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 set_task_cpu(p, cpu);
1335 activate_task(dst_rq, p, 0);
1336 check_preempt_curr(dst_rq, p, 0);
1339 * Task isn't running anymore; make it appear like we migrated
1340 * it before it went to sleep. This means on wakeup we make the
1341 * previous cpu our targer instead of where it really is.
1347 struct migration_swap_arg {
1348 struct task_struct *src_task, *dst_task;
1349 int src_cpu, dst_cpu;
1352 static int migrate_swap_stop(void *data)
1354 struct migration_swap_arg *arg = data;
1355 struct rq *src_rq, *dst_rq;
1358 if (!cpu_active(arg->src_cpu) || !cpu_active(arg->dst_cpu))
1361 src_rq = cpu_rq(arg->src_cpu);
1362 dst_rq = cpu_rq(arg->dst_cpu);
1364 double_raw_lock(&arg->src_task->pi_lock,
1365 &arg->dst_task->pi_lock);
1366 double_rq_lock(src_rq, dst_rq);
1368 if (task_cpu(arg->dst_task) != arg->dst_cpu)
1371 if (task_cpu(arg->src_task) != arg->src_cpu)
1374 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task)))
1377 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task)))
1380 __migrate_swap_task(arg->src_task, arg->dst_cpu);
1381 __migrate_swap_task(arg->dst_task, arg->src_cpu);
1386 double_rq_unlock(src_rq, dst_rq);
1387 raw_spin_unlock(&arg->dst_task->pi_lock);
1388 raw_spin_unlock(&arg->src_task->pi_lock);
1394 * Cross migrate two tasks
1396 int migrate_swap(struct task_struct *cur, struct task_struct *p)
1398 struct migration_swap_arg arg;
1401 arg = (struct migration_swap_arg){
1403 .src_cpu = task_cpu(cur),
1405 .dst_cpu = task_cpu(p),
1408 if (arg.src_cpu == arg.dst_cpu)
1412 * These three tests are all lockless; this is OK since all of them
1413 * will be re-checked with proper locks held further down the line.
1415 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu))
1418 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task)))
1421 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task)))
1424 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu);
1425 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg);
1432 * wait_task_inactive - wait for a thread to unschedule.
1434 * If @match_state is nonzero, it's the @p->state value just checked and
1435 * not expected to change. If it changes, i.e. @p might have woken up,
1436 * then return zero. When we succeed in waiting for @p to be off its CPU,
1437 * we return a positive number (its total switch count). If a second call
1438 * a short while later returns the same number, the caller can be sure that
1439 * @p has remained unscheduled the whole time.
1441 * The caller must ensure that the task *will* unschedule sometime soon,
1442 * else this function might spin for a *long* time. This function can't
1443 * be called with interrupts off, or it may introduce deadlock with
1444 * smp_call_function() if an IPI is sent by the same process we are
1445 * waiting to become inactive.
1447 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1449 unsigned long flags;
1450 int running, queued;
1456 * We do the initial early heuristics without holding
1457 * any task-queue locks at all. We'll only try to get
1458 * the runqueue lock when things look like they will
1464 * If the task is actively running on another CPU
1465 * still, just relax and busy-wait without holding
1468 * NOTE! Since we don't hold any locks, it's not
1469 * even sure that "rq" stays as the right runqueue!
1470 * But we don't care, since "task_running()" will
1471 * return false if the runqueue has changed and p
1472 * is actually now running somewhere else!
1474 while (task_running(rq, p)) {
1475 if (match_state && unlikely(p->state != match_state))
1481 * Ok, time to look more closely! We need the rq
1482 * lock now, to be *sure*. If we're wrong, we'll
1483 * just go back and repeat.
1485 rq = task_rq_lock(p, &flags);
1486 trace_sched_wait_task(p);
1487 running = task_running(rq, p);
1488 queued = task_on_rq_queued(p);
1490 if (!match_state || p->state == match_state)
1491 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1492 task_rq_unlock(rq, p, &flags);
1495 * If it changed from the expected state, bail out now.
1497 if (unlikely(!ncsw))
1501 * Was it really running after all now that we
1502 * checked with the proper locks actually held?
1504 * Oops. Go back and try again..
1506 if (unlikely(running)) {
1512 * It's not enough that it's not actively running,
1513 * it must be off the runqueue _entirely_, and not
1516 * So if it was still runnable (but just not actively
1517 * running right now), it's preempted, and we should
1518 * yield - it could be a while.
1520 if (unlikely(queued)) {
1521 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
1523 set_current_state(TASK_UNINTERRUPTIBLE);
1524 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
1529 * Ahh, all good. It wasn't running, and it wasn't
1530 * runnable, which means that it will never become
1531 * running in the future either. We're all done!
1540 * kick_process - kick a running thread to enter/exit the kernel
1541 * @p: the to-be-kicked thread
1543 * Cause a process which is running on another CPU to enter
1544 * kernel-mode, without any delay. (to get signals handled.)
1546 * NOTE: this function doesn't have to take the runqueue lock,
1547 * because all it wants to ensure is that the remote task enters
1548 * the kernel. If the IPI races and the task has been migrated
1549 * to another CPU then no harm is done and the purpose has been
1552 void kick_process(struct task_struct *p)
1558 if ((cpu != smp_processor_id()) && task_curr(p))
1559 smp_send_reschedule(cpu);
1562 EXPORT_SYMBOL_GPL(kick_process);
1565 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
1567 static int select_fallback_rq(int cpu, struct task_struct *p)
1569 int nid = cpu_to_node(cpu);
1570 const struct cpumask *nodemask = NULL;
1571 enum { cpuset, possible, fail } state = cpuset;
1575 * If the node that the cpu is on has been offlined, cpu_to_node()
1576 * will return -1. There is no cpu on the node, and we should
1577 * select the cpu on the other node.
1580 nodemask = cpumask_of_node(nid);
1582 /* Look for allowed, online CPU in same node. */
1583 for_each_cpu(dest_cpu, nodemask) {
1584 if (!cpu_online(dest_cpu))
1586 if (!cpu_active(dest_cpu))
1588 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p)))
1594 /* Any allowed, online CPU? */
1595 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) {
1596 if (!cpu_online(dest_cpu))
1598 if (!cpu_active(dest_cpu))
1603 /* No more Mr. Nice Guy. */
1606 if (IS_ENABLED(CONFIG_CPUSETS)) {
1607 cpuset_cpus_allowed_fallback(p);
1613 do_set_cpus_allowed(p, cpu_possible_mask);
1624 if (state != cpuset) {
1626 * Don't tell them about moving exiting tasks or
1627 * kernel threads (both mm NULL), since they never
1630 if (p->mm && printk_ratelimit()) {
1631 printk_deferred("process %d (%s) no longer affine to cpu%d\n",
1632 task_pid_nr(p), p->comm, cpu);
1640 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
1643 int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags)
1645 lockdep_assert_held(&p->pi_lock);
1647 if (p->nr_cpus_allowed > 1)
1648 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags);
1651 * In order not to call set_task_cpu() on a blocking task we need
1652 * to rely on ttwu() to place the task on a valid ->cpus_allowed
1655 * Since this is common to all placement strategies, this lives here.
1657 * [ this allows ->select_task() to simply return task_cpu(p) and
1658 * not worry about this generic constraint ]
1660 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) ||
1662 cpu = select_fallback_rq(task_cpu(p), p);
1667 static void update_avg(u64 *avg, u64 sample)
1669 s64 diff = sample - *avg;
1675 static inline int __set_cpus_allowed_ptr(struct task_struct *p,
1676 const struct cpumask *new_mask, bool check)
1678 return set_cpus_allowed_ptr(p, new_mask);
1681 #endif /* CONFIG_SMP */
1684 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
1686 #ifdef CONFIG_SCHEDSTATS
1687 struct rq *rq = this_rq();
1690 int this_cpu = smp_processor_id();
1692 if (cpu == this_cpu) {
1693 schedstat_inc(rq, ttwu_local);
1694 schedstat_inc(p, se.statistics.nr_wakeups_local);
1696 struct sched_domain *sd;
1698 schedstat_inc(p, se.statistics.nr_wakeups_remote);
1700 for_each_domain(this_cpu, sd) {
1701 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
1702 schedstat_inc(sd, ttwu_wake_remote);
1709 if (wake_flags & WF_MIGRATED)
1710 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
1712 #endif /* CONFIG_SMP */
1714 schedstat_inc(rq, ttwu_count);
1715 schedstat_inc(p, se.statistics.nr_wakeups);
1717 if (wake_flags & WF_SYNC)
1718 schedstat_inc(p, se.statistics.nr_wakeups_sync);
1720 #endif /* CONFIG_SCHEDSTATS */
1723 static inline void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
1725 activate_task(rq, p, en_flags);
1726 p->on_rq = TASK_ON_RQ_QUEUED;
1728 /* if a worker is waking up, notify workqueue */
1729 if (p->flags & PF_WQ_WORKER)
1730 wq_worker_waking_up(p, cpu_of(rq));
1734 * Mark the task runnable and perform wakeup-preemption.
1737 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
1739 check_preempt_curr(rq, p, wake_flags);
1740 p->state = TASK_RUNNING;
1741 trace_sched_wakeup(p);
1744 if (p->sched_class->task_woken) {
1746 * Our task @p is fully woken up and running; so its safe to
1747 * drop the rq->lock, hereafter rq is only used for statistics.
1749 lockdep_unpin_lock(&rq->lock);
1750 p->sched_class->task_woken(rq, p);
1751 lockdep_pin_lock(&rq->lock);
1754 if (rq->idle_stamp) {
1755 u64 delta = rq_clock(rq) - rq->idle_stamp;
1756 u64 max = 2*rq->max_idle_balance_cost;
1758 update_avg(&rq->avg_idle, delta);
1760 if (rq->avg_idle > max)
1769 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
1771 lockdep_assert_held(&rq->lock);
1774 if (p->sched_contributes_to_load)
1775 rq->nr_uninterruptible--;
1778 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
1779 ttwu_do_wakeup(rq, p, wake_flags);
1783 * Called in case the task @p isn't fully descheduled from its runqueue,
1784 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
1785 * since all we need to do is flip p->state to TASK_RUNNING, since
1786 * the task is still ->on_rq.
1788 static int ttwu_remote(struct task_struct *p, int wake_flags)
1793 rq = __task_rq_lock(p);
1794 if (task_on_rq_queued(p)) {
1795 /* check_preempt_curr() may use rq clock */
1796 update_rq_clock(rq);
1797 ttwu_do_wakeup(rq, p, wake_flags);
1800 __task_rq_unlock(rq);
1806 void sched_ttwu_pending(void)
1808 struct rq *rq = this_rq();
1809 struct llist_node *llist = llist_del_all(&rq->wake_list);
1810 struct task_struct *p;
1811 unsigned long flags;
1816 raw_spin_lock_irqsave(&rq->lock, flags);
1817 lockdep_pin_lock(&rq->lock);
1820 p = llist_entry(llist, struct task_struct, wake_entry);
1821 llist = llist_next(llist);
1822 ttwu_do_activate(rq, p, 0);
1825 lockdep_unpin_lock(&rq->lock);
1826 raw_spin_unlock_irqrestore(&rq->lock, flags);
1829 void scheduler_ipi(void)
1832 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting
1833 * TIF_NEED_RESCHED remotely (for the first time) will also send
1836 preempt_fold_need_resched();
1838 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick())
1842 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
1843 * traditionally all their work was done from the interrupt return
1844 * path. Now that we actually do some work, we need to make sure
1847 * Some archs already do call them, luckily irq_enter/exit nest
1850 * Arguably we should visit all archs and update all handlers,
1851 * however a fair share of IPIs are still resched only so this would
1852 * somewhat pessimize the simple resched case.
1855 sched_ttwu_pending();
1858 * Check if someone kicked us for doing the nohz idle load balance.
1860 if (unlikely(got_nohz_idle_kick())) {
1861 this_rq()->idle_balance = 1;
1862 raise_softirq_irqoff(SCHED_SOFTIRQ);
1867 static void ttwu_queue_remote(struct task_struct *p, int cpu)
1869 struct rq *rq = cpu_rq(cpu);
1871 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) {
1872 if (!set_nr_if_polling(rq->idle))
1873 smp_send_reschedule(cpu);
1875 trace_sched_wake_idle_without_ipi(cpu);
1879 void wake_up_if_idle(int cpu)
1881 struct rq *rq = cpu_rq(cpu);
1882 unsigned long flags;
1886 if (!is_idle_task(rcu_dereference(rq->curr)))
1889 if (set_nr_if_polling(rq->idle)) {
1890 trace_sched_wake_idle_without_ipi(cpu);
1892 raw_spin_lock_irqsave(&rq->lock, flags);
1893 if (is_idle_task(rq->curr))
1894 smp_send_reschedule(cpu);
1895 /* Else cpu is not in idle, do nothing here */
1896 raw_spin_unlock_irqrestore(&rq->lock, flags);
1903 bool cpus_share_cache(int this_cpu, int that_cpu)
1905 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu);
1907 #endif /* CONFIG_SMP */
1909 static void ttwu_queue(struct task_struct *p, int cpu)
1911 struct rq *rq = cpu_rq(cpu);
1913 #if defined(CONFIG_SMP)
1914 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) {
1915 sched_clock_cpu(cpu); /* sync clocks x-cpu */
1916 ttwu_queue_remote(p, cpu);
1921 raw_spin_lock(&rq->lock);
1922 lockdep_pin_lock(&rq->lock);
1923 ttwu_do_activate(rq, p, 0);
1924 lockdep_unpin_lock(&rq->lock);
1925 raw_spin_unlock(&rq->lock);
1929 * try_to_wake_up - wake up a thread
1930 * @p: the thread to be awakened
1931 * @state: the mask of task states that can be woken
1932 * @wake_flags: wake modifier flags (WF_*)
1934 * Put it on the run-queue if it's not already there. The "current"
1935 * thread is always on the run-queue (except when the actual
1936 * re-schedule is in progress), and as such you're allowed to do
1937 * the simpler "current->state = TASK_RUNNING" to mark yourself
1938 * runnable without the overhead of this.
1940 * Return: %true if @p was woken up, %false if it was already running.
1941 * or @state didn't match @p's state.
1944 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
1946 unsigned long flags;
1947 int cpu, success = 0;
1954 * If we are going to wake up a thread waiting for CONDITION we
1955 * need to ensure that CONDITION=1 done by the caller can not be
1956 * reordered with p->state check below. This pairs with mb() in
1957 * set_current_state() the waiting thread does.
1959 smp_mb__before_spinlock();
1960 raw_spin_lock_irqsave(&p->pi_lock, flags);
1961 if (!(p->state & state))
1964 trace_sched_waking(p);
1966 success = 1; /* we're going to change ->state */
1969 if (p->on_rq && ttwu_remote(p, wake_flags))
1974 * Ensure we load p->on_cpu _after_ p->on_rq, otherwise it would be
1975 * possible to, falsely, observe p->on_cpu == 0.
1977 * One must be running (->on_cpu == 1) in order to remove oneself
1978 * from the runqueue.
1980 * [S] ->on_cpu = 1; [L] ->on_rq
1984 * [S] ->on_rq = 0; [L] ->on_cpu
1986 * Pairs with the full barrier implied in the UNLOCK+LOCK on rq->lock
1987 * from the consecutive calls to schedule(); the first switching to our
1988 * task, the second putting it to sleep.
1993 * If the owning (remote) cpu is still in the middle of schedule() with
1994 * this task as prev, wait until its done referencing the task.
1999 * Combined with the control dependency above, we have an effective
2000 * smp_load_acquire() without the need for full barriers.
2002 * Pairs with the smp_store_release() in finish_lock_switch().
2004 * This ensures that tasks getting woken will be fully ordered against
2005 * their previous state and preserve Program Order.
2009 rq = cpu_rq(task_cpu(p));
2011 raw_spin_lock(&rq->lock);
2012 wallclock = walt_ktime_clock();
2013 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2014 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2015 raw_spin_unlock(&rq->lock);
2017 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2018 p->state = TASK_WAKING;
2020 if (p->sched_class->task_waking)
2021 p->sched_class->task_waking(p);
2023 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags);
2025 if (task_cpu(p) != cpu) {
2026 wake_flags |= WF_MIGRATED;
2027 set_task_cpu(p, cpu);
2030 #endif /* CONFIG_SMP */
2034 ttwu_stat(p, cpu, wake_flags);
2036 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2042 * try_to_wake_up_local - try to wake up a local task with rq lock held
2043 * @p: the thread to be awakened
2045 * Put @p on the run-queue if it's not already there. The caller must
2046 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2049 static void try_to_wake_up_local(struct task_struct *p)
2051 struct rq *rq = task_rq(p);
2053 if (WARN_ON_ONCE(rq != this_rq()) ||
2054 WARN_ON_ONCE(p == current))
2057 lockdep_assert_held(&rq->lock);
2059 if (!raw_spin_trylock(&p->pi_lock)) {
2061 * This is OK, because current is on_cpu, which avoids it being
2062 * picked for load-balance and preemption/IRQs are still
2063 * disabled avoiding further scheduler activity on it and we've
2064 * not yet picked a replacement task.
2066 lockdep_unpin_lock(&rq->lock);
2067 raw_spin_unlock(&rq->lock);
2068 raw_spin_lock(&p->pi_lock);
2069 raw_spin_lock(&rq->lock);
2070 lockdep_pin_lock(&rq->lock);
2073 if (!(p->state & TASK_NORMAL))
2076 trace_sched_waking(p);
2078 if (!task_on_rq_queued(p)) {
2079 u64 wallclock = walt_ktime_clock();
2081 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE, wallclock, 0);
2082 walt_update_task_ravg(p, rq, TASK_WAKE, wallclock, 0);
2083 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2086 ttwu_do_wakeup(rq, p, 0);
2087 ttwu_stat(p, smp_processor_id(), 0);
2089 raw_spin_unlock(&p->pi_lock);
2093 * wake_up_process - Wake up a specific process
2094 * @p: The process to be woken up.
2096 * Attempt to wake up the nominated process and move it to the set of runnable
2099 * Return: 1 if the process was woken up, 0 if it was already running.
2101 * It may be assumed that this function implies a write memory barrier before
2102 * changing the task state if and only if any tasks are woken up.
2104 int wake_up_process(struct task_struct *p)
2106 return try_to_wake_up(p, TASK_NORMAL, 0);
2108 EXPORT_SYMBOL(wake_up_process);
2110 int wake_up_state(struct task_struct *p, unsigned int state)
2112 return try_to_wake_up(p, state, 0);
2116 * This function clears the sched_dl_entity static params.
2118 void __dl_clear_params(struct task_struct *p)
2120 struct sched_dl_entity *dl_se = &p->dl;
2122 dl_se->dl_runtime = 0;
2123 dl_se->dl_deadline = 0;
2124 dl_se->dl_period = 0;
2128 dl_se->dl_throttled = 0;
2130 dl_se->dl_yielded = 0;
2134 * Perform scheduler related setup for a newly forked process p.
2135 * p is forked by current.
2137 * __sched_fork() is basic setup used by init_idle() too:
2139 static void __sched_fork(unsigned long clone_flags, struct task_struct *p)
2144 p->se.exec_start = 0;
2145 p->se.sum_exec_runtime = 0;
2146 p->se.prev_sum_exec_runtime = 0;
2147 p->se.nr_migrations = 0;
2149 INIT_LIST_HEAD(&p->se.group_node);
2150 walt_init_new_task_load(p);
2152 #ifdef CONFIG_SCHEDSTATS
2153 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2156 RB_CLEAR_NODE(&p->dl.rb_node);
2157 init_dl_task_timer(&p->dl);
2158 __dl_clear_params(p);
2160 INIT_LIST_HEAD(&p->rt.run_list);
2162 #ifdef CONFIG_PREEMPT_NOTIFIERS
2163 INIT_HLIST_HEAD(&p->preempt_notifiers);
2166 #ifdef CONFIG_NUMA_BALANCING
2167 if (p->mm && atomic_read(&p->mm->mm_users) == 1) {
2168 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay);
2169 p->mm->numa_scan_seq = 0;
2172 if (clone_flags & CLONE_VM)
2173 p->numa_preferred_nid = current->numa_preferred_nid;
2175 p->numa_preferred_nid = -1;
2177 p->node_stamp = 0ULL;
2178 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0;
2179 p->numa_scan_period = sysctl_numa_balancing_scan_delay;
2180 p->numa_work.next = &p->numa_work;
2181 p->numa_faults = NULL;
2182 p->last_task_numa_placement = 0;
2183 p->last_sum_exec_runtime = 0;
2185 p->numa_group = NULL;
2186 #endif /* CONFIG_NUMA_BALANCING */
2189 DEFINE_STATIC_KEY_FALSE(sched_numa_balancing);
2191 #ifdef CONFIG_NUMA_BALANCING
2193 void set_numabalancing_state(bool enabled)
2196 static_branch_enable(&sched_numa_balancing);
2198 static_branch_disable(&sched_numa_balancing);
2201 #ifdef CONFIG_PROC_SYSCTL
2202 int sysctl_numa_balancing(struct ctl_table *table, int write,
2203 void __user *buffer, size_t *lenp, loff_t *ppos)
2207 int state = static_branch_likely(&sched_numa_balancing);
2209 if (write && !capable(CAP_SYS_ADMIN))
2214 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos);
2218 set_numabalancing_state(state);
2225 * fork()/clone()-time setup:
2227 int sched_fork(unsigned long clone_flags, struct task_struct *p)
2229 unsigned long flags;
2230 int cpu = get_cpu();
2232 __sched_fork(clone_flags, p);
2234 * We mark the process as running here. This guarantees that
2235 * nobody will actually run it, and a signal or other external
2236 * event cannot wake it up and insert it on the runqueue either.
2238 p->state = TASK_RUNNING;
2241 * Make sure we do not leak PI boosting priority to the child.
2243 p->prio = current->normal_prio;
2246 * Revert to default priority/policy on fork if requested.
2248 if (unlikely(p->sched_reset_on_fork)) {
2249 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
2250 p->policy = SCHED_NORMAL;
2251 p->static_prio = NICE_TO_PRIO(0);
2253 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2254 p->static_prio = NICE_TO_PRIO(0);
2256 p->prio = p->normal_prio = __normal_prio(p);
2260 * We don't need the reset flag anymore after the fork. It has
2261 * fulfilled its duty:
2263 p->sched_reset_on_fork = 0;
2266 if (dl_prio(p->prio)) {
2269 } else if (rt_prio(p->prio)) {
2270 p->sched_class = &rt_sched_class;
2272 p->sched_class = &fair_sched_class;
2275 if (p->sched_class->task_fork)
2276 p->sched_class->task_fork(p);
2279 * The child is not yet in the pid-hash so no cgroup attach races,
2280 * and the cgroup is pinned to this child due to cgroup_fork()
2281 * is ran before sched_fork().
