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>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
94 DEFINE_MUTEX(sched_domains_mutex);
95 DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
97 static void update_rq_clock_task(struct rq *rq, s64 delta);
99 void update_rq_clock(struct rq *rq)
103 lockdep_assert_held(&rq->lock);
105 if (rq->clock_skip_update & RQCF_ACT_SKIP)
108 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
112 update_rq_clock_task(rq, delta);
116 * Debugging: various feature bits
119 #define SCHED_FEAT(name, enabled) \
120 (1UL << __SCHED_FEAT_##name) * enabled |
122 const_debug unsigned int sysctl_sched_features =
123 #include "features.h"
128 #ifdef CONFIG_SCHED_DEBUG
129 #define SCHED_FEAT(name, enabled) \
132 static const char * const sched_feat_names[] = {
133 #include "features.h"
138 static int sched_feat_show(struct seq_file *m, void *v)
142 for (i = 0; i < __SCHED_FEAT_NR; i++) {
143 if (!(sysctl_sched_features & (1UL << i)))
145 seq_printf(m, "%s ", sched_feat_names[i]);
152 #ifdef HAVE_JUMP_LABEL
154 #define jump_label_key__true STATIC_KEY_INIT_TRUE
155 #define jump_label_key__false STATIC_KEY_INIT_FALSE
157 #define SCHED_FEAT(name, enabled) \
158 jump_label_key__##enabled ,
160 struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
161 #include "features.h"
166 static void sched_feat_disable(int i)
168 static_key_disable(&sched_feat_keys[i]);
171 static void sched_feat_enable(int i)
173 static_key_enable(&sched_feat_keys[i]);
176 static void sched_feat_disable(int i) { };
177 static void sched_feat_enable(int i) { };
178 #endif /* HAVE_JUMP_LABEL */
180 static int sched_feat_set(char *cmp)
185 if (strncmp(cmp, "NO_", 3) == 0) {
190 for (i = 0; i < __SCHED_FEAT_NR; i++) {
191 if (strcmp(cmp, sched_feat_names[i]) == 0) {
193 sysctl_sched_features &= ~(1UL << i);
194 sched_feat_disable(i);
196 sysctl_sched_features |= (1UL << i);
197 sched_feat_enable(i);
207 sched_feat_write(struct file *filp, const char __user *ubuf,
208 size_t cnt, loff_t *ppos)
218 if (copy_from_user(&buf, ubuf, cnt))
224 /* Ensure the static_key remains in a consistent state */
225 inode = file_inode(filp);
226 mutex_lock(&inode->i_mutex);
227 i = sched_feat_set(cmp);
228 mutex_unlock(&inode->i_mutex);
229 if (i == __SCHED_FEAT_NR)
237 static int sched_feat_open(struct inode *inode, struct file *filp)
239 return single_open(filp, sched_feat_show, NULL);
242 static const struct file_operations sched_feat_fops = {
243 .open = sched_feat_open,
244 .write = sched_feat_write,
247 .release = single_release,
250 static __init int sched_init_debug(void)
252 debugfs_create_file("sched_features", 0644, NULL, NULL,
257 late_initcall(sched_init_debug);
258 #endif /* CONFIG_SCHED_DEBUG */
261 * Number of tasks to iterate in a single balance run.
262 * Limited because this is done with IRQs disabled.
264 const_debug unsigned int sysctl_sched_nr_migrate = 32;
267 * period over which we average the RT time consumption, measured
272 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
275 * period over which we measure -rt task cpu usage in us.
278 unsigned int sysctl_sched_rt_period = 1000000;
280 __read_mostly int scheduler_running;
283 * part of the period that we allow rt tasks to run in us.
286 int sysctl_sched_rt_runtime = 950000;
288 /* cpus with isolated domains */
289 cpumask_var_t cpu_isolated_map;
292 lock_rq_of(struct task_struct *p, unsigned long *flags)
294 return task_rq_lock(p, flags);
298 unlock_rq_of(struct rq *rq, struct task_struct *p, unsigned long *flags)
300 task_rq_unlock(rq, p, flags);
304 * this_rq_lock - lock this runqueue and disable interrupts.
306 static struct rq *this_rq_lock(void)
313 raw_spin_lock(&rq->lock);
318 #ifdef CONFIG_SCHED_HRTICK
320 * Use HR-timers to deliver accurate preemption points.
323 static void hrtick_clear(struct rq *rq)
325 if (hrtimer_active(&rq->hrtick_timer))
326 hrtimer_cancel(&rq->hrtick_timer);
330 * High-resolution timer tick.
331 * Runs from hardirq context with interrupts disabled.
333 static enum hrtimer_restart hrtick(struct hrtimer *timer)
335 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
337 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
339 raw_spin_lock(&rq->lock);
341 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
342 raw_spin_unlock(&rq->lock);
344 return HRTIMER_NORESTART;
349 static void __hrtick_restart(struct rq *rq)
351 struct hrtimer *timer = &rq->hrtick_timer;
353 hrtimer_start_expires(timer, HRTIMER_MODE_ABS_PINNED);
357 * called from hardirq (IPI) context
359 static void __hrtick_start(void *arg)
363 raw_spin_lock(&rq->lock);
364 __hrtick_restart(rq);
365 rq->hrtick_csd_pending = 0;
366 raw_spin_unlock(&rq->lock);
370 * Called to set the hrtick timer state.
372 * called with rq->lock held and irqs disabled
374 void hrtick_start(struct rq *rq, u64 delay)
376 struct hrtimer *timer = &rq->hrtick_timer;
381 * Don't schedule slices shorter than 10000ns, that just
382 * doesn't make sense and can cause timer DoS.
384 delta = max_t(s64, delay, 10000LL);
385 time = ktime_add_ns(timer->base->get_time(), delta);
387 hrtimer_set_expires(timer, time);
389 if (rq == this_rq()) {
390 __hrtick_restart(rq);
391 } else if (!rq->hrtick_csd_pending) {
392 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd);
393 rq->hrtick_csd_pending = 1;
398 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
400 int cpu = (int)(long)hcpu;
403 case CPU_UP_CANCELED:
404 case CPU_UP_CANCELED_FROZEN:
405 case CPU_DOWN_PREPARE:
406 case CPU_DOWN_PREPARE_FROZEN:
408 case CPU_DEAD_FROZEN:
409 hrtick_clear(cpu_rq(cpu));
416 static __init void init_hrtick(void)
418 hotcpu_notifier(hotplug_hrtick, 0);
422 * Called to set the hrtick timer state.
424 * called with rq->lock held and irqs disabled
426 void hrtick_start(struct rq *rq, u64 delay)
429 * Don't schedule slices shorter than 10000ns, that just
430 * doesn't make sense. Rely on vruntime for fairness.
432 delay = max_t(u64, delay, 10000LL);
433 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay),
434 HRTIMER_MODE_REL_PINNED);
437 static inline void init_hrtick(void)
440 #endif /* CONFIG_SMP */
442 static void init_rq_hrtick(struct rq *rq)
445 rq->hrtick_csd_pending = 0;
447 rq->hrtick_csd.flags = 0;
448 rq->hrtick_csd.func = __hrtick_start;
449 rq->hrtick_csd.info = rq;
452 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
453 rq->hrtick_timer.function = hrtick;
455 #else /* CONFIG_SCHED_HRTICK */
456 static inline void hrtick_clear(struct rq *rq)
460 static inline void init_rq_hrtick(struct rq *rq)
464 static inline void init_hrtick(void)
467 #endif /* CONFIG_SCHED_HRTICK */
470 * cmpxchg based fetch_or, macro so it works for different integer types
472 #define fetch_or(ptr, val) \
473 ({ typeof(*(ptr)) __old, __val = *(ptr); \
475 __old = cmpxchg((ptr), __val, __val | (val)); \
476 if (__old == __val) \
483 #if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG)
485 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG,
486 * this avoids any races wrt polling state changes and thereby avoids
489 static bool set_nr_and_not_polling(struct task_struct *p)
491 struct thread_info *ti = task_thread_info(p);
492 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG);
496 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set.
498 * If this returns true, then the idle task promises to call
499 * sched_ttwu_pending() and reschedule soon.
501 static bool set_nr_if_polling(struct task_struct *p)
503 struct thread_info *ti = task_thread_info(p);
504 typeof(ti->flags) old, val = READ_ONCE(ti->flags);
507 if (!(val & _TIF_POLLING_NRFLAG))
509 if (val & _TIF_NEED_RESCHED)
511 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED);
520 static bool set_nr_and_not_polling(struct task_struct *p)
522 set_tsk_need_resched(p);
527 static bool set_nr_if_polling(struct task_struct *p)
534 void wake_q_add(struct wake_q_head *head, struct task_struct *task)
536 struct wake_q_node *node = &task->wake_q;
539 * Atomically grab the task, if ->wake_q is !nil already it means
540 * its already queued (either by us or someone else) and will get the
541 * wakeup due to that.
543 * This cmpxchg() implies a full barrier, which pairs with the write
544 * barrier implied by the wakeup in wake_up_list().
546 if (cmpxchg(&node->next, NULL, WAKE_Q_TAIL))
551 get_task_struct(task);
554 * The head is context local, there can be no concurrency.
557 head->lastp = &node->next;
561 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
562 int sibling_count_hint);
564 void wake_up_q(struct wake_q_head *head)
566 struct wake_q_node *node = head->first;
568 while (node != WAKE_Q_TAIL) {
569 struct task_struct *task;
571 task = container_of(node, struct task_struct, wake_q);
573 /* task can safely be re-inserted now */
575 task->wake_q.next = NULL;
578 * try_to_wake_up() implies a wmb() to pair with the queueing
579 * in wake_q_add() so as not to miss wakeups.
581 try_to_wake_up(task, TASK_NORMAL, 0, head->count);
582 put_task_struct(task);
587 * resched_curr - mark rq's current task 'to be rescheduled now'.
589 * On UP this means the setting of the need_resched flag, on SMP it
590 * might also involve a cross-CPU call to trigger the scheduler on
593 void resched_curr(struct rq *rq)
595 struct task_struct *curr = rq->curr;
598 lockdep_assert_held(&rq->lock);
600 if (test_tsk_need_resched(curr))
605 if (cpu == smp_processor_id()) {
606 set_tsk_need_resched(curr);
607 set_preempt_need_resched();
611 if (set_nr_and_not_polling(curr))
612 smp_send_reschedule(cpu);
614 trace_sched_wake_idle_without_ipi(cpu);
617 void resched_cpu(int cpu)
619 struct rq *rq = cpu_rq(cpu);
622 raw_spin_lock_irqsave(&rq->lock, flags);
624 raw_spin_unlock_irqrestore(&rq->lock, flags);
628 #ifdef CONFIG_NO_HZ_COMMON
630 * In the semi idle case, use the nearest busy cpu for migrating timers
631 * from an idle cpu. This is good for power-savings.
633 * We don't do similar optimization for completely idle system, as
634 * selecting an idle cpu will add more delays to the timers than intended
635 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
637 int get_nohz_timer_target(void)
639 int i, cpu = smp_processor_id();
640 struct sched_domain *sd;
642 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
646 for_each_domain(cpu, sd) {
647 for_each_cpu(i, sched_domain_span(sd)) {
651 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
658 if (!is_housekeeping_cpu(cpu))
659 cpu = housekeeping_any_cpu();
665 * When add_timer_on() enqueues a timer into the timer wheel of an
666 * idle CPU then this timer might expire before the next timer event
667 * which is scheduled to wake up that CPU. In case of a completely
668 * idle system the next event might even be infinite time into the
669 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
670 * leaves the inner idle loop so the newly added timer is taken into
671 * account when the CPU goes back to idle and evaluates the timer
672 * wheel for the next timer event.
674 static void wake_up_idle_cpu(int cpu)
676 struct rq *rq = cpu_rq(cpu);
678 if (cpu == smp_processor_id())
681 if (set_nr_and_not_polling(rq->idle))
682 smp_send_reschedule(cpu);
684 trace_sched_wake_idle_without_ipi(cpu);
687 static bool wake_up_full_nohz_cpu(int cpu)
690 * We just need the target to call irq_exit() and re-evaluate
691 * the next tick. The nohz full kick at least implies that.
692 * If needed we can still optimize that later with an
695 if (tick_nohz_full_cpu(cpu)) {
696 if (cpu != smp_processor_id() ||
697 tick_nohz_tick_stopped())
698 tick_nohz_full_kick_cpu(cpu);
705 void wake_up_nohz_cpu(int cpu)
707 if (!wake_up_full_nohz_cpu(cpu))
708 wake_up_idle_cpu(cpu);
711 static inline bool got_nohz_idle_kick(void)
713 int cpu = smp_processor_id();
715 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
718 if (idle_cpu(cpu) && !need_resched())
722 * We can't run Idle Load Balance on this CPU for this time so we
723 * cancel it and clear NOHZ_BALANCE_KICK
725 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
729 #else /* CONFIG_NO_HZ_COMMON */
731 static inline bool got_nohz_idle_kick(void)
736 #endif /* CONFIG_NO_HZ_COMMON */
738 #ifdef CONFIG_NO_HZ_FULL
739 bool sched_can_stop_tick(void)
742 * FIFO realtime policy runs the highest priority task. Other runnable
743 * tasks are of a lower priority. The scheduler tick does nothing.
745 if (current->policy == SCHED_FIFO)
749 * Round-robin realtime tasks time slice with other tasks at the same
750 * realtime priority. Is this task the only one at this priority?
752 if (current->policy == SCHED_RR) {
753 struct sched_rt_entity *rt_se = ¤t->rt;
755 return rt_se->run_list.prev == rt_se->run_list.next;
759 * More than one running task need preemption.
760 * nr_running update is assumed to be visible
761 * after IPI is sent from wakers.
763 if (this_rq()->nr_running > 1)
768 #endif /* CONFIG_NO_HZ_FULL */
770 void sched_avg_update(struct rq *rq)
772 s64 period = sched_avg_period();
774 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
776 * Inline assembly required to prevent the compiler
777 * optimising this loop into a divmod call.
778 * See __iter_div_u64_rem() for another example of this.
780 asm("" : "+rm" (rq->age_stamp));
781 rq->age_stamp += period;
786 #endif /* CONFIG_SMP */
788 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
789 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
791 * Iterate task_group tree rooted at *from, calling @down when first entering a
792 * node and @up when leaving it for the final time.
794 * Caller must hold rcu_lock or sufficient equivalent.
796 int walk_tg_tree_from(struct task_group *from,
797 tg_visitor down, tg_visitor up, void *data)
799 struct task_group *parent, *child;
805 ret = (*down)(parent, data);
808 list_for_each_entry_rcu(child, &parent->children, siblings) {
815 ret = (*up)(parent, data);
816 if (ret || parent == from)
820 parent = parent->parent;
827 int tg_nop(struct task_group *tg, void *data)
833 static void set_load_weight(struct task_struct *p)
835 int prio = p->static_prio - MAX_RT_PRIO;
836 struct load_weight *load = &p->se.load;
839 * SCHED_IDLE tasks get minimal weight:
841 if (idle_policy(p->policy)) {
842 load->weight = scale_load(WEIGHT_IDLEPRIO);
843 load->inv_weight = WMULT_IDLEPRIO;
847 load->weight = scale_load(prio_to_weight[prio]);
848 load->inv_weight = prio_to_wmult[prio];
851 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
854 if (!(flags & ENQUEUE_RESTORE))
855 sched_info_queued(rq, p);
856 p->sched_class->enqueue_task(rq, p, flags);
859 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
862 if (!(flags & DEQUEUE_SAVE))
863 sched_info_dequeued(rq, p);
864 p->sched_class->dequeue_task(rq, p, flags);
867 void activate_task(struct rq *rq, struct task_struct *p, int flags)
869 if (task_contributes_to_load(p))
870 rq->nr_uninterruptible--;
872 enqueue_task(rq, p, flags);
875 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
877 if (task_contributes_to_load(p))
878 rq->nr_uninterruptible++;
880 dequeue_task(rq, p, flags);
883 static void update_rq_clock_task(struct rq *rq, s64 delta)
886 * In theory, the compile should just see 0 here, and optimize out the call
887 * to sched_rt_avg_update. But I don't trust it...
889 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
890 s64 steal = 0, irq_delta = 0;
892 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
893 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
896 * Since irq_time is only updated on {soft,}irq_exit, we might run into
897 * this case when a previous update_rq_clock() happened inside a
900 * When this happens, we stop ->clock_task and only update the
901 * prev_irq_time stamp to account for the part that fit, so that a next
902 * update will consume the rest. This ensures ->clock_task is
905 * It does however cause some slight miss-attribution of {soft,}irq
906 * time, a more accurate solution would be to update the irq_time using
907 * the current rq->clock timestamp, except that would require using
910 if (irq_delta > delta)
913 rq->prev_irq_time += irq_delta;
916 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
917 if (static_key_false((¶virt_steal_rq_enabled))) {
918 steal = paravirt_steal_clock(cpu_of(rq));
919 steal -= rq->prev_steal_time_rq;
921 if (unlikely(steal > delta))
924 rq->prev_steal_time_rq += steal;
929 rq->clock_task += delta;
931 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
932 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
933 sched_rt_avg_update(rq, irq_delta + steal);
937 void sched_set_stop_task(int cpu, struct task_struct *stop)
939 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
940 struct task_struct *old_stop = cpu_rq(cpu)->stop;
944 * Make it appear like a SCHED_FIFO task, its something
945 * userspace knows about and won't get confused about.
947 * Also, it will make PI more or less work without too
948 * much confusion -- but then, stop work should not
949 * rely on PI working anyway.
951 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
953 stop->sched_class = &stop_sched_class;
956 cpu_rq(cpu)->stop = stop;
960 * Reset it back to a normal scheduling class so that
961 * it can die in pieces.
963 old_stop->sched_class = &rt_sched_class;
968 * __normal_prio - return the priority that is based on the static prio
970 static inline int __normal_prio(struct task_struct *p)
972 return p->static_prio;
976 * Calculate the expected normal priority: i.e. priority
977 * without taking RT-inheritance into account. Might be
978 * boosted by interactivity modifiers. Changes upon fork,
979 * setprio syscalls, and whenever the interactivity
980 * estimator recalculates.
982 static inline int normal_prio(struct task_struct *p)
986 if (task_has_dl_policy(p))
987 prio = MAX_DL_PRIO-1;
988 else if (task_has_rt_policy(p))
989 prio = MAX_RT_PRIO-1 - p->rt_priority;
991 prio = __normal_prio(p);
996 * Calculate the current priority, i.e. the priority
997 * taken into account by the scheduler. This value might
998 * be boosted by RT tasks, or might be boosted by
999 * interactivity modifiers. Will be RT if the task got
1000 * RT-boosted. If not then it returns p->normal_prio.
1002 static int effective_prio(struct task_struct *p)
1004 p->normal_prio = normal_prio(p);
1006 * If we are RT tasks or we were boosted to RT priority,
1007 * keep the priority unchanged. Otherwise, update priority
1008 * to the normal priority:
1010 if (!rt_prio(p->prio))
1011 return p->normal_prio;
1016 * task_curr - is this task currently executing on a CPU?
1017 * @p: the task in question.
1019 * Return: 1 if the task is currently executing. 0 otherwise.
1021 inline int task_curr(const struct task_struct *p)
1023 return cpu_curr(task_cpu(p)) == p;
1027 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1028 * use the balance_callback list if you want balancing.
1030 * this means any call to check_class_changed() must be followed by a call to
1031 * balance_callback().
1033 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1034 const struct sched_class *prev_class,
1037 if (prev_class != p->sched_class) {
1038 if (prev_class->switched_from)
1039 prev_class->switched_from(rq, p);
1041 p->sched_class->switched_to(rq, p);
1042 } else if (oldprio != p->prio || dl_task(p))
1043 p->sched_class->prio_changed(rq, p, oldprio);
1046 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1048 const struct sched_class *class;
1050 if (p->sched_class == rq->curr->sched_class) {
1051 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1053 for_each_class(class) {
1054 if (class == rq->curr->sched_class)
1056 if (class == p->sched_class) {
1064 * A queue event has occurred, and we're going to schedule. In
1065 * this case, we can save a useless back to back clock update.
1067 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1068 rq_clock_skip_update(rq, true);
1073 * This is how migration works:
1075 * 1) we invoke migration_cpu_stop() on the target CPU using
1077 * 2) stopper starts to run (implicitly forcing the migrated thread
1079 * 3) it checks whether the migrated task is still in the wrong runqueue.
1080 * 4) if it's in the wrong runqueue then the migration thread removes
1081 * it and puts it into the right queue.
1082 * 5) stopper completes and stop_one_cpu() returns and the migration
1087 * move_queued_task - move a queued task to new rq.
1089 * Returns (locked) new rq. Old rq's lock is released.
1091 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1093 lockdep_assert_held(&rq->lock);
1095 dequeue_task(rq, p, 0);
1096 p->on_rq = TASK_ON_RQ_MIGRATING;
1097 double_lock_balance(rq, cpu_rq(new_cpu));
1098 set_task_cpu(p, new_cpu);
1099 double_unlock_balance(rq, cpu_rq(new_cpu));
1100 raw_spin_unlock(&rq->lock);
1102 rq = cpu_rq(new_cpu);
1104 raw_spin_lock(&rq->lock);
1105 BUG_ON(task_cpu(p) != new_cpu);
1106 p->on_rq = TASK_ON_RQ_QUEUED;
1107 enqueue_task(rq, p, 0);
1108 check_preempt_curr(rq, p, 0);
1113 struct migration_arg {
1114 struct task_struct *task;
1119 * Move (not current) task off this cpu, onto dest cpu. We're doing
1120 * this because either it can't run here any more (set_cpus_allowed()
1121 * away from this CPU, or CPU going down), or because we're
1122 * attempting to rebalance this task on exec (sched_exec).
1124 * So we race with normal scheduler movements, but that's OK, as long
1125 * as the task is no longer on this CPU.
1127 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1129 if (unlikely(!cpu_active(dest_cpu)))
1132 /* Affinity changed (again). */
1133 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1136 rq = move_queued_task(rq, p, dest_cpu);
1142 * migration_cpu_stop - this will be executed by a highprio stopper thread
1143 * and performs thread migration by bumping thread off CPU then
1144 * 'pushing' onto another runqueue.
1146 static int migration_cpu_stop(void *data)
1148 struct migration_arg *arg = data;
1149 struct task_struct *p = arg->task;
1150 struct rq *rq = this_rq();
1153 * The original target cpu might have gone down and we might
1154 * be on another cpu but it doesn't matter.
1156 local_irq_disable();
1158 * We need to explicitly wake pending tasks before running
1159 * __migrate_task() such that we will not miss enforcing cpus_allowed
1160 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1162 sched_ttwu_pending();
1164 raw_spin_lock(&p->pi_lock);
1165 raw_spin_lock(&rq->lock);
1167 * If task_rq(p) != rq, it cannot be migrated here, because we're
1168 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1169 * we're holding p->pi_lock.
1171 if (task_rq(p) == rq && task_on_rq_queued(p))
1172 rq = __migrate_task(rq, p, arg->dest_cpu);
1173 raw_spin_unlock(&rq->lock);
1174 raw_spin_unlock(&p->pi_lock);
1181 * sched_class::set_cpus_allowed must do the below, but is not required to
1182 * actually call this function.
1184 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1186 cpumask_copy(&p->cpus_allowed, new_mask);
1187 p->nr_cpus_allowed = cpumask_weight(new_mask);
1190 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1192 struct rq *rq = task_rq(p);
1193 bool queued, running;
1195 lockdep_assert_held(&p->pi_lock);
1197 queued = task_on_rq_queued(p);
1198 running = task_current(rq, p);
1202 * Because __kthread_bind() calls this on blocked tasks without
1205 lockdep_assert_held(&rq->lock);
1206 dequeue_task(rq, p, DEQUEUE_SAVE);
1209 put_prev_task(rq, p);
1211 p->sched_class->set_cpus_allowed(p, new_mask);
1214 p->sched_class->set_curr_task(rq);
1216 enqueue_task(rq, p, ENQUEUE_RESTORE);
1220 * Change a given task's CPU affinity. Migrate the thread to a
1221 * proper CPU and schedule it away if the CPU it's executing on
1222 * is removed from the allowed bitmask.
1224 * NOTE: the caller must have a valid reference to the task, the
1225 * task must not exit() & deallocate itself prematurely. The
1226 * call is not atomic; no spinlocks may be held.
