered);
6535 for_each_cpu(i, span) {
6536 struct sched_group *sg;
6539 if (cpumask_test_cpu(i, covered))
6542 group = get_group(i, sdd, &sg);
6543 cpumask_setall(sched_group_mask(sg));
6545 for_each_cpu(j, span) {
6546 if (get_group(j, sdd, NULL) != group)
6549 cpumask_set_cpu(j, covered);
6550 cpumask_set_cpu(j, sched_group_cpus(sg));
6565 * Initialize sched groups cpu_capacity.
6567 * cpu_capacity indicates the capacity of sched group, which is used while
6568 * distributing the load between different sched groups in a sched domain.
6569 * Typically cpu_capacity for all the groups in a sched domain will be same
6570 * unless there are asymmetries in the topology. If there are asymmetries,
6571 * group having more cpu_capacity will pickup more load compared to the
6572 * group having less cpu_capacity.
6574 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6576 struct sched_group *sg = sd->groups;
6581 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6583 } while (sg != sd->groups);
6585 if (cpu != group_balance_cpu(sg))
6588 update_group_capacity(sd, cpu);
6589 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6593 * Check that the per-cpu provided sd energy data is consistent for all cpus
6596 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6597 const struct cpumask *cpumask)
6599 const struct sched_group_energy * const sge = fn(cpu);
6600 struct cpumask mask;
6603 if (cpumask_weight(cpumask) <= 1)
6606 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6608 for_each_cpu(i, &mask) {
6609 const struct sched_group_energy * const e = fn(i);
6612 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6614 for (y = 0; y < (e->nr_idle_states); y++) {
6615 BUG_ON(e->idle_states[y].power !=
6616 sge->idle_states[y].power);
6619 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6621 for (y = 0; y < (e->nr_cap_states); y++) {
6622 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6623 BUG_ON(e->cap_states[y].power !=
6624 sge->cap_states[y].power);
6629 static void init_sched_energy(int cpu, struct sched_domain *sd,
6630 sched_domain_energy_f fn)
6632 if (!(fn && fn(cpu)))
6635 if (cpu != group_balance_cpu(sd->groups))
6638 if (sd->child && !sd->child->groups->sge) {
6639 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6640 #ifdef CONFIG_SCHED_DEBUG
6641 pr_err(" energy data on %s but not on %s domain\n",
6642 sd->name, sd->child->name);
6647 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6649 sd->groups->sge = fn(cpu);
6653 * Initializers for schedule domains
6654 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6657 static int default_relax_domain_level = -1;
6658 int sched_domain_level_max;
6660 static int __init setup_relax_domain_level(char *str)
6662 if (kstrtoint(str, 0, &default_relax_domain_level))
6663 pr_warn("Unable to set relax_domain_level\n");
6667 __setup("relax_domain_level=", setup_relax_domain_level);
6669 static void set_domain_attribute(struct sched_domain *sd,
6670 struct sched_domain_attr *attr)
6674 if (!attr || attr->relax_domain_level < 0) {
6675 if (default_relax_domain_level < 0)
6678 request = default_relax_domain_level;
6680 request = attr->relax_domain_level;
6681 if (request < sd->level) {
6682 /* turn off idle balance on this domain */
6683 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6685 /* turn on idle balance on this domain */
6686 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6690 static void __sdt_free(const struct cpumask *cpu_map);
6691 static int __sdt_alloc(const struct cpumask *cpu_map);
6693 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6694 const struct cpumask *cpu_map)
6698 if (!atomic_read(&d->rd->refcount))
6699 free_rootdomain(&d->rd->rcu); /* fall through */
6701 free_percpu(d->sd); /* fall through */
6703 __sdt_free(cpu_map); /* fall through */
6709 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6710 const struct cpumask *cpu_map)
6712 memset(d, 0, sizeof(*d));
6714 if (__sdt_alloc(cpu_map))
6715 return sa_sd_storage;
6716 d->sd = alloc_percpu(struct sched_domain *);
6718 return sa_sd_storage;
6719 d->rd = alloc_rootdomain();
6722 return sa_rootdomain;
6726 * NULL the sd_data elements we've used to build the sched_domain and
6727 * sched_group structure so that the subsequent __free_domain_allocs()
6728 * will not free the data we're using.
6730 static void claim_allocations(int cpu, struct sched_domain *sd)
6732 struct sd_data *sdd = sd->private;
6734 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6735 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6737 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6738 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6740 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6741 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6745 static int sched_domains_numa_levels;
6746 enum numa_topology_type sched_numa_topology_type;
6747 static int *sched_domains_numa_distance;
6748 int sched_max_numa_distance;
6749 static struct cpumask ***sched_domains_numa_masks;
6750 static int sched_domains_curr_level;
6754 * SD_flags allowed in topology descriptions.
6756 * These flags are purely descriptive of the topology and do not prescribe
6757 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6760 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6761 * SD_SHARE_PKG_RESOURCES - describes shared caches
6762 * SD_NUMA - describes NUMA topologies
6763 * SD_SHARE_POWERDOMAIN - describes shared power domain
6764 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6765 * SD_SHARE_CAP_STATES - describes shared capacity states
6767 * Odd one out, which beside describing the topology has a quirk also
6768 * prescribes the desired behaviour that goes along with it:
6771 * SD_ASYM_PACKING - describes SMT quirks
6773 #define TOPOLOGY_SD_FLAGS \
6774 (SD_SHARE_CPUCAPACITY | \
6775 SD_SHARE_PKG_RESOURCES | \
6778 SD_ASYM_CPUCAPACITY | \
6779 SD_SHARE_POWERDOMAIN | \
6780 SD_SHARE_CAP_STATES)
6782 static struct sched_domain *
6783 sd_init(struct sched_domain_topology_level *tl,
6784 struct sched_domain *child, int cpu)
6786 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6787 int sd_weight, sd_flags = 0;
6791 * Ugly hack to pass state to sd_numa_mask()...
6793 sched_domains_curr_level = tl->numa_level;
6796 sd_weight = cpumask_weight(tl->mask(cpu));
6799 sd_flags = (*tl->sd_flags)();
6800 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6801 "wrong sd_flags in topology description\n"))
6802 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6804 *sd = (struct sched_domain){
6805 .min_interval = sd_weight,
6806 .max_interval = 2*sd_weight,
6808 .imbalance_pct = 125,
6810 .cache_nice_tries = 0,
6817 .flags = 1*SD_LOAD_BALANCE
6818 | 1*SD_BALANCE_NEWIDLE
6823 | 0*SD_SHARE_CPUCAPACITY
6824 | 0*SD_SHARE_PKG_RESOURCES
6826 | 0*SD_PREFER_SIBLING
6831 .last_balance = jiffies,
6832 .balance_interval = sd_weight,
6834 .max_newidle_lb_cost = 0,
6835 .next_decay_max_lb_cost = jiffies,
6837 #ifdef CONFIG_SCHED_DEBUG
6843 * Convert topological properties into behaviour.