2283 * Silence PROVE_RCU.
2285 raw_spin_lock_irqsave(&p->pi_lock, flags);
2286 set_task_cpu(p, cpu);
2287 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2289 #ifdef CONFIG_SCHED_INFO
2290 if (likely(sched_info_on()))
2291 memset(&p->sched_info, 0, sizeof(p->sched_info));
2293 #if defined(CONFIG_SMP)
2296 init_task_preempt_count(p);
2298 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2299 RB_CLEAR_NODE(&p->pushable_dl_tasks);
2306 unsigned long to_ratio(u64 period, u64 runtime)
2308 if (runtime == RUNTIME_INF)
2312 * Doing this here saves a lot of checks in all
2313 * the calling paths, and returning zero seems
2314 * safe for them anyway.
2319 return div64_u64(runtime << 20, period);
2323 inline struct dl_bw *dl_bw_of(int i)
2325 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2326 "sched RCU must be held");
2327 return &cpu_rq(i)->rd->dl_bw;
2330 static inline int dl_bw_cpus(int i)
2332 struct root_domain *rd = cpu_rq(i)->rd;
2335 RCU_LOCKDEP_WARN(!rcu_read_lock_sched_held(),
2336 "sched RCU must be held");
2337 for_each_cpu_and(i, rd->span, cpu_active_mask)
2343 inline struct dl_bw *dl_bw_of(int i)
2345 return &cpu_rq(i)->dl.dl_bw;
2348 static inline int dl_bw_cpus(int i)
2355 * We must be sure that accepting a new task (or allowing changing the
2356 * parameters of an existing one) is consistent with the bandwidth
2357 * constraints. If yes, this function also accordingly updates the currently
2358 * allocated bandwidth to reflect the new situation.
2360 * This function is called while holding p's rq->lock.
2362 * XXX we should delay bw change until the task's 0-lag point, see
2365 static int dl_overflow(struct task_struct *p, int policy,
2366 const struct sched_attr *attr)
2369 struct dl_bw *dl_b = dl_bw_of(task_cpu(p));
2370 u64 period = attr->sched_period ?: attr->sched_deadline;
2371 u64 runtime = attr->sched_runtime;
2372 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0;
2375 if (new_bw == p->dl.dl_bw)
2379 * Either if a task, enters, leave, or stays -deadline but changes
2380 * its parameters, we may need to update accordingly the total
2381 * allocated bandwidth of the container.
2383 raw_spin_lock(&dl_b->lock);
2384 cpus = dl_bw_cpus(task_cpu(p));
2385 if (dl_policy(policy) && !task_has_dl_policy(p) &&
2386 !__dl_overflow(dl_b, cpus, 0, new_bw)) {
2387 __dl_add(dl_b, new_bw);
2389 } else if (dl_policy(policy) && task_has_dl_policy(p) &&
2390 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) {
2391 __dl_clear(dl_b, p->dl.dl_bw);
2392 __dl_add(dl_b, new_bw);
2394 } else if (!dl_policy(policy) && task_has_dl_policy(p)) {
2395 __dl_clear(dl_b, p->dl.dl_bw);
2398 raw_spin_unlock(&dl_b->lock);
2403 extern void init_dl_bw(struct dl_bw *dl_b);
2406 * wake_up_new_task - wake up a newly created task for the first time.
2408 * This function will do some initial scheduler statistics housekeeping
2409 * that must be done for every newly created context, then puts the task
2410 * on the runqueue and wakes it.
2412 void wake_up_new_task(struct task_struct *p)
2414 unsigned long flags;
2417 raw_spin_lock_irqsave(&p->pi_lock, flags);
2419 walt_init_new_task_load(p);
2421 /* Initialize new task's runnable average */
2422 init_entity_runnable_average(&p->se);
2425 * Fork balancing, do it here and not earlier because:
2426 * - cpus_allowed can change in the fork path
2427 * - any previously selected cpu might disappear through hotplug
2429 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0));
2432 rq = __task_rq_lock(p);
2433 walt_mark_task_starting(p);
2434 activate_task(rq, p, ENQUEUE_WAKEUP_NEW);
2435 p->on_rq = TASK_ON_RQ_QUEUED;
2436 trace_sched_wakeup_new(p);
2437 check_preempt_curr(rq, p, WF_FORK);
2439 if (p->sched_class->task_woken) {
2441 * Nothing relies on rq->lock after this, so its fine to
2444 lockdep_unpin_lock(&rq->lock);
2445 p->sched_class->task_woken(rq, p);
2446 lockdep_pin_lock(&rq->lock);
2449 task_rq_unlock(rq, p, &flags);
2452 #ifdef CONFIG_PREEMPT_NOTIFIERS
2454 static struct static_key preempt_notifier_key = STATIC_KEY_INIT_FALSE;
2456 void preempt_notifier_inc(void)
2458 static_key_slow_inc(&preempt_notifier_key);
2460 EXPORT_SYMBOL_GPL(preempt_notifier_inc);
2462 void preempt_notifier_dec(void)
2464 static_key_slow_dec(&preempt_notifier_key);
2466 EXPORT_SYMBOL_GPL(preempt_notifier_dec);
2469 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2470 * @notifier: notifier struct to register
2472 void preempt_notifier_register(struct preempt_notifier *notifier)
2474 if (!static_key_false(&preempt_notifier_key))
2475 WARN(1, "registering preempt_notifier while notifiers disabled\n");
2477 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2479 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2482 * preempt_notifier_unregister - no longer interested in preemption notifications
2483 * @notifier: notifier struct to unregister
2485 * This is *not* safe to call from within a preemption notifier.
2487 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2489 hlist_del(¬ifier->link);
2491 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2493 static void __fire_sched_in_preempt_notifiers(struct task_struct *curr)
2495 struct preempt_notifier *notifier;
2497 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2498 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2501 static __always_inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2503 if (static_key_false(&preempt_notifier_key))
2504 __fire_sched_in_preempt_notifiers(curr);
2508 __fire_sched_out_preempt_notifiers(struct task_struct *curr,
2509 struct task_struct *next)
2511 struct preempt_notifier *notifier;
2513 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link)
2514 notifier->ops->sched_out(notifier, next);
2517 static __always_inline void
2518 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2519 struct task_struct *next)
2521 if (static_key_false(&preempt_notifier_key))
2522 __fire_sched_out_preempt_notifiers(curr, next);
2525 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2527 static inline void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2532 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2533 struct task_struct *next)
2537 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2540 * prepare_task_switch - prepare to switch tasks
2541 * @rq: the runqueue preparing to switch
2542 * @prev: the current task that is being switched out
2543 * @next: the task we are going to switch to.
2545 * This is called with the rq lock held and interrupts off. It must
2546 * be paired with a subsequent finish_task_switch after the context
2549 * prepare_task_switch sets up locking and calls architecture specific
2553 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2554 struct task_struct *next)
2556 sched_info_switch(rq, prev, next);
2557 perf_event_task_sched_out(prev, next);
2558 fire_sched_out_preempt_notifiers(prev, next);
2559 prepare_lock_switch(rq, next);
2560 prepare_arch_switch(next);
2564 * finish_task_switch - clean up after a task-switch
2565 * @prev: the thread we just switched away from.
2567 * finish_task_switch must be called after the context switch, paired
2568 * with a prepare_task_switch call before the context switch.
2569 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2570 * and do any other architecture-specific cleanup actions.
2572 * Note that we may have delayed dropping an mm in context_switch(). If
2573 * so, we finish that here outside of the runqueue lock. (Doing it
2574 * with the lock held can cause deadlocks; see schedule() for
2577 * The context switch have flipped the stack from under us and restored the
2578 * local variables which were saved when this task called schedule() in the
2579 * past. prev == current is still correct but we need to recalculate this_rq
2580 * because prev may have moved to another CPU.
2582 static struct rq *finish_task_switch(struct task_struct *prev)
2583 __releases(rq->lock)
2585 struct rq *rq = this_rq();
2586 struct mm_struct *mm = rq->prev_mm;
2590 * The previous task will have left us with a preempt_count of 2
2591 * because it left us after:
2594 * preempt_disable(); // 1
2596 * raw_spin_lock_irq(&rq->lock) // 2
2598 * Also, see FORK_PREEMPT_COUNT.
2600 if (WARN_ONCE(preempt_count() != 2*PREEMPT_DISABLE_OFFSET,
2601 "corrupted preempt_count: %s/%d/0x%x\n",
2602 current->comm, current->pid, preempt_count()))
2603 preempt_count_set(FORK_PREEMPT_COUNT);
2608 * A task struct has one reference for the use as "current".
2609 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2610 * schedule one last time. The schedule call will never return, and
2611 * the scheduled task must drop that reference.
2613 * We must observe prev->state before clearing prev->on_cpu (in
2614 * finish_lock_switch), otherwise a concurrent wakeup can get prev
2615 * running on another CPU and we could rave with its RUNNING -> DEAD
2616 * transition, resulting in a double drop.
2618 prev_state = prev->state;
2619 vtime_task_switch(prev);
2620 perf_event_task_sched_in(prev, current);
2621 finish_lock_switch(rq, prev);
2622 finish_arch_post_lock_switch();
2624 fire_sched_in_preempt_notifiers(current);
2627 if (unlikely(prev_state == TASK_DEAD)) {
2628 if (prev->sched_class->task_dead)
2629 prev->sched_class->task_dead(prev);
2632 * Remove function-return probe instances associated with this
2633 * task and put them back on the free list.
2635 kprobe_flush_task(prev);
2636 put_task_struct(prev);
2639 tick_nohz_task_switch();
2645 /* rq->lock is NOT held, but preemption is disabled */
2646 static void __balance_callback(struct rq *rq)
2648 struct callback_head *head, *next;
2649 void (*func)(struct rq *rq);
2650 unsigned long flags;
2652 raw_spin_lock_irqsave(&rq->lock, flags);
2653 head = rq->balance_callback;
2654 rq->balance_callback = NULL;
2656 func = (void (*)(struct rq *))head->func;
2663 raw_spin_unlock_irqrestore(&rq->lock, flags);
2666 static inline void balance_callback(struct rq *rq)
2668 if (unlikely(rq->balance_callback))
2669 __balance_callback(rq);
2674 static inline void balance_callback(struct rq *rq)
2681 * schedule_tail - first thing a freshly forked thread must call.
2682 * @prev: the thread we just switched away from.
2684 asmlinkage __visible void schedule_tail(struct task_struct *prev)
2685 __releases(rq->lock)
2690 * New tasks start with FORK_PREEMPT_COUNT, see there and
2691 * finish_task_switch() for details.
2693 * finish_task_switch() will drop rq->lock() and lower preempt_count
2694 * and the preempt_enable() will end up enabling preemption (on
2695 * PREEMPT_COUNT kernels).
2698 rq = finish_task_switch(prev);
2699 balance_callback(rq);
2702 if (current->set_child_tid)
2703 put_user(task_pid_vnr(current), current->set_child_tid);
2707 * context_switch - switch to the new MM and the new thread's register state.
2709 static inline struct rq *
2710 context_switch(struct rq *rq, struct task_struct *prev,
2711 struct task_struct *next)
2713 struct mm_struct *mm, *oldmm;
2715 prepare_task_switch(rq, prev, next);
2718 oldmm = prev->active_mm;
2720 * For paravirt, this is coupled with an exit in switch_to to
2721 * combine the page table reload and the switch backend into
2724 arch_start_context_switch(prev);
2727 next->active_mm = oldmm;
2728 atomic_inc(&oldmm->mm_count);
2729 enter_lazy_tlb(oldmm, next);
2731 switch_mm(oldmm, mm, next);
2734 prev->active_mm = NULL;
2735 rq->prev_mm = oldmm;
2738 * Since the runqueue lock will be released by the next
2739 * task (which is an invalid locking op but in the case
2740 * of the scheduler it's an obvious special-case), so we
2741 * do an early lockdep release here:
2743 lockdep_unpin_lock(&rq->lock);
2744 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2746 /* Here we just switch the register state and the stack. */
2747 switch_to(prev, next, prev);
2750 return finish_task_switch(prev);
2754 * nr_running and nr_context_switches:
2756 * externally visible scheduler statistics: current number of runnable
2757 * threads, total number of context switches performed since bootup.
2759 unsigned long nr_running(void)
2761 unsigned long i, sum = 0;
2763 for_each_online_cpu(i)
2764 sum += cpu_rq(i)->nr_running;
2770 * Check if only the current task is running on the cpu.
2772 * Caution: this function does not check that the caller has disabled
2773 * preemption, thus the result might have a time-of-check-to-time-of-use
2774 * race. The caller is responsible to use it correctly, for example:
2776 * - from a non-preemptable section (of course)
2778 * - from a thread that is bound to a single CPU
2780 * - in a loop with very short iterations (e.g. a polling loop)
2782 bool single_task_running(void)
2784 return raw_rq()->nr_running == 1;
2786 EXPORT_SYMBOL(single_task_running);
2788 unsigned long long nr_context_switches(void)
2791 unsigned long long sum = 0;
2793 for_each_possible_cpu(i)
2794 sum += cpu_rq(i)->nr_switches;
2799 unsigned long nr_iowait(void)
2801 unsigned long i, sum = 0;
2803 for_each_possible_cpu(i)
2804 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2809 unsigned long nr_iowait_cpu(int cpu)
2811 struct rq *this = cpu_rq(cpu);
2812 return atomic_read(&this->nr_iowait);
2815 #ifdef CONFIG_CPU_QUIET
2816 u64 nr_running_integral(unsigned int cpu)
2818 unsigned int seqcnt;
2822 if (cpu >= nr_cpu_ids)
2828 * Update average to avoid reading stalled value if there were
2829 * no run-queue changes for a long time. On the other hand if
2830 * the changes are happening right now, just read current value
2834 seqcnt = read_seqcount_begin(&q->ave_seqcnt);
2835 integral = do_nr_running_integral(q);
2836 if (read_seqcount_retry(&q->ave_seqcnt, seqcnt)) {
2837 read_seqcount_begin(&q->ave_seqcnt);
2838 integral = q->nr_running_integral;
2845 void get_iowait_load(unsigned long *nr_waiters, unsigned long *load)
2847 struct rq *rq = this_rq();
2848 *nr_waiters = atomic_read(&rq->nr_iowait);
2849 *load = rq->load.weight;
2855 * sched_exec - execve() is a valuable balancing opportunity, because at
2856 * this point the task has the smallest effective memory and cache footprint.
2858 void sched_exec(void)
2860 struct task_struct *p = current;
2861 unsigned long flags;
2864 raw_spin_lock_irqsave(&p->pi_lock, flags);
2865 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0);
2866 if (dest_cpu == smp_processor_id())
2869 if (likely(cpu_active(dest_cpu))) {
2870 struct migration_arg arg = { p, dest_cpu };
2872 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2873 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
2877 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2882 DEFINE_PER_CPU(struct kernel_stat, kstat);
2883 DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat);
2885 EXPORT_PER_CPU_SYMBOL(kstat);
2886 EXPORT_PER_CPU_SYMBOL(kernel_cpustat);
2889 * Return accounted runtime for the task.
2890 * In case the task is currently running, return the runtime plus current's
2891 * pending runtime that have not been accounted yet.
2893 unsigned long long task_sched_runtime(struct task_struct *p)
2895 unsigned long flags;
2899 #if defined(CONFIG_64BIT) && defined(CONFIG_SMP)
2901 * 64-bit doesn't need locks to atomically read a 64bit value.
2902 * So we have a optimization chance when the task's delta_exec is 0.
2903 * Reading ->on_cpu is racy, but this is ok.
2905 * If we race with it leaving cpu, we'll take a lock. So we're correct.
2906 * If we race with it entering cpu, unaccounted time is 0. This is
2907 * indistinguishable from the read occurring a few cycles earlier.
2908 * If we see ->on_cpu without ->on_rq, the task is leaving, and has
2909 * been accounted, so we're correct here as well.
2911 if (!p->on_cpu || !task_on_rq_queued(p))
2912 return p->se.sum_exec_runtime;
2915 rq = task_rq_lock(p, &flags);
2917 * Must be ->curr _and_ ->on_rq. If dequeued, we would
2918 * project cycles that may never be accounted to this
2919 * thread, breaking clock_gettime().
2921 if (task_current(rq, p) && task_on_rq_queued(p)) {
2922 update_rq_clock(rq);
2923 p->sched_class->update_curr(rq);
2925 ns = p->se.sum_exec_runtime;
2926 task_rq_unlock(rq, p, &flags);
2931 #ifdef CONFIG_CPU_FREQ_GOV_SCHED
2934 unsigned long add_capacity_margin(unsigned long cpu_capacity)
2936 cpu_capacity = cpu_capacity * capacity_margin;
2937 cpu_capacity /= SCHED_CAPACITY_SCALE;
2938 return cpu_capacity;
2942 unsigned long sum_capacity_reqs(unsigned long cfs_cap,
2943 struct sched_capacity_reqs *scr)
2945 unsigned long total = add_capacity_margin(cfs_cap + scr->rt);
2946 return total += scr->dl;
2949 static void sched_freq_tick_pelt(int cpu)
2951 unsigned long cpu_utilization = capacity_max;
2952 unsigned long capacity_curr = capacity_curr_of(cpu);
2953 struct sched_capacity_reqs *scr;
2955 scr = &per_cpu(cpu_sched_capacity_reqs, cpu);
2956 if (sum_capacity_reqs(cpu_utilization, scr) < capacity_curr)
2960 * To make free room for a task that is building up its "real"
2961 * utilization and to harm its performance the least, request
2962 * a jump to a higher OPP as soon as the margin of free capacity
2963 * is impacted (specified by capacity_margin).
2965 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
2968 #ifdef CONFIG_SCHED_WALT
2969 static void sched_freq_tick_walt(int cpu)
2971 unsigned long cpu_utilization = cpu_util(cpu);
2972 unsigned long capacity_curr = capacity_curr_of(cpu);
2974 if (walt_disabled || !sysctl_sched_use_walt_cpu_util)
2975 return sched_freq_tick_pelt(cpu);
2978 * Add a margin to the WALT utilization.
2979 * NOTE: WALT tracks a single CPU signal for all the scheduling
2980 * classes, thus this margin is going to be added to the DL class as
2981 * well, which is something we do not do in sched_freq_tick_pelt case.
2983 cpu_utilization = add_capacity_margin(cpu_utilization);
2984 if (cpu_utilization <= capacity_curr)
2988 * It is likely that the load is growing so we
2989 * keep the added margin in our request as an
2992 set_cfs_cpu_capacity(cpu, true, cpu_utilization);
2995 #define _sched_freq_tick(cpu) sched_freq_tick_walt(cpu)
2997 #define _sched_freq_tick(cpu) sched_freq_tick_pelt(cpu)
2998 #endif /* CONFIG_SCHED_WALT */
3000 static void sched_freq_tick(int cpu)
3002 unsigned long capacity_orig, capacity_curr;
3007 capacity_orig = capacity_orig_of(cpu);
3008 capacity_curr = capacity_curr_of(cpu);
3009 if (capacity_curr == capacity_orig)
3012 _sched_freq_tick(cpu);
3015 static inline void sched_freq_tick(int cpu) { }
3016 #endif /* CONFIG_CPU_FREQ_GOV_SCHED */
3019 * This function gets called by the timer code, with HZ frequency.
3020 * We call it with interrupts disabled.
3022 void scheduler_tick(void)
3024 int cpu = smp_processor_id();
3025 struct rq *rq = cpu_rq(cpu);
3026 struct task_struct *curr = rq->curr;
3030 raw_spin_lock(&rq->lock);
3031 walt_set_window_start(rq);
3032 update_rq_clock(rq);
3033 curr->sched_class->task_tick(rq, curr, 0);
3034 update_cpu_load_active(rq);
3035 walt_update_task_ravg(rq->curr, rq, TASK_UPDATE,
3036 walt_ktime_clock(), 0);
3037 calc_global_load_tick(rq);
3038 sched_freq_tick(cpu);
3039 raw_spin_unlock(&rq->lock);
3041 perf_event_task_tick();
3044 rq->idle_balance = idle_cpu(cpu);
3045 trigger_load_balance(rq);
3047 rq_last_tick_reset(rq);
3050 #ifdef CONFIG_NO_HZ_FULL
3052 * scheduler_tick_max_deferment
3054 * Keep at least one tick per second when a single
3055 * active task is running because the scheduler doesn't
3056 * yet completely support full dynticks environment.
3058 * This makes sure that uptime, CFS vruntime, load
3059 * balancing, etc... continue to move forward, even
3060 * with a very low granularity.
3062 * Return: Maximum deferment in nanoseconds.
3064 u64 scheduler_tick_max_deferment(void)
3066 struct rq *rq = this_rq();
3067 unsigned long next, now = READ_ONCE(jiffies);
3069 next = rq->last_sched_tick + HZ;
3071 if (time_before_eq(next, now))
3074 return jiffies_to_nsecs(next - now);
3078 notrace unsigned long get_parent_ip(unsigned long addr)
3080 if (in_lock_functions(addr)) {
3081 addr = CALLER_ADDR2;
3082 if (in_lock_functions(addr))
3083 addr = CALLER_ADDR3;
3088 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3089 defined(CONFIG_PREEMPT_TRACER))
3091 void preempt_count_add(int val)
3093 #ifdef CONFIG_DEBUG_PREEMPT
3097 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3100 __preempt_count_add(val);
3101 #ifdef CONFIG_DEBUG_PREEMPT
3103 * Spinlock count overflowing soon?
3105 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3108 if (preempt_count() == val) {
3109 unsigned long ip = get_parent_ip(CALLER_ADDR1);
3110 #ifdef CONFIG_DEBUG_PREEMPT
3111 current->preempt_disable_ip = ip;
3113 trace_preempt_off(CALLER_ADDR0, ip);
3116 EXPORT_SYMBOL(preempt_count_add);
3117 NOKPROBE_SYMBOL(preempt_count_add);
3119 void preempt_count_sub(int val)
3121 #ifdef CONFIG_DEBUG_PREEMPT
3125 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3128 * Is the spinlock portion underflowing?
