4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/unistd.h>
66 #include <linux/pagemap.h>
67 #include <linux/hrtimer.h>
68 #include <linux/tick.h>
69 #include <linux/debugfs.h>
70 #include <linux/ctype.h>
71 #include <linux/ftrace.h>
72 #include <linux/slab.h>
73 #include <linux/init_task.h>
74 #include <linux/binfmts.h>
75 #include <linux/context_tracking.h>
76 #include <linux/compiler.h>
78 #include <asm/switch_to.h>
80 #include <asm/irq_regs.h>
81 #include <asm/mutex.h>
82 #ifdef CONFIG_PARAVIRT
83 #include <asm/paravirt.h>
87 #include "../workqueue_internal.h"
88 #include "../smpboot.h"
90 #define CREATE_TRACE_POINTS
91 #include <trace/events/sched.h>
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))
549 get_task_struct(task);
552 * The head is context local, there can be no concurrency.
555 head->lastp = &node->next;
558 void wake_up_q(struct wake_q_head *head)
560 struct wake_q_node *node = head->first;
562 while (node != WAKE_Q_TAIL) {
563 struct task_struct *task;
565 task = container_of(node, struct task_struct, wake_q);
567 /* task can safely be re-inserted now */
569 task->wake_q.next = NULL;
572 * wake_up_process() implies a wmb() to pair with the queueing
573 * in wake_q_add() so as not to miss wakeups.
575 wake_up_process(task);
576 put_task_struct(task);
581 * resched_curr - mark rq's current task 'to be rescheduled now'.
583 * On UP this means the setting of the need_resched flag, on SMP it
584 * might also involve a cross-CPU call to trigger the scheduler on
587 void resched_curr(struct rq *rq)
589 struct task_struct *curr = rq->curr;
592 lockdep_assert_held(&rq->lock);
594 if (test_tsk_need_resched(curr))
599 if (cpu == smp_processor_id()) {
600 set_tsk_need_resched(curr);
601 set_preempt_need_resched();
605 if (set_nr_and_not_polling(curr))
606 smp_send_reschedule(cpu);
608 trace_sched_wake_idle_without_ipi(cpu);
611 void resched_cpu(int cpu)
613 struct rq *rq = cpu_rq(cpu);
616 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
619 raw_spin_unlock_irqrestore(&rq->lock, flags);
623 #ifdef CONFIG_NO_HZ_COMMON
625 * In the semi idle case, use the nearest busy cpu for migrating timers
626 * from an idle cpu. This is good for power-savings.
628 * We don't do similar optimization for completely idle system, as
629 * selecting an idle cpu will add more delays to the timers than intended
630 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
632 int get_nohz_timer_target(void)
634 int i, cpu = smp_processor_id();
635 struct sched_domain *sd;
637 if (!idle_cpu(cpu) && is_housekeeping_cpu(cpu))
641 for_each_domain(cpu, sd) {
642 for_each_cpu(i, sched_domain_span(sd)) {
646 if (!idle_cpu(i) && is_housekeeping_cpu(i)) {
653 if (!is_housekeeping_cpu(cpu))
654 cpu = housekeeping_any_cpu();
660 * When add_timer_on() enqueues a timer into the timer wheel of an
661 * idle CPU then this timer might expire before the next timer event
662 * which is scheduled to wake up that CPU. In case of a completely
663 * idle system the next event might even be infinite time into the
664 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
665 * leaves the inner idle loop so the newly added timer is taken into
666 * account when the CPU goes back to idle and evaluates the timer
667 * wheel for the next timer event.
669 static void wake_up_idle_cpu(int cpu)
671 struct rq *rq = cpu_rq(cpu);
673 if (cpu == smp_processor_id())
676 if (set_nr_and_not_polling(rq->idle))
677 smp_send_reschedule(cpu);
679 trace_sched_wake_idle_without_ipi(cpu);
682 static bool wake_up_full_nohz_cpu(int cpu)
685 * We just need the target to call irq_exit() and re-evaluate
686 * the next tick. The nohz full kick at least implies that.
687 * If needed we can still optimize that later with an
690 if (tick_nohz_full_cpu(cpu)) {
691 if (cpu != smp_processor_id() ||
692 tick_nohz_tick_stopped())
693 tick_nohz_full_kick_cpu(cpu);
700 void wake_up_nohz_cpu(int cpu)
702 if (!wake_up_full_nohz_cpu(cpu))
703 wake_up_idle_cpu(cpu);
706 static inline bool got_nohz_idle_kick(void)
708 int cpu = smp_processor_id();
710 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)))
713 if (idle_cpu(cpu) && !need_resched())
717 * We can't run Idle Load Balance on this CPU for this time so we
718 * cancel it and clear NOHZ_BALANCE_KICK
720 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu));
724 #else /* CONFIG_NO_HZ_COMMON */
726 static inline bool got_nohz_idle_kick(void)
731 #endif /* CONFIG_NO_HZ_COMMON */
733 #ifdef CONFIG_NO_HZ_FULL
734 bool sched_can_stop_tick(void)
737 * FIFO realtime policy runs the highest priority task. Other runnable
738 * tasks are of a lower priority. The scheduler tick does nothing.
740 if (current->policy == SCHED_FIFO)
744 * Round-robin realtime tasks time slice with other tasks at the same
745 * realtime priority. Is this task the only one at this priority?
747 if (current->policy == SCHED_RR) {
748 struct sched_rt_entity *rt_se = ¤t->rt;
750 return rt_se->run_list.prev == rt_se->run_list.next;
754 * More than one running task need preemption.
755 * nr_running update is assumed to be visible
756 * after IPI is sent from wakers.
758 if (this_rq()->nr_running > 1)
763 #endif /* CONFIG_NO_HZ_FULL */
765 void sched_avg_update(struct rq *rq)
767 s64 period = sched_avg_period();
769 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) {
771 * Inline assembly required to prevent the compiler
772 * optimising this loop into a divmod call.
773 * See __iter_div_u64_rem() for another example of this.
775 asm("" : "+rm" (rq->age_stamp));
776 rq->age_stamp += period;
781 #endif /* CONFIG_SMP */
783 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
784 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
786 * Iterate task_group tree rooted at *from, calling @down when first entering a
787 * node and @up when leaving it for the final time.
789 * Caller must hold rcu_lock or sufficient equivalent.
791 int walk_tg_tree_from(struct task_group *from,
792 tg_visitor down, tg_visitor up, void *data)
794 struct task_group *parent, *child;
800 ret = (*down)(parent, data);
803 list_for_each_entry_rcu(child, &parent->children, siblings) {
810 ret = (*up)(parent, data);
811 if (ret || parent == from)
815 parent = parent->parent;
822 int tg_nop(struct task_group *tg, void *data)
828 static void set_load_weight(struct task_struct *p)
830 int prio = p->static_prio - MAX_RT_PRIO;
831 struct load_weight *load = &p->se.load;
834 * SCHED_IDLE tasks get minimal weight:
836 if (idle_policy(p->policy)) {
837 load->weight = scale_load(WEIGHT_IDLEPRIO);
838 load->inv_weight = WMULT_IDLEPRIO;
842 load->weight = scale_load(prio_to_weight[prio]);
843 load->inv_weight = prio_to_wmult[prio];
846 static inline void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
849 if (!(flags & ENQUEUE_RESTORE))
850 sched_info_queued(rq, p);
851 p->sched_class->enqueue_task(rq, p, flags);
854 static inline void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
857 if (!(flags & DEQUEUE_SAVE))
858 sched_info_dequeued(rq, p);
859 p->sched_class->dequeue_task(rq, p, flags);
862 void activate_task(struct rq *rq, struct task_struct *p, int flags)
864 if (task_contributes_to_load(p))
865 rq->nr_uninterruptible--;
867 enqueue_task(rq, p, flags);
870 void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
872 if (task_contributes_to_load(p))
873 rq->nr_uninterruptible++;
875 dequeue_task(rq, p, flags);
878 static void update_rq_clock_task(struct rq *rq, s64 delta)
881 * In theory, the compile should just see 0 here, and optimize out the call
882 * to sched_rt_avg_update. But I don't trust it...
884 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
885 s64 steal = 0, irq_delta = 0;
887 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
888 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
891 * Since irq_time is only updated on {soft,}irq_exit, we might run into
892 * this case when a previous update_rq_clock() happened inside a
895 * When this happens, we stop ->clock_task and only update the
896 * prev_irq_time stamp to account for the part that fit, so that a next
897 * update will consume the rest. This ensures ->clock_task is
900 * It does however cause some slight miss-attribution of {soft,}irq
901 * time, a more accurate solution would be to update the irq_time using
902 * the current rq->clock timestamp, except that would require using
905 if (irq_delta > delta)
908 rq->prev_irq_time += irq_delta;
911 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
912 if (static_key_false((¶virt_steal_rq_enabled))) {
913 steal = paravirt_steal_clock(cpu_of(rq));
914 steal -= rq->prev_steal_time_rq;
916 if (unlikely(steal > delta))
919 rq->prev_steal_time_rq += steal;
924 rq->clock_task += delta;
926 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
927 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY))
928 sched_rt_avg_update(rq, irq_delta + steal);
932 void sched_set_stop_task(int cpu, struct task_struct *stop)
934 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
935 struct task_struct *old_stop = cpu_rq(cpu)->stop;
939 * Make it appear like a SCHED_FIFO task, its something
940 * userspace knows about and won't get confused about.
942 * Also, it will make PI more or less work without too
943 * much confusion -- but then, stop work should not
944 * rely on PI working anyway.
946 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
948 stop->sched_class = &stop_sched_class;
951 cpu_rq(cpu)->stop = stop;
955 * Reset it back to a normal scheduling class so that
956 * it can die in pieces.
958 old_stop->sched_class = &rt_sched_class;
963 * __normal_prio - return the priority that is based on the static prio
965 static inline int __normal_prio(struct task_struct *p)
967 return p->static_prio;
971 * Calculate the expected normal priority: i.e. priority
972 * without taking RT-inheritance into account. Might be
973 * boosted by interactivity modifiers. Changes upon fork,
974 * setprio syscalls, and whenever the interactivity
975 * estimator recalculates.
977 static inline int normal_prio(struct task_struct *p)
981 if (task_has_dl_policy(p))
982 prio = MAX_DL_PRIO-1;
983 else if (task_has_rt_policy(p))
984 prio = MAX_RT_PRIO-1 - p->rt_priority;
986 prio = __normal_prio(p);
991 * Calculate the current priority, i.e. the priority
992 * taken into account by the scheduler. This value might
993 * be boosted by RT tasks, or might be boosted by
994 * interactivity modifiers. Will be RT if the task got
995 * RT-boosted. If not then it returns p->normal_prio.
997 static int effective_prio(struct task_struct *p)
999 p->normal_prio = normal_prio(p);
1001 * If we are RT tasks or we were boosted to RT priority,
1002 * keep the priority unchanged. Otherwise, update priority
1003 * to the normal priority:
1005 if (!rt_prio(p->prio))
1006 return p->normal_prio;
1011 * task_curr - is this task currently executing on a CPU?
1012 * @p: the task in question.
1014 * Return: 1 if the task is currently executing. 0 otherwise.
1016 inline int task_curr(const struct task_struct *p)
1018 return cpu_curr(task_cpu(p)) == p;
1022 * switched_from, switched_to and prio_changed must _NOT_ drop rq->lock,
1023 * use the balance_callback list if you want balancing.
1025 * this means any call to check_class_changed() must be followed by a call to
1026 * balance_callback().
1028 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1029 const struct sched_class *prev_class,
1032 if (prev_class != p->sched_class) {
1033 if (prev_class->switched_from)
1034 prev_class->switched_from(rq, p);
1036 p->sched_class->switched_to(rq, p);
1037 } else if (oldprio != p->prio || dl_task(p))
1038 p->sched_class->prio_changed(rq, p, oldprio);
1041 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
1043 const struct sched_class *class;
1045 if (p->sched_class == rq->curr->sched_class) {
1046 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
1048 for_each_class(class) {
1049 if (class == rq->curr->sched_class)
1051 if (class == p->sched_class) {
1059 * A queue event has occurred, and we're going to schedule. In
1060 * this case, we can save a useless back to back clock update.
1062 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr))
1063 rq_clock_skip_update(rq, true);
1068 * This is how migration works:
1070 * 1) we invoke migration_cpu_stop() on the target CPU using
1072 * 2) stopper starts to run (implicitly forcing the migrated thread
1074 * 3) it checks whether the migrated task is still in the wrong runqueue.
1075 * 4) if it's in the wrong runqueue then the migration thread removes
1076 * it and puts it into the right queue.
1077 * 5) stopper completes and stop_one_cpu() returns and the migration
1082 * move_queued_task - move a queued task to new rq.
1084 * Returns (locked) new rq. Old rq's lock is released.
1086 static struct rq *move_queued_task(struct rq *rq, struct task_struct *p, int new_cpu)
1088 lockdep_assert_held(&rq->lock);
1090 dequeue_task(rq, p, 0);
1091 p->on_rq = TASK_ON_RQ_MIGRATING;
1092 double_lock_balance(rq, cpu_rq(new_cpu));
1093 set_task_cpu(p, new_cpu);
1094 double_unlock_balance(rq, cpu_rq(new_cpu));
1095 raw_spin_unlock(&rq->lock);
1097 rq = cpu_rq(new_cpu);
1099 raw_spin_lock(&rq->lock);
1100 BUG_ON(task_cpu(p) != new_cpu);
1101 p->on_rq = TASK_ON_RQ_QUEUED;
1102 enqueue_task(rq, p, 0);
1103 check_preempt_curr(rq, p, 0);
1108 struct migration_arg {
1109 struct task_struct *task;
1114 * Move (not current) task off this cpu, onto dest cpu. We're doing
1115 * this because either it can't run here any more (set_cpus_allowed()
1116 * away from this CPU, or CPU going down), or because we're
1117 * attempting to rebalance this task on exec (sched_exec).
1119 * So we race with normal scheduler movements, but that's OK, as long
1120 * as the task is no longer on this CPU.
1122 static struct rq *__migrate_task(struct rq *rq, struct task_struct *p, int dest_cpu)
1124 if (unlikely(!cpu_active(dest_cpu)))
1127 /* Affinity changed (again). */
1128 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1131 rq = move_queued_task(rq, p, dest_cpu);
1137 * migration_cpu_stop - this will be executed by a highprio stopper thread
1138 * and performs thread migration by bumping thread off CPU then
1139 * 'pushing' onto another runqueue.
1141 static int migration_cpu_stop(void *data)
1143 struct migration_arg *arg = data;
1144 struct task_struct *p = arg->task;
1145 struct rq *rq = this_rq();
1148 * The original target cpu might have gone down and we might
1149 * be on another cpu but it doesn't matter.
1151 local_irq_disable();
1153 * We need to explicitly wake pending tasks before running
1154 * __migrate_task() such that we will not miss enforcing cpus_allowed
1155 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test.
1157 sched_ttwu_pending();
1159 raw_spin_lock(&p->pi_lock);
1160 raw_spin_lock(&rq->lock);
1162 * If task_rq(p) != rq, it cannot be migrated here, because we're
1163 * holding rq->lock, if p->on_rq == 0 it cannot get enqueued because
1164 * we're holding p->pi_lock.
1166 if (task_rq(p) == rq && task_on_rq_queued(p))
1167 rq = __migrate_task(rq, p, arg->dest_cpu);
1168 raw_spin_unlock(&rq->lock);
1169 raw_spin_unlock(&p->pi_lock);
1176 * sched_class::set_cpus_allowed must do the below, but is not required to
1177 * actually call this function.
1179 void set_cpus_allowed_common(struct task_struct *p, const struct cpumask *new_mask)
1181 cpumask_copy(&p->cpus_allowed, new_mask);
1182 p->nr_cpus_allowed = cpumask_weight(new_mask);
1185 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
1187 struct rq *rq = task_rq(p);
1188 bool queued, running;
1190 lockdep_assert_held(&p->pi_lock);
1192 queued = task_on_rq_queued(p);
1193 running = task_current(rq, p);
1197 * Because __kthread_bind() calls this on blocked tasks without
1200 lockdep_assert_held(&rq->lock);
1201 dequeue_task(rq, p, DEQUEUE_SAVE);
1204 put_prev_task(rq, p);
1206 p->sched_class->set_cpus_allowed(p, new_mask);
1209 p->sched_class->set_curr_task(rq);
1211 enqueue_task(rq, p, ENQUEUE_RESTORE);
1215 * Change a given task's CPU affinity. Migrate the thread to a
1216 * proper CPU and schedule it away if the CPU it's executing on
1217 * is removed from the allowed bitmask.
1219 * NOTE: the caller must have a valid reference to the task, the
1220 * task must not exit() & deallocate itself prematurely. The
1221 * call is not atomic; no spinlocks may be held.
1223 static int __set_cpus_allowed_ptr(struct task_struct *p,
1224 const struct cpumask *new_mask, bool check)
1226 unsigned long flags;
1228 unsigned int dest_cpu;
1231 rq = task_rq_lock(p, &flags);
1234 * Must re-check here, to close a race against __kthread_bind(),
1235 * sched_setaffinity() is not guaranteed to observe the flag.
1237 if (check && (p->flags & PF_NO_SETAFFINITY)) {
1242 if (cpumask_equal(&p->cpus_allowed, new_mask))
1245 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
1250 do_set_cpus_allowed(p, new_mask);
1252 /* Can the task run on the task's current CPU? If so, we're done */
1253 if (cpumask_test_cpu(task_cpu(p), new_mask))
1256 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
1257 if (task_running(rq, p) || p->state == TASK_WAKING) {
1258 struct migration_arg arg = { p, dest_cpu };
1259 /* Need help from migration thread: drop lock and wait. */
1260 task_rq_unlock(rq, p, &flags);
1261 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
1262 tlb_migrate_finish(p->mm);
1264 } else if (task_on_rq_queued(p)) {
1266 * OK, since we're going to drop the lock immediately
1267 * afterwards anyway.
1269 lockdep_unpin_lock(&rq->lock);
1270 rq = move_queued_task(rq, p, dest_cpu);
1271 lockdep_pin_lock(&rq->lock);
1274 task_rq_unlock(rq, p, &flags);
1279 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
1281 return __set_cpus_allowed_ptr(p, new_mask, false);
1283 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
1285 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1287 #ifdef CONFIG_SCHED_DEBUG
1289 * We should never call set_task_cpu() on a blocked task,
1290 * ttwu() will sort out the placement.
1292 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
1295 #ifdef CONFIG_LOCKDEP
1297 * The caller should hold either p->pi_lock or rq->lock, when changing
1298 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
1300 * sched_move_task() holds both and thus holding either pins the cgroup,
1303 * Furthermore, all task_rq users should acquire both locks, see
1306 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
1307 lockdep_is_held(&task_rq(p)->lock)));
1311 trace_sched_migrate_task(p, new_cpu);
1313 if (task_cpu(p) != new_cpu) {
1314 if (p->sched_class->migrate_task_rq)
1315 p->sched_class->migrate_task_rq(p);
1316 p->se.nr_migrations++;
1317 perf_event_task_migrate(p);
1319 walt_fixup_busy_time(p, new_cpu);
1322 __set_task_cpu(p, new_cpu);
1325 static void __migrate_swap_task(struct task_struct *p, int cpu)
1327 if (task_on_rq_queued(p)) {
1328 struct rq *src_rq, *dst_rq;
1330 src_rq = task_rq(p);
1331 dst_rq = cpu_rq(cpu);
1333 deactivate_task(src_rq, p, 0);
1334 p->on_rq = TASK_ON_RQ_MIGRATING;
1335 set_task_cpu(p, cpu);
1336 p->on_rq = TASK_ON_RQ_QUEUED;
1337 activate_task(dst_rq, p, 0);
1338 check_preempt_curr(dst_rq, p, 0);
1341 * Task isn't running anymore; make it appear like we migrated
1342 * it before it went to sleep. This means on wakeup we make the
1343 * previous cpu our targer instead of where it really is.
1349 struct migration_swap_arg {
1350 struct task_struct *src_task, *dst_task;
1351 int src_cpu, dst_cpu;
1354 static int migrate_swap_stop(void *data)
1356 struct migration_swap_arg *arg = data;
1357 struct rq *src_rq, *dst_rq;
1360 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1363 src_rq = cpu_rq(arg->src_cpu);
1364 dst_rq = cpu_rq(arg->dst_cpu);
1366 double_raw_lock(&arg->src_task->pi_lock,
1367 &arg->dst_task->pi_lock);
1368 double_rq_lock(src_rq, dst_rq);
1370 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1373 if (task_cpu(arg->src_task) != arg->src_cpu)
1376 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1379 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1382 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1383 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1388 double_rq_unlock(src_rq, dst_rq);
1389 raw_spin_unlock(&arg->dst_task->pi_lock);
1390 raw_spin_unlock(&arg->src_task->pi_lock);
1396 * Cross migrate two tasks
1398 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1400 struct migration_swap_arg arg;
1403 arg = (struct migration_swap_arg){
1405 .src_cpu = task_cpu(cur),
1407 .dst_cpu = task_cpu(p),
1410 if (arg.src_cpu == arg.dst_cpu)
1414 * These three tests are all lockless; this is OK since all of them
1415 * will be re-checked with proper locks held further down the line.
1417 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1420 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1423 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1426 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1427 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1434 * wait_task_inactive - wait for a thread to unschedule.
1436 * If @match_state is nonzero, it's the @p->state value just checked and
1437 * not expected to change. If it changes, i.e. @p might have woken up,
1438 * then return zero. When we succeed in waiting for @p to be off its CPU,
1439 * we return a positive number (its total switch count). If a second call
1440 * a short while later returns the same number, the caller can be sure that
1441 * @p has remained unscheduled the whole time.
1443 * The caller must ensure that the task *will* unschedule sometime soon,
1444 * else this function might spin for a *long* time. This function can't
1445 * be called with interrupts off, or it may introduce deadlock with
1446 * smp_call_function() if an IPI is sent by the same process we are
1447 * waiting to become inactive.
1449 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1451 unsigned long flags;
1452 int running, queued;
1458 * We do the initial early heuristics without holding
1459 * any task-queue locks at all. We'll only try to get
1460 * the runqueue lock when things look like they will
1466 * If the task is actively running on another CPU
1467 * still, just relax and busy-wait without holding
1470 * NOTE! Since we don't hold any locks, it's not
1471 * even sure that "rq" stays as the right runqueue!
1472 * But we don't care, since "task_running()" will
1473 * return false if the runqueue has changed and p
1474 * is actually now running somewhere else!
1476 while (task_running(rq, p)) {
1477 if (match_state && unlikely(p->state != match_state))
1483 * Ok, time to look more closely! We need the rq
1484 * lock now, to be *sure*. If we're wrong, we'll
1485 * just go back and repeat.
1487 rq = task_rq_lock(p, &flags);
1488 trace_sched_wait_task(p);
1489 running = task_running(rq, p);
1490 queued = task_on_rq_queued(p);
1492 if (!match_state || p->state == match_state)
1493 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1494 task_rq_unlock(rq, p, &flags);
1497 * If it changed from the expected state, bail out now.
1499 if (unlikely(!ncsw))
1503 * Was it really running after all now that we
1504 * checked with the proper locks actually held?
1506 * Oops. Go back and try again..
1508 if (unlikely(running)) {
1514 * It's not enough that it's not actively running,
1515 * it must be off the runqueue _entirely_, and not
1518 * So if it was still runnable (but just not actively
1519 * running right now), it's preempted, and we should
1520 * yield - it could be a while.