6846 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6847 struct sched_domain *t = sd;
6849 for_each_lower_domain(t)
6850 t->flags |= SD_BALANCE_WAKE;
6853 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6854 sd->flags |= SD_PREFER_SIBLING;
6855 sd->imbalance_pct = 110;
6856 sd->smt_gain = 1178; /* ~15% */
6858 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6859 sd->imbalance_pct = 117;
6860 sd->cache_nice_tries = 1;
6864 } else if (sd->flags & SD_NUMA) {
6865 sd->cache_nice_tries = 2;
6869 sd->flags |= SD_SERIALIZE;
6870 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6871 sd->flags &= ~(SD_BALANCE_EXEC |
6878 sd->flags |= SD_PREFER_SIBLING;
6879 sd->cache_nice_tries = 1;
6884 sd->private = &tl->data;
6890 * Topology list, bottom-up.
6892 static struct sched_domain_topology_level default_topology[] = {
6893 #ifdef CONFIG_SCHED_SMT
6894 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6896 #ifdef CONFIG_SCHED_MC
6897 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6899 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6903 static struct sched_domain_topology_level *sched_domain_topology =
6906 #define for_each_sd_topology(tl) \
6907 for (tl = sched_domain_topology; tl->mask; tl++)
6909 void set_sched_topology(struct sched_domain_topology_level *tl)
6911 sched_domain_topology = tl;
6916 static const struct cpumask *sd_numa_mask(int cpu)
6918 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6921 static void sched_numa_warn(const char *str)
6923 static int done = false;
6931 printk(KERN_WARNING "ERROR: %s\n\n", str);
6933 for (i = 0; i < nr_node_ids; i++) {
6934 printk(KERN_WARNING " ");
6935 for (j = 0; j < nr_node_ids; j++)
6936 printk(KERN_CONT "%02d ", node_distance(i,j));
6937 printk(KERN_CONT "\n");
6939 printk(KERN_WARNING "\n");
6942 bool find_numa_distance(int distance)
6946 if (distance == node_distance(0, 0))
6949 for (i = 0; i < sched_domains_numa_levels; i++) {
6950 if (sched_domains_numa_distance[i] == distance)
6958 * A system can have three types of NUMA topology:
6959 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6960 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6961 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6963 * The difference between a glueless mesh topology and a backplane
6964 * topology lies in whether communication between not directly
6965 * connected nodes goes through intermediary nodes (where programs
6966 * could run), or through backplane controllers. This affects
6967 * placement of programs.
6969 * The type of topology can be discerned with the following tests:
6970 * - If the maximum distance between any nodes is 1 hop, the system
6971 * is directly connected.
6972 * - If for two nodes A and B, located N > 1 hops away from each other,
6973 * there is an intermediary node C, which is < N hops away from both
6974 * nodes A and B, the system is a glueless mesh.
6976 static void init_numa_topology_type(void)
6980 n = sched_max_numa_distance;
6982 if (sched_domains_numa_levels <= 1) {
6983 sched_numa_topology_type = NUMA_DIRECT;
6987 for_each_online_node(a) {
6988 for_each_online_node(b) {
6989 /* Find two nodes furthest removed from each other. */
6990 if (node_distance(a, b) < n)
6993 /* Is there an intermediary node between a and b? */
6994 for_each_online_node(c) {
6995 if (node_distance(a, c) < n &&
6996 node_distance(b, c) < n) {
6997 sched_numa_topology_type =
7003 sched_numa_topology_type = NUMA_BACKPLANE;
7009 static void sched_init_numa(void)
7011 int next_distance, curr_distance = node_distance(0, 0);
7012 struct sched_domain_topology_level *tl;
7016 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
7017 if (!sched_domains_numa_distance)
7021 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
7022 * unique distances in the node_distance() table.
7024 * Assumes node_distance(0,j) includes all distances in
7025 * node_distance(i,j) in order to avoid cubic time.
7027 next_distance = curr_distance;
7028 for (i = 0; i < nr_node_ids; i++) {
7029 for (j = 0; j < nr_node_ids; j++) {
7030 for (k = 0; k < nr_node_ids; k++) {
7031 int distance = node_distance(i, k);
7033 if (distance > curr_distance &&
7034 (distance < next_distance ||
7035 next_distance == curr_distance))
7036 next_distance = distance;
7039 * While not a strong assumption it would be nice to know
7040 * about cases where if node A is connected to B, B is not
7041 * equally connected to A.
7043 if (sched_debug() && node_distance(k, i) != distance)
7044 sched_numa_warn("Node-distance not symmetric");
7046 if (sched_debug() && i && !find_numa_distance(distance))
7047 sched_numa_warn("Node-0 not representative");
7049 if (next_distance != curr_distance) {
7050 sched_domains_numa_distance[level++] = next_distance;
7051 sched_domains_numa_levels = level;
7052 curr_distance = next_distance;
7057 * In case of sched_debug() we verify the above assumption.
7067 * 'level' contains the number of unique distances, excluding the
7068 * identity distance node_distance(i,i).
7070 * The sched_domains_numa_distance[] array includes the actual distance
7075 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7076 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7077 * the array will contain less then 'level' members. This could be
7078 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7079 * in other functions.
7081 * We reset it to 'level' at the end of this function.
7083 sched_domains_numa_levels = 0;
7085 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7086 if (!sched_domains_numa_masks)
7090 * Now for each level, construct a mask per node which contains all
7091 * cpus of nodes that are that many hops away from us.
7093 for (i = 0; i < level; i++) {
7094 sched_domains_numa_masks[i] =
7095 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7096 if (!sched_domains_numa_masks[i])
7099 for (j = 0; j < nr_node_ids; j++) {
7100 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7104 sched_domains_numa_masks[i][j] = mask;
7107 if (node_distance(j, k) > sched_domains_numa_distance[i])
7110 cpumask_or(mask, mask, cpumask_of_node(k));
7115 /* Compute default topology size */
7116 for (i = 0; sched_domain_topology[i].mask; i++);
7118 tl = kzalloc((i + level + 1) *
7119 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7124 * Copy the default topology bits..
7126 for (i = 0; sched_domain_topology[i].mask; i++)
7127 tl[i] = sched_domain_topology[i];
7130 * .. and append 'j' levels of NUMA goodness.