3130 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3131 !(preempt_count() & PREEMPT_MASK)))
3135 if (preempt_count() == val)
3136 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3137 __preempt_count_sub(val);
3139 EXPORT_SYMBOL(preempt_count_sub);
3140 NOKPROBE_SYMBOL(preempt_count_sub);
3145 * Print scheduling while atomic bug:
3147 static noinline void __schedule_bug(struct task_struct *prev)
3149 if (oops_in_progress)
3152 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3153 prev->comm, prev->pid, preempt_count());
3155 debug_show_held_locks(prev);
3157 if (irqs_disabled())
3158 print_irqtrace_events(prev);
3159 #ifdef CONFIG_DEBUG_PREEMPT
3160 if (in_atomic_preempt_off()) {
3161 pr_err("Preemption disabled at:");
3162 print_ip_sym(current->preempt_disable_ip);
3167 add_taint(TAINT_WARN, LOCKDEP_STILL_OK);
3171 * Various schedule()-time debugging checks and statistics:
3173 static inline void schedule_debug(struct task_struct *prev)
3175 #ifdef CONFIG_SCHED_STACK_END_CHECK
3176 if (task_stack_end_corrupted(prev))
3177 panic("corrupted stack end detected inside scheduler\n");
3180 if (unlikely(in_atomic_preempt_off())) {
3181 __schedule_bug(prev);
3182 preempt_count_set(PREEMPT_DISABLED);
3186 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3188 schedstat_inc(this_rq(), sched_count);
3192 * Pick up the highest-prio task:
3194 static inline struct task_struct *
3195 pick_next_task(struct rq *rq, struct task_struct *prev)
3197 const struct sched_class *class = &fair_sched_class;
3198 struct task_struct *p;
3201 * Optimization: we know that if all tasks are in
3202 * the fair class we can call that function directly:
3204 if (likely(prev->sched_class == class &&
3205 rq->nr_running == rq->cfs.h_nr_running)) {
3206 p = fair_sched_class.pick_next_task(rq, prev);
3207 if (unlikely(p == RETRY_TASK))
3210 /* assumes fair_sched_class->next == idle_sched_class */
3212 p = idle_sched_class.pick_next_task(rq, prev);
3218 for_each_class(class) {
3219 p = class->pick_next_task(rq, prev);
3221 if (unlikely(p == RETRY_TASK))
3227 BUG(); /* the idle class will always have a runnable task */
3231 * __schedule() is the main scheduler function.
3233 * The main means of driving the scheduler and thus entering this function are:
3235 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc.
3237 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return
3238 * paths. For example, see arch/x86/entry_64.S.
3240 * To drive preemption between tasks, the scheduler sets the flag in timer
3241 * interrupt handler scheduler_tick().
3243 * 3. Wakeups don't really cause entry into schedule(). They add a
3244 * task to the run-queue and that's it.
3246 * Now, if the new task added to the run-queue preempts the current
3247 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets
3248 * called on the nearest possible occasion:
3250 * - If the kernel is preemptible (CONFIG_PREEMPT=y):
3252 * - in syscall or exception context, at the next outmost
3253 * preempt_enable(). (this might be as soon as the wake_up()'s
3256 * - in IRQ context, return from interrupt-handler to
3257 * preemptible context
3259 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set)
3262 * - cond_resched() call
3263 * - explicit schedule() call
3264 * - return from syscall or exception to user-space
3265 * - return from interrupt-handler to user-space
3267 * WARNING: must be called with preemption disabled!
3269 static void __sched notrace __schedule(bool preempt)
3271 struct task_struct *prev, *next;
3272 unsigned long *switch_count;
3277 cpu = smp_processor_id();
3279 rcu_note_context_switch();
3283 * do_exit() calls schedule() with preemption disabled as an exception;
3284 * however we must fix that up, otherwise the next task will see an
3285 * inconsistent (higher) preempt count.
3287 * It also avoids the below schedule_debug() test from complaining
3290 if (unlikely(prev->state == TASK_DEAD))
3291 preempt_enable_no_resched_notrace();
3293 schedule_debug(prev);
3295 if (sched_feat(HRTICK))
3299 * Make sure that signal_pending_state()->signal_pending() below
3300 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE)
3301 * done by the caller to avoid the race with signal_wake_up().
3303 smp_mb__before_spinlock();
3304 raw_spin_lock_irq(&rq->lock);
3305 lockdep_pin_lock(&rq->lock);
3307 rq->clock_skip_update <<= 1; /* promote REQ to ACT */
3309 switch_count = &prev->nivcsw;
3310 if (!preempt && prev->state) {
3311 if (unlikely(signal_pending_state(prev->state, prev))) {
3312 prev->state = TASK_RUNNING;
3314 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3318 * If a worker went to sleep, notify and ask workqueue
3319 * whether it wants to wake up a task to maintain
3322 if (prev->flags & PF_WQ_WORKER) {
3323 struct task_struct *to_wakeup;
3325 to_wakeup = wq_worker_sleeping(prev, cpu);
3327 try_to_wake_up_local(to_wakeup);
3330 switch_count = &prev->nvcsw;
3333 if (task_on_rq_queued(prev))
3334 update_rq_clock(rq);
3336 next = pick_next_task(rq, prev);
3337 wallclock = walt_ktime_clock();
3338 walt_update_task_ravg(prev, rq, PUT_PREV_TASK, wallclock, 0);
3339 walt_update_task_ravg(next, rq, PICK_NEXT_TASK, wallclock, 0);
3340 clear_tsk_need_resched(prev);
3341 clear_preempt_need_resched();
3342 rq->clock_skip_update = 0;
3344 if (likely(prev != next)) {
3349 trace_sched_switch(preempt, prev, next);
3350 rq = context_switch(rq, prev, next); /* unlocks the rq */
3353 lockdep_unpin_lock(&rq->lock);
3354 raw_spin_unlock_irq(&rq->lock);
3357 balance_callback(rq);
3360 static inline void sched_submit_work(struct task_struct *tsk)
3362 if (!tsk->state || tsk_is_pi_blocked(tsk))
3365 * If we are going to sleep and we have plugged IO queued,
3366 * make sure to submit it to avoid deadlocks.
3368 if (blk_needs_flush_plug(tsk))
3369 blk_schedule_flush_plug(tsk);
3372 asmlinkage __visible void __sched schedule(void)
3374 struct task_struct *tsk = current;
3376 sched_submit_work(tsk);
3380 sched_preempt_enable_no_resched();
3381 } while (need_resched());
3383 EXPORT_SYMBOL(schedule);
3385 #ifdef CONFIG_CONTEXT_TRACKING
3386 asmlinkage __visible void __sched schedule_user(void)
3389 * If we come here after a random call to set_need_resched(),
3390 * or we have been woken up remotely but the IPI has not yet arrived,
3391 * we haven't yet exited the RCU idle mode. Do it here manually until
3392 * we find a better solution.
3394 * NB: There are buggy callers of this function. Ideally we
3395 * should warn if prev_state != CONTEXT_USER, but that will trigger
3396 * too frequently to make sense yet.
3398 enum ctx_state prev_state = exception_enter();
3400 exception_exit(prev_state);
3405 * schedule_preempt_disabled - called with preemption disabled
3407 * Returns with preemption disabled. Note: preempt_count must be 1
3409 void __sched schedule_preempt_disabled(void)
3411 sched_preempt_enable_no_resched();
3416 static void __sched notrace preempt_schedule_common(void)
3419 preempt_disable_notrace();
3421 preempt_enable_no_resched_notrace();
3424 * Check again in case we missed a preemption opportunity
3425 * between schedule and now.
3427 } while (need_resched());
3430 #ifdef CONFIG_PREEMPT
3432 * this is the entry point to schedule() from in-kernel preemption
3433 * off of preempt_enable. Kernel preemptions off return from interrupt
3434 * occur there and call schedule directly.
3436 asmlinkage __visible void __sched notrace preempt_schedule(void)
3439 * If there is a non-zero preempt_count or interrupts are disabled,
3440 * we do not want to preempt the current task. Just return..
3442 if (likely(!preemptible()))
3445 preempt_schedule_common();
3447 NOKPROBE_SYMBOL(preempt_schedule);
3448 EXPORT_SYMBOL(preempt_schedule);
3451 * preempt_schedule_notrace - preempt_schedule called by tracing
3453 * The tracing infrastructure uses preempt_enable_notrace to prevent
3454 * recursion and tracing preempt enabling caused by the tracing
3455 * infrastructure itself. But as tracing can happen in areas coming
3456 * from userspace or just about to enter userspace, a preempt enable
3457 * can occur before user_exit() is called. This will cause the scheduler
3458 * to be called when the system is still in usermode.
3460 * To prevent this, the preempt_enable_notrace will use this function
3461 * instead of preempt_schedule() to exit user context if needed before
3462 * calling the scheduler.
3464 asmlinkage __visible void __sched notrace preempt_schedule_notrace(void)
3466 enum ctx_state prev_ctx;
3468 if (likely(!preemptible()))
3472 preempt_disable_notrace();
3474 * Needs preempt disabled in case user_exit() is traced
3475 * and the tracer calls preempt_enable_notrace() causing
3476 * an infinite recursion.
3478 prev_ctx = exception_enter();
3480 exception_exit(prev_ctx);
3482 preempt_enable_no_resched_notrace();
3483 } while (need_resched());
3485 EXPORT_SYMBOL_GPL(preempt_schedule_notrace);
3487 #endif /* CONFIG_PREEMPT */
3490 * this is the entry point to schedule() from kernel preemption
3491 * off of irq context.
3492 * Note, that this is called and return with irqs disabled. This will
3493 * protect us against recursive calling from irq.
3495 asmlinkage __visible void __sched preempt_schedule_irq(void)
3497 enum ctx_state prev_state;
3499 /* Catch callers which need to be fixed */
3500 BUG_ON(preempt_count() || !irqs_disabled());
3502 prev_state = exception_enter();
3508 local_irq_disable();
3509 sched_preempt_enable_no_resched();
3510 } while (need_resched());
3512 exception_exit(prev_state);
3515 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3518 return try_to_wake_up(curr->private, mode, wake_flags);
3520 EXPORT_SYMBOL(default_wake_function);
3522 #ifdef CONFIG_RT_MUTEXES
3525 * rt_mutex_setprio - set the current priority of a task
3527 * @prio: prio value (kernel-internal form)
3529 * This function changes the 'effective' priority of a task. It does
3530 * not touch ->normal_prio like __setscheduler().
3532 * Used by the rt_mutex code to implement priority inheritance
3533 * logic. Call site only calls if the priority of the task changed.
3535 void rt_mutex_setprio(struct task_struct *p, int prio)
3537 int oldprio, queued, running, enqueue_flag = ENQUEUE_RESTORE;
3539 const struct sched_class *prev_class;
3541 BUG_ON(prio > MAX_PRIO);
3543 rq = __task_rq_lock(p);
3546 * Idle task boosting is a nono in general. There is one
3547 * exception, when PREEMPT_RT and NOHZ is active:
3549 * The idle task calls get_next_timer_interrupt() and holds
3550 * the timer wheel base->lock on the CPU and another CPU wants
3551 * to access the timer (probably to cancel it). We can safely
3552 * ignore the boosting request, as the idle CPU runs this code
3553 * with interrupts disabled and will complete the lock
3554 * protected section without being interrupted. So there is no
3555 * real need to boost.
3557 if (unlikely(p == rq->idle)) {
3558 WARN_ON(p != rq->curr);
3559 WARN_ON(p->pi_blocked_on);
3563 trace_sched_pi_setprio(p, prio);
3565 prev_class = p->sched_class;
3566 queued = task_on_rq_queued(p);
3567 running = task_current(rq, p);
3569 dequeue_task(rq, p, DEQUEUE_SAVE);
3571 put_prev_task(rq, p);
3574 * Boosting condition are:
3575 * 1. -rt task is running and holds mutex A
3576 * --> -dl task blocks on mutex A
3578 * 2. -dl task is running and holds mutex A
3579 * --> -dl task blocks on mutex A and could preempt the
3582 if (dl_prio(prio)) {
3583 struct task_struct *pi_task = rt_mutex_get_top_task(p);
3584 if (!dl_prio(p->normal_prio) ||
3585 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) {
3586 p->dl.dl_boosted = 1;
3587 enqueue_flag |= ENQUEUE_REPLENISH;
3589 p->dl.dl_boosted = 0;
3590 p->sched_class = &dl_sched_class;
3591 } else if (rt_prio(prio)) {
3592 if (dl_prio(oldprio))
3593 p->dl.dl_boosted = 0;
3595 enqueue_flag |= ENQUEUE_HEAD;
3596 p->sched_class = &rt_sched_class;
3598 if (dl_prio(oldprio))
3599 p->dl.dl_boosted = 0;
3600 if (rt_prio(oldprio))
3602 p->sched_class = &fair_sched_class;
3608 p->sched_class->set_curr_task(rq);
3610 enqueue_task(rq, p, enqueue_flag);
3612 check_class_changed(rq, p, prev_class, oldprio);
3614 preempt_disable(); /* avoid rq from going away on us */
3615 __task_rq_unlock(rq);
3617 balance_callback(rq);
3622 void set_user_nice(struct task_struct *p, long nice)
3624 int old_prio, delta, queued;
3625 unsigned long flags;
3628 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE)
3631 * We have to be careful, if called from sys_setpriority(),
3632 * the task might be in the middle of scheduling on another CPU.
3634 rq = task_rq_lock(p, &flags);
3636 * The RT priorities are set via sched_setscheduler(), but we still
3637 * allow the 'normal' nice value to be set - but as expected
3638 * it wont have any effect on scheduling until the task is
3639 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR:
3641 if (task_has_dl_policy(p) || task_has_rt_policy(p)) {
3642 p->static_prio = NICE_TO_PRIO(nice);
3645 queued = task_on_rq_queued(p);
3647 dequeue_task(rq, p, DEQUEUE_SAVE);
3649 p->static_prio = NICE_TO_PRIO(nice);
3652 p->prio = effective_prio(p);
3653 delta = p->prio - old_prio;
3656 enqueue_task(rq, p, ENQUEUE_RESTORE);
3658 * If the task increased its priority or is running and
3659 * lowered its priority, then reschedule its CPU:
3661 if (delta < 0 || (delta > 0 && task_running(rq, p)))
3665 task_rq_unlock(rq, p, &flags);
3667 EXPORT_SYMBOL(set_user_nice);
3670 * can_nice - check if a task can reduce its nice value
3674 int can_nice(const struct task_struct *p, const int nice)
3676 /* convert nice value [19,-20] to rlimit style value [1,40] */
3677 int nice_rlim = nice_to_rlimit(nice);
3679 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
3680 capable(CAP_SYS_NICE));
3683 #ifdef __ARCH_WANT_SYS_NICE
3686 * sys_nice - change the priority of the current process.
3687 * @increment: priority increment
3689 * sys_setpriority is a more generic, but much slower function that
3690 * does similar things.
3692 SYSCALL_DEFINE1(nice, int, increment)
3697 * Setpriority might change our priority at the same moment.
3698 * We don't have to worry. Conceptually one call occurs first
3699 * and we have a single winner.
3701 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH);
3702 nice = task_nice(current) + increment;
3704 nice = clamp_val(nice, MIN_NICE, MAX_NICE);
3705 if (increment < 0 && !can_nice(current, nice))
3708 retval = security_task_setnice(current, nice);
3712 set_user_nice(current, nice);
3719 * task_prio - return the priority value of a given task.
3720 * @p: the task in question.
3722 * Return: The priority value as seen by users in /proc.
3723 * RT tasks are offset by -200. Normal tasks are centered
3724 * around 0, value goes from -16 to +15.
3726 int task_prio(const struct task_struct *p)
3728 return p->prio - MAX_RT_PRIO;
3732 * idle_cpu - is a given cpu idle currently?
3733 * @cpu: the processor in question.
3735 * Return: 1 if the CPU is currently idle. 0 otherwise.
3737 int idle_cpu(int cpu)
3739 struct rq *rq = cpu_rq(cpu);
3741 if (rq->curr != rq->idle)
3748 if (!llist_empty(&rq->wake_list))
3756 * idle_task - return the idle task for a given cpu.
3757 * @cpu: the processor in question.
3759 * Return: The idle task for the cpu @cpu.
3761 struct task_struct *idle_task(int cpu)
3763 return cpu_rq(cpu)->idle;
3767 * find_process_by_pid - find a process with a matching PID value.
3768 * @pid: the pid in question.
3770 * The task of @pid, if found. %NULL otherwise.
3772 static struct task_struct *find_process_by_pid(pid_t pid)
3774 return pid ? find_task_by_vpid(pid) : current;
3778 * This function initializes the sched_dl_entity of a newly becoming
3779 * SCHED_DEADLINE task.
3781 * Only the static values are considered here, the actual runtime and the
3782 * absolute deadline will be properly calculated when the task is enqueued
3783 * for the first time with its new policy.
3786 __setparam_dl(struct task_struct *p, const struct sched_attr *attr)
3788 struct sched_dl_entity *dl_se = &p->dl;
3790 dl_se->dl_runtime = attr->sched_runtime;
3791 dl_se->dl_deadline = attr->sched_deadline;
3792 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline;
3793 dl_se->flags = attr->sched_flags;
3794 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime);
3797 * Changing the parameters of a task is 'tricky' and we're not doing
3798 * the correct thing -- also see task_dead_dl() and switched_from_dl().
3800 * What we SHOULD do is delay the bandwidth release until the 0-lag
3801 * point. This would include retaining the task_struct until that time
3802 * and change dl_overflow() to not immediately decrement the current
3805 * Instead we retain the current runtime/deadline and let the new
3806 * parameters take effect after the current reservation period lapses.
3807 * This is safe (albeit pessimistic) because the 0-lag point is always
3808 * before the current scheduling deadline.
3810 * We can still have temporary overloads because we do not delay the
3811 * change in bandwidth until that time; so admission control is
3812 * not on the safe side. It does however guarantee tasks will never
3813 * consume more than promised.
3818 * sched_setparam() passes in -1 for its policy, to let the functions
3819 * it calls know not to change it.
3821 #define SETPARAM_POLICY -1
3823 static void __setscheduler_params(struct task_struct *p,
3824 const struct sched_attr *attr)
3826 int policy = attr->sched_policy;
3828 if (policy == SETPARAM_POLICY)
3833 if (dl_policy(policy))
3834 __setparam_dl(p, attr);
3835 else if (fair_policy(policy))
3836 p->static_prio = NICE_TO_PRIO(attr->sched_nice);
3839 * __sched_setscheduler() ensures attr->sched_priority == 0 when
3840 * !rt_policy. Always setting this ensures that things like
3841 * getparam()/getattr() don't report silly values for !rt tasks.
3843 p->rt_priority = attr->sched_priority;
3844 p->normal_prio = normal_prio(p);
3848 /* Actually do priority change: must hold pi & rq lock. */
3849 static void __setscheduler(struct rq *rq, struct task_struct *p,
3850 const struct sched_attr *attr, bool keep_boost)
3852 __setscheduler_params(p, attr);
3855 * Keep a potential priority boosting if called from
3856 * sched_setscheduler().
3859 p->prio = rt_mutex_get_effective_prio(p, normal_prio(p));
3861 p->prio = normal_prio(p);
3863 if (dl_prio(p->prio))
3864 p->sched_class = &dl_sched_class;
3865 else if (rt_prio(p->prio))
3866 p->sched_class = &rt_sched_class;
3868 p->sched_class = &fair_sched_class;
3872 __getparam_dl(struct task_struct *p, struct sched_attr *attr)
3874 struct sched_dl_entity *dl_se = &p->dl;
3876 attr->sched_priority = p->rt_priority;
3877 attr->sched_runtime = dl_se->dl_runtime;
3878 attr->sched_deadline = dl_se->dl_deadline;
3879 attr->sched_period = dl_se->dl_period;
3880 attr->sched_flags = dl_se->flags;
3884 * This function validates the new parameters of a -deadline task.
3885 * We ask for the deadline not being zero, and greater or equal
3886 * than the runtime, as well as the period of being zero or
3887 * greater than deadline. Furthermore, we have to be sure that
3888 * user parameters are above the internal resolution of 1us (we
3889 * check sched_runtime only since it is always the smaller one) and
3890 * below 2^63 ns (we have to check both sched_deadline and
3891 * sched_period, as the latter can be zero).
3894 __checkparam_dl(const struct sched_attr *attr)
3897 if (attr->sched_deadline == 0)
3901 * Since we truncate DL_SCALE bits, make sure we're at least
3904 if (attr->sched_runtime < (1ULL << DL_SCALE))
3908 * Since we use the MSB for wrap-around and sign issues, make
3909 * sure it's not set (mind that period can be equal to zero).
3911 if (attr->sched_deadline & (1ULL << 63) ||
3912 attr->sched_period & (1ULL << 63))
3915 /* runtime <= deadline <= period (if period != 0) */
3916 if ((attr->sched_period != 0 &&
3917 attr->sched_period < attr->sched_deadline) ||
3918 attr->sched_deadline < attr->sched_runtime)
3925 * check the target process has a UID that matches the current process's
3927 static bool check_same_owner(struct task_struct *p)
3929 const struct cred *cred = current_cred(), *pcred;
3933 pcred = __task_cred(p);
3934 match = (uid_eq(cred->euid, pcred->euid) ||
3935 uid_eq(cred->euid, pcred->uid));
3940 static bool dl_param_changed(struct task_struct *p,
3941 const struct sched_attr *attr)
3943 struct sched_dl_entity *dl_se = &p->dl;
3945 if (dl_se->dl_runtime != attr->sched_runtime ||
3946 dl_se->dl_deadline != attr->sched_deadline ||
3947 dl_se->dl_period != attr->sched_period ||
3948 dl_se->flags != attr->sched_flags)
3954 static int __sched_setscheduler(struct task_struct *p,
3955 const struct sched_attr *attr,
3958 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 :
3959 MAX_RT_PRIO - 1 - attr->sched_priority;
3960 int retval, oldprio, oldpolicy = -1, queued, running;
3961 int new_effective_prio, policy = attr->sched_policy;
3962 unsigned long flags;
3963 const struct sched_class *prev_class;
3967 /* may grab non-irq protected spin_locks */
3968 BUG_ON(in_interrupt());
3970 /* double check policy once rq lock held */
3972 reset_on_fork = p->sched_reset_on_fork;
3973 policy = oldpolicy = p->policy;
3975 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK);
3977 if (!valid_policy(policy))
3981 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK))
3985 * Valid priorities for SCHED_FIFO and SCHED_RR are
3986 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
3987 * SCHED_BATCH and SCHED_IDLE is 0.
