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 INIT_LIST_HEAD(&p->se.group_node);
2174 walt_init_new_task_load(p);
2176 #ifdef CONFIG_FAIR_GROUP_SCHED
2177 p->se.cfs_rq = NULL;
2180 #ifdef CONFIG_SCHEDSTATS
2181 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2184 RB_CLEAR_NODE(&p->dl.rb_node);
2185 init_dl_task_timer(&p->dl);
2186 __dl_clear_params(p);
2188 INIT_LIST_HEAD(&p->rt.run_list);
2190 #ifdef CONFIG_PREEMPT_NOTIFIERS
2191 INIT_HLIST_HEAD(&p->preempt_notifiers);
2194 #ifdef CONFIG_NUMA_BALANCING
2195 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2196 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2197 p->mm->numa_scan_seq = 0;
2200 if (clone_flags & CLONE_VM)
2201 p->numa_preferred_nid = current->numa_preferred_nid;
2203 p->numa_preferred_nid = -1;
2205 p->node_stamp = 0ULL;
2206 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2207 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2208 p->numa_work.next = &p->numa_work;
2209 p->numa_faults = NULL;
2210 p->last_task_numa_placement = 0;
2211 p->last_sum_exec_runtime = 0;
2213 p->numa_group = NULL;
2214 #endif /* CONFIG_NUMA_BALANCING */
2217 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2219 #ifdef CONFIG_NUMA_BALANCING
2221 void set_numabalancing_state(bool enabled)
2224 static_branch_enable(&sched_numa_balancing);
2226 static_branch_disable(&sched_numa_balancing);
2229 #ifdef CONFIG_PROC_SYSCTL
2230 int sysctl_numa_balancing(struct ctl_table *table, int write,
2231 void __user *buffer, size_t *lenp, loff_t *ppos)
2235 int state = static_branch_likely(&sched_numa_balancing);
2237 if (write && !capable(CAP_SYS_ADMIN))
2242 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2246 set_numabalancing_state(state);
2253 * fork()/clone()-time setup:
2255 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2257 unsigned long flags;
2258 int cpu = get_cpu();
2260 __sched_fork(clone_flags, p);
2262 * We mark the process as NEW here. This guarantees that
2263 * nobody will actually run it, and a signal or other external
2264 * event cannot wake it up and insert it on the runqueue either.
2266 p->state = TASK_NEW;
2269 * Make sure we do not leak PI boosting priority to the child.
2271 p->prio = current->normal_prio;
2274 * Revert to default priority/policy on fork if requested.
2276 if (unlikely(p->sched_reset_on_fork)) {
2277 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2278 p->policy = SCHED_NORMAL;
2279 p->static_prio = NICE_TO_PRIO(0);
2281 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2282 p->static_prio = NICE_TO_PRIO(0);
2284 p->prio = p->normal_prio = __normal_prio(p);
2288 * We don't need the reset flag anymore after the fork. It has
2289 * fulfilled its duty:
2291 p->sched_reset_on_fork = 0;
2294 if (dl_prio(p->prio)) {
2297 } else if (rt_prio(p->prio)) {
2298 p->sched_class = &rt_sched_class;
2300 p->sched_class = &fair_sched_class;
2303 init_entity_runnable_average(&p->se);
2306 * The child is not yet in the pid-hash so no cgroup attach races,
2307 * and the cgroup is pinned to this child due to cgroup_fork()
2308 * is ran before sched_fork().
2310 * Silence PROVE_RCU.
2312 raw_spin_lock_irqsave(&p->pi_lock, flags);
2314 * We're setting the cpu for the first time, we don't migrate,
2315 * so use __set_task_cpu().
2317 __set_task_cpu(p, cpu);
2318 if (p->sched_class->task_fork)
2319 p->sched_class->task_fork(p);
2320 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2322 #ifdef CONFIG_SCHED_INFO
2323 if (likely(sched_info_on()))
2324 memset(&p->sched_info, 0, sizeof(p->sched_info));
2326 #if defined(CONFIG_SMP)
2329 init_task_preempt_count(p);
2331 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2332 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2339 unsigned long to_ratio(u64 period, u64 runtime)
2341 if (runtime == RUNTIME_INF)
2345 * Doing this here saves a lot of checks in all
2346 * the calling paths, and returning zero seems
2347 * safe for them anyway.
2352 return div64_u64(runtime << 20, period);
2356 inline struct dl_bw *dl_bw_of(int i)
2358 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2359 "sched RCU must be held");
2360 return &cpu_rq(i)->rd->dl_bw;
2363 static inline int dl_bw_cpus(int i)
2365 struct root_domain *rd = cpu_rq(i)->rd;
2368 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2369 "sched RCU must be held");
2370 for_each_cpu_and(i, rd->span, cpu_active_mask)
2376 inline struct dl_bw *dl_bw_of(int i)
2378 return &cpu_rq(i)->dl.dl_bw;
2381 static inline int dl_bw_cpus(int i)
2388 * We must be sure that accepting a new task (or allowing changing the
2389 * parameters of an existing one) is consistent with the bandwidth
2390 * constraints. If yes, this function also accordingly updates the currently
2391 * allocated bandwidth to reflect the new situation.
2393 * This function is called while holding p's rq->lock.
2395 * XXX we should delay bw change until the task's 0-lag point, see
2398 static int dl_overflow(struct task_struct *p, int policy,
2399 const struct sched_attr *attr)
2402 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2403 u64 period = attr->sched_period ?: attr->sched_deadline;
2404 u64 runtime = attr->sched_runtime;
2405 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2408 if (new_bw == p->dl.dl_bw)
2412 * Either if a task, enters, leave, or stays -deadline but changes
2413 * its parameters, we may need to update accordingly the total
2414 * allocated bandwidth of the container.
2416 raw_spin_lock(&dl_b->lock);
2417 cpus = dl_bw_cpus(task_cpu(p));
2418 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2419 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2420 __dl_add(dl_b, new_bw);
2422 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2423 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2424 __dl_clear(dl_b, p->dl.dl_bw);
2425 __dl_add(dl_b, new_bw);
2427 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2428 __dl_clear(dl_b, p->dl.dl_bw);
2431 raw_spin_unlock(&dl_b->lock);
2436 extern void init_dl_bw(struct dl_bw *dl_b);
2439 * wake_up_new_task - wake up a newly created task for the first time.
2441 * This function will do some initial scheduler statistics housekeeping
2442 * that must be done for every newly created context, then puts the task
2443 * on the runqueue and wakes it.
2445 void wake_up_new_task(struct task_struct *p)
2447 unsigned long flags;
2450 raw_spin_lock_irqsave(&p->pi_lock, flags);
2451 p->state = TASK_RUNNING;
2453 walt_init_new_task_load(p);
2455 /* Initialize new task's runnable average */
2456 init_entity_runnable_average(&p->se);
2459 * Fork balancing, do it here and not earlier because:
2460 * - cpus_allowed can change in the fork path
2461 * - any previously selected cpu might disappear through hotplug
2463 * Use __set_task_cpu() to avoid calling sched_class::migrate_task_rq,
2464 * as we're not fully set-up yet.
2466 __set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2468 rq = __task_rq_lock(p);
2469 update_rq_clock(rq);
2470 post_init_entity_util_avg(&p->se);
2472 walt_mark_task_starting(p);
2473 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2474 p->on_rq = TASK_ON_RQ_QUEUED;
2475 trace_sched_wakeup_new(p);
2476 check_preempt_curr(rq, p, WF_FORK);
2478 if (p->sched_class->task_woken) {
2480 * Nothing relies on rq->lock after this, so its fine to
2483 lockdep_unpin_lock(&rq->lock);
2484 p->sched_class->task_woken(rq, p);
2485 lockdep_pin_lock(&rq->lock);
2488 task_rq_unlock(rq, p, &flags);
2491 #ifdef CONFIG_PREEMPT_NOTIFIERS
2493 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2495 void preempt_notifier_inc(void)
2497 static_key_slow_inc(&preempt_notifier_key);
2499 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2501 void preempt_notifier_dec(void)
2503 static_key_slow_dec(&preempt_notifier_key);
2505 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2508 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2509 * @notifier: notifier struct to register
2511 void preempt_notifier_register(struct preempt_notifier *notifier)
2513 if (!static_key_false(&preempt_notifier_key))
2514 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2516 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2518 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2521 * preempt_notifier_unregister - no longer interested in preemption notifications
2522 * @notifier: notifier struct to unregister
2524 * This is *not* safe to call from within a preemption notifier.
2526 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2528 hlist_del(¬ifier->link);
2530 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2532 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2534 struct preempt_notifier *notifier;
2536 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2537 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2540 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2542 if (static_key_false(&preempt_notifier_key))
2543 __fire_sched_in_preempt_notifiers(curr);
2547 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2548 struct task_struct *next)
2550 struct preempt_notifier *notifier;
2552 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2553 notifier->ops->sched_out(notifier, next);
2556 static __always_inline void
2557 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2558 struct task_struct *next)
2560 if (static_key_false(&preempt_notifier_key))
2561 __fire_sched_out_preempt_notifiers(curr, next);
2564 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2566 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2571 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2572 struct task_struct *next)
2576 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2579 * prepare_task_switch - prepare to switch tasks
2580 * @rq: the runqueue preparing to switch
2581 * @prev: the current task that is being switched out
2582 * @next: the task we are going to switch to.
2584 * This is called with the rq lock held and interrupts off. It must
2585 * be paired with a subsequent finish_task_switch after the context
2588 * prepare_task_switch sets up locking and calls architecture specific
2592 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2593 struct task_struct *next)
2595 sched_info_switch(rq, prev, next);
2596 perf_event_task_sched_out(prev, next);
2597 fire_sched_out_preempt_notifiers(prev, next);
2598 prepare_lock_switch(rq, next);
2599 prepare_arch_switch(next);
2603 * finish_task_switch - clean up after a task-switch
2604 * @prev: the thread we just switched away from.
2606 * finish_task_switch must be called after the context switch, paired
2607 * with a prepare_task_switch call before the context switch.
2608 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2609 * and do any other architecture-specific cleanup actions.
2611 * Note that we may have delayed dropping an mm in context_switch(). If
2612 * so, we finish that here outside of the runqueue lock. (Doing it
2613 * with the lock held can cause deadlocks; see schedule() for
2616 * The context switch have flipped the stack from under us and restored the
2617 * local variables which were saved when this task called schedule() in the
2618 * past. prev == current is still correct but we need to recalculate this_rq
2619 * because prev may have moved to another CPU.
2621 static struct rq *finish_task_switch(struct task_struct *prev)
2622 __releases(rq->lock)
2624 struct rq *rq = this_rq();
2625 struct mm_struct *mm = rq->prev_mm;
2629 * The previous task will have left us with a preempt_count of 2
2630 * because it left us after:
2633 * preempt_disable(); // 1
2635 * raw_spin_lock_irq(&rq->lock) // 2
2637 * Also, see FORK_PREEMPT_COUNT.
2639 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2640 "corrupted preempt_count: %s/%d/0x%x\n",
2641 current->comm, current->pid, preempt_count()))
2642 preempt_count_set(FORK_PREEMPT_COUNT);
2647 * A task struct has one reference for the use as "current".
2648 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2649 * schedule one last time. The schedule call will never return, and
2650 * the scheduled task must drop that reference.
2652 * We must observe prev->state before clearing prev->on_cpu (in
2653 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2654 * running on another CPU and we could rave with its RUNNING -> DEAD
2655 * transition, resulting in a double drop.
2657 prev_state = prev->state;
2658 vtime_task_switch(prev);
2659 perf_event_task_sched_in(prev, current);
2660 finish_lock_switch(rq, prev);
2661 finish_arch_post_lock_switch();
2663 fire_sched_in_preempt_notifiers(current);
2666 if (unlikely(prev_state == TASK_DEAD)) {
2667 if (prev->sched_class->task_dead)
2668 prev->sched_class->task_dead(prev);
2671 * Remove function-return probe instances associated with this
2672 * task and put them back on the free list.
2674 kprobe_flush_task(prev);
2675 put_task_struct(prev);
2678 tick_nohz_task_switch();
2684 /* rq->lock is NOT held, but preemption is disabled */
2685 static void __balance_callback(struct rq *rq)
2687 struct callback_head *head, *next;
2688 void (*func)(struct rq *rq);
2689 unsigned long flags;
2691 raw_spin_lock_irqsave(&rq->lock, flags);
2692 head = rq->balance_callback;
2693 rq->balance_callback = NULL;
2695 func = (void (*)(struct rq *))head->func;
2702 raw_spin_unlock_irqrestore(&rq->lock, flags);
2705 static inline void balance_callback(struct rq *rq)
2707 if (unlikely(rq->balance_callback))
2708 __balance_callback(rq);
2713 static inline void balance_callback(struct rq *rq)
2720 * schedule_tail - first thing a freshly forked thread must call.
2721 * @prev: the thread we just switched away from.
2723 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2724 __releases(rq->lock)
2729 * New tasks start with FORK_PREEMPT_COUNT, see there and
2730 * finish_task_switch() for details.
2732 * finish_task_switch() will drop rq->lock() and lower preempt_count
2733 * and the preempt_enable() will end up enabling preemption (on
2734 * PREEMPT_COUNT kernels).
2737 rq = finish_task_switch(prev);
2738 balance_callback(rq);
2741 if (current->set_child_tid)
2742 put_user(task_pid_vnr(current), current->set_child_tid);
2746 * context_switch - switch to the new MM and the new thread's register state.
2748 static inline struct rq *
2749 context_switch(struct rq *rq, struct task_struct *prev,
2750 struct task_struct *next)
2752 struct mm_struct *mm, *oldmm;
2754 prepare_task_switch(rq, prev, next);
2757 oldmm = prev->active_mm;
2759 * For paravirt, this is coupled with an exit in switch_to to
2760 * combine the page table reload and the switch backend into
2763 arch_start_context_switch(prev);
2766 next->active_mm = oldmm;
2767 atomic_inc(&oldmm->mm_count);
2768 enter_lazy_tlb(oldmm, next);
2770 switch_mm(oldmm, mm, next);
2773 prev->active_mm = NULL;
2774 rq->prev_mm = oldmm;
2777 * Since the runqueue lock will be released by the next
2778 * task (which is an invalid locking op but in the case
2779 * of the scheduler it's an obvious special-case), so we
2780 * do an early lockdep release here:
2782 lockdep_unpin_lock(&rq->lock);
2783 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2785 /* Here we just switch the register state and the stack. */
2786 switch_to(prev, next, prev);
2789 return finish_task_switch(prev);
2793 * nr_running and nr_context_switches:
2795 * externally visible scheduler statistics: current number of runnable
2796 * threads, total number of context switches performed since bootup.
2798 unsigned long nr_running(void)
2800 unsigned long i, sum = 0;
2802 for_each_online_cpu(i)
2803 sum += cpu_rq(i)->nr_running;
2809 * Check if only the current task is running on the cpu.
2811 * Caution: this function does not check that the caller has disabled
2812 * preemption, thus the result might have a time-of-check-to-time-of-use
2813 * race. The caller is responsible to use it correctly, for example:
2815 * - from a non-preemptable section (of course)
2817 * - from a thread that is bound to a single CPU
2819 * - in a loop with very short iterations (e.g. a polling loop)
2821 bool single_task_running(void)
2823 return raw_rq()->nr_running == 1;
2825 EXPORT_SYMBOL(single_task_running);
2827 unsigned long long nr_context_switches(void)
2830 unsigned long long sum = 0;
2832 for_each_possible_cpu(i)
2833 sum += cpu_rq(i)->nr_switches;
2838 unsigned long nr_iowait(void)
2840 unsigned long i, sum = 0;
2842 for_each_possible_cpu(i)
2843 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2848 unsigned long nr_iowait_cpu(int cpu)
2850 struct rq *this = cpu_rq(cpu);
2851 return atomic_read(&this->nr_iowait);
2854 #ifdef CONFIG_CPU_QUIET
2855 u64 nr_running_integral(unsigned int cpu)
2857 unsigned int seqcnt;
2861 if (cpu >= nr_cpu_ids)
2867 * Update average to avoid reading stalled value if there were
2868 * no run-queue changes for a long time. On the other hand if
2869 * the changes are happening right now, just read current value
2873 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2874 integral = do_nr_running_integral(q);
2875 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2876 read_seqcount_begin(&q->ave_seqcnt);
2877 integral = q->nr_running_integral;
2884 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2886 struct rq *rq = this_rq();
2887 *nr_waiters = atomic_read(&rq->nr_iowait);
2888 *load = rq->load.weight;
2894 * sched_exec - execve() is a valuable balancing opportunity, because at
2895 * this point the task has the smallest effective memory and cache footprint.
2897 void sched_exec(void)
2899 struct task_struct *p = current;
2900 unsigned long flags;
2903 raw_spin_lock_irqsave(&p->pi_lock, flags);
2904 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2905 if (dest_cpu == smp_processor_id())
2908 if (likely(cpu_active(dest_cpu))) {
2909 struct migration_arg arg = { p, dest_cpu };
2911 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2912 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2916 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2921 DEFINE_PER_CPU(struct kernel_stat, kstat);
2922 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2924 EXPORT_PER_CPU_SYMBOL(kstat);
2925 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2928 * Return accounted runtime for the task.
2929 * In case the task is currently running, return the runtime plus current's
2930 * pending runtime that have not been accounted yet.
2932 unsigned long long task_sched_runtime(struct task_struct *p)
2934 unsigned long flags;
2938 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2940 * 64-bit doesn't need locks to atomically read a 64bit value.
2941 * So we have a optimization chance when the task's delta_exec is 0.
2942 * Reading ->on_cpu is racy, but this is ok.
2944 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2945 * If we race with it entering cpu, unaccounted time is 0. This is
2946 * indistinguishable from the read occurring a few cycles earlier.
2947 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2948 * been accounted, so we're correct here as well.
2950 if (!p->on_cpu || !task_on_rq_queued(p))
2951 return p->se.sum_exec_runtime;
2954 rq = task_rq_lock(p, &flags);
2956 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2957 * project cycles that may never be accounted to this
2958 * thread, breaking clock_gettime().
2960 if (task_current(rq, p) && task_on_rq_queued(p)) {
2961 update_rq_clock(rq);
2962 p->sched_class->update_curr(rq);
2964 ns = p->se.sum_exec_runtime;
2965 task_rq_unlock(rq, p, &flags);
2970 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2973 unsigned long add_capacity_margin(unsigned long cpu_capacity)
2975 cpu_capacity = cpu_capacity * capacity_margin;
2976 cpu_capacity /= SCHED_CAPACITY_SCALE;
2977 return cpu_capacity;
2981 unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2982 struct sched_capacity_reqs *scr)
2984 unsigned long total = add_capacity_margin(cfs_cap + scr->rt);
2985 return total += scr->dl;
2988 unsigned long boosted_cpu_util(int cpu);
2989 static void sched_freq_tick_pelt(int cpu)
2991 unsigned long cpu_utilization = boosted_cpu_util(cpu);
2992 unsigned long capacity_curr = capacity_curr_of(cpu);
2993 struct sched_capacity_reqs *scr;
2995 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2996 if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)
3000 * To make free room for a task that is building up its "real"
3001 * utilization and to harm its performance the least, request
3002 * a jump to a higher OPP as soon as the margin of free capacity
3003 * is impacted (specified by capacity_margin).