7132 for (j = 0; j < level; i++, j++) {
7133 tl[i] = (struct sched_domain_topology_level){
7134 .mask = sd_numa_mask,
7135 .sd_flags = cpu_numa_flags,
7136 .flags = SDTL_OVERLAP,
7142 sched_domain_topology = tl;
7144 sched_domains_numa_levels = level;
7145 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7147 init_numa_topology_type();
7150 static void sched_domains_numa_masks_set(int cpu)
7153 int node = cpu_to_node(cpu);
7155 for (i = 0; i < sched_domains_numa_levels; i++) {
7156 for (j = 0; j < nr_node_ids; j++) {
7157 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7158 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7163 static void sched_domains_numa_masks_clear(int cpu)
7166 for (i = 0; i < sched_domains_numa_levels; i++) {
7167 for (j = 0; j < nr_node_ids; j++)
7168 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7173 * Update sched_domains_numa_masks[level][node] array when new cpus
7176 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7177 unsigned long action,
7180 int cpu = (long)hcpu;
7182 switch (action & ~CPU_TASKS_FROZEN) {
7184 sched_domains_numa_masks_set(cpu);
7188 sched_domains_numa_masks_clear(cpu);
7198 static inline void sched_init_numa(void)
7202 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7203 unsigned long action,
7208 #endif /* CONFIG_NUMA */
7210 static int __sdt_alloc(const struct cpumask *cpu_map)
7212 struct sched_domain_topology_level *tl;
7215 for_each_sd_topology(tl) {
7216 struct sd_data *sdd = &tl->data;
7218 sdd->sd = alloc_percpu(struct sched_domain *);
7222 sdd->sg = alloc_percpu(struct sched_group *);
7226 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7230 for_each_cpu(j, cpu_map) {
7231 struct sched_domain *sd;
7232 struct sched_group *sg;
7233 struct sched_group_capacity *sgc;
7235 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7236 GFP_KERNEL, cpu_to_node(j));
7240 *per_cpu_ptr(sdd->sd, j) = sd;
7242 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7243 GFP_KERNEL, cpu_to_node(j));
7249 *per_cpu_ptr(sdd->sg, j) = sg;
7251 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7252 GFP_KERNEL, cpu_to_node(j));
7256 *per_cpu_ptr(sdd->sgc, j) = sgc;
7263 static void __sdt_free(const struct cpumask *cpu_map)
7265 struct sched_domain_topology_level *tl;
7268 for_each_sd_topology(tl) {
7269 struct sd_data *sdd = &tl->data;
7271 for_each_cpu(j, cpu_map) {
7272 struct sched_domain *sd;
7275 sd = *per_cpu_ptr(sdd->sd, j);
7276 if (sd && (sd->flags & SD_OVERLAP))
7277 free_sched_groups(sd->groups, 0);
7278 kfree(*per_cpu_ptr(sdd->sd, j));
7282 kfree(*per_cpu_ptr(sdd->sg, j));
7284 kfree(*per_cpu_ptr(sdd->sgc, j));
7286 free_percpu(sdd->sd);
7288 free_percpu(sdd->sg);
7290 free_percpu(sdd->sgc);
7295 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7296 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7297 struct sched_domain *child, int cpu)
7299 struct sched_domain *sd = sd_init(tl, child, cpu);
7301 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7303 sd->level = child->level + 1;
7304 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7307 if (!cpumask_subset(sched_domain_span(child),
7308 sched_domain_span(sd))) {
7309 pr_err("BUG: arch topology borken\n");
7310 #ifdef CONFIG_SCHED_DEBUG
7311 pr_err(" the %s domain not a subset of the %s domain\n",
7312 child->name, sd->name);
7314 /* Fixup, ensure @sd has at least @child cpus. */
7315 cpumask_or(sched_domain_span(sd),
7316 sched_domain_span(sd),
7317 sched_domain_span(child));
7321 set_domain_attribute(sd, attr);
7327 * Build sched domains for a given set of cpus and attach the sched domains
7328 * to the individual cpus
7330 static int build_sched_domains(const struct cpumask *cpu_map,
7331 struct sched_domain_attr *attr)
7333 enum s_alloc alloc_state;
7334 struct sched_domain *sd;
7336 int i, ret = -ENOMEM;
7338 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7339 if (alloc_state != sa_rootdomain)
7342 /* Set up domains for cpus specified by the cpu_map. */
7343 for_each_cpu(i, cpu_map) {
7344 struct sched_domain_topology_level *tl;
7347 for_each_sd_topology(tl) {
7348 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7349 if (tl == sched_domain_topology)
7350 *per_cpu_ptr(d.sd, i) = sd;
7351 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7352 sd->flags |= SD_OVERLAP;
7356 /* Build the groups for the domains */
7357 for_each_cpu(i, cpu_map) {
7358 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7359 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7360 if (sd->flags & SD_OVERLAP) {
7361 if (build_overlap_sched_groups(sd, i))
7364 if (build_sched_groups(sd, i))
7370 /* Calculate CPU capacity for physical packages and nodes */
7371 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7372 struct sched_domain_topology_level *tl = sched_domain_topology;
7374 if (!cpumask_test_cpu(i, cpu_map))
7377 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7378 init_sched_energy(i, sd, tl->energy);
7379 claim_allocations(i, sd);
7380 init_sched_groups_capacity(i, sd);
7384 /* Attach the domains */
7386 for_each_cpu(i, cpu_map) {
7387 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu);
7388 int min_cpu = READ_ONCE(d.rd->min_cap_orig_cpu);
7390 if ((max_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig >
7391 cpu_rq(max_cpu)->cpu_capacity_orig))
7392 WRITE_ONCE(d.rd->max_cap_orig_cpu, i);
7394 if ((min_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig <
7395 cpu_rq(min_cpu)->cpu_capacity_orig))
7396 WRITE_ONCE(d.rd->min_cap_orig_cpu, i);
7398 sd = *per_cpu_ptr(d.sd, i);
7400 cpu_attach_domain(sd, d.rd, i);
7406 __free_domain_allocs(&d, alloc_state, cpu_map);
7410 static cpumask_var_t *doms_cur; /* current sched domains */
7411 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7412 static struct sched_domain_attr *dattr_cur;
7413 /* attribues of custom domains in 'doms_cur' */
7416 * Special case: If a kmalloc of a doms_cur partition (array of
7417 * cpumask) fails, then fallback to a single sched domain,
7418 * as determined by the single cpumask fallback_doms.
7420 static cpumask_var_t fallback_doms;
7423 * arch_update_cpu_topology lets virtualized architectures update the
7424 * cpu core maps. It is supposed to return 1 if the topology changed
7425 * or 0 if it stayed the same.
7427 int __weak arch_update_cpu_topology(void)
7432 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7435 cpumask_var_t *doms;
7437 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7440 for (i = 0; i < ndoms; i++) {
7441 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7442 free_sched_domains(doms, i);
7449 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7452 for (i = 0; i < ndoms; i++)
7453 free_cpumask_var(doms[i]);
7458 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7459 * For now this just excludes isolated cpus, but could be used to
7460 * exclude other special cases in the future.
7462 static int init_sched_domains(const struct cpumask *cpu_map)
7466 arch_update_cpu_topology();
7468 doms_cur = alloc_sched_domains(ndoms_cur);
7470 doms_cur = &fallback_doms;
7471 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7472 err = build_sched_domains(doms_cur[0], NULL);
7473 register_sched_domain_sysctl();
7479 * Detach sched domains from a group of cpus specified in cpu_map
7480 * These cpus will now be attached to the NULL domain
7482 static void detach_destroy_domains(const struct cpumask *cpu_map)
7487 for_each_cpu(i, cpu_map)
7488 cpu_attach_domain(NULL, &def_root_domain, i);
7492 /* handle null as "default" */
7493 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7494 struct sched_domain_attr *new, int idx_new)
7496 struct sched_domain_attr tmp;
7503 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7504 new ? (new + idx_new) : &tmp,
7505 sizeof(struct sched_domain_attr));
7509 * Partition sched domains as specified by the 'ndoms_new'
7510 * cpumasks in the array doms_new[] of cpumasks. This compares
7511 * doms_new[] to the current sched domain partitioning, doms_cur[].