1522 if (unlikely(queued)) {
1523 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1525 set_current_state(TASK_UNINTERRUPTIBLE);
1526 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1531 * Ahh, all good. It wasn't running, and it wasn't
1532 * runnable, which means that it will never become
1533 * running in the future either. We're all done!
1542 * kick_process - kick a running thread to enter/exit the kernel
1543 * @p: the to-be-kicked thread
1545 * Cause a process which is running on another CPU to enter
1546 * kernel-mode, without any delay. (to get signals handled.)
1548 * NOTE: this function doesn't have to take the runqueue lock,
1549 * because all it wants to ensure is that the remote task enters
1550 * the kernel. If the IPI races and the task has been migrated
1551 * to another CPU then no harm is done and the purpose has been
1554 void kick_process(struct task_struct *p)
1560 if ((cpu != smp_processor_id()) && task_curr(p))
1561 smp_send_reschedule(cpu);
1564 EXPORT_SYMBOL_GPL(kick_process);
1567 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1569 static int select_fallback_rq(int cpu, struct task_struct *p)
1571 int nid = cpu_to_node(cpu);
1572 const struct cpumask *nodemask = NULL;
1573 enum { cpuset, possible, fail } state = cpuset;
1577 * If the node that the cpu is on has been offlined, cpu_to_node()
1578 * will return -1. There is no cpu on the node, and we should
1579 * select the cpu on the other node.
1582 nodemask = cpumask_of_node(nid);
1584 /* Look for allowed, online CPU in same node. */
1585 for_each_cpu(dest_cpu, nodemask) {
1586 if (!cpu_online(dest_cpu))
1588 if (!cpu_active(dest_cpu))
1590 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1596 /* Any allowed, online CPU? */
1597 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1598 if (!cpu_online(dest_cpu))
1600 if (!cpu_active(dest_cpu))
1605 /* No more Mr. Nice Guy. */
1608 if (IS_ENABLED(CONFIG_CPUSETS)) {
1609 cpuset_cpus_allowed_fallback(p);
1615 do_set_cpus_allowed(p, cpu_possible_mask);
1626 if (state != cpuset) {
1628 * Don't tell them about moving exiting tasks or
1629 * kernel threads (both mm NULL), since they never
1632 if (p->mm && printk_ratelimit()) {
1633 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1634 task_pid_nr(p), p->comm, cpu);
1642 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1645 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1647 lockdep_assert_held(&p->pi_lock);
1649 if (p->nr_cpus_allowed > 1)
1650 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1653 * In order not to call set_task_cpu() on a blocking task we need
1654 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1657 * Since this is common to all placement strategies, this lives here.
1659 * [ this allows ->select_task() to simply return task_cpu(p) and
1660 * not worry about this generic constraint ]
1662 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1664 cpu = select_fallback_rq(task_cpu(p), p);
1669 static void update_avg(u64 *avg, u64 sample)
1671 s64 diff = sample - *avg;
1677 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1678 const struct cpumask *new_mask, bool check)
1680 return set_cpus_allowed_ptr(p, new_mask);
1683 #endif /* CONFIG_SMP */
1686 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1688 #ifdef CONFIG_SCHEDSTATS
1689 struct rq *rq = this_rq();
1692 int this_cpu = smp_processor_id();
1694 if (cpu == this_cpu) {
1695 schedstat_inc(rq, ttwu_local);
1696 schedstat_inc(p, se.statistics.nr_wakeups_local);
1698 struct sched_domain *sd;
1700 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1702 for_each_domain(this_cpu, sd) {
1703 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1704 schedstat_inc(sd, ttwu_wake_remote);
1711 if (wake_flags & WF_MIGRATED)
1712 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1714 #endif /* CONFIG_SMP */
1716 schedstat_inc(rq, ttwu_count);
1717 schedstat_inc(p, se.statistics.nr_wakeups);
1719 if (wake_flags & WF_SYNC)
1720 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1722 #endif /* CONFIG_SCHEDSTATS */
1725 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1727 activate_task(rq, p, en_flags);
1728 p->on_rq = TASK_ON_RQ_QUEUED;
1730 /* if a worker is waking up, notify workqueue */
1731 if (p->flags & PF_WQ_WORKER)
1732 wq_worker_waking_up(p, cpu_of(rq));
1736 * Mark the task runnable and perform wakeup-preemption.
1739 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1741 check_preempt_curr(rq, p, wake_flags);
1742 p->state = TASK_RUNNING;
1743 trace_sched_wakeup(p);
1746 if (p->sched_class->task_woken) {
1748 * Our task @p is fully woken up and running; so its safe to
1749 * drop the rq->lock, hereafter rq is only used for statistics.
1751 lockdep_unpin_lock(&rq->lock);
1752 p->sched_class->task_woken(rq, p);
1753 lockdep_pin_lock(&rq->lock);
1756 if (rq->idle_stamp) {
1757 u64 delta = rq_clock(rq) - rq->idle_stamp;
1758 u64 max = 2*rq->max_idle_balance_cost;
1760 update_avg(&rq->avg_idle, delta);
1762 if (rq->avg_idle > max)
1771 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1773 lockdep_assert_held(&rq->lock);
1776 if (p->sched_contributes_to_load)
1777 rq->nr_uninterruptible--;
1780 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1781 ttwu_do_wakeup(rq, p, wake_flags);
1785 * Called in case the task @p isn't fully descheduled from its runqueue,
1786 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1787 * since all we need to do is flip p->state to TASK_RUNNING, since
1788 * the task is still ->on_rq.
1790 static int ttwu_remote(struct task_struct *p, int wake_flags)
1795 rq = __task_rq_lock(p);
1796 if (task_on_rq_queued(p)) {
1797 /* check_preempt_curr() may use rq clock */
1798 update_rq_clock(rq);
1799 ttwu_do_wakeup(rq, p, wake_flags);
1802 __task_rq_unlock(rq);
1808 void sched_ttwu_pending(void)
1810 struct rq *rq = this_rq();
1811 struct llist_node *llist = llist_del_all(&rq->wake_list);
1812 struct task_struct *p;
1813 unsigned long flags;
1818 raw_spin_lock_irqsave(&rq->lock, flags);
1819 lockdep_pin_lock(&rq->lock);
1822 p = llist_entry(llist, struct task_struct, wake_entry);
1823 llist = llist_next(llist);
1824 ttwu_do_activate(rq, p, 0);
1827 lockdep_unpin_lock(&rq->lock);
1828 raw_spin_unlock_irqrestore(&rq->lock, flags);
1831 void scheduler_ipi(void)
1834 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1835 * TIF_NEED_RESCHED remotely (for the first time) will also send
1838 preempt_fold_need_resched();
1840 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1844 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1845 * traditionally all their work was done from the interrupt return
1846 * path. Now that we actually do some work, we need to make sure
1849 * Some archs already do call them, luckily irq_enter/exit nest
1852 * Arguably we should visit all archs and update all handlers,
1853 * however a fair share of IPIs are still resched only so this would
1854 * somewhat pessimize the simple resched case.
1857 sched_ttwu_pending();
1860 * Check if someone kicked us for doing the nohz idle load balance.
1862 if (unlikely(got_nohz_idle_kick())) {
1863 this_rq()->idle_balance = 1;
1864 raise_softirq_irqoff(SCHED_SOFTIRQ);
1869 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1871 struct rq *rq = cpu_rq(cpu);
1873 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1874 if (!set_nr_if_polling(rq->idle))
1875 smp_send_reschedule(cpu);
1877 trace_sched_wake_idle_without_ipi(cpu);
1881 void wake_up_if_idle(int cpu)
1883 struct rq *rq = cpu_rq(cpu);
1884 unsigned long flags;
1888 if (!is_idle_task(rcu_dereference(rq->curr)))
1891 if (set_nr_if_polling(rq->idle)) {
1892 trace_sched_wake_idle_without_ipi(cpu);
1894 raw_spin_lock_irqsave(&rq->lock, flags);
1895 if (is_idle_task(rq->curr))
1896 smp_send_reschedule(cpu);
1897 /* Else cpu is not in idle, do nothing here */
1898 raw_spin_unlock_irqrestore(&rq->lock, flags);
1905 bool cpus_share_cache(int this_cpu, int that_cpu)
1907 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1909 #endif /* CONFIG_SMP */
1911 static void ttwu_queue(struct task_struct *p, int cpu)
1913 struct rq *rq = cpu_rq(cpu);
1915 #if defined(CONFIG_SMP)
1916 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1917 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1918 ttwu_queue_remote(p, cpu);
1923 raw_spin_lock(&rq->lock);
1924 lockdep_pin_lock(&rq->lock);
1925 ttwu_do_activate(rq, p, 0);
1926 lockdep_unpin_lock(&rq->lock);
1927 raw_spin_unlock(&rq->lock);
1931 * try_to_wake_up - wake up a thread
1932 * @p: the thread to be awakened
1933 * @state: the mask of task states that can be woken
1934 * @wake_flags: wake modifier flags (WF_*)
1936 * Put it on the run-queue if it's not already there. The "current"
1937 * thread is always on the run-queue (except when the actual
1938 * re-schedule is in progress), and as such you're allowed to do
1939 * the simpler "current->state = TASK_RUNNING" to mark yourself
1940 * runnable without the overhead of this.
1942 * Return: %true if @p was woken up, %false if it was already running.
1943 * or @state didn't match @p's state.
1946 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1948 unsigned long flags;
1949 int cpu, success = 0;
1956 * If we are going to wake up a thread waiting for CONDITION we
1957 * need to ensure that CONDITION=1 done by the caller can not be
1958 * reordered with p->state check below. This pairs with mb() in
1959 * set_current_state() the waiting thread does.
1961 smp_mb__before_spinlock();
1962 raw_spin_lock_irqsave(&p->pi_lock, flags);
1963 if (!(p->state & state))
1966 trace_sched_waking(p);
1968 success = 1; /* we're going to change ->state */
1972 * Ensure we load p->on_rq _after_ p->state, otherwise it would
1973 * be possible to, falsely, observe p->on_rq == 0 and get stuck
1974 * in smp_cond_load_acquire() below.
1976 * sched_ttwu_pending() try_to_wake_up()
1977 * [S] p->on_rq = 1; [L] P->state
1978 * UNLOCK rq->lock -----.
1982 * LOCK rq->lock -----'
1986 * [S] p->state = UNINTERRUPTIBLE [L] p->on_rq
1988 * Pairs with the UNLOCK+LOCK on rq->lock from the
1989 * last wakeup of our task and the schedule that got our task
1993 if (p->on_rq && ttwu_remote(p, wake_flags))
1998 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1999 * possible to, falsely, observe p->on_cpu == 0.
2001 * One must be running (->on_cpu == 1) in order to remove oneself
2002 * from the runqueue.
2004 * [S] ->on_cpu = 1; [L] ->on_rq
2008 * [S] ->on_rq = 0; [L] ->on_cpu
2010 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
2011 * from the consecutive calls to schedule(); the first switching to our
2012 * task, the second putting it to sleep.
2017 * If the owning (remote) cpu is still in the middle of schedule() with
2018 * this task as prev, wait until its done referencing the task.
2023 * Combined with the control dependency above, we have an effective
2024 * smp_load_acquire() without the need for full barriers.
2026 * Pairs with the smp_store_release() in finish_lock_switch().
2028 * This ensures that tasks getting woken will be fully ordered against
2029 * their previous state and preserve Program Order.
2033 rq = cpu_rq(task_cpu(p));
2035 raw_spin_lock(&rq->lock);
2036 wallclock = walt_ktime_clock();
2037 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2038 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2039 raw_spin_unlock(&rq->lock);
2041 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2042 p->state = TASK_WAKING;
2044 if (p->sched_class->task_waking)
2045 p->sched_class->task_waking(p);
2047 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2049 if (task_cpu(p) != cpu) {
2050 wake_flags |= WF_MIGRATED;
2051 set_task_cpu(p, cpu);
2054 #endif /* CONFIG_SMP */
2058 ttwu_stat(p, cpu, wake_flags);
2060 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2066 * try_to_wake_up_local - try to wake up a local task with rq lock held
2067 * @p: the thread to be awakened
2069 * Put @p on the run-queue if it's not already there. The caller must
2070 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2073 static void try_to_wake_up_local(struct task_struct *p)
2075 struct rq *rq = task_rq(p);
2077 if (WARN_ON_ONCE(rq != this_rq()) ||
2078 WARN_ON_ONCE(p == current))
2081 lockdep_assert_held(&rq->lock);
2083 if (!raw_spin_trylock(&p->pi_lock)) {
2085 * This is OK, because current is on_cpu, which avoids it being
2086 * picked for load-balance and preemption/IRQs are still
2087 * disabled avoiding further scheduler activity on it and we've
2088 * not yet picked a replacement task.
2090 lockdep_unpin_lock(&rq->lock);
2091 raw_spin_unlock(&rq->lock);
2092 raw_spin_lock(&p->pi_lock);
2093 raw_spin_lock(&rq->lock);
2094 lockdep_pin_lock(&rq->lock);
2097 if (!(p->state & TASK_NORMAL))
2100 trace_sched_waking(p);
2102 if (!task_on_rq_queued(p)) {
2103 u64 wallclock = walt_ktime_clock();
2105 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2106 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2107 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2110 ttwu_do_wakeup(rq, p, 0);
2111 ttwu_stat(p, smp_processor_id(), 0);
2113 raw_spin_unlock(&p->pi_lock);
2117 * wake_up_process - Wake up a specific process
2118 * @p: The process to be woken up.
2120 * Attempt to wake up the nominated process and move it to the set of runnable
2123 * Return: 1 if the process was woken up, 0 if it was already running.
2125 * It may be assumed that this function implies a write memory barrier before
2126 * changing the task state if and only if any tasks are woken up.
2128 int wake_up_process(struct task_struct *p)
2130 return try_to_wake_up(p, TASK_NORMAL, 0);
2132 EXPORT_SYMBOL(wake_up_process);
2134 int wake_up_state(struct task_struct *p, unsigned int state)
2136 return try_to_wake_up(p, state, 0);
2140 * This function clears the sched_dl_entity static params.
2142 void __dl_clear_params(struct task_struct *p)
2144 struct sched_dl_entity *dl_se = &p->dl;
2146 dl_se->dl_runtime = 0;
2147 dl_se->dl_deadline = 0;
2148 dl_se->dl_period = 0;
2152 dl_se->dl_throttled = 0;
2154 dl_se->dl_yielded = 0;
2158 * Perform scheduler related setup for a newly forked process p.
2159 * p is forked by current.
2161 * __sched_fork() is basic setup used by init_idle() too:
2163 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2168 p->se.exec_start = 0;
2169 p->se.sum_exec_runtime = 0;
2170 p->se.prev_sum_exec_runtime = 0;
2171 p->se.nr_migrations = 0;
2173 #ifdef CONFIG_SCHED_WALT
2174 p->last_sleep_ts = 0;
2177 INIT_LIST_HEAD(&p->se.group_node);
2178 walt_init_new_task_load(p);
2180 #ifdef CONFIG_FAIR_GROUP_SCHED
2181 p->se.cfs_rq = NULL;
2184 #ifdef CONFIG_SCHEDSTATS
2185 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2188 RB_CLEAR_NODE(&p->dl.rb_node);
2189 init_dl_task_timer(&p->dl);
2190 __dl_clear_params(p);
2192 INIT_LIST_HEAD(&p->rt.run_list);
2194 #ifdef CONFIG_PREEMPT_NOTIFIERS
2195 INIT_HLIST_HEAD(&p->preempt_notifiers);
2198 #ifdef CONFIG_NUMA_BALANCING
2199 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2200 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2201 p->mm->numa_scan_seq = 0;
2204 if (clone_flags & CLONE_VM)
2205 p->numa_preferred_nid = current->numa_preferred_nid;
2207 p->numa_preferred_nid = -1;
2209 p->node_stamp = 0ULL;
2210 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2211 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2212 p->numa_work.next = &p->numa_work;
2213 p->numa_faults = NULL;
2214 p->last_task_numa_placement = 0;
2215 p->last_sum_exec_runtime = 0;
2217 p->numa_group = NULL;
2218 #endif /* CONFIG_NUMA_BALANCING */
2221 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2223 #ifdef CONFIG_NUMA_BALANCING
2225 void set_numabalancing_state(bool enabled)
2228 static_branch_enable(&sched_numa_balancing);
2230 static_branch_disable(&sched_numa_balancing);
2233 #ifdef CONFIG_PROC_SYSCTL
2234 int sysctl_numa_balancing(struct ctl_table *table, int write,
2235 void __user *buffer, size_t *lenp, loff_t *ppos)
2239 int state = static_branch_likely(&sched_numa_balancing);
2241 if (write && !capable(CAP_SYS_ADMIN))
2246 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2250 set_numabalancing_state(state);
2257 * fork()/clone()-time setup:
2259 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2261 unsigned long flags;
2262 int cpu = get_cpu();
2264 __sched_fork(clone_flags, p);
2266 * We mark the process as NEW here. This guarantees that
2267 * nobody will actually run it, and a signal or other external
2268 * event cannot wake it up and insert it on the runqueue either.
2270 p->state = TASK_NEW;
2273 * Make sure we do not leak PI boosting priority to the child.
2275 p->prio = current->normal_prio;
2278 * Revert to default priority/policy on fork if requested.
2280 if (unlikely(p->sched_reset_on_fork)) {
2281 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2282 p->policy = SCHED_NORMAL;
2283 p->static_prio = NICE_TO_PRIO(0);
2285 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2286 p->static_prio = NICE_TO_PRIO(0);
2288 p->prio = p->normal_prio = __normal_prio(p);
2292 * We don't need the reset flag anymore after the fork. It has
2293 * fulfilled its duty:
2295 p->sched_reset_on_fork = 0;
2298 if (dl_prio(p->prio)) {
2301 } else if (rt_prio(p->prio)) {
2302 p->sched_class = &rt_sched_class;
2304 p->sched_class = &fair_sched_class;
2307 init_entity_runnable_average(&p->se);
2310 * The child is not yet in the pid-hash so no cgroup attach races,
2311 * and the cgroup is pinned to this child due to cgroup_fork()
2312 * is ran before sched_fork().
2314 * Silence PROVE_RCU.
2316 raw_spin_lock_irqsave(&p->pi_lock, flags);
2318 * We're setting the cpu for the first time, we don't migrate,
2319 * so use __set_task_cpu().
2321 __set_task_cpu(p, cpu);
2322 if (p->sched_class->task_fork)
2323 p->sched_class->task_fork(p);
2324 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2326 #ifdef CONFIG_SCHED_INFO
2327 if (likely(sched_info_on()))
2328 memset(&p->sched_info, 0, sizeof(p->sched_info));
2330 #if defined(CONFIG_SMP)
2333 init_task_preempt_count(p);
2335 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2336 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2343 unsigned long to_ratio(u64 period, u64 runtime)
2345 if (runtime == RUNTIME_INF)
2349 * Doing this here saves a lot of checks in all
2350 * the calling paths, and returning zero seems
2351 * safe for them anyway.
2356 return div64_u64(runtime << 20, period);
2360 inline struct dl_bw *dl_bw_of(int i)
2362 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2363 "sched RCU must be held");
2364 return &cpu_rq(i)->rd->dl_bw;
2367 static inline int dl_bw_cpus(int i)
2369 struct root_domain *rd = cpu_rq(i)->rd;
2372 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2373 "sched RCU must be held");
2374 for_each_cpu_and(i, rd->span, cpu_active_mask)
2380 inline struct dl_bw *dl_bw_of(int i)
2382 return &cpu_rq(i)->dl.dl_bw;
2385 static inline int dl_bw_cpus(int i)
2392 * We must be sure that accepting a new task (or allowing changing the
2393 * parameters of an existing one) is consistent with the bandwidth
2394 * constraints. If yes, this function also accordingly updates the currently
2395 * allocated bandwidth to reflect the new situation.
2397 * This function is called while holding p's rq->lock.
2399 * XXX we should delay bw change until the task's 0-lag point, see
2402 static int dl_overflow(struct task_struct *p, int policy,
2403 const struct sched_attr *attr)
2406 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2407 u64 period = attr->sched_period ?: attr->sched_deadline;
2408 u64 runtime = attr->sched_runtime;
2409 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2412 if (new_bw == p->dl.dl_bw)
2416 * Either if a task, enters, leave, or stays -deadline but changes
2417 * its parameters, we may need to update accordingly the total
2418 * allocated bandwidth of the container.
2420 raw_spin_lock(&dl_b->lock);
2421 cpus = dl_bw_cpus(task_cpu(p));
2422 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2423 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2424 __dl_add(dl_b, new_bw);
2426 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2427 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2428 __dl_clear(dl_b, p->dl.dl_bw);
2429 __dl_add(dl_b, new_bw);
2431 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2432 __dl_clear(dl_b, p->dl.dl_bw);
2435 raw_spin_unlock(&dl_b->lock);
2440 extern void init_dl_bw(struct dl_bw *dl_b);
2443 * wake_up_new_task - wake up a newly created task for the first time.
2445 * This function will do some initial scheduler statistics housekeeping
2446 * that must be done for every newly created context, then puts the task
2447 * on the runqueue and wakes it.
2449 void wake_up_new_task(struct task_struct *p)
2451 unsigned long flags;
2454 raw_spin_lock_irqsave(&p->pi_lock, flags);
2455 p->state = TASK_RUNNING;
2457 walt_init_new_task_load(p);
2459 /* Initialize new task's runnable average */
2460 init_entity_runnable_average(&p->se);
2463 * Fork balancing, do it here and not earlier because:
2464 * - cpus_allowed can change in the fork path
2465 * - any previously selected cpu might disappear through hotplug
2467 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2468 * as we're not fully set-up yet.
2470 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2472 rq = __task_rq_lock(p);
2473 update_rq_clock(rq);
2474 post_init_entity_util_avg(&p->se);
2476 walt_mark_task_starting(p);
2477 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2478 p->on_rq = TASK_ON_RQ_QUEUED;
2479 trace_sched_wakeup_new(p);
2480 check_preempt_curr(rq, p, WF_FORK);
2482 if (p->sched_class->task_woken) {
2484 * Nothing relies on rq->lock after this, so its fine to
2487 lockdep_unpin_lock(&rq->lock);
2488 p->sched_class->task_woken(rq, p);
2489 lockdep_pin_lock(&rq->lock);
2492 task_rq_unlock(rq, p, &flags);
2495 #ifdef CONFIG_PREEMPT_NOTIFIERS
2497 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2499 void preempt_notifier_inc(void)
2501 static_key_slow_inc(&preempt_notifier_key);
2503 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2505 void preempt_notifier_dec(void)
2507 static_key_slow_dec(&preempt_notifier_key);
2509 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2512 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2513 * @notifier: notifier struct to register
2515 void preempt_notifier_register(struct preempt_notifier *notifier)
2517 if (!static_key_false(&preempt_notifier_key))
2518 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2520 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2522 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2525 * preempt_notifier_unregister - no longer interested in preemption notifications
2526 * @notifier: notifier struct to unregister
2528 * This is *not* safe to call from within a preemption notifier.
2530 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2532 hlist_del(¬ifier->link);
2534 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2536 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2538 struct preempt_notifier *notifier;
2540 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2541 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2544 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2546 if (static_key_false(&preempt_notifier_key))
2547 __fire_sched_in_preempt_notifiers(curr);
2551 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2552 struct task_struct *next)
2554 struct preempt_notifier *notifier;
2556 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2557 notifier->ops->sched_out(notifier, next);
2560 static __always_inline void
2561 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2562 struct task_struct *next)
2564 if (static_key_false(&preempt_notifier_key))
2565 __fire_sched_out_preempt_notifiers(curr, next);
2568 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2570 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2575 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2576 struct task_struct *next)
2580 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2583 * prepare_task_switch - prepare to switch tasks
2584 * @rq: the runqueue preparing to switch
2585 * @prev: the current task that is being switched out
2586 * @next: the task we are going to switch to.