3004 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3006 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3007 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3010 #ifdef CONFIG_SCHED_WALT
3011 static void sched_freq_tick_walt(int cpu)
3013 unsigned long cpu_utilization = cpu_util_freq(cpu);
3014 unsigned long capacity_curr = capacity_curr_of(cpu);
3016 if (walt_disabled || !sysctl_sched_use_walt_cpu_util)
3017 return sched_freq_tick_pelt(cpu);
3020 * Add a margin to the WALT utilization to check if we will need to
3021 * increase frequency.
3022 * NOTE: WALT tracks a single CPU signal for all the scheduling
3023 * classes, thus this margin is going to be added to the DL class as
3024 * well, which is something we do not do in sched_freq_tick_pelt case.
3026 if (add_capacity_margin(cpu_utilization) <= capacity_curr)
3030 * It is likely that the load is growing so we
3031 * keep the added margin in our request as an
3033 * Remember CPU utilization in sched_capacity_reqs should be normalised.
3035 cpu_utilization = cpu_utilization * SCHED_CAPACITY_SCALE / capacity_orig_of(cpu);
3036 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
3039 #define _sched_freq_tick(cpu) sched_freq_tick_walt(cpu)
3041 #define _sched_freq_tick(cpu) sched_freq_tick_pelt(cpu)
3042 #endif /* CONFIG_SCHED_WALT */
3044 static void sched_freq_tick(int cpu)
3049 _sched_freq_tick(cpu);
3052 static inline void sched_freq_tick(int cpu) { }
3053 #endif /* CONFIG_CPU_FREQ_GOV_SCHED */
3056 * This function gets called by the timer code, with HZ frequency.
3057 * We call it with interrupts disabled.
3059 void scheduler_tick(void)
3061 int cpu = smp_processor_id();
3062 struct rq *rq = cpu_rq(cpu);
3063 struct task_struct *curr = rq->curr;
3067 raw_spin_lock(&rq->lock);
3068 walt_set_window_start(rq);
3069 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3070 walt_ktime_clock(), 0);
3071 update_rq_clock(rq);
3072 curr->sched_class->task_tick(rq, curr, 0);
3073 update_cpu_load_active(rq);
3074 calc_global_load_tick(rq);
3075 sched_freq_tick(cpu);
3076 raw_spin_unlock(&rq->lock);
3078 perf_event_task_tick();
3081 rq->idle_balance = idle_cpu(cpu);
3082 trigger_load_balance(rq);
3084 rq_last_tick_reset(rq);
3087 #ifdef CONFIG_NO_HZ_FULL
3089 * scheduler_tick_max_deferment
3091 * Keep at least one tick per second when a single
3092 * active task is running because the scheduler doesn't
3093 * yet completely support full dynticks environment.
3095 * This makes sure that uptime, CFS vruntime, load
3096 * balancing, etc... continue to move forward, even
3097 * with a very low granularity.
3099 * Return: Maximum deferment in nanoseconds.
3101 u64 scheduler_tick_max_deferment(void)
3103 struct rq *rq = this_rq();
3104 unsigned long next, now = READ_ONCE(jiffies);
3106 next = rq->last_sched_tick + HZ;
3108 if (time_before_eq(next, now))
3111 return jiffies_to_nsecs(next - now);
3115 notrace unsigned long get_parent_ip(unsigned long addr)
3117 if (in_lock_functions(addr)) {
3118 addr = CALLER_ADDR2;
3119 if (in_lock_functions(addr))
3120 addr = CALLER_ADDR3;
3125 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3126 defined(CONFIG_PREEMPT_TRACER))
3128 void preempt_count_add(int val)
3130 #ifdef CONFIG_DEBUG_PREEMPT
3134 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3137 __preempt_count_add(val);
3138 #ifdef CONFIG_DEBUG_PREEMPT
3140 * Spinlock count overflowing soon?
3142 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3145 if (preempt_count() == val) {
3146 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3147 #ifdef CONFIG_DEBUG_PREEMPT
3148 current->preempt_disable_ip = ip;
3150 trace_preempt_off(CALLER_ADDR0, ip);
3153 EXPORT_SYMBOL(preempt_count_add);
3154 NOKPROBE_SYMBOL(preempt_count_add);
3156 void preempt_count_sub(int val)
3158 #ifdef CONFIG_DEBUG_PREEMPT
3162 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3165 * Is the spinlock portion underflowing?
3167 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3168 !(preempt_count() & PREEMPT_MASK)))
3172 if (preempt_count() == val)
3173 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3174 __preempt_count_sub(val);
3176 EXPORT_SYMBOL(preempt_count_sub);
3177 NOKPROBE_SYMBOL(preempt_count_sub);
3182 * Print scheduling while atomic bug:
3184 static noinline void __schedule_bug(struct task_struct *prev)
3186 if (oops_in_progress)
3189 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3190 prev->comm, prev->pid, preempt_count());
3192 debug_show_held_locks(prev);
3194 if (irqs_disabled())
3195 print_irqtrace_events(prev);
3196 #ifdef CONFIG_DEBUG_PREEMPT
3197 if (in_atomic_preempt_off()) {
3198 pr_err("Preemption disabled at:");
3199 print_ip_sym(current->preempt_disable_ip);
3204 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3208 * Various schedule()-time debugging checks and statistics:
3210 static inline void schedule_debug(struct task_struct *prev)
3212 #ifdef CONFIG_SCHED_STACK_END_CHECK
3213 if (task_stack_end_corrupted(prev))
3214 panic("corrupted stack end detected inside scheduler\n");
3217 if (unlikely(in_atomic_preempt_off())) {
3218 __schedule_bug(prev);
3219 preempt_count_set(PREEMPT_DISABLED);
3223 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3225 schedstat_inc(this_rq(), sched_count);
3229 * Pick up the highest-prio task:
3231 static inline struct task_struct *
3232 pick_next_task(struct rq *rq, struct task_struct *prev)
3234 const struct sched_class *class = &fair_sched_class;
3235 struct task_struct *p;
3238 * Optimization: we know that if all tasks are in
3239 * the fair class we can call that function directly:
3241 if (likely(prev->sched_class == class &&
3242 rq->nr_running == rq->cfs.h_nr_running)) {
3243 p = fair_sched_class.pick_next_task(rq, prev);
3244 if (unlikely(p == RETRY_TASK))
3247 /* assumes fair_sched_class->next == idle_sched_class */
3249 p = idle_sched_class.pick_next_task(rq, prev);
3255 for_each_class(class) {
3256 p = class->pick_next_task(rq, prev);
3258 if (unlikely(p == RETRY_TASK))
3264 BUG(); /* the idle class will always have a runnable task */
3268 * __schedule() is the main scheduler function.
3270 * The main means of driving the scheduler and thus entering this function are:
3272 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3274 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3275 * paths. For example, see arch/x86/entry_64.S.
3277 * To drive preemption between tasks, the scheduler sets the flag in timer
3278 * interrupt handler scheduler_tick().
3280 * 3. Wakeups don't really cause entry into schedule(). They add a
3281 * task to the run-queue and that's it.
3283 * Now, if the new task added to the run-queue preempts the current
3284 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3285 * called on the nearest possible occasion:
3287 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3289 * - in syscall or exception context, at the next outmost
3290 * preempt_enable(). (this might be as soon as the wake_up()'s
3293 * - in IRQ context, return from interrupt-handler to
3294 * preemptible context
3296 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3299 * - cond_resched() call
3300 * - explicit schedule() call
3301 * - return from syscall or exception to user-space
3302 * - return from interrupt-handler to user-space
3304 * WARNING: must be called with preemption disabled!
3306 static void __sched notrace __schedule(bool preempt)
3308 struct task_struct *prev, *next;
3309 unsigned long *switch_count;
3314 cpu = smp_processor_id();
3316 rcu_note_context_switch();
3320 * do_exit() calls schedule() with preemption disabled as an exception;
3321 * however we must fix that up, otherwise the next task will see an
3322 * inconsistent (higher) preempt count.
3324 * It also avoids the below schedule_debug() test from complaining
3327 if (unlikely(prev->state == TASK_DEAD))
3328 preempt_enable_no_resched_notrace();
3330 schedule_debug(prev);
3332 if (sched_feat(HRTICK))
3336 * Make sure that signal_pending_state()->signal_pending() below
3337 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3338 * done by the caller to avoid the race with signal_wake_up().
3340 smp_mb__before_spinlock();
3341 raw_spin_lock_irq(&rq->lock);
3342 lockdep_pin_lock(&rq->lock);
3344 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3346 switch_count = &prev->nivcsw;
3347 if (!preempt && prev->state) {
3348 if (unlikely(signal_pending_state(prev->state, prev))) {
3349 prev->state = TASK_RUNNING;
3351 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3355 * If a worker went to sleep, notify and ask workqueue
3356 * whether it wants to wake up a task to maintain
3359 if (prev->flags & PF_WQ_WORKER) {
3360 struct task_struct *to_wakeup;
3362 to_wakeup = wq_worker_sleeping(prev, cpu);
3364 try_to_wake_up_local(to_wakeup);
3367 switch_count = &prev->nvcsw;
3370 if (task_on_rq_queued(prev))
3371 update_rq_clock(rq);
3373 next = pick_next_task(rq, prev);
3374 wallclock = walt_ktime_clock();
3375 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3376 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3377 clear_tsk_need_resched(prev);
3378 clear_preempt_need_resched();
3379 rq->clock_skip_update = 0;
3381 if (likely(prev != next)) {
3386 trace_sched_switch(preempt, prev, next);
3387 rq = context_switch(rq, prev, next); /* unlocks the rq */
3390 lockdep_unpin_lock(&rq->lock);
3391 raw_spin_unlock_irq(&rq->lock);
3394 balance_callback(rq);
3397 static inline void sched_submit_work(struct task_struct *tsk)
3399 if (!tsk->state || tsk_is_pi_blocked(tsk))
3402 * If we are going to sleep and we have plugged IO queued,
3403 * make sure to submit it to avoid deadlocks.
3405 if (blk_needs_flush_plug(tsk))
3406 blk_schedule_flush_plug(tsk);
3409 asmlinkage __visible void __sched schedule(void)
3411 struct task_struct *tsk = current;
3413 sched_submit_work(tsk);
3417 sched_preempt_enable_no_resched();
3418 } while (need_resched());
3420 EXPORT_SYMBOL(schedule);
3422 #ifdef CONFIG_CONTEXT_TRACKING
3423 asmlinkage __visible void __sched schedule_user(void)
3426 * If we come here after a random call to set_need_resched(),
3427 * or we have been woken up remotely but the IPI has not yet arrived,
3428 * we haven't yet exited the RCU idle mode. Do it here manually until
3429 * we find a better solution.
3431 * NB: There are buggy callers of this function. Ideally we
3432 * should warn if prev_state != CONTEXT_USER, but that will trigger
3433 * too frequently to make sense yet.
3435 enum ctx_state prev_state = exception_enter();
3437 exception_exit(prev_state);
3442 * schedule_preempt_disabled - called with preemption disabled
3444 * Returns with preemption disabled. Note: preempt_count must be 1
3446 void __sched schedule_preempt_disabled(void)
3448 sched_preempt_enable_no_resched();
3453 static void __sched notrace preempt_schedule_common(void)
3456 preempt_disable_notrace();
3458 preempt_enable_no_resched_notrace();
3461 * Check again in case we missed a preemption opportunity
3462 * between schedule and now.
3464 } while (need_resched());
3467 #ifdef CONFIG_PREEMPT
3469 * this is the entry point to schedule() from in-kernel preemption
3470 * off of preempt_enable. Kernel preemptions off return from interrupt
3471 * occur there and call schedule directly.
3473 asmlinkage __visible void __sched notrace preempt_schedule(void)
3476 * If there is a non-zero preempt_count or interrupts are disabled,
3477 * we do not want to preempt the current task. Just return..
3479 if (likely(!preemptible()))
3482 preempt_schedule_common();
3484 NOKPROBE_SYMBOL(preempt_schedule);
3485 EXPORT_SYMBOL(preempt_schedule);
3488 * preempt_schedule_notrace - preempt_schedule called by tracing
3490 * The tracing infrastructure uses preempt_enable_notrace to prevent
3491 * recursion and tracing preempt enabling caused by the tracing
3492 * infrastructure itself. But as tracing can happen in areas coming
3493 * from userspace or just about to enter userspace, a preempt enable
3494 * can occur before user_exit() is called. This will cause the scheduler
3495 * to be called when the system is still in usermode.
3497 * To prevent this, the preempt_enable_notrace will use this function
3498 * instead of preempt_schedule() to exit user context if needed before
3499 * calling the scheduler.
3501 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3503 enum ctx_state prev_ctx;
3505 if (likely(!preemptible()))
3509 preempt_disable_notrace();
3511 * Needs preempt disabled in case user_exit() is traced
3512 * and the tracer calls preempt_enable_notrace() causing
3513 * an infinite recursion.
3515 prev_ctx = exception_enter();
3517 exception_exit(prev_ctx);
3519 preempt_enable_no_resched_notrace();
3520 } while (need_resched());
3522 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3524 #endif /* CONFIG_PREEMPT */
3527 * this is the entry point to schedule() from kernel preemption
3528 * off of irq context.
3529 * Note, that this is called and return with irqs disabled. This will
3530 * protect us against recursive calling from irq.
3532 asmlinkage __visible void __sched preempt_schedule_irq(void)
3534 enum ctx_state prev_state;
3536 /* Catch callers which need to be fixed */
3537 BUG_ON(preempt_count() || !irqs_disabled());
3539 prev_state = exception_enter();
3545 local_irq_disable();
3546 sched_preempt_enable_no_resched();
3547 } while (need_resched());
3549 exception_exit(prev_state);
3552 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3555 return try_to_wake_up(curr->private, mode, wake_flags);
3557 EXPORT_SYMBOL(default_wake_function);
3559 #ifdef CONFIG_RT_MUTEXES
3562 * rt_mutex_setprio - set the current priority of a task
3564 * @prio: prio value (kernel-internal form)
3566 * This function changes the 'effective' priority of a task. It does
3567 * not touch ->normal_prio like __setscheduler().
3569 * Used by the rt_mutex code to implement priority inheritance
3570 * logic. Call site only calls if the priority of the task changed.
3572 void rt_mutex_setprio(struct task_struct *p, int prio)
3574 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3576 const struct sched_class *prev_class;
3578 BUG_ON(prio > MAX_PRIO);
3580 rq = __task_rq_lock(p);
3581 update_rq_clock(rq);
3584 * Idle task boosting is a nono in general. There is one
3585 * exception, when PREEMPT_RT and NOHZ is active:
3587 * The idle task calls get_next_timer_interrupt() and holds
3588 * the timer wheel base->lock on the CPU and another CPU wants
3589 * to access the timer (probably to cancel it). We can safely
3590 * ignore the boosting request, as the idle CPU runs this code
3591 * with interrupts disabled and will complete the lock
3592 * protected section without being interrupted. So there is no
3593 * real need to boost.
3595 if (unlikely(p == rq->idle)) {
3596 WARN_ON(p != rq->curr);
3597 WARN_ON(p->pi_blocked_on);
3601 trace_sched_pi_setprio(p, prio);
3603 prev_class = p->sched_class;
3604 queued = task_on_rq_queued(p);
3605 running = task_current(rq, p);
3607 dequeue_task(rq, p, DEQUEUE_SAVE);
3609 put_prev_task(rq, p);
3612 * Boosting condition are:
3613 * 1. -rt task is running and holds mutex A
3614 * --> -dl task blocks on mutex A
3616 * 2. -dl task is running and holds mutex A
3617 * --> -dl task blocks on mutex A and could preempt the
3620 if (dl_prio(prio)) {
3621 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3622 if (!dl_prio(p->normal_prio) ||
3623 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3624 p->dl.dl_boosted = 1;
3625 enqueue_flag |= ENQUEUE_REPLENISH;
3627 p->dl.dl_boosted = 0;
3628 p->sched_class = &dl_sched_class;
3629 } else if (rt_prio(prio)) {
3630 if (dl_prio(oldprio))
3631 p->dl.dl_boosted = 0;
3633 enqueue_flag |= ENQUEUE_HEAD;
3634 p->sched_class = &rt_sched_class;
3636 if (dl_prio(oldprio))
3637 p->dl.dl_boosted = 0;
3638 if (rt_prio(oldprio))
3640 p->sched_class = &fair_sched_class;
3646 p->sched_class->set_curr_task(rq);
3648 enqueue_task(rq, p, enqueue_flag);
3650 check_class_changed(rq, p, prev_class, oldprio);
3652 preempt_disable(); /* avoid rq from going away on us */
3653 __task_rq_unlock(rq);
3655 balance_callback(rq);
3660 void set_user_nice(struct task_struct *p, long nice)
3662 int old_prio, delta, queued;
3663 unsigned long flags;
3666 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3669 * We have to be careful, if called from sys_setpriority(),
3670 * the task might be in the middle of scheduling on another CPU.
3672 rq = task_rq_lock(p, &flags);
3674 * The RT priorities are set via sched_setscheduler(), but we still
3675 * allow the 'normal' nice value to be set - but as expected
3676 * it wont have any effect on scheduling until the task is
3677 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3679 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3680 p->static_prio = NICE_TO_PRIO(nice);
3683 queued = task_on_rq_queued(p);
3685 dequeue_task(rq, p, DEQUEUE_SAVE);
3687 p->static_prio = NICE_TO_PRIO(nice);
3690 p->prio = effective_prio(p);
3691 delta = p->prio - old_prio;
3694 enqueue_task(rq, p, ENQUEUE_RESTORE);
3696 * If the task increased its priority or is running and
3697 * lowered its priority, then reschedule its CPU:
3699 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3703 task_rq_unlock(rq, p, &flags);
3705 EXPORT_SYMBOL(set_user_nice);
3708 * can_nice - check if a task can reduce its nice value
3712 int can_nice(const struct task_struct *p, const int nice)
3714 /* convert nice value [19,-20] to rlimit style value [1,40] */
3715 int nice_rlim = nice_to_rlimit(nice);
3717 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3718 capable(CAP_SYS_NICE));
3721 #ifdef __ARCH_WANT_SYS_NICE
3724 * sys_nice - change the priority of the current process.
3725 * @increment: priority increment
3727 * sys_setpriority is a more generic, but much slower function that
3728 * does similar things.
3730 SYSCALL_DEFINE1(nice, int, increment)
3735 * Setpriority might change our priority at the same moment.
3736 * We don't have to worry. Conceptually one call occurs first
3737 * and we have a single winner.
3739 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3740 nice = task_nice(current) + increment;
3742 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3743 if (increment < 0 && !can_nice(current, nice))
3746 retval = security_task_setnice(current, nice);
3750 set_user_nice(current, nice);
3757 * task_prio - return the priority value of a given task.