7512 * It destroys each deleted domain and builds each new domain.
7514 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7515 * The masks don't intersect (don't overlap.) We should setup one
7516 * sched domain for each mask. CPUs not in any of the cpumasks will
7517 * not be load balanced. If the same cpumask appears both in the
7518 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7521 * The passed in 'doms_new' should be allocated using
7522 * alloc_sched_domains. This routine takes ownership of it and will
7523 * free_sched_domains it when done with it. If the caller failed the
7524 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7525 * and partition_sched_domains() will fallback to the single partition
7526 * 'fallback_doms', it also forces the domains to be rebuilt.
7528 * If doms_new == NULL it will be replaced with cpu_online_mask.
7529 * ndoms_new == 0 is a special case for destroying existing domains,
7530 * and it will not create the default domain.
7532 * Call with hotplug lock held
7534 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7535 struct sched_domain_attr *dattr_new)
7540 mutex_lock(&sched_domains_mutex);
7542 /* always unregister in case we don't destroy any domains */
7543 unregister_sched_domain_sysctl();
7545 /* Let architecture update cpu core mappings. */
7546 new_topology = arch_update_cpu_topology();
7548 n = doms_new ? ndoms_new : 0;
7550 /* Destroy deleted domains */
7551 for (i = 0; i < ndoms_cur; i++) {
7552 for (j = 0; j < n && !new_topology; j++) {
7553 if (cpumask_equal(doms_cur[i], doms_new[j])
7554 && dattrs_equal(dattr_cur, i, dattr_new, j))
7557 /* no match - a current sched domain not in new doms_new[] */
7558 detach_destroy_domains(doms_cur[i]);
7564 if (doms_new == NULL) {
7566 doms_new = &fallback_doms;
7567 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7568 WARN_ON_ONCE(dattr_new);
7571 /* Build new domains */
7572 for (i = 0; i < ndoms_new; i++) {
7573 for (j = 0; j < n && !new_topology; j++) {
7574 if (cpumask_equal(doms_new[i], doms_cur[j])
7575 && dattrs_equal(dattr_new, i, dattr_cur, j))
7578 /* no match - add a new doms_new */
7579 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7584 /* Remember the new sched domains */
7585 if (doms_cur != &fallback_doms)
7586 free_sched_domains(doms_cur, ndoms_cur);
7587 kfree(dattr_cur); /* kfree(NULL) is safe */
7588 doms_cur = doms_new;
7589 dattr_cur = dattr_new;
7590 ndoms_cur = ndoms_new;
7592 register_sched_domain_sysctl();
7594 mutex_unlock(&sched_domains_mutex);
7597 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7600 * Update cpusets according to cpu_active mask. If cpusets are
7601 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7602 * around partition_sched_domains().
7604 * If we come here as part of a suspend/resume, don't touch cpusets because we
7605 * want to restore it back to its original state upon resume anyway.
7607 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7611 case CPU_ONLINE_FROZEN:
7612 case CPU_DOWN_FAILED_FROZEN:
7615 * num_cpus_frozen tracks how many CPUs are involved in suspend
7616 * resume sequence. As long as this is not the last online
7617 * operation in the resume sequence, just build a single sched
7618 * domain, ignoring cpusets.
7620 partition_sched_domains(1, NULL, NULL);
7621 if (--num_cpus_frozen)
7625 * This is the last CPU online operation. So fall through and
7626 * restore the original sched domains by considering the
7627 * cpuset configurations.
7629 cpuset_force_rebuild();
7632 cpuset_update_active_cpus(true);
7640 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7643 unsigned long flags;
7644 long cpu = (long)hcpu;
7650 case CPU_DOWN_PREPARE:
7651 rcu_read_lock_sched();
7652 dl_b = dl_bw_of(cpu);
7654 raw_spin_lock_irqsave(&dl_b->lock, flags);
7655 cpus = dl_bw_cpus(cpu);
7656 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7657 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7659 rcu_read_unlock_sched();
7662 return notifier_from_errno(-EBUSY);
7663 cpuset_update_active_cpus(false);
7665 case CPU_DOWN_PREPARE_FROZEN:
7667 partition_sched_domains(1, NULL, NULL);
7675 void __init sched_init_smp(void)
7677 cpumask_var_t non_isolated_cpus;
7679 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7680 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7685 * There's no userspace yet to cause hotplug operations; hence all the
7686 * cpu masks are stable and all blatant races in the below code cannot
7689 mutex_lock(&sched_domains_mutex);
7690 init_sched_domains(cpu_active_mask);
7691 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7692 if (cpumask_empty(non_isolated_cpus))
7693 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7694 mutex_unlock(&sched_domains_mutex);
7696 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7697 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7698 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7702 /* Move init over to a non-isolated CPU */
7703 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7705 sched_init_granularity();
7706 free_cpumask_var(non_isolated_cpus);
7708 init_sched_rt_class();
7709 init_sched_dl_class();
7712 void __init sched_init_smp(void)
7714 sched_init_granularity();
7716 #endif /* CONFIG_SMP */
7718 int in_sched_functions(unsigned long addr)
7720 return in_lock_functions(addr) ||
7721 (addr >= (unsigned long)__sched_text_start
7722 && addr < (unsigned long)__sched_text_end);
7725 #ifdef CONFIG_CGROUP_SCHED
7727 * Default task group.
7728 * Every task in system belongs to this group at bootup.
7730 struct task_group root_task_group;
7731 LIST_HEAD(task_groups);
7734 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7736 void __init sched_init(void)
7739 unsigned long alloc_size = 0, ptr;
7741 #ifdef CONFIG_FAIR_GROUP_SCHED
7742 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7744 #ifdef CONFIG_RT_GROUP_SCHED
7745 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7748 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7750 #ifdef CONFIG_FAIR_GROUP_SCHED
7751 root_task_group.se = (struct sched_entity **)ptr;
7752 ptr += nr_cpu_ids * sizeof(void **);
7754 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7755 ptr += nr_cpu_ids * sizeof(void **);
7757 #endif /* CONFIG_FAIR_GROUP_SCHED */
7758 #ifdef CONFIG_RT_GROUP_SCHED
7759 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7760 ptr += nr_cpu_ids * sizeof(void **);
7762 root_task_group.rt_rq = (struct rt_rq **)ptr;
7763 ptr += nr_cpu_ids * sizeof(void **);
7765 #endif /* CONFIG_RT_GROUP_SCHED */
7767 #ifdef CONFIG_CPUMASK_OFFSTACK
7768 for_each_possible_cpu(i) {
7769 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7770 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7772 #endif /* CONFIG_CPUMASK_OFFSTACK */
7774 init_rt_bandwidth(&def_rt_bandwidth,
7775 global_rt_period(), global_rt_runtime());
7776 init_dl_bandwidth(&def_dl_bandwidth,
7777 global_rt_period(), global_rt_runtime());
7780 init_defrootdomain();
7783 #ifdef CONFIG_RT_GROUP_SCHED
7784 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7785 global_rt_period(), global_rt_runtime());
7786 #endif /* CONFIG_RT_GROUP_SCHED */
7788 #ifdef CONFIG_CGROUP_SCHED
7789 list_add(&root_task_group.list, &task_groups);
7790 INIT_LIST_HEAD(&root_task_group.children);
7791 INIT_LIST_HEAD(&root_task_group.siblings);
7792 autogroup_init(&init_task);
7794 #endif /* CONFIG_CGROUP_SCHED */
7796 for_each_possible_cpu(i) {
7800 raw_spin_lock_init(&rq->lock);
7802 rq->calc_load_active = 0;
7803 rq->calc_load_update = jiffies + LOAD_FREQ;
7804 init_cfs_rq(&rq->cfs);
7805 init_rt_rq(&rq->rt);
7806 init_dl_rq(&rq->dl);
7807 #ifdef CONFIG_FAIR_GROUP_SCHED
7808 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7809 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7810 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7812 * How much cpu bandwidth does root_task_group get?