1228 static int __set_cpus_allowed_ptr(struct task_struct *p,
1229 const struct cpumask *new_mask, bool check)
1231 unsigned long flags;
1233 unsigned int dest_cpu;
1236 rq = task_rq_lock(p, &flags);
1239 * Must re-check here, to close a race against __kthread_bind(),
1240 * sched_setaffinity() is not guaranteed to observe the flag.
1242 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1247 if (cpumask_equal(&p->cpus_allowed, new_mask))
1250 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1255 do_set_cpus_allowed(p, new_mask);
1257 /* Can the task run on the task's current CPU? If so, we're done */
1258 if (cpumask_test_cpu(task_cpu(p), new_mask))
1261 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1262 if (task_running(rq, p) || p->state == TASK_WAKING) {
1263 struct migration_arg arg = { p, dest_cpu };
1264 /* Need help from migration thread: drop lock and wait. */
1265 task_rq_unlock(rq, p, &flags);
1266 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1267 tlb_migrate_finish(p->mm);
1269 } else if (task_on_rq_queued(p)) {
1271 * OK, since we're going to drop the lock immediately
1272 * afterwards anyway.
1274 lockdep_unpin_lock(&rq->lock);
1275 rq = move_queued_task(rq, p, dest_cpu);
1276 lockdep_pin_lock(&rq->lock);
1279 task_rq_unlock(rq, p, &flags);
1284 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1286 return __set_cpus_allowed_ptr(p, new_mask, false);
1288 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1290 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1292 #ifdef CONFIG_SCHED_DEBUG
1294 * We should never call set_task_cpu() on a blocked task,
1295 * ttwu() will sort out the placement.
1297 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1300 #ifdef CONFIG_LOCKDEP
1302 * The caller should hold either p->pi_lock or rq->lock, when changing
1303 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1305 * sched_move_task() holds both and thus holding either pins the cgroup,
1308 * Furthermore, all task_rq users should acquire both locks, see
1311 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1312 lockdep_is_held(&task_rq(p)->lock)));
1316 trace_sched_migrate_task(p, new_cpu);
1318 if (task_cpu(p) != new_cpu) {
1319 if (p->sched_class->migrate_task_rq)
1320 p->sched_class->migrate_task_rq(p);
1321 p->se.nr_migrations++;
1322 perf_event_task_migrate(p);
1324 walt_fixup_busy_time(p, new_cpu);
1327 __set_task_cpu(p, new_cpu);
1330 static void __migrate_swap_task(struct task_struct *p, int cpu)
1332 if (task_on_rq_queued(p)) {
1333 struct rq *src_rq, *dst_rq;
1335 src_rq = task_rq(p);
1336 dst_rq = cpu_rq(cpu);
1338 deactivate_task(src_rq, p, 0);
1339 p->on_rq = TASK_ON_RQ_MIGRATING;
1340 set_task_cpu(p, cpu);
1341 p->on_rq = TASK_ON_RQ_QUEUED;
1342 activate_task(dst_rq, p, 0);
1343 check_preempt_curr(dst_rq, p, 0);
1346 * Task isn't running anymore; make it appear like we migrated
1347 * it before it went to sleep. This means on wakeup we make the
1348 * previous cpu our targer instead of where it really is.
1354 struct migration_swap_arg {
1355 struct task_struct *src_task, *dst_task;
1356 int src_cpu, dst_cpu;
1359 static int migrate_swap_stop(void *data)
1361 struct migration_swap_arg *arg = data;
1362 struct rq *src_rq, *dst_rq;
1365 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1368 src_rq = cpu_rq(arg->src_cpu);
1369 dst_rq = cpu_rq(arg->dst_cpu);
1371 double_raw_lock(&arg->src_task->pi_lock,
1372 &arg->dst_task->pi_lock);
1373 double_rq_lock(src_rq, dst_rq);
1375 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1378 if (task_cpu(arg->src_task) != arg->src_cpu)
1381 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1384 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1387 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1388 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1393 double_rq_unlock(src_rq, dst_rq);
1394 raw_spin_unlock(&arg->dst_task->pi_lock);
1395 raw_spin_unlock(&arg->src_task->pi_lock);
1401 * Cross migrate two tasks
1403 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1405 struct migration_swap_arg arg;
1408 arg = (struct migration_swap_arg){
1410 .src_cpu = task_cpu(cur),
1412 .dst_cpu = task_cpu(p),
1415 if (arg.src_cpu == arg.dst_cpu)
1419 * These three tests are all lockless; this is OK since all of them
1420 * will be re-checked with proper locks held further down the line.
1422 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1425 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1428 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1431 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1432 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1439 * wait_task_inactive - wait for a thread to unschedule.
1441 * If @match_state is nonzero, it's the @p->state value just checked and
1442 * not expected to change. If it changes, i.e. @p might have woken up,
1443 * then return zero. When we succeed in waiting for @p to be off its CPU,
1444 * we return a positive number (its total switch count). If a second call
1445 * a short while later returns the same number, the caller can be sure that
1446 * @p has remained unscheduled the whole time.
1448 * The caller must ensure that the task *will* unschedule sometime soon,
1449 * else this function might spin for a *long* time. This function can't
1450 * be called with interrupts off, or it may introduce deadlock with
1451 * smp_call_function() if an IPI is sent by the same process we are
1452 * waiting to become inactive.
1454 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1456 unsigned long flags;
1457 int running, queued;
1463 * We do the initial early heuristics without holding
1464 * any task-queue locks at all. We'll only try to get
1465 * the runqueue lock when things look like they will
1471 * If the task is actively running on another CPU
1472 * still, just relax and busy-wait without holding
1475 * NOTE! Since we don't hold any locks, it's not
1476 * even sure that "rq" stays as the right runqueue!
1477 * But we don't care, since "task_running()" will
1478 * return false if the runqueue has changed and p
1479 * is actually now running somewhere else!
1481 while (task_running(rq, p)) {
1482 if (match_state && unlikely(p->state != match_state))
1488 * Ok, time to look more closely! We need the rq
1489 * lock now, to be *sure*. If we're wrong, we'll
1490 * just go back and repeat.
1492 rq = task_rq_lock(p, &flags);
1493 trace_sched_wait_task(p);
1494 running = task_running(rq, p);
1495 queued = task_on_rq_queued(p);
1497 if (!match_state || p->state == match_state)
1498 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1499 task_rq_unlock(rq, p, &flags);
1502 * If it changed from the expected state, bail out now.
1504 if (unlikely(!ncsw))
1508 * Was it really running after all now that we
1509 * checked with the proper locks actually held?
1511 * Oops. Go back and try again..
1513 if (unlikely(running)) {
1519 * It's not enough that it's not actively running,
1520 * it must be off the runqueue _entirely_, and not
1523 * So if it was still runnable (but just not actively
1524 * running right now), it's preempted, and we should
1525 * yield - it could be a while.
1527 if (unlikely(queued)) {
1528 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1530 set_current_state(TASK_UNINTERRUPTIBLE);
1531 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1536 * Ahh, all good. It wasn't running, and it wasn't
1537 * runnable, which means that it will never become
1538 * running in the future either. We're all done!
1547 * kick_process - kick a running thread to enter/exit the kernel
1548 * @p: the to-be-kicked thread
1550 * Cause a process which is running on another CPU to enter
1551 * kernel-mode, without any delay. (to get signals handled.)
1553 * NOTE: this function doesn't have to take the runqueue lock,
1554 * because all it wants to ensure is that the remote task enters
1555 * the kernel. If the IPI races and the task has been migrated
1556 * to another CPU then no harm is done and the purpose has been
1559 void kick_process(struct task_struct *p)
1565 if ((cpu != smp_processor_id()) && task_curr(p))
1566 smp_send_reschedule(cpu);
1569 EXPORT_SYMBOL_GPL(kick_process);
1572 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1574 static int select_fallback_rq(int cpu, struct task_struct *p)
1576 int nid = cpu_to_node(cpu);
1577 const struct cpumask *nodemask = NULL;
1578 enum { cpuset, possible, fail } state = cpuset;
1582 * If the node that the cpu is on has been offlined, cpu_to_node()
1583 * will return -1. There is no cpu on the node, and we should
1584 * select the cpu on the other node.
1587 nodemask = cpumask_of_node(nid);
1589 /* Look for allowed, online CPU in same node. */
1590 for_each_cpu(dest_cpu, nodemask) {
1591 if (!cpu_online(dest_cpu))
1593 if (!cpu_active(dest_cpu))
1595 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1601 /* Any allowed, online CPU? */
1602 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1603 if (!cpu_online(dest_cpu))
1605 if (!cpu_active(dest_cpu))
1610 /* No more Mr. Nice Guy. */
1613 if (IS_ENABLED(CONFIG_CPUSETS)) {
1614 cpuset_cpus_allowed_fallback(p);
1620 do_set_cpus_allowed(p, cpu_possible_mask);
1631 if (state != cpuset) {
1633 * Don't tell them about moving exiting tasks or
1634 * kernel threads (both mm NULL), since they never
1637 if (p->mm && printk_ratelimit()) {
1638 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1639 task_pid_nr(p), p->comm, cpu);
1647 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1650 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags,
1651 int sibling_count_hint)
1653 lockdep_assert_held(&p->pi_lock);
1655 if (p->nr_cpus_allowed > 1)
1656 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags,
1657 sibling_count_hint);
1660 * In order not to call set_task_cpu() on a blocking task we need
1661 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1664 * Since this is common to all placement strategies, this lives here.
1666 * [ this allows ->select_task() to simply return task_cpu(p) and
1667 * not worry about this generic constraint ]
1669 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1671 cpu = select_fallback_rq(task_cpu(p), p);
1676 static void update_avg(u64 *avg, u64 sample)
1678 s64 diff = sample - *avg;
1684 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1685 const struct cpumask *new_mask, bool check)
1687 return set_cpus_allowed_ptr(p, new_mask);
1690 #endif /* CONFIG_SMP */
1693 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1695 #ifdef CONFIG_SCHEDSTATS
1696 struct rq *rq = this_rq();
1699 int this_cpu = smp_processor_id();
1701 if (cpu == this_cpu) {
1702 schedstat_inc(rq, ttwu_local);
1703 schedstat_inc(p, se.statistics.nr_wakeups_local);
1705 struct sched_domain *sd;
1707 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1709 for_each_domain(this_cpu, sd) {
1710 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1711 schedstat_inc(sd, ttwu_wake_remote);
1718 if (wake_flags & WF_MIGRATED)
1719 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1721 #endif /* CONFIG_SMP */
1723 schedstat_inc(rq, ttwu_count);
1724 schedstat_inc(p, se.statistics.nr_wakeups);
1726 if (wake_flags & WF_SYNC)
1727 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1729 #endif /* CONFIG_SCHEDSTATS */
1732 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1734 activate_task(rq, p, en_flags);
1735 p->on_rq = TASK_ON_RQ_QUEUED;
1737 /* if a worker is waking up, notify workqueue */
1738 if (p->flags & PF_WQ_WORKER)
1739 wq_worker_waking_up(p, cpu_of(rq));
1743 * Mark the task runnable and perform wakeup-preemption.
1746 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1748 check_preempt_curr(rq, p, wake_flags);
1749 p->state = TASK_RUNNING;
1750 trace_sched_wakeup(p);
1753 if (p->sched_class->task_woken) {
1755 * Our task @p is fully woken up and running; so its safe to
1756 * drop the rq->lock, hereafter rq is only used for statistics.
1758 lockdep_unpin_lock(&rq->lock);
1759 p->sched_class->task_woken(rq, p);
1760 lockdep_pin_lock(&rq->lock);
1763 if (rq->idle_stamp) {
1764 u64 delta = rq_clock(rq) - rq->idle_stamp;
1765 u64 max = 2*rq->max_idle_balance_cost;
1767 update_avg(&rq->avg_idle, delta);
1769 if (rq->avg_idle > max)
1778 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1780 lockdep_assert_held(&rq->lock);
1783 if (p->sched_contributes_to_load)
1784 rq->nr_uninterruptible--;
1787 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1788 ttwu_do_wakeup(rq, p, wake_flags);
1792 * Called in case the task @p isn't fully descheduled from its runqueue,
1793 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1794 * since all we need to do is flip p->state to TASK_RUNNING, since
1795 * the task is still ->on_rq.
1797 static int ttwu_remote(struct task_struct *p, int wake_flags)
1802 rq = __task_rq_lock(p);
1803 if (task_on_rq_queued(p)) {
1804 /* check_preempt_curr() may use rq clock */
1805 update_rq_clock(rq);
1806 ttwu_do_wakeup(rq, p, wake_flags);
1809 __task_rq_unlock(rq);
1815 void sched_ttwu_pending(void)
1817 struct rq *rq = this_rq();
1818 struct llist_node *llist = llist_del_all(&rq->wake_list);
1819 struct task_struct *p;
1820 unsigned long flags;
1825 raw_spin_lock_irqsave(&rq->lock, flags);
1826 lockdep_pin_lock(&rq->lock);
1829 p = llist_entry(llist, struct task_struct, wake_entry);
1830 llist = llist_next(llist);
1831 ttwu_do_activate(rq, p, 0);
1834 lockdep_unpin_lock(&rq->lock);
1835 raw_spin_unlock_irqrestore(&rq->lock, flags);
1838 void scheduler_ipi(void)
1841 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1842 * TIF_NEED_RESCHED remotely (for the first time) will also send
1845 preempt_fold_need_resched();
1847 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1851 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1852 * traditionally all their work was done from the interrupt return
1853 * path. Now that we actually do some work, we need to make sure
1856 * Some archs already do call them, luckily irq_enter/exit nest
1859 * Arguably we should visit all archs and update all handlers,
1860 * however a fair share of IPIs are still resched only so this would
1861 * somewhat pessimize the simple resched case.
1864 sched_ttwu_pending();
1867 * Check if someone kicked us for doing the nohz idle load balance.
1869 if (unlikely(got_nohz_idle_kick())) {
1870 this_rq()->idle_balance = 1;
1871 raise_softirq_irqoff(SCHED_SOFTIRQ);
1876 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1878 struct rq *rq = cpu_rq(cpu);
1880 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1881 if (!set_nr_if_polling(rq->idle))
1882 smp_send_reschedule(cpu);
1884 trace_sched_wake_idle_without_ipi(cpu);
1888 void wake_up_if_idle(int cpu)
1890 struct rq *rq = cpu_rq(cpu);
1891 unsigned long flags;
1895 if (!is_idle_task(rcu_dereference(rq->curr)))
1898 if (set_nr_if_polling(rq->idle)) {
1899 trace_sched_wake_idle_without_ipi(cpu);
1901 raw_spin_lock_irqsave(&rq->lock, flags);
1902 if (is_idle_task(rq->curr))
1903 smp_send_reschedule(cpu);
1904 /* Else cpu is not in idle, do nothing here */
1905 raw_spin_unlock_irqrestore(&rq->lock, flags);
1912 bool cpus_share_cache(int this_cpu, int that_cpu)
1914 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1916 #endif /* CONFIG_SMP */
1918 static void ttwu_queue(struct task_struct *p, int cpu)
1920 struct rq *rq = cpu_rq(cpu);
1922 #if defined(CONFIG_SMP)
1923 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1924 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1925 ttwu_queue_remote(p, cpu);
1930 raw_spin_lock(&rq->lock);
1931 lockdep_pin_lock(&rq->lock);
1932 ttwu_do_activate(rq, p, 0);
1933 lockdep_unpin_lock(&rq->lock);
1934 raw_spin_unlock(&rq->lock);
1938 * try_to_wake_up - wake up a thread
1939 * @p: the thread to be awakened
1940 * @state: the mask of task states that can be woken
1941 * @wake_flags: wake modifier flags (WF_*)
1942 * @sibling_count_hint: A hint at the number of threads that are being woken up
1945 * Put it on the run-queue if it's not already there. The "current"
1946 * thread is always on the run-queue (except when the actual
1947 * re-schedule is in progress), and as such you're allowed to do
1948 * the simpler "current->state = TASK_RUNNING" to mark yourself
1949 * runnable without the overhead of this.
1951 * Return: %true if @p was woken up, %false if it was already running.
1952 * or @state didn't match @p's state.
1955 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags,
1956 int sibling_count_hint)
1958 unsigned long flags;
1959 int cpu, success = 0;
1966 * If we are going to wake up a thread waiting for CONDITION we
1967 * need to ensure that CONDITION=1 done by the caller can not be
1968 * reordered with p->state check below. This pairs with mb() in
1969 * set_current_state() the waiting thread does.
1971 smp_mb__before_spinlock();
1972 raw_spin_lock_irqsave(&p->pi_lock, flags);
1973 if (!(p->state & state))
1976 trace_sched_waking(p);
1978 success = 1; /* we're going to change ->state */
1982 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1983 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1984 * in smp_cond_load_acquire() below.
1986 * sched_ttwu_pending() try_to_wake_up()
1987 * [S] p->on_rq = 1; [L] P->state
1988 * UNLOCK rq->lock -----.
1992 * LOCK rq->lock -----'
1996 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1998 * Pairs with the UNLOCK+LOCK on rq->lock from the
1999 * last wakeup of our task and the schedule that got our task
2003 if (p->on_rq && ttwu_remote(p, wake_flags))
2008 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
2009 * possible to, falsely, observe p->on_cpu == 0.
2011 * One must be running (->on_cpu == 1) in order to remove oneself
2012 * from the runqueue.
2014 * [S] ->on_cpu = 1; [L] ->on_rq
2018 * [S] ->on_rq = 0; [L] ->on_cpu
2020 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2021 * from the consecutive calls to schedule(); the first switching to our
2022 * task, the second putting it to sleep.
2027 * If the owning (remote) cpu is still in the middle of schedule() with
2028 * this task as prev, wait until its done referencing the task.
2033 * Combined with the control dependency above, we have an effective
2034 * smp_load_acquire() without the need for full barriers.
2036 * Pairs with the smp_store_release() in finish_lock_switch().
2038 * This ensures that tasks getting woken will be fully ordered against
2039 * their previous state and preserve Program Order.
2043 rq = cpu_rq(task_cpu(p));
2045 raw_spin_lock(&rq->lock);
2046 wallclock = walt_ktime_clock();
2047 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2048 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2049 raw_spin_unlock(&rq->lock);
2051 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2052 p->state = TASK_WAKING;
2054 if (p->sched_class->task_waking)
2055 p->sched_class->task_waking(p);
2057 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags,
2058 sibling_count_hint);
2059 if (task_cpu(p) != cpu) {
2060 wake_flags |= WF_MIGRATED;
2061 set_task_cpu(p, cpu);
2064 #endif /* CONFIG_SMP */
2068 ttwu_stat(p, cpu, wake_flags);
2070 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2076 * try_to_wake_up_local - try to wake up a local task with rq lock held
2077 * @p: the thread to be awakened
2079 * Put @p on the run-queue if it's not already there. The caller must
2080 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2083 static void try_to_wake_up_local(struct task_struct *p)
2085 struct rq *rq = task_rq(p);
2087 if (WARN_ON_ONCE(rq != this_rq()) ||
2088 WARN_ON_ONCE(p == current))
2091 lockdep_assert_held(&rq->lock);
2093 if (!raw_spin_trylock(&p->pi_lock)) {
2095 * This is OK, because current is on_cpu, which avoids it being
2096 * picked for load-balance and preemption/IRQs are still
2097 * disabled avoiding further scheduler activity on it and we've
2098 * not yet picked a replacement task.
2100 lockdep_unpin_lock(&rq->lock);
2101 raw_spin_unlock(&rq->lock);
2102 raw_spin_lock(&p->pi_lock);
2103 raw_spin_lock(&rq->lock);
2104 lockdep_pin_lock(&rq->lock);
2107 if (!(p->state & TASK_NORMAL))
2110 trace_sched_waking(p);
2112 if (!task_on_rq_queued(p)) {
2113 u64 wallclock = walt_ktime_clock();
2115 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2116 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2117 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2120 ttwu_do_wakeup(rq, p, 0);
2121 ttwu_stat(p, smp_processor_id(), 0);
2123 raw_spin_unlock(&p->pi_lock);
2127 * wake_up_process - Wake up a specific process
2128 * @p: The process to be woken up.
2130 * Attempt to wake up the nominated process and move it to the set of runnable
2133 * Return: 1 if the process was woken up, 0 if it was already running.
2135 * It may be assumed that this function implies a write memory barrier before
2136 * changing the task state if and only if any tasks are woken up.
2138 int wake_up_process(struct task_struct *p)
2140 return try_to_wake_up(p, TASK_NORMAL, 0, 1);
2142 EXPORT_SYMBOL(wake_up_process);
2144 int wake_up_state(struct task_struct *p, unsigned int state)
2146 return try_to_wake_up(p, state, 0, 1);
2150 * This function clears the sched_dl_entity static params.
2152 void __dl_clear_params(struct task_struct *p)
2154 struct sched_dl_entity *dl_se = &p->dl;
2156 dl_se->dl_runtime = 0;
2157 dl_se->dl_deadline = 0;
2158 dl_se->dl_period = 0;
2162 dl_se->dl_throttled = 0;
2164 dl_se->dl_yielded = 0;
2168 * Perform scheduler related setup for a newly forked process p.
2169 * p is forked by current.
2171 * __sched_fork() is basic setup used by init_idle() too:
2173 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2178 p->se.exec_start = 0;
2179 p->se.sum_exec_runtime = 0;
2180 p->se.prev_sum_exec_runtime = 0;
2181 p->se.nr_migrations = 0;
2183 #ifdef CONFIG_SCHED_WALT
2184 p->last_sleep_ts = 0;
2187 INIT_LIST_HEAD(&p->se.group_node);
2188 walt_init_new_task_load(p);
2190 #ifdef CONFIG_FAIR_GROUP_SCHED
2191 p->se.cfs_rq = NULL;
2194 #ifdef CONFIG_SCHEDSTATS
2195 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2198 RB_CLEAR_NODE(&p->dl.rb_node);
2199 init_dl_task_timer(&p->dl);
2200 __dl_clear_params(p);
2202 INIT_LIST_HEAD(&p->rt.run_list);
2204 #ifdef CONFIG_PREEMPT_NOTIFIERS
2205 INIT_HLIST_HEAD(&p->preempt_notifiers);
2208 #ifdef CONFIG_NUMA_BALANCING
2209 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2210 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2211 p->mm->numa_scan_seq = 0;
2214 if (clone_flags & CLONE_VM)
2215 p->numa_preferred_nid = current->numa_preferred_nid;
2217 p->numa_preferred_nid = -1;
2219 p->node_stamp = 0ULL;
2220 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2221 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2222 p->numa_work.next = &p->numa_work;
2223 p->numa_faults = NULL;
2224 p->last_task_numa_placement = 0;
2225 p->last_sum_exec_runtime = 0;
2227 p->numa_group = NULL;
2228 #endif /* CONFIG_NUMA_BALANCING */
2231 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2233 #ifdef CONFIG_NUMA_BALANCING
2235 void set_numabalancing_state(bool enabled)
2238 static_branch_enable(&sched_numa_balancing);
2240 static_branch_disable(&sched_numa_balancing);
2243 #ifdef CONFIG_PROC_SYSCTL
2244 int sysctl_numa_balancing(struct ctl_table *table, int write,
2245 void __user *buffer, size_t *lenp, loff_t *ppos)
2249 int state = static_branch_likely(&sched_numa_balancing);
2251 if (write && !capable(CAP_SYS_ADMIN))
2256 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2260 set_numabalancing_state(state);
2267 * fork()/clone()-time setup:
2269 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2271 unsigned long flags;
2272 int cpu = get_cpu();
2274 __sched_fork(clone_flags, p);
2276 * We mark the process as NEW here. This guarantees that
2277 * nobody will actually run it, and a signal or other external
2278 * event cannot wake it up and insert it on the runqueue either.
2280 p->state = TASK_NEW;
2283 * Make sure we do not leak PI boosting priority to the child.
2285 p->prio = current->normal_prio;
2288 * Revert to default priority/policy on fork if requested.