3758 * @p: the task in question.
3760 * Return: The priority value as seen by users in /proc.
3761 * RT tasks are offset by -200. Normal tasks are centered
3762 * around 0, value goes from -16 to +15.
3764 int task_prio(const struct task_struct *p)
3766 return p->prio - MAX_RT_PRIO;
3770 * idle_cpu - is a given cpu idle currently?
3771 * @cpu: the processor in question.
3773 * Return: 1 if the CPU is currently idle. 0 otherwise.
3775 int idle_cpu(int cpu)
3777 struct rq *rq = cpu_rq(cpu);
3779 if (rq->curr != rq->idle)
3786 if (!llist_empty(&rq->wake_list))
3794 * idle_task - return the idle task for a given cpu.
3795 * @cpu: the processor in question.
3797 * Return: The idle task for the cpu @cpu.
3799 struct task_struct *idle_task(int cpu)
3801 return cpu_rq(cpu)->idle;
3805 * find_process_by_pid - find a process with a matching PID value.
3806 * @pid: the pid in question.
3808 * The task of @pid, if found. %NULL otherwise.
3810 static struct task_struct *find_process_by_pid(pid_t pid)
3812 return pid ? find_task_by_vpid(pid) : current;
3816 * This function initializes the sched_dl_entity of a newly becoming
3817 * SCHED_DEADLINE task.
3819 * Only the static values are considered here, the actual runtime and the
3820 * absolute deadline will be properly calculated when the task is enqueued
3821 * for the first time with its new policy.
3824 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3826 struct sched_dl_entity *dl_se = &p->dl;
3828 dl_se->dl_runtime = attr->sched_runtime;
3829 dl_se->dl_deadline = attr->sched_deadline;
3830 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3831 dl_se->flags = attr->sched_flags;
3832 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3835 * Changing the parameters of a task is 'tricky' and we're not doing
3836 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3838 * What we SHOULD do is delay the bandwidth release until the 0-lag
3839 * point. This would include retaining the task_struct until that time
3840 * and change dl_overflow() to not immediately decrement the current
3843 * Instead we retain the current runtime/deadline and let the new
3844 * parameters take effect after the current reservation period lapses.
3845 * This is safe (albeit pessimistic) because the 0-lag point is always
3846 * before the current scheduling deadline.
3848 * We can still have temporary overloads because we do not delay the
3849 * change in bandwidth until that time; so admission control is
3850 * not on the safe side. It does however guarantee tasks will never
3851 * consume more than promised.
3856 * sched_setparam() passes in -1 for its policy, to let the functions
3857 * it calls know not to change it.
3859 #define SETPARAM_POLICY -1
3861 static void __setscheduler_params(struct task_struct *p,
3862 const struct sched_attr *attr)
3864 int policy = attr->sched_policy;
3866 if (policy == SETPARAM_POLICY)
3871 if (dl_policy(policy))
3872 __setparam_dl(p, attr);
3873 else if (fair_policy(policy))
3874 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3877 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3878 * !rt_policy. Always setting this ensures that things like
3879 * getparam()/getattr() don't report silly values for !rt tasks.
3881 p->rt_priority = attr->sched_priority;
3882 p->normal_prio = normal_prio(p);
3886 /* Actually do priority change: must hold pi & rq lock. */
3887 static void __setscheduler(struct rq *rq, struct task_struct *p,
3888 const struct sched_attr *attr, bool keep_boost)
3890 __setscheduler_params(p, attr);
3893 * Keep a potential priority boosting if called from
3894 * sched_setscheduler().
3897 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3899 p->prio = normal_prio(p);
3901 if (dl_prio(p->prio))
3902 p->sched_class = &dl_sched_class;
3903 else if (rt_prio(p->prio))
3904 p->sched_class = &rt_sched_class;
3906 p->sched_class = &fair_sched_class;
3910 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3912 struct sched_dl_entity *dl_se = &p->dl;
3914 attr->sched_priority = p->rt_priority;
3915 attr->sched_runtime = dl_se->dl_runtime;
3916 attr->sched_deadline = dl_se->dl_deadline;
3917 attr->sched_period = dl_se->dl_period;
3918 attr->sched_flags = dl_se->flags;
3922 * This function validates the new parameters of a -deadline task.
3923 * We ask for the deadline not being zero, and greater or equal
3924 * than the runtime, as well as the period of being zero or
3925 * greater than deadline. Furthermore, we have to be sure that
3926 * user parameters are above the internal resolution of 1us (we
3927 * check sched_runtime only since it is always the smaller one) and
3928 * below 2^63 ns (we have to check both sched_deadline and
3929 * sched_period, as the latter can be zero).
3932 __checkparam_dl(const struct sched_attr *attr)
3935 if (attr->sched_deadline == 0)
3939 * Since we truncate DL_SCALE bits, make sure we're at least
3942 if (attr->sched_runtime < (1ULL << DL_SCALE))
3946 * Since we use the MSB for wrap-around and sign issues, make
3947 * sure it's not set (mind that period can be equal to zero).
3949 if (attr->sched_deadline & (1ULL << 63) ||
3950 attr->sched_period & (1ULL << 63))
3953 /* runtime <= deadline <= period (if period != 0) */
3954 if ((attr->sched_period != 0 &&
3955 attr->sched_period < attr->sched_deadline) ||
3956 attr->sched_deadline < attr->sched_runtime)
3963 * check the target process has a UID that matches the current process's
3965 static bool check_same_owner(struct task_struct *p)
3967 const struct cred *cred = current_cred(), *pcred;
3971 pcred = __task_cred(p);
3972 match = (uid_eq(cred->euid, pcred->euid) ||
3973 uid_eq(cred->euid, pcred->uid));
3978 static bool dl_param_changed(struct task_struct *p,
3979 const struct sched_attr *attr)
3981 struct sched_dl_entity *dl_se = &p->dl;
3983 if (dl_se->dl_runtime != attr->sched_runtime ||
3984 dl_se->dl_deadline != attr->sched_deadline ||
3985 dl_se->dl_period != attr->sched_period ||
3986 dl_se->flags != attr->sched_flags)
3992 static int __sched_setscheduler(struct task_struct *p,
3993 const struct sched_attr *attr,
3996 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3997 MAX_RT_PRIO - 1 - attr->sched_priority;
3998 int retval, oldprio, oldpolicy = -1, queued, running;
3999 int new_effective_prio, policy = attr->sched_policy;
4000 unsigned long flags;
4001 const struct sched_class *prev_class;
4005 /* may grab non-irq protected spin_locks */
4006 BUG_ON(in_interrupt());
4008 /* double check policy once rq lock held */
4010 reset_on_fork = p->sched_reset_on_fork;
4011 policy = oldpolicy = p->policy;
4013 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
4015 if (!valid_policy(policy))
4019 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
4023 * Valid priorities for SCHED_FIFO and SCHED_RR are
4024 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4025 * SCHED_BATCH and SCHED_IDLE is 0.
4027 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
4028 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
4030 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
4031 (rt_policy(policy) != (attr->sched_priority != 0)))
4035 * Allow unprivileged RT tasks to decrease priority:
4037 if (user && !capable(CAP_SYS_NICE)) {
4038 if (fair_policy(policy)) {
4039 if (attr->sched_nice < task_nice(p) &&
4040 !can_nice(p, attr->sched_nice))
4044 if (rt_policy(policy)) {
4045 unsigned long rlim_rtprio =
4046 task_rlimit(p, RLIMIT_RTPRIO);
4048 /* can't set/change the rt policy */
4049 if (policy != p->policy && !rlim_rtprio)
4052 /* can't increase priority */
4053 if (attr->sched_priority > p->rt_priority &&
4054 attr->sched_priority > rlim_rtprio)
4059 * Can't set/change SCHED_DEADLINE policy at all for now
4060 * (safest behavior); in the future we would like to allow
4061 * unprivileged DL tasks to increase their relative deadline
4062 * or reduce their runtime (both ways reducing utilization)
4064 if (dl_policy(policy))
4068 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4069 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4071 if (idle_policy(p->policy) && !idle_policy(policy)) {
4072 if (!can_nice(p, task_nice(p)))
4076 /* can't change other user's priorities */
4077 if (!check_same_owner(p))
4080 /* Normal users shall not reset the sched_reset_on_fork flag */
4081 if (p->sched_reset_on_fork && !reset_on_fork)
4086 retval = security_task_setscheduler(p);
4092 * make sure no PI-waiters arrive (or leave) while we are
4093 * changing the priority of the task:
4095 * To be able to change p->policy safely, the appropriate
4096 * runqueue lock must be held.
4098 rq = task_rq_lock(p, &flags);
4099 update_rq_clock(rq);
4102 * Changing the policy of the stop threads its a very bad idea
4104 if (p == rq->stop) {
4105 task_rq_unlock(rq, p, &flags);
4110 * If not changing anything there's no need to proceed further,
4111 * but store a possible modification of reset_on_fork.
4113 if (unlikely(policy == p->policy)) {
4114 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4116 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4118 if (dl_policy(policy) && dl_param_changed(p, attr))
4121 p->sched_reset_on_fork = reset_on_fork;
4122 task_rq_unlock(rq, p, &flags);
4128 #ifdef CONFIG_RT_GROUP_SCHED
4130 * Do not allow realtime tasks into groups that have no runtime
4133 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4134 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4135 !task_group_is_autogroup(task_group(p))) {
4136 task_rq_unlock(rq, p, &flags);
4141 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4142 cpumask_t *span = rq->rd->span;
4145 * Don't allow tasks with an affinity mask smaller than
4146 * the entire root_domain to become SCHED_DEADLINE. We
4147 * will also fail if there's no bandwidth available.
4149 if (!cpumask_subset(span, &p->cpus_allowed) ||
4150 rq->rd->dl_bw.bw == 0) {
4151 task_rq_unlock(rq, p, &flags);
4158 /* recheck policy now with rq lock held */
4159 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4160 policy = oldpolicy = -1;
4161 task_rq_unlock(rq, p, &flags);
4166 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4167 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4170 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4171 task_rq_unlock(rq, p, &flags);
4175 p->sched_reset_on_fork = reset_on_fork;
4180 * Take priority boosted tasks into account. If the new
4181 * effective priority is unchanged, we just store the new
4182 * normal parameters and do not touch the scheduler class and
4183 * the runqueue. This will be done when the task deboost
4186 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4187 if (new_effective_prio == oldprio) {
4188 __setscheduler_params(p, attr);
4189 task_rq_unlock(rq, p, &flags);
4194 queued = task_on_rq_queued(p);
4195 running = task_current(rq, p);
4197 dequeue_task(rq, p, DEQUEUE_SAVE);
4199 put_prev_task(rq, p);
4201 prev_class = p->sched_class;
4202 __setscheduler(rq, p, attr, pi);
4205 p->sched_class->set_curr_task(rq);
4207 int enqueue_flags = ENQUEUE_RESTORE;
4209 * We enqueue to tail when the priority of a task is
4210 * increased (user space view).
4212 if (oldprio <= p->prio)
4213 enqueue_flags |= ENQUEUE_HEAD;
4215 enqueue_task(rq, p, enqueue_flags);
4218 check_class_changed(rq, p, prev_class, oldprio);
4219 preempt_disable(); /* avoid rq from going away on us */
4220 task_rq_unlock(rq, p, &flags);
4223 rt_mutex_adjust_pi(p);
4226 * Run balance callbacks after we've adjusted the PI chain.
4228 balance_callback(rq);
4234 static int _sched_setscheduler(struct task_struct *p, int policy,
4235 const struct sched_param *param, bool check)
4237 struct sched_attr attr = {
4238 .sched_policy = policy,
4239 .sched_priority = param->sched_priority,
4240 .sched_nice = PRIO_TO_NICE(p->static_prio),
4243 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4244 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4245 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4246 policy &= ~SCHED_RESET_ON_FORK;
4247 attr.sched_policy = policy;
4250 return __sched_setscheduler(p, &attr, check, true);
4253 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4254 * @p: the task in question.
4255 * @policy: new policy.
4256 * @param: structure containing the new RT priority.
4258 * Return: 0 on success. An error code otherwise.
4260 * NOTE that the task may be already dead.
4262 int sched_setscheduler(struct task_struct *p, int policy,
4263 const struct sched_param *param)
4265 return _sched_setscheduler(p, policy, param, true);
4267 EXPORT_SYMBOL_GPL(sched_setscheduler);
4269 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4271 return __sched_setscheduler(p, attr, true, true);
4273 EXPORT_SYMBOL_GPL(sched_setattr);
4276 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4277 * @p: the task in question.
4278 * @policy: new policy.
4279 * @param: structure containing the new RT priority.
4281 * Just like sched_setscheduler, only don't bother checking if the
4282 * current context has permission. For example, this is needed in
4283 * stop_machine(): we create temporary high priority worker threads,
4284 * but our caller might not have that capability.
4286 * Return: 0 on success. An error code otherwise.
4288 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4289 const struct sched_param *param)
4291 return _sched_setscheduler(p, policy, param, false);
4293 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4296 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4298 struct sched_param lparam;
4299 struct task_struct *p;
4302 if (!param || pid < 0)
4304 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4309 p = find_process_by_pid(pid);
4311 retval = sched_setscheduler(p, policy, &lparam);
4318 * Mimics kernel/events/core.c perf_copy_attr().
4320 static int sched_copy_attr(struct sched_attr __user *uattr,
4321 struct sched_attr *attr)
4326 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4330 * zero the full structure, so that a short copy will be nice.
4332 memset(attr, 0, sizeof(*attr));
4334 ret = get_user(size, &uattr->size);
4338 if (size > PAGE_SIZE) /* silly large */
4341 if (!size) /* abi compat */
4342 size = SCHED_ATTR_SIZE_VER0;
4344 if (size < SCHED_ATTR_SIZE_VER0)
4348 * If we're handed a bigger struct than we know of,
4349 * ensure all the unknown bits are 0 - i.e. new
4350 * user-space does not rely on any kernel feature
4351 * extensions we dont know about yet.
4353 if (size > sizeof(*attr)) {
4354 unsigned char __user *addr;
4355 unsigned char __user *end;
4358 addr = (void __user *)uattr + sizeof(*attr);
4359 end = (void __user *)uattr + size;
4361 for (; addr < end; addr++) {
4362 ret = get_user(val, addr);
4368 size = sizeof(*attr);
4371 ret = copy_from_user(attr, uattr, size);
4376 * XXX: do we want to be lenient like existing syscalls; or do we want
4377 * to be strict and return an error on out-of-bounds values?
4379 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4384 put_user(sizeof(*attr), &uattr->size);
4389 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4390 * @pid: the pid in question.
4391 * @policy: new policy.
4392 * @param: structure containing the new RT priority.
4394 * Return: 0 on success. An error code otherwise.
4396 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4397 struct sched_param __user *, param)
4399 /* negative values for policy are not valid */
4403 return do_sched_setscheduler(pid, policy, param);
4407 * sys_sched_setparam - set/change the RT priority of a thread
4408 * @pid: the pid in question.
4409 * @param: structure containing the new RT priority.
4411 * Return: 0 on success. An error code otherwise.
4413 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4415 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4419 * sys_sched_setattr - same as above, but with extended sched_attr
4420 * @pid: the pid in question.
4421 * @uattr: structure containing the extended parameters.
4422 * @flags: for future extension.
4424 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4425 unsigned int, flags)
4427 struct sched_attr attr;
4428 struct task_struct *p;
4431 if (!uattr || pid < 0 || flags)
4434 retval = sched_copy_attr(uattr, &attr);
4438 if ((int)attr.sched_policy < 0)
4443 p = find_process_by_pid(pid);
4445 retval = sched_setattr(p, &attr);
4452 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4453 * @pid: the pid in question.
4455 * Return: On success, the policy of the thread. Otherwise, a negative error
4458 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4460 struct task_struct *p;
4468 p = find_process_by_pid(pid);
4470 retval = security_task_getscheduler(p);
4473 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4480 * sys_sched_getparam - get the RT priority of a thread
4481 * @pid: the pid in question.
4482 * @param: structure containing the RT priority.
4484 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4487 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4489 struct sched_param lp = { .sched_priority = 0 };
4490 struct task_struct *p;
4493 if (!param || pid < 0)
4497 p = find_process_by_pid(pid);
4502 retval = security_task_getscheduler(p);
4506 if (task_has_rt_policy(p))
4507 lp.sched_priority = p->rt_priority;
4511 * This one might sleep, we cannot do it with a spinlock held ...
4513 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4522 static int sched_read_attr(struct sched_attr __user *uattr,
4523 struct sched_attr *attr,
4528 if (!access_ok(VERIFY_WRITE, uattr, usize))
4532 * If we're handed a smaller struct than we know of,
4533 * ensure all the unknown bits are 0 - i.e. old
4534 * user-space does not get uncomplete information.
4536 if (usize < sizeof(*attr)) {
4537 unsigned char *addr;
4540 addr = (void *)attr + usize;
4541 end = (void *)attr + sizeof(*attr);
4543 for (; addr < end; addr++) {
4551 ret = copy_to_user(uattr, attr, attr->size);
4559 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4560 * @pid: the pid in question.
4561 * @uattr: structure containing the extended parameters.
4562 * @size: sizeof(attr) for fwd/bwd comp.
4563 * @flags: for future extension.
4565 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4566 unsigned int, size, unsigned int, flags)
4568 struct sched_attr attr = {
4569 .size = sizeof(struct sched_attr),
4571 struct task_struct *p;
4574 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4575 size < SCHED_ATTR_SIZE_VER0 || flags)
4579 p = find_process_by_pid(pid);
4584 retval = security_task_getscheduler(p);
4588 attr.sched_policy = p->policy;
4589 if (p->sched_reset_on_fork)
4590 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4591 if (task_has_dl_policy(p))
4592 __getparam_dl(p, &attr);
4593 else if (task_has_rt_policy(p))
4594 attr.sched_priority = p->rt_priority;
4596 attr.sched_nice = task_nice(p);
4600 retval = sched_read_attr(uattr, &attr, size);
4608 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4610 cpumask_var_t cpus_allowed, new_mask;
4611 struct task_struct *p;
4616 p = find_process_by_pid(pid);
4622 /* Prevent p going away */
4626 if (p->flags & PF_NO_SETAFFINITY) {
4630 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4634 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4636 goto out_free_cpus_allowed;
4639 if (!check_same_owner(p)) {
4641 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4643 goto out_free_new_mask;
4648 retval = security_task_setscheduler(p);
4650 goto out_free_new_mask;
4653 cpuset_cpus_allowed(p, cpus_allowed);
4654 cpumask_and(new_mask, in_mask, cpus_allowed);
4657 * Since bandwidth control happens on root_domain basis,
4658 * if admission test is enabled, we only admit -deadline
4659 * tasks allowed to run on all the CPUs in the task's
4663 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4665 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4668 goto out_free_new_mask;
4674 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4677 cpuset_cpus_allowed(p, cpus_allowed);
4678 if (!cpumask_subset(new_mask, cpus_allowed)) {
4680 * We must have raced with a concurrent cpuset
4681 * update. Just reset the cpus_allowed to the
4682 * cpuset's cpus_allowed
4684 cpumask_copy(new_mask, cpus_allowed);
4689 free_cpumask_var(new_mask);
4690 out_free_cpus_allowed:
4691 free_cpumask_var(cpus_allowed);
4697 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4698 struct cpumask *new_mask)
4700 if (len < cpumask_size())
4701 cpumask_clear(new_mask);
4702 else if (len > cpumask_size())
4703 len = cpumask_size();
4705 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4709 * sys_sched_setaffinity - set the cpu affinity of a process
4710 * @pid: pid of the process
4711 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4712 * @user_mask_ptr: user-space pointer to the new cpu mask
4714 * Return: 0 on success. An error code otherwise.