7814 * In case of task-groups formed thr' the cgroup filesystem, it
7815 * gets 100% of the cpu resources in the system. This overall
7816 * system cpu resource is divided among the tasks of
7817 * root_task_group and its child task-groups in a fair manner,
7818 * based on each entity's (task or task-group's) weight
7819 * (se->load.weight).
7821 * In other words, if root_task_group has 10 tasks of weight
7822 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7823 * then A0's share of the cpu resource is:
7825 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7827 * We achieve this by letting root_task_group's tasks sit
7828 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7830 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7831 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7832 #endif /* CONFIG_FAIR_GROUP_SCHED */
7834 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7835 #ifdef CONFIG_RT_GROUP_SCHED
7836 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7839 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7840 rq->cpu_load[j] = 0;
7842 rq->last_load_update_tick = jiffies;
7847 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7848 rq->balance_callback = NULL;
7849 rq->active_balance = 0;
7850 rq->next_balance = jiffies;
7852 rq->push_task = NULL;
7856 rq->avg_idle = 2*sysctl_sched_migration_cost;
7857 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7858 #ifdef CONFIG_SCHED_WALT
7859 rq->cur_irqload = 0;
7860 rq->avg_irqload = 0;
7864 INIT_LIST_HEAD(&rq->cfs_tasks);
7866 rq_attach_root(rq, &def_root_domain);
7867 #ifdef CONFIG_NO_HZ_COMMON
7870 #ifdef CONFIG_NO_HZ_FULL
7871 rq->last_sched_tick = 0;
7875 atomic_set(&rq->nr_iowait, 0);
7878 set_load_weight(&init_task);
7880 #ifdef CONFIG_PREEMPT_NOTIFIERS
7881 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7885 * The boot idle thread does lazy MMU switching as well:
7887 atomic_inc(&init_mm.mm_count);
7888 enter_lazy_tlb(&init_mm, current);
7891 * During early bootup we pretend to be a normal task:
7893 current->sched_class = &fair_sched_class;
7896 * Make us the idle thread. Technically, schedule() should not be
7897 * called from this thread, however somewhere below it might be,
7898 * but because we are the idle thread, we just pick up running again
7899 * when this runqueue becomes "idle".
7901 init_idle(current, smp_processor_id());
7903 calc_load_update = jiffies + LOAD_FREQ;
7906 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7907 /* May be allocated at isolcpus cmdline parse time */
7908 if (cpu_isolated_map == NULL)
7909 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7910 idle_thread_set_boot_cpu();
7911 set_cpu_rq_start_time();
7913 init_sched_fair_class();
7915 scheduler_running = 1;
7918 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7919 static inline int preempt_count_equals(int preempt_offset)
7921 int nested = preempt_count() + rcu_preempt_depth();
7923 return (nested == preempt_offset);
7926 static int __might_sleep_init_called;
7927 int __init __might_sleep_init(void)
7929 __might_sleep_init_called = 1;
7932 early_initcall(__might_sleep_init);
7934 void __might_sleep(const char *file, int line, int preempt_offset)
7937 * Blocking primitives will set (and therefore destroy) current->state,
7938 * since we will exit with TASK_RUNNING make sure we enter with it,
7939 * otherwise we will destroy state.
7941 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7942 "do not call blocking ops when !TASK_RUNNING; "
7943 "state=%lx set at [<%p>] %pS\n",
7945 (void *)current->task_state_change,
7946 (void *)current->task_state_change);
7948 ___might_sleep(file, line, preempt_offset);
7950 EXPORT_SYMBOL(__might_sleep);
7952 void ___might_sleep(const char *file, int line, int preempt_offset)
7954 static unsigned long prev_jiffy; /* ratelimiting */
7956 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7957 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7958 !is_idle_task(current)) || oops_in_progress)
7960 if (system_state != SYSTEM_RUNNING &&
7961 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7963 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7965 prev_jiffy = jiffies;
7968 "BUG: sleeping function called from invalid context at %s:%d\n",
7971 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7972 in_atomic(), irqs_disabled(),
7973 current->pid, current->comm);
7975 if (task_stack_end_corrupted(current))
7976 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7978 debug_show_held_locks(current);
7979 if (irqs_disabled())
7980 print_irqtrace_events(current);
7981 #ifdef CONFIG_DEBUG_PREEMPT
7982 if (!preempt_count_equals(preempt_offset)) {
7983 pr_err("Preemption disabled at:");
7984 print_ip_sym(current->preempt_disable_ip);
7990 EXPORT_SYMBOL(___might_sleep);
7993 #ifdef CONFIG_MAGIC_SYSRQ
7994 void normalize_rt_tasks(void)
7996 struct task_struct *g, *p;
7997 struct sched_attr attr = {
7998 .sched_policy = SCHED_NORMAL,
8001 read_lock(&tasklist_lock);
8002 for_each_process_thread(g, p) {
8004 * Only normalize user tasks:
8006 if (p->flags & PF_KTHREAD)
8009 p->se.exec_start = 0;
8010 #ifdef CONFIG_SCHEDSTATS
8011 p->se.statistics.wait_start = 0;
8012 p->se.statistics.sleep_start = 0;
8013 p->se.statistics.block_start = 0;
8016 if (!dl_task(p) && !rt_task(p)) {
8018 * Renice negative nice level userspace
8021 if (task_nice(p) < 0)
8022 set_user_nice(p, 0);
8026 __sched_setscheduler(p, &attr, false, false);
8028 read_unlock(&tasklist_lock);
8031 #endif /* CONFIG_MAGIC_SYSRQ */
8033 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8035 * These functions are only useful for the IA64 MCA handling, or kdb.
8037 * They can only be called when the whole system has been
8038 * stopped - every CPU needs to be quiescent, and no scheduling
8039 * activity can take place. Using them for anything else would
8040 * be a serious bug, and as a result, they aren't even visible
8041 * under any other configuration.
8045 * curr_task - return the current task for a given cpu.
8046 * @cpu: the processor in question.
8048 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8050 * Return: The current task for @cpu.