3989 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) ||
3990 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1))
3992 if ((dl_policy(policy) && !__checkparam_dl(attr)) ||
3993 (rt_policy(policy) != (attr->sched_priority != 0)))
3997 * Allow unprivileged RT tasks to decrease priority:
3999 if (user && !capable(CAP_SYS_NICE)) {
4000 if (fair_policy(policy)) {
4001 if (attr->sched_nice < task_nice(p) &&
4002 !can_nice(p, attr->sched_nice))
4006 if (rt_policy(policy)) {
4007 unsigned long rlim_rtprio =
4008 task_rlimit(p, RLIMIT_RTPRIO);
4010 /* can't set/change the rt policy */
4011 if (policy != p->policy && !rlim_rtprio)
4014 /* can't increase priority */
4015 if (attr->sched_priority > p->rt_priority &&
4016 attr->sched_priority > rlim_rtprio)
4021 * Can't set/change SCHED_DEADLINE policy at all for now
4022 * (safest behavior); in the future we would like to allow
4023 * unprivileged DL tasks to increase their relative deadline
4024 * or reduce their runtime (both ways reducing utilization)
4026 if (dl_policy(policy))
4030 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4031 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4033 if (idle_policy(p->policy) && !idle_policy(policy)) {
4034 if (!can_nice(p, task_nice(p)))
4038 /* can't change other user's priorities */
4039 if (!check_same_owner(p))
4042 /* Normal users shall not reset the sched_reset_on_fork flag */
4043 if (p->sched_reset_on_fork && !reset_on_fork)
4048 retval = security_task_setscheduler(p);
4054 * make sure no PI-waiters arrive (or leave) while we are
4055 * changing the priority of the task:
4057 * To be able to change p->policy safely, the appropriate
4058 * runqueue lock must be held.
4060 rq = task_rq_lock(p, &flags);
4063 * Changing the policy of the stop threads its a very bad idea
4065 if (p == rq->stop) {
4066 task_rq_unlock(rq, p, &flags);
4071 * If not changing anything there's no need to proceed further,
4072 * but store a possible modification of reset_on_fork.
4074 if (unlikely(policy == p->policy)) {
4075 if (fair_policy(policy) && attr->sched_nice != task_nice(p))
4077 if (rt_policy(policy) && attr->sched_priority != p->rt_priority)
4079 if (dl_policy(policy) && dl_param_changed(p, attr))
4082 p->sched_reset_on_fork = reset_on_fork;
4083 task_rq_unlock(rq, p, &flags);
4089 #ifdef CONFIG_RT_GROUP_SCHED
4091 * Do not allow realtime tasks into groups that have no runtime
4094 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4095 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
4096 !task_group_is_autogroup(task_group(p))) {
4097 task_rq_unlock(rq, p, &flags);
4102 if (dl_bandwidth_enabled() && dl_policy(policy)) {
4103 cpumask_t *span = rq->rd->span;
4106 * Don't allow tasks with an affinity mask smaller than
4107 * the entire root_domain to become SCHED_DEADLINE. We
4108 * will also fail if there's no bandwidth available.
4110 if (!cpumask_subset(span, &p->cpus_allowed) ||
4111 rq->rd->dl_bw.bw == 0) {
4112 task_rq_unlock(rq, p, &flags);
4119 /* recheck policy now with rq lock held */
4120 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4121 policy = oldpolicy = -1;
4122 task_rq_unlock(rq, p, &flags);
4127 * If setscheduling to SCHED_DEADLINE (or changing the parameters
4128 * of a SCHED_DEADLINE task) we need to check if enough bandwidth
4131 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) {
4132 task_rq_unlock(rq, p, &flags);
4136 p->sched_reset_on_fork = reset_on_fork;
4141 * Take priority boosted tasks into account. If the new
4142 * effective priority is unchanged, we just store the new
4143 * normal parameters and do not touch the scheduler class and
4144 * the runqueue. This will be done when the task deboost
4147 new_effective_prio = rt_mutex_get_effective_prio(p, newprio);
4148 if (new_effective_prio == oldprio) {
4149 __setscheduler_params(p, attr);
4150 task_rq_unlock(rq, p, &flags);
4155 queued = task_on_rq_queued(p);
4156 running = task_current(rq, p);
4158 dequeue_task(rq, p, DEQUEUE_SAVE);
4160 put_prev_task(rq, p);
4162 prev_class = p->sched_class;
4163 __setscheduler(rq, p, attr, pi);
4166 p->sched_class->set_curr_task(rq);
4168 int enqueue_flags = ENQUEUE_RESTORE;
4170 * We enqueue to tail when the priority of a task is
4171 * increased (user space view).
4173 if (oldprio <= p->prio)
4174 enqueue_flags |= ENQUEUE_HEAD;
4176 enqueue_task(rq, p, enqueue_flags);
4179 check_class_changed(rq, p, prev_class, oldprio);
4180 preempt_disable(); /* avoid rq from going away on us */
4181 task_rq_unlock(rq, p, &flags);
4184 rt_mutex_adjust_pi(p);
4187 * Run balance callbacks after we've adjusted the PI chain.
4189 balance_callback(rq);
4195 static int _sched_setscheduler(struct task_struct *p, int policy,
4196 const struct sched_param *param, bool check)
4198 struct sched_attr attr = {
4199 .sched_policy = policy,
4200 .sched_priority = param->sched_priority,
4201 .sched_nice = PRIO_TO_NICE(p->static_prio),
4204 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */
4205 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) {
4206 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4207 policy &= ~SCHED_RESET_ON_FORK;
4208 attr.sched_policy = policy;
4211 return __sched_setscheduler(p, &attr, check, true);
4214 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4215 * @p: the task in question.
4216 * @policy: new policy.
4217 * @param: structure containing the new RT priority.
4219 * Return: 0 on success. An error code otherwise.
4221 * NOTE that the task may be already dead.
4223 int sched_setscheduler(struct task_struct *p, int policy,
4224 const struct sched_param *param)
4226 return _sched_setscheduler(p, policy, param, true);
4228 EXPORT_SYMBOL_GPL(sched_setscheduler);
4230 int sched_setattr(struct task_struct *p, const struct sched_attr *attr)
4232 return __sched_setscheduler(p, attr, true, true);
4234 EXPORT_SYMBOL_GPL(sched_setattr);
4237 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4238 * @p: the task in question.
4239 * @policy: new policy.
4240 * @param: structure containing the new RT priority.
4242 * Just like sched_setscheduler, only don't bother checking if the
4243 * current context has permission. For example, this is needed in
4244 * stop_machine(): we create temporary high priority worker threads,
4245 * but our caller might not have that capability.
4247 * Return: 0 on success. An error code otherwise.
4249 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4250 const struct sched_param *param)
4252 return _sched_setscheduler(p, policy, param, false);
4254 EXPORT_SYMBOL_GPL(sched_setscheduler_nocheck);
4257 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4259 struct sched_param lparam;
4260 struct task_struct *p;
4263 if (!param || pid < 0)
4265 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4270 p = find_process_by_pid(pid);
4272 retval = sched_setscheduler(p, policy, &lparam);
4279 * Mimics kernel/events/core.c perf_copy_attr().
4281 static int sched_copy_attr(struct sched_attr __user *uattr,
4282 struct sched_attr *attr)
4287 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0))
4291 * zero the full structure, so that a short copy will be nice.
4293 memset(attr, 0, sizeof(*attr));
4295 ret = get_user(size, &uattr->size);
4299 if (size > PAGE_SIZE) /* silly large */
4302 if (!size) /* abi compat */
4303 size = SCHED_ATTR_SIZE_VER0;
4305 if (size < SCHED_ATTR_SIZE_VER0)
4309 * If we're handed a bigger struct than we know of,
4310 * ensure all the unknown bits are 0 - i.e. new
4311 * user-space does not rely on any kernel feature
4312 * extensions we dont know about yet.
4314 if (size > sizeof(*attr)) {
4315 unsigned char __user *addr;
4316 unsigned char __user *end;
4319 addr = (void __user *)uattr + sizeof(*attr);
4320 end = (void __user *)uattr + size;
4322 for (; addr < end; addr++) {
4323 ret = get_user(val, addr);
4329 size = sizeof(*attr);
4332 ret = copy_from_user(attr, uattr, size);
4337 * XXX: do we want to be lenient like existing syscalls; or do we want
4338 * to be strict and return an error on out-of-bounds values?
4340 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE);
4345 put_user(sizeof(*attr), &uattr->size);
4350 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4351 * @pid: the pid in question.
4352 * @policy: new policy.
4353 * @param: structure containing the new RT priority.
4355 * Return: 0 on success. An error code otherwise.
4357 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4358 struct sched_param __user *, param)
4360 /* negative values for policy are not valid */
4364 return do_sched_setscheduler(pid, policy, param);
4368 * sys_sched_setparam - set/change the RT priority of a thread
4369 * @pid: the pid in question.
4370 * @param: structure containing the new RT priority.
4372 * Return: 0 on success. An error code otherwise.
4374 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4376 return do_sched_setscheduler(pid, SETPARAM_POLICY, param);
4380 * sys_sched_setattr - same as above, but with extended sched_attr
4381 * @pid: the pid in question.
4382 * @uattr: structure containing the extended parameters.
4383 * @flags: for future extension.
4385 SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr,
4386 unsigned int, flags)
4388 struct sched_attr attr;
4389 struct task_struct *p;
4392 if (!uattr || pid < 0 || flags)
4395 retval = sched_copy_attr(uattr, &attr);
4399 if ((int)attr.sched_policy < 0)
4404 p = find_process_by_pid(pid);
4406 retval = sched_setattr(p, &attr);
4413 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4414 * @pid: the pid in question.
4416 * Return: On success, the policy of the thread. Otherwise, a negative error
4419 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4421 struct task_struct *p;
4429 p = find_process_by_pid(pid);
4431 retval = security_task_getscheduler(p);
4434 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4441 * sys_sched_getparam - get the RT priority of a thread
4442 * @pid: the pid in question.
4443 * @param: structure containing the RT priority.
4445 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error
4448 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4450 struct sched_param lp = { .sched_priority = 0 };
4451 struct task_struct *p;
4454 if (!param || pid < 0)
4458 p = find_process_by_pid(pid);
4463 retval = security_task_getscheduler(p);
4467 if (task_has_rt_policy(p))
4468 lp.sched_priority = p->rt_priority;
4472 * This one might sleep, we cannot do it with a spinlock held ...
4474 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4483 static int sched_read_attr(struct sched_attr __user *uattr,
4484 struct sched_attr *attr,
4489 if (!access_ok(VERIFY_WRITE, uattr, usize))
4493 * If we're handed a smaller struct than we know of,
4494 * ensure all the unknown bits are 0 - i.e. old
4495 * user-space does not get uncomplete information.
4497 if (usize < sizeof(*attr)) {
4498 unsigned char *addr;
4501 addr = (void *)attr + usize;
4502 end = (void *)attr + sizeof(*attr);
4504 for (; addr < end; addr++) {
4512 ret = copy_to_user(uattr, attr, attr->size);
4520 * sys_sched_getattr - similar to sched_getparam, but with sched_attr
4521 * @pid: the pid in question.
4522 * @uattr: structure containing the extended parameters.
4523 * @size: sizeof(attr) for fwd/bwd comp.
4524 * @flags: for future extension.
4526 SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr,
4527 unsigned int, size, unsigned int, flags)
4529 struct sched_attr attr = {
4530 .size = sizeof(struct sched_attr),
4532 struct task_struct *p;
4535 if (!uattr || pid < 0 || size > PAGE_SIZE ||
4536 size < SCHED_ATTR_SIZE_VER0 || flags)
4540 p = find_process_by_pid(pid);
4545 retval = security_task_getscheduler(p);
4549 attr.sched_policy = p->policy;
4550 if (p->sched_reset_on_fork)
4551 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK;
4552 if (task_has_dl_policy(p))
4553 __getparam_dl(p, &attr);
4554 else if (task_has_rt_policy(p))
4555 attr.sched_priority = p->rt_priority;
4557 attr.sched_nice = task_nice(p);
4561 retval = sched_read_attr(uattr, &attr, size);
4569 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4571 cpumask_var_t cpus_allowed, new_mask;
4572 struct task_struct *p;
4577 p = find_process_by_pid(pid);
4583 /* Prevent p going away */
4587 if (p->flags & PF_NO_SETAFFINITY) {
4591 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4595 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4597 goto out_free_cpus_allowed;
4600 if (!check_same_owner(p)) {
4602 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) {
4604 goto out_free_new_mask;
4609 retval = security_task_setscheduler(p);
4611 goto out_free_new_mask;
4614 cpuset_cpus_allowed(p, cpus_allowed);
4615 cpumask_and(new_mask, in_mask, cpus_allowed);
4618 * Since bandwidth control happens on root_domain basis,
4619 * if admission test is enabled, we only admit -deadline
4620 * tasks allowed to run on all the CPUs in the task's
4624 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) {
4626 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) {
4629 goto out_free_new_mask;
4635 retval = __set_cpus_allowed_ptr(p, new_mask, true);
4638 cpuset_cpus_allowed(p, cpus_allowed);
4639 if (!cpumask_subset(new_mask, cpus_allowed)) {
4641 * We must have raced with a concurrent cpuset
4642 * update. Just reset the cpus_allowed to the
4643 * cpuset's cpus_allowed
4645 cpumask_copy(new_mask, cpus_allowed);
4650 free_cpumask_var(new_mask);
4651 out_free_cpus_allowed:
4652 free_cpumask_var(cpus_allowed);
4658 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4659 struct cpumask *new_mask)
4661 if (len < cpumask_size())
4662 cpumask_clear(new_mask);
4663 else if (len > cpumask_size())
4664 len = cpumask_size();
4666 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4670 * sys_sched_setaffinity - set the cpu affinity of a process
4671 * @pid: pid of the process
4672 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4673 * @user_mask_ptr: user-space pointer to the new cpu mask
4675 * Return: 0 on success. An error code otherwise.
4677 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4678 unsigned long __user *, user_mask_ptr)
4680 cpumask_var_t new_mask;
4683 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4686 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4688 retval = sched_setaffinity(pid, new_mask);
4689 free_cpumask_var(new_mask);
4693 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4695 struct task_struct *p;
4696 unsigned long flags;
4702 p = find_process_by_pid(pid);
4706 retval = security_task_getscheduler(p);
4710 raw_spin_lock_irqsave(&p->pi_lock, flags);
4711 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask);
4712 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4721 * sys_sched_getaffinity - get the cpu affinity of a process
4722 * @pid: pid of the process
4723 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4724 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4726 * Return: 0 on success. An error code otherwise.
4728 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4729 unsigned long __user *, user_mask_ptr)
4734 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4736 if (len & (sizeof(unsigned long)-1))
4739 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4742 ret = sched_getaffinity(pid, mask);
4744 size_t retlen = min_t(size_t, len, cpumask_size());
4746 if (copy_to_user(user_mask_ptr, mask, retlen))
4751 free_cpumask_var(mask);
4757 * sys_sched_yield - yield the current processor to other threads.
4759 * This function yields the current CPU to other tasks. If there are no
4760 * other threads running on this CPU then this function will return.
4764 SYSCALL_DEFINE0(sched_yield)
4766 struct rq *rq = this_rq_lock();
4768 schedstat_inc(rq, yld_count);
4769 current->sched_class->yield_task(rq);
4772 * Since we are going to call schedule() anyway, there's
4773 * no need to preempt or enable interrupts:
4775 __release(rq->lock);
4776 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4777 do_raw_spin_unlock(&rq->lock);
4778 sched_preempt_enable_no_resched();
4785 int __sched _cond_resched(void)
4787 if (should_resched(0)) {
4788 preempt_schedule_common();
4793 EXPORT_SYMBOL(_cond_resched);
4796 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4797 * call schedule, and on return reacquire the lock.
4799 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4800 * operations here to prevent schedule() from being called twice (once via
4801 * spin_unlock(), once by hand).
4803 int __cond_resched_lock(spinlock_t *lock)
4805 int resched = should_resched(PREEMPT_LOCK_OFFSET);
4808 lockdep_assert_held(lock);
4810 if (spin_needbreak(lock) || resched) {
4813 preempt_schedule_common();
4821 EXPORT_SYMBOL(__cond_resched_lock);
4823 int __sched __cond_resched_softirq(void)
4825 BUG_ON(!in_softirq());
4827 if (should_resched(SOFTIRQ_DISABLE_OFFSET)) {
4829 preempt_schedule_common();
4835 EXPORT_SYMBOL(__cond_resched_softirq);
4838 * yield - yield the current processor to other threads.
4840 * Do not ever use this function, there's a 99% chance you're doing it wrong.
4842 * The scheduler is at all times free to pick the calling task as the most
4843 * eligible task to run, if removing the yield() call from your code breaks
4844 * it, its already broken.
4846 * Typical broken usage is:
4851 * where one assumes that yield() will let 'the other' process run that will
4852 * make event true. If the current task is a SCHED_FIFO task that will never
4853 * happen. Never use yield() as a progress guarantee!!
4855 * If you want to use yield() to wait for something, use wait_event().
4856 * If you want to use yield() to be 'nice' for others, use cond_resched().
4857 * If you still want to use yield(), do not!
4859 void __sched yield(void)
4861 set_current_state(TASK_RUNNING);
4864 EXPORT_SYMBOL(yield);
4867 * yield_to - yield the current processor to another thread in
4868 * your thread group, or accelerate that thread toward the
4869 * processor it's on.
4871 * @preempt: whether task preemption is allowed or not
4873 * It's the caller's job to ensure that the target task struct
4874 * can't go away on us before we can do any checks.
4877 * true (>0) if we indeed boosted the target task.
4878 * false (0) if we failed to boost the target.
4879 * -ESRCH if there's no task to yield to.
4881 int __sched yield_to(struct task_struct *p, bool preempt)
4883 struct task_struct *curr = current;
4884 struct rq *rq, *p_rq;
4885 unsigned long flags;
4888 local_irq_save(flags);
4894 * If we're the only runnable task on the rq and target rq also
4895 * has only one task, there's absolutely no point in yielding.
4897 if (rq->nr_running == 1 && p_rq->nr_running == 1) {
4902 double_rq_lock(rq, p_rq);
4903 if (task_rq(p) != p_rq) {
4904 double_rq_unlock(rq, p_rq);
4908 if (!curr->sched_class->yield_to_task)
4911 if (curr->sched_class != p->sched_class)
4914 if (task_running(p_rq, p) || p->state)
4917 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
4919 schedstat_inc(rq, yld_count);
4921 * Make p's CPU reschedule; pick_next_entity takes care of
4924 if (preempt && rq != p_rq)
4929 double_rq_unlock(rq, p_rq);
4931 local_irq_restore(flags);
4938 EXPORT_SYMBOL_GPL(yield_to);
4941 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4942 * that process accounting knows that this is a task in IO wait state.
4944 long __sched io_schedule_timeout(long timeout)
4946 int old_iowait = current->in_iowait;
4950 current->in_iowait = 1;
4951 blk_schedule_flush_plug(current);
4953 delayacct_blkio_start();
4955 atomic_inc(&rq->nr_iowait);
4956 ret = schedule_timeout(timeout);
4957 current->in_iowait = old_iowait;
4958 atomic_dec(&rq->nr_iowait);
4959 delayacct_blkio_end();
4963 EXPORT_SYMBOL(io_schedule_timeout);
4966 * sys_sched_get_priority_max - return maximum RT priority.
4967 * @policy: scheduling class.
4969 * Return: On success, this syscall returns the maximum
4970 * rt_priority that can be used by a given scheduling class.
4971 * On failure, a negative error code is returned.
4973 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
4980 ret = MAX_USER_RT_PRIO-1;
4982 case SCHED_DEADLINE:
4993 * sys_sched_get_priority_min - return minimum RT priority.
4994 * @policy: scheduling class.
4996 * Return: On success, this syscall returns the minimum
4997 * rt_priority that can be used by a given scheduling class.
4998 * On failure, a negative error code is returned.
5000 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5009 case SCHED_DEADLINE:
5019 * sys_sched_rr_get_interval - return the default timeslice of a process.
5020 * @pid: pid of the process.
5021 * @interval: userspace pointer to the timeslice value.
5023 * this syscall writes the default timeslice value of a given process
5024 * into the user-space timespec buffer. A value of '0' means infinity.
5026 * Return: On success, 0 and the timeslice is in @interval. Otherwise,
5029 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5030 struct timespec __user *, interval)
5032 struct task_struct *p;
5033 unsigned int time_slice;
5034 unsigned long flags;
5044 p = find_process_by_pid(pid);
5048 retval = security_task_getscheduler(p);
5052 rq = task_rq_lock(p, &flags);
5054 if (p->sched_class->get_rr_interval)
5055 time_slice = p->sched_class->get_rr_interval(rq, p);
5056 task_rq_unlock(rq, p, &flags);
5059 jiffies_to_timespec(time_slice, &t);
5060 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5068 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5070 void sched_show_task(struct task_struct *p)
5072 unsigned long free = 0;
5074 unsigned long state = p->state;
5077 state = __ffs(state) + 1;
5078 printk(KERN_INFO "%-15.15s %c", p->comm,
5079 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5080 #if BITS_PER_LONG == 32
5081 if (state == TASK_RUNNING)
5082 printk(KERN_CONT " running ");
5084 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5086 if (state == TASK_RUNNING)
5087 printk(KERN_CONT " running task ");
5089 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5091 #ifdef CONFIG_DEBUG_STACK_USAGE
5092 free = stack_not_used(p);
5097 ppid = task_pid_nr(rcu_dereference(p->real_parent));
5099 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5100 task_pid_nr(p), ppid,
5101 (unsigned long)task_thread_info(p)->flags);
5103 print_worker_info(KERN_INFO, p);
5104 show_stack(p, NULL);
5107 void show_state_filter(unsigned long state_filter)
5109 struct task_struct *g, *p;
5111 #if BITS_PER_LONG == 32
5113 " task PC stack pid father\n");
5116 " task PC stack pid father\n");
5119 for_each_process_thread(g, p) {
5121 * reset the NMI-timeout, listing all files on a slow
5122 * console might take a lot of time:
5123 * Also, reset softlockup watchdogs on all CPUs, because
5124 * another CPU might be blocked waiting for us to process
5127 touch_nmi_watchdog();
5128 touch_all_softlockup_watchdogs();
5129 if (!state_filter || (p->state & state_filter))
5133 #ifdef CONFIG_SCHED_DEBUG
5134 sysrq_sched_debug_show();
5138 * Only show locks if all tasks are dumped:
5141 debug_show_all_locks();
5144 void init_idle_bootup_task(struct task_struct *idle)
5146 idle->sched_class = &idle_sched_class;
5150 * init_idle - set up an idle thread for a given CPU
5151 * @idle: task in question
5152 * @cpu: cpu the idle task belongs to
5154 * NOTE: this function does not set the idle thread's NEED_RESCHED
5155 * flag, to make booting more robust.