2588 * This is called with the rq lock held and interrupts off. It must
2589 * be paired with a subsequent finish_task_switch after the context
2592 * prepare_task_switch sets up locking and calls architecture specific
2596 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2597 struct task_struct *next)
2599 sched_info_switch(rq, prev, next);
2600 perf_event_task_sched_out(prev, next);
2601 fire_sched_out_preempt_notifiers(prev, next);
2602 prepare_lock_switch(rq, next);
2603 prepare_arch_switch(next);
2607 * finish_task_switch - clean up after a task-switch
2608 * @prev: the thread we just switched away from.
2610 * finish_task_switch must be called after the context switch, paired
2611 * with a prepare_task_switch call before the context switch.
2612 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2613 * and do any other architecture-specific cleanup actions.
2615 * Note that we may have delayed dropping an mm in context_switch(). If
2616 * so, we finish that here outside of the runqueue lock. (Doing it
2617 * with the lock held can cause deadlocks; see schedule() for
2620 * The context switch have flipped the stack from under us and restored the
2621 * local variables which were saved when this task called schedule() in the
2622 * past. prev == current is still correct but we need to recalculate this_rq
2623 * because prev may have moved to another CPU.
2625 static struct rq *finish_task_switch(struct task_struct *prev)
2626 __releases(rq->lock)
2628 struct rq *rq = this_rq();
2629 struct mm_struct *mm = rq->prev_mm;
2633 * The previous task will have left us with a preempt_count of 2
2634 * because it left us after:
2637 * preempt_disable(); // 1
2639 * raw_spin_lock_irq(&rq->lock) // 2
2641 * Also, see FORK_PREEMPT_COUNT.
2643 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2644 "corrupted preempt_count: %s/%d/0x%x\n",
2645 current->comm, current->pid, preempt_count()))
2646 preempt_count_set(FORK_PREEMPT_COUNT);
2651 * A task struct has one reference for the use as "current".
2652 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2653 * schedule one last time. The schedule call will never return, and
2654 * the scheduled task must drop that reference.
2656 * We must observe prev->state before clearing prev->on_cpu (in
2657 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2658 * running on another CPU and we could rave with its RUNNING -> DEAD
2659 * transition, resulting in a double drop.
2661 prev_state = prev->state;
2662 vtime_task_switch(prev);
2663 perf_event_task_sched_in(prev, current);
2664 finish_lock_switch(rq, prev);
2665 finish_arch_post_lock_switch();
2667 fire_sched_in_preempt_notifiers(current);
2670 if (unlikely(prev_state == TASK_DEAD)) {
2671 if (prev->sched_class->task_dead)
2672 prev->sched_class->task_dead(prev);
2675 * Remove function-return probe instances associated with this
2676 * task and put them back on the free list.
2678 kprobe_flush_task(prev);
2679 put_task_struct(prev);
2682 tick_nohz_task_switch();
2688 /* rq->lock is NOT held, but preemption is disabled */
2689 static void __balance_callback(struct rq *rq)
2691 struct callback_head *head, *next;
2692 void (*func)(struct rq *rq);
2693 unsigned long flags;
2695 raw_spin_lock_irqsave(&rq->lock, flags);
2696 head = rq->balance_callback;
2697 rq->balance_callback = NULL;
2699 func = (void (*)(struct rq *))head->func;
2706 raw_spin_unlock_irqrestore(&rq->lock, flags);
2709 static inline void balance_callback(struct rq *rq)
2711 if (unlikely(rq->balance_callback))
2712 __balance_callback(rq);
2717 static inline void balance_callback(struct rq *rq)
2724 * schedule_tail - first thing a freshly forked thread must call.
2725 * @prev: the thread we just switched away from.
2727 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2728 __releases(rq->lock)
2733 * New tasks start with FORK_PREEMPT_COUNT, see there and
2734 * finish_task_switch() for details.
2736 * finish_task_switch() will drop rq->lock() and lower preempt_count
2737 * and the preempt_enable() will end up enabling preemption (on
2738 * PREEMPT_COUNT kernels).
2741 rq = finish_task_switch(prev);
2742 balance_callback(rq);
2745 if (current->set_child_tid)
2746 put_user(task_pid_vnr(current), current->set_child_tid);
2750 * context_switch - switch to the new MM and the new thread's register state.
2752 static inline struct rq *
2753 context_switch(struct rq *rq, struct task_struct *prev,
2754 struct task_struct *next)
2756 struct mm_struct *mm, *oldmm;
2758 prepare_task_switch(rq, prev, next);
2761 oldmm = prev->active_mm;
2763 * For paravirt, this is coupled with an exit in switch_to to
2764 * combine the page table reload and the switch backend into
2767 arch_start_context_switch(prev);
2770 next->active_mm = oldmm;
2771 atomic_inc(&oldmm->mm_count);
2772 enter_lazy_tlb(oldmm, next);
2774 switch_mm(oldmm, mm, next);
2777 prev->active_mm = NULL;
2778 rq->prev_mm = oldmm;
2781 * Since the runqueue lock will be released by the next
2782 * task (which is an invalid locking op but in the case
2783 * of the scheduler it's an obvious special-case), so we
2784 * do an early lockdep release here:
2786 lockdep_unpin_lock(&rq->lock);
2787 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2789 /* Here we just switch the register state and the stack. */
2790 switch_to(prev, next, prev);
2793 return finish_task_switch(prev);
2797 * nr_running and nr_context_switches:
2799 * externally visible scheduler statistics: current number of runnable
2800 * threads, total number of context switches performed since bootup.
2802 unsigned long nr_running(void)
2804 unsigned long i, sum = 0;
2806 for_each_online_cpu(i)
2807 sum += cpu_rq(i)->nr_running;
2813 * Check if only the current task is running on the cpu.
2815 * Caution: this function does not check that the caller has disabled
2816 * preemption, thus the result might have a time-of-check-to-time-of-use
2817 * race. The caller is responsible to use it correctly, for example:
2819 * - from a non-preemptable section (of course)
2821 * - from a thread that is bound to a single CPU
2823 * - in a loop with very short iterations (e.g. a polling loop)
2825 bool single_task_running(void)
2827 return raw_rq()->nr_running == 1;
2829 EXPORT_SYMBOL(single_task_running);
2831 unsigned long long nr_context_switches(void)
2834 unsigned long long sum = 0;
2836 for_each_possible_cpu(i)
2837 sum += cpu_rq(i)->nr_switches;
2842 unsigned long nr_iowait(void)
2844 unsigned long i, sum = 0;
2846 for_each_possible_cpu(i)
2847 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2852 unsigned long nr_iowait_cpu(int cpu)
2854 struct rq *this = cpu_rq(cpu);
2855 return atomic_read(&this->nr_iowait);
2858 #ifdef CONFIG_CPU_QUIET
2859 u64 nr_running_integral(unsigned int cpu)
2861 unsigned int seqcnt;
2865 if (cpu >= nr_cpu_ids)
2871 * Update average to avoid reading stalled value if there were
2872 * no run-queue changes for a long time. On the other hand if
2873 * the changes are happening right now, just read current value
2877 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2878 integral = do_nr_running_integral(q);
2879 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2880 read_seqcount_begin(&q->ave_seqcnt);
2881 integral = q->nr_running_integral;
2888 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2890 struct rq *rq = this_rq();
2891 *nr_waiters = atomic_read(&rq->nr_iowait);
2892 *load = rq->load.weight;
2898 * sched_exec - execve() is a valuable balancing opportunity, because at
2899 * this point the task has the smallest effective memory and cache footprint.
2901 void sched_exec(void)
2903 struct task_struct *p = current;
2904 unsigned long flags;
2907 raw_spin_lock_irqsave(&p->pi_lock, flags);
2908 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2909 if (dest_cpu == smp_processor_id())
2912 if (likely(cpu_active(dest_cpu))) {
2913 struct migration_arg arg = { p, dest_cpu };
2915 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2916 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2920 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2925 DEFINE_PER_CPU(struct kernel_stat, kstat);
2926 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2928 EXPORT_PER_CPU_SYMBOL(kstat);
2929 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2932 * Return accounted runtime for the task.
2933 * In case the task is currently running, return the runtime plus current's
2934 * pending runtime that have not been accounted yet.
2936 unsigned long long task_sched_runtime(struct task_struct *p)
2938 unsigned long flags;
2942 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2944 * 64-bit doesn't need locks to atomically read a 64bit value.
2945 * So we have a optimization chance when the task's delta_exec is 0.
2946 * Reading ->on_cpu is racy, but this is ok.
2948 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2949 * If we race with it entering cpu, unaccounted time is 0. This is
2950 * indistinguishable from the read occurring a few cycles earlier.
2951 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2952 * been accounted, so we're correct here as well.
2954 if (!p->on_cpu || !task_on_rq_queued(p))
2955 return p->se.sum_exec_runtime;
2958 rq = task_rq_lock(p, &flags);
2960 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2961 * project cycles that may never be accounted to this
2962 * thread, breaking clock_gettime().
2964 if (task_current(rq, p) && task_on_rq_queued(p)) {
2965 update_rq_clock(rq);
2966 p->sched_class->update_curr(rq);
2968 ns = p->se.sum_exec_runtime;
2969 task_rq_unlock(rq, p, &flags);
2974 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2977 unsigned long add_capacity_margin(unsigned long cpu_capacity)
2979 cpu_capacity = cpu_capacity * capacity_margin;
2980 cpu_capacity /= SCHED_CAPACITY_SCALE;
2981 return cpu_capacity;
2985 unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2986 struct sched_capacity_reqs *scr)
2988 unsigned long total = add_capacity_margin(cfs_cap + scr->rt);
2989 return total += scr->dl;
2992 unsigned long boosted_cpu_util(int cpu);
2993 static void sched_freq_tick_pelt(int cpu)
2995 unsigned long cpu_utilization = boosted_cpu_util(cpu);
2996 unsigned long capacity_curr = capacity_curr_of(cpu);
2997 struct sched_capacity_reqs *scr;
2999 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
3000 if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)
3004 * To make free room for a task that is building up its "real"
3005 * utilization and to harm its performance the least, request
3006 * a jump to a higher OPP as soon as the margin of free capacity
3007 * is impacted (specified by capacity_margin).
3008 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3010 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3011 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3014 #ifdef CONFIG_SCHED_WALT
3015 static void sched_freq_tick_walt(int cpu)
3017 unsigned long cpu_utilization = cpu_util_freq(cpu);
3018 unsigned long capacity_curr = capacity_curr_of(cpu);
3020 if (walt_disabled || !sysctl_sched_use_walt_cpu_util)
3021 return sched_freq_tick_pelt(cpu);
3024 * Add a margin to the WALT utilization to check if we will need to
3025 * increase frequency.
3026 * NOTE: WALT tracks a single CPU signal for all the scheduling
3027 * classes, thus this margin is going to be added to the DL class as
3028 * well, which is something we do not do in sched_freq_tick_pelt case.
3030 if (add_capacity_margin(cpu_utilization) <= capacity_curr)
3034 * It is likely that the load is growing so we
3035 * keep the added margin in our request as an
3037 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3039 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3040 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3043 #define _sched_freq_tick(cpu) sched_freq_tick_walt(cpu)
3045 #define _sched_freq_tick(cpu) sched_freq_tick_pelt(cpu)
3046 #endif /* CONFIG_SCHED_WALT */
3048 static void sched_freq_tick(int cpu)
3053 _sched_freq_tick(cpu);
3056 static inline void sched_freq_tick(int cpu) { }
3057 #endif /* CONFIG_CPU_FREQ_GOV_SCHED */
3060 * This function gets called by the timer code, with HZ frequency.
3061 * We call it with interrupts disabled.
3063 void scheduler_tick(void)
3065 int cpu = smp_processor_id();
3066 struct rq *rq = cpu_rq(cpu);
3067 struct task_struct *curr = rq->curr;
3071 raw_spin_lock(&rq->lock);
3072 walt_set_window_start(rq);
3073 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3074 walt_ktime_clock(), 0);
3075 update_rq_clock(rq);
3076 curr->sched_class->task_tick(rq, curr, 0);
3077 update_cpu_load_active(rq);
3078 calc_global_load_tick(rq);
3079 sched_freq_tick(cpu);
3080 raw_spin_unlock(&rq->lock);
3082 perf_event_task_tick();
3085 rq->idle_balance = idle_cpu(cpu);
3086 trigger_load_balance(rq);
3088 rq_last_tick_reset(rq);
3091 #ifdef CONFIG_NO_HZ_FULL
3093 * scheduler_tick_max_deferment
3095 * Keep at least one tick per second when a single
3096 * active task is running because the scheduler doesn't
3097 * yet completely support full dynticks environment.
3099 * This makes sure that uptime, CFS vruntime, load
3100 * balancing, etc... continue to move forward, even
3101 * with a very low granularity.
3103 * Return: Maximum deferment in nanoseconds.
3105 u64 scheduler_tick_max_deferment(void)
3107 struct rq *rq = this_rq();
3108 unsigned long next, now = READ_ONCE(jiffies);
3110 next = rq->last_sched_tick + HZ;
3112 if (time_before_eq(next, now))
3115 return jiffies_to_nsecs(next - now);
3119 notrace unsigned long get_parent_ip(unsigned long addr)
3121 if (in_lock_functions(addr)) {
3122 addr = CALLER_ADDR2;
3123 if (in_lock_functions(addr))
3124 addr = CALLER_ADDR3;
3129 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3130 defined(CONFIG_PREEMPT_TRACER))
3132 void preempt_count_add(int val)
3134 #ifdef CONFIG_DEBUG_PREEMPT
3138 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3141 __preempt_count_add(val);
3142 #ifdef CONFIG_DEBUG_PREEMPT
3144 * Spinlock count overflowing soon?
3146 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3149 if (preempt_count() == val) {
3150 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3151 #ifdef CONFIG_DEBUG_PREEMPT
3152 current->preempt_disable_ip = ip;
3154 trace_preempt_off(CALLER_ADDR0, ip);
3157 EXPORT_SYMBOL(preempt_count_add);
3158 NOKPROBE_SYMBOL(preempt_count_add);
3160 void preempt_count_sub(int val)
3162 #ifdef CONFIG_DEBUG_PREEMPT
3166 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3169 * Is the spinlock portion underflowing?
3171 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3172 !(preempt_count() & PREEMPT_MASK)))
3176 if (preempt_count() == val)
3177 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3178 __preempt_count_sub(val);
3180 EXPORT_SYMBOL(preempt_count_sub);
3181 NOKPROBE_SYMBOL(preempt_count_sub);
3186 * Print scheduling while atomic bug:
3188 static noinline void __schedule_bug(struct task_struct *prev)
3190 if (oops_in_progress)
3193 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3194 prev->comm, prev->pid, preempt_count());
3196 debug_show_held_locks(prev);
3198 if (irqs_disabled())
3199 print_irqtrace_events(prev);
3200 #ifdef CONFIG_DEBUG_PREEMPT
3201 if (in_atomic_preempt_off()) {
3202 pr_err("Preemption disabled at:");
3203 print_ip_sym(current->preempt_disable_ip);
3208 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3212 * Various schedule()-time debugging checks and statistics:
3214 static inline void schedule_debug(struct task_struct *prev)
3216 #ifdef CONFIG_SCHED_STACK_END_CHECK
3217 if (task_stack_end_corrupted(prev))
3218 panic("corrupted stack end detected inside scheduler\n");
3221 if (unlikely(in_atomic_preempt_off())) {
3222 __schedule_bug(prev);
3223 preempt_count_set(PREEMPT_DISABLED);
3227 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3229 schedstat_inc(this_rq(), sched_count);
3233 * Pick up the highest-prio task:
3235 static inline struct task_struct *
3236 pick_next_task(struct rq *rq, struct task_struct *prev)
3238 const struct sched_class *class = &fair_sched_class;
3239 struct task_struct *p;
3242 * Optimization: we know that if all tasks are in
3243 * the fair class we can call that function directly:
3245 if (likely(prev->sched_class == class &&
3246 rq->nr_running == rq->cfs.h_nr_running)) {
3247 p = fair_sched_class.pick_next_task(rq, prev);
3248 if (unlikely(p == RETRY_TASK))
3251 /* assumes fair_sched_class->next == idle_sched_class */
3253 p = idle_sched_class.pick_next_task(rq, prev);
3259 for_each_class(class) {
3260 p = class->pick_next_task(rq, prev);
3262 if (unlikely(p == RETRY_TASK))
3268 BUG(); /* the idle class will always have a runnable task */
3272 * __schedule() is the main scheduler function.
3274 * The main means of driving the scheduler and thus entering this function are:
3276 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3278 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3279 * paths. For example, see arch/x86/entry_64.S.
3281 * To drive preemption between tasks, the scheduler sets the flag in timer
3282 * interrupt handler scheduler_tick().
3284 * 3. Wakeups don't really cause entry into schedule(). They add a
3285 * task to the run-queue and that's it.
3287 * Now, if the new task added to the run-queue preempts the current
3288 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3289 * called on the nearest possible occasion:
3291 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3293 * - in syscall or exception context, at the next outmost
3294 * preempt_enable(). (this might be as soon as the wake_up()'s
3297 * - in IRQ context, return from interrupt-handler to
3298 * preemptible context
3300 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3303 * - cond_resched() call
3304 * - explicit schedule() call
3305 * - return from syscall or exception to user-space
3306 * - return from interrupt-handler to user-space
3308 * WARNING: must be called with preemption disabled!
3310 static void __sched notrace __schedule(bool preempt)
3312 struct task_struct *prev, *next;
3313 unsigned long *switch_count;
3318 cpu = smp_processor_id();
3320 rcu_note_context_switch();
3324 * do_exit() calls schedule() with preemption disabled as an exception;
3325 * however we must fix that up, otherwise the next task will see an
3326 * inconsistent (higher) preempt count.
3328 * It also avoids the below schedule_debug() test from complaining
3331 if (unlikely(prev->state == TASK_DEAD))
3332 preempt_enable_no_resched_notrace();
3334 schedule_debug(prev);
3336 if (sched_feat(HRTICK))
3340 * Make sure that signal_pending_state()->signal_pending() below
3341 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3342 * done by the caller to avoid the race with signal_wake_up().
3344 smp_mb__before_spinlock();
3345 raw_spin_lock_irq(&rq->lock);
3346 lockdep_pin_lock(&rq->lock);
3348 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3350 switch_count = &prev->nivcsw;
3351 if (!preempt && prev->state) {
3352 if (unlikely(signal_pending_state(prev->state, prev))) {
3353 prev->state = TASK_RUNNING;
3355 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3359 * If a worker went to sleep, notify and ask workqueue
3360 * whether it wants to wake up a task to maintain
3363 if (prev->flags & PF_WQ_WORKER) {
3364 struct task_struct *to_wakeup;
3366 to_wakeup = wq_worker_sleeping(prev, cpu);
3368 try_to_wake_up_local(to_wakeup);
3371 switch_count = &prev->nvcsw;
3374 if (task_on_rq_queued(prev))
3375 update_rq_clock(rq);
3377 next = pick_next_task(rq, prev);
3378 wallclock = walt_ktime_clock();
3379 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3380 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3381 clear_tsk_need_resched(prev);
3382 clear_preempt_need_resched();
3383 rq->clock_skip_update = 0;
3385 if (likely(prev != next)) {
3386 #ifdef CONFIG_SCHED_WALT
3388 prev->last_sleep_ts = wallclock;
3394 trace_sched_switch(preempt, prev, next);
3395 rq = context_switch(rq, prev, next); /* unlocks the rq */
3398 lockdep_unpin_lock(&rq->lock);
3399 raw_spin_unlock_irq(&rq->lock);
3402 balance_callback(rq);
3405 static inline void sched_submit_work(struct task_struct *tsk)
3407 if (!tsk->state || tsk_is_pi_blocked(tsk))
3410 * If we are going to sleep and we have plugged IO queued,
3411 * make sure to submit it to avoid deadlocks.
3413 if (blk_needs_flush_plug(tsk))
3414 blk_schedule_flush_plug(tsk);
3417 asmlinkage __visible void __sched schedule(void)
3419 struct task_struct *tsk = current;
3421 sched_submit_work(tsk);
3425 sched_preempt_enable_no_resched();
3426 } while (need_resched());
3428 EXPORT_SYMBOL(schedule);
3430 #ifdef CONFIG_CONTEXT_TRACKING
3431 asmlinkage __visible void __sched schedule_user(void)
3434 * If we come here after a random call to set_need_resched(),
3435 * or we have been woken up remotely but the IPI has not yet arrived,
3436 * we haven't yet exited the RCU idle mode. Do it here manually until
3437 * we find a better solution.
3439 * NB: There are buggy callers of this function. Ideally we
3440 * should warn if prev_state != CONTEXT_USER, but that will trigger
3441 * too frequently to make sense yet.
3443 enum ctx_state prev_state = exception_enter();
3445 exception_exit(prev_state);
3450 * schedule_preempt_disabled - called with preemption disabled
3452 * Returns with preemption disabled. Note: preempt_count must be 1
3454 void __sched schedule_preempt_disabled(void)
3456 sched_preempt_enable_no_resched();
3461 static void __sched notrace preempt_schedule_common(void)
3464 preempt_disable_notrace();
3466 preempt_enable_no_resched_notrace();
3469 * Check again in case we missed a preemption opportunity
3470 * between schedule and now.
3472 } while (need_resched());
3475 #ifdef CONFIG_PREEMPT
3477 * this is the entry point to schedule() from in-kernel preemption
3478 * off of preempt_enable. Kernel preemptions off return from interrupt
3479 * occur there and call schedule directly.
3481 asmlinkage __visible void __sched notrace preempt_schedule(void)
3484 * If there is a non-zero preempt_count or interrupts are disabled,
3485 * we do not want to preempt the current task. Just return..
3487 if (likely(!preemptible()))
3490 preempt_schedule_common();
3492 NOKPROBE_SYMBOL(preempt_schedule);
3493 EXPORT_SYMBOL(preempt_schedule);
3496 * preempt_schedule_notrace - preempt_schedule called by tracing
3498 * The tracing infrastructure uses preempt_enable_notrace to prevent
3499 * recursion and tracing preempt enabling caused by the tracing
3500 * infrastructure itself. But as tracing can happen in areas coming
3501 * from userspace or just about to enter userspace, a preempt enable
3502 * can occur before user_exit() is called. This will cause the scheduler
3503 * to be called when the system is still in usermode.
3505 * To prevent this, the preempt_enable_notrace will use this function
3506 * instead of preempt_schedule() to exit user context if needed before
3507 * calling the scheduler.
3509 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3511 enum ctx_state prev_ctx;
3513 if (likely(!preemptible()))
3517 preempt_disable_notrace();
3519 * Needs preempt disabled in case user_exit() is traced
3520 * and the tracer calls preempt_enable_notrace() causing
3521 * an infinite recursion.
3523 prev_ctx = exception_enter();
3525 exception_exit(prev_ctx);
3527 preempt_enable_no_resched_notrace();
3528 } while (need_resched());
3530 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3532 #endif /* CONFIG_PREEMPT */
3535 * this is the entry point to schedule() from kernel preemption
3536 * off of irq context.
3537 * Note, that this is called and return with irqs disabled. This will
3538 * protect us against recursive calling from irq.