4716 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4717 unsigned long __user *, user_mask_ptr)
4719 cpumask_var_t new_mask;
4722 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4725 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4727 retval = sched_setaffinity(pid, new_mask);
4728 free_cpumask_var(new_mask);
4732 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4734 struct task_struct *p;
4735 unsigned long flags;
4741 p = find_process_by_pid(pid);
4745 retval = security_task_getscheduler(p);
4749 raw_spin_lock_irqsave(&p->pi_lock, flags);
4750 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4751 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4760 * sys_sched_getaffinity - get the cpu affinity of a process
4761 * @pid: pid of the process
4762 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4763 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4765 * Return: 0 on success. An error code otherwise.
4767 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4768 unsigned long __user *, user_mask_ptr)
4773 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4775 if (len & (sizeof(unsigned long)-1))
4778 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4781 ret = sched_getaffinity(pid, mask);
4783 size_t retlen = min_t(size_t, len, cpumask_size());
4785 if (copy_to_user(user_mask_ptr, mask, retlen))
4790 free_cpumask_var(mask);
4796 * sys_sched_yield - yield the current processor to other threads.
4798 * This function yields the current CPU to other tasks. If there are no
4799 * other threads running on this CPU then this function will return.
4803 SYSCALL_DEFINE0(sched_yield)
4805 struct rq *rq = this_rq_lock();
4807 schedstat_inc(rq, yld_count);
4808 current->sched_class->yield_task(rq);
4811 * Since we are going to call schedule() anyway, there's
4812 * no need to preempt or enable interrupts:
4814 __release(rq->lock);
4815 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4816 do_raw_spin_unlock(&rq->lock);
4817 sched_preempt_enable_no_resched();
4824 int __sched _cond_resched(void)
4826 if (should_resched(0)) {
4827 preempt_schedule_common();
4832 EXPORT_SYMBOL(_cond_resched);
4835 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4836 * call schedule, and on return reacquire the lock.
4838 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4839 * operations here to prevent schedule() from being called twice (once via
4840 * spin_unlock(), once by hand).
4842 int __cond_resched_lock(spinlock_t *lock)
4844 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4847 lockdep_assert_held(lock);
4849 if (spin_needbreak(lock) || resched) {
4852 preempt_schedule_common();
4860 EXPORT_SYMBOL(__cond_resched_lock);
4862 int __sched __cond_resched_softirq(void)
4864 BUG_ON(!in_softirq());
4866 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4868 preempt_schedule_common();
4874 EXPORT_SYMBOL(__cond_resched_softirq);
4877 * yield - yield the current processor to other threads.
4879 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4881 * The scheduler is at all times free to pick the calling task as the most
4882 * eligible task to run, if removing the yield() call from your code breaks
4883 * it, its already broken.
4885 * Typical broken usage is:
4890 * where one assumes that yield() will let 'the other' process run that will
4891 * make event true. If the current task is a SCHED_FIFO task that will never
4892 * happen. Never use yield() as a progress guarantee!!
4894 * If you want to use yield() to wait for something, use wait_event().
4895 * If you want to use yield() to be 'nice' for others, use cond_resched().
4896 * If you still want to use yield(), do not!
4898 void __sched yield(void)
4900 set_current_state(TASK_RUNNING);
4903 EXPORT_SYMBOL(yield);
4906 * yield_to - yield the current processor to another thread in
4907 * your thread group, or accelerate that thread toward the
4908 * processor it's on.
4910 * @preempt: whether task preemption is allowed or not
4912 * It's the caller's job to ensure that the target task struct
4913 * can't go away on us before we can do any checks.
4916 * true (>0) if we indeed boosted the target task.
4917 * false (0) if we failed to boost the target.
4918 * -ESRCH if there's no task to yield to.
4920 int __sched yield_to(struct task_struct *p, bool preempt)
4922 struct task_struct *curr = current;
4923 struct rq *rq, *p_rq;
4924 unsigned long flags;
4927 local_irq_save(flags);
4933 * If we're the only runnable task on the rq and target rq also
4934 * has only one task, there's absolutely no point in yielding.
4936 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4941 double_rq_lock(rq, p_rq);
4942 if (task_rq(p) != p_rq) {
4943 double_rq_unlock(rq, p_rq);
4947 if (!curr->sched_class->yield_to_task)
4950 if (curr->sched_class != p->sched_class)
4953 if (task_running(p_rq, p) || p->state)
4956 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4958 schedstat_inc(rq, yld_count);
4960 * Make p's CPU reschedule; pick_next_entity takes care of
4963 if (preempt && rq != p_rq)
4968 double_rq_unlock(rq, p_rq);
4970 local_irq_restore(flags);
4977 EXPORT_SYMBOL_GPL(yield_to);
4980 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4981 * that process accounting knows that this is a task in IO wait state.
4983 long __sched io_schedule_timeout(long timeout)
4985 int old_iowait = current->in_iowait;
4989 current->in_iowait = 1;
4990 blk_schedule_flush_plug(current);
4992 delayacct_blkio_start();
4994 atomic_inc(&rq->nr_iowait);
4995 ret = schedule_timeout(timeout);
4996 current->in_iowait = old_iowait;
4997 atomic_dec(&rq->nr_iowait);
4998 delayacct_blkio_end();
5002 EXPORT_SYMBOL(io_schedule_timeout);
5005 * sys_sched_get_priority_max - return maximum RT priority.
5006 * @policy: scheduling class.
5008 * Return: On success, this syscall returns the maximum
5009 * rt_priority that can be used by a given scheduling class.
5010 * On failure, a negative error code is returned.
5012 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5019 ret = MAX_USER_RT_PRIO-1;
5021 case SCHED_DEADLINE:
5032 * sys_sched_get_priority_min - return minimum RT priority.
5033 * @policy: scheduling class.
5035 * Return: On success, this syscall returns the minimum
5036 * rt_priority that can be used by a given scheduling class.
5037 * On failure, a negative error code is returned.
5039 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5048 case SCHED_DEADLINE:
5058 * sys_sched_rr_get_interval - return the default timeslice of a process.
5059 * @pid: pid of the process.
5060 * @interval: userspace pointer to the timeslice value.
5062 * this syscall writes the default timeslice value of a given process
5063 * into the user-space timespec buffer. A value of '0' means infinity.
5065 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5068 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5069 struct timespec __user *, interval)
5071 struct task_struct *p;
5072 unsigned int time_slice;
5073 unsigned long flags;
5083 p = find_process_by_pid(pid);
5087 retval = security_task_getscheduler(p);
5091 rq = task_rq_lock(p, &flags);
5093 if (p->sched_class->get_rr_interval)
5094 time_slice = p->sched_class->get_rr_interval(rq, p);
5095 task_rq_unlock(rq, p, &flags);
5098 jiffies_to_timespec(time_slice, &t);
5099 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5107 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5109 void sched_show_task(struct task_struct *p)
5111 unsigned long free = 0;
5113 unsigned long state = p->state;
5116 state = __ffs(state) + 1;
5117 printk(KERN_INFO "%-15.15s %c", p->comm,
5118 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5119 #if BITS_PER_LONG == 32
5120 if (state == TASK_RUNNING)
5121 printk(KERN_CONT " running ");
5123 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5125 if (state == TASK_RUNNING)
5126 printk(KERN_CONT " running task ");
5128 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5130 #ifdef CONFIG_DEBUG_STACK_USAGE
5131 free = stack_not_used(p);
5136 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5138 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5139 task_pid_nr(p), ppid,
5140 (unsigned long)task_thread_info(p)->flags);
5142 print_worker_info(KERN_INFO, p);
5143 show_stack(p, NULL);
5146 void show_state_filter(unsigned long state_filter)
5148 struct task_struct *g, *p;
5150 #if BITS_PER_LONG == 32
5152 " task PC stack pid father\n");
5155 " task PC stack pid father\n");
5158 for_each_process_thread(g, p) {
5160 * reset the NMI-timeout, listing all files on a slow
5161 * console might take a lot of time:
5162 * Also, reset softlockup watchdogs on all CPUs, because
5163 * another CPU might be blocked waiting for us to process
5166 touch_nmi_watchdog();
5167 touch_all_softlockup_watchdogs();
5168 if (!state_filter || (p->state & state_filter))
5172 #ifdef CONFIG_SCHED_DEBUG
5173 sysrq_sched_debug_show();
5177 * Only show locks if all tasks are dumped:
5180 debug_show_all_locks();
5183 void init_idle_bootup_task(struct task_struct *idle)
5185 idle->sched_class = &idle_sched_class;
5189 * init_idle - set up an idle thread for a given CPU
5190 * @idle: task in question
5191 * @cpu: cpu the idle task belongs to
5193 * NOTE: this function does not set the idle thread's NEED_RESCHED
5194 * flag, to make booting more robust.
5196 void init_idle(struct task_struct *idle, int cpu)
5198 struct rq *rq = cpu_rq(cpu);
5199 unsigned long flags;
5201 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5202 raw_spin_lock(&rq->lock);
5204 __sched_fork(0, idle);
5206 idle->state = TASK_RUNNING;
5207 idle->se.exec_start = sched_clock();
5211 * Its possible that init_idle() gets called multiple times on a task,
5212 * in that case do_set_cpus_allowed() will not do the right thing.
5214 * And since this is boot we can forgo the serialization.
5216 set_cpus_allowed_common(idle, cpumask_of(cpu));
5219 * We're having a chicken and egg problem, even though we are
5220 * holding rq->lock, the cpu isn't yet set to this cpu so the
5221 * lockdep check in task_group() will fail.
5223 * Similar case to sched_fork(). / Alternatively we could
5224 * use task_rq_lock() here and obtain the other rq->lock.
5229 __set_task_cpu(idle, cpu);
5232 rq->curr = rq->idle = idle;
5233 idle->on_rq = TASK_ON_RQ_QUEUED;
5237 raw_spin_unlock(&rq->lock);
5238 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5240 /* Set the preempt count _outside_ the spinlocks! */
5241 init_idle_preempt_count(idle, cpu);
5244 * The idle tasks have their own, simple scheduling class:
5246 idle->sched_class = &idle_sched_class;
5247 ftrace_graph_init_idle_task(idle, cpu);
5248 vtime_init_idle(idle, cpu);
5250 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5254 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5255 const struct cpumask *trial)
5257 int ret = 1, trial_cpus;
5258 struct dl_bw *cur_dl_b;
5259 unsigned long flags;
5261 if (!cpumask_weight(cur))
5264 rcu_read_lock_sched();
5265 cur_dl_b = dl_bw_of(cpumask_any(cur));
5266 trial_cpus = cpumask_weight(trial);
5268 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5269 if (cur_dl_b->bw != -1 &&
5270 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5272 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5273 rcu_read_unlock_sched();
5278 int task_can_attach(struct task_struct *p,
5279 const struct cpumask *cs_cpus_allowed)
5284 * Kthreads which disallow setaffinity shouldn't be moved
5285 * to a new cpuset; we don't want to change their cpu
5286 * affinity and isolating such threads by their set of
5287 * allowed nodes is unnecessary. Thus, cpusets are not
5288 * applicable for such threads. This prevents checking for
5289 * success of set_cpus_allowed_ptr() on all attached tasks
5290 * before cpus_allowed may be changed.
5292 if (p->flags & PF_NO_SETAFFINITY) {
5298 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5300 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5305 unsigned long flags;
5307 rcu_read_lock_sched();
5308 dl_b = dl_bw_of(dest_cpu);
5309 raw_spin_lock_irqsave(&dl_b->lock, flags);
5310 cpus = dl_bw_cpus(dest_cpu);
5311 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5316 * We reserve space for this task in the destination
5317 * root_domain, as we can't fail after this point.
5318 * We will free resources in the source root_domain
5319 * later on (see set_cpus_allowed_dl()).
5321 __dl_add(dl_b, p->dl.dl_bw);
5323 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5324 rcu_read_unlock_sched();
5334 #ifdef CONFIG_NUMA_BALANCING
5335 /* Migrate current task p to target_cpu */
5336 int migrate_task_to(struct task_struct *p, int target_cpu)
5338 struct migration_arg arg = { p, target_cpu };
5339 int curr_cpu = task_cpu(p);
5341 if (curr_cpu == target_cpu)
5344 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5347 /* TODO: This is not properly updating schedstats */
5349 trace_sched_move_numa(p, curr_cpu, target_cpu);
5350 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5354 * Requeue a task on a given node and accurately track the number of NUMA
5355 * tasks on the runqueues
5357 void sched_setnuma(struct task_struct *p, int nid)
5360 unsigned long flags;
5361 bool queued, running;
5363 rq = task_rq_lock(p, &flags);
5364 queued = task_on_rq_queued(p);
5365 running = task_current(rq, p);
5368 dequeue_task(rq, p, DEQUEUE_SAVE);
5370 put_prev_task(rq, p);
5372 p->numa_preferred_nid = nid;
5375 p->sched_class->set_curr_task(rq);
5377 enqueue_task(rq, p, ENQUEUE_RESTORE);
5378 task_rq_unlock(rq, p, &flags);
5380 #endif /* CONFIG_NUMA_BALANCING */
5382 #ifdef CONFIG_HOTPLUG_CPU
5384 * Ensures that the idle task is using init_mm right before its cpu goes
5387 void idle_task_exit(void)
5389 struct mm_struct *mm = current->active_mm;
5391 BUG_ON(cpu_online(smp_processor_id()));
5393 if (mm != &init_mm) {
5394 switch_mm(mm, &init_mm, current);
5395 finish_arch_post_lock_switch();
5401 * Since this CPU is going 'away' for a while, fold any nr_active delta
5402 * we might have. Assumes we're called after migrate_tasks() so that the
5403 * nr_active count is stable.
5405 * Also see the comment "Global load-average calculations".
5407 static void calc_load_migrate(struct rq *rq)
5409 long delta = calc_load_fold_active(rq);
5411 atomic_long_add(delta, &calc_load_tasks);
5414 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5418 static const struct sched_class fake_sched_class = {
5419 .put_prev_task = put_prev_task_fake,
5422 static struct task_struct fake_task = {
5424 * Avoid pull_{rt,dl}_task()
5426 .prio = MAX_PRIO + 1,
5427 .sched_class = &fake_sched_class,
5431 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5432 * try_to_wake_up()->select_task_rq().
5434 * Called with rq->lock held even though we'er in stop_machine() and
5435 * there's no concurrency possible, we hold the required locks anyway
5436 * because of lock validation efforts.
5438 static void migrate_tasks(struct rq *dead_rq)
5440 struct rq *rq = dead_rq;
5441 struct task_struct *next, *stop = rq->stop;
5445 * Fudge the rq selection such that the below task selection loop
5446 * doesn't get stuck on the currently eligible stop task.
5448 * We're currently inside stop_machine() and the rq is either stuck
5449 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5450 * either way we should never end up calling schedule() until we're
5456 * put_prev_task() and pick_next_task() sched
5457 * class method both need to have an up-to-date
5458 * value of rq->clock[_task]
5460 update_rq_clock(rq);
5464 * There's this thread running, bail when that's the only
5467 if (rq->nr_running == 1)
5471 * pick_next_task assumes pinned rq->lock.
5473 lockdep_pin_lock(&rq->lock);
5474 next = pick_next_task(rq, &fake_task);
5476 next->sched_class->put_prev_task(rq, next);
5479 * Rules for changing task_struct::cpus_allowed are holding
5480 * both pi_lock and rq->lock, such that holding either
5481 * stabilizes the mask.
5483 * Drop rq->lock is not quite as disastrous as it usually is
5484 * because !cpu_active at this point, which means load-balance
5485 * will not interfere. Also, stop-machine.
5487 lockdep_unpin_lock(&rq->lock);
5488 raw_spin_unlock(&rq->lock);
5489 raw_spin_lock(&next->pi_lock);
5490 raw_spin_lock(&rq->lock);
5493 * Since we're inside stop-machine, _nothing_ should have
5494 * changed the task, WARN if weird stuff happened, because in
5495 * that case the above rq->lock drop is a fail too.
5497 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5498 raw_spin_unlock(&next->pi_lock);
5502 /* Find suitable destination for @next, with force if needed. */
5503 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5505 rq = __migrate_task(rq, next, dest_cpu);
5506 if (rq != dead_rq) {
5507 raw_spin_unlock(&rq->lock);
5509 raw_spin_lock(&rq->lock);
5511 raw_spin_unlock(&next->pi_lock);
5516 #endif /* CONFIG_HOTPLUG_CPU */
5518 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5520 static struct ctl_table sd_ctl_dir[] = {
5522 .procname = "sched_domain",
5528 static struct ctl_table sd_ctl_root[] = {
5530 .procname = "kernel",
5532 .child = sd_ctl_dir,
5537 static struct ctl_table *sd_alloc_ctl_entry(int n)
5539 struct ctl_table *entry =
5540 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5545 static void sd_free_ctl_entry(struct ctl_table **tablep)
5547 struct ctl_table *entry;
5550 * In the intermediate directories, both the child directory and
5551 * procname are dynamically allocated and could fail but the mode
5552 * will always be set. In the lowest directory the names are
5553 * static strings and all have proc handlers.