2290 if (unlikely(p->sched_reset_on_fork)) {
2291 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2292 p->policy = SCHED_NORMAL;
2293 p->static_prio = NICE_TO_PRIO(0);
2295 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2296 p->static_prio = NICE_TO_PRIO(0);
2298 p->prio = p->normal_prio = __normal_prio(p);
2302 * We don't need the reset flag anymore after the fork. It has
2303 * fulfilled its duty:
2305 p->sched_reset_on_fork = 0;
2308 if (dl_prio(p->prio)) {
2311 } else if (rt_prio(p->prio)) {
2312 p->sched_class = &rt_sched_class;
2314 p->sched_class = &fair_sched_class;
2317 init_entity_runnable_average(&p->se);
2320 * The child is not yet in the pid-hash so no cgroup attach races,
2321 * and the cgroup is pinned to this child due to cgroup_fork()
2322 * is ran before sched_fork().
2324 * Silence PROVE_RCU.
2326 raw_spin_lock_irqsave(&p->pi_lock, flags);
2328 * We're setting the cpu for the first time, we don't migrate,
2329 * so use __set_task_cpu().
2331 __set_task_cpu(p, cpu);
2332 if (p->sched_class->task_fork)
2333 p->sched_class->task_fork(p);
2334 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2336 #ifdef CONFIG_SCHED_INFO
2337 if (likely(sched_info_on()))
2338 memset(&p->sched_info, 0, sizeof(p->sched_info));
2340 #if defined(CONFIG_SMP)
2343 init_task_preempt_count(p);
2345 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2346 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2353 unsigned long to_ratio(u64 period, u64 runtime)
2355 if (runtime == RUNTIME_INF)
2359 * Doing this here saves a lot of checks in all
2360 * the calling paths, and returning zero seems
2361 * safe for them anyway.
2366 return div64_u64(runtime << 20, period);
2370 inline struct dl_bw *dl_bw_of(int i)
2372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2373 "sched RCU must be held");
2374 return &cpu_rq(i)->rd->dl_bw;
2377 static inline int dl_bw_cpus(int i)
2379 struct root_domain *rd = cpu_rq(i)->rd;
2382 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2383 "sched RCU must be held");
2384 for_each_cpu_and(i, rd->span, cpu_active_mask)
2390 inline struct dl_bw *dl_bw_of(int i)
2392 return &cpu_rq(i)->dl.dl_bw;
2395 static inline int dl_bw_cpus(int i)
2402 * We must be sure that accepting a new task (or allowing changing the
2403 * parameters of an existing one) is consistent with the bandwidth
2404 * constraints. If yes, this function also accordingly updates the currently
2405 * allocated bandwidth to reflect the new situation.
2407 * This function is called while holding p's rq->lock.
2409 * XXX we should delay bw change until the task's 0-lag point, see
2412 static int dl_overflow(struct task_struct *p, int policy,
2413 const struct sched_attr *attr)
2416 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2417 u64 period = attr->sched_period ?: attr->sched_deadline;
2418 u64 runtime = attr->sched_runtime;
2419 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2422 if (new_bw == p->dl.dl_bw)
2426 * Either if a task, enters, leave, or stays -deadline but changes
2427 * its parameters, we may need to update accordingly the total
2428 * allocated bandwidth of the container.
2430 raw_spin_lock(&dl_b->lock);
2431 cpus = dl_bw_cpus(task_cpu(p));
2432 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2433 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2434 __dl_add(dl_b, new_bw);
2436 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2437 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2438 __dl_clear(dl_b, p->dl.dl_bw);
2439 __dl_add(dl_b, new_bw);
2441 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2442 __dl_clear(dl_b, p->dl.dl_bw);
2445 raw_spin_unlock(&dl_b->lock);
2450 extern void init_dl_bw(struct dl_bw *dl_b);
2453 * wake_up_new_task - wake up a newly created task for the first time.
2455 * This function will do some initial scheduler statistics housekeeping
2456 * that must be done for every newly created context, then puts the task
2457 * on the runqueue and wakes it.
2459 void wake_up_new_task(struct task_struct *p)
2461 unsigned long flags;
2464 raw_spin_lock_irqsave(&p->pi_lock, flags);
2465 p->state = TASK_RUNNING;
2467 walt_init_new_task_load(p);
2469 /* Initialize new task's runnable average */
2470 init_entity_runnable_average(&p->se);
2473 * Fork balancing, do it here and not earlier because:
2474 * - cpus_allowed can change in the fork path
2475 * - any previously selected cpu might disappear through hotplug
2477 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2478 * as we're not fully set-up yet.
2480 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0, 1));
2482 rq = __task_rq_lock(p);
2483 update_rq_clock(rq);
2484 post_init_entity_util_avg(&p->se);
2486 walt_mark_task_starting(p);
2487 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2488 p->on_rq = TASK_ON_RQ_QUEUED;
2489 trace_sched_wakeup_new(p);
2490 check_preempt_curr(rq, p, WF_FORK);
2492 if (p->sched_class->task_woken) {
2494 * Nothing relies on rq->lock after this, so its fine to
2497 lockdep_unpin_lock(&rq->lock);
2498 p->sched_class->task_woken(rq, p);
2499 lockdep_pin_lock(&rq->lock);
2502 task_rq_unlock(rq, p, &flags);
2505 #ifdef CONFIG_PREEMPT_NOTIFIERS
2507 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2509 void preempt_notifier_inc(void)
2511 static_key_slow_inc(&preempt_notifier_key);
2513 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2515 void preempt_notifier_dec(void)
2517 static_key_slow_dec(&preempt_notifier_key);
2519 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2522 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2523 * @notifier: notifier struct to register
2525 void preempt_notifier_register(struct preempt_notifier *notifier)
2527 if (!static_key_false(&preempt_notifier_key))
2528 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2530 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2532 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2535 * preempt_notifier_unregister - no longer interested in preemption notifications
2536 * @notifier: notifier struct to unregister
2538 * This is *not* safe to call from within a preemption notifier.
2540 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2542 hlist_del(¬ifier->link);
2544 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2546 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2548 struct preempt_notifier *notifier;
2550 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2551 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2554 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2556 if (static_key_false(&preempt_notifier_key))
2557 __fire_sched_in_preempt_notifiers(curr);
2561 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2562 struct task_struct *next)
2564 struct preempt_notifier *notifier;
2566 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2567 notifier->ops->sched_out(notifier, next);
2570 static __always_inline void
2571 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2572 struct task_struct *next)
2574 if (static_key_false(&preempt_notifier_key))
2575 __fire_sched_out_preempt_notifiers(curr, next);
2578 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2580 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2585 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2586 struct task_struct *next)
2590 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2593 * prepare_task_switch - prepare to switch tasks
2594 * @rq: the runqueue preparing to switch
2595 * @prev: the current task that is being switched out
2596 * @next: the task we are going to switch to.
2598 * This is called with the rq lock held and interrupts off. It must
2599 * be paired with a subsequent finish_task_switch after the context
2602 * prepare_task_switch sets up locking and calls architecture specific
2606 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2607 struct task_struct *next)
2609 sched_info_switch(rq, prev, next);
2610 perf_event_task_sched_out(prev, next);
2611 fire_sched_out_preempt_notifiers(prev, next);
2612 prepare_lock_switch(rq, next);
2613 prepare_arch_switch(next);
2617 * finish_task_switch - clean up after a task-switch
2618 * @prev: the thread we just switched away from.
2620 * finish_task_switch must be called after the context switch, paired
2621 * with a prepare_task_switch call before the context switch.
2622 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2623 * and do any other architecture-specific cleanup actions.
2625 * Note that we may have delayed dropping an mm in context_switch(). If
2626 * so, we finish that here outside of the runqueue lock. (Doing it
2627 * with the lock held can cause deadlocks; see schedule() for
2630 * The context switch have flipped the stack from under us and restored the
2631 * local variables which were saved when this task called schedule() in the
2632 * past. prev == current is still correct but we need to recalculate this_rq
2633 * because prev may have moved to another CPU.
2635 static struct rq *finish_task_switch(struct task_struct *prev)
2636 __releases(rq->lock)
2638 struct rq *rq = this_rq();
2639 struct mm_struct *mm = rq->prev_mm;
2643 * The previous task will have left us with a preempt_count of 2
2644 * because it left us after:
2647 * preempt_disable(); // 1
2649 * raw_spin_lock_irq(&rq->lock) // 2
2651 * Also, see FORK_PREEMPT_COUNT.
2653 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2654 "corrupted preempt_count: %s/%d/0x%x\n",
2655 current->comm, current->pid, preempt_count()))
2656 preempt_count_set(FORK_PREEMPT_COUNT);
2661 * A task struct has one reference for the use as "current".
2662 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2663 * schedule one last time. The schedule call will never return, and
2664 * the scheduled task must drop that reference.
2666 * We must observe prev->state before clearing prev->on_cpu (in
2667 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2668 * running on another CPU and we could rave with its RUNNING -> DEAD
2669 * transition, resulting in a double drop.
2671 prev_state = prev->state;
2672 vtime_task_switch(prev);
2673 perf_event_task_sched_in(prev, current);
2674 finish_lock_switch(rq, prev);
2675 finish_arch_post_lock_switch();
2677 fire_sched_in_preempt_notifiers(current);
2680 if (unlikely(prev_state == TASK_DEAD)) {
2681 if (prev->sched_class->task_dead)
2682 prev->sched_class->task_dead(prev);
2685 * Remove function-return probe instances associated with this
2686 * task and put them back on the free list.
2688 kprobe_flush_task(prev);
2689 put_task_struct(prev);
2692 tick_nohz_task_switch();
2698 /* rq->lock is NOT held, but preemption is disabled */
2699 static void __balance_callback(struct rq *rq)
2701 struct callback_head *head, *next;
2702 void (*func)(struct rq *rq);
2703 unsigned long flags;
2705 raw_spin_lock_irqsave(&rq->lock, flags);
2706 head = rq->balance_callback;
2707 rq->balance_callback = NULL;
2709 func = (void (*)(struct rq *))head->func;
2716 raw_spin_unlock_irqrestore(&rq->lock, flags);
2719 static inline void balance_callback(struct rq *rq)
2721 if (unlikely(rq->balance_callback))
2722 __balance_callback(rq);
2727 static inline void balance_callback(struct rq *rq)
2734 * schedule_tail - first thing a freshly forked thread must call.
2735 * @prev: the thread we just switched away from.
2737 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2738 __releases(rq->lock)
2743 * New tasks start with FORK_PREEMPT_COUNT, see there and
2744 * finish_task_switch() for details.
2746 * finish_task_switch() will drop rq->lock() and lower preempt_count
2747 * and the preempt_enable() will end up enabling preemption (on
2748 * PREEMPT_COUNT kernels).
2751 rq = finish_task_switch(prev);
2752 balance_callback(rq);
2755 if (current->set_child_tid)
2756 put_user(task_pid_vnr(current), current->set_child_tid);
2760 * context_switch - switch to the new MM and the new thread's register state.
2762 static inline struct rq *
2763 context_switch(struct rq *rq, struct task_struct *prev,
2764 struct task_struct *next)
2766 struct mm_struct *mm, *oldmm;
2768 prepare_task_switch(rq, prev, next);
2771 oldmm = prev->active_mm;
2773 * For paravirt, this is coupled with an exit in switch_to to
2774 * combine the page table reload and the switch backend into
2777 arch_start_context_switch(prev);
2780 next->active_mm = oldmm;
2781 atomic_inc(&oldmm->mm_count);
2782 enter_lazy_tlb(oldmm, next);
2784 switch_mm_irqs_off(oldmm, mm, next);
2787 prev->active_mm = NULL;
2788 rq->prev_mm = oldmm;
2791 * Since the runqueue lock will be released by the next
2792 * task (which is an invalid locking op but in the case
2793 * of the scheduler it's an obvious special-case), so we
2794 * do an early lockdep release here:
2796 lockdep_unpin_lock(&rq->lock);
2797 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2799 /* Here we just switch the register state and the stack. */
2800 switch_to(prev, next, prev);
2803 return finish_task_switch(prev);
2807 * nr_running and nr_context_switches:
2809 * externally visible scheduler statistics: current number of runnable
2810 * threads, total number of context switches performed since bootup.
2812 unsigned long nr_running(void)
2814 unsigned long i, sum = 0;
2816 for_each_online_cpu(i)
2817 sum += cpu_rq(i)->nr_running;
2823 * Check if only the current task is running on the cpu.
2825 * Caution: this function does not check that the caller has disabled
2826 * preemption, thus the result might have a time-of-check-to-time-of-use
2827 * race. The caller is responsible to use it correctly, for example:
2829 * - from a non-preemptable section (of course)
2831 * - from a thread that is bound to a single CPU
2833 * - in a loop with very short iterations (e.g. a polling loop)
2835 bool single_task_running(void)
2837 return raw_rq()->nr_running == 1;
2839 EXPORT_SYMBOL(single_task_running);
2841 unsigned long long nr_context_switches(void)
2844 unsigned long long sum = 0;
2846 for_each_possible_cpu(i)
2847 sum += cpu_rq(i)->nr_switches;
2852 unsigned long nr_iowait(void)
2854 unsigned long i, sum = 0;
2856 for_each_possible_cpu(i)
2857 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2862 unsigned long nr_iowait_cpu(int cpu)
2864 struct rq *this = cpu_rq(cpu);
2865 return atomic_read(&this->nr_iowait);
2868 #ifdef CONFIG_CPU_QUIET
2869 u64 nr_running_integral(unsigned int cpu)
2871 unsigned int seqcnt;
2875 if (cpu >= nr_cpu_ids)
2881 * Update average to avoid reading stalled value if there were
2882 * no run-queue changes for a long time. On the other hand if
2883 * the changes are happening right now, just read current value
2887 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2888 integral = do_nr_running_integral(q);
2889 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2890 read_seqcount_begin(&q->ave_seqcnt);
2891 integral = q->nr_running_integral;
2898 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2900 struct rq *rq = this_rq();
2901 *nr_waiters = atomic_read(&rq->nr_iowait);
2902 *load = rq->load.weight;
2908 * sched_exec - execve() is a valuable balancing opportunity, because at
2909 * this point the task has the smallest effective memory and cache footprint.
2911 void sched_exec(void)
2913 struct task_struct *p = current;
2914 unsigned long flags;
2917 raw_spin_lock_irqsave(&p->pi_lock, flags);
2918 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0, 1);
2919 if (dest_cpu == smp_processor_id())
2922 if (likely(cpu_active(dest_cpu))) {
2923 struct migration_arg arg = { p, dest_cpu };
2925 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2926 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2930 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2935 DEFINE_PER_CPU(struct kernel_stat, kstat);
2936 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2938 EXPORT_PER_CPU_SYMBOL(kstat);
2939 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2942 * Return accounted runtime for the task.
2943 * In case the task is currently running, return the runtime plus current's
2944 * pending runtime that have not been accounted yet.
2946 unsigned long long task_sched_runtime(struct task_struct *p)
2948 unsigned long flags;
2952 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2954 * 64-bit doesn't need locks to atomically read a 64bit value.
2955 * So we have a optimization chance when the task's delta_exec is 0.
2956 * Reading ->on_cpu is racy, but this is ok.
2958 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2959 * If we race with it entering cpu, unaccounted time is 0. This is
2960 * indistinguishable from the read occurring a few cycles earlier.
2961 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2962 * been accounted, so we're correct here as well.
2964 if (!p->on_cpu || !task_on_rq_queued(p))
2965 return p->se.sum_exec_runtime;
2968 rq = task_rq_lock(p, &flags);
2970 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2971 * project cycles that may never be accounted to this
2972 * thread, breaking clock_gettime().
2974 if (task_current(rq, p) && task_on_rq_queued(p)) {
2975 update_rq_clock(rq);
2976 p->sched_class->update_curr(rq);
2978 ns = p->se.sum_exec_runtime;
2979 task_rq_unlock(rq, p, &flags);
2985 * This function gets called by the timer code, with HZ frequency.
2986 * We call it with interrupts disabled.
2988 void scheduler_tick(void)
2990 int cpu = smp_processor_id();
2991 struct rq *rq = cpu_rq(cpu);
2992 struct task_struct *curr = rq->curr;
2996 raw_spin_lock(&rq->lock);
2997 walt_set_window_start(rq);
2998 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
2999 walt_ktime_clock(), 0);
3000 update_rq_clock(rq);
3001 curr->sched_class->task_tick(rq, curr, 0);
3002 update_cpu_load_active(rq);
3003 calc_global_load_tick(rq);
3004 raw_spin_unlock(&rq->lock);
3006 perf_event_task_tick();
3009 rq->idle_balance = idle_cpu(cpu);
3010 trigger_load_balance(rq);
3012 rq_last_tick_reset(rq);
3014 if (curr->sched_class == &fair_sched_class)
3015 check_for_migration(rq, curr);
3018 #ifdef CONFIG_NO_HZ_FULL
3020 * scheduler_tick_max_deferment
3022 * Keep at least one tick per second when a single
3023 * active task is running because the scheduler doesn't
3024 * yet completely support full dynticks environment.
3026 * This makes sure that uptime, CFS vruntime, load
3027 * balancing, etc... continue to move forward, even
3028 * with a very low granularity.
3030 * Return: Maximum deferment in nanoseconds.
3032 u64 scheduler_tick_max_deferment(void)
3034 struct rq *rq = this_rq();
3035 unsigned long next, now = READ_ONCE(jiffies);
3037 next = rq->last_sched_tick + HZ;
3039 if (time_before_eq(next, now))
3042 return jiffies_to_nsecs(next - now);
3046 notrace unsigned long get_parent_ip(unsigned long addr)
3048 if (in_lock_functions(addr)) {
3049 addr = CALLER_ADDR2;
3050 if (in_lock_functions(addr))
3051 addr = CALLER_ADDR3;
3056 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3057 defined(CONFIG_PREEMPT_TRACER))
3059 void preempt_count_add(int val)
3061 #ifdef CONFIG_DEBUG_PREEMPT
3065 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3068 __preempt_count_add(val);
3069 #ifdef CONFIG_DEBUG_PREEMPT
3071 * Spinlock count overflowing soon?
3073 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3076 if (preempt_count() == val) {
3077 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3078 #ifdef CONFIG_DEBUG_PREEMPT
3079 current->preempt_disable_ip = ip;
3081 trace_preempt_off(CALLER_ADDR0, ip);
3084 EXPORT_SYMBOL(preempt_count_add);
3085 NOKPROBE_SYMBOL(preempt_count_add);
3087 void preempt_count_sub(int val)
3089 #ifdef CONFIG_DEBUG_PREEMPT
3093 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3096 * Is the spinlock portion underflowing?
3098 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3099 !(preempt_count() & PREEMPT_MASK)))
3103 if (preempt_count() == val)
3104 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3105 __preempt_count_sub(val);
3107 EXPORT_SYMBOL(preempt_count_sub);
3108 NOKPROBE_SYMBOL(preempt_count_sub);
3113 * Print scheduling while atomic bug:
3115 static noinline void __schedule_bug(struct task_struct *prev)
3117 if (oops_in_progress)
3120 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3121 prev->comm, prev->pid, preempt_count());
3123 debug_show_held_locks(prev);
3125 if (irqs_disabled())
3126 print_irqtrace_events(prev);
3127 #ifdef CONFIG_DEBUG_PREEMPT
3128 if (in_atomic_preempt_off()) {
3129 pr_err("Preemption disabled at:");
3130 print_ip_sym(current->preempt_disable_ip);
3135 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3139 * Various schedule()-time debugging checks and statistics:
3141 static inline void schedule_debug(struct task_struct *prev)
3143 #ifdef CONFIG_SCHED_STACK_END_CHECK
3144 if (task_stack_end_corrupted(prev))
3145 panic("corrupted stack end detected inside scheduler\n");
3148 if (unlikely(in_atomic_preempt_off())) {
3149 __schedule_bug(prev);
3150 preempt_count_set(PREEMPT_DISABLED);
3154 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3156 schedstat_inc(this_rq(), sched_count);
3160 * Pick up the highest-prio task:
3162 static inline struct task_struct *
3163 pick_next_task(struct rq *rq, struct task_struct *prev)
3165 const struct sched_class *class = &fair_sched_class;
3166 struct task_struct *p;
3169 * Optimization: we know that if all tasks are in
3170 * the fair class we can call that function directly:
3172 if (likely(prev->sched_class == class &&
3173 rq->nr_running == rq->cfs.h_nr_running)) {
3174 p = fair_sched_class.pick_next_task(rq, prev);
3175 if (unlikely(p == RETRY_TASK))
3178 /* assumes fair_sched_class->next == idle_sched_class */
3180 p = idle_sched_class.pick_next_task(rq, prev);
3186 for_each_class(class) {
3187 p = class->pick_next_task(rq, prev);
3189 if (unlikely(p == RETRY_TASK))
3195 BUG(); /* the idle class will always have a runnable task */
3199 * __schedule() is the main scheduler function.
3201 * The main means of driving the scheduler and thus entering this function are:
3203 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3205 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3206 * paths. For example, see arch/x86/entry_64.S.
3208 * To drive preemption between tasks, the scheduler sets the flag in timer
3209 * interrupt handler scheduler_tick().
3211 * 3. Wakeups don't really cause entry into schedule(). They add a
3212 * task to the run-queue and that's it.
3214 * Now, if the new task added to the run-queue preempts the current
3215 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3216 * called on the nearest possible occasion:
3218 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3220 * - in syscall or exception context, at the next outmost
3221 * preempt_enable(). (this might be as soon as the wake_up()'s
3224 * - in IRQ context, return from interrupt-handler to
3225 * preemptible context
3227 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3230 * - cond_resched() call
3231 * - explicit schedule() call
3232 * - return from syscall or exception to user-space
3233 * - return from interrupt-handler to user-space
3235 * WARNING: must be called with preemption disabled!
3237 static void __sched notrace __schedule(bool preempt)
3239 struct task_struct *prev, *next;
3240 unsigned long *switch_count;
3245 cpu = smp_processor_id();
3247 rcu_note_context_switch();
3251 * do_exit() calls schedule() with preemption disabled as an exception;
3252 * however we must fix that up, otherwise the next task will see an
3253 * inconsistent (higher) preempt count.
3255 * It also avoids the below schedule_debug() test from complaining
3258 if (unlikely(prev->state == TASK_DEAD))
3259 preempt_enable_no_resched_notrace();
3261 schedule_debug(prev);
3263 if (sched_feat(HRTICK))
3267 * Make sure that signal_pending_state()->signal_pending() below
3268 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3269 * done by the caller to avoid the race with signal_wake_up().
3271 smp_mb__before_spinlock();
3272 raw_spin_lock_irq(&rq->lock);
3273 lockdep_pin_lock(&rq->lock);
3275 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3277 switch_count = &prev->nivcsw;
3278 if (!preempt && prev->state) {
3279 if (unlikely(signal_pending_state(prev->state, prev))) {
3280 prev->state = TASK_RUNNING;
3282 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3286 * If a worker went to sleep, notify and ask workqueue
3287 * whether it wants to wake up a task to maintain
3290 if (prev->flags & PF_WQ_WORKER) {
3291 struct task_struct *to_wakeup;
3293 to_wakeup = wq_worker_sleeping(prev, cpu);
3295 try_to_wake_up_local(to_wakeup);
3298 switch_count = &prev->nvcsw;
3301 if (task_on_rq_queued(prev))
3302 update_rq_clock(rq);
3304 next = pick_next_task(rq, prev);
3305 wallclock = walt_ktime_clock();
3306 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3307 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3308 clear_tsk_need_resched(prev);
3309 clear_preempt_need_resched();
3310 rq->clock_skip_update = 0;
3312 if (likely(prev != next)) {
3313 #ifdef CONFIG_SCHED_WALT
3315 prev->last_sleep_ts = wallclock;
3321 trace_sched_switch(preempt, prev, next);
3322 rq = context_switch(rq, prev, next); /* unlocks the rq */
3325 lockdep_unpin_lock(&rq->lock);
3326 raw_spin_unlock_irq(&rq->lock);
3329 balance_callback(rq);
3332 static inline void sched_submit_work(struct task_struct *tsk)
3334 if (!tsk->state || tsk_is_pi_blocked(tsk))
3337 * If we are going to sleep and we have plugged IO queued,
3338 * make sure to submit it to avoid deadlocks.