5157 void init_idle(struct task_struct *idle, int cpu)
5159 struct rq *rq = cpu_rq(cpu);
5160 unsigned long flags;
5162 raw_spin_lock_irqsave(&idle->pi_lock, flags);
5163 raw_spin_lock(&rq->lock);
5165 __sched_fork(0, idle);
5167 idle->state = TASK_RUNNING;
5168 idle->se.exec_start = sched_clock();
5172 * Its possible that init_idle() gets called multiple times on a task,
5173 * in that case do_set_cpus_allowed() will not do the right thing.
5175 * And since this is boot we can forgo the serialization.
5177 set_cpus_allowed_common(idle, cpumask_of(cpu));
5180 * We're having a chicken and egg problem, even though we are
5181 * holding rq->lock, the cpu isn't yet set to this cpu so the
5182 * lockdep check in task_group() will fail.
5184 * Similar case to sched_fork(). / Alternatively we could
5185 * use task_rq_lock() here and obtain the other rq->lock.
5190 __set_task_cpu(idle, cpu);
5193 rq->curr = rq->idle = idle;
5194 idle->on_rq = TASK_ON_RQ_QUEUED;
5198 raw_spin_unlock(&rq->lock);
5199 raw_spin_unlock_irqrestore(&idle->pi_lock, flags);
5201 /* Set the preempt count _outside_ the spinlocks! */
5202 init_idle_preempt_count(idle, cpu);
5205 * The idle tasks have their own, simple scheduling class:
5207 idle->sched_class = &idle_sched_class;
5208 ftrace_graph_init_idle_task(idle, cpu);
5209 vtime_init_idle(idle, cpu);
5211 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu);
5215 int cpuset_cpumask_can_shrink(const struct cpumask *cur,
5216 const struct cpumask *trial)
5218 int ret = 1, trial_cpus;
5219 struct dl_bw *cur_dl_b;
5220 unsigned long flags;
5222 if (!cpumask_weight(cur))
5225 rcu_read_lock_sched();
5226 cur_dl_b = dl_bw_of(cpumask_any(cur));
5227 trial_cpus = cpumask_weight(trial);
5229 raw_spin_lock_irqsave(&cur_dl_b->lock, flags);
5230 if (cur_dl_b->bw != -1 &&
5231 cur_dl_b->bw * trial_cpus < cur_dl_b->total_bw)
5233 raw_spin_unlock_irqrestore(&cur_dl_b->lock, flags);
5234 rcu_read_unlock_sched();
5239 int task_can_attach(struct task_struct *p,
5240 const struct cpumask *cs_cpus_allowed)
5245 * Kthreads which disallow setaffinity shouldn't be moved
5246 * to a new cpuset; we don't want to change their cpu
5247 * affinity and isolating such threads by their set of
5248 * allowed nodes is unnecessary. Thus, cpusets are not
5249 * applicable for such threads. This prevents checking for
5250 * success of set_cpus_allowed_ptr() on all attached tasks
5251 * before cpus_allowed may be changed.
5253 if (p->flags & PF_NO_SETAFFINITY) {
5259 if (dl_task(p) && !cpumask_intersects(task_rq(p)->rd->span,
5261 unsigned int dest_cpu = cpumask_any_and(cpu_active_mask,
5266 unsigned long flags;
5268 rcu_read_lock_sched();
5269 dl_b = dl_bw_of(dest_cpu);
5270 raw_spin_lock_irqsave(&dl_b->lock, flags);
5271 cpus = dl_bw_cpus(dest_cpu);
5272 overflow = __dl_overflow(dl_b, cpus, 0, p->dl.dl_bw);
5277 * We reserve space for this task in the destination
5278 * root_domain, as we can't fail after this point.
5279 * We will free resources in the source root_domain
5280 * later on (see set_cpus_allowed_dl()).
5282 __dl_add(dl_b, p->dl.dl_bw);
5284 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
5285 rcu_read_unlock_sched();
5295 #ifdef CONFIG_NUMA_BALANCING
5296 /* Migrate current task p to target_cpu */
5297 int migrate_task_to(struct task_struct *p, int target_cpu)
5299 struct migration_arg arg = { p, target_cpu };
5300 int curr_cpu = task_cpu(p);
5302 if (curr_cpu == target_cpu)
5305 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p)))
5308 /* TODO: This is not properly updating schedstats */
5310 trace_sched_move_numa(p, curr_cpu, target_cpu);
5311 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg);
5315 * Requeue a task on a given node and accurately track the number of NUMA
5316 * tasks on the runqueues
5318 void sched_setnuma(struct task_struct *p, int nid)
5321 unsigned long flags;
5322 bool queued, running;
5324 rq = task_rq_lock(p, &flags);
5325 queued = task_on_rq_queued(p);
5326 running = task_current(rq, p);
5329 dequeue_task(rq, p, DEQUEUE_SAVE);
5331 put_prev_task(rq, p);
5333 p->numa_preferred_nid = nid;
5336 p->sched_class->set_curr_task(rq);
5338 enqueue_task(rq, p, ENQUEUE_RESTORE);
5339 task_rq_unlock(rq, p, &flags);
5341 #endif /* CONFIG_NUMA_BALANCING */
5343 #ifdef CONFIG_HOTPLUG_CPU
5345 * Ensures that the idle task is using init_mm right before its cpu goes
5348 void idle_task_exit(void)
5350 struct mm_struct *mm = current->active_mm;
5352 BUG_ON(cpu_online(smp_processor_id()));
5354 if (mm != &init_mm) {
5355 switch_mm(mm, &init_mm, current);
5356 finish_arch_post_lock_switch();
5362 * Since this CPU is going 'away' for a while, fold any nr_active delta
5363 * we might have. Assumes we're called after migrate_tasks() so that the
5364 * nr_active count is stable.
5366 * Also see the comment "Global load-average calculations".
5368 static void calc_load_migrate(struct rq *rq)
5370 long delta = calc_load_fold_active(rq);
5372 atomic_long_add(delta, &calc_load_tasks);
5375 static void put_prev_task_fake(struct rq *rq, struct task_struct *prev)
5379 static const struct sched_class fake_sched_class = {
5380 .put_prev_task = put_prev_task_fake,
5383 static struct task_struct fake_task = {
5385 * Avoid pull_{rt,dl}_task()
5387 .prio = MAX_PRIO + 1,
5388 .sched_class = &fake_sched_class,
5392 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5393 * try_to_wake_up()->select_task_rq().
5395 * Called with rq->lock held even though we'er in stop_machine() and
5396 * there's no concurrency possible, we hold the required locks anyway
5397 * because of lock validation efforts.
5399 static void migrate_tasks(struct rq *dead_rq)
5401 struct rq *rq = dead_rq;
5402 struct task_struct *next, *stop = rq->stop;
5406 * Fudge the rq selection such that the below task selection loop
5407 * doesn't get stuck on the currently eligible stop task.
5409 * We're currently inside stop_machine() and the rq is either stuck
5410 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5411 * either way we should never end up calling schedule() until we're
5417 * put_prev_task() and pick_next_task() sched
5418 * class method both need to have an up-to-date
5419 * value of rq->clock[_task]
5421 update_rq_clock(rq);
5425 * There's this thread running, bail when that's the only
5428 if (rq->nr_running == 1)
5432 * pick_next_task assumes pinned rq->lock.
5434 lockdep_pin_lock(&rq->lock);
5435 next = pick_next_task(rq, &fake_task);
5437 next->sched_class->put_prev_task(rq, next);
5440 * Rules for changing task_struct::cpus_allowed are holding
5441 * both pi_lock and rq->lock, such that holding either
5442 * stabilizes the mask.
5444 * Drop rq->lock is not quite as disastrous as it usually is
5445 * because !cpu_active at this point, which means load-balance
5446 * will not interfere. Also, stop-machine.
5448 lockdep_unpin_lock(&rq->lock);
5449 raw_spin_unlock(&rq->lock);
5450 raw_spin_lock(&next->pi_lock);
5451 raw_spin_lock(&rq->lock);
5454 * Since we're inside stop-machine, _nothing_ should have
5455 * changed the task, WARN if weird stuff happened, because in
5456 * that case the above rq->lock drop is a fail too.
5458 if (WARN_ON(task_rq(next) != rq || !task_on_rq_queued(next))) {
5459 raw_spin_unlock(&next->pi_lock);
5463 /* Find suitable destination for @next, with force if needed. */
5464 dest_cpu = select_fallback_rq(dead_rq->cpu, next);
5466 rq = __migrate_task(rq, next, dest_cpu);
5467 if (rq != dead_rq) {
5468 raw_spin_unlock(&rq->lock);
5470 raw_spin_lock(&rq->lock);
5472 raw_spin_unlock(&next->pi_lock);
5477 #endif /* CONFIG_HOTPLUG_CPU */
5479 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5481 static struct ctl_table sd_ctl_dir[] = {
5483 .procname = "sched_domain",
5489 static struct ctl_table sd_ctl_root[] = {
5491 .procname = "kernel",
5493 .child = sd_ctl_dir,
5498 static struct ctl_table *sd_alloc_ctl_entry(int n)
5500 struct ctl_table *entry =
5501 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5506 static void sd_free_ctl_entry(struct ctl_table **tablep)
5508 struct ctl_table *entry;
5511 * In the intermediate directories, both the child directory and
5512 * procname are dynamically allocated and could fail but the mode
5513 * will always be set. In the lowest directory the names are
5514 * static strings and all have proc handlers.
5516 for (entry = *tablep; entry->mode; entry++) {
5518 sd_free_ctl_entry(&entry->child);
5519 if (entry->proc_handler == NULL)
5520 kfree(entry->procname);
5527 static int min_load_idx = 0;
5528 static int max_load_idx = CPU_LOAD_IDX_MAX-1;
5531 set_table_entry(struct ctl_table *entry,
5532 const char *procname, void *data, int maxlen,
5533 umode_t mode, proc_handler *proc_handler,
5536 entry->procname = procname;
5538 entry->maxlen = maxlen;
5540 entry->proc_handler = proc_handler;
5543 entry->extra1 = &min_load_idx;
5544 entry->extra2 = &max_load_idx;
5548 static struct ctl_table *
5549 sd_alloc_ctl_energy_table(struct sched_group_energy *sge)
5551 struct ctl_table *table = sd_alloc_ctl_entry(5);
5556 set_table_entry(&table[0], "nr_idle_states", &sge->nr_idle_states,
5557 sizeof(int), 0644, proc_dointvec_minmax, false);
5558 set_table_entry(&table[1], "idle_states", &sge->idle_states[0].power,
5559 sge->nr_idle_states*sizeof(struct idle_state), 0644,
5560 proc_doulongvec_minmax, false);
5561 set_table_entry(&table[2], "nr_cap_states", &sge->nr_cap_states,
5562 sizeof(int), 0644, proc_dointvec_minmax, false);
5563 set_table_entry(&table[3], "cap_states", &sge->cap_states[0].cap,
5564 sge->nr_cap_states*sizeof(struct capacity_state), 0644,
5565 proc_doulongvec_minmax, false);
5570 static struct ctl_table *
5571 sd_alloc_ctl_group_table(struct sched_group *sg)
5573 struct ctl_table *table = sd_alloc_ctl_entry(2);
5578 table->procname = kstrdup("energy", GFP_KERNEL);
5580 table->child = sd_alloc_ctl_energy_table((struct sched_group_energy *)sg->sge);
5585 static struct ctl_table *
5586 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5588 struct ctl_table *table;
5589 unsigned int nr_entries = 14;
5592 struct sched_group *sg = sd->groups;
5597 do {} while (nr_sgs++, sg = sg->next, sg != sd->groups);
5599 nr_entries += nr_sgs;
5602 table = sd_alloc_ctl_entry(nr_entries);
5607 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5608 sizeof(long), 0644, proc_doulongvec_minmax, false);
5609 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5610 sizeof(long), 0644, proc_doulongvec_minmax, false);
5611 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5612 sizeof(int), 0644, proc_dointvec_minmax, true);
5613 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5614 sizeof(int), 0644, proc_dointvec_minmax, true);
5615 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5616 sizeof(int), 0644, proc_dointvec_minmax, true);
5617 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5618 sizeof(int), 0644, proc_dointvec_minmax, true);
5619 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5620 sizeof(int), 0644, proc_dointvec_minmax, true);
5621 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5622 sizeof(int), 0644, proc_dointvec_minmax, false);
5623 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5624 sizeof(int), 0644, proc_dointvec_minmax, false);
5625 set_table_entry(&table[9], "cache_nice_tries",
5626 &sd->cache_nice_tries,
5627 sizeof(int), 0644, proc_dointvec_minmax, false);
5628 set_table_entry(&table[10], "flags", &sd->flags,
5629 sizeof(int), 0644, proc_dointvec_minmax, false);
5630 set_table_entry(&table[11], "max_newidle_lb_cost",
5631 &sd->max_newidle_lb_cost,
5632 sizeof(long), 0644, proc_doulongvec_minmax, false);
5633 set_table_entry(&table[12], "name", sd->name,
5634 CORENAME_MAX_SIZE, 0444, proc_dostring, false);
5638 struct ctl_table *entry = &table[13];
5641 snprintf(buf, 32, "group%d", i);
5642 entry->procname = kstrdup(buf, GFP_KERNEL);
5644 entry->child = sd_alloc_ctl_group_table(sg);
5645 } while (entry++, i++, sg = sg->next, sg != sd->groups);
5647 /* &table[nr_entries-1] is terminator */
5652 static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5654 struct ctl_table *entry, *table;
5655 struct sched_domain *sd;
5656 int domain_num = 0, i;
5659 for_each_domain(cpu, sd)
5661 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5666 for_each_domain(cpu, sd) {
5667 snprintf(buf, 32, "domain%d", i);
5668 entry->procname = kstrdup(buf, GFP_KERNEL);
5670 entry->child = sd_alloc_ctl_domain_table(sd);
5677 static struct ctl_table_header *sd_sysctl_header;
5678 static void register_sched_domain_sysctl(void)
5680 int i, cpu_num = num_possible_cpus();
5681 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5684 WARN_ON(sd_ctl_dir[0].child);
5685 sd_ctl_dir[0].child = entry;
5690 for_each_possible_cpu(i) {
5691 snprintf(buf, 32, "cpu%d", i);
5692 entry->procname = kstrdup(buf, GFP_KERNEL);
5694 entry->child = sd_alloc_ctl_cpu_table(i);
5698 WARN_ON(sd_sysctl_header);
5699 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5702 /* may be called multiple times per register */
5703 static void unregister_sched_domain_sysctl(void)
5705 unregister_sysctl_table(sd_sysctl_header);
5706 sd_sysctl_header = NULL;
5707 if (sd_ctl_dir[0].child)
5708 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5711 static void register_sched_domain_sysctl(void)
5714 static void unregister_sched_domain_sysctl(void)
5717 #endif /* CONFIG_SCHED_DEBUG && CONFIG_SYSCTL */
5719 static void set_rq_online(struct rq *rq)
5722 const struct sched_class *class;
5724 cpumask_set_cpu(rq->cpu, rq->rd->online);
5727 for_each_class(class) {
5728 if (class->rq_online)
5729 class->rq_online(rq);
5734 static void set_rq_offline(struct rq *rq)
5737 const struct sched_class *class;
5739 for_each_class(class) {
5740 if (class->rq_offline)
5741 class->rq_offline(rq);
5744 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5750 * migration_call - callback that gets triggered when a CPU is added.
5751 * Here we can start up the necessary migration thread for the new CPU.
5754 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5756 int cpu = (long)hcpu;
5757 unsigned long flags;
5758 struct rq *rq = cpu_rq(cpu);
5760 switch (action & ~CPU_TASKS_FROZEN) {
5762 case CPU_UP_PREPARE:
5763 raw_spin_lock_irqsave(&rq->lock, flags);
5764 walt_set_window_start(rq);
5765 raw_spin_unlock_irqrestore(&rq->lock, flags);
5766 rq->calc_load_update = calc_load_update;
5767 account_reset_rq(rq);
5771 /* Update our root-domain */
5772 raw_spin_lock_irqsave(&rq->lock, flags);
5774 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5778 raw_spin_unlock_irqrestore(&rq->lock, flags);
5781 #ifdef CONFIG_HOTPLUG_CPU
5783 sched_ttwu_pending();
5784 /* Update our root-domain */
5785 raw_spin_lock_irqsave(&rq->lock, flags);
5786 walt_migrate_sync_cpu(cpu);
5788 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5792 BUG_ON(rq->nr_running != 1); /* the migration thread */
5793 raw_spin_unlock_irqrestore(&rq->lock, flags);
5797 calc_load_migrate(rq);
5802 update_max_interval();
5808 * Register at high priority so that task migration (migrate_all_tasks)
5809 * happens before everything else. This has to be lower priority than
5810 * the notifier in the perf_event subsystem, though.
5812 static struct notifier_block migration_notifier = {
5813 .notifier_call = migration_call,
5814 .priority = CPU_PRI_MIGRATION,
5817 static void set_cpu_rq_start_time(void)
5819 int cpu = smp_processor_id();
5820 struct rq *rq = cpu_rq(cpu);
5821 rq->age_stamp = sched_clock_cpu(cpu);
5824 static int sched_cpu_active(struct notifier_block *nfb,
5825 unsigned long action, void *hcpu)
5827 int cpu = (long)hcpu;
5829 switch (action & ~CPU_TASKS_FROZEN) {
5831 set_cpu_rq_start_time();
5836 * At this point a starting CPU has marked itself as online via
5837 * set_cpu_online(). But it might not yet have marked itself
5838 * as active, which is essential from here on.
5840 set_cpu_active(cpu, true);
5841 stop_machine_unpark(cpu);
5844 case CPU_DOWN_FAILED:
5845 set_cpu_active(cpu, true);
5853 static int sched_cpu_inactive(struct notifier_block *nfb,
5854 unsigned long action, void *hcpu)
5856 switch (action & ~CPU_TASKS_FROZEN) {
5857 case CPU_DOWN_PREPARE:
5858 set_cpu_active((long)hcpu, false);
5865 static int __init migration_init(void)
5867 void *cpu = (void *)(long)smp_processor_id();
5870 /* Initialize migration for the boot CPU */
5871 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5872 BUG_ON(err == NOTIFY_BAD);
5873 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5874 register_cpu_notifier(&migration_notifier);
5876 /* Register cpu active notifiers */
5877 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5878 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5882 early_initcall(migration_init);
5884 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
5886 #ifdef CONFIG_SCHED_DEBUG
5888 static __read_mostly int sched_debug_enabled;
5890 static int __init sched_debug_setup(char *str)
5892 sched_debug_enabled = 1;
5896 early_param("sched_debug", sched_debug_setup);
5898 static inline bool sched_debug(void)
5900 return sched_debug_enabled;
5903 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5904 struct cpumask *groupmask)
5906 struct sched_group *group = sd->groups;
5908 cpumask_clear(groupmask);
5910 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5912 if (!(sd->flags & SD_LOAD_BALANCE)) {
5913 printk("does not load-balance\n");
5915 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5920 printk(KERN_CONT "span %*pbl level %s\n",
5921 cpumask_pr_args(sched_domain_span(sd)), sd->name);
5923 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5924 printk(KERN_ERR "ERROR: domain->span does not contain "
5927 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5928 printk(KERN_ERR "ERROR: domain->groups does not contain"
5932 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5936 printk(KERN_ERR "ERROR: group is NULL\n");
5940 if (!cpumask_weight(sched_group_cpus(group))) {
5941 printk(KERN_CONT "\n");
5942 printk(KERN_ERR "ERROR: empty group\n");
5946 if (!(sd->flags & SD_OVERLAP) &&
5947 cpumask_intersects(groupmask, sched_group_cpus(group))) {
5948 printk(KERN_CONT "\n");
5949 printk(KERN_ERR "ERROR: repeated CPUs\n");
5953 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
5955 printk(KERN_CONT " %*pbl",
5956 cpumask_pr_args(sched_group_cpus(group)));
5957 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) {
5958 printk(KERN_CONT " (cpu_capacity = %lu)",
5959 group->sgc->capacity);
5962 group = group->next;
5963 } while (group != sd->groups);
5964 printk(KERN_CONT "\n");
5966 if (!cpumask_equal(sched_domain_span(sd), groupmask))
5967 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
5970 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
5971 printk(KERN_ERR "ERROR: parent span is not a superset "
5972 "of domain->span\n");
5976 static void sched_domain_debug(struct sched_domain *sd, int cpu)
5980 if (!sched_debug_enabled)
5984 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
5988 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
5991 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
5999 #else /* !CONFIG_SCHED_DEBUG */
6000 # define sched_domain_debug(sd, cpu) do { } while (0)
6001 static inline bool sched_debug(void)
6005 #endif /* CONFIG_SCHED_DEBUG */
6007 static int sd_degenerate(struct sched_domain *sd)
6009 if (cpumask_weight(sched_domain_span(sd)) == 1)
6012 /* Following flags need at least 2 groups */
6013 if (sd->flags & (SD_LOAD_BALANCE |
6014 SD_BALANCE_NEWIDLE |
6017 SD_SHARE_CPUCAPACITY |
6018 SD_SHARE_PKG_RESOURCES |
6019 SD_SHARE_POWERDOMAIN |
6020 SD_SHARE_CAP_STATES)) {
6021 if (sd->groups != sd->groups->next)
6025 /* Following flags don't use groups */
6026 if (sd->flags & (SD_WAKE_AFFINE))
6033 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6035 unsigned long cflags = sd->flags, pflags = parent->flags;
6037 if (sd_degenerate(parent))
6040 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6043 /* Flags needing groups don't count if only 1 group in parent */
6044 if (parent->groups == parent->groups->next) {
6045 pflags &= ~(SD_LOAD_BALANCE |
6046 SD_BALANCE_NEWIDLE |
6049 SD_SHARE_CPUCAPACITY |
6050 SD_SHARE_PKG_RESOURCES |
6052 SD_SHARE_POWERDOMAIN |
6053 SD_SHARE_CAP_STATES);
6054 if (nr_node_ids == 1)
6055 pflags &= ~SD_SERIALIZE;
6057 if (~cflags & pflags)
6063 static void free_rootdomain(struct rcu_head *rcu)
6065 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6067 cpupri_cleanup(&rd->cpupri);
6068 cpudl_cleanup(&rd->cpudl);
6069 free_cpumask_var(rd->dlo_mask);
6070 free_cpumask_var(rd->rto_mask);
6071 free_cpumask_var(rd->online);
6072 free_cpumask_var(rd->span);
6076 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6078 struct root_domain *old_rd = NULL;
6079 unsigned long flags;
6081 raw_spin_lock_irqsave(&rq->lock, flags);
6086 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6089 cpumask_clear_cpu(rq->cpu, old_rd->span);
6092 * If we dont want to free the old_rd yet then
6093 * set old_rd to NULL to skip the freeing later
6096 if (!atomic_dec_and_test(&old_rd->refcount))
6100 atomic_inc(&rd->refcount);
6103 cpumask_set_cpu(rq->cpu, rd->span);
6104 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6107 raw_spin_unlock_irqrestore(&rq->lock, flags);
6110 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6113 static int init_rootdomain(struct root_domain *rd)
6115 memset(rd, 0, sizeof(*rd));
6117 if (!zalloc_cpumask_var(&rd->span, GFP_KERNEL))
6119 if (!zalloc_cpumask_var(&rd->online, GFP_KERNEL))
6121 if (!zalloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL))
6123 if (!zalloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6126 init_dl_bw(&rd->dl_bw);
6127 if (cpudl_init(&rd->cpudl) != 0)
6130 if (cpupri_init(&rd->cpupri) != 0)
6133 init_max_cpu_capacity(&rd->max_cpu_capacity);
6137 free_cpumask_var(rd->rto_mask);
6139 free_cpumask_var(rd->dlo_mask);
6141 free_cpumask_var(rd->online);
6143 free_cpumask_var(rd->span);
6149 * By default the system creates a single root-domain with all cpus as
6150 * members (mimicking the global state we have today).