5555 for (entry = *tablep; entry->mode; entry++) {
5557 sd_free_ctl_entry(&entry->child);
5558 if (entry->proc_handler == NULL)
5559 kfree(entry->procname);
5566 static int min_load_idx = 0;
5567 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5570 set_table_entry(struct ctl_table *entry,
5571 const char *procname, void *data, int maxlen,
5572 umode_t mode, proc_handler *proc_handler,
5575 entry->procname = procname;
5577 entry->maxlen = maxlen;
5579 entry->proc_handler = proc_handler;
5582 entry->extra1 = &min_load_idx;
5583 entry->extra2 = &max_load_idx;
5587 static struct ctl_table *
5588 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5590 struct ctl_table *table = sd_alloc_ctl_entry(5);
5595 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5596 sizeof(int), 0644, proc_dointvec_minmax, false);
5597 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5598 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5599 proc_doulongvec_minmax, false);
5600 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5601 sizeof(int), 0644, proc_dointvec_minmax, false);
5602 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5603 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5604 proc_doulongvec_minmax, false);
5609 static struct ctl_table *
5610 sd_alloc_ctl_group_table(struct sched_group *sg)
5612 struct ctl_table *table = sd_alloc_ctl_entry(2);
5617 table->procname = kstrdup("energy", GFP_KERNEL);
5619 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5624 static struct ctl_table *
5625 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5627 struct ctl_table *table;
5628 unsigned int nr_entries = 14;
5631 struct sched_group *sg = sd->groups;
5636 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5638 nr_entries += nr_sgs;
5641 table = sd_alloc_ctl_entry(nr_entries);
5646 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5647 sizeof(long), 0644, proc_doulongvec_minmax, false);
5648 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5649 sizeof(long), 0644, proc_doulongvec_minmax, false);
5650 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5651 sizeof(int), 0644, proc_dointvec_minmax, true);
5652 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5653 sizeof(int), 0644, proc_dointvec_minmax, true);
5654 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5655 sizeof(int), 0644, proc_dointvec_minmax, true);
5656 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5657 sizeof(int), 0644, proc_dointvec_minmax, true);
5658 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5659 sizeof(int), 0644, proc_dointvec_minmax, true);
5660 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5661 sizeof(int), 0644, proc_dointvec_minmax, false);
5662 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5663 sizeof(int), 0644, proc_dointvec_minmax, false);
5664 set_table_entry(&table[9], "cache_nice_tries",
5665 &sd->cache_nice_tries,
5666 sizeof(int), 0644, proc_dointvec_minmax, false);
5667 set_table_entry(&table[10], "flags", &sd->flags,
5668 sizeof(int), 0644, proc_dointvec_minmax, false);
5669 set_table_entry(&table[11], "max_newidle_lb_cost",
5670 &sd->max_newidle_lb_cost,
5671 sizeof(long), 0644, proc_doulongvec_minmax, false);
5672 set_table_entry(&table[12], "name", sd->name,
5673 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5677 struct ctl_table *entry = &table[13];
5680 snprintf(buf, 32, "group%d", i);
5681 entry->procname = kstrdup(buf, GFP_KERNEL);
5683 entry->child = sd_alloc_ctl_group_table(sg);
5684 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5686 /* &table[nr_entries-1] is terminator */
5691 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5693 struct ctl_table *entry, *table;
5694 struct sched_domain *sd;
5695 int domain_num = 0, i;
5698 for_each_domain(cpu, sd)
5700 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5705 for_each_domain(cpu, sd) {
5706 snprintf(buf, 32, "domain%d", i);
5707 entry->procname = kstrdup(buf, GFP_KERNEL);
5709 entry->child = sd_alloc_ctl_domain_table(sd);
5716 static struct ctl_table_header *sd_sysctl_header;
5717 static void register_sched_domain_sysctl(void)
5719 int i, cpu_num = num_possible_cpus();
5720 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5723 WARN_ON(sd_ctl_dir[0].child);
5724 sd_ctl_dir[0].child = entry;
5729 for_each_possible_cpu(i) {
5730 snprintf(buf, 32, "cpu%d", i);
5731 entry->procname = kstrdup(buf, GFP_KERNEL);
5733 entry->child = sd_alloc_ctl_cpu_table(i);
5737 WARN_ON(sd_sysctl_header);
5738 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5741 /* may be called multiple times per register */
5742 static void unregister_sched_domain_sysctl(void)
5744 unregister_sysctl_table(sd_sysctl_header);
5745 sd_sysctl_header = NULL;
5746 if (sd_ctl_dir[0].child)
5747 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5750 static void register_sched_domain_sysctl(void)
5753 static void unregister_sched_domain_sysctl(void)
5756 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5758 static void set_rq_online(struct rq *rq)
5761 const struct sched_class *class;
5763 cpumask_set_cpu(rq->cpu, rq->rd->online);
5766 for_each_class(class) {
5767 if (class->rq_online)
5768 class->rq_online(rq);
5773 static void set_rq_offline(struct rq *rq)
5776 const struct sched_class *class;
5778 for_each_class(class) {
5779 if (class->rq_offline)
5780 class->rq_offline(rq);
5783 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5789 * migration_call - callback that gets triggered when a CPU is added.
5790 * Here we can start up the necessary migration thread for the new CPU.
5793 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5795 int cpu = (long)hcpu;
5796 unsigned long flags;
5797 struct rq *rq = cpu_rq(cpu);
5799 switch (action & ~CPU_TASKS_FROZEN) {
5801 case CPU_UP_PREPARE:
5802 raw_spin_lock_irqsave(&rq->lock, flags);
5803 walt_set_window_start(rq);
5804 raw_spin_unlock_irqrestore(&rq->lock, flags);
5805 rq->calc_load_update = calc_load_update;
5809 /* Update our root-domain */
5810 raw_spin_lock_irqsave(&rq->lock, flags);
5812 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5816 raw_spin_unlock_irqrestore(&rq->lock, flags);
5819 #ifdef CONFIG_HOTPLUG_CPU
5821 sched_ttwu_pending();
5822 /* Update our root-domain */
5823 raw_spin_lock_irqsave(&rq->lock, flags);
5824 walt_migrate_sync_cpu(cpu);
5826 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5830 BUG_ON(rq->nr_running != 1); /* the migration thread */
5831 raw_spin_unlock_irqrestore(&rq->lock, flags);
5835 calc_load_migrate(rq);
5840 update_max_interval();
5846 * Register at high priority so that task migration (migrate_all_tasks)
5847 * happens before everything else. This has to be lower priority than
5848 * the notifier in the perf_event subsystem, though.
5850 static struct notifier_block migration_notifier = {
5851 .notifier_call = migration_call,
5852 .priority = CPU_PRI_MIGRATION,
5855 static void set_cpu_rq_start_time(void)
5857 int cpu = smp_processor_id();
5858 struct rq *rq = cpu_rq(cpu);
5859 rq->age_stamp = sched_clock_cpu(cpu);
5862 static int sched_cpu_active(struct notifier_block *nfb,
5863 unsigned long action, void *hcpu)
5865 int cpu = (long)hcpu;
5867 switch (action & ~CPU_TASKS_FROZEN) {
5869 set_cpu_rq_start_time();
5874 * At this point a starting CPU has marked itself as online via
5875 * set_cpu_online(). But it might not yet have marked itself
5876 * as active, which is essential from here on.
5878 set_cpu_active(cpu, true);
5879 stop_machine_unpark(cpu);
5882 case CPU_DOWN_FAILED:
5883 set_cpu_active(cpu, true);
5891 static int sched_cpu_inactive(struct notifier_block *nfb,
5892 unsigned long action, void *hcpu)
5894 switch (action & ~CPU_TASKS_FROZEN) {
5895 case CPU_DOWN_PREPARE:
5896 set_cpu_active((long)hcpu, false);
5903 static int __init migration_init(void)
5905 void *cpu = (void *)(long)smp_processor_id();
5908 /* Initialize migration for the boot CPU */
5909 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5910 BUG_ON(err == NOTIFY_BAD);
5911 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5912 register_cpu_notifier(&migration_notifier);
5914 /* Register cpu active notifiers */
5915 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5916 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5920 early_initcall(migration_init);
5922 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5924 #ifdef CONFIG_SCHED_DEBUG
5926 static __read_mostly int sched_debug_enabled;
5928 static int __init sched_debug_setup(char *str)
5930 sched_debug_enabled = 1;
5934 early_param("sched_debug", sched_debug_setup);
5936 static inline bool sched_debug(void)
5938 return sched_debug_enabled;
5941 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5942 struct cpumask *groupmask)
5944 struct sched_group *group = sd->groups;
5946 cpumask_clear(groupmask);
5948 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5950 if (!(sd->flags & SD_LOAD_BALANCE)) {
5951 printk("does not load-balance\n");
5955 printk(KERN_CONT "span %*pbl level %s\n",
5956 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5958 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5959 printk(KERN_ERR "ERROR: domain->span does not contain "
5962 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5963 printk(KERN_ERR "ERROR: domain->groups does not contain"
5967 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5971 printk(KERN_ERR "ERROR: group is NULL\n");
5975 if (!cpumask_weight(sched_group_cpus(group))) {
5976 printk(KERN_CONT "\n");
5977 printk(KERN_ERR "ERROR: empty group\n");
5981 if (!(sd->flags & SD_OVERLAP) &&
5982 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5983 printk(KERN_CONT "\n");
5984 printk(KERN_ERR "ERROR: repeated CPUs\n");
5988 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5990 printk(KERN_CONT " %*pbl",
5991 cpumask_pr_args(sched_group_cpus(group)));
5992 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5993 printk(KERN_CONT " (cpu_capacity = %lu)",
5994 group->sgc->capacity);
5997 group = group->next;
5998 } while (group != sd->groups);
5999 printk(KERN_CONT "\n");
6001 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6002 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6005 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6006 printk(KERN_ERR "ERROR: parent span is not a superset "
6007 "of domain->span\n");
6011 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6015 if (!sched_debug_enabled)
6019 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6023 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6026 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6034 #else /* !CONFIG_SCHED_DEBUG */
6035 # define sched_domain_debug(sd, cpu) do { } while (0)
6036 static inline bool sched_debug(void)
6040 #endif /* CONFIG_SCHED_DEBUG */
6042 static int sd_degenerate(struct sched_domain *sd)
6044 if (cpumask_weight(sched_domain_span(sd)) == 1) {
6045 if (sd->groups->sge)
6046 sd->flags &= ~SD_LOAD_BALANCE;
6051 /* Following flags need at least 2 groups */
6052 if (sd->flags & (SD_LOAD_BALANCE |
6053 SD_BALANCE_NEWIDLE |
6056 SD_SHARE_CPUCAPACITY |
6057 SD_ASYM_CPUCAPACITY |
6058 SD_SHARE_PKG_RESOURCES |
6059 SD_SHARE_POWERDOMAIN |
6060 SD_SHARE_CAP_STATES)) {
6061 if (sd->groups != sd->groups->next)
6065 /* Following flags don't use groups */
6066 if (sd->flags & (SD_WAKE_AFFINE))
6073 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6075 unsigned long cflags = sd->flags, pflags = parent->flags;
6077 if (sd_degenerate(parent))
6080 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6083 /* Flags needing groups don't count if only 1 group in parent */
6084 if (parent->groups == parent->groups->next) {
6085 pflags &= ~(SD_LOAD_BALANCE |
6086 SD_BALANCE_NEWIDLE |
6089 SD_ASYM_CPUCAPACITY |
6090 SD_SHARE_CPUCAPACITY |
6091 SD_SHARE_PKG_RESOURCES |
6093 SD_SHARE_POWERDOMAIN |
6094 SD_SHARE_CAP_STATES);
6095 if (parent->groups->sge) {
6096 parent->flags &= ~SD_LOAD_BALANCE;
6099 if (nr_node_ids == 1)
6100 pflags &= ~SD_SERIALIZE;
6102 if (~cflags & pflags)
6108 static void free_rootdomain(struct rcu_head *rcu)
6110 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6112 cpupri_cleanup(&rd->cpupri);
6113 cpudl_cleanup(&rd->cpudl);
6114 free_cpumask_var(rd->dlo_mask);
6115 free_cpumask_var(rd->rto_mask);
6116 free_cpumask_var(rd->online);
6117 free_cpumask_var(rd->span);
6121 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6123 struct root_domain *old_rd = NULL;
6124 unsigned long flags;
6126 raw_spin_lock_irqsave(&rq->lock, flags);
6131 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6134 cpumask_clear_cpu(rq->cpu, old_rd->span);
6137 * If we dont want to free the old_rd yet then
6138 * set old_rd to NULL to skip the freeing later
6141 if (!atomic_dec_and_test(&old_rd->refcount))
6145 atomic_inc(&rd->refcount);
6148 cpumask_set_cpu(rq->cpu, rd->span);
6149 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6152 raw_spin_unlock_irqrestore(&rq->lock, flags);
6155 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6158 static int init_rootdomain(struct root_domain *rd)
6160 memset(rd, 0, sizeof(*rd));
6162 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6164 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6166 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6168 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6171 init_dl_bw(&rd->dl_bw);
6172 if (cpudl_init(&rd->cpudl) != 0)
6175 if (cpupri_init(&rd->cpupri) != 0)
6178 init_max_cpu_capacity(&rd->max_cpu_capacity);
6180 rd->max_cap_orig_cpu = rd->min_cap_orig_cpu = -1;
6185 free_cpumask_var(rd->rto_mask);
6187 free_cpumask_var(rd->dlo_mask);
6189 free_cpumask_var(rd->online);
6191 free_cpumask_var(rd->span);
6197 * By default the system creates a single root-domain with all cpus as
6198 * members (mimicking the global state we have today).
6200 struct root_domain def_root_domain;
6202 static void init_defrootdomain(void)
6204 init_rootdomain(&def_root_domain);
6206 atomic_set(&def_root_domain.refcount, 1);
6209 static struct root_domain *alloc_rootdomain(void)
6211 struct root_domain *rd;
6213 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6217 if (init_rootdomain(rd) != 0) {
6225 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6227 struct sched_group *tmp, *first;
6236 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6241 } while (sg != first);
6244 static void free_sched_domain(struct rcu_head *rcu)
6246 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6249 * If its an overlapping domain it has private groups, iterate and
6252 if (sd->flags & SD_OVERLAP) {
6253 free_sched_groups(sd->groups, 1);
6254 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6255 kfree(sd->groups->sgc);
6261 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6263 call_rcu(&sd->rcu, free_sched_domain);
6266 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6268 for (; sd; sd = sd->parent)
6269 destroy_sched_domain(sd, cpu);
6273 * Keep a special pointer to the highest sched_domain that has
6274 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6275 * allows us to avoid some pointer chasing select_idle_sibling().
6277 * Also keep a unique ID per domain (we use the first cpu number in
6278 * the cpumask of the domain), this allows us to quickly tell if
6279 * two cpus are in the same cache domain, see cpus_share_cache().
6281 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6282 DEFINE_PER_CPU(int, sd_llc_size);
6283 DEFINE_PER_CPU(int, sd_llc_id);
6284 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6285 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6286 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6287 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6288 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6290 static void update_top_cache_domain(int cpu)
6292 struct sched_domain *sd;
6293 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6297 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6299 id = cpumask_first(sched_domain_span(sd));
6300 size = cpumask_weight(sched_domain_span(sd));
6301 busy_sd = sd->parent; /* sd_busy */
6303 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6305 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6306 per_cpu(sd_llc_size, cpu) = size;
6307 per_cpu(sd_llc_id, cpu) = id;
6309 sd = lowest_flag_domain(cpu, SD_NUMA);
6310 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6312 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6313 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6315 for_each_domain(cpu, sd) {
6316 if (sd->groups->sge)
6321 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6323 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6324 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6328 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6329 * hold the hotplug lock.
6332 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6334 struct rq *rq = cpu_rq(cpu);
6335 struct sched_domain *tmp;
6337 /* Remove the sched domains which do not contribute to scheduling. */
6338 for (tmp = sd; tmp; ) {
6339 struct sched_domain *parent = tmp->parent;
6343 if (sd_parent_degenerate(tmp, parent)) {
6344 tmp->parent = parent->parent;
6346 parent->parent->child = tmp;
6348 * Transfer SD_PREFER_SIBLING down in case of a
6349 * degenerate parent; the spans match for this
6350 * so the property transfers.
6352 if (parent->flags & SD_PREFER_SIBLING)
6353 tmp->flags |= SD_PREFER_SIBLING;
6354 destroy_sched_domain(parent, cpu);
6359 if (sd && sd_degenerate(sd)) {
6362 destroy_sched_domain(tmp, cpu);
6367 sched_domain_debug(sd, cpu);
6369 rq_attach_root(rq, rd);
6371 rcu_assign_pointer(rq->sd, sd);
6372 destroy_sched_domains(tmp, cpu);
6374 update_top_cache_domain(cpu);
6377 /* Setup the mask of cpus configured for isolated domains */
6378 static int __init isolated_cpu_setup(char *str)
6380 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6381 cpulist_parse(str, cpu_isolated_map);
6385 __setup("isolcpus=", isolated_cpu_setup);
6388 struct sched_domain ** __percpu sd;
6389 struct root_domain *rd;
6400 * Build an iteration mask that can exclude certain CPUs from the upwards
6403 * Only CPUs that can arrive at this group should be considered to continue
6406 * Asymmetric node setups can result in situations where the domain tree is of
6407 * unequal depth, make sure to skip domains that already cover the entire
6410 * In that case build_sched_domains() will have terminated the iteration early
6411 * and our sibling sd spans will be empty. Domains should always include the
6412 * cpu they're built on, so check that.
6415 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6417 const struct cpumask *sg_span = sched_group_cpus(sg);
6418 struct sd_data *sdd = sd->private;
6419 struct sched_domain *sibling;
6422 for_each_cpu(i, sg_span) {
6423 sibling = *per_cpu_ptr(sdd->sd, i);
6426 * Can happen in the asymmetric case, where these siblings are
6427 * unused. The mask will not be empty because those CPUs that
6428 * do have the top domain _should_ span the domain.
6430 if (!sibling->child)
6433 /* If we would not end up here, we can't continue from here */
6434 if (!cpumask_equal(sg_span, sched_domain_span(sibling->child)))
6437 cpumask_set_cpu(i, sched_group_mask(sg));
6440 /* We must not have empty masks here */
6441 WARN_ON_ONCE(cpumask_empty(sched_group_mask(sg)));
6445 * Return the canonical balance cpu for this group, this is the first cpu
6446 * of this group that's also in the iteration mask.
6448 int group_balance_cpu(struct sched_group *sg)
6450 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6454 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6456 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6457 const struct cpumask *span = sched_domain_span(sd);
6458 struct cpumask *covered = sched_domains_tmpmask;
6459 struct sd_data *sdd = sd->private;
6460 struct sched_domain *sibling;
6463 cpumask_clear(covered);
6465 for_each_cpu(i, span) {
6466 struct cpumask *sg_span;
6468 if (cpumask_test_cpu(i, covered))
6471 sibling = *per_cpu_ptr(sdd->sd, i);
6473 /* See the comment near build_group_mask(). */
6474 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6477 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6478 GFP_KERNEL, cpu_to_node(cpu));
6483 sg_span = sched_group_cpus(sg);
6485 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6487 cpumask_set_cpu(i, sg_span);
6489 cpumask_or(covered, covered, sg_span);
6491 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6492 if (atomic_inc_return(&sg->sgc->ref) == 1)
6493 build_group_mask(sd, sg);
6496 * Initialize sgc->capacity such that even if we mess up the
6497 * domains and no possible iteration will get us here, we won't
6500 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6501 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6502 sg->sgc->min_capacity = SCHED_CAPACITY_SCALE;
6505 * Make sure the first group of this domain contains the
6506 * canonical balance cpu. Otherwise the sched_domain iteration
6507 * breaks. See update_sg_lb_stats().
6509 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6510 group_balance_cpu(sg) == cpu)
6520 sd->groups = groups;
6525 free_sched_groups(first, 0);
6530 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6532 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6533 struct sched_domain *child = sd->child;
6536 cpu = cpumask_first(sched_domain_span(child));
6539 *sg = *per_cpu_ptr(sdd->sg, cpu);
6540 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6541 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6548 * build_sched_groups will build a circular linked list of the groups
6549 * covered by the given span, and will set each group's ->cpumask correctly,
6550 * and ->cpu_capacity to 0.