3340 if (blk_needs_flush_plug(tsk))
3341 blk_schedule_flush_plug(tsk);
3344 asmlinkage __visible void __sched schedule(void)
3346 struct task_struct *tsk = current;
3348 sched_submit_work(tsk);
3352 sched_preempt_enable_no_resched();
3353 } while (need_resched());
3355 EXPORT_SYMBOL(schedule);
3357 #ifdef CONFIG_CONTEXT_TRACKING
3358 asmlinkage __visible void __sched schedule_user(void)
3361 * If we come here after a random call to set_need_resched(),
3362 * or we have been woken up remotely but the IPI has not yet arrived,
3363 * we haven't yet exited the RCU idle mode. Do it here manually until
3364 * we find a better solution.
3366 * NB: There are buggy callers of this function. Ideally we
3367 * should warn if prev_state != CONTEXT_USER, but that will trigger
3368 * too frequently to make sense yet.
3370 enum ctx_state prev_state = exception_enter();
3372 exception_exit(prev_state);
3377 * schedule_preempt_disabled - called with preemption disabled
3379 * Returns with preemption disabled. Note: preempt_count must be 1
3381 void __sched schedule_preempt_disabled(void)
3383 sched_preempt_enable_no_resched();
3388 static void __sched notrace preempt_schedule_common(void)
3391 preempt_disable_notrace();
3393 preempt_enable_no_resched_notrace();
3396 * Check again in case we missed a preemption opportunity
3397 * between schedule and now.
3399 } while (need_resched());
3402 #ifdef CONFIG_PREEMPT
3404 * this is the entry point to schedule() from in-kernel preemption
3405 * off of preempt_enable. Kernel preemptions off return from interrupt
3406 * occur there and call schedule directly.
3408 asmlinkage __visible void __sched notrace preempt_schedule(void)
3411 * If there is a non-zero preempt_count or interrupts are disabled,
3412 * we do not want to preempt the current task. Just return..
3414 if (likely(!preemptible()))
3417 preempt_schedule_common();
3419 NOKPROBE_SYMBOL(preempt_schedule);
3420 EXPORT_SYMBOL(preempt_schedule);
3423 * preempt_schedule_notrace - preempt_schedule called by tracing
3425 * The tracing infrastructure uses preempt_enable_notrace to prevent
3426 * recursion and tracing preempt enabling caused by the tracing
3427 * infrastructure itself. But as tracing can happen in areas coming
3428 * from userspace or just about to enter userspace, a preempt enable
3429 * can occur before user_exit() is called. This will cause the scheduler
3430 * to be called when the system is still in usermode.
3432 * To prevent this, the preempt_enable_notrace will use this function
3433 * instead of preempt_schedule() to exit user context if needed before
3434 * calling the scheduler.
3436 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3438 enum ctx_state prev_ctx;
3440 if (likely(!preemptible()))
3444 preempt_disable_notrace();
3446 * Needs preempt disabled in case user_exit() is traced
3447 * and the tracer calls preempt_enable_notrace() causing
3448 * an infinite recursion.
3450 prev_ctx = exception_enter();
3452 exception_exit(prev_ctx);
3454 preempt_enable_no_resched_notrace();
3455 } while (need_resched());
3457 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3459 #endif /* CONFIG_PREEMPT */
3462 * this is the entry point to schedule() from kernel preemption
3463 * off of irq context.
3464 * Note, that this is called and return with irqs disabled. This will
3465 * protect us against recursive calling from irq.
3467 asmlinkage __visible void __sched preempt_schedule_irq(void)
3469 enum ctx_state prev_state;
3471 /* Catch callers which need to be fixed */
3472 BUG_ON(preempt_count() || !irqs_disabled());
3474 prev_state = exception_enter();
3480 local_irq_disable();
3481 sched_preempt_enable_no_resched();
3482 } while (need_resched());
3484 exception_exit(prev_state);
3487 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3490 return try_to_wake_up(curr->private, mode, wake_flags, 1);
3492 EXPORT_SYMBOL(default_wake_function);
3494 #ifdef CONFIG_RT_MUTEXES
3497 * rt_mutex_setprio - set the current priority of a task
3499 * @prio: prio value (kernel-internal form)
3501 * This function changes the 'effective' priority of a task. It does
3502 * not touch ->normal_prio like __setscheduler().
3504 * Used by the rt_mutex code to implement priority inheritance
3505 * logic. Call site only calls if the priority of the task changed.
3507 void rt_mutex_setprio(struct task_struct *p, int prio)
3509 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3511 const struct sched_class *prev_class;
3513 BUG_ON(prio > MAX_PRIO);
3515 rq = __task_rq_lock(p);
3516 update_rq_clock(rq);
3519 * Idle task boosting is a nono in general. There is one
3520 * exception, when PREEMPT_RT and NOHZ is active:
3522 * The idle task calls get_next_timer_interrupt() and holds
3523 * the timer wheel base->lock on the CPU and another CPU wants
3524 * to access the timer (probably to cancel it). We can safely
3525 * ignore the boosting request, as the idle CPU runs this code
3526 * with interrupts disabled and will complete the lock
3527 * protected section without being interrupted. So there is no
3528 * real need to boost.
3530 if (unlikely(p == rq->idle)) {
3531 WARN_ON(p != rq->curr);
3532 WARN_ON(p->pi_blocked_on);
3536 trace_sched_pi_setprio(p, prio);
3538 prev_class = p->sched_class;
3539 queued = task_on_rq_queued(p);
3540 running = task_current(rq, p);
3542 dequeue_task(rq, p, DEQUEUE_SAVE);
3544 put_prev_task(rq, p);
3547 * Boosting condition are:
3548 * 1. -rt task is running and holds mutex A
3549 * --> -dl task blocks on mutex A
3551 * 2. -dl task is running and holds mutex A
3552 * --> -dl task blocks on mutex A and could preempt the
3555 if (dl_prio(prio)) {
3556 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3557 if (!dl_prio(p->normal_prio) ||
3558 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3559 p->dl.dl_boosted = 1;
3560 enqueue_flag |= ENQUEUE_REPLENISH;
3562 p->dl.dl_boosted = 0;
3563 p->sched_class = &dl_sched_class;
3564 } else if (rt_prio(prio)) {
3565 if (dl_prio(oldprio))
3566 p->dl.dl_boosted = 0;
3568 enqueue_flag |= ENQUEUE_HEAD;
3569 p->sched_class = &rt_sched_class;
3571 if (dl_prio(oldprio))
3572 p->dl.dl_boosted = 0;
3573 if (rt_prio(oldprio))
3575 p->sched_class = &fair_sched_class;
3581 p->sched_class->set_curr_task(rq);
3583 enqueue_task(rq, p, enqueue_flag);
3585 check_class_changed(rq, p, prev_class, oldprio);
3587 preempt_disable(); /* avoid rq from going away on us */
3588 __task_rq_unlock(rq);
3590 balance_callback(rq);
3595 void set_user_nice(struct task_struct *p, long nice)
3597 int old_prio, delta, queued;
3598 unsigned long flags;
3601 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3604 * We have to be careful, if called from sys_setpriority(),
3605 * the task might be in the middle of scheduling on another CPU.
3607 rq = task_rq_lock(p, &flags);
3608 update_rq_clock(rq);
3611 * The RT priorities are set via sched_setscheduler(), but we still
3612 * allow the 'normal' nice value to be set - but as expected
3613 * it wont have any effect on scheduling until the task is
3614 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3616 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3617 p->static_prio = NICE_TO_PRIO(nice);
3620 queued = task_on_rq_queued(p);
3622 dequeue_task(rq, p, DEQUEUE_SAVE);
3624 p->static_prio = NICE_TO_PRIO(nice);
3627 p->prio = effective_prio(p);
3628 delta = p->prio - old_prio;
3631 enqueue_task(rq, p, ENQUEUE_RESTORE);
3633 * If the task increased its priority or is running and
3634 * lowered its priority, then reschedule its CPU:
3636 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3640 task_rq_unlock(rq, p, &flags);
3642 EXPORT_SYMBOL(set_user_nice);
3645 * can_nice - check if a task can reduce its nice value
3649 int can_nice(const struct task_struct *p, const int nice)
3651 /* convert nice value [19,-20] to rlimit style value [1,40] */
3652 int nice_rlim = nice_to_rlimit(nice);
3654 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3655 capable(CAP_SYS_NICE));
3658 #ifdef __ARCH_WANT_SYS_NICE
3661 * sys_nice - change the priority of the current process.
3662 * @increment: priority increment
3664 * sys_setpriority is a more generic, but much slower function that
3665 * does similar things.
3667 SYSCALL_DEFINE1(nice, int, increment)
3672 * Setpriority might change our priority at the same moment.
3673 * We don't have to worry. Conceptually one call occurs first
3674 * and we have a single winner.
3676 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3677 nice = task_nice(current) + increment;
3679 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3680 if (increment < 0 && !can_nice(current, nice))
3683 retval = security_task_setnice(current, nice);
3687 set_user_nice(current, nice);
3694 * task_prio - return the priority value of a given task.
3695 * @p: the task in question.
3697 * Return: The priority value as seen by users in /proc.
3698 * RT tasks are offset by -200. Normal tasks are centered
3699 * around 0, value goes from -16 to +15.
3701 int task_prio(const struct task_struct *p)
3703 return p->prio - MAX_RT_PRIO;
3707 * idle_cpu - is a given cpu idle currently?
3708 * @cpu: the processor in question.
3710 * Return: 1 if the CPU is currently idle. 0 otherwise.
3712 int idle_cpu(int cpu)
3714 struct rq *rq = cpu_rq(cpu);
3716 if (rq->curr != rq->idle)
3723 if (!llist_empty(&rq->wake_list))
3731 * idle_task - return the idle task for a given cpu.
3732 * @cpu: the processor in question.
3734 * Return: The idle task for the cpu @cpu.
3736 struct task_struct *idle_task(int cpu)
3738 return cpu_rq(cpu)->idle;
3742 * find_process_by_pid - find a process with a matching PID value.
3743 * @pid: the pid in question.
3745 * The task of @pid, if found. %NULL otherwise.
3747 static struct task_struct *find_process_by_pid(pid_t pid)
3749 return pid ? find_task_by_vpid(pid) : current;
3753 * This function initializes the sched_dl_entity of a newly becoming
3754 * SCHED_DEADLINE task.
3756 * Only the static values are considered here, the actual runtime and the
3757 * absolute deadline will be properly calculated when the task is enqueued
3758 * for the first time with its new policy.
3761 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3763 struct sched_dl_entity *dl_se = &p->dl;
3765 dl_se->dl_runtime = attr->sched_runtime;
3766 dl_se->dl_deadline = attr->sched_deadline;
3767 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3768 dl_se->flags = attr->sched_flags;
3769 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3772 * Changing the parameters of a task is 'tricky' and we're not doing
3773 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3775 * What we SHOULD do is delay the bandwidth release until the 0-lag
3776 * point. This would include retaining the task_struct until that time
3777 * and change dl_overflow() to not immediately decrement the current
3780 * Instead we retain the current runtime/deadline and let the new
3781 * parameters take effect after the current reservation period lapses.
3782 * This is safe (albeit pessimistic) because the 0-lag point is always
3783 * before the current scheduling deadline.
3785 * We can still have temporary overloads because we do not delay the
3786 * change in bandwidth until that time; so admission control is
3787 * not on the safe side. It does however guarantee tasks will never
3788 * consume more than promised.
3793 * sched_setparam() passes in -1 for its policy, to let the functions
3794 * it calls know not to change it.
3796 #define SETPARAM_POLICY -1
3798 static void __setscheduler_params(struct task_struct *p,
3799 const struct sched_attr *attr)
3801 int policy = attr->sched_policy;
3803 if (policy == SETPARAM_POLICY)
3808 if (dl_policy(policy))
3809 __setparam_dl(p, attr);
3810 else if (fair_policy(policy))
3811 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3814 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3815 * !rt_policy. Always setting this ensures that things like
3816 * getparam()/getattr() don't report silly values for !rt tasks.
3818 p->rt_priority = attr->sched_priority;
3819 p->normal_prio = normal_prio(p);
3823 /* Actually do priority change: must hold pi & rq lock. */
3824 static void __setscheduler(struct rq *rq, struct task_struct *p,
3825 const struct sched_attr *attr, bool keep_boost)
3827 __setscheduler_params(p, attr);
3830 * Keep a potential priority boosting if called from
3831 * sched_setscheduler().
3834 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3836 p->prio = normal_prio(p);
3838 if (dl_prio(p->prio))
3839 p->sched_class = &dl_sched_class;
3840 else if (rt_prio(p->prio))
3841 p->sched_class = &rt_sched_class;
3843 p->sched_class = &fair_sched_class;
3847 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3849 struct sched_dl_entity *dl_se = &p->dl;
3851 attr->sched_priority = p->rt_priority;
3852 attr->sched_runtime = dl_se->dl_runtime;
3853 attr->sched_deadline = dl_se->dl_deadline;
3854 attr->sched_period = dl_se->dl_period;
3855 attr->sched_flags = dl_se->flags;
3859 * This function validates the new parameters of a -deadline task.
3860 * We ask for the deadline not being zero, and greater or equal
3861 * than the runtime, as well as the period of being zero or
3862 * greater than deadline. Furthermore, we have to be sure that
3863 * user parameters are above the internal resolution of 1us (we
3864 * check sched_runtime only since it is always the smaller one) and
3865 * below 2^63 ns (we have to check both sched_deadline and
3866 * sched_period, as the latter can be zero).
3869 __checkparam_dl(const struct sched_attr *attr)
3872 if (attr->sched_deadline == 0)
3876 * Since we truncate DL_SCALE bits, make sure we're at least
3879 if (attr->sched_runtime < (1ULL << DL_SCALE))
3883 * Since we use the MSB for wrap-around and sign issues, make
3884 * sure it's not set (mind that period can be equal to zero).
3886 if (attr->sched_deadline & (1ULL << 63) ||
3887 attr->sched_period & (1ULL << 63))
3890 /* runtime <= deadline <= period (if period != 0) */
3891 if ((attr->sched_period != 0 &&
3892 attr->sched_period < attr->sched_deadline) ||
3893 attr->sched_deadline < attr->sched_runtime)
3900 * check the target process has a UID that matches the current process's
3902 static bool check_same_owner(struct task_struct *p)
3904 const struct cred *cred = current_cred(), *pcred;
3908 pcred = __task_cred(p);
3909 match = (uid_eq(cred->euid, pcred->euid) ||
3910 uid_eq(cred->euid, pcred->uid));
3915 static bool dl_param_changed(struct task_struct *p,
3916 const struct sched_attr *attr)
3918 struct sched_dl_entity *dl_se = &p->dl;
3920 if (dl_se->dl_runtime != attr->sched_runtime ||
3921 dl_se->dl_deadline != attr->sched_deadline ||
3922 dl_se->dl_period != attr->sched_period ||
3923 dl_se->flags != attr->sched_flags)
3929 static int __sched_setscheduler(struct task_struct *p,
3930 const struct sched_attr *attr,
3933 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3934 MAX_RT_PRIO - 1 - attr->sched_priority;
3935 int retval, oldprio, oldpolicy = -1, queued, running;
3936 int new_effective_prio, policy = attr->sched_policy;
3937 unsigned long flags;
3938 const struct sched_class *prev_class;
3942 /* may grab non-irq protected spin_locks */
3943 BUG_ON(in_interrupt());
3945 /* double check policy once rq lock held */
3947 reset_on_fork = p->sched_reset_on_fork;
3948 policy = oldpolicy = p->policy;
3950 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3952 if (!valid_policy(policy))
3956 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3960 * Valid priorities for SCHED_FIFO and SCHED_RR are
3961 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3962 * SCHED_BATCH and SCHED_IDLE is 0.
3964 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3965 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3967 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3968 (rt_policy(policy) != (attr->sched_priority != 0)))
3972 * Allow unprivileged RT tasks to decrease priority:
3974 if (user && !capable(CAP_SYS_NICE)) {
3975 if (fair_policy(policy)) {
3976 if (attr->sched_nice < task_nice(p) &&
3977 !can_nice(p, attr->sched_nice))
3981 if (rt_policy(policy)) {
3982 unsigned long rlim_rtprio =
3983 task_rlimit(p, RLIMIT_RTPRIO);
3985 /* can't set/change the rt policy */
3986 if (policy != p->policy && !rlim_rtprio)
3989 /* can't increase priority */
3990 if (attr->sched_priority > p->rt_priority &&
3991 attr->sched_priority > rlim_rtprio)
3996 * Can't set/change SCHED_DEADLINE policy at all for now
3997 * (safest behavior); in the future we would like to allow
3998 * unprivileged DL tasks to increase their relative deadline
3999 * or reduce their runtime (both ways reducing utilization)
4001 if (dl_policy(policy))
4005 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4006 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4008 if (idle_policy(p->policy) && !idle_policy(policy)) {
4009 if (!can_nice(p, task_nice(p)))
4013 /* can't change other user's priorities */
4014 if (!check_same_owner(p))
4017 /* Normal users shall not reset the sched_reset_on_fork flag */
4018 if (p->sched_reset_on_fork && !reset_on_fork)
4023 retval = security_task_setscheduler(p);
4029 * make sure no PI-waiters arrive (or leave) while we are
4030 * changing the priority of the task:
4032 * To be able to change p->policy safely, the appropriate
4033 * runqueue lock must be held.
4035 rq = task_rq_lock(p, &flags);
4036 update_rq_clock(rq);
4039 * Changing the policy of the stop threads its a very bad idea
4041 if (p == rq->stop) {
4042 task_rq_unlock(rq, p, &flags);
4047 * If not changing anything there's no need to proceed further,
4048 * but store a possible modification of reset_on_fork.
4050 if (unlikely(policy == p->policy)) {
4051 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4053 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4055 if (dl_policy(policy) && dl_param_changed(p, attr))
4058 p->sched_reset_on_fork = reset_on_fork;
4059 task_rq_unlock(rq, p, &flags);
4065 #ifdef CONFIG_RT_GROUP_SCHED
4067 * Do not allow realtime tasks into groups that have no runtime
4070 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4071 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4072 !task_group_is_autogroup(task_group(p))) {
4073 task_rq_unlock(rq, p, &flags);
4078 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4079 cpumask_t *span = rq->rd->span;
4082 * Don't allow tasks with an affinity mask smaller than
4083 * the entire root_domain to become SCHED_DEADLINE. We
4084 * will also fail if there's no bandwidth available.
4086 if (!cpumask_subset(span, &p->cpus_allowed) ||
4087 rq->rd->dl_bw.bw == 0) {
4088 task_rq_unlock(rq, p, &flags);
4095 /* recheck policy now with rq lock held */
4096 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4097 policy = oldpolicy = -1;
4098 task_rq_unlock(rq, p, &flags);
4103 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4104 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4107 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4108 task_rq_unlock(rq, p, &flags);
4112 p->sched_reset_on_fork = reset_on_fork;
4117 * Take priority boosted tasks into account. If the new
4118 * effective priority is unchanged, we just store the new
4119 * normal parameters and do not touch the scheduler class and
4120 * the runqueue. This will be done when the task deboost
4123 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4124 if (new_effective_prio == oldprio) {
4125 __setscheduler_params(p, attr);
4126 task_rq_unlock(rq, p, &flags);
4131 queued = task_on_rq_queued(p);
4132 running = task_current(rq, p);
4134 dequeue_task(rq, p, DEQUEUE_SAVE);
4136 put_prev_task(rq, p);
4138 prev_class = p->sched_class;
4139 __setscheduler(rq, p, attr, pi);
4142 p->sched_class->set_curr_task(rq);
4144 int enqueue_flags = ENQUEUE_RESTORE;
4146 * We enqueue to tail when the priority of a task is
4147 * increased (user space view).
4149 if (oldprio <= p->prio)
4150 enqueue_flags |= ENQUEUE_HEAD;
4152 enqueue_task(rq, p, enqueue_flags);
4155 check_class_changed(rq, p, prev_class, oldprio);
4156 preempt_disable(); /* avoid rq from going away on us */
4157 task_rq_unlock(rq, p, &flags);
4160 rt_mutex_adjust_pi(p);
4163 * Run balance callbacks after we've adjusted the PI chain.
4165 balance_callback(rq);
4171 static int _sched_setscheduler(struct task_struct *p, int policy,
4172 const struct sched_param *param, bool check)
4174 struct sched_attr attr = {
4175 .sched_policy = policy,
4176 .sched_priority = param->sched_priority,
4177 .sched_nice = PRIO_TO_NICE(p->static_prio),
4180 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4181 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4182 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4183 policy &= ~SCHED_RESET_ON_FORK;
4184 attr.sched_policy = policy;
4187 return __sched_setscheduler(p, &attr, check, true);
4190 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4191 * @p: the task in question.
4192 * @policy: new policy.
4193 * @param: structure containing the new RT priority.
4195 * Return: 0 on success. An error code otherwise.
4197 * NOTE that the task may be already dead.
4199 int sched_setscheduler(struct task_struct *p, int policy,
4200 const struct sched_param *param)
4202 return _sched_setscheduler(p, policy, param, true);
4204 EXPORT_SYMBOL_GPL(sched_setscheduler);
4206 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4208 return __sched_setscheduler(p, attr, true, true);
4210 EXPORT_SYMBOL_GPL(sched_setattr);
4213 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4214 * @p: the task in question.
4215 * @policy: new policy.
4216 * @param: structure containing the new RT priority.
4218 * Just like sched_setscheduler, only don't bother checking if the
4219 * current context has permission. For example, this is needed in
4220 * stop_machine(): we create temporary high priority worker threads,
4221 * but our caller might not have that capability.
4223 * Return: 0 on success. An error code otherwise.
4225 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4226 const struct sched_param *param)
4228 return _sched_setscheduler(p, policy, param, false);
4230 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4233 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4235 struct sched_param lparam;
4236 struct task_struct *p;
4239 if (!param || pid < 0)
4241 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4246 p = find_process_by_pid(pid);
4248 retval = sched_setscheduler(p, policy, &lparam);
4255 * Mimics kernel/events/core.c perf_copy_attr().
4257 static int sched_copy_attr(struct sched_attr __user *uattr,
4258 struct sched_attr *attr)
4263 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4267 * zero the full structure, so that a short copy will be nice.
4269 memset(attr, 0, sizeof(*attr));
4271 ret = get_user(size, &uattr->size);
4275 if (size > PAGE_SIZE) /* silly large */
4278 if (!size) /* abi compat */
4279 size = SCHED_ATTR_SIZE_VER0;
4281 if (size < SCHED_ATTR_SIZE_VER0)
4285 * If we're handed a bigger struct than we know of,
4286 * ensure all the unknown bits are 0 - i.e. new
4287 * user-space does not rely on any kernel feature
4288 * extensions we dont know about yet.
4290 if (size > sizeof(*attr)) {
4291 unsigned char __user *addr;
4292 unsigned char __user *end;
4295 addr = (void __user *)uattr + sizeof(*attr);
4296 end = (void __user *)uattr + size;
4298 for (; addr < end; addr++) {
4299 ret = get_user(val, addr);
4305 size = sizeof(*attr);
4308 ret = copy_from_user(attr, uattr, size);
4313 * XXX: do we want to be lenient like existing syscalls; or do we want
4314 * to be strict and return an error on out-of-bounds values?
4316 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4321 put_user(sizeof(*attr), &uattr->size);
4326 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4327 * @pid: the pid in question.
4328 * @policy: new policy.
4329 * @param: structure containing the new RT priority.
4331 * Return: 0 on success. An error code otherwise.
4333 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4334 struct sched_param __user *, param)
4336 /* negative values for policy are not valid */
4340 return do_sched_setscheduler(pid, policy, param);
4344 * sys_sched_setparam - set/change the RT priority of a thread
4345 * @pid: the pid in question.
4346 * @param: structure containing the new RT priority.
4348 * Return: 0 on success. An error code otherwise.
4350 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4352 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4356 * sys_sched_setattr - same as above, but with extended sched_attr
4357 * @pid: the pid in question.
4358 * @uattr: structure containing the extended parameters.
4359 * @flags: for future extension.
4361 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4362 unsigned int, flags)
4364 struct sched_attr attr;
4365 struct task_struct *p;
4368 if (!uattr || pid < 0 || flags)
4371 retval = sched_copy_attr(uattr, &attr);
4375 if ((int)attr.sched_policy < 0)
4380 p = find_process_by_pid(pid);
4382 retval = sched_setattr(p, &attr);
4389 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4390 * @pid: the pid in question.
4392 * Return: On success, the policy of the thread. Otherwise, a negative error
4395 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4397 struct task_struct *p;
4405 p = find_process_by_pid(pid);
4407 retval = security_task_getscheduler(p);
4410 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4417 * sys_sched_getparam - get the RT priority of a thread
4418 * @pid: the pid in question.