3540 asmlinkage __visible void __sched preempt_schedule_irq(void)
3542 enum ctx_state prev_state;
3544 /* Catch callers which need to be fixed */
3545 BUG_ON(preempt_count() || !irqs_disabled());
3547 prev_state = exception_enter();
3553 local_irq_disable();
3554 sched_preempt_enable_no_resched();
3555 } while (need_resched());
3557 exception_exit(prev_state);
3560 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3563 return try_to_wake_up(curr->private, mode, wake_flags);
3565 EXPORT_SYMBOL(default_wake_function);
3567 #ifdef CONFIG_RT_MUTEXES
3570 * rt_mutex_setprio - set the current priority of a task
3572 * @prio: prio value (kernel-internal form)
3574 * This function changes the 'effective' priority of a task. It does
3575 * not touch ->normal_prio like __setscheduler().
3577 * Used by the rt_mutex code to implement priority inheritance
3578 * logic. Call site only calls if the priority of the task changed.
3580 void rt_mutex_setprio(struct task_struct *p, int prio)
3582 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3584 const struct sched_class *prev_class;
3586 BUG_ON(prio > MAX_PRIO);
3588 rq = __task_rq_lock(p);
3589 update_rq_clock(rq);
3592 * Idle task boosting is a nono in general. There is one
3593 * exception, when PREEMPT_RT and NOHZ is active:
3595 * The idle task calls get_next_timer_interrupt() and holds
3596 * the timer wheel base->lock on the CPU and another CPU wants
3597 * to access the timer (probably to cancel it). We can safely
3598 * ignore the boosting request, as the idle CPU runs this code
3599 * with interrupts disabled and will complete the lock
3600 * protected section without being interrupted. So there is no
3601 * real need to boost.
3603 if (unlikely(p == rq->idle)) {
3604 WARN_ON(p != rq->curr);
3605 WARN_ON(p->pi_blocked_on);
3609 trace_sched_pi_setprio(p, prio);
3611 prev_class = p->sched_class;
3612 queued = task_on_rq_queued(p);
3613 running = task_current(rq, p);
3615 dequeue_task(rq, p, DEQUEUE_SAVE);
3617 put_prev_task(rq, p);
3620 * Boosting condition are:
3621 * 1. -rt task is running and holds mutex A
3622 * --> -dl task blocks on mutex A
3624 * 2. -dl task is running and holds mutex A
3625 * --> -dl task blocks on mutex A and could preempt the
3628 if (dl_prio(prio)) {
3629 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3630 if (!dl_prio(p->normal_prio) ||
3631 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3632 p->dl.dl_boosted = 1;
3633 enqueue_flag |= ENQUEUE_REPLENISH;
3635 p->dl.dl_boosted = 0;
3636 p->sched_class = &dl_sched_class;
3637 } else if (rt_prio(prio)) {
3638 if (dl_prio(oldprio))
3639 p->dl.dl_boosted = 0;
3641 enqueue_flag |= ENQUEUE_HEAD;
3642 p->sched_class = &rt_sched_class;
3644 if (dl_prio(oldprio))
3645 p->dl.dl_boosted = 0;
3646 if (rt_prio(oldprio))
3648 p->sched_class = &fair_sched_class;
3654 p->sched_class->set_curr_task(rq);
3656 enqueue_task(rq, p, enqueue_flag);
3658 check_class_changed(rq, p, prev_class, oldprio);
3660 preempt_disable(); /* avoid rq from going away on us */
3661 __task_rq_unlock(rq);
3663 balance_callback(rq);
3668 void set_user_nice(struct task_struct *p, long nice)
3670 int old_prio, delta, queued;
3671 unsigned long flags;
3674 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3677 * We have to be careful, if called from sys_setpriority(),
3678 * the task might be in the middle of scheduling on another CPU.
3680 rq = task_rq_lock(p, &flags);
3681 update_rq_clock(rq);
3684 * The RT priorities are set via sched_setscheduler(), but we still
3685 * allow the 'normal' nice value to be set - but as expected
3686 * it wont have any effect on scheduling until the task is
3687 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3689 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3690 p->static_prio = NICE_TO_PRIO(nice);
3693 queued = task_on_rq_queued(p);
3695 dequeue_task(rq, p, DEQUEUE_SAVE);
3697 p->static_prio = NICE_TO_PRIO(nice);
3700 p->prio = effective_prio(p);
3701 delta = p->prio - old_prio;
3704 enqueue_task(rq, p, ENQUEUE_RESTORE);
3706 * If the task increased its priority or is running and
3707 * lowered its priority, then reschedule its CPU:
3709 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3713 task_rq_unlock(rq, p, &flags);
3715 EXPORT_SYMBOL(set_user_nice);
3718 * can_nice - check if a task can reduce its nice value
3722 int can_nice(const struct task_struct *p, const int nice)
3724 /* convert nice value [19,-20] to rlimit style value [1,40] */
3725 int nice_rlim = nice_to_rlimit(nice);
3727 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3728 capable(CAP_SYS_NICE));
3731 #ifdef __ARCH_WANT_SYS_NICE
3734 * sys_nice - change the priority of the current process.
3735 * @increment: priority increment
3737 * sys_setpriority is a more generic, but much slower function that
3738 * does similar things.
3740 SYSCALL_DEFINE1(nice, int, increment)
3745 * Setpriority might change our priority at the same moment.
3746 * We don't have to worry. Conceptually one call occurs first
3747 * and we have a single winner.
3749 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3750 nice = task_nice(current) + increment;
3752 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3753 if (increment < 0 && !can_nice(current, nice))
3756 retval = security_task_setnice(current, nice);
3760 set_user_nice(current, nice);
3767 * task_prio - return the priority value of a given task.
3768 * @p: the task in question.
3770 * Return: The priority value as seen by users in /proc.
3771 * RT tasks are offset by -200. Normal tasks are centered
3772 * around 0, value goes from -16 to +15.
3774 int task_prio(const struct task_struct *p)
3776 return p->prio - MAX_RT_PRIO;
3780 * idle_cpu - is a given cpu idle currently?
3781 * @cpu: the processor in question.
3783 * Return: 1 if the CPU is currently idle. 0 otherwise.
3785 int idle_cpu(int cpu)
3787 struct rq *rq = cpu_rq(cpu);
3789 if (rq->curr != rq->idle)
3796 if (!llist_empty(&rq->wake_list))
3804 * idle_task - return the idle task for a given cpu.
3805 * @cpu: the processor in question.
3807 * Return: The idle task for the cpu @cpu.
3809 struct task_struct *idle_task(int cpu)
3811 return cpu_rq(cpu)->idle;
3815 * find_process_by_pid - find a process with a matching PID value.
3816 * @pid: the pid in question.
3818 * The task of @pid, if found. %NULL otherwise.
3820 static struct task_struct *find_process_by_pid(pid_t pid)
3822 return pid ? find_task_by_vpid(pid) : current;
3826 * This function initializes the sched_dl_entity of a newly becoming
3827 * SCHED_DEADLINE task.
3829 * Only the static values are considered here, the actual runtime and the
3830 * absolute deadline will be properly calculated when the task is enqueued
3831 * for the first time with its new policy.
3834 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3836 struct sched_dl_entity *dl_se = &p->dl;
3838 dl_se->dl_runtime = attr->sched_runtime;
3839 dl_se->dl_deadline = attr->sched_deadline;
3840 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3841 dl_se->flags = attr->sched_flags;
3842 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3845 * Changing the parameters of a task is 'tricky' and we're not doing
3846 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3848 * What we SHOULD do is delay the bandwidth release until the 0-lag
3849 * point. This would include retaining the task_struct until that time
3850 * and change dl_overflow() to not immediately decrement the current
3853 * Instead we retain the current runtime/deadline and let the new
3854 * parameters take effect after the current reservation period lapses.
3855 * This is safe (albeit pessimistic) because the 0-lag point is always
3856 * before the current scheduling deadline.
3858 * We can still have temporary overloads because we do not delay the
3859 * change in bandwidth until that time; so admission control is
3860 * not on the safe side. It does however guarantee tasks will never
3861 * consume more than promised.
3866 * sched_setparam() passes in -1 for its policy, to let the functions
3867 * it calls know not to change it.
3869 #define SETPARAM_POLICY -1
3871 static void __setscheduler_params(struct task_struct *p,
3872 const struct sched_attr *attr)
3874 int policy = attr->sched_policy;
3876 if (policy == SETPARAM_POLICY)
3881 if (dl_policy(policy))
3882 __setparam_dl(p, attr);
3883 else if (fair_policy(policy))
3884 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3887 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3888 * !rt_policy. Always setting this ensures that things like
3889 * getparam()/getattr() don't report silly values for !rt tasks.
3891 p->rt_priority = attr->sched_priority;
3892 p->normal_prio = normal_prio(p);
3896 /* Actually do priority change: must hold pi & rq lock. */
3897 static void __setscheduler(struct rq *rq, struct task_struct *p,
3898 const struct sched_attr *attr, bool keep_boost)
3900 __setscheduler_params(p, attr);
3903 * Keep a potential priority boosting if called from
3904 * sched_setscheduler().
3907 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3909 p->prio = normal_prio(p);
3911 if (dl_prio(p->prio))
3912 p->sched_class = &dl_sched_class;
3913 else if (rt_prio(p->prio))
3914 p->sched_class = &rt_sched_class;
3916 p->sched_class = &fair_sched_class;
3920 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3922 struct sched_dl_entity *dl_se = &p->dl;
3924 attr->sched_priority = p->rt_priority;
3925 attr->sched_runtime = dl_se->dl_runtime;
3926 attr->sched_deadline = dl_se->dl_deadline;
3927 attr->sched_period = dl_se->dl_period;
3928 attr->sched_flags = dl_se->flags;
3932 * This function validates the new parameters of a -deadline task.
3933 * We ask for the deadline not being zero, and greater or equal
3934 * than the runtime, as well as the period of being zero or
3935 * greater than deadline. Furthermore, we have to be sure that
3936 * user parameters are above the internal resolution of 1us (we
3937 * check sched_runtime only since it is always the smaller one) and
3938 * below 2^63 ns (we have to check both sched_deadline and
3939 * sched_period, as the latter can be zero).
3942 __checkparam_dl(const struct sched_attr *attr)
3945 if (attr->sched_deadline == 0)
3949 * Since we truncate DL_SCALE bits, make sure we're at least
3952 if (attr->sched_runtime < (1ULL << DL_SCALE))
3956 * Since we use the MSB for wrap-around and sign issues, make
3957 * sure it's not set (mind that period can be equal to zero).
3959 if (attr->sched_deadline & (1ULL << 63) ||
3960 attr->sched_period & (1ULL << 63))
3963 /* runtime <= deadline <= period (if period != 0) */
3964 if ((attr->sched_period != 0 &&
3965 attr->sched_period < attr->sched_deadline) ||
3966 attr->sched_deadline < attr->sched_runtime)
3973 * check the target process has a UID that matches the current process's
3975 static bool check_same_owner(struct task_struct *p)
3977 const struct cred *cred = current_cred(), *pcred;
3981 pcred = __task_cred(p);
3982 match = (uid_eq(cred->euid, pcred->euid) ||
3983 uid_eq(cred->euid, pcred->uid));
3988 static bool dl_param_changed(struct task_struct *p,
3989 const struct sched_attr *attr)
3991 struct sched_dl_entity *dl_se = &p->dl;
3993 if (dl_se->dl_runtime != attr->sched_runtime ||
3994 dl_se->dl_deadline != attr->sched_deadline ||
3995 dl_se->dl_period != attr->sched_period ||
3996 dl_se->flags != attr->sched_flags)
4002 static int __sched_setscheduler(struct task_struct *p,
4003 const struct sched_attr *attr,
4006 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
4007 MAX_RT_PRIO - 1 - attr->sched_priority;
4008 int retval, oldprio, oldpolicy = -1, queued, running;
4009 int new_effective_prio, policy = attr->sched_policy;
4010 unsigned long flags;
4011 const struct sched_class *prev_class;
4015 /* may grab non-irq protected spin_locks */
4016 BUG_ON(in_interrupt());
4018 /* double check policy once rq lock held */
4020 reset_on_fork = p->sched_reset_on_fork;
4021 policy = oldpolicy = p->policy;
4023 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4025 if (!valid_policy(policy))
4029 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4033 * Valid priorities for SCHED_FIFO and SCHED_RR are
4034 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4035 * SCHED_BATCH and SCHED_IDLE is 0.
4037 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4038 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4040 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4041 (rt_policy(policy) != (attr->sched_priority != 0)))
4045 * Allow unprivileged RT tasks to decrease priority:
4047 if (user && !capable(CAP_SYS_NICE)) {
4048 if (fair_policy(policy)) {
4049 if (attr->sched_nice < task_nice(p) &&
4050 !can_nice(p, attr->sched_nice))
4054 if (rt_policy(policy)) {
4055 unsigned long rlim_rtprio =
4056 task_rlimit(p, RLIMIT_RTPRIO);
4058 /* can't set/change the rt policy */
4059 if (policy != p->policy && !rlim_rtprio)
4062 /* can't increase priority */
4063 if (attr->sched_priority > p->rt_priority &&
4064 attr->sched_priority > rlim_rtprio)
4069 * Can't set/change SCHED_DEADLINE policy at all for now
4070 * (safest behavior); in the future we would like to allow
4071 * unprivileged DL tasks to increase their relative deadline
4072 * or reduce their runtime (both ways reducing utilization)
4074 if (dl_policy(policy))
4078 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4079 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4081 if (idle_policy(p->policy) && !idle_policy(policy)) {
4082 if (!can_nice(p, task_nice(p)))
4086 /* can't change other user's priorities */
4087 if (!check_same_owner(p))
4090 /* Normal users shall not reset the sched_reset_on_fork flag */
4091 if (p->sched_reset_on_fork && !reset_on_fork)
4096 retval = security_task_setscheduler(p);
4102 * make sure no PI-waiters arrive (or leave) while we are
4103 * changing the priority of the task:
4105 * To be able to change p->policy safely, the appropriate
4106 * runqueue lock must be held.
4108 rq = task_rq_lock(p, &flags);
4109 update_rq_clock(rq);
4112 * Changing the policy of the stop threads its a very bad idea
4114 if (p == rq->stop) {
4115 task_rq_unlock(rq, p, &flags);
4120 * If not changing anything there's no need to proceed further,
4121 * but store a possible modification of reset_on_fork.
4123 if (unlikely(policy == p->policy)) {
4124 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4126 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4128 if (dl_policy(policy) && dl_param_changed(p, attr))
4131 p->sched_reset_on_fork = reset_on_fork;
4132 task_rq_unlock(rq, p, &flags);
4138 #ifdef CONFIG_RT_GROUP_SCHED
4140 * Do not allow realtime tasks into groups that have no runtime
4143 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4144 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4145 !task_group_is_autogroup(task_group(p))) {
4146 task_rq_unlock(rq, p, &flags);
4151 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4152 cpumask_t *span = rq->rd->span;
4155 * Don't allow tasks with an affinity mask smaller than
4156 * the entire root_domain to become SCHED_DEADLINE. We
4157 * will also fail if there's no bandwidth available.
4159 if (!cpumask_subset(span, &p->cpus_allowed) ||
4160 rq->rd->dl_bw.bw == 0) {
4161 task_rq_unlock(rq, p, &flags);
4168 /* recheck policy now with rq lock held */
4169 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4170 policy = oldpolicy = -1;
4171 task_rq_unlock(rq, p, &flags);
4176 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4177 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4180 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4181 task_rq_unlock(rq, p, &flags);
4185 p->sched_reset_on_fork = reset_on_fork;
4190 * Take priority boosted tasks into account. If the new
4191 * effective priority is unchanged, we just store the new
4192 * normal parameters and do not touch the scheduler class and
4193 * the runqueue. This will be done when the task deboost
4196 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4197 if (new_effective_prio == oldprio) {
4198 __setscheduler_params(p, attr);
4199 task_rq_unlock(rq, p, &flags);
4204 queued = task_on_rq_queued(p);
4205 running = task_current(rq, p);
4207 dequeue_task(rq, p, DEQUEUE_SAVE);
4209 put_prev_task(rq, p);
4211 prev_class = p->sched_class;
4212 __setscheduler(rq, p, attr, pi);
4215 p->sched_class->set_curr_task(rq);
4217 int enqueue_flags = ENQUEUE_RESTORE;
4219 * We enqueue to tail when the priority of a task is
4220 * increased (user space view).
4222 if (oldprio <= p->prio)
4223 enqueue_flags |= ENQUEUE_HEAD;
4225 enqueue_task(rq, p, enqueue_flags);
4228 check_class_changed(rq, p, prev_class, oldprio);
4229 preempt_disable(); /* avoid rq from going away on us */
4230 task_rq_unlock(rq, p, &flags);
4233 rt_mutex_adjust_pi(p);
4236 * Run balance callbacks after we've adjusted the PI chain.
4238 balance_callback(rq);
4244 static int _sched_setscheduler(struct task_struct *p, int policy,
4245 const struct sched_param *param, bool check)
4247 struct sched_attr attr = {
4248 .sched_policy = policy,
4249 .sched_priority = param->sched_priority,
4250 .sched_nice = PRIO_TO_NICE(p->static_prio),
4253 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4254 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4255 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4256 policy &= ~SCHED_RESET_ON_FORK;
4257 attr.sched_policy = policy;
4260 return __sched_setscheduler(p, &attr, check, true);
4263 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4264 * @p: the task in question.
4265 * @policy: new policy.
4266 * @param: structure containing the new RT priority.
4268 * Return: 0 on success. An error code otherwise.
4270 * NOTE that the task may be already dead.
4272 int sched_setscheduler(struct task_struct *p, int policy,
4273 const struct sched_param *param)
4275 return _sched_setscheduler(p, policy, param, true);
4277 EXPORT_SYMBOL_GPL(sched_setscheduler);
4279 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4281 return __sched_setscheduler(p, attr, true, true);
4283 EXPORT_SYMBOL_GPL(sched_setattr);
4286 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4287 * @p: the task in question.
4288 * @policy: new policy.
4289 * @param: structure containing the new RT priority.
4291 * Just like sched_setscheduler, only don't bother checking if the
4292 * current context has permission. For example, this is needed in
4293 * stop_machine(): we create temporary high priority worker threads,
4294 * but our caller might not have that capability.
4296 * Return: 0 on success. An error code otherwise.
4298 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4299 const struct sched_param *param)
4301 return _sched_setscheduler(p, policy, param, false);
4303 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4306 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4308 struct sched_param lparam;
4309 struct task_struct *p;
4312 if (!param || pid < 0)
4314 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4319 p = find_process_by_pid(pid);
4321 retval = sched_setscheduler(p, policy, &lparam);
4328 * Mimics kernel/events/core.c perf_copy_attr().
4330 static int sched_copy_attr(struct sched_attr __user *uattr,
4331 struct sched_attr *attr)
4336 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4340 * zero the full structure, so that a short copy will be nice.
4342 memset(attr, 0, sizeof(*attr));
4344 ret = get_user(size, &uattr->size);
4348 if (size > PAGE_SIZE) /* silly large */
4351 if (!size) /* abi compat */
4352 size = SCHED_ATTR_SIZE_VER0;
4354 if (size < SCHED_ATTR_SIZE_VER0)
4358 * If we're handed a bigger struct than we know of,
4359 * ensure all the unknown bits are 0 - i.e. new
4360 * user-space does not rely on any kernel feature
4361 * extensions we dont know about yet.
4363 if (size > sizeof(*attr)) {
4364 unsigned char __user *addr;
4365 unsigned char __user *end;
4368 addr = (void __user *)uattr + sizeof(*attr);
4369 end = (void __user *)uattr + size;
4371 for (; addr < end; addr++) {
4372 ret = get_user(val, addr);
4378 size = sizeof(*attr);
4381 ret = copy_from_user(attr, uattr, size);
4386 * XXX: do we want to be lenient like existing syscalls; or do we want
4387 * to be strict and return an error on out-of-bounds values?
4389 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4394 put_user(sizeof(*attr), &uattr->size);
4399 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4400 * @pid: the pid in question.
4401 * @policy: new policy.
4402 * @param: structure containing the new RT priority.
4404 * Return: 0 on success. An error code otherwise.
4406 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4407 struct sched_param __user *, param)
4409 /* negative values for policy are not valid */
4413 return do_sched_setscheduler(pid, policy, param);
4417 * sys_sched_setparam - set/change the RT priority of a thread
4418 * @pid: the pid in question.
4419 * @param: structure containing the new RT priority.
4421 * Return: 0 on success. An error code otherwise.
4423 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4425 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4429 * sys_sched_setattr - same as above, but with extended sched_attr
4430 * @pid: the pid in question.
4431 * @uattr: structure containing the extended parameters.
4432 * @flags: for future extension.
4434 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4435 unsigned int, flags)
4437 struct sched_attr attr;
4438 struct task_struct *p;
4441 if (!uattr || pid < 0 || flags)
4444 retval = sched_copy_attr(uattr, &attr);
4448 if ((int)attr.sched_policy < 0)
4453 p = find_process_by_pid(pid);
4455 retval = sched_setattr(p, &attr);
4462 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4463 * @pid: the pid in question.
4465 * Return: On success, the policy of the thread. Otherwise, a negative error
4468 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4470 struct task_struct *p;
4478 p = find_process_by_pid(pid);
4480 retval = security_task_getscheduler(p);
4483 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4490 * sys_sched_getparam - get the RT priority of a thread
4491 * @pid: the pid in question.
4492 * @param: structure containing the RT priority.
4494 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4497 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4499 struct sched_param lp = { .sched_priority = 0 };
4500 struct task_struct *p;
4503 if (!param || pid < 0)
4507 p = find_process_by_pid(pid);
4512 retval = security_task_getscheduler(p);
4516 if (task_has_rt_policy(p))
4517 lp.sched_priority = p->rt_priority;
4521 * This one might sleep, we cannot do it with a spinlock held ...
4523 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4532 static int sched_read_attr(struct sched_attr __user *uattr,
4533 struct sched_attr *attr,
4538 if (!access_ok(VERIFY_WRITE, uattr, usize))
4542 * If we're handed a smaller struct than we know of,
4543 * ensure all the unknown bits are 0 - i.e. old
4544 * user-space does not get uncomplete information.
4546 if (usize < sizeof(*attr)) {
4547 unsigned char *addr;
4550 addr = (void *)attr + usize;
4551 end = (void *)attr + sizeof(*attr);
4553 for (; addr < end; addr++) {
4561 ret = copy_to_user(uattr, attr, attr->size);
4569 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4570 * @pid: the pid in question.
4571 * @uattr: structure containing the extended parameters.
4572 * @size: sizeof(attr) for fwd/bwd comp.
4573 * @flags: for future extension.