6552 * Assumes the sched_domain tree is fully constructed
6555 build_sched_groups(struct sched_domain *sd, int cpu)
6557 struct sched_group *first = NULL, *last = NULL;
6558 struct sd_data *sdd = sd->private;
6559 const struct cpumask *span = sched_domain_span(sd);
6560 struct cpumask *covered;
6563 get_group(cpu, sdd, &sd->groups);
6564 atomic_inc(&sd->groups->ref);
6566 if (cpu != cpumask_first(span))
6569 lockdep_assert_held(&sched_domains_mutex);
6570 covered = sched_domains_tmpmask;
6572 cpumask_clear(covered);
6574 for_each_cpu(i, span) {
6575 struct sched_group *sg;
6578 if (cpumask_test_cpu(i, covered))
6581 group = get_group(i, sdd, &sg);
6582 cpumask_setall(sched_group_mask(sg));
6584 for_each_cpu(j, span) {
6585 if (get_group(j, sdd, NULL) != group)
6588 cpumask_set_cpu(j, covered);
6589 cpumask_set_cpu(j, sched_group_cpus(sg));
6604 * Initialize sched groups cpu_capacity.
6606 * cpu_capacity indicates the capacity of sched group, which is used while
6607 * distributing the load between different sched groups in a sched domain.
6608 * Typically cpu_capacity for all the groups in a sched domain will be same
6609 * unless there are asymmetries in the topology. If there are asymmetries,
6610 * group having more cpu_capacity will pickup more load compared to the
6611 * group having less cpu_capacity.
6613 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6615 struct sched_group *sg = sd->groups;
6620 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6622 } while (sg != sd->groups);
6624 if (cpu != group_balance_cpu(sg))
6627 update_group_capacity(sd, cpu);
6628 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6632 * Check that the per-cpu provided sd energy data is consistent for all cpus
6635 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6636 const struct cpumask *cpumask)
6638 const struct sched_group_energy * const sge = fn(cpu);
6639 struct cpumask mask;
6642 if (cpumask_weight(cpumask) <= 1)
6645 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6647 for_each_cpu(i, &mask) {
6648 const struct sched_group_energy * const e = fn(i);
6651 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6653 for (y = 0; y < (e->nr_idle_states); y++) {
6654 BUG_ON(e->idle_states[y].power !=
6655 sge->idle_states[y].power);
6658 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6660 for (y = 0; y < (e->nr_cap_states); y++) {
6661 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6662 BUG_ON(e->cap_states[y].power !=
6663 sge->cap_states[y].power);
6668 static void init_sched_energy(int cpu, struct sched_domain *sd,
6669 sched_domain_energy_f fn)
6671 if (!(fn && fn(cpu)))
6674 if (cpu != group_balance_cpu(sd->groups))
6677 if (sd->child && !sd->child->groups->sge) {
6678 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6679 #ifdef CONFIG_SCHED_DEBUG
6680 pr_err(" energy data on %s but not on %s domain\n",
6681 sd->name, sd->child->name);
6686 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6688 sd->groups->sge = fn(cpu);
6692 * Initializers for schedule domains
6693 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6696 static int default_relax_domain_level = -1;
6697 int sched_domain_level_max;
6699 static int __init setup_relax_domain_level(char *str)
6701 if (kstrtoint(str, 0, &default_relax_domain_level))
6702 pr_warn("Unable to set relax_domain_level\n");
6706 __setup("relax_domain_level=", setup_relax_domain_level);
6708 static void set_domain_attribute(struct sched_domain *sd,
6709 struct sched_domain_attr *attr)
6713 if (!attr || attr->relax_domain_level < 0) {
6714 if (default_relax_domain_level < 0)
6717 request = default_relax_domain_level;
6719 request = attr->relax_domain_level;
6720 if (request < sd->level) {
6721 /* turn off idle balance on this domain */
6722 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6724 /* turn on idle balance on this domain */
6725 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6729 static void __sdt_free(const struct cpumask *cpu_map);
6730 static int __sdt_alloc(const struct cpumask *cpu_map);
6732 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6733 const struct cpumask *cpu_map)
6737 if (!atomic_read(&d->rd->refcount))
6738 free_rootdomain(&d->rd->rcu); /* fall through */
6740 free_percpu(d->sd); /* fall through */
6742 __sdt_free(cpu_map); /* fall through */
6748 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6749 const struct cpumask *cpu_map)
6751 memset(d, 0, sizeof(*d));
6753 if (__sdt_alloc(cpu_map))
6754 return sa_sd_storage;
6755 d->sd = alloc_percpu(struct sched_domain *);
6757 return sa_sd_storage;
6758 d->rd = alloc_rootdomain();
6761 return sa_rootdomain;
6765 * NULL the sd_data elements we've used to build the sched_domain and
6766 * sched_group structure so that the subsequent __free_domain_allocs()
6767 * will not free the data we're using.
6769 static void claim_allocations(int cpu, struct sched_domain *sd)
6771 struct sd_data *sdd = sd->private;
6773 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6774 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6776 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6777 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6779 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6780 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6784 static int sched_domains_numa_levels;
6785 enum numa_topology_type sched_numa_topology_type;
6786 static int *sched_domains_numa_distance;
6787 int sched_max_numa_distance;
6788 static struct cpumask ***sched_domains_numa_masks;
6789 static int sched_domains_curr_level;
6793 * SD_flags allowed in topology descriptions.
6795 * These flags are purely descriptive of the topology and do not prescribe
6796 * behaviour. Behaviour is artificial and mapped in the below sd_init()
6799 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6800 * SD_SHARE_PKG_RESOURCES - describes shared caches
6801 * SD_NUMA - describes NUMA topologies
6802 * SD_SHARE_POWERDOMAIN - describes shared power domain
6803 * SD_ASYM_CPUCAPACITY - describes mixed capacity topologies
6804 * SD_SHARE_CAP_STATES - describes shared capacity states
6806 * Odd one out, which beside describing the topology has a quirk also
6807 * prescribes the desired behaviour that goes along with it:
6810 * SD_ASYM_PACKING - describes SMT quirks
6812 #define TOPOLOGY_SD_FLAGS \
6813 (SD_SHARE_CPUCAPACITY | \
6814 SD_SHARE_PKG_RESOURCES | \
6817 SD_ASYM_CPUCAPACITY | \
6818 SD_SHARE_POWERDOMAIN | \
6819 SD_SHARE_CAP_STATES)
6821 static struct sched_domain *
6822 sd_init(struct sched_domain_topology_level *tl,
6823 struct sched_domain *child, int cpu)
6825 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6826 int sd_weight, sd_flags = 0;
6830 * Ugly hack to pass state to sd_numa_mask()...
6832 sched_domains_curr_level = tl->numa_level;
6835 sd_weight = cpumask_weight(tl->mask(cpu));
6838 sd_flags = (*tl->sd_flags)();
6839 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6840 "wrong sd_flags in topology description\n"))
6841 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6843 *sd = (struct sched_domain){
6844 .min_interval = sd_weight,
6845 .max_interval = 2*sd_weight,
6847 .imbalance_pct = 125,
6849 .cache_nice_tries = 0,
6856 .flags = 1*SD_LOAD_BALANCE
6857 | 1*SD_BALANCE_NEWIDLE
6862 | 0*SD_SHARE_CPUCAPACITY
6863 | 0*SD_SHARE_PKG_RESOURCES
6865 | 0*SD_PREFER_SIBLING
6870 .last_balance = jiffies,
6871 .balance_interval = sd_weight,
6873 .max_newidle_lb_cost = 0,
6874 .next_decay_max_lb_cost = jiffies,
6876 #ifdef CONFIG_SCHED_DEBUG
6882 * Convert topological properties into behaviour.
6885 if (sd->flags & SD_ASYM_CPUCAPACITY) {
6886 struct sched_domain *t = sd;
6888 for_each_lower_domain(t)
6889 t->flags |= SD_BALANCE_WAKE;
6892 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6893 sd->flags |= SD_PREFER_SIBLING;
6894 sd->imbalance_pct = 110;
6895 sd->smt_gain = 1178; /* ~15% */
6897 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6898 sd->imbalance_pct = 117;
6899 sd->cache_nice_tries = 1;
6903 } else if (sd->flags & SD_NUMA) {
6904 sd->cache_nice_tries = 2;
6908 sd->flags |= SD_SERIALIZE;
6909 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6910 sd->flags &= ~(SD_BALANCE_EXEC |
6917 sd->flags |= SD_PREFER_SIBLING;
6918 sd->cache_nice_tries = 1;
6923 sd->private = &tl->data;
6929 * Topology list, bottom-up.
6931 static struct sched_domain_topology_level default_topology[] = {
6932 #ifdef CONFIG_SCHED_SMT
6933 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6935 #ifdef CONFIG_SCHED_MC
6936 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6938 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6942 static struct sched_domain_topology_level *sched_domain_topology =
6945 #define for_each_sd_topology(tl) \
6946 for (tl = sched_domain_topology; tl->mask; tl++)
6948 void set_sched_topology(struct sched_domain_topology_level *tl)
6950 sched_domain_topology = tl;
6955 static const struct cpumask *sd_numa_mask(int cpu)
6957 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6960 static void sched_numa_warn(const char *str)
6962 static int done = false;
6970 printk(KERN_WARNING "ERROR: %s\n\n", str);
6972 for (i = 0; i < nr_node_ids; i++) {
6973 printk(KERN_WARNING " ");
6974 for (j = 0; j < nr_node_ids; j++)
6975 printk(KERN_CONT "%02d ", node_distance(i,j));
6976 printk(KERN_CONT "\n");
6978 printk(KERN_WARNING "\n");
6981 bool find_numa_distance(int distance)
6985 if (distance == node_distance(0, 0))
6988 for (i = 0; i < sched_domains_numa_levels; i++) {
6989 if (sched_domains_numa_distance[i] == distance)
6997 * A system can have three types of NUMA topology:
6998 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6999 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
7000 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
7002 * The difference between a glueless mesh topology and a backplane
7003 * topology lies in whether communication between not directly
7004 * connected nodes goes through intermediary nodes (where programs
7005 * could run), or through backplane controllers. This affects
7006 * placement of programs.
7008 * The type of topology can be discerned with the following tests:
7009 * - If the maximum distance between any nodes is 1 hop, the system
7010 * is directly connected.
7011 * - If for two nodes A and B, located N > 1 hops away from each other,
7012 * there is an intermediary node C, which is < N hops away from both
7013 * nodes A and B, the system is a glueless mesh.
7015 static void init_numa_topology_type(void)
7019 n = sched_max_numa_distance;
7021 if (sched_domains_numa_levels <= 1) {
7022 sched_numa_topology_type = NUMA_DIRECT;
7026 for_each_online_node(a) {
7027 for_each_online_node(b) {
7028 /* Find two nodes furthest removed from each other. */
7029 if (node_distance(a, b) < n)
7032 /* Is there an intermediary node between a and b? */
7033 for_each_online_node(c) {
7034 if (node_distance(a, c) < n &&
7035 node_distance(b, c) < n) {
7036 sched_numa_topology_type =
7042 sched_numa_topology_type = NUMA_BACKPLANE;
7048 static void sched_init_numa(void)
7050 int next_distance, curr_distance = node_distance(0, 0);
7051 struct sched_domain_topology_level *tl;
7055 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
7056 if (!sched_domains_numa_distance)
7060 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
7061 * unique distances in the node_distance() table.
7063 * Assumes node_distance(0,j) includes all distances in
7064 * node_distance(i,j) in order to avoid cubic time.
7066 next_distance = curr_distance;
7067 for (i = 0; i < nr_node_ids; i++) {
7068 for (j = 0; j < nr_node_ids; j++) {
7069 for (k = 0; k < nr_node_ids; k++) {
7070 int distance = node_distance(i, k);
7072 if (distance > curr_distance &&
7073 (distance < next_distance ||
7074 next_distance == curr_distance))
7075 next_distance = distance;
7078 * While not a strong assumption it would be nice to know
7079 * about cases where if node A is connected to B, B is not
7080 * equally connected to A.
7082 if (sched_debug() && node_distance(k, i) != distance)
7083 sched_numa_warn("Node-distance not symmetric");
7085 if (sched_debug() && i && !find_numa_distance(distance))
7086 sched_numa_warn("Node-0 not representative");
7088 if (next_distance != curr_distance) {
7089 sched_domains_numa_distance[level++] = next_distance;
7090 sched_domains_numa_levels = level;
7091 curr_distance = next_distance;
7096 * In case of sched_debug() we verify the above assumption.
7106 * 'level' contains the number of unique distances, excluding the
7107 * identity distance node_distance(i,i).
7109 * The sched_domains_numa_distance[] array includes the actual distance
7114 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7115 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7116 * the array will contain less then 'level' members. This could be
7117 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7118 * in other functions.
7120 * We reset it to 'level' at the end of this function.
7122 sched_domains_numa_levels = 0;
7124 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7125 if (!sched_domains_numa_masks)
7129 * Now for each level, construct a mask per node which contains all
7130 * cpus of nodes that are that many hops away from us.
7132 for (i = 0; i < level; i++) {
7133 sched_domains_numa_masks[i] =
7134 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7135 if (!sched_domains_numa_masks[i])
7138 for (j = 0; j < nr_node_ids; j++) {
7139 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7143 sched_domains_numa_masks[i][j] = mask;
7146 if (node_distance(j, k) > sched_domains_numa_distance[i])
7149 cpumask_or(mask, mask, cpumask_of_node(k));
7154 /* Compute default topology size */
7155 for (i = 0; sched_domain_topology[i].mask; i++);
7157 tl = kzalloc((i + level + 1) *
7158 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7163 * Copy the default topology bits..
7165 for (i = 0; sched_domain_topology[i].mask; i++)
7166 tl[i] = sched_domain_topology[i];
7169 * .. and append 'j' levels of NUMA goodness.
7171 for (j = 0; j < level; i++, j++) {
7172 tl[i] = (struct sched_domain_topology_level){
7173 .mask = sd_numa_mask,
7174 .sd_flags = cpu_numa_flags,
7175 .flags = SDTL_OVERLAP,
7181 sched_domain_topology = tl;
7183 sched_domains_numa_levels = level;
7184 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7186 init_numa_topology_type();
7189 static void sched_domains_numa_masks_set(int cpu)
7192 int node = cpu_to_node(cpu);
7194 for (i = 0; i < sched_domains_numa_levels; i++) {
7195 for (j = 0; j < nr_node_ids; j++) {
7196 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7197 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7202 static void sched_domains_numa_masks_clear(int cpu)
7205 for (i = 0; i < sched_domains_numa_levels; i++) {
7206 for (j = 0; j < nr_node_ids; j++)
7207 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7212 * Update sched_domains_numa_masks[level][node] array when new cpus
7215 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7216 unsigned long action,
7219 int cpu = (long)hcpu;
7221 switch (action & ~CPU_TASKS_FROZEN) {
7223 sched_domains_numa_masks_set(cpu);
7227 sched_domains_numa_masks_clear(cpu);
7237 static inline void sched_init_numa(void)
7241 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7242 unsigned long action,
7247 #endif /* CONFIG_NUMA */
7249 static int __sdt_alloc(const struct cpumask *cpu_map)
7251 struct sched_domain_topology_level *tl;
7254 for_each_sd_topology(tl) {
7255 struct sd_data *sdd = &tl->data;
7257 sdd->sd = alloc_percpu(struct sched_domain *);
7261 sdd->sg = alloc_percpu(struct sched_group *);
7265 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7269 for_each_cpu(j, cpu_map) {
7270 struct sched_domain *sd;
7271 struct sched_group *sg;
7272 struct sched_group_capacity *sgc;
7274 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7275 GFP_KERNEL, cpu_to_node(j));
7279 *per_cpu_ptr(sdd->sd, j) = sd;
7281 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7282 GFP_KERNEL, cpu_to_node(j));
7288 *per_cpu_ptr(sdd->sg, j) = sg;
7290 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7291 GFP_KERNEL, cpu_to_node(j));
7295 *per_cpu_ptr(sdd->sgc, j) = sgc;
7302 static void __sdt_free(const struct cpumask *cpu_map)
7304 struct sched_domain_topology_level *tl;
7307 for_each_sd_topology(tl) {
7308 struct sd_data *sdd = &tl->data;
7310 for_each_cpu(j, cpu_map) {
7311 struct sched_domain *sd;
7314 sd = *per_cpu_ptr(sdd->sd, j);
7315 if (sd && (sd->flags & SD_OVERLAP))
7316 free_sched_groups(sd->groups, 0);
7317 kfree(*per_cpu_ptr(sdd->sd, j));
7321 kfree(*per_cpu_ptr(sdd->sg, j));
7323 kfree(*per_cpu_ptr(sdd->sgc, j));
7325 free_percpu(sdd->sd);
7327 free_percpu(sdd->sg);
7329 free_percpu(sdd->sgc);
7334 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7335 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7336 struct sched_domain *child, int cpu)
7338 struct sched_domain *sd = sd_init(tl, child, cpu);
7340 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7342 sd->level = child->level + 1;
7343 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7346 if (!cpumask_subset(sched_domain_span(child),
7347 sched_domain_span(sd))) {
7348 pr_err("BUG: arch topology borken\n");
7349 #ifdef CONFIG_SCHED_DEBUG
7350 pr_err(" the %s domain not a subset of the %s domain\n",
7351 child->name, sd->name);
7353 /* Fixup, ensure @sd has at least @child cpus. */
7354 cpumask_or(sched_domain_span(sd),
7355 sched_domain_span(sd),
7356 sched_domain_span(child));
7360 set_domain_attribute(sd, attr);
7366 * Build sched domains for a given set of cpus and attach the sched domains
7367 * to the individual cpus
7369 static int build_sched_domains(const struct cpumask *cpu_map,
7370 struct sched_domain_attr *attr)
7372 enum s_alloc alloc_state;
7373 struct sched_domain *sd;
7375 int i, ret = -ENOMEM;
7377 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7378 if (alloc_state != sa_rootdomain)
7381 /* Set up domains for cpus specified by the cpu_map. */
7382 for_each_cpu(i, cpu_map) {
7383 struct sched_domain_topology_level *tl;
7386 for_each_sd_topology(tl) {
7387 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7388 if (tl == sched_domain_topology)
7389 *per_cpu_ptr(d.sd, i) = sd;
7390 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7391 sd->flags |= SD_OVERLAP;
7395 /* Build the groups for the domains */
7396 for_each_cpu(i, cpu_map) {
7397 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7398 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7399 if (sd->flags & SD_OVERLAP) {
7400 if (build_overlap_sched_groups(sd, i))
7403 if (build_sched_groups(sd, i))
7409 /* Calculate CPU capacity for physical packages and nodes */
7410 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7411 struct sched_domain_topology_level *tl = sched_domain_topology;
7413 if (!cpumask_test_cpu(i, cpu_map))
7416 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7417 init_sched_energy(i, sd, tl->energy);
7418 claim_allocations(i, sd);
7419 init_sched_groups_capacity(i, sd);
7423 /* Attach the domains */
7425 for_each_cpu(i, cpu_map) {
7426 int max_cpu = READ_ONCE(d.rd->max_cap_orig_cpu);
7427 int min_cpu = READ_ONCE(d.rd->min_cap_orig_cpu);
7429 if ((max_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig >
7430 cpu_rq(max_cpu)->cpu_capacity_orig))
7431 WRITE_ONCE(d.rd->max_cap_orig_cpu, i);
7433 if ((min_cpu < 0) || (cpu_rq(i)->cpu_capacity_orig <
7434 cpu_rq(min_cpu)->cpu_capacity_orig))
7435 WRITE_ONCE(d.rd->min_cap_orig_cpu, i);
7437 sd = *per_cpu_ptr(d.sd, i);
7439 cpu_attach_domain(sd, d.rd, i);
7445 __free_domain_allocs(&d, alloc_state, cpu_map);
7449 static cpumask_var_t *doms_cur; /* current sched domains */
7450 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7451 static struct sched_domain_attr *dattr_cur;
7452 /* attribues of custom domains in 'doms_cur' */
7455 * Special case: If a kmalloc of a doms_cur partition (array of
7456 * cpumask) fails, then fallback to a single sched domain,
7457 * as determined by the single cpumask fallback_doms.