8052 struct task_struct *curr_task(int cpu)
8054 return cpu_curr(cpu);
8057 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8061 * set_curr_task - set the current task for a given cpu.
8062 * @cpu: the processor in question.
8063 * @p: the task pointer to set.
8065 * Description: This function must only be used when non-maskable interrupts
8066 * are serviced on a separate stack. It allows the architecture to switch the
8067 * notion of the current task on a cpu in a non-blocking manner. This function
8068 * must be called with all CPU's synchronized, and interrupts disabled, the
8069 * and caller must save the original value of the current task (see
8070 * curr_task() above) and restore that value before reenabling interrupts and
8071 * re-starting the system.
8073 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8075 void set_curr_task(int cpu, struct task_struct *p)
8082 #ifdef CONFIG_CGROUP_SCHED
8083 /* task_group_lock serializes the addition/removal of task groups */
8084 static DEFINE_SPINLOCK(task_group_lock);
8086 static void sched_free_group(struct task_group *tg)
8088 free_fair_sched_group(tg);
8089 free_rt_sched_group(tg);
8094 /* allocate runqueue etc for a new task group */
8095 struct task_group *sched_create_group(struct task_group *parent)
8097 struct task_group *tg;
8099 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8101 return ERR_PTR(-ENOMEM);
8103 if (!alloc_fair_sched_group(tg, parent))
8106 if (!alloc_rt_sched_group(tg, parent))
8112 sched_free_group(tg);
8113 return ERR_PTR(-ENOMEM);
8116 void sched_online_group(struct task_group *tg, struct task_group *parent)
8118 unsigned long flags;
8120 spin_lock_irqsave(&task_group_lock, flags);
8121 list_add_rcu(&tg->list, &task_groups);
8123 WARN_ON(!parent); /* root should already exist */
8125 tg->parent = parent;
8126 INIT_LIST_HEAD(&tg->children);
8127 list_add_rcu(&tg->siblings, &parent->children);
8128 spin_unlock_irqrestore(&task_group_lock, flags);
8131 /* rcu callback to free various structures associated with a task group */
8132 static void sched_free_group_rcu(struct rcu_head *rhp)
8134 /* now it should be safe to free those cfs_rqs */
8135 sched_free_group(container_of(rhp, struct task_group, rcu));
8138 void sched_destroy_group(struct task_group *tg)
8140 /* wait for possible concurrent references to cfs_rqs complete */
8141 call_rcu(&tg->rcu, sched_free_group_rcu);
8144 void sched_offline_group(struct task_group *tg)
8146 unsigned long flags;
8148 /* end participation in shares distribution */
8149 unregister_fair_sched_group(tg);
8151 spin_lock_irqsave(&task_group_lock, flags);
8152 list_del_rcu(&tg->list);
8153 list_del_rcu(&tg->siblings);
8154 spin_unlock_irqrestore(&task_group_lock, flags);
8157 static void sched_change_group(struct task_struct *tsk, int type)
8159 struct task_group *tg;
8162 * All callers are synchronized by task_rq_lock(); we do not use RCU
8163 * which is pointless here. Thus, we pass "true" to task_css_check()
8164 * to prevent lockdep warnings.
8166 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8167 struct task_group, css);
8168 tg = autogroup_task_group(tsk, tg);
8169 tsk->sched_task_group = tg;
8171 #ifdef CONFIG_FAIR_GROUP_SCHED
8172 if (tsk->sched_class->task_change_group)
8173 tsk->sched_class->task_change_group(tsk, type);
8176 set_task_rq(tsk, task_cpu(tsk));
8180 * Change task's runqueue when it moves between groups.
8182 * The caller of this function should have put the task in its new group by
8183 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8186 void sched_move_task(struct task_struct *tsk)
8188 int queued, running;
8189 unsigned long flags;
8192 rq = task_rq_lock(tsk, &flags);
8194 running = task_current(rq, tsk);
8195 queued = task_on_rq_queued(tsk);
8198 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8199 if (unlikely(running))
8200 put_prev_task(rq, tsk);
8202 sched_change_group(tsk, TASK_MOVE_GROUP);
8204 if (unlikely(running))
8205 tsk->sched_class->set_curr_task(rq);
8207 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8209 task_rq_unlock(rq, tsk, &flags);
8211 #endif /* CONFIG_CGROUP_SCHED */
8213 #ifdef CONFIG_RT_GROUP_SCHED
8215 * Ensure that the real time constraints are schedulable.
8217 static DEFINE_MUTEX(rt_constraints_mutex);
8219 /* Must be called with tasklist_lock held */
8220 static inline int tg_has_rt_tasks(struct task_group *tg)
8222 struct task_struct *g, *p;
8225 * Autogroups do not have RT tasks; see autogroup_create().
8227 if (task_group_is_autogroup(tg))
8230 for_each_process_thread(g, p) {
8231 if (rt_task(p) && task_group(p) == tg)
8238 struct rt_schedulable_data {
8239 struct task_group *tg;
8244 static int tg_rt_schedulable(struct task_group *tg, void *data)
8246 struct rt_schedulable_data *d = data;
8247 struct task_group *child;
8248 unsigned long total, sum = 0;
8249 u64 period, runtime;
8251 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8252 runtime = tg->rt_bandwidth.rt_runtime;
8255 period = d->rt_period;
8256 runtime = d->rt_runtime;
8260 * Cannot have more runtime than the period.
8262 if (runtime > period && runtime != RUNTIME_INF)
8266 * Ensure we don't starve existing RT tasks.
8268 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8271 total = to_ratio(period, runtime);
8274 * Nobody can have more than the global setting allows.
8276 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8280 * The sum of our children's runtime should not exceed our own.
8282 list_for_each_entry_rcu(child, &tg->children, siblings) {
8283 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8284 runtime = child->rt_bandwidth.rt_runtime;
8286 if (child == d->tg) {
8287 period = d->rt_period;
8288 runtime = d->rt_runtime;
8291 sum += to_ratio(period, runtime);
8300 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8304 struct rt_schedulable_data data = {
8306 .rt_period = period,
8307 .rt_runtime = runtime,
8311 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8317 static int tg_set_rt_bandwidth(struct task_group *tg,
8318 u64 rt_period, u64 rt_runtime)
8323 * Disallowing the root group RT runtime is BAD, it would disallow the
8324 * kernel creating (and or operating) RT threads.