4575 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4576 unsigned int, size, unsigned int, flags)
4578 struct sched_attr attr = {
4579 .size = sizeof(struct sched_attr),
4581 struct task_struct *p;
4584 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4585 size < SCHED_ATTR_SIZE_VER0 || flags)
4589 p = find_process_by_pid(pid);
4594 retval = security_task_getscheduler(p);
4598 attr.sched_policy = p->policy;
4599 if (p->sched_reset_on_fork)
4600 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4601 if (task_has_dl_policy(p))
4602 __getparam_dl(p, &attr);
4603 else if (task_has_rt_policy(p))
4604 attr.sched_priority = p->rt_priority;
4606 attr.sched_nice = task_nice(p);
4610 retval = sched_read_attr(uattr, &attr, size);
4618 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4620 cpumask_var_t cpus_allowed, new_mask;
4621 struct task_struct *p;
4626 p = find_process_by_pid(pid);
4632 /* Prevent p going away */
4636 if (p->flags & PF_NO_SETAFFINITY) {
4640 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4644 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4646 goto out_free_cpus_allowed;
4649 if (!check_same_owner(p)) {
4651 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4653 goto out_free_new_mask;
4658 retval = security_task_setscheduler(p);
4660 goto out_free_new_mask;
4663 cpuset_cpus_allowed(p, cpus_allowed);
4664 cpumask_and(new_mask, in_mask, cpus_allowed);
4667 * Since bandwidth control happens on root_domain basis,
4668 * if admission test is enabled, we only admit -deadline
4669 * tasks allowed to run on all the CPUs in the task's
4673 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4675 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4678 goto out_free_new_mask;
4684 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4687 cpuset_cpus_allowed(p, cpus_allowed);
4688 if (!cpumask_subset(new_mask, cpus_allowed)) {
4690 * We must have raced with a concurrent cpuset
4691 * update. Just reset the cpus_allowed to the
4692 * cpuset's cpus_allowed
4694 cpumask_copy(new_mask, cpus_allowed);
4699 free_cpumask_var(new_mask);
4700 out_free_cpus_allowed:
4701 free_cpumask_var(cpus_allowed);
4707 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4708 struct cpumask *new_mask)
4710 if (len < cpumask_size())
4711 cpumask_clear(new_mask);
4712 else if (len > cpumask_size())
4713 len = cpumask_size();
4715 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4719 * sys_sched_setaffinity - set the cpu affinity of a process
4720 * @pid: pid of the process
4721 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4722 * @user_mask_ptr: user-space pointer to the new cpu mask
4724 * Return: 0 on success. An error code otherwise.
4726 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4727 unsigned long __user *, user_mask_ptr)
4729 cpumask_var_t new_mask;
4732 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4735 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4737 retval = sched_setaffinity(pid, new_mask);
4738 free_cpumask_var(new_mask);
4742 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4744 struct task_struct *p;
4745 unsigned long flags;
4751 p = find_process_by_pid(pid);
4755 retval = security_task_getscheduler(p);
4759 raw_spin_lock_irqsave(&p->pi_lock, flags);
4760 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4761 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4770 * sys_sched_getaffinity - get the cpu affinity of a process
4771 * @pid: pid of the process
4772 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4773 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4775 * Return: 0 on success. An error code otherwise.
4777 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4778 unsigned long __user *, user_mask_ptr)
4783 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4785 if (len & (sizeof(unsigned long)-1))
4788 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4791 ret = sched_getaffinity(pid, mask);
4793 size_t retlen = min_t(size_t, len, cpumask_size());
4795 if (copy_to_user(user_mask_ptr, mask, retlen))
4800 free_cpumask_var(mask);
4806 * sys_sched_yield - yield the current processor to other threads.
4808 * This function yields the current CPU to other tasks. If there are no
4809 * other threads running on this CPU then this function will return.
4813 SYSCALL_DEFINE0(sched_yield)
4815 struct rq *rq = this_rq_lock();
4817 schedstat_inc(rq, yld_count);
4818 current->sched_class->yield_task(rq);
4821 * Since we are going to call schedule() anyway, there's
4822 * no need to preempt or enable interrupts:
4824 __release(rq->lock);
4825 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4826 do_raw_spin_unlock(&rq->lock);
4827 sched_preempt_enable_no_resched();
4834 int __sched _cond_resched(void)
4836 if (should_resched(0)) {
4837 preempt_schedule_common();
4842 EXPORT_SYMBOL(_cond_resched);
4845 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4846 * call schedule, and on return reacquire the lock.
4848 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4849 * operations here to prevent schedule() from being called twice (once via
4850 * spin_unlock(), once by hand).
4852 int __cond_resched_lock(spinlock_t *lock)
4854 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4857 lockdep_assert_held(lock);
4859 if (spin_needbreak(lock) || resched) {
4862 preempt_schedule_common();
4870 EXPORT_SYMBOL(__cond_resched_lock);
4872 int __sched __cond_resched_softirq(void)
4874 BUG_ON(!in_softirq());
4876 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4878 preempt_schedule_common();
4884 EXPORT_SYMBOL(__cond_resched_softirq);
4887 * yield - yield the current processor to other threads.
4889 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4891 * The scheduler is at all times free to pick the calling task as the most
4892 * eligible task to run, if removing the yield() call from your code breaks
4893 * it, its already broken.
4895 * Typical broken usage is:
4900 * where one assumes that yield() will let 'the other' process run that will
4901 * make event true. If the current task is a SCHED_FIFO task that will never
4902 * happen. Never use yield() as a progress guarantee!!
4904 * If you want to use yield() to wait for something, use wait_event().
4905 * If you want to use yield() to be 'nice' for others, use cond_resched().
4906 * If you still want to use yield(), do not!
4908 void __sched yield(void)
4910 set_current_state(TASK_RUNNING);
4913 EXPORT_SYMBOL(yield);
4916 * yield_to - yield the current processor to another thread in
4917 * your thread group, or accelerate that thread toward the
4918 * processor it's on.
4920 * @preempt: whether task preemption is allowed or not
4922 * It's the caller's job to ensure that the target task struct
4923 * can't go away on us before we can do any checks.
4926 * true (>0) if we indeed boosted the target task.
4927 * false (0) if we failed to boost the target.
4928 * -ESRCH if there's no task to yield to.
4930 int __sched yield_to(struct task_struct *p, bool preempt)
4932 struct task_struct *curr = current;
4933 struct rq *rq, *p_rq;
4934 unsigned long flags;
4937 local_irq_save(flags);
4943 * If we're the only runnable task on the rq and target rq also
4944 * has only one task, there's absolutely no point in yielding.
4946 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4951 double_rq_lock(rq, p_rq);
4952 if (task_rq(p) != p_rq) {
4953 double_rq_unlock(rq, p_rq);
4957 if (!curr->sched_class->yield_to_task)
4960 if (curr->sched_class != p->sched_class)
4963 if (task_running(p_rq, p) || p->state)
4966 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4968 schedstat_inc(rq, yld_count);
4970 * Make p's CPU reschedule; pick_next_entity takes care of
4973 if (preempt && rq != p_rq)
4978 double_rq_unlock(rq, p_rq);
4980 local_irq_restore(flags);
4987 EXPORT_SYMBOL_GPL(yield_to);
4990 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4991 * that process accounting knows that this is a task in IO wait state.
4993 long __sched io_schedule_timeout(long timeout)
4995 int old_iowait = current->in_iowait;
4999 current->in_iowait = 1;
5000 blk_schedule_flush_plug(current);
5002 delayacct_blkio_start();
5004 atomic_inc(&rq->nr_iowait);
5005 ret = schedule_timeout(timeout);
5006 current->in_iowait = old_iowait;
5007 atomic_dec(&rq->nr_iowait);
5008 delayacct_blkio_end();
5012 EXPORT_SYMBOL(io_schedule_timeout);
5015 * sys_sched_get_priority_max - return maximum RT priority.
5016 * @policy: scheduling class.
5018 * Return: On success, this syscall returns the maximum
5019 * rt_priority that can be used by a given scheduling class.
5020 * On failure, a negative error code is returned.
5022 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5029 ret = MAX_USER_RT_PRIO-1;
5031 case SCHED_DEADLINE:
5042 * sys_sched_get_priority_min - return minimum RT priority.
5043 * @policy: scheduling class.
5045 * Return: On success, this syscall returns the minimum
5046 * rt_priority that can be used by a given scheduling class.
5047 * On failure, a negative error code is returned.
5049 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5058 case SCHED_DEADLINE:
5068 * sys_sched_rr_get_interval - return the default timeslice of a process.
5069 * @pid: pid of the process.
5070 * @interval: userspace pointer to the timeslice value.
5072 * this syscall writes the default timeslice value of a given process
5073 * into the user-space timespec buffer. A value of '0' means infinity.
5075 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5078 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5079 struct timespec __user *, interval)
5081 struct task_struct *p;
5082 unsigned int time_slice;
5083 unsigned long flags;
5093 p = find_process_by_pid(pid);
5097 retval = security_task_getscheduler(p);
5101 rq = task_rq_lock(p, &flags);
5103 if (p->sched_class->get_rr_interval)
5104 time_slice = p->sched_class->get_rr_interval(rq, p);
5105 task_rq_unlock(rq, p, &flags);
5108 jiffies_to_timespec(time_slice, &t);
5109 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5117 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5119 void sched_show_task(struct task_struct *p)
5121 unsigned long free = 0;
5123 unsigned long state = p->state;
5126 state = __ffs(state) + 1;
5127 printk(KERN_INFO "%-15.15s %c", p->comm,
5128 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5129 #if BITS_PER_LONG == 32
5130 if (state == TASK_RUNNING)
5131 printk(KERN_CONT " running ");
5133 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5135 if (state == TASK_RUNNING)
5136 printk(KERN_CONT " running task ");
5138 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5140 #ifdef CONFIG_DEBUG_STACK_USAGE
5141 free = stack_not_used(p);
5146 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5148 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5149 task_pid_nr(p), ppid,
5150 (unsigned long)task_thread_info(p)->flags);
5152 print_worker_info(KERN_INFO, p);
5153 show_stack(p, NULL);
5156 void show_state_filter(unsigned long state_filter)
5158 struct task_struct *g, *p;
5160 #if BITS_PER_LONG == 32
5162 " task PC stack pid father\n");
5165 " task PC stack pid father\n");
5168 for_each_process_thread(g, p) {
5170 * reset the NMI-timeout, listing all files on a slow
5171 * console might take a lot of time:
5172 * Also, reset softlockup watchdogs on all CPUs, because
5173 * another CPU might be blocked waiting for us to process
5176 touch_nmi_watchdog();
5177 touch_all_softlockup_watchdogs();
5178 if (!state_filter || (p->state & state_filter))
5182 #ifdef CONFIG_SCHED_DEBUG
5183 sysrq_sched_debug_show();
5187 * Only show locks if all tasks are dumped:
5190 debug_show_all_locks();
5193 void init_idle_bootup_task(struct task_struct *idle)
5195 idle->sched_class = &idle_sched_class;
5199 * init_idle - set up an idle thread for a given CPU
5200 * @idle: task in question
5201 * @cpu: cpu the idle task belongs to
5203 * NOTE: this function does not set the idle thread's NEED_RESCHED
5204 * flag, to make booting more robust.
5206 void init_idle(struct task_struct *idle, int cpu)
5208 struct rq *rq = cpu_rq(cpu);
5209 unsigned long flags;
5211 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5212 raw_spin_lock(&rq->lock);
5214 __sched_fork(0, idle);
5216 idle->state = TASK_RUNNING;
5217 idle->se.exec_start = sched_clock();
5221 * Its possible that init_idle() gets called multiple times on a task,
5222 * in that case do_set_cpus_allowed() will not do the right thing.
5224 * And since this is boot we can forgo the serialization.
5226 set_cpus_allowed_common(idle, cpumask_of(cpu));
5229 * We're having a chicken and egg problem, even though we are
5230 * holding rq->lock, the cpu isn't yet set to this cpu so the
5231 * lockdep check in task_group() will fail.
5233 * Similar case to sched_fork(). / Alternatively we could
5234 * use task_rq_lock() here and obtain the other rq->lock.
5239 __set_task_cpu(idle, cpu);
5242 rq->curr = rq->idle = idle;
5243 idle->on_rq = TASK_ON_RQ_QUEUED;
5247 raw_spin_unlock(&rq->lock);
5248 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5250 /* Set the preempt count _outside_ the spinlocks! */
5251 init_idle_preempt_count(idle, cpu);
5254 * The idle tasks have their own, simple scheduling class:
5256 idle->sched_class = &idle_sched_class;
5257 ftrace_graph_init_idle_task(idle, cpu);
5258 vtime_init_idle(idle, cpu);
5260 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5264 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5265 const struct cpumask *trial)
5267 int ret = 1, trial_cpus;
5268 struct dl_bw *cur_dl_b;
5269 unsigned long flags;
5271 if (!cpumask_weight(cur))
5274 rcu_read_lock_sched();
5275 cur_dl_b = dl_bw_of(cpumask_any(cur));
5276 trial_cpus = cpumask_weight(trial);
5278 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5279 if (cur_dl_b->bw != -1 &&
5280 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5282 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5283 rcu_read_unlock_sched();
5288 int task_can_attach(struct task_struct *p,
5289 const struct cpumask *cs_cpus_allowed)
5294 * Kthreads which disallow setaffinity shouldn't be moved
5295 * to a new cpuset; we don't want to change their cpu
5296 * affinity and isolating such threads by their set of
5297 * allowed nodes is unnecessary. Thus, cpusets are not
5298 * applicable for such threads. This prevents checking for
5299 * success of set_cpus_allowed_ptr() on all attached tasks
5300 * before cpus_allowed may be changed.
5302 if (p->flags & PF_NO_SETAFFINITY) {
5308 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5310 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5315 unsigned long flags;
5317 rcu_read_lock_sched();
5318 dl_b = dl_bw_of(dest_cpu);
5319 raw_spin_lock_irqsave(&dl_b->lock, flags);
5320 cpus = dl_bw_cpus(dest_cpu);
5321 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5326 * We reserve space for this task in the destination
5327 * root_domain, as we can't fail after this point.
5328 * We will free resources in the source root_domain
5329 * later on (see set_cpus_allowed_dl()).
5331 __dl_add(dl_b, p->dl.dl_bw);
5333 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5334 rcu_read_unlock_sched();
5344 #ifdef CONFIG_NUMA_BALANCING
5345 /* Migrate current task p to target_cpu */
5346 int migrate_task_to(struct task_struct *p, int target_cpu)
5348 struct migration_arg arg = { p, target_cpu };
5349 int curr_cpu = task_cpu(p);
5351 if (curr_cpu == target_cpu)
5354 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5357 /* TODO: This is not properly updating schedstats */
5359 trace_sched_move_numa(p, curr_cpu, target_cpu);
5360 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5364 * Requeue a task on a given node and accurately track the number of NUMA
5365 * tasks on the runqueues
5367 void sched_setnuma(struct task_struct *p, int nid)
5370 unsigned long flags;
5371 bool queued, running;
5373 rq = task_rq_lock(p, &flags);
5374 queued = task_on_rq_queued(p);
5375 running = task_current(rq, p);
5378 dequeue_task(rq, p, DEQUEUE_SAVE);
5380 put_prev_task(rq, p);
5382 p->numa_preferred_nid = nid;
5385 p->sched_class->set_curr_task(rq);
5387 enqueue_task(rq, p, ENQUEUE_RESTORE);
5388 task_rq_unlock(rq, p, &flags);
5390 #endif /* CONFIG_NUMA_BALANCING */
5392 #ifdef CONFIG_HOTPLUG_CPU
5394 * Ensures that the idle task is using init_mm right before its cpu goes
5397 void idle_task_exit(void)
5399 struct mm_struct *mm = current->active_mm;
5401 BUG_ON(cpu_online(smp_processor_id()));
5403 if (mm != &init_mm) {
5404 switch_mm(mm, &init_mm, current);
5405 finish_arch_post_lock_switch();
5411 * Since this CPU is going 'away' for a while, fold any nr_active delta
5412 * we might have. Assumes we're called after migrate_tasks() so that the
5413 * nr_active count is stable.
5415 * Also see the comment "Global load-average calculations".
5417 static void calc_load_migrate(struct rq *rq)
5419 long delta = calc_load_fold_active(rq);
5421 atomic_long_add(delta, &calc_load_tasks);
5424 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5428 static const struct sched_class fake_sched_class = {
5429 .put_prev_task = put_prev_task_fake,
5432 static struct task_struct fake_task = {
5434 * Avoid pull_{rt,dl}_task()
5436 .prio = MAX_PRIO + 1,
5437 .sched_class = &fake_sched_class,
5441 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5442 * try_to_wake_up()->select_task_rq().
5444 * Called with rq->lock held even though we'er in stop_machine() and
5445 * there's no concurrency possible, we hold the required locks anyway
5446 * because of lock validation efforts.
5448 static void migrate_tasks(struct rq *dead_rq)
5450 struct rq *rq = dead_rq;
5451 struct task_struct *next, *stop = rq->stop;
5455 * Fudge the rq selection such that the below task selection loop
5456 * doesn't get stuck on the currently eligible stop task.
5458 * We're currently inside stop_machine() and the rq is either stuck
5459 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5460 * either way we should never end up calling schedule() until we're
5466 * put_prev_task() and pick_next_task() sched
5467 * class method both need to have an up-to-date
5468 * value of rq->clock[_task]
5470 update_rq_clock(rq);
5474 * There's this thread running, bail when that's the only
5477 if (rq->nr_running == 1)
5481 * pick_next_task assumes pinned rq->lock.
5483 lockdep_pin_lock(&rq->lock);
5484 next = pick_next_task(rq, &fake_task);
5486 next->sched_class->put_prev_task(rq, next);
5489 * Rules for changing task_struct::cpus_allowed are holding
5490 * both pi_lock and rq->lock, such that holding either
5491 * stabilizes the mask.
5493 * Drop rq->lock is not quite as disastrous as it usually is
5494 * because !cpu_active at this point, which means load-balance
5495 * will not interfere. Also, stop-machine.
5497 lockdep_unpin_lock(&rq->lock);
5498 raw_spin_unlock(&rq->lock);
5499 raw_spin_lock(&next->pi_lock);
5500 raw_spin_lock(&rq->lock);
5503 * Since we're inside stop-machine, _nothing_ should have
5504 * changed the task, WARN if weird stuff happened, because in
5505 * that case the above rq->lock drop is a fail too.
5507 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5508 raw_spin_unlock(&next->pi_lock);
5512 /* Find suitable destination for @next, with force if needed. */
5513 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5515 rq = __migrate_task(rq, next, dest_cpu);
5516 if (rq != dead_rq) {
5517 raw_spin_unlock(&rq->lock);
5519 raw_spin_lock(&rq->lock);
5521 raw_spin_unlock(&next->pi_lock);
5526 #endif /* CONFIG_HOTPLUG_CPU */
5528 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5530 static struct ctl_table sd_ctl_dir[] = {
5532 .procname = "sched_domain",
5538 static struct ctl_table sd_ctl_root[] = {
5540 .procname = "kernel",
5542 .child = sd_ctl_dir,
5547 static struct ctl_table *sd_alloc_ctl_entry(int n)
5549 struct ctl_table *entry =
5550 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5555 static void sd_free_ctl_entry(struct ctl_table **tablep)
5557 struct ctl_table *entry;
5560 * In the intermediate directories, both the child directory and
5561 * procname are dynamically allocated and could fail but the mode
5562 * will always be set. In the lowest directory the names are
5563 * static strings and all have proc handlers.
5565 for (entry = *tablep; entry->mode; entry++) {
5567 sd_free_ctl_entry(&entry->child);
5568 if (entry->proc_handler == NULL)
5569 kfree(entry->procname);
5576 static int min_load_idx = 0;
5577 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5580 set_table_entry(struct ctl_table *entry,
5581 const char *procname, void *data, int maxlen,
5582 umode_t mode, proc_handler *proc_handler,
5585 entry->procname = procname;
5587 entry->maxlen = maxlen;
5589 entry->proc_handler = proc_handler;
5592 entry->extra1 = &min_load_idx;
5593 entry->extra2 = &max_load_idx;
5597 static struct ctl_table *
5598 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5600 struct ctl_table *table = sd_alloc_ctl_entry(5);
5605 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5606 sizeof(int), 0644, proc_dointvec_minmax, false);
5607 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5608 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5609 proc_doulongvec_minmax, false);
5610 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5611 sizeof(int), 0644, proc_dointvec_minmax, false);
5612 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5613 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5614 proc_doulongvec_minmax, false);
5619 static struct ctl_table *
5620 sd_alloc_ctl_group_table(struct sched_group *sg)
5622 struct ctl_table *table = sd_alloc_ctl_entry(2);
5627 table->procname = kstrdup("energy", GFP_KERNEL);
5629 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5634 static struct ctl_table *
5635 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5637 struct ctl_table *table;
5638 unsigned int nr_entries = 14;
5641 struct sched_group *sg = sd->groups;
5646 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5648 nr_entries += nr_sgs;
5651 table = sd_alloc_ctl_entry(nr_entries);
5656 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5657 sizeof(long), 0644, proc_doulongvec_minmax, false);
5658 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5659 sizeof(long), 0644, proc_doulongvec_minmax, false);
5660 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5661 sizeof(int), 0644, proc_dointvec_minmax, true);
5662 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5663 sizeof(int), 0644, proc_dointvec_minmax, true);
5664 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5665 sizeof(int), 0644, proc_dointvec_minmax, true);
5666 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5667 sizeof(int), 0644, proc_dointvec_minmax, true);
5668 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5669 sizeof(int), 0644, proc_dointvec_minmax, true);
5670 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5671 sizeof(int), 0644, proc_dointvec_minmax, false);
5672 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5673 sizeof(int), 0644, proc_dointvec_minmax, false);
5674 set_table_entry(&table[9], "cache_nice_tries",
5675 &sd->cache_nice_tries,
5676 sizeof(int), 0644, proc_dointvec_minmax, false);
5677 set_table_entry(&table[10], "flags", &sd->flags,
5678 sizeof(int), 0644, proc_dointvec_minmax, false);
5679 set_table_entry(&table[11], "max_newidle_lb_cost",
5680 &sd->max_newidle_lb_cost,
5681 sizeof(long), 0644, proc_doulongvec_minmax, false);
5682 set_table_entry(&table[12], "name", sd->name,
5683 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5687 struct ctl_table *entry = &table[13];
5690 snprintf(buf, 32, "group%d", i);
5691 entry->procname = kstrdup(buf, GFP_KERNEL);
5693 entry->child = sd_alloc_ctl_group_table(sg);
5694 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5696 /* &table[nr_entries-1] is terminator */
5701 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5703 struct ctl_table *entry, *table;
5704 struct sched_domain *sd;
5705 int domain_num = 0, i;
5708 for_each_domain(cpu, sd)
5710 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5715 for_each_domain(cpu, sd) {
5716 snprintf(buf, 32, "domain%d", i);
5717 entry->procname = kstrdup(buf, GFP_KERNEL);
5719 entry->child = sd_alloc_ctl_domain_table(sd);
5726 static struct ctl_table_header *sd_sysctl_header;
5727 static void register_sched_domain_sysctl(void)
5729 int i, cpu_num = num_possible_cpus();
5730 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5733 WARN_ON(sd_ctl_dir[0].child);
5734 sd_ctl_dir[0].child = entry;
5739 for_each_possible_cpu(i) {
5740 snprintf(buf, 32, "cpu%d", i);
5741 entry->procname = kstrdup(buf, GFP_KERNEL);
5743 entry->child = sd_alloc_ctl_cpu_table(i);
5747 WARN_ON(sd_sysctl_header);
5748 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5751 /* may be called multiple times per register */
5752 static void unregister_sched_domain_sysctl(void)
5754 unregister_sysctl_table(sd_sysctl_header);
5755 sd_sysctl_header = NULL;
5756 if (sd_ctl_dir[0].child)
5757 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5760 static void register_sched_domain_sysctl(void)
5763 static void unregister_sched_domain_sysctl(void)
5766 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5768 static void set_rq_online(struct rq *rq)
5771 const struct sched_class *class;
5773 cpumask_set_cpu(rq->cpu, rq->rd->online);
5776 for_each_class(class) {
5777 if (class->rq_online)
5778 class->rq_online(rq);
5783 static void set_rq_offline(struct rq *rq)
5786 const struct sched_class *class;
5788 for_each_class(class) {
5789 if (class->rq_offline)
5790 class->rq_offline(rq);
5793 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5799 * migration_call - callback that gets triggered when a CPU is added.
5800 * Here we can start up the necessary migration thread for the new CPU.