7459 static cpumask_var_t fallback_doms;
7462 * arch_update_cpu_topology lets virtualized architectures update the
7463 * cpu core maps. It is supposed to return 1 if the topology changed
7464 * or 0 if it stayed the same.
7466 int __weak arch_update_cpu_topology(void)
7471 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7474 cpumask_var_t *doms;
7476 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7479 for (i = 0; i < ndoms; i++) {
7480 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7481 free_sched_domains(doms, i);
7488 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7491 for (i = 0; i < ndoms; i++)
7492 free_cpumask_var(doms[i]);
7497 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7498 * For now this just excludes isolated cpus, but could be used to
7499 * exclude other special cases in the future.
7501 static int init_sched_domains(const struct cpumask *cpu_map)
7505 arch_update_cpu_topology();
7507 doms_cur = alloc_sched_domains(ndoms_cur);
7509 doms_cur = &fallback_doms;
7510 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7511 err = build_sched_domains(doms_cur[0], NULL);
7512 register_sched_domain_sysctl();
7518 * Detach sched domains from a group of cpus specified in cpu_map
7519 * These cpus will now be attached to the NULL domain
7521 static void detach_destroy_domains(const struct cpumask *cpu_map)
7526 for_each_cpu(i, cpu_map)
7527 cpu_attach_domain(NULL, &def_root_domain, i);
7531 /* handle null as "default" */
7532 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7533 struct sched_domain_attr *new, int idx_new)
7535 struct sched_domain_attr tmp;
7542 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7543 new ? (new + idx_new) : &tmp,
7544 sizeof(struct sched_domain_attr));
7548 * Partition sched domains as specified by the 'ndoms_new'
7549 * cpumasks in the array doms_new[] of cpumasks. This compares
7550 * doms_new[] to the current sched domain partitioning, doms_cur[].
7551 * It destroys each deleted domain and builds each new domain.
7553 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7554 * The masks don't intersect (don't overlap.) We should setup one
7555 * sched domain for each mask. CPUs not in any of the cpumasks will
7556 * not be load balanced. If the same cpumask appears both in the
7557 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7560 * The passed in 'doms_new' should be allocated using
7561 * alloc_sched_domains. This routine takes ownership of it and will
7562 * free_sched_domains it when done with it. If the caller failed the
7563 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7564 * and partition_sched_domains() will fallback to the single partition
7565 * 'fallback_doms', it also forces the domains to be rebuilt.
7567 * If doms_new == NULL it will be replaced with cpu_online_mask.
7568 * ndoms_new == 0 is a special case for destroying existing domains,
7569 * and it will not create the default domain.
7571 * Call with hotplug lock held
7573 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7574 struct sched_domain_attr *dattr_new)
7579 mutex_lock(&sched_domains_mutex);
7581 /* always unregister in case we don't destroy any domains */
7582 unregister_sched_domain_sysctl();
7584 /* Let architecture update cpu core mappings. */
7585 new_topology = arch_update_cpu_topology();
7587 n = doms_new ? ndoms_new : 0;
7589 /* Destroy deleted domains */
7590 for (i = 0; i < ndoms_cur; i++) {
7591 for (j = 0; j < n && !new_topology; j++) {
7592 if (cpumask_equal(doms_cur[i], doms_new[j])
7593 && dattrs_equal(dattr_cur, i, dattr_new, j))
7596 /* no match - a current sched domain not in new doms_new[] */
7597 detach_destroy_domains(doms_cur[i]);
7603 if (doms_new == NULL) {
7605 doms_new = &fallback_doms;
7606 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7607 WARN_ON_ONCE(dattr_new);
7610 /* Build new domains */
7611 for (i = 0; i < ndoms_new; i++) {
7612 for (j = 0; j < n && !new_topology; j++) {
7613 if (cpumask_equal(doms_new[i], doms_cur[j])
7614 && dattrs_equal(dattr_new, i, dattr_cur, j))
7617 /* no match - add a new doms_new */
7618 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7623 /* Remember the new sched domains */
7624 if (doms_cur != &fallback_doms)
7625 free_sched_domains(doms_cur, ndoms_cur);
7626 kfree(dattr_cur); /* kfree(NULL) is safe */
7627 doms_cur = doms_new;
7628 dattr_cur = dattr_new;
7629 ndoms_cur = ndoms_new;
7631 register_sched_domain_sysctl();
7633 mutex_unlock(&sched_domains_mutex);
7636 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7639 * Update cpusets according to cpu_active mask. If cpusets are
7640 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7641 * around partition_sched_domains().
7643 * If we come here as part of a suspend/resume, don't touch cpusets because we
7644 * want to restore it back to its original state upon resume anyway.
7646 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7650 case CPU_ONLINE_FROZEN:
7651 case CPU_DOWN_FAILED_FROZEN:
7654 * num_cpus_frozen tracks how many CPUs are involved in suspend
7655 * resume sequence. As long as this is not the last online
7656 * operation in the resume sequence, just build a single sched
7657 * domain, ignoring cpusets.
7659 partition_sched_domains(1, NULL, NULL);
7660 if (--num_cpus_frozen)
7664 * This is the last CPU online operation. So fall through and
7665 * restore the original sched domains by considering the
7666 * cpuset configurations.
7668 cpuset_force_rebuild();
7671 cpuset_update_active_cpus(true);
7679 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7682 unsigned long flags;
7683 long cpu = (long)hcpu;
7689 case CPU_DOWN_PREPARE:
7690 rcu_read_lock_sched();
7691 dl_b = dl_bw_of(cpu);
7693 raw_spin_lock_irqsave(&dl_b->lock, flags);
7694 cpus = dl_bw_cpus(cpu);
7695 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7696 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7698 rcu_read_unlock_sched();
7701 return notifier_from_errno(-EBUSY);
7702 cpuset_update_active_cpus(false);
7704 case CPU_DOWN_PREPARE_FROZEN:
7706 partition_sched_domains(1, NULL, NULL);
7714 void __init sched_init_smp(void)
7716 cpumask_var_t non_isolated_cpus;
7718 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7719 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7724 * There's no userspace yet to cause hotplug operations; hence all the
7725 * cpu masks are stable and all blatant races in the below code cannot
7728 mutex_lock(&sched_domains_mutex);
7729 init_sched_domains(cpu_active_mask);
7730 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7731 if (cpumask_empty(non_isolated_cpus))
7732 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7733 mutex_unlock(&sched_domains_mutex);
7735 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7736 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7737 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7741 /* Move init over to a non-isolated CPU */
7742 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7744 sched_init_granularity();
7745 free_cpumask_var(non_isolated_cpus);
7747 init_sched_rt_class();
7748 init_sched_dl_class();
7751 void __init sched_init_smp(void)
7753 sched_init_granularity();
7755 #endif /* CONFIG_SMP */
7757 int in_sched_functions(unsigned long addr)
7759 return in_lock_functions(addr) ||
7760 (addr >= (unsigned long)__sched_text_start
7761 && addr < (unsigned long)__sched_text_end);
7764 #ifdef CONFIG_CGROUP_SCHED
7766 * Default task group.
7767 * Every task in system belongs to this group at bootup.
7769 struct task_group root_task_group;
7770 LIST_HEAD(task_groups);
7773 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7775 void __init sched_init(void)
7778 unsigned long alloc_size = 0, ptr;
7780 #ifdef CONFIG_FAIR_GROUP_SCHED
7781 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7783 #ifdef CONFIG_RT_GROUP_SCHED
7784 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7787 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7789 #ifdef CONFIG_FAIR_GROUP_SCHED
7790 root_task_group.se = (struct sched_entity **)ptr;
7791 ptr += nr_cpu_ids * sizeof(void **);
7793 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7794 ptr += nr_cpu_ids * sizeof(void **);
7796 #endif /* CONFIG_FAIR_GROUP_SCHED */
7797 #ifdef CONFIG_RT_GROUP_SCHED
7798 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7799 ptr += nr_cpu_ids * sizeof(void **);
7801 root_task_group.rt_rq = (struct rt_rq **)ptr;
7802 ptr += nr_cpu_ids * sizeof(void **);
7804 #endif /* CONFIG_RT_GROUP_SCHED */
7806 #ifdef CONFIG_CPUMASK_OFFSTACK
7807 for_each_possible_cpu(i) {
7808 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7809 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7811 #endif /* CONFIG_CPUMASK_OFFSTACK */
7813 init_rt_bandwidth(&def_rt_bandwidth,
7814 global_rt_period(), global_rt_runtime());
7815 init_dl_bandwidth(&def_dl_bandwidth,
7816 global_rt_period(), global_rt_runtime());
7819 init_defrootdomain();
7822 #ifdef CONFIG_RT_GROUP_SCHED
7823 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7824 global_rt_period(), global_rt_runtime());
7825 #endif /* CONFIG_RT_GROUP_SCHED */
7827 #ifdef CONFIG_CGROUP_SCHED
7828 list_add(&root_task_group.list, &task_groups);
7829 INIT_LIST_HEAD(&root_task_group.children);
7830 INIT_LIST_HEAD(&root_task_group.siblings);
7831 autogroup_init(&init_task);
7833 #endif /* CONFIG_CGROUP_SCHED */
7835 for_each_possible_cpu(i) {
7839 raw_spin_lock_init(&rq->lock);
7841 rq->calc_load_active = 0;
7842 rq->calc_load_update = jiffies + LOAD_FREQ;
7843 init_cfs_rq(&rq->cfs);
7844 init_rt_rq(&rq->rt);
7845 init_dl_rq(&rq->dl);
7846 #ifdef CONFIG_FAIR_GROUP_SCHED
7847 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7848 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7849 rq->tmp_alone_branch = &rq->leaf_cfs_rq_list;
7851 * How much cpu bandwidth does root_task_group get?
7853 * In case of task-groups formed thr' the cgroup filesystem, it
7854 * gets 100% of the cpu resources in the system. This overall
7855 * system cpu resource is divided among the tasks of
7856 * root_task_group and its child task-groups in a fair manner,
7857 * based on each entity's (task or task-group's) weight
7858 * (se->load.weight).
7860 * In other words, if root_task_group has 10 tasks of weight
7861 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7862 * then A0's share of the cpu resource is:
7864 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7866 * We achieve this by letting root_task_group's tasks sit
7867 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7869 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7870 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7871 #endif /* CONFIG_FAIR_GROUP_SCHED */
7873 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7874 #ifdef CONFIG_RT_GROUP_SCHED
7875 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7878 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7879 rq->cpu_load[j] = 0;
7881 rq->last_load_update_tick = jiffies;
7886 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7887 rq->balance_callback = NULL;
7888 rq->active_balance = 0;
7889 rq->next_balance = jiffies;
7894 rq->avg_idle = 2*sysctl_sched_migration_cost;
7895 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7896 #ifdef CONFIG_SCHED_WALT
7897 rq->cur_irqload = 0;
7898 rq->avg_irqload = 0;
7902 INIT_LIST_HEAD(&rq->cfs_tasks);
7904 rq_attach_root(rq, &def_root_domain);
7905 #ifdef CONFIG_NO_HZ_COMMON
7908 #ifdef CONFIG_NO_HZ_FULL
7909 rq->last_sched_tick = 0;
7913 atomic_set(&rq->nr_iowait, 0);
7916 set_load_weight(&init_task);
7918 #ifdef CONFIG_PREEMPT_NOTIFIERS
7919 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7923 * The boot idle thread does lazy MMU switching as well:
7925 atomic_inc(&init_mm.mm_count);
7926 enter_lazy_tlb(&init_mm, current);
7929 * During early bootup we pretend to be a normal task:
7931 current->sched_class = &fair_sched_class;
7934 * Make us the idle thread. Technically, schedule() should not be
7935 * called from this thread, however somewhere below it might be,
7936 * but because we are the idle thread, we just pick up running again
7937 * when this runqueue becomes "idle".
7939 init_idle(current, smp_processor_id());
7941 calc_load_update = jiffies + LOAD_FREQ;
7944 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7945 /* May be allocated at isolcpus cmdline parse time */
7946 if (cpu_isolated_map == NULL)
7947 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7948 idle_thread_set_boot_cpu();
7949 set_cpu_rq_start_time();
7951 init_sched_fair_class();
7953 scheduler_running = 1;
7956 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7957 static inline int preempt_count_equals(int preempt_offset)
7959 int nested = preempt_count() + rcu_preempt_depth();
7961 return (nested == preempt_offset);
7964 static int __might_sleep_init_called;
7965 int __init __might_sleep_init(void)
7967 __might_sleep_init_called = 1;
7970 early_initcall(__might_sleep_init);
7972 void __might_sleep(const char *file, int line, int preempt_offset)
7975 * Blocking primitives will set (and therefore destroy) current->state,
7976 * since we will exit with TASK_RUNNING make sure we enter with it,
7977 * otherwise we will destroy state.
7979 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7980 "do not call blocking ops when !TASK_RUNNING; "
7981 "state=%lx set at [<%p>] %pS\n",
7983 (void *)current->task_state_change,
7984 (void *)current->task_state_change);
7986 ___might_sleep(file, line, preempt_offset);
7988 EXPORT_SYMBOL(__might_sleep);
7990 void ___might_sleep(const char *file, int line, int preempt_offset)
7992 static unsigned long prev_jiffy; /* ratelimiting */
7994 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7995 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7996 !is_idle_task(current)) || oops_in_progress)
7998 if (system_state != SYSTEM_RUNNING &&
7999 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8001 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8003 prev_jiffy = jiffies;
8006 "BUG: sleeping function called from invalid context at %s:%d\n",
8009 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8010 in_atomic(), irqs_disabled(),
8011 current->pid, current->comm);
8013 if (task_stack_end_corrupted(current))
8014 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
8016 debug_show_held_locks(current);
8017 if (irqs_disabled())
8018 print_irqtrace_events(current);
8019 #ifdef CONFIG_DEBUG_PREEMPT
8020 if (!preempt_count_equals(preempt_offset)) {
8021 pr_err("Preemption disabled at:");
8022 print_ip_sym(current->preempt_disable_ip);
8028 EXPORT_SYMBOL(___might_sleep);
8031 #ifdef CONFIG_MAGIC_SYSRQ
8032 void normalize_rt_tasks(void)
8034 struct task_struct *g, *p;
8035 struct sched_attr attr = {
8036 .sched_policy = SCHED_NORMAL,
8039 read_lock(&tasklist_lock);
8040 for_each_process_thread(g, p) {
8042 * Only normalize user tasks:
8044 if (p->flags & PF_KTHREAD)
8047 p->se.exec_start = 0;
8048 #ifdef CONFIG_SCHEDSTATS
8049 p->se.statistics.wait_start = 0;
8050 p->se.statistics.sleep_start = 0;
8051 p->se.statistics.block_start = 0;
8054 if (!dl_task(p) && !rt_task(p)) {
8056 * Renice negative nice level userspace
8059 if (task_nice(p) < 0)
8060 set_user_nice(p, 0);
8064 __sched_setscheduler(p, &attr, false, false);
8066 read_unlock(&tasklist_lock);
8069 #endif /* CONFIG_MAGIC_SYSRQ */
8071 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8073 * These functions are only useful for the IA64 MCA handling, or kdb.
8075 * They can only be called when the whole system has been
8076 * stopped - every CPU needs to be quiescent, and no scheduling
8077 * activity can take place. Using them for anything else would
8078 * be a serious bug, and as a result, they aren't even visible
8079 * under any other configuration.
8083 * curr_task - return the current task for a given cpu.
8084 * @cpu: the processor in question.
8086 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8088 * Return: The current task for @cpu.
8090 struct task_struct *curr_task(int cpu)
8092 return cpu_curr(cpu);
8095 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8099 * set_curr_task - set the current task for a given cpu.
8100 * @cpu: the processor in question.
8101 * @p: the task pointer to set.
8103 * Description: This function must only be used when non-maskable interrupts
8104 * are serviced on a separate stack. It allows the architecture to switch the
8105 * notion of the current task on a cpu in a non-blocking manner. This function
8106 * must be called with all CPU's synchronized, and interrupts disabled, the
8107 * and caller must save the original value of the current task (see
8108 * curr_task() above) and restore that value before reenabling interrupts and
8109 * re-starting the system.
8111 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8113 void set_curr_task(int cpu, struct task_struct *p)
8120 #ifdef CONFIG_CGROUP_SCHED
8121 /* task_group_lock serializes the addition/removal of task groups */
8122 static DEFINE_SPINLOCK(task_group_lock);
8124 static void sched_free_group(struct task_group *tg)
8126 free_fair_sched_group(tg);
8127 free_rt_sched_group(tg);
8132 /* allocate runqueue etc for a new task group */
8133 struct task_group *sched_create_group(struct task_group *parent)
8135 struct task_group *tg;
8137 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8139 return ERR_PTR(-ENOMEM);
8141 if (!alloc_fair_sched_group(tg, parent))
8144 if (!alloc_rt_sched_group(tg, parent))
8150 sched_free_group(tg);
8151 return ERR_PTR(-ENOMEM);
8154 void sched_online_group(struct task_group *tg, struct task_group *parent)
8156 unsigned long flags;
8158 spin_lock_irqsave(&task_group_lock, flags);
8159 list_add_rcu(&tg->list, &task_groups);
8161 WARN_ON(!parent); /* root should already exist */
8163 tg->parent = parent;
8164 INIT_LIST_HEAD(&tg->children);
8165 list_add_rcu(&tg->siblings, &parent->children);
8166 spin_unlock_irqrestore(&task_group_lock, flags);
8169 /* rcu callback to free various structures associated with a task group */
8170 static void sched_free_group_rcu(struct rcu_head *rhp)
8172 /* now it should be safe to free those cfs_rqs */
8173 sched_free_group(container_of(rhp, struct task_group, rcu));
8176 void sched_destroy_group(struct task_group *tg)
8178 /* wait for possible concurrent references to cfs_rqs complete */
8179 call_rcu(&tg->rcu, sched_free_group_rcu);
8182 void sched_offline_group(struct task_group *tg)
8184 unsigned long flags;
8187 /* end participation in shares distribution */
8188 for_each_possible_cpu(i)
8189 unregister_fair_sched_group(tg, i);
8191 spin_lock_irqsave(&task_group_lock, flags);
8192 list_del_rcu(&tg->list);
8193 list_del_rcu(&tg->siblings);
8194 spin_unlock_irqrestore(&task_group_lock, flags);
8197 static void sched_change_group(struct task_struct *tsk, int type)
8199 struct task_group *tg;
8202 * All callers are synchronized by task_rq_lock(); we do not use RCU
8203 * which is pointless here. Thus, we pass "true" to task_css_check()
8204 * to prevent lockdep warnings.