6152 struct root_domain def_root_domain;
6154 static void init_defrootdomain(void)
6156 init_rootdomain(&def_root_domain);
6158 atomic_set(&def_root_domain.refcount, 1);
6161 static struct root_domain *alloc_rootdomain(void)
6163 struct root_domain *rd;
6165 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6169 if (init_rootdomain(rd) != 0) {
6177 static void free_sched_groups(struct sched_group *sg, int free_sgc)
6179 struct sched_group *tmp, *first;
6188 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref))
6193 } while (sg != first);
6196 static void free_sched_domain(struct rcu_head *rcu)
6198 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6201 * If its an overlapping domain it has private groups, iterate and
6204 if (sd->flags & SD_OVERLAP) {
6205 free_sched_groups(sd->groups, 1);
6206 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6207 kfree(sd->groups->sgc);
6213 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6215 call_rcu(&sd->rcu, free_sched_domain);
6218 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6220 for (; sd; sd = sd->parent)
6221 destroy_sched_domain(sd, cpu);
6225 * Keep a special pointer to the highest sched_domain that has
6226 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
6227 * allows us to avoid some pointer chasing select_idle_sibling().
6229 * Also keep a unique ID per domain (we use the first cpu number in
6230 * the cpumask of the domain), this allows us to quickly tell if
6231 * two cpus are in the same cache domain, see cpus_share_cache().
6233 DEFINE_PER_CPU(struct sched_domain *, sd_llc);
6234 DEFINE_PER_CPU(int, sd_llc_size);
6235 DEFINE_PER_CPU(int, sd_llc_id);
6236 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
6237 DEFINE_PER_CPU(struct sched_domain *, sd_busy);
6238 DEFINE_PER_CPU(struct sched_domain *, sd_asym);
6239 DEFINE_PER_CPU(struct sched_domain *, sd_ea);
6240 DEFINE_PER_CPU(struct sched_domain *, sd_scs);
6242 static void update_top_cache_domain(int cpu)
6244 struct sched_domain *sd;
6245 struct sched_domain *busy_sd = NULL, *ea_sd = NULL;
6249 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
6251 id = cpumask_first(sched_domain_span(sd));
6252 size = cpumask_weight(sched_domain_span(sd));
6253 busy_sd = sd->parent; /* sd_busy */
6255 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd);
6257 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);
6258 per_cpu(sd_llc_size, cpu) = size;
6259 per_cpu(sd_llc_id, cpu) = id;
6261 sd = lowest_flag_domain(cpu, SD_NUMA);
6262 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
6264 sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
6265 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
6267 for_each_domain(cpu, sd) {
6268 if (sd->groups->sge)
6273 rcu_assign_pointer(per_cpu(sd_ea, cpu), ea_sd);
6275 sd = highest_flag_domain(cpu, SD_SHARE_CAP_STATES);
6276 rcu_assign_pointer(per_cpu(sd_scs, cpu), sd);
6280 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6281 * hold the hotplug lock.
6284 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6286 struct rq *rq = cpu_rq(cpu);
6287 struct sched_domain *tmp;
6289 /* Remove the sched domains which do not contribute to scheduling. */
6290 for (tmp = sd; tmp; ) {
6291 struct sched_domain *parent = tmp->parent;
6295 if (sd_parent_degenerate(tmp, parent)) {
6296 tmp->parent = parent->parent;
6298 parent->parent->child = tmp;
6300 * Transfer SD_PREFER_SIBLING down in case of a
6301 * degenerate parent; the spans match for this
6302 * so the property transfers.
6304 if (parent->flags & SD_PREFER_SIBLING)
6305 tmp->flags |= SD_PREFER_SIBLING;
6306 destroy_sched_domain(parent, cpu);
6311 if (sd && sd_degenerate(sd)) {
6314 destroy_sched_domain(tmp, cpu);
6319 sched_domain_debug(sd, cpu);
6321 rq_attach_root(rq, rd);
6323 rcu_assign_pointer(rq->sd, sd);
6324 destroy_sched_domains(tmp, cpu);
6326 update_top_cache_domain(cpu);
6329 /* Setup the mask of cpus configured for isolated domains */
6330 static int __init isolated_cpu_setup(char *str)
6332 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6333 cpulist_parse(str, cpu_isolated_map);
6337 __setup("isolcpus=", isolated_cpu_setup);
6340 struct sched_domain ** __percpu sd;
6341 struct root_domain *rd;
6352 * Build an iteration mask that can exclude certain CPUs from the upwards
6355 * Asymmetric node setups can result in situations where the domain tree is of
6356 * unequal depth, make sure to skip domains that already cover the entire
6359 * In that case build_sched_domains() will have terminated the iteration early
6360 * and our sibling sd spans will be empty. Domains should always include the
6361 * cpu they're built on, so check that.
6364 static void build_group_mask(struct sched_domain *sd, struct sched_group *sg)
6366 const struct cpumask *span = sched_domain_span(sd);
6367 struct sd_data *sdd = sd->private;
6368 struct sched_domain *sibling;
6371 for_each_cpu(i, span) {
6372 sibling = *per_cpu_ptr(sdd->sd, i);
6373 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6376 cpumask_set_cpu(i, sched_group_mask(sg));
6381 * Return the canonical balance cpu for this group, this is the first cpu
6382 * of this group that's also in the iteration mask.
6384 int group_balance_cpu(struct sched_group *sg)
6386 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg));
6390 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
6392 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
6393 const struct cpumask *span = sched_domain_span(sd);
6394 struct cpumask *covered = sched_domains_tmpmask;
6395 struct sd_data *sdd = sd->private;
6396 struct sched_domain *sibling;
6399 cpumask_clear(covered);
6401 for_each_cpu(i, span) {
6402 struct cpumask *sg_span;
6404 if (cpumask_test_cpu(i, covered))
6407 sibling = *per_cpu_ptr(sdd->sd, i);
6409 /* See the comment near build_group_mask(). */
6410 if (!cpumask_test_cpu(i, sched_domain_span(sibling)))
6413 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
6414 GFP_KERNEL, cpu_to_node(cpu));
6419 sg_span = sched_group_cpus(sg);
6421 cpumask_copy(sg_span, sched_domain_span(sibling->child));
6423 cpumask_set_cpu(i, sg_span);
6425 cpumask_or(covered, covered, sg_span);
6427 sg->sgc = *per_cpu_ptr(sdd->sgc, i);
6428 if (atomic_inc_return(&sg->sgc->ref) == 1)
6429 build_group_mask(sd, sg);
6432 * Initialize sgc->capacity such that even if we mess up the
6433 * domains and no possible iteration will get us here, we won't
6436 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span);
6437 sg->sgc->max_capacity = SCHED_CAPACITY_SCALE;
6440 * Make sure the first group of this domain contains the
6441 * canonical balance cpu. Otherwise the sched_domain iteration
6442 * breaks. See update_sg_lb_stats().
6444 if ((!groups && cpumask_test_cpu(cpu, sg_span)) ||
6445 group_balance_cpu(sg) == cpu)
6455 sd->groups = groups;
6460 free_sched_groups(first, 0);
6465 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
6467 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
6468 struct sched_domain *child = sd->child;
6471 cpu = cpumask_first(sched_domain_span(child));
6474 *sg = *per_cpu_ptr(sdd->sg, cpu);
6475 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu);
6476 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */
6483 * build_sched_groups will build a circular linked list of the groups
6484 * covered by the given span, and will set each group's ->cpumask correctly,
6485 * and ->cpu_capacity to 0.
6487 * Assumes the sched_domain tree is fully constructed
6490 build_sched_groups(struct sched_domain *sd, int cpu)
6492 struct sched_group *first = NULL, *last = NULL;
6493 struct sd_data *sdd = sd->private;
6494 const struct cpumask *span = sched_domain_span(sd);
6495 struct cpumask *covered;
6498 get_group(cpu, sdd, &sd->groups);
6499 atomic_inc(&sd->groups->ref);
6501 if (cpu != cpumask_first(span))
6504 lockdep_assert_held(&sched_domains_mutex);
6505 covered = sched_domains_tmpmask;
6507 cpumask_clear(covered);
6509 for_each_cpu(i, span) {
6510 struct sched_group *sg;
6513 if (cpumask_test_cpu(i, covered))
6516 group = get_group(i, sdd, &sg);
6517 cpumask_setall(sched_group_mask(sg));
6519 for_each_cpu(j, span) {
6520 if (get_group(j, sdd, NULL) != group)
6523 cpumask_set_cpu(j, covered);
6524 cpumask_set_cpu(j, sched_group_cpus(sg));
6539 * Initialize sched groups cpu_capacity.
6541 * cpu_capacity indicates the capacity of sched group, which is used while
6542 * distributing the load between different sched groups in a sched domain.
6543 * Typically cpu_capacity for all the groups in a sched domain will be same
6544 * unless there are asymmetries in the topology. If there are asymmetries,
6545 * group having more cpu_capacity will pickup more load compared to the
6546 * group having less cpu_capacity.
6548 static void init_sched_groups_capacity(int cpu, struct sched_domain *sd)
6550 struct sched_group *sg = sd->groups;
6555 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
6557 } while (sg != sd->groups);
6559 if (cpu != group_balance_cpu(sg))
6562 update_group_capacity(sd, cpu);
6563 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight);
6567 * Check that the per-cpu provided sd energy data is consistent for all cpus
6570 static inline void check_sched_energy_data(int cpu, sched_domain_energy_f fn,
6571 const struct cpumask *cpumask)
6573 const struct sched_group_energy * const sge = fn(cpu);
6574 struct cpumask mask;
6577 if (cpumask_weight(cpumask) <= 1)
6580 cpumask_xor(&mask, cpumask, get_cpu_mask(cpu));
6582 for_each_cpu(i, &mask) {
6583 const struct sched_group_energy * const e = fn(i);
6586 BUG_ON(e->nr_idle_states != sge->nr_idle_states);
6588 for (y = 0; y < (e->nr_idle_states); y++) {
6589 BUG_ON(e->idle_states[y].power !=
6590 sge->idle_states[y].power);
6593 BUG_ON(e->nr_cap_states != sge->nr_cap_states);
6595 for (y = 0; y < (e->nr_cap_states); y++) {
6596 BUG_ON(e->cap_states[y].cap != sge->cap_states[y].cap);
6597 BUG_ON(e->cap_states[y].power !=
6598 sge->cap_states[y].power);
6603 static void init_sched_energy(int cpu, struct sched_domain *sd,
6604 sched_domain_energy_f fn)
6606 if (!(fn && fn(cpu)))
6609 if (cpu != group_balance_cpu(sd->groups))
6612 if (sd->child && !sd->child->groups->sge) {
6613 pr_err("BUG: EAS setup broken for CPU%d\n", cpu);
6614 #ifdef CONFIG_SCHED_DEBUG
6615 pr_err(" energy data on %s but not on %s domain\n",
6616 sd->name, sd->child->name);
6621 check_sched_energy_data(cpu, fn, sched_group_cpus(sd->groups));
6623 sd->groups->sge = fn(cpu);
6627 * Initializers for schedule domains
6628 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6631 static int default_relax_domain_level = -1;
6632 int sched_domain_level_max;
6634 static int __init setup_relax_domain_level(char *str)
6636 if (kstrtoint(str, 0, &default_relax_domain_level))
6637 pr_warn("Unable to set relax_domain_level\n");
6641 __setup("relax_domain_level=", setup_relax_domain_level);
6643 static void set_domain_attribute(struct sched_domain *sd,
6644 struct sched_domain_attr *attr)
6648 if (!attr || attr->relax_domain_level < 0) {
6649 if (default_relax_domain_level < 0)
6652 request = default_relax_domain_level;
6654 request = attr->relax_domain_level;
6655 if (request < sd->level) {
6656 /* turn off idle balance on this domain */
6657 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6659 /* turn on idle balance on this domain */
6660 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6664 static void __sdt_free(const struct cpumask *cpu_map);
6665 static int __sdt_alloc(const struct cpumask *cpu_map);
6667 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6668 const struct cpumask *cpu_map)
6672 if (!atomic_read(&d->rd->refcount))
6673 free_rootdomain(&d->rd->rcu); /* fall through */
6675 free_percpu(d->sd); /* fall through */
6677 __sdt_free(cpu_map); /* fall through */
6683 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6684 const struct cpumask *cpu_map)
6686 memset(d, 0, sizeof(*d));
6688 if (__sdt_alloc(cpu_map))
6689 return sa_sd_storage;
6690 d->sd = alloc_percpu(struct sched_domain *);
6692 return sa_sd_storage;
6693 d->rd = alloc_rootdomain();
6696 return sa_rootdomain;
6700 * NULL the sd_data elements we've used to build the sched_domain and
6701 * sched_group structure so that the subsequent __free_domain_allocs()
6702 * will not free the data we're using.
6704 static void claim_allocations(int cpu, struct sched_domain *sd)
6706 struct sd_data *sdd = sd->private;
6708 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
6709 *per_cpu_ptr(sdd->sd, cpu) = NULL;
6711 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
6712 *per_cpu_ptr(sdd->sg, cpu) = NULL;
6714 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref))
6715 *per_cpu_ptr(sdd->sgc, cpu) = NULL;
6719 static int sched_domains_numa_levels;
6720 enum numa_topology_type sched_numa_topology_type;
6721 static int *sched_domains_numa_distance;
6722 int sched_max_numa_distance;
6723 static struct cpumask ***sched_domains_numa_masks;
6724 static int sched_domains_curr_level;
6728 * SD_flags allowed in topology descriptions.
6730 * SD_SHARE_CPUCAPACITY - describes SMT topologies
6731 * SD_SHARE_PKG_RESOURCES - describes shared caches
6732 * SD_NUMA - describes NUMA topologies
6733 * SD_SHARE_POWERDOMAIN - describes shared power domain
6734 * SD_SHARE_CAP_STATES - describes shared capacity states
6737 * SD_ASYM_PACKING - describes SMT quirks
6739 #define TOPOLOGY_SD_FLAGS \
6740 (SD_SHARE_CPUCAPACITY | \
6741 SD_SHARE_PKG_RESOURCES | \
6744 SD_SHARE_POWERDOMAIN | \
6745 SD_SHARE_CAP_STATES)
6747 static struct sched_domain *
6748 sd_init(struct sched_domain_topology_level *tl, int cpu)
6750 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu);
6751 int sd_weight, sd_flags = 0;
6755 * Ugly hack to pass state to sd_numa_mask()...
6757 sched_domains_curr_level = tl->numa_level;
6760 sd_weight = cpumask_weight(tl->mask(cpu));
6763 sd_flags = (*tl->sd_flags)();
6764 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS,
6765 "wrong sd_flags in topology description\n"))
6766 sd_flags &= ~TOPOLOGY_SD_FLAGS;
6768 *sd = (struct sched_domain){
6769 .min_interval = sd_weight,
6770 .max_interval = 2*sd_weight,
6772 .imbalance_pct = 125,
6774 .cache_nice_tries = 0,
6781 .flags = 1*SD_LOAD_BALANCE
6782 | 1*SD_BALANCE_NEWIDLE
6787 | 0*SD_SHARE_CPUCAPACITY
6788 | 0*SD_SHARE_PKG_RESOURCES
6790 | 0*SD_PREFER_SIBLING
6795 .last_balance = jiffies,
6796 .balance_interval = sd_weight,
6798 .max_newidle_lb_cost = 0,
6799 .next_decay_max_lb_cost = jiffies,
6800 #ifdef CONFIG_SCHED_DEBUG
6806 * Convert topological properties into behaviour.
6809 if (sd->flags & SD_SHARE_CPUCAPACITY) {
6810 sd->flags |= SD_PREFER_SIBLING;
6811 sd->imbalance_pct = 110;
6812 sd->smt_gain = 1178; /* ~15% */
6814 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) {
6815 sd->imbalance_pct = 117;
6816 sd->cache_nice_tries = 1;
6820 } else if (sd->flags & SD_NUMA) {
6821 sd->cache_nice_tries = 2;
6825 sd->flags |= SD_SERIALIZE;
6826 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) {
6827 sd->flags &= ~(SD_BALANCE_EXEC |
6834 sd->flags |= SD_PREFER_SIBLING;
6835 sd->cache_nice_tries = 1;
6840 sd->private = &tl->data;
6846 * Topology list, bottom-up.
6848 static struct sched_domain_topology_level default_topology[] = {
6849 #ifdef CONFIG_SCHED_SMT
6850 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
6852 #ifdef CONFIG_SCHED_MC
6853 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
6855 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
6859 static struct sched_domain_topology_level *sched_domain_topology =
6862 #define for_each_sd_topology(tl) \
6863 for (tl = sched_domain_topology; tl->mask; tl++)
6865 void set_sched_topology(struct sched_domain_topology_level *tl)
6867 sched_domain_topology = tl;
6872 static const struct cpumask *sd_numa_mask(int cpu)
6874 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)];
6877 static void sched_numa_warn(const char *str)
6879 static int done = false;
6887 printk(KERN_WARNING "ERROR: %s\n\n", str);
6889 for (i = 0; i < nr_node_ids; i++) {
6890 printk(KERN_WARNING " ");
6891 for (j = 0; j < nr_node_ids; j++)
6892 printk(KERN_CONT "%02d ", node_distance(i,j));
6893 printk(KERN_CONT "\n");
6895 printk(KERN_WARNING "\n");
6898 bool find_numa_distance(int distance)
6902 if (distance == node_distance(0, 0))
6905 for (i = 0; i < sched_domains_numa_levels; i++) {
6906 if (sched_domains_numa_distance[i] == distance)
6914 * A system can have three types of NUMA topology:
6915 * NUMA_DIRECT: all nodes are directly connected, or not a NUMA system
6916 * NUMA_GLUELESS_MESH: some nodes reachable through intermediary nodes
6917 * NUMA_BACKPLANE: nodes can reach other nodes through a backplane
6919 * The difference between a glueless mesh topology and a backplane
6920 * topology lies in whether communication between not directly
6921 * connected nodes goes through intermediary nodes (where programs
6922 * could run), or through backplane controllers. This affects
6923 * placement of programs.
6925 * The type of topology can be discerned with the following tests:
6926 * - If the maximum distance between any nodes is 1 hop, the system
6927 * is directly connected.
6928 * - If for two nodes A and B, located N > 1 hops away from each other,
6929 * there is an intermediary node C, which is < N hops away from both
6930 * nodes A and B, the system is a glueless mesh.
6932 static void init_numa_topology_type(void)
6936 n = sched_max_numa_distance;
6938 if (sched_domains_numa_levels <= 1) {
6939 sched_numa_topology_type = NUMA_DIRECT;
6943 for_each_online_node(a) {
6944 for_each_online_node(b) {
6945 /* Find two nodes furthest removed from each other. */
6946 if (node_distance(a, b) < n)
6949 /* Is there an intermediary node between a and b? */
6950 for_each_online_node(c) {
6951 if (node_distance(a, c) < n &&
6952 node_distance(b, c) < n) {
6953 sched_numa_topology_type =
6959 sched_numa_topology_type = NUMA_BACKPLANE;
6965 static void sched_init_numa(void)
6967 int next_distance, curr_distance = node_distance(0, 0);
6968 struct sched_domain_topology_level *tl;
6972 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL);
6973 if (!sched_domains_numa_distance)
6977 * O(nr_nodes^2) deduplicating selection sort -- in order to find the
6978 * unique distances in the node_distance() table.
6980 * Assumes node_distance(0,j) includes all distances in
6981 * node_distance(i,j) in order to avoid cubic time.
6983 next_distance = curr_distance;
6984 for (i = 0; i < nr_node_ids; i++) {
6985 for (j = 0; j < nr_node_ids; j++) {
6986 for (k = 0; k < nr_node_ids; k++) {
6987 int distance = node_distance(i, k);
6989 if (distance > curr_distance &&
6990 (distance < next_distance ||
6991 next_distance == curr_distance))
6992 next_distance = distance;
6995 * While not a strong assumption it would be nice to know
6996 * about cases where if node A is connected to B, B is not
6997 * equally connected to A.
6999 if (sched_debug() && node_distance(k, i) != distance)
7000 sched_numa_warn("Node-distance not symmetric");
7002 if (sched_debug() && i && !find_numa_distance(distance))
7003 sched_numa_warn("Node-0 not representative");
7005 if (next_distance != curr_distance) {
7006 sched_domains_numa_distance[level++] = next_distance;
7007 sched_domains_numa_levels = level;
7008 curr_distance = next_distance;
7013 * In case of sched_debug() we verify the above assumption.
7023 * 'level' contains the number of unique distances, excluding the
7024 * identity distance node_distance(i,i).
7026 * The sched_domains_numa_distance[] array includes the actual distance
7031 * Here, we should temporarily reset sched_domains_numa_levels to 0.
7032 * If it fails to allocate memory for array sched_domains_numa_masks[][],
7033 * the array will contain less then 'level' members. This could be
7034 * dangerous when we use it to iterate array sched_domains_numa_masks[][]
7035 * in other functions.
7037 * We reset it to 'level' at the end of this function.