5803 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5805 int cpu = (long)hcpu;
5806 unsigned long flags;
5807 struct rq *rq = cpu_rq(cpu);
5809 switch (action & ~CPU_TASKS_FROZEN) {
5811 case CPU_UP_PREPARE:
5812 raw_spin_lock_irqsave(&rq->lock, flags);
5813 walt_set_window_start(rq);
5814 raw_spin_unlock_irqrestore(&rq->lock, flags);
5815 rq->calc_load_update = calc_load_update;
5819 /* Update our root-domain */
5820 raw_spin_lock_irqsave(&rq->lock, flags);
5822 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5826 raw_spin_unlock_irqrestore(&rq->lock, flags);
5829 #ifdef CONFIG_HOTPLUG_CPU
5831 sched_ttwu_pending();
5832 /* Update our root-domain */
5833 raw_spin_lock_irqsave(&rq->lock, flags);
5834 walt_migrate_sync_cpu(cpu);
5836 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5840 BUG_ON(rq->nr_running != 1); /* the migration thread */
5841 raw_spin_unlock_irqrestore(&rq->lock, flags);
5845 calc_load_migrate(rq);
5850 update_max_interval();
5856 * Register at high priority so that task migration (migrate_all_tasks)
5857 * happens before everything else. This has to be lower priority than
5858 * the notifier in the perf_event subsystem, though.
5860 static struct notifier_block migration_notifier = {
5861 .notifier_call = migration_call,
5862 .priority = CPU_PRI_MIGRATION,
5865 static void set_cpu_rq_start_time(void)
5867 int cpu = smp_processor_id();
5868 struct rq *rq = cpu_rq(cpu);
5869 rq->age_stamp = sched_clock_cpu(cpu);
5872 static int sched_cpu_active(struct notifier_block *nfb,
5873 unsigned long action, void *hcpu)
5875 int cpu = (long)hcpu;
5877 switch (action & ~CPU_TASKS_FROZEN) {
5879 set_cpu_rq_start_time();
5884 * At this point a starting CPU has marked itself as online via
5885 * set_cpu_online(). But it might not yet have marked itself
5886 * as active, which is essential from here on.
5888 set_cpu_active(cpu, true);
5889 stop_machine_unpark(cpu);
5892 case CPU_DOWN_FAILED:
5893 set_cpu_active(cpu, true);
5901 static int sched_cpu_inactive(struct notifier_block *nfb,
5902 unsigned long action, void *hcpu)
5904 switch (action & ~CPU_TASKS_FROZEN) {
5905 case CPU_DOWN_PREPARE:
5906 set_cpu_active((long)hcpu, false);
5913 static int __init migration_init(void)
5915 void *cpu = (void *)(long)smp_processor_id();
5918 /* Initialize migration for the boot CPU */
5919 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5920 BUG_ON(err == NOTIFY_BAD);
5921 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5922 register_cpu_notifier(&migration_notifier);
5924 /* Register cpu active notifiers */
5925 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5926 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5930 early_initcall(migration_init);
5932 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5934 #ifdef CONFIG_SCHED_DEBUG
5936 static __read_mostly int sched_debug_enabled;
5938 static int __init sched_debug_setup(char *str)
5940 sched_debug_enabled = 1;
5944 early_param("sched_debug", sched_debug_setup);
5946 static inline bool sched_debug(void)
5948 return sched_debug_enabled;
5951 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5952 struct cpumask *groupmask)
5954 struct sched_group *group = sd->groups;
5956 cpumask_clear(groupmask);
5958 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5960 if (!(sd->flags & SD_LOAD_BALANCE)) {
5961 printk("does not load-balance\n");
5965 printk(KERN_CONT "span %*pbl level %s\n",
5966 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5968 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5969 printk(KERN_ERR "ERROR: domain->span does not contain "
5972 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5973 printk(KERN_ERR "ERROR: domain->groups does not contain"
5977 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5981 printk(KERN_ERR "ERROR: group is NULL\n");
5985 if (!cpumask_weight(sched_group_cpus(group))) {
5986 printk(KERN_CONT "\n");
5987 printk(KERN_ERR "ERROR: empty group\n");
5991 if (!(sd->flags & SD_OVERLAP) &&
5992 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5993 printk(KERN_CONT "\n");
5994 printk(KERN_ERR "ERROR: repeated CPUs\n");
5998 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6000 printk(KERN_CONT " %*pbl",
6001 cpumask_pr_args(sched_group_cpus(group)));
6002 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
6003 printk(KERN_CONT " (cpu_capacity = %lu)",
6004 group->sgc->capacity);
6007 group = group->next;
6008 } while (group != sd->groups);
6009 printk(KERN_CONT "\n");
6011 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6012 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6015 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6016 printk(KERN_ERR "ERROR: parent span is not a superset "
6017 "of domain->span\n");
6021 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6025 if (!sched_debug_enabled)
6029 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6033 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6036 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6044 #else /* !CONFIG_SCHED_DEBUG */
6045 # define sched_domain_debug(sd, cpu) do { } while (0)
6046 static inline bool sched_debug(void)
6050 #endif /* CONFIG_SCHED_DEBUG */
6052 static int sd_degenerate(struct sched_domain *sd)
6054 if (cpumask_weight(sched_domain_span(sd)) == 1) {
6055 if (sd->groups->sge)
6056 sd->flags &= ~SD_LOAD_BALANCE;
6061 /* Following flags need at least 2 groups */
6062 if (sd->flags & (SD_LOAD_BALANCE |
6063 SD_BALANCE_NEWIDLE |
6066 SD_SHARE_CPUCAPACITY |
6067 SD_ASYM_CPUCAPACITY |
6068 SD_SHARE_PKG_RESOURCES |
6069 SD_SHARE_POWERDOMAIN |
6070 SD_SHARE_CAP_STATES)) {
6071 if (sd->groups != sd->groups->next)
6075 /* Following flags don't use groups */
6076 if (sd->flags & (SD_WAKE_AFFINE))
6083 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6085 unsigned long cflags = sd->flags, pflags = parent->flags;
6087 if (sd_degenerate(parent))
6090 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6093 /* Flags needing groups don't count if only 1 group in parent */
6094 if (parent->groups == parent->groups->next) {
6095 pflags &= ~(SD_LOAD_BALANCE |
6096 SD_BALANCE_NEWIDLE |
6099 SD_ASYM_CPUCAPACITY |
6100 SD_SHARE_CPUCAPACITY |
6101 SD_SHARE_PKG_RESOURCES |
6103 SD_SHARE_POWERDOMAIN |
6104 SD_SHARE_CAP_STATES);
6105 if (parent->groups->sge) {
6106 parent->flags &= ~SD_LOAD_BALANCE;
6109 if (nr_node_ids == 1)
6110 pflags &= ~SD_SERIALIZE;
6112 if (~cflags & pflags)
6118 static void free_rootdomain(struct rcu_head *rcu)
6120 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6122 cpupri_cleanup(&rd->cpupri);
6123 cpudl_cleanup(&rd->cpudl);
6124 free_cpumask_var(rd->dlo_mask);
6125 free_cpumask_var(rd->rto_mask);
6126 free_cpumask_var(rd->online);
6127 free_cpumask_var(rd->span);
6131 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6133 struct root_domain *old_rd = NULL;
6134 unsigned long flags;
6136 raw_spin_lock_irqsave(&rq->lock, flags);
6141 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6144 cpumask_clear_cpu(rq->cpu, old_rd->span);
6147 * If we dont want to free the old_rd yet then
6148 * set old_rd to NULL to skip the freeing later
6151 if (!atomic_dec_and_test(&old_rd->refcount))
6155 atomic_inc(&rd->refcount);
6158 cpumask_set_cpu(rq->cpu, rd->span);
6159 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6162 raw_spin_unlock_irqrestore(&rq->lock, flags);
6165 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6168 static int init_rootdomain(struct root_domain *rd)
6170 memset(rd, 0, sizeof(*rd));
6172 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6174 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6176 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6178 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6181 init_dl_bw(&rd->dl_bw);
6182 if (cpudl_init(&rd->cpudl) != 0)
6185 if (cpupri_init(&rd->cpupri) != 0)
6188 init_max_cpu_capacity(&rd->max_cpu_capacity);
6190 rd->max_cap_orig_cpu = rd->min_cap_orig_cpu = -1;
6195 free_cpumask_var(rd->rto_mask);
6197 free_cpumask_var(rd->dlo_mask);
6199 free_cpumask_var(rd->online);
6201 free_cpumask_var(rd->span);
6207 * By default the system creates a single root-domain with all cpus as
6208 * members (mimicking the global state we have today).
6210 struct root_domain def_root_domain;
6212 static void init_defrootdomain(void)
6214 init_rootdomain(&def_root_domain);
6216 atomic_set(&def_root_domain.refcount, 1);
6219 static struct root_domain *alloc_rootdomain(void)
6221 struct root_domain *rd;
6223 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6227 if (init_rootdomain(rd) != 0) {
6235 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6237 struct sched_group *tmp, *first;
6246 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6251 } while (sg != first);
6254 static void free_sched_domain(struct rcu_head *rcu)
6256 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6259 * If its an overlapping domain it has private groups, iterate and
6262 if (sd->flags & SD_OVERLAP) {
6263 free_sched_groups(sd->groups, 1);
6264 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6265 kfree(sd->groups->sgc);
6271 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6273 call_rcu(&sd->rcu, free_sched_domain);
6276 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6278 for (; sd; sd = sd->parent)
6279 destroy_sched_domain(sd, cpu);
6283 * Keep a special pointer to the highest sched_domain that has
6284 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6285 * allows us to avoid some pointer chasing select_idle_sibling().
6287 * Also keep a unique ID per domain (we use the first cpu number in
6288 * the cpumask of the domain), this allows us to quickly tell if
6289 * two cpus are in the same cache domain, see cpus_share_cache().
6291 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6292 DEFINE_PER_CPU(int, sd_llc_size);
6293 DEFINE_PER_CPU(int, sd_llc_id);
6294 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6295 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6296 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6297 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6298 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6300 static void update_top_cache_domain(int cpu)
6302 struct sched_domain *sd;
6303 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6307 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6309 id = cpumask_first(sched_domain_span(sd));
6310 size = cpumask_weight(sched_domain_span(sd));
6311 busy_sd = sd->parent; /* sd_busy */
6313 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6315 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6316 per_cpu(sd_llc_size, cpu) = size;
6317 per_cpu(sd_llc_id, cpu) = id;
6319 sd = lowest_flag_domain(cpu, SD_NUMA);
6320 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6322 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6323 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6325 for_each_domain(cpu, sd) {
6326 if (sd->groups->sge)
6331 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6333 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6334 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6338 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6339 * hold the hotplug lock.
6342 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6344 struct rq *rq = cpu_rq(cpu);
6345 struct sched_domain *tmp;
6347 /* Remove the sched domains which do not contribute to scheduling. */
6348 for (tmp = sd; tmp; ) {
6349 struct sched_domain *parent = tmp->parent;
6353 if (sd_parent_degenerate(tmp, parent)) {
6354 tmp->parent = parent->parent;
6356 parent->parent->child = tmp;
6358 * Transfer SD_PREFER_SIBLING down in case of a
6359 * degenerate parent; the spans match for this
6360 * so the property transfers.
6362 if (parent->flags & SD_PREFER_SIBLING)
6363 tmp->flags |= SD_PREFER_SIBLING;
6364 destroy_sched_domain(parent, cpu);
6369 if (sd && sd_degenerate(sd)) {
6372 destroy_sched_domain(tmp, cpu);
6377 sched_domain_debug(sd, cpu);
6379 rq_attach_root(rq, rd);
6381 rcu_assign_pointer(rq->sd, sd);
6382 destroy_sched_domains(tmp, cpu);
6384 update_top_cache_domain(cpu);
6387 /* Setup the mask of cpus configured for isolated domains */
6388 static int __init isolated_cpu_setup(char *str)
6390 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6391 cpulist_parse(str, cpu_isolated_map);
6395 __setup("isolcpus=", isolated_cpu_setup);
6398 struct sched_domain ** __percpu sd;
6399 struct root_domain *rd;
6410 * Build an iteration mask that can exclude certain CPUs from the upwards
6413 * Only CPUs that can arrive at this group should be considered to continue
6416 * Asymmetric node setups can result in situations where the domain tree is of
6417 * unequal depth, make sure to skip domains that already cover the entire
6420 * In that case build_sched_domains() will have terminated the iteration early
6421 * and our sibling sd spans will be empty. Domains should always include the
6422 * cpu they're built on, so check that.
6425 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6427 const struct cpumask *sg_span = sched_group_cpus(sg);
6428 struct sd_data *sdd = sd->private;
6429 struct sched_domain *sibling;
6432 for_each_cpu(i, sg_span) {
6433 sibling = *per_cpu_ptr(sdd->sd, i);
6436 * Can happen in the asymmetric case, where these siblings are
6437 * unused. The mask will not be empty because those CPUs that
6438 * do have the top domain _should_ span the domain.
6440 if (!sibling->child)
6443 /* If we would not end up here, we can't continue from here */
6444 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6447 cpumask_set_cpu(i, sched_group_mask(sg));
6450 /* We must not have empty masks here */
6451 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6455 * Return the canonical balance cpu for this group, this is the first cpu
6456 * of this group that's also in the iteration mask.
6458 int group_balance_cpu(struct sched_group *sg)
6460 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6464 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6466 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6467 const struct cpumask *span = sched_domain_span(sd);
6468 struct cpumask *covered = sched_domains_tmpmask;
6469 struct sd_data *sdd = sd->private;
6470 struct sched_domain *sibling;
6473 cpumask_clear(covered);
6475 for_each_cpu(i, span) {
6476 struct cpumask *sg_span;
6478 if (cpumask_test_cpu(i, covered))
6481 sibling = *per_cpu_ptr(sdd->sd, i);
6483 /* See the comment near build_group_mask(). */
6484 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6487 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6488 GFP_KERNEL, cpu_to_node(cpu));
6493 sg_span = sched_group_cpus(sg);
6495 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6497 cpumask_set_cpu(i, sg_span);
6499 cpumask_or(covered, covered, sg_span);
6501 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6502 if (atomic_inc_return(&sg->sgc->ref) == 1)
6503 build_group_mask(sd, sg);
6506 * Initialize sgc->capacity such that even if we mess up the
6507 * domains and no possible iteration will get us here, we won't
6510 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6511 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6512 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6515 * Make sure the first group of this domain contains the
6516 * canonical balance cpu. Otherwise the sched_domain iteration
6517 * breaks. See update_sg_lb_stats().
6519 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6520 group_balance_cpu(sg) == cpu)
6530 sd->groups = groups;
6535 free_sched_groups(first, 0);
6540 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6542 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6543 struct sched_domain *child = sd->child;
6546 cpu = cpumask_first(sched_domain_span(child));
6549 *sg = *per_cpu_ptr(sdd->sg, cpu);
6550 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6551 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6558 * build_sched_groups will build a circular linked list of the groups
6559 * covered by the given span, and will set each group's ->cpumask correctly,
6560 * and ->cpu_capacity to 0.
6562 * Assumes the sched_domain tree is fully constructed
6565 build_sched_groups(struct sched_domain *sd, int cpu)
6567 struct sched_group *first = NULL, *last = NULL;
6568 struct sd_data *sdd = sd->private;
6569 const struct cpumask *span = sched_domain_span(sd);
6570 struct cpumask *covered;
6573 get_group(cpu, sdd, &sd->groups);
6574 atomic_inc(&sd->groups->ref);
6576 if (cpu != cpumask_first(span))
6579 lockdep_assert_held(&sched_domains_mutex);
6580 covered = sched_domains_tmpmask;
6582 cpumask_clear(covered);
6584 for_each_cpu(i, span) {
6585 struct sched_group *sg;
6588 if (cpumask_test_cpu(i, covered))
6591 group = get_group(i, sdd, &sg);
6592 cpumask_setall(sched_group_mask(sg));
6594 for_each_cpu(j, span) {
6595 if (get_group(j, sdd, NULL) != group)
6598 cpumask_set_cpu(j, covered);
6599 cpumask_set_cpu(j, sched_group_cpus(sg));
6614 * Initialize sched groups cpu_capacity.
6616 * cpu_capacity indicates the capacity of sched group, which is used while
6617 * distributing the load between different sched groups in a sched domain.
6618 * Typically cpu_capacity for all the groups in a sched domain will be same
6619 * unless there are asymmetries in the topology. If there are asymmetries,
6620 * group having more cpu_capacity will pickup more load compared to the
6621 * group having less cpu_capacity.
6623 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6625 struct sched_group *sg = sd->groups;
6630 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6632 } while (sg != sd->groups);
6634 if (cpu != group_balance_cpu(sg))
6637 update_group_capacity(sd, cpu);
6638 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6642 * Check that the per-cpu provided sd energy data is consistent for all cpus
6645 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6646 const struct cpumask *cpumask)
6648 const struct sched_group_energy * const sge = fn(cpu);
6649 struct cpumask mask;
6652 if (cpumask_weight(cpumask) <= 1)
6655 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6657 for_each_cpu(i, &mask) {
6658 const struct sched_group_energy * const e = fn(i);
6661 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6663 for (y = 0; y < (e->nr_idle_states); y++) {
6664 BUG_ON(e->idle_states[y].power !=
6665 sge->idle_states[y].power);
6668 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6670 for (y = 0; y < (e->nr_cap_states); y++) {
6671 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6672 BUG_ON(e->cap_states[y].power !=
6673 sge->cap_states[y].power);
6678 static void init_sched_energy(int cpu, struct sched_domain *sd,
6679 sched_domain_energy_f fn)
6681 if (!(fn && fn(cpu)))
6684 if (cpu != group_balance_cpu(sd->groups))
6687 if (sd->child && !sd->child->groups->sge) {
6688 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6689 #ifdef CONFIG_SCHED_DEBUG
6690 pr_err(" energy data on %s but not on %s domain\n",
6691 sd->name, sd->child->name);
6696 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6698 sd->groups->sge = fn(cpu);
6702 * Initializers for schedule domains
6703 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6706 static int default_relax_domain_level = -1;
6707 int sched_domain_level_max;
6709 static int __init setup_relax_domain_level(char *str)
6711 if (kstrtoint(str, 0, &default_relax_domain_level))
6712 pr_warn("Unable to set relax_domain_level\n");
6716 __setup("relax_domain_level=", setup_relax_domain_level);
6718 static void set_domain_attribute(struct sched_domain *sd,
6719 struct sched_domain_attr *attr)
6723 if (!attr || attr->relax_domain_level < 0) {
6724 if (default_relax_domain_level < 0)
6727 request = default_relax_domain_level;
6729 request = attr->relax_domain_level;
6730 if (request < sd->level) {
6731 /* turn off idle balance on this domain */
6732 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6734 /* turn on idle balance on this domain */
6735 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6739 static void __sdt_free(const struct cpumask *cpu_map);
6740 static int __sdt_alloc(const struct cpumask *cpu_map);
6742 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6743 const struct cpumask *cpu_map)
6747 if (!atomic_read(&d->rd->refcount))
6748 free_rootdomain(&d->rd->rcu); /* fall through */
6750 free_percpu(d->sd); /* fall through */
6752 __sdt_free(cpu_map); /* fall through */
6758 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6759 const struct cpumask *cpu_map)
6761 memset(d, 0, sizeof(*d));
6763 if (__sdt_alloc(cpu_map))
6764 return sa_sd_storage;
6765 d->sd = alloc_percpu(struct sched_domain *);
6767 return sa_sd_storage;
6768 d->rd = alloc_rootdomain();
6771 return sa_rootdomain;
6775 * NULL the sd_data elements we've used to build the sched_domain and
6776 * sched_group structure so that the subsequent __free_domain_allocs()
6777 * will not free the data we're using.
6779 static void claim_allocations(int cpu, struct sched_domain *sd)
6781 struct sd_data *sdd = sd->private;
6783 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6784 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6786 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6787 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6789 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6790 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6794 static int sched_domains_numa_levels;
6795 enum numa_topology_type sched_numa_topology_type;
6796 static int *sched_domains_numa_distance;
6797 int sched_max_numa_distance;
6798 static struct cpumask ***sched_domains_numa_masks;
6799 static int sched_domains_curr_level;
6803 * SD_flags allowed in topology descriptions.
6805 * These flags are purely descriptive of the topology and do not prescribe
6806 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6809 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6810 * SD_SHARE_PKG_RESOURCES - describes shared caches
6811 * SD_NUMA - describes NUMA topologies
6812 * SD_SHARE_POWERDOMAIN - describes shared power domain
6813 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6814 * SD_SHARE_CAP_STATES - describes shared capacity states
6816 * Odd one out, which beside describing the topology has a quirk also
6817 * prescribes the desired behaviour that goes along with it:
6820 * SD_ASYM_PACKING - describes SMT quirks
6822 #define TOPOLOGY_SD_FLAGS \
6823 (SD_SHARE_CPUCAPACITY | \
6824 SD_SHARE_PKG_RESOURCES | \
6827 SD_ASYM_CPUCAPACITY | \
6828 SD_SHARE_POWERDOMAIN | \
6829 SD_SHARE_CAP_STATES)
6831 static struct sched_domain *
6832 sd_init(struct sched_domain_topology_level *tl,
6833 struct sched_domain *child, int cpu)
6835 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6836 int sd_weight, sd_flags = 0;
6840 * Ugly hack to pass state to sd_numa_mask()...
6842 sched_domains_curr_level = tl->numa_level;
6845 sd_weight = cpumask_weight(tl->mask(cpu));
6848 sd_flags = (*tl->sd_flags)();
6849 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6850 "wrong sd_flags in topology description\n"))
6851 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6853 *sd = (struct sched_domain){
6854 .min_interval = sd_weight,
6855 .max_interval = 2*sd_weight,
6857 .imbalance_pct = 125,
6859 .cache_nice_tries = 0,
6866 .flags = 1*SD_LOAD_BALANCE
6867 | 1*SD_BALANCE_NEWIDLE
6872 | 0*SD_SHARE_CPUCAPACITY
6873 | 0*SD_SHARE_PKG_RESOURCES
6875 | 0*SD_PREFER_SIBLING
6880 .last_balance = jiffies,
6881 .balance_interval = sd_weight,
6883 .max_newidle_lb_cost = 0,
6884 .next_decay_max_lb_cost = jiffies,
6886 #ifdef CONFIG_SCHED_DEBUG
6892 * Convert topological properties into behaviour.
6895 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6896 struct sched_domain *t = sd;
6898 for_each_lower_domain(t)
6899 t->flags |= SD_BALANCE_WAKE;
6902 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6903 sd->flags |= SD_PREFER_SIBLING;
6904 sd->imbalance_pct = 110;
6905 sd->smt_gain = 1178; /* ~15% */
6907 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6908 sd->imbalance_pct = 117;
6909 sd->cache_nice_tries = 1;
6913 } else if (sd->flags & SD_NUMA) {
6914 sd->cache_nice_tries = 2;
6918 sd->flags |= SD_SERIALIZE;
6919 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6920 sd->flags &= ~(SD_BALANCE_EXEC |
6927 sd->flags |= SD_PREFER_SIBLING;
6928 sd->cache_nice_tries = 1;
6933 sd->private = &tl->data;
6939 * Topology list, bottom-up.