4419 * @param: structure containing the RT priority.
4421 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4424 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4426 struct sched_param lp = { .sched_priority = 0 };
4427 struct task_struct *p;
4430 if (!param || pid < 0)
4434 p = find_process_by_pid(pid);
4439 retval = security_task_getscheduler(p);
4443 if (task_has_rt_policy(p))
4444 lp.sched_priority = p->rt_priority;
4448 * This one might sleep, we cannot do it with a spinlock held ...
4450 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4459 static int sched_read_attr(struct sched_attr __user *uattr,
4460 struct sched_attr *attr,
4465 if (!access_ok(VERIFY_WRITE, uattr, usize))
4469 * If we're handed a smaller struct than we know of,
4470 * ensure all the unknown bits are 0 - i.e. old
4471 * user-space does not get uncomplete information.
4473 if (usize < sizeof(*attr)) {
4474 unsigned char *addr;
4477 addr = (void *)attr + usize;
4478 end = (void *)attr + sizeof(*attr);
4480 for (; addr < end; addr++) {
4488 ret = copy_to_user(uattr, attr, attr->size);
4496 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4497 * @pid: the pid in question.
4498 * @uattr: structure containing the extended parameters.
4499 * @size: sizeof(attr) for fwd/bwd comp.
4500 * @flags: for future extension.
4502 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4503 unsigned int, size, unsigned int, flags)
4505 struct sched_attr attr = {
4506 .size = sizeof(struct sched_attr),
4508 struct task_struct *p;
4511 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4512 size < SCHED_ATTR_SIZE_VER0 || flags)
4516 p = find_process_by_pid(pid);
4521 retval = security_task_getscheduler(p);
4525 attr.sched_policy = p->policy;
4526 if (p->sched_reset_on_fork)
4527 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4528 if (task_has_dl_policy(p))
4529 __getparam_dl(p, &attr);
4530 else if (task_has_rt_policy(p))
4531 attr.sched_priority = p->rt_priority;
4533 attr.sched_nice = task_nice(p);
4537 retval = sched_read_attr(uattr, &attr, size);
4545 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4547 cpumask_var_t cpus_allowed, new_mask;
4548 struct task_struct *p;
4553 p = find_process_by_pid(pid);
4559 /* Prevent p going away */
4563 if (p->flags & PF_NO_SETAFFINITY) {
4567 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4571 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4573 goto out_free_cpus_allowed;
4576 if (!check_same_owner(p)) {
4578 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4580 goto out_free_new_mask;
4585 retval = security_task_setscheduler(p);
4587 goto out_free_new_mask;
4590 cpuset_cpus_allowed(p, cpus_allowed);
4591 cpumask_and(new_mask, in_mask, cpus_allowed);
4594 * Since bandwidth control happens on root_domain basis,
4595 * if admission test is enabled, we only admit -deadline
4596 * tasks allowed to run on all the CPUs in the task's
4600 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4602 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4605 goto out_free_new_mask;
4611 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4614 cpuset_cpus_allowed(p, cpus_allowed);
4615 if (!cpumask_subset(new_mask, cpus_allowed)) {
4617 * We must have raced with a concurrent cpuset
4618 * update. Just reset the cpus_allowed to the
4619 * cpuset's cpus_allowed
4621 cpumask_copy(new_mask, cpus_allowed);
4626 free_cpumask_var(new_mask);
4627 out_free_cpus_allowed:
4628 free_cpumask_var(cpus_allowed);
4634 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4635 struct cpumask *new_mask)
4637 if (len < cpumask_size())
4638 cpumask_clear(new_mask);
4639 else if (len > cpumask_size())
4640 len = cpumask_size();
4642 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4646 * sys_sched_setaffinity - set the cpu affinity of a process
4647 * @pid: pid of the process
4648 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4649 * @user_mask_ptr: user-space pointer to the new cpu mask
4651 * Return: 0 on success. An error code otherwise.
4653 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4654 unsigned long __user *, user_mask_ptr)
4656 cpumask_var_t new_mask;
4659 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4662 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4664 retval = sched_setaffinity(pid, new_mask);
4665 free_cpumask_var(new_mask);
4669 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4671 struct task_struct *p;
4672 unsigned long flags;
4678 p = find_process_by_pid(pid);
4682 retval = security_task_getscheduler(p);
4686 raw_spin_lock_irqsave(&p->pi_lock, flags);
4687 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4688 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4697 * sys_sched_getaffinity - get the cpu affinity of a process
4698 * @pid: pid of the process
4699 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4700 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4702 * Return: 0 on success. An error code otherwise.
4704 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4705 unsigned long __user *, user_mask_ptr)
4710 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4712 if (len & (sizeof(unsigned long)-1))
4715 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4718 ret = sched_getaffinity(pid, mask);
4720 size_t retlen = min_t(size_t, len, cpumask_size());
4722 if (copy_to_user(user_mask_ptr, mask, retlen))
4727 free_cpumask_var(mask);
4733 * sys_sched_yield - yield the current processor to other threads.
4735 * This function yields the current CPU to other tasks. If there are no
4736 * other threads running on this CPU then this function will return.
4740 SYSCALL_DEFINE0(sched_yield)
4742 struct rq *rq = this_rq_lock();
4744 schedstat_inc(rq, yld_count);
4745 current->sched_class->yield_task(rq);
4748 * Since we are going to call schedule() anyway, there's
4749 * no need to preempt or enable interrupts:
4751 __release(rq->lock);
4752 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4753 do_raw_spin_unlock(&rq->lock);
4754 sched_preempt_enable_no_resched();
4761 int __sched _cond_resched(void)
4763 if (should_resched(0)) {
4764 preempt_schedule_common();
4769 EXPORT_SYMBOL(_cond_resched);
4772 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4773 * call schedule, and on return reacquire the lock.
4775 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4776 * operations here to prevent schedule() from being called twice (once via
4777 * spin_unlock(), once by hand).
4779 int __cond_resched_lock(spinlock_t *lock)
4781 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4784 lockdep_assert_held(lock);
4786 if (spin_needbreak(lock) || resched) {
4789 preempt_schedule_common();
4797 EXPORT_SYMBOL(__cond_resched_lock);
4799 int __sched __cond_resched_softirq(void)
4801 BUG_ON(!in_softirq());
4803 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4805 preempt_schedule_common();
4811 EXPORT_SYMBOL(__cond_resched_softirq);
4814 * yield - yield the current processor to other threads.
4816 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4818 * The scheduler is at all times free to pick the calling task as the most
4819 * eligible task to run, if removing the yield() call from your code breaks
4820 * it, its already broken.
4822 * Typical broken usage is:
4827 * where one assumes that yield() will let 'the other' process run that will
4828 * make event true. If the current task is a SCHED_FIFO task that will never
4829 * happen. Never use yield() as a progress guarantee!!
4831 * If you want to use yield() to wait for something, use wait_event().
4832 * If you want to use yield() to be 'nice' for others, use cond_resched().
4833 * If you still want to use yield(), do not!
4835 void __sched yield(void)
4837 set_current_state(TASK_RUNNING);
4840 EXPORT_SYMBOL(yield);
4843 * yield_to - yield the current processor to another thread in
4844 * your thread group, or accelerate that thread toward the
4845 * processor it's on.
4847 * @preempt: whether task preemption is allowed or not
4849 * It's the caller's job to ensure that the target task struct
4850 * can't go away on us before we can do any checks.
4853 * true (>0) if we indeed boosted the target task.
4854 * false (0) if we failed to boost the target.
4855 * -ESRCH if there's no task to yield to.
4857 int __sched yield_to(struct task_struct *p, bool preempt)
4859 struct task_struct *curr = current;
4860 struct rq *rq, *p_rq;
4861 unsigned long flags;
4864 local_irq_save(flags);
4870 * If we're the only runnable task on the rq and target rq also
4871 * has only one task, there's absolutely no point in yielding.
4873 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4878 double_rq_lock(rq, p_rq);
4879 if (task_rq(p) != p_rq) {
4880 double_rq_unlock(rq, p_rq);
4884 if (!curr->sched_class->yield_to_task)
4887 if (curr->sched_class != p->sched_class)
4890 if (task_running(p_rq, p) || p->state)
4893 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4895 schedstat_inc(rq, yld_count);
4897 * Make p's CPU reschedule; pick_next_entity takes care of
4900 if (preempt && rq != p_rq)
4905 double_rq_unlock(rq, p_rq);
4907 local_irq_restore(flags);
4914 EXPORT_SYMBOL_GPL(yield_to);
4917 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4918 * that process accounting knows that this is a task in IO wait state.
4920 long __sched io_schedule_timeout(long timeout)
4922 int old_iowait = current->in_iowait;
4926 current->in_iowait = 1;
4927 blk_schedule_flush_plug(current);
4929 delayacct_blkio_start();
4931 atomic_inc(&rq->nr_iowait);
4932 ret = schedule_timeout(timeout);
4933 current->in_iowait = old_iowait;
4934 atomic_dec(&rq->nr_iowait);
4935 delayacct_blkio_end();
4939 EXPORT_SYMBOL(io_schedule_timeout);
4942 * sys_sched_get_priority_max - return maximum RT priority.
4943 * @policy: scheduling class.
4945 * Return: On success, this syscall returns the maximum
4946 * rt_priority that can be used by a given scheduling class.
4947 * On failure, a negative error code is returned.
4949 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4956 ret = MAX_USER_RT_PRIO-1;
4958 case SCHED_DEADLINE:
4969 * sys_sched_get_priority_min - return minimum RT priority.
4970 * @policy: scheduling class.
4972 * Return: On success, this syscall returns the minimum
4973 * rt_priority that can be used by a given scheduling class.
4974 * On failure, a negative error code is returned.
4976 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
4985 case SCHED_DEADLINE:
4995 * sys_sched_rr_get_interval - return the default timeslice of a process.
4996 * @pid: pid of the process.
4997 * @interval: userspace pointer to the timeslice value.
4999 * this syscall writes the default timeslice value of a given process
5000 * into the user-space timespec buffer. A value of '0' means infinity.
5002 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5005 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5006 struct timespec __user *, interval)
5008 struct task_struct *p;
5009 unsigned int time_slice;
5010 unsigned long flags;
5020 p = find_process_by_pid(pid);
5024 retval = security_task_getscheduler(p);
5028 rq = task_rq_lock(p, &flags);
5030 if (p->sched_class->get_rr_interval)
5031 time_slice = p->sched_class->get_rr_interval(rq, p);
5032 task_rq_unlock(rq, p, &flags);
5035 jiffies_to_timespec(time_slice, &t);
5036 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5044 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5046 void sched_show_task(struct task_struct *p)
5048 unsigned long free = 0;
5050 unsigned long state = p->state;
5053 state = __ffs(state) + 1;
5054 printk(KERN_INFO "%-15.15s %c", p->comm,
5055 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5056 #if BITS_PER_LONG == 32
5057 if (state == TASK_RUNNING)
5058 printk(KERN_CONT " running ");
5060 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5062 if (state == TASK_RUNNING)
5063 printk(KERN_CONT " running task ");
5065 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5067 #ifdef CONFIG_DEBUG_STACK_USAGE
5068 free = stack_not_used(p);
5073 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5075 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5076 task_pid_nr(p), ppid,
5077 (unsigned long)task_thread_info(p)->flags);
5079 print_worker_info(KERN_INFO, p);
5080 show_stack(p, NULL);
5083 void show_state_filter(unsigned long state_filter)
5085 struct task_struct *g, *p;
5087 #if BITS_PER_LONG == 32
5089 " task PC stack pid father\n");
5092 " task PC stack pid father\n");
5095 for_each_process_thread(g, p) {
5097 * reset the NMI-timeout, listing all files on a slow
5098 * console might take a lot of time:
5099 * Also, reset softlockup watchdogs on all CPUs, because
5100 * another CPU might be blocked waiting for us to process
5103 touch_nmi_watchdog();
5104 touch_all_softlockup_watchdogs();
5105 if (!state_filter || (p->state & state_filter))
5109 #ifdef CONFIG_SCHED_DEBUG
5110 sysrq_sched_debug_show();
5114 * Only show locks if all tasks are dumped:
5117 debug_show_all_locks();
5120 void init_idle_bootup_task(struct task_struct *idle)
5122 idle->sched_class = &idle_sched_class;
5126 * init_idle - set up an idle thread for a given CPU
5127 * @idle: task in question
5128 * @cpu: cpu the idle task belongs to
5130 * NOTE: this function does not set the idle thread's NEED_RESCHED
5131 * flag, to make booting more robust.
5133 void init_idle(struct task_struct *idle, int cpu)
5135 struct rq *rq = cpu_rq(cpu);
5136 unsigned long flags;
5138 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5139 raw_spin_lock(&rq->lock);
5141 __sched_fork(0, idle);
5143 idle->state = TASK_RUNNING;
5144 idle->se.exec_start = sched_clock();
5148 * Its possible that init_idle() gets called multiple times on a task,
5149 * in that case do_set_cpus_allowed() will not do the right thing.
5151 * And since this is boot we can forgo the serialization.
5153 set_cpus_allowed_common(idle, cpumask_of(cpu));
5156 * We're having a chicken and egg problem, even though we are
5157 * holding rq->lock, the cpu isn't yet set to this cpu so the
5158 * lockdep check in task_group() will fail.
5160 * Similar case to sched_fork(). / Alternatively we could
5161 * use task_rq_lock() here and obtain the other rq->lock.
5166 __set_task_cpu(idle, cpu);
5169 rq->curr = rq->idle = idle;
5170 idle->on_rq = TASK_ON_RQ_QUEUED;
5174 raw_spin_unlock(&rq->lock);
5175 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5177 /* Set the preempt count _outside_ the spinlocks! */
5178 init_idle_preempt_count(idle, cpu);
5181 * The idle tasks have their own, simple scheduling class:
5183 idle->sched_class = &idle_sched_class;
5184 ftrace_graph_init_idle_task(idle, cpu);
5185 vtime_init_idle(idle, cpu);
5187 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5191 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5192 const struct cpumask *trial)
5194 int ret = 1, trial_cpus;
5195 struct dl_bw *cur_dl_b;
5196 unsigned long flags;
5198 if (!cpumask_weight(cur))
5201 rcu_read_lock_sched();
5202 cur_dl_b = dl_bw_of(cpumask_any(cur));
5203 trial_cpus = cpumask_weight(trial);
5205 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5206 if (cur_dl_b->bw != -1 &&
5207 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5209 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5210 rcu_read_unlock_sched();
5215 int task_can_attach(struct task_struct *p,
5216 const struct cpumask *cs_cpus_allowed)
5221 * Kthreads which disallow setaffinity shouldn't be moved
5222 * to a new cpuset; we don't want to change their cpu
5223 * affinity and isolating such threads by their set of
5224 * allowed nodes is unnecessary. Thus, cpusets are not
5225 * applicable for such threads. This prevents checking for
5226 * success of set_cpus_allowed_ptr() on all attached tasks
5227 * before cpus_allowed may be changed.
5229 if (p->flags & PF_NO_SETAFFINITY) {
5235 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5237 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5242 unsigned long flags;
5244 rcu_read_lock_sched();
5245 dl_b = dl_bw_of(dest_cpu);
5246 raw_spin_lock_irqsave(&dl_b->lock, flags);
5247 cpus = dl_bw_cpus(dest_cpu);
5248 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5253 * We reserve space for this task in the destination
5254 * root_domain, as we can't fail after this point.
5255 * We will free resources in the source root_domain
5256 * later on (see set_cpus_allowed_dl()).
5258 __dl_add(dl_b, p->dl.dl_bw);
5260 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5261 rcu_read_unlock_sched();
5271 #ifdef CONFIG_NUMA_BALANCING
5272 /* Migrate current task p to target_cpu */
5273 int migrate_task_to(struct task_struct *p, int target_cpu)
5275 struct migration_arg arg = { p, target_cpu };
5276 int curr_cpu = task_cpu(p);
5278 if (curr_cpu == target_cpu)
5281 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5284 /* TODO: This is not properly updating schedstats */
5286 trace_sched_move_numa(p, curr_cpu, target_cpu);
5287 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5291 * Requeue a task on a given node and accurately track the number of NUMA
5292 * tasks on the runqueues
5294 void sched_setnuma(struct task_struct *p, int nid)
5297 unsigned long flags;
5298 bool queued, running;
5300 rq = task_rq_lock(p, &flags);
5301 queued = task_on_rq_queued(p);
5302 running = task_current(rq, p);
5305 dequeue_task(rq, p, DEQUEUE_SAVE);
5307 put_prev_task(rq, p);
5309 p->numa_preferred_nid = nid;
5312 p->sched_class->set_curr_task(rq);
5314 enqueue_task(rq, p, ENQUEUE_RESTORE);
5315 task_rq_unlock(rq, p, &flags);
5317 #endif /* CONFIG_NUMA_BALANCING */
5319 #ifdef CONFIG_HOTPLUG_CPU
5321 * Ensures that the idle task is using init_mm right before its cpu goes
5324 void idle_task_exit(void)
5326 struct mm_struct *mm = current->active_mm;
5328 BUG_ON(cpu_online(smp_processor_id()));
5330 if (mm != &init_mm) {
5331 switch_mm(mm, &init_mm, current);
5332 finish_arch_post_lock_switch();
5338 * Since this CPU is going 'away' for a while, fold any nr_active delta
5339 * we might have. Assumes we're called after migrate_tasks() so that the
5340 * nr_active count is stable.
5342 * Also see the comment "Global load-average calculations".
5344 static void calc_load_migrate(struct rq *rq)
5346 long delta = calc_load_fold_active(rq);
5348 atomic_long_add(delta, &calc_load_tasks);
5351 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5355 static const struct sched_class fake_sched_class = {
5356 .put_prev_task = put_prev_task_fake,
5359 static struct task_struct fake_task = {
5361 * Avoid pull_{rt,dl}_task()
5363 .prio = MAX_PRIO + 1,
5364 .sched_class = &fake_sched_class,
5368 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5369 * try_to_wake_up()->select_task_rq().
5371 * Called with rq->lock held even though we'er in stop_machine() and
5372 * there's no concurrency possible, we hold the required locks anyway
5373 * because of lock validation efforts.
5375 static void migrate_tasks(struct rq *dead_rq)
5377 struct rq *rq = dead_rq;
5378 struct task_struct *next, *stop = rq->stop;
5382 * Fudge the rq selection such that the below task selection loop
5383 * doesn't get stuck on the currently eligible stop task.
5385 * We're currently inside stop_machine() and the rq is either stuck
5386 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5387 * either way we should never end up calling schedule() until we're
5393 * put_prev_task() and pick_next_task() sched
5394 * class method both need to have an up-to-date
5395 * value of rq->clock[_task]
5397 update_rq_clock(rq);
5401 * There's this thread running, bail when that's the only
5404 if (rq->nr_running == 1)
5408 * pick_next_task assumes pinned rq->lock.
5410 lockdep_pin_lock(&rq->lock);
5411 next = pick_next_task(rq, &fake_task);
5413 next->sched_class->put_prev_task(rq, next);
5416 * Rules for changing task_struct::cpus_allowed are holding
5417 * both pi_lock and rq->lock, such that holding either
5418 * stabilizes the mask.
5420 * Drop rq->lock is not quite as disastrous as it usually is
5421 * because !cpu_active at this point, which means load-balance
5422 * will not interfere. Also, stop-machine.
5424 lockdep_unpin_lock(&rq->lock);
5425 raw_spin_unlock(&rq->lock);
5426 raw_spin_lock(&next->pi_lock);
5427 raw_spin_lock(&rq->lock);
5430 * Since we're inside stop-machine, _nothing_ should have
5431 * changed the task, WARN if weird stuff happened, because in
5432 * that case the above rq->lock drop is a fail too.
5434 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5435 raw_spin_unlock(&next->pi_lock);
5439 /* Find suitable destination for @next, with force if needed. */
5440 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5442 rq = __migrate_task(rq, next, dest_cpu);
5443 if (rq != dead_rq) {
5444 raw_spin_unlock(&rq->lock);
5446 raw_spin_lock(&rq->lock);
5448 raw_spin_unlock(&next->pi_lock);
5453 #endif /* CONFIG_HOTPLUG_CPU */
5455 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5457 static struct ctl_table sd_ctl_dir[] = {
5459 .procname = "sched_domain",
5465 static struct ctl_table sd_ctl_root[] = {
5467 .procname = "kernel",
5469 .child = sd_ctl_dir,
5474 static struct ctl_table *sd_alloc_ctl_entry(int n)
5476 struct ctl_table *entry =
5477 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5482 static void sd_free_ctl_entry(struct ctl_table **tablep)
5484 struct ctl_table *entry;
5487 * In the intermediate directories, both the child directory and
5488 * procname are dynamically allocated and could fail but the mode
5489 * will always be set. In the lowest directory the names are
5490 * static strings and all have proc handlers.
5492 for (entry = *tablep; entry->mode; entry++) {
5494 sd_free_ctl_entry(&entry->child);
5495 if (entry->proc_handler == NULL)
5496 kfree(entry->procname);
5503 static int min_load_idx = 0;
5504 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5507 set_table_entry(struct ctl_table *entry,
5508 const char *procname, void *data, int maxlen,
5509 umode_t mode, proc_handler *proc_handler,
5512 entry->procname = procname;
5514 entry->maxlen = maxlen;
5516 entry->proc_handler = proc_handler;
5519 entry->extra1 = &min_load_idx;
5520 entry->extra2 = &max_load_idx;
5524 static struct ctl_table *
5525 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5527 struct ctl_table *table = sd_alloc_ctl_entry(5);
5532 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5533 sizeof(int), 0644, proc_dointvec_minmax, false);
5534 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5535 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5536 proc_doulongvec_minmax, false);
5537 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5538 sizeof(int), 0644, proc_dointvec_minmax, false);
5539 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5540 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5541 proc_doulongvec_minmax, false);
5546 static struct ctl_table *
5547 sd_alloc_ctl_group_table(struct sched_group *sg)
5549 struct ctl_table *table = sd_alloc_ctl_entry(2);
5554 table->procname = kstrdup("energy", GFP_KERNEL);
5556 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5561 static struct ctl_table *
5562 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5564 struct ctl_table *table;
5565 unsigned int nr_entries = 14;
5568 struct sched_group *sg = sd->groups;
5573 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5575 nr_entries += nr_sgs;
5578 table = sd_alloc_ctl_entry(nr_entries);
5583 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5584 sizeof(long), 0644, proc_doulongvec_minmax, false);
5585 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5586 sizeof(long), 0644, proc_doulongvec_minmax, false);
5587 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5588 sizeof(int), 0644, proc_dointvec_minmax, true);
5589 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5590 sizeof(int), 0644, proc_dointvec_minmax, true);
5591 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5592 sizeof(int), 0644, proc_dointvec_minmax, true);
5593 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5594 sizeof(int), 0644, proc_dointvec_minmax, true);
5595 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5596 sizeof(int), 0644, proc_dointvec_minmax, true);
5597 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5598 sizeof(int), 0644, proc_dointvec_minmax, false);
5599 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5600 sizeof(int), 0644, proc_dointvec_minmax, false);
5601 set_table_entry(&table[9], "cache_nice_tries",
5602 &sd->cache_nice_tries,
5603 sizeof(int), 0644, proc_dointvec_minmax, false);
5604 set_table_entry(&table[10], "flags", &sd->flags,
5605 sizeof(int), 0644, proc_dointvec_minmax, false);
5606 set_table_entry(&table[11], "max_newidle_lb_cost",
5607 &sd->max_newidle_lb_cost,
5608 sizeof(long), 0644, proc_doulongvec_minmax, false);
5609 set_table_entry(&table[12], "name", sd->name,
5610 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5614 struct ctl_table *entry = &table[13];
5617 snprintf(buf, 32, "group%d", i);
5618 entry->procname = kstrdup(buf, GFP_KERNEL);
5620 entry->child = sd_alloc_ctl_group_table(sg);
5621 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5623 /* &table[nr_entries-1] is terminator */
5628 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5630 struct ctl_table *entry, *table;
5631 struct sched_domain *sd;
5632 int domain_num = 0, i;
5635 for_each_domain(cpu, sd)
5637 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5642 for_each_domain(cpu, sd) {
5643 snprintf(buf, 32, "domain%d", i);
5644 entry->procname = kstrdup(buf, GFP_KERNEL);
5646 entry->child = sd_alloc_ctl_domain_table(sd);
5653 static struct ctl_table_header *sd_sysctl_header;
5654 static void register_sched_domain_sysctl(void)
5656 int i, cpu_num = num_possible_cpus();
5657 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5660 WARN_ON(sd_ctl_dir[0].child);
5661 sd_ctl_dir[0].child = entry;
5666 for_each_possible_cpu(i) {
5667 snprintf(buf, 32, "cpu%d", i);
5668 entry->procname = kstrdup(buf, GFP_KERNEL);
5670 entry->child = sd_alloc_ctl_cpu_table(i);
5674 WARN_ON(sd_sysctl_header);
5675 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5678 /* may be called multiple times per register */
5679 static void unregister_sched_domain_sysctl(void)
5681 unregister_sysctl_table(sd_sysctl_header);
5682 sd_sysctl_header = NULL;
5683 if (sd_ctl_dir[0].child)
5684 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5687 static void register_sched_domain_sysctl(void)
5690 static void unregister_sched_domain_sysctl(void)
5693 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5695 static void set_rq_online(struct rq *rq)
5698 const struct sched_class *class;
5700 cpumask_set_cpu(rq->cpu, rq->rd->online);
5703 for_each_class(class) {
5704 if (class->rq_online)
5705 class->rq_online(rq);
5710 static void set_rq_offline(struct rq *rq)
5713 const struct sched_class *class;
5715 for_each_class(class) {
5716 if (class->rq_offline)
5717 class->rq_offline(rq);
5720 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5726 * migration_call - callback that gets triggered when a CPU is added.