7039 sched_domains_numa_levels = 0;
7041 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL);
7042 if (!sched_domains_numa_masks)
7046 * Now for each level, construct a mask per node which contains all
7047 * cpus of nodes that are that many hops away from us.
7049 for (i = 0; i < level; i++) {
7050 sched_domains_numa_masks[i] =
7051 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL);
7052 if (!sched_domains_numa_masks[i])
7055 for (j = 0; j < nr_node_ids; j++) {
7056 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL);
7060 sched_domains_numa_masks[i][j] = mask;
7063 if (node_distance(j, k) > sched_domains_numa_distance[i])
7066 cpumask_or(mask, mask, cpumask_of_node(k));
7071 /* Compute default topology size */
7072 for (i = 0; sched_domain_topology[i].mask; i++);
7074 tl = kzalloc((i + level + 1) *
7075 sizeof(struct sched_domain_topology_level), GFP_KERNEL);
7080 * Copy the default topology bits..
7082 for (i = 0; sched_domain_topology[i].mask; i++)
7083 tl[i] = sched_domain_topology[i];
7086 * .. and append 'j' levels of NUMA goodness.
7088 for (j = 0; j < level; i++, j++) {
7089 tl[i] = (struct sched_domain_topology_level){
7090 .mask = sd_numa_mask,
7091 .sd_flags = cpu_numa_flags,
7092 .flags = SDTL_OVERLAP,
7098 sched_domain_topology = tl;
7100 sched_domains_numa_levels = level;
7101 sched_max_numa_distance = sched_domains_numa_distance[level - 1];
7103 init_numa_topology_type();
7106 static void sched_domains_numa_masks_set(int cpu)
7109 int node = cpu_to_node(cpu);
7111 for (i = 0; i < sched_domains_numa_levels; i++) {
7112 for (j = 0; j < nr_node_ids; j++) {
7113 if (node_distance(j, node) <= sched_domains_numa_distance[i])
7114 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]);
7119 static void sched_domains_numa_masks_clear(int cpu)
7122 for (i = 0; i < sched_domains_numa_levels; i++) {
7123 for (j = 0; j < nr_node_ids; j++)
7124 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]);
7129 * Update sched_domains_numa_masks[level][node] array when new cpus
7132 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7133 unsigned long action,
7136 int cpu = (long)hcpu;
7138 switch (action & ~CPU_TASKS_FROZEN) {
7140 sched_domains_numa_masks_set(cpu);
7144 sched_domains_numa_masks_clear(cpu);
7154 static inline void sched_init_numa(void)
7158 static int sched_domains_numa_masks_update(struct notifier_block *nfb,
7159 unsigned long action,
7164 #endif /* CONFIG_NUMA */
7166 static int __sdt_alloc(const struct cpumask *cpu_map)
7168 struct sched_domain_topology_level *tl;
7171 for_each_sd_topology(tl) {
7172 struct sd_data *sdd = &tl->data;
7174 sdd->sd = alloc_percpu(struct sched_domain *);
7178 sdd->sg = alloc_percpu(struct sched_group *);
7182 sdd->sgc = alloc_percpu(struct sched_group_capacity *);
7186 for_each_cpu(j, cpu_map) {
7187 struct sched_domain *sd;
7188 struct sched_group *sg;
7189 struct sched_group_capacity *sgc;
7191 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7192 GFP_KERNEL, cpu_to_node(j));
7196 *per_cpu_ptr(sdd->sd, j) = sd;
7198 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7199 GFP_KERNEL, cpu_to_node(j));
7205 *per_cpu_ptr(sdd->sg, j) = sg;
7207 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(),
7208 GFP_KERNEL, cpu_to_node(j));
7212 *per_cpu_ptr(sdd->sgc, j) = sgc;
7219 static void __sdt_free(const struct cpumask *cpu_map)
7221 struct sched_domain_topology_level *tl;
7224 for_each_sd_topology(tl) {
7225 struct sd_data *sdd = &tl->data;
7227 for_each_cpu(j, cpu_map) {
7228 struct sched_domain *sd;
7231 sd = *per_cpu_ptr(sdd->sd, j);
7232 if (sd && (sd->flags & SD_OVERLAP))
7233 free_sched_groups(sd->groups, 0);
7234 kfree(*per_cpu_ptr(sdd->sd, j));
7238 kfree(*per_cpu_ptr(sdd->sg, j));
7240 kfree(*per_cpu_ptr(sdd->sgc, j));
7242 free_percpu(sdd->sd);
7244 free_percpu(sdd->sg);
7246 free_percpu(sdd->sgc);
7251 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7252 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7253 struct sched_domain *child, int cpu)
7255 struct sched_domain *sd = sd_init(tl, cpu);
7259 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7261 sd->level = child->level + 1;
7262 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7266 if (!cpumask_subset(sched_domain_span(child),
7267 sched_domain_span(sd))) {
7268 pr_err("BUG: arch topology borken\n");
7269 #ifdef CONFIG_SCHED_DEBUG
7270 pr_err(" the %s domain not a subset of the %s domain\n",
7271 child->name, sd->name);
7273 /* Fixup, ensure @sd has at least @child cpus. */
7274 cpumask_or(sched_domain_span(sd),
7275 sched_domain_span(sd),
7276 sched_domain_span(child));
7280 set_domain_attribute(sd, attr);
7286 * Build sched domains for a given set of cpus and attach the sched domains
7287 * to the individual cpus
7289 static int build_sched_domains(const struct cpumask *cpu_map,
7290 struct sched_domain_attr *attr)
7292 enum s_alloc alloc_state;
7293 struct sched_domain *sd;
7295 struct rq *rq = NULL;
7296 int i, ret = -ENOMEM;
7298 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7299 if (alloc_state != sa_rootdomain)
7302 /* Set up domains for cpus specified by the cpu_map. */
7303 for_each_cpu(i, cpu_map) {
7304 struct sched_domain_topology_level *tl;
7307 for_each_sd_topology(tl) {
7308 sd = build_sched_domain(tl, cpu_map, attr, sd, i);
7309 if (tl == sched_domain_topology)
7310 *per_cpu_ptr(d.sd, i) = sd;
7311 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7312 sd->flags |= SD_OVERLAP;
7313 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7318 /* Build the groups for the domains */
7319 for_each_cpu(i, cpu_map) {
7320 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7321 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7322 if (sd->flags & SD_OVERLAP) {
7323 if (build_overlap_sched_groups(sd, i))
7326 if (build_sched_groups(sd, i))
7332 /* Calculate CPU capacity for physical packages and nodes */
7333 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7334 struct sched_domain_topology_level *tl = sched_domain_topology;
7336 if (!cpumask_test_cpu(i, cpu_map))
7339 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent, tl++) {
7340 init_sched_energy(i, sd, tl->energy);
7341 claim_allocations(i, sd);
7342 init_sched_groups_capacity(i, sd);
7346 /* Attach the domains */
7348 for_each_cpu(i, cpu_map) {
7350 sd = *per_cpu_ptr(d.sd, i);
7351 cpu_attach_domain(sd, d.rd, i);
7357 __free_domain_allocs(&d, alloc_state, cpu_map);
7361 static cpumask_var_t *doms_cur; /* current sched domains */
7362 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7363 static struct sched_domain_attr *dattr_cur;
7364 /* attribues of custom domains in 'doms_cur' */
7367 * Special case: If a kmalloc of a doms_cur partition (array of
7368 * cpumask) fails, then fallback to a single sched domain,
7369 * as determined by the single cpumask fallback_doms.
7371 static cpumask_var_t fallback_doms;
7374 * arch_update_cpu_topology lets virtualized architectures update the
7375 * cpu core maps. It is supposed to return 1 if the topology changed
7376 * or 0 if it stayed the same.
7378 int __weak arch_update_cpu_topology(void)
7383 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7386 cpumask_var_t *doms;
7388 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7391 for (i = 0; i < ndoms; i++) {
7392 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7393 free_sched_domains(doms, i);
7400 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7403 for (i = 0; i < ndoms; i++)
7404 free_cpumask_var(doms[i]);
7409 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7410 * For now this just excludes isolated cpus, but could be used to
7411 * exclude other special cases in the future.
7413 static int init_sched_domains(const struct cpumask *cpu_map)
7417 arch_update_cpu_topology();
7419 doms_cur = alloc_sched_domains(ndoms_cur);
7421 doms_cur = &fallback_doms;
7422 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7423 err = build_sched_domains(doms_cur[0], NULL);
7424 register_sched_domain_sysctl();
7430 * Detach sched domains from a group of cpus specified in cpu_map
7431 * These cpus will now be attached to the NULL domain
7433 static void detach_destroy_domains(const struct cpumask *cpu_map)
7438 for_each_cpu(i, cpu_map)
7439 cpu_attach_domain(NULL, &def_root_domain, i);
7443 /* handle null as "default" */
7444 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7445 struct sched_domain_attr *new, int idx_new)
7447 struct sched_domain_attr tmp;
7454 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7455 new ? (new + idx_new) : &tmp,
7456 sizeof(struct sched_domain_attr));
7460 * Partition sched domains as specified by the 'ndoms_new'
7461 * cpumasks in the array doms_new[] of cpumasks. This compares
7462 * doms_new[] to the current sched domain partitioning, doms_cur[].
7463 * It destroys each deleted domain and builds each new domain.
7465 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7466 * The masks don't intersect (don't overlap.) We should setup one
7467 * sched domain for each mask. CPUs not in any of the cpumasks will
7468 * not be load balanced. If the same cpumask appears both in the
7469 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7472 * The passed in 'doms_new' should be allocated using
7473 * alloc_sched_domains. This routine takes ownership of it and will
7474 * free_sched_domains it when done with it. If the caller failed the
7475 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7476 * and partition_sched_domains() will fallback to the single partition
7477 * 'fallback_doms', it also forces the domains to be rebuilt.
7479 * If doms_new == NULL it will be replaced with cpu_online_mask.
7480 * ndoms_new == 0 is a special case for destroying existing domains,
7481 * and it will not create the default domain.
7483 * Call with hotplug lock held
7485 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7486 struct sched_domain_attr *dattr_new)
7491 mutex_lock(&sched_domains_mutex);
7493 /* always unregister in case we don't destroy any domains */
7494 unregister_sched_domain_sysctl();
7496 /* Let architecture update cpu core mappings. */
7497 new_topology = arch_update_cpu_topology();
7499 n = doms_new ? ndoms_new : 0;
7501 /* Destroy deleted domains */
7502 for (i = 0; i < ndoms_cur; i++) {
7503 for (j = 0; j < n && !new_topology; j++) {
7504 if (cpumask_equal(doms_cur[i], doms_new[j])
7505 && dattrs_equal(dattr_cur, i, dattr_new, j))
7508 /* no match - a current sched domain not in new doms_new[] */
7509 detach_destroy_domains(doms_cur[i]);
7515 if (doms_new == NULL) {
7517 doms_new = &fallback_doms;
7518 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7519 WARN_ON_ONCE(dattr_new);
7522 /* Build new domains */
7523 for (i = 0; i < ndoms_new; i++) {
7524 for (j = 0; j < n && !new_topology; j++) {
7525 if (cpumask_equal(doms_new[i], doms_cur[j])
7526 && dattrs_equal(dattr_new, i, dattr_cur, j))
7529 /* no match - add a new doms_new */
7530 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7535 /* Remember the new sched domains */
7536 if (doms_cur != &fallback_doms)
7537 free_sched_domains(doms_cur, ndoms_cur);
7538 kfree(dattr_cur); /* kfree(NULL) is safe */
7539 doms_cur = doms_new;
7540 dattr_cur = dattr_new;
7541 ndoms_cur = ndoms_new;
7543 register_sched_domain_sysctl();
7545 mutex_unlock(&sched_domains_mutex);
7548 static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */
7551 * Update cpusets according to cpu_active mask. If cpusets are
7552 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7553 * around partition_sched_domains().
7555 * If we come here as part of a suspend/resume, don't touch cpusets because we
7556 * want to restore it back to its original state upon resume anyway.
7558 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7562 case CPU_ONLINE_FROZEN:
7563 case CPU_DOWN_FAILED_FROZEN:
7566 * num_cpus_frozen tracks how many CPUs are involved in suspend
7567 * resume sequence. As long as this is not the last online
7568 * operation in the resume sequence, just build a single sched
7569 * domain, ignoring cpusets.
7572 if (likely(num_cpus_frozen)) {
7573 partition_sched_domains(1, NULL, NULL);
7578 * This is the last CPU online operation. So fall through and
7579 * restore the original sched domains by considering the
7580 * cpuset configurations.
7584 cpuset_update_active_cpus(true);
7592 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7595 unsigned long flags;
7596 long cpu = (long)hcpu;
7602 case CPU_DOWN_PREPARE:
7603 rcu_read_lock_sched();
7604 dl_b = dl_bw_of(cpu);
7606 raw_spin_lock_irqsave(&dl_b->lock, flags);
7607 cpus = dl_bw_cpus(cpu);
7608 overflow = __dl_overflow(dl_b, cpus, 0, 0);
7609 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
7611 rcu_read_unlock_sched();
7614 return notifier_from_errno(-EBUSY);
7615 cpuset_update_active_cpus(false);
7617 case CPU_DOWN_PREPARE_FROZEN:
7619 partition_sched_domains(1, NULL, NULL);
7627 void __init sched_init_smp(void)
7629 cpumask_var_t non_isolated_cpus;
7631 walt_init_cpu_efficiency();
7632 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7633 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7638 * There's no userspace yet to cause hotplug operations; hence all the
7639 * cpu masks are stable and all blatant races in the below code cannot
7642 mutex_lock(&sched_domains_mutex);
7643 init_sched_domains(cpu_active_mask);
7644 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7645 if (cpumask_empty(non_isolated_cpus))
7646 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7647 mutex_unlock(&sched_domains_mutex);
7649 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE);
7650 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7651 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7655 /* Move init over to a non-isolated CPU */
7656 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7658 sched_init_granularity();
7659 free_cpumask_var(non_isolated_cpus);
7661 init_sched_rt_class();
7662 init_sched_dl_class();
7665 void __init sched_init_smp(void)
7667 sched_init_granularity();
7669 #endif /* CONFIG_SMP */
7671 int in_sched_functions(unsigned long addr)
7673 return in_lock_functions(addr) ||
7674 (addr >= (unsigned long)__sched_text_start
7675 && addr < (unsigned long)__sched_text_end);
7678 #ifdef CONFIG_CGROUP_SCHED
7680 * Default task group.
7681 * Every task in system belongs to this group at bootup.
7683 struct task_group root_task_group;
7684 LIST_HEAD(task_groups);
7687 DECLARE_PER_CPU(cpumask_var_t, load_balance_mask);
7689 void __init sched_init(void)
7692 unsigned long alloc_size = 0, ptr;
7694 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7697 #ifdef CONFIG_RT_GROUP_SCHED
7698 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7701 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7703 #ifdef CONFIG_FAIR_GROUP_SCHED
7704 root_task_group.se = (struct sched_entity **)ptr;
7705 ptr += nr_cpu_ids * sizeof(void **);
7707 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7708 ptr += nr_cpu_ids * sizeof(void **);
7710 #endif /* CONFIG_FAIR_GROUP_SCHED */
7711 #ifdef CONFIG_RT_GROUP_SCHED
7712 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7713 ptr += nr_cpu_ids * sizeof(void **);
7715 root_task_group.rt_rq = (struct rt_rq **)ptr;
7716 ptr += nr_cpu_ids * sizeof(void **);
7718 #endif /* CONFIG_RT_GROUP_SCHED */
7720 #ifdef CONFIG_CPUMASK_OFFSTACK
7721 for_each_possible_cpu(i) {
7722 per_cpu(load_balance_mask, i) = (cpumask_var_t)kzalloc_node(
7723 cpumask_size(), GFP_KERNEL, cpu_to_node(i));
7725 #endif /* CONFIG_CPUMASK_OFFSTACK */
7727 init_rt_bandwidth(&def_rt_bandwidth,
7728 global_rt_period(), global_rt_runtime());
7729 init_dl_bandwidth(&def_dl_bandwidth,
7730 global_rt_period(), global_rt_runtime());
7733 init_defrootdomain();
7736 #ifdef CONFIG_RT_GROUP_SCHED
7737 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7738 global_rt_period(), global_rt_runtime());
7739 #endif /* CONFIG_RT_GROUP_SCHED */
7741 #ifdef CONFIG_CGROUP_SCHED
7742 list_add(&root_task_group.list, &task_groups);
7743 INIT_LIST_HEAD(&root_task_group.children);
7744 INIT_LIST_HEAD(&root_task_group.siblings);
7745 autogroup_init(&init_task);
7747 #endif /* CONFIG_CGROUP_SCHED */
7749 for_each_possible_cpu(i) {
7753 raw_spin_lock_init(&rq->lock);
7755 rq->calc_load_active = 0;
7756 rq->calc_load_update = jiffies + LOAD_FREQ;
7757 init_cfs_rq(&rq->cfs);
7758 init_rt_rq(&rq->rt);
7759 init_dl_rq(&rq->dl);
7760 #ifdef CONFIG_FAIR_GROUP_SCHED
7761 root_task_group.shares = ROOT_TASK_GROUP_LOAD;
7762 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7764 * How much cpu bandwidth does root_task_group get?
7766 * In case of task-groups formed thr' the cgroup filesystem, it
7767 * gets 100% of the cpu resources in the system. This overall
7768 * system cpu resource is divided among the tasks of
7769 * root_task_group and its child task-groups in a fair manner,
7770 * based on each entity's (task or task-group's) weight
7771 * (se->load.weight).
7773 * In other words, if root_task_group has 10 tasks of weight
7774 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7775 * then A0's share of the cpu resource is:
7777 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7779 * We achieve this by letting root_task_group's tasks sit
7780 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7782 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
7783 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
7784 #endif /* CONFIG_FAIR_GROUP_SCHED */
7786 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7787 #ifdef CONFIG_RT_GROUP_SCHED
7788 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
7791 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7792 rq->cpu_load[j] = 0;
7794 rq->last_load_update_tick = jiffies;
7799 rq->cpu_capacity = rq->cpu_capacity_orig = SCHED_CAPACITY_SCALE;
7800 rq->balance_callback = NULL;
7801 rq->active_balance = 0;
7802 rq->next_balance = jiffies;
7807 rq->avg_idle = 2*sysctl_sched_migration_cost;
7808 rq->max_idle_balance_cost = sysctl_sched_migration_cost;
7809 #ifdef CONFIG_SCHED_WALT
7810 rq->cur_irqload = 0;
7811 rq->avg_irqload = 0;
7815 INIT_LIST_HEAD(&rq->cfs_tasks);
7817 rq_attach_root(rq, &def_root_domain);
7818 #ifdef CONFIG_NO_HZ_COMMON
7821 #ifdef CONFIG_NO_HZ_FULL
7822 rq->last_sched_tick = 0;
7826 atomic_set(&rq->nr_iowait, 0);
7829 set_load_weight(&init_task);
7831 #ifdef CONFIG_PREEMPT_NOTIFIERS
7832 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7836 * The boot idle thread does lazy MMU switching as well:
7838 atomic_inc(&init_mm.mm_count);
7839 enter_lazy_tlb(&init_mm, current);
7842 * During early bootup we pretend to be a normal task:
7844 current->sched_class = &fair_sched_class;
7847 * Make us the idle thread. Technically, schedule() should not be
7848 * called from this thread, however somewhere below it might be,
7849 * but because we are the idle thread, we just pick up running again
7850 * when this runqueue becomes "idle".
7852 init_idle(current, smp_processor_id());
7854 calc_load_update = jiffies + LOAD_FREQ;
7857 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
7858 /* May be allocated at isolcpus cmdline parse time */
7859 if (cpu_isolated_map == NULL)
7860 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7861 idle_thread_set_boot_cpu();
7862 set_cpu_rq_start_time();
7864 init_sched_fair_class();
7866 scheduler_running = 1;
7869 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
7870 static inline int preempt_count_equals(int preempt_offset)
7872 int nested = preempt_count() + rcu_preempt_depth();
7874 return (nested == preempt_offset);
7877 static int __might_sleep_init_called;
7878 int __init __might_sleep_init(void)
7880 __might_sleep_init_called = 1;
7883 early_initcall(__might_sleep_init);
7885 void __might_sleep(const char *file, int line, int preempt_offset)
7888 * Blocking primitives will set (and therefore destroy) current->state,
7889 * since we will exit with TASK_RUNNING make sure we enter with it,
7890 * otherwise we will destroy state.
7892 WARN_ONCE(current->state != TASK_RUNNING && current->task_state_change,
7893 "do not call blocking ops when !TASK_RUNNING; "
7894 "state=%lx set at [<%p>] %pS\n",
7896 (void *)current->task_state_change,
7897 (void *)current->task_state_change);
7899 ___might_sleep(file, line, preempt_offset);
7901 EXPORT_SYMBOL(__might_sleep);
7903 void ___might_sleep(const char *file, int line, int preempt_offset)
7905 static unsigned long prev_jiffy; /* ratelimiting */
7907 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */
7908 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() &&
7909 !is_idle_task(current)) || oops_in_progress)
7911 if (system_state != SYSTEM_RUNNING &&
7912 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7914 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7916 prev_jiffy = jiffies;
7919 "BUG: sleeping function called from invalid context at %s:%d\n",
7922 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7923 in_atomic(), irqs_disabled(),
7924 current->pid, current->comm);
7926 if (task_stack_end_corrupted(current))
7927 printk(KERN_EMERG "Thread overran stack, or stack corrupted\n");
7929 debug_show_held_locks(current);
7930 if (irqs_disabled())
7931 print_irqtrace_events(current);
7932 #ifdef CONFIG_DEBUG_PREEMPT
7933 if (!preempt_count_equals(preempt_offset)) {
7934 pr_err("Preemption disabled at:");
7935 print_ip_sym(current->preempt_disable_ip);
7941 EXPORT_SYMBOL(___might_sleep);
7944 #ifdef CONFIG_MAGIC_SYSRQ
7945 void normalize_rt_tasks(void)
7947 struct task_struct *g, *p;
7948 struct sched_attr attr = {
7949 .sched_policy = SCHED_NORMAL,
7952 read_lock(&tasklist_lock);
7953 for_each_process_thread(g, p) {
7955 * Only normalize user tasks:
7957 if (p->flags & PF_KTHREAD)
7960 p->se.exec_start = 0;
7961 #ifdef CONFIG_SCHEDSTATS
7962 p->se.statistics.wait_start = 0;
7963 p->se.statistics.sleep_start = 0;
7964 p->se.statistics.block_start = 0;
7967 if (!dl_task(p) && !rt_task(p)) {
7969 * Renice negative nice level userspace
7972 if (task_nice(p) < 0)
7973 set_user_nice(p, 0);
7977 __sched_setscheduler(p, &attr, false, false);
7979 read_unlock(&tasklist_lock);
7982 #endif /* CONFIG_MAGIC_SYSRQ */
7984 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7986 * These functions are only useful for the IA64 MCA handling, or kdb.