8326 if (tg == &root_task_group && rt_runtime == 0)
8329 /* No period doesn't make any sense. */
8333 mutex_lock(&rt_constraints_mutex);
8334 read_lock(&tasklist_lock);
8335 err = __rt_schedulable(tg, rt_period, rt_runtime);
8339 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8340 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8341 tg->rt_bandwidth.rt_runtime = rt_runtime;
8343 for_each_possible_cpu(i) {
8344 struct rt_rq *rt_rq = tg->rt_rq[i];
8346 raw_spin_lock(&rt_rq->rt_runtime_lock);
8347 rt_rq->rt_runtime = rt_runtime;
8348 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8350 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8352 read_unlock(&tasklist_lock);
8353 mutex_unlock(&rt_constraints_mutex);
8358 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8360 u64 rt_runtime, rt_period;
8362 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8363 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8364 if (rt_runtime_us < 0)
8365 rt_runtime = RUNTIME_INF;
8367 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8370 static long sched_group_rt_runtime(struct task_group *tg)
8374 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8377 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8378 do_div(rt_runtime_us, NSEC_PER_USEC);
8379 return rt_runtime_us;
8382 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8384 u64 rt_runtime, rt_period;
8386 rt_period = rt_period_us * NSEC_PER_USEC;
8387 rt_runtime = tg->rt_bandwidth.rt_runtime;
8389 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8392 static long sched_group_rt_period(struct task_group *tg)
8396 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8397 do_div(rt_period_us, NSEC_PER_USEC);
8398 return rt_period_us;
8400 #endif /* CONFIG_RT_GROUP_SCHED */
8402 #ifdef CONFIG_RT_GROUP_SCHED
8403 static int sched_rt_global_constraints(void)
8407 mutex_lock(&rt_constraints_mutex);
8408 read_lock(&tasklist_lock);
8409 ret = __rt_schedulable(NULL, 0, 0);
8410 read_unlock(&tasklist_lock);
8411 mutex_unlock(&rt_constraints_mutex);
8416 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8418 /* Don't accept realtime tasks when there is no way for them to run */
8419 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8425 #else /* !CONFIG_RT_GROUP_SCHED */
8426 static int sched_rt_global_constraints(void)
8428 unsigned long flags;
8431 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8432 for_each_possible_cpu(i) {
8433 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8435 raw_spin_lock(&rt_rq->rt_runtime_lock);
8436 rt_rq->rt_runtime = global_rt_runtime();
8437 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8439 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8443 #endif /* CONFIG_RT_GROUP_SCHED */
8445 static int sched_dl_global_validate(void)
8447 u64 runtime = global_rt_runtime();
8448 u64 period = global_rt_period();
8449 u64 new_bw = to_ratio(period, runtime);
8452 unsigned long flags;
8455 * Here we want to check the bandwidth not being set to some
8456 * value smaller than the currently allocated bandwidth in
8457 * any of the root_domains.
8459 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8460 * cycling on root_domains... Discussion on different/better
8461 * solutions is welcome!
8463 for_each_possible_cpu(cpu) {
8464 rcu_read_lock_sched();
8465 dl_b = dl_bw_of(cpu);
8467 raw_spin_lock_irqsave(&dl_b->lock, flags);
8468 if (new_bw < dl_b->total_bw)
8470 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8472 rcu_read_unlock_sched();
8481 static void sched_dl_do_global(void)
8486 unsigned long flags;
8488 def_dl_bandwidth.dl_period = global_rt_period();
8489 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8491 if (global_rt_runtime() != RUNTIME_INF)
8492 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8495 * FIXME: As above...
8497 for_each_possible_cpu(cpu) {
8498 rcu_read_lock_sched();
8499 dl_b = dl_bw_of(cpu);
8501 raw_spin_lock_irqsave(&dl_b->lock, flags);
8503 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8505 rcu_read_unlock_sched();
8509 static int sched_rt_global_validate(void)
8511 if (sysctl_sched_rt_period <= 0)
8514 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8515 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8521 static void sched_rt_do_global(void)
8523 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8524 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8527 int sched_rt_handler(struct ctl_table *table, int write,
8528 void __user *buffer, size_t *lenp,
8531 int old_period, old_runtime;
8532 static DEFINE_MUTEX(mutex);
8536 old_period = sysctl_sched_rt_period;
8537 old_runtime = sysctl_sched_rt_runtime;
8539 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8541 if (!ret && write) {
8542 ret = sched_rt_global_validate();
8546 ret = sched_dl_global_validate();
8550 ret = sched_rt_global_constraints();
8554 sched_rt_do_global();
8555 sched_dl_do_global();
8559 sysctl_sched_rt_period = old_period;
8560 sysctl_sched_rt_runtime = old_runtime;
8562 mutex_unlock(&mutex);
8567 int sched_rr_handler(struct ctl_table *table, int write,
8568 void __user *buffer, size_t *lenp,
8572 static DEFINE_MUTEX(mutex);
8575 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8576 /* make sure that internally we keep jiffies */
8577 /* also, writing zero resets timeslice to default */
8578 if (!ret && write) {
8579 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8580 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8582 mutex_unlock(&mutex);
8586 #ifdef CONFIG_CGROUP_SCHED
8588 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8590 return css ? container_of(css, struct task_group, css) : NULL;
8593 static struct cgroup_subsys_state *
8594 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8596 struct task_group *parent = css_tg(parent_css);
8597 struct task_group *tg;
8600 /* This is early initialization for the top cgroup */
8601 return &root_task_group.css;
8604 tg = sched_create_group(parent);
8606 return ERR_PTR(-ENOMEM);
8611 /* Expose task group only after completing cgroup initialization */
8612 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8614 struct task_group *tg = css_tg(css);
8615 struct task_group *parent = css_tg(css->parent);
8618 sched_online_group(tg, parent);
8622 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8624 struct task_group *tg = css_tg(css);
8626 sched_offline_group(tg);
8629 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8631 struct task_group *tg = css_tg(css);
8634 * Relies on the RCU grace period between css_released() and this.
8636 sched_free_group(tg);
8640 * This is called before wake_up_new_task(), therefore we really only
8641 * have to set its group bits, all the other stuff does not apply.
8643 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8645 unsigned long flags;
8648 rq = task_rq_lock(task, &flags);
8650 update_rq_clock(rq);
8651 sched_change_group(task, TASK_SET_GROUP);
8653 task_rq_unlock(rq, task, &flags);
8656 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8658 struct task_struct *task;
8659 struct cgroup_subsys_state *css;
8662 cgroup_taskset_for_each(task, css, tset) {
8663 #ifdef CONFIG_RT_GROUP_SCHED
8664 if (!sched_rt_can_attach(css_tg(css), task))
8667 /* We don't support RT-tasks being in separate groups */
8668 if (task->sched_class != &fair_sched_class)
8672 * Serialize against wake_up_new_task() such that if its
8673 * running, we're sure to observe its full state.
8675 raw_spin_lock_irq(&task->pi_lock);
8677 * Avoid calling sched_move_task() before wake_up_new_task()
8678 * has happened. This would lead to problems with PELT, due to
8679 * move wanting to detach+attach while we're not attached yet.
8681 if (task->state == TASK_NEW)
8683 raw_spin_unlock_irq(&task->pi_lock);
8691 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8693 struct task_struct *task;
8694 struct cgroup_subsys_state *css;
8696 cgroup_taskset_for_each(task, css, tset)
8697 sched_move_task(task);
8700 #ifdef CONFIG_FAIR_GROUP_SCHED
8701 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8702 struct cftype *cftype, u64 shareval)
8704 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8707 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8710 struct task_group *tg = css_tg(css);
8712 return (u64) scale_load_down(tg->shares);
8715 #ifdef CONFIG_CFS_BANDWIDTH
8716 static DEFINE_MUTEX(cfs_constraints_mutex);
8718 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8719 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8721 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8723 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8725 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8726 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8728 if (tg == &root_task_group)
8732 * Ensure we have at some amount of bandwidth every period. This is
8733 * to prevent reaching a state of large arrears when throttled via
8734 * entity_tick() resulting in prolonged exit starvation.