5727 * Here we can start up the necessary migration thread for the new CPU.
5730 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5732 int cpu = (long)hcpu;
5733 unsigned long flags;
5734 struct rq *rq = cpu_rq(cpu);
5736 switch (action & ~CPU_TASKS_FROZEN) {
5738 case CPU_UP_PREPARE:
5739 raw_spin_lock_irqsave(&rq->lock, flags);
5740 walt_set_window_start(rq);
5741 raw_spin_unlock_irqrestore(&rq->lock, flags);
5742 rq->calc_load_update = calc_load_update;
5746 /* Update our root-domain */
5747 raw_spin_lock_irqsave(&rq->lock, flags);
5749 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5753 raw_spin_unlock_irqrestore(&rq->lock, flags);
5756 #ifdef CONFIG_HOTPLUG_CPU
5758 sched_ttwu_pending();
5759 /* Update our root-domain */
5760 raw_spin_lock_irqsave(&rq->lock, flags);
5761 walt_migrate_sync_cpu(cpu);
5763 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5767 BUG_ON(rq->nr_running != 1); /* the migration thread */
5768 raw_spin_unlock_irqrestore(&rq->lock, flags);
5772 calc_load_migrate(rq);
5777 update_max_interval();
5783 * Register at high priority so that task migration (migrate_all_tasks)
5784 * happens before everything else. This has to be lower priority than
5785 * the notifier in the perf_event subsystem, though.
5787 static struct notifier_block migration_notifier = {
5788 .notifier_call = migration_call,
5789 .priority = CPU_PRI_MIGRATION,
5792 static void set_cpu_rq_start_time(void)
5794 int cpu = smp_processor_id();
5795 struct rq *rq = cpu_rq(cpu);
5796 rq->age_stamp = sched_clock_cpu(cpu);
5799 static int sched_cpu_active(struct notifier_block *nfb,
5800 unsigned long action, void *hcpu)
5802 int cpu = (long)hcpu;
5804 switch (action & ~CPU_TASKS_FROZEN) {
5806 set_cpu_rq_start_time();
5811 * At this point a starting CPU has marked itself as online via
5812 * set_cpu_online(). But it might not yet have marked itself
5813 * as active, which is essential from here on.
5815 set_cpu_active(cpu, true);
5816 stop_machine_unpark(cpu);
5819 case CPU_DOWN_FAILED:
5820 set_cpu_active(cpu, true);
5828 static int sched_cpu_inactive(struct notifier_block *nfb,
5829 unsigned long action, void *hcpu)
5831 switch (action & ~CPU_TASKS_FROZEN) {
5832 case CPU_DOWN_PREPARE:
5833 set_cpu_active((long)hcpu, false);
5840 static int __init migration_init(void)
5842 void *cpu = (void *)(long)smp_processor_id();
5845 /* Initialize migration for the boot CPU */
5846 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5847 BUG_ON(err == NOTIFY_BAD);
5848 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5849 register_cpu_notifier(&migration_notifier);
5851 /* Register cpu active notifiers */
5852 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5853 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5857 early_initcall(migration_init);
5859 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5861 #ifdef CONFIG_SCHED_DEBUG
5863 static __read_mostly int sched_debug_enabled;
5865 static int __init sched_debug_setup(char *str)
5867 sched_debug_enabled = 1;
5871 early_param("sched_debug", sched_debug_setup);
5873 static inline bool sched_debug(void)
5875 return sched_debug_enabled;
5878 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5879 struct cpumask *groupmask)
5881 struct sched_group *group = sd->groups;
5883 cpumask_clear(groupmask);
5885 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5887 if (!(sd->flags & SD_LOAD_BALANCE)) {
5888 printk("does not load-balance\n");
5892 printk(KERN_CONT "span %*pbl level %s\n",
5893 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5895 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5896 printk(KERN_ERR "ERROR: domain->span does not contain "
5899 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5900 printk(KERN_ERR "ERROR: domain->groups does not contain"
5904 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5908 printk(KERN_ERR "ERROR: group is NULL\n");
5912 if (!cpumask_weight(sched_group_cpus(group))) {
5913 printk(KERN_CONT "\n");
5914 printk(KERN_ERR "ERROR: empty group\n");
5918 if (!(sd->flags & SD_OVERLAP) &&
5919 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5920 printk(KERN_CONT "\n");
5921 printk(KERN_ERR "ERROR: repeated CPUs\n");
5925 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5927 printk(KERN_CONT " %*pbl",
5928 cpumask_pr_args(sched_group_cpus(group)));
5929 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5930 printk(KERN_CONT " (cpu_capacity = %lu)",
5931 group->sgc->capacity);
5934 group = group->next;
5935 } while (group != sd->groups);
5936 printk(KERN_CONT "\n");
5938 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5939 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5942 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5943 printk(KERN_ERR "ERROR: parent span is not a superset "
5944 "of domain->span\n");
5948 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5952 if (!sched_debug_enabled)
5956 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5960 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5963 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5971 #else /* !CONFIG_SCHED_DEBUG */
5972 # define sched_domain_debug(sd, cpu) do { } while (0)
5973 static inline bool sched_debug(void)
5977 #endif /* CONFIG_SCHED_DEBUG */
5979 static int sd_degenerate(struct sched_domain *sd)
5981 if (cpumask_weight(sched_domain_span(sd)) == 1) {
5982 if (sd->groups->sge)
5983 sd->flags &= ~SD_LOAD_BALANCE;
5988 /* Following flags need at least 2 groups */
5989 if (sd->flags & (SD_LOAD_BALANCE |
5990 SD_BALANCE_NEWIDLE |
5993 SD_SHARE_CPUCAPACITY |
5994 SD_ASYM_CPUCAPACITY |
5995 SD_SHARE_PKG_RESOURCES |
5996 SD_SHARE_POWERDOMAIN |
5997 SD_SHARE_CAP_STATES)) {
5998 if (sd->groups != sd->groups->next)
6002 /* Following flags don't use groups */
6003 if (sd->flags & (SD_WAKE_AFFINE))
6010 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6012 unsigned long cflags = sd->flags, pflags = parent->flags;
6014 if (sd_degenerate(parent))
6017 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6020 /* Flags needing groups don't count if only 1 group in parent */
6021 if (parent->groups == parent->groups->next) {
6022 pflags &= ~(SD_LOAD_BALANCE |
6023 SD_BALANCE_NEWIDLE |
6026 SD_ASYM_CPUCAPACITY |
6027 SD_SHARE_CPUCAPACITY |
6028 SD_SHARE_PKG_RESOURCES |
6030 SD_SHARE_POWERDOMAIN |
6031 SD_SHARE_CAP_STATES);
6032 if (parent->groups->sge) {
6033 parent->flags &= ~SD_LOAD_BALANCE;
6036 if (nr_node_ids == 1)
6037 pflags &= ~SD_SERIALIZE;
6039 if (~cflags & pflags)
6045 static void free_rootdomain(struct rcu_head *rcu)
6047 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6049 cpupri_cleanup(&rd->cpupri);
6050 cpudl_cleanup(&rd->cpudl);
6051 free_cpumask_var(rd->dlo_mask);
6052 free_cpumask_var(rd->rto_mask);
6053 free_cpumask_var(rd->online);
6054 free_cpumask_var(rd->span);
6058 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6060 struct root_domain *old_rd = NULL;
6061 unsigned long flags;
6063 raw_spin_lock_irqsave(&rq->lock, flags);
6068 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6071 cpumask_clear_cpu(rq->cpu, old_rd->span);
6074 * If we dont want to free the old_rd yet then
6075 * set old_rd to NULL to skip the freeing later
6078 if (!atomic_dec_and_test(&old_rd->refcount))
6082 atomic_inc(&rd->refcount);
6085 cpumask_set_cpu(rq->cpu, rd->span);
6086 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6089 raw_spin_unlock_irqrestore(&rq->lock, flags);
6092 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6095 static int init_rootdomain(struct root_domain *rd)
6097 memset(rd, 0, sizeof(*rd));
6099 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6101 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6103 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6105 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6108 #ifdef HAVE_RT_PUSH_IPI
6110 raw_spin_lock_init(&rd->rto_lock);
6111 init_irq_work(&rd->rto_push_work, rto_push_irq_work_func);
6114 init_dl_bw(&rd->dl_bw);
6115 if (cpudl_init(&rd->cpudl) != 0)
6118 if (cpupri_init(&rd->cpupri) != 0)
6121 init_max_cpu_capacity(&rd->max_cpu_capacity);
6123 rd->max_cap_orig_cpu = rd->min_cap_orig_cpu = -1;
6128 free_cpumask_var(rd->rto_mask);
6130 free_cpumask_var(rd->dlo_mask);
6132 free_cpumask_var(rd->online);
6134 free_cpumask_var(rd->span);
6140 * By default the system creates a single root-domain with all cpus as
6141 * members (mimicking the global state we have today).
6143 struct root_domain def_root_domain;
6145 static void init_defrootdomain(void)
6147 init_rootdomain(&def_root_domain);
6149 atomic_set(&def_root_domain.refcount, 1);
6152 static struct root_domain *alloc_rootdomain(void)
6154 struct root_domain *rd;
6156 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6160 if (init_rootdomain(rd) != 0) {
6168 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6170 struct sched_group *tmp, *first;
6179 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6184 } while (sg != first);
6187 static void free_sched_domain(struct rcu_head *rcu)
6189 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6192 * If its an overlapping domain it has private groups, iterate and
6195 if (sd->flags & SD_OVERLAP) {
6196 free_sched_groups(sd->groups, 1);
6197 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6198 kfree(sd->groups->sgc);
6204 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6206 call_rcu(&sd->rcu, free_sched_domain);
6209 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6211 for (; sd; sd = sd->parent)
6212 destroy_sched_domain(sd, cpu);
6216 * Keep a special pointer to the highest sched_domain that has
6217 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6218 * allows us to avoid some pointer chasing select_idle_sibling().
6220 * Also keep a unique ID per domain (we use the first cpu number in
6221 * the cpumask of the domain), this allows us to quickly tell if
6222 * two cpus are in the same cache domain, see cpus_share_cache().
6224 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6225 DEFINE_PER_CPU(int, sd_llc_size);
6226 DEFINE_PER_CPU(int, sd_llc_id);
6227 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6228 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6229 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6230 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6231 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6233 static void update_top_cache_domain(int cpu)
6235 struct sched_domain *sd;
6236 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6240 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6242 id = cpumask_first(sched_domain_span(sd));
6243 size = cpumask_weight(sched_domain_span(sd));
6244 busy_sd = sd->parent; /* sd_busy */
6246 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6248 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6249 per_cpu(sd_llc_size, cpu) = size;
6250 per_cpu(sd_llc_id, cpu) = id;
6252 sd = lowest_flag_domain(cpu, SD_NUMA);
6253 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6255 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6256 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6258 for_each_domain(cpu, sd) {
6259 if (sd->groups->sge)
6264 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6266 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6267 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6271 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6272 * hold the hotplug lock.
6275 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6277 struct rq *rq = cpu_rq(cpu);
6278 struct sched_domain *tmp;
6280 /* Remove the sched domains which do not contribute to scheduling. */
6281 for (tmp = sd; tmp; ) {
6282 struct sched_domain *parent = tmp->parent;
6286 if (sd_parent_degenerate(tmp, parent)) {
6287 tmp->parent = parent->parent;
6289 parent->parent->child = tmp;
6291 * Transfer SD_PREFER_SIBLING down in case of a
6292 * degenerate parent; the spans match for this
6293 * so the property transfers.
6295 if (parent->flags & SD_PREFER_SIBLING)
6296 tmp->flags |= SD_PREFER_SIBLING;
6297 destroy_sched_domain(parent, cpu);
6302 if (sd && sd_degenerate(sd)) {
6305 destroy_sched_domain(tmp, cpu);
6310 sched_domain_debug(sd, cpu);
6312 rq_attach_root(rq, rd);
6314 rcu_assign_pointer(rq->sd, sd);
6315 destroy_sched_domains(tmp, cpu);
6317 update_top_cache_domain(cpu);
6320 /* Setup the mask of cpus configured for isolated domains */
6321 static int __init isolated_cpu_setup(char *str)
6323 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6324 cpulist_parse(str, cpu_isolated_map);
6328 __setup("isolcpus=", isolated_cpu_setup);
6331 struct sched_domain ** __percpu sd;
6332 struct root_domain *rd;
6343 * Build an iteration mask that can exclude certain CPUs from the upwards
6346 * Only CPUs that can arrive at this group should be considered to continue
6349 * Asymmetric node setups can result in situations where the domain tree is of
6350 * unequal depth, make sure to skip domains that already cover the entire
6353 * In that case build_sched_domains() will have terminated the iteration early
6354 * and our sibling sd spans will be empty. Domains should always include the
6355 * cpu they're built on, so check that.
6358 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6360 const struct cpumask *sg_span = sched_group_cpus(sg);
6361 struct sd_data *sdd = sd->private;
6362 struct sched_domain *sibling;
6365 for_each_cpu(i, sg_span) {
6366 sibling = *per_cpu_ptr(sdd->sd, i);
6369 * Can happen in the asymmetric case, where these siblings are
6370 * unused. The mask will not be empty because those CPUs that
6371 * do have the top domain _should_ span the domain.
6373 if (!sibling->child)
6376 /* If we would not end up here, we can't continue from here */
6377 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6380 cpumask_set_cpu(i, sched_group_mask(sg));
6383 /* We must not have empty masks here */
6384 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6388 * Return the canonical balance cpu for this group, this is the first cpu
6389 * of this group that's also in the iteration mask.
6391 int group_balance_cpu(struct sched_group *sg)
6393 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6397 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6399 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6400 const struct cpumask *span = sched_domain_span(sd);
6401 struct cpumask *covered = sched_domains_tmpmask;
6402 struct sd_data *sdd = sd->private;
6403 struct sched_domain *sibling;
6406 cpumask_clear(covered);
6408 for_each_cpu(i, span) {
6409 struct cpumask *sg_span;
6411 if (cpumask_test_cpu(i, covered))
6414 sibling = *per_cpu_ptr(sdd->sd, i);
6416 /* See the comment near build_group_mask(). */
6417 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6420 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6421 GFP_KERNEL, cpu_to_node(cpu));
6426 sg_span = sched_group_cpus(sg);
6428 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6430 cpumask_set_cpu(i, sg_span);
6432 cpumask_or(covered, covered, sg_span);
6434 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6435 if (atomic_inc_return(&sg->sgc->ref) == 1)
6436 build_group_mask(sd, sg);
6439 * Initialize sgc->capacity such that even if we mess up the
6440 * domains and no possible iteration will get us here, we won't
6443 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6444 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6445 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6448 * Make sure the first group of this domain contains the
6449 * canonical balance cpu. Otherwise the sched_domain iteration
6450 * breaks. See update_sg_lb_stats().
6452 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6453 group_balance_cpu(sg) == cpu)
6463 sd->groups = groups;
6468 free_sched_groups(first, 0);
6473 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6475 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6476 struct sched_domain *child = sd->child;
6479 cpu = cpumask_first(sched_domain_span(child));
6482 *sg = *per_cpu_ptr(sdd->sg, cpu);
6483 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6484 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6491 * build_sched_groups will build a circular linked list of the groups
6492 * covered by the given span, and will set each group's ->cpumask correctly,
6493 * and ->cpu_capacity to 0.
6495 * Assumes the sched_domain tree is fully constructed
6498 build_sched_groups(struct sched_domain *sd, int cpu)
6500 struct sched_group *first = NULL, *last = NULL;
6501 struct sd_data *sdd = sd->private;
6502 const struct cpumask *span = sched_domain_span(sd);
6503 struct cpumask *covered;
6506 get_group(cpu, sdd, &sd->groups);
6507 atomic_inc(&sd->groups->ref);
6509 if (cpu != cpumask_first(span))
6512 lockdep_assert_held(&sched_domains_mutex);
6513 covered = sched_domains_tmpmask;
6515 cpumask_clear(covered);
6517 for_each_cpu(i, span) {
6518 struct sched_group *sg;
6521 if (cpumask_test_cpu(i, covered))
6524 group = get_group(i, sdd, &sg);
6525 cpumask_setall(sched_group_mask(sg));
6527 for_each_cpu(j, span) {
6528 if (get_group(j, sdd, NULL) != group)
6531 cpumask_set_cpu(j, covered);
6532 cpumask_set_cpu(j, sched_group_cpus(sg));
6547 * Initialize sched groups cpu_capacity.
6549 * cpu_capacity indicates the capacity of sched group, which is used while
6550 * distributing the load between different sched groups in a sched domain.
6551 * Typically cpu_capacity for all the groups in a sched domain will be same
6552 * unless there are asymmetries in the topology. If there are asymmetries,
6553 * group having more cpu_capacity will pickup more load compared to the
6554 * group having less cpu_capacity.
6556 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6558 struct sched_group *sg = sd->groups;
6563 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6565 } while (sg != sd->groups);
6567 if (cpu != group_balance_cpu(sg))
6570 update_group_capacity(sd, cpu);
6571 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6575 * Check that the per-cpu provided sd energy data is consistent for all cpus
6578 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6579 const struct cpumask *cpumask)
6581 const struct sched_group_energy * const sge = fn(cpu);
6582 struct cpumask mask;
6585 if (cpumask_weight(cpumask) <= 1)
6588 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6590 for_each_cpu(i, &mask) {
6591 const struct sched_group_energy * const e = fn(i);
6594 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6596 for (y = 0; y < (e->nr_idle_states); y++) {
6597 BUG_ON(e->idle_states[y].power !=
6598 sge->idle_states[y].power);
6601 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6603 for (y = 0; y < (e->nr_cap_states); y++) {
6604 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6605 BUG_ON(e->cap_states[y].power !=
6606 sge->cap_states[y].power);
6611 static void init_sched_energy(int cpu, struct sched_domain *sd,
6612 sched_domain_energy_f fn)
6614 if (!(fn && fn(cpu)))
6617 if (cpu != group_balance_cpu(sd->groups))
6620 if (sd->child && !sd->child->groups->sge) {
6621 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6622 #ifdef CONFIG_SCHED_DEBUG
6623 pr_err(" energy data on %s but not on %s domain\n",
6624 sd->name, sd->child->name);
6629 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6631 sd->groups->sge = fn(cpu);
6635 * Initializers for schedule domains
6636 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6639 static int default_relax_domain_level = -1;
6640 int sched_domain_level_max;
6642 static int __init setup_relax_domain_level(char *str)
6644 if (kstrtoint(str, 0, &default_relax_domain_level))
6645 pr_warn("Unable to set relax_domain_level\n");
6649 __setup("relax_domain_level=", setup_relax_domain_level);
6651 static void set_domain_attribute(struct sched_domain *sd,
6652 struct sched_domain_attr *attr)
6656 if (!attr || attr->relax_domain_level < 0) {
6657 if (default_relax_domain_level < 0)
6660 request = default_relax_domain_level;
6662 request = attr->relax_domain_level;
6663 if (request < sd->level) {
6664 /* turn off idle balance on this domain */
6665 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6667 /* turn on idle balance on this domain */
6668 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6672 static void __sdt_free(const struct cpumask *cpu_map);
6673 static int __sdt_alloc(const struct cpumask *cpu_map);
6675 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6676 const struct cpumask *cpu_map)
6680 if (!atomic_read(&d->rd->refcount))
6681 free_rootdomain(&d->rd->rcu); /* fall through */
6683 free_percpu(d->sd); /* fall through */
6685 __sdt_free(cpu_map); /* fall through */
6691 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6692 const struct cpumask *cpu_map)
6694 memset(d, 0, sizeof(*d));
6696 if (__sdt_alloc(cpu_map))
6697 return sa_sd_storage;
6698 d->sd = alloc_percpu(struct sched_domain *);
6700 return sa_sd_storage;
6701 d->rd = alloc_rootdomain();
6704 return sa_rootdomain;
6708 * NULL the sd_data elements we've used to build the sched_domain and
6709 * sched_group structure so that the subsequent __free_domain_allocs()
6710 * will not free the data we're using.
6712 static void claim_allocations(int cpu, struct sched_domain *sd)
6714 struct sd_data *sdd = sd->private;
6716 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6717 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6719 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6720 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6722 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6723 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6727 static int sched_domains_numa_levels;
6728 enum numa_topology_type sched_numa_topology_type;
6729 static int *sched_domains_numa_distance;
6730 int sched_max_numa_distance;
6731 static struct cpumask ***sched_domains_numa_masks;
6732 static int sched_domains_curr_level;
6736 * SD_flags allowed in topology descriptions.
6738 * These flags are purely descriptive of the topology and do not prescribe
6739 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6742 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6743 * SD_SHARE_PKG_RESOURCES - describes shared caches
6744 * SD_NUMA - describes NUMA topologies
6745 * SD_SHARE_POWERDOMAIN - describes shared power domain
6746 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6747 * SD_SHARE_CAP_STATES - describes shared capacity states
6749 * Odd one out, which beside describing the topology has a quirk also
6750 * prescribes the desired behaviour that goes along with it:
6753 * SD_ASYM_PACKING - describes SMT quirks
6755 #define TOPOLOGY_SD_FLAGS \
6756 (SD_SHARE_CPUCAPACITY | \
6757 SD_SHARE_PKG_RESOURCES | \
6760 SD_ASYM_CPUCAPACITY | \
6761 SD_SHARE_POWERDOMAIN | \
6762 SD_SHARE_CAP_STATES)
6764 static struct sched_domain *
6765 sd_init(struct sched_domain_topology_level *tl,
6766 struct sched_domain *child, int cpu)
6768 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6769 int sd_weight, sd_flags = 0;
6773 * Ugly hack to pass state to sd_numa_mask()...
6775 sched_domains_curr_level = tl->numa_level;
6778 sd_weight = cpumask_weight(tl->mask(cpu));
6781 sd_flags = (*tl->sd_flags)();
6782 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6783 "wrong sd_flags in topology description\n"))
6784 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6786 *sd = (struct sched_domain){
6787 .min_interval = sd_weight,
6788 .max_interval = 2*sd_weight,
6790 .imbalance_pct = 125,
6792 .cache_nice_tries = 0,
6799 .flags = 1*SD_LOAD_BALANCE
6800 | 1*SD_BALANCE_NEWIDLE
6805 | 0*SD_SHARE_CPUCAPACITY
6806 | 0*SD_SHARE_PKG_RESOURCES
6808 | 0*SD_PREFER_SIBLING
6813 .last_balance = jiffies,
6814 .balance_interval = sd_weight,
6816 .max_newidle_lb_cost = 0,
6817 .next_decay_max_lb_cost = jiffies,
6819 #ifdef CONFIG_SCHED_DEBUG
6825 * Convert topological properties into behaviour.