7988 * They can only be called when the whole system has been
7989 * stopped - every CPU needs to be quiescent, and no scheduling
7990 * activity can take place. Using them for anything else would
7991 * be a serious bug, and as a result, they aren't even visible
7992 * under any other configuration.
7996 * curr_task - return the current task for a given cpu.
7997 * @cpu: the processor in question.
7999 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8001 * Return: The current task for @cpu.
8003 struct task_struct *curr_task(int cpu)
8005 return cpu_curr(cpu);
8008 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8012 * set_curr_task - set the current task for a given cpu.
8013 * @cpu: the processor in question.
8014 * @p: the task pointer to set.
8016 * Description: This function must only be used when non-maskable interrupts
8017 * are serviced on a separate stack. It allows the architecture to switch the
8018 * notion of the current task on a cpu in a non-blocking manner. This function
8019 * must be called with all CPU's synchronized, and interrupts disabled, the
8020 * and caller must save the original value of the current task (see
8021 * curr_task() above) and restore that value before reenabling interrupts and
8022 * re-starting the system.
8024 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8026 void set_curr_task(int cpu, struct task_struct *p)
8033 #ifdef CONFIG_CGROUP_SCHED
8034 /* task_group_lock serializes the addition/removal of task groups */
8035 static DEFINE_SPINLOCK(task_group_lock);
8037 static void sched_free_group(struct task_group *tg)
8039 free_fair_sched_group(tg);
8040 free_rt_sched_group(tg);
8045 /* allocate runqueue etc for a new task group */
8046 struct task_group *sched_create_group(struct task_group *parent)
8048 struct task_group *tg;
8050 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8052 return ERR_PTR(-ENOMEM);
8054 if (!alloc_fair_sched_group(tg, parent))
8057 if (!alloc_rt_sched_group(tg, parent))
8063 sched_free_group(tg);
8064 return ERR_PTR(-ENOMEM);
8067 void sched_online_group(struct task_group *tg, struct task_group *parent)
8069 unsigned long flags;
8071 spin_lock_irqsave(&task_group_lock, flags);
8072 list_add_rcu(&tg->list, &task_groups);
8074 WARN_ON(!parent); /* root should already exist */
8076 tg->parent = parent;
8077 INIT_LIST_HEAD(&tg->children);
8078 list_add_rcu(&tg->siblings, &parent->children);
8079 spin_unlock_irqrestore(&task_group_lock, flags);
8082 /* rcu callback to free various structures associated with a task group */
8083 static void sched_free_group_rcu(struct rcu_head *rhp)
8085 /* now it should be safe to free those cfs_rqs */
8086 sched_free_group(container_of(rhp, struct task_group, rcu));
8089 void sched_destroy_group(struct task_group *tg)
8091 /* wait for possible concurrent references to cfs_rqs complete */
8092 call_rcu(&tg->rcu, sched_free_group_rcu);
8095 void sched_offline_group(struct task_group *tg)
8097 unsigned long flags;
8100 /* end participation in shares distribution */
8101 for_each_possible_cpu(i)
8102 unregister_fair_sched_group(tg, i);
8104 spin_lock_irqsave(&task_group_lock, flags);
8105 list_del_rcu(&tg->list);
8106 list_del_rcu(&tg->siblings);
8107 spin_unlock_irqrestore(&task_group_lock, flags);
8110 /* change task's runqueue when it moves between groups.
8111 * The caller of this function should have put the task in its new group
8112 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8113 * reflect its new group.
8115 void sched_move_task(struct task_struct *tsk)
8117 struct task_group *tg;
8118 int queued, running;
8119 unsigned long flags;
8122 rq = task_rq_lock(tsk, &flags);
8124 running = task_current(rq, tsk);
8125 queued = task_on_rq_queued(tsk);
8128 dequeue_task(rq, tsk, DEQUEUE_SAVE);
8129 if (unlikely(running))
8130 put_prev_task(rq, tsk);
8133 * All callers are synchronized by task_rq_lock(); we do not use RCU
8134 * which is pointless here. Thus, we pass "true" to task_css_check()
8135 * to prevent lockdep warnings.
8137 tg = container_of(task_css_check(tsk, cpu_cgrp_id, true),
8138 struct task_group, css);
8139 tg = autogroup_task_group(tsk, tg);
8140 tsk->sched_task_group = tg;
8142 #ifdef CONFIG_FAIR_GROUP_SCHED
8143 if (tsk->sched_class->task_move_group)
8144 tsk->sched_class->task_move_group(tsk);
8147 set_task_rq(tsk, task_cpu(tsk));
8149 if (unlikely(running))
8150 tsk->sched_class->set_curr_task(rq);
8152 enqueue_task(rq, tsk, ENQUEUE_RESTORE);
8154 task_rq_unlock(rq, tsk, &flags);
8156 #endif /* CONFIG_CGROUP_SCHED */
8158 #ifdef CONFIG_RT_GROUP_SCHED
8160 * Ensure that the real time constraints are schedulable.
8162 static DEFINE_MUTEX(rt_constraints_mutex);
8164 /* Must be called with tasklist_lock held */
8165 static inline int tg_has_rt_tasks(struct task_group *tg)
8167 struct task_struct *g, *p;
8170 * Autogroups do not have RT tasks; see autogroup_create().
8172 if (task_group_is_autogroup(tg))
8175 for_each_process_thread(g, p) {
8176 if (rt_task(p) && task_group(p) == tg)
8183 struct rt_schedulable_data {
8184 struct task_group *tg;
8189 static int tg_rt_schedulable(struct task_group *tg, void *data)
8191 struct rt_schedulable_data *d = data;
8192 struct task_group *child;
8193 unsigned long total, sum = 0;
8194 u64 period, runtime;
8196 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8197 runtime = tg->rt_bandwidth.rt_runtime;
8200 period = d->rt_period;
8201 runtime = d->rt_runtime;
8205 * Cannot have more runtime than the period.
8207 if (runtime > period && runtime != RUNTIME_INF)
8211 * Ensure we don't starve existing RT tasks.
8213 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8216 total = to_ratio(period, runtime);
8219 * Nobody can have more than the global setting allows.
8221 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8225 * The sum of our children's runtime should not exceed our own.
8227 list_for_each_entry_rcu(child, &tg->children, siblings) {
8228 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8229 runtime = child->rt_bandwidth.rt_runtime;
8231 if (child == d->tg) {
8232 period = d->rt_period;
8233 runtime = d->rt_runtime;
8236 sum += to_ratio(period, runtime);
8245 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8249 struct rt_schedulable_data data = {
8251 .rt_period = period,
8252 .rt_runtime = runtime,
8256 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8262 static int tg_set_rt_bandwidth(struct task_group *tg,
8263 u64 rt_period, u64 rt_runtime)
8268 * Disallowing the root group RT runtime is BAD, it would disallow the
8269 * kernel creating (and or operating) RT threads.
8271 if (tg == &root_task_group && rt_runtime == 0)
8274 /* No period doesn't make any sense. */
8278 mutex_lock(&rt_constraints_mutex);
8279 read_lock(&tasklist_lock);
8280 err = __rt_schedulable(tg, rt_period, rt_runtime);
8284 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8285 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8286 tg->rt_bandwidth.rt_runtime = rt_runtime;
8288 for_each_possible_cpu(i) {
8289 struct rt_rq *rt_rq = tg->rt_rq[i];
8291 raw_spin_lock(&rt_rq->rt_runtime_lock);
8292 rt_rq->rt_runtime = rt_runtime;
8293 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8295 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8297 read_unlock(&tasklist_lock);
8298 mutex_unlock(&rt_constraints_mutex);
8303 static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8305 u64 rt_runtime, rt_period;
8307 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8308 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8309 if (rt_runtime_us < 0)
8310 rt_runtime = RUNTIME_INF;
8312 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8315 static long sched_group_rt_runtime(struct task_group *tg)
8319 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8322 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8323 do_div(rt_runtime_us, NSEC_PER_USEC);
8324 return rt_runtime_us;
8327 static int sched_group_set_rt_period(struct task_group *tg, u64 rt_period_us)
8329 u64 rt_runtime, rt_period;
8331 rt_period = rt_period_us * NSEC_PER_USEC;
8332 rt_runtime = tg->rt_bandwidth.rt_runtime;
8334 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8337 static long sched_group_rt_period(struct task_group *tg)
8341 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8342 do_div(rt_period_us, NSEC_PER_USEC);
8343 return rt_period_us;
8345 #endif /* CONFIG_RT_GROUP_SCHED */
8347 #ifdef CONFIG_RT_GROUP_SCHED
8348 static int sched_rt_global_constraints(void)
8352 mutex_lock(&rt_constraints_mutex);
8353 read_lock(&tasklist_lock);
8354 ret = __rt_schedulable(NULL, 0, 0);
8355 read_unlock(&tasklist_lock);
8356 mutex_unlock(&rt_constraints_mutex);
8361 static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8363 /* Don't accept realtime tasks when there is no way for them to run */
8364 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8370 #else /* !CONFIG_RT_GROUP_SCHED */
8371 static int sched_rt_global_constraints(void)
8373 unsigned long flags;
8376 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8377 for_each_possible_cpu(i) {
8378 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8380 raw_spin_lock(&rt_rq->rt_runtime_lock);
8381 rt_rq->rt_runtime = global_rt_runtime();
8382 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8384 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8388 #endif /* CONFIG_RT_GROUP_SCHED */
8390 static int sched_dl_global_validate(void)
8392 u64 runtime = global_rt_runtime();
8393 u64 period = global_rt_period();
8394 u64 new_bw = to_ratio(period, runtime);
8397 unsigned long flags;
8400 * Here we want to check the bandwidth not being set to some
8401 * value smaller than the currently allocated bandwidth in
8402 * any of the root_domains.
8404 * FIXME: Cycling on all the CPUs is overdoing, but simpler than
8405 * cycling on root_domains... Discussion on different/better
8406 * solutions is welcome!
8408 for_each_possible_cpu(cpu) {
8409 rcu_read_lock_sched();
8410 dl_b = dl_bw_of(cpu);
8412 raw_spin_lock_irqsave(&dl_b->lock, flags);
8413 if (new_bw < dl_b->total_bw)
8415 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8417 rcu_read_unlock_sched();
8426 static void sched_dl_do_global(void)
8431 unsigned long flags;
8433 def_dl_bandwidth.dl_period = global_rt_period();
8434 def_dl_bandwidth.dl_runtime = global_rt_runtime();
8436 if (global_rt_runtime() != RUNTIME_INF)
8437 new_bw = to_ratio(global_rt_period(), global_rt_runtime());
8440 * FIXME: As above...
8442 for_each_possible_cpu(cpu) {
8443 rcu_read_lock_sched();
8444 dl_b = dl_bw_of(cpu);
8446 raw_spin_lock_irqsave(&dl_b->lock, flags);
8448 raw_spin_unlock_irqrestore(&dl_b->lock, flags);
8450 rcu_read_unlock_sched();
8454 static int sched_rt_global_validate(void)
8456 if (sysctl_sched_rt_period <= 0)
8459 if ((sysctl_sched_rt_runtime != RUNTIME_INF) &&
8460 (sysctl_sched_rt_runtime > sysctl_sched_rt_period))
8466 static void sched_rt_do_global(void)
8468 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8469 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period());
8472 int sched_rt_handler(struct ctl_table *table, int write,
8473 void __user *buffer, size_t *lenp,
8476 int old_period, old_runtime;
8477 static DEFINE_MUTEX(mutex);
8481 old_period = sysctl_sched_rt_period;
8482 old_runtime = sysctl_sched_rt_runtime;
8484 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8486 if (!ret && write) {
8487 ret = sched_rt_global_validate();
8491 ret = sched_dl_global_validate();
8495 ret = sched_rt_global_constraints();
8499 sched_rt_do_global();
8500 sched_dl_do_global();
8504 sysctl_sched_rt_period = old_period;
8505 sysctl_sched_rt_runtime = old_runtime;
8507 mutex_unlock(&mutex);
8512 int sched_rr_handler(struct ctl_table *table, int write,
8513 void __user *buffer, size_t *lenp,
8517 static DEFINE_MUTEX(mutex);
8520 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8521 /* make sure that internally we keep jiffies */
8522 /* also, writing zero resets timeslice to default */
8523 if (!ret && write) {
8524 sched_rr_timeslice = sched_rr_timeslice <= 0 ?
8525 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice);
8527 mutex_unlock(&mutex);
8531 #ifdef CONFIG_CGROUP_SCHED
8533 static inline struct task_group *css_tg(struct cgroup_subsys_state *css)
8535 return css ? container_of(css, struct task_group, css) : NULL;
8538 static struct cgroup_subsys_state *
8539 cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css)
8541 struct task_group *parent = css_tg(parent_css);
8542 struct task_group *tg;
8545 /* This is early initialization for the top cgroup */
8546 return &root_task_group.css;
8549 tg = sched_create_group(parent);
8551 return ERR_PTR(-ENOMEM);
8553 sched_online_group(tg, parent);
8558 static void cpu_cgroup_css_released(struct cgroup_subsys_state *css)
8560 struct task_group *tg = css_tg(css);
8562 sched_offline_group(tg);
8565 static void cpu_cgroup_css_free(struct cgroup_subsys_state *css)
8567 struct task_group *tg = css_tg(css);
8570 * Relies on the RCU grace period between css_released() and this.
8572 sched_free_group(tg);
8575 static void cpu_cgroup_fork(struct task_struct *task, void *private)
8577 sched_move_task(task);
8580 static int cpu_cgroup_can_attach(struct cgroup_taskset *tset)
8582 struct task_struct *task;
8583 struct cgroup_subsys_state *css;
8585 cgroup_taskset_for_each(task, css, tset) {
8586 #ifdef CONFIG_RT_GROUP_SCHED
8587 if (!sched_rt_can_attach(css_tg(css), task))
8590 /* We don't support RT-tasks being in separate groups */
8591 if (task->sched_class != &fair_sched_class)
8598 static void cpu_cgroup_attach(struct cgroup_taskset *tset)
8600 struct task_struct *task;
8601 struct cgroup_subsys_state *css;
8603 cgroup_taskset_for_each(task, css, tset)
8604 sched_move_task(task);
8607 #ifdef CONFIG_FAIR_GROUP_SCHED
8608 static int cpu_shares_write_u64(struct cgroup_subsys_state *css,
8609 struct cftype *cftype, u64 shareval)
8611 return sched_group_set_shares(css_tg(css), scale_load(shareval));
8614 static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css,
8617 struct task_group *tg = css_tg(css);
8619 return (u64) scale_load_down(tg->shares);
8622 #ifdef CONFIG_CFS_BANDWIDTH
8623 static DEFINE_MUTEX(cfs_constraints_mutex);
8625 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
8626 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
8628 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
8630 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
8632 int i, ret = 0, runtime_enabled, runtime_was_enabled;
8633 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8635 if (tg == &root_task_group)
8639 * Ensure we have at some amount of bandwidth every period. This is
8640 * to prevent reaching a state of large arrears when throttled via
8641 * entity_tick() resulting in prolonged exit starvation.
8643 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
8647 * Likewise, bound things on the otherside by preventing insane quota
8648 * periods. This also allows us to normalize in computing quota
8651 if (period > max_cfs_quota_period)
8655 * Prevent race between setting of cfs_rq->runtime_enabled and
8656 * unthrottle_offline_cfs_rqs().
8659 mutex_lock(&cfs_constraints_mutex);
8660 ret = __cfs_schedulable(tg, period, quota);
8664 runtime_enabled = quota != RUNTIME_INF;
8665 runtime_was_enabled = cfs_b->quota != RUNTIME_INF;
8667 * If we need to toggle cfs_bandwidth_used, off->on must occur
8668 * before making related changes, and on->off must occur afterwards
8670 if (runtime_enabled && !runtime_was_enabled)
8671 cfs_bandwidth_usage_inc();
8672 raw_spin_lock_irq(&cfs_b->lock);
8673 cfs_b->period = ns_to_ktime(period);
8674 cfs_b->quota = quota;
8676 __refill_cfs_bandwidth_runtime(cfs_b);
8677 /* restart the period timer (if active) to handle new period expiry */
8678 if (runtime_enabled)
8679 start_cfs_bandwidth(cfs_b);
8680 raw_spin_unlock_irq(&cfs_b->lock);
8682 for_each_online_cpu(i) {
8683 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
8684 struct rq *rq = cfs_rq->rq;
8686 raw_spin_lock_irq(&rq->lock);
8687 cfs_rq->runtime_enabled = runtime_enabled;
8688 cfs_rq->runtime_remaining = 0;
8690 if (cfs_rq->throttled)
8691 unthrottle_cfs_rq(cfs_rq);
8692 raw_spin_unlock_irq(&rq->lock);
8694 if (runtime_was_enabled && !runtime_enabled)
8695 cfs_bandwidth_usage_dec();
8697 mutex_unlock(&cfs_constraints_mutex);
8703 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
8707 period = ktime_to_ns(tg->cfs_bandwidth.period);
8708 if (cfs_quota_us < 0)
8709 quota = RUNTIME_INF;
8711 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
8713 return tg_set_cfs_bandwidth(tg, period, quota);
8716 long tg_get_cfs_quota(struct task_group *tg)
8720 if (tg->cfs_bandwidth.quota == RUNTIME_INF)
8723 quota_us = tg->cfs_bandwidth.quota;
8724 do_div(quota_us, NSEC_PER_USEC);
8729 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
8733 period = (u64)cfs_period_us * NSEC_PER_USEC;
8734 quota = tg->cfs_bandwidth.quota;
8736 return tg_set_cfs_bandwidth(tg, period, quota);
8739 long tg_get_cfs_period(struct task_group *tg)
8743 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period);
8744 do_div(cfs_period_us, NSEC_PER_USEC);
8746 return cfs_period_us;
8749 static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css,
8752 return tg_get_cfs_quota(css_tg(css));
8755 static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css,
8756 struct cftype *cftype, s64 cfs_quota_us)
8758 return tg_set_cfs_quota(css_tg(css), cfs_quota_us);
8761 static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css,
8764 return tg_get_cfs_period(css_tg(css));
8767 static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css,
8768 struct cftype *cftype, u64 cfs_period_us)
8770 return tg_set_cfs_period(css_tg(css), cfs_period_us);
8773 struct cfs_schedulable_data {
8774 struct task_group *tg;
8779 * normalize group quota/period to be quota/max_period
8780 * note: units are usecs
8782 static u64 normalize_cfs_quota(struct task_group *tg,
8783 struct cfs_schedulable_data *d)
8791 period = tg_get_cfs_period(tg);
8792 quota = tg_get_cfs_quota(tg);
8795 /* note: these should typically be equivalent */
8796 if (quota == RUNTIME_INF || quota == -1)
8799 return to_ratio(period, quota);
8802 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
8804 struct cfs_schedulable_data *d = data;
8805 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8806 s64 quota = 0, parent_quota = -1;
8809 quota = RUNTIME_INF;
8811 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth;
8813 quota = normalize_cfs_quota(tg, d);
8814 parent_quota = parent_b->hierarchical_quota;
8817 * ensure max(child_quota) <= parent_quota, inherit when no
8820 if (quota == RUNTIME_INF)
8821 quota = parent_quota;
8822 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
8825 cfs_b->hierarchical_quota = quota;
8830 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
8833 struct cfs_schedulable_data data = {
8839 if (quota != RUNTIME_INF) {
8840 do_div(data.period, NSEC_PER_USEC);
8841 do_div(data.quota, NSEC_PER_USEC);
8845 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
8851 static int cpu_stats_show(struct seq_file *sf, void *v)
8853 struct task_group *tg = css_tg(seq_css(sf));
8854 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth;
8856 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods);
8857 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled);
8858 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time);
8862 #endif /* CONFIG_CFS_BANDWIDTH */
8863 #endif /* CONFIG_FAIR_GROUP_SCHED */
8865 #ifdef CONFIG_RT_GROUP_SCHED
8866 static int cpu_rt_runtime_write(struct cgroup_subsys_state *css,
8867 struct cftype *cft, s64 val)
8869 return sched_group_set_rt_runtime(css_tg(css), val);
8872 static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css,
8875 return sched_group_rt_runtime(css_tg(css));
8878 static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css,
8879 struct cftype *cftype, u64 rt_period_us)
8881 return sched_group_set_rt_period(css_tg(css), rt_period_us);
8884 static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css,
8887 return sched_group_rt_period(css_tg(css));
8889 #endif /* CONFIG_RT_GROUP_SCHED */
8891 static struct cftype cpu_files[] = {
8892 #ifdef CONFIG_FAIR_GROUP_SCHED
8895 .read_u64 = cpu_shares_read_u64,
8896 .write_u64 = cpu_shares_write_u64,
8899 #ifdef CONFIG_CFS_BANDWIDTH
8901 .name = "cfs_quota_us",
8902 .read_s64 = cpu_cfs_quota_read_s64,
8903 .write_s64 = cpu_cfs_quota_write_s64,
8906 .name = "cfs_period_us",
8907 .read_u64 = cpu_cfs_period_read_u64,
8908 .write_u64 = cpu_cfs_period_write_u64,
8912 .seq_show = cpu_stats_show,
8915 #ifdef CONFIG_RT_GROUP_SCHED
8917 .name = "rt_runtime_us",
8918 .read_s64 = cpu_rt_runtime_read,
8919 .write_s64 = cpu_rt_runtime_write,
8922 .name = "rt_period_us",
8923 .read_u64 = cpu_rt_period_read_uint,
8924 .write_u64 = cpu_rt_period_write_uint,
8930 struct cgroup_subsys cpu_cgrp_subsys = {
8931 .css_alloc = cpu_cgroup_css_alloc,
8932 .css_released = cpu_cgroup_css_released,
8933 .css_free = cpu_cgroup_css_free,
8934 .fork = cpu_cgroup_fork,
8935 .can_attach = cpu_cgroup_can_attach,
8936 .attach = cpu_cgroup_attach,
8937 .allow_attach = subsys_cgroup_allow_attach,
8938 .legacy_cftypes = cpu_files,
8942 #endif /* CONFIG_CGROUP_SCHED */
8944 void dump_cpu_task(int cpu)
8946 pr_info("Task dump for CPU %d:\n", cpu);
8947 sched_show_task(cpu_curr(cpu));