8736 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8740 * Likewise, bound things on the otherside by preventing insane quota
8741 * periods. This also allows us to normalize in computing quota
8744 if (period > max_cfs_quota_period)
8748 * Prevent race between setting of cfs_rq->runtime_enabled and
8749 * unthrottle_offline_cfs_rqs().
8752 mutex_lock(&cfs_constraints_mutex);
8753 ret = __cfs_schedulable(tg, period, quota);
8757 runtime_enabled = quota != RUNTIME_INF;
8758 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8760 * If we need to toggle cfs_bandwidth_used, off->on must occur
8761 * before making related changes, and on->off must occur afterwards
8763 if (runtime_enabled && !runtime_was_enabled)
8764 cfs_bandwidth_usage_inc();
8765 raw_spin_lock_irq(&cfs_b->lock);
8766 cfs_b->period = ns_to_ktime(period);
8767 cfs_b->quota = quota;
8769 __refill_cfs_bandwidth_runtime(cfs_b);
8770 /* restart the period timer (if active) to handle new period expiry */
8771 if (runtime_enabled)
8772 start_cfs_bandwidth(cfs_b);
8773 raw_spin_unlock_irq(&cfs_b->lock);
8775 for_each_online_cpu(i) {
8776 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8777 struct rq *rq = cfs_rq->rq;
8779 raw_spin_lock_irq(&rq->lock);
8780 cfs_rq->runtime_enabled = runtime_enabled;
8781 cfs_rq->runtime_remaining = 0;
8783 if (cfs_rq->throttled)
8784 unthrottle_cfs_rq(cfs_rq);
8785 raw_spin_unlock_irq(&rq->lock);
8787 if (runtime_was_enabled && !runtime_enabled)
8788 cfs_bandwidth_usage_dec();
8790 mutex_unlock(&cfs_constraints_mutex);
8796 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8800 period = ktime_to_ns(tg->cfs_bandwidth.period);
8801 if (cfs_quota_us < 0)
8802 quota = RUNTIME_INF;
8804 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8806 return tg_set_cfs_bandwidth(tg, period, quota);
8809 long tg_get_cfs_quota(struct task_group *tg)
8813 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8816 quota_us = tg->cfs_bandwidth.quota;
8817 do_div(quota_us, NSEC_PER_USEC);
8822 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8826 period = (u64)cfs_period_us * NSEC_PER_USEC;
8827 quota = tg->cfs_bandwidth.quota;
8829 return tg_set_cfs_bandwidth(tg, period, quota);
8832 long tg_get_cfs_period(struct task_group *tg)
8836 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8837 do_div(cfs_period_us, NSEC_PER_USEC);
8839 return cfs_period_us;
8842 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8845 return tg_get_cfs_quota(css_tg(css));
8848 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8849 struct cftype *cftype, s64 cfs_quota_us)
8851 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8854 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8857 return tg_get_cfs_period(css_tg(css));
8860 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8861 struct cftype *cftype, u64 cfs_period_us)
8863 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8866 struct cfs_schedulable_data {
8867 struct task_group *tg;
8872 * normalize group quota/period to be quota/max_period
8873 * note: units are usecs
8875 static u64 normalize_cfs_quota(struct task_group *tg,
8876 struct cfs_schedulable_data *d)
8884 period = tg_get_cfs_period(tg);
8885 quota = tg_get_cfs_quota(tg);
8888 /* note: these should typically be equivalent */
8889 if (quota == RUNTIME_INF || quota == -1)
8892 return to_ratio(period, quota);
8895 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8897 struct cfs_schedulable_data *d = data;
8898 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8899 s64 quota = 0, parent_quota = -1;
8902 quota = RUNTIME_INF;
8904 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8906 quota = normalize_cfs_quota(tg, d);
8907 parent_quota = parent_b->hierarchical_quota;
8910 * ensure max(child_quota) <= parent_quota, inherit when no
8913 if (quota == RUNTIME_INF)
8914 quota = parent_quota;
8915 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8918 cfs_b->hierarchical_quota = quota;
8923 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8926 struct cfs_schedulable_data data = {
8932 if (quota != RUNTIME_INF) {
8933 do_div(data.period, NSEC_PER_USEC);
8934 do_div(data.quota, NSEC_PER_USEC);
8938 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8944 static int cpu_stats_show(struct seq_file *sf, void *v)
8946 struct task_group *tg = css_tg(seq_css(sf));
8947 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8949 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8950 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8951 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8955 #endif /* CONFIG_CFS_BANDWIDTH */
8956 #endif /* CONFIG_FAIR_GROUP_SCHED */
8958 #ifdef CONFIG_RT_GROUP_SCHED
8959 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8960 struct cftype *cft, s64 val)
8962 return sched_group_set_rt_runtime(css_tg(css), val);
8965 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8968 return sched_group_rt_runtime(css_tg(css));
8971 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8972 struct cftype *cftype, u64 rt_period_us)
8974 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8977 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8980 return sched_group_rt_period(css_tg(css));
8982 #endif /* CONFIG_RT_GROUP_SCHED */
8984 static struct cftype cpu_files[] = {
8985 #ifdef CONFIG_FAIR_GROUP_SCHED
8988 .read_u64 = cpu_shares_read_u64,
8989 .write_u64 = cpu_shares_write_u64,
8992 #ifdef CONFIG_CFS_BANDWIDTH
8994 .name = "cfs_quota_us",
8995 .read_s64 = cpu_cfs_quota_read_s64,
8996 .write_s64 = cpu_cfs_quota_write_s64,
8999 .name = "cfs_period_us",
9000 .read_u64 = cpu_cfs_period_read_u64,
9001 .write_u64 = cpu_cfs_period_write_u64,
9005 .seq_show = cpu_stats_show,
9008 #ifdef CONFIG_RT_GROUP_SCHED
9010 .name = "rt_runtime_us",
9011 .read_s64 = cpu_rt_runtime_read,
9012 .write_s64 = cpu_rt_runtime_write,
9015 .name = "rt_period_us",
9016 .read_u64 = cpu_rt_period_read_uint,
9017 .write_u64 = cpu_rt_period_write_uint,
9023 struct cgroup_subsys cpu_cgrp_subsys = {
9024 .css_alloc = cpu_cgroup_css_alloc,
9025 .css_online = cpu_cgroup_css_online,
9026 .css_released = cpu_cgroup_css_released,
9027 .css_free = cpu_cgroup_css_free,
9028 .fork = cpu_cgroup_fork,
9029 .can_attach = cpu_cgroup_can_attach,
9030 .attach = cpu_cgroup_attach,
9031 .legacy_cftypes = cpu_files,
9035 #endif /* CONFIG_CGROUP_SCHED */
9037 void dump_cpu_task(int cpu)
9039 pr_info("Task dump for CPU %d:\n", cpu);
9040 sched_show_task(cpu_curr(cpu));