6941 static struct sched_domain_topology_level default_topology[] = {
6942 #ifdef CONFIG_SCHED_SMT
6943 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6945 #ifdef CONFIG_SCHED_MC
6946 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6948 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6952 static struct sched_domain_topology_level *sched_domain_topology =
6955 #define for_each_sd_topology(tl) \
6956 for (tl = sched_domain_topology; tl->mask; tl++)
6958 void set_sched_topology(struct sched_domain_topology_level *tl)
6960 sched_domain_topology = tl;
6965 static const struct cpumask *sd_numa_mask(int cpu)
6967 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6970 static void sched_numa_warn(const char *str)
6972 static int done = false;
6980 printk(KERN_WARNING "ERROR: %s\n\n", str);
6982 for (i = 0; i < nr_node_ids; i++) {
6983 printk(KERN_WARNING " ");
6984 for (j = 0; j < nr_node_ids; j++)
6985 printk(KERN_CONT "%02d ", node_distance(i,j));
6986 printk(KERN_CONT "\n");
6988 printk(KERN_WARNING "\n");
6991 bool find_numa_distance(int distance)
6995 if (distance == node_distance(0, 0))
6998 for (i = 0; i < sched_domains_numa_levels; i++) {
6999 if (sched_domains_numa_distance[i] == distance)
7007 * A system can have three types of NUMA topology:
7008 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
7009 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
7010 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
7012 * The difference between a glueless mesh topology and a backplane
7013 * topology lies in whether communication between not directly
7014 * connected nodes goes through intermediary nodes (where programs
7015 * could run), or through backplane controllers. This affects
7016 * placement of programs.
7018 * The type of topology can be discerned with the following tests:
7019 * - If the maximum distance between any nodes is 1 hop, the system
7020 * is directly connected.
7021 * - If for two nodes A and B, located N > 1 hops away from each other,
7022 * there is an intermediary node C, which is < N hops away from both
7023 * nodes A and B, the system is a glueless mesh.
7025 static void init_numa_topology_type(void)
7029 n = sched_max_numa_distance;
7031 if (sched_domains_numa_levels <= 1) {
7032 sched_numa_topology_type = NUMA_DIRECT;
7036 for_each_online_node(a) {
7037 for_each_online_node(b) {
7038 /* Find two nodes furthest removed from each other. */
7039 if (node_distance(a, b) < n)
7042 /* Is there an intermediary node between a and b? */
7043 for_each_online_node(c) {
7044 if (node_distance(a, c) < n &&
7045 node_distance(b, c) < n) {
7046 sched_numa_topology_type =
7052 sched_numa_topology_type = NUMA_BACKPLANE;
7058 static void sched_init_numa(void)
7060 int next_distance, curr_distance = node_distance(0, 0);
7061 struct sched_domain_topology_level *tl;
7065 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
7066 if (!sched_domains_numa_distance)
7070 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
7071 * unique distances in the node_distance() table.
7073 * Assumes node_distance(0,j) includes all distances in
7074 * node_distance(i,j) in order to avoid cubic time.
7076 next_distance = curr_distance;
7077 for (i = 0; i < nr_node_ids; i++) {
7078 for (j = 0; j < nr_node_ids; j++) {
7079 for (k = 0; k < nr_node_ids; k++) {
7080 int distance = node_distance(i, k);
7082 if (distance > curr_distance &&
7083 (distance < next_distance ||
7084 next_distance == curr_distance))
7085 next_distance = distance;
7088 * While not a strong assumption it would be nice to know
7089 * about cases where if node A is connected to B, B is not
7090 * equally connected to A.
7092 if (sched_debug() && node_distance(k, i) != distance)
7093 sched_numa_warn("Node-distance not symmetric");
7095 if (sched_debug() && i && !find_numa_distance(distance))
7096 sched_numa_warn("Node-0 not representative");
7098 if (next_distance != curr_distance) {
7099 sched_domains_numa_distance[level++] = next_distance;
7100 sched_domains_numa_levels = level;
7101 curr_distance = next_distance;
7106 * In case of sched_debug() we verify the above assumption.
7116 * 'level' contains the number of unique distances, excluding the
7117 * identity distance node_distance(i,i).
7119 * The sched_domains_numa_distance[] array includes the actual distance
7124 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7125 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7126 * the array will contain less then 'level' members. This could be
7127 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7128 * in other functions.
7130 * We reset it to 'level' at the end of this function.
7132 sched_domains_numa_levels = 0;
7134 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7135 if (!sched_domains_numa_masks)
7139 * Now for each level, construct a mask per node which contains all
7140 * cpus of nodes that are that many hops away from us.
7142 for (i = 0; i < level; i++) {
7143 sched_domains_numa_masks[i] =
7144 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7145 if (!sched_domains_numa_masks[i])
7148 for (j = 0; j < nr_node_ids; j++) {
7149 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7153 sched_domains_numa_masks[i][j] = mask;
7156 if (node_distance(j, k) > sched_domains_numa_distance[i])
7159 cpumask_or(mask, mask, cpumask_of_node(k));
7164 /* Compute default topology size */
7165 for (i = 0; sched_domain_topology[i].mask; i++);
7167 tl = kzalloc((i + level + 1) *
7168 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7173 * Copy the default topology bits..
7175 for (i = 0; sched_domain_topology[i].mask; i++)
7176 tl[i] = sched_domain_topology[i];
7179 * .. and append 'j' levels of NUMA goodness.
7181 for (j = 0; j < level; i++, j++) {
7182 tl[i] = (struct sched_domain_topology_level){
7183 .mask = sd_numa_mask,
7184 .sd_flags = cpu_numa_flags,
7185 .flags = SDTL_OVERLAP,
7191 sched_domain_topology = tl;
7193 sched_domains_numa_levels = level;
7194 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7196 init_numa_topology_type();
7199 static void sched_domains_numa_masks_set(int cpu)
7202 int node = cpu_to_node(cpu);
7204 for (i = 0; i < sched_domains_numa_levels; i++) {
7205 for (j = 0; j < nr_node_ids; j++) {
7206 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7207 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7212 static void sched_domains_numa_masks_clear(int cpu)
7215 for (i = 0; i < sched_domains_numa_levels; i++) {
7216 for (j = 0; j < nr_node_ids; j++)
7217 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7222 * Update sched_domains_numa_masks[level][node] array when new cpus
7225 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7226 unsigned long action,
7229 int cpu = (long)hcpu;
7231 switch (action & ~CPU_TASKS_FROZEN) {
7233 sched_domains_numa_masks_set(cpu);
7237 sched_domains_numa_masks_clear(cpu);
7247 static inline void sched_init_numa(void)
7251 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7252 unsigned long action,
7257 #endif /* CONFIG_NUMA */
7259 static int __sdt_alloc(const struct cpumask *cpu_map)
7261 struct sched_domain_topology_level *tl;
7264 for_each_sd_topology(tl) {
7265 struct sd_data *sdd = &tl->data;
7267 sdd->sd = alloc_percpu(struct sched_domain *);
7271 sdd->sg = alloc_percpu(struct sched_group *);
7275 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7279 for_each_cpu(j, cpu_map) {
7280 struct sched_domain *sd;
7281 struct sched_group *sg;
7282 struct sched_group_capacity *sgc;
7284 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7285 GFP_KERNEL, cpu_to_node(j));
7289 *per_cpu_ptr(sdd->sd, j) = sd;
7291 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7292 GFP_KERNEL, cpu_to_node(j));
7298 *per_cpu_ptr(sdd->sg, j) = sg;
7300 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7301 GFP_KERNEL, cpu_to_node(j));
7305 *per_cpu_ptr(sdd->sgc, j) = sgc;
7312 static void __sdt_free(const struct cpumask *cpu_map)
7314 struct sched_domain_topology_level *tl;
7317 for_each_sd_topology(tl) {
7318 struct sd_data *sdd = &tl->data;
7320 for_each_cpu(j, cpu_map) {
7321 struct sched_domain *sd;
7324 sd = *per_cpu_ptr(sdd->sd, j);
7325 if (sd && (sd->flags & SD_OVERLAP))
7326 free_sched_groups(sd->groups, 0);
7327 kfree(*per_cpu_ptr(sdd->sd, j));
7331 kfree(*per_cpu_ptr(sdd->sg, j));
7333 kfree(*per_cpu_ptr(sdd->sgc, j));
7335 free_percpu(sdd->sd);
7337 free_percpu(sdd->sg);
7339 free_percpu(sdd->sgc);
7344 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7345 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7346 struct sched_domain *child, int cpu)
7348 struct sched_domain *sd = sd_init(tl, child, cpu);
7350 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7352 sd->level = child->level + 1;
7353 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7356 if (!cpumask_subset(sched_domain_span(child),
7357 sched_domain_span(sd))) {
7358 pr_err("BUG: arch topology borken\n");
7359 #ifdef CONFIG_SCHED_DEBUG
7360 pr_err(" the %s domain not a subset of the %s domain\n",
7361 child->name, sd->name);
7363 /* Fixup, ensure @sd has at least @child cpus. */
7364 cpumask_or(sched_domain_span(sd),
7365 sched_domain_span(sd),
7366 sched_domain_span(child));
7370 set_domain_attribute(sd, attr);
7376 * Build sched domains for a given set of cpus and attach the sched domains
7377 * to the individual cpus
7379 static int build_sched_domains(const struct cpumask *cpu_map,
7380 struct sched_domain_attr *attr)
7382 enum s_alloc alloc_state;
7383 struct sched_domain *sd;
7385 int i, ret = -ENOMEM;
7387 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7388 if (alloc_state != sa_rootdomain)
7391 /* Set up domains for cpus specified by the cpu_map. */
7392 for_each_cpu(i, cpu_map) {
7393 struct sched_domain_topology_level *tl;
7396 for_each_sd_topology(tl) {
7397 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7398 if (tl == sched_domain_topology)
7399 *per_cpu_ptr(d.sd, i) = sd;
7400 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7401 sd->flags |= SD_OVERLAP;
7405 /* Build the groups for the domains */
7406 for_each_cpu(i, cpu_map) {
7407 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7408 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7409 if (sd->flags & SD_OVERLAP) {
7410 if (build_overlap_sched_groups(sd, i))
7413 if (build_sched_groups(sd, i))
7419 /* Calculate CPU capacity for physical packages and nodes */
7420 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7421 struct sched_domain_topology_level *tl = sched_domain_topology;
7423 if (!cpumask_test_cpu(i, cpu_map))
7426 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7427 init_sched_energy(i, sd, tl->energy);
7428 claim_allocations(i, sd);
7429 init_sched_groups_capacity(i, sd);
7433 /* Attach the domains */
7435 for_each_cpu(i, cpu_map) {
7436 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu);
7437 int min_cpu = READ_ONCE(d.rd->min_cap_orig_cpu);
7439 if ((max_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig >
7440 cpu_rq(max_cpu)->cpu_capacity_orig))
7441 WRITE_ONCE(d.rd->max_cap_orig_cpu, i);
7443 if ((min_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig <
7444 cpu_rq(min_cpu)->cpu_capacity_orig))
7445 WRITE_ONCE(d.rd->min_cap_orig_cpu, i);
7447 sd = *per_cpu_ptr(d.sd, i);
7449 cpu_attach_domain(sd, d.rd, i);
7455 __free_domain_allocs(&d, alloc_state, cpu_map);
7459 static cpumask_var_t *doms_cur; /* current sched domains */
7460 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7461 static struct sched_domain_attr *dattr_cur;
7462 /* attribues of custom domains in 'doms_cur' */
7465 * Special case: If a kmalloc of a doms_cur partition (array of
7466 * cpumask) fails, then fallback to a single sched domain,
7467 * as determined by the single cpumask fallback_doms.
7469 static cpumask_var_t fallback_doms;
7472 * arch_update_cpu_topology lets virtualized architectures update the
7473 * cpu core maps. It is supposed to return 1 if the topology changed
7474 * or 0 if it stayed the same.
7476 int __weak arch_update_cpu_topology(void)
7481 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7484 cpumask_var_t *doms;
7486 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7489 for (i = 0; i < ndoms; i++) {
7490 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7491 free_sched_domains(doms, i);
7498 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7501 for (i = 0; i < ndoms; i++)
7502 free_cpumask_var(doms[i]);
7507 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7508 * For now this just excludes isolated cpus, but could be used to
7509 * exclude other special cases in the future.
7511 static int init_sched_domains(const struct cpumask *cpu_map)
7515 arch_update_cpu_topology();
7517 doms_cur = alloc_sched_domains(ndoms_cur);
7519 doms_cur = &fallback_doms;
7520 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7521 err = build_sched_domains(doms_cur[0], NULL);
7522 register_sched_domain_sysctl();
7528 * Detach sched domains from a group of cpus specified in cpu_map
7529 * These cpus will now be attached to the NULL domain
7531 static void detach_destroy_domains(const struct cpumask *cpu_map)
7536 for_each_cpu(i, cpu_map)
7537 cpu_attach_domain(NULL, &def_root_domain, i);
7541 /* handle null as "default" */
7542 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7543 struct sched_domain_attr *new, int idx_new)
7545 struct sched_domain_attr tmp;
7552 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7553 new ? (new + idx_new) : &tmp,
7554 sizeof(struct sched_domain_attr));
7558 * Partition sched domains as specified by the 'ndoms_new'
7559 * cpumasks in the array doms_new[] of cpumasks. This compares
7560 * doms_new[] to the current sched domain partitioning, doms_cur[].
7561 * It destroys each deleted domain and builds each new domain.
7563 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7564 * The masks don't intersect (don't overlap.) We should setup one
7565 * sched domain for each mask. CPUs not in any of the cpumasks will
7566 * not be load balanced. If the same cpumask appears both in the
7567 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7570 * The passed in 'doms_new' should be allocated using
7571 * alloc_sched_domains. This routine takes ownership of it and will
7572 * free_sched_domains it when done with it. If the caller failed the
7573 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7574 * and partition_sched_domains() will fallback to the single partition
7575 * 'fallback_doms', it also forces the domains to be rebuilt.
7577 * If doms_new == NULL it will be replaced with cpu_online_mask.
7578 * ndoms_new == 0 is a special case for destroying existing domains,
7579 * and it will not create the default domain.
7581 * Call with hotplug lock held
7583 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7584 struct sched_domain_attr *dattr_new)
7589 mutex_lock(&sched_domains_mutex);
7591 /* always unregister in case we don't destroy any domains */
7592 unregister_sched_domain_sysctl();
7594 /* Let architecture update cpu core mappings. */
7595 new_topology = arch_update_cpu_topology();
7597 n = doms_new ? ndoms_new : 0;
7599 /* Destroy deleted domains */
7600 for (i = 0; i < ndoms_cur; i++) {
7601 for (j = 0; j < n && !new_topology; j++) {
7602 if (cpumask_equal(doms_cur[i], doms_new[j])
7603 && dattrs_equal(dattr_cur, i, dattr_new, j))
7606 /* no match - a current sched domain not in new doms_new[] */
7607 detach_destroy_domains(doms_cur[i]);
7613 if (doms_new == NULL) {
7615 doms_new = &fallback_doms;
7616 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7617 WARN_ON_ONCE(dattr_new);
7620 /* Build new domains */
7621 for (i = 0; i < ndoms_new; i++) {
7622 for (j = 0; j < n && !new_topology; j++) {
7623 if (cpumask_equal(doms_new[i], doms_cur[j])
7624 && dattrs_equal(dattr_new, i, dattr_cur, j))
7627 /* no match - add a new doms_new */
7628 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7633 /* Remember the new sched domains */
7634 if (doms_cur != &fallback_doms)
7635 free_sched_domains(doms_cur, ndoms_cur);
7636 kfree(dattr_cur); /* kfree(NULL) is safe */
7637 doms_cur = doms_new;
7638 dattr_cur = dattr_new;
7639 ndoms_cur = ndoms_new;
7641 register_sched_domain_sysctl();
7643 mutex_unlock(&sched_domains_mutex);
7646 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7649 * Update cpusets according to cpu_active mask. If cpusets are
7650 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7651 * around partition_sched_domains().
7653 * If we come here as part of a suspend/resume, don't touch cpusets because we
7654 * want to restore it back to its original state upon resume anyway.
7656 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7660 case CPU_ONLINE_FROZEN:
7661 case CPU_DOWN_FAILED_FROZEN:
7664 * num_cpus_frozen tracks how many CPUs are involved in suspend
7665 * resume sequence. As long as this is not the last online
7666 * operation in the resume sequence, just build a single sched
7667 * domain, ignoring cpusets.
7669 partition_sched_domains(1, NULL, NULL);
7670 if (--num_cpus_frozen)
7674 * This is the last CPU online operation. So fall through and
7675 * restore the original sched domains by considering the
7676 * cpuset configurations.
7678 cpuset_force_rebuild();
7681 cpuset_update_active_cpus(true);
7689 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7692 unsigned long flags;
7693 long cpu = (long)hcpu;
7699 case CPU_DOWN_PREPARE:
7700 rcu_read_lock_sched();
7701 dl_b = dl_bw_of(cpu);
7703 raw_spin_lock_irqsave(&dl_b->lock, flags);
7704 cpus = dl_bw_cpus(cpu);
7705 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7706 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7708 rcu_read_unlock_sched();
7711 return notifier_from_errno(-EBUSY);
7712 cpuset_update_active_cpus(false);
7714 case CPU_DOWN_PREPARE_FROZEN:
7716 partition_sched_domains(1, NULL, NULL);
7724 void __init sched_init_smp(void)
7726 cpumask_var_t non_isolated_cpus;
7728 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7729 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7734 * There's no userspace yet to cause hotplug operations; hence all the
7735 * cpu masks are stable and all blatant races in the below code cannot
7738 mutex_lock(&sched_domains_mutex);
7739 init_sched_domains(cpu_active_mask);
7740 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7741 if (cpumask_empty(non_isolated_cpus))
7742 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7743 mutex_unlock(&sched_domains_mutex);
7745 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7746 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7747 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7751 /* Move init over to a non-isolated CPU */
7752 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7754 sched_init_granularity();
7755 free_cpumask_var(non_isolated_cpus);
7757 init_sched_rt_class();
7758 init_sched_dl_class();
7761 void __init sched_init_smp(void)
7763 sched_init_granularity();
7765 #endif /* CONFIG_SMP */
7767 int in_sched_functions(unsigned long addr)
7769 return in_lock_functions(addr) ||
7770 (addr >= (unsigned long)__sched_text_start
7771 && addr < (unsigned long)__sched_text_end);
7774 #ifdef CONFIG_CGROUP_SCHED
7776 * Default task group.
7777 * Every task in system belongs to this group at bootup.
7779 struct task_group root_task_group;
7780 LIST_HEAD(task_groups);
7783 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7785 void __init sched_init(void)
7788 unsigned long alloc_size = 0, ptr;
7790 #ifdef CONFIG_FAIR_GROUP_SCHED
7791 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7793 #ifdef CONFIG_RT_GROUP_SCHED
7794 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7797 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7799 #ifdef CONFIG_FAIR_GROUP_SCHED
7800 root_task_group.se = (struct sched_entity **)ptr;
7801 ptr += nr_cpu_ids * sizeof(void **);
7803 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7804 ptr += nr_cpu_ids * sizeof(void **);
7806 #endif /* CONFIG_FAIR_GROUP_SCHED */
7807 #ifdef CONFIG_RT_GROUP_SCHED
7808 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7809 ptr += nr_cpu_ids * sizeof(void **);
7811 root_task_group.rt_rq = (struct rt_rq **)ptr;
7812 ptr += nr_cpu_ids * sizeof(void **);
7814 #endif /* CONFIG_RT_GROUP_SCHED */
7816 #ifdef CONFIG_CPUMASK_OFFSTACK
7817 for_each_possible_cpu(i) {
7818 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7819 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7821 #endif /* CONFIG_CPUMASK_OFFSTACK */
7823 init_rt_bandwidth(&def_rt_bandwidth,
7824 global_rt_period(), global_rt_runtime());
7825 init_dl_bandwidth(&def_dl_bandwidth,
7826 global_rt_period(), global_rt_runtime());
7829 init_defrootdomain();
7832 #ifdef CONFIG_RT_GROUP_SCHED
7833 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7834 global_rt_period(), global_rt_runtime());
7835 #endif /* CONFIG_RT_GROUP_SCHED */
7837 #ifdef CONFIG_CGROUP_SCHED
7838 list_add(&root_task_group.list, &task_groups);
7839 INIT_LIST_HEAD(&root_task_group.children);
7840 INIT_LIST_HEAD(&root_task_group.siblings);
7841 autogroup_init(&init_task);
7843 #endif /* CONFIG_CGROUP_SCHED */
7845 for_each_possible_cpu(i) {
7849 raw_spin_lock_init(&rq->lock);
7851 rq->calc_load_active = 0;
7852 rq->calc_load_update = jiffies + LOAD_FREQ;
7853 init_cfs_rq(&rq->cfs);
7854 init_rt_rq(&rq->rt);
7855 init_dl_rq(&rq->dl);
7856 #ifdef CONFIG_FAIR_GROUP_SCHED
7857 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7858 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7859 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7861 * How much cpu bandwidth does root_task_group get?
7863 * In case of task-groups formed thr' the cgroup filesystem, it
7864 * gets 100% of the cpu resources in the system. This overall
7865 * system cpu resource is divided among the tasks of
7866 * root_task_group and its child task-groups in a fair manner,
7867 * based on each entity's (task or task-group's) weight
7868 * (se->load.weight).
7870 * In other words, if root_task_group has 10 tasks of weight
7871 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7872 * then A0's share of the cpu resource is:
7874 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7876 * We achieve this by letting root_task_group's tasks sit
7877 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7879 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7880 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7881 #endif /* CONFIG_FAIR_GROUP_SCHED */
7883 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7884 #ifdef CONFIG_RT_GROUP_SCHED
7885 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7888 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7889 rq->cpu_load[j] = 0;
7891 rq->last_load_update_tick = jiffies;
7896 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7897 rq->balance_callback = NULL;
7898 rq->active_balance = 0;
7899 rq->next_balance = jiffies;
7904 rq->avg_idle = 2*sysctl_sched_migration_cost;
7905 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7906 #ifdef CONFIG_SCHED_WALT
7907 rq->cur_irqload = 0;
7908 rq->avg_irqload = 0;
7912 INIT_LIST_HEAD(&rq->cfs_tasks);
7914 rq_attach_root(rq, &def_root_domain);
7915 #ifdef CONFIG_NO_HZ_COMMON
7918 #ifdef CONFIG_NO_HZ_FULL
7919 rq->last_sched_tick = 0;
7923 atomic_set(&rq->nr_iowait, 0);
7926 set_load_weight(&init_task);
7928 #ifdef CONFIG_PREEMPT_NOTIFIERS
7929 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7933 * The boot idle thread does lazy MMU switching as well:
7935 atomic_inc(&init_mm.mm_count);
7936 enter_lazy_tlb(&init_mm, current);
7939 * During early bootup we pretend to be a normal task:
7941 current->sched_class = &fair_sched_class;
7944 * Make us the idle thread. Technically, schedule() should not be
7945 * called from this thread, however somewhere below it might be,
7946 * but because we are the idle thread, we just pick up running again
7947 * when this runqueue becomes "idle".
7949 init_idle(current, smp_processor_id());
7951 calc_load_update = jiffies + LOAD_FREQ;
7954 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7955 /* May be allocated at isolcpus cmdline parse time */
7956 if (cpu_isolated_map == NULL)
7957 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7958 idle_thread_set_boot_cpu();
7959 set_cpu_rq_start_time();
7961 init_sched_fair_class();
7963 scheduler_running = 1;
7966 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7967 static inline int preempt_count_equals(int preempt_offset)
7969 int nested = preempt_count() + rcu_preempt_depth();
7971 return (nested == preempt_offset);
7974 static int __might_sleep_init_called;
7975 int __init __might_sleep_init(void)
7977 __might_sleep_init_called = 1;
7980 early_initcall(__might_sleep_init);
7982 void __might_sleep(const char *file, int line, int preempt_offset)
7985 * Blocking primitives will set (and therefore destroy) current->state,
7986 * since we will exit with TASK_RUNNING make sure we enter with it,
7987 * otherwise we will destroy state.