6828 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6829 struct sched_domain *t = sd;
6831 for_each_lower_domain(t)
6832 t->flags |= SD_BALANCE_WAKE;
6835 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6836 sd->flags |= SD_PREFER_SIBLING;
6837 sd->imbalance_pct = 110;
6838 sd->smt_gain = 1178; /* ~15% */
6840 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6841 sd->imbalance_pct = 117;
6842 sd->cache_nice_tries = 1;
6846 } else if (sd->flags & SD_NUMA) {
6847 sd->cache_nice_tries = 2;
6851 sd->flags |= SD_SERIALIZE;
6852 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6853 sd->flags &= ~(SD_BALANCE_EXEC |
6860 sd->flags |= SD_PREFER_SIBLING;
6861 sd->cache_nice_tries = 1;
6866 sd->private = &tl->data;
6872 * Topology list, bottom-up.
6874 static struct sched_domain_topology_level default_topology[] = {
6875 #ifdef CONFIG_SCHED_SMT
6876 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6878 #ifdef CONFIG_SCHED_MC
6879 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6881 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6885 static struct sched_domain_topology_level *sched_domain_topology =
6888 #define for_each_sd_topology(tl) \
6889 for (tl = sched_domain_topology; tl->mask; tl++)
6891 void set_sched_topology(struct sched_domain_topology_level *tl)
6893 sched_domain_topology = tl;
6898 static const struct cpumask *sd_numa_mask(int cpu)
6900 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6903 static void sched_numa_warn(const char *str)
6905 static int done = false;
6913 printk(KERN_WARNING "ERROR: %s\n\n", str);
6915 for (i = 0; i < nr_node_ids; i++) {
6916 printk(KERN_WARNING " ");
6917 for (j = 0; j < nr_node_ids; j++)
6918 printk(KERN_CONT "%02d ", node_distance(i,j));
6919 printk(KERN_CONT "\n");
6921 printk(KERN_WARNING "\n");
6924 bool find_numa_distance(int distance)
6928 if (distance == node_distance(0, 0))
6931 for (i = 0; i < sched_domains_numa_levels; i++) {
6932 if (sched_domains_numa_distance[i] == distance)
6940 * A system can have three types of NUMA topology:
6941 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6942 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6943 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6945 * The difference between a glueless mesh topology and a backplane
6946 * topology lies in whether communication between not directly
6947 * connected nodes goes through intermediary nodes (where programs
6948 * could run), or through backplane controllers. This affects
6949 * placement of programs.
6951 * The type of topology can be discerned with the following tests:
6952 * - If the maximum distance between any nodes is 1 hop, the system
6953 * is directly connected.
6954 * - If for two nodes A and B, located N > 1 hops away from each other,
6955 * there is an intermediary node C, which is < N hops away from both
6956 * nodes A and B, the system is a glueless mesh.
6958 static void init_numa_topology_type(void)
6962 n = sched_max_numa_distance;
6964 if (sched_domains_numa_levels <= 1) {
6965 sched_numa_topology_type = NUMA_DIRECT;
6969 for_each_online_node(a) {
6970 for_each_online_node(b) {
6971 /* Find two nodes furthest removed from each other. */
6972 if (node_distance(a, b) < n)
6975 /* Is there an intermediary node between a and b? */
6976 for_each_online_node(c) {
6977 if (node_distance(a, c) < n &&
6978 node_distance(b, c) < n) {
6979 sched_numa_topology_type =
6985 sched_numa_topology_type = NUMA_BACKPLANE;
6991 static void sched_init_numa(void)
6993 int next_distance, curr_distance = node_distance(0, 0);
6994 struct sched_domain_topology_level *tl;
6998 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6999 if (!sched_domains_numa_distance)
7003 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
7004 * unique distances in the node_distance() table.
7006 * Assumes node_distance(0,j) includes all distances in
7007 * node_distance(i,j) in order to avoid cubic time.
7009 next_distance = curr_distance;
7010 for (i = 0; i < nr_node_ids; i++) {
7011 for (j = 0; j < nr_node_ids; j++) {
7012 for (k = 0; k < nr_node_ids; k++) {
7013 int distance = node_distance(i, k);
7015 if (distance > curr_distance &&
7016 (distance < next_distance ||
7017 next_distance == curr_distance))
7018 next_distance = distance;
7021 * While not a strong assumption it would be nice to know
7022 * about cases where if node A is connected to B, B is not
7023 * equally connected to A.
7025 if (sched_debug() && node_distance(k, i) != distance)
7026 sched_numa_warn("Node-distance not symmetric");
7028 if (sched_debug() && i && !find_numa_distance(distance))
7029 sched_numa_warn("Node-0 not representative");
7031 if (next_distance != curr_distance) {
7032 sched_domains_numa_distance[level++] = next_distance;
7033 sched_domains_numa_levels = level;
7034 curr_distance = next_distance;
7039 * In case of sched_debug() we verify the above assumption.
7049 * 'level' contains the number of unique distances, excluding the
7050 * identity distance node_distance(i,i).
7052 * The sched_domains_numa_distance[] array includes the actual distance
7057 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7058 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7059 * the array will contain less then 'level' members. This could be
7060 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7061 * in other functions.
7063 * We reset it to 'level' at the end of this function.
7065 sched_domains_numa_levels = 0;
7067 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7068 if (!sched_domains_numa_masks)
7072 * Now for each level, construct a mask per node which contains all
7073 * cpus of nodes that are that many hops away from us.
7075 for (i = 0; i < level; i++) {
7076 sched_domains_numa_masks[i] =
7077 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7078 if (!sched_domains_numa_masks[i])
7081 for (j = 0; j < nr_node_ids; j++) {
7082 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7086 sched_domains_numa_masks[i][j] = mask;
7089 if (node_distance(j, k) > sched_domains_numa_distance[i])
7092 cpumask_or(mask, mask, cpumask_of_node(k));
7097 /* Compute default topology size */
7098 for (i = 0; sched_domain_topology[i].mask; i++);
7100 tl = kzalloc((i + level + 1) *
7101 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7106 * Copy the default topology bits..
7108 for (i = 0; sched_domain_topology[i].mask; i++)
7109 tl[i] = sched_domain_topology[i];
7112 * .. and append 'j' levels of NUMA goodness.
7114 for (j = 0; j < level; i++, j++) {
7115 tl[i] = (struct sched_domain_topology_level){
7116 .mask = sd_numa_mask,
7117 .sd_flags = cpu_numa_flags,
7118 .flags = SDTL_OVERLAP,
7124 sched_domain_topology = tl;
7126 sched_domains_numa_levels = level;
7127 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7129 init_numa_topology_type();
7132 static void sched_domains_numa_masks_set(int cpu)
7135 int node = cpu_to_node(cpu);
7137 for (i = 0; i < sched_domains_numa_levels; i++) {
7138 for (j = 0; j < nr_node_ids; j++) {
7139 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7140 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7145 static void sched_domains_numa_masks_clear(int cpu)
7148 for (i = 0; i < sched_domains_numa_levels; i++) {
7149 for (j = 0; j < nr_node_ids; j++)
7150 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7155 * Update sched_domains_numa_masks[level][node] array when new cpus
7158 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7159 unsigned long action,
7162 int cpu = (long)hcpu;
7164 switch (action & ~CPU_TASKS_FROZEN) {
7166 sched_domains_numa_masks_set(cpu);
7170 sched_domains_numa_masks_clear(cpu);
7180 static inline void sched_init_numa(void)
7184 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7185 unsigned long action,
7190 #endif /* CONFIG_NUMA */
7192 static int __sdt_alloc(const struct cpumask *cpu_map)
7194 struct sched_domain_topology_level *tl;
7197 for_each_sd_topology(tl) {
7198 struct sd_data *sdd = &tl->data;
7200 sdd->sd = alloc_percpu(struct sched_domain *);
7204 sdd->sg = alloc_percpu(struct sched_group *);
7208 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7212 for_each_cpu(j, cpu_map) {
7213 struct sched_domain *sd;
7214 struct sched_group *sg;
7215 struct sched_group_capacity *sgc;
7217 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7218 GFP_KERNEL, cpu_to_node(j));
7222 *per_cpu_ptr(sdd->sd, j) = sd;
7224 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7225 GFP_KERNEL, cpu_to_node(j));
7231 *per_cpu_ptr(sdd->sg, j) = sg;
7233 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7234 GFP_KERNEL, cpu_to_node(j));
7238 *per_cpu_ptr(sdd->sgc, j) = sgc;
7245 static void __sdt_free(const struct cpumask *cpu_map)
7247 struct sched_domain_topology_level *tl;
7250 for_each_sd_topology(tl) {
7251 struct sd_data *sdd = &tl->data;
7253 for_each_cpu(j, cpu_map) {
7254 struct sched_domain *sd;
7257 sd = *per_cpu_ptr(sdd->sd, j);
7258 if (sd && (sd->flags & SD_OVERLAP))
7259 free_sched_groups(sd->groups, 0);
7260 kfree(*per_cpu_ptr(sdd->sd, j));
7264 kfree(*per_cpu_ptr(sdd->sg, j));
7266 kfree(*per_cpu_ptr(sdd->sgc, j));
7268 free_percpu(sdd->sd);
7270 free_percpu(sdd->sg);
7272 free_percpu(sdd->sgc);
7277 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7278 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7279 struct sched_domain *child, int cpu)
7281 struct sched_domain *sd = sd_init(tl, child, cpu);
7283 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7285 sd->level = child->level + 1;
7286 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7289 if (!cpumask_subset(sched_domain_span(child),
7290 sched_domain_span(sd))) {
7291 pr_err("BUG: arch topology borken\n");
7292 #ifdef CONFIG_SCHED_DEBUG
7293 pr_err(" the %s domain not a subset of the %s domain\n",
7294 child->name, sd->name);
7296 /* Fixup, ensure @sd has at least @child cpus. */
7297 cpumask_or(sched_domain_span(sd),
7298 sched_domain_span(sd),
7299 sched_domain_span(child));
7303 set_domain_attribute(sd, attr);
7309 * Build sched domains for a given set of cpus and attach the sched domains
7310 * to the individual cpus
7312 static int build_sched_domains(const struct cpumask *cpu_map,
7313 struct sched_domain_attr *attr)
7315 enum s_alloc alloc_state;
7316 struct sched_domain *sd;
7318 int i, ret = -ENOMEM;
7320 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7321 if (alloc_state != sa_rootdomain)
7324 /* Set up domains for cpus specified by the cpu_map. */
7325 for_each_cpu(i, cpu_map) {
7326 struct sched_domain_topology_level *tl;
7329 for_each_sd_topology(tl) {
7330 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7331 if (tl == sched_domain_topology)
7332 *per_cpu_ptr(d.sd, i) = sd;
7333 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7334 sd->flags |= SD_OVERLAP;
7338 /* Build the groups for the domains */
7339 for_each_cpu(i, cpu_map) {
7340 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7341 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7342 if (sd->flags & SD_OVERLAP) {
7343 if (build_overlap_sched_groups(sd, i))
7346 if (build_sched_groups(sd, i))
7352 /* Calculate CPU capacity for physical packages and nodes */
7353 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7354 struct sched_domain_topology_level *tl = sched_domain_topology;
7356 if (!cpumask_test_cpu(i, cpu_map))
7359 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7360 init_sched_energy(i, sd, tl->energy);
7361 claim_allocations(i, sd);
7362 init_sched_groups_capacity(i, sd);
7366 /* Attach the domains */
7368 for_each_cpu(i, cpu_map) {
7369 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu);
7370 int min_cpu = READ_ONCE(d.rd->min_cap_orig_cpu);
7372 if ((max_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig >
7373 cpu_rq(max_cpu)->cpu_capacity_orig))
7374 WRITE_ONCE(d.rd->max_cap_orig_cpu, i);
7376 if ((min_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig <
7377 cpu_rq(min_cpu)->cpu_capacity_orig))
7378 WRITE_ONCE(d.rd->min_cap_orig_cpu, i);
7380 sd = *per_cpu_ptr(d.sd, i);
7382 cpu_attach_domain(sd, d.rd, i);
7388 __free_domain_allocs(&d, alloc_state, cpu_map);
7392 static cpumask_var_t *doms_cur; /* current sched domains */
7393 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7394 static struct sched_domain_attr *dattr_cur;
7395 /* attribues of custom domains in 'doms_cur' */
7398 * Special case: If a kmalloc of a doms_cur partition (array of
7399 * cpumask) fails, then fallback to a single sched domain,
7400 * as determined by the single cpumask fallback_doms.
7402 static cpumask_var_t fallback_doms;
7405 * arch_update_cpu_topology lets virtualized architectures update the
7406 * cpu core maps. It is supposed to return 1 if the topology changed
7407 * or 0 if it stayed the same.
7409 int __weak arch_update_cpu_topology(void)
7414 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7417 cpumask_var_t *doms;
7419 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7422 for (i = 0; i < ndoms; i++) {
7423 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7424 free_sched_domains(doms, i);
7431 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7434 for (i = 0; i < ndoms; i++)
7435 free_cpumask_var(doms[i]);
7440 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7441 * For now this just excludes isolated cpus, but could be used to
7442 * exclude other special cases in the future.
7444 static int init_sched_domains(const struct cpumask *cpu_map)
7448 arch_update_cpu_topology();
7450 doms_cur = alloc_sched_domains(ndoms_cur);
7452 doms_cur = &fallback_doms;
7453 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7454 err = build_sched_domains(doms_cur[0], NULL);
7455 register_sched_domain_sysctl();
7461 * Detach sched domains from a group of cpus specified in cpu_map
7462 * These cpus will now be attached to the NULL domain
7464 static void detach_destroy_domains(const struct cpumask *cpu_map)
7469 for_each_cpu(i, cpu_map)
7470 cpu_attach_domain(NULL, &def_root_domain, i);
7474 /* handle null as "default" */
7475 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7476 struct sched_domain_attr *new, int idx_new)
7478 struct sched_domain_attr tmp;
7485 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7486 new ? (new + idx_new) : &tmp,
7487 sizeof(struct sched_domain_attr));
7491 * Partition sched domains as specified by the 'ndoms_new'
7492 * cpumasks in the array doms_new[] of cpumasks. This compares
7493 * doms_new[] to the current sched domain partitioning, doms_cur[].
7494 * It destroys each deleted domain and builds each new domain.
7496 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7497 * The masks don't intersect (don't overlap.) We should setup one
7498 * sched domain for each mask. CPUs not in any of the cpumasks will
7499 * not be load balanced. If the same cpumask appears both in the
7500 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7503 * The passed in 'doms_new' should be allocated using
7504 * alloc_sched_domains. This routine takes ownership of it and will
7505 * free_sched_domains it when done with it. If the caller failed the
7506 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7507 * and partition_sched_domains() will fallback to the single partition
7508 * 'fallback_doms', it also forces the domains to be rebuilt.
7510 * If doms_new == NULL it will be replaced with cpu_online_mask.
7511 * ndoms_new == 0 is a special case for destroying existing domains,
7512 * and it will not create the default domain.
7514 * Call with hotplug lock held
7516 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7517 struct sched_domain_attr *dattr_new)
7522 mutex_lock(&sched_domains_mutex);
7524 /* always unregister in case we don't destroy any domains */
7525 unregister_sched_domain_sysctl();
7527 /* Let architecture update cpu core mappings. */
7528 new_topology = arch_update_cpu_topology();
7530 n = doms_new ? ndoms_new : 0;
7532 /* Destroy deleted domains */
7533 for (i = 0; i < ndoms_cur; i++) {
7534 for (j = 0; j < n && !new_topology; j++) {
7535 if (cpumask_equal(doms_cur[i], doms_new[j])
7536 && dattrs_equal(dattr_cur, i, dattr_new, j))
7539 /* no match - a current sched domain not in new doms_new[] */
7540 detach_destroy_domains(doms_cur[i]);
7546 if (doms_new == NULL) {
7548 doms_new = &fallback_doms;
7549 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7550 WARN_ON_ONCE(dattr_new);
7553 /* Build new domains */
7554 for (i = 0; i < ndoms_new; i++) {
7555 for (j = 0; j < n && !new_topology; j++) {
7556 if (cpumask_equal(doms_new[i], doms_cur[j])
7557 && dattrs_equal(dattr_new, i, dattr_cur, j))
7560 /* no match - add a new doms_new */
7561 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7566 /* Remember the new sched domains */
7567 if (doms_cur != &fallback_doms)
7568 free_sched_domains(doms_cur, ndoms_cur);
7569 kfree(dattr_cur); /* kfree(NULL) is safe */
7570 doms_cur = doms_new;
7571 dattr_cur = dattr_new;
7572 ndoms_cur = ndoms_new;
7574 register_sched_domain_sysctl();
7576 mutex_unlock(&sched_domains_mutex);
7579 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7582 * Update cpusets according to cpu_active mask. If cpusets are
7583 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7584 * around partition_sched_domains().
7586 * If we come here as part of a suspend/resume, don't touch cpusets because we
7587 * want to restore it back to its original state upon resume anyway.
7589 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7593 case CPU_ONLINE_FROZEN:
7594 case CPU_DOWN_FAILED_FROZEN:
7597 * num_cpus_frozen tracks how many CPUs are involved in suspend
7598 * resume sequence. As long as this is not the last online
7599 * operation in the resume sequence, just build a single sched
7600 * domain, ignoring cpusets.
7602 partition_sched_domains(1, NULL, NULL);
7603 if (--num_cpus_frozen)
7607 * This is the last CPU online operation. So fall through and
7608 * restore the original sched domains by considering the
7609 * cpuset configurations.
7611 cpuset_force_rebuild();
7614 cpuset_update_active_cpus(true);
7622 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7625 unsigned long flags;
7626 long cpu = (long)hcpu;
7632 case CPU_DOWN_PREPARE:
7633 rcu_read_lock_sched();
7634 dl_b = dl_bw_of(cpu);
7636 raw_spin_lock_irqsave(&dl_b->lock, flags);
7637 cpus = dl_bw_cpus(cpu);
7638 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7639 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7641 rcu_read_unlock_sched();
7644 return notifier_from_errno(-EBUSY);
7645 cpuset_update_active_cpus(false);
7647 case CPU_DOWN_PREPARE_FROZEN:
7649 partition_sched_domains(1, NULL, NULL);
7657 void __init sched_init_smp(void)
7659 cpumask_var_t non_isolated_cpus;
7661 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7662 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7667 * There's no userspace yet to cause hotplug operations; hence all the
7668 * cpu masks are stable and all blatant races in the below code cannot
7671 mutex_lock(&sched_domains_mutex);
7672 init_sched_domains(cpu_active_mask);
7673 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7674 if (cpumask_empty(non_isolated_cpus))
7675 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7676 mutex_unlock(&sched_domains_mutex);
7678 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7679 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7680 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7684 /* Move init over to a non-isolated CPU */
7685 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7687 sched_init_granularity();
7688 free_cpumask_var(non_isolated_cpus);
7690 init_sched_rt_class();
7691 init_sched_dl_class();
7694 void __init sched_init_smp(void)
7696 sched_init_granularity();
7698 #endif /* CONFIG_SMP */
7700 int in_sched_functions(unsigned long addr)
7702 return in_lock_functions(addr) ||
7703 (addr >= (unsigned long)__sched_text_start
7704 && addr < (unsigned long)__sched_text_end);
7707 #ifdef CONFIG_CGROUP_SCHED
7709 * Default task group.
7710 * Every task in system belongs to this group at bootup.
7712 struct task_group root_task_group;
7713 LIST_HEAD(task_groups);
7716 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7718 void __init sched_init(void)
7721 unsigned long alloc_size = 0, ptr;
7723 #ifdef CONFIG_FAIR_GROUP_SCHED
7724 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7726 #ifdef CONFIG_RT_GROUP_SCHED
7727 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7730 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7732 #ifdef CONFIG_FAIR_GROUP_SCHED
7733 root_task_group.se = (struct sched_entity **)ptr;
7734 ptr += nr_cpu_ids * sizeof(void **);
7736 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7737 ptr += nr_cpu_ids * sizeof(void **);
7739 #endif /* CONFIG_FAIR_GROUP_SCHED */
7740 #ifdef CONFIG_RT_GROUP_SCHED
7741 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7742 ptr += nr_cpu_ids * sizeof(void **);
7744 root_task_group.rt_rq = (struct rt_rq **)ptr;
7745 ptr += nr_cpu_ids * sizeof(void **);
7747 #endif /* CONFIG_RT_GROUP_SCHED */
7749 #ifdef CONFIG_CPUMASK_OFFSTACK
7750 for_each_possible_cpu(i) {
7751 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7752 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7754 #endif /* CONFIG_CPUMASK_OFFSTACK */
7756 init_rt_bandwidth(&def_rt_bandwidth,
7757 global_rt_period(), global_rt_runtime());
7758 init_dl_bandwidth(&def_dl_bandwidth,
7759 global_rt_period(), global_rt_runtime());
7762 init_defrootdomain();
7765 #ifdef CONFIG_RT_GROUP_SCHED
7766 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7767 global_rt_period(), global_rt_runtime());
7768 #endif /* CONFIG_RT_GROUP_SCHED */
7770 #ifdef CONFIG_CGROUP_SCHED
7771 list_add(&root_task_group.list, &task_groups);
7772 INIT_LIST_HEAD(&root_task_group.children);
7773 INIT_LIST_HEAD(&root_task_group.siblings);
7774 autogroup_init(&init_task);
7776 #endif /* CONFIG_CGROUP_SCHED */
7778 for_each_possible_cpu(i) {
7782 raw_spin_lock_init(&rq->lock);
7784 rq->calc_load_active = 0;
7785 rq->calc_load_update = jiffies + LOAD_FREQ;
7786 init_cfs_rq(&rq->cfs);
7787 init_rt_rq(&rq->rt);
7788 init_dl_rq(&rq->dl);
7789 #ifdef CONFIG_FAIR_GROUP_SCHED
7790 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7791 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7792 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7794 * How much cpu bandwidth does root_task_group get?
7796 * In case of task-groups formed thr' the cgroup filesystem, it
7797 * gets 100% of the cpu resources in the system. This overall
7798 * system cpu resource is divided among the tasks of
7799 * root_task_group and its child task-groups in a fair manner,
7800 * based on each entity's (task or task-group's) weight
7801 * (se->load.weight).
7803 * In other words, if root_task_group has 10 tasks of weight
7804 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7805 * then A0's share of the cpu resource is:
7807 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7809 * We achieve this by letting root_task_group's tasks sit
7810 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7812 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7813 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7814 #endif /* CONFIG_FAIR_GROUP_SCHED */
7816 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7817 #ifdef CONFIG_RT_GROUP_SCHED
7818 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7821 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7822 rq->cpu_load[j] = 0;
7824 rq->last_load_update_tick = jiffies;
7829 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7830 rq->balance_callback = NULL;
7831 rq->active_balance = 0;
7832 rq->next_balance = jiffies;
7834 rq->push_task = NULL;
7838 rq->avg_idle = 2*sysctl_sched_migration_cost;
7839 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7840 #ifdef CONFIG_SCHED_WALT
7841 rq->cur_irqload = 0;
7842 rq->avg_irqload = 0;
7846 INIT_LIST_HEAD(&rq->cfs_tasks);
7848 rq_attach_root(rq, &def_root_domain);
7849 #ifdef CONFIG_NO_HZ_COMMON
7852 #ifdef CONFIG_NO_HZ_FULL
7853 rq->last_sched_tick = 0;
7857 atomic_set(&rq->nr_iowait, 0);
7860 set_load_weight(&init_task);
7862 #ifdef CONFIG_PREEMPT_NOTIFIERS
7863 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7867 * The boot idle thread does lazy MMU switching as well:
7869 atomic_inc(&init_mm.mm_count);
7870 enter_lazy_tlb(&init_mm, current);
7873 * During early bootup we pretend to be a normal task:
7875 current->sched_class = &fair_sched_class;
7878 * Make us the idle thread. Technically, schedule() should not be
7879 * called from this thread, however somewhere below it might be,
7880 * but because we are the idle thread, we just pick up running again
7881 * when this runqueue becomes "idle".