8206 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8207 struct task_group, css);
8208 tg = autogroup_task_group(tsk, tg);
8209 tsk->sched_task_group = tg;
8211 #ifdef CONFIG_FAIR_GROUP_SCHED
8212 if (tsk->sched_class->task_change_group)
8213 tsk->sched_class->task_change_group(tsk, type);
8216 set_task_rq(tsk, task_cpu(tsk));
8220 * Change task's runqueue when it moves between groups.
8222 * The caller of this function should have put the task in its new group by
8223 * now. This function just updates tsk->se.cfs_rq and tsk->se.parent to reflect
8226 void sched_move_task(struct task_struct *tsk)
8228 int queued, running;
8229 unsigned long flags;
8232 rq = task_rq_lock(tsk, &flags);
8234 running = task_current(rq, tsk);
8235 queued = task_on_rq_queued(tsk);
8238 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8239 if (unlikely(running))
8240 put_prev_task(rq, tsk);
8242 sched_change_group(tsk, TASK_MOVE_GROUP);
8244 if (unlikely(running))
8245 tsk->sched_class->set_curr_task(rq);
8247 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8249 task_rq_unlock(rq, tsk, &flags);
8251 #endif /* CONFIG_CGROUP_SCHED */
8253 #ifdef CONFIG_RT_GROUP_SCHED
8255 * Ensure that the real time constraints are schedulable.
8257 static DEFINE_MUTEX(rt_constraints_mutex);
8259 /* Must be called with tasklist_lock held */
8260 static inline int tg_has_rt_tasks(struct task_group *tg)
8262 struct task_struct *g, *p;
8265 * Autogroups do not have RT tasks; see autogroup_create().
8267 if (task_group_is_autogroup(tg))
8270 for_each_process_thread(g, p) {
8271 if (rt_task(p) && task_group(p) == tg)
8278 struct rt_schedulable_data {
8279 struct task_group *tg;
8284 static int tg_rt_schedulable(struct task_group *tg, void *data)
8286 struct rt_schedulable_data *d = data;
8287 struct task_group *child;
8288 unsigned long total, sum = 0;
8289 u64 period, runtime;
8291 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8292 runtime = tg->rt_bandwidth.rt_runtime;
8295 period = d->rt_period;
8296 runtime = d->rt_runtime;
8300 * Cannot have more runtime than the period.
8302 if (runtime > period && runtime != RUNTIME_INF)
8306 * Ensure we don't starve existing RT tasks.
8308 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8311 total = to_ratio(period, runtime);
8314 * Nobody can have more than the global setting allows.
8316 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8320 * The sum of our children's runtime should not exceed our own.
8322 list_for_each_entry_rcu(child, &tg->children, siblings) {
8323 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8324 runtime = child->rt_bandwidth.rt_runtime;
8326 if (child == d->tg) {
8327 period = d->rt_period;
8328 runtime = d->rt_runtime;
8331 sum += to_ratio(period, runtime);
8340 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8344 struct rt_schedulable_data data = {
8346 .rt_period = period,
8347 .rt_runtime = runtime,
8351 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8357 static int tg_set_rt_bandwidth(struct task_group *tg,
8358 u64 rt_period, u64 rt_runtime)
8363 * Disallowing the root group RT runtime is BAD, it would disallow the
8364 * kernel creating (and or operating) RT threads.
8366 if (tg == &root_task_group && rt_runtime == 0)
8369 /* No period doesn't make any sense. */
8373 mutex_lock(&rt_constraints_mutex);
8374 read_lock(&tasklist_lock);
8375 err = __rt_schedulable(tg, rt_period, rt_runtime);
8379 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8380 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8381 tg->rt_bandwidth.rt_runtime = rt_runtime;
8383 for_each_possible_cpu(i) {
8384 struct rt_rq *rt_rq = tg->rt_rq[i];
8386 raw_spin_lock(&rt_rq->rt_runtime_lock);
8387 rt_rq->rt_runtime = rt_runtime;
8388 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8390 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8392 read_unlock(&tasklist_lock);
8393 mutex_unlock(&rt_constraints_mutex);
8398 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8400 u64 rt_runtime, rt_period;
8402 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8403 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8404 if (rt_runtime_us < 0)
8405 rt_runtime = RUNTIME_INF;
8407 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8410 static long sched_group_rt_runtime(struct task_group *tg)
8414 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8417 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8418 do_div(rt_runtime_us, NSEC_PER_USEC);
8419 return rt_runtime_us;
8422 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8424 u64 rt_runtime, rt_period;
8426 rt_period = rt_period_us * NSEC_PER_USEC;
8427 rt_runtime = tg->rt_bandwidth.rt_runtime;
8429 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8432 static long sched_group_rt_period(struct task_group *tg)
8436 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8437 do_div(rt_period_us, NSEC_PER_USEC);
8438 return rt_period_us;
8440 #endif /* CONFIG_RT_GROUP_SCHED */
8442 #ifdef CONFIG_RT_GROUP_SCHED
8443 static int sched_rt_global_constraints(void)
8447 mutex_lock(&rt_constraints_mutex);
8448 read_lock(&tasklist_lock);
8449 ret = __rt_schedulable(NULL, 0, 0);
8450 read_unlock(&tasklist_lock);
8451 mutex_unlock(&rt_constraints_mutex);
8456 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8458 /* Don't accept realtime tasks when there is no way for them to run */
8459 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8465 #else /* !CONFIG_RT_GROUP_SCHED */
8466 static int sched_rt_global_constraints(void)
8468 unsigned long flags;
8471 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8472 for_each_possible_cpu(i) {
8473 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8475 raw_spin_lock(&rt_rq->rt_runtime_lock);
8476 rt_rq->rt_runtime = global_rt_runtime();
8477 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8479 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8483 #endif /* CONFIG_RT_GROUP_SCHED */
8485 static int sched_dl_global_validate(void)
8487 u64 runtime = global_rt_runtime();
8488 u64 period = global_rt_period();
8489 u64 new_bw = to_ratio(period, runtime);
8492 unsigned long flags;
8495 * Here we want to check the bandwidth not being set to some
8496 * value smaller than the currently allocated bandwidth in
8497 * any of the root_domains.
8499 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8500 * cycling on root_domains... Discussion on different/better
8501 * solutions is welcome!
8503 for_each_possible_cpu(cpu) {
8504 rcu_read_lock_sched();
8505 dl_b = dl_bw_of(cpu);
8507 raw_spin_lock_irqsave(&dl_b->lock, flags);
8508 if (new_bw < dl_b->total_bw)
8510 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8512 rcu_read_unlock_sched();
8521 static void sched_dl_do_global(void)
8526 unsigned long flags;
8528 def_dl_bandwidth.dl_period = global_rt_period();
8529 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8531 if (global_rt_runtime() != RUNTIME_INF)
8532 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8535 * FIXME: As above...
8537 for_each_possible_cpu(cpu) {
8538 rcu_read_lock_sched();
8539 dl_b = dl_bw_of(cpu);
8541 raw_spin_lock_irqsave(&dl_b->lock, flags);
8543 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8545 rcu_read_unlock_sched();
8549 static int sched_rt_global_validate(void)
8551 if (sysctl_sched_rt_period <= 0)
8554 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8555 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8561 static void sched_rt_do_global(void)
8563 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8564 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8567 int sched_rt_handler(struct ctl_table *table, int write,
8568 void __user *buffer, size_t *lenp,
8571 int old_period, old_runtime;
8572 static DEFINE_MUTEX(mutex);
8576 old_period = sysctl_sched_rt_period;
8577 old_runtime = sysctl_sched_rt_runtime;
8579 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8581 if (!ret && write) {
8582 ret = sched_rt_global_validate();
8586 ret = sched_dl_global_validate();
8590 ret = sched_rt_global_constraints();
8594 sched_rt_do_global();
8595 sched_dl_do_global();
8599 sysctl_sched_rt_period = old_period;
8600 sysctl_sched_rt_runtime = old_runtime;
8602 mutex_unlock(&mutex);
8607 int sched_rr_handler(struct ctl_table *table, int write,
8608 void __user *buffer, size_t *lenp,
8612 static DEFINE_MUTEX(mutex);
8615 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8616 /* make sure that internally we keep jiffies */
8617 /* also, writing zero resets timeslice to default */
8618 if (!ret && write) {
8619 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8620 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8622 mutex_unlock(&mutex);
8626 #ifdef CONFIG_CGROUP_SCHED
8628 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8630 return css ? container_of(css, struct task_group, css) : NULL;
8633 static struct cgroup_subsys_state *
8634 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8636 struct task_group *parent = css_tg(parent_css);
8637 struct task_group *tg;
8640 /* This is early initialization for the top cgroup */
8641 return &root_task_group.css;
8644 tg = sched_create_group(parent);
8646 return ERR_PTR(-ENOMEM);
8651 /* Expose task group only after completing cgroup initialization */
8652 static int cpu_cgroup_css_online(struct cgroup_subsys_state *css)
8654 struct task_group *tg = css_tg(css);
8655 struct task_group *parent = css_tg(css->parent);
8658 sched_online_group(tg, parent);
8662 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8664 struct task_group *tg = css_tg(css);
8666 sched_offline_group(tg);
8669 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8671 struct task_group *tg = css_tg(css);
8674 * Relies on the RCU grace period between css_released() and this.
8676 sched_free_group(tg);
8680 * This is called before wake_up_new_task(), therefore we really only
8681 * have to set its group bits, all the other stuff does not apply.
8683 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8685 unsigned long flags;
8688 rq = task_rq_lock(task, &flags);
8690 update_rq_clock(rq);
8691 sched_change_group(task, TASK_SET_GROUP);
8693 task_rq_unlock(rq, task, &flags);
8696 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8698 struct task_struct *task;
8699 struct cgroup_subsys_state *css;
8702 cgroup_taskset_for_each(task, css, tset) {
8703 #ifdef CONFIG_RT_GROUP_SCHED
8704 if (!sched_rt_can_attach(css_tg(css), task))
8707 /* We don't support RT-tasks being in separate groups */
8708 if (task->sched_class != &fair_sched_class)
8712 * Serialize against wake_up_new_task() such that if its
8713 * running, we're sure to observe its full state.
8715 raw_spin_lock_irq(&task->pi_lock);
8717 * Avoid calling sched_move_task() before wake_up_new_task()
8718 * has happened. This would lead to problems with PELT, due to
8719 * move wanting to detach+attach while we're not attached yet.
8721 if (task->state == TASK_NEW)
8723 raw_spin_unlock_irq(&task->pi_lock);
8731 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8733 struct task_struct *task;
8734 struct cgroup_subsys_state *css;
8736 cgroup_taskset_for_each(task, css, tset)
8737 sched_move_task(task);
8740 #ifdef CONFIG_FAIR_GROUP_SCHED
8741 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8742 struct cftype *cftype, u64 shareval)
8744 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8747 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8750 struct task_group *tg = css_tg(css);
8752 return (u64) scale_load_down(tg->shares);
8755 #ifdef CONFIG_CFS_BANDWIDTH
8756 static DEFINE_MUTEX(cfs_constraints_mutex);
8758 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8759 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8761 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8763 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8765 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8766 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8768 if (tg == &root_task_group)
8772 * Ensure we have at some amount of bandwidth every period. This is
8773 * to prevent reaching a state of large arrears when throttled via
8774 * entity_tick() resulting in prolonged exit starvation.
8776 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8780 * Likewise, bound things on the otherside by preventing insane quota
8781 * periods. This also allows us to normalize in computing quota
8784 if (period > max_cfs_quota_period)
8788 * Prevent race between setting of cfs_rq->runtime_enabled and
8789 * unthrottle_offline_cfs_rqs().
8792 mutex_lock(&cfs_constraints_mutex);
8793 ret = __cfs_schedulable(tg, period, quota);
8797 runtime_enabled = quota != RUNTIME_INF;
8798 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8800 * If we need to toggle cfs_bandwidth_used, off->on must occur
8801 * before making related changes, and on->off must occur afterwards
8803 if (runtime_enabled && !runtime_was_enabled)
8804 cfs_bandwidth_usage_inc();
8805 raw_spin_lock_irq(&cfs_b->lock);
8806 cfs_b->period = ns_to_ktime(period);
8807 cfs_b->quota = quota;
8809 __refill_cfs_bandwidth_runtime(cfs_b);
8810 /* restart the period timer (if active) to handle new period expiry */
8811 if (runtime_enabled)
8812 start_cfs_bandwidth(cfs_b);
8813 raw_spin_unlock_irq(&cfs_b->lock);
8815 for_each_online_cpu(i) {
8816 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8817 struct rq *rq = cfs_rq->rq;
8819 raw_spin_lock_irq(&rq->lock);
8820 cfs_rq->runtime_enabled = runtime_enabled;
8821 cfs_rq->runtime_remaining = 0;
8823 if (cfs_rq->throttled)
8824 unthrottle_cfs_rq(cfs_rq);
8825 raw_spin_unlock_irq(&rq->lock);
8827 if (runtime_was_enabled && !runtime_enabled)
8828 cfs_bandwidth_usage_dec();
8830 mutex_unlock(&cfs_constraints_mutex);
8836 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8840 period = ktime_to_ns(tg->cfs_bandwidth.period);
8841 if (cfs_quota_us < 0)
8842 quota = RUNTIME_INF;
8844 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8846 return tg_set_cfs_bandwidth(tg, period, quota);
8849 long tg_get_cfs_quota(struct task_group *tg)
8853 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8856 quota_us = tg->cfs_bandwidth.quota;
8857 do_div(quota_us, NSEC_PER_USEC);
8862 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8866 period = (u64)cfs_period_us * NSEC_PER_USEC;
8867 quota = tg->cfs_bandwidth.quota;
8869 return tg_set_cfs_bandwidth(tg, period, quota);
8872 long tg_get_cfs_period(struct task_group *tg)
8876 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8877 do_div(cfs_period_us, NSEC_PER_USEC);
8879 return cfs_period_us;
8882 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8885 return tg_get_cfs_quota(css_tg(css));
8888 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8889 struct cftype *cftype, s64 cfs_quota_us)
8891 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8894 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8897 return tg_get_cfs_period(css_tg(css));
8900 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8901 struct cftype *cftype, u64 cfs_period_us)
8903 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8906 struct cfs_schedulable_data {
8907 struct task_group *tg;
8912 * normalize group quota/period to be quota/max_period
8913 * note: units are usecs
8915 static u64 normalize_cfs_quota(struct task_group *tg,
8916 struct cfs_schedulable_data *d)
8924 period = tg_get_cfs_period(tg);
8925 quota = tg_get_cfs_quota(tg);
8928 /* note: these should typically be equivalent */
8929 if (quota == RUNTIME_INF || quota == -1)
8932 return to_ratio(period, quota);
8935 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8937 struct cfs_schedulable_data *d = data;
8938 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8939 s64 quota = 0, parent_quota = -1;
8942 quota = RUNTIME_INF;
8944 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8946 quota = normalize_cfs_quota(tg, d);
8947 parent_quota = parent_b->hierarchical_quota;
8950 * ensure max(child_quota) <= parent_quota, inherit when no
8953 if (quota == RUNTIME_INF)
8954 quota = parent_quota;
8955 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8958 cfs_b->hierarchical_quota = quota;
8963 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8966 struct cfs_schedulable_data data = {
8972 if (quota != RUNTIME_INF) {
8973 do_div(data.period, NSEC_PER_USEC);
8974 do_div(data.quota, NSEC_PER_USEC);
8978 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8984 static int cpu_stats_show(struct seq_file *sf, void *v)
8986 struct task_group *tg = css_tg(seq_css(sf));
8987 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8989 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8990 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8991 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8995 #endif /* CONFIG_CFS_BANDWIDTH */
8996 #endif /* CONFIG_FAIR_GROUP_SCHED */
8998 #ifdef CONFIG_RT_GROUP_SCHED
8999 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
9000 struct cftype *cft, s64 val)
9002 return sched_group_set_rt_runtime(css_tg(css), val);
9005 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
9008 return sched_group_rt_runtime(css_tg(css));
9011 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
9012 struct cftype *cftype, u64 rt_period_us)
9014 return sched_group_set_rt_period(css_tg(css), rt_period_us);
9017 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
9020 return sched_group_rt_period(css_tg(css));
9022 #endif /* CONFIG_RT_GROUP_SCHED */
9024 static struct cftype cpu_files[] = {
9025 #ifdef CONFIG_FAIR_GROUP_SCHED
9028 .read_u64 = cpu_shares_read_u64,
9029 .write_u64 = cpu_shares_write_u64,
9032 #ifdef CONFIG_CFS_BANDWIDTH
9034 .name = "cfs_quota_us",
9035 .read_s64 = cpu_cfs_quota_read_s64,
9036 .write_s64 = cpu_cfs_quota_write_s64,
9039 .name = "cfs_period_us",
9040 .read_u64 = cpu_cfs_period_read_u64,
9041 .write_u64 = cpu_cfs_period_write_u64,
9045 .seq_show = cpu_stats_show,
9048 #ifdef CONFIG_RT_GROUP_SCHED
9050 .name = "rt_runtime_us",
9051 .read_s64 = cpu_rt_runtime_read,
9052 .write_s64 = cpu_rt_runtime_write,
9055 .name = "rt_period_us",
9056 .read_u64 = cpu_rt_period_read_uint,
9057 .write_u64 = cpu_rt_period_write_uint,
9063 struct cgroup_subsys cpu_cgrp_subsys = {
9064 .css_alloc = cpu_cgroup_css_alloc,
9065 .css_online = cpu_cgroup_css_online,
9066 .css_released = cpu_cgroup_css_released,
9067 .css_free = cpu_cgroup_css_free,
9068 .fork = cpu_cgroup_fork,
9069 .can_attach = cpu_cgroup_can_attach,
9070 .attach = cpu_cgroup_attach,
9071 .legacy_cftypes = cpu_files,
9075 #endif /* CONFIG_CGROUP_SCHED */
9077 void dump_cpu_task(int cpu)
9079 pr_info("Task dump for CPU %d:\n", cpu);
9080 sched_show_task(cpu_curr(cpu));