7989 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7990 "do not call blocking ops when !TASK_RUNNING; "
7991 "state=%lx set at [<%p>] %pS\n",
7993 (void *)current->task_state_change,
7994 (void *)current->task_state_change);
7996 ___might_sleep(file, line, preempt_offset);
7998 EXPORT_SYMBOL(__might_sleep);
8000 void ___might_sleep(const char *file, int line, int preempt_offset)
8002 static unsigned long prev_jiffy; /* ratelimiting */
8004 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
8005 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
8006 !is_idle_task(current)) || oops_in_progress)
8008 if (system_state != SYSTEM_RUNNING &&
8009 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8011 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8013 prev_jiffy = jiffies;
8016 "BUG: sleeping function called from invalid context at %s:%d\n",
8019 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8020 in_atomic(), irqs_disabled(),
8021 current->pid, current->comm);
8023 if (task_stack_end_corrupted(current))
8024 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
8026 debug_show_held_locks(current);
8027 if (irqs_disabled())
8028 print_irqtrace_events(current);
8029 #ifdef CONFIG_DEBUG_PREEMPT
8030 if (!preempt_count_equals(preempt_offset)) {
8031 pr_err("Preemption disabled at:");
8032 print_ip_sym(current->preempt_disable_ip);
8038 EXPORT_SYMBOL(___might_sleep);
8041 #ifdef CONFIG_MAGIC_SYSRQ
8042 void normalize_rt_tasks(void)
8044 struct task_struct *g, *p;
8045 struct sched_attr attr = {
8046 .sched_policy = SCHED_NORMAL,
8049 read_lock(&tasklist_lock);
8050 for_each_process_thread(g, p) {
8052 * Only normalize user tasks:
8054 if (p->flags & PF_KTHREAD)
8057 p->se.exec_start = 0;
8058 #ifdef CONFIG_SCHEDSTATS
8059 p->se.statistics.wait_start = 0;
8060 p->se.statistics.sleep_start = 0;
8061 p->se.statistics.block_start = 0;
8064 if (!dl_task(p) && !rt_task(p)) {
8066 * Renice negative nice level userspace
8069 if (task_nice(p) < 0)
8070 set_user_nice(p, 0);
8074 __sched_setscheduler(p, &attr, false, false);
8076 read_unlock(&tasklist_lock);
8079 #endif /* CONFIG_MAGIC_SYSRQ */
8081 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8083 * These functions are only useful for the IA64 MCA handling, or kdb.
8085 * They can only be called when the whole system has been
8086 * stopped - every CPU needs to be quiescent, and no scheduling
8087 * activity can take place. Using them for anything else would
8088 * be a serious bug, and as a result, they aren't even visible
8089 * under any other configuration.
8093 * curr_task - return the current task for a given cpu.
8094 * @cpu: the processor in question.
8096 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8098 * Return: The current task for @cpu.
8100 struct task_struct *curr_task(int cpu)
8102 return cpu_curr(cpu);
8105 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8109 * set_curr_task - set the current task for a given cpu.
8110 * @cpu: the processor in question.
8111 * @p: the task pointer to set.
8113 * Description: This function must only be used when non-maskable interrupts
8114 * are serviced on a separate stack. It allows the architecture to switch the
8115 * notion of the current task on a cpu in a non-blocking manner. This function
8116 * must be called with all CPU's synchronized, and interrupts disabled, the
8117 * and caller must save the original value of the current task (see
8118 * curr_task() above) and restore that value before reenabling interrupts and
8119 * re-starting the system.
8121 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8123 void set_curr_task(int cpu, struct task_struct *p)
8130 #ifdef CONFIG_CGROUP_SCHED
8131 /* task_group_lock serializes the addition/removal of task groups */
8132 static DEFINE_SPINLOCK(task_group_lock);
8134 static void sched_free_group(struct task_group *tg)
8136 free_fair_sched_group(tg);
8137 free_rt_sched_group(tg);
8142 /* allocate runqueue etc for a new task group */
8143 struct task_group *sched_create_group(struct task_group *parent)
8145 struct task_group *tg;
8147 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8149 return ERR_PTR(-ENOMEM);
8151 if (!alloc_fair_sched_group(tg, parent))
8154 if (!alloc_rt_sched_group(tg, parent))
8160 sched_free_group(tg);
8161 return ERR_PTR(-ENOMEM);
8164 void sched_online_group(struct task_group *tg, struct task_group *parent)
8166 unsigned long flags;
8168 spin_lock_irqsave(&task_group_lock, flags);
8169 list_add_rcu(&tg->list, &task_groups);
8171 WARN_ON(!parent); /* root should already exist */
8173 tg->parent = parent;
8174 INIT_LIST_HEAD(&tg->children);
8175 list_add_rcu(&tg->siblings, &parent->children);
8176 spin_unlock_irqrestore(&task_group_lock, flags);
8179 /* rcu callback to free various structures associated with a task group */
8180 static void sched_free_group_rcu(struct rcu_head *rhp)
8182 /* now it should be safe to free those cfs_rqs */
8183 sched_free_group(container_of(rhp, struct task_group, rcu));
8186 void sched_destroy_group(struct task_group *tg)
8188 /* wait for possible concurrent references to cfs_rqs complete */
8189 call_rcu(&tg->rcu, sched_free_group_rcu);
8192 void sched_offline_group(struct task_group *tg)
8194 unsigned long flags;
8197 /* end participation in shares distribution */
8198 for_each_possible_cpu(i)
8199 unregister_fair_sched_group(tg, i);
8201 spin_lock_irqsave(&task_group_lock, flags);
8202 list_del_rcu(&tg->list);
8203 list_del_rcu(&tg->siblings);
8204 spin_unlock_irqrestore(&task_group_lock, flags);
8207 static void sched_change_group(struct task_struct *tsk, int type)
8209 struct task_group *tg;
8212 * All callers are synchronized by task_rq_lock(); we do not use RCU
8213 * which is pointless here. Thus, we pass "true" to task_css_check()
8214 * to prevent lockdep warnings.
8216 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8217 struct task_group, css);
8218 tg = autogroup_task_group(tsk, tg);
8219 tsk->sched_task_group = tg;
8221 #ifdef CONFIG_FAIR_GROUP_SCHED
8222 if (tsk->sched_class->task_change_group)
8223 tsk->sched_class->task_change_group(tsk, type);
8226 set_task_rq(tsk, task_cpu(tsk));
8230 * Change task's runqueue when it moves between groups.
8232 * The caller of this function should have put the task in its new group by
8233 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8236 void sched_move_task(struct task_struct *tsk)
8238 int queued, running;
8239 unsigned long flags;
8242 rq = task_rq_lock(tsk, &flags);
8244 running = task_current(rq, tsk);
8245 queued = task_on_rq_queued(tsk);
8248 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8249 if (unlikely(running))
8250 put_prev_task(rq, tsk);
8252 sched_change_group(tsk, TASK_MOVE_GROUP);
8254 if (unlikely(running))
8255 tsk->sched_class->set_curr_task(rq);
8257 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8259 task_rq_unlock(rq, tsk, &flags);
8261 #endif /* CONFIG_CGROUP_SCHED */
8263 #ifdef CONFIG_RT_GROUP_SCHED
8265 * Ensure that the real time constraints are schedulable.
8267 static DEFINE_MUTEX(rt_constraints_mutex);
8269 /* Must be called with tasklist_lock held */
8270 static inline int tg_has_rt_tasks(struct task_group *tg)
8272 struct task_struct *g, *p;
8275 * Autogroups do not have RT tasks; see autogroup_create().
8277 if (task_group_is_autogroup(tg))
8280 for_each_process_thread(g, p) {
8281 if (rt_task(p) && task_group(p) == tg)
8288 struct rt_schedulable_data {
8289 struct task_group *tg;
8294 static int tg_rt_schedulable(struct task_group *tg, void *data)
8296 struct rt_schedulable_data *d = data;
8297 struct task_group *child;
8298 unsigned long total, sum = 0;
8299 u64 period, runtime;
8301 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8302 runtime = tg->rt_bandwidth.rt_runtime;
8305 period = d->rt_period;
8306 runtime = d->rt_runtime;
8310 * Cannot have more runtime than the period.
8312 if (runtime > period && runtime != RUNTIME_INF)
8316 * Ensure we don't starve existing RT tasks.
8318 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8321 total = to_ratio(period, runtime);
8324 * Nobody can have more than the global setting allows.
8326 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8330 * The sum of our children's runtime should not exceed our own.
8332 list_for_each_entry_rcu(child, &tg->children, siblings) {
8333 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8334 runtime = child->rt_bandwidth.rt_runtime;
8336 if (child == d->tg) {
8337 period = d->rt_period;
8338 runtime = d->rt_runtime;
8341 sum += to_ratio(period, runtime);
8350 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8354 struct rt_schedulable_data data = {
8356 .rt_period = period,
8357 .rt_runtime = runtime,
8361 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8367 static int tg_set_rt_bandwidth(struct task_group *tg,
8368 u64 rt_period, u64 rt_runtime)
8373 * Disallowing the root group RT runtime is BAD, it would disallow the
8374 * kernel creating (and or operating) RT threads.
8376 if (tg == &root_task_group && rt_runtime == 0)
8379 /* No period doesn't make any sense. */
8383 mutex_lock(&rt_constraints_mutex);
8384 read_lock(&tasklist_lock);
8385 err = __rt_schedulable(tg, rt_period, rt_runtime);
8389 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8390 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8391 tg->rt_bandwidth.rt_runtime = rt_runtime;
8393 for_each_possible_cpu(i) {
8394 struct rt_rq *rt_rq = tg->rt_rq[i];
8396 raw_spin_lock(&rt_rq->rt_runtime_lock);
8397 rt_rq->rt_runtime = rt_runtime;
8398 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8400 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8402 read_unlock(&tasklist_lock);
8403 mutex_unlock(&rt_constraints_mutex);
8408 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8410 u64 rt_runtime, rt_period;
8412 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8413 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8414 if (rt_runtime_us < 0)
8415 rt_runtime = RUNTIME_INF;
8417 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8420 static long sched_group_rt_runtime(struct task_group *tg)
8424 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8427 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8428 do_div(rt_runtime_us, NSEC_PER_USEC);
8429 return rt_runtime_us;
8432 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8434 u64 rt_runtime, rt_period;
8436 rt_period = rt_period_us * NSEC_PER_USEC;
8437 rt_runtime = tg->rt_bandwidth.rt_runtime;
8439 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8442 static long sched_group_rt_period(struct task_group *tg)
8446 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8447 do_div(rt_period_us, NSEC_PER_USEC);
8448 return rt_period_us;
8450 #endif /* CONFIG_RT_GROUP_SCHED */
8452 #ifdef CONFIG_RT_GROUP_SCHED
8453 static int sched_rt_global_constraints(void)
8457 mutex_lock(&rt_constraints_mutex);
8458 read_lock(&tasklist_lock);
8459 ret = __rt_schedulable(NULL, 0, 0);
8460 read_unlock(&tasklist_lock);
8461 mutex_unlock(&rt_constraints_mutex);
8466 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8468 /* Don't accept realtime tasks when there is no way for them to run */
8469 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8475 #else /* !CONFIG_RT_GROUP_SCHED */
8476 static int sched_rt_global_constraints(void)
8478 unsigned long flags;
8481 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8482 for_each_possible_cpu(i) {
8483 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8485 raw_spin_lock(&rt_rq->rt_runtime_lock);
8486 rt_rq->rt_runtime = global_rt_runtime();
8487 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8489 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8493 #endif /* CONFIG_RT_GROUP_SCHED */
8495 static int sched_dl_global_validate(void)
8497 u64 runtime = global_rt_runtime();
8498 u64 period = global_rt_period();
8499 u64 new_bw = to_ratio(period, runtime);
8502 unsigned long flags;
8505 * Here we want to check the bandwidth not being set to some
8506 * value smaller than the currently allocated bandwidth in
8507 * any of the root_domains.
8509 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8510 * cycling on root_domains... Discussion on different/better
8511 * solutions is welcome!
8513 for_each_possible_cpu(cpu) {
8514 rcu_read_lock_sched();
8515 dl_b = dl_bw_of(cpu);
8517 raw_spin_lock_irqsave(&dl_b->lock, flags);
8518 if (new_bw < dl_b->total_bw)
8520 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8522 rcu_read_unlock_sched();
8531 static void sched_dl_do_global(void)
8536 unsigned long flags;
8538 def_dl_bandwidth.dl_period = global_rt_period();
8539 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8541 if (global_rt_runtime() != RUNTIME_INF)
8542 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8545 * FIXME: As above...
8547 for_each_possible_cpu(cpu) {
8548 rcu_read_lock_sched();
8549 dl_b = dl_bw_of(cpu);
8551 raw_spin_lock_irqsave(&dl_b->lock, flags);
8553 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8555 rcu_read_unlock_sched();
8559 static int sched_rt_global_validate(void)
8561 if (sysctl_sched_rt_period <= 0)
8564 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8565 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8571 static void sched_rt_do_global(void)
8573 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8574 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8577 int sched_rt_handler(struct ctl_table *table, int write,
8578 void __user *buffer, size_t *lenp,
8581 int old_period, old_runtime;
8582 static DEFINE_MUTEX(mutex);
8586 old_period = sysctl_sched_rt_period;
8587 old_runtime = sysctl_sched_rt_runtime;
8589 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8591 if (!ret && write) {
8592 ret = sched_rt_global_validate();
8596 ret = sched_dl_global_validate();
8600 ret = sched_rt_global_constraints();
8604 sched_rt_do_global();
8605 sched_dl_do_global();
8609 sysctl_sched_rt_period = old_period;
8610 sysctl_sched_rt_runtime = old_runtime;
8612 mutex_unlock(&mutex);
8617 int sched_rr_handler(struct ctl_table *table, int write,
8618 void __user *buffer, size_t *lenp,
8622 static DEFINE_MUTEX(mutex);
8625 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8626 /* make sure that internally we keep jiffies */
8627 /* also, writing zero resets timeslice to default */
8628 if (!ret && write) {
8629 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8630 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8632 mutex_unlock(&mutex);
8636 #ifdef CONFIG_CGROUP_SCHED
8638 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8640 return css ? container_of(css, struct task_group, css) : NULL;
8643 static struct cgroup_subsys_state *
8644 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8646 struct task_group *parent = css_tg(parent_css);
8647 struct task_group *tg;
8650 /* This is early initialization for the top cgroup */
8651 return &root_task_group.css;
8654 tg = sched_create_group(parent);
8656 return ERR_PTR(-ENOMEM);
8661 /* Expose task group only after completing cgroup initialization */
8662 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8664 struct task_group *tg = css_tg(css);
8665 struct task_group *parent = css_tg(css->parent);
8668 sched_online_group(tg, parent);
8672 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8674 struct task_group *tg = css_tg(css);
8676 sched_offline_group(tg);
8679 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8681 struct task_group *tg = css_tg(css);
8684 * Relies on the RCU grace period between css_released() and this.
8686 sched_free_group(tg);
8690 * This is called before wake_up_new_task(), therefore we really only
8691 * have to set its group bits, all the other stuff does not apply.
8693 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8695 unsigned long flags;
8698 rq = task_rq_lock(task, &flags);
8700 update_rq_clock(rq);
8701 sched_change_group(task, TASK_SET_GROUP);
8703 task_rq_unlock(rq, task, &flags);
8706 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8708 struct task_struct *task;
8709 struct cgroup_subsys_state *css;
8712 cgroup_taskset_for_each(task, css, tset) {
8713 #ifdef CONFIG_RT_GROUP_SCHED
8714 if (!sched_rt_can_attach(css_tg(css), task))
8717 /* We don't support RT-tasks being in separate groups */
8718 if (task->sched_class != &fair_sched_class)
8722 * Serialize against wake_up_new_task() such that if its
8723 * running, we're sure to observe its full state.
8725 raw_spin_lock_irq(&task->pi_lock);
8727 * Avoid calling sched_move_task() before wake_up_new_task()
8728 * has happened. This would lead to problems with PELT, due to
8729 * move wanting to detach+attach while we're not attached yet.
8731 if (task->state == TASK_NEW)
8733 raw_spin_unlock_irq(&task->pi_lock);
8741 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8743 struct task_struct *task;
8744 struct cgroup_subsys_state *css;
8746 cgroup_taskset_for_each(task, css, tset)
8747 sched_move_task(task);
8750 #ifdef CONFIG_FAIR_GROUP_SCHED
8751 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8752 struct cftype *cftype, u64 shareval)
8754 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8757 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8760 struct task_group *tg = css_tg(css);
8762 return (u64) scale_load_down(tg->shares);
8765 #ifdef CONFIG_CFS_BANDWIDTH
8766 static DEFINE_MUTEX(cfs_constraints_mutex);
8768 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8769 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8771 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8773 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8775 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8776 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8778 if (tg == &root_task_group)
8782 * Ensure we have at some amount of bandwidth every period. This is
8783 * to prevent reaching a state of large arrears when throttled via
8784 * entity_tick() resulting in prolonged exit starvation.
8786 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8790 * Likewise, bound things on the otherside by preventing insane quota
8791 * periods. This also allows us to normalize in computing quota
8794 if (period > max_cfs_quota_period)
8798 * Prevent race between setting of cfs_rq->runtime_enabled and
8799 * unthrottle_offline_cfs_rqs().
8802 mutex_lock(&cfs_constraints_mutex);
8803 ret = __cfs_schedulable(tg, period, quota);
8807 runtime_enabled = quota != RUNTIME_INF;
8808 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8810 * If we need to toggle cfs_bandwidth_used, off->on must occur
8811 * before making related changes, and on->off must occur afterwards
8813 if (runtime_enabled && !runtime_was_enabled)
8814 cfs_bandwidth_usage_inc();
8815 raw_spin_lock_irq(&cfs_b->lock);
8816 cfs_b->period = ns_to_ktime(period);
8817 cfs_b->quota = quota;
8819 __refill_cfs_bandwidth_runtime(cfs_b);
8820 /* restart the period timer (if active) to handle new period expiry */
8821 if (runtime_enabled)
8822 start_cfs_bandwidth(cfs_b);
8823 raw_spin_unlock_irq(&cfs_b->lock);
8825 for_each_online_cpu(i) {
8826 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8827 struct rq *rq = cfs_rq->rq;
8829 raw_spin_lock_irq(&rq->lock);
8830 cfs_rq->runtime_enabled = runtime_enabled;
8831 cfs_rq->runtime_remaining = 0;
8833 if (cfs_rq->throttled)
8834 unthrottle_cfs_rq(cfs_rq);
8835 raw_spin_unlock_irq(&rq->lock);
8837 if (runtime_was_enabled && !runtime_enabled)
8838 cfs_bandwidth_usage_dec();
8840 mutex_unlock(&cfs_constraints_mutex);
8846 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8850 period = ktime_to_ns(tg->cfs_bandwidth.period);
8851 if (cfs_quota_us < 0)
8852 quota = RUNTIME_INF;
8854 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8856 return tg_set_cfs_bandwidth(tg, period, quota);
8859 long tg_get_cfs_quota(struct task_group *tg)
8863 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8866 quota_us = tg->cfs_bandwidth.quota;
8867 do_div(quota_us, NSEC_PER_USEC);
8872 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8876 period = (u64)cfs_period_us * NSEC_PER_USEC;
8877 quota = tg->cfs_bandwidth.quota;
8879 return tg_set_cfs_bandwidth(tg, period, quota);
8882 long tg_get_cfs_period(struct task_group *tg)
8886 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8887 do_div(cfs_period_us, NSEC_PER_USEC);
8889 return cfs_period_us;
8892 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8895 return tg_get_cfs_quota(css_tg(css));
8898 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8899 struct cftype *cftype, s64 cfs_quota_us)
8901 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8904 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8907 return tg_get_cfs_period(css_tg(css));
8910 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8911 struct cftype *cftype, u64 cfs_period_us)
8913 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8916 struct cfs_schedulable_data {
8917 struct task_group *tg;
8922 * normalize group quota/period to be quota/max_period
8923 * note: units are usecs
8925 static u64 normalize_cfs_quota(struct task_group *tg,
8926 struct cfs_schedulable_data *d)
8934 period = tg_get_cfs_period(tg);
8935 quota = tg_get_cfs_quota(tg);
8938 /* note: these should typically be equivalent */
8939 if (quota == RUNTIME_INF || quota == -1)
8942 return to_ratio(period, quota);
8945 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8947 struct cfs_schedulable_data *d = data;
8948 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8949 s64 quota = 0, parent_quota = -1;
8952 quota = RUNTIME_INF;
8954 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8956 quota = normalize_cfs_quota(tg, d);
8957 parent_quota = parent_b->hierarchical_quota;
8960 * ensure max(child_quota) <= parent_quota, inherit when no
8963 if (quota == RUNTIME_INF)
8964 quota = parent_quota;
8965 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8968 cfs_b->hierarchical_quota = quota;
8973 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8976 struct cfs_schedulable_data data = {
8982 if (quota != RUNTIME_INF) {
8983 do_div(data.period, NSEC_PER_USEC);
8984 do_div(data.quota, NSEC_PER_USEC);
8988 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8994 static int cpu_stats_show(struct seq_file *sf, void *v)
8996 struct task_group *tg = css_tg(seq_css(sf));
8997 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8999 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
9000 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
9001 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
9005 #endif /* CONFIG_CFS_BANDWIDTH */
9006 #endif /* CONFIG_FAIR_GROUP_SCHED */
9008 #ifdef CONFIG_RT_GROUP_SCHED
9009 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
9010 struct cftype *cft, s64 val)
9012 return sched_group_set_rt_runtime(css_tg(css), val);
9015 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
9018 return sched_group_rt_runtime(css_tg(css));
9021 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
9022 struct cftype *cftype, u64 rt_period_us)
9024 return sched_group_set_rt_period(css_tg(css), rt_period_us);
9027 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
9030 return sched_group_rt_period(css_tg(css));
9032 #endif /* CONFIG_RT_GROUP_SCHED */
9034 static struct cftype cpu_files[] = {
9035 #ifdef CONFIG_FAIR_GROUP_SCHED
9038 .read_u64 = cpu_shares_read_u64,
9039 .write_u64 = cpu_shares_write_u64,
9042 #ifdef CONFIG_CFS_BANDWIDTH
9044 .name = "cfs_quota_us",
9045 .read_s64 = cpu_cfs_quota_read_s64,
9046 .write_s64 = cpu_cfs_quota_write_s64,
9049 .name = "cfs_period_us",
9050 .read_u64 = cpu_cfs_period_read_u64,
9051 .write_u64 = cpu_cfs_period_write_u64,
9055 .seq_show = cpu_stats_show,
9058 #ifdef CONFIG_RT_GROUP_SCHED
9060 .name = "rt_runtime_us",
9061 .read_s64 = cpu_rt_runtime_read,
9062 .write_s64 = cpu_rt_runtime_write,
9065 .name = "rt_period_us",
9066 .read_u64 = cpu_rt_period_read_uint,
9067 .write_u64 = cpu_rt_period_write_uint,
9073 struct cgroup_subsys cpu_cgrp_subsys = {
9074 .css_alloc = cpu_cgroup_css_alloc,
9075 .css_online = cpu_cgroup_css_online,
9076 .css_released = cpu_cgroup_css_released,
9077 .css_free = cpu_cgroup_css_free,
9078 .fork = cpu_cgroup_fork,
9079 .can_attach = cpu_cgroup_can_attach,
9080 .attach = cpu_cgroup_attach,
9081 .legacy_cftypes = cpu_files,
9085 #endif /* CONFIG_CGROUP_SCHED */
9087 void dump_cpu_task(int cpu)
9089 pr_info("Task dump for CPU %d:\n", cpu);
9090 sched_show_task(cpu_curr(cpu));