7883 init_idle(current, smp_processor_id());
7885 calc_load_update = jiffies + LOAD_FREQ;
7888 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7889 /* May be allocated at isolcpus cmdline parse time */
7890 if (cpu_isolated_map == NULL)
7891 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7892 idle_thread_set_boot_cpu();
7893 set_cpu_rq_start_time();
7895 init_sched_fair_class();
7897 scheduler_running = 1;
7900 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7901 static inline int preempt_count_equals(int preempt_offset)
7903 int nested = preempt_count() + rcu_preempt_depth();
7905 return (nested == preempt_offset);
7908 static int __might_sleep_init_called;
7909 int __init __might_sleep_init(void)
7911 __might_sleep_init_called = 1;
7914 early_initcall(__might_sleep_init);
7916 void __might_sleep(const char *file, int line, int preempt_offset)
7919 * Blocking primitives will set (and therefore destroy) current->state,
7920 * since we will exit with TASK_RUNNING make sure we enter with it,
7921 * otherwise we will destroy state.
7923 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7924 "do not call blocking ops when !TASK_RUNNING; "
7925 "state=%lx set at [<%p>] %pS\n",
7927 (void *)current->task_state_change,
7928 (void *)current->task_state_change);
7930 ___might_sleep(file, line, preempt_offset);
7932 EXPORT_SYMBOL(__might_sleep);
7934 void ___might_sleep(const char *file, int line, int preempt_offset)
7936 static unsigned long prev_jiffy; /* ratelimiting */
7938 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7939 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7940 !is_idle_task(current)) || oops_in_progress)
7942 if (system_state != SYSTEM_RUNNING &&
7943 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7945 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7947 prev_jiffy = jiffies;
7950 "BUG: sleeping function called from invalid context at %s:%d\n",
7953 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7954 in_atomic(), irqs_disabled(),
7955 current->pid, current->comm);
7957 if (task_stack_end_corrupted(current))
7958 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7960 debug_show_held_locks(current);
7961 if (irqs_disabled())
7962 print_irqtrace_events(current);
7963 #ifdef CONFIG_DEBUG_PREEMPT
7964 if (!preempt_count_equals(preempt_offset)) {
7965 pr_err("Preemption disabled at:");
7966 print_ip_sym(current->preempt_disable_ip);
7972 EXPORT_SYMBOL(___might_sleep);
7975 #ifdef CONFIG_MAGIC_SYSRQ
7976 void normalize_rt_tasks(void)
7978 struct task_struct *g, *p;
7979 struct sched_attr attr = {
7980 .sched_policy = SCHED_NORMAL,
7983 read_lock(&tasklist_lock);
7984 for_each_process_thread(g, p) {
7986 * Only normalize user tasks:
7988 if (p->flags & PF_KTHREAD)
7991 p->se.exec_start = 0;
7992 #ifdef CONFIG_SCHEDSTATS
7993 p->se.statistics.wait_start = 0;
7994 p->se.statistics.sleep_start = 0;
7995 p->se.statistics.block_start = 0;
7998 if (!dl_task(p) && !rt_task(p)) {
8000 * Renice negative nice level userspace
8003 if (task_nice(p) < 0)
8004 set_user_nice(p, 0);
8008 __sched_setscheduler(p, &attr, false, false);
8010 read_unlock(&tasklist_lock);
8013 #endif /* CONFIG_MAGIC_SYSRQ */
8015 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8017 * These functions are only useful for the IA64 MCA handling, or kdb.
8019 * They can only be called when the whole system has been
8020 * stopped - every CPU needs to be quiescent, and no scheduling
8021 * activity can take place. Using them for anything else would
8022 * be a serious bug, and as a result, they aren't even visible
8023 * under any other configuration.
8027 * curr_task - return the current task for a given cpu.
8028 * @cpu: the processor in question.
8030 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8032 * Return: The current task for @cpu.
8034 struct task_struct *curr_task(int cpu)
8036 return cpu_curr(cpu);
8039 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8043 * set_curr_task - set the current task for a given cpu.
8044 * @cpu: the processor in question.
8045 * @p: the task pointer to set.
8047 * Description: This function must only be used when non-maskable interrupts
8048 * are serviced on a separate stack. It allows the architecture to switch the
8049 * notion of the current task on a cpu in a non-blocking manner. This function
8050 * must be called with all CPU's synchronized, and interrupts disabled, the
8051 * and caller must save the original value of the current task (see
8052 * curr_task() above) and restore that value before reenabling interrupts and
8053 * re-starting the system.
8055 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8057 void set_curr_task(int cpu, struct task_struct *p)
8064 #ifdef CONFIG_CGROUP_SCHED
8065 /* task_group_lock serializes the addition/removal of task groups */
8066 static DEFINE_SPINLOCK(task_group_lock);
8068 static void sched_free_group(struct task_group *tg)
8070 free_fair_sched_group(tg);
8071 free_rt_sched_group(tg);
8076 /* allocate runqueue etc for a new task group */
8077 struct task_group *sched_create_group(struct task_group *parent)
8079 struct task_group *tg;
8081 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8083 return ERR_PTR(-ENOMEM);
8085 if (!alloc_fair_sched_group(tg, parent))
8088 if (!alloc_rt_sched_group(tg, parent))
8094 sched_free_group(tg);
8095 return ERR_PTR(-ENOMEM);
8098 void sched_online_group(struct task_group *tg, struct task_group *parent)
8100 unsigned long flags;
8102 spin_lock_irqsave(&task_group_lock, flags);
8103 list_add_rcu(&tg->list, &task_groups);
8105 WARN_ON(!parent); /* root should already exist */
8107 tg->parent = parent;
8108 INIT_LIST_HEAD(&tg->children);
8109 list_add_rcu(&tg->siblings, &parent->children);
8110 spin_unlock_irqrestore(&task_group_lock, flags);
8113 /* rcu callback to free various structures associated with a task group */
8114 static void sched_free_group_rcu(struct rcu_head *rhp)
8116 /* now it should be safe to free those cfs_rqs */
8117 sched_free_group(container_of(rhp, struct task_group, rcu));
8120 void sched_destroy_group(struct task_group *tg)
8122 /* wait for possible concurrent references to cfs_rqs complete */
8123 call_rcu(&tg->rcu, sched_free_group_rcu);
8126 void sched_offline_group(struct task_group *tg)
8128 unsigned long flags;
8131 /* end participation in shares distribution */
8132 for_each_possible_cpu(i)
8133 unregister_fair_sched_group(tg, i);
8135 spin_lock_irqsave(&task_group_lock, flags);
8136 list_del_rcu(&tg->list);
8137 list_del_rcu(&tg->siblings);
8138 spin_unlock_irqrestore(&task_group_lock, flags);
8141 static void sched_change_group(struct task_struct *tsk, int type)
8143 struct task_group *tg;
8146 * All callers are synchronized by task_rq_lock(); we do not use RCU
8147 * which is pointless here. Thus, we pass "true" to task_css_check()
8148 * to prevent lockdep warnings.
8150 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8151 struct task_group, css);
8152 tg = autogroup_task_group(tsk, tg);
8153 tsk->sched_task_group = tg;
8155 #ifdef CONFIG_FAIR_GROUP_SCHED
8156 if (tsk->sched_class->task_change_group)
8157 tsk->sched_class->task_change_group(tsk, type);
8160 set_task_rq(tsk, task_cpu(tsk));
8164 * Change task's runqueue when it moves between groups.
8166 * The caller of this function should have put the task in its new group by
8167 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8170 void sched_move_task(struct task_struct *tsk)
8172 int queued, running;
8173 unsigned long flags;
8176 rq = task_rq_lock(tsk, &flags);
8178 running = task_current(rq, tsk);
8179 queued = task_on_rq_queued(tsk);
8182 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8183 if (unlikely(running))
8184 put_prev_task(rq, tsk);
8186 sched_change_group(tsk, TASK_MOVE_GROUP);
8188 if (unlikely(running))
8189 tsk->sched_class->set_curr_task(rq);
8191 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8193 task_rq_unlock(rq, tsk, &flags);
8195 #endif /* CONFIG_CGROUP_SCHED */
8197 #ifdef CONFIG_RT_GROUP_SCHED
8199 * Ensure that the real time constraints are schedulable.
8201 static DEFINE_MUTEX(rt_constraints_mutex);
8203 /* Must be called with tasklist_lock held */
8204 static inline int tg_has_rt_tasks(struct task_group *tg)
8206 struct task_struct *g, *p;
8209 * Autogroups do not have RT tasks; see autogroup_create().
8211 if (task_group_is_autogroup(tg))
8214 for_each_process_thread(g, p) {
8215 if (rt_task(p) && task_group(p) == tg)
8222 struct rt_schedulable_data {
8223 struct task_group *tg;
8228 static int tg_rt_schedulable(struct task_group *tg, void *data)
8230 struct rt_schedulable_data *d = data;
8231 struct task_group *child;
8232 unsigned long total, sum = 0;
8233 u64 period, runtime;
8235 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8236 runtime = tg->rt_bandwidth.rt_runtime;
8239 period = d->rt_period;
8240 runtime = d->rt_runtime;
8244 * Cannot have more runtime than the period.
8246 if (runtime > period && runtime != RUNTIME_INF)
8250 * Ensure we don't starve existing RT tasks.
8252 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8255 total = to_ratio(period, runtime);
8258 * Nobody can have more than the global setting allows.
8260 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8264 * The sum of our children's runtime should not exceed our own.
8266 list_for_each_entry_rcu(child, &tg->children, siblings) {
8267 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8268 runtime = child->rt_bandwidth.rt_runtime;
8270 if (child == d->tg) {
8271 period = d->rt_period;
8272 runtime = d->rt_runtime;
8275 sum += to_ratio(period, runtime);
8284 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8288 struct rt_schedulable_data data = {
8290 .rt_period = period,
8291 .rt_runtime = runtime,
8295 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8301 static int tg_set_rt_bandwidth(struct task_group *tg,
8302 u64 rt_period, u64 rt_runtime)
8307 * Disallowing the root group RT runtime is BAD, it would disallow the
8308 * kernel creating (and or operating) RT threads.
8310 if (tg == &root_task_group && rt_runtime == 0)
8313 /* No period doesn't make any sense. */
8317 mutex_lock(&rt_constraints_mutex);
8318 read_lock(&tasklist_lock);
8319 err = __rt_schedulable(tg, rt_period, rt_runtime);
8323 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8324 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8325 tg->rt_bandwidth.rt_runtime = rt_runtime;
8327 for_each_possible_cpu(i) {
8328 struct rt_rq *rt_rq = tg->rt_rq[i];
8330 raw_spin_lock(&rt_rq->rt_runtime_lock);
8331 rt_rq->rt_runtime = rt_runtime;
8332 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8334 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8336 read_unlock(&tasklist_lock);
8337 mutex_unlock(&rt_constraints_mutex);
8342 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8344 u64 rt_runtime, rt_period;
8346 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8347 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8348 if (rt_runtime_us < 0)
8349 rt_runtime = RUNTIME_INF;
8351 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8354 static long sched_group_rt_runtime(struct task_group *tg)
8358 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8361 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8362 do_div(rt_runtime_us, NSEC_PER_USEC);
8363 return rt_runtime_us;
8366 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8368 u64 rt_runtime, rt_period;
8370 rt_period = rt_period_us * NSEC_PER_USEC;
8371 rt_runtime = tg->rt_bandwidth.rt_runtime;
8373 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8376 static long sched_group_rt_period(struct task_group *tg)
8380 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8381 do_div(rt_period_us, NSEC_PER_USEC);
8382 return rt_period_us;
8384 #endif /* CONFIG_RT_GROUP_SCHED */
8386 #ifdef CONFIG_RT_GROUP_SCHED
8387 static int sched_rt_global_constraints(void)
8391 mutex_lock(&rt_constraints_mutex);
8392 read_lock(&tasklist_lock);
8393 ret = __rt_schedulable(NULL, 0, 0);
8394 read_unlock(&tasklist_lock);
8395 mutex_unlock(&rt_constraints_mutex);
8400 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8402 /* Don't accept realtime tasks when there is no way for them to run */
8403 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8409 #else /* !CONFIG_RT_GROUP_SCHED */
8410 static int sched_rt_global_constraints(void)
8412 unsigned long flags;
8415 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8416 for_each_possible_cpu(i) {
8417 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8419 raw_spin_lock(&rt_rq->rt_runtime_lock);
8420 rt_rq->rt_runtime = global_rt_runtime();
8421 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8423 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8427 #endif /* CONFIG_RT_GROUP_SCHED */
8429 static int sched_dl_global_validate(void)
8431 u64 runtime = global_rt_runtime();
8432 u64 period = global_rt_period();
8433 u64 new_bw = to_ratio(period, runtime);
8436 unsigned long flags;
8439 * Here we want to check the bandwidth not being set to some
8440 * value smaller than the currently allocated bandwidth in
8441 * any of the root_domains.
8443 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8444 * cycling on root_domains... Discussion on different/better
8445 * solutions is welcome!
8447 for_each_possible_cpu(cpu) {
8448 rcu_read_lock_sched();
8449 dl_b = dl_bw_of(cpu);
8451 raw_spin_lock_irqsave(&dl_b->lock, flags);
8452 if (new_bw < dl_b->total_bw)
8454 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8456 rcu_read_unlock_sched();
8465 static void sched_dl_do_global(void)
8470 unsigned long flags;
8472 def_dl_bandwidth.dl_period = global_rt_period();
8473 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8475 if (global_rt_runtime() != RUNTIME_INF)
8476 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8479 * FIXME: As above...
8481 for_each_possible_cpu(cpu) {
8482 rcu_read_lock_sched();
8483 dl_b = dl_bw_of(cpu);
8485 raw_spin_lock_irqsave(&dl_b->lock, flags);
8487 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8489 rcu_read_unlock_sched();
8493 static int sched_rt_global_validate(void)
8495 if (sysctl_sched_rt_period <= 0)
8498 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8499 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8505 static void sched_rt_do_global(void)
8507 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8508 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8511 int sched_rt_handler(struct ctl_table *table, int write,
8512 void __user *buffer, size_t *lenp,
8515 int old_period, old_runtime;
8516 static DEFINE_MUTEX(mutex);
8520 old_period = sysctl_sched_rt_period;
8521 old_runtime = sysctl_sched_rt_runtime;
8523 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8525 if (!ret && write) {
8526 ret = sched_rt_global_validate();
8530 ret = sched_dl_global_validate();
8534 ret = sched_rt_global_constraints();
8538 sched_rt_do_global();
8539 sched_dl_do_global();
8543 sysctl_sched_rt_period = old_period;
8544 sysctl_sched_rt_runtime = old_runtime;
8546 mutex_unlock(&mutex);
8551 int sched_rr_handler(struct ctl_table *table, int write,
8552 void __user *buffer, size_t *lenp,
8556 static DEFINE_MUTEX(mutex);
8559 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8560 /* make sure that internally we keep jiffies */
8561 /* also, writing zero resets timeslice to default */
8562 if (!ret && write) {
8563 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8564 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8566 mutex_unlock(&mutex);
8570 #ifdef CONFIG_CGROUP_SCHED
8572 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8574 return css ? container_of(css, struct task_group, css) : NULL;
8577 static struct cgroup_subsys_state *
8578 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8580 struct task_group *parent = css_tg(parent_css);
8581 struct task_group *tg;
8584 /* This is early initialization for the top cgroup */
8585 return &root_task_group.css;
8588 tg = sched_create_group(parent);
8590 return ERR_PTR(-ENOMEM);
8595 /* Expose task group only after completing cgroup initialization */
8596 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8598 struct task_group *tg = css_tg(css);
8599 struct task_group *parent = css_tg(css->parent);
8602 sched_online_group(tg, parent);
8606 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8608 struct task_group *tg = css_tg(css);
8610 sched_offline_group(tg);
8613 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8615 struct task_group *tg = css_tg(css);
8618 * Relies on the RCU grace period between css_released() and this.
8620 sched_free_group(tg);
8624 * This is called before wake_up_new_task(), therefore we really only
8625 * have to set its group bits, all the other stuff does not apply.
8627 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8629 unsigned long flags;
8632 rq = task_rq_lock(task, &flags);
8634 update_rq_clock(rq);
8635 sched_change_group(task, TASK_SET_GROUP);
8637 task_rq_unlock(rq, task, &flags);
8640 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8642 struct task_struct *task;
8643 struct cgroup_subsys_state *css;
8646 cgroup_taskset_for_each(task, css, tset) {
8647 #ifdef CONFIG_RT_GROUP_SCHED
8648 if (!sched_rt_can_attach(css_tg(css), task))
8651 /* We don't support RT-tasks being in separate groups */
8652 if (task->sched_class != &fair_sched_class)
8656 * Serialize against wake_up_new_task() such that if its
8657 * running, we're sure to observe its full state.
8659 raw_spin_lock_irq(&task->pi_lock);
8661 * Avoid calling sched_move_task() before wake_up_new_task()
8662 * has happened. This would lead to problems with PELT, due to
8663 * move wanting to detach+attach while we're not attached yet.
8665 if (task->state == TASK_NEW)
8667 raw_spin_unlock_irq(&task->pi_lock);
8675 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8677 struct task_struct *task;
8678 struct cgroup_subsys_state *css;
8680 cgroup_taskset_for_each(task, css, tset)
8681 sched_move_task(task);
8684 #ifdef CONFIG_FAIR_GROUP_SCHED
8685 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8686 struct cftype *cftype, u64 shareval)
8688 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8691 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8694 struct task_group *tg = css_tg(css);
8696 return (u64) scale_load_down(tg->shares);
8699 #ifdef CONFIG_CFS_BANDWIDTH
8700 static DEFINE_MUTEX(cfs_constraints_mutex);
8702 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8703 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8705 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8707 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8709 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8710 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8712 if (tg == &root_task_group)
8716 * Ensure we have at some amount of bandwidth every period. This is
8717 * to prevent reaching a state of large arrears when throttled via
8718 * entity_tick() resulting in prolonged exit starvation.
8720 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8724 * Likewise, bound things on the otherside by preventing insane quota
8725 * periods. This also allows us to normalize in computing quota
8728 if (period > max_cfs_quota_period)
8732 * Prevent race between setting of cfs_rq->runtime_enabled and
8733 * unthrottle_offline_cfs_rqs().
8736 mutex_lock(&cfs_constraints_mutex);
8737 ret = __cfs_schedulable(tg, period, quota);
8741 runtime_enabled = quota != RUNTIME_INF;
8742 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8744 * If we need to toggle cfs_bandwidth_used, off->on must occur
8745 * before making related changes, and on->off must occur afterwards
8747 if (runtime_enabled && !runtime_was_enabled)
8748 cfs_bandwidth_usage_inc();
8749 raw_spin_lock_irq(&cfs_b->lock);
8750 cfs_b->period = ns_to_ktime(period);
8751 cfs_b->quota = quota;
8753 __refill_cfs_bandwidth_runtime(cfs_b);
8754 /* restart the period timer (if active) to handle new period expiry */
8755 if (runtime_enabled)
8756 start_cfs_bandwidth(cfs_b);
8757 raw_spin_unlock_irq(&cfs_b->lock);
8759 for_each_online_cpu(i) {
8760 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8761 struct rq *rq = cfs_rq->rq;
8763 raw_spin_lock_irq(&rq->lock);
8764 cfs_rq->runtime_enabled = runtime_enabled;
8765 cfs_rq->runtime_remaining = 0;
8767 if (cfs_rq->throttled)
8768 unthrottle_cfs_rq(cfs_rq);
8769 raw_spin_unlock_irq(&rq->lock);
8771 if (runtime_was_enabled && !runtime_enabled)
8772 cfs_bandwidth_usage_dec();
8774 mutex_unlock(&cfs_constraints_mutex);
8780 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8784 period = ktime_to_ns(tg->cfs_bandwidth.period);
8785 if (cfs_quota_us < 0)
8786 quota = RUNTIME_INF;
8788 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8790 return tg_set_cfs_bandwidth(tg, period, quota);
8793 long tg_get_cfs_quota(struct task_group *tg)
8797 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8800 quota_us = tg->cfs_bandwidth.quota;
8801 do_div(quota_us, NSEC_PER_USEC);
8806 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8810 period = (u64)cfs_period_us * NSEC_PER_USEC;
8811 quota = tg->cfs_bandwidth.quota;
8813 return tg_set_cfs_bandwidth(tg, period, quota);
8816 long tg_get_cfs_period(struct task_group *tg)
8820 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8821 do_div(cfs_period_us, NSEC_PER_USEC);
8823 return cfs_period_us;
8826 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8829 return tg_get_cfs_quota(css_tg(css));
8832 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8833 struct cftype *cftype, s64 cfs_quota_us)
8835 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8838 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8841 return tg_get_cfs_period(css_tg(css));
8844 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8845 struct cftype *cftype, u64 cfs_period_us)
8847 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8850 struct cfs_schedulable_data {
8851 struct task_group *tg;
8856 * normalize group quota/period to be quota/max_period
8857 * note: units are usecs
8859 static u64 normalize_cfs_quota(struct task_group *tg,
8860 struct cfs_schedulable_data *d)
8868 period = tg_get_cfs_period(tg);
8869 quota = tg_get_cfs_quota(tg);
8872 /* note: these should typically be equivalent */
8873 if (quota == RUNTIME_INF || quota == -1)
8876 return to_ratio(period, quota);
8879 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8881 struct cfs_schedulable_data *d = data;
8882 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8883 s64 quota = 0, parent_quota = -1;
8886 quota = RUNTIME_INF;
8888 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8890 quota = normalize_cfs_quota(tg, d);
8891 parent_quota = parent_b->hierarchical_quota;
8894 * ensure max(child_quota) <= parent_quota, inherit when no
8897 if (quota == RUNTIME_INF)
8898 quota = parent_quota;
8899 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8902 cfs_b->hierarchical_quota = quota;
8907 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8910 struct cfs_schedulable_data data = {
8916 if (quota != RUNTIME_INF) {
8917 do_div(data.period, NSEC_PER_USEC);
8918 do_div(data.quota, NSEC_PER_USEC);
8922 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8928 static int cpu_stats_show(struct seq_file *sf, void *v)
8930 struct task_group *tg = css_tg(seq_css(sf));
8931 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8933 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8934 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8935 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8939 #endif /* CONFIG_CFS_BANDWIDTH */
8940 #endif /* CONFIG_FAIR_GROUP_SCHED */
8942 #ifdef CONFIG_RT_GROUP_SCHED
8943 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8944 struct cftype *cft, s64 val)
8946 return sched_group_set_rt_runtime(css_tg(css), val);
8949 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8952 return sched_group_rt_runtime(css_tg(css));
8955 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8956 struct cftype *cftype, u64 rt_period_us)
8958 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8961 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8964 return sched_group_rt_period(css_tg(css));
8966 #endif /* CONFIG_RT_GROUP_SCHED */
8968 static struct cftype cpu_files[] = {
8969 #ifdef CONFIG_FAIR_GROUP_SCHED
8972 .read_u64 = cpu_shares_read_u64,
8973 .write_u64 = cpu_shares_write_u64,
8976 #ifdef CONFIG_CFS_BANDWIDTH
8978 .name = "cfs_quota_us",
8979 .read_s64 = cpu_cfs_quota_read_s64,
8980 .write_s64 = cpu_cfs_quota_write_s64,
8983 .name = "cfs_period_us",
8984 .read_u64 = cpu_cfs_period_read_u64,
8985 .write_u64 = cpu_cfs_period_write_u64,
8989 .seq_show = cpu_stats_show,
8992 #ifdef CONFIG_RT_GROUP_SCHED
8994 .name = "rt_runtime_us",
8995 .read_s64 = cpu_rt_runtime_read,
8996 .write_s64 = cpu_rt_runtime_write,
8999 .name = "rt_period_us",
9000 .read_u64 = cpu_rt_period_read_uint,
9001 .write_u64 = cpu_rt_period_write_uint,
9007 struct cgroup_subsys cpu_cgrp_subsys = {
9008 .css_alloc = cpu_cgroup_css_alloc,
9009 .css_online = cpu_cgroup_css_online,
9010 .css_released = cpu_cgroup_css_released,
9011 .css_free = cpu_cgroup_css_free,
9012 .fork = cpu_cgroup_fork,
9013 .can_attach = cpu_cgroup_can_attach,
9014 .attach = cpu_cgroup_attach,
9015 .legacy_cftypes = cpu_files,
9019 #endif /* CONFIG_CGROUP_SCHED */
9021 void dump_cpu_task(int cpu)
9023 pr_info("Task dump for CPU %d:\n", cpu);
9024 sched_show_task(cpu_curr(